ceph/src/mds/CDir.h

   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
  16 #ifndef CEPH_CDIR_H
  17 #define CEPH_CDIR_H
  18
  19 #include <iosfwd>
  20 #include <list>
  21 #include <map>
  22 #include <set>
  23 #include <string>
  24 #include <string_view>
  25
  26 #include "common/bloom_filter.hpp"
  27 #include "common/config.h"
  28 #include "include/buffer_fwd.h"
  29 #include "include/counter.h"
  30 #include "include/types.h"
  31
  32 #include "CInode.h"
  33 #include "MDSCacheObject.h"
  34 #include "MDSContext.h"
  35 #include "cephfs_features.h"
  36 #include "SessionMap.h"
  37 #include "messages/MClientReply.h"
  38
  39 class CDentry;
  40 class MDCache;
  41
  42 std::ostream& operator<<(std::ostream& out, const class CDir& dir);
  43
  44 class CDir : public MDSCacheObject, public Counter<CDir> {
  45 public:
  46   MEMPOOL_CLASS_HELPERS();
  47
  48   typedef mempool::mds_co::map<dentry_key_t, CDentry*> dentry_key_map;
  49   typedef mempool::mds_co::set<dentry_key_t> dentry_key_set;
  50
  51   using fnode_ptr = std::shared_ptr<fnode_t>;
  52   using fnode_const_ptr = std::shared_ptr<const fnode_t>;
  53
  54   template <typename ...Args>
  55   static fnode_ptr allocate_fnode(Args && ...args) {
  56     static mempool::mds_co::pool_allocator<fnode_t> allocator;
  57     return std::allocate_shared<fnode_t>(allocator, std::forward<Args>(args)...);
  58   }
  59
  60   struct dentry_commit_item {
  61     string key;
  62     snapid_t first;
  63     bool is_remote = false;
  64
  65     inodeno_t ino;
  66     unsigned char d_type;
  67     mempool::mds_co::string alternate_name;
  68
  69     bool snaprealm = false;
  70     sr_t srnode;
  71
  72     mempool::mds_co::string symlink;
  73     uint64_t features;
  74     uint64_t dft_len;
  75     CInode::inode_const_ptr inode;
  76     CInode::xattr_map_const_ptr xattrs;
  77     CInode::old_inode_map_const_ptr old_inodes;
  78     snapid_t oldest_snap;
  79     damage_flags_t damage_flags;
  80   };
  81
  82   // -- freezing --
  83   struct freeze_tree_state_t {
  84     CDir *dir; // freezing/frozen tree root
  85     int auth_pins = 0;
  86     bool frozen = false;
  87     freeze_tree_state_t(CDir *d) : dir(d) {}
  88   };
  89
  90   class scrub_info_t {
  91   public:
  92     MEMPOOL_CLASS_HELPERS();
  93     struct scrub_stamps {
  94       version_t version = 0;
  95       utime_t time;
  96     };
  97
  98     scrub_info_t() {}
  99
 100     scrub_stamps last_recursive; // when we last finished a recursive scrub
 101     scrub_stamps last_local; // when we last did a local scrub
 102
 103     bool directory_scrubbing = false; /// safety check
 104     bool last_scrub_dirty = false; /// is scrub info dirty or is it flushed to fnode?
 105
 106     ScrubHeaderRef header;
 107   };
 108
 109   // -- pins --
 110   static const int PIN_DNWAITER =     1;
 111   static const int PIN_INOWAITER =    2;
 112   static const int PIN_CHILD =        3;
 113   static const int PIN_FROZEN =       4;
 114   static const int PIN_SUBTREE =      5;
 115   static const int PIN_IMPORTING =    7;
 116   static const int PIN_IMPORTBOUND =  9;
 117   static const int PIN_EXPORTBOUND = 10;
 118   static const int PIN_STICKY =      11;
 119   static const int PIN_SUBTREETEMP = 12;  // used by MDCache::trim_non_auth()
 120
 121   // -- state --
 122   static const unsigned STATE_COMPLETE =      (1<< 0);  // the complete contents are in cache
 123   static const unsigned STATE_FROZENTREE =    (1<< 1);  // root of tree (bounded by exports)
 124   static const unsigned STATE_FREEZINGTREE =  (1<< 2);  // in process of freezing
 125   static const unsigned STATE_FROZENDIR =     (1<< 3);
 126   static const unsigned STATE_FREEZINGDIR =   (1<< 4);
 127   static const unsigned STATE_COMMITTING =    (1<< 5);  // mid-commit
 128   static const unsigned STATE_FETCHING =      (1<< 6);  // currenting fetching
 129   static const unsigned STATE_CREATING =      (1<< 7);
 130   static const unsigned STATE_IMPORTBOUND =   (1<< 8);
 131   static const unsigned STATE_EXPORTBOUND =   (1<< 9);
 132   static const unsigned STATE_EXPORTING =     (1<<10);
 133   static const unsigned STATE_IMPORTING =     (1<<11);
 134   static const unsigned STATE_FRAGMENTING =   (1<<12);
 135   static const unsigned STATE_STICKY =        (1<<13);  // sticky pin due to inode stickydirs
 136   static const unsigned STATE_DNPINNEDFRAG =  (1<<14);  // dir is refragmenting
 137   static const unsigned STATE_ASSIMRSTAT =    (1<<15);  // assimilating inode->frag rstats
 138   static const unsigned STATE_DIRTYDFT =      (1<<16);  // dirty dirfragtree
 139   static const unsigned STATE_BADFRAG =       (1<<17);  // bad dirfrag
 140   static const unsigned STATE_TRACKEDBYOFT =  (1<<18);  // tracked by open file table
 141   static const unsigned STATE_AUXSUBTREE =    (1<<19);  // no subtree merge
 142
 143   // common states
 144   static const unsigned STATE_CLEAN =  0;
 145
 146   // these state bits are preserved by an import/export
 147   // ...except if the directory is hashed, in which case none of them are!
 148   static const unsigned MASK_STATE_EXPORTED =
 149   (STATE_COMPLETE|STATE_DIRTY|STATE_DIRTYDFT|STATE_BADFRAG);
 150   static const unsigned MASK_STATE_IMPORT_KEPT =
 151   (
 152    STATE_IMPORTING |
 153    STATE_IMPORTBOUND |
 154    STATE_EXPORTBOUND |
 155    STATE_FROZENTREE |
 156    STATE_STICKY |
 157    STATE_TRACKEDBYOFT);
 158   static const unsigned MASK_STATE_EXPORT_KEPT =
 159   (STATE_EXPORTING |
 160    STATE_IMPORTBOUND |
 161    STATE_EXPORTBOUND |
 162    STATE_FROZENTREE |
 163    STATE_FROZENDIR |
 164    STATE_STICKY |
 165    STATE_TRACKEDBYOFT);
 166   static const unsigned MASK_STATE_FRAGMENT_KEPT =
 167   (STATE_DIRTY |
 168    STATE_EXPORTBOUND |
 169    STATE_IMPORTBOUND |
 170    STATE_AUXSUBTREE |
 171    STATE_REJOINUNDEF);
 172
 173   // -- rep spec --
 174   static const int REP_NONE =     0;
 175   static const int REP_ALL =      1;
 176   static const int REP_LIST =     2;
 177
 178   static const unsigned EXPORT_NONCE  = 1;
 179
 180   // -- wait masks --
 181   static const uint64_t WAIT_DENTRY       = (1<<0);  // wait for item to be in cache
 182   static const uint64_t WAIT_COMPLETE     = (1<<1);  // wait for complete dir contents
 183   static const uint64_t WAIT_FROZEN       = (1<<2);  // auth pins removed
 184   static const uint64_t WAIT_CREATED      = (1<<3);  // new dirfrag is logged
 185
 186   static const int WAIT_DNLOCK_OFFSET = 4;
 187
 188   static const uint64_t WAIT_ANY_MASK = (uint64_t)(-1);
 189   static const uint64_t WAIT_ATSUBTREEROOT = (WAIT_SINGLEAUTH);
 190
 191   // -- dump flags --
 192   static const int DUMP_PATH             = (1 << 0);
 193   static const int DUMP_DIRFRAG          = (1 << 1);
 194   static const int DUMP_SNAPID_FIRST     = (1 << 2);
 195   static const int DUMP_VERSIONS         = (1 << 3);
 196   static const int DUMP_REP              = (1 << 4);
 197   static const int DUMP_DIR_AUTH         = (1 << 5);
 198   static const int DUMP_STATES           = (1 << 6);
 199   static const int DUMP_MDS_CACHE_OBJECT = (1 << 7);
 200   static const int DUMP_ITEMS            = (1 << 8);
 201   static const int DUMP_ALL              = (-1);
 202   static const int DUMP_DEFAULT          = DUMP_ALL & (~DUMP_ITEMS);
 203
 204   CDir(CInode *in, frag_t fg, MDCache *mdc, bool auth);
 205
 206   std::string_view pin_name(int p) const override {
 207     switch (p) {
 208     case PIN_DNWAITER: return "dnwaiter";
 209     case PIN_INOWAITER: return "inowaiter";
 210     case PIN_CHILD: return "child";
 211     case PIN_FROZEN: return "frozen";
 212     case PIN_SUBTREE: return "subtree";
 213     case PIN_IMPORTING: return "importing";
 214     case PIN_IMPORTBOUND: return "importbound";
 215     case PIN_EXPORTBOUND: return "exportbound";
 216     case PIN_STICKY: return "sticky";
 217     case PIN_SUBTREETEMP: return "subtreetemp";
 218     default: return generic_pin_name(p);
 219     }
 220   }
 221
 222   bool is_lt(const MDSCacheObject *r) const override {
 223     return dirfrag() < (static_cast<const CDir*>(r))->dirfrag();
 224   }
 225
 226   void resync_accounted_fragstat();
 227   void resync_accounted_rstat();
 228   void assimilate_dirty_rstat_inodes(MutationRef& mut);
 229   void assimilate_dirty_rstat_inodes_finish(EMetaBlob *blob);
 230
 231   void mark_exporting() {
 232     state_set(CDir::STATE_EXPORTING);
 233     inode->num_exporting_dirs++;
 234   }
 235   void clear_exporting() {
 236     state_clear(CDir::STATE_EXPORTING);
 237     inode->num_exporting_dirs--;
 238   }
 239
 240   version_t get_version() const { return fnode->version; }
 241   void update_projected_version() {
 242     ceph_assert(projected_fnode.empty());
 243     projected_version = fnode->version;
 244   }
 245   version_t get_projected_version() const { return projected_version; }
 246
 247   void reset_fnode(fnode_const_ptr&& ptr) {
 248     fnode = std::move(ptr);
 249   }
 250
 251   const fnode_const_ptr& get_fnode() const {
 252     return fnode;
 253   }
 254
 255   // only used for updating newly allocated CDir
 256   fnode_t* _get_fnode() {
 257     if (fnode == empty_fnode)
 258       reset_fnode(allocate_fnode());
 259     return const_cast<fnode_t*>(fnode.get());
 260   }
 261
 262   const fnode_const_ptr& get_projected_fnode() const {
 263     if (projected_fnode.empty())
 264       return fnode;
 265     else
 266       return projected_fnode.back();
 267   }
 268
 269   // fnode should have already been projected in caller's context
 270   fnode_t* _get_projected_fnode() {
 271     ceph_assert(!projected_fnode.empty());
 272     return const_cast<fnode_t*>(projected_fnode.back().get());
 273   }
 274
 275   fnode_ptr project_fnode(const MutationRef& mut);
 276
 277   void pop_and_dirty_projected_fnode(LogSegment *ls, const MutationRef& mut);
 278   bool is_projected() const { return !projected_fnode.empty(); }
 279   version_t pre_dirty(version_t min=0);
 280   void _mark_dirty(LogSegment *ls);
 281   void _set_dirty_flag() {
 282     if (!state_test(STATE_DIRTY)) {
 283       state_set(STATE_DIRTY);
 284       get(PIN_DIRTY);
 285     }
 286   }
 287   void mark_dirty(LogSegment *ls, version_t pv=0);
 288   void mark_clean();
 289
 290   bool is_new() { return item_new.is_on_list(); }
 291   void mark_new(LogSegment *ls);
 292
 293   bool is_bad() { return state_test(STATE_BADFRAG); }
 294
 295   /**
 296    * Call to start this CDir on a new scrub.
 297    * @pre It is not currently scrubbing
 298    * @pre The CDir is marked complete.
 299    * @post It has set up its internal scrubbing state.
 300    */
 301   void scrub_initialize(const ScrubHeaderRef& header);
 302   const ScrubHeaderRef& get_scrub_header() {
 303     static const ScrubHeaderRef nullref;
 304     return scrub_infop ? scrub_infop->header : nullref;
 305   }
 306
 307   bool scrub_is_in_progress() const {
 308     return (scrub_infop && scrub_infop->directory_scrubbing);
 309   }
 310
 311   /**
 312    * Call this once all CDentries have been scrubbed, according to
 313    * scrub_dentry_next's listing. It finalizes the scrub statistics.
 314    */
 315   void scrub_finished();
 316
 317   void scrub_aborted();
 318   /**
 319    * Tell the CDir to do a local scrub of itself.
 320    * @pre The CDir is_complete().
 321    * @returns true if the rstats and directory contents match, false otherwise.
 322    */
 323   bool scrub_local();
 324
 325   const scrub_info_t *scrub_info() const {
 326     if (!scrub_infop)
 327       scrub_info_create();
 328     return scrub_infop.get();
 329   }
 330
 331   // -- accessors --
 332   inodeno_t ino()     const { return inode->ino(); }          // deprecate me?
 333   frag_t    get_frag()    const { return frag; }
 334   dirfrag_t dirfrag() const { return dirfrag_t(inode->ino(), frag); }
 335
 336   CInode *get_inode()    { return inode; }
 337   const CInode *get_inode() const { return inode; }
 338   CDir *get_parent_dir() { return inode->get_parent_dir(); }
 339
 340   dentry_key_map::iterator begin() { return items.begin(); }
 341   dentry_key_map::iterator end() { return items.end(); }
 342   dentry_key_map::iterator lower_bound(dentry_key_t key) { return items.lower_bound(key); }
 343
 344   unsigned get_num_head_items() const { return num_head_items; }
 345   unsigned get_num_head_null() const { return num_head_null; }
 346   unsigned get_num_snap_items() const { return num_snap_items; }
 347   unsigned get_num_snap_null() const { return num_snap_null; }
 348   unsigned get_num_any() const { return num_head_items + num_head_null + num_snap_items + num_snap_null; }
 349
 350   bool check_rstats(bool scrub=false);
 351
 352   void inc_num_dirty() { num_dirty++; }
 353   void dec_num_dirty() {
 354     ceph_assert(num_dirty > 0);
 355     num_dirty--;
 356   }
 357   int get_num_dirty() const {
 358     return num_dirty;
 359   }
 360
 361   void adjust_num_inodes_with_caps(int d);
 362
 363   int64_t get_frag_size() const {
 364     return get_projected_fnode()->fragstat.size();
 365   }
 366
 367   // -- dentries and inodes --
 368   CDentry* lookup_exact_snap(std::string_view dname, snapid_t last);
 369   CDentry* lookup(std::string_view n, snapid_t snap=CEPH_NOSNAP);
 370
 371   CDentry* add_null_dentry(std::string_view dname,
 372                            snapid_t first=2, snapid_t last=CEPH_NOSNAP);
 373   CDentry* add_primary_dentry(std::string_view dname, CInode *in, mempool::mds_co::string alternate_name,
 374                               snapid_t first=2, snapid_t last=CEPH_NOSNAP);
 375   CDentry* add_remote_dentry(std::string_view dname, inodeno_t ino, unsigned char d_type,
 376                              mempool::mds_co::string alternate_name,
 377                              snapid_t first=2, snapid_t last=CEPH_NOSNAP);
 378   void remove_dentry( CDentry *dn );         // delete dentry
 379   void link_remote_inode( CDentry *dn, inodeno_t ino, unsigned char d_type);
 380   void link_remote_inode( CDentry *dn, CInode *in );
 381   void link_primary_inode( CDentry *dn, CInode *in );
 382   void unlink_inode(CDentry *dn, bool adjust_lru=true);
 383   void try_remove_unlinked_dn(CDentry *dn);
 384
 385   void add_to_bloom(CDentry *dn);
 386   bool is_in_bloom(std::string_view name);
 387   bool has_bloom() { return (bloom ? true : false); }
 388   void remove_bloom() {
 389     bloom.reset();
 390   }
 391
 392   void try_remove_dentries_for_stray();
 393   bool try_trim_snap_dentry(CDentry *dn, const std::set<snapid_t>& snaps);
 394
 395   void split(int bits, std::vector<CDir*>* subs, MDSContext::vec& waiters, bool replay);
 396   void merge(const std::vector<CDir*>& subs, MDSContext::vec& waiters, bool replay);
 397
 398   bool should_split() const {
 399     return g_conf()->mds_bal_split_size > 0 &&
 400            (int)get_frag_size() > g_conf()->mds_bal_split_size;
 401   }
 402   bool should_split_fast() const;
 403   bool should_merge() const;
 404
 405   mds_authority_t authority() const override;
 406   mds_authority_t get_dir_auth() const { return dir_auth; }
 407   void set_dir_auth(const mds_authority_t &a);
 408   void set_dir_auth(mds_rank_t a) { set_dir_auth(mds_authority_t(a, CDIR_AUTH_UNKNOWN)); }
 409   bool is_ambiguous_dir_auth() const {
 410     return dir_auth.second != CDIR_AUTH_UNKNOWN;
 411   }
 412   bool is_full_dir_auth() const {
 413     return is_auth() && !is_ambiguous_dir_auth();
 414   }
 415   bool is_full_dir_nonauth() const {
 416     return !is_auth() && !is_ambiguous_dir_auth();
 417   }
 418
 419   bool is_subtree_root() const {
 420     return dir_auth != CDIR_AUTH_DEFAULT;
 421   }
 422
 423   bool contains(CDir *x);  // true if we are x or an ancestor of x
 424
 425   // for giving to clients
 426   void get_dist_spec(std::set<mds_rank_t>& ls, mds_rank_t auth) {
 427     if (is_auth()) {
 428       list_replicas(ls);
 429       if (!ls.empty())
 430         ls.insert(auth);
 431     }
 432   }
 433
 434   static void encode_dirstat(ceph::buffer::list& bl, const session_info_t& info, const DirStat& ds);
 435
 436   void _encode_base(ceph::buffer::list& bl) {
 437     ENCODE_START(1, 1, bl);
 438     encode(first, bl);
 439     encode(*fnode, bl);
 440     encode(dir_rep, bl);
 441     encode(dir_rep_by, bl);
 442     ENCODE_FINISH(bl);
 443   }
 444   void _decode_base(ceph::buffer::list::const_iterator& p) {
 445     DECODE_START(1, p);
 446     decode(first, p);
 447     {
 448       auto _fnode = allocate_fnode();
 449       decode(*_fnode, p);
 450       reset_fnode(std::move(_fnode));
 451     }
 452     decode(dir_rep, p);
 453     decode(dir_rep_by, p);
 454     DECODE_FINISH(p);
 455   }
 456
 457   // -- state --
 458   bool is_complete() { return state & STATE_COMPLETE; }
 459   bool is_exporting() { return state & STATE_EXPORTING; }
 460   bool is_importing() { return state & STATE_IMPORTING; }
 461   bool is_dirty_dft() { return state & STATE_DIRTYDFT; }
 462
 463   int get_dir_rep() const { return dir_rep; }
 464   bool is_rep() const {
 465     if (dir_rep == REP_NONE) return false;
 466     return true;
 467   }
 468   bool can_rep() const;
 469
 470   // -- fetch --
 471   object_t get_ondisk_object() {
 472     return file_object_t(ino(), frag);
 473   }
 474   void fetch(MDSContext *c, bool ignore_authpinnability=false);
 475   void fetch(MDSContext *c, std::string_view want_dn, bool ignore_authpinnability=false);
 476   void fetch(MDSContext *c, const std::set<dentry_key_t>& keys);
 477
 478 #if 0  // unused?
 479   void wait_for_commit(Context *c, version_t v=0);
 480 #endif
 481   void commit_to(version_t want);
 482   void commit(version_t want, MDSContext *c,
 483               bool ignore_authpinnability=false, int op_prio=-1);
 484
 485   // -- dirtyness --
 486   version_t get_committing_version() const { return committing_version; }
 487   version_t get_committed_version() const { return committed_version; }
 488   void set_committed_version(version_t v) { committed_version = v; }
 489
 490   void mark_complete();
 491
 492   // -- reference counting --
 493   void first_get() override;
 494   void last_put() override;
 495
 496   bool is_waiting_for_dentry(std::string_view dname, snapid_t snap) {
 497     return waiting_on_dentry.count(string_snap_t(dname, snap));
 498   }
 499   void add_dentry_waiter(std::string_view dentry, snapid_t snap, MDSContext *c);
 500   void take_dentry_waiting(std::string_view dentry, snapid_t first, snapid_t last, MDSContext::vec& ls);
 501   void take_sub_waiting(MDSContext::vec& ls);  // dentry or ino
 502
 503   void add_waiter(uint64_t mask, MDSContext *c) override;
 504   void take_waiting(uint64_t mask, MDSContext::vec& ls) override;  // may include dentry waiters
 505   void finish_waiting(uint64_t mask, int result = 0);    // ditto
 506
 507   // -- import/export --
 508   mds_rank_t get_export_pin(bool inherit=true) const;
 509   bool is_exportable(mds_rank_t dest) const;
 510
 511   void encode_export(ceph::buffer::list& bl);
 512   void finish_export();
 513   void abort_export() {
 514     put(PIN_TEMPEXPORTING);
 515   }
 516   void decode_import(ceph::buffer::list::const_iterator& blp, LogSegment *ls);
 517   void abort_import();
 518
 519   // -- auth pins --
 520   bool can_auth_pin(int *err_ret=nullptr) const override;
 521   int get_auth_pins() const { return auth_pins; }
 522   int get_dir_auth_pins() const { return dir_auth_pins; }
 523   void auth_pin(void *who) override;
 524   void auth_unpin(void *who) override;
 525
 526   void adjust_nested_auth_pins(int dirinc, void *by);
 527   void verify_fragstat();
 528
 529   void _walk_tree(std::function<bool(CDir*)> cb);
 530
 531   bool freeze_tree();
 532   void _freeze_tree();
 533   void unfreeze_tree();
 534   void adjust_freeze_after_rename(CDir *dir);
 535
 536   bool freeze_dir();
 537   void _freeze_dir();
 538   void unfreeze_dir();
 539
 540   void maybe_finish_freeze();
 541
 542   std::pair<bool,bool> is_freezing_or_frozen_tree() const {
 543     if (freeze_tree_state) {
 544       if (freeze_tree_state->frozen)
 545         return std::make_pair(false, true);
 546       return std::make_pair(true, false);
 547     }
 548     return std::make_pair(false, false);
 549   }
 550
 551   bool is_freezing() const override { return is_freezing_dir() || is_freezing_tree(); }
 552   bool is_freezing_tree() const {
 553     if (!num_freezing_trees)
 554       return false;
 555     return is_freezing_or_frozen_tree().first;
 556   }
 557   bool is_freezing_tree_root() const { return state & STATE_FREEZINGTREE; }
 558   bool is_freezing_dir() const { return state & STATE_FREEZINGDIR; }
 559
 560   bool is_frozen() const override { return is_frozen_dir() || is_frozen_tree(); }
 561   bool is_frozen_tree() const {
 562     if (!num_frozen_trees)
 563       return false;
 564     return is_freezing_or_frozen_tree().second;
 565   }
 566   bool is_frozen_tree_root() const { return state & STATE_FROZENTREE; }
 567   bool is_frozen_dir() const { return state & STATE_FROZENDIR; }
 568
 569   bool is_freezeable(bool freezing=false) const {
 570     // no nested auth pins.
 571     if (auth_pins - (freezing ? 1 : 0) > 0 ||
 572         (freeze_tree_state && freeze_tree_state->auth_pins != auth_pins))
 573       return false;
 574
 575     // inode must not be frozen.
 576     if (!is_subtree_root() && inode->is_frozen())
 577       return false;
 578
 579     return true;
 580   }
 581
 582   bool is_freezeable_dir(bool freezing=false) const {
 583     if ((auth_pins - freezing) > 0 || dir_auth_pins > 0)
 584       return false;
 585
 586     // if not subtree root, inode must not be frozen (tree--frozen_dir is okay).
 587     if (!is_subtree_root() && inode->is_frozen() && !inode->is_frozen_dir())
 588       return false;
 589
 590     return true;
 591   }
 592
 593   bool is_any_freezing_or_frozen_inode() const {
 594     return num_frozen_inodes || !freezing_inodes.empty();
 595   }
 596   bool is_auth_pinned_by_lock_cache() const {
 597     return frozen_inode_suppressed;
 598   }
 599   void disable_frozen_inode() {
 600     ceph_assert(num_frozen_inodes == 0);
 601     frozen_inode_suppressed++;
 602   }
 603   void enable_frozen_inode();
 604
 605   std::ostream& print_db_line_prefix(std::ostream& out) override;
 606   void print(std::ostream& out) override;
 607   void dump(ceph::Formatter *f, int flags = DUMP_DEFAULT) const;
 608   void dump_load(ceph::Formatter *f);
 609
 610   // context
 611   MDCache *mdcache;
 612
 613   CInode *inode;  // my inode
 614   frag_t frag;   // my frag
 615
 616   snapid_t first = 2;
 617   mempool::mds_co::compact_map<snapid_t,old_rstat_t> dirty_old_rstat;  // [value.first,key]
 618
 619   // my inodes with dirty rstat data
 620   elist<CInode*> dirty_rstat_inodes;
 621
 622   elist<CDentry*> dirty_dentries;
 623   elist<CDir*>::item item_dirty, item_new;
 624
 625   // lock caches that auth-pin me
 626   elist<MDLockCache::DirItem*> lock_caches_with_auth_pins;
 627
 628   // all dirfrags within freezing/frozen tree reference the 'state'
 629   std::shared_ptr<freeze_tree_state_t> freeze_tree_state;
 630
 631 protected:
 632   // friends
 633   friend class Migrator;
 634   friend class CInode;
 635   friend class MDCache;
 636   friend class MDiscover;
 637   friend class MDBalancer;
 638
 639   friend class CDirDiscover;
 640   friend class CDirExport;
 641   friend class C_IO_Dir_TMAP_Fetched;
 642   friend class C_IO_Dir_OMAP_Fetched;
 643   friend class C_IO_Dir_OMAP_FetchedMore;
 644   friend class C_IO_Dir_Committed;
 645   friend class C_IO_Dir_Commit_Ops;
 646
 647   void _omap_fetch(MDSContext *fin, const std::set<dentry_key_t>& keys);
 648   void _omap_fetch_more(version_t omap_version, bufferlist& hdrbl,
 649                         map<string, bufferlist>& omap, MDSContext *fin);
 650   CDentry *_load_dentry(
 651       std::string_view key,
 652       std::string_view dname,
 653       snapid_t last,
 654       ceph::buffer::list &bl,
 655       int pos,
 656       const std::set<snapid_t> *snaps,
 657       double rand_threshold,
 658       bool *force_dirty);
 659
 660   /**
 661    * Go bad due to a damaged dentry (register with damagetable and go BADFRAG)
 662    */
 663   void go_bad_dentry(snapid_t last, std::string_view dname);
 664
 665   /**
 666    * Go bad due to a damaged header (register with damagetable and go BADFRAG)
 667    */
 668   void go_bad(bool complete);
 669
 670   void _omap_fetched(ceph::buffer::list& hdrbl, std::map<std::string, ceph::buffer::list>& omap,
 671                      bool complete, int r);
 672
 673   // -- commit --
 674   void _commit(version_t want, int op_prio);
 675   void _omap_commit_ops(int r, int op_prio, int64_t metapool, version_t version, bool _new,
 676                         vector<dentry_commit_item> &to_set, bufferlist &dfts,
 677                         vector<string> &to_remove,
 678                         mempool::mds_co::compact_set<mempool::mds_co::string> &_stale);
 679   void _encode_primary_inode_base(dentry_commit_item &item, bufferlist &dfts,
 680                                   bufferlist &bl);
 681   void _omap_commit(int op_prio);
 682   void _parse_dentry(CDentry *dn, dentry_commit_item &item,
 683                      const set<snapid_t> *snaps, bufferlist &bl);
 684   void _committed(int r, version_t v);
 685
 686   static fnode_const_ptr empty_fnode;
 687   // fnode is a pointer to constant fnode_t, the constant fnode_t can be shared
 688   // by CDir and log events. To update fnode, read-copy-update should be used.
 689
 690   fnode_const_ptr fnode = empty_fnode;
 691
 692   version_t projected_version = 0;
 693   mempool::mds_co::list<fnode_const_ptr> projected_fnode;
 694
 695   std::unique_ptr<scrub_info_t> scrub_infop;
 696
 697   // contents of this directory
 698   dentry_key_map items;       // non-null AND null
 699   unsigned num_head_items = 0;
 700   unsigned num_head_null = 0;
 701   unsigned num_snap_items = 0;
 702   unsigned num_snap_null = 0;
 703
 704   int num_dirty = 0;
 705
 706   int num_inodes_with_caps = 0;
 707
 708   // state
 709   version_t committing_version = 0;
 710   version_t committed_version = 0;
 711
 712   mempool::mds_co::compact_set<mempool::mds_co::string> stale_items;
 713
 714   // lock nesting, freeze
 715   static int num_frozen_trees;
 716   static int num_freezing_trees;
 717
 718   // freezing/frozen inodes in this dirfrag
 719   int num_frozen_inodes = 0;
 720   int frozen_inode_suppressed = 0;
 721   elist<CInode*> freezing_inodes;
 722
 723   int dir_auth_pins = 0;
 724
 725   // cache control  (defined for authority; hints for replicas)
 726   __s32      dir_rep;
 727   mempool::mds_co::compact_set<__s32> dir_rep_by;      // if dir_rep == REP_LIST
 728
 729   // popularity
 730   dirfrag_load_vec_t pop_me;
 731   dirfrag_load_vec_t pop_nested;
 732   dirfrag_load_vec_t pop_auth_subtree;
 733   dirfrag_load_vec_t pop_auth_subtree_nested;
 734
 735   ceph::coarse_mono_time last_popularity_sample = ceph::coarse_mono_clock::zero();
 736
 737   load_spread_t pop_spread;
 738
 739   elist<CInode*> pop_lru_subdirs;
 740
 741   std::unique_ptr<bloom_filter> bloom; // XXX not part of mempool::mds_co
 742   /* If you set up the bloom filter, you must keep it accurate!
 743    * It's deleted when you mark_complete() and is deliberately not serialized.*/
 744
 745   mempool::mds_co::compact_set<mempool::mds_co::string> wanted_items;
 746   mempool::mds_co::compact_map<version_t, MDSContext::vec_alloc<mempool::mds_co::pool_allocator> > waiting_for_commit;
 747
 748   // -- waiters --
 749   mempool::mds_co::compact_map< string_snap_t, MDSContext::vec_alloc<mempool::mds_co::pool_allocator> > waiting_on_dentry; // FIXME string_snap_t not in mempool
 750
 751 private:
 752   friend std::ostream& operator<<(std::ostream& out, const class CDir& dir);
 753
 754   void log_mark_dirty();
 755
 756   /**
 757    * Create a scrub_info_t struct for the scrub_infop pointer.
 758    */
 759   void scrub_info_create() const;
 760   /**
 761    * Delete the scrub_infop if it's not got any useful data.
 762    */
 763   void scrub_maybe_delete_info();
 764
 765   void link_inode_work( CDentry *dn, CInode *in );
 766   void unlink_inode_work( CDentry *dn );
 767   void remove_null_dentries();
 768   void purge_stale_snap_data(const std::set<snapid_t>& snaps);
 769
 770   void prepare_new_fragment(bool replay);
 771   void prepare_old_fragment(std::map<string_snap_t, MDSContext::vec >& dentry_waiters, bool replay);
 772   void steal_dentry(CDentry *dn);  // from another dir.  used by merge/split.
 773   void finish_old_fragment(MDSContext::vec& waiters, bool replay);
 774   void init_fragment_pins();
 775   std::string get_path() const;
 776
 777   // -- authority --
 778   /*
 779    *     normal: <parent,unknown>   !subtree_root
 780    * delegation: <mds,unknown>       subtree_root
 781    *  ambiguous: <mds1,mds2>         subtree_root
 782    *             <parent,mds2>       subtree_root
 783    */
 784   mds_authority_t dir_auth;
 785 };
 786
 787 #endif