ceph/src/librbd/Journal.h

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3
   4 #ifndef CEPH_LIBRBD_JOURNAL_H
   5 #define CEPH_LIBRBD_JOURNAL_H
   6
   7 #include "include/int_types.h"
   8 #include "include/Context.h"
   9 #include "include/interval_set.h"
  10 #include "include/rados/librados_fwd.hpp"
  11 #include "common/Cond.h"
  12 #include "common/Mutex.h"
  13 #include "common/Cond.h"
  14 #include "common/WorkQueue.h"
  15 #include "journal/Future.h"
  16 #include "journal/JournalMetadataListener.h"
  17 #include "journal/ReplayEntry.h"
  18 #include "journal/ReplayHandler.h"
  19 #include "librbd/Utils.h"
  20 #include "librbd/journal/Types.h"
  21 #include "librbd/journal/TypeTraits.h"
  22
  23 #include <algorithm>
  24 #include <list>
  25 #include <string>
  26 #include <atomic>
  27 #include <unordered_map>
  28
  29 class SafeTimer;
  30 namespace journal {
  31 class Journaler;
  32 }
  33
  34 namespace librbd {
  35
  36 class ImageCtx;
  37
  38 namespace journal { template <typename> class Replay; }
  39
  40 template <typename ImageCtxT = ImageCtx>
  41 class Journal {
  42 public:
  43   /**
  44    * @verbatim
  45    *
  46    * <start>
  47    *    |
  48    *    v
  49    * UNINITIALIZED ---> INITIALIZING ---> REPLAYING ------> FLUSHING ---> READY
  50    *    |                 *  .  ^             *  .              *           |
  51    *    |                 *  .  |             *  .              *           |
  52    *    |                 *  .  |    (error)  *  . . . . . . .  *           |
  53    *    |                 *  .  |             *              .  *           |
  54    *    |                 *  .  |             v              .  *           |
  55    *    |                 *  .  |         FLUSHING_RESTART   .  *           |
  56    *    |                 *  .  |             |              .  *           |
  57    *    |                 *  .  |             |              .  *           |
  58    *    |                 *  .  |             v              .  *           v
  59    *    |                 *  .  |         RESTARTING  < * * * * *       STOPPING
  60    *    |                 *  .  |             |              .              |
  61    *    |                 *  .  |             |              .              |
  62    *    |       * * * * * *  .  \-------------/              .              |
  63    *    |       * (error)    .                               .              |
  64    *    |       *            .   . . . . . . . . . . . . . . .              |
  65    *    |       *            .   .                                          |
  66    *    |       v            v   v                                          |
  67    *    |     CLOSED <----- CLOSING <---------------------------------------/
  68    *    |       |
  69    *    |       v
  70    *    \---> <finish>
  71    *
  72    * @endverbatim
  73    */
  74   enum State {
  75     STATE_UNINITIALIZED,
  76     STATE_INITIALIZING,
  77     STATE_REPLAYING,
  78     STATE_FLUSHING_RESTART,
  79     STATE_RESTARTING_REPLAY,
  80     STATE_FLUSHING_REPLAY,
  81     STATE_READY,
  82     STATE_STOPPING,
  83     STATE_CLOSING,
  84     STATE_CLOSED
  85   };
  86
  87   static const std::string IMAGE_CLIENT_ID;
  88   static const std::string LOCAL_MIRROR_UUID;
  89   static const std::string ORPHAN_MIRROR_UUID;
  90
  91   Journal(ImageCtxT &image_ctx);
  92   ~Journal();
  93
  94   static bool is_journal_supported(ImageCtxT &image_ctx);
  95   static int create(librados::IoCtx &io_ctx, const std::string &image_id,
  96                     uint8_t order, uint8_t splay_width,
  97                     const std::string &object_pool);
  98   static int remove(librados::IoCtx &io_ctx, const std::string &image_id);
  99   static int reset(librados::IoCtx &io_ctx, const std::string &image_id);
 100
 101   static void is_tag_owner(ImageCtxT *image_ctx, bool *is_tag_owner,
 102                            Context *on_finish);
 103   static void is_tag_owner(librados::IoCtx& io_ctx, std::string& image_id,
 104                            bool *is_tag_owner, ContextWQ *op_work_queue,
 105                            Context *on_finish);
 106   static void get_tag_owner(librados::IoCtx& io_ctx, std::string& image_id,
 107                             std::string *mirror_uuid,
 108                             ContextWQ *op_work_queue, Context *on_finish);
 109   static int request_resync(ImageCtxT *image_ctx);
 110   static void promote(ImageCtxT *image_ctx, Context *on_finish);
 111   static void demote(ImageCtxT *image_ctx, Context *on_finish);
 112
 113   bool is_journal_ready() const;
 114   bool is_journal_replaying() const;
 115   bool is_journal_appending() const;
 116
 117   void wait_for_journal_ready(Context *on_ready);
 118
 119   void open(Context *on_finish);
 120   void close(Context *on_finish);
 121
 122   bool is_tag_owner() const;
 123   uint64_t get_tag_tid() const;
 124   journal::TagData get_tag_data() const;
 125
 126   void allocate_local_tag(Context *on_finish);
 127   void allocate_tag(const std::string &mirror_uuid,
 128                     const journal::TagPredecessor &predecessor,
 129                     Context *on_finish);
 130
 131   void flush_commit_position(Context *on_finish);
 132
 133   uint64_t append_write_event(uint64_t offset, size_t length,
 134                               const bufferlist &bl,
 135                               bool flush_entry);
 136   uint64_t append_io_event(journal::EventEntry &&event_entry,
 137                            uint64_t offset, size_t length,
 138                            bool flush_entry, int filter_ret_val);
 139   void commit_io_event(uint64_t tid, int r);
 140   void commit_io_event_extent(uint64_t tid, uint64_t offset, uint64_t length,
 141                               int r);
 142
 143   void append_op_event(uint64_t op_tid, journal::EventEntry &&event_entry,
 144                        Context *on_safe);
 145   void commit_op_event(uint64_t tid, int r, Context *on_safe);
 146   void replay_op_ready(uint64_t op_tid, Context *on_resume);
 147
 148   void flush_event(uint64_t tid, Context *on_safe);
 149   void wait_event(uint64_t tid, Context *on_safe);
 150
 151   uint64_t allocate_op_tid() {
 152     uint64_t op_tid = ++m_op_tid;
 153     ceph_assert(op_tid != 0);
 154     return op_tid;
 155   }
 156
 157   void start_external_replay(journal::Replay<ImageCtxT> **journal_replay,
 158                              Context *on_start);
 159   void stop_external_replay();
 160
 161   void add_listener(journal::Listener *listener);
 162   void remove_listener(journal::Listener *listener);
 163
 164   int is_resync_requested(bool *do_resync);
 165
 166   inline ContextWQ *get_work_queue() {
 167     return m_work_queue;
 168   }
 169
 170 private:
 171   ImageCtxT &m_image_ctx;
 172
 173   // mock unit testing support
 174   typedef journal::TypeTraits<ImageCtxT> TypeTraits;
 175   typedef typename TypeTraits::Journaler Journaler;
 176   typedef typename TypeTraits::Future Future;
 177   typedef typename TypeTraits::ReplayEntry ReplayEntry;
 178
 179   typedef std::list<bufferlist> Bufferlists;
 180   typedef std::list<Context *> Contexts;
 181   typedef std::list<Future> Futures;
 182   typedef interval_set<uint64_t> ExtentInterval;
 183
 184   struct Event {
 185     Futures futures;
 186     Contexts on_safe_contexts;
 187     ExtentInterval pending_extents;
 188     int filter_ret_val = 0;
 189     bool committed_io = false;
 190     bool safe = false;
 191     int ret_val = 0;
 192
 193     Event() {
 194     }
 195     Event(const Futures &_futures, uint64_t offset, size_t length,
 196           int filter_ret_val)
 197       : futures(_futures), filter_ret_val(filter_ret_val) {
 198       if (length > 0) {
 199         pending_extents.insert(offset, length);
 200       }
 201     }
 202   };
 203
 204   typedef std::unordered_map<uint64_t, Event> Events;
 205   typedef std::unordered_map<uint64_t, Future> TidToFutures;
 206
 207   struct C_IOEventSafe : public Context {
 208     Journal *journal;
 209     uint64_t tid;
 210
 211     C_IOEventSafe(Journal *_journal, uint64_t _tid)
 212       : journal(_journal), tid(_tid) {
 213     }
 214
 215     void finish(int r) override {
 216       journal->handle_io_event_safe(r, tid);
 217     }
 218   };
 219
 220   struct C_OpEventSafe : public Context {
 221     Journal *journal;
 222     uint64_t tid;
 223     Future op_start_future;
 224     Future op_finish_future;
 225     Context *on_safe;
 226
 227     C_OpEventSafe(Journal *journal, uint64_t tid, const Future &op_start_future,
 228                   const Future &op_finish_future, Context *on_safe)
 229       : journal(journal), tid(tid), op_start_future(op_start_future),
 230         op_finish_future(op_finish_future), on_safe(on_safe) {
 231     }
 232
 233     void finish(int r) override {
 234       journal->handle_op_event_safe(r, tid, op_start_future, op_finish_future,
 235                                     on_safe);
 236     }
 237   };
 238
 239   struct C_ReplayProcessSafe : public Context {
 240     Journal *journal;
 241     ReplayEntry replay_entry;
 242
 243     C_ReplayProcessSafe(Journal *journal, ReplayEntry &&replay_entry) :
 244       journal(journal), replay_entry(std::move(replay_entry)) {
 245     }
 246     void finish(int r) override {
 247       journal->handle_replay_process_safe(replay_entry, r);
 248     }
 249   };
 250
 251   struct ReplayHandler : public ::journal::ReplayHandler {
 252     Journal *journal;
 253     ReplayHandler(Journal *_journal) : journal(_journal) {
 254     }
 255
 256     void get() override {
 257       // TODO
 258     }
 259     void put() override {
 260       // TODO
 261     }
 262
 263     void handle_entries_available() override {
 264       journal->handle_replay_ready();
 265     }
 266     void handle_complete(int r) override {
 267       journal->handle_replay_complete(r);
 268     }
 269   };
 270
 271   ContextWQ *m_work_queue = nullptr;
 272   SafeTimer *m_timer = nullptr;
 273   Mutex *m_timer_lock = nullptr;
 274
 275   Journaler *m_journaler;
 276   mutable Mutex m_lock;
 277   State m_state;
 278   uint64_t m_max_append_size = 0;
 279   uint64_t m_tag_class = 0;
 280   uint64_t m_tag_tid = 0;
 281   journal::ImageClientMeta m_client_meta;
 282   journal::TagData m_tag_data;
 283
 284   int m_error_result;
 285   Contexts m_wait_for_state_contexts;
 286
 287   ReplayHandler m_replay_handler;
 288   bool m_close_pending;
 289
 290   Mutex m_event_lock;
 291   uint64_t m_event_tid;
 292   Events m_events;
 293
 294   std::atomic<uint64_t> m_op_tid = { 0 };
 295   TidToFutures m_op_futures;
 296
 297   bool m_processing_entry = false;
 298   bool m_blocking_writes;
 299
 300   journal::Replay<ImageCtxT> *m_journal_replay;
 301
 302   util::AsyncOpTracker m_async_journal_op_tracker;
 303
 304   struct MetadataListener : public ::journal::JournalMetadataListener {
 305     Journal<ImageCtxT> *journal;
 306
 307     MetadataListener(Journal<ImageCtxT> *journal) : journal(journal) { }
 308
 309     void handle_update(::journal::JournalMetadata *) override {
 310       FunctionContext *ctx = new FunctionContext([this](int r) {
 311         journal->handle_metadata_updated();
 312       });
 313       journal->m_work_queue->queue(ctx, 0);
 314     }
 315   } m_metadata_listener;
 316
 317   typedef std::set<journal::Listener *> Listeners;
 318   Listeners m_listeners;
 319   Cond m_listener_cond;
 320   bool m_listener_notify = false;
 321
 322   uint64_t m_refresh_sequence = 0;
 323
 324   bool is_journal_replaying(const Mutex &) const;
 325   bool is_tag_owner(const Mutex &) const;
 326
 327   uint64_t append_io_events(journal::EventType event_type,
 328                             const Bufferlists &bufferlists,
 329                             uint64_t offset, size_t length, bool flush_entry,
 330                             int filter_ret_val);
 331   Future wait_event(Mutex &lock, uint64_t tid, Context *on_safe);
 332
 333   void create_journaler();
 334   void destroy_journaler(int r);
 335   void recreate_journaler(int r);
 336
 337   void complete_event(typename Events::iterator it, int r);
 338
 339   void start_append();
 340
 341   void handle_open(int r);
 342
 343   void handle_replay_ready();
 344   void handle_replay_complete(int r);
 345   void handle_replay_process_ready(int r);
 346   void handle_replay_process_safe(ReplayEntry replay_entry, int r);
 347
 348   void handle_start_external_replay(int r,
 349                                     journal::Replay<ImageCtxT> **journal_replay,
 350                                     Context *on_finish);
 351
 352   void handle_flushing_restart(int r);
 353   void handle_flushing_replay();
 354
 355   void handle_recording_stopped(int r);
 356
 357   void handle_journal_destroyed(int r);
 358
 359   void handle_io_event_safe(int r, uint64_t tid);
 360   void handle_op_event_safe(int r, uint64_t tid, const Future &op_start_future,
 361                             const Future &op_finish_future, Context *on_safe);
 362
 363   void stop_recording();
 364
 365   void transition_state(State state, int r);
 366
 367   bool is_steady_state() const;
 368   void wait_for_steady_state(Context *on_state);
 369
 370   int check_resync_requested(bool *do_resync);
 371
 372   void handle_metadata_updated();
 373   void handle_refresh_metadata(uint64_t refresh_sequence, uint64_t tag_tid,
 374                                journal::TagData tag_data, int r);
 375
 376 };
 377
 378 } // namespace librbd
 379
 380 extern template class librbd::Journal<librbd::ImageCtx>;
 381
 382 #endif // CEPH_LIBRBD_JOURNAL_H