ceph/src/mds/CDir.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #include <string_view>
  16 #include <algorithm>
  17
  18 #include "include/types.h"
  19
  20 #include "CDir.h"
  21 #include "CDentry.h"
  22 #include "CInode.h"
  23 #include "Mutation.h"
  24
  25 #include "MDSMap.h"
  26 #include "MDSRank.h"
  27 #include "MDCache.h"
  28 #include "Locker.h"
  29 #include "MDLog.h"
  30 #include "LogSegment.h"
  31 #include "MDBalancer.h"
  32
  33 #include "common/bloom_filter.hpp"
  34 #include "include/Context.h"
  35 #include "common/Clock.h"
  36
  37 #include "osdc/Objecter.h"
  38
  39 #include "common/config.h"
  40 #include "include/ceph_assert.h"
  41 #include "include/compat.h"
  42
  43 #define dout_context g_ceph_context
  44 #define dout_subsys ceph_subsys_mds
  45 #undef dout_prefix
  46 #define dout_prefix *_dout << "mds." << mdcache->mds->get_nodeid() << ".cache.dir(" << this->dirfrag() << ") "
  47
  48 using namespace std;
  49
  50 int CDir::num_frozen_trees = 0;
  51 int CDir::num_freezing_trees = 0;
  52
  53 CDir::fnode_const_ptr CDir::empty_fnode = CDir::allocate_fnode();
  54
  55 class CDirContext : public MDSContext
  56 {
  57 protected:
  58   CDir *dir;
  59   MDSRank* get_mds() override {return dir->mdcache->mds;}
  60
  61 public:
  62   explicit CDirContext(CDir *d) : dir(d) {
  63     ceph_assert(dir != NULL);
  64   }
  65 };
  66
  67
  68 class CDirIOContext : public MDSIOContextBase
  69 {
  70 protected:
  71   CDir *dir;
  72   MDSRank* get_mds() override {return dir->mdcache->mds;}
  73
  74 public:
  75   explicit CDirIOContext(CDir *d) : dir(d) {
  76     ceph_assert(dir != NULL);
  77   }
  78 };
  79
  80
  81 // PINS
  82 //int cdir_pins[CDIR_NUM_PINS] = { 0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
  83
  84
  85 ostream& operator<<(ostream& out, const CDir& dir)
  86 {
  87   out << "[dir " << dir.dirfrag() << " " << dir.get_path() << "/"
  88       << " [" << dir.first << ",head]";
  89   if (dir.is_auth()) {
  90     out << " auth";
  91     if (dir.is_replicated())
  92       out << dir.get_replicas();
  93
  94     if (dir.is_projected())
  95       out << " pv=" << dir.get_projected_version();
  96     out << " v=" << dir.get_version();
  97     out << " cv=" << dir.get_committing_version();
  98     out << "/" << dir.get_committed_version();
  99   } else {
 100     mds_authority_t a = dir.authority();
 101     out << " rep@" << a.first;
 102     if (a.second != CDIR_AUTH_UNKNOWN)
 103       out << "," << a.second;
 104     out << "." << dir.get_replica_nonce();
 105   }
 106
 107   if (dir.is_rep()) out << " REP";
 108
 109   if (dir.get_dir_auth() != CDIR_AUTH_DEFAULT) {
 110     if (dir.get_dir_auth().second == CDIR_AUTH_UNKNOWN)
 111       out << " dir_auth=" << dir.get_dir_auth().first;
 112     else
 113       out << " dir_auth=" << dir.get_dir_auth();
 114   }
 115
 116   if (dir.get_auth_pins() || dir.get_dir_auth_pins()) {
 117     out << " ap=" << dir.get_auth_pins()
 118         << "+" << dir.get_dir_auth_pins();
 119 #ifdef MDS_AUTHPIN_SET
 120     dir.print_authpin_set(out);
 121 #endif
 122   }
 123
 124   out << " state=" << dir.get_state();
 125   if (dir.state_test(CDir::STATE_COMPLETE)) out << "|complete";
 126   if (dir.state_test(CDir::STATE_FREEZINGTREE)) out << "|freezingtree";
 127   if (dir.state_test(CDir::STATE_FROZENTREE)) out << "|frozentree";
 128   if (dir.state_test(CDir::STATE_AUXSUBTREE)) out << "|auxsubtree";
 129   if (dir.state_test(CDir::STATE_FROZENDIR)) out << "|frozendir";
 130   if (dir.state_test(CDir::STATE_FREEZINGDIR)) out << "|freezingdir";
 131   if (dir.state_test(CDir::STATE_EXPORTBOUND)) out << "|exportbound";
 132   if (dir.state_test(CDir::STATE_IMPORTBOUND)) out << "|importbound";
 133   if (dir.state_test(CDir::STATE_BADFRAG)) out << "|badfrag";
 134   if (dir.state_test(CDir::STATE_FRAGMENTING)) out << "|fragmenting";
 135   if (dir.state_test(CDir::STATE_CREATING)) out << "|creating";
 136   if (dir.state_test(CDir::STATE_COMMITTING)) out << "|committing";
 137   if (dir.state_test(CDir::STATE_FETCHING)) out << "|fetching";
 138   if (dir.state_test(CDir::STATE_EXPORTING)) out << "|exporting";
 139   if (dir.state_test(CDir::STATE_IMPORTING)) out << "|importing";
 140   if (dir.state_test(CDir::STATE_STICKY)) out << "|sticky";
 141   if (dir.state_test(CDir::STATE_DNPINNEDFRAG)) out << "|dnpinnedfrag";
 142   if (dir.state_test(CDir::STATE_ASSIMRSTAT)) out << "|assimrstat";
 143
 144   // fragstat
 145   out << " " << dir.get_fnode()->fragstat;
 146   if (!(dir.get_fnode()->fragstat == dir.get_fnode()->accounted_fragstat))
 147     out << "/" << dir.get_fnode()->accounted_fragstat;
 148   if (g_conf()->mds_debug_scatterstat && dir.is_projected()) {
 149     const auto& pf = dir.get_projected_fnode();
 150     out << "->" << pf->fragstat;
 151     if (!(pf->fragstat == pf->accounted_fragstat))
 152       out << "/" << pf->accounted_fragstat;
 153   }
 154
 155   // rstat
 156   out << " " << dir.get_fnode()->rstat;
 157   if (!(dir.get_fnode()->rstat == dir.get_fnode()->accounted_rstat))
 158     out << "/" << dir.get_fnode()->accounted_rstat;
 159   if (g_conf()->mds_debug_scatterstat && dir.is_projected()) {
 160     const auto& pf = dir.get_projected_fnode();
 161     out << "->" << pf->rstat;
 162     if (!(pf->rstat == pf->accounted_rstat))
 163       out << "/" << pf->accounted_rstat;
 164  }
 165
 166   out << " hs=" << dir.get_num_head_items() << "+" << dir.get_num_head_null();
 167   out << ",ss=" << dir.get_num_snap_items() << "+" << dir.get_num_snap_null();
 168   if (dir.get_num_dirty())
 169     out << " dirty=" << dir.get_num_dirty();
 170
 171   if (dir.get_num_ref()) {
 172     out << " |";
 173     dir.print_pin_set(out);
 174   }
 175
 176   out << " " << &dir;
 177   return out << "]";
 178 }
 179
 180
 181 void CDir::print(ostream& out)
 182 {
 183   out << *this;
 184 }
 185
 186
 187
 188
 189 ostream& CDir::print_db_line_prefix(ostream& out)
 190 {
 191   return out << ceph_clock_now() << " mds." << mdcache->mds->get_nodeid() << ".cache.dir(" << this->dirfrag() << ") ";
 192 }
 193
 194
 195
 196 // -------------------------------------------------------------------
 197 // CDir
 198
 199 CDir::CDir(CInode *in, frag_t fg, MDCache *mdc, bool auth) :
 200   mdcache(mdc), inode(in), frag(fg),
 201   dirty_rstat_inodes(member_offset(CInode, dirty_rstat_item)),
 202   dirty_dentries(member_offset(CDentry, item_dir_dirty)),
 203   item_dirty(this), item_new(this),
 204   lock_caches_with_auth_pins(member_offset(MDLockCache::DirItem, item_dir)),
 205   freezing_inodes(member_offset(CInode, item_freezing_inode)),
 206   dir_rep(REP_NONE),
 207   pop_me(mdc->decayrate),
 208   pop_nested(mdc->decayrate),
 209   pop_auth_subtree(mdc->decayrate),
 210   pop_auth_subtree_nested(mdc->decayrate),
 211   pop_spread(mdc->decayrate),
 212   pop_lru_subdirs(member_offset(CInode, item_pop_lru)),
 213   dir_auth(CDIR_AUTH_DEFAULT)
 214 {
 215   // auth
 216   ceph_assert(in->is_dir());
 217   if (auth)
 218     state_set(STATE_AUTH);
 219 }
 220
 221 /**
 222  * Check the recursive statistics on size for consistency.
 223  * If mds_debug_scatterstat is enabled, assert for correctness,
 224  * otherwise just print out the mismatch and continue.
 225  */
 226 bool CDir::check_rstats(bool scrub)
 227 {
 228   if (!g_conf()->mds_debug_scatterstat && !scrub)
 229     return true;
 230
 231   dout(25) << "check_rstats on " << this << dendl;
 232   if (!is_complete() || !is_auth() || is_frozen()) {
 233     dout(3) << "check_rstats " << (scrub ? "(scrub) " : "")
 234             << "bailing out -- incomplete or non-auth or frozen dir on "
 235             << *this << dendl;
 236     return !scrub;
 237   }
 238
 239   frag_info_t frag_info;
 240   nest_info_t nest_info;
 241   for (auto i = items.begin(); i != items.end(); ++i) {
 242     if (i->second->last != CEPH_NOSNAP)
 243       continue;
 244     CDentry::linkage_t *dnl = i->second->get_linkage();
 245     if (dnl->is_primary()) {
 246       CInode *in = dnl->get_inode();
 247       nest_info.add(in->get_inode()->accounted_rstat);
 248       if (in->is_dir())
 249         frag_info.nsubdirs++;
 250       else
 251         frag_info.nfiles++;
 252     } else if (dnl->is_remote())
 253       frag_info.nfiles++;
 254   }
 255
 256   bool good = true;
 257   // fragstat
 258   if(!frag_info.same_sums(fnode->fragstat)) {
 259     dout(1) << "mismatch between head items and fnode.fragstat! printing dentries" << dendl;
 260     dout(1) << "get_num_head_items() = " << get_num_head_items()
 261              << "; fnode.fragstat.nfiles=" << fnode->fragstat.nfiles
 262              << " fnode.fragstat.nsubdirs=" << fnode->fragstat.nsubdirs << dendl;
 263     good = false;
 264   } else {
 265     dout(20) << "get_num_head_items() = " << get_num_head_items()
 266              << "; fnode.fragstat.nfiles=" << fnode->fragstat.nfiles
 267              << " fnode.fragstat.nsubdirs=" << fnode->fragstat.nsubdirs << dendl;
 268   }
 269
 270   // rstat
 271   if (!nest_info.same_sums(fnode->rstat)) {
 272     dout(1) << "mismatch between child accounted_rstats and my rstats!" << dendl;
 273     dout(1) << "total of child dentries: " << nest_info << dendl;
 274     dout(1) << "my rstats:              " << fnode->rstat << dendl;
 275     good = false;
 276   } else {
 277     dout(20) << "total of child dentries: " << nest_info << dendl;
 278     dout(20) << "my rstats:              " << fnode->rstat << dendl;
 279   }
 280
 281   if (!good) {
 282     if (!scrub) {
 283       for (auto i = items.begin(); i != items.end(); ++i) {
 284         CDentry *dn = i->second;
 285         if (dn->get_linkage()->is_primary()) {
 286           CInode *in = dn->get_linkage()->inode;
 287           dout(1) << *dn << " rstat " << in->get_inode()->accounted_rstat << dendl;
 288         } else {
 289           dout(1) << *dn << dendl;
 290         }
 291       }
 292
 293       ceph_assert(frag_info.nfiles == fnode->fragstat.nfiles);
 294       ceph_assert(frag_info.nsubdirs == fnode->fragstat.nsubdirs);
 295       ceph_assert(nest_info.rbytes == fnode->rstat.rbytes);
 296       ceph_assert(nest_info.rfiles == fnode->rstat.rfiles);
 297       ceph_assert(nest_info.rsubdirs == fnode->rstat.rsubdirs);
 298     }
 299   }
 300   dout(10) << "check_rstats complete on " << this << dendl;
 301   return good;
 302 }
 303
 304 void CDir::adjust_num_inodes_with_caps(int d)
 305 {
 306   // FIXME: smarter way to decide if adding 'this' to open file table
 307   if (num_inodes_with_caps == 0 && d > 0)
 308     mdcache->open_file_table.add_dirfrag(this);
 309   else if (num_inodes_with_caps > 0 && num_inodes_with_caps == -d)
 310     mdcache->open_file_table.remove_dirfrag(this);
 311
 312   num_inodes_with_caps += d;
 313   ceph_assert(num_inodes_with_caps >= 0);
 314 }
 315
 316 CDentry *CDir::lookup(std::string_view name, snapid_t snap)
 317 {
 318   dout(20) << "lookup (" << snap << ", '" << name << "')" << dendl;
 319   auto iter = items.lower_bound(dentry_key_t(snap, name, inode->hash_dentry_name(name)));
 320   if (iter == items.end())
 321     return 0;
 322   if (iter->second->get_name() == name &&
 323       iter->second->first <= snap &&
 324       iter->second->last >= snap) {
 325     dout(20) << "  hit -> " << iter->first << dendl;
 326     return iter->second;
 327   }
 328   dout(20) << "  miss -> " << iter->first << dendl;
 329   return 0;
 330 }
 331
 332 CDentry *CDir::lookup_exact_snap(std::string_view name, snapid_t last) {
 333   dout(20) << __func__ << " (" << last << ", '" << name << "')" << dendl;
 334   auto p = items.find(dentry_key_t(last, name, inode->hash_dentry_name(name)));
 335   if (p == items.end())
 336     return NULL;
 337   return p->second;
 338 }
 339
 340 /***
 341  * linking fun
 342  */
 343
 344 CDentry* CDir::add_null_dentry(std::string_view dname,
 345                                snapid_t first, snapid_t last)
 346 {
 347   // foreign
 348   ceph_assert(lookup_exact_snap(dname, last) == 0);
 349
 350   // create dentry
 351   CDentry* dn = new CDentry(dname, inode->hash_dentry_name(dname), "", first, last);
 352   if (is_auth())
 353     dn->state_set(CDentry::STATE_AUTH);
 354
 355   mdcache->bottom_lru.lru_insert_mid(dn);
 356   dn->state_set(CDentry::STATE_BOTTOMLRU);
 357
 358   dn->dir = this;
 359   dn->version = get_projected_version();
 360
 361   // add to dir
 362   ceph_assert(items.count(dn->key()) == 0);
 363   //assert(null_items.count(dn->get_name()) == 0);
 364
 365   items[dn->key()] = dn;
 366   if (last == CEPH_NOSNAP)
 367     num_head_null++;
 368   else
 369     num_snap_null++;
 370
 371   if (state_test(CDir::STATE_DNPINNEDFRAG)) {
 372     dn->get(CDentry::PIN_FRAGMENTING);
 373     dn->state_set(CDentry::STATE_FRAGMENTING);
 374   }
 375
 376   dout(12) << __func__ << " " << *dn << dendl;
 377
 378   // pin?
 379   if (get_num_any() == 1)
 380     get(PIN_CHILD);
 381
 382   ceph_assert(get_num_any() == items.size());
 383   return dn;
 384 }
 385
 386
 387 CDentry* CDir::add_primary_dentry(std::string_view dname, CInode *in,
 388                                   mempool::mds_co::string alternate_name,
 389                                   snapid_t first, snapid_t last)
 390 {
 391   // primary
 392   ceph_assert(lookup_exact_snap(dname, last) == 0);
 393
 394   // create dentry
 395   CDentry* dn = new CDentry(dname, inode->hash_dentry_name(dname), std::move(alternate_name), first, last);
 396   if (is_auth())
 397     dn->state_set(CDentry::STATE_AUTH);
 398   if (is_auth() || !inode->is_stray()) {
 399     mdcache->lru.lru_insert_mid(dn);
 400   } else {
 401     mdcache->bottom_lru.lru_insert_mid(dn);
 402     dn->state_set(CDentry::STATE_BOTTOMLRU);
 403   }
 404
 405   dn->dir = this;
 406   dn->version = get_projected_version();
 407
 408   // add to dir
 409   ceph_assert(items.count(dn->key()) == 0);
 410   //assert(null_items.count(dn->get_name()) == 0);
 411
 412   items[dn->key()] = dn;
 413
 414   dn->get_linkage()->inode = in;
 415
 416   link_inode_work(dn, in);
 417
 418   if (dn->last == CEPH_NOSNAP)
 419     num_head_items++;
 420   else
 421     num_snap_items++;
 422
 423   if (state_test(CDir::STATE_DNPINNEDFRAG)) {
 424     dn->get(CDentry::PIN_FRAGMENTING);
 425     dn->state_set(CDentry::STATE_FRAGMENTING);
 426   }
 427
 428   dout(12) << __func__ << " " << *dn << dendl;
 429
 430   // pin?
 431   if (get_num_any() == 1)
 432     get(PIN_CHILD);
 433   ceph_assert(get_num_any() == items.size());
 434   return dn;
 435 }
 436
 437 CDentry* CDir::add_remote_dentry(std::string_view dname, inodeno_t ino, unsigned char d_type,
 438                                  mempool::mds_co::string alternate_name,
 439                                  snapid_t first, snapid_t last)
 440 {
 441   // foreign
 442   ceph_assert(lookup_exact_snap(dname, last) == 0);
 443
 444   // create dentry
 445   CDentry* dn = new CDentry(dname, inode->hash_dentry_name(dname), std::move(alternate_name), ino, d_type, first, last);
 446   if (is_auth())
 447     dn->state_set(CDentry::STATE_AUTH);
 448   mdcache->lru.lru_insert_mid(dn);
 449
 450   dn->dir = this;
 451   dn->version = get_projected_version();
 452
 453   // add to dir
 454   ceph_assert(items.count(dn->key()) == 0);
 455   //assert(null_items.count(dn->get_name()) == 0);
 456
 457   items[dn->key()] = dn;
 458   if (last == CEPH_NOSNAP)
 459     num_head_items++;
 460   else
 461     num_snap_items++;
 462
 463   if (state_test(CDir::STATE_DNPINNEDFRAG)) {
 464     dn->get(CDentry::PIN_FRAGMENTING);
 465     dn->state_set(CDentry::STATE_FRAGMENTING);
 466   }
 467
 468   dout(12) << __func__ << " " << *dn << dendl;
 469
 470   // pin?
 471   if (get_num_any() == 1)
 472     get(PIN_CHILD);
 473
 474   ceph_assert(get_num_any() == items.size());
 475   return dn;
 476 }
 477
 478
 479
 480 void CDir::remove_dentry(CDentry *dn)
 481 {
 482   dout(12) << __func__ << " " << *dn << dendl;
 483
 484   // there should be no client leases at this point!
 485   ceph_assert(dn->client_lease_map.empty());
 486
 487   if (state_test(CDir::STATE_DNPINNEDFRAG)) {
 488     dn->put(CDentry::PIN_FRAGMENTING);
 489     dn->state_clear(CDentry::STATE_FRAGMENTING);
 490   }
 491
 492   if (dn->get_linkage()->is_null()) {
 493     if (dn->last == CEPH_NOSNAP)
 494       num_head_null--;
 495     else
 496       num_snap_null--;
 497   } else {
 498     if (dn->last == CEPH_NOSNAP)
 499       num_head_items--;
 500     else
 501       num_snap_items--;
 502   }
 503
 504   if (!dn->get_linkage()->is_null())
 505     // detach inode and dentry
 506     unlink_inode_work(dn);
 507
 508   // remove from list
 509   ceph_assert(items.count(dn->key()) == 1);
 510   items.erase(dn->key());
 511
 512   // clean?
 513   if (dn->is_dirty())
 514     dn->mark_clean();
 515
 516   if (dn->state_test(CDentry::STATE_BOTTOMLRU))
 517     mdcache->bottom_lru.lru_remove(dn);
 518   else
 519     mdcache->lru.lru_remove(dn);
 520   delete dn;
 521
 522   // unpin?
 523   if (get_num_any() == 0)
 524     put(PIN_CHILD);
 525   ceph_assert(get_num_any() == items.size());
 526 }
 527
 528 void CDir::link_remote_inode(CDentry *dn, CInode *in)
 529 {
 530   link_remote_inode(dn, in->ino(), IFTODT(in->get_projected_inode()->mode));
 531 }
 532
 533 void CDir::link_remote_inode(CDentry *dn, inodeno_t ino, unsigned char d_type)
 534 {
 535   dout(12) << __func__ << " " << *dn << " remote " << ino << dendl;
 536   ceph_assert(dn->get_linkage()->is_null());
 537
 538   dn->get_linkage()->set_remote(ino, d_type);
 539
 540   if (dn->state_test(CDentry::STATE_BOTTOMLRU)) {
 541     mdcache->bottom_lru.lru_remove(dn);
 542     mdcache->lru.lru_insert_mid(dn);
 543     dn->state_clear(CDentry::STATE_BOTTOMLRU);
 544   }
 545
 546   if (dn->last == CEPH_NOSNAP) {
 547     num_head_items++;
 548     num_head_null--;
 549   } else {
 550     num_snap_items++;
 551     num_snap_null--;
 552   }
 553   ceph_assert(get_num_any() == items.size());
 554 }
 555
 556 void CDir::link_primary_inode(CDentry *dn, CInode *in)
 557 {
 558   dout(12) << __func__ << " " << *dn << " " << *in << dendl;
 559   ceph_assert(dn->get_linkage()->is_null());
 560
 561   dn->get_linkage()->inode = in;
 562
 563   link_inode_work(dn, in);
 564
 565   if (dn->state_test(CDentry::STATE_BOTTOMLRU) &&
 566       (is_auth() || !inode->is_stray())) {
 567     mdcache->bottom_lru.lru_remove(dn);
 568     mdcache->lru.lru_insert_mid(dn);
 569     dn->state_clear(CDentry::STATE_BOTTOMLRU);
 570   }
 571
 572   if (dn->last == CEPH_NOSNAP) {
 573     num_head_items++;
 574     num_head_null--;
 575   } else {
 576     num_snap_items++;
 577     num_snap_null--;
 578   }
 579
 580   ceph_assert(get_num_any() == items.size());
 581 }
 582
 583 void CDir::link_inode_work( CDentry *dn, CInode *in)
 584 {
 585   ceph_assert(dn->get_linkage()->get_inode() == in);
 586   in->set_primary_parent(dn);
 587
 588   // set inode version
 589   //in->inode.version = dn->get_version();
 590
 591   // pin dentry?
 592   if (in->get_num_ref())
 593     dn->get(CDentry::PIN_INODEPIN);
 594
 595   if (in->state_test(CInode::STATE_TRACKEDBYOFT))
 596     mdcache->open_file_table.notify_link(in);
 597   if (in->is_any_caps())
 598     adjust_num_inodes_with_caps(1);
 599
 600   // adjust auth pin count
 601   if (in->auth_pins)
 602     dn->adjust_nested_auth_pins(in->auth_pins, NULL);
 603
 604   if (in->is_freezing_inode())
 605     freezing_inodes.push_back(&in->item_freezing_inode);
 606   else if (in->is_frozen_inode() || in->is_frozen_auth_pin())
 607     num_frozen_inodes++;
 608
 609   // verify open snaprealm parent
 610   if (in->snaprealm)
 611     in->snaprealm->adjust_parent();
 612   else if (in->is_any_caps())
 613     in->move_to_realm(inode->find_snaprealm());
 614 }
 615
 616 void CDir::unlink_inode(CDentry *dn, bool adjust_lru)
 617 {
 618   if (dn->get_linkage()->is_primary()) {
 619     dout(12) << __func__ << " " << *dn << " " << *dn->get_linkage()->get_inode() << dendl;
 620   } else {
 621     dout(12) << __func__ << " " << *dn << dendl;
 622   }
 623
 624   unlink_inode_work(dn);
 625
 626   if (adjust_lru && !dn->state_test(CDentry::STATE_BOTTOMLRU)) {
 627     mdcache->lru.lru_remove(dn);
 628     mdcache->bottom_lru.lru_insert_mid(dn);
 629     dn->state_set(CDentry::STATE_BOTTOMLRU);
 630   }
 631
 632   if (dn->last == CEPH_NOSNAP) {
 633     num_head_items--;
 634     num_head_null++;
 635   } else {
 636     num_snap_items--;
 637     num_snap_null++;
 638   }
 639   ceph_assert(get_num_any() == items.size());
 640 }
 641
 642
 643 void CDir::try_remove_unlinked_dn(CDentry *dn)
 644 {
 645   ceph_assert(dn->dir == this);
 646   ceph_assert(dn->get_linkage()->is_null());
 647
 648   // no pins (besides dirty)?
 649   if (dn->get_num_ref() != dn->is_dirty())
 650     return;
 651
 652   // was the dn new?
 653   if (dn->is_new()) {
 654     dout(10) << __func__ << " " << *dn << " in " << *this << dendl;
 655     if (dn->is_dirty())
 656       dn->mark_clean();
 657     remove_dentry(dn);
 658
 659     // NOTE: we may not have any more dirty dentries, but the fnode
 660     // still changed, so the directory must remain dirty.
 661   }
 662 }
 663
 664
 665 void CDir::unlink_inode_work(CDentry *dn)
 666 {
 667   CInode *in = dn->get_linkage()->get_inode();
 668
 669   if (dn->get_linkage()->is_remote()) {
 670     // remote
 671     if (in)
 672       dn->unlink_remote(dn->get_linkage());
 673
 674     dn->get_linkage()->set_remote(0, 0);
 675   } else if (dn->get_linkage()->is_primary()) {
 676     // primary
 677     // unpin dentry?
 678     if (in->get_num_ref())
 679       dn->put(CDentry::PIN_INODEPIN);
 680
 681     if (in->state_test(CInode::STATE_TRACKEDBYOFT))
 682       mdcache->open_file_table.notify_unlink(in);
 683     if (in->is_any_caps())
 684       adjust_num_inodes_with_caps(-1);
 685
 686     // unlink auth_pin count
 687     if (in->auth_pins)
 688       dn->adjust_nested_auth_pins(-in->auth_pins, nullptr);
 689
 690     if (in->is_freezing_inode())
 691       in->item_freezing_inode.remove_myself();
 692     else if (in->is_frozen_inode() || in->is_frozen_auth_pin())
 693       num_frozen_inodes--;
 694
 695     // detach inode
 696     in->remove_primary_parent(dn);
 697     if (in->is_dir())
 698       in->item_pop_lru.remove_myself();
 699     dn->get_linkage()->inode = 0;
 700   } else {
 701     ceph_assert(!dn->get_linkage()->is_null());
 702   }
 703 }
 704
 705 void CDir::add_to_bloom(CDentry *dn)
 706 {
 707   ceph_assert(dn->last == CEPH_NOSNAP);
 708   if (!bloom) {
 709     /* not create bloom filter for incomplete dir that was added by log replay */
 710     if (!is_complete())
 711       return;
 712
 713     /* don't maintain bloom filters in standby replay (saves cycles, and also
 714      * avoids need to implement clearing it in EExport for #16924) */
 715     if (mdcache->mds->is_standby_replay()) {
 716       return;
 717     }
 718
 719     unsigned size = get_num_head_items() + get_num_snap_items();
 720     if (size < 100) size = 100;
 721     bloom.reset(new bloom_filter(size, 1.0 / size, 0));
 722   }
 723   /* This size and false positive probability is completely random.*/
 724   bloom->insert(dn->get_name().data(), dn->get_name().size());
 725 }
 726
 727 bool CDir::is_in_bloom(std::string_view name)
 728 {
 729   if (!bloom)
 730     return false;
 731   return bloom->contains(name.data(), name.size());
 732 }
 733
 734 void CDir::remove_null_dentries() {
 735   dout(12) << __func__ << " " << *this << dendl;
 736
 737   auto p = items.begin();
 738   while (p != items.end()) {
 739     CDentry *dn = p->second;
 740     ++p;
 741     if (dn->get_linkage()->is_null() && !dn->is_projected())
 742       remove_dentry(dn);
 743   }
 744
 745   ceph_assert(num_snap_null == 0);
 746   ceph_assert(num_head_null == 0);
 747   ceph_assert(get_num_any() == items.size());
 748 }
 749
 750 /** remove dirty null dentries for deleted directory. the dirfrag will be
 751  *  deleted soon, so it's safe to not commit dirty dentries.
 752  *
 753  *  This is called when a directory is being deleted, a prerequisite
 754  *  of which is that its children have been unlinked: we expect to only see
 755  *  null, unprojected dentries here.
 756  */
 757 void CDir::try_remove_dentries_for_stray()
 758 {
 759   dout(10) << __func__ << dendl;
 760   ceph_assert(get_parent_dir()->inode->is_stray());
 761
 762   // clear dirty only when the directory was not snapshotted
 763   bool clear_dirty = !inode->snaprealm;
 764
 765   auto p = items.begin();
 766   while (p != items.end()) {
 767     CDentry *dn = p->second;
 768     ++p;
 769     if (dn->last == CEPH_NOSNAP) {
 770       ceph_assert(!dn->is_projected());
 771       ceph_assert(dn->get_linkage()->is_null());
 772       if (clear_dirty && dn->is_dirty())
 773         dn->mark_clean();
 774       // It's OK to remove lease prematurely because we will never link
 775       // the dentry to inode again.
 776       if (dn->is_any_leases())
 777         dn->remove_client_leases(mdcache->mds->locker);
 778       if (dn->get_num_ref() == 0)
 779         remove_dentry(dn);
 780     } else {
 781       ceph_assert(!dn->is_projected());
 782       CDentry::linkage_t *dnl= dn->get_linkage();
 783       CInode *in = NULL;
 784       if (dnl->is_primary()) {
 785         in = dnl->get_inode();
 786         if (clear_dirty && in->is_dirty())
 787           in->mark_clean();
 788       }
 789       if (clear_dirty && dn->is_dirty())
 790         dn->mark_clean();
 791       if (dn->get_num_ref() == 0) {
 792         remove_dentry(dn);
 793         if (in)
 794           mdcache->remove_inode(in);
 795       }
 796     }
 797   }
 798
 799   if (clear_dirty && is_dirty())
 800     mark_clean();
 801 }
 802
 803 bool CDir::try_trim_snap_dentry(CDentry *dn, const set<snapid_t>& snaps)
 804 {
 805   ceph_assert(dn->last != CEPH_NOSNAP);
 806   set<snapid_t>::const_iterator p = snaps.lower_bound(dn->first);
 807   CDentry::linkage_t *dnl= dn->get_linkage();
 808   CInode *in = 0;
 809   if (dnl->is_primary())
 810     in = dnl->get_inode();
 811   if ((p == snaps.end() || *p > dn->last) &&
 812       (dn->get_num_ref() == dn->is_dirty()) &&
 813       (!in || in->get_num_ref() == in->is_dirty())) {
 814     dout(10) << " purging snapped " << *dn << dendl;
 815     if (in && in->is_dirty())
 816       in->mark_clean();
 817     remove_dentry(dn);
 818     if (in) {
 819       dout(10) << " purging snapped " << *in << dendl;
 820       mdcache->remove_inode(in);
 821     }
 822     return true;
 823   }
 824   return false;
 825 }
 826
 827
 828 void CDir::purge_stale_snap_data(const set<snapid_t>& snaps)
 829 {
 830   dout(10) << __func__ << " " << snaps << dendl;
 831
 832   auto p = items.begin();
 833   while (p != items.end()) {
 834     CDentry *dn = p->second;
 835     ++p;
 836
 837     if (dn->last == CEPH_NOSNAP)
 838       continue;
 839
 840     try_trim_snap_dentry(dn, snaps);
 841   }
 842 }
 843
 844
 845 /**
 846  * steal_dentry -- semi-violently move a dentry from one CDir to another
 847  * (*) violently, in that nitems, most pins, etc. are not correctly maintained
 848  * on the old CDir corpse; must call finish_old_fragment() when finished.
 849  */
 850 void CDir::steal_dentry(CDentry *dn)
 851 {
 852   dout(15) << __func__ << " " << *dn << dendl;
 853
 854   items[dn->key()] = dn;
 855
 856   dn->dir->items.erase(dn->key());
 857   if (dn->dir->items.empty())
 858     dn->dir->put(PIN_CHILD);
 859
 860   if (get_num_any() == 0)
 861     get(PIN_CHILD);
 862   if (dn->get_linkage()->is_null()) {
 863     if (dn->last == CEPH_NOSNAP)
 864       num_head_null++;
 865     else
 866       num_snap_null++;
 867   } else if (dn->last == CEPH_NOSNAP) {
 868       num_head_items++;
 869
 870     auto _fnode = _get_fnode();
 871
 872     if (dn->get_linkage()->is_primary()) {
 873       CInode *in = dn->get_linkage()->get_inode();
 874       const auto& pi = in->get_projected_inode();
 875       if (in->is_dir()) {
 876         _fnode->fragstat.nsubdirs++;
 877         if (in->item_pop_lru.is_on_list())
 878           pop_lru_subdirs.push_back(&in->item_pop_lru);
 879       } else {
 880         _fnode->fragstat.nfiles++;
 881       }
 882       _fnode->rstat.rbytes += pi->accounted_rstat.rbytes;
 883       _fnode->rstat.rfiles += pi->accounted_rstat.rfiles;
 884       _fnode->rstat.rsubdirs += pi->accounted_rstat.rsubdirs;
 885       _fnode->rstat.rsnaps += pi->accounted_rstat.rsnaps;
 886       if (pi->accounted_rstat.rctime > fnode->rstat.rctime)
 887         _fnode->rstat.rctime = pi->accounted_rstat.rctime;
 888
 889       if (in->is_any_caps())
 890         adjust_num_inodes_with_caps(1);
 891
 892       // move dirty inode rstat to new dirfrag
 893       if (in->is_dirty_rstat())
 894         dirty_rstat_inodes.push_back(&in->dirty_rstat_item);
 895     } else if (dn->get_linkage()->is_remote()) {
 896       if (dn->get_linkage()->get_remote_d_type() == DT_DIR)
 897         _fnode->fragstat.nsubdirs++;
 898       else
 899         _fnode->fragstat.nfiles++;
 900     }
 901   } else {
 902     num_snap_items++;
 903     if (dn->get_linkage()->is_primary()) {
 904       CInode *in = dn->get_linkage()->get_inode();
 905       if (in->is_dirty_rstat())
 906         dirty_rstat_inodes.push_back(&in->dirty_rstat_item);
 907     }
 908   }
 909
 910   {
 911     int dap = dn->get_num_dir_auth_pins();
 912     if (dap) {
 913       adjust_nested_auth_pins(dap, NULL);
 914       dn->dir->adjust_nested_auth_pins(-dap, NULL);
 915     }
 916   }
 917
 918   if (dn->is_dirty()) {
 919     dirty_dentries.push_back(&dn->item_dir_dirty);
 920     num_dirty++;
 921   }
 922
 923   dn->dir = this;
 924 }
 925
 926 void CDir::prepare_old_fragment(map<string_snap_t, MDSContext::vec >& dentry_waiters, bool replay)
 927 {
 928   // auth_pin old fragment for duration so that any auth_pinning
 929   // during the dentry migration doesn't trigger side effects
 930   if (!replay && is_auth())
 931     auth_pin(this);
 932
 933   if (!waiting_on_dentry.empty()) {
 934     for (const auto &p : waiting_on_dentry) {
 935       auto &e = dentry_waiters[p.first];
 936       for (const auto &waiter : p.second) {
 937         e.push_back(waiter);
 938       }
 939     }
 940     waiting_on_dentry.clear();
 941     put(PIN_DNWAITER);
 942   }
 943 }
 944
 945 void CDir::prepare_new_fragment(bool replay)
 946 {
 947   if (!replay && is_auth()) {
 948     _freeze_dir();
 949     mark_complete();
 950   }
 951   inode->add_dirfrag(this);
 952 }
 953
 954 void CDir::finish_old_fragment(MDSContext::vec& waiters, bool replay)
 955 {
 956   // take waiters _before_ unfreeze...
 957   if (!replay) {
 958     take_waiting(WAIT_ANY_MASK, waiters);
 959     if (is_auth()) {
 960       auth_unpin(this);  // pinned in prepare_old_fragment
 961       ceph_assert(is_frozen_dir());
 962       unfreeze_dir();
 963     }
 964   }
 965
 966   ceph_assert(dir_auth_pins == 0);
 967   ceph_assert(auth_pins == 0);
 968
 969   num_head_items = num_head_null = 0;
 970   num_snap_items = num_snap_null = 0;
 971   adjust_num_inodes_with_caps(-num_inodes_with_caps);
 972
 973   // this mirrors init_fragment_pins()
 974   if (is_auth())
 975     clear_replica_map();
 976   if (is_dirty())
 977     mark_clean();
 978   if (state_test(STATE_IMPORTBOUND))
 979     put(PIN_IMPORTBOUND);
 980   if (state_test(STATE_EXPORTBOUND))
 981     put(PIN_EXPORTBOUND);
 982   if (is_subtree_root())
 983     put(PIN_SUBTREE);
 984
 985   if (auth_pins > 0)
 986     put(PIN_AUTHPIN);
 987
 988   ceph_assert(get_num_ref() == (state_test(STATE_STICKY) ? 1:0));
 989 }
 990
 991 void CDir::init_fragment_pins()
 992 {
 993   if (is_replicated())
 994     get(PIN_REPLICATED);
 995   if (state_test(STATE_DIRTY))
 996     get(PIN_DIRTY);
 997   if (state_test(STATE_EXPORTBOUND))
 998     get(PIN_EXPORTBOUND);
 999   if (state_test(STATE_IMPORTBOUND))
1000     get(PIN_IMPORTBOUND);
1001   if (is_subtree_root())
1002     get(PIN_SUBTREE);
1003 }
1004
1005 void CDir::split(int bits, std::vector<CDir*>* subs, MDSContext::vec& waiters, bool replay)
1006 {
1007   dout(10) << "split by " << bits << " bits on " << *this << dendl;
1008
1009   ceph_assert(replay || is_complete() || !is_auth());
1010
1011   frag_vec_t frags;
1012   frag.split(bits, frags);
1013
1014   vector<CDir*> subfrags(1 << bits);
1015
1016   double fac = 1.0 / (double)(1 << bits);  // for scaling load vecs
1017
1018   version_t rstat_version = inode->get_projected_inode()->rstat.version;
1019   version_t dirstat_version = inode->get_projected_inode()->dirstat.version;
1020
1021   nest_info_t rstatdiff;
1022   frag_info_t fragstatdiff;
1023   if (fnode->accounted_rstat.version == rstat_version)
1024     rstatdiff.add_delta(fnode->accounted_rstat, fnode->rstat);
1025   if (fnode->accounted_fragstat.version == dirstat_version)
1026     fragstatdiff.add_delta(fnode->accounted_fragstat, fnode->fragstat);
1027   dout(10) << " rstatdiff " << rstatdiff << " fragstatdiff " << fragstatdiff << dendl;
1028
1029   map<string_snap_t, MDSContext::vec > dentry_waiters;
1030   prepare_old_fragment(dentry_waiters, replay);
1031
1032   // create subfrag dirs
1033   int n = 0;
1034   for (const auto& fg : frags) {
1035     CDir *f = new CDir(inode, fg, mdcache, is_auth());
1036     f->state_set(state & (MASK_STATE_FRAGMENT_KEPT | STATE_COMPLETE));
1037     f->get_replicas() = get_replicas();
1038     f->pop_me = pop_me;
1039     f->pop_me.scale(fac);
1040
1041     // FIXME; this is an approximation
1042     f->pop_nested = pop_nested;
1043     f->pop_nested.scale(fac);
1044     f->pop_auth_subtree = pop_auth_subtree;
1045     f->pop_auth_subtree.scale(fac);
1046     f->pop_auth_subtree_nested = pop_auth_subtree_nested;
1047     f->pop_auth_subtree_nested.scale(fac);
1048
1049     dout(10) << " subfrag " << fg << " " << *f << dendl;
1050     subfrags[n++] = f;
1051     subs->push_back(f);
1052
1053     f->set_dir_auth(get_dir_auth());
1054     f->freeze_tree_state = freeze_tree_state;
1055     f->prepare_new_fragment(replay);
1056     f->init_fragment_pins();
1057   }
1058
1059   // repartition dentries
1060   while (!items.empty()) {
1061     auto p = items.begin();
1062
1063     CDentry *dn = p->second;
1064     frag_t subfrag = inode->pick_dirfrag(dn->get_name());
1065     int n = (subfrag.value() & (subfrag.mask() ^ frag.mask())) >> subfrag.mask_shift();
1066     dout(15) << " subfrag " << subfrag << " n=" << n << " for " << p->first << dendl;
1067     CDir *f = subfrags[n];
1068     f->steal_dentry(dn);
1069   }
1070
1071   for (const auto &p : dentry_waiters) {
1072     frag_t subfrag = inode->pick_dirfrag(p.first.name);
1073     int n = (subfrag.value() & (subfrag.mask() ^ frag.mask())) >> subfrag.mask_shift();
1074     CDir *f = subfrags[n];
1075
1076     if (f->waiting_on_dentry.empty())
1077       f->get(PIN_DNWAITER);
1078     auto &e = f->waiting_on_dentry[p.first];
1079     for (const auto &waiter : p.second) {
1080       e.push_back(waiter);
1081     }
1082   }
1083
1084   // FIXME: handle dirty old rstat
1085
1086   // fix up new frag fragstats
1087   for (int i = 0; i < n; i++) {
1088     CDir *f = subfrags[i];
1089     auto _fnode = f->_get_fnode();
1090     _fnode->version = f->projected_version = get_version();
1091     _fnode->rstat.version = rstat_version;
1092     _fnode->accounted_rstat = _fnode->rstat;
1093     _fnode->fragstat.version = dirstat_version;
1094     _fnode->accounted_fragstat = _fnode->fragstat;
1095     dout(10) << " rstat " << _fnode->rstat << " fragstat " << _fnode->fragstat
1096              << " on " << *f << dendl;
1097
1098     if (i == 0) {
1099       // give any outstanding frag stat differential to first frag
1100       dout(10) << " giving rstatdiff " << rstatdiff << " fragstatdiff" << fragstatdiff
1101                << " to " << *subfrags[0] << dendl;
1102       _fnode->accounted_rstat.add(rstatdiff);
1103       _fnode->accounted_fragstat.add(fragstatdiff);
1104     }
1105   }
1106
1107   finish_old_fragment(waiters, replay);
1108 }
1109
1110 void CDir::merge(const std::vector<CDir*>& subs, MDSContext::vec& waiters, bool replay)
1111 {
1112   dout(10) << "merge " << subs << dendl;
1113
1114   ceph_assert(subs.size() > 0);
1115
1116   set_dir_auth(subs.front()->get_dir_auth());
1117   freeze_tree_state = subs.front()->freeze_tree_state;
1118
1119   for (const auto& dir : subs) {
1120     ceph_assert(get_dir_auth() == dir->get_dir_auth());
1121     ceph_assert(freeze_tree_state == dir->freeze_tree_state);
1122   }
1123
1124   prepare_new_fragment(replay);
1125
1126   auto _fnode = _get_fnode();
1127
1128   nest_info_t rstatdiff;
1129   frag_info_t fragstatdiff;
1130   bool touched_mtime, touched_chattr;
1131   version_t rstat_version = inode->get_projected_inode()->rstat.version;
1132   version_t dirstat_version = inode->get_projected_inode()->dirstat.version;
1133
1134   map<string_snap_t, MDSContext::vec > dentry_waiters;
1135
1136   for (const auto& dir : subs) {
1137     dout(10) << " subfrag " << dir->get_frag() << " " << *dir << dendl;
1138     ceph_assert(!dir->is_auth() || dir->is_complete() || replay);
1139
1140     if (dir->get_fnode()->accounted_rstat.version == rstat_version)
1141       rstatdiff.add_delta(dir->get_fnode()->accounted_rstat, dir->get_fnode()->rstat);
1142     if (dir->get_fnode()->accounted_fragstat.version == dirstat_version)
1143       fragstatdiff.add_delta(dir->get_fnode()->accounted_fragstat, dir->get_fnode()->fragstat,
1144                              &touched_mtime, &touched_chattr);
1145
1146     dir->prepare_old_fragment(dentry_waiters, replay);
1147
1148     // steal dentries
1149     while (!dir->items.empty())
1150       steal_dentry(dir->items.begin()->second);
1151
1152     // merge replica map
1153     for (const auto &p : dir->get_replicas()) {
1154       unsigned cur = get_replicas()[p.first];
1155       if (p.second > cur)
1156         get_replicas()[p.first] = p.second;
1157     }
1158
1159     // merge version
1160     if (dir->get_version() > _fnode->version)
1161       _fnode->version = projected_version = dir->get_version();
1162
1163     // merge state
1164     state_set(dir->get_state() & MASK_STATE_FRAGMENT_KEPT);
1165
1166     dir->finish_old_fragment(waiters, replay);
1167     inode->close_dirfrag(dir->get_frag());
1168   }
1169
1170   if (!dentry_waiters.empty()) {
1171     get(PIN_DNWAITER);
1172     for (const auto &p : dentry_waiters) {
1173       auto &e = waiting_on_dentry[p.first];
1174       for (const auto &waiter : p.second) {
1175         e.push_back(waiter);
1176       }
1177     }
1178   }
1179
1180   if (is_auth() && !replay)
1181     mark_complete();
1182
1183   // FIXME: merge dirty old rstat
1184   _fnode->rstat.version = rstat_version;
1185   _fnode->accounted_rstat = _fnode->rstat;
1186   _fnode->accounted_rstat.add(rstatdiff);
1187
1188   _fnode->fragstat.version = dirstat_version;
1189   _fnode->accounted_fragstat = _fnode->fragstat;
1190   _fnode->accounted_fragstat.add(fragstatdiff);
1191
1192   init_fragment_pins();
1193 }
1194
1195
1196
1197
1198 void CDir::resync_accounted_fragstat()
1199 {
1200   auto pf = _get_projected_fnode();
1201   const auto& pi = inode->get_projected_inode();
1202
1203   if (pf->accounted_fragstat.version != pi->dirstat.version) {
1204     pf->fragstat.version = pi->dirstat.version;
1205     dout(10) << __func__ << " " << pf->accounted_fragstat << " -> " << pf->fragstat << dendl;
1206     pf->accounted_fragstat = pf->fragstat;
1207   }
1208 }
1209
1210 /*
1211  * resync rstat and accounted_rstat with inode
1212  */
1213 void CDir::resync_accounted_rstat()
1214 {
1215   auto pf = _get_projected_fnode();
1216   const auto& pi = inode->get_projected_inode();
1217
1218   if (pf->accounted_rstat.version != pi->rstat.version) {
1219     pf->rstat.version = pi->rstat.version;
1220     dout(10) << __func__ << " " << pf->accounted_rstat << " -> " << pf->rstat << dendl;
1221     pf->accounted_rstat = pf->rstat;
1222     dirty_old_rstat.clear();
1223   }
1224 }
1225
1226 void CDir::assimilate_dirty_rstat_inodes(MutationRef& mut)
1227 {
1228   dout(10) << __func__ << dendl;
1229   for (elist<CInode*>::iterator p = dirty_rstat_inodes.begin_use_current();
1230        !p.end(); ++p) {
1231     CInode *in = *p;
1232     ceph_assert(in->is_auth());
1233     if (in->is_frozen())
1234       continue;
1235
1236     mut->auth_pin(in);
1237
1238     auto pi = in->project_inode(mut);
1239     pi.inode->version = in->pre_dirty();
1240
1241     mdcache->project_rstat_inode_to_frag(mut, in, this, 0, 0, nullptr);
1242   }
1243   state_set(STATE_ASSIMRSTAT);
1244   dout(10) << __func__ << " done" << dendl;
1245 }
1246
1247 void CDir::assimilate_dirty_rstat_inodes_finish(EMetaBlob *blob)
1248 {
1249   if (!state_test(STATE_ASSIMRSTAT))
1250     return;
1251   state_clear(STATE_ASSIMRSTAT);
1252   dout(10) << __func__ << dendl;
1253   elist<CInode*>::iterator p = dirty_rstat_inodes.begin_use_current();
1254   while (!p.end()) {
1255     CInode *in = *p;
1256     ++p;
1257
1258     if (in->is_frozen())
1259       continue;
1260
1261     CDentry *dn = in->get_projected_parent_dn();
1262
1263     in->clear_dirty_rstat();
1264     blob->add_primary_dentry(dn, in, true);
1265   }
1266
1267   if (!dirty_rstat_inodes.empty())
1268     mdcache->mds->locker->mark_updated_scatterlock(&inode->nestlock);
1269 }
1270
1271
1272
1273
1274 /****************************************
1275  * WAITING
1276  */
1277
1278 void CDir::add_dentry_waiter(std::string_view dname, snapid_t snapid, MDSContext *c)
1279 {
1280   if (waiting_on_dentry.empty())
1281     get(PIN_DNWAITER);
1282   waiting_on_dentry[string_snap_t(dname, snapid)].push_back(c);
1283   dout(10) << __func__ << " dentry " << dname
1284            << " snap " << snapid
1285            << " " << c << " on " << *this << dendl;
1286 }
1287
1288 void CDir::take_dentry_waiting(std::string_view dname, snapid_t first, snapid_t last,
1289                                MDSContext::vec& ls)
1290 {
1291   if (waiting_on_dentry.empty())
1292     return;
1293
1294   string_snap_t lb(dname, first);
1295   string_snap_t ub(dname, last);
1296   auto it = waiting_on_dentry.lower_bound(lb);
1297   while (it != waiting_on_dentry.end() &&
1298          !(ub < it->first)) {
1299     dout(10) << __func__ << " " << dname
1300              << " [" << first << "," << last << "] found waiter on snap "
1301              << it->first.snapid
1302              << " on " << *this << dendl;
1303     for (const auto &waiter : it->second) {
1304       ls.push_back(waiter);
1305     }
1306     waiting_on_dentry.erase(it++);
1307   }
1308
1309   if (waiting_on_dentry.empty())
1310     put(PIN_DNWAITER);
1311 }
1312
1313 void CDir::take_sub_waiting(MDSContext::vec& ls)
1314 {
1315   dout(10) << __func__ << dendl;
1316   if (!waiting_on_dentry.empty()) {
1317     for (const auto &p : waiting_on_dentry) {
1318       for (const auto &waiter : p.second) {
1319         ls.push_back(waiter);
1320       }
1321     }
1322     waiting_on_dentry.clear();
1323     put(PIN_DNWAITER);
1324   }
1325 }
1326
1327
1328
1329 void CDir::add_waiter(uint64_t tag, MDSContext *c)
1330 {
1331   // hierarchical?
1332
1333   // at subtree root?
1334   if (tag & WAIT_ATSUBTREEROOT) {
1335     if (!is_subtree_root()) {
1336       // try parent
1337       dout(10) << "add_waiter " << std::hex << tag << std::dec << " " << c << " should be ATSUBTREEROOT, " << *this << " is not root, trying parent" << dendl;
1338       inode->parent->dir->add_waiter(tag, c);
1339       return;
1340     }
1341   }
1342
1343   ceph_assert(!(tag & WAIT_CREATED) || state_test(STATE_CREATING));
1344
1345   MDSCacheObject::add_waiter(tag, c);
1346 }
1347
1348
1349
1350 /* NOTE: this checks dentry waiters too */
1351 void CDir::take_waiting(uint64_t mask, MDSContext::vec& ls)
1352 {
1353   if ((mask & WAIT_DENTRY) && !waiting_on_dentry.empty()) {
1354     // take all dentry waiters
1355     for (const auto &p : waiting_on_dentry) {
1356       dout(10) << "take_waiting dentry " << p.first.name
1357                << " snap " << p.first.snapid << " on " << *this << dendl;
1358       for (const auto &waiter : p.second) {
1359         ls.push_back(waiter);
1360       }
1361     }
1362     waiting_on_dentry.clear();
1363     put(PIN_DNWAITER);
1364   }
1365
1366   // waiting
1367   MDSCacheObject::take_waiting(mask, ls);
1368 }
1369
1370
1371 void CDir::finish_waiting(uint64_t mask, int result)
1372 {
1373   dout(11) << __func__ << " mask " << hex << mask << dec << " result " << result << " on " << *this << dendl;
1374
1375   MDSContext::vec finished;
1376   take_waiting(mask, finished);
1377   if (result < 0)
1378     finish_contexts(g_ceph_context, finished, result);
1379   else
1380     mdcache->mds->queue_waiters(finished);
1381 }
1382
1383
1384
1385 // dirty/clean
1386
1387 CDir::fnode_ptr CDir::project_fnode(const MutationRef& mut)
1388 {
1389   ceph_assert(get_version() != 0);
1390
1391   if (mut && mut->is_projected(this))
1392     return std::const_pointer_cast<fnode_t>(projected_fnode.back());
1393
1394   auto pf = allocate_fnode(*get_projected_fnode());
1395
1396   if (scrub_infop && scrub_infop->last_scrub_dirty) {
1397     pf->localized_scrub_stamp = scrub_infop->last_local.time;
1398     pf->localized_scrub_version = scrub_infop->last_local.version;
1399     pf->recursive_scrub_stamp = scrub_infop->last_recursive.time;
1400     pf->recursive_scrub_version = scrub_infop->last_recursive.version;
1401     scrub_infop->last_scrub_dirty = false;
1402     scrub_maybe_delete_info();
1403   }
1404
1405   projected_fnode.emplace_back(pf);
1406   if (mut)
1407     mut->add_projected_node(this);
1408   dout(10) << __func__ <<  " " << pf.get() << dendl;
1409   return pf;
1410 }
1411
1412 void CDir::pop_and_dirty_projected_fnode(LogSegment *ls, const MutationRef& mut)
1413 {
1414   ceph_assert(!projected_fnode.empty());
1415   auto pf = std::move(projected_fnode.front());
1416   dout(15) << __func__ << " " << pf.get() << " v" << pf->version << dendl;
1417
1418   projected_fnode.pop_front();
1419   if (mut)
1420     mut->remove_projected_node(this);
1421
1422   reset_fnode(std::move(pf));
1423   _mark_dirty(ls);
1424 }
1425
1426 version_t CDir::pre_dirty(version_t min)
1427 {
1428   if (min > projected_version)
1429     projected_version = min;
1430   ++projected_version;
1431   dout(10) << __func__ << " " << projected_version << dendl;
1432   return projected_version;
1433 }
1434
1435 void CDir::mark_dirty(LogSegment *ls, version_t pv)
1436 {
1437   ceph_assert(is_auth());
1438
1439   if (pv) {
1440     ceph_assert(get_version() < pv);
1441     ceph_assert(pv <= projected_version);
1442     ceph_assert(!projected_fnode.empty() &&
1443                 pv <= projected_fnode.front()->version);
1444   }
1445
1446   _mark_dirty(ls);
1447 }
1448
1449 void CDir::_mark_dirty(LogSegment *ls)
1450 {
1451   if (!state_test(STATE_DIRTY)) {
1452     dout(10) << __func__ << " (was clean) " << *this << " version " << get_version() << dendl;
1453     _set_dirty_flag();
1454     ceph_assert(ls);
1455   } else {
1456     dout(10) << __func__ << " (already dirty) " << *this << " version " << get_version() << dendl;
1457   }
1458   if (ls) {
1459     ls->dirty_dirfrags.push_back(&item_dirty);
1460
1461     // if i've never committed, i need to be before _any_ mention of me is trimmed from the journal.
1462     if (committed_version == 0 && !item_new.is_on_list())
1463       ls->new_dirfrags.push_back(&item_new);
1464   }
1465 }
1466
1467 void CDir::mark_new(LogSegment *ls)
1468 {
1469   ls->new_dirfrags.push_back(&item_new);
1470   state_clear(STATE_CREATING);
1471
1472   MDSContext::vec waiters;
1473   take_waiting(CDir::WAIT_CREATED, waiters);
1474   mdcache->mds->queue_waiters(waiters);
1475 }
1476
1477 void CDir::mark_clean()
1478 {
1479   dout(10) << __func__ << " " << *this << " version " << get_version() << dendl;
1480   if (state_test(STATE_DIRTY)) {
1481     item_dirty.remove_myself();
1482     item_new.remove_myself();
1483
1484     state_clear(STATE_DIRTY);
1485     put(PIN_DIRTY);
1486   }
1487 }
1488
1489 // caller should hold auth pin of this
1490 void CDir::log_mark_dirty()
1491 {
1492   if (is_dirty() || projected_version > get_version())
1493     return; // noop if it is already dirty or will be dirty
1494
1495   auto _fnode = allocate_fnode(*get_fnode());
1496   _fnode->version = pre_dirty();
1497   reset_fnode(std::move(_fnode));
1498   mark_dirty(mdcache->mds->mdlog->get_current_segment());
1499 }
1500
1501 void CDir::mark_complete() {
1502   state_set(STATE_COMPLETE);
1503   bloom.reset();
1504 }
1505
1506 void CDir::first_get()
1507 {
1508   inode->get(CInode::PIN_DIRFRAG);
1509 }
1510
1511 void CDir::last_put()
1512 {
1513   inode->put(CInode::PIN_DIRFRAG);
1514 }
1515
1516
1517
1518 /******************************************************************************
1519  * FETCH and COMMIT
1520  */
1521
1522 // -----------------------
1523 // FETCH
1524 void CDir::fetch(MDSContext *c, bool ignore_authpinnability)
1525 {
1526   string want;
1527   return fetch(c, want, ignore_authpinnability);
1528 }
1529
1530 void CDir::fetch(MDSContext *c, std::string_view want_dn, bool ignore_authpinnability)
1531 {
1532   dout(10) << "fetch on " << *this << dendl;
1533
1534   ceph_assert(is_auth());
1535   ceph_assert(!is_complete());
1536
1537   if (!can_auth_pin() && !ignore_authpinnability) {
1538     if (c) {
1539       dout(7) << "fetch waiting for authpinnable" << dendl;
1540       add_waiter(WAIT_UNFREEZE, c);
1541     } else
1542       dout(7) << "fetch not authpinnable and no context" << dendl;
1543     return;
1544   }
1545
1546   // unlinked directory inode shouldn't have any entry
1547   if (!inode->is_base() && get_parent_dir()->inode->is_stray() &&
1548       !inode->snaprealm) {
1549     dout(7) << "fetch dirfrag for unlinked directory, mark complete" << dendl;
1550     if (get_version() == 0) {
1551       ceph_assert(inode->is_auth());
1552       auto _fnode = allocate_fnode();
1553       _fnode->version = 1;
1554       reset_fnode(std::move(_fnode));
1555
1556       if (state_test(STATE_REJOINUNDEF)) {
1557         ceph_assert(mdcache->mds->is_rejoin());
1558         state_clear(STATE_REJOINUNDEF);
1559         mdcache->opened_undef_dirfrag(this);
1560       }
1561     }
1562     mark_complete();
1563
1564     if (c)
1565       mdcache->mds->queue_waiter(c);
1566     return;
1567   }
1568
1569   if (c) add_waiter(WAIT_COMPLETE, c);
1570   if (!want_dn.empty()) wanted_items.insert(mempool::mds_co::string(want_dn));
1571
1572   // already fetching?
1573   if (state_test(CDir::STATE_FETCHING)) {
1574     dout(7) << "already fetching; waiting" << dendl;
1575     return;
1576   }
1577
1578   auth_pin(this);
1579   state_set(CDir::STATE_FETCHING);
1580
1581   if (mdcache->mds->logger) mdcache->mds->logger->inc(l_mds_dir_fetch);
1582
1583   mdcache->mds->balancer->hit_dir(this, META_POP_FETCH);
1584
1585   std::set<dentry_key_t> empty;
1586   _omap_fetch(NULL, empty);
1587 }
1588
1589 void CDir::fetch(MDSContext *c, const std::set<dentry_key_t>& keys)
1590 {
1591   dout(10) << "fetch " << keys.size() << " keys on " << *this << dendl;
1592
1593   ceph_assert(is_auth());
1594   ceph_assert(!is_complete());
1595
1596   if (!can_auth_pin()) {
1597     dout(7) << "fetch keys waiting for authpinnable" << dendl;
1598     add_waiter(WAIT_UNFREEZE, c);
1599     return;
1600   }
1601   if (state_test(CDir::STATE_FETCHING)) {
1602     dout(7) << "fetch keys waiting for full fetch" << dendl;
1603     add_waiter(WAIT_COMPLETE, c);
1604     return;
1605   }
1606
1607   auth_pin(this);
1608   if (mdcache->mds->logger) mdcache->mds->logger->inc(l_mds_dir_fetch);
1609
1610   mdcache->mds->balancer->hit_dir(this, META_POP_FETCH);
1611
1612   _omap_fetch(c, keys);
1613 }
1614
1615 class C_IO_Dir_OMAP_FetchedMore : public CDirIOContext {
1616   MDSContext *fin;
1617 public:
1618   const version_t omap_version;
1619   bufferlist hdrbl;
1620   bool more = false;
1621   map<string, bufferlist> omap;      ///< carry-over from before
1622   map<string, bufferlist> omap_more; ///< new batch
1623   int ret;
1624   C_IO_Dir_OMAP_FetchedMore(CDir *d, version_t v, MDSContext *f) :
1625     CDirIOContext(d), fin(f), omap_version(v), ret(0) { }
1626   void finish(int r) {
1627     if (omap_version < dir->get_committed_version()) {
1628       omap.clear();
1629       dir->_omap_fetch(fin, {});
1630       return;
1631     }
1632
1633     // merge results
1634     if (omap.empty()) {
1635       omap.swap(omap_more);
1636     } else {
1637       omap.insert(omap_more.begin(), omap_more.end());
1638     }
1639     if (more) {
1640       dir->_omap_fetch_more(omap_version, hdrbl, omap, fin);
1641     } else {
1642       dir->_omap_fetched(hdrbl, omap, !fin, r);
1643       if (fin)
1644         fin->complete(r);
1645     }
1646   }
1647   void print(ostream& out) const override {
1648     out << "dirfrag_fetch_more(" << dir->dirfrag() << ")";
1649   }
1650 };
1651
1652 class C_IO_Dir_OMAP_Fetched : public CDirIOContext {
1653   MDSContext *fin;
1654 public:
1655   const version_t omap_version;
1656   bufferlist hdrbl;
1657   bool more = false;
1658   map<string, bufferlist> omap;
1659   bufferlist btbl;
1660   int ret1, ret2, ret3;
1661
1662   C_IO_Dir_OMAP_Fetched(CDir *d, MDSContext *f) :
1663     CDirIOContext(d), fin(f),
1664     omap_version(d->get_committing_version()),
1665     ret1(0), ret2(0), ret3(0) { }
1666   void finish(int r) override {
1667     // check the correctness of backtrace
1668     if (r >= 0 && ret3 != -CEPHFS_ECANCELED)
1669       dir->inode->verify_diri_backtrace(btbl, ret3);
1670     if (r >= 0) r = ret1;
1671     if (r >= 0) r = ret2;
1672
1673     if (more) {
1674       if (omap_version < dir->get_committed_version()) {
1675         omap.clear();
1676         dir->_omap_fetch(fin, {});
1677       } else {
1678         dir->_omap_fetch_more(omap_version, hdrbl, omap, fin);
1679       }
1680       return;
1681     }
1682
1683     dir->_omap_fetched(hdrbl, omap, !fin, r);
1684     if (fin)
1685       fin->complete(r);
1686
1687   }
1688   void print(ostream& out) const override {
1689     out << "dirfrag_fetch(" << dir->dirfrag() << ")";
1690   }
1691 };
1692
1693 void CDir::_omap_fetch(MDSContext *c, const std::set<dentry_key_t>& keys)
1694 {
1695   C_IO_Dir_OMAP_Fetched *fin = new C_IO_Dir_OMAP_Fetched(this, c);
1696   object_t oid = get_ondisk_object();
1697   object_locator_t oloc(mdcache->mds->mdsmap->get_metadata_pool());
1698   ObjectOperation rd;
1699   rd.omap_get_header(&fin->hdrbl, &fin->ret1);
1700   if (keys.empty()) {
1701     ceph_assert(!c);
1702     rd.omap_get_vals("", "", g_conf()->mds_dir_keys_per_op,
1703                      &fin->omap, &fin->more, &fin->ret2);
1704   } else {
1705     ceph_assert(c);
1706     std::set<std::string> str_keys;
1707     for (auto p : keys) {
1708       string str;
1709       p.encode(str);
1710       str_keys.insert(str);
1711     }
1712     rd.omap_get_vals_by_keys(str_keys, &fin->omap, &fin->ret2);
1713   }
1714   // check the correctness of backtrace
1715   if (g_conf()->mds_verify_backtrace > 0 && frag == frag_t()) {
1716     rd.getxattr("parent", &fin->btbl, &fin->ret3);
1717     rd.set_last_op_flags(CEPH_OSD_OP_FLAG_FAILOK);
1718   } else {
1719     fin->ret3 = -CEPHFS_ECANCELED;
1720   }
1721
1722   mdcache->mds->objecter->read(oid, oloc, rd, CEPH_NOSNAP, NULL, 0,
1723                              new C_OnFinisher(fin, mdcache->mds->finisher));
1724 }
1725
1726 void CDir::_omap_fetch_more(version_t omap_version, bufferlist& hdrbl,
1727                             map<string, bufferlist>& omap, MDSContext *c)
1728 {
1729   // we have more omap keys to fetch!
1730   object_t oid = get_ondisk_object();
1731   object_locator_t oloc(mdcache->mds->mdsmap->get_metadata_pool());
1732   auto fin = new C_IO_Dir_OMAP_FetchedMore(this, omap_version, c);
1733   fin->hdrbl = std::move(hdrbl);
1734   fin->omap.swap(omap);
1735   ObjectOperation rd;
1736   rd.omap_get_vals(fin->omap.rbegin()->first,
1737                    "", /* filter prefix */
1738                    g_conf()->mds_dir_keys_per_op,
1739                    &fin->omap_more,
1740                    &fin->more,
1741                    &fin->ret);
1742   mdcache->mds->objecter->read(oid, oloc, rd, CEPH_NOSNAP, NULL, 0,
1743                              new C_OnFinisher(fin, mdcache->mds->finisher));
1744 }
1745
1746 CDentry *CDir::_load_dentry(
1747     std::string_view key,
1748     std::string_view dname,
1749     const snapid_t last,
1750     bufferlist &bl,
1751     const int pos,
1752     const std::set<snapid_t> *snaps,
1753     double rand_threshold,
1754     bool *force_dirty)
1755 {
1756   auto q = bl.cbegin();
1757
1758   snapid_t first;
1759   decode(first, q);
1760
1761   // marker
1762   char type;
1763   decode(type, q);
1764
1765   dout(20) << "_fetched pos " << pos << " marker '" << type << "' dname '" << dname
1766            << " [" << first << "," << last << "]"
1767            << dendl;
1768
1769   bool stale = false;
1770   if (snaps && last != CEPH_NOSNAP) {
1771     set<snapid_t>::const_iterator p = snaps->lower_bound(first);
1772     if (p == snaps->end() || *p > last) {
1773       dout(10) << " skipping stale dentry on [" << first << "," << last << "]" << dendl;
1774       stale = true;
1775     }
1776   }
1777
1778   /*
1779    * look for existing dentry for _last_ snap, because unlink +
1780    * create may leave a "hole" (epochs during which the dentry
1781    * doesn't exist) but for which no explicit negative dentry is in
1782    * the cache.
1783    */
1784   CDentry *dn;
1785   if (stale)
1786     dn = lookup_exact_snap(dname, last);
1787   else
1788     dn = lookup(dname, last);
1789
1790   if (type == 'L' || type == 'l') {
1791     // hard link
1792     inodeno_t ino;
1793     unsigned char d_type;
1794     mempool::mds_co::string alternate_name;
1795
1796     CDentry::decode_remote(type, ino, d_type, alternate_name, q);
1797
1798     if (stale) {
1799       if (!dn) {
1800         stale_items.insert(mempool::mds_co::string(key));
1801         *force_dirty = true;
1802       }
1803       return dn;
1804     }
1805
1806     if (dn) {
1807       CDentry::linkage_t *dnl = dn->get_linkage();
1808       dout(12) << "_fetched had " << (dnl->is_null() ? "NEG" : "") << " dentry " << *dn << dendl;
1809       if (committed_version == 0 &&
1810           dnl->is_remote() &&
1811           dn->is_dirty() &&
1812           ino == dnl->get_remote_ino() &&
1813           d_type == dnl->get_remote_d_type() &&
1814           alternate_name == dn->get_alternate_name()) {
1815         // see comment below
1816         dout(10) << "_fetched  had underwater dentry " << *dn << ", marking clean" << dendl;
1817         dn->mark_clean();
1818       }
1819     } else {
1820       // (remote) link
1821       dn = add_remote_dentry(dname, ino, d_type, std::move(alternate_name), first, last);
1822
1823       // link to inode?
1824       CInode *in = mdcache->get_inode(ino);   // we may or may not have it.
1825       if (in) {
1826         dn->link_remote(dn->get_linkage(), in);
1827         dout(12) << "_fetched  got remote link " << ino << " which we have " << *in << dendl;
1828       } else {
1829         dout(12) << "_fetched  got remote link " << ino << " (don't have it)" << dendl;
1830       }
1831     }
1832   }
1833   else if (type == 'I' || type == 'i') {
1834     InodeStore inode_data;
1835     mempool::mds_co::string alternate_name;
1836     // inode
1837     // Load inode data before looking up or constructing CInode
1838     if (type == 'i') {
1839       DECODE_START(2, q);
1840       if (struct_v >= 2) {
1841         decode(alternate_name, q);
1842       }
1843       inode_data.decode(q);
1844       DECODE_FINISH(q);
1845     } else {
1846       inode_data.decode_bare(q);
1847     }
1848
1849     if (stale) {
1850       if (!dn) {
1851         stale_items.insert(mempool::mds_co::string(key));
1852         *force_dirty = true;
1853       }
1854       return dn;
1855     }
1856
1857     bool undef_inode = false;
1858     if (dn) {
1859       CDentry::linkage_t *dnl = dn->get_linkage();
1860       dout(12) << "_fetched had " << (dnl->is_null() ? "NEG" : "") << " dentry " << *dn << dendl;
1861
1862       if (dnl->is_primary()) {
1863         CInode *in = dnl->get_inode();
1864         if (in->state_test(CInode::STATE_REJOINUNDEF)) {
1865           undef_inode = true;
1866         } else if (committed_version == 0 &&
1867                    dn->is_dirty() &&
1868                    inode_data.inode->ino == in->ino() &&
1869                    inode_data.inode->version == in->get_version()) {
1870           /* clean underwater item?
1871            * Underwater item is something that is dirty in our cache from
1872            * journal replay, but was previously flushed to disk before the
1873            * mds failed.
1874            *
1875            * We only do this is committed_version == 0. that implies either
1876            * - this is a fetch after from a clean/empty CDir is created
1877            *   (and has no effect, since the dn won't exist); or
1878            * - this is a fetch after _recovery_, which is what we're worried
1879            *   about.  Items that are marked dirty from the journal should be
1880            *   marked clean if they appear on disk.
1881            */
1882           dout(10) << "_fetched  had underwater dentry " << *dn << ", marking clean" << dendl;
1883           dn->mark_clean();
1884           dout(10) << "_fetched  had underwater inode " << *dnl->get_inode() << ", marking clean" << dendl;
1885           in->mark_clean();
1886         }
1887       }
1888     }
1889
1890     if (!dn || undef_inode) {
1891       // add inode
1892       CInode *in = mdcache->get_inode(inode_data.inode->ino, last);
1893       if (!in || undef_inode) {
1894         if (undef_inode && in)
1895           in->first = first;
1896         else
1897           in = new CInode(mdcache, true, first, last);
1898
1899         in->reset_inode(std::move(inode_data.inode));
1900         in->reset_xattrs(std::move(inode_data.xattrs));
1901         // symlink?
1902         if (in->is_symlink())
1903           in->symlink = inode_data.symlink;
1904
1905         in->dirfragtree.swap(inode_data.dirfragtree);
1906         in->reset_old_inodes(std::move(inode_data.old_inodes));
1907         if (in->is_any_old_inodes()) {
1908           snapid_t min_first = in->get_old_inodes()->rbegin()->first + 1;
1909           if (min_first > in->first)
1910             in->first = min_first;
1911         }
1912
1913         in->oldest_snap = inode_data.oldest_snap;
1914         in->decode_snap_blob(inode_data.snap_blob);
1915         if (snaps && !in->snaprealm)
1916           in->purge_stale_snap_data(*snaps);
1917
1918         if (!undef_inode) {
1919           mdcache->add_inode(in); // add
1920           dn = add_primary_dentry(dname, in, std::move(alternate_name), first, last); // link
1921         }
1922         dout(12) << "_fetched  got " << *dn << " " << *in << dendl;
1923
1924         if (in->get_inode()->is_dirty_rstat())
1925           in->mark_dirty_rstat();
1926
1927         in->maybe_ephemeral_rand(rand_threshold);
1928         //in->hack_accessed = false;
1929         //in->hack_load_stamp = ceph_clock_now();
1930         //num_new_inodes_loaded++;
1931       } else if (g_conf().get_val<bool>("mds_hack_allow_loading_invalid_metadata")) {
1932         dout(20) << "hack: adding duplicate dentry for " << *in << dendl;
1933         dn = add_primary_dentry(dname, in, std::move(alternate_name), first, last);
1934       } else {
1935         dout(0) << "_fetched  badness: got (but i already had) " << *in
1936                 << " mode " << in->get_inode()->mode
1937                 << " mtime " << in->get_inode()->mtime << dendl;
1938         string dirpath, inopath;
1939         this->inode->make_path_string(dirpath);
1940         in->make_path_string(inopath);
1941         mdcache->mds->clog->error() << "loaded dup inode " << inode_data.inode->ino
1942           << " [" << first << "," << last << "] v" << inode_data.inode->version
1943           << " at " << dirpath << "/" << dname
1944           << ", but inode " << in->vino() << " v" << in->get_version()
1945           << " already exists at " << inopath;
1946         return dn;
1947       }
1948     }
1949   } else {
1950     CachedStackStringStream css;
1951     *css << "Invalid tag char '" << type << "' pos " << pos;
1952     throw buffer::malformed_input(css->str());
1953   }
1954
1955   return dn;
1956 }
1957
1958 void CDir::_omap_fetched(bufferlist& hdrbl, map<string, bufferlist>& omap,
1959                          bool complete, int r)
1960 {
1961   LogChannelRef clog = mdcache->mds->clog;
1962   dout(10) << "_fetched header " << hdrbl.length() << " bytes "
1963            << omap.size() << " keys for " << *this << dendl;
1964
1965   ceph_assert(r == 0 || r == -CEPHFS_ENOENT || r == -CEPHFS_ENODATA);
1966   ceph_assert(is_auth());
1967   ceph_assert(!is_frozen());
1968
1969   if (hdrbl.length() == 0) {
1970     dout(0) << "_fetched missing object for " << *this << dendl;
1971
1972     clog->error() << "dir " << dirfrag() << " object missing on disk; some "
1973                      "files may be lost (" << get_path() << ")";
1974
1975     go_bad(complete);
1976     return;
1977   }
1978
1979   fnode_t got_fnode;
1980   {
1981     auto p = hdrbl.cbegin();
1982     try {
1983       decode(got_fnode, p);
1984     } catch (const buffer::error &err) {
1985       derr << "Corrupt fnode in dirfrag " << dirfrag()
1986            << ": " << err.what() << dendl;
1987       clog->warn() << "Corrupt fnode header in " << dirfrag() << ": "
1988                    << err.what() << " (" << get_path() << ")";
1989       go_bad(complete);
1990       return;
1991     }
1992     if (!p.end()) {
1993       clog->warn() << "header buffer of dir " << dirfrag() << " has "
1994                   << hdrbl.length() - p.get_off() << " extra bytes ("
1995                   << get_path() << ")";
1996       go_bad(complete);
1997       return;
1998     }
1999   }
2000
2001   dout(10) << "_fetched version " << got_fnode.version << dendl;
2002
2003   // take the loaded fnode?
2004   // only if we are a fresh CDir* with no prior state.
2005   if (get_version() == 0) {
2006     ceph_assert(!is_projected());
2007     ceph_assert(!state_test(STATE_COMMITTING));
2008     auto _fnode = allocate_fnode(got_fnode);
2009     reset_fnode(std::move(_fnode));
2010     projected_version = committing_version = committed_version = get_version();
2011
2012     if (state_test(STATE_REJOINUNDEF)) {
2013       ceph_assert(mdcache->mds->is_rejoin());
2014       state_clear(STATE_REJOINUNDEF);
2015       mdcache->opened_undef_dirfrag(this);
2016     }
2017   }
2018
2019   list<CInode*> undef_inodes;
2020
2021   // purge stale snaps?
2022   bool force_dirty = false;
2023   const set<snapid_t> *snaps = NULL;
2024   SnapRealm *realm = inode->find_snaprealm();
2025   if (fnode->snap_purged_thru < realm->get_last_destroyed()) {
2026     snaps = &realm->get_snaps();
2027     dout(10) << " snap_purged_thru " << fnode->snap_purged_thru
2028              << " < " << realm->get_last_destroyed()
2029              << ", snap purge based on " << *snaps << dendl;
2030     if (get_num_snap_items() == 0) {
2031       const_cast<snapid_t&>(fnode->snap_purged_thru) = realm->get_last_destroyed();
2032       force_dirty = true;
2033     }
2034   }
2035
2036   unsigned pos = omap.size() - 1;
2037   double rand_threshold = get_inode()->get_ephemeral_rand();
2038   for (map<string, bufferlist>::reverse_iterator p = omap.rbegin();
2039        p != omap.rend();
2040        ++p, --pos) {
2041     string dname;
2042     snapid_t last;
2043     dentry_key_t::decode_helper(p->first, dname, last);
2044
2045     CDentry *dn = NULL;
2046     try {
2047       dn = _load_dentry(
2048             p->first, dname, last, p->second, pos, snaps,
2049             rand_threshold, &force_dirty);
2050     } catch (const buffer::error &err) {
2051       mdcache->mds->clog->warn() << "Corrupt dentry '" << dname << "' in "
2052                                   "dir frag " << dirfrag() << ": "
2053                                << err.what() << "(" << get_path() << ")";
2054
2055       // Remember that this dentry is damaged.  Subsequent operations
2056       // that try to act directly on it will get their CEPHFS_EIOs, but this
2057       // dirfrag as a whole will continue to look okay (minus the
2058       // mysteriously-missing dentry)
2059       go_bad_dentry(last, dname);
2060
2061       // Anyone who was WAIT_DENTRY for this guy will get kicked
2062       // to RetryRequest, and hit the DamageTable-interrogating path.
2063       // Stats will now be bogus because we will think we're complete,
2064       // but have 1 or more missing dentries.
2065       continue;
2066     }
2067
2068     if (!dn)
2069       continue;
2070
2071     CDentry::linkage_t *dnl = dn->get_linkage();
2072     if (dnl->is_primary() && dnl->get_inode()->state_test(CInode::STATE_REJOINUNDEF))
2073       undef_inodes.push_back(dnl->get_inode());
2074
2075     if (wanted_items.count(mempool::mds_co::string(dname)) > 0 || !complete) {
2076       dout(10) << " touching wanted dn " << *dn << dendl;
2077       mdcache->touch_dentry(dn);
2078     }
2079   }
2080
2081   //cache->mds->logger->inc("newin", num_new_inodes_loaded);
2082
2083   // mark complete, !fetching
2084   if (complete) {
2085     wanted_items.clear();
2086     mark_complete();
2087     state_clear(STATE_FETCHING);
2088   }
2089
2090   // open & force frags
2091   while (!undef_inodes.empty()) {
2092     CInode *in = undef_inodes.front();
2093     undef_inodes.pop_front();
2094     in->state_clear(CInode::STATE_REJOINUNDEF);
2095     mdcache->opened_undef_inode(in);
2096   }
2097
2098   // dirty myself to remove stale snap dentries
2099   if (force_dirty && !mdcache->is_readonly())
2100     log_mark_dirty();
2101
2102   auth_unpin(this);
2103
2104   if (complete) {
2105     // kick waiters
2106     finish_waiting(WAIT_COMPLETE, 0);
2107   }
2108 }
2109
2110 void CDir::go_bad_dentry(snapid_t last, std::string_view dname)
2111 {
2112   dout(10) << __func__ << " " << dname << dendl;
2113   std::string path(get_path());
2114   path += "/";
2115   path += dname;
2116   const bool fatal = mdcache->mds->damage_table.notify_dentry(
2117       inode->ino(), frag, last, dname, path);
2118   if (fatal) {
2119     mdcache->mds->damaged();
2120     ceph_abort();  // unreachable, damaged() respawns us
2121   }
2122 }
2123
2124 void CDir::go_bad(bool complete)
2125 {
2126   dout(10) << __func__ << " " << frag << dendl;
2127   const bool fatal = mdcache->mds->damage_table.notify_dirfrag(
2128       inode->ino(), frag, get_path());
2129   if (fatal) {
2130     mdcache->mds->damaged();
2131     ceph_abort();  // unreachable, damaged() respawns us
2132   }
2133
2134   if (complete) {
2135     if (get_version() == 0) {
2136       auto _fnode = allocate_fnode();
2137       _fnode->version = 1;
2138       reset_fnode(std::move(_fnode));
2139     }
2140
2141     state_set(STATE_BADFRAG);
2142     mark_complete();
2143   }
2144
2145   state_clear(STATE_FETCHING);
2146   auth_unpin(this);
2147   finish_waiting(WAIT_COMPLETE, -CEPHFS_EIO);
2148 }
2149
2150 // -----------------------
2151 // COMMIT
2152
2153 /**
2154  * commit
2155  *
2156  * @param want - min version i want committed
2157  * @param c - callback for completion
2158  */
2159 void CDir::commit(version_t want, MDSContext *c, bool ignore_authpinnability, int op_prio)
2160 {
2161   dout(10) << "commit want " << want << " on " << *this << dendl;
2162   if (want == 0) want = get_version();
2163
2164   // preconditions
2165   ceph_assert(want <= get_version() || get_version() == 0);    // can't commit the future
2166   ceph_assert(want > committed_version); // the caller is stupid
2167   ceph_assert(is_auth());
2168   ceph_assert(ignore_authpinnability || can_auth_pin());
2169
2170   // note: queue up a noop if necessary, so that we always
2171   // get an auth_pin.
2172   if (!c)
2173     c = new C_MDSInternalNoop;
2174
2175   // auth_pin on first waiter
2176   if (waiting_for_commit.empty())
2177     auth_pin(this);
2178   waiting_for_commit[want].push_back(c);
2179
2180   // ok.
2181   _commit(want, op_prio);
2182 }
2183
2184 class C_IO_Dir_Committed : public CDirIOContext {
2185   version_t version;
2186 public:
2187   C_IO_Dir_Committed(CDir *d, version_t v) : CDirIOContext(d), version(v) { }
2188   void finish(int r) override {
2189     dir->_committed(r, version);
2190   }
2191   void print(ostream& out) const override {
2192     out << "dirfrag_committed(" << dir->dirfrag() << ")";
2193   }
2194 };
2195
2196 class C_IO_Dir_Commit_Ops : public Context {
2197 public:
2198   C_IO_Dir_Commit_Ops(CDir *d, int pr,
2199                       vector<CDir::dentry_commit_item> &&s, bufferlist &&bl,
2200                       vector<string> &&r,
2201                       mempool::mds_co::compact_set<mempool::mds_co::string> &&stales) :
2202     dir(d), op_prio(pr) {
2203     metapool = dir->mdcache->mds->get_metadata_pool();
2204     version = dir->get_version();
2205     is_new = dir->is_new();
2206     to_set.swap(s);
2207     dfts.swap(bl);
2208     to_remove.swap(r);
2209     stale_items.swap(stales);
2210   }
2211
2212   void finish(int r) override {
2213     dir->_omap_commit_ops(r, op_prio, metapool, version, is_new, to_set, dfts,
2214                           to_remove, stale_items);
2215   }
2216
2217 private:
2218   CDir *dir;
2219   int op_prio;
2220   int64_t metapool;
2221   version_t version;
2222   bool is_new;
2223   vector<CDir::dentry_commit_item> to_set;
2224   bufferlist dfts;
2225   vector<string> to_remove;
2226   mempool::mds_co::compact_set<mempool::mds_co::string> stale_items;
2227 };
2228
2229 // This is doing the same thing with the InodeStoreBase::encode()
2230 void CDir::_encode_primary_inode_base(dentry_commit_item &item, bufferlist &dfts,
2231                                       bufferlist &bl)
2232 {
2233   ENCODE_START(6, 4, bl);
2234   encode(*item.inode, bl, item.features);
2235
2236   if (!item.symlink.empty())
2237     encode(item.symlink, bl);
2238
2239   // dirfragtree
2240   dfts.splice(0, item.dft_len, &bl);
2241
2242   if (item.xattrs)
2243     encode(*item.xattrs, bl);
2244   else
2245     encode((__u32)0, bl);
2246
2247   if (item.snaprealm) {
2248     bufferlist snapr_bl;
2249     encode(item.srnode, snapr_bl);
2250     encode(snapr_bl, bl);
2251   } else {
2252     encode(bufferlist(), bl);
2253   }
2254
2255   if (item.old_inodes)
2256     encode(*item.old_inodes, bl, item.features);
2257   else
2258     encode((__u32)0, bl);
2259
2260   encode(item.oldest_snap, bl);
2261   encode(item.damage_flags, bl);
2262   ENCODE_FINISH(bl);
2263 }
2264
2265 // This is not locked by mds_lock
2266 void CDir::_omap_commit_ops(int r, int op_prio, int64_t metapool, version_t version, bool _new,
2267                             vector<dentry_commit_item> &to_set, bufferlist &dfts,
2268                             vector<string>& to_remove,
2269                             mempool::mds_co::compact_set<mempool::mds_co::string> &stales)
2270 {
2271   dout(10) << __func__ << dendl;
2272
2273   if (r < 0) {
2274     mdcache->mds->handle_write_error_with_lock(r);
2275     return;
2276   }
2277
2278   C_GatherBuilder gather(g_ceph_context,
2279                          new C_OnFinisher(new C_IO_Dir_Committed(this, version),
2280                          mdcache->mds->finisher));
2281
2282   SnapContext snapc;
2283   object_t oid = get_ondisk_object();
2284   object_locator_t oloc(metapool);
2285
2286   map<string, bufferlist> _set;
2287   set<string> _rm;
2288
2289   unsigned max_write_size = mdcache->max_dir_commit_size;
2290   unsigned write_size = 0;
2291
2292   auto commit_one = [&](bool header=false) {
2293     ObjectOperation op;
2294
2295     // don't create new dirfrag blindly
2296     if (!_new)
2297       op.stat(nullptr, nullptr, nullptr);
2298
2299     /*
2300      * save the header at the last moment.. If we were to send it off before
2301      * other updates, but die before sending them all, we'd think that the
2302      * on-disk state was fully committed even though it wasn't! However, since
2303      * the messages are strictly ordered between the MDS and the OSD, and
2304      * since messages to a given PG are strictly ordered, if we simply send
2305      * the message containing the header off last, we cannot get our header
2306      * into an incorrect state.
2307      */
2308     if (header) {
2309       bufferlist header;
2310       encode(*fnode, header);
2311       op.omap_set_header(header);
2312     }
2313
2314     op.priority = op_prio;
2315     if (!_set.empty())
2316       op.omap_set(_set);
2317     if (!_rm.empty())
2318       op.omap_rm_keys(_rm);
2319     mdcache->mds->objecter->mutate(oid, oloc, op, snapc,
2320                                    ceph::real_clock::now(),
2321                                    0, gather.new_sub());
2322     write_size = 0;
2323     _set.clear();
2324     _rm.clear();
2325   };
2326
2327   for (auto &key : stales) {
2328     unsigned size = key.length() + sizeof(__u32);
2329     if (write_size + size > max_write_size)
2330       commit_one();
2331
2332     write_size += size;
2333     _rm.emplace(key);
2334   }
2335
2336   for (auto &key : to_remove) {
2337     unsigned size = key.length() + sizeof(__u32);
2338     if (write_size + size > max_write_size)
2339       commit_one();
2340
2341     write_size += size;
2342     _rm.emplace(std::move(key));
2343   }
2344
2345   bufferlist bl;
2346   using ceph::encode;
2347   for (auto &item : to_set) {
2348     encode(item.first, bl);
2349     if (item.is_remote) {
2350       // remote link
2351       CDentry::encode_remote(item.ino, item.d_type, item.alternate_name, bl);
2352     } else {
2353       // marker, name, inode, [symlink string]
2354       bl.append('i');         // inode
2355
2356       ENCODE_START(2, 1, bl);
2357       encode(item.alternate_name, bl);
2358       _encode_primary_inode_base(item, dfts, bl);
2359       ENCODE_FINISH(bl);
2360     }
2361
2362     unsigned size = item.key.length() + bl.length() + 2 * sizeof(__u32);
2363     if (write_size + size > max_write_size)
2364       commit_one();
2365
2366     write_size += size;
2367     _set[std::move(item.key)].swap(bl);
2368   }
2369
2370   commit_one(true);
2371   gather.activate();
2372 }
2373
2374 /**
2375  * Flush out the modified dentries in this dir. Keep the bufferlist
2376  * below max_write_size;
2377  */
2378 void CDir::_omap_commit(int op_prio)
2379 {
2380   dout(10) << __func__ << dendl;
2381
2382   if (op_prio < 0)
2383     op_prio = CEPH_MSG_PRIO_DEFAULT;
2384
2385   // snap purge?
2386   const set<snapid_t> *snaps = NULL;
2387   SnapRealm *realm = inode->find_snaprealm();
2388   if (fnode->snap_purged_thru < realm->get_last_destroyed()) {
2389     snaps = &realm->get_snaps();
2390     dout(10) << " snap_purged_thru " << fnode->snap_purged_thru
2391              << " < " << realm->get_last_destroyed()
2392              << ", snap purge based on " << *snaps << dendl;
2393     // fnode.snap_purged_thru = realm->get_last_destroyed();
2394   }
2395
2396   size_t count = 0;
2397   if (state_test(CDir::STATE_FRAGMENTING) && is_new()) {
2398     count = get_num_head_items() + get_num_snap_items();
2399   } else {
2400     for (elist<CDentry*>::iterator it = dirty_dentries.begin(); !it.end(); ++it)
2401       ++count;
2402   }
2403
2404   vector<string> to_remove;
2405   // reverve enough memories, which maybe larger than the actually needed
2406   to_remove.reserve(count);
2407
2408   vector<dentry_commit_item> to_set;
2409   // reverve enough memories, which maybe larger than the actually needed
2410   to_set.reserve(count);
2411
2412   // for dir fragtrees
2413   bufferlist dfts(CEPH_PAGE_SIZE);
2414
2415   auto write_one = [&](CDentry *dn) {
2416     string key;
2417     dn->key().encode(key);
2418
2419     if (dn->last != CEPH_NOSNAP &&
2420         snaps && try_trim_snap_dentry(dn, *snaps)) {
2421       dout(10) << " rm " << key << dendl;
2422       to_remove.emplace_back(std::move(key));
2423       return;
2424     }
2425
2426     if (dn->get_linkage()->is_null()) {
2427       dout(10) << " rm " << dn->get_name() << " " << *dn << dendl;
2428       to_remove.emplace_back(std::move(key));
2429     } else {
2430       dout(10) << " set " << dn->get_name() << " " << *dn << dendl;
2431
2432       uint64_t off = dfts.length();
2433       // try to reserve new size if there has less
2434       // than 1/8 page space
2435       uint64_t left = CEPH_PAGE_SIZE - off % CEPH_PAGE_SIZE;
2436       if (left < CEPH_PAGE_SIZE / 8)
2437         dfts.reserve(left + CEPH_PAGE_SIZE);
2438
2439       auto& item = to_set.emplace_back();
2440       item.key = std::move(key);
2441       _parse_dentry(dn, item, snaps, dfts);
2442       item.dft_len = dfts.length() - off;
2443     }
2444   };
2445
2446   if (state_test(CDir::STATE_FRAGMENTING) && is_new()) {
2447     ceph_assert(committed_version == 0);
2448     for (auto p = items.begin(); p != items.end(); ) {
2449       CDentry *dn = p->second;
2450       ++p;
2451       if (dn->get_linkage()->is_null())
2452         continue;
2453       write_one(dn);
2454     }
2455   } else {
2456     for (auto p = dirty_dentries.begin(); !p.end(); ) {
2457       CDentry *dn = *p;
2458       ++p;
2459       write_one(dn);
2460     }
2461   }
2462
2463   auto c = new C_IO_Dir_Commit_Ops(this, op_prio, std::move(to_set), std::move(dfts),
2464                                    std::move(to_remove), std::move(stale_items));
2465   stale_items.clear();
2466   mdcache->mds->finisher->queue(c);
2467 }
2468
2469 void CDir::_parse_dentry(CDentry *dn, dentry_commit_item &item,
2470                          const set<snapid_t> *snaps, bufferlist &bl)
2471 {
2472   // clear dentry NEW flag, if any.  we can no longer silently drop it.
2473   dn->clear_new();
2474
2475   item.first = dn->first;
2476
2477   // primary or remote?
2478   auto& linkage = dn->linkage;
2479   item.alternate_name = dn->get_alternate_name();
2480   if (linkage.is_remote()) {
2481     item.is_remote = true;
2482     item.ino = linkage.get_remote_ino();
2483     item.d_type = linkage.get_remote_d_type();
2484     dout(14) << " dn '" << dn->get_name() << "' remote ino " << item.ino << dendl;
2485   } else if (linkage.is_primary()) {
2486     // primary link
2487     CInode *in = linkage.get_inode();
2488     ceph_assert(in);
2489
2490     dout(14) << " dn '" << dn->get_name() << "' inode " << *in << dendl;
2491
2492     if (in->is_multiversion()) {
2493       if (!in->snaprealm) {
2494         if (snaps)
2495           in->purge_stale_snap_data(*snaps);
2496       } else {
2497         in->purge_stale_snap_data(in->snaprealm->get_snaps());
2498       }
2499     }
2500
2501     if (in->snaprealm) {
2502       item.snaprealm = true;
2503       item.srnode = in->snaprealm->srnode;
2504     }
2505     item.features = mdcache->mds->mdsmap->get_up_features();
2506     item.inode = in->inode;
2507     if (in->inode->is_symlink())
2508       item.symlink = in->symlink;
2509     using ceph::encode;
2510     encode(in->dirfragtree, bl);
2511     item.xattrs = in->xattrs;
2512     item.old_inodes = in->old_inodes;
2513     item.oldest_snap = in->oldest_snap;
2514     item.damage_flags = in->damage_flags;
2515   } else {
2516     ceph_assert(!linkage.is_null());
2517   }
2518 }
2519
2520 void CDir::_commit(version_t want, int op_prio)
2521 {
2522   dout(10) << "_commit want " << want << " on " << *this << dendl;
2523
2524   // we can't commit things in the future.
2525   // (even the projected future.)
2526   ceph_assert(want <= get_version() || get_version() == 0);
2527
2528   // check pre+postconditions.
2529   ceph_assert(is_auth());
2530
2531   // already committed?
2532   if (committed_version >= want) {
2533     dout(10) << "already committed " << committed_version << " >= " << want << dendl;
2534     return;
2535   }
2536   // already committing >= want?
2537   if (committing_version >= want) {
2538     dout(10) << "already committing " << committing_version << " >= " << want << dendl;
2539     ceph_assert(state_test(STATE_COMMITTING));
2540     return;
2541   }
2542
2543   // alrady committed an older version?
2544   if (committing_version > committed_version) {
2545     dout(10) << "already committing older " << committing_version << ", waiting for that to finish" << dendl;
2546     return;
2547   }
2548
2549   // commit.
2550   committing_version = get_version();
2551
2552   // mark committing (if not already)
2553   if (!state_test(STATE_COMMITTING)) {
2554     dout(10) << "marking committing" << dendl;
2555     state_set(STATE_COMMITTING);
2556   }
2557
2558   if (mdcache->mds->logger) mdcache->mds->logger->inc(l_mds_dir_commit);
2559
2560   mdcache->mds->balancer->hit_dir(this, META_POP_STORE);
2561
2562   _omap_commit(op_prio);
2563 }
2564
2565
2566 /**
2567  * _committed
2568  *
2569  * @param v version i just committed
2570  */
2571 void CDir::_committed(int r, version_t v)
2572 {
2573   if (r < 0) {
2574     // the directory could be partly purged during MDS failover
2575     if (r == -CEPHFS_ENOENT && committed_version == 0 &&
2576         !inode->is_base() && get_parent_dir()->inode->is_stray()) {
2577       r = 0;
2578       if (inode->snaprealm)
2579         inode->state_set(CInode::STATE_MISSINGOBJS);
2580     }
2581     if (r < 0) {
2582       dout(1) << "commit error " << r << " v " << v << dendl;
2583       mdcache->mds->clog->error() << "failed to commit dir " << dirfrag() << " object,"
2584                                 << " errno " << r;
2585       mdcache->mds->handle_write_error(r);
2586       return;
2587     }
2588   }
2589
2590   dout(10) << "_committed v " << v << " on " << *this << dendl;
2591   ceph_assert(is_auth());
2592
2593   bool stray = inode->is_stray();
2594
2595   // take note.
2596   ceph_assert(v > committed_version);
2597   ceph_assert(v <= committing_version);
2598   committed_version = v;
2599
2600   // _all_ commits done?
2601   if (committing_version == committed_version)
2602     state_clear(CDir::STATE_COMMITTING);
2603
2604   // _any_ commit, even if we've been redirtied, means we're no longer new.
2605   item_new.remove_myself();
2606
2607   // dir clean?
2608   if (committed_version == get_version())
2609     mark_clean();
2610
2611   // dentries clean?
2612   for (auto p = dirty_dentries.begin(); !p.end(); ) {
2613     CDentry *dn = *p;
2614     ++p;
2615
2616     // inode?
2617     if (dn->linkage.is_primary()) {
2618       CInode *in = dn->linkage.get_inode();
2619       ceph_assert(in);
2620       ceph_assert(in->is_auth());
2621
2622       if (committed_version >= in->get_version()) {
2623         if (in->is_dirty()) {
2624           dout(15) << " dir " << committed_version << " >= inode " << in->get_version() << " now clean " << *in << dendl;
2625           in->mark_clean();
2626         }
2627       } else {
2628         dout(15) << " dir " << committed_version << " < inode " << in->get_version() << " still dirty " << *in << dendl;
2629         ceph_assert(in->is_dirty() || in->last < CEPH_NOSNAP);  // special case for cow snap items (not predirtied)
2630       }
2631     }
2632
2633     // dentry
2634     if (committed_version >= dn->get_version()) {
2635       dout(15) << " dir " << committed_version << " >= dn " << dn->get_version() << " now clean " << *dn << dendl;
2636       dn->mark_clean();
2637
2638       // drop clean null stray dentries immediately
2639       if (stray &&
2640           dn->get_num_ref() == 0 &&
2641           !dn->is_projected() &&
2642           dn->get_linkage()->is_null())
2643         remove_dentry(dn);
2644     } else {
2645       dout(15) << " dir " << committed_version << " < dn " << dn->get_version() << " still dirty " << *dn << dendl;
2646       ceph_assert(dn->is_dirty());
2647     }
2648   }
2649
2650   // finishers?
2651   bool were_waiters = !waiting_for_commit.empty();
2652
2653   auto it = waiting_for_commit.begin();
2654   while (it != waiting_for_commit.end()) {
2655     auto _it = it;
2656     ++_it;
2657     if (it->first > committed_version) {
2658       dout(10) << " there are waiters for " << it->first << ", committing again" << dendl;
2659       _commit(it->first, -1);
2660       break;
2661     }
2662     MDSContext::vec t;
2663     for (const auto &waiter : it->second)
2664       t.push_back(waiter);
2665     mdcache->mds->queue_waiters(t);
2666     waiting_for_commit.erase(it);
2667     it = _it;
2668   }
2669
2670   // try drop dentries in this dirfrag if it's about to be purged
2671   if (!inode->is_base() && get_parent_dir()->inode->is_stray() &&
2672       inode->snaprealm)
2673     mdcache->maybe_eval_stray(inode, true);
2674
2675   // unpin if we kicked the last waiter.
2676   if (were_waiters &&
2677       waiting_for_commit.empty())
2678     auth_unpin(this);
2679 }
2680
2681
2682
2683
2684 // IMPORT/EXPORT
2685
2686 mds_rank_t CDir::get_export_pin(bool inherit) const
2687 {
2688   mds_rank_t export_pin = inode->get_export_pin(inherit);
2689   if (export_pin == MDS_RANK_EPHEMERAL_DIST)
2690     export_pin = mdcache->hash_into_rank_bucket(ino(), get_frag());
2691   else if (export_pin == MDS_RANK_EPHEMERAL_RAND)
2692     export_pin = mdcache->hash_into_rank_bucket(ino());
2693   return export_pin;
2694 }
2695
2696 bool CDir::is_exportable(mds_rank_t dest) const
2697 {
2698   mds_rank_t export_pin = get_export_pin();
2699   if (export_pin == dest)
2700     return true;
2701   if (export_pin >= 0)
2702     return false;
2703   return true;
2704 }
2705
2706 void CDir::encode_export(bufferlist& bl)
2707 {
2708   ENCODE_START(1, 1, bl);
2709   ceph_assert(!is_projected());
2710   encode(first, bl);
2711   encode(*fnode, bl);
2712   encode(dirty_old_rstat, bl);
2713   encode(committed_version, bl);
2714
2715   encode(state, bl);
2716   encode(dir_rep, bl);
2717
2718   encode(pop_me, bl);
2719   encode(pop_auth_subtree, bl);
2720
2721   encode(dir_rep_by, bl);
2722   encode(get_replicas(), bl);
2723
2724   get(PIN_TEMPEXPORTING);
2725   ENCODE_FINISH(bl);
2726 }
2727
2728 void CDir::finish_export()
2729 {
2730   state &= MASK_STATE_EXPORT_KEPT;
2731   pop_nested.sub(pop_auth_subtree);
2732   pop_auth_subtree_nested.sub(pop_auth_subtree);
2733   pop_me.zero();
2734   pop_auth_subtree.zero();
2735   put(PIN_TEMPEXPORTING);
2736   dirty_old_rstat.clear();
2737 }
2738
2739 void CDir::decode_import(bufferlist::const_iterator& blp, LogSegment *ls)
2740 {
2741   DECODE_START(1, blp);
2742   decode(first, blp);
2743   {
2744     auto _fnode = allocate_fnode();
2745     decode(*_fnode, blp);
2746     reset_fnode(std::move(_fnode));
2747   }
2748   update_projected_version();
2749
2750   decode(dirty_old_rstat, blp);
2751   decode(committed_version, blp);
2752   committing_version = committed_version;
2753
2754   unsigned s;
2755   decode(s, blp);
2756   state &= MASK_STATE_IMPORT_KEPT;
2757   state_set(STATE_AUTH | (s & MASK_STATE_EXPORTED));
2758
2759   if (is_dirty()) {
2760     get(PIN_DIRTY);
2761     _mark_dirty(ls);
2762   }
2763
2764   decode(dir_rep, blp);
2765
2766   decode(pop_me, blp);
2767   decode(pop_auth_subtree, blp);
2768   pop_nested.add(pop_auth_subtree);
2769   pop_auth_subtree_nested.add(pop_auth_subtree);
2770
2771   decode(dir_rep_by, blp);
2772   decode(get_replicas(), blp);
2773   if (is_replicated()) get(PIN_REPLICATED);
2774
2775   replica_nonce = 0;  // no longer defined
2776
2777   // did we import some dirty scatterlock data?
2778   if (dirty_old_rstat.size() ||
2779       !(fnode->rstat == fnode->accounted_rstat)) {
2780     mdcache->mds->locker->mark_updated_scatterlock(&inode->nestlock);
2781     ls->dirty_dirfrag_nest.push_back(&inode->item_dirty_dirfrag_nest);
2782   }
2783   if (!(fnode->fragstat == fnode->accounted_fragstat)) {
2784     mdcache->mds->locker->mark_updated_scatterlock(&inode->filelock);
2785     ls->dirty_dirfrag_dir.push_back(&inode->item_dirty_dirfrag_dir);
2786   }
2787   if (is_dirty_dft()) {
2788     if (inode->dirfragtreelock.get_state() != LOCK_MIX &&
2789         inode->dirfragtreelock.is_stable()) {
2790       // clear stale dirtydft
2791       state_clear(STATE_DIRTYDFT);
2792     } else {
2793       mdcache->mds->locker->mark_updated_scatterlock(&inode->dirfragtreelock);
2794       ls->dirty_dirfrag_dirfragtree.push_back(&inode->item_dirty_dirfrag_dirfragtree);
2795     }
2796   }
2797   DECODE_FINISH(blp);
2798 }
2799
2800 void CDir::abort_import()
2801 {
2802   ceph_assert(is_auth());
2803   state_clear(CDir::STATE_AUTH);
2804   remove_bloom();
2805   clear_replica_map();
2806   set_replica_nonce(CDir::EXPORT_NONCE);
2807   if (is_dirty())
2808     mark_clean();
2809
2810   pop_nested.sub(pop_auth_subtree);
2811   pop_auth_subtree_nested.sub(pop_auth_subtree);
2812   pop_me.zero();
2813   pop_auth_subtree.zero();
2814 }
2815
2816 void CDir::encode_dirstat(bufferlist& bl, const session_info_t& info, const DirStat& ds) {
2817   if (info.has_feature(CEPHFS_FEATURE_REPLY_ENCODING)) {
2818     ENCODE_START(1, 1, bl);
2819     encode(ds.frag, bl);
2820     encode(ds.auth, bl);
2821     encode(ds.dist, bl);
2822     ENCODE_FINISH(bl);
2823   }
2824   else {
2825     encode(ds.frag, bl);
2826     encode(ds.auth, bl);
2827     encode(ds.dist, bl);
2828   }
2829 }
2830
2831 /********************************
2832  * AUTHORITY
2833  */
2834
2835 /*
2836  * if dir_auth.first == parent, auth is same as inode.
2837  * unless .second != unknown, in which case that sticks.
2838  */
2839 mds_authority_t CDir::authority() const
2840 {
2841   if (is_subtree_root())
2842     return dir_auth;
2843   else
2844     return inode->authority();
2845 }
2846
2847 /** is_subtree_root()
2848  * true if this is an auth delegation point.
2849  * that is, dir_auth != default (parent,unknown)
2850  *
2851  * some key observations:
2852  *  if i am auth:
2853  *    - any region bound will be an export, or frozen.
2854  *
2855  * note that this DOES heed dir_auth.pending
2856  */
2857 /*
2858 bool CDir::is_subtree_root()
2859 {
2860   if (dir_auth == CDIR_AUTH_DEFAULT) {
2861     //dout(10) << "is_subtree_root false " << dir_auth << " != " << CDIR_AUTH_DEFAULT
2862     //<< " on " << ino() << dendl;
2863     return false;
2864   } else {
2865     //dout(10) << "is_subtree_root true " << dir_auth << " != " << CDIR_AUTH_DEFAULT
2866     //<< " on " << ino() << dendl;
2867     return true;
2868   }
2869 }
2870 */
2871
2872 /** contains(x)
2873  * true if we are x, or an ancestor of x
2874  */
2875 bool CDir::contains(CDir *x)
2876 {
2877   while (1) {
2878     if (x == this)
2879       return true;
2880     x = x->get_inode()->get_projected_parent_dir();
2881     if (x == 0)
2882       return false;
2883   }
2884 }
2885
2886 bool CDir::can_rep() const
2887 {
2888   if (!is_rep())
2889     return true;
2890
2891   unsigned mds_num = mdcache->mds->get_mds_map()->get_num_mds(MDSMap::STATE_ACTIVE);
2892   if ((mds_num - 1) > get_replicas().size())
2893     return true;
2894
2895   return false;
2896 }
2897
2898
2899 /** set_dir_auth
2900  */
2901 void CDir::set_dir_auth(const mds_authority_t &a)
2902 {
2903   dout(10) << "setting dir_auth=" << a
2904            << " from " << dir_auth
2905            << " on " << *this << dendl;
2906
2907   bool was_subtree = is_subtree_root();
2908   bool was_ambiguous = dir_auth.second >= 0;
2909
2910   // set it.
2911   dir_auth = a;
2912
2913   // new subtree root?
2914   if (!was_subtree && is_subtree_root()) {
2915     dout(10) << " new subtree root, adjusting auth_pins" << dendl;
2916
2917     if (freeze_tree_state) {
2918       // only by CDir::_freeze_tree()
2919       ceph_assert(is_freezing_tree_root());
2920     }
2921
2922     inode->num_subtree_roots++;
2923
2924     // unpin parent of frozen dir/tree?
2925     if (inode->is_auth()) {
2926       ceph_assert(!is_frozen_tree_root());
2927       if (is_frozen_dir())
2928         inode->auth_unpin(this);
2929     }
2930   }
2931   if (was_subtree && !is_subtree_root()) {
2932     dout(10) << " old subtree root, adjusting auth_pins" << dendl;
2933
2934     inode->num_subtree_roots--;
2935
2936     // pin parent of frozen dir/tree?
2937     if (inode->is_auth()) {
2938       ceph_assert(!is_frozen_tree_root());
2939       if (is_frozen_dir())
2940         inode->auth_pin(this);
2941     }
2942   }
2943
2944   // newly single auth?
2945   if (was_ambiguous && dir_auth.second == CDIR_AUTH_UNKNOWN) {
2946     MDSContext::vec ls;
2947     take_waiting(WAIT_SINGLEAUTH, ls);
2948     mdcache->mds->queue_waiters(ls);
2949   }
2950 }
2951
2952 /*****************************************
2953  * AUTH PINS and FREEZING
2954  *
2955  * the basic plan is that auth_pins only exist in auth regions, and they
2956  * prevent a freeze (and subsequent auth change).
2957  *
2958  * however, we also need to prevent a parent from freezing if a child is frozen.
2959  * for that reason, the parent inode of a frozen directory is auth_pinned.
2960  *
2961  * the oddity is when the frozen directory is a subtree root.  if that's the case,
2962  * the parent inode isn't frozen.  which means that when subtree authority is adjusted
2963  * at the bounds, inodes for any frozen bound directories need to get auth_pins at that
2964  * time.
2965  *
2966  */
2967
2968 void CDir::auth_pin(void *by)
2969 {
2970   if (auth_pins == 0)
2971     get(PIN_AUTHPIN);
2972   auth_pins++;
2973
2974 #ifdef MDS_AUTHPIN_SET
2975   auth_pin_set.insert(by);
2976 #endif
2977
2978   dout(10) << "auth_pin by " << by << " on " << *this << " count now " << auth_pins << dendl;
2979
2980   if (freeze_tree_state)
2981     freeze_tree_state->auth_pins += 1;
2982 }
2983
2984 void CDir::auth_unpin(void *by)
2985 {
2986   auth_pins--;
2987
2988 #ifdef MDS_AUTHPIN_SET
2989   {
2990     auto it = auth_pin_set.find(by);
2991     ceph_assert(it != auth_pin_set.end());
2992     auth_pin_set.erase(it);
2993   }
2994 #endif
2995   if (auth_pins == 0)
2996     put(PIN_AUTHPIN);
2997
2998   dout(10) << "auth_unpin by " << by << " on " << *this << " count now " << auth_pins << dendl;
2999   ceph_assert(auth_pins >= 0);
3000
3001   if (freeze_tree_state)
3002     freeze_tree_state->auth_pins -= 1;
3003
3004   maybe_finish_freeze();  // pending freeze?
3005 }
3006
3007 void CDir::adjust_nested_auth_pins(int dirinc, void *by)
3008 {
3009   ceph_assert(dirinc);
3010   dir_auth_pins += dirinc;
3011
3012   dout(15) << __func__ << " " << dirinc << " on " << *this
3013            << " by " << by << " count now "
3014            << auth_pins << "/" << dir_auth_pins << dendl;
3015   ceph_assert(dir_auth_pins >= 0);
3016
3017   if (freeze_tree_state)
3018     freeze_tree_state->auth_pins += dirinc;
3019
3020   if (dirinc < 0)
3021     maybe_finish_freeze();  // pending freeze?
3022 }
3023
3024 #ifdef MDS_VERIFY_FRAGSTAT
3025 void CDir::verify_fragstat()
3026 {
3027   ceph_assert(is_complete());
3028   if (inode->is_stray())
3029     return;
3030
3031   frag_info_t c;
3032   memset(&c, 0, sizeof(c));
3033
3034   for (auto it = items.begin();
3035        it != items.end();
3036        ++it) {
3037     CDentry *dn = it->second;
3038     if (dn->is_null())
3039       continue;
3040
3041     dout(10) << " " << *dn << dendl;
3042     if (dn->is_primary())
3043       dout(10) << "     " << *dn->inode << dendl;
3044
3045     if (dn->is_primary()) {
3046       if (dn->inode->is_dir())
3047         c.nsubdirs++;
3048       else
3049         c.nfiles++;
3050     }
3051     if (dn->is_remote()) {
3052       if (dn->get_remote_d_type() == DT_DIR)
3053         c.nsubdirs++;
3054       else
3055         c.nfiles++;
3056     }
3057   }
3058
3059   if (c.nsubdirs != fnode->fragstat.nsubdirs ||
3060       c.nfiles != fnode->fragstat.nfiles) {
3061     dout(0) << "verify_fragstat failed " << fnode->fragstat << " on " << *this << dendl;
3062     dout(0) << "               i count " << c << dendl;
3063     ceph_abort();
3064   } else {
3065     dout(0) << "verify_fragstat ok " << fnode->fragstat << " on " << *this << dendl;
3066   }
3067 }
3068 #endif
3069
3070 /*****************************************************************************
3071  * FREEZING
3072  */
3073
3074 // FREEZE TREE
3075
3076 void CDir::_walk_tree(std::function<bool(CDir*)> callback)
3077 {
3078   deque<CDir*> dfq;
3079   dfq.push_back(this);
3080
3081   while (!dfq.empty()) {
3082     CDir *dir = dfq.front();
3083     dfq.pop_front();
3084
3085     for (auto& p : *dir) {
3086       CDentry *dn = p.second;
3087       if (!dn->get_linkage()->is_primary())
3088         continue;
3089       CInode *in = dn->get_linkage()->get_inode();
3090       if (!in->is_dir())
3091         continue;
3092
3093       auto&& dfv = in->get_nested_dirfrags();
3094       for (auto& dir : dfv) {
3095         auto ret = callback(dir);
3096         if (ret)
3097           dfq.push_back(dir);
3098       }
3099     }
3100   }
3101 }
3102
3103 bool CDir::freeze_tree()
3104 {
3105   ceph_assert(!is_frozen());
3106   ceph_assert(!is_freezing());
3107   ceph_assert(!freeze_tree_state);
3108
3109   auth_pin(this);
3110
3111   // Travese the subtree to mark dirfrags as 'freezing' (set freeze_tree_state)
3112   // and to accumulate auth pins and record total count in freeze_tree_state.
3113   // when auth unpin an 'freezing' object, the counter in freeze_tree_state also
3114   // gets decreased. Subtree become 'frozen' when the counter reaches zero.
3115   freeze_tree_state = std::make_shared<freeze_tree_state_t>(this);
3116   freeze_tree_state->auth_pins += get_auth_pins() + get_dir_auth_pins();
3117   if (!lock_caches_with_auth_pins.empty())
3118     mdcache->mds->locker->invalidate_lock_caches(this);
3119
3120   _walk_tree([this](CDir *dir) {
3121       if (dir->freeze_tree_state)
3122         return false;
3123       dir->freeze_tree_state = freeze_tree_state;
3124       freeze_tree_state->auth_pins += dir->get_auth_pins() + dir->get_dir_auth_pins();
3125       if (!dir->lock_caches_with_auth_pins.empty())
3126         mdcache->mds->locker->invalidate_lock_caches(dir);
3127       return true;
3128     }
3129   );
3130
3131   if (is_freezeable(true)) {
3132     _freeze_tree();
3133     auth_unpin(this);
3134     return true;
3135   } else {
3136     state_set(STATE_FREEZINGTREE);
3137     ++num_freezing_trees;
3138     dout(10) << "freeze_tree waiting " << *this << dendl;
3139     return false;
3140   }
3141 }
3142
3143 void CDir::_freeze_tree()
3144 {
3145   dout(10) << __func__ << " " << *this << dendl;
3146   ceph_assert(is_freezeable(true));
3147
3148   if (freeze_tree_state) {
3149     ceph_assert(is_auth());
3150   } else {
3151     ceph_assert(!is_auth());
3152     freeze_tree_state = std::make_shared<freeze_tree_state_t>(this);
3153   }
3154   freeze_tree_state->frozen = true;
3155
3156   if (is_auth()) {
3157     mds_authority_t auth;
3158     bool was_subtree = is_subtree_root();
3159     if (was_subtree) {
3160       auth = get_dir_auth();
3161     } else {
3162       // temporarily prevent parent subtree from becoming frozen.
3163       inode->auth_pin(this);
3164       // create new subtree
3165       auth = authority();
3166     }
3167
3168     _walk_tree([this, &auth] (CDir *dir) {
3169         if (dir->freeze_tree_state != freeze_tree_state) {
3170           mdcache->adjust_subtree_auth(dir, auth);
3171           return false;
3172         }
3173         return true;
3174       }
3175     );
3176
3177     ceph_assert(auth.first >= 0);
3178     ceph_assert(auth.second == CDIR_AUTH_UNKNOWN);
3179     auth.second = auth.first;
3180     mdcache->adjust_subtree_auth(this, auth);
3181     if (!was_subtree)
3182       inode->auth_unpin(this);
3183   } else {
3184     // importing subtree ?
3185     _walk_tree([this] (CDir *dir) {
3186         ceph_assert(!dir->freeze_tree_state);
3187         dir->freeze_tree_state = freeze_tree_state;
3188         return true;
3189       }
3190     );
3191   }
3192
3193   // twiddle state
3194   if (state_test(STATE_FREEZINGTREE)) {
3195     state_clear(STATE_FREEZINGTREE);
3196     --num_freezing_trees;
3197   }
3198
3199   state_set(STATE_FROZENTREE);
3200   ++num_frozen_trees;
3201   get(PIN_FROZEN);
3202 }
3203
3204 void CDir::unfreeze_tree()
3205 {
3206   dout(10) << __func__ << " " << *this << dendl;
3207
3208   MDSContext::vec unfreeze_waiters;
3209   take_waiting(WAIT_UNFREEZE, unfreeze_waiters);
3210
3211   if (freeze_tree_state) {
3212     _walk_tree([this, &unfreeze_waiters](CDir *dir) {
3213         if (dir->freeze_tree_state != freeze_tree_state)
3214           return false;
3215         dir->freeze_tree_state.reset();
3216         dir->take_waiting(WAIT_UNFREEZE, unfreeze_waiters);
3217         return true;
3218       }
3219     );
3220   }
3221
3222   if (state_test(STATE_FROZENTREE)) {
3223     // frozen.  unfreeze.
3224     state_clear(STATE_FROZENTREE);
3225     --num_frozen_trees;
3226
3227     put(PIN_FROZEN);
3228
3229     if (is_auth()) {
3230       // must be subtree
3231       ceph_assert(is_subtree_root());
3232       // for debug purpose, caller should ensure 'dir_auth.second == dir_auth.first'
3233       mds_authority_t auth = get_dir_auth();
3234       ceph_assert(auth.first >= 0);
3235       ceph_assert(auth.second == auth.first);
3236       auth.second = CDIR_AUTH_UNKNOWN;
3237       mdcache->adjust_subtree_auth(this, auth);
3238     }
3239     freeze_tree_state.reset();
3240   } else {
3241     ceph_assert(state_test(STATE_FREEZINGTREE));
3242
3243     // freezing.  stop it.
3244     state_clear(STATE_FREEZINGTREE);
3245     --num_freezing_trees;
3246     freeze_tree_state.reset();
3247
3248     finish_waiting(WAIT_FROZEN, -1);
3249     auth_unpin(this);
3250   }
3251
3252   mdcache->mds->queue_waiters(unfreeze_waiters);
3253 }
3254
3255 void CDir::adjust_freeze_after_rename(CDir *dir)
3256 {
3257   if (!freeze_tree_state || dir->freeze_tree_state != freeze_tree_state)
3258     return;
3259   CDir *newdir = dir->get_inode()->get_parent_dir();
3260   if (newdir == this || newdir->freeze_tree_state == freeze_tree_state)
3261     return;
3262
3263   ceph_assert(!freeze_tree_state->frozen);
3264   ceph_assert(get_dir_auth_pins() > 0);
3265
3266   MDSContext::vec unfreeze_waiters;
3267
3268   auto unfreeze = [this, &unfreeze_waiters](CDir *dir) {
3269     if (dir->freeze_tree_state != freeze_tree_state)
3270       return false;
3271     int dec = dir->get_auth_pins() + dir->get_dir_auth_pins();
3272     // shouldn't become zero because srcdn of rename was auth pinned
3273     ceph_assert(freeze_tree_state->auth_pins > dec);
3274     freeze_tree_state->auth_pins -= dec;
3275     dir->freeze_tree_state.reset();
3276     dir->take_waiting(WAIT_UNFREEZE, unfreeze_waiters);
3277     return true;
3278   };
3279
3280   unfreeze(dir);
3281   dir->_walk_tree(unfreeze);
3282
3283   mdcache->mds->queue_waiters(unfreeze_waiters);
3284 }
3285
3286 bool CDir::can_auth_pin(int *err_ret) const
3287 {
3288   int err;
3289   if (!is_auth()) {
3290     err = ERR_NOT_AUTH;
3291   } else if (is_freezing_dir() || is_frozen_dir()) {
3292     err = ERR_FRAGMENTING_DIR;
3293   } else {
3294     auto p = is_freezing_or_frozen_tree();
3295     if (p.first || p.second) {
3296       err = ERR_EXPORTING_TREE;
3297     } else {
3298       err = 0;
3299     }
3300   }
3301   if (err && err_ret)
3302     *err_ret = err;
3303   return !err;
3304 }
3305
3306 class C_Dir_AuthUnpin : public CDirContext {
3307   public:
3308   explicit C_Dir_AuthUnpin(CDir *d) : CDirContext(d) {}
3309   void finish(int r) override {
3310     dir->auth_unpin(dir->get_inode());
3311   }
3312 };
3313
3314 void CDir::maybe_finish_freeze()
3315 {
3316   if (dir_auth_pins != 0)
3317     return;
3318
3319   // we can freeze the _dir_ even with nested pins...
3320   if (state_test(STATE_FREEZINGDIR)) {
3321     if (auth_pins == 1) {
3322       _freeze_dir();
3323       auth_unpin(this);
3324       finish_waiting(WAIT_FROZEN);
3325     }
3326   }
3327
3328   if (freeze_tree_state) {
3329     if (freeze_tree_state->frozen ||
3330         freeze_tree_state->auth_pins != 1)
3331       return;
3332
3333     if (freeze_tree_state->dir != this) {
3334       freeze_tree_state->dir->maybe_finish_freeze();
3335       return;
3336     }
3337
3338     ceph_assert(state_test(STATE_FREEZINGTREE));
3339
3340     if (!is_subtree_root() && inode->is_frozen()) {
3341       dout(10) << __func__ << " !subtree root and frozen inode, waiting for unfreeze on " << inode << dendl;
3342       // retake an auth_pin...
3343       auth_pin(inode);
3344       // and release it when the parent inode unfreezes
3345       inode->add_waiter(WAIT_UNFREEZE, new C_Dir_AuthUnpin(this));
3346       return;
3347     }
3348
3349     _freeze_tree();
3350     auth_unpin(this);
3351     finish_waiting(WAIT_FROZEN);
3352   }
3353 }
3354
3355
3356
3357 // FREEZE DIR
3358
3359 bool CDir::freeze_dir()
3360 {
3361   ceph_assert(!is_frozen());
3362   ceph_assert(!is_freezing());
3363
3364   auth_pin(this);
3365   if (is_freezeable_dir(true)) {
3366     _freeze_dir();
3367     auth_unpin(this);
3368     return true;
3369   } else {
3370     state_set(STATE_FREEZINGDIR);
3371     if (!lock_caches_with_auth_pins.empty())
3372       mdcache->mds->locker->invalidate_lock_caches(this);
3373     dout(10) << "freeze_dir + wait " << *this << dendl;
3374     return false;
3375   }
3376 }
3377
3378 void CDir::_freeze_dir()
3379 {
3380   dout(10) << __func__ << " " << *this << dendl;
3381   //assert(is_freezeable_dir(true));
3382   // not always true during split because the original fragment may have frozen a while
3383   // ago and we're just now getting around to breaking it up.
3384
3385   state_clear(STATE_FREEZINGDIR);
3386   state_set(STATE_FROZENDIR);
3387   get(PIN_FROZEN);
3388
3389   if (is_auth() && !is_subtree_root())
3390     inode->auth_pin(this);  // auth_pin for duration of freeze
3391 }
3392
3393
3394 void CDir::unfreeze_dir()
3395 {
3396   dout(10) << __func__ << " " << *this << dendl;
3397
3398   if (state_test(STATE_FROZENDIR)) {
3399     state_clear(STATE_FROZENDIR);
3400     put(PIN_FROZEN);
3401
3402     // unpin  (may => FREEZEABLE)   FIXME: is this order good?
3403     if (is_auth() && !is_subtree_root())
3404       inode->auth_unpin(this);
3405
3406     finish_waiting(WAIT_UNFREEZE);
3407   } else {
3408     finish_waiting(WAIT_FROZEN, -1);
3409
3410     // still freezing. stop.
3411     ceph_assert(state_test(STATE_FREEZINGDIR));
3412     state_clear(STATE_FREEZINGDIR);
3413     auth_unpin(this);
3414
3415     finish_waiting(WAIT_UNFREEZE);
3416   }
3417 }
3418
3419 void CDir::enable_frozen_inode()
3420 {
3421   ceph_assert(frozen_inode_suppressed > 0);
3422   if (--frozen_inode_suppressed == 0) {
3423     for (auto p = freezing_inodes.begin(); !p.end(); ) {
3424       CInode *in = *p;
3425       ++p;
3426       ceph_assert(in->is_freezing_inode());
3427       in->maybe_finish_freeze_inode();
3428     }
3429   }
3430 }
3431
3432 /**
3433  * Slightly less complete than operator<<, because this is intended
3434  * for identifying a directory and its state rather than for dumping
3435  * debug output.
3436  */
3437 void CDir::dump(Formatter *f, int flags) const
3438 {
3439   ceph_assert(f != NULL);
3440   if (flags & DUMP_PATH) {
3441     f->dump_stream("path") << get_path();
3442   }
3443   if (flags & DUMP_DIRFRAG) {
3444     f->dump_stream("dirfrag") << dirfrag();
3445   }
3446   if (flags & DUMP_SNAPID_FIRST) {
3447     f->dump_int("snapid_first", first);
3448   }
3449   if (flags & DUMP_VERSIONS) {
3450     f->dump_stream("projected_version") << get_projected_version();
3451     f->dump_stream("version") << get_version();
3452     f->dump_stream("committing_version") << get_committing_version();
3453     f->dump_stream("committed_version") << get_committed_version();
3454   }
3455   if (flags & DUMP_REP) {
3456     f->dump_bool("is_rep", is_rep());
3457   }
3458   if (flags & DUMP_DIR_AUTH) {
3459     if (get_dir_auth() != CDIR_AUTH_DEFAULT) {
3460       if (get_dir_auth().second == CDIR_AUTH_UNKNOWN) {
3461         f->dump_stream("dir_auth") << get_dir_auth().first;
3462       } else {
3463         f->dump_stream("dir_auth") << get_dir_auth();
3464       }
3465     } else {
3466       f->dump_string("dir_auth", "");
3467     }
3468   }
3469   if (flags & DUMP_STATES) {
3470     f->open_array_section("states");
3471     MDSCacheObject::dump_states(f);
3472     if (state_test(CDir::STATE_COMPLETE)) f->dump_string("state", "complete");
3473     if (state_test(CDir::STATE_FREEZINGTREE)) f->dump_string("state", "freezingtree");
3474     if (state_test(CDir::STATE_FROZENTREE)) f->dump_string("state", "frozentree");
3475     if (state_test(CDir::STATE_FROZENDIR)) f->dump_string("state", "frozendir");
3476     if (state_test(CDir::STATE_FREEZINGDIR)) f->dump_string("state", "freezingdir");
3477     if (state_test(CDir::STATE_EXPORTBOUND)) f->dump_string("state", "exportbound");
3478     if (state_test(CDir::STATE_IMPORTBOUND)) f->dump_string("state", "importbound");
3479     if (state_test(CDir::STATE_BADFRAG)) f->dump_string("state", "badfrag");
3480     f->close_section();
3481   }
3482   if (flags & DUMP_MDS_CACHE_OBJECT) {
3483     MDSCacheObject::dump(f);
3484   }
3485   if (flags & DUMP_ITEMS) {
3486     f->open_array_section("dentries");
3487     for (auto &p : items) {
3488       CDentry *dn = p.second;
3489       f->open_object_section("dentry");
3490       dn->dump(f);
3491       f->close_section();
3492     }
3493     f->close_section();
3494   }
3495 }
3496
3497 void CDir::dump_load(Formatter *f)
3498 {
3499   f->dump_stream("path") << get_path();
3500   f->dump_stream("dirfrag") << dirfrag();
3501
3502   f->open_object_section("pop_me");
3503   pop_me.dump(f);
3504   f->close_section();
3505
3506   f->open_object_section("pop_nested");
3507   pop_nested.dump(f);
3508   f->close_section();
3509
3510   f->open_object_section("pop_auth_subtree");
3511   pop_auth_subtree.dump(f);
3512   f->close_section();
3513
3514   f->open_object_section("pop_auth_subtree_nested");
3515   pop_auth_subtree_nested.dump(f);
3516   f->close_section();
3517 }
3518
3519 /****** Scrub Stuff *******/
3520
3521 void CDir::scrub_info_create() const
3522 {
3523   ceph_assert(!scrub_infop);
3524
3525   // break out of const-land to set up implicit initial state
3526   CDir *me = const_cast<CDir*>(this);
3527   const auto& pf = me->get_projected_fnode();
3528
3529   std::unique_ptr<scrub_info_t> si(new scrub_info_t());
3530
3531   si->last_recursive.version = pf->recursive_scrub_version;
3532   si->last_recursive.time = pf->recursive_scrub_stamp;
3533
3534   si->last_local.version = pf->localized_scrub_version;
3535   si->last_local.time = pf->localized_scrub_stamp;
3536
3537   me->scrub_infop.swap(si);
3538 }
3539
3540 void CDir::scrub_initialize(const ScrubHeaderRef& header)
3541 {
3542   ceph_assert(header);
3543   // FIXME: weird implicit construction, is someone else meant
3544   // to be calling scrub_info_create first?
3545   scrub_info();
3546   scrub_infop->directory_scrubbing = true;
3547   scrub_infop->header = header;
3548   header->inc_num_pending();
3549 }
3550
3551 void CDir::scrub_aborted() {
3552   dout(20) << __func__ << dendl;
3553   ceph_assert(scrub_is_in_progress());
3554
3555   scrub_infop->last_scrub_dirty = false;
3556   scrub_infop->directory_scrubbing = false;
3557   scrub_infop->header->dec_num_pending();
3558   scrub_infop.reset();
3559 }
3560
3561 void CDir::scrub_finished()
3562 {
3563   dout(20) << __func__ << dendl;
3564   ceph_assert(scrub_is_in_progress());
3565
3566   scrub_infop->last_local.time = ceph_clock_now();
3567   scrub_infop->last_local.version = get_version();
3568   if (scrub_infop->header->get_recursive())
3569     scrub_infop->last_recursive = scrub_infop->last_local;
3570
3571   scrub_infop->last_scrub_dirty = true;
3572
3573   scrub_infop->directory_scrubbing = false;
3574   scrub_infop->header->dec_num_pending();
3575 }
3576
3577 void CDir::scrub_maybe_delete_info()
3578 {
3579   if (scrub_infop &&
3580       !scrub_infop->directory_scrubbing &&
3581       !scrub_infop->last_scrub_dirty)
3582     scrub_infop.reset();
3583 }
3584
3585 bool CDir::scrub_local()
3586 {
3587   ceph_assert(is_complete());
3588   bool good = check_rstats(true);
3589   if (!good && scrub_infop->header->get_repair()) {
3590     mdcache->repair_dirfrag_stats(this);
3591     scrub_infop->header->set_repaired();
3592   }
3593   return good;
3594 }
3595
3596 std::string CDir::get_path() const
3597 {
3598   std::string path;
3599   get_inode()->make_path_string(path, true);
3600   return path;
3601 }
3602
3603 bool CDir::should_split_fast() const
3604 {
3605   // Max size a fragment can be before trigger fast splitting
3606   int fast_limit = g_conf()->mds_bal_split_size * g_conf()->mds_bal_fragment_fast_factor;
3607
3608   // Fast path: the sum of accounted size and null dentries does not
3609   // exceed threshold: we definitely are not over it.
3610   if (get_frag_size() + get_num_head_null() <= fast_limit) {
3611     return false;
3612   }
3613
3614   // Fast path: the accounted size of the frag exceeds threshold: we
3615   // definitely are over it
3616   if (get_frag_size() > fast_limit) {
3617     return true;
3618   }
3619
3620   int64_t effective_size = 0;
3621
3622   for (const auto &p : items) {
3623     const CDentry *dn = p.second;
3624     if (!dn->get_projected_linkage()->is_null()) {
3625       effective_size++;
3626     }
3627   }
3628
3629   return effective_size > fast_limit;
3630 }
3631
3632 bool CDir::should_merge() const
3633 {
3634   if (get_frag() == frag_t())
3635     return false;
3636
3637   if (inode->is_ephemeral_dist()) {
3638     unsigned min_frag_bits = mdcache->get_ephemeral_dist_frag_bits();
3639     if (min_frag_bits > 0 && get_frag().bits() < min_frag_bits + 1)
3640       return false;
3641   }
3642
3643   return (int)get_frag_size() < g_conf()->mds_bal_merge_size;
3644 }
3645
3646 MEMPOOL_DEFINE_OBJECT_FACTORY(CDir, co_dir, mds_co);
3647 MEMPOOL_DEFINE_OBJECT_FACTORY(CDir::scrub_info_t, scrub_info_t, mds_co)