ceph/src/mds/Migrator.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  * Handles the import and export of  mds authorities and actual cache data.
  14  * See src/doc/exports.txt for a description.
  15  */
  16
  17 #ifndef CEPH_MDS_MIGRATOR_H
  18 #define CEPH_MDS_MIGRATOR_H
  19
  20 #include "include/types.h"
  21
  22 #include "MDSContext.h"
  23
  24 #include <map>
  25 #include <list>
  26 #include <set>
  27 #include <string_view>
  28
  29 class MDSRank;
  30 class CDir;
  31 class CInode;
  32 class CDentry;
  33 class Session;
  34
  35 #include "messages/MExportCaps.h"
  36 #include "messages/MExportCapsAck.h"
  37 #include "messages/MExportDir.h"
  38 #include "messages/MExportDirAck.h"
  39 #include "messages/MExportDirCancel.h"
  40 #include "messages/MExportDirDiscover.h"
  41 #include "messages/MExportDirDiscoverAck.h"
  42 #include "messages/MExportDirFinish.h"
  43 #include "messages/MExportDirNotify.h"
  44 #include "messages/MExportDirNotifyAck.h"
  45 #include "messages/MExportDirPrep.h"
  46 #include "messages/MExportDirPrepAck.h"
  47 #include "messages/MGatherCaps.h"
  48
  49 class EImportStart;
  50
  51 class Migrator {
  52 public:
  53   // export stages.  used to clean up intelligently if there's a failure.
  54   const static int EXPORT_CANCELLED     = 0;  // cancelled
  55   const static int EXPORT_CANCELLING    = 1;  // waiting for cancel notifyacks
  56   const static int EXPORT_LOCKING       = 2;  // acquiring locks
  57   const static int EXPORT_DISCOVERING   = 3;  // dest is disovering export dir
  58   const static int EXPORT_FREEZING      = 4;  // we're freezing the dir tree
  59   const static int EXPORT_PREPPING      = 5;  // sending dest spanning tree to export bounds
  60   const static int EXPORT_WARNING       = 6;  // warning bystanders of dir_auth_pending
  61   const static int EXPORT_EXPORTING     = 7;  // sent actual export, waiting for ack
  62   const static int EXPORT_LOGGINGFINISH = 8;  // logging EExportFinish
  63   const static int EXPORT_NOTIFYING     = 9;  // waiting for notifyacks
  64   static std::string_view get_export_statename(int s) {
  65     switch (s) {
  66     case EXPORT_CANCELLING: return "cancelling";
  67     case EXPORT_LOCKING: return "locking";
  68     case EXPORT_DISCOVERING: return "discovering";
  69     case EXPORT_FREEZING: return "freezing";
  70     case EXPORT_PREPPING: return "prepping";
  71     case EXPORT_WARNING: return "warning";
  72     case EXPORT_EXPORTING: return "exporting";
  73     case EXPORT_LOGGINGFINISH: return "loggingfinish";
  74     case EXPORT_NOTIFYING: return "notifying";
  75     default: ceph_abort(); return std::string_view();
  76     }
  77   }
  78
  79   // -- imports --
  80   const static int IMPORT_DISCOVERING   = 1; // waiting for prep
  81   const static int IMPORT_DISCOVERED    = 2; // waiting for prep
  82   const static int IMPORT_PREPPING      = 3; // opening dirs on bounds
  83   const static int IMPORT_PREPPED       = 4; // opened bounds, waiting for import
  84   const static int IMPORT_LOGGINGSTART  = 5; // got import, logging EImportStart
  85   const static int IMPORT_ACKING        = 6; // logged EImportStart, sent ack, waiting for finish
  86   const static int IMPORT_FINISHING     = 7; // sent cap imports, waiting for finish
  87   const static int IMPORT_ABORTING      = 8; // notifying bystanders of an abort before unfreezing
  88   static std::string_view get_import_statename(int s) {
  89     switch (s) {
  90     case IMPORT_DISCOVERING: return "discovering";
  91     case IMPORT_DISCOVERED: return "discovered";
  92     case IMPORT_PREPPING: return "prepping";
  93     case IMPORT_PREPPED: return "prepped";
  94     case IMPORT_LOGGINGSTART: return "loggingstart";
  95     case IMPORT_ACKING: return "acking";
  96     case IMPORT_FINISHING: return "finishing";
  97     case IMPORT_ABORTING: return "aborting";
  98     default: ceph_abort(); return std::string_view();
  99     }
 100   }
 101
 102   // -- cons --
 103   Migrator(MDSRank *m, MDCache *c);
 104
 105   void handle_conf_change(const ConfigProxy& conf,
 106                           const std::set <std::string> &changed,
 107                           const MDSMap &mds_map);
 108
 109 protected:
 110   struct export_base_t {
 111     dirfrag_t dirfrag;
 112     mds_rank_t dest;
 113     unsigned pending_children;
 114     uint64_t export_queue_gen;
 115     bool restart = false;
 116     export_base_t(dirfrag_t df, mds_rank_t d, unsigned c, uint64_t g) :
 117       dirfrag(df), dest(d), pending_children(c), export_queue_gen(g) {}
 118   };
 119
 120   // export fun
 121   struct export_state_t {
 122     int state = 0;
 123     mds_rank_t peer = MDS_RANK_NONE;
 124     uint64_t tid = 0;
 125     std::set<mds_rank_t> warning_ack_waiting;
 126     std::set<mds_rank_t> notify_ack_waiting;
 127     std::map<inodeno_t,std::map<client_t,Capability::Import> > peer_imported;
 128     MutationRef mut;
 129     size_t approx_size = 0;
 130     // for freeze tree deadlock detection
 131     utime_t last_cum_auth_pins_change;
 132     int last_cum_auth_pins = 0;
 133     int num_remote_waiters = 0; // number of remote authpin waiters
 134     export_state_t() {}
 135
 136     std::shared_ptr<export_base_t> parent;
 137   };
 138   std::map<CDir*, export_state_t>  export_state;
 139   typedef map<CDir*, export_state_t>::iterator export_state_iterator;
 140
 141   uint64_t total_exporting_size = 0;
 142   unsigned num_locking_exports = 0; // exports in locking state (approx_size == 0)
 143
 144   std::list<pair<dirfrag_t,mds_rank_t> >  export_queue;
 145   uint64_t export_queue_gen = 1;
 146
 147   // import fun
 148   struct import_state_t {
 149     int state;
 150     mds_rank_t peer;
 151     uint64_t tid;
 152     std::set<mds_rank_t> bystanders;
 153     std::list<dirfrag_t> bound_ls;
 154     std::list<ScatterLock*> updated_scatterlocks;
 155     std::map<client_t,pair<Session*,uint64_t> > session_map;
 156     std::map<CInode*, std::map<client_t,Capability::Export> > peer_exports;
 157     MutationRef mut;
 158     import_state_t() : state(0), peer(0), tid(0), mut() {}
 159   };
 160
 161   std::map<dirfrag_t, import_state_t>  import_state;
 162
 163   void handle_export_discover_ack(const MExportDirDiscoverAck::const_ref &m);
 164   void export_frozen(CDir *dir, uint64_t tid);
 165   void handle_export_prep_ack(const MExportDirPrepAck::const_ref &m);
 166   void export_sessions_flushed(CDir *dir, uint64_t tid);
 167   void export_go(CDir *dir);
 168   void export_go_synced(CDir *dir, uint64_t tid);
 169   void export_try_cancel(CDir *dir, bool notify_peer=true);
 170   void export_cancel_finish(export_state_iterator& it);
 171   void export_reverse(CDir *dir, export_state_t& stat);
 172   void export_notify_abort(CDir *dir, export_state_t& stat, std::set<CDir*>& bounds);
 173   void handle_export_ack(const MExportDirAck::const_ref &m);
 174   void export_logged_finish(CDir *dir);
 175   void handle_export_notify_ack(const MExportDirNotifyAck::const_ref &m);
 176   void export_finish(CDir *dir);
 177
 178   void handle_gather_caps(const MGatherCaps::const_ref &m);
 179
 180   friend class C_MDC_ExportFreeze;
 181   friend class C_MDS_ExportFinishLogged;
 182   friend class C_M_ExportGo;
 183   friend class C_M_ExportSessionsFlushed;
 184   friend class C_MDS_ExportDiscover;
 185   friend class C_MDS_ExportPrep;
 186   friend class MigratorContext;
 187   friend class MigratorLogContext;
 188
 189   // importer
 190   void handle_export_discover(const MExportDirDiscover::const_ref &m, bool started=false);
 191   void handle_export_cancel(const MExportDirCancel::const_ref &m);
 192   void handle_export_prep(const MExportDirPrep::const_ref &m, bool did_assim=false);
 193   void handle_export_dir(const MExportDir::const_ref &m);
 194
 195   void import_reverse_discovering(dirfrag_t df);
 196   void import_reverse_discovered(dirfrag_t df, CInode *diri);
 197   void import_reverse_prepping(CDir *dir, import_state_t& stat);
 198   void import_remove_pins(CDir *dir, std::set<CDir*>& bounds);
 199   void import_reverse_unfreeze(CDir *dir);
 200   void import_reverse_final(CDir *dir);
 201   void import_notify_abort(CDir *dir, std::set<CDir*>& bounds);
 202   void import_notify_finish(CDir *dir, std::set<CDir*>& bounds);
 203   void import_logged_start(dirfrag_t df, CDir *dir, mds_rank_t from,
 204                            std::map<client_t,pair<Session*,uint64_t> >& imported_session_map);
 205   void handle_export_finish(const MExportDirFinish::const_ref &m);
 206
 207   void handle_export_caps(const MExportCaps::const_ref &m);
 208   void handle_export_caps_ack(const MExportCapsAck::const_ref &m);
 209   void logged_import_caps(CInode *in,
 210                           mds_rank_t from,
 211                           std::map<client_t,pair<Session*,uint64_t> >& imported_session_map,
 212                           std::map<CInode*, std::map<client_t,Capability::Export> >& cap_imports);
 213
 214
 215   friend class C_MDS_ImportDirLoggedStart;
 216   friend class C_MDS_ImportDirLoggedFinish;
 217   friend class C_M_LoggedImportCaps;
 218
 219   // bystander
 220   void handle_export_notify(const MExportDirNotify::const_ref &m);
 221
 222
 223 public:
 224
 225   void dispatch(const Message::const_ref &);
 226
 227   void show_importing();
 228   void show_exporting();
 229
 230   int get_num_exporting() const { return export_state.size(); }
 231   int get_export_queue_size() const { return export_queue.size(); }
 232
 233   // -- status --
 234   int is_exporting(CDir *dir) const {
 235     auto it = export_state.find(dir);
 236     if (it != export_state.end()) return it->second.state;
 237     return 0;
 238   }
 239   bool is_exporting() const { return !export_state.empty(); }
 240   int is_importing(dirfrag_t df) const {
 241     auto it = import_state.find(df);
 242     if (it != import_state.end()) return it->second.state;
 243     return 0;
 244   }
 245   bool is_importing() const { return !import_state.empty(); }
 246
 247   bool is_ambiguous_import(dirfrag_t df) const {
 248     auto it = import_state.find(df);
 249     if (it == import_state.end())
 250       return false;
 251     if (it->second.state >= IMPORT_LOGGINGSTART &&
 252         it->second.state < IMPORT_ABORTING)
 253       return true;
 254     return false;
 255   }
 256
 257   int get_import_state(dirfrag_t df) const {
 258     auto it = import_state.find(df);
 259     ceph_assert(it != import_state.end());
 260     return it->second.state;
 261   }
 262   int get_import_peer(dirfrag_t df) const {
 263     auto it = import_state.find(df);
 264     ceph_assert(it != import_state.end());
 265     return it->second.peer;
 266   }
 267
 268   int get_export_state(CDir *dir) const {
 269     auto it = export_state.find(dir);
 270     ceph_assert(it != export_state.end());
 271     return it->second.state;
 272   }
 273   // this returns true if we are export @dir,
 274   // and are not waiting for @who to be
 275   // be warned of ambiguous auth.
 276   // only returns meaningful results during EXPORT_WARNING state.
 277   bool export_has_warned(CDir *dir, mds_rank_t who) {
 278     auto it = export_state.find(dir);
 279     ceph_assert(it != export_state.end());
 280     ceph_assert(it->second.state == EXPORT_WARNING);
 281     return (it->second.warning_ack_waiting.count(who) == 0);
 282   }
 283
 284   bool export_has_notified(CDir *dir, mds_rank_t who) const {
 285     auto it = export_state.find(dir);
 286     ceph_assert(it != export_state.end());
 287     ceph_assert(it->second.state == EXPORT_NOTIFYING);
 288     return (it->second.notify_ack_waiting.count(who) == 0);
 289   }
 290
 291   void export_freeze_inc_num_waiters(CDir *dir) {
 292     auto it = export_state.find(dir);
 293     ceph_assert(it != export_state.end());
 294     it->second.num_remote_waiters++;
 295   }
 296   void find_stale_export_freeze();
 297
 298   // -- misc --
 299   void handle_mds_failure_or_stop(mds_rank_t who);
 300
 301   void audit();
 302
 303   // -- import/export --
 304   // exporter
 305   void dispatch_export_dir(MDRequestRef& mdr, int count);
 306   void export_dir(CDir *dir, mds_rank_t dest);
 307   void export_empty_import(CDir *dir);
 308
 309   void export_dir_nicely(CDir *dir, mds_rank_t dest);
 310   void maybe_do_queued_export();
 311   void clear_export_queue() {
 312     export_queue.clear();
 313     export_queue_gen++;
 314   }
 315
 316   void maybe_split_export(CDir* dir, uint64_t max_size, bool null_okay,
 317                           vector<pair<CDir*, size_t> >& results);
 318   void child_export_finish(std::shared_ptr<export_base_t>& parent, bool success);
 319
 320   void get_export_lock_set(CDir *dir, MutationImpl::LockOpVec& lov);
 321   void get_export_client_set(CDir *dir, std::set<client_t> &client_set);
 322   void get_export_client_set(CInode *in, std::set<client_t> &client_set);
 323
 324   void encode_export_inode(CInode *in, bufferlist& bl,
 325                            std::map<client_t,entity_inst_t>& exported_client_map,
 326                            std::map<client_t,client_metadata_t>& exported_client_metadata_map);
 327   void encode_export_inode_caps(CInode *in, bool auth_cap, bufferlist& bl,
 328                                 std::map<client_t,entity_inst_t>& exported_client_map,
 329                                 std::map<client_t,client_metadata_t>& exported_client_metadata_map);
 330   void finish_export_inode(CInode *in, mds_rank_t target,
 331                            std::map<client_t,Capability::Import>& peer_imported,
 332                            MDSContext::vec& finished);
 333   void finish_export_inode_caps(CInode *in, mds_rank_t target,
 334                                 std::map<client_t,Capability::Import>& peer_imported);
 335
 336
 337   uint64_t encode_export_dir(bufferlist& exportbl,
 338                         CDir *dir,
 339                         std::map<client_t,entity_inst_t>& exported_client_map,
 340                         std::map<client_t,client_metadata_t>& exported_client_metadata_map);
 341   void finish_export_dir(CDir *dir, mds_rank_t target,
 342                          std::map<inodeno_t,std::map<client_t,Capability::Import> >& peer_imported,
 343                          MDSContext::vec& finished, int *num_dentries);
 344
 345   void clear_export_proxy_pins(CDir *dir);
 346
 347   void export_caps(CInode *in);
 348
 349   void decode_import_inode(CDentry *dn, bufferlist::const_iterator& blp,
 350                            mds_rank_t oldauth, LogSegment *ls,
 351                            std::map<CInode*, std::map<client_t,Capability::Export> >& cap_imports,
 352                            std::list<ScatterLock*>& updated_scatterlocks);
 353   void decode_import_inode_caps(CInode *in, bool auth_cap, bufferlist::const_iterator &blp,
 354                                 std::map<CInode*, std::map<client_t,Capability::Export> >& cap_imports);
 355   void finish_import_inode_caps(CInode *in, mds_rank_t from, bool auth_cap,
 356                                 const std::map<client_t,pair<Session*,uint64_t> >& smap,
 357                                 const std::map<client_t,Capability::Export> &export_map,
 358                                 std::map<client_t,Capability::Import> &import_map);
 359   int decode_import_dir(bufferlist::const_iterator& blp,
 360                         mds_rank_t oldauth,
 361                         CDir *import_root,
 362                         EImportStart *le,
 363                         LogSegment *ls,
 364                         std::map<CInode*, std::map<client_t,Capability::Export> >& cap_imports,
 365                         std::list<ScatterLock*>& updated_scatterlocks);
 366
 367   void import_reverse(CDir *dir);
 368
 369   void import_finish(CDir *dir, bool notify, bool last=true);
 370
 371 private:
 372   MDSRank *mds;
 373   MDCache *cache;
 374   uint64_t max_export_size = 0;
 375   bool inject_session_race = false;
 376 };
 377
 378 #endif