ceph/src/osd/OSD.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 #ifndef CEPH_OSD_H
  16 #define CEPH_OSD_H
  17
  18 #include "PG.h"
  19
  20 #include "msg/Dispatcher.h"
  21
  22 #include "common/Mutex.h"
  23 #include "common/RWLock.h"
  24 #include "common/Timer.h"
  25 #include "common/WorkQueue.h"
  26 #include "common/AsyncReserver.h"
  27 #include "common/ceph_context.h"
  28 #include "common/zipkin_trace.h"
  29
  30 #include "mgr/MgrClient.h"
  31
  32 #include "os/ObjectStore.h"
  33 #include "OSDCap.h"
  34
  35 #include "auth/KeyRing.h"
  36 #include "osd/ClassHandler.h"
  37
  38 #include "include/CompatSet.h"
  39
  40 #include "OpRequest.h"
  41 #include "Session.h"
  42
  43 #include "osd/PGQueueable.h"
  44
  45 #include <atomic>
  46 #include <map>
  47 #include <memory>
  48 #include "include/memory.h"
  49 using namespace std;
  50
  51 #include "include/unordered_map.h"
  52
  53 #include "common/shared_cache.hpp"
  54 #include "common/simple_cache.hpp"
  55 #include "common/sharedptr_registry.hpp"
  56 #include "common/WeightedPriorityQueue.h"
  57 #include "common/PrioritizedQueue.h"
  58 #include "osd/mClockOpClassQueue.h"
  59 #include "osd/mClockClientQueue.h"
  60 #include "messages/MOSDOp.h"
  61 #include "include/Spinlock.h"
  62 #include "common/EventTrace.h"
  63
  64 #define CEPH_OSD_PROTOCOL    10 /* cluster internal */
  65
  66
  67 enum {
  68   l_osd_first = 10000,
  69   l_osd_op_wip,
  70   l_osd_op,
  71   l_osd_op_inb,
  72   l_osd_op_outb,
  73   l_osd_op_lat,
  74   l_osd_op_process_lat,
  75   l_osd_op_prepare_lat,
  76   l_osd_op_r,
  77   l_osd_op_r_outb,
  78   l_osd_op_r_lat,
  79   l_osd_op_r_lat_outb_hist,
  80   l_osd_op_r_process_lat,
  81   l_osd_op_r_prepare_lat,
  82   l_osd_op_w,
  83   l_osd_op_w_inb,
  84   l_osd_op_w_lat,
  85   l_osd_op_w_lat_inb_hist,
  86   l_osd_op_w_process_lat,
  87   l_osd_op_w_prepare_lat,
  88   l_osd_op_rw,
  89   l_osd_op_rw_inb,
  90   l_osd_op_rw_outb,
  91   l_osd_op_rw_lat,
  92   l_osd_op_rw_lat_inb_hist,
  93   l_osd_op_rw_lat_outb_hist,
  94   l_osd_op_rw_process_lat,
  95   l_osd_op_rw_prepare_lat,
  96
  97   l_osd_op_before_queue_op_lat,
  98   l_osd_op_before_dequeue_op_lat,
  99
 100   l_osd_sop,
 101   l_osd_sop_inb,
 102   l_osd_sop_lat,
 103   l_osd_sop_w,
 104   l_osd_sop_w_inb,
 105   l_osd_sop_w_lat,
 106   l_osd_sop_pull,
 107   l_osd_sop_pull_lat,
 108   l_osd_sop_push,
 109   l_osd_sop_push_inb,
 110   l_osd_sop_push_lat,
 111
 112   l_osd_pull,
 113   l_osd_push,
 114   l_osd_push_outb,
 115
 116   l_osd_rop,
 117
 118   l_osd_loadavg,
 119   l_osd_buf,
 120   l_osd_history_alloc_bytes,
 121   l_osd_history_alloc_num,
 122   l_osd_cached_crc,
 123   l_osd_cached_crc_adjusted,
 124   l_osd_missed_crc,
 125
 126   l_osd_pg,
 127   l_osd_pg_primary,
 128   l_osd_pg_replica,
 129   l_osd_pg_stray,
 130   l_osd_hb_to,
 131   l_osd_map,
 132   l_osd_mape,
 133   l_osd_mape_dup,
 134
 135   l_osd_waiting_for_map,
 136
 137   l_osd_map_cache_hit,
 138   l_osd_map_cache_miss,
 139   l_osd_map_cache_miss_low,
 140   l_osd_map_cache_miss_low_avg,
 141   l_osd_map_bl_cache_hit,
 142   l_osd_map_bl_cache_miss,
 143
 144   l_osd_stat_bytes,
 145   l_osd_stat_bytes_used,
 146   l_osd_stat_bytes_avail,
 147
 148   l_osd_copyfrom,
 149
 150   l_osd_tier_promote,
 151   l_osd_tier_flush,
 152   l_osd_tier_flush_fail,
 153   l_osd_tier_try_flush,
 154   l_osd_tier_try_flush_fail,
 155   l_osd_tier_evict,
 156   l_osd_tier_whiteout,
 157   l_osd_tier_dirty,
 158   l_osd_tier_clean,
 159   l_osd_tier_delay,
 160   l_osd_tier_proxy_read,
 161   l_osd_tier_proxy_write,
 162
 163   l_osd_agent_wake,
 164   l_osd_agent_skip,
 165   l_osd_agent_flush,
 166   l_osd_agent_evict,
 167
 168   l_osd_object_ctx_cache_hit,
 169   l_osd_object_ctx_cache_total,
 170
 171   l_osd_op_cache_hit,
 172   l_osd_tier_flush_lat,
 173   l_osd_tier_promote_lat,
 174   l_osd_tier_r_lat,
 175
 176   l_osd_pg_info,
 177   l_osd_pg_fastinfo,
 178   l_osd_pg_biginfo,
 179
 180   l_osd_last,
 181 };
 182
 183 // RecoveryState perf counters
 184 enum {
 185   rs_first = 20000,
 186   rs_initial_latency,
 187   rs_started_latency,
 188   rs_reset_latency,
 189   rs_start_latency,
 190   rs_primary_latency,
 191   rs_peering_latency,
 192   rs_backfilling_latency,
 193   rs_waitremotebackfillreserved_latency,
 194   rs_waitlocalbackfillreserved_latency,
 195   rs_notbackfilling_latency,
 196   rs_repnotrecovering_latency,
 197   rs_repwaitrecoveryreserved_latency,
 198   rs_repwaitbackfillreserved_latency,
 199   rs_reprecovering_latency,
 200   rs_activating_latency,
 201   rs_waitlocalrecoveryreserved_latency,
 202   rs_waitremoterecoveryreserved_latency,
 203   rs_recovering_latency,
 204   rs_recovered_latency,
 205   rs_clean_latency,
 206   rs_active_latency,
 207   rs_replicaactive_latency,
 208   rs_stray_latency,
 209   rs_getinfo_latency,
 210   rs_getlog_latency,
 211   rs_waitactingchange_latency,
 212   rs_incomplete_latency,
 213   rs_down_latency,
 214   rs_getmissing_latency,
 215   rs_waitupthru_latency,
 216   rs_notrecovering_latency,
 217   rs_last,
 218 };
 219
 220 class Messenger;
 221 class Message;
 222 class MonClient;
 223 class PerfCounters;
 224 class ObjectStore;
 225 class FuseStore;
 226 class OSDMap;
 227 class MLog;
 228 class Objecter;
 229
 230 class Watch;
 231 class PrimaryLogPG;
 232
 233 class AuthAuthorizeHandlerRegistry;
 234
 235 class TestOpsSocketHook;
 236 struct C_CompleteSplits;
 237 struct C_OpenPGs;
 238 class LogChannel;
 239 class CephContext;
 240 typedef ceph::shared_ptr<ObjectStore::Sequencer> SequencerRef;
 241 class MOSDOp;
 242
 243 class DeletingState {
 244   Mutex lock;
 245   Cond cond;
 246   enum {
 247     QUEUED,
 248     CLEARING_DIR,
 249     CLEARING_WAITING,
 250     DELETING_DIR,
 251     DELETED_DIR,
 252     CANCELED,
 253   } status;
 254   bool stop_deleting;
 255 public:
 256   const spg_t pgid;
 257   const PGRef old_pg_state;
 258   explicit DeletingState(const pair<spg_t, PGRef> &in) :
 259     lock("DeletingState::lock"), status(QUEUED), stop_deleting(false),
 260     pgid(in.first), old_pg_state(in.second) {
 261     }
 262
 263   /// transition status to CLEARING_WAITING
 264   bool pause_clearing() {
 265     Mutex::Locker l(lock);
 266     assert(status == CLEARING_DIR);
 267     if (stop_deleting) {
 268       status = CANCELED;
 269       cond.Signal();
 270       return false;
 271     }
 272     status = CLEARING_WAITING;
 273     return true;
 274   } ///< @return false if we should cancel deletion
 275
 276   /// start or resume the clearing - transition the status to CLEARING_DIR
 277   bool start_or_resume_clearing() {
 278     Mutex::Locker l(lock);
 279     assert(
 280       status == QUEUED ||
 281       status == DELETED_DIR ||
 282       status == CLEARING_WAITING);
 283     if (stop_deleting) {
 284       status = CANCELED;
 285       cond.Signal();
 286       return false;
 287     }
 288     status = CLEARING_DIR;
 289     return true;
 290   } ///< @return false if we should cancel the deletion
 291
 292   /// transition status to CLEARING_DIR
 293   bool resume_clearing() {
 294     Mutex::Locker l(lock);
 295     assert(status == CLEARING_WAITING);
 296     if (stop_deleting) {
 297       status = CANCELED;
 298       cond.Signal();
 299       return false;
 300     }
 301     status = CLEARING_DIR;
 302     return true;
 303   } ///< @return false if we should cancel deletion
 304
 305   /// transition status to deleting
 306   bool start_deleting() {
 307     Mutex::Locker l(lock);
 308     assert(status == CLEARING_DIR);
 309     if (stop_deleting) {
 310       status = CANCELED;
 311       cond.Signal();
 312       return false;
 313     }
 314     status = DELETING_DIR;
 315     return true;
 316   } ///< @return false if we should cancel deletion
 317
 318   /// signal collection removal queued
 319   void finish_deleting() {
 320     Mutex::Locker l(lock);
 321     assert(status == DELETING_DIR);
 322     status = DELETED_DIR;
 323     cond.Signal();
 324   }
 325
 326   /// try to halt the deletion
 327   bool try_stop_deletion() {
 328     Mutex::Locker l(lock);
 329     stop_deleting = true;
 330     /**
 331      * If we are in DELETING_DIR or CLEARING_DIR, there are in progress
 332      * operations we have to wait for before continuing on.  States
 333      * CLEARING_WAITING and QUEUED indicate that the remover will check
 334      * stop_deleting before queueing any further operations.  CANCELED
 335      * indicates that the remover has already halted.  DELETED_DIR
 336      * indicates that the deletion has been fully queued.
 337      */
 338     while (status == DELETING_DIR || status == CLEARING_DIR)
 339       cond.Wait(lock);
 340     return status != DELETED_DIR;
 341   } ///< @return true if we don't need to recreate the collection
 342 };
 343 typedef ceph::shared_ptr<DeletingState> DeletingStateRef;
 344
 345 class OSD;
 346
 347 class OSDService {
 348 public:
 349   OSD *osd;
 350   CephContext *cct;
 351   SharedPtrRegistry<spg_t, ObjectStore::Sequencer> osr_registry;
 352   ceph::shared_ptr<ObjectStore::Sequencer> meta_osr;
 353   SharedPtrRegistry<spg_t, DeletingState> deleting_pgs;
 354   const int whoami;
 355   ObjectStore *&store;
 356   LogClient &log_client;
 357   LogChannelRef clog;
 358   PGRecoveryStats &pg_recovery_stats;
 359 private:
 360   Messenger *&cluster_messenger;
 361   Messenger *&client_messenger;
 362 public:
 363   PerfCounters *&logger;
 364   PerfCounters *&recoverystate_perf;
 365   MonClient   *&monc;
 366   ThreadPool::BatchWorkQueue<PG> &peering_wq;
 367   GenContextWQ recovery_gen_wq;
 368   ClassHandler  *&class_handler;
 369
 370   void enqueue_back(spg_t pgid, PGQueueable qi);
 371   void enqueue_front(spg_t pgid, PGQueueable qi);
 372
 373   void maybe_inject_dispatch_delay() {
 374     if (g_conf->osd_debug_inject_dispatch_delay_probability > 0) {
 375       if (rand() % 10000 <
 376           g_conf->osd_debug_inject_dispatch_delay_probability * 10000) {
 377         utime_t t;
 378         t.set_from_double(g_conf->osd_debug_inject_dispatch_delay_duration);
 379         t.sleep();
 380       }
 381     }
 382   }
 383
 384 private:
 385   // -- map epoch lower bound --
 386   Mutex pg_epoch_lock;
 387   multiset<epoch_t> pg_epochs;
 388   map<spg_t,epoch_t> pg_epoch;
 389
 390 public:
 391   void pg_add_epoch(spg_t pgid, epoch_t epoch) {
 392     Mutex::Locker l(pg_epoch_lock);
 393     map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
 394     assert(t == pg_epoch.end());
 395     pg_epoch[pgid] = epoch;
 396     pg_epochs.insert(epoch);
 397   }
 398   void pg_update_epoch(spg_t pgid, epoch_t epoch) {
 399     Mutex::Locker l(pg_epoch_lock);
 400     map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
 401     assert(t != pg_epoch.end());
 402     pg_epochs.erase(pg_epochs.find(t->second));
 403     t->second = epoch;
 404     pg_epochs.insert(epoch);
 405   }
 406   void pg_remove_epoch(spg_t pgid) {
 407     Mutex::Locker l(pg_epoch_lock);
 408     map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
 409     if (t != pg_epoch.end()) {
 410       pg_epochs.erase(pg_epochs.find(t->second));
 411       pg_epoch.erase(t);
 412     }
 413   }
 414   epoch_t get_min_pg_epoch() {
 415     Mutex::Locker l(pg_epoch_lock);
 416     if (pg_epochs.empty())
 417       return 0;
 418     else
 419       return *pg_epochs.begin();
 420   }
 421
 422 private:
 423   // -- superblock --
 424   Mutex publish_lock, pre_publish_lock; // pre-publish orders before publish
 425   OSDSuperblock superblock;
 426
 427 public:
 428   OSDSuperblock get_superblock() {
 429     Mutex::Locker l(publish_lock);
 430     return superblock;
 431   }
 432   void publish_superblock(const OSDSuperblock &block) {
 433     Mutex::Locker l(publish_lock);
 434     superblock = block;
 435   }
 436
 437   int get_nodeid() const { return whoami; }
 438
 439   std::atomic<epoch_t> max_oldest_map;
 440 private:
 441   OSDMapRef osdmap;
 442
 443 public:
 444   OSDMapRef get_osdmap() {
 445     Mutex::Locker l(publish_lock);
 446     return osdmap;
 447   }
 448   epoch_t get_osdmap_epoch() {
 449     Mutex::Locker l(publish_lock);
 450     return osdmap ? osdmap->get_epoch() : 0;
 451   }
 452   void publish_map(OSDMapRef map) {
 453     Mutex::Locker l(publish_lock);
 454     osdmap = map;
 455   }
 456
 457   /*
 458    * osdmap - current published map
 459    * next_osdmap - pre_published map that is about to be published.
 460    *
 461    * We use the next_osdmap to send messages and initiate connections,
 462    * but only if the target is the same instance as the one in the map
 463    * epoch the current user is working from (i.e., the result is
 464    * equivalent to what is in next_osdmap).
 465    *
 466    * This allows the helpers to start ignoring osds that are about to
 467    * go down, and let OSD::handle_osd_map()/note_down_osd() mark them
 468    * down, without worrying about reopening connections from threads
 469    * working from old maps.
 470    */
 471 private:
 472   OSDMapRef next_osdmap;
 473   Cond pre_publish_cond;
 474
 475 public:
 476   void pre_publish_map(OSDMapRef map) {
 477     Mutex::Locker l(pre_publish_lock);
 478     next_osdmap = std::move(map);
 479   }
 480
 481   void activate_map();
 482   /// map epochs reserved below
 483   map<epoch_t, unsigned> map_reservations;
 484
 485   /// gets ref to next_osdmap and registers the epoch as reserved
 486   OSDMapRef get_nextmap_reserved() {
 487     Mutex::Locker l(pre_publish_lock);
 488     if (!next_osdmap)
 489       return OSDMapRef();
 490     epoch_t e = next_osdmap->get_epoch();
 491     map<epoch_t, unsigned>::iterator i =
 492       map_reservations.insert(make_pair(e, 0)).first;
 493     i->second++;
 494     return next_osdmap;
 495   }
 496   /// releases reservation on map
 497   void release_map(OSDMapRef osdmap) {
 498     Mutex::Locker l(pre_publish_lock);
 499     map<epoch_t, unsigned>::iterator i =
 500       map_reservations.find(osdmap->get_epoch());
 501     assert(i != map_reservations.end());
 502     assert(i->second > 0);
 503     if (--(i->second) == 0) {
 504       map_reservations.erase(i);
 505     }
 506     pre_publish_cond.Signal();
 507   }
 508   /// blocks until there are no reserved maps prior to next_osdmap
 509   void await_reserved_maps() {
 510     Mutex::Locker l(pre_publish_lock);
 511     assert(next_osdmap);
 512     while (true) {
 513       map<epoch_t, unsigned>::const_iterator i = map_reservations.cbegin();
 514       if (i == map_reservations.cend() || i->first >= next_osdmap->get_epoch()) {
 515         break;
 516       } else {
 517         pre_publish_cond.Wait(pre_publish_lock);
 518       }
 519     }
 520   }
 521
 522 private:
 523   Mutex peer_map_epoch_lock;
 524   map<int, epoch_t> peer_map_epoch;
 525 public:
 526   epoch_t get_peer_epoch(int p);
 527   epoch_t note_peer_epoch(int p, epoch_t e);
 528   void forget_peer_epoch(int p, epoch_t e);
 529
 530   void send_map(class MOSDMap *m, Connection *con);
 531   void send_incremental_map(epoch_t since, Connection *con, OSDMapRef& osdmap);
 532   MOSDMap *build_incremental_map_msg(epoch_t from, epoch_t to,
 533                                        OSDSuperblock& superblock);
 534   bool should_share_map(entity_name_t name, Connection *con, epoch_t epoch,
 535                         const OSDMapRef& osdmap, const epoch_t *sent_epoch_p);
 536   void share_map(entity_name_t name, Connection *con, epoch_t epoch,
 537                  OSDMapRef& osdmap, epoch_t *sent_epoch_p);
 538   void share_map_peer(int peer, Connection *con,
 539                       OSDMapRef map = OSDMapRef());
 540
 541   ConnectionRef get_con_osd_cluster(int peer, epoch_t from_epoch);
 542   pair<ConnectionRef,ConnectionRef> get_con_osd_hb(int peer, epoch_t from_epoch);  // (back, front)
 543   void send_message_osd_cluster(int peer, Message *m, epoch_t from_epoch);
 544   void send_message_osd_cluster(Message *m, Connection *con) {
 545     con->send_message(m);
 546   }
 547   void send_message_osd_cluster(Message *m, const ConnectionRef& con) {
 548     con->send_message(m);
 549   }
 550   void send_message_osd_client(Message *m, Connection *con) {
 551     con->send_message(m);
 552   }
 553   void send_message_osd_client(Message *m, const ConnectionRef& con) {
 554     con->send_message(m);
 555   }
 556   entity_name_t get_cluster_msgr_name() {
 557     return cluster_messenger->get_myname();
 558   }
 559
 560 private:
 561   // -- scrub scheduling --
 562   Mutex sched_scrub_lock;
 563   int scrubs_pending;
 564   int scrubs_active;
 565
 566 public:
 567   struct ScrubJob {
 568     CephContext* cct;
 569     /// pg to be scrubbed
 570     spg_t pgid;
 571     /// a time scheduled for scrub. but the scrub could be delayed if system
 572     /// load is too high or it fails to fall in the scrub hours
 573     utime_t sched_time;
 574     /// the hard upper bound of scrub time
 575     utime_t deadline;
 576     ScrubJob() : cct(nullptr) {}
 577     explicit ScrubJob(CephContext* cct, const spg_t& pg,
 578                       const utime_t& timestamp,
 579                       double pool_scrub_min_interval = 0,
 580                       double pool_scrub_max_interval = 0, bool must = true);
 581     /// order the jobs by sched_time
 582     bool operator<(const ScrubJob& rhs) const;
 583   };
 584   set<ScrubJob> sched_scrub_pg;
 585
 586   /// @returns the scrub_reg_stamp used for unregister the scrub job
 587   utime_t reg_pg_scrub(spg_t pgid, utime_t t, double pool_scrub_min_interval,
 588                        double pool_scrub_max_interval, bool must) {
 589     ScrubJob scrub(cct, pgid, t, pool_scrub_min_interval, pool_scrub_max_interval,
 590                    must);
 591     Mutex::Locker l(sched_scrub_lock);
 592     sched_scrub_pg.insert(scrub);
 593     return scrub.sched_time;
 594   }
 595   void unreg_pg_scrub(spg_t pgid, utime_t t) {
 596     Mutex::Locker l(sched_scrub_lock);
 597     size_t removed = sched_scrub_pg.erase(ScrubJob(cct, pgid, t));
 598     assert(removed);
 599   }
 600   bool first_scrub_stamp(ScrubJob *out) {
 601     Mutex::Locker l(sched_scrub_lock);
 602     if (sched_scrub_pg.empty())
 603       return false;
 604     set<ScrubJob>::iterator iter = sched_scrub_pg.begin();
 605     *out = *iter;
 606     return true;
 607   }
 608   bool next_scrub_stamp(const ScrubJob& next,
 609                         ScrubJob *out) {
 610     Mutex::Locker l(sched_scrub_lock);
 611     if (sched_scrub_pg.empty())
 612       return false;
 613     set<ScrubJob>::const_iterator iter = sched_scrub_pg.lower_bound(next);
 614     if (iter == sched_scrub_pg.cend())
 615       return false;
 616     ++iter;
 617     if (iter == sched_scrub_pg.cend())
 618       return false;
 619     *out = *iter;
 620     return true;
 621   }
 622
 623   void dumps_scrub(Formatter *f) {
 624     assert(f != nullptr);
 625     Mutex::Locker l(sched_scrub_lock);
 626
 627     f->open_array_section("scrubs");
 628     for (const auto &i: sched_scrub_pg) {
 629       f->open_object_section("scrub");
 630       f->dump_stream("pgid") << i.pgid;
 631       f->dump_stream("sched_time") << i.sched_time;
 632       f->dump_stream("deadline") << i.deadline;
 633       f->dump_bool("forced", i.sched_time == i.deadline);
 634       f->close_section();
 635     }
 636     f->close_section();
 637   }
 638
 639   bool can_inc_scrubs_pending();
 640   bool inc_scrubs_pending();
 641   void inc_scrubs_active(bool reserved);
 642   void dec_scrubs_pending();
 643   void dec_scrubs_active();
 644
 645   void reply_op_error(OpRequestRef op, int err);
 646   void reply_op_error(OpRequestRef op, int err, eversion_t v, version_t uv);
 647   void handle_misdirected_op(PG *pg, OpRequestRef op);
 648
 649
 650 private:
 651   // -- agent shared state --
 652   Mutex agent_lock;
 653   Cond agent_cond;
 654   map<uint64_t, set<PGRef> > agent_queue;
 655   set<PGRef>::iterator agent_queue_pos;
 656   bool agent_valid_iterator;
 657   int agent_ops;
 658   int flush_mode_high_count; //once have one pg with FLUSH_MODE_HIGH then flush objects with high speed
 659   set<hobject_t> agent_oids;
 660   bool agent_active;
 661   struct AgentThread : public Thread {
 662     OSDService *osd;
 663     explicit AgentThread(OSDService *o) : osd(o) {}
 664     void *entry() override {
 665       osd->agent_entry();
 666       return NULL;
 667     }
 668   } agent_thread;
 669   bool agent_stop_flag;
 670   Mutex agent_timer_lock;
 671   SafeTimer agent_timer;
 672
 673 public:
 674   void agent_entry();
 675   void agent_stop();
 676
 677   void _enqueue(PG *pg, uint64_t priority) {
 678     if (!agent_queue.empty() &&
 679         agent_queue.rbegin()->first < priority)
 680       agent_valid_iterator = false;  // inserting higher-priority queue
 681     set<PGRef>& nq = agent_queue[priority];
 682     if (nq.empty())
 683       agent_cond.Signal();
 684     nq.insert(pg);
 685   }
 686
 687   void _dequeue(PG *pg, uint64_t old_priority) {
 688     set<PGRef>& oq = agent_queue[old_priority];
 689     set<PGRef>::iterator p = oq.find(pg);
 690     assert(p != oq.end());
 691     if (p == agent_queue_pos)
 692       ++agent_queue_pos;
 693     oq.erase(p);
 694     if (oq.empty()) {
 695       if (agent_queue.rbegin()->first == old_priority)
 696         agent_valid_iterator = false;
 697       agent_queue.erase(old_priority);
 698     }
 699   }
 700
 701   /// enable agent for a pg
 702   void agent_enable_pg(PG *pg, uint64_t priority) {
 703     Mutex::Locker l(agent_lock);
 704     _enqueue(pg, priority);
 705   }
 706
 707   /// adjust priority for an enagled pg
 708   void agent_adjust_pg(PG *pg, uint64_t old_priority, uint64_t new_priority) {
 709     Mutex::Locker l(agent_lock);
 710     assert(new_priority != old_priority);
 711     _enqueue(pg, new_priority);
 712     _dequeue(pg, old_priority);
 713   }
 714
 715   /// disable agent for a pg
 716   void agent_disable_pg(PG *pg, uint64_t old_priority) {
 717     Mutex::Locker l(agent_lock);
 718     _dequeue(pg, old_priority);
 719   }
 720
 721   /// note start of an async (evict) op
 722   void agent_start_evict_op() {
 723     Mutex::Locker l(agent_lock);
 724     ++agent_ops;
 725   }
 726
 727   /// note finish or cancellation of an async (evict) op
 728   void agent_finish_evict_op() {
 729     Mutex::Locker l(agent_lock);
 730     assert(agent_ops > 0);
 731     --agent_ops;
 732     agent_cond.Signal();
 733   }
 734
 735   /// note start of an async (flush) op
 736   void agent_start_op(const hobject_t& oid) {
 737     Mutex::Locker l(agent_lock);
 738     ++agent_ops;
 739     assert(agent_oids.count(oid) == 0);
 740     agent_oids.insert(oid);
 741   }
 742
 743   /// note finish or cancellation of an async (flush) op
 744   void agent_finish_op(const hobject_t& oid) {
 745     Mutex::Locker l(agent_lock);
 746     assert(agent_ops > 0);
 747     --agent_ops;
 748     assert(agent_oids.count(oid) == 1);
 749     agent_oids.erase(oid);
 750     agent_cond.Signal();
 751   }
 752
 753   /// check if we are operating on an object
 754   bool agent_is_active_oid(const hobject_t& oid) {
 755     Mutex::Locker l(agent_lock);
 756     return agent_oids.count(oid);
 757   }
 758
 759   /// get count of active agent ops
 760   int agent_get_num_ops() {
 761     Mutex::Locker l(agent_lock);
 762     return agent_ops;
 763   }
 764
 765   void agent_inc_high_count() {
 766     Mutex::Locker l(agent_lock);
 767     flush_mode_high_count ++;
 768   }
 769
 770   void agent_dec_high_count() {
 771     Mutex::Locker l(agent_lock);
 772     flush_mode_high_count --;
 773   }
 774
 775 private:
 776   /// throttle promotion attempts
 777   std::atomic_uint promote_probability_millis{1000}; ///< probability thousands. one word.
 778   PromoteCounter promote_counter;
 779   utime_t last_recalibrate;
 780   unsigned long promote_max_objects, promote_max_bytes;
 781
 782 public:
 783   bool promote_throttle() {
 784     // NOTE: lockless!  we rely on the probability being a single word.
 785     promote_counter.attempt();
 786     if ((unsigned)rand() % 1000 > promote_probability_millis)
 787       return true;  // yes throttle (no promote)
 788     if (promote_max_objects &&
 789         promote_counter.objects > promote_max_objects)
 790       return true;  // yes throttle
 791     if (promote_max_bytes &&
 792         promote_counter.bytes > promote_max_bytes)
 793       return true;  // yes throttle
 794     return false;   //  no throttle (promote)
 795   }
 796   void promote_finish(uint64_t bytes) {
 797     promote_counter.finish(bytes);
 798   }
 799   void promote_throttle_recalibrate();
 800
 801   // -- Objecter, for tiering reads/writes from/to other OSDs --
 802   Objecter *objecter;
 803   Finisher objecter_finisher;
 804
 805   // -- Watch --
 806   Mutex watch_lock;
 807   SafeTimer watch_timer;
 808   uint64_t next_notif_id;
 809   uint64_t get_next_id(epoch_t cur_epoch) {
 810     Mutex::Locker l(watch_lock);
 811     return (((uint64_t)cur_epoch) << 32) | ((uint64_t)(next_notif_id++));
 812   }
 813
 814   // -- Recovery/Backfill Request Scheduling --
 815   Mutex recovery_request_lock;
 816   SafeTimer recovery_request_timer;
 817
 818   // For async recovery sleep
 819   bool recovery_needs_sleep = true;
 820   utime_t recovery_schedule_time = utime_t();
 821
 822   Mutex recovery_sleep_lock;
 823   SafeTimer recovery_sleep_timer;
 824
 825   // -- tids --
 826   // for ops i issue
 827   std::atomic_uint last_tid{0};
 828   ceph_tid_t get_tid() {
 829     return (ceph_tid_t)last_tid++;
 830   }
 831
 832   // -- backfill_reservation --
 833   Finisher reserver_finisher;
 834   AsyncReserver<spg_t> local_reserver;
 835   AsyncReserver<spg_t> remote_reserver;
 836
 837   // -- pg_temp --
 838 private:
 839   Mutex pg_temp_lock;
 840   map<pg_t, vector<int> > pg_temp_wanted;
 841   map<pg_t, vector<int> > pg_temp_pending;
 842   void _sent_pg_temp();
 843 public:
 844   void queue_want_pg_temp(pg_t pgid, vector<int>& want);
 845   void remove_want_pg_temp(pg_t pgid);
 846   void requeue_pg_temp();
 847   void send_pg_temp();
 848
 849   void send_pg_created(pg_t pgid);
 850
 851   void queue_for_peering(PG *pg);
 852
 853   Mutex snap_sleep_lock;
 854   SafeTimer snap_sleep_timer;
 855
 856   Mutex scrub_sleep_lock;
 857   SafeTimer scrub_sleep_timer;
 858
 859   AsyncReserver<spg_t> snap_reserver;
 860   void queue_for_snap_trim(PG *pg);
 861
 862   void queue_for_scrub(PG *pg, bool with_high_priority) {
 863     unsigned scrub_queue_priority = pg->scrubber.priority;
 864     if (with_high_priority && scrub_queue_priority < cct->_conf->osd_client_op_priority) {
 865       scrub_queue_priority = cct->_conf->osd_client_op_priority;
 866     }
 867     enqueue_back(
 868       pg->info.pgid,
 869       PGQueueable(
 870         PGScrub(pg->get_osdmap()->get_epoch()),
 871         cct->_conf->osd_scrub_cost,
 872         scrub_queue_priority,
 873         ceph_clock_now(),
 874         entity_inst_t(),
 875         pg->get_osdmap()->get_epoch()));
 876   }
 877
 878 private:
 879   // -- pg recovery and associated throttling --
 880   Mutex recovery_lock;
 881   list<pair<epoch_t, PGRef> > awaiting_throttle;
 882
 883   utime_t defer_recovery_until;
 884   uint64_t recovery_ops_active;
 885   uint64_t recovery_ops_reserved;
 886   bool recovery_paused;
 887 #ifdef DEBUG_RECOVERY_OIDS
 888   map<spg_t, set<hobject_t> > recovery_oids;
 889 #endif
 890   bool _recover_now(uint64_t *available_pushes);
 891   void _maybe_queue_recovery();
 892   void _queue_for_recovery(
 893     pair<epoch_t, PGRef> p, uint64_t reserved_pushes) {
 894     assert(recovery_lock.is_locked_by_me());
 895     enqueue_back(
 896       p.second->info.pgid,
 897       PGQueueable(
 898         PGRecovery(p.first, reserved_pushes),
 899         cct->_conf->osd_recovery_cost,
 900         cct->_conf->osd_recovery_priority,
 901         ceph_clock_now(),
 902         entity_inst_t(),
 903         p.first));
 904   }
 905 public:
 906   void start_recovery_op(PG *pg, const hobject_t& soid);
 907   void finish_recovery_op(PG *pg, const hobject_t& soid, bool dequeue);
 908   bool is_recovery_active();
 909   void release_reserved_pushes(uint64_t pushes) {
 910     Mutex::Locker l(recovery_lock);
 911     assert(recovery_ops_reserved >= pushes);
 912     recovery_ops_reserved -= pushes;
 913     _maybe_queue_recovery();
 914   }
 915   void defer_recovery(float defer_for) {
 916     defer_recovery_until = ceph_clock_now();
 917     defer_recovery_until += defer_for;
 918   }
 919   void pause_recovery() {
 920     Mutex::Locker l(recovery_lock);
 921     recovery_paused = true;
 922   }
 923   bool recovery_is_paused() {
 924     Mutex::Locker l(recovery_lock);
 925     return recovery_paused;
 926   }
 927   void unpause_recovery() {
 928     Mutex::Locker l(recovery_lock);
 929     recovery_paused = false;
 930     _maybe_queue_recovery();
 931   }
 932   void kick_recovery_queue() {
 933     Mutex::Locker l(recovery_lock);
 934     _maybe_queue_recovery();
 935   }
 936   void clear_queued_recovery(PG *pg) {
 937     Mutex::Locker l(recovery_lock);
 938     for (list<pair<epoch_t, PGRef> >::iterator i = awaiting_throttle.begin();
 939          i != awaiting_throttle.end();
 940       ) {
 941       if (i->second.get() == pg) {
 942         awaiting_throttle.erase(i);
 943         return;
 944       } else {
 945         ++i;
 946       }
 947     }
 948   }
 949   // delayed pg activation
 950   void queue_for_recovery(PG *pg) {
 951     Mutex::Locker l(recovery_lock);
 952
 953     if (pg->get_state() & (PG_STATE_FORCED_RECOVERY | PG_STATE_FORCED_BACKFILL)) {
 954       awaiting_throttle.push_front(make_pair(pg->get_osdmap()->get_epoch(), pg));
 955     } else {
 956       awaiting_throttle.push_back(make_pair(pg->get_osdmap()->get_epoch(), pg));
 957     }
 958     _maybe_queue_recovery();
 959   }
 960   void queue_recovery_after_sleep(PG *pg, epoch_t queued, uint64_t reserved_pushes) {
 961     Mutex::Locker l(recovery_lock);
 962     _queue_for_recovery(make_pair(queued, pg), reserved_pushes);
 963   }
 964
 965   void adjust_pg_priorities(const vector<PGRef>& pgs, int newflags);
 966
 967   // osd map cache (past osd maps)
 968   Mutex map_cache_lock;
 969   SharedLRU<epoch_t, const OSDMap> map_cache;
 970   SimpleLRU<epoch_t, bufferlist> map_bl_cache;
 971   SimpleLRU<epoch_t, bufferlist> map_bl_inc_cache;
 972
 973   OSDMapRef try_get_map(epoch_t e);
 974   OSDMapRef get_map(epoch_t e) {
 975     OSDMapRef ret(try_get_map(e));
 976     assert(ret);
 977     return ret;
 978   }
 979   OSDMapRef add_map(OSDMap *o) {
 980     Mutex::Locker l(map_cache_lock);
 981     return _add_map(o);
 982   }
 983   OSDMapRef _add_map(OSDMap *o);
 984
 985   void add_map_bl(epoch_t e, bufferlist& bl) {
 986     Mutex::Locker l(map_cache_lock);
 987     return _add_map_bl(e, bl);
 988   }
 989   void pin_map_bl(epoch_t e, bufferlist &bl);
 990   void _add_map_bl(epoch_t e, bufferlist& bl);
 991   bool get_map_bl(epoch_t e, bufferlist& bl) {
 992     Mutex::Locker l(map_cache_lock);
 993     return _get_map_bl(e, bl);
 994   }
 995   bool _get_map_bl(epoch_t e, bufferlist& bl);
 996
 997   void add_map_inc_bl(epoch_t e, bufferlist& bl) {
 998     Mutex::Locker l(map_cache_lock);
 999     return _add_map_inc_bl(e, bl);
1000   }
1001   void pin_map_inc_bl(epoch_t e, bufferlist &bl);
1002   void _add_map_inc_bl(epoch_t e, bufferlist& bl);
1003   bool get_inc_map_bl(epoch_t e, bufferlist& bl);
1004
1005   void clear_map_bl_cache_pins(epoch_t e);
1006
1007   void need_heartbeat_peer_update();
1008
1009   void pg_stat_queue_enqueue(PG *pg);
1010   void pg_stat_queue_dequeue(PG *pg);
1011
1012   void init();
1013   void final_init();
1014   void start_shutdown();
1015   void shutdown_reserver();
1016   void shutdown();
1017
1018 private:
1019   // split
1020   Mutex in_progress_split_lock;
1021   map<spg_t, spg_t> pending_splits; // child -> parent
1022   map<spg_t, set<spg_t> > rev_pending_splits; // parent -> [children]
1023   set<spg_t> in_progress_splits;       // child
1024
1025 public:
1026   void _start_split(spg_t parent, const set<spg_t> &children);
1027   void start_split(spg_t parent, const set<spg_t> &children) {
1028     Mutex::Locker l(in_progress_split_lock);
1029     return _start_split(parent, children);
1030   }
1031   void mark_split_in_progress(spg_t parent, const set<spg_t> &pgs);
1032   void complete_split(const set<spg_t> &pgs);
1033   void cancel_pending_splits_for_parent(spg_t parent);
1034   void _cancel_pending_splits_for_parent(spg_t parent);
1035   bool splitting(spg_t pgid);
1036   void expand_pg_num(OSDMapRef old_map,
1037                      OSDMapRef new_map);
1038   void _maybe_split_pgid(OSDMapRef old_map,
1039                          OSDMapRef new_map,
1040                          spg_t pgid);
1041   void init_splits_between(spg_t pgid, OSDMapRef frommap, OSDMapRef tomap);
1042
1043   // -- stats --
1044   Mutex stat_lock;
1045   osd_stat_t osd_stat;
1046   uint32_t seq = 0;
1047
1048   void update_osd_stat(vector<int>& hb_peers);
1049   osd_stat_t set_osd_stat(const struct store_statfs_t &stbuf,
1050                           vector<int>& hb_peers);
1051   osd_stat_t get_osd_stat() {
1052     Mutex::Locker l(stat_lock);
1053     ++seq;
1054     osd_stat.up_from = up_epoch;
1055     osd_stat.seq = ((uint64_t)osd_stat.up_from << 32) + seq;
1056     return osd_stat;
1057   }
1058   uint64_t get_osd_stat_seq() {
1059     Mutex::Locker l(stat_lock);
1060     return osd_stat.seq;
1061   }
1062
1063   // -- OSD Full Status --
1064 private:
1065   friend TestOpsSocketHook;
1066   mutable Mutex full_status_lock;
1067   enum s_names { INVALID = -1, NONE, NEARFULL, BACKFILLFULL, FULL, FAILSAFE } cur_state;  // ascending
1068   const char *get_full_state_name(s_names s) const {
1069     switch (s) {
1070     case NONE: return "none";
1071     case NEARFULL: return "nearfull";
1072     case BACKFILLFULL: return "backfillfull";
1073     case FULL: return "full";
1074     case FAILSAFE: return "failsafe";
1075     default: return "???";
1076     }
1077   }
1078   s_names get_full_state(string type) const {
1079     if (type == "none")
1080       return NONE;
1081     else if (type == "failsafe")
1082       return FAILSAFE;
1083     else if (type == "full")
1084       return FULL;
1085     else if (type == "backfillfull")
1086       return BACKFILLFULL;
1087     else if (type == "nearfull")
1088       return NEARFULL;
1089     else
1090       return INVALID;
1091   }
1092   double cur_ratio;  ///< current utilization
1093   mutable int64_t injectfull = 0;
1094   s_names injectfull_state = NONE;
1095   float get_failsafe_full_ratio();
1096   void check_full_status(float ratio);
1097   bool _check_full(s_names type, ostream &ss) const;
1098 public:
1099   bool check_failsafe_full(ostream &ss) const;
1100   bool check_full(ostream &ss) const;
1101   bool check_backfill_full(ostream &ss) const;
1102   bool check_nearfull(ostream &ss) const;
1103   bool is_failsafe_full() const;
1104   bool is_full() const;
1105   bool is_backfillfull() const;
1106   bool is_nearfull() const;
1107   bool need_fullness_update();  ///< osdmap state needs update
1108   void set_injectfull(s_names type, int64_t count);
1109   bool check_osdmap_full(const set<pg_shard_t> &missing_on);
1110
1111
1112   // -- epochs --
1113 private:
1114   mutable Mutex epoch_lock; // protects access to boot_epoch, up_epoch, bind_epoch
1115   epoch_t boot_epoch;  // _first_ epoch we were marked up (after this process started)
1116   epoch_t up_epoch;    // _most_recent_ epoch we were marked up
1117   epoch_t bind_epoch;  // epoch we last did a bind to new ip:ports
1118 public:
1119   /**
1120    * Retrieve the boot_, up_, and bind_ epochs the OSD has set. The params
1121    * can be NULL if you don't care about them.
1122    */
1123   void retrieve_epochs(epoch_t *_boot_epoch, epoch_t *_up_epoch,
1124                        epoch_t *_bind_epoch) const;
1125   /**
1126    * Set the boot, up, and bind epochs. Any NULL params will not be set.
1127    */
1128   void set_epochs(const epoch_t *_boot_epoch, const epoch_t *_up_epoch,
1129                   const epoch_t *_bind_epoch);
1130   epoch_t get_boot_epoch() const {
1131     epoch_t ret;
1132     retrieve_epochs(&ret, NULL, NULL);
1133     return ret;
1134   }
1135   epoch_t get_up_epoch() const {
1136     epoch_t ret;
1137     retrieve_epochs(NULL, &ret, NULL);
1138     return ret;
1139   }
1140   epoch_t get_bind_epoch() const {
1141     epoch_t ret;
1142     retrieve_epochs(NULL, NULL, &ret);
1143     return ret;
1144   }
1145
1146   // -- stopping --
1147   Mutex is_stopping_lock;
1148   Cond is_stopping_cond;
1149   enum {
1150     NOT_STOPPING,
1151     PREPARING_TO_STOP,
1152     STOPPING };
1153   std::atomic_int state{NOT_STOPPING};
1154   int get_state() {
1155     return state;
1156   }
1157   void set_state(int s) {
1158     state = s;
1159   }
1160   bool is_stopping() const {
1161     return state == STOPPING;
1162   }
1163   bool is_preparing_to_stop() const {
1164     return state == PREPARING_TO_STOP;
1165   }
1166   bool prepare_to_stop();
1167   void got_stop_ack();
1168
1169
1170 #ifdef PG_DEBUG_REFS
1171   Mutex pgid_lock;
1172   map<spg_t, int> pgid_tracker;
1173   map<spg_t, PG*> live_pgs;
1174   void add_pgid(spg_t pgid, PG *pg);
1175   void remove_pgid(spg_t pgid, PG *pg);
1176   void dump_live_pgids();
1177 #endif
1178
1179   explicit OSDService(OSD *osd);
1180   ~OSDService();
1181 };
1182
1183 class OSD : public Dispatcher,
1184             public md_config_obs_t {
1185   /** OSD **/
1186   Mutex osd_lock;                       // global lock
1187   SafeTimer tick_timer;    // safe timer (osd_lock)
1188
1189   // Tick timer for those stuff that do not need osd_lock
1190   Mutex tick_timer_lock;
1191   SafeTimer tick_timer_without_osd_lock;
1192 public:
1193   // config observer bits
1194   const char** get_tracked_conf_keys() const override;
1195   void handle_conf_change(const struct md_config_t *conf,
1196                           const std::set <std::string> &changed) override;
1197   void update_log_config();
1198   void check_config();
1199
1200 protected:
1201
1202   static const double OSD_TICK_INTERVAL; // tick interval for tick_timer and tick_timer_without_osd_lock
1203
1204   AuthAuthorizeHandlerRegistry *authorize_handler_cluster_registry;
1205   AuthAuthorizeHandlerRegistry *authorize_handler_service_registry;
1206
1207   Messenger   *cluster_messenger;
1208   Messenger   *client_messenger;
1209   Messenger   *objecter_messenger;
1210   MonClient   *monc; // check the "monc helpers" list before accessing directly
1211   MgrClient   mgrc;
1212   PerfCounters      *logger;
1213   PerfCounters      *recoverystate_perf;
1214   ObjectStore *store;
1215 #ifdef HAVE_LIBFUSE
1216   FuseStore *fuse_store = nullptr;
1217 #endif
1218   LogClient log_client;
1219   LogChannelRef clog;
1220
1221   int whoami;
1222   std::string dev_path, journal_path;
1223
1224   bool store_is_rotational = true;
1225   bool journal_is_rotational = true;
1226
1227   ZTracer::Endpoint trace_endpoint;
1228   void create_logger();
1229   void create_recoverystate_perf();
1230   void tick();
1231   void tick_without_osd_lock();
1232   void _dispatch(Message *m);
1233   void dispatch_op(OpRequestRef op);
1234
1235   void check_osdmap_features(ObjectStore *store);
1236
1237   // asok
1238   friend class OSDSocketHook;
1239   class OSDSocketHook *asok_hook;
1240   bool asok_command(string admin_command, cmdmap_t& cmdmap, string format, ostream& ss);
1241
1242 public:
1243   ClassHandler  *class_handler = nullptr;
1244   int get_nodeid() { return whoami; }
1245
1246   static ghobject_t get_osdmap_pobject_name(epoch_t epoch) {
1247     char foo[20];
1248     snprintf(foo, sizeof(foo), "osdmap.%d", epoch);
1249     return ghobject_t(hobject_t(sobject_t(object_t(foo), 0)));
1250   }
1251   static ghobject_t get_inc_osdmap_pobject_name(epoch_t epoch) {
1252     char foo[22];
1253     snprintf(foo, sizeof(foo), "inc_osdmap.%d", epoch);
1254     return ghobject_t(hobject_t(sobject_t(object_t(foo), 0)));
1255   }
1256
1257   static ghobject_t make_snapmapper_oid() {
1258     return ghobject_t(hobject_t(
1259       sobject_t(
1260         object_t("snapmapper"),
1261         0)));
1262   }
1263
1264   static ghobject_t make_pg_log_oid(spg_t pg) {
1265     stringstream ss;
1266     ss << "pglog_" << pg;
1267     string s;
1268     getline(ss, s);
1269     return ghobject_t(hobject_t(sobject_t(object_t(s.c_str()), 0)));
1270   }
1271
1272   static ghobject_t make_pg_biginfo_oid(spg_t pg) {
1273     stringstream ss;
1274     ss << "pginfo_" << pg;
1275     string s;
1276     getline(ss, s);
1277     return ghobject_t(hobject_t(sobject_t(object_t(s.c_str()), 0)));
1278   }
1279   static ghobject_t make_infos_oid() {
1280     hobject_t oid(sobject_t("infos", CEPH_NOSNAP));
1281     return ghobject_t(oid);
1282   }
1283   static void recursive_remove_collection(CephContext* cct,
1284                                           ObjectStore *store,
1285                                           spg_t pgid,
1286                                           coll_t tmp);
1287
1288   /**
1289    * get_osd_initial_compat_set()
1290    *
1291    * Get the initial feature set for this OSD.  Features
1292    * here are automatically upgraded.
1293    *
1294    * Return value: Initial osd CompatSet
1295    */
1296   static CompatSet get_osd_initial_compat_set();
1297
1298   /**
1299    * get_osd_compat_set()
1300    *
1301    * Get all features supported by this OSD
1302    *
1303    * Return value: CompatSet of all supported features
1304    */
1305   static CompatSet get_osd_compat_set();
1306
1307
1308 private:
1309   class C_Tick;
1310   class C_Tick_WithoutOSDLock;
1311
1312   // -- superblock --
1313   OSDSuperblock superblock;
1314
1315   void write_superblock();
1316   void write_superblock(ObjectStore::Transaction& t);
1317   int read_superblock();
1318
1319   void clear_temp_objects();
1320
1321   CompatSet osd_compat;
1322
1323   // -- state --
1324 public:
1325   typedef enum {
1326     STATE_INITIALIZING = 1,
1327     STATE_PREBOOT,
1328     STATE_BOOTING,
1329     STATE_ACTIVE,
1330     STATE_STOPPING,
1331     STATE_WAITING_FOR_HEALTHY
1332   } osd_state_t;
1333
1334   static const char *get_state_name(int s) {
1335     switch (s) {
1336     case STATE_INITIALIZING: return "initializing";
1337     case STATE_PREBOOT: return "preboot";
1338     case STATE_BOOTING: return "booting";
1339     case STATE_ACTIVE: return "active";
1340     case STATE_STOPPING: return "stopping";
1341     case STATE_WAITING_FOR_HEALTHY: return "waiting_for_healthy";
1342     default: return "???";
1343     }
1344   }
1345
1346 private:
1347   std::atomic_int state{STATE_INITIALIZING};
1348   bool waiting_for_luminous_mons = false;
1349
1350 public:
1351   int get_state() const {
1352     return state;
1353   }
1354   void set_state(int s) {
1355     state = s;
1356   }
1357   bool is_initializing() const {
1358     return state == STATE_INITIALIZING;
1359   }
1360   bool is_preboot() const {
1361     return state == STATE_PREBOOT;
1362   }
1363   bool is_booting() const {
1364     return state == STATE_BOOTING;
1365   }
1366   bool is_active() const {
1367     return state == STATE_ACTIVE;
1368   }
1369   bool is_stopping() const {
1370     return state == STATE_STOPPING;
1371   }
1372   bool is_waiting_for_healthy() const {
1373     return state == STATE_WAITING_FOR_HEALTHY;
1374   }
1375
1376 private:
1377
1378   ThreadPool peering_tp;
1379   ShardedThreadPool osd_op_tp;
1380   ThreadPool disk_tp;
1381   ThreadPool command_tp;
1382
1383   void set_disk_tp_priority();
1384   void get_latest_osdmap();
1385
1386   // -- sessions --
1387 private:
1388   void dispatch_session_waiting(Session *session, OSDMapRef osdmap);
1389   void maybe_share_map(Session *session, OpRequestRef op, OSDMapRef osdmap);
1390
1391   Mutex session_waiting_lock;
1392   set<Session*> session_waiting_for_map;
1393
1394   /// Caller assumes refs for included Sessions
1395   void get_sessions_waiting_for_map(set<Session*> *out) {
1396     Mutex::Locker l(session_waiting_lock);
1397     out->swap(session_waiting_for_map);
1398   }
1399   void register_session_waiting_on_map(Session *session) {
1400     Mutex::Locker l(session_waiting_lock);
1401     if (session_waiting_for_map.insert(session).second) {
1402       session->get();
1403     }
1404   }
1405   void clear_session_waiting_on_map(Session *session) {
1406     Mutex::Locker l(session_waiting_lock);
1407     set<Session*>::iterator i = session_waiting_for_map.find(session);
1408     if (i != session_waiting_for_map.end()) {
1409       (*i)->put();
1410       session_waiting_for_map.erase(i);
1411     }
1412   }
1413   void dispatch_sessions_waiting_on_map() {
1414     set<Session*> sessions_to_check;
1415     get_sessions_waiting_for_map(&sessions_to_check);
1416     for (set<Session*>::iterator i = sessions_to_check.begin();
1417          i != sessions_to_check.end();
1418          sessions_to_check.erase(i++)) {
1419       (*i)->session_dispatch_lock.Lock();
1420       dispatch_session_waiting(*i, osdmap);
1421       (*i)->session_dispatch_lock.Unlock();
1422       (*i)->put();
1423     }
1424   }
1425   void session_handle_reset(Session *session) {
1426     Mutex::Locker l(session->session_dispatch_lock);
1427     clear_session_waiting_on_map(session);
1428
1429     session->clear_backoffs();
1430
1431     /* Messages have connection refs, we need to clear the
1432      * connection->session->message->connection
1433      * cycles which result.
1434      * Bug #12338
1435      */
1436     session->waiting_on_map.clear_and_dispose(TrackedOp::Putter());
1437   }
1438
1439 private:
1440   /**
1441    * @defgroup monc helpers
1442    * @{
1443    * Right now we only have the one
1444    */
1445
1446   /**
1447    * Ask the Monitors for a sequence of OSDMaps.
1448    *
1449    * @param epoch The epoch to start with when replying
1450    * @param force_request True if this request forces a new subscription to
1451    * the monitors; false if an outstanding request that encompasses it is
1452    * sufficient.
1453    */
1454   void osdmap_subscribe(version_t epoch, bool force_request);
1455   /** @} monc helpers */
1456
1457   // -- heartbeat --
1458   /// information about a heartbeat peer
1459   struct HeartbeatInfo {
1460     int peer;           ///< peer
1461     ConnectionRef con_front;   ///< peer connection (front)
1462     ConnectionRef con_back;    ///< peer connection (back)
1463     utime_t first_tx;   ///< time we sent our first ping request
1464     utime_t last_tx;    ///< last time we sent a ping request
1465     utime_t last_rx_front;  ///< last time we got a ping reply on the front side
1466     utime_t last_rx_back;   ///< last time we got a ping reply on the back side
1467     epoch_t epoch;      ///< most recent epoch we wanted this peer
1468
1469     bool is_unhealthy(utime_t cutoff) const {
1470       return
1471         ! ((last_rx_front > cutoff ||
1472             (last_rx_front == utime_t() && (last_tx == utime_t() ||
1473                                             first_tx > cutoff))) &&
1474            (last_rx_back > cutoff ||
1475             (last_rx_back == utime_t() && (last_tx == utime_t() ||
1476                                            first_tx > cutoff))));
1477     }
1478     bool is_healthy(utime_t cutoff) const {
1479       return last_rx_front > cutoff && last_rx_back > cutoff;
1480     }
1481
1482   };
1483   /// state attached to outgoing heartbeat connections
1484   struct HeartbeatSession : public RefCountedObject {
1485     int peer;
1486     explicit HeartbeatSession(int p) : peer(p) {}
1487   };
1488   Mutex heartbeat_lock;
1489   map<int, int> debug_heartbeat_drops_remaining;
1490   Cond heartbeat_cond;
1491   bool heartbeat_stop;
1492   std::atomic_bool heartbeat_need_update;
1493   map<int,HeartbeatInfo> heartbeat_peers;  ///< map of osd id to HeartbeatInfo
1494   utime_t last_mon_heartbeat;
1495   Messenger *hb_front_client_messenger;
1496   Messenger *hb_back_client_messenger;
1497   Messenger *hb_front_server_messenger;
1498   Messenger *hb_back_server_messenger;
1499   utime_t last_heartbeat_resample;   ///< last time we chose random peers in waiting-for-healthy state
1500   double daily_loadavg;
1501
1502   void _add_heartbeat_peer(int p);
1503   void _remove_heartbeat_peer(int p);
1504   bool heartbeat_reset(Connection *con);
1505   void maybe_update_heartbeat_peers();
1506   void reset_heartbeat_peers();
1507   bool heartbeat_peers_need_update() {
1508     return heartbeat_need_update.load();
1509   }
1510   void heartbeat_set_peers_need_update() {
1511     heartbeat_need_update.store(true);
1512   }
1513   void heartbeat_clear_peers_need_update() {
1514     heartbeat_need_update.store(false);
1515   }
1516   void heartbeat();
1517   void heartbeat_check();
1518   void heartbeat_entry();
1519   void need_heartbeat_peer_update();
1520
1521   void heartbeat_kick() {
1522     Mutex::Locker l(heartbeat_lock);
1523     heartbeat_cond.Signal();
1524   }
1525
1526   struct T_Heartbeat : public Thread {
1527     OSD *osd;
1528     explicit T_Heartbeat(OSD *o) : osd(o) {}
1529     void *entry() override {
1530       osd->heartbeat_entry();
1531       return 0;
1532     }
1533   } heartbeat_thread;
1534
1535 public:
1536   bool heartbeat_dispatch(Message *m);
1537
1538   struct HeartbeatDispatcher : public Dispatcher {
1539     OSD *osd;
1540     explicit HeartbeatDispatcher(OSD *o) : Dispatcher(o->cct), osd(o) {}
1541
1542     bool ms_can_fast_dispatch_any() const override { return true; }
1543     bool ms_can_fast_dispatch(const Message *m) const override {
1544       switch (m->get_type()) {
1545         case CEPH_MSG_PING:
1546         case MSG_OSD_PING:
1547           return true;
1548         default:
1549           return false;
1550         }
1551     }
1552     void ms_fast_dispatch(Message *m) override {
1553       osd->heartbeat_dispatch(m);
1554     }
1555     bool ms_dispatch(Message *m) override {
1556       return osd->heartbeat_dispatch(m);
1557     }
1558     bool ms_handle_reset(Connection *con) override {
1559       return osd->heartbeat_reset(con);
1560     }
1561     void ms_handle_remote_reset(Connection *con) override {}
1562     bool ms_handle_refused(Connection *con) override {
1563       return osd->ms_handle_refused(con);
1564     }
1565     bool ms_verify_authorizer(Connection *con, int peer_type,
1566                               int protocol, bufferlist& authorizer_data, bufferlist& authorizer_reply,
1567                               bool& isvalid, CryptoKey& session_key) override {
1568       isvalid = true;
1569       return true;
1570     }
1571   } heartbeat_dispatcher;
1572
1573 private:
1574   // -- waiters --
1575   list<OpRequestRef> finished;
1576
1577   void take_waiters(list<OpRequestRef>& ls) {
1578     assert(osd_lock.is_locked());
1579     finished.splice(finished.end(), ls);
1580   }
1581   void do_waiters();
1582
1583   // -- op tracking --
1584   OpTracker op_tracker;
1585   void check_ops_in_flight();
1586   void test_ops(std::string command, std::string args, ostream& ss);
1587   friend class TestOpsSocketHook;
1588   TestOpsSocketHook *test_ops_hook;
1589   friend struct C_CompleteSplits;
1590   friend struct C_OpenPGs;
1591
1592   // -- op queue --
1593   enum class io_queue {
1594     prioritized,
1595     weightedpriority,
1596     mclock_opclass,
1597     mclock_client,
1598   };
1599   friend std::ostream& operator<<(std::ostream& out, const OSD::io_queue& q);
1600
1601   const io_queue op_queue;
1602   const unsigned int op_prio_cutoff;
1603
1604   /*
1605    * The ordered op delivery chain is:
1606    *
1607    *   fast dispatch -> pqueue back
1608    *                    pqueue front <-> to_process back
1609    *                                     to_process front  -> RunVis(item)
1610    *                                                      <- queue_front()
1611    *
1612    * The pqueue is per-shard, and to_process is per pg_slot.  Items can be
1613    * pushed back up into to_process and/or pqueue while order is preserved.
1614    *
1615    * Multiple worker threads can operate on each shard.
1616    *
1617    * Under normal circumstances, num_running == to_proces.size().  There are
1618    * two times when that is not true: (1) when waiting_for_pg == true and
1619    * to_process is accumulating requests that are waiting for the pg to be
1620    * instantiated; in that case they will all get requeued together by
1621    * wake_pg_waiters, and (2) when wake_pg_waiters just ran, waiting_for_pg
1622    * and already requeued the items.
1623    */
1624   friend class PGQueueable;
1625
1626   class ShardedOpWQ
1627     : public ShardedThreadPool::ShardedWQ<pair<spg_t,PGQueueable>>
1628   {
1629     struct ShardData {
1630       Mutex sdata_lock;
1631       Cond sdata_cond;
1632
1633       Mutex sdata_op_ordering_lock;   ///< protects all members below
1634
1635       OSDMapRef waiting_for_pg_osdmap;
1636       struct pg_slot {
1637         PGRef pg;                     ///< cached pg reference [optional]
1638         list<PGQueueable> to_process; ///< order items for this slot
1639         int num_running = 0;          ///< _process threads doing pg lookup/lock
1640
1641         /// true if pg does/did not exist. if so all new items go directly to
1642         /// to_process.  cleared by prune_pg_waiters.
1643         bool waiting_for_pg = false;
1644
1645         /// incremented by wake_pg_waiters; indicates racing _process threads
1646         /// should bail out (their op has been requeued)
1647         uint64_t requeue_seq = 0;
1648       };
1649
1650       /// map of slots for each spg_t.  maintains ordering of items dequeued
1651       /// from pqueue while _process thread drops shard lock to acquire the
1652       /// pg lock.  slots are removed only by prune_pg_waiters.
1653       unordered_map<spg_t,pg_slot> pg_slots;
1654
1655       /// priority queue
1656       std::unique_ptr<OpQueue< pair<spg_t, PGQueueable>, entity_inst_t>> pqueue;
1657
1658       void _enqueue_front(pair<spg_t, PGQueueable> item, unsigned cutoff) {
1659         unsigned priority = item.second.get_priority();
1660         unsigned cost = item.second.get_cost();
1661         if (priority >= cutoff)
1662           pqueue->enqueue_strict_front(
1663             item.second.get_owner(),
1664             priority, item);
1665         else
1666           pqueue->enqueue_front(
1667             item.second.get_owner(),
1668             priority, cost, item);
1669       }
1670
1671       ShardData(
1672         string lock_name, string ordering_lock,
1673         uint64_t max_tok_per_prio, uint64_t min_cost, CephContext *cct,
1674         io_queue opqueue)
1675         : sdata_lock(lock_name.c_str(), false, true, false, cct),
1676           sdata_op_ordering_lock(ordering_lock.c_str(), false, true,
1677                                  false, cct) {
1678         if (opqueue == io_queue::weightedpriority) {
1679           pqueue = std::unique_ptr
1680             <WeightedPriorityQueue<pair<spg_t,PGQueueable>,entity_inst_t>>(
1681               new WeightedPriorityQueue<pair<spg_t,PGQueueable>,entity_inst_t>(
1682                 max_tok_per_prio, min_cost));
1683         } else if (opqueue == io_queue::prioritized) {
1684           pqueue = std::unique_ptr
1685             <PrioritizedQueue<pair<spg_t,PGQueueable>,entity_inst_t>>(
1686               new PrioritizedQueue<pair<spg_t,PGQueueable>,entity_inst_t>(
1687                 max_tok_per_prio, min_cost));
1688         } else if (opqueue == io_queue::mclock_opclass) {
1689           pqueue = std::unique_ptr
1690             <ceph::mClockOpClassQueue>(new ceph::mClockOpClassQueue(cct));
1691         } else if (opqueue == io_queue::mclock_client) {
1692           pqueue = std::unique_ptr
1693             <ceph::mClockClientQueue>(new ceph::mClockClientQueue(cct));
1694         }
1695       }
1696     }; // struct ShardData
1697
1698     vector<ShardData*> shard_list;
1699     OSD *osd;
1700     uint32_t num_shards;
1701
1702   public:
1703     ShardedOpWQ(uint32_t pnum_shards,
1704                 OSD *o,
1705                 time_t ti,
1706                 time_t si,
1707                 ShardedThreadPool* tp)
1708       : ShardedThreadPool::ShardedWQ<pair<spg_t,PGQueueable>>(ti, si, tp),
1709         osd(o),
1710         num_shards(pnum_shards) {
1711       for (uint32_t i = 0; i < num_shards; i++) {
1712         char lock_name[32] = {0};
1713         snprintf(lock_name, sizeof(lock_name), "%s.%d", "OSD:ShardedOpWQ:", i);
1714         char order_lock[32] = {0};
1715         snprintf(order_lock, sizeof(order_lock), "%s.%d",
1716                  "OSD:ShardedOpWQ:order:", i);
1717         ShardData* one_shard = new ShardData(
1718           lock_name, order_lock,
1719           osd->cct->_conf->osd_op_pq_max_tokens_per_priority,
1720           osd->cct->_conf->osd_op_pq_min_cost, osd->cct, osd->op_queue);
1721         shard_list.push_back(one_shard);
1722       }
1723     }
1724     ~ShardedOpWQ() override {
1725       while (!shard_list.empty()) {
1726         delete shard_list.back();
1727         shard_list.pop_back();
1728       }
1729     }
1730
1731     /// wake any pg waiters after a PG is created/instantiated
1732     void wake_pg_waiters(spg_t pgid);
1733
1734     /// prune ops (and possiblye pg_slots) for pgs that shouldn't be here
1735     void prune_pg_waiters(OSDMapRef osdmap, int whoami);
1736
1737     /// clear cached PGRef on pg deletion
1738     void clear_pg_pointer(spg_t pgid);
1739
1740     /// clear pg_slots on shutdown
1741     void clear_pg_slots();
1742
1743     /// try to do some work
1744     void _process(uint32_t thread_index, heartbeat_handle_d *hb) override;
1745
1746     /// enqueue a new item
1747     void _enqueue(pair <spg_t, PGQueueable> item) override;
1748
1749     /// requeue an old item (at the front of the line)
1750     void _enqueue_front(pair <spg_t, PGQueueable> item) override;
1751
1752     void return_waiting_threads() override {
1753       for(uint32_t i = 0; i < num_shards; i++) {
1754         ShardData* sdata = shard_list[i];
1755         assert (NULL != sdata);
1756         sdata->sdata_lock.Lock();
1757         sdata->sdata_cond.Signal();
1758         sdata->sdata_lock.Unlock();
1759       }
1760     }
1761
1762     void dump(Formatter *f) {
1763       for(uint32_t i = 0; i < num_shards; i++) {
1764         ShardData* sdata = shard_list[i];
1765         char lock_name[32] = {0};
1766         snprintf(lock_name, sizeof(lock_name), "%s%d", "OSD:ShardedOpWQ:", i);
1767         assert (NULL != sdata);
1768         sdata->sdata_op_ordering_lock.Lock();
1769         f->open_object_section(lock_name);
1770         sdata->pqueue->dump(f);
1771         f->close_section();
1772         sdata->sdata_op_ordering_lock.Unlock();
1773       }
1774     }
1775
1776     /// Must be called on ops queued back to front
1777     struct Pred {
1778       spg_t pgid;
1779       list<OpRequestRef> *out_ops;
1780       uint64_t reserved_pushes_to_free;
1781       Pred(spg_t pg, list<OpRequestRef> *out_ops = 0)
1782         : pgid(pg), out_ops(out_ops), reserved_pushes_to_free(0) {}
1783       void accumulate(const PGQueueable &op) {
1784         reserved_pushes_to_free += op.get_reserved_pushes();
1785         if (out_ops) {
1786           boost::optional<OpRequestRef> mop = op.maybe_get_op();
1787           if (mop)
1788             out_ops->push_front(*mop);
1789         }
1790       }
1791       bool operator()(const pair<spg_t, PGQueueable> &op) {
1792         if (op.first == pgid) {
1793           accumulate(op.second);
1794           return true;
1795         } else {
1796           return false;
1797         }
1798       }
1799       uint64_t get_reserved_pushes_to_free() const {
1800         return reserved_pushes_to_free;
1801       }
1802     };
1803
1804     bool is_shard_empty(uint32_t thread_index) override {
1805       uint32_t shard_index = thread_index % num_shards;
1806       ShardData* sdata = shard_list[shard_index];
1807       assert(NULL != sdata);
1808       Mutex::Locker l(sdata->sdata_op_ordering_lock);
1809       return sdata->pqueue->empty();
1810     }
1811   } op_shardedwq;
1812
1813
1814   void enqueue_op(spg_t pg, OpRequestRef& op, epoch_t epoch);
1815   void dequeue_op(
1816     PGRef pg, OpRequestRef op,
1817     ThreadPool::TPHandle &handle);
1818
1819   // -- peering queue --
1820   struct PeeringWQ : public ThreadPool::BatchWorkQueue<PG> {
1821     list<PG*> peering_queue;
1822     OSD *osd;
1823     set<PG*> in_use;
1824     PeeringWQ(OSD *o, time_t ti, time_t si, ThreadPool *tp)
1825       : ThreadPool::BatchWorkQueue<PG>(
1826         "OSD::PeeringWQ", ti, si, tp), osd(o) {}
1827
1828     void _dequeue(PG *pg) override {
1829       for (list<PG*>::iterator i = peering_queue.begin();
1830            i != peering_queue.end();
1831            ) {
1832         if (*i == pg) {
1833           peering_queue.erase(i++);
1834           pg->put("PeeringWQ");
1835         } else {
1836           ++i;
1837         }
1838       }
1839     }
1840     bool _enqueue(PG *pg) override {
1841       pg->get("PeeringWQ");
1842       peering_queue.push_back(pg);
1843       return true;
1844     }
1845     bool _empty() override {
1846       return peering_queue.empty();
1847     }
1848     void _dequeue(list<PG*> *out) override;
1849     void _process(
1850       const list<PG *> &pgs,
1851       ThreadPool::TPHandle &handle) override {
1852       assert(!pgs.empty());
1853       osd->process_peering_events(pgs, handle);
1854       for (list<PG *>::const_iterator i = pgs.begin();
1855            i != pgs.end();
1856            ++i) {
1857         (*i)->put("PeeringWQ");
1858       }
1859     }
1860     void _process_finish(const list<PG *> &pgs) override {
1861       for (list<PG*>::const_iterator i = pgs.begin();
1862            i != pgs.end();
1863            ++i) {
1864         in_use.erase(*i);
1865       }
1866     }
1867     void _clear() override {
1868       assert(peering_queue.empty());
1869     }
1870   } peering_wq;
1871
1872   void process_peering_events(
1873     const list<PG*> &pg,
1874     ThreadPool::TPHandle &handle);
1875
1876   friend class PG;
1877   friend class PrimaryLogPG;
1878
1879
1880  protected:
1881
1882   // -- osd map --
1883   OSDMapRef       osdmap;
1884   OSDMapRef get_osdmap() {
1885     return osdmap;
1886   }
1887   epoch_t get_osdmap_epoch() const {
1888     return osdmap ? osdmap->get_epoch() : 0;
1889   }
1890
1891   utime_t         had_map_since;
1892   RWLock          map_lock;
1893   list<OpRequestRef>  waiting_for_osdmap;
1894   deque<utime_t> osd_markdown_log;
1895
1896   friend struct send_map_on_destruct;
1897
1898   void wait_for_new_map(OpRequestRef op);
1899   void handle_osd_map(class MOSDMap *m);
1900   void _committed_osd_maps(epoch_t first, epoch_t last, class MOSDMap *m);
1901   void trim_maps(epoch_t oldest, int nreceived, bool skip_maps);
1902   void note_down_osd(int osd);
1903   void note_up_osd(int osd);
1904   friend class C_OnMapCommit;
1905
1906   bool advance_pg(
1907     epoch_t advance_to, PG *pg,
1908     ThreadPool::TPHandle &handle,
1909     PG::RecoveryCtx *rctx,
1910     set<PGRef> *split_pgs
1911   );
1912   void consume_map();
1913   void activate_map();
1914
1915   // osd map cache (past osd maps)
1916   OSDMapRef get_map(epoch_t e) {
1917     return service.get_map(e);
1918   }
1919   OSDMapRef add_map(OSDMap *o) {
1920     return service.add_map(o);
1921   }
1922   void add_map_bl(epoch_t e, bufferlist& bl) {
1923     return service.add_map_bl(e, bl);
1924   }
1925   void pin_map_bl(epoch_t e, bufferlist &bl) {
1926     return service.pin_map_bl(e, bl);
1927   }
1928   bool get_map_bl(epoch_t e, bufferlist& bl) {
1929     return service.get_map_bl(e, bl);
1930   }
1931   void add_map_inc_bl(epoch_t e, bufferlist& bl) {
1932     return service.add_map_inc_bl(e, bl);
1933   }
1934   void pin_map_inc_bl(epoch_t e, bufferlist &bl) {
1935     return service.pin_map_inc_bl(e, bl);
1936   }
1937
1938 protected:
1939   // -- placement groups --
1940   RWLock pg_map_lock; // this lock orders *above* individual PG _locks
1941   ceph::unordered_map<spg_t, PG*> pg_map; // protected by pg_map lock
1942
1943   map<spg_t, list<PG::CephPeeringEvtRef> > peering_wait_for_split;
1944   PGRecoveryStats pg_recovery_stats;
1945
1946   PGPool _get_pool(int id, OSDMapRef createmap);
1947
1948   PG   *_lookup_lock_pg_with_map_lock_held(spg_t pgid);
1949   PG   *_lookup_lock_pg(spg_t pgid);
1950
1951 public:
1952   PG   *lookup_lock_pg(spg_t pgid);
1953
1954 protected:
1955   PG   *_open_lock_pg(OSDMapRef createmap,
1956                       spg_t pg, bool no_lockdep_check=false);
1957   enum res_result {
1958     RES_PARENT,    // resurrected a parent
1959     RES_SELF,      // resurrected self
1960     RES_NONE       // nothing relevant deleting
1961   };
1962   res_result _try_resurrect_pg(
1963     OSDMapRef curmap, spg_t pgid, spg_t *resurrected, PGRef *old_pg_state);
1964
1965   PG   *_create_lock_pg(
1966     OSDMapRef createmap,
1967     spg_t pgid,
1968     bool hold_map_lock,
1969     bool backfill,
1970     int role,
1971     vector<int>& up, int up_primary,
1972     vector<int>& acting, int acting_primary,
1973     pg_history_t history,
1974     const PastIntervals& pi,
1975     ObjectStore::Transaction& t);
1976
1977   PG* _make_pg(OSDMapRef createmap, spg_t pgid);
1978   void add_newly_split_pg(PG *pg,
1979                           PG::RecoveryCtx *rctx);
1980
1981   int handle_pg_peering_evt(
1982     spg_t pgid,
1983     const pg_history_t& orig_history,
1984     const PastIntervals& pi,
1985     epoch_t epoch,
1986     PG::CephPeeringEvtRef evt);
1987
1988   void load_pgs();
1989   void build_past_intervals_parallel();
1990
1991   /// build initial pg history and intervals on create
1992   void build_initial_pg_history(
1993     spg_t pgid,
1994     epoch_t created,
1995     utime_t created_stamp,
1996     pg_history_t *h,
1997     PastIntervals *pi);
1998
1999   /// project pg history from from to now
2000   bool project_pg_history(
2001     spg_t pgid, pg_history_t& h, epoch_t from,
2002     const vector<int>& lastup,
2003     int lastupprimary,
2004     const vector<int>& lastacting,
2005     int lastactingprimary
2006     ); ///< @return false if there was a map gap between from and now
2007
2008   // this must be called with pg->lock held on any pg addition to pg_map
2009   void wake_pg_waiters(PGRef pg) {
2010     assert(pg->is_locked());
2011     op_shardedwq.wake_pg_waiters(pg->info.pgid);
2012   }
2013   epoch_t last_pg_create_epoch;
2014
2015   void handle_pg_create(OpRequestRef op);
2016
2017   void split_pgs(
2018     PG *parent,
2019     const set<spg_t> &childpgids, set<PGRef> *out_pgs,
2020     OSDMapRef curmap,
2021     OSDMapRef nextmap,
2022     PG::RecoveryCtx *rctx);
2023
2024   // == monitor interaction ==
2025   Mutex mon_report_lock;
2026   utime_t last_mon_report;
2027   utime_t last_pg_stats_sent;
2028
2029   /* if our monitor dies, we want to notice it and reconnect.
2030    *  So we keep track of when it last acked our stat updates,
2031    *  and if too much time passes (and we've been sending
2032    *  more updates) then we can call it dead and reconnect
2033    *  elsewhere.
2034    */
2035   utime_t last_pg_stats_ack;
2036   float stats_ack_timeout;
2037   set<uint64_t> outstanding_pg_stats; // how many stat updates haven't been acked yet
2038
2039   // -- boot --
2040   void start_boot();
2041   void _got_mon_epochs(epoch_t oldest, epoch_t newest);
2042   void _preboot(epoch_t oldest, epoch_t newest);
2043   void _send_boot();
2044   void _collect_metadata(map<string,string> *pmeta);
2045
2046   void start_waiting_for_healthy();
2047   bool _is_healthy();
2048
2049   void send_full_update();
2050
2051   friend struct C_OSD_GetVersion;
2052
2053   // -- alive --
2054   epoch_t up_thru_wanted;
2055
2056   void queue_want_up_thru(epoch_t want);
2057   void send_alive();
2058
2059   // -- full map requests --
2060   epoch_t requested_full_first, requested_full_last;
2061
2062   void request_full_map(epoch_t first, epoch_t last);
2063   void rerequest_full_maps() {
2064     epoch_t first = requested_full_first;
2065     epoch_t last = requested_full_last;
2066     requested_full_first = 0;
2067     requested_full_last = 0;
2068     request_full_map(first, last);
2069   }
2070   void got_full_map(epoch_t e);
2071
2072   // -- failures --
2073   map<int,utime_t> failure_queue;
2074   map<int,pair<utime_t,entity_inst_t> > failure_pending;
2075
2076   void requeue_failures();
2077   void send_failures();
2078   void send_still_alive(epoch_t epoch, const entity_inst_t &i);
2079
2080   // -- pg stats --
2081   Mutex pg_stat_queue_lock;
2082   Cond pg_stat_queue_cond;
2083   xlist<PG*> pg_stat_queue;
2084   bool osd_stat_updated;
2085   uint64_t pg_stat_tid, pg_stat_tid_flushed;
2086
2087   void send_pg_stats(const utime_t &now);
2088   void handle_pg_stats_ack(class MPGStatsAck *ack);
2089   void flush_pg_stats();
2090
2091   ceph::coarse_mono_clock::time_point last_sent_beacon;
2092   Mutex min_last_epoch_clean_lock{"OSD::min_last_epoch_clean_lock"};
2093   epoch_t min_last_epoch_clean = 0;
2094   // which pgs were scanned for min_lec
2095   std::vector<pg_t> min_last_epoch_clean_pgs;
2096   void send_beacon(const ceph::coarse_mono_clock::time_point& now);
2097
2098   void pg_stat_queue_enqueue(PG *pg) {
2099     pg_stat_queue_lock.Lock();
2100     if (pg->is_primary() && !pg->stat_queue_item.is_on_list()) {
2101       pg->get("pg_stat_queue");
2102       pg_stat_queue.push_back(&pg->stat_queue_item);
2103     }
2104     osd_stat_updated = true;
2105     pg_stat_queue_lock.Unlock();
2106   }
2107   void pg_stat_queue_dequeue(PG *pg) {
2108     pg_stat_queue_lock.Lock();
2109     if (pg->stat_queue_item.remove_myself())
2110       pg->put("pg_stat_queue");
2111     pg_stat_queue_lock.Unlock();
2112   }
2113   void clear_pg_stat_queue() {
2114     pg_stat_queue_lock.Lock();
2115     while (!pg_stat_queue.empty()) {
2116       PG *pg = pg_stat_queue.front();
2117       pg_stat_queue.pop_front();
2118       pg->put("pg_stat_queue");
2119     }
2120     pg_stat_queue_lock.Unlock();
2121   }
2122   void clear_outstanding_pg_stats(){
2123     Mutex::Locker l(pg_stat_queue_lock);
2124     outstanding_pg_stats.clear();
2125   }
2126
2127   ceph_tid_t get_tid() {
2128     return service.get_tid();
2129   }
2130
2131   // -- generic pg peering --
2132   PG::RecoveryCtx create_context();
2133   void dispatch_context(PG::RecoveryCtx &ctx, PG *pg, OSDMapRef curmap,
2134                         ThreadPool::TPHandle *handle = NULL);
2135   void dispatch_context_transaction(PG::RecoveryCtx &ctx, PG *pg,
2136                                     ThreadPool::TPHandle *handle = NULL);
2137   void do_notifies(map<int,
2138                        vector<pair<pg_notify_t, PastIntervals> > >&
2139                        notify_list,
2140                    OSDMapRef map);
2141   void do_queries(map<int, map<spg_t,pg_query_t> >& query_map,
2142                   OSDMapRef map);
2143   void do_infos(map<int,
2144                     vector<pair<pg_notify_t, PastIntervals> > >& info_map,
2145                 OSDMapRef map);
2146
2147   bool require_mon_peer(const Message *m);
2148   bool require_mon_or_mgr_peer(const Message *m);
2149   bool require_osd_peer(const Message *m);
2150   /***
2151    * Verifies that we were alive in the given epoch, and that
2152    * still are.
2153    */
2154   bool require_self_aliveness(const Message *m, epoch_t alive_since);
2155   /**
2156    * Verifies that the OSD who sent the given op has the same
2157    * address as in the given map.
2158    * @pre op was sent by an OSD using the cluster messenger
2159    */
2160   bool require_same_peer_instance(const Message *m, OSDMapRef& map,
2161                                   bool is_fast_dispatch);
2162
2163   bool require_same_or_newer_map(OpRequestRef& op, epoch_t e,
2164                                  bool is_fast_dispatch);
2165
2166   void handle_pg_query(OpRequestRef op);
2167   void handle_pg_notify(OpRequestRef op);
2168   void handle_pg_log(OpRequestRef op);
2169   void handle_pg_info(OpRequestRef op);
2170   void handle_pg_trim(OpRequestRef op);
2171
2172   void handle_pg_backfill_reserve(OpRequestRef op);
2173   void handle_pg_recovery_reserve(OpRequestRef op);
2174
2175   void handle_force_recovery(Message *m);
2176
2177   void handle_pg_remove(OpRequestRef op);
2178   void _remove_pg(PG *pg);
2179
2180   // -- commands --
2181   struct Command {
2182     vector<string> cmd;
2183     ceph_tid_t tid;
2184     bufferlist indata;
2185     ConnectionRef con;
2186
2187     Command(vector<string>& c, ceph_tid_t t, bufferlist& bl, Connection *co)
2188       : cmd(c), tid(t), indata(bl), con(co) {}
2189   };
2190   list<Command*> command_queue;
2191   struct CommandWQ : public ThreadPool::WorkQueue<Command> {
2192     OSD *osd;
2193     CommandWQ(OSD *o, time_t ti, time_t si, ThreadPool *tp)
2194       : ThreadPool::WorkQueue<Command>("OSD::CommandWQ", ti, si, tp), osd(o) {}
2195
2196     bool _empty() override {
2197       return osd->command_queue.empty();
2198     }
2199     bool _enqueue(Command *c) override {
2200       osd->command_queue.push_back(c);
2201       return true;
2202     }
2203     void _dequeue(Command *pg) override {
2204       ceph_abort();
2205     }
2206     Command *_dequeue() override {
2207       if (osd->command_queue.empty())
2208         return NULL;
2209       Command *c = osd->command_queue.front();
2210       osd->command_queue.pop_front();
2211       return c;
2212     }
2213     void _process(Command *c, ThreadPool::TPHandle &) override {
2214       osd->osd_lock.Lock();
2215       if (osd->is_stopping()) {
2216         osd->osd_lock.Unlock();
2217         delete c;
2218         return;
2219       }
2220       osd->do_command(c->con.get(), c->tid, c->cmd, c->indata);
2221       osd->osd_lock.Unlock();
2222       delete c;
2223     }
2224     void _clear() override {
2225       while (!osd->command_queue.empty()) {
2226         Command *c = osd->command_queue.front();
2227         osd->command_queue.pop_front();
2228         delete c;
2229       }
2230     }
2231   } command_wq;
2232
2233   void handle_command(class MMonCommand *m);
2234   void handle_command(class MCommand *m);
2235   void do_command(Connection *con, ceph_tid_t tid, vector<string>& cmd, bufferlist& data);
2236
2237   // -- pg recovery --
2238   void do_recovery(PG *pg, epoch_t epoch_queued, uint64_t pushes_reserved,
2239                    ThreadPool::TPHandle &handle);
2240
2241
2242   // -- scrubbing --
2243   void sched_scrub();
2244   bool scrub_random_backoff();
2245   bool scrub_load_below_threshold();
2246   bool scrub_time_permit(utime_t now);
2247
2248   // -- removing --
2249   struct RemoveWQ :
2250     public ThreadPool::WorkQueueVal<pair<PGRef, DeletingStateRef> > {
2251     CephContext* cct;
2252     ObjectStore *&store;
2253     list<pair<PGRef, DeletingStateRef> > remove_queue;
2254     RemoveWQ(CephContext* cct, ObjectStore *&o, time_t ti, time_t si,
2255              ThreadPool *tp)
2256       : ThreadPool::WorkQueueVal<pair<PGRef, DeletingStateRef> >(
2257         "OSD::RemoveWQ", ti, si, tp), cct(cct), store(o) {}
2258
2259     bool _empty() override {
2260       return remove_queue.empty();
2261     }
2262     void _enqueue(pair<PGRef, DeletingStateRef> item) override {
2263       remove_queue.push_back(item);
2264     }
2265     void _enqueue_front(pair<PGRef, DeletingStateRef> item) override {
2266       remove_queue.push_front(item);
2267     }
2268     bool _dequeue(pair<PGRef, DeletingStateRef> item) {
2269       ceph_abort();
2270     }
2271     pair<PGRef, DeletingStateRef> _dequeue() override {
2272       assert(!remove_queue.empty());
2273       pair<PGRef, DeletingStateRef> item = remove_queue.front();
2274       remove_queue.pop_front();
2275       return item;
2276     }
2277     void _process(pair<PGRef, DeletingStateRef>,
2278                   ThreadPool::TPHandle &) override;
2279     void _clear() override {
2280       remove_queue.clear();
2281     }
2282   } remove_wq;
2283
2284 private:
2285   bool ms_can_fast_dispatch_any() const override { return true; }
2286   bool ms_can_fast_dispatch(const Message *m) const override {
2287     switch (m->get_type()) {
2288     case CEPH_MSG_OSD_OP:
2289     case CEPH_MSG_OSD_BACKOFF:
2290     case MSG_OSD_SUBOP:
2291     case MSG_OSD_REPOP:
2292     case MSG_OSD_SUBOPREPLY:
2293     case MSG_OSD_REPOPREPLY:
2294     case MSG_OSD_PG_PUSH:
2295     case MSG_OSD_PG_PULL:
2296     case MSG_OSD_PG_PUSH_REPLY:
2297     case MSG_OSD_PG_SCAN:
2298     case MSG_OSD_PG_BACKFILL:
2299     case MSG_OSD_PG_BACKFILL_REMOVE:
2300     case MSG_OSD_EC_WRITE:
2301     case MSG_OSD_EC_WRITE_REPLY:
2302     case MSG_OSD_EC_READ:
2303     case MSG_OSD_EC_READ_REPLY:
2304     case MSG_OSD_SCRUB_RESERVE:
2305     case MSG_OSD_REP_SCRUB:
2306     case MSG_OSD_REP_SCRUBMAP:
2307     case MSG_OSD_PG_UPDATE_LOG_MISSING:
2308     case MSG_OSD_PG_UPDATE_LOG_MISSING_REPLY:
2309     case MSG_OSD_PG_RECOVERY_DELETE:
2310     case MSG_OSD_PG_RECOVERY_DELETE_REPLY:
2311       return true;
2312     default:
2313       return false;
2314     }
2315   }
2316   void ms_fast_dispatch(Message *m) override;
2317   void ms_fast_preprocess(Message *m) override;
2318   bool ms_dispatch(Message *m) override;
2319   bool ms_get_authorizer(int dest_type, AuthAuthorizer **authorizer, bool force_new) override;
2320   bool ms_verify_authorizer(Connection *con, int peer_type,
2321                             int protocol, bufferlist& authorizer, bufferlist& authorizer_reply,
2322                             bool& isvalid, CryptoKey& session_key) override;
2323   void ms_handle_connect(Connection *con) override;
2324   void ms_handle_fast_connect(Connection *con) override;
2325   void ms_handle_fast_accept(Connection *con) override;
2326   bool ms_handle_reset(Connection *con) override;
2327   void ms_handle_remote_reset(Connection *con) override {}
2328   bool ms_handle_refused(Connection *con) override;
2329
2330   io_queue get_io_queue() const {
2331     if (cct->_conf->osd_op_queue == "debug_random") {
2332       static io_queue index_lookup[] = { io_queue::prioritized,
2333                                          io_queue::weightedpriority,
2334                                          io_queue::mclock_opclass,
2335                                          io_queue::mclock_client };
2336       srand(time(NULL));
2337       unsigned which = rand() % (sizeof(index_lookup) / sizeof(index_lookup[0]));
2338       return index_lookup[which];
2339     } else if (cct->_conf->osd_op_queue == "prioritized") {
2340       return io_queue::prioritized;
2341     } else if (cct->_conf->osd_op_queue == "mclock_opclass") {
2342       return io_queue::mclock_opclass;
2343     } else if (cct->_conf->osd_op_queue == "mclock_client") {
2344       return io_queue::mclock_client;
2345     } else {
2346       // default / catch-all is 'wpq'
2347       return io_queue::weightedpriority;
2348     }
2349   }
2350
2351   unsigned int get_io_prio_cut() const {
2352     if (cct->_conf->osd_op_queue_cut_off == "debug_random") {
2353       srand(time(NULL));
2354       return (rand() % 2 < 1) ? CEPH_MSG_PRIO_HIGH : CEPH_MSG_PRIO_LOW;
2355     } else if (cct->_conf->osd_op_queue_cut_off == "high") {
2356       return CEPH_MSG_PRIO_HIGH;
2357     } else {
2358       // default / catch-all is 'low'
2359       return CEPH_MSG_PRIO_LOW;
2360     }
2361   }
2362
2363  public:
2364   /* internal and external can point to the same messenger, they will still
2365    * be cleaned up properly*/
2366   OSD(CephContext *cct_,
2367       ObjectStore *store_,
2368       int id,
2369       Messenger *internal,
2370       Messenger *external,
2371       Messenger *hb_front_client,
2372       Messenger *hb_back_client,
2373       Messenger *hb_front_server,
2374       Messenger *hb_back_server,
2375       Messenger *osdc_messenger,
2376       MonClient *mc, const std::string &dev, const std::string &jdev);
2377   ~OSD() override;
2378
2379   // static bits
2380   static int mkfs(CephContext *cct, ObjectStore *store,
2381                   const string& dev,
2382                   uuid_d fsid, int whoami);
2383   /* remove any non-user xattrs from a map of them */
2384   void filter_xattrs(map<string, bufferptr>& attrs) {
2385     for (map<string, bufferptr>::iterator iter = attrs.begin();
2386          iter != attrs.end();
2387          ) {
2388       if (('_' != iter->first.at(0)) || (iter->first.size() == 1))
2389         attrs.erase(iter++);
2390       else ++iter;
2391     }
2392   }
2393
2394 private:
2395   int mon_cmd_maybe_osd_create(string &cmd);
2396   int update_crush_device_class();
2397   int update_crush_location();
2398
2399   static int write_meta(ObjectStore *store,
2400                         uuid_d& cluster_fsid, uuid_d& osd_fsid, int whoami);
2401
2402   void handle_pg_scrub(struct MOSDScrub *m, PG* pg);
2403   void handle_scrub(struct MOSDScrub *m);
2404   void handle_osd_ping(class MOSDPing *m);
2405
2406   int init_op_flags(OpRequestRef& op);
2407
2408   int get_num_op_shards();
2409   int get_num_op_threads();
2410
2411   float get_osd_recovery_sleep();
2412
2413 public:
2414   static int peek_meta(ObjectStore *store, string& magic,
2415                        uuid_d& cluster_fsid, uuid_d& osd_fsid, int& whoami);
2416
2417
2418   // startup/shutdown
2419   int pre_init();
2420   int init();
2421   void final_init();
2422
2423   int enable_disable_fuse(bool stop);
2424
2425   void suicide(int exitcode);
2426   int shutdown();
2427
2428   void handle_signal(int signum);
2429
2430   /// check if we can throw out op from a disconnected client
2431   static bool op_is_discardable(const MOSDOp *m);
2432
2433 public:
2434   OSDService service;
2435   friend class OSDService;
2436 };
2437
2438
2439 std::ostream& operator<<(std::ostream& out, const OSD::io_queue& q);
2440
2441
2442 //compatibility of the executable
2443 extern const CompatSet::Feature ceph_osd_feature_compat[];
2444 extern const CompatSet::Feature ceph_osd_feature_ro_compat[];
2445 extern const CompatSet::Feature ceph_osd_feature_incompat[];
2446
2447 #endif // CEPH_OSD_H