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 "
74 class ServerContext
: public MDSInternalContextBase
{
77 MDSRank
*get_mds() override
83 explicit ServerContext(Server
*s
) : server(s
) {
84 assert(server
!= NULL
);
88 class ServerLogContext
: public MDSLogContextBase
{
91 MDSRank
*get_mds() override
97 void pre_finish(int r
) override
{
99 mdr
->mark_event("journal_committed: ");
102 explicit ServerLogContext(Server
*s
) : server(s
) {
103 assert(server
!= NULL
);
105 explicit ServerLogContext(Server
*s
, MDRequestRef
& r
) : server(s
), mdr(r
) {
106 assert(server
!= NULL
);
110 void Server::create_logger()
112 PerfCountersBuilder
plb(g_ceph_context
, "mds_server", l_mdss_first
, l_mdss_last
);
113 plb
.add_u64_counter(l_mdss_handle_client_request
,"handle_client_request",
114 "Client requests", "hcr", PerfCountersBuilder::PRIO_INTERESTING
);
115 plb
.add_u64_counter(l_mdss_handle_slave_request
, "handle_slave_request",
116 "Slave requests", "hsr", PerfCountersBuilder::PRIO_INTERESTING
);
117 plb
.add_u64_counter(l_mdss_handle_client_session
, "handle_client_session",
118 "Client session messages", "hcs", PerfCountersBuilder::PRIO_INTERESTING
);
119 plb
.add_u64_counter(l_mdss_dispatch_client_request
, "dispatch_client_request", "Client requests dispatched");
120 plb
.add_u64_counter(l_mdss_dispatch_slave_request
, "dispatch_server_request", "Server requests dispatched");
121 plb
.add_u64_counter(l_mdss_req_lookuphash
, "req_lookuphash",
122 "Request type lookup hash of inode");
123 plb
.add_u64_counter(l_mdss_req_lookupino
, "req_lookupino",
124 "Request type lookup inode");
125 plb
.add_u64_counter(l_mdss_req_lookupparent
, "req_lookupparent",
126 "Request type lookup parent");
127 plb
.add_u64_counter(l_mdss_req_lookupname
, "req_lookupname",
128 "Request type lookup name");
129 plb
.add_u64_counter(l_mdss_req_lookup
, "req_lookup",
130 "Request type lookup");
131 plb
.add_u64_counter(l_mdss_req_lookupsnap
, "req_lookupsnap",
132 "Request type lookup snapshot");
133 plb
.add_u64_counter(l_mdss_req_getattr
, "req_getattr",
134 "Request type get attribute");
135 plb
.add_u64_counter(l_mdss_req_setattr
, "req_setattr",
136 "Request type set attribute");
137 plb
.add_u64_counter(l_mdss_req_setlayout
, "req_setlayout",
138 "Request type set file layout");
139 plb
.add_u64_counter(l_mdss_req_setdirlayout
, "req_setdirlayout",
140 "Request type set directory layout");
141 plb
.add_u64_counter(l_mdss_req_setxattr
, "req_setxattr",
142 "Request type set extended attribute");
143 plb
.add_u64_counter(l_mdss_req_rmxattr
, "req_rmxattr",
144 "Request type remove extended attribute");
145 plb
.add_u64_counter(l_mdss_req_readdir
, "req_readdir",
146 "Request type read directory");
147 plb
.add_u64_counter(l_mdss_req_setfilelock
, "req_setfilelock",
148 "Request type set file lock");
149 plb
.add_u64_counter(l_mdss_req_getfilelock
, "req_getfilelock",
150 "Request type get file lock");
151 plb
.add_u64_counter(l_mdss_req_create
, "req_create",
152 "Request type create");
153 plb
.add_u64_counter(l_mdss_req_open
, "req_open",
154 "Request type open");
155 plb
.add_u64_counter(l_mdss_req_mknod
, "req_mknod",
156 "Request type make node");
157 plb
.add_u64_counter(l_mdss_req_link
, "req_link",
158 "Request type link");
159 plb
.add_u64_counter(l_mdss_req_unlink
, "req_unlink",
160 "Request type unlink");
161 plb
.add_u64_counter(l_mdss_req_rmdir
, "req_rmdir",
162 "Request type remove directory");
163 plb
.add_u64_counter(l_mdss_req_rename
, "req_rename",
164 "Request type rename");
165 plb
.add_u64_counter(l_mdss_req_mkdir
, "req_mkdir",
166 "Request type make directory");
167 plb
.add_u64_counter(l_mdss_req_symlink
, "req_symlink",
168 "Request type symbolic link");
169 plb
.add_u64_counter(l_mdss_req_lssnap
, "req_lssnap",
170 "Request type list snapshot");
171 plb
.add_u64_counter(l_mdss_req_mksnap
, "req_mksnap",
172 "Request type make snapshot");
173 plb
.add_u64_counter(l_mdss_req_rmsnap
, "req_rmsnap",
174 "Request type remove snapshot");
175 plb
.add_u64_counter(l_mdss_req_renamesnap
, "req_renamesnap",
176 "Request type rename snapshot");
177 logger
= plb
.create_perf_counters();
178 g_ceph_context
->get_perfcounters_collection()->add(logger
);
181 Server::Server(MDSRank
*m
) :
183 mdcache(mds
->mdcache
), mdlog(mds
->mdlog
),
186 reconnect_done(NULL
),
187 failed_reconnects(0),
188 reconnect_evicting(false),
189 terminating_sessions(false)
194 /* This function DOES put the passed message before returning*/
195 void Server::dispatch(Message
*m
)
197 switch (m
->get_type()) {
198 case CEPH_MSG_CLIENT_RECONNECT
:
199 handle_client_reconnect(static_cast<MClientReconnect
*>(m
));
204 if (!mds
->is_active() &&
205 !(mds
->is_stopping() && m
->get_source().is_mds())) {
206 if (m
->get_type() == CEPH_MSG_CLIENT_REQUEST
&&
207 (mds
->is_reconnect() || mds
->get_want_state() == CEPH_MDS_STATE_RECONNECT
)) {
208 MClientRequest
*req
= static_cast<MClientRequest
*>(m
);
209 Session
*session
= get_session(req
);
210 if (!session
|| session
->is_closed()) {
211 dout(5) << "session is closed, dropping " << req
->get_reqid() << dendl
;
215 bool queue_replay
= false;
216 if (req
->is_replay()) {
217 dout(3) << "queuing replayed op" << dendl
;
219 } else if (req
->get_retry_attempt()) {
220 // process completed request in clientreplay stage. The completed request
221 // might have created new file/directorie. This guarantees MDS sends a reply
222 // to client before other request modifies the new file/directorie.
223 if (session
->have_completed_request(req
->get_reqid().tid
, NULL
)) {
224 dout(3) << "queuing completed op" << dendl
;
227 // this request was created before the cap reconnect message, drop any embedded
229 req
->releases
.clear();
232 req
->mark_queued_for_replay();
233 mds
->enqueue_replay(new C_MDS_RetryMessage(mds
, m
));
238 bool wait_for_active
= true;
239 if (m
->get_type() == MSG_MDS_SLAVE_REQUEST
) {
240 // handle_slave_request() will wait if necessary
241 wait_for_active
= false;
242 } else if (mds
->is_clientreplay()) {
243 // session open requests need to be handled during replay,
244 // close requests need to be delayed
245 if ((m
->get_type() == CEPH_MSG_CLIENT_SESSION
&&
246 (static_cast<MClientSession
*>(m
))->get_op() != CEPH_SESSION_REQUEST_CLOSE
)) {
247 wait_for_active
= false;
248 } else if (m
->get_type() == CEPH_MSG_CLIENT_REQUEST
) {
249 MClientRequest
*req
= static_cast<MClientRequest
*>(m
);
250 if (req
->is_queued_for_replay()) {
251 wait_for_active
= false;
255 if (wait_for_active
) {
256 dout(3) << "not active yet, waiting" << dendl
;
257 mds
->wait_for_active(new C_MDS_RetryMessage(mds
, m
));
262 switch (m
->get_type()) {
263 case CEPH_MSG_CLIENT_SESSION
:
264 handle_client_session(static_cast<MClientSession
*>(m
));
266 case CEPH_MSG_CLIENT_REQUEST
:
267 handle_client_request(static_cast<MClientRequest
*>(m
));
269 case MSG_MDS_SLAVE_REQUEST
:
270 handle_slave_request(static_cast<MMDSSlaveRequest
*>(m
));
273 derr
<< "server unknown message " << m
->get_type() << dendl
;
274 assert(0 == "server unknown message");
280 // ----------------------------------------------------------
281 // SESSION management
283 class C_MDS_session_finish
: public ServerLogContext
{
288 interval_set
<inodeno_t
> inos
;
292 C_MDS_session_finish(Server
*srv
, Session
*se
, uint64_t sseq
, bool s
, version_t mv
, Context
*fin_
= NULL
) :
293 ServerLogContext(srv
), session(se
), state_seq(sseq
), open(s
), cmapv(mv
), inotablev(0), fin(fin_
) { }
294 C_MDS_session_finish(Server
*srv
, Session
*se
, uint64_t sseq
, bool s
, version_t mv
, interval_set
<inodeno_t
>& i
, version_t iv
, Context
*fin_
= NULL
) :
295 ServerLogContext(srv
), session(se
), state_seq(sseq
), open(s
), cmapv(mv
), inos(i
), inotablev(iv
), fin(fin_
) { }
296 void finish(int r
) override
{
298 server
->_session_logged(session
, state_seq
, open
, cmapv
, inos
, inotablev
);
305 Session
*Server::get_session(Message
*m
)
307 Session
*session
= static_cast<Session
*>(m
->get_connection()->get_priv());
309 dout(20) << "get_session have " << session
<< " " << session
->info
.inst
310 << " state " << session
->get_state_name() << dendl
;
311 session
->put(); // not carry ref
313 dout(20) << "get_session dne for " << m
->get_source_inst() << dendl
;
318 /* This function DOES put the passed message before returning*/
319 void Server::handle_client_session(MClientSession
*m
)
322 bool blacklisted
= false;
323 Session
*session
= get_session(m
);
325 dout(3) << "handle_client_session " << *m
<< " from " << m
->get_source() << dendl
;
326 assert(m
->get_source().is_client()); // should _not_ come from an mds!
329 dout(0) << " ignoring sessionless msg " << *m
<< dendl
;
335 logger
->inc(l_mdss_handle_client_session
);
338 switch (m
->get_op()) {
339 case CEPH_SESSION_REQUEST_OPEN
:
340 if (session
->is_opening() ||
341 session
->is_open() ||
342 session
->is_stale() ||
343 session
->is_killing()) {
344 dout(10) << "currently open|opening|stale|killing, dropping this req" << dendl
;
348 assert(session
->is_closed() ||
349 session
->is_closing());
351 blacklisted
= mds
->objecter
->with_osdmap(
352 [session
](const OSDMap
&osd_map
) -> bool {
353 return osd_map
.is_blacklisted(session
->info
.inst
.addr
);
357 dout(10) << "ignoring blacklisted client " << session
->info
.inst
.addr
<< dendl
;
362 session
->set_client_metadata(m
->client_meta
);
363 dout(20) << __func__
<< " CEPH_SESSION_REQUEST_OPEN "
364 << session
->info
.client_metadata
.size() << " metadata entries:" << dendl
;
365 for (map
<string
, string
>::iterator i
= session
->info
.client_metadata
.begin();
366 i
!= session
->info
.client_metadata
.end(); ++i
) {
367 dout(20) << " " << i
->first
<< ": " << i
->second
<< dendl
;
370 // Special case for the 'root' metadata path; validate that the claimed
371 // root is actually within the caps of the session
372 if (session
->info
.client_metadata
.count("root")) {
373 const auto claimed_root
= session
->info
.client_metadata
.at("root");
374 // claimed_root has a leading "/" which we strip before passing
376 if (claimed_root
.empty() || claimed_root
[0] != '/' ||
377 !session
->auth_caps
.path_capable(claimed_root
.substr(1))) {
378 derr
<< __func__
<< " forbidden path claimed as mount root: "
379 << claimed_root
<< " by " << m
->get_source() << dendl
;
380 // Tell the client we're rejecting their open
381 mds
->send_message_client(new MClientSession(CEPH_SESSION_REJECT
), session
);
382 mds
->clog
->warn() << "client session with invalid root '" <<
383 claimed_root
<< "' denied (" << session
->info
.inst
<< ")";
385 // Drop out; don't record this session in SessionMap or journal it.
390 if (session
->is_closed())
391 mds
->sessionmap
.add_session(session
);
393 pv
= mds
->sessionmap
.mark_projected(session
);
394 sseq
= mds
->sessionmap
.set_state(session
, Session::STATE_OPENING
);
395 mds
->sessionmap
.touch_session(session
);
396 mdlog
->start_submit_entry(new ESession(m
->get_source_inst(), true, pv
, m
->client_meta
),
397 new C_MDS_session_finish(this, session
, sseq
, true, pv
));
401 case CEPH_SESSION_REQUEST_RENEWCAPS
:
402 if (session
->is_open() ||
403 session
->is_stale()) {
404 mds
->sessionmap
.touch_session(session
);
405 if (session
->is_stale()) {
406 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
407 mds
->locker
->resume_stale_caps(session
);
408 mds
->sessionmap
.touch_session(session
);
410 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_RENEWCAPS
, m
->get_seq()));
412 dout(10) << "ignoring renewcaps on non open|stale session (" << session
->get_state_name() << ")" << dendl
;
416 case CEPH_SESSION_REQUEST_CLOSE
:
418 if (session
->is_closed() ||
419 session
->is_closing() ||
420 session
->is_killing()) {
421 dout(10) << "already closed|closing|killing, dropping this req" << dendl
;
425 if (session
->is_importing()) {
426 dout(10) << "ignoring close req on importing session" << dendl
;
430 assert(session
->is_open() ||
431 session
->is_stale() ||
432 session
->is_opening());
433 if (m
->get_seq() < session
->get_push_seq()) {
434 dout(10) << "old push seq " << m
->get_seq() << " < " << session
->get_push_seq()
435 << ", dropping" << dendl
;
439 // We are getting a seq that is higher than expected.
440 // Handle the same as any other seqn error.
442 if (m
->get_seq() != session
->get_push_seq()) {
443 dout(0) << "old push seq " << m
->get_seq() << " != " << session
->get_push_seq()
444 << ", BUGGY!" << dendl
;
445 mds
->clog
->warn() << "incorrect push seq " << m
->get_seq() << " != "
446 << session
->get_push_seq() << ", dropping" << " from client : " << session
->get_human_name();
450 journal_close_session(session
, Session::STATE_CLOSING
, NULL
);
454 case CEPH_SESSION_FLUSHMSG_ACK
:
455 finish_flush_session(session
, m
->get_seq());
458 case CEPH_SESSION_REQUEST_FLUSH_MDLOG
:
468 void Server::flush_client_sessions(set
<client_t
>& client_set
, MDSGatherBuilder
& gather
)
470 for (set
<client_t
>::iterator p
= client_set
.begin(); p
!= client_set
.end(); ++p
) {
471 Session
*session
= mds
->sessionmap
.get_session(entity_name_t::CLIENT(p
->v
));
473 if (!session
->is_open() ||
474 !session
->connection
.get() ||
475 !session
->connection
->has_feature(CEPH_FEATURE_EXPORT_PEER
))
477 version_t seq
= session
->wait_for_flush(gather
.new_sub());
478 mds
->send_message_client(new MClientSession(CEPH_SESSION_FLUSHMSG
, seq
), session
);
482 void Server::finish_flush_session(Session
*session
, version_t seq
)
484 list
<MDSInternalContextBase
*> finished
;
485 session
->finish_flush(seq
, finished
);
486 mds
->queue_waiters(finished
);
489 void Server::_session_logged(Session
*session
, uint64_t state_seq
, bool open
, version_t pv
,
490 interval_set
<inodeno_t
>& inos
, version_t piv
)
492 dout(10) << "_session_logged " << session
->info
.inst
<< " state_seq " << state_seq
<< " " << (open
? "open":"close")
493 << " " << pv
<< dendl
;
496 assert(session
->is_closing() || session
->is_killing() ||
497 session
->is_opening()); // re-open closing session
498 session
->info
.prealloc_inos
.subtract(inos
);
499 mds
->inotable
->apply_release_ids(inos
);
500 assert(mds
->inotable
->get_version() == piv
);
503 mds
->sessionmap
.mark_dirty(session
);
506 if (session
->get_state_seq() != state_seq
) {
507 dout(10) << " journaled state_seq " << state_seq
<< " != current " << session
->get_state_seq()
508 << ", noop" << dendl
;
509 // close must have been canceled (by an import?), or any number of other things..
511 assert(session
->is_opening());
512 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
513 mds
->sessionmap
.touch_session(session
);
514 assert(session
->connection
!= NULL
);
515 session
->connection
->send_message(new MClientSession(CEPH_SESSION_OPEN
));
516 if (mdcache
->is_readonly())
517 session
->connection
->send_message(new MClientSession(CEPH_SESSION_FORCE_RO
));
518 } else if (session
->is_closing() ||
519 session
->is_killing()) {
520 // kill any lingering capabilities, leases, requests
521 while (!session
->caps
.empty()) {
522 Capability
*cap
= session
->caps
.front();
523 CInode
*in
= cap
->get_inode();
524 dout(20) << " killing capability " << ccap_string(cap
->issued()) << " on " << *in
<< dendl
;
525 mds
->locker
->remove_client_cap(in
, session
->info
.inst
.name
.num());
527 while (!session
->leases
.empty()) {
528 ClientLease
*r
= session
->leases
.front();
529 CDentry
*dn
= static_cast<CDentry
*>(r
->parent
);
530 dout(20) << " killing client lease of " << *dn
<< dendl
;
531 dn
->remove_client_lease(r
, mds
->locker
);
533 if (client_reconnect_gather
.count(session
->info
.get_client())) {
534 dout(20) << " removing client from reconnect set" << dendl
;
535 client_reconnect_gather
.erase(session
->info
.get_client());
537 if (client_reconnect_gather
.empty()) {
538 dout(7) << " client " << session
->info
.inst
<< " was last reconnect, finishing" << dendl
;
539 reconnect_gather_finish();
543 if (session
->is_closing()) {
544 // mark con disposable. if there is a fault, we will get a
545 // reset and clean it up. if the client hasn't received the
546 // CLOSE message yet, they will reconnect and get an
547 // ms_handle_remote_reset() and realize they had in fact closed.
548 // do this *before* sending the message to avoid a possible
550 if (session
->connection
!= NULL
) {
551 // Conditional because terminate_sessions will indiscrimately
552 // put sessions in CLOSING whether they ever had a conn or not.
553 session
->connection
->mark_disposable();
557 mds
->send_message_client(new MClientSession(CEPH_SESSION_CLOSE
), session
);
558 mds
->sessionmap
.set_state(session
, Session::STATE_CLOSED
);
560 mds
->sessionmap
.remove_session(session
);
561 } else if (session
->is_killing()) {
562 // destroy session, close connection
563 if (session
->connection
!= NULL
) {
564 session
->connection
->mark_down();
566 mds
->sessionmap
.remove_session(session
);
576 * Inject sessions from some source other than actual connections.
579 * - sessions inferred from journal replay
580 * - sessions learned from other MDSs during rejoin
581 * - sessions learned from other MDSs during dir/caps migration
582 * - sessions learned from other MDSs during a cross-MDS rename
584 version_t
Server::prepare_force_open_sessions(map
<client_t
,entity_inst_t
>& cm
,
585 map
<client_t
,uint64_t>& sseqmap
)
587 version_t pv
= mds
->sessionmap
.get_projected();
589 dout(10) << "prepare_force_open_sessions " << pv
590 << " on " << cm
.size() << " clients"
592 for (map
<client_t
,entity_inst_t
>::iterator p
= cm
.begin(); p
!= cm
.end(); ++p
) {
594 Session
*session
= mds
->sessionmap
.get_or_add_session(p
->second
);
595 pv
= mds
->sessionmap
.mark_projected(session
);
596 if (session
->is_closed() ||
597 session
->is_closing() ||
598 session
->is_killing())
599 sseqmap
[p
->first
] = mds
->sessionmap
.set_state(session
, Session::STATE_OPENING
);
601 assert(session
->is_open() ||
602 session
->is_opening() ||
603 session
->is_stale());
604 session
->inc_importing();
609 void Server::finish_force_open_sessions(map
<client_t
,entity_inst_t
>& cm
,
610 map
<client_t
,uint64_t>& sseqmap
,
614 * FIXME: need to carefully consider the race conditions between a
615 * client trying to close a session and an MDS doing an import
616 * trying to force open a session...
618 dout(10) << "finish_force_open_sessions on " << cm
.size() << " clients,"
619 << " initial v " << mds
->sessionmap
.get_version() << dendl
;
622 int sessions_inserted
= 0;
623 for (map
<client_t
,entity_inst_t
>::iterator p
= cm
.begin(); p
!= cm
.end(); ++p
) {
626 Session
*session
= mds
->sessionmap
.get_session(p
->second
.name
);
629 if (sseqmap
.count(p
->first
)) {
630 uint64_t sseq
= sseqmap
[p
->first
];
631 if (session
->get_state_seq() != sseq
) {
632 dout(10) << "force_open_sessions skipping changed " << session
->info
.inst
<< dendl
;
634 dout(10) << "force_open_sessions opened " << session
->info
.inst
<< dendl
;
635 mds
->sessionmap
.set_state(session
, Session::STATE_OPEN
);
636 mds
->sessionmap
.touch_session(session
);
637 mds
->send_message_client(new MClientSession(CEPH_SESSION_OPEN
), session
);
638 if (mdcache
->is_readonly())
639 mds
->send_message_client(new MClientSession(CEPH_SESSION_FORCE_RO
), session
);
642 dout(10) << "force_open_sessions skipping already-open " << session
->info
.inst
<< dendl
;
643 assert(session
->is_open() || session
->is_stale());
647 session
->dec_importing();
650 mds
->sessionmap
.mark_dirty(session
);
653 dout(10) << __func__
<< ": final v " << mds
->sessionmap
.get_version() << dendl
;
656 class C_MDS_TerminatedSessions
: public ServerContext
{
657 void finish(int r
) override
{
658 server
->terminating_sessions
= false;
661 explicit C_MDS_TerminatedSessions(Server
*s
) : ServerContext(s
) {}
664 void Server::terminate_sessions()
666 dout(2) << "terminate_sessions" << dendl
;
668 terminating_sessions
= true;
670 // kill them off. clients will retry etc.
671 set
<Session
*> sessions
;
672 mds
->sessionmap
.get_client_session_set(sessions
);
673 for (set
<Session
*>::const_iterator p
= sessions
.begin();
676 Session
*session
= *p
;
677 if (session
->is_closing() ||
678 session
->is_killing() ||
679 session
->is_closed())
681 journal_close_session(session
, Session::STATE_CLOSING
, NULL
);
684 mdlog
->wait_for_safe(new C_MDS_TerminatedSessions(this));
688 void Server::find_idle_sessions()
690 dout(10) << "find_idle_sessions. laggy until " << mds
->get_laggy_until() << dendl
;
693 // (caps go stale, lease die)
694 utime_t now
= ceph_clock_now();
695 utime_t cutoff
= now
;
696 cutoff
-= g_conf
->mds_session_timeout
;
698 Session
*session
= mds
->sessionmap
.get_oldest_session(Session::STATE_OPEN
);
700 dout(20) << "laggiest active session is " << session
->info
.inst
<< dendl
;
701 if (session
->last_cap_renew
>= cutoff
) {
702 dout(20) << "laggiest active session is " << session
->info
.inst
<< " and sufficiently new ("
703 << session
->last_cap_renew
<< ")" << dendl
;
707 dout(10) << "new stale session " << session
->info
.inst
<< " last " << session
->last_cap_renew
<< dendl
;
708 mds
->sessionmap
.set_state(session
, Session::STATE_STALE
);
709 mds
->locker
->revoke_stale_caps(session
);
710 mds
->locker
->remove_stale_leases(session
);
711 mds
->send_message_client(new MClientSession(CEPH_SESSION_STALE
, session
->get_push_seq()), session
);
712 finish_flush_session(session
, session
->get_push_seq());
717 cutoff
-= g_conf
->mds_session_autoclose
;
719 // don't kick clients if we've been laggy
720 if (mds
->get_laggy_until() > cutoff
) {
721 dout(10) << " laggy_until " << mds
->get_laggy_until() << " > cutoff " << cutoff
722 << ", not kicking any clients to be safe" << dendl
;
726 if (mds
->sessionmap
.get_sessions().size() == 1 &&
727 mds
->mdsmap
->get_num_in_mds() == 1) {
728 dout(20) << "not evicting a slow client, because there is only one"
733 // Collect a list of sessions exceeding the autoclose threshold
734 std::vector
<Session
*> to_evict
;
735 const auto sessions_p
= mds
->sessionmap
.by_state
.find(Session::STATE_STALE
);
736 if (sessions_p
== mds
->sessionmap
.by_state
.end() || sessions_p
->second
->empty()) {
739 const auto &stale_sessions
= sessions_p
->second
;
740 assert(stale_sessions
!= nullptr);
742 for (const auto &session
: *stale_sessions
) {
743 if (session
->is_importing()) {
744 dout(10) << "stopping at importing session " << session
->info
.inst
<< dendl
;
747 assert(session
->is_stale());
748 if (session
->last_cap_renew
>= cutoff
) {
749 dout(20) << "oldest stale session is " << session
->info
.inst
<< " and sufficiently new ("
750 << session
->last_cap_renew
<< ")" << dendl
;
754 to_evict
.push_back(session
);
757 for (const auto &session
: to_evict
) {
759 age
-= session
->last_cap_renew
;
760 mds
->clog
->warn() << "evicting unresponsive client " << *session
761 << ", after " << age
<< " seconds";
762 dout(10) << "autoclosing stale session " << session
->info
.inst
<< " last "
763 << session
->last_cap_renew
<< dendl
;
765 if (g_conf
->mds_session_blacklist_on_timeout
) {
766 std::stringstream ss
;
767 mds
->evict_client(session
->info
.inst
.name
.num(), false, true,
770 kill_session(session
, NULL
);
776 * XXX bump in the interface here, not using an MDSInternalContextBase here
777 * because all the callers right now happen to use a SaferCond
779 void Server::kill_session(Session
*session
, Context
*on_safe
)
781 assert(mds
->mds_lock
.is_locked_by_me());
783 if ((session
->is_opening() ||
784 session
->is_open() ||
785 session
->is_stale()) &&
786 !session
->is_importing()) {
787 dout(10) << "kill_session " << session
<< dendl
;
788 journal_close_session(session
, Session::STATE_KILLING
, on_safe
);
790 dout(10) << "kill_session importing or already closing/killing " << session
<< dendl
;
791 assert(session
->is_closing() ||
792 session
->is_closed() ||
793 session
->is_killing() ||
794 session
->is_importing());
796 on_safe
->complete(0);
801 size_t Server::apply_blacklist(const std::set
<entity_addr_t
> &blacklist
)
803 std::list
<Session
*> victims
;
804 const auto sessions
= mds
->sessionmap
.get_sessions();
805 for (const auto p
: sessions
) {
806 if (!p
.first
.is_client()) {
807 // Do not apply OSDMap blacklist to MDS daemons, we find out
808 // about their death via MDSMap.
812 Session
*s
= p
.second
;
813 if (blacklist
.count(s
->info
.inst
.addr
)) {
814 victims
.push_back(s
);
818 for (const auto s
: victims
) {
819 kill_session(s
, nullptr);
822 dout(10) << "apply_blacklist: killed " << victims
.size() << dendl
;
824 return victims
.size();
827 void Server::journal_close_session(Session
*session
, int state
, Context
*on_safe
)
829 uint64_t sseq
= mds
->sessionmap
.set_state(session
, state
);
830 version_t pv
= mds
->sessionmap
.mark_projected(session
);
833 // release alloc and pending-alloc inos for this session
834 // and wipe out session state, in case the session close aborts for some reason
835 interval_set
<inodeno_t
> both
;
836 both
.insert(session
->info
.prealloc_inos
);
837 both
.insert(session
->pending_prealloc_inos
);
839 mds
->inotable
->project_release_ids(both
);
840 piv
= mds
->inotable
->get_projected_version();
844 mdlog
->start_submit_entry(new ESession(session
->info
.inst
, false, pv
, both
, piv
),
845 new C_MDS_session_finish(this, session
, sseq
, false, pv
, both
, piv
, on_safe
));
848 // clean up requests, too
849 elist
<MDRequestImpl
*>::iterator p
=
850 session
->requests
.begin(member_offset(MDRequestImpl
,
851 item_session_request
));
853 MDRequestRef mdr
= mdcache
->request_get((*p
)->reqid
);
855 mdcache
->request_kill(mdr
);
858 finish_flush_session(session
, session
->get_push_seq());
861 void Server::reconnect_clients(MDSInternalContext
*reconnect_done_
)
863 reconnect_done
= reconnect_done_
;
864 mds
->sessionmap
.get_client_set(client_reconnect_gather
);
866 if (client_reconnect_gather
.empty()) {
867 dout(7) << "reconnect_clients -- no sessions, doing nothing." << dendl
;
868 reconnect_gather_finish();
872 // clients will get the mdsmap and discover we're reconnecting via the monitor.
874 reconnect_start
= ceph_clock_now();
875 dout(1) << "reconnect_clients -- " << client_reconnect_gather
.size() << " sessions" << dendl
;
876 mds
->sessionmap
.dump();
879 /* This function DOES put the passed message before returning*/
880 void Server::handle_client_reconnect(MClientReconnect
*m
)
882 dout(7) << "handle_client_reconnect " << m
->get_source() << dendl
;
883 client_t from
= m
->get_source().num();
884 Session
*session
= get_session(m
);
887 if (!mds
->is_reconnect() && mds
->get_want_state() == CEPH_MDS_STATE_RECONNECT
) {
888 dout(10) << " we're almost in reconnect state (mdsmap delivery race?); waiting" << dendl
;
889 mds
->wait_for_reconnect(new C_MDS_RetryMessage(mds
, m
));
893 utime_t delay
= ceph_clock_now();
894 delay
-= reconnect_start
;
895 dout(10) << " reconnect_start " << reconnect_start
<< " delay " << delay
<< dendl
;
898 if (!mds
->is_reconnect()) {
899 // XXX maybe in the future we can do better than this?
900 dout(1) << " no longer in reconnect state, ignoring reconnect, sending close" << dendl
;
901 mds
->clog
->info() << "denied reconnect attempt (mds is "
902 << ceph_mds_state_name(mds
->get_state())
903 << ") from " << m
->get_source_inst()
904 << " after " << delay
<< " (allowed interval " << g_conf
->mds_reconnect_timeout
<< ")";
906 } else if (session
->is_closed()) {
907 dout(1) << " session is closed, ignoring reconnect, sending close" << dendl
;
908 mds
->clog
->info() << "denied reconnect attempt (mds is "
909 << ceph_mds_state_name(mds
->get_state())
910 << ") from " << m
->get_source_inst() << " (session is closed)";
912 } else if (mdcache
->is_readonly()) {
913 dout(1) << " read-only FS, ignoring reconnect, sending close" << dendl
;
914 mds
->clog
->info() << "denied reconnect attempt (mds is read-only)";
919 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_CLOSE
));
924 // notify client of success with an OPEN
925 m
->get_connection()->send_message(new MClientSession(CEPH_SESSION_OPEN
));
926 session
->last_cap_renew
= ceph_clock_now();
927 mds
->clog
->debug() << "reconnect by " << session
->info
.inst
<< " after " << delay
;
930 for (vector
<ceph_mds_snaprealm_reconnect
>::iterator p
= m
->realms
.begin();
931 p
!= m
->realms
.end();
933 CInode
*in
= mdcache
->get_inode(inodeno_t(p
->ino
));
934 if (in
&& in
->state_test(CInode::STATE_PURGING
))
937 assert(in
->snaprealm
);
938 if (in
->snaprealm
->have_past_parents_open()) {
939 dout(15) << "open snaprealm (w/ past parents) on " << *in
<< dendl
;
940 mdcache
->finish_snaprealm_reconnect(from
, in
->snaprealm
, snapid_t(p
->seq
));
942 dout(15) << "open snaprealm (w/o past parents) on " << *in
<< dendl
;
943 mdcache
->add_reconnected_snaprealm(from
, inodeno_t(p
->ino
), snapid_t(p
->seq
));
946 dout(15) << "open snaprealm (w/o inode) on " << inodeno_t(p
->ino
)
947 << " seq " << p
->seq
<< dendl
;
948 mdcache
->add_reconnected_snaprealm(from
, inodeno_t(p
->ino
), snapid_t(p
->seq
));
953 for (map
<inodeno_t
, cap_reconnect_t
>::iterator p
= m
->caps
.begin();
956 // make sure our last_cap_id is MAX over all issued caps
957 if (p
->second
.capinfo
.cap_id
> mdcache
->last_cap_id
)
958 mdcache
->last_cap_id
= p
->second
.capinfo
.cap_id
;
960 CInode
*in
= mdcache
->get_inode(p
->first
);
961 if (in
&& in
->state_test(CInode::STATE_PURGING
))
963 if (in
&& in
->is_auth()) {
964 // we recovered it, and it's ours. take note.
965 dout(15) << "open cap realm " << inodeno_t(p
->second
.capinfo
.snaprealm
)
966 << " on " << *in
<< dendl
;
967 in
->reconnect_cap(from
, p
->second
, session
);
968 mdcache
->add_reconnected_cap(from
, p
->first
, p
->second
);
969 recover_filelocks(in
, p
->second
.flockbl
, m
->get_orig_source().num());
973 if (in
&& !in
->is_auth()) {
975 dout(10) << "non-auth " << *in
<< ", will pass off to authority" << dendl
;
976 // add to cap export list.
977 p
->second
.path
.clear(); // we don't need path
978 mdcache
->rejoin_export_caps(p
->first
, from
, p
->second
,
979 in
->authority().first
);
981 // don't know if the inode is mine
982 dout(10) << "missing ino " << p
->first
<< ", will load later" << dendl
;
983 p
->second
.path
.clear(); // we don't need path
984 mdcache
->rejoin_recovered_caps(p
->first
, from
, p
->second
, MDS_RANK_NONE
);
988 // remove from gather set
989 client_reconnect_gather
.erase(from
);
990 if (client_reconnect_gather
.empty())
991 reconnect_gather_finish();
998 void Server::reconnect_gather_finish()
1000 dout(7) << "reconnect_gather_finish. failed on " << failed_reconnects
<< " clients" << dendl
;
1001 assert(reconnect_done
);
1002 reconnect_done
->complete(0);
1003 reconnect_done
= NULL
;
1006 void Server::reconnect_tick()
1008 if (reconnect_evicting
) {
1009 dout(4) << "reconnect_tick: waiting for evictions" << dendl
;
1013 utime_t reconnect_end
= reconnect_start
;
1014 reconnect_end
+= g_conf
->mds_reconnect_timeout
;
1015 if (ceph_clock_now() >= reconnect_end
&&
1016 !client_reconnect_gather
.empty()) {
1017 dout(10) << "reconnect timed out" << dendl
;
1019 // If we're doing blacklist evictions, use this to wait for them before
1020 // proceeding to reconnect_gather_finish
1021 MDSGatherBuilder
gather(g_ceph_context
);
1023 for (set
<client_t
>::iterator p
= client_reconnect_gather
.begin();
1024 p
!= client_reconnect_gather
.end();
1026 Session
*session
= mds
->sessionmap
.get_session(entity_name_t::CLIENT(p
->v
));
1028 dout(1) << "reconnect gave up on " << session
->info
.inst
<< dendl
;
1030 mds
->clog
->warn() << "evicting unresponsive client " << *session
1031 << ", after waiting " << g_conf
->mds_reconnect_timeout
1032 << " seconds during MDS startup";
1034 if (g_conf
->mds_session_blacklist_on_timeout
) {
1035 std::stringstream ss
;
1036 mds
->evict_client(session
->info
.inst
.name
.num(), false, true, ss
,
1039 kill_session(session
, NULL
);
1042 failed_reconnects
++;
1044 client_reconnect_gather
.clear();
1046 if (gather
.has_subs()) {
1047 dout(1) << "reconnect will complete once clients are evicted" << dendl
;
1048 gather
.set_finisher(new MDSInternalContextWrapper(mds
, new FunctionContext(
1049 [this](int r
){reconnect_gather_finish();})));
1051 reconnect_evicting
= true;
1053 reconnect_gather_finish();
1058 void Server::recover_filelocks(CInode
*in
, bufferlist locks
, int64_t client
)
1060 if (!locks
.length()) return;
1063 bufferlist::iterator p
= locks
.begin();
1064 ::decode(numlocks
, p
);
1065 for (int i
= 0; i
< numlocks
; ++i
) {
1067 lock
.client
= client
;
1068 in
->get_fcntl_lock_state()->held_locks
.insert(pair
<uint64_t, ceph_filelock
>(lock
.start
, lock
));
1069 ++in
->get_fcntl_lock_state()->client_held_lock_counts
[client
];
1071 ::decode(numlocks
, p
);
1072 for (int i
= 0; i
< numlocks
; ++i
) {
1074 lock
.client
= client
;
1075 in
->get_flock_lock_state()->held_locks
.insert(pair
<uint64_t, ceph_filelock
> (lock
.start
, lock
));
1076 ++in
->get_flock_lock_state()->client_held_lock_counts
[client
];
1082 * Call this when the MDCache is oversized, to send requests to the clients
1083 * to trim some caps, and consequently unpin some inodes in the MDCache so
1084 * that it can trim too.
1086 void Server::recall_client_state(float ratio
)
1088 int max_caps_per_client
= (int)(g_conf
->mds_cache_size
* .8);
1089 int min_caps_per_client
= 100;
1091 dout(10) << "recall_client_state " << ratio
1092 << ", caps per client " << min_caps_per_client
<< "-" << max_caps_per_client
1095 set
<Session
*> sessions
;
1096 mds
->sessionmap
.get_client_session_set(sessions
);
1097 for (set
<Session
*>::const_iterator p
= sessions
.begin();
1098 p
!= sessions
.end();
1100 Session
*session
= *p
;
1101 if (!session
->is_open() ||
1102 !session
->info
.inst
.name
.is_client())
1105 dout(10) << " session " << session
->info
.inst
1106 << " caps " << session
->caps
.size()
1107 << ", leases " << session
->leases
.size()
1110 if (session
->caps
.size() > min_caps_per_client
) {
1111 int newlim
= MIN((int)(session
->caps
.size() * ratio
), max_caps_per_client
);
1112 if (session
->caps
.size() > newlim
) {
1113 MClientSession
*m
= new MClientSession(CEPH_SESSION_RECALL_STATE
);
1114 m
->head
.max_caps
= newlim
;
1115 mds
->send_message_client(m
, session
);
1116 session
->notify_recall_sent(newlim
);
1122 void Server::force_clients_readonly()
1124 dout(10) << "force_clients_readonly" << dendl
;
1125 set
<Session
*> sessions
;
1126 mds
->sessionmap
.get_client_session_set(sessions
);
1127 for (set
<Session
*>::const_iterator p
= sessions
.begin();
1128 p
!= sessions
.end();
1130 Session
*session
= *p
;
1131 if (!session
->info
.inst
.name
.is_client() ||
1132 !(session
->is_open() || session
->is_stale()))
1134 mds
->send_message_client(new MClientSession(CEPH_SESSION_FORCE_RO
), session
);
1139 * some generic stuff for finishing off requests
1141 void Server::journal_and_reply(MDRequestRef
& mdr
, CInode
*in
, CDentry
*dn
, LogEvent
*le
, MDSLogContextBase
*fin
)
1143 dout(10) << "journal_and_reply tracei " << in
<< " tracedn " << dn
<< dendl
;
1144 assert(!mdr
->has_completed
);
1146 // note trace items for eventual reply.
1155 early_reply(mdr
, in
, dn
);
1157 mdr
->committing
= true;
1158 submit_mdlog_entry(le
, fin
, mdr
, __func__
);
1160 if (mdr
->client_request
&& mdr
->client_request
->is_queued_for_replay()) {
1161 if (mds
->queue_one_replay()) {
1162 dout(10) << " queued next replay op" << dendl
;
1164 dout(10) << " journaled last replay op, flushing" << dendl
;
1167 } else if (mdr
->did_early_reply
)
1168 mds
->locker
->drop_rdlocks(mdr
.get());
1173 void Server::submit_mdlog_entry(LogEvent
*le
, MDSLogContextBase
*fin
, MDRequestRef
& mdr
,
1177 string
event_str("submit entry: ");
1179 mdr
->mark_event_string(event_str
);
1181 mdlog
->submit_entry(le
, fin
);
1185 * send response built from mdr contents and error code; clean up mdr
1187 void Server::respond_to_request(MDRequestRef
& mdr
, int r
)
1189 if (mdr
->client_request
) {
1190 reply_client_request(mdr
, new MClientReply(mdr
->client_request
, r
));
1192 // add here to avoid counting ops multiple times (e.g., locks, loading)
1193 switch(mdr
->client_request
->get_op()) {
1194 case CEPH_MDS_OP_LOOKUPHASH
:
1195 logger
->inc(l_mdss_req_lookuphash
);
1197 case CEPH_MDS_OP_LOOKUPINO
:
1198 logger
->inc(l_mdss_req_lookupino
);
1200 case CEPH_MDS_OP_LOOKUPPARENT
:
1201 logger
->inc(l_mdss_req_lookupparent
);
1203 case CEPH_MDS_OP_LOOKUPNAME
:
1204 logger
->inc(l_mdss_req_lookupname
);
1206 case CEPH_MDS_OP_LOOKUP
:
1207 logger
->inc(l_mdss_req_lookup
);
1209 case CEPH_MDS_OP_LOOKUPSNAP
:
1210 logger
->inc(l_mdss_req_lookupsnap
);
1212 case CEPH_MDS_OP_GETATTR
:
1213 logger
->inc(l_mdss_req_getattr
);
1215 case CEPH_MDS_OP_SETATTR
:
1216 logger
->inc(l_mdss_req_setattr
);
1218 case CEPH_MDS_OP_SETLAYOUT
:
1219 logger
->inc(l_mdss_req_setlayout
);
1221 case CEPH_MDS_OP_SETDIRLAYOUT
:
1222 logger
->inc(l_mdss_req_setdirlayout
);
1224 case CEPH_MDS_OP_SETXATTR
:
1225 logger
->inc(l_mdss_req_setxattr
);
1227 case CEPH_MDS_OP_RMXATTR
:
1228 logger
->inc(l_mdss_req_rmxattr
);
1230 case CEPH_MDS_OP_READDIR
:
1231 logger
->inc(l_mdss_req_readdir
);
1233 case CEPH_MDS_OP_SETFILELOCK
:
1234 logger
->inc(l_mdss_req_setfilelock
);
1236 case CEPH_MDS_OP_GETFILELOCK
:
1237 logger
->inc(l_mdss_req_getfilelock
);
1239 case CEPH_MDS_OP_CREATE
:
1240 logger
->inc(l_mdss_req_create
);
1241 case CEPH_MDS_OP_OPEN
:
1242 logger
->inc(l_mdss_req_open
);
1244 case CEPH_MDS_OP_MKNOD
:
1245 logger
->inc(l_mdss_req_mknod
);
1247 case CEPH_MDS_OP_LINK
:
1248 logger
->inc(l_mdss_req_link
);
1250 case CEPH_MDS_OP_UNLINK
:
1251 logger
->inc(l_mdss_req_unlink
);
1253 case CEPH_MDS_OP_RMDIR
:
1254 logger
->inc(l_mdss_req_rmdir
);
1256 case CEPH_MDS_OP_RENAME
:
1257 logger
->inc(l_mdss_req_rename
);
1259 case CEPH_MDS_OP_MKDIR
:
1260 logger
->inc(l_mdss_req_mkdir
);
1262 case CEPH_MDS_OP_SYMLINK
:
1263 logger
->inc(l_mdss_req_symlink
);
1265 case CEPH_MDS_OP_LSSNAP
:
1266 logger
->inc(l_mdss_req_lssnap
);
1268 case CEPH_MDS_OP_MKSNAP
:
1269 logger
->inc(l_mdss_req_mksnap
);
1271 case CEPH_MDS_OP_RMSNAP
:
1272 logger
->inc(l_mdss_req_rmsnap
);
1274 case CEPH_MDS_OP_RENAMESNAP
:
1275 logger
->inc(l_mdss_req_renamesnap
);
1278 } else if (mdr
->internal_op
> -1) {
1279 dout(10) << "respond_to_request on internal request " << mdr
<< dendl
;
1280 if (!mdr
->internal_op_finish
)
1281 assert(0 == "trying to respond to internal op without finisher");
1282 mdr
->internal_op_finish
->complete(r
);
1283 mdcache
->request_finish(mdr
);
1287 void Server::early_reply(MDRequestRef
& mdr
, CInode
*tracei
, CDentry
*tracedn
)
1289 if (!g_conf
->mds_early_reply
)
1292 if (mdr
->has_more() && mdr
->more()->has_journaled_slaves
) {
1293 dout(10) << "early_reply - there are journaled slaves, not allowed." << dendl
;
1297 if (mdr
->alloc_ino
) {
1298 dout(10) << "early_reply - allocated ino, not allowed" << dendl
;
1302 MClientRequest
*req
= mdr
->client_request
;
1303 entity_inst_t client_inst
= req
->get_source_inst();
1304 if (client_inst
.name
.is_mds())
1307 if (req
->is_replay()) {
1308 dout(10) << " no early reply on replay op" << dendl
;
1313 MClientReply
*reply
= new MClientReply(req
, 0);
1314 reply
->set_unsafe();
1316 // mark xlocks "done", indicating that we are exposing uncommitted changes.
1318 //_rename_finish() does not send dentry link/unlink message to replicas.
1319 // so do not set xlocks on dentries "done", the xlocks prevent dentries
1320 // that have projected linkages from getting new replica.
1321 mds
->locker
->set_xlocks_done(mdr
.get(), req
->get_op() == CEPH_MDS_OP_RENAME
);
1323 dout(10) << "early_reply " << reply
->get_result()
1324 << " (" << cpp_strerror(reply
->get_result())
1325 << ") " << *req
<< dendl
;
1327 if (tracei
|| tracedn
) {
1329 mdr
->cap_releases
.erase(tracei
->vino());
1331 mdr
->cap_releases
.erase(tracedn
->get_dir()->get_inode()->vino());
1333 set_trace_dist(mdr
->session
, reply
, tracei
, tracedn
, mdr
->snapid
,
1334 req
->get_dentry_wanted(), mdr
);
1337 reply
->set_extra_bl(mdr
->reply_extra_bl
);
1338 req
->get_connection()->send_message(reply
);
1340 mdr
->did_early_reply
= true;
1342 mds
->logger
->inc(l_mds_reply
);
1343 utime_t lat
= ceph_clock_now() - req
->get_recv_stamp();
1344 mds
->logger
->tinc(l_mds_reply_latency
, lat
);
1345 dout(20) << "lat " << lat
<< dendl
;
1347 mdr
->mark_event("early_replied");
1352 * include a trace to tracei
1355 void Server::reply_client_request(MDRequestRef
& mdr
, MClientReply
*reply
)
1358 MClientRequest
*req
= mdr
->client_request
;
1360 dout(7) << "reply_client_request " << reply
->get_result()
1361 << " (" << cpp_strerror(reply
->get_result())
1362 << ") " << *req
<< dendl
;
1364 mdr
->mark_event("replying");
1366 Session
*session
= mdr
->session
;
1368 // note successful request in session map?
1370 // setfilelock requests are special, they only modify states in MDS memory.
1371 // The states get lost when MDS fails. If Client re-send a completed
1372 // setfilelock request, it means that client did not receive corresponding
1373 // setfilelock reply. So MDS should re-execute the setfilelock request.
1374 if (req
->may_write() && req
->get_op() != CEPH_MDS_OP_SETFILELOCK
&&
1375 reply
->get_result() == 0 && session
) {
1376 inodeno_t created
= mdr
->alloc_ino
? mdr
->alloc_ino
: mdr
->used_prealloc_ino
;
1377 session
->add_completed_request(mdr
->reqid
.tid
, created
);
1379 mdr
->ls
->touched_sessions
.insert(session
->info
.inst
.name
);
1383 // give any preallocated inos to the session
1384 apply_allocated_inos(mdr
, session
);
1386 // get tracei/tracedn from mdr?
1387 snapid_t snapid
= mdr
->snapid
;
1388 CInode
*tracei
= mdr
->tracei
;
1389 CDentry
*tracedn
= mdr
->tracedn
;
1391 bool is_replay
= mdr
->client_request
->is_replay();
1392 bool did_early_reply
= mdr
->did_early_reply
;
1393 entity_inst_t client_inst
= req
->get_source_inst();
1394 int dentry_wanted
= req
->get_dentry_wanted();
1396 if (!did_early_reply
&& !is_replay
) {
1398 mds
->logger
->inc(l_mds_reply
);
1399 utime_t lat
= ceph_clock_now() - mdr
->client_request
->get_recv_stamp();
1400 mds
->logger
->tinc(l_mds_reply_latency
, lat
);
1401 dout(20) << "lat " << lat
<< dendl
;
1404 mdr
->cap_releases
.erase(tracei
->vino());
1406 mdr
->cap_releases
.erase(tracedn
->get_dir()->get_inode()->vino());
1409 // drop non-rdlocks before replying, so that we can issue leases
1410 mdcache
->request_drop_non_rdlocks(mdr
);
1413 if (client_inst
.name
.is_mds() || !session
) {
1414 reply
->put(); // mds doesn't need a reply
1418 if (!did_early_reply
&& // don't issue leases if we sent an earlier reply already
1419 (tracei
|| tracedn
)) {
1422 mdcache
->try_reconnect_cap(tracei
, session
);
1424 // include metadata in reply
1425 set_trace_dist(session
, reply
, tracei
, tracedn
,
1426 snapid
, dentry_wanted
,
1431 // We can set the extra bl unconditionally: if it's already been sent in the
1432 // early_reply, set_extra_bl will have claimed it and reply_extra_bl is empty
1433 reply
->set_extra_bl(mdr
->reply_extra_bl
);
1435 reply
->set_mdsmap_epoch(mds
->mdsmap
->get_epoch());
1436 req
->get_connection()->send_message(reply
);
1439 if (req
->is_queued_for_replay() &&
1440 (mdr
->has_completed
|| reply
->get_result() < 0)) {
1441 if (reply
->get_result() < 0) {
1442 int r
= reply
->get_result();
1443 derr
<< "reply_client_request: failed to replay " << *req
1444 << " error " << r
<< " (" << cpp_strerror(r
) << ")" << dendl
;
1445 mds
->clog
->warn() << "failed to replay " << req
->get_reqid() << " error " << r
;
1447 mds
->queue_one_replay();
1451 mdcache
->request_finish(mdr
);
1453 // take a closer look at tracei, if it happens to be a remote link
1456 tracedn
->get_projected_linkage()->is_remote()) {
1457 mdcache
->eval_remote(tracedn
);
1462 void Server::encode_empty_dirstat(bufferlist
& bl
)
1464 static DirStat empty
;
1468 void Server::encode_infinite_lease(bufferlist
& bl
)
1475 dout(20) << "encode_infinite_lease " << e
<< dendl
;
1478 void Server::encode_null_lease(bufferlist
& bl
)
1485 dout(20) << "encode_null_lease " << e
<< dendl
;
1490 * pass inode OR dentry (not both, or we may get confused)
1492 * trace is in reverse order (i.e. root inode comes last)
1494 void Server::set_trace_dist(Session
*session
, MClientReply
*reply
,
1495 CInode
*in
, CDentry
*dn
,
1500 // skip doing this for debugging purposes?
1501 if (g_conf
->mds_inject_traceless_reply_probability
&&
1502 mdr
->ls
&& !mdr
->o_trunc
&&
1503 (rand() % 10000 < g_conf
->mds_inject_traceless_reply_probability
* 10000.0)) {
1504 dout(5) << "deliberately skipping trace for " << *reply
<< dendl
;
1508 // inode, dentry, dir, ..., inode
1510 mds_rank_t whoami
= mds
->get_nodeid();
1511 client_t client
= session
->get_client();
1512 utime_t now
= ceph_clock_now();
1514 dout(20) << "set_trace_dist snapid " << snapid
<< dendl
;
1516 //assert((bool)dn == (bool)dentry_wanted); // not true for snapshot lookups
1519 if (snapid
== CEPH_NOSNAP
) {
1522 realm
= in
->find_snaprealm();
1524 realm
= dn
->get_dir()->get_inode()->find_snaprealm();
1525 reply
->snapbl
= realm
->get_snap_trace();
1526 dout(10) << "set_trace_dist snaprealm " << *realm
<< " len=" << reply
->snapbl
.length() << dendl
;
1531 reply
->head
.is_dentry
= 1;
1532 CDir
*dir
= dn
->get_dir();
1533 CInode
*diri
= dir
->get_inode();
1535 diri
->encode_inodestat(bl
, session
, NULL
, snapid
);
1536 dout(20) << "set_trace_dist added diri " << *diri
<< dendl
;
1538 #ifdef MDS_VERIFY_FRAGSTAT
1539 if (dir
->is_complete())
1540 dir
->verify_fragstat();
1542 dir
->encode_dirstat(bl
, whoami
);
1543 dout(20) << "set_trace_dist added dir " << *dir
<< dendl
;
1545 ::encode(dn
->get_name(), bl
);
1546 if (snapid
== CEPH_NOSNAP
)
1547 mds
->locker
->issue_client_lease(dn
, client
, bl
, now
, session
);
1549 encode_null_lease(bl
);
1550 dout(20) << "set_trace_dist added dn " << snapid
<< " " << *dn
<< dendl
;
1552 reply
->head
.is_dentry
= 0;
1556 in
->encode_inodestat(bl
, session
, NULL
, snapid
, 0, mdr
->getattr_caps
);
1557 dout(20) << "set_trace_dist added in " << *in
<< dendl
;
1558 reply
->head
.is_target
= 1;
1560 reply
->head
.is_target
= 0;
1562 reply
->set_trace(bl
);
1569 * process a client request
1570 * This function DOES put the passed message before returning
1572 void Server::handle_client_request(MClientRequest
*req
)
1574 dout(4) << "handle_client_request " << *req
<< dendl
;
1577 mds
->logger
->inc(l_mds_request
);
1579 logger
->inc(l_mdss_handle_client_request
);
1581 if (!mdcache
->is_open()) {
1582 dout(5) << "waiting for root" << dendl
;
1583 mdcache
->wait_for_open(new C_MDS_RetryMessage(mds
, req
));
1588 Session
*session
= 0;
1589 if (req
->get_source().is_client()) {
1590 session
= get_session(req
);
1592 dout(5) << "no session for " << req
->get_source() << ", dropping" << dendl
;
1593 } else if (session
->is_closed() ||
1594 session
->is_closing() ||
1595 session
->is_killing()) {
1596 dout(5) << "session closed|closing|killing, dropping" << dendl
;
1600 if (req
->is_queued_for_replay())
1601 mds
->queue_one_replay();
1608 if (req
->get_mdsmap_epoch() < mds
->mdsmap
->get_epoch()) {
1609 // send it? hrm, this isn't ideal; they may get a lot of copies if
1610 // they have a high request rate.
1613 // completed request?
1614 bool has_completed
= false;
1615 if (req
->is_replay() || req
->get_retry_attempt()) {
1618 if (session
->have_completed_request(req
->get_reqid().tid
, &created
)) {
1619 has_completed
= true;
1620 // Don't send traceless reply if the completed request has created
1621 // new inode. Treat the request as lookup request instead.
1622 if (req
->is_replay() ||
1623 ((created
== inodeno_t() || !mds
->is_clientreplay()) &&
1624 req
->get_op() != CEPH_MDS_OP_OPEN
&&
1625 req
->get_op() != CEPH_MDS_OP_CREATE
)) {
1626 dout(5) << "already completed " << req
->get_reqid() << dendl
;
1627 MClientReply
*reply
= new MClientReply(req
, 0);
1628 if (created
!= inodeno_t()) {
1630 ::encode(created
, extra
);
1631 reply
->set_extra_bl(extra
);
1633 req
->get_connection()->send_message(reply
);
1635 if (req
->is_queued_for_replay())
1636 mds
->queue_one_replay();
1641 if (req
->get_op() != CEPH_MDS_OP_OPEN
&&
1642 req
->get_op() != CEPH_MDS_OP_CREATE
) {
1643 dout(10) << " completed request which created new inode " << created
1644 << ", convert it to lookup request" << dendl
;
1645 req
->head
.op
= req
->get_dentry_wanted() ? CEPH_MDS_OP_LOOKUP
: CEPH_MDS_OP_GETATTR
;
1646 req
->head
.args
.getattr
.mask
= CEPH_STAT_CAP_INODE_ALL
;
1651 // trim completed_request list
1652 if (req
->get_oldest_client_tid() > 0) {
1653 dout(15) << " oldest_client_tid=" << req
->get_oldest_client_tid() << dendl
;
1655 if (session
->trim_completed_requests(req
->get_oldest_client_tid())) {
1656 // Sessions 'completed_requests' was dirtied, mark it to be
1657 // potentially flushed at segment expiry.
1658 mdlog
->get_current_segment()->touched_sessions
.insert(session
->info
.inst
.name
);
1660 if (session
->get_num_trim_requests_warnings() > 0 &&
1661 session
->get_num_completed_requests() * 2 < g_conf
->mds_max_completed_requests
)
1662 session
->reset_num_trim_requests_warnings();
1664 if (session
->get_num_completed_requests() >=
1665 (g_conf
->mds_max_completed_requests
<< session
->get_num_trim_requests_warnings())) {
1666 session
->inc_num_trim_requests_warnings();
1668 ss
<< "client." << session
->get_client() << " does not advance its oldest_client_tid ("
1669 << req
->get_oldest_client_tid() << "), "
1670 << session
->get_num_completed_requests()
1671 << " completed requests recorded in session\n";
1672 mds
->clog
->warn() << ss
.str();
1673 dout(20) << __func__
<< " " << ss
.str() << dendl
;
1678 // register + dispatch
1679 MDRequestRef mdr
= mdcache
->request_start(req
);
1684 mdr
->session
= session
;
1685 session
->requests
.push_back(&mdr
->item_session_request
);
1689 mdr
->has_completed
= true;
1691 // process embedded cap releases?
1692 // (only if NOT replay!)
1693 if (!req
->releases
.empty() && req
->get_source().is_client() && !req
->is_replay()) {
1694 client_t client
= req
->get_source().num();
1695 for (vector
<MClientRequest::Release
>::iterator p
= req
->releases
.begin();
1696 p
!= req
->releases
.end();
1698 mds
->locker
->process_request_cap_release(mdr
, client
, p
->item
, p
->dname
);
1699 req
->releases
.clear();
1702 dispatch_client_request(mdr
);
1706 void Server::handle_osd_map()
1708 /* Note that we check the OSDMAP_FULL flag directly rather than
1709 * using osdmap_full_flag(), because we want to know "is the flag set"
1710 * rather than "does the flag apply to us?" */
1711 mds
->objecter
->with_osdmap([this](const OSDMap
& o
) {
1712 is_full
= o
.test_flag(CEPH_OSDMAP_FULL
);
1713 dout(7) << __func__
<< ": full = " << is_full
<< " epoch = "
1714 << o
.get_epoch() << dendl
;
1718 void Server::dispatch_client_request(MDRequestRef
& mdr
)
1720 // we shouldn't be waiting on anyone.
1721 assert(!mdr
->has_more() || mdr
->more()->waiting_on_slave
.empty());
1724 dout(10) << "request " << *mdr
<< " was killed" << dendl
;
1728 MClientRequest
*req
= mdr
->client_request
;
1730 if (logger
) logger
->inc(l_mdss_dispatch_client_request
);
1732 dout(7) << "dispatch_client_request " << *req
<< dendl
;
1734 if (req
->may_write()) {
1735 if (mdcache
->is_readonly()) {
1736 dout(10) << " read-only FS" << dendl
;
1737 respond_to_request(mdr
, -EROFS
);
1740 if (mdr
->has_more() && mdr
->more()->slave_error
) {
1741 dout(10) << " got error from slaves" << dendl
;
1742 respond_to_request(mdr
, mdr
->more()->slave_error
);
1748 if (req
->get_op() == CEPH_MDS_OP_SETLAYOUT
||
1749 req
->get_op() == CEPH_MDS_OP_SETDIRLAYOUT
||
1750 req
->get_op() == CEPH_MDS_OP_SETLAYOUT
||
1751 req
->get_op() == CEPH_MDS_OP_RMXATTR
||
1752 req
->get_op() == CEPH_MDS_OP_SETXATTR
||
1753 req
->get_op() == CEPH_MDS_OP_CREATE
||
1754 req
->get_op() == CEPH_MDS_OP_SYMLINK
||
1755 req
->get_op() == CEPH_MDS_OP_MKSNAP
||
1756 ((req
->get_op() == CEPH_MDS_OP_LINK
||
1757 req
->get_op() == CEPH_MDS_OP_RENAME
) &&
1758 (!mdr
->has_more() || mdr
->more()->witnessed
.empty())) // haven't started slave request
1761 dout(20) << __func__
<< ": full, responding ENOSPC to op " << ceph_mds_op_name(req
->get_op()) << dendl
;
1762 respond_to_request(mdr
, -ENOSPC
);
1765 dout(20) << __func__
<< ": full, permitting op " << ceph_mds_op_name(req
->get_op()) << dendl
;
1769 switch (req
->get_op()) {
1770 case CEPH_MDS_OP_LOOKUPHASH
:
1771 case CEPH_MDS_OP_LOOKUPINO
:
1772 handle_client_lookup_ino(mdr
, false, false);
1774 case CEPH_MDS_OP_LOOKUPPARENT
:
1775 handle_client_lookup_ino(mdr
, true, false);
1777 case CEPH_MDS_OP_LOOKUPNAME
:
1778 handle_client_lookup_ino(mdr
, false, true);
1782 case CEPH_MDS_OP_LOOKUP
:
1783 handle_client_getattr(mdr
, true);
1786 case CEPH_MDS_OP_LOOKUPSNAP
:
1787 // lookupsnap does not reference a CDentry; treat it as a getattr
1788 case CEPH_MDS_OP_GETATTR
:
1789 handle_client_getattr(mdr
, false);
1792 case CEPH_MDS_OP_SETATTR
:
1793 handle_client_setattr(mdr
);
1795 case CEPH_MDS_OP_SETLAYOUT
:
1796 handle_client_setlayout(mdr
);
1798 case CEPH_MDS_OP_SETDIRLAYOUT
:
1799 handle_client_setdirlayout(mdr
);
1801 case CEPH_MDS_OP_SETXATTR
:
1802 handle_client_setxattr(mdr
);
1804 case CEPH_MDS_OP_RMXATTR
:
1805 handle_client_removexattr(mdr
);
1808 case CEPH_MDS_OP_READDIR
:
1809 handle_client_readdir(mdr
);
1812 case CEPH_MDS_OP_SETFILELOCK
:
1813 handle_client_file_setlock(mdr
);
1816 case CEPH_MDS_OP_GETFILELOCK
:
1817 handle_client_file_readlock(mdr
);
1821 case CEPH_MDS_OP_CREATE
:
1822 if (mdr
->has_completed
)
1823 handle_client_open(mdr
); // already created.. just open
1825 handle_client_openc(mdr
);
1828 case CEPH_MDS_OP_OPEN
:
1829 handle_client_open(mdr
);
1834 case CEPH_MDS_OP_MKNOD
:
1835 handle_client_mknod(mdr
);
1837 case CEPH_MDS_OP_LINK
:
1838 handle_client_link(mdr
);
1840 case CEPH_MDS_OP_UNLINK
:
1841 case CEPH_MDS_OP_RMDIR
:
1842 handle_client_unlink(mdr
);
1844 case CEPH_MDS_OP_RENAME
:
1845 handle_client_rename(mdr
);
1847 case CEPH_MDS_OP_MKDIR
:
1848 handle_client_mkdir(mdr
);
1850 case CEPH_MDS_OP_SYMLINK
:
1851 handle_client_symlink(mdr
);
1856 case CEPH_MDS_OP_LSSNAP
:
1857 handle_client_lssnap(mdr
);
1859 case CEPH_MDS_OP_MKSNAP
:
1860 handle_client_mksnap(mdr
);
1862 case CEPH_MDS_OP_RMSNAP
:
1863 handle_client_rmsnap(mdr
);
1865 case CEPH_MDS_OP_RENAMESNAP
:
1866 handle_client_renamesnap(mdr
);
1870 dout(1) << " unknown client op " << req
->get_op() << dendl
;
1871 respond_to_request(mdr
, -EOPNOTSUPP
);
1876 // ---------------------------------------
1879 /* This function DOES put the passed message before returning*/
1880 void Server::handle_slave_request(MMDSSlaveRequest
*m
)
1882 dout(4) << "handle_slave_request " << m
->get_reqid() << " from " << m
->get_source() << dendl
;
1883 mds_rank_t from
= mds_rank_t(m
->get_source().num());
1885 if (logger
) logger
->inc(l_mdss_handle_slave_request
);
1889 return handle_slave_request_reply(m
);
1891 // the purpose of rename notify is enforcing causal message ordering. making sure
1892 // bystanders have received all messages from rename srcdn's auth MDS.
1893 if (m
->get_op() == MMDSSlaveRequest::OP_RENAMENOTIFY
) {
1894 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(m
->get_reqid(), m
->get_attempt(),
1895 MMDSSlaveRequest::OP_RENAMENOTIFYACK
);
1896 mds
->send_message(reply
, m
->get_connection());
1901 CDentry
*straydn
= NULL
;
1902 if (m
->stray
.length() > 0) {
1903 straydn
= mdcache
->add_replica_stray(m
->stray
, from
);
1908 // am i a new slave?
1910 if (mdcache
->have_request(m
->get_reqid())) {
1912 mdr
= mdcache
->request_get(m
->get_reqid());
1914 // is my request newer?
1915 if (mdr
->attempt
> m
->get_attempt()) {
1916 dout(10) << "local request " << *mdr
<< " attempt " << mdr
->attempt
<< " > " << m
->get_attempt()
1917 << ", dropping " << *m
<< dendl
;
1923 if (mdr
->attempt
< m
->get_attempt()) {
1924 // mine is old, close it out
1925 dout(10) << "local request " << *mdr
<< " attempt " << mdr
->attempt
<< " < " << m
->get_attempt()
1926 << ", closing out" << dendl
;
1927 mdcache
->request_finish(mdr
);
1929 } else if (mdr
->slave_to_mds
!= from
) {
1930 dout(10) << "local request " << *mdr
<< " not slave to mds." << from
<< dendl
;
1935 if (m
->get_op() == MMDSSlaveRequest::OP_FINISH
&& m
->is_abort()) {
1936 mdr
->aborted
= true;
1937 if (mdr
->slave_request
) {
1938 // only abort on-going xlock, wrlock and auth pin
1939 assert(!mdr
->slave_did_prepare());
1941 mdcache
->request_finish(mdr
);
1948 if (m
->get_op() == MMDSSlaveRequest::OP_FINISH
) {
1949 dout(10) << "missing slave request for " << m
->get_reqid()
1950 << " OP_FINISH, must have lost race with a forward" << dendl
;
1954 mdr
= mdcache
->request_start_slave(m
->get_reqid(), m
->get_attempt(), m
);
1955 mdr
->set_op_stamp(m
->op_stamp
);
1957 assert(mdr
->slave_request
== 0); // only one at a time, please!
1961 mdr
->straydn
= straydn
;
1964 if (!mds
->is_clientreplay() && !mds
->is_active() && !mds
->is_stopping()) {
1965 dout(3) << "not clientreplay|active yet, waiting" << dendl
;
1966 mds
->wait_for_replay(new C_MDS_RetryMessage(mds
, m
));
1968 } else if (mds
->is_clientreplay() && !mds
->mdsmap
->is_clientreplay(from
) &&
1969 mdr
->locks
.empty()) {
1970 dout(3) << "not active yet, waiting" << dendl
;
1971 mds
->wait_for_active(new C_MDS_RetryMessage(mds
, m
));
1975 mdr
->slave_request
= m
;
1977 dispatch_slave_request(mdr
);
1980 /* This function DOES put the passed message before returning*/
1981 void Server::handle_slave_request_reply(MMDSSlaveRequest
*m
)
1983 mds_rank_t from
= mds_rank_t(m
->get_source().num());
1985 if (!mds
->is_clientreplay() && !mds
->is_active() && !mds
->is_stopping()) {
1986 metareqid_t r
= m
->get_reqid();
1987 if (!mdcache
->have_uncommitted_master(r
, from
)) {
1988 dout(10) << "handle_slave_request_reply ignoring slave reply from mds."
1989 << from
<< " reqid " << r
<< dendl
;
1993 dout(3) << "not clientreplay|active yet, waiting" << dendl
;
1994 mds
->wait_for_replay(new C_MDS_RetryMessage(mds
, m
));
1998 if (m
->get_op() == MMDSSlaveRequest::OP_COMMITTED
) {
1999 metareqid_t r
= m
->get_reqid();
2000 mdcache
->committed_master_slave(r
, from
);
2005 MDRequestRef mdr
= mdcache
->request_get(m
->get_reqid());
2006 if (m
->get_attempt() != mdr
->attempt
) {
2007 dout(10) << "handle_slave_request_reply " << *mdr
<< " ignoring reply from other attempt "
2008 << m
->get_attempt() << dendl
;
2013 switch (m
->get_op()) {
2014 case MMDSSlaveRequest::OP_XLOCKACK
:
2016 // identify lock, master request
2017 SimpleLock
*lock
= mds
->locker
->get_lock(m
->get_lock_type(),
2018 m
->get_object_info());
2019 mdr
->more()->slaves
.insert(from
);
2020 dout(10) << "got remote xlock on " << *lock
<< " on " << *lock
->get_parent() << dendl
;
2021 mdr
->xlocks
.insert(lock
);
2022 mdr
->locks
.insert(lock
);
2023 mdr
->finish_locking(lock
);
2024 lock
->get_xlock(mdr
, mdr
->get_client());
2026 assert(mdr
->more()->waiting_on_slave
.count(from
));
2027 mdr
->more()->waiting_on_slave
.erase(from
);
2028 assert(mdr
->more()->waiting_on_slave
.empty());
2029 mdcache
->dispatch_request(mdr
);
2033 case MMDSSlaveRequest::OP_WRLOCKACK
:
2035 // identify lock, master request
2036 SimpleLock
*lock
= mds
->locker
->get_lock(m
->get_lock_type(),
2037 m
->get_object_info());
2038 mdr
->more()->slaves
.insert(from
);
2039 dout(10) << "got remote wrlock on " << *lock
<< " on " << *lock
->get_parent() << dendl
;
2040 mdr
->remote_wrlocks
[lock
] = from
;
2041 mdr
->locks
.insert(lock
);
2042 mdr
->finish_locking(lock
);
2044 assert(mdr
->more()->waiting_on_slave
.count(from
));
2045 mdr
->more()->waiting_on_slave
.erase(from
);
2046 assert(mdr
->more()->waiting_on_slave
.empty());
2047 mdcache
->dispatch_request(mdr
);
2051 case MMDSSlaveRequest::OP_AUTHPINACK
:
2052 handle_slave_auth_pin_ack(mdr
, m
);
2055 case MMDSSlaveRequest::OP_LINKPREPACK
:
2056 handle_slave_link_prep_ack(mdr
, m
);
2059 case MMDSSlaveRequest::OP_RMDIRPREPACK
:
2060 handle_slave_rmdir_prep_ack(mdr
, m
);
2063 case MMDSSlaveRequest::OP_RENAMEPREPACK
:
2064 handle_slave_rename_prep_ack(mdr
, m
);
2067 case MMDSSlaveRequest::OP_RENAMENOTIFYACK
:
2068 handle_slave_rename_notify_ack(mdr
, m
);
2079 /* This function DOES put the mdr->slave_request before returning*/
2080 void Server::dispatch_slave_request(MDRequestRef
& mdr
)
2082 dout(7) << "dispatch_slave_request " << *mdr
<< " " << *mdr
->slave_request
<< dendl
;
2085 dout(7) << " abort flag set, finishing" << dendl
;
2086 mdcache
->request_finish(mdr
);
2090 if (logger
) logger
->inc(l_mdss_dispatch_slave_request
);
2092 int op
= mdr
->slave_request
->get_op();
2094 case MMDSSlaveRequest::OP_XLOCK
:
2095 case MMDSSlaveRequest::OP_WRLOCK
:
2098 SimpleLock
*lock
= mds
->locker
->get_lock(mdr
->slave_request
->get_lock_type(),
2099 mdr
->slave_request
->get_object_info());
2102 dout(10) << "don't have object, dropping" << dendl
;
2103 ceph_abort(); // can this happen, if we auth pinned properly.
2105 if (op
== MMDSSlaveRequest::OP_XLOCK
&& !lock
->get_parent()->is_auth()) {
2106 dout(10) << "not auth for remote xlock attempt, dropping on "
2107 << *lock
<< " on " << *lock
->get_parent() << dendl
;
2109 // use acquire_locks so that we get auth_pinning.
2110 set
<SimpleLock
*> rdlocks
;
2111 set
<SimpleLock
*> wrlocks
= mdr
->wrlocks
;
2112 set
<SimpleLock
*> xlocks
= mdr
->xlocks
;
2116 case MMDSSlaveRequest::OP_XLOCK
:
2117 xlocks
.insert(lock
);
2118 replycode
= MMDSSlaveRequest::OP_XLOCKACK
;
2120 case MMDSSlaveRequest::OP_WRLOCK
:
2121 wrlocks
.insert(lock
);
2122 replycode
= MMDSSlaveRequest::OP_WRLOCKACK
;
2126 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
2130 MMDSSlaveRequest
*r
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, replycode
);
2131 r
->set_lock_type(lock
->get_type());
2132 lock
->get_parent()->set_object_info(r
->get_object_info());
2133 mds
->send_message(r
, mdr
->slave_request
->get_connection());
2137 mdr
->slave_request
->put();
2138 mdr
->slave_request
= 0;
2142 case MMDSSlaveRequest::OP_UNXLOCK
:
2143 case MMDSSlaveRequest::OP_UNWRLOCK
:
2145 SimpleLock
*lock
= mds
->locker
->get_lock(mdr
->slave_request
->get_lock_type(),
2146 mdr
->slave_request
->get_object_info());
2148 bool need_issue
= false;
2150 case MMDSSlaveRequest::OP_UNXLOCK
:
2151 mds
->locker
->xlock_finish(lock
, mdr
.get(), &need_issue
);
2153 case MMDSSlaveRequest::OP_UNWRLOCK
:
2154 mds
->locker
->wrlock_finish(lock
, mdr
.get(), &need_issue
);
2158 mds
->locker
->issue_caps(static_cast<CInode
*>(lock
->get_parent()));
2160 // done. no ack necessary.
2161 mdr
->slave_request
->put();
2162 mdr
->slave_request
= 0;
2166 case MMDSSlaveRequest::OP_DROPLOCKS
:
2167 mds
->locker
->drop_locks(mdr
.get());
2168 mdr
->slave_request
->put();
2169 mdr
->slave_request
= 0;
2172 case MMDSSlaveRequest::OP_AUTHPIN
:
2173 handle_slave_auth_pin(mdr
);
2176 case MMDSSlaveRequest::OP_LINKPREP
:
2177 case MMDSSlaveRequest::OP_UNLINKPREP
:
2178 handle_slave_link_prep(mdr
);
2181 case MMDSSlaveRequest::OP_RMDIRPREP
:
2182 handle_slave_rmdir_prep(mdr
);
2185 case MMDSSlaveRequest::OP_RENAMEPREP
:
2186 handle_slave_rename_prep(mdr
);
2189 case MMDSSlaveRequest::OP_FINISH
:
2190 // information about rename imported caps
2191 if (mdr
->slave_request
->inode_export
.length() > 0)
2192 mdr
->more()->inode_import
.claim(mdr
->slave_request
->inode_export
);
2193 // finish off request.
2194 mdcache
->request_finish(mdr
);
2202 /* This function DOES put the mdr->slave_request before returning*/
2203 void Server::handle_slave_auth_pin(MDRequestRef
& mdr
)
2205 dout(10) << "handle_slave_auth_pin " << *mdr
<< dendl
;
2207 // build list of objects
2208 list
<MDSCacheObject
*> objects
;
2209 CInode
*auth_pin_freeze
= NULL
;
2210 bool fail
= false, wouldblock
= false, readonly
= false;
2212 if (mdcache
->is_readonly()) {
2213 dout(10) << " read-only FS" << dendl
;
2219 for (vector
<MDSCacheObjectInfo
>::iterator p
= mdr
->slave_request
->get_authpins().begin();
2220 p
!= mdr
->slave_request
->get_authpins().end();
2222 MDSCacheObject
*object
= mdcache
->get_object(*p
);
2224 dout(10) << " don't have " << *p
<< dendl
;
2229 objects
.push_back(object
);
2230 if (*p
== mdr
->slave_request
->get_authpin_freeze())
2231 auth_pin_freeze
= static_cast<CInode
*>(object
);
2235 // can we auth pin them?
2237 for (list
<MDSCacheObject
*>::iterator p
= objects
.begin();
2240 if (!(*p
)->is_auth()) {
2241 dout(10) << " not auth for " << **p
<< dendl
;
2245 if (mdr
->is_auth_pinned(*p
))
2247 if (!mdr
->can_auth_pin(*p
)) {
2248 if (mdr
->slave_request
->is_nonblock()) {
2249 dout(10) << " can't auth_pin (freezing?) " << **p
<< " nonblocking" << dendl
;
2255 dout(10) << " waiting for authpinnable on " << **p
<< dendl
;
2256 (*p
)->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2257 mdr
->drop_local_auth_pins();
2259 mds
->locker
->notify_freeze_waiter(*p
);
2267 mdr
->drop_local_auth_pins(); // just in case
2269 /* freeze authpin wrong inode */
2270 if (mdr
->has_more() && mdr
->more()->is_freeze_authpin
&&
2271 mdr
->more()->rename_inode
!= auth_pin_freeze
)
2272 mdr
->unfreeze_auth_pin(true);
2274 /* handle_slave_rename_prep() call freeze_inode() to wait for all other operations
2275 * on the source inode to complete. This happens after all locks for the rename
2276 * operation are acquired. But to acquire locks, we need auth pin locks' parent
2277 * objects first. So there is an ABBA deadlock if someone auth pins the source inode
2278 * after locks are acquired and before Server::handle_slave_rename_prep() is called.
2279 * The solution is freeze the inode and prevent other MDRequests from getting new
2282 if (auth_pin_freeze
) {
2283 dout(10) << " freezing auth pin on " << *auth_pin_freeze
<< dendl
;
2284 if (!mdr
->freeze_auth_pin(auth_pin_freeze
)) {
2285 auth_pin_freeze
->add_waiter(CInode::WAIT_FROZEN
, new C_MDS_RetryRequest(mdcache
, mdr
));
2286 mds
->mdlog
->flush();
2290 for (list
<MDSCacheObject
*>::iterator p
= objects
.begin();
2293 dout(10) << "auth_pinning " << **p
<< dendl
;
2299 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_AUTHPINACK
);
2301 // return list of my auth_pins (if any)
2302 for (set
<MDSCacheObject
*>::iterator p
= mdr
->auth_pins
.begin();
2303 p
!= mdr
->auth_pins
.end();
2305 MDSCacheObjectInfo info
;
2306 (*p
)->set_object_info(info
);
2307 reply
->get_authpins().push_back(info
);
2308 if (*p
== (MDSCacheObject
*)auth_pin_freeze
)
2309 auth_pin_freeze
->set_object_info(reply
->get_authpin_freeze());
2313 reply
->mark_error_wouldblock();
2315 reply
->mark_error_rofs();
2317 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
2319 // clean up this request
2320 mdr
->slave_request
->put();
2321 mdr
->slave_request
= 0;
2325 /* This function DOES NOT put the passed ack before returning*/
2326 void Server::handle_slave_auth_pin_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
2328 dout(10) << "handle_slave_auth_pin_ack on " << *mdr
<< " " << *ack
<< dendl
;
2329 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
2332 set
<MDSCacheObject
*> pinned
;
2333 for (vector
<MDSCacheObjectInfo
>::iterator p
= ack
->get_authpins().begin();
2334 p
!= ack
->get_authpins().end();
2336 MDSCacheObject
*object
= mdcache
->get_object(*p
);
2337 assert(object
); // we pinned it
2338 dout(10) << " remote has pinned " << *object
<< dendl
;
2339 if (!mdr
->is_auth_pinned(object
))
2340 mdr
->remote_auth_pins
[object
] = from
;
2341 if (*p
== ack
->get_authpin_freeze())
2342 mdr
->set_remote_frozen_auth_pin(static_cast<CInode
*>(object
));
2343 pinned
.insert(object
);
2346 // removed frozen auth pin ?
2347 if (mdr
->more()->is_remote_frozen_authpin
&&
2348 ack
->get_authpin_freeze() == MDSCacheObjectInfo()) {
2349 auto p
= mdr
->remote_auth_pins
.find(mdr
->more()->rename_inode
);
2350 assert(p
!= mdr
->remote_auth_pins
.end());
2351 if (p
->second
== from
) {
2352 mdr
->more()->is_remote_frozen_authpin
= false;
2356 // removed auth pins?
2357 map
<MDSCacheObject
*, mds_rank_t
>::iterator p
= mdr
->remote_auth_pins
.begin();
2358 while (p
!= mdr
->remote_auth_pins
.end()) {
2359 MDSCacheObject
* object
= p
->first
;
2360 if (p
->second
== from
&& pinned
.count(object
) == 0) {
2361 dout(10) << " remote has unpinned " << *object
<< dendl
;
2362 mdr
->remote_auth_pins
.erase(p
++);
2368 if (ack
->is_error_rofs()) {
2369 mdr
->more()->slave_error
= -EROFS
;
2370 mdr
->aborted
= true;
2371 } else if (ack
->is_error_wouldblock()) {
2372 mdr
->more()->slave_error
= -EWOULDBLOCK
;
2373 mdr
->aborted
= true;
2377 mdr
->more()->slaves
.insert(from
);
2379 // clear from waiting list
2380 assert(mdr
->more()->waiting_on_slave
.count(from
));
2381 mdr
->more()->waiting_on_slave
.erase(from
);
2384 if (mdr
->more()->waiting_on_slave
.empty())
2385 mdcache
->dispatch_request(mdr
);
2387 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
2391 // ---------------------------------------
2396 * check whether we are permitted to complete a request
2398 * Check whether we have permission to perform the operation specified
2399 * by mask on the given inode, based on the capability in the mdr's
2402 bool Server::check_access(MDRequestRef
& mdr
, CInode
*in
, unsigned mask
)
2405 int r
= mdr
->session
->check_access(
2407 mdr
->client_request
->get_caller_uid(),
2408 mdr
->client_request
->get_caller_gid(),
2409 &mdr
->client_request
->get_caller_gid_list(),
2410 mdr
->client_request
->head
.args
.setattr
.uid
,
2411 mdr
->client_request
->head
.args
.setattr
.gid
);
2413 respond_to_request(mdr
, r
);
2421 * check whether fragment has reached maximum size
2424 bool Server::check_fragment_space(MDRequestRef
&mdr
, CDir
*in
)
2426 const auto size
= in
->get_frag_size();
2427 if (size
>= g_conf
->mds_bal_fragment_size_max
) {
2428 dout(10) << "fragment " << *in
<< " size exceeds " << g_conf
->mds_bal_fragment_size_max
<< " (ENOSPC)" << dendl
;
2429 respond_to_request(mdr
, -ENOSPC
);
2437 /** validate_dentry_dir
2439 * verify that the dir exists and would own the dname.
2440 * do not check if the dentry exists.
2442 CDir
*Server::validate_dentry_dir(MDRequestRef
& mdr
, CInode
*diri
, const string
& dname
)
2444 // make sure parent is a dir?
2445 if (!diri
->is_dir()) {
2446 dout(7) << "validate_dentry_dir: not a dir" << dendl
;
2447 respond_to_request(mdr
, -ENOTDIR
);
2452 frag_t fg
= diri
->pick_dirfrag(dname
);
2453 CDir
*dir
= try_open_auth_dirfrag(diri
, fg
, mdr
);
2458 if (dir
->is_frozen()) {
2459 dout(7) << "dir is frozen " << *dir
<< dendl
;
2460 dir
->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2468 /** prepare_null_dentry
2469 * prepare a null (or existing) dentry in given dir.
2470 * wait for any dn lock.
2472 CDentry
* Server::prepare_null_dentry(MDRequestRef
& mdr
, CDir
*dir
, const string
& dname
, bool okexist
)
2474 dout(10) << "prepare_null_dentry " << dname
<< " in " << *dir
<< dendl
;
2475 assert(dir
->is_auth());
2477 client_t client
= mdr
->get_client();
2479 // does it already exist?
2480 CDentry
*dn
= dir
->lookup(dname
);
2483 if (dn->lock.is_xlocked_by_other(mdr)) {
2484 dout(10) << "waiting on xlocked dentry " << *dn << dendl;
2485 dn->lock.add_waiter(SimpleLock::WAIT_RD, new C_MDS_RetryRequest(mdcache, mdr));
2489 if (!dn
->get_linkage(client
, mdr
)->is_null()) {
2490 // name already exists
2491 dout(10) << "dentry " << dname
<< " exists in " << *dir
<< dendl
;
2493 respond_to_request(mdr
, -EEXIST
);
2497 dn
->first
= dir
->inode
->find_snaprealm()->get_newest_seq() + 1;
2503 // make sure dir is complete
2504 if (!dir
->is_complete() && (!dir
->has_bloom() || dir
->is_in_bloom(dname
))) {
2505 dout(7) << " incomplete dir contents for " << *dir
<< ", fetching" << dendl
;
2506 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
));
2511 dn
= dir
->add_null_dentry(dname
, dir
->inode
->find_snaprealm()->get_newest_seq() + 1);
2513 dout(10) << "prepare_null_dentry added " << *dn
<< dendl
;
2517 CDentry
* Server::prepare_stray_dentry(MDRequestRef
& mdr
, CInode
*in
)
2519 CDentry
*straydn
= mdr
->straydn
;
2522 in
->name_stray_dentry(straydname
);
2523 if (straydn
->get_name() == straydname
)
2526 assert(!mdr
->done_locking
);
2527 mdr
->unpin(straydn
);
2530 CDir
*straydir
= mdcache
->get_stray_dir(in
);
2532 if (!mdr
->client_request
->is_replay() &&
2533 !check_fragment_space(mdr
, straydir
))
2536 straydn
= mdcache
->get_or_create_stray_dentry(in
);
2537 mdr
->straydn
= straydn
;
2542 /** prepare_new_inode
2544 * create a new inode. set c/m/atime. hit dir pop.
2546 CInode
* Server::prepare_new_inode(MDRequestRef
& mdr
, CDir
*dir
, inodeno_t useino
, unsigned mode
,
2547 file_layout_t
*layout
)
2549 CInode
*in
= new CInode(mdcache
);
2551 // Server::prepare_force_open_sessions() can re-open session in closing
2552 // state. In that corner case, session's prealloc_inos are being freed.
2553 // To simplify the code, we disallow using/refilling session's prealloc_ino
2554 // while session is opening.
2555 bool allow_prealloc_inos
= !mdr
->session
->is_opening();
2558 if (allow_prealloc_inos
&&
2559 mdr
->session
->info
.prealloc_inos
.size()) {
2560 mdr
->used_prealloc_ino
=
2561 in
->inode
.ino
= mdr
->session
->take_ino(useino
); // prealloc -> used
2562 mds
->sessionmap
.mark_projected(mdr
->session
);
2564 dout(10) << "prepare_new_inode used_prealloc " << mdr
->used_prealloc_ino
2565 << " (" << mdr
->session
->info
.prealloc_inos
2566 << ", " << mdr
->session
->info
.prealloc_inos
.size() << " left)"
2570 in
->inode
.ino
= mds
->inotable
->project_alloc_id();
2571 dout(10) << "prepare_new_inode alloc " << mdr
->alloc_ino
<< dendl
;
2574 if (useino
&& useino
!= in
->inode
.ino
) {
2575 dout(0) << "WARNING: client specified " << useino
<< " and i allocated " << in
->inode
.ino
<< dendl
;
2576 mds
->clog
->error() << mdr
->client_request
->get_source()
2577 << " specified ino " << useino
2578 << " but mds." << mds
->get_nodeid() << " allocated " << in
->inode
.ino
;
2579 //ceph_abort(); // just for now.
2582 if (allow_prealloc_inos
&&
2583 mdr
->session
->get_num_projected_prealloc_inos() < g_conf
->mds_client_prealloc_inos
/ 2) {
2584 int need
= g_conf
->mds_client_prealloc_inos
- mdr
->session
->get_num_projected_prealloc_inos();
2585 mds
->inotable
->project_alloc_ids(mdr
->prealloc_inos
, need
);
2586 assert(mdr
->prealloc_inos
.size()); // or else fix projected increment semantics
2587 mdr
->session
->pending_prealloc_inos
.insert(mdr
->prealloc_inos
);
2588 mds
->sessionmap
.mark_projected(mdr
->session
);
2589 dout(10) << "prepare_new_inode prealloc " << mdr
->prealloc_inos
<< dendl
;
2592 in
->inode
.version
= 1;
2593 in
->inode
.xattr_version
= 1;
2594 in
->inode
.nlink
= 1; // FIXME
2596 in
->inode
.mode
= mode
;
2598 memset(&in
->inode
.dir_layout
, 0, sizeof(in
->inode
.dir_layout
));
2599 if (in
->inode
.is_dir()) {
2600 in
->inode
.dir_layout
.dl_dir_hash
= g_conf
->mds_default_dir_hash
;
2601 } else if (layout
) {
2602 in
->inode
.layout
= *layout
;
2604 in
->inode
.layout
= mdcache
->default_file_layout
;
2607 in
->inode
.truncate_size
= -1ull; // not truncated, yet!
2608 in
->inode
.truncate_seq
= 1; /* starting with 1, 0 is kept for no-truncation logic */
2610 CInode
*diri
= dir
->get_inode();
2612 dout(10) << oct
<< " dir mode 0" << diri
->inode
.mode
<< " new mode 0" << mode
<< dec
<< dendl
;
2614 if (diri
->inode
.mode
& S_ISGID
) {
2615 dout(10) << " dir is sticky" << dendl
;
2616 in
->inode
.gid
= diri
->inode
.gid
;
2617 if (S_ISDIR(mode
)) {
2618 dout(10) << " new dir also sticky" << dendl
;
2619 in
->inode
.mode
|= S_ISGID
;
2622 in
->inode
.gid
= mdr
->client_request
->get_caller_gid();
2624 in
->inode
.uid
= mdr
->client_request
->get_caller_uid();
2626 in
->inode
.btime
= in
->inode
.ctime
= in
->inode
.mtime
= in
->inode
.atime
=
2627 mdr
->get_op_stamp();
2629 in
->inode
.change_attr
= 0;
2631 MClientRequest
*req
= mdr
->client_request
;
2632 if (req
->get_data().length()) {
2633 bufferlist::iterator p
= req
->get_data().begin();
2635 // xattrs on new inode?
2636 map
<string
,bufferptr
> xattrs
;
2637 ::decode(xattrs
, p
);
2638 for (map
<string
,bufferptr
>::iterator p
= xattrs
.begin(); p
!= xattrs
.end(); ++p
) {
2639 dout(10) << "prepare_new_inode setting xattr " << p
->first
<< dendl
;
2640 in
->xattrs
[p
->first
] = p
->second
;
2644 if (!mds
->mdsmap
->get_inline_data_enabled() ||
2645 !mdr
->session
->connection
->has_feature(CEPH_FEATURE_MDS_INLINE_DATA
))
2646 in
->inode
.inline_data
.version
= CEPH_INLINE_NONE
;
2648 mdcache
->add_inode(in
); // add
2649 dout(10) << "prepare_new_inode " << *in
<< dendl
;
2653 void Server::journal_allocated_inos(MDRequestRef
& mdr
, EMetaBlob
*blob
)
2655 dout(20) << "journal_allocated_inos sessionmapv " << mds
->sessionmap
.get_projected()
2656 << " inotablev " << mds
->inotable
->get_projected_version()
2658 blob
->set_ino_alloc(mdr
->alloc_ino
,
2659 mdr
->used_prealloc_ino
,
2661 mdr
->client_request
->get_source(),
2662 mds
->sessionmap
.get_projected(),
2663 mds
->inotable
->get_projected_version());
2666 void Server::apply_allocated_inos(MDRequestRef
& mdr
, Session
*session
)
2668 dout(10) << "apply_allocated_inos " << mdr
->alloc_ino
2669 << " / " << mdr
->prealloc_inos
2670 << " / " << mdr
->used_prealloc_ino
<< dendl
;
2672 if (mdr
->alloc_ino
) {
2673 mds
->inotable
->apply_alloc_id(mdr
->alloc_ino
);
2675 if (mdr
->prealloc_inos
.size()) {
2677 session
->pending_prealloc_inos
.subtract(mdr
->prealloc_inos
);
2678 session
->info
.prealloc_inos
.insert(mdr
->prealloc_inos
);
2679 mds
->sessionmap
.mark_dirty(session
);
2680 mds
->inotable
->apply_alloc_ids(mdr
->prealloc_inos
);
2682 if (mdr
->used_prealloc_ino
) {
2684 session
->info
.used_inos
.erase(mdr
->used_prealloc_ino
);
2685 mds
->sessionmap
.mark_dirty(session
);
2689 class C_MDS_TryFindInode
: public ServerContext
{
2692 C_MDS_TryFindInode(Server
*s
, MDRequestRef
& r
) : ServerContext(s
), mdr(r
) {}
2693 void finish(int r
) override
{
2694 if (r
== -ESTALE
) // :( find_ino_peers failed
2695 server
->respond_to_request(mdr
, r
);
2697 server
->dispatch_client_request(mdr
);
2701 CDir
*Server::traverse_to_auth_dir(MDRequestRef
& mdr
, vector
<CDentry
*> &trace
, filepath refpath
)
2703 // figure parent dir vs dname
2704 if (refpath
.depth() == 0) {
2705 dout(7) << "can't do that to root" << dendl
;
2706 respond_to_request(mdr
, -EINVAL
);
2709 string dname
= refpath
.last_dentry();
2710 refpath
.pop_dentry();
2712 dout(10) << "traverse_to_auth_dir dirpath " << refpath
<< " dname " << dname
<< dendl
;
2714 // traverse to parent dir
2716 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, refpath
, &trace
, &diri
, MDS_TRAVERSE_FORWARD
);
2717 if (r
> 0) return 0; // delayed
2720 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
2721 mdcache
->find_ino_peers(refpath
.get_ino(), new C_MDS_TryFindInode(this, mdr
));
2724 respond_to_request(mdr
, r
);
2728 // is it an auth dir?
2729 CDir
*dir
= validate_dentry_dir(mdr
, diri
, dname
);
2731 return 0; // forwarded or waiting for freeze
2733 dout(10) << "traverse_to_auth_dir " << *dir
<< dendl
;
2737 /* If this returns null, the request has been handled
2738 * as appropriate: forwarded on, or the client's been replied to */
2739 CInode
* Server::rdlock_path_pin_ref(MDRequestRef
& mdr
, int n
,
2740 set
<SimpleLock
*> &rdlocks
,
2742 bool no_want_auth
, /* for readdir, who doesn't want auth _even_if_ it's
2744 file_layout_t
**layout
,
2745 bool no_lookup
) // true if we cannot return a null dentry lease
2747 const filepath
& refpath
= n
? mdr
->get_filepath2() : mdr
->get_filepath();
2748 dout(10) << "rdlock_path_pin_ref " << *mdr
<< " " << refpath
<< dendl
;
2750 if (mdr
->done_locking
)
2754 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, refpath
, &mdr
->dn
[n
], &mdr
->in
[n
], MDS_TRAVERSE_FORWARD
);
2756 return NULL
; // delayed
2757 if (r
< 0) { // error
2758 if (r
== -ENOENT
&& n
== 0 && mdr
->dn
[n
].size()) {
2760 mdr
->tracedn
= mdr
->dn
[n
][mdr
->dn
[n
].size()-1];
2761 respond_to_request(mdr
, r
);
2762 } else if (r
== -ESTALE
) {
2763 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
2764 MDSInternalContextBase
*c
= new C_MDS_TryFindInode(this, mdr
);
2765 mdcache
->find_ino_peers(refpath
.get_ino(), c
);
2767 dout(10) << "FAIL on error " << r
<< dendl
;
2768 respond_to_request(mdr
, r
);
2772 CInode
*ref
= mdr
->in
[n
];
2773 dout(10) << "ref is " << *ref
<< dendl
;
2775 // fw to inode auth?
2776 if (mdr
->snapid
!= CEPH_NOSNAP
&& !no_want_auth
)
2780 if (ref
->is_ambiguous_auth()) {
2781 dout(10) << "waiting for single auth on " << *ref
<< dendl
;
2782 ref
->add_waiter(CInode::WAIT_SINGLEAUTH
, new C_MDS_RetryRequest(mdcache
, mdr
));
2785 if (!ref
->is_auth()) {
2786 dout(10) << "fw to auth for " << *ref
<< dendl
;
2787 mdcache
->request_forward(mdr
, ref
->authority().first
);
2792 // do NOT proceed if freezing, as cap release may defer in that case, and
2793 // we could deadlock when we try to lock @ref.
2794 // if we're already auth_pinned, continue; the release has already been processed.
2795 if (ref
->is_frozen() || ref
->is_frozen_auth_pin() ||
2796 (ref
->is_freezing() && !mdr
->is_auth_pinned(ref
))) {
2797 dout(7) << "waiting for !frozen/authpinnable on " << *ref
<< dendl
;
2798 ref
->add_waiter(CInode::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2799 /* If we have any auth pins, this will deadlock.
2800 * But the only way to get here if we've already got auth pins
2801 * is because we're on an inode with snapshots that got updated
2802 * between dispatches of this request. So we're going to drop
2803 * our locks and our auth pins and reacquire them later.
2805 * This is safe since we're only in this function when working on
2806 * a single MDS request; otherwise we'd be in
2807 * rdlock_path_xlock_dentry.
2809 mds
->locker
->drop_locks(mdr
.get(), NULL
);
2810 mdr
->drop_local_auth_pins();
2811 if (!mdr
->remote_auth_pins
.empty())
2812 mds
->locker
->notify_freeze_waiter(ref
);
2819 for (int i
=0; i
<(int)mdr
->dn
[n
].size(); i
++)
2820 rdlocks
.insert(&mdr
->dn
[n
][i
]->lock
);
2822 mds
->locker
->include_snap_rdlocks_wlayout(rdlocks
, ref
, layout
);
2824 mds
->locker
->include_snap_rdlocks(rdlocks
, ref
);
2832 /** rdlock_path_xlock_dentry
2833 * traverse path to the directory that could/would contain dentry.
2834 * make sure i am auth for that dentry, forward as necessary.
2835 * create null dentry in place (or use existing if okexist).
2836 * get rdlocks on traversed dentries, xlock on new dentry.
2838 CDentry
* Server::rdlock_path_xlock_dentry(MDRequestRef
& mdr
, int n
,
2839 set
<SimpleLock
*>& rdlocks
, set
<SimpleLock
*>& wrlocks
, set
<SimpleLock
*>& xlocks
,
2840 bool okexist
, bool mustexist
, bool alwaysxlock
,
2841 file_layout_t
**layout
)
2843 const filepath
& refpath
= n
? mdr
->get_filepath2() : mdr
->get_filepath();
2845 dout(10) << "rdlock_path_xlock_dentry " << *mdr
<< " " << refpath
<< dendl
;
2847 client_t client
= mdr
->get_client();
2849 if (mdr
->done_locking
)
2850 return mdr
->dn
[n
].back();
2852 CDir
*dir
= traverse_to_auth_dir(mdr
, mdr
->dn
[n
], refpath
);
2855 CInode
*diri
= dir
->get_inode();
2856 if (!mdr
->reqid
.name
.is_mds()) {
2857 if (diri
->is_system() && !diri
->is_root()) {
2858 respond_to_request(mdr
, -EROFS
);
2862 if (!diri
->is_base() && diri
->get_projected_parent_dir()->inode
->is_stray()) {
2863 respond_to_request(mdr
, -ENOENT
);
2867 // make a null dentry?
2868 const string
&dname
= refpath
.last_dentry();
2871 dn
= dir
->lookup(dname
);
2873 // make sure dir is complete
2874 if (!dn
&& !dir
->is_complete() &&
2875 (!dir
->has_bloom() || dir
->is_in_bloom(dname
))) {
2876 dout(7) << " incomplete dir contents for " << *dir
<< ", fetching" << dendl
;
2877 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
));
2882 if (dn
&& !dn
->lock
.can_read(client
) && dn
->lock
.get_xlock_by() != mdr
) {
2883 dout(10) << "waiting on xlocked dentry " << *dn
<< dendl
;
2884 dn
->lock
.add_waiter(SimpleLock::WAIT_RD
, new C_MDS_RetryRequest(mdcache
, mdr
));
2889 if (!dn
|| dn
->get_linkage(client
, mdr
)->is_null()) {
2890 dout(7) << "dentry " << dname
<< " dne in " << *dir
<< dendl
;
2891 respond_to_request(mdr
, -ENOENT
);
2895 dn
= prepare_null_dentry(mdr
, dir
, dname
, okexist
);
2900 mdr
->dn
[n
].push_back(dn
);
2901 CDentry::linkage_t
*dnl
= dn
->get_linkage(client
, mdr
);
2902 mdr
->in
[n
] = dnl
->get_inode();
2905 // NOTE: rename takes the same set of locks for srcdn
2906 for (int i
=0; i
<(int)mdr
->dn
[n
].size(); i
++)
2907 rdlocks
.insert(&mdr
->dn
[n
][i
]->lock
);
2908 if (alwaysxlock
|| dnl
->is_null())
2909 xlocks
.insert(&dn
->lock
); // new dn, xlock
2911 rdlocks
.insert(&dn
->lock
); // existing dn, rdlock
2912 wrlocks
.insert(&dn
->get_dir()->inode
->filelock
); // also, wrlock on dir mtime
2913 wrlocks
.insert(&dn
->get_dir()->inode
->nestlock
); // also, wrlock on dir mtime
2915 mds
->locker
->include_snap_rdlocks_wlayout(rdlocks
, dn
->get_dir()->inode
, layout
);
2917 mds
->locker
->include_snap_rdlocks(rdlocks
, dn
->get_dir()->inode
);
2927 * try_open_auth_dirfrag -- open dirfrag, or forward to dirfrag auth
2929 * @param diri base inode
2930 * @param fg the exact frag we want
2931 * @param mdr request
2932 * @returns the pointer, or NULL if it had to be delayed (but mdr is taken care of)
2934 CDir
* Server::try_open_auth_dirfrag(CInode
*diri
, frag_t fg
, MDRequestRef
& mdr
)
2936 CDir
*dir
= diri
->get_dirfrag(fg
);
2938 // not open and inode not mine?
2939 if (!dir
&& !diri
->is_auth()) {
2940 mds_rank_t inauth
= diri
->authority().first
;
2941 dout(7) << "try_open_auth_dirfrag: not open, not inode auth, fw to mds." << inauth
<< dendl
;
2942 mdcache
->request_forward(mdr
, inauth
);
2946 // not open and inode frozen?
2947 if (!dir
&& diri
->is_frozen()) {
2948 dout(10) << "try_open_auth_dirfrag: dir inode is frozen, waiting " << *diri
<< dendl
;
2949 assert(diri
->get_parent_dir());
2950 diri
->add_waiter(CInode::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
2956 dir
= diri
->get_or_open_dirfrag(mdcache
, fg
);
2958 // am i auth for the dirfrag?
2959 if (!dir
->is_auth()) {
2960 mds_rank_t auth
= dir
->authority().first
;
2961 dout(7) << "try_open_auth_dirfrag: not auth for " << *dir
2962 << ", fw to mds." << auth
<< dendl
;
2963 mdcache
->request_forward(mdr
, auth
);
2971 // ===============================================================================
2974 void Server::handle_client_getattr(MDRequestRef
& mdr
, bool is_lookup
)
2976 MClientRequest
*req
= mdr
->client_request
;
2977 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
2979 if (req
->get_filepath().depth() == 0 && is_lookup
) {
2980 // refpath can't be empty for lookup but it can for
2981 // getattr (we do getattr with empty refpath for mount of '/')
2982 respond_to_request(mdr
, -EINVAL
);
2986 CInode
*ref
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, false, false, NULL
, !is_lookup
);
2990 * if client currently holds the EXCL cap on a field, do not rdlock
2991 * it; client's stat() will result in valid info if _either_ EXCL
2992 * cap is held or MDS rdlocks and reads the value here.
2994 * handling this case here is easier than weakening rdlock
2995 * semantics... that would cause problems elsewhere.
2997 client_t client
= mdr
->get_client();
2999 Capability
*cap
= ref
->get_client_cap(client
);
3000 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
||
3001 mdr
->snapid
<= cap
->client_follows
))
3002 issued
= cap
->issued();
3004 int mask
= req
->head
.args
.getattr
.mask
;
3005 if ((mask
& CEPH_CAP_LINK_SHARED
) && (issued
& CEPH_CAP_LINK_EXCL
) == 0) rdlocks
.insert(&ref
->linklock
);
3006 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0) rdlocks
.insert(&ref
->authlock
);
3007 if ((mask
& CEPH_CAP_FILE_SHARED
) && (issued
& CEPH_CAP_FILE_EXCL
) == 0) rdlocks
.insert(&ref
->filelock
);
3008 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0) rdlocks
.insert(&ref
->xattrlock
);
3010 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3013 if (!check_access(mdr
, ref
, MAY_READ
))
3016 // note which caps are requested, so we return at least a snapshot
3017 // value for them. (currently this matters for xattrs and inline data)
3018 mdr
->getattr_caps
= mask
;
3020 mds
->balancer
->hit_inode(ceph_clock_now(), ref
, META_POP_IRD
,
3021 req
->get_source().num());
3024 dout(10) << "reply to stat on " << *req
<< dendl
;
3027 mdr
->tracedn
= mdr
->dn
[0].back();
3028 respond_to_request(mdr
, 0);
3031 struct C_MDS_LookupIno2
: public ServerContext
{
3033 C_MDS_LookupIno2(Server
*s
, MDRequestRef
& r
) : ServerContext(s
), mdr(r
) {}
3034 void finish(int r
) override
{
3035 server
->_lookup_ino_2(mdr
, r
);
3039 /* This function DOES clean up the mdr before returning*/
3043 void Server::handle_client_lookup_ino(MDRequestRef
& mdr
,
3044 bool want_parent
, bool want_dentry
)
3046 MClientRequest
*req
= mdr
->client_request
;
3048 inodeno_t ino
= req
->get_filepath().get_ino();
3049 CInode
*in
= mdcache
->get_inode(ino
);
3050 if (in
&& in
->state_test(CInode::STATE_PURGING
)) {
3051 respond_to_request(mdr
, -ESTALE
);
3055 mdcache
->open_ino(ino
, (int64_t)-1, new C_MDS_LookupIno2(this, mdr
), false);
3059 if (mdr
&& in
->snaprealm
&& !in
->snaprealm
->is_open() &&
3060 !in
->snaprealm
->open_parents(new C_MDS_RetryRequest(mdcache
, mdr
))) {
3064 // check for nothing (not read or write); this still applies the
3066 if (!check_access(mdr
, in
, 0))
3069 CDentry
*dn
= in
->get_projected_parent_dn();
3070 CInode
*diri
= dn
? dn
->get_dir()->inode
: NULL
;
3072 set
<SimpleLock
*> rdlocks
;
3073 if (dn
&& (want_parent
|| want_dentry
)) {
3075 rdlocks
.insert(&dn
->lock
);
3078 unsigned mask
= req
->head
.args
.getattr
.mask
;
3080 Capability
*cap
= in
->get_client_cap(mdr
->get_client());
3082 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
|| mdr
->snapid
<= cap
->client_follows
))
3083 issued
= cap
->issued();
3084 // permission bits, ACL/security xattrs
3085 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0)
3086 rdlocks
.insert(&in
->authlock
);
3087 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0)
3088 rdlocks
.insert(&in
->xattrlock
);
3090 mdr
->getattr_caps
= mask
;
3093 if (!rdlocks
.empty()) {
3094 set
<SimpleLock
*> wrlocks
, xlocks
;
3095 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3098 // need read access to directory inode
3099 if (!check_access(mdr
, diri
, MAY_READ
))
3104 if (in
->is_base()) {
3105 respond_to_request(mdr
, -EINVAL
);
3108 if (!diri
|| diri
->is_stray()) {
3109 respond_to_request(mdr
, -ESTALE
);
3112 dout(10) << "reply to lookup_parent " << *in
<< dendl
;
3114 respond_to_request(mdr
, 0);
3117 inodeno_t dirino
= req
->get_filepath2().get_ino();
3118 if (!diri
|| (dirino
!= inodeno_t() && diri
->ino() != dirino
)) {
3119 respond_to_request(mdr
, -ENOENT
);
3122 dout(10) << "reply to lookup_name " << *in
<< dendl
;
3124 dout(10) << "reply to lookup_ino " << *in
<< dendl
;
3129 respond_to_request(mdr
, 0);
3133 void Server::_lookup_ino_2(MDRequestRef
& mdr
, int r
)
3135 inodeno_t ino
= mdr
->client_request
->get_filepath().get_ino();
3136 dout(10) << "_lookup_ino_2 " << mdr
.get() << " ino " << ino
<< " r=" << r
<< dendl
;
3138 // `r` is a rank if >=0, else an error code
3140 mds_rank_t
dest_rank(r
);
3141 if (dest_rank
== mds
->get_nodeid())
3142 dispatch_client_request(mdr
);
3144 mdcache
->request_forward(mdr
, dest_rank
);
3149 if (r
== -ENOENT
|| r
== -ENODATA
)
3151 respond_to_request(mdr
, r
);
3155 /* This function takes responsibility for the passed mdr*/
3156 void Server::handle_client_open(MDRequestRef
& mdr
)
3158 MClientRequest
*req
= mdr
->client_request
;
3159 dout(7) << "open on " << req
->get_filepath() << dendl
;
3161 int flags
= req
->head
.args
.open
.flags
;
3162 int cmode
= ceph_flags_to_mode(flags
);
3164 respond_to_request(mdr
, -EINVAL
);
3168 bool need_auth
= !file_mode_is_readonly(cmode
) || (flags
& CEPH_O_TRUNC
);
3170 if ((cmode
& CEPH_FILE_MODE_WR
) && mdcache
->is_readonly()) {
3171 dout(7) << "read-only FS" << dendl
;
3172 respond_to_request(mdr
, -EROFS
);
3176 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3177 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, need_auth
);
3181 if (cur
->is_frozen() || cur
->state_test(CInode::STATE_EXPORTINGCAPS
)) {
3183 mdr
->done_locking
= false;
3184 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3189 if (!cur
->inode
.is_file()) {
3190 // can only open non-regular inode with mode FILE_MODE_PIN, at least for now.
3191 cmode
= CEPH_FILE_MODE_PIN
;
3192 // the inode is symlink and client wants to follow it, ignore the O_TRUNC flag.
3193 if (cur
->inode
.is_symlink() && !(flags
& CEPH_O_NOFOLLOW
))
3194 flags
&= ~CEPH_O_TRUNC
;
3197 dout(10) << "open flags = " << flags
3198 << ", filemode = " << cmode
3199 << ", need_auth = " << need_auth
3203 /*if (!cur->inode.is_file() && !cur->inode.is_dir()) {
3204 dout(7) << "not a file or dir " << *cur << dendl;
3205 respond_to_request(mdr, -ENXIO); // FIXME what error do we want?
3208 if ((flags
& CEPH_O_DIRECTORY
) && !cur
->inode
.is_dir() && !cur
->inode
.is_symlink()) {
3209 dout(7) << "specified O_DIRECTORY on non-directory " << *cur
<< dendl
;
3210 respond_to_request(mdr
, -EINVAL
);
3214 if ((flags
& CEPH_O_TRUNC
) && !cur
->inode
.is_file()) {
3215 dout(7) << "specified O_TRUNC on !(file|symlink) " << *cur
<< dendl
;
3216 // we should return -EISDIR for directory, return -EINVAL for other non-regular
3217 respond_to_request(mdr
, cur
->inode
.is_dir() ? -EISDIR
: -EINVAL
);
3221 if (cur
->inode
.inline_data
.version
!= CEPH_INLINE_NONE
&&
3222 !mdr
->session
->connection
->has_feature(CEPH_FEATURE_MDS_INLINE_DATA
)) {
3223 dout(7) << "old client cannot open inline data file " << *cur
<< dendl
;
3224 respond_to_request(mdr
, -EPERM
);
3228 // snapped data is read only
3229 if (mdr
->snapid
!= CEPH_NOSNAP
&&
3230 ((cmode
& CEPH_FILE_MODE_WR
) || req
->may_write())) {
3231 dout(7) << "snap " << mdr
->snapid
<< " is read-only " << *cur
<< dendl
;
3232 respond_to_request(mdr
, -EROFS
);
3236 unsigned mask
= req
->head
.args
.open
.mask
;
3238 Capability
*cap
= cur
->get_client_cap(mdr
->get_client());
3240 if (cap
&& (mdr
->snapid
== CEPH_NOSNAP
|| mdr
->snapid
<= cap
->client_follows
))
3241 issued
= cap
->issued();
3242 // permission bits, ACL/security xattrs
3243 if ((mask
& CEPH_CAP_AUTH_SHARED
) && (issued
& CEPH_CAP_AUTH_EXCL
) == 0)
3244 rdlocks
.insert(&cur
->authlock
);
3245 if ((mask
& CEPH_CAP_XATTR_SHARED
) && (issued
& CEPH_CAP_XATTR_EXCL
) == 0)
3246 rdlocks
.insert(&cur
->xattrlock
);
3248 mdr
->getattr_caps
= mask
;
3252 if ((flags
& CEPH_O_TRUNC
) && !mdr
->has_completed
) {
3253 assert(cur
->is_auth());
3255 xlocks
.insert(&cur
->filelock
);
3256 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3259 if (!check_access(mdr
, cur
, MAY_WRITE
))
3262 // wait for pending truncate?
3263 const inode_t
*pi
= cur
->get_projected_inode();
3264 if (pi
->is_truncating()) {
3265 dout(10) << " waiting for pending truncate from " << pi
->truncate_from
3266 << " to " << pi
->truncate_size
<< " to complete on " << *cur
<< dendl
;
3267 mds
->locker
->drop_locks(mdr
.get());
3268 mdr
->drop_local_auth_pins();
3269 cur
->add_waiter(CInode::WAIT_TRUNC
, new C_MDS_RetryRequest(mdcache
, mdr
));
3273 do_open_truncate(mdr
, cmode
);
3277 // sync filelock if snapped.
3278 // this makes us wait for writers to flushsnaps, ensuring we get accurate metadata,
3279 // and that data itself is flushed so that we can read the snapped data off disk.
3280 if (mdr
->snapid
!= CEPH_NOSNAP
&& !cur
->is_dir()) {
3281 rdlocks
.insert(&cur
->filelock
);
3284 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3288 if (cmode
& CEPH_FILE_MODE_WR
)
3290 if (!check_access(mdr
, cur
, mask
))
3293 if (cur
->is_file() || cur
->is_dir()) {
3294 if (mdr
->snapid
== CEPH_NOSNAP
) {
3296 Capability
*cap
= mds
->locker
->issue_new_caps(cur
, cmode
, mdr
->session
, 0, req
->is_replay());
3298 dout(12) << "open issued caps " << ccap_string(cap
->pending())
3299 << " for " << req
->get_source()
3300 << " on " << *cur
<< dendl
;
3302 int caps
= ceph_caps_for_mode(cmode
);
3303 dout(12) << "open issued IMMUTABLE SNAP caps " << ccap_string(caps
)
3304 << " for " << req
->get_source()
3305 << " snapid " << mdr
->snapid
3306 << " on " << *cur
<< dendl
;
3307 mdr
->snap_caps
= caps
;
3311 // increase max_size?
3312 if (cmode
& CEPH_FILE_MODE_WR
)
3313 mds
->locker
->check_inode_max_size(cur
);
3315 // make sure this inode gets into the journal
3316 if (cur
->is_auth() && cur
->last
== CEPH_NOSNAP
&&
3317 !cur
->item_open_file
.is_on_list()) {
3318 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
3319 EOpen
*le
= new EOpen(mds
->mdlog
);
3320 mdlog
->start_entry(le
);
3321 le
->add_clean_inode(cur
);
3322 ls
->open_files
.push_back(&cur
->item_open_file
);
3323 mdlog
->submit_entry(le
);
3327 if (cmode
& CEPH_FILE_MODE_WR
)
3328 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), cur
, META_POP_IWR
);
3330 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), cur
, META_POP_IRD
,
3331 mdr
->client_request
->get_source().num());
3334 if (req
->get_dentry_wanted()) {
3335 assert(mdr
->dn
[0].size());
3336 dn
= mdr
->dn
[0].back();
3341 respond_to_request(mdr
, 0);
3344 class C_MDS_openc_finish
: public ServerLogContext
{
3349 C_MDS_openc_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ni
, snapid_t f
) :
3350 ServerLogContext(s
, r
), dn(d
), newi(ni
), follows(f
) {}
3351 void finish(int r
) override
{
3354 dn
->pop_projected_linkage();
3356 // dirty inode, dn, dir
3357 newi
->inode
.version
--; // a bit hacky, see C_MDS_mknod_finish
3358 newi
->mark_dirty(newi
->inode
.version
+1, mdr
->ls
);
3359 newi
->_mark_dirty_parent(mdr
->ls
, true);
3363 get_mds()->locker
->share_inode_max_size(newi
);
3365 MDRequestRef null_ref
;
3366 get_mds()->mdcache
->send_dentry_link(dn
, null_ref
);
3368 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), newi
, META_POP_IWR
);
3370 server
->respond_to_request(mdr
, 0);
3372 assert(g_conf
->mds_kill_openc_at
!= 1);
3376 /* This function takes responsibility for the passed mdr*/
3377 void Server::handle_client_openc(MDRequestRef
& mdr
)
3379 MClientRequest
*req
= mdr
->client_request
;
3380 client_t client
= mdr
->get_client();
3382 dout(7) << "open w/ O_CREAT on " << req
->get_filepath() << dendl
;
3384 int cmode
= ceph_flags_to_mode(req
->head
.args
.open
.flags
);
3386 respond_to_request(mdr
, -EINVAL
);
3390 bool excl
= req
->head
.args
.open
.flags
& CEPH_O_EXCL
;
3393 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, req
->get_filepath(),
3394 &mdr
->dn
[0], NULL
, MDS_TRAVERSE_FORWARD
);
3398 handle_client_open(mdr
);
3401 if (r
< 0 && r
!= -ENOENT
) {
3403 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
3404 MDSInternalContextBase
*c
= new C_MDS_TryFindInode(this, mdr
);
3405 mdcache
->find_ino_peers(req
->get_filepath().get_ino(), c
);
3407 dout(10) << "FAIL on error " << r
<< dendl
;
3408 respond_to_request(mdr
, r
);
3414 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3415 file_layout_t
*dir_layout
= NULL
;
3416 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
,
3417 !excl
, false, false, &dir_layout
);
3419 if (mdr
->snapid
!= CEPH_NOSNAP
) {
3420 respond_to_request(mdr
, -EROFS
);
3424 file_layout_t layout
;
3426 layout
= *dir_layout
;
3428 layout
= mdcache
->default_file_layout
;
3430 // What kind of client caps are required to complete this operation
3431 uint64_t access
= MAY_WRITE
;
3433 const auto default_layout
= layout
;
3435 // fill in any special params from client
3436 if (req
->head
.args
.open
.stripe_unit
)
3437 layout
.stripe_unit
= req
->head
.args
.open
.stripe_unit
;
3438 if (req
->head
.args
.open
.stripe_count
)
3439 layout
.stripe_count
= req
->head
.args
.open
.stripe_count
;
3440 if (req
->head
.args
.open
.object_size
)
3441 layout
.object_size
= req
->head
.args
.open
.object_size
;
3442 if (req
->get_connection()->has_feature(CEPH_FEATURE_CREATEPOOLID
) &&
3443 (__s32
)req
->head
.args
.open
.pool
>= 0) {
3444 layout
.pool_id
= req
->head
.args
.open
.pool
;
3446 // make sure we have as new a map as the client
3447 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
3448 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
3453 // If client doesn't have capability to modify layout pools, then
3454 // only permit this request if the requested pool matches what the
3455 // file would have inherited anyway from its parent.
3456 if (default_layout
!= layout
) {
3457 access
|= MAY_SET_VXATTR
;
3460 if (!layout
.is_valid()) {
3461 dout(10) << " invalid initial file layout" << dendl
;
3462 respond_to_request(mdr
, -EINVAL
);
3465 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
3466 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
3467 respond_to_request(mdr
, -EINVAL
);
3472 CDir
*dir
= dn
->get_dir();
3473 CInode
*diri
= dir
->get_inode();
3474 rdlocks
.insert(&diri
->authlock
);
3475 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3478 if (!check_access(mdr
, diri
, access
))
3481 if (!check_fragment_space(mdr
, dir
))
3484 CDentry::linkage_t
*dnl
= dn
->get_projected_linkage();
3486 if (!dnl
->is_null()) {
3488 assert(req
->head
.args
.open
.flags
& CEPH_O_EXCL
);
3489 dout(10) << "O_EXCL, target exists, failing with -EEXIST" << dendl
;
3490 mdr
->tracei
= dnl
->get_inode();
3492 respond_to_request(mdr
, -EEXIST
);
3497 SnapRealm
*realm
= diri
->find_snaprealm(); // use directory's realm; inode isn't attached yet.
3498 snapid_t follows
= realm
->get_newest_seq();
3500 CInode
*in
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
),
3501 req
->head
.args
.open
.mode
| S_IFREG
, &layout
);
3505 dn
->push_projected_linkage(in
);
3507 in
->inode
.version
= dn
->pre_dirty();
3508 if (layout
.pool_id
!= mdcache
->default_file_layout
.pool_id
)
3509 in
->inode
.add_old_pool(mdcache
->default_file_layout
.pool_id
);
3510 in
->inode
.update_backtrace();
3511 if (cmode
& CEPH_FILE_MODE_WR
) {
3512 in
->inode
.client_ranges
[client
].range
.first
= 0;
3513 in
->inode
.client_ranges
[client
].range
.last
= in
->inode
.get_layout_size_increment();
3514 in
->inode
.client_ranges
[client
].follows
= follows
;
3516 in
->inode
.rstat
.rfiles
= 1;
3518 assert(dn
->first
== follows
+1);
3519 in
->first
= dn
->first
;
3522 mdr
->ls
= mdlog
->get_current_segment();
3523 EUpdate
*le
= new EUpdate(mdlog
, "openc");
3524 mdlog
->start_entry(le
);
3525 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
3526 journal_allocated_inos(mdr
, &le
->metablob
);
3527 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
3528 le
->metablob
.add_primary_dentry(dn
, in
, true, true, true);
3531 mds
->locker
->issue_new_caps(in
, cmode
, mdr
->session
, realm
, req
->is_replay());
3532 in
->authlock
.set_state(LOCK_EXCL
);
3533 in
->xattrlock
.set_state(LOCK_EXCL
);
3535 // make sure this inode gets into the journal
3536 le
->metablob
.add_opened_ino(in
->ino());
3537 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
3538 ls
->open_files
.push_back(&in
->item_open_file
);
3540 C_MDS_openc_finish
*fin
= new C_MDS_openc_finish(this, mdr
, dn
, in
, follows
);
3542 if (mdr
->client_request
->get_connection()->has_feature(CEPH_FEATURE_REPLY_CREATE_INODE
)) {
3543 dout(10) << "adding ino to reply to indicate inode was created" << dendl
;
3544 // add the file created flag onto the reply if create_flags features is supported
3545 ::encode(in
->inode
.ino
, mdr
->reply_extra_bl
);
3548 journal_and_reply(mdr
, in
, dn
, le
, fin
);
3550 // We hit_dir (via hit_inode) in our finish callback, but by then we might
3551 // have overshot the split size (multiple opencs in flight), so here is
3552 // an early chance to split the dir if this openc makes it oversized.
3553 mds
->balancer
->maybe_fragment(dir
, false);
3558 void Server::handle_client_readdir(MDRequestRef
& mdr
)
3560 MClientRequest
*req
= mdr
->client_request
;
3561 client_t client
= req
->get_source().num();
3562 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3563 CInode
*diri
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, false, true);
3566 // it's a directory, right?
3567 if (!diri
->is_dir()) {
3569 dout(10) << "reply to " << *req
<< " readdir -ENOTDIR" << dendl
;
3570 respond_to_request(mdr
, -ENOTDIR
);
3574 rdlocks
.insert(&diri
->filelock
);
3575 rdlocks
.insert(&diri
->dirfragtreelock
);
3577 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
3580 if (!check_access(mdr
, diri
, MAY_READ
))
3584 frag_t fg
= (__u32
)req
->head
.args
.readdir
.frag
;
3585 unsigned req_flags
= (__u32
)req
->head
.args
.readdir
.flags
;
3586 string offset_str
= req
->get_path2();
3588 __u32 offset_hash
= 0;
3589 if (!offset_str
.empty())
3590 offset_hash
= ceph_frag_value(diri
->hash_dentry_name(offset_str
));
3592 offset_hash
= (__u32
)req
->head
.args
.readdir
.offset_hash
;
3594 dout(10) << " frag " << fg
<< " offset '" << offset_str
<< "'"
3595 << " offset_hash " << offset_hash
<< " flags " << req_flags
<< dendl
;
3597 // does the frag exist?
3598 if (diri
->dirfragtree
[fg
.value()] != fg
) {
3600 if (req_flags
& CEPH_READDIR_REPLY_BITFLAGS
) {
3601 if (fg
.contains((unsigned)offset_hash
)) {
3602 newfg
= diri
->dirfragtree
[offset_hash
];
3604 // client actually wants next frag
3605 newfg
= diri
->dirfragtree
[fg
.value()];
3609 newfg
= diri
->dirfragtree
[fg
.value()];
3611 dout(10) << " adjust frag " << fg
<< " -> " << newfg
<< " " << diri
->dirfragtree
<< dendl
;
3615 CDir
*dir
= try_open_auth_dirfrag(diri
, fg
, mdr
);
3619 dout(10) << "handle_client_readdir on " << *dir
<< dendl
;
3620 assert(dir
->is_auth());
3622 if (!dir
->is_complete()) {
3623 if (dir
->is_frozen()) {
3624 dout(7) << "dir is frozen " << *dir
<< dendl
;
3625 mds
->locker
->drop_locks(mdr
.get());
3626 mdr
->drop_local_auth_pins();
3627 dir
->add_waiter(CDir::WAIT_UNFREEZE
, new C_MDS_RetryRequest(mdcache
, mdr
));
3631 dout(10) << " incomplete dir contents for readdir on " << *dir
<< ", fetching" << dendl
;
3632 dir
->fetch(new C_MDS_RetryRequest(mdcache
, mdr
), true);
3636 #ifdef MDS_VERIFY_FRAGSTAT
3637 dir
->verify_fragstat();
3640 utime_t now
= ceph_clock_now();
3641 mdr
->set_mds_stamp(now
);
3643 snapid_t snapid
= mdr
->snapid
;
3644 dout(10) << "snapid " << snapid
<< dendl
;
3646 SnapRealm
*realm
= diri
->find_snaprealm();
3648 unsigned max
= req
->head
.args
.readdir
.max_entries
;
3650 max
= dir
->get_num_any(); // whatever, something big.
3651 unsigned max_bytes
= req
->head
.args
.readdir
.max_bytes
;
3653 // make sure at least one item can be encoded
3654 max_bytes
= (512 << 10) + g_conf
->mds_max_xattr_pairs_size
;
3658 dir
->encode_dirstat(dirbl
, mds
->get_nodeid());
3660 // count bytes available.
3661 // this isn't perfect, but we should capture the main variable/unbounded size items!
3662 int front_bytes
= dirbl
.length() + sizeof(__u32
) + sizeof(__u8
)*2;
3663 int bytes_left
= max_bytes
- front_bytes
;
3664 bytes_left
-= realm
->get_snap_trace().length();
3666 // build dir contents
3669 bool start
= !offset_hash
&& offset_str
.empty();
3670 bool end
= (dir
->begin() == dir
->end());
3671 // skip all dns < dentry_key_t(snapid, offset_str, offset_hash)
3672 dentry_key_t
skip_key(snapid
, offset_str
.c_str(), offset_hash
);
3673 for (CDir::map_t::iterator it
= start
? dir
->begin() : dir
->lower_bound(skip_key
);
3674 !end
&& numfiles
< max
;
3675 end
= (it
== dir
->end())) {
3676 CDentry
*dn
= it
->second
;
3679 if (dn
->state_test(CDentry::STATE_PURGING
))
3682 bool dnp
= dn
->use_projected(client
, mdr
);
3683 CDentry::linkage_t
*dnl
= dnp
? dn
->get_projected_linkage() : dn
->get_linkage();
3688 if (dn
->last
< snapid
|| dn
->first
> snapid
) {
3689 dout(20) << "skipping non-overlapping snap " << *dn
<< dendl
;
3694 dentry_key_t
offset_key(dn
->last
, offset_str
.c_str(), offset_hash
);
3695 if (!(offset_key
< dn
->key()))
3699 CInode
*in
= dnl
->get_inode();
3701 if (in
&& in
->ino() == CEPH_INO_CEPH
)
3705 // better for the MDS to do the work, if we think the client will stat any of these files.
3706 if (dnl
->is_remote() && !in
) {
3707 in
= mdcache
->get_inode(dnl
->get_remote_ino());
3709 dn
->link_remote(dnl
, in
);
3710 } else if (dn
->state_test(CDentry::STATE_BADREMOTEINO
)) {
3711 dout(10) << "skipping bad remote ino on " << *dn
<< dendl
;
3714 // touch everything i _do_ have
3715 for (CDir::map_t::iterator p
= dir
->begin(); p
!= dir
->end(); ++p
)
3716 if (!p
->second
->get_linkage()->is_null())
3717 mdcache
->lru
.lru_touch(p
->second
);
3719 // already issued caps and leases, reply immediately.
3720 if (dnbl
.length() > 0) {
3721 mdcache
->open_remote_dentry(dn
, dnp
, new C_MDSInternalNoop
);
3722 dout(10) << " open remote dentry after caps were issued, stopping at "
3723 << dnbl
.length() << " < " << bytes_left
<< dendl
;
3727 mds
->locker
->drop_locks(mdr
.get());
3728 mdr
->drop_local_auth_pins();
3729 mdcache
->open_remote_dentry(dn
, dnp
, new C_MDS_RetryRequest(mdcache
, mdr
));
3735 if ((int)(dnbl
.length() + dn
->name
.length() + sizeof(__u32
) + sizeof(LeaseStat
)) > bytes_left
) {
3736 dout(10) << " ran out of room, stopping at " << dnbl
.length() << " < " << bytes_left
<< dendl
;
3740 unsigned start_len
= dnbl
.length();
3743 dout(12) << "including dn " << *dn
<< dendl
;
3744 ::encode(dn
->name
, dnbl
);
3745 mds
->locker
->issue_client_lease(dn
, client
, dnbl
, now
, mdr
->session
);
3748 dout(12) << "including inode " << *in
<< dendl
;
3749 int r
= in
->encode_inodestat(dnbl
, mdr
->session
, realm
, snapid
, bytes_left
- (int)dnbl
.length());
3751 // chop off dn->name, lease
3752 dout(10) << " ran out of room, stopping at " << start_len
<< " < " << bytes_left
<< dendl
;
3754 keep
.substr_of(dnbl
, 0, start_len
);
3762 mdcache
->lru
.lru_touch(dn
);
3767 flags
= CEPH_READDIR_FRAG_END
;
3769 flags
|= CEPH_READDIR_FRAG_COMPLETE
; // FIXME: what purpose does this serve
3771 // client only understand END and COMPLETE flags ?
3772 if (req_flags
& CEPH_READDIR_REPLY_BITFLAGS
) {
3773 flags
|= CEPH_READDIR_HASH_ORDER
| CEPH_READDIR_OFFSET_HASH
;
3776 // finish final blob
3777 ::encode(numfiles
, dirbl
);
3778 ::encode(flags
, dirbl
);
3779 dirbl
.claim_append(dnbl
);
3782 dout(10) << "reply to " << *req
<< " readdir num=" << numfiles
3783 << " bytes=" << dirbl
.length()
3784 << " start=" << (int)start
3785 << " end=" << (int)end
3787 mdr
->reply_extra_bl
= dirbl
;
3789 // bump popularity. NOTE: this doesn't quite capture it.
3790 mds
->balancer
->hit_dir(now
, dir
, META_POP_IRD
, -1, numfiles
);
3794 respond_to_request(mdr
, 0);
3799 // ===============================================================================
3804 * finisher for basic inode updates
3806 class C_MDS_inode_update_finish
: public ServerLogContext
{
3808 bool truncating_smaller
, changed_ranges
;
3810 C_MDS_inode_update_finish(Server
*s
, MDRequestRef
& r
, CInode
*i
,
3811 bool sm
=false, bool cr
=false) :
3812 ServerLogContext(s
, r
), in(i
), truncating_smaller(sm
), changed_ranges(cr
) { }
3813 void finish(int r
) override
{
3817 in
->pop_and_dirty_projected_inode(mdr
->ls
);
3820 // notify any clients
3821 if (truncating_smaller
&& in
->inode
.is_truncating()) {
3822 get_mds()->locker
->issue_truncate(in
);
3823 get_mds()->mdcache
->truncate_inode(in
, mdr
->ls
);
3826 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
3828 server
->respond_to_request(mdr
, 0);
3831 get_mds()->locker
->share_inode_max_size(in
);
3835 void Server::handle_client_file_setlock(MDRequestRef
& mdr
)
3837 MClientRequest
*req
= mdr
->client_request
;
3838 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3840 // get the inode to operate on, and set up any locks needed for that
3841 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3845 xlocks
.insert(&cur
->flocklock
);
3846 /* acquire_locks will return true if it gets the locks. If it fails,
3847 it will redeliver this request at a later date, so drop the request.
3849 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
)) {
3850 dout(10) << "handle_client_file_setlock could not get locks!" << dendl
;
3854 // copy the lock change into a ceph_filelock so we can store/apply it
3855 ceph_filelock set_lock
;
3856 set_lock
.start
= req
->head
.args
.filelock_change
.start
;
3857 set_lock
.length
= req
->head
.args
.filelock_change
.length
;
3858 set_lock
.client
= req
->get_orig_source().num();
3859 set_lock
.owner
= req
->head
.args
.filelock_change
.owner
;
3860 set_lock
.pid
= req
->head
.args
.filelock_change
.pid
;
3861 set_lock
.type
= req
->head
.args
.filelock_change
.type
;
3862 bool will_wait
= req
->head
.args
.filelock_change
.wait
;
3864 dout(10) << "handle_client_file_setlock: " << set_lock
<< dendl
;
3866 ceph_lock_state_t
*lock_state
= NULL
;
3867 bool interrupt
= false;
3869 // get the appropriate lock state
3870 switch (req
->head
.args
.filelock_change
.rule
) {
3871 case CEPH_LOCK_FLOCK_INTR
:
3874 case CEPH_LOCK_FLOCK
:
3875 lock_state
= cur
->get_flock_lock_state();
3878 case CEPH_LOCK_FCNTL_INTR
:
3881 case CEPH_LOCK_FCNTL
:
3882 lock_state
= cur
->get_fcntl_lock_state();
3886 dout(10) << "got unknown lock type " << set_lock
.type
3887 << ", dropping request!" << dendl
;
3888 respond_to_request(mdr
, -EOPNOTSUPP
);
3892 dout(10) << " state prior to lock change: " << *lock_state
<< dendl
;
3893 if (CEPH_LOCK_UNLOCK
== set_lock
.type
) {
3894 list
<ceph_filelock
> activated_locks
;
3895 list
<MDSInternalContextBase
*> waiters
;
3896 if (lock_state
->is_waiting(set_lock
)) {
3897 dout(10) << " unlock removing waiting lock " << set_lock
<< dendl
;
3898 lock_state
->remove_waiting(set_lock
);
3899 cur
->take_waiting(CInode::WAIT_FLOCK
, waiters
);
3900 } else if (!interrupt
) {
3901 dout(10) << " unlock attempt on " << set_lock
<< dendl
;
3902 lock_state
->remove_lock(set_lock
, activated_locks
);
3903 cur
->take_waiting(CInode::WAIT_FLOCK
, waiters
);
3905 mds
->queue_waiters(waiters
);
3907 respond_to_request(mdr
, 0);
3909 dout(10) << " lock attempt on " << set_lock
<< dendl
;
3910 bool deadlock
= false;
3911 if (mdr
->more()->flock_was_waiting
&&
3912 !lock_state
->is_waiting(set_lock
)) {
3913 dout(10) << " was waiting for lock but not anymore, must have been canceled " << set_lock
<< dendl
;
3914 respond_to_request(mdr
, -EINTR
);
3915 } else if (!lock_state
->add_lock(set_lock
, will_wait
, mdr
->more()->flock_was_waiting
, &deadlock
)) {
3916 dout(10) << " it failed on this attempt" << dendl
;
3917 // couldn't set lock right now
3919 respond_to_request(mdr
, -EDEADLK
);
3920 } else if (!will_wait
) {
3921 respond_to_request(mdr
, -EWOULDBLOCK
);
3923 dout(10) << " added to waiting list" << dendl
;
3924 assert(lock_state
->is_waiting(set_lock
));
3925 mdr
->more()->flock_was_waiting
= true;
3926 mds
->locker
->drop_locks(mdr
.get());
3927 mdr
->drop_local_auth_pins();
3928 cur
->add_waiter(CInode::WAIT_FLOCK
, new C_MDS_RetryRequest(mdcache
, mdr
));
3931 respond_to_request(mdr
, 0);
3933 dout(10) << " state after lock change: " << *lock_state
<< dendl
;
3936 void Server::handle_client_file_readlock(MDRequestRef
& mdr
)
3938 MClientRequest
*req
= mdr
->client_request
;
3939 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3941 // get the inode to operate on, and set up any locks needed for that
3942 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3946 /* acquire_locks will return true if it gets the locks. If it fails,
3947 it will redeliver this request at a later date, so drop the request.
3949 rdlocks
.insert(&cur
->flocklock
);
3950 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
)) {
3951 dout(10) << "handle_client_file_readlock could not get locks!" << dendl
;
3955 // copy the lock change into a ceph_filelock so we can store/apply it
3956 ceph_filelock checking_lock
;
3957 checking_lock
.start
= req
->head
.args
.filelock_change
.start
;
3958 checking_lock
.length
= req
->head
.args
.filelock_change
.length
;
3959 checking_lock
.client
= req
->get_orig_source().num();
3960 checking_lock
.owner
= req
->head
.args
.filelock_change
.owner
;
3961 checking_lock
.pid
= req
->head
.args
.filelock_change
.pid
;
3962 checking_lock
.type
= req
->head
.args
.filelock_change
.type
;
3964 // get the appropriate lock state
3965 ceph_lock_state_t
*lock_state
= NULL
;
3966 switch (req
->head
.args
.filelock_change
.rule
) {
3967 case CEPH_LOCK_FLOCK
:
3968 lock_state
= cur
->get_flock_lock_state();
3971 case CEPH_LOCK_FCNTL
:
3972 lock_state
= cur
->get_fcntl_lock_state();
3976 dout(10) << "got unknown lock type " << checking_lock
.type
<< dendl
;
3977 respond_to_request(mdr
, -EINVAL
);
3980 lock_state
->look_for_lock(checking_lock
);
3983 ::encode(checking_lock
, lock_bl
);
3985 mdr
->reply_extra_bl
= lock_bl
;
3986 respond_to_request(mdr
, 0);
3989 void Server::handle_client_setattr(MDRequestRef
& mdr
)
3991 MClientRequest
*req
= mdr
->client_request
;
3992 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
3993 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
3996 if (mdr
->snapid
!= CEPH_NOSNAP
) {
3997 respond_to_request(mdr
, -EROFS
);
4000 if (cur
->ino() < MDS_INO_SYSTEM_BASE
&& !cur
->is_base()) {
4001 respond_to_request(mdr
, -EPERM
);
4005 __u32 mask
= req
->head
.args
.setattr
.mask
;
4006 __u32 access_mask
= MAY_WRITE
;
4009 if (mask
& (CEPH_SETATTR_MODE
|CEPH_SETATTR_UID
|CEPH_SETATTR_GID
|CEPH_SETATTR_BTIME
|CEPH_SETATTR_KILL_SGUID
))
4010 xlocks
.insert(&cur
->authlock
);
4011 if (mask
& (CEPH_SETATTR_MTIME
|CEPH_SETATTR_ATIME
|CEPH_SETATTR_SIZE
))
4012 xlocks
.insert(&cur
->filelock
);
4013 if (mask
& CEPH_SETATTR_CTIME
)
4014 wrlocks
.insert(&cur
->versionlock
);
4016 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4019 if ((mask
& CEPH_SETATTR_UID
) && (cur
->inode
.uid
!= req
->head
.args
.setattr
.uid
))
4020 access_mask
|= MAY_CHOWN
;
4022 if ((mask
& CEPH_SETATTR_GID
) && (cur
->inode
.gid
!= req
->head
.args
.setattr
.gid
))
4023 access_mask
|= MAY_CHGRP
;
4025 if (!check_access(mdr
, cur
, access_mask
))
4028 // trunc from bigger -> smaller?
4029 inode_t
*pi
= cur
->get_projected_inode();
4031 uint64_t old_size
= MAX(pi
->size
, req
->head
.args
.setattr
.old_size
);
4033 // ENOSPC on growing file while full, but allow shrinks
4034 if (is_full
&& req
->head
.args
.setattr
.size
> old_size
) {
4035 dout(20) << __func__
<< ": full, responding ENOSPC to setattr with larger size" << dendl
;
4036 respond_to_request(mdr
, -ENOSPC
);
4040 bool truncating_smaller
= false;
4041 if (mask
& CEPH_SETATTR_SIZE
) {
4042 truncating_smaller
= req
->head
.args
.setattr
.size
< old_size
;
4043 if (truncating_smaller
&& pi
->is_truncating()) {
4044 dout(10) << " waiting for pending truncate from " << pi
->truncate_from
4045 << " to " << pi
->truncate_size
<< " to complete on " << *cur
<< dendl
;
4046 mds
->locker
->drop_locks(mdr
.get());
4047 mdr
->drop_local_auth_pins();
4048 cur
->add_waiter(CInode::WAIT_TRUNC
, new C_MDS_RetryRequest(mdcache
, mdr
));
4053 bool changed_ranges
= false;
4056 mdr
->ls
= mdlog
->get_current_segment();
4057 EUpdate
*le
= new EUpdate(mdlog
, "setattr");
4058 mdlog
->start_entry(le
);
4060 pi
= cur
->project_inode();
4062 if (mask
& CEPH_SETATTR_UID
)
4063 pi
->uid
= req
->head
.args
.setattr
.uid
;
4064 if (mask
& CEPH_SETATTR_GID
)
4065 pi
->gid
= req
->head
.args
.setattr
.gid
;
4067 if (mask
& CEPH_SETATTR_MODE
)
4068 pi
->mode
= (pi
->mode
& ~07777) | (req
->head
.args
.setattr
.mode
& 07777);
4069 else if ((mask
& (CEPH_SETATTR_UID
|CEPH_SETATTR_GID
|CEPH_SETATTR_KILL_SGUID
)) &&
4070 S_ISREG(pi
->mode
)) {
4071 pi
->mode
&= ~S_ISUID
;
4072 if ((pi
->mode
& (S_ISGID
|S_IXGRP
)) == (S_ISGID
|S_IXGRP
))
4073 pi
->mode
&= ~S_ISGID
;
4076 if (mask
& CEPH_SETATTR_MTIME
)
4077 pi
->mtime
= req
->head
.args
.setattr
.mtime
;
4078 if (mask
& CEPH_SETATTR_ATIME
)
4079 pi
->atime
= req
->head
.args
.setattr
.atime
;
4080 if (mask
& CEPH_SETATTR_BTIME
)
4081 pi
->btime
= req
->head
.args
.setattr
.btime
;
4082 if (mask
& (CEPH_SETATTR_ATIME
| CEPH_SETATTR_MTIME
| CEPH_SETATTR_BTIME
))
4083 pi
->time_warp_seq
++; // maybe not a timewarp, but still a serialization point.
4084 if (mask
& CEPH_SETATTR_SIZE
) {
4085 if (truncating_smaller
) {
4086 pi
->truncate(old_size
, req
->head
.args
.setattr
.size
);
4087 le
->metablob
.add_truncate_start(cur
->ino());
4089 pi
->size
= req
->head
.args
.setattr
.size
;
4090 pi
->rstat
.rbytes
= pi
->size
;
4092 pi
->mtime
= mdr
->get_op_stamp();
4094 // adjust client's max_size?
4095 map
<client_t
,client_writeable_range_t
> new_ranges
;
4096 bool max_increased
= false;
4097 mds
->locker
->calc_new_client_ranges(cur
, pi
->size
, &new_ranges
, &max_increased
);
4098 if (pi
->client_ranges
!= new_ranges
) {
4099 dout(10) << " client_ranges " << pi
->client_ranges
<< " -> " << new_ranges
<< dendl
;
4100 pi
->client_ranges
= new_ranges
;
4101 changed_ranges
= true;
4105 pi
->version
= cur
->pre_dirty();
4106 pi
->ctime
= mdr
->get_op_stamp();
4110 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4111 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4112 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4114 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
,
4115 truncating_smaller
, changed_ranges
));
4117 // flush immediately if there are readers/writers waiting
4118 if (xlocks
.count(&cur
->filelock
) &&
4119 (cur
->get_caps_wanted() & (CEPH_CAP_FILE_RD
|CEPH_CAP_FILE_WR
)))
4120 mds
->mdlog
->flush();
4123 /* Takes responsibility for mdr */
4124 void Server::do_open_truncate(MDRequestRef
& mdr
, int cmode
)
4126 CInode
*in
= mdr
->in
[0];
4127 client_t client
= mdr
->get_client();
4130 dout(10) << "do_open_truncate " << *in
<< dendl
;
4132 SnapRealm
*realm
= in
->find_snaprealm();
4133 mds
->locker
->issue_new_caps(in
, cmode
, mdr
->session
, realm
, mdr
->client_request
->is_replay());
4135 mdr
->ls
= mdlog
->get_current_segment();
4136 EUpdate
*le
= new EUpdate(mdlog
, "open_truncate");
4137 mdlog
->start_entry(le
);
4140 inode_t
*pi
= in
->project_inode();
4141 pi
->version
= in
->pre_dirty();
4142 pi
->mtime
= pi
->ctime
= mdr
->get_op_stamp();
4145 uint64_t old_size
= MAX(pi
->size
, mdr
->client_request
->head
.args
.open
.old_size
);
4147 pi
->truncate(old_size
, 0);
4148 le
->metablob
.add_truncate_start(in
->ino());
4151 bool changed_ranges
= false;
4152 if (cmode
& CEPH_FILE_MODE_WR
) {
4153 pi
->client_ranges
[client
].range
.first
= 0;
4154 pi
->client_ranges
[client
].range
.last
= pi
->get_layout_size_increment();
4155 pi
->client_ranges
[client
].follows
= in
->find_snaprealm()->get_newest_seq();
4156 changed_ranges
= true;
4159 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
4161 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, 0, PREDIRTY_PRIMARY
);
4162 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, in
);
4164 // make sure ino gets into the journal
4165 le
->metablob
.add_opened_ino(in
->ino());
4166 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
4167 ls
->open_files
.push_back(&in
->item_open_file
);
4169 mdr
->o_trunc
= true;
4172 if (mdr
->client_request
->get_dentry_wanted()) {
4173 assert(mdr
->dn
[0].size());
4174 dn
= mdr
->dn
[0].back();
4177 journal_and_reply(mdr
, in
, dn
, le
, new C_MDS_inode_update_finish(this, mdr
, in
, old_size
> 0,
4179 // Although the `open` part can give an early reply, the truncation won't
4180 // happen until our EUpdate is persistent, to give the client a prompt
4181 // response we must also flush that event.
4186 /* This function cleans up the passed mdr */
4187 void Server::handle_client_setlayout(MDRequestRef
& mdr
)
4189 MClientRequest
*req
= mdr
->client_request
;
4190 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4191 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4194 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4195 respond_to_request(mdr
, -EROFS
);
4198 if (!cur
->is_file()) {
4199 respond_to_request(mdr
, -EINVAL
);
4202 if (cur
->get_projected_inode()->size
||
4203 cur
->get_projected_inode()->truncate_seq
> 1) {
4204 respond_to_request(mdr
, -ENOTEMPTY
);
4209 file_layout_t layout
= cur
->get_projected_inode()->layout
;
4210 // save existing layout for later
4211 const auto old_layout
= layout
;
4213 int access
= MAY_WRITE
;
4215 if (req
->head
.args
.setlayout
.layout
.fl_object_size
> 0)
4216 layout
.object_size
= req
->head
.args
.setlayout
.layout
.fl_object_size
;
4217 if (req
->head
.args
.setlayout
.layout
.fl_stripe_unit
> 0)
4218 layout
.stripe_unit
= req
->head
.args
.setlayout
.layout
.fl_stripe_unit
;
4219 if (req
->head
.args
.setlayout
.layout
.fl_stripe_count
> 0)
4220 layout
.stripe_count
=req
->head
.args
.setlayout
.layout
.fl_stripe_count
;
4221 if (req
->head
.args
.setlayout
.layout
.fl_pg_pool
> 0) {
4222 layout
.pool_id
= req
->head
.args
.setlayout
.layout
.fl_pg_pool
;
4224 // make sure we have as new a map as the client
4225 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
4226 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
4231 // Don't permit layout modifications without 'p' caps
4232 if (layout
!= old_layout
) {
4233 access
|= MAY_SET_VXATTR
;
4236 if (!layout
.is_valid()) {
4237 dout(10) << "bad layout" << dendl
;
4238 respond_to_request(mdr
, -EINVAL
);
4241 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
4242 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
4243 respond_to_request(mdr
, -EINVAL
);
4247 xlocks
.insert(&cur
->filelock
);
4248 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4251 if (!check_access(mdr
, cur
, access
))
4255 inode_t
*pi
= cur
->project_inode();
4256 pi
->layout
= layout
;
4257 // add the old pool to the inode
4258 pi
->add_old_pool(old_layout
.pool_id
);
4259 pi
->version
= cur
->pre_dirty();
4260 pi
->ctime
= mdr
->get_op_stamp();
4264 mdr
->ls
= mdlog
->get_current_segment();
4265 EUpdate
*le
= new EUpdate(mdlog
, "setlayout");
4266 mdlog
->start_entry(le
);
4267 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4268 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4269 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4271 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4274 void Server::handle_client_setdirlayout(MDRequestRef
& mdr
)
4276 MClientRequest
*req
= mdr
->client_request
;
4277 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4278 file_layout_t
*dir_layout
= NULL
;
4279 CInode
*cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4282 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4283 respond_to_request(mdr
, -EROFS
);
4287 if (!cur
->is_dir()) {
4288 respond_to_request(mdr
, -ENOTDIR
);
4292 xlocks
.insert(&cur
->policylock
);
4293 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4297 const inode_t
*old_pi
= cur
->get_projected_inode();
4298 file_layout_t layout
;
4299 if (old_pi
->has_layout())
4300 layout
= old_pi
->layout
;
4301 else if (dir_layout
)
4302 layout
= *dir_layout
;
4304 layout
= mdcache
->default_file_layout
;
4306 // Level of access required to complete
4307 int access
= MAY_WRITE
;
4309 const auto old_layout
= layout
;
4311 if (req
->head
.args
.setlayout
.layout
.fl_object_size
> 0)
4312 layout
.object_size
= req
->head
.args
.setlayout
.layout
.fl_object_size
;
4313 if (req
->head
.args
.setlayout
.layout
.fl_stripe_unit
> 0)
4314 layout
.stripe_unit
= req
->head
.args
.setlayout
.layout
.fl_stripe_unit
;
4315 if (req
->head
.args
.setlayout
.layout
.fl_stripe_count
> 0)
4316 layout
.stripe_count
=req
->head
.args
.setlayout
.layout
.fl_stripe_count
;
4317 if (req
->head
.args
.setlayout
.layout
.fl_pg_pool
> 0) {
4318 layout
.pool_id
= req
->head
.args
.setlayout
.layout
.fl_pg_pool
;
4319 // make sure we have as new a map as the client
4320 if (req
->get_mdsmap_epoch() > mds
->mdsmap
->get_epoch()) {
4321 mds
->wait_for_mdsmap(req
->get_mdsmap_epoch(), new C_MDS_RetryRequest(mdcache
, mdr
));
4326 if (layout
!= old_layout
) {
4327 access
|= MAY_SET_VXATTR
;
4330 if (!layout
.is_valid()) {
4331 dout(10) << "bad layout" << dendl
;
4332 respond_to_request(mdr
, -EINVAL
);
4335 if (!mds
->mdsmap
->is_data_pool(layout
.pool_id
)) {
4336 dout(10) << " invalid data pool " << layout
.pool_id
<< dendl
;
4337 respond_to_request(mdr
, -EINVAL
);
4341 if (!check_access(mdr
, cur
, access
))
4344 inode_t
*pi
= cur
->project_inode();
4345 pi
->layout
= layout
;
4346 pi
->version
= cur
->pre_dirty();
4349 mdr
->ls
= mdlog
->get_current_segment();
4350 EUpdate
*le
= new EUpdate(mdlog
, "setlayout");
4351 mdlog
->start_entry(le
);
4352 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4353 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4354 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4356 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4361 int Server::parse_layout_vxattr(string name
, string value
, const OSDMap
& osdmap
,
4362 file_layout_t
*layout
, bool validate
)
4364 dout(20) << "parse_layout_vxattr name " << name
<< " value '" << value
<< "'" << dendl
;
4366 if (name
== "layout") {
4367 string::iterator begin
= value
.begin();
4368 string::iterator end
= value
.end();
4369 keys_and_values
<string::iterator
> p
; // create instance of parser
4370 std::map
<string
, string
> m
; // map to receive results
4371 if (!qi::parse(begin
, end
, p
, m
)) { // returns true if successful
4374 string
left(begin
, end
);
4375 dout(10) << " parsed " << m
<< " left '" << left
<< "'" << dendl
;
4378 for (map
<string
,string
>::iterator q
= m
.begin(); q
!= m
.end(); ++q
) {
4379 // Skip validation on each attr, we do it once at the end (avoid
4380 // rejecting intermediate states if the overall result is ok)
4381 int r
= parse_layout_vxattr(string("layout.") + q
->first
, q
->second
,
4382 osdmap
, layout
, false);
4386 } else if (name
== "layout.object_size") {
4387 layout
->object_size
= boost::lexical_cast
<unsigned>(value
);
4388 } else if (name
== "layout.stripe_unit") {
4389 layout
->stripe_unit
= boost::lexical_cast
<unsigned>(value
);
4390 } else if (name
== "layout.stripe_count") {
4391 layout
->stripe_count
= boost::lexical_cast
<unsigned>(value
);
4392 } else if (name
== "layout.pool") {
4394 layout
->pool_id
= boost::lexical_cast
<unsigned>(value
);
4395 } catch (boost::bad_lexical_cast
const&) {
4396 int64_t pool
= osdmap
.lookup_pg_pool_name(value
);
4398 dout(10) << " unknown pool " << value
<< dendl
;
4401 layout
->pool_id
= pool
;
4403 } else if (name
== "layout.pool_namespace") {
4404 layout
->pool_ns
= value
;
4406 dout(10) << " unknown layout vxattr " << name
<< dendl
;
4409 } catch (boost::bad_lexical_cast
const&) {
4410 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4414 if (validate
&& !layout
->is_valid()) {
4415 dout(10) << "bad layout" << dendl
;
4418 if (!mds
->mdsmap
->is_data_pool(layout
->pool_id
)) {
4419 dout(10) << " invalid data pool " << layout
->pool_id
<< dendl
;
4425 int Server::parse_quota_vxattr(string name
, string value
, quota_info_t
*quota
)
4427 dout(20) << "parse_quota_vxattr name " << name
<< " value '" << value
<< "'" << dendl
;
4429 if (name
== "quota") {
4430 string::iterator begin
= value
.begin();
4431 string::iterator end
= value
.end();
4432 keys_and_values
<string::iterator
> p
; // create instance of parser
4433 std::map
<string
, string
> m
; // map to receive results
4434 if (!qi::parse(begin
, end
, p
, m
)) { // returns true if successful
4437 string
left(begin
, end
);
4438 dout(10) << " parsed " << m
<< " left '" << left
<< "'" << dendl
;
4441 for (map
<string
,string
>::iterator q
= m
.begin(); q
!= m
.end(); ++q
) {
4442 int r
= parse_quota_vxattr(string("quota.") + q
->first
, q
->second
, quota
);
4446 } else if (name
== "quota.max_bytes") {
4447 int64_t q
= boost::lexical_cast
<int64_t>(value
);
4450 quota
->max_bytes
= q
;
4451 } else if (name
== "quota.max_files") {
4452 int64_t q
= boost::lexical_cast
<int64_t>(value
);
4455 quota
->max_files
= q
;
4457 dout(10) << " unknown quota vxattr " << name
<< dendl
;
4460 } catch (boost::bad_lexical_cast
const&) {
4461 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4465 if (!quota
->is_valid()) {
4466 dout(10) << "bad quota" << dendl
;
4473 * Verify that the file layout attribute carried by client
4474 * is well-formatted.
4475 * Return 0 on success, otherwise this function takes
4476 * responsibility for the passed mdr.
4478 int Server::check_layout_vxattr(MDRequestRef
& mdr
,
4481 file_layout_t
*layout
)
4483 MClientRequest
*req
= mdr
->client_request
;
4487 mds
->objecter
->with_osdmap([&](const OSDMap
& osdmap
) {
4488 r
= parse_layout_vxattr(name
, value
, osdmap
, layout
);
4489 epoch
= osdmap
.get_epoch();
4494 // we don't have the specified pool, make sure our map
4495 // is newer than or as new as the client.
4496 epoch_t req_epoch
= req
->get_osdmap_epoch();
4498 if (req_epoch
> epoch
) {
4500 // well, our map is older. consult mds.
4501 Context
*fin
= new C_IO_Wrapper(mds
, new C_MDS_RetryRequest(mdcache
, mdr
));
4503 if (!mds
->objecter
->wait_for_map(req_epoch
, fin
))
4504 return r
; // wait, fin will retry this request later
4508 // now we have at least as new a map as the client, try again.
4509 mds
->objecter
->with_osdmap([&](const OSDMap
& osdmap
) {
4510 r
= parse_layout_vxattr(name
, value
, osdmap
, layout
);
4511 epoch
= osdmap
.get_epoch();
4514 assert(epoch
>= req_epoch
); // otherwise wait_for_map() told a lie
4516 } else if (req_epoch
== 0 && !mdr
->waited_for_osdmap
) {
4518 // For compatibility with client w/ old code, we still need get the
4519 // latest map. One day if COMPACT_VERSION of MClientRequest >=3,
4520 // we can remove those code.
4521 mdr
->waited_for_osdmap
= true;
4522 mds
->objecter
->wait_for_latest_osdmap(new C_IO_Wrapper(
4523 mds
, new C_MDS_RetryRequest(mdcache
, mdr
)));
4533 respond_to_request(mdr
, r
);
4541 void Server::handle_set_vxattr(MDRequestRef
& mdr
, CInode
*cur
,
4542 file_layout_t
*dir_layout
,
4543 set
<SimpleLock
*> rdlocks
,
4544 set
<SimpleLock
*> wrlocks
,
4545 set
<SimpleLock
*> xlocks
)
4547 MClientRequest
*req
= mdr
->client_request
;
4548 string
name(req
->get_path2());
4549 bufferlist bl
= req
->get_data();
4550 string
value (bl
.c_str(), bl
.length());
4551 dout(10) << "handle_set_vxattr " << name
4552 << " val " << value
.length()
4553 << " bytes on " << *cur
4559 if (!check_access(mdr
, cur
, MAY_SET_VXATTR
)) {
4563 if (name
.compare(0, 15, "ceph.dir.layout") == 0) {
4564 if (!cur
->is_dir()) {
4565 respond_to_request(mdr
, -EINVAL
);
4569 file_layout_t layout
;
4570 if (cur
->get_projected_inode()->has_layout())
4571 layout
= cur
->get_projected_inode()->layout
;
4572 else if (dir_layout
)
4573 layout
= *dir_layout
;
4575 layout
= mdcache
->default_file_layout
;
4577 rest
= name
.substr(name
.find("layout"));
4578 if (check_layout_vxattr(mdr
, rest
, value
, &layout
) < 0)
4581 xlocks
.insert(&cur
->policylock
);
4582 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4585 pi
= cur
->project_inode();
4586 pi
->layout
= layout
;
4587 } else if (name
.compare(0, 16, "ceph.file.layout") == 0) {
4588 if (!cur
->is_file()) {
4589 respond_to_request(mdr
, -EINVAL
);
4592 if (cur
->get_projected_inode()->size
||
4593 cur
->get_projected_inode()->truncate_seq
> 1) {
4594 respond_to_request(mdr
, -ENOTEMPTY
);
4597 file_layout_t layout
= cur
->get_projected_inode()->layout
;
4598 rest
= name
.substr(name
.find("layout"));
4599 if (check_layout_vxattr(mdr
, rest
, value
, &layout
) < 0)
4602 xlocks
.insert(&cur
->filelock
);
4603 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4606 pi
= cur
->project_inode();
4607 int64_t old_pool
= pi
->layout
.pool_id
;
4608 pi
->add_old_pool(old_pool
);
4609 pi
->layout
= layout
;
4610 pi
->ctime
= mdr
->get_op_stamp();
4611 } else if (name
.compare(0, 10, "ceph.quota") == 0) {
4612 if (!cur
->is_dir() || cur
->is_root()) {
4613 respond_to_request(mdr
, -EINVAL
);
4617 quota_info_t quota
= cur
->get_projected_inode()->quota
;
4619 rest
= name
.substr(name
.find("quota"));
4620 int r
= parse_quota_vxattr(rest
, value
, "a
);
4622 respond_to_request(mdr
, r
);
4626 xlocks
.insert(&cur
->policylock
);
4627 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4630 pi
= cur
->project_inode();
4632 } else if (name
.find("ceph.dir.pin") == 0) {
4633 if (!cur
->is_dir() || cur
->is_root()) {
4634 respond_to_request(mdr
, -EINVAL
);
4640 rank
= boost::lexical_cast
<mds_rank_t
>(value
);
4641 if (rank
< 0) rank
= MDS_RANK_NONE
;
4642 } catch (boost::bad_lexical_cast
const&) {
4643 dout(10) << "bad vxattr value, unable to parse int for " << name
<< dendl
;
4644 respond_to_request(mdr
, -EINVAL
);
4648 xlocks
.insert(&cur
->policylock
);
4649 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4652 pi
= cur
->project_inode();
4653 cur
->set_export_pin(rank
);
4655 dout(10) << " unknown vxattr " << name
<< dendl
;
4656 respond_to_request(mdr
, -EINVAL
);
4661 pi
->ctime
= mdr
->get_op_stamp();
4662 pi
->version
= cur
->pre_dirty();
4664 pi
->update_backtrace();
4667 mdr
->ls
= mdlog
->get_current_segment();
4668 EUpdate
*le
= new EUpdate(mdlog
, "set vxattr layout");
4669 mdlog
->start_entry(le
);
4670 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4671 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4672 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4674 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4678 void Server::handle_remove_vxattr(MDRequestRef
& mdr
, CInode
*cur
,
4679 file_layout_t
*dir_layout
,
4680 set
<SimpleLock
*> rdlocks
,
4681 set
<SimpleLock
*> wrlocks
,
4682 set
<SimpleLock
*> xlocks
)
4684 MClientRequest
*req
= mdr
->client_request
;
4685 string
name(req
->get_path2());
4687 dout(10) << __func__
<< " " << name
<< " on " << *cur
<< dendl
;
4689 if (name
== "ceph.dir.layout") {
4690 if (!cur
->is_dir()) {
4691 respond_to_request(mdr
, -ENODATA
);
4694 if (cur
->is_root()) {
4695 dout(10) << "can't remove layout policy on the root directory" << dendl
;
4696 respond_to_request(mdr
, -EINVAL
);
4700 if (!cur
->get_projected_inode()->has_layout()) {
4701 respond_to_request(mdr
, -ENODATA
);
4705 xlocks
.insert(&cur
->policylock
);
4706 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4709 inode_t
*pi
= cur
->project_inode();
4711 pi
->version
= cur
->pre_dirty();
4714 mdr
->ls
= mdlog
->get_current_segment();
4715 EUpdate
*le
= new EUpdate(mdlog
, "remove dir layout vxattr");
4716 mdlog
->start_entry(le
);
4717 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4718 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4719 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4721 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4723 } else if (name
== "ceph.dir.layout.pool_namespace"
4724 || name
== "ceph.file.layout.pool_namespace") {
4725 // Namespace is the only layout field that has a meaningful
4726 // null/none value (empty string, means default layout). Is equivalent
4727 // to a setxattr with empty string: pass through the empty payload of
4728 // the rmxattr request to do this.
4729 handle_set_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4733 respond_to_request(mdr
, -ENODATA
);
4736 class C_MDS_inode_xattr_update_finish
: public ServerLogContext
{
4740 C_MDS_inode_xattr_update_finish(Server
*s
, MDRequestRef
& r
, CInode
*i
) :
4741 ServerLogContext(s
, r
), in(i
) { }
4742 void finish(int r
) override
{
4746 in
->pop_and_dirty_projected_inode(mdr
->ls
);
4750 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
4752 server
->respond_to_request(mdr
, 0);
4756 void Server::handle_client_setxattr(MDRequestRef
& mdr
)
4758 MClientRequest
*req
= mdr
->client_request
;
4759 string
name(req
->get_path2());
4760 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4763 file_layout_t
*dir_layout
= NULL
;
4764 if (name
.compare(0, 15, "ceph.dir.layout") == 0)
4765 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4767 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4771 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4772 respond_to_request(mdr
, -EROFS
);
4776 int flags
= req
->head
.args
.setxattr
.flags
;
4778 // magic ceph.* namespace?
4779 if (name
.compare(0, 5, "ceph.") == 0) {
4780 handle_set_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4784 xlocks
.insert(&cur
->xattrlock
);
4785 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4788 if (!check_access(mdr
, cur
, MAY_WRITE
))
4791 map
<string
, bufferptr
> *pxattrs
= cur
->get_projected_xattrs();
4792 size_t len
= req
->get_data().length();
4793 size_t inc
= len
+ name
.length();
4795 // check xattrs kv pairs size
4796 size_t cur_xattrs_size
= 0;
4797 for (const auto& p
: *pxattrs
) {
4798 if ((flags
& CEPH_XATTR_REPLACE
) && (name
.compare(p
.first
) == 0)) {
4801 cur_xattrs_size
+= p
.first
.length() + p
.second
.length();
4804 if (((cur_xattrs_size
+ inc
) > g_conf
->mds_max_xattr_pairs_size
)) {
4805 dout(10) << "xattr kv pairs size too big. cur_xattrs_size "
4806 << cur_xattrs_size
<< ", inc " << inc
<< dendl
;
4807 respond_to_request(mdr
, -ENOSPC
);
4811 if ((flags
& CEPH_XATTR_CREATE
) && pxattrs
->count(name
)) {
4812 dout(10) << "setxattr '" << name
<< "' XATTR_CREATE and EEXIST on " << *cur
<< dendl
;
4813 respond_to_request(mdr
, -EEXIST
);
4816 if ((flags
& CEPH_XATTR_REPLACE
) && !pxattrs
->count(name
)) {
4817 dout(10) << "setxattr '" << name
<< "' XATTR_REPLACE and ENODATA on " << *cur
<< dendl
;
4818 respond_to_request(mdr
, -ENODATA
);
4822 dout(10) << "setxattr '" << name
<< "' len " << len
<< " on " << *cur
<< dendl
;
4825 map
<string
,bufferptr
> *px
= new map
<string
,bufferptr
>;
4826 inode_t
*pi
= cur
->project_inode(px
);
4827 pi
->version
= cur
->pre_dirty();
4828 pi
->ctime
= mdr
->get_op_stamp();
4830 pi
->xattr_version
++;
4832 if (!(flags
& CEPH_XATTR_REMOVE
)) {
4833 (*px
)[name
] = buffer::create(len
);
4835 req
->get_data().copy(0, len
, (*px
)[name
].c_str());
4839 mdr
->ls
= mdlog
->get_current_segment();
4840 EUpdate
*le
= new EUpdate(mdlog
, "setxattr");
4841 mdlog
->start_entry(le
);
4842 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4843 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4844 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4846 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4849 void Server::handle_client_removexattr(MDRequestRef
& mdr
)
4851 MClientRequest
*req
= mdr
->client_request
;
4852 string
name(req
->get_path2());
4853 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4854 file_layout_t
*dir_layout
= NULL
;
4856 if (name
== "ceph.dir.layout")
4857 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true, false, &dir_layout
);
4859 cur
= rdlock_path_pin_ref(mdr
, 0, rdlocks
, true);
4863 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4864 respond_to_request(mdr
, -EROFS
);
4868 if (name
.compare(0, 5, "ceph.") == 0) {
4869 handle_remove_vxattr(mdr
, cur
, dir_layout
, rdlocks
, wrlocks
, xlocks
);
4873 xlocks
.insert(&cur
->xattrlock
);
4874 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4877 map
<string
, bufferptr
> *pxattrs
= cur
->get_projected_xattrs();
4878 if (pxattrs
->count(name
) == 0) {
4879 dout(10) << "removexattr '" << name
<< "' and ENODATA on " << *cur
<< dendl
;
4880 respond_to_request(mdr
, -ENODATA
);
4884 dout(10) << "removexattr '" << name
<< "' on " << *cur
<< dendl
;
4887 map
<string
,bufferptr
> *px
= new map
<string
,bufferptr
>;
4888 inode_t
*pi
= cur
->project_inode(px
);
4889 pi
->version
= cur
->pre_dirty();
4890 pi
->ctime
= mdr
->get_op_stamp();
4892 pi
->xattr_version
++;
4896 mdr
->ls
= mdlog
->get_current_segment();
4897 EUpdate
*le
= new EUpdate(mdlog
, "removexattr");
4898 mdlog
->start_entry(le
);
4899 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
4900 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, cur
, 0, PREDIRTY_PRIMARY
);
4901 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, cur
);
4903 journal_and_reply(mdr
, cur
, 0, le
, new C_MDS_inode_update_finish(this, mdr
, cur
));
4907 // =================================================================
4908 // DIRECTORY and NAMESPACE OPS
4911 // ------------------------------------------------
4915 class C_MDS_mknod_finish
: public ServerLogContext
{
4919 C_MDS_mknod_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ni
) :
4920 ServerLogContext(s
, r
), dn(d
), newi(ni
) {}
4921 void finish(int r
) override
{
4925 dn
->pop_projected_linkage();
4927 // be a bit hacky with the inode version, here.. we decrement it
4928 // just to keep mark_dirty() happen. (we didn't bother projecting
4929 // a new version of hte inode since it's just been created)
4930 newi
->inode
.version
--;
4931 newi
->mark_dirty(newi
->inode
.version
+ 1, mdr
->ls
);
4932 newi
->_mark_dirty_parent(mdr
->ls
, true);
4935 if (newi
->inode
.is_dir()) {
4936 CDir
*dir
= newi
->get_dirfrag(frag_t());
4938 dir
->fnode
.version
--;
4939 dir
->mark_dirty(dir
->fnode
.version
+ 1, mdr
->ls
);
4940 dir
->mark_new(mdr
->ls
);
4945 MDRequestRef null_ref
;
4946 get_mds()->mdcache
->send_dentry_link(dn
, null_ref
);
4948 if (newi
->inode
.is_file())
4949 get_mds()->locker
->share_inode_max_size(newi
);
4952 get_mds()->balancer
->hit_inode(mdr
->get_mds_stamp(), newi
, META_POP_IWR
);
4955 server
->respond_to_request(mdr
, 0);
4960 void Server::handle_client_mknod(MDRequestRef
& mdr
)
4962 MClientRequest
*req
= mdr
->client_request
;
4963 client_t client
= mdr
->get_client();
4964 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
4965 file_layout_t
*dir_layout
= NULL
;
4966 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false,
4969 if (mdr
->snapid
!= CEPH_NOSNAP
) {
4970 respond_to_request(mdr
, -EROFS
);
4973 CInode
*diri
= dn
->get_dir()->get_inode();
4974 rdlocks
.insert(&diri
->authlock
);
4975 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
4978 if (!check_access(mdr
, diri
, MAY_WRITE
))
4981 if (!check_fragment_space(mdr
, dn
->get_dir()))
4984 unsigned mode
= req
->head
.args
.mknod
.mode
;
4985 if ((mode
& S_IFMT
) == 0)
4989 file_layout_t layout
;
4990 if (dir_layout
&& S_ISREG(mode
))
4991 layout
= *dir_layout
;
4993 layout
= mdcache
->default_file_layout
;
4995 SnapRealm
*realm
= dn
->get_dir()->inode
->find_snaprealm();
4996 snapid_t follows
= realm
->get_newest_seq();
4997 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
),
5001 dn
->push_projected_linkage(newi
);
5003 newi
->inode
.rdev
= req
->head
.args
.mknod
.rdev
;
5004 newi
->inode
.version
= dn
->pre_dirty();
5005 newi
->inode
.rstat
.rfiles
= 1;
5006 if (layout
.pool_id
!= mdcache
->default_file_layout
.pool_id
)
5007 newi
->inode
.add_old_pool(mdcache
->default_file_layout
.pool_id
);
5008 newi
->inode
.update_backtrace();
5010 // if the client created a _regular_ file via MKNOD, it's highly likely they'll
5011 // want to write to it (e.g., if they are reexporting NFS)
5012 if (S_ISREG(newi
->inode
.mode
)) {
5013 dout(15) << " setting a client_range too, since this is a regular file" << dendl
;
5014 newi
->inode
.client_ranges
[client
].range
.first
= 0;
5015 newi
->inode
.client_ranges
[client
].range
.last
= newi
->inode
.get_layout_size_increment();
5016 newi
->inode
.client_ranges
[client
].follows
= follows
;
5018 // issue a cap on the file
5019 int cmode
= CEPH_FILE_MODE_RDWR
;
5020 Capability
*cap
= mds
->locker
->issue_new_caps(newi
, cmode
, mdr
->session
, realm
, req
->is_replay());
5024 // put locks in excl mode
5025 newi
->filelock
.set_state(LOCK_EXCL
);
5026 newi
->authlock
.set_state(LOCK_EXCL
);
5027 newi
->xattrlock
.set_state(LOCK_EXCL
);
5031 assert(dn
->first
== follows
+ 1);
5032 newi
->first
= dn
->first
;
5034 dout(10) << "mknod mode " << newi
->inode
.mode
<< " rdev " << newi
->inode
.rdev
<< dendl
;
5037 mdr
->ls
= mdlog
->get_current_segment();
5038 EUpdate
*le
= new EUpdate(mdlog
, "mknod");
5039 mdlog
->start_entry(le
);
5040 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5041 journal_allocated_inos(mdr
, &le
->metablob
);
5043 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(),
5044 PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5045 le
->metablob
.add_primary_dentry(dn
, newi
, true, true, true);
5047 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5053 /* This function takes responsibility for the passed mdr*/
5054 void Server::handle_client_mkdir(MDRequestRef
& mdr
)
5056 MClientRequest
*req
= mdr
->client_request
;
5057 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5058 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5060 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5061 respond_to_request(mdr
, -EROFS
);
5064 CDir
*dir
= dn
->get_dir();
5065 CInode
*diri
= dir
->get_inode();
5066 rdlocks
.insert(&diri
->authlock
);
5067 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5070 // mkdir check access
5071 if (!check_access(mdr
, diri
, MAY_WRITE
))
5074 if (!check_fragment_space(mdr
, dir
))
5078 SnapRealm
*realm
= dn
->get_dir()->inode
->find_snaprealm();
5079 snapid_t follows
= realm
->get_newest_seq();
5081 unsigned mode
= req
->head
.args
.mkdir
.mode
;
5084 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
), mode
);
5087 // it's a directory.
5088 dn
->push_projected_linkage(newi
);
5090 newi
->inode
.version
= dn
->pre_dirty();
5091 newi
->inode
.rstat
.rsubdirs
= 1;
5092 newi
->inode
.update_backtrace();
5094 dout(12) << " follows " << follows
<< dendl
;
5095 assert(dn
->first
== follows
+ 1);
5096 newi
->first
= dn
->first
;
5098 // ...and that new dir is empty.
5099 CDir
*newdir
= newi
->get_or_open_dirfrag(mdcache
, frag_t());
5100 newdir
->state_set(CDir::STATE_CREATING
);
5101 newdir
->mark_complete();
5102 newdir
->fnode
.version
= newdir
->pre_dirty();
5105 mdr
->ls
= mdlog
->get_current_segment();
5106 EUpdate
*le
= new EUpdate(mdlog
, "mkdir");
5107 mdlog
->start_entry(le
);
5108 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5109 journal_allocated_inos(mdr
, &le
->metablob
);
5110 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5111 le
->metablob
.add_primary_dentry(dn
, newi
, true, true);
5112 le
->metablob
.add_new_dir(newdir
); // dirty AND complete AND new
5114 // issue a cap on the directory
5115 int cmode
= CEPH_FILE_MODE_RDWR
;
5116 Capability
*cap
= mds
->locker
->issue_new_caps(newi
, cmode
, mdr
->session
, realm
, req
->is_replay());
5120 // put locks in excl mode
5121 newi
->filelock
.set_state(LOCK_EXCL
);
5122 newi
->authlock
.set_state(LOCK_EXCL
);
5123 newi
->xattrlock
.set_state(LOCK_EXCL
);
5126 // make sure this inode gets into the journal
5127 le
->metablob
.add_opened_ino(newi
->ino());
5128 LogSegment
*ls
= mds
->mdlog
->get_current_segment();
5129 ls
->open_files
.push_back(&newi
->item_open_file
);
5131 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5137 void Server::handle_client_symlink(MDRequestRef
& mdr
)
5139 MClientRequest
*req
= mdr
->client_request
;
5140 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5141 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5143 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5144 respond_to_request(mdr
, -EROFS
);
5147 CDir
*dir
= dn
->get_dir();
5148 CInode
*diri
= dir
->get_inode();
5149 rdlocks
.insert(&diri
->authlock
);
5150 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5153 if (!check_access(mdr
, diri
, MAY_WRITE
))
5156 if (!check_fragment_space(mdr
, dir
))
5159 unsigned mode
= S_IFLNK
| 0777;
5160 CInode
*newi
= prepare_new_inode(mdr
, dn
->get_dir(), inodeno_t(req
->head
.ino
), mode
);
5164 dn
->push_projected_linkage(newi
);
5166 newi
->symlink
= req
->get_path2();
5167 newi
->inode
.size
= newi
->symlink
.length();
5168 newi
->inode
.rstat
.rbytes
= newi
->inode
.size
;
5169 newi
->inode
.rstat
.rfiles
= 1;
5170 newi
->inode
.version
= dn
->pre_dirty();
5171 newi
->inode
.update_backtrace();
5173 newi
->first
= dn
->first
;
5176 mdr
->ls
= mdlog
->get_current_segment();
5177 EUpdate
*le
= new EUpdate(mdlog
, "symlink");
5178 mdlog
->start_entry(le
);
5179 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
5180 journal_allocated_inos(mdr
, &le
->metablob
);
5181 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, newi
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5182 le
->metablob
.add_primary_dentry(dn
, newi
, true, true);
5184 journal_and_reply(mdr
, newi
, dn
, le
, new C_MDS_mknod_finish(this, mdr
, dn
, newi
));
5193 void Server::handle_client_link(MDRequestRef
& mdr
)
5195 MClientRequest
*req
= mdr
->client_request
;
5197 dout(7) << "handle_client_link " << req
->get_filepath()
5198 << " to " << req
->get_filepath2()
5201 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5203 CDentry
*dn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, false, false, false);
5205 CInode
*targeti
= rdlock_path_pin_ref(mdr
, 1, rdlocks
, false);
5206 if (!targeti
) return;
5207 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5208 respond_to_request(mdr
, -EROFS
);
5212 CDir
*dir
= dn
->get_dir();
5213 dout(7) << "handle_client_link link " << dn
->get_name() << " in " << *dir
<< dendl
;
5214 dout(7) << "target is " << *targeti
<< dendl
;
5215 if (targeti
->is_dir()) {
5216 dout(7) << "target is a dir, failing..." << dendl
;
5217 respond_to_request(mdr
, -EINVAL
);
5221 xlocks
.insert(&targeti
->linklock
);
5223 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5226 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
5227 if (!check_access(mdr
, targeti
, MAY_WRITE
))
5230 if (!check_access(mdr
, dir
->get_inode(), MAY_WRITE
))
5233 if (!check_fragment_space(mdr
, dir
))
5238 assert(g_conf
->mds_kill_link_at
!= 1);
5241 if (targeti
->is_auth())
5242 _link_local(mdr
, dn
, targeti
);
5244 _link_remote(mdr
, true, dn
, targeti
);
5248 class C_MDS_link_local_finish
: public ServerLogContext
{
5254 C_MDS_link_local_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CInode
*ti
,
5255 version_t dnpv_
, version_t tipv_
) :
5256 ServerLogContext(s
, r
), dn(d
), targeti(ti
),
5257 dnpv(dnpv_
), tipv(tipv_
) { }
5258 void finish(int r
) override
{
5260 server
->_link_local_finish(mdr
, dn
, targeti
, dnpv
, tipv
);
5265 void Server::_link_local(MDRequestRef
& mdr
, CDentry
*dn
, CInode
*targeti
)
5267 dout(10) << "_link_local " << *dn
<< " to " << *targeti
<< dendl
;
5269 mdr
->ls
= mdlog
->get_current_segment();
5271 // predirty NEW dentry
5272 version_t dnpv
= dn
->pre_dirty();
5273 version_t tipv
= targeti
->pre_dirty();
5275 // project inode update
5276 inode_t
*pi
= targeti
->project_inode();
5278 pi
->ctime
= mdr
->get_op_stamp();
5283 EUpdate
*le
= new EUpdate(mdlog
, "link_local");
5284 mdlog
->start_entry(le
);
5285 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5286 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, 1); // new dn
5287 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, 0, PREDIRTY_PRIMARY
); // targeti
5288 le
->metablob
.add_remote_dentry(dn
, true, targeti
->ino(), targeti
->d_type()); // new remote
5289 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, targeti
);
5291 // do this after predirty_*, to avoid funky extra dnl arg
5292 dn
->push_projected_linkage(targeti
->ino(), targeti
->d_type());
5294 journal_and_reply(mdr
, targeti
, dn
, le
, new C_MDS_link_local_finish(this, mdr
, dn
, targeti
, dnpv
, tipv
));
5297 void Server::_link_local_finish(MDRequestRef
& mdr
, CDentry
*dn
, CInode
*targeti
,
5298 version_t dnpv
, version_t tipv
)
5300 dout(10) << "_link_local_finish " << *dn
<< " to " << *targeti
<< dendl
;
5302 // link and unlock the NEW dentry
5303 CDentry::linkage_t
*dnl
= dn
->pop_projected_linkage();
5304 if (!dnl
->get_inode())
5305 dn
->link_remote(dnl
, targeti
);
5306 dn
->mark_dirty(dnpv
, mdr
->ls
);
5309 targeti
->pop_and_dirty_projected_inode(mdr
->ls
);
5313 MDRequestRef null_ref
;
5314 mdcache
->send_dentry_link(dn
, null_ref
);
5316 // bump target popularity
5317 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5318 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
5321 respond_to_request(mdr
, 0);
5325 // link / unlink remote
5327 class C_MDS_link_remote_finish
: public ServerLogContext
{
5333 C_MDS_link_remote_finish(Server
*s
, MDRequestRef
& r
, bool i
, CDentry
*d
, CInode
*ti
) :
5334 ServerLogContext(s
, r
), inc(i
), dn(d
), targeti(ti
),
5335 dpv(d
->get_projected_version()) {}
5336 void finish(int r
) override
{
5338 server
->_link_remote_finish(mdr
, inc
, dn
, targeti
, dpv
);
5342 void Server::_link_remote(MDRequestRef
& mdr
, bool inc
, CDentry
*dn
, CInode
*targeti
)
5344 dout(10) << "_link_remote "
5345 << (inc
? "link ":"unlink ")
5346 << *dn
<< " to " << *targeti
<< dendl
;
5348 // 1. send LinkPrepare to dest (journal nlink++ prepare)
5349 mds_rank_t linkauth
= targeti
->authority().first
;
5350 if (mdr
->more()->witnessed
.count(linkauth
) == 0) {
5351 if (mds
->is_cluster_degraded() &&
5352 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(linkauth
)) {
5353 dout(10) << " targeti auth mds." << linkauth
<< " is not active" << dendl
;
5354 if (mdr
->more()->waiting_on_slave
.empty())
5355 mds
->wait_for_active_peer(linkauth
, new C_MDS_RetryRequest(mdcache
, mdr
));
5359 dout(10) << " targeti auth must prepare nlink++/--" << dendl
;
5362 op
= MMDSSlaveRequest::OP_LINKPREP
;
5364 op
= MMDSSlaveRequest::OP_UNLINKPREP
;
5365 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, op
);
5366 targeti
->set_object_info(req
->get_object_info());
5367 req
->op_stamp
= mdr
->get_op_stamp();
5368 mds
->send_message_mds(req
, linkauth
);
5370 assert(mdr
->more()->waiting_on_slave
.count(linkauth
) == 0);
5371 mdr
->more()->waiting_on_slave
.insert(linkauth
);
5374 dout(10) << " targeti auth has prepared nlink++/--" << dendl
;
5376 assert(g_conf
->mds_kill_link_at
!= 2);
5378 mdr
->set_mds_stamp(ceph_clock_now());
5381 mdr
->ls
= mdlog
->get_current_segment();
5382 EUpdate
*le
= new EUpdate(mdlog
, inc
? "link_remote":"unlink_remote");
5383 mdlog
->start_entry(le
);
5384 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5385 if (!mdr
->more()->witnessed
.empty()) {
5386 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
5387 le
->reqid
= mdr
->reqid
;
5388 le
->had_slaves
= true;
5389 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
5394 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, 1);
5395 le
->metablob
.add_remote_dentry(dn
, true, targeti
->ino(), targeti
->d_type()); // new remote
5396 dn
->push_projected_linkage(targeti
->ino(), targeti
->d_type());
5399 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, targeti
, dn
->get_dir(), PREDIRTY_DIR
, -1);
5400 mdcache
->journal_cow_dentry(mdr
.get(), &le
->metablob
, dn
);
5401 le
->metablob
.add_null_dentry(dn
, true);
5402 dn
->push_projected_linkage();
5405 journal_and_reply(mdr
, targeti
, dn
, le
, new C_MDS_link_remote_finish(this, mdr
, inc
, dn
, targeti
));
5408 void Server::_link_remote_finish(MDRequestRef
& mdr
, bool inc
,
5409 CDentry
*dn
, CInode
*targeti
,
5412 dout(10) << "_link_remote_finish "
5413 << (inc
? "link ":"unlink ")
5414 << *dn
<< " to " << *targeti
<< dendl
;
5416 assert(g_conf
->mds_kill_link_at
!= 3);
5418 if (!mdr
->more()->witnessed
.empty())
5419 mdcache
->logged_master_update(mdr
->reqid
);
5422 // link the new dentry
5423 CDentry::linkage_t
*dnl
= dn
->pop_projected_linkage();
5424 if (!dnl
->get_inode())
5425 dn
->link_remote(dnl
, targeti
);
5426 dn
->mark_dirty(dpv
, mdr
->ls
);
5428 // unlink main dentry
5429 dn
->get_dir()->unlink_inode(dn
);
5430 dn
->pop_projected_linkage();
5431 dn
->mark_dirty(dn
->get_projected_version(), mdr
->ls
); // dirty old dentry
5436 MDRequestRef null_ref
;
5438 mdcache
->send_dentry_link(dn
, null_ref
);
5440 mdcache
->send_dentry_unlink(dn
, NULL
, null_ref
);
5442 // bump target popularity
5443 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5444 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
5447 respond_to_request(mdr
, 0);
5450 // removing a new dn?
5451 dn
->get_dir()->try_remove_unlinked_dn(dn
);
5455 // remote linking/unlinking
5457 class C_MDS_SlaveLinkPrep
: public ServerLogContext
{
5460 C_MDS_SlaveLinkPrep(Server
*s
, MDRequestRef
& r
, CInode
*t
) :
5461 ServerLogContext(s
, r
), targeti(t
) { }
5462 void finish(int r
) override
{
5464 server
->_logged_slave_link(mdr
, targeti
);
5468 class C_MDS_SlaveLinkCommit
: public ServerContext
{
5472 C_MDS_SlaveLinkCommit(Server
*s
, MDRequestRef
& r
, CInode
*t
) :
5473 ServerContext(s
), mdr(r
), targeti(t
) { }
5474 void finish(int r
) override
{
5475 server
->_commit_slave_link(mdr
, r
, targeti
);
5479 /* This function DOES put the mdr->slave_request before returning*/
5480 void Server::handle_slave_link_prep(MDRequestRef
& mdr
)
5482 dout(10) << "handle_slave_link_prep " << *mdr
5483 << " on " << mdr
->slave_request
->get_object_info()
5486 assert(g_conf
->mds_kill_link_at
!= 4);
5488 CInode
*targeti
= mdcache
->get_inode(mdr
->slave_request
->get_object_info().ino
);
5490 dout(10) << "targeti " << *targeti
<< dendl
;
5491 CDentry
*dn
= targeti
->get_parent_dn();
5492 CDentry::linkage_t
*dnl
= dn
->get_linkage();
5493 assert(dnl
->is_primary());
5495 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
5497 mdr
->auth_pin(targeti
);
5499 //ceph_abort(); // test hack: make sure master can handle a slave that fails to prepare...
5500 assert(g_conf
->mds_kill_link_at
!= 5);
5503 mdr
->ls
= mdlog
->get_current_segment();
5504 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_prep", mdr
->reqid
, mdr
->slave_to_mds
,
5505 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::LINK
);
5506 mdlog
->start_entry(le
);
5508 inode_t
*pi
= dnl
->get_inode()->project_inode();
5510 // update journaled target inode
5512 if (mdr
->slave_request
->get_op() == MMDSSlaveRequest::OP_LINKPREP
) {
5520 link_rollback rollback
;
5521 rollback
.reqid
= mdr
->reqid
;
5522 rollback
.ino
= targeti
->ino();
5523 rollback
.old_ctime
= targeti
->inode
.ctime
; // we hold versionlock xlock; no concorrent projections
5524 const fnode_t
*pf
= targeti
->get_parent_dn()->get_dir()->get_projected_fnode();
5525 rollback
.old_dir_mtime
= pf
->fragstat
.mtime
;
5526 rollback
.old_dir_rctime
= pf
->rstat
.rctime
;
5527 rollback
.was_inc
= inc
;
5528 ::encode(rollback
, le
->rollback
);
5529 mdr
->more()->rollback_bl
= le
->rollback
;
5531 pi
->ctime
= mdr
->get_op_stamp();
5532 pi
->version
= targeti
->pre_dirty();
5534 dout(10) << " projected inode " << pi
<< " v " << pi
->version
<< dendl
;
5537 mdcache
->predirty_journal_parents(mdr
, &le
->commit
, dnl
->get_inode(), 0, PREDIRTY_SHALLOW
|PREDIRTY_PRIMARY
);
5538 mdcache
->journal_dirty_inode(mdr
.get(), &le
->commit
, targeti
);
5540 // set up commit waiter
5541 mdr
->more()->slave_commit
= new C_MDS_SlaveLinkCommit(this, mdr
, targeti
);
5543 mdr
->more()->slave_update_journaled
= true;
5544 submit_mdlog_entry(le
, new C_MDS_SlaveLinkPrep(this, mdr
, targeti
),
5549 void Server::_logged_slave_link(MDRequestRef
& mdr
, CInode
*targeti
)
5551 dout(10) << "_logged_slave_link " << *mdr
5552 << " " << *targeti
<< dendl
;
5554 assert(g_conf
->mds_kill_link_at
!= 6);
5556 // update the target
5557 targeti
->pop_and_dirty_projected_inode(mdr
->ls
);
5561 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), targeti
, META_POP_IWR
);
5564 mdr
->slave_request
->put();
5565 mdr
->slave_request
= 0;
5568 if (!mdr
->aborted
) {
5569 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
5570 MMDSSlaveRequest::OP_LINKPREPACK
);
5571 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
5573 dout(10) << " abort flag set, finishing" << dendl
;
5574 mdcache
->request_finish(mdr
);
5579 struct C_MDS_CommittedSlave
: public ServerLogContext
{
5580 C_MDS_CommittedSlave(Server
*s
, MDRequestRef
& m
) : ServerLogContext(s
, m
) {}
5581 void finish(int r
) override
{
5582 server
->_committed_slave(mdr
);
5586 void Server::_commit_slave_link(MDRequestRef
& mdr
, int r
, CInode
*targeti
)
5588 dout(10) << "_commit_slave_link " << *mdr
5590 << " " << *targeti
<< dendl
;
5592 assert(g_conf
->mds_kill_link_at
!= 7);
5595 // drop our pins, etc.
5598 // write a commit to the journal
5599 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_commit", mdr
->reqid
, mdr
->slave_to_mds
,
5600 ESlaveUpdate::OP_COMMIT
, ESlaveUpdate::LINK
);
5601 mdlog
->start_entry(le
);
5602 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
5605 do_link_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
);
5609 void Server::_committed_slave(MDRequestRef
& mdr
)
5611 dout(10) << "_committed_slave " << *mdr
<< dendl
;
5613 assert(g_conf
->mds_kill_link_at
!= 8);
5615 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
5616 MMDSSlaveRequest::OP_COMMITTED
);
5617 mds
->send_message_mds(req
, mdr
->slave_to_mds
);
5618 mdcache
->request_finish(mdr
);
5621 struct C_MDS_LoggedLinkRollback
: public ServerLogContext
{
5623 C_MDS_LoggedLinkRollback(Server
*s
, MutationRef
& m
, MDRequestRef
& r
) : ServerLogContext(s
, r
), mut(m
) {}
5624 void finish(int r
) override
{
5625 server
->_link_rollback_finish(mut
, mdr
);
5629 void Server::do_link_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
)
5631 link_rollback rollback
;
5632 bufferlist::iterator p
= rbl
.begin();
5633 ::decode(rollback
, p
);
5635 dout(10) << "do_link_rollback on " << rollback
.reqid
5636 << (rollback
.was_inc
? " inc":" dec")
5637 << " ino " << rollback
.ino
5640 assert(g_conf
->mds_kill_link_at
!= 9);
5642 mdcache
->add_rollback(rollback
.reqid
, master
); // need to finish this update before resolve finishes
5643 assert(mdr
|| mds
->is_resolve());
5645 MutationRef
mut(new MutationImpl(nullptr, utime_t(), rollback
.reqid
));
5646 mut
->ls
= mds
->mdlog
->get_current_segment();
5648 CInode
*in
= mdcache
->get_inode(rollback
.ino
);
5650 dout(10) << " target is " << *in
<< dendl
;
5651 assert(!in
->is_projected()); // live slave request hold versionlock xlock.
5653 inode_t
*pi
= in
->project_inode();
5654 pi
->version
= in
->pre_dirty();
5655 mut
->add_projected_inode(in
);
5657 // parent dir rctime
5658 CDir
*parent
= in
->get_projected_parent_dn()->get_dir();
5659 fnode_t
*pf
= parent
->project_fnode();
5660 mut
->add_projected_fnode(parent
);
5661 pf
->version
= parent
->pre_dirty();
5662 if (pf
->fragstat
.mtime
== pi
->ctime
) {
5663 pf
->fragstat
.mtime
= rollback
.old_dir_mtime
;
5664 if (pf
->rstat
.rctime
== pi
->ctime
)
5665 pf
->rstat
.rctime
= rollback
.old_dir_rctime
;
5666 mut
->add_updated_lock(&parent
->get_inode()->filelock
);
5667 mut
->add_updated_lock(&parent
->get_inode()->nestlock
);
5671 pi
->ctime
= rollback
.old_ctime
;
5672 if (rollback
.was_inc
)
5678 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_link_rollback", rollback
.reqid
, master
,
5679 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::LINK
);
5680 mdlog
->start_entry(le
);
5681 le
->commit
.add_dir_context(parent
);
5682 le
->commit
.add_dir(parent
, true);
5683 le
->commit
.add_primary_dentry(in
->get_projected_parent_dn(), 0, true);
5685 submit_mdlog_entry(le
, new C_MDS_LoggedLinkRollback(this, mut
, mdr
),
5690 void Server::_link_rollback_finish(MutationRef
& mut
, MDRequestRef
& mdr
)
5692 dout(10) << "_link_rollback_finish" << dendl
;
5694 assert(g_conf
->mds_kill_link_at
!= 10);
5698 mdcache
->request_finish(mdr
);
5700 mdcache
->finish_rollback(mut
->reqid
);
5706 /* This function DOES NOT put the passed message before returning*/
5707 void Server::handle_slave_link_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*m
)
5709 dout(10) << "handle_slave_link_prep_ack " << *mdr
5710 << " " << *m
<< dendl
;
5711 mds_rank_t from
= mds_rank_t(m
->get_source().num());
5713 assert(g_conf
->mds_kill_link_at
!= 11);
5716 mdr
->more()->slaves
.insert(from
);
5719 assert(mdr
->more()->witnessed
.count(from
) == 0);
5720 mdr
->more()->witnessed
.insert(from
);
5721 assert(!m
->is_not_journaled());
5722 mdr
->more()->has_journaled_slaves
= true;
5724 // remove from waiting list
5725 assert(mdr
->more()->waiting_on_slave
.count(from
));
5726 mdr
->more()->waiting_on_slave
.erase(from
);
5728 assert(mdr
->more()->waiting_on_slave
.empty());
5730 dispatch_client_request(mdr
); // go again!
5739 void Server::handle_client_unlink(MDRequestRef
& mdr
)
5741 MClientRequest
*req
= mdr
->client_request
;
5742 client_t client
= mdr
->get_client();
5746 if (req
->get_op() == CEPH_MDS_OP_RMDIR
) rmdir
= true;
5748 if (req
->get_filepath().depth() == 0) {
5749 respond_to_request(mdr
, -EINVAL
);
5754 vector
<CDentry
*> trace
;
5756 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, req
->get_filepath(), &trace
, &in
, MDS_TRAVERSE_FORWARD
);
5760 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
5761 mdcache
->find_ino_peers(req
->get_filepath().get_ino(), new C_MDS_TryFindInode(this, mdr
));
5764 respond_to_request(mdr
, r
);
5767 if (mdr
->snapid
!= CEPH_NOSNAP
) {
5768 respond_to_request(mdr
, -EROFS
);
5772 CDentry
*dn
= trace
[trace
.size()-1];
5774 if (!dn
->is_auth()) {
5775 mdcache
->request_forward(mdr
, dn
->authority().first
);
5779 CInode
*diri
= dn
->get_dir()->get_inode();
5781 CDentry::linkage_t
*dnl
= dn
->get_linkage(client
, mdr
);
5782 assert(!dnl
->is_null());
5785 dout(7) << "handle_client_rmdir on " << *dn
<< dendl
;
5787 dout(7) << "handle_client_unlink on " << *dn
<< dendl
;
5789 dout(7) << "dn links to " << *in
<< dendl
;
5794 // do empty directory checks
5795 if (_dir_is_nonempty_unlocked(mdr
, in
)) {
5796 respond_to_request(mdr
, -ENOTEMPTY
);
5800 dout(7) << "handle_client_unlink on dir " << *in
<< ", returning error" << dendl
;
5801 respond_to_request(mdr
, -EISDIR
);
5807 dout(7) << "handle_client_rmdir on non-dir " << *in
<< ", returning error" << dendl
;
5808 respond_to_request(mdr
, -ENOTDIR
);
5813 // -- create stray dentry? --
5814 CDentry
*straydn
= NULL
;
5815 if (dnl
->is_primary()) {
5816 straydn
= prepare_stray_dentry(mdr
, dnl
->get_inode());
5819 dout(10) << " straydn is " << *straydn
<< dendl
;
5820 } else if (mdr
->straydn
) {
5821 mdr
->unpin(mdr
->straydn
);
5822 mdr
->straydn
= NULL
;
5826 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
5828 for (int i
=0; i
<(int)trace
.size()-1; i
++)
5829 rdlocks
.insert(&trace
[i
]->lock
);
5830 xlocks
.insert(&dn
->lock
);
5831 wrlocks
.insert(&diri
->filelock
);
5832 wrlocks
.insert(&diri
->nestlock
);
5833 xlocks
.insert(&in
->linklock
);
5835 wrlocks
.insert(&straydn
->get_dir()->inode
->filelock
);
5836 wrlocks
.insert(&straydn
->get_dir()->inode
->nestlock
);
5837 xlocks
.insert(&straydn
->lock
);
5840 rdlocks
.insert(&in
->filelock
); // to verify it's empty
5841 mds
->locker
->include_snap_rdlocks(rdlocks
, dnl
->get_inode());
5843 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
5847 _dir_is_nonempty(mdr
, in
)) {
5848 respond_to_request(mdr
, -ENOTEMPTY
);
5852 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
5853 if (!check_access(mdr
, diri
, MAY_WRITE
))
5858 if (in
->is_dir() && in
->has_subtree_root_dirfrag()) {
5859 // subtree root auths need to be witnesses
5860 set
<mds_rank_t
> witnesses
;
5861 in
->list_replicas(witnesses
);
5862 dout(10) << " witnesses " << witnesses
<< ", have " << mdr
->more()->witnessed
<< dendl
;
5864 for (set
<mds_rank_t
>::iterator p
= witnesses
.begin();
5865 p
!= witnesses
.end();
5867 if (mdr
->more()->witnessed
.count(*p
)) {
5868 dout(10) << " already witnessed by mds." << *p
<< dendl
;
5869 } else if (mdr
->more()->waiting_on_slave
.count(*p
)) {
5870 dout(10) << " already waiting on witness mds." << *p
<< dendl
;
5872 if (!_rmdir_prepare_witness(mdr
, *p
, trace
, straydn
))
5876 if (!mdr
->more()->waiting_on_slave
.empty())
5877 return; // we're waiting for a witness.
5881 if (dnl
->is_remote() && !dnl
->get_inode()->is_auth())
5882 _link_remote(mdr
, false, dn
, dnl
->get_inode());
5884 _unlink_local(mdr
, dn
, straydn
);
5887 class C_MDS_unlink_local_finish
: public ServerLogContext
{
5890 version_t dnpv
; // deleted dentry
5892 C_MDS_unlink_local_finish(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CDentry
*sd
) :
5893 ServerLogContext(s
, r
), dn(d
), straydn(sd
),
5894 dnpv(d
->get_projected_version()) {}
5895 void finish(int r
) override
{
5897 server
->_unlink_local_finish(mdr
, dn
, straydn
, dnpv
);
5901 void Server::_unlink_local(MDRequestRef
& mdr
, CDentry
*dn
, CDentry
*straydn
)
5903 dout(10) << "_unlink_local " << *dn
<< dendl
;
5905 CDentry::linkage_t
*dnl
= dn
->get_projected_linkage();
5906 CInode
*in
= dnl
->get_inode();
5908 SnapRealm
*realm
= in
->find_snaprealm();
5909 snapid_t follows
= realm
->get_newest_seq();
5912 mdr
->ls
= mdlog
->get_current_segment();
5914 // prepare log entry
5915 EUpdate
*le
= new EUpdate(mdlog
, "unlink_local");
5916 mdlog
->start_entry(le
);
5917 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
5918 if (!mdr
->more()->witnessed
.empty()) {
5919 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
5920 le
->reqid
= mdr
->reqid
;
5921 le
->had_slaves
= true;
5922 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
5926 assert(dnl
->is_primary());
5927 straydn
->push_projected_linkage(in
);
5928 straydn
->first
= follows
+ 1;
5931 // the unlinked dentry
5934 inode_t
*pi
= in
->project_inode();
5935 dn
->make_path_string(pi
->stray_prior_path
, true);
5936 mdr
->add_projected_inode(in
); // do this _after_ my dn->pre_dirty().. we apply that one manually.
5937 pi
->version
= in
->pre_dirty();
5938 pi
->ctime
= mdr
->get_op_stamp();
5942 in
->state_set(CInode::STATE_ORPHAN
);
5944 if (dnl
->is_primary()) {
5945 // primary link. add stray dentry.
5947 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, -1);
5948 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, straydn
->get_dir(), PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
5950 // project snaprealm, too
5951 if (in
->snaprealm
|| follows
+ 1 > in
->get_oldest_snap())
5952 in
->project_past_snaprealm_parent(straydn
->get_dir()->inode
->find_snaprealm());
5954 pi
->update_backtrace();
5955 le
->metablob
.add_primary_dentry(straydn
, in
, true, true);
5957 // remote link. update remote inode.
5958 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, dn
->get_dir(), PREDIRTY_DIR
, -1);
5959 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, in
, 0, PREDIRTY_PRIMARY
);
5960 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, in
);
5963 mdcache
->journal_cow_dentry(mdr
.get(), &le
->metablob
, dn
);
5964 le
->metablob
.add_null_dentry(dn
, true);
5967 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
5968 le
->metablob
.renamed_dirino
= in
->ino();
5971 dn
->push_projected_linkage();
5975 mdcache
->project_subtree_rename(in
, dn
->get_dir(), straydn
->get_dir());
5978 journal_and_reply(mdr
, 0, dn
, le
, new C_MDS_unlink_local_finish(this, mdr
, dn
, straydn
));
5981 void Server::_unlink_local_finish(MDRequestRef
& mdr
,
5982 CDentry
*dn
, CDentry
*straydn
,
5985 dout(10) << "_unlink_local_finish " << *dn
<< dendl
;
5987 if (!mdr
->more()->witnessed
.empty())
5988 mdcache
->logged_master_update(mdr
->reqid
);
5990 // unlink main dentry
5991 dn
->get_dir()->unlink_inode(dn
);
5992 dn
->pop_projected_linkage();
5994 // relink as stray? (i.e. was primary link?)
5995 CInode
*strayin
= NULL
;
5996 bool snap_is_new
= false;
5998 dout(20) << " straydn is " << *straydn
<< dendl
;
5999 CDentry::linkage_t
*straydnl
= straydn
->pop_projected_linkage();
6000 strayin
= straydnl
->get_inode();
6002 snap_is_new
= strayin
->snaprealm
? true : false;
6003 mdcache
->touch_dentry_bottom(straydn
);
6006 dn
->mark_dirty(dnpv
, mdr
->ls
);
6009 if (snap_is_new
) //only new if strayin exists
6010 mdcache
->do_realm_invalidate_and_update_notify(strayin
, CEPH_SNAP_OP_SPLIT
, true);
6012 mdcache
->send_dentry_unlink(dn
, straydn
, mdr
);
6014 // update subtree map?
6015 if (straydn
&& strayin
->is_dir())
6016 mdcache
->adjust_subtree_after_rename(strayin
, dn
->get_dir(), true);
6019 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), dn
->get_dir(), META_POP_IWR
);
6022 respond_to_request(mdr
, 0);
6024 // removing a new dn?
6025 dn
->get_dir()->try_remove_unlinked_dn(dn
);
6028 // respond_to_request() drops locks. So stray reintegration can race with us.
6029 if (straydn
&& !straydn
->get_projected_linkage()->is_null()) {
6030 // Tip off the MDCache that this dentry is a stray that
6031 // might be elegible for purge.
6032 mdcache
->notify_stray(straydn
);
6036 bool Server::_rmdir_prepare_witness(MDRequestRef
& mdr
, mds_rank_t who
, vector
<CDentry
*>& trace
, CDentry
*straydn
)
6038 if (mds
->is_cluster_degraded() &&
6039 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(who
)) {
6040 dout(10) << "_rmdir_prepare_witness mds." << who
<< " is not active" << dendl
;
6041 if (mdr
->more()->waiting_on_slave
.empty())
6042 mds
->wait_for_active_peer(who
, new C_MDS_RetryRequest(mdcache
, mdr
));
6046 dout(10) << "_rmdir_prepare_witness mds." << who
<< dendl
;
6047 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6048 MMDSSlaveRequest::OP_RMDIRPREP
);
6049 req
->srcdnpath
= filepath(trace
.front()->get_dir()->ino());
6050 for (auto dn
: trace
)
6051 req
->srcdnpath
.push_dentry(dn
->name
);
6052 mdcache
->replicate_stray(straydn
, who
, req
->stray
);
6054 req
->op_stamp
= mdr
->get_op_stamp();
6055 mds
->send_message_mds(req
, who
);
6057 assert(mdr
->more()->waiting_on_slave
.count(who
) == 0);
6058 mdr
->more()->waiting_on_slave
.insert(who
);
6062 struct C_MDS_SlaveRmdirPrep
: public ServerLogContext
{
6063 CDentry
*dn
, *straydn
;
6064 C_MDS_SlaveRmdirPrep(Server
*s
, MDRequestRef
& r
, CDentry
*d
, CDentry
*st
)
6065 : ServerLogContext(s
, r
), dn(d
), straydn(st
) {}
6066 void finish(int r
) override
{
6067 server
->_logged_slave_rmdir(mdr
, dn
, straydn
);
6071 struct C_MDS_SlaveRmdirCommit
: public ServerContext
{
6074 C_MDS_SlaveRmdirCommit(Server
*s
, MDRequestRef
& r
, CDentry
*sd
)
6075 : ServerContext(s
), mdr(r
), straydn(sd
) { }
6076 void finish(int r
) override
{
6077 server
->_commit_slave_rmdir(mdr
, r
, straydn
);
6081 void Server::handle_slave_rmdir_prep(MDRequestRef
& mdr
)
6083 dout(10) << "handle_slave_rmdir_prep " << *mdr
6084 << " " << mdr
->slave_request
->srcdnpath
6085 << " to " << mdr
->slave_request
->destdnpath
6088 vector
<CDentry
*> trace
;
6089 filepath
srcpath(mdr
->slave_request
->srcdnpath
);
6090 dout(10) << " src " << srcpath
<< dendl
;
6092 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &trace
, &in
, MDS_TRAVERSE_DISCOVERXLOCK
);
6095 mdcache
->find_ino_peers(srcpath
.get_ino(), new C_MDS_RetryRequest(mdcache
, mdr
),
6100 CDentry
*dn
= trace
[trace
.size()-1];
6101 dout(10) << " dn " << *dn
<< dendl
;
6104 assert(mdr
->straydn
);
6105 CDentry
*straydn
= mdr
->straydn
;
6106 dout(10) << " straydn " << *straydn
<< dendl
;
6108 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
6110 rmdir_rollback rollback
;
6111 rollback
.reqid
= mdr
->reqid
;
6112 rollback
.src_dir
= dn
->get_dir()->dirfrag();
6113 rollback
.src_dname
= dn
->name
;
6114 rollback
.dest_dir
= straydn
->get_dir()->dirfrag();
6115 rollback
.dest_dname
= straydn
->name
;
6116 ::encode(rollback
, mdr
->more()->rollback_bl
);
6117 dout(20) << " rollback is " << mdr
->more()->rollback_bl
.length() << " bytes" << dendl
;
6119 // set up commit waiter
6120 mdr
->more()->slave_commit
= new C_MDS_SlaveRmdirCommit(this, mdr
, straydn
);
6122 if (!in
->has_subtree_root_dirfrag(mds
->get_nodeid())) {
6123 dout(10) << " no auth subtree in " << *in
<< ", skipping journal" << dendl
;
6124 dn
->get_dir()->unlink_inode(dn
);
6125 straydn
->get_dir()->link_primary_inode(straydn
, in
);
6127 assert(straydn
->first
>= in
->first
);
6128 in
->first
= straydn
->first
;
6130 mdcache
->adjust_subtree_after_rename(in
, dn
->get_dir(), false);
6132 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6133 MMDSSlaveRequest::OP_RMDIRPREPACK
);
6134 reply
->mark_not_journaled();
6135 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
6137 // send caps to auth (if we're not already)
6138 if (in
->is_any_caps() && !in
->state_test(CInode::STATE_EXPORTINGCAPS
))
6139 mdcache
->migrator
->export_caps(in
);
6141 mdcache
->touch_dentry_bottom(straydn
); // move stray to end of lru
6143 mdr
->slave_request
->put();
6144 mdr
->slave_request
= 0;
6149 straydn
->push_projected_linkage(in
);
6150 dn
->push_projected_linkage();
6152 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir", mdr
->reqid
, mdr
->slave_to_mds
,
6153 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::RMDIR
);
6154 mdlog
->start_entry(le
);
6155 le
->rollback
= mdr
->more()->rollback_bl
;
6157 le
->commit
.add_dir_context(straydn
->get_dir());
6158 le
->commit
.add_primary_dentry(straydn
, in
, true);
6159 // slave: no need to journal original dentry
6161 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
6162 le
->commit
.renamed_dirino
= in
->ino();
6164 mdcache
->project_subtree_rename(in
, dn
->get_dir(), straydn
->get_dir());
6166 mdr
->more()->slave_update_journaled
= true;
6167 submit_mdlog_entry(le
, new C_MDS_SlaveRmdirPrep(this, mdr
, dn
, straydn
),
6172 void Server::_logged_slave_rmdir(MDRequestRef
& mdr
, CDentry
*dn
, CDentry
*straydn
)
6174 dout(10) << "_logged_slave_rmdir " << *mdr
<< " on " << *dn
<< dendl
;
6176 // update our cache now, so we are consistent with what is in the journal
6177 // when we journal a subtree map
6178 CInode
*in
= dn
->get_linkage()->get_inode();
6179 dn
->get_dir()->unlink_inode(dn
);
6180 straydn
->pop_projected_linkage();
6181 dn
->pop_projected_linkage();
6182 mdcache
->adjust_subtree_after_rename(in
, dn
->get_dir(), true);
6185 mdr
->slave_request
->put();
6186 mdr
->slave_request
= 0;
6189 if (!mdr
->aborted
) {
6190 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6191 MMDSSlaveRequest::OP_RMDIRPREPACK
);
6192 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
6194 dout(10) << " abort flag set, finishing" << dendl
;
6195 mdcache
->request_finish(mdr
);
6199 void Server::handle_slave_rmdir_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
6201 dout(10) << "handle_slave_rmdir_prep_ack " << *mdr
6202 << " " << *ack
<< dendl
;
6204 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
6206 mdr
->more()->slaves
.insert(from
);
6207 mdr
->more()->witnessed
.insert(from
);
6208 if (!ack
->is_not_journaled())
6209 mdr
->more()->has_journaled_slaves
= true;
6211 // remove from waiting list
6212 assert(mdr
->more()->waiting_on_slave
.count(from
));
6213 mdr
->more()->waiting_on_slave
.erase(from
);
6215 if (mdr
->more()->waiting_on_slave
.empty())
6216 dispatch_client_request(mdr
); // go again!
6218 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
6221 void Server::_commit_slave_rmdir(MDRequestRef
& mdr
, int r
, CDentry
*straydn
)
6223 dout(10) << "_commit_slave_rmdir " << *mdr
<< " r=" << r
<< dendl
;
6226 if (mdr
->more()->slave_update_journaled
) {
6227 CInode
*strayin
= straydn
->get_projected_linkage()->get_inode();
6228 if (strayin
&& !strayin
->snaprealm
)
6229 mdcache
->clear_dirty_bits_for_stray(strayin
);
6234 if (mdr
->more()->slave_update_journaled
) {
6235 // write a commit to the journal
6236 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir_commit", mdr
->reqid
,
6237 mdr
->slave_to_mds
, ESlaveUpdate::OP_COMMIT
,
6238 ESlaveUpdate::RMDIR
);
6239 mdlog
->start_entry(le
);
6240 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
6243 _committed_slave(mdr
);
6247 do_rmdir_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
);
6251 struct C_MDS_LoggedRmdirRollback
: public ServerLogContext
{
6255 C_MDS_LoggedRmdirRollback(Server
*s
, MDRequestRef
& m
, metareqid_t mr
, CDentry
*d
, CDentry
*st
)
6256 : ServerLogContext(s
, m
), reqid(mr
), dn(d
), straydn(st
) {}
6257 void finish(int r
) override
{
6258 server
->_rmdir_rollback_finish(mdr
, reqid
, dn
, straydn
);
6262 void Server::do_rmdir_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
)
6264 // unlink the other rollback methods, the rmdir rollback is only
6265 // needed to record the subtree changes in the journal for inode
6266 // replicas who are auth for empty dirfrags. no actual changes to
6267 // the file system are taking place here, so there is no Mutation.
6269 rmdir_rollback rollback
;
6270 bufferlist::iterator p
= rbl
.begin();
6271 ::decode(rollback
, p
);
6273 dout(10) << "do_rmdir_rollback on " << rollback
.reqid
<< dendl
;
6274 mdcache
->add_rollback(rollback
.reqid
, master
); // need to finish this update before resolve finishes
6275 assert(mdr
|| mds
->is_resolve());
6277 CDir
*dir
= mdcache
->get_dirfrag(rollback
.src_dir
);
6279 dir
= mdcache
->get_dirfrag(rollback
.src_dir
.ino
, rollback
.src_dname
);
6281 CDentry
*dn
= dir
->lookup(rollback
.src_dname
);
6283 dout(10) << " dn " << *dn
<< dendl
;
6284 dir
= mdcache
->get_dirfrag(rollback
.dest_dir
);
6286 CDentry
*straydn
= dir
->lookup(rollback
.dest_dname
);
6288 dout(10) << " straydn " << *dn
<< dendl
;
6289 CInode
*in
= straydn
->get_linkage()->get_inode();
6291 if (mdr
&& !mdr
->more()->slave_update_journaled
) {
6292 assert(!in
->has_subtree_root_dirfrag(mds
->get_nodeid()));
6294 straydn
->get_dir()->unlink_inode(straydn
);
6295 dn
->get_dir()->link_primary_inode(dn
, in
);
6297 mdcache
->adjust_subtree_after_rename(in
, straydn
->get_dir(), false);
6299 mdcache
->request_finish(mdr
);
6300 mdcache
->finish_rollback(rollback
.reqid
);
6304 dn
->push_projected_linkage(in
);
6305 straydn
->push_projected_linkage();
6307 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rmdir_rollback", rollback
.reqid
, master
,
6308 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::RMDIR
);
6309 mdlog
->start_entry(le
);
6311 le
->commit
.add_dir_context(dn
->get_dir());
6312 le
->commit
.add_primary_dentry(dn
, in
, true);
6313 // slave: no need to journal straydn
6315 dout(10) << " noting renamed (unlinked) dir ino " << in
->ino() << " in metablob" << dendl
;
6316 le
->commit
.renamed_dirino
= in
->ino();
6318 mdcache
->project_subtree_rename(in
, straydn
->get_dir(), dn
->get_dir());
6320 submit_mdlog_entry(le
,
6321 new C_MDS_LoggedRmdirRollback(this, mdr
,rollback
.reqid
,
6327 void Server::_rmdir_rollback_finish(MDRequestRef
& mdr
, metareqid_t reqid
, CDentry
*dn
, CDentry
*straydn
)
6329 dout(10) << "_rmdir_rollback_finish " << reqid
<< dendl
;
6331 straydn
->get_dir()->unlink_inode(straydn
);
6332 dn
->pop_projected_linkage();
6333 straydn
->pop_projected_linkage();
6335 CInode
*in
= dn
->get_linkage()->get_inode();
6336 mdcache
->adjust_subtree_after_rename(in
, straydn
->get_dir(), true);
6337 if (mds
->is_resolve()) {
6338 CDir
*root
= mdcache
->get_subtree_root(straydn
->get_dir());
6339 mdcache
->try_trim_non_auth_subtree(root
);
6343 mdcache
->request_finish(mdr
);
6345 mdcache
->finish_rollback(reqid
);
6349 /** _dir_is_nonempty[_unlocked]
6351 * check if a directory is non-empty (i.e. we can rmdir it).
6353 * the unlocked varient this is a fastpath check. we can't really be
6354 * sure until we rdlock the filelock.
6356 bool Server::_dir_is_nonempty_unlocked(MDRequestRef
& mdr
, CInode
*in
)
6358 dout(10) << "dir_is_nonempty_unlocked " << *in
<< dendl
;
6359 assert(in
->is_auth());
6361 if (in
->snaprealm
&& in
->snaprealm
->srnode
.snaps
.size())
6362 return true; // in a snapshot!
6365 in
->get_dirfrags(ls
);
6366 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6368 // is the frag obviously non-empty?
6369 if (dir
->is_auth()) {
6370 if (dir
->get_projected_fnode()->fragstat
.size()) {
6371 dout(10) << "dir_is_nonempty_unlocked dirstat has "
6372 << dir
->get_projected_fnode()->fragstat
.size() << " items " << *dir
<< dendl
;
6381 bool Server::_dir_is_nonempty(MDRequestRef
& mdr
, CInode
*in
)
6383 dout(10) << "dir_is_nonempty " << *in
<< dendl
;
6384 assert(in
->is_auth());
6385 assert(in
->filelock
.can_read(mdr
->get_client()));
6387 frag_info_t dirstat
;
6388 version_t dirstat_version
= in
->get_projected_inode()->dirstat
.version
;
6391 in
->get_dirfrags(ls
);
6392 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6394 const fnode_t
*pf
= dir
->get_projected_fnode();
6395 if (pf
->fragstat
.size()) {
6396 dout(10) << "dir_is_nonempty dirstat has "
6397 << pf
->fragstat
.size() << " items " << *dir
<< dendl
;
6401 if (pf
->accounted_fragstat
.version
== dirstat_version
)
6402 dirstat
.add(pf
->accounted_fragstat
);
6404 dirstat
.add(pf
->fragstat
);
6407 return dirstat
.size() != in
->get_projected_inode()->dirstat
.size();
6411 // ======================================================
6414 class C_MDS_rename_finish
: public ServerLogContext
{
6419 C_MDS_rename_finish(Server
*s
, MDRequestRef
& r
,
6420 CDentry
*sdn
, CDentry
*ddn
, CDentry
*stdn
) :
6421 ServerLogContext(s
, r
),
6422 srcdn(sdn
), destdn(ddn
), straydn(stdn
) { }
6423 void finish(int r
) override
{
6425 server
->_rename_finish(mdr
, srcdn
, destdn
, straydn
);
6430 /** handle_client_rename
6432 * rename master is the destdn auth. this is because cached inodes
6433 * must remain connected. thus, any replica of srci, must also
6434 * replicate destdn, and possibly straydn, so that srci (and
6435 * destdn->inode) remain connected during the rename.
6437 * to do this, we freeze srci, then master (destdn auth) verifies that
6438 * all other nodes have also replciated destdn and straydn. note that
6439 * destdn replicas need not also replicate srci. this only works when
6442 * This function takes responsibility for the passed mdr.
6444 void Server::handle_client_rename(MDRequestRef
& mdr
)
6446 MClientRequest
*req
= mdr
->client_request
;
6447 dout(7) << "handle_client_rename " << *req
<< dendl
;
6449 filepath destpath
= req
->get_filepath();
6450 filepath srcpath
= req
->get_filepath2();
6451 if (destpath
.depth() == 0 || srcpath
.depth() == 0) {
6452 respond_to_request(mdr
, -EINVAL
);
6455 const string
&destname
= destpath
.last_dentry();
6457 vector
<CDentry
*>& srctrace
= mdr
->dn
[1];
6458 vector
<CDentry
*>& desttrace
= mdr
->dn
[0];
6460 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
6462 CDentry
*destdn
= rdlock_path_xlock_dentry(mdr
, 0, rdlocks
, wrlocks
, xlocks
, true, false, true);
6463 if (!destdn
) return;
6464 dout(10) << " destdn " << *destdn
<< dendl
;
6465 if (mdr
->snapid
!= CEPH_NOSNAP
) {
6466 respond_to_request(mdr
, -EROFS
);
6469 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
6470 CDir
*destdir
= destdn
->get_dir();
6471 assert(destdir
->is_auth());
6473 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &srctrace
, NULL
, MDS_TRAVERSE_DISCOVER
);
6478 dout(10) << "FAIL on ESTALE but attempting recovery" << dendl
;
6479 mdcache
->find_ino_peers(srcpath
.get_ino(), new C_MDS_TryFindInode(this, mdr
));
6481 dout(10) << "FAIL on error " << r
<< dendl
;
6482 respond_to_request(mdr
, r
);
6487 assert(!srctrace
.empty());
6488 CDentry
*srcdn
= srctrace
[srctrace
.size()-1];
6489 dout(10) << " srcdn " << *srcdn
<< dendl
;
6490 if (srcdn
->last
!= CEPH_NOSNAP
) {
6491 respond_to_request(mdr
, -EROFS
);
6494 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
6495 CInode
*srci
= srcdnl
->get_inode();
6496 dout(10) << " srci " << *srci
<< dendl
;
6499 if (!destdnl
->is_null()) {
6500 //dout(10) << "dest dn exists " << *destdn << dendl;
6501 oldin
= mdcache
->get_dentry_inode(destdn
, mdr
, true);
6503 dout(10) << " oldin " << *oldin
<< dendl
;
6505 // mv /some/thing /to/some/existing_other_thing
6506 if (oldin
->is_dir() && !srci
->is_dir()) {
6507 respond_to_request(mdr
, -EISDIR
);
6510 if (!oldin
->is_dir() && srci
->is_dir()) {
6511 respond_to_request(mdr
, -ENOTDIR
);
6515 // non-empty dir? do trivial fast unlocked check, do another check later with read locks
6516 if (oldin
->is_dir() && _dir_is_nonempty_unlocked(mdr
, oldin
)) {
6517 respond_to_request(mdr
, -ENOTEMPTY
);
6520 if (srci
== oldin
&& !srcdn
->get_dir()->inode
->is_stray()) {
6521 respond_to_request(mdr
, 0); // no-op. POSIX makes no sense.
6526 // -- some sanity checks --
6528 // src+dest traces _must_ share a common ancestor for locking to prevent orphans
6529 if (destpath
.get_ino() != srcpath
.get_ino() &&
6530 !(req
->get_source().is_mds() &&
6531 MDS_INO_IS_MDSDIR(srcpath
.get_ino()))) { // <-- mds 'rename' out of stray dir is ok!
6532 CInode
*srcbase
= srctrace
[0]->get_dir()->get_inode();
6533 CInode
*destbase
= desttrace
[0]->get_dir()->get_inode();
6534 // ok, extend srctrace toward root until it is an ancestor of desttrace.
6535 while (srcbase
!= destbase
&&
6536 !srcbase
->is_projected_ancestor_of(destbase
)) {
6537 CDentry
*pdn
= srcbase
->get_projected_parent_dn();
6538 srctrace
.insert(srctrace
.begin(), pdn
);
6539 dout(10) << "rename prepending srctrace with " << *pdn
<< dendl
;
6540 srcbase
= pdn
->get_dir()->get_inode();
6543 // then, extend destpath until it shares the same parent inode as srcpath.
6544 while (destbase
!= srcbase
) {
6545 CDentry
*pdn
= destbase
->get_projected_parent_dn();
6546 desttrace
.insert(desttrace
.begin(), pdn
);
6547 rdlocks
.insert(&pdn
->lock
);
6548 dout(10) << "rename prepending desttrace with " << *pdn
<< dendl
;
6549 destbase
= pdn
->get_dir()->get_inode();
6551 dout(10) << "rename src and dest traces now share common ancestor " << *destbase
<< dendl
;
6555 if (srcdn
->get_dir() == destdir
&& srcdn
->name
== destname
) {
6556 dout(7) << "rename src=dest, noop" << dendl
;
6557 respond_to_request(mdr
, 0);
6561 // dest a child of src?
6562 // e.g. mv /usr /usr/foo
6563 CDentry
*pdn
= destdir
->inode
->get_projected_parent_dn();
6566 dout(7) << "cannot rename item to be a child of itself" << dendl
;
6567 respond_to_request(mdr
, -EINVAL
);
6570 pdn
= pdn
->get_dir()->inode
->parent
;
6573 // is this a stray migration, reintegration or merge? (sanity checks!)
6574 if (mdr
->reqid
.name
.is_mds() &&
6575 !(MDS_INO_IS_MDSDIR(srcpath
.get_ino()) &&
6576 MDS_INO_IS_MDSDIR(destpath
.get_ino())) &&
6577 !(destdnl
->is_remote() &&
6578 destdnl
->get_remote_ino() == srci
->ino())) {
6579 respond_to_request(mdr
, -EINVAL
); // actually, this won't reply, but whatev.
6583 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
6584 (srcdnl
->is_primary() || destdnl
->is_primary()));
6586 dout(10) << " this is a link merge" << dendl
;
6588 // -- create stray dentry? --
6589 CDentry
*straydn
= NULL
;
6590 if (destdnl
->is_primary() && !linkmerge
) {
6591 straydn
= prepare_stray_dentry(mdr
, destdnl
->get_inode());
6594 dout(10) << " straydn is " << *straydn
<< dendl
;
6595 } else if (mdr
->straydn
) {
6596 mdr
->unpin(mdr
->straydn
);
6597 mdr
->straydn
= NULL
;
6600 // -- prepare witness list --
6602 * NOTE: we use _all_ replicas as witnesses.
6603 * this probably isn't totally necessary (esp for file renames),
6604 * but if/when we change that, we have to make sure rejoin is
6605 * sufficiently robust to handle strong rejoins from survivors
6606 * with totally wrong dentry->inode linkage.
6607 * (currently, it can ignore rename effects, because the resolve
6608 * stage will sort them out.)
6610 set
<mds_rank_t
> witnesses
= mdr
->more()->extra_witnesses
;
6611 if (srcdn
->is_auth())
6612 srcdn
->list_replicas(witnesses
);
6614 witnesses
.insert(srcdn
->authority().first
);
6615 if (srcdnl
->is_remote() && !srci
->is_auth())
6616 witnesses
.insert(srci
->authority().first
);
6617 destdn
->list_replicas(witnesses
);
6618 if (destdnl
->is_remote() && !oldin
->is_auth())
6619 witnesses
.insert(oldin
->authority().first
);
6620 dout(10) << " witnesses " << witnesses
<< ", have " << mdr
->more()->witnessed
<< dendl
;
6624 map
<SimpleLock
*, mds_rank_t
> remote_wrlocks
;
6626 // srctrace items. this mirrors locks taken in rdlock_path_xlock_dentry
6627 for (int i
=0; i
<(int)srctrace
.size(); i
++)
6628 rdlocks
.insert(&srctrace
[i
]->lock
);
6629 xlocks
.insert(&srcdn
->lock
);
6630 mds_rank_t srcdirauth
= srcdn
->get_dir()->authority().first
;
6631 if (srcdirauth
!= mds
->get_nodeid()) {
6632 dout(10) << " will remote_wrlock srcdir scatterlocks on mds." << srcdirauth
<< dendl
;
6633 remote_wrlocks
[&srcdn
->get_dir()->inode
->filelock
] = srcdirauth
;
6634 remote_wrlocks
[&srcdn
->get_dir()->inode
->nestlock
] = srcdirauth
;
6636 rdlocks
.insert(&srci
->dirfragtreelock
);
6638 wrlocks
.insert(&srcdn
->get_dir()->inode
->filelock
);
6639 wrlocks
.insert(&srcdn
->get_dir()->inode
->nestlock
);
6641 mds
->locker
->include_snap_rdlocks(rdlocks
, srcdn
->get_dir()->inode
);
6645 wrlocks
.insert(&straydn
->get_dir()->inode
->filelock
);
6646 wrlocks
.insert(&straydn
->get_dir()->inode
->nestlock
);
6647 xlocks
.insert(&straydn
->lock
);
6650 // xlock versionlock on dentries if there are witnesses.
6651 // replicas can't see projected dentry linkages, and will get
6652 // confused if we try to pipeline things.
6653 if (!witnesses
.empty()) {
6654 // take xlock on all projected ancestor dentries for srcdn and destdn.
6655 // this ensures the srcdn and destdn can be traversed to by the witnesses.
6656 for (int i
= 0; i
<(int)srctrace
.size(); i
++) {
6657 if (srctrace
[i
]->is_auth() && srctrace
[i
]->is_projected())
6658 xlocks
.insert(&srctrace
[i
]->versionlock
);
6660 for (int i
=0; i
<(int)desttrace
.size(); i
++) {
6661 if (desttrace
[i
]->is_auth() && desttrace
[i
]->is_projected())
6662 xlocks
.insert(&desttrace
[i
]->versionlock
);
6664 // xlock srci and oldin's primary dentries, so witnesses can call
6665 // open_remote_ino() with 'want_locked=true' when the srcdn or destdn
6667 if (srcdnl
->is_remote())
6668 xlocks
.insert(&srci
->get_projected_parent_dn()->lock
);
6669 if (destdnl
->is_remote())
6670 xlocks
.insert(&oldin
->get_projected_parent_dn()->lock
);
6673 // we need to update srci's ctime. xlock its least contended lock to do that...
6674 xlocks
.insert(&srci
->linklock
);
6676 // xlock oldin (for nlink--)
6678 xlocks
.insert(&oldin
->linklock
);
6679 if (oldin
->is_dir())
6680 rdlocks
.insert(&oldin
->filelock
);
6682 if (srcdnl
->is_primary() && srci
->is_dir())
6683 // FIXME: this should happen whenever we are renamning between
6684 // realms, regardless of the file type
6685 // FIXME: If/when this changes, make sure to update the
6686 // "allowance" in handle_slave_rename_prep
6687 xlocks
.insert(&srci
->snaplock
); // FIXME: an auth bcast could be sufficient?
6689 rdlocks
.insert(&srci
->snaplock
);
6691 CInode
*auth_pin_freeze
= !srcdn
->is_auth() && srcdnl
->is_primary() ? srci
: NULL
;
6692 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
,
6693 &remote_wrlocks
, auth_pin_freeze
))
6696 if ((!mdr
->has_more() || mdr
->more()->witnessed
.empty())) {
6697 if (!check_access(mdr
, srcdn
->get_dir()->get_inode(), MAY_WRITE
))
6700 if (!check_access(mdr
, destdn
->get_dir()->get_inode(), MAY_WRITE
))
6703 if (!check_fragment_space(mdr
, destdn
->get_dir()))
6706 if (!check_access(mdr
, srci
, MAY_WRITE
))
6710 // with read lock, really verify oldin is empty
6713 _dir_is_nonempty(mdr
, oldin
)) {
6714 respond_to_request(mdr
, -ENOTEMPTY
);
6718 /* project_past_snaprealm_parent() will do this job
6720 // moving between snaprealms?
6721 if (srcdnl->is_primary() && srci->is_multiversion() && !srci->snaprealm) {
6722 SnapRealm *srcrealm = srci->find_snaprealm();
6723 SnapRealm *destrealm = destdn->get_dir()->inode->find_snaprealm();
6724 if (srcrealm != destrealm &&
6725 (srcrealm->get_newest_seq() + 1 > srcdn->first ||
6726 destrealm->get_newest_seq() + 1 > srcdn->first)) {
6727 dout(10) << " renaming between snaprealms, creating snaprealm for " << *srci << dendl;
6728 mdcache->snaprealm_create(mdr, srci);
6734 assert(g_conf
->mds_kill_rename_at
!= 1);
6736 // -- open all srcdn inode frags, if any --
6737 // we need these open so that auth can properly delegate from inode to dirfrags
6738 // after the inode is _ours_.
6739 if (srcdnl
->is_primary() &&
6740 !srcdn
->is_auth() &&
6742 dout(10) << "srci is remote dir, setting stickydirs and opening all frags" << dendl
;
6743 mdr
->set_stickydirs(srci
);
6746 srci
->dirfragtree
.get_leaves(frags
);
6747 for (list
<frag_t
>::iterator p
= frags
.begin();
6750 CDir
*dir
= srci
->get_dirfrag(*p
);
6752 dout(10) << " opening " << *p
<< " under " << *srci
<< dendl
;
6753 mdcache
->open_remote_dirfrag(srci
, *p
, new C_MDS_RetryRequest(mdcache
, mdr
));
6759 // -- prepare witnesses --
6761 // do srcdn auth last
6762 mds_rank_t last
= MDS_RANK_NONE
;
6763 if (!srcdn
->is_auth()) {
6764 last
= srcdn
->authority().first
;
6765 mdr
->more()->srcdn_auth_mds
= last
;
6766 // ask auth of srci to mark srci as ambiguous auth if more than two MDS
6767 // are involved in the rename operation.
6768 if (srcdnl
->is_primary() && !mdr
->more()->is_ambiguous_auth
) {
6769 dout(10) << " preparing ambiguous auth for srci" << dendl
;
6770 assert(mdr
->more()->is_remote_frozen_authpin
);
6771 assert(mdr
->more()->rename_inode
== srci
);
6772 _rename_prepare_witness(mdr
, last
, witnesses
, srctrace
, desttrace
, straydn
);
6777 for (set
<mds_rank_t
>::iterator p
= witnesses
.begin();
6778 p
!= witnesses
.end();
6780 if (*p
== last
) continue; // do it last!
6781 if (mdr
->more()->witnessed
.count(*p
)) {
6782 dout(10) << " already witnessed by mds." << *p
<< dendl
;
6783 } else if (mdr
->more()->waiting_on_slave
.count(*p
)) {
6784 dout(10) << " already waiting on witness mds." << *p
<< dendl
;
6786 if (!_rename_prepare_witness(mdr
, *p
, witnesses
, srctrace
, desttrace
, straydn
))
6790 if (!mdr
->more()->waiting_on_slave
.empty())
6791 return; // we're waiting for a witness.
6793 if (last
!= MDS_RANK_NONE
&& mdr
->more()->witnessed
.count(last
) == 0) {
6794 dout(10) << " preparing last witness (srcdn auth)" << dendl
;
6795 assert(mdr
->more()->waiting_on_slave
.count(last
) == 0);
6796 _rename_prepare_witness(mdr
, last
, witnesses
, srctrace
, desttrace
, straydn
);
6800 // test hack: bail after slave does prepare, so we can verify it's _live_ rollback.
6801 if (!mdr
->more()->slaves
.empty() && !srci
->is_dir())
6802 assert(g_conf
->mds_kill_rename_at
!= 3);
6803 if (!mdr
->more()->slaves
.empty() && srci
->is_dir())
6804 assert(g_conf
->mds_kill_rename_at
!= 4);
6806 // -- declare now --
6807 mdr
->set_mds_stamp(ceph_clock_now());
6809 // -- prepare journal entry --
6810 mdr
->ls
= mdlog
->get_current_segment();
6811 EUpdate
*le
= new EUpdate(mdlog
, "rename");
6812 mdlog
->start_entry(le
);
6813 le
->metablob
.add_client_req(mdr
->reqid
, mdr
->client_request
->get_oldest_client_tid());
6814 if (!mdr
->more()->witnessed
.empty()) {
6815 dout(20) << " noting uncommitted_slaves " << mdr
->more()->witnessed
<< dendl
;
6817 le
->reqid
= mdr
->reqid
;
6818 le
->had_slaves
= true;
6820 mdcache
->add_uncommitted_master(mdr
->reqid
, mdr
->ls
, mdr
->more()->witnessed
);
6821 // no need to send frozen auth pin to recovring auth MDS of srci
6822 mdr
->more()->is_remote_frozen_authpin
= false;
6825 _rename_prepare(mdr
, &le
->metablob
, &le
->client_map
, srcdn
, destdn
, straydn
);
6826 if (le
->client_map
.length())
6827 le
->cmapv
= mds
->sessionmap
.get_projected();
6829 // -- commit locally --
6830 C_MDS_rename_finish
*fin
= new C_MDS_rename_finish(this, mdr
, srcdn
, destdn
, straydn
);
6832 journal_and_reply(mdr
, srci
, destdn
, le
, fin
);
6836 void Server::_rename_finish(MDRequestRef
& mdr
, CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
6838 dout(10) << "_rename_finish " << *mdr
<< dendl
;
6840 if (!mdr
->more()->witnessed
.empty())
6841 mdcache
->logged_master_update(mdr
->reqid
);
6844 _rename_apply(mdr
, srcdn
, destdn
, straydn
);
6846 mdcache
->send_dentry_link(destdn
, mdr
);
6848 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
6849 CInode
*in
= destdnl
->get_inode();
6850 bool need_eval
= mdr
->more()->cap_imports
.count(in
);
6852 // test hack: test slave commit
6853 if (!mdr
->more()->slaves
.empty() && !in
->is_dir())
6854 assert(g_conf
->mds_kill_rename_at
!= 5);
6855 if (!mdr
->more()->slaves
.empty() && in
->is_dir())
6856 assert(g_conf
->mds_kill_rename_at
!= 6);
6859 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), srcdn
->get_dir(), META_POP_IWR
);
6860 if (destdnl
->is_remote() && in
->is_auth())
6861 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), in
, META_POP_IWR
);
6863 // did we import srci? if so, explicitly ack that import that, before we unlock and reply.
6865 assert(g_conf
->mds_kill_rename_at
!= 7);
6868 respond_to_request(mdr
, 0);
6871 mds
->locker
->eval(in
, CEPH_CAP_LOCKS
, true);
6874 // respond_to_request() drops locks. So stray reintegration can race with us.
6875 if (straydn
&& !straydn
->get_projected_linkage()->is_null()) {
6876 mdcache
->notify_stray(straydn
);
6884 bool Server::_rename_prepare_witness(MDRequestRef
& mdr
, mds_rank_t who
, set
<mds_rank_t
> &witnesse
,
6885 vector
<CDentry
*>& srctrace
, vector
<CDentry
*>& dsttrace
, CDentry
*straydn
)
6887 if (mds
->is_cluster_degraded() &&
6888 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(who
)) {
6889 dout(10) << "_rename_prepare_witness mds." << who
<< " is not active" << dendl
;
6890 if (mdr
->more()->waiting_on_slave
.empty())
6891 mds
->wait_for_active_peer(who
, new C_MDS_RetryRequest(mdcache
, mdr
));
6895 dout(10) << "_rename_prepare_witness mds." << who
<< dendl
;
6896 MMDSSlaveRequest
*req
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
6897 MMDSSlaveRequest::OP_RENAMEPREP
);
6899 req
->srcdnpath
= filepath(srctrace
.front()->get_dir()->ino());
6900 for (auto dn
: srctrace
)
6901 req
->srcdnpath
.push_dentry(dn
->name
);
6902 req
->destdnpath
= filepath(dsttrace
.front()->get_dir()->ino());
6903 for (auto dn
: dsttrace
)
6904 req
->destdnpath
.push_dentry(dn
->name
);
6906 mdcache
->replicate_stray(straydn
, who
, req
->stray
);
6908 req
->srcdn_auth
= mdr
->more()->srcdn_auth_mds
;
6910 // srcdn auth will verify our current witness list is sufficient
6911 req
->witnesses
= witnesse
;
6913 req
->op_stamp
= mdr
->get_op_stamp();
6914 mds
->send_message_mds(req
, who
);
6916 assert(mdr
->more()->waiting_on_slave
.count(who
) == 0);
6917 mdr
->more()->waiting_on_slave
.insert(who
);
6921 version_t
Server::_rename_prepare_import(MDRequestRef
& mdr
, CDentry
*srcdn
, bufferlist
*client_map_bl
)
6923 version_t oldpv
= mdr
->more()->inode_import_v
;
6925 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
6928 bufferlist::iterator blp
= mdr
->more()->inode_import
.begin();
6931 ::decode(mdr
->more()->imported_client_map
, blp
);
6932 ::encode(mdr
->more()->imported_client_map
, *client_map_bl
,
6933 mds
->mdsmap
->get_up_features());
6934 prepare_force_open_sessions(mdr
->more()->imported_client_map
, mdr
->more()->sseq_map
);
6936 list
<ScatterLock
*> updated_scatterlocks
;
6937 mdcache
->migrator
->decode_import_inode(srcdn
, blp
, srcdn
->authority().first
, mdr
->ls
,
6938 mdr
->more()->cap_imports
, updated_scatterlocks
);
6940 // hack: force back to !auth and clean, temporarily
6941 srcdnl
->get_inode()->state_clear(CInode::STATE_AUTH
);
6942 srcdnl
->get_inode()->mark_clean();
6947 bool Server::_need_force_journal(CInode
*diri
, bool empty
)
6950 diri
->get_dirfrags(ls
);
6952 bool force_journal
= false;
6954 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6955 if ((*p
)->is_subtree_root() && (*p
)->get_dir_auth().first
== mds
->get_nodeid()) {
6956 dout(10) << " frag " << (*p
)->get_frag() << " is auth subtree dirfrag, will force journal" << dendl
;
6957 force_journal
= true;
6960 dout(20) << " frag " << (*p
)->get_frag() << " is not auth subtree dirfrag" << dendl
;
6963 // see if any children of our frags are auth subtrees.
6964 list
<CDir
*> subtrees
;
6965 mdcache
->list_subtrees(subtrees
);
6966 dout(10) << " subtrees " << subtrees
<< " frags " << ls
<< dendl
;
6967 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
6969 for (list
<CDir
*>::iterator q
= subtrees
.begin(); q
!= subtrees
.end(); ++q
) {
6970 if (dir
->contains(*q
)) {
6971 if ((*q
)->get_dir_auth().first
== mds
->get_nodeid()) {
6972 dout(10) << " frag " << (*p
)->get_frag() << " contains (maybe) auth subtree, will force journal "
6974 force_journal
= true;
6977 dout(20) << " frag " << (*p
)->get_frag() << " contains but isn't auth for " << **q
<< dendl
;
6979 dout(20) << " frag " << (*p
)->get_frag() << " does not contain " << **q
<< dendl
;
6985 return force_journal
;
6988 void Server::_rename_prepare(MDRequestRef
& mdr
,
6989 EMetaBlob
*metablob
, bufferlist
*client_map_bl
,
6990 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
6992 dout(10) << "_rename_prepare " << *mdr
<< " " << *srcdn
<< " " << *destdn
<< dendl
;
6994 dout(10) << " straydn " << *straydn
<< dendl
;
6996 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
6997 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
6998 CInode
*srci
= srcdnl
->get_inode();
6999 CInode
*oldin
= destdnl
->get_inode();
7001 // primary+remote link merge?
7002 bool linkmerge
= (srci
== destdnl
->get_inode() &&
7003 (srcdnl
->is_primary() || destdnl
->is_primary()));
7004 bool silent
= srcdn
->get_dir()->inode
->is_stray();
7006 bool force_journal_dest
= false;
7007 if (srci
->is_dir() && !destdn
->is_auth()) {
7008 if (srci
->is_auth()) {
7009 // if we are auth for srci and exporting it, force journal because journal replay needs
7010 // the source inode to create auth subtrees.
7011 dout(10) << " we are exporting srci, will force journal destdn" << dendl
;
7012 force_journal_dest
= true;
7014 force_journal_dest
= _need_force_journal(srci
, false);
7017 bool force_journal_stray
= false;
7018 if (oldin
&& oldin
->is_dir() && straydn
&& !straydn
->is_auth())
7019 force_journal_stray
= _need_force_journal(oldin
, true);
7022 dout(10) << " merging remote and primary links to the same inode" << dendl
;
7024 dout(10) << " reintegrating stray; will avoid changing nlink or dir mtime" << dendl
;
7025 if (force_journal_dest
)
7026 dout(10) << " forcing journal destdn because we (will) have auth subtrees nested beneath it" << dendl
;
7027 if (force_journal_stray
)
7028 dout(10) << " forcing journal straydn because we (will) have auth subtrees nested beneath it" << dendl
;
7030 if (srci
->is_dir() && (destdn
->is_auth() || force_journal_dest
)) {
7031 dout(10) << " noting renamed dir ino " << srci
->ino() << " in metablob" << dendl
;
7032 metablob
->renamed_dirino
= srci
->ino();
7033 } else if (oldin
&& oldin
->is_dir() && force_journal_stray
) {
7034 dout(10) << " noting rename target dir " << oldin
->ino() << " in metablob" << dendl
;
7035 metablob
->renamed_dirino
= oldin
->ino();
7039 inode_t
*pi
= 0; // renamed inode
7040 inode_t
*tpi
= 0; // target/overwritten inode
7044 if (destdnl
->is_primary()) {
7045 assert(straydn
); // moving to straydn.
7046 // link--, and move.
7047 if (destdn
->is_auth()) {
7048 tpi
= oldin
->project_inode(); //project_snaprealm
7049 tpi
->version
= straydn
->pre_dirty(tpi
->version
);
7050 tpi
->update_backtrace();
7052 straydn
->push_projected_linkage(oldin
);
7053 } else if (destdnl
->is_remote()) {
7055 if (oldin
->is_auth()) {
7056 tpi
= oldin
->project_inode();
7057 tpi
->version
= oldin
->pre_dirty();
7063 if (srcdnl
->is_remote()) {
7066 if (destdn
->is_auth())
7067 mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty();
7068 destdn
->push_projected_linkage(srcdnl
->get_remote_ino(), srcdnl
->get_remote_d_type());
7070 if (srci
->is_auth()) {
7071 pi
= srci
->project_inode();
7072 pi
->version
= srci
->pre_dirty();
7075 dout(10) << " will merge remote onto primary link" << dendl
;
7076 if (destdn
->is_auth()) {
7077 pi
= oldin
->project_inode();
7078 pi
->version
= mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty(oldin
->inode
.version
);
7082 if (destdn
->is_auth()) {
7084 if (srcdn
->is_auth())
7085 oldpv
= srci
->get_projected_version();
7087 oldpv
= _rename_prepare_import(mdr
, srcdn
, client_map_bl
);
7089 // note which dirfrags have child subtrees in the journal
7090 // event, so that we can open those (as bounds) during replay.
7091 if (srci
->is_dir()) {
7093 srci
->get_dirfrags(ls
);
7094 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
7096 if (!dir
->is_auth())
7097 metablob
->renamed_dir_frags
.push_back(dir
->get_frag());
7099 dout(10) << " noting renamed dir open frags " << metablob
->renamed_dir_frags
<< dendl
;
7102 pi
= srci
->project_inode(); // project snaprealm if srcdnl->is_primary
7103 // & srcdnl->snaprealm
7104 pi
->version
= mdr
->more()->pvmap
[destdn
] = destdn
->pre_dirty(oldpv
);
7105 pi
->update_backtrace();
7107 destdn
->push_projected_linkage(srci
);
7111 if (srcdn
->is_auth())
7112 mdr
->more()->pvmap
[srcdn
] = srcdn
->pre_dirty();
7113 srcdn
->push_projected_linkage(); // push null linkage
7117 pi
->ctime
= mdr
->get_op_stamp();
7123 tpi
->ctime
= mdr
->get_op_stamp();
7125 destdn
->make_path_string(tpi
->stray_prior_path
, true);
7127 if (tpi
->nlink
== 0)
7128 oldin
->state_set(CInode::STATE_ORPHAN
);
7132 // prepare nesting, mtime updates
7133 int predirty_dir
= silent
? 0:PREDIRTY_DIR
;
7135 // guarantee stray dir is processed first during journal replay. unlink the old inode,
7136 // then link the source inode to destdn
7137 if (destdnl
->is_primary()) {
7139 if (straydn
->is_auth()) {
7140 metablob
->add_dir_context(straydn
->get_dir());
7141 metablob
->add_dir(straydn
->get_dir(), true);
7146 if (destdn
->is_auth() && !destdnl
->is_null()) {
7147 mdcache
->predirty_journal_parents(mdr
, metablob
, oldin
, destdn
->get_dir(),
7148 (destdnl
->is_primary() ? PREDIRTY_PRIMARY
:0)|predirty_dir
, -1);
7149 if (destdnl
->is_primary()) {
7151 mdcache
->predirty_journal_parents(mdr
, metablob
, oldin
, straydn
->get_dir(),
7152 PREDIRTY_PRIMARY
|PREDIRTY_DIR
, 1);
7157 int predirty_primary
= (srcdnl
->is_primary() && srcdn
->get_dir() != destdn
->get_dir()) ? PREDIRTY_PRIMARY
:0;
7158 int flags
= predirty_dir
| predirty_primary
;
7159 if (srcdn
->is_auth())
7160 mdcache
->predirty_journal_parents(mdr
, metablob
, srci
, srcdn
->get_dir(), PREDIRTY_SHALLOW
|flags
, -1);
7161 if (destdn
->is_auth())
7162 mdcache
->predirty_journal_parents(mdr
, metablob
, srci
, destdn
->get_dir(), flags
, 1);
7164 SnapRealm
*src_realm
= srci
->find_snaprealm();
7165 SnapRealm
*dest_realm
= destdn
->get_dir()->inode
->find_snaprealm();
7166 snapid_t next_dest_snap
= dest_realm
->get_newest_seq() + 1;
7168 // add it all to the metablob
7171 if (destdnl
->is_primary()) {
7173 if (destdn
->is_auth()) {
7174 // project snaprealm, too
7175 if (oldin
->snaprealm
|| dest_realm
->get_newest_seq() + 1 > oldin
->get_oldest_snap())
7176 oldin
->project_past_snaprealm_parent(straydn
->get_dir()->inode
->find_snaprealm());
7177 straydn
->first
= MAX(oldin
->first
, next_dest_snap
);
7178 metablob
->add_primary_dentry(straydn
, oldin
, true, true);
7179 } else if (force_journal_stray
) {
7180 dout(10) << " forced journaling straydn " << *straydn
<< dendl
;
7181 metablob
->add_dir_context(straydn
->get_dir());
7182 metablob
->add_primary_dentry(straydn
, oldin
, true);
7184 } else if (destdnl
->is_remote()) {
7185 if (oldin
->is_auth()) {
7187 metablob
->add_dir_context(oldin
->get_projected_parent_dir());
7188 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, oldin
->get_projected_parent_dn(),
7189 CEPH_NOSNAP
, 0, destdnl
);
7190 metablob
->add_primary_dentry(oldin
->get_projected_parent_dn(), oldin
, true);
7196 if (srcdnl
->is_remote()) {
7198 if (destdn
->is_auth() && !destdnl
->is_null())
7199 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7201 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7203 if (destdn
->is_auth())
7204 metablob
->add_remote_dentry(destdn
, true, srcdnl
->get_remote_ino(), srcdnl
->get_remote_d_type());
7205 if (srci
->get_projected_parent_dn()->is_auth()) { // it's remote
7206 metablob
->add_dir_context(srci
->get_projected_parent_dir());
7207 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, srci
->get_projected_parent_dn(), CEPH_NOSNAP
, 0, srcdnl
);
7208 metablob
->add_primary_dentry(srci
->get_projected_parent_dn(), srci
, true);
7211 if (destdn
->is_auth() && !destdnl
->is_null())
7212 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7214 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7216 if (destdn
->is_auth())
7217 metablob
->add_primary_dentry(destdn
, destdnl
->get_inode(), true, true);
7219 } else if (srcdnl
->is_primary()) {
7220 // project snap parent update?
7221 if (destdn
->is_auth() && src_realm
!= dest_realm
&&
7222 (srci
->snaprealm
|| src_realm
->get_newest_seq() + 1 > srci
->get_oldest_snap()))
7223 srci
->project_past_snaprealm_parent(dest_realm
);
7225 if (destdn
->is_auth() && !destdnl
->is_null())
7226 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, destdn
, CEPH_NOSNAP
, 0, destdnl
);
7228 destdn
->first
= MAX(destdn
->first
, next_dest_snap
);
7230 if (destdn
->is_auth())
7231 metablob
->add_primary_dentry(destdn
, srci
, true, true);
7232 else if (force_journal_dest
) {
7233 dout(10) << " forced journaling destdn " << *destdn
<< dendl
;
7234 metablob
->add_dir_context(destdn
->get_dir());
7235 metablob
->add_primary_dentry(destdn
, srci
, true);
7236 if (srcdn
->is_auth() && srci
->is_dir()) {
7237 // journal new subtrees root dirfrags
7239 srci
->get_dirfrags(ls
);
7240 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
7243 metablob
->add_dir(dir
, true);
7250 if (srcdn
->is_auth()) {
7251 dout(10) << " journaling srcdn " << *srcdn
<< dendl
;
7252 mdcache
->journal_cow_dentry(mdr
.get(), metablob
, srcdn
, CEPH_NOSNAP
, 0, srcdnl
);
7253 // also journal the inode in case we need do slave rename rollback. It is Ok to add
7254 // both primary and NULL dentries. Because during journal replay, null dentry is
7255 // processed after primary dentry.
7256 if (srcdnl
->is_primary() && !srci
->is_dir() && !destdn
->is_auth())
7257 metablob
->add_primary_dentry(srcdn
, srci
, true);
7258 metablob
->add_null_dentry(srcdn
, true);
7260 dout(10) << " NOT journaling srcdn " << *srcdn
<< dendl
;
7262 // make renamed inode first track the dn
7263 if (srcdnl
->is_primary() && destdn
->is_auth())
7264 srci
->first
= destdn
->first
;
7266 if (oldin
&& oldin
->is_dir()) {
7268 mdcache
->project_subtree_rename(oldin
, destdn
->get_dir(), straydn
->get_dir());
7271 mdcache
->project_subtree_rename(srci
, srcdn
->get_dir(), destdn
->get_dir());
7276 void Server::_rename_apply(MDRequestRef
& mdr
, CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7278 dout(10) << "_rename_apply " << *mdr
<< " " << *srcdn
<< " " << *destdn
<< dendl
;
7279 dout(10) << " pvs " << mdr
->more()->pvmap
<< dendl
;
7281 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
7282 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
7284 CInode
*oldin
= destdnl
->get_inode();
7286 // primary+remote link merge?
7287 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
7288 (srcdnl
->is_primary() || destdnl
->is_primary()));
7292 if (destdnl
->is_primary()) {
7294 dout(10) << "straydn is " << *straydn
<< dendl
;
7295 destdn
->get_dir()->unlink_inode(destdn
, false);
7297 straydn
->pop_projected_linkage();
7298 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7299 assert(!straydn
->is_projected()); // no other projected
7301 mdcache
->touch_dentry_bottom(straydn
); // drop dn as quickly as possible.
7304 if (destdn
->is_auth()) {
7305 bool hadrealm
= (oldin
->snaprealm
? true : false);
7306 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7307 if (oldin
->snaprealm
&& !hadrealm
)
7308 mdcache
->do_realm_invalidate_and_update_notify(oldin
, CEPH_SNAP_OP_SPLIT
);
7310 // FIXME this snaprealm is not filled out correctly
7311 //oldin->open_snaprealm(); might be sufficient..
7313 } else if (destdnl
->is_remote()) {
7314 destdn
->get_dir()->unlink_inode(destdn
, false);
7315 if (oldin
->is_auth())
7316 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7320 // unlink src before we relink it at dest
7321 CInode
*in
= srcdnl
->get_inode();
7324 bool srcdn_was_remote
= srcdnl
->is_remote();
7325 srcdn
->get_dir()->unlink_inode(srcdn
);
7328 if (srcdn_was_remote
) {
7331 destdnl
= destdn
->pop_projected_linkage();
7332 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7333 assert(!destdn
->is_projected()); // no other projected
7335 destdn
->link_remote(destdnl
, in
);
7336 if (destdn
->is_auth())
7337 destdn
->mark_dirty(mdr
->more()->pvmap
[destdn
], mdr
->ls
);
7340 in
->pop_and_dirty_projected_inode(mdr
->ls
);
7342 dout(10) << "merging remote onto primary link" << dendl
;
7343 oldin
->pop_and_dirty_projected_inode(mdr
->ls
);
7347 dout(10) << "merging primary onto remote link" << dendl
;
7348 destdn
->get_dir()->unlink_inode(destdn
, false);
7350 destdnl
= destdn
->pop_projected_linkage();
7351 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7352 assert(!destdn
->is_projected()); // no other projected
7354 // srcdn inode import?
7355 if (!srcdn
->is_auth() && destdn
->is_auth()) {
7356 assert(mdr
->more()->inode_import
.length() > 0);
7358 map
<client_t
,Capability::Import
> imported_caps
;
7360 // finish cap imports
7361 finish_force_open_sessions(mdr
->more()->imported_client_map
, mdr
->more()->sseq_map
);
7362 if (mdr
->more()->cap_imports
.count(destdnl
->get_inode())) {
7363 mdcache
->migrator
->finish_import_inode_caps(destdnl
->get_inode(),
7364 mdr
->more()->srcdn_auth_mds
, true,
7365 mdr
->more()->cap_imports
[destdnl
->get_inode()],
7369 mdr
->more()->inode_import
.clear();
7370 ::encode(imported_caps
, mdr
->more()->inode_import
);
7372 /* hack: add an auth pin for each xlock we hold. These were
7373 * remote xlocks previously but now they're local and
7374 * we're going to try and unpin when we xlock_finish. */
7375 for (set
<SimpleLock
*>::iterator i
= mdr
->xlocks
.begin();
7376 i
!= mdr
->xlocks
.end();
7378 if ((*i
)->get_parent() == destdnl
->get_inode() &&
7379 !(*i
)->is_locallock())
7380 mds
->locker
->xlock_import(*i
);
7382 // hack: fix auth bit
7383 in
->state_set(CInode::STATE_AUTH
);
7385 mdr
->clear_ambiguous_auth();
7388 if (destdn
->is_auth()) {
7389 in
->pop_and_dirty_projected_inode(mdr
->ls
);
7392 // FIXME: fix up snaprealm!
7397 if (srcdn
->is_auth())
7398 srcdn
->mark_dirty(mdr
->more()->pvmap
[srcdn
], mdr
->ls
);
7399 srcdn
->pop_projected_linkage();
7400 if (mdr
->is_slave() && !mdr
->more()->slave_update_journaled
)
7401 assert(!srcdn
->is_projected()); // no other projected
7403 // apply remaining projected inodes (nested)
7406 // update subtree map?
7407 if (destdnl
->is_primary() && in
->is_dir())
7408 mdcache
->adjust_subtree_after_rename(in
, srcdn
->get_dir(), true);
7410 if (straydn
&& oldin
->is_dir())
7411 mdcache
->adjust_subtree_after_rename(oldin
, destdn
->get_dir(), true);
7413 // removing a new dn?
7414 if (srcdn
->is_auth())
7415 srcdn
->get_dir()->try_remove_unlinked_dn(srcdn
);
7423 class C_MDS_SlaveRenamePrep
: public ServerLogContext
{
7424 CDentry
*srcdn
, *destdn
, *straydn
;
7426 C_MDS_SlaveRenamePrep(Server
*s
, MDRequestRef
& m
, CDentry
*sr
, CDentry
*de
, CDentry
*st
) :
7427 ServerLogContext(s
, m
), srcdn(sr
), destdn(de
), straydn(st
) {}
7428 void finish(int r
) override
{
7429 server
->_logged_slave_rename(mdr
, srcdn
, destdn
, straydn
);
7433 class C_MDS_SlaveRenameCommit
: public ServerContext
{
7435 CDentry
*srcdn
, *destdn
, *straydn
;
7437 C_MDS_SlaveRenameCommit(Server
*s
, MDRequestRef
& m
, CDentry
*sr
, CDentry
*de
, CDentry
*st
) :
7438 ServerContext(s
), mdr(m
), srcdn(sr
), destdn(de
), straydn(st
) {}
7439 void finish(int r
) override
{
7440 server
->_commit_slave_rename(mdr
, r
, srcdn
, destdn
, straydn
);
7444 class C_MDS_SlaveRenameSessionsFlushed
: public ServerContext
{
7447 C_MDS_SlaveRenameSessionsFlushed(Server
*s
, MDRequestRef
& r
) :
7448 ServerContext(s
), mdr(r
) {}
7449 void finish(int r
) override
{
7450 server
->_slave_rename_sessions_flushed(mdr
);
7454 /* This function DOES put the mdr->slave_request before returning*/
7455 void Server::handle_slave_rename_prep(MDRequestRef
& mdr
)
7457 dout(10) << "handle_slave_rename_prep " << *mdr
7458 << " " << mdr
->slave_request
->srcdnpath
7459 << " to " << mdr
->slave_request
->destdnpath
7462 if (mdr
->slave_request
->is_interrupted()) {
7463 dout(10) << " slave request interrupted, sending noop reply" << dendl
;
7464 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_RENAMEPREPACK
);
7465 reply
->mark_interrupted();
7466 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7467 mdr
->slave_request
->put();
7468 mdr
->slave_request
= 0;
7473 filepath
destpath(mdr
->slave_request
->destdnpath
);
7474 dout(10) << " dest " << destpath
<< dendl
;
7475 vector
<CDentry
*> trace
;
7476 int r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, destpath
, &trace
, NULL
, MDS_TRAVERSE_DISCOVERXLOCK
);
7479 mdcache
->find_ino_peers(destpath
.get_ino(), new C_MDS_RetryRequest(mdcache
, mdr
),
7483 assert(r
== 0); // we shouldn't get an error here!
7485 CDentry
*destdn
= trace
[trace
.size()-1];
7486 CDentry::linkage_t
*destdnl
= destdn
->get_projected_linkage();
7487 dout(10) << " destdn " << *destdn
<< dendl
;
7491 filepath
srcpath(mdr
->slave_request
->srcdnpath
);
7492 dout(10) << " src " << srcpath
<< dendl
;
7493 CInode
*srci
= nullptr;
7494 r
= mdcache
->path_traverse(mdr
, NULL
, NULL
, srcpath
, &trace
, &srci
, MDS_TRAVERSE_DISCOVERXLOCK
);
7498 // srcpath must not point to a null dentry
7499 assert(srci
!= nullptr);
7501 CDentry
*srcdn
= trace
[trace
.size()-1];
7502 CDentry::linkage_t
*srcdnl
= srcdn
->get_projected_linkage();
7503 dout(10) << " srcdn " << *srcdn
<< dendl
;
7508 bool linkmerge
= (srcdnl
->get_inode() == destdnl
->get_inode() &&
7509 (srcdnl
->is_primary() || destdnl
->is_primary()));
7510 CDentry
*straydn
= mdr
->straydn
;
7511 if (destdnl
->is_primary() && !linkmerge
)
7514 mdr
->set_op_stamp(mdr
->slave_request
->op_stamp
);
7515 mdr
->more()->srcdn_auth_mds
= srcdn
->authority().first
;
7517 // set up commit waiter (early, to clean up any freezing etc we do)
7518 if (!mdr
->more()->slave_commit
)
7519 mdr
->more()->slave_commit
= new C_MDS_SlaveRenameCommit(this, mdr
, srcdn
, destdn
, straydn
);
7522 if (srcdn
->is_auth()) {
7523 set
<mds_rank_t
> srcdnrep
;
7524 srcdn
->list_replicas(srcdnrep
);
7526 bool reply_witness
= false;
7527 if (srcdnl
->is_primary() && !srcdnl
->get_inode()->state_test(CInode::STATE_AMBIGUOUSAUTH
)) {
7530 // - avoid conflicting lock state changes
7531 // - avoid concurrent updates to the inode
7532 // (this could also be accomplished with the versionlock)
7533 int allowance
= 2; // 1 for the mdr auth_pin, 1 for the link lock
7534 allowance
+= srcdnl
->get_inode()->is_dir(); // for the snap lock
7535 dout(10) << " freezing srci " << *srcdnl
->get_inode() << " with allowance " << allowance
<< dendl
;
7536 bool frozen_inode
= srcdnl
->get_inode()->freeze_inode(allowance
);
7538 // unfreeze auth pin after freezing the inode to avoid queueing waiters
7539 if (srcdnl
->get_inode()->is_frozen_auth_pin())
7540 mdr
->unfreeze_auth_pin();
7542 if (!frozen_inode
) {
7543 srcdnl
->get_inode()->add_waiter(CInode::WAIT_FROZEN
, new C_MDS_RetryRequest(mdcache
, mdr
));
7548 * set ambiguous auth for srci
7549 * NOTE: we don't worry about ambiguous cache expire as we do
7550 * with subtree migrations because all slaves will pin
7551 * srcdn->get_inode() for duration of this rename.
7553 mdr
->set_ambiguous_auth(srcdnl
->get_inode());
7555 // just mark the source inode as ambiguous auth if more than two MDS are involved.
7556 // the master will send another OP_RENAMEPREP slave request later.
7557 if (mdr
->slave_request
->witnesses
.size() > 1) {
7558 dout(10) << " set srci ambiguous auth; providing srcdn replica list" << dendl
;
7559 reply_witness
= true;
7562 // make sure bystanders have received all lock related messages
7563 for (set
<mds_rank_t
>::iterator p
= srcdnrep
.begin(); p
!= srcdnrep
.end(); ++p
) {
7564 if (*p
== mdr
->slave_to_mds
||
7565 (mds
->is_cluster_degraded() &&
7566 !mds
->mdsmap
->is_clientreplay_or_active_or_stopping(*p
)))
7568 MMDSSlaveRequest
*notify
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
7569 MMDSSlaveRequest::OP_RENAMENOTIFY
);
7570 mds
->send_message_mds(notify
, *p
);
7571 mdr
->more()->waiting_on_slave
.insert(*p
);
7574 // make sure clients have received all cap related messages
7575 set
<client_t
> export_client_set
;
7576 mdcache
->migrator
->get_export_client_set(srcdnl
->get_inode(), export_client_set
);
7578 MDSGatherBuilder
gather(g_ceph_context
);
7579 flush_client_sessions(export_client_set
, gather
);
7580 if (gather
.has_subs()) {
7581 mdr
->more()->waiting_on_slave
.insert(MDS_RANK_NONE
);
7582 gather
.set_finisher(new C_MDS_SlaveRenameSessionsFlushed(this, mdr
));
7587 // is witness list sufficient?
7588 for (set
<mds_rank_t
>::iterator p
= srcdnrep
.begin(); p
!= srcdnrep
.end(); ++p
) {
7589 if (*p
== mdr
->slave_to_mds
||
7590 mdr
->slave_request
->witnesses
.count(*p
)) continue;
7591 dout(10) << " witness list insufficient; providing srcdn replica list" << dendl
;
7592 reply_witness
= true;
7596 if (reply_witness
) {
7597 assert(!srcdnrep
.empty());
7598 MMDSSlaveRequest
*reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
,
7599 MMDSSlaveRequest::OP_RENAMEPREPACK
);
7600 reply
->witnesses
.swap(srcdnrep
);
7601 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7602 mdr
->slave_request
->put();
7603 mdr
->slave_request
= 0;
7606 dout(10) << " witness list sufficient: includes all srcdn replicas" << dendl
;
7607 if (!mdr
->more()->waiting_on_slave
.empty()) {
7608 dout(10) << " still waiting for rename notify acks from "
7609 << mdr
->more()->waiting_on_slave
<< dendl
;
7612 } else if (srcdnl
->is_primary() && srcdn
->authority() != destdn
->authority()) {
7613 // set ambiguous auth for srci on witnesses
7614 mdr
->set_ambiguous_auth(srcdnl
->get_inode());
7617 // encode everything we'd need to roll this back... basically, just the original state.
7618 rename_rollback rollback
;
7620 rollback
.reqid
= mdr
->reqid
;
7622 rollback
.orig_src
.dirfrag
= srcdn
->get_dir()->dirfrag();
7623 rollback
.orig_src
.dirfrag_old_mtime
= srcdn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7624 rollback
.orig_src
.dirfrag_old_rctime
= srcdn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7625 rollback
.orig_src
.dname
= srcdn
->name
;
7626 if (srcdnl
->is_primary())
7627 rollback
.orig_src
.ino
= srcdnl
->get_inode()->ino();
7629 assert(srcdnl
->is_remote());
7630 rollback
.orig_src
.remote_ino
= srcdnl
->get_remote_ino();
7631 rollback
.orig_src
.remote_d_type
= srcdnl
->get_remote_d_type();
7634 rollback
.orig_dest
.dirfrag
= destdn
->get_dir()->dirfrag();
7635 rollback
.orig_dest
.dirfrag_old_mtime
= destdn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7636 rollback
.orig_dest
.dirfrag_old_rctime
= destdn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7637 rollback
.orig_dest
.dname
= destdn
->name
;
7638 if (destdnl
->is_primary())
7639 rollback
.orig_dest
.ino
= destdnl
->get_inode()->ino();
7640 else if (destdnl
->is_remote()) {
7641 rollback
.orig_dest
.remote_ino
= destdnl
->get_remote_ino();
7642 rollback
.orig_dest
.remote_d_type
= destdnl
->get_remote_d_type();
7646 rollback
.stray
.dirfrag
= straydn
->get_dir()->dirfrag();
7647 rollback
.stray
.dirfrag_old_mtime
= straydn
->get_dir()->get_projected_fnode()->fragstat
.mtime
;
7648 rollback
.stray
.dirfrag_old_rctime
= straydn
->get_dir()->get_projected_fnode()->rstat
.rctime
;
7649 rollback
.stray
.dname
= straydn
->name
;
7651 ::encode(rollback
, mdr
->more()->rollback_bl
);
7652 dout(20) << " rollback is " << mdr
->more()->rollback_bl
.length() << " bytes" << dendl
;
7655 mdr
->ls
= mdlog
->get_current_segment();
7656 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_prep", mdr
->reqid
, mdr
->slave_to_mds
,
7657 ESlaveUpdate::OP_PREPARE
, ESlaveUpdate::RENAME
);
7658 mdlog
->start_entry(le
);
7659 le
->rollback
= mdr
->more()->rollback_bl
;
7661 bufferlist blah
; // inode import data... obviously not used if we're the slave
7662 _rename_prepare(mdr
, &le
->commit
, &blah
, srcdn
, destdn
, straydn
);
7664 if (le
->commit
.empty()) {
7665 dout(10) << " empty metablob, skipping journal" << dendl
;
7666 mdlog
->cancel_entry(le
);
7668 _logged_slave_rename(mdr
, srcdn
, destdn
, straydn
);
7670 mdr
->more()->slave_update_journaled
= true;
7671 submit_mdlog_entry(le
, new C_MDS_SlaveRenamePrep(this, mdr
, srcdn
, destdn
, straydn
),
7677 void Server::_logged_slave_rename(MDRequestRef
& mdr
,
7678 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7680 dout(10) << "_logged_slave_rename " << *mdr
<< dendl
;
7683 MMDSSlaveRequest
*reply
= NULL
;
7684 if (!mdr
->aborted
) {
7685 reply
= new MMDSSlaveRequest(mdr
->reqid
, mdr
->attempt
, MMDSSlaveRequest::OP_RENAMEPREPACK
);
7686 if (!mdr
->more()->slave_update_journaled
)
7687 reply
->mark_not_journaled();
7690 CDentry::linkage_t
*srcdnl
= srcdn
->get_linkage();
7691 CDentry::linkage_t
*destdnl
= NULL
;
7692 //CDentry::linkage_t *straydnl = straydn ? straydn->get_linkage() : 0;
7695 if (srcdn
->is_auth() && srcdnl
->is_primary()) {
7696 // set export bounds for CInode::encode_export()
7698 if (srcdnl
->get_inode()->is_dir()) {
7699 srcdnl
->get_inode()->get_dirfrags(bounds
);
7700 for (list
<CDir
*>::iterator p
= bounds
.begin(); p
!= bounds
.end(); ++p
)
7701 (*p
)->state_set(CDir::STATE_EXPORTBOUND
);
7704 map
<client_t
,entity_inst_t
> exported_client_map
;
7706 mdcache
->migrator
->encode_export_inode(srcdnl
->get_inode(), inodebl
,
7707 exported_client_map
);
7709 for (list
<CDir
*>::iterator p
= bounds
.begin(); p
!= bounds
.end(); ++p
)
7710 (*p
)->state_clear(CDir::STATE_EXPORTBOUND
);
7713 ::encode(exported_client_map
, reply
->inode_export
, mds
->mdsmap
->get_up_features());
7714 reply
->inode_export
.claim_append(inodebl
);
7715 reply
->inode_export_v
= srcdnl
->get_inode()->inode
.version
;
7718 // remove mdr auth pin
7719 mdr
->auth_unpin(srcdnl
->get_inode());
7720 mdr
->more()->is_inode_exporter
= true;
7722 if (srcdnl
->get_inode()->is_dirty())
7723 srcdnl
->get_inode()->mark_clean();
7725 dout(10) << " exported srci " << *srcdnl
->get_inode() << dendl
;
7729 _rename_apply(mdr
, srcdn
, destdn
, straydn
);
7731 destdnl
= destdn
->get_linkage();
7734 mds
->balancer
->hit_dir(mdr
->get_mds_stamp(), srcdn
->get_dir(), META_POP_IWR
);
7735 if (destdnl
->get_inode() && destdnl
->get_inode()->is_auth())
7736 mds
->balancer
->hit_inode(mdr
->get_mds_stamp(), destdnl
->get_inode(),
7740 mdr
->slave_request
->put();
7741 mdr
->slave_request
= 0;
7745 mds
->send_message_mds(reply
, mdr
->slave_to_mds
);
7747 assert(mdr
->aborted
);
7748 dout(10) << " abort flag set, finishing" << dendl
;
7749 mdcache
->request_finish(mdr
);
7753 void Server::_commit_slave_rename(MDRequestRef
& mdr
, int r
,
7754 CDentry
*srcdn
, CDentry
*destdn
, CDentry
*straydn
)
7756 dout(10) << "_commit_slave_rename " << *mdr
<< " r=" << r
<< dendl
;
7758 CDentry::linkage_t
*destdnl
= destdn
->get_linkage();
7760 list
<MDSInternalContextBase
*> finished
;
7762 // unfreeze+singleauth inode
7763 // hmm, do i really need to delay this?
7764 if (mdr
->more()->is_inode_exporter
) {
7766 CInode
*in
= destdnl
->get_inode();
7769 // we exported, clear out any xlocks that we moved to another MDS
7770 set
<SimpleLock
*>::iterator i
= mdr
->xlocks
.begin();
7771 while (i
!= mdr
->xlocks
.end()) {
7772 SimpleLock
*lock
= *i
++;
7774 // we only care about xlocks on the exported inode
7775 if (lock
->get_parent() == in
&&
7776 !lock
->is_locallock())
7777 mds
->locker
->xlock_export(lock
, mdr
.get());
7780 map
<client_t
,Capability::Import
> peer_imported
;
7781 bufferlist::iterator bp
= mdr
->more()->inode_import
.begin();
7782 ::decode(peer_imported
, bp
);
7784 dout(10) << " finishing inode export on " << *destdnl
->get_inode() << dendl
;
7785 mdcache
->migrator
->finish_export_inode(destdnl
->get_inode(),
7786 mdr
->get_mds_stamp(),
7787 mdr
->slave_to_mds
, peer_imported
, finished
);
7788 mds
->queue_waiters(finished
); // this includes SINGLEAUTH waiters.
7791 assert(destdnl
->get_inode()->is_frozen_inode());
7792 destdnl
->get_inode()->unfreeze_inode(finished
);
7796 if (mdr
->more()->is_ambiguous_auth
) {
7797 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
7798 mdr
->more()->is_ambiguous_auth
= false;
7801 if (straydn
&& mdr
->more()->slave_update_journaled
) {
7802 CInode
*strayin
= straydn
->get_projected_linkage()->get_inode();
7803 if (strayin
&& !strayin
->snaprealm
)
7804 mdcache
->clear_dirty_bits_for_stray(strayin
);
7807 mds
->queue_waiters(finished
);
7810 if (mdr
->more()->slave_update_journaled
) {
7811 // write a commit to the journal
7812 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_commit", mdr
->reqid
,
7813 mdr
->slave_to_mds
, ESlaveUpdate::OP_COMMIT
,
7814 ESlaveUpdate::RENAME
);
7815 mdlog
->start_entry(le
);
7816 submit_mdlog_entry(le
, new C_MDS_CommittedSlave(this, mdr
), mdr
, __func__
);
7819 _committed_slave(mdr
);
7824 // rollback_bl may be empty if we froze the inode but had to provide an expanded
7825 // witness list from the master, and they failed before we tried prep again.
7826 if (mdr
->more()->rollback_bl
.length()) {
7827 if (mdr
->more()->is_inode_exporter
) {
7828 dout(10) << " reversing inode export of " << *destdnl
->get_inode() << dendl
;
7829 destdnl
->get_inode()->abort_export();
7831 if (mdcache
->is_ambiguous_slave_update(mdr
->reqid
, mdr
->slave_to_mds
)) {
7832 mdcache
->remove_ambiguous_slave_update(mdr
->reqid
, mdr
->slave_to_mds
);
7833 // rollback but preserve the slave request
7834 do_rename_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
, false);
7835 mdr
->more()->rollback_bl
.clear();
7837 do_rename_rollback(mdr
->more()->rollback_bl
, mdr
->slave_to_mds
, mdr
, true);
7839 dout(10) << " rollback_bl empty, not rollback back rename (master failed after getting extra witnesses?)" << dendl
;
7841 if (mdr
->more()->is_ambiguous_auth
) {
7842 if (srcdn
->is_auth())
7843 mdr
->more()->rename_inode
->unfreeze_inode(finished
);
7845 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
7846 mdr
->more()->is_ambiguous_auth
= false;
7848 mds
->queue_waiters(finished
);
7849 mdcache
->request_finish(mdr
);
7854 void _rollback_repair_dir(MutationRef
& mut
, CDir
*dir
, rename_rollback::drec
&r
, utime_t ctime
,
7855 bool isdir
, int linkunlink
, nest_info_t
&rstat
)
7858 pf
= dir
->project_fnode();
7859 mut
->add_projected_fnode(dir
);
7860 pf
->version
= dir
->pre_dirty();
7863 pf
->fragstat
.nsubdirs
+= linkunlink
;
7865 pf
->fragstat
.nfiles
+= linkunlink
;
7868 pf
->rstat
.rbytes
+= linkunlink
* rstat
.rbytes
;
7869 pf
->rstat
.rfiles
+= linkunlink
* rstat
.rfiles
;
7870 pf
->rstat
.rsubdirs
+= linkunlink
* rstat
.rsubdirs
;
7871 pf
->rstat
.rsnaprealms
+= linkunlink
* rstat
.rsnaprealms
;
7873 if (pf
->fragstat
.mtime
== ctime
) {
7874 pf
->fragstat
.mtime
= r
.dirfrag_old_mtime
;
7875 if (pf
->rstat
.rctime
== ctime
)
7876 pf
->rstat
.rctime
= r
.dirfrag_old_rctime
;
7878 mut
->add_updated_lock(&dir
->get_inode()->filelock
);
7879 mut
->add_updated_lock(&dir
->get_inode()->nestlock
);
7882 struct C_MDS_LoggedRenameRollback
: public ServerLogContext
{
7889 C_MDS_LoggedRenameRollback(Server
*s
, MutationRef
& m
, MDRequestRef
& r
,
7890 CDentry
*sd
, version_t pv
, CDentry
*dd
,
7891 CDentry
*st
, bool f
) :
7892 ServerLogContext(s
, r
), mut(m
), srcdn(sd
), srcdnpv(pv
), destdn(dd
),
7893 straydn(st
), finish_mdr(f
) {}
7894 void finish(int r
) override
{
7895 server
->_rename_rollback_finish(mut
, mdr
, srcdn
, srcdnpv
,
7896 destdn
, straydn
, finish_mdr
);
7900 void Server::do_rename_rollback(bufferlist
&rbl
, mds_rank_t master
, MDRequestRef
& mdr
,
7903 rename_rollback rollback
;
7904 bufferlist::iterator p
= rbl
.begin();
7905 ::decode(rollback
, p
);
7907 dout(10) << "do_rename_rollback on " << rollback
.reqid
<< dendl
;
7908 // need to finish this update before sending resolve to claim the subtree
7909 mdcache
->add_rollback(rollback
.reqid
, master
);
7911 MutationRef
mut(new MutationImpl(nullptr, utime_t(), rollback
.reqid
));
7912 mut
->ls
= mds
->mdlog
->get_current_segment();
7914 CDentry
*srcdn
= NULL
;
7915 CDir
*srcdir
= mdcache
->get_dirfrag(rollback
.orig_src
.dirfrag
);
7917 srcdir
= mdcache
->get_dirfrag(rollback
.orig_src
.dirfrag
.ino
, rollback
.orig_src
.dname
);
7919 dout(10) << " srcdir " << *srcdir
<< dendl
;
7920 srcdn
= srcdir
->lookup(rollback
.orig_src
.dname
);
7922 dout(10) << " srcdn " << *srcdn
<< dendl
;
7923 assert(srcdn
->get_linkage()->is_null());
7925 dout(10) << " srcdn not found" << dendl
;
7927 dout(10) << " srcdir not found" << dendl
;
7929 CDentry
*destdn
= NULL
;
7930 CDir
*destdir
= mdcache
->get_dirfrag(rollback
.orig_dest
.dirfrag
);
7932 destdir
= mdcache
->get_dirfrag(rollback
.orig_dest
.dirfrag
.ino
, rollback
.orig_dest
.dname
);
7934 dout(10) << " destdir " << *destdir
<< dendl
;
7935 destdn
= destdir
->lookup(rollback
.orig_dest
.dname
);
7937 dout(10) << " destdn " << *destdn
<< dendl
;
7939 dout(10) << " destdn not found" << dendl
;
7941 dout(10) << " destdir not found" << dendl
;
7944 if (rollback
.orig_src
.ino
) {
7945 in
= mdcache
->get_inode(rollback
.orig_src
.ino
);
7946 if (in
&& in
->is_dir())
7947 assert(srcdn
&& destdn
);
7949 in
= mdcache
->get_inode(rollback
.orig_src
.remote_ino
);
7951 CDir
*straydir
= NULL
;
7952 CDentry
*straydn
= NULL
;
7953 if (rollback
.stray
.dirfrag
.ino
) {
7954 straydir
= mdcache
->get_dirfrag(rollback
.stray
.dirfrag
);
7956 dout(10) << "straydir " << *straydir
<< dendl
;
7957 straydn
= straydir
->lookup(rollback
.stray
.dname
);
7959 dout(10) << " straydn " << *straydn
<< dendl
;
7960 assert(straydn
->get_linkage()->is_primary());
7962 dout(10) << " straydn not found" << dendl
;
7964 dout(10) << "straydir not found" << dendl
;
7967 CInode
*target
= NULL
;
7968 if (rollback
.orig_dest
.ino
) {
7969 target
= mdcache
->get_inode(rollback
.orig_dest
.ino
);
7971 assert(destdn
&& straydn
);
7972 } else if (rollback
.orig_dest
.remote_ino
)
7973 target
= mdcache
->get_inode(rollback
.orig_dest
.remote_ino
);
7975 // can't use is_auth() in the resolve stage
7976 mds_rank_t whoami
= mds
->get_nodeid();
7978 assert(!destdn
|| destdn
->authority().first
!= whoami
);
7979 assert(!straydn
|| straydn
->authority().first
!= whoami
);
7981 bool force_journal_src
= false;
7982 bool force_journal_dest
= false;
7983 if (in
&& in
->is_dir() && srcdn
->authority().first
!= whoami
)
7984 force_journal_src
= _need_force_journal(in
, false);
7985 if (in
&& target
&& target
->is_dir())
7986 force_journal_dest
= _need_force_journal(in
, true);
7988 version_t srcdnpv
= 0;
7991 if (srcdn
->authority().first
== whoami
)
7992 srcdnpv
= srcdn
->pre_dirty();
7993 if (rollback
.orig_src
.ino
) {
7995 srcdn
->push_projected_linkage(in
);
7997 srcdn
->push_projected_linkage(rollback
.orig_src
.remote_ino
,
7998 rollback
.orig_src
.remote_d_type
);
8003 if (in
->authority().first
== whoami
) {
8004 pi
= in
->project_inode();
8005 mut
->add_projected_inode(in
);
8006 pi
->version
= in
->pre_dirty();
8008 pi
= in
->get_projected_inode();
8009 if (pi
->ctime
== rollback
.ctime
)
8010 pi
->ctime
= rollback
.orig_src
.old_ctime
;
8013 if (srcdn
&& srcdn
->authority().first
== whoami
) {
8015 _rollback_repair_dir(mut
, srcdir
, rollback
.orig_src
, rollback
.ctime
,
8016 in
? in
->is_dir() : false, 1, pi
? pi
->accounted_rstat
: blah
);
8021 if (rollback
.orig_dest
.ino
&& target
) {
8022 destdn
->push_projected_linkage(target
);
8023 } else if (rollback
.orig_dest
.remote_ino
) {
8024 destdn
->push_projected_linkage(rollback
.orig_dest
.remote_ino
,
8025 rollback
.orig_dest
.remote_d_type
);
8027 // the dentry will be trimmed soon, it's ok to have wrong linkage
8028 if (rollback
.orig_dest
.ino
)
8029 assert(mds
->is_resolve());
8030 destdn
->push_projected_linkage();
8035 straydn
->push_projected_linkage();
8039 if (target
->authority().first
== whoami
) {
8040 ti
= target
->project_inode();
8041 mut
->add_projected_inode(target
);
8042 ti
->version
= target
->pre_dirty();
8044 ti
= target
->get_projected_inode();
8045 if (ti
->ctime
== rollback
.ctime
)
8046 ti
->ctime
= rollback
.orig_dest
.old_ctime
;
8047 if (MDS_INO_IS_STRAY(rollback
.orig_src
.dirfrag
.ino
)) {
8048 if (MDS_INO_IS_STRAY(rollback
.orig_dest
.dirfrag
.ino
))
8049 assert(!rollback
.orig_dest
.ino
&& !rollback
.orig_dest
.remote_ino
);
8051 assert(rollback
.orig_dest
.remote_ino
&&
8052 rollback
.orig_dest
.remote_ino
== rollback
.orig_src
.ino
);
8058 dout(0) << " srcdn back to " << *srcdn
<< dendl
;
8060 dout(0) << " srci back to " << *in
<< dendl
;
8062 dout(0) << " destdn back to " << *destdn
<< dendl
;
8064 dout(0) << " desti back to " << *target
<< dendl
;
8067 ESlaveUpdate
*le
= new ESlaveUpdate(mdlog
, "slave_rename_rollback", rollback
.reqid
, master
,
8068 ESlaveUpdate::OP_ROLLBACK
, ESlaveUpdate::RENAME
);
8069 mdlog
->start_entry(le
);
8071 if (srcdn
&& (srcdn
->authority().first
== whoami
|| force_journal_src
)) {
8072 le
->commit
.add_dir_context(srcdir
);
8073 if (rollback
.orig_src
.ino
)
8074 le
->commit
.add_primary_dentry(srcdn
, 0, true);
8076 le
->commit
.add_remote_dentry(srcdn
, true);
8079 if (!rollback
.orig_src
.ino
&& // remote linkage
8080 in
&& in
->authority().first
== whoami
) {
8081 le
->commit
.add_dir_context(in
->get_projected_parent_dir());
8082 le
->commit
.add_primary_dentry(in
->get_projected_parent_dn(), in
, true);
8085 if (force_journal_dest
) {
8086 assert(rollback
.orig_dest
.ino
);
8087 le
->commit
.add_dir_context(destdir
);
8088 le
->commit
.add_primary_dentry(destdn
, 0, true);
8091 // slave: no need to journal straydn
8093 if (target
&& target
!= in
&& target
->authority().first
== whoami
) {
8094 assert(rollback
.orig_dest
.remote_ino
);
8095 le
->commit
.add_dir_context(target
->get_projected_parent_dir());
8096 le
->commit
.add_primary_dentry(target
->get_projected_parent_dn(), target
, true);
8099 if (in
&& in
->is_dir() && (srcdn
->authority().first
== whoami
|| force_journal_src
)) {
8100 dout(10) << " noting renamed dir ino " << in
->ino() << " in metablob" << dendl
;
8101 le
->commit
.renamed_dirino
= in
->ino();
8102 if (srcdn
->authority().first
== whoami
) {
8104 in
->get_dirfrags(ls
);
8105 for (list
<CDir
*>::iterator p
= ls
.begin(); p
!= ls
.end(); ++p
) {
8107 if (!dir
->is_auth())
8108 le
->commit
.renamed_dir_frags
.push_back(dir
->get_frag());
8110 dout(10) << " noting renamed dir open frags " << le
->commit
.renamed_dir_frags
<< dendl
;
8112 } else if (force_journal_dest
) {
8113 dout(10) << " noting rename target ino " << target
->ino() << " in metablob" << dendl
;
8114 le
->commit
.renamed_dirino
= target
->ino();
8117 if (target
&& target
->is_dir()) {
8119 mdcache
->project_subtree_rename(target
, straydir
, destdir
);
8122 if (in
&& in
->is_dir()) {
8124 mdcache
->project_subtree_rename(in
, destdir
, srcdir
);
8127 if (mdr
&& !mdr
->more()->slave_update_journaled
) {
8128 assert(le
->commit
.empty());
8129 mdlog
->cancel_entry(le
);
8131 _rename_rollback_finish(mut
, mdr
, srcdn
, srcdnpv
, destdn
, straydn
, finish_mdr
);
8133 assert(!le
->commit
.empty());
8135 mdr
->more()->slave_update_journaled
= false;
8136 MDSLogContextBase
*fin
= new C_MDS_LoggedRenameRollback(this, mut
, mdr
, srcdn
, srcdnpv
,
8137 destdn
, straydn
, finish_mdr
);
8138 submit_mdlog_entry(le
, fin
, mdr
, __func__
);
8143 void Server::_rename_rollback_finish(MutationRef
& mut
, MDRequestRef
& mdr
, CDentry
*srcdn
,
8144 version_t srcdnpv
, CDentry
*destdn
,
8145 CDentry
*straydn
, bool finish_mdr
)
8147 dout(10) << "_rename_rollback_finish " << mut
->reqid
<< dendl
;
8150 straydn
->get_dir()->unlink_inode(straydn
);
8151 straydn
->pop_projected_linkage();
8154 destdn
->get_dir()->unlink_inode(destdn
);
8155 destdn
->pop_projected_linkage();
8158 srcdn
->pop_projected_linkage();
8159 if (srcdn
->authority().first
== mds
->get_nodeid())
8160 srcdn
->mark_dirty(srcdnpv
, mut
->ls
);
8165 if (srcdn
&& srcdn
->get_linkage()->is_primary()) {
8166 CInode
*in
= srcdn
->get_linkage()->get_inode();
8167 if (srcdn
->authority().first
== mds
->get_nodeid())
8168 in
->state_set(CInode::STATE_AUTH
);
8169 // update subtree map?
8170 if (in
&& in
->is_dir()) {
8172 mdcache
->adjust_subtree_after_rename(in
, destdn
->get_dir(), true);
8177 CInode
*oldin
= destdn
->get_linkage()->get_inode();
8178 // update subtree map?
8179 if (oldin
&& oldin
->is_dir()) {
8181 mdcache
->adjust_subtree_after_rename(oldin
, straydn
->get_dir(), true);
8185 if (mds
->is_resolve()) {
8188 root
= mdcache
->get_subtree_root(straydn
->get_dir());
8190 root
= mdcache
->get_subtree_root(destdn
->get_dir());
8192 mdcache
->try_trim_non_auth_subtree(root
);
8196 list
<MDSInternalContextBase
*> finished
;
8197 if (mdr
->more()->is_ambiguous_auth
) {
8198 if (srcdn
->is_auth())
8199 mdr
->more()->rename_inode
->unfreeze_inode(finished
);
8201 mdr
->more()->rename_inode
->clear_ambiguous_auth(finished
);
8202 mdr
->more()->is_ambiguous_auth
= false;
8204 mds
->queue_waiters(finished
);
8205 if (finish_mdr
|| mdr
->aborted
)
8206 mdcache
->request_finish(mdr
);
8208 mdr
->more()->slave_rolling_back
= false;
8211 mdcache
->finish_rollback(mut
->reqid
);
8216 /* This function DOES put the passed message before returning*/
8217 void Server::handle_slave_rename_prep_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
8219 dout(10) << "handle_slave_rename_prep_ack " << *mdr
8220 << " witnessed by " << ack
->get_source()
8221 << " " << *ack
<< dendl
;
8222 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
8225 mdr
->more()->slaves
.insert(from
);
8226 if (mdr
->more()->srcdn_auth_mds
== from
&&
8227 mdr
->more()->is_remote_frozen_authpin
&&
8228 !mdr
->more()->is_ambiguous_auth
) {
8229 mdr
->set_ambiguous_auth(mdr
->more()->rename_inode
);
8232 // witnessed? or add extra witnesses?
8233 assert(mdr
->more()->witnessed
.count(from
) == 0);
8234 if (ack
->is_interrupted()) {
8235 dout(10) << " slave request interrupted, noop" << dendl
;
8236 } else if (ack
->witnesses
.empty()) {
8237 mdr
->more()->witnessed
.insert(from
);
8238 if (!ack
->is_not_journaled())
8239 mdr
->more()->has_journaled_slaves
= true;
8241 dout(10) << " extra witnesses (srcdn replicas) are " << ack
->witnesses
<< dendl
;
8242 mdr
->more()->extra_witnesses
.swap(ack
->witnesses
);
8243 mdr
->more()->extra_witnesses
.erase(mds
->get_nodeid()); // not me!
8247 if (ack
->inode_export
.length()) {
8248 dout(10) << " got srci import" << dendl
;
8249 mdr
->more()->inode_import
.claim(ack
->inode_export
);
8250 mdr
->more()->inode_import_v
= ack
->inode_export_v
;
8253 // remove from waiting list
8254 assert(mdr
->more()->waiting_on_slave
.count(from
));
8255 mdr
->more()->waiting_on_slave
.erase(from
);
8257 if (mdr
->more()->waiting_on_slave
.empty())
8258 dispatch_client_request(mdr
); // go again!
8260 dout(10) << "still waiting on slaves " << mdr
->more()->waiting_on_slave
<< dendl
;
8263 void Server::handle_slave_rename_notify_ack(MDRequestRef
& mdr
, MMDSSlaveRequest
*ack
)
8265 dout(10) << "handle_slave_rename_notify_ack " << *mdr
<< " from mds."
8266 << ack
->get_source() << dendl
;
8267 assert(mdr
->is_slave());
8268 mds_rank_t from
= mds_rank_t(ack
->get_source().num());
8270 if (mdr
->more()->waiting_on_slave
.count(from
)) {
8271 mdr
->more()->waiting_on_slave
.erase(from
);
8273 if (mdr
->more()->waiting_on_slave
.empty()) {
8274 if (mdr
->slave_request
)
8275 dispatch_slave_request(mdr
);
8277 dout(10) << " still waiting for rename notify acks from "
8278 << mdr
->more()->waiting_on_slave
<< dendl
;
8282 void Server::_slave_rename_sessions_flushed(MDRequestRef
& mdr
)
8284 dout(10) << "_slave_rename_sessions_flushed " << *mdr
<< dendl
;
8286 if (mdr
->more()->waiting_on_slave
.count(MDS_RANK_NONE
)) {
8287 mdr
->more()->waiting_on_slave
.erase(MDS_RANK_NONE
);
8289 if (mdr
->more()->waiting_on_slave
.empty()) {
8290 if (mdr
->slave_request
)
8291 dispatch_slave_request(mdr
);
8293 dout(10) << " still waiting for rename notify acks from "
8294 << mdr
->more()->waiting_on_slave
<< dendl
;
8299 /* This function takes responsibility for the passed mdr*/
8300 void Server::handle_client_lssnap(MDRequestRef
& mdr
)
8302 MClientRequest
*req
= mdr
->client_request
;
8305 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8306 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8307 respond_to_request(mdr
, -ESTALE
);
8310 if (!diri
->is_auth()) {
8311 mdcache
->request_forward(mdr
, diri
->authority().first
);
8314 if (!diri
->is_dir()) {
8315 respond_to_request(mdr
, -ENOTDIR
);
8318 dout(10) << "lssnap on " << *diri
<< dendl
;
8321 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8322 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8323 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8326 if (!check_access(mdr
, diri
, MAY_READ
))
8329 SnapRealm
*realm
= diri
->find_snaprealm();
8330 map
<snapid_t
,SnapInfo
*> infomap
;
8331 realm
->get_snap_info(infomap
, diri
->get_oldest_snap());
8333 unsigned max_entries
= req
->head
.args
.readdir
.max_entries
;
8335 max_entries
= infomap
.size();
8336 int max_bytes
= req
->head
.args
.readdir
.max_bytes
;
8338 // make sure at least one item can be encoded
8339 max_bytes
= (512 << 10) + g_conf
->mds_max_xattr_pairs_size
;
8341 __u64 last_snapid
= 0;
8342 string offset_str
= req
->get_path2();
8343 if (!offset_str
.empty())
8344 last_snapid
= realm
->resolve_snapname(offset_str
, diri
->ino());
8347 encode_empty_dirstat(dirbl
);
8349 max_bytes
-= dirbl
.length() - sizeof(__u32
) + sizeof(__u8
) * 2;
8353 map
<snapid_t
,SnapInfo
*>::iterator p
= infomap
.upper_bound(last_snapid
);
8354 for (; p
!= infomap
.end() && num
< max_entries
; ++p
) {
8355 dout(10) << p
->first
<< " -> " << *p
->second
<< dendl
;
8359 if (p
->second
->ino
== diri
->ino())
8360 snap_name
= p
->second
->name
;
8362 snap_name
= p
->second
->get_long_name();
8364 unsigned start_len
= dnbl
.length();
8365 if (int(start_len
+ snap_name
.length() + sizeof(__u32
) + sizeof(LeaseStat
)) > max_bytes
)
8368 ::encode(snap_name
, dnbl
);
8369 encode_infinite_lease(dnbl
);
8371 int r
= diri
->encode_inodestat(dnbl
, mdr
->session
, realm
, p
->first
, max_bytes
- (int)dnbl
.length());
8374 keep
.substr_of(dnbl
, 0, start_len
);
8381 ::encode(num
, dirbl
);
8383 if (p
== infomap
.end()) {
8384 flags
= CEPH_READDIR_FRAG_END
;
8385 if (last_snapid
== 0)
8386 flags
|= CEPH_READDIR_FRAG_COMPLETE
;
8388 ::encode(flags
, dirbl
);
8389 dirbl
.claim_append(dnbl
);
8391 mdr
->reply_extra_bl
= dirbl
;
8393 respond_to_request(mdr
, 0);
8399 struct C_MDS_mksnap_finish
: public ServerLogContext
{
8402 C_MDS_mksnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, SnapInfo
&i
) :
8403 ServerLogContext(s
, r
), diri(di
), info(i
) {}
8404 void finish(int r
) override
{
8405 server
->_mksnap_finish(mdr
, diri
, info
);
8409 /* This function takes responsibility for the passed mdr*/
8410 void Server::handle_client_mksnap(MDRequestRef
& mdr
)
8412 if (!mds
->mdsmap
->allows_snaps()) {
8413 // you can't make snapshots until you set an option right now
8414 respond_to_request(mdr
, -EPERM
);
8418 MClientRequest
*req
= mdr
->client_request
;
8419 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8420 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8421 respond_to_request(mdr
, -ESTALE
);
8425 if (!diri
->is_auth()) { // fw to auth?
8426 mdcache
->request_forward(mdr
, diri
->authority().first
);
8431 if (!diri
->is_dir()) {
8432 respond_to_request(mdr
, -ENOTDIR
);
8435 if (diri
->is_system() && !diri
->is_root()) {
8436 // no snaps in system dirs (root is ok)
8437 respond_to_request(mdr
, -EPERM
);
8441 const string
&snapname
= req
->get_filepath().last_dentry();
8443 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
) {
8444 dout(20) << "mksnap " << snapname
<< " on " << *diri
<< " denied to uid " << mdr
->client_request
->get_caller_uid() << dendl
;
8445 respond_to_request(mdr
, -EPERM
);
8449 dout(10) << "mksnap " << snapname
<< " on " << *diri
<< dendl
;
8452 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8454 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8455 rdlocks
.erase(&diri
->snaplock
);
8456 xlocks
.insert(&diri
->snaplock
);
8458 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8461 if (!check_access(mdr
, diri
, MAY_WRITE
))
8464 // make sure name is unique
8465 if (diri
->snaprealm
&&
8466 diri
->snaprealm
->exists(snapname
)) {
8467 respond_to_request(mdr
, -EEXIST
);
8470 if (snapname
.length() == 0 ||
8471 snapname
[0] == '_') {
8472 respond_to_request(mdr
, -EINVAL
);
8476 // allocate a snapid
8477 if (!mdr
->more()->stid
) {
8479 mds
->snapclient
->prepare_create(diri
->ino(), snapname
,
8480 mdr
->get_mds_stamp(),
8481 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8482 new C_MDS_RetryRequest(mdcache
, mdr
));
8486 version_t stid
= mdr
->more()->stid
;
8488 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8489 ::decode(snapid
, p
);
8490 dout(10) << " stid " << stid
<< " snapid " << snapid
<< dendl
;
8494 info
.ino
= diri
->ino();
8495 info
.snapid
= snapid
;
8496 info
.name
= snapname
;
8497 info
.stamp
= mdr
->get_op_stamp();
8499 inode_t
*pi
= diri
->project_inode();
8500 pi
->ctime
= info
.stamp
;
8501 pi
->version
= diri
->pre_dirty();
8503 // project the snaprealm
8504 sr_t
*newsnap
= diri
->project_snaprealm(snapid
);
8505 newsnap
->snaps
[snapid
] = info
;
8506 newsnap
->seq
= snapid
;
8507 newsnap
->last_created
= snapid
;
8509 // journal the inode changes
8510 mdr
->ls
= mdlog
->get_current_segment();
8511 EUpdate
*le
= new EUpdate(mdlog
, "mksnap");
8512 mdlog
->start_entry(le
);
8514 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8515 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8516 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8517 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8519 // journal the snaprealm changes
8520 submit_mdlog_entry(le
, new C_MDS_mksnap_finish(this, mdr
, diri
, info
),
8525 void Server::_mksnap_finish(MDRequestRef
& mdr
, CInode
*diri
, SnapInfo
&info
)
8527 dout(10) << "_mksnap_finish " << *mdr
<< " " << info
<< dendl
;
8529 int op
= (diri
->snaprealm
? CEPH_SNAP_OP_CREATE
: CEPH_SNAP_OP_SPLIT
);
8531 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8534 mds
->snapclient
->commit(mdr
->more()->stid
, mdr
->ls
);
8537 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8539 mdcache
->do_realm_invalidate_and_update_notify(diri
, op
);
8543 mdr
->snapid
= info
.snapid
;
8545 respond_to_request(mdr
, 0);
8551 struct C_MDS_rmsnap_finish
: public ServerLogContext
{
8554 C_MDS_rmsnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, snapid_t sn
) :
8555 ServerLogContext(s
, r
), diri(di
), snapid(sn
) {}
8556 void finish(int r
) override
{
8557 server
->_rmsnap_finish(mdr
, diri
, snapid
);
8561 /* This function takes responsibility for the passed mdr*/
8562 void Server::handle_client_rmsnap(MDRequestRef
& mdr
)
8564 MClientRequest
*req
= mdr
->client_request
;
8566 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8567 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8568 respond_to_request(mdr
, -ESTALE
);
8571 if (!diri
->is_auth()) { // fw to auth?
8572 mdcache
->request_forward(mdr
, diri
->authority().first
);
8575 if (!diri
->is_dir()) {
8576 respond_to_request(mdr
, -ENOTDIR
);
8580 const string
&snapname
= req
->get_filepath().last_dentry();
8582 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
) {
8583 dout(20) << "rmsnap " << snapname
<< " on " << *diri
<< " denied to uid " << mdr
->client_request
->get_caller_uid() << dendl
;
8584 respond_to_request(mdr
, -EPERM
);
8588 dout(10) << "rmsnap " << snapname
<< " on " << *diri
<< dendl
;
8591 if (snapname
.length() == 0 || snapname
[0] == '_') {
8592 respond_to_request(mdr
, -EINVAL
); // can't prune a parent snap, currently.
8595 if (!diri
->snaprealm
|| !diri
->snaprealm
->exists(snapname
)) {
8596 respond_to_request(mdr
, -ENOENT
);
8599 snapid_t snapid
= diri
->snaprealm
->resolve_snapname(snapname
, diri
->ino());
8600 dout(10) << " snapname " << snapname
<< " is " << snapid
<< dendl
;
8602 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8603 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8604 rdlocks
.erase(&diri
->snaplock
);
8605 xlocks
.insert(&diri
->snaplock
);
8607 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8610 if (!check_access(mdr
, diri
, MAY_WRITE
))
8614 if (!mdr
->more()->stid
) {
8615 mds
->snapclient
->prepare_destroy(diri
->ino(), snapid
,
8616 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8617 new C_MDS_RetryRequest(mdcache
, mdr
));
8620 version_t stid
= mdr
->more()->stid
;
8621 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8624 dout(10) << " stid is " << stid
<< ", seq is " << seq
<< dendl
;
8627 inode_t
*pi
= diri
->project_inode();
8628 pi
->version
= diri
->pre_dirty();
8629 pi
->ctime
= mdr
->get_op_stamp();
8631 mdr
->ls
= mdlog
->get_current_segment();
8632 EUpdate
*le
= new EUpdate(mdlog
, "rmsnap");
8633 mdlog
->start_entry(le
);
8635 // project the snaprealm
8636 sr_t
*newnode
= diri
->project_snaprealm();
8637 newnode
->snaps
.erase(snapid
);
8639 newnode
->last_destroyed
= seq
;
8641 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8642 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8643 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8644 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8646 submit_mdlog_entry(le
, new C_MDS_rmsnap_finish(this, mdr
, diri
, snapid
),
8651 void Server::_rmsnap_finish(MDRequestRef
& mdr
, CInode
*diri
, snapid_t snapid
)
8653 dout(10) << "_rmsnap_finish " << *mdr
<< " " << snapid
<< dendl
;
8654 snapid_t stid
= mdr
->more()->stid
;
8655 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8659 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8662 mds
->snapclient
->commit(stid
, mdr
->ls
);
8664 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8666 mdcache
->do_realm_invalidate_and_update_notify(diri
, CEPH_SNAP_OP_DESTROY
);
8670 respond_to_request(mdr
, 0);
8672 // purge snapshot data
8673 if (diri
->snaprealm
->have_past_parents_open())
8674 diri
->purge_stale_snap_data(diri
->snaprealm
->get_snaps());
8677 struct C_MDS_renamesnap_finish
: public ServerLogContext
{
8680 C_MDS_renamesnap_finish(Server
*s
, MDRequestRef
& r
, CInode
*di
, snapid_t sn
) :
8681 ServerLogContext(s
, r
), diri(di
), snapid(sn
) {}
8682 void finish(int r
) override
{
8683 server
->_renamesnap_finish(mdr
, diri
, snapid
);
8687 /* This function takes responsibility for the passed mdr*/
8688 void Server::handle_client_renamesnap(MDRequestRef
& mdr
)
8690 MClientRequest
*req
= mdr
->client_request
;
8691 if (req
->get_filepath().get_ino() != req
->get_filepath2().get_ino()) {
8692 respond_to_request(mdr
, -EINVAL
);
8696 CInode
*diri
= mdcache
->get_inode(req
->get_filepath().get_ino());
8697 if (!diri
|| diri
->state_test(CInode::STATE_PURGING
)) {
8698 respond_to_request(mdr
, -ESTALE
);
8702 if (!diri
->is_auth()) { // fw to auth?
8703 mdcache
->request_forward(mdr
, diri
->authority().first
);
8707 if (!diri
->is_dir()) { // dir only
8708 respond_to_request(mdr
, -ENOTDIR
);
8712 if (mdr
->client_request
->get_caller_uid() < g_conf
->mds_snap_min_uid
||
8713 mdr
->client_request
->get_caller_uid() > g_conf
->mds_snap_max_uid
) {
8714 respond_to_request(mdr
, -EPERM
);
8718 const string
&dstname
= req
->get_filepath().last_dentry();
8719 const string
&srcname
= req
->get_filepath2().last_dentry();
8720 dout(10) << "renamesnap " << srcname
<< "->" << dstname
<< " on " << *diri
<< dendl
;
8722 if (srcname
.length() == 0 || srcname
[0] == '_') {
8723 respond_to_request(mdr
, -EINVAL
); // can't rename a parent snap.
8726 if (!diri
->snaprealm
|| !diri
->snaprealm
->exists(srcname
)) {
8727 respond_to_request(mdr
, -ENOENT
);
8730 if (dstname
.length() == 0 || dstname
[0] == '_') {
8731 respond_to_request(mdr
, -EINVAL
);
8734 if (diri
->snaprealm
->exists(dstname
)) {
8735 respond_to_request(mdr
, -EEXIST
);
8739 snapid_t snapid
= diri
->snaprealm
->resolve_snapname(srcname
, diri
->ino());
8740 dout(10) << " snapname " << srcname
<< " is " << snapid
<< dendl
;
8743 set
<SimpleLock
*> rdlocks
, wrlocks
, xlocks
;
8745 mds
->locker
->include_snap_rdlocks(rdlocks
, diri
);
8746 rdlocks
.erase(&diri
->snaplock
);
8747 xlocks
.insert(&diri
->snaplock
);
8749 if (!mds
->locker
->acquire_locks(mdr
, rdlocks
, wrlocks
, xlocks
))
8752 if (!check_access(mdr
, diri
, MAY_WRITE
))
8756 if (!mdr
->more()->stid
) {
8757 mds
->snapclient
->prepare_update(diri
->ino(), snapid
, dstname
, utime_t(),
8758 &mdr
->more()->stid
, &mdr
->more()->snapidbl
,
8759 new C_MDS_RetryRequest(mdcache
, mdr
));
8763 version_t stid
= mdr
->more()->stid
;
8764 bufferlist::iterator p
= mdr
->more()->snapidbl
.begin();
8767 dout(10) << " stid is " << stid
<< ", seq is " << seq
<< dendl
;
8770 inode_t
*pi
= diri
->project_inode();
8771 pi
->ctime
= mdr
->get_op_stamp();
8772 pi
->version
= diri
->pre_dirty();
8774 // project the snaprealm
8775 sr_t
*newsnap
= diri
->project_snaprealm();
8776 assert(newsnap
->snaps
.count(snapid
));
8777 newsnap
->snaps
[snapid
].name
= dstname
;
8779 // journal the inode changes
8780 mdr
->ls
= mdlog
->get_current_segment();
8781 EUpdate
*le
= new EUpdate(mdlog
, "renamesnap");
8782 mdlog
->start_entry(le
);
8784 le
->metablob
.add_client_req(req
->get_reqid(), req
->get_oldest_client_tid());
8785 le
->metablob
.add_table_transaction(TABLE_SNAP
, stid
);
8786 mdcache
->predirty_journal_parents(mdr
, &le
->metablob
, diri
, 0, PREDIRTY_PRIMARY
, false);
8787 mdcache
->journal_dirty_inode(mdr
.get(), &le
->metablob
, diri
);
8789 // journal the snaprealm changes
8790 submit_mdlog_entry(le
, new C_MDS_renamesnap_finish(this, mdr
, diri
, snapid
),
8795 void Server::_renamesnap_finish(MDRequestRef
& mdr
, CInode
*diri
, snapid_t snapid
)
8797 dout(10) << "_renamesnap_finish " << *mdr
<< " " << snapid
<< dendl
;
8799 diri
->pop_and_dirty_projected_inode(mdr
->ls
);
8802 mds
->snapclient
->commit(mdr
->more()->stid
, mdr
->ls
);
8804 dout(10) << "snaprealm now " << *diri
->snaprealm
<< dendl
;
8806 mdcache
->do_realm_invalidate_and_update_notify(diri
, CEPH_SNAP_OP_UPDATE
, true);
8811 mdr
->snapid
= snapid
;
8812 respond_to_request(mdr
, 0);
8816 * Return true if server is in state RECONNECT and this
8817 * client has not yet reconnected.
8819 bool Server::waiting_for_reconnect(client_t c
) const
8821 return client_reconnect_gather
.count(c
) > 0;
8824 void Server::dump_reconnect_status(Formatter
*f
) const
8826 f
->open_object_section("reconnect_status");
8827 f
->dump_stream("client_reconnect_gather") << client_reconnect_gather
;