ceph/src/librbd/io/ImageRequest.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3
   4 #include "librbd/io/ImageRequest.h"
   5 #include "librbd/ImageCtx.h"
   6 #include "librbd/internal.h"
   7 #include "librbd/Journal.h"
   8 #include "librbd/Utils.h"
   9 #include "librbd/cache/ImageCache.h"
  10 #include "librbd/io/AioCompletion.h"
  11 #include "librbd/io/ObjectRequest.h"
  12 #include "librbd/journal/Types.h"
  13 #include "include/rados/librados.hpp"
  14 #include "common/WorkQueue.h"
  15 #include "osdc/Striper.h"
  16
  17 #define dout_subsys ceph_subsys_rbd
  18 #undef dout_prefix
  19 #define dout_prefix *_dout << "librbd::io::ImageRequest: " << this \
  20                            << " " << __func__ << ": "
  21
  22 namespace librbd {
  23 namespace io {
  24
  25 using util::get_image_ctx;
  26
  27 namespace {
  28
  29 template <typename ImageCtxT = ImageCtx>
  30 struct C_DiscardJournalCommit : public Context {
  31   typedef std::vector<ObjectExtent> ObjectExtents;
  32
  33   ImageCtxT &image_ctx;
  34   AioCompletion *aio_comp;
  35   ObjectExtents object_extents;
  36
  37   C_DiscardJournalCommit(ImageCtxT &_image_ctx, AioCompletion *_aio_comp,
  38                          const ObjectExtents &_object_extents, uint64_t tid)
  39     : image_ctx(_image_ctx), aio_comp(_aio_comp),
  40       object_extents(_object_extents) {
  41     CephContext *cct = image_ctx.cct;
  42     ldout(cct, 20) << "delaying cache discard until journal tid " << tid << " "
  43                    << "safe" << dendl;
  44
  45     aio_comp->add_request();
  46   }
  47
  48   void finish(int r) override {
  49     CephContext *cct = image_ctx.cct;
  50     ldout(cct, 20) << "C_DiscardJournalCommit: "
  51                    << "journal committed: discarding from cache" << dendl;
  52
  53     Mutex::Locker cache_locker(image_ctx.cache_lock);
  54     image_ctx.object_cacher->discard_set(image_ctx.object_set, object_extents);
  55     aio_comp->complete_request(r);
  56   }
  57 };
  58
  59 template <typename ImageCtxT = ImageCtx>
  60 struct C_FlushJournalCommit : public Context {
  61   ImageCtxT &image_ctx;
  62   AioCompletion *aio_comp;
  63
  64   C_FlushJournalCommit(ImageCtxT &_image_ctx, AioCompletion *_aio_comp,
  65                        uint64_t tid)
  66     : image_ctx(_image_ctx), aio_comp(_aio_comp) {
  67     CephContext *cct = image_ctx.cct;
  68     ldout(cct, 20) << "delaying flush until journal tid " << tid << " "
  69                    << "safe" << dendl;
  70
  71     aio_comp->add_request();
  72   }
  73
  74   void finish(int r) override {
  75     CephContext *cct = image_ctx.cct;
  76     ldout(cct, 20) << "C_FlushJournalCommit: journal committed" << dendl;
  77     aio_comp->complete_request(r);
  78   }
  79 };
  80
  81 template <typename ImageCtxT>
  82 class C_ObjectCacheRead : public Context {
  83 public:
  84   explicit C_ObjectCacheRead(ImageCtxT &ictx, ObjectReadRequest<ImageCtxT> *req)
  85     : m_image_ctx(ictx), m_req(req), m_enqueued(false) {}
  86
  87   void complete(int r) override {
  88     if (!m_enqueued) {
  89       // cache_lock creates a lock ordering issue -- so re-execute this context
  90       // outside the cache_lock
  91       m_enqueued = true;
  92       m_image_ctx.op_work_queue->queue(this, r);
  93       return;
  94     }
  95     Context::complete(r);
  96   }
  97
  98 protected:
  99   void finish(int r) override {
 100     m_req->complete(r);
 101   }
 102
 103 private:
 104   ImageCtxT &m_image_ctx;
 105   ObjectReadRequest<ImageCtxT> *m_req;
 106   bool m_enqueued;
 107 };
 108
 109 } // anonymous namespace
 110
 111 template <typename I>
 112 ImageRequest<I>* ImageRequest<I>::create_read_request(
 113     I &image_ctx, AioCompletion *aio_comp, Extents &&image_extents,
 114     ReadResult &&read_result, int op_flags,
 115     const ZTracer::Trace &parent_trace) {
 116   return new ImageReadRequest<I>(image_ctx, aio_comp,
 117                                  std::move(image_extents),
 118                                  std::move(read_result), op_flags,
 119                                  parent_trace);
 120 }
 121
 122 template <typename I>
 123 ImageRequest<I>* ImageRequest<I>::create_write_request(
 124     I &image_ctx, AioCompletion *aio_comp, Extents &&image_extents,
 125     bufferlist &&bl, int op_flags, const ZTracer::Trace &parent_trace) {
 126   return new ImageWriteRequest<I>(image_ctx, aio_comp, std::move(image_extents),
 127                                   std::move(bl), op_flags, parent_trace);
 128 }
 129
 130 template <typename I>
 131 ImageRequest<I>* ImageRequest<I>::create_discard_request(
 132     I &image_ctx, AioCompletion *aio_comp, uint64_t off, uint64_t len,
 133     bool skip_partial_discard, const ZTracer::Trace &parent_trace) {
 134   return new ImageDiscardRequest<I>(image_ctx, aio_comp, off, len,
 135                                     skip_partial_discard, parent_trace);
 136 }
 137
 138 template <typename I>
 139 ImageRequest<I>* ImageRequest<I>::create_flush_request(
 140     I &image_ctx, AioCompletion *aio_comp,
 141     const ZTracer::Trace &parent_trace) {
 142   return new ImageFlushRequest<I>(image_ctx, aio_comp, parent_trace);
 143 }
 144
 145 template <typename I>
 146 ImageRequest<I>* ImageRequest<I>::create_writesame_request(
 147     I &image_ctx, AioCompletion *aio_comp, uint64_t off, uint64_t len,
 148     bufferlist &&bl, int op_flags, const ZTracer::Trace &parent_trace) {
 149   return new ImageWriteSameRequest<I>(image_ctx, aio_comp, off, len,
 150                                       std::move(bl), op_flags, parent_trace);
 151 }
 152
 153 template <typename I>
 154 ImageRequest<I>* ImageRequest<I>::create_compare_and_write_request(
 155     I &image_ctx, AioCompletion *c, Extents &&image_extents,
 156     bufferlist &&cmp_bl, bufferlist &&bl, uint64_t *mismatch_offset,
 157     int op_flags, const ZTracer::Trace &parent_trace) {
 158   return new ImageCompareAndWriteRequest<I>(image_ctx, c,
 159                                             std::move(image_extents),
 160                                             std::move(cmp_bl),
 161                                             std::move(bl), mismatch_offset,
 162                                             op_flags, parent_trace);
 163 }
 164
 165 template <typename I>
 166 void ImageRequest<I>::aio_read(I *ictx, AioCompletion *c,
 167                                Extents &&image_extents,
 168                                ReadResult &&read_result, int op_flags,
 169                                const ZTracer::Trace &parent_trace) {
 170   ImageReadRequest<I> req(*ictx, c, std::move(image_extents),
 171                           std::move(read_result), op_flags, parent_trace);
 172   req.send();
 173 }
 174
 175 template <typename I>
 176 void ImageRequest<I>::aio_write(I *ictx, AioCompletion *c,
 177                                 Extents &&image_extents, bufferlist &&bl,
 178                                 int op_flags,
 179                                 const ZTracer::Trace &parent_trace) {
 180   ImageWriteRequest<I> req(*ictx, c, std::move(image_extents), std::move(bl),
 181                            op_flags, parent_trace);
 182   req.send();
 183 }
 184
 185 template <typename I>
 186 void ImageRequest<I>::aio_discard(I *ictx, AioCompletion *c,
 187                                   uint64_t off, uint64_t len,
 188                                   bool skip_partial_discard,
 189                                   const ZTracer::Trace &parent_trace) {
 190   ImageDiscardRequest<I> req(*ictx, c, off, len, skip_partial_discard,
 191                              parent_trace);
 192   req.send();
 193 }
 194
 195 template <typename I>
 196 void ImageRequest<I>::aio_flush(I *ictx, AioCompletion *c,
 197                                 const ZTracer::Trace &parent_trace) {
 198   ImageFlushRequest<I> req(*ictx, c, parent_trace);
 199   req.send();
 200 }
 201
 202 template <typename I>
 203 void ImageRequest<I>::aio_writesame(I *ictx, AioCompletion *c,
 204                                     uint64_t off, uint64_t len,
 205                                     bufferlist &&bl, int op_flags,
 206                                     const ZTracer::Trace &parent_trace) {
 207   ImageWriteSameRequest<I> req(*ictx, c, off, len, std::move(bl), op_flags,
 208                                parent_trace);
 209   req.send();
 210 }
 211
 212 template <typename I>
 213 void ImageRequest<I>::aio_compare_and_write(I *ictx, AioCompletion *c,
 214                                             Extents &&image_extents,
 215                                             bufferlist &&cmp_bl,
 216                                             bufferlist &&bl,
 217                                             uint64_t *mismatch_offset,
 218                                             int op_flags,
 219                                             const ZTracer::Trace &parent_trace) {
 220   ImageCompareAndWriteRequest<I> req(*ictx, c, std::move(image_extents),
 221                                      std::move(cmp_bl), std::move(bl),
 222                                      mismatch_offset, op_flags, parent_trace);
 223   req.send();
 224 }
 225
 226
 227 template <typename I>
 228 void ImageRequest<I>::send() {
 229   I &image_ctx = this->m_image_ctx;
 230   assert(m_aio_comp->is_initialized(get_aio_type()));
 231   assert(m_aio_comp->is_started() ^ (get_aio_type() == AIO_TYPE_FLUSH));
 232
 233   CephContext *cct = image_ctx.cct;
 234   AioCompletion *aio_comp = this->m_aio_comp;
 235   ldout(cct, 20) << get_request_type() << ": ictx=" << &image_ctx << ", "
 236                  << "completion=" << aio_comp << dendl;
 237
 238   aio_comp->get();
 239   int r = clip_request();
 240   if (r < 0) {
 241     m_aio_comp->fail(r);
 242     return;
 243   }
 244
 245   if (m_bypass_image_cache || m_image_ctx.image_cache == nullptr) {
 246     send_request();
 247   } else {
 248     send_image_cache_request();
 249   }
 250 }
 251
 252 template <typename I>
 253 int ImageRequest<I>::clip_request() {
 254   RWLock::RLocker snap_locker(m_image_ctx.snap_lock);
 255   for (auto &image_extent : m_image_extents) {
 256     auto clip_len = image_extent.second;
 257     int r = clip_io(get_image_ctx(&m_image_ctx), image_extent.first, &clip_len);
 258     if (r < 0) {
 259       return r;
 260     }
 261
 262     image_extent.second = clip_len;
 263   }
 264   return 0;
 265 }
 266
 267 template <typename I>
 268 void ImageRequest<I>::start_op() {
 269   m_aio_comp->start_op();
 270 }
 271
 272 template <typename I>
 273 void ImageRequest<I>::fail(int r) {
 274   AioCompletion *aio_comp = this->m_aio_comp;
 275   aio_comp->get();
 276   aio_comp->fail(r);
 277 }
 278
 279 template <typename I>
 280 ImageReadRequest<I>::ImageReadRequest(I &image_ctx, AioCompletion *aio_comp,
 281                                       Extents &&image_extents,
 282                                       ReadResult &&read_result, int op_flags,
 283                                       const ZTracer::Trace &parent_trace)
 284   : ImageRequest<I>(image_ctx, aio_comp, std::move(image_extents), "read",
 285                     parent_trace),
 286     m_op_flags(op_flags) {
 287   aio_comp->read_result = std::move(read_result);
 288 }
 289
 290 template <typename I>
 291 int ImageReadRequest<I>::clip_request() {
 292   int r = ImageRequest<I>::clip_request();
 293   if (r < 0) {
 294     return r;
 295   }
 296
 297   uint64_t buffer_length = 0;
 298   auto &image_extents = this->m_image_extents;
 299   for (auto &image_extent : image_extents) {
 300     buffer_length += image_extent.second;
 301   }
 302   this->m_aio_comp->read_result.set_clip_length(buffer_length);
 303   return 0;
 304 }
 305
 306 template <typename I>
 307 void ImageReadRequest<I>::send_request() {
 308   I &image_ctx = this->m_image_ctx;
 309   CephContext *cct = image_ctx.cct;
 310
 311   auto &image_extents = this->m_image_extents;
 312   if (image_ctx.object_cacher && image_ctx.readahead_max_bytes > 0 &&
 313       !(m_op_flags & LIBRADOS_OP_FLAG_FADVISE_RANDOM)) {
 314     readahead(get_image_ctx(&image_ctx), image_extents);
 315   }
 316
 317   AioCompletion *aio_comp = this->m_aio_comp;
 318   librados::snap_t snap_id;
 319   map<object_t,vector<ObjectExtent> > object_extents;
 320   uint64_t buffer_ofs = 0;
 321   {
 322     // prevent image size from changing between computing clip and recording
 323     // pending async operation
 324     RWLock::RLocker snap_locker(image_ctx.snap_lock);
 325     snap_id = image_ctx.snap_id;
 326
 327     // map image extents to object extents
 328     for (auto &extent : image_extents) {
 329       if (extent.second == 0) {
 330         continue;
 331       }
 332
 333       Striper::file_to_extents(cct, image_ctx.format_string, &image_ctx.layout,
 334                                extent.first, extent.second, 0, object_extents,
 335                                buffer_ofs);
 336       buffer_ofs += extent.second;
 337     }
 338   }
 339
 340   // pre-calculate the expected number of read requests
 341   uint32_t request_count = 0;
 342   for (auto &object_extent : object_extents) {
 343     request_count += object_extent.second.size();
 344   }
 345   aio_comp->set_request_count(request_count);
 346
 347   // issue the requests
 348   for (auto &object_extent : object_extents) {
 349     for (auto &extent : object_extent.second) {
 350       ldout(cct, 20) << "oid " << extent.oid << " " << extent.offset << "~"
 351                      << extent.length << " from " << extent.buffer_extents
 352                      << dendl;
 353
 354       auto req_comp = new io::ReadResult::C_SparseReadRequest<I>(
 355         aio_comp);
 356       ObjectReadRequest<I> *req = ObjectReadRequest<I>::create(
 357         &image_ctx, extent.oid.name, extent.objectno, extent.offset,
 358         extent.length, extent.buffer_extents, snap_id, true, m_op_flags,
 359         this->m_trace, req_comp);
 360       req_comp->request = req;
 361
 362       if (image_ctx.object_cacher) {
 363         C_ObjectCacheRead<I> *cache_comp = new C_ObjectCacheRead<I>(image_ctx,
 364                                                                     req);
 365         image_ctx.aio_read_from_cache(
 366           extent.oid, extent.objectno, &req->data(), extent.length,
 367           extent.offset, cache_comp, m_op_flags,
 368           (this->m_trace.valid() ? &this->m_trace : nullptr));
 369       } else {
 370         req->send();
 371       }
 372     }
 373   }
 374
 375   aio_comp->put();
 376
 377   image_ctx.perfcounter->inc(l_librbd_rd);
 378   image_ctx.perfcounter->inc(l_librbd_rd_bytes, buffer_ofs);
 379 }
 380
 381 template <typename I>
 382 void ImageReadRequest<I>::send_image_cache_request() {
 383   I &image_ctx = this->m_image_ctx;
 384   assert(image_ctx.image_cache != nullptr);
 385
 386   AioCompletion *aio_comp = this->m_aio_comp;
 387   aio_comp->set_request_count(1);
 388
 389   auto *req_comp = new io::ReadResult::C_ImageReadRequest(
 390     aio_comp, this->m_image_extents);
 391   image_ctx.image_cache->aio_read(std::move(this->m_image_extents),
 392                                   &req_comp->bl, m_op_flags,
 393                                   req_comp);
 394 }
 395
 396 template <typename I>
 397 void AbstractImageWriteRequest<I>::send_request() {
 398   I &image_ctx = this->m_image_ctx;
 399   CephContext *cct = image_ctx.cct;
 400
 401   RWLock::RLocker md_locker(image_ctx.md_lock);
 402
 403   bool journaling = false;
 404
 405   AioCompletion *aio_comp = this->m_aio_comp;
 406   uint64_t clip_len = 0;
 407   ObjectExtents object_extents;
 408   ::SnapContext snapc;
 409   {
 410     // prevent image size from changing between computing clip and recording
 411     // pending async operation
 412     RWLock::RLocker snap_locker(image_ctx.snap_lock);
 413     if (image_ctx.snap_id != CEPH_NOSNAP || image_ctx.read_only) {
 414       aio_comp->fail(-EROFS);
 415       return;
 416     }
 417
 418     for (auto &extent : this->m_image_extents) {
 419       if (extent.second == 0) {
 420         continue;
 421       }
 422
 423       // map to object extents
 424       Striper::file_to_extents(cct, image_ctx.format_string, &image_ctx.layout,
 425                                extent.first, extent.second, 0, object_extents);
 426       clip_len += extent.second;
 427     }
 428
 429     snapc = image_ctx.snapc;
 430     journaling = (image_ctx.journal != nullptr &&
 431                   image_ctx.journal->is_journal_appending());
 432   }
 433
 434   int ret = prune_object_extents(object_extents);
 435   if (ret < 0) {
 436     aio_comp->fail(ret);
 437     return;
 438   }
 439
 440   if (!object_extents.empty()) {
 441     uint64_t journal_tid = 0;
 442     aio_comp->set_request_count(
 443       object_extents.size() + get_object_cache_request_count(journaling));
 444
 445     ObjectRequests requests;
 446     send_object_requests(object_extents, snapc,
 447                          (journaling ? &requests : nullptr));
 448
 449     if (journaling) {
 450       // in-flight ops are flushed prior to closing the journal
 451       assert(image_ctx.journal != NULL);
 452       journal_tid = append_journal_event(requests, m_synchronous);
 453     }
 454
 455     if (image_ctx.object_cacher != NULL) {
 456       send_object_cache_requests(object_extents, journal_tid);
 457     }
 458   } else {
 459     // no IO to perform -- fire completion
 460     aio_comp->unblock();
 461   }
 462
 463   update_stats(clip_len);
 464   aio_comp->put();
 465 }
 466
 467 template <typename I>
 468 void AbstractImageWriteRequest<I>::send_object_requests(
 469     const ObjectExtents &object_extents, const ::SnapContext &snapc,
 470     ObjectRequests *object_requests) {
 471   I &image_ctx = this->m_image_ctx;
 472   CephContext *cct = image_ctx.cct;
 473
 474   AioCompletion *aio_comp = this->m_aio_comp;
 475   for (ObjectExtents::const_iterator p = object_extents.begin();
 476        p != object_extents.end(); ++p) {
 477     ldout(cct, 20) << "oid " << p->oid << " " << p->offset << "~" << p->length
 478                    << " from " << p->buffer_extents << dendl;
 479     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 480     ObjectRequestHandle *request = create_object_request(*p, snapc,
 481                                                             req_comp);
 482
 483     // if journaling, stash the request for later; otherwise send
 484     if (request != NULL) {
 485       if (object_requests != NULL) {
 486         object_requests->push_back(request);
 487       } else {
 488         request->send();
 489       }
 490     }
 491   }
 492 }
 493
 494 template <typename I>
 495 void ImageWriteRequest<I>::assemble_extent(const ObjectExtent &object_extent,
 496                                            bufferlist *bl) {
 497   for (auto q = object_extent.buffer_extents.begin();
 498        q != object_extent.buffer_extents.end(); ++q) {
 499     bufferlist sub_bl;
 500     sub_bl.substr_of(m_bl, q->first, q->second);
 501     bl->claim_append(sub_bl);
 502   }
 503 }
 504
 505 template <typename I>
 506 uint64_t ImageWriteRequest<I>::append_journal_event(
 507     const ObjectRequests &requests, bool synchronous) {
 508   I &image_ctx = this->m_image_ctx;
 509
 510   uint64_t tid = 0;
 511   uint64_t buffer_offset = 0;
 512   assert(!this->m_image_extents.empty());
 513   for (auto &extent : this->m_image_extents) {
 514     bufferlist sub_bl;
 515     sub_bl.substr_of(m_bl, buffer_offset, extent.second);
 516     buffer_offset += extent.second;
 517
 518     tid = image_ctx.journal->append_write_event(extent.first, extent.second,
 519                                                 sub_bl, requests, synchronous);
 520   }
 521
 522   if (image_ctx.object_cacher == NULL) {
 523     AioCompletion *aio_comp = this->m_aio_comp;
 524     aio_comp->associate_journal_event(tid);
 525   }
 526   return tid;
 527 }
 528
 529 template <typename I>
 530 void ImageWriteRequest<I>::send_image_cache_request() {
 531   I &image_ctx = this->m_image_ctx;
 532   assert(image_ctx.image_cache != nullptr);
 533
 534   AioCompletion *aio_comp = this->m_aio_comp;
 535   aio_comp->set_request_count(1);
 536   C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 537   image_ctx.image_cache->aio_write(std::move(this->m_image_extents),
 538                                    std::move(m_bl), m_op_flags, req_comp);
 539 }
 540
 541 template <typename I>
 542 void ImageWriteRequest<I>::send_object_cache_requests(
 543     const ObjectExtents &object_extents, uint64_t journal_tid) {
 544   I &image_ctx = this->m_image_ctx;
 545   for (auto p = object_extents.begin(); p != object_extents.end(); ++p) {
 546     const ObjectExtent &object_extent = *p;
 547
 548     bufferlist bl;
 549     assemble_extent(object_extent, &bl);
 550
 551     AioCompletion *aio_comp = this->m_aio_comp;
 552     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 553     image_ctx.write_to_cache(
 554       object_extent.oid, bl, object_extent.length, object_extent.offset,
 555       req_comp, m_op_flags, journal_tid,
 556       (this->m_trace.valid() ? &this->m_trace : nullptr));
 557   }
 558 }
 559
 560 template <typename I>
 561 void ImageWriteRequest<I>::send_object_requests(
 562     const ObjectExtents &object_extents, const ::SnapContext &snapc,
 563     ObjectRequests *object_requests) {
 564   I &image_ctx = this->m_image_ctx;
 565
 566   // cache handles creating object requests during writeback
 567   if (image_ctx.object_cacher == NULL) {
 568     AbstractImageWriteRequest<I>::send_object_requests(object_extents, snapc,
 569                                                        object_requests);
 570   }
 571 }
 572
 573 template <typename I>
 574 ObjectRequestHandle *ImageWriteRequest<I>::create_object_request(
 575     const ObjectExtent &object_extent, const ::SnapContext &snapc,
 576     Context *on_finish) {
 577   I &image_ctx = this->m_image_ctx;
 578   assert(image_ctx.object_cacher == NULL);
 579
 580   bufferlist bl;
 581   assemble_extent(object_extent, &bl);
 582   ObjectRequest<I> *req = ObjectRequest<I>::create_write(
 583     &image_ctx, object_extent.oid.name, object_extent.objectno,
 584     object_extent.offset, bl, snapc, m_op_flags, this->m_trace, on_finish);
 585   return req;
 586 }
 587
 588 template <typename I>
 589 void ImageWriteRequest<I>::update_stats(size_t length) {
 590   I &image_ctx = this->m_image_ctx;
 591   image_ctx.perfcounter->inc(l_librbd_wr);
 592   image_ctx.perfcounter->inc(l_librbd_wr_bytes, length);
 593 }
 594
 595 template <typename I>
 596 uint64_t ImageDiscardRequest<I>::append_journal_event(
 597     const ObjectRequests &requests, bool synchronous) {
 598   I &image_ctx = this->m_image_ctx;
 599
 600   uint64_t tid = 0;
 601   assert(!this->m_image_extents.empty());
 602   for (auto &extent : this->m_image_extents) {
 603     journal::EventEntry event_entry(journal::AioDiscardEvent(extent.first,
 604                                                              extent.second,
 605                                                              this->m_skip_partial_discard));
 606     tid = image_ctx.journal->append_io_event(std::move(event_entry),
 607                                              requests, extent.first,
 608                                              extent.second, synchronous);
 609   }
 610
 611   AioCompletion *aio_comp = this->m_aio_comp;
 612   aio_comp->associate_journal_event(tid);
 613   return tid;
 614 }
 615
 616 template <typename I>
 617 int ImageDiscardRequest<I>::prune_object_extents(ObjectExtents &object_extents) {
 618   I &image_ctx = this->m_image_ctx;
 619   CephContext *cct = image_ctx.cct;
 620   if (!this->m_skip_partial_discard) {
 621     return 0;
 622   }
 623
 624   for (auto p = object_extents.begin(); p != object_extents.end(); ) {
 625     if (p->offset + p->length < image_ctx.layout.object_size) {
 626       ldout(cct, 20) << "oid " << p->oid << " " << p->offset << "~"
 627                      << p->length << " from " << p->buffer_extents
 628                      << ": skip partial discard" << dendl;
 629       p = object_extents.erase(p);
 630     } else {
 631       ++p;
 632     }
 633   }
 634
 635   return 0;
 636 }
 637
 638 template <typename I>
 639 uint32_t ImageDiscardRequest<I>::get_object_cache_request_count(bool journaling) const {
 640   // extra completion request is required for tracking journal commit
 641   I &image_ctx = this->m_image_ctx;
 642   return (image_ctx.object_cacher != nullptr && journaling ? 1 : 0);
 643 }
 644
 645 template <typename I>
 646 void ImageDiscardRequest<I>::send_image_cache_request() {
 647   I &image_ctx = this->m_image_ctx;
 648   assert(image_ctx.image_cache != nullptr);
 649
 650   AioCompletion *aio_comp = this->m_aio_comp;
 651   aio_comp->set_request_count(this->m_image_extents.size());
 652   for (auto &extent : this->m_image_extents) {
 653     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 654     image_ctx.image_cache->aio_discard(extent.first, extent.second,
 655                                        this->m_skip_partial_discard, req_comp);
 656   }
 657 }
 658
 659 template <typename I>
 660 void ImageDiscardRequest<I>::send_object_cache_requests(
 661     const ObjectExtents &object_extents, uint64_t journal_tid) {
 662   I &image_ctx = this->m_image_ctx;
 663   if (journal_tid == 0) {
 664     Mutex::Locker cache_locker(image_ctx.cache_lock);
 665     image_ctx.object_cacher->discard_set(image_ctx.object_set,
 666                                          object_extents);
 667   } else {
 668     // cannot discard from cache until journal has committed
 669     assert(image_ctx.journal != NULL);
 670     AioCompletion *aio_comp = this->m_aio_comp;
 671     image_ctx.journal->wait_event(
 672       journal_tid, new C_DiscardJournalCommit<I>(image_ctx, aio_comp,
 673                                                  object_extents, journal_tid));
 674   }
 675 }
 676
 677 template <typename I>
 678 ObjectRequestHandle *ImageDiscardRequest<I>::create_object_request(
 679     const ObjectExtent &object_extent, const ::SnapContext &snapc,
 680     Context *on_finish) {
 681   I &image_ctx = this->m_image_ctx;
 682
 683   ObjectRequest<I> *req;
 684   if (object_extent.length == image_ctx.layout.object_size) {
 685     req = ObjectRequest<I>::create_remove(
 686       &image_ctx, object_extent.oid.name, object_extent.objectno, snapc,
 687       this->m_trace, on_finish);
 688   } else if (object_extent.offset + object_extent.length ==
 689                image_ctx.layout.object_size) {
 690     req = ObjectRequest<I>::create_truncate(
 691       &image_ctx, object_extent.oid.name, object_extent.objectno,
 692       object_extent.offset, snapc, this->m_trace, on_finish);
 693   } else {
 694     req = ObjectRequest<I>::create_zero(
 695       &image_ctx, object_extent.oid.name, object_extent.objectno,
 696       object_extent.offset, object_extent.length, snapc,
 697       this->m_trace, on_finish);
 698   }
 699   return req;
 700 }
 701
 702 template <typename I>
 703 void ImageDiscardRequest<I>::update_stats(size_t length) {
 704   I &image_ctx = this->m_image_ctx;
 705   image_ctx.perfcounter->inc(l_librbd_discard);
 706   image_ctx.perfcounter->inc(l_librbd_discard_bytes, length);
 707 }
 708
 709 template <typename I>
 710 void ImageFlushRequest<I>::send_request() {
 711   I &image_ctx = this->m_image_ctx;
 712   image_ctx.user_flushed();
 713
 714   bool journaling = false;
 715   {
 716     RWLock::RLocker snap_locker(image_ctx.snap_lock);
 717     journaling = (image_ctx.journal != nullptr &&
 718                   image_ctx.journal->is_journal_appending());
 719   }
 720
 721   AioCompletion *aio_comp = this->m_aio_comp;
 722   if (journaling) {
 723     // in-flight ops are flushed prior to closing the journal
 724     uint64_t journal_tid = image_ctx.journal->append_io_event(
 725       journal::EventEntry(journal::AioFlushEvent()),
 726       ObjectRequests(), 0, 0, false);
 727
 728     aio_comp->set_request_count(1);
 729     aio_comp->associate_journal_event(journal_tid);
 730
 731     FunctionContext *flush_ctx = new FunctionContext(
 732       [aio_comp, &image_ctx, journal_tid] (int r) {
 733         auto ctx = new C_FlushJournalCommit<I>(image_ctx, aio_comp,
 734                                                journal_tid);
 735         image_ctx.journal->flush_event(journal_tid, ctx);
 736
 737         // track flush op for block writes
 738         aio_comp->start_op(true);
 739         aio_comp->put();
 740     });
 741
 742     image_ctx.flush_async_operations(flush_ctx);
 743   } else {
 744     // flush rbd cache only when journaling is not enabled
 745     aio_comp->set_request_count(1);
 746     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 747     image_ctx.flush(req_comp);
 748
 749     aio_comp->start_op(true);
 750     aio_comp->put();
 751   }
 752
 753   image_ctx.perfcounter->inc(l_librbd_aio_flush);
 754 }
 755
 756 template <typename I>
 757 void ImageFlushRequest<I>::send_image_cache_request() {
 758   I &image_ctx = this->m_image_ctx;
 759   assert(image_ctx.image_cache != nullptr);
 760
 761   AioCompletion *aio_comp = this->m_aio_comp;
 762   aio_comp->set_request_count(1);
 763   C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 764   image_ctx.image_cache->aio_flush(req_comp);
 765 }
 766
 767 template <typename I>
 768 bool ImageWriteSameRequest<I>::assemble_writesame_extent(const ObjectExtent &object_extent,
 769                                                          bufferlist *bl, bool force_write) {
 770   size_t m_data_len = m_data_bl.length();
 771
 772   if (!force_write) {
 773     bool may_writesame = true;
 774
 775     for (auto q = object_extent.buffer_extents.begin();
 776          q != object_extent.buffer_extents.end(); ++q) {
 777       if (!(q->first % m_data_len == 0 && q->second % m_data_len == 0)) {
 778         may_writesame = false;
 779         break;
 780       }
 781     }
 782
 783     if (may_writesame) {
 784       bl->append(m_data_bl);
 785       return true;
 786     }
 787   }
 788
 789   for (auto q = object_extent.buffer_extents.begin();
 790        q != object_extent.buffer_extents.end(); ++q) {
 791     bufferlist sub_bl;
 792     uint64_t sub_off = q->first % m_data_len;
 793     uint64_t sub_len = m_data_len - sub_off;
 794     uint64_t extent_left = q->second;
 795     while (extent_left >= sub_len) {
 796       sub_bl.substr_of(m_data_bl, sub_off, sub_len);
 797       bl->claim_append(sub_bl);
 798       extent_left -= sub_len;
 799       if (sub_off) {
 800         sub_off = 0;
 801         sub_len = m_data_len;
 802       }
 803     }
 804     if (extent_left) {
 805       sub_bl.substr_of(m_data_bl, sub_off, extent_left);
 806       bl->claim_append(sub_bl);
 807     }
 808   }
 809   return false;
 810 }
 811
 812 template <typename I>
 813 uint64_t ImageWriteSameRequest<I>::append_journal_event(
 814     const ObjectRequests &requests, bool synchronous) {
 815   I &image_ctx = this->m_image_ctx;
 816
 817   uint64_t tid = 0;
 818   assert(!this->m_image_extents.empty());
 819   for (auto &extent : this->m_image_extents) {
 820     journal::EventEntry event_entry(journal::AioWriteSameEvent(extent.first,
 821                                                                extent.second,
 822                                                                m_data_bl));
 823     tid = image_ctx.journal->append_io_event(std::move(event_entry),
 824                                              requests, extent.first,
 825                                              extent.second, synchronous);
 826   }
 827
 828   if (image_ctx.object_cacher == NULL) {
 829     AioCompletion *aio_comp = this->m_aio_comp;
 830     aio_comp->associate_journal_event(tid);
 831   }
 832   return tid;
 833 }
 834
 835 template <typename I>
 836 void ImageWriteSameRequest<I>::send_image_cache_request() {
 837   I &image_ctx = this->m_image_ctx;
 838   assert(image_ctx.image_cache != nullptr);
 839
 840   AioCompletion *aio_comp = this->m_aio_comp;
 841   aio_comp->set_request_count(this->m_image_extents.size());
 842   for (auto &extent : this->m_image_extents) {
 843     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 844     image_ctx.image_cache->aio_writesame(extent.first, extent.second,
 845                                          std::move(m_data_bl), m_op_flags,
 846                                          req_comp);
 847   }
 848 }
 849
 850 template <typename I>
 851 void ImageWriteSameRequest<I>::send_object_cache_requests(
 852     const ObjectExtents &object_extents, uint64_t journal_tid) {
 853   I &image_ctx = this->m_image_ctx;
 854   for (auto p = object_extents.begin(); p != object_extents.end(); ++p) {
 855     const ObjectExtent &object_extent = *p;
 856
 857     bufferlist bl;
 858     assemble_writesame_extent(object_extent, &bl, true);
 859
 860     AioCompletion *aio_comp = this->m_aio_comp;
 861     C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 862     image_ctx.write_to_cache(
 863       object_extent.oid, bl, object_extent.length, object_extent.offset,
 864       req_comp, m_op_flags, journal_tid,
 865       (this->m_trace.valid() ? &this->m_trace : nullptr));
 866   }
 867 }
 868
 869 template <typename I>
 870 void ImageWriteSameRequest<I>::send_object_requests(
 871     const ObjectExtents &object_extents, const ::SnapContext &snapc,
 872     ObjectRequests *object_requests) {
 873   I &image_ctx = this->m_image_ctx;
 874
 875   // cache handles creating object requests during writeback
 876   if (image_ctx.object_cacher == NULL) {
 877     AbstractImageWriteRequest<I>::send_object_requests(object_extents, snapc,
 878                                                        object_requests);
 879   }
 880 }
 881
 882 template <typename I>
 883 ObjectRequestHandle *ImageWriteSameRequest<I>::create_object_request(
 884     const ObjectExtent &object_extent, const ::SnapContext &snapc,
 885     Context *on_finish) {
 886   I &image_ctx = this->m_image_ctx;
 887   assert(image_ctx.object_cacher == NULL);
 888
 889   bufferlist bl;
 890   ObjectRequest<I> *req;
 891
 892   if (assemble_writesame_extent(object_extent, &bl, false)) {
 893     req = ObjectRequest<I>::create_writesame(
 894       &image_ctx, object_extent.oid.name, object_extent.objectno,
 895       object_extent.offset, object_extent.length,
 896       bl, snapc, m_op_flags, this->m_trace, on_finish);
 897     return req;
 898   }
 899   req = ObjectRequest<I>::create_write(
 900     &image_ctx, object_extent.oid.name, object_extent.objectno,
 901     object_extent.offset, bl, snapc, m_op_flags, this->m_trace, on_finish);
 902   return req;
 903 }
 904
 905 template <typename I>
 906 void ImageWriteSameRequest<I>::update_stats(size_t length) {
 907   I &image_ctx = this->m_image_ctx;
 908   image_ctx.perfcounter->inc(l_librbd_ws);
 909   image_ctx.perfcounter->inc(l_librbd_ws_bytes, length);
 910 }
 911
 912 template <typename I>
 913 uint64_t ImageCompareAndWriteRequest<I>::append_journal_event(
 914     const ObjectRequests &requests, bool synchronous) {
 915
 916   I &image_ctx = this->m_image_ctx;
 917
 918   uint64_t tid = 0;
 919   assert(this->m_image_extents.size() == 1);
 920   auto &extent = this->m_image_extents.front();
 921   journal::EventEntry event_entry(journal::AioCompareAndWriteEvent(extent.first,
 922                                                                    extent.second,
 923                                                                    m_cmp_bl, m_bl));
 924   tid = image_ctx.journal->append_io_event(std::move(event_entry),
 925                                            requests, extent.first,
 926                                            extent.second, synchronous);
 927
 928   AioCompletion *aio_comp = this->m_aio_comp;
 929   aio_comp->associate_journal_event(tid);
 930
 931   return tid;
 932 }
 933
 934 template <typename I>
 935 void ImageCompareAndWriteRequest<I>::send_object_cache_requests(
 936   const ObjectExtents &object_extents, uint64_t journal_tid) {
 937   I &image_ctx = this->m_image_ctx;
 938
 939   if (image_ctx.object_cacher != NULL) {
 940     Mutex::Locker cache_locker(image_ctx.cache_lock);
 941     image_ctx.object_cacher->discard_set(image_ctx.object_set,
 942                                          object_extents);
 943   }
 944 }
 945
 946 template <typename I>
 947 void ImageCompareAndWriteRequest<I>::assemble_extent(
 948   const ObjectExtent &object_extent, bufferlist *bl) {
 949   for (auto q = object_extent.buffer_extents.begin();
 950        q != object_extent.buffer_extents.end(); ++q) {
 951     bufferlist sub_bl;
 952     sub_bl.substr_of(m_bl, q->first, q->second);
 953     bl->claim_append(sub_bl);
 954   }
 955 }
 956
 957 template <typename I>
 958 void ImageCompareAndWriteRequest<I>::send_image_cache_request() {
 959   I &image_ctx = this->m_image_ctx;
 960   assert(image_ctx.image_cache != nullptr);
 961
 962   AioCompletion *aio_comp = this->m_aio_comp;
 963   aio_comp->set_request_count(1);
 964   C_AioRequest *req_comp = new C_AioRequest(aio_comp);
 965   image_ctx.image_cache->aio_compare_and_write(
 966     std::move(this->m_image_extents), std::move(m_cmp_bl), std::move(m_bl),
 967     m_mismatch_offset, m_op_flags, req_comp);
 968 }
 969
 970 template <typename I>
 971 ObjectRequestHandle *ImageCompareAndWriteRequest<I>::create_object_request(
 972     const ObjectExtent &object_extent,
 973     const ::SnapContext &snapc,
 974     Context *on_finish) {
 975   I &image_ctx = this->m_image_ctx;
 976
 977   bufferlist bl;
 978   assemble_extent(object_extent, &bl);
 979   ObjectRequest<I> *req = ObjectRequest<I>::create_compare_and_write(
 980                                   &image_ctx, object_extent.oid.name,
 981                                   object_extent.objectno, object_extent.offset,
 982                                   m_cmp_bl, bl, snapc, m_mismatch_offset,
 983                                   m_op_flags, this->m_trace, on_finish);
 984   return req;
 985 }
 986
 987 template <typename I>
 988 void ImageCompareAndWriteRequest<I>::update_stats(size_t length) {
 989   I &image_ctx = this->m_image_ctx;
 990   image_ctx.perfcounter->inc(l_librbd_cmp);
 991   image_ctx.perfcounter->inc(l_librbd_cmp_bytes, length);
 992 }
 993
 994 template <typename I>
 995 int ImageCompareAndWriteRequest<I>::prune_object_extents(ObjectExtents &object_extents) {
 996   if (object_extents.size() > 1)
 997     return -EINVAL;
 998
 999   I &image_ctx = this->m_image_ctx;
1000   uint64_t sector_size = 512ULL;
1001   uint64_t su = image_ctx.layout.stripe_unit;
1002   ObjectExtent object_extent = object_extents.front();
1003   if (object_extent.offset % sector_size + object_extent.length > sector_size ||
1004       (su != 0 && (object_extent.offset % su + object_extent.length > su)))
1005     return -EINVAL;
1006
1007   return 0;
1008 }
1009
1010 } // namespace io
1011 } // namespace librbd
1012
1013 template class librbd::io::ImageRequest<librbd::ImageCtx>;
1014 template class librbd::io::ImageReadRequest<librbd::ImageCtx>;
1015 template class librbd::io::AbstractImageWriteRequest<librbd::ImageCtx>;
1016 template class librbd::io::ImageWriteRequest<librbd::ImageCtx>;
1017 template class librbd::io::ImageDiscardRequest<librbd::ImageCtx>;
1018 template class librbd::io::ImageFlushRequest<librbd::ImageCtx>;
1019 template class librbd::io::ImageWriteSameRequest<librbd::ImageCtx>;
1020 template class librbd::io::ImageCompareAndWriteRequest<librbd::ImageCtx>;