1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
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.
15 #include <boost/lexical_cast.hpp>
16 #include "include/assert.h" // lexical_cast includes system assert.h
18 #include <boost/config/warning_disable.hpp>
19 #include <boost/fusion/include/std_pair.hpp>
27 #include "MDBalancer.h"
29 #include "SnapClient.h"
32 #include "msg/Messenger.h"
34 #include "osdc/Objecter.h"
36 #include "messages/MClientSession.h"
37 #include "messages/MClientRequest.h"
38 #include "messages/MClientReply.h"
39 #include "messages/MClientReconnect.h"
40 #include "messages/MClientCaps.h"
41 #include "messages/MClientSnap.h"
43 #include "messages/MMDSSlaveRequest.h"
45 #include "messages/MLock.h"
47 #include "events/EUpdate.h"
48 #include "events/ESlaveUpdate.h"
49 #include "events/ESession.h"
50 #include "events/EOpen.h"
51 #include "events/ECommitted.h"
53 #include "include/filepath.h"
54 #include "common/errno.h"
55 #include "common/Timer.h"
56 #include "common/perf_counters.h"
57 #include "include/compat.h"
58 #include "osd/OSDMap.h"
66 #include "common/config.h"
68 #define dout_context g_ceph_context
69 #define dout_subsys ceph_subsys_mds
71 #define dout_prefix *_dout << "mds." << mds->get_nodeid() << ".server "
73 class ServerContext
: public MDSInternalContextBase
{
76 MDSRank
*get_mds() override
82 explicit ServerContext(Server
*s
) : server(s
) {
83 assert(server
!= NULL
);
87 class ServerLogContext
: public MDSLogContextBase
{
90 MDSRank
*get_mds() override
96 void pre_finish(int r
) override
{
98 mdr
->mark_event("journal_committed: ");
101 explicit ServerLogContext(Server
*s
) : server(s
) {
102 assert(server
!= NULL
);
104 explicit ServerLogContext(Server
*s
, MDRequestRef
& r
) : server(s
), mdr(r
) {
105 assert(server
!= NULL
);
109 void Server::create_logger()
111 PerfCountersBuilder
plb(g_ceph_context
, "mds_server", l_mdss_first
, l_mdss_last
);
112 plb
.add_u64_counter(l_mdss_handle_client_request
,"handle_client_request",
113 "Client requests", "hcr", PerfCountersBuilder::PRIO_INTERESTING
);
114 plb
.add_u64_counter(l_mdss_handle_slave_request
, "handle_slave_request",
115 "Slave requests", "hsr", PerfCountersBuilder::PRIO_INTERESTING
);
116 plb
.add_u64_counter(l_mdss_handle_client_session
, "handle_client_session",
117 "Client session messages", "hcs", PerfCountersBuilder::PRIO_INTERESTING
);
118 plb
.add_u64_counter(l_mdss_dispatch_client_request
, "dispatch_client_request", "Client requests dispatched");
119 plb
.add_u64_counter(l_mdss_dispatch_slave_request
, "dispatch_server_request", "Server requests dispatched");
120 plb
.add_u64_counter(l_mdss_req_lookuphash
, "req_lookuphash",
121 "Request type lookup hash of inode");
122 plb
.add_u64_counter(l_mdss_req_lookupino
, "req_lookupino",
123 "Request type lookup inode");
124 plb
.add_u64_counter(l_mdss_req_lookupparent
, "req_lookupparent",
125 "Request type lookup parent");
126 plb
.add_u64_counter(l_mdss_req_lookupname
, "req_lookupname",
127 "Request type lookup name");
128 plb
.add_u64_counter(l_mdss_req_lookup
, "req_lookup",
129 "Request type lookup");
130 plb
.add_u64_counter(l_mdss_req_lookupsnap
, "req_lookupsnap",
131 "Request type lookup snapshot");
132 plb
.add_u64_counter(l_mdss_req_getattr
, "req_getattr",
133 "Request type get attribute");
134 plb
.add_u64_counter(l_mdss_req_setattr
, "req_setattr",
135 "Request type set attribute");
136 plb
.add_u64_counter(l_mdss_req_setlayout
, "req_setlayout",
137 "Request type set file layout");
138 plb
.add_u64_counter(l_mdss_req_setdirlayout
, "req_setdirlayout",
139 "Request type set directory layout");
140 plb
.add_u64_counter(l_mdss_req_setxattr
, "req_setxattr",
141 "Request type set extended attribute");
142 plb
.add_u64_counter(l_mdss_req_rmxattr
, "req_rmxattr",
143 "Request type remove extended attribute");
144 plb
.add_u64_counter(l_mdss_req_readdir
, "req_readdir",
145 "Request type read directory");
146 plb
.add_u64_counter(l_mdss_req_setfilelock
, "req_setfilelock",
147 "Request type set file lock");
148 plb
.add_u64_counter(l_mdss_req_getfilelock
, "req_getfilelock",
149 "Request type get file lock");
150 plb
.add_u64_counter(l_mdss_req_create
, "req_create",
151 "Request type create");
152 plb
.add_u64_counter(l_mdss_req_open
, "req_open",
153 "Request type open");
154 plb
.add_u64_counter(l_mdss_req_mknod
, "req_mknod",
155 "Request type make node");
156 plb
.add_u64_counter(l_mdss_req_link
, "req_link",
157 "Request type link");
158 plb
.add_u64_counter(l_mdss_req_unlink
, "req_unlink",
159 "Request type unlink");
160 plb
.add_u64_counter(l_mdss_req_rmdir
, "req_rmdir",
161 "Request type remove directory");
162 plb
.add_u64_counter(l_mdss_req_rename
, "req_rename",
163 "Request type rename");
164 plb
.add_u64_counter(l_mdss_req_mkdir
, "req_mkdir",
165 "Request type make directory");
166 plb
.add_u64_counter(l_mdss_req_symlink
, "req_symlink",
167 "Request type symbolic link");
168 plb
.add_u64_counter(l_mdss_req_lssnap
, "req_lssnap",
169 "Request type list snapshot");
170 plb
.add_u64_counter(l_mdss_req_mksnap
, "req_mksnap",
171 "Request type make snapshot");
172 plb
.add_u64_counter(l_mdss_req_rmsnap
, "req_rmsnap",
173 "Request type remove snapshot");
174 plb
.add_u64_counter(l_mdss_req_renamesnap
, "req_renamesnap",
175 "Request type rename snapshot");
176 logger
= plb
.create_perf_counters();
177 g_ceph_context
->get_perfcounters_collection()->add(logger
);
180 Server::Server(MDSRank
*m
) :
182 mdcache(mds
->mdcache
), mdlog(mds
->mdlog
),
185 reconnect_done(NULL
),
186 failed_reconnects(0),
187 reconnect_evicting(false),
188 terminating_sessions(false)
193 /* This function DOES put the passed message before returning*/
194 void Server::dispatch(Message
*m
)
196 switch (m
->get_type()) {
197 case CEPH_MSG_CLIENT_RECONNECT
:
198 handle_client_reconnect(static_cast<MClientReconnect
*>(m
));
203 if (!mds
->is_active() &&
204 !(mds
->is_stopping() && m
->get_source().is_mds())) {
205 if (m
->get_type() == CEPH_MSG_CLIENT_REQUEST
&&
206 (mds
->is_reconnect() || mds
->get_want_state() == CEPH_MDS_STATE_RECONNECT
)) {
207 MClientRequest
*req
= static_cast<MClientRequest
*>(m
);
208 Session
*session
= get_session(req
);
209 if (!session
|| session
->is_closed()) {
210 dout(5) << "session is closed, dropping " << req
->get_reqid() << dendl
;
214 bool queue_replay
= false;
215 if (req
->is_replay()) {
216 dout(3) << "queuing replayed op" << dendl
;
218 } else if (req
->get_retry_attempt()) {
219 // process completed request in clientreplay stage. The completed request
220 // might have created new file/directorie. This guarantees MDS sends a reply
221 // to client before other request modifies the new file/directorie.
222 if (session
->have_completed_request(req
->get_reqid().tid
, NULL
)) {
223 dout(3) << "queuing completed op" << dendl
;
226 // this request was created before the cap reconnect message, drop any embedded
228 req
->releases
.clear();
231 req
->mark_queued_for_replay();
232 mds
->enqueue_replay(new C_MDS_RetryMessage(mds
, m
));
237 bool wait_for_active
= true;
238 if (m
->get_type() == MSG_MDS_SLAVE_REQUEST
) {
239 // handle_slave_request() will wait if necessary
240 wait_for_active
= false;
241 } else if (mds
->is_clientreplay()) {
242 // session open requests need to be handled during replay,
243 // close requests need to be delayed
244 if ((m
->get_type() == CEPH_MSG_CLIENT_SESSION
&&
245 (static_cast<MClientSession
*>(m
))->get_op() != CEPH_SESSION_REQUEST_CLOSE
)) {
246 wait_for_active
= false;
247 } else if (m
->get_type() == CEPH_MSG_CLIENT_REQUEST
) {
248 MClientRequest
*req
= static_cast<MClientRequest
*>(m
);
249 if (req
->is_queued_for_replay()) {
250 wait_for_active
= false;
254 if (wait_for_active
) {
255 dout(3) << "not active yet, waiting" << dendl
;
256 mds
->wait_for_active(new C_MDS_RetryMessage(mds
, m
));
261 switch (m
->get_type()) {
262 case CEPH_MSG_CLIENT_SESSION
:
263 handle_client_session(static_cast<MClientSession
*>(m
));
265 case CEPH_MSG_CLIENT_REQUEST
:
266 handle_client_request(static_cast<MClientRequest
*>(m
));
268 case MSG_MDS_SLAVE_REQUEST
:
269 handle_slave_request(static_cast<MMDSSlaveRequest
*>(m
));
272 derr
<< "server unknown message " << m
->get_type() << dendl
;
273 assert(0 == "server unknown message");
279 // ----------------------------------------------------------
280 // SESSION management
282 class C_MDS_session_finish
: public ServerLogContext
{
287 interval_set
<inodeno_t
> inos
;
291 C_MDS_session_finish(Server
*srv
, Session
*se
, uint64_t sseq
, bool s
, version_t mv
, Context
*fin_
= NULL
) :
292 ServerLogContext(srv
), session(se
), state_seq(sseq
), open(s
), cmapv(mv
), inotablev(0), fin(fin_
) { }
293 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
) :
294 ServerLogContext(srv
), session(se
), state_seq(sseq
), open(s
), cmapv(mv
), inos(i
), inotablev(iv
), fin(fin_
) { }
295 void finish(int r
) override
{
297 server
->_session_logged(session
, state_seq
, open
, cmapv
, inos
, inotablev
);
304 Session
*Server::get_session(Message
*m
)
306 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
308 dout(20) << "get_session have " << session
<< " " << session
->info
.inst
309 << " state " << session
->get_state_name() << dendl
;
310 session
->put(); // not carry ref
312 dout(20) << "get_session dne for " << m
->get_source_inst() << dendl
;
317 /* This function DOES put the passed message before returning*/
318 void Server::handle_client_session(MClientSession
*m
)
321 bool blacklisted
= false;
322 Session
*session
= get_session(m
);
324 dout(3) << "handle_client_session " << *m
<< " from " << m
->get_source() << dendl
;
325 assert(m
->get_source().is_client()); // should _not_ come from an mds!
328 dout(0) << " ignoring sessionless msg " << *m
<< dendl
;
334 logger
->inc(l_mdss_handle_client_session
);
337 switch (m
->get_op()) {
338 case CEPH_SESSION_REQUEST_OPEN
:
339 if (session
->is_opening() ||
340 session
->is_open() ||
341 session
->is_stale() ||
342 session
->is_killing()) {
343 dout(10) << "currently open|opening|stale|killing, dropping this req" << dendl
;
347 assert(session
->is_closed() ||
348 session
->is_closing());
350 blacklisted
= mds
->objecter
->with_osdmap(
351 [session
](const OSDMap
&osd_map
) -> bool {
352 return osd_map
.is_blacklisted(session
->info
.inst
.addr
);
356 dout(10) << "ignoring blacklisted client " << session
->info
.inst
.addr
<< dendl
;
361 session
->set_client_metadata(m
->client_meta
);
362 dout(20) << __func__
<< " CEPH_SESSION_REQUEST_OPEN "
363 << session
->info
.client_metadata
.size() << " metadata entries:" << dendl
;
364 for (map
<string
, string
>::iterator i
= session
->info
.client_metadata
.begin();
365 i
!= session
->info
.client_metadata
.end(); ++i
) {
366 dout(20) << " " << i
->first
<< ": " << i
->second
<< dendl
;
369 // Special case for the 'root' metadata path; validate that the claimed
370 // root is actually within the caps of the session
371 if (session
->info
.client_metadata
.count("root")) {
372 const auto claimed_root
= session
->info
.client_metadata
.at("root");
373 // claimed_root has a leading "/" which we strip before passing
375 if (claimed_root
.empty() || claimed_root
[0] != '/' ||
376 !session
->auth_caps
.path_capable(claimed_root
.substr(1))) {
377 derr
<< __func__
<< " forbidden path claimed as mount root: "
378 << claimed_root
<< " by " << m
->get_source() << dendl
;
379 // Tell the client we're rejecting their open
380 mds
->send_message_client(new MClientSession(CEPH_SESSION_REJECT
), session
);
381 mds
->clog
->warn() << "client session with invalid root '" <<
382 claimed_root
<< "' denied (" << session
->info
.inst
<< ")";
384 // Drop out; don't record this session in SessionMap or journal it.
389 if (session
->is_closed())
390 mds
->sessionmap
.add_session(session
);
392 pv
= mds
->sessionmap
.mark_projected(session
);
393 sseq
= mds
->sessionmap
.set_state(session
, Session::STATE_OPENING
);
394 mds
->sessionmap
.touch_session(session
);
395 mdlog
->start_submit_entry(new ESession(m
->get_source_inst(), true, pv
, m
->client_meta
),
396 new C_MDS_session_finish(this, session
, sseq
, true, pv
));
400 case CEPH_SESSION_REQUEST_RENEWCAPS
:
401 if (session
->is_open() ||
402 session
->is_stale()) {
403 mds
->sessionmap
.touch_session(session
);
404 if (session
->is_stale()) {
405 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
406 mds
->locker
->resume_stale_caps(session
);
407 mds
->sessionmap
.touch_session(session
);
409 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_RENEWCAPS
, m
->get_seq()));
411 dout(10) << "ignoring renewcaps on non open|stale session (" << session
->get_state_name() << ")" << dendl
;
415 case CEPH_SESSION_REQUEST_CLOSE
:
417 if (session
->is_closed() ||
418 session
->is_closing() ||
419 session
->is_killing()) {
420 dout(10) << "already closed|closing|killing, dropping this req" << dendl
;
424 if (session
->is_importing()) {
425 dout(10) << "ignoring close req on importing session" << dendl
;
429 assert(session
->is_open() ||
430 session
->is_stale() ||
431 session
->is_opening());
432 if (m
->get_seq() < session
->get_push_seq()) {
433 dout(10) << "old push seq " << m
->get_seq() << " < " << session
->get_push_seq()
434 << ", dropping" << dendl
;
438 // We are getting a seq that is higher than expected.
439 // Handle the same as any other seqn error.
441 if (m
->get_seq() != session
->get_push_seq()) {
442 dout(0) << "old push seq " << m
->get_seq() << " != " << session
->get_push_seq()
443 << ", BUGGY!" << dendl
;
444 mds
->clog
->warn() << "incorrect push seq " << m
->get_seq() << " != "
445 << session
->get_push_seq() << ", dropping" << " from client : " << session
->get_human_name();
449 journal_close_session(session
, Session::STATE_CLOSING
, NULL
);
453 case CEPH_SESSION_FLUSHMSG_ACK
:
454 finish_flush_session(session
, m
->get_seq());
457 case CEPH_SESSION_REQUEST_FLUSH_MDLOG
:
467 void Server::flush_client_sessions(set
<client_t
>& client_set
, MDSGatherBuilder
& gather
)
469 for (set
<client_t
>::iterator p
= client_set
.begin(); p
!= client_set
.end(); ++p
) {
470 Session
*session
= mds
->sessionmap
.get_session(entity_name_t::CLIENT(p
->v
));
472 if (!session
->is_open() ||
473 !session
->connection
.get() ||
474 !session
->connection
->has_feature(CEPH_FEATURE_EXPORT_PEER
))
476 version_t seq
= session
->wait_for_flush(gather
.new_sub());
477 mds
->send_message_client(new MClientSession(CEPH_SESSION_FLUSHMSG
, seq
), session
);
481 void Server::finish_flush_session(Session
*session
, version_t seq
)
483 list
<MDSInternalContextBase
*> finished
;
484 session
->finish_flush(seq
, finished
);
485 mds
->queue_waiters(finished
);
488 void Server::_session_logged(Session
*session
, uint64_t state_seq
, bool open
, version_t pv
,
489 interval_set
<inodeno_t
>& inos
, version_t piv
)
491 dout(10) << "_session_logged " << session
->info
.inst
<< " state_seq " << state_seq
<< " " << (open
? "open":"close")
492 << " " << pv
<< dendl
;
495 assert(session
->is_closing() || session
->is_killing() ||
496 session
->is_opening()); // re-open closing session
497 session
->info
.prealloc_inos
.subtract(inos
);
498 mds
->inotable
->apply_release_ids(inos
);
499 assert(mds
->inotable
->get_version() == piv
);
502 mds
->sessionmap
.mark_dirty(session
);
505 if (session
->get_state_seq() != state_seq
) {
506 dout(10) << " journaled state_seq " << state_seq
<< " != current " << session
->get_state_seq()
507 << ", noop" << dendl
;
508 // close must have been canceled (by an import?), or any number of other things..
510 assert(session
->is_opening());
511 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
512 mds
->sessionmap
.touch_session(session
);
513 assert(session
->connection
!= NULL
);
514 session
->connection
->send_message(new MClientSession(CEPH_SESSION_OPEN
));
515 if (mdcache
->is_readonly())
516 session
->connection
->send_message(new MClientSession(CEPH_SESSION_FORCE_RO
));
517 } else if (session
->is_closing() ||
518 session
->is_killing()) {
519 // kill any lingering capabilities, leases, requests
520 while (!session
->caps
.empty()) {
521 Capability
*cap
= session
->caps
.front();
522 CInode
*in
= cap
->get_inode();
523 dout(20) << " killing capability " << ccap_string(cap
->issued()) << " on " << *in
<< dendl
;
524 mds
->locker
->remove_client_cap(in
, session
->info
.inst
.name
.num());
526 while (!session
->leases
.empty()) {
527 ClientLease
*r
= session
->leases
.front();
528 CDentry
*dn
= static_cast<CDentry
*>(r
->parent
);
529 dout(20) << " killing client lease of " << *dn
<< dendl
;
530 dn
->remove_client_lease(r
, mds
->locker
);
532 if (client_reconnect_gather
.count(session
->info
.get_client())) {
533 dout(20) << " removing client from reconnect set" << dendl
;
534 client_reconnect_gather
.erase(session
->info
.get_client());
536 if (client_reconnect_gather
.empty()) {
537 dout(7) << " client " << session
->info
.inst
<< " was last reconnect, finishing" << dendl
;
538 reconnect_gather_finish();
542 if (session
->is_closing()) {
543 // mark con disposable. if there is a fault, we will get a
544 // reset and clean it up. if the client hasn't received the
545 // CLOSE message yet, they will reconnect and get an
546 // ms_handle_remote_reset() and realize they had in fact closed.
547 // do this *before* sending the message to avoid a possible
549 if (session
->connection
!= NULL
) {
550 // Conditional because terminate_sessions will indiscrimately
551 // put sessions in CLOSING whether they ever had a conn or not.
552 session
->connection
->mark_disposable();
556 mds
->send_message_client(new MClientSession(CEPH_SESSION_CLOSE
), session
);
557 mds
->sessionmap
.set_state(session
, Session::STATE_CLOSED
);
559 mds
->sessionmap
.remove_session(session
);
560 } else if (session
->is_killing()) {
561 // destroy session, close connection
562 if (session
->connection
!= NULL
) {
563 session
->connection
->mark_down();
565 mds
->sessionmap
.remove_session(session
);
575 * Inject sessions from some source other than actual connections.
578 * - sessions inferred from journal replay
579 * - sessions learned from other MDSs during rejoin
580 * - sessions learned from other MDSs during dir/caps migration
581 * - sessions learned from other MDSs during a cross-MDS rename
583 version_t
Server::prepare_force_open_sessions(map
<client_t
,entity_inst_t
>& cm
,
584 map
<client_t
,uint64_t>& sseqmap
)
586 version_t pv
= mds
->sessionmap
.get_projected();
588 dout(10) << "prepare_force_open_sessions " << pv
589 << " on " << cm
.size() << " clients"
591 for (map
<client_t
,entity_inst_t
>::iterator p
= cm
.begin(); p
!= cm
.end(); ++p
) {
593 Session
*session
= mds
->sessionmap
.get_or_add_session(p
->second
);
594 pv
= mds
->sessionmap
.mark_projected(session
);
595 if (session
->is_closed() ||
596 session
->is_closing() ||
597 session
->is_killing())
598 sseqmap
[p
->first
] = mds
->sessionmap
.set_state(session
, Session::STATE_OPENING
);
600 assert(session
->is_open() ||
601 session
->is_opening() ||
602 session
->is_stale());
603 session
->inc_importing();
608 void Server::finish_force_open_sessions(map
<client_t
,entity_inst_t
>& cm
,
609 map
<client_t
,uint64_t>& sseqmap
,
613 * FIXME: need to carefully consider the race conditions between a
614 * client trying to close a session and an MDS doing an import
615 * trying to force open a session...
617 dout(10) << "finish_force_open_sessions on " << cm
.size() << " clients,"
618 << " initial v " << mds
->sessionmap
.get_version() << dendl
;
621 int sessions_inserted
= 0;
622 for (map
<client_t
,entity_inst_t
>::iterator p
= cm
.begin(); p
!= cm
.end(); ++p
) {
625 Session
*session
= mds
->sessionmap
.get_session(p
->second
.name
);
628 if (sseqmap
.count(p
->first
)) {
629 uint64_t sseq
= sseqmap
[p
->first
];
630 if (session
->get_state_seq() != sseq
) {
631 dout(10) << "force_open_sessions skipping changed " << session
->info
.inst
<< dendl
;
633 dout(10) << "force_open_sessions opened " << session
->info
.inst
<< dendl
;
634 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
635 mds
->sessionmap
.touch_session(session
);
636 mds
->send_message_client(new MClientSession(CEPH_SESSION_OPEN
), session
);
637 if (mdcache
->is_readonly())
638 mds
->send_message_client(new MClientSession(CEPH_SESSION_FORCE_RO
), session
);
641 dout(10) << "force_open_sessions skipping already-open " << session
->info
.inst
<< dendl
;
642 assert(session
->is_open() || session
->is_stale());
646 session
->dec_importing();
649 mds
->sessionmap
.mark_dirty(session
);
652 dout(10) << __func__
<< ": final v " << mds
->sessionmap
.get_version() << dendl
;
655 class C_MDS_TerminatedSessions
: public ServerContext
{
656 void finish(int r
) override
{
657 server
->terminating_sessions
= false;
660 explicit C_MDS_TerminatedSessions(Server
*s
) : ServerContext(s
) {}
663 void Server::terminate_sessions()
665 dout(2) << "terminate_sessions" << dendl
;
667 terminating_sessions
= true;
669 // kill them off. clients will retry etc.
670 set
<Session
*> sessions
;
671 mds
->sessionmap
.get_client_session_set(sessions
);
672 for (set
<Session
*>::const_iterator p
= sessions
.begin();
675 Session
*session
= *p
;
676 if (session
->is_closing() ||
677 session
->is_killing() ||
678 session
->is_closed())
680 journal_close_session(session
, Session::STATE_CLOSING
, NULL
);
683 mdlog
->wait_for_safe(new C_MDS_TerminatedSessions(this));
687 void Server::find_idle_sessions()
689 dout(10) << "find_idle_sessions. laggy until " << mds
->get_laggy_until() << dendl
;
692 // (caps go stale, lease die)
693 utime_t now
= ceph_clock_now();
694 utime_t cutoff
= now
;
695 cutoff
-= g_conf
->mds_session_timeout
;
697 Session
*session
= mds
->sessionmap
.get_oldest_session(Session::STATE_OPEN
);
699 dout(20) << "laggiest active session is " << session
->info
.inst
<< dendl
;
700 if (session
->last_cap_renew
>= cutoff
) {
701 dout(20) << "laggiest active session is " << session
->info
.inst
<< " and sufficiently new ("
702 << session
->last_cap_renew
<< ")" << dendl
;
706 dout(10) << "new stale session " << session
->info
.inst
<< " last " << session
->last_cap_renew
<< dendl
;
707 mds
->sessionmap
.set_state(session
, Session::STATE_STALE
);
708 mds
->locker
->revoke_stale_caps(session
);
709 mds
->locker
->remove_stale_leases(session
);
710 mds
->send_message_client(new MClientSession(CEPH_SESSION_STALE
, session
->get_push_seq()), session
);
711 finish_flush_session(session
, session
->get_push_seq());
716 cutoff
-= g_conf
->mds_session_autoclose
;
718 // don't kick clients if we've been laggy
719 if (mds
->get_laggy_until() > cutoff
) {
720 dout(10) << " laggy_until " << mds
->get_laggy_until() << " > cutoff " << cutoff
721 << ", not kicking any clients to be safe" << dendl
;
725 if (mds
->sessionmap
.get_sessions().size() == 1 &&
726 mds
->mdsmap
->get_num_in_mds() == 1) {
727 dout(20) << "not evicting a slow client, because there is only one"
732 // Collect a list of sessions exceeding the autoclose threshold
733 std::vector
<Session
*> to_evict
;
734 const auto sessions_p
= mds
->sessionmap
.by_state
.find(Session::STATE_STALE
);
735 if (sessions_p
== mds
->sessionmap
.by_state
.end() || sessions_p
->second
->empty()) {
738 const auto &stale_sessions
= sessions_p
->second
;
739 assert(stale_sessions
!= nullptr);
741 for (const auto &session
: *stale_sessions
) {
742 if (session
->is_importing()) {
743 dout(10) << "stopping at importing session " << session
->info
.inst
<< dendl
;
746 assert(session
->is_stale());
747 if (session
->last_cap_renew
>= cutoff
) {
748 dout(20) << "oldest stale session is " << session
->info
.inst
<< " and sufficiently new ("
749 << session
->last_cap_renew
<< ")" << dendl
;
753 to_evict
.push_back(session
);
756 for (const auto &session
: to_evict
) {
758 age
-= session
->last_cap_renew
;
759 mds
->clog
->warn() << "evicting unresponsive client " << *session
760 << ", after " << age
<< " seconds";
761 dout(10) << "autoclosing stale session " << session
->info
.inst
<< " last "
762 << session
->last_cap_renew
<< dendl
;
764 if (g_conf
->mds_session_blacklist_on_timeout
) {
765 std::stringstream ss
;
766 mds
->evict_client(session
->info
.inst
.name
.num(), false, true,
769 kill_session(session
, NULL
);
775 * XXX bump in the interface here, not using an MDSInternalContextBase here
776 * because all the callers right now happen to use a SaferCond
778 void Server::kill_session(Session
*session
, Context
*on_safe
)
780 assert(mds
->mds_lock
.is_locked_by_me());
782 if ((session
->is_opening() ||
783 session
->is_open() ||
784 session
->is_stale()) &&
785 !session
->is_importing()) {
786 dout(10) << "kill_session " << session
<< dendl
;
787 journal_close_session(session
, Session::STATE_KILLING
, on_safe
);
789 dout(10) << "kill_session importing or already closing/killing " << session
<< dendl
;
790 assert(session
->is_closing() ||
791 session
->is_closed() ||
792 session
->is_killing() ||
793 session
->is_importing());
795 on_safe
->complete(0);
800 size_t Server::apply_blacklist(const std::set
<entity_addr_t
> &blacklist
)
802 std::list
<Session
*> victims
;
803 const auto sessions
= mds
->sessionmap
.get_sessions();
804 for (const auto p
: sessions
) {
805 if (!p
.first
.is_client()) {
806 // Do not apply OSDMap blacklist to MDS daemons, we find out
807 // about their death via MDSMap.
811 Session
*s
= p
.second
;
812 if (blacklist
.count(s
->info
.inst
.addr
)) {
813 victims
.push_back(s
);
817 for (const auto s
: victims
) {
818 kill_session(s
, nullptr);
821 dout(10) << "apply_blacklist: killed " << victims
.size() << dendl
;
823 return victims
.size();
826 void Server::journal_close_session(Session
*session
, int state
, Context
*on_safe
)
828 uint64_t sseq
= mds
->sessionmap
.set_state(session
, state
);
829 version_t pv
= mds
->sessionmap
.mark_projected(session
);
832 // release alloc and pending-alloc inos for this session
833 // and wipe out session state, in case the session close aborts for some reason
834 interval_set
<inodeno_t
> both
;
835 both
.insert(session
->info
.prealloc_inos
);
836 both
.insert(session
->pending_prealloc_inos
);
838 mds
->inotable
->project_release_ids(both
);
839 piv
= mds
->inotable
->get_projected_version();
843 mdlog
->start_submit_entry(new ESession(session
->info
.inst
, false, pv
, both
, piv
),
844 new C_MDS_session_finish(this, session
, sseq
, false, pv
, both
, piv
, on_safe
));
847 // clean up requests, too
848 elist
<MDRequestImpl
*>::iterator p
=
849 session
->requests
.begin(member_offset(MDRequestImpl
,
850 item_session_request
));
852 MDRequestRef mdr
= mdcache
->request_get((*p
)->reqid
);
854 mdcache
->request_kill(mdr
);
857 finish_flush_session(session
, session
->get_push_seq());
860 void Server::reconnect_clients(MDSInternalContext
*reconnect_done_
)
862 reconnect_done
= reconnect_done_
;
863 mds
->sessionmap
.get_client_set(client_reconnect_gather
);
865 if (client_reconnect_gather
.empty()) {
866 dout(7) << "reconnect_clients -- no sessions, doing nothing." << dendl
;
867 reconnect_gather_finish();
871 // clients will get the mdsmap and discover we're reconnecting via the monitor.
873 reconnect_start
= ceph_clock_now();
874 dout(1) << "reconnect_clients -- " << client_reconnect_gather
.size() << " sessions" << dendl
;
875 mds
->sessionmap
.dump();
878 /* This function DOES put the passed message before returning*/
879 void Server::handle_client_reconnect(MClientReconnect
*m
)
881 dout(7) << "handle_client_reconnect " << m
->get_source() << dendl
;
882 client_t from
= m
->get_source().num();
883 Session
*session
= get_session(m
);
886 if (!mds
->is_reconnect() && mds
->get_want_state() == CEPH_MDS_STATE_RECONNECT
) {
887 dout(10) << " we're almost in reconnect state (mdsmap delivery race?); waiting" << dendl
;
888 mds
->wait_for_reconnect(new C_MDS_RetryMessage(mds
, m
));
892 utime_t delay
= ceph_clock_now();
893 delay
-= reconnect_start
;
894 dout(10) << " reconnect_start " << reconnect_start
<< " delay " << delay
<< dendl
;
897 if (!mds
->is_reconnect()) {
898 // XXX maybe in the future we can do better than this?
899 dout(1) << " no longer in reconnect state, ignoring reconnect, sending close" << dendl
;
900 mds
->clog
->info() << "denied reconnect attempt (mds is "
901 << ceph_mds_state_name(mds
->get_state())
902 << ") from " << m
->get_source_inst()
903 << " after " << delay
<< " (allowed interval " << g_conf
->mds_reconnect_timeout
<< ")";
905 } else if (session
->is_closed()) {
906 dout(1) << " session is closed, ignoring reconnect, sending close" << dendl
;
907 mds
->clog
->info() << "denied reconnect attempt (mds is "
908 << ceph_mds_state_name(mds
->get_state())
909 << ") from " << m
->get_source_inst() << " (session is closed)";
911 } else if (mdcache
->is_readonly()) {
912 dout(1) << " read-only FS, ignoring reconnect, sending close" << dendl
;
913 mds
->clog
->info() << "denied reconnect attempt (mds is read-only)";
918 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_CLOSE
));
923 // notify client of success with an OPEN
924 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_OPEN
));
925 session
->last_cap_renew
= ceph_clock_now();
926 mds
->clog
->debug() << "reconnect by " << session
->info
.inst
<< " after " << delay
;
929 for (vector
<ceph_mds_snaprealm_reconnect
>::iterator p
= m
->realms
.begin();
930 p
!= m
->realms
.end();
932 CInode
*in
= mdcache
->get_inode(inodeno_t(p
->ino
));
933 if (in
&& in
->state_test(CInode::STATE_PURGING
))
936 assert(in
->snaprealm
);
937 if (in
->snaprealm
->have_past_parents_open()) {
938 dout(15) << "open snaprealm (w/ past parents) on " << *in
<< dendl
;
939 mdcache
->finish_snaprealm_reconnect(from
, in
->snaprealm
, snapid_t(p
->seq
));
941 dout(15) << "open snaprealm (w/o past parents) on " << *in
<< dendl
;
942 mdcache
->add_reconnected_snaprealm(from
, inodeno_t(p
->ino
), snapid_t(p
->seq
));
945 dout(15) << "open snaprealm (w/o inode) on " << inodeno_t(p
->ino
)
946 << " seq " << p
->seq
<< dendl
;
947 mdcache
->add_reconnected_snaprealm(from
, inodeno_t(p
->ino
), snapid_t(p
->seq
));
952 for (map
<inodeno_t
, cap_reconnect_t
>::iterator p
= m
->caps
.begin();
955 // make sure our last_cap_id is MAX over all issued caps
956 if (p
->second
.capinfo
.cap_id
> mdcache
->last_cap_id
)
957 mdcache
->last_cap_id
= p
->second
.capinfo
.cap_id
;
959 CInode
*in
= mdcache
->get_inode(p
->first
);
960 if (in
&& in
->state_test(CInode::STATE_PURGING
))
962 if (in
&& in
->is_auth()) {
963 // we recovered it, and it's ours. take note.
964 dout(15) << "open cap realm " << inodeno_t(p
->second
.capinfo
.snaprealm
)
965 << " on " << *in
<< dendl
;
966 in
->reconnect_cap(from
, p
->second
, session
);
967 mdcache
->add_reconnected_cap(from
, p
->first
, p
->second
);
968 recover_filelocks(in
, p
->second
.flockbl
, m
->get_orig_source().num());
972 if (in
&& !in
->is_auth()) {
974 dout(10) << "non-auth " << *in
<< ", will pass off to authority" << dendl
;
975 // add to cap export list.
976 p
->second
.path
.clear(); // we don't need path
977 mdcache
->rejoin_export_caps(p
->first
, from
, p
->second
,
978 in
->authority().first
);
980 // don't know if the inode is mine
981 dout(10) << "missing ino " << p
->first
<< ", will load later" << dendl
;
982 p
->second
.path
.clear(); // we don't need path
983 mdcache
->rejoin_recovered_caps(p
->first
, from
, p
->second
, MDS_RANK_NONE
);
987 // remove from gather set
988 client_reconnect_gather
.erase(from
);
989 if (client_reconnect_gather
.empty())
990 reconnect_gather_finish();
997 void Server::reconnect_gather_finish()
999 dout(7) << "reconnect_gather_finish. failed on " << failed_reconnects
<< " clients" << dendl
;
1000 assert(reconnect_done
);
1001 reconnect_done
->complete(0);
1002 reconnect_done
= NULL
;
1005 void Server::reconnect_tick()
1007 if (reconnect_evicting
) {
1008 dout(4) << "reconnect_tick: waiting for evictions" << dendl
;
1012 utime_t reconnect_end
= reconnect_start
;
1013 reconnect_end
+= g_conf
->mds_reconnect_timeout
;
1014 if (ceph_clock_now() >= reconnect_end
&&
1015 !client_reconnect_gather
.empty()) {
1016 dout(10) << "reconnect timed out" << dendl
;
1018 // If we're doing blacklist evictions, use this to wait for them before
1019 // proceeding to reconnect_gather_finish
1020 MDSGatherBuilder
gather(g_ceph_context
);
1022 for (set
<client_t
>::iterator p
= client_reconnect_gather
.begin();
1023 p
!= client_reconnect_gather
.end();
1025 Session
*session
= mds
->sessionmap
.get_session(entity_name_t::CLIENT(p
->v
));
1027 dout(1) << "reconnect gave up on " << session
->info
.inst
<< dendl
;
1029 mds
->clog
->warn() << "evicting unresponsive client " << *session
1030 << ", after waiting " << g_conf
->mds_reconnect_timeout
1031 << " seconds during MDS startup";
1033 if (g_conf
->mds_session_blacklist_on_timeout
) {
1034 std::stringstream ss
;
1035 mds
->evict_client(session
->info
.inst
.name
.num(), false, true, ss
,
1038 kill_session(session
, NULL
);
1041 failed_reconnects
++;
1043 client_reconnect_gather
.clear();
1045 if (gather
.has_subs()) {
1046 dout(1) << "reconnect will complete once clients are evicted" << dendl
;
1047 gather
.set_finisher(new MDSInternalContextWrapper(mds
, new FunctionContext(
1048 [this](int r
){reconnect_gather_finish();})));
1050 reconnect_evicting
= true;
1052 reconnect_gather_finish();
1057 void Server::recover_filelocks(CInode
*in
, bufferlist locks
, int64_t client
)
1059 if (!locks
.length()) return;
1062 bufferlist::iterator p
= locks
.begin();
1063 ::decode(numlocks
, p
);
1064 for (int i
= 0; i
< numlocks
; ++i
) {
1066 lock
.client
= client
;
1067 in
->get_fcntl_lock_state()->held_locks
.insert(pair
<uint64_t, ceph_filelock
>(lock
.start
, lock
));
1068 ++in
->get_fcntl_lock_state()->client_held_lock_counts
[client
];
1070 ::decode(numlocks
, p
);
1071 for (int i
= 0; i
< numlocks
; ++i
) {
1073 lock
.client
= client
;
1074 in
->get_flock_lock_state()->held_locks
.insert(pair
<uint64_t, ceph_filelock
> (lock
.start
, lock
));
1075 ++in
->get_flock_lock_state()->client_held_lock_counts
[client
];
1081 * Call this when the MDCache is oversized, to send requests to the clients
1082 * to trim some caps, and consequently unpin some inodes in the MDCache so
1083 * that it can trim too.
1085 void Server::recall_client_state(float ratio
)
1087 int max_caps_per_client
= (int)(g_conf
->mds_cache_size
* .8);
1088 int min_caps_per_client
= 100;
1090 dout(10) << "recall_client_state " << ratio
1091 << ", caps per client " << min_caps_per_client
<< "-" << max_caps_per_client
1094 set
<Session
*> sessions
;
1095 mds
->sessionmap
.get_client_session_set(sessions
);
1096 for (set
<Session
*>::const_iterator p
= sessions
.begin();
1097 p
!= sessions
.end();
1099 Session
*session
= *p
;
1100 if (!session
->is_open() ||
1101 !session
->info
.inst
.name
.is_client())
1104 dout(10) << " session " << session
->info
.inst
1105 << " caps " << session
->caps
.size()
1106 << ", leases " << session
->leases
.size()
1109 if (session
->caps
.size() > min_caps_per_client
) {
1110 int newlim
= MIN((int)(session
->caps
.size() * ratio
), max_caps_per_client
);
1111 if (session
->caps
.size() > newlim
) {
1112 MClientSession
*m
= new MClientSession(CEPH_SESSION_RECALL_STATE
);
1113 m
->head
.max_caps
= newlim
;
1114 mds
->send_message_client(m
, session
);
1115 session
->notify_recall_sent(newlim
);
1121 void Server::force_clients_readonly()
1123 dout(10) << "force_clients_readonly" << dendl
;
1124 set
<Session
*> sessions
;
1125 mds
->sessionmap
.get_client_session_set(sessions
);
1126 for (set
<Session
*>::const_iterator p
= sessions
.begin();
1127 p
!= sessions
.end();
1129 Session
*session
= *p
;
1130 if (!session
->info
.inst
.name
.is_client() ||
1131 !(session
->is_open() || session
->is_stale()))
1133 mds
->send_message_client(new MClientSession(CEPH_SESSION_FORCE_RO
), session
);
1138 * some generic stuff for finishing off requests
1140 void Server::journal_and_reply(MDRequestRef
& mdr
, CInode
*in
, CDentry
*dn
, LogEvent
*le
, MDSLogContextBase
*fin
)
1142 dout(10) << "journal_and_reply tracei " << in
<< " tracedn " << dn
<< dendl
;
1143 assert(!mdr
->has_completed
);
1145 // note trace items for eventual reply.
1154 early_reply(mdr
, in
, dn
);
1156 mdr
->committing
= true;
1157 submit_mdlog_entry(le
, fin
, mdr
, __func__
);
1159 if (mdr
->client_request
&& mdr
->client_request
->is_queued_for_replay()) {
1160 if (mds
->queue_one_replay()) {
1161 dout(10) << " queued next replay op" << dendl
;
1163 dout(10) << " journaled last replay op, flushing" << dendl
;
1166 } else if (mdr
->did_early_reply
)
1167 mds
->locker
->drop_rdlocks(mdr
.get());
1172 void Server::submit_mdlog_entry(LogEvent
*le
, MDSLogContextBase
*fin
, MDRequestRef
& mdr
,
1176 string
event_str("submit entry: ");
1178 mdr
->mark_event_string(event_str
);
1180 mdlog
->submit_entry(le
, fin
);
1184 * send response built from mdr contents and error code; clean up mdr
1186 void Server::respond_to_request(MDRequestRef
& mdr
, int r
)
1188 if (mdr
->client_request
) {
1189 reply_client_request(mdr
, new MClientReply(mdr
->client_request
, r
));
1191 // add here to avoid counting ops multiple times (e.g., locks, loading)
1192 switch(mdr
->client_request
->get_op()) {
1193 case CEPH_MDS_OP_LOOKUPHASH
:
1194 logger
->inc(l_mdss_req_lookuphash
);
1196 case CEPH_MDS_OP_LOOKUPINO
:
1197 logger
->inc(l_mdss_req_lookupino
);
1199 case CEPH_MDS_OP_LOOKUPPARENT
:
1200 logger
->inc(l_mdss_req_lookupparent
);
1202 case CEPH_MDS_OP_LOOKUPNAME
:
1203 logger
->inc(l_mdss_req_lookupname
);
1205 case CEPH_MDS_OP_LOOKUP
:
1206 logger
->inc(l_mdss_req_lookup
);
1208 case CEPH_MDS_OP_LOOKUPSNAP
:
1209 logger
->inc(l_mdss_req_lookupsnap
);
1211 case CEPH_MDS_OP_GETATTR
:
1212 logger
->inc(l_mdss_req_getattr
);
1214 case CEPH_MDS_OP_SETATTR
:
1215 logger
->inc(l_mdss_req_setattr
);
1217 case CEPH_MDS_OP_SETLAYOUT
:
1218 logger
->inc(l_mdss_req_setlayout
);
1220 case CEPH_MDS_OP_SETDIRLAYOUT
:
1221 logger
->inc(l_mdss_req_setdirlayout
);
1223 case CEPH_MDS_OP_SETXATTR
:
1224 logger
->inc(l_mdss_req_setxattr
);
1226 case CEPH_MDS_OP_RMXATTR
:
1227 logger
->inc(l_mdss_req_rmxattr
);
1229 case CEPH_MDS_OP_READDIR
:
1230 logger
->inc(l_mdss_req_readdir
);
1232 case CEPH_MDS_OP_SETFILELOCK
:
1233 logger
->inc(l_mdss_req_setfilelock
);
1235 case CEPH_MDS_OP_GETFILELOCK
:
1236 logger
->inc(l_mdss_req_getfilelock
);
1238 case CEPH_MDS_OP_CREATE
:
1239 logger
->inc(l_mdss_req_create
);
1240 case CEPH_MDS_OP_OPEN
:
1241 logger
->inc(l_mdss_req_open
);
1243 case CEPH_MDS_OP_MKNOD
:
1244 logger
->inc(l_mdss_req_mknod
);
1246 case CEPH_MDS_OP_LINK
:
1247 logger
->inc(l_mdss_req_link
);
1249 case CEPH_MDS_OP_UNLINK
:
1250 logger
->inc(l_mdss_req_unlink
);
1252 case CEPH_MDS_OP_RMDIR
:
1253 logger
->inc(l_mdss_req_rmdir
);
1255 case CEPH_MDS_OP_RENAME
:
1256 logger
->inc(l_mdss_req_rename
);
1258 case CEPH_MDS_OP_MKDIR
:
1259 logger
->inc(l_mdss_req_mkdir
);
1261 case CEPH_MDS_OP_SYMLINK
:
1262 logger
->inc(l_mdss_req_symlink
);
1264 case CEPH_MDS_OP_LSSNAP
:
1265 logger
->inc(l_mdss_req_lssnap
);
1267 case CEPH_MDS_OP_MKSNAP
:
1268 logger
->inc(l_mdss_req_mksnap
);
1270 case CEPH_MDS_OP_RMSNAP
:
1271 logger
->inc(l_mdss_req_rmsnap
);
1273 case CEPH_MDS_OP_RENAMESNAP
:
1274 logger
->inc(l_mdss_req_renamesnap
);
1277 } else if (mdr
->internal_op
> -1) {
1278 dout(10) << "respond_to_request on internal request " << mdr
<< dendl
;
1279 if (!mdr
->internal_op_finish
)
1280 assert(0 == "trying to respond to internal op without finisher");
1281 mdr
->internal_op_finish
->complete(r
);
1282 mdcache
->request_finish(mdr
);
1286 void Server::early_reply(MDRequestRef
& mdr
, CInode
*tracei
, CDentry
*tracedn
)
1288 if (!g_conf
->mds_early_reply
)
1291 if (mdr
->has_more() && mdr
->more()->has_journaled_slaves
) {
1292 dout(10) << "early_reply - there are journaled slaves, not allowed." << dendl
;
1296 if (mdr
->alloc_ino
) {
1297 dout(10) << "early_reply - allocated ino, not allowed" << dendl
;
1301 MClientRequest
*req
= mdr
->client_request
;
1302 entity_inst_t client_inst
= req
->get_source_inst();
1303 if (client_inst
.name
.is_mds())
1306 if (req
->is_replay()) {
1307 dout(10) << " no early reply on replay op" << dendl
;
1312 MClientReply
*reply
= new MClientReply(req
, 0);
1313 reply
->set_unsafe();
1315 // mark xlocks "done", indicating that we are exposing uncommitted changes.
1317 //_rename_finish() does not send dentry link/unlink message to replicas.
1318 // so do not set xlocks on dentries "done", the xlocks prevent dentries
1319 // that have projected linkages from getting new replica.
1320 mds
->locker
->set_xlocks_done(mdr
.get(), req
->get_op() == CEPH_MDS_OP_RENAME
);
1322 dout(10) << "early_reply " << reply
->get_result()
1323 << " (" << cpp_strerror(reply
->get_result())
1324 << ") " << *req
<< dendl
;
1326 if (tracei
|| tracedn
) {
1328 mdr
->cap_releases
.erase(tracei
->vino());
1330 mdr
->cap_releases
.erase(tracedn
->get_dir()->get_inode()->vino());
1332 set_trace_dist(mdr
->session
, reply
, tracei
, tracedn
, mdr
->snapid
,
1333 req
->get_dentry_wanted(), mdr
);
1336 reply
->set_extra_bl(mdr
->reply_extra_bl
);
1337 req
->get_connection()->send_message(reply
);
1339 mdr
->did_early_reply
= true;
1341 mds
->logger
->inc(l_mds_reply
);
1342 utime_t lat
= ceph_clock_now() - req
->get_recv_stamp();
1343 mds
->logger
->tinc(l_mds_reply_latency
, lat
);
1344 dout(20) << "lat " << lat
<< dendl
;
1346 mdr
->mark_event("early_replied");
1351 * include a trace to tracei
1354 void Server::reply_client_request(MDRequestRef
& mdr
, MClientReply
*reply
)
1357 MClientRequest
*req
= mdr
->client_request
;
1359 dout(7) << "reply_client_request " << reply
->get_result()
1360 << " (" << cpp_strerror(reply
->get_result())
1361 << ") " << *req
<< dendl
;
1363 mdr
->mark_event("replying");
1365 Session
*session
= mdr
->session
;
1367 // note successful request in session map?
1369 // setfilelock requests are special, they only modify states in MDS memory.
1370 // The states get lost when MDS fails. If Client re-send a completed
1371 // setfilelock request, it means that client did not receive corresponding
1372 // setfilelock reply. So MDS should re-execute the setfilelock request.
1373 if (req
->may_write() && req
->get_op() != CEPH_MDS_OP_SETFILELOCK
&&
1374 reply
->get_result() == 0 && session
) {
1375 inodeno_t created
= mdr
->alloc_ino
? mdr
->alloc_ino
: mdr
->used_prealloc_ino
;
1376 session
->add_completed_request(mdr
->reqid
.tid
, created
);
1378 mdr
->ls
->touched_sessions
.insert(session
->info
.inst
.name
);
1382 // give any preallocated inos to the session
1383 apply_allocated_inos(mdr
, session
);
1385 // get tracei/tracedn from mdr?
1386 snapid_t snapid
= mdr
->snapid
;
1387 CInode
*tracei
= mdr
->tracei
;
1388 CDentry
*tracedn
= mdr
->tracedn
;
1390 bool is_replay
= mdr
->client_request
->is_replay();
1391 bool did_early_reply
= mdr
->did_early_reply
;
1392 entity_inst_t client_inst
= req
->get_source_inst();
1393 int dentry_wanted
= req
->get_dentry_wanted();
1395 if (!did_early_reply
&& !is_replay
) {
1397 mds
->logger
->inc(l_mds_reply
);
1398 utime_t lat
= ceph_clock_now() - mdr
->client_request
->get_recv_stamp();
1399 mds
->logger
->tinc(l_mds_reply_latency
, lat
);
1400 dout(20) << "lat " << lat
<< dendl
;
1403 mdr
->cap_releases
.erase(tracei
->vino());
1405 mdr
->cap_releases
.erase(tracedn
->get_dir()->get_inode()->vino());
1408 // drop non-rdlocks before replying, so that we can issue leases
1409 mdcache
->request_drop_non_rdlocks(mdr
);
1412 if (client_inst
.name
.is_mds() || !session
) {
1413 reply
->put(); // mds doesn't need a reply
1417 if (!did_early_reply
&& // don't issue leases if we sent an earlier reply already
1418 (tracei
|| tracedn
)) {
1421 mdcache
->try_reconnect_cap(tracei
, session
);
1423 // include metadata in reply
1424 set_trace_dist(session
, reply
, tracei
, tracedn
,
1425 snapid
, dentry_wanted
,
1430 // We can set the extra bl unconditionally: if it's already been sent in the
1431 // early_reply, set_extra_bl will have claimed it and reply_extra_bl is empty
1432 reply
->set_extra_bl(mdr
->reply_extra_bl
);
1434 reply
->set_mdsmap_epoch(mds
->mdsmap
->get_epoch());
1435 req
->get_connection()->send_message(reply
);
1438 if (req
->is_queued_for_replay() &&
1439 (mdr
->has_completed
|| reply
->get_result() < 0)) {
1440 if (reply
->get_result() < 0) {
1441 int r
= reply
->get_result();
1442 derr
<< "reply_client_request: failed to replay " << *req
1443 << " error " << r
<< " (" << cpp_strerror(r
) << ")" << dendl
;
1444 mds
->clog
->warn() << "failed to replay " << req
->get_reqid() << " error " << r
;
1446 mds
->queue_one_replay();
1450 mdcache
->request_finish(mdr
);
1452 // take a closer look at tracei, if it happens to be a remote link
1455 tracedn
->get_projected_linkage()->is_remote()) {
1456 mdcache
->eval_remote(tracedn
);
1461 void Server::encode_empty_dirstat(bufferlist
& bl
)
1463 static DirStat empty
;
1467 void Server::encode_infinite_lease(bufferlist
& bl
)
1474 dout(20) << "encode_infinite_lease " << e
<< dendl
;
1477 void Server::encode_null_lease(bufferlist
& bl
)
1484 dout(20) << "encode_null_lease " << e
<< dendl
;
1489 * pass inode OR dentry (not both, or we may get confused)
1491 * trace is in reverse order (i.e. root inode comes last)
1493 void Server::set_trace_dist(Session
*session
, MClientReply
*reply
,
1494 CInode
*in
, CDentry
*dn
,
1499 // skip doing this for debugging purposes?
1500 if (g_conf
->mds_inject_traceless_reply_probability
&&
1501 mdr
->ls
&& !mdr
->o_trunc
&&
1502 (rand() % 10000 < g_conf
->mds_inject_traceless_reply_probability
* 10000.0)) {
1503 dout(5) << "deliberately skipping trace for " << *reply
<< dendl
;
1507 // inode, dentry, dir, ..., inode
1509 mds_rank_t whoami
= mds
->get_nodeid();
1510 client_t client
= session
->get_client();
1511 utime_t now
= ceph_clock_now();
1513 dout(20) << "set_trace_dist snapid " << snapid
<< dendl
;
1515 //assert((bool)dn == (bool)dentry_wanted); // not true for snapshot lookups
1518 if (snapid
== CEPH_NOSNAP
) {
1521 realm
= in
->find_snaprealm();
1523 realm
= dn
->get_dir()->get_inode()->find_snaprealm();
1524 reply
->snapbl
= realm
->get_snap_trace();
1525 dout(10) << "set_trace_dist snaprealm " << *realm
<< " len=" << reply
->snapbl
.length() << dendl
;
1530 reply
->head
.is_dentry
= 1;
1531 CDir
*dir
= dn
->get_dir();
1532 CInode
*diri
= dir
->get_inode();
1534 diri
->encode_inodestat(bl
, session
, NULL
, snapid
);
1535 dout(20) << "set_trace_dist added diri " << *diri
<< dendl
;
1537 #ifdef MDS_VERIFY_FRAGSTAT
1538 if (dir
->is_complete())
1539 dir
->verify_fragstat();
1541 dir
->encode_dirstat(bl
, whoami
);
1542 dout(20) << "set_trace_dist added dir " << *dir
<< dendl
;
1544 ::encode(dn
->get_name(), bl
);
1545 if (snapid
== CEPH_NOSNAP
)
1546 mds
->locker
->issue_client_lease(dn
, client
, bl
, now
, session
);
1548 encode_null_lease(bl
);
1549 dout(20) << "set_trace_dist added dn " << snapid
<< " " << *dn
<< dendl
;
1551 reply
->head
.is_dentry
= 0;
1555 in
->encode_inodestat(bl
, session
, NULL
, snapid
, 0, mdr
->getattr_caps
);
1556 dout(20) << "set_trace_dist added in " << *in
<< dendl
;
1557 reply
->head
.is_target
= 1;
1559 reply
->head
.is_target
= 0;
1561 reply
->set_trace(bl
);
1568 * process a client request
1569 * This function DOES put the passed message before returning
1571 void Server::handle_client_request(MClientRequest
*req
)
1573 dout(4) << "handle_client_request " << *req
<< dendl
;
1576 mds
->logger
->inc(l_mds_request
);
1578 logger
->inc(l_mdss_handle_client_request
);
1580 if (!mdcache
->is_open()) {
1581 dout(5) << "waiting for root" << dendl
;
1582 mdcache
->wait_for_open(new C_MDS_RetryMessage(mds
, req
));
1587 Session
*session
= 0;
1588 if (req
->get_source().is_client()) {
1589 session
= get_session(req
);
1591 dout(5) << "no session for " << req
->get_source() << ", dropping" << dendl
;
1592 } else if (session
->is_closed() ||
1593 session
->is_closing() ||
1594 session
->is_killing()) {
1595 dout(5) << "session closed|closing|killing, dropping" << dendl
;
1599 if (req
->is_queued_for_replay())
1600 mds
->queue_one_replay();
1607 if (req
->get_mdsmap_epoch() < mds
->mdsmap
->get_epoch()) {
1608 // send it? hrm, this isn't ideal; they may get a lot of copies if
1609 // they have a high request rate.
1612 // completed request?
1613 bool has_completed
= false;
1614 if (req
->is_replay() || req
->get_retry_attempt()) {
1617 if (session
->have_completed_request(req
->get_reqid().tid
, &created
)) {
1618 has_completed
= true;
1619 // Don't send traceless reply if the completed request has created
1620 // new inode. Treat the request as lookup request instead.
1621 if (req
->is_replay() ||
1622 ((created
== inodeno_t() || !mds
->is_clientreplay()) &&
1623 req
->get_op() != CEPH_MDS_OP_OPEN
&&
1624 req
->get_op() != CEPH_MDS_OP_CREATE
)) {
1625 dout(5) << "already completed " << req
->get_reqid() << dendl
;
1626 MClientReply
*reply
= new MClientReply(req
, 0);
1627 if (created
!= inodeno_t()) {
1629 ::encode(created
, extra
);
1630 reply
->set_extra_bl(extra
);
1632 req
->get_connection()->send_message(reply
);
1634 if (req
->is_queued_for_replay())
1635 mds
->queue_one_replay();
1640 if (req
->get_op() != CEPH_MDS_OP_OPEN
&&
1641 req
->get_op() != CEPH_MDS_OP_CREATE
) {
1642 dout(10) << " completed request which created new inode " << created
1643 << ", convert it to lookup request" << dendl
;
1644 req
->head
.op
= req
->get_dentry_wanted() ? CEPH_MDS_OP_LOOKUP
: CEPH_MDS_OP_GETATTR
;
1645 req
->head
.args
.getattr
.mask
= CEPH_STAT_CAP_INODE_ALL
;
1650 // trim completed_request list
1651 if (req
->get_oldest_client_tid() > 0) {
1652 dout(15) << " oldest_client_tid=" << req
->get_oldest_client_tid() << dendl
;
1654 if (session
->trim_completed_requests(req
->get_oldest_client_tid())) {
1655 // Sessions 'completed_requests' was dirtied, mark it to be
1656 // potentially flushed at segment expiry.
1657 mdlog
->get_current_segment()->touched_sessions
.insert(session
->info
.inst
.name
);
1659 if (session
->get_num_trim_requests_warnings() > 0 &&
1660 session
->get_num_completed_requests() * 2 < g_conf
->mds_max_completed_requests
)
1661 session
->reset_num_trim_requests_warnings();
1663 if (session
->get_num_completed_requests() >=
1664 (g_conf
->mds_max_completed_requests
<< session
->get_num_trim_requests_warnings())) {
1665 session
->inc_num_trim_requests_warnings();
1667 ss
<< "client." << session
->get_client() << " does not advance its oldest_client_tid ("
1668 << req
->get_oldest_client_tid() << "), "
1669 << session
->get_num_completed_requests()
1670 << " completed requests recorded in session\n";
1671 mds
->clog
->warn() << ss
.str();
1672 dout(20) << __func__
<< " " << ss
.str() << dendl
;
1677 // register + dispatch
1678 MDRequestRef mdr
= mdcache
->request_start(req
);
1683 mdr
->session
= session
;
1684 session
->requests
.push_back(&mdr
->item_session_request
);
1688 mdr
->has_completed
= true;
1690 // process embedded cap releases?
1691 // (only if NOT replay!)
1692 if (!req
->releases
.empty() && req
->get_source().is_client() && !req
->is_replay()) {
1693 client_t client
= req
->get_source().num();
1694 for (vector
<MClientRequest::Release
>::iterator p
= req
->releases
.begin();
1695 p
!= req
->releases
.end();
1697 mds
->locker
->process_request_cap_release(mdr
, client
, p
->item
, p
->dname
);
1698 req
->releases
.clear();
1701 dispatch_client_request(mdr
);
1705 void Server::handle_osd_map()
1707 /* Note that we check the OSDMAP_FULL flag directly rather than
1708 * using osdmap_full_flag(), because we want to know "is the flag set"
1709 * rather than "does the flag apply to us?" */
1710 mds
->objecter
->with_osdmap([this](const OSDMap
& o
) {
1711 is_full
= o
.test_flag(CEPH_OSDMAP_FULL
);
1712 dout(7) << __func__
<< ": full = " << is_full
<< " epoch = "
1713 << o
.get_epoch() << dendl
;
1717 void Server::dispatch_client_request(MDRequestRef
& mdr
)
1719 // we shouldn't be waiting on anyone.
1720 assert(!mdr
->has_more() || mdr
->more()->waiting_on_slave
.empty());
1723 dout(10) << "request " << *mdr
<< " was killed" << dendl
;
1727 MClientRequest
*req
= mdr
->client_request
;
1729 if (logger
) logger
->inc(l_mdss_dispatch_client_request
);
1731 dout(7) << "dispatch_client_request " << *req
<< dendl
;
1733 if (req
->may_write()) {
1734 if (mdcache
->is_readonly()) {
1735 dout(10) << " read-only FS" << dendl
;
1736 respond_to_request(mdr
, -EROFS
);
1739 if (mdr
->has_more() && mdr
->more()->slave_error
) {
1740 dout(10) << " got error from slaves" << dendl
;
1741 respond_to_request(mdr
, mdr
->more()->slave_error
);
1747 if (req
->get_op() == CEPH_MDS_OP_SETLAYOUT
||
1748 req
->get_op() == CEPH_MDS_OP_SETDIRLAYOUT
||
1749 req
->get_op() == CEPH_MDS_OP_SETLAYOUT
||
1750 req
->get_op() == CEPH_MDS_OP_RMXATTR
||
1751 req
->get_op() == CEPH_MDS_OP_SETXATTR
||
1752 req
->get_op() == CEPH_MDS_OP_CREATE
||
1753 req
->get_op() == CEPH_MDS_OP_SYMLINK
||
1754 req
->get_op() == CEPH_MDS_OP_MKSNAP
||
1755 ((req
->get_op() == CEPH_MDS_OP_LINK
||
1756 req
->get_op() == CEPH_MDS_OP_RENAME
) &&
1757 (!mdr
->has_more() || mdr
->more()->witnessed
.empty())) // haven't started slave request
1760 dout(20) << __func__
<< ": full, responding ENOSPC to op " << ceph_mds_op_name(req
->get_op()) << dendl
;
1761 respond_to_request(mdr
, -ENOSPC
);
1764 dout(20) << __func__
<< ": full, permitting op " << ceph_mds_op_name(req
->get_op()) << dendl
;
1768 switch (req
->get_op()) {
1769 case CEPH_MDS_OP_LOOKUPHASH
:
1770 case CEPH_MDS_OP_LOOKUPINO
:
1771 handle_client_lookup_ino(mdr
, false, false);
1773 case CEPH_MDS_OP_LOOKUPPARENT
:
1774 handle_client_lookup_ino(mdr
, true, false);
1776 case CEPH_MDS_OP_LOOKUPNAME
:
1777 handle_client_lookup_ino(mdr
, false, true);
1781 case CEPH_MDS_OP_LOOKUP
:
1782 handle_client_getattr(mdr
, true);
1785 case CEPH_MDS_OP_LOOKUPSNAP
:
1786 // lookupsnap does not reference a CDentry; treat it as a getattr
1787 case CEPH_MDS_OP_GETATTR
:
1788 handle_client_getattr(mdr
, false);
1791 case CEPH_MDS_OP_SETATTR
:
1792 handle_client_setattr(mdr
);
1794 case CEPH_MDS_OP_SETLAYOUT
:
1795 handle_client_setlayout(mdr
);
1797 case CEPH_MDS_OP_SETDIRLAYOUT
:
1798 handle_client_setdirlayout(mdr
);
1800 case CEPH_MDS_OP_SETXATTR
:
1801 handle_client_setxattr(mdr
);
1803 case CEPH_MDS_OP_RMXATTR
:
1804 handle_client_removexattr(mdr
);
1807 case CEPH_MDS_OP_READDIR
:
1808 handle_client_readdir(mdr
);
1811 case CEPH_MDS_OP_SETFILELOCK
:
1812 handle_client_file_setlock(mdr
);
1815 case CEPH_MDS_OP_GETFILELOCK
:
1816 handle_client_file_readlock(mdr
);
1820 case CEPH_MDS_OP_CREATE
:
1821 if (mdr
->has_completed
)
1822 handle_client_open(mdr
); // already created.. just open
1824 handle_client_openc(mdr
);
1827 case CEPH_MDS_OP_OPEN
:
1828 handle_client_open(mdr
);
1833 case CEPH_MDS_OP_MKNOD
:
1834 handle_client_mknod(mdr
);
1836 case CEPH_MDS_OP_LINK
:
1837 handle_client_link(mdr
);
1839 case CEPH_MDS_OP_UNLINK
:
1840 case CEPH_MDS_OP_RMDIR
:
1841 handle_client_unlink(mdr
);
1843 case CEPH_MDS_OP_RENAME
:
1844 handle_client_rename(mdr
);
1846 case CEPH_MDS_OP_MKDIR
:
1847 handle_client_mkdir(mdr
);
1849 case CEPH_MDS_OP_SYMLINK
:
1850 handle_client_symlink(mdr
);
1855 case CEPH_MDS_OP_LSSNAP
:
1856 handle_client_lssnap(mdr
);
1858 case CEPH_MDS_OP_MKSNAP
:
1859 handle_client_mksnap(mdr
);
1861 case CEPH_MDS_OP_RMSNAP
:
1862 handle_client_rmsnap(mdr
);
1864 case CEPH_MDS_OP_RENAMESNAP
:
1865 handle_client_renamesnap(mdr
);
1869 dout(1) << " unknown client op " << req
->get_op() << dendl
;
1870 respond_to_request(mdr
, -EOPNOTSUPP
);
1875 // ---------------------------------------
1878 /* This function DOES put the passed message before returning*/
1879 void Server::handle_slave_request(MMDSSlaveRequest
*m
)
1881 dout(4) << "handle_slave_request " << m
->get_reqid() << " from " << m
->get_source() << dendl
;
1882 mds_rank_t from
= mds_rank_t(m
->get_source().num());
1884 if (logger
) logger
->inc(l_mdss_handle_slave_request
);
1888 return handle_slave_request_reply(m
);
1890 // the purpose of rename notify is enforcing causal message ordering. making sure
1891 // bystanders have received all messages from rename srcdn's auth MDS.
1892 if (m
->get_op() == MMDSSlaveRequest::OP_RENAMENOTIFY
) {
1893 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(m
->get_reqid(), m
->get_attempt(),
1894 MMDSSlaveRequest::OP_RENAMENOTIFYACK
);
1895 mds
->send_message(reply
, m
->get_connection());
1900 CDentry
*straydn
= NULL
;
1901 if (m
->stray
.length() > 0) {
1902 straydn
= mdcache
->add_replica_stray(m
->stray
, from
);
1907 // am i a new slave?
1909 if (mdcache
->have_request(m
->get_reqid())) {
1911 mdr
= mdcache
->request_get(m
->get_reqid());
1913 // is my request newer?
1914 if (mdr
->attempt
> m
->get_attempt()) {
1915 dout(10) << "local request " << *mdr
<< " attempt " << mdr
->attempt
<< " > " << m
->get_attempt()
1916 << ", dropping " << *m
<< dendl
;
1922 if (mdr
->attempt
< m
->get_attempt()) {
1923 // mine is old, close it out
1924 dout(10) << "local request " << *mdr
<< " attempt " << mdr
->attempt
<< " < " << m
->get_attempt()
1925 << ", closing out" << dendl
;
1926 mdcache
->request_finish(mdr
);
1928 } else if (mdr
->slave_to_mds
!= from
) {
1929 dout(10) << "local request " << *mdr
<< " not slave to mds." << from
<< dendl
;
1934 if (m
->get_op() == MMDSSlaveRequest::OP_FINISH
&& m
->is_abort()) {
1935 mdr
->aborted
= true;
1936 if (mdr
->slave_request
) {
1937 // only abort on-going xlock, wrlock and auth pin
1938 assert(!mdr
->slave_did_prepare());
1940 mdcache
->request_finish(mdr
);
1947 if (m
->get_op() == MMDSSlaveRequest::OP_FINISH
) {
1948 dout(10) << "missing slave request for " << m
->get_reqid()
1949 << " OP_FINISH, must have lost race with a forward" << dendl
;
1953 mdr
= mdcache
->request_start_slave(m
->get_reqid(), m
->get_attempt(), m
);
1954 mdr
->set_op_stamp(m
->op_stamp
);
1956 assert(mdr
->slave_request
== 0); // only one at a time, please!
1960 mdr
->straydn
= straydn
;
1963 if (!mds
->is_clientreplay() && !mds
->is_active() && !mds
->is_stopping()) {
1964 dout(3) << "not clientreplay|active yet, waiting" << dendl
;
1965 mds
->wait_for_replay(new C_MDS_RetryMessage(mds
, m
));
1967 } else if (mds
->is_clientreplay() && !mds
->mdsmap
->is_clientreplay(from
) &&
1968 mdr
->locks
.empty()) {
1969 dout(3) << "not active yet, waiting" << dendl
;
1970 mds
->wait_for_active(new C_MDS_RetryMessage(mds
, m
));
1974 mdr
->slave_request
= m
;
1976 dispatch_slave_request(mdr
);
1979 /* This function DOES put the passed message before returning*/
1980 void Server::handle_slave_request_reply(MMDSSlaveRequest
*m
)
1982 mds_rank_t from
= mds_rank_t(m
->get_source().num());
1984 if (!mds
->is_clientreplay() && !mds
->is_active() && !mds
->is_stopping()) {
1985 metareqid_t r
= m
->get_reqid();
1986 if (!mdcache
->have_uncommitted_master(r
, from
)) {
1987 dout(10) << "handle_slave_request_reply ignoring slave reply from mds."
1988 << from
<< " reqid " << r
<< dendl
;
1992 dout(3) << "not clientreplay|active yet, waiting" << dendl
;
1993 mds
->wait_for_replay(new C_MDS_RetryMessage(mds
, m
));
1997 if (m
->get_op() == MMDSSlaveRequest::OP_COMMITTED
) {
1998 metareqid_t r
= m
->get_reqid();
1999 mdcache
->committed_master_slave(r
, from
);
2004 MDRequestRef mdr
= mdcache
->request_get(m
->get_reqid());
2005 if (m
->get_attempt() != mdr
->attempt
) {
2006 dout(10) << "handle_slave_request_reply " << *mdr
<< " ignoring reply from other attempt "
2007 << m
->get_attempt() << dendl
;
2012 switch (m
->get_op()) {
2013 case MMDSSlaveRequest::OP_XLOCKACK
:
2015 // identify lock, master request
2016 SimpleLock
*lock
= mds
->locker
->get_lock(m
->get_lock_type(),
2017 m
->get_object_info());
2018 mdr
->more()->slaves
.insert(from
);
2019 dout(10) << "got remote xlock on " << *lock
<< " on " << *lock
->get_parent() << dendl
;
2020 mdr
->xlocks
.insert(lock
);
2021 mdr
->locks
.insert(lock
);
2022 mdr
->finish_locking(lock
);
2023 lock
->get_xlock(mdr
, mdr
->get_client());
2025 assert(mdr
->more()->waiting_on_slave
.count(from
));
2026 mdr
->more()->waiting_on_slave
.erase(from
);
2027 assert(mdr
->more()->waiting_on_slave
.empty());
2028 mdcache
->dispatch_request(mdr
);
2032 case MMDSSlaveRequest::OP_WRLOCKACK
:
2034 // identify lock, master request
2035 SimpleLock
*lock
= mds
->locker
->get_lock(m
->get_lock_type(),
2036 m
->get_object_info());
2037 mdr
->more()->slaves
.insert(from
);
2038 dout(10) << "got remote wrlock on " << *lock
<< " on " << *lock
->get_parent() << dendl
;
2039 mdr
->remote_wrlocks
[lock
] = from
;
2040 mdr
->locks
.insert(lock
);
2041 mdr
->finish_locking(lock
);
2043 assert(mdr
->more()->waiting_on_slave
.count(from
));
2044 mdr
->more()->waiting_on_slave
.erase(from
);
2045 assert(mdr
->more()->waiting_on_slave
.empty());
2046 mdcache
->dispatch_request(mdr
);
2050 case MMDSSlaveRequest::OP_AUTHPINACK
:
2051 handle_slave_auth_pin_ack(mdr
, m
);
2054 case MMDSSlaveRequest::OP_LINKPREPACK
:
2055 handle_slave_link_prep_ack(mdr
, m
);
2058 case MMDSSlaveRequest::OP_RMDIRPREPACK
:
2059 handle_slave_rmdir_prep_ack(mdr
, m
);
2062 case MMDSSlaveRequest::OP_RENAMEPREPACK
:
2063 handle_slave_rename_prep_ack(mdr
, m
);
2066 case MMDSSlaveRequest::OP_RENAMENOTIFYACK
:
2067 handle_slave_rename_notify_ack(mdr
, m
);
2078 /* This function DOES put the mdr->slave_request before returning*/
2079 void Server::dispatch_slave_request(MDRequestRef
& mdr
)
2081 dout(7) << "dispatch_slave_request " << *mdr
<< " " << *mdr
->slave_request
<< dendl
;
2084 dout(7) << " abort flag set, finishing" << dendl
;
2085 mdcache
->request_finish(mdr
);
2089 if (logger
) logger
->inc(l_mdss_dispatch_slave_request
);
2091 int op
= mdr
->slave_request
->get_op();
2093 case MMDSSlaveRequest::OP_XLOCK
:
2094 case MMDSSlaveRequest::OP_WRLOCK
:
2097 SimpleLock
*lock
= mds
->locker
->get_lock(mdr
->slave_request
->get_lock_type(),
2098 mdr
->slave_request
->get_object_info());
2101 dout(10) << "don't have object, dropping" << dendl
;
2102 ceph_abort(); // can this happen, if we auth pinned properly.
2104 if (op
== MMDSSlaveRequest::OP_XLOCK
&& !lock
->get_parent()->is_auth()) {
2105 dout(10) << "not auth for remote xlock attempt, dropping on "
2106 << *lock
<< " on " << *lock
->get_parent() << dendl
;
2108 // use acquire_locks so that we get auth_pinning.
2109 set
<SimpleLock
*> rdlocks
;
2110 set
<SimpleLock
*> wrlocks
= mdr
->wrlocks
;
2111 set
<SimpleLock
*> xlocks
= mdr
->xlocks
;
2115 case MMDSSlaveRequest::OP_XLOCK
:
2116 xlocks
.insert(lock
);
2117 replycode
= MMDSSlaveRequest::OP_XLOCKACK
;
2119 case MMDSSlaveRequest::OP_WRLOCK
:
2120 wrlocks
.insert(lock
);
2121 replycode
= MMDSSlaveRequest::OP_WRLOCKACK
;
2125 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
2129 MMDSSlaveRequest
*r
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, replycode
);
2130 r
->set_lock_type(lock
->get_type());
2131 lock
->get_parent()->set_object_info(r
->get_object_info());
2132 mds
->send_message(r
, mdr
->slave_request
->get_connection());
2136 mdr
->slave_request
->put();
2137 mdr
->slave_request
= 0;
2141 case MMDSSlaveRequest::OP_UNXLOCK
:
2142 case MMDSSlaveRequest::OP_UNWRLOCK
:
2144 SimpleLock
*lock
= mds
->locker
->get_lock(mdr
->slave_request
->get_lock_type(),
2145 mdr
->slave_request
->get_object_info());
2147 bool need_issue
= false;
2149 case MMDSSlaveRequest::OP_UNXLOCK
:
2150 mds
->locker
->xlock_finish(lock
, mdr
.get(), &need_issue
);
2152 case MMDSSlaveRequest::OP_UNWRLOCK
:
2153 mds
->locker
->wrlock_finish(lock
, mdr
.get(), &need_issue
);
2157 mds
->locker
->issue_caps(static_cast<CInode
*>(lock
->get_parent()));
2159 // done. no ack necessary.
2160 mdr
->slave_request
->put();
2161 mdr
->slave_request
= 0;
2165 case MMDSSlaveRequest::OP_DROPLOCKS
:
2166 mds
->locker
->drop_locks(mdr
.get());
2167 mdr
->slave_request
->put();
2168 mdr
->slave_request
= 0;
2171 case MMDSSlaveRequest::OP_AUTHPIN
:
2172 handle_slave_auth_pin(mdr
);
2175 case MMDSSlaveRequest::OP_LINKPREP
:
2176 case MMDSSlaveRequest::OP_UNLINKPREP
:
2177 handle_slave_link_prep(mdr
);
2180 case MMDSSlaveRequest::OP_RMDIRPREP
:
2181 handle_slave_rmdir_prep(mdr
);
2184 case MMDSSlaveRequest::OP_RENAMEPREP
:
2185 handle_slave_rename_prep(mdr
);
2188 case MMDSSlaveRequest::OP_FINISH
:
2189 // information about rename imported caps
2190 if (mdr
->slave_request
->inode_export
.length() > 0)
2191 mdr
->more()->inode_import
.claim(mdr
->slave_request
->inode_export
);
2192 // finish off request.
2193 mdcache
->request_finish(mdr
);
2201 /* This function DOES put the mdr->slave_request before returning*/
2202 void Server::handle_slave_auth_pin(MDRequestRef
& mdr
)
2204 dout(10) << "handle_slave_auth_pin " << *mdr
<< dendl
;
2206 // build list of objects
2207 list
<MDSCacheObject
*> objects
;
2208 CInode
*auth_pin_freeze
= NULL
;
2209 bool fail
= false, wouldblock
= false, readonly
= false;
2211 if (mdcache
->is_readonly()) {
2212 dout(10) << " read-only FS" << dendl
;
2218 for (vector
<MDSCacheObjectInfo
>::iterator p
= mdr
->slave_request
->get_authpins().begin();
2219 p
!= mdr
->slave_request
->get_authpins().end();
2221 MDSCacheObject
*object
= mdcache
->get_object(*p
);
2223 dout(10) << " don't have " << *p
<< dendl
;
2228 objects
.push_back(object
);
2229 if (*p
== mdr
->slave_request
->get_authpin_freeze())
2230 auth_pin_freeze
= static_cast<CInode
*>(object
);
2234 // can we auth pin them?
2236 for (list
<MDSCacheObject
*>::iterator p
= objects
.begin();
2239 if (!(*p
)->is_auth()) {
2240 dout(10) << " not auth for " << **p
<< dendl
;
2244 if (mdr
->is_auth_pinned(*p
))
2246 if (!mdr
->can_auth_pin(*p
)) {
2247 if (mdr
->slave_request
->is_nonblock()) {
2248 dout(10) << " can't auth_pin (freezing?) " << **p
<< " nonblocking" << dendl
;
2254 dout(10) << " waiting for authpinnable on " << **p
<< dendl
;
2255 (*p
)->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2256 mdr
->drop_local_auth_pins();
2258 mds
->locker
->notify_freeze_waiter(*p
);
2266 mdr
->drop_local_auth_pins(); // just in case
2268 /* freeze authpin wrong inode */
2269 if (mdr
->has_more() && mdr
->more()->is_freeze_authpin
&&
2270 mdr
->more()->rename_inode
!= auth_pin_freeze
)
2271 mdr
->unfreeze_auth_pin(true);
2273 /* handle_slave_rename_prep() call freeze_inode() to wait for all other operations
2274 * on the source inode to complete. This happens after all locks for the rename
2275 * operation are acquired. But to acquire locks, we need auth pin locks' parent
2276 * objects first. So there is an ABBA deadlock if someone auth pins the source inode
2277 * after locks are acquired and before Server::handle_slave_rename_prep() is called.
2278 * The solution is freeze the inode and prevent other MDRequests from getting new
2281 if (auth_pin_freeze
) {
2282 dout(10) << " freezing auth pin on " << *auth_pin_freeze
<< dendl
;
2283 if (!mdr
->freeze_auth_pin(auth_pin_freeze
)) {
2284 auth_pin_freeze
->add_waiter(CInode::WAIT_FROZEN
, new C_MDS_RetryRequest(mdcache
, mdr
));
2285 mds
->mdlog
->flush();
2289 for (list
<MDSCacheObject
*>::iterator p
= objects
.begin();
2292 dout(10) << "auth_pinning " << **p
<< dendl
;
2298 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_AUTHPINACK
);
2300 // return list of my auth_pins (if any)
2301 for (set
<MDSCacheObject
*>::iterator p
= mdr
->auth_pins
.begin();
2302 p
!= mdr
->auth_pins
.end();
2304 MDSCacheObjectInfo info
;
2305 (*p
)->set_object_info(info
);
2306 reply
->get_authpins().push_back(info
);
2307 if (*p
== (MDSCacheObject
*)auth_pin_freeze
)
2308 auth_pin_freeze
->set_object_info(reply
->get_authpin_freeze());
2312 reply
->mark_error_wouldblock();
2314 reply
->mark_error_rofs();
2316 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
2318 // clean up this request
2319 mdr
->slave_request
->put();
2320 mdr
->slave_request
= 0;
2324 /* This function DOES NOT put the passed ack before returning*/
2325 void Server::handle_slave_auth_pin_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
2327 dout(10) << "handle_slave_auth_pin_ack on " << *mdr
<< " " << *ack
<< dendl
;
2328 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
2331 set
<MDSCacheObject
*> pinned
;
2332 for (vector
<MDSCacheObjectInfo
>::iterator p
= ack
->get_authpins().begin();
2333 p
!= ack
->get_authpins().end();
2335 MDSCacheObject
*object
= mdcache
->get_object(*p
);
2336 assert(object
); // we pinned it
2337 dout(10) << " remote has pinned " << *object
<< dendl
;
2338 if (!mdr
->is_auth_pinned(object
))
2339 mdr
->remote_auth_pins
[object
] = from
;
2340 if (*p
== ack
->get_authpin_freeze())
2341 mdr
->set_remote_frozen_auth_pin(static_cast<CInode
*>(object
));
2342 pinned
.insert(object
);
2345 // removed frozen auth pin ?
2346 if (mdr
->more()->is_remote_frozen_authpin
&&
2347 ack
->get_authpin_freeze() == MDSCacheObjectInfo()) {
2348 auto p
= mdr
->remote_auth_pins
.find(mdr
->more()->rename_inode
);
2349 assert(p
!= mdr
->remote_auth_pins
.end());
2350 if (p
->second
== from
) {
2351 mdr
->more()->is_remote_frozen_authpin
= false;
2355 // removed auth pins?
2356 map
<MDSCacheObject
*, mds_rank_t
>::iterator p
= mdr
->remote_auth_pins
.begin();
2357 while (p
!= mdr
->remote_auth_pins
.end()) {
2358 MDSCacheObject
* object
= p
->first
;
2359 if (p
->second
== from
&& pinned
.count(object
) == 0) {
2360 dout(10) << " remote has unpinned " << *object
<< dendl
;
2361 mdr
->remote_auth_pins
.erase(p
++);
2367 if (ack
->is_error_rofs()) {
2368 mdr
->more()->slave_error
= -EROFS
;
2369 mdr
->aborted
= true;
2370 } else if (ack
->is_error_wouldblock()) {
2371 mdr
->more()->slave_error
= -EWOULDBLOCK
;
2372 mdr
->aborted
= true;
2376 mdr
->more()->slaves
.insert(from
);
2378 // clear from waiting list
2379 assert(mdr
->more()->waiting_on_slave
.count(from
));
2380 mdr
->more()->waiting_on_slave
.erase(from
);
2383 if (mdr
->more()->waiting_on_slave
.empty())
2384 mdcache
->dispatch_request(mdr
);
2386 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
2390 // ---------------------------------------
2395 * check whether we are permitted to complete a request
2397 * Check whether we have permission to perform the operation specified
2398 * by mask on the given inode, based on the capability in the mdr's
2401 bool Server::check_access(MDRequestRef
& mdr
, CInode
*in
, unsigned mask
)
2404 int r
= mdr
->session
->check_access(
2406 mdr
->client_request
->get_caller_uid(),
2407 mdr
->client_request
->get_caller_gid(),
2408 &mdr
->client_request
->get_caller_gid_list(),
2409 mdr
->client_request
->head
.args
.setattr
.uid
,
2410 mdr
->client_request
->head
.args
.setattr
.gid
);
2412 respond_to_request(mdr
, r
);
2420 * check whether fragment has reached maximum size
2423 bool Server::check_fragment_space(MDRequestRef
&mdr
, CDir
*in
)
2425 const auto size
= in
->get_frag_size();
2426 if (size
>= g_conf
->mds_bal_fragment_size_max
) {
2427 dout(10) << "fragment " << *in
<< " size exceeds " << g_conf
->mds_bal_fragment_size_max
<< " (ENOSPC)" << dendl
;
2428 respond_to_request(mdr
, -ENOSPC
);
2436 /** validate_dentry_dir
2438 * verify that the dir exists and would own the dname.
2439 * do not check if the dentry exists.
2441 CDir
*Server::validate_dentry_dir(MDRequestRef
& mdr
, CInode
*diri
, const string
& dname
)
2443 // make sure parent is a dir?
2444 if (!diri
->is_dir()) {
2445 dout(7) << "validate_dentry_dir: not a dir" << dendl
;
2446 respond_to_request(mdr
, -ENOTDIR
);
2451 frag_t fg
= diri
->pick_dirfrag(dname
);
2452 CDir
*dir
= try_open_auth_dirfrag(diri
, fg
, mdr
);
2457 if (dir
->is_frozen()) {
2458 dout(7) << "dir is frozen " << *dir
<< dendl
;
2459 dir
->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2467 /** prepare_null_dentry
2468 * prepare a null (or existing) dentry in given dir.
2469 * wait for any dn lock.
2471 CDentry
* Server::prepare_null_dentry(MDRequestRef
& mdr
, CDir
*dir
, const string
& dname
, bool okexist
)
2473 dout(10) << "prepare_null_dentry " << dname
<< " in " << *dir
<< dendl
;
2474 assert(dir
->is_auth());
2476 client_t client
= mdr
->get_client();
2478 // does it already exist?
2479 CDentry
*dn
= dir
->lookup(dname
);
2482 if (dn->lock.is_xlocked_by_other(mdr)) {
2483 dout(10) << "waiting on xlocked dentry " << *dn << dendl;
2484 dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryRequest(mdcache, mdr));
2488 if (!dn
->get_linkage(client
, mdr
)->is_null()) {
2489 // name already exists
2490 dout(10) << "dentry " << dname
<< " exists in " << *dir
<< dendl
;
2492 respond_to_request(mdr
, -EEXIST
);
2496 dn
->first
= dir
->inode
->find_snaprealm()->get_newest_seq() + 1;
2502 // make sure dir is complete
2503 if (!dir
->is_complete() && (!dir
->has_bloom() || dir
->is_in_bloom(dname
))) {
2504 dout(7) << " incomplete dir contents for " << *dir
<< ", fetching" << dendl
;
2505 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
));
2510 dn
= dir
->add_null_dentry(dname
, dir
->inode
->find_snaprealm()->get_newest_seq() + 1);
2512 dout(10) << "prepare_null_dentry added " << *dn
<< dendl
;
2516 CDentry
* Server::prepare_stray_dentry(MDRequestRef
& mdr
, CInode
*in
)
2518 CDentry
*straydn
= mdr
->straydn
;
2521 in
->name_stray_dentry(straydname
);
2522 if (straydn
->get_name() == straydname
)
2525 assert(!mdr
->done_locking
);
2526 mdr
->unpin(straydn
);
2529 CDir
*straydir
= mdcache
->get_stray_dir(in
);
2531 if (!mdr
->client_request
->is_replay() &&
2532 !check_fragment_space(mdr
, straydir
))
2535 straydn
= mdcache
->get_or_create_stray_dentry(in
);
2536 mdr
->straydn
= straydn
;
2541 /** prepare_new_inode
2543 * create a new inode. set c/m/atime. hit dir pop.
2545 CInode
* Server::prepare_new_inode(MDRequestRef
& mdr
, CDir
*dir
, inodeno_t useino
, unsigned mode
,
2546 file_layout_t
*layout
)
2548 CInode
*in
= new CInode(mdcache
);
2550 // Server::prepare_force_open_sessions() can re-open session in closing
2551 // state. In that corner case, session's prealloc_inos are being freed.
2552 // To simplify the code, we disallow using/refilling session's prealloc_ino
2553 // while session is opening.
2554 bool allow_prealloc_inos
= !mdr
->session
->is_opening();
2557 if (allow_prealloc_inos
&&
2558 mdr
->session
->info
.prealloc_inos
.size()) {
2559 mdr
->used_prealloc_ino
=
2560 in
->inode
.ino
= mdr
->session
->take_ino(useino
); // prealloc -> used
2561 mds
->sessionmap
.mark_projected(mdr
->session
);
2563 dout(10) << "prepare_new_inode used_prealloc " << mdr
->used_prealloc_ino
2564 << " (" << mdr
->session
->info
.prealloc_inos
2565 << ", " << mdr
->session
->info
.prealloc_inos
.size() << " left)"
2569 in
->inode
.ino
= mds
->inotable
->project_alloc_id();
2570 dout(10) << "prepare_new_inode alloc " << mdr
->alloc_ino
<< dendl
;
2573 if (useino
&& useino
!= in
->inode
.ino
) {
2574 dout(0) << "WARNING: client specified " << useino
<< " and i allocated " << in
->inode
.ino
<< dendl
;
2575 mds
->clog
->error() << mdr
->client_request
->get_source()
2576 << " specified ino " << useino
2577 << " but mds." << mds
->get_nodeid() << " allocated " << in
->inode
.ino
;
2578 //ceph_abort(); // just for now.
2581 if (allow_prealloc_inos
&&
2582 mdr
->session
->get_num_projected_prealloc_inos() < g_conf
->mds_client_prealloc_inos
/ 2) {
2583 int need
= g_conf
->mds_client_prealloc_inos
- mdr
->session
->get_num_projected_prealloc_inos();
2584 mds
->inotable
->project_alloc_ids(mdr
->prealloc_inos
, need
);
2585 assert(mdr
->prealloc_inos
.size()); // or else fix projected increment semantics
2586 mdr
->session
->pending_prealloc_inos
.insert(mdr
->prealloc_inos
);
2587 mds
->sessionmap
.mark_projected(mdr
->session
);
2588 dout(10) << "prepare_new_inode prealloc " << mdr
->prealloc_inos
<< dendl
;
2591 in
->inode
.version
= 1;
2592 in
->inode
.xattr_version
= 1;
2593 in
->inode
.nlink
= 1; // FIXME
2595 in
->inode
.mode
= mode
;
2597 memset(&in
->inode
.dir_layout
, 0, sizeof(in
->inode
.dir_layout
));
2598 if (in
->inode
.is_dir()) {
2599 in
->inode
.dir_layout
.dl_dir_hash
= g_conf
->mds_default_dir_hash
;
2600 } else if (layout
) {
2601 in
->inode
.layout
= *layout
;
2603 in
->inode
.layout
= mdcache
->default_file_layout
;
2606 in
->inode
.truncate_size
= -1ull; // not truncated, yet!
2607 in
->inode
.truncate_seq
= 1; /* starting with 1, 0 is kept for no-truncation logic */
2609 CInode
*diri
= dir
->get_inode();
2611 dout(10) << oct
<< " dir mode 0" << diri
->inode
.mode
<< " new mode 0" << mode
<< dec
<< dendl
;
2613 if (diri
->inode
.mode
& S_ISGID
) {
2614 dout(10) << " dir is sticky" << dendl
;
2615 in
->inode
.gid
= diri
->inode
.gid
;
2616 if (S_ISDIR(mode
)) {
2617 dout(10) << " new dir also sticky" << dendl
;
2618 in
->inode
.mode
|= S_ISGID
;
2621 in
->inode
.gid
= mdr
->client_request
->get_caller_gid();
2623 in
->inode
.uid
= mdr
->client_request
->get_caller_uid();
2625 in
->inode
.btime
= in
->inode
.ctime
= in
->inode
.mtime
= in
->inode
.atime
=
2626 mdr
->get_op_stamp();
2628 in
->inode
.change_attr
= 0;
2630 MClientRequest
*req
= mdr
->client_request
;
2631 if (req
->get_data().length()) {
2632 bufferlist::iterator p
= req
->get_data().begin();
2634 // xattrs on new inode?
2635 map
<string
,bufferptr
> xattrs
;
2636 ::decode(xattrs
, p
);
2637 for (map
<string
,bufferptr
>::iterator p
= xattrs
.begin(); p
!= xattrs
.end(); ++p
) {
2638 dout(10) << "prepare_new_inode setting xattr " << p
->first
<< dendl
;
2639 in
->xattrs
[p
->first
] = p
->second
;
2643 if (!mds
->mdsmap
->get_inline_data_enabled() ||
2644 !mdr
->session
->connection
->has_feature(CEPH_FEATURE_MDS_INLINE_DATA
))
2645 in
->inode
.inline_data
.version
= CEPH_INLINE_NONE
;
2647 mdcache
->add_inode(in
); // add
2648 dout(10) << "prepare_new_inode " << *in
<< dendl
;
2652 void Server::journal_allocated_inos(MDRequestRef
& mdr
, EMetaBlob
*blob
)
2654 dout(20) << "journal_allocated_inos sessionmapv " << mds
->sessionmap
.get_projected()
2655 << " inotablev " << mds
->inotable
->get_projected_version()
2657 blob
->set_ino_alloc(mdr
->alloc_ino
,
2658 mdr
->used_prealloc_ino
,
2660 mdr
->client_request
->get_source(),
2661 mds
->sessionmap
.get_projected(),
2662 mds
->inotable
->get_projected_version());
2665 void Server::apply_allocated_inos(MDRequestRef
& mdr
, Session
*session
)
2667 dout(10) << "apply_allocated_inos " << mdr
->alloc_ino
2668 << " / " << mdr
->prealloc_inos
2669 << " / " << mdr
->used_prealloc_ino
<< dendl
;
2671 if (mdr
->alloc_ino
) {
2672 mds
->inotable
->apply_alloc_id(mdr
->alloc_ino
);
2674 if (mdr
->prealloc_inos
.size()) {
2676 session
->pending_prealloc_inos
.subtract(mdr
->prealloc_inos
);
2677 session
->info
.prealloc_inos
.insert(mdr
->prealloc_inos
);
2678 mds
->sessionmap
.mark_dirty(session
);
2679 mds
->inotable
->apply_alloc_ids(mdr
->prealloc_inos
);
2681 if (mdr
->used_prealloc_ino
) {
2683 session
->info
.used_inos
.erase(mdr
->used_prealloc_ino
);
2684 mds
->sessionmap
.mark_dirty(session
);
2688 class C_MDS_TryFindInode
: public ServerContext
{
2691 C_MDS_TryFindInode(Server
*s
, MDRequestRef
& r
) : ServerContext(s
), mdr(r
) {}
2692 void finish(int r
) override
{
2693 if (r
== -ESTALE
) // :( find_ino_peers failed
2694 server
->respond_to_request(mdr
, r
);
2696 server
->dispatch_client_request(mdr
);
2700 CDir
*Server::traverse_to_auth_dir(MDRequestRef
& mdr
, vector
<CDentry
*> &trace
, filepath refpath
)
2702 // figure parent dir vs dname
2703 if (refpath
.depth() == 0) {
2704 dout(7) << "can't do that to root" << dendl
;
2705 respond_to_request(mdr
, -EINVAL
);
2708 string dname
= refpath
.last_dentry();
2709 refpath
.pop_dentry();
2711 dout(10) << "traverse_to_auth_dir dirpath " << refpath
<< " dname " << dname
<< dendl
;
2713 // traverse to parent dir
2715 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, refpath
, &trace
, &diri
, MDS_TRAVERSE_FORWARD
);
2716 if (r
> 0) return 0; // delayed
2719 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
2720 mdcache
->find_ino_peers(refpath
.get_ino(), new C_MDS_TryFindInode(this, mdr
));
2723 respond_to_request(mdr
, r
);
2727 // is it an auth dir?
2728 CDir
*dir
= validate_dentry_dir(mdr
, diri
, dname
);
2730 return 0; // forwarded or waiting for freeze
2732 dout(10) << "traverse_to_auth_dir " << *dir
<< dendl
;
2736 /* If this returns null, the request has been handled
2737 * as appropriate: forwarded on, or the client's been replied to */
2738 CInode
* Server::rdlock_path_pin_ref(MDRequestRef
& mdr
, int n
,
2739 set
<SimpleLock
*> &rdlocks
,
2741 bool no_want_auth
, /* for readdir, who doesn't want auth _even_if_ it's
2743 file_layout_t
**layout
,
2744 bool no_lookup
) // true if we cannot return a null dentry lease
2746 const filepath
& refpath
= n
? mdr
->get_filepath2() : mdr
->get_filepath();
2747 dout(10) << "rdlock_path_pin_ref " << *mdr
<< " " << refpath
<< dendl
;
2749 if (mdr
->done_locking
)
2753 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, refpath
, &mdr
->dn
[n
], &mdr
->in
[n
], MDS_TRAVERSE_FORWARD
);
2755 return NULL
; // delayed
2756 if (r
< 0) { // error
2757 if (r
== -ENOENT
&& n
== 0 && mdr
->dn
[n
].size()) {
2759 mdr
->tracedn
= mdr
->dn
[n
][mdr
->dn
[n
].size()-1];
2760 respond_to_request(mdr
, r
);
2761 } else if (r
== -ESTALE
) {
2762 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
2763 MDSInternalContextBase
*c
= new C_MDS_TryFindInode(this, mdr
);
2764 mdcache
->find_ino_peers(refpath
.get_ino(), c
);
2766 dout(10) << "FAIL on error " << r
<< dendl
;
2767 respond_to_request(mdr
, r
);
2771 CInode
*ref
= mdr
->in
[n
];
2772 dout(10) << "ref is " << *ref
<< dendl
;
2774 // fw to inode auth?
2775 if (mdr
->snapid
!= CEPH_NOSNAP
&& !no_want_auth
)
2779 if (ref
->is_ambiguous_auth()) {
2780 dout(10) << "waiting for single auth on " << *ref
<< dendl
;
2781 ref
->add_waiter(CInode::WAIT_SINGLEAUTH
, new C_MDS_RetryRequest(mdcache
, mdr
));
2784 if (!ref
->is_auth()) {
2785 dout(10) << "fw to auth for " << *ref
<< dendl
;
2786 mdcache
->request_forward(mdr
, ref
->authority().first
);
2791 // do NOT proceed if freezing, as cap release may defer in that case, and
2792 // we could deadlock when we try to lock @ref.
2793 // if we're already auth_pinned, continue; the release has already been processed.
2794 if (ref
->is_frozen() || ref
->is_frozen_auth_pin() ||
2795 (ref
->is_freezing() && !mdr
->is_auth_pinned(ref
))) {
2796 dout(7) << "waiting for !frozen/authpinnable on " << *ref
<< dendl
;
2797 ref
->add_waiter(CInode::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2798 /* If we have any auth pins, this will deadlock.
2799 * But the only way to get here if we've already got auth pins
2800 * is because we're on an inode with snapshots that got updated
2801 * between dispatches of this request. So we're going to drop
2802 * our locks and our auth pins and reacquire them later.
2804 * This is safe since we're only in this function when working on
2805 * a single MDS request; otherwise we'd be in
2806 * rdlock_path_xlock_dentry.
2808 mds
->locker
->drop_locks(mdr
.get(), NULL
);
2809 mdr
->drop_local_auth_pins();
2810 if (!mdr
->remote_auth_pins
.empty())
2811 mds
->locker
->notify_freeze_waiter(ref
);
2818 for (int i
=0; i
<(int)mdr
->dn
[n
].size(); i
++)
2819 rdlocks
.insert(&mdr
->dn
[n
][i
]->lock
);
2821 mds
->locker
->include_snap_rdlocks_wlayout(rdlocks
, ref
, layout
);
2823 mds
->locker
->include_snap_rdlocks(rdlocks
, ref
);
2831 /** rdlock_path_xlock_dentry
2832 * traverse path to the directory that could/would contain dentry.
2833 * make sure i am auth for that dentry, forward as necessary.
2834 * create null dentry in place (or use existing if okexist).
2835 * get rdlocks on traversed dentries, xlock on new dentry.
2837 CDentry
* Server::rdlock_path_xlock_dentry(MDRequestRef
& mdr
, int n
,
2838 set
<SimpleLock
*>& rdlocks
, set
<SimpleLock
*>& wrlocks
, set
<SimpleLock
*>& xlocks
,
2839 bool okexist
, bool mustexist
, bool alwaysxlock
,
2840 file_layout_t
**layout
)
2842 const filepath
& refpath
= n
? mdr
->get_filepath2() : mdr
->get_filepath();
2844 dout(10) << "rdlock_path_xlock_dentry " << *mdr
<< " " << refpath
<< dendl
;
2846 client_t client
= mdr
->get_client();
2848 if (mdr
->done_locking
)
2849 return mdr
->dn
[n
].back();
2851 CDir
*dir
= traverse_to_auth_dir(mdr
, mdr
->dn
[n
], refpath
);
2854 CInode
*diri
= dir
->get_inode();
2855 if (!mdr
->reqid
.name
.is_mds()) {
2856 if (diri
->is_system() && !diri
->is_root()) {
2857 respond_to_request(mdr
, -EROFS
);
2861 if (!diri
->is_base() && diri
->get_projected_parent_dir()->inode
->is_stray()) {
2862 respond_to_request(mdr
, -ENOENT
);
2866 // make a null dentry?
2867 const string
&dname
= refpath
.last_dentry();
2870 dn
= dir
->lookup(dname
);
2872 // make sure dir is complete
2873 if (!dn
&& !dir
->is_complete() &&
2874 (!dir
->has_bloom() || dir
->is_in_bloom(dname
))) {
2875 dout(7) << " incomplete dir contents for " << *dir
<< ", fetching" << dendl
;
2876 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
));
2881 if (dn
&& !dn
->lock
.can_read(client
) && dn
->lock
.get_xlock_by() != mdr
) {
2882 dout(10) << "waiting on xlocked dentry " << *dn
<< dendl
;
2883 dn
->lock
.add_waiter(SimpleLock::WAIT_RD
, new C_MDS_RetryRequest(mdcache
, mdr
));
2888 if (!dn
|| dn
->get_linkage(client
, mdr
)->is_null()) {
2889 dout(7) << "dentry " << dname
<< " dne in " << *dir
<< dendl
;
2890 respond_to_request(mdr
, -ENOENT
);
2894 dn
= prepare_null_dentry(mdr
, dir
, dname
, okexist
);
2899 mdr
->dn
[n
].push_back(dn
);
2900 CDentry::linkage_t
*dnl
= dn
->get_linkage(client
, mdr
);
2901 mdr
->in
[n
] = dnl
->get_inode();
2904 // NOTE: rename takes the same set of locks for srcdn
2905 for (int i
=0; i
<(int)mdr
->dn
[n
].size(); i
++)
2906 rdlocks
.insert(&mdr
->dn
[n
][i
]->lock
);
2907 if (alwaysxlock
|| dnl
->is_null())
2908 xlocks
.insert(&dn
->lock
); // new dn, xlock
2910 rdlocks
.insert(&dn
->lock
); // existing dn, rdlock
2911 wrlocks
.insert(&dn
->get_dir()->inode
->filelock
); // also, wrlock on dir mtime
2912 wrlocks
.insert(&dn
->get_dir()->inode
->nestlock
); // also, wrlock on dir mtime
2914 mds
->locker
->include_snap_rdlocks_wlayout(rdlocks
, dn
->get_dir()->inode
, layout
);
2916 mds
->locker
->include_snap_rdlocks(rdlocks
, dn
->get_dir()->inode
);
2926 * try_open_auth_dirfrag -- open dirfrag, or forward to dirfrag auth
2928 * @param diri base inode
2929 * @param fg the exact frag we want
2930 * @param mdr request
2931 * @returns the pointer, or NULL if it had to be delayed (but mdr is taken care of)
2933 CDir
* Server::try_open_auth_dirfrag(CInode
*diri
, frag_t fg
, MDRequestRef
& mdr
)
2935 CDir
*dir
= diri
->get_dirfrag(fg
);
2937 // not open and inode not mine?
2938 if (!dir
&& !diri
->is_auth()) {
2939 mds_rank_t inauth
= diri
->authority().first
;
2940 dout(7) << "try_open_auth_dirfrag: not open, not inode auth, fw to mds." << inauth
<< dendl
;
2941 mdcache
->request_forward(mdr
, inauth
);
2945 // not open and inode frozen?
2946 if (!dir
&& diri
->is_frozen()) {
2947 dout(10) << "try_open_auth_dirfrag: dir inode is frozen, waiting " << *diri
<< dendl
;
2948 assert(diri
->get_parent_dir());
2949 diri
->add_waiter(CInode::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2955 dir
= diri
->get_or_open_dirfrag(mdcache
, fg
);
2957 // am i auth for the dirfrag?
2958 if (!dir
->is_auth()) {
2959 mds_rank_t auth
= dir
->authority().first
;
2960 dout(7) << "try_open_auth_dirfrag: not auth for " << *dir
2961 << ", fw to mds." << auth
<< dendl
;
2962 mdcache
->request_forward(mdr
, auth
);
2970 // ===============================================================================
2973 void Server::handle_client_getattr(MDRequestRef
& mdr
, bool is_lookup
)
2975 MClientRequest
*req
= mdr
->client_request
;
2976 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
2978 if (req
->get_filepath().depth() == 0 && is_lookup
) {
2979 // refpath can't be empty for lookup but it can for
2980 // getattr (we do getattr with empty refpath for mount of '/')
2981 respond_to_request(mdr
, -EINVAL
);
2985 CInode
*ref
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, false, false, NULL
, !is_lookup
);
2989 * if client currently holds the EXCL cap on a field, do not rdlock
2990 * it; client's stat() will result in valid info if _either_ EXCL
2991 * cap is held or MDS rdlocks and reads the value here.
2993 * handling this case here is easier than weakening rdlock
2994 * semantics... that would cause problems elsewhere.
2996 client_t client
= mdr
->get_client();
2998 Capability
*cap
= ref
->get_client_cap(client
);
2999 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
||
3000 mdr
->snapid
<= cap
->client_follows
))
3001 issued
= cap
->issued();
3003 int mask
= req
->head
.args
.getattr
.mask
;
3004 if ((mask
& CEPH_CAP_LINK_SHARED
) && (issued
& CEPH_CAP_LINK_EXCL
) == 0) rdlocks
.insert(&ref
->linklock
);
3005 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0) rdlocks
.insert(&ref
->authlock
);
3006 if ((mask
& CEPH_CAP_FILE_SHARED
) && (issued
& CEPH_CAP_FILE_EXCL
) == 0) rdlocks
.insert(&ref
->filelock
);
3007 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0) rdlocks
.insert(&ref
->xattrlock
);
3009 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3012 if (!check_access(mdr
, ref
, MAY_READ
))
3015 // note which caps are requested, so we return at least a snapshot
3016 // value for them. (currently this matters for xattrs and inline data)
3017 mdr
->getattr_caps
= mask
;
3019 mds
->balancer
->hit_inode(ceph_clock_now(), ref
, META_POP_IRD
,
3020 req
->get_source().num());
3023 dout(10) << "reply to stat on " << *req
<< dendl
;
3026 mdr
->tracedn
= mdr
->dn
[0].back();
3027 respond_to_request(mdr
, 0);
3030 struct C_MDS_LookupIno2
: public ServerContext
{
3032 C_MDS_LookupIno2(Server
*s
, MDRequestRef
& r
) : ServerContext(s
), mdr(r
) {}
3033 void finish(int r
) override
{
3034 server
->_lookup_ino_2(mdr
, r
);
3038 /* This function DOES clean up the mdr before returning*/
3042 void Server::handle_client_lookup_ino(MDRequestRef
& mdr
,
3043 bool want_parent
, bool want_dentry
)
3045 MClientRequest
*req
= mdr
->client_request
;
3047 inodeno_t ino
= req
->get_filepath().get_ino();
3048 CInode
*in
= mdcache
->get_inode(ino
);
3049 if (in
&& in
->state_test(CInode::STATE_PURGING
)) {
3050 respond_to_request(mdr
, -ESTALE
);
3054 mdcache
->open_ino(ino
, (int64_t)-1, new C_MDS_LookupIno2(this, mdr
), false);
3058 if (mdr
&& in
->snaprealm
&& !in
->snaprealm
->is_open() &&
3059 !in
->snaprealm
->open_parents(new C_MDS_RetryRequest(mdcache
, mdr
))) {
3063 // check for nothing (not read or write); this still applies the
3065 if (!check_access(mdr
, in
, 0))
3068 CDentry
*dn
= in
->get_projected_parent_dn();
3069 CInode
*diri
= dn
? dn
->get_dir()->inode
: NULL
;
3071 set
<SimpleLock
*> rdlocks
;
3072 if (dn
&& (want_parent
|| want_dentry
)) {
3074 rdlocks
.insert(&dn
->lock
);
3077 unsigned mask
= req
->head
.args
.getattr
.mask
;
3079 Capability
*cap
= in
->get_client_cap(mdr
->get_client());
3081 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
|| mdr
->snapid
<= cap
->client_follows
))
3082 issued
= cap
->issued();
3083 // permission bits, ACL/security xattrs
3084 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0)
3085 rdlocks
.insert(&in
->authlock
);
3086 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0)
3087 rdlocks
.insert(&in
->xattrlock
);
3089 mdr
->getattr_caps
= mask
;
3092 if (!rdlocks
.empty()) {
3093 set
<SimpleLock
*> wrlocks
, xlocks
;
3094 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3098 // need read access to directory inode
3099 if (!check_access(mdr
, diri
, MAY_READ
))
3105 if (in
->is_base()) {
3106 respond_to_request(mdr
, -EINVAL
);
3109 if (!diri
|| diri
->is_stray()) {
3110 respond_to_request(mdr
, -ESTALE
);
3113 dout(10) << "reply to lookup_parent " << *in
<< dendl
;
3115 respond_to_request(mdr
, 0);
3118 inodeno_t dirino
= req
->get_filepath2().get_ino();
3119 if (!diri
|| (dirino
!= inodeno_t() && diri
->ino() != dirino
)) {
3120 respond_to_request(mdr
, -ENOENT
);
3123 dout(10) << "reply to lookup_name " << *in
<< dendl
;
3125 dout(10) << "reply to lookup_ino " << *in
<< dendl
;
3130 respond_to_request(mdr
, 0);
3134 void Server::_lookup_ino_2(MDRequestRef
& mdr
, int r
)
3136 inodeno_t ino
= mdr
->client_request
->get_filepath().get_ino();
3137 dout(10) << "_lookup_ino_2 " << mdr
.get() << " ino " << ino
<< " r=" << r
<< dendl
;
3139 // `r` is a rank if >=0, else an error code
3141 mds_rank_t
dest_rank(r
);
3142 if (dest_rank
== mds
->get_nodeid())
3143 dispatch_client_request(mdr
);
3145 mdcache
->request_forward(mdr
, dest_rank
);
3150 if (r
== -ENOENT
|| r
== -ENODATA
)
3152 respond_to_request(mdr
, r
);
3156 /* This function takes responsibility for the passed mdr*/
3157 void Server::handle_client_open(MDRequestRef
& mdr
)
3159 MClientRequest
*req
= mdr
->client_request
;
3160 dout(7) << "open on " << req
->get_filepath() << dendl
;
3162 int flags
= req
->head
.args
.open
.flags
;
3163 int cmode
= ceph_flags_to_mode(flags
);
3165 respond_to_request(mdr
, -EINVAL
);
3169 bool need_auth
= !file_mode_is_readonly(cmode
) || (flags
& CEPH_O_TRUNC
);
3171 if ((cmode
& CEPH_FILE_MODE_WR
) && mdcache
->is_readonly()) {
3172 dout(7) << "read-only FS" << dendl
;
3173 respond_to_request(mdr
, -EROFS
);
3177 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3178 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, need_auth
);
3182 if (cur
->is_frozen() || cur
->state_test(CInode::STATE_EXPORTINGCAPS
)) {
3184 mdr
->done_locking
= false;
3185 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3190 if (!cur
->inode
.is_file()) {
3191 // can only open non-regular inode with mode FILE_MODE_PIN, at least for now.
3192 cmode
= CEPH_FILE_MODE_PIN
;
3193 // the inode is symlink and client wants to follow it, ignore the O_TRUNC flag.
3194 if (cur
->inode
.is_symlink() && !(flags
& CEPH_O_NOFOLLOW
))
3195 flags
&= ~CEPH_O_TRUNC
;
3198 dout(10) << "open flags = " << flags
3199 << ", filemode = " << cmode
3200 << ", need_auth = " << need_auth
3204 /*if (!cur->inode.is_file() && !cur->inode.is_dir()) {
3205 dout(7) << "not a file or dir " << *cur << dendl;
3206 respond_to_request(mdr, -ENXIO); // FIXME what error do we want?
3209 if ((flags
& CEPH_O_DIRECTORY
) && !cur
->inode
.is_dir() && !cur
->inode
.is_symlink()) {
3210 dout(7) << "specified O_DIRECTORY on non-directory " << *cur
<< dendl
;
3211 respond_to_request(mdr
, -EINVAL
);
3215 if ((flags
& CEPH_O_TRUNC
) && !cur
->inode
.is_file()) {
3216 dout(7) << "specified O_TRUNC on !(file|symlink) " << *cur
<< dendl
;
3217 // we should return -EISDIR for directory, return -EINVAL for other non-regular
3218 respond_to_request(mdr
, cur
->inode
.is_dir() ? -EISDIR
: -EINVAL
);
3222 if (cur
->inode
.inline_data
.version
!= CEPH_INLINE_NONE
&&
3223 !mdr
->session
->connection
->has_feature(CEPH_FEATURE_MDS_INLINE_DATA
)) {
3224 dout(7) << "old client cannot open inline data file " << *cur
<< dendl
;
3225 respond_to_request(mdr
, -EPERM
);
3229 // snapped data is read only
3230 if (mdr
->snapid
!= CEPH_NOSNAP
&&
3231 ((cmode
& CEPH_FILE_MODE_WR
) || req
->may_write())) {
3232 dout(7) << "snap " << mdr
->snapid
<< " is read-only " << *cur
<< dendl
;
3233 respond_to_request(mdr
, -EROFS
);
3237 unsigned mask
= req
->head
.args
.open
.mask
;
3239 Capability
*cap
= cur
->get_client_cap(mdr
->get_client());
3241 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
|| mdr
->snapid
<= cap
->client_follows
))
3242 issued
= cap
->issued();
3243 // permission bits, ACL/security xattrs
3244 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0)
3245 rdlocks
.insert(&cur
->authlock
);
3246 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0)
3247 rdlocks
.insert(&cur
->xattrlock
);
3249 mdr
->getattr_caps
= mask
;
3253 if ((flags
& CEPH_O_TRUNC
) && !mdr
->has_completed
) {
3254 assert(cur
->is_auth());
3256 xlocks
.insert(&cur
->filelock
);
3257 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3260 if (!check_access(mdr
, cur
, MAY_WRITE
))
3263 // wait for pending truncate?
3264 const inode_t
*pi
= cur
->get_projected_inode();
3265 if (pi
->is_truncating()) {
3266 dout(10) << " waiting for pending truncate from " << pi
->truncate_from
3267 << " to " << pi
->truncate_size
<< " to complete on " << *cur
<< dendl
;
3268 mds
->locker
->drop_locks(mdr
.get());
3269 mdr
->drop_local_auth_pins();
3270 cur
->add_waiter(CInode::WAIT_TRUNC
, new C_MDS_RetryRequest(mdcache
, mdr
));
3274 do_open_truncate(mdr
, cmode
);
3278 // sync filelock if snapped.
3279 // this makes us wait for writers to flushsnaps, ensuring we get accurate metadata,
3280 // and that data itself is flushed so that we can read the snapped data off disk.
3281 if (mdr
->snapid
!= CEPH_NOSNAP
&& !cur
->is_dir()) {
3282 rdlocks
.insert(&cur
->filelock
);
3285 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3289 if (cmode
& CEPH_FILE_MODE_WR
)
3291 if (!check_access(mdr
, cur
, mask
))
3294 if (cur
->is_file() || cur
->is_dir()) {
3295 if (mdr
->snapid
== CEPH_NOSNAP
) {
3297 Capability
*cap
= mds
->locker
->issue_new_caps(cur
, cmode
, mdr
->session
, 0, req
->is_replay());
3299 dout(12) << "open issued caps " << ccap_string(cap
->pending())
3300 << " for " << req
->get_source()
3301 << " on " << *cur
<< dendl
;
3303 int caps
= ceph_caps_for_mode(cmode
);
3304 dout(12) << "open issued IMMUTABLE SNAP caps " << ccap_string(caps
)
3305 << " for " << req
->get_source()
3306 << " snapid " << mdr
->snapid
3307 << " on " << *cur
<< dendl
;
3308 mdr
->snap_caps
= caps
;
3312 // increase max_size?
3313 if (cmode
& CEPH_FILE_MODE_WR
)
3314 mds
->locker
->check_inode_max_size(cur
);
3316 // make sure this inode gets into the journal
3317 if (cur
->is_auth() && cur
->last
== CEPH_NOSNAP
&&
3318 !cur
->item_open_file
.is_on_list()) {
3319 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
3320 EOpen
*le
= new EOpen(mds
->mdlog
);
3321 mdlog
->start_entry(le
);
3322 le
->add_clean_inode(cur
);
3323 ls
->open_files
.push_back(&cur
->item_open_file
);
3324 mdlog
->submit_entry(le
);
3328 if (cmode
& CEPH_FILE_MODE_WR
)
3329 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), cur
, META_POP_IWR
);
3331 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), cur
, META_POP_IRD
,
3332 mdr
->client_request
->get_source().num());
3335 if (req
->get_dentry_wanted()) {
3336 assert(mdr
->dn
[0].size());
3337 dn
= mdr
->dn
[0].back();
3342 respond_to_request(mdr
, 0);
3345 class C_MDS_openc_finish
: public ServerLogContext
{
3350 C_MDS_openc_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ni
, snapid_t f
) :
3351 ServerLogContext(s
, r
), dn(d
), newi(ni
), follows(f
) {}
3352 void finish(int r
) override
{
3355 dn
->pop_projected_linkage();
3357 // dirty inode, dn, dir
3358 newi
->inode
.version
--; // a bit hacky, see C_MDS_mknod_finish
3359 newi
->mark_dirty(newi
->inode
.version
+1, mdr
->ls
);
3360 newi
->_mark_dirty_parent(mdr
->ls
, true);
3364 get_mds()->locker
->share_inode_max_size(newi
);
3366 MDRequestRef null_ref
;
3367 get_mds()->mdcache
->send_dentry_link(dn
, null_ref
);
3369 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), newi
, META_POP_IWR
);
3371 server
->respond_to_request(mdr
, 0);
3373 assert(g_conf
->mds_kill_openc_at
!= 1);
3377 /* This function takes responsibility for the passed mdr*/
3378 void Server::handle_client_openc(MDRequestRef
& mdr
)
3380 MClientRequest
*req
= mdr
->client_request
;
3381 client_t client
= mdr
->get_client();
3383 dout(7) << "open w/ O_CREAT on " << req
->get_filepath() << dendl
;
3385 int cmode
= ceph_flags_to_mode(req
->head
.args
.open
.flags
);
3387 respond_to_request(mdr
, -EINVAL
);
3391 bool excl
= req
->head
.args
.open
.flags
& CEPH_O_EXCL
;
3394 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, req
->get_filepath(),
3395 &mdr
->dn
[0], NULL
, MDS_TRAVERSE_FORWARD
);
3399 handle_client_open(mdr
);
3402 if (r
< 0 && r
!= -ENOENT
) {
3404 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
3405 MDSInternalContextBase
*c
= new C_MDS_TryFindInode(this, mdr
);
3406 mdcache
->find_ino_peers(req
->get_filepath().get_ino(), c
);
3408 dout(10) << "FAIL on error " << r
<< dendl
;
3409 respond_to_request(mdr
, r
);
3415 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3416 file_layout_t
*dir_layout
= NULL
;
3417 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
,
3418 !excl
, false, false, &dir_layout
);
3420 if (mdr
->snapid
!= CEPH_NOSNAP
) {
3421 respond_to_request(mdr
, -EROFS
);
3425 file_layout_t layout
;
3427 layout
= *dir_layout
;
3429 layout
= mdcache
->default_file_layout
;
3431 // What kind of client caps are required to complete this operation
3432 uint64_t access
= MAY_WRITE
;
3434 const auto default_layout
= layout
;
3436 // fill in any special params from client
3437 if (req
->head
.args
.open
.stripe_unit
)
3438 layout
.stripe_unit
= req
->head
.args
.open
.stripe_unit
;
3439 if (req
->head
.args
.open
.stripe_count
)
3440 layout
.stripe_count
= req
->head
.args
.open
.stripe_count
;
3441 if (req
->head
.args
.open
.object_size
)
3442 layout
.object_size
= req
->head
.args
.open
.object_size
;
3443 if (req
->get_connection()->has_feature(CEPH_FEATURE_CREATEPOOLID
) &&
3444 (__s32
)req
->head
.args
.open
.pool
>= 0) {
3445 layout
.pool_id
= req
->head
.args
.open
.pool
;
3447 // make sure we have as new a map as the client
3448 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
3449 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
3454 // If client doesn't have capability to modify layout pools, then
3455 // only permit this request if the requested pool matches what the
3456 // file would have inherited anyway from its parent.
3457 if (default_layout
!= layout
) {
3458 access
|= MAY_SET_VXATTR
;
3461 if (!layout
.is_valid()) {
3462 dout(10) << " invalid initial file layout" << dendl
;
3463 respond_to_request(mdr
, -EINVAL
);
3466 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
3467 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
3468 respond_to_request(mdr
, -EINVAL
);
3473 CDir
*dir
= dn
->get_dir();
3474 CInode
*diri
= dir
->get_inode();
3475 rdlocks
.insert(&diri
->authlock
);
3476 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3479 if (!check_access(mdr
, diri
, access
))
3482 if (!check_fragment_space(mdr
, dir
))
3485 CDentry::linkage_t
*dnl
= dn
->get_projected_linkage();
3487 if (!dnl
->is_null()) {
3489 assert(req
->head
.args
.open
.flags
& CEPH_O_EXCL
);
3490 dout(10) << "O_EXCL, target exists, failing with -EEXIST" << dendl
;
3491 mdr
->tracei
= dnl
->get_inode();
3493 respond_to_request(mdr
, -EEXIST
);
3498 SnapRealm
*realm
= diri
->find_snaprealm(); // use directory's realm; inode isn't attached yet.
3499 snapid_t follows
= realm
->get_newest_seq();
3501 CInode
*in
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
),
3502 req
->head
.args
.open
.mode
| S_IFREG
, &layout
);
3506 dn
->push_projected_linkage(in
);
3508 in
->inode
.version
= dn
->pre_dirty();
3509 if (layout
.pool_id
!= mdcache
->default_file_layout
.pool_id
)
3510 in
->inode
.add_old_pool(mdcache
->default_file_layout
.pool_id
);
3511 in
->inode
.update_backtrace();
3512 if (cmode
& CEPH_FILE_MODE_WR
) {
3513 in
->inode
.client_ranges
[client
].range
.first
= 0;
3514 in
->inode
.client_ranges
[client
].range
.last
= in
->inode
.get_layout_size_increment();
3515 in
->inode
.client_ranges
[client
].follows
= follows
;
3517 in
->inode
.rstat
.rfiles
= 1;
3519 assert(dn
->first
== follows
+1);
3520 in
->first
= dn
->first
;
3523 mdr
->ls
= mdlog
->get_current_segment();
3524 EUpdate
*le
= new EUpdate(mdlog
, "openc");
3525 mdlog
->start_entry(le
);
3526 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
3527 journal_allocated_inos(mdr
, &le
->metablob
);
3528 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
3529 le
->metablob
.add_primary_dentry(dn
, in
, true, true, true);
3532 mds
->locker
->issue_new_caps(in
, cmode
, mdr
->session
, realm
, req
->is_replay());
3533 in
->authlock
.set_state(LOCK_EXCL
);
3534 in
->xattrlock
.set_state(LOCK_EXCL
);
3536 // make sure this inode gets into the journal
3537 le
->metablob
.add_opened_ino(in
->ino());
3538 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
3539 ls
->open_files
.push_back(&in
->item_open_file
);
3541 C_MDS_openc_finish
*fin
= new C_MDS_openc_finish(this, mdr
, dn
, in
, follows
);
3543 if (mdr
->client_request
->get_connection()->has_feature(CEPH_FEATURE_REPLY_CREATE_INODE
)) {
3544 dout(10) << "adding ino to reply to indicate inode was created" << dendl
;
3545 // add the file created flag onto the reply if create_flags features is supported
3546 ::encode(in
->inode
.ino
, mdr
->reply_extra_bl
);
3549 journal_and_reply(mdr
, in
, dn
, le
, fin
);
3551 // We hit_dir (via hit_inode) in our finish callback, but by then we might
3552 // have overshot the split size (multiple opencs in flight), so here is
3553 // an early chance to split the dir if this openc makes it oversized.
3554 mds
->balancer
->maybe_fragment(dir
, false);
3559 void Server::handle_client_readdir(MDRequestRef
& mdr
)
3561 MClientRequest
*req
= mdr
->client_request
;
3562 client_t client
= req
->get_source().num();
3563 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3564 CInode
*diri
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, false, true);
3567 // it's a directory, right?
3568 if (!diri
->is_dir()) {
3570 dout(10) << "reply to " << *req
<< " readdir -ENOTDIR" << dendl
;
3571 respond_to_request(mdr
, -ENOTDIR
);
3575 rdlocks
.insert(&diri
->filelock
);
3576 rdlocks
.insert(&diri
->dirfragtreelock
);
3578 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3581 if (!check_access(mdr
, diri
, MAY_READ
))
3585 frag_t fg
= (__u32
)req
->head
.args
.readdir
.frag
;
3586 unsigned req_flags
= (__u32
)req
->head
.args
.readdir
.flags
;
3587 string offset_str
= req
->get_path2();
3589 __u32 offset_hash
= 0;
3590 if (!offset_str
.empty())
3591 offset_hash
= ceph_frag_value(diri
->hash_dentry_name(offset_str
));
3593 offset_hash
= (__u32
)req
->head
.args
.readdir
.offset_hash
;
3595 dout(10) << " frag " << fg
<< " offset '" << offset_str
<< "'"
3596 << " offset_hash " << offset_hash
<< " flags " << req_flags
<< dendl
;
3598 // does the frag exist?
3599 if (diri
->dirfragtree
[fg
.value()] != fg
) {
3601 if (req_flags
& CEPH_READDIR_REPLY_BITFLAGS
) {
3602 if (fg
.contains((unsigned)offset_hash
)) {
3603 newfg
= diri
->dirfragtree
[offset_hash
];
3605 // client actually wants next frag
3606 newfg
= diri
->dirfragtree
[fg
.value()];
3610 newfg
= diri
->dirfragtree
[fg
.value()];
3612 dout(10) << " adjust frag " << fg
<< " -> " << newfg
<< " " << diri
->dirfragtree
<< dendl
;
3616 CDir
*dir
= try_open_auth_dirfrag(diri
, fg
, mdr
);
3620 dout(10) << "handle_client_readdir on " << *dir
<< dendl
;
3621 assert(dir
->is_auth());
3623 if (!dir
->is_complete()) {
3624 if (dir
->is_frozen()) {
3625 dout(7) << "dir is frozen " << *dir
<< dendl
;
3626 mds
->locker
->drop_locks(mdr
.get());
3627 mdr
->drop_local_auth_pins();
3628 dir
->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
3632 dout(10) << " incomplete dir contents for readdir on " << *dir
<< ", fetching" << dendl
;
3633 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
), true);
3637 #ifdef MDS_VERIFY_FRAGSTAT
3638 dir
->verify_fragstat();
3641 utime_t now
= ceph_clock_now();
3642 mdr
->set_mds_stamp(now
);
3644 snapid_t snapid
= mdr
->snapid
;
3645 dout(10) << "snapid " << snapid
<< dendl
;
3647 SnapRealm
*realm
= diri
->find_snaprealm();
3649 unsigned max
= req
->head
.args
.readdir
.max_entries
;
3651 max
= dir
->get_num_any(); // whatever, something big.
3652 unsigned max_bytes
= req
->head
.args
.readdir
.max_bytes
;
3654 // make sure at least one item can be encoded
3655 max_bytes
= (512 << 10) + g_conf
->mds_max_xattr_pairs_size
;
3659 dir
->encode_dirstat(dirbl
, mds
->get_nodeid());
3661 // count bytes available.
3662 // this isn't perfect, but we should capture the main variable/unbounded size items!
3663 int front_bytes
= dirbl
.length() + sizeof(__u32
) + sizeof(__u8
)*2;
3664 int bytes_left
= max_bytes
- front_bytes
;
3665 bytes_left
-= realm
->get_snap_trace().length();
3667 // build dir contents
3670 bool start
= !offset_hash
&& offset_str
.empty();
3671 bool end
= (dir
->begin() == dir
->end());
3672 // skip all dns < dentry_key_t(snapid, offset_str, offset_hash)
3673 dentry_key_t
skip_key(snapid
, offset_str
.c_str(), offset_hash
);
3674 for (CDir::map_t::iterator it
= start
? dir
->begin() : dir
->lower_bound(skip_key
);
3675 !end
&& numfiles
< max
;
3676 end
= (it
== dir
->end())) {
3677 CDentry
*dn
= it
->second
;
3680 if (dn
->state_test(CDentry::STATE_PURGING
))
3683 bool dnp
= dn
->use_projected(client
, mdr
);
3684 CDentry::linkage_t
*dnl
= dnp
? dn
->get_projected_linkage() : dn
->get_linkage();
3689 if (dn
->last
< snapid
|| dn
->first
> snapid
) {
3690 dout(20) << "skipping non-overlapping snap " << *dn
<< dendl
;
3695 dentry_key_t
offset_key(dn
->last
, offset_str
.c_str(), offset_hash
);
3696 if (!(offset_key
< dn
->key()))
3700 CInode
*in
= dnl
->get_inode();
3702 if (in
&& in
->ino() == CEPH_INO_CEPH
)
3706 // better for the MDS to do the work, if we think the client will stat any of these files.
3707 if (dnl
->is_remote() && !in
) {
3708 in
= mdcache
->get_inode(dnl
->get_remote_ino());
3710 dn
->link_remote(dnl
, in
);
3711 } else if (dn
->state_test(CDentry::STATE_BADREMOTEINO
)) {
3712 dout(10) << "skipping bad remote ino on " << *dn
<< dendl
;
3715 // touch everything i _do_ have
3716 for (CDir::map_t::iterator p
= dir
->begin(); p
!= dir
->end(); ++p
)
3717 if (!p
->second
->get_linkage()->is_null())
3718 mdcache
->lru
.lru_touch(p
->second
);
3720 // already issued caps and leases, reply immediately.
3721 if (dnbl
.length() > 0) {
3722 mdcache
->open_remote_dentry(dn
, dnp
, new C_MDSInternalNoop
);
3723 dout(10) << " open remote dentry after caps were issued, stopping at "
3724 << dnbl
.length() << " < " << bytes_left
<< dendl
;
3728 mds
->locker
->drop_locks(mdr
.get());
3729 mdr
->drop_local_auth_pins();
3730 mdcache
->open_remote_dentry(dn
, dnp
, new C_MDS_RetryRequest(mdcache
, mdr
));
3736 if ((int)(dnbl
.length() + dn
->name
.length() + sizeof(__u32
) + sizeof(LeaseStat
)) > bytes_left
) {
3737 dout(10) << " ran out of room, stopping at " << dnbl
.length() << " < " << bytes_left
<< dendl
;
3741 unsigned start_len
= dnbl
.length();
3744 dout(12) << "including dn " << *dn
<< dendl
;
3745 ::encode(dn
->name
, dnbl
);
3746 mds
->locker
->issue_client_lease(dn
, client
, dnbl
, now
, mdr
->session
);
3749 dout(12) << "including inode " << *in
<< dendl
;
3750 int r
= in
->encode_inodestat(dnbl
, mdr
->session
, realm
, snapid
, bytes_left
- (int)dnbl
.length());
3752 // chop off dn->name, lease
3753 dout(10) << " ran out of room, stopping at " << start_len
<< " < " << bytes_left
<< dendl
;
3755 keep
.substr_of(dnbl
, 0, start_len
);
3763 mdcache
->lru
.lru_touch(dn
);
3768 flags
= CEPH_READDIR_FRAG_END
;
3770 flags
|= CEPH_READDIR_FRAG_COMPLETE
; // FIXME: what purpose does this serve
3772 // client only understand END and COMPLETE flags ?
3773 if (req_flags
& CEPH_READDIR_REPLY_BITFLAGS
) {
3774 flags
|= CEPH_READDIR_HASH_ORDER
| CEPH_READDIR_OFFSET_HASH
;
3777 // finish final blob
3778 ::encode(numfiles
, dirbl
);
3779 ::encode(flags
, dirbl
);
3780 dirbl
.claim_append(dnbl
);
3783 dout(10) << "reply to " << *req
<< " readdir num=" << numfiles
3784 << " bytes=" << dirbl
.length()
3785 << " start=" << (int)start
3786 << " end=" << (int)end
3788 mdr
->reply_extra_bl
= dirbl
;
3790 // bump popularity. NOTE: this doesn't quite capture it.
3791 mds
->balancer
->hit_dir(now
, dir
, META_POP_IRD
, -1, numfiles
);
3795 respond_to_request(mdr
, 0);
3800 // ===============================================================================
3805 * finisher for basic inode updates
3807 class C_MDS_inode_update_finish
: public ServerLogContext
{
3809 bool truncating_smaller
, changed_ranges
;
3811 C_MDS_inode_update_finish(Server
*s
, MDRequestRef
& r
, CInode
*i
,
3812 bool sm
=false, bool cr
=false) :
3813 ServerLogContext(s
, r
), in(i
), truncating_smaller(sm
), changed_ranges(cr
) { }
3814 void finish(int r
) override
{
3818 in
->pop_and_dirty_projected_inode(mdr
->ls
);
3821 // notify any clients
3822 if (truncating_smaller
&& in
->inode
.is_truncating()) {
3823 get_mds()->locker
->issue_truncate(in
);
3824 get_mds()->mdcache
->truncate_inode(in
, mdr
->ls
);
3827 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
3829 server
->respond_to_request(mdr
, 0);
3832 get_mds()->locker
->share_inode_max_size(in
);
3836 void Server::handle_client_file_setlock(MDRequestRef
& mdr
)
3838 MClientRequest
*req
= mdr
->client_request
;
3839 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3841 // get the inode to operate on, and set up any locks needed for that
3842 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3846 xlocks
.insert(&cur
->flocklock
);
3847 /* acquire_locks will return true if it gets the locks. If it fails,
3848 it will redeliver this request at a later date, so drop the request.
3850 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
)) {
3851 dout(10) << "handle_client_file_setlock could not get locks!" << dendl
;
3855 // copy the lock change into a ceph_filelock so we can store/apply it
3856 ceph_filelock set_lock
;
3857 set_lock
.start
= req
->head
.args
.filelock_change
.start
;
3858 set_lock
.length
= req
->head
.args
.filelock_change
.length
;
3859 set_lock
.client
= req
->get_orig_source().num();
3860 set_lock
.owner
= req
->head
.args
.filelock_change
.owner
;
3861 set_lock
.pid
= req
->head
.args
.filelock_change
.pid
;
3862 set_lock
.type
= req
->head
.args
.filelock_change
.type
;
3863 bool will_wait
= req
->head
.args
.filelock_change
.wait
;
3865 dout(10) << "handle_client_file_setlock: " << set_lock
<< dendl
;
3867 ceph_lock_state_t
*lock_state
= NULL
;
3868 bool interrupt
= false;
3870 // get the appropriate lock state
3871 switch (req
->head
.args
.filelock_change
.rule
) {
3872 case CEPH_LOCK_FLOCK_INTR
:
3875 case CEPH_LOCK_FLOCK
:
3876 lock_state
= cur
->get_flock_lock_state();
3879 case CEPH_LOCK_FCNTL_INTR
:
3882 case CEPH_LOCK_FCNTL
:
3883 lock_state
= cur
->get_fcntl_lock_state();
3887 dout(10) << "got unknown lock type " << set_lock
.type
3888 << ", dropping request!" << dendl
;
3889 respond_to_request(mdr
, -EOPNOTSUPP
);
3893 dout(10) << " state prior to lock change: " << *lock_state
<< dendl
;
3894 if (CEPH_LOCK_UNLOCK
== set_lock
.type
) {
3895 list
<ceph_filelock
> activated_locks
;
3896 list
<MDSInternalContextBase
*> waiters
;
3897 if (lock_state
->is_waiting(set_lock
)) {
3898 dout(10) << " unlock removing waiting lock " << set_lock
<< dendl
;
3899 lock_state
->remove_waiting(set_lock
);
3900 cur
->take_waiting(CInode::WAIT_FLOCK
, waiters
);
3901 } else if (!interrupt
) {
3902 dout(10) << " unlock attempt on " << set_lock
<< dendl
;
3903 lock_state
->remove_lock(set_lock
, activated_locks
);
3904 cur
->take_waiting(CInode::WAIT_FLOCK
, waiters
);
3906 mds
->queue_waiters(waiters
);
3908 respond_to_request(mdr
, 0);
3910 dout(10) << " lock attempt on " << set_lock
<< dendl
;
3911 bool deadlock
= false;
3912 if (mdr
->more()->flock_was_waiting
&&
3913 !lock_state
->is_waiting(set_lock
)) {
3914 dout(10) << " was waiting for lock but not anymore, must have been canceled " << set_lock
<< dendl
;
3915 respond_to_request(mdr
, -EINTR
);
3916 } else if (!lock_state
->add_lock(set_lock
, will_wait
, mdr
->more()->flock_was_waiting
, &deadlock
)) {
3917 dout(10) << " it failed on this attempt" << dendl
;
3918 // couldn't set lock right now
3920 respond_to_request(mdr
, -EDEADLK
);
3921 } else if (!will_wait
) {
3922 respond_to_request(mdr
, -EWOULDBLOCK
);
3924 dout(10) << " added to waiting list" << dendl
;
3925 assert(lock_state
->is_waiting(set_lock
));
3926 mdr
->more()->flock_was_waiting
= true;
3927 mds
->locker
->drop_locks(mdr
.get());
3928 mdr
->drop_local_auth_pins();
3929 cur
->add_waiter(CInode::WAIT_FLOCK
, new C_MDS_RetryRequest(mdcache
, mdr
));
3932 respond_to_request(mdr
, 0);
3934 dout(10) << " state after lock change: " << *lock_state
<< dendl
;
3937 void Server::handle_client_file_readlock(MDRequestRef
& mdr
)
3939 MClientRequest
*req
= mdr
->client_request
;
3940 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3942 // get the inode to operate on, and set up any locks needed for that
3943 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3947 /* acquire_locks will return true if it gets the locks. If it fails,
3948 it will redeliver this request at a later date, so drop the request.
3950 rdlocks
.insert(&cur
->flocklock
);
3951 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
)) {
3952 dout(10) << "handle_client_file_readlock could not get locks!" << dendl
;
3956 // copy the lock change into a ceph_filelock so we can store/apply it
3957 ceph_filelock checking_lock
;
3958 checking_lock
.start
= req
->head
.args
.filelock_change
.start
;
3959 checking_lock
.length
= req
->head
.args
.filelock_change
.length
;
3960 checking_lock
.client
= req
->get_orig_source().num();
3961 checking_lock
.owner
= req
->head
.args
.filelock_change
.owner
;
3962 checking_lock
.pid
= req
->head
.args
.filelock_change
.pid
;
3963 checking_lock
.type
= req
->head
.args
.filelock_change
.type
;
3965 // get the appropriate lock state
3966 ceph_lock_state_t
*lock_state
= NULL
;
3967 switch (req
->head
.args
.filelock_change
.rule
) {
3968 case CEPH_LOCK_FLOCK
:
3969 lock_state
= cur
->get_flock_lock_state();
3972 case CEPH_LOCK_FCNTL
:
3973 lock_state
= cur
->get_fcntl_lock_state();
3977 dout(10) << "got unknown lock type " << checking_lock
.type
<< dendl
;
3978 respond_to_request(mdr
, -EINVAL
);
3981 lock_state
->look_for_lock(checking_lock
);
3984 ::encode(checking_lock
, lock_bl
);
3986 mdr
->reply_extra_bl
= lock_bl
;
3987 respond_to_request(mdr
, 0);
3990 void Server::handle_client_setattr(MDRequestRef
& mdr
)
3992 MClientRequest
*req
= mdr
->client_request
;
3993 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3994 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3997 if (mdr
->snapid
!= CEPH_NOSNAP
) {
3998 respond_to_request(mdr
, -EROFS
);
4001 if (cur
->ino() < MDS_INO_SYSTEM_BASE
&& !cur
->is_base()) {
4002 respond_to_request(mdr
, -EPERM
);
4006 __u32 mask
= req
->head
.args
.setattr
.mask
;
4007 __u32 access_mask
= MAY_WRITE
;
4010 if (mask
& (CEPH_SETATTR_MODE
|CEPH_SETATTR_UID
|CEPH_SETATTR_GID
|CEPH_SETATTR_BTIME
|CEPH_SETATTR_KILL_SGUID
))
4011 xlocks
.insert(&cur
->authlock
);
4012 if (mask
& (CEPH_SETATTR_MTIME
|CEPH_SETATTR_ATIME
|CEPH_SETATTR_SIZE
))
4013 xlocks
.insert(&cur
->filelock
);
4014 if (mask
& CEPH_SETATTR_CTIME
)
4015 wrlocks
.insert(&cur
->versionlock
);
4017 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4020 if ((mask
& CEPH_SETATTR_UID
) && (cur
->inode
.uid
!= req
->head
.args
.setattr
.uid
))
4021 access_mask
|= MAY_CHOWN
;
4023 if ((mask
& CEPH_SETATTR_GID
) && (cur
->inode
.gid
!= req
->head
.args
.setattr
.gid
))
4024 access_mask
|= MAY_CHGRP
;
4026 if (!check_access(mdr
, cur
, access_mask
))
4029 // trunc from bigger -> smaller?
4030 inode_t
*pi
= cur
->get_projected_inode();
4032 uint64_t old_size
= MAX(pi
->size
, req
->head
.args
.setattr
.old_size
);
4034 // ENOSPC on growing file while full, but allow shrinks
4035 if (is_full
&& req
->head
.args
.setattr
.size
> old_size
) {
4036 dout(20) << __func__
<< ": full, responding ENOSPC to setattr with larger size" << dendl
;
4037 respond_to_request(mdr
, -ENOSPC
);
4041 bool truncating_smaller
= false;
4042 if (mask
& CEPH_SETATTR_SIZE
) {
4043 truncating_smaller
= req
->head
.args
.setattr
.size
< old_size
;
4044 if (truncating_smaller
&& pi
->is_truncating()) {
4045 dout(10) << " waiting for pending truncate from " << pi
->truncate_from
4046 << " to " << pi
->truncate_size
<< " to complete on " << *cur
<< dendl
;
4047 mds
->locker
->drop_locks(mdr
.get());
4048 mdr
->drop_local_auth_pins();
4049 cur
->add_waiter(CInode::WAIT_TRUNC
, new C_MDS_RetryRequest(mdcache
, mdr
));
4054 bool changed_ranges
= false;
4057 mdr
->ls
= mdlog
->get_current_segment();
4058 EUpdate
*le
= new EUpdate(mdlog
, "setattr");
4059 mdlog
->start_entry(le
);
4061 pi
= cur
->project_inode();
4063 if (mask
& CEPH_SETATTR_UID
)
4064 pi
->uid
= req
->head
.args
.setattr
.uid
;
4065 if (mask
& CEPH_SETATTR_GID
)
4066 pi
->gid
= req
->head
.args
.setattr
.gid
;
4068 if (mask
& CEPH_SETATTR_MODE
)
4069 pi
->mode
= (pi
->mode
& ~07777) | (req
->head
.args
.setattr
.mode
& 07777);
4070 else if ((mask
& (CEPH_SETATTR_UID
|CEPH_SETATTR_GID
|CEPH_SETATTR_KILL_SGUID
)) &&
4071 S_ISREG(pi
->mode
)) {
4072 pi
->mode
&= ~S_ISUID
;
4073 if ((pi
->mode
& (S_ISGID
|S_IXGRP
)) == (S_ISGID
|S_IXGRP
))
4074 pi
->mode
&= ~S_ISGID
;
4077 if (mask
& CEPH_SETATTR_MTIME
)
4078 pi
->mtime
= req
->head
.args
.setattr
.mtime
;
4079 if (mask
& CEPH_SETATTR_ATIME
)
4080 pi
->atime
= req
->head
.args
.setattr
.atime
;
4081 if (mask
& CEPH_SETATTR_BTIME
)
4082 pi
->btime
= req
->head
.args
.setattr
.btime
;
4083 if (mask
& (CEPH_SETATTR_ATIME
| CEPH_SETATTR_MTIME
| CEPH_SETATTR_BTIME
))
4084 pi
->time_warp_seq
++; // maybe not a timewarp, but still a serialization point.
4085 if (mask
& CEPH_SETATTR_SIZE
) {
4086 if (truncating_smaller
) {
4087 pi
->truncate(old_size
, req
->head
.args
.setattr
.size
);
4088 le
->metablob
.add_truncate_start(cur
->ino());
4090 pi
->size
= req
->head
.args
.setattr
.size
;
4091 pi
->rstat
.rbytes
= pi
->size
;
4093 pi
->mtime
= mdr
->get_op_stamp();
4095 // adjust client's max_size?
4096 map
<client_t
,client_writeable_range_t
> new_ranges
;
4097 bool max_increased
= false;
4098 mds
->locker
->calc_new_client_ranges(cur
, pi
->size
, &new_ranges
, &max_increased
);
4099 if (pi
->client_ranges
!= new_ranges
) {
4100 dout(10) << " client_ranges " << pi
->client_ranges
<< " -> " << new_ranges
<< dendl
;
4101 pi
->client_ranges
= new_ranges
;
4102 changed_ranges
= true;
4106 pi
->version
= cur
->pre_dirty();
4107 pi
->ctime
= mdr
->get_op_stamp();
4111 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4112 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4113 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4115 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
,
4116 truncating_smaller
, changed_ranges
));
4118 // flush immediately if there are readers/writers waiting
4119 if (xlocks
.count(&cur
->filelock
) &&
4120 (cur
->get_caps_wanted() & (CEPH_CAP_FILE_RD
|CEPH_CAP_FILE_WR
)))
4121 mds
->mdlog
->flush();
4124 /* Takes responsibility for mdr */
4125 void Server::do_open_truncate(MDRequestRef
& mdr
, int cmode
)
4127 CInode
*in
= mdr
->in
[0];
4128 client_t client
= mdr
->get_client();
4131 dout(10) << "do_open_truncate " << *in
<< dendl
;
4133 SnapRealm
*realm
= in
->find_snaprealm();
4134 mds
->locker
->issue_new_caps(in
, cmode
, mdr
->session
, realm
, mdr
->client_request
->is_replay());
4136 mdr
->ls
= mdlog
->get_current_segment();
4137 EUpdate
*le
= new EUpdate(mdlog
, "open_truncate");
4138 mdlog
->start_entry(le
);
4141 inode_t
*pi
= in
->project_inode();
4142 pi
->version
= in
->pre_dirty();
4143 pi
->mtime
= pi
->ctime
= mdr
->get_op_stamp();
4146 uint64_t old_size
= MAX(pi
->size
, mdr
->client_request
->head
.args
.open
.old_size
);
4148 pi
->truncate(old_size
, 0);
4149 le
->metablob
.add_truncate_start(in
->ino());
4152 bool changed_ranges
= false;
4153 if (cmode
& CEPH_FILE_MODE_WR
) {
4154 pi
->client_ranges
[client
].range
.first
= 0;
4155 pi
->client_ranges
[client
].range
.last
= pi
->get_layout_size_increment();
4156 pi
->client_ranges
[client
].follows
= in
->find_snaprealm()->get_newest_seq();
4157 changed_ranges
= true;
4160 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
4162 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, 0, PREDIRTY_PRIMARY
);
4163 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, in
);
4165 // make sure ino gets into the journal
4166 le
->metablob
.add_opened_ino(in
->ino());
4167 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
4168 ls
->open_files
.push_back(&in
->item_open_file
);
4170 mdr
->o_trunc
= true;
4173 if (mdr
->client_request
->get_dentry_wanted()) {
4174 assert(mdr
->dn
[0].size());
4175 dn
= mdr
->dn
[0].back();
4178 journal_and_reply(mdr
, in
, dn
, le
, new C_MDS_inode_update_finish(this, mdr
, in
, old_size
> 0,
4180 // Although the `open` part can give an early reply, the truncation won't
4181 // happen until our EUpdate is persistent, to give the client a prompt
4182 // response we must also flush that event.
4187 /* This function cleans up the passed mdr */
4188 void Server::handle_client_setlayout(MDRequestRef
& mdr
)
4190 MClientRequest
*req
= mdr
->client_request
;
4191 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4192 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4195 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4196 respond_to_request(mdr
, -EROFS
);
4199 if (!cur
->is_file()) {
4200 respond_to_request(mdr
, -EINVAL
);
4203 if (cur
->get_projected_inode()->size
||
4204 cur
->get_projected_inode()->truncate_seq
> 1) {
4205 respond_to_request(mdr
, -ENOTEMPTY
);
4210 file_layout_t layout
= cur
->get_projected_inode()->layout
;
4211 // save existing layout for later
4212 const auto old_layout
= layout
;
4214 int access
= MAY_WRITE
;
4216 if (req
->head
.args
.setlayout
.layout
.fl_object_size
> 0)
4217 layout
.object_size
= req
->head
.args
.setlayout
.layout
.fl_object_size
;
4218 if (req
->head
.args
.setlayout
.layout
.fl_stripe_unit
> 0)
4219 layout
.stripe_unit
= req
->head
.args
.setlayout
.layout
.fl_stripe_unit
;
4220 if (req
->head
.args
.setlayout
.layout
.fl_stripe_count
> 0)
4221 layout
.stripe_count
=req
->head
.args
.setlayout
.layout
.fl_stripe_count
;
4222 if (req
->head
.args
.setlayout
.layout
.fl_pg_pool
> 0) {
4223 layout
.pool_id
= req
->head
.args
.setlayout
.layout
.fl_pg_pool
;
4225 // make sure we have as new a map as the client
4226 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
4227 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
4232 // Don't permit layout modifications without 'p' caps
4233 if (layout
!= old_layout
) {
4234 access
|= MAY_SET_VXATTR
;
4237 if (!layout
.is_valid()) {
4238 dout(10) << "bad layout" << dendl
;
4239 respond_to_request(mdr
, -EINVAL
);
4242 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
4243 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
4244 respond_to_request(mdr
, -EINVAL
);
4248 xlocks
.insert(&cur
->filelock
);
4249 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4252 if (!check_access(mdr
, cur
, access
))
4256 inode_t
*pi
= cur
->project_inode();
4257 pi
->layout
= layout
;
4258 // add the old pool to the inode
4259 pi
->add_old_pool(old_layout
.pool_id
);
4260 pi
->version
= cur
->pre_dirty();
4261 pi
->ctime
= mdr
->get_op_stamp();
4265 mdr
->ls
= mdlog
->get_current_segment();
4266 EUpdate
*le
= new EUpdate(mdlog
, "setlayout");
4267 mdlog
->start_entry(le
);
4268 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4269 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4270 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4272 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4275 void Server::handle_client_setdirlayout(MDRequestRef
& mdr
)
4277 MClientRequest
*req
= mdr
->client_request
;
4278 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4279 file_layout_t
*dir_layout
= NULL
;
4280 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4283 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4284 respond_to_request(mdr
, -EROFS
);
4288 if (!cur
->is_dir()) {
4289 respond_to_request(mdr
, -ENOTDIR
);
4293 xlocks
.insert(&cur
->policylock
);
4294 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4298 const inode_t
*old_pi
= cur
->get_projected_inode();
4299 file_layout_t layout
;
4300 if (old_pi
->has_layout())
4301 layout
= old_pi
->layout
;
4302 else if (dir_layout
)
4303 layout
= *dir_layout
;
4305 layout
= mdcache
->default_file_layout
;
4307 // Level of access required to complete
4308 int access
= MAY_WRITE
;
4310 const auto old_layout
= layout
;
4312 if (req
->head
.args
.setlayout
.layout
.fl_object_size
> 0)
4313 layout
.object_size
= req
->head
.args
.setlayout
.layout
.fl_object_size
;
4314 if (req
->head
.args
.setlayout
.layout
.fl_stripe_unit
> 0)
4315 layout
.stripe_unit
= req
->head
.args
.setlayout
.layout
.fl_stripe_unit
;
4316 if (req
->head
.args
.setlayout
.layout
.fl_stripe_count
> 0)
4317 layout
.stripe_count
=req
->head
.args
.setlayout
.layout
.fl_stripe_count
;
4318 if (req
->head
.args
.setlayout
.layout
.fl_pg_pool
> 0) {
4319 layout
.pool_id
= req
->head
.args
.setlayout
.layout
.fl_pg_pool
;
4320 // make sure we have as new a map as the client
4321 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
4322 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
4327 if (layout
!= old_layout
) {
4328 access
|= MAY_SET_VXATTR
;
4331 if (!layout
.is_valid()) {
4332 dout(10) << "bad layout" << dendl
;
4333 respond_to_request(mdr
, -EINVAL
);
4336 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
4337 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
4338 respond_to_request(mdr
, -EINVAL
);
4342 if (!check_access(mdr
, cur
, access
))
4345 inode_t
*pi
= cur
->project_inode();
4346 pi
->layout
= layout
;
4347 pi
->version
= cur
->pre_dirty();
4350 mdr
->ls
= mdlog
->get_current_segment();
4351 EUpdate
*le
= new EUpdate(mdlog
, "setlayout");
4352 mdlog
->start_entry(le
);
4353 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4354 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4355 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4357 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4362 int Server::parse_layout_vxattr(string name
, string value
, const OSDMap
& osdmap
,
4363 file_layout_t
*layout
, bool validate
)
4365 dout(20) << "parse_layout_vxattr name " << name
<< " value '" << value
<< "'" << dendl
;
4367 if (name
== "layout") {
4368 string::iterator begin
= value
.begin();
4369 string::iterator end
= value
.end();
4370 keys_and_values
<string::iterator
> p
; // create instance of parser
4371 std::map
<string
, string
> m
; // map to receive results
4372 if (!qi::parse(begin
, end
, p
, m
)) { // returns true if successful
4375 string
left(begin
, end
);
4376 dout(10) << " parsed " << m
<< " left '" << left
<< "'" << dendl
;
4379 for (map
<string
,string
>::iterator q
= m
.begin(); q
!= m
.end(); ++q
) {
4380 // Skip validation on each attr, we do it once at the end (avoid
4381 // rejecting intermediate states if the overall result is ok)
4382 int r
= parse_layout_vxattr(string("layout.") + q
->first
, q
->second
,
4383 osdmap
, layout
, false);
4387 } else if (name
== "layout.object_size") {
4388 layout
->object_size
= boost::lexical_cast
<unsigned>(value
);
4389 } else if (name
== "layout.stripe_unit") {
4390 layout
->stripe_unit
= boost::lexical_cast
<unsigned>(value
);
4391 } else if (name
== "layout.stripe_count") {
4392 layout
->stripe_count
= boost::lexical_cast
<unsigned>(value
);
4393 } else if (name
== "layout.pool") {
4395 layout
->pool_id
= boost::lexical_cast
<unsigned>(value
);
4396 } catch (boost::bad_lexical_cast
const&) {
4397 int64_t pool
= osdmap
.lookup_pg_pool_name(value
);
4399 dout(10) << " unknown pool " << value
<< dendl
;
4402 layout
->pool_id
= pool
;
4404 } else if (name
== "layout.pool_namespace") {
4405 layout
->pool_ns
= value
;
4407 dout(10) << " unknown layout vxattr " << name
<< dendl
;
4410 } catch (boost::bad_lexical_cast
const&) {
4411 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4415 if (validate
&& !layout
->is_valid()) {
4416 dout(10) << "bad layout" << dendl
;
4419 if (!mds
->mdsmap
->is_data_pool(layout
->pool_id
)) {
4420 dout(10) << " invalid data pool " << layout
->pool_id
<< dendl
;
4426 int Server::parse_quota_vxattr(string name
, string value
, quota_info_t
*quota
)
4428 dout(20) << "parse_quota_vxattr name " << name
<< " value '" << value
<< "'" << dendl
;
4430 if (name
== "quota") {
4431 string::iterator begin
= value
.begin();
4432 string::iterator end
= value
.end();
4433 keys_and_values
<string::iterator
> p
; // create instance of parser
4434 std::map
<string
, string
> m
; // map to receive results
4435 if (!qi::parse(begin
, end
, p
, m
)) { // returns true if successful
4438 string
left(begin
, end
);
4439 dout(10) << " parsed " << m
<< " left '" << left
<< "'" << dendl
;
4442 for (map
<string
,string
>::iterator q
= m
.begin(); q
!= m
.end(); ++q
) {
4443 int r
= parse_quota_vxattr(string("quota.") + q
->first
, q
->second
, quota
);
4447 } else if (name
== "quota.max_bytes") {
4448 int64_t q
= boost::lexical_cast
<int64_t>(value
);
4451 quota
->max_bytes
= q
;
4452 } else if (name
== "quota.max_files") {
4453 int64_t q
= boost::lexical_cast
<int64_t>(value
);
4456 quota
->max_files
= q
;
4458 dout(10) << " unknown quota vxattr " << name
<< dendl
;
4461 } catch (boost::bad_lexical_cast
const&) {
4462 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4466 if (!quota
->is_valid()) {
4467 dout(10) << "bad quota" << dendl
;
4474 * Verify that the file layout attribute carried by client
4475 * is well-formatted.
4476 * Return 0 on success, otherwise this function takes
4477 * responsibility for the passed mdr.
4479 int Server::check_layout_vxattr(MDRequestRef
& mdr
,
4482 file_layout_t
*layout
)
4484 MClientRequest
*req
= mdr
->client_request
;
4488 mds
->objecter
->with_osdmap([&](const OSDMap
& osdmap
) {
4489 r
= parse_layout_vxattr(name
, value
, osdmap
, layout
);
4490 epoch
= osdmap
.get_epoch();
4495 // we don't have the specified pool, make sure our map
4496 // is newer than or as new as the client.
4497 epoch_t req_epoch
= req
->get_osdmap_epoch();
4499 if (req_epoch
> epoch
) {
4501 // well, our map is older. consult mds.
4502 Context
*fin
= new C_IO_Wrapper(mds
, new C_MDS_RetryRequest(mdcache
, mdr
));
4504 if (!mds
->objecter
->wait_for_map(req_epoch
, fin
))
4505 return r
; // wait, fin will retry this request later
4509 // now we have at least as new a map as the client, try again.
4510 mds
->objecter
->with_osdmap([&](const OSDMap
& osdmap
) {
4511 r
= parse_layout_vxattr(name
, value
, osdmap
, layout
);
4512 epoch
= osdmap
.get_epoch();
4515 assert(epoch
>= req_epoch
); // otherwise wait_for_map() told a lie
4517 } else if (req_epoch
== 0 && !mdr
->waited_for_osdmap
) {
4519 // For compatibility with client w/ old code, we still need get the
4520 // latest map. One day if COMPACT_VERSION of MClientRequest >=3,
4521 // we can remove those code.
4522 mdr
->waited_for_osdmap
= true;
4523 mds
->objecter
->wait_for_latest_osdmap(new C_IO_Wrapper(
4524 mds
, new C_MDS_RetryRequest(mdcache
, mdr
)));
4534 respond_to_request(mdr
, r
);
4542 void Server::handle_set_vxattr(MDRequestRef
& mdr
, CInode
*cur
,
4543 file_layout_t
*dir_layout
,
4544 set
<SimpleLock
*> rdlocks
,
4545 set
<SimpleLock
*> wrlocks
,
4546 set
<SimpleLock
*> xlocks
)
4548 MClientRequest
*req
= mdr
->client_request
;
4549 string
name(req
->get_path2());
4550 bufferlist bl
= req
->get_data();
4551 string
value (bl
.c_str(), bl
.length());
4552 dout(10) << "handle_set_vxattr " << name
4553 << " val " << value
.length()
4554 << " bytes on " << *cur
4560 if (!check_access(mdr
, cur
, MAY_SET_VXATTR
)) {
4564 if (name
.compare(0, 15, "ceph.dir.layout") == 0) {
4565 if (!cur
->is_dir()) {
4566 respond_to_request(mdr
, -EINVAL
);
4570 file_layout_t layout
;
4571 if (cur
->get_projected_inode()->has_layout())
4572 layout
= cur
->get_projected_inode()->layout
;
4573 else if (dir_layout
)
4574 layout
= *dir_layout
;
4576 layout
= mdcache
->default_file_layout
;
4578 rest
= name
.substr(name
.find("layout"));
4579 if (check_layout_vxattr(mdr
, rest
, value
, &layout
) < 0)
4582 xlocks
.insert(&cur
->policylock
);
4583 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4586 pi
= cur
->project_inode();
4587 pi
->layout
= layout
;
4588 } else if (name
.compare(0, 16, "ceph.file.layout") == 0) {
4589 if (!cur
->is_file()) {
4590 respond_to_request(mdr
, -EINVAL
);
4593 if (cur
->get_projected_inode()->size
||
4594 cur
->get_projected_inode()->truncate_seq
> 1) {
4595 respond_to_request(mdr
, -ENOTEMPTY
);
4598 file_layout_t layout
= cur
->get_projected_inode()->layout
;
4599 rest
= name
.substr(name
.find("layout"));
4600 if (check_layout_vxattr(mdr
, rest
, value
, &layout
) < 0)
4603 xlocks
.insert(&cur
->filelock
);
4604 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4607 pi
= cur
->project_inode();
4608 int64_t old_pool
= pi
->layout
.pool_id
;
4609 pi
->add_old_pool(old_pool
);
4610 pi
->layout
= layout
;
4611 pi
->ctime
= mdr
->get_op_stamp();
4612 } else if (name
.compare(0, 10, "ceph.quota") == 0) {
4613 if (!cur
->is_dir() || cur
->is_root()) {
4614 respond_to_request(mdr
, -EINVAL
);
4618 quota_info_t quota
= cur
->get_projected_inode()->quota
;
4620 rest
= name
.substr(name
.find("quota"));
4621 int r
= parse_quota_vxattr(rest
, value
, "a
);
4623 respond_to_request(mdr
, r
);
4627 xlocks
.insert(&cur
->policylock
);
4628 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4631 pi
= cur
->project_inode();
4633 } else if (name
.find("ceph.dir.pin") == 0) {
4634 if (!cur
->is_dir() || cur
->is_root()) {
4635 respond_to_request(mdr
, -EINVAL
);
4641 rank
= boost::lexical_cast
<mds_rank_t
>(value
);
4642 if (rank
< 0) rank
= MDS_RANK_NONE
;
4643 } catch (boost::bad_lexical_cast
const&) {
4644 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4645 respond_to_request(mdr
, -EINVAL
);
4649 xlocks
.insert(&cur
->policylock
);
4650 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4653 pi
= cur
->project_inode();
4654 cur
->set_export_pin(rank
);
4656 dout(10) << " unknown vxattr " << name
<< dendl
;
4657 respond_to_request(mdr
, -EINVAL
);
4662 pi
->ctime
= mdr
->get_op_stamp();
4663 pi
->version
= cur
->pre_dirty();
4665 pi
->update_backtrace();
4668 mdr
->ls
= mdlog
->get_current_segment();
4669 EUpdate
*le
= new EUpdate(mdlog
, "set vxattr layout");
4670 mdlog
->start_entry(le
);
4671 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4672 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4673 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4675 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4679 void Server::handle_remove_vxattr(MDRequestRef
& mdr
, CInode
*cur
,
4680 file_layout_t
*dir_layout
,
4681 set
<SimpleLock
*> rdlocks
,
4682 set
<SimpleLock
*> wrlocks
,
4683 set
<SimpleLock
*> xlocks
)
4685 MClientRequest
*req
= mdr
->client_request
;
4686 string
name(req
->get_path2());
4688 dout(10) << __func__
<< " " << name
<< " on " << *cur
<< dendl
;
4690 if (name
== "ceph.dir.layout") {
4691 if (!cur
->is_dir()) {
4692 respond_to_request(mdr
, -ENODATA
);
4695 if (cur
->is_root()) {
4696 dout(10) << "can't remove layout policy on the root directory" << dendl
;
4697 respond_to_request(mdr
, -EINVAL
);
4701 if (!cur
->get_projected_inode()->has_layout()) {
4702 respond_to_request(mdr
, -ENODATA
);
4706 xlocks
.insert(&cur
->policylock
);
4707 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4710 inode_t
*pi
= cur
->project_inode();
4712 pi
->version
= cur
->pre_dirty();
4715 mdr
->ls
= mdlog
->get_current_segment();
4716 EUpdate
*le
= new EUpdate(mdlog
, "remove dir layout vxattr");
4717 mdlog
->start_entry(le
);
4718 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4719 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4720 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4722 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4724 } else if (name
== "ceph.dir.layout.pool_namespace"
4725 || name
== "ceph.file.layout.pool_namespace") {
4726 // Namespace is the only layout field that has a meaningful
4727 // null/none value (empty string, means default layout). Is equivalent
4728 // to a setxattr with empty string: pass through the empty payload of
4729 // the rmxattr request to do this.
4730 handle_set_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4734 respond_to_request(mdr
, -ENODATA
);
4737 class C_MDS_inode_xattr_update_finish
: public ServerLogContext
{
4741 C_MDS_inode_xattr_update_finish(Server
*s
, MDRequestRef
& r
, CInode
*i
) :
4742 ServerLogContext(s
, r
), in(i
) { }
4743 void finish(int r
) override
{
4747 in
->pop_and_dirty_projected_inode(mdr
->ls
);
4751 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
4753 server
->respond_to_request(mdr
, 0);
4757 void Server::handle_client_setxattr(MDRequestRef
& mdr
)
4759 MClientRequest
*req
= mdr
->client_request
;
4760 string
name(req
->get_path2());
4761 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4764 file_layout_t
*dir_layout
= NULL
;
4765 if (name
.compare(0, 15, "ceph.dir.layout") == 0)
4766 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4768 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4772 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4773 respond_to_request(mdr
, -EROFS
);
4777 int flags
= req
->head
.args
.setxattr
.flags
;
4779 // magic ceph.* namespace?
4780 if (name
.compare(0, 5, "ceph.") == 0) {
4781 handle_set_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4785 xlocks
.insert(&cur
->xattrlock
);
4786 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4789 if (!check_access(mdr
, cur
, MAY_WRITE
))
4792 map
<string
, bufferptr
> *pxattrs
= cur
->get_projected_xattrs();
4793 size_t len
= req
->get_data().length();
4794 size_t inc
= len
+ name
.length();
4796 // check xattrs kv pairs size
4797 size_t cur_xattrs_size
= 0;
4798 for (const auto& p
: *pxattrs
) {
4799 if ((flags
& CEPH_XATTR_REPLACE
) && (name
.compare(p
.first
) == 0)) {
4802 cur_xattrs_size
+= p
.first
.length() + p
.second
.length();
4805 if (((cur_xattrs_size
+ inc
) > g_conf
->mds_max_xattr_pairs_size
)) {
4806 dout(10) << "xattr kv pairs size too big. cur_xattrs_size "
4807 << cur_xattrs_size
<< ", inc " << inc
<< dendl
;
4808 respond_to_request(mdr
, -ENOSPC
);
4812 if ((flags
& CEPH_XATTR_CREATE
) && pxattrs
->count(name
)) {
4813 dout(10) << "setxattr '" << name
<< "' XATTR_CREATE and EEXIST on " << *cur
<< dendl
;
4814 respond_to_request(mdr
, -EEXIST
);
4817 if ((flags
& CEPH_XATTR_REPLACE
) && !pxattrs
->count(name
)) {
4818 dout(10) << "setxattr '" << name
<< "' XATTR_REPLACE and ENODATA on " << *cur
<< dendl
;
4819 respond_to_request(mdr
, -ENODATA
);
4823 dout(10) << "setxattr '" << name
<< "' len " << len
<< " on " << *cur
<< dendl
;
4826 map
<string
,bufferptr
> *px
= new map
<string
,bufferptr
>;
4827 inode_t
*pi
= cur
->project_inode(px
);
4828 pi
->version
= cur
->pre_dirty();
4829 pi
->ctime
= mdr
->get_op_stamp();
4831 pi
->xattr_version
++;
4833 if (!(flags
& CEPH_XATTR_REMOVE
)) {
4834 (*px
)[name
] = buffer::create(len
);
4836 req
->get_data().copy(0, len
, (*px
)[name
].c_str());
4840 mdr
->ls
= mdlog
->get_current_segment();
4841 EUpdate
*le
= new EUpdate(mdlog
, "setxattr");
4842 mdlog
->start_entry(le
);
4843 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4844 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4845 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4847 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4850 void Server::handle_client_removexattr(MDRequestRef
& mdr
)
4852 MClientRequest
*req
= mdr
->client_request
;
4853 string
name(req
->get_path2());
4854 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4855 file_layout_t
*dir_layout
= NULL
;
4857 if (name
== "ceph.dir.layout")
4858 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4860 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4864 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4865 respond_to_request(mdr
, -EROFS
);
4869 if (name
.compare(0, 5, "ceph.") == 0) {
4870 handle_remove_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4874 xlocks
.insert(&cur
->xattrlock
);
4875 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4878 map
<string
, bufferptr
> *pxattrs
= cur
->get_projected_xattrs();
4879 if (pxattrs
->count(name
) == 0) {
4880 dout(10) << "removexattr '" << name
<< "' and ENODATA on " << *cur
<< dendl
;
4881 respond_to_request(mdr
, -ENODATA
);
4885 dout(10) << "removexattr '" << name
<< "' on " << *cur
<< dendl
;
4888 map
<string
,bufferptr
> *px
= new map
<string
,bufferptr
>;
4889 inode_t
*pi
= cur
->project_inode(px
);
4890 pi
->version
= cur
->pre_dirty();
4891 pi
->ctime
= mdr
->get_op_stamp();
4893 pi
->xattr_version
++;
4897 mdr
->ls
= mdlog
->get_current_segment();
4898 EUpdate
*le
= new EUpdate(mdlog
, "removexattr");
4899 mdlog
->start_entry(le
);
4900 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4901 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4902 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4904 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4908 // =================================================================
4909 // DIRECTORY and NAMESPACE OPS
4912 // ------------------------------------------------
4916 class C_MDS_mknod_finish
: public ServerLogContext
{
4920 C_MDS_mknod_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ni
) :
4921 ServerLogContext(s
, r
), dn(d
), newi(ni
) {}
4922 void finish(int r
) override
{
4926 dn
->pop_projected_linkage();
4928 // be a bit hacky with the inode version, here.. we decrement it
4929 // just to keep mark_dirty() happen. (we didn't bother projecting
4930 // a new version of hte inode since it's just been created)
4931 newi
->inode
.version
--;
4932 newi
->mark_dirty(newi
->inode
.version
+ 1, mdr
->ls
);
4933 newi
->_mark_dirty_parent(mdr
->ls
, true);
4936 if (newi
->inode
.is_dir()) {
4937 CDir
*dir
= newi
->get_dirfrag(frag_t());
4939 dir
->fnode
.version
--;
4940 dir
->mark_dirty(dir
->fnode
.version
+ 1, mdr
->ls
);
4941 dir
->mark_new(mdr
->ls
);
4946 MDRequestRef null_ref
;
4947 get_mds()->mdcache
->send_dentry_link(dn
, null_ref
);
4949 if (newi
->inode
.is_file())
4950 get_mds()->locker
->share_inode_max_size(newi
);
4953 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), newi
, META_POP_IWR
);
4956 server
->respond_to_request(mdr
, 0);
4961 void Server::handle_client_mknod(MDRequestRef
& mdr
)
4963 MClientRequest
*req
= mdr
->client_request
;
4964 client_t client
= mdr
->get_client();
4965 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4966 file_layout_t
*dir_layout
= NULL
;
4967 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false,
4970 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4971 respond_to_request(mdr
, -EROFS
);
4974 CInode
*diri
= dn
->get_dir()->get_inode();
4975 rdlocks
.insert(&diri
->authlock
);
4976 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4979 if (!check_access(mdr
, diri
, MAY_WRITE
))
4982 if (!check_fragment_space(mdr
, dn
->get_dir()))
4985 unsigned mode
= req
->head
.args
.mknod
.mode
;
4986 if ((mode
& S_IFMT
) == 0)
4990 file_layout_t layout
;
4991 if (dir_layout
&& S_ISREG(mode
))
4992 layout
= *dir_layout
;
4994 layout
= mdcache
->default_file_layout
;
4996 SnapRealm
*realm
= dn
->get_dir()->inode
->find_snaprealm();
4997 snapid_t follows
= realm
->get_newest_seq();
4998 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
),
5002 dn
->push_projected_linkage(newi
);
5004 newi
->inode
.rdev
= req
->head
.args
.mknod
.rdev
;
5005 newi
->inode
.version
= dn
->pre_dirty();
5006 newi
->inode
.rstat
.rfiles
= 1;
5007 if (layout
.pool_id
!= mdcache
->default_file_layout
.pool_id
)
5008 newi
->inode
.add_old_pool(mdcache
->default_file_layout
.pool_id
);
5009 newi
->inode
.update_backtrace();
5011 // if the client created a _regular_ file via MKNOD, it's highly likely they'll
5012 // want to write to it (e.g., if they are reexporting NFS)
5013 if (S_ISREG(newi
->inode
.mode
)) {
5014 dout(15) << " setting a client_range too, since this is a regular file" << dendl
;
5015 newi
->inode
.client_ranges
[client
].range
.first
= 0;
5016 newi
->inode
.client_ranges
[client
].range
.last
= newi
->inode
.get_layout_size_increment();
5017 newi
->inode
.client_ranges
[client
].follows
= follows
;
5019 // issue a cap on the file
5020 int cmode
= CEPH_FILE_MODE_RDWR
;
5021 Capability
*cap
= mds
->locker
->issue_new_caps(newi
, cmode
, mdr
->session
, realm
, req
->is_replay());
5025 // put locks in excl mode
5026 newi
->filelock
.set_state(LOCK_EXCL
);
5027 newi
->authlock
.set_state(LOCK_EXCL
);
5028 newi
->xattrlock
.set_state(LOCK_EXCL
);
5032 assert(dn
->first
== follows
+ 1);
5033 newi
->first
= dn
->first
;
5035 dout(10) << "mknod mode " << newi
->inode
.mode
<< " rdev " << newi
->inode
.rdev
<< dendl
;
5038 mdr
->ls
= mdlog
->get_current_segment();
5039 EUpdate
*le
= new EUpdate(mdlog
, "mknod");
5040 mdlog
->start_entry(le
);
5041 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5042 journal_allocated_inos(mdr
, &le
->metablob
);
5044 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(),
5045 PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5046 le
->metablob
.add_primary_dentry(dn
, newi
, true, true, true);
5048 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5054 /* This function takes responsibility for the passed mdr*/
5055 void Server::handle_client_mkdir(MDRequestRef
& mdr
)
5057 MClientRequest
*req
= mdr
->client_request
;
5058 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5059 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5061 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5062 respond_to_request(mdr
, -EROFS
);
5065 CDir
*dir
= dn
->get_dir();
5066 CInode
*diri
= dir
->get_inode();
5067 rdlocks
.insert(&diri
->authlock
);
5068 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5071 // mkdir check access
5072 if (!check_access(mdr
, diri
, MAY_WRITE
))
5075 if (!check_fragment_space(mdr
, dir
))
5079 SnapRealm
*realm
= dn
->get_dir()->inode
->find_snaprealm();
5080 snapid_t follows
= realm
->get_newest_seq();
5082 unsigned mode
= req
->head
.args
.mkdir
.mode
;
5085 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
), mode
);
5088 // it's a directory.
5089 dn
->push_projected_linkage(newi
);
5091 newi
->inode
.version
= dn
->pre_dirty();
5092 newi
->inode
.rstat
.rsubdirs
= 1;
5093 newi
->inode
.update_backtrace();
5095 dout(12) << " follows " << follows
<< dendl
;
5096 assert(dn
->first
== follows
+ 1);
5097 newi
->first
= dn
->first
;
5099 // ...and that new dir is empty.
5100 CDir
*newdir
= newi
->get_or_open_dirfrag(mdcache
, frag_t());
5101 newdir
->state_set(CDir::STATE_CREATING
);
5102 newdir
->mark_complete();
5103 newdir
->fnode
.version
= newdir
->pre_dirty();
5106 mdr
->ls
= mdlog
->get_current_segment();
5107 EUpdate
*le
= new EUpdate(mdlog
, "mkdir");
5108 mdlog
->start_entry(le
);
5109 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5110 journal_allocated_inos(mdr
, &le
->metablob
);
5111 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5112 le
->metablob
.add_primary_dentry(dn
, newi
, true, true);
5113 le
->metablob
.add_new_dir(newdir
); // dirty AND complete AND new
5115 // issue a cap on the directory
5116 int cmode
= CEPH_FILE_MODE_RDWR
;
5117 Capability
*cap
= mds
->locker
->issue_new_caps(newi
, cmode
, mdr
->session
, realm
, req
->is_replay());
5121 // put locks in excl mode
5122 newi
->filelock
.set_state(LOCK_EXCL
);
5123 newi
->authlock
.set_state(LOCK_EXCL
);
5124 newi
->xattrlock
.set_state(LOCK_EXCL
);
5127 // make sure this inode gets into the journal
5128 le
->metablob
.add_opened_ino(newi
->ino());
5129 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
5130 ls
->open_files
.push_back(&newi
->item_open_file
);
5132 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5138 void Server::handle_client_symlink(MDRequestRef
& mdr
)
5140 MClientRequest
*req
= mdr
->client_request
;
5141 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5142 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5144 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5145 respond_to_request(mdr
, -EROFS
);
5148 CDir
*dir
= dn
->get_dir();
5149 CInode
*diri
= dir
->get_inode();
5150 rdlocks
.insert(&diri
->authlock
);
5151 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5154 if (!check_access(mdr
, diri
, MAY_WRITE
))
5157 if (!check_fragment_space(mdr
, dir
))
5160 unsigned mode
= S_IFLNK
| 0777;
5161 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
), mode
);
5165 dn
->push_projected_linkage(newi
);
5167 newi
->symlink
= req
->get_path2();
5168 newi
->inode
.size
= newi
->symlink
.length();
5169 newi
->inode
.rstat
.rbytes
= newi
->inode
.size
;
5170 newi
->inode
.rstat
.rfiles
= 1;
5171 newi
->inode
.version
= dn
->pre_dirty();
5172 newi
->inode
.update_backtrace();
5174 newi
->first
= dn
->first
;
5177 mdr
->ls
= mdlog
->get_current_segment();
5178 EUpdate
*le
= new EUpdate(mdlog
, "symlink");
5179 mdlog
->start_entry(le
);
5180 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5181 journal_allocated_inos(mdr
, &le
->metablob
);
5182 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5183 le
->metablob
.add_primary_dentry(dn
, newi
, true, true);
5185 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5194 void Server::handle_client_link(MDRequestRef
& mdr
)
5196 MClientRequest
*req
= mdr
->client_request
;
5198 dout(7) << "handle_client_link " << req
->get_filepath()
5199 << " to " << req
->get_filepath2()
5202 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5204 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5206 CInode
*targeti
= rdlock_path_pin_ref(mdr
, 1, rdlocks
, false);
5207 if (!targeti
) return;
5208 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5209 respond_to_request(mdr
, -EROFS
);
5213 CDir
*dir
= dn
->get_dir();
5214 dout(7) << "handle_client_link link " << dn
->get_name() << " in " << *dir
<< dendl
;
5215 dout(7) << "target is " << *targeti
<< dendl
;
5216 if (targeti
->is_dir()) {
5217 dout(7) << "target is a dir, failing..." << dendl
;
5218 respond_to_request(mdr
, -EINVAL
);
5222 xlocks
.insert(&targeti
->linklock
);
5224 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5227 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
5228 if (!check_access(mdr
, targeti
, MAY_WRITE
))
5231 if (!check_access(mdr
, dir
->get_inode(), MAY_WRITE
))
5234 if (!check_fragment_space(mdr
, dir
))
5239 assert(g_conf
->mds_kill_link_at
!= 1);
5242 if (targeti
->is_auth())
5243 _link_local(mdr
, dn
, targeti
);
5245 _link_remote(mdr
, true, dn
, targeti
);
5249 class C_MDS_link_local_finish
: public ServerLogContext
{
5255 C_MDS_link_local_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ti
,
5256 version_t dnpv_
, version_t tipv_
) :
5257 ServerLogContext(s
, r
), dn(d
), targeti(ti
),
5258 dnpv(dnpv_
), tipv(tipv_
) { }
5259 void finish(int r
) override
{
5261 server
->_link_local_finish(mdr
, dn
, targeti
, dnpv
, tipv
);
5266 void Server::_link_local(MDRequestRef
& mdr
, CDentry
*dn
, CInode
*targeti
)
5268 dout(10) << "_link_local " << *dn
<< " to " << *targeti
<< dendl
;
5270 mdr
->ls
= mdlog
->get_current_segment();
5272 // predirty NEW dentry
5273 version_t dnpv
= dn
->pre_dirty();
5274 version_t tipv
= targeti
->pre_dirty();
5276 // project inode update
5277 inode_t
*pi
= targeti
->project_inode();
5279 pi
->ctime
= mdr
->get_op_stamp();
5284 EUpdate
*le
= new EUpdate(mdlog
, "link_local");
5285 mdlog
->start_entry(le
);
5286 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5287 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, 1); // new dn
5288 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, 0, PREDIRTY_PRIMARY
); // targeti
5289 le
->metablob
.add_remote_dentry(dn
, true, targeti
->ino(), targeti
->d_type()); // new remote
5290 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, targeti
);
5292 // do this after predirty_*, to avoid funky extra dnl arg
5293 dn
->push_projected_linkage(targeti
->ino(), targeti
->d_type());
5295 journal_and_reply(mdr
, targeti
, dn
, le
, new C_MDS_link_local_finish(this, mdr
, dn
, targeti
, dnpv
, tipv
));
5298 void Server::_link_local_finish(MDRequestRef
& mdr
, CDentry
*dn
, CInode
*targeti
,
5299 version_t dnpv
, version_t tipv
)
5301 dout(10) << "_link_local_finish " << *dn
<< " to " << *targeti
<< dendl
;
5303 // link and unlock the NEW dentry
5304 CDentry::linkage_t
*dnl
= dn
->pop_projected_linkage();
5305 if (!dnl
->get_inode())
5306 dn
->link_remote(dnl
, targeti
);
5307 dn
->mark_dirty(dnpv
, mdr
->ls
);
5310 targeti
->pop_and_dirty_projected_inode(mdr
->ls
);
5314 MDRequestRef null_ref
;
5315 mdcache
->send_dentry_link(dn
, null_ref
);
5317 // bump target popularity
5318 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5319 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
5322 respond_to_request(mdr
, 0);
5326 // link / unlink remote
5328 class C_MDS_link_remote_finish
: public ServerLogContext
{
5334 C_MDS_link_remote_finish(Server
*s
, MDRequestRef
& r
, bool i
, CDentry
*d
, CInode
*ti
) :
5335 ServerLogContext(s
, r
), inc(i
), dn(d
), targeti(ti
),
5336 dpv(d
->get_projected_version()) {}
5337 void finish(int r
) override
{
5339 server
->_link_remote_finish(mdr
, inc
, dn
, targeti
, dpv
);
5343 void Server::_link_remote(MDRequestRef
& mdr
, bool inc
, CDentry
*dn
, CInode
*targeti
)
5345 dout(10) << "_link_remote "
5346 << (inc
? "link ":"unlink ")
5347 << *dn
<< " to " << *targeti
<< dendl
;
5349 // 1. send LinkPrepare to dest (journal nlink++ prepare)
5350 mds_rank_t linkauth
= targeti
->authority().first
;
5351 if (mdr
->more()->witnessed
.count(linkauth
) == 0) {
5352 if (mds
->is_cluster_degraded() &&
5353 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(linkauth
)) {
5354 dout(10) << " targeti auth mds." << linkauth
<< " is not active" << dendl
;
5355 if (mdr
->more()->waiting_on_slave
.empty())
5356 mds
->wait_for_active_peer(linkauth
, new C_MDS_RetryRequest(mdcache
, mdr
));
5360 dout(10) << " targeti auth must prepare nlink++/--" << dendl
;
5363 op
= MMDSSlaveRequest::OP_LINKPREP
;
5365 op
= MMDSSlaveRequest::OP_UNLINKPREP
;
5366 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, op
);
5367 targeti
->set_object_info(req
->get_object_info());
5368 req
->op_stamp
= mdr
->get_op_stamp();
5369 mds
->send_message_mds(req
, linkauth
);
5371 assert(mdr
->more()->waiting_on_slave
.count(linkauth
) == 0);
5372 mdr
->more()->waiting_on_slave
.insert(linkauth
);
5375 dout(10) << " targeti auth has prepared nlink++/--" << dendl
;
5377 assert(g_conf
->mds_kill_link_at
!= 2);
5379 mdr
->set_mds_stamp(ceph_clock_now());
5382 mdr
->ls
= mdlog
->get_current_segment();
5383 EUpdate
*le
= new EUpdate(mdlog
, inc
? "link_remote":"unlink_remote");
5384 mdlog
->start_entry(le
);
5385 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5386 if (!mdr
->more()->witnessed
.empty()) {
5387 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
5388 le
->reqid
= mdr
->reqid
;
5389 le
->had_slaves
= true;
5390 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
5395 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, 1);
5396 le
->metablob
.add_remote_dentry(dn
, true, targeti
->ino(), targeti
->d_type()); // new remote
5397 dn
->push_projected_linkage(targeti
->ino(), targeti
->d_type());
5400 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, -1);
5401 mdcache
->journal_cow_dentry(mdr
.get(), &le
->metablob
, dn
);
5402 le
->metablob
.add_null_dentry(dn
, true);
5403 dn
->push_projected_linkage();
5406 journal_and_reply(mdr
, targeti
, dn
, le
, new C_MDS_link_remote_finish(this, mdr
, inc
, dn
, targeti
));
5409 void Server::_link_remote_finish(MDRequestRef
& mdr
, bool inc
,
5410 CDentry
*dn
, CInode
*targeti
,
5413 dout(10) << "_link_remote_finish "
5414 << (inc
? "link ":"unlink ")
5415 << *dn
<< " to " << *targeti
<< dendl
;
5417 assert(g_conf
->mds_kill_link_at
!= 3);
5419 if (!mdr
->more()->witnessed
.empty())
5420 mdcache
->logged_master_update(mdr
->reqid
);
5423 // link the new dentry
5424 CDentry::linkage_t
*dnl
= dn
->pop_projected_linkage();
5425 if (!dnl
->get_inode())
5426 dn
->link_remote(dnl
, targeti
);
5427 dn
->mark_dirty(dpv
, mdr
->ls
);
5429 // unlink main dentry
5430 dn
->get_dir()->unlink_inode(dn
);
5431 dn
->pop_projected_linkage();
5432 dn
->mark_dirty(dn
->get_projected_version(), mdr
->ls
); // dirty old dentry
5437 MDRequestRef null_ref
;
5439 mdcache
->send_dentry_link(dn
, null_ref
);
5441 mdcache
->send_dentry_unlink(dn
, NULL
, null_ref
);
5443 // bump target popularity
5444 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5445 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
5448 respond_to_request(mdr
, 0);
5451 // removing a new dn?
5452 dn
->get_dir()->try_remove_unlinked_dn(dn
);
5456 // remote linking/unlinking
5458 class C_MDS_SlaveLinkPrep
: public ServerLogContext
{
5461 C_MDS_SlaveLinkPrep(Server
*s
, MDRequestRef
& r
, CInode
*t
) :
5462 ServerLogContext(s
, r
), targeti(t
) { }
5463 void finish(int r
) override
{
5465 server
->_logged_slave_link(mdr
, targeti
);
5469 class C_MDS_SlaveLinkCommit
: public ServerContext
{
5473 C_MDS_SlaveLinkCommit(Server
*s
, MDRequestRef
& r
, CInode
*t
) :
5474 ServerContext(s
), mdr(r
), targeti(t
) { }
5475 void finish(int r
) override
{
5476 server
->_commit_slave_link(mdr
, r
, targeti
);
5480 /* This function DOES put the mdr->slave_request before returning*/
5481 void Server::handle_slave_link_prep(MDRequestRef
& mdr
)
5483 dout(10) << "handle_slave_link_prep " << *mdr
5484 << " on " << mdr
->slave_request
->get_object_info()
5487 assert(g_conf
->mds_kill_link_at
!= 4);
5489 CInode
*targeti
= mdcache
->get_inode(mdr
->slave_request
->get_object_info().ino
);
5491 dout(10) << "targeti " << *targeti
<< dendl
;
5492 CDentry
*dn
= targeti
->get_parent_dn();
5493 CDentry::linkage_t
*dnl
= dn
->get_linkage();
5494 assert(dnl
->is_primary());
5496 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
5498 mdr
->auth_pin(targeti
);
5500 //ceph_abort(); // test hack: make sure master can handle a slave that fails to prepare...
5501 assert(g_conf
->mds_kill_link_at
!= 5);
5504 mdr
->ls
= mdlog
->get_current_segment();
5505 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_prep", mdr
->reqid
, mdr
->slave_to_mds
,
5506 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::LINK
);
5507 mdlog
->start_entry(le
);
5509 inode_t
*pi
= dnl
->get_inode()->project_inode();
5511 // update journaled target inode
5513 if (mdr
->slave_request
->get_op() == MMDSSlaveRequest::OP_LINKPREP
) {
5521 link_rollback rollback
;
5522 rollback
.reqid
= mdr
->reqid
;
5523 rollback
.ino
= targeti
->ino();
5524 rollback
.old_ctime
= targeti
->inode
.ctime
; // we hold versionlock xlock; no concorrent projections
5525 const fnode_t
*pf
= targeti
->get_parent_dn()->get_dir()->get_projected_fnode();
5526 rollback
.old_dir_mtime
= pf
->fragstat
.mtime
;
5527 rollback
.old_dir_rctime
= pf
->rstat
.rctime
;
5528 rollback
.was_inc
= inc
;
5529 ::encode(rollback
, le
->rollback
);
5530 mdr
->more()->rollback_bl
= le
->rollback
;
5532 pi
->ctime
= mdr
->get_op_stamp();
5533 pi
->version
= targeti
->pre_dirty();
5535 dout(10) << " projected inode " << pi
<< " v " << pi
->version
<< dendl
;
5538 mdcache
->predirty_journal_parents(mdr
, &le
->commit
, dnl
->get_inode(), 0, PREDIRTY_SHALLOW
|PREDIRTY_PRIMARY
);
5539 mdcache
->journal_dirty_inode(mdr
.get(), &le
->commit
, targeti
);
5541 // set up commit waiter
5542 mdr
->more()->slave_commit
= new C_MDS_SlaveLinkCommit(this, mdr
, targeti
);
5544 mdr
->more()->slave_update_journaled
= true;
5545 submit_mdlog_entry(le
, new C_MDS_SlaveLinkPrep(this, mdr
, targeti
),
5550 void Server::_logged_slave_link(MDRequestRef
& mdr
, CInode
*targeti
)
5552 dout(10) << "_logged_slave_link " << *mdr
5553 << " " << *targeti
<< dendl
;
5555 assert(g_conf
->mds_kill_link_at
!= 6);
5557 // update the target
5558 targeti
->pop_and_dirty_projected_inode(mdr
->ls
);
5562 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5565 mdr
->slave_request
->put();
5566 mdr
->slave_request
= 0;
5569 if (!mdr
->aborted
) {
5570 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
5571 MMDSSlaveRequest::OP_LINKPREPACK
);
5572 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
5574 dout(10) << " abort flag set, finishing" << dendl
;
5575 mdcache
->request_finish(mdr
);
5580 struct C_MDS_CommittedSlave
: public ServerLogContext
{
5581 C_MDS_CommittedSlave(Server
*s
, MDRequestRef
& m
) : ServerLogContext(s
, m
) {}
5582 void finish(int r
) override
{
5583 server
->_committed_slave(mdr
);
5587 void Server::_commit_slave_link(MDRequestRef
& mdr
, int r
, CInode
*targeti
)
5589 dout(10) << "_commit_slave_link " << *mdr
5591 << " " << *targeti
<< dendl
;
5593 assert(g_conf
->mds_kill_link_at
!= 7);
5596 // drop our pins, etc.
5599 // write a commit to the journal
5600 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_commit", mdr
->reqid
, mdr
->slave_to_mds
,
5601 ESlaveUpdate::OP_COMMIT
, ESlaveUpdate::LINK
);
5602 mdlog
->start_entry(le
);
5603 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
5606 do_link_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
);
5610 void Server::_committed_slave(MDRequestRef
& mdr
)
5612 dout(10) << "_committed_slave " << *mdr
<< dendl
;
5614 assert(g_conf
->mds_kill_link_at
!= 8);
5616 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
5617 MMDSSlaveRequest::OP_COMMITTED
);
5618 mds
->send_message_mds(req
, mdr
->slave_to_mds
);
5619 mdcache
->request_finish(mdr
);
5622 struct C_MDS_LoggedLinkRollback
: public ServerLogContext
{
5624 C_MDS_LoggedLinkRollback(Server
*s
, MutationRef
& m
, MDRequestRef
& r
) : ServerLogContext(s
, r
), mut(m
) {}
5625 void finish(int r
) override
{
5626 server
->_link_rollback_finish(mut
, mdr
);
5630 void Server::do_link_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
)
5632 link_rollback rollback
;
5633 bufferlist::iterator p
= rbl
.begin();
5634 ::decode(rollback
, p
);
5636 dout(10) << "do_link_rollback on " << rollback
.reqid
5637 << (rollback
.was_inc
? " inc":" dec")
5638 << " ino " << rollback
.ino
5641 assert(g_conf
->mds_kill_link_at
!= 9);
5643 mdcache
->add_rollback(rollback
.reqid
, master
); // need to finish this update before resolve finishes
5644 assert(mdr
|| mds
->is_resolve());
5646 MutationRef
mut(new MutationImpl(nullptr, utime_t(), rollback
.reqid
));
5647 mut
->ls
= mds
->mdlog
->get_current_segment();
5649 CInode
*in
= mdcache
->get_inode(rollback
.ino
);
5651 dout(10) << " target is " << *in
<< dendl
;
5652 assert(!in
->is_projected()); // live slave request hold versionlock xlock.
5654 inode_t
*pi
= in
->project_inode();
5655 pi
->version
= in
->pre_dirty();
5656 mut
->add_projected_inode(in
);
5658 // parent dir rctime
5659 CDir
*parent
= in
->get_projected_parent_dn()->get_dir();
5660 fnode_t
*pf
= parent
->project_fnode();
5661 mut
->add_projected_fnode(parent
);
5662 pf
->version
= parent
->pre_dirty();
5663 if (pf
->fragstat
.mtime
== pi
->ctime
) {
5664 pf
->fragstat
.mtime
= rollback
.old_dir_mtime
;
5665 if (pf
->rstat
.rctime
== pi
->ctime
)
5666 pf
->rstat
.rctime
= rollback
.old_dir_rctime
;
5667 mut
->add_updated_lock(&parent
->get_inode()->filelock
);
5668 mut
->add_updated_lock(&parent
->get_inode()->nestlock
);
5672 pi
->ctime
= rollback
.old_ctime
;
5673 if (rollback
.was_inc
)
5679 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_rollback", rollback
.reqid
, master
,
5680 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::LINK
);
5681 mdlog
->start_entry(le
);
5682 le
->commit
.add_dir_context(parent
);
5683 le
->commit
.add_dir(parent
, true);
5684 le
->commit
.add_primary_dentry(in
->get_projected_parent_dn(), 0, true);
5686 submit_mdlog_entry(le
, new C_MDS_LoggedLinkRollback(this, mut
, mdr
),
5691 void Server::_link_rollback_finish(MutationRef
& mut
, MDRequestRef
& mdr
)
5693 dout(10) << "_link_rollback_finish" << dendl
;
5695 assert(g_conf
->mds_kill_link_at
!= 10);
5699 mdcache
->request_finish(mdr
);
5701 mdcache
->finish_rollback(mut
->reqid
);
5707 /* This function DOES NOT put the passed message before returning*/
5708 void Server::handle_slave_link_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*m
)
5710 dout(10) << "handle_slave_link_prep_ack " << *mdr
5711 << " " << *m
<< dendl
;
5712 mds_rank_t from
= mds_rank_t(m
->get_source().num());
5714 assert(g_conf
->mds_kill_link_at
!= 11);
5717 mdr
->more()->slaves
.insert(from
);
5720 assert(mdr
->more()->witnessed
.count(from
) == 0);
5721 mdr
->more()->witnessed
.insert(from
);
5722 assert(!m
->is_not_journaled());
5723 mdr
->more()->has_journaled_slaves
= true;
5725 // remove from waiting list
5726 assert(mdr
->more()->waiting_on_slave
.count(from
));
5727 mdr
->more()->waiting_on_slave
.erase(from
);
5729 assert(mdr
->more()->waiting_on_slave
.empty());
5731 dispatch_client_request(mdr
); // go again!
5740 void Server::handle_client_unlink(MDRequestRef
& mdr
)
5742 MClientRequest
*req
= mdr
->client_request
;
5743 client_t client
= mdr
->get_client();
5747 if (req
->get_op() == CEPH_MDS_OP_RMDIR
) rmdir
= true;
5749 if (req
->get_filepath().depth() == 0) {
5750 respond_to_request(mdr
, -EINVAL
);
5755 vector
<CDentry
*> trace
;
5757 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, req
->get_filepath(), &trace
, &in
, MDS_TRAVERSE_FORWARD
);
5761 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
5762 mdcache
->find_ino_peers(req
->get_filepath().get_ino(), new C_MDS_TryFindInode(this, mdr
));
5765 respond_to_request(mdr
, r
);
5768 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5769 respond_to_request(mdr
, -EROFS
);
5773 CDentry
*dn
= trace
[trace
.size()-1];
5775 if (!dn
->is_auth()) {
5776 mdcache
->request_forward(mdr
, dn
->authority().first
);
5780 CInode
*diri
= dn
->get_dir()->get_inode();
5782 CDentry::linkage_t
*dnl
= dn
->get_linkage(client
, mdr
);
5783 assert(!dnl
->is_null());
5786 dout(7) << "handle_client_rmdir on " << *dn
<< dendl
;
5788 dout(7) << "handle_client_unlink on " << *dn
<< dendl
;
5790 dout(7) << "dn links to " << *in
<< dendl
;
5795 // do empty directory checks
5796 if (_dir_is_nonempty_unlocked(mdr
, in
)) {
5797 respond_to_request(mdr
, -ENOTEMPTY
);
5801 dout(7) << "handle_client_unlink on dir " << *in
<< ", returning error" << dendl
;
5802 respond_to_request(mdr
, -EISDIR
);
5808 dout(7) << "handle_client_rmdir on non-dir " << *in
<< ", returning error" << dendl
;
5809 respond_to_request(mdr
, -ENOTDIR
);
5814 // -- create stray dentry? --
5815 CDentry
*straydn
= NULL
;
5816 if (dnl
->is_primary()) {
5817 straydn
= prepare_stray_dentry(mdr
, dnl
->get_inode());
5820 dout(10) << " straydn is " << *straydn
<< dendl
;
5821 } else if (mdr
->straydn
) {
5822 mdr
->unpin(mdr
->straydn
);
5823 mdr
->straydn
= NULL
;
5827 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5829 for (int i
=0; i
<(int)trace
.size()-1; i
++)
5830 rdlocks
.insert(&trace
[i
]->lock
);
5831 xlocks
.insert(&dn
->lock
);
5832 wrlocks
.insert(&diri
->filelock
);
5833 wrlocks
.insert(&diri
->nestlock
);
5834 xlocks
.insert(&in
->linklock
);
5836 wrlocks
.insert(&straydn
->get_dir()->inode
->filelock
);
5837 wrlocks
.insert(&straydn
->get_dir()->inode
->nestlock
);
5838 xlocks
.insert(&straydn
->lock
);
5841 rdlocks
.insert(&in
->filelock
); // to verify it's empty
5842 mds
->locker
->include_snap_rdlocks(rdlocks
, dnl
->get_inode());
5844 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5848 _dir_is_nonempty(mdr
, in
)) {
5849 respond_to_request(mdr
, -ENOTEMPTY
);
5853 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
5854 if (!check_access(mdr
, diri
, MAY_WRITE
))
5859 if (in
->is_dir() && in
->has_subtree_root_dirfrag()) {
5860 // subtree root auths need to be witnesses
5861 set
<mds_rank_t
> witnesses
;
5862 in
->list_replicas(witnesses
);
5863 dout(10) << " witnesses " << witnesses
<< ", have " << mdr
->more()->witnessed
<< dendl
;
5865 for (set
<mds_rank_t
>::iterator p
= witnesses
.begin();
5866 p
!= witnesses
.end();
5868 if (mdr
->more()->witnessed
.count(*p
)) {
5869 dout(10) << " already witnessed by mds." << *p
<< dendl
;
5870 } else if (mdr
->more()->waiting_on_slave
.count(*p
)) {
5871 dout(10) << " already waiting on witness mds." << *p
<< dendl
;
5873 if (!_rmdir_prepare_witness(mdr
, *p
, trace
, straydn
))
5877 if (!mdr
->more()->waiting_on_slave
.empty())
5878 return; // we're waiting for a witness.
5882 if (dnl
->is_remote() && !dnl
->get_inode()->is_auth())
5883 _link_remote(mdr
, false, dn
, dnl
->get_inode());
5885 _unlink_local(mdr
, dn
, straydn
);
5888 class C_MDS_unlink_local_finish
: public ServerLogContext
{
5891 version_t dnpv
; // deleted dentry
5893 C_MDS_unlink_local_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CDentry
*sd
) :
5894 ServerLogContext(s
, r
), dn(d
), straydn(sd
),
5895 dnpv(d
->get_projected_version()) {}
5896 void finish(int r
) override
{
5898 server
->_unlink_local_finish(mdr
, dn
, straydn
, dnpv
);
5902 void Server::_unlink_local(MDRequestRef
& mdr
, CDentry
*dn
, CDentry
*straydn
)
5904 dout(10) << "_unlink_local " << *dn
<< dendl
;
5906 CDentry::linkage_t
*dnl
= dn
->get_projected_linkage();
5907 CInode
*in
= dnl
->get_inode();
5909 SnapRealm
*realm
= in
->find_snaprealm();
5910 snapid_t follows
= realm
->get_newest_seq();
5913 mdr
->ls
= mdlog
->get_current_segment();
5915 // prepare log entry
5916 EUpdate
*le
= new EUpdate(mdlog
, "unlink_local");
5917 mdlog
->start_entry(le
);
5918 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5919 if (!mdr
->more()->witnessed
.empty()) {
5920 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
5921 le
->reqid
= mdr
->reqid
;
5922 le
->had_slaves
= true;
5923 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
5927 assert(dnl
->is_primary());
5928 straydn
->push_projected_linkage(in
);
5929 straydn
->first
= follows
+ 1;
5932 // the unlinked dentry
5935 inode_t
*pi
= in
->project_inode();
5936 dn
->make_path_string(pi
->stray_prior_path
, true);
5937 mdr
->add_projected_inode(in
); // do this _after_ my dn->pre_dirty().. we apply that one manually.
5938 pi
->version
= in
->pre_dirty();
5939 pi
->ctime
= mdr
->get_op_stamp();
5943 in
->state_set(CInode::STATE_ORPHAN
);
5945 if (dnl
->is_primary()) {
5946 // primary link. add stray dentry.
5948 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, -1);
5949 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, straydn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5951 // project snaprealm, too
5952 if (in
->snaprealm
|| follows
+ 1 > in
->get_oldest_snap())
5953 in
->project_past_snaprealm_parent(straydn
->get_dir()->inode
->find_snaprealm());
5955 pi
->update_backtrace();
5956 le
->metablob
.add_primary_dentry(straydn
, in
, true, true);
5958 // remote link. update remote inode.
5959 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_DIR
, -1);
5960 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, 0, PREDIRTY_PRIMARY
);
5961 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, in
);
5964 mdcache
->journal_cow_dentry(mdr
.get(), &le
->metablob
, dn
);
5965 le
->metablob
.add_null_dentry(dn
, true);
5968 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
5969 le
->metablob
.renamed_dirino
= in
->ino();
5972 dn
->push_projected_linkage();
5976 mdcache
->project_subtree_rename(in
, dn
->get_dir(), straydn
->get_dir());
5979 journal_and_reply(mdr
, 0, dn
, le
, new C_MDS_unlink_local_finish(this, mdr
, dn
, straydn
));
5982 void Server::_unlink_local_finish(MDRequestRef
& mdr
,
5983 CDentry
*dn
, CDentry
*straydn
,
5986 dout(10) << "_unlink_local_finish " << *dn
<< dendl
;
5988 if (!mdr
->more()->witnessed
.empty())
5989 mdcache
->logged_master_update(mdr
->reqid
);
5991 // unlink main dentry
5992 dn
->get_dir()->unlink_inode(dn
);
5993 dn
->pop_projected_linkage();
5995 // relink as stray? (i.e. was primary link?)
5996 CInode
*strayin
= NULL
;
5997 bool snap_is_new
= false;
5999 dout(20) << " straydn is " << *straydn
<< dendl
;
6000 CDentry::linkage_t
*straydnl
= straydn
->pop_projected_linkage();
6001 strayin
= straydnl
->get_inode();
6003 snap_is_new
= strayin
->snaprealm
? true : false;
6004 mdcache
->touch_dentry_bottom(straydn
);
6007 dn
->mark_dirty(dnpv
, mdr
->ls
);
6010 if (snap_is_new
) //only new if strayin exists
6011 mdcache
->do_realm_invalidate_and_update_notify(strayin
, CEPH_SNAP_OP_SPLIT
, true);
6013 mdcache
->send_dentry_unlink(dn
, straydn
, mdr
);
6015 // update subtree map?
6016 if (straydn
&& strayin
->is_dir())
6017 mdcache
->adjust_subtree_after_rename(strayin
, dn
->get_dir(), true);
6020 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
6023 respond_to_request(mdr
, 0);
6025 // removing a new dn?
6026 dn
->get_dir()->try_remove_unlinked_dn(dn
);
6029 // respond_to_request() drops locks. So stray reintegration can race with us.
6030 if (straydn
&& !straydn
->get_projected_linkage()->is_null()) {
6031 // Tip off the MDCache that this dentry is a stray that
6032 // might be elegible for purge.
6033 mdcache
->notify_stray(straydn
);
6037 bool Server::_rmdir_prepare_witness(MDRequestRef
& mdr
, mds_rank_t who
, vector
<CDentry
*>& trace
, CDentry
*straydn
)
6039 if (mds
->is_cluster_degraded() &&
6040 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(who
)) {
6041 dout(10) << "_rmdir_prepare_witness mds." << who
<< " is not active" << dendl
;
6042 if (mdr
->more()->waiting_on_slave
.empty())
6043 mds
->wait_for_active_peer(who
, new C_MDS_RetryRequest(mdcache
, mdr
));
6047 dout(10) << "_rmdir_prepare_witness mds." << who
<< dendl
;
6048 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6049 MMDSSlaveRequest::OP_RMDIRPREP
);
6050 req
->srcdnpath
= filepath(trace
.front()->get_dir()->ino());
6051 for (auto dn
: trace
)
6052 req
->srcdnpath
.push_dentry(dn
->name
);
6053 mdcache
->replicate_stray(straydn
, who
, req
->stray
);
6055 req
->op_stamp
= mdr
->get_op_stamp();
6056 mds
->send_message_mds(req
, who
);
6058 assert(mdr
->more()->waiting_on_slave
.count(who
) == 0);
6059 mdr
->more()->waiting_on_slave
.insert(who
);
6063 struct C_MDS_SlaveRmdirPrep
: public ServerLogContext
{
6064 CDentry
*dn
, *straydn
;
6065 C_MDS_SlaveRmdirPrep(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CDentry
*st
)
6066 : ServerLogContext(s
, r
), dn(d
), straydn(st
) {}
6067 void finish(int r
) override
{
6068 server
->_logged_slave_rmdir(mdr
, dn
, straydn
);
6072 struct C_MDS_SlaveRmdirCommit
: public ServerContext
{
6075 C_MDS_SlaveRmdirCommit(Server
*s
, MDRequestRef
& r
, CDentry
*sd
)
6076 : ServerContext(s
), mdr(r
), straydn(sd
) { }
6077 void finish(int r
) override
{
6078 server
->_commit_slave_rmdir(mdr
, r
, straydn
);
6082 void Server::handle_slave_rmdir_prep(MDRequestRef
& mdr
)
6084 dout(10) << "handle_slave_rmdir_prep " << *mdr
6085 << " " << mdr
->slave_request
->srcdnpath
6086 << " to " << mdr
->slave_request
->destdnpath
6089 vector
<CDentry
*> trace
;
6090 filepath
srcpath(mdr
->slave_request
->srcdnpath
);
6091 dout(10) << " src " << srcpath
<< dendl
;
6093 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &trace
, &in
, MDS_TRAVERSE_DISCOVERXLOCK
);
6096 mdcache
->find_ino_peers(srcpath
.get_ino(), new C_MDS_RetryRequest(mdcache
, mdr
),
6101 CDentry
*dn
= trace
[trace
.size()-1];
6102 dout(10) << " dn " << *dn
<< dendl
;
6105 assert(mdr
->straydn
);
6106 CDentry
*straydn
= mdr
->straydn
;
6107 dout(10) << " straydn " << *straydn
<< dendl
;
6109 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
6111 rmdir_rollback rollback
;
6112 rollback
.reqid
= mdr
->reqid
;
6113 rollback
.src_dir
= dn
->get_dir()->dirfrag();
6114 rollback
.src_dname
= dn
->name
;
6115 rollback
.dest_dir
= straydn
->get_dir()->dirfrag();
6116 rollback
.dest_dname
= straydn
->name
;
6117 ::encode(rollback
, mdr
->more()->rollback_bl
);
6118 dout(20) << " rollback is " << mdr
->more()->rollback_bl
.length() << " bytes" << dendl
;
6120 // set up commit waiter
6121 mdr
->more()->slave_commit
= new C_MDS_SlaveRmdirCommit(this, mdr
, straydn
);
6123 if (!in
->has_subtree_root_dirfrag(mds
->get_nodeid())) {
6124 dout(10) << " no auth subtree in " << *in
<< ", skipping journal" << dendl
;
6125 dn
->get_dir()->unlink_inode(dn
);
6126 straydn
->get_dir()->link_primary_inode(straydn
, in
);
6128 assert(straydn
->first
>= in
->first
);
6129 in
->first
= straydn
->first
;
6131 mdcache
->adjust_subtree_after_rename(in
, dn
->get_dir(), false);
6133 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6134 MMDSSlaveRequest::OP_RMDIRPREPACK
);
6135 reply
->mark_not_journaled();
6136 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
6138 // send caps to auth (if we're not already)
6139 if (in
->is_any_caps() && !in
->state_test(CInode::STATE_EXPORTINGCAPS
))
6140 mdcache
->migrator
->export_caps(in
);
6142 mdcache
->touch_dentry_bottom(straydn
); // move stray to end of lru
6144 mdr
->slave_request
->put();
6145 mdr
->slave_request
= 0;
6150 straydn
->push_projected_linkage(in
);
6151 dn
->push_projected_linkage();
6153 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir", mdr
->reqid
, mdr
->slave_to_mds
,
6154 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::RMDIR
);
6155 mdlog
->start_entry(le
);
6156 le
->rollback
= mdr
->more()->rollback_bl
;
6158 le
->commit
.add_dir_context(straydn
->get_dir());
6159 le
->commit
.add_primary_dentry(straydn
, in
, true);
6160 // slave: no need to journal original dentry
6162 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
6163 le
->commit
.renamed_dirino
= in
->ino();
6165 mdcache
->project_subtree_rename(in
, dn
->get_dir(), straydn
->get_dir());
6167 mdr
->more()->slave_update_journaled
= true;
6168 submit_mdlog_entry(le
, new C_MDS_SlaveRmdirPrep(this, mdr
, dn
, straydn
),
6173 void Server::_logged_slave_rmdir(MDRequestRef
& mdr
, CDentry
*dn
, CDentry
*straydn
)
6175 dout(10) << "_logged_slave_rmdir " << *mdr
<< " on " << *dn
<< dendl
;
6177 // update our cache now, so we are consistent with what is in the journal
6178 // when we journal a subtree map
6179 CInode
*in
= dn
->get_linkage()->get_inode();
6180 dn
->get_dir()->unlink_inode(dn
);
6181 straydn
->pop_projected_linkage();
6182 dn
->pop_projected_linkage();
6183 mdcache
->adjust_subtree_after_rename(in
, dn
->get_dir(), true);
6186 mdr
->slave_request
->put();
6187 mdr
->slave_request
= 0;
6190 if (!mdr
->aborted
) {
6191 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6192 MMDSSlaveRequest::OP_RMDIRPREPACK
);
6193 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
6195 dout(10) << " abort flag set, finishing" << dendl
;
6196 mdcache
->request_finish(mdr
);
6200 void Server::handle_slave_rmdir_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
6202 dout(10) << "handle_slave_rmdir_prep_ack " << *mdr
6203 << " " << *ack
<< dendl
;
6205 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
6207 mdr
->more()->slaves
.insert(from
);
6208 mdr
->more()->witnessed
.insert(from
);
6209 if (!ack
->is_not_journaled())
6210 mdr
->more()->has_journaled_slaves
= true;
6212 // remove from waiting list
6213 assert(mdr
->more()->waiting_on_slave
.count(from
));
6214 mdr
->more()->waiting_on_slave
.erase(from
);
6216 if (mdr
->more()->waiting_on_slave
.empty())
6217 dispatch_client_request(mdr
); // go again!
6219 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
6222 void Server::_commit_slave_rmdir(MDRequestRef
& mdr
, int r
, CDentry
*straydn
)
6224 dout(10) << "_commit_slave_rmdir " << *mdr
<< " r=" << r
<< dendl
;
6227 if (mdr
->more()->slave_update_journaled
) {
6228 CInode
*strayin
= straydn
->get_projected_linkage()->get_inode();
6229 if (strayin
&& !strayin
->snaprealm
)
6230 mdcache
->clear_dirty_bits_for_stray(strayin
);
6235 if (mdr
->more()->slave_update_journaled
) {
6236 // write a commit to the journal
6237 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir_commit", mdr
->reqid
,
6238 mdr
->slave_to_mds
, ESlaveUpdate::OP_COMMIT
,
6239 ESlaveUpdate::RMDIR
);
6240 mdlog
->start_entry(le
);
6241 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
6244 _committed_slave(mdr
);
6248 do_rmdir_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
);
6252 struct C_MDS_LoggedRmdirRollback
: public ServerLogContext
{
6256 C_MDS_LoggedRmdirRollback(Server
*s
, MDRequestRef
& m
, metareqid_t mr
, CDentry
*d
, CDentry
*st
)
6257 : ServerLogContext(s
, m
), reqid(mr
), dn(d
), straydn(st
) {}
6258 void finish(int r
) override
{
6259 server
->_rmdir_rollback_finish(mdr
, reqid
, dn
, straydn
);
6263 void Server::do_rmdir_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
)
6265 // unlink the other rollback methods, the rmdir rollback is only
6266 // needed to record the subtree changes in the journal for inode
6267 // replicas who are auth for empty dirfrags. no actual changes to
6268 // the file system are taking place here, so there is no Mutation.
6270 rmdir_rollback rollback
;
6271 bufferlist::iterator p
= rbl
.begin();
6272 ::decode(rollback
, p
);
6274 dout(10) << "do_rmdir_rollback on " << rollback
.reqid
<< dendl
;
6275 mdcache
->add_rollback(rollback
.reqid
, master
); // need to finish this update before resolve finishes
6276 assert(mdr
|| mds
->is_resolve());
6278 CDir
*dir
= mdcache
->get_dirfrag(rollback
.src_dir
);
6280 dir
= mdcache
->get_dirfrag(rollback
.src_dir
.ino
, rollback
.src_dname
);
6282 CDentry
*dn
= dir
->lookup(rollback
.src_dname
);
6284 dout(10) << " dn " << *dn
<< dendl
;
6285 dir
= mdcache
->get_dirfrag(rollback
.dest_dir
);
6287 CDentry
*straydn
= dir
->lookup(rollback
.dest_dname
);
6289 dout(10) << " straydn " << *dn
<< dendl
;
6290 CInode
*in
= straydn
->get_linkage()->get_inode();
6292 if (mdr
&& !mdr
->more()->slave_update_journaled
) {
6293 assert(!in
->has_subtree_root_dirfrag(mds
->get_nodeid()));
6295 straydn
->get_dir()->unlink_inode(straydn
);
6296 dn
->get_dir()->link_primary_inode(dn
, in
);
6298 mdcache
->adjust_subtree_after_rename(in
, straydn
->get_dir(), false);
6300 mdcache
->request_finish(mdr
);
6301 mdcache
->finish_rollback(rollback
.reqid
);
6305 dn
->push_projected_linkage(in
);
6306 straydn
->push_projected_linkage();
6308 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir_rollback", rollback
.reqid
, master
,
6309 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::RMDIR
);
6310 mdlog
->start_entry(le
);
6312 le
->commit
.add_dir_context(dn
->get_dir());
6313 le
->commit
.add_primary_dentry(dn
, in
, true);
6314 // slave: no need to journal straydn
6316 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
6317 le
->commit
.renamed_dirino
= in
->ino();
6319 mdcache
->project_subtree_rename(in
, straydn
->get_dir(), dn
->get_dir());
6321 submit_mdlog_entry(le
,
6322 new C_MDS_LoggedRmdirRollback(this, mdr
,rollback
.reqid
,
6328 void Server::_rmdir_rollback_finish(MDRequestRef
& mdr
, metareqid_t reqid
, CDentry
*dn
, CDentry
*straydn
)
6330 dout(10) << "_rmdir_rollback_finish " << reqid
<< dendl
;
6332 straydn
->get_dir()->unlink_inode(straydn
);
6333 dn
->pop_projected_linkage();
6334 straydn
->pop_projected_linkage();
6336 CInode
*in
= dn
->get_linkage()->get_inode();
6337 mdcache
->adjust_subtree_after_rename(in
, straydn
->get_dir(), true);
6338 if (mds
->is_resolve()) {
6339 CDir
*root
= mdcache
->get_subtree_root(straydn
->get_dir());
6340 mdcache
->try_trim_non_auth_subtree(root
);
6344 mdcache
->request_finish(mdr
);
6346 mdcache
->finish_rollback(reqid
);
6350 /** _dir_is_nonempty[_unlocked]
6352 * check if a directory is non-empty (i.e. we can rmdir it).
6354 * the unlocked varient this is a fastpath check. we can't really be
6355 * sure until we rdlock the filelock.
6357 bool Server::_dir_is_nonempty_unlocked(MDRequestRef
& mdr
, CInode
*in
)
6359 dout(10) << "dir_is_nonempty_unlocked " << *in
<< dendl
;
6360 assert(in
->is_auth());
6362 if (in
->snaprealm
&& in
->snaprealm
->srnode
.snaps
.size())
6363 return true; // in a snapshot!
6366 in
->get_dirfrags(ls
);
6367 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6369 // is the frag obviously non-empty?
6370 if (dir
->is_auth()) {
6371 if (dir
->get_projected_fnode()->fragstat
.size()) {
6372 dout(10) << "dir_is_nonempty_unlocked dirstat has "
6373 << dir
->get_projected_fnode()->fragstat
.size() << " items " << *dir
<< dendl
;
6382 bool Server::_dir_is_nonempty(MDRequestRef
& mdr
, CInode
*in
)
6384 dout(10) << "dir_is_nonempty " << *in
<< dendl
;
6385 assert(in
->is_auth());
6386 assert(in
->filelock
.can_read(mdr
->get_client()));
6388 frag_info_t dirstat
;
6389 version_t dirstat_version
= in
->get_projected_inode()->dirstat
.version
;
6392 in
->get_dirfrags(ls
);
6393 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6395 const fnode_t
*pf
= dir
->get_projected_fnode();
6396 if (pf
->fragstat
.size()) {
6397 dout(10) << "dir_is_nonempty dirstat has "
6398 << pf
->fragstat
.size() << " items " << *dir
<< dendl
;
6402 if (pf
->accounted_fragstat
.version
== dirstat_version
)
6403 dirstat
.add(pf
->accounted_fragstat
);
6405 dirstat
.add(pf
->fragstat
);
6408 return dirstat
.size() != in
->get_projected_inode()->dirstat
.size();
6412 // ======================================================
6415 class C_MDS_rename_finish
: public ServerLogContext
{
6420 C_MDS_rename_finish(Server
*s
, MDRequestRef
& r
,
6421 CDentry
*sdn
, CDentry
*ddn
, CDentry
*stdn
) :
6422 ServerLogContext(s
, r
),
6423 srcdn(sdn
), destdn(ddn
), straydn(stdn
) { }
6424 void finish(int r
) override
{
6426 server
->_rename_finish(mdr
, srcdn
, destdn
, straydn
);
6431 /** handle_client_rename
6433 * rename master is the destdn auth. this is because cached inodes
6434 * must remain connected. thus, any replica of srci, must also
6435 * replicate destdn, and possibly straydn, so that srci (and
6436 * destdn->inode) remain connected during the rename.
6438 * to do this, we freeze srci, then master (destdn auth) verifies that
6439 * all other nodes have also replciated destdn and straydn. note that
6440 * destdn replicas need not also replicate srci. this only works when
6443 * This function takes responsibility for the passed mdr.
6445 void Server::handle_client_rename(MDRequestRef
& mdr
)
6447 MClientRequest
*req
= mdr
->client_request
;
6448 dout(7) << "handle_client_rename " << *req
<< dendl
;
6450 filepath destpath
= req
->get_filepath();
6451 filepath srcpath
= req
->get_filepath2();
6452 if (destpath
.depth() == 0 || srcpath
.depth() == 0) {
6453 respond_to_request(mdr
, -EINVAL
);
6456 const string
&destname
= destpath
.last_dentry();
6458 vector
<CDentry
*>& srctrace
= mdr
->dn
[1];
6459 vector
<CDentry
*>& desttrace
= mdr
->dn
[0];
6461 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
6463 CDentry
*destdn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, true, false, true);
6464 if (!destdn
) return;
6465 dout(10) << " destdn " << *destdn
<< dendl
;
6466 if (mdr
->snapid
!= CEPH_NOSNAP
) {
6467 respond_to_request(mdr
, -EROFS
);
6470 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
6471 CDir
*destdir
= destdn
->get_dir();
6472 assert(destdir
->is_auth());
6474 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &srctrace
, NULL
, MDS_TRAVERSE_DISCOVER
);
6479 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
6480 mdcache
->find_ino_peers(srcpath
.get_ino(), new C_MDS_TryFindInode(this, mdr
));
6482 dout(10) << "FAIL on error " << r
<< dendl
;
6483 respond_to_request(mdr
, r
);
6488 assert(!srctrace
.empty());
6489 CDentry
*srcdn
= srctrace
[srctrace
.size()-1];
6490 dout(10) << " srcdn " << *srcdn
<< dendl
;
6491 if (srcdn
->last
!= CEPH_NOSNAP
) {
6492 respond_to_request(mdr
, -EROFS
);
6495 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
6496 CInode
*srci
= srcdnl
->get_inode();
6497 dout(10) << " srci " << *srci
<< dendl
;
6500 if (!destdnl
->is_null()) {
6501 //dout(10) << "dest dn exists " << *destdn << dendl;
6502 oldin
= mdcache
->get_dentry_inode(destdn
, mdr
, true);
6504 dout(10) << " oldin " << *oldin
<< dendl
;
6506 // mv /some/thing /to/some/existing_other_thing
6507 if (oldin
->is_dir() && !srci
->is_dir()) {
6508 respond_to_request(mdr
, -EISDIR
);
6511 if (!oldin
->is_dir() && srci
->is_dir()) {
6512 respond_to_request(mdr
, -ENOTDIR
);
6516 // non-empty dir? do trivial fast unlocked check, do another check later with read locks
6517 if (oldin
->is_dir() && _dir_is_nonempty_unlocked(mdr
, oldin
)) {
6518 respond_to_request(mdr
, -ENOTEMPTY
);
6521 if (srci
== oldin
&& !srcdn
->get_dir()->inode
->is_stray()) {
6522 respond_to_request(mdr
, 0); // no-op. POSIX makes no sense.
6527 // -- some sanity checks --
6529 // src+dest traces _must_ share a common ancestor for locking to prevent orphans
6530 if (destpath
.get_ino() != srcpath
.get_ino() &&
6531 !(req
->get_source().is_mds() &&
6532 MDS_INO_IS_MDSDIR(srcpath
.get_ino()))) { // <-- mds 'rename' out of stray dir is ok!
6533 CInode
*srcbase
= srctrace
[0]->get_dir()->get_inode();
6534 CInode
*destbase
= desttrace
[0]->get_dir()->get_inode();
6535 // ok, extend srctrace toward root until it is an ancestor of desttrace.
6536 while (srcbase
!= destbase
&&
6537 !srcbase
->is_projected_ancestor_of(destbase
)) {
6538 CDentry
*pdn
= srcbase
->get_projected_parent_dn();
6539 srctrace
.insert(srctrace
.begin(), pdn
);
6540 dout(10) << "rename prepending srctrace with " << *pdn
<< dendl
;
6541 srcbase
= pdn
->get_dir()->get_inode();
6544 // then, extend destpath until it shares the same parent inode as srcpath.
6545 while (destbase
!= srcbase
) {
6546 CDentry
*pdn
= destbase
->get_projected_parent_dn();
6547 desttrace
.insert(desttrace
.begin(), pdn
);
6548 rdlocks
.insert(&pdn
->lock
);
6549 dout(10) << "rename prepending desttrace with " << *pdn
<< dendl
;
6550 destbase
= pdn
->get_dir()->get_inode();
6552 dout(10) << "rename src and dest traces now share common ancestor " << *destbase
<< dendl
;
6556 if (srcdn
->get_dir() == destdir
&& srcdn
->name
== destname
) {
6557 dout(7) << "rename src=dest, noop" << dendl
;
6558 respond_to_request(mdr
, 0);
6562 // dest a child of src?
6563 // e.g. mv /usr /usr/foo
6564 CDentry
*pdn
= destdir
->inode
->get_projected_parent_dn();
6567 dout(7) << "cannot rename item to be a child of itself" << dendl
;
6568 respond_to_request(mdr
, -EINVAL
);
6571 pdn
= pdn
->get_dir()->inode
->parent
;
6574 // is this a stray migration, reintegration or merge? (sanity checks!)
6575 if (mdr
->reqid
.name
.is_mds() &&
6576 !(MDS_INO_IS_MDSDIR(srcpath
.get_ino()) &&
6577 MDS_INO_IS_MDSDIR(destpath
.get_ino())) &&
6578 !(destdnl
->is_remote() &&
6579 destdnl
->get_remote_ino() == srci
->ino())) {
6580 respond_to_request(mdr
, -EINVAL
); // actually, this won't reply, but whatev.
6584 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
6585 (srcdnl
->is_primary() || destdnl
->is_primary()));
6587 dout(10) << " this is a link merge" << dendl
;
6589 // -- create stray dentry? --
6590 CDentry
*straydn
= NULL
;
6591 if (destdnl
->is_primary() && !linkmerge
) {
6592 straydn
= prepare_stray_dentry(mdr
, destdnl
->get_inode());
6595 dout(10) << " straydn is " << *straydn
<< dendl
;
6596 } else if (mdr
->straydn
) {
6597 mdr
->unpin(mdr
->straydn
);
6598 mdr
->straydn
= NULL
;
6601 // -- prepare witness list --
6603 * NOTE: we use _all_ replicas as witnesses.
6604 * this probably isn't totally necessary (esp for file renames),
6605 * but if/when we change that, we have to make sure rejoin is
6606 * sufficiently robust to handle strong rejoins from survivors
6607 * with totally wrong dentry->inode linkage.
6608 * (currently, it can ignore rename effects, because the resolve
6609 * stage will sort them out.)
6611 set
<mds_rank_t
> witnesses
= mdr
->more()->extra_witnesses
;
6612 if (srcdn
->is_auth())
6613 srcdn
->list_replicas(witnesses
);
6615 witnesses
.insert(srcdn
->authority().first
);
6616 if (srcdnl
->is_remote() && !srci
->is_auth())
6617 witnesses
.insert(srci
->authority().first
);
6618 destdn
->list_replicas(witnesses
);
6619 if (destdnl
->is_remote() && !oldin
->is_auth())
6620 witnesses
.insert(oldin
->authority().first
);
6621 dout(10) << " witnesses " << witnesses
<< ", have " << mdr
->more()->witnessed
<< dendl
;
6625 map
<SimpleLock
*, mds_rank_t
> remote_wrlocks
;
6627 // srctrace items. this mirrors locks taken in rdlock_path_xlock_dentry
6628 for (int i
=0; i
<(int)srctrace
.size(); i
++)
6629 rdlocks
.insert(&srctrace
[i
]->lock
);
6630 xlocks
.insert(&srcdn
->lock
);
6631 mds_rank_t srcdirauth
= srcdn
->get_dir()->authority().first
;
6632 if (srcdirauth
!= mds
->get_nodeid()) {
6633 dout(10) << " will remote_wrlock srcdir scatterlocks on mds." << srcdirauth
<< dendl
;
6634 remote_wrlocks
[&srcdn
->get_dir()->inode
->filelock
] = srcdirauth
;
6635 remote_wrlocks
[&srcdn
->get_dir()->inode
->nestlock
] = srcdirauth
;
6637 rdlocks
.insert(&srci
->dirfragtreelock
);
6639 wrlocks
.insert(&srcdn
->get_dir()->inode
->filelock
);
6640 wrlocks
.insert(&srcdn
->get_dir()->inode
->nestlock
);
6642 mds
->locker
->include_snap_rdlocks(rdlocks
, srcdn
->get_dir()->inode
);
6646 wrlocks
.insert(&straydn
->get_dir()->inode
->filelock
);
6647 wrlocks
.insert(&straydn
->get_dir()->inode
->nestlock
);
6648 xlocks
.insert(&straydn
->lock
);
6651 // xlock versionlock on dentries if there are witnesses.
6652 // replicas can't see projected dentry linkages, and will get
6653 // confused if we try to pipeline things.
6654 if (!witnesses
.empty()) {
6655 // take xlock on all projected ancestor dentries for srcdn and destdn.
6656 // this ensures the srcdn and destdn can be traversed to by the witnesses.
6657 for (int i
= 0; i
<(int)srctrace
.size(); i
++) {
6658 if (srctrace
[i
]->is_auth() && srctrace
[i
]->is_projected())
6659 xlocks
.insert(&srctrace
[i
]->versionlock
);
6661 for (int i
=0; i
<(int)desttrace
.size(); i
++) {
6662 if (desttrace
[i
]->is_auth() && desttrace
[i
]->is_projected())
6663 xlocks
.insert(&desttrace
[i
]->versionlock
);
6665 // xlock srci and oldin's primary dentries, so witnesses can call
6666 // open_remote_ino() with 'want_locked=true' when the srcdn or destdn
6668 if (srcdnl
->is_remote())
6669 xlocks
.insert(&srci
->get_projected_parent_dn()->lock
);
6670 if (destdnl
->is_remote())
6671 xlocks
.insert(&oldin
->get_projected_parent_dn()->lock
);
6674 // we need to update srci's ctime. xlock its least contended lock to do that...
6675 xlocks
.insert(&srci
->linklock
);
6677 // xlock oldin (for nlink--)
6679 xlocks
.insert(&oldin
->linklock
);
6680 if (oldin
->is_dir())
6681 rdlocks
.insert(&oldin
->filelock
);
6683 if (srcdnl
->is_primary() && srci
->is_dir())
6684 // FIXME: this should happen whenever we are renamning between
6685 // realms, regardless of the file type
6686 // FIXME: If/when this changes, make sure to update the
6687 // "allowance" in handle_slave_rename_prep
6688 xlocks
.insert(&srci
->snaplock
); // FIXME: an auth bcast could be sufficient?
6690 rdlocks
.insert(&srci
->snaplock
);
6692 CInode
*auth_pin_freeze
= !srcdn
->is_auth() && srcdnl
->is_primary() ? srci
: NULL
;
6693 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
,
6694 &remote_wrlocks
, auth_pin_freeze
))
6697 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
6698 if (!check_access(mdr
, srcdn
->get_dir()->get_inode(), MAY_WRITE
))
6701 if (!check_access(mdr
, destdn
->get_dir()->get_inode(), MAY_WRITE
))
6704 if (!check_fragment_space(mdr
, destdn
->get_dir()))
6707 if (!check_access(mdr
, srci
, MAY_WRITE
))
6711 // with read lock, really verify oldin is empty
6714 _dir_is_nonempty(mdr
, oldin
)) {
6715 respond_to_request(mdr
, -ENOTEMPTY
);
6719 /* project_past_snaprealm_parent() will do this job
6721 // moving between snaprealms?
6722 if (srcdnl->is_primary() && srci->is_multiversion() && !srci->snaprealm) {
6723 SnapRealm *srcrealm = srci->find_snaprealm();
6724 SnapRealm *destrealm = destdn->get_dir()->inode->find_snaprealm();
6725 if (srcrealm != destrealm &&
6726 (srcrealm->get_newest_seq() + 1 > srcdn->first ||
6727 destrealm->get_newest_seq() + 1 > srcdn->first)) {
6728 dout(10) << " renaming between snaprealms, creating snaprealm for " << *srci << dendl;
6729 mdcache->snaprealm_create(mdr, srci);
6735 assert(g_conf
->mds_kill_rename_at
!= 1);
6737 // -- open all srcdn inode frags, if any --
6738 // we need these open so that auth can properly delegate from inode to dirfrags
6739 // after the inode is _ours_.
6740 if (srcdnl
->is_primary() &&
6741 !srcdn
->is_auth() &&
6743 dout(10) << "srci is remote dir, setting stickydirs and opening all frags" << dendl
;
6744 mdr
->set_stickydirs(srci
);
6747 srci
->dirfragtree
.get_leaves(frags
);
6748 for (list
<frag_t
>::iterator p
= frags
.begin();
6751 CDir
*dir
= srci
->get_dirfrag(*p
);
6753 dout(10) << " opening " << *p
<< " under " << *srci
<< dendl
;
6754 mdcache
->open_remote_dirfrag(srci
, *p
, new C_MDS_RetryRequest(mdcache
, mdr
));
6760 // -- prepare witnesses --
6762 // do srcdn auth last
6763 mds_rank_t last
= MDS_RANK_NONE
;
6764 if (!srcdn
->is_auth()) {
6765 last
= srcdn
->authority().first
;
6766 mdr
->more()->srcdn_auth_mds
= last
;
6767 // ask auth of srci to mark srci as ambiguous auth if more than two MDS
6768 // are involved in the rename operation.
6769 if (srcdnl
->is_primary() && !mdr
->more()->is_ambiguous_auth
) {
6770 dout(10) << " preparing ambiguous auth for srci" << dendl
;
6771 assert(mdr
->more()->is_remote_frozen_authpin
);
6772 assert(mdr
->more()->rename_inode
== srci
);
6773 _rename_prepare_witness(mdr
, last
, witnesses
, srctrace
, desttrace
, straydn
);
6778 for (set
<mds_rank_t
>::iterator p
= witnesses
.begin();
6779 p
!= witnesses
.end();
6781 if (*p
== last
) continue; // do it last!
6782 if (mdr
->more()->witnessed
.count(*p
)) {
6783 dout(10) << " already witnessed by mds." << *p
<< dendl
;
6784 } else if (mdr
->more()->waiting_on_slave
.count(*p
)) {
6785 dout(10) << " already waiting on witness mds." << *p
<< dendl
;
6787 if (!_rename_prepare_witness(mdr
, *p
, witnesses
, srctrace
, desttrace
, straydn
))
6791 if (!mdr
->more()->waiting_on_slave
.empty())
6792 return; // we're waiting for a witness.
6794 if (last
!= MDS_RANK_NONE
&& mdr
->more()->witnessed
.count(last
) == 0) {
6795 dout(10) << " preparing last witness (srcdn auth)" << dendl
;
6796 assert(mdr
->more()->waiting_on_slave
.count(last
) == 0);
6797 _rename_prepare_witness(mdr
, last
, witnesses
, srctrace
, desttrace
, straydn
);
6801 // test hack: bail after slave does prepare, so we can verify it's _live_ rollback.
6802 if (!mdr
->more()->slaves
.empty() && !srci
->is_dir())
6803 assert(g_conf
->mds_kill_rename_at
!= 3);
6804 if (!mdr
->more()->slaves
.empty() && srci
->is_dir())
6805 assert(g_conf
->mds_kill_rename_at
!= 4);
6807 // -- declare now --
6808 mdr
->set_mds_stamp(ceph_clock_now());
6810 // -- prepare journal entry --
6811 mdr
->ls
= mdlog
->get_current_segment();
6812 EUpdate
*le
= new EUpdate(mdlog
, "rename");
6813 mdlog
->start_entry(le
);
6814 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
6815 if (!mdr
->more()->witnessed
.empty()) {
6816 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
6818 le
->reqid
= mdr
->reqid
;
6819 le
->had_slaves
= true;
6821 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
6822 // no need to send frozen auth pin to recovring auth MDS of srci
6823 mdr
->more()->is_remote_frozen_authpin
= false;
6826 _rename_prepare(mdr
, &le
->metablob
, &le
->client_map
, srcdn
, destdn
, straydn
);
6827 if (le
->client_map
.length())
6828 le
->cmapv
= mds
->sessionmap
.get_projected();
6830 // -- commit locally --
6831 C_MDS_rename_finish
*fin
= new C_MDS_rename_finish(this, mdr
, srcdn
, destdn
, straydn
);
6833 journal_and_reply(mdr
, srci
, destdn
, le
, fin
);
6837 void Server::_rename_finish(MDRequestRef
& mdr
, CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
6839 dout(10) << "_rename_finish " << *mdr
<< dendl
;
6841 if (!mdr
->more()->witnessed
.empty())
6842 mdcache
->logged_master_update(mdr
->reqid
);
6845 _rename_apply(mdr
, srcdn
, destdn
, straydn
);
6847 mdcache
->send_dentry_link(destdn
, mdr
);
6849 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
6850 CInode
*in
= destdnl
->get_inode();
6851 bool need_eval
= mdr
->more()->cap_imports
.count(in
);
6853 // test hack: test slave commit
6854 if (!mdr
->more()->slaves
.empty() && !in
->is_dir())
6855 assert(g_conf
->mds_kill_rename_at
!= 5);
6856 if (!mdr
->more()->slaves
.empty() && in
->is_dir())
6857 assert(g_conf
->mds_kill_rename_at
!= 6);
6860 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), srcdn
->get_dir(), META_POP_IWR
);
6861 if (destdnl
->is_remote() && in
->is_auth())
6862 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
6864 // did we import srci? if so, explicitly ack that import that, before we unlock and reply.
6866 assert(g_conf
->mds_kill_rename_at
!= 7);
6869 respond_to_request(mdr
, 0);
6872 mds
->locker
->eval(in
, CEPH_CAP_LOCKS
, true);
6875 // respond_to_request() drops locks. So stray reintegration can race with us.
6876 if (straydn
&& !straydn
->get_projected_linkage()->is_null()) {
6877 mdcache
->notify_stray(straydn
);
6885 bool Server::_rename_prepare_witness(MDRequestRef
& mdr
, mds_rank_t who
, set
<mds_rank_t
> &witnesse
,
6886 vector
<CDentry
*>& srctrace
, vector
<CDentry
*>& dsttrace
, CDentry
*straydn
)
6888 if (mds
->is_cluster_degraded() &&
6889 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(who
)) {
6890 dout(10) << "_rename_prepare_witness mds." << who
<< " is not active" << dendl
;
6891 if (mdr
->more()->waiting_on_slave
.empty())
6892 mds
->wait_for_active_peer(who
, new C_MDS_RetryRequest(mdcache
, mdr
));
6896 dout(10) << "_rename_prepare_witness mds." << who
<< dendl
;
6897 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6898 MMDSSlaveRequest::OP_RENAMEPREP
);
6900 req
->srcdnpath
= filepath(srctrace
.front()->get_dir()->ino());
6901 for (auto dn
: srctrace
)
6902 req
->srcdnpath
.push_dentry(dn
->name
);
6903 req
->destdnpath
= filepath(dsttrace
.front()->get_dir()->ino());
6904 for (auto dn
: dsttrace
)
6905 req
->destdnpath
.push_dentry(dn
->name
);
6907 mdcache
->replicate_stray(straydn
, who
, req
->stray
);
6909 req
->srcdn_auth
= mdr
->more()->srcdn_auth_mds
;
6911 // srcdn auth will verify our current witness list is sufficient
6912 req
->witnesses
= witnesse
;
6914 req
->op_stamp
= mdr
->get_op_stamp();
6915 mds
->send_message_mds(req
, who
);
6917 assert(mdr
->more()->waiting_on_slave
.count(who
) == 0);
6918 mdr
->more()->waiting_on_slave
.insert(who
);
6922 version_t
Server::_rename_prepare_import(MDRequestRef
& mdr
, CDentry
*srcdn
, bufferlist
*client_map_bl
)
6924 version_t oldpv
= mdr
->more()->inode_import_v
;
6926 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
6929 bufferlist::iterator blp
= mdr
->more()->inode_import
.begin();
6932 ::decode(mdr
->more()->imported_client_map
, blp
);
6933 ::encode(mdr
->more()->imported_client_map
, *client_map_bl
,
6934 mds
->mdsmap
->get_up_features());
6935 prepare_force_open_sessions(mdr
->more()->imported_client_map
, mdr
->more()->sseq_map
);
6937 list
<ScatterLock
*> updated_scatterlocks
;
6938 mdcache
->migrator
->decode_import_inode(srcdn
, blp
, srcdn
->authority().first
, mdr
->ls
,
6939 mdr
->more()->cap_imports
, updated_scatterlocks
);
6941 // hack: force back to !auth and clean, temporarily
6942 srcdnl
->get_inode()->state_clear(CInode::STATE_AUTH
);
6943 srcdnl
->get_inode()->mark_clean();
6948 bool Server::_need_force_journal(CInode
*diri
, bool empty
)
6951 diri
->get_dirfrags(ls
);
6953 bool force_journal
= false;
6955 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6956 if ((*p
)->is_subtree_root() && (*p
)->get_dir_auth().first
== mds
->get_nodeid()) {
6957 dout(10) << " frag " << (*p
)->get_frag() << " is auth subtree dirfrag, will force journal" << dendl
;
6958 force_journal
= true;
6961 dout(20) << " frag " << (*p
)->get_frag() << " is not auth subtree dirfrag" << dendl
;
6964 // see if any children of our frags are auth subtrees.
6965 list
<CDir
*> subtrees
;
6966 mdcache
->list_subtrees(subtrees
);
6967 dout(10) << " subtrees " << subtrees
<< " frags " << ls
<< dendl
;
6968 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6970 for (list
<CDir
*>::iterator q
= subtrees
.begin(); q
!= subtrees
.end(); ++q
) {
6971 if (dir
->contains(*q
)) {
6972 if ((*q
)->get_dir_auth().first
== mds
->get_nodeid()) {
6973 dout(10) << " frag " << (*p
)->get_frag() << " contains (maybe) auth subtree, will force journal "
6975 force_journal
= true;
6978 dout(20) << " frag " << (*p
)->get_frag() << " contains but isn't auth for " << **q
<< dendl
;
6980 dout(20) << " frag " << (*p
)->get_frag() << " does not contain " << **q
<< dendl
;
6986 return force_journal
;
6989 void Server::_rename_prepare(MDRequestRef
& mdr
,
6990 EMetaBlob
*metablob
, bufferlist
*client_map_bl
,
6991 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
6993 dout(10) << "_rename_prepare " << *mdr
<< " " << *srcdn
<< " " << *destdn
<< dendl
;
6995 dout(10) << " straydn " << *straydn
<< dendl
;
6997 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
6998 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
6999 CInode
*srci
= srcdnl
->get_inode();
7000 CInode
*oldin
= destdnl
->get_inode();
7002 // primary+remote link merge?
7003 bool linkmerge
= (srci
== destdnl
->get_inode() &&
7004 (srcdnl
->is_primary() || destdnl
->is_primary()));
7005 bool silent
= srcdn
->get_dir()->inode
->is_stray();
7007 bool force_journal_dest
= false;
7008 if (srci
->is_dir() && !destdn
->is_auth()) {
7009 if (srci
->is_auth()) {
7010 // if we are auth for srci and exporting it, force journal because journal replay needs
7011 // the source inode to create auth subtrees.
7012 dout(10) << " we are exporting srci, will force journal destdn" << dendl
;
7013 force_journal_dest
= true;
7015 force_journal_dest
= _need_force_journal(srci
, false);
7018 bool force_journal_stray
= false;
7019 if (oldin
&& oldin
->is_dir() && straydn
&& !straydn
->is_auth())
7020 force_journal_stray
= _need_force_journal(oldin
, true);
7023 dout(10) << " merging remote and primary links to the same inode" << dendl
;
7025 dout(10) << " reintegrating stray; will avoid changing nlink or dir mtime" << dendl
;
7026 if (force_journal_dest
)
7027 dout(10) << " forcing journal destdn because we (will) have auth subtrees nested beneath it" << dendl
;
7028 if (force_journal_stray
)
7029 dout(10) << " forcing journal straydn because we (will) have auth subtrees nested beneath it" << dendl
;
7031 if (srci
->is_dir() && (destdn
->is_auth() || force_journal_dest
)) {
7032 dout(10) << " noting renamed dir ino " << srci
->ino() << " in metablob" << dendl
;
7033 metablob
->renamed_dirino
= srci
->ino();
7034 } else if (oldin
&& oldin
->is_dir() && force_journal_stray
) {
7035 dout(10) << " noting rename target dir " << oldin
->ino() << " in metablob" << dendl
;
7036 metablob
->renamed_dirino
= oldin
->ino();
7040 inode_t
*pi
= 0; // renamed inode
7041 inode_t
*tpi
= 0; // target/overwritten inode
7045 if (destdnl
->is_primary()) {
7046 assert(straydn
); // moving to straydn.
7047 // link--, and move.
7048 if (destdn
->is_auth()) {
7049 tpi
= oldin
->project_inode(); //project_snaprealm
7050 tpi
->version
= straydn
->pre_dirty(tpi
->version
);
7051 tpi
->update_backtrace();
7053 straydn
->push_projected_linkage(oldin
);
7054 } else if (destdnl
->is_remote()) {
7056 if (oldin
->is_auth()) {
7057 tpi
= oldin
->project_inode();
7058 tpi
->version
= oldin
->pre_dirty();
7064 if (srcdnl
->is_remote()) {
7067 if (destdn
->is_auth())
7068 mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty();
7069 destdn
->push_projected_linkage(srcdnl
->get_remote_ino(), srcdnl
->get_remote_d_type());
7071 if (srci
->is_auth()) {
7072 pi
= srci
->project_inode();
7073 pi
->version
= srci
->pre_dirty();
7076 dout(10) << " will merge remote onto primary link" << dendl
;
7077 if (destdn
->is_auth()) {
7078 pi
= oldin
->project_inode();
7079 pi
->version
= mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty(oldin
->inode
.version
);
7083 if (destdn
->is_auth()) {
7085 if (srcdn
->is_auth())
7086 oldpv
= srci
->get_projected_version();
7088 oldpv
= _rename_prepare_import(mdr
, srcdn
, client_map_bl
);
7090 // note which dirfrags have child subtrees in the journal
7091 // event, so that we can open those (as bounds) during replay.
7092 if (srci
->is_dir()) {
7094 srci
->get_dirfrags(ls
);
7095 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
7097 if (!dir
->is_auth())
7098 metablob
->renamed_dir_frags
.push_back(dir
->get_frag());
7100 dout(10) << " noting renamed dir open frags " << metablob
->renamed_dir_frags
<< dendl
;
7103 pi
= srci
->project_inode(); // project snaprealm if srcdnl->is_primary
7104 // & srcdnl->snaprealm
7105 pi
->version
= mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty(oldpv
);
7106 pi
->update_backtrace();
7108 destdn
->push_projected_linkage(srci
);
7112 if (srcdn
->is_auth())
7113 mdr
->more()->pvmap
[srcdn
] = srcdn
->pre_dirty();
7114 srcdn
->push_projected_linkage(); // push null linkage
7118 pi
->ctime
= mdr
->get_op_stamp();
7124 tpi
->ctime
= mdr
->get_op_stamp();
7126 destdn
->make_path_string(tpi
->stray_prior_path
, true);
7128 if (tpi
->nlink
== 0)
7129 oldin
->state_set(CInode::STATE_ORPHAN
);
7133 // prepare nesting, mtime updates
7134 int predirty_dir
= silent
? 0:PREDIRTY_DIR
;
7136 // guarantee stray dir is processed first during journal replay. unlink the old inode,
7137 // then link the source inode to destdn
7138 if (destdnl
->is_primary()) {
7140 if (straydn
->is_auth()) {
7141 metablob
->add_dir_context(straydn
->get_dir());
7142 metablob
->add_dir(straydn
->get_dir(), true);
7147 if (destdn
->is_auth() && !destdnl
->is_null()) {
7148 mdcache
->predirty_journal_parents(mdr
, metablob
, oldin
, destdn
->get_dir(),
7149 (destdnl
->is_primary() ? PREDIRTY_PRIMARY
:0)|predirty_dir
, -1);
7150 if (destdnl
->is_primary()) {
7152 mdcache
->predirty_journal_parents(mdr
, metablob
, oldin
, straydn
->get_dir(),
7153 PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
7158 int predirty_primary
= (srcdnl
->is_primary() && srcdn
->get_dir() != destdn
->get_dir()) ? PREDIRTY_PRIMARY
:0;
7159 int flags
= predirty_dir
| predirty_primary
;
7160 if (srcdn
->is_auth())
7161 mdcache
->predirty_journal_parents(mdr
, metablob
, srci
, srcdn
->get_dir(), PREDIRTY_SHALLOW
|flags
, -1);
7162 if (destdn
->is_auth())
7163 mdcache
->predirty_journal_parents(mdr
, metablob
, srci
, destdn
->get_dir(), flags
, 1);
7165 SnapRealm
*src_realm
= srci
->find_snaprealm();
7166 SnapRealm
*dest_realm
= destdn
->get_dir()->inode
->find_snaprealm();
7167 snapid_t next_dest_snap
= dest_realm
->get_newest_seq() + 1;
7169 // add it all to the metablob
7172 if (destdnl
->is_primary()) {
7174 if (destdn
->is_auth()) {
7175 // project snaprealm, too
7176 if (oldin
->snaprealm
|| dest_realm
->get_newest_seq() + 1 > oldin
->get_oldest_snap())
7177 oldin
->project_past_snaprealm_parent(straydn
->get_dir()->inode
->find_snaprealm());
7178 straydn
->first
= MAX(oldin
->first
, next_dest_snap
);
7179 metablob
->add_primary_dentry(straydn
, oldin
, true, true);
7180 } else if (force_journal_stray
) {
7181 dout(10) << " forced journaling straydn " << *straydn
<< dendl
;
7182 metablob
->add_dir_context(straydn
->get_dir());
7183 metablob
->add_primary_dentry(straydn
, oldin
, true);
7185 } else if (destdnl
->is_remote()) {
7186 if (oldin
->is_auth()) {
7188 metablob
->add_dir_context(oldin
->get_projected_parent_dir());
7189 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, oldin
->get_projected_parent_dn(),
7190 CEPH_NOSNAP
, 0, destdnl
);
7191 metablob
->add_primary_dentry(oldin
->get_projected_parent_dn(), oldin
, true);
7197 if (srcdnl
->is_remote()) {
7199 if (destdn
->is_auth() && !destdnl
->is_null())
7200 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7202 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7204 if (destdn
->is_auth())
7205 metablob
->add_remote_dentry(destdn
, true, srcdnl
->get_remote_ino(), srcdnl
->get_remote_d_type());
7206 if (srci
->get_projected_parent_dn()->is_auth()) { // it's remote
7207 metablob
->add_dir_context(srci
->get_projected_parent_dir());
7208 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, srci
->get_projected_parent_dn(), CEPH_NOSNAP
, 0, srcdnl
);
7209 metablob
->add_primary_dentry(srci
->get_projected_parent_dn(), srci
, true);
7212 if (destdn
->is_auth() && !destdnl
->is_null())
7213 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7215 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7217 if (destdn
->is_auth())
7218 metablob
->add_primary_dentry(destdn
, destdnl
->get_inode(), true, true);
7220 } else if (srcdnl
->is_primary()) {
7221 // project snap parent update?
7222 if (destdn
->is_auth() && src_realm
!= dest_realm
&&
7223 (srci
->snaprealm
|| src_realm
->get_newest_seq() + 1 > srci
->get_oldest_snap()))
7224 srci
->project_past_snaprealm_parent(dest_realm
);
7226 if (destdn
->is_auth() && !destdnl
->is_null())
7227 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7229 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7231 if (destdn
->is_auth())
7232 metablob
->add_primary_dentry(destdn
, srci
, true, true);
7233 else if (force_journal_dest
) {
7234 dout(10) << " forced journaling destdn " << *destdn
<< dendl
;
7235 metablob
->add_dir_context(destdn
->get_dir());
7236 metablob
->add_primary_dentry(destdn
, srci
, true);
7237 if (srcdn
->is_auth() && srci
->is_dir()) {
7238 // journal new subtrees root dirfrags
7240 srci
->get_dirfrags(ls
);
7241 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
7244 metablob
->add_dir(dir
, true);
7251 if (srcdn
->is_auth()) {
7252 dout(10) << " journaling srcdn " << *srcdn
<< dendl
;
7253 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, srcdn
, CEPH_NOSNAP
, 0, srcdnl
);
7254 // also journal the inode in case we need do slave rename rollback. It is Ok to add
7255 // both primary and NULL dentries. Because during journal replay, null dentry is
7256 // processed after primary dentry.
7257 if (srcdnl
->is_primary() && !srci
->is_dir() && !destdn
->is_auth())
7258 metablob
->add_primary_dentry(srcdn
, srci
, true);
7259 metablob
->add_null_dentry(srcdn
, true);
7261 dout(10) << " NOT journaling srcdn " << *srcdn
<< dendl
;
7263 // make renamed inode first track the dn
7264 if (srcdnl
->is_primary() && destdn
->is_auth())
7265 srci
->first
= destdn
->first
;
7267 if (oldin
&& oldin
->is_dir()) {
7269 mdcache
->project_subtree_rename(oldin
, destdn
->get_dir(), straydn
->get_dir());
7272 mdcache
->project_subtree_rename(srci
, srcdn
->get_dir(), destdn
->get_dir());
7277 void Server::_rename_apply(MDRequestRef
& mdr
, CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7279 dout(10) << "_rename_apply " << *mdr
<< " " << *srcdn
<< " " << *destdn
<< dendl
;
7280 dout(10) << " pvs " << mdr
->more()->pvmap
<< dendl
;
7282 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
7283 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
7285 CInode
*oldin
= destdnl
->get_inode();
7287 // primary+remote link merge?
7288 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
7289 (srcdnl
->is_primary() || destdnl
->is_primary()));
7293 if (destdnl
->is_primary()) {
7295 dout(10) << "straydn is " << *straydn
<< dendl
;
7296 destdn
->get_dir()->unlink_inode(destdn
, false);
7298 straydn
->pop_projected_linkage();
7299 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7300 assert(!straydn
->is_projected()); // no other projected
7302 mdcache
->touch_dentry_bottom(straydn
); // drop dn as quickly as possible.
7305 if (destdn
->is_auth()) {
7306 bool hadrealm
= (oldin
->snaprealm
? true : false);
7307 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7308 if (oldin
->snaprealm
&& !hadrealm
)
7309 mdcache
->do_realm_invalidate_and_update_notify(oldin
, CEPH_SNAP_OP_SPLIT
);
7311 // FIXME this snaprealm is not filled out correctly
7312 //oldin->open_snaprealm(); might be sufficient..
7314 } else if (destdnl
->is_remote()) {
7315 destdn
->get_dir()->unlink_inode(destdn
, false);
7316 if (oldin
->is_auth())
7317 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7321 // unlink src before we relink it at dest
7322 CInode
*in
= srcdnl
->get_inode();
7325 bool srcdn_was_remote
= srcdnl
->is_remote();
7326 srcdn
->get_dir()->unlink_inode(srcdn
);
7329 if (srcdn_was_remote
) {
7332 destdnl
= destdn
->pop_projected_linkage();
7333 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7334 assert(!destdn
->is_projected()); // no other projected
7336 destdn
->link_remote(destdnl
, in
);
7337 if (destdn
->is_auth())
7338 destdn
->mark_dirty(mdr
->more()->pvmap
[destdn
], mdr
->ls
);
7341 in
->pop_and_dirty_projected_inode(mdr
->ls
);
7343 dout(10) << "merging remote onto primary link" << dendl
;
7344 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7348 dout(10) << "merging primary onto remote link" << dendl
;
7349 destdn
->get_dir()->unlink_inode(destdn
, false);
7351 destdnl
= destdn
->pop_projected_linkage();
7352 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7353 assert(!destdn
->is_projected()); // no other projected
7355 // srcdn inode import?
7356 if (!srcdn
->is_auth() && destdn
->is_auth()) {
7357 assert(mdr
->more()->inode_import
.length() > 0);
7359 map
<client_t
,Capability::Import
> imported_caps
;
7361 // finish cap imports
7362 finish_force_open_sessions(mdr
->more()->imported_client_map
, mdr
->more()->sseq_map
);
7363 if (mdr
->more()->cap_imports
.count(destdnl
->get_inode())) {
7364 mdcache
->migrator
->finish_import_inode_caps(destdnl
->get_inode(),
7365 mdr
->more()->srcdn_auth_mds
, true,
7366 mdr
->more()->cap_imports
[destdnl
->get_inode()],
7370 mdr
->more()->inode_import
.clear();
7371 ::encode(imported_caps
, mdr
->more()->inode_import
);
7373 /* hack: add an auth pin for each xlock we hold. These were
7374 * remote xlocks previously but now they're local and
7375 * we're going to try and unpin when we xlock_finish. */
7376 for (set
<SimpleLock
*>::iterator i
= mdr
->xlocks
.begin();
7377 i
!= mdr
->xlocks
.end();
7379 if ((*i
)->get_parent() == destdnl
->get_inode() &&
7380 !(*i
)->is_locallock())
7381 mds
->locker
->xlock_import(*i
);
7383 // hack: fix auth bit
7384 in
->state_set(CInode::STATE_AUTH
);
7386 mdr
->clear_ambiguous_auth();
7389 if (destdn
->is_auth()) {
7390 in
->pop_and_dirty_projected_inode(mdr
->ls
);
7393 // FIXME: fix up snaprealm!
7398 if (srcdn
->is_auth())
7399 srcdn
->mark_dirty(mdr
->more()->pvmap
[srcdn
], mdr
->ls
);
7400 srcdn
->pop_projected_linkage();
7401 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7402 assert(!srcdn
->is_projected()); // no other projected
7404 // apply remaining projected inodes (nested)
7407 // update subtree map?
7408 if (destdnl
->is_primary() && in
->is_dir())
7409 mdcache
->adjust_subtree_after_rename(in
, srcdn
->get_dir(), true);
7411 if (straydn
&& oldin
->is_dir())
7412 mdcache
->adjust_subtree_after_rename(oldin
, destdn
->get_dir(), true);
7414 // removing a new dn?
7415 if (srcdn
->is_auth())
7416 srcdn
->get_dir()->try_remove_unlinked_dn(srcdn
);
7424 class C_MDS_SlaveRenamePrep
: public ServerLogContext
{
7425 CDentry
*srcdn
, *destdn
, *straydn
;
7427 C_MDS_SlaveRenamePrep(Server
*s
, MDRequestRef
& m
, CDentry
*sr
, CDentry
*de
, CDentry
*st
) :
7428 ServerLogContext(s
, m
), srcdn(sr
), destdn(de
), straydn(st
) {}
7429 void finish(int r
) override
{
7430 server
->_logged_slave_rename(mdr
, srcdn
, destdn
, straydn
);
7434 class C_MDS_SlaveRenameCommit
: public ServerContext
{
7436 CDentry
*srcdn
, *destdn
, *straydn
;
7438 C_MDS_SlaveRenameCommit(Server
*s
, MDRequestRef
& m
, CDentry
*sr
, CDentry
*de
, CDentry
*st
) :
7439 ServerContext(s
), mdr(m
), srcdn(sr
), destdn(de
), straydn(st
) {}
7440 void finish(int r
) override
{
7441 server
->_commit_slave_rename(mdr
, r
, srcdn
, destdn
, straydn
);
7445 class C_MDS_SlaveRenameSessionsFlushed
: public ServerContext
{
7448 C_MDS_SlaveRenameSessionsFlushed(Server
*s
, MDRequestRef
& r
) :
7449 ServerContext(s
), mdr(r
) {}
7450 void finish(int r
) override
{
7451 server
->_slave_rename_sessions_flushed(mdr
);
7455 /* This function DOES put the mdr->slave_request before returning*/
7456 void Server::handle_slave_rename_prep(MDRequestRef
& mdr
)
7458 dout(10) << "handle_slave_rename_prep " << *mdr
7459 << " " << mdr
->slave_request
->srcdnpath
7460 << " to " << mdr
->slave_request
->destdnpath
7463 if (mdr
->slave_request
->is_interrupted()) {
7464 dout(10) << " slave request interrupted, sending noop reply" << dendl
;
7465 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_RENAMEPREPACK
);
7466 reply
->mark_interrupted();
7467 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7468 mdr
->slave_request
->put();
7469 mdr
->slave_request
= 0;
7474 filepath
destpath(mdr
->slave_request
->destdnpath
);
7475 dout(10) << " dest " << destpath
<< dendl
;
7476 vector
<CDentry
*> trace
;
7477 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, destpath
, &trace
, NULL
, MDS_TRAVERSE_DISCOVERXLOCK
);
7480 mdcache
->find_ino_peers(destpath
.get_ino(), new C_MDS_RetryRequest(mdcache
, mdr
),
7484 assert(r
== 0); // we shouldn't get an error here!
7486 CDentry
*destdn
= trace
[trace
.size()-1];
7487 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
7488 dout(10) << " destdn " << *destdn
<< dendl
;
7492 filepath
srcpath(mdr
->slave_request
->srcdnpath
);
7493 dout(10) << " src " << srcpath
<< dendl
;
7494 CInode
*srci
= nullptr;
7495 r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &trace
, &srci
, MDS_TRAVERSE_DISCOVERXLOCK
);
7499 // srcpath must not point to a null dentry
7500 assert(srci
!= nullptr);
7502 CDentry
*srcdn
= trace
[trace
.size()-1];
7503 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
7504 dout(10) << " srcdn " << *srcdn
<< dendl
;
7509 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
7510 (srcdnl
->is_primary() || destdnl
->is_primary()));
7511 CDentry
*straydn
= mdr
->straydn
;
7512 if (destdnl
->is_primary() && !linkmerge
)
7515 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
7516 mdr
->more()->srcdn_auth_mds
= srcdn
->authority().first
;
7518 // set up commit waiter (early, to clean up any freezing etc we do)
7519 if (!mdr
->more()->slave_commit
)
7520 mdr
->more()->slave_commit
= new C_MDS_SlaveRenameCommit(this, mdr
, srcdn
, destdn
, straydn
);
7523 if (srcdn
->is_auth()) {
7524 set
<mds_rank_t
> srcdnrep
;
7525 srcdn
->list_replicas(srcdnrep
);
7527 bool reply_witness
= false;
7528 if (srcdnl
->is_primary() && !srcdnl
->get_inode()->state_test(CInode::STATE_AMBIGUOUSAUTH
)) {
7531 // - avoid conflicting lock state changes
7532 // - avoid concurrent updates to the inode
7533 // (this could also be accomplished with the versionlock)
7534 int allowance
= 2; // 1 for the mdr auth_pin, 1 for the link lock
7535 allowance
+= srcdnl
->get_inode()->is_dir(); // for the snap lock
7536 dout(10) << " freezing srci " << *srcdnl
->get_inode() << " with allowance " << allowance
<< dendl
;
7537 bool frozen_inode
= srcdnl
->get_inode()->freeze_inode(allowance
);
7539 // unfreeze auth pin after freezing the inode to avoid queueing waiters
7540 if (srcdnl
->get_inode()->is_frozen_auth_pin())
7541 mdr
->unfreeze_auth_pin();
7543 if (!frozen_inode
) {
7544 srcdnl
->get_inode()->add_waiter(CInode::WAIT_FROZEN
, new C_MDS_RetryRequest(mdcache
, mdr
));
7549 * set ambiguous auth for srci
7550 * NOTE: we don't worry about ambiguous cache expire as we do
7551 * with subtree migrations because all slaves will pin
7552 * srcdn->get_inode() for duration of this rename.
7554 mdr
->set_ambiguous_auth(srcdnl
->get_inode());
7556 // just mark the source inode as ambiguous auth if more than two MDS are involved.
7557 // the master will send another OP_RENAMEPREP slave request later.
7558 if (mdr
->slave_request
->witnesses
.size() > 1) {
7559 dout(10) << " set srci ambiguous auth; providing srcdn replica list" << dendl
;
7560 reply_witness
= true;
7563 // make sure bystanders have received all lock related messages
7564 for (set
<mds_rank_t
>::iterator p
= srcdnrep
.begin(); p
!= srcdnrep
.end(); ++p
) {
7565 if (*p
== mdr
->slave_to_mds
||
7566 (mds
->is_cluster_degraded() &&
7567 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(*p
)))
7569 MMDSSlaveRequest
*notify
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
7570 MMDSSlaveRequest::OP_RENAMENOTIFY
);
7571 mds
->send_message_mds(notify
, *p
);
7572 mdr
->more()->waiting_on_slave
.insert(*p
);
7575 // make sure clients have received all cap related messages
7576 set
<client_t
> export_client_set
;
7577 mdcache
->migrator
->get_export_client_set(srcdnl
->get_inode(), export_client_set
);
7579 MDSGatherBuilder
gather(g_ceph_context
);
7580 flush_client_sessions(export_client_set
, gather
);
7581 if (gather
.has_subs()) {
7582 mdr
->more()->waiting_on_slave
.insert(MDS_RANK_NONE
);
7583 gather
.set_finisher(new C_MDS_SlaveRenameSessionsFlushed(this, mdr
));
7588 // is witness list sufficient?
7589 for (set
<mds_rank_t
>::iterator p
= srcdnrep
.begin(); p
!= srcdnrep
.end(); ++p
) {
7590 if (*p
== mdr
->slave_to_mds
||
7591 mdr
->slave_request
->witnesses
.count(*p
)) continue;
7592 dout(10) << " witness list insufficient; providing srcdn replica list" << dendl
;
7593 reply_witness
= true;
7597 if (reply_witness
) {
7598 assert(!srcdnrep
.empty());
7599 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
7600 MMDSSlaveRequest::OP_RENAMEPREPACK
);
7601 reply
->witnesses
.swap(srcdnrep
);
7602 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7603 mdr
->slave_request
->put();
7604 mdr
->slave_request
= 0;
7607 dout(10) << " witness list sufficient: includes all srcdn replicas" << dendl
;
7608 if (!mdr
->more()->waiting_on_slave
.empty()) {
7609 dout(10) << " still waiting for rename notify acks from "
7610 << mdr
->more()->waiting_on_slave
<< dendl
;
7613 } else if (srcdnl
->is_primary() && srcdn
->authority() != destdn
->authority()) {
7614 // set ambiguous auth for srci on witnesses
7615 mdr
->set_ambiguous_auth(srcdnl
->get_inode());
7618 // encode everything we'd need to roll this back... basically, just the original state.
7619 rename_rollback rollback
;
7621 rollback
.reqid
= mdr
->reqid
;
7623 rollback
.orig_src
.dirfrag
= srcdn
->get_dir()->dirfrag();
7624 rollback
.orig_src
.dirfrag_old_mtime
= srcdn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7625 rollback
.orig_src
.dirfrag_old_rctime
= srcdn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7626 rollback
.orig_src
.dname
= srcdn
->name
;
7627 if (srcdnl
->is_primary())
7628 rollback
.orig_src
.ino
= srcdnl
->get_inode()->ino();
7630 assert(srcdnl
->is_remote());
7631 rollback
.orig_src
.remote_ino
= srcdnl
->get_remote_ino();
7632 rollback
.orig_src
.remote_d_type
= srcdnl
->get_remote_d_type();
7635 rollback
.orig_dest
.dirfrag
= destdn
->get_dir()->dirfrag();
7636 rollback
.orig_dest
.dirfrag_old_mtime
= destdn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7637 rollback
.orig_dest
.dirfrag_old_rctime
= destdn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7638 rollback
.orig_dest
.dname
= destdn
->name
;
7639 if (destdnl
->is_primary())
7640 rollback
.orig_dest
.ino
= destdnl
->get_inode()->ino();
7641 else if (destdnl
->is_remote()) {
7642 rollback
.orig_dest
.remote_ino
= destdnl
->get_remote_ino();
7643 rollback
.orig_dest
.remote_d_type
= destdnl
->get_remote_d_type();
7647 rollback
.stray
.dirfrag
= straydn
->get_dir()->dirfrag();
7648 rollback
.stray
.dirfrag_old_mtime
= straydn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7649 rollback
.stray
.dirfrag_old_rctime
= straydn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7650 rollback
.stray
.dname
= straydn
->name
;
7652 ::encode(rollback
, mdr
->more()->rollback_bl
);
7653 dout(20) << " rollback is " << mdr
->more()->rollback_bl
.length() << " bytes" << dendl
;
7656 mdr
->ls
= mdlog
->get_current_segment();
7657 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_prep", mdr
->reqid
, mdr
->slave_to_mds
,
7658 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::RENAME
);
7659 mdlog
->start_entry(le
);
7660 le
->rollback
= mdr
->more()->rollback_bl
;
7662 bufferlist blah
; // inode import data... obviously not used if we're the slave
7663 _rename_prepare(mdr
, &le
->commit
, &blah
, srcdn
, destdn
, straydn
);
7665 if (le
->commit
.empty()) {
7666 dout(10) << " empty metablob, skipping journal" << dendl
;
7667 mdlog
->cancel_entry(le
);
7669 _logged_slave_rename(mdr
, srcdn
, destdn
, straydn
);
7671 mdr
->more()->slave_update_journaled
= true;
7672 submit_mdlog_entry(le
, new C_MDS_SlaveRenamePrep(this, mdr
, srcdn
, destdn
, straydn
),
7678 void Server::_logged_slave_rename(MDRequestRef
& mdr
,
7679 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7681 dout(10) << "_logged_slave_rename " << *mdr
<< dendl
;
7684 MMDSSlaveRequest
*reply
= NULL
;
7685 if (!mdr
->aborted
) {
7686 reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_RENAMEPREPACK
);
7687 if (!mdr
->more()->slave_update_journaled
)
7688 reply
->mark_not_journaled();
7691 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
7692 CDentry::linkage_t
*destdnl
= NULL
;
7693 //CDentry::linkage_t *straydnl = straydn ? straydn->get_linkage() : 0;
7696 if (srcdn
->is_auth() && srcdnl
->is_primary()) {
7697 // set export bounds for CInode::encode_export()
7699 if (srcdnl
->get_inode()->is_dir()) {
7700 srcdnl
->get_inode()->get_dirfrags(bounds
);
7701 for (list
<CDir
*>::iterator p
= bounds
.begin(); p
!= bounds
.end(); ++p
)
7702 (*p
)->state_set(CDir::STATE_EXPORTBOUND
);
7705 map
<client_t
,entity_inst_t
> exported_client_map
;
7707 mdcache
->migrator
->encode_export_inode(srcdnl
->get_inode(), inodebl
,
7708 exported_client_map
);
7710 for (list
<CDir
*>::iterator p
= bounds
.begin(); p
!= bounds
.end(); ++p
)
7711 (*p
)->state_clear(CDir::STATE_EXPORTBOUND
);
7714 ::encode(exported_client_map
, reply
->inode_export
, mds
->mdsmap
->get_up_features());
7715 reply
->inode_export
.claim_append(inodebl
);
7716 reply
->inode_export_v
= srcdnl
->get_inode()->inode
.version
;
7719 // remove mdr auth pin
7720 mdr
->auth_unpin(srcdnl
->get_inode());
7721 mdr
->more()->is_inode_exporter
= true;
7723 if (srcdnl
->get_inode()->is_dirty())
7724 srcdnl
->get_inode()->mark_clean();
7726 dout(10) << " exported srci " << *srcdnl
->get_inode() << dendl
;
7730 _rename_apply(mdr
, srcdn
, destdn
, straydn
);
7732 destdnl
= destdn
->get_linkage();
7735 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), srcdn
->get_dir(), META_POP_IWR
);
7736 if (destdnl
->get_inode() && destdnl
->get_inode()->is_auth())
7737 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), destdnl
->get_inode(),
7741 mdr
->slave_request
->put();
7742 mdr
->slave_request
= 0;
7746 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7748 assert(mdr
->aborted
);
7749 dout(10) << " abort flag set, finishing" << dendl
;
7750 mdcache
->request_finish(mdr
);
7754 void Server::_commit_slave_rename(MDRequestRef
& mdr
, int r
,
7755 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7757 dout(10) << "_commit_slave_rename " << *mdr
<< " r=" << r
<< dendl
;
7759 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
7761 list
<MDSInternalContextBase
*> finished
;
7763 // unfreeze+singleauth inode
7764 // hmm, do i really need to delay this?
7765 if (mdr
->more()->is_inode_exporter
) {
7767 CInode
*in
= destdnl
->get_inode();
7770 // we exported, clear out any xlocks that we moved to another MDS
7771 set
<SimpleLock
*>::iterator i
= mdr
->xlocks
.begin();
7772 while (i
!= mdr
->xlocks
.end()) {
7773 SimpleLock
*lock
= *i
++;
7775 // we only care about xlocks on the exported inode
7776 if (lock
->get_parent() == in
&&
7777 !lock
->is_locallock())
7778 mds
->locker
->xlock_export(lock
, mdr
.get());
7781 map
<client_t
,Capability::Import
> peer_imported
;
7782 bufferlist::iterator bp
= mdr
->more()->inode_import
.begin();
7783 ::decode(peer_imported
, bp
);
7785 dout(10) << " finishing inode export on " << *destdnl
->get_inode() << dendl
;
7786 mdcache
->migrator
->finish_export_inode(destdnl
->get_inode(),
7787 mdr
->get_mds_stamp(),
7788 mdr
->slave_to_mds
, peer_imported
, finished
);
7789 mds
->queue_waiters(finished
); // this includes SINGLEAUTH waiters.
7792 assert(destdnl
->get_inode()->is_frozen_inode());
7793 destdnl
->get_inode()->unfreeze_inode(finished
);
7797 if (mdr
->more()->is_ambiguous_auth
) {
7798 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
7799 mdr
->more()->is_ambiguous_auth
= false;
7802 if (straydn
&& mdr
->more()->slave_update_journaled
) {
7803 CInode
*strayin
= straydn
->get_projected_linkage()->get_inode();
7804 if (strayin
&& !strayin
->snaprealm
)
7805 mdcache
->clear_dirty_bits_for_stray(strayin
);
7808 mds
->queue_waiters(finished
);
7811 if (mdr
->more()->slave_update_journaled
) {
7812 // write a commit to the journal
7813 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_commit", mdr
->reqid
,
7814 mdr
->slave_to_mds
, ESlaveUpdate::OP_COMMIT
,
7815 ESlaveUpdate::RENAME
);
7816 mdlog
->start_entry(le
);
7817 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
7820 _committed_slave(mdr
);
7825 // rollback_bl may be empty if we froze the inode but had to provide an expanded
7826 // witness list from the master, and they failed before we tried prep again.
7827 if (mdr
->more()->rollback_bl
.length()) {
7828 if (mdr
->more()->is_inode_exporter
) {
7829 dout(10) << " reversing inode export of " << *destdnl
->get_inode() << dendl
;
7830 destdnl
->get_inode()->abort_export();
7832 if (mdcache
->is_ambiguous_slave_update(mdr
->reqid
, mdr
->slave_to_mds
)) {
7833 mdcache
->remove_ambiguous_slave_update(mdr
->reqid
, mdr
->slave_to_mds
);
7834 // rollback but preserve the slave request
7835 do_rename_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
, false);
7836 mdr
->more()->rollback_bl
.clear();
7838 do_rename_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
, true);
7840 dout(10) << " rollback_bl empty, not rollback back rename (master failed after getting extra witnesses?)" << dendl
;
7842 if (mdr
->more()->is_ambiguous_auth
) {
7843 if (srcdn
->is_auth())
7844 mdr
->more()->rename_inode
->unfreeze_inode(finished
);
7846 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
7847 mdr
->more()->is_ambiguous_auth
= false;
7849 mds
->queue_waiters(finished
);
7850 mdcache
->request_finish(mdr
);
7855 void _rollback_repair_dir(MutationRef
& mut
, CDir
*dir
, rename_rollback::drec
&r
, utime_t ctime
,
7856 bool isdir
, int linkunlink
, nest_info_t
&rstat
)
7859 pf
= dir
->project_fnode();
7860 mut
->add_projected_fnode(dir
);
7861 pf
->version
= dir
->pre_dirty();
7864 pf
->fragstat
.nsubdirs
+= linkunlink
;
7866 pf
->fragstat
.nfiles
+= linkunlink
;
7869 pf
->rstat
.rbytes
+= linkunlink
* rstat
.rbytes
;
7870 pf
->rstat
.rfiles
+= linkunlink
* rstat
.rfiles
;
7871 pf
->rstat
.rsubdirs
+= linkunlink
* rstat
.rsubdirs
;
7872 pf
->rstat
.rsnaprealms
+= linkunlink
* rstat
.rsnaprealms
;
7874 if (pf
->fragstat
.mtime
== ctime
) {
7875 pf
->fragstat
.mtime
= r
.dirfrag_old_mtime
;
7876 if (pf
->rstat
.rctime
== ctime
)
7877 pf
->rstat
.rctime
= r
.dirfrag_old_rctime
;
7879 mut
->add_updated_lock(&dir
->get_inode()->filelock
);
7880 mut
->add_updated_lock(&dir
->get_inode()->nestlock
);
7883 struct C_MDS_LoggedRenameRollback
: public ServerLogContext
{
7890 C_MDS_LoggedRenameRollback(Server
*s
, MutationRef
& m
, MDRequestRef
& r
,
7891 CDentry
*sd
, version_t pv
, CDentry
*dd
,
7892 CDentry
*st
, bool f
) :
7893 ServerLogContext(s
, r
), mut(m
), srcdn(sd
), srcdnpv(pv
), destdn(dd
),
7894 straydn(st
), finish_mdr(f
) {}
7895 void finish(int r
) override
{
7896 server
->_rename_rollback_finish(mut
, mdr
, srcdn
, srcdnpv
,
7897 destdn
, straydn
, finish_mdr
);
7901 void Server::do_rename_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
,
7904 rename_rollback rollback
;
7905 bufferlist::iterator p
= rbl
.begin();
7906 ::decode(rollback
, p
);
7908 dout(10) << "do_rename_rollback on " << rollback
.reqid
<< dendl
;
7909 // need to finish this update before sending resolve to claim the subtree
7910 mdcache
->add_rollback(rollback
.reqid
, master
);
7912 MutationRef
mut(new MutationImpl(nullptr, utime_t(), rollback
.reqid
));
7913 mut
->ls
= mds
->mdlog
->get_current_segment();
7915 CDentry
*srcdn
= NULL
;
7916 CDir
*srcdir
= mdcache
->get_dirfrag(rollback
.orig_src
.dirfrag
);
7918 srcdir
= mdcache
->get_dirfrag(rollback
.orig_src
.dirfrag
.ino
, rollback
.orig_src
.dname
);
7920 dout(10) << " srcdir " << *srcdir
<< dendl
;
7921 srcdn
= srcdir
->lookup(rollback
.orig_src
.dname
);
7923 dout(10) << " srcdn " << *srcdn
<< dendl
;
7924 assert(srcdn
->get_linkage()->is_null());
7926 dout(10) << " srcdn not found" << dendl
;
7928 dout(10) << " srcdir not found" << dendl
;
7930 CDentry
*destdn
= NULL
;
7931 CDir
*destdir
= mdcache
->get_dirfrag(rollback
.orig_dest
.dirfrag
);
7933 destdir
= mdcache
->get_dirfrag(rollback
.orig_dest
.dirfrag
.ino
, rollback
.orig_dest
.dname
);
7935 dout(10) << " destdir " << *destdir
<< dendl
;
7936 destdn
= destdir
->lookup(rollback
.orig_dest
.dname
);
7938 dout(10) << " destdn " << *destdn
<< dendl
;
7940 dout(10) << " destdn not found" << dendl
;
7942 dout(10) << " destdir not found" << dendl
;
7945 if (rollback
.orig_src
.ino
) {
7946 in
= mdcache
->get_inode(rollback
.orig_src
.ino
);
7947 if (in
&& in
->is_dir())
7948 assert(srcdn
&& destdn
);
7950 in
= mdcache
->get_inode(rollback
.orig_src
.remote_ino
);
7952 CDir
*straydir
= NULL
;
7953 CDentry
*straydn
= NULL
;
7954 if (rollback
.stray
.dirfrag
.ino
) {
7955 straydir
= mdcache
->get_dirfrag(rollback
.stray
.dirfrag
);
7957 dout(10) << "straydir " << *straydir
<< dendl
;
7958 straydn
= straydir
->lookup(rollback
.stray
.dname
);
7960 dout(10) << " straydn " << *straydn
<< dendl
;
7961 assert(straydn
->get_linkage()->is_primary());
7963 dout(10) << " straydn not found" << dendl
;
7965 dout(10) << "straydir not found" << dendl
;
7968 CInode
*target
= NULL
;
7969 if (rollback
.orig_dest
.ino
) {
7970 target
= mdcache
->get_inode(rollback
.orig_dest
.ino
);
7972 assert(destdn
&& straydn
);
7973 } else if (rollback
.orig_dest
.remote_ino
)
7974 target
= mdcache
->get_inode(rollback
.orig_dest
.remote_ino
);
7976 // can't use is_auth() in the resolve stage
7977 mds_rank_t whoami
= mds
->get_nodeid();
7979 assert(!destdn
|| destdn
->authority().first
!= whoami
);
7980 assert(!straydn
|| straydn
->authority().first
!= whoami
);
7982 bool force_journal_src
= false;
7983 bool force_journal_dest
= false;
7984 if (in
&& in
->is_dir() && srcdn
->authority().first
!= whoami
)
7985 force_journal_src
= _need_force_journal(in
, false);
7986 if (in
&& target
&& target
->is_dir())
7987 force_journal_dest
= _need_force_journal(in
, true);
7989 version_t srcdnpv
= 0;
7992 if (srcdn
->authority().first
== whoami
)
7993 srcdnpv
= srcdn
->pre_dirty();
7994 if (rollback
.orig_src
.ino
) {
7996 srcdn
->push_projected_linkage(in
);
7998 srcdn
->push_projected_linkage(rollback
.orig_src
.remote_ino
,
7999 rollback
.orig_src
.remote_d_type
);
8004 if (in
->authority().first
== whoami
) {
8005 pi
= in
->project_inode();
8006 mut
->add_projected_inode(in
);
8007 pi
->version
= in
->pre_dirty();
8009 pi
= in
->get_projected_inode();
8010 if (pi
->ctime
== rollback
.ctime
)
8011 pi
->ctime
= rollback
.orig_src
.old_ctime
;
8014 if (srcdn
&& srcdn
->authority().first
== whoami
) {
8016 _rollback_repair_dir(mut
, srcdir
, rollback
.orig_src
, rollback
.ctime
,
8017 in
? in
->is_dir() : false, 1, pi
? pi
->accounted_rstat
: blah
);
8022 if (rollback
.orig_dest
.ino
&& target
) {
8023 destdn
->push_projected_linkage(target
);
8024 } else if (rollback
.orig_dest
.remote_ino
) {
8025 destdn
->push_projected_linkage(rollback
.orig_dest
.remote_ino
,
8026 rollback
.orig_dest
.remote_d_type
);
8028 // the dentry will be trimmed soon, it's ok to have wrong linkage
8029 if (rollback
.orig_dest
.ino
)
8030 assert(mds
->is_resolve());
8031 destdn
->push_projected_linkage();
8036 straydn
->push_projected_linkage();
8040 if (target
->authority().first
== whoami
) {
8041 ti
= target
->project_inode();
8042 mut
->add_projected_inode(target
);
8043 ti
->version
= target
->pre_dirty();
8045 ti
= target
->get_projected_inode();
8046 if (ti
->ctime
== rollback
.ctime
)
8047 ti
->ctime
= rollback
.orig_dest
.old_ctime
;
8048 if (MDS_INO_IS_STRAY(rollback
.orig_src
.dirfrag
.ino
)) {
8049 if (MDS_INO_IS_STRAY(rollback
.orig_dest
.dirfrag
.ino
))
8050 assert(!rollback
.orig_dest
.ino
&& !rollback
.orig_dest
.remote_ino
);
8052 assert(rollback
.orig_dest
.remote_ino
&&
8053 rollback
.orig_dest
.remote_ino
== rollback
.orig_src
.ino
);
8059 dout(0) << " srcdn back to " << *srcdn
<< dendl
;
8061 dout(0) << " srci back to " << *in
<< dendl
;
8063 dout(0) << " destdn back to " << *destdn
<< dendl
;
8065 dout(0) << " desti back to " << *target
<< dendl
;
8068 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_rollback", rollback
.reqid
, master
,
8069 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::RENAME
);
8070 mdlog
->start_entry(le
);
8072 if (srcdn
&& (srcdn
->authority().first
== whoami
|| force_journal_src
)) {
8073 le
->commit
.add_dir_context(srcdir
);
8074 if (rollback
.orig_src
.ino
)
8075 le
->commit
.add_primary_dentry(srcdn
, 0, true);
8077 le
->commit
.add_remote_dentry(srcdn
, true);
8080 if (!rollback
.orig_src
.ino
&& // remote linkage
8081 in
&& in
->authority().first
== whoami
) {
8082 le
->commit
.add_dir_context(in
->get_projected_parent_dir());
8083 le
->commit
.add_primary_dentry(in
->get_projected_parent_dn(), in
, true);
8086 if (force_journal_dest
) {
8087 assert(rollback
.orig_dest
.ino
);
8088 le
->commit
.add_dir_context(destdir
);
8089 le
->commit
.add_primary_dentry(destdn
, 0, true);
8092 // slave: no need to journal straydn
8094 if (target
&& target
!= in
&& target
->authority().first
== whoami
) {
8095 assert(rollback
.orig_dest
.remote_ino
);
8096 le
->commit
.add_dir_context(target
->get_projected_parent_dir());
8097 le
->commit
.add_primary_dentry(target
->get_projected_parent_dn(), target
, true);
8100 if (in
&& in
->is_dir() && (srcdn
->authority().first
== whoami
|| force_journal_src
)) {
8101 dout(10) << " noting renamed dir ino " << in
->ino() << " in metablob" << dendl
;
8102 le
->commit
.renamed_dirino
= in
->ino();
8103 if (srcdn
->authority().first
== whoami
) {
8105 in
->get_dirfrags(ls
);
8106 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
8108 if (!dir
->is_auth())
8109 le
->commit
.renamed_dir_frags
.push_back(dir
->get_frag());
8111 dout(10) << " noting renamed dir open frags " << le
->commit
.renamed_dir_frags
<< dendl
;
8113 } else if (force_journal_dest
) {
8114 dout(10) << " noting rename target ino " << target
->ino() << " in metablob" << dendl
;
8115 le
->commit
.renamed_dirino
= target
->ino();
8118 if (target
&& target
->is_dir()) {
8120 mdcache
->project_subtree_rename(target
, straydir
, destdir
);
8123 if (in
&& in
->is_dir()) {
8125 mdcache
->project_subtree_rename(in
, destdir
, srcdir
);
8128 if (mdr
&& !mdr
->more()->slave_update_journaled
) {
8129 assert(le
->commit
.empty());
8130 mdlog
->cancel_entry(le
);
8132 _rename_rollback_finish(mut
, mdr
, srcdn
, srcdnpv
, destdn
, straydn
, finish_mdr
);
8134 assert(!le
->commit
.empty());
8136 mdr
->more()->slave_update_journaled
= false;
8137 MDSLogContextBase
*fin
= new C_MDS_LoggedRenameRollback(this, mut
, mdr
, srcdn
, srcdnpv
,
8138 destdn
, straydn
, finish_mdr
);
8139 submit_mdlog_entry(le
, fin
, mdr
, __func__
);
8144 void Server::_rename_rollback_finish(MutationRef
& mut
, MDRequestRef
& mdr
, CDentry
*srcdn
,
8145 version_t srcdnpv
, CDentry
*destdn
,
8146 CDentry
*straydn
, bool finish_mdr
)
8148 dout(10) << "_rename_rollback_finish " << mut
->reqid
<< dendl
;
8151 straydn
->get_dir()->unlink_inode(straydn
);
8152 straydn
->pop_projected_linkage();
8155 destdn
->get_dir()->unlink_inode(destdn
);
8156 destdn
->pop_projected_linkage();
8159 srcdn
->pop_projected_linkage();
8160 if (srcdn
->authority().first
== mds
->get_nodeid())
8161 srcdn
->mark_dirty(srcdnpv
, mut
->ls
);
8166 if (srcdn
&& srcdn
->get_linkage()->is_primary()) {
8167 CInode
*in
= srcdn
->get_linkage()->get_inode();
8168 if (srcdn
->authority().first
== mds
->get_nodeid())
8169 in
->state_set(CInode::STATE_AUTH
);
8170 // update subtree map?
8171 if (in
&& in
->is_dir()) {
8173 mdcache
->adjust_subtree_after_rename(in
, destdn
->get_dir(), true);
8178 CInode
*oldin
= destdn
->get_linkage()->get_inode();
8179 // update subtree map?
8180 if (oldin
&& oldin
->is_dir()) {
8182 mdcache
->adjust_subtree_after_rename(oldin
, straydn
->get_dir(), true);
8186 if (mds
->is_resolve()) {
8189 root
= mdcache
->get_subtree_root(straydn
->get_dir());
8191 root
= mdcache
->get_subtree_root(destdn
->get_dir());
8193 mdcache
->try_trim_non_auth_subtree(root
);
8197 list
<MDSInternalContextBase
*> finished
;
8198 if (mdr
->more()->is_ambiguous_auth
) {
8199 if (srcdn
->is_auth())
8200 mdr
->more()->rename_inode
->unfreeze_inode(finished
);
8202 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
8203 mdr
->more()->is_ambiguous_auth
= false;
8205 mds
->queue_waiters(finished
);
8206 if (finish_mdr
|| mdr
->aborted
)
8207 mdcache
->request_finish(mdr
);
8209 mdr
->more()->slave_rolling_back
= false;
8212 mdcache
->finish_rollback(mut
->reqid
);
8217 /* This function DOES put the passed message before returning*/
8218 void Server::handle_slave_rename_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
8220 dout(10) << "handle_slave_rename_prep_ack " << *mdr
8221 << " witnessed by " << ack
->get_source()
8222 << " " << *ack
<< dendl
;
8223 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
8226 mdr
->more()->slaves
.insert(from
);
8227 if (mdr
->more()->srcdn_auth_mds
== from
&&
8228 mdr
->more()->is_remote_frozen_authpin
&&
8229 !mdr
->more()->is_ambiguous_auth
) {
8230 mdr
->set_ambiguous_auth(mdr
->more()->rename_inode
);
8233 // witnessed? or add extra witnesses?
8234 assert(mdr
->more()->witnessed
.count(from
) == 0);
8235 if (ack
->is_interrupted()) {
8236 dout(10) << " slave request interrupted, noop" << dendl
;
8237 } else if (ack
->witnesses
.empty()) {
8238 mdr
->more()->witnessed
.insert(from
);
8239 if (!ack
->is_not_journaled())
8240 mdr
->more()->has_journaled_slaves
= true;
8242 dout(10) << " extra witnesses (srcdn replicas) are " << ack
->witnesses
<< dendl
;
8243 mdr
->more()->extra_witnesses
.swap(ack
->witnesses
);
8244 mdr
->more()->extra_witnesses
.erase(mds
->get_nodeid()); // not me!
8248 if (ack
->inode_export
.length()) {
8249 dout(10) << " got srci import" << dendl
;
8250 mdr
->more()->inode_import
.claim(ack
->inode_export
);
8251 mdr
->more()->inode_import_v
= ack
->inode_export_v
;
8254 // remove from waiting list
8255 assert(mdr
->more()->waiting_on_slave
.count(from
));
8256 mdr
->more()->waiting_on_slave
.erase(from
);
8258 if (mdr
->more()->waiting_on_slave
.empty())
8259 dispatch_client_request(mdr
); // go again!
8261 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
8264 void Server::handle_slave_rename_notify_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
8266 dout(10) << "handle_slave_rename_notify_ack " << *mdr
<< " from mds."
8267 << ack
->get_source() << dendl
;
8268 assert(mdr
->is_slave());
8269 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
8271 if (mdr
->more()->waiting_on_slave
.count(from
)) {
8272 mdr
->more()->waiting_on_slave
.erase(from
);
8274 if (mdr
->more()->waiting_on_slave
.empty()) {
8275 if (mdr
->slave_request
)
8276 dispatch_slave_request(mdr
);
8278 dout(10) << " still waiting for rename notify acks from "
8279 << mdr
->more()->waiting_on_slave
<< dendl
;
8283 void Server::_slave_rename_sessions_flushed(MDRequestRef
& mdr
)
8285 dout(10) << "_slave_rename_sessions_flushed " << *mdr
<< dendl
;
8287 if (mdr
->more()->waiting_on_slave
.count(MDS_RANK_NONE
)) {
8288 mdr
->more()->waiting_on_slave
.erase(MDS_RANK_NONE
);
8290 if (mdr
->more()->waiting_on_slave
.empty()) {
8291 if (mdr
->slave_request
)
8292 dispatch_slave_request(mdr
);
8294 dout(10) << " still waiting for rename notify acks from "
8295 << mdr
->more()->waiting_on_slave
<< dendl
;
8300 /* This function takes responsibility for the passed mdr*/
8301 void Server::handle_client_lssnap(MDRequestRef
& mdr
)
8303 MClientRequest
*req
= mdr
->client_request
;
8306 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8307 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8308 respond_to_request(mdr
, -ESTALE
);
8311 if (!diri
->is_auth()) {
8312 mdcache
->request_forward(mdr
, diri
->authority().first
);
8315 if (!diri
->is_dir()) {
8316 respond_to_request(mdr
, -ENOTDIR
);
8319 dout(10) << "lssnap on " << *diri
<< dendl
;
8322 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8323 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8324 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8327 if (!check_access(mdr
, diri
, MAY_READ
))
8330 SnapRealm
*realm
= diri
->find_snaprealm();
8331 map
<snapid_t
,SnapInfo
*> infomap
;
8332 realm
->get_snap_info(infomap
, diri
->get_oldest_snap());
8334 unsigned max_entries
= req
->head
.args
.readdir
.max_entries
;
8336 max_entries
= infomap
.size();
8337 int max_bytes
= req
->head
.args
.readdir
.max_bytes
;
8339 // make sure at least one item can be encoded
8340 max_bytes
= (512 << 10) + g_conf
->mds_max_xattr_pairs_size
;
8342 __u64 last_snapid
= 0;
8343 string offset_str
= req
->get_path2();
8344 if (!offset_str
.empty())
8345 last_snapid
= realm
->resolve_snapname(offset_str
, diri
->ino());
8348 encode_empty_dirstat(dirbl
);
8350 max_bytes
-= dirbl
.length() - sizeof(__u32
) + sizeof(__u8
) * 2;
8354 map
<snapid_t
,SnapInfo
*>::iterator p
= infomap
.upper_bound(last_snapid
);
8355 for (; p
!= infomap
.end() && num
< max_entries
; ++p
) {
8356 dout(10) << p
->first
<< " -> " << *p
->second
<< dendl
;
8360 if (p
->second
->ino
== diri
->ino())
8361 snap_name
= p
->second
->name
;
8363 snap_name
= p
->second
->get_long_name();
8365 unsigned start_len
= dnbl
.length();
8366 if (int(start_len
+ snap_name
.length() + sizeof(__u32
) + sizeof(LeaseStat
)) > max_bytes
)
8369 ::encode(snap_name
, dnbl
);
8370 encode_infinite_lease(dnbl
);
8372 int r
= diri
->encode_inodestat(dnbl
, mdr
->session
, realm
, p
->first
, max_bytes
- (int)dnbl
.length());
8375 keep
.substr_of(dnbl
, 0, start_len
);
8382 ::encode(num
, dirbl
);
8384 if (p
== infomap
.end()) {
8385 flags
= CEPH_READDIR_FRAG_END
;
8386 if (last_snapid
== 0)
8387 flags
|= CEPH_READDIR_FRAG_COMPLETE
;
8389 ::encode(flags
, dirbl
);
8390 dirbl
.claim_append(dnbl
);
8392 mdr
->reply_extra_bl
= dirbl
;
8394 respond_to_request(mdr
, 0);
8400 struct C_MDS_mksnap_finish
: public ServerLogContext
{
8403 C_MDS_mksnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, SnapInfo
&i
) :
8404 ServerLogContext(s
, r
), diri(di
), info(i
) {}
8405 void finish(int r
) override
{
8406 server
->_mksnap_finish(mdr
, diri
, info
);
8410 /* This function takes responsibility for the passed mdr*/
8411 void Server::handle_client_mksnap(MDRequestRef
& mdr
)
8413 if (!mds
->mdsmap
->allows_snaps()) {
8414 // you can't make snapshots until you set an option right now
8415 respond_to_request(mdr
, -EPERM
);
8419 MClientRequest
*req
= mdr
->client_request
;
8420 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8421 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8422 respond_to_request(mdr
, -ESTALE
);
8426 if (!diri
->is_auth()) { // fw to auth?
8427 mdcache
->request_forward(mdr
, diri
->authority().first
);
8432 if (!diri
->is_dir()) {
8433 respond_to_request(mdr
, -ENOTDIR
);
8436 if (diri
->is_system() && !diri
->is_root()) {
8437 // no snaps in system dirs (root is ok)
8438 respond_to_request(mdr
, -EPERM
);
8442 const string
&snapname
= req
->get_filepath().last_dentry();
8444 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
) {
8445 dout(20) << "mksnap " << snapname
<< " on " << *diri
<< " denied to uid " << mdr
->client_request
->get_caller_uid() << dendl
;
8446 respond_to_request(mdr
, -EPERM
);
8450 dout(10) << "mksnap " << snapname
<< " on " << *diri
<< dendl
;
8453 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8455 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8456 rdlocks
.erase(&diri
->snaplock
);
8457 xlocks
.insert(&diri
->snaplock
);
8459 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8462 if (!check_access(mdr
, diri
, MAY_WRITE
))
8465 // make sure name is unique
8466 if (diri
->snaprealm
&&
8467 diri
->snaprealm
->exists(snapname
)) {
8468 respond_to_request(mdr
, -EEXIST
);
8471 if (snapname
.length() == 0 ||
8472 snapname
[0] == '_') {
8473 respond_to_request(mdr
, -EINVAL
);
8477 // allocate a snapid
8478 if (!mdr
->more()->stid
) {
8480 mds
->snapclient
->prepare_create(diri
->ino(), snapname
,
8481 mdr
->get_mds_stamp(),
8482 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8483 new C_MDS_RetryRequest(mdcache
, mdr
));
8487 version_t stid
= mdr
->more()->stid
;
8489 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8490 ::decode(snapid
, p
);
8491 dout(10) << " stid " << stid
<< " snapid " << snapid
<< dendl
;
8495 info
.ino
= diri
->ino();
8496 info
.snapid
= snapid
;
8497 info
.name
= snapname
;
8498 info
.stamp
= mdr
->get_op_stamp();
8500 inode_t
*pi
= diri
->project_inode();
8501 pi
->ctime
= info
.stamp
;
8502 pi
->version
= diri
->pre_dirty();
8504 // project the snaprealm
8505 sr_t
*newsnap
= diri
->project_snaprealm(snapid
);
8506 newsnap
->snaps
[snapid
] = info
;
8507 newsnap
->seq
= snapid
;
8508 newsnap
->last_created
= snapid
;
8510 // journal the inode changes
8511 mdr
->ls
= mdlog
->get_current_segment();
8512 EUpdate
*le
= new EUpdate(mdlog
, "mksnap");
8513 mdlog
->start_entry(le
);
8515 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8516 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8517 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8518 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8520 // journal the snaprealm changes
8521 submit_mdlog_entry(le
, new C_MDS_mksnap_finish(this, mdr
, diri
, info
),
8526 void Server::_mksnap_finish(MDRequestRef
& mdr
, CInode
*diri
, SnapInfo
&info
)
8528 dout(10) << "_mksnap_finish " << *mdr
<< " " << info
<< dendl
;
8530 int op
= (diri
->snaprealm
? CEPH_SNAP_OP_CREATE
: CEPH_SNAP_OP_SPLIT
);
8532 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8535 mds
->snapclient
->commit(mdr
->more()->stid
, mdr
->ls
);
8538 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8540 mdcache
->do_realm_invalidate_and_update_notify(diri
, op
);
8544 mdr
->snapid
= info
.snapid
;
8546 respond_to_request(mdr
, 0);
8552 struct C_MDS_rmsnap_finish
: public ServerLogContext
{
8555 C_MDS_rmsnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, snapid_t sn
) :
8556 ServerLogContext(s
, r
), diri(di
), snapid(sn
) {}
8557 void finish(int r
) override
{
8558 server
->_rmsnap_finish(mdr
, diri
, snapid
);
8562 /* This function takes responsibility for the passed mdr*/
8563 void Server::handle_client_rmsnap(MDRequestRef
& mdr
)
8565 MClientRequest
*req
= mdr
->client_request
;
8567 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8568 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8569 respond_to_request(mdr
, -ESTALE
);
8572 if (!diri
->is_auth()) { // fw to auth?
8573 mdcache
->request_forward(mdr
, diri
->authority().first
);
8576 if (!diri
->is_dir()) {
8577 respond_to_request(mdr
, -ENOTDIR
);
8581 const string
&snapname
= req
->get_filepath().last_dentry();
8583 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
) {
8584 dout(20) << "rmsnap " << snapname
<< " on " << *diri
<< " denied to uid " << mdr
->client_request
->get_caller_uid() << dendl
;
8585 respond_to_request(mdr
, -EPERM
);
8589 dout(10) << "rmsnap " << snapname
<< " on " << *diri
<< dendl
;
8592 if (snapname
.length() == 0 || snapname
[0] == '_') {
8593 respond_to_request(mdr
, -EINVAL
); // can't prune a parent snap, currently.
8596 if (!diri
->snaprealm
|| !diri
->snaprealm
->exists(snapname
)) {
8597 respond_to_request(mdr
, -ENOENT
);
8600 snapid_t snapid
= diri
->snaprealm
->resolve_snapname(snapname
, diri
->ino());
8601 dout(10) << " snapname " << snapname
<< " is " << snapid
<< dendl
;
8603 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8604 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8605 rdlocks
.erase(&diri
->snaplock
);
8606 xlocks
.insert(&diri
->snaplock
);
8608 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8611 if (!check_access(mdr
, diri
, MAY_WRITE
))
8615 if (!mdr
->more()->stid
) {
8616 mds
->snapclient
->prepare_destroy(diri
->ino(), snapid
,
8617 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8618 new C_MDS_RetryRequest(mdcache
, mdr
));
8621 version_t stid
= mdr
->more()->stid
;
8622 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8625 dout(10) << " stid is " << stid
<< ", seq is " << seq
<< dendl
;
8628 inode_t
*pi
= diri
->project_inode();
8629 pi
->version
= diri
->pre_dirty();
8630 pi
->ctime
= mdr
->get_op_stamp();
8632 mdr
->ls
= mdlog
->get_current_segment();
8633 EUpdate
*le
= new EUpdate(mdlog
, "rmsnap");
8634 mdlog
->start_entry(le
);
8636 // project the snaprealm
8637 sr_t
*newnode
= diri
->project_snaprealm();
8638 newnode
->snaps
.erase(snapid
);
8640 newnode
->last_destroyed
= seq
;
8642 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8643 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8644 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8645 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8647 submit_mdlog_entry(le
, new C_MDS_rmsnap_finish(this, mdr
, diri
, snapid
),
8652 void Server::_rmsnap_finish(MDRequestRef
& mdr
, CInode
*diri
, snapid_t snapid
)
8654 dout(10) << "_rmsnap_finish " << *mdr
<< " " << snapid
<< dendl
;
8655 snapid_t stid
= mdr
->more()->stid
;
8656 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8660 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8663 mds
->snapclient
->commit(stid
, mdr
->ls
);
8665 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8667 mdcache
->do_realm_invalidate_and_update_notify(diri
, CEPH_SNAP_OP_DESTROY
);
8671 respond_to_request(mdr
, 0);
8673 // purge snapshot data
8674 if (diri
->snaprealm
->have_past_parents_open())
8675 diri
->purge_stale_snap_data(diri
->snaprealm
->get_snaps());
8678 struct C_MDS_renamesnap_finish
: public ServerLogContext
{
8681 C_MDS_renamesnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, snapid_t sn
) :
8682 ServerLogContext(s
, r
), diri(di
), snapid(sn
) {}
8683 void finish(int r
) override
{
8684 server
->_renamesnap_finish(mdr
, diri
, snapid
);
8688 /* This function takes responsibility for the passed mdr*/
8689 void Server::handle_client_renamesnap(MDRequestRef
& mdr
)
8691 MClientRequest
*req
= mdr
->client_request
;
8692 if (req
->get_filepath().get_ino() != req
->get_filepath2().get_ino()) {
8693 respond_to_request(mdr
, -EINVAL
);
8697 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8698 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8699 respond_to_request(mdr
, -ESTALE
);
8703 if (!diri
->is_auth()) { // fw to auth?
8704 mdcache
->request_forward(mdr
, diri
->authority().first
);
8708 if (!diri
->is_dir()) { // dir only
8709 respond_to_request(mdr
, -ENOTDIR
);
8713 if (mdr
->client_request
->get_caller_uid() < g_conf
->mds_snap_min_uid
||
8714 mdr
->client_request
->get_caller_uid() > g_conf
->mds_snap_max_uid
) {
8715 respond_to_request(mdr
, -EPERM
);
8719 const string
&dstname
= req
->get_filepath().last_dentry();
8720 const string
&srcname
= req
->get_filepath2().last_dentry();
8721 dout(10) << "renamesnap " << srcname
<< "->" << dstname
<< " on " << *diri
<< dendl
;
8723 if (srcname
.length() == 0 || srcname
[0] == '_') {
8724 respond_to_request(mdr
, -EINVAL
); // can't rename a parent snap.
8727 if (!diri
->snaprealm
|| !diri
->snaprealm
->exists(srcname
)) {
8728 respond_to_request(mdr
, -ENOENT
);
8731 if (dstname
.length() == 0 || dstname
[0] == '_') {
8732 respond_to_request(mdr
, -EINVAL
);
8735 if (diri
->snaprealm
->exists(dstname
)) {
8736 respond_to_request(mdr
, -EEXIST
);
8740 snapid_t snapid
= diri
->snaprealm
->resolve_snapname(srcname
, diri
->ino());
8741 dout(10) << " snapname " << srcname
<< " is " << snapid
<< dendl
;
8744 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8746 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8747 rdlocks
.erase(&diri
->snaplock
);
8748 xlocks
.insert(&diri
->snaplock
);
8750 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8753 if (!check_access(mdr
, diri
, MAY_WRITE
))
8757 if (!mdr
->more()->stid
) {
8758 mds
->snapclient
->prepare_update(diri
->ino(), snapid
, dstname
, utime_t(),
8759 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8760 new C_MDS_RetryRequest(mdcache
, mdr
));
8764 version_t stid
= mdr
->more()->stid
;
8765 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8768 dout(10) << " stid is " << stid
<< ", seq is " << seq
<< dendl
;
8771 inode_t
*pi
= diri
->project_inode();
8772 pi
->ctime
= mdr
->get_op_stamp();
8773 pi
->version
= diri
->pre_dirty();
8775 // project the snaprealm
8776 sr_t
*newsnap
= diri
->project_snaprealm();
8777 assert(newsnap
->snaps
.count(snapid
));
8778 newsnap
->snaps
[snapid
].name
= dstname
;
8780 // journal the inode changes
8781 mdr
->ls
= mdlog
->get_current_segment();
8782 EUpdate
*le
= new EUpdate(mdlog
, "renamesnap");
8783 mdlog
->start_entry(le
);
8785 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8786 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8787 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8788 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8790 // journal the snaprealm changes
8791 submit_mdlog_entry(le
, new C_MDS_renamesnap_finish(this, mdr
, diri
, snapid
),
8796 void Server::_renamesnap_finish(MDRequestRef
& mdr
, CInode
*diri
, snapid_t snapid
)
8798 dout(10) << "_renamesnap_finish " << *mdr
<< " " << snapid
<< dendl
;
8800 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8803 mds
->snapclient
->commit(mdr
->more()->stid
, mdr
->ls
);
8805 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8807 mdcache
->do_realm_invalidate_and_update_notify(diri
, CEPH_SNAP_OP_UPDATE
, true);
8812 mdr
->snapid
= snapid
;
8813 respond_to_request(mdr
, 0);
8817 * Return true if server is in state RECONNECT and this
8818 * client has not yet reconnected.
8820 bool Server::waiting_for_reconnect(client_t c
) const
8822 return client_reconnect_gather
.count(c
) > 0;
8825 void Server::dump_reconnect_status(Formatter
*f
) const
8827 f
->open_object_section("reconnect_status");
8828 f
->dump_stream("client_reconnect_gather") << client_reconnect_gather
;