ceph/src/librbd/ImageCtx.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 #include <errno.h>
   4 #include <boost/assign/list_of.hpp>
   5 #include <stddef.h>
   6
   7 #include "common/ceph_context.h"
   8 #include "common/dout.h"
   9 #include "common/errno.h"
  10 #include "common/perf_counters.h"
  11 #include "common/WorkQueue.h"
  12 #include "common/Timer.h"
  13
  14 #include "librbd/AsyncRequest.h"
  15 #include "librbd/ExclusiveLock.h"
  16 #include "librbd/internal.h"
  17 #include "librbd/ImageCtx.h"
  18 #include "librbd/ImageState.h"
  19 #include "librbd/ImageWatcher.h"
  20 #include "librbd/Journal.h"
  21 #include "librbd/LibrbdAdminSocketHook.h"
  22 #include "librbd/ObjectMap.h"
  23 #include "librbd/Operations.h"
  24 #include "librbd/operation/ResizeRequest.h"
  25 #include "librbd/Utils.h"
  26 #include "librbd/LibrbdWriteback.h"
  27 #include "librbd/exclusive_lock/AutomaticPolicy.h"
  28 #include "librbd/exclusive_lock/StandardPolicy.h"
  29 #include "librbd/io/AioCompletion.h"
  30 #include "librbd/io/AsyncOperation.h"
  31 #include "librbd/io/ImageRequestWQ.h"
  32 #include "librbd/journal/StandardPolicy.h"
  33
  34 #include "osdc/Striper.h"
  35 #include <boost/bind.hpp>
  36
  37 #define dout_subsys ceph_subsys_rbd
  38 #undef dout_prefix
  39 #define dout_prefix *_dout << "librbd::ImageCtx: "
  40
  41 using std::map;
  42 using std::pair;
  43 using std::set;
  44 using std::string;
  45 using std::vector;
  46
  47 using ceph::bufferlist;
  48 using librados::snap_t;
  49 using librados::IoCtx;
  50
  51 namespace librbd {
  52
  53 namespace {
  54
  55 class ThreadPoolSingleton : public ThreadPool {
  56 public:
  57   ContextWQ *op_work_queue;
  58
  59   explicit ThreadPoolSingleton(CephContext *cct)
  60     : ThreadPool(cct, "librbd::thread_pool", "tp_librbd", 1,
  61                  "rbd_op_threads"),
  62       op_work_queue(new ContextWQ("librbd::op_work_queue",
  63                                   cct->_conf->rbd_op_thread_timeout,
  64                                   this)) {
  65     start();
  66   }
  67   ~ThreadPoolSingleton() override {
  68     op_work_queue->drain();
  69     delete op_work_queue;
  70
  71     stop();
  72   }
  73 };
  74
  75 class SafeTimerSingleton : public SafeTimer {
  76 public:
  77   Mutex lock;
  78
  79   explicit SafeTimerSingleton(CephContext *cct)
  80       : SafeTimer(cct, lock, true),
  81         lock("librbd::Journal::SafeTimerSingleton::lock") {
  82     init();
  83   }
  84   ~SafeTimerSingleton() {
  85     Mutex::Locker locker(lock);
  86     shutdown();
  87   }
  88 };
  89
  90 struct C_FlushCache : public Context {
  91   ImageCtx *image_ctx;
  92   Context *on_safe;
  93
  94   C_FlushCache(ImageCtx *_image_ctx, Context *_on_safe)
  95     : image_ctx(_image_ctx), on_safe(_on_safe) {
  96   }
  97   void finish(int r) override {
  98     // successful cache flush indicates all IO is now safe
  99     image_ctx->flush_cache(on_safe);
 100   }
 101 };
 102
 103 struct C_ShutDownCache : public Context {
 104   ImageCtx *image_ctx;
 105   Context *on_finish;
 106
 107   C_ShutDownCache(ImageCtx *_image_ctx, Context *_on_finish)
 108     : image_ctx(_image_ctx), on_finish(_on_finish) {
 109   }
 110   void finish(int r) override {
 111     image_ctx->object_cacher->stop();
 112     on_finish->complete(r);
 113   }
 114 };
 115
 116 struct C_InvalidateCache : public Context {
 117   ImageCtx *image_ctx;
 118   bool purge_on_error;
 119   bool reentrant_safe;
 120   Context *on_finish;
 121
 122   C_InvalidateCache(ImageCtx *_image_ctx, bool _purge_on_error,
 123                     bool _reentrant_safe, Context *_on_finish)
 124     : image_ctx(_image_ctx), purge_on_error(_purge_on_error),
 125       reentrant_safe(_reentrant_safe), on_finish(_on_finish) {
 126   }
 127   void finish(int r) override {
 128     assert(image_ctx->cache_lock.is_locked());
 129     CephContext *cct = image_ctx->cct;
 130
 131     if (r == -EBLACKLISTED) {
 132       lderr(cct) << "Blacklisted during flush!  Purging cache..." << dendl;
 133       image_ctx->object_cacher->purge_set(image_ctx->object_set);
 134     } else if (r != 0 && purge_on_error) {
 135       lderr(cct) << "invalidate cache encountered error "
 136                  << cpp_strerror(r) << " !Purging cache..." << dendl;
 137       image_ctx->object_cacher->purge_set(image_ctx->object_set);
 138     } else if (r != 0) {
 139       lderr(cct) << "flush_cache returned " << r << dendl;
 140     }
 141
 142     loff_t unclean = image_ctx->object_cacher->release_set(
 143       image_ctx->object_set);
 144     if (unclean == 0) {
 145       r = 0;
 146     } else {
 147       lderr(cct) << "could not release all objects from cache: "
 148                  << unclean << " bytes remain" << dendl;
 149       if (r == 0) {
 150         r = -EBUSY;
 151       }
 152     }
 153
 154     if (reentrant_safe) {
 155       on_finish->complete(r);
 156     } else {
 157       image_ctx->op_work_queue->queue(on_finish, r);
 158     }
 159   }
 160
 161 };
 162
 163 } // anonymous namespace
 164
 165   const string ImageCtx::METADATA_CONF_PREFIX = "conf_";
 166
 167   ImageCtx::ImageCtx(const string &image_name, const string &image_id,
 168                      const char *snap, IoCtx& p, bool ro)
 169     : cct((CephContext*)p.cct()),
 170       perfcounter(NULL),
 171       snap_id(CEPH_NOSNAP),
 172       snap_exists(true),
 173       read_only(ro),
 174       flush_encountered(false),
 175       exclusive_locked(false),
 176       name(image_name),
 177       image_watcher(NULL),
 178       journal(NULL),
 179       owner_lock(util::unique_lock_name("librbd::ImageCtx::owner_lock", this)),
 180       md_lock(util::unique_lock_name("librbd::ImageCtx::md_lock", this)),
 181       cache_lock(util::unique_lock_name("librbd::ImageCtx::cache_lock", this)),
 182       snap_lock(util::unique_lock_name("librbd::ImageCtx::snap_lock", this)),
 183       parent_lock(util::unique_lock_name("librbd::ImageCtx::parent_lock", this)),
 184       object_map_lock(util::unique_lock_name("librbd::ImageCtx::object_map_lock", this)),
 185       async_ops_lock(util::unique_lock_name("librbd::ImageCtx::async_ops_lock", this)),
 186       copyup_list_lock(util::unique_lock_name("librbd::ImageCtx::copyup_list_lock", this)),
 187       completed_reqs_lock(util::unique_lock_name("librbd::ImageCtx::completed_reqs_lock", this)),
 188       extra_read_flags(0),
 189       old_format(true),
 190       order(0), size(0), features(0),
 191       format_string(NULL),
 192       id(image_id), parent(NULL),
 193       stripe_unit(0), stripe_count(0), flags(0),
 194       object_cacher(NULL), writeback_handler(NULL), object_set(NULL),
 195       readahead(),
 196       total_bytes_read(0),
 197       state(new ImageState<>(this)),
 198       operations(new Operations<>(*this)),
 199       exclusive_lock(nullptr), object_map(nullptr),
 200       io_work_queue(nullptr), op_work_queue(nullptr),
 201       asok_hook(nullptr),
 202       trace_endpoint("librbd")
 203   {
 204     md_ctx.dup(p);
 205     data_ctx.dup(p);
 206     if (snap)
 207       snap_name = snap;
 208
 209     memset(&header, 0, sizeof(header));
 210
 211     ThreadPool *thread_pool;
 212     get_thread_pool_instance(cct, &thread_pool, &op_work_queue);
 213     io_work_queue = new io::ImageRequestWQ<>(
 214       this, "librbd::io_work_queue", cct->_conf->rbd_op_thread_timeout,
 215       thread_pool);
 216
 217     if (cct->_conf->rbd_auto_exclusive_lock_until_manual_request) {
 218       exclusive_lock_policy = new exclusive_lock::AutomaticPolicy(this);
 219     } else {
 220       exclusive_lock_policy = new exclusive_lock::StandardPolicy(this);
 221     }
 222     journal_policy = new journal::StandardPolicy<ImageCtx>(this);
 223   }
 224
 225   ImageCtx::~ImageCtx() {
 226     assert(image_watcher == NULL);
 227     assert(exclusive_lock == NULL);
 228     assert(object_map == NULL);
 229     assert(journal == NULL);
 230     assert(asok_hook == NULL);
 231
 232     if (perfcounter) {
 233       perf_stop();
 234     }
 235     if (object_cacher) {
 236       delete object_cacher;
 237       object_cacher = NULL;
 238     }
 239     if (writeback_handler) {
 240       delete writeback_handler;
 241       writeback_handler = NULL;
 242     }
 243     if (object_set) {
 244       delete object_set;
 245       object_set = NULL;
 246     }
 247     delete[] format_string;
 248
 249     md_ctx.aio_flush();
 250     data_ctx.aio_flush();
 251     io_work_queue->drain();
 252
 253     delete journal_policy;
 254     delete exclusive_lock_policy;
 255     delete io_work_queue;
 256     delete operations;
 257     delete state;
 258   }
 259
 260   void ImageCtx::init() {
 261     assert(!header_oid.empty());
 262     assert(old_format || !id.empty());
 263
 264     asok_hook = new LibrbdAdminSocketHook(this);
 265
 266     string pname = string("librbd-") + id + string("-") +
 267       data_ctx.get_pool_name() + string("-") + name;
 268     if (!snap_name.empty()) {
 269       pname += "-";
 270       pname += snap_name;
 271     }
 272
 273     trace_endpoint.copy_name(pname);
 274     perf_start(pname);
 275
 276     if (cache) {
 277       Mutex::Locker l(cache_lock);
 278       ldout(cct, 20) << "enabling caching..." << dendl;
 279       writeback_handler = new LibrbdWriteback(this, cache_lock);
 280
 281       uint64_t init_max_dirty = cache_max_dirty;
 282       if (cache_writethrough_until_flush)
 283         init_max_dirty = 0;
 284       ldout(cct, 20) << "Initial cache settings:"
 285                      << " size=" << cache_size
 286                      << " num_objects=" << 10
 287                      << " max_dirty=" << init_max_dirty
 288                      << " target_dirty=" << cache_target_dirty
 289                      << " max_dirty_age="
 290                      << cache_max_dirty_age << dendl;
 291
 292       object_cacher = new ObjectCacher(cct, pname, *writeback_handler, cache_lock,
 293                                        NULL, NULL,
 294                                        cache_size,
 295                                        10,  /* reset this in init */
 296                                        init_max_dirty,
 297                                        cache_target_dirty,
 298                                        cache_max_dirty_age,
 299                                        cache_block_writes_upfront);
 300
 301       // size object cache appropriately
 302       uint64_t obj = cache_max_dirty_object;
 303       if (!obj) {
 304         obj = MIN(2000, MAX(10, cache_size / 100 / sizeof(ObjectCacher::Object)));
 305       }
 306       ldout(cct, 10) << " cache bytes " << cache_size
 307         << " -> about " << obj << " objects" << dendl;
 308       object_cacher->set_max_objects(obj);
 309
 310       object_set = new ObjectCacher::ObjectSet(NULL, data_ctx.get_id(), 0);
 311       object_set->return_enoent = true;
 312       object_cacher->start();
 313     }
 314
 315     readahead.set_trigger_requests(readahead_trigger_requests);
 316     readahead.set_max_readahead_size(readahead_max_bytes);
 317   }
 318
 319   void ImageCtx::shutdown() {
 320     delete image_watcher;
 321     image_watcher = nullptr;
 322
 323     delete asok_hook;
 324     asok_hook = nullptr;
 325   }
 326
 327   void ImageCtx::init_layout()
 328   {
 329     if (stripe_unit == 0 || stripe_count == 0) {
 330       stripe_unit = 1ull << order;
 331       stripe_count = 1;
 332     }
 333
 334     vector<uint64_t> alignments;
 335     alignments.push_back(stripe_count << order); // object set (in file striping terminology)
 336     alignments.push_back(stripe_unit * stripe_count); // stripe
 337     alignments.push_back(stripe_unit); // stripe unit
 338     readahead.set_alignments(alignments);
 339
 340     layout = file_layout_t();
 341     layout.stripe_unit = stripe_unit;
 342     layout.stripe_count = stripe_count;
 343     layout.object_size = 1ull << order;
 344     layout.pool_id = data_ctx.get_id();  // FIXME: pool id overflow?
 345
 346     delete[] format_string;
 347     size_t len = object_prefix.length() + 16;
 348     format_string = new char[len];
 349     if (old_format) {
 350       snprintf(format_string, len, "%s.%%012llx", object_prefix.c_str());
 351     } else {
 352       snprintf(format_string, len, "%s.%%016llx", object_prefix.c_str());
 353     }
 354
 355     ldout(cct, 10) << "init_layout stripe_unit " << stripe_unit
 356                    << " stripe_count " << stripe_count
 357                    << " object_size " << layout.object_size
 358                    << " prefix " << object_prefix
 359                    << " format " << format_string
 360                    << dendl;
 361   }
 362
 363   void ImageCtx::perf_start(string name) {
 364     PerfCountersBuilder plb(cct, name, l_librbd_first, l_librbd_last);
 365
 366     plb.add_u64_counter(l_librbd_rd, "rd", "Reads");
 367     plb.add_u64_counter(l_librbd_rd_bytes, "rd_bytes", "Data size in reads");
 368     plb.add_time_avg(l_librbd_rd_latency, "rd_latency", "Latency of reads");
 369     plb.add_u64_counter(l_librbd_wr, "wr", "Writes");
 370     plb.add_u64_counter(l_librbd_wr_bytes, "wr_bytes", "Written data");
 371     plb.add_time_avg(l_librbd_wr_latency, "wr_latency", "Write latency");
 372     plb.add_u64_counter(l_librbd_discard, "discard", "Discards");
 373     plb.add_u64_counter(l_librbd_discard_bytes, "discard_bytes", "Discarded data");
 374     plb.add_time_avg(l_librbd_discard_latency, "discard_latency", "Discard latency");
 375     plb.add_u64_counter(l_librbd_flush, "flush", "Flushes");
 376     plb.add_u64_counter(l_librbd_aio_flush, "aio_flush", "Async flushes");
 377     plb.add_time_avg(l_librbd_aio_flush_latency, "aio_flush_latency", "Latency of async flushes");
 378     plb.add_u64_counter(l_librbd_ws, "ws", "WriteSames");
 379     plb.add_u64_counter(l_librbd_ws_bytes, "ws_bytes", "WriteSame data");
 380     plb.add_time_avg(l_librbd_ws_latency, "ws_latency", "WriteSame latency");
 381     plb.add_u64_counter(l_librbd_cmp, "cmp", "CompareAndWrites");
 382     plb.add_u64_counter(l_librbd_cmp_bytes, "cmp_bytes", "Data size in cmps");
 383     plb.add_time_avg(l_librbd_cmp_latency, "cmp_latency", "Latency of cmps");
 384     plb.add_u64_counter(l_librbd_snap_create, "snap_create", "Snap creations");
 385     plb.add_u64_counter(l_librbd_snap_remove, "snap_remove", "Snap removals");
 386     plb.add_u64_counter(l_librbd_snap_rollback, "snap_rollback", "Snap rollbacks");
 387     plb.add_u64_counter(l_librbd_snap_rename, "snap_rename", "Snap rename");
 388     plb.add_u64_counter(l_librbd_notify, "notify", "Updated header notifications");
 389     plb.add_u64_counter(l_librbd_resize, "resize", "Resizes");
 390     plb.add_u64_counter(l_librbd_readahead, "readahead", "Read ahead");
 391     plb.add_u64_counter(l_librbd_readahead_bytes, "readahead_bytes", "Data size in read ahead");
 392     plb.add_u64_counter(l_librbd_invalidate_cache, "invalidate_cache", "Cache invalidates");
 393
 394     perfcounter = plb.create_perf_counters();
 395     cct->get_perfcounters_collection()->add(perfcounter);
 396   }
 397
 398   void ImageCtx::perf_stop() {
 399     assert(perfcounter);
 400     cct->get_perfcounters_collection()->remove(perfcounter);
 401     delete perfcounter;
 402   }
 403
 404   void ImageCtx::set_read_flag(unsigned flag) {
 405     extra_read_flags |= flag;
 406   }
 407
 408   int ImageCtx::get_read_flags(snap_t snap_id) {
 409     int flags = librados::OPERATION_NOFLAG | extra_read_flags;
 410     if (snap_id == LIBRADOS_SNAP_HEAD)
 411       return flags;
 412
 413     if (balance_snap_reads)
 414       flags |= librados::OPERATION_BALANCE_READS;
 415     else if (localize_snap_reads)
 416       flags |= librados::OPERATION_LOCALIZE_READS;
 417     return flags;
 418   }
 419
 420   int ImageCtx::snap_set(cls::rbd::SnapshotNamespace in_snap_namespace,
 421                          string in_snap_name)
 422   {
 423     assert(snap_lock.is_wlocked());
 424     snap_t in_snap_id = get_snap_id(in_snap_namespace, in_snap_name);
 425     if (in_snap_id != CEPH_NOSNAP) {
 426       snap_id = in_snap_id;
 427       snap_namespace = in_snap_namespace;
 428       snap_name = in_snap_name;
 429       snap_exists = true;
 430       data_ctx.snap_set_read(snap_id);
 431       return 0;
 432     }
 433     return -ENOENT;
 434   }
 435
 436   void ImageCtx::snap_unset()
 437   {
 438     assert(snap_lock.is_wlocked());
 439     snap_id = CEPH_NOSNAP;
 440     snap_namespace = {};
 441     snap_name = "";
 442     snap_exists = true;
 443     data_ctx.snap_set_read(snap_id);
 444   }
 445
 446   snap_t ImageCtx::get_snap_id(cls::rbd::SnapshotNamespace in_snap_namespace,
 447                                string in_snap_name) const
 448   {
 449     assert(snap_lock.is_locked());
 450     auto it = snap_ids.find({in_snap_namespace, in_snap_name});
 451     if (it != snap_ids.end())
 452       return it->second;
 453     return CEPH_NOSNAP;
 454   }
 455
 456   const SnapInfo* ImageCtx::get_snap_info(snap_t in_snap_id) const
 457   {
 458     assert(snap_lock.is_locked());
 459     map<snap_t, SnapInfo>::const_iterator it =
 460       snap_info.find(in_snap_id);
 461     if (it != snap_info.end())
 462       return &it->second;
 463     return NULL;
 464   }
 465
 466   int ImageCtx::get_snap_name(snap_t in_snap_id,
 467                               string *out_snap_name) const
 468   {
 469     assert(snap_lock.is_locked());
 470     const SnapInfo *info = get_snap_info(in_snap_id);
 471     if (info) {
 472       *out_snap_name = info->name;
 473       return 0;
 474     }
 475     return -ENOENT;
 476   }
 477
 478   int ImageCtx::get_snap_namespace(snap_t in_snap_id,
 479                                    cls::rbd::SnapshotNamespace *out_snap_namespace) const
 480   {
 481     assert(snap_lock.is_locked());
 482     const SnapInfo *info = get_snap_info(in_snap_id);
 483     if (info) {
 484       *out_snap_namespace = info->snap_namespace;
 485       return 0;
 486     }
 487     return -ENOENT;
 488   }
 489
 490   int ImageCtx::get_parent_spec(snap_t in_snap_id,
 491                                 ParentSpec *out_pspec) const
 492   {
 493     const SnapInfo *info = get_snap_info(in_snap_id);
 494     if (info) {
 495       *out_pspec = info->parent.spec;
 496       return 0;
 497     }
 498     return -ENOENT;
 499   }
 500
 501   uint64_t ImageCtx::get_current_size() const
 502   {
 503     assert(snap_lock.is_locked());
 504     return size;
 505   }
 506
 507   uint64_t ImageCtx::get_object_size() const
 508   {
 509     return 1ull << order;
 510   }
 511
 512   string ImageCtx::get_object_name(uint64_t num) const {
 513     char buf[object_prefix.length() + 32];
 514     snprintf(buf, sizeof(buf), format_string, num);
 515     return string(buf);
 516   }
 517
 518   uint64_t ImageCtx::get_stripe_unit() const
 519   {
 520     return stripe_unit;
 521   }
 522
 523   uint64_t ImageCtx::get_stripe_count() const
 524   {
 525     return stripe_count;
 526   }
 527
 528   uint64_t ImageCtx::get_stripe_period() const
 529   {
 530     return stripe_count * (1ull << order);
 531   }
 532
 533   utime_t ImageCtx::get_create_timestamp() const
 534   {
 535     return create_timestamp;
 536   }
 537
 538   int ImageCtx::is_snap_protected(snap_t in_snap_id,
 539                                   bool *is_protected) const
 540   {
 541     assert(snap_lock.is_locked());
 542     const SnapInfo *info = get_snap_info(in_snap_id);
 543     if (info) {
 544       *is_protected =
 545         (info->protection_status == RBD_PROTECTION_STATUS_PROTECTED);
 546       return 0;
 547     }
 548     return -ENOENT;
 549   }
 550
 551   int ImageCtx::is_snap_unprotected(snap_t in_snap_id,
 552                                     bool *is_unprotected) const
 553   {
 554     assert(snap_lock.is_locked());
 555     const SnapInfo *info = get_snap_info(in_snap_id);
 556     if (info) {
 557       *is_unprotected =
 558         (info->protection_status == RBD_PROTECTION_STATUS_UNPROTECTED);
 559       return 0;
 560     }
 561     return -ENOENT;
 562   }
 563
 564   void ImageCtx::add_snap(cls::rbd::SnapshotNamespace in_snap_namespace,
 565                           string in_snap_name,
 566                           snap_t id, uint64_t in_size,
 567                           const ParentInfo &parent, uint8_t protection_status,
 568                           uint64_t flags, utime_t timestamp)
 569   {
 570     assert(snap_lock.is_wlocked());
 571     snaps.push_back(id);
 572     SnapInfo info(in_snap_name, in_snap_namespace,
 573                   in_size, parent, protection_status, flags, timestamp);
 574     snap_info.insert({id, info});
 575     snap_ids.insert({{in_snap_namespace, in_snap_name}, id});
 576   }
 577
 578   void ImageCtx::rm_snap(cls::rbd::SnapshotNamespace in_snap_namespace,
 579                          string in_snap_name,
 580                          snap_t id)
 581   {
 582     assert(snap_lock.is_wlocked());
 583     snaps.erase(std::remove(snaps.begin(), snaps.end(), id), snaps.end());
 584     snap_info.erase(id);
 585     snap_ids.erase({in_snap_namespace, in_snap_name});
 586   }
 587
 588   uint64_t ImageCtx::get_image_size(snap_t in_snap_id) const
 589   {
 590     assert(snap_lock.is_locked());
 591     if (in_snap_id == CEPH_NOSNAP) {
 592       if (!resize_reqs.empty() &&
 593           resize_reqs.front()->shrinking()) {
 594         return resize_reqs.front()->get_image_size();
 595       }
 596       return size;
 597     }
 598
 599     const SnapInfo *info = get_snap_info(in_snap_id);
 600     if (info) {
 601       return info->size;
 602     }
 603     return 0;
 604   }
 605
 606   uint64_t ImageCtx::get_object_count(snap_t in_snap_id) const {
 607     assert(snap_lock.is_locked());
 608     uint64_t image_size = get_image_size(in_snap_id);
 609     return Striper::get_num_objects(layout, image_size);
 610   }
 611
 612   bool ImageCtx::test_features(uint64_t features) const
 613   {
 614     RWLock::RLocker l(snap_lock);
 615     return test_features(features, snap_lock);
 616   }
 617
 618   bool ImageCtx::test_features(uint64_t in_features,
 619                                const RWLock &in_snap_lock) const
 620   {
 621     assert(snap_lock.is_locked());
 622     return ((features & in_features) == in_features);
 623   }
 624
 625   int ImageCtx::get_flags(librados::snap_t _snap_id, uint64_t *_flags) const
 626   {
 627     assert(snap_lock.is_locked());
 628     if (_snap_id == CEPH_NOSNAP) {
 629       *_flags = flags;
 630       return 0;
 631     }
 632     const SnapInfo *info = get_snap_info(_snap_id);
 633     if (info) {
 634       *_flags = info->flags;
 635       return 0;
 636     }
 637     return -ENOENT;
 638   }
 639
 640   int ImageCtx::test_flags(uint64_t flags, bool *flags_set) const
 641   {
 642     RWLock::RLocker l(snap_lock);
 643     return test_flags(flags, snap_lock, flags_set);
 644   }
 645
 646   int ImageCtx::test_flags(uint64_t flags, const RWLock &in_snap_lock,
 647                            bool *flags_set) const
 648   {
 649     assert(snap_lock.is_locked());
 650     uint64_t snap_flags;
 651     int r = get_flags(snap_id, &snap_flags);
 652     if (r < 0) {
 653       return r;
 654     }
 655     *flags_set = ((snap_flags & flags) == flags);
 656     return 0;
 657   }
 658
 659   int ImageCtx::update_flags(snap_t in_snap_id, uint64_t flag, bool enabled)
 660   {
 661     assert(snap_lock.is_wlocked());
 662     uint64_t *_flags;
 663     if (in_snap_id == CEPH_NOSNAP) {
 664       _flags = &flags;
 665     } else {
 666       map<snap_t, SnapInfo>::iterator it = snap_info.find(in_snap_id);
 667       if (it == snap_info.end()) {
 668         return -ENOENT;
 669       }
 670       _flags = &it->second.flags;
 671     }
 672
 673     if (enabled) {
 674       (*_flags) |= flag;
 675     } else {
 676       (*_flags) &= ~flag;
 677     }
 678     return 0;
 679   }
 680
 681   const ParentInfo* ImageCtx::get_parent_info(snap_t in_snap_id) const
 682   {
 683     assert(snap_lock.is_locked());
 684     assert(parent_lock.is_locked());
 685     if (in_snap_id == CEPH_NOSNAP)
 686       return &parent_md;
 687     const SnapInfo *info = get_snap_info(in_snap_id);
 688     if (info)
 689       return &info->parent;
 690     return NULL;
 691   }
 692
 693   int64_t ImageCtx::get_parent_pool_id(snap_t in_snap_id) const
 694   {
 695     const ParentInfo *info = get_parent_info(in_snap_id);
 696     if (info)
 697       return info->spec.pool_id;
 698     return -1;
 699   }
 700
 701   string ImageCtx::get_parent_image_id(snap_t in_snap_id) const
 702   {
 703     const ParentInfo *info = get_parent_info(in_snap_id);
 704     if (info)
 705       return info->spec.image_id;
 706     return "";
 707   }
 708
 709   uint64_t ImageCtx::get_parent_snap_id(snap_t in_snap_id) const
 710   {
 711     const ParentInfo *info = get_parent_info(in_snap_id);
 712     if (info)
 713       return info->spec.snap_id;
 714     return CEPH_NOSNAP;
 715   }
 716
 717   int ImageCtx::get_parent_overlap(snap_t in_snap_id, uint64_t *overlap) const
 718   {
 719     assert(snap_lock.is_locked());
 720     const ParentInfo *info = get_parent_info(in_snap_id);
 721     if (info) {
 722       *overlap = info->overlap;
 723       return 0;
 724     }
 725     return -ENOENT;
 726   }
 727
 728   void ImageCtx::aio_read_from_cache(object_t o, uint64_t object_no,
 729                                      bufferlist *bl, size_t len,
 730                                      uint64_t off, Context *onfinish,
 731                                      int fadvise_flags, ZTracer::Trace *trace) {
 732     snap_lock.get_read();
 733     ObjectCacher::OSDRead *rd = object_cacher->prepare_read(snap_id, bl, fadvise_flags);
 734     snap_lock.put_read();
 735     ObjectExtent extent(o, object_no, off, len, 0);
 736     extent.oloc.pool = data_ctx.get_id();
 737     extent.buffer_extents.push_back(make_pair(0, len));
 738     rd->extents.push_back(extent);
 739     cache_lock.Lock();
 740     int r = object_cacher->readx(rd, object_set, onfinish, trace);
 741     cache_lock.Unlock();
 742     if (r != 0)
 743       onfinish->complete(r);
 744   }
 745
 746   void ImageCtx::write_to_cache(object_t o, const bufferlist& bl, size_t len,
 747                                 uint64_t off, Context *onfinish,
 748                                 int fadvise_flags, uint64_t journal_tid,
 749                                 ZTracer::Trace *trace) {
 750     snap_lock.get_read();
 751     ObjectCacher::OSDWrite *wr = object_cacher->prepare_write(
 752       snapc, bl, ceph::real_time::min(), fadvise_flags, journal_tid);
 753     snap_lock.put_read();
 754     ObjectExtent extent(o, 0, off, len, 0);
 755     extent.oloc.pool = data_ctx.get_id();
 756     // XXX: nspace is always default, io_ctx_impl field private
 757     //extent.oloc.nspace = data_ctx.io_ctx_impl->oloc.nspace;
 758     extent.buffer_extents.push_back(make_pair(0, len));
 759     wr->extents.push_back(extent);
 760     {
 761       Mutex::Locker l(cache_lock);
 762       object_cacher->writex(wr, object_set, onfinish, trace);
 763     }
 764   }
 765
 766   void ImageCtx::user_flushed() {
 767     if (object_cacher && cache_writethrough_until_flush) {
 768       md_lock.get_read();
 769       bool flushed_before = flush_encountered;
 770       md_lock.put_read();
 771
 772       uint64_t max_dirty = cache_max_dirty;
 773       if (!flushed_before && max_dirty > 0) {
 774         md_lock.get_write();
 775         flush_encountered = true;
 776         md_lock.put_write();
 777
 778         ldout(cct, 10) << "saw first user flush, enabling writeback" << dendl;
 779         Mutex::Locker l(cache_lock);
 780         object_cacher->set_max_dirty(max_dirty);
 781       }
 782     }
 783   }
 784
 785   void ImageCtx::flush_cache(Context *onfinish) {
 786     cache_lock.Lock();
 787     object_cacher->flush_set(object_set, onfinish);
 788     cache_lock.Unlock();
 789   }
 790
 791   void ImageCtx::shut_down_cache(Context *on_finish) {
 792     if (object_cacher == NULL) {
 793       on_finish->complete(0);
 794       return;
 795     }
 796
 797     cache_lock.Lock();
 798     object_cacher->release_set(object_set);
 799     cache_lock.Unlock();
 800
 801     C_ShutDownCache *shut_down = new C_ShutDownCache(this, on_finish);
 802     flush_cache(new C_InvalidateCache(this, true, false, shut_down));
 803   }
 804
 805   int ImageCtx::invalidate_cache(bool purge_on_error) {
 806     flush_async_operations();
 807     if (object_cacher == NULL) {
 808       return 0;
 809     }
 810
 811     cache_lock.Lock();
 812     object_cacher->release_set(object_set);
 813     cache_lock.Unlock();
 814
 815     C_SaferCond ctx;
 816     flush_cache(new C_InvalidateCache(this, purge_on_error, true, &ctx));
 817
 818     int result = ctx.wait();
 819     return result;
 820   }
 821
 822   void ImageCtx::invalidate_cache(bool purge_on_error, Context *on_finish) {
 823     if (object_cacher == NULL) {
 824       op_work_queue->queue(on_finish, 0);
 825       return;
 826     }
 827
 828     cache_lock.Lock();
 829     object_cacher->release_set(object_set);
 830     cache_lock.Unlock();
 831
 832     flush_cache(new C_InvalidateCache(this, purge_on_error, false, on_finish));
 833   }
 834
 835   void ImageCtx::clear_nonexistence_cache() {
 836     assert(cache_lock.is_locked());
 837     if (!object_cacher)
 838       return;
 839     object_cacher->clear_nonexistence(object_set);
 840   }
 841
 842   bool ImageCtx::is_cache_empty() {
 843     Mutex::Locker locker(cache_lock);
 844     return object_cacher->set_is_empty(object_set);
 845   }
 846
 847   void ImageCtx::register_watch(Context *on_finish) {
 848     assert(image_watcher == NULL);
 849     image_watcher = new ImageWatcher<>(*this);
 850     image_watcher->register_watch(on_finish);
 851   }
 852
 853   uint64_t ImageCtx::prune_parent_extents(vector<pair<uint64_t,uint64_t> >& objectx,
 854                                           uint64_t overlap)
 855   {
 856     // drop extents completely beyond the overlap
 857     while (!objectx.empty() && objectx.back().first >= overlap)
 858       objectx.pop_back();
 859
 860     // trim final overlapping extent
 861     if (!objectx.empty() && objectx.back().first + objectx.back().second > overlap)
 862       objectx.back().second = overlap - objectx.back().first;
 863
 864     uint64_t len = 0;
 865     for (vector<pair<uint64_t,uint64_t> >::iterator p = objectx.begin();
 866          p != objectx.end();
 867          ++p)
 868       len += p->second;
 869     ldout(cct, 10) << "prune_parent_extents image overlap " << overlap
 870                    << ", object overlap " << len
 871                    << " from image extents " << objectx << dendl;
 872     return len;
 873   }
 874
 875   void ImageCtx::flush_async_operations() {
 876     C_SaferCond ctx;
 877     flush_async_operations(&ctx);
 878     ctx.wait();
 879   }
 880
 881   void ImageCtx::flush_async_operations(Context *on_finish) {
 882     {
 883       Mutex::Locker l(async_ops_lock);
 884       if (!async_ops.empty()) {
 885         ldout(cct, 20) << "flush async operations: " << on_finish << " "
 886                        << "count=" << async_ops.size() << dendl;
 887         async_ops.front()->add_flush_context(on_finish);
 888         return;
 889       }
 890     }
 891     on_finish->complete(0);
 892   }
 893
 894   int ImageCtx::flush() {
 895     C_SaferCond cond_ctx;
 896     flush(&cond_ctx);
 897     return cond_ctx.wait();
 898   }
 899
 900   void ImageCtx::flush(Context *on_safe) {
 901     // ensure no locks are held when flush is complete
 902     on_safe = util::create_async_context_callback(*this, on_safe);
 903
 904     if (object_cacher != NULL) {
 905       // flush cache after completing all in-flight AIO ops
 906       on_safe = new C_FlushCache(this, on_safe);
 907     }
 908     flush_async_operations(on_safe);
 909   }
 910
 911   void ImageCtx::cancel_async_requests() {
 912     C_SaferCond ctx;
 913     cancel_async_requests(&ctx);
 914     ctx.wait();
 915   }
 916
 917   void ImageCtx::cancel_async_requests(Context *on_finish) {
 918     {
 919       Mutex::Locker async_ops_locker(async_ops_lock);
 920       if (!async_requests.empty()) {
 921         ldout(cct, 10) << "canceling async requests: count="
 922                        << async_requests.size() << dendl;
 923         for (auto req : async_requests) {
 924           ldout(cct, 10) << "canceling async request: " << req << dendl;
 925           req->cancel();
 926         }
 927         async_requests_waiters.push_back(on_finish);
 928         return;
 929       }
 930     }
 931
 932     on_finish->complete(0);
 933   }
 934
 935   void ImageCtx::clear_pending_completions() {
 936     Mutex::Locker l(completed_reqs_lock);
 937     ldout(cct, 10) << "clear pending AioCompletion: count="
 938                    << completed_reqs.size() << dendl;
 939     completed_reqs.clear();
 940   }
 941
 942   bool ImageCtx::_filter_metadata_confs(const string &prefix,
 943                                         map<string, bool> &configs,
 944                                         const map<string, bufferlist> &pairs,
 945                                         map<string, bufferlist> *res) {
 946     size_t conf_prefix_len = prefix.size();
 947
 948     for (auto it : pairs) {
 949       if (it.first.compare(0, MIN(conf_prefix_len, it.first.size()), prefix) > 0)
 950         return false;
 951
 952       if (it.first.size() <= conf_prefix_len)
 953         continue;
 954
 955       string key = it.first.substr(conf_prefix_len, it.first.size() - conf_prefix_len);
 956       auto cit = configs.find(key);
 957       if (cit != configs.end()) {
 958         cit->second = true;
 959         res->insert(make_pair(key, it.second));
 960       }
 961     }
 962     return true;
 963   }
 964
 965   void ImageCtx::apply_metadata(const std::map<std::string, bufferlist> &meta) {
 966     ldout(cct, 20) << __func__ << dendl;
 967     std::map<string, bool> configs = boost::assign::map_list_of(
 968         "rbd_non_blocking_aio", false)(
 969         "rbd_cache", false)(
 970         "rbd_cache_writethrough_until_flush", false)(
 971         "rbd_cache_size", false)(
 972         "rbd_cache_max_dirty", false)(
 973         "rbd_cache_target_dirty", false)(
 974         "rbd_cache_max_dirty_age", false)(
 975         "rbd_cache_max_dirty_object", false)(
 976         "rbd_cache_block_writes_upfront", false)(
 977         "rbd_concurrent_management_ops", false)(
 978         "rbd_balance_snap_reads", false)(
 979         "rbd_localize_snap_reads", false)(
 980         "rbd_balance_parent_reads", false)(
 981         "rbd_localize_parent_reads", false)(
 982         "rbd_readahead_trigger_requests", false)(
 983         "rbd_readahead_max_bytes", false)(
 984         "rbd_readahead_disable_after_bytes", false)(
 985         "rbd_clone_copy_on_read", false)(
 986         "rbd_blacklist_on_break_lock", false)(
 987         "rbd_blacklist_expire_seconds", false)(
 988         "rbd_request_timed_out_seconds", false)(
 989         "rbd_journal_order", false)(
 990         "rbd_journal_splay_width", false)(
 991         "rbd_journal_commit_age", false)(
 992         "rbd_journal_object_flush_interval", false)(
 993         "rbd_journal_object_flush_bytes", false)(
 994         "rbd_journal_object_flush_age", false)(
 995         "rbd_journal_pool", false)(
 996         "rbd_journal_max_payload_bytes", false)(
 997         "rbd_journal_max_concurrent_object_sets", false)(
 998         "rbd_mirroring_resync_after_disconnect", false)(
 999         "rbd_mirroring_replay_delay", false)(
1000         "rbd_skip_partial_discard", false);
1001
1002     md_config_t local_config_t;
1003     std::map<std::string, bufferlist> res;
1004
1005     _filter_metadata_confs(METADATA_CONF_PREFIX, configs, meta, &res);
1006     for (auto it : res) {
1007       std::string val(it.second.c_str(), it.second.length());
1008       int j = local_config_t.set_val(it.first.c_str(), val);
1009       if (j < 0) {
1010         lderr(cct) << __func__ << " failed to set config " << it.first
1011                    << " with value " << it.second.c_str() << ": " << j
1012                    << dendl;
1013       }
1014     }
1015
1016 #define ASSIGN_OPTION(config)                                                  \
1017     do {                                                                       \
1018       string key = "rbd_";                                                     \
1019       key = key + #config;                                                     \
1020       if (configs[key])                                                        \
1021         config = local_config_t.rbd_##config;                                  \
1022       else                                                                     \
1023         config = cct->_conf->rbd_##config;                                     \
1024     } while (0);
1025
1026     ASSIGN_OPTION(non_blocking_aio);
1027     ASSIGN_OPTION(cache);
1028     ASSIGN_OPTION(cache_writethrough_until_flush);
1029     ASSIGN_OPTION(cache_size);
1030     ASSIGN_OPTION(cache_max_dirty);
1031     ASSIGN_OPTION(cache_target_dirty);
1032     ASSIGN_OPTION(cache_max_dirty_age);
1033     ASSIGN_OPTION(cache_max_dirty_object);
1034     ASSIGN_OPTION(cache_block_writes_upfront);
1035     ASSIGN_OPTION(concurrent_management_ops);
1036     ASSIGN_OPTION(balance_snap_reads);
1037     ASSIGN_OPTION(localize_snap_reads);
1038     ASSIGN_OPTION(balance_parent_reads);
1039     ASSIGN_OPTION(localize_parent_reads);
1040     ASSIGN_OPTION(readahead_trigger_requests);
1041     ASSIGN_OPTION(readahead_max_bytes);
1042     ASSIGN_OPTION(readahead_disable_after_bytes);
1043     ASSIGN_OPTION(clone_copy_on_read);
1044     ASSIGN_OPTION(blacklist_on_break_lock);
1045     ASSIGN_OPTION(blacklist_expire_seconds);
1046     ASSIGN_OPTION(request_timed_out_seconds);
1047     ASSIGN_OPTION(enable_alloc_hint);
1048     ASSIGN_OPTION(journal_order);
1049     ASSIGN_OPTION(journal_splay_width);
1050     ASSIGN_OPTION(journal_commit_age);
1051     ASSIGN_OPTION(journal_object_flush_interval);
1052     ASSIGN_OPTION(journal_object_flush_bytes);
1053     ASSIGN_OPTION(journal_object_flush_age);
1054     ASSIGN_OPTION(journal_pool);
1055     ASSIGN_OPTION(journal_max_payload_bytes);
1056     ASSIGN_OPTION(journal_max_concurrent_object_sets);
1057     ASSIGN_OPTION(mirroring_resync_after_disconnect);
1058     ASSIGN_OPTION(mirroring_replay_delay);
1059     ASSIGN_OPTION(skip_partial_discard);
1060   }
1061
1062   ExclusiveLock<ImageCtx> *ImageCtx::create_exclusive_lock() {
1063     return new ExclusiveLock<ImageCtx>(*this);
1064   }
1065
1066   ObjectMap<ImageCtx> *ImageCtx::create_object_map(uint64_t snap_id) {
1067     return new ObjectMap<ImageCtx>(*this, snap_id);
1068   }
1069
1070   Journal<ImageCtx> *ImageCtx::create_journal() {
1071     return new Journal<ImageCtx>(*this);
1072   }
1073
1074   void ImageCtx::set_image_name(const std::string &image_name) {
1075     // update the name so rename can be invoked repeatedly
1076     RWLock::RLocker owner_locker(owner_lock);
1077     RWLock::WLocker snap_locker(snap_lock);
1078     name = image_name;
1079     if (old_format) {
1080       header_oid = util::old_header_name(image_name);
1081     }
1082   }
1083
1084   void ImageCtx::notify_update() {
1085     state->handle_update_notification();
1086     ImageWatcher<>::notify_header_update(md_ctx, header_oid);
1087   }
1088
1089   void ImageCtx::notify_update(Context *on_finish) {
1090     state->handle_update_notification();
1091     image_watcher->notify_header_update(on_finish);
1092   }
1093
1094   exclusive_lock::Policy *ImageCtx::get_exclusive_lock_policy() const {
1095     assert(owner_lock.is_locked());
1096     assert(exclusive_lock_policy != nullptr);
1097     return exclusive_lock_policy;
1098   }
1099
1100   void ImageCtx::set_exclusive_lock_policy(exclusive_lock::Policy *policy) {
1101     assert(owner_lock.is_wlocked());
1102     assert(policy != nullptr);
1103     delete exclusive_lock_policy;
1104     exclusive_lock_policy = policy;
1105   }
1106
1107   journal::Policy *ImageCtx::get_journal_policy() const {
1108     assert(snap_lock.is_locked());
1109     assert(journal_policy != nullptr);
1110     return journal_policy;
1111   }
1112
1113   void ImageCtx::set_journal_policy(journal::Policy *policy) {
1114     assert(snap_lock.is_wlocked());
1115     assert(policy != nullptr);
1116     delete journal_policy;
1117     journal_policy = policy;
1118   }
1119
1120   void ImageCtx::get_thread_pool_instance(CephContext *cct,
1121                                           ThreadPool **thread_pool,
1122                                           ContextWQ **op_work_queue) {
1123     ThreadPoolSingleton *thread_pool_singleton;
1124     cct->lookup_or_create_singleton_object<ThreadPoolSingleton>(
1125       thread_pool_singleton, "librbd::thread_pool");
1126     *thread_pool = thread_pool_singleton;
1127     *op_work_queue = thread_pool_singleton->op_work_queue;
1128   }
1129
1130   void ImageCtx::get_timer_instance(CephContext *cct, SafeTimer **timer,
1131                                     Mutex **timer_lock) {
1132     SafeTimerSingleton *safe_timer_singleton;
1133     cct->lookup_or_create_singleton_object<SafeTimerSingleton>(
1134       safe_timer_singleton, "librbd::journal::safe_timer");
1135     *timer = safe_timer_singleton;
1136     *timer_lock = &safe_timer_singleton->lock;
1137   }
1138 }