ceph/src/osd/PGBackend.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) 2013,2014 Inktank Storage, Inc.
   7  * Copyright (C) 2013,2014 Cloudwatt <libre.licensing@cloudwatt.com>
   8  *
   9  * Author: Loic Dachary <loic@dachary.org>
  10  *
  11  * This is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License version 2.1, as published by the Free Software
  14  * Foundation.  See file COPYING.
  15  *
  16  */
  17
  18 #ifndef PGBACKEND_H
  19 #define PGBACKEND_H
  20
  21 #include "osd_types.h"
  22 #include "common/WorkQueue.h"
  23 #include "include/Context.h"
  24 #include "os/ObjectStore.h"
  25 #include "common/LogClient.h"
  26 #include <string>
  27 #include "PGTransaction.h"
  28
  29 namespace Scrub {
  30   class Store;
  31 }
  32 struct shard_info_wrapper;
  33 struct inconsistent_obj_wrapper;
  34
  35 //forward declaration
  36 class OSDMap;
  37 class PGLog;
  38 typedef ceph::shared_ptr<const OSDMap> OSDMapRef;
  39
  40  /**
  41   * PGBackend
  42   *
  43   * PGBackend defines an interface for logic handling IO and
  44   * replication on RADOS objects.  The PGBackend implementation
  45   * is responsible for:
  46   *
  47   * 1) Handling client operations
  48   * 2) Handling object recovery
  49   * 3) Handling object access
  50   * 4) Handling scrub, deep-scrub, repair
  51   */
  52  class PGBackend {
  53  public:
  54    CephContext* cct;
  55  protected:
  56    ObjectStore *store;
  57    const coll_t coll;
  58    ObjectStore::CollectionHandle &ch;
  59  public:
  60    /**
  61     * Provides interfaces for PGBackend callbacks
  62     *
  63     * The intention is that the parent calls into the PGBackend
  64     * implementation holding a lock and that the callbacks are
  65     * called under the same locks.
  66     */
  67    class Listener {
  68    public:
  69      /// Debugging
  70      virtual DoutPrefixProvider *get_dpp() = 0;
  71
  72      /// Recovery
  73
  74      /**
  75       * Called with the transaction recovering oid
  76       */
  77      virtual void on_local_recover(
  78        const hobject_t &oid,
  79        const ObjectRecoveryInfo &recovery_info,
  80        ObjectContextRef obc,
  81        ObjectStore::Transaction *t
  82        ) = 0;
  83
  84      /**
  85       * Called when transaction recovering oid is durable and
  86       * applied on all replicas
  87       */
  88      virtual void on_global_recover(
  89        const hobject_t &oid,
  90        const object_stat_sum_t &stat_diff
  91        ) = 0;
  92
  93      /**
  94       * Called when peer is recovered
  95       */
  96      virtual void on_peer_recover(
  97        pg_shard_t peer,
  98        const hobject_t &oid,
  99        const ObjectRecoveryInfo &recovery_info
 100        ) = 0;
 101
 102      virtual void begin_peer_recover(
 103        pg_shard_t peer,
 104        const hobject_t oid) = 0;
 105
 106      virtual void failed_push(const list<pg_shard_t> &from, const hobject_t &soid) = 0;
 107      virtual void primary_failed(const hobject_t &soid) = 0;
 108      virtual bool primary_error(const hobject_t& soid, eversion_t v) = 0;
 109
 110      virtual void cancel_pull(const hobject_t &soid) = 0;
 111
 112      virtual void apply_stats(
 113        const hobject_t &soid,
 114        const object_stat_sum_t &delta_stats) = 0;
 115
 116      /**
 117       * Called when a read on the primary fails when pushing
 118       */
 119      virtual void on_primary_error(
 120        const hobject_t &oid,
 121        eversion_t v
 122        ) = 0;
 123
 124
 125      /**
 126       * Bless a context
 127       *
 128       * Wraps a context in whatever outer layers the parent usually
 129       * uses to call into the PGBackend
 130       */
 131      virtual Context *bless_context(Context *c) = 0;
 132      virtual GenContext<ThreadPool::TPHandle&> *bless_gencontext(
 133        GenContext<ThreadPool::TPHandle&> *c) = 0;
 134
 135      virtual void send_message(int to_osd, Message *m) = 0;
 136      virtual void queue_transaction(
 137        ObjectStore::Transaction&& t,
 138        OpRequestRef op = OpRequestRef()
 139        ) = 0;
 140      virtual void queue_transactions(
 141        vector<ObjectStore::Transaction>& tls,
 142        OpRequestRef op = OpRequestRef()
 143        ) = 0;
 144      virtual epoch_t get_epoch() const = 0;
 145      virtual epoch_t get_interval_start_epoch() const = 0;
 146      virtual epoch_t get_last_peering_reset_epoch() const = 0;
 147
 148      virtual const set<pg_shard_t> &get_actingbackfill_shards() const = 0;
 149      virtual const set<pg_shard_t> &get_acting_shards() const = 0;
 150      virtual const set<pg_shard_t> &get_backfill_shards() const = 0;
 151
 152      virtual std::string gen_dbg_prefix() const = 0;
 153
 154      virtual const map<hobject_t, set<pg_shard_t>> &get_missing_loc_shards()
 155        const = 0;
 156
 157      virtual const pg_missing_tracker_t &get_local_missing() const = 0;
 158      virtual const map<pg_shard_t, pg_missing_t> &get_shard_missing()
 159        const = 0;
 160      virtual boost::optional<const pg_missing_const_i &> maybe_get_shard_missing(
 161        pg_shard_t peer) const {
 162        if (peer == primary_shard()) {
 163          return get_local_missing();
 164        } else {
 165          map<pg_shard_t, pg_missing_t>::const_iterator i =
 166            get_shard_missing().find(peer);
 167          if (i == get_shard_missing().end()) {
 168            return boost::optional<const pg_missing_const_i &>();
 169          } else {
 170            return i->second;
 171          }
 172        }
 173      }
 174      virtual const pg_missing_const_i &get_shard_missing(pg_shard_t peer) const {
 175        auto m = maybe_get_shard_missing(peer);
 176        assert(m);
 177        return *m;
 178      }
 179
 180      virtual const map<pg_shard_t, pg_info_t> &get_shard_info() const = 0;
 181      virtual const pg_info_t &get_shard_info(pg_shard_t peer) const {
 182        if (peer == primary_shard()) {
 183          return get_info();
 184        } else {
 185          map<pg_shard_t, pg_info_t>::const_iterator i =
 186            get_shard_info().find(peer);
 187          assert(i != get_shard_info().end());
 188          return i->second;
 189        }
 190      }
 191
 192      virtual const PGLog &get_log() const = 0;
 193      virtual bool pgb_is_primary() const = 0;
 194      virtual OSDMapRef pgb_get_osdmap() const = 0;
 195      virtual const pg_info_t &get_info() const = 0;
 196      virtual const pg_pool_t &get_pool() const = 0;
 197
 198      virtual ObjectContextRef get_obc(
 199        const hobject_t &hoid,
 200        const map<string, bufferlist> &attrs) = 0;
 201
 202      virtual bool try_lock_for_read(
 203        const hobject_t &hoid,
 204        ObcLockManager &manager) = 0;
 205
 206      virtual void release_locks(ObcLockManager &manager) = 0;
 207
 208      virtual void op_applied(
 209        const eversion_t &applied_version) = 0;
 210
 211      virtual bool should_send_op(
 212        pg_shard_t peer,
 213        const hobject_t &hoid) = 0;
 214
 215      virtual void log_operation(
 216        const vector<pg_log_entry_t> &logv,
 217        const boost::optional<pg_hit_set_history_t> &hset_history,
 218        const eversion_t &trim_to,
 219        const eversion_t &roll_forward_to,
 220        bool transaction_applied,
 221        ObjectStore::Transaction &t) = 0;
 222
 223      virtual void pgb_set_object_snap_mapping(
 224        const hobject_t &soid,
 225        const set<snapid_t> &snaps,
 226        ObjectStore::Transaction *t) = 0;
 227
 228      virtual void pgb_clear_object_snap_mapping(
 229        const hobject_t &soid,
 230        ObjectStore::Transaction *t) = 0;
 231
 232      virtual void update_peer_last_complete_ondisk(
 233        pg_shard_t fromosd,
 234        eversion_t lcod) = 0;
 235
 236      virtual void update_last_complete_ondisk(
 237        eversion_t lcod) = 0;
 238
 239      virtual void update_stats(
 240        const pg_stat_t &stat) = 0;
 241
 242      virtual void schedule_recovery_work(
 243        GenContext<ThreadPool::TPHandle&> *c) = 0;
 244
 245      virtual pg_shard_t whoami_shard() const = 0;
 246      int whoami() const {
 247        return whoami_shard().osd;
 248      }
 249      spg_t whoami_spg_t() const {
 250        return get_info().pgid;
 251      }
 252
 253      virtual spg_t primary_spg_t() const = 0;
 254      virtual pg_shard_t primary_shard() const = 0;
 255
 256      virtual uint64_t min_peer_features() const = 0;
 257
 258      virtual hobject_t get_temp_recovery_object(const hobject_t& target,
 259                                                 eversion_t version) = 0;
 260
 261      virtual void send_message_osd_cluster(
 262        int peer, Message *m, epoch_t from_epoch) = 0;
 263      virtual void send_message_osd_cluster(
 264        Message *m, Connection *con) = 0;
 265      virtual void send_message_osd_cluster(
 266        Message *m, const ConnectionRef& con) = 0;
 267      virtual ConnectionRef get_con_osd_cluster(int peer, epoch_t from_epoch) = 0;
 268      virtual entity_name_t get_cluster_msgr_name() = 0;
 269
 270      virtual PerfCounters *get_logger() = 0;
 271
 272      virtual ceph_tid_t get_tid() = 0;
 273
 274      virtual LogClientTemp clog_error() = 0;
 275
 276      virtual bool check_failsafe_full(ostream &ss) = 0;
 277
 278      virtual bool check_osdmap_full(const set<pg_shard_t> &missing_on) = 0;
 279
 280      virtual ~Listener() {}
 281    };
 282    Listener *parent;
 283    Listener *get_parent() const { return parent; }
 284    PGBackend(CephContext* cct, Listener *l, ObjectStore *store, coll_t coll,
 285              ObjectStore::CollectionHandle &ch) :
 286      cct(cct),
 287      store(store),
 288      coll(coll),
 289      ch(ch),
 290      parent(l) {}
 291    bool is_primary() const { return get_parent()->pgb_is_primary(); }
 292    OSDMapRef get_osdmap() const { return get_parent()->pgb_get_osdmap(); }
 293    const pg_info_t &get_info() { return get_parent()->get_info(); }
 294
 295    std::string gen_prefix() const {
 296      return parent->gen_dbg_prefix();
 297    }
 298
 299    /**
 300     * RecoveryHandle
 301     *
 302     * We may want to recover multiple objects in the same set of
 303     * messages.  RecoveryHandle is an interface for the opaque
 304     * object used by the implementation to store the details of
 305     * the pending recovery operations.
 306     */
 307    struct RecoveryHandle {
 308      bool cache_dont_need;
 309
 310      RecoveryHandle(): cache_dont_need(false) {}
 311      virtual ~RecoveryHandle() {}
 312    };
 313
 314    /// Get a fresh recovery operation
 315    virtual RecoveryHandle *open_recovery_op() = 0;
 316
 317    /// run_recovery_op: finish the operation represented by h
 318    virtual void run_recovery_op(
 319      RecoveryHandle *h,     ///< [in] op to finish
 320      int priority           ///< [in] msg priority
 321      ) = 0;
 322
 323    /**
 324     * recover_object
 325     *
 326     * Triggers a recovery operation on the specified hobject_t
 327     * onreadable must be called before onwriteable
 328     *
 329     * On each replica (primary included), get_parent()->on_not_missing()
 330     * must be called when the transaction finalizing the recovery
 331     * is queued.  Similarly, get_parent()->on_readable() must be called
 332     * when the transaction is applied in the backing store.
 333     *
 334     * get_parent()->on_not_degraded() should be called on the primary
 335     * when writes can resume on the object.
 336     *
 337     * obc may be NULL if the primary lacks the object.
 338     *
 339     * head may be NULL only if the head/snapdir is missing
 340     *
 341     * @param missing [in] set of info, missing pairs for queried nodes
 342     * @param overlaps [in] mapping of object to file offset overlaps
 343     */
 344    virtual int recover_object(
 345      const hobject_t &hoid, ///< [in] object to recover
 346      eversion_t v,          ///< [in] version to recover
 347      ObjectContextRef head,  ///< [in] context of the head/snapdir object
 348      ObjectContextRef obc,  ///< [in] context of the object
 349      RecoveryHandle *h      ///< [in,out] handle to attach recovery op to
 350      ) = 0;
 351
 352    /**
 353     * true if PGBackend can handle this message while inactive
 354     *
 355     * If it returns true, handle_message *must* also return true
 356     */
 357    virtual bool can_handle_while_inactive(OpRequestRef op) = 0;
 358
 359    /// gives PGBackend a crack at an incoming message
 360    virtual bool handle_message(
 361      OpRequestRef op ///< [in] message received
 362      ) = 0; ///< @return true if the message was handled
 363
 364    virtual void check_recovery_sources(const OSDMapRef& osdmap) = 0;
 365
 366
 367    /**
 368     * clean up any temporary on-disk state due to a pg interval change
 369     */
 370    void on_change_cleanup(ObjectStore::Transaction *t);
 371    /**
 372     * implementation should clear itself, contexts blessed prior to on_change
 373     * won't be called after on_change()
 374     */
 375    virtual void on_change() = 0;
 376    virtual void clear_recovery_state() = 0;
 377
 378    virtual void on_flushed() = 0;
 379
 380    virtual IsPGRecoverablePredicate *get_is_recoverable_predicate() = 0;
 381    virtual IsPGReadablePredicate *get_is_readable_predicate() = 0;
 382
 383    virtual void dump_recovery_info(Formatter *f) const = 0;
 384
 385  private:
 386    set<hobject_t> temp_contents;
 387  public:
 388    // Track contents of temp collection, clear on reset
 389    void add_temp_obj(const hobject_t &oid) {
 390      temp_contents.insert(oid);
 391    }
 392    void add_temp_objs(const set<hobject_t> &oids) {
 393      temp_contents.insert(oids.begin(), oids.end());
 394    }
 395    void clear_temp_obj(const hobject_t &oid) {
 396      temp_contents.erase(oid);
 397    }
 398    void clear_temp_objs(const set<hobject_t> &oids) {
 399      for (set<hobject_t>::const_iterator i = oids.begin();
 400           i != oids.end();
 401           ++i) {
 402        temp_contents.erase(*i);
 403      }
 404    }
 405
 406    virtual ~PGBackend() {}
 407
 408    /// execute implementation specific transaction
 409    virtual void submit_transaction(
 410      const hobject_t &hoid,               ///< [in] object
 411      const object_stat_sum_t &delta_stats,///< [in] stat change
 412      const eversion_t &at_version,        ///< [in] version
 413      PGTransactionUPtr &&t,               ///< [in] trans to execute (move)
 414      const eversion_t &trim_to,           ///< [in] trim log to here
 415      const eversion_t &roll_forward_to,  ///< [in] trim rollback info to here
 416      const vector<pg_log_entry_t> &log_entries, ///< [in] log entries for t
 417      /// [in] hitset history (if updated with this transaction)
 418      boost::optional<pg_hit_set_history_t> &hset_history,
 419      Context *on_local_applied_sync,      ///< [in] called when applied locally
 420      Context *on_all_applied,             ///< [in] called when all acked
 421      Context *on_all_commit,              ///< [in] called when all commit
 422      ceph_tid_t tid,                      ///< [in] tid
 423      osd_reqid_t reqid,                   ///< [in] reqid
 424      OpRequestRef op                      ///< [in] op
 425      ) = 0;
 426
 427    /// submit callback to be called in order with pending writes
 428    virtual void call_write_ordered(std::function<void(void)> &&cb) = 0;
 429
 430    void try_stash(
 431      const hobject_t &hoid,
 432      version_t v,
 433      ObjectStore::Transaction *t);
 434
 435    void rollback(
 436      const pg_log_entry_t &entry,
 437      ObjectStore::Transaction *t);
 438
 439    friend class LRBTrimmer;
 440    void rollforward(
 441      const pg_log_entry_t &entry,
 442      ObjectStore::Transaction *t);
 443
 444    void trim(
 445      const pg_log_entry_t &entry,
 446      ObjectStore::Transaction *t);
 447
 448    void remove(
 449      const hobject_t &hoid,
 450      ObjectStore::Transaction *t);
 451
 452  protected:
 453    /// Reapply old attributes
 454    void rollback_setattrs(
 455      const hobject_t &hoid,
 456      map<string, boost::optional<bufferlist> > &old_attrs,
 457      ObjectStore::Transaction *t);
 458
 459    /// Truncate object to rollback append
 460    virtual void rollback_append(
 461      const hobject_t &hoid,
 462      uint64_t old_size,
 463      ObjectStore::Transaction *t);
 464
 465    /// Unstash object to rollback stash
 466    void rollback_stash(
 467      const hobject_t &hoid,
 468      version_t old_version,
 469      ObjectStore::Transaction *t);
 470
 471    /// Unstash object to rollback stash
 472    void rollback_try_stash(
 473      const hobject_t &hoid,
 474      version_t old_version,
 475      ObjectStore::Transaction *t);
 476
 477    /// Delete object to rollback create
 478    void rollback_create(
 479      const hobject_t &hoid,
 480      ObjectStore::Transaction *t) {
 481      remove(hoid, t);
 482    }
 483
 484    /// Clone the extents back into place
 485    void rollback_extents(
 486      version_t gen,
 487      const vector<pair<uint64_t, uint64_t> > &extents,
 488      const hobject_t &hoid,
 489      ObjectStore::Transaction *t);
 490  public:
 491
 492    /// Trim object stashed at version
 493    void trim_rollback_object(
 494      const hobject_t &hoid,
 495      version_t gen,
 496      ObjectStore::Transaction *t);
 497
 498    /// List objects in collection
 499    int objects_list_partial(
 500      const hobject_t &begin,
 501      int min,
 502      int max,
 503      vector<hobject_t> *ls,
 504      hobject_t *next);
 505
 506    int objects_list_range(
 507      const hobject_t &start,
 508      const hobject_t &end,
 509      snapid_t seq,
 510      vector<hobject_t> *ls,
 511      vector<ghobject_t> *gen_obs=0);
 512
 513    int objects_get_attr(
 514      const hobject_t &hoid,
 515      const string &attr,
 516      bufferlist *out);
 517
 518    virtual int objects_get_attrs(
 519      const hobject_t &hoid,
 520      map<string, bufferlist> *out);
 521
 522    virtual int objects_read_sync(
 523      const hobject_t &hoid,
 524      uint64_t off,
 525      uint64_t len,
 526      uint32_t op_flags,
 527      bufferlist *bl) = 0;
 528
 529    virtual void objects_read_async(
 530      const hobject_t &hoid,
 531      const list<pair<boost::tuple<uint64_t, uint64_t, uint32_t>,
 532                 pair<bufferlist*, Context*> > > &to_read,
 533      Context *on_complete, bool fast_read = false) = 0;
 534
 535    virtual bool scrub_supported() = 0;
 536    virtual bool auto_repair_supported() const = 0;
 537    void be_scan_list(
 538      ScrubMap &map, const vector<hobject_t> &ls, bool deep, uint32_t seed,
 539      ThreadPool::TPHandle &handle);
 540    bool be_compare_scrub_objects(
 541      pg_shard_t auth_shard,
 542      const ScrubMap::object &auth,
 543      const object_info_t& auth_oi,
 544      const ScrubMap::object &candidate,
 545      shard_info_wrapper& shard_error,
 546      inconsistent_obj_wrapper &result,
 547      ostream &errorstream);
 548    map<pg_shard_t, ScrubMap *>::const_iterator be_select_auth_object(
 549      const hobject_t &obj,
 550      const map<pg_shard_t,ScrubMap*> &maps,
 551      object_info_t *auth_oi,
 552      map<pg_shard_t, shard_info_wrapper> &shard_map,
 553      inconsistent_obj_wrapper &object_error);
 554    void be_compare_scrubmaps(
 555      const map<pg_shard_t,ScrubMap*> &maps,
 556      bool repair,
 557      map<hobject_t, set<pg_shard_t>> &missing,
 558      map<hobject_t, set<pg_shard_t>> &inconsistent,
 559      map<hobject_t, list<pg_shard_t>> &authoritative,
 560      map<hobject_t, pair<uint32_t,uint32_t>> &missing_digest,
 561      int &shallow_errors, int &deep_errors,
 562      Scrub::Store *store,
 563      const spg_t& pgid,
 564      const vector<int> &acting,
 565      ostream &errorstream);
 566    virtual uint64_t be_get_ondisk_size(
 567      uint64_t logical_size) = 0;
 568    virtual void be_deep_scrub(
 569      const hobject_t &poid,
 570      uint32_t seed,
 571      ScrubMap::object &o,
 572      ThreadPool::TPHandle &handle) = 0;
 573
 574    static PGBackend *build_pg_backend(
 575      const pg_pool_t &pool,
 576      const OSDMapRef curmap,
 577      Listener *l,
 578      coll_t coll,
 579      ObjectStore::CollectionHandle &ch,
 580      ObjectStore *store,
 581      CephContext *cct);
 582 };
 583
 584 #endif