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/utility/string_view.hpp>
17 #include "include/types.h"
29 #include "LogSegment.h"
31 #include "common/bloom_filter.hpp"
32 #include "include/Context.h"
33 #include "common/Clock.h"
35 #include "osdc/Objecter.h"
37 #include "common/config.h"
38 #include "include/assert.h"
39 #include "include/compat.h"
41 #define dout_context g_ceph_context
42 #define dout_subsys ceph_subsys_mds
44 #define dout_prefix *_dout << "mds." << cache->mds->get_nodeid() << ".cache.dir(" << this->dirfrag() << ") "
46 int CDir::num_frozen_trees
= 0;
47 int CDir::num_freezing_trees
= 0;
49 class CDirContext
: public MDSInternalContextBase
53 MDSRank
* get_mds() override
{return dir
->cache
->mds
;}
56 explicit CDirContext(CDir
*d
) : dir(d
) {
62 class CDirIOContext
: public MDSIOContextBase
66 MDSRank
* get_mds() override
{return dir
->cache
->mds
;}
69 explicit CDirIOContext(CDir
*d
) : dir(d
) {
76 //int cdir_pins[CDIR_NUM_PINS] = { 0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
79 ostream
& operator<<(ostream
& out
, const CDir
& dir
)
81 out
<< "[dir " << dir
.dirfrag() << " " << dir
.get_path() << "/"
82 << " [" << dir
.first
<< ",head]";
85 if (dir
.is_replicated())
86 out
<< dir
.get_replicas();
88 if (dir
.is_projected())
89 out
<< " pv=" << dir
.get_projected_version();
90 out
<< " v=" << dir
.get_version();
91 out
<< " cv=" << dir
.get_committing_version();
92 out
<< "/" << dir
.get_committed_version();
94 mds_authority_t a
= dir
.authority();
95 out
<< " rep@" << a
.first
;
96 if (a
.second
!= CDIR_AUTH_UNKNOWN
)
97 out
<< "," << a
.second
;
98 out
<< "." << dir
.get_replica_nonce();
101 if (dir
.is_rep()) out
<< " REP";
103 if (dir
.get_dir_auth() != CDIR_AUTH_DEFAULT
) {
104 if (dir
.get_dir_auth().second
== CDIR_AUTH_UNKNOWN
)
105 out
<< " dir_auth=" << dir
.get_dir_auth().first
;
107 out
<< " dir_auth=" << dir
.get_dir_auth();
110 if (dir
.get_cum_auth_pins())
111 out
<< " ap=" << dir
.get_auth_pins()
112 << "+" << dir
.get_dir_auth_pins()
113 << "+" << dir
.get_nested_auth_pins();
115 out
<< " state=" << dir
.get_state();
116 if (dir
.state_test(CDir::STATE_COMPLETE
)) out
<< "|complete";
117 if (dir
.state_test(CDir::STATE_FREEZINGTREE
)) out
<< "|freezingtree";
118 if (dir
.state_test(CDir::STATE_FROZENTREE
)) out
<< "|frozentree";
119 if (dir
.state_test(CDir::STATE_AUXSUBTREE
)) out
<< "|auxsubtree";
120 //if (dir.state_test(CDir::STATE_FROZENTREELEAF)) out << "|frozentreeleaf";
121 if (dir
.state_test(CDir::STATE_FROZENDIR
)) out
<< "|frozendir";
122 if (dir
.state_test(CDir::STATE_FREEZINGDIR
)) out
<< "|freezingdir";
123 if (dir
.state_test(CDir::STATE_EXPORTBOUND
)) out
<< "|exportbound";
124 if (dir
.state_test(CDir::STATE_IMPORTBOUND
)) out
<< "|importbound";
125 if (dir
.state_test(CDir::STATE_BADFRAG
)) out
<< "|badfrag";
126 if (dir
.state_test(CDir::STATE_FRAGMENTING
)) out
<< "|fragmenting";
129 out
<< " " << dir
.fnode
.fragstat
;
130 if (!(dir
.fnode
.fragstat
== dir
.fnode
.accounted_fragstat
))
131 out
<< "/" << dir
.fnode
.accounted_fragstat
;
132 if (g_conf
->mds_debug_scatterstat
&& dir
.is_projected()) {
133 const fnode_t
*pf
= dir
.get_projected_fnode();
134 out
<< "->" << pf
->fragstat
;
135 if (!(pf
->fragstat
== pf
->accounted_fragstat
))
136 out
<< "/" << pf
->accounted_fragstat
;
140 out
<< " " << dir
.fnode
.rstat
;
141 if (!(dir
.fnode
.rstat
== dir
.fnode
.accounted_rstat
))
142 out
<< "/" << dir
.fnode
.accounted_rstat
;
143 if (g_conf
->mds_debug_scatterstat
&& dir
.is_projected()) {
144 const fnode_t
*pf
= dir
.get_projected_fnode();
145 out
<< "->" << pf
->rstat
;
146 if (!(pf
->rstat
== pf
->accounted_rstat
))
147 out
<< "/" << pf
->accounted_rstat
;
150 out
<< " hs=" << dir
.get_num_head_items() << "+" << dir
.get_num_head_null();
151 out
<< ",ss=" << dir
.get_num_snap_items() << "+" << dir
.get_num_snap_null();
152 if (dir
.get_num_dirty())
153 out
<< " dirty=" << dir
.get_num_dirty();
155 if (dir
.get_num_ref()) {
157 dir
.print_pin_set(out
);
165 void CDir::print(ostream
& out
)
173 ostream
& CDir::print_db_line_prefix(ostream
& out
)
175 return out
<< ceph_clock_now() << " mds." << cache
->mds
->get_nodeid() << ".cache.dir(" << this->dirfrag() << ") ";
180 // -------------------------------------------------------------------
183 CDir::CDir(CInode
*in
, frag_t fg
, MDCache
*mdcache
, bool auth
) :
184 cache(mdcache
), inode(in
), frag(fg
),
186 dirty_rstat_inodes(member_offset(CInode
, dirty_rstat_item
)),
187 projected_version(0),
188 dirty_dentries(member_offset(CDentry
, item_dir_dirty
)),
189 item_dirty(this), item_new(this),
190 num_head_items(0), num_head_null(0),
191 num_snap_items(0), num_snap_null(0),
192 num_dirty(0), committing_version(0), committed_version(0),
193 dir_auth_pins(0), request_pins(0),
195 pop_me(ceph_clock_now()),
196 pop_nested(ceph_clock_now()),
197 pop_auth_subtree(ceph_clock_now()),
198 pop_auth_subtree_nested(ceph_clock_now()),
199 pop_lru_subdirs(member_offset(CInode
, item_pop_lru
)),
200 num_dentries_nested(0), num_dentries_auth_subtree(0),
201 num_dentries_auth_subtree_nested(0),
202 dir_auth(CDIR_AUTH_DEFAULT
)
204 memset(&fnode
, 0, sizeof(fnode
));
207 assert(in
->is_dir());
208 if (auth
) state_set(STATE_AUTH
);
212 * Check the recursive statistics on size for consistency.
213 * If mds_debug_scatterstat is enabled, assert for correctness,
214 * otherwise just print out the mismatch and continue.
216 bool CDir::check_rstats(bool scrub
)
218 if (!g_conf
->mds_debug_scatterstat
&& !scrub
)
221 dout(25) << "check_rstats on " << this << dendl
;
222 if (!is_complete() || !is_auth() || is_frozen()) {
224 dout(10) << "check_rstats bailing out -- incomplete or non-auth or frozen dir!" << dendl
;
228 frag_info_t frag_info
;
229 nest_info_t nest_info
;
230 for (auto i
= items
.begin(); i
!= items
.end(); ++i
) {
231 if (i
->second
->last
!= CEPH_NOSNAP
)
233 CDentry::linkage_t
*dnl
= i
->second
->get_linkage();
234 if (dnl
->is_primary()) {
235 CInode
*in
= dnl
->get_inode();
236 nest_info
.add(in
->inode
.accounted_rstat
);
238 frag_info
.nsubdirs
++;
241 } else if (dnl
->is_remote())
247 if(!frag_info
.same_sums(fnode
.fragstat
)) {
248 dout(1) << "mismatch between head items and fnode.fragstat! printing dentries" << dendl
;
249 dout(1) << "get_num_head_items() = " << get_num_head_items()
250 << "; fnode.fragstat.nfiles=" << fnode
.fragstat
.nfiles
251 << " fnode.fragstat.nsubdirs=" << fnode
.fragstat
.nsubdirs
<< dendl
;
254 dout(20) << "get_num_head_items() = " << get_num_head_items()
255 << "; fnode.fragstat.nfiles=" << fnode
.fragstat
.nfiles
256 << " fnode.fragstat.nsubdirs=" << fnode
.fragstat
.nsubdirs
<< dendl
;
260 if (!nest_info
.same_sums(fnode
.rstat
)) {
261 dout(1) << "mismatch between child accounted_rstats and my rstats!" << dendl
;
262 dout(1) << "total of child dentrys: " << nest_info
<< dendl
;
263 dout(1) << "my rstats: " << fnode
.rstat
<< dendl
;
266 dout(20) << "total of child dentrys: " << nest_info
<< dendl
;
267 dout(20) << "my rstats: " << fnode
.rstat
<< dendl
;
272 for (auto i
= items
.begin(); i
!= items
.end(); ++i
) {
273 CDentry
*dn
= i
->second
;
274 if (dn
->get_linkage()->is_primary()) {
275 CInode
*in
= dn
->get_linkage()->inode
;
276 dout(1) << *dn
<< " rstat " << in
->inode
.accounted_rstat
<< dendl
;
278 dout(1) << *dn
<< dendl
;
282 assert(frag_info
.nfiles
== fnode
.fragstat
.nfiles
);
283 assert(frag_info
.nsubdirs
== fnode
.fragstat
.nsubdirs
);
284 assert(nest_info
.rbytes
== fnode
.rstat
.rbytes
);
285 assert(nest_info
.rfiles
== fnode
.rstat
.rfiles
);
286 assert(nest_info
.rsubdirs
== fnode
.rstat
.rsubdirs
);
289 dout(10) << "check_rstats complete on " << this << dendl
;
293 CDentry
*CDir::lookup(boost::string_view name
, snapid_t snap
)
295 dout(20) << "lookup (" << snap
<< ", '" << name
<< "')" << dendl
;
296 auto iter
= items
.lower_bound(dentry_key_t(snap
, name
, inode
->hash_dentry_name(name
)));
297 if (iter
== items
.end())
299 if (iter
->second
->get_name() == name
&&
300 iter
->second
->first
<= snap
&&
301 iter
->second
->last
>= snap
) {
302 dout(20) << " hit -> " << iter
->first
<< dendl
;
305 dout(20) << " miss -> " << iter
->first
<< dendl
;
309 CDentry
*CDir::lookup_exact_snap(boost::string_view name
, snapid_t last
) {
310 auto p
= items
.find(dentry_key_t(last
, name
, inode
->hash_dentry_name(name
)));
311 if (p
== items
.end())
320 CDentry
* CDir::add_null_dentry(boost::string_view dname
,
321 snapid_t first
, snapid_t last
)
324 assert(lookup_exact_snap(dname
, last
) == 0);
327 CDentry
* dn
= new CDentry(dname
, inode
->hash_dentry_name(dname
), first
, last
);
329 dn
->state_set(CDentry::STATE_AUTH
);
331 cache
->bottom_lru
.lru_insert_mid(dn
);
332 dn
->state_set(CDentry::STATE_BOTTOMLRU
);
335 dn
->version
= get_projected_version();
338 assert(items
.count(dn
->key()) == 0);
339 //assert(null_items.count(dn->get_name()) == 0);
341 items
[dn
->key()] = dn
;
342 if (last
== CEPH_NOSNAP
)
347 if (state_test(CDir::STATE_DNPINNEDFRAG
)) {
348 dn
->get(CDentry::PIN_FRAGMENTING
);
349 dn
->state_set(CDentry::STATE_FRAGMENTING
);
352 dout(12) << "add_null_dentry " << *dn
<< dendl
;
355 if (get_num_any() == 1)
358 assert(get_num_any() == items
.size());
363 CDentry
* CDir::add_primary_dentry(boost::string_view dname
, CInode
*in
,
364 snapid_t first
, snapid_t last
)
367 assert(lookup_exact_snap(dname
, last
) == 0);
370 CDentry
* dn
= new CDentry(dname
, inode
->hash_dentry_name(dname
), first
, last
);
372 dn
->state_set(CDentry::STATE_AUTH
);
373 if (is_auth() || !inode
->is_stray()) {
374 cache
->lru
.lru_insert_mid(dn
);
376 cache
->bottom_lru
.lru_insert_mid(dn
);
377 dn
->state_set(CDentry::STATE_BOTTOMLRU
);
381 dn
->version
= get_projected_version();
384 assert(items
.count(dn
->key()) == 0);
385 //assert(null_items.count(dn->get_name()) == 0);
387 items
[dn
->key()] = dn
;
389 dn
->get_linkage()->inode
= in
;
391 link_inode_work(dn
, in
);
393 if (dn
->last
== CEPH_NOSNAP
)
398 if (state_test(CDir::STATE_DNPINNEDFRAG
)) {
399 dn
->get(CDentry::PIN_FRAGMENTING
);
400 dn
->state_set(CDentry::STATE_FRAGMENTING
);
403 dout(12) << "add_primary_dentry " << *dn
<< dendl
;
406 if (get_num_any() == 1)
408 assert(get_num_any() == items
.size());
412 CDentry
* CDir::add_remote_dentry(boost::string_view dname
, inodeno_t ino
, unsigned char d_type
,
413 snapid_t first
, snapid_t last
)
416 assert(lookup_exact_snap(dname
, last
) == 0);
419 CDentry
* dn
= new CDentry(dname
, inode
->hash_dentry_name(dname
), ino
, d_type
, first
, last
);
421 dn
->state_set(CDentry::STATE_AUTH
);
422 cache
->lru
.lru_insert_mid(dn
);
425 dn
->version
= get_projected_version();
428 assert(items
.count(dn
->key()) == 0);
429 //assert(null_items.count(dn->get_name()) == 0);
431 items
[dn
->key()] = dn
;
432 if (last
== CEPH_NOSNAP
)
437 if (state_test(CDir::STATE_DNPINNEDFRAG
)) {
438 dn
->get(CDentry::PIN_FRAGMENTING
);
439 dn
->state_set(CDentry::STATE_FRAGMENTING
);
442 dout(12) << "add_remote_dentry " << *dn
<< dendl
;
445 if (get_num_any() == 1)
448 assert(get_num_any() == items
.size());
454 void CDir::remove_dentry(CDentry
*dn
)
456 dout(12) << "remove_dentry " << *dn
<< dendl
;
458 // there should be no client leases at this point!
459 assert(dn
->client_lease_map
.empty());
461 if (state_test(CDir::STATE_DNPINNEDFRAG
)) {
462 dn
->put(CDentry::PIN_FRAGMENTING
);
463 dn
->state_clear(CDentry::STATE_FRAGMENTING
);
466 if (dn
->get_linkage()->is_null()) {
467 if (dn
->last
== CEPH_NOSNAP
)
472 if (dn
->last
== CEPH_NOSNAP
)
478 if (!dn
->get_linkage()->is_null())
479 // detach inode and dentry
480 unlink_inode_work(dn
);
483 assert(items
.count(dn
->key()) == 1);
484 items
.erase(dn
->key());
490 if (dn
->state_test(CDentry::STATE_BOTTOMLRU
))
491 cache
->bottom_lru
.lru_remove(dn
);
493 cache
->lru
.lru_remove(dn
);
497 if (get_num_any() == 0)
499 assert(get_num_any() == items
.size());
502 void CDir::link_remote_inode(CDentry
*dn
, CInode
*in
)
504 link_remote_inode(dn
, in
->ino(), IFTODT(in
->get_projected_inode()->mode
));
507 void CDir::link_remote_inode(CDentry
*dn
, inodeno_t ino
, unsigned char d_type
)
509 dout(12) << "link_remote_inode " << *dn
<< " remote " << ino
<< dendl
;
510 assert(dn
->get_linkage()->is_null());
512 dn
->get_linkage()->set_remote(ino
, d_type
);
514 if (dn
->state_test(CDentry::STATE_BOTTOMLRU
)) {
515 cache
->bottom_lru
.lru_remove(dn
);
516 cache
->lru
.lru_insert_mid(dn
);
517 dn
->state_clear(CDentry::STATE_BOTTOMLRU
);
520 if (dn
->last
== CEPH_NOSNAP
) {
527 assert(get_num_any() == items
.size());
530 void CDir::link_primary_inode(CDentry
*dn
, CInode
*in
)
532 dout(12) << "link_primary_inode " << *dn
<< " " << *in
<< dendl
;
533 assert(dn
->get_linkage()->is_null());
535 dn
->get_linkage()->inode
= in
;
537 link_inode_work(dn
, in
);
539 if (dn
->state_test(CDentry::STATE_BOTTOMLRU
) &&
540 (is_auth() || !inode
->is_stray())) {
541 cache
->bottom_lru
.lru_remove(dn
);
542 cache
->lru
.lru_insert_mid(dn
);
543 dn
->state_clear(CDentry::STATE_BOTTOMLRU
);
546 if (dn
->last
== CEPH_NOSNAP
) {
554 assert(get_num_any() == items
.size());
557 void CDir::link_inode_work( CDentry
*dn
, CInode
*in
)
559 assert(dn
->get_linkage()->get_inode() == in
);
560 in
->set_primary_parent(dn
);
563 //in->inode.version = dn->get_version();
566 if (in
->get_num_ref())
567 dn
->get(CDentry::PIN_INODEPIN
);
569 // adjust auth pin count
570 if (in
->auth_pins
+ in
->nested_auth_pins
)
571 dn
->adjust_nested_auth_pins(in
->auth_pins
+ in
->nested_auth_pins
, in
->auth_pins
, NULL
);
573 // verify open snaprealm parent
575 in
->snaprealm
->adjust_parent();
576 else if (in
->is_any_caps())
577 in
->move_to_realm(inode
->find_snaprealm());
580 void CDir::unlink_inode(CDentry
*dn
, bool adjust_lru
)
582 if (dn
->get_linkage()->is_primary()) {
583 dout(12) << "unlink_inode " << *dn
<< " " << *dn
->get_linkage()->get_inode() << dendl
;
585 dout(12) << "unlink_inode " << *dn
<< dendl
;
588 unlink_inode_work(dn
);
590 if (adjust_lru
&& !dn
->state_test(CDentry::STATE_BOTTOMLRU
)) {
591 cache
->lru
.lru_remove(dn
);
592 cache
->bottom_lru
.lru_insert_mid(dn
);
593 dn
->state_set(CDentry::STATE_BOTTOMLRU
);
596 if (dn
->last
== CEPH_NOSNAP
) {
603 assert(get_num_any() == items
.size());
607 void CDir::try_remove_unlinked_dn(CDentry
*dn
)
609 assert(dn
->dir
== this);
610 assert(dn
->get_linkage()->is_null());
612 // no pins (besides dirty)?
613 if (dn
->get_num_ref() != dn
->is_dirty())
618 dout(10) << "try_remove_unlinked_dn " << *dn
<< " in " << *this << dendl
;
623 // NOTE: we may not have any more dirty dentries, but the fnode
624 // still changed, so the directory must remain dirty.
629 void CDir::unlink_inode_work( CDentry
*dn
)
631 CInode
*in
= dn
->get_linkage()->get_inode();
633 if (dn
->get_linkage()->is_remote()) {
636 dn
->unlink_remote(dn
->get_linkage());
638 dn
->get_linkage()->set_remote(0, 0);
639 } else if (dn
->get_linkage()->is_primary()) {
642 if (in
->get_num_ref())
643 dn
->put(CDentry::PIN_INODEPIN
);
645 // unlink auth_pin count
646 if (in
->auth_pins
+ in
->nested_auth_pins
)
647 dn
->adjust_nested_auth_pins(0 - (in
->auth_pins
+ in
->nested_auth_pins
), 0 - in
->auth_pins
, NULL
);
650 in
->remove_primary_parent(dn
);
652 in
->item_pop_lru
.remove_myself();
653 dn
->get_linkage()->inode
= 0;
655 assert(!dn
->get_linkage()->is_null());
659 void CDir::add_to_bloom(CDentry
*dn
)
661 assert(dn
->last
== CEPH_NOSNAP
);
663 /* not create bloom filter for incomplete dir that was added by log replay */
667 /* don't maintain bloom filters in standby replay (saves cycles, and also
668 * avoids need to implement clearing it in EExport for #16924) */
669 if (cache
->mds
->is_standby_replay()) {
673 unsigned size
= get_num_head_items() + get_num_snap_items();
674 if (size
< 100) size
= 100;
675 bloom
.reset(new bloom_filter(size
, 1.0 / size
, 0));
677 /* This size and false positive probability is completely random.*/
678 bloom
->insert(dn
->get_name().data(), dn
->get_name().size());
681 bool CDir::is_in_bloom(boost::string_view name
)
685 return bloom
->contains(name
.data(), name
.size());
688 void CDir::remove_null_dentries() {
689 dout(12) << "remove_null_dentries " << *this << dendl
;
691 auto p
= items
.begin();
692 while (p
!= items
.end()) {
693 CDentry
*dn
= p
->second
;
695 if (dn
->get_linkage()->is_null() && !dn
->is_projected())
699 assert(num_snap_null
== 0);
700 assert(num_head_null
== 0);
701 assert(get_num_any() == items
.size());
704 /** remove dirty null dentries for deleted directory. the dirfrag will be
705 * deleted soon, so it's safe to not commit dirty dentries.
707 * This is called when a directory is being deleted, a prerequisite
708 * of which is that its children have been unlinked: we expect to only see
709 * null, unprojected dentries here.
711 void CDir::try_remove_dentries_for_stray()
713 dout(10) << __func__
<< dendl
;
714 assert(get_parent_dir()->inode
->is_stray());
716 // clear dirty only when the directory was not snapshotted
717 bool clear_dirty
= !inode
->snaprealm
;
719 auto p
= items
.begin();
720 while (p
!= items
.end()) {
721 CDentry
*dn
= p
->second
;
723 if (dn
->last
== CEPH_NOSNAP
) {
724 assert(!dn
->is_projected());
725 assert(dn
->get_linkage()->is_null());
726 if (clear_dirty
&& dn
->is_dirty())
728 // It's OK to remove lease prematurely because we will never link
729 // the dentry to inode again.
730 if (dn
->is_any_leases())
731 dn
->remove_client_leases(cache
->mds
->locker
);
732 if (dn
->get_num_ref() == 0)
735 assert(!dn
->is_projected());
736 CDentry::linkage_t
*dnl
= dn
->get_linkage();
738 if (dnl
->is_primary()) {
739 in
= dnl
->get_inode();
740 if (clear_dirty
&& in
->is_dirty())
743 if (clear_dirty
&& dn
->is_dirty())
745 if (dn
->get_num_ref() == 0) {
748 cache
->remove_inode(in
);
753 if (clear_dirty
&& is_dirty())
757 bool CDir::try_trim_snap_dentry(CDentry
*dn
, const set
<snapid_t
>& snaps
)
759 assert(dn
->last
!= CEPH_NOSNAP
);
760 set
<snapid_t
>::const_iterator p
= snaps
.lower_bound(dn
->first
);
761 CDentry::linkage_t
*dnl
= dn
->get_linkage();
763 if (dnl
->is_primary())
764 in
= dnl
->get_inode();
765 if ((p
== snaps
.end() || *p
> dn
->last
) &&
766 (dn
->get_num_ref() == dn
->is_dirty()) &&
767 (!in
|| in
->get_num_ref() == in
->is_dirty())) {
768 dout(10) << " purging snapped " << *dn
<< dendl
;
769 if (in
&& in
->is_dirty())
773 dout(10) << " purging snapped " << *in
<< dendl
;
774 cache
->remove_inode(in
);
782 void CDir::purge_stale_snap_data(const set
<snapid_t
>& snaps
)
784 dout(10) << "purge_stale_snap_data " << snaps
<< dendl
;
786 auto p
= items
.begin();
787 while (p
!= items
.end()) {
788 CDentry
*dn
= p
->second
;
791 if (dn
->last
== CEPH_NOSNAP
)
794 try_trim_snap_dentry(dn
, snaps
);
800 * steal_dentry -- semi-violently move a dentry from one CDir to another
801 * (*) violently, in that nitems, most pins, etc. are not correctly maintained
802 * on the old CDir corpse; must call finish_old_fragment() when finished.
804 void CDir::steal_dentry(CDentry
*dn
)
806 dout(15) << "steal_dentry " << *dn
<< dendl
;
808 items
[dn
->key()] = dn
;
810 dn
->dir
->items
.erase(dn
->key());
811 if (dn
->dir
->items
.empty())
812 dn
->dir
->put(PIN_CHILD
);
814 if (get_num_any() == 0)
816 if (dn
->get_linkage()->is_null()) {
817 if (dn
->last
== CEPH_NOSNAP
)
821 } else if (dn
->last
== CEPH_NOSNAP
) {
824 if (dn
->get_linkage()->is_primary()) {
825 CInode
*in
= dn
->get_linkage()->get_inode();
826 auto pi
= in
->get_projected_inode();
828 fnode
.fragstat
.nsubdirs
++;
829 if (in
->item_pop_lru
.is_on_list())
830 pop_lru_subdirs
.push_back(&in
->item_pop_lru
);
832 fnode
.fragstat
.nfiles
++;
834 fnode
.rstat
.rbytes
+= pi
->accounted_rstat
.rbytes
;
835 fnode
.rstat
.rfiles
+= pi
->accounted_rstat
.rfiles
;
836 fnode
.rstat
.rsubdirs
+= pi
->accounted_rstat
.rsubdirs
;
837 fnode
.rstat
.rsnaprealms
+= pi
->accounted_rstat
.rsnaprealms
;
838 if (pi
->accounted_rstat
.rctime
> fnode
.rstat
.rctime
)
839 fnode
.rstat
.rctime
= pi
->accounted_rstat
.rctime
;
841 // move dirty inode rstat to new dirfrag
842 if (in
->is_dirty_rstat())
843 dirty_rstat_inodes
.push_back(&in
->dirty_rstat_item
);
844 } else if (dn
->get_linkage()->is_remote()) {
845 if (dn
->get_linkage()->get_remote_d_type() == DT_DIR
)
846 fnode
.fragstat
.nsubdirs
++;
848 fnode
.fragstat
.nfiles
++;
852 if (dn
->get_linkage()->is_primary()) {
853 CInode
*in
= dn
->get_linkage()->get_inode();
854 if (in
->is_dirty_rstat())
855 dirty_rstat_inodes
.push_back(&in
->dirty_rstat_item
);
859 if (dn
->auth_pins
|| dn
->nested_auth_pins
) {
860 // use the helpers here to maintain the auth_pin invariants on the dir inode
861 int ap
= dn
->get_num_auth_pins() + dn
->get_num_nested_auth_pins();
862 int dap
= dn
->get_num_dir_auth_pins();
864 adjust_nested_auth_pins(ap
, dap
, NULL
);
865 dn
->dir
->adjust_nested_auth_pins(-ap
, -dap
, NULL
);
868 if (dn
->is_dirty()) {
869 dirty_dentries
.push_back(&dn
->item_dir_dirty
);
876 void CDir::prepare_old_fragment(map
<string_snap_t
, std::list
<MDSInternalContextBase
*> >& dentry_waiters
, bool replay
)
878 // auth_pin old fragment for duration so that any auth_pinning
879 // during the dentry migration doesn't trigger side effects
880 if (!replay
&& is_auth())
883 if (!waiting_on_dentry
.empty()) {
884 for (const auto &p
: waiting_on_dentry
) {
885 auto &e
= dentry_waiters
[p
.first
];
886 for (const auto &waiter
: p
.second
) {
890 waiting_on_dentry
.clear();
895 void CDir::prepare_new_fragment(bool replay
)
897 if (!replay
&& is_auth()) {
901 inode
->add_dirfrag(this);
904 void CDir::finish_old_fragment(list
<MDSInternalContextBase
*>& waiters
, bool replay
)
906 // take waiters _before_ unfreeze...
908 take_waiting(WAIT_ANY_MASK
, waiters
);
910 auth_unpin(this); // pinned in prepare_old_fragment
911 assert(is_frozen_dir());
916 assert(nested_auth_pins
== 0);
917 assert(dir_auth_pins
== 0);
918 assert(auth_pins
== 0);
920 num_head_items
= num_head_null
= 0;
921 num_snap_items
= num_snap_null
= 0;
923 // this mirrors init_fragment_pins()
928 if (state_test(STATE_IMPORTBOUND
))
929 put(PIN_IMPORTBOUND
);
930 if (state_test(STATE_EXPORTBOUND
))
931 put(PIN_EXPORTBOUND
);
932 if (is_subtree_root())
938 assert(get_num_ref() == (state_test(STATE_STICKY
) ? 1:0));
941 void CDir::init_fragment_pins()
945 if (state_test(STATE_DIRTY
))
947 if (state_test(STATE_EXPORTBOUND
))
948 get(PIN_EXPORTBOUND
);
949 if (state_test(STATE_IMPORTBOUND
))
950 get(PIN_IMPORTBOUND
);
951 if (is_subtree_root())
955 void CDir::split(int bits
, list
<CDir
*>& subs
, list
<MDSInternalContextBase
*>& waiters
, bool replay
)
957 dout(10) << "split by " << bits
<< " bits on " << *this << dendl
;
959 assert(replay
|| is_complete() || !is_auth());
962 frag
.split(bits
, frags
);
964 vector
<CDir
*> subfrags(1 << bits
);
966 double fac
= 1.0 / (double)(1 << bits
); // for scaling load vecs
968 version_t rstat_version
= inode
->get_projected_inode()->rstat
.version
;
969 version_t dirstat_version
= inode
->get_projected_inode()->dirstat
.version
;
971 nest_info_t rstatdiff
;
972 frag_info_t fragstatdiff
;
973 if (fnode
.accounted_rstat
.version
== rstat_version
)
974 rstatdiff
.add_delta(fnode
.accounted_rstat
, fnode
.rstat
);
975 if (fnode
.accounted_fragstat
.version
== dirstat_version
)
976 fragstatdiff
.add_delta(fnode
.accounted_fragstat
, fnode
.fragstat
);
977 dout(10) << " rstatdiff " << rstatdiff
<< " fragstatdiff " << fragstatdiff
<< dendl
;
979 map
<string_snap_t
, std::list
<MDSInternalContextBase
*> > dentry_waiters
;
980 prepare_old_fragment(dentry_waiters
, replay
);
982 // create subfrag dirs
984 for (list
<frag_t
>::iterator p
= frags
.begin(); p
!= frags
.end(); ++p
) {
985 CDir
*f
= new CDir(inode
, *p
, cache
, is_auth());
986 f
->state_set(state
& (MASK_STATE_FRAGMENT_KEPT
| STATE_COMPLETE
));
987 f
->get_replicas() = get_replicas();
988 f
->dir_auth
= dir_auth
;
989 f
->init_fragment_pins();
990 f
->set_version(get_version());
993 f
->pop_me
.scale(fac
);
995 // FIXME; this is an approximation
996 f
->pop_nested
= pop_nested
;
997 f
->pop_nested
.scale(fac
);
998 f
->pop_auth_subtree
= pop_auth_subtree
;
999 f
->pop_auth_subtree
.scale(fac
);
1000 f
->pop_auth_subtree_nested
= pop_auth_subtree_nested
;
1001 f
->pop_auth_subtree_nested
.scale(fac
);
1003 dout(10) << " subfrag " << *p
<< " " << *f
<< dendl
;
1007 f
->set_dir_auth(get_dir_auth());
1008 f
->prepare_new_fragment(replay
);
1011 // repartition dentries
1012 while (!items
.empty()) {
1013 auto p
= items
.begin();
1015 CDentry
*dn
= p
->second
;
1016 frag_t subfrag
= inode
->pick_dirfrag(dn
->get_name());
1017 int n
= (subfrag
.value() & (subfrag
.mask() ^ frag
.mask())) >> subfrag
.mask_shift();
1018 dout(15) << " subfrag " << subfrag
<< " n=" << n
<< " for " << p
->first
<< dendl
;
1019 CDir
*f
= subfrags
[n
];
1020 f
->steal_dentry(dn
);
1023 for (const auto &p
: dentry_waiters
) {
1024 frag_t subfrag
= inode
->pick_dirfrag(p
.first
.name
);
1025 int n
= (subfrag
.value() & (subfrag
.mask() ^ frag
.mask())) >> subfrag
.mask_shift();
1026 CDir
*f
= subfrags
[n
];
1028 if (f
->waiting_on_dentry
.empty())
1029 f
->get(PIN_DNWAITER
);
1030 auto &e
= f
->waiting_on_dentry
[p
.first
];
1031 for (const auto &waiter
: p
.second
) {
1032 e
.push_back(waiter
);
1036 // FIXME: handle dirty old rstat
1038 // fix up new frag fragstats
1039 for (int i
=0; i
<n
; i
++) {
1040 CDir
*f
= subfrags
[i
];
1041 f
->fnode
.rstat
.version
= rstat_version
;
1042 f
->fnode
.accounted_rstat
= f
->fnode
.rstat
;
1043 f
->fnode
.fragstat
.version
= dirstat_version
;
1044 f
->fnode
.accounted_fragstat
= f
->fnode
.fragstat
;
1045 dout(10) << " rstat " << f
->fnode
.rstat
<< " fragstat " << f
->fnode
.fragstat
1046 << " on " << *f
<< dendl
;
1049 // give any outstanding frag stat differential to first frag
1050 dout(10) << " giving rstatdiff " << rstatdiff
<< " fragstatdiff" << fragstatdiff
1051 << " to " << *subfrags
[0] << dendl
;
1052 subfrags
[0]->fnode
.accounted_rstat
.add(rstatdiff
);
1053 subfrags
[0]->fnode
.accounted_fragstat
.add(fragstatdiff
);
1055 finish_old_fragment(waiters
, replay
);
1058 void CDir::merge(list
<CDir
*>& subs
, list
<MDSInternalContextBase
*>& waiters
, bool replay
)
1060 dout(10) << "merge " << subs
<< dendl
;
1062 mds_authority_t new_auth
= CDIR_AUTH_DEFAULT
;
1063 for (auto dir
: subs
) {
1064 if (dir
->get_dir_auth() != CDIR_AUTH_DEFAULT
&&
1065 dir
->get_dir_auth() != new_auth
) {
1066 assert(new_auth
== CDIR_AUTH_DEFAULT
);
1067 new_auth
= dir
->get_dir_auth();
1071 set_dir_auth(new_auth
);
1072 prepare_new_fragment(replay
);
1074 nest_info_t rstatdiff
;
1075 frag_info_t fragstatdiff
;
1076 bool touched_mtime
, touched_chattr
;
1077 version_t rstat_version
= inode
->get_projected_inode()->rstat
.version
;
1078 version_t dirstat_version
= inode
->get_projected_inode()->dirstat
.version
;
1080 map
<string_snap_t
, std::list
<MDSInternalContextBase
*> > dentry_waiters
;
1082 for (auto dir
: subs
) {
1083 dout(10) << " subfrag " << dir
->get_frag() << " " << *dir
<< dendl
;
1084 assert(!dir
->is_auth() || dir
->is_complete() || replay
);
1086 if (dir
->fnode
.accounted_rstat
.version
== rstat_version
)
1087 rstatdiff
.add_delta(dir
->fnode
.accounted_rstat
, dir
->fnode
.rstat
);
1088 if (dir
->fnode
.accounted_fragstat
.version
== dirstat_version
)
1089 fragstatdiff
.add_delta(dir
->fnode
.accounted_fragstat
, dir
->fnode
.fragstat
,
1090 &touched_mtime
, &touched_chattr
);
1092 dir
->prepare_old_fragment(dentry_waiters
, replay
);
1095 while (!dir
->items
.empty())
1096 steal_dentry(dir
->items
.begin()->second
);
1098 // merge replica map
1099 for (const auto &p
: dir
->get_replicas()) {
1100 unsigned cur
= get_replicas()[p
.first
];
1102 get_replicas()[p
.first
] = p
.second
;
1106 if (dir
->get_version() > get_version())
1107 set_version(dir
->get_version());
1110 state_set(dir
->get_state() & MASK_STATE_FRAGMENT_KEPT
);
1112 dir
->finish_old_fragment(waiters
, replay
);
1113 inode
->close_dirfrag(dir
->get_frag());
1116 if (!dentry_waiters
.empty()) {
1118 for (const auto &p
: dentry_waiters
) {
1119 auto &e
= waiting_on_dentry
[p
.first
];
1120 for (const auto &waiter
: p
.second
) {
1121 e
.push_back(waiter
);
1126 if (is_auth() && !replay
)
1129 // FIXME: merge dirty old rstat
1130 fnode
.rstat
.version
= rstat_version
;
1131 fnode
.accounted_rstat
= fnode
.rstat
;
1132 fnode
.accounted_rstat
.add(rstatdiff
);
1134 fnode
.fragstat
.version
= dirstat_version
;
1135 fnode
.accounted_fragstat
= fnode
.fragstat
;
1136 fnode
.accounted_fragstat
.add(fragstatdiff
);
1138 init_fragment_pins();
1144 void CDir::resync_accounted_fragstat()
1146 fnode_t
*pf
= get_projected_fnode();
1147 auto pi
= inode
->get_projected_inode();
1149 if (pf
->accounted_fragstat
.version
!= pi
->dirstat
.version
) {
1150 pf
->fragstat
.version
= pi
->dirstat
.version
;
1151 dout(10) << "resync_accounted_fragstat " << pf
->accounted_fragstat
<< " -> " << pf
->fragstat
<< dendl
;
1152 pf
->accounted_fragstat
= pf
->fragstat
;
1157 * resync rstat and accounted_rstat with inode
1159 void CDir::resync_accounted_rstat()
1161 fnode_t
*pf
= get_projected_fnode();
1162 auto pi
= inode
->get_projected_inode();
1164 if (pf
->accounted_rstat
.version
!= pi
->rstat
.version
) {
1165 pf
->rstat
.version
= pi
->rstat
.version
;
1166 dout(10) << "resync_accounted_rstat " << pf
->accounted_rstat
<< " -> " << pf
->rstat
<< dendl
;
1167 pf
->accounted_rstat
= pf
->rstat
;
1168 dirty_old_rstat
.clear();
1172 void CDir::assimilate_dirty_rstat_inodes()
1174 dout(10) << "assimilate_dirty_rstat_inodes" << dendl
;
1175 for (elist
<CInode
*>::iterator p
= dirty_rstat_inodes
.begin_use_current();
1178 assert(in
->is_auth());
1179 if (in
->is_frozen())
1182 auto &pi
= in
->project_inode();
1183 pi
.inode
.version
= in
->pre_dirty();
1185 inode
->mdcache
->project_rstat_inode_to_frag(in
, this, 0, 0, NULL
);
1187 state_set(STATE_ASSIMRSTAT
);
1188 dout(10) << "assimilate_dirty_rstat_inodes done" << dendl
;
1191 void CDir::assimilate_dirty_rstat_inodes_finish(MutationRef
& mut
, EMetaBlob
*blob
)
1193 if (!state_test(STATE_ASSIMRSTAT
))
1195 state_clear(STATE_ASSIMRSTAT
);
1196 dout(10) << "assimilate_dirty_rstat_inodes_finish" << dendl
;
1197 elist
<CInode
*>::iterator p
= dirty_rstat_inodes
.begin_use_current();
1202 if (in
->is_frozen())
1205 CDentry
*dn
= in
->get_projected_parent_dn();
1208 mut
->add_projected_inode(in
);
1210 in
->clear_dirty_rstat();
1211 blob
->add_primary_dentry(dn
, in
, true);
1214 if (!dirty_rstat_inodes
.empty())
1215 inode
->mdcache
->mds
->locker
->mark_updated_scatterlock(&inode
->nestlock
);
1221 /****************************************
1225 void CDir::add_dentry_waiter(boost::string_view dname
, snapid_t snapid
, MDSInternalContextBase
*c
)
1227 if (waiting_on_dentry
.empty())
1229 waiting_on_dentry
[string_snap_t(dname
, snapid
)].push_back(c
);
1230 dout(10) << "add_dentry_waiter dentry " << dname
1231 << " snap " << snapid
1232 << " " << c
<< " on " << *this << dendl
;
1235 void CDir::take_dentry_waiting(boost::string_view dname
, snapid_t first
, snapid_t last
,
1236 list
<MDSInternalContextBase
*>& ls
)
1238 if (waiting_on_dentry
.empty())
1241 string_snap_t
lb(dname
, first
);
1242 string_snap_t
ub(dname
, last
);
1243 auto it
= waiting_on_dentry
.lower_bound(lb
);
1244 while (it
!= waiting_on_dentry
.end() &&
1245 !(ub
< it
->first
)) {
1246 dout(10) << "take_dentry_waiting dentry " << dname
1247 << " [" << first
<< "," << last
<< "] found waiter on snap "
1249 << " on " << *this << dendl
;
1250 for (const auto &waiter
: it
->second
) {
1251 ls
.push_back(waiter
);
1253 waiting_on_dentry
.erase(it
++);
1256 if (waiting_on_dentry
.empty())
1260 void CDir::take_sub_waiting(list
<MDSInternalContextBase
*>& ls
)
1262 dout(10) << "take_sub_waiting" << dendl
;
1263 if (!waiting_on_dentry
.empty()) {
1264 for (const auto &p
: waiting_on_dentry
) {
1265 for (const auto &waiter
: p
.second
) {
1266 ls
.push_back(waiter
);
1269 waiting_on_dentry
.clear();
1276 void CDir::add_waiter(uint64_t tag
, MDSInternalContextBase
*c
)
1281 if (tag
& WAIT_ATFREEZEROOT
) {
1282 if (!(is_freezing_tree_root() || is_frozen_tree_root() ||
1283 is_freezing_dir() || is_frozen_dir())) {
1285 dout(10) << "add_waiter " << std::hex
<< tag
<< std::dec
<< " " << c
<< " should be ATFREEZEROOT, " << *this << " is not root, trying parent" << dendl
;
1286 inode
->parent
->dir
->add_waiter(tag
, c
);
1292 if (tag
& WAIT_ATSUBTREEROOT
) {
1293 if (!is_subtree_root()) {
1295 dout(10) << "add_waiter " << std::hex
<< tag
<< std::dec
<< " " << c
<< " should be ATSUBTREEROOT, " << *this << " is not root, trying parent" << dendl
;
1296 inode
->parent
->dir
->add_waiter(tag
, c
);
1301 assert(!(tag
& WAIT_CREATED
) || state_test(STATE_CREATING
));
1303 MDSCacheObject::add_waiter(tag
, c
);
1308 /* NOTE: this checks dentry waiters too */
1309 void CDir::take_waiting(uint64_t mask
, list
<MDSInternalContextBase
*>& ls
)
1311 if ((mask
& WAIT_DENTRY
) && !waiting_on_dentry
.empty()) {
1312 // take all dentry waiters
1313 for (const auto &p
: waiting_on_dentry
) {
1314 dout(10) << "take_waiting dentry " << p
.first
.name
1315 << " snap " << p
.first
.snapid
<< " on " << *this << dendl
;
1316 for (const auto &waiter
: p
.second
) {
1317 ls
.push_back(waiter
);
1320 waiting_on_dentry
.clear();
1325 MDSCacheObject::take_waiting(mask
, ls
);
1329 void CDir::finish_waiting(uint64_t mask
, int result
)
1331 dout(11) << "finish_waiting mask " << hex
<< mask
<< dec
<< " result " << result
<< " on " << *this << dendl
;
1333 list
<MDSInternalContextBase
*> finished
;
1334 take_waiting(mask
, finished
);
1336 finish_contexts(g_ceph_context
, finished
, result
);
1338 cache
->mds
->queue_waiters(finished
);
1345 fnode_t
*CDir::project_fnode()
1347 assert(get_version() != 0);
1348 projected_fnode
.emplace_back(*get_projected_fnode());
1349 auto &p
= projected_fnode
.back();
1351 if (scrub_infop
&& scrub_infop
->last_scrub_dirty
) {
1352 p
.localized_scrub_stamp
= scrub_infop
->last_local
.time
;
1353 p
.localized_scrub_version
= scrub_infop
->last_local
.version
;
1354 p
.recursive_scrub_stamp
= scrub_infop
->last_recursive
.time
;
1355 p
.recursive_scrub_version
= scrub_infop
->last_recursive
.version
;
1356 scrub_infop
->last_scrub_dirty
= false;
1357 scrub_maybe_delete_info();
1360 dout(10) << __func__
<< " " << &p
<< dendl
;
1364 void CDir::pop_and_dirty_projected_fnode(LogSegment
*ls
)
1366 assert(!projected_fnode
.empty());
1367 auto &front
= projected_fnode
.front();
1368 dout(15) << __func__
<< " " << &front
<< " v" << front
.version
<< dendl
;
1371 projected_fnode
.pop_front();
1375 version_t
CDir::pre_dirty(version_t min
)
1377 if (min
> projected_version
)
1378 projected_version
= min
;
1379 ++projected_version
;
1380 dout(10) << "pre_dirty " << projected_version
<< dendl
;
1381 return projected_version
;
1384 void CDir::mark_dirty(version_t pv
, LogSegment
*ls
)
1386 assert(get_version() < pv
);
1387 assert(pv
<= projected_version
);
1392 void CDir::_mark_dirty(LogSegment
*ls
)
1394 if (!state_test(STATE_DIRTY
)) {
1395 dout(10) << "mark_dirty (was clean) " << *this << " version " << get_version() << dendl
;
1399 dout(10) << "mark_dirty (already dirty) " << *this << " version " << get_version() << dendl
;
1402 ls
->dirty_dirfrags
.push_back(&item_dirty
);
1404 // if i've never committed, i need to be before _any_ mention of me is trimmed from the journal.
1405 if (committed_version
== 0 && !item_new
.is_on_list())
1406 ls
->new_dirfrags
.push_back(&item_new
);
1410 void CDir::mark_new(LogSegment
*ls
)
1412 ls
->new_dirfrags
.push_back(&item_new
);
1413 state_clear(STATE_CREATING
);
1415 list
<MDSInternalContextBase
*> waiters
;
1416 take_waiting(CDir::WAIT_CREATED
, waiters
);
1417 cache
->mds
->queue_waiters(waiters
);
1420 void CDir::mark_clean()
1422 dout(10) << "mark_clean " << *this << " version " << get_version() << dendl
;
1423 if (state_test(STATE_DIRTY
)) {
1424 item_dirty
.remove_myself();
1425 item_new
.remove_myself();
1427 state_clear(STATE_DIRTY
);
1432 // caller should hold auth pin of this
1433 void CDir::log_mark_dirty()
1435 if (is_dirty() || projected_version
> get_version())
1436 return; // noop if it is already dirty or will be dirty
1438 version_t pv
= pre_dirty();
1439 mark_dirty(pv
, cache
->mds
->mdlog
->get_current_segment());
1442 void CDir::mark_complete() {
1443 state_set(STATE_COMPLETE
);
1447 void CDir::first_get()
1449 inode
->get(CInode::PIN_DIRFRAG
);
1452 void CDir::last_put()
1454 inode
->put(CInode::PIN_DIRFRAG
);
1459 /******************************************************************************
1463 // -----------------------
1465 void CDir::fetch(MDSInternalContextBase
*c
, bool ignore_authpinnability
)
1468 return fetch(c
, want
, ignore_authpinnability
);
1471 void CDir::fetch(MDSInternalContextBase
*c
, boost::string_view want_dn
, bool ignore_authpinnability
)
1473 dout(10) << "fetch on " << *this << dendl
;
1476 assert(!is_complete());
1478 if (!can_auth_pin() && !ignore_authpinnability
) {
1480 dout(7) << "fetch waiting for authpinnable" << dendl
;
1481 add_waiter(WAIT_UNFREEZE
, c
);
1483 dout(7) << "fetch not authpinnable and no context" << dendl
;
1487 // unlinked directory inode shouldn't have any entry
1488 if (!inode
->is_base() && get_parent_dir()->inode
->is_stray() &&
1489 !inode
->snaprealm
) {
1490 dout(7) << "fetch dirfrag for unlinked directory, mark complete" << dendl
;
1491 if (get_version() == 0) {
1492 assert(inode
->is_auth());
1495 if (state_test(STATE_REJOINUNDEF
)) {
1496 assert(cache
->mds
->is_rejoin());
1497 state_clear(STATE_REJOINUNDEF
);
1498 cache
->opened_undef_dirfrag(this);
1504 cache
->mds
->queue_waiter(c
);
1508 if (c
) add_waiter(WAIT_COMPLETE
, c
);
1509 if (!want_dn
.empty()) wanted_items
.insert(mempool::mds_co::string(want_dn
));
1511 // already fetching?
1512 if (state_test(CDir::STATE_FETCHING
)) {
1513 dout(7) << "already fetching; waiting" << dendl
;
1518 state_set(CDir::STATE_FETCHING
);
1520 if (cache
->mds
->logger
) cache
->mds
->logger
->inc(l_mds_dir_fetch
);
1522 std::set
<dentry_key_t
> empty
;
1523 _omap_fetch(NULL
, empty
);
1526 void CDir::fetch(MDSInternalContextBase
*c
, const std::set
<dentry_key_t
>& keys
)
1528 dout(10) << "fetch " << keys
.size() << " keys on " << *this << dendl
;
1531 assert(!is_complete());
1533 if (!can_auth_pin()) {
1534 dout(7) << "fetch keys waiting for authpinnable" << dendl
;
1535 add_waiter(WAIT_UNFREEZE
, c
);
1538 if (state_test(CDir::STATE_FETCHING
)) {
1539 dout(7) << "fetch keys waiting for full fetch" << dendl
;
1540 add_waiter(WAIT_COMPLETE
, c
);
1545 if (cache
->mds
->logger
) cache
->mds
->logger
->inc(l_mds_dir_fetch
);
1547 _omap_fetch(c
, keys
);
1550 class C_IO_Dir_OMAP_FetchedMore
: public CDirIOContext
{
1551 MDSInternalContextBase
*fin
;
1555 map
<string
, bufferlist
> omap
; ///< carry-over from before
1556 map
<string
, bufferlist
> omap_more
; ///< new batch
1558 C_IO_Dir_OMAP_FetchedMore(CDir
*d
, MDSInternalContextBase
*f
) :
1559 CDirIOContext(d
), fin(f
), ret(0) { }
1560 void finish(int r
) {
1563 omap
.swap(omap_more
);
1565 omap
.insert(omap_more
.begin(), omap_more
.end());
1568 dir
->_omap_fetch_more(hdrbl
, omap
, fin
);
1570 dir
->_omap_fetched(hdrbl
, omap
, !fin
, r
);
1577 class C_IO_Dir_OMAP_Fetched
: public CDirIOContext
{
1578 MDSInternalContextBase
*fin
;
1582 map
<string
, bufferlist
> omap
;
1584 int ret1
, ret2
, ret3
;
1586 C_IO_Dir_OMAP_Fetched(CDir
*d
, MDSInternalContextBase
*f
) :
1587 CDirIOContext(d
), fin(f
), ret1(0), ret2(0), ret3(0) { }
1588 void finish(int r
) override
{
1589 // check the correctness of backtrace
1590 if (r
>= 0 && ret3
!= -ECANCELED
)
1591 dir
->inode
->verify_diri_backtrace(btbl
, ret3
);
1592 if (r
>= 0) r
= ret1
;
1593 if (r
>= 0) r
= ret2
;
1595 dir
->_omap_fetch_more(hdrbl
, omap
, fin
);
1597 dir
->_omap_fetched(hdrbl
, omap
, !fin
, r
);
1604 void CDir::_omap_fetch(MDSInternalContextBase
*c
, const std::set
<dentry_key_t
>& keys
)
1606 C_IO_Dir_OMAP_Fetched
*fin
= new C_IO_Dir_OMAP_Fetched(this, c
);
1607 object_t oid
= get_ondisk_object();
1608 object_locator_t
oloc(cache
->mds
->mdsmap
->get_metadata_pool());
1610 rd
.omap_get_header(&fin
->hdrbl
, &fin
->ret1
);
1613 rd
.omap_get_vals("", "", g_conf
->mds_dir_keys_per_op
,
1614 &fin
->omap
, &fin
->more
, &fin
->ret2
);
1617 std::set
<std::string
> str_keys
;
1618 for (auto p
: keys
) {
1621 str_keys
.insert(str
);
1623 rd
.omap_get_vals_by_keys(str_keys
, &fin
->omap
, &fin
->ret2
);
1625 // check the correctness of backtrace
1626 if (g_conf
->mds_verify_backtrace
> 0 && frag
== frag_t()) {
1627 rd
.getxattr("parent", &fin
->btbl
, &fin
->ret3
);
1628 rd
.set_last_op_flags(CEPH_OSD_OP_FLAG_FAILOK
);
1630 fin
->ret3
= -ECANCELED
;
1633 cache
->mds
->objecter
->read(oid
, oloc
, rd
, CEPH_NOSNAP
, NULL
, 0,
1634 new C_OnFinisher(fin
, cache
->mds
->finisher
));
1637 void CDir::_omap_fetch_more(
1639 map
<string
, bufferlist
>& omap
,
1640 MDSInternalContextBase
*c
)
1642 // we have more omap keys to fetch!
1643 object_t oid
= get_ondisk_object();
1644 object_locator_t
oloc(cache
->mds
->mdsmap
->get_metadata_pool());
1645 C_IO_Dir_OMAP_FetchedMore
*fin
= new C_IO_Dir_OMAP_FetchedMore(this, c
);
1646 fin
->hdrbl
.claim(hdrbl
);
1647 fin
->omap
.swap(omap
);
1649 rd
.omap_get_vals(fin
->omap
.rbegin()->first
,
1650 "", /* filter prefix */
1651 g_conf
->mds_dir_keys_per_op
,
1655 cache
->mds
->objecter
->read(oid
, oloc
, rd
, CEPH_NOSNAP
, NULL
, 0,
1656 new C_OnFinisher(fin
, cache
->mds
->finisher
));
1659 CDentry
*CDir::_load_dentry(
1660 boost::string_view key
,
1661 boost::string_view dname
,
1662 const snapid_t last
,
1665 const std::set
<snapid_t
> *snaps
,
1668 bufferlist::iterator q
= bl
.begin();
1677 dout(20) << "_fetched pos " << pos
<< " marker '" << type
<< "' dname '" << dname
1678 << " [" << first
<< "," << last
<< "]"
1682 if (snaps
&& last
!= CEPH_NOSNAP
) {
1683 set
<snapid_t
>::const_iterator p
= snaps
->lower_bound(first
);
1684 if (p
== snaps
->end() || *p
> last
) {
1685 dout(10) << " skipping stale dentry on [" << first
<< "," << last
<< "]" << dendl
;
1691 * look for existing dentry for _last_ snap, because unlink +
1692 * create may leave a "hole" (epochs during which the dentry
1693 * doesn't exist) but for which no explicit negative dentry is in
1698 dn
= lookup_exact_snap(dname
, last
);
1700 dn
= lookup(dname
, last
);
1705 unsigned char d_type
;
1707 ::decode(d_type
, q
);
1711 stale_items
.insert(mempool::mds_co::string(key
));
1712 *force_dirty
= true;
1718 CDentry::linkage_t
*dnl
= dn
->get_linkage();
1719 dout(12) << "_fetched had " << (dnl
->is_null() ? "NEG" : "") << " dentry " << *dn
<< dendl
;
1720 if (committed_version
== 0 &&
1723 ino
== dnl
->get_remote_ino() &&
1724 d_type
== dnl
->get_remote_d_type()) {
1725 // see comment below
1726 dout(10) << "_fetched had underwater dentry " << *dn
<< ", marking clean" << dendl
;
1731 dn
= add_remote_dentry(dname
, ino
, d_type
, first
, last
);
1734 CInode
*in
= cache
->get_inode(ino
); // we may or may not have it.
1736 dn
->link_remote(dn
->get_linkage(), in
);
1737 dout(12) << "_fetched got remote link " << ino
<< " which we have " << *in
<< dendl
;
1739 dout(12) << "_fetched got remote link " << ino
<< " (dont' have it)" << dendl
;
1743 else if (type
== 'I') {
1746 // Load inode data before looking up or constructing CInode
1747 InodeStore inode_data
;
1748 inode_data
.decode_bare(q
);
1752 stale_items
.insert(mempool::mds_co::string(key
));
1753 *force_dirty
= true;
1758 bool undef_inode
= false;
1760 CDentry::linkage_t
*dnl
= dn
->get_linkage();
1761 dout(12) << "_fetched had " << (dnl
->is_null() ? "NEG" : "") << " dentry " << *dn
<< dendl
;
1763 if (dnl
->is_primary()) {
1764 CInode
*in
= dnl
->get_inode();
1765 if (in
->state_test(CInode::STATE_REJOINUNDEF
)) {
1767 } else if (committed_version
== 0 &&
1769 inode_data
.inode
.ino
== in
->ino() &&
1770 inode_data
.inode
.version
== in
->get_version()) {
1771 /* clean underwater item?
1772 * Underwater item is something that is dirty in our cache from
1773 * journal replay, but was previously flushed to disk before the
1776 * We only do this is committed_version == 0. that implies either
1777 * - this is a fetch after from a clean/empty CDir is created
1778 * (and has no effect, since the dn won't exist); or
1779 * - this is a fetch after _recovery_, which is what we're worried
1780 * about. Items that are marked dirty from the journal should be
1781 * marked clean if they appear on disk.
1783 dout(10) << "_fetched had underwater dentry " << *dn
<< ", marking clean" << dendl
;
1785 dout(10) << "_fetched had underwater inode " << *dnl
->get_inode() << ", marking clean" << dendl
;
1791 if (!dn
|| undef_inode
) {
1793 CInode
*in
= cache
->get_inode(inode_data
.inode
.ino
, last
);
1794 if (!in
|| undef_inode
) {
1795 if (undef_inode
&& in
)
1798 in
= new CInode(cache
, true, first
, last
);
1800 in
->inode
= inode_data
.inode
;
1802 if (in
->is_symlink())
1803 in
->symlink
= inode_data
.symlink
;
1805 in
->dirfragtree
.swap(inode_data
.dirfragtree
);
1806 in
->xattrs
.swap(inode_data
.xattrs
);
1807 in
->old_inodes
.swap(inode_data
.old_inodes
);
1808 if (!in
->old_inodes
.empty()) {
1809 snapid_t min_first
= in
->old_inodes
.rbegin()->first
+ 1;
1810 if (min_first
> in
->first
)
1811 in
->first
= min_first
;
1814 in
->oldest_snap
= inode_data
.oldest_snap
;
1815 in
->decode_snap_blob(inode_data
.snap_blob
);
1816 if (snaps
&& !in
->snaprealm
)
1817 in
->purge_stale_snap_data(*snaps
);
1820 cache
->add_inode(in
); // add
1821 dn
= add_primary_dentry(dname
, in
, first
, last
); // link
1823 dout(12) << "_fetched got " << *dn
<< " " << *in
<< dendl
;
1825 if (in
->inode
.is_dirty_rstat())
1826 in
->mark_dirty_rstat();
1828 //in->hack_accessed = false;
1829 //in->hack_load_stamp = ceph_clock_now();
1830 //num_new_inodes_loaded++;
1831 } else if (g_conf
->get_val
<bool>("mds_hack_allow_loading_invalid_metadata")) {
1832 dout(20) << "hack: adding duplicate dentry for " << *in
<< dendl
;
1833 dn
= add_primary_dentry(dname
, in
, first
, last
);
1835 dout(0) << "_fetched badness: got (but i already had) " << *in
1836 << " mode " << in
->inode
.mode
1837 << " mtime " << in
->inode
.mtime
<< dendl
;
1838 string dirpath
, inopath
;
1839 this->inode
->make_path_string(dirpath
);
1840 in
->make_path_string(inopath
);
1841 cache
->mds
->clog
->error() << "loaded dup inode " << inode_data
.inode
.ino
1842 << " [" << first
<< "," << last
<< "] v" << inode_data
.inode
.version
1843 << " at " << dirpath
<< "/" << dname
1844 << ", but inode " << in
->vino() << " v" << in
->inode
.version
1845 << " already exists at " << inopath
;
1850 std::ostringstream oss
;
1851 oss
<< "Invalid tag char '" << type
<< "' pos " << pos
;
1852 throw buffer::malformed_input(oss
.str());
1858 void CDir::_omap_fetched(bufferlist
& hdrbl
, map
<string
, bufferlist
>& omap
,
1859 bool complete
, int r
)
1861 LogChannelRef clog
= cache
->mds
->clog
;
1862 dout(10) << "_fetched header " << hdrbl
.length() << " bytes "
1863 << omap
.size() << " keys for " << *this << dendl
;
1865 assert(r
== 0 || r
== -ENOENT
|| r
== -ENODATA
);
1867 assert(!is_frozen());
1869 if (hdrbl
.length() == 0) {
1870 dout(0) << "_fetched missing object for " << *this << dendl
;
1872 clog
->error() << "dir " << dirfrag() << " object missing on disk; some "
1873 "files may be lost (" << get_path() << ")";
1881 bufferlist::iterator p
= hdrbl
.begin();
1883 ::decode(got_fnode
, p
);
1884 } catch (const buffer::error
&err
) {
1885 derr
<< "Corrupt fnode in dirfrag " << dirfrag()
1886 << ": " << err
<< dendl
;
1887 clog
->warn() << "Corrupt fnode header in " << dirfrag() << ": "
1888 << err
<< " (" << get_path() << ")";
1893 clog
->warn() << "header buffer of dir " << dirfrag() << " has "
1894 << hdrbl
.length() - p
.get_off() << " extra bytes ("
1895 << get_path() << ")";
1901 dout(10) << "_fetched version " << got_fnode
.version
<< dendl
;
1903 // take the loaded fnode?
1904 // only if we are a fresh CDir* with no prior state.
1905 if (get_version() == 0) {
1906 assert(!is_projected());
1907 assert(!state_test(STATE_COMMITTING
));
1909 projected_version
= committing_version
= committed_version
= got_fnode
.version
;
1911 if (state_test(STATE_REJOINUNDEF
)) {
1912 assert(cache
->mds
->is_rejoin());
1913 state_clear(STATE_REJOINUNDEF
);
1914 cache
->opened_undef_dirfrag(this);
1918 list
<CInode
*> undef_inodes
;
1920 // purge stale snaps?
1921 // only if we have past_parents open!
1922 bool force_dirty
= false;
1923 const set
<snapid_t
> *snaps
= NULL
;
1924 SnapRealm
*realm
= inode
->find_snaprealm();
1925 if (!realm
->have_past_parents_open()) {
1926 dout(10) << " no snap purge, one or more past parents NOT open" << dendl
;
1927 } else if (fnode
.snap_purged_thru
< realm
->get_last_destroyed()) {
1928 snaps
= &realm
->get_snaps();
1929 dout(10) << " snap_purged_thru " << fnode
.snap_purged_thru
1930 << " < " << realm
->get_last_destroyed()
1931 << ", snap purge based on " << *snaps
<< dendl
;
1932 if (get_num_snap_items() == 0) {
1933 fnode
.snap_purged_thru
= realm
->get_last_destroyed();
1938 unsigned pos
= omap
.size() - 1;
1939 for (map
<string
, bufferlist
>::reverse_iterator p
= omap
.rbegin();
1944 dentry_key_t::decode_helper(p
->first
, dname
, last
);
1949 p
->first
, dname
, last
, p
->second
, pos
, snaps
,
1951 } catch (const buffer::error
&err
) {
1952 cache
->mds
->clog
->warn() << "Corrupt dentry '" << dname
<< "' in "
1953 "dir frag " << dirfrag() << ": "
1954 << err
<< "(" << get_path() << ")";
1956 // Remember that this dentry is damaged. Subsequent operations
1957 // that try to act directly on it will get their EIOs, but this
1958 // dirfrag as a whole will continue to look okay (minus the
1959 // mysteriously-missing dentry)
1960 go_bad_dentry(last
, dname
);
1962 // Anyone who was WAIT_DENTRY for this guy will get kicked
1963 // to RetryRequest, and hit the DamageTable-interrogating path.
1964 // Stats will now be bogus because we will think we're complete,
1965 // but have 1 or more missing dentries.
1972 CDentry::linkage_t
*dnl
= dn
->get_linkage();
1973 if (dnl
->is_primary() && dnl
->get_inode()->state_test(CInode::STATE_REJOINUNDEF
))
1974 undef_inodes
.push_back(dnl
->get_inode());
1976 if (!complete
|| wanted_items
.count(mempool::mds_co::string(boost::string_view(dname
))) > 0) {
1977 dout(10) << " touching wanted dn " << *dn
<< dendl
;
1978 inode
->mdcache
->touch_dentry(dn
);
1982 //cache->mds->logger->inc("newin", num_new_inodes_loaded);
1984 // mark complete, !fetching
1986 wanted_items
.clear();
1988 state_clear(STATE_FETCHING
);
1990 if (scrub_infop
&& scrub_infop
->need_scrub_local
) {
1991 scrub_infop
->need_scrub_local
= false;
1996 // open & force frags
1997 while (!undef_inodes
.empty()) {
1998 CInode
*in
= undef_inodes
.front();
1999 undef_inodes
.pop_front();
2000 in
->state_clear(CInode::STATE_REJOINUNDEF
);
2001 cache
->opened_undef_inode(in
);
2004 // dirty myself to remove stale snap dentries
2005 if (force_dirty
&& !inode
->mdcache
->is_readonly())
2012 finish_waiting(WAIT_COMPLETE
, 0);
2016 void CDir::_go_bad()
2018 if (get_version() == 0)
2020 state_set(STATE_BADFRAG
);
2021 // mark complete, !fetching
2023 state_clear(STATE_FETCHING
);
2027 finish_waiting(WAIT_COMPLETE
, -EIO
);
2030 void CDir::go_bad_dentry(snapid_t last
, boost::string_view dname
)
2032 dout(10) << __func__
<< " " << dname
<< dendl
;
2033 std::string
path(get_path());
2035 path
+= std::string(dname
);
2036 const bool fatal
= cache
->mds
->damage_table
.notify_dentry(
2037 inode
->ino(), frag
, last
, dname
, path
);
2039 cache
->mds
->damaged();
2040 ceph_abort(); // unreachable, damaged() respawns us
2044 void CDir::go_bad(bool complete
)
2046 dout(10) << "go_bad " << frag
<< dendl
;
2047 const bool fatal
= cache
->mds
->damage_table
.notify_dirfrag(
2048 inode
->ino(), frag
, get_path());
2050 cache
->mds
->damaged();
2051 ceph_abort(); // unreachable, damaged() respawns us
2060 // -----------------------
2066 * @param want - min version i want committed
2067 * @param c - callback for completion
2069 void CDir::commit(version_t want
, MDSInternalContextBase
*c
, bool ignore_authpinnability
, int op_prio
)
2071 dout(10) << "commit want " << want
<< " on " << *this << dendl
;
2072 if (want
== 0) want
= get_version();
2075 assert(want
<= get_version() || get_version() == 0); // can't commit the future
2076 assert(want
> committed_version
); // the caller is stupid
2078 assert(ignore_authpinnability
|| can_auth_pin());
2080 // note: queue up a noop if necessary, so that we always
2083 c
= new C_MDSInternalNoop
;
2085 // auth_pin on first waiter
2086 if (waiting_for_commit
.empty())
2088 waiting_for_commit
[want
].push_back(c
);
2091 _commit(want
, op_prio
);
2094 class C_IO_Dir_Committed
: public CDirIOContext
{
2097 C_IO_Dir_Committed(CDir
*d
, version_t v
) : CDirIOContext(d
), version(v
) { }
2098 void finish(int r
) override
{
2099 dir
->_committed(r
, version
);
2104 * Flush out the modified dentries in this dir. Keep the bufferlist
2105 * below max_write_size;
2107 void CDir::_omap_commit(int op_prio
)
2109 dout(10) << "_omap_commit" << dendl
;
2111 unsigned max_write_size
= cache
->max_dir_commit_size
;
2112 unsigned write_size
= 0;
2115 op_prio
= CEPH_MSG_PRIO_DEFAULT
;
2118 const set
<snapid_t
> *snaps
= NULL
;
2119 SnapRealm
*realm
= inode
->find_snaprealm();
2120 if (!realm
->have_past_parents_open()) {
2121 dout(10) << " no snap purge, one or more past parents NOT open" << dendl
;
2122 } else if (fnode
.snap_purged_thru
< realm
->get_last_destroyed()) {
2123 snaps
= &realm
->get_snaps();
2124 dout(10) << " snap_purged_thru " << fnode
.snap_purged_thru
2125 << " < " << realm
->get_last_destroyed()
2126 << ", snap purge based on " << *snaps
<< dendl
;
2127 // fnode.snap_purged_thru = realm->get_last_destroyed();
2130 set
<string
> to_remove
;
2131 map
<string
, bufferlist
> to_set
;
2133 C_GatherBuilder
gather(g_ceph_context
,
2134 new C_OnFinisher(new C_IO_Dir_Committed(this,
2136 cache
->mds
->finisher
));
2139 object_t oid
= get_ondisk_object();
2140 object_locator_t
oloc(cache
->mds
->mdsmap
->get_metadata_pool());
2142 if (!stale_items
.empty()) {
2143 for (const auto &p
: stale_items
) {
2144 to_remove
.insert(std::string(boost::string_view(p
)));
2145 write_size
+= p
.length();
2147 stale_items
.clear();
2150 auto write_one
= [&](CDentry
*dn
) {
2152 dn
->key().encode(key
);
2154 if (dn
->last
!= CEPH_NOSNAP
&&
2155 snaps
&& try_trim_snap_dentry(dn
, *snaps
)) {
2156 dout(10) << " rm " << key
<< dendl
;
2157 write_size
+= key
.length();
2158 to_remove
.insert(key
);
2162 if (dn
->get_linkage()->is_null()) {
2163 dout(10) << " rm " << dn
->get_name() << " " << *dn
<< dendl
;
2164 write_size
+= key
.length();
2165 to_remove
.insert(key
);
2167 dout(10) << " set " << dn
->get_name() << " " << *dn
<< dendl
;
2169 _encode_dentry(dn
, dnbl
, snaps
);
2170 write_size
+= key
.length() + dnbl
.length();
2171 to_set
[key
].swap(dnbl
);
2174 if (write_size
>= max_write_size
) {
2176 op
.priority
= op_prio
;
2178 // don't create new dirfrag blindly
2179 if (!is_new() && !state_test(CDir::STATE_FRAGMENTING
))
2180 op
.stat(NULL
, (ceph::real_time
*) NULL
, NULL
);
2182 if (!to_set
.empty())
2183 op
.omap_set(to_set
);
2184 if (!to_remove
.empty())
2185 op
.omap_rm_keys(to_remove
);
2187 cache
->mds
->objecter
->mutate(oid
, oloc
, op
, snapc
,
2188 ceph::real_clock::now(),
2189 0, gather
.new_sub());
2197 if (state_test(CDir::STATE_FRAGMENTING
)) {
2198 for (auto p
= items
.begin(); p
!= items
.end(); ) {
2199 CDentry
*dn
= p
->second
;
2201 if (!dn
->is_dirty() && dn
->get_linkage()->is_null())
2206 for (auto p
= dirty_dentries
.begin(); !p
.end(); ) {
2214 op
.priority
= op_prio
;
2216 // don't create new dirfrag blindly
2217 if (!is_new() && !state_test(CDir::STATE_FRAGMENTING
))
2218 op
.stat(NULL
, (ceph::real_time
*)NULL
, NULL
);
2221 * save the header at the last moment.. If we were to send it off before other
2222 * updates, but die before sending them all, we'd think that the on-disk state
2223 * was fully committed even though it wasn't! However, since the messages are
2224 * strictly ordered between the MDS and the OSD, and since messages to a given
2225 * PG are strictly ordered, if we simply send the message containing the header
2226 * off last, we cannot get our header into an incorrect state.
2229 ::encode(fnode
, header
);
2230 op
.omap_set_header(header
);
2232 if (!to_set
.empty())
2233 op
.omap_set(to_set
);
2234 if (!to_remove
.empty())
2235 op
.omap_rm_keys(to_remove
);
2237 cache
->mds
->objecter
->mutate(oid
, oloc
, op
, snapc
,
2238 ceph::real_clock::now(),
2239 0, gather
.new_sub());
2244 void CDir::_encode_dentry(CDentry
*dn
, bufferlist
& bl
,
2245 const set
<snapid_t
> *snaps
)
2247 // clear dentry NEW flag, if any. we can no longer silently drop it.
2250 ::encode(dn
->first
, bl
);
2252 // primary or remote?
2253 if (dn
->linkage
.is_remote()) {
2254 inodeno_t ino
= dn
->linkage
.get_remote_ino();
2255 unsigned char d_type
= dn
->linkage
.get_remote_d_type();
2256 dout(14) << " pos " << bl
.length() << " dn '" << dn
->get_name() << "' remote ino " << ino
<< dendl
;
2258 // marker, name, ino
2259 bl
.append('L'); // remote link
2261 ::encode(d_type
, bl
);
2262 } else if (dn
->linkage
.is_primary()) {
2264 CInode
*in
= dn
->linkage
.get_inode();
2267 dout(14) << " pos " << bl
.length() << " dn '" << dn
->get_name() << "' inode " << *in
<< dendl
;
2269 // marker, name, inode, [symlink string]
2270 bl
.append('I'); // inode
2272 if (in
->is_multiversion()) {
2273 if (!in
->snaprealm
) {
2275 in
->purge_stale_snap_data(*snaps
);
2276 } else if (in
->snaprealm
->have_past_parents_open()) {
2277 in
->purge_stale_snap_data(in
->snaprealm
->get_snaps());
2281 bufferlist snap_blob
;
2282 in
->encode_snap_blob(snap_blob
);
2283 in
->encode_bare(bl
, cache
->mds
->mdsmap
->get_up_features(), &snap_blob
);
2285 assert(!dn
->linkage
.is_null());
2289 void CDir::_commit(version_t want
, int op_prio
)
2291 dout(10) << "_commit want " << want
<< " on " << *this << dendl
;
2293 // we can't commit things in the future.
2294 // (even the projected future.)
2295 assert(want
<= get_version() || get_version() == 0);
2297 // check pre+postconditions.
2300 // already committed?
2301 if (committed_version
>= want
) {
2302 dout(10) << "already committed " << committed_version
<< " >= " << want
<< dendl
;
2305 // already committing >= want?
2306 if (committing_version
>= want
) {
2307 dout(10) << "already committing " << committing_version
<< " >= " << want
<< dendl
;
2308 assert(state_test(STATE_COMMITTING
));
2312 // alrady committed an older version?
2313 if (committing_version
> committed_version
) {
2314 dout(10) << "already committing older " << committing_version
<< ", waiting for that to finish" << dendl
;
2319 committing_version
= get_version();
2321 // mark committing (if not already)
2322 if (!state_test(STATE_COMMITTING
)) {
2323 dout(10) << "marking committing" << dendl
;
2324 state_set(STATE_COMMITTING
);
2327 if (cache
->mds
->logger
) cache
->mds
->logger
->inc(l_mds_dir_commit
);
2329 _omap_commit(op_prio
);
2336 * @param v version i just committed
2338 void CDir::_committed(int r
, version_t v
)
2341 // the directory could be partly purged during MDS failover
2342 if (r
== -ENOENT
&& committed_version
== 0 &&
2343 !inode
->is_base() && get_parent_dir()->inode
->is_stray()) {
2345 if (inode
->snaprealm
)
2346 inode
->state_set(CInode::STATE_MISSINGOBJS
);
2349 dout(1) << "commit error " << r
<< " v " << v
<< dendl
;
2350 cache
->mds
->clog
->error() << "failed to commit dir " << dirfrag() << " object,"
2352 cache
->mds
->handle_write_error(r
);
2357 dout(10) << "_committed v " << v
<< " on " << *this << dendl
;
2360 bool stray
= inode
->is_stray();
2363 assert(v
> committed_version
);
2364 assert(v
<= committing_version
);
2365 committed_version
= v
;
2367 // _all_ commits done?
2368 if (committing_version
== committed_version
)
2369 state_clear(CDir::STATE_COMMITTING
);
2371 // _any_ commit, even if we've been redirtied, means we're no longer new.
2372 item_new
.remove_myself();
2375 if (committed_version
== get_version())
2379 for (auto p
= dirty_dentries
.begin(); !p
.end(); ) {
2384 if (dn
->linkage
.is_primary()) {
2385 CInode
*in
= dn
->linkage
.get_inode();
2387 assert(in
->is_auth());
2389 if (committed_version
>= in
->get_version()) {
2390 if (in
->is_dirty()) {
2391 dout(15) << " dir " << committed_version
<< " >= inode " << in
->get_version() << " now clean " << *in
<< dendl
;
2395 dout(15) << " dir " << committed_version
<< " < inode " << in
->get_version() << " still dirty " << *in
<< dendl
;
2396 assert(in
->is_dirty() || in
->last
< CEPH_NOSNAP
); // special case for cow snap items (not predirtied)
2401 if (committed_version
>= dn
->get_version()) {
2402 dout(15) << " dir " << committed_version
<< " >= dn " << dn
->get_version() << " now clean " << *dn
<< dendl
;
2405 // drop clean null stray dentries immediately
2407 dn
->get_num_ref() == 0 &&
2408 !dn
->is_projected() &&
2409 dn
->get_linkage()->is_null())
2412 dout(15) << " dir " << committed_version
<< " < dn " << dn
->get_version() << " still dirty " << *dn
<< dendl
;
2413 assert(dn
->is_dirty());
2418 bool were_waiters
= !waiting_for_commit
.empty();
2420 auto it
= waiting_for_commit
.begin();
2421 while (it
!= waiting_for_commit
.end()) {
2424 if (it
->first
> committed_version
) {
2425 dout(10) << " there are waiters for " << it
->first
<< ", committing again" << dendl
;
2426 _commit(it
->first
, -1);
2429 std::list
<MDSInternalContextBase
*> t
;
2430 for (const auto &waiter
: it
->second
)
2431 t
.push_back(waiter
);
2432 cache
->mds
->queue_waiters(t
);
2433 waiting_for_commit
.erase(it
);
2437 // try drop dentries in this dirfrag if it's about to be purged
2438 if (!inode
->is_base() && get_parent_dir()->inode
->is_stray() &&
2440 cache
->maybe_eval_stray(inode
, true);
2442 // unpin if we kicked the last waiter.
2444 waiting_for_commit
.empty())
2453 void CDir::encode_export(bufferlist
& bl
)
2455 assert(!is_projected());
2456 ::encode(first
, bl
);
2457 ::encode(fnode
, bl
);
2458 ::encode(dirty_old_rstat
, bl
);
2459 ::encode(committed_version
, bl
);
2461 ::encode(state
, bl
);
2462 ::encode(dir_rep
, bl
);
2464 ::encode(pop_me
, bl
);
2465 ::encode(pop_auth_subtree
, bl
);
2467 ::encode(dir_rep_by
, bl
);
2468 ::encode(get_replicas(), bl
);
2470 get(PIN_TEMPEXPORTING
);
2473 void CDir::finish_export(utime_t now
)
2475 state
&= MASK_STATE_EXPORT_KEPT
;
2476 pop_auth_subtree_nested
.sub(now
, cache
->decayrate
, pop_auth_subtree
);
2478 pop_auth_subtree
.zero(now
);
2479 put(PIN_TEMPEXPORTING
);
2480 dirty_old_rstat
.clear();
2483 void CDir::decode_import(bufferlist::iterator
& blp
, utime_t now
, LogSegment
*ls
)
2485 ::decode(first
, blp
);
2486 ::decode(fnode
, blp
);
2487 ::decode(dirty_old_rstat
, blp
);
2488 projected_version
= fnode
.version
;
2489 ::decode(committed_version
, blp
);
2490 committing_version
= committed_version
;
2494 state
&= MASK_STATE_IMPORT_KEPT
;
2495 state_set(STATE_AUTH
| (s
& MASK_STATE_EXPORTED
));
2502 ::decode(dir_rep
, blp
);
2504 ::decode(pop_me
, now
, blp
);
2505 ::decode(pop_auth_subtree
, now
, blp
);
2506 pop_auth_subtree_nested
.add(now
, cache
->decayrate
, pop_auth_subtree
);
2508 ::decode(dir_rep_by
, blp
);
2509 ::decode(get_replicas(), blp
);
2510 if (is_replicated()) get(PIN_REPLICATED
);
2512 replica_nonce
= 0; // no longer defined
2514 // did we import some dirty scatterlock data?
2515 if (dirty_old_rstat
.size() ||
2516 !(fnode
.rstat
== fnode
.accounted_rstat
)) {
2517 cache
->mds
->locker
->mark_updated_scatterlock(&inode
->nestlock
);
2518 ls
->dirty_dirfrag_nest
.push_back(&inode
->item_dirty_dirfrag_nest
);
2520 if (!(fnode
.fragstat
== fnode
.accounted_fragstat
)) {
2521 cache
->mds
->locker
->mark_updated_scatterlock(&inode
->filelock
);
2522 ls
->dirty_dirfrag_dir
.push_back(&inode
->item_dirty_dirfrag_dir
);
2524 if (is_dirty_dft()) {
2525 if (inode
->dirfragtreelock
.get_state() != LOCK_MIX
&&
2526 inode
->dirfragtreelock
.is_stable()) {
2527 // clear stale dirtydft
2528 state_clear(STATE_DIRTYDFT
);
2530 cache
->mds
->locker
->mark_updated_scatterlock(&inode
->dirfragtreelock
);
2531 ls
->dirty_dirfrag_dirfragtree
.push_back(&inode
->item_dirty_dirfrag_dirfragtree
);
2539 /********************************
2544 * if dir_auth.first == parent, auth is same as inode.
2545 * unless .second != unknown, in which case that sticks.
2547 mds_authority_t
CDir::authority() const
2549 if (is_subtree_root())
2552 return inode
->authority();
2555 /** is_subtree_root()
2556 * true if this is an auth delegation point.
2557 * that is, dir_auth != default (parent,unknown)
2559 * some key observations:
2561 * - any region bound will be an export, or frozen.
2563 * note that this DOES heed dir_auth.pending
2566 bool CDir::is_subtree_root()
2568 if (dir_auth == CDIR_AUTH_DEFAULT) {
2569 //dout(10) << "is_subtree_root false " << dir_auth << " != " << CDIR_AUTH_DEFAULT
2570 //<< " on " << ino() << dendl;
2573 //dout(10) << "is_subtree_root true " << dir_auth << " != " << CDIR_AUTH_DEFAULT
2574 //<< " on " << ino() << dendl;
2581 * true if we are x, or an ancestor of x
2583 bool CDir::contains(CDir
*x
)
2588 x
= x
->get_inode()->get_projected_parent_dir();
2598 void CDir::set_dir_auth(mds_authority_t a
)
2600 dout(10) << "setting dir_auth=" << a
2601 << " from " << dir_auth
2602 << " on " << *this << dendl
;
2604 bool was_subtree
= is_subtree_root();
2605 bool was_ambiguous
= dir_auth
.second
>= 0;
2610 // new subtree root?
2611 if (!was_subtree
&& is_subtree_root()) {
2612 dout(10) << " new subtree root, adjusting auth_pins" << dendl
;
2614 // adjust nested auth pins
2615 if (get_cum_auth_pins())
2616 inode
->adjust_nested_auth_pins(-1, NULL
);
2618 // unpin parent of frozen dir/tree?
2619 if (inode
->is_auth()) {
2620 assert(!is_frozen_tree_root());
2621 if (is_frozen_dir())
2622 inode
->auth_unpin(this);
2625 if (was_subtree
&& !is_subtree_root()) {
2626 dout(10) << " old subtree root, adjusting auth_pins" << dendl
;
2628 // adjust nested auth pins
2629 if (get_cum_auth_pins())
2630 inode
->adjust_nested_auth_pins(1, NULL
);
2632 // pin parent of frozen dir/tree?
2633 if (inode
->is_auth()) {
2634 assert(!is_frozen_tree_root());
2635 if (is_frozen_dir())
2636 inode
->auth_pin(this);
2640 // newly single auth?
2641 if (was_ambiguous
&& dir_auth
.second
== CDIR_AUTH_UNKNOWN
) {
2642 list
<MDSInternalContextBase
*> ls
;
2643 take_waiting(WAIT_SINGLEAUTH
, ls
);
2644 cache
->mds
->queue_waiters(ls
);
2649 /*****************************************
2650 * AUTH PINS and FREEZING
2652 * the basic plan is that auth_pins only exist in auth regions, and they
2653 * prevent a freeze (and subsequent auth change).
2655 * however, we also need to prevent a parent from freezing if a child is frozen.
2656 * for that reason, the parent inode of a frozen directory is auth_pinned.
2658 * the oddity is when the frozen directory is a subtree root. if that's the case,
2659 * the parent inode isn't frozen. which means that when subtree authority is adjusted
2660 * at the bounds, inodes for any frozen bound directories need to get auth_pins at that
2665 void CDir::auth_pin(void *by
)
2671 #ifdef MDS_AUTHPIN_SET
2672 auth_pin_set
.insert(by
);
2675 dout(10) << "auth_pin by " << by
2677 << " count now " << auth_pins
<< " + " << nested_auth_pins
<< dendl
;
2680 if (!is_subtree_root() &&
2681 get_cum_auth_pins() == 1)
2682 inode
->adjust_nested_auth_pins(1, by
);
2685 void CDir::auth_unpin(void *by
)
2689 #ifdef MDS_AUTHPIN_SET
2690 assert(auth_pin_set
.count(by
));
2691 auth_pin_set
.erase(auth_pin_set
.find(by
));
2696 dout(10) << "auth_unpin by " << by
2698 << " count now " << auth_pins
<< " + " << nested_auth_pins
<< dendl
;
2699 assert(auth_pins
>= 0);
2701 int newcum
= get_cum_auth_pins();
2703 maybe_finish_freeze(); // pending freeze?
2706 if (!is_subtree_root() &&
2708 inode
->adjust_nested_auth_pins(-1, by
);
2711 void CDir::adjust_nested_auth_pins(int inc
, int dirinc
, void *by
)
2714 nested_auth_pins
+= inc
;
2715 dir_auth_pins
+= dirinc
;
2717 dout(15) << "adjust_nested_auth_pins " << inc
<< "/" << dirinc
<< " on " << *this
2718 << " by " << by
<< " count now "
2719 << auth_pins
<< " + " << nested_auth_pins
<< dendl
;
2720 assert(nested_auth_pins
>= 0);
2721 assert(dir_auth_pins
>= 0);
2723 int newcum
= get_cum_auth_pins();
2725 maybe_finish_freeze(); // pending freeze?
2728 if (!is_subtree_root()) {
2730 inode
->adjust_nested_auth_pins(-1, by
);
2731 else if (newcum
== inc
)
2732 inode
->adjust_nested_auth_pins(1, by
);
2736 #ifdef MDS_VERIFY_FRAGSTAT
2737 void CDir::verify_fragstat()
2739 assert(is_complete());
2740 if (inode
->is_stray())
2744 memset(&c
, 0, sizeof(c
));
2746 for (auto it
= items
.begin();
2749 CDentry
*dn
= it
->second
;
2753 dout(10) << " " << *dn
<< dendl
;
2754 if (dn
->is_primary())
2755 dout(10) << " " << *dn
->inode
<< dendl
;
2757 if (dn
->is_primary()) {
2758 if (dn
->inode
->is_dir())
2763 if (dn
->is_remote()) {
2764 if (dn
->get_remote_d_type() == DT_DIR
)
2771 if (c
.nsubdirs
!= fnode
.fragstat
.nsubdirs
||
2772 c
.nfiles
!= fnode
.fragstat
.nfiles
) {
2773 dout(0) << "verify_fragstat failed " << fnode
.fragstat
<< " on " << *this << dendl
;
2774 dout(0) << " i count " << c
<< dendl
;
2777 dout(0) << "verify_fragstat ok " << fnode
.fragstat
<< " on " << *this << dendl
;
2782 /*****************************************************************************
2788 bool CDir::freeze_tree()
2790 assert(!is_frozen());
2791 assert(!is_freezing());
2794 if (is_freezeable(true)) {
2799 state_set(STATE_FREEZINGTREE
);
2800 ++num_freezing_trees
;
2801 dout(10) << "freeze_tree waiting " << *this << dendl
;
2806 void CDir::_freeze_tree()
2808 dout(10) << "_freeze_tree " << *this << dendl
;
2809 assert(is_freezeable(true));
2812 if (state_test(STATE_FREEZINGTREE
)) {
2813 state_clear(STATE_FREEZINGTREE
); // actually, this may get set again by next context?
2814 --num_freezing_trees
;
2818 mds_authority_t auth
;
2819 bool was_subtree
= is_subtree_root();
2821 auth
= get_dir_auth();
2823 // temporarily prevent parent subtree from becoming frozen.
2824 inode
->auth_pin(this);
2825 // create new subtree
2829 assert(auth
.first
>= 0);
2830 assert(auth
.second
== CDIR_AUTH_UNKNOWN
);
2831 auth
.second
= auth
.first
;
2832 inode
->mdcache
->adjust_subtree_auth(this, auth
);
2834 inode
->auth_unpin(this);
2837 state_set(STATE_FROZENTREE
);
2842 void CDir::unfreeze_tree()
2844 dout(10) << "unfreeze_tree " << *this << dendl
;
2846 if (state_test(STATE_FROZENTREE
)) {
2847 // frozen. unfreeze.
2848 state_clear(STATE_FROZENTREE
);
2855 assert(is_subtree_root());
2856 // for debug purpose, caller should ensure 'dir_auth.second == dir_auth.first'
2857 mds_authority_t auth
= get_dir_auth();
2858 assert(auth
.first
>= 0);
2859 assert(auth
.second
== auth
.first
);
2860 auth
.second
= CDIR_AUTH_UNKNOWN
;
2861 inode
->mdcache
->adjust_subtree_auth(this, auth
);
2865 finish_waiting(WAIT_UNFREEZE
);
2867 finish_waiting(WAIT_FROZEN
, -1);
2869 // freezing. stop it.
2870 assert(state_test(STATE_FREEZINGTREE
));
2871 state_clear(STATE_FREEZINGTREE
);
2872 --num_freezing_trees
;
2875 finish_waiting(WAIT_UNFREEZE
);
2879 bool CDir::is_freezing_tree() const
2881 if (num_freezing_trees
== 0)
2883 const CDir
*dir
= this;
2885 if (dir
->is_freezing_tree_root()) return true;
2886 if (dir
->is_subtree_root()) return false;
2887 if (dir
->inode
->parent
)
2888 dir
= dir
->inode
->parent
->dir
;
2890 return false; // root on replica
2894 bool CDir::is_frozen_tree() const
2896 if (num_frozen_trees
== 0)
2898 const CDir
*dir
= this;
2900 if (dir
->is_frozen_tree_root()) return true;
2901 if (dir
->is_subtree_root()) return false;
2902 if (dir
->inode
->parent
)
2903 dir
= dir
->inode
->parent
->dir
;
2905 return false; // root on replica
2909 CDir
*CDir::get_frozen_tree_root()
2911 assert(is_frozen());
2914 if (dir
->is_frozen_tree_root())
2916 if (dir
->inode
->parent
)
2917 dir
= dir
->inode
->parent
->dir
;
2923 class C_Dir_AuthUnpin
: public CDirContext
{
2925 explicit C_Dir_AuthUnpin(CDir
*d
) : CDirContext(d
) {}
2926 void finish(int r
) override
{
2927 dir
->auth_unpin(dir
->get_inode());
2931 void CDir::maybe_finish_freeze()
2933 if (auth_pins
!= 1 || dir_auth_pins
!= 0)
2936 // we can freeze the _dir_ even with nested pins...
2937 if (state_test(STATE_FREEZINGDIR
)) {
2940 finish_waiting(WAIT_FROZEN
);
2943 if (nested_auth_pins
!= 0)
2946 if (state_test(STATE_FREEZINGTREE
)) {
2947 if (!is_subtree_root() && inode
->is_frozen()) {
2948 dout(10) << "maybe_finish_freeze !subtree root and frozen inode, waiting for unfreeze on " << inode
<< dendl
;
2949 // retake an auth_pin...
2951 // and release it when the parent inode unfreezes
2952 inode
->add_waiter(WAIT_UNFREEZE
, new C_Dir_AuthUnpin(this));
2958 finish_waiting(WAIT_FROZEN
);
2966 bool CDir::freeze_dir()
2968 assert(!is_frozen());
2969 assert(!is_freezing());
2972 if (is_freezeable_dir(true)) {
2977 state_set(STATE_FREEZINGDIR
);
2978 dout(10) << "freeze_dir + wait " << *this << dendl
;
2983 void CDir::_freeze_dir()
2985 dout(10) << "_freeze_dir " << *this << dendl
;
2986 //assert(is_freezeable_dir(true));
2987 // not always true during split because the original fragment may have frozen a while
2988 // ago and we're just now getting around to breaking it up.
2990 state_clear(STATE_FREEZINGDIR
);
2991 state_set(STATE_FROZENDIR
);
2994 if (is_auth() && !is_subtree_root())
2995 inode
->auth_pin(this); // auth_pin for duration of freeze
2999 void CDir::unfreeze_dir()
3001 dout(10) << "unfreeze_dir " << *this << dendl
;
3003 if (state_test(STATE_FROZENDIR
)) {
3004 state_clear(STATE_FROZENDIR
);
3007 // unpin (may => FREEZEABLE) FIXME: is this order good?
3008 if (is_auth() && !is_subtree_root())
3009 inode
->auth_unpin(this);
3011 finish_waiting(WAIT_UNFREEZE
);
3013 finish_waiting(WAIT_FROZEN
, -1);
3015 // still freezing. stop.
3016 assert(state_test(STATE_FREEZINGDIR
));
3017 state_clear(STATE_FREEZINGDIR
);
3020 finish_waiting(WAIT_UNFREEZE
);
3025 * Slightly less complete than operator<<, because this is intended
3026 * for identifying a directory and its state rather than for dumping
3029 void CDir::dump(Formatter
*f
) const
3033 f
->dump_stream("path") << get_path();
3035 f
->dump_stream("dirfrag") << dirfrag();
3036 f
->dump_int("snapid_first", first
);
3038 f
->dump_stream("projected_version") << get_projected_version();
3039 f
->dump_stream("version") << get_version();
3040 f
->dump_stream("committing_version") << get_committing_version();
3041 f
->dump_stream("committed_version") << get_committed_version();
3043 f
->dump_bool("is_rep", is_rep());
3045 if (get_dir_auth() != CDIR_AUTH_DEFAULT
) {
3046 if (get_dir_auth().second
== CDIR_AUTH_UNKNOWN
) {
3047 f
->dump_stream("dir_auth") << get_dir_auth().first
;
3049 f
->dump_stream("dir_auth") << get_dir_auth();
3052 f
->dump_string("dir_auth", "");
3055 f
->open_array_section("states");
3056 MDSCacheObject::dump_states(f
);
3057 if (state_test(CDir::STATE_COMPLETE
)) f
->dump_string("state", "complete");
3058 if (state_test(CDir::STATE_FREEZINGTREE
)) f
->dump_string("state", "freezingtree");
3059 if (state_test(CDir::STATE_FROZENTREE
)) f
->dump_string("state", "frozentree");
3060 if (state_test(CDir::STATE_FROZENDIR
)) f
->dump_string("state", "frozendir");
3061 if (state_test(CDir::STATE_FREEZINGDIR
)) f
->dump_string("state", "freezingdir");
3062 if (state_test(CDir::STATE_EXPORTBOUND
)) f
->dump_string("state", "exportbound");
3063 if (state_test(CDir::STATE_IMPORTBOUND
)) f
->dump_string("state", "importbound");
3064 if (state_test(CDir::STATE_BADFRAG
)) f
->dump_string("state", "badfrag");
3067 MDSCacheObject::dump(f
);
3070 void CDir::dump_load(Formatter
*f
, utime_t now
, const DecayRate
& rate
)
3072 f
->dump_stream("path") << get_path();
3073 f
->dump_stream("dirfrag") << dirfrag();
3075 f
->open_object_section("pop_me");
3076 pop_me
.dump(f
, now
, rate
);
3079 f
->open_object_section("pop_nested");
3080 pop_nested
.dump(f
, now
, rate
);
3083 f
->open_object_section("pop_auth_subtree");
3084 pop_auth_subtree
.dump(f
, now
, rate
);
3087 f
->open_object_section("pop_auth_subtree_nested");
3088 pop_auth_subtree_nested
.dump(f
, now
, rate
);
3092 /****** Scrub Stuff *******/
3094 void CDir::scrub_info_create() const
3096 assert(!scrub_infop
);
3098 // break out of const-land to set up implicit initial state
3099 CDir
*me
= const_cast<CDir
*>(this);
3100 fnode_t
*fn
= me
->get_projected_fnode();
3102 std::unique_ptr
<scrub_info_t
> si(new scrub_info_t());
3104 si
->last_recursive
.version
= si
->recursive_start
.version
=
3105 fn
->recursive_scrub_version
;
3106 si
->last_recursive
.time
= si
->recursive_start
.time
=
3107 fn
->recursive_scrub_stamp
;
3109 si
->last_local
.version
= fn
->localized_scrub_version
;
3110 si
->last_local
.time
= fn
->localized_scrub_stamp
;
3112 me
->scrub_infop
.swap(si
);
3115 void CDir::scrub_initialize(const ScrubHeaderRefConst
& header
)
3117 dout(20) << __func__
<< dendl
;
3118 assert(is_complete());
3119 assert(header
!= nullptr);
3121 // FIXME: weird implicit construction, is someone else meant
3122 // to be calling scrub_info_create first?
3124 assert(scrub_infop
&& !scrub_infop
->directory_scrubbing
);
3126 scrub_infop
->recursive_start
.version
= get_projected_version();
3127 scrub_infop
->recursive_start
.time
= ceph_clock_now();
3129 scrub_infop
->directories_to_scrub
.clear();
3130 scrub_infop
->directories_scrubbing
.clear();
3131 scrub_infop
->directories_scrubbed
.clear();
3132 scrub_infop
->others_to_scrub
.clear();
3133 scrub_infop
->others_scrubbing
.clear();
3134 scrub_infop
->others_scrubbed
.clear();
3136 for (auto i
= items
.begin();
3139 // TODO: handle snapshot scrubbing
3140 if (i
->first
.snapid
!= CEPH_NOSNAP
)
3143 CDentry::linkage_t
*dnl
= i
->second
->get_projected_linkage();
3144 if (dnl
->is_primary()) {
3145 if (dnl
->get_inode()->is_dir())
3146 scrub_infop
->directories_to_scrub
.insert(i
->first
);
3148 scrub_infop
->others_to_scrub
.insert(i
->first
);
3149 } else if (dnl
->is_remote()) {
3150 // TODO: check remote linkage
3153 scrub_infop
->directory_scrubbing
= true;
3154 scrub_infop
->header
= header
;
3157 void CDir::scrub_finished()
3159 dout(20) << __func__
<< dendl
;
3160 assert(scrub_infop
&& scrub_infop
->directory_scrubbing
);
3162 assert(scrub_infop
->directories_to_scrub
.empty());
3163 assert(scrub_infop
->directories_scrubbing
.empty());
3164 scrub_infop
->directories_scrubbed
.clear();
3165 assert(scrub_infop
->others_to_scrub
.empty());
3166 assert(scrub_infop
->others_scrubbing
.empty());
3167 scrub_infop
->others_scrubbed
.clear();
3168 scrub_infop
->directory_scrubbing
= false;
3170 scrub_infop
->last_recursive
= scrub_infop
->recursive_start
;
3171 scrub_infop
->last_scrub_dirty
= true;
3174 int CDir::_next_dentry_on_set(dentry_key_set
&dns
, bool missing_okay
,
3175 MDSInternalContext
*cb
, CDentry
**dnout
)
3180 while (!dns
.empty()) {
3181 set
<dentry_key_t
>::iterator front
= dns
.begin();
3183 dn
= lookup(dnkey
.name
);
3185 if (!is_complete() &&
3186 (!has_bloom() || is_in_bloom(dnkey
.name
))) {
3187 // need to re-read this dirfrag
3193 dout(15) << " we no longer have directory dentry "
3194 << dnkey
.name
<< ", assuming it got renamed" << dendl
;
3198 dout(5) << " we lost dentry " << dnkey
.name
3199 << ", bailing out because that's impossible!" << dendl
;
3203 // okay, we got a dentry
3206 if (dn
->get_projected_version() < scrub_infop
->last_recursive
.version
&&
3207 !(scrub_infop
->header
->get_force())) {
3208 dout(15) << " skip dentry " << dnkey
.name
3209 << ", no change since last scrub" << dendl
;
3213 if (!dn
->get_linkage()->is_primary()) {
3214 dout(15) << " skip dentry " << dnkey
.name
3215 << ", no longer primary" << dendl
;
3226 int CDir::scrub_dentry_next(MDSInternalContext
*cb
, CDentry
**dnout
)
3228 dout(20) << __func__
<< dendl
;
3229 assert(scrub_infop
&& scrub_infop
->directory_scrubbing
);
3231 dout(20) << "trying to scrub directories underneath us" << dendl
;
3232 int rval
= _next_dentry_on_set(scrub_infop
->directories_to_scrub
, true,
3235 dout(20) << __func__
<< " inserted to directories scrubbing: "
3237 scrub_infop
->directories_scrubbing
.insert((*dnout
)->key());
3238 } else if (rval
== EAGAIN
) {
3239 // we don't need to do anything else
3240 } else { // we emptied out the directory scrub set
3241 assert(rval
== ENOENT
);
3242 dout(20) << "no directories left, moving on to other kinds of dentries"
3245 rval
= _next_dentry_on_set(scrub_infop
->others_to_scrub
, false, cb
, dnout
);
3247 dout(20) << __func__
<< " inserted to others scrubbing: "
3249 scrub_infop
->others_scrubbing
.insert((*dnout
)->key());
3252 dout(20) << " returning " << rval
<< " with dn=" << *dnout
<< dendl
;
3256 void CDir::scrub_dentries_scrubbing(list
<CDentry
*> *out_dentries
)
3258 dout(20) << __func__
<< dendl
;
3259 assert(scrub_infop
&& scrub_infop
->directory_scrubbing
);
3261 for (set
<dentry_key_t
>::iterator i
=
3262 scrub_infop
->directories_scrubbing
.begin();
3263 i
!= scrub_infop
->directories_scrubbing
.end();
3265 CDentry
*d
= lookup(i
->name
, i
->snapid
);
3267 out_dentries
->push_back(d
);
3269 for (set
<dentry_key_t
>::iterator i
= scrub_infop
->others_scrubbing
.begin();
3270 i
!= scrub_infop
->others_scrubbing
.end();
3272 CDentry
*d
= lookup(i
->name
, i
->snapid
);
3274 out_dentries
->push_back(d
);
3278 void CDir::scrub_dentry_finished(CDentry
*dn
)
3280 dout(20) << __func__
<< " on dn " << *dn
<< dendl
;
3281 assert(scrub_infop
&& scrub_infop
->directory_scrubbing
);
3282 dentry_key_t dn_key
= dn
->key();
3283 if (scrub_infop
->directories_scrubbing
.erase(dn_key
)) {
3284 scrub_infop
->directories_scrubbed
.insert(dn_key
);
3286 assert(scrub_infop
->others_scrubbing
.count(dn_key
));
3287 scrub_infop
->others_scrubbing
.erase(dn_key
);
3288 scrub_infop
->others_scrubbed
.insert(dn_key
);
3292 void CDir::scrub_maybe_delete_info()
3295 !scrub_infop
->directory_scrubbing
&&
3296 !scrub_infop
->need_scrub_local
&&
3297 !scrub_infop
->last_scrub_dirty
&&
3298 !scrub_infop
->pending_scrub_error
&&
3299 scrub_infop
->dirty_scrub_stamps
.empty()) {
3300 scrub_infop
.reset();
3304 bool CDir::scrub_local()
3306 assert(is_complete());
3307 bool rval
= check_rstats(true);
3311 scrub_infop
->last_local
.time
= ceph_clock_now();
3312 scrub_infop
->last_local
.version
= get_projected_version();
3313 scrub_infop
->pending_scrub_error
= false;
3314 scrub_infop
->last_scrub_dirty
= true;
3316 scrub_infop
->pending_scrub_error
= true;
3317 if (scrub_infop
->header
->get_repair())
3318 cache
->repair_dirfrag_stats(this);
3323 std::string
CDir::get_path() const
3326 get_inode()->make_path_string(path
, true);
3330 bool CDir::should_split_fast() const
3332 // Max size a fragment can be before trigger fast splitting
3333 int fast_limit
= g_conf
->mds_bal_split_size
* g_conf
->mds_bal_fragment_fast_factor
;
3335 // Fast path: the sum of accounted size and null dentries does not
3336 // exceed threshold: we definitely are not over it.
3337 if (get_frag_size() + get_num_head_null() <= fast_limit
) {
3341 // Fast path: the accounted size of the frag exceeds threshold: we
3342 // definitely are over it
3343 if (get_frag_size() > fast_limit
) {
3347 int64_t effective_size
= 0;
3349 for (const auto &p
: items
) {
3350 const CDentry
*dn
= p
.second
;
3351 if (!dn
->get_projected_linkage()->is_null()) {
3356 return effective_size
> fast_limit
;
3359 MEMPOOL_DEFINE_OBJECT_FACTORY(CDir
, co_dir
, mds_co
);