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