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        bool is_delete,
  82        ObjectStore::Transaction *t
  83        ) = 0;
  84
  85      /**
  86       * Called when transaction recovering oid is durable and
  87       * applied on all replicas
  88       */
  89      virtual void on_global_recover(
  90        const hobject_t &oid,
  91        const object_stat_sum_t &stat_diff,
  92        bool is_delete
  93        ) = 0;
  94
  95      /**
  96       * Called when peer is recovered
  97       */
  98      virtual void on_peer_recover(
  99        pg_shard_t peer,
 100        const hobject_t &oid,
 101        const ObjectRecoveryInfo &recovery_info
 102        ) = 0;
 103
 104      virtual void begin_peer_recover(
 105        pg_shard_t peer,
 106        const hobject_t oid) = 0;
 107
 108      virtual void failed_push(const list<pg_shard_t> &from, const hobject_t &soid) = 0;
 109      virtual void primary_failed(const hobject_t &soid) = 0;
 110      virtual bool primary_error(const hobject_t& soid, eversion_t v) = 0;
 111      virtual void cancel_pull(const hobject_t &soid) = 0;
 112
 113      virtual void apply_stats(
 114        const hobject_t &soid,
 115        const object_stat_sum_t &delta_stats) = 0;
 116
 117      /**
 118       * Called when a read on the primary fails when pushing
 119       */
 120      virtual void on_primary_error(
 121        const hobject_t &oid,
 122        eversion_t v
 123        ) = 0;
 124
 125      virtual void remove_missing_object(const hobject_t &oid,
 126                                         eversion_t v,
 127                                         Context *on_complete) = 0;
 128
 129      /**
 130       * Bless a context
 131       *
 132       * Wraps a context in whatever outer layers the parent usually
 133       * uses to call into the PGBackend
 134       */
 135      virtual Context *bless_context(Context *c) = 0;
 136      virtual GenContext<ThreadPool::TPHandle&> *bless_gencontext(
 137        GenContext<ThreadPool::TPHandle&> *c) = 0;
 138
 139      virtual void send_message(int to_osd, Message *m) = 0;
 140      virtual void queue_transaction(
 141        ObjectStore::Transaction&& t,
 142        OpRequestRef op = OpRequestRef()
 143        ) = 0;
 144      virtual void queue_transactions(
 145        vector<ObjectStore::Transaction>& tls,
 146        OpRequestRef op = OpRequestRef()
 147        ) = 0;
 148      virtual epoch_t get_epoch() const = 0;
 149      virtual epoch_t get_interval_start_epoch() const = 0;
 150      virtual epoch_t get_last_peering_reset_epoch() const = 0;
 151
 152      virtual const set<pg_shard_t> &get_actingbackfill_shards() const = 0;
 153      virtual const set<pg_shard_t> &get_acting_shards() const = 0;
 154      virtual const set<pg_shard_t> &get_backfill_shards() const = 0;
 155
 156      virtual std::string gen_dbg_prefix() const = 0;
 157
 158      virtual const map<hobject_t, set<pg_shard_t>> &get_missing_loc_shards()
 159        const = 0;
 160
 161      virtual const pg_missing_tracker_t &get_local_missing() const = 0;
 162      virtual const map<pg_shard_t, pg_missing_t> &get_shard_missing()
 163        const = 0;
 164      virtual boost::optional<const pg_missing_const_i &> maybe_get_shard_missing(
 165        pg_shard_t peer) const {
 166        if (peer == primary_shard()) {
 167          return get_local_missing();
 168        } else {
 169          map<pg_shard_t, pg_missing_t>::const_iterator i =
 170            get_shard_missing().find(peer);
 171          if (i == get_shard_missing().end()) {
 172            return boost::optional<const pg_missing_const_i &>();
 173          } else {
 174            return i->second;
 175          }
 176        }
 177      }
 178      virtual const pg_missing_const_i &get_shard_missing(pg_shard_t peer) const {
 179        auto m = maybe_get_shard_missing(peer);
 180        assert(m);
 181        return *m;
 182      }
 183
 184      virtual const map<pg_shard_t, pg_info_t> &get_shard_info() const = 0;
 185      virtual const pg_info_t &get_shard_info(pg_shard_t peer) const {
 186        if (peer == primary_shard()) {
 187          return get_info();
 188        } else {
 189          map<pg_shard_t, pg_info_t>::const_iterator i =
 190            get_shard_info().find(peer);
 191          assert(i != get_shard_info().end());
 192          return i->second;
 193        }
 194      }
 195
 196      virtual const PGLog &get_log() const = 0;
 197      virtual bool pgb_is_primary() const = 0;
 198      virtual OSDMapRef pgb_get_osdmap() const = 0;
 199      virtual const pg_info_t &get_info() const = 0;
 200      virtual const pg_pool_t &get_pool() const = 0;
 201
 202      virtual ObjectContextRef get_obc(
 203        const hobject_t &hoid,
 204        const map<string, bufferlist> &attrs) = 0;
 205
 206      virtual bool try_lock_for_read(
 207        const hobject_t &hoid,
 208        ObcLockManager &manager) = 0;
 209
 210      virtual void release_locks(ObcLockManager &manager) = 0;
 211
 212      virtual void op_applied(
 213        const eversion_t &applied_version) = 0;
 214
 215      virtual bool should_send_op(
 216        pg_shard_t peer,
 217        const hobject_t &hoid) = 0;
 218
 219      virtual void log_operation(
 220        const vector<pg_log_entry_t> &logv,
 221        const boost::optional<pg_hit_set_history_t> &hset_history,
 222        const eversion_t &trim_to,
 223        const eversion_t &roll_forward_to,
 224        bool transaction_applied,
 225        ObjectStore::Transaction &t) = 0;
 226
 227      virtual void pgb_set_object_snap_mapping(
 228        const hobject_t &soid,
 229        const set<snapid_t> &snaps,
 230        ObjectStore::Transaction *t) = 0;
 231
 232      virtual void pgb_clear_object_snap_mapping(
 233        const hobject_t &soid,
 234        ObjectStore::Transaction *t) = 0;
 235
 236      virtual void update_peer_last_complete_ondisk(
 237        pg_shard_t fromosd,
 238        eversion_t lcod) = 0;
 239
 240      virtual void update_last_complete_ondisk(
 241        eversion_t lcod) = 0;
 242
 243      virtual void update_stats(
 244        const pg_stat_t &stat) = 0;
 245
 246      virtual void schedule_recovery_work(
 247        GenContext<ThreadPool::TPHandle&> *c) = 0;
 248
 249      virtual pg_shard_t whoami_shard() const = 0;
 250      int whoami() const {
 251        return whoami_shard().osd;
 252      }
 253      spg_t whoami_spg_t() const {
 254        return get_info().pgid;
 255      }
 256
 257      virtual spg_t primary_spg_t() const = 0;
 258      virtual pg_shard_t primary_shard() const = 0;
 259
 260      virtual uint64_t min_peer_features() const = 0;
 261
 262      virtual hobject_t get_temp_recovery_object(const hobject_t& target,
 263                                                 eversion_t version) = 0;
 264
 265      virtual void send_message_osd_cluster(
 266        int peer, Message *m, epoch_t from_epoch) = 0;
 267      virtual void send_message_osd_cluster(
 268        Message *m, Connection *con) = 0;
 269      virtual void send_message_osd_cluster(
 270        Message *m, const ConnectionRef& con) = 0;
 271      virtual ConnectionRef get_con_osd_cluster(int peer, epoch_t from_epoch) = 0;
 272      virtual entity_name_t get_cluster_msgr_name() = 0;
 273
 274      virtual PerfCounters *get_logger() = 0;
 275
 276      virtual ceph_tid_t get_tid() = 0;
 277
 278      virtual LogClientTemp clog_error() = 0;
 279      virtual LogClientTemp clog_warn() = 0;
 280
 281      virtual bool check_failsafe_full(ostream &ss) = 0;
 282
 283      virtual bool check_osdmap_full(const set<pg_shard_t> &missing_on) = 0;
 284
 285      virtual ~Listener() {}
 286    };
 287    Listener *parent;
 288    Listener *get_parent() const { return parent; }
 289    PGBackend(CephContext* cct, Listener *l, ObjectStore *store, coll_t coll,
 290              ObjectStore::CollectionHandle &ch) :
 291      cct(cct),
 292      store(store),
 293      coll(coll),
 294      ch(ch),
 295      parent(l) {}
 296    bool is_primary() const { return get_parent()->pgb_is_primary(); }
 297    OSDMapRef get_osdmap() const { return get_parent()->pgb_get_osdmap(); }
 298    const pg_info_t &get_info() { return get_parent()->get_info(); }
 299
 300    std::string gen_prefix() const {
 301      return parent->gen_dbg_prefix();
 302    }
 303
 304    /**
 305     * RecoveryHandle
 306     *
 307     * We may want to recover multiple objects in the same set of
 308     * messages.  RecoveryHandle is an interface for the opaque
 309     * object used by the implementation to store the details of
 310     * the pending recovery operations.
 311     */
 312    struct RecoveryHandle {
 313      bool cache_dont_need;
 314      map<pg_shard_t, vector<pair<hobject_t, eversion_t> > > deletes;
 315
 316      RecoveryHandle(): cache_dont_need(false) {}
 317      virtual ~RecoveryHandle() {}
 318    };
 319
 320    /// Get a fresh recovery operation
 321    virtual RecoveryHandle *open_recovery_op() = 0;
 322
 323    /// run_recovery_op: finish the operation represented by h
 324    virtual void run_recovery_op(
 325      RecoveryHandle *h,     ///< [in] op to finish
 326      int priority           ///< [in] msg priority
 327      ) = 0;
 328
 329    void recover_delete_object(const hobject_t &oid, eversion_t v,
 330                               RecoveryHandle *h);
 331    void send_recovery_deletes(int prio,
 332                               const map<pg_shard_t, vector<pair<hobject_t, eversion_t> > > &deletes);
 333
 334    /**
 335     * recover_object
 336     *
 337     * Triggers a recovery operation on the specified hobject_t
 338     * onreadable must be called before onwriteable
 339     *
 340     * On each replica (primary included), get_parent()->on_not_missing()
 341     * must be called when the transaction finalizing the recovery
 342     * is queued.  Similarly, get_parent()->on_readable() must be called
 343     * when the transaction is applied in the backing store.
 344     *
 345     * get_parent()->on_not_degraded() should be called on the primary
 346     * when writes can resume on the object.
 347     *
 348     * obc may be NULL if the primary lacks the object.
 349     *
 350     * head may be NULL only if the head/snapdir is missing
 351     *
 352     * @param missing [in] set of info, missing pairs for queried nodes
 353     * @param overlaps [in] mapping of object to file offset overlaps
 354     */
 355    virtual int recover_object(
 356      const hobject_t &hoid, ///< [in] object to recover
 357      eversion_t v,          ///< [in] version to recover
 358      ObjectContextRef head,  ///< [in] context of the head/snapdir object
 359      ObjectContextRef obc,  ///< [in] context of the object
 360      RecoveryHandle *h      ///< [in,out] handle to attach recovery op to
 361      ) = 0;
 362
 363    /**
 364     * true if PGBackend can handle this message while inactive
 365     *
 366     * If it returns true, handle_message *must* also return true
 367     */
 368    virtual bool can_handle_while_inactive(OpRequestRef op) = 0;
 369
 370    /// gives PGBackend a crack at an incoming message
 371    bool handle_message(
 372      OpRequestRef op ///< [in] message received
 373      ); ///< @return true if the message was handled
 374
 375    /// the variant of handle_message that is overridden by child classes
 376    virtual bool _handle_message(OpRequestRef op) = 0;
 377
 378    virtual void check_recovery_sources(const OSDMapRef& osdmap) = 0;
 379
 380
 381    /**
 382     * clean up any temporary on-disk state due to a pg interval change
 383     */
 384    void on_change_cleanup(ObjectStore::Transaction *t);
 385    /**
 386     * implementation should clear itself, contexts blessed prior to on_change
 387     * won't be called after on_change()
 388     */
 389    virtual void on_change() = 0;
 390    virtual void clear_recovery_state() = 0;
 391
 392    virtual void on_flushed() = 0;
 393
 394    virtual IsPGRecoverablePredicate *get_is_recoverable_predicate() = 0;
 395    virtual IsPGReadablePredicate *get_is_readable_predicate() = 0;
 396
 397    virtual void dump_recovery_info(Formatter *f) const = 0;
 398
 399  private:
 400    set<hobject_t> temp_contents;
 401  public:
 402    // Track contents of temp collection, clear on reset
 403    void add_temp_obj(const hobject_t &oid) {
 404      temp_contents.insert(oid);
 405    }
 406    void add_temp_objs(const set<hobject_t> &oids) {
 407      temp_contents.insert(oids.begin(), oids.end());
 408    }
 409    void clear_temp_obj(const hobject_t &oid) {
 410      temp_contents.erase(oid);
 411    }
 412    void clear_temp_objs(const set<hobject_t> &oids) {
 413      for (set<hobject_t>::const_iterator i = oids.begin();
 414           i != oids.end();
 415           ++i) {
 416        temp_contents.erase(*i);
 417      }
 418    }
 419
 420    virtual ~PGBackend() {}
 421
 422    /// execute implementation specific transaction
 423    virtual void submit_transaction(
 424      const hobject_t &hoid,               ///< [in] object
 425      const object_stat_sum_t &delta_stats,///< [in] stat change
 426      const eversion_t &at_version,        ///< [in] version
 427      PGTransactionUPtr &&t,               ///< [in] trans to execute (move)
 428      const eversion_t &trim_to,           ///< [in] trim log to here
 429      const eversion_t &roll_forward_to,  ///< [in] trim rollback info to here
 430      const vector<pg_log_entry_t> &log_entries, ///< [in] log entries for t
 431      /// [in] hitset history (if updated with this transaction)
 432      boost::optional<pg_hit_set_history_t> &hset_history,
 433      Context *on_local_applied_sync,      ///< [in] called when applied locally
 434      Context *on_all_applied,             ///< [in] called when all acked
 435      Context *on_all_commit,              ///< [in] called when all commit
 436      ceph_tid_t tid,                      ///< [in] tid
 437      osd_reqid_t reqid,                   ///< [in] reqid
 438      OpRequestRef op                      ///< [in] op
 439      ) = 0;
 440
 441    /// submit callback to be called in order with pending writes
 442    virtual void call_write_ordered(std::function<void(void)> &&cb) = 0;
 443
 444    void try_stash(
 445      const hobject_t &hoid,
 446      version_t v,
 447      ObjectStore::Transaction *t);
 448
 449    void rollback(
 450      const pg_log_entry_t &entry,
 451      ObjectStore::Transaction *t);
 452
 453    friend class LRBTrimmer;
 454    void rollforward(
 455      const pg_log_entry_t &entry,
 456      ObjectStore::Transaction *t);
 457
 458    void trim(
 459      const pg_log_entry_t &entry,
 460      ObjectStore::Transaction *t);
 461
 462    void remove(
 463      const hobject_t &hoid,
 464      ObjectStore::Transaction *t);
 465
 466  protected:
 467
 468    void handle_recovery_delete(OpRequestRef op);
 469    void handle_recovery_delete_reply(OpRequestRef op);
 470
 471    /// Reapply old attributes
 472    void rollback_setattrs(
 473      const hobject_t &hoid,
 474      map<string, boost::optional<bufferlist> > &old_attrs,
 475      ObjectStore::Transaction *t);
 476
 477    /// Truncate object to rollback append
 478    virtual void rollback_append(
 479      const hobject_t &hoid,
 480      uint64_t old_size,
 481      ObjectStore::Transaction *t);
 482
 483    /// Unstash object to rollback stash
 484    void rollback_stash(
 485      const hobject_t &hoid,
 486      version_t old_version,
 487      ObjectStore::Transaction *t);
 488
 489    /// Unstash object to rollback stash
 490    void rollback_try_stash(
 491      const hobject_t &hoid,
 492      version_t old_version,
 493      ObjectStore::Transaction *t);
 494
 495    /// Delete object to rollback create
 496    void rollback_create(
 497      const hobject_t &hoid,
 498      ObjectStore::Transaction *t) {
 499      remove(hoid, t);
 500    }
 501
 502    /// Clone the extents back into place
 503    void rollback_extents(
 504      version_t gen,
 505      const vector<pair<uint64_t, uint64_t> > &extents,
 506      const hobject_t &hoid,
 507      ObjectStore::Transaction *t);
 508  public:
 509
 510    /// Trim object stashed at version
 511    void trim_rollback_object(
 512      const hobject_t &hoid,
 513      version_t gen,
 514      ObjectStore::Transaction *t);
 515
 516    /// List objects in collection
 517    int objects_list_partial(
 518      const hobject_t &begin,
 519      int min,
 520      int max,
 521      vector<hobject_t> *ls,
 522      hobject_t *next);
 523
 524    int objects_list_range(
 525      const hobject_t &start,
 526      const hobject_t &end,
 527      snapid_t seq,
 528      vector<hobject_t> *ls,
 529      vector<ghobject_t> *gen_obs=0);
 530
 531    int objects_get_attr(
 532      const hobject_t &hoid,
 533      const string &attr,
 534      bufferlist *out);
 535
 536    virtual int objects_get_attrs(
 537      const hobject_t &hoid,
 538      map<string, bufferlist> *out);
 539
 540    virtual int objects_read_sync(
 541      const hobject_t &hoid,
 542      uint64_t off,
 543      uint64_t len,
 544      uint32_t op_flags,
 545      bufferlist *bl) = 0;
 546
 547    virtual void objects_read_async(
 548      const hobject_t &hoid,
 549      const list<pair<boost::tuple<uint64_t, uint64_t, uint32_t>,
 550                 pair<bufferlist*, Context*> > > &to_read,
 551      Context *on_complete, bool fast_read = false) = 0;
 552
 553    virtual bool scrub_supported() = 0;
 554    virtual bool auto_repair_supported() const = 0;
 555    void be_scan_list(
 556      ScrubMap &map, const vector<hobject_t> &ls, bool deep, uint32_t seed,
 557      ThreadPool::TPHandle &handle);
 558    bool be_compare_scrub_objects(
 559      pg_shard_t auth_shard,
 560      const ScrubMap::object &auth,
 561      const object_info_t& auth_oi,
 562      const ScrubMap::object &candidate,
 563      shard_info_wrapper& shard_error,
 564      inconsistent_obj_wrapper &result,
 565      ostream &errorstream);
 566    map<pg_shard_t, ScrubMap *>::const_iterator be_select_auth_object(
 567      const hobject_t &obj,
 568      const map<pg_shard_t,ScrubMap*> &maps,
 569      object_info_t *auth_oi,
 570      map<pg_shard_t, shard_info_wrapper> &shard_map,
 571      inconsistent_obj_wrapper &object_error);
 572    void be_compare_scrubmaps(
 573      const map<pg_shard_t,ScrubMap*> &maps,
 574      bool repair,
 575      map<hobject_t, set<pg_shard_t>> &missing,
 576      map<hobject_t, set<pg_shard_t>> &inconsistent,
 577      map<hobject_t, list<pg_shard_t>> &authoritative,
 578      map<hobject_t, pair<uint32_t,uint32_t>> &missing_digest,
 579      int &shallow_errors, int &deep_errors,
 580      Scrub::Store *store,
 581      const spg_t& pgid,
 582      const vector<int> &acting,
 583      ostream &errorstream);
 584    virtual uint64_t be_get_ondisk_size(
 585      uint64_t logical_size) = 0;
 586    virtual void be_deep_scrub(
 587      const hobject_t &poid,
 588      uint32_t seed,
 589      ScrubMap::object &o,
 590      ThreadPool::TPHandle &handle) = 0;
 591
 592    static PGBackend *build_pg_backend(
 593      const pg_pool_t &pool,
 594      const OSDMapRef curmap,
 595      Listener *l,
 596      coll_t coll,
 597      ObjectStore::CollectionHandle &ch,
 598      ObjectStore *store,
 599      CephContext *cct);
 600 };
 601
 602 #endif