ceph/src/osdc/Journaler.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
   7  *
   8  * This is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License version 2.1, as published by the Free Software
  11  * Foundation.  See file COPYING.
  12  *
  13  */
  14
  15 #include "common/perf_counters.h"
  16 #include "common/dout.h"
  17 #include "include/Context.h"
  18 #include "msg/Messenger.h"
  19 #include "osdc/Journaler.h"
  20 #include "common/errno.h"
  21 #include "include/assert.h"
  22 #include "common/Finisher.h"
  23
  24 #define dout_subsys ceph_subsys_journaler
  25 #undef dout_prefix
  26 #define dout_prefix *_dout << objecter->messenger->get_myname() \
  27   << ".journaler." << name << (readonly ? "(ro) ":"(rw) ")
  28
  29 using std::chrono::seconds;
  30
  31
  32 class Journaler::C_DelayFlush : public Context {
  33   Journaler *journaler;
  34   public:
  35   C_DelayFlush(Journaler *j) : journaler(j) {}
  36   void finish(int r) override {
  37     journaler->_do_delayed_flush();
  38   }
  39 };
  40
  41 void Journaler::set_readonly()
  42 {
  43   lock_guard l(lock);
  44
  45   ldout(cct, 1) << "set_readonly" << dendl;
  46   readonly = true;
  47 }
  48
  49 void Journaler::set_writeable()
  50 {
  51   lock_guard l(lock);
  52
  53   ldout(cct, 1) << "set_writeable" << dendl;
  54   readonly = false;
  55 }
  56
  57 void Journaler::create(file_layout_t *l, stream_format_t const sf)
  58 {
  59   lock_guard lk(lock);
  60
  61   assert(!readonly);
  62   state = STATE_ACTIVE;
  63
  64   stream_format = sf;
  65   journal_stream.set_format(sf);
  66   _set_layout(l);
  67
  68   prezeroing_pos = prezero_pos = write_pos = flush_pos =
  69     safe_pos = read_pos = requested_pos = received_pos =
  70     expire_pos = trimming_pos = trimmed_pos =
  71     next_safe_pos = layout.get_period();
  72
  73   ldout(cct, 1) << "created blank journal at inode 0x" << std::hex << ino
  74                 << std::dec << ", format=" << stream_format << dendl;
  75 }
  76
  77 void Journaler::set_layout(file_layout_t const *l)
  78 {
  79     lock_guard lk(lock);
  80     _set_layout(l);
  81 }
  82
  83 void Journaler::_set_layout(file_layout_t const *l)
  84 {
  85   layout = *l;
  86
  87   assert(layout.pool_id == pg_pool);
  88   last_written.layout = layout;
  89   last_committed.layout = layout;
  90
  91   // prefetch intelligently.
  92   // (watch out, this is big if you use big objects or weird striping)
  93   uint64_t periods = cct->_conf->journaler_prefetch_periods;
  94   if (periods < 2)
  95     periods = 2;  // we need at least 2 periods to make progress.
  96   fetch_len = layout.get_period() * periods;
  97 }
  98
  99
 100 /***************** HEADER *******************/
 101
 102 ostream& operator<<(ostream &out, const Journaler::Header &h)
 103 {
 104   return out << "loghead(trim " << h.trimmed_pos
 105              << ", expire " << h.expire_pos
 106              << ", write " << h.write_pos
 107              << ", stream_format " << (int)(h.stream_format)
 108              << ")";
 109 }
 110
 111 class Journaler::C_ReadHead : public Context {
 112   Journaler *ls;
 113 public:
 114   bufferlist bl;
 115   explicit C_ReadHead(Journaler *l) : ls(l) {}
 116   void finish(int r) override {
 117     ls->_finish_read_head(r, bl);
 118   }
 119 };
 120
 121 class Journaler::C_RereadHead : public Context {
 122   Journaler *ls;
 123   Context *onfinish;
 124 public:
 125   bufferlist bl;
 126   C_RereadHead(Journaler *l, Context *onfinish_) : ls (l),
 127                                                    onfinish(onfinish_) {}
 128   void finish(int r) override {
 129     ls->_finish_reread_head(r, bl, onfinish);
 130   }
 131 };
 132
 133 class Journaler::C_ProbeEnd : public Context {
 134   Journaler *ls;
 135 public:
 136   uint64_t end;
 137   explicit C_ProbeEnd(Journaler *l) : ls(l), end(-1) {}
 138   void finish(int r) override {
 139     ls->_finish_probe_end(r, end);
 140   }
 141 };
 142
 143 class Journaler::C_ReProbe : public Context {
 144   Journaler *ls;
 145   C_OnFinisher *onfinish;
 146 public:
 147   uint64_t end;
 148   C_ReProbe(Journaler *l, C_OnFinisher *onfinish_) :
 149     ls(l), onfinish(onfinish_), end(0) {}
 150   void finish(int r) override {
 151     ls->_finish_reprobe(r, end, onfinish);
 152   }
 153 };
 154
 155 void Journaler::recover(Context *onread)
 156 {
 157   lock_guard l(lock);
 158   if (stopping) {
 159     onread->complete(-EAGAIN);
 160     return;
 161   }
 162
 163   ldout(cct, 1) << "recover start" << dendl;
 164   assert(state != STATE_ACTIVE);
 165   assert(readonly);
 166
 167   if (onread)
 168     waitfor_recover.push_back(wrap_finisher(onread));
 169
 170   if (state != STATE_UNDEF) {
 171     ldout(cct, 1) << "recover - already recovering" << dendl;
 172     return;
 173   }
 174
 175   ldout(cct, 1) << "read_head" << dendl;
 176   state = STATE_READHEAD;
 177   C_ReadHead *fin = new C_ReadHead(this);
 178   _read_head(fin, &fin->bl);
 179 }
 180
 181 void Journaler::_read_head(Context *on_finish, bufferlist *bl)
 182 {
 183   // lock is locked
 184   assert(state == STATE_READHEAD || state == STATE_REREADHEAD);
 185
 186   object_t oid = file_object_t(ino, 0);
 187   object_locator_t oloc(pg_pool);
 188   objecter->read_full(oid, oloc, CEPH_NOSNAP, bl, 0, wrap_finisher(on_finish));
 189 }
 190
 191 void Journaler::reread_head(Context *onfinish)
 192 {
 193   lock_guard l(lock);
 194   _reread_head(wrap_finisher(onfinish));
 195 }
 196
 197 /**
 198  * Re-read the head from disk, and set the write_pos, expire_pos, trimmed_pos
 199  * from the on-disk header. This switches the state to STATE_REREADHEAD for
 200  * the duration, and you shouldn't start a re-read while other operations are
 201  * in-flight, nor start other operations while a re-read is in progress.
 202  * Also, don't call this until the Journaler has finished its recovery and has
 203  * gone STATE_ACTIVE!
 204  */
 205 void Journaler::_reread_head(Context *onfinish)
 206 {
 207   ldout(cct, 10) << "reread_head" << dendl;
 208   assert(state == STATE_ACTIVE);
 209
 210   state = STATE_REREADHEAD;
 211   C_RereadHead *fin = new C_RereadHead(this, onfinish);
 212   _read_head(fin, &fin->bl);
 213 }
 214
 215 void Journaler::_finish_reread_head(int r, bufferlist& bl, Context *finish)
 216 {
 217   lock_guard l(lock);
 218
 219   //read on-disk header into
 220   assert(bl.length() || r < 0 );
 221
 222   // unpack header
 223   if (r == 0) {
 224     Header h;
 225     bufferlist::iterator p = bl.begin();
 226     try {
 227       ::decode(h, p);
 228     } catch (const buffer::error &e) {
 229       finish->complete(-EINVAL);
 230       return;
 231     }
 232     prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos
 233       = h.write_pos;
 234     expire_pos = h.expire_pos;
 235     trimmed_pos = trimming_pos = h.trimmed_pos;
 236     init_headers(h);
 237     state = STATE_ACTIVE;
 238   }
 239
 240   finish->complete(r);
 241 }
 242
 243 void Journaler::_finish_read_head(int r, bufferlist& bl)
 244 {
 245   lock_guard l(lock);
 246
 247   assert(state == STATE_READHEAD);
 248
 249   if (r!=0) {
 250     ldout(cct, 0) << "error getting journal off disk" << dendl;
 251     list<Context*> ls;
 252     ls.swap(waitfor_recover);
 253     finish_contexts(cct, ls, r);
 254     return;
 255   }
 256
 257   if (bl.length() == 0) {
 258     ldout(cct, 1) << "_finish_read_head r=" << r
 259                   << " read 0 bytes, assuming empty log" << dendl;
 260     state = STATE_ACTIVE;
 261     list<Context*> ls;
 262     ls.swap(waitfor_recover);
 263     finish_contexts(cct, ls, 0);
 264     return;
 265   }
 266
 267   // unpack header
 268   bool corrupt = false;
 269   Header h;
 270   bufferlist::iterator p = bl.begin();
 271   try {
 272     ::decode(h, p);
 273
 274     if (h.magic != magic) {
 275       ldout(cct, 0) << "on disk magic '" << h.magic << "' != my magic '"
 276                     << magic << "'" << dendl;
 277       corrupt = true;
 278     } else if (h.write_pos < h.expire_pos || h.expire_pos < h.trimmed_pos) {
 279       ldout(cct, 0) << "Corrupt header (bad offsets): " << h << dendl;
 280       corrupt = true;
 281     }
 282   } catch (const buffer::error &e) {
 283     corrupt = true;
 284   }
 285
 286   if (corrupt) {
 287     list<Context*> ls;
 288     ls.swap(waitfor_recover);
 289     finish_contexts(cct, ls, -EINVAL);
 290     return;
 291   }
 292
 293   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos
 294     = h.write_pos;
 295   read_pos = requested_pos = received_pos = expire_pos = h.expire_pos;
 296   trimmed_pos = trimming_pos = h.trimmed_pos;
 297
 298   init_headers(h);
 299   _set_layout(&h.layout);
 300   stream_format = h.stream_format;
 301   journal_stream.set_format(h.stream_format);
 302
 303   ldout(cct, 1) << "_finish_read_head " << h
 304                 << ".  probing for end of log (from " << write_pos << ")..."
 305                 << dendl;
 306   C_ProbeEnd *fin = new C_ProbeEnd(this);
 307   state = STATE_PROBING;
 308   _probe(fin, &fin->end);
 309 }
 310
 311 void Journaler::_probe(Context *finish, uint64_t *end)
 312 {
 313   // lock is locked
 314   ldout(cct, 1) << "probing for end of the log" << dendl;
 315   assert(state == STATE_PROBING || state == STATE_REPROBING);
 316   // probe the log
 317   filer.probe(ino, &layout, CEPH_NOSNAP,
 318               write_pos, end, true, 0, wrap_finisher(finish));
 319 }
 320
 321 void Journaler::_reprobe(C_OnFinisher *finish)
 322 {
 323   ldout(cct, 10) << "reprobe" << dendl;
 324   assert(state == STATE_ACTIVE);
 325
 326   state = STATE_REPROBING;
 327   C_ReProbe *fin = new C_ReProbe(this, finish);
 328   _probe(fin, &fin->end);
 329 }
 330
 331
 332 void Journaler::_finish_reprobe(int r, uint64_t new_end,
 333                                 C_OnFinisher *onfinish)
 334 {
 335   lock_guard l(lock);
 336
 337   assert(new_end >= write_pos || r < 0);
 338   ldout(cct, 1) << "_finish_reprobe new_end = " << new_end
 339           << " (header had " << write_pos << ")."
 340           << dendl;
 341   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = new_end;
 342   state = STATE_ACTIVE;
 343   onfinish->complete(r);
 344 }
 345
 346 void Journaler::_finish_probe_end(int r, uint64_t end)
 347 {
 348   lock_guard l(lock);
 349
 350   assert(state == STATE_PROBING);
 351   if (r < 0) { // error in probing
 352     goto out;
 353   }
 354   if (((int64_t)end) == -1) {
 355     end = write_pos;
 356     ldout(cct, 1) << "_finish_probe_end write_pos = " << end << " (header had "
 357                   << write_pos << "). log was empty. recovered." << dendl;
 358     ceph_abort(); // hrm.
 359   } else {
 360     assert(end >= write_pos);
 361     ldout(cct, 1) << "_finish_probe_end write_pos = " << end
 362                   << " (header had " << write_pos << "). recovered."
 363                   << dendl;
 364   }
 365
 366   state = STATE_ACTIVE;
 367
 368   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = end;
 369
 370 out:
 371   // done.
 372   list<Context*> ls;
 373   ls.swap(waitfor_recover);
 374   finish_contexts(cct, ls, r);
 375 }
 376
 377 class Journaler::C_RereadHeadProbe : public Context
 378 {
 379   Journaler *ls;
 380   C_OnFinisher *final_finish;
 381 public:
 382   C_RereadHeadProbe(Journaler *l, C_OnFinisher *finish) :
 383     ls(l), final_finish(finish) {}
 384   void finish(int r) override {
 385     ls->_finish_reread_head_and_probe(r, final_finish);
 386   }
 387 };
 388
 389 void Journaler::reread_head_and_probe(Context *onfinish)
 390 {
 391   lock_guard l(lock);
 392
 393   assert(state == STATE_ACTIVE);
 394   _reread_head(new C_RereadHeadProbe(this, wrap_finisher(onfinish)));
 395 }
 396
 397 void Journaler::_finish_reread_head_and_probe(int r, C_OnFinisher *onfinish)
 398 {
 399   // Expect to be called back from finish_reread_head, which already takes lock
 400   // lock is locked
 401
 402   assert(!r); //if we get an error, we're boned
 403   _reprobe(onfinish);
 404 }
 405
 406
 407 // WRITING
 408
 409 class Journaler::C_WriteHead : public Context {
 410 public:
 411   Journaler *ls;
 412   Header h;
 413   C_OnFinisher *oncommit;
 414   C_WriteHead(Journaler *l, Header& h_, C_OnFinisher *c) : ls(l), h(h_),
 415                                                            oncommit(c) {}
 416   void finish(int r) override {
 417     ls->_finish_write_head(r, h, oncommit);
 418   }
 419 };
 420
 421 void Journaler::write_head(Context *oncommit)
 422 {
 423   lock_guard l(lock);
 424   _write_head(oncommit);
 425 }
 426
 427
 428 void Journaler::_write_head(Context *oncommit)
 429 {
 430   assert(!readonly);
 431   assert(state == STATE_ACTIVE);
 432   last_written.trimmed_pos = trimmed_pos;
 433   last_written.expire_pos = expire_pos;
 434   last_written.unused_field = expire_pos;
 435   last_written.write_pos = safe_pos;
 436   last_written.stream_format = stream_format;
 437   ldout(cct, 10) << "write_head " << last_written << dendl;
 438
 439   // Avoid persisting bad pointers in case of bugs
 440   assert(last_written.write_pos >= last_written.expire_pos);
 441   assert(last_written.expire_pos >= last_written.trimmed_pos);
 442
 443   last_wrote_head = ceph::real_clock::now();
 444
 445   bufferlist bl;
 446   ::encode(last_written, bl);
 447   SnapContext snapc;
 448
 449   object_t oid = file_object_t(ino, 0);
 450   object_locator_t oloc(pg_pool);
 451   objecter->write_full(oid, oloc, snapc, bl, ceph::real_clock::now(), 0,
 452                        wrap_finisher(new C_WriteHead(
 453                                              this, last_written,
 454                                              wrap_finisher(oncommit))),
 455                        0, 0, write_iohint);
 456 }
 457
 458 void Journaler::_finish_write_head(int r, Header &wrote,
 459                                    C_OnFinisher *oncommit)
 460 {
 461   lock_guard l(lock);
 462
 463   if (r < 0) {
 464     lderr(cct) << "_finish_write_head got " << cpp_strerror(r) << dendl;
 465     handle_write_error(r);
 466     return;
 467   }
 468   assert(!readonly);
 469   ldout(cct, 10) << "_finish_write_head " << wrote << dendl;
 470   last_committed = wrote;
 471   if (oncommit) {
 472     oncommit->complete(r);
 473   }
 474
 475   _trim();  // trim?
 476 }
 477
 478
 479 /***************** WRITING *******************/
 480
 481 class Journaler::C_Flush : public Context {
 482   Journaler *ls;
 483   uint64_t start;
 484   ceph::real_time stamp;
 485 public:
 486   C_Flush(Journaler *l, int64_t s, ceph::real_time st)
 487     : ls(l), start(s), stamp(st) {}
 488   void finish(int r) override {
 489     ls->_finish_flush(r, start, stamp);
 490   }
 491 };
 492
 493 void Journaler::_finish_flush(int r, uint64_t start, ceph::real_time stamp)
 494 {
 495   lock_guard l(lock);
 496   assert(!readonly);
 497
 498   if (r < 0) {
 499     lderr(cct) << "_finish_flush got " << cpp_strerror(r) << dendl;
 500     handle_write_error(r);
 501     return;
 502   }
 503
 504   assert(start < flush_pos);
 505
 506   // calc latency?
 507   if (logger) {
 508     ceph::timespan lat = ceph::real_clock::now() - stamp;
 509     logger->tinc(logger_key_lat, lat);
 510   }
 511
 512   // adjust safe_pos
 513   auto it = pending_safe.find(start);
 514   assert(it != pending_safe.end());
 515   pending_safe.erase(it);
 516   if (pending_safe.empty())
 517     safe_pos = next_safe_pos;
 518   else
 519     safe_pos = pending_safe.begin()->second;
 520
 521   ldout(cct, 10) << "_finish_flush safe from " << start
 522                  << ", pending_safe " << pending_safe
 523                  << ", (prezeroing/prezero)/write/flush/safe positions now "
 524                  << "(" << prezeroing_pos << "/" << prezero_pos << ")/"
 525                  << write_pos << "/" << flush_pos << "/" << safe_pos
 526                  << dendl;
 527
 528   // kick waiters <= safe_pos
 529   while (!waitfor_safe.empty()) {
 530     if (waitfor_safe.begin()->first > safe_pos)
 531       break;
 532     finish_contexts(cct, waitfor_safe.begin()->second);
 533     waitfor_safe.erase(waitfor_safe.begin());
 534   }
 535 }
 536
 537
 538
 539 uint64_t Journaler::append_entry(bufferlist& bl)
 540 {
 541   unique_lock l(lock);
 542
 543   assert(!readonly);
 544   uint32_t s = bl.length();
 545
 546   // append
 547   size_t delta = bl.length() + journal_stream.get_envelope_size();
 548   // write_buf space is nearly full
 549   if (!write_buf_throttle.get_or_fail(delta)) {
 550     l.unlock();
 551     ldout(cct, 10) << "write_buf_throttle wait, delta " << delta << dendl;
 552     write_buf_throttle.get(delta);
 553     l.lock();
 554   }
 555   ldout(cct, 20) << "write_buf_throttle get, delta " << delta << dendl;
 556   size_t wrote = journal_stream.write(bl, &write_buf, write_pos);
 557   ldout(cct, 10) << "append_entry len " << s << " to " << write_pos << "~"
 558                  << wrote << dendl;
 559   write_pos += wrote;
 560
 561   // flush previous object?
 562   uint64_t su = get_layout_period();
 563   assert(su > 0);
 564   uint64_t write_off = write_pos % su;
 565   uint64_t write_obj = write_pos / su;
 566   uint64_t flush_obj = flush_pos / su;
 567   if (write_obj != flush_obj) {
 568     ldout(cct, 10) << " flushing completed object(s) (su " << su << " wro "
 569                    << write_obj << " flo " << flush_obj << ")" << dendl;
 570     _do_flush(write_buf.length() - write_off);
 571     if (write_off) {
 572       // current entry isn't being flushed, set next_safe_pos to the end of previous entry
 573       next_safe_pos = write_pos - wrote;
 574     }
 575   }
 576
 577   return write_pos;
 578 }
 579
 580
 581 void Journaler::_do_flush(unsigned amount)
 582 {
 583   if (write_pos == flush_pos)
 584     return;
 585   assert(write_pos > flush_pos);
 586   assert(!readonly);
 587
 588   // flush
 589   uint64_t len = write_pos - flush_pos;
 590   assert(len == write_buf.length());
 591   if (amount && amount < len)
 592     len = amount;
 593
 594   // zero at least two full periods ahead.  this ensures
 595   // that the next object will not exist.
 596   uint64_t period = get_layout_period();
 597   if (flush_pos + len + 2*period > prezero_pos) {
 598     _issue_prezero();
 599
 600     int64_t newlen = prezero_pos - flush_pos - period;
 601     if (newlen <= 0) {
 602       ldout(cct, 10) << "_do_flush wanted to do " << flush_pos << "~" << len
 603                      << " already too close to prezero_pos " << prezero_pos
 604                      << ", zeroing first" << dendl;
 605       waiting_for_zero = true;
 606       return;
 607     }
 608     if (static_cast<uint64_t>(newlen) < len) {
 609       ldout(cct, 10) << "_do_flush wanted to do " << flush_pos << "~" << len
 610                      << " but hit prezero_pos " << prezero_pos
 611                      << ", will do " << flush_pos << "~" << newlen << dendl;
 612       len = newlen;
 613     } else {
 614       waiting_for_zero = false;
 615     }
 616   } else {
 617     waiting_for_zero = false;
 618   }
 619   ldout(cct, 10) << "_do_flush flushing " << flush_pos << "~" << len << dendl;
 620
 621   // submit write for anything pending
 622   // flush _start_ pos to _finish_flush
 623   ceph::real_time now = ceph::real_clock::now();
 624   SnapContext snapc;
 625
 626   Context *onsafe = new C_Flush(this, flush_pos, now);  // on COMMIT
 627   pending_safe[flush_pos] = next_safe_pos;
 628
 629   bufferlist write_bl;
 630
 631   // adjust pointers
 632   if (len == write_buf.length()) {
 633     write_bl.swap(write_buf);
 634     next_safe_pos = write_pos;
 635   } else {
 636     write_buf.splice(0, len, &write_bl);
 637   }
 638
 639   filer.write(ino, &layout, snapc,
 640               flush_pos, len, write_bl, ceph::real_clock::now(),
 641               0,
 642               wrap_finisher(onsafe), write_iohint);
 643
 644   flush_pos += len;
 645   assert(write_buf.length() == write_pos - flush_pos);
 646   write_buf_throttle.put(len);
 647   ldout(cct, 20) << "write_buf_throttle put, len " << len << dendl;
 648
 649   ldout(cct, 10)
 650     << "_do_flush (prezeroing/prezero)/write/flush/safe pointers now at "
 651     << "(" << prezeroing_pos << "/" << prezero_pos << ")/" << write_pos
 652     << "/" << flush_pos << "/" << safe_pos << dendl;
 653
 654   _issue_prezero();
 655 }
 656
 657
 658 void Journaler::wait_for_flush(Context *onsafe)
 659 {
 660   lock_guard l(lock);
 661   if (stopping) {
 662     onsafe->complete(-EAGAIN);
 663     return;
 664   }
 665   _wait_for_flush(onsafe);
 666 }
 667
 668 void Journaler::_wait_for_flush(Context *onsafe)
 669 {
 670   assert(!readonly);
 671
 672   // all flushed and safe?
 673   if (write_pos == safe_pos) {
 674     assert(write_buf.length() == 0);
 675     ldout(cct, 10)
 676       << "flush nothing to flush, (prezeroing/prezero)/write/flush/safe "
 677       "pointers at " << "(" << prezeroing_pos << "/" << prezero_pos << ")/"
 678       << write_pos << "/" << flush_pos << "/" << safe_pos << dendl;
 679     if (onsafe) {
 680       finisher->queue(onsafe, 0);
 681     }
 682     return;
 683   }
 684
 685   // queue waiter
 686   if (onsafe) {
 687     waitfor_safe[write_pos].push_back(wrap_finisher(onsafe));
 688   }
 689 }
 690
 691 void Journaler::flush(Context *onsafe)
 692 {
 693   lock_guard l(lock);
 694   _flush(wrap_finisher(onsafe));
 695 }
 696
 697 void Journaler::_flush(C_OnFinisher *onsafe)
 698 {
 699   assert(!readonly);
 700
 701   if (write_pos == flush_pos) {
 702     assert(write_buf.length() == 0);
 703     ldout(cct, 10) << "flush nothing to flush, (prezeroing/prezero)/write/"
 704       "flush/safe pointers at " << "(" << prezeroing_pos << "/" << prezero_pos
 705                    << ")/" << write_pos << "/" << flush_pos << "/" << safe_pos
 706                    << dendl;
 707     if (onsafe) {
 708       onsafe->complete(0);
 709     }
 710   } else {
 711     _do_flush();
 712     _wait_for_flush(onsafe);
 713   }
 714
 715   // write head?
 716   if (_write_head_needed()) {
 717     _write_head();
 718   }
 719 }
 720
 721 bool Journaler::_write_head_needed()
 722 {
 723   return last_wrote_head + seconds(cct->_conf->journaler_write_head_interval)
 724       < ceph::real_clock::now();
 725 }
 726
 727
 728 /*************** prezeroing ******************/
 729
 730 struct C_Journaler_Prezero : public Context {
 731   Journaler *journaler;
 732   uint64_t from, len;
 733   C_Journaler_Prezero(Journaler *j, uint64_t f, uint64_t l)
 734     : journaler(j), from(f), len(l) {}
 735   void finish(int r) override {
 736     journaler->_finish_prezero(r, from, len);
 737   }
 738 };
 739
 740 void Journaler::_issue_prezero()
 741 {
 742   assert(prezeroing_pos >= flush_pos);
 743
 744   // we need to zero at least two periods, minimum, to ensure that we
 745   // have a full empty object/period in front of us.
 746   uint64_t num_periods = MAX(2, cct->_conf->journaler_prezero_periods);
 747
 748   /*
 749    * issue zero requests based on write_pos, even though the invariant
 750    * is that we zero ahead of flush_pos.
 751    */
 752   uint64_t period = get_layout_period();
 753   uint64_t to = write_pos + period * num_periods  + period - 1;
 754   to -= to % period;
 755
 756   if (prezeroing_pos >= to) {
 757     ldout(cct, 20) << "_issue_prezero target " << to << " <= prezeroing_pos "
 758                    << prezeroing_pos << dendl;
 759     return;
 760   }
 761
 762   while (prezeroing_pos < to) {
 763     uint64_t len;
 764     if (prezeroing_pos % period == 0) {
 765       len = period;
 766       ldout(cct, 10) << "_issue_prezero removing " << prezeroing_pos << "~"
 767                      << period << " (full period)" << dendl;
 768     } else {
 769       len = period - (prezeroing_pos % period);
 770       ldout(cct, 10) << "_issue_prezero zeroing " << prezeroing_pos << "~"
 771                      << len << " (partial period)" << dendl;
 772     }
 773     SnapContext snapc;
 774     Context *c = wrap_finisher(new C_Journaler_Prezero(this, prezeroing_pos,
 775                                                        len));
 776     filer.zero(ino, &layout, snapc, prezeroing_pos, len,
 777                ceph::real_clock::now(), 0, c);
 778     prezeroing_pos += len;
 779   }
 780 }
 781
 782 // Lock cycle because we get called out of objecter callback (holding
 783 // objecter read lock), but there are also cases where we take the journaler
 784 // lock before calling into objecter to do I/O.
 785 void Journaler::_finish_prezero(int r, uint64_t start, uint64_t len)
 786 {
 787   lock_guard l(lock);
 788
 789   ldout(cct, 10) << "_prezeroed to " << start << "~" << len
 790                  << ", prezeroing/prezero was " << prezeroing_pos << "/"
 791                  << prezero_pos << ", pending " << pending_zero
 792                  << dendl;
 793   if (r < 0 && r != -ENOENT) {
 794     lderr(cct) << "_prezeroed got " << cpp_strerror(r) << dendl;
 795     handle_write_error(r);
 796     return;
 797   }
 798
 799   assert(r == 0 || r == -ENOENT);
 800
 801   if (start == prezero_pos) {
 802     prezero_pos += len;
 803     while (!pending_zero.empty() &&
 804            pending_zero.begin().get_start() == prezero_pos) {
 805       interval_set<uint64_t>::iterator b(pending_zero.begin());
 806       prezero_pos += b.get_len();
 807       pending_zero.erase(b);
 808     }
 809
 810     if (waiting_for_zero) {
 811       _do_flush();
 812     }
 813   } else {
 814     pending_zero.insert(start, len);
 815   }
 816   ldout(cct, 10) << "_prezeroed prezeroing/prezero now " << prezeroing_pos
 817                  << "/" << prezero_pos
 818                  << ", pending " << pending_zero
 819                  << dendl;
 820 }
 821
 822
 823
 824 /***************** READING *******************/
 825
 826
 827 class Journaler::C_Read : public Context {
 828   Journaler *ls;
 829   uint64_t offset;
 830   uint64_t length;
 831 public:
 832   bufferlist bl;
 833   C_Read(Journaler *j, uint64_t o, uint64_t l) : ls(j), offset(o), length(l) {}
 834   void finish(int r) override {
 835     ls->_finish_read(r, offset, length, bl);
 836   }
 837 };
 838
 839 class Journaler::C_RetryRead : public Context {
 840   Journaler *ls;
 841 public:
 842   explicit C_RetryRead(Journaler *l) : ls(l) {}
 843
 844   void finish(int r) override {
 845     // Should only be called from waitfor_safe i.e. already inside lock
 846     // (ls->lock is locked
 847     ls->_prefetch();
 848   }
 849 };
 850
 851 void Journaler::_finish_read(int r, uint64_t offset, uint64_t length,
 852                              bufferlist& bl)
 853 {
 854   lock_guard l(lock);
 855
 856   if (r < 0) {
 857     ldout(cct, 0) << "_finish_read got error " << r << dendl;
 858     error = r;
 859   } else {
 860     ldout(cct, 10) << "_finish_read got " << offset << "~" << bl.length()
 861                    << dendl;
 862     if (bl.length() < length) {
 863       ldout(cct, 0) << "_finish_read got less than expected (" << length << ")"
 864                     << dendl;
 865       error = -EINVAL;
 866     }
 867   }
 868
 869   if (error) {
 870     if (on_readable) {
 871       C_OnFinisher *f = on_readable;
 872       on_readable = 0;
 873       f->complete(error);
 874     }
 875     return;
 876   }
 877
 878   prefetch_buf[offset].swap(bl);
 879
 880   try {
 881     _assimilate_prefetch();
 882   } catch (const buffer::error &err) {
 883     lderr(cct) << "_decode error from assimilate_prefetch" << dendl;
 884     error = -EINVAL;
 885     if (on_readable) {
 886       C_OnFinisher *f = on_readable;
 887       on_readable = 0;
 888       f->complete(error);
 889     }
 890     return;
 891   }
 892   _prefetch();
 893 }
 894
 895 void Journaler::_assimilate_prefetch()
 896 {
 897   bool was_readable = readable;
 898
 899   bool got_any = false;
 900   while (!prefetch_buf.empty()) {
 901     map<uint64_t,bufferlist>::iterator p = prefetch_buf.begin();
 902     if (p->first != received_pos) {
 903       uint64_t gap = p->first - received_pos;
 904       ldout(cct, 10) << "_assimilate_prefetch gap of " << gap
 905                      << " from received_pos " << received_pos
 906                      << " to first prefetched buffer " << p->first << dendl;
 907       break;
 908     }
 909
 910     ldout(cct, 10) << "_assimilate_prefetch " << p->first << "~"
 911                    << p->second.length() << dendl;
 912     received_pos += p->second.length();
 913     read_buf.claim_append(p->second);
 914     assert(received_pos <= requested_pos);
 915     prefetch_buf.erase(p);
 916     got_any = true;
 917   }
 918
 919   if (got_any) {
 920     ldout(cct, 10) << "_assimilate_prefetch read_buf now " << read_pos << "~"
 921                    << read_buf.length() << ", read pointers " << read_pos
 922                    << "/" << received_pos << "/" << requested_pos
 923                    << dendl;
 924
 925     // Update readability (this will also hit any decode errors resulting
 926     // from bad data)
 927     readable = _is_readable();
 928   }
 929
 930   if ((got_any && !was_readable && readable) || read_pos == write_pos) {
 931     // readable!
 932     ldout(cct, 10) << "_finish_read now readable (or at journal end) readable="
 933                    << readable << " read_pos=" << read_pos << " write_pos="
 934                    << write_pos << dendl;
 935     if (on_readable) {
 936       C_OnFinisher *f = on_readable;
 937       on_readable = 0;
 938       f->complete(0);
 939     }
 940   }
 941 }
 942
 943 void Journaler::_issue_read(uint64_t len)
 944 {
 945   // stuck at safe_pos?  (this is needed if we are reading the tail of
 946   // a journal we are also writing to)
 947   assert(requested_pos <= safe_pos);
 948   if (requested_pos == safe_pos) {
 949     ldout(cct, 10) << "_issue_read requested_pos = safe_pos = " << safe_pos
 950                    << ", waiting" << dendl;
 951     assert(write_pos > requested_pos);
 952     if (pending_safe.empty()) {
 953       _flush(NULL);
 954     }
 955     waitfor_safe[flush_pos].push_back(new C_RetryRead(this));
 956     return;
 957   }
 958
 959   // don't read too much
 960   if (requested_pos + len > safe_pos) {
 961     len = safe_pos - requested_pos;
 962     ldout(cct, 10) << "_issue_read reading only up to safe_pos " << safe_pos
 963                    << dendl;
 964   }
 965
 966   // go.
 967   ldout(cct, 10) << "_issue_read reading " << requested_pos << "~" << len
 968                  << ", read pointers " << read_pos << "/" << received_pos
 969                  << "/" << (requested_pos+len) << dendl;
 970
 971   // step by period (object).  _don't_ do a single big filer.read()
 972   // here because it will wait for all object reads to complete before
 973   // giving us back any data.  this way we can process whatever bits
 974   // come in that are contiguous.
 975   uint64_t period = get_layout_period();
 976   while (len > 0) {
 977     uint64_t e = requested_pos + period;
 978     e -= e % period;
 979     uint64_t l = e - requested_pos;
 980     if (l > len)
 981       l = len;
 982     C_Read *c = new C_Read(this, requested_pos, l);
 983     filer.read(ino, &layout, CEPH_NOSNAP, requested_pos, l, &c->bl, 0,
 984                wrap_finisher(c), CEPH_OSD_OP_FLAG_FADVISE_DONTNEED);
 985     requested_pos += l;
 986     len -= l;
 987   }
 988 }
 989
 990 void Journaler::_prefetch()
 991 {
 992   ldout(cct, 10) << "_prefetch" << dendl;
 993   // prefetch
 994   uint64_t pf;
 995   if (temp_fetch_len) {
 996     ldout(cct, 10) << "_prefetch temp_fetch_len " << temp_fetch_len << dendl;
 997     pf = temp_fetch_len;
 998     temp_fetch_len = 0;
 999   } else {
1000     pf = fetch_len;
1001   }
1002
1003   uint64_t raw_target = read_pos + pf;
1004
1005   // read full log segments, so increase if necessary
1006   uint64_t period = get_layout_period();
1007   uint64_t remainder = raw_target % period;
1008   uint64_t adjustment = remainder ? period - remainder : 0;
1009   uint64_t target = raw_target + adjustment;
1010
1011   // don't read past the log tail
1012   if (target > write_pos)
1013     target = write_pos;
1014
1015   if (requested_pos < target) {
1016     uint64_t len = target - requested_pos;
1017     ldout(cct, 10) << "_prefetch " << pf << " requested_pos " << requested_pos
1018                    << " < target " << target << " (" << raw_target
1019                    << "), prefetching " << len << dendl;
1020
1021     if (pending_safe.empty() && write_pos > safe_pos) {
1022       // If we are reading and writing the journal, then we may need
1023       // to issue a flush if one isn't already in progress.
1024       // Avoid doing a flush every time so that if we do write/read/write/read
1025       // we don't end up flushing after every write.
1026       ldout(cct, 10) << "_prefetch: requested_pos=" << requested_pos
1027                      << ", read_pos=" << read_pos
1028                      << ", write_pos=" << write_pos
1029                      << ", safe_pos=" << safe_pos << dendl;
1030       _do_flush();
1031     }
1032
1033     _issue_read(len);
1034   }
1035 }
1036
1037
1038 /*
1039  * _is_readable() - return true if next entry is ready.
1040  */
1041 bool Journaler::_is_readable()
1042 {
1043   // anything to read?
1044   if (read_pos == write_pos)
1045     return false;
1046
1047   // Check if the retrieve bytestream has enough for an entry
1048   uint64_t need;
1049   if (journal_stream.readable(read_buf, &need)) {
1050     return true;
1051   }
1052
1053   ldout (cct, 10) << "_is_readable read_buf.length() == " << read_buf.length()
1054                   << ", but need " << need << " for next entry; fetch_len is "
1055                   << fetch_len << dendl;
1056
1057   // partial fragment at the end?
1058   if (received_pos == write_pos) {
1059     ldout(cct, 10) << "is_readable() detected partial entry at tail, "
1060       "adjusting write_pos to " << read_pos << dendl;
1061
1062     // adjust write_pos
1063     prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = read_pos;
1064     assert(write_buf.length() == 0);
1065
1066     // reset read state
1067     requested_pos = received_pos = read_pos;
1068     read_buf.clear();
1069
1070     // FIXME: truncate on disk?
1071
1072     return false;
1073   }
1074
1075   if (need > fetch_len) {
1076     temp_fetch_len = need;
1077     ldout(cct, 10) << "_is_readable noting temp_fetch_len " << temp_fetch_len
1078                    << dendl;
1079   }
1080
1081   ldout(cct, 10) << "_is_readable: not readable, returning false" << dendl;
1082   return false;
1083 }
1084
1085 /*
1086  * is_readable() - kickstart prefetch, too
1087  */
1088 bool Journaler::is_readable()
1089 {
1090   lock_guard l(lock);
1091
1092   if (error != 0) {
1093     return false;
1094   }
1095
1096   bool r = readable;
1097   _prefetch();
1098   return r;
1099 }
1100
1101 class Journaler::C_EraseFinish : public Context {
1102   Journaler *journaler;
1103   C_OnFinisher *completion;
1104   public:
1105   C_EraseFinish(Journaler *j, C_OnFinisher *c) : journaler(j), completion(c) {}
1106   void finish(int r) override {
1107     journaler->_finish_erase(r, completion);
1108   }
1109 };
1110
1111 /**
1112  * Entirely erase the journal, including header.  For use when you
1113  * have already made a copy of the journal somewhere else.
1114  */
1115 void Journaler::erase(Context *completion)
1116 {
1117   lock_guard l(lock);
1118
1119   // Async delete the journal data
1120   uint64_t first = trimmed_pos / get_layout_period();
1121   uint64_t num = (write_pos - trimmed_pos) / get_layout_period() + 2;
1122   filer.purge_range(ino, &layout, SnapContext(), first, num,
1123                     ceph::real_clock::now(), 0,
1124                     wrap_finisher(new C_EraseFinish(
1125                                     this, wrap_finisher(completion))));
1126
1127   // We will not start the operation to delete the header until
1128   // _finish_erase has seen the data deletion succeed: otherwise if
1129   // there was an error deleting data we might prematurely delete the
1130   // header thereby lose our reference to the data.
1131 }
1132
1133 void Journaler::_finish_erase(int data_result, C_OnFinisher *completion)
1134 {
1135   lock_guard l(lock);
1136
1137   if (data_result == 0) {
1138     // Async delete the journal header
1139     filer.purge_range(ino, &layout, SnapContext(), 0, 1,
1140                       ceph::real_clock::now(),
1141                       0, wrap_finisher(completion));
1142   } else {
1143     lderr(cct) << "Failed to delete journal " << ino << " data: "
1144                << cpp_strerror(data_result) << dendl;
1145     completion->complete(data_result);
1146   }
1147 }
1148
1149 /* try_read_entry(bl)
1150  *  read entry into bl if it's ready.
1151  *  otherwise, do nothing.
1152  */
1153 bool Journaler::try_read_entry(bufferlist& bl)
1154 {
1155   lock_guard l(lock);
1156
1157   if (!readable) {
1158     ldout(cct, 10) << "try_read_entry at " << read_pos << " not readable"
1159                    << dendl;
1160     return false;
1161   }
1162
1163   uint64_t start_ptr;
1164   size_t consumed;
1165   try {
1166     consumed = journal_stream.read(read_buf, &bl, &start_ptr);
1167     if (stream_format >= JOURNAL_FORMAT_RESILIENT) {
1168       assert(start_ptr == read_pos);
1169     }
1170   } catch (const buffer::error &e) {
1171     lderr(cct) << __func__ << ": decode error from journal_stream" << dendl;
1172     error = -EINVAL;
1173     return false;
1174   }
1175
1176   ldout(cct, 10) << "try_read_entry at " << read_pos << " read "
1177                  << read_pos << "~" << consumed << " (have "
1178                  << read_buf.length() << ")" << dendl;
1179
1180   read_pos += consumed;
1181   try {
1182     // We were readable, we might not be any more
1183     readable = _is_readable();
1184   } catch (const buffer::error &e) {
1185     lderr(cct) << __func__ << ": decode error from _is_readable" << dendl;
1186     error = -EINVAL;
1187     return false;
1188   }
1189
1190   // prefetch?
1191   _prefetch();
1192   return true;
1193 }
1194
1195 void Journaler::wait_for_readable(Context *onreadable)
1196 {
1197   lock_guard l(lock);
1198   if (stopping) {
1199     onreadable->complete(-EAGAIN);
1200     return;
1201   }
1202
1203   assert(on_readable == 0);
1204   if (!readable) {
1205     ldout(cct, 10) << "wait_for_readable at " << read_pos << " onreadable "
1206                    << onreadable << dendl;
1207     on_readable = wrap_finisher(onreadable);
1208   } else {
1209     // race with OSD reply
1210     finisher->queue(onreadable, 0);
1211   }
1212 }
1213
1214 bool Journaler::have_waiter() const
1215 {
1216   return on_readable != nullptr;
1217 }
1218
1219
1220
1221
1222 /***************** TRIMMING *******************/
1223
1224
1225 class Journaler::C_Trim : public Context {
1226   Journaler *ls;
1227   uint64_t to;
1228 public:
1229   C_Trim(Journaler *l, int64_t t) : ls(l), to(t) {}
1230   void finish(int r) override {
1231     ls->_finish_trim(r, to);
1232   }
1233 };
1234
1235 void Journaler::trim()
1236 {
1237   lock_guard l(lock);
1238   _trim();
1239 }
1240
1241 void Journaler::_trim()
1242 {
1243   assert(!readonly);
1244   uint64_t period = get_layout_period();
1245   uint64_t trim_to = last_committed.expire_pos;
1246   trim_to -= trim_to % period;
1247   ldout(cct, 10) << "trim last_commited head was " << last_committed
1248            << ", can trim to " << trim_to
1249            << dendl;
1250   if (trim_to == 0 || trim_to == trimming_pos) {
1251     ldout(cct, 10) << "trim already trimmed/trimming to "
1252                    << trimmed_pos << "/" << trimming_pos << dendl;
1253     return;
1254   }
1255
1256   if (trimming_pos > trimmed_pos) {
1257     ldout(cct, 10) << "trim already trimming atm, try again later.  "
1258       "trimmed/trimming is " << trimmed_pos << "/" << trimming_pos << dendl;
1259     return;
1260   }
1261
1262   // trim
1263   assert(trim_to <= write_pos);
1264   assert(trim_to <= expire_pos);
1265   assert(trim_to > trimming_pos);
1266   ldout(cct, 10) << "trim trimming to " << trim_to
1267                  << ", trimmed/trimming/expire are "
1268                  << trimmed_pos << "/" << trimming_pos << "/" << expire_pos
1269                  << dendl;
1270
1271   // delete range of objects
1272   uint64_t first = trimming_pos / period;
1273   uint64_t num = (trim_to - trimming_pos) / period;
1274   SnapContext snapc;
1275   filer.purge_range(ino, &layout, snapc, first, num,
1276                     ceph::real_clock::now(), 0,
1277                     wrap_finisher(new C_Trim(this, trim_to)));
1278   trimming_pos = trim_to;
1279 }
1280
1281 void Journaler::_finish_trim(int r, uint64_t to)
1282 {
1283   lock_guard l(lock);
1284
1285   assert(!readonly);
1286   ldout(cct, 10) << "_finish_trim trimmed_pos was " << trimmed_pos
1287            << ", trimmed/trimming/expire now "
1288            << to << "/" << trimming_pos << "/" << expire_pos
1289            << dendl;
1290   if (r < 0 && r != -ENOENT) {
1291     lderr(cct) << "_finish_trim got " << cpp_strerror(r) << dendl;
1292     handle_write_error(r);
1293     return;
1294   }
1295
1296   assert(r >= 0 || r == -ENOENT);
1297
1298   assert(to <= trimming_pos);
1299   assert(to > trimmed_pos);
1300   trimmed_pos = to;
1301 }
1302
1303 void Journaler::handle_write_error(int r)
1304 {
1305   // lock is locked
1306
1307   lderr(cct) << "handle_write_error " << cpp_strerror(r) << dendl;
1308   if (on_write_error) {
1309     on_write_error->complete(r);
1310     on_write_error = NULL;
1311     called_write_error = true;
1312   } else if (called_write_error) {
1313     /* We don't call error handler more than once, subsequent errors
1314      * are dropped -- this is okay as long as the error handler does
1315      * something dramatic like respawn */
1316     lderr(cct) << __func__ << ": multiple write errors, handler already called"
1317                << dendl;
1318   } else {
1319     assert(0 == "unhandled write error");
1320   }
1321 }
1322
1323
1324 /**
1325  * Test whether the 'read_buf' byte stream has enough data to read
1326  * an entry
1327  *
1328  * sets 'next_envelope_size' to the number of bytes needed to advance (enough
1329  * to get the next header if header was unavailable, or enough to get the whole
1330  * next entry if the header was available but the body wasn't).
1331  */
1332 bool JournalStream::readable(bufferlist &read_buf, uint64_t *need) const
1333 {
1334   assert(need != NULL);
1335
1336   uint32_t entry_size = 0;
1337   uint64_t entry_sentinel = 0;
1338   bufferlist::iterator p = read_buf.begin();
1339
1340   // Do we have enough data to decode an entry prefix?
1341   if (format >= JOURNAL_FORMAT_RESILIENT) {
1342     *need = sizeof(entry_size) + sizeof(entry_sentinel);
1343   } else {
1344     *need = sizeof(entry_size);
1345   }
1346   if (read_buf.length() >= *need) {
1347     if (format >= JOURNAL_FORMAT_RESILIENT) {
1348       ::decode(entry_sentinel, p);
1349       if (entry_sentinel != sentinel) {
1350         throw buffer::malformed_input("Invalid sentinel");
1351       }
1352     }
1353
1354     ::decode(entry_size, p);
1355   } else {
1356     return false;
1357   }
1358
1359   // Do we have enough data to decode an entry prefix, payload and suffix?
1360   if (format >= JOURNAL_FORMAT_RESILIENT) {
1361     *need = JOURNAL_ENVELOPE_RESILIENT + entry_size;
1362   } else {
1363     *need = JOURNAL_ENVELOPE_LEGACY + entry_size;
1364   }
1365   if (read_buf.length() >= *need) {
1366     return true;  // No more bytes needed
1367   }
1368
1369   return false;
1370 }
1371
1372
1373 /**
1374  * Consume one entry from a journal byte stream 'from', splicing a
1375  * serialized LogEvent blob into 'entry'.
1376  *
1377  * 'entry' must be non null and point to an empty bufferlist.
1378  *
1379  * 'from' must contain sufficient valid data (i.e. readable is true).
1380  *
1381  * 'start_ptr' will be set to the entry's start pointer, if the collection
1382  * format provides it.  It may not be null.
1383  *
1384  * @returns The number of bytes consumed from the `from` byte stream.  Note
1385  *          that this is not equal to the length of `entry`, which contains
1386  *          the inner serialized LogEvent and not the envelope.
1387  */
1388 size_t JournalStream::read(bufferlist &from, bufferlist *entry,
1389                            uint64_t *start_ptr)
1390 {
1391   assert(start_ptr != NULL);
1392   assert(entry != NULL);
1393   assert(entry->length() == 0);
1394
1395   uint32_t entry_size = 0;
1396
1397   // Consume envelope prefix: entry_size and entry_sentinel
1398   bufferlist::iterator from_ptr = from.begin();
1399   if (format >= JOURNAL_FORMAT_RESILIENT) {
1400     uint64_t entry_sentinel = 0;
1401     ::decode(entry_sentinel, from_ptr);
1402     // Assertion instead of clean check because of precondition of this
1403     // fn is that readable() already passed
1404     assert(entry_sentinel == sentinel);
1405   }
1406   ::decode(entry_size, from_ptr);
1407
1408   // Read out the payload
1409   from_ptr.copy(entry_size, *entry);
1410
1411   // Consume the envelope suffix (start_ptr)
1412   if (format >= JOURNAL_FORMAT_RESILIENT) {
1413     ::decode(*start_ptr, from_ptr);
1414   } else {
1415     *start_ptr = 0;
1416   }
1417
1418   // Trim the input buffer to discard the bytes we have consumed
1419   from.splice(0, from_ptr.get_off());
1420
1421   return from_ptr.get_off();
1422 }
1423
1424
1425 /**
1426  * Append one entry
1427  */
1428 size_t JournalStream::write(bufferlist &entry, bufferlist *to,
1429                             uint64_t const &start_ptr)
1430 {
1431   assert(to != NULL);
1432
1433   uint32_t const entry_size = entry.length();
1434   if (format >= JOURNAL_FORMAT_RESILIENT) {
1435     ::encode(sentinel, *to);
1436   }
1437   ::encode(entry_size, *to);
1438   to->claim_append(entry);
1439   if (format >= JOURNAL_FORMAT_RESILIENT) {
1440     ::encode(start_ptr, *to);
1441   }
1442
1443   if (format >= JOURNAL_FORMAT_RESILIENT) {
1444     return JOURNAL_ENVELOPE_RESILIENT + entry_size;
1445   } else {
1446     return JOURNAL_ENVELOPE_LEGACY + entry_size;
1447   }
1448 }
1449
1450 /**
1451  * set write error callback
1452  *
1453  * Set a callback/context to trigger if we get a write error from
1454  * the objecter.  This may be from an explicit request (e.g., flush)
1455  * or something async the journaler did on its own (e.g., journal
1456  * header update).
1457  *
1458  * It is only used once; if the caller continues to use the
1459  * Journaler and wants to hear about errors, it needs to reset the
1460  * error_handler.
1461  *
1462  * @param c callback/context to trigger on error
1463  */
1464 void Journaler::set_write_error_handler(Context *c) {
1465   lock_guard l(lock);
1466   assert(!on_write_error);
1467   on_write_error = wrap_finisher(c);
1468   called_write_error = false;
1469 }
1470
1471
1472 /**
1473  * Wrap a context in a C_OnFinisher, if it is non-NULL
1474  *
1475  * Utility function to avoid lots of error-prone and verbose
1476  * NULL checking on contexts passed in.
1477  */
1478 C_OnFinisher *Journaler::wrap_finisher(Context *c)
1479 {
1480   if (c != NULL) {
1481     return new C_OnFinisher(c, finisher);
1482   } else {
1483     return NULL;
1484   }
1485 }
1486
1487 void Journaler::shutdown()
1488 {
1489   lock_guard l(lock);
1490
1491   ldout(cct, 1) << __func__ << dendl;
1492
1493   readable = false;
1494   stopping = true;
1495
1496   // Kick out anyone reading from journal
1497   error = -EAGAIN;
1498   if (on_readable) {
1499     C_OnFinisher *f = on_readable;
1500     on_readable = 0;
1501     f->complete(-EAGAIN);
1502   }
1503
1504   finish_contexts(cct, waitfor_recover, -ESHUTDOWN);
1505
1506   std::map<uint64_t, std::list<Context*> >::iterator i;
1507   for (i = waitfor_safe.begin(); i != waitfor_safe.end(); ++i) {
1508     finish_contexts(cct, i->second, -EAGAIN);
1509   }
1510   waitfor_safe.clear();
1511 }
1512