]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * | |
8 | * This is free software; you can redistribute it and/or | |
9 | * modify it under the terms of the GNU Lesser General Public | |
10 | * License version 2.1, as published by the Free Software | |
11 | * Foundation. See file COPYING. | |
12 | * | |
13 | */ | |
14 | ||
15 | #include <boost/lexical_cast.hpp> | |
11fdf7f2 | 16 | #include "include/ceph_assert.h" // lexical_cast includes system assert.h |
7c673cae FG |
17 | |
18 | #include <boost/config/warning_disable.hpp> | |
19 | #include <boost/fusion/include/std_pair.hpp> | |
a8e16298 | 20 | #include <boost/range/adaptor/reversed.hpp> |
7c673cae FG |
21 | |
22 | #include "MDSRank.h" | |
23 | #include "Server.h" | |
24 | #include "Locker.h" | |
25 | #include "MDCache.h" | |
26 | #include "MDLog.h" | |
27 | #include "Migrator.h" | |
28 | #include "MDBalancer.h" | |
29 | #include "InoTable.h" | |
30 | #include "SnapClient.h" | |
31 | #include "Mutation.h" | |
11fdf7f2 | 32 | #include "cephfs_features.h" |
7c673cae FG |
33 | |
34 | #include "msg/Messenger.h" | |
35 | ||
36 | #include "osdc/Objecter.h" | |
37 | ||
7c673cae FG |
38 | #include "events/EUpdate.h" |
39 | #include "events/ESlaveUpdate.h" | |
40 | #include "events/ESession.h" | |
41 | #include "events/EOpen.h" | |
42 | #include "events/ECommitted.h" | |
43 | ||
11fdf7f2 | 44 | #include "include/stringify.h" |
7c673cae FG |
45 | #include "include/filepath.h" |
46 | #include "common/errno.h" | |
47 | #include "common/Timer.h" | |
48 | #include "common/perf_counters.h" | |
49 | #include "include/compat.h" | |
50 | #include "osd/OSDMap.h" | |
51 | ||
52 | #include <errno.h> | |
a8e16298 | 53 | #include <math.h> |
7c673cae FG |
54 | |
55 | #include <list> | |
56 | #include <iostream> | |
11fdf7f2 | 57 | #include <string_view> |
7c673cae FG |
58 | |
59 | #include "common/config.h" | |
60 | ||
61 | #define dout_context g_ceph_context | |
62 | #define dout_subsys ceph_subsys_mds | |
63 | #undef dout_prefix | |
64 | #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".server " | |
65 | ||
11fdf7f2 | 66 | class ServerContext : public MDSContext { |
7c673cae FG |
67 | protected: |
68 | Server *server; | |
69 | MDSRank *get_mds() override | |
70 | { | |
71 | return server->mds; | |
72 | } | |
73 | ||
74 | public: | |
75 | explicit ServerContext(Server *s) : server(s) { | |
11fdf7f2 | 76 | ceph_assert(server != NULL); |
7c673cae FG |
77 | } |
78 | }; | |
79 | ||
80 | class ServerLogContext : public MDSLogContextBase { | |
81 | protected: | |
82 | Server *server; | |
83 | MDSRank *get_mds() override | |
84 | { | |
85 | return server->mds; | |
86 | } | |
87 | ||
88 | MDRequestRef mdr; | |
89 | void pre_finish(int r) override { | |
90 | if (mdr) | |
91 | mdr->mark_event("journal_committed: "); | |
92 | } | |
93 | public: | |
94 | explicit ServerLogContext(Server *s) : server(s) { | |
11fdf7f2 | 95 | ceph_assert(server != NULL); |
7c673cae FG |
96 | } |
97 | explicit ServerLogContext(Server *s, MDRequestRef& r) : server(s), mdr(r) { | |
11fdf7f2 | 98 | ceph_assert(server != NULL); |
7c673cae FG |
99 | } |
100 | }; | |
101 | ||
102 | void Server::create_logger() | |
103 | { | |
104 | PerfCountersBuilder plb(g_ceph_context, "mds_server", l_mdss_first, l_mdss_last); | |
91327a77 AA |
105 | |
106 | plb.add_u64_counter(l_mdss_handle_client_request, "handle_client_request", | |
107 | "Client requests", "hcr", PerfCountersBuilder::PRIO_INTERESTING); | |
7c673cae | 108 | plb.add_u64_counter(l_mdss_handle_slave_request, "handle_slave_request", |
91327a77 AA |
109 | "Slave requests", "hsr", PerfCountersBuilder::PRIO_INTERESTING); |
110 | plb.add_u64_counter(l_mdss_handle_client_session, | |
111 | "handle_client_session", "Client session messages", "hcs", | |
112 | PerfCountersBuilder::PRIO_INTERESTING); | |
11fdf7f2 TL |
113 | plb.add_u64_counter(l_mdss_cap_revoke_eviction, "cap_revoke_eviction", |
114 | "Cap Revoke Client Eviction", "cre", PerfCountersBuilder::PRIO_INTERESTING); | |
91327a77 AA |
115 | |
116 | // fop latencies are useful | |
117 | plb.set_prio_default(PerfCountersBuilder::PRIO_USEFUL); | |
118 | plb.add_time_avg(l_mdss_req_lookuphash_latency, "req_lookuphash_latency", | |
119 | "Request type lookup hash of inode latency"); | |
120 | plb.add_time_avg(l_mdss_req_lookupino_latency, "req_lookupino_latency", | |
121 | "Request type lookup inode latency"); | |
122 | plb.add_time_avg(l_mdss_req_lookupparent_latency, "req_lookupparent_latency", | |
123 | "Request type lookup parent latency"); | |
124 | plb.add_time_avg(l_mdss_req_lookupname_latency, "req_lookupname_latency", | |
125 | "Request type lookup name latency"); | |
126 | plb.add_time_avg(l_mdss_req_lookup_latency, "req_lookup_latency", | |
127 | "Request type lookup latency"); | |
128 | plb.add_time_avg(l_mdss_req_lookupsnap_latency, "req_lookupsnap_latency", | |
129 | "Request type lookup snapshot latency"); | |
130 | plb.add_time_avg(l_mdss_req_getattr_latency, "req_getattr_latency", | |
131 | "Request type get attribute latency"); | |
132 | plb.add_time_avg(l_mdss_req_setattr_latency, "req_setattr_latency", | |
133 | "Request type set attribute latency"); | |
134 | plb.add_time_avg(l_mdss_req_setlayout_latency, "req_setlayout_latency", | |
135 | "Request type set file layout latency"); | |
136 | plb.add_time_avg(l_mdss_req_setdirlayout_latency, "req_setdirlayout_latency", | |
137 | "Request type set directory layout latency"); | |
138 | plb.add_time_avg(l_mdss_req_setxattr_latency, "req_setxattr_latency", | |
139 | "Request type set extended attribute latency"); | |
140 | plb.add_time_avg(l_mdss_req_rmxattr_latency, "req_rmxattr_latency", | |
141 | "Request type remove extended attribute latency"); | |
142 | plb.add_time_avg(l_mdss_req_readdir_latency, "req_readdir_latency", | |
143 | "Request type read directory latency"); | |
144 | plb.add_time_avg(l_mdss_req_setfilelock_latency, "req_setfilelock_latency", | |
145 | "Request type set file lock latency"); | |
146 | plb.add_time_avg(l_mdss_req_getfilelock_latency, "req_getfilelock_latency", | |
147 | "Request type get file lock latency"); | |
148 | plb.add_time_avg(l_mdss_req_create_latency, "req_create_latency", | |
149 | "Request type create latency"); | |
150 | plb.add_time_avg(l_mdss_req_open_latency, "req_open_latency", | |
151 | "Request type open latency"); | |
152 | plb.add_time_avg(l_mdss_req_mknod_latency, "req_mknod_latency", | |
153 | "Request type make node latency"); | |
154 | plb.add_time_avg(l_mdss_req_link_latency, "req_link_latency", | |
155 | "Request type link latency"); | |
156 | plb.add_time_avg(l_mdss_req_unlink_latency, "req_unlink_latency", | |
157 | "Request type unlink latency"); | |
158 | plb.add_time_avg(l_mdss_req_rmdir_latency, "req_rmdir_latency", | |
159 | "Request type remove directory latency"); | |
160 | plb.add_time_avg(l_mdss_req_rename_latency, "req_rename_latency", | |
161 | "Request type rename latency"); | |
162 | plb.add_time_avg(l_mdss_req_mkdir_latency, "req_mkdir_latency", | |
163 | "Request type make directory latency"); | |
164 | plb.add_time_avg(l_mdss_req_symlink_latency, "req_symlink_latency", | |
165 | "Request type symbolic link latency"); | |
166 | plb.add_time_avg(l_mdss_req_lssnap_latency, "req_lssnap_latency", | |
167 | "Request type list snapshot latency"); | |
168 | plb.add_time_avg(l_mdss_req_mksnap_latency, "req_mksnap_latency", | |
169 | "Request type make snapshot latency"); | |
170 | plb.add_time_avg(l_mdss_req_rmsnap_latency, "req_rmsnap_latency", | |
171 | "Request type remove snapshot latency"); | |
172 | plb.add_time_avg(l_mdss_req_renamesnap_latency, "req_renamesnap_latency", | |
173 | "Request type rename snapshot latency"); | |
174 | ||
91327a77 AA |
175 | plb.set_prio_default(PerfCountersBuilder::PRIO_DEBUGONLY); |
176 | plb.add_u64_counter(l_mdss_dispatch_client_request, "dispatch_client_request", | |
177 | "Client requests dispatched"); | |
178 | plb.add_u64_counter(l_mdss_dispatch_slave_request, "dispatch_server_request", | |
179 | "Server requests dispatched"); | |
180 | ||
7c673cae FG |
181 | logger = plb.create_perf_counters(); |
182 | g_ceph_context->get_perfcounters_collection()->add(logger); | |
183 | } | |
184 | ||
185 | Server::Server(MDSRank *m) : | |
186 | mds(m), | |
187 | mdcache(mds->mdcache), mdlog(mds->mdlog), | |
188 | logger(0), | |
189 | is_full(false), | |
190 | reconnect_done(NULL), | |
191 | failed_reconnects(0), | |
31f18b77 | 192 | reconnect_evicting(false), |
a8e16298 | 193 | terminating_sessions(false), |
11fdf7f2 | 194 | recall_throttle(g_conf().get_val<double>("mds_recall_max_decay_rate")) |
7c673cae | 195 | { |
11fdf7f2 | 196 | supported_features = feature_bitset_t(CEPHFS_FEATURES_MDS_SUPPORTED); |
7c673cae FG |
197 | } |
198 | ||
11fdf7f2 | 199 | void Server::dispatch(const Message::const_ref &m) |
7c673cae FG |
200 | { |
201 | switch (m->get_type()) { | |
202 | case CEPH_MSG_CLIENT_RECONNECT: | |
11fdf7f2 | 203 | handle_client_reconnect(MClientReconnect::msgref_cast(m)); |
7c673cae FG |
204 | return; |
205 | } | |
206 | ||
207 | // active? | |
94b18763 FG |
208 | // handle_slave_request()/handle_client_session() will wait if necessary |
209 | if (m->get_type() == CEPH_MSG_CLIENT_REQUEST && !mds->is_active()) { | |
11fdf7f2 | 210 | const auto &req = MClientRequest::msgref_cast(m); |
94b18763 FG |
211 | if (mds->is_reconnect() || mds->get_want_state() == CEPH_MDS_STATE_RECONNECT) { |
212 | Session *session = mds->get_session(req); | |
7c673cae FG |
213 | if (!session || session->is_closed()) { |
214 | dout(5) << "session is closed, dropping " << req->get_reqid() << dendl; | |
7c673cae FG |
215 | return; |
216 | } | |
217 | bool queue_replay = false; | |
218 | if (req->is_replay()) { | |
219 | dout(3) << "queuing replayed op" << dendl; | |
220 | queue_replay = true; | |
11fdf7f2 TL |
221 | if (req->head.ino && |
222 | !session->have_completed_request(req->get_reqid().tid, nullptr)) { | |
223 | mdcache->add_replay_ino_alloc(inodeno_t(req->head.ino)); | |
224 | } | |
7c673cae FG |
225 | } else if (req->get_retry_attempt()) { |
226 | // process completed request in clientreplay stage. The completed request | |
227 | // might have created new file/directorie. This guarantees MDS sends a reply | |
228 | // to client before other request modifies the new file/directorie. | |
229 | if (session->have_completed_request(req->get_reqid().tid, NULL)) { | |
230 | dout(3) << "queuing completed op" << dendl; | |
231 | queue_replay = true; | |
232 | } | |
233 | // this request was created before the cap reconnect message, drop any embedded | |
234 | // cap releases. | |
235 | req->releases.clear(); | |
236 | } | |
237 | if (queue_replay) { | |
238 | req->mark_queued_for_replay(); | |
239 | mds->enqueue_replay(new C_MDS_RetryMessage(mds, m)); | |
240 | return; | |
241 | } | |
242 | } | |
243 | ||
244 | bool wait_for_active = true; | |
94b18763 | 245 | if (mds->is_stopping()) { |
28e407b8 | 246 | wait_for_active = false; |
7c673cae | 247 | } else if (mds->is_clientreplay()) { |
94b18763 | 248 | if (req->is_queued_for_replay()) { |
7c673cae | 249 | wait_for_active = false; |
7c673cae FG |
250 | } |
251 | } | |
252 | if (wait_for_active) { | |
253 | dout(3) << "not active yet, waiting" << dendl; | |
254 | mds->wait_for_active(new C_MDS_RetryMessage(mds, m)); | |
255 | return; | |
256 | } | |
257 | } | |
258 | ||
259 | switch (m->get_type()) { | |
260 | case CEPH_MSG_CLIENT_SESSION: | |
11fdf7f2 | 261 | handle_client_session(MClientSession::msgref_cast(m)); |
7c673cae FG |
262 | return; |
263 | case CEPH_MSG_CLIENT_REQUEST: | |
11fdf7f2 TL |
264 | handle_client_request(MClientRequest::msgref_cast(m)); |
265 | return; | |
266 | case CEPH_MSG_CLIENT_RECLAIM: | |
267 | handle_client_reclaim(MClientReclaim::msgref_cast(m)); | |
7c673cae FG |
268 | return; |
269 | case MSG_MDS_SLAVE_REQUEST: | |
11fdf7f2 | 270 | handle_slave_request(MMDSSlaveRequest::msgref_cast(m)); |
7c673cae FG |
271 | return; |
272 | default: | |
273 | derr << "server unknown message " << m->get_type() << dendl; | |
11fdf7f2 | 274 | ceph_abort_msg("server unknown message"); |
7c673cae FG |
275 | } |
276 | } | |
277 | ||
278 | ||
279 | ||
280 | // ---------------------------------------------------------- | |
281 | // SESSION management | |
282 | ||
283 | class C_MDS_session_finish : public ServerLogContext { | |
284 | Session *session; | |
285 | uint64_t state_seq; | |
286 | bool open; | |
287 | version_t cmapv; | |
288 | interval_set<inodeno_t> inos; | |
289 | version_t inotablev; | |
290 | Context *fin; | |
291 | public: | |
292 | C_MDS_session_finish(Server *srv, Session *se, uint64_t sseq, bool s, version_t mv, Context *fin_ = NULL) : | |
293 | ServerLogContext(srv), session(se), state_seq(sseq), open(s), cmapv(mv), inotablev(0), fin(fin_) { } | |
294 | C_MDS_session_finish(Server *srv, Session *se, uint64_t sseq, bool s, version_t mv, interval_set<inodeno_t>& i, version_t iv, Context *fin_ = NULL) : | |
295 | ServerLogContext(srv), session(se), state_seq(sseq), open(s), cmapv(mv), inos(i), inotablev(iv), fin(fin_) { } | |
296 | void finish(int r) override { | |
11fdf7f2 | 297 | ceph_assert(r == 0); |
7c673cae FG |
298 | server->_session_logged(session, state_seq, open, cmapv, inos, inotablev); |
299 | if (fin) { | |
300 | fin->complete(r); | |
301 | } | |
302 | } | |
303 | }; | |
304 | ||
11fdf7f2 TL |
305 | Session* Server::find_session_by_uuid(std::string_view uuid) |
306 | { | |
307 | Session* session = nullptr; | |
308 | for (auto& it : mds->sessionmap.get_sessions()) { | |
309 | auto& metadata = it.second->info.client_metadata; | |
310 | ||
311 | auto p = metadata.find("uuid"); | |
312 | if (p == metadata.end() || p->second != uuid) | |
313 | continue; | |
314 | ||
315 | if (!session) { | |
316 | session = it.second; | |
317 | } else if (!session->reclaiming_from) { | |
318 | assert(it.second->reclaiming_from == session); | |
319 | session = it.second; | |
320 | } else { | |
321 | assert(session->reclaiming_from == it.second); | |
322 | } | |
323 | } | |
324 | return session; | |
325 | } | |
326 | ||
327 | void Server::reclaim_session(Session *session, const MClientReclaim::const_ref &m) | |
328 | { | |
329 | if (!session->is_open() && !session->is_stale()) { | |
330 | dout(10) << "session not open, dropping this req" << dendl; | |
331 | return; | |
332 | } | |
333 | ||
334 | auto reply = MClientReclaimReply::create(0); | |
335 | if (m->get_uuid().empty()) { | |
336 | dout(10) << __func__ << " invalid message (no uuid)" << dendl; | |
337 | reply->set_result(-EINVAL); | |
338 | mds->send_message_client(reply, session); | |
339 | return; | |
340 | } | |
341 | ||
342 | unsigned flags = m->get_flags(); | |
343 | if (flags != CEPH_RECLAIM_RESET) { // currently only support reset | |
344 | dout(10) << __func__ << " unsupported flags" << dendl; | |
345 | reply->set_result(-EOPNOTSUPP); | |
346 | mds->send_message_client(reply, session); | |
347 | return; | |
348 | } | |
349 | ||
350 | Session* target = find_session_by_uuid(m->get_uuid()); | |
351 | if (target) { | |
352 | if (session->info.auth_name != target->info.auth_name) { | |
353 | dout(10) << __func__ << " session auth_name " << session->info.auth_name | |
354 | << " != target auth_name " << target->info.auth_name << dendl; | |
355 | reply->set_result(-EPERM); | |
356 | mds->send_message_client(reply, session); | |
357 | } | |
358 | ||
359 | assert(!target->reclaiming_from); | |
360 | assert(!session->reclaiming_from); | |
361 | session->reclaiming_from = target; | |
362 | reply->set_addrs(entity_addrvec_t(target->info.inst.addr)); | |
363 | } | |
364 | ||
365 | if (flags & CEPH_RECLAIM_RESET) { | |
366 | finish_reclaim_session(session, reply); | |
367 | return; | |
368 | } | |
369 | ||
370 | ceph_abort(); | |
371 | } | |
372 | ||
373 | void Server::finish_reclaim_session(Session *session, const MClientReclaimReply::ref &reply) | |
374 | { | |
375 | Session *target = session->reclaiming_from; | |
376 | if (target) { | |
377 | session->reclaiming_from = nullptr; | |
378 | ||
379 | Context *send_reply; | |
380 | if (reply) { | |
381 | int64_t session_id = session->get_client().v; | |
382 | send_reply = new FunctionContext([this, session_id, reply](int r) { | |
383 | assert(mds->mds_lock.is_locked_by_me()); | |
384 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(session_id)); | |
385 | if (!session) { | |
386 | return; | |
387 | } | |
388 | auto epoch = mds->objecter->with_osdmap([](const OSDMap &map){ return map.get_epoch(); }); | |
389 | reply->set_epoch(epoch); | |
390 | mds->send_message_client(reply, session); | |
391 | }); | |
392 | } else { | |
393 | send_reply = nullptr; | |
394 | } | |
395 | ||
396 | bool blacklisted = mds->objecter->with_osdmap([target](const OSDMap &map) { | |
397 | return map.is_blacklisted(target->info.inst.addr); | |
398 | }); | |
399 | ||
400 | if (blacklisted || !g_conf()->mds_session_blacklist_on_evict) { | |
401 | kill_session(target, send_reply); | |
402 | } else { | |
403 | std::stringstream ss; | |
404 | mds->evict_client(target->get_client().v, false, true, ss, send_reply); | |
405 | } | |
406 | } else if (reply) { | |
407 | mds->send_message_client(reply, session); | |
408 | } | |
409 | } | |
410 | ||
411 | void Server::handle_client_reclaim(const MClientReclaim::const_ref &m) | |
412 | { | |
413 | Session *session = mds->get_session(m); | |
414 | dout(3) << __func__ << " " << *m << " from " << m->get_source() << dendl; | |
415 | assert(m->get_source().is_client()); // should _not_ come from an mds! | |
416 | ||
417 | if (!session) { | |
418 | dout(0) << " ignoring sessionless msg " << *m << dendl; | |
419 | return; | |
420 | } | |
421 | ||
422 | if (mds->get_state() < MDSMap::STATE_CLIENTREPLAY) { | |
423 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, m)); | |
424 | return; | |
425 | } | |
426 | ||
427 | if (m->get_flags() & MClientReclaim::FLAG_FINISH) { | |
428 | finish_reclaim_session(session); | |
429 | } else { | |
430 | reclaim_session(session, m); | |
431 | } | |
432 | } | |
433 | ||
434 | void Server::handle_client_session(const MClientSession::const_ref &m) | |
7c673cae FG |
435 | { |
436 | version_t pv; | |
94b18763 | 437 | Session *session = mds->get_session(m); |
7c673cae FG |
438 | |
439 | dout(3) << "handle_client_session " << *m << " from " << m->get_source() << dendl; | |
11fdf7f2 | 440 | ceph_assert(m->get_source().is_client()); // should _not_ come from an mds! |
7c673cae FG |
441 | |
442 | if (!session) { | |
443 | dout(0) << " ignoring sessionless msg " << *m << dendl; | |
7c673cae FG |
444 | return; |
445 | } | |
446 | ||
94b18763 FG |
447 | if (m->get_op() == CEPH_SESSION_REQUEST_RENEWCAPS) { |
448 | // always handle renewcaps (state >= MDSMap::STATE_RECONNECT) | |
449 | } else if (m->get_op() == CEPH_SESSION_REQUEST_CLOSE) { | |
450 | // close requests need to be handled when mds is active | |
451 | if (mds->get_state() < MDSMap::STATE_ACTIVE) { | |
452 | mds->wait_for_active(new C_MDS_RetryMessage(mds, m)); | |
453 | return; | |
454 | } | |
455 | } else { | |
456 | if (mds->get_state() < MDSMap::STATE_CLIENTREPLAY) { | |
457 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, m)); | |
458 | return; | |
459 | } | |
460 | } | |
461 | ||
7c673cae FG |
462 | if (logger) |
463 | logger->inc(l_mdss_handle_client_session); | |
464 | ||
465 | uint64_t sseq = 0; | |
466 | switch (m->get_op()) { | |
467 | case CEPH_SESSION_REQUEST_OPEN: | |
468 | if (session->is_opening() || | |
469 | session->is_open() || | |
470 | session->is_stale() || | |
28e407b8 AA |
471 | session->is_killing() || |
472 | terminating_sessions) { | |
7c673cae | 473 | dout(10) << "currently open|opening|stale|killing, dropping this req" << dendl; |
7c673cae FG |
474 | return; |
475 | } | |
11fdf7f2 | 476 | ceph_assert(session->is_closed() || session->is_closing()); |
7c673cae | 477 | |
b32b8144 FG |
478 | if (mds->is_stopping()) { |
479 | dout(10) << "mds is stopping, dropping open req" << dendl; | |
b32b8144 FG |
480 | return; |
481 | } | |
482 | ||
a8e16298 TL |
483 | { |
484 | auto& addr = session->info.inst.addr; | |
11fdf7f2 | 485 | session->set_client_metadata(client_metadata_t(m->metadata, m->supported_features)); |
a8e16298 TL |
486 | auto& client_metadata = session->info.client_metadata; |
487 | ||
11fdf7f2 | 488 | auto log_session_status = [this, m, session](std::string_view status, std::string_view err) { |
a8e16298 TL |
489 | auto now = ceph_clock_now(); |
490 | auto throttle_elapsed = m->get_recv_complete_stamp() - m->get_throttle_stamp(); | |
491 | auto elapsed = now - m->get_recv_stamp(); | |
11fdf7f2 TL |
492 | CachedStackStringStream css; |
493 | *css << "New client session:" | |
a8e16298 TL |
494 | << " addr=\"" << session->info.inst.addr << "\"" |
495 | << ",elapsed=" << elapsed | |
496 | << ",throttled=" << throttle_elapsed | |
497 | << ",status=\"" << status << "\""; | |
498 | if (!err.empty()) { | |
11fdf7f2 | 499 | *css << ",error=\"" << err << "\""; |
a8e16298 TL |
500 | } |
501 | const auto& metadata = session->info.client_metadata; | |
11fdf7f2 TL |
502 | if (auto it = metadata.find("root"); it != metadata.end()) { |
503 | *css << ",root=\"" << it->second << "\""; | |
a8e16298 | 504 | } |
11fdf7f2 TL |
505 | dout(2) << css->strv() << dendl; |
506 | }; | |
507 | ||
508 | auto send_reject_message = [this, &session, &log_session_status](std::string_view err_str) { | |
509 | auto m = MClientSession::create(CEPH_SESSION_REJECT); | |
510 | if (session->info.has_feature(CEPHFS_FEATURE_MIMIC)) | |
511 | m->metadata["error_string"] = err_str; | |
512 | mds->send_message_client(m, session); | |
513 | log_session_status("REJECTED", err_str); | |
a8e16298 | 514 | }; |
7c673cae | 515 | |
11fdf7f2 TL |
516 | bool blacklisted = mds->objecter->with_osdmap( |
517 | [&addr](const OSDMap &osd_map) -> bool { | |
518 | return osd_map.is_blacklisted(addr); | |
519 | }); | |
520 | ||
a8e16298 | 521 | if (blacklisted) { |
11fdf7f2 TL |
522 | dout(10) << "rejecting blacklisted client " << addr << dendl; |
523 | send_reject_message("blacklisted"); | |
524 | session->clear(); | |
525 | break; | |
7c673cae | 526 | } |
7c673cae | 527 | |
11fdf7f2 TL |
528 | if (client_metadata.features.empty()) |
529 | infer_supported_features(session, client_metadata); | |
530 | ||
531 | dout(20) << __func__ << " CEPH_SESSION_REQUEST_OPEN metadata entries:" << dendl; | |
532 | dout(20) << " features: '" << client_metadata.features << dendl; | |
533 | for (const auto& p : client_metadata) { | |
534 | dout(20) << " " << p.first << ": " << p.second << dendl; | |
535 | } | |
536 | ||
537 | feature_bitset_t missing_features = required_client_features; | |
538 | missing_features -= client_metadata.features; | |
539 | if (!missing_features.empty()) { | |
540 | stringstream ss; | |
541 | ss << "missing required features '" << missing_features << "'"; | |
542 | send_reject_message(ss.str()); | |
543 | mds->clog->warn() << "client session lacks required features '" | |
544 | << missing_features << "' denied (" << session->info.inst << ")"; | |
545 | session->clear(); | |
546 | break; | |
a8e16298 | 547 | } |
7c673cae | 548 | |
a8e16298 TL |
549 | // Special case for the 'root' metadata path; validate that the claimed |
550 | // root is actually within the caps of the session | |
11fdf7f2 TL |
551 | if (auto it = client_metadata.find("root"); it != client_metadata.end()) { |
552 | auto claimed_root = it->second; | |
553 | stringstream ss; | |
554 | bool denied = false; | |
555 | // claimed_root has a leading "/" which we strip before passing | |
556 | // into caps check | |
557 | if (claimed_root.empty() || claimed_root[0] != '/') { | |
558 | denied = true; | |
559 | ss << "invalue root '" << claimed_root << "'"; | |
560 | } else if (!session->auth_caps.path_capable(claimed_root.substr(1))) { | |
561 | denied = true; | |
562 | ss << "non-allowable root '" << claimed_root << "'"; | |
563 | } | |
564 | ||
565 | if (denied) { | |
566 | // Tell the client we're rejecting their open | |
567 | send_reject_message(ss.str()); | |
568 | mds->clog->warn() << "client session with " << ss.str() | |
569 | << " denied (" << session->info.inst << ")"; | |
570 | session->clear(); | |
571 | break; | |
572 | } | |
573 | } | |
574 | ||
575 | if (auto it = client_metadata.find("uuid"); it != client_metadata.end()) { | |
576 | if (find_session_by_uuid(it->second)) { | |
577 | send_reject_message("duplicated session uuid"); | |
578 | mds->clog->warn() << "client session with duplicated session uuid '" | |
579 | << it->second << "' denied (" << session->info.inst << ")"; | |
580 | session->clear(); | |
581 | break; | |
582 | } | |
a8e16298 TL |
583 | } |
584 | ||
585 | if (session->is_closed()) | |
11fdf7f2 | 586 | mds->sessionmap.add_session(session); |
a8e16298 TL |
587 | |
588 | pv = mds->sessionmap.mark_projected(session); | |
589 | sseq = mds->sessionmap.set_state(session, Session::STATE_OPENING); | |
590 | mds->sessionmap.touch_session(session); | |
591 | auto fin = new FunctionContext([log_session_status = std::move(log_session_status)](int r){ | |
11fdf7f2 | 592 | ceph_assert(r == 0); |
a8e16298 TL |
593 | log_session_status("ACCEPTED", ""); |
594 | }); | |
595 | mdlog->start_submit_entry(new ESession(m->get_source_inst(), true, pv, client_metadata), | |
596 | new C_MDS_session_finish(this, session, sseq, true, pv, fin)); | |
11fdf7f2 | 597 | mdlog->flush(); |
a8e16298 | 598 | } |
7c673cae FG |
599 | break; |
600 | ||
601 | case CEPH_SESSION_REQUEST_RENEWCAPS: | |
11fdf7f2 | 602 | if (session->is_open() || session->is_stale()) { |
7c673cae FG |
603 | mds->sessionmap.touch_session(session); |
604 | if (session->is_stale()) { | |
605 | mds->sessionmap.set_state(session, Session::STATE_OPEN); | |
606 | mds->locker->resume_stale_caps(session); | |
607 | mds->sessionmap.touch_session(session); | |
608 | } | |
11fdf7f2 TL |
609 | auto reply = MClientSession::create(CEPH_SESSION_RENEWCAPS, m->get_seq()); |
610 | mds->send_message_client(reply, session); | |
7c673cae FG |
611 | } else { |
612 | dout(10) << "ignoring renewcaps on non open|stale session (" << session->get_state_name() << ")" << dendl; | |
613 | } | |
614 | break; | |
615 | ||
616 | case CEPH_SESSION_REQUEST_CLOSE: | |
617 | { | |
618 | if (session->is_closed() || | |
619 | session->is_closing() || | |
620 | session->is_killing()) { | |
621 | dout(10) << "already closed|closing|killing, dropping this req" << dendl; | |
7c673cae FG |
622 | return; |
623 | } | |
624 | if (session->is_importing()) { | |
625 | dout(10) << "ignoring close req on importing session" << dendl; | |
7c673cae FG |
626 | return; |
627 | } | |
11fdf7f2 | 628 | ceph_assert(session->is_open() || |
7c673cae FG |
629 | session->is_stale() || |
630 | session->is_opening()); | |
631 | if (m->get_seq() < session->get_push_seq()) { | |
632 | dout(10) << "old push seq " << m->get_seq() << " < " << session->get_push_seq() | |
633 | << ", dropping" << dendl; | |
7c673cae FG |
634 | return; |
635 | } | |
636 | // We are getting a seq that is higher than expected. | |
637 | // Handle the same as any other seqn error. | |
638 | // | |
639 | if (m->get_seq() != session->get_push_seq()) { | |
640 | dout(0) << "old push seq " << m->get_seq() << " != " << session->get_push_seq() | |
641 | << ", BUGGY!" << dendl; | |
642 | mds->clog->warn() << "incorrect push seq " << m->get_seq() << " != " | |
643 | << session->get_push_seq() << ", dropping" << " from client : " << session->get_human_name(); | |
7c673cae FG |
644 | return; |
645 | } | |
646 | journal_close_session(session, Session::STATE_CLOSING, NULL); | |
647 | } | |
648 | break; | |
649 | ||
650 | case CEPH_SESSION_FLUSHMSG_ACK: | |
651 | finish_flush_session(session, m->get_seq()); | |
652 | break; | |
653 | ||
31f18b77 | 654 | case CEPH_SESSION_REQUEST_FLUSH_MDLOG: |
b32b8144 FG |
655 | if (mds->is_active()) |
656 | mdlog->flush(); | |
31f18b77 FG |
657 | break; |
658 | ||
7c673cae FG |
659 | default: |
660 | ceph_abort(); | |
661 | } | |
7c673cae FG |
662 | } |
663 | ||
f64942e4 AA |
664 | |
665 | void Server::flush_session(Session *session, MDSGatherBuilder *gather) { | |
666 | if (!session->is_open() || | |
11fdf7f2 TL |
667 | !session->get_connection() || |
668 | !session->get_connection()->has_feature(CEPH_FEATURE_EXPORT_PEER)) { | |
f64942e4 AA |
669 | return; |
670 | } | |
671 | ||
672 | version_t seq = session->wait_for_flush(gather->new_sub()); | |
11fdf7f2 TL |
673 | mds->send_message_client( |
674 | MClientSession::create(CEPH_SESSION_FLUSHMSG, seq), session); | |
f64942e4 AA |
675 | } |
676 | ||
7c673cae FG |
677 | void Server::flush_client_sessions(set<client_t>& client_set, MDSGatherBuilder& gather) |
678 | { | |
679 | for (set<client_t>::iterator p = client_set.begin(); p != client_set.end(); ++p) { | |
680 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(p->v)); | |
11fdf7f2 | 681 | ceph_assert(session); |
f64942e4 | 682 | flush_session(session, &gather); |
7c673cae FG |
683 | } |
684 | } | |
685 | ||
686 | void Server::finish_flush_session(Session *session, version_t seq) | |
687 | { | |
11fdf7f2 | 688 | MDSContext::vec finished; |
7c673cae FG |
689 | session->finish_flush(seq, finished); |
690 | mds->queue_waiters(finished); | |
691 | } | |
692 | ||
693 | void Server::_session_logged(Session *session, uint64_t state_seq, bool open, version_t pv, | |
694 | interval_set<inodeno_t>& inos, version_t piv) | |
695 | { | |
696 | dout(10) << "_session_logged " << session->info.inst << " state_seq " << state_seq << " " << (open ? "open":"close") | |
697 | << " " << pv << dendl; | |
698 | ||
699 | if (piv) { | |
11fdf7f2 | 700 | ceph_assert(session->is_closing() || session->is_killing() || |
7c673cae FG |
701 | session->is_opening()); // re-open closing session |
702 | session->info.prealloc_inos.subtract(inos); | |
703 | mds->inotable->apply_release_ids(inos); | |
11fdf7f2 | 704 | ceph_assert(mds->inotable->get_version() == piv); |
7c673cae FG |
705 | } |
706 | ||
707 | mds->sessionmap.mark_dirty(session); | |
708 | ||
709 | // apply | |
710 | if (session->get_state_seq() != state_seq) { | |
711 | dout(10) << " journaled state_seq " << state_seq << " != current " << session->get_state_seq() | |
712 | << ", noop" << dendl; | |
713 | // close must have been canceled (by an import?), or any number of other things.. | |
714 | } else if (open) { | |
11fdf7f2 | 715 | ceph_assert(session->is_opening()); |
7c673cae FG |
716 | mds->sessionmap.set_state(session, Session::STATE_OPEN); |
717 | mds->sessionmap.touch_session(session); | |
11fdf7f2 TL |
718 | ceph_assert(session->get_connection()); |
719 | auto reply = MClientSession::create(CEPH_SESSION_OPEN); | |
720 | if (session->info.has_feature(CEPHFS_FEATURE_MIMIC)) | |
721 | reply->supported_features = supported_features; | |
722 | mds->send_message_client(reply, session); | |
723 | if (mdcache->is_readonly()) { | |
724 | auto m = MClientSession::create(CEPH_SESSION_FORCE_RO); | |
725 | mds->send_message_client(m, session); | |
726 | } | |
7c673cae FG |
727 | } else if (session->is_closing() || |
728 | session->is_killing()) { | |
729 | // kill any lingering capabilities, leases, requests | |
730 | while (!session->caps.empty()) { | |
731 | Capability *cap = session->caps.front(); | |
732 | CInode *in = cap->get_inode(); | |
733 | dout(20) << " killing capability " << ccap_string(cap->issued()) << " on " << *in << dendl; | |
a8e16298 | 734 | mds->locker->remove_client_cap(in, cap); |
7c673cae FG |
735 | } |
736 | while (!session->leases.empty()) { | |
737 | ClientLease *r = session->leases.front(); | |
738 | CDentry *dn = static_cast<CDentry*>(r->parent); | |
739 | dout(20) << " killing client lease of " << *dn << dendl; | |
740 | dn->remove_client_lease(r, mds->locker); | |
741 | } | |
11fdf7f2 | 742 | if (client_reconnect_gather.erase(session->info.get_client())) { |
7c673cae | 743 | dout(20) << " removing client from reconnect set" << dendl; |
7c673cae FG |
744 | if (client_reconnect_gather.empty()) { |
745 | dout(7) << " client " << session->info.inst << " was last reconnect, finishing" << dendl; | |
746 | reconnect_gather_finish(); | |
747 | } | |
748 | } | |
11fdf7f2 TL |
749 | if (client_reclaim_gather.erase(session->info.get_client())) { |
750 | dout(20) << " removing client from reclaim set" << dendl; | |
751 | if (client_reclaim_gather.empty()) { | |
752 | dout(7) << " client " << session->info.inst << " was last reclaimed, finishing" << dendl; | |
753 | mds->maybe_clientreplay_done(); | |
754 | } | |
755 | } | |
7c673cae FG |
756 | |
757 | if (session->is_closing()) { | |
758 | // mark con disposable. if there is a fault, we will get a | |
759 | // reset and clean it up. if the client hasn't received the | |
760 | // CLOSE message yet, they will reconnect and get an | |
761 | // ms_handle_remote_reset() and realize they had in fact closed. | |
762 | // do this *before* sending the message to avoid a possible | |
763 | // race. | |
11fdf7f2 | 764 | if (session->get_connection()) { |
7c673cae FG |
765 | // Conditional because terminate_sessions will indiscrimately |
766 | // put sessions in CLOSING whether they ever had a conn or not. | |
11fdf7f2 | 767 | session->get_connection()->mark_disposable(); |
7c673cae FG |
768 | } |
769 | ||
770 | // reset session | |
11fdf7f2 | 771 | mds->send_message_client(MClientSession::create(CEPH_SESSION_CLOSE), session); |
7c673cae FG |
772 | mds->sessionmap.set_state(session, Session::STATE_CLOSED); |
773 | session->clear(); | |
774 | mds->sessionmap.remove_session(session); | |
775 | } else if (session->is_killing()) { | |
776 | // destroy session, close connection | |
11fdf7f2 TL |
777 | if (session->get_connection()) { |
778 | session->get_connection()->mark_down(); | |
779 | session->get_connection()->set_priv(NULL); | |
7c673cae FG |
780 | } |
781 | mds->sessionmap.remove_session(session); | |
782 | } else { | |
783 | ceph_abort(); | |
784 | } | |
785 | } else { | |
786 | ceph_abort(); | |
787 | } | |
788 | } | |
789 | ||
790 | /** | |
791 | * Inject sessions from some source other than actual connections. | |
792 | * | |
793 | * For example: | |
794 | * - sessions inferred from journal replay | |
795 | * - sessions learned from other MDSs during rejoin | |
796 | * - sessions learned from other MDSs during dir/caps migration | |
797 | * - sessions learned from other MDSs during a cross-MDS rename | |
798 | */ | |
799 | version_t Server::prepare_force_open_sessions(map<client_t,entity_inst_t>& cm, | |
11fdf7f2 | 800 | map<client_t,client_metadata_t>& cmm, |
28e407b8 | 801 | map<client_t, pair<Session*,uint64_t> >& smap) |
7c673cae FG |
802 | { |
803 | version_t pv = mds->sessionmap.get_projected(); | |
804 | ||
805 | dout(10) << "prepare_force_open_sessions " << pv | |
806 | << " on " << cm.size() << " clients" | |
807 | << dendl; | |
7c673cae | 808 | |
28e407b8 | 809 | mds->objecter->with_osdmap( |
11fdf7f2 | 810 | [this, &cm, &cmm](const OSDMap &osd_map) { |
28e407b8 AA |
811 | for (auto p = cm.begin(); p != cm.end(); ) { |
812 | if (osd_map.is_blacklisted(p->second.addr)) { | |
813 | dout(10) << " ignoring blacklisted client." << p->first | |
814 | << " (" << p->second.addr << ")" << dendl; | |
11fdf7f2 | 815 | cmm.erase(p->first); |
28e407b8 AA |
816 | cm.erase(p++); |
817 | } else { | |
818 | ++p; | |
819 | } | |
820 | } | |
821 | }); | |
822 | ||
823 | for (map<client_t,entity_inst_t>::iterator p = cm.begin(); p != cm.end(); ++p) { | |
7c673cae FG |
824 | Session *session = mds->sessionmap.get_or_add_session(p->second); |
825 | pv = mds->sessionmap.mark_projected(session); | |
28e407b8 | 826 | uint64_t sseq; |
7c673cae FG |
827 | if (session->is_closed() || |
828 | session->is_closing() || | |
28e407b8 AA |
829 | session->is_killing()) { |
830 | sseq = mds->sessionmap.set_state(session, Session::STATE_OPENING); | |
11fdf7f2 TL |
831 | auto q = cmm.find(p->first); |
832 | if (q != cmm.end()) | |
833 | session->info.client_metadata.merge(q->second); | |
28e407b8 | 834 | } else { |
11fdf7f2 | 835 | ceph_assert(session->is_open() || |
7c673cae FG |
836 | session->is_opening() || |
837 | session->is_stale()); | |
28e407b8 AA |
838 | sseq = 0; |
839 | } | |
840 | smap[p->first] = make_pair(session, sseq); | |
7c673cae FG |
841 | session->inc_importing(); |
842 | } | |
843 | return pv; | |
844 | } | |
845 | ||
28e407b8 | 846 | void Server::finish_force_open_sessions(const map<client_t,pair<Session*,uint64_t> >& smap, |
7c673cae FG |
847 | bool dec_import) |
848 | { | |
849 | /* | |
850 | * FIXME: need to carefully consider the race conditions between a | |
851 | * client trying to close a session and an MDS doing an import | |
852 | * trying to force open a session... | |
853 | */ | |
28e407b8 | 854 | dout(10) << "finish_force_open_sessions on " << smap.size() << " clients," |
7c673cae | 855 | << " initial v " << mds->sessionmap.get_version() << dendl; |
7c673cae | 856 | |
28e407b8 AA |
857 | for (auto &it : smap) { |
858 | Session *session = it.second.first; | |
859 | uint64_t sseq = it.second.second; | |
860 | if (sseq > 0) { | |
7c673cae FG |
861 | if (session->get_state_seq() != sseq) { |
862 | dout(10) << "force_open_sessions skipping changed " << session->info.inst << dendl; | |
863 | } else { | |
864 | dout(10) << "force_open_sessions opened " << session->info.inst << dendl; | |
865 | mds->sessionmap.set_state(session, Session::STATE_OPEN); | |
866 | mds->sessionmap.touch_session(session); | |
11fdf7f2 TL |
867 | |
868 | auto reply = MClientSession::create(CEPH_SESSION_OPEN); | |
869 | if (session->info.has_feature(CEPHFS_FEATURE_MIMIC)) | |
870 | reply->supported_features = supported_features; | |
871 | mds->send_message_client(reply, session); | |
872 | ||
7c673cae | 873 | if (mdcache->is_readonly()) |
11fdf7f2 | 874 | mds->send_message_client(MClientSession::create(CEPH_SESSION_FORCE_RO), session); |
7c673cae FG |
875 | } |
876 | } else { | |
877 | dout(10) << "force_open_sessions skipping already-open " << session->info.inst << dendl; | |
11fdf7f2 | 878 | ceph_assert(session->is_open() || session->is_stale()); |
7c673cae FG |
879 | } |
880 | ||
881 | if (dec_import) { | |
882 | session->dec_importing(); | |
883 | } | |
884 | ||
885 | mds->sessionmap.mark_dirty(session); | |
886 | } | |
887 | ||
888 | dout(10) << __func__ << ": final v " << mds->sessionmap.get_version() << dendl; | |
889 | } | |
890 | ||
891 | class C_MDS_TerminatedSessions : public ServerContext { | |
892 | void finish(int r) override { | |
893 | server->terminating_sessions = false; | |
894 | } | |
895 | public: | |
896 | explicit C_MDS_TerminatedSessions(Server *s) : ServerContext(s) {} | |
897 | }; | |
898 | ||
899 | void Server::terminate_sessions() | |
900 | { | |
a8e16298 | 901 | dout(5) << "terminating all sessions..." << dendl; |
7c673cae FG |
902 | |
903 | terminating_sessions = true; | |
904 | ||
905 | // kill them off. clients will retry etc. | |
906 | set<Session*> sessions; | |
907 | mds->sessionmap.get_client_session_set(sessions); | |
908 | for (set<Session*>::const_iterator p = sessions.begin(); | |
909 | p != sessions.end(); | |
910 | ++p) { | |
911 | Session *session = *p; | |
912 | if (session->is_closing() || | |
913 | session->is_killing() || | |
914 | session->is_closed()) | |
915 | continue; | |
916 | journal_close_session(session, Session::STATE_CLOSING, NULL); | |
917 | } | |
918 | ||
919 | mdlog->wait_for_safe(new C_MDS_TerminatedSessions(this)); | |
920 | } | |
921 | ||
922 | ||
923 | void Server::find_idle_sessions() | |
924 | { | |
91327a77 AA |
925 | auto now = clock::now(); |
926 | auto last_cleared_laggy = mds->last_cleared_laggy(); | |
927 | ||
928 | dout(10) << "find_idle_sessions. last cleared laggy state " << last_cleared_laggy << "s ago" << dendl; | |
7c673cae FG |
929 | |
930 | // timeout/stale | |
931 | // (caps go stale, lease die) | |
91327a77 AA |
932 | double queue_max_age = mds->get_dispatch_queue_max_age(ceph_clock_now()); |
933 | double cutoff = queue_max_age + mds->mdsmap->get_session_timeout(); | |
f64942e4 | 934 | |
11fdf7f2 TL |
935 | std::vector<Session*> to_evict; |
936 | ||
f64942e4 AA |
937 | const auto sessions_p1 = mds->sessionmap.by_state.find(Session::STATE_OPEN); |
938 | if (sessions_p1 != mds->sessionmap.by_state.end() && !sessions_p1->second->empty()) { | |
939 | std::vector<Session*> new_stale; | |
940 | ||
941 | for (auto session : *(sessions_p1->second)) { | |
942 | auto last_cap_renew_span = std::chrono::duration<double>(now - session->last_cap_renew).count(); | |
943 | if (last_cap_renew_span < cutoff) { | |
944 | dout(20) << "laggiest active session is " << session->info.inst | |
945 | << " and renewed caps recently (" << last_cap_renew_span << "s ago)" << dendl; | |
946 | break; | |
947 | } | |
948 | ||
949 | if (session->last_seen > session->last_cap_renew) { | |
950 | last_cap_renew_span = std::chrono::duration<double>(now - session->last_seen).count(); | |
951 | if (last_cap_renew_span < cutoff) { | |
952 | dout(20) << "laggiest active session is " << session->info.inst | |
953 | << " and renewed caps recently (" << last_cap_renew_span << "s ago)" << dendl; | |
954 | continue; | |
955 | } | |
956 | } | |
957 | ||
11fdf7f2 TL |
958 | auto it = session->info.client_metadata.find("timeout"); |
959 | if (it != session->info.client_metadata.end()) { | |
960 | unsigned timeout = strtoul(it->second.c_str(), nullptr, 0); | |
961 | if (timeout == 0) { | |
962 | dout(10) << "skipping session " << session->info.inst | |
963 | << ", infinite timeout specified" << dendl; | |
964 | continue; | |
965 | } | |
966 | double cutoff = queue_max_age + timeout; | |
967 | if (last_cap_renew_span < cutoff) { | |
968 | dout(10) << "skipping session " << session->info.inst | |
969 | << ", timeout (" << timeout << ") specified" | |
970 | << " and renewed caps recently (" << last_cap_renew_span << "s ago)" << dendl; | |
971 | continue; | |
972 | } | |
973 | ||
974 | // do not go through stale, evict it directly. | |
975 | to_evict.push_back(session); | |
976 | } else { | |
977 | dout(10) << "new stale session " << session->info.inst | |
978 | << " last renewed caps " << last_cap_renew_span << "s ago" << dendl; | |
979 | new_stale.push_back(session); | |
980 | } | |
7c673cae FG |
981 | } |
982 | ||
f64942e4 AA |
983 | for (auto session : new_stale) { |
984 | mds->sessionmap.set_state(session, Session::STATE_STALE); | |
985 | mds->locker->revoke_stale_caps(session); | |
986 | mds->locker->remove_stale_leases(session); | |
11fdf7f2 | 987 | mds->send_message_client(MClientSession::create(CEPH_SESSION_STALE, session->get_push_seq()), session); |
f64942e4 AA |
988 | finish_flush_session(session, session->get_push_seq()); |
989 | } | |
7c673cae FG |
990 | } |
991 | ||
992 | // autoclose | |
91327a77 | 993 | cutoff = queue_max_age + mds->mdsmap->get_session_autoclose(); |
7c673cae FG |
994 | |
995 | // don't kick clients if we've been laggy | |
91327a77 AA |
996 | if (last_cleared_laggy < cutoff) { |
997 | dout(10) << " last cleared laggy " << last_cleared_laggy << "s ago (< cutoff " << cutoff | |
998 | << "), not kicking any clients to be safe" << dendl; | |
7c673cae FG |
999 | return; |
1000 | } | |
1001 | ||
91327a77 AA |
1002 | if (mds->sessionmap.get_sessions().size() == 1 && mds->mdsmap->get_num_in_mds() == 1) { |
1003 | dout(20) << "skipping client eviction because there is only one" << dendl; | |
7c673cae FG |
1004 | return; |
1005 | } | |
1006 | ||
31f18b77 | 1007 | // Collect a list of sessions exceeding the autoclose threshold |
f64942e4 | 1008 | const auto sessions_p2 = mds->sessionmap.by_state.find(Session::STATE_STALE); |
11fdf7f2 TL |
1009 | if (sessions_p2 != mds->sessionmap.by_state.end() && !sessions_p2->second->empty()) { |
1010 | for (auto session : *(sessions_p2->second)) { | |
1011 | assert(session->is_stale()); | |
1012 | auto last_cap_renew_span = std::chrono::duration<double>(now - session->last_cap_renew).count(); | |
1013 | if (last_cap_renew_span < cutoff) { | |
1014 | dout(20) << "oldest stale session is " << session->info.inst | |
1015 | << " and recently renewed caps " << last_cap_renew_span << "s ago" << dendl; | |
1016 | break; | |
1017 | } | |
1018 | to_evict.push_back(session); | |
1019 | } | |
31f18b77 | 1020 | } |
31f18b77 | 1021 | |
11fdf7f2 | 1022 | for (auto session: to_evict) { |
7c673cae | 1023 | if (session->is_importing()) { |
11fdf7f2 TL |
1024 | dout(10) << "skipping session " << session->info.inst << ", it's being imported" << dendl; |
1025 | continue; | |
7c673cae | 1026 | } |
31f18b77 | 1027 | |
11fdf7f2 TL |
1028 | auto last_cap_renew_span = std::chrono::duration<double>(now - session->last_cap_renew).count(); |
1029 | mds->clog->warn() << "evicting unresponsive client " << *session | |
1030 | << ", after " << last_cap_renew_span << " seconds"; | |
1031 | dout(10) << "autoclosing stale session " << session->info.inst | |
1032 | << " last renewed caps " << last_cap_renew_span << "s ago" << dendl; | |
31f18b77 | 1033 | |
11fdf7f2 | 1034 | if (g_conf()->mds_session_blacklist_on_timeout) { |
31f18b77 | 1035 | std::stringstream ss; |
11fdf7f2 | 1036 | mds->evict_client(session->get_client().v, false, true, ss, nullptr); |
31f18b77 FG |
1037 | } else { |
1038 | kill_session(session, NULL); | |
1039 | } | |
7c673cae FG |
1040 | } |
1041 | } | |
1042 | ||
91327a77 AA |
1043 | void Server::evict_cap_revoke_non_responders() { |
1044 | if (!cap_revoke_eviction_timeout) { | |
1045 | return; | |
1046 | } | |
1047 | ||
1048 | std::list<client_t> to_evict; | |
1049 | mds->locker->get_late_revoking_clients(&to_evict, cap_revoke_eviction_timeout); | |
1050 | ||
1051 | for (auto const &client: to_evict) { | |
1052 | mds->clog->warn() << "client id " << client << " has not responded to" | |
1053 | << " cap revoke by MDS for over " << cap_revoke_eviction_timeout | |
1054 | << " seconds, evicting"; | |
1055 | dout(1) << __func__ << ": evicting cap revoke non-responder client id " | |
1056 | << client << dendl; | |
1057 | ||
1058 | std::stringstream ss; | |
1059 | bool evicted = mds->evict_client(client.v, false, | |
11fdf7f2 | 1060 | g_conf()->mds_session_blacklist_on_evict, |
91327a77 AA |
1061 | ss, nullptr); |
1062 | if (evicted && logger) { | |
1063 | logger->inc(l_mdss_cap_revoke_eviction); | |
1064 | } | |
1065 | } | |
1066 | } | |
1067 | ||
11fdf7f2 | 1068 | void Server::handle_conf_change(const ConfigProxy& conf, |
91327a77 AA |
1069 | const std::set <std::string> &changed) { |
1070 | if (changed.count("mds_cap_revoke_eviction_timeout")) { | |
11fdf7f2 | 1071 | cap_revoke_eviction_timeout = g_conf().get_val<double>("mds_cap_revoke_eviction_timeout"); |
91327a77 AA |
1072 | dout(20) << __func__ << " cap revoke eviction timeout changed to " |
1073 | << cap_revoke_eviction_timeout << dendl; | |
1074 | } | |
a8e16298 | 1075 | if (changed.count("mds_recall_max_decay_rate")) { |
11fdf7f2 | 1076 | recall_throttle = DecayCounter(g_conf().get_val<double>("mds_recall_max_decay_rate")); |
a8e16298 | 1077 | } |
91327a77 AA |
1078 | } |
1079 | ||
7c673cae | 1080 | /* |
11fdf7f2 | 1081 | * XXX bump in the interface here, not using an MDSContext here |
7c673cae FG |
1082 | * because all the callers right now happen to use a SaferCond |
1083 | */ | |
1084 | void Server::kill_session(Session *session, Context *on_safe) | |
1085 | { | |
11fdf7f2 | 1086 | ceph_assert(mds->mds_lock.is_locked_by_me()); |
31f18b77 | 1087 | |
7c673cae FG |
1088 | if ((session->is_opening() || |
1089 | session->is_open() || | |
1090 | session->is_stale()) && | |
1091 | !session->is_importing()) { | |
1092 | dout(10) << "kill_session " << session << dendl; | |
1093 | journal_close_session(session, Session::STATE_KILLING, on_safe); | |
1094 | } else { | |
1095 | dout(10) << "kill_session importing or already closing/killing " << session << dendl; | |
11fdf7f2 TL |
1096 | if (session->is_closing() || |
1097 | session->is_killing()) { | |
1098 | if (on_safe) | |
1099 | mdlog->wait_for_safe(new MDSInternalContextWrapper(mds, on_safe)); | |
1100 | } else { | |
1101 | ceph_assert(session->is_closed() || | |
1102 | session->is_importing()); | |
1103 | if (on_safe) | |
1104 | on_safe->complete(0); | |
7c673cae FG |
1105 | } |
1106 | } | |
1107 | } | |
1108 | ||
31f18b77 FG |
1109 | size_t Server::apply_blacklist(const std::set<entity_addr_t> &blacklist) |
1110 | { | |
1111 | std::list<Session*> victims; | |
11fdf7f2 TL |
1112 | const auto& sessions = mds->sessionmap.get_sessions(); |
1113 | for (const auto& p : sessions) { | |
31f18b77 FG |
1114 | if (!p.first.is_client()) { |
1115 | // Do not apply OSDMap blacklist to MDS daemons, we find out | |
1116 | // about their death via MDSMap. | |
1117 | continue; | |
1118 | } | |
1119 | ||
1120 | Session *s = p.second; | |
1121 | if (blacklist.count(s->info.inst.addr)) { | |
1122 | victims.push_back(s); | |
1123 | } | |
1124 | } | |
1125 | ||
1126 | for (const auto s : victims) { | |
1127 | kill_session(s, nullptr); | |
1128 | } | |
1129 | ||
1130 | dout(10) << "apply_blacklist: killed " << victims.size() << dendl; | |
1131 | ||
1132 | return victims.size(); | |
1133 | } | |
1134 | ||
7c673cae FG |
1135 | void Server::journal_close_session(Session *session, int state, Context *on_safe) |
1136 | { | |
1137 | uint64_t sseq = mds->sessionmap.set_state(session, state); | |
1138 | version_t pv = mds->sessionmap.mark_projected(session); | |
1139 | version_t piv = 0; | |
1140 | ||
1141 | // release alloc and pending-alloc inos for this session | |
1142 | // and wipe out session state, in case the session close aborts for some reason | |
1143 | interval_set<inodeno_t> both; | |
1144 | both.insert(session->info.prealloc_inos); | |
1145 | both.insert(session->pending_prealloc_inos); | |
1146 | if (both.size()) { | |
1147 | mds->inotable->project_release_ids(both); | |
1148 | piv = mds->inotable->get_projected_version(); | |
1149 | } else | |
1150 | piv = 0; | |
1151 | ||
1152 | mdlog->start_submit_entry(new ESession(session->info.inst, false, pv, both, piv), | |
1153 | new C_MDS_session_finish(this, session, sseq, false, pv, both, piv, on_safe)); | |
1154 | mdlog->flush(); | |
1155 | ||
1156 | // clean up requests, too | |
1157 | elist<MDRequestImpl*>::iterator p = | |
1158 | session->requests.begin(member_offset(MDRequestImpl, | |
1159 | item_session_request)); | |
1160 | while (!p.end()) { | |
1161 | MDRequestRef mdr = mdcache->request_get((*p)->reqid); | |
1162 | ++p; | |
1163 | mdcache->request_kill(mdr); | |
1164 | } | |
1165 | ||
1166 | finish_flush_session(session, session->get_push_seq()); | |
1167 | } | |
1168 | ||
11fdf7f2 | 1169 | void Server::reconnect_clients(MDSContext *reconnect_done_) |
7c673cae FG |
1170 | { |
1171 | reconnect_done = reconnect_done_; | |
28e407b8 | 1172 | |
11fdf7f2 | 1173 | auto now = clock::now(); |
28e407b8 AA |
1174 | set<Session*> sessions; |
1175 | mds->sessionmap.get_client_session_set(sessions); | |
1176 | for (auto session : sessions) { | |
11fdf7f2 TL |
1177 | if (session->is_open()) { |
1178 | client_reconnect_gather.insert(session->get_client()); | |
1179 | session->last_cap_renew = now; | |
1180 | } | |
28e407b8 | 1181 | } |
7c673cae FG |
1182 | |
1183 | if (client_reconnect_gather.empty()) { | |
1184 | dout(7) << "reconnect_clients -- no sessions, doing nothing." << dendl; | |
1185 | reconnect_gather_finish(); | |
1186 | return; | |
1187 | } | |
1188 | ||
1189 | // clients will get the mdsmap and discover we're reconnecting via the monitor. | |
1190 | ||
11fdf7f2 | 1191 | reconnect_start = now; |
7c673cae FG |
1192 | dout(1) << "reconnect_clients -- " << client_reconnect_gather.size() << " sessions" << dendl; |
1193 | mds->sessionmap.dump(); | |
1194 | } | |
1195 | ||
11fdf7f2 | 1196 | void Server::handle_client_reconnect(const MClientReconnect::const_ref &m) |
7c673cae | 1197 | { |
11fdf7f2 TL |
1198 | dout(7) << "handle_client_reconnect " << m->get_source() |
1199 | << (m->has_more() ? " (more)" : "") << dendl; | |
7c673cae | 1200 | client_t from = m->get_source().num(); |
94b18763 | 1201 | Session *session = mds->get_session(m); |
11fdf7f2 | 1202 | ceph_assert(session); |
7c673cae FG |
1203 | |
1204 | if (!mds->is_reconnect() && mds->get_want_state() == CEPH_MDS_STATE_RECONNECT) { | |
1205 | dout(10) << " we're almost in reconnect state (mdsmap delivery race?); waiting" << dendl; | |
1206 | mds->wait_for_reconnect(new C_MDS_RetryMessage(mds, m)); | |
1207 | return; | |
1208 | } | |
1209 | ||
f64942e4 | 1210 | auto delay = std::chrono::duration<double>(clock::now() - reconnect_start).count(); |
7c673cae FG |
1211 | dout(10) << " reconnect_start " << reconnect_start << " delay " << delay << dendl; |
1212 | ||
1213 | bool deny = false; | |
b32b8144 | 1214 | if (!mds->is_reconnect() || mds->get_want_state() != CEPH_MDS_STATE_RECONNECT || reconnect_evicting) { |
7c673cae FG |
1215 | // XXX maybe in the future we can do better than this? |
1216 | dout(1) << " no longer in reconnect state, ignoring reconnect, sending close" << dendl; | |
1217 | mds->clog->info() << "denied reconnect attempt (mds is " | |
1218 | << ceph_mds_state_name(mds->get_state()) | |
1219 | << ") from " << m->get_source_inst() | |
11fdf7f2 | 1220 | << " after " << delay << " (allowed interval " << g_conf()->mds_reconnect_timeout << ")"; |
7c673cae | 1221 | deny = true; |
11fdf7f2 TL |
1222 | } else { |
1223 | std::string error_str; | |
1224 | if (!session->is_open()) { | |
1225 | error_str = "session is closed"; | |
1226 | } else if (mdcache->is_readonly()) { | |
1227 | error_str = "mds is readonly"; | |
1228 | } else { | |
1229 | if (session->info.client_metadata.features.empty()) | |
1230 | infer_supported_features(session, session->info.client_metadata); | |
1231 | ||
1232 | feature_bitset_t missing_features = required_client_features; | |
1233 | missing_features -= session->info.client_metadata.features; | |
1234 | if (!missing_features.empty()) { | |
1235 | stringstream ss; | |
1236 | ss << "missing required features '" << missing_features << "'"; | |
1237 | error_str = ss.str(); | |
1238 | } | |
1239 | } | |
1240 | ||
1241 | if (!error_str.empty()) { | |
1242 | deny = true; | |
1243 | dout(1) << " " << error_str << ", ignoring reconnect, sending close" << dendl; | |
1244 | mds->clog->info() << "denied reconnect attempt from " | |
1245 | << m->get_source_inst() << " (" << error_str << ")"; | |
1246 | } | |
7c673cae FG |
1247 | } |
1248 | ||
1249 | if (deny) { | |
11fdf7f2 TL |
1250 | auto r = MClientSession::create(CEPH_SESSION_CLOSE); |
1251 | mds->send_message_client(r, session); | |
1252 | if (session->is_open()) | |
1253 | kill_session(session, nullptr); | |
7c673cae FG |
1254 | return; |
1255 | } | |
1256 | ||
11fdf7f2 TL |
1257 | if (!m->has_more()) { |
1258 | // notify client of success with an OPEN | |
1259 | auto reply = MClientSession::create(CEPH_SESSION_OPEN); | |
1260 | if (session->info.has_feature(CEPHFS_FEATURE_MIMIC)) | |
1261 | reply->supported_features = supported_features; | |
1262 | mds->send_message_client(reply, session); | |
1263 | mds->clog->debug() << "reconnect by " << session->info.inst << " after " << delay; | |
1264 | } | |
1265 | ||
91327a77 | 1266 | session->last_cap_renew = clock::now(); |
7c673cae FG |
1267 | |
1268 | // snaprealms | |
11fdf7f2 TL |
1269 | for (const auto &r : m->realms) { |
1270 | CInode *in = mdcache->get_inode(inodeno_t(r.realm.ino)); | |
7c673cae FG |
1271 | if (in && in->state_test(CInode::STATE_PURGING)) |
1272 | continue; | |
1273 | if (in) { | |
11fdf7f2 TL |
1274 | if (in->snaprealm) { |
1275 | dout(15) << "open snaprealm (w inode) on " << *in << dendl; | |
7c673cae | 1276 | } else { |
11fdf7f2 TL |
1277 | // this can happen if we are non-auth or we rollback snaprealm |
1278 | dout(15) << "open snaprealm (null snaprealm) on " << *in << dendl; | |
7c673cae | 1279 | } |
11fdf7f2 | 1280 | mdcache->add_reconnected_snaprealm(from, inodeno_t(r.realm.ino), snapid_t(r.realm.seq)); |
7c673cae | 1281 | } else { |
11fdf7f2 TL |
1282 | dout(15) << "open snaprealm (w/o inode) on " << inodeno_t(r.realm.ino) |
1283 | << " seq " << r.realm.seq << dendl; | |
1284 | mdcache->add_reconnected_snaprealm(from, inodeno_t(r.realm.ino), snapid_t(r.realm.seq)); | |
7c673cae FG |
1285 | } |
1286 | } | |
1287 | ||
1288 | // caps | |
11fdf7f2 | 1289 | for (const auto &p : m->caps) { |
7c673cae | 1290 | // make sure our last_cap_id is MAX over all issued caps |
11fdf7f2 TL |
1291 | if (p.second.capinfo.cap_id > mdcache->last_cap_id) |
1292 | mdcache->last_cap_id = p.second.capinfo.cap_id; | |
7c673cae | 1293 | |
11fdf7f2 | 1294 | CInode *in = mdcache->get_inode(p.first); |
7c673cae FG |
1295 | if (in && in->state_test(CInode::STATE_PURGING)) |
1296 | continue; | |
1297 | if (in && in->is_auth()) { | |
1298 | // we recovered it, and it's ours. take note. | |
11fdf7f2 | 1299 | dout(15) << "open cap realm " << inodeno_t(p.second.capinfo.snaprealm) |
7c673cae | 1300 | << " on " << *in << dendl; |
11fdf7f2 TL |
1301 | in->reconnect_cap(from, p.second, session); |
1302 | mdcache->add_reconnected_cap(from, p.first, p.second); | |
1303 | recover_filelocks(in, p.second.flockbl, m->get_orig_source().num()); | |
7c673cae FG |
1304 | continue; |
1305 | } | |
1306 | ||
1307 | if (in && !in->is_auth()) { | |
1308 | // not mine. | |
1309 | dout(10) << "non-auth " << *in << ", will pass off to authority" << dendl; | |
1310 | // add to cap export list. | |
11fdf7f2 TL |
1311 | mdcache->rejoin_export_caps(p.first, from, p.second, |
1312 | in->authority().first, true); | |
7c673cae FG |
1313 | } else { |
1314 | // don't know if the inode is mine | |
11fdf7f2 TL |
1315 | dout(10) << "missing ino " << p.first << ", will load later" << dendl; |
1316 | mdcache->rejoin_recovered_caps(p.first, from, p.second, MDS_RANK_NONE); | |
7c673cae FG |
1317 | } |
1318 | } | |
1319 | ||
f64942e4 AA |
1320 | reconnect_last_seen = clock::now(); |
1321 | ||
11fdf7f2 TL |
1322 | if (!m->has_more()) { |
1323 | mdcache->rejoin_recovered_client(session->get_client(), session->info.inst); | |
1324 | ||
1325 | // remove from gather set | |
1326 | client_reconnect_gather.erase(from); | |
1327 | if (client_reconnect_gather.empty()) | |
1328 | reconnect_gather_finish(); | |
1329 | } | |
1330 | } | |
1331 | ||
1332 | void Server::infer_supported_features(Session *session, client_metadata_t& client_metadata) | |
1333 | { | |
1334 | int supported = -1; | |
1335 | auto it = client_metadata.find("ceph_version"); | |
1336 | if (it != client_metadata.end()) { | |
1337 | // user space client | |
1338 | if (it->second.compare(0, 16, "ceph version 12.") == 0) | |
1339 | supported = CEPHFS_FEATURE_LUMINOUS; | |
1340 | else if (session->get_connection()->has_feature(CEPH_FEATURE_FS_CHANGE_ATTR)) | |
1341 | supported = CEPHFS_FEATURE_KRAKEN; | |
1342 | } else { | |
1343 | it = client_metadata.find("kernel_version"); | |
1344 | if (it != client_metadata.end()) { | |
1345 | // kernel client | |
1346 | if (session->get_connection()->has_feature(CEPH_FEATURE_NEW_OSDOP_ENCODING)) | |
1347 | supported = CEPHFS_FEATURE_LUMINOUS; | |
1348 | } | |
1349 | } | |
1350 | if (supported == -1 && | |
1351 | session->get_connection()->has_feature(CEPH_FEATURE_FS_FILE_LAYOUT_V2)) | |
1352 | supported = CEPHFS_FEATURE_JEWEL; | |
7c673cae | 1353 | |
11fdf7f2 TL |
1354 | if (supported >= 0) { |
1355 | unsigned long value = (1UL << (supported + 1)) - 1; | |
1356 | client_metadata.features = feature_bitset_t(value); | |
1357 | dout(10) << __func__ << " got '" << client_metadata.features << "'" << dendl; | |
1358 | } | |
7c673cae FG |
1359 | } |
1360 | ||
11fdf7f2 TL |
1361 | void Server::update_required_client_features() |
1362 | { | |
1363 | vector<size_t> bits = CEPHFS_FEATURES_MDS_REQUIRED; | |
1364 | ||
1365 | int min_compat = mds->mdsmap->get_min_compat_client(); | |
1366 | if (min_compat >= CEPH_RELEASE_NAUTILUS) { | |
1367 | bits.push_back(CEPHFS_FEATURE_NAUTILUS); | |
1368 | } else if (min_compat >= CEPH_RELEASE_MIMIC) | |
1369 | bits.push_back(CEPHFS_FEATURE_MIMIC); | |
1370 | else if (min_compat >= CEPH_RELEASE_LUMINOUS) | |
1371 | bits.push_back(CEPHFS_FEATURE_LUMINOUS); | |
1372 | else if (min_compat >= CEPH_RELEASE_KRAKEN) | |
1373 | bits.push_back(CEPHFS_FEATURE_KRAKEN); | |
1374 | else if (min_compat >= CEPH_RELEASE_JEWEL) | |
1375 | bits.push_back(CEPHFS_FEATURE_JEWEL); | |
1376 | ||
1377 | std::sort(bits.begin(), bits.end()); | |
1378 | required_client_features = feature_bitset_t(bits); | |
1379 | dout(7) << "required_client_features: " << required_client_features << dendl; | |
1380 | ||
1381 | if (mds->get_state() >= MDSMap::STATE_RECONNECT) { | |
1382 | set<Session*> sessions; | |
1383 | mds->sessionmap.get_client_session_set(sessions); | |
1384 | for (auto session : sessions) { | |
1385 | feature_bitset_t missing_features = required_client_features; | |
1386 | missing_features -= session->info.client_metadata.features; | |
1387 | if (!missing_features.empty()) { | |
1388 | bool blacklisted = mds->objecter->with_osdmap( | |
1389 | [session](const OSDMap &osd_map) -> bool { | |
1390 | return osd_map.is_blacklisted(session->info.inst.addr); | |
1391 | }); | |
1392 | if (blacklisted) | |
1393 | continue; | |
7c673cae | 1394 | |
11fdf7f2 TL |
1395 | mds->clog->warn() << "evicting session " << *session << ", missing required features '" |
1396 | << missing_features << "'"; | |
1397 | std::stringstream ss; | |
1398 | mds->evict_client(session->get_client().v, false, | |
1399 | g_conf()->mds_session_blacklist_on_evict, ss); | |
1400 | } | |
1401 | } | |
1402 | } | |
1403 | } | |
7c673cae FG |
1404 | |
1405 | void Server::reconnect_gather_finish() | |
1406 | { | |
1407 | dout(7) << "reconnect_gather_finish. failed on " << failed_reconnects << " clients" << dendl; | |
11fdf7f2 TL |
1408 | ceph_assert(reconnect_done); |
1409 | ||
1410 | if (!mds->snapclient->is_synced()) { | |
1411 | // make sure snaptable cache is populated. snaprealms will be | |
1412 | // extensively used in rejoin stage. | |
1413 | dout(7) << " snaptable cache isn't synced, delaying state transition" << dendl; | |
1414 | mds->snapclient->wait_for_sync(reconnect_done); | |
1415 | } else { | |
1416 | reconnect_done->complete(0); | |
1417 | } | |
7c673cae FG |
1418 | reconnect_done = NULL; |
1419 | } | |
1420 | ||
1421 | void Server::reconnect_tick() | |
1422 | { | |
31f18b77 | 1423 | if (reconnect_evicting) { |
f64942e4 | 1424 | dout(7) << "reconnect_tick: waiting for evictions" << dendl; |
31f18b77 FG |
1425 | return; |
1426 | } | |
1427 | ||
f64942e4 AA |
1428 | if (client_reconnect_gather.empty()) |
1429 | return; | |
31f18b77 | 1430 | |
f64942e4 AA |
1431 | auto now = clock::now(); |
1432 | auto elapse1 = std::chrono::duration<double>(now - reconnect_start).count(); | |
11fdf7f2 | 1433 | if (elapse1 < g_conf()->mds_reconnect_timeout) |
f64942e4 | 1434 | return; |
31f18b77 | 1435 | |
f64942e4 AA |
1436 | vector<Session*> remaining_sessions; |
1437 | remaining_sessions.reserve(client_reconnect_gather.size()); | |
1438 | for (auto c : client_reconnect_gather) { | |
1439 | Session *session = mds->sessionmap.get_session(entity_name_t::CLIENT(c.v)); | |
1440 | ceph_assert(session); | |
1441 | remaining_sessions.push_back(session); | |
1442 | // client re-sends cap flush messages before the reconnect message | |
1443 | if (session->last_seen > reconnect_last_seen) | |
1444 | reconnect_last_seen = session->last_seen; | |
1445 | } | |
31f18b77 | 1446 | |
f64942e4 | 1447 | auto elapse2 = std::chrono::duration<double>(now - reconnect_last_seen).count(); |
11fdf7f2 | 1448 | if (elapse2 < g_conf()->mds_reconnect_timeout / 2) { |
f64942e4 AA |
1449 | dout(7) << "reconnect_tick: last seen " << elapse2 |
1450 | << " seconds ago, extending reconnect interval" << dendl; | |
1451 | return; | |
1452 | } | |
1453 | ||
1454 | dout(7) << "reconnect timed out, " << remaining_sessions.size() | |
1455 | << " clients have not reconnected in time" << dendl; | |
1456 | ||
1457 | // If we're doing blacklist evictions, use this to wait for them before | |
1458 | // proceeding to reconnect_gather_finish | |
1459 | MDSGatherBuilder gather(g_ceph_context); | |
1460 | ||
1461 | for (auto session : remaining_sessions) { | |
11fdf7f2 TL |
1462 | // Keep sessions that have specified timeout. These sessions will prevent |
1463 | // mds from going to active. MDS goes to active after they all have been | |
1464 | // killed or reclaimed. | |
1465 | if (session->info.client_metadata.find("timeout") != | |
1466 | session->info.client_metadata.end()) { | |
1467 | dout(1) << "reconnect keeps " << session->info.inst | |
1468 | << ", need to be reclaimed" << dendl; | |
1469 | client_reclaim_gather.insert(session->get_client()); | |
1470 | continue; | |
1471 | } | |
1472 | ||
f64942e4 | 1473 | dout(1) << "reconnect gives up on " << session->info.inst << dendl; |
31f18b77 | 1474 | |
f64942e4 AA |
1475 | mds->clog->warn() << "evicting unresponsive client " << *session |
1476 | << ", after waiting " << elapse1 | |
1477 | << " seconds during MDS startup"; | |
1478 | ||
11fdf7f2 | 1479 | if (g_conf()->mds_session_blacklist_on_timeout) { |
f64942e4 AA |
1480 | std::stringstream ss; |
1481 | mds->evict_client(session->get_client().v, false, true, ss, | |
1482 | gather.new_sub()); | |
31f18b77 | 1483 | } else { |
f64942e4 | 1484 | kill_session(session, NULL); |
31f18b77 | 1485 | } |
f64942e4 AA |
1486 | |
1487 | failed_reconnects++; | |
1488 | } | |
1489 | client_reconnect_gather.clear(); | |
1490 | ||
1491 | if (gather.has_subs()) { | |
1492 | dout(1) << "reconnect will complete once clients are evicted" << dendl; | |
1493 | gather.set_finisher(new MDSInternalContextWrapper(mds, new FunctionContext( | |
1494 | [this](int r){reconnect_gather_finish();}))); | |
1495 | gather.activate(); | |
1496 | reconnect_evicting = true; | |
1497 | } else { | |
1498 | reconnect_gather_finish(); | |
7c673cae FG |
1499 | } |
1500 | } | |
1501 | ||
1502 | void Server::recover_filelocks(CInode *in, bufferlist locks, int64_t client) | |
1503 | { | |
1504 | if (!locks.length()) return; | |
1505 | int numlocks; | |
1506 | ceph_filelock lock; | |
11fdf7f2 TL |
1507 | auto p = locks.cbegin(); |
1508 | decode(numlocks, p); | |
7c673cae | 1509 | for (int i = 0; i < numlocks; ++i) { |
11fdf7f2 | 1510 | decode(lock, p); |
7c673cae FG |
1511 | lock.client = client; |
1512 | in->get_fcntl_lock_state()->held_locks.insert(pair<uint64_t, ceph_filelock>(lock.start, lock)); | |
1513 | ++in->get_fcntl_lock_state()->client_held_lock_counts[client]; | |
1514 | } | |
11fdf7f2 | 1515 | decode(numlocks, p); |
7c673cae | 1516 | for (int i = 0; i < numlocks; ++i) { |
11fdf7f2 | 1517 | decode(lock, p); |
7c673cae FG |
1518 | lock.client = client; |
1519 | in->get_flock_lock_state()->held_locks.insert(pair<uint64_t, ceph_filelock> (lock.start, lock)); | |
1520 | ++in->get_flock_lock_state()->client_held_lock_counts[client]; | |
1521 | } | |
1522 | } | |
1523 | ||
7c673cae FG |
1524 | /** |
1525 | * Call this when the MDCache is oversized, to send requests to the clients | |
1526 | * to trim some caps, and consequently unpin some inodes in the MDCache so | |
1527 | * that it can trim too. | |
1528 | */ | |
a8e16298 TL |
1529 | std::pair<bool, uint64_t> Server::recall_client_state(MDSGatherBuilder* gather, RecallFlags flags) |
1530 | { | |
1531 | const auto now = clock::now(); | |
1532 | const bool steady = flags&RecallFlags::STEADY; | |
1533 | const bool enforce_max = flags&RecallFlags::ENFORCE_MAX; | |
1534 | ||
11fdf7f2 TL |
1535 | const auto max_caps_per_client = g_conf().get_val<uint64_t>("mds_max_caps_per_client"); |
1536 | const auto min_caps_per_client = g_conf().get_val<uint64_t>("mds_min_caps_per_client"); | |
1537 | const auto recall_global_max_decay_threshold = g_conf().get_val<Option::size_t>("mds_recall_global_max_decay_threshold"); | |
1538 | const auto recall_max_caps = g_conf().get_val<Option::size_t>("mds_recall_max_caps"); | |
1539 | const auto recall_max_decay_threshold = g_conf().get_val<Option::size_t>("mds_recall_max_decay_threshold"); | |
a8e16298 TL |
1540 | |
1541 | dout(7) << __func__ << ":" | |
1542 | << " min=" << min_caps_per_client | |
1543 | << " max=" << max_caps_per_client | |
1544 | << " total=" << Capability::count() | |
1545 | << " flags=0x" << std::hex << flags | |
1546 | << dendl; | |
f64942e4 | 1547 | |
a8e16298 TL |
1548 | /* trim caps of sessions with the most caps first */ |
1549 | std::multimap<uint64_t, Session*> caps_session; | |
11fdf7f2 | 1550 | auto f = [&caps_session, enforce_max, max_caps_per_client](auto& s) { |
a8e16298 TL |
1551 | auto num_caps = s->caps.size(); |
1552 | if (!enforce_max || num_caps > max_caps_per_client) { | |
1553 | caps_session.emplace(std::piecewise_construct, std::forward_as_tuple(num_caps), std::forward_as_tuple(s)); | |
1554 | } | |
1555 | }; | |
1556 | mds->sessionmap.get_client_sessions(std::move(f)); | |
1557 | ||
1558 | std::pair<bool, uint64_t> result = {false, 0}; | |
11fdf7f2 | 1559 | auto& [throttled, caps_recalled] = result; |
a8e16298 | 1560 | last_recall_state = now; |
11fdf7f2 | 1561 | for (const auto& [num_caps, session] : boost::adaptors::reverse(caps_session)) { |
7c673cae | 1562 | if (!session->is_open() || |
11fdf7f2 | 1563 | !session->get_connection() || |
7c673cae FG |
1564 | !session->info.inst.name.is_client()) |
1565 | continue; | |
1566 | ||
a8e16298 TL |
1567 | dout(10) << __func__ << ":" |
1568 | << " session " << session->info.inst | |
1569 | << " caps " << num_caps | |
7c673cae FG |
1570 | << ", leases " << session->leases.size() |
1571 | << dendl; | |
1572 | ||
a8e16298 TL |
1573 | uint64_t newlim; |
1574 | if (num_caps < recall_max_caps || (num_caps-recall_max_caps) < min_caps_per_client) { | |
1575 | newlim = min_caps_per_client; | |
1576 | } else { | |
1577 | newlim = num_caps-recall_max_caps; | |
1578 | } | |
1579 | if (num_caps > newlim) { | |
1580 | /* now limit the number of caps we recall at a time to prevent overloading ourselves */ | |
1581 | uint64_t recall = std::min<uint64_t>(recall_max_caps, num_caps-newlim); | |
1582 | newlim = num_caps-recall; | |
1583 | const uint64_t session_recall_throttle = session->get_recall_caps_throttle(); | |
11fdf7f2 TL |
1584 | const uint64_t session_recall_throttle2o = session->get_recall_caps_throttle2o(); |
1585 | const uint64_t global_recall_throttle = recall_throttle.get(); | |
a8e16298 TL |
1586 | if (session_recall_throttle+recall > recall_max_decay_threshold) { |
1587 | dout(15) << " session recall threshold (" << recall_max_decay_threshold << ") hit at " << session_recall_throttle << "; skipping!" << dendl; | |
1588 | throttled = true; | |
1589 | continue; | |
11fdf7f2 TL |
1590 | } else if (session_recall_throttle2o+recall > recall_max_caps*2) { |
1591 | dout(15) << " session recall 2nd-order threshold (" << 2*recall_max_caps << ") hit at " << session_recall_throttle2o << "; skipping!" << dendl; | |
1592 | throttled = true; | |
1593 | continue; | |
a8e16298 TL |
1594 | } else if (global_recall_throttle+recall > recall_global_max_decay_threshold) { |
1595 | dout(15) << " global recall threshold (" << recall_global_max_decay_threshold << ") hit at " << global_recall_throttle << "; skipping!" << dendl; | |
1596 | throttled = true; | |
1597 | break; | |
1598 | } | |
1599 | ||
1600 | // now check if we've recalled caps recently and the client is unlikely to satisfy a new recall | |
1601 | if (steady) { | |
1602 | const auto session_recall = session->get_recall_caps(); | |
1603 | const auto session_release = session->get_release_caps(); | |
1604 | if (2*session_release < session_recall && 2*session_recall > recall_max_decay_threshold) { | |
1605 | /* The session has been unable to keep up with the number of caps | |
1606 | * recalled (by half); additionally, to prevent marking sessions | |
1607 | * we've just begun to recall from, the session_recall counter | |
1608 | * (decayed count of caps recently recalled) is **greater** than the | |
1609 | * session threshold for the session's cap recall throttle. | |
1610 | */ | |
1611 | dout(15) << " 2*session_release < session_recall" | |
11fdf7f2 TL |
1612 | " (2*" << session_release << " < " << session_recall << ") &&" |
1613 | " 2*session_recall < recall_max_decay_threshold" | |
1614 | " (2*" << session_recall << " > " << recall_max_decay_threshold << ")" | |
a8e16298 TL |
1615 | " Skipping because we are unlikely to get more released." << dendl; |
1616 | continue; | |
1617 | } else if (recall < recall_max_caps && 2*recall < session_recall) { | |
1618 | /* The number of caps recalled is less than the number we *could* | |
1619 | * recall (so there isn't much left to recall?) and the number of | |
1620 | * caps is less than the current recall_caps counter (decayed count | |
1621 | * of caps recently recalled). | |
1622 | */ | |
1623 | dout(15) << " 2*recall < session_recall " | |
1624 | " (2*" << recall << " < " << session_recall << ") &&" | |
1625 | " recall < recall_max_caps (" << recall << " < " << recall_max_caps << ");" | |
1626 | " Skipping because we are unlikely to get more released." << dendl; | |
1627 | continue; | |
1628 | } | |
1629 | } | |
1630 | ||
1631 | dout(7) << " recalling " << recall << " caps; session_recall_throttle = " << session_recall_throttle << "; global_recall_throttle = " << global_recall_throttle << dendl; | |
1632 | ||
11fdf7f2 | 1633 | auto m = MClientSession::create(CEPH_SESSION_RECALL_STATE); |
3efd9988 FG |
1634 | m->head.max_caps = newlim; |
1635 | mds->send_message_client(m, session); | |
a8e16298 | 1636 | if (gather) { |
f64942e4 AA |
1637 | flush_session(session, gather); |
1638 | } | |
a8e16298 | 1639 | caps_recalled += session->notify_recall_sent(newlim); |
11fdf7f2 | 1640 | recall_throttle.hit(recall); |
7c673cae FG |
1641 | } |
1642 | } | |
a8e16298 TL |
1643 | |
1644 | dout(7) << "recalled" << (throttled ? " (throttled)" : "") << " " << caps_recalled << " client caps." << dendl; | |
1645 | ||
1646 | return result; | |
7c673cae FG |
1647 | } |
1648 | ||
1649 | void Server::force_clients_readonly() | |
1650 | { | |
1651 | dout(10) << "force_clients_readonly" << dendl; | |
1652 | set<Session*> sessions; | |
1653 | mds->sessionmap.get_client_session_set(sessions); | |
1654 | for (set<Session*>::const_iterator p = sessions.begin(); | |
1655 | p != sessions.end(); | |
1656 | ++p) { | |
1657 | Session *session = *p; | |
1658 | if (!session->info.inst.name.is_client() || | |
1659 | !(session->is_open() || session->is_stale())) | |
1660 | continue; | |
11fdf7f2 | 1661 | mds->send_message_client(MClientSession::create(CEPH_SESSION_FORCE_RO), session); |
7c673cae FG |
1662 | } |
1663 | } | |
1664 | ||
1665 | /******* | |
1666 | * some generic stuff for finishing off requests | |
1667 | */ | |
1668 | void Server::journal_and_reply(MDRequestRef& mdr, CInode *in, CDentry *dn, LogEvent *le, MDSLogContextBase *fin) | |
1669 | { | |
1670 | dout(10) << "journal_and_reply tracei " << in << " tracedn " << dn << dendl; | |
11fdf7f2 | 1671 | ceph_assert(!mdr->has_completed); |
7c673cae FG |
1672 | |
1673 | // note trace items for eventual reply. | |
1674 | mdr->tracei = in; | |
1675 | if (in) | |
1676 | mdr->pin(in); | |
1677 | ||
1678 | mdr->tracedn = dn; | |
1679 | if (dn) | |
1680 | mdr->pin(dn); | |
1681 | ||
1682 | early_reply(mdr, in, dn); | |
1683 | ||
1684 | mdr->committing = true; | |
1685 | submit_mdlog_entry(le, fin, mdr, __func__); | |
1686 | ||
1687 | if (mdr->client_request && mdr->client_request->is_queued_for_replay()) { | |
1688 | if (mds->queue_one_replay()) { | |
1689 | dout(10) << " queued next replay op" << dendl; | |
1690 | } else { | |
11fdf7f2 | 1691 | dout(10) << " journaled last replay op" << dendl; |
7c673cae FG |
1692 | } |
1693 | } else if (mdr->did_early_reply) | |
b32b8144 | 1694 | mds->locker->drop_rdlocks_for_early_reply(mdr.get()); |
7c673cae FG |
1695 | else |
1696 | mdlog->flush(); | |
1697 | } | |
1698 | ||
1699 | void Server::submit_mdlog_entry(LogEvent *le, MDSLogContextBase *fin, MDRequestRef& mdr, | |
11fdf7f2 | 1700 | std::string_view event) |
7c673cae FG |
1701 | { |
1702 | if (mdr) { | |
1703 | string event_str("submit entry: "); | |
1704 | event_str += event; | |
11fdf7f2 | 1705 | mdr->mark_event(event_str); |
7c673cae FG |
1706 | } |
1707 | mdlog->submit_entry(le, fin); | |
1708 | } | |
1709 | ||
1710 | /* | |
1711 | * send response built from mdr contents and error code; clean up mdr | |
1712 | */ | |
1713 | void Server::respond_to_request(MDRequestRef& mdr, int r) | |
1714 | { | |
1715 | if (mdr->client_request) { | |
11fdf7f2 | 1716 | reply_client_request(mdr, MClientReply::create(*mdr->client_request, r)); |
7c673cae FG |
1717 | } else if (mdr->internal_op > -1) { |
1718 | dout(10) << "respond_to_request on internal request " << mdr << dendl; | |
1719 | if (!mdr->internal_op_finish) | |
11fdf7f2 | 1720 | ceph_abort_msg("trying to respond to internal op without finisher"); |
7c673cae FG |
1721 | mdr->internal_op_finish->complete(r); |
1722 | mdcache->request_finish(mdr); | |
1723 | } | |
1724 | } | |
1725 | ||
91327a77 | 1726 | // statistics mds req op number and latency |
11fdf7f2 | 1727 | void Server::perf_gather_op_latency(const MClientRequest::const_ref &req, utime_t lat) |
91327a77 AA |
1728 | { |
1729 | int code = l_mdss_first; | |
1730 | switch(req->get_op()) { | |
1731 | case CEPH_MDS_OP_LOOKUPHASH: | |
1732 | code = l_mdss_req_lookuphash_latency; | |
1733 | break; | |
1734 | case CEPH_MDS_OP_LOOKUPINO: | |
1735 | code = l_mdss_req_lookupino_latency; | |
1736 | break; | |
1737 | case CEPH_MDS_OP_LOOKUPPARENT: | |
1738 | code = l_mdss_req_lookupparent_latency; | |
1739 | break; | |
1740 | case CEPH_MDS_OP_LOOKUPNAME: | |
1741 | code = l_mdss_req_lookupname_latency; | |
1742 | break; | |
1743 | case CEPH_MDS_OP_LOOKUP: | |
1744 | code = l_mdss_req_lookup_latency; | |
1745 | break; | |
1746 | case CEPH_MDS_OP_LOOKUPSNAP: | |
1747 | code = l_mdss_req_lookupsnap_latency; | |
1748 | break; | |
1749 | case CEPH_MDS_OP_GETATTR: | |
1750 | code = l_mdss_req_getattr_latency; | |
1751 | break; | |
1752 | case CEPH_MDS_OP_SETATTR: | |
1753 | code = l_mdss_req_setattr_latency; | |
1754 | break; | |
1755 | case CEPH_MDS_OP_SETLAYOUT: | |
1756 | code = l_mdss_req_setlayout_latency; | |
1757 | break; | |
1758 | case CEPH_MDS_OP_SETDIRLAYOUT: | |
1759 | code = l_mdss_req_setdirlayout_latency; | |
1760 | break; | |
1761 | case CEPH_MDS_OP_SETXATTR: | |
1762 | code = l_mdss_req_setxattr_latency; | |
1763 | break; | |
1764 | case CEPH_MDS_OP_RMXATTR: | |
1765 | code = l_mdss_req_rmxattr_latency; | |
1766 | break; | |
1767 | case CEPH_MDS_OP_READDIR: | |
1768 | code = l_mdss_req_readdir_latency; | |
1769 | break; | |
1770 | case CEPH_MDS_OP_SETFILELOCK: | |
1771 | code = l_mdss_req_setfilelock_latency; | |
1772 | break; | |
1773 | case CEPH_MDS_OP_GETFILELOCK: | |
1774 | code = l_mdss_req_getfilelock_latency; | |
1775 | break; | |
1776 | case CEPH_MDS_OP_CREATE: | |
1777 | code = l_mdss_req_create_latency; | |
1778 | break; | |
1779 | case CEPH_MDS_OP_OPEN: | |
1780 | code = l_mdss_req_open_latency; | |
1781 | break; | |
1782 | case CEPH_MDS_OP_MKNOD: | |
1783 | code = l_mdss_req_mknod_latency; | |
1784 | break; | |
1785 | case CEPH_MDS_OP_LINK: | |
1786 | code = l_mdss_req_link_latency; | |
1787 | break; | |
1788 | case CEPH_MDS_OP_UNLINK: | |
1789 | code = l_mdss_req_unlink_latency; | |
1790 | break; | |
1791 | case CEPH_MDS_OP_RMDIR: | |
1792 | code = l_mdss_req_rmdir_latency; | |
1793 | break; | |
1794 | case CEPH_MDS_OP_RENAME: | |
1795 | code = l_mdss_req_rename_latency; | |
1796 | break; | |
1797 | case CEPH_MDS_OP_MKDIR: | |
1798 | code = l_mdss_req_mkdir_latency; | |
1799 | break; | |
1800 | case CEPH_MDS_OP_SYMLINK: | |
1801 | code = l_mdss_req_symlink_latency; | |
1802 | break; | |
1803 | case CEPH_MDS_OP_LSSNAP: | |
1804 | code = l_mdss_req_lssnap_latency; | |
1805 | break; | |
1806 | case CEPH_MDS_OP_MKSNAP: | |
1807 | code = l_mdss_req_mksnap_latency; | |
1808 | break; | |
1809 | case CEPH_MDS_OP_RMSNAP: | |
1810 | code = l_mdss_req_rmsnap_latency; | |
1811 | break; | |
1812 | case CEPH_MDS_OP_RENAMESNAP: | |
1813 | code = l_mdss_req_renamesnap_latency; | |
1814 | break; | |
1815 | default: ceph_abort(); | |
1816 | } | |
1817 | logger->tinc(code, lat); | |
1818 | } | |
1819 | ||
7c673cae FG |
1820 | void Server::early_reply(MDRequestRef& mdr, CInode *tracei, CDentry *tracedn) |
1821 | { | |
11fdf7f2 | 1822 | if (!g_conf()->mds_early_reply) |
7c673cae FG |
1823 | return; |
1824 | ||
b32b8144 FG |
1825 | if (mdr->no_early_reply) { |
1826 | dout(10) << "early_reply - flag no_early_reply is set, not allowed." << dendl; | |
1827 | return; | |
1828 | } | |
1829 | ||
7c673cae FG |
1830 | if (mdr->has_more() && mdr->more()->has_journaled_slaves) { |
1831 | dout(10) << "early_reply - there are journaled slaves, not allowed." << dendl; | |
1832 | return; | |
1833 | } | |
1834 | ||
1835 | if (mdr->alloc_ino) { | |
1836 | dout(10) << "early_reply - allocated ino, not allowed" << dendl; | |
1837 | return; | |
1838 | } | |
1839 | ||
11fdf7f2 | 1840 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
1841 | entity_inst_t client_inst = req->get_source_inst(); |
1842 | if (client_inst.name.is_mds()) | |
1843 | return; | |
1844 | ||
1845 | if (req->is_replay()) { | |
1846 | dout(10) << " no early reply on replay op" << dendl; | |
1847 | return; | |
1848 | } | |
1849 | ||
1850 | ||
11fdf7f2 | 1851 | auto reply = MClientReply::create(*req, 0); |
7c673cae FG |
1852 | reply->set_unsafe(); |
1853 | ||
1854 | // mark xlocks "done", indicating that we are exposing uncommitted changes. | |
1855 | // | |
1856 | //_rename_finish() does not send dentry link/unlink message to replicas. | |
1857 | // so do not set xlocks on dentries "done", the xlocks prevent dentries | |
1858 | // that have projected linkages from getting new replica. | |
1859 | mds->locker->set_xlocks_done(mdr.get(), req->get_op() == CEPH_MDS_OP_RENAME); | |
1860 | ||
1861 | dout(10) << "early_reply " << reply->get_result() | |
1862 | << " (" << cpp_strerror(reply->get_result()) | |
1863 | << ") " << *req << dendl; | |
1864 | ||
1865 | if (tracei || tracedn) { | |
1866 | if (tracei) | |
1867 | mdr->cap_releases.erase(tracei->vino()); | |
1868 | if (tracedn) | |
1869 | mdr->cap_releases.erase(tracedn->get_dir()->get_inode()->vino()); | |
1870 | ||
1871 | set_trace_dist(mdr->session, reply, tracei, tracedn, mdr->snapid, | |
1872 | req->get_dentry_wanted(), mdr); | |
1873 | } | |
1874 | ||
1875 | reply->set_extra_bl(mdr->reply_extra_bl); | |
11fdf7f2 | 1876 | mds->send_message_client(reply, mdr->session); |
7c673cae FG |
1877 | |
1878 | mdr->did_early_reply = true; | |
1879 | ||
1880 | mds->logger->inc(l_mds_reply); | |
1881 | utime_t lat = ceph_clock_now() - req->get_recv_stamp(); | |
1882 | mds->logger->tinc(l_mds_reply_latency, lat); | |
91327a77 AA |
1883 | if (client_inst.name.is_client()) { |
1884 | mds->sessionmap.hit_session(mdr->session); | |
1885 | } | |
1886 | perf_gather_op_latency(req, lat); | |
7c673cae FG |
1887 | dout(20) << "lat " << lat << dendl; |
1888 | ||
1889 | mdr->mark_event("early_replied"); | |
1890 | } | |
1891 | ||
1892 | /* | |
1893 | * send given reply | |
1894 | * include a trace to tracei | |
1895 | * Clean up mdr | |
1896 | */ | |
11fdf7f2 | 1897 | void Server::reply_client_request(MDRequestRef& mdr, const MClientReply::ref &reply) |
7c673cae | 1898 | { |
11fdf7f2 TL |
1899 | ceph_assert(mdr.get()); |
1900 | const MClientRequest::const_ref &req = mdr->client_request; | |
7c673cae FG |
1901 | |
1902 | dout(7) << "reply_client_request " << reply->get_result() | |
1903 | << " (" << cpp_strerror(reply->get_result()) | |
1904 | << ") " << *req << dendl; | |
1905 | ||
1906 | mdr->mark_event("replying"); | |
1907 | ||
1908 | Session *session = mdr->session; | |
1909 | ||
1910 | // note successful request in session map? | |
1911 | // | |
1912 | // setfilelock requests are special, they only modify states in MDS memory. | |
1913 | // The states get lost when MDS fails. If Client re-send a completed | |
1914 | // setfilelock request, it means that client did not receive corresponding | |
1915 | // setfilelock reply. So MDS should re-execute the setfilelock request. | |
1916 | if (req->may_write() && req->get_op() != CEPH_MDS_OP_SETFILELOCK && | |
1917 | reply->get_result() == 0 && session) { | |
1918 | inodeno_t created = mdr->alloc_ino ? mdr->alloc_ino : mdr->used_prealloc_ino; | |
1919 | session->add_completed_request(mdr->reqid.tid, created); | |
1920 | if (mdr->ls) { | |
1921 | mdr->ls->touched_sessions.insert(session->info.inst.name); | |
1922 | } | |
1923 | } | |
1924 | ||
1925 | // give any preallocated inos to the session | |
1926 | apply_allocated_inos(mdr, session); | |
1927 | ||
1928 | // get tracei/tracedn from mdr? | |
1929 | snapid_t snapid = mdr->snapid; | |
1930 | CInode *tracei = mdr->tracei; | |
1931 | CDentry *tracedn = mdr->tracedn; | |
1932 | ||
1933 | bool is_replay = mdr->client_request->is_replay(); | |
1934 | bool did_early_reply = mdr->did_early_reply; | |
1935 | entity_inst_t client_inst = req->get_source_inst(); | |
1936 | int dentry_wanted = req->get_dentry_wanted(); | |
1937 | ||
1938 | if (!did_early_reply && !is_replay) { | |
1939 | ||
1940 | mds->logger->inc(l_mds_reply); | |
1941 | utime_t lat = ceph_clock_now() - mdr->client_request->get_recv_stamp(); | |
1942 | mds->logger->tinc(l_mds_reply_latency, lat); | |
91327a77 AA |
1943 | if (client_inst.name.is_client()) { |
1944 | mds->sessionmap.hit_session(session); | |
1945 | } | |
1946 | perf_gather_op_latency(req, lat); | |
7c673cae FG |
1947 | dout(20) << "lat " << lat << dendl; |
1948 | ||
1949 | if (tracei) | |
1950 | mdr->cap_releases.erase(tracei->vino()); | |
1951 | if (tracedn) | |
1952 | mdr->cap_releases.erase(tracedn->get_dir()->get_inode()->vino()); | |
1953 | } | |
1954 | ||
1955 | // drop non-rdlocks before replying, so that we can issue leases | |
1956 | mdcache->request_drop_non_rdlocks(mdr); | |
1957 | ||
1958 | // reply at all? | |
11fdf7f2 | 1959 | if (!(client_inst.name.is_mds() || !session)) { |
7c673cae FG |
1960 | // send reply. |
1961 | if (!did_early_reply && // don't issue leases if we sent an earlier reply already | |
1962 | (tracei || tracedn)) { | |
1963 | if (is_replay) { | |
1964 | if (tracei) | |
1965 | mdcache->try_reconnect_cap(tracei, session); | |
1966 | } else { | |
1967 | // include metadata in reply | |
1968 | set_trace_dist(session, reply, tracei, tracedn, | |
1969 | snapid, dentry_wanted, | |
1970 | mdr); | |
1971 | } | |
1972 | } | |
1973 | ||
1974 | // We can set the extra bl unconditionally: if it's already been sent in the | |
1975 | // early_reply, set_extra_bl will have claimed it and reply_extra_bl is empty | |
1976 | reply->set_extra_bl(mdr->reply_extra_bl); | |
1977 | ||
1978 | reply->set_mdsmap_epoch(mds->mdsmap->get_epoch()); | |
11fdf7f2 | 1979 | mds->send_message_client(reply, session); |
7c673cae FG |
1980 | } |
1981 | ||
1982 | if (req->is_queued_for_replay() && | |
1983 | (mdr->has_completed || reply->get_result() < 0)) { | |
1984 | if (reply->get_result() < 0) { | |
1985 | int r = reply->get_result(); | |
1986 | derr << "reply_client_request: failed to replay " << *req | |
1987 | << " error " << r << " (" << cpp_strerror(r) << ")" << dendl; | |
1988 | mds->clog->warn() << "failed to replay " << req->get_reqid() << " error " << r; | |
1989 | } | |
1990 | mds->queue_one_replay(); | |
1991 | } | |
1992 | ||
1993 | // clean up request | |
1994 | mdcache->request_finish(mdr); | |
1995 | ||
1996 | // take a closer look at tracei, if it happens to be a remote link | |
1997 | if (tracei && | |
1998 | tracedn && | |
1999 | tracedn->get_projected_linkage()->is_remote()) { | |
2000 | mdcache->eval_remote(tracedn); | |
2001 | } | |
2002 | } | |
2003 | ||
7c673cae FG |
2004 | /* |
2005 | * pass inode OR dentry (not both, or we may get confused) | |
2006 | * | |
2007 | * trace is in reverse order (i.e. root inode comes last) | |
2008 | */ | |
11fdf7f2 | 2009 | void Server::set_trace_dist(Session *session, const MClientReply::ref &reply, |
7c673cae FG |
2010 | CInode *in, CDentry *dn, |
2011 | snapid_t snapid, | |
2012 | int dentry_wanted, | |
2013 | MDRequestRef& mdr) | |
2014 | { | |
2015 | // skip doing this for debugging purposes? | |
11fdf7f2 | 2016 | if (g_conf()->mds_inject_traceless_reply_probability && |
7c673cae | 2017 | mdr->ls && !mdr->o_trunc && |
11fdf7f2 | 2018 | (rand() % 10000 < g_conf()->mds_inject_traceless_reply_probability * 10000.0)) { |
7c673cae FG |
2019 | dout(5) << "deliberately skipping trace for " << *reply << dendl; |
2020 | return; | |
2021 | } | |
2022 | ||
2023 | // inode, dentry, dir, ..., inode | |
2024 | bufferlist bl; | |
2025 | mds_rank_t whoami = mds->get_nodeid(); | |
2026 | client_t client = session->get_client(); | |
2027 | utime_t now = ceph_clock_now(); | |
2028 | ||
2029 | dout(20) << "set_trace_dist snapid " << snapid << dendl; | |
2030 | ||
2031 | //assert((bool)dn == (bool)dentry_wanted); // not true for snapshot lookups | |
2032 | ||
2033 | // realm | |
2034 | if (snapid == CEPH_NOSNAP) { | |
2035 | SnapRealm *realm; | |
2036 | if (in) | |
2037 | realm = in->find_snaprealm(); | |
2038 | else | |
2039 | realm = dn->get_dir()->get_inode()->find_snaprealm(); | |
2040 | reply->snapbl = realm->get_snap_trace(); | |
2041 | dout(10) << "set_trace_dist snaprealm " << *realm << " len=" << reply->snapbl.length() << dendl; | |
2042 | } | |
2043 | ||
2044 | // dir + dentry? | |
2045 | if (dn) { | |
2046 | reply->head.is_dentry = 1; | |
2047 | CDir *dir = dn->get_dir(); | |
2048 | CInode *diri = dir->get_inode(); | |
2049 | ||
2050 | diri->encode_inodestat(bl, session, NULL, snapid); | |
2051 | dout(20) << "set_trace_dist added diri " << *diri << dendl; | |
2052 | ||
2053 | #ifdef MDS_VERIFY_FRAGSTAT | |
2054 | if (dir->is_complete()) | |
2055 | dir->verify_fragstat(); | |
2056 | #endif | |
11fdf7f2 TL |
2057 | DirStat ds; |
2058 | ds.frag = dir->get_frag(); | |
2059 | ds.auth = dir->get_dir_auth().first; | |
2060 | if (dir->is_auth()) | |
2061 | dir->get_dist_spec(ds.dist, whoami); | |
2062 | ||
2063 | dir->encode_dirstat(bl, session->info, ds); | |
7c673cae FG |
2064 | dout(20) << "set_trace_dist added dir " << *dir << dendl; |
2065 | ||
11fdf7f2 | 2066 | encode(dn->get_name(), bl); |
7c673cae FG |
2067 | if (snapid == CEPH_NOSNAP) |
2068 | mds->locker->issue_client_lease(dn, client, bl, now, session); | |
11fdf7f2 TL |
2069 | else { |
2070 | //null lease | |
2071 | LeaseStat e; | |
2072 | mds->locker->encode_lease(bl, session->info, e); | |
2073 | } | |
7c673cae FG |
2074 | dout(20) << "set_trace_dist added dn " << snapid << " " << *dn << dendl; |
2075 | } else | |
2076 | reply->head.is_dentry = 0; | |
2077 | ||
2078 | // inode | |
2079 | if (in) { | |
2080 | in->encode_inodestat(bl, session, NULL, snapid, 0, mdr->getattr_caps); | |
2081 | dout(20) << "set_trace_dist added in " << *in << dendl; | |
2082 | reply->head.is_target = 1; | |
2083 | } else | |
2084 | reply->head.is_target = 0; | |
2085 | ||
2086 | reply->set_trace(bl); | |
2087 | } | |
2088 | ||
11fdf7f2 | 2089 | void Server::handle_client_request(const MClientRequest::const_ref &req) |
7c673cae FG |
2090 | { |
2091 | dout(4) << "handle_client_request " << *req << dendl; | |
2092 | ||
2093 | if (mds->logger) | |
2094 | mds->logger->inc(l_mds_request); | |
2095 | if (logger) | |
2096 | logger->inc(l_mdss_handle_client_request); | |
2097 | ||
2098 | if (!mdcache->is_open()) { | |
2099 | dout(5) << "waiting for root" << dendl; | |
2100 | mdcache->wait_for_open(new C_MDS_RetryMessage(mds, req)); | |
2101 | return; | |
2102 | } | |
2103 | ||
2104 | // active session? | |
2105 | Session *session = 0; | |
2106 | if (req->get_source().is_client()) { | |
94b18763 | 2107 | session = mds->get_session(req); |
7c673cae FG |
2108 | if (!session) { |
2109 | dout(5) << "no session for " << req->get_source() << ", dropping" << dendl; | |
2110 | } else if (session->is_closed() || | |
2111 | session->is_closing() || | |
2112 | session->is_killing()) { | |
2113 | dout(5) << "session closed|closing|killing, dropping" << dendl; | |
2114 | session = NULL; | |
2115 | } | |
2116 | if (!session) { | |
2117 | if (req->is_queued_for_replay()) | |
2118 | mds->queue_one_replay(); | |
7c673cae FG |
2119 | return; |
2120 | } | |
2121 | } | |
2122 | ||
2123 | // old mdsmap? | |
2124 | if (req->get_mdsmap_epoch() < mds->mdsmap->get_epoch()) { | |
2125 | // send it? hrm, this isn't ideal; they may get a lot of copies if | |
2126 | // they have a high request rate. | |
2127 | } | |
2128 | ||
2129 | // completed request? | |
2130 | bool has_completed = false; | |
2131 | if (req->is_replay() || req->get_retry_attempt()) { | |
11fdf7f2 | 2132 | ceph_assert(session); |
7c673cae FG |
2133 | inodeno_t created; |
2134 | if (session->have_completed_request(req->get_reqid().tid, &created)) { | |
2135 | has_completed = true; | |
2136 | // Don't send traceless reply if the completed request has created | |
2137 | // new inode. Treat the request as lookup request instead. | |
2138 | if (req->is_replay() || | |
2139 | ((created == inodeno_t() || !mds->is_clientreplay()) && | |
2140 | req->get_op() != CEPH_MDS_OP_OPEN && | |
2141 | req->get_op() != CEPH_MDS_OP_CREATE)) { | |
2142 | dout(5) << "already completed " << req->get_reqid() << dendl; | |
11fdf7f2 | 2143 | auto reply = MClientReply::create(*req, 0); |
7c673cae FG |
2144 | if (created != inodeno_t()) { |
2145 | bufferlist extra; | |
11fdf7f2 | 2146 | encode(created, extra); |
7c673cae FG |
2147 | reply->set_extra_bl(extra); |
2148 | } | |
11fdf7f2 | 2149 | mds->send_message_client(reply, session); |
7c673cae FG |
2150 | |
2151 | if (req->is_queued_for_replay()) | |
2152 | mds->queue_one_replay(); | |
2153 | ||
7c673cae FG |
2154 | return; |
2155 | } | |
2156 | if (req->get_op() != CEPH_MDS_OP_OPEN && | |
2157 | req->get_op() != CEPH_MDS_OP_CREATE) { | |
2158 | dout(10) << " completed request which created new inode " << created | |
2159 | << ", convert it to lookup request" << dendl; | |
2160 | req->head.op = req->get_dentry_wanted() ? CEPH_MDS_OP_LOOKUP : CEPH_MDS_OP_GETATTR; | |
2161 | req->head.args.getattr.mask = CEPH_STAT_CAP_INODE_ALL; | |
2162 | } | |
2163 | } | |
2164 | } | |
2165 | ||
2166 | // trim completed_request list | |
2167 | if (req->get_oldest_client_tid() > 0) { | |
2168 | dout(15) << " oldest_client_tid=" << req->get_oldest_client_tid() << dendl; | |
11fdf7f2 | 2169 | ceph_assert(session); |
7c673cae FG |
2170 | if (session->trim_completed_requests(req->get_oldest_client_tid())) { |
2171 | // Sessions 'completed_requests' was dirtied, mark it to be | |
2172 | // potentially flushed at segment expiry. | |
2173 | mdlog->get_current_segment()->touched_sessions.insert(session->info.inst.name); | |
2174 | ||
2175 | if (session->get_num_trim_requests_warnings() > 0 && | |
11fdf7f2 | 2176 | session->get_num_completed_requests() * 2 < g_conf()->mds_max_completed_requests) |
7c673cae FG |
2177 | session->reset_num_trim_requests_warnings(); |
2178 | } else { | |
2179 | if (session->get_num_completed_requests() >= | |
11fdf7f2 | 2180 | (g_conf()->mds_max_completed_requests << session->get_num_trim_requests_warnings())) { |
7c673cae FG |
2181 | session->inc_num_trim_requests_warnings(); |
2182 | stringstream ss; | |
2183 | ss << "client." << session->get_client() << " does not advance its oldest_client_tid (" | |
2184 | << req->get_oldest_client_tid() << "), " | |
2185 | << session->get_num_completed_requests() | |
2186 | << " completed requests recorded in session\n"; | |
2187 | mds->clog->warn() << ss.str(); | |
2188 | dout(20) << __func__ << " " << ss.str() << dendl; | |
2189 | } | |
2190 | } | |
2191 | } | |
2192 | ||
2193 | // register + dispatch | |
2194 | MDRequestRef mdr = mdcache->request_start(req); | |
2195 | if (!mdr.get()) | |
2196 | return; | |
2197 | ||
2198 | if (session) { | |
2199 | mdr->session = session; | |
2200 | session->requests.push_back(&mdr->item_session_request); | |
2201 | } | |
2202 | ||
2203 | if (has_completed) | |
2204 | mdr->has_completed = true; | |
2205 | ||
2206 | // process embedded cap releases? | |
2207 | // (only if NOT replay!) | |
2208 | if (!req->releases.empty() && req->get_source().is_client() && !req->is_replay()) { | |
2209 | client_t client = req->get_source().num(); | |
11fdf7f2 TL |
2210 | for (const auto &r : req->releases) { |
2211 | mds->locker->process_request_cap_release(mdr, client, r.item, r.dname); | |
2212 | } | |
7c673cae FG |
2213 | req->releases.clear(); |
2214 | } | |
2215 | ||
2216 | dispatch_client_request(mdr); | |
2217 | return; | |
2218 | } | |
2219 | ||
2220 | void Server::handle_osd_map() | |
2221 | { | |
2222 | /* Note that we check the OSDMAP_FULL flag directly rather than | |
2223 | * using osdmap_full_flag(), because we want to know "is the flag set" | |
2224 | * rather than "does the flag apply to us?" */ | |
2225 | mds->objecter->with_osdmap([this](const OSDMap& o) { | |
b32b8144 FG |
2226 | auto pi = o.get_pg_pool(mds->mdsmap->get_metadata_pool()); |
2227 | is_full = pi && pi->has_flag(pg_pool_t::FLAG_FULL); | |
7c673cae FG |
2228 | dout(7) << __func__ << ": full = " << is_full << " epoch = " |
2229 | << o.get_epoch() << dendl; | |
2230 | }); | |
2231 | } | |
2232 | ||
2233 | void Server::dispatch_client_request(MDRequestRef& mdr) | |
2234 | { | |
2235 | // we shouldn't be waiting on anyone. | |
11fdf7f2 | 2236 | ceph_assert(!mdr->has_more() || mdr->more()->waiting_on_slave.empty()); |
7c673cae FG |
2237 | |
2238 | if (mdr->killed) { | |
2239 | dout(10) << "request " << *mdr << " was killed" << dendl; | |
2240 | return; | |
94b18763 FG |
2241 | } else if (mdr->aborted) { |
2242 | mdr->aborted = false; | |
2243 | mdcache->request_kill(mdr); | |
2244 | return; | |
7c673cae FG |
2245 | } |
2246 | ||
11fdf7f2 | 2247 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
2248 | |
2249 | if (logger) logger->inc(l_mdss_dispatch_client_request); | |
2250 | ||
2251 | dout(7) << "dispatch_client_request " << *req << dendl; | |
2252 | ||
2253 | if (req->may_write()) { | |
2254 | if (mdcache->is_readonly()) { | |
2255 | dout(10) << " read-only FS" << dendl; | |
2256 | respond_to_request(mdr, -EROFS); | |
2257 | return; | |
2258 | } | |
2259 | if (mdr->has_more() && mdr->more()->slave_error) { | |
2260 | dout(10) << " got error from slaves" << dendl; | |
2261 | respond_to_request(mdr, mdr->more()->slave_error); | |
2262 | return; | |
2263 | } | |
2264 | } | |
2265 | ||
2266 | if (is_full) { | |
2267 | if (req->get_op() == CEPH_MDS_OP_SETLAYOUT || | |
2268 | req->get_op() == CEPH_MDS_OP_SETDIRLAYOUT || | |
2269 | req->get_op() == CEPH_MDS_OP_SETLAYOUT || | |
2270 | req->get_op() == CEPH_MDS_OP_RMXATTR || | |
2271 | req->get_op() == CEPH_MDS_OP_SETXATTR || | |
2272 | req->get_op() == CEPH_MDS_OP_CREATE || | |
2273 | req->get_op() == CEPH_MDS_OP_SYMLINK || | |
2274 | req->get_op() == CEPH_MDS_OP_MKSNAP || | |
2275 | ((req->get_op() == CEPH_MDS_OP_LINK || | |
2276 | req->get_op() == CEPH_MDS_OP_RENAME) && | |
2277 | (!mdr->has_more() || mdr->more()->witnessed.empty())) // haven't started slave request | |
2278 | ) { | |
2279 | ||
2280 | dout(20) << __func__ << ": full, responding ENOSPC to op " << ceph_mds_op_name(req->get_op()) << dendl; | |
2281 | respond_to_request(mdr, -ENOSPC); | |
2282 | return; | |
2283 | } else { | |
2284 | dout(20) << __func__ << ": full, permitting op " << ceph_mds_op_name(req->get_op()) << dendl; | |
2285 | } | |
2286 | } | |
2287 | ||
2288 | switch (req->get_op()) { | |
2289 | case CEPH_MDS_OP_LOOKUPHASH: | |
2290 | case CEPH_MDS_OP_LOOKUPINO: | |
2291 | handle_client_lookup_ino(mdr, false, false); | |
2292 | break; | |
2293 | case CEPH_MDS_OP_LOOKUPPARENT: | |
2294 | handle_client_lookup_ino(mdr, true, false); | |
2295 | break; | |
2296 | case CEPH_MDS_OP_LOOKUPNAME: | |
2297 | handle_client_lookup_ino(mdr, false, true); | |
2298 | break; | |
2299 | ||
2300 | // inodes ops. | |
2301 | case CEPH_MDS_OP_LOOKUP: | |
2302 | handle_client_getattr(mdr, true); | |
2303 | break; | |
2304 | ||
2305 | case CEPH_MDS_OP_LOOKUPSNAP: | |
2306 | // lookupsnap does not reference a CDentry; treat it as a getattr | |
2307 | case CEPH_MDS_OP_GETATTR: | |
2308 | handle_client_getattr(mdr, false); | |
2309 | break; | |
2310 | ||
2311 | case CEPH_MDS_OP_SETATTR: | |
2312 | handle_client_setattr(mdr); | |
2313 | break; | |
2314 | case CEPH_MDS_OP_SETLAYOUT: | |
2315 | handle_client_setlayout(mdr); | |
2316 | break; | |
2317 | case CEPH_MDS_OP_SETDIRLAYOUT: | |
2318 | handle_client_setdirlayout(mdr); | |
2319 | break; | |
2320 | case CEPH_MDS_OP_SETXATTR: | |
2321 | handle_client_setxattr(mdr); | |
2322 | break; | |
2323 | case CEPH_MDS_OP_RMXATTR: | |
2324 | handle_client_removexattr(mdr); | |
2325 | break; | |
2326 | ||
2327 | case CEPH_MDS_OP_READDIR: | |
2328 | handle_client_readdir(mdr); | |
2329 | break; | |
2330 | ||
2331 | case CEPH_MDS_OP_SETFILELOCK: | |
2332 | handle_client_file_setlock(mdr); | |
2333 | break; | |
2334 | ||
2335 | case CEPH_MDS_OP_GETFILELOCK: | |
2336 | handle_client_file_readlock(mdr); | |
2337 | break; | |
2338 | ||
2339 | // funky. | |
2340 | case CEPH_MDS_OP_CREATE: | |
2341 | if (mdr->has_completed) | |
2342 | handle_client_open(mdr); // already created.. just open | |
2343 | else | |
2344 | handle_client_openc(mdr); | |
2345 | break; | |
2346 | ||
2347 | case CEPH_MDS_OP_OPEN: | |
2348 | handle_client_open(mdr); | |
2349 | break; | |
2350 | ||
2351 | // namespace. | |
2352 | // no prior locks. | |
2353 | case CEPH_MDS_OP_MKNOD: | |
2354 | handle_client_mknod(mdr); | |
2355 | break; | |
2356 | case CEPH_MDS_OP_LINK: | |
2357 | handle_client_link(mdr); | |
2358 | break; | |
2359 | case CEPH_MDS_OP_UNLINK: | |
2360 | case CEPH_MDS_OP_RMDIR: | |
2361 | handle_client_unlink(mdr); | |
2362 | break; | |
2363 | case CEPH_MDS_OP_RENAME: | |
2364 | handle_client_rename(mdr); | |
2365 | break; | |
2366 | case CEPH_MDS_OP_MKDIR: | |
2367 | handle_client_mkdir(mdr); | |
2368 | break; | |
2369 | case CEPH_MDS_OP_SYMLINK: | |
2370 | handle_client_symlink(mdr); | |
2371 | break; | |
2372 | ||
2373 | ||
2374 | // snaps | |
2375 | case CEPH_MDS_OP_LSSNAP: | |
2376 | handle_client_lssnap(mdr); | |
2377 | break; | |
2378 | case CEPH_MDS_OP_MKSNAP: | |
2379 | handle_client_mksnap(mdr); | |
2380 | break; | |
2381 | case CEPH_MDS_OP_RMSNAP: | |
2382 | handle_client_rmsnap(mdr); | |
2383 | break; | |
2384 | case CEPH_MDS_OP_RENAMESNAP: | |
2385 | handle_client_renamesnap(mdr); | |
2386 | break; | |
2387 | ||
2388 | default: | |
2389 | dout(1) << " unknown client op " << req->get_op() << dendl; | |
2390 | respond_to_request(mdr, -EOPNOTSUPP); | |
2391 | } | |
2392 | } | |
2393 | ||
2394 | ||
2395 | // --------------------------------------- | |
2396 | // SLAVE REQUESTS | |
2397 | ||
11fdf7f2 | 2398 | void Server::handle_slave_request(const MMDSSlaveRequest::const_ref &m) |
7c673cae FG |
2399 | { |
2400 | dout(4) << "handle_slave_request " << m->get_reqid() << " from " << m->get_source() << dendl; | |
2401 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
2402 | ||
2403 | if (logger) logger->inc(l_mdss_handle_slave_request); | |
2404 | ||
2405 | // reply? | |
2406 | if (m->is_reply()) | |
2407 | return handle_slave_request_reply(m); | |
2408 | ||
2409 | // the purpose of rename notify is enforcing causal message ordering. making sure | |
2410 | // bystanders have received all messages from rename srcdn's auth MDS. | |
2411 | if (m->get_op() == MMDSSlaveRequest::OP_RENAMENOTIFY) { | |
11fdf7f2 | 2412 | auto reply = MMDSSlaveRequest::create(m->get_reqid(), m->get_attempt(), MMDSSlaveRequest::OP_RENAMENOTIFYACK); |
7c673cae | 2413 | mds->send_message(reply, m->get_connection()); |
7c673cae FG |
2414 | return; |
2415 | } | |
2416 | ||
2417 | CDentry *straydn = NULL; | |
11fdf7f2 TL |
2418 | if (m->straybl.length() > 0) { |
2419 | straydn = mdcache->add_replica_stray(m->straybl, from); | |
2420 | ceph_assert(straydn); | |
2421 | m->straybl.clear(); | |
7c673cae FG |
2422 | } |
2423 | ||
2424 | // am i a new slave? | |
2425 | MDRequestRef mdr; | |
2426 | if (mdcache->have_request(m->get_reqid())) { | |
2427 | // existing? | |
2428 | mdr = mdcache->request_get(m->get_reqid()); | |
2429 | ||
2430 | // is my request newer? | |
2431 | if (mdr->attempt > m->get_attempt()) { | |
2432 | dout(10) << "local request " << *mdr << " attempt " << mdr->attempt << " > " << m->get_attempt() | |
2433 | << ", dropping " << *m << dendl; | |
7c673cae FG |
2434 | return; |
2435 | } | |
2436 | ||
2437 | ||
2438 | if (mdr->attempt < m->get_attempt()) { | |
2439 | // mine is old, close it out | |
2440 | dout(10) << "local request " << *mdr << " attempt " << mdr->attempt << " < " << m->get_attempt() | |
2441 | << ", closing out" << dendl; | |
2442 | mdcache->request_finish(mdr); | |
2443 | mdr.reset(); | |
2444 | } else if (mdr->slave_to_mds != from) { | |
2445 | dout(10) << "local request " << *mdr << " not slave to mds." << from << dendl; | |
7c673cae FG |
2446 | return; |
2447 | } | |
2448 | ||
2449 | if (m->get_op() == MMDSSlaveRequest::OP_FINISH && m->is_abort()) { | |
2450 | mdr->aborted = true; | |
2451 | if (mdr->slave_request) { | |
2452 | // only abort on-going xlock, wrlock and auth pin | |
11fdf7f2 | 2453 | ceph_assert(!mdr->slave_did_prepare()); |
7c673cae FG |
2454 | } else { |
2455 | mdcache->request_finish(mdr); | |
2456 | } | |
2457 | return; | |
2458 | } | |
2459 | } | |
2460 | if (!mdr.get()) { | |
2461 | // new? | |
2462 | if (m->get_op() == MMDSSlaveRequest::OP_FINISH) { | |
2463 | dout(10) << "missing slave request for " << m->get_reqid() | |
2464 | << " OP_FINISH, must have lost race with a forward" << dendl; | |
7c673cae FG |
2465 | return; |
2466 | } | |
2467 | mdr = mdcache->request_start_slave(m->get_reqid(), m->get_attempt(), m); | |
2468 | mdr->set_op_stamp(m->op_stamp); | |
2469 | } | |
11fdf7f2 | 2470 | ceph_assert(mdr->slave_request == 0); // only one at a time, please! |
7c673cae FG |
2471 | |
2472 | if (straydn) { | |
2473 | mdr->pin(straydn); | |
2474 | mdr->straydn = straydn; | |
2475 | } | |
2476 | ||
2477 | if (!mds->is_clientreplay() && !mds->is_active() && !mds->is_stopping()) { | |
2478 | dout(3) << "not clientreplay|active yet, waiting" << dendl; | |
2479 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, m)); | |
2480 | return; | |
2481 | } else if (mds->is_clientreplay() && !mds->mdsmap->is_clientreplay(from) && | |
2482 | mdr->locks.empty()) { | |
2483 | dout(3) << "not active yet, waiting" << dendl; | |
2484 | mds->wait_for_active(new C_MDS_RetryMessage(mds, m)); | |
2485 | return; | |
2486 | } | |
2487 | ||
91327a77 | 2488 | mdr->reset_slave_request(m); |
7c673cae FG |
2489 | |
2490 | dispatch_slave_request(mdr); | |
2491 | } | |
2492 | ||
11fdf7f2 | 2493 | void Server::handle_slave_request_reply(const MMDSSlaveRequest::const_ref &m) |
7c673cae FG |
2494 | { |
2495 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
2496 | ||
2497 | if (!mds->is_clientreplay() && !mds->is_active() && !mds->is_stopping()) { | |
2498 | metareqid_t r = m->get_reqid(); | |
2499 | if (!mdcache->have_uncommitted_master(r, from)) { | |
2500 | dout(10) << "handle_slave_request_reply ignoring slave reply from mds." | |
2501 | << from << " reqid " << r << dendl; | |
7c673cae FG |
2502 | return; |
2503 | } | |
2504 | dout(3) << "not clientreplay|active yet, waiting" << dendl; | |
2505 | mds->wait_for_replay(new C_MDS_RetryMessage(mds, m)); | |
2506 | return; | |
2507 | } | |
2508 | ||
2509 | if (m->get_op() == MMDSSlaveRequest::OP_COMMITTED) { | |
2510 | metareqid_t r = m->get_reqid(); | |
2511 | mdcache->committed_master_slave(r, from); | |
7c673cae FG |
2512 | return; |
2513 | } | |
2514 | ||
2515 | MDRequestRef mdr = mdcache->request_get(m->get_reqid()); | |
2516 | if (m->get_attempt() != mdr->attempt) { | |
2517 | dout(10) << "handle_slave_request_reply " << *mdr << " ignoring reply from other attempt " | |
2518 | << m->get_attempt() << dendl; | |
7c673cae FG |
2519 | return; |
2520 | } | |
2521 | ||
2522 | switch (m->get_op()) { | |
2523 | case MMDSSlaveRequest::OP_XLOCKACK: | |
2524 | { | |
2525 | // identify lock, master request | |
2526 | SimpleLock *lock = mds->locker->get_lock(m->get_lock_type(), | |
2527 | m->get_object_info()); | |
2528 | mdr->more()->slaves.insert(from); | |
11fdf7f2 | 2529 | lock->decode_locked_state(m->get_lock_data()); |
7c673cae | 2530 | dout(10) << "got remote xlock on " << *lock << " on " << *lock->get_parent() << dendl; |
11fdf7f2 | 2531 | mdr->locks.emplace_hint(mdr->locks.end(), lock, MutationImpl::LockOp::XLOCK); |
7c673cae FG |
2532 | mdr->finish_locking(lock); |
2533 | lock->get_xlock(mdr, mdr->get_client()); | |
2534 | ||
11fdf7f2 | 2535 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae | 2536 | mdr->more()->waiting_on_slave.erase(from); |
11fdf7f2 | 2537 | ceph_assert(mdr->more()->waiting_on_slave.empty()); |
7c673cae FG |
2538 | mdcache->dispatch_request(mdr); |
2539 | } | |
2540 | break; | |
2541 | ||
2542 | case MMDSSlaveRequest::OP_WRLOCKACK: | |
2543 | { | |
2544 | // identify lock, master request | |
2545 | SimpleLock *lock = mds->locker->get_lock(m->get_lock_type(), | |
2546 | m->get_object_info()); | |
2547 | mdr->more()->slaves.insert(from); | |
2548 | dout(10) << "got remote wrlock on " << *lock << " on " << *lock->get_parent() << dendl; | |
11fdf7f2 TL |
2549 | auto it = mdr->locks.emplace_hint(mdr->locks.end(), |
2550 | lock, MutationImpl::LockOp::REMOTE_WRLOCK, from); | |
2551 | ceph_assert(it->is_remote_wrlock()); | |
2552 | ceph_assert(it->wrlock_target == from); | |
2553 | ||
7c673cae FG |
2554 | mdr->finish_locking(lock); |
2555 | ||
11fdf7f2 | 2556 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae | 2557 | mdr->more()->waiting_on_slave.erase(from); |
11fdf7f2 | 2558 | ceph_assert(mdr->more()->waiting_on_slave.empty()); |
7c673cae FG |
2559 | mdcache->dispatch_request(mdr); |
2560 | } | |
2561 | break; | |
2562 | ||
2563 | case MMDSSlaveRequest::OP_AUTHPINACK: | |
2564 | handle_slave_auth_pin_ack(mdr, m); | |
2565 | break; | |
2566 | ||
2567 | case MMDSSlaveRequest::OP_LINKPREPACK: | |
2568 | handle_slave_link_prep_ack(mdr, m); | |
2569 | break; | |
2570 | ||
2571 | case MMDSSlaveRequest::OP_RMDIRPREPACK: | |
2572 | handle_slave_rmdir_prep_ack(mdr, m); | |
2573 | break; | |
2574 | ||
2575 | case MMDSSlaveRequest::OP_RENAMEPREPACK: | |
2576 | handle_slave_rename_prep_ack(mdr, m); | |
2577 | break; | |
2578 | ||
2579 | case MMDSSlaveRequest::OP_RENAMENOTIFYACK: | |
2580 | handle_slave_rename_notify_ack(mdr, m); | |
2581 | break; | |
2582 | ||
2583 | default: | |
2584 | ceph_abort(); | |
2585 | } | |
7c673cae FG |
2586 | } |
2587 | ||
7c673cae FG |
2588 | void Server::dispatch_slave_request(MDRequestRef& mdr) |
2589 | { | |
2590 | dout(7) << "dispatch_slave_request " << *mdr << " " << *mdr->slave_request << dendl; | |
2591 | ||
2592 | if (mdr->aborted) { | |
2593 | dout(7) << " abort flag set, finishing" << dendl; | |
2594 | mdcache->request_finish(mdr); | |
2595 | return; | |
2596 | } | |
2597 | ||
2598 | if (logger) logger->inc(l_mdss_dispatch_slave_request); | |
2599 | ||
2600 | int op = mdr->slave_request->get_op(); | |
2601 | switch (op) { | |
2602 | case MMDSSlaveRequest::OP_XLOCK: | |
2603 | case MMDSSlaveRequest::OP_WRLOCK: | |
2604 | { | |
2605 | // identify object | |
2606 | SimpleLock *lock = mds->locker->get_lock(mdr->slave_request->get_lock_type(), | |
2607 | mdr->slave_request->get_object_info()); | |
2608 | ||
2609 | if (!lock) { | |
2610 | dout(10) << "don't have object, dropping" << dendl; | |
2611 | ceph_abort(); // can this happen, if we auth pinned properly. | |
2612 | } | |
2613 | if (op == MMDSSlaveRequest::OP_XLOCK && !lock->get_parent()->is_auth()) { | |
2614 | dout(10) << "not auth for remote xlock attempt, dropping on " | |
2615 | << *lock << " on " << *lock->get_parent() << dendl; | |
2616 | } else { | |
2617 | // use acquire_locks so that we get auth_pinning. | |
11fdf7f2 TL |
2618 | MutationImpl::LockOpVec lov; |
2619 | for (const auto& p : mdr->locks) { | |
2620 | if (p.is_xlock()) | |
2621 | lov.add_xlock(p.lock); | |
2622 | else if (p.is_wrlock()) | |
2623 | lov.add_wrlock(p.lock); | |
2624 | } | |
7c673cae FG |
2625 | |
2626 | int replycode = 0; | |
2627 | switch (op) { | |
2628 | case MMDSSlaveRequest::OP_XLOCK: | |
11fdf7f2 | 2629 | lov.add_xlock(lock); |
7c673cae FG |
2630 | replycode = MMDSSlaveRequest::OP_XLOCKACK; |
2631 | break; | |
2632 | case MMDSSlaveRequest::OP_WRLOCK: | |
11fdf7f2 | 2633 | lov.add_wrlock(lock); |
7c673cae FG |
2634 | replycode = MMDSSlaveRequest::OP_WRLOCKACK; |
2635 | break; | |
2636 | } | |
2637 | ||
11fdf7f2 | 2638 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
2639 | return; |
2640 | ||
2641 | // ack | |
11fdf7f2 | 2642 | auto r = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, replycode); |
7c673cae FG |
2643 | r->set_lock_type(lock->get_type()); |
2644 | lock->get_parent()->set_object_info(r->get_object_info()); | |
11fdf7f2 TL |
2645 | if (replycode == MMDSSlaveRequest::OP_XLOCKACK) |
2646 | lock->encode_locked_state(r->get_lock_data()); | |
7c673cae FG |
2647 | mds->send_message(r, mdr->slave_request->get_connection()); |
2648 | } | |
2649 | ||
2650 | // done. | |
91327a77 | 2651 | mdr->reset_slave_request(); |
7c673cae FG |
2652 | } |
2653 | break; | |
2654 | ||
2655 | case MMDSSlaveRequest::OP_UNXLOCK: | |
2656 | case MMDSSlaveRequest::OP_UNWRLOCK: | |
2657 | { | |
2658 | SimpleLock *lock = mds->locker->get_lock(mdr->slave_request->get_lock_type(), | |
2659 | mdr->slave_request->get_object_info()); | |
11fdf7f2 TL |
2660 | ceph_assert(lock); |
2661 | auto it = mdr->locks.find(lock); | |
2662 | ceph_assert(it != mdr->locks.end()); | |
7c673cae FG |
2663 | bool need_issue = false; |
2664 | switch (op) { | |
2665 | case MMDSSlaveRequest::OP_UNXLOCK: | |
11fdf7f2 | 2666 | mds->locker->xlock_finish(it, mdr.get(), &need_issue); |
7c673cae FG |
2667 | break; |
2668 | case MMDSSlaveRequest::OP_UNWRLOCK: | |
11fdf7f2 | 2669 | mds->locker->wrlock_finish(it, mdr.get(), &need_issue); |
7c673cae FG |
2670 | break; |
2671 | } | |
2672 | if (need_issue) | |
2673 | mds->locker->issue_caps(static_cast<CInode*>(lock->get_parent())); | |
2674 | ||
2675 | // done. no ack necessary. | |
91327a77 | 2676 | mdr->reset_slave_request(); |
7c673cae FG |
2677 | } |
2678 | break; | |
2679 | ||
2680 | case MMDSSlaveRequest::OP_DROPLOCKS: | |
2681 | mds->locker->drop_locks(mdr.get()); | |
91327a77 | 2682 | mdr->reset_slave_request(); |
7c673cae FG |
2683 | break; |
2684 | ||
2685 | case MMDSSlaveRequest::OP_AUTHPIN: | |
2686 | handle_slave_auth_pin(mdr); | |
2687 | break; | |
2688 | ||
2689 | case MMDSSlaveRequest::OP_LINKPREP: | |
2690 | case MMDSSlaveRequest::OP_UNLINKPREP: | |
2691 | handle_slave_link_prep(mdr); | |
2692 | break; | |
2693 | ||
2694 | case MMDSSlaveRequest::OP_RMDIRPREP: | |
2695 | handle_slave_rmdir_prep(mdr); | |
2696 | break; | |
2697 | ||
2698 | case MMDSSlaveRequest::OP_RENAMEPREP: | |
2699 | handle_slave_rename_prep(mdr); | |
2700 | break; | |
2701 | ||
2702 | case MMDSSlaveRequest::OP_FINISH: | |
2703 | // information about rename imported caps | |
2704 | if (mdr->slave_request->inode_export.length() > 0) | |
11fdf7f2 | 2705 | mdr->more()->inode_import = mdr->slave_request->inode_export; |
7c673cae FG |
2706 | // finish off request. |
2707 | mdcache->request_finish(mdr); | |
2708 | break; | |
2709 | ||
2710 | default: | |
2711 | ceph_abort(); | |
2712 | } | |
2713 | } | |
2714 | ||
7c673cae FG |
2715 | void Server::handle_slave_auth_pin(MDRequestRef& mdr) |
2716 | { | |
2717 | dout(10) << "handle_slave_auth_pin " << *mdr << dendl; | |
2718 | ||
2719 | // build list of objects | |
2720 | list<MDSCacheObject*> objects; | |
2721 | CInode *auth_pin_freeze = NULL; | |
2722 | bool fail = false, wouldblock = false, readonly = false; | |
2723 | ||
2724 | if (mdcache->is_readonly()) { | |
2725 | dout(10) << " read-only FS" << dendl; | |
2726 | readonly = true; | |
2727 | fail = true; | |
2728 | } | |
2729 | ||
2730 | if (!fail) { | |
11fdf7f2 TL |
2731 | for (const auto &oi : mdr->slave_request->get_authpins()) { |
2732 | MDSCacheObject *object = mdcache->get_object(oi); | |
7c673cae | 2733 | if (!object) { |
11fdf7f2 | 2734 | dout(10) << " don't have " << oi << dendl; |
7c673cae FG |
2735 | fail = true; |
2736 | break; | |
2737 | } | |
2738 | ||
2739 | objects.push_back(object); | |
11fdf7f2 | 2740 | if (oi == mdr->slave_request->get_authpin_freeze()) |
7c673cae FG |
2741 | auth_pin_freeze = static_cast<CInode*>(object); |
2742 | } | |
2743 | } | |
2744 | ||
2745 | // can we auth pin them? | |
2746 | if (!fail) { | |
2747 | for (list<MDSCacheObject*>::iterator p = objects.begin(); | |
2748 | p != objects.end(); | |
2749 | ++p) { | |
2750 | if (!(*p)->is_auth()) { | |
2751 | dout(10) << " not auth for " << **p << dendl; | |
2752 | fail = true; | |
2753 | break; | |
2754 | } | |
2755 | if (mdr->is_auth_pinned(*p)) | |
2756 | continue; | |
2757 | if (!mdr->can_auth_pin(*p)) { | |
2758 | if (mdr->slave_request->is_nonblock()) { | |
2759 | dout(10) << " can't auth_pin (freezing?) " << **p << " nonblocking" << dendl; | |
2760 | fail = true; | |
2761 | wouldblock = true; | |
2762 | break; | |
2763 | } | |
2764 | // wait | |
2765 | dout(10) << " waiting for authpinnable on " << **p << dendl; | |
2766 | (*p)->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); | |
2767 | mdr->drop_local_auth_pins(); | |
2768 | ||
224ce89b | 2769 | mds->locker->notify_freeze_waiter(*p); |
7c673cae FG |
2770 | return; |
2771 | } | |
2772 | } | |
2773 | } | |
2774 | ||
2775 | // auth pin! | |
2776 | if (fail) { | |
2777 | mdr->drop_local_auth_pins(); // just in case | |
2778 | } else { | |
2779 | /* freeze authpin wrong inode */ | |
2780 | if (mdr->has_more() && mdr->more()->is_freeze_authpin && | |
2781 | mdr->more()->rename_inode != auth_pin_freeze) | |
2782 | mdr->unfreeze_auth_pin(true); | |
2783 | ||
2784 | /* handle_slave_rename_prep() call freeze_inode() to wait for all other operations | |
2785 | * on the source inode to complete. This happens after all locks for the rename | |
2786 | * operation are acquired. But to acquire locks, we need auth pin locks' parent | |
2787 | * objects first. So there is an ABBA deadlock if someone auth pins the source inode | |
2788 | * after locks are acquired and before Server::handle_slave_rename_prep() is called. | |
2789 | * The solution is freeze the inode and prevent other MDRequests from getting new | |
2790 | * auth pins. | |
2791 | */ | |
2792 | if (auth_pin_freeze) { | |
2793 | dout(10) << " freezing auth pin on " << *auth_pin_freeze << dendl; | |
2794 | if (!mdr->freeze_auth_pin(auth_pin_freeze)) { | |
2795 | auth_pin_freeze->add_waiter(CInode::WAIT_FROZEN, new C_MDS_RetryRequest(mdcache, mdr)); | |
2796 | mds->mdlog->flush(); | |
2797 | return; | |
2798 | } | |
2799 | } | |
2800 | for (list<MDSCacheObject*>::iterator p = objects.begin(); | |
2801 | p != objects.end(); | |
2802 | ++p) { | |
2803 | dout(10) << "auth_pinning " << **p << dendl; | |
2804 | mdr->auth_pin(*p); | |
2805 | } | |
2806 | } | |
2807 | ||
2808 | // ack! | |
11fdf7f2 | 2809 | auto reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_AUTHPINACK); |
7c673cae FG |
2810 | |
2811 | // return list of my auth_pins (if any) | |
11fdf7f2 | 2812 | for (const auto &p : mdr->auth_pins) { |
7c673cae | 2813 | MDSCacheObjectInfo info; |
11fdf7f2 | 2814 | p->set_object_info(info); |
7c673cae | 2815 | reply->get_authpins().push_back(info); |
11fdf7f2 | 2816 | if (p == (MDSCacheObject*)auth_pin_freeze) |
7c673cae FG |
2817 | auth_pin_freeze->set_object_info(reply->get_authpin_freeze()); |
2818 | } | |
2819 | ||
2820 | if (wouldblock) | |
2821 | reply->mark_error_wouldblock(); | |
2822 | if (readonly) | |
2823 | reply->mark_error_rofs(); | |
2824 | ||
2825 | mds->send_message_mds(reply, mdr->slave_to_mds); | |
2826 | ||
2827 | // clean up this request | |
91327a77 | 2828 | mdr->reset_slave_request(); |
7c673cae FG |
2829 | return; |
2830 | } | |
2831 | ||
11fdf7f2 | 2832 | void Server::handle_slave_auth_pin_ack(MDRequestRef& mdr, const MMDSSlaveRequest::const_ref &ack) |
7c673cae FG |
2833 | { |
2834 | dout(10) << "handle_slave_auth_pin_ack on " << *mdr << " " << *ack << dendl; | |
2835 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
2836 | ||
2837 | // added auth pins? | |
2838 | set<MDSCacheObject*> pinned; | |
11fdf7f2 TL |
2839 | for (const auto &oi : ack->get_authpins()) { |
2840 | MDSCacheObject *object = mdcache->get_object(oi); | |
2841 | ceph_assert(object); // we pinned it | |
7c673cae FG |
2842 | dout(10) << " remote has pinned " << *object << dendl; |
2843 | if (!mdr->is_auth_pinned(object)) | |
2844 | mdr->remote_auth_pins[object] = from; | |
11fdf7f2 | 2845 | if (oi == ack->get_authpin_freeze()) |
7c673cae FG |
2846 | mdr->set_remote_frozen_auth_pin(static_cast<CInode *>(object)); |
2847 | pinned.insert(object); | |
2848 | } | |
2849 | ||
2850 | // removed frozen auth pin ? | |
2851 | if (mdr->more()->is_remote_frozen_authpin && | |
2852 | ack->get_authpin_freeze() == MDSCacheObjectInfo()) { | |
2853 | auto p = mdr->remote_auth_pins.find(mdr->more()->rename_inode); | |
11fdf7f2 | 2854 | ceph_assert(p != mdr->remote_auth_pins.end()); |
7c673cae FG |
2855 | if (p->second == from) { |
2856 | mdr->more()->is_remote_frozen_authpin = false; | |
2857 | } | |
2858 | } | |
2859 | ||
2860 | // removed auth pins? | |
11fdf7f2 | 2861 | auto p = mdr->remote_auth_pins.begin(); |
7c673cae FG |
2862 | while (p != mdr->remote_auth_pins.end()) { |
2863 | MDSCacheObject* object = p->first; | |
2864 | if (p->second == from && pinned.count(object) == 0) { | |
2865 | dout(10) << " remote has unpinned " << *object << dendl; | |
2866 | mdr->remote_auth_pins.erase(p++); | |
2867 | } else { | |
2868 | ++p; | |
2869 | } | |
2870 | } | |
2871 | ||
2872 | if (ack->is_error_rofs()) { | |
2873 | mdr->more()->slave_error = -EROFS; | |
2874 | mdr->aborted = true; | |
2875 | } else if (ack->is_error_wouldblock()) { | |
2876 | mdr->more()->slave_error = -EWOULDBLOCK; | |
2877 | mdr->aborted = true; | |
2878 | } | |
2879 | ||
2880 | // note slave | |
2881 | mdr->more()->slaves.insert(from); | |
2882 | ||
2883 | // clear from waiting list | |
11fdf7f2 | 2884 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae FG |
2885 | mdr->more()->waiting_on_slave.erase(from); |
2886 | ||
2887 | // go again? | |
2888 | if (mdr->more()->waiting_on_slave.empty()) | |
2889 | mdcache->dispatch_request(mdr); | |
2890 | else | |
2891 | dout(10) << "still waiting on slaves " << mdr->more()->waiting_on_slave << dendl; | |
2892 | } | |
2893 | ||
2894 | ||
2895 | // --------------------------------------- | |
2896 | // HELPERS | |
2897 | ||
2898 | ||
2899 | /** | |
2900 | * check whether we are permitted to complete a request | |
2901 | * | |
2902 | * Check whether we have permission to perform the operation specified | |
2903 | * by mask on the given inode, based on the capability in the mdr's | |
2904 | * session. | |
2905 | */ | |
2906 | bool Server::check_access(MDRequestRef& mdr, CInode *in, unsigned mask) | |
2907 | { | |
2908 | if (mdr->session) { | |
2909 | int r = mdr->session->check_access( | |
2910 | in, mask, | |
2911 | mdr->client_request->get_caller_uid(), | |
2912 | mdr->client_request->get_caller_gid(), | |
2913 | &mdr->client_request->get_caller_gid_list(), | |
2914 | mdr->client_request->head.args.setattr.uid, | |
2915 | mdr->client_request->head.args.setattr.gid); | |
2916 | if (r < 0) { | |
2917 | respond_to_request(mdr, r); | |
2918 | return false; | |
2919 | } | |
2920 | } | |
2921 | return true; | |
2922 | } | |
2923 | ||
2924 | /** | |
2925 | * check whether fragment has reached maximum size | |
2926 | * | |
2927 | */ | |
2928 | bool Server::check_fragment_space(MDRequestRef &mdr, CDir *in) | |
2929 | { | |
2930 | const auto size = in->get_frag_size(); | |
11fdf7f2 TL |
2931 | if (size >= g_conf()->mds_bal_fragment_size_max) { |
2932 | dout(10) << "fragment " << *in << " size exceeds " << g_conf()->mds_bal_fragment_size_max << " (ENOSPC)" << dendl; | |
7c673cae FG |
2933 | respond_to_request(mdr, -ENOSPC); |
2934 | return false; | |
2935 | } | |
2936 | ||
2937 | return true; | |
2938 | } | |
2939 | ||
2940 | ||
2941 | /** validate_dentry_dir | |
2942 | * | |
2943 | * verify that the dir exists and would own the dname. | |
2944 | * do not check if the dentry exists. | |
2945 | */ | |
11fdf7f2 | 2946 | CDir *Server::validate_dentry_dir(MDRequestRef& mdr, CInode *diri, std::string_view dname) |
7c673cae FG |
2947 | { |
2948 | // make sure parent is a dir? | |
2949 | if (!diri->is_dir()) { | |
2950 | dout(7) << "validate_dentry_dir: not a dir" << dendl; | |
2951 | respond_to_request(mdr, -ENOTDIR); | |
2952 | return NULL; | |
2953 | } | |
2954 | ||
2955 | // which dirfrag? | |
2956 | frag_t fg = diri->pick_dirfrag(dname); | |
2957 | CDir *dir = try_open_auth_dirfrag(diri, fg, mdr); | |
2958 | if (!dir) | |
2959 | return 0; | |
2960 | ||
2961 | // frozen? | |
2962 | if (dir->is_frozen()) { | |
2963 | dout(7) << "dir is frozen " << *dir << dendl; | |
2964 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); | |
2965 | return NULL; | |
2966 | } | |
2967 | ||
2968 | return dir; | |
2969 | } | |
2970 | ||
2971 | ||
2972 | /** prepare_null_dentry | |
2973 | * prepare a null (or existing) dentry in given dir. | |
2974 | * wait for any dn lock. | |
2975 | */ | |
11fdf7f2 | 2976 | CDentry* Server::prepare_null_dentry(MDRequestRef& mdr, CDir *dir, std::string_view dname, bool okexist) |
7c673cae FG |
2977 | { |
2978 | dout(10) << "prepare_null_dentry " << dname << " in " << *dir << dendl; | |
11fdf7f2 | 2979 | ceph_assert(dir->is_auth()); |
7c673cae FG |
2980 | |
2981 | client_t client = mdr->get_client(); | |
2982 | ||
2983 | // does it already exist? | |
2984 | CDentry *dn = dir->lookup(dname); | |
2985 | if (dn) { | |
2986 | /* | |
2987 | if (dn->lock.is_xlocked_by_other(mdr)) { | |
2988 | dout(10) << "waiting on xlocked dentry " << *dn << dendl; | |
2989 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryRequest(mdcache, mdr)); | |
2990 | return 0; | |
2991 | } | |
2992 | */ | |
2993 | if (!dn->get_linkage(client, mdr)->is_null()) { | |
2994 | // name already exists | |
2995 | dout(10) << "dentry " << dname << " exists in " << *dir << dendl; | |
2996 | if (!okexist) { | |
2997 | respond_to_request(mdr, -EEXIST); | |
2998 | return 0; | |
2999 | } | |
3000 | } else { | |
11fdf7f2 TL |
3001 | snapid_t next_snap = mdcache->get_global_snaprealm()->get_newest_seq() + 1; |
3002 | dn->first = std::max(dn->first, next_snap); | |
7c673cae | 3003 | } |
7c673cae FG |
3004 | return dn; |
3005 | } | |
3006 | ||
3007 | // make sure dir is complete | |
3008 | if (!dir->is_complete() && (!dir->has_bloom() || dir->is_in_bloom(dname))) { | |
3009 | dout(7) << " incomplete dir contents for " << *dir << ", fetching" << dendl; | |
3010 | dir->fetch(new C_MDS_RetryRequest(mdcache, mdr)); | |
3011 | return 0; | |
3012 | } | |
3013 | ||
3014 | // create | |
11fdf7f2 | 3015 | dn = dir->add_null_dentry(dname, mdcache->get_global_snaprealm()->get_newest_seq() + 1); |
7c673cae FG |
3016 | dn->mark_new(); |
3017 | dout(10) << "prepare_null_dentry added " << *dn << dendl; | |
3018 | return dn; | |
3019 | } | |
3020 | ||
3021 | CDentry* Server::prepare_stray_dentry(MDRequestRef& mdr, CInode *in) | |
3022 | { | |
3023 | CDentry *straydn = mdr->straydn; | |
3024 | if (straydn) { | |
3025 | string straydname; | |
3026 | in->name_stray_dentry(straydname); | |
3027 | if (straydn->get_name() == straydname) | |
3028 | return straydn; | |
3029 | ||
11fdf7f2 | 3030 | ceph_assert(!mdr->done_locking); |
7c673cae FG |
3031 | mdr->unpin(straydn); |
3032 | } | |
3033 | ||
3034 | CDir *straydir = mdcache->get_stray_dir(in); | |
3035 | ||
3036 | if (!mdr->client_request->is_replay() && | |
3037 | !check_fragment_space(mdr, straydir)) | |
3038 | return NULL; | |
3039 | ||
3040 | straydn = mdcache->get_or_create_stray_dentry(in); | |
3041 | mdr->straydn = straydn; | |
3042 | mdr->pin(straydn); | |
3043 | return straydn; | |
3044 | } | |
3045 | ||
3046 | /** prepare_new_inode | |
3047 | * | |
3048 | * create a new inode. set c/m/atime. hit dir pop. | |
3049 | */ | |
3050 | CInode* Server::prepare_new_inode(MDRequestRef& mdr, CDir *dir, inodeno_t useino, unsigned mode, | |
3051 | file_layout_t *layout) | |
3052 | { | |
3053 | CInode *in = new CInode(mdcache); | |
3054 | ||
3055 | // Server::prepare_force_open_sessions() can re-open session in closing | |
3056 | // state. In that corner case, session's prealloc_inos are being freed. | |
3057 | // To simplify the code, we disallow using/refilling session's prealloc_ino | |
3058 | // while session is opening. | |
3059 | bool allow_prealloc_inos = !mdr->session->is_opening(); | |
3060 | ||
3061 | // assign ino | |
3062 | if (allow_prealloc_inos && | |
3063 | mdr->session->info.prealloc_inos.size()) { | |
3064 | mdr->used_prealloc_ino = | |
3065 | in->inode.ino = mdr->session->take_ino(useino); // prealloc -> used | |
3066 | mds->sessionmap.mark_projected(mdr->session); | |
3067 | ||
3068 | dout(10) << "prepare_new_inode used_prealloc " << mdr->used_prealloc_ino | |
3069 | << " (" << mdr->session->info.prealloc_inos | |
3070 | << ", " << mdr->session->info.prealloc_inos.size() << " left)" | |
3071 | << dendl; | |
3072 | } else { | |
3073 | mdr->alloc_ino = | |
3074 | in->inode.ino = mds->inotable->project_alloc_id(); | |
3075 | dout(10) << "prepare_new_inode alloc " << mdr->alloc_ino << dendl; | |
3076 | } | |
3077 | ||
3078 | if (useino && useino != in->inode.ino) { | |
3079 | dout(0) << "WARNING: client specified " << useino << " and i allocated " << in->inode.ino << dendl; | |
3080 | mds->clog->error() << mdr->client_request->get_source() | |
3081 | << " specified ino " << useino | |
3082 | << " but mds." << mds->get_nodeid() << " allocated " << in->inode.ino; | |
3083 | //ceph_abort(); // just for now. | |
3084 | } | |
3085 | ||
3086 | if (allow_prealloc_inos && | |
11fdf7f2 TL |
3087 | mdr->session->get_num_projected_prealloc_inos() < g_conf()->mds_client_prealloc_inos / 2) { |
3088 | int need = g_conf()->mds_client_prealloc_inos - mdr->session->get_num_projected_prealloc_inos(); | |
7c673cae | 3089 | mds->inotable->project_alloc_ids(mdr->prealloc_inos, need); |
11fdf7f2 | 3090 | ceph_assert(mdr->prealloc_inos.size()); // or else fix projected increment semantics |
7c673cae FG |
3091 | mdr->session->pending_prealloc_inos.insert(mdr->prealloc_inos); |
3092 | mds->sessionmap.mark_projected(mdr->session); | |
3093 | dout(10) << "prepare_new_inode prealloc " << mdr->prealloc_inos << dendl; | |
3094 | } | |
3095 | ||
3096 | in->inode.version = 1; | |
3097 | in->inode.xattr_version = 1; | |
3098 | in->inode.nlink = 1; // FIXME | |
3099 | ||
3100 | in->inode.mode = mode; | |
3101 | ||
3102 | memset(&in->inode.dir_layout, 0, sizeof(in->inode.dir_layout)); | |
3103 | if (in->inode.is_dir()) { | |
11fdf7f2 | 3104 | in->inode.dir_layout.dl_dir_hash = g_conf()->mds_default_dir_hash; |
7c673cae FG |
3105 | } else if (layout) { |
3106 | in->inode.layout = *layout; | |
3107 | } else { | |
3108 | in->inode.layout = mdcache->default_file_layout; | |
3109 | } | |
3110 | ||
3111 | in->inode.truncate_size = -1ull; // not truncated, yet! | |
3112 | in->inode.truncate_seq = 1; /* starting with 1, 0 is kept for no-truncation logic */ | |
3113 | ||
3114 | CInode *diri = dir->get_inode(); | |
3115 | ||
3116 | dout(10) << oct << " dir mode 0" << diri->inode.mode << " new mode 0" << mode << dec << dendl; | |
3117 | ||
3118 | if (diri->inode.mode & S_ISGID) { | |
3119 | dout(10) << " dir is sticky" << dendl; | |
3120 | in->inode.gid = diri->inode.gid; | |
3121 | if (S_ISDIR(mode)) { | |
3122 | dout(10) << " new dir also sticky" << dendl; | |
3123 | in->inode.mode |= S_ISGID; | |
3124 | } | |
3125 | } else | |
3126 | in->inode.gid = mdr->client_request->get_caller_gid(); | |
3127 | ||
3128 | in->inode.uid = mdr->client_request->get_caller_uid(); | |
3129 | ||
3130 | in->inode.btime = in->inode.ctime = in->inode.mtime = in->inode.atime = | |
3131 | mdr->get_op_stamp(); | |
3132 | ||
3133 | in->inode.change_attr = 0; | |
3134 | ||
11fdf7f2 | 3135 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae | 3136 | if (req->get_data().length()) { |
11fdf7f2 | 3137 | auto p = req->get_data().cbegin(); |
7c673cae FG |
3138 | |
3139 | // xattrs on new inode? | |
94b18763 | 3140 | CInode::mempool_xattr_map xattrs; |
11fdf7f2 | 3141 | decode(xattrs, p); |
94b18763 FG |
3142 | for (const auto &p : xattrs) { |
3143 | dout(10) << "prepare_new_inode setting xattr " << p.first << dendl; | |
3144 | auto em = in->xattrs.emplace(std::piecewise_construct, std::forward_as_tuple(p.first), std::forward_as_tuple(p.second)); | |
3145 | if (!em.second) | |
3146 | em.first->second = p.second; | |
7c673cae FG |
3147 | } |
3148 | } | |
3149 | ||
3150 | if (!mds->mdsmap->get_inline_data_enabled() || | |
11fdf7f2 | 3151 | !mdr->session->get_connection()->has_feature(CEPH_FEATURE_MDS_INLINE_DATA)) |
7c673cae FG |
3152 | in->inode.inline_data.version = CEPH_INLINE_NONE; |
3153 | ||
3154 | mdcache->add_inode(in); // add | |
3155 | dout(10) << "prepare_new_inode " << *in << dendl; | |
3156 | return in; | |
3157 | } | |
3158 | ||
3159 | void Server::journal_allocated_inos(MDRequestRef& mdr, EMetaBlob *blob) | |
3160 | { | |
3161 | dout(20) << "journal_allocated_inos sessionmapv " << mds->sessionmap.get_projected() | |
3162 | << " inotablev " << mds->inotable->get_projected_version() | |
3163 | << dendl; | |
3164 | blob->set_ino_alloc(mdr->alloc_ino, | |
3165 | mdr->used_prealloc_ino, | |
3166 | mdr->prealloc_inos, | |
3167 | mdr->client_request->get_source(), | |
3168 | mds->sessionmap.get_projected(), | |
3169 | mds->inotable->get_projected_version()); | |
3170 | } | |
3171 | ||
3172 | void Server::apply_allocated_inos(MDRequestRef& mdr, Session *session) | |
3173 | { | |
3174 | dout(10) << "apply_allocated_inos " << mdr->alloc_ino | |
3175 | << " / " << mdr->prealloc_inos | |
3176 | << " / " << mdr->used_prealloc_ino << dendl; | |
3177 | ||
3178 | if (mdr->alloc_ino) { | |
3179 | mds->inotable->apply_alloc_id(mdr->alloc_ino); | |
3180 | } | |
3181 | if (mdr->prealloc_inos.size()) { | |
11fdf7f2 | 3182 | ceph_assert(session); |
7c673cae FG |
3183 | session->pending_prealloc_inos.subtract(mdr->prealloc_inos); |
3184 | session->info.prealloc_inos.insert(mdr->prealloc_inos); | |
3185 | mds->sessionmap.mark_dirty(session); | |
3186 | mds->inotable->apply_alloc_ids(mdr->prealloc_inos); | |
3187 | } | |
3188 | if (mdr->used_prealloc_ino) { | |
11fdf7f2 | 3189 | ceph_assert(session); |
7c673cae FG |
3190 | session->info.used_inos.erase(mdr->used_prealloc_ino); |
3191 | mds->sessionmap.mark_dirty(session); | |
3192 | } | |
3193 | } | |
3194 | ||
3195 | class C_MDS_TryFindInode : public ServerContext { | |
3196 | MDRequestRef mdr; | |
3197 | public: | |
3198 | C_MDS_TryFindInode(Server *s, MDRequestRef& r) : ServerContext(s), mdr(r) {} | |
3199 | void finish(int r) override { | |
3200 | if (r == -ESTALE) // :( find_ino_peers failed | |
3201 | server->respond_to_request(mdr, r); | |
3202 | else | |
3203 | server->dispatch_client_request(mdr); | |
3204 | } | |
3205 | }; | |
3206 | ||
11fdf7f2 TL |
3207 | class CF_MDS_MDRContextFactory : public MDSContextFactory { |
3208 | public: | |
3209 | CF_MDS_MDRContextFactory(MDCache *cache, MDRequestRef &mdr) : cache(cache), mdr(mdr) {} | |
3210 | MDSContext *build() { | |
3211 | return new C_MDS_RetryRequest(cache, mdr); | |
3212 | } | |
3213 | private: | |
3214 | MDCache *cache; | |
3215 | MDRequestRef mdr; | |
3216 | }; | |
3217 | ||
7c673cae FG |
3218 | CDir *Server::traverse_to_auth_dir(MDRequestRef& mdr, vector<CDentry*> &trace, filepath refpath) |
3219 | { | |
3220 | // figure parent dir vs dname | |
3221 | if (refpath.depth() == 0) { | |
3222 | dout(7) << "can't do that to root" << dendl; | |
3223 | respond_to_request(mdr, -EINVAL); | |
3224 | return 0; | |
3225 | } | |
3226 | string dname = refpath.last_dentry(); | |
3227 | refpath.pop_dentry(); | |
3228 | ||
3229 | dout(10) << "traverse_to_auth_dir dirpath " << refpath << " dname " << dname << dendl; | |
3230 | ||
3231 | // traverse to parent dir | |
3232 | CInode *diri; | |
11fdf7f2 TL |
3233 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
3234 | int r = mdcache->path_traverse(mdr, cf, refpath, &trace, &diri, MDS_TRAVERSE_FORWARD); | |
7c673cae FG |
3235 | if (r > 0) return 0; // delayed |
3236 | if (r < 0) { | |
3237 | if (r == -ESTALE) { | |
3238 | dout(10) << "FAIL on ESTALE but attempting recovery" << dendl; | |
3239 | mdcache->find_ino_peers(refpath.get_ino(), new C_MDS_TryFindInode(this, mdr)); | |
3240 | return 0; | |
3241 | } | |
3242 | respond_to_request(mdr, r); | |
3243 | return 0; | |
3244 | } | |
3245 | ||
3246 | // is it an auth dir? | |
3247 | CDir *dir = validate_dentry_dir(mdr, diri, dname); | |
3248 | if (!dir) | |
3249 | return 0; // forwarded or waiting for freeze | |
3250 | ||
3251 | dout(10) << "traverse_to_auth_dir " << *dir << dendl; | |
3252 | return dir; | |
3253 | } | |
3254 | ||
3255 | /* If this returns null, the request has been handled | |
3256 | * as appropriate: forwarded on, or the client's been replied to */ | |
3257 | CInode* Server::rdlock_path_pin_ref(MDRequestRef& mdr, int n, | |
11fdf7f2 | 3258 | MutationImpl::LockOpVec& lov, |
7c673cae FG |
3259 | bool want_auth, |
3260 | bool no_want_auth, /* for readdir, who doesn't want auth _even_if_ it's | |
3261 | a snapped dir */ | |
3262 | file_layout_t **layout, | |
3263 | bool no_lookup) // true if we cannot return a null dentry lease | |
3264 | { | |
3265 | const filepath& refpath = n ? mdr->get_filepath2() : mdr->get_filepath(); | |
3266 | dout(10) << "rdlock_path_pin_ref " << *mdr << " " << refpath << dendl; | |
3267 | ||
3268 | if (mdr->done_locking) | |
3269 | return mdr->in[n]; | |
3270 | ||
3271 | // traverse | |
11fdf7f2 TL |
3272 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
3273 | int r = mdcache->path_traverse(mdr, cf, refpath, &mdr->dn[n], &mdr->in[n], MDS_TRAVERSE_FORWARD); | |
7c673cae FG |
3274 | if (r > 0) |
3275 | return NULL; // delayed | |
3276 | if (r < 0) { // error | |
91327a77 AA |
3277 | if (r == -ENOENT && n == 0 && !mdr->dn[n].empty()) { |
3278 | if (!no_lookup) { | |
3279 | mdr->tracedn = mdr->dn[n].back(); | |
3280 | } | |
7c673cae FG |
3281 | respond_to_request(mdr, r); |
3282 | } else if (r == -ESTALE) { | |
3283 | dout(10) << "FAIL on ESTALE but attempting recovery" << dendl; | |
11fdf7f2 | 3284 | MDSContext *c = new C_MDS_TryFindInode(this, mdr); |
7c673cae FG |
3285 | mdcache->find_ino_peers(refpath.get_ino(), c); |
3286 | } else { | |
3287 | dout(10) << "FAIL on error " << r << dendl; | |
3288 | respond_to_request(mdr, r); | |
3289 | } | |
3290 | return 0; | |
3291 | } | |
3292 | CInode *ref = mdr->in[n]; | |
3293 | dout(10) << "ref is " << *ref << dendl; | |
3294 | ||
3295 | // fw to inode auth? | |
3296 | if (mdr->snapid != CEPH_NOSNAP && !no_want_auth) | |
3297 | want_auth = true; | |
3298 | ||
3299 | if (want_auth) { | |
3300 | if (ref->is_ambiguous_auth()) { | |
3301 | dout(10) << "waiting for single auth on " << *ref << dendl; | |
3302 | ref->add_waiter(CInode::WAIT_SINGLEAUTH, new C_MDS_RetryRequest(mdcache, mdr)); | |
3303 | return 0; | |
3304 | } | |
3305 | if (!ref->is_auth()) { | |
3306 | dout(10) << "fw to auth for " << *ref << dendl; | |
3307 | mdcache->request_forward(mdr, ref->authority().first); | |
3308 | return 0; | |
3309 | } | |
3310 | ||
3311 | // auth_pin? | |
3312 | // do NOT proceed if freezing, as cap release may defer in that case, and | |
3313 | // we could deadlock when we try to lock @ref. | |
3314 | // if we're already auth_pinned, continue; the release has already been processed. | |
3315 | if (ref->is_frozen() || ref->is_frozen_auth_pin() || | |
3316 | (ref->is_freezing() && !mdr->is_auth_pinned(ref))) { | |
3317 | dout(7) << "waiting for !frozen/authpinnable on " << *ref << dendl; | |
3318 | ref->add_waiter(CInode::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); | |
3319 | /* If we have any auth pins, this will deadlock. | |
3320 | * But the only way to get here if we've already got auth pins | |
3321 | * is because we're on an inode with snapshots that got updated | |
3322 | * between dispatches of this request. So we're going to drop | |
3323 | * our locks and our auth pins and reacquire them later. | |
3324 | * | |
3325 | * This is safe since we're only in this function when working on | |
3326 | * a single MDS request; otherwise we'd be in | |
3327 | * rdlock_path_xlock_dentry. | |
3328 | */ | |
3329 | mds->locker->drop_locks(mdr.get(), NULL); | |
3330 | mdr->drop_local_auth_pins(); | |
224ce89b WB |
3331 | if (!mdr->remote_auth_pins.empty()) |
3332 | mds->locker->notify_freeze_waiter(ref); | |
7c673cae FG |
3333 | return 0; |
3334 | } | |
3335 | ||
3336 | mdr->auth_pin(ref); | |
3337 | } | |
3338 | ||
3339 | for (int i=0; i<(int)mdr->dn[n].size(); i++) | |
11fdf7f2 | 3340 | lov.add_rdlock(&mdr->dn[n][i]->lock); |
7c673cae | 3341 | if (layout) |
11fdf7f2 | 3342 | mds->locker->include_snap_rdlocks_wlayout(ref, lov, layout); |
7c673cae | 3343 | else |
11fdf7f2 | 3344 | mds->locker->include_snap_rdlocks(ref, lov); |
7c673cae FG |
3345 | |
3346 | // set and pin ref | |
3347 | mdr->pin(ref); | |
3348 | return ref; | |
3349 | } | |
3350 | ||
3351 | ||
3352 | /** rdlock_path_xlock_dentry | |
3353 | * traverse path to the directory that could/would contain dentry. | |
3354 | * make sure i am auth for that dentry, forward as necessary. | |
3355 | * create null dentry in place (or use existing if okexist). | |
3356 | * get rdlocks on traversed dentries, xlock on new dentry. | |
3357 | */ | |
3358 | CDentry* Server::rdlock_path_xlock_dentry(MDRequestRef& mdr, int n, | |
11fdf7f2 | 3359 | MutationImpl::LockOpVec& lov, |
7c673cae FG |
3360 | bool okexist, bool mustexist, bool alwaysxlock, |
3361 | file_layout_t **layout) | |
3362 | { | |
3363 | const filepath& refpath = n ? mdr->get_filepath2() : mdr->get_filepath(); | |
3364 | ||
3365 | dout(10) << "rdlock_path_xlock_dentry " << *mdr << " " << refpath << dendl; | |
3366 | ||
3367 | client_t client = mdr->get_client(); | |
3368 | ||
3369 | if (mdr->done_locking) | |
3370 | return mdr->dn[n].back(); | |
3371 | ||
3372 | CDir *dir = traverse_to_auth_dir(mdr, mdr->dn[n], refpath); | |
3373 | if (!dir) return 0; | |
7c673cae | 3374 | |
7c673cae FG |
3375 | CInode *diri = dir->get_inode(); |
3376 | if (!mdr->reqid.name.is_mds()) { | |
3377 | if (diri->is_system() && !diri->is_root()) { | |
3378 | respond_to_request(mdr, -EROFS); | |
3379 | return 0; | |
3380 | } | |
3381 | } | |
3382 | if (!diri->is_base() && diri->get_projected_parent_dir()->inode->is_stray()) { | |
3383 | respond_to_request(mdr, -ENOENT); | |
3384 | return 0; | |
3385 | } | |
3386 | ||
3387 | // make a null dentry? | |
11fdf7f2 | 3388 | std::string_view dname = refpath.last_dentry(); |
7c673cae FG |
3389 | CDentry *dn; |
3390 | if (mustexist) { | |
3391 | dn = dir->lookup(dname); | |
3392 | ||
3393 | // make sure dir is complete | |
3394 | if (!dn && !dir->is_complete() && | |
3395 | (!dir->has_bloom() || dir->is_in_bloom(dname))) { | |
3396 | dout(7) << " incomplete dir contents for " << *dir << ", fetching" << dendl; | |
3397 | dir->fetch(new C_MDS_RetryRequest(mdcache, mdr)); | |
3398 | return 0; | |
3399 | } | |
3400 | ||
3401 | // readable? | |
3402 | if (dn && !dn->lock.can_read(client) && dn->lock.get_xlock_by() != mdr) { | |
3403 | dout(10) << "waiting on xlocked dentry " << *dn << dendl; | |
3404 | dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryRequest(mdcache, mdr)); | |
3405 | return 0; | |
3406 | } | |
3407 | ||
3408 | // exists? | |
3409 | if (!dn || dn->get_linkage(client, mdr)->is_null()) { | |
3410 | dout(7) << "dentry " << dname << " dne in " << *dir << dendl; | |
3411 | respond_to_request(mdr, -ENOENT); | |
3412 | return 0; | |
3413 | } | |
3414 | } else { | |
3415 | dn = prepare_null_dentry(mdr, dir, dname, okexist); | |
3416 | if (!dn) | |
3417 | return 0; | |
3418 | } | |
3419 | ||
3420 | mdr->dn[n].push_back(dn); | |
3421 | CDentry::linkage_t *dnl = dn->get_linkage(client, mdr); | |
3422 | mdr->in[n] = dnl->get_inode(); | |
3423 | ||
3424 | // -- lock -- | |
3425 | // NOTE: rename takes the same set of locks for srcdn | |
3426 | for (int i=0; i<(int)mdr->dn[n].size(); i++) | |
11fdf7f2 | 3427 | lov.add_rdlock(&mdr->dn[n][i]->lock); |
7c673cae | 3428 | if (alwaysxlock || dnl->is_null()) |
11fdf7f2 | 3429 | lov.add_xlock(&dn->lock); // new dn, xlock |
7c673cae | 3430 | else |
11fdf7f2 TL |
3431 | lov.add_rdlock(&dn->lock); // existing dn, rdlock |
3432 | lov.add_wrlock(&dn->get_dir()->inode->filelock); // also, wrlock on dir mtime | |
3433 | lov.add_wrlock(&dn->get_dir()->inode->nestlock); // also, wrlock on dir mtime | |
7c673cae | 3434 | if (layout) |
11fdf7f2 | 3435 | mds->locker->include_snap_rdlocks_wlayout(dn->get_dir()->inode, lov, layout); |
7c673cae | 3436 | else |
11fdf7f2 | 3437 | mds->locker->include_snap_rdlocks(dn->get_dir()->inode, lov); |
7c673cae FG |
3438 | |
3439 | return dn; | |
3440 | } | |
3441 | ||
3442 | ||
3443 | ||
3444 | ||
3445 | ||
3446 | /** | |
3447 | * try_open_auth_dirfrag -- open dirfrag, or forward to dirfrag auth | |
3448 | * | |
3449 | * @param diri base inode | |
3450 | * @param fg the exact frag we want | |
3451 | * @param mdr request | |
3452 | * @returns the pointer, or NULL if it had to be delayed (but mdr is taken care of) | |
3453 | */ | |
3454 | CDir* Server::try_open_auth_dirfrag(CInode *diri, frag_t fg, MDRequestRef& mdr) | |
3455 | { | |
3456 | CDir *dir = diri->get_dirfrag(fg); | |
3457 | ||
3458 | // not open and inode not mine? | |
3459 | if (!dir && !diri->is_auth()) { | |
3460 | mds_rank_t inauth = diri->authority().first; | |
3461 | dout(7) << "try_open_auth_dirfrag: not open, not inode auth, fw to mds." << inauth << dendl; | |
3462 | mdcache->request_forward(mdr, inauth); | |
3463 | return 0; | |
3464 | } | |
3465 | ||
3466 | // not open and inode frozen? | |
3467 | if (!dir && diri->is_frozen()) { | |
3468 | dout(10) << "try_open_auth_dirfrag: dir inode is frozen, waiting " << *diri << dendl; | |
11fdf7f2 | 3469 | ceph_assert(diri->get_parent_dir()); |
224ce89b | 3470 | diri->add_waiter(CInode::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); |
7c673cae FG |
3471 | return 0; |
3472 | } | |
3473 | ||
3474 | // invent? | |
3475 | if (!dir) | |
3476 | dir = diri->get_or_open_dirfrag(mdcache, fg); | |
3477 | ||
3478 | // am i auth for the dirfrag? | |
3479 | if (!dir->is_auth()) { | |
3480 | mds_rank_t auth = dir->authority().first; | |
3481 | dout(7) << "try_open_auth_dirfrag: not auth for " << *dir | |
3482 | << ", fw to mds." << auth << dendl; | |
3483 | mdcache->request_forward(mdr, auth); | |
3484 | return 0; | |
3485 | } | |
3486 | ||
3487 | return dir; | |
3488 | } | |
3489 | ||
3490 | ||
3491 | // =============================================================================== | |
3492 | // STAT | |
3493 | ||
3494 | void Server::handle_client_getattr(MDRequestRef& mdr, bool is_lookup) | |
3495 | { | |
11fdf7f2 | 3496 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
3497 | |
3498 | if (req->get_filepath().depth() == 0 && is_lookup) { | |
3499 | // refpath can't be empty for lookup but it can for | |
3500 | // getattr (we do getattr with empty refpath for mount of '/') | |
3501 | respond_to_request(mdr, -EINVAL); | |
3502 | return; | |
3503 | } | |
3504 | ||
28e407b8 AA |
3505 | bool want_auth = false; |
3506 | int mask = req->head.args.getattr.mask; | |
3507 | if (mask & CEPH_STAT_RSTAT) | |
3508 | want_auth = true; // set want_auth for CEPH_STAT_RSTAT mask | |
3509 | ||
11fdf7f2 TL |
3510 | MutationImpl::LockOpVec lov; |
3511 | CInode *ref = rdlock_path_pin_ref(mdr, 0, lov, want_auth, false, NULL, | |
28e407b8 | 3512 | !is_lookup); |
7c673cae FG |
3513 | if (!ref) return; |
3514 | ||
3515 | /* | |
3516 | * if client currently holds the EXCL cap on a field, do not rdlock | |
3517 | * it; client's stat() will result in valid info if _either_ EXCL | |
3518 | * cap is held or MDS rdlocks and reads the value here. | |
3519 | * | |
3520 | * handling this case here is easier than weakening rdlock | |
3521 | * semantics... that would cause problems elsewhere. | |
3522 | */ | |
3523 | client_t client = mdr->get_client(); | |
3524 | int issued = 0; | |
3525 | Capability *cap = ref->get_client_cap(client); | |
3526 | if (cap && (mdr->snapid == CEPH_NOSNAP || | |
3527 | mdr->snapid <= cap->client_follows)) | |
3528 | issued = cap->issued(); | |
3529 | ||
94b18763 | 3530 | if ((mask & CEPH_CAP_LINK_SHARED) && !(issued & CEPH_CAP_LINK_EXCL)) |
11fdf7f2 | 3531 | lov.add_rdlock(&ref->linklock); |
94b18763 | 3532 | if ((mask & CEPH_CAP_AUTH_SHARED) && !(issued & CEPH_CAP_AUTH_EXCL)) |
11fdf7f2 | 3533 | lov.add_rdlock(&ref->authlock); |
94b18763 | 3534 | if ((mask & CEPH_CAP_XATTR_SHARED) && !(issued & CEPH_CAP_XATTR_EXCL)) |
11fdf7f2 | 3535 | lov.add_rdlock(&ref->xattrlock); |
94b18763 FG |
3536 | if ((mask & CEPH_CAP_FILE_SHARED) && !(issued & CEPH_CAP_FILE_EXCL)) { |
3537 | // Don't wait on unstable filelock if client is allowed to read file size. | |
3538 | // This can reduce the response time of getattr in the case that multiple | |
3539 | // clients do stat(2) and there are writers. | |
3540 | // The downside of this optimization is that mds may not issue Fs caps along | |
3541 | // with getattr reply. Client may need to send more getattr requests. | |
11fdf7f2 TL |
3542 | if (mdr->is_rdlocked(&ref->filelock)) { |
3543 | lov.add_rdlock(&ref->filelock); | |
94b18763 FG |
3544 | } else if (ref->filelock.is_stable() || |
3545 | ref->filelock.get_num_wrlocks() > 0 || | |
3546 | !ref->filelock.can_read(mdr->get_client())) { | |
11fdf7f2 | 3547 | lov.add_rdlock(&ref->filelock); |
94b18763 FG |
3548 | mdr->done_locking = false; |
3549 | } | |
3550 | } | |
7c673cae | 3551 | |
11fdf7f2 | 3552 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
3553 | return; |
3554 | ||
3555 | if (!check_access(mdr, ref, MAY_READ)) | |
3556 | return; | |
3557 | ||
28e407b8 AA |
3558 | utime_t now = ceph_clock_now(); |
3559 | mdr->set_mds_stamp(now); | |
3560 | ||
7c673cae FG |
3561 | // note which caps are requested, so we return at least a snapshot |
3562 | // value for them. (currently this matters for xattrs and inline data) | |
3563 | mdr->getattr_caps = mask; | |
3564 | ||
11fdf7f2 | 3565 | mds->balancer->hit_inode(ref, META_POP_IRD, req->get_source().num()); |
7c673cae FG |
3566 | |
3567 | // reply | |
3568 | dout(10) << "reply to stat on " << *req << dendl; | |
3569 | mdr->tracei = ref; | |
3570 | if (is_lookup) | |
3571 | mdr->tracedn = mdr->dn[0].back(); | |
3572 | respond_to_request(mdr, 0); | |
3573 | } | |
3574 | ||
3575 | struct C_MDS_LookupIno2 : public ServerContext { | |
3576 | MDRequestRef mdr; | |
3577 | C_MDS_LookupIno2(Server *s, MDRequestRef& r) : ServerContext(s), mdr(r) {} | |
3578 | void finish(int r) override { | |
3579 | server->_lookup_ino_2(mdr, r); | |
3580 | } | |
3581 | }; | |
3582 | ||
7c673cae FG |
3583 | /* |
3584 | * filepath: ino | |
3585 | */ | |
3586 | void Server::handle_client_lookup_ino(MDRequestRef& mdr, | |
3587 | bool want_parent, bool want_dentry) | |
3588 | { | |
11fdf7f2 TL |
3589 | const MClientRequest::const_ref &req = mdr->client_request; |
3590 | ||
3591 | if ((uint64_t)req->head.args.lookupino.snapid > 0) | |
3592 | return _lookup_snap_ino(mdr); | |
7c673cae FG |
3593 | |
3594 | inodeno_t ino = req->get_filepath().get_ino(); | |
3595 | CInode *in = mdcache->get_inode(ino); | |
3596 | if (in && in->state_test(CInode::STATE_PURGING)) { | |
3597 | respond_to_request(mdr, -ESTALE); | |
3598 | return; | |
3599 | } | |
3600 | if (!in) { | |
3601 | mdcache->open_ino(ino, (int64_t)-1, new C_MDS_LookupIno2(this, mdr), false); | |
3602 | return; | |
3603 | } | |
3604 | ||
11fdf7f2 | 3605 | if (mdr && in->snaprealm && !in->snaprealm->have_past_parents_open() && |
7c673cae FG |
3606 | !in->snaprealm->open_parents(new C_MDS_RetryRequest(mdcache, mdr))) { |
3607 | return; | |
3608 | } | |
3609 | ||
3610 | // check for nothing (not read or write); this still applies the | |
3611 | // path check. | |
3612 | if (!check_access(mdr, in, 0)) | |
3613 | return; | |
3614 | ||
3615 | CDentry *dn = in->get_projected_parent_dn(); | |
3616 | CInode *diri = dn ? dn->get_dir()->inode : NULL; | |
3617 | ||
11fdf7f2 | 3618 | MutationImpl::LockOpVec lov; |
7c673cae FG |
3619 | if (dn && (want_parent || want_dentry)) { |
3620 | mdr->pin(dn); | |
11fdf7f2 | 3621 | lov.add_rdlock(&dn->lock); |
7c673cae FG |
3622 | } |
3623 | ||
11fdf7f2 | 3624 | unsigned mask = req->head.args.lookupino.mask; |
7c673cae FG |
3625 | if (mask) { |
3626 | Capability *cap = in->get_client_cap(mdr->get_client()); | |
3627 | int issued = 0; | |
3628 | if (cap && (mdr->snapid == CEPH_NOSNAP || mdr->snapid <= cap->client_follows)) | |
3629 | issued = cap->issued(); | |
3630 | // permission bits, ACL/security xattrs | |
3631 | if ((mask & CEPH_CAP_AUTH_SHARED) && (issued & CEPH_CAP_AUTH_EXCL) == 0) | |
11fdf7f2 | 3632 | lov.add_rdlock(&in->authlock); |
7c673cae | 3633 | if ((mask & CEPH_CAP_XATTR_SHARED) && (issued & CEPH_CAP_XATTR_EXCL) == 0) |
11fdf7f2 | 3634 | lov.add_rdlock(&in->xattrlock); |
7c673cae FG |
3635 | |
3636 | mdr->getattr_caps = mask; | |
3637 | } | |
3638 | ||
11fdf7f2 TL |
3639 | if (!lov.empty()) { |
3640 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
3641 | return; |
3642 | ||
d2e6a577 FG |
3643 | if (diri != NULL) { |
3644 | // need read access to directory inode | |
3645 | if (!check_access(mdr, diri, MAY_READ)) | |
3646 | return; | |
3647 | } | |
7c673cae FG |
3648 | } |
3649 | ||
3650 | if (want_parent) { | |
3651 | if (in->is_base()) { | |
3652 | respond_to_request(mdr, -EINVAL); | |
3653 | return; | |
3654 | } | |
3655 | if (!diri || diri->is_stray()) { | |
3656 | respond_to_request(mdr, -ESTALE); | |
3657 | return; | |
3658 | } | |
3659 | dout(10) << "reply to lookup_parent " << *in << dendl; | |
3660 | mdr->tracei = diri; | |
3661 | respond_to_request(mdr, 0); | |
3662 | } else { | |
3663 | if (want_dentry) { | |
3664 | inodeno_t dirino = req->get_filepath2().get_ino(); | |
3665 | if (!diri || (dirino != inodeno_t() && diri->ino() != dirino)) { | |
3666 | respond_to_request(mdr, -ENOENT); | |
3667 | return; | |
3668 | } | |
3669 | dout(10) << "reply to lookup_name " << *in << dendl; | |
3670 | } else | |
3671 | dout(10) << "reply to lookup_ino " << *in << dendl; | |
3672 | ||
3673 | mdr->tracei = in; | |
3674 | if (want_dentry) | |
3675 | mdr->tracedn = dn; | |
3676 | respond_to_request(mdr, 0); | |
3677 | } | |
3678 | } | |
3679 | ||
11fdf7f2 TL |
3680 | void Server::_lookup_snap_ino(MDRequestRef& mdr) |
3681 | { | |
3682 | const MClientRequest::const_ref &req = mdr->client_request; | |
3683 | ||
3684 | vinodeno_t vino; | |
3685 | vino.ino = req->get_filepath().get_ino(); | |
3686 | vino.snapid = (__u64)req->head.args.lookupino.snapid; | |
3687 | inodeno_t parent_ino = (__u64)req->head.args.lookupino.parent; | |
3688 | __u32 hash = req->head.args.lookupino.hash; | |
3689 | ||
3690 | dout(7) << "lookup_snap_ino " << vino << " parent " << parent_ino << " hash " << hash << dendl; | |
3691 | ||
3692 | CInode *in = mdcache->lookup_snap_inode(vino); | |
3693 | if (!in) { | |
3694 | in = mdcache->get_inode(vino.ino); | |
3695 | if (in) { | |
3696 | if (in->state_test(CInode::STATE_PURGING) || | |
3697 | !in->has_snap_data(vino.snapid)) { | |
3698 | if (in->is_dir() || !parent_ino) { | |
3699 | respond_to_request(mdr, -ESTALE); | |
3700 | return; | |
3701 | } | |
3702 | in = NULL; | |
3703 | } | |
3704 | } | |
3705 | } | |
3706 | ||
3707 | if (in) { | |
3708 | dout(10) << "reply to lookup_snap_ino " << *in << dendl; | |
3709 | mdr->snapid = vino.snapid; | |
3710 | mdr->tracei = in; | |
3711 | respond_to_request(mdr, 0); | |
3712 | return; | |
3713 | } | |
3714 | ||
3715 | CInode *diri = NULL; | |
3716 | if (parent_ino) { | |
3717 | diri = mdcache->get_inode(parent_ino); | |
3718 | if (!diri) { | |
3719 | mdcache->open_ino(parent_ino, mds->mdsmap->get_metadata_pool(), new C_MDS_LookupIno2(this, mdr)); | |
3720 | return; | |
3721 | } | |
3722 | ||
3723 | if (!diri->is_dir()) { | |
3724 | respond_to_request(mdr, -EINVAL); | |
3725 | return; | |
3726 | } | |
3727 | ||
3728 | MutationImpl::LockOpVec lov; | |
3729 | lov.add_rdlock(&diri->dirfragtreelock); | |
3730 | if (!mds->locker->acquire_locks(mdr, lov)) | |
3731 | return; | |
3732 | ||
3733 | frag_t frag = diri->dirfragtree[hash]; | |
3734 | CDir *dir = try_open_auth_dirfrag(diri, frag, mdr); | |
3735 | if (!dir) | |
3736 | return; | |
3737 | ||
3738 | if (!dir->is_complete()) { | |
3739 | if (dir->is_frozen()) { | |
3740 | mds->locker->drop_locks(mdr.get()); | |
3741 | mdr->drop_local_auth_pins(); | |
3742 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); | |
3743 | return; | |
3744 | } | |
3745 | dir->fetch(new C_MDS_RetryRequest(mdcache, mdr), true); | |
3746 | return; | |
3747 | } | |
3748 | ||
3749 | respond_to_request(mdr, -ESTALE); | |
3750 | } else { | |
3751 | mdcache->open_ino(vino.ino, mds->mdsmap->get_metadata_pool(), new C_MDS_LookupIno2(this, mdr), false); | |
3752 | } | |
3753 | } | |
3754 | ||
7c673cae FG |
3755 | void Server::_lookup_ino_2(MDRequestRef& mdr, int r) |
3756 | { | |
3757 | inodeno_t ino = mdr->client_request->get_filepath().get_ino(); | |
3758 | dout(10) << "_lookup_ino_2 " << mdr.get() << " ino " << ino << " r=" << r << dendl; | |
3759 | ||
3760 | // `r` is a rank if >=0, else an error code | |
3761 | if (r >= 0) { | |
3762 | mds_rank_t dest_rank(r); | |
3763 | if (dest_rank == mds->get_nodeid()) | |
3764 | dispatch_client_request(mdr); | |
3765 | else | |
3766 | mdcache->request_forward(mdr, dest_rank); | |
3767 | return; | |
3768 | } | |
3769 | ||
3770 | // give up | |
3771 | if (r == -ENOENT || r == -ENODATA) | |
3772 | r = -ESTALE; | |
3773 | respond_to_request(mdr, r); | |
3774 | } | |
3775 | ||
3776 | ||
3777 | /* This function takes responsibility for the passed mdr*/ | |
3778 | void Server::handle_client_open(MDRequestRef& mdr) | |
3779 | { | |
11fdf7f2 | 3780 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
3781 | dout(7) << "open on " << req->get_filepath() << dendl; |
3782 | ||
3783 | int flags = req->head.args.open.flags; | |
3784 | int cmode = ceph_flags_to_mode(flags); | |
3785 | if (cmode < 0) { | |
3786 | respond_to_request(mdr, -EINVAL); | |
3787 | return; | |
3788 | } | |
3789 | ||
181888fb FG |
3790 | bool need_auth = !file_mode_is_readonly(cmode) || |
3791 | (flags & (CEPH_O_TRUNC | CEPH_O_DIRECTORY)); | |
7c673cae FG |
3792 | |
3793 | if ((cmode & CEPH_FILE_MODE_WR) && mdcache->is_readonly()) { | |
3794 | dout(7) << "read-only FS" << dendl; | |
3795 | respond_to_request(mdr, -EROFS); | |
3796 | return; | |
3797 | } | |
3798 | ||
11fdf7f2 TL |
3799 | MutationImpl::LockOpVec lov; |
3800 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, need_auth); | |
7c673cae FG |
3801 | if (!cur) |
3802 | return; | |
3803 | ||
3804 | if (cur->is_frozen() || cur->state_test(CInode::STATE_EXPORTINGCAPS)) { | |
11fdf7f2 | 3805 | ceph_assert(!need_auth); |
7c673cae | 3806 | mdr->done_locking = false; |
11fdf7f2 | 3807 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true); |
7c673cae FG |
3808 | if (!cur) |
3809 | return; | |
3810 | } | |
3811 | ||
3812 | if (!cur->inode.is_file()) { | |
3813 | // can only open non-regular inode with mode FILE_MODE_PIN, at least for now. | |
3814 | cmode = CEPH_FILE_MODE_PIN; | |
3815 | // the inode is symlink and client wants to follow it, ignore the O_TRUNC flag. | |
3816 | if (cur->inode.is_symlink() && !(flags & CEPH_O_NOFOLLOW)) | |
3817 | flags &= ~CEPH_O_TRUNC; | |
3818 | } | |
3819 | ||
3820 | dout(10) << "open flags = " << flags | |
3821 | << ", filemode = " << cmode | |
3822 | << ", need_auth = " << need_auth | |
3823 | << dendl; | |
3824 | ||
3825 | // regular file? | |
3826 | /*if (!cur->inode.is_file() && !cur->inode.is_dir()) { | |
3827 | dout(7) << "not a file or dir " << *cur << dendl; | |
3828 | respond_to_request(mdr, -ENXIO); // FIXME what error do we want? | |
3829 | return; | |
3830 | }*/ | |
3831 | if ((flags & CEPH_O_DIRECTORY) && !cur->inode.is_dir() && !cur->inode.is_symlink()) { | |
3832 | dout(7) << "specified O_DIRECTORY on non-directory " << *cur << dendl; | |
3833 | respond_to_request(mdr, -EINVAL); | |
3834 | return; | |
3835 | } | |
3836 | ||
3837 | if ((flags & CEPH_O_TRUNC) && !cur->inode.is_file()) { | |
3838 | dout(7) << "specified O_TRUNC on !(file|symlink) " << *cur << dendl; | |
3839 | // we should return -EISDIR for directory, return -EINVAL for other non-regular | |
3840 | respond_to_request(mdr, cur->inode.is_dir() ? -EISDIR : -EINVAL); | |
3841 | return; | |
3842 | } | |
3843 | ||
3844 | if (cur->inode.inline_data.version != CEPH_INLINE_NONE && | |
11fdf7f2 | 3845 | !mdr->session->get_connection()->has_feature(CEPH_FEATURE_MDS_INLINE_DATA)) { |
7c673cae FG |
3846 | dout(7) << "old client cannot open inline data file " << *cur << dendl; |
3847 | respond_to_request(mdr, -EPERM); | |
3848 | return; | |
3849 | } | |
3850 | ||
3851 | // snapped data is read only | |
3852 | if (mdr->snapid != CEPH_NOSNAP && | |
3853 | ((cmode & CEPH_FILE_MODE_WR) || req->may_write())) { | |
3854 | dout(7) << "snap " << mdr->snapid << " is read-only " << *cur << dendl; | |
3855 | respond_to_request(mdr, -EROFS); | |
3856 | return; | |
3857 | } | |
3858 | ||
3859 | unsigned mask = req->head.args.open.mask; | |
3860 | if (mask) { | |
3861 | Capability *cap = cur->get_client_cap(mdr->get_client()); | |
3862 | int issued = 0; | |
3863 | if (cap && (mdr->snapid == CEPH_NOSNAP || mdr->snapid <= cap->client_follows)) | |
3864 | issued = cap->issued(); | |
3865 | // permission bits, ACL/security xattrs | |
3866 | if ((mask & CEPH_CAP_AUTH_SHARED) && (issued & CEPH_CAP_AUTH_EXCL) == 0) | |
11fdf7f2 | 3867 | lov.add_rdlock(&cur->authlock); |
7c673cae | 3868 | if ((mask & CEPH_CAP_XATTR_SHARED) && (issued & CEPH_CAP_XATTR_EXCL) == 0) |
11fdf7f2 | 3869 | lov.add_rdlock(&cur->xattrlock); |
7c673cae FG |
3870 | |
3871 | mdr->getattr_caps = mask; | |
3872 | } | |
3873 | ||
3874 | // O_TRUNC | |
3875 | if ((flags & CEPH_O_TRUNC) && !mdr->has_completed) { | |
11fdf7f2 | 3876 | ceph_assert(cur->is_auth()); |
7c673cae | 3877 | |
11fdf7f2 TL |
3878 | lov.add_xlock(&cur->filelock); |
3879 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
3880 | return; |
3881 | ||
3882 | if (!check_access(mdr, cur, MAY_WRITE)) | |
3883 | return; | |
3884 | ||
3885 | // wait for pending truncate? | |
94b18763 | 3886 | const auto pi = cur->get_projected_inode(); |
7c673cae FG |
3887 | if (pi->is_truncating()) { |
3888 | dout(10) << " waiting for pending truncate from " << pi->truncate_from | |
3889 | << " to " << pi->truncate_size << " to complete on " << *cur << dendl; | |
3890 | mds->locker->drop_locks(mdr.get()); | |
3891 | mdr->drop_local_auth_pins(); | |
3892 | cur->add_waiter(CInode::WAIT_TRUNC, new C_MDS_RetryRequest(mdcache, mdr)); | |
3893 | return; | |
3894 | } | |
3895 | ||
3896 | do_open_truncate(mdr, cmode); | |
3897 | return; | |
3898 | } | |
3899 | ||
3900 | // sync filelock if snapped. | |
3901 | // this makes us wait for writers to flushsnaps, ensuring we get accurate metadata, | |
3902 | // and that data itself is flushed so that we can read the snapped data off disk. | |
3903 | if (mdr->snapid != CEPH_NOSNAP && !cur->is_dir()) { | |
11fdf7f2 | 3904 | lov.add_rdlock(&cur->filelock); |
7c673cae FG |
3905 | } |
3906 | ||
11fdf7f2 | 3907 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
3908 | return; |
3909 | ||
3910 | mask = MAY_READ; | |
3911 | if (cmode & CEPH_FILE_MODE_WR) | |
3912 | mask |= MAY_WRITE; | |
3913 | if (!check_access(mdr, cur, mask)) | |
3914 | return; | |
3915 | ||
28e407b8 AA |
3916 | utime_t now = ceph_clock_now(); |
3917 | mdr->set_mds_stamp(now); | |
3918 | ||
7c673cae FG |
3919 | if (cur->is_file() || cur->is_dir()) { |
3920 | if (mdr->snapid == CEPH_NOSNAP) { | |
3921 | // register new cap | |
3922 | Capability *cap = mds->locker->issue_new_caps(cur, cmode, mdr->session, 0, req->is_replay()); | |
3923 | if (cap) | |
3924 | dout(12) << "open issued caps " << ccap_string(cap->pending()) | |
3925 | << " for " << req->get_source() | |
3926 | << " on " << *cur << dendl; | |
3927 | } else { | |
3928 | int caps = ceph_caps_for_mode(cmode); | |
3929 | dout(12) << "open issued IMMUTABLE SNAP caps " << ccap_string(caps) | |
3930 | << " for " << req->get_source() | |
3931 | << " snapid " << mdr->snapid | |
3932 | << " on " << *cur << dendl; | |
3933 | mdr->snap_caps = caps; | |
3934 | } | |
3935 | } | |
3936 | ||
3937 | // increase max_size? | |
3938 | if (cmode & CEPH_FILE_MODE_WR) | |
3939 | mds->locker->check_inode_max_size(cur); | |
3940 | ||
3941 | // make sure this inode gets into the journal | |
3942 | if (cur->is_auth() && cur->last == CEPH_NOSNAP && | |
11fdf7f2 | 3943 | mdcache->open_file_table.should_log_open(cur)) { |
7c673cae FG |
3944 | EOpen *le = new EOpen(mds->mdlog); |
3945 | mdlog->start_entry(le); | |
3946 | le->add_clean_inode(cur); | |
7c673cae FG |
3947 | mdlog->submit_entry(le); |
3948 | } | |
3949 | ||
3950 | // hit pop | |
3951 | if (cmode & CEPH_FILE_MODE_WR) | |
11fdf7f2 | 3952 | mds->balancer->hit_inode(cur, META_POP_IWR); |
7c673cae | 3953 | else |
11fdf7f2 | 3954 | mds->balancer->hit_inode(cur, META_POP_IRD, |
7c673cae FG |
3955 | mdr->client_request->get_source().num()); |
3956 | ||
3957 | CDentry *dn = 0; | |
3958 | if (req->get_dentry_wanted()) { | |
11fdf7f2 | 3959 | ceph_assert(mdr->dn[0].size()); |
7c673cae FG |
3960 | dn = mdr->dn[0].back(); |
3961 | } | |
3962 | ||
3963 | mdr->tracei = cur; | |
3964 | mdr->tracedn = dn; | |
3965 | respond_to_request(mdr, 0); | |
3966 | } | |
3967 | ||
3968 | class C_MDS_openc_finish : public ServerLogContext { | |
3969 | CDentry *dn; | |
3970 | CInode *newi; | |
7c673cae | 3971 | public: |
11fdf7f2 TL |
3972 | C_MDS_openc_finish(Server *s, MDRequestRef& r, CDentry *d, CInode *ni) : |
3973 | ServerLogContext(s, r), dn(d), newi(ni) {} | |
7c673cae | 3974 | void finish(int r) override { |
11fdf7f2 | 3975 | ceph_assert(r == 0); |
7c673cae FG |
3976 | |
3977 | dn->pop_projected_linkage(); | |
3978 | ||
3979 | // dirty inode, dn, dir | |
3980 | newi->inode.version--; // a bit hacky, see C_MDS_mknod_finish | |
3981 | newi->mark_dirty(newi->inode.version+1, mdr->ls); | |
28e407b8 | 3982 | newi->mark_dirty_parent(mdr->ls, true); |
7c673cae FG |
3983 | |
3984 | mdr->apply(); | |
3985 | ||
3986 | get_mds()->locker->share_inode_max_size(newi); | |
3987 | ||
3988 | MDRequestRef null_ref; | |
3989 | get_mds()->mdcache->send_dentry_link(dn, null_ref); | |
3990 | ||
11fdf7f2 | 3991 | get_mds()->balancer->hit_inode(newi, META_POP_IWR); |
7c673cae FG |
3992 | |
3993 | server->respond_to_request(mdr, 0); | |
3994 | ||
11fdf7f2 | 3995 | ceph_assert(g_conf()->mds_kill_openc_at != 1); |
7c673cae FG |
3996 | } |
3997 | }; | |
3998 | ||
3999 | /* This function takes responsibility for the passed mdr*/ | |
4000 | void Server::handle_client_openc(MDRequestRef& mdr) | |
4001 | { | |
11fdf7f2 | 4002 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
4003 | client_t client = mdr->get_client(); |
4004 | ||
4005 | dout(7) << "open w/ O_CREAT on " << req->get_filepath() << dendl; | |
4006 | ||
4007 | int cmode = ceph_flags_to_mode(req->head.args.open.flags); | |
4008 | if (cmode < 0) { | |
4009 | respond_to_request(mdr, -EINVAL); | |
4010 | return; | |
4011 | } | |
4012 | ||
c07f9fc5 FG |
4013 | bool excl = req->head.args.open.flags & CEPH_O_EXCL; |
4014 | ||
4015 | if (!excl) { | |
11fdf7f2 TL |
4016 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
4017 | int r = mdcache->path_traverse(mdr, cf, req->get_filepath(), | |
7c673cae FG |
4018 | &mdr->dn[0], NULL, MDS_TRAVERSE_FORWARD); |
4019 | if (r > 0) return; | |
4020 | if (r == 0) { | |
4021 | // it existed. | |
4022 | handle_client_open(mdr); | |
4023 | return; | |
4024 | } | |
4025 | if (r < 0 && r != -ENOENT) { | |
4026 | if (r == -ESTALE) { | |
4027 | dout(10) << "FAIL on ESTALE but attempting recovery" << dendl; | |
11fdf7f2 | 4028 | MDSContext *c = new C_MDS_TryFindInode(this, mdr); |
7c673cae FG |
4029 | mdcache->find_ino_peers(req->get_filepath().get_ino(), c); |
4030 | } else { | |
4031 | dout(10) << "FAIL on error " << r << dendl; | |
4032 | respond_to_request(mdr, r); | |
4033 | } | |
4034 | return; | |
4035 | } | |
7c673cae FG |
4036 | } |
4037 | ||
11fdf7f2 TL |
4038 | MutationImpl::LockOpVec lov; |
4039 | file_layout_t *dir_layout = nullptr; | |
4040 | CDentry *dn = rdlock_path_xlock_dentry(mdr, 0, lov, | |
7c673cae FG |
4041 | !excl, false, false, &dir_layout); |
4042 | if (!dn) return; | |
4043 | if (mdr->snapid != CEPH_NOSNAP) { | |
4044 | respond_to_request(mdr, -EROFS); | |
4045 | return; | |
4046 | } | |
4047 | // set layout | |
4048 | file_layout_t layout; | |
4049 | if (dir_layout) | |
4050 | layout = *dir_layout; | |
4051 | else | |
4052 | layout = mdcache->default_file_layout; | |
4053 | ||
4054 | // What kind of client caps are required to complete this operation | |
4055 | uint64_t access = MAY_WRITE; | |
4056 | ||
4057 | const auto default_layout = layout; | |
4058 | ||
4059 | // fill in any special params from client | |
4060 | if (req->head.args.open.stripe_unit) | |
4061 | layout.stripe_unit = req->head.args.open.stripe_unit; | |
4062 | if (req->head.args.open.stripe_count) | |
4063 | layout.stripe_count = req->head.args.open.stripe_count; | |
4064 | if (req->head.args.open.object_size) | |
4065 | layout.object_size = req->head.args.open.object_size; | |
4066 | if (req->get_connection()->has_feature(CEPH_FEATURE_CREATEPOOLID) && | |
4067 | (__s32)req->head.args.open.pool >= 0) { | |
4068 | layout.pool_id = req->head.args.open.pool; | |
4069 | ||
4070 | // make sure we have as new a map as the client | |
4071 | if (req->get_mdsmap_epoch() > mds->mdsmap->get_epoch()) { | |
4072 | mds->wait_for_mdsmap(req->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache, mdr)); | |
4073 | return; | |
4074 | } | |
4075 | } | |
4076 | ||
4077 | // If client doesn't have capability to modify layout pools, then | |
4078 | // only permit this request if the requested pool matches what the | |
4079 | // file would have inherited anyway from its parent. | |
4080 | if (default_layout != layout) { | |
4081 | access |= MAY_SET_VXATTR; | |
4082 | } | |
4083 | ||
4084 | if (!layout.is_valid()) { | |
4085 | dout(10) << " invalid initial file layout" << dendl; | |
4086 | respond_to_request(mdr, -EINVAL); | |
4087 | return; | |
4088 | } | |
4089 | if (!mds->mdsmap->is_data_pool(layout.pool_id)) { | |
4090 | dout(10) << " invalid data pool " << layout.pool_id << dendl; | |
4091 | respond_to_request(mdr, -EINVAL); | |
4092 | return; | |
4093 | } | |
4094 | ||
c07f9fc5 | 4095 | // created null dn. |
7c673cae FG |
4096 | CDir *dir = dn->get_dir(); |
4097 | CInode *diri = dir->get_inode(); | |
11fdf7f2 TL |
4098 | lov.add_rdlock(&diri->authlock); |
4099 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
4100 | return; |
4101 | ||
4102 | if (!check_access(mdr, diri, access)) | |
4103 | return; | |
4104 | ||
4105 | if (!check_fragment_space(mdr, dir)) | |
4106 | return; | |
4107 | ||
4108 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
4109 | ||
4110 | if (!dnl->is_null()) { | |
4111 | // it existed. | |
11fdf7f2 | 4112 | ceph_assert(req->head.args.open.flags & CEPH_O_EXCL); |
7c673cae FG |
4113 | dout(10) << "O_EXCL, target exists, failing with -EEXIST" << dendl; |
4114 | mdr->tracei = dnl->get_inode(); | |
4115 | mdr->tracedn = dn; | |
4116 | respond_to_request(mdr, -EEXIST); | |
4117 | return; | |
4118 | } | |
4119 | ||
7c673cae | 4120 | // create inode. |
7c673cae FG |
4121 | CInode *in = prepare_new_inode(mdr, dn->get_dir(), inodeno_t(req->head.ino), |
4122 | req->head.args.open.mode | S_IFREG, &layout); | |
11fdf7f2 | 4123 | ceph_assert(in); |
7c673cae FG |
4124 | |
4125 | // it's a file. | |
4126 | dn->push_projected_linkage(in); | |
4127 | ||
4128 | in->inode.version = dn->pre_dirty(); | |
4129 | if (layout.pool_id != mdcache->default_file_layout.pool_id) | |
4130 | in->inode.add_old_pool(mdcache->default_file_layout.pool_id); | |
4131 | in->inode.update_backtrace(); | |
a8e16298 TL |
4132 | in->inode.rstat.rfiles = 1; |
4133 | ||
4134 | SnapRealm *realm = diri->find_snaprealm(); | |
11fdf7f2 TL |
4135 | snapid_t follows = mdcache->get_global_snaprealm()->get_newest_seq(); |
4136 | ceph_assert(follows >= realm->get_newest_seq()); | |
a8e16298 TL |
4137 | |
4138 | ceph_assert(dn->first == follows+1); | |
4139 | in->first = dn->first; | |
4140 | ||
4141 | // do the open | |
4142 | Capability *cap = mds->locker->issue_new_caps(in, cmode, mdr->session, realm, req->is_replay()); | |
4143 | in->authlock.set_state(LOCK_EXCL); | |
4144 | in->xattrlock.set_state(LOCK_EXCL); | |
4145 | ||
4146 | if (cap && (cmode & CEPH_FILE_MODE_WR)) { | |
7c673cae FG |
4147 | in->inode.client_ranges[client].range.first = 0; |
4148 | in->inode.client_ranges[client].range.last = in->inode.get_layout_size_increment(); | |
4149 | in->inode.client_ranges[client].follows = follows; | |
a8e16298 | 4150 | cap->mark_clientwriteable(); |
7c673cae | 4151 | } |
7c673cae FG |
4152 | |
4153 | // prepare finisher | |
4154 | mdr->ls = mdlog->get_current_segment(); | |
4155 | EUpdate *le = new EUpdate(mdlog, "openc"); | |
4156 | mdlog->start_entry(le); | |
4157 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
4158 | journal_allocated_inos(mdr, &le->metablob); | |
4159 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, dn->get_dir(), PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
4160 | le->metablob.add_primary_dentry(dn, in, true, true, true); | |
4161 | ||
7c673cae FG |
4162 | // make sure this inode gets into the journal |
4163 | le->metablob.add_opened_ino(in->ino()); | |
7c673cae | 4164 | |
11fdf7f2 | 4165 | C_MDS_openc_finish *fin = new C_MDS_openc_finish(this, mdr, dn, in); |
7c673cae FG |
4166 | |
4167 | if (mdr->client_request->get_connection()->has_feature(CEPH_FEATURE_REPLY_CREATE_INODE)) { | |
4168 | dout(10) << "adding ino to reply to indicate inode was created" << dendl; | |
4169 | // add the file created flag onto the reply if create_flags features is supported | |
11fdf7f2 | 4170 | encode(in->inode.ino, mdr->reply_extra_bl); |
7c673cae FG |
4171 | } |
4172 | ||
4173 | journal_and_reply(mdr, in, dn, le, fin); | |
4174 | ||
4175 | // We hit_dir (via hit_inode) in our finish callback, but by then we might | |
4176 | // have overshot the split size (multiple opencs in flight), so here is | |
4177 | // an early chance to split the dir if this openc makes it oversized. | |
4178 | mds->balancer->maybe_fragment(dir, false); | |
4179 | } | |
4180 | ||
4181 | ||
4182 | ||
4183 | void Server::handle_client_readdir(MDRequestRef& mdr) | |
4184 | { | |
11fdf7f2 | 4185 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae | 4186 | client_t client = req->get_source().num(); |
11fdf7f2 TL |
4187 | MutationImpl::LockOpVec lov; |
4188 | CInode *diri = rdlock_path_pin_ref(mdr, 0, lov, false, true); | |
7c673cae FG |
4189 | if (!diri) return; |
4190 | ||
4191 | // it's a directory, right? | |
4192 | if (!diri->is_dir()) { | |
4193 | // not a dir | |
4194 | dout(10) << "reply to " << *req << " readdir -ENOTDIR" << dendl; | |
4195 | respond_to_request(mdr, -ENOTDIR); | |
4196 | return; | |
4197 | } | |
4198 | ||
11fdf7f2 TL |
4199 | lov.add_rdlock(&diri->filelock); |
4200 | lov.add_rdlock(&diri->dirfragtreelock); | |
7c673cae | 4201 | |
11fdf7f2 | 4202 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
4203 | return; |
4204 | ||
4205 | if (!check_access(mdr, diri, MAY_READ)) | |
4206 | return; | |
4207 | ||
4208 | // which frag? | |
4209 | frag_t fg = (__u32)req->head.args.readdir.frag; | |
4210 | unsigned req_flags = (__u32)req->head.args.readdir.flags; | |
4211 | string offset_str = req->get_path2(); | |
4212 | ||
4213 | __u32 offset_hash = 0; | |
4214 | if (!offset_str.empty()) | |
4215 | offset_hash = ceph_frag_value(diri->hash_dentry_name(offset_str)); | |
4216 | else | |
4217 | offset_hash = (__u32)req->head.args.readdir.offset_hash; | |
4218 | ||
4219 | dout(10) << " frag " << fg << " offset '" << offset_str << "'" | |
4220 | << " offset_hash " << offset_hash << " flags " << req_flags << dendl; | |
4221 | ||
4222 | // does the frag exist? | |
4223 | if (diri->dirfragtree[fg.value()] != fg) { | |
4224 | frag_t newfg; | |
4225 | if (req_flags & CEPH_READDIR_REPLY_BITFLAGS) { | |
4226 | if (fg.contains((unsigned)offset_hash)) { | |
4227 | newfg = diri->dirfragtree[offset_hash]; | |
4228 | } else { | |
4229 | // client actually wants next frag | |
4230 | newfg = diri->dirfragtree[fg.value()]; | |
4231 | } | |
4232 | } else { | |
4233 | offset_str.clear(); | |
4234 | newfg = diri->dirfragtree[fg.value()]; | |
4235 | } | |
4236 | dout(10) << " adjust frag " << fg << " -> " << newfg << " " << diri->dirfragtree << dendl; | |
4237 | fg = newfg; | |
4238 | } | |
4239 | ||
4240 | CDir *dir = try_open_auth_dirfrag(diri, fg, mdr); | |
4241 | if (!dir) return; | |
4242 | ||
4243 | // ok! | |
4244 | dout(10) << "handle_client_readdir on " << *dir << dendl; | |
11fdf7f2 | 4245 | ceph_assert(dir->is_auth()); |
7c673cae FG |
4246 | |
4247 | if (!dir->is_complete()) { | |
4248 | if (dir->is_frozen()) { | |
4249 | dout(7) << "dir is frozen " << *dir << dendl; | |
4250 | mds->locker->drop_locks(mdr.get()); | |
4251 | mdr->drop_local_auth_pins(); | |
4252 | dir->add_waiter(CDir::WAIT_UNFREEZE, new C_MDS_RetryRequest(mdcache, mdr)); | |
4253 | return; | |
4254 | } | |
4255 | // fetch | |
4256 | dout(10) << " incomplete dir contents for readdir on " << *dir << ", fetching" << dendl; | |
4257 | dir->fetch(new C_MDS_RetryRequest(mdcache, mdr), true); | |
4258 | return; | |
4259 | } | |
4260 | ||
4261 | #ifdef MDS_VERIFY_FRAGSTAT | |
4262 | dir->verify_fragstat(); | |
4263 | #endif | |
4264 | ||
4265 | utime_t now = ceph_clock_now(); | |
4266 | mdr->set_mds_stamp(now); | |
4267 | ||
4268 | snapid_t snapid = mdr->snapid; | |
4269 | dout(10) << "snapid " << snapid << dendl; | |
4270 | ||
4271 | SnapRealm *realm = diri->find_snaprealm(); | |
4272 | ||
4273 | unsigned max = req->head.args.readdir.max_entries; | |
4274 | if (!max) | |
4275 | max = dir->get_num_any(); // whatever, something big. | |
4276 | unsigned max_bytes = req->head.args.readdir.max_bytes; | |
4277 | if (!max_bytes) | |
4278 | // make sure at least one item can be encoded | |
11fdf7f2 | 4279 | max_bytes = (512 << 10) + g_conf()->mds_max_xattr_pairs_size; |
7c673cae FG |
4280 | |
4281 | // start final blob | |
4282 | bufferlist dirbl; | |
11fdf7f2 TL |
4283 | DirStat ds; |
4284 | ds.frag = dir->get_frag(); | |
4285 | ds.auth = dir->get_dir_auth().first; | |
4286 | if (dir->is_auth()) | |
4287 | dir->get_dist_spec(ds.dist, mds->get_nodeid()); | |
4288 | ||
4289 | dir->encode_dirstat(dirbl, mdr->session->info, ds); | |
7c673cae FG |
4290 | |
4291 | // count bytes available. | |
4292 | // this isn't perfect, but we should capture the main variable/unbounded size items! | |
4293 | int front_bytes = dirbl.length() + sizeof(__u32) + sizeof(__u8)*2; | |
4294 | int bytes_left = max_bytes - front_bytes; | |
4295 | bytes_left -= realm->get_snap_trace().length(); | |
4296 | ||
4297 | // build dir contents | |
4298 | bufferlist dnbl; | |
4299 | __u32 numfiles = 0; | |
4300 | bool start = !offset_hash && offset_str.empty(); | |
7c673cae FG |
4301 | // skip all dns < dentry_key_t(snapid, offset_str, offset_hash) |
4302 | dentry_key_t skip_key(snapid, offset_str.c_str(), offset_hash); | |
181888fb FG |
4303 | auto it = start ? dir->begin() : dir->lower_bound(skip_key); |
4304 | bool end = (it == dir->end()); | |
4305 | for (; !end && numfiles < max; end = (it == dir->end())) { | |
7c673cae FG |
4306 | CDentry *dn = it->second; |
4307 | ++it; | |
4308 | ||
4309 | if (dn->state_test(CDentry::STATE_PURGING)) | |
4310 | continue; | |
4311 | ||
4312 | bool dnp = dn->use_projected(client, mdr); | |
4313 | CDentry::linkage_t *dnl = dnp ? dn->get_projected_linkage() : dn->get_linkage(); | |
4314 | ||
4315 | if (dnl->is_null()) | |
4316 | continue; | |
4317 | ||
4318 | if (dn->last < snapid || dn->first > snapid) { | |
4319 | dout(20) << "skipping non-overlapping snap " << *dn << dendl; | |
4320 | continue; | |
4321 | } | |
4322 | ||
4323 | if (!start) { | |
4324 | dentry_key_t offset_key(dn->last, offset_str.c_str(), offset_hash); | |
4325 | if (!(offset_key < dn->key())) | |
4326 | continue; | |
4327 | } | |
4328 | ||
4329 | CInode *in = dnl->get_inode(); | |
4330 | ||
4331 | if (in && in->ino() == CEPH_INO_CEPH) | |
4332 | continue; | |
4333 | ||
4334 | // remote link? | |
4335 | // better for the MDS to do the work, if we think the client will stat any of these files. | |
4336 | if (dnl->is_remote() && !in) { | |
4337 | in = mdcache->get_inode(dnl->get_remote_ino()); | |
4338 | if (in) { | |
4339 | dn->link_remote(dnl, in); | |
4340 | } else if (dn->state_test(CDentry::STATE_BADREMOTEINO)) { | |
4341 | dout(10) << "skipping bad remote ino on " << *dn << dendl; | |
4342 | continue; | |
4343 | } else { | |
4344 | // touch everything i _do_ have | |
94b18763 FG |
4345 | for (auto &p : *dir) { |
4346 | if (!p.second->get_linkage()->is_null()) | |
4347 | mdcache->lru.lru_touch(p.second); | |
4348 | } | |
7c673cae FG |
4349 | |
4350 | // already issued caps and leases, reply immediately. | |
4351 | if (dnbl.length() > 0) { | |
4352 | mdcache->open_remote_dentry(dn, dnp, new C_MDSInternalNoop); | |
4353 | dout(10) << " open remote dentry after caps were issued, stopping at " | |
4354 | << dnbl.length() << " < " << bytes_left << dendl; | |
4355 | break; | |
4356 | } | |
4357 | ||
4358 | mds->locker->drop_locks(mdr.get()); | |
4359 | mdr->drop_local_auth_pins(); | |
4360 | mdcache->open_remote_dentry(dn, dnp, new C_MDS_RetryRequest(mdcache, mdr)); | |
4361 | return; | |
4362 | } | |
4363 | } | |
11fdf7f2 | 4364 | ceph_assert(in); |
7c673cae | 4365 | |
94b18763 | 4366 | if ((int)(dnbl.length() + dn->get_name().length() + sizeof(__u32) + sizeof(LeaseStat)) > bytes_left) { |
7c673cae FG |
4367 | dout(10) << " ran out of room, stopping at " << dnbl.length() << " < " << bytes_left << dendl; |
4368 | break; | |
4369 | } | |
4370 | ||
4371 | unsigned start_len = dnbl.length(); | |
4372 | ||
4373 | // dentry | |
4374 | dout(12) << "including dn " << *dn << dendl; | |
11fdf7f2 | 4375 | encode(dn->get_name(), dnbl); |
7c673cae FG |
4376 | mds->locker->issue_client_lease(dn, client, dnbl, now, mdr->session); |
4377 | ||
4378 | // inode | |
4379 | dout(12) << "including inode " << *in << dendl; | |
4380 | int r = in->encode_inodestat(dnbl, mdr->session, realm, snapid, bytes_left - (int)dnbl.length()); | |
4381 | if (r < 0) { | |
4382 | // chop off dn->name, lease | |
4383 | dout(10) << " ran out of room, stopping at " << start_len << " < " << bytes_left << dendl; | |
4384 | bufferlist keep; | |
4385 | keep.substr_of(dnbl, 0, start_len); | |
4386 | dnbl.swap(keep); | |
4387 | break; | |
4388 | } | |
11fdf7f2 | 4389 | ceph_assert(r >= 0); |
7c673cae FG |
4390 | numfiles++; |
4391 | ||
4392 | // touch dn | |
4393 | mdcache->lru.lru_touch(dn); | |
4394 | } | |
4395 | ||
4396 | __u16 flags = 0; | |
4397 | if (end) { | |
4398 | flags = CEPH_READDIR_FRAG_END; | |
4399 | if (start) | |
4400 | flags |= CEPH_READDIR_FRAG_COMPLETE; // FIXME: what purpose does this serve | |
4401 | } | |
4402 | // client only understand END and COMPLETE flags ? | |
4403 | if (req_flags & CEPH_READDIR_REPLY_BITFLAGS) { | |
4404 | flags |= CEPH_READDIR_HASH_ORDER | CEPH_READDIR_OFFSET_HASH; | |
4405 | } | |
4406 | ||
4407 | // finish final blob | |
11fdf7f2 TL |
4408 | encode(numfiles, dirbl); |
4409 | encode(flags, dirbl); | |
7c673cae FG |
4410 | dirbl.claim_append(dnbl); |
4411 | ||
4412 | // yay, reply | |
4413 | dout(10) << "reply to " << *req << " readdir num=" << numfiles | |
4414 | << " bytes=" << dirbl.length() | |
4415 | << " start=" << (int)start | |
4416 | << " end=" << (int)end | |
4417 | << dendl; | |
4418 | mdr->reply_extra_bl = dirbl; | |
4419 | ||
4420 | // bump popularity. NOTE: this doesn't quite capture it. | |
11fdf7f2 | 4421 | mds->balancer->hit_dir(dir, META_POP_IRD, -1, numfiles); |
7c673cae FG |
4422 | |
4423 | // reply | |
4424 | mdr->tracei = diri; | |
4425 | respond_to_request(mdr, 0); | |
4426 | } | |
4427 | ||
4428 | ||
4429 | ||
4430 | // =============================================================================== | |
4431 | // INODE UPDATES | |
4432 | ||
4433 | ||
4434 | /* | |
4435 | * finisher for basic inode updates | |
4436 | */ | |
4437 | class C_MDS_inode_update_finish : public ServerLogContext { | |
4438 | CInode *in; | |
11fdf7f2 | 4439 | bool truncating_smaller, changed_ranges, new_realm; |
7c673cae FG |
4440 | public: |
4441 | C_MDS_inode_update_finish(Server *s, MDRequestRef& r, CInode *i, | |
11fdf7f2 TL |
4442 | bool sm=false, bool cr=false, bool nr=false) : |
4443 | ServerLogContext(s, r), in(i), | |
4444 | truncating_smaller(sm), changed_ranges(cr), new_realm(nr) { } | |
7c673cae | 4445 | void finish(int r) override { |
11fdf7f2 | 4446 | ceph_assert(r == 0); |
7c673cae FG |
4447 | |
4448 | // apply | |
4449 | in->pop_and_dirty_projected_inode(mdr->ls); | |
4450 | mdr->apply(); | |
4451 | ||
11fdf7f2 TL |
4452 | MDSRank *mds = get_mds(); |
4453 | ||
7c673cae FG |
4454 | // notify any clients |
4455 | if (truncating_smaller && in->inode.is_truncating()) { | |
11fdf7f2 TL |
4456 | mds->locker->issue_truncate(in); |
4457 | mds->mdcache->truncate_inode(in, mdr->ls); | |
4458 | } | |
4459 | ||
4460 | if (new_realm) { | |
4461 | int op = CEPH_SNAP_OP_SPLIT; | |
4462 | mds->mdcache->send_snap_update(in, 0, op); | |
4463 | mds->mdcache->do_realm_invalidate_and_update_notify(in, op); | |
7c673cae FG |
4464 | } |
4465 | ||
11fdf7f2 | 4466 | get_mds()->balancer->hit_inode(in, META_POP_IWR); |
7c673cae FG |
4467 | |
4468 | server->respond_to_request(mdr, 0); | |
4469 | ||
4470 | if (changed_ranges) | |
4471 | get_mds()->locker->share_inode_max_size(in); | |
4472 | } | |
4473 | }; | |
4474 | ||
4475 | void Server::handle_client_file_setlock(MDRequestRef& mdr) | |
4476 | { | |
11fdf7f2 TL |
4477 | const MClientRequest::const_ref &req = mdr->client_request; |
4478 | MutationImpl::LockOpVec lov; | |
7c673cae FG |
4479 | |
4480 | // get the inode to operate on, and set up any locks needed for that | |
11fdf7f2 | 4481 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true); |
7c673cae FG |
4482 | if (!cur) |
4483 | return; | |
4484 | ||
11fdf7f2 | 4485 | lov.add_xlock(&cur->flocklock); |
7c673cae FG |
4486 | /* acquire_locks will return true if it gets the locks. If it fails, |
4487 | it will redeliver this request at a later date, so drop the request. | |
4488 | */ | |
11fdf7f2 | 4489 | if (!mds->locker->acquire_locks(mdr, lov)) { |
7c673cae FG |
4490 | dout(10) << "handle_client_file_setlock could not get locks!" << dendl; |
4491 | return; | |
4492 | } | |
4493 | ||
4494 | // copy the lock change into a ceph_filelock so we can store/apply it | |
4495 | ceph_filelock set_lock; | |
4496 | set_lock.start = req->head.args.filelock_change.start; | |
4497 | set_lock.length = req->head.args.filelock_change.length; | |
4498 | set_lock.client = req->get_orig_source().num(); | |
4499 | set_lock.owner = req->head.args.filelock_change.owner; | |
4500 | set_lock.pid = req->head.args.filelock_change.pid; | |
4501 | set_lock.type = req->head.args.filelock_change.type; | |
4502 | bool will_wait = req->head.args.filelock_change.wait; | |
4503 | ||
4504 | dout(10) << "handle_client_file_setlock: " << set_lock << dendl; | |
4505 | ||
4506 | ceph_lock_state_t *lock_state = NULL; | |
4507 | bool interrupt = false; | |
4508 | ||
4509 | // get the appropriate lock state | |
4510 | switch (req->head.args.filelock_change.rule) { | |
4511 | case CEPH_LOCK_FLOCK_INTR: | |
4512 | interrupt = true; | |
4513 | // fall-thru | |
4514 | case CEPH_LOCK_FLOCK: | |
4515 | lock_state = cur->get_flock_lock_state(); | |
4516 | break; | |
4517 | ||
4518 | case CEPH_LOCK_FCNTL_INTR: | |
4519 | interrupt = true; | |
4520 | // fall-thru | |
4521 | case CEPH_LOCK_FCNTL: | |
4522 | lock_state = cur->get_fcntl_lock_state(); | |
4523 | break; | |
4524 | ||
4525 | default: | |
4526 | dout(10) << "got unknown lock type " << set_lock.type | |
4527 | << ", dropping request!" << dendl; | |
4528 | respond_to_request(mdr, -EOPNOTSUPP); | |
4529 | return; | |
4530 | } | |
4531 | ||
4532 | dout(10) << " state prior to lock change: " << *lock_state << dendl; | |
4533 | if (CEPH_LOCK_UNLOCK == set_lock.type) { | |
4534 | list<ceph_filelock> activated_locks; | |
11fdf7f2 | 4535 | MDSContext::vec waiters; |
7c673cae FG |
4536 | if (lock_state->is_waiting(set_lock)) { |
4537 | dout(10) << " unlock removing waiting lock " << set_lock << dendl; | |
4538 | lock_state->remove_waiting(set_lock); | |
4539 | cur->take_waiting(CInode::WAIT_FLOCK, waiters); | |
4540 | } else if (!interrupt) { | |
4541 | dout(10) << " unlock attempt on " << set_lock << dendl; | |
4542 | lock_state->remove_lock(set_lock, activated_locks); | |
4543 | cur->take_waiting(CInode::WAIT_FLOCK, waiters); | |
4544 | } | |
4545 | mds->queue_waiters(waiters); | |
4546 | ||
4547 | respond_to_request(mdr, 0); | |
4548 | } else { | |
4549 | dout(10) << " lock attempt on " << set_lock << dendl; | |
4550 | bool deadlock = false; | |
4551 | if (mdr->more()->flock_was_waiting && | |
4552 | !lock_state->is_waiting(set_lock)) { | |
4553 | dout(10) << " was waiting for lock but not anymore, must have been canceled " << set_lock << dendl; | |
4554 | respond_to_request(mdr, -EINTR); | |
4555 | } else if (!lock_state->add_lock(set_lock, will_wait, mdr->more()->flock_was_waiting, &deadlock)) { | |
4556 | dout(10) << " it failed on this attempt" << dendl; | |
4557 | // couldn't set lock right now | |
4558 | if (deadlock) { | |
4559 | respond_to_request(mdr, -EDEADLK); | |
4560 | } else if (!will_wait) { | |
4561 | respond_to_request(mdr, -EWOULDBLOCK); | |
4562 | } else { | |
4563 | dout(10) << " added to waiting list" << dendl; | |
11fdf7f2 | 4564 | ceph_assert(lock_state->is_waiting(set_lock)); |
7c673cae FG |
4565 | mdr->more()->flock_was_waiting = true; |
4566 | mds->locker->drop_locks(mdr.get()); | |
4567 | mdr->drop_local_auth_pins(); | |
1adf2230 AA |
4568 | mdr->mark_event("failed to add lock, waiting"); |
4569 | mdr->mark_nowarn(); | |
7c673cae FG |
4570 | cur->add_waiter(CInode::WAIT_FLOCK, new C_MDS_RetryRequest(mdcache, mdr)); |
4571 | } | |
4572 | } else | |
4573 | respond_to_request(mdr, 0); | |
4574 | } | |
4575 | dout(10) << " state after lock change: " << *lock_state << dendl; | |
4576 | } | |
4577 | ||
4578 | void Server::handle_client_file_readlock(MDRequestRef& mdr) | |
4579 | { | |
11fdf7f2 TL |
4580 | const MClientRequest::const_ref &req = mdr->client_request; |
4581 | MutationImpl::LockOpVec lov; | |
7c673cae FG |
4582 | |
4583 | // get the inode to operate on, and set up any locks needed for that | |
11fdf7f2 | 4584 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true); |
7c673cae FG |
4585 | if (!cur) |
4586 | return; | |
4587 | ||
4588 | /* acquire_locks will return true if it gets the locks. If it fails, | |
4589 | it will redeliver this request at a later date, so drop the request. | |
4590 | */ | |
11fdf7f2 TL |
4591 | lov.add_rdlock(&cur->flocklock); |
4592 | if (!mds->locker->acquire_locks(mdr, lov)) { | |
7c673cae FG |
4593 | dout(10) << "handle_client_file_readlock could not get locks!" << dendl; |
4594 | return; | |
4595 | } | |
4596 | ||
4597 | // copy the lock change into a ceph_filelock so we can store/apply it | |
4598 | ceph_filelock checking_lock; | |
4599 | checking_lock.start = req->head.args.filelock_change.start; | |
4600 | checking_lock.length = req->head.args.filelock_change.length; | |
4601 | checking_lock.client = req->get_orig_source().num(); | |
4602 | checking_lock.owner = req->head.args.filelock_change.owner; | |
4603 | checking_lock.pid = req->head.args.filelock_change.pid; | |
4604 | checking_lock.type = req->head.args.filelock_change.type; | |
4605 | ||
4606 | // get the appropriate lock state | |
4607 | ceph_lock_state_t *lock_state = NULL; | |
4608 | switch (req->head.args.filelock_change.rule) { | |
4609 | case CEPH_LOCK_FLOCK: | |
4610 | lock_state = cur->get_flock_lock_state(); | |
4611 | break; | |
4612 | ||
4613 | case CEPH_LOCK_FCNTL: | |
4614 | lock_state = cur->get_fcntl_lock_state(); | |
4615 | break; | |
4616 | ||
4617 | default: | |
4618 | dout(10) << "got unknown lock type " << checking_lock.type << dendl; | |
4619 | respond_to_request(mdr, -EINVAL); | |
4620 | return; | |
4621 | } | |
4622 | lock_state->look_for_lock(checking_lock); | |
4623 | ||
4624 | bufferlist lock_bl; | |
11fdf7f2 | 4625 | encode(checking_lock, lock_bl); |
7c673cae FG |
4626 | |
4627 | mdr->reply_extra_bl = lock_bl; | |
4628 | respond_to_request(mdr, 0); | |
4629 | } | |
4630 | ||
4631 | void Server::handle_client_setattr(MDRequestRef& mdr) | |
4632 | { | |
11fdf7f2 TL |
4633 | const MClientRequest::const_ref &req = mdr->client_request; |
4634 | MutationImpl::LockOpVec lov; | |
4635 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true); | |
7c673cae FG |
4636 | if (!cur) return; |
4637 | ||
4638 | if (mdr->snapid != CEPH_NOSNAP) { | |
4639 | respond_to_request(mdr, -EROFS); | |
4640 | return; | |
4641 | } | |
4642 | if (cur->ino() < MDS_INO_SYSTEM_BASE && !cur->is_base()) { | |
4643 | respond_to_request(mdr, -EPERM); | |
4644 | return; | |
4645 | } | |
4646 | ||
4647 | __u32 mask = req->head.args.setattr.mask; | |
4648 | __u32 access_mask = MAY_WRITE; | |
4649 | ||
4650 | // xlock inode | |
4651 | if (mask & (CEPH_SETATTR_MODE|CEPH_SETATTR_UID|CEPH_SETATTR_GID|CEPH_SETATTR_BTIME|CEPH_SETATTR_KILL_SGUID)) | |
11fdf7f2 | 4652 | lov.add_xlock(&cur->authlock); |
7c673cae | 4653 | if (mask & (CEPH_SETATTR_MTIME|CEPH_SETATTR_ATIME|CEPH_SETATTR_SIZE)) |
11fdf7f2 | 4654 | lov.add_xlock(&cur->filelock); |
7c673cae | 4655 | if (mask & CEPH_SETATTR_CTIME) |
11fdf7f2 | 4656 | lov.add_wrlock(&cur->versionlock); |
7c673cae | 4657 | |
11fdf7f2 | 4658 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
4659 | return; |
4660 | ||
4661 | if ((mask & CEPH_SETATTR_UID) && (cur->inode.uid != req->head.args.setattr.uid)) | |
4662 | access_mask |= MAY_CHOWN; | |
4663 | ||
4664 | if ((mask & CEPH_SETATTR_GID) && (cur->inode.gid != req->head.args.setattr.gid)) | |
4665 | access_mask |= MAY_CHGRP; | |
4666 | ||
4667 | if (!check_access(mdr, cur, access_mask)) | |
4668 | return; | |
4669 | ||
4670 | // trunc from bigger -> smaller? | |
94b18763 | 4671 | auto pip = cur->get_projected_inode(); |
7c673cae | 4672 | |
94b18763 | 4673 | uint64_t old_size = std::max<uint64_t>(pip->size, req->head.args.setattr.old_size); |
7c673cae FG |
4674 | |
4675 | // ENOSPC on growing file while full, but allow shrinks | |
4676 | if (is_full && req->head.args.setattr.size > old_size) { | |
4677 | dout(20) << __func__ << ": full, responding ENOSPC to setattr with larger size" << dendl; | |
4678 | respond_to_request(mdr, -ENOSPC); | |
4679 | return; | |
4680 | } | |
4681 | ||
4682 | bool truncating_smaller = false; | |
4683 | if (mask & CEPH_SETATTR_SIZE) { | |
4684 | truncating_smaller = req->head.args.setattr.size < old_size; | |
94b18763 FG |
4685 | if (truncating_smaller && pip->is_truncating()) { |
4686 | dout(10) << " waiting for pending truncate from " << pip->truncate_from | |
4687 | << " to " << pip->truncate_size << " to complete on " << *cur << dendl; | |
7c673cae FG |
4688 | mds->locker->drop_locks(mdr.get()); |
4689 | mdr->drop_local_auth_pins(); | |
4690 | cur->add_waiter(CInode::WAIT_TRUNC, new C_MDS_RetryRequest(mdcache, mdr)); | |
4691 | return; | |
4692 | } | |
4693 | } | |
4694 | ||
4695 | bool changed_ranges = false; | |
4696 | ||
4697 | // project update | |
4698 | mdr->ls = mdlog->get_current_segment(); | |
4699 | EUpdate *le = new EUpdate(mdlog, "setattr"); | |
4700 | mdlog->start_entry(le); | |
4701 | ||
94b18763 | 4702 | auto &pi = cur->project_inode(); |
7c673cae FG |
4703 | |
4704 | if (mask & CEPH_SETATTR_UID) | |
94b18763 | 4705 | pi.inode.uid = req->head.args.setattr.uid; |
7c673cae | 4706 | if (mask & CEPH_SETATTR_GID) |
94b18763 | 4707 | pi.inode.gid = req->head.args.setattr.gid; |
7c673cae FG |
4708 | |
4709 | if (mask & CEPH_SETATTR_MODE) | |
94b18763 | 4710 | pi.inode.mode = (pi.inode.mode & ~07777) | (req->head.args.setattr.mode & 07777); |
7c673cae | 4711 | else if ((mask & (CEPH_SETATTR_UID|CEPH_SETATTR_GID|CEPH_SETATTR_KILL_SGUID)) && |
94b18763 FG |
4712 | S_ISREG(pi.inode.mode) && |
4713 | (pi.inode.mode & (S_IXUSR|S_IXGRP|S_IXOTH))) { | |
4714 | pi.inode.mode &= ~(S_ISUID|S_ISGID); | |
7c673cae FG |
4715 | } |
4716 | ||
4717 | if (mask & CEPH_SETATTR_MTIME) | |
94b18763 | 4718 | pi.inode.mtime = req->head.args.setattr.mtime; |
7c673cae | 4719 | if (mask & CEPH_SETATTR_ATIME) |
94b18763 | 4720 | pi.inode.atime = req->head.args.setattr.atime; |
7c673cae | 4721 | if (mask & CEPH_SETATTR_BTIME) |
94b18763 | 4722 | pi.inode.btime = req->head.args.setattr.btime; |
7c673cae | 4723 | if (mask & (CEPH_SETATTR_ATIME | CEPH_SETATTR_MTIME | CEPH_SETATTR_BTIME)) |
94b18763 | 4724 | pi.inode.time_warp_seq++; // maybe not a timewarp, but still a serialization point. |
7c673cae FG |
4725 | if (mask & CEPH_SETATTR_SIZE) { |
4726 | if (truncating_smaller) { | |
94b18763 | 4727 | pi.inode.truncate(old_size, req->head.args.setattr.size); |
7c673cae FG |
4728 | le->metablob.add_truncate_start(cur->ino()); |
4729 | } else { | |
94b18763 FG |
4730 | pi.inode.size = req->head.args.setattr.size; |
4731 | pi.inode.rstat.rbytes = pi.inode.size; | |
7c673cae | 4732 | } |
94b18763 | 4733 | pi.inode.mtime = mdr->get_op_stamp(); |
7c673cae FG |
4734 | |
4735 | // adjust client's max_size? | |
94b18763 | 4736 | CInode::mempool_inode::client_range_map new_ranges; |
7c673cae | 4737 | bool max_increased = false; |
a8e16298 | 4738 | mds->locker->calc_new_client_ranges(cur, pi.inode.size, true, &new_ranges, &max_increased); |
94b18763 FG |
4739 | if (pi.inode.client_ranges != new_ranges) { |
4740 | dout(10) << " client_ranges " << pi.inode.client_ranges << " -> " << new_ranges << dendl; | |
4741 | pi.inode.client_ranges = new_ranges; | |
7c673cae FG |
4742 | changed_ranges = true; |
4743 | } | |
4744 | } | |
4745 | ||
94b18763 | 4746 | pi.inode.version = cur->pre_dirty(); |
91327a77 AA |
4747 | pi.inode.ctime = mdr->get_op_stamp(); |
4748 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
4749 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 4750 | pi.inode.change_attr++; |
7c673cae FG |
4751 | |
4752 | // log + wait | |
4753 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
4754 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
4755 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
4756 | ||
4757 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur, | |
4758 | truncating_smaller, changed_ranges)); | |
4759 | ||
4760 | // flush immediately if there are readers/writers waiting | |
11fdf7f2 | 4761 | if (mdr->is_xlocked(&cur->filelock) && |
7c673cae FG |
4762 | (cur->get_caps_wanted() & (CEPH_CAP_FILE_RD|CEPH_CAP_FILE_WR))) |
4763 | mds->mdlog->flush(); | |
4764 | } | |
4765 | ||
4766 | /* Takes responsibility for mdr */ | |
4767 | void Server::do_open_truncate(MDRequestRef& mdr, int cmode) | |
4768 | { | |
4769 | CInode *in = mdr->in[0]; | |
4770 | client_t client = mdr->get_client(); | |
11fdf7f2 | 4771 | ceph_assert(in); |
7c673cae FG |
4772 | |
4773 | dout(10) << "do_open_truncate " << *in << dendl; | |
4774 | ||
4775 | SnapRealm *realm = in->find_snaprealm(); | |
a8e16298 | 4776 | Capability *cap = mds->locker->issue_new_caps(in, cmode, mdr->session, realm, mdr->client_request->is_replay()); |
7c673cae FG |
4777 | |
4778 | mdr->ls = mdlog->get_current_segment(); | |
4779 | EUpdate *le = new EUpdate(mdlog, "open_truncate"); | |
4780 | mdlog->start_entry(le); | |
4781 | ||
4782 | // prepare | |
94b18763 FG |
4783 | auto &pi = in->project_inode(); |
4784 | pi.inode.version = in->pre_dirty(); | |
91327a77 AA |
4785 | pi.inode.mtime = pi.inode.ctime = mdr->get_op_stamp(); |
4786 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
4787 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 4788 | pi.inode.change_attr++; |
7c673cae | 4789 | |
94b18763 | 4790 | uint64_t old_size = std::max<uint64_t>(pi.inode.size, mdr->client_request->head.args.open.old_size); |
7c673cae | 4791 | if (old_size > 0) { |
94b18763 | 4792 | pi.inode.truncate(old_size, 0); |
7c673cae FG |
4793 | le->metablob.add_truncate_start(in->ino()); |
4794 | } | |
4795 | ||
4796 | bool changed_ranges = false; | |
a8e16298 | 4797 | if (cap && (cmode & CEPH_FILE_MODE_WR)) { |
94b18763 FG |
4798 | pi.inode.client_ranges[client].range.first = 0; |
4799 | pi.inode.client_ranges[client].range.last = pi.inode.get_layout_size_increment(); | |
11fdf7f2 | 4800 | pi.inode.client_ranges[client].follows = realm->get_newest_seq(); |
7c673cae | 4801 | changed_ranges = true; |
a8e16298 | 4802 | cap->mark_clientwriteable(); |
7c673cae FG |
4803 | } |
4804 | ||
4805 | le->metablob.add_client_req(mdr->reqid, mdr->client_request->get_oldest_client_tid()); | |
4806 | ||
4807 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
4808 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, in); | |
4809 | ||
4810 | // make sure ino gets into the journal | |
4811 | le->metablob.add_opened_ino(in->ino()); | |
7c673cae FG |
4812 | |
4813 | mdr->o_trunc = true; | |
4814 | ||
4815 | CDentry *dn = 0; | |
4816 | if (mdr->client_request->get_dentry_wanted()) { | |
11fdf7f2 | 4817 | ceph_assert(mdr->dn[0].size()); |
7c673cae FG |
4818 | dn = mdr->dn[0].back(); |
4819 | } | |
4820 | ||
4821 | journal_and_reply(mdr, in, dn, le, new C_MDS_inode_update_finish(this, mdr, in, old_size > 0, | |
4822 | changed_ranges)); | |
4823 | // Although the `open` part can give an early reply, the truncation won't | |
4824 | // happen until our EUpdate is persistent, to give the client a prompt | |
4825 | // response we must also flush that event. | |
4826 | mdlog->flush(); | |
4827 | } | |
4828 | ||
4829 | ||
4830 | /* This function cleans up the passed mdr */ | |
4831 | void Server::handle_client_setlayout(MDRequestRef& mdr) | |
4832 | { | |
11fdf7f2 TL |
4833 | const MClientRequest::const_ref &req = mdr->client_request; |
4834 | MutationImpl::LockOpVec lov; | |
4835 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true); | |
7c673cae FG |
4836 | if (!cur) return; |
4837 | ||
4838 | if (mdr->snapid != CEPH_NOSNAP) { | |
4839 | respond_to_request(mdr, -EROFS); | |
4840 | return; | |
4841 | } | |
4842 | if (!cur->is_file()) { | |
4843 | respond_to_request(mdr, -EINVAL); | |
4844 | return; | |
4845 | } | |
4846 | if (cur->get_projected_inode()->size || | |
4847 | cur->get_projected_inode()->truncate_seq > 1) { | |
4848 | respond_to_request(mdr, -ENOTEMPTY); | |
4849 | return; | |
4850 | } | |
4851 | ||
4852 | // validate layout | |
4853 | file_layout_t layout = cur->get_projected_inode()->layout; | |
4854 | // save existing layout for later | |
4855 | const auto old_layout = layout; | |
4856 | ||
4857 | int access = MAY_WRITE; | |
4858 | ||
4859 | if (req->head.args.setlayout.layout.fl_object_size > 0) | |
4860 | layout.object_size = req->head.args.setlayout.layout.fl_object_size; | |
4861 | if (req->head.args.setlayout.layout.fl_stripe_unit > 0) | |
4862 | layout.stripe_unit = req->head.args.setlayout.layout.fl_stripe_unit; | |
4863 | if (req->head.args.setlayout.layout.fl_stripe_count > 0) | |
4864 | layout.stripe_count=req->head.args.setlayout.layout.fl_stripe_count; | |
4865 | if (req->head.args.setlayout.layout.fl_pg_pool > 0) { | |
4866 | layout.pool_id = req->head.args.setlayout.layout.fl_pg_pool; | |
4867 | ||
4868 | // make sure we have as new a map as the client | |
4869 | if (req->get_mdsmap_epoch() > mds->mdsmap->get_epoch()) { | |
4870 | mds->wait_for_mdsmap(req->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache, mdr)); | |
4871 | return; | |
4872 | } | |
4873 | } | |
4874 | ||
4875 | // Don't permit layout modifications without 'p' caps | |
4876 | if (layout != old_layout) { | |
4877 | access |= MAY_SET_VXATTR; | |
4878 | } | |
4879 | ||
4880 | if (!layout.is_valid()) { | |
4881 | dout(10) << "bad layout" << dendl; | |
4882 | respond_to_request(mdr, -EINVAL); | |
4883 | return; | |
4884 | } | |
4885 | if (!mds->mdsmap->is_data_pool(layout.pool_id)) { | |
4886 | dout(10) << " invalid data pool " << layout.pool_id << dendl; | |
4887 | respond_to_request(mdr, -EINVAL); | |
4888 | return; | |
4889 | } | |
4890 | ||
11fdf7f2 TL |
4891 | lov.add_xlock(&cur->filelock); |
4892 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
4893 | return; |
4894 | ||
4895 | if (!check_access(mdr, cur, access)) | |
4896 | return; | |
4897 | ||
4898 | // project update | |
94b18763 FG |
4899 | auto &pi = cur->project_inode(); |
4900 | pi.inode.layout = layout; | |
7c673cae | 4901 | // add the old pool to the inode |
94b18763 FG |
4902 | pi.inode.add_old_pool(old_layout.pool_id); |
4903 | pi.inode.version = cur->pre_dirty(); | |
91327a77 AA |
4904 | pi.inode.ctime = mdr->get_op_stamp(); |
4905 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
4906 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 4907 | pi.inode.change_attr++; |
7c673cae FG |
4908 | |
4909 | // log + wait | |
4910 | mdr->ls = mdlog->get_current_segment(); | |
4911 | EUpdate *le = new EUpdate(mdlog, "setlayout"); | |
4912 | mdlog->start_entry(le); | |
4913 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
4914 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
4915 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
4916 | ||
4917 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur)); | |
4918 | } | |
4919 | ||
4920 | void Server::handle_client_setdirlayout(MDRequestRef& mdr) | |
4921 | { | |
11fdf7f2 TL |
4922 | const MClientRequest::const_ref &req = mdr->client_request; |
4923 | MutationImpl::LockOpVec lov; | |
4924 | file_layout_t *dir_layout = nullptr; | |
4925 | CInode *cur = rdlock_path_pin_ref(mdr, 0, lov, true, false, &dir_layout); | |
7c673cae FG |
4926 | if (!cur) return; |
4927 | ||
4928 | if (mdr->snapid != CEPH_NOSNAP) { | |
4929 | respond_to_request(mdr, -EROFS); | |
4930 | return; | |
4931 | } | |
4932 | ||
4933 | if (!cur->is_dir()) { | |
4934 | respond_to_request(mdr, -ENOTDIR); | |
4935 | return; | |
4936 | } | |
4937 | ||
11fdf7f2 TL |
4938 | lov.add_xlock(&cur->policylock); |
4939 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
4940 | return; |
4941 | ||
4942 | // validate layout | |
94b18763 | 4943 | const auto old_pi = cur->get_projected_inode(); |
7c673cae FG |
4944 | file_layout_t layout; |
4945 | if (old_pi->has_layout()) | |
4946 | layout = old_pi->layout; | |
4947 | else if (dir_layout) | |
4948 | layout = *dir_layout; | |
4949 | else | |
4950 | layout = mdcache->default_file_layout; | |
4951 | ||
4952 | // Level of access required to complete | |
4953 | int access = MAY_WRITE; | |
4954 | ||
4955 | const auto old_layout = layout; | |
4956 | ||
4957 | if (req->head.args.setlayout.layout.fl_object_size > 0) | |
4958 | layout.object_size = req->head.args.setlayout.layout.fl_object_size; | |
4959 | if (req->head.args.setlayout.layout.fl_stripe_unit > 0) | |
4960 | layout.stripe_unit = req->head.args.setlayout.layout.fl_stripe_unit; | |
4961 | if (req->head.args.setlayout.layout.fl_stripe_count > 0) | |
4962 | layout.stripe_count=req->head.args.setlayout.layout.fl_stripe_count; | |
4963 | if (req->head.args.setlayout.layout.fl_pg_pool > 0) { | |
4964 | layout.pool_id = req->head.args.setlayout.layout.fl_pg_pool; | |
4965 | // make sure we have as new a map as the client | |
4966 | if (req->get_mdsmap_epoch() > mds->mdsmap->get_epoch()) { | |
4967 | mds->wait_for_mdsmap(req->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache, mdr)); | |
4968 | return; | |
4969 | } | |
4970 | } | |
4971 | ||
4972 | if (layout != old_layout) { | |
4973 | access |= MAY_SET_VXATTR; | |
4974 | } | |
4975 | ||
4976 | if (!layout.is_valid()) { | |
4977 | dout(10) << "bad layout" << dendl; | |
4978 | respond_to_request(mdr, -EINVAL); | |
4979 | return; | |
4980 | } | |
4981 | if (!mds->mdsmap->is_data_pool(layout.pool_id)) { | |
4982 | dout(10) << " invalid data pool " << layout.pool_id << dendl; | |
4983 | respond_to_request(mdr, -EINVAL); | |
4984 | return; | |
4985 | } | |
4986 | ||
4987 | if (!check_access(mdr, cur, access)) | |
4988 | return; | |
4989 | ||
94b18763 FG |
4990 | auto &pi = cur->project_inode(); |
4991 | pi.inode.layout = layout; | |
4992 | pi.inode.version = cur->pre_dirty(); | |
7c673cae FG |
4993 | |
4994 | // log + wait | |
4995 | mdr->ls = mdlog->get_current_segment(); | |
4996 | EUpdate *le = new EUpdate(mdlog, "setlayout"); | |
4997 | mdlog->start_entry(le); | |
4998 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
4999 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
5000 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
5001 | ||
b32b8144 | 5002 | mdr->no_early_reply = true; |
7c673cae FG |
5003 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur)); |
5004 | } | |
5005 | ||
5006 | // XATTRS | |
5007 | ||
5008 | int Server::parse_layout_vxattr(string name, string value, const OSDMap& osdmap, | |
5009 | file_layout_t *layout, bool validate) | |
5010 | { | |
5011 | dout(20) << "parse_layout_vxattr name " << name << " value '" << value << "'" << dendl; | |
5012 | try { | |
5013 | if (name == "layout") { | |
5014 | string::iterator begin = value.begin(); | |
5015 | string::iterator end = value.end(); | |
5016 | keys_and_values<string::iterator> p; // create instance of parser | |
5017 | std::map<string, string> m; // map to receive results | |
5018 | if (!qi::parse(begin, end, p, m)) { // returns true if successful | |
5019 | return -EINVAL; | |
5020 | } | |
5021 | string left(begin, end); | |
5022 | dout(10) << " parsed " << m << " left '" << left << "'" << dendl; | |
5023 | if (begin != end) | |
5024 | return -EINVAL; | |
5025 | for (map<string,string>::iterator q = m.begin(); q != m.end(); ++q) { | |
5026 | // Skip validation on each attr, we do it once at the end (avoid | |
5027 | // rejecting intermediate states if the overall result is ok) | |
5028 | int r = parse_layout_vxattr(string("layout.") + q->first, q->second, | |
5029 | osdmap, layout, false); | |
5030 | if (r < 0) | |
5031 | return r; | |
5032 | } | |
5033 | } else if (name == "layout.object_size") { | |
5034 | layout->object_size = boost::lexical_cast<unsigned>(value); | |
5035 | } else if (name == "layout.stripe_unit") { | |
5036 | layout->stripe_unit = boost::lexical_cast<unsigned>(value); | |
5037 | } else if (name == "layout.stripe_count") { | |
5038 | layout->stripe_count = boost::lexical_cast<unsigned>(value); | |
5039 | } else if (name == "layout.pool") { | |
5040 | try { | |
5041 | layout->pool_id = boost::lexical_cast<unsigned>(value); | |
5042 | } catch (boost::bad_lexical_cast const&) { | |
5043 | int64_t pool = osdmap.lookup_pg_pool_name(value); | |
5044 | if (pool < 0) { | |
5045 | dout(10) << " unknown pool " << value << dendl; | |
5046 | return -ENOENT; | |
5047 | } | |
5048 | layout->pool_id = pool; | |
5049 | } | |
5050 | } else if (name == "layout.pool_namespace") { | |
5051 | layout->pool_ns = value; | |
5052 | } else { | |
5053 | dout(10) << " unknown layout vxattr " << name << dendl; | |
5054 | return -EINVAL; | |
5055 | } | |
5056 | } catch (boost::bad_lexical_cast const&) { | |
5057 | dout(10) << "bad vxattr value, unable to parse int for " << name << dendl; | |
5058 | return -EINVAL; | |
5059 | } | |
5060 | ||
5061 | if (validate && !layout->is_valid()) { | |
5062 | dout(10) << "bad layout" << dendl; | |
5063 | return -EINVAL; | |
5064 | } | |
5065 | if (!mds->mdsmap->is_data_pool(layout->pool_id)) { | |
5066 | dout(10) << " invalid data pool " << layout->pool_id << dendl; | |
5067 | return -EINVAL; | |
5068 | } | |
5069 | return 0; | |
5070 | } | |
5071 | ||
5072 | int Server::parse_quota_vxattr(string name, string value, quota_info_t *quota) | |
5073 | { | |
5074 | dout(20) << "parse_quota_vxattr name " << name << " value '" << value << "'" << dendl; | |
5075 | try { | |
5076 | if (name == "quota") { | |
5077 | string::iterator begin = value.begin(); | |
5078 | string::iterator end = value.end(); | |
11fdf7f2 TL |
5079 | if (begin == end) { |
5080 | // keep quota unchanged. (for create_quota_realm()) | |
5081 | return 0; | |
5082 | } | |
7c673cae FG |
5083 | keys_and_values<string::iterator> p; // create instance of parser |
5084 | std::map<string, string> m; // map to receive results | |
5085 | if (!qi::parse(begin, end, p, m)) { // returns true if successful | |
5086 | return -EINVAL; | |
5087 | } | |
5088 | string left(begin, end); | |
5089 | dout(10) << " parsed " << m << " left '" << left << "'" << dendl; | |
5090 | if (begin != end) | |
5091 | return -EINVAL; | |
5092 | for (map<string,string>::iterator q = m.begin(); q != m.end(); ++q) { | |
5093 | int r = parse_quota_vxattr(string("quota.") + q->first, q->second, quota); | |
5094 | if (r < 0) | |
5095 | return r; | |
5096 | } | |
5097 | } else if (name == "quota.max_bytes") { | |
5098 | int64_t q = boost::lexical_cast<int64_t>(value); | |
5099 | if (q < 0) | |
5100 | return -EINVAL; | |
5101 | quota->max_bytes = q; | |
5102 | } else if (name == "quota.max_files") { | |
5103 | int64_t q = boost::lexical_cast<int64_t>(value); | |
5104 | if (q < 0) | |
5105 | return -EINVAL; | |
5106 | quota->max_files = q; | |
5107 | } else { | |
5108 | dout(10) << " unknown quota vxattr " << name << dendl; | |
5109 | return -EINVAL; | |
5110 | } | |
5111 | } catch (boost::bad_lexical_cast const&) { | |
5112 | dout(10) << "bad vxattr value, unable to parse int for " << name << dendl; | |
5113 | return -EINVAL; | |
5114 | } | |
5115 | ||
5116 | if (!quota->is_valid()) { | |
5117 | dout(10) << "bad quota" << dendl; | |
5118 | return -EINVAL; | |
5119 | } | |
5120 | return 0; | |
5121 | } | |
5122 | ||
11fdf7f2 TL |
5123 | void Server::create_quota_realm(CInode *in) |
5124 | { | |
5125 | dout(10) << __func__ << " " << *in << dendl; | |
5126 | ||
5127 | auto req = MClientRequest::create(CEPH_MDS_OP_SETXATTR); | |
5128 | req->set_filepath(filepath(in->ino())); | |
5129 | req->set_string2("ceph.quota"); | |
5130 | // empty vxattr value | |
5131 | req->set_tid(mds->issue_tid()); | |
5132 | ||
5133 | mds->send_message_mds(req, in->authority().first); | |
5134 | } | |
5135 | ||
7c673cae FG |
5136 | /* |
5137 | * Verify that the file layout attribute carried by client | |
5138 | * is well-formatted. | |
5139 | * Return 0 on success, otherwise this function takes | |
5140 | * responsibility for the passed mdr. | |
5141 | */ | |
5142 | int Server::check_layout_vxattr(MDRequestRef& mdr, | |
5143 | string name, | |
5144 | string value, | |
5145 | file_layout_t *layout) | |
5146 | { | |
11fdf7f2 | 5147 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
5148 | epoch_t epoch; |
5149 | int r; | |
5150 | ||
5151 | mds->objecter->with_osdmap([&](const OSDMap& osdmap) { | |
5152 | r = parse_layout_vxattr(name, value, osdmap, layout); | |
5153 | epoch = osdmap.get_epoch(); | |
5154 | }); | |
5155 | ||
5156 | if (r == -ENOENT) { | |
5157 | ||
5158 | // we don't have the specified pool, make sure our map | |
5159 | // is newer than or as new as the client. | |
5160 | epoch_t req_epoch = req->get_osdmap_epoch(); | |
5161 | ||
5162 | if (req_epoch > epoch) { | |
5163 | ||
5164 | // well, our map is older. consult mds. | |
5165 | Context *fin = new C_IO_Wrapper(mds, new C_MDS_RetryRequest(mdcache, mdr)); | |
5166 | ||
5167 | if (!mds->objecter->wait_for_map(req_epoch, fin)) | |
5168 | return r; // wait, fin will retry this request later | |
5169 | ||
5170 | delete fin; | |
5171 | ||
5172 | // now we have at least as new a map as the client, try again. | |
5173 | mds->objecter->with_osdmap([&](const OSDMap& osdmap) { | |
5174 | r = parse_layout_vxattr(name, value, osdmap, layout); | |
5175 | epoch = osdmap.get_epoch(); | |
5176 | }); | |
5177 | ||
11fdf7f2 | 5178 | ceph_assert(epoch >= req_epoch); // otherwise wait_for_map() told a lie |
7c673cae FG |
5179 | |
5180 | } else if (req_epoch == 0 && !mdr->waited_for_osdmap) { | |
5181 | ||
5182 | // For compatibility with client w/ old code, we still need get the | |
5183 | // latest map. One day if COMPACT_VERSION of MClientRequest >=3, | |
5184 | // we can remove those code. | |
5185 | mdr->waited_for_osdmap = true; | |
5186 | mds->objecter->wait_for_latest_osdmap(new C_IO_Wrapper( | |
5187 | mds, new C_MDS_RetryRequest(mdcache, mdr))); | |
5188 | return r; | |
5189 | } | |
5190 | } | |
5191 | ||
5192 | if (r < 0) { | |
5193 | ||
5194 | if (r == -ENOENT) | |
5195 | r = -EINVAL; | |
5196 | ||
5197 | respond_to_request(mdr, r); | |
5198 | return r; | |
5199 | } | |
5200 | ||
5201 | // all is well | |
5202 | return 0; | |
5203 | } | |
5204 | ||
5205 | void Server::handle_set_vxattr(MDRequestRef& mdr, CInode *cur, | |
5206 | file_layout_t *dir_layout, | |
11fdf7f2 | 5207 | MutationImpl::LockOpVec& lov) |
7c673cae | 5208 | { |
11fdf7f2 | 5209 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
5210 | string name(req->get_path2()); |
5211 | bufferlist bl = req->get_data(); | |
5212 | string value (bl.c_str(), bl.length()); | |
5213 | dout(10) << "handle_set_vxattr " << name | |
5214 | << " val " << value.length() | |
5215 | << " bytes on " << *cur | |
5216 | << dendl; | |
5217 | ||
94b18763 | 5218 | CInode::mempool_inode *pip = nullptr; |
7c673cae FG |
5219 | string rest; |
5220 | ||
5221 | if (!check_access(mdr, cur, MAY_SET_VXATTR)) { | |
5222 | return; | |
5223 | } | |
5224 | ||
11fdf7f2 | 5225 | bool new_realm = false; |
7c673cae FG |
5226 | if (name.compare(0, 15, "ceph.dir.layout") == 0) { |
5227 | if (!cur->is_dir()) { | |
5228 | respond_to_request(mdr, -EINVAL); | |
5229 | return; | |
5230 | } | |
5231 | ||
5232 | file_layout_t layout; | |
5233 | if (cur->get_projected_inode()->has_layout()) | |
5234 | layout = cur->get_projected_inode()->layout; | |
5235 | else if (dir_layout) | |
5236 | layout = *dir_layout; | |
5237 | else | |
5238 | layout = mdcache->default_file_layout; | |
5239 | ||
5240 | rest = name.substr(name.find("layout")); | |
5241 | if (check_layout_vxattr(mdr, rest, value, &layout) < 0) | |
5242 | return; | |
5243 | ||
11fdf7f2 TL |
5244 | lov.add_xlock(&cur->policylock); |
5245 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5246 | return; |
5247 | ||
94b18763 FG |
5248 | auto &pi = cur->project_inode(); |
5249 | pi.inode.layout = layout; | |
b32b8144 | 5250 | mdr->no_early_reply = true; |
94b18763 | 5251 | pip = &pi.inode; |
7c673cae FG |
5252 | } else if (name.compare(0, 16, "ceph.file.layout") == 0) { |
5253 | if (!cur->is_file()) { | |
5254 | respond_to_request(mdr, -EINVAL); | |
5255 | return; | |
5256 | } | |
5257 | if (cur->get_projected_inode()->size || | |
5258 | cur->get_projected_inode()->truncate_seq > 1) { | |
5259 | respond_to_request(mdr, -ENOTEMPTY); | |
5260 | return; | |
5261 | } | |
5262 | file_layout_t layout = cur->get_projected_inode()->layout; | |
5263 | rest = name.substr(name.find("layout")); | |
5264 | if (check_layout_vxattr(mdr, rest, value, &layout) < 0) | |
5265 | return; | |
5266 | ||
11fdf7f2 TL |
5267 | lov.add_xlock(&cur->filelock); |
5268 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5269 | return; |
5270 | ||
94b18763 FG |
5271 | auto &pi = cur->project_inode(); |
5272 | int64_t old_pool = pi.inode.layout.pool_id; | |
5273 | pi.inode.add_old_pool(old_pool); | |
5274 | pi.inode.layout = layout; | |
94b18763 | 5275 | pip = &pi.inode; |
7c673cae FG |
5276 | } else if (name.compare(0, 10, "ceph.quota") == 0) { |
5277 | if (!cur->is_dir() || cur->is_root()) { | |
5278 | respond_to_request(mdr, -EINVAL); | |
5279 | return; | |
5280 | } | |
5281 | ||
5282 | quota_info_t quota = cur->get_projected_inode()->quota; | |
5283 | ||
5284 | rest = name.substr(name.find("quota")); | |
5285 | int r = parse_quota_vxattr(rest, value, "a); | |
5286 | if (r < 0) { | |
5287 | respond_to_request(mdr, r); | |
5288 | return; | |
5289 | } | |
5290 | ||
11fdf7f2 TL |
5291 | lov.add_xlock(&cur->policylock); |
5292 | if (quota.is_enable() && !cur->get_projected_srnode()) { | |
5293 | lov.add_xlock(&cur->snaplock); | |
5294 | new_realm = true; | |
5295 | } | |
5296 | ||
5297 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5298 | return; |
5299 | ||
11fdf7f2 | 5300 | auto &pi = cur->project_inode(false, new_realm); |
94b18763 FG |
5301 | pi.inode.quota = quota; |
5302 | ||
11fdf7f2 TL |
5303 | if (new_realm) { |
5304 | SnapRealm *realm = cur->find_snaprealm(); | |
5305 | auto seq = realm->get_newest_seq(); | |
5306 | auto &newsnap = *pi.snapnode; | |
5307 | newsnap.created = seq; | |
5308 | newsnap.seq = seq; | |
5309 | } | |
b32b8144 | 5310 | mdr->no_early_reply = true; |
94b18763 | 5311 | pip = &pi.inode; |
28e407b8 AA |
5312 | |
5313 | client_t exclude_ct = mdr->get_client(); | |
a8e16298 | 5314 | mdcache->broadcast_quota_to_client(cur, exclude_ct, true); |
7c673cae FG |
5315 | } else if (name.find("ceph.dir.pin") == 0) { |
5316 | if (!cur->is_dir() || cur->is_root()) { | |
5317 | respond_to_request(mdr, -EINVAL); | |
5318 | return; | |
5319 | } | |
5320 | ||
5321 | mds_rank_t rank; | |
5322 | try { | |
5323 | rank = boost::lexical_cast<mds_rank_t>(value); | |
5324 | if (rank < 0) rank = MDS_RANK_NONE; | |
5325 | } catch (boost::bad_lexical_cast const&) { | |
5326 | dout(10) << "bad vxattr value, unable to parse int for " << name << dendl; | |
5327 | respond_to_request(mdr, -EINVAL); | |
5328 | return; | |
5329 | } | |
5330 | ||
11fdf7f2 TL |
5331 | lov.add_xlock(&cur->policylock); |
5332 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5333 | return; |
5334 | ||
94b18763 | 5335 | auto &pi = cur->project_inode(); |
7c673cae | 5336 | cur->set_export_pin(rank); |
94b18763 | 5337 | pip = &pi.inode; |
7c673cae FG |
5338 | } else { |
5339 | dout(10) << " unknown vxattr " << name << dendl; | |
5340 | respond_to_request(mdr, -EINVAL); | |
5341 | return; | |
5342 | } | |
5343 | ||
94b18763 | 5344 | pip->change_attr++; |
91327a77 AA |
5345 | pip->ctime = mdr->get_op_stamp(); |
5346 | if (mdr->get_op_stamp() > pip->rstat.rctime) | |
5347 | pip->rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 5348 | pip->version = cur->pre_dirty(); |
7c673cae | 5349 | if (cur->is_file()) |
94b18763 | 5350 | pip->update_backtrace(); |
7c673cae FG |
5351 | |
5352 | // log + wait | |
5353 | mdr->ls = mdlog->get_current_segment(); | |
5354 | EUpdate *le = new EUpdate(mdlog, "set vxattr layout"); | |
5355 | mdlog->start_entry(le); | |
5356 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5357 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
5358 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
5359 | ||
11fdf7f2 TL |
5360 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur, |
5361 | false, false, new_realm)); | |
7c673cae FG |
5362 | return; |
5363 | } | |
5364 | ||
5365 | void Server::handle_remove_vxattr(MDRequestRef& mdr, CInode *cur, | |
5366 | file_layout_t *dir_layout, | |
11fdf7f2 | 5367 | MutationImpl::LockOpVec& lov) |
7c673cae | 5368 | { |
11fdf7f2 | 5369 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
5370 | string name(req->get_path2()); |
5371 | ||
5372 | dout(10) << __func__ << " " << name << " on " << *cur << dendl; | |
5373 | ||
5374 | if (name == "ceph.dir.layout") { | |
5375 | if (!cur->is_dir()) { | |
5376 | respond_to_request(mdr, -ENODATA); | |
5377 | return; | |
5378 | } | |
5379 | if (cur->is_root()) { | |
5380 | dout(10) << "can't remove layout policy on the root directory" << dendl; | |
5381 | respond_to_request(mdr, -EINVAL); | |
5382 | return; | |
5383 | } | |
5384 | ||
5385 | if (!cur->get_projected_inode()->has_layout()) { | |
5386 | respond_to_request(mdr, -ENODATA); | |
5387 | return; | |
5388 | } | |
5389 | ||
11fdf7f2 TL |
5390 | lov.add_xlock(&cur->policylock); |
5391 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5392 | return; |
5393 | ||
94b18763 FG |
5394 | auto &pi = cur->project_inode(); |
5395 | pi.inode.clear_layout(); | |
5396 | pi.inode.version = cur->pre_dirty(); | |
7c673cae FG |
5397 | |
5398 | // log + wait | |
5399 | mdr->ls = mdlog->get_current_segment(); | |
5400 | EUpdate *le = new EUpdate(mdlog, "remove dir layout vxattr"); | |
5401 | mdlog->start_entry(le); | |
5402 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5403 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
5404 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
5405 | ||
b32b8144 | 5406 | mdr->no_early_reply = true; |
7c673cae FG |
5407 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur)); |
5408 | return; | |
5409 | } else if (name == "ceph.dir.layout.pool_namespace" | |
5410 | || name == "ceph.file.layout.pool_namespace") { | |
5411 | // Namespace is the only layout field that has a meaningful | |
5412 | // null/none value (empty string, means default layout). Is equivalent | |
5413 | // to a setxattr with empty string: pass through the empty payload of | |
5414 | // the rmxattr request to do this. | |
11fdf7f2 | 5415 | handle_set_vxattr(mdr, cur, dir_layout, lov); |
7c673cae FG |
5416 | return; |
5417 | } | |
5418 | ||
5419 | respond_to_request(mdr, -ENODATA); | |
5420 | } | |
5421 | ||
5422 | class C_MDS_inode_xattr_update_finish : public ServerLogContext { | |
5423 | CInode *in; | |
5424 | public: | |
5425 | ||
5426 | C_MDS_inode_xattr_update_finish(Server *s, MDRequestRef& r, CInode *i) : | |
5427 | ServerLogContext(s, r), in(i) { } | |
5428 | void finish(int r) override { | |
11fdf7f2 | 5429 | ceph_assert(r == 0); |
7c673cae FG |
5430 | |
5431 | // apply | |
5432 | in->pop_and_dirty_projected_inode(mdr->ls); | |
5433 | ||
5434 | mdr->apply(); | |
5435 | ||
11fdf7f2 | 5436 | get_mds()->balancer->hit_inode(in, META_POP_IWR); |
7c673cae FG |
5437 | |
5438 | server->respond_to_request(mdr, 0); | |
5439 | } | |
5440 | }; | |
5441 | ||
5442 | void Server::handle_client_setxattr(MDRequestRef& mdr) | |
5443 | { | |
11fdf7f2 | 5444 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae | 5445 | string name(req->get_path2()); |
11fdf7f2 | 5446 | MutationImpl::LockOpVec lov; |
7c673cae FG |
5447 | CInode *cur; |
5448 | ||
5449 | file_layout_t *dir_layout = NULL; | |
5450 | if (name.compare(0, 15, "ceph.dir.layout") == 0) | |
11fdf7f2 | 5451 | cur = rdlock_path_pin_ref(mdr, 0, lov, true, false, &dir_layout); |
7c673cae | 5452 | else |
11fdf7f2 | 5453 | cur = rdlock_path_pin_ref(mdr, 0, lov, true); |
7c673cae FG |
5454 | if (!cur) |
5455 | return; | |
5456 | ||
5457 | if (mdr->snapid != CEPH_NOSNAP) { | |
5458 | respond_to_request(mdr, -EROFS); | |
5459 | return; | |
5460 | } | |
5461 | ||
5462 | int flags = req->head.args.setxattr.flags; | |
5463 | ||
5464 | // magic ceph.* namespace? | |
5465 | if (name.compare(0, 5, "ceph.") == 0) { | |
11fdf7f2 | 5466 | handle_set_vxattr(mdr, cur, dir_layout, lov); |
7c673cae FG |
5467 | return; |
5468 | } | |
5469 | ||
11fdf7f2 TL |
5470 | lov.add_xlock(&cur->xattrlock); |
5471 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5472 | return; |
5473 | ||
5474 | if (!check_access(mdr, cur, MAY_WRITE)) | |
5475 | return; | |
5476 | ||
94b18763 | 5477 | auto pxattrs = cur->get_projected_xattrs(); |
7c673cae FG |
5478 | size_t len = req->get_data().length(); |
5479 | size_t inc = len + name.length(); | |
5480 | ||
5481 | // check xattrs kv pairs size | |
5482 | size_t cur_xattrs_size = 0; | |
5483 | for (const auto& p : *pxattrs) { | |
11fdf7f2 | 5484 | if ((flags & CEPH_XATTR_REPLACE) && (name.compare(p.first) == 0)) { |
7c673cae FG |
5485 | continue; |
5486 | } | |
5487 | cur_xattrs_size += p.first.length() + p.second.length(); | |
5488 | } | |
5489 | ||
11fdf7f2 | 5490 | if (((cur_xattrs_size + inc) > g_conf()->mds_max_xattr_pairs_size)) { |
7c673cae FG |
5491 | dout(10) << "xattr kv pairs size too big. cur_xattrs_size " |
5492 | << cur_xattrs_size << ", inc " << inc << dendl; | |
5493 | respond_to_request(mdr, -ENOSPC); | |
5494 | return; | |
5495 | } | |
5496 | ||
11fdf7f2 | 5497 | if ((flags & CEPH_XATTR_CREATE) && pxattrs->count(mempool::mds_co::string(name))) { |
7c673cae FG |
5498 | dout(10) << "setxattr '" << name << "' XATTR_CREATE and EEXIST on " << *cur << dendl; |
5499 | respond_to_request(mdr, -EEXIST); | |
5500 | return; | |
5501 | } | |
11fdf7f2 | 5502 | if ((flags & CEPH_XATTR_REPLACE) && !pxattrs->count(mempool::mds_co::string(name))) { |
7c673cae FG |
5503 | dout(10) << "setxattr '" << name << "' XATTR_REPLACE and ENODATA on " << *cur << dendl; |
5504 | respond_to_request(mdr, -ENODATA); | |
5505 | return; | |
5506 | } | |
5507 | ||
5508 | dout(10) << "setxattr '" << name << "' len " << len << " on " << *cur << dendl; | |
5509 | ||
5510 | // project update | |
94b18763 FG |
5511 | auto &pi = cur->project_inode(true); |
5512 | pi.inode.version = cur->pre_dirty(); | |
91327a77 AA |
5513 | pi.inode.ctime = mdr->get_op_stamp(); |
5514 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
5515 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 FG |
5516 | pi.inode.change_attr++; |
5517 | pi.inode.xattr_version++; | |
5518 | auto &px = *pi.xattrs; | |
5519 | if ((flags & CEPH_XATTR_REMOVE)) { | |
11fdf7f2 | 5520 | px.erase(mempool::mds_co::string(name)); |
94b18763 FG |
5521 | } else { |
5522 | bufferptr b = buffer::create(len); | |
7c673cae | 5523 | if (len) |
94b18763 | 5524 | req->get_data().copy(0, len, b.c_str()); |
11fdf7f2 | 5525 | auto em = px.emplace(std::piecewise_construct, std::forward_as_tuple(mempool::mds_co::string(name)), std::forward_as_tuple(b)); |
94b18763 FG |
5526 | if (!em.second) |
5527 | em.first->second = b; | |
7c673cae FG |
5528 | } |
5529 | ||
5530 | // log + wait | |
5531 | mdr->ls = mdlog->get_current_segment(); | |
5532 | EUpdate *le = new EUpdate(mdlog, "setxattr"); | |
5533 | mdlog->start_entry(le); | |
5534 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5535 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
5536 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
5537 | ||
5538 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur)); | |
5539 | } | |
5540 | ||
5541 | void Server::handle_client_removexattr(MDRequestRef& mdr) | |
5542 | { | |
11fdf7f2 | 5543 | const MClientRequest::const_ref &req = mdr->client_request; |
94b18763 | 5544 | std::string name(req->get_path2()); |
11fdf7f2 TL |
5545 | |
5546 | MutationImpl::LockOpVec lov; | |
5547 | file_layout_t *dir_layout = nullptr; | |
7c673cae FG |
5548 | CInode *cur; |
5549 | if (name == "ceph.dir.layout") | |
11fdf7f2 | 5550 | cur = rdlock_path_pin_ref(mdr, 0, lov, true, false, &dir_layout); |
7c673cae | 5551 | else |
11fdf7f2 | 5552 | cur = rdlock_path_pin_ref(mdr, 0, lov, true); |
7c673cae FG |
5553 | if (!cur) |
5554 | return; | |
5555 | ||
5556 | if (mdr->snapid != CEPH_NOSNAP) { | |
5557 | respond_to_request(mdr, -EROFS); | |
5558 | return; | |
5559 | } | |
5560 | ||
5561 | if (name.compare(0, 5, "ceph.") == 0) { | |
11fdf7f2 | 5562 | handle_remove_vxattr(mdr, cur, dir_layout, lov); |
7c673cae FG |
5563 | return; |
5564 | } | |
5565 | ||
11fdf7f2 TL |
5566 | lov.add_xlock(&cur->xattrlock); |
5567 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5568 | return; |
5569 | ||
94b18763 | 5570 | auto pxattrs = cur->get_projected_xattrs(); |
11fdf7f2 | 5571 | if (pxattrs->count(mempool::mds_co::string(name)) == 0) { |
7c673cae FG |
5572 | dout(10) << "removexattr '" << name << "' and ENODATA on " << *cur << dendl; |
5573 | respond_to_request(mdr, -ENODATA); | |
5574 | return; | |
5575 | } | |
5576 | ||
5577 | dout(10) << "removexattr '" << name << "' on " << *cur << dendl; | |
5578 | ||
5579 | // project update | |
94b18763 FG |
5580 | auto &pi = cur->project_inode(true); |
5581 | auto &px = *pi.xattrs; | |
5582 | pi.inode.version = cur->pre_dirty(); | |
91327a77 AA |
5583 | pi.inode.ctime = mdr->get_op_stamp(); |
5584 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
5585 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 FG |
5586 | pi.inode.change_attr++; |
5587 | pi.inode.xattr_version++; | |
11fdf7f2 | 5588 | px.erase(mempool::mds_co::string(name)); |
7c673cae FG |
5589 | |
5590 | // log + wait | |
5591 | mdr->ls = mdlog->get_current_segment(); | |
5592 | EUpdate *le = new EUpdate(mdlog, "removexattr"); | |
5593 | mdlog->start_entry(le); | |
5594 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5595 | mdcache->predirty_journal_parents(mdr, &le->metablob, cur, 0, PREDIRTY_PRIMARY); | |
5596 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, cur); | |
5597 | ||
5598 | journal_and_reply(mdr, cur, 0, le, new C_MDS_inode_update_finish(this, mdr, cur)); | |
5599 | } | |
5600 | ||
5601 | ||
5602 | // ================================================================= | |
5603 | // DIRECTORY and NAMESPACE OPS | |
5604 | ||
5605 | ||
5606 | // ------------------------------------------------ | |
5607 | ||
5608 | // MKNOD | |
5609 | ||
5610 | class C_MDS_mknod_finish : public ServerLogContext { | |
5611 | CDentry *dn; | |
5612 | CInode *newi; | |
5613 | public: | |
5614 | C_MDS_mknod_finish(Server *s, MDRequestRef& r, CDentry *d, CInode *ni) : | |
5615 | ServerLogContext(s, r), dn(d), newi(ni) {} | |
5616 | void finish(int r) override { | |
11fdf7f2 | 5617 | ceph_assert(r == 0); |
7c673cae FG |
5618 | |
5619 | // link the inode | |
5620 | dn->pop_projected_linkage(); | |
5621 | ||
5622 | // be a bit hacky with the inode version, here.. we decrement it | |
5623 | // just to keep mark_dirty() happen. (we didn't bother projecting | |
5624 | // a new version of hte inode since it's just been created) | |
5625 | newi->inode.version--; | |
5626 | newi->mark_dirty(newi->inode.version + 1, mdr->ls); | |
28e407b8 | 5627 | newi->mark_dirty_parent(mdr->ls, true); |
7c673cae FG |
5628 | |
5629 | // mkdir? | |
5630 | if (newi->inode.is_dir()) { | |
5631 | CDir *dir = newi->get_dirfrag(frag_t()); | |
11fdf7f2 | 5632 | ceph_assert(dir); |
7c673cae FG |
5633 | dir->fnode.version--; |
5634 | dir->mark_dirty(dir->fnode.version + 1, mdr->ls); | |
5635 | dir->mark_new(mdr->ls); | |
5636 | } | |
5637 | ||
5638 | mdr->apply(); | |
5639 | ||
5640 | MDRequestRef null_ref; | |
5641 | get_mds()->mdcache->send_dentry_link(dn, null_ref); | |
5642 | ||
5643 | if (newi->inode.is_file()) | |
5644 | get_mds()->locker->share_inode_max_size(newi); | |
5645 | ||
5646 | // hit pop | |
11fdf7f2 | 5647 | get_mds()->balancer->hit_inode(newi, META_POP_IWR); |
7c673cae FG |
5648 | |
5649 | // reply | |
5650 | server->respond_to_request(mdr, 0); | |
5651 | } | |
5652 | }; | |
5653 | ||
5654 | ||
5655 | void Server::handle_client_mknod(MDRequestRef& mdr) | |
5656 | { | |
11fdf7f2 | 5657 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae | 5658 | client_t client = mdr->get_client(); |
11fdf7f2 TL |
5659 | MutationImpl::LockOpVec lov; |
5660 | file_layout_t *dir_layout = nullptr; | |
5661 | CDentry *dn = rdlock_path_xlock_dentry(mdr, 0, lov, false, false, false, | |
7c673cae FG |
5662 | &dir_layout); |
5663 | if (!dn) return; | |
5664 | if (mdr->snapid != CEPH_NOSNAP) { | |
5665 | respond_to_request(mdr, -EROFS); | |
5666 | return; | |
5667 | } | |
5668 | CInode *diri = dn->get_dir()->get_inode(); | |
11fdf7f2 TL |
5669 | lov.add_rdlock(&diri->authlock); |
5670 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5671 | return; |
5672 | ||
5673 | if (!check_access(mdr, diri, MAY_WRITE)) | |
5674 | return; | |
5675 | ||
5676 | if (!check_fragment_space(mdr, dn->get_dir())) | |
5677 | return; | |
5678 | ||
5679 | unsigned mode = req->head.args.mknod.mode; | |
5680 | if ((mode & S_IFMT) == 0) | |
5681 | mode |= S_IFREG; | |
5682 | ||
5683 | // set layout | |
5684 | file_layout_t layout; | |
5685 | if (dir_layout && S_ISREG(mode)) | |
5686 | layout = *dir_layout; | |
5687 | else | |
5688 | layout = mdcache->default_file_layout; | |
5689 | ||
11fdf7f2 TL |
5690 | CInode *newi = prepare_new_inode(mdr, dn->get_dir(), inodeno_t(req->head.ino), mode, &layout); |
5691 | ceph_assert(newi); | |
7c673cae FG |
5692 | |
5693 | dn->push_projected_linkage(newi); | |
5694 | ||
5695 | newi->inode.rdev = req->head.args.mknod.rdev; | |
5696 | newi->inode.version = dn->pre_dirty(); | |
5697 | newi->inode.rstat.rfiles = 1; | |
5698 | if (layout.pool_id != mdcache->default_file_layout.pool_id) | |
5699 | newi->inode.add_old_pool(mdcache->default_file_layout.pool_id); | |
5700 | newi->inode.update_backtrace(); | |
5701 | ||
11fdf7f2 TL |
5702 | snapid_t follows = mdcache->get_global_snaprealm()->get_newest_seq(); |
5703 | SnapRealm *realm = dn->get_dir()->inode->find_snaprealm(); | |
5704 | ceph_assert(follows >= realm->get_newest_seq()); | |
5705 | ||
7c673cae FG |
5706 | // if the client created a _regular_ file via MKNOD, it's highly likely they'll |
5707 | // want to write to it (e.g., if they are reexporting NFS) | |
5708 | if (S_ISREG(newi->inode.mode)) { | |
7c673cae FG |
5709 | // issue a cap on the file |
5710 | int cmode = CEPH_FILE_MODE_RDWR; | |
5711 | Capability *cap = mds->locker->issue_new_caps(newi, cmode, mdr->session, realm, req->is_replay()); | |
5712 | if (cap) { | |
5713 | cap->set_wanted(0); | |
5714 | ||
5715 | // put locks in excl mode | |
5716 | newi->filelock.set_state(LOCK_EXCL); | |
5717 | newi->authlock.set_state(LOCK_EXCL); | |
5718 | newi->xattrlock.set_state(LOCK_EXCL); | |
a8e16298 TL |
5719 | |
5720 | dout(15) << " setting a client_range too, since this is a regular file" << dendl; | |
5721 | newi->inode.client_ranges[client].range.first = 0; | |
5722 | newi->inode.client_ranges[client].range.last = newi->inode.get_layout_size_increment(); | |
5723 | newi->inode.client_ranges[client].follows = follows; | |
5724 | cap->mark_clientwriteable(); | |
7c673cae FG |
5725 | } |
5726 | } | |
5727 | ||
11fdf7f2 | 5728 | ceph_assert(dn->first == follows + 1); |
7c673cae FG |
5729 | newi->first = dn->first; |
5730 | ||
5731 | dout(10) << "mknod mode " << newi->inode.mode << " rdev " << newi->inode.rdev << dendl; | |
5732 | ||
5733 | // prepare finisher | |
5734 | mdr->ls = mdlog->get_current_segment(); | |
5735 | EUpdate *le = new EUpdate(mdlog, "mknod"); | |
5736 | mdlog->start_entry(le); | |
5737 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5738 | journal_allocated_inos(mdr, &le->metablob); | |
5739 | ||
5740 | mdcache->predirty_journal_parents(mdr, &le->metablob, newi, dn->get_dir(), | |
5741 | PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
5742 | le->metablob.add_primary_dentry(dn, newi, true, true, true); | |
5743 | ||
5744 | journal_and_reply(mdr, newi, dn, le, new C_MDS_mknod_finish(this, mdr, dn, newi)); | |
5745 | } | |
5746 | ||
5747 | ||
5748 | ||
5749 | // MKDIR | |
5750 | /* This function takes responsibility for the passed mdr*/ | |
5751 | void Server::handle_client_mkdir(MDRequestRef& mdr) | |
5752 | { | |
11fdf7f2 | 5753 | const MClientRequest::const_ref &req = mdr->client_request; |
91327a77 AA |
5754 | if (req->get_filepath().is_last_dot_or_dotdot()) { |
5755 | respond_to_request(mdr, -EEXIST); | |
5756 | return; | |
5757 | } | |
5758 | ||
11fdf7f2 TL |
5759 | MutationImpl::LockOpVec lov; |
5760 | CDentry *dn = rdlock_path_xlock_dentry(mdr, 0, lov, false, false, false); | |
7c673cae FG |
5761 | if (!dn) return; |
5762 | if (mdr->snapid != CEPH_NOSNAP) { | |
5763 | respond_to_request(mdr, -EROFS); | |
5764 | return; | |
5765 | } | |
5766 | CDir *dir = dn->get_dir(); | |
5767 | CInode *diri = dir->get_inode(); | |
11fdf7f2 TL |
5768 | lov.add_rdlock(&diri->authlock); |
5769 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5770 | return; |
5771 | ||
5772 | // mkdir check access | |
5773 | if (!check_access(mdr, diri, MAY_WRITE)) | |
5774 | return; | |
5775 | ||
5776 | if (!check_fragment_space(mdr, dir)) | |
5777 | return; | |
5778 | ||
5779 | // new inode | |
7c673cae FG |
5780 | unsigned mode = req->head.args.mkdir.mode; |
5781 | mode &= ~S_IFMT; | |
5782 | mode |= S_IFDIR; | |
5783 | CInode *newi = prepare_new_inode(mdr, dn->get_dir(), inodeno_t(req->head.ino), mode); | |
11fdf7f2 | 5784 | ceph_assert(newi); |
7c673cae FG |
5785 | |
5786 | // it's a directory. | |
5787 | dn->push_projected_linkage(newi); | |
5788 | ||
5789 | newi->inode.version = dn->pre_dirty(); | |
5790 | newi->inode.rstat.rsubdirs = 1; | |
5791 | newi->inode.update_backtrace(); | |
5792 | ||
11fdf7f2 TL |
5793 | snapid_t follows = mdcache->get_global_snaprealm()->get_newest_seq(); |
5794 | SnapRealm *realm = dn->get_dir()->inode->find_snaprealm(); | |
5795 | ceph_assert(follows >= realm->get_newest_seq()); | |
5796 | ||
7c673cae | 5797 | dout(12) << " follows " << follows << dendl; |
11fdf7f2 | 5798 | ceph_assert(dn->first == follows + 1); |
7c673cae FG |
5799 | newi->first = dn->first; |
5800 | ||
5801 | // ...and that new dir is empty. | |
5802 | CDir *newdir = newi->get_or_open_dirfrag(mdcache, frag_t()); | |
5803 | newdir->state_set(CDir::STATE_CREATING); | |
5804 | newdir->mark_complete(); | |
5805 | newdir->fnode.version = newdir->pre_dirty(); | |
5806 | ||
5807 | // prepare finisher | |
5808 | mdr->ls = mdlog->get_current_segment(); | |
5809 | EUpdate *le = new EUpdate(mdlog, "mkdir"); | |
5810 | mdlog->start_entry(le); | |
5811 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5812 | journal_allocated_inos(mdr, &le->metablob); | |
5813 | mdcache->predirty_journal_parents(mdr, &le->metablob, newi, dn->get_dir(), PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
5814 | le->metablob.add_primary_dentry(dn, newi, true, true); | |
5815 | le->metablob.add_new_dir(newdir); // dirty AND complete AND new | |
5816 | ||
5817 | // issue a cap on the directory | |
5818 | int cmode = CEPH_FILE_MODE_RDWR; | |
5819 | Capability *cap = mds->locker->issue_new_caps(newi, cmode, mdr->session, realm, req->is_replay()); | |
5820 | if (cap) { | |
5821 | cap->set_wanted(0); | |
5822 | ||
5823 | // put locks in excl mode | |
5824 | newi->filelock.set_state(LOCK_EXCL); | |
5825 | newi->authlock.set_state(LOCK_EXCL); | |
5826 | newi->xattrlock.set_state(LOCK_EXCL); | |
5827 | } | |
5828 | ||
5829 | // make sure this inode gets into the journal | |
5830 | le->metablob.add_opened_ino(newi->ino()); | |
7c673cae FG |
5831 | |
5832 | journal_and_reply(mdr, newi, dn, le, new C_MDS_mknod_finish(this, mdr, dn, newi)); | |
5833 | } | |
5834 | ||
5835 | ||
5836 | // SYMLINK | |
5837 | ||
5838 | void Server::handle_client_symlink(MDRequestRef& mdr) | |
5839 | { | |
11fdf7f2 TL |
5840 | const MClientRequest::const_ref &req = mdr->client_request; |
5841 | MutationImpl::LockOpVec lov; | |
5842 | CDentry *dn = rdlock_path_xlock_dentry(mdr, 0, lov, false, false, false); | |
7c673cae FG |
5843 | if (!dn) return; |
5844 | if (mdr->snapid != CEPH_NOSNAP) { | |
5845 | respond_to_request(mdr, -EROFS); | |
5846 | return; | |
5847 | } | |
5848 | CDir *dir = dn->get_dir(); | |
5849 | CInode *diri = dir->get_inode(); | |
11fdf7f2 TL |
5850 | lov.add_rdlock(&diri->authlock); |
5851 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
5852 | return; |
5853 | ||
5854 | if (!check_access(mdr, diri, MAY_WRITE)) | |
5855 | return; | |
5856 | ||
5857 | if (!check_fragment_space(mdr, dir)) | |
5858 | return; | |
5859 | ||
5860 | unsigned mode = S_IFLNK | 0777; | |
5861 | CInode *newi = prepare_new_inode(mdr, dn->get_dir(), inodeno_t(req->head.ino), mode); | |
11fdf7f2 | 5862 | ceph_assert(newi); |
7c673cae FG |
5863 | |
5864 | // it's a symlink | |
5865 | dn->push_projected_linkage(newi); | |
5866 | ||
11fdf7f2 | 5867 | newi->symlink = req->get_path2(); |
7c673cae FG |
5868 | newi->inode.size = newi->symlink.length(); |
5869 | newi->inode.rstat.rbytes = newi->inode.size; | |
5870 | newi->inode.rstat.rfiles = 1; | |
5871 | newi->inode.version = dn->pre_dirty(); | |
5872 | newi->inode.update_backtrace(); | |
5873 | ||
5874 | newi->first = dn->first; | |
5875 | ||
5876 | // prepare finisher | |
5877 | mdr->ls = mdlog->get_current_segment(); | |
5878 | EUpdate *le = new EUpdate(mdlog, "symlink"); | |
5879 | mdlog->start_entry(le); | |
5880 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
5881 | journal_allocated_inos(mdr, &le->metablob); | |
5882 | mdcache->predirty_journal_parents(mdr, &le->metablob, newi, dn->get_dir(), PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
5883 | le->metablob.add_primary_dentry(dn, newi, true, true); | |
5884 | ||
5885 | journal_and_reply(mdr, newi, dn, le, new C_MDS_mknod_finish(this, mdr, dn, newi)); | |
5886 | } | |
5887 | ||
5888 | ||
5889 | ||
5890 | ||
5891 | ||
5892 | // LINK | |
5893 | ||
5894 | void Server::handle_client_link(MDRequestRef& mdr) | |
5895 | { | |
11fdf7f2 | 5896 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
5897 | |
5898 | dout(7) << "handle_client_link " << req->get_filepath() | |
5899 | << " to " << req->get_filepath2() | |
5900 | << dendl; | |
5901 | ||
11fdf7f2 | 5902 | MutationImpl::LockOpVec lov; |
7c673cae | 5903 | |
11fdf7f2 | 5904 | CDentry *dn = rdlock_path_xlock_dentry(mdr, 0, lov, false, false, false); |
7c673cae | 5905 | if (!dn) return; |
11fdf7f2 | 5906 | CInode *targeti = rdlock_path_pin_ref(mdr, 1, lov, false); |
7c673cae FG |
5907 | if (!targeti) return; |
5908 | if (mdr->snapid != CEPH_NOSNAP) { | |
5909 | respond_to_request(mdr, -EROFS); | |
5910 | return; | |
5911 | } | |
5912 | ||
5913 | CDir *dir = dn->get_dir(); | |
5914 | dout(7) << "handle_client_link link " << dn->get_name() << " in " << *dir << dendl; | |
5915 | dout(7) << "target is " << *targeti << dendl; | |
5916 | if (targeti->is_dir()) { | |
181888fb FG |
5917 | // if srcdn is replica, need to make sure its linkage is correct |
5918 | vector<CDentry*>& trace = mdr->dn[1]; | |
5919 | if (trace.empty() || | |
5920 | trace.back()->is_auth() || | |
5921 | trace.back()->lock.can_read(mdr->get_client())) { | |
5922 | dout(7) << "target is a dir, failing..." << dendl; | |
5923 | respond_to_request(mdr, -EINVAL); | |
5924 | return; | |
5925 | } | |
7c673cae FG |
5926 | } |
5927 | ||
11fdf7f2 TL |
5928 | lov.erase_rdlock(&targeti->snaplock); |
5929 | lov.add_xlock(&targeti->snaplock); | |
5930 | lov.add_xlock(&targeti->linklock); | |
7c673cae | 5931 | |
11fdf7f2 | 5932 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
5933 | return; |
5934 | ||
5935 | if ((!mdr->has_more() || mdr->more()->witnessed.empty())) { | |
5936 | if (!check_access(mdr, targeti, MAY_WRITE)) | |
5937 | return; | |
5938 | ||
5939 | if (!check_access(mdr, dir->get_inode(), MAY_WRITE)) | |
5940 | return; | |
5941 | ||
5942 | if (!check_fragment_space(mdr, dir)) | |
5943 | return; | |
5944 | } | |
5945 | ||
5946 | // go! | |
11fdf7f2 | 5947 | ceph_assert(g_conf()->mds_kill_link_at != 1); |
7c673cae FG |
5948 | |
5949 | // local or remote? | |
5950 | if (targeti->is_auth()) | |
5951 | _link_local(mdr, dn, targeti); | |
5952 | else | |
5953 | _link_remote(mdr, true, dn, targeti); | |
5954 | } | |
5955 | ||
5956 | ||
5957 | class C_MDS_link_local_finish : public ServerLogContext { | |
5958 | CDentry *dn; | |
5959 | CInode *targeti; | |
5960 | version_t dnpv; | |
5961 | version_t tipv; | |
11fdf7f2 | 5962 | bool adjust_realm; |
7c673cae FG |
5963 | public: |
5964 | C_MDS_link_local_finish(Server *s, MDRequestRef& r, CDentry *d, CInode *ti, | |
11fdf7f2 | 5965 | version_t dnpv_, version_t tipv_, bool ar) : |
7c673cae | 5966 | ServerLogContext(s, r), dn(d), targeti(ti), |
11fdf7f2 | 5967 | dnpv(dnpv_), tipv(tipv_), adjust_realm(ar) { } |
7c673cae | 5968 | void finish(int r) override { |
11fdf7f2 TL |
5969 | ceph_assert(r == 0); |
5970 | server->_link_local_finish(mdr, dn, targeti, dnpv, tipv, adjust_realm); | |
7c673cae FG |
5971 | } |
5972 | }; | |
5973 | ||
5974 | ||
5975 | void Server::_link_local(MDRequestRef& mdr, CDentry *dn, CInode *targeti) | |
5976 | { | |
5977 | dout(10) << "_link_local " << *dn << " to " << *targeti << dendl; | |
5978 | ||
5979 | mdr->ls = mdlog->get_current_segment(); | |
5980 | ||
5981 | // predirty NEW dentry | |
5982 | version_t dnpv = dn->pre_dirty(); | |
5983 | version_t tipv = targeti->pre_dirty(); | |
5984 | ||
5985 | // project inode update | |
94b18763 FG |
5986 | auto &pi = targeti->project_inode(); |
5987 | pi.inode.nlink++; | |
91327a77 AA |
5988 | pi.inode.ctime = mdr->get_op_stamp(); |
5989 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
5990 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 FG |
5991 | pi.inode.change_attr++; |
5992 | pi.inode.version = tipv; | |
7c673cae | 5993 | |
11fdf7f2 TL |
5994 | bool adjust_realm = false; |
5995 | if (!targeti->is_projected_snaprealm_global()) { | |
5996 | sr_t *newsnap = targeti->project_snaprealm(); | |
5997 | targeti->mark_snaprealm_global(newsnap); | |
5998 | targeti->record_snaprealm_parent_dentry(newsnap, NULL, targeti->get_projected_parent_dn(), true); | |
5999 | adjust_realm = true; | |
6000 | } | |
6001 | ||
7c673cae FG |
6002 | // log + wait |
6003 | EUpdate *le = new EUpdate(mdlog, "link_local"); | |
6004 | mdlog->start_entry(le); | |
6005 | le->metablob.add_client_req(mdr->reqid, mdr->client_request->get_oldest_client_tid()); | |
6006 | mdcache->predirty_journal_parents(mdr, &le->metablob, targeti, dn->get_dir(), PREDIRTY_DIR, 1); // new dn | |
6007 | mdcache->predirty_journal_parents(mdr, &le->metablob, targeti, 0, PREDIRTY_PRIMARY); // targeti | |
6008 | le->metablob.add_remote_dentry(dn, true, targeti->ino(), targeti->d_type()); // new remote | |
6009 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, targeti); | |
6010 | ||
6011 | // do this after predirty_*, to avoid funky extra dnl arg | |
6012 | dn->push_projected_linkage(targeti->ino(), targeti->d_type()); | |
6013 | ||
11fdf7f2 TL |
6014 | journal_and_reply(mdr, targeti, dn, le, |
6015 | new C_MDS_link_local_finish(this, mdr, dn, targeti, dnpv, tipv, adjust_realm)); | |
7c673cae FG |
6016 | } |
6017 | ||
6018 | void Server::_link_local_finish(MDRequestRef& mdr, CDentry *dn, CInode *targeti, | |
11fdf7f2 | 6019 | version_t dnpv, version_t tipv, bool adjust_realm) |
7c673cae FG |
6020 | { |
6021 | dout(10) << "_link_local_finish " << *dn << " to " << *targeti << dendl; | |
6022 | ||
6023 | // link and unlock the NEW dentry | |
31f18b77 FG |
6024 | CDentry::linkage_t *dnl = dn->pop_projected_linkage(); |
6025 | if (!dnl->get_inode()) | |
6026 | dn->link_remote(dnl, targeti); | |
7c673cae FG |
6027 | dn->mark_dirty(dnpv, mdr->ls); |
6028 | ||
6029 | // target inode | |
6030 | targeti->pop_and_dirty_projected_inode(mdr->ls); | |
6031 | ||
6032 | mdr->apply(); | |
6033 | ||
6034 | MDRequestRef null_ref; | |
6035 | mdcache->send_dentry_link(dn, null_ref); | |
6036 | ||
11fdf7f2 TL |
6037 | if (adjust_realm) { |
6038 | int op = CEPH_SNAP_OP_SPLIT; | |
6039 | mds->mdcache->send_snap_update(targeti, 0, op); | |
6040 | mds->mdcache->do_realm_invalidate_and_update_notify(targeti, op); | |
6041 | } | |
6042 | ||
7c673cae | 6043 | // bump target popularity |
11fdf7f2 TL |
6044 | mds->balancer->hit_inode(targeti, META_POP_IWR); |
6045 | mds->balancer->hit_dir(dn->get_dir(), META_POP_IWR); | |
7c673cae FG |
6046 | |
6047 | // reply | |
6048 | respond_to_request(mdr, 0); | |
6049 | } | |
6050 | ||
6051 | ||
6052 | // link / unlink remote | |
6053 | ||
6054 | class C_MDS_link_remote_finish : public ServerLogContext { | |
6055 | bool inc; | |
6056 | CDentry *dn; | |
6057 | CInode *targeti; | |
6058 | version_t dpv; | |
6059 | public: | |
6060 | C_MDS_link_remote_finish(Server *s, MDRequestRef& r, bool i, CDentry *d, CInode *ti) : | |
6061 | ServerLogContext(s, r), inc(i), dn(d), targeti(ti), | |
6062 | dpv(d->get_projected_version()) {} | |
6063 | void finish(int r) override { | |
11fdf7f2 | 6064 | ceph_assert(r == 0); |
7c673cae FG |
6065 | server->_link_remote_finish(mdr, inc, dn, targeti, dpv); |
6066 | } | |
6067 | }; | |
6068 | ||
6069 | void Server::_link_remote(MDRequestRef& mdr, bool inc, CDentry *dn, CInode *targeti) | |
6070 | { | |
6071 | dout(10) << "_link_remote " | |
6072 | << (inc ? "link ":"unlink ") | |
6073 | << *dn << " to " << *targeti << dendl; | |
6074 | ||
6075 | // 1. send LinkPrepare to dest (journal nlink++ prepare) | |
6076 | mds_rank_t linkauth = targeti->authority().first; | |
6077 | if (mdr->more()->witnessed.count(linkauth) == 0) { | |
6078 | if (mds->is_cluster_degraded() && | |
6079 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(linkauth)) { | |
6080 | dout(10) << " targeti auth mds." << linkauth << " is not active" << dendl; | |
6081 | if (mdr->more()->waiting_on_slave.empty()) | |
6082 | mds->wait_for_active_peer(linkauth, new C_MDS_RetryRequest(mdcache, mdr)); | |
6083 | return; | |
6084 | } | |
6085 | ||
6086 | dout(10) << " targeti auth must prepare nlink++/--" << dendl; | |
6087 | int op; | |
6088 | if (inc) | |
6089 | op = MMDSSlaveRequest::OP_LINKPREP; | |
6090 | else | |
6091 | op = MMDSSlaveRequest::OP_UNLINKPREP; | |
11fdf7f2 | 6092 | auto req = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, op); |
7c673cae FG |
6093 | targeti->set_object_info(req->get_object_info()); |
6094 | req->op_stamp = mdr->get_op_stamp(); | |
11fdf7f2 TL |
6095 | if (auto& desti_srnode = mdr->more()->desti_srnode) |
6096 | encode(*desti_srnode, req->desti_snapbl); | |
7c673cae FG |
6097 | mds->send_message_mds(req, linkauth); |
6098 | ||
11fdf7f2 | 6099 | ceph_assert(mdr->more()->waiting_on_slave.count(linkauth) == 0); |
7c673cae FG |
6100 | mdr->more()->waiting_on_slave.insert(linkauth); |
6101 | return; | |
6102 | } | |
6103 | dout(10) << " targeti auth has prepared nlink++/--" << dendl; | |
6104 | ||
11fdf7f2 TL |
6105 | ceph_assert(g_conf()->mds_kill_link_at != 2); |
6106 | ||
6107 | if (auto& desti_srnode = mdr->more()->desti_srnode) { | |
6108 | delete desti_srnode; | |
6109 | desti_srnode = NULL; | |
6110 | } | |
7c673cae FG |
6111 | |
6112 | mdr->set_mds_stamp(ceph_clock_now()); | |
6113 | ||
6114 | // add to event | |
6115 | mdr->ls = mdlog->get_current_segment(); | |
6116 | EUpdate *le = new EUpdate(mdlog, inc ? "link_remote":"unlink_remote"); | |
6117 | mdlog->start_entry(le); | |
6118 | le->metablob.add_client_req(mdr->reqid, mdr->client_request->get_oldest_client_tid()); | |
6119 | if (!mdr->more()->witnessed.empty()) { | |
6120 | dout(20) << " noting uncommitted_slaves " << mdr->more()->witnessed << dendl; | |
6121 | le->reqid = mdr->reqid; | |
6122 | le->had_slaves = true; | |
6123 | mdcache->add_uncommitted_master(mdr->reqid, mdr->ls, mdr->more()->witnessed); | |
6124 | } | |
6125 | ||
6126 | if (inc) { | |
6127 | dn->pre_dirty(); | |
6128 | mdcache->predirty_journal_parents(mdr, &le->metablob, targeti, dn->get_dir(), PREDIRTY_DIR, 1); | |
6129 | le->metablob.add_remote_dentry(dn, true, targeti->ino(), targeti->d_type()); // new remote | |
6130 | dn->push_projected_linkage(targeti->ino(), targeti->d_type()); | |
6131 | } else { | |
6132 | dn->pre_dirty(); | |
6133 | mdcache->predirty_journal_parents(mdr, &le->metablob, targeti, dn->get_dir(), PREDIRTY_DIR, -1); | |
6134 | mdcache->journal_cow_dentry(mdr.get(), &le->metablob, dn); | |
6135 | le->metablob.add_null_dentry(dn, true); | |
31f18b77 | 6136 | dn->push_projected_linkage(); |
7c673cae FG |
6137 | } |
6138 | ||
6139 | journal_and_reply(mdr, targeti, dn, le, new C_MDS_link_remote_finish(this, mdr, inc, dn, targeti)); | |
6140 | } | |
6141 | ||
6142 | void Server::_link_remote_finish(MDRequestRef& mdr, bool inc, | |
6143 | CDentry *dn, CInode *targeti, | |
6144 | version_t dpv) | |
6145 | { | |
6146 | dout(10) << "_link_remote_finish " | |
6147 | << (inc ? "link ":"unlink ") | |
6148 | << *dn << " to " << *targeti << dendl; | |
6149 | ||
11fdf7f2 | 6150 | ceph_assert(g_conf()->mds_kill_link_at != 3); |
7c673cae FG |
6151 | |
6152 | if (!mdr->more()->witnessed.empty()) | |
6153 | mdcache->logged_master_update(mdr->reqid); | |
6154 | ||
6155 | if (inc) { | |
6156 | // link the new dentry | |
31f18b77 FG |
6157 | CDentry::linkage_t *dnl = dn->pop_projected_linkage(); |
6158 | if (!dnl->get_inode()) | |
6159 | dn->link_remote(dnl, targeti); | |
7c673cae FG |
6160 | dn->mark_dirty(dpv, mdr->ls); |
6161 | } else { | |
6162 | // unlink main dentry | |
6163 | dn->get_dir()->unlink_inode(dn); | |
31f18b77 | 6164 | dn->pop_projected_linkage(); |
7c673cae FG |
6165 | dn->mark_dirty(dn->get_projected_version(), mdr->ls); // dirty old dentry |
6166 | } | |
6167 | ||
6168 | mdr->apply(); | |
6169 | ||
6170 | MDRequestRef null_ref; | |
6171 | if (inc) | |
6172 | mdcache->send_dentry_link(dn, null_ref); | |
6173 | else | |
6174 | mdcache->send_dentry_unlink(dn, NULL, null_ref); | |
6175 | ||
6176 | // bump target popularity | |
11fdf7f2 TL |
6177 | mds->balancer->hit_inode(targeti, META_POP_IWR); |
6178 | mds->balancer->hit_dir(dn->get_dir(), META_POP_IWR); | |
7c673cae FG |
6179 | |
6180 | // reply | |
6181 | respond_to_request(mdr, 0); | |
6182 | ||
6183 | if (!inc) | |
6184 | // removing a new dn? | |
6185 | dn->get_dir()->try_remove_unlinked_dn(dn); | |
6186 | } | |
6187 | ||
6188 | ||
6189 | // remote linking/unlinking | |
6190 | ||
6191 | class C_MDS_SlaveLinkPrep : public ServerLogContext { | |
6192 | CInode *targeti; | |
11fdf7f2 | 6193 | bool adjust_realm; |
7c673cae | 6194 | public: |
11fdf7f2 TL |
6195 | C_MDS_SlaveLinkPrep(Server *s, MDRequestRef& r, CInode *t, bool ar) : |
6196 | ServerLogContext(s, r), targeti(t), adjust_realm(ar) { } | |
7c673cae | 6197 | void finish(int r) override { |
11fdf7f2 TL |
6198 | ceph_assert(r == 0); |
6199 | server->_logged_slave_link(mdr, targeti, adjust_realm); | |
7c673cae FG |
6200 | } |
6201 | }; | |
6202 | ||
6203 | class C_MDS_SlaveLinkCommit : public ServerContext { | |
6204 | MDRequestRef mdr; | |
6205 | CInode *targeti; | |
6206 | public: | |
6207 | C_MDS_SlaveLinkCommit(Server *s, MDRequestRef& r, CInode *t) : | |
6208 | ServerContext(s), mdr(r), targeti(t) { } | |
6209 | void finish(int r) override { | |
6210 | server->_commit_slave_link(mdr, r, targeti); | |
6211 | } | |
6212 | }; | |
6213 | ||
7c673cae FG |
6214 | void Server::handle_slave_link_prep(MDRequestRef& mdr) |
6215 | { | |
6216 | dout(10) << "handle_slave_link_prep " << *mdr | |
6217 | << " on " << mdr->slave_request->get_object_info() | |
6218 | << dendl; | |
6219 | ||
11fdf7f2 | 6220 | ceph_assert(g_conf()->mds_kill_link_at != 4); |
7c673cae FG |
6221 | |
6222 | CInode *targeti = mdcache->get_inode(mdr->slave_request->get_object_info().ino); | |
11fdf7f2 | 6223 | ceph_assert(targeti); |
7c673cae FG |
6224 | dout(10) << "targeti " << *targeti << dendl; |
6225 | CDentry *dn = targeti->get_parent_dn(); | |
6226 | CDentry::linkage_t *dnl = dn->get_linkage(); | |
11fdf7f2 | 6227 | ceph_assert(dnl->is_primary()); |
7c673cae FG |
6228 | |
6229 | mdr->set_op_stamp(mdr->slave_request->op_stamp); | |
6230 | ||
6231 | mdr->auth_pin(targeti); | |
6232 | ||
6233 | //ceph_abort(); // test hack: make sure master can handle a slave that fails to prepare... | |
11fdf7f2 | 6234 | ceph_assert(g_conf()->mds_kill_link_at != 5); |
7c673cae FG |
6235 | |
6236 | // journal it | |
6237 | mdr->ls = mdlog->get_current_segment(); | |
6238 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_link_prep", mdr->reqid, mdr->slave_to_mds, | |
6239 | ESlaveUpdate::OP_PREPARE, ESlaveUpdate::LINK); | |
6240 | mdlog->start_entry(le); | |
6241 | ||
94b18763 | 6242 | auto &pi = dnl->get_inode()->project_inode(); |
7c673cae FG |
6243 | |
6244 | // update journaled target inode | |
6245 | bool inc; | |
11fdf7f2 TL |
6246 | bool adjust_realm = false; |
6247 | bool realm_projected = false; | |
7c673cae FG |
6248 | if (mdr->slave_request->get_op() == MMDSSlaveRequest::OP_LINKPREP) { |
6249 | inc = true; | |
94b18763 | 6250 | pi.inode.nlink++; |
11fdf7f2 TL |
6251 | if (!targeti->is_projected_snaprealm_global()) { |
6252 | sr_t *newsnap = targeti->project_snaprealm(); | |
6253 | targeti->mark_snaprealm_global(newsnap); | |
6254 | targeti->record_snaprealm_parent_dentry(newsnap, NULL, targeti->get_projected_parent_dn(), true); | |
6255 | adjust_realm = true; | |
6256 | realm_projected = true; | |
6257 | } | |
7c673cae FG |
6258 | } else { |
6259 | inc = false; | |
94b18763 | 6260 | pi.inode.nlink--; |
11fdf7f2 TL |
6261 | if (targeti->is_projected_snaprealm_global()) { |
6262 | ceph_assert(mdr->slave_request->desti_snapbl.length()); | |
6263 | auto p = mdr->slave_request->desti_snapbl.cbegin(); | |
6264 | ||
6265 | sr_t *newsnap = targeti->project_snaprealm(); | |
6266 | decode(*newsnap, p); | |
6267 | ||
6268 | if (pi.inode.nlink == 0) | |
6269 | ceph_assert(!newsnap->is_parent_global()); | |
6270 | ||
6271 | realm_projected = true; | |
6272 | } else { | |
6273 | ceph_assert(mdr->slave_request->desti_snapbl.length() == 0); | |
6274 | } | |
7c673cae FG |
6275 | } |
6276 | ||
6277 | link_rollback rollback; | |
6278 | rollback.reqid = mdr->reqid; | |
6279 | rollback.ino = targeti->ino(); | |
6280 | rollback.old_ctime = targeti->inode.ctime; // we hold versionlock xlock; no concorrent projections | |
6281 | const fnode_t *pf = targeti->get_parent_dn()->get_dir()->get_projected_fnode(); | |
6282 | rollback.old_dir_mtime = pf->fragstat.mtime; | |
6283 | rollback.old_dir_rctime = pf->rstat.rctime; | |
6284 | rollback.was_inc = inc; | |
11fdf7f2 TL |
6285 | if (realm_projected) { |
6286 | if (targeti->snaprealm) { | |
6287 | encode(true, rollback.snapbl); | |
6288 | targeti->encode_snap_blob(rollback.snapbl); | |
6289 | } else { | |
6290 | encode(false, rollback.snapbl); | |
6291 | } | |
6292 | } | |
6293 | encode(rollback, le->rollback); | |
7c673cae FG |
6294 | mdr->more()->rollback_bl = le->rollback; |
6295 | ||
94b18763 FG |
6296 | pi.inode.ctime = mdr->get_op_stamp(); |
6297 | pi.inode.version = targeti->pre_dirty(); | |
7c673cae | 6298 | |
94b18763 | 6299 | dout(10) << " projected inode " << pi.inode.ino << " v " << pi.inode.version << dendl; |
7c673cae FG |
6300 | |
6301 | // commit case | |
6302 | mdcache->predirty_journal_parents(mdr, &le->commit, dnl->get_inode(), 0, PREDIRTY_SHALLOW|PREDIRTY_PRIMARY); | |
6303 | mdcache->journal_dirty_inode(mdr.get(), &le->commit, targeti); | |
6304 | ||
6305 | // set up commit waiter | |
6306 | mdr->more()->slave_commit = new C_MDS_SlaveLinkCommit(this, mdr, targeti); | |
6307 | ||
6308 | mdr->more()->slave_update_journaled = true; | |
11fdf7f2 | 6309 | submit_mdlog_entry(le, new C_MDS_SlaveLinkPrep(this, mdr, targeti, adjust_realm), |
7c673cae FG |
6310 | mdr, __func__); |
6311 | mdlog->flush(); | |
6312 | } | |
6313 | ||
11fdf7f2 | 6314 | void Server::_logged_slave_link(MDRequestRef& mdr, CInode *targeti, bool adjust_realm) |
7c673cae FG |
6315 | { |
6316 | dout(10) << "_logged_slave_link " << *mdr | |
6317 | << " " << *targeti << dendl; | |
6318 | ||
11fdf7f2 | 6319 | ceph_assert(g_conf()->mds_kill_link_at != 6); |
7c673cae FG |
6320 | |
6321 | // update the target | |
6322 | targeti->pop_and_dirty_projected_inode(mdr->ls); | |
6323 | mdr->apply(); | |
6324 | ||
6325 | // hit pop | |
11fdf7f2 | 6326 | mds->balancer->hit_inode(targeti, META_POP_IWR); |
7c673cae FG |
6327 | |
6328 | // done. | |
91327a77 | 6329 | mdr->reset_slave_request(); |
7c673cae | 6330 | |
11fdf7f2 TL |
6331 | if (adjust_realm) { |
6332 | int op = CEPH_SNAP_OP_SPLIT; | |
6333 | mds->mdcache->send_snap_update(targeti, 0, op); | |
6334 | mds->mdcache->do_realm_invalidate_and_update_notify(targeti, op); | |
6335 | } | |
6336 | ||
7c673cae FG |
6337 | // ack |
6338 | if (!mdr->aborted) { | |
11fdf7f2 | 6339 | auto reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_LINKPREPACK); |
7c673cae FG |
6340 | mds->send_message_mds(reply, mdr->slave_to_mds); |
6341 | } else { | |
6342 | dout(10) << " abort flag set, finishing" << dendl; | |
6343 | mdcache->request_finish(mdr); | |
6344 | } | |
6345 | } | |
6346 | ||
6347 | ||
6348 | struct C_MDS_CommittedSlave : public ServerLogContext { | |
6349 | C_MDS_CommittedSlave(Server *s, MDRequestRef& m) : ServerLogContext(s, m) {} | |
6350 | void finish(int r) override { | |
6351 | server->_committed_slave(mdr); | |
6352 | } | |
6353 | }; | |
6354 | ||
6355 | void Server::_commit_slave_link(MDRequestRef& mdr, int r, CInode *targeti) | |
6356 | { | |
6357 | dout(10) << "_commit_slave_link " << *mdr | |
6358 | << " r=" << r | |
6359 | << " " << *targeti << dendl; | |
6360 | ||
11fdf7f2 | 6361 | ceph_assert(g_conf()->mds_kill_link_at != 7); |
7c673cae FG |
6362 | |
6363 | if (r == 0) { | |
6364 | // drop our pins, etc. | |
6365 | mdr->cleanup(); | |
6366 | ||
6367 | // write a commit to the journal | |
6368 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_link_commit", mdr->reqid, mdr->slave_to_mds, | |
6369 | ESlaveUpdate::OP_COMMIT, ESlaveUpdate::LINK); | |
6370 | mdlog->start_entry(le); | |
6371 | submit_mdlog_entry(le, new C_MDS_CommittedSlave(this, mdr), mdr, __func__); | |
6372 | mdlog->flush(); | |
6373 | } else { | |
6374 | do_link_rollback(mdr->more()->rollback_bl, mdr->slave_to_mds, mdr); | |
6375 | } | |
6376 | } | |
6377 | ||
6378 | void Server::_committed_slave(MDRequestRef& mdr) | |
6379 | { | |
6380 | dout(10) << "_committed_slave " << *mdr << dendl; | |
6381 | ||
11fdf7f2 | 6382 | ceph_assert(g_conf()->mds_kill_link_at != 8); |
7c673cae | 6383 | |
11fdf7f2 | 6384 | auto req = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_COMMITTED); |
7c673cae FG |
6385 | mds->send_message_mds(req, mdr->slave_to_mds); |
6386 | mdcache->request_finish(mdr); | |
6387 | } | |
6388 | ||
6389 | struct C_MDS_LoggedLinkRollback : public ServerLogContext { | |
6390 | MutationRef mut; | |
11fdf7f2 TL |
6391 | map<client_t,MClientSnap::ref> splits; |
6392 | C_MDS_LoggedLinkRollback(Server *s, MutationRef& m, MDRequestRef& r, | |
6393 | map<client_t,MClientSnap::ref>&& _splits) : | |
6394 | ServerLogContext(s, r), mut(m), splits(std::move(_splits)) { | |
6395 | } | |
7c673cae | 6396 | void finish(int r) override { |
11fdf7f2 | 6397 | server->_link_rollback_finish(mut, mdr, splits); |
7c673cae FG |
6398 | } |
6399 | }; | |
6400 | ||
6401 | void Server::do_link_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr) | |
6402 | { | |
6403 | link_rollback rollback; | |
11fdf7f2 TL |
6404 | auto p = rbl.cbegin(); |
6405 | decode(rollback, p); | |
7c673cae FG |
6406 | |
6407 | dout(10) << "do_link_rollback on " << rollback.reqid | |
6408 | << (rollback.was_inc ? " inc":" dec") | |
6409 | << " ino " << rollback.ino | |
6410 | << dendl; | |
6411 | ||
11fdf7f2 | 6412 | ceph_assert(g_conf()->mds_kill_link_at != 9); |
7c673cae FG |
6413 | |
6414 | mdcache->add_rollback(rollback.reqid, master); // need to finish this update before resolve finishes | |
11fdf7f2 | 6415 | ceph_assert(mdr || mds->is_resolve()); |
7c673cae FG |
6416 | |
6417 | MutationRef mut(new MutationImpl(nullptr, utime_t(), rollback.reqid)); | |
6418 | mut->ls = mds->mdlog->get_current_segment(); | |
6419 | ||
6420 | CInode *in = mdcache->get_inode(rollback.ino); | |
11fdf7f2 | 6421 | ceph_assert(in); |
7c673cae | 6422 | dout(10) << " target is " << *in << dendl; |
11fdf7f2 | 6423 | ceph_assert(!in->is_projected()); // live slave request hold versionlock xlock. |
7c673cae | 6424 | |
94b18763 FG |
6425 | auto &pi = in->project_inode(); |
6426 | pi.inode.version = in->pre_dirty(); | |
7c673cae FG |
6427 | mut->add_projected_inode(in); |
6428 | ||
6429 | // parent dir rctime | |
6430 | CDir *parent = in->get_projected_parent_dn()->get_dir(); | |
6431 | fnode_t *pf = parent->project_fnode(); | |
6432 | mut->add_projected_fnode(parent); | |
6433 | pf->version = parent->pre_dirty(); | |
94b18763 | 6434 | if (pf->fragstat.mtime == pi.inode.ctime) { |
7c673cae | 6435 | pf->fragstat.mtime = rollback.old_dir_mtime; |
94b18763 | 6436 | if (pf->rstat.rctime == pi.inode.ctime) |
7c673cae FG |
6437 | pf->rstat.rctime = rollback.old_dir_rctime; |
6438 | mut->add_updated_lock(&parent->get_inode()->filelock); | |
6439 | mut->add_updated_lock(&parent->get_inode()->nestlock); | |
6440 | } | |
6441 | ||
6442 | // inode | |
91327a77 | 6443 | pi.inode.ctime = rollback.old_ctime; |
7c673cae | 6444 | if (rollback.was_inc) |
94b18763 | 6445 | pi.inode.nlink--; |
7c673cae | 6446 | else |
94b18763 | 6447 | pi.inode.nlink++; |
7c673cae | 6448 | |
11fdf7f2 TL |
6449 | map<client_t,MClientSnap::ref> splits; |
6450 | if (rollback.snapbl.length() && in->snaprealm) { | |
6451 | bool hadrealm; | |
6452 | auto p = rollback.snapbl.cbegin(); | |
6453 | decode(hadrealm, p); | |
6454 | if (hadrealm) { | |
6455 | if (!mds->is_resolve()) { | |
6456 | sr_t *new_srnode = new sr_t(); | |
6457 | decode(*new_srnode, p); | |
6458 | in->project_snaprealm(new_srnode); | |
6459 | } else { | |
6460 | decode(in->snaprealm->srnode, p); | |
6461 | } | |
6462 | } else { | |
6463 | SnapRealm *realm = parent->get_inode()->find_snaprealm(); | |
6464 | if (!mds->is_resolve()) | |
6465 | mdcache->prepare_realm_merge(in->snaprealm, realm, splits); | |
6466 | in->project_snaprealm(NULL); | |
6467 | } | |
6468 | } | |
6469 | ||
7c673cae FG |
6470 | // journal it |
6471 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_link_rollback", rollback.reqid, master, | |
6472 | ESlaveUpdate::OP_ROLLBACK, ESlaveUpdate::LINK); | |
6473 | mdlog->start_entry(le); | |
6474 | le->commit.add_dir_context(parent); | |
6475 | le->commit.add_dir(parent, true); | |
6476 | le->commit.add_primary_dentry(in->get_projected_parent_dn(), 0, true); | |
6477 | ||
11fdf7f2 | 6478 | submit_mdlog_entry(le, new C_MDS_LoggedLinkRollback(this, mut, mdr, std::move(splits)), |
7c673cae FG |
6479 | mdr, __func__); |
6480 | mdlog->flush(); | |
6481 | } | |
6482 | ||
11fdf7f2 TL |
6483 | void Server::_link_rollback_finish(MutationRef& mut, MDRequestRef& mdr, |
6484 | map<client_t,MClientSnap::ref>& splits) | |
7c673cae FG |
6485 | { |
6486 | dout(10) << "_link_rollback_finish" << dendl; | |
6487 | ||
11fdf7f2 | 6488 | ceph_assert(g_conf()->mds_kill_link_at != 10); |
7c673cae FG |
6489 | |
6490 | mut->apply(); | |
11fdf7f2 TL |
6491 | |
6492 | if (!mds->is_resolve()) | |
6493 | mdcache->send_snaps(splits); | |
6494 | ||
7c673cae FG |
6495 | if (mdr) |
6496 | mdcache->request_finish(mdr); | |
6497 | ||
6498 | mdcache->finish_rollback(mut->reqid); | |
6499 | ||
6500 | mut->cleanup(); | |
6501 | } | |
6502 | ||
6503 | ||
11fdf7f2 | 6504 | void Server::handle_slave_link_prep_ack(MDRequestRef& mdr, const MMDSSlaveRequest::const_ref &m) |
7c673cae FG |
6505 | { |
6506 | dout(10) << "handle_slave_link_prep_ack " << *mdr | |
6507 | << " " << *m << dendl; | |
6508 | mds_rank_t from = mds_rank_t(m->get_source().num()); | |
6509 | ||
11fdf7f2 | 6510 | ceph_assert(g_conf()->mds_kill_link_at != 11); |
7c673cae FG |
6511 | |
6512 | // note slave | |
6513 | mdr->more()->slaves.insert(from); | |
6514 | ||
6515 | // witnessed! | |
11fdf7f2 | 6516 | ceph_assert(mdr->more()->witnessed.count(from) == 0); |
7c673cae | 6517 | mdr->more()->witnessed.insert(from); |
11fdf7f2 | 6518 | ceph_assert(!m->is_not_journaled()); |
7c673cae FG |
6519 | mdr->more()->has_journaled_slaves = true; |
6520 | ||
6521 | // remove from waiting list | |
11fdf7f2 | 6522 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae FG |
6523 | mdr->more()->waiting_on_slave.erase(from); |
6524 | ||
11fdf7f2 | 6525 | ceph_assert(mdr->more()->waiting_on_slave.empty()); |
7c673cae FG |
6526 | |
6527 | dispatch_client_request(mdr); // go again! | |
6528 | } | |
6529 | ||
6530 | ||
6531 | ||
6532 | ||
6533 | ||
6534 | // UNLINK | |
6535 | ||
6536 | void Server::handle_client_unlink(MDRequestRef& mdr) | |
6537 | { | |
11fdf7f2 | 6538 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
6539 | client_t client = mdr->get_client(); |
6540 | ||
6541 | // rmdir or unlink? | |
6542 | bool rmdir = false; | |
6543 | if (req->get_op() == CEPH_MDS_OP_RMDIR) rmdir = true; | |
6544 | ||
91327a77 AA |
6545 | const filepath& refpath = req->get_filepath(); |
6546 | if (refpath.depth() == 0) { | |
7c673cae FG |
6547 | respond_to_request(mdr, -EINVAL); |
6548 | return; | |
91327a77 AA |
6549 | } |
6550 | if (refpath.is_last_dot_or_dotdot()) { | |
6551 | respond_to_request(mdr, -ENOTEMPTY); | |
6552 | return; | |
6553 | } | |
7c673cae FG |
6554 | |
6555 | // traverse to path | |
6556 | vector<CDentry*> trace; | |
6557 | CInode *in; | |
11fdf7f2 TL |
6558 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
6559 | int r = mdcache->path_traverse(mdr, cf, refpath, &trace, &in, MDS_TRAVERSE_FORWARD); | |
7c673cae FG |
6560 | if (r > 0) return; |
6561 | if (r < 0) { | |
6562 | if (r == -ESTALE) { | |
6563 | dout(10) << "FAIL on ESTALE but attempting recovery" << dendl; | |
91327a77 | 6564 | mdcache->find_ino_peers(refpath.get_ino(), new C_MDS_TryFindInode(this, mdr)); |
7c673cae FG |
6565 | return; |
6566 | } | |
6567 | respond_to_request(mdr, r); | |
6568 | return; | |
6569 | } | |
6570 | if (mdr->snapid != CEPH_NOSNAP) { | |
6571 | respond_to_request(mdr, -EROFS); | |
6572 | return; | |
6573 | } | |
6574 | ||
91327a77 | 6575 | CDentry *dn = trace.back(); |
11fdf7f2 | 6576 | ceph_assert(dn); |
7c673cae FG |
6577 | if (!dn->is_auth()) { |
6578 | mdcache->request_forward(mdr, dn->authority().first); | |
6579 | return; | |
6580 | } | |
6581 | ||
6582 | CInode *diri = dn->get_dir()->get_inode(); | |
6583 | ||
6584 | CDentry::linkage_t *dnl = dn->get_linkage(client, mdr); | |
11fdf7f2 | 6585 | ceph_assert(!dnl->is_null()); |
7c673cae FG |
6586 | |
6587 | if (rmdir) { | |
6588 | dout(7) << "handle_client_rmdir on " << *dn << dendl; | |
6589 | } else { | |
6590 | dout(7) << "handle_client_unlink on " << *dn << dendl; | |
6591 | } | |
6592 | dout(7) << "dn links to " << *in << dendl; | |
6593 | ||
6594 | // rmdir vs is_dir | |
6595 | if (in->is_dir()) { | |
6596 | if (rmdir) { | |
6597 | // do empty directory checks | |
6598 | if (_dir_is_nonempty_unlocked(mdr, in)) { | |
6599 | respond_to_request(mdr, -ENOTEMPTY); | |
6600 | return; | |
6601 | } | |
6602 | } else { | |
6603 | dout(7) << "handle_client_unlink on dir " << *in << ", returning error" << dendl; | |
6604 | respond_to_request(mdr, -EISDIR); | |
6605 | return; | |
6606 | } | |
6607 | } else { | |
6608 | if (rmdir) { | |
6609 | // unlink | |
6610 | dout(7) << "handle_client_rmdir on non-dir " << *in << ", returning error" << dendl; | |
6611 | respond_to_request(mdr, -ENOTDIR); | |
6612 | return; | |
6613 | } | |
6614 | } | |
6615 | ||
6616 | // -- create stray dentry? -- | |
6617 | CDentry *straydn = NULL; | |
6618 | if (dnl->is_primary()) { | |
6619 | straydn = prepare_stray_dentry(mdr, dnl->get_inode()); | |
6620 | if (!straydn) | |
6621 | return; | |
6622 | dout(10) << " straydn is " << *straydn << dendl; | |
6623 | } else if (mdr->straydn) { | |
6624 | mdr->unpin(mdr->straydn); | |
6625 | mdr->straydn = NULL; | |
6626 | } | |
6627 | ||
6628 | // lock | |
11fdf7f2 TL |
6629 | MutationImpl::LockOpVec lov; |
6630 | ||
6631 | for (int i=0; i<(int)trace.size()-1; i++) | |
6632 | lov.add_rdlock(&trace[i]->lock); | |
6633 | lov.add_xlock(&dn->lock); | |
6634 | lov.add_wrlock(&diri->filelock); | |
6635 | lov.add_wrlock(&diri->nestlock); | |
6636 | lov.add_xlock(&in->linklock); | |
7c673cae | 6637 | if (straydn) { |
11fdf7f2 TL |
6638 | lov.add_wrlock(&straydn->get_dir()->inode->filelock); |
6639 | lov.add_wrlock(&straydn->get_dir()->inode->nestlock); | |
6640 | lov.add_xlock(&straydn->lock); | |
7c673cae | 6641 | } |
11fdf7f2 TL |
6642 | |
6643 | mds->locker->include_snap_rdlocks(diri, lov); | |
6644 | lov.add_xlock(&in->snaplock); | |
7c673cae | 6645 | if (in->is_dir()) |
11fdf7f2 | 6646 | lov.add_rdlock(&in->filelock); // to verify it's empty |
7c673cae | 6647 | |
11fdf7f2 | 6648 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
6649 | return; |
6650 | ||
6651 | if (in->is_dir() && | |
6652 | _dir_is_nonempty(mdr, in)) { | |
6653 | respond_to_request(mdr, -ENOTEMPTY); | |
6654 | return; | |
6655 | } | |
6656 | ||
6657 | if ((!mdr->has_more() || mdr->more()->witnessed.empty())) { | |
6658 | if (!check_access(mdr, diri, MAY_WRITE)) | |
6659 | return; | |
6660 | } | |
6661 | ||
11fdf7f2 TL |
6662 | if (straydn) |
6663 | straydn->first = mdcache->get_global_snaprealm()->get_newest_seq() + 1; | |
6664 | ||
6665 | if (!mdr->more()->desti_srnode) { | |
6666 | if (in->is_projected_snaprealm_global()) { | |
6667 | sr_t *new_srnode = in->prepare_new_srnode(0); | |
6668 | in->record_snaprealm_parent_dentry(new_srnode, NULL, dn, dnl->is_primary()); | |
6669 | // dropping the last linkage or dropping the last remote linkage, | |
6670 | // detch the inode from global snaprealm | |
6671 | auto nlink = in->get_projected_inode()->nlink; | |
6672 | if (nlink == 1 || | |
6673 | (nlink == 2 && !dnl->is_primary() && | |
6674 | !in->get_projected_parent_dir()->inode->is_stray())) | |
6675 | in->clear_snaprealm_global(new_srnode); | |
6676 | mdr->more()->desti_srnode = new_srnode; | |
6677 | } else if (dnl->is_primary()) { | |
6678 | // prepare snaprealm blob for slave request | |
6679 | SnapRealm *realm = in->find_snaprealm(); | |
6680 | snapid_t follows = realm->get_newest_seq(); | |
6681 | if (in->snaprealm || follows + 1 > in->get_oldest_snap()) { | |
6682 | sr_t *new_srnode = in->prepare_new_srnode(follows); | |
6683 | in->record_snaprealm_past_parent(new_srnode, straydn->get_dir()->inode->find_snaprealm()); | |
6684 | mdr->more()->desti_srnode = new_srnode; | |
6685 | } | |
6686 | } | |
6687 | } | |
6688 | ||
7c673cae FG |
6689 | // yay! |
6690 | if (in->is_dir() && in->has_subtree_root_dirfrag()) { | |
6691 | // subtree root auths need to be witnesses | |
6692 | set<mds_rank_t> witnesses; | |
6693 | in->list_replicas(witnesses); | |
6694 | dout(10) << " witnesses " << witnesses << ", have " << mdr->more()->witnessed << dendl; | |
6695 | ||
6696 | for (set<mds_rank_t>::iterator p = witnesses.begin(); | |
6697 | p != witnesses.end(); | |
6698 | ++p) { | |
6699 | if (mdr->more()->witnessed.count(*p)) { | |
6700 | dout(10) << " already witnessed by mds." << *p << dendl; | |
6701 | } else if (mdr->more()->waiting_on_slave.count(*p)) { | |
6702 | dout(10) << " already waiting on witness mds." << *p << dendl; | |
6703 | } else { | |
6704 | if (!_rmdir_prepare_witness(mdr, *p, trace, straydn)) | |
6705 | return; | |
6706 | } | |
6707 | } | |
6708 | if (!mdr->more()->waiting_on_slave.empty()) | |
6709 | return; // we're waiting for a witness. | |
6710 | } | |
6711 | ||
6712 | // ok! | |
6713 | if (dnl->is_remote() && !dnl->get_inode()->is_auth()) | |
6714 | _link_remote(mdr, false, dn, dnl->get_inode()); | |
6715 | else | |
6716 | _unlink_local(mdr, dn, straydn); | |
6717 | } | |
6718 | ||
6719 | class C_MDS_unlink_local_finish : public ServerLogContext { | |
6720 | CDentry *dn; | |
6721 | CDentry *straydn; | |
6722 | version_t dnpv; // deleted dentry | |
6723 | public: | |
6724 | C_MDS_unlink_local_finish(Server *s, MDRequestRef& r, CDentry *d, CDentry *sd) : | |
6725 | ServerLogContext(s, r), dn(d), straydn(sd), | |
6726 | dnpv(d->get_projected_version()) {} | |
6727 | void finish(int r) override { | |
11fdf7f2 | 6728 | ceph_assert(r == 0); |
7c673cae FG |
6729 | server->_unlink_local_finish(mdr, dn, straydn, dnpv); |
6730 | } | |
6731 | }; | |
6732 | ||
6733 | void Server::_unlink_local(MDRequestRef& mdr, CDentry *dn, CDentry *straydn) | |
6734 | { | |
6735 | dout(10) << "_unlink_local " << *dn << dendl; | |
6736 | ||
6737 | CDentry::linkage_t *dnl = dn->get_projected_linkage(); | |
6738 | CInode *in = dnl->get_inode(); | |
6739 | ||
7c673cae FG |
6740 | |
6741 | // ok, let's do it. | |
6742 | mdr->ls = mdlog->get_current_segment(); | |
6743 | ||
6744 | // prepare log entry | |
6745 | EUpdate *le = new EUpdate(mdlog, "unlink_local"); | |
6746 | mdlog->start_entry(le); | |
6747 | le->metablob.add_client_req(mdr->reqid, mdr->client_request->get_oldest_client_tid()); | |
6748 | if (!mdr->more()->witnessed.empty()) { | |
6749 | dout(20) << " noting uncommitted_slaves " << mdr->more()->witnessed << dendl; | |
6750 | le->reqid = mdr->reqid; | |
6751 | le->had_slaves = true; | |
6752 | mdcache->add_uncommitted_master(mdr->reqid, mdr->ls, mdr->more()->witnessed); | |
6753 | } | |
6754 | ||
6755 | if (straydn) { | |
11fdf7f2 | 6756 | ceph_assert(dnl->is_primary()); |
7c673cae | 6757 | straydn->push_projected_linkage(in); |
7c673cae FG |
6758 | } |
6759 | ||
6760 | // the unlinked dentry | |
6761 | dn->pre_dirty(); | |
6762 | ||
94b18763 FG |
6763 | auto &pi = in->project_inode(); |
6764 | { | |
6765 | std::string t; | |
6766 | dn->make_path_string(t, true); | |
11fdf7f2 | 6767 | pi.inode.stray_prior_path = std::move(t); |
94b18763 | 6768 | } |
94b18763 | 6769 | pi.inode.version = in->pre_dirty(); |
91327a77 AA |
6770 | pi.inode.ctime = mdr->get_op_stamp(); |
6771 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
6772 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 FG |
6773 | pi.inode.change_attr++; |
6774 | pi.inode.nlink--; | |
6775 | if (pi.inode.nlink == 0) | |
7c673cae FG |
6776 | in->state_set(CInode::STATE_ORPHAN); |
6777 | ||
11fdf7f2 TL |
6778 | if (mdr->more()->desti_srnode) { |
6779 | auto& desti_srnode = mdr->more()->desti_srnode; | |
6780 | in->project_snaprealm(desti_srnode); | |
6781 | desti_srnode = NULL; | |
6782 | } | |
6783 | ||
6784 | if (straydn) { | |
6785 | // will manually pop projected inode | |
6786 | ||
7c673cae | 6787 | // primary link. add stray dentry. |
7c673cae FG |
6788 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, dn->get_dir(), PREDIRTY_PRIMARY|PREDIRTY_DIR, -1); |
6789 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, straydn->get_dir(), PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
6790 | ||
94b18763 | 6791 | pi.inode.update_backtrace(); |
7c673cae FG |
6792 | le->metablob.add_primary_dentry(straydn, in, true, true); |
6793 | } else { | |
11fdf7f2 | 6794 | mdr->add_projected_inode(in); |
7c673cae FG |
6795 | // remote link. update remote inode. |
6796 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, dn->get_dir(), PREDIRTY_DIR, -1); | |
6797 | mdcache->predirty_journal_parents(mdr, &le->metablob, in, 0, PREDIRTY_PRIMARY); | |
6798 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, in); | |
6799 | } | |
6800 | ||
6801 | mdcache->journal_cow_dentry(mdr.get(), &le->metablob, dn); | |
6802 | le->metablob.add_null_dentry(dn, true); | |
6803 | ||
6804 | if (in->is_dir()) { | |
6805 | dout(10) << " noting renamed (unlinked) dir ino " << in->ino() << " in metablob" << dendl; | |
6806 | le->metablob.renamed_dirino = in->ino(); | |
6807 | } | |
6808 | ||
6809 | dn->push_projected_linkage(); | |
6810 | ||
11fdf7f2 TL |
6811 | if (straydn) { |
6812 | ceph_assert(in->first <= straydn->first); | |
6813 | in->first = straydn->first; | |
6814 | } | |
6815 | ||
7c673cae | 6816 | if (in->is_dir()) { |
11fdf7f2 | 6817 | ceph_assert(straydn); |
7c673cae FG |
6818 | mdcache->project_subtree_rename(in, dn->get_dir(), straydn->get_dir()); |
6819 | } | |
6820 | ||
6821 | journal_and_reply(mdr, 0, dn, le, new C_MDS_unlink_local_finish(this, mdr, dn, straydn)); | |
6822 | } | |
6823 | ||
6824 | void Server::_unlink_local_finish(MDRequestRef& mdr, | |
6825 | CDentry *dn, CDentry *straydn, | |
6826 | version_t dnpv) | |
6827 | { | |
6828 | dout(10) << "_unlink_local_finish " << *dn << dendl; | |
6829 | ||
6830 | if (!mdr->more()->witnessed.empty()) | |
6831 | mdcache->logged_master_update(mdr->reqid); | |
6832 | ||
11fdf7f2 TL |
6833 | CInode *strayin = NULL; |
6834 | bool hadrealm = false; | |
6835 | if (straydn) { | |
6836 | // if there is newly created snaprealm, need to split old snaprealm's | |
6837 | // inodes_with_caps. So pop snaprealm before linkage changes. | |
6838 | strayin = dn->get_linkage()->get_inode(); | |
6839 | hadrealm = strayin->snaprealm ? true : false; | |
6840 | strayin->early_pop_projected_snaprealm(); | |
6841 | } | |
6842 | ||
7c673cae FG |
6843 | // unlink main dentry |
6844 | dn->get_dir()->unlink_inode(dn); | |
6845 | dn->pop_projected_linkage(); | |
6846 | ||
6847 | // relink as stray? (i.e. was primary link?) | |
7c673cae FG |
6848 | if (straydn) { |
6849 | dout(20) << " straydn is " << *straydn << dendl; | |
11fdf7f2 TL |
6850 | straydn->pop_projected_linkage(); |
6851 | ||
6852 | strayin->pop_and_dirty_projected_inode(mdr->ls); | |
7c673cae | 6853 | |
7c673cae FG |
6854 | mdcache->touch_dentry_bottom(straydn); |
6855 | } | |
6856 | ||
6857 | dn->mark_dirty(dnpv, mdr->ls); | |
6858 | mdr->apply(); | |
7c673cae FG |
6859 | |
6860 | mdcache->send_dentry_unlink(dn, straydn, mdr); | |
6861 | ||
11fdf7f2 TL |
6862 | if (straydn) { |
6863 | // update subtree map? | |
6864 | if (strayin->is_dir()) | |
6865 | mdcache->adjust_subtree_after_rename(strayin, dn->get_dir(), true); | |
6866 | ||
6867 | if (strayin->snaprealm && !hadrealm) | |
6868 | mdcache->do_realm_invalidate_and_update_notify(strayin, CEPH_SNAP_OP_SPLIT, false); | |
6869 | } | |
7c673cae FG |
6870 | |
6871 | // bump pop | |
11fdf7f2 | 6872 | mds->balancer->hit_dir(dn->get_dir(), META_POP_IWR); |
7c673cae FG |
6873 | |
6874 | // reply | |
6875 | respond_to_request(mdr, 0); | |
6876 | ||
6877 | // removing a new dn? | |
6878 | dn->get_dir()->try_remove_unlinked_dn(dn); | |
6879 | ||
6880 | // clean up ? | |
6881 | // respond_to_request() drops locks. So stray reintegration can race with us. | |
6882 | if (straydn && !straydn->get_projected_linkage()->is_null()) { | |
6883 | // Tip off the MDCache that this dentry is a stray that | |
6884 | // might be elegible for purge. | |
6885 | mdcache->notify_stray(straydn); | |
6886 | } | |
6887 | } | |
6888 | ||
6889 | bool Server::_rmdir_prepare_witness(MDRequestRef& mdr, mds_rank_t who, vector<CDentry*>& trace, CDentry *straydn) | |
6890 | { | |
6891 | if (mds->is_cluster_degraded() && | |
6892 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(who)) { | |
6893 | dout(10) << "_rmdir_prepare_witness mds." << who << " is not active" << dendl; | |
6894 | if (mdr->more()->waiting_on_slave.empty()) | |
6895 | mds->wait_for_active_peer(who, new C_MDS_RetryRequest(mdcache, mdr)); | |
6896 | return false; | |
6897 | } | |
6898 | ||
6899 | dout(10) << "_rmdir_prepare_witness mds." << who << dendl; | |
11fdf7f2 | 6900 | auto req = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RMDIRPREP); |
7c673cae FG |
6901 | req->srcdnpath = filepath(trace.front()->get_dir()->ino()); |
6902 | for (auto dn : trace) | |
94b18763 | 6903 | req->srcdnpath.push_dentry(dn->get_name()); |
11fdf7f2 TL |
6904 | mdcache->replicate_stray(straydn, who, req->straybl); |
6905 | if (mdr->more()->desti_srnode) | |
6906 | encode(*mdr->more()->desti_srnode, req->desti_snapbl); | |
7c673cae FG |
6907 | |
6908 | req->op_stamp = mdr->get_op_stamp(); | |
6909 | mds->send_message_mds(req, who); | |
6910 | ||
11fdf7f2 | 6911 | ceph_assert(mdr->more()->waiting_on_slave.count(who) == 0); |
7c673cae FG |
6912 | mdr->more()->waiting_on_slave.insert(who); |
6913 | return true; | |
6914 | } | |
6915 | ||
6916 | struct C_MDS_SlaveRmdirPrep : public ServerLogContext { | |
6917 | CDentry *dn, *straydn; | |
6918 | C_MDS_SlaveRmdirPrep(Server *s, MDRequestRef& r, CDentry *d, CDentry *st) | |
6919 | : ServerLogContext(s, r), dn(d), straydn(st) {} | |
6920 | void finish(int r) override { | |
6921 | server->_logged_slave_rmdir(mdr, dn, straydn); | |
6922 | } | |
6923 | }; | |
6924 | ||
6925 | struct C_MDS_SlaveRmdirCommit : public ServerContext { | |
6926 | MDRequestRef mdr; | |
31f18b77 FG |
6927 | CDentry *straydn; |
6928 | C_MDS_SlaveRmdirCommit(Server *s, MDRequestRef& r, CDentry *sd) | |
6929 | : ServerContext(s), mdr(r), straydn(sd) { } | |
7c673cae | 6930 | void finish(int r) override { |
31f18b77 | 6931 | server->_commit_slave_rmdir(mdr, r, straydn); |
7c673cae FG |
6932 | } |
6933 | }; | |
6934 | ||
6935 | void Server::handle_slave_rmdir_prep(MDRequestRef& mdr) | |
6936 | { | |
6937 | dout(10) << "handle_slave_rmdir_prep " << *mdr | |
6938 | << " " << mdr->slave_request->srcdnpath | |
6939 | << " to " << mdr->slave_request->destdnpath | |
6940 | << dendl; | |
6941 | ||
6942 | vector<CDentry*> trace; | |
6943 | filepath srcpath(mdr->slave_request->srcdnpath); | |
6944 | dout(10) << " src " << srcpath << dendl; | |
6945 | CInode *in; | |
11fdf7f2 TL |
6946 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
6947 | int r = mdcache->path_traverse(mdr, cf, srcpath, &trace, &in, MDS_TRAVERSE_DISCOVERXLOCK); | |
7c673cae FG |
6948 | if (r > 0) return; |
6949 | if (r == -ESTALE) { | |
6950 | mdcache->find_ino_peers(srcpath.get_ino(), new C_MDS_RetryRequest(mdcache, mdr), | |
6951 | mdr->slave_to_mds); | |
6952 | return; | |
6953 | } | |
11fdf7f2 | 6954 | ceph_assert(r == 0); |
91327a77 | 6955 | CDentry *dn = trace.back(); |
7c673cae FG |
6956 | dout(10) << " dn " << *dn << dendl; |
6957 | mdr->pin(dn); | |
6958 | ||
11fdf7f2 | 6959 | ceph_assert(mdr->straydn); |
7c673cae FG |
6960 | CDentry *straydn = mdr->straydn; |
6961 | dout(10) << " straydn " << *straydn << dendl; | |
6962 | ||
6963 | mdr->set_op_stamp(mdr->slave_request->op_stamp); | |
6964 | ||
6965 | rmdir_rollback rollback; | |
6966 | rollback.reqid = mdr->reqid; | |
6967 | rollback.src_dir = dn->get_dir()->dirfrag(); | |
11fdf7f2 | 6968 | rollback.src_dname = dn->get_name(); |
7c673cae | 6969 | rollback.dest_dir = straydn->get_dir()->dirfrag(); |
11fdf7f2 TL |
6970 | rollback.dest_dname = straydn->get_name(); |
6971 | if (mdr->slave_request->desti_snapbl.length()) { | |
6972 | if (in->snaprealm) { | |
6973 | encode(true, rollback.snapbl); | |
6974 | in->encode_snap_blob(rollback.snapbl); | |
6975 | } else { | |
6976 | encode(false, rollback.snapbl); | |
6977 | } | |
6978 | } | |
6979 | encode(rollback, mdr->more()->rollback_bl); | |
6980 | // FIXME: rollback snaprealm | |
7c673cae FG |
6981 | dout(20) << " rollback is " << mdr->more()->rollback_bl.length() << " bytes" << dendl; |
6982 | ||
6983 | // set up commit waiter | |
31f18b77 | 6984 | mdr->more()->slave_commit = new C_MDS_SlaveRmdirCommit(this, mdr, straydn); |
7c673cae | 6985 | |
11fdf7f2 TL |
6986 | straydn->push_projected_linkage(in); |
6987 | dn->push_projected_linkage(); | |
7c673cae | 6988 | |
11fdf7f2 TL |
6989 | ceph_assert(straydn->first >= in->first); |
6990 | in->first = straydn->first; | |
7c673cae | 6991 | |
11fdf7f2 TL |
6992 | if (!in->has_subtree_root_dirfrag(mds->get_nodeid())) { |
6993 | dout(10) << " no auth subtree in " << *in << ", skipping journal" << dendl; | |
6994 | _logged_slave_rmdir(mdr, dn, straydn); | |
7c673cae FG |
6995 | return; |
6996 | } | |
6997 | ||
7c673cae FG |
6998 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rmdir", mdr->reqid, mdr->slave_to_mds, |
6999 | ESlaveUpdate::OP_PREPARE, ESlaveUpdate::RMDIR); | |
7000 | mdlog->start_entry(le); | |
7001 | le->rollback = mdr->more()->rollback_bl; | |
7002 | ||
7003 | le->commit.add_dir_context(straydn->get_dir()); | |
7004 | le->commit.add_primary_dentry(straydn, in, true); | |
7005 | // slave: no need to journal original dentry | |
7006 | ||
7007 | dout(10) << " noting renamed (unlinked) dir ino " << in->ino() << " in metablob" << dendl; | |
7008 | le->commit.renamed_dirino = in->ino(); | |
7009 | ||
7010 | mdcache->project_subtree_rename(in, dn->get_dir(), straydn->get_dir()); | |
7011 | ||
7012 | mdr->more()->slave_update_journaled = true; | |
7013 | submit_mdlog_entry(le, new C_MDS_SlaveRmdirPrep(this, mdr, dn, straydn), | |
7014 | mdr, __func__); | |
7015 | mdlog->flush(); | |
7016 | } | |
7017 | ||
7018 | void Server::_logged_slave_rmdir(MDRequestRef& mdr, CDentry *dn, CDentry *straydn) | |
7019 | { | |
7020 | dout(10) << "_logged_slave_rmdir " << *mdr << " on " << *dn << dendl; | |
11fdf7f2 TL |
7021 | CInode *in = dn->get_linkage()->get_inode(); |
7022 | ||
7023 | bool new_realm; | |
7024 | if (mdr->slave_request->desti_snapbl.length()) { | |
7025 | new_realm = !in->snaprealm; | |
7026 | in->decode_snap_blob(mdr->slave_request->desti_snapbl); | |
7027 | ceph_assert(in->snaprealm); | |
7028 | ceph_assert(in->snaprealm->have_past_parents_open()); | |
7029 | } else { | |
7030 | new_realm = false; | |
7031 | } | |
7c673cae FG |
7032 | |
7033 | // update our cache now, so we are consistent with what is in the journal | |
7034 | // when we journal a subtree map | |
7c673cae FG |
7035 | dn->get_dir()->unlink_inode(dn); |
7036 | straydn->pop_projected_linkage(); | |
7037 | dn->pop_projected_linkage(); | |
11fdf7f2 TL |
7038 | |
7039 | mdcache->adjust_subtree_after_rename(in, dn->get_dir(), mdr->more()->slave_update_journaled); | |
7040 | ||
7041 | if (new_realm) | |
7042 | mdcache->do_realm_invalidate_and_update_notify(in, CEPH_SNAP_OP_SPLIT, false); | |
7c673cae FG |
7043 | |
7044 | // done. | |
91327a77 | 7045 | mdr->reset_slave_request(); |
7c673cae FG |
7046 | mdr->straydn = 0; |
7047 | ||
7048 | if (!mdr->aborted) { | |
11fdf7f2 TL |
7049 | auto reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RMDIRPREPACK); |
7050 | if (!mdr->more()->slave_update_journaled) | |
7051 | reply->mark_not_journaled(); | |
7c673cae FG |
7052 | mds->send_message_mds(reply, mdr->slave_to_mds); |
7053 | } else { | |
7054 | dout(10) << " abort flag set, finishing" << dendl; | |
7055 | mdcache->request_finish(mdr); | |
7056 | } | |
7057 | } | |
7058 | ||
11fdf7f2 | 7059 | void Server::handle_slave_rmdir_prep_ack(MDRequestRef& mdr, const MMDSSlaveRequest::const_ref &ack) |
7c673cae FG |
7060 | { |
7061 | dout(10) << "handle_slave_rmdir_prep_ack " << *mdr | |
7062 | << " " << *ack << dendl; | |
7063 | ||
7064 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
7065 | ||
7066 | mdr->more()->slaves.insert(from); | |
7067 | mdr->more()->witnessed.insert(from); | |
7068 | if (!ack->is_not_journaled()) | |
7069 | mdr->more()->has_journaled_slaves = true; | |
7070 | ||
7071 | // remove from waiting list | |
11fdf7f2 | 7072 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae FG |
7073 | mdr->more()->waiting_on_slave.erase(from); |
7074 | ||
7075 | if (mdr->more()->waiting_on_slave.empty()) | |
7076 | dispatch_client_request(mdr); // go again! | |
7077 | else | |
7078 | dout(10) << "still waiting on slaves " << mdr->more()->waiting_on_slave << dendl; | |
7079 | } | |
7080 | ||
31f18b77 | 7081 | void Server::_commit_slave_rmdir(MDRequestRef& mdr, int r, CDentry *straydn) |
7c673cae FG |
7082 | { |
7083 | dout(10) << "_commit_slave_rmdir " << *mdr << " r=" << r << dendl; | |
7084 | ||
7085 | if (r == 0) { | |
31f18b77 FG |
7086 | if (mdr->more()->slave_update_journaled) { |
7087 | CInode *strayin = straydn->get_projected_linkage()->get_inode(); | |
7088 | if (strayin && !strayin->snaprealm) | |
7089 | mdcache->clear_dirty_bits_for_stray(strayin); | |
7090 | } | |
7091 | ||
7c673cae FG |
7092 | mdr->cleanup(); |
7093 | ||
7094 | if (mdr->more()->slave_update_journaled) { | |
7095 | // write a commit to the journal | |
7096 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rmdir_commit", mdr->reqid, | |
7097 | mdr->slave_to_mds, ESlaveUpdate::OP_COMMIT, | |
7098 | ESlaveUpdate::RMDIR); | |
7099 | mdlog->start_entry(le); | |
7100 | submit_mdlog_entry(le, new C_MDS_CommittedSlave(this, mdr), mdr, __func__); | |
7101 | mdlog->flush(); | |
7102 | } else { | |
7103 | _committed_slave(mdr); | |
7104 | } | |
7105 | } else { | |
7106 | // abort | |
7107 | do_rmdir_rollback(mdr->more()->rollback_bl, mdr->slave_to_mds, mdr); | |
7108 | } | |
7109 | } | |
7110 | ||
7111 | struct C_MDS_LoggedRmdirRollback : public ServerLogContext { | |
7112 | metareqid_t reqid; | |
7113 | CDentry *dn; | |
7114 | CDentry *straydn; | |
7115 | C_MDS_LoggedRmdirRollback(Server *s, MDRequestRef& m, metareqid_t mr, CDentry *d, CDentry *st) | |
7116 | : ServerLogContext(s, m), reqid(mr), dn(d), straydn(st) {} | |
7117 | void finish(int r) override { | |
7118 | server->_rmdir_rollback_finish(mdr, reqid, dn, straydn); | |
7119 | } | |
7120 | }; | |
7121 | ||
7122 | void Server::do_rmdir_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr) | |
7123 | { | |
7124 | // unlink the other rollback methods, the rmdir rollback is only | |
7125 | // needed to record the subtree changes in the journal for inode | |
7126 | // replicas who are auth for empty dirfrags. no actual changes to | |
7127 | // the file system are taking place here, so there is no Mutation. | |
7128 | ||
7129 | rmdir_rollback rollback; | |
11fdf7f2 TL |
7130 | auto p = rbl.cbegin(); |
7131 | decode(rollback, p); | |
7c673cae FG |
7132 | |
7133 | dout(10) << "do_rmdir_rollback on " << rollback.reqid << dendl; | |
7134 | mdcache->add_rollback(rollback.reqid, master); // need to finish this update before resolve finishes | |
11fdf7f2 | 7135 | ceph_assert(mdr || mds->is_resolve()); |
7c673cae FG |
7136 | |
7137 | CDir *dir = mdcache->get_dirfrag(rollback.src_dir); | |
7138 | if (!dir) | |
7139 | dir = mdcache->get_dirfrag(rollback.src_dir.ino, rollback.src_dname); | |
11fdf7f2 | 7140 | ceph_assert(dir); |
7c673cae | 7141 | CDentry *dn = dir->lookup(rollback.src_dname); |
11fdf7f2 | 7142 | ceph_assert(dn); |
7c673cae | 7143 | dout(10) << " dn " << *dn << dendl; |
11fdf7f2 TL |
7144 | CDir *straydir = mdcache->get_dirfrag(rollback.dest_dir); |
7145 | ceph_assert(straydir); | |
7146 | CDentry *straydn = straydir->lookup(rollback.dest_dname); | |
7147 | ceph_assert(straydn); | |
7148 | dout(10) << " straydn " << *straydn << dendl; | |
7c673cae FG |
7149 | CInode *in = straydn->get_linkage()->get_inode(); |
7150 | ||
11fdf7f2 TL |
7151 | dn->push_projected_linkage(in); |
7152 | straydn->push_projected_linkage(); | |
7c673cae | 7153 | |
11fdf7f2 TL |
7154 | if (rollback.snapbl.length() && in->snaprealm) { |
7155 | bool hadrealm; | |
7156 | auto p = rollback.snapbl.cbegin(); | |
7157 | decode(hadrealm, p); | |
7158 | if (hadrealm) { | |
7159 | decode(in->snaprealm->srnode, p); | |
7160 | } else { | |
7161 | in->snaprealm->merge_to(dir->get_inode()->find_snaprealm()); | |
7162 | } | |
7163 | } | |
7c673cae | 7164 | |
11fdf7f2 TL |
7165 | if (mdr && !mdr->more()->slave_update_journaled) { |
7166 | ceph_assert(!in->has_subtree_root_dirfrag(mds->get_nodeid())); | |
7c673cae | 7167 | |
11fdf7f2 | 7168 | _rmdir_rollback_finish(mdr, rollback.reqid, dn, straydn); |
7c673cae FG |
7169 | return; |
7170 | } | |
7171 | ||
7c673cae FG |
7172 | |
7173 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rmdir_rollback", rollback.reqid, master, | |
7174 | ESlaveUpdate::OP_ROLLBACK, ESlaveUpdate::RMDIR); | |
7175 | mdlog->start_entry(le); | |
7176 | ||
7177 | le->commit.add_dir_context(dn->get_dir()); | |
7178 | le->commit.add_primary_dentry(dn, in, true); | |
7179 | // slave: no need to journal straydn | |
7180 | ||
7181 | dout(10) << " noting renamed (unlinked) dir ino " << in->ino() << " in metablob" << dendl; | |
7182 | le->commit.renamed_dirino = in->ino(); | |
7183 | ||
7184 | mdcache->project_subtree_rename(in, straydn->get_dir(), dn->get_dir()); | |
7185 | ||
7186 | submit_mdlog_entry(le, | |
7187 | new C_MDS_LoggedRmdirRollback(this, mdr,rollback.reqid, | |
7188 | dn, straydn), | |
7189 | mdr, __func__); | |
7190 | mdlog->flush(); | |
7191 | } | |
7192 | ||
7193 | void Server::_rmdir_rollback_finish(MDRequestRef& mdr, metareqid_t reqid, CDentry *dn, CDentry *straydn) | |
7194 | { | |
7195 | dout(10) << "_rmdir_rollback_finish " << reqid << dendl; | |
7196 | ||
7197 | straydn->get_dir()->unlink_inode(straydn); | |
7198 | dn->pop_projected_linkage(); | |
7199 | straydn->pop_projected_linkage(); | |
7200 | ||
7201 | CInode *in = dn->get_linkage()->get_inode(); | |
11fdf7f2 TL |
7202 | mdcache->adjust_subtree_after_rename(in, straydn->get_dir(), |
7203 | !mdr || mdr->more()->slave_update_journaled); | |
7204 | ||
7c673cae FG |
7205 | if (mds->is_resolve()) { |
7206 | CDir *root = mdcache->get_subtree_root(straydn->get_dir()); | |
7207 | mdcache->try_trim_non_auth_subtree(root); | |
7208 | } | |
7209 | ||
7210 | if (mdr) | |
7211 | mdcache->request_finish(mdr); | |
7212 | ||
7213 | mdcache->finish_rollback(reqid); | |
7214 | } | |
7215 | ||
7216 | ||
7217 | /** _dir_is_nonempty[_unlocked] | |
7218 | * | |
7219 | * check if a directory is non-empty (i.e. we can rmdir it). | |
7220 | * | |
7221 | * the unlocked varient this is a fastpath check. we can't really be | |
7222 | * sure until we rdlock the filelock. | |
7223 | */ | |
7224 | bool Server::_dir_is_nonempty_unlocked(MDRequestRef& mdr, CInode *in) | |
7225 | { | |
7226 | dout(10) << "dir_is_nonempty_unlocked " << *in << dendl; | |
11fdf7f2 | 7227 | ceph_assert(in->is_auth()); |
7c673cae FG |
7228 | |
7229 | if (in->snaprealm && in->snaprealm->srnode.snaps.size()) | |
7230 | return true; // in a snapshot! | |
7231 | ||
7232 | list<CDir*> ls; | |
7233 | in->get_dirfrags(ls); | |
7234 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
7235 | CDir *dir = *p; | |
7236 | // is the frag obviously non-empty? | |
7237 | if (dir->is_auth()) { | |
7238 | if (dir->get_projected_fnode()->fragstat.size()) { | |
7239 | dout(10) << "dir_is_nonempty_unlocked dirstat has " | |
7240 | << dir->get_projected_fnode()->fragstat.size() << " items " << *dir << dendl; | |
7241 | return true; | |
7242 | } | |
7243 | } | |
7244 | } | |
7245 | ||
7246 | return false; | |
7247 | } | |
7248 | ||
7249 | bool Server::_dir_is_nonempty(MDRequestRef& mdr, CInode *in) | |
7250 | { | |
7251 | dout(10) << "dir_is_nonempty " << *in << dendl; | |
11fdf7f2 TL |
7252 | ceph_assert(in->is_auth()); |
7253 | ceph_assert(in->filelock.can_read(mdr->get_client())); | |
7c673cae FG |
7254 | |
7255 | frag_info_t dirstat; | |
7256 | version_t dirstat_version = in->get_projected_inode()->dirstat.version; | |
7257 | ||
7258 | list<CDir*> ls; | |
7259 | in->get_dirfrags(ls); | |
7260 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
7261 | CDir *dir = *p; | |
7262 | const fnode_t *pf = dir->get_projected_fnode(); | |
7263 | if (pf->fragstat.size()) { | |
7264 | dout(10) << "dir_is_nonempty dirstat has " | |
7265 | << pf->fragstat.size() << " items " << *dir << dendl; | |
7266 | return true; | |
7267 | } | |
7268 | ||
7269 | if (pf->accounted_fragstat.version == dirstat_version) | |
7270 | dirstat.add(pf->accounted_fragstat); | |
7271 | else | |
7272 | dirstat.add(pf->fragstat); | |
7273 | } | |
7274 | ||
7275 | return dirstat.size() != in->get_projected_inode()->dirstat.size(); | |
7276 | } | |
7277 | ||
7278 | ||
7279 | // ====================================================== | |
7280 | ||
7281 | ||
7282 | class C_MDS_rename_finish : public ServerLogContext { | |
7283 | CDentry *srcdn; | |
7284 | CDentry *destdn; | |
7285 | CDentry *straydn; | |
7286 | public: | |
7287 | C_MDS_rename_finish(Server *s, MDRequestRef& r, | |
7288 | CDentry *sdn, CDentry *ddn, CDentry *stdn) : | |
7289 | ServerLogContext(s, r), | |
7290 | srcdn(sdn), destdn(ddn), straydn(stdn) { } | |
7291 | void finish(int r) override { | |
11fdf7f2 | 7292 | ceph_assert(r == 0); |
7c673cae FG |
7293 | server->_rename_finish(mdr, srcdn, destdn, straydn); |
7294 | } | |
7295 | }; | |
7296 | ||
7297 | ||
7298 | /** handle_client_rename | |
7299 | * | |
7300 | * rename master is the destdn auth. this is because cached inodes | |
7301 | * must remain connected. thus, any replica of srci, must also | |
7302 | * replicate destdn, and possibly straydn, so that srci (and | |
7303 | * destdn->inode) remain connected during the rename. | |
7304 | * | |
7305 | * to do this, we freeze srci, then master (destdn auth) verifies that | |
7306 | * all other nodes have also replciated destdn and straydn. note that | |
7307 | * destdn replicas need not also replicate srci. this only works when | |
7308 | * destdn is master. | |
7309 | * | |
7310 | * This function takes responsibility for the passed mdr. | |
7311 | */ | |
7312 | void Server::handle_client_rename(MDRequestRef& mdr) | |
7313 | { | |
11fdf7f2 | 7314 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
7315 | dout(7) << "handle_client_rename " << *req << dendl; |
7316 | ||
7317 | filepath destpath = req->get_filepath(); | |
7318 | filepath srcpath = req->get_filepath2(); | |
7319 | if (destpath.depth() == 0 || srcpath.depth() == 0) { | |
7320 | respond_to_request(mdr, -EINVAL); | |
7321 | return; | |
7322 | } | |
91327a77 AA |
7323 | if (srcpath.is_last_dot_or_dotdot() || destpath.is_last_dot_or_dotdot()) { |
7324 | respond_to_request(mdr, -EBUSY); | |
7325 | return; | |
7326 | } | |
7327 | ||
11fdf7f2 | 7328 | std::string_view destname = destpath.last_dentry(); |
7c673cae FG |
7329 | |
7330 | vector<CDentry*>& srctrace = mdr->dn[1]; | |
7331 | vector<CDentry*>& desttrace = mdr->dn[0]; | |
7332 | ||
11fdf7f2 | 7333 | MutationImpl::LockOpVec lov; |
7c673cae | 7334 | |
11fdf7f2 | 7335 | CDentry *destdn = rdlock_path_xlock_dentry(mdr, 0, lov, true, false, true); |
7c673cae FG |
7336 | if (!destdn) return; |
7337 | dout(10) << " destdn " << *destdn << dendl; | |
7338 | if (mdr->snapid != CEPH_NOSNAP) { | |
7339 | respond_to_request(mdr, -EROFS); | |
7340 | return; | |
7341 | } | |
7342 | CDentry::linkage_t *destdnl = destdn->get_projected_linkage(); | |
7343 | CDir *destdir = destdn->get_dir(); | |
11fdf7f2 | 7344 | ceph_assert(destdir->is_auth()); |
7c673cae | 7345 | |
11fdf7f2 TL |
7346 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
7347 | int r = mdcache->path_traverse(mdr, cf, srcpath, &srctrace, NULL, MDS_TRAVERSE_DISCOVER); | |
7c673cae FG |
7348 | if (r > 0) |
7349 | return; // delayed | |
7350 | if (r < 0) { | |
7351 | if (r == -ESTALE) { | |
7352 | dout(10) << "FAIL on ESTALE but attempting recovery" << dendl; | |
7353 | mdcache->find_ino_peers(srcpath.get_ino(), new C_MDS_TryFindInode(this, mdr)); | |
7354 | } else { | |
7355 | dout(10) << "FAIL on error " << r << dendl; | |
7356 | respond_to_request(mdr, r); | |
7357 | } | |
7358 | return; | |
7359 | ||
7360 | } | |
11fdf7f2 | 7361 | ceph_assert(!srctrace.empty()); |
91327a77 | 7362 | CDentry *srcdn = srctrace.back(); |
7c673cae FG |
7363 | dout(10) << " srcdn " << *srcdn << dendl; |
7364 | if (srcdn->last != CEPH_NOSNAP) { | |
7365 | respond_to_request(mdr, -EROFS); | |
7366 | return; | |
7367 | } | |
11fdf7f2 | 7368 | CDir *srcdir = srcdn->get_dir(); |
7c673cae FG |
7369 | CDentry::linkage_t *srcdnl = srcdn->get_projected_linkage(); |
7370 | CInode *srci = srcdnl->get_inode(); | |
7371 | dout(10) << " srci " << *srci << dendl; | |
7372 | ||
7373 | CInode *oldin = 0; | |
7374 | if (!destdnl->is_null()) { | |
7375 | //dout(10) << "dest dn exists " << *destdn << dendl; | |
7376 | oldin = mdcache->get_dentry_inode(destdn, mdr, true); | |
7377 | if (!oldin) return; | |
7378 | dout(10) << " oldin " << *oldin << dendl; | |
7c673cae FG |
7379 | |
7380 | // non-empty dir? do trivial fast unlocked check, do another check later with read locks | |
7381 | if (oldin->is_dir() && _dir_is_nonempty_unlocked(mdr, oldin)) { | |
7382 | respond_to_request(mdr, -ENOTEMPTY); | |
7383 | return; | |
7384 | } | |
181888fb FG |
7385 | |
7386 | // if srcdn is replica, need to make sure its linkage is correct | |
7387 | if (srcdn->is_auth() || | |
7388 | srcdn->lock.can_read(mdr->get_client()) || | |
7389 | (srcdn->lock.is_xlocked() && srcdn->lock.get_xlock_by() == mdr)) { | |
7390 | // mv /some/thing /to/some/existing_other_thing | |
7391 | if (oldin->is_dir() && !srci->is_dir()) { | |
7392 | respond_to_request(mdr, -EISDIR); | |
7393 | return; | |
7394 | } | |
7395 | if (!oldin->is_dir() && srci->is_dir()) { | |
7396 | respond_to_request(mdr, -ENOTDIR); | |
7397 | return; | |
7398 | } | |
11fdf7f2 | 7399 | if (srci == oldin && !srcdir->inode->is_stray()) { |
181888fb FG |
7400 | respond_to_request(mdr, 0); // no-op. POSIX makes no sense. |
7401 | return; | |
7402 | } | |
7c673cae FG |
7403 | } |
7404 | } | |
7405 | ||
7406 | // -- some sanity checks -- | |
7407 | ||
7408 | // src+dest traces _must_ share a common ancestor for locking to prevent orphans | |
7409 | if (destpath.get_ino() != srcpath.get_ino() && | |
7410 | !(req->get_source().is_mds() && | |
7411 | MDS_INO_IS_MDSDIR(srcpath.get_ino()))) { // <-- mds 'rename' out of stray dir is ok! | |
7412 | CInode *srcbase = srctrace[0]->get_dir()->get_inode(); | |
7413 | CInode *destbase = desttrace[0]->get_dir()->get_inode(); | |
7414 | // ok, extend srctrace toward root until it is an ancestor of desttrace. | |
7415 | while (srcbase != destbase && | |
7416 | !srcbase->is_projected_ancestor_of(destbase)) { | |
7417 | CDentry *pdn = srcbase->get_projected_parent_dn(); | |
7418 | srctrace.insert(srctrace.begin(), pdn); | |
7419 | dout(10) << "rename prepending srctrace with " << *pdn << dendl; | |
7420 | srcbase = pdn->get_dir()->get_inode(); | |
7421 | } | |
7422 | ||
7423 | // then, extend destpath until it shares the same parent inode as srcpath. | |
7424 | while (destbase != srcbase) { | |
7425 | CDentry *pdn = destbase->get_projected_parent_dn(); | |
7426 | desttrace.insert(desttrace.begin(), pdn); | |
11fdf7f2 | 7427 | lov.add_rdlock(&pdn->lock); |
7c673cae FG |
7428 | dout(10) << "rename prepending desttrace with " << *pdn << dendl; |
7429 | destbase = pdn->get_dir()->get_inode(); | |
7430 | } | |
7431 | dout(10) << "rename src and dest traces now share common ancestor " << *destbase << dendl; | |
7432 | } | |
7433 | ||
7434 | // src == dest? | |
11fdf7f2 | 7435 | if (srcdir == destdir && srcdn->get_name() == destname) { |
7c673cae FG |
7436 | dout(7) << "rename src=dest, noop" << dendl; |
7437 | respond_to_request(mdr, 0); | |
7438 | return; | |
7439 | } | |
7440 | ||
7441 | // dest a child of src? | |
7442 | // e.g. mv /usr /usr/foo | |
7443 | CDentry *pdn = destdir->inode->get_projected_parent_dn(); | |
7444 | while (pdn) { | |
7445 | if (pdn == srcdn) { | |
7446 | dout(7) << "cannot rename item to be a child of itself" << dendl; | |
7447 | respond_to_request(mdr, -EINVAL); | |
7448 | return; | |
7449 | } | |
7450 | pdn = pdn->get_dir()->inode->parent; | |
7451 | } | |
7452 | ||
7453 | // is this a stray migration, reintegration or merge? (sanity checks!) | |
7454 | if (mdr->reqid.name.is_mds() && | |
7455 | !(MDS_INO_IS_MDSDIR(srcpath.get_ino()) && | |
7456 | MDS_INO_IS_MDSDIR(destpath.get_ino())) && | |
7457 | !(destdnl->is_remote() && | |
7458 | destdnl->get_remote_ino() == srci->ino())) { | |
7459 | respond_to_request(mdr, -EINVAL); // actually, this won't reply, but whatev. | |
7460 | return; | |
7461 | } | |
7462 | ||
11fdf7f2 | 7463 | bool linkmerge = srcdnl->get_inode() == destdnl->get_inode(); |
7c673cae FG |
7464 | if (linkmerge) |
7465 | dout(10) << " this is a link merge" << dendl; | |
7466 | ||
7467 | // -- create stray dentry? -- | |
7468 | CDentry *straydn = NULL; | |
7469 | if (destdnl->is_primary() && !linkmerge) { | |
7470 | straydn = prepare_stray_dentry(mdr, destdnl->get_inode()); | |
7471 | if (!straydn) | |
7472 | return; | |
7473 | dout(10) << " straydn is " << *straydn << dendl; | |
7474 | } else if (mdr->straydn) { | |
7475 | mdr->unpin(mdr->straydn); | |
7476 | mdr->straydn = NULL; | |
7477 | } | |
7478 | ||
7479 | // -- prepare witness list -- | |
7480 | /* | |
7481 | * NOTE: we use _all_ replicas as witnesses. | |
7482 | * this probably isn't totally necessary (esp for file renames), | |
7483 | * but if/when we change that, we have to make sure rejoin is | |
7484 | * sufficiently robust to handle strong rejoins from survivors | |
7485 | * with totally wrong dentry->inode linkage. | |
7486 | * (currently, it can ignore rename effects, because the resolve | |
7487 | * stage will sort them out.) | |
7488 | */ | |
7489 | set<mds_rank_t> witnesses = mdr->more()->extra_witnesses; | |
7490 | if (srcdn->is_auth()) | |
7491 | srcdn->list_replicas(witnesses); | |
7492 | else | |
7493 | witnesses.insert(srcdn->authority().first); | |
7494 | if (srcdnl->is_remote() && !srci->is_auth()) | |
7495 | witnesses.insert(srci->authority().first); | |
7496 | destdn->list_replicas(witnesses); | |
7497 | if (destdnl->is_remote() && !oldin->is_auth()) | |
7498 | witnesses.insert(oldin->authority().first); | |
7499 | dout(10) << " witnesses " << witnesses << ", have " << mdr->more()->witnessed << dendl; | |
7500 | ||
7501 | ||
7502 | // -- locks -- | |
7c673cae FG |
7503 | |
7504 | // srctrace items. this mirrors locks taken in rdlock_path_xlock_dentry | |
7505 | for (int i=0; i<(int)srctrace.size(); i++) | |
11fdf7f2 TL |
7506 | lov.add_rdlock(&srctrace[i]->lock); |
7507 | lov.add_xlock(&srcdn->lock); | |
7508 | mds_rank_t srcdirauth = srcdir->authority().first; | |
7c673cae FG |
7509 | if (srcdirauth != mds->get_nodeid()) { |
7510 | dout(10) << " will remote_wrlock srcdir scatterlocks on mds." << srcdirauth << dendl; | |
11fdf7f2 TL |
7511 | lov.add_remote_wrlock(&srcdir->inode->filelock, srcdirauth); |
7512 | lov.add_remote_wrlock(&srcdir->inode->nestlock, srcdirauth); | |
7c673cae | 7513 | if (srci->is_dir()) |
11fdf7f2 | 7514 | lov.add_rdlock(&srci->dirfragtreelock); |
7c673cae | 7515 | } else { |
11fdf7f2 TL |
7516 | lov.add_wrlock(&srcdir->inode->filelock); |
7517 | lov.add_wrlock(&srcdir->inode->nestlock); | |
7c673cae | 7518 | } |
11fdf7f2 | 7519 | mds->locker->include_snap_rdlocks(srcdir->inode, lov); |
7c673cae FG |
7520 | |
7521 | // straydn? | |
7522 | if (straydn) { | |
11fdf7f2 TL |
7523 | lov.add_wrlock(&straydn->get_dir()->inode->filelock); |
7524 | lov.add_wrlock(&straydn->get_dir()->inode->nestlock); | |
7525 | lov.add_xlock(&straydn->lock); | |
7c673cae FG |
7526 | } |
7527 | ||
7528 | // xlock versionlock on dentries if there are witnesses. | |
7529 | // replicas can't see projected dentry linkages, and will get | |
7530 | // confused if we try to pipeline things. | |
7531 | if (!witnesses.empty()) { | |
7532 | // take xlock on all projected ancestor dentries for srcdn and destdn. | |
7533 | // this ensures the srcdn and destdn can be traversed to by the witnesses. | |
7534 | for (int i= 0; i<(int)srctrace.size(); i++) { | |
7535 | if (srctrace[i]->is_auth() && srctrace[i]->is_projected()) | |
11fdf7f2 | 7536 | lov.add_xlock(&srctrace[i]->versionlock); |
7c673cae FG |
7537 | } |
7538 | for (int i=0; i<(int)desttrace.size(); i++) { | |
7539 | if (desttrace[i]->is_auth() && desttrace[i]->is_projected()) | |
11fdf7f2 | 7540 | lov.add_xlock(&desttrace[i]->versionlock); |
7c673cae FG |
7541 | } |
7542 | // xlock srci and oldin's primary dentries, so witnesses can call | |
7543 | // open_remote_ino() with 'want_locked=true' when the srcdn or destdn | |
7544 | // is traversed. | |
7545 | if (srcdnl->is_remote()) | |
11fdf7f2 | 7546 | lov.add_xlock(&srci->get_projected_parent_dn()->lock); |
7c673cae | 7547 | if (destdnl->is_remote()) |
11fdf7f2 | 7548 | lov.add_xlock(&oldin->get_projected_parent_dn()->lock); |
7c673cae FG |
7549 | } |
7550 | ||
7551 | // we need to update srci's ctime. xlock its least contended lock to do that... | |
11fdf7f2 TL |
7552 | lov.add_xlock(&srci->linklock); |
7553 | lov.add_xlock(&srci->snaplock); | |
7c673cae | 7554 | |
7c673cae | 7555 | if (oldin) { |
11fdf7f2 TL |
7556 | // xlock oldin (for nlink--) |
7557 | lov.add_xlock(&oldin->linklock); | |
7558 | lov.add_xlock(&oldin->snaplock); | |
7c673cae | 7559 | if (oldin->is_dir()) |
11fdf7f2 TL |
7560 | lov.add_rdlock(&oldin->filelock); // to verify it's empty |
7561 | } | |
7c673cae FG |
7562 | |
7563 | CInode *auth_pin_freeze = !srcdn->is_auth() && srcdnl->is_primary() ? srci : NULL; | |
11fdf7f2 | 7564 | if (!mds->locker->acquire_locks(mdr, lov, auth_pin_freeze)) |
7c673cae FG |
7565 | return; |
7566 | ||
11fdf7f2 TL |
7567 | if (linkmerge) |
7568 | ceph_assert(srcdir->inode->is_stray() && srcdnl->is_primary() && destdnl->is_remote()); | |
7569 | ||
7c673cae | 7570 | if ((!mdr->has_more() || mdr->more()->witnessed.empty())) { |
11fdf7f2 | 7571 | if (!check_access(mdr, srcdir->get_inode(), MAY_WRITE)) |
7c673cae FG |
7572 | return; |
7573 | ||
7574 | if (!check_access(mdr, destdn->get_dir()->get_inode(), MAY_WRITE)) | |
7575 | return; | |
7576 | ||
7577 | if (!check_fragment_space(mdr, destdn->get_dir())) | |
7578 | return; | |
7579 | ||
7580 | if (!check_access(mdr, srci, MAY_WRITE)) | |
7581 | return; | |
7582 | } | |
7583 | ||
7584 | // with read lock, really verify oldin is empty | |
7585 | if (oldin && | |
7586 | oldin->is_dir() && | |
7587 | _dir_is_nonempty(mdr, oldin)) { | |
7588 | respond_to_request(mdr, -ENOTEMPTY); | |
7589 | return; | |
7590 | } | |
7591 | ||
11fdf7f2 | 7592 | /* project_snaprealm_past_parent() will do this job |
7c673cae FG |
7593 | * |
7594 | // moving between snaprealms? | |
7595 | if (srcdnl->is_primary() && srci->is_multiversion() && !srci->snaprealm) { | |
7596 | SnapRealm *srcrealm = srci->find_snaprealm(); | |
7597 | SnapRealm *destrealm = destdn->get_dir()->inode->find_snaprealm(); | |
7598 | if (srcrealm != destrealm && | |
7599 | (srcrealm->get_newest_seq() + 1 > srcdn->first || | |
7600 | destrealm->get_newest_seq() + 1 > srcdn->first)) { | |
7601 | dout(10) << " renaming between snaprealms, creating snaprealm for " << *srci << dendl; | |
7602 | mdcache->snaprealm_create(mdr, srci); | |
7603 | return; | |
7604 | } | |
7605 | } | |
7606 | */ | |
7607 | ||
11fdf7f2 | 7608 | ceph_assert(g_conf()->mds_kill_rename_at != 1); |
7c673cae FG |
7609 | |
7610 | // -- open all srcdn inode frags, if any -- | |
7611 | // we need these open so that auth can properly delegate from inode to dirfrags | |
7612 | // after the inode is _ours_. | |
7613 | if (srcdnl->is_primary() && | |
7614 | !srcdn->is_auth() && | |
7615 | srci->is_dir()) { | |
7616 | dout(10) << "srci is remote dir, setting stickydirs and opening all frags" << dendl; | |
7617 | mdr->set_stickydirs(srci); | |
7618 | ||
11fdf7f2 TL |
7619 | frag_vec_t leaves; |
7620 | srci->dirfragtree.get_leaves(leaves); | |
7621 | for (const auto& leaf : leaves) { | |
7622 | CDir *dir = srci->get_dirfrag(leaf); | |
7c673cae | 7623 | if (!dir) { |
11fdf7f2 TL |
7624 | dout(10) << " opening " << leaf << " under " << *srci << dendl; |
7625 | mdcache->open_remote_dirfrag(srci, leaf, new C_MDS_RetryRequest(mdcache, mdr)); | |
7c673cae FG |
7626 | return; |
7627 | } | |
7628 | } | |
7629 | } | |
7630 | ||
11fdf7f2 TL |
7631 | // -- prepare snaprealm --- |
7632 | ||
7633 | if (linkmerge) { | |
7634 | if (!mdr->more()->srci_srnode && | |
7635 | srci->get_projected_inode()->nlink == 1 && | |
7636 | srci->is_projected_snaprealm_global()) { | |
7637 | sr_t *new_srnode = srci->prepare_new_srnode(0); | |
7638 | srci->record_snaprealm_parent_dentry(new_srnode, NULL, destdn, false); | |
7639 | ||
7640 | srci->clear_snaprealm_global(new_srnode); | |
7641 | mdr->more()->srci_srnode = new_srnode; | |
7642 | } | |
7643 | } else { | |
7644 | if (oldin && !mdr->more()->desti_srnode) { | |
7645 | if (oldin->is_projected_snaprealm_global()) { | |
7646 | sr_t *new_srnode = oldin->prepare_new_srnode(0); | |
7647 | oldin->record_snaprealm_parent_dentry(new_srnode, NULL, destdn, destdnl->is_primary()); | |
7648 | // dropping the last linkage or dropping the last remote linkage, | |
7649 | // detch the inode from global snaprealm | |
7650 | auto nlink = oldin->get_projected_inode()->nlink; | |
7651 | if (nlink == 1 || | |
7652 | (nlink == 2 && !destdnl->is_primary() && | |
7653 | !oldin->get_projected_parent_dir()->inode->is_stray())) | |
7654 | oldin->clear_snaprealm_global(new_srnode); | |
7655 | mdr->more()->desti_srnode = new_srnode; | |
7656 | } else if (destdnl->is_primary()) { | |
7657 | SnapRealm *dest_realm = destdir->inode->find_snaprealm(); | |
7658 | snapid_t follows = dest_realm->get_newest_seq(); | |
7659 | if (oldin->snaprealm || follows + 1 > oldin->get_oldest_snap()) { | |
7660 | sr_t *new_srnode = oldin->prepare_new_srnode(follows); | |
7661 | oldin->record_snaprealm_past_parent(new_srnode, straydn->get_dir()->inode->find_snaprealm()); | |
7662 | mdr->more()->desti_srnode = new_srnode; | |
7663 | } | |
7664 | } | |
7665 | } | |
7666 | if (!mdr->more()->srci_srnode) { | |
7667 | SnapRealm *dest_realm = destdir->inode->find_snaprealm(); | |
7668 | if (srci->is_projected_snaprealm_global()) { | |
7669 | sr_t *new_srnode = srci->prepare_new_srnode(0); | |
7670 | srci->record_snaprealm_parent_dentry(new_srnode, dest_realm, srcdn, srcdnl->is_primary()); | |
7671 | mdr->more()->srci_srnode = new_srnode; | |
7672 | } else if (srcdnl->is_primary()) { | |
7673 | SnapRealm *src_realm = srcdir->inode->find_snaprealm(); | |
7674 | snapid_t follows = src_realm->get_newest_seq(); | |
7675 | if (src_realm != dest_realm && | |
7676 | (srci->snaprealm || follows + 1 > srci->get_oldest_snap())) { | |
7677 | sr_t *new_srnode = srci->prepare_new_srnode(follows); | |
7678 | srci->record_snaprealm_past_parent(new_srnode, dest_realm); | |
7679 | mdr->more()->srci_srnode = new_srnode; | |
7680 | } | |
7681 | } | |
7682 | } | |
7683 | } | |
7684 | ||
7c673cae FG |
7685 | // -- prepare witnesses -- |
7686 | ||
7687 | // do srcdn auth last | |
7688 | mds_rank_t last = MDS_RANK_NONE; | |
7689 | if (!srcdn->is_auth()) { | |
7690 | last = srcdn->authority().first; | |
7691 | mdr->more()->srcdn_auth_mds = last; | |
7692 | // ask auth of srci to mark srci as ambiguous auth if more than two MDS | |
7693 | // are involved in the rename operation. | |
7694 | if (srcdnl->is_primary() && !mdr->more()->is_ambiguous_auth) { | |
7695 | dout(10) << " preparing ambiguous auth for srci" << dendl; | |
11fdf7f2 TL |
7696 | ceph_assert(mdr->more()->is_remote_frozen_authpin); |
7697 | ceph_assert(mdr->more()->rename_inode == srci); | |
7c673cae FG |
7698 | _rename_prepare_witness(mdr, last, witnesses, srctrace, desttrace, straydn); |
7699 | return; | |
7700 | } | |
7701 | } | |
7702 | ||
7703 | for (set<mds_rank_t>::iterator p = witnesses.begin(); | |
7704 | p != witnesses.end(); | |
7705 | ++p) { | |
7706 | if (*p == last) continue; // do it last! | |
7707 | if (mdr->more()->witnessed.count(*p)) { | |
7708 | dout(10) << " already witnessed by mds." << *p << dendl; | |
7709 | } else if (mdr->more()->waiting_on_slave.count(*p)) { | |
7710 | dout(10) << " already waiting on witness mds." << *p << dendl; | |
7711 | } else { | |
7712 | if (!_rename_prepare_witness(mdr, *p, witnesses, srctrace, desttrace, straydn)) | |
7713 | return; | |
7714 | } | |
7715 | } | |
7716 | if (!mdr->more()->waiting_on_slave.empty()) | |
7717 | return; // we're waiting for a witness. | |
7718 | ||
7719 | if (last != MDS_RANK_NONE && mdr->more()->witnessed.count(last) == 0) { | |
7720 | dout(10) << " preparing last witness (srcdn auth)" << dendl; | |
11fdf7f2 | 7721 | ceph_assert(mdr->more()->waiting_on_slave.count(last) == 0); |
7c673cae FG |
7722 | _rename_prepare_witness(mdr, last, witnesses, srctrace, desttrace, straydn); |
7723 | return; | |
7724 | } | |
7725 | ||
7726 | // test hack: bail after slave does prepare, so we can verify it's _live_ rollback. | |
7727 | if (!mdr->more()->slaves.empty() && !srci->is_dir()) | |
11fdf7f2 | 7728 | ceph_assert(g_conf()->mds_kill_rename_at != 3); |
7c673cae | 7729 | if (!mdr->more()->slaves.empty() && srci->is_dir()) |
11fdf7f2 | 7730 | ceph_assert(g_conf()->mds_kill_rename_at != 4); |
7c673cae FG |
7731 | |
7732 | // -- declare now -- | |
7733 | mdr->set_mds_stamp(ceph_clock_now()); | |
7734 | ||
7735 | // -- prepare journal entry -- | |
7736 | mdr->ls = mdlog->get_current_segment(); | |
7737 | EUpdate *le = new EUpdate(mdlog, "rename"); | |
7738 | mdlog->start_entry(le); | |
7739 | le->metablob.add_client_req(mdr->reqid, mdr->client_request->get_oldest_client_tid()); | |
7740 | if (!mdr->more()->witnessed.empty()) { | |
7741 | dout(20) << " noting uncommitted_slaves " << mdr->more()->witnessed << dendl; | |
7742 | ||
7743 | le->reqid = mdr->reqid; | |
7744 | le->had_slaves = true; | |
7745 | ||
7746 | mdcache->add_uncommitted_master(mdr->reqid, mdr->ls, mdr->more()->witnessed); | |
7747 | // no need to send frozen auth pin to recovring auth MDS of srci | |
7748 | mdr->more()->is_remote_frozen_authpin = false; | |
7749 | } | |
7750 | ||
7751 | _rename_prepare(mdr, &le->metablob, &le->client_map, srcdn, destdn, straydn); | |
7752 | if (le->client_map.length()) | |
7753 | le->cmapv = mds->sessionmap.get_projected(); | |
7754 | ||
7755 | // -- commit locally -- | |
7756 | C_MDS_rename_finish *fin = new C_MDS_rename_finish(this, mdr, srcdn, destdn, straydn); | |
7757 | ||
7758 | journal_and_reply(mdr, srci, destdn, le, fin); | |
7759 | } | |
7760 | ||
7761 | ||
7762 | void Server::_rename_finish(MDRequestRef& mdr, CDentry *srcdn, CDentry *destdn, CDentry *straydn) | |
7763 | { | |
7764 | dout(10) << "_rename_finish " << *mdr << dendl; | |
7765 | ||
7766 | if (!mdr->more()->witnessed.empty()) | |
7767 | mdcache->logged_master_update(mdr->reqid); | |
7768 | ||
7769 | // apply | |
7770 | _rename_apply(mdr, srcdn, destdn, straydn); | |
7771 | ||
7772 | mdcache->send_dentry_link(destdn, mdr); | |
7773 | ||
7774 | CDentry::linkage_t *destdnl = destdn->get_linkage(); | |
7775 | CInode *in = destdnl->get_inode(); | |
7776 | bool need_eval = mdr->more()->cap_imports.count(in); | |
7777 | ||
7778 | // test hack: test slave commit | |
7779 | if (!mdr->more()->slaves.empty() && !in->is_dir()) | |
11fdf7f2 | 7780 | ceph_assert(g_conf()->mds_kill_rename_at != 5); |
7c673cae | 7781 | if (!mdr->more()->slaves.empty() && in->is_dir()) |
11fdf7f2 | 7782 | ceph_assert(g_conf()->mds_kill_rename_at != 6); |
7c673cae FG |
7783 | |
7784 | // bump popularity | |
11fdf7f2 | 7785 | mds->balancer->hit_dir(srcdn->get_dir(), META_POP_IWR); |
7c673cae | 7786 | if (destdnl->is_remote() && in->is_auth()) |
11fdf7f2 | 7787 | mds->balancer->hit_inode(in, META_POP_IWR); |
7c673cae FG |
7788 | |
7789 | // did we import srci? if so, explicitly ack that import that, before we unlock and reply. | |
7790 | ||
11fdf7f2 | 7791 | ceph_assert(g_conf()->mds_kill_rename_at != 7); |
7c673cae FG |
7792 | |
7793 | // reply | |
7794 | respond_to_request(mdr, 0); | |
7795 | ||
7796 | if (need_eval) | |
7797 | mds->locker->eval(in, CEPH_CAP_LOCKS, true); | |
7798 | ||
7799 | // clean up? | |
7800 | // respond_to_request() drops locks. So stray reintegration can race with us. | |
7801 | if (straydn && !straydn->get_projected_linkage()->is_null()) { | |
7802 | mdcache->notify_stray(straydn); | |
7803 | } | |
7804 | } | |
7805 | ||
7806 | ||
7807 | ||
7808 | // helpers | |
7809 | ||
7810 | bool Server::_rename_prepare_witness(MDRequestRef& mdr, mds_rank_t who, set<mds_rank_t> &witnesse, | |
7811 | vector<CDentry*>& srctrace, vector<CDentry*>& dsttrace, CDentry *straydn) | |
7812 | { | |
7813 | if (mds->is_cluster_degraded() && | |
7814 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(who)) { | |
7815 | dout(10) << "_rename_prepare_witness mds." << who << " is not active" << dendl; | |
7816 | if (mdr->more()->waiting_on_slave.empty()) | |
7817 | mds->wait_for_active_peer(who, new C_MDS_RetryRequest(mdcache, mdr)); | |
7818 | return false; | |
7819 | } | |
7820 | ||
7821 | dout(10) << "_rename_prepare_witness mds." << who << dendl; | |
11fdf7f2 | 7822 | auto req = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RENAMEPREP); |
7c673cae FG |
7823 | |
7824 | req->srcdnpath = filepath(srctrace.front()->get_dir()->ino()); | |
7825 | for (auto dn : srctrace) | |
94b18763 | 7826 | req->srcdnpath.push_dentry(dn->get_name()); |
7c673cae FG |
7827 | req->destdnpath = filepath(dsttrace.front()->get_dir()->ino()); |
7828 | for (auto dn : dsttrace) | |
94b18763 | 7829 | req->destdnpath.push_dentry(dn->get_name()); |
7c673cae | 7830 | if (straydn) |
11fdf7f2 TL |
7831 | mdcache->replicate_stray(straydn, who, req->straybl); |
7832 | ||
7833 | if (mdr->more()->srci_srnode) | |
7834 | encode(*mdr->more()->srci_srnode, req->srci_snapbl); | |
7835 | if (mdr->more()->desti_srnode) | |
7836 | encode(*mdr->more()->desti_srnode, req->desti_snapbl); | |
31f18b77 FG |
7837 | |
7838 | req->srcdn_auth = mdr->more()->srcdn_auth_mds; | |
7c673cae FG |
7839 | |
7840 | // srcdn auth will verify our current witness list is sufficient | |
7841 | req->witnesses = witnesse; | |
7842 | ||
7843 | req->op_stamp = mdr->get_op_stamp(); | |
7844 | mds->send_message_mds(req, who); | |
7845 | ||
11fdf7f2 | 7846 | ceph_assert(mdr->more()->waiting_on_slave.count(who) == 0); |
7c673cae FG |
7847 | mdr->more()->waiting_on_slave.insert(who); |
7848 | return true; | |
7849 | } | |
7850 | ||
7851 | version_t Server::_rename_prepare_import(MDRequestRef& mdr, CDentry *srcdn, bufferlist *client_map_bl) | |
7852 | { | |
7853 | version_t oldpv = mdr->more()->inode_import_v; | |
7854 | ||
7855 | CDentry::linkage_t *srcdnl = srcdn->get_linkage(); | |
7856 | ||
7857 | /* import node */ | |
11fdf7f2 | 7858 | auto blp = mdr->more()->inode_import.cbegin(); |
7c673cae FG |
7859 | |
7860 | // imported caps | |
28e407b8 | 7861 | map<client_t,entity_inst_t> client_map; |
11fdf7f2 | 7862 | map<client_t, client_metadata_t> client_metadata_map; |
28e407b8 | 7863 | decode(client_map, blp); |
11fdf7f2 TL |
7864 | decode(client_metadata_map, blp); |
7865 | prepare_force_open_sessions(client_map, client_metadata_map, | |
7866 | mdr->more()->imported_session_map); | |
28e407b8 | 7867 | encode(client_map, *client_map_bl, mds->mdsmap->get_up_features()); |
11fdf7f2 | 7868 | encode(client_metadata_map, *client_map_bl); |
7c673cae FG |
7869 | |
7870 | list<ScatterLock*> updated_scatterlocks; | |
7871 | mdcache->migrator->decode_import_inode(srcdn, blp, srcdn->authority().first, mdr->ls, | |
7872 | mdr->more()->cap_imports, updated_scatterlocks); | |
7873 | ||
7874 | // hack: force back to !auth and clean, temporarily | |
7875 | srcdnl->get_inode()->state_clear(CInode::STATE_AUTH); | |
7876 | srcdnl->get_inode()->mark_clean(); | |
7877 | ||
7878 | return oldpv; | |
7879 | } | |
7880 | ||
7881 | bool Server::_need_force_journal(CInode *diri, bool empty) | |
7882 | { | |
11fdf7f2 TL |
7883 | std::vector<CDir*> dirs; |
7884 | diri->get_dirfrags(dirs); | |
7c673cae FG |
7885 | |
7886 | bool force_journal = false; | |
7887 | if (empty) { | |
11fdf7f2 TL |
7888 | for (const auto& dir : dirs) { |
7889 | if (dir->is_subtree_root() && dir->get_dir_auth().first == mds->get_nodeid()) { | |
7890 | dout(10) << " frag " << dir->get_frag() << " is auth subtree dirfrag, will force journal" << dendl; | |
7c673cae FG |
7891 | force_journal = true; |
7892 | break; | |
7893 | } else | |
11fdf7f2 | 7894 | dout(20) << " frag " << dir->get_frag() << " is not auth subtree dirfrag" << dendl; |
7c673cae FG |
7895 | } |
7896 | } else { | |
7897 | // see if any children of our frags are auth subtrees. | |
11fdf7f2 TL |
7898 | std::vector<CDir*> subtrees; |
7899 | mdcache->get_subtrees(subtrees); | |
7900 | dout(10) << " subtrees " << subtrees << " frags " << dirs << dendl; | |
7901 | for (const auto& dir : dirs) { | |
7902 | for (const auto& subtree : subtrees) { | |
7903 | if (dir->contains(subtree)) { | |
7904 | if (subtree->get_dir_auth().first == mds->get_nodeid()) { | |
7905 | dout(10) << " frag " << dir->get_frag() << " contains (maybe) auth subtree, will force journal " | |
7906 | << *subtree << dendl; | |
7c673cae FG |
7907 | force_journal = true; |
7908 | break; | |
7909 | } else | |
11fdf7f2 | 7910 | dout(20) << " frag " << dir->get_frag() << " contains but isn't auth for " << *subtree << dendl; |
7c673cae | 7911 | } else |
11fdf7f2 | 7912 | dout(20) << " frag " << dir->get_frag() << " does not contain " << *subtree << dendl; |
7c673cae FG |
7913 | } |
7914 | if (force_journal) | |
7915 | break; | |
7916 | } | |
7917 | } | |
7918 | return force_journal; | |
7919 | } | |
7920 | ||
7921 | void Server::_rename_prepare(MDRequestRef& mdr, | |
7922 | EMetaBlob *metablob, bufferlist *client_map_bl, | |
7923 | CDentry *srcdn, CDentry *destdn, CDentry *straydn) | |
7924 | { | |
7925 | dout(10) << "_rename_prepare " << *mdr << " " << *srcdn << " " << *destdn << dendl; | |
7926 | if (straydn) | |
7927 | dout(10) << " straydn " << *straydn << dendl; | |
7928 | ||
7929 | CDentry::linkage_t *srcdnl = srcdn->get_projected_linkage(); | |
7930 | CDentry::linkage_t *destdnl = destdn->get_projected_linkage(); | |
7931 | CInode *srci = srcdnl->get_inode(); | |
7932 | CInode *oldin = destdnl->get_inode(); | |
7933 | ||
7934 | // primary+remote link merge? | |
11fdf7f2 TL |
7935 | bool linkmerge = (srci == oldin); |
7936 | if (linkmerge) | |
7937 | ceph_assert(srcdnl->is_primary() && destdnl->is_remote()); | |
7c673cae FG |
7938 | bool silent = srcdn->get_dir()->inode->is_stray(); |
7939 | ||
7940 | bool force_journal_dest = false; | |
7941 | if (srci->is_dir() && !destdn->is_auth()) { | |
7942 | if (srci->is_auth()) { | |
7943 | // if we are auth for srci and exporting it, force journal because journal replay needs | |
7944 | // the source inode to create auth subtrees. | |
7945 | dout(10) << " we are exporting srci, will force journal destdn" << dendl; | |
7946 | force_journal_dest = true; | |
7947 | } else | |
7948 | force_journal_dest = _need_force_journal(srci, false); | |
7949 | } | |
7950 | ||
7951 | bool force_journal_stray = false; | |
7952 | if (oldin && oldin->is_dir() && straydn && !straydn->is_auth()) | |
7953 | force_journal_stray = _need_force_journal(oldin, true); | |
7954 | ||
7955 | if (linkmerge) | |
7956 | dout(10) << " merging remote and primary links to the same inode" << dendl; | |
7957 | if (silent) | |
7958 | dout(10) << " reintegrating stray; will avoid changing nlink or dir mtime" << dendl; | |
7959 | if (force_journal_dest) | |
7960 | dout(10) << " forcing journal destdn because we (will) have auth subtrees nested beneath it" << dendl; | |
7961 | if (force_journal_stray) | |
7962 | dout(10) << " forcing journal straydn because we (will) have auth subtrees nested beneath it" << dendl; | |
7963 | ||
7964 | if (srci->is_dir() && (destdn->is_auth() || force_journal_dest)) { | |
7965 | dout(10) << " noting renamed dir ino " << srci->ino() << " in metablob" << dendl; | |
7966 | metablob->renamed_dirino = srci->ino(); | |
7967 | } else if (oldin && oldin->is_dir() && force_journal_stray) { | |
7968 | dout(10) << " noting rename target dir " << oldin->ino() << " in metablob" << dendl; | |
7969 | metablob->renamed_dirino = oldin->ino(); | |
7970 | } | |
7971 | ||
7972 | // prepare | |
94b18763 FG |
7973 | CInode::mempool_inode *spi = 0; // renamed inode |
7974 | CInode::mempool_inode *tpi = 0; // target/overwritten inode | |
7c673cae FG |
7975 | |
7976 | // target inode | |
7977 | if (!linkmerge) { | |
7978 | if (destdnl->is_primary()) { | |
11fdf7f2 | 7979 | ceph_assert(straydn); // moving to straydn. |
7c673cae FG |
7980 | // link--, and move. |
7981 | if (destdn->is_auth()) { | |
94b18763 FG |
7982 | auto &pi= oldin->project_inode(); //project_snaprealm |
7983 | pi.inode.version = straydn->pre_dirty(pi.inode.version); | |
7984 | pi.inode.update_backtrace(); | |
7985 | tpi = &pi.inode; | |
7c673cae FG |
7986 | } |
7987 | straydn->push_projected_linkage(oldin); | |
7988 | } else if (destdnl->is_remote()) { | |
7989 | // nlink-- targeti | |
7990 | if (oldin->is_auth()) { | |
94b18763 FG |
7991 | auto &pi = oldin->project_inode(); |
7992 | pi.inode.version = oldin->pre_dirty(); | |
7993 | tpi = &pi.inode; | |
7c673cae FG |
7994 | } |
7995 | } | |
7996 | } | |
7997 | ||
7998 | // dest | |
7999 | if (srcdnl->is_remote()) { | |
8000 | if (!linkmerge) { | |
8001 | // destdn | |
8002 | if (destdn->is_auth()) | |
8003 | mdr->more()->pvmap[destdn] = destdn->pre_dirty(); | |
8004 | destdn->push_projected_linkage(srcdnl->get_remote_ino(), srcdnl->get_remote_d_type()); | |
8005 | // srci | |
8006 | if (srci->is_auth()) { | |
94b18763 FG |
8007 | auto &pi = srci->project_inode(); |
8008 | pi.inode.version = srci->pre_dirty(); | |
8009 | spi = &pi.inode; | |
7c673cae FG |
8010 | } |
8011 | } else { | |
8012 | dout(10) << " will merge remote onto primary link" << dendl; | |
8013 | if (destdn->is_auth()) { | |
94b18763 FG |
8014 | auto &pi = oldin->project_inode(); |
8015 | pi.inode.version = mdr->more()->pvmap[destdn] = destdn->pre_dirty(oldin->inode.version); | |
8016 | spi = &pi.inode; | |
7c673cae FG |
8017 | } |
8018 | } | |
8019 | } else { // primary | |
8020 | if (destdn->is_auth()) { | |
8021 | version_t oldpv; | |
8022 | if (srcdn->is_auth()) | |
8023 | oldpv = srci->get_projected_version(); | |
8024 | else { | |
8025 | oldpv = _rename_prepare_import(mdr, srcdn, client_map_bl); | |
8026 | ||
8027 | // note which dirfrags have child subtrees in the journal | |
8028 | // event, so that we can open those (as bounds) during replay. | |
8029 | if (srci->is_dir()) { | |
8030 | list<CDir*> ls; | |
8031 | srci->get_dirfrags(ls); | |
8032 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
8033 | CDir *dir = *p; | |
8034 | if (!dir->is_auth()) | |
8035 | metablob->renamed_dir_frags.push_back(dir->get_frag()); | |
8036 | } | |
8037 | dout(10) << " noting renamed dir open frags " << metablob->renamed_dir_frags << dendl; | |
8038 | } | |
8039 | } | |
94b18763 | 8040 | auto &pi = srci->project_inode(); // project snaprealm if srcdnl->is_primary |
7c673cae | 8041 | // & srcdnl->snaprealm |
94b18763 FG |
8042 | pi.inode.version = mdr->more()->pvmap[destdn] = destdn->pre_dirty(oldpv); |
8043 | pi.inode.update_backtrace(); | |
8044 | spi = &pi.inode; | |
7c673cae FG |
8045 | } |
8046 | destdn->push_projected_linkage(srci); | |
8047 | } | |
8048 | ||
8049 | // src | |
8050 | if (srcdn->is_auth()) | |
8051 | mdr->more()->pvmap[srcdn] = srcdn->pre_dirty(); | |
8052 | srcdn->push_projected_linkage(); // push null linkage | |
8053 | ||
8054 | if (!silent) { | |
94b18763 | 8055 | if (spi) { |
91327a77 AA |
8056 | spi->ctime = mdr->get_op_stamp(); |
8057 | if (mdr->get_op_stamp() > spi->rstat.rctime) | |
8058 | spi->rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 8059 | spi->change_attr++; |
7c673cae | 8060 | if (linkmerge) |
94b18763 | 8061 | spi->nlink--; |
7c673cae FG |
8062 | } |
8063 | if (tpi) { | |
91327a77 AA |
8064 | tpi->ctime = mdr->get_op_stamp(); |
8065 | if (mdr->get_op_stamp() > tpi->rstat.rctime) | |
8066 | tpi->rstat.rctime = mdr->get_op_stamp(); | |
7c673cae | 8067 | tpi->change_attr++; |
94b18763 FG |
8068 | { |
8069 | std::string t; | |
8070 | destdn->make_path_string(t, true); | |
11fdf7f2 | 8071 | tpi->stray_prior_path = std::move(t); |
94b18763 | 8072 | } |
7c673cae FG |
8073 | tpi->nlink--; |
8074 | if (tpi->nlink == 0) | |
8075 | oldin->state_set(CInode::STATE_ORPHAN); | |
8076 | } | |
8077 | } | |
8078 | ||
8079 | // prepare nesting, mtime updates | |
8080 | int predirty_dir = silent ? 0:PREDIRTY_DIR; | |
8081 | ||
8082 | // guarantee stray dir is processed first during journal replay. unlink the old inode, | |
8083 | // then link the source inode to destdn | |
8084 | if (destdnl->is_primary()) { | |
11fdf7f2 | 8085 | ceph_assert(straydn); |
7c673cae FG |
8086 | if (straydn->is_auth()) { |
8087 | metablob->add_dir_context(straydn->get_dir()); | |
8088 | metablob->add_dir(straydn->get_dir(), true); | |
8089 | } | |
8090 | } | |
8091 | ||
8092 | // sub off target | |
8093 | if (destdn->is_auth() && !destdnl->is_null()) { | |
8094 | mdcache->predirty_journal_parents(mdr, metablob, oldin, destdn->get_dir(), | |
8095 | (destdnl->is_primary() ? PREDIRTY_PRIMARY:0)|predirty_dir, -1); | |
224ce89b | 8096 | if (destdnl->is_primary()) { |
11fdf7f2 | 8097 | ceph_assert(straydn); |
7c673cae FG |
8098 | mdcache->predirty_journal_parents(mdr, metablob, oldin, straydn->get_dir(), |
8099 | PREDIRTY_PRIMARY|PREDIRTY_DIR, 1); | |
224ce89b | 8100 | } |
7c673cae FG |
8101 | } |
8102 | ||
8103 | // move srcdn | |
8104 | int predirty_primary = (srcdnl->is_primary() && srcdn->get_dir() != destdn->get_dir()) ? PREDIRTY_PRIMARY:0; | |
8105 | int flags = predirty_dir | predirty_primary; | |
8106 | if (srcdn->is_auth()) | |
8107 | mdcache->predirty_journal_parents(mdr, metablob, srci, srcdn->get_dir(), PREDIRTY_SHALLOW|flags, -1); | |
8108 | if (destdn->is_auth()) | |
8109 | mdcache->predirty_journal_parents(mdr, metablob, srci, destdn->get_dir(), flags, 1); | |
8110 | ||
7c673cae FG |
8111 | // add it all to the metablob |
8112 | // target inode | |
8113 | if (!linkmerge) { | |
8114 | if (destdnl->is_primary()) { | |
11fdf7f2 | 8115 | ceph_assert(straydn); |
7c673cae FG |
8116 | if (destdn->is_auth()) { |
8117 | // project snaprealm, too | |
11fdf7f2 TL |
8118 | if (auto& desti_srnode = mdr->more()->desti_srnode) { |
8119 | oldin->project_snaprealm(desti_srnode); | |
8120 | if (tpi->nlink == 0) | |
8121 | ceph_assert(!desti_srnode->is_parent_global()); | |
8122 | desti_srnode = NULL; | |
8123 | } | |
8124 | straydn->first = mdcache->get_global_snaprealm()->get_newest_seq() + 1; | |
7c673cae FG |
8125 | metablob->add_primary_dentry(straydn, oldin, true, true); |
8126 | } else if (force_journal_stray) { | |
8127 | dout(10) << " forced journaling straydn " << *straydn << dendl; | |
8128 | metablob->add_dir_context(straydn->get_dir()); | |
8129 | metablob->add_primary_dentry(straydn, oldin, true); | |
8130 | } | |
8131 | } else if (destdnl->is_remote()) { | |
8132 | if (oldin->is_auth()) { | |
11fdf7f2 TL |
8133 | sr_t *new_srnode = NULL; |
8134 | if (mdr->slave_request) { | |
8135 | if (mdr->slave_request->desti_snapbl.length() > 0) { | |
8136 | new_srnode = new sr_t(); | |
8137 | auto p = mdr->slave_request->desti_snapbl.cbegin(); | |
8138 | decode(*new_srnode, p); | |
8139 | } | |
8140 | } else if (auto& desti_srnode = mdr->more()->desti_srnode) { | |
8141 | new_srnode = desti_srnode; | |
8142 | desti_srnode = NULL; | |
8143 | } | |
8144 | if (new_srnode) { | |
8145 | oldin->project_snaprealm(new_srnode); | |
8146 | if (tpi->nlink == 0) | |
8147 | ceph_assert(!new_srnode->is_parent_global()); | |
8148 | } | |
7c673cae FG |
8149 | // auth for targeti |
8150 | metablob->add_dir_context(oldin->get_projected_parent_dir()); | |
8151 | mdcache->journal_cow_dentry(mdr.get(), metablob, oldin->get_projected_parent_dn(), | |
8152 | CEPH_NOSNAP, 0, destdnl); | |
8153 | metablob->add_primary_dentry(oldin->get_projected_parent_dn(), oldin, true); | |
8154 | } | |
8155 | } | |
8156 | } | |
8157 | ||
8158 | // dest | |
8159 | if (srcdnl->is_remote()) { | |
11fdf7f2 TL |
8160 | ceph_assert(!linkmerge); |
8161 | if (destdn->is_auth() && !destdnl->is_null()) | |
8162 | mdcache->journal_cow_dentry(mdr.get(), metablob, destdn, CEPH_NOSNAP, 0, destdnl); | |
8163 | else | |
8164 | destdn->first = mdcache->get_global_snaprealm()->get_newest_seq() + 1; | |
7c673cae | 8165 | |
11fdf7f2 TL |
8166 | if (destdn->is_auth()) |
8167 | metablob->add_remote_dentry(destdn, true, srcdnl->get_remote_ino(), srcdnl->get_remote_d_type()); | |
8168 | ||
8169 | if (srci->is_auth() ) { // it's remote | |
8170 | if (mdr->slave_request) { | |
8171 | if (mdr->slave_request->srci_snapbl.length() > 0) { | |
8172 | sr_t *new_srnode = new sr_t(); | |
8173 | auto p = mdr->slave_request->srci_snapbl.cbegin(); | |
8174 | decode(*new_srnode, p); | |
8175 | srci->project_snaprealm(new_srnode); | |
8176 | } | |
8177 | } else if (auto& srci_srnode = mdr->more()->srci_srnode) { | |
8178 | srci->project_snaprealm(srci_srnode); | |
8179 | srci_srnode = NULL; | |
7c673cae | 8180 | } |
7c673cae | 8181 | |
11fdf7f2 TL |
8182 | CDentry *srci_pdn = srci->get_projected_parent_dn(); |
8183 | metablob->add_dir_context(srci_pdn->get_dir()); | |
8184 | mdcache->journal_cow_dentry(mdr.get(), metablob, srci_pdn, CEPH_NOSNAP, 0, srcdnl); | |
8185 | metablob->add_primary_dentry(srci_pdn, srci, true); | |
7c673cae FG |
8186 | } |
8187 | } else if (srcdnl->is_primary()) { | |
8188 | // project snap parent update? | |
11fdf7f2 TL |
8189 | if (destdn->is_auth()) { |
8190 | if (auto& srci_srnode = mdr->more()->srci_srnode) { | |
8191 | srci->project_snaprealm(srci_srnode); | |
8192 | srci_srnode = NULL; | |
8193 | } | |
8194 | } | |
7c673cae FG |
8195 | |
8196 | if (destdn->is_auth() && !destdnl->is_null()) | |
8197 | mdcache->journal_cow_dentry(mdr.get(), metablob, destdn, CEPH_NOSNAP, 0, destdnl); | |
11fdf7f2 TL |
8198 | |
8199 | destdn->first = mdcache->get_global_snaprealm()->get_newest_seq() + 1; | |
7c673cae FG |
8200 | |
8201 | if (destdn->is_auth()) | |
8202 | metablob->add_primary_dentry(destdn, srci, true, true); | |
8203 | else if (force_journal_dest) { | |
8204 | dout(10) << " forced journaling destdn " << *destdn << dendl; | |
8205 | metablob->add_dir_context(destdn->get_dir()); | |
8206 | metablob->add_primary_dentry(destdn, srci, true); | |
8207 | if (srcdn->is_auth() && srci->is_dir()) { | |
8208 | // journal new subtrees root dirfrags | |
8209 | list<CDir*> ls; | |
8210 | srci->get_dirfrags(ls); | |
8211 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
8212 | CDir *dir = *p; | |
8213 | if (dir->is_auth()) | |
8214 | metablob->add_dir(dir, true); | |
8215 | } | |
8216 | } | |
8217 | } | |
8218 | } | |
8219 | ||
8220 | // src | |
8221 | if (srcdn->is_auth()) { | |
8222 | dout(10) << " journaling srcdn " << *srcdn << dendl; | |
8223 | mdcache->journal_cow_dentry(mdr.get(), metablob, srcdn, CEPH_NOSNAP, 0, srcdnl); | |
8224 | // also journal the inode in case we need do slave rename rollback. It is Ok to add | |
8225 | // both primary and NULL dentries. Because during journal replay, null dentry is | |
8226 | // processed after primary dentry. | |
8227 | if (srcdnl->is_primary() && !srci->is_dir() && !destdn->is_auth()) | |
8228 | metablob->add_primary_dentry(srcdn, srci, true); | |
8229 | metablob->add_null_dentry(srcdn, true); | |
8230 | } else | |
8231 | dout(10) << " NOT journaling srcdn " << *srcdn << dendl; | |
8232 | ||
8233 | // make renamed inode first track the dn | |
11fdf7f2 TL |
8234 | if (srcdnl->is_primary() && destdn->is_auth()) { |
8235 | ceph_assert(srci->first <= destdn->first); | |
8236 | srci->first = destdn->first; | |
8237 | } | |
8238 | // make stray inode first track the straydn | |
8239 | if (straydn && straydn->is_auth()) { | |
8240 | ceph_assert(oldin->first <= straydn->first); | |
8241 | oldin->first = straydn->first; | |
8242 | } | |
7c673cae | 8243 | |
224ce89b | 8244 | if (oldin && oldin->is_dir()) { |
11fdf7f2 | 8245 | ceph_assert(straydn); |
7c673cae | 8246 | mdcache->project_subtree_rename(oldin, destdn->get_dir(), straydn->get_dir()); |
224ce89b | 8247 | } |
7c673cae FG |
8248 | if (srci->is_dir()) |
8249 | mdcache->project_subtree_rename(srci, srcdn->get_dir(), destdn->get_dir()); | |
8250 | ||
8251 | } | |
8252 | ||
8253 | ||
8254 | void Server::_rename_apply(MDRequestRef& mdr, CDentry *srcdn, CDentry *destdn, CDentry *straydn) | |
8255 | { | |
8256 | dout(10) << "_rename_apply " << *mdr << " " << *srcdn << " " << *destdn << dendl; | |
8257 | dout(10) << " pvs " << mdr->more()->pvmap << dendl; | |
8258 | ||
8259 | CDentry::linkage_t *srcdnl = srcdn->get_linkage(); | |
8260 | CDentry::linkage_t *destdnl = destdn->get_linkage(); | |
8261 | ||
8262 | CInode *oldin = destdnl->get_inode(); | |
7c673cae FG |
8263 | |
8264 | // primary+remote link merge? | |
11fdf7f2 TL |
8265 | bool linkmerge = (srcdnl->get_inode() == oldin); |
8266 | if (linkmerge) | |
8267 | ceph_assert(srcdnl->is_primary() || destdnl->is_remote()); | |
8268 | ||
8269 | bool new_in_snaprealm = false; | |
8270 | bool new_oldin_snaprealm = false; | |
7c673cae FG |
8271 | |
8272 | // target inode | |
8273 | if (!linkmerge) { | |
8274 | if (destdnl->is_primary()) { | |
11fdf7f2 | 8275 | ceph_assert(straydn); |
7c673cae | 8276 | dout(10) << "straydn is " << *straydn << dendl; |
11fdf7f2 TL |
8277 | |
8278 | // if there is newly created snaprealm, need to split old snaprealm's | |
8279 | // inodes_with_caps. So pop snaprealm before linkage changes. | |
8280 | if (destdn->is_auth()) { | |
8281 | bool hadrealm = (oldin->snaprealm ? true : false); | |
8282 | oldin->early_pop_projected_snaprealm(); | |
8283 | new_oldin_snaprealm = (oldin->snaprealm && !hadrealm); | |
8284 | } else { | |
8285 | ceph_assert(mdr->slave_request); | |
8286 | if (mdr->slave_request->desti_snapbl.length()) { | |
8287 | new_oldin_snaprealm = !oldin->snaprealm; | |
8288 | oldin->decode_snap_blob(mdr->slave_request->desti_snapbl); | |
8289 | ceph_assert(oldin->snaprealm); | |
8290 | ceph_assert(oldin->snaprealm->have_past_parents_open()); | |
8291 | } | |
8292 | } | |
8293 | ||
31f18b77 | 8294 | destdn->get_dir()->unlink_inode(destdn, false); |
7c673cae FG |
8295 | |
8296 | straydn->pop_projected_linkage(); | |
8297 | if (mdr->is_slave() && !mdr->more()->slave_update_journaled) | |
11fdf7f2 | 8298 | ceph_assert(!straydn->is_projected()); // no other projected |
7c673cae FG |
8299 | |
8300 | // nlink-- targeti | |
11fdf7f2 | 8301 | if (destdn->is_auth()) |
7c673cae | 8302 | oldin->pop_and_dirty_projected_inode(mdr->ls); |
11fdf7f2 TL |
8303 | |
8304 | mdcache->touch_dentry_bottom(straydn); // drop dn as quickly as possible. | |
7c673cae | 8305 | } else if (destdnl->is_remote()) { |
31f18b77 | 8306 | destdn->get_dir()->unlink_inode(destdn, false); |
11fdf7f2 TL |
8307 | if (oldin->is_auth()) { |
8308 | oldin->pop_and_dirty_projected_inode(mdr->ls); | |
8309 | } else if (mdr->slave_request) { | |
8310 | if (mdr->slave_request->desti_snapbl.length() > 0) { | |
8311 | ceph_assert(oldin->snaprealm); | |
8312 | oldin->decode_snap_blob(mdr->slave_request->desti_snapbl); | |
8313 | } | |
8314 | } else if (auto& desti_srnode = mdr->more()->desti_srnode) { | |
8315 | delete desti_srnode; | |
8316 | desti_srnode = NULL; | |
8317 | } | |
7c673cae FG |
8318 | } |
8319 | } | |
8320 | ||
8321 | // unlink src before we relink it at dest | |
8322 | CInode *in = srcdnl->get_inode(); | |
11fdf7f2 | 8323 | ceph_assert(in); |
7c673cae FG |
8324 | |
8325 | bool srcdn_was_remote = srcdnl->is_remote(); | |
11fdf7f2 TL |
8326 | if (!srcdn_was_remote) { |
8327 | // if there is newly created snaprealm, need to split old snaprealm's | |
8328 | // inodes_with_caps. So pop snaprealm before linkage changes. | |
8329 | if (destdn->is_auth()) { | |
8330 | bool hadrealm = (in->snaprealm ? true : false); | |
8331 | in->early_pop_projected_snaprealm(); | |
8332 | new_in_snaprealm = (in->snaprealm && !hadrealm); | |
8333 | } else { | |
8334 | ceph_assert(mdr->slave_request); | |
8335 | if (mdr->slave_request->srci_snapbl.length()) { | |
8336 | new_in_snaprealm = !in->snaprealm; | |
8337 | in->decode_snap_blob(mdr->slave_request->srci_snapbl); | |
8338 | ceph_assert(in->snaprealm); | |
8339 | ceph_assert(in->snaprealm->have_past_parents_open()); | |
8340 | } | |
8341 | } | |
8342 | } | |
8343 | ||
7c673cae FG |
8344 | srcdn->get_dir()->unlink_inode(srcdn); |
8345 | ||
8346 | // dest | |
8347 | if (srcdn_was_remote) { | |
8348 | if (!linkmerge) { | |
8349 | // destdn | |
8350 | destdnl = destdn->pop_projected_linkage(); | |
8351 | if (mdr->is_slave() && !mdr->more()->slave_update_journaled) | |
11fdf7f2 | 8352 | ceph_assert(!destdn->is_projected()); // no other projected |
7c673cae FG |
8353 | |
8354 | destdn->link_remote(destdnl, in); | |
8355 | if (destdn->is_auth()) | |
8356 | destdn->mark_dirty(mdr->more()->pvmap[destdn], mdr->ls); | |
8357 | // in | |
11fdf7f2 | 8358 | if (in->is_auth()) { |
7c673cae | 8359 | in->pop_and_dirty_projected_inode(mdr->ls); |
11fdf7f2 TL |
8360 | } else if (mdr->slave_request) { |
8361 | if (mdr->slave_request->srci_snapbl.length() > 0) { | |
8362 | ceph_assert(in->snaprealm); | |
8363 | in->decode_snap_blob(mdr->slave_request->srci_snapbl); | |
8364 | } | |
8365 | } else if (auto& srci_srnode = mdr->more()->srci_srnode) { | |
8366 | delete srci_srnode; | |
8367 | srci_srnode = NULL; | |
8368 | } | |
7c673cae FG |
8369 | } else { |
8370 | dout(10) << "merging remote onto primary link" << dendl; | |
8371 | oldin->pop_and_dirty_projected_inode(mdr->ls); | |
8372 | } | |
8373 | } else { // primary | |
8374 | if (linkmerge) { | |
8375 | dout(10) << "merging primary onto remote link" << dendl; | |
31f18b77 | 8376 | destdn->get_dir()->unlink_inode(destdn, false); |
7c673cae FG |
8377 | } |
8378 | destdnl = destdn->pop_projected_linkage(); | |
8379 | if (mdr->is_slave() && !mdr->more()->slave_update_journaled) | |
11fdf7f2 | 8380 | ceph_assert(!destdn->is_projected()); // no other projected |
7c673cae FG |
8381 | |
8382 | // srcdn inode import? | |
8383 | if (!srcdn->is_auth() && destdn->is_auth()) { | |
11fdf7f2 | 8384 | ceph_assert(mdr->more()->inode_import.length() > 0); |
7c673cae FG |
8385 | |
8386 | map<client_t,Capability::Import> imported_caps; | |
8387 | ||
8388 | // finish cap imports | |
28e407b8 | 8389 | finish_force_open_sessions(mdr->more()->imported_session_map); |
7c673cae FG |
8390 | if (mdr->more()->cap_imports.count(destdnl->get_inode())) { |
8391 | mdcache->migrator->finish_import_inode_caps(destdnl->get_inode(), | |
28e407b8 AA |
8392 | mdr->more()->srcdn_auth_mds, true, |
8393 | mdr->more()->imported_session_map, | |
8394 | mdr->more()->cap_imports[destdnl->get_inode()], | |
8395 | imported_caps); | |
7c673cae FG |
8396 | } |
8397 | ||
8398 | mdr->more()->inode_import.clear(); | |
11fdf7f2 | 8399 | encode(imported_caps, mdr->more()->inode_import); |
7c673cae FG |
8400 | |
8401 | /* hack: add an auth pin for each xlock we hold. These were | |
8402 | * remote xlocks previously but now they're local and | |
8403 | * we're going to try and unpin when we xlock_finish. */ | |
11fdf7f2 TL |
8404 | |
8405 | for (auto i = mdr->locks.lower_bound(&destdnl->get_inode()->versionlock); | |
8406 | i != mdr->locks.end(); | |
8407 | ++i) { | |
8408 | SimpleLock *lock = i->lock; | |
8409 | if (lock->get_parent() != destdnl->get_inode()) | |
8410 | break; | |
8411 | if (i->is_xlock() && !lock->is_locallock()) | |
8412 | mds->locker->xlock_import(lock); | |
8413 | } | |
7c673cae FG |
8414 | |
8415 | // hack: fix auth bit | |
8416 | in->state_set(CInode::STATE_AUTH); | |
7c673cae FG |
8417 | |
8418 | mdr->clear_ambiguous_auth(); | |
8419 | } | |
8420 | ||
11fdf7f2 | 8421 | if (destdn->is_auth()) |
7c673cae | 8422 | in->pop_and_dirty_projected_inode(mdr->ls); |
7c673cae FG |
8423 | } |
8424 | ||
8425 | // src | |
8426 | if (srcdn->is_auth()) | |
8427 | srcdn->mark_dirty(mdr->more()->pvmap[srcdn], mdr->ls); | |
8428 | srcdn->pop_projected_linkage(); | |
8429 | if (mdr->is_slave() && !mdr->more()->slave_update_journaled) | |
11fdf7f2 | 8430 | ceph_assert(!srcdn->is_projected()); // no other projected |
7c673cae FG |
8431 | |
8432 | // apply remaining projected inodes (nested) | |
8433 | mdr->apply(); | |
8434 | ||
8435 | // update subtree map? | |
11fdf7f2 | 8436 | if (destdnl->is_primary() && in->is_dir()) |
224ce89b | 8437 | mdcache->adjust_subtree_after_rename(in, srcdn->get_dir(), true); |
7c673cae FG |
8438 | |
8439 | if (straydn && oldin->is_dir()) | |
8440 | mdcache->adjust_subtree_after_rename(oldin, destdn->get_dir(), true); | |
8441 | ||
11fdf7f2 TL |
8442 | if (new_oldin_snaprealm) |
8443 | mdcache->do_realm_invalidate_and_update_notify(oldin, CEPH_SNAP_OP_SPLIT, false); | |
8444 | if (new_in_snaprealm) | |
8445 | mdcache->do_realm_invalidate_and_update_notify(in, CEPH_SNAP_OP_SPLIT, true); | |
8446 | ||
7c673cae FG |
8447 | // removing a new dn? |
8448 | if (srcdn->is_auth()) | |
8449 | srcdn->get_dir()->try_remove_unlinked_dn(srcdn); | |
8450 | } | |
8451 | ||
8452 | ||
8453 | ||
8454 | // ------------ | |
8455 | // SLAVE | |
8456 | ||
8457 | class C_MDS_SlaveRenamePrep : public ServerLogContext { | |
8458 | CDentry *srcdn, *destdn, *straydn; | |
8459 | public: | |
8460 | C_MDS_SlaveRenamePrep(Server *s, MDRequestRef& m, CDentry *sr, CDentry *de, CDentry *st) : | |
8461 | ServerLogContext(s, m), srcdn(sr), destdn(de), straydn(st) {} | |
8462 | void finish(int r) override { | |
8463 | server->_logged_slave_rename(mdr, srcdn, destdn, straydn); | |
8464 | } | |
8465 | }; | |
8466 | ||
8467 | class C_MDS_SlaveRenameCommit : public ServerContext { | |
8468 | MDRequestRef mdr; | |
8469 | CDentry *srcdn, *destdn, *straydn; | |
8470 | public: | |
8471 | C_MDS_SlaveRenameCommit(Server *s, MDRequestRef& m, CDentry *sr, CDentry *de, CDentry *st) : | |
8472 | ServerContext(s), mdr(m), srcdn(sr), destdn(de), straydn(st) {} | |
8473 | void finish(int r) override { | |
8474 | server->_commit_slave_rename(mdr, r, srcdn, destdn, straydn); | |
8475 | } | |
8476 | }; | |
8477 | ||
8478 | class C_MDS_SlaveRenameSessionsFlushed : public ServerContext { | |
8479 | MDRequestRef mdr; | |
8480 | public: | |
8481 | C_MDS_SlaveRenameSessionsFlushed(Server *s, MDRequestRef& r) : | |
8482 | ServerContext(s), mdr(r) {} | |
8483 | void finish(int r) override { | |
8484 | server->_slave_rename_sessions_flushed(mdr); | |
8485 | } | |
8486 | }; | |
8487 | ||
7c673cae FG |
8488 | void Server::handle_slave_rename_prep(MDRequestRef& mdr) |
8489 | { | |
8490 | dout(10) << "handle_slave_rename_prep " << *mdr | |
8491 | << " " << mdr->slave_request->srcdnpath | |
8492 | << " to " << mdr->slave_request->destdnpath | |
8493 | << dendl; | |
31f18b77 FG |
8494 | |
8495 | if (mdr->slave_request->is_interrupted()) { | |
8496 | dout(10) << " slave request interrupted, sending noop reply" << dendl; | |
11fdf7f2 | 8497 | auto reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RENAMEPREPACK); |
31f18b77 FG |
8498 | reply->mark_interrupted(); |
8499 | mds->send_message_mds(reply, mdr->slave_to_mds); | |
91327a77 | 8500 | mdr->reset_slave_request(); |
31f18b77 FG |
8501 | return; |
8502 | } | |
8503 | ||
7c673cae FG |
8504 | // discover destdn |
8505 | filepath destpath(mdr->slave_request->destdnpath); | |
8506 | dout(10) << " dest " << destpath << dendl; | |
8507 | vector<CDentry*> trace; | |
11fdf7f2 TL |
8508 | CF_MDS_MDRContextFactory cf(mdcache, mdr); |
8509 | int r = mdcache->path_traverse(mdr, cf, destpath, &trace, NULL, MDS_TRAVERSE_DISCOVERXLOCK); | |
7c673cae FG |
8510 | if (r > 0) return; |
8511 | if (r == -ESTALE) { | |
8512 | mdcache->find_ino_peers(destpath.get_ino(), new C_MDS_RetryRequest(mdcache, mdr), | |
8513 | mdr->slave_to_mds); | |
8514 | return; | |
8515 | } | |
11fdf7f2 | 8516 | ceph_assert(r == 0); // we shouldn't get an error here! |
7c673cae | 8517 | |
91327a77 | 8518 | CDentry *destdn = trace.back(); |
7c673cae FG |
8519 | CDentry::linkage_t *destdnl = destdn->get_projected_linkage(); |
8520 | dout(10) << " destdn " << *destdn << dendl; | |
8521 | mdr->pin(destdn); | |
8522 | ||
8523 | // discover srcdn | |
8524 | filepath srcpath(mdr->slave_request->srcdnpath); | |
8525 | dout(10) << " src " << srcpath << dendl; | |
8526 | CInode *srci = nullptr; | |
11fdf7f2 | 8527 | r = mdcache->path_traverse(mdr, cf, srcpath, &trace, &srci, MDS_TRAVERSE_DISCOVERXLOCK); |
7c673cae | 8528 | if (r > 0) return; |
11fdf7f2 | 8529 | ceph_assert(r == 0); |
7c673cae FG |
8530 | |
8531 | // srcpath must not point to a null dentry | |
11fdf7f2 | 8532 | ceph_assert(srci != nullptr); |
7c673cae | 8533 | |
91327a77 | 8534 | CDentry *srcdn = trace.back(); |
7c673cae FG |
8535 | CDentry::linkage_t *srcdnl = srcdn->get_projected_linkage(); |
8536 | dout(10) << " srcdn " << *srcdn << dendl; | |
8537 | mdr->pin(srcdn); | |
8538 | mdr->pin(srci); | |
8539 | ||
8540 | // stray? | |
11fdf7f2 TL |
8541 | bool linkmerge = srcdnl->get_inode() == destdnl->get_inode(); |
8542 | if (linkmerge) | |
8543 | ceph_assert(srcdnl->is_primary() && destdnl->is_remote()); | |
7c673cae FG |
8544 | CDentry *straydn = mdr->straydn; |
8545 | if (destdnl->is_primary() && !linkmerge) | |
11fdf7f2 | 8546 | ceph_assert(straydn); |
7c673cae FG |
8547 | |
8548 | mdr->set_op_stamp(mdr->slave_request->op_stamp); | |
8549 | mdr->more()->srcdn_auth_mds = srcdn->authority().first; | |
8550 | ||
8551 | // set up commit waiter (early, to clean up any freezing etc we do) | |
8552 | if (!mdr->more()->slave_commit) | |
8553 | mdr->more()->slave_commit = new C_MDS_SlaveRenameCommit(this, mdr, srcdn, destdn, straydn); | |
8554 | ||
8555 | // am i srcdn auth? | |
8556 | if (srcdn->is_auth()) { | |
8557 | set<mds_rank_t> srcdnrep; | |
8558 | srcdn->list_replicas(srcdnrep); | |
8559 | ||
8560 | bool reply_witness = false; | |
8561 | if (srcdnl->is_primary() && !srcdnl->get_inode()->state_test(CInode::STATE_AMBIGUOUSAUTH)) { | |
8562 | // freeze? | |
8563 | // we need this to | |
8564 | // - avoid conflicting lock state changes | |
8565 | // - avoid concurrent updates to the inode | |
8566 | // (this could also be accomplished with the versionlock) | |
11fdf7f2 | 8567 | int allowance = 3; // 1 for the mdr auth_pin, 1 for the link lock, 1 for the snap lock |
7c673cae FG |
8568 | dout(10) << " freezing srci " << *srcdnl->get_inode() << " with allowance " << allowance << dendl; |
8569 | bool frozen_inode = srcdnl->get_inode()->freeze_inode(allowance); | |
8570 | ||
8571 | // unfreeze auth pin after freezing the inode to avoid queueing waiters | |
8572 | if (srcdnl->get_inode()->is_frozen_auth_pin()) | |
8573 | mdr->unfreeze_auth_pin(); | |
8574 | ||
8575 | if (!frozen_inode) { | |
8576 | srcdnl->get_inode()->add_waiter(CInode::WAIT_FROZEN, new C_MDS_RetryRequest(mdcache, mdr)); | |
8577 | return; | |
8578 | } | |
8579 | ||
8580 | /* | |
8581 | * set ambiguous auth for srci | |
8582 | * NOTE: we don't worry about ambiguous cache expire as we do | |
8583 | * with subtree migrations because all slaves will pin | |
8584 | * srcdn->get_inode() for duration of this rename. | |
8585 | */ | |
8586 | mdr->set_ambiguous_auth(srcdnl->get_inode()); | |
8587 | ||
8588 | // just mark the source inode as ambiguous auth if more than two MDS are involved. | |
8589 | // the master will send another OP_RENAMEPREP slave request later. | |
8590 | if (mdr->slave_request->witnesses.size() > 1) { | |
8591 | dout(10) << " set srci ambiguous auth; providing srcdn replica list" << dendl; | |
8592 | reply_witness = true; | |
8593 | } | |
8594 | ||
8595 | // make sure bystanders have received all lock related messages | |
8596 | for (set<mds_rank_t>::iterator p = srcdnrep.begin(); p != srcdnrep.end(); ++p) { | |
8597 | if (*p == mdr->slave_to_mds || | |
8598 | (mds->is_cluster_degraded() && | |
8599 | !mds->mdsmap->is_clientreplay_or_active_or_stopping(*p))) | |
8600 | continue; | |
11fdf7f2 | 8601 | auto notify = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RENAMENOTIFY); |
7c673cae FG |
8602 | mds->send_message_mds(notify, *p); |
8603 | mdr->more()->waiting_on_slave.insert(*p); | |
8604 | } | |
8605 | ||
8606 | // make sure clients have received all cap related messages | |
8607 | set<client_t> export_client_set; | |
8608 | mdcache->migrator->get_export_client_set(srcdnl->get_inode(), export_client_set); | |
8609 | ||
8610 | MDSGatherBuilder gather(g_ceph_context); | |
8611 | flush_client_sessions(export_client_set, gather); | |
8612 | if (gather.has_subs()) { | |
8613 | mdr->more()->waiting_on_slave.insert(MDS_RANK_NONE); | |
8614 | gather.set_finisher(new C_MDS_SlaveRenameSessionsFlushed(this, mdr)); | |
8615 | gather.activate(); | |
8616 | } | |
8617 | } | |
8618 | ||
8619 | // is witness list sufficient? | |
8620 | for (set<mds_rank_t>::iterator p = srcdnrep.begin(); p != srcdnrep.end(); ++p) { | |
8621 | if (*p == mdr->slave_to_mds || | |
8622 | mdr->slave_request->witnesses.count(*p)) continue; | |
8623 | dout(10) << " witness list insufficient; providing srcdn replica list" << dendl; | |
8624 | reply_witness = true; | |
8625 | break; | |
8626 | } | |
8627 | ||
8628 | if (reply_witness) { | |
11fdf7f2 TL |
8629 | ceph_assert(!srcdnrep.empty()); |
8630 | auto reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RENAMEPREPACK); | |
7c673cae FG |
8631 | reply->witnesses.swap(srcdnrep); |
8632 | mds->send_message_mds(reply, mdr->slave_to_mds); | |
91327a77 | 8633 | mdr->reset_slave_request(); |
7c673cae FG |
8634 | return; |
8635 | } | |
8636 | dout(10) << " witness list sufficient: includes all srcdn replicas" << dendl; | |
8637 | if (!mdr->more()->waiting_on_slave.empty()) { | |
8638 | dout(10) << " still waiting for rename notify acks from " | |
8639 | << mdr->more()->waiting_on_slave << dendl; | |
8640 | return; | |
8641 | } | |
8642 | } else if (srcdnl->is_primary() && srcdn->authority() != destdn->authority()) { | |
8643 | // set ambiguous auth for srci on witnesses | |
8644 | mdr->set_ambiguous_auth(srcdnl->get_inode()); | |
8645 | } | |
8646 | ||
8647 | // encode everything we'd need to roll this back... basically, just the original state. | |
8648 | rename_rollback rollback; | |
8649 | ||
8650 | rollback.reqid = mdr->reqid; | |
8651 | ||
8652 | rollback.orig_src.dirfrag = srcdn->get_dir()->dirfrag(); | |
8653 | rollback.orig_src.dirfrag_old_mtime = srcdn->get_dir()->get_projected_fnode()->fragstat.mtime; | |
8654 | rollback.orig_src.dirfrag_old_rctime = srcdn->get_dir()->get_projected_fnode()->rstat.rctime; | |
11fdf7f2 | 8655 | rollback.orig_src.dname = srcdn->get_name(); |
7c673cae FG |
8656 | if (srcdnl->is_primary()) |
8657 | rollback.orig_src.ino = srcdnl->get_inode()->ino(); | |
8658 | else { | |
11fdf7f2 | 8659 | ceph_assert(srcdnl->is_remote()); |
7c673cae FG |
8660 | rollback.orig_src.remote_ino = srcdnl->get_remote_ino(); |
8661 | rollback.orig_src.remote_d_type = srcdnl->get_remote_d_type(); | |
8662 | } | |
8663 | ||
8664 | rollback.orig_dest.dirfrag = destdn->get_dir()->dirfrag(); | |
8665 | rollback.orig_dest.dirfrag_old_mtime = destdn->get_dir()->get_projected_fnode()->fragstat.mtime; | |
8666 | rollback.orig_dest.dirfrag_old_rctime = destdn->get_dir()->get_projected_fnode()->rstat.rctime; | |
11fdf7f2 | 8667 | rollback.orig_dest.dname = destdn->get_name(); |
7c673cae FG |
8668 | if (destdnl->is_primary()) |
8669 | rollback.orig_dest.ino = destdnl->get_inode()->ino(); | |
8670 | else if (destdnl->is_remote()) { | |
8671 | rollback.orig_dest.remote_ino = destdnl->get_remote_ino(); | |
8672 | rollback.orig_dest.remote_d_type = destdnl->get_remote_d_type(); | |
8673 | } | |
8674 | ||
8675 | if (straydn) { | |
8676 | rollback.stray.dirfrag = straydn->get_dir()->dirfrag(); | |
8677 | rollback.stray.dirfrag_old_mtime = straydn->get_dir()->get_projected_fnode()->fragstat.mtime; | |
8678 | rollback.stray.dirfrag_old_rctime = straydn->get_dir()->get_projected_fnode()->rstat.rctime; | |
11fdf7f2 TL |
8679 | rollback.stray.dname = straydn->get_name(); |
8680 | } | |
8681 | if (mdr->slave_request->desti_snapbl.length()) { | |
8682 | CInode *oldin = destdnl->get_inode(); | |
8683 | if (oldin->snaprealm) { | |
8684 | encode(true, rollback.desti_snapbl); | |
8685 | oldin->encode_snap_blob(rollback.desti_snapbl); | |
8686 | } else { | |
8687 | encode(false, rollback.desti_snapbl); | |
8688 | } | |
8689 | } | |
8690 | if (mdr->slave_request->srci_snapbl.length()) { | |
8691 | if (srci->snaprealm) { | |
8692 | encode(true, rollback.srci_snapbl); | |
8693 | srci->encode_snap_blob(rollback.srci_snapbl); | |
8694 | } else { | |
8695 | encode(false, rollback.srci_snapbl); | |
8696 | } | |
7c673cae | 8697 | } |
11fdf7f2 TL |
8698 | encode(rollback, mdr->more()->rollback_bl); |
8699 | // FIXME: rollback snaprealm | |
7c673cae FG |
8700 | dout(20) << " rollback is " << mdr->more()->rollback_bl.length() << " bytes" << dendl; |
8701 | ||
8702 | // journal. | |
8703 | mdr->ls = mdlog->get_current_segment(); | |
8704 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rename_prep", mdr->reqid, mdr->slave_to_mds, | |
8705 | ESlaveUpdate::OP_PREPARE, ESlaveUpdate::RENAME); | |
8706 | mdlog->start_entry(le); | |
8707 | le->rollback = mdr->more()->rollback_bl; | |
8708 | ||
8709 | bufferlist blah; // inode import data... obviously not used if we're the slave | |
8710 | _rename_prepare(mdr, &le->commit, &blah, srcdn, destdn, straydn); | |
8711 | ||
8712 | if (le->commit.empty()) { | |
8713 | dout(10) << " empty metablob, skipping journal" << dendl; | |
8714 | mdlog->cancel_entry(le); | |
8715 | mdr->ls = NULL; | |
8716 | _logged_slave_rename(mdr, srcdn, destdn, straydn); | |
8717 | } else { | |
8718 | mdr->more()->slave_update_journaled = true; | |
8719 | submit_mdlog_entry(le, new C_MDS_SlaveRenamePrep(this, mdr, srcdn, destdn, straydn), | |
8720 | mdr, __func__); | |
8721 | mdlog->flush(); | |
8722 | } | |
8723 | } | |
8724 | ||
8725 | void Server::_logged_slave_rename(MDRequestRef& mdr, | |
8726 | CDentry *srcdn, CDentry *destdn, CDentry *straydn) | |
8727 | { | |
8728 | dout(10) << "_logged_slave_rename " << *mdr << dendl; | |
8729 | ||
8730 | // prepare ack | |
11fdf7f2 | 8731 | MMDSSlaveRequest::ref reply; |
7c673cae | 8732 | if (!mdr->aborted) { |
11fdf7f2 | 8733 | reply = MMDSSlaveRequest::create(mdr->reqid, mdr->attempt, MMDSSlaveRequest::OP_RENAMEPREPACK); |
7c673cae FG |
8734 | if (!mdr->more()->slave_update_journaled) |
8735 | reply->mark_not_journaled(); | |
8736 | } | |
8737 | ||
8738 | CDentry::linkage_t *srcdnl = srcdn->get_linkage(); | |
7c673cae FG |
8739 | //CDentry::linkage_t *straydnl = straydn ? straydn->get_linkage() : 0; |
8740 | ||
8741 | // export srci? | |
8742 | if (srcdn->is_auth() && srcdnl->is_primary()) { | |
8743 | // set export bounds for CInode::encode_export() | |
11fdf7f2 TL |
8744 | if (reply) { |
8745 | list<CDir*> bounds; | |
8746 | if (srcdnl->get_inode()->is_dir()) { | |
8747 | srcdnl->get_inode()->get_dirfrags(bounds); | |
8748 | for (list<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) | |
8749 | (*p)->state_set(CDir::STATE_EXPORTBOUND); | |
8750 | } | |
7c673cae | 8751 | |
11fdf7f2 TL |
8752 | map<client_t,entity_inst_t> exported_client_map; |
8753 | map<client_t, client_metadata_t> exported_client_metadata_map; | |
8754 | bufferlist inodebl; | |
8755 | mdcache->migrator->encode_export_inode(srcdnl->get_inode(), inodebl, | |
8756 | exported_client_map, | |
8757 | exported_client_metadata_map); | |
7c673cae | 8758 | |
11fdf7f2 TL |
8759 | for (list<CDir*>::iterator p = bounds.begin(); p != bounds.end(); ++p) |
8760 | (*p)->state_clear(CDir::STATE_EXPORTBOUND); | |
7c673cae | 8761 | |
11fdf7f2 TL |
8762 | encode(exported_client_map, reply->inode_export, mds->mdsmap->get_up_features()); |
8763 | encode(exported_client_metadata_map, reply->inode_export); | |
7c673cae FG |
8764 | reply->inode_export.claim_append(inodebl); |
8765 | reply->inode_export_v = srcdnl->get_inode()->inode.version; | |
8766 | } | |
8767 | ||
8768 | // remove mdr auth pin | |
8769 | mdr->auth_unpin(srcdnl->get_inode()); | |
8770 | mdr->more()->is_inode_exporter = true; | |
8771 | ||
8772 | if (srcdnl->get_inode()->is_dirty()) | |
8773 | srcdnl->get_inode()->mark_clean(); | |
8774 | ||
8775 | dout(10) << " exported srci " << *srcdnl->get_inode() << dendl; | |
8776 | } | |
8777 | ||
8778 | // apply | |
8779 | _rename_apply(mdr, srcdn, destdn, straydn); | |
11fdf7f2 TL |
8780 | |
8781 | CDentry::linkage_t *destdnl = destdn->get_linkage(); | |
7c673cae FG |
8782 | |
8783 | // bump popularity | |
11fdf7f2 | 8784 | mds->balancer->hit_dir(srcdn->get_dir(), META_POP_IWR); |
7c673cae | 8785 | if (destdnl->get_inode() && destdnl->get_inode()->is_auth()) |
11fdf7f2 | 8786 | mds->balancer->hit_inode(destdnl->get_inode(), META_POP_IWR); |
7c673cae FG |
8787 | |
8788 | // done. | |
91327a77 | 8789 | mdr->reset_slave_request(); |
7c673cae FG |
8790 | mdr->straydn = 0; |
8791 | ||
8792 | if (reply) { | |
8793 | mds->send_message_mds(reply, mdr->slave_to_mds); | |
8794 | } else { | |
11fdf7f2 | 8795 | ceph_assert(mdr->aborted); |
7c673cae FG |
8796 | dout(10) << " abort flag set, finishing" << dendl; |
8797 | mdcache->request_finish(mdr); | |
8798 | } | |
8799 | } | |
8800 | ||
8801 | void Server::_commit_slave_rename(MDRequestRef& mdr, int r, | |
8802 | CDentry *srcdn, CDentry *destdn, CDentry *straydn) | |
8803 | { | |
8804 | dout(10) << "_commit_slave_rename " << *mdr << " r=" << r << dendl; | |
8805 | ||
f64942e4 AA |
8806 | CInode *in = destdn->get_linkage()->get_inode(); |
8807 | ||
8808 | inodeno_t migrated_stray; | |
8809 | if (srcdn->is_auth() && srcdn->get_dir()->inode->is_stray()) | |
8810 | migrated_stray = in->ino(); | |
7c673cae | 8811 | |
11fdf7f2 | 8812 | MDSContext::vec finished; |
7c673cae FG |
8813 | if (r == 0) { |
8814 | // unfreeze+singleauth inode | |
8815 | // hmm, do i really need to delay this? | |
8816 | if (mdr->more()->is_inode_exporter) { | |
7c673cae FG |
8817 | // drop our pins |
8818 | // we exported, clear out any xlocks that we moved to another MDS | |
7c673cae | 8819 | |
11fdf7f2 TL |
8820 | for (auto i = mdr->locks.lower_bound(&in->versionlock); |
8821 | i != mdr->locks.end(); ) { | |
8822 | SimpleLock *lock = i->lock; | |
8823 | if (lock->get_parent() != in) | |
8824 | break; | |
7c673cae | 8825 | // we only care about xlocks on the exported inode |
11fdf7f2 TL |
8826 | if (i->is_xlock() && !lock->is_locallock()) |
8827 | mds->locker->xlock_export(i++, mdr.get()); | |
8828 | else | |
8829 | ++i; | |
7c673cae FG |
8830 | } |
8831 | ||
8832 | map<client_t,Capability::Import> peer_imported; | |
11fdf7f2 TL |
8833 | auto bp = mdr->more()->inode_import.cbegin(); |
8834 | decode(peer_imported, bp); | |
7c673cae | 8835 | |
f64942e4 | 8836 | dout(10) << " finishing inode export on " << *in << dendl; |
11fdf7f2 | 8837 | mdcache->migrator->finish_export_inode(in, mdr->slave_to_mds, peer_imported, finished); |
7c673cae FG |
8838 | mds->queue_waiters(finished); // this includes SINGLEAUTH waiters. |
8839 | ||
8840 | // unfreeze | |
11fdf7f2 | 8841 | ceph_assert(in->is_frozen_inode()); |
f64942e4 | 8842 | in->unfreeze_inode(finished); |
7c673cae FG |
8843 | } |
8844 | ||
8845 | // singleauth | |
8846 | if (mdr->more()->is_ambiguous_auth) { | |
8847 | mdr->more()->rename_inode->clear_ambiguous_auth(finished); | |
8848 | mdr->more()->is_ambiguous_auth = false; | |
8849 | } | |
8850 | ||
31f18b77 FG |
8851 | if (straydn && mdr->more()->slave_update_journaled) { |
8852 | CInode *strayin = straydn->get_projected_linkage()->get_inode(); | |
8853 | if (strayin && !strayin->snaprealm) | |
8854 | mdcache->clear_dirty_bits_for_stray(strayin); | |
8855 | } | |
7c673cae FG |
8856 | |
8857 | mds->queue_waiters(finished); | |
8858 | mdr->cleanup(); | |
8859 | ||
8860 | if (mdr->more()->slave_update_journaled) { | |
8861 | // write a commit to the journal | |
8862 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rename_commit", mdr->reqid, | |
8863 | mdr->slave_to_mds, ESlaveUpdate::OP_COMMIT, | |
8864 | ESlaveUpdate::RENAME); | |
8865 | mdlog->start_entry(le); | |
8866 | submit_mdlog_entry(le, new C_MDS_CommittedSlave(this, mdr), mdr, __func__); | |
8867 | mdlog->flush(); | |
8868 | } else { | |
8869 | _committed_slave(mdr); | |
8870 | } | |
8871 | } else { | |
8872 | ||
8873 | // abort | |
8874 | // rollback_bl may be empty if we froze the inode but had to provide an expanded | |
8875 | // witness list from the master, and they failed before we tried prep again. | |
8876 | if (mdr->more()->rollback_bl.length()) { | |
8877 | if (mdr->more()->is_inode_exporter) { | |
f64942e4 AA |
8878 | dout(10) << " reversing inode export of " << *in << dendl; |
8879 | in->abort_export(); | |
7c673cae FG |
8880 | } |
8881 | if (mdcache->is_ambiguous_slave_update(mdr->reqid, mdr->slave_to_mds)) { | |
8882 | mdcache->remove_ambiguous_slave_update(mdr->reqid, mdr->slave_to_mds); | |
8883 | // rollback but preserve the slave request | |
8884 | do_rename_rollback(mdr->more()->rollback_bl, mdr->slave_to_mds, mdr, false); | |
8885 | mdr->more()->rollback_bl.clear(); | |
8886 | } else | |
8887 | do_rename_rollback(mdr->more()->rollback_bl, mdr->slave_to_mds, mdr, true); | |
8888 | } else { | |
8889 | dout(10) << " rollback_bl empty, not rollback back rename (master failed after getting extra witnesses?)" << dendl; | |
8890 | // singleauth | |
8891 | if (mdr->more()->is_ambiguous_auth) { | |
8892 | if (srcdn->is_auth()) | |
8893 | mdr->more()->rename_inode->unfreeze_inode(finished); | |
8894 | ||
8895 | mdr->more()->rename_inode->clear_ambiguous_auth(finished); | |
8896 | mdr->more()->is_ambiguous_auth = false; | |
8897 | } | |
8898 | mds->queue_waiters(finished); | |
8899 | mdcache->request_finish(mdr); | |
8900 | } | |
8901 | } | |
f64942e4 AA |
8902 | |
8903 | if (migrated_stray && mds->is_stopping()) | |
8904 | mdcache->shutdown_export_stray_finish(migrated_stray); | |
7c673cae FG |
8905 | } |
8906 | ||
8907 | void _rollback_repair_dir(MutationRef& mut, CDir *dir, rename_rollback::drec &r, utime_t ctime, | |
8908 | bool isdir, int linkunlink, nest_info_t &rstat) | |
8909 | { | |
8910 | fnode_t *pf; | |
8911 | pf = dir->project_fnode(); | |
8912 | mut->add_projected_fnode(dir); | |
8913 | pf->version = dir->pre_dirty(); | |
8914 | ||
8915 | if (isdir) { | |
8916 | pf->fragstat.nsubdirs += linkunlink; | |
8917 | } else { | |
8918 | pf->fragstat.nfiles += linkunlink; | |
8919 | } | |
8920 | if (r.ino) { | |
8921 | pf->rstat.rbytes += linkunlink * rstat.rbytes; | |
8922 | pf->rstat.rfiles += linkunlink * rstat.rfiles; | |
8923 | pf->rstat.rsubdirs += linkunlink * rstat.rsubdirs; | |
11fdf7f2 | 8924 | pf->rstat.rsnaps += linkunlink * rstat.rsnaps; |
7c673cae FG |
8925 | } |
8926 | if (pf->fragstat.mtime == ctime) { | |
8927 | pf->fragstat.mtime = r.dirfrag_old_mtime; | |
8928 | if (pf->rstat.rctime == ctime) | |
8929 | pf->rstat.rctime = r.dirfrag_old_rctime; | |
8930 | } | |
8931 | mut->add_updated_lock(&dir->get_inode()->filelock); | |
8932 | mut->add_updated_lock(&dir->get_inode()->nestlock); | |
8933 | } | |
8934 | ||
8935 | struct C_MDS_LoggedRenameRollback : public ServerLogContext { | |
8936 | MutationRef mut; | |
8937 | CDentry *srcdn; | |
8938 | version_t srcdnpv; | |
8939 | CDentry *destdn; | |
8940 | CDentry *straydn; | |
11fdf7f2 | 8941 | map<client_t,MClientSnap::ref> splits[2]; |
7c673cae FG |
8942 | bool finish_mdr; |
8943 | C_MDS_LoggedRenameRollback(Server *s, MutationRef& m, MDRequestRef& r, | |
11fdf7f2 TL |
8944 | CDentry *sd, version_t pv, CDentry *dd, CDentry *st, |
8945 | map<client_t,MClientSnap::ref> _splits[2], bool f) : | |
7c673cae | 8946 | ServerLogContext(s, r), mut(m), srcdn(sd), srcdnpv(pv), destdn(dd), |
11fdf7f2 TL |
8947 | straydn(st), finish_mdr(f) { |
8948 | splits[0].swap(_splits[0]); | |
8949 | splits[1].swap(_splits[1]); | |
8950 | } | |
7c673cae FG |
8951 | void finish(int r) override { |
8952 | server->_rename_rollback_finish(mut, mdr, srcdn, srcdnpv, | |
11fdf7f2 | 8953 | destdn, straydn, splits, finish_mdr); |
7c673cae FG |
8954 | } |
8955 | }; | |
8956 | ||
8957 | void Server::do_rename_rollback(bufferlist &rbl, mds_rank_t master, MDRequestRef& mdr, | |
8958 | bool finish_mdr) | |
8959 | { | |
8960 | rename_rollback rollback; | |
11fdf7f2 TL |
8961 | auto p = rbl.cbegin(); |
8962 | decode(rollback, p); | |
7c673cae FG |
8963 | |
8964 | dout(10) << "do_rename_rollback on " << rollback.reqid << dendl; | |
8965 | // need to finish this update before sending resolve to claim the subtree | |
8966 | mdcache->add_rollback(rollback.reqid, master); | |
8967 | ||
8968 | MutationRef mut(new MutationImpl(nullptr, utime_t(), rollback.reqid)); | |
8969 | mut->ls = mds->mdlog->get_current_segment(); | |
8970 | ||
8971 | CDentry *srcdn = NULL; | |
8972 | CDir *srcdir = mdcache->get_dirfrag(rollback.orig_src.dirfrag); | |
8973 | if (!srcdir) | |
8974 | srcdir = mdcache->get_dirfrag(rollback.orig_src.dirfrag.ino, rollback.orig_src.dname); | |
8975 | if (srcdir) { | |
8976 | dout(10) << " srcdir " << *srcdir << dendl; | |
8977 | srcdn = srcdir->lookup(rollback.orig_src.dname); | |
8978 | if (srcdn) { | |
8979 | dout(10) << " srcdn " << *srcdn << dendl; | |
11fdf7f2 | 8980 | ceph_assert(srcdn->get_linkage()->is_null()); |
7c673cae FG |
8981 | } else |
8982 | dout(10) << " srcdn not found" << dendl; | |
8983 | } else | |
8984 | dout(10) << " srcdir not found" << dendl; | |
8985 | ||
8986 | CDentry *destdn = NULL; | |
8987 | CDir *destdir = mdcache->get_dirfrag(rollback.orig_dest.dirfrag); | |
8988 | if (!destdir) | |
8989 | destdir = mdcache->get_dirfrag(rollback.orig_dest.dirfrag.ino, rollback.orig_dest.dname); | |
8990 | if (destdir) { | |
8991 | dout(10) << " destdir " << *destdir << dendl; | |
8992 | destdn = destdir->lookup(rollback.orig_dest.dname); | |
8993 | if (destdn) | |
8994 | dout(10) << " destdn " << *destdn << dendl; | |
8995 | else | |
8996 | dout(10) << " destdn not found" << dendl; | |
8997 | } else | |
8998 | dout(10) << " destdir not found" << dendl; | |
8999 | ||
9000 | CInode *in = NULL; | |
9001 | if (rollback.orig_src.ino) { | |
9002 | in = mdcache->get_inode(rollback.orig_src.ino); | |
9003 | if (in && in->is_dir()) | |
11fdf7f2 | 9004 | ceph_assert(srcdn && destdn); |
7c673cae FG |
9005 | } else |
9006 | in = mdcache->get_inode(rollback.orig_src.remote_ino); | |
9007 | ||
9008 | CDir *straydir = NULL; | |
9009 | CDentry *straydn = NULL; | |
9010 | if (rollback.stray.dirfrag.ino) { | |
9011 | straydir = mdcache->get_dirfrag(rollback.stray.dirfrag); | |
9012 | if (straydir) { | |
9013 | dout(10) << "straydir " << *straydir << dendl; | |
9014 | straydn = straydir->lookup(rollback.stray.dname); | |
9015 | if (straydn) { | |
9016 | dout(10) << " straydn " << *straydn << dendl; | |
11fdf7f2 | 9017 | ceph_assert(straydn->get_linkage()->is_primary()); |
7c673cae FG |
9018 | } else |
9019 | dout(10) << " straydn not found" << dendl; | |
9020 | } else | |
9021 | dout(10) << "straydir not found" << dendl; | |
9022 | } | |
9023 | ||
9024 | CInode *target = NULL; | |
9025 | if (rollback.orig_dest.ino) { | |
9026 | target = mdcache->get_inode(rollback.orig_dest.ino); | |
9027 | if (target) | |
11fdf7f2 | 9028 | ceph_assert(destdn && straydn); |
7c673cae FG |
9029 | } else if (rollback.orig_dest.remote_ino) |
9030 | target = mdcache->get_inode(rollback.orig_dest.remote_ino); | |
9031 | ||
9032 | // can't use is_auth() in the resolve stage | |
9033 | mds_rank_t whoami = mds->get_nodeid(); | |
9034 | // slave | |
11fdf7f2 TL |
9035 | ceph_assert(!destdn || destdn->authority().first != whoami); |
9036 | ceph_assert(!straydn || straydn->authority().first != whoami); | |
7c673cae FG |
9037 | |
9038 | bool force_journal_src = false; | |
9039 | bool force_journal_dest = false; | |
9040 | if (in && in->is_dir() && srcdn->authority().first != whoami) | |
9041 | force_journal_src = _need_force_journal(in, false); | |
9042 | if (in && target && target->is_dir()) | |
9043 | force_journal_dest = _need_force_journal(in, true); | |
9044 | ||
9045 | version_t srcdnpv = 0; | |
9046 | // repair src | |
9047 | if (srcdn) { | |
9048 | if (srcdn->authority().first == whoami) | |
9049 | srcdnpv = srcdn->pre_dirty(); | |
9050 | if (rollback.orig_src.ino) { | |
11fdf7f2 | 9051 | ceph_assert(in); |
7c673cae FG |
9052 | srcdn->push_projected_linkage(in); |
9053 | } else | |
9054 | srcdn->push_projected_linkage(rollback.orig_src.remote_ino, | |
9055 | rollback.orig_src.remote_d_type); | |
9056 | } | |
9057 | ||
11fdf7f2 TL |
9058 | map<client_t,MClientSnap::ref> splits[2]; |
9059 | ||
9060 | CInode::mempool_inode *pip = nullptr; | |
7c673cae | 9061 | if (in) { |
11fdf7f2 TL |
9062 | bool projected; |
9063 | if (in->get_projected_parent_dn()->authority().first == whoami) { | |
94b18763 | 9064 | auto &pi = in->project_inode(); |
94b18763 | 9065 | pip = &pi.inode; |
11fdf7f2 TL |
9066 | mut->add_projected_inode(in); |
9067 | pip->version = in->pre_dirty(); | |
9068 | projected = true; | |
9069 | } else { | |
94b18763 | 9070 | pip = in->get_projected_inode(); |
11fdf7f2 TL |
9071 | projected = false; |
9072 | } | |
94b18763 | 9073 | if (pip->ctime == rollback.ctime) |
91327a77 | 9074 | pip->ctime = rollback.orig_src.old_ctime; |
11fdf7f2 TL |
9075 | |
9076 | if (rollback.srci_snapbl.length() && in->snaprealm) { | |
9077 | bool hadrealm; | |
9078 | auto p = rollback.srci_snapbl.cbegin(); | |
9079 | decode(hadrealm, p); | |
9080 | if (hadrealm) { | |
9081 | if (projected && !mds->is_resolve()) { | |
9082 | sr_t *new_srnode = new sr_t(); | |
9083 | decode(*new_srnode, p); | |
9084 | in->project_snaprealm(new_srnode); | |
9085 | } else | |
9086 | decode(in->snaprealm->srnode, p); | |
9087 | } else { | |
9088 | SnapRealm *realm; | |
9089 | if (rollback.orig_src.ino) { | |
9090 | ceph_assert(srcdir); | |
9091 | realm = srcdir->get_inode()->find_snaprealm(); | |
9092 | } else { | |
9093 | realm = in->snaprealm->parent; | |
9094 | } | |
9095 | if (!mds->is_resolve()) | |
9096 | mdcache->prepare_realm_merge(in->snaprealm, realm, splits[0]); | |
9097 | if (projected) | |
9098 | in->project_snaprealm(NULL); | |
9099 | else | |
9100 | in->snaprealm->merge_to(realm); | |
9101 | } | |
9102 | } | |
7c673cae FG |
9103 | } |
9104 | ||
9105 | if (srcdn && srcdn->authority().first == whoami) { | |
9106 | nest_info_t blah; | |
9107 | _rollback_repair_dir(mut, srcdir, rollback.orig_src, rollback.ctime, | |
94b18763 | 9108 | in ? in->is_dir() : false, 1, pip ? pip->accounted_rstat : blah); |
7c673cae FG |
9109 | } |
9110 | ||
9111 | // repair dest | |
9112 | if (destdn) { | |
9113 | if (rollback.orig_dest.ino && target) { | |
9114 | destdn->push_projected_linkage(target); | |
9115 | } else if (rollback.orig_dest.remote_ino) { | |
9116 | destdn->push_projected_linkage(rollback.orig_dest.remote_ino, | |
9117 | rollback.orig_dest.remote_d_type); | |
9118 | } else { | |
9119 | // the dentry will be trimmed soon, it's ok to have wrong linkage | |
9120 | if (rollback.orig_dest.ino) | |
11fdf7f2 | 9121 | ceph_assert(mds->is_resolve()); |
7c673cae FG |
9122 | destdn->push_projected_linkage(); |
9123 | } | |
9124 | } | |
9125 | ||
9126 | if (straydn) | |
9127 | straydn->push_projected_linkage(); | |
9128 | ||
9129 | if (target) { | |
11fdf7f2 TL |
9130 | bool projected; |
9131 | CInode::mempool_inode *ti = nullptr; | |
9132 | if (target->get_projected_parent_dn()->authority().first == whoami) { | |
94b18763 | 9133 | auto &pi = target->project_inode(); |
94b18763 | 9134 | ti = &pi.inode; |
11fdf7f2 TL |
9135 | mut->add_projected_inode(target); |
9136 | ti->version = target->pre_dirty(); | |
9137 | projected = true; | |
9138 | } else { | |
7c673cae | 9139 | ti = target->get_projected_inode(); |
11fdf7f2 TL |
9140 | projected = false; |
9141 | } | |
7c673cae | 9142 | if (ti->ctime == rollback.ctime) |
91327a77 | 9143 | ti->ctime = rollback.orig_dest.old_ctime; |
7c673cae FG |
9144 | if (MDS_INO_IS_STRAY(rollback.orig_src.dirfrag.ino)) { |
9145 | if (MDS_INO_IS_STRAY(rollback.orig_dest.dirfrag.ino)) | |
11fdf7f2 | 9146 | ceph_assert(!rollback.orig_dest.ino && !rollback.orig_dest.remote_ino); |
7c673cae | 9147 | else |
11fdf7f2 | 9148 | ceph_assert(rollback.orig_dest.remote_ino && |
7c673cae FG |
9149 | rollback.orig_dest.remote_ino == rollback.orig_src.ino); |
9150 | } else | |
9151 | ti->nlink++; | |
11fdf7f2 TL |
9152 | |
9153 | if (rollback.desti_snapbl.length() && target->snaprealm) { | |
9154 | bool hadrealm; | |
9155 | auto p = rollback.desti_snapbl.cbegin(); | |
9156 | decode(hadrealm, p); | |
9157 | if (hadrealm) { | |
9158 | if (projected && !mds->is_resolve()) { | |
9159 | sr_t *new_srnode = new sr_t(); | |
9160 | decode(*new_srnode, p); | |
9161 | target->project_snaprealm(new_srnode); | |
9162 | } else | |
9163 | decode(target->snaprealm->srnode, p); | |
9164 | } else { | |
9165 | SnapRealm *realm; | |
9166 | if (rollback.orig_dest.ino) { | |
9167 | ceph_assert(destdir); | |
9168 | realm = destdir->get_inode()->find_snaprealm(); | |
9169 | } else { | |
9170 | realm = target->snaprealm->parent; | |
9171 | } | |
9172 | if (!mds->is_resolve()) | |
9173 | mdcache->prepare_realm_merge(target->snaprealm, realm, splits[1]); | |
9174 | if (projected) | |
9175 | target->project_snaprealm(NULL); | |
9176 | else | |
9177 | target->snaprealm->merge_to(realm); | |
9178 | } | |
9179 | } | |
7c673cae FG |
9180 | } |
9181 | ||
9182 | if (srcdn) | |
9183 | dout(0) << " srcdn back to " << *srcdn << dendl; | |
9184 | if (in) | |
9185 | dout(0) << " srci back to " << *in << dendl; | |
9186 | if (destdn) | |
9187 | dout(0) << " destdn back to " << *destdn << dendl; | |
9188 | if (target) | |
9189 | dout(0) << " desti back to " << *target << dendl; | |
9190 | ||
9191 | // journal it | |
9192 | ESlaveUpdate *le = new ESlaveUpdate(mdlog, "slave_rename_rollback", rollback.reqid, master, | |
9193 | ESlaveUpdate::OP_ROLLBACK, ESlaveUpdate::RENAME); | |
9194 | mdlog->start_entry(le); | |
9195 | ||
9196 | if (srcdn && (srcdn->authority().first == whoami || force_journal_src)) { | |
9197 | le->commit.add_dir_context(srcdir); | |
9198 | if (rollback.orig_src.ino) | |
9199 | le->commit.add_primary_dentry(srcdn, 0, true); | |
9200 | else | |
9201 | le->commit.add_remote_dentry(srcdn, true); | |
9202 | } | |
9203 | ||
9204 | if (!rollback.orig_src.ino && // remote linkage | |
9205 | in && in->authority().first == whoami) { | |
9206 | le->commit.add_dir_context(in->get_projected_parent_dir()); | |
9207 | le->commit.add_primary_dentry(in->get_projected_parent_dn(), in, true); | |
9208 | } | |
9209 | ||
9210 | if (force_journal_dest) { | |
11fdf7f2 | 9211 | ceph_assert(rollback.orig_dest.ino); |
7c673cae FG |
9212 | le->commit.add_dir_context(destdir); |
9213 | le->commit.add_primary_dentry(destdn, 0, true); | |
9214 | } | |
9215 | ||
9216 | // slave: no need to journal straydn | |
9217 | ||
9218 | if (target && target != in && target->authority().first == whoami) { | |
11fdf7f2 | 9219 | ceph_assert(rollback.orig_dest.remote_ino); |
7c673cae FG |
9220 | le->commit.add_dir_context(target->get_projected_parent_dir()); |
9221 | le->commit.add_primary_dentry(target->get_projected_parent_dn(), target, true); | |
9222 | } | |
9223 | ||
9224 | if (in && in->is_dir() && (srcdn->authority().first == whoami || force_journal_src)) { | |
9225 | dout(10) << " noting renamed dir ino " << in->ino() << " in metablob" << dendl; | |
9226 | le->commit.renamed_dirino = in->ino(); | |
9227 | if (srcdn->authority().first == whoami) { | |
9228 | list<CDir*> ls; | |
9229 | in->get_dirfrags(ls); | |
9230 | for (list<CDir*>::iterator p = ls.begin(); p != ls.end(); ++p) { | |
9231 | CDir *dir = *p; | |
9232 | if (!dir->is_auth()) | |
9233 | le->commit.renamed_dir_frags.push_back(dir->get_frag()); | |
9234 | } | |
9235 | dout(10) << " noting renamed dir open frags " << le->commit.renamed_dir_frags << dendl; | |
9236 | } | |
9237 | } else if (force_journal_dest) { | |
9238 | dout(10) << " noting rename target ino " << target->ino() << " in metablob" << dendl; | |
9239 | le->commit.renamed_dirino = target->ino(); | |
9240 | } | |
9241 | ||
9242 | if (target && target->is_dir()) { | |
11fdf7f2 | 9243 | ceph_assert(destdn); |
7c673cae FG |
9244 | mdcache->project_subtree_rename(target, straydir, destdir); |
9245 | } | |
9246 | ||
9247 | if (in && in->is_dir()) { | |
11fdf7f2 | 9248 | ceph_assert(srcdn); |
7c673cae FG |
9249 | mdcache->project_subtree_rename(in, destdir, srcdir); |
9250 | } | |
9251 | ||
9252 | if (mdr && !mdr->more()->slave_update_journaled) { | |
11fdf7f2 | 9253 | ceph_assert(le->commit.empty()); |
7c673cae FG |
9254 | mdlog->cancel_entry(le); |
9255 | mut->ls = NULL; | |
11fdf7f2 | 9256 | _rename_rollback_finish(mut, mdr, srcdn, srcdnpv, destdn, straydn, splits, finish_mdr); |
7c673cae | 9257 | } else { |
11fdf7f2 | 9258 | ceph_assert(!le->commit.empty()); |
7c673cae FG |
9259 | if (mdr) |
9260 | mdr->more()->slave_update_journaled = false; | |
11fdf7f2 TL |
9261 | MDSLogContextBase *fin = new C_MDS_LoggedRenameRollback(this, mut, mdr, |
9262 | srcdn, srcdnpv, destdn, straydn, | |
9263 | splits, finish_mdr); | |
7c673cae FG |
9264 | submit_mdlog_entry(le, fin, mdr, __func__); |
9265 | mdlog->flush(); | |
9266 | } | |
9267 | } | |
9268 | ||
9269 | void Server::_rename_rollback_finish(MutationRef& mut, MDRequestRef& mdr, CDentry *srcdn, | |
11fdf7f2 TL |
9270 | version_t srcdnpv, CDentry *destdn, CDentry *straydn, |
9271 | map<client_t,MClientSnap::ref> splits[2], bool finish_mdr) | |
7c673cae FG |
9272 | { |
9273 | dout(10) << "_rename_rollback_finish " << mut->reqid << dendl; | |
9274 | ||
9275 | if (straydn) { | |
9276 | straydn->get_dir()->unlink_inode(straydn); | |
9277 | straydn->pop_projected_linkage(); | |
9278 | } | |
9279 | if (destdn) { | |
9280 | destdn->get_dir()->unlink_inode(destdn); | |
9281 | destdn->pop_projected_linkage(); | |
9282 | } | |
9283 | if (srcdn) { | |
9284 | srcdn->pop_projected_linkage(); | |
11fdf7f2 | 9285 | if (srcdn->authority().first == mds->get_nodeid()) { |
7c673cae | 9286 | srcdn->mark_dirty(srcdnpv, mut->ls); |
11fdf7f2 TL |
9287 | if (srcdn->get_linkage()->is_primary()) |
9288 | srcdn->get_linkage()->get_inode()->state_set(CInode::STATE_AUTH); | |
9289 | } | |
7c673cae FG |
9290 | } |
9291 | ||
9292 | mut->apply(); | |
9293 | ||
9294 | if (srcdn && srcdn->get_linkage()->is_primary()) { | |
9295 | CInode *in = srcdn->get_linkage()->get_inode(); | |
7c673cae | 9296 | if (in && in->is_dir()) { |
11fdf7f2 | 9297 | ceph_assert(destdn); |
7c673cae FG |
9298 | mdcache->adjust_subtree_after_rename(in, destdn->get_dir(), true); |
9299 | } | |
9300 | } | |
9301 | ||
9302 | if (destdn) { | |
9303 | CInode *oldin = destdn->get_linkage()->get_inode(); | |
9304 | // update subtree map? | |
9305 | if (oldin && oldin->is_dir()) { | |
11fdf7f2 | 9306 | ceph_assert(straydn); |
7c673cae FG |
9307 | mdcache->adjust_subtree_after_rename(oldin, straydn->get_dir(), true); |
9308 | } | |
9309 | } | |
9310 | ||
9311 | if (mds->is_resolve()) { | |
9312 | CDir *root = NULL; | |
9313 | if (straydn) | |
9314 | root = mdcache->get_subtree_root(straydn->get_dir()); | |
9315 | else if (destdn) | |
9316 | root = mdcache->get_subtree_root(destdn->get_dir()); | |
9317 | if (root) | |
9318 | mdcache->try_trim_non_auth_subtree(root); | |
11fdf7f2 TL |
9319 | } else { |
9320 | mdcache->send_snaps(splits[1]); | |
9321 | mdcache->send_snaps(splits[0]); | |
7c673cae FG |
9322 | } |
9323 | ||
9324 | if (mdr) { | |
11fdf7f2 | 9325 | MDSContext::vec finished; |
7c673cae FG |
9326 | if (mdr->more()->is_ambiguous_auth) { |
9327 | if (srcdn->is_auth()) | |
9328 | mdr->more()->rename_inode->unfreeze_inode(finished); | |
9329 | ||
9330 | mdr->more()->rename_inode->clear_ambiguous_auth(finished); | |
9331 | mdr->more()->is_ambiguous_auth = false; | |
9332 | } | |
9333 | mds->queue_waiters(finished); | |
9334 | if (finish_mdr || mdr->aborted) | |
9335 | mdcache->request_finish(mdr); | |
9336 | else | |
9337 | mdr->more()->slave_rolling_back = false; | |
9338 | } | |
9339 | ||
9340 | mdcache->finish_rollback(mut->reqid); | |
9341 | ||
9342 | mut->cleanup(); | |
9343 | } | |
9344 | ||
11fdf7f2 | 9345 | void Server::handle_slave_rename_prep_ack(MDRequestRef& mdr, const MMDSSlaveRequest::const_ref &ack) |
7c673cae FG |
9346 | { |
9347 | dout(10) << "handle_slave_rename_prep_ack " << *mdr | |
9348 | << " witnessed by " << ack->get_source() | |
9349 | << " " << *ack << dendl; | |
9350 | mds_rank_t from = mds_rank_t(ack->get_source().num()); | |
9351 | ||
9352 | // note slave | |
9353 | mdr->more()->slaves.insert(from); | |
9354 | if (mdr->more()->srcdn_auth_mds == from && | |
9355 | mdr->more()->is_remote_frozen_authpin && | |
9356 | !mdr->more()->is_ambiguous_auth) { | |
9357 | mdr->set_ambiguous_auth(mdr->more()->rename_inode); | |
9358 | } | |
9359 | ||
9360 | // witnessed? or add extra witnesses? | |
11fdf7f2 | 9361 | ceph_assert(mdr->more()->witnessed.count(from) == 0); |
31f18b77 FG |
9362 | if (ack->is_interrupted()) { |
9363 | dout(10) << " slave request interrupted, noop" << dendl; | |
9364 | } else if (ack->witnesses.empty()) { | |
7c673cae FG |
9365 | mdr->more()->witnessed.insert(from); |
9366 | if (!ack->is_not_journaled()) | |
9367 | mdr->more()->has_journaled_slaves = true; | |
9368 | } else { | |
9369 | dout(10) << " extra witnesses (srcdn replicas) are " << ack->witnesses << dendl; | |
11fdf7f2 | 9370 | mdr->more()->extra_witnesses = ack->witnesses; |
7c673cae FG |
9371 | mdr->more()->extra_witnesses.erase(mds->get_nodeid()); // not me! |
9372 | } | |
9373 | ||
9374 | // srci import? | |
9375 | if (ack->inode_export.length()) { | |
9376 | dout(10) << " got srci import" << dendl; | |
11fdf7f2 | 9377 | mdr->more()->inode_import.share(ack->inode_export); |
7c673cae FG |
9378 | mdr->more()->inode_import_v = ack->inode_export_v; |
9379 | } | |
9380 | ||
9381 | // remove from waiting list | |
11fdf7f2 | 9382 | ceph_assert(mdr->more()->waiting_on_slave.count(from)); |
7c673cae FG |
9383 | mdr->more()->waiting_on_slave.erase(from); |
9384 | ||
9385 | if (mdr->more()->waiting_on_slave.empty()) | |
9386 | dispatch_client_request(mdr); // go again! | |
9387 | else | |
9388 | dout(10) << "still waiting on slaves " << mdr->more()->waiting_on_slave << dendl; | |
9389 | } | |
9390 | ||
11fdf7f2 | 9391 | void Server::handle_slave_rename_notify_ack(MDRequestRef& mdr, const MMDSSlaveRequest::const_ref &ack) |
7c673cae FG |
9392 | { |
9393 | dout(10) << "handle_slave_rename_notify_ack " << *mdr << " from mds." | |
9394 | << ack->get_source() << dendl; | |
11fdf7f2 | 9395 | ceph_assert(mdr->is_slave()); |
7c673cae FG |
9396 | mds_rank_t from = mds_rank_t(ack->get_source().num()); |
9397 | ||
9398 | if (mdr->more()->waiting_on_slave.count(from)) { | |
9399 | mdr->more()->waiting_on_slave.erase(from); | |
9400 | ||
9401 | if (mdr->more()->waiting_on_slave.empty()) { | |
9402 | if (mdr->slave_request) | |
9403 | dispatch_slave_request(mdr); | |
9404 | } else | |
9405 | dout(10) << " still waiting for rename notify acks from " | |
9406 | << mdr->more()->waiting_on_slave << dendl; | |
9407 | } | |
9408 | } | |
9409 | ||
9410 | void Server::_slave_rename_sessions_flushed(MDRequestRef& mdr) | |
9411 | { | |
9412 | dout(10) << "_slave_rename_sessions_flushed " << *mdr << dendl; | |
9413 | ||
9414 | if (mdr->more()->waiting_on_slave.count(MDS_RANK_NONE)) { | |
9415 | mdr->more()->waiting_on_slave.erase(MDS_RANK_NONE); | |
9416 | ||
9417 | if (mdr->more()->waiting_on_slave.empty()) { | |
9418 | if (mdr->slave_request) | |
9419 | dispatch_slave_request(mdr); | |
9420 | } else | |
9421 | dout(10) << " still waiting for rename notify acks from " | |
9422 | << mdr->more()->waiting_on_slave << dendl; | |
9423 | } | |
9424 | } | |
9425 | ||
9426 | // snaps | |
9427 | /* This function takes responsibility for the passed mdr*/ | |
9428 | void Server::handle_client_lssnap(MDRequestRef& mdr) | |
9429 | { | |
11fdf7f2 | 9430 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
9431 | |
9432 | // traverse to path | |
9433 | CInode *diri = mdcache->get_inode(req->get_filepath().get_ino()); | |
9434 | if (!diri || diri->state_test(CInode::STATE_PURGING)) { | |
9435 | respond_to_request(mdr, -ESTALE); | |
9436 | return; | |
9437 | } | |
9438 | if (!diri->is_auth()) { | |
9439 | mdcache->request_forward(mdr, diri->authority().first); | |
9440 | return; | |
9441 | } | |
9442 | if (!diri->is_dir()) { | |
9443 | respond_to_request(mdr, -ENOTDIR); | |
9444 | return; | |
9445 | } | |
9446 | dout(10) << "lssnap on " << *diri << dendl; | |
9447 | ||
9448 | // lock snap | |
11fdf7f2 TL |
9449 | MutationImpl::LockOpVec lov; |
9450 | mds->locker->include_snap_rdlocks(diri, lov); | |
9451 | if (!mds->locker->acquire_locks(mdr, lov)) | |
7c673cae FG |
9452 | return; |
9453 | ||
9454 | if (!check_access(mdr, diri, MAY_READ)) | |
9455 | return; | |
9456 | ||
9457 | SnapRealm *realm = diri->find_snaprealm(); | |
11fdf7f2 | 9458 | map<snapid_t,const SnapInfo*> infomap; |
7c673cae FG |
9459 | realm->get_snap_info(infomap, diri->get_oldest_snap()); |
9460 | ||
9461 | unsigned max_entries = req->head.args.readdir.max_entries; | |
9462 | if (!max_entries) | |
9463 | max_entries = infomap.size(); | |
9464 | int max_bytes = req->head.args.readdir.max_bytes; | |
9465 | if (!max_bytes) | |
9466 | // make sure at least one item can be encoded | |
11fdf7f2 | 9467 | max_bytes = (512 << 10) + g_conf()->mds_max_xattr_pairs_size; |
7c673cae FG |
9468 | |
9469 | __u64 last_snapid = 0; | |
9470 | string offset_str = req->get_path2(); | |
9471 | if (!offset_str.empty()) | |
9472 | last_snapid = realm->resolve_snapname(offset_str, diri->ino()); | |
9473 | ||
11fdf7f2 | 9474 | //Empty DirStat |
7c673cae | 9475 | bufferlist dirbl; |
11fdf7f2 TL |
9476 | static DirStat empty; |
9477 | CDir::encode_dirstat(dirbl, mdr->session->info, empty); | |
7c673cae FG |
9478 | |
9479 | max_bytes -= dirbl.length() - sizeof(__u32) + sizeof(__u8) * 2; | |
9480 | ||
9481 | __u32 num = 0; | |
9482 | bufferlist dnbl; | |
11fdf7f2 | 9483 | auto p = infomap.upper_bound(last_snapid); |
7c673cae FG |
9484 | for (; p != infomap.end() && num < max_entries; ++p) { |
9485 | dout(10) << p->first << " -> " << *p->second << dendl; | |
9486 | ||
9487 | // actual | |
9488 | string snap_name; | |
9489 | if (p->second->ino == diri->ino()) | |
11fdf7f2 | 9490 | snap_name = p->second->name; |
7c673cae | 9491 | else |
11fdf7f2 | 9492 | snap_name = p->second->get_long_name(); |
7c673cae FG |
9493 | |
9494 | unsigned start_len = dnbl.length(); | |
9495 | if (int(start_len + snap_name.length() + sizeof(__u32) + sizeof(LeaseStat)) > max_bytes) | |
9496 | break; | |
9497 | ||
11fdf7f2 TL |
9498 | encode(snap_name, dnbl); |
9499 | //infinite lease | |
9500 | LeaseStat e(-1, -1, 0); | |
9501 | mds->locker->encode_lease(dnbl, mdr->session->info, e); | |
9502 | dout(20) << "encode_infinite_lease" << dendl; | |
7c673cae FG |
9503 | |
9504 | int r = diri->encode_inodestat(dnbl, mdr->session, realm, p->first, max_bytes - (int)dnbl.length()); | |
9505 | if (r < 0) { | |
9506 | bufferlist keep; | |
9507 | keep.substr_of(dnbl, 0, start_len); | |
9508 | dnbl.swap(keep); | |
9509 | break; | |
9510 | } | |
9511 | ++num; | |
9512 | } | |
9513 | ||
11fdf7f2 | 9514 | encode(num, dirbl); |
7c673cae FG |
9515 | __u16 flags = 0; |
9516 | if (p == infomap.end()) { | |
9517 | flags = CEPH_READDIR_FRAG_END; | |
9518 | if (last_snapid == 0) | |
9519 | flags |= CEPH_READDIR_FRAG_COMPLETE; | |
9520 | } | |
11fdf7f2 | 9521 | encode(flags, dirbl); |
7c673cae FG |
9522 | dirbl.claim_append(dnbl); |
9523 | ||
9524 | mdr->reply_extra_bl = dirbl; | |
9525 | mdr->tracei = diri; | |
9526 | respond_to_request(mdr, 0); | |
9527 | } | |
9528 | ||
9529 | ||
9530 | // MKSNAP | |
9531 | ||
9532 | struct C_MDS_mksnap_finish : public ServerLogContext { | |
9533 | CInode *diri; | |
9534 | SnapInfo info; | |
9535 | C_MDS_mksnap_finish(Server *s, MDRequestRef& r, CInode *di, SnapInfo &i) : | |
9536 | ServerLogContext(s, r), diri(di), info(i) {} | |
9537 | void finish(int r) override { | |
9538 | server->_mksnap_finish(mdr, diri, info); | |
9539 | } | |
9540 | }; | |
9541 | ||
9542 | /* This function takes responsibility for the passed mdr*/ | |
9543 | void Server::handle_client_mksnap(MDRequestRef& mdr) | |
9544 | { | |
11fdf7f2 TL |
9545 | const MClientRequest::const_ref &req = mdr->client_request; |
9546 | // make sure we have as new a map as the client | |
9547 | if (req->get_mdsmap_epoch() > mds->mdsmap->get_epoch()) { | |
9548 | mds->wait_for_mdsmap(req->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache, mdr)); | |
9549 | return; | |
9550 | } | |
7c673cae FG |
9551 | if (!mds->mdsmap->allows_snaps()) { |
9552 | // you can't make snapshots until you set an option right now | |
9553 | respond_to_request(mdr, -EPERM); | |
9554 | return; | |
9555 | } | |
9556 | ||
7c673cae FG |
9557 | CInode *diri = mdcache->get_inode(req->get_filepath().get_ino()); |
9558 | if (!diri || diri->state_test(CInode::STATE_PURGING)) { | |
9559 | respond_to_request(mdr, -ESTALE); | |
9560 | return; | |
9561 | } | |
9562 | ||
9563 | if (!diri->is_auth()) { // fw to auth? | |
9564 | mdcache->request_forward(mdr, diri->authority().first); | |
9565 | return; | |
9566 | } | |
9567 | ||
9568 | // dir only | |
9569 | if (!diri->is_dir()) { | |
9570 | respond_to_request(mdr, -ENOTDIR); | |
9571 | return; | |
9572 | } | |
9573 | if (diri->is_system() && !diri->is_root()) { | |
9574 | // no snaps in system dirs (root is ok) | |
9575 | respond_to_request(mdr, -EPERM); | |
9576 | return; | |
9577 | } | |
9578 | ||
11fdf7f2 | 9579 | std::string_view snapname = req->get_filepath().last_dentry(); |
7c673cae | 9580 | |
11fdf7f2 | 9581 | if (mdr->client_request->get_caller_uid() < g_conf()->mds_snap_min_uid || mdr->client_request->get_caller_uid() > g_conf()->mds_snap_max_uid) { |
7c673cae FG |
9582 | dout(20) << "mksnap " << snapname << " on " << *diri << " denied to uid " << mdr->client_request->get_caller_uid() << dendl; |
9583 | respond_to_request(mdr, -EPERM); | |
9584 | return; | |
9585 | } | |
9586 | ||
9587 | dout(10) << "mksnap " << snapname << " on " << *diri << dendl; | |
9588 | ||
9589 | // lock snap | |
11fdf7f2 | 9590 | MutationImpl::LockOpVec lov; |
7c673cae | 9591 | |
11fdf7f2 TL |
9592 | mds->locker->include_snap_rdlocks(diri, lov); |
9593 | lov.erase_rdlock(&diri->snaplock); | |
9594 | lov.add_xlock(&diri->snaplock); | |
7c673cae | 9595 | |
11fdf7f2 | 9596 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
9597 | return; |
9598 | ||
11fdf7f2 | 9599 | if (!check_access(mdr, diri, MAY_WRITE|MAY_SNAPSHOT)) |
7c673cae FG |
9600 | return; |
9601 | ||
9602 | // make sure name is unique | |
9603 | if (diri->snaprealm && | |
9604 | diri->snaprealm->exists(snapname)) { | |
9605 | respond_to_request(mdr, -EEXIST); | |
9606 | return; | |
9607 | } | |
9608 | if (snapname.length() == 0 || | |
9609 | snapname[0] == '_') { | |
9610 | respond_to_request(mdr, -EINVAL); | |
9611 | return; | |
9612 | } | |
9613 | ||
9614 | // allocate a snapid | |
9615 | if (!mdr->more()->stid) { | |
9616 | // prepare an stid | |
9617 | mds->snapclient->prepare_create(diri->ino(), snapname, | |
9618 | mdr->get_mds_stamp(), | |
9619 | &mdr->more()->stid, &mdr->more()->snapidbl, | |
9620 | new C_MDS_RetryRequest(mdcache, mdr)); | |
9621 | return; | |
9622 | } | |
9623 | ||
9624 | version_t stid = mdr->more()->stid; | |
9625 | snapid_t snapid; | |
11fdf7f2 TL |
9626 | auto p = mdr->more()->snapidbl.cbegin(); |
9627 | decode(snapid, p); | |
7c673cae FG |
9628 | dout(10) << " stid " << stid << " snapid " << snapid << dendl; |
9629 | ||
11fdf7f2 TL |
9630 | ceph_assert(mds->snapclient->get_cached_version() >= stid); |
9631 | ||
7c673cae FG |
9632 | // journal |
9633 | SnapInfo info; | |
9634 | info.ino = diri->ino(); | |
9635 | info.snapid = snapid; | |
11fdf7f2 | 9636 | info.name = snapname; |
7c673cae FG |
9637 | info.stamp = mdr->get_op_stamp(); |
9638 | ||
94b18763 | 9639 | auto &pi = diri->project_inode(false, true); |
91327a77 AA |
9640 | pi.inode.ctime = info.stamp; |
9641 | if (info.stamp > pi.inode.rstat.rctime) | |
9642 | pi.inode.rstat.rctime = info.stamp; | |
11fdf7f2 | 9643 | pi.inode.rstat.rsnaps++; |
94b18763 | 9644 | pi.inode.version = diri->pre_dirty(); |
7c673cae FG |
9645 | |
9646 | // project the snaprealm | |
94b18763 FG |
9647 | auto &newsnap = *pi.snapnode; |
9648 | newsnap.created = snapid; | |
9649 | auto em = newsnap.snaps.emplace(std::piecewise_construct, std::forward_as_tuple(snapid), std::forward_as_tuple(info)); | |
9650 | if (!em.second) | |
9651 | em.first->second = info; | |
9652 | newsnap.seq = snapid; | |
9653 | newsnap.last_created = snapid; | |
7c673cae FG |
9654 | |
9655 | // journal the inode changes | |
9656 | mdr->ls = mdlog->get_current_segment(); | |
9657 | EUpdate *le = new EUpdate(mdlog, "mksnap"); | |
9658 | mdlog->start_entry(le); | |
9659 | ||
9660 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
9661 | le->metablob.add_table_transaction(TABLE_SNAP, stid); | |
9662 | mdcache->predirty_journal_parents(mdr, &le->metablob, diri, 0, PREDIRTY_PRIMARY, false); | |
9663 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, diri); | |
9664 | ||
9665 | // journal the snaprealm changes | |
9666 | submit_mdlog_entry(le, new C_MDS_mksnap_finish(this, mdr, diri, info), | |
9667 | mdr, __func__); | |
9668 | mdlog->flush(); | |
9669 | } | |
9670 | ||
9671 | void Server::_mksnap_finish(MDRequestRef& mdr, CInode *diri, SnapInfo &info) | |
9672 | { | |
9673 | dout(10) << "_mksnap_finish " << *mdr << " " << info << dendl; | |
9674 | ||
9675 | int op = (diri->snaprealm? CEPH_SNAP_OP_CREATE : CEPH_SNAP_OP_SPLIT); | |
9676 | ||
9677 | diri->pop_and_dirty_projected_inode(mdr->ls); | |
9678 | mdr->apply(); | |
9679 | ||
9680 | mds->snapclient->commit(mdr->more()->stid, mdr->ls); | |
9681 | ||
9682 | // create snap | |
9683 | dout(10) << "snaprealm now " << *diri->snaprealm << dendl; | |
9684 | ||
11fdf7f2 TL |
9685 | // notify other mds |
9686 | mdcache->send_snap_update(diri, mdr->more()->stid, op); | |
9687 | ||
7c673cae FG |
9688 | mdcache->do_realm_invalidate_and_update_notify(diri, op); |
9689 | ||
9690 | // yay | |
9691 | mdr->in[0] = diri; | |
9692 | mdr->snapid = info.snapid; | |
9693 | mdr->tracei = diri; | |
9694 | respond_to_request(mdr, 0); | |
9695 | } | |
9696 | ||
9697 | ||
9698 | // RMSNAP | |
9699 | ||
9700 | struct C_MDS_rmsnap_finish : public ServerLogContext { | |
9701 | CInode *diri; | |
9702 | snapid_t snapid; | |
9703 | C_MDS_rmsnap_finish(Server *s, MDRequestRef& r, CInode *di, snapid_t sn) : | |
9704 | ServerLogContext(s, r), diri(di), snapid(sn) {} | |
9705 | void finish(int r) override { | |
9706 | server->_rmsnap_finish(mdr, diri, snapid); | |
9707 | } | |
9708 | }; | |
9709 | ||
9710 | /* This function takes responsibility for the passed mdr*/ | |
9711 | void Server::handle_client_rmsnap(MDRequestRef& mdr) | |
9712 | { | |
11fdf7f2 | 9713 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
9714 | |
9715 | CInode *diri = mdcache->get_inode(req->get_filepath().get_ino()); | |
9716 | if (!diri || diri->state_test(CInode::STATE_PURGING)) { | |
9717 | respond_to_request(mdr, -ESTALE); | |
9718 | return; | |
9719 | } | |
9720 | if (!diri->is_auth()) { // fw to auth? | |
9721 | mdcache->request_forward(mdr, diri->authority().first); | |
9722 | return; | |
9723 | } | |
9724 | if (!diri->is_dir()) { | |
9725 | respond_to_request(mdr, -ENOTDIR); | |
9726 | return; | |
9727 | } | |
9728 | ||
11fdf7f2 | 9729 | std::string_view snapname = req->get_filepath().last_dentry(); |
7c673cae | 9730 | |
11fdf7f2 | 9731 | if (mdr->client_request->get_caller_uid() < g_conf()->mds_snap_min_uid || mdr->client_request->get_caller_uid() > g_conf()->mds_snap_max_uid) { |
7c673cae FG |
9732 | dout(20) << "rmsnap " << snapname << " on " << *diri << " denied to uid " << mdr->client_request->get_caller_uid() << dendl; |
9733 | respond_to_request(mdr, -EPERM); | |
9734 | return; | |
9735 | } | |
9736 | ||
9737 | dout(10) << "rmsnap " << snapname << " on " << *diri << dendl; | |
9738 | ||
9739 | // does snap exist? | |
9740 | if (snapname.length() == 0 || snapname[0] == '_') { | |
9741 | respond_to_request(mdr, -EINVAL); // can't prune a parent snap, currently. | |
9742 | return; | |
9743 | } | |
9744 | if (!diri->snaprealm || !diri->snaprealm->exists(snapname)) { | |
9745 | respond_to_request(mdr, -ENOENT); | |
9746 | return; | |
9747 | } | |
9748 | snapid_t snapid = diri->snaprealm->resolve_snapname(snapname, diri->ino()); | |
9749 | dout(10) << " snapname " << snapname << " is " << snapid << dendl; | |
9750 | ||
11fdf7f2 TL |
9751 | MutationImpl::LockOpVec lov; |
9752 | mds->locker->include_snap_rdlocks(diri, lov); | |
9753 | lov.erase_rdlock(&diri->snaplock); | |
9754 | lov.add_xlock(&diri->snaplock); | |
7c673cae | 9755 | |
11fdf7f2 | 9756 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
9757 | return; |
9758 | ||
11fdf7f2 | 9759 | if (!check_access(mdr, diri, MAY_WRITE|MAY_SNAPSHOT)) |
7c673cae FG |
9760 | return; |
9761 | ||
9762 | // prepare | |
9763 | if (!mdr->more()->stid) { | |
9764 | mds->snapclient->prepare_destroy(diri->ino(), snapid, | |
9765 | &mdr->more()->stid, &mdr->more()->snapidbl, | |
9766 | new C_MDS_RetryRequest(mdcache, mdr)); | |
9767 | return; | |
9768 | } | |
9769 | version_t stid = mdr->more()->stid; | |
11fdf7f2 | 9770 | auto p = mdr->more()->snapidbl.cbegin(); |
7c673cae | 9771 | snapid_t seq; |
11fdf7f2 | 9772 | decode(seq, p); |
7c673cae FG |
9773 | dout(10) << " stid is " << stid << ", seq is " << seq << dendl; |
9774 | ||
11fdf7f2 TL |
9775 | ceph_assert(mds->snapclient->get_cached_version() >= stid); |
9776 | ||
7c673cae | 9777 | // journal |
94b18763 FG |
9778 | auto &pi = diri->project_inode(false, true); |
9779 | pi.inode.version = diri->pre_dirty(); | |
91327a77 AA |
9780 | pi.inode.ctime = mdr->get_op_stamp(); |
9781 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
9782 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
11fdf7f2 | 9783 | pi.inode.rstat.rsnaps--; |
7c673cae FG |
9784 | |
9785 | mdr->ls = mdlog->get_current_segment(); | |
9786 | EUpdate *le = new EUpdate(mdlog, "rmsnap"); | |
9787 | mdlog->start_entry(le); | |
9788 | ||
9789 | // project the snaprealm | |
94b18763 FG |
9790 | auto &newnode = *pi.snapnode; |
9791 | newnode.snaps.erase(snapid); | |
9792 | newnode.seq = seq; | |
9793 | newnode.last_destroyed = seq; | |
7c673cae FG |
9794 | |
9795 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
9796 | le->metablob.add_table_transaction(TABLE_SNAP, stid); | |
9797 | mdcache->predirty_journal_parents(mdr, &le->metablob, diri, 0, PREDIRTY_PRIMARY, false); | |
9798 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, diri); | |
9799 | ||
9800 | submit_mdlog_entry(le, new C_MDS_rmsnap_finish(this, mdr, diri, snapid), | |
9801 | mdr, __func__); | |
9802 | mdlog->flush(); | |
9803 | } | |
9804 | ||
9805 | void Server::_rmsnap_finish(MDRequestRef& mdr, CInode *diri, snapid_t snapid) | |
9806 | { | |
9807 | dout(10) << "_rmsnap_finish " << *mdr << " " << snapid << dendl; | |
9808 | snapid_t stid = mdr->more()->stid; | |
11fdf7f2 | 9809 | auto p = mdr->more()->snapidbl.cbegin(); |
7c673cae | 9810 | snapid_t seq; |
11fdf7f2 | 9811 | decode(seq, p); |
7c673cae FG |
9812 | |
9813 | diri->pop_and_dirty_projected_inode(mdr->ls); | |
9814 | mdr->apply(); | |
9815 | ||
9816 | mds->snapclient->commit(stid, mdr->ls); | |
9817 | ||
9818 | dout(10) << "snaprealm now " << *diri->snaprealm << dendl; | |
9819 | ||
11fdf7f2 TL |
9820 | // notify other mds |
9821 | mdcache->send_snap_update(diri, mdr->more()->stid, CEPH_SNAP_OP_DESTROY); | |
9822 | ||
7c673cae FG |
9823 | mdcache->do_realm_invalidate_and_update_notify(diri, CEPH_SNAP_OP_DESTROY); |
9824 | ||
9825 | // yay | |
9826 | mdr->in[0] = diri; | |
9827 | respond_to_request(mdr, 0); | |
9828 | ||
9829 | // purge snapshot data | |
9830 | if (diri->snaprealm->have_past_parents_open()) | |
9831 | diri->purge_stale_snap_data(diri->snaprealm->get_snaps()); | |
9832 | } | |
9833 | ||
9834 | struct C_MDS_renamesnap_finish : public ServerLogContext { | |
9835 | CInode *diri; | |
9836 | snapid_t snapid; | |
9837 | C_MDS_renamesnap_finish(Server *s, MDRequestRef& r, CInode *di, snapid_t sn) : | |
9838 | ServerLogContext(s, r), diri(di), snapid(sn) {} | |
9839 | void finish(int r) override { | |
9840 | server->_renamesnap_finish(mdr, diri, snapid); | |
9841 | } | |
9842 | }; | |
9843 | ||
9844 | /* This function takes responsibility for the passed mdr*/ | |
9845 | void Server::handle_client_renamesnap(MDRequestRef& mdr) | |
9846 | { | |
11fdf7f2 | 9847 | const MClientRequest::const_ref &req = mdr->client_request; |
7c673cae FG |
9848 | if (req->get_filepath().get_ino() != req->get_filepath2().get_ino()) { |
9849 | respond_to_request(mdr, -EINVAL); | |
9850 | return; | |
9851 | } | |
9852 | ||
9853 | CInode *diri = mdcache->get_inode(req->get_filepath().get_ino()); | |
9854 | if (!diri || diri->state_test(CInode::STATE_PURGING)) { | |
9855 | respond_to_request(mdr, -ESTALE); | |
9856 | return; | |
9857 | } | |
9858 | ||
9859 | if (!diri->is_auth()) { // fw to auth? | |
9860 | mdcache->request_forward(mdr, diri->authority().first); | |
9861 | return; | |
9862 | } | |
9863 | ||
9864 | if (!diri->is_dir()) { // dir only | |
9865 | respond_to_request(mdr, -ENOTDIR); | |
9866 | return; | |
9867 | } | |
9868 | ||
11fdf7f2 TL |
9869 | if (mdr->client_request->get_caller_uid() < g_conf()->mds_snap_min_uid || |
9870 | mdr->client_request->get_caller_uid() > g_conf()->mds_snap_max_uid) { | |
7c673cae FG |
9871 | respond_to_request(mdr, -EPERM); |
9872 | return; | |
9873 | } | |
9874 | ||
11fdf7f2 TL |
9875 | std::string_view dstname = req->get_filepath().last_dentry(); |
9876 | std::string_view srcname = req->get_filepath2().last_dentry(); | |
7c673cae FG |
9877 | dout(10) << "renamesnap " << srcname << "->" << dstname << " on " << *diri << dendl; |
9878 | ||
9879 | if (srcname.length() == 0 || srcname[0] == '_') { | |
9880 | respond_to_request(mdr, -EINVAL); // can't rename a parent snap. | |
9881 | return; | |
9882 | } | |
9883 | if (!diri->snaprealm || !diri->snaprealm->exists(srcname)) { | |
9884 | respond_to_request(mdr, -ENOENT); | |
9885 | return; | |
9886 | } | |
9887 | if (dstname.length() == 0 || dstname[0] == '_') { | |
9888 | respond_to_request(mdr, -EINVAL); | |
9889 | return; | |
9890 | } | |
9891 | if (diri->snaprealm->exists(dstname)) { | |
9892 | respond_to_request(mdr, -EEXIST); | |
9893 | return; | |
9894 | } | |
9895 | ||
9896 | snapid_t snapid = diri->snaprealm->resolve_snapname(srcname, diri->ino()); | |
9897 | dout(10) << " snapname " << srcname << " is " << snapid << dendl; | |
9898 | ||
9899 | // lock snap | |
11fdf7f2 | 9900 | MutationImpl::LockOpVec lov; |
7c673cae | 9901 | |
11fdf7f2 TL |
9902 | mds->locker->include_snap_rdlocks(diri, lov); |
9903 | lov.erase_rdlock(&diri->snaplock); | |
9904 | lov.add_xlock(&diri->snaplock); | |
7c673cae | 9905 | |
11fdf7f2 | 9906 | if (!mds->locker->acquire_locks(mdr, lov)) |
7c673cae FG |
9907 | return; |
9908 | ||
11fdf7f2 | 9909 | if (!check_access(mdr, diri, MAY_WRITE|MAY_SNAPSHOT)) |
7c673cae FG |
9910 | return; |
9911 | ||
9912 | // prepare | |
9913 | if (!mdr->more()->stid) { | |
9914 | mds->snapclient->prepare_update(diri->ino(), snapid, dstname, utime_t(), | |
11fdf7f2 | 9915 | &mdr->more()->stid, |
7c673cae FG |
9916 | new C_MDS_RetryRequest(mdcache, mdr)); |
9917 | return; | |
9918 | } | |
9919 | ||
9920 | version_t stid = mdr->more()->stid; | |
11fdf7f2 TL |
9921 | dout(10) << " stid is " << stid << dendl; |
9922 | ||
9923 | ceph_assert(mds->snapclient->get_cached_version() >= stid); | |
7c673cae FG |
9924 | |
9925 | // journal | |
94b18763 | 9926 | auto &pi = diri->project_inode(false, true); |
91327a77 AA |
9927 | pi.inode.ctime = mdr->get_op_stamp(); |
9928 | if (mdr->get_op_stamp() > pi.inode.rstat.rctime) | |
9929 | pi.inode.rstat.rctime = mdr->get_op_stamp(); | |
94b18763 | 9930 | pi.inode.version = diri->pre_dirty(); |
7c673cae FG |
9931 | |
9932 | // project the snaprealm | |
94b18763 FG |
9933 | auto &newsnap = *pi.snapnode; |
9934 | auto it = newsnap.snaps.find(snapid); | |
11fdf7f2 TL |
9935 | ceph_assert(it != newsnap.snaps.end()); |
9936 | it->second.name = dstname; | |
7c673cae FG |
9937 | |
9938 | // journal the inode changes | |
9939 | mdr->ls = mdlog->get_current_segment(); | |
9940 | EUpdate *le = new EUpdate(mdlog, "renamesnap"); | |
9941 | mdlog->start_entry(le); | |
9942 | ||
9943 | le->metablob.add_client_req(req->get_reqid(), req->get_oldest_client_tid()); | |
9944 | le->metablob.add_table_transaction(TABLE_SNAP, stid); | |
9945 | mdcache->predirty_journal_parents(mdr, &le->metablob, diri, 0, PREDIRTY_PRIMARY, false); | |
9946 | mdcache->journal_dirty_inode(mdr.get(), &le->metablob, diri); | |
9947 | ||
9948 | // journal the snaprealm changes | |
9949 | submit_mdlog_entry(le, new C_MDS_renamesnap_finish(this, mdr, diri, snapid), | |
9950 | mdr, __func__); | |
9951 | mdlog->flush(); | |
9952 | } | |
9953 | ||
9954 | void Server::_renamesnap_finish(MDRequestRef& mdr, CInode *diri, snapid_t snapid) | |
9955 | { | |
9956 | dout(10) << "_renamesnap_finish " << *mdr << " " << snapid << dendl; | |
9957 | ||
9958 | diri->pop_and_dirty_projected_inode(mdr->ls); | |
9959 | mdr->apply(); | |
9960 | ||
9961 | mds->snapclient->commit(mdr->more()->stid, mdr->ls); | |
9962 | ||
9963 | dout(10) << "snaprealm now " << *diri->snaprealm << dendl; | |
9964 | ||
11fdf7f2 TL |
9965 | // notify other mds |
9966 | mdcache->send_snap_update(diri, mdr->more()->stid, CEPH_SNAP_OP_UPDATE); | |
9967 | ||
9968 | mdcache->do_realm_invalidate_and_update_notify(diri, CEPH_SNAP_OP_UPDATE); | |
7c673cae FG |
9969 | |
9970 | // yay | |
9971 | mdr->in[0] = diri; | |
9972 | mdr->tracei = diri; | |
9973 | mdr->snapid = snapid; | |
9974 | respond_to_request(mdr, 0); | |
9975 | } | |
9976 | ||
9977 | /** | |
9978 | * Return true if server is in state RECONNECT and this | |
9979 | * client has not yet reconnected. | |
9980 | */ | |
9981 | bool Server::waiting_for_reconnect(client_t c) const | |
9982 | { | |
9983 | return client_reconnect_gather.count(c) > 0; | |
9984 | } | |
9985 | ||
9986 | void Server::dump_reconnect_status(Formatter *f) const | |
9987 | { | |
9988 | f->open_object_section("reconnect_status"); | |
9989 | f->dump_stream("client_reconnect_gather") << client_reconnect_gather; | |
9990 | f->close_section(); | |
9991 | } |