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   if (layout.pool_id != pg_pool) {
  88     // user can reset pool id through cephfs-journal-tool
  89     lderr(cct) << "may got older pool id from header layout" << dendl;
  90     ceph_abort();
  91   }
  92   last_written.layout = layout;
  93   last_committed.layout = layout;
  94
  95   // prefetch intelligently.
  96   // (watch out, this is big if you use big objects or weird striping)
  97   uint64_t periods = cct->_conf->journaler_prefetch_periods;
  98   if (periods < 2)
  99     periods = 2;  // we need at least 2 periods to make progress.
 100   fetch_len = layout.get_period() * periods;
 101 }
 102
 103
 104 /***************** HEADER *******************/
 105
 106 ostream& operator<<(ostream &out, const Journaler::Header &h)
 107 {
 108   return out << "loghead(trim " << h.trimmed_pos
 109              << ", expire " << h.expire_pos
 110              << ", write " << h.write_pos
 111              << ", stream_format " << (int)(h.stream_format)
 112              << ")";
 113 }
 114
 115 class Journaler::C_ReadHead : public Context {
 116   Journaler *ls;
 117 public:
 118   bufferlist bl;
 119   explicit C_ReadHead(Journaler *l) : ls(l) {}
 120   void finish(int r) override {
 121     ls->_finish_read_head(r, bl);
 122   }
 123 };
 124
 125 class Journaler::C_RereadHead : public Context {
 126   Journaler *ls;
 127   Context *onfinish;
 128 public:
 129   bufferlist bl;
 130   C_RereadHead(Journaler *l, Context *onfinish_) : ls (l),
 131                                                    onfinish(onfinish_) {}
 132   void finish(int r) override {
 133     ls->_finish_reread_head(r, bl, onfinish);
 134   }
 135 };
 136
 137 class Journaler::C_ProbeEnd : public Context {
 138   Journaler *ls;
 139 public:
 140   uint64_t end;
 141   explicit C_ProbeEnd(Journaler *l) : ls(l), end(-1) {}
 142   void finish(int r) override {
 143     ls->_finish_probe_end(r, end);
 144   }
 145 };
 146
 147 class Journaler::C_ReProbe : public Context {
 148   Journaler *ls;
 149   C_OnFinisher *onfinish;
 150 public:
 151   uint64_t end;
 152   C_ReProbe(Journaler *l, C_OnFinisher *onfinish_) :
 153     ls(l), onfinish(onfinish_), end(0) {}
 154   void finish(int r) override {
 155     ls->_finish_reprobe(r, end, onfinish);
 156   }
 157 };
 158
 159 void Journaler::recover(Context *onread)
 160 {
 161   lock_guard l(lock);
 162   if (is_stopping()) {
 163     onread->complete(-EAGAIN);
 164     return;
 165   }
 166
 167   ldout(cct, 1) << "recover start" << dendl;
 168   assert(state != STATE_ACTIVE);
 169   assert(readonly);
 170
 171   if (onread)
 172     waitfor_recover.push_back(wrap_finisher(onread));
 173
 174   if (state != STATE_UNDEF) {
 175     ldout(cct, 1) << "recover - already recovering" << dendl;
 176     return;
 177   }
 178
 179   ldout(cct, 1) << "read_head" << dendl;
 180   state = STATE_READHEAD;
 181   C_ReadHead *fin = new C_ReadHead(this);
 182   _read_head(fin, &fin->bl);
 183 }
 184
 185 void Journaler::_read_head(Context *on_finish, bufferlist *bl)
 186 {
 187   // lock is locked
 188   assert(state == STATE_READHEAD || state == STATE_REREADHEAD);
 189
 190   object_t oid = file_object_t(ino, 0);
 191   object_locator_t oloc(pg_pool);
 192   objecter->read_full(oid, oloc, CEPH_NOSNAP, bl, 0, wrap_finisher(on_finish));
 193 }
 194
 195 void Journaler::reread_head(Context *onfinish)
 196 {
 197   lock_guard l(lock);
 198   _reread_head(wrap_finisher(onfinish));
 199 }
 200
 201 /**
 202  * Re-read the head from disk, and set the write_pos, expire_pos, trimmed_pos
 203  * from the on-disk header. This switches the state to STATE_REREADHEAD for
 204  * the duration, and you shouldn't start a re-read while other operations are
 205  * in-flight, nor start other operations while a re-read is in progress.
 206  * Also, don't call this until the Journaler has finished its recovery and has
 207  * gone STATE_ACTIVE!
 208  */
 209 void Journaler::_reread_head(Context *onfinish)
 210 {
 211   ldout(cct, 10) << "reread_head" << dendl;
 212   assert(state == STATE_ACTIVE);
 213
 214   state = STATE_REREADHEAD;
 215   C_RereadHead *fin = new C_RereadHead(this, onfinish);
 216   _read_head(fin, &fin->bl);
 217 }
 218
 219 void Journaler::_finish_reread_head(int r, bufferlist& bl, Context *finish)
 220 {
 221   lock_guard l(lock);
 222   if (is_stopping()) {
 223     finish->complete(-EAGAIN);
 224     return;
 225   }
 226
 227   //read on-disk header into
 228   assert(bl.length() || r < 0 );
 229
 230   // unpack header
 231   if (r == 0) {
 232     Header h;
 233     bufferlist::iterator p = bl.begin();
 234     try {
 235       ::decode(h, p);
 236     } catch (const buffer::error &e) {
 237       finish->complete(-EINVAL);
 238       return;
 239     }
 240     prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos
 241       = h.write_pos;
 242     expire_pos = h.expire_pos;
 243     trimmed_pos = trimming_pos = h.trimmed_pos;
 244     init_headers(h);
 245     state = STATE_ACTIVE;
 246   }
 247
 248   finish->complete(r);
 249 }
 250
 251 void Journaler::_finish_read_head(int r, bufferlist& bl)
 252 {
 253   lock_guard l(lock);
 254   if (is_stopping())
 255     return;
 256
 257   assert(state == STATE_READHEAD);
 258
 259   if (r!=0) {
 260     ldout(cct, 0) << "error getting journal off disk" << dendl;
 261     list<Context*> ls;
 262     ls.swap(waitfor_recover);
 263     finish_contexts(cct, ls, r);
 264     return;
 265   }
 266
 267   if (bl.length() == 0) {
 268     ldout(cct, 1) << "_finish_read_head r=" << r
 269                   << " read 0 bytes, assuming empty log" << dendl;
 270     state = STATE_ACTIVE;
 271     list<Context*> ls;
 272     ls.swap(waitfor_recover);
 273     finish_contexts(cct, ls, 0);
 274     return;
 275   }
 276
 277   // unpack header
 278   bool corrupt = false;
 279   Header h;
 280   bufferlist::iterator p = bl.begin();
 281   try {
 282     ::decode(h, p);
 283
 284     if (h.magic != magic) {
 285       ldout(cct, 0) << "on disk magic '" << h.magic << "' != my magic '"
 286                     << magic << "'" << dendl;
 287       corrupt = true;
 288     } else if (h.write_pos < h.expire_pos || h.expire_pos < h.trimmed_pos) {
 289       ldout(cct, 0) << "Corrupt header (bad offsets): " << h << dendl;
 290       corrupt = true;
 291     }
 292   } catch (const buffer::error &e) {
 293     corrupt = true;
 294   }
 295
 296   if (corrupt) {
 297     list<Context*> ls;
 298     ls.swap(waitfor_recover);
 299     finish_contexts(cct, ls, -EINVAL);
 300     return;
 301   }
 302
 303   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos
 304     = h.write_pos;
 305   read_pos = requested_pos = received_pos = expire_pos = h.expire_pos;
 306   trimmed_pos = trimming_pos = h.trimmed_pos;
 307
 308   init_headers(h);
 309   _set_layout(&h.layout);
 310   stream_format = h.stream_format;
 311   journal_stream.set_format(h.stream_format);
 312
 313   ldout(cct, 1) << "_finish_read_head " << h
 314                 << ".  probing for end of log (from " << write_pos << ")..."
 315                 << dendl;
 316   C_ProbeEnd *fin = new C_ProbeEnd(this);
 317   state = STATE_PROBING;
 318   _probe(fin, &fin->end);
 319 }
 320
 321 void Journaler::_probe(Context *finish, uint64_t *end)
 322 {
 323   // lock is locked
 324   ldout(cct, 1) << "probing for end of the log" << dendl;
 325   assert(state == STATE_PROBING || state == STATE_REPROBING);
 326   // probe the log
 327   filer.probe(ino, &layout, CEPH_NOSNAP,
 328               write_pos, end, true, 0, wrap_finisher(finish));
 329 }
 330
 331 void Journaler::_reprobe(C_OnFinisher *finish)
 332 {
 333   ldout(cct, 10) << "reprobe" << dendl;
 334   assert(state == STATE_ACTIVE);
 335
 336   state = STATE_REPROBING;
 337   C_ReProbe *fin = new C_ReProbe(this, finish);
 338   _probe(fin, &fin->end);
 339 }
 340
 341
 342 void Journaler::_finish_reprobe(int r, uint64_t new_end,
 343                                 C_OnFinisher *onfinish)
 344 {
 345   lock_guard l(lock);
 346   if (is_stopping()) {
 347     onfinish->complete(-EAGAIN);
 348     return;
 349   }
 350
 351   assert(new_end >= write_pos || r < 0);
 352   ldout(cct, 1) << "_finish_reprobe new_end = " << new_end
 353           << " (header had " << write_pos << ")."
 354           << dendl;
 355   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = new_end;
 356   state = STATE_ACTIVE;
 357   onfinish->complete(r);
 358 }
 359
 360 void Journaler::_finish_probe_end(int r, uint64_t end)
 361 {
 362   lock_guard l(lock);
 363   if (is_stopping())
 364     return;
 365
 366   assert(state == STATE_PROBING);
 367   if (r < 0) { // error in probing
 368     goto out;
 369   }
 370   if (((int64_t)end) == -1) {
 371     end = write_pos;
 372     ldout(cct, 1) << "_finish_probe_end write_pos = " << end << " (header had "
 373                   << write_pos << "). log was empty. recovered." << dendl;
 374     ceph_abort(); // hrm.
 375   } else {
 376     assert(end >= write_pos);
 377     ldout(cct, 1) << "_finish_probe_end write_pos = " << end
 378                   << " (header had " << write_pos << "). recovered."
 379                   << dendl;
 380   }
 381
 382   state = STATE_ACTIVE;
 383
 384   prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = end;
 385
 386 out:
 387   // done.
 388   list<Context*> ls;
 389   ls.swap(waitfor_recover);
 390   finish_contexts(cct, ls, r);
 391 }
 392
 393 class Journaler::C_RereadHeadProbe : public Context
 394 {
 395   Journaler *ls;
 396   C_OnFinisher *final_finish;
 397 public:
 398   C_RereadHeadProbe(Journaler *l, C_OnFinisher *finish) :
 399     ls(l), final_finish(finish) {}
 400   void finish(int r) override {
 401     ls->_finish_reread_head_and_probe(r, final_finish);
 402   }
 403 };
 404
 405 void Journaler::reread_head_and_probe(Context *onfinish)
 406 {
 407   lock_guard l(lock);
 408
 409   assert(state == STATE_ACTIVE);
 410   _reread_head(new C_RereadHeadProbe(this, wrap_finisher(onfinish)));
 411 }
 412
 413 void Journaler::_finish_reread_head_and_probe(int r, C_OnFinisher *onfinish)
 414 {
 415   // Expect to be called back from finish_reread_head, which already takes lock
 416   // lock is locked
 417   if (is_stopping()) {
 418     onfinish->complete(-EAGAIN);
 419     return;
 420   }
 421
 422   assert(!r); //if we get an error, we're boned
 423   _reprobe(onfinish);
 424 }
 425
 426
 427 // WRITING
 428
 429 class Journaler::C_WriteHead : public Context {
 430 public:
 431   Journaler *ls;
 432   Header h;
 433   C_OnFinisher *oncommit;
 434   C_WriteHead(Journaler *l, Header& h_, C_OnFinisher *c) : ls(l), h(h_),
 435                                                            oncommit(c) {}
 436   void finish(int r) override {
 437     ls->_finish_write_head(r, h, oncommit);
 438   }
 439 };
 440
 441 void Journaler::write_head(Context *oncommit)
 442 {
 443   lock_guard l(lock);
 444   _write_head(oncommit);
 445 }
 446
 447
 448 void Journaler::_write_head(Context *oncommit)
 449 {
 450   assert(!readonly);
 451   assert(state == STATE_ACTIVE);
 452   last_written.trimmed_pos = trimmed_pos;
 453   last_written.expire_pos = expire_pos;
 454   last_written.unused_field = expire_pos;
 455   last_written.write_pos = safe_pos;
 456   last_written.stream_format = stream_format;
 457   ldout(cct, 10) << "write_head " << last_written << dendl;
 458
 459   // Avoid persisting bad pointers in case of bugs
 460   assert(last_written.write_pos >= last_written.expire_pos);
 461   assert(last_written.expire_pos >= last_written.trimmed_pos);
 462
 463   last_wrote_head = ceph::real_clock::now();
 464
 465   bufferlist bl;
 466   ::encode(last_written, bl);
 467   SnapContext snapc;
 468
 469   object_t oid = file_object_t(ino, 0);
 470   object_locator_t oloc(pg_pool);
 471   objecter->write_full(oid, oloc, snapc, bl, ceph::real_clock::now(), 0,
 472                        wrap_finisher(new C_WriteHead(
 473                                              this, last_written,
 474                                              wrap_finisher(oncommit))),
 475                        0, 0, write_iohint);
 476 }
 477
 478 void Journaler::_finish_write_head(int r, Header &wrote,
 479                                    C_OnFinisher *oncommit)
 480 {
 481   lock_guard l(lock);
 482
 483   if (r < 0) {
 484     lderr(cct) << "_finish_write_head got " << cpp_strerror(r) << dendl;
 485     handle_write_error(r);
 486     return;
 487   }
 488   assert(!readonly);
 489   ldout(cct, 10) << "_finish_write_head " << wrote << dendl;
 490   last_committed = wrote;
 491   if (oncommit) {
 492     oncommit->complete(r);
 493   }
 494
 495   _trim();  // trim?
 496 }
 497
 498
 499 /***************** WRITING *******************/
 500
 501 class Journaler::C_Flush : public Context {
 502   Journaler *ls;
 503   uint64_t start;
 504   ceph::real_time stamp;
 505 public:
 506   C_Flush(Journaler *l, int64_t s, ceph::real_time st)
 507     : ls(l), start(s), stamp(st) {}
 508   void finish(int r) override {
 509     ls->_finish_flush(r, start, stamp);
 510   }
 511 };
 512
 513 void Journaler::_finish_flush(int r, uint64_t start, ceph::real_time stamp)
 514 {
 515   lock_guard l(lock);
 516   assert(!readonly);
 517
 518   if (r < 0) {
 519     lderr(cct) << "_finish_flush got " << cpp_strerror(r) << dendl;
 520     handle_write_error(r);
 521     return;
 522   }
 523
 524   assert(start < flush_pos);
 525
 526   // calc latency?
 527   if (logger) {
 528     ceph::timespan lat = ceph::real_clock::now() - stamp;
 529     logger->tinc(logger_key_lat, lat);
 530   }
 531
 532   // adjust safe_pos
 533   auto it = pending_safe.find(start);
 534   assert(it != pending_safe.end());
 535   pending_safe.erase(it);
 536   if (pending_safe.empty())
 537     safe_pos = next_safe_pos;
 538   else
 539     safe_pos = pending_safe.begin()->second;
 540
 541   ldout(cct, 10) << "_finish_flush safe from " << start
 542                  << ", pending_safe " << pending_safe
 543                  << ", (prezeroing/prezero)/write/flush/safe positions now "
 544                  << "(" << prezeroing_pos << "/" << prezero_pos << ")/"
 545                  << write_pos << "/" << flush_pos << "/" << safe_pos
 546                  << dendl;
 547
 548   // kick waiters <= safe_pos
 549   if (!waitfor_safe.empty()) {
 550     list<Context*> ls;
 551     while (!waitfor_safe.empty()) {
 552       auto it = waitfor_safe.begin();
 553       if (it->first > safe_pos)
 554         break;
 555       ls.splice(ls.end(), it->second);
 556       waitfor_safe.erase(it);
 557     }
 558     finish_contexts(cct, ls);
 559   }
 560 }
 561
 562
 563
 564 uint64_t Journaler::append_entry(bufferlist& bl)
 565 {
 566   unique_lock l(lock);
 567
 568   assert(!readonly);
 569   uint32_t s = bl.length();
 570
 571   // append
 572   size_t delta = bl.length() + journal_stream.get_envelope_size();
 573   // write_buf space is nearly full
 574   if (!write_buf_throttle.get_or_fail(delta)) {
 575     l.unlock();
 576     ldout(cct, 10) << "write_buf_throttle wait, delta " << delta << dendl;
 577     write_buf_throttle.get(delta);
 578     l.lock();
 579   }
 580   ldout(cct, 20) << "write_buf_throttle get, delta " << delta << dendl;
 581   size_t wrote = journal_stream.write(bl, &write_buf, write_pos);
 582   ldout(cct, 10) << "append_entry len " << s << " to " << write_pos << "~"
 583                  << wrote << dendl;
 584   write_pos += wrote;
 585
 586   // flush previous object?
 587   uint64_t su = get_layout_period();
 588   assert(su > 0);
 589   uint64_t write_off = write_pos % su;
 590   uint64_t write_obj = write_pos / su;
 591   uint64_t flush_obj = flush_pos / su;
 592   if (write_obj != flush_obj) {
 593     ldout(cct, 10) << " flushing completed object(s) (su " << su << " wro "
 594                    << write_obj << " flo " << flush_obj << ")" << dendl;
 595     _do_flush(write_buf.length() - write_off);
 596
 597     // if _do_flush() skips flushing some data, it does do a best effort to
 598     // update next_safe_pos.
 599     if (write_buf.length() > 0 &&
 600         write_buf.length() <= wrote) { // the unflushed data are within this entry
 601       // set next_safe_pos to end of previous entry
 602       next_safe_pos = write_pos - wrote;
 603     }
 604   }
 605
 606   return write_pos;
 607 }
 608
 609
 610 void Journaler::_do_flush(unsigned amount)
 611 {
 612   if (is_stopping())
 613     return;
 614   if (write_pos == flush_pos)
 615     return;
 616   assert(write_pos > flush_pos);
 617   assert(!readonly);
 618
 619   // flush
 620   uint64_t len = write_pos - flush_pos;
 621   assert(len == write_buf.length());
 622   if (amount && amount < len)
 623     len = amount;
 624
 625   // zero at least two full periods ahead.  this ensures
 626   // that the next object will not exist.
 627   uint64_t period = get_layout_period();
 628   if (flush_pos + len + 2*period > prezero_pos) {
 629     _issue_prezero();
 630
 631     int64_t newlen = prezero_pos - flush_pos - period;
 632     if (newlen <= 0) {
 633       ldout(cct, 10) << "_do_flush wanted to do " << flush_pos << "~" << len
 634                      << " already too close to prezero_pos " << prezero_pos
 635                      << ", zeroing first" << dendl;
 636       waiting_for_zero_pos = flush_pos + len;
 637       return;
 638     }
 639     if (static_cast<uint64_t>(newlen) < len) {
 640       ldout(cct, 10) << "_do_flush wanted to do " << flush_pos << "~" << len
 641                      << " but hit prezero_pos " << prezero_pos
 642                      << ", will do " << flush_pos << "~" << newlen << dendl;
 643       waiting_for_zero_pos = flush_pos + len;
 644       len = newlen;
 645     }
 646   }
 647   ldout(cct, 10) << "_do_flush flushing " << flush_pos << "~" << len << dendl;
 648
 649   // submit write for anything pending
 650   // flush _start_ pos to _finish_flush
 651   ceph::real_time now = ceph::real_clock::now();
 652   SnapContext snapc;
 653
 654   Context *onsafe = new C_Flush(this, flush_pos, now);  // on COMMIT
 655   pending_safe[flush_pos] = next_safe_pos;
 656
 657   bufferlist write_bl;
 658
 659   // adjust pointers
 660   if (len == write_buf.length()) {
 661     write_bl.swap(write_buf);
 662     next_safe_pos = write_pos;
 663   } else {
 664     write_buf.splice(0, len, &write_bl);
 665     // Keys of waitfor_safe map are journal entry boundaries.
 666     // Try finding a journal entry that we are actually flushing
 667     // and set next_safe_pos to end of it. This is best effort.
 668     // The one we found may not be the lastest flushing entry.
 669     auto p = waitfor_safe.lower_bound(flush_pos + len);
 670     if (p != waitfor_safe.end()) {
 671       if (p->first > flush_pos + len && p != waitfor_safe.begin())
 672        --p;
 673       if (p->first <= flush_pos + len && p->first > next_safe_pos)
 674        next_safe_pos = p->first;
 675     }
 676   }
 677
 678   filer.write(ino, &layout, snapc,
 679               flush_pos, len, write_bl, ceph::real_clock::now(),
 680               0,
 681               wrap_finisher(onsafe), write_iohint);
 682
 683   flush_pos += len;
 684   assert(write_buf.length() == write_pos - flush_pos);
 685   write_buf_throttle.put(len);
 686   ldout(cct, 20) << "write_buf_throttle put, len " << len << dendl;
 687
 688   ldout(cct, 10)
 689     << "_do_flush (prezeroing/prezero)/write/flush/safe pointers now at "
 690     << "(" << prezeroing_pos << "/" << prezero_pos << ")/" << write_pos
 691     << "/" << flush_pos << "/" << safe_pos << dendl;
 692
 693   _issue_prezero();
 694 }
 695
 696
 697 void Journaler::wait_for_flush(Context *onsafe)
 698 {
 699   lock_guard l(lock);
 700   if (is_stopping()) {
 701     onsafe->complete(-EAGAIN);
 702     return;
 703   }
 704   _wait_for_flush(onsafe);
 705 }
 706
 707 void Journaler::_wait_for_flush(Context *onsafe)
 708 {
 709   assert(!readonly);
 710
 711   // all flushed and safe?
 712   if (write_pos == safe_pos) {
 713     assert(write_buf.length() == 0);
 714     ldout(cct, 10)
 715       << "flush nothing to flush, (prezeroing/prezero)/write/flush/safe "
 716       "pointers at " << "(" << prezeroing_pos << "/" << prezero_pos << ")/"
 717       << write_pos << "/" << flush_pos << "/" << safe_pos << dendl;
 718     if (onsafe) {
 719       finisher->queue(onsafe, 0);
 720     }
 721     return;
 722   }
 723
 724   // queue waiter
 725   if (onsafe) {
 726     waitfor_safe[write_pos].push_back(wrap_finisher(onsafe));
 727   }
 728 }
 729
 730 void Journaler::flush(Context *onsafe)
 731 {
 732   lock_guard l(lock);
 733   if (is_stopping()) {
 734     onsafe->complete(-EAGAIN);
 735     return;
 736   }
 737   _flush(wrap_finisher(onsafe));
 738 }
 739
 740 void Journaler::_flush(C_OnFinisher *onsafe)
 741 {
 742   assert(!readonly);
 743
 744   if (write_pos == flush_pos) {
 745     assert(write_buf.length() == 0);
 746     ldout(cct, 10) << "flush nothing to flush, (prezeroing/prezero)/write/"
 747       "flush/safe pointers at " << "(" << prezeroing_pos << "/" << prezero_pos
 748                    << ")/" << write_pos << "/" << flush_pos << "/" << safe_pos
 749                    << dendl;
 750     if (onsafe) {
 751       onsafe->complete(0);
 752     }
 753   } else {
 754     _do_flush();
 755     _wait_for_flush(onsafe);
 756   }
 757
 758   // write head?
 759   if (_write_head_needed()) {
 760     _write_head();
 761   }
 762 }
 763
 764 bool Journaler::_write_head_needed()
 765 {
 766   return last_wrote_head + seconds(cct->_conf->journaler_write_head_interval)
 767       < ceph::real_clock::now();
 768 }
 769
 770
 771 /*************** prezeroing ******************/
 772
 773 struct C_Journaler_Prezero : public Context {
 774   Journaler *journaler;
 775   uint64_t from, len;
 776   C_Journaler_Prezero(Journaler *j, uint64_t f, uint64_t l)
 777     : journaler(j), from(f), len(l) {}
 778   void finish(int r) override {
 779     journaler->_finish_prezero(r, from, len);
 780   }
 781 };
 782
 783 void Journaler::_issue_prezero()
 784 {
 785   assert(prezeroing_pos >= flush_pos);
 786
 787   // we need to zero at least two periods, minimum, to ensure that we
 788   // have a full empty object/period in front of us.
 789   uint64_t num_periods = MAX(2, cct->_conf->journaler_prezero_periods);
 790
 791   /*
 792    * issue zero requests based on write_pos, even though the invariant
 793    * is that we zero ahead of flush_pos.
 794    */
 795   uint64_t period = get_layout_period();
 796   uint64_t to = write_pos + period * num_periods  + period - 1;
 797   to -= to % period;
 798
 799   if (prezeroing_pos >= to) {
 800     ldout(cct, 20) << "_issue_prezero target " << to << " <= prezeroing_pos "
 801                    << prezeroing_pos << dendl;
 802     return;
 803   }
 804
 805   while (prezeroing_pos < to) {
 806     uint64_t len;
 807     if (prezeroing_pos % period == 0) {
 808       len = period;
 809       ldout(cct, 10) << "_issue_prezero removing " << prezeroing_pos << "~"
 810                      << period << " (full period)" << dendl;
 811     } else {
 812       len = period - (prezeroing_pos % period);
 813       ldout(cct, 10) << "_issue_prezero zeroing " << prezeroing_pos << "~"
 814                      << len << " (partial period)" << dendl;
 815     }
 816     SnapContext snapc;
 817     Context *c = wrap_finisher(new C_Journaler_Prezero(this, prezeroing_pos,
 818                                                        len));
 819     filer.zero(ino, &layout, snapc, prezeroing_pos, len,
 820                ceph::real_clock::now(), 0, c);
 821     prezeroing_pos += len;
 822   }
 823 }
 824
 825 // Lock cycle because we get called out of objecter callback (holding
 826 // objecter read lock), but there are also cases where we take the journaler
 827 // lock before calling into objecter to do I/O.
 828 void Journaler::_finish_prezero(int r, uint64_t start, uint64_t len)
 829 {
 830   lock_guard l(lock);
 831
 832   ldout(cct, 10) << "_prezeroed to " << start << "~" << len
 833                  << ", prezeroing/prezero was " << prezeroing_pos << "/"
 834                  << prezero_pos << ", pending " << pending_zero
 835                  << dendl;
 836   if (r < 0 && r != -ENOENT) {
 837     lderr(cct) << "_prezeroed got " << cpp_strerror(r) << dendl;
 838     handle_write_error(r);
 839     return;
 840   }
 841
 842   assert(r == 0 || r == -ENOENT);
 843
 844   if (start == prezero_pos) {
 845     prezero_pos += len;
 846     while (!pending_zero.empty() &&
 847            pending_zero.begin().get_start() == prezero_pos) {
 848       interval_set<uint64_t>::iterator b(pending_zero.begin());
 849       prezero_pos += b.get_len();
 850       pending_zero.erase(b);
 851     }
 852
 853     if (waiting_for_zero_pos > flush_pos) {
 854       _do_flush(waiting_for_zero_pos - flush_pos);
 855     }
 856   } else {
 857     pending_zero.insert(start, len);
 858   }
 859   ldout(cct, 10) << "_prezeroed prezeroing/prezero now " << prezeroing_pos
 860                  << "/" << prezero_pos
 861                  << ", pending " << pending_zero
 862                  << dendl;
 863 }
 864
 865
 866
 867 /***************** READING *******************/
 868
 869
 870 class Journaler::C_Read : public Context {
 871   Journaler *ls;
 872   uint64_t offset;
 873   uint64_t length;
 874 public:
 875   bufferlist bl;
 876   C_Read(Journaler *j, uint64_t o, uint64_t l) : ls(j), offset(o), length(l) {}
 877   void finish(int r) override {
 878     ls->_finish_read(r, offset, length, bl);
 879   }
 880 };
 881
 882 class Journaler::C_RetryRead : public Context {
 883   Journaler *ls;
 884 public:
 885   explicit C_RetryRead(Journaler *l) : ls(l) {}
 886
 887   void finish(int r) override {
 888     // Should only be called from waitfor_safe i.e. already inside lock
 889     // (ls->lock is locked
 890     ls->_prefetch();
 891   }
 892 };
 893
 894 void Journaler::_finish_read(int r, uint64_t offset, uint64_t length,
 895                              bufferlist& bl)
 896 {
 897   lock_guard l(lock);
 898
 899   if (r < 0) {
 900     ldout(cct, 0) << "_finish_read got error " << r << dendl;
 901     error = r;
 902   } else {
 903     ldout(cct, 10) << "_finish_read got " << offset << "~" << bl.length()
 904                    << dendl;
 905     if (bl.length() < length) {
 906       ldout(cct, 0) << "_finish_read got less than expected (" << length << ")"
 907                     << dendl;
 908       error = -EINVAL;
 909     }
 910   }
 911
 912   if (error) {
 913     if (on_readable) {
 914       C_OnFinisher *f = on_readable;
 915       on_readable = 0;
 916       f->complete(error);
 917     }
 918     return;
 919   }
 920
 921   prefetch_buf[offset].swap(bl);
 922
 923   try {
 924     _assimilate_prefetch();
 925   } catch (const buffer::error &err) {
 926     lderr(cct) << "_decode error from assimilate_prefetch" << dendl;
 927     error = -EINVAL;
 928     if (on_readable) {
 929       C_OnFinisher *f = on_readable;
 930       on_readable = 0;
 931       f->complete(error);
 932     }
 933     return;
 934   }
 935   _prefetch();
 936 }
 937
 938 void Journaler::_assimilate_prefetch()
 939 {
 940   bool was_readable = readable;
 941
 942   bool got_any = false;
 943   while (!prefetch_buf.empty()) {
 944     map<uint64_t,bufferlist>::iterator p = prefetch_buf.begin();
 945     if (p->first != received_pos) {
 946       uint64_t gap = p->first - received_pos;
 947       ldout(cct, 10) << "_assimilate_prefetch gap of " << gap
 948                      << " from received_pos " << received_pos
 949                      << " to first prefetched buffer " << p->first << dendl;
 950       break;
 951     }
 952
 953     ldout(cct, 10) << "_assimilate_prefetch " << p->first << "~"
 954                    << p->second.length() << dendl;
 955     received_pos += p->second.length();
 956     read_buf.claim_append(p->second);
 957     assert(received_pos <= requested_pos);
 958     prefetch_buf.erase(p);
 959     got_any = true;
 960   }
 961
 962   if (got_any) {
 963     ldout(cct, 10) << "_assimilate_prefetch read_buf now " << read_pos << "~"
 964                    << read_buf.length() << ", read pointers " << read_pos
 965                    << "/" << received_pos << "/" << requested_pos
 966                    << dendl;
 967
 968     // Update readability (this will also hit any decode errors resulting
 969     // from bad data)
 970     readable = _is_readable();
 971   }
 972
 973   if ((got_any && !was_readable && readable) || read_pos == write_pos) {
 974     // readable!
 975     ldout(cct, 10) << "_finish_read now readable (or at journal end) readable="
 976                    << readable << " read_pos=" << read_pos << " write_pos="
 977                    << write_pos << dendl;
 978     if (on_readable) {
 979       C_OnFinisher *f = on_readable;
 980       on_readable = 0;
 981       f->complete(0);
 982     }
 983   }
 984 }
 985
 986 void Journaler::_issue_read(uint64_t len)
 987 {
 988   // stuck at safe_pos?  (this is needed if we are reading the tail of
 989   // a journal we are also writing to)
 990   assert(requested_pos <= safe_pos);
 991   if (requested_pos == safe_pos) {
 992     ldout(cct, 10) << "_issue_read requested_pos = safe_pos = " << safe_pos
 993                    << ", waiting" << dendl;
 994     assert(write_pos > requested_pos);
 995     if (pending_safe.empty()) {
 996       _flush(NULL);
 997     }
 998
 999     // Make sure keys of waitfor_safe map are journal entry boundaries.
1000     // The key we used here is either next_safe_pos or old value of
1001     // next_safe_pos. next_safe_pos is always set to journal entry
1002     // boundary.
1003     auto p = pending_safe.rbegin();
1004     if (p != pending_safe.rend())
1005       waitfor_safe[p->second].push_back(new C_RetryRead(this));
1006     else
1007       waitfor_safe[next_safe_pos].push_back(new C_RetryRead(this));
1008     return;
1009   }
1010
1011   // don't read too much
1012   if (requested_pos + len > safe_pos) {
1013     len = safe_pos - requested_pos;
1014     ldout(cct, 10) << "_issue_read reading only up to safe_pos " << safe_pos
1015                    << dendl;
1016   }
1017
1018   // go.
1019   ldout(cct, 10) << "_issue_read reading " << requested_pos << "~" << len
1020                  << ", read pointers " << read_pos << "/" << received_pos
1021                  << "/" << (requested_pos+len) << dendl;
1022
1023   // step by period (object).  _don't_ do a single big filer.read()
1024   // here because it will wait for all object reads to complete before
1025   // giving us back any data.  this way we can process whatever bits
1026   // come in that are contiguous.
1027   uint64_t period = get_layout_period();
1028   while (len > 0) {
1029     uint64_t e = requested_pos + period;
1030     e -= e % period;
1031     uint64_t l = e - requested_pos;
1032     if (l > len)
1033       l = len;
1034     C_Read *c = new C_Read(this, requested_pos, l);
1035     filer.read(ino, &layout, CEPH_NOSNAP, requested_pos, l, &c->bl, 0,
1036                wrap_finisher(c), CEPH_OSD_OP_FLAG_FADVISE_DONTNEED);
1037     requested_pos += l;
1038     len -= l;
1039   }
1040 }
1041
1042 void Journaler::_prefetch()
1043 {
1044   if (is_stopping())
1045     return;
1046
1047   ldout(cct, 10) << "_prefetch" << dendl;
1048   // prefetch
1049   uint64_t pf;
1050   if (temp_fetch_len) {
1051     ldout(cct, 10) << "_prefetch temp_fetch_len " << temp_fetch_len << dendl;
1052     pf = temp_fetch_len;
1053     temp_fetch_len = 0;
1054   } else {
1055     pf = fetch_len;
1056   }
1057
1058   uint64_t raw_target = read_pos + pf;
1059
1060   // read full log segments, so increase if necessary
1061   uint64_t period = get_layout_period();
1062   uint64_t remainder = raw_target % period;
1063   uint64_t adjustment = remainder ? period - remainder : 0;
1064   uint64_t target = raw_target + adjustment;
1065
1066   // don't read past the log tail
1067   if (target > write_pos)
1068     target = write_pos;
1069
1070   if (requested_pos < target) {
1071     uint64_t len = target - requested_pos;
1072     ldout(cct, 10) << "_prefetch " << pf << " requested_pos " << requested_pos
1073                    << " < target " << target << " (" << raw_target
1074                    << "), prefetching " << len << dendl;
1075
1076     if (pending_safe.empty() && write_pos > safe_pos) {
1077       // If we are reading and writing the journal, then we may need
1078       // to issue a flush if one isn't already in progress.
1079       // Avoid doing a flush every time so that if we do write/read/write/read
1080       // we don't end up flushing after every write.
1081       ldout(cct, 10) << "_prefetch: requested_pos=" << requested_pos
1082                      << ", read_pos=" << read_pos
1083                      << ", write_pos=" << write_pos
1084                      << ", safe_pos=" << safe_pos << dendl;
1085       _do_flush();
1086     }
1087
1088     _issue_read(len);
1089   }
1090 }
1091
1092
1093 /*
1094  * _is_readable() - return true if next entry is ready.
1095  */
1096 bool Journaler::_is_readable()
1097 {
1098   // anything to read?
1099   if (read_pos == write_pos)
1100     return false;
1101
1102   // Check if the retrieve bytestream has enough for an entry
1103   uint64_t need;
1104   if (journal_stream.readable(read_buf, &need)) {
1105     return true;
1106   }
1107
1108   ldout (cct, 10) << "_is_readable read_buf.length() == " << read_buf.length()
1109                   << ", but need " << need << " for next entry; fetch_len is "
1110                   << fetch_len << dendl;
1111
1112   // partial fragment at the end?
1113   if (received_pos == write_pos) {
1114     ldout(cct, 10) << "is_readable() detected partial entry at tail, "
1115       "adjusting write_pos to " << read_pos << dendl;
1116
1117     // adjust write_pos
1118     prezeroing_pos = prezero_pos = write_pos = flush_pos = safe_pos = next_safe_pos = read_pos;
1119     assert(write_buf.length() == 0);
1120     assert(waitfor_safe.empty());
1121
1122     // reset read state
1123     requested_pos = received_pos = read_pos;
1124     read_buf.clear();
1125
1126     // FIXME: truncate on disk?
1127
1128     return false;
1129   }
1130
1131   if (need > fetch_len) {
1132     temp_fetch_len = need;
1133     ldout(cct, 10) << "_is_readable noting temp_fetch_len " << temp_fetch_len
1134                    << dendl;
1135   }
1136
1137   ldout(cct, 10) << "_is_readable: not readable, returning false" << dendl;
1138   return false;
1139 }
1140
1141 /*
1142  * is_readable() - kickstart prefetch, too
1143  */
1144 bool Journaler::is_readable()
1145 {
1146   lock_guard l(lock);
1147
1148   if (error != 0) {
1149     return false;
1150   }
1151
1152   bool r = readable;
1153   _prefetch();
1154   return r;
1155 }
1156
1157 class Journaler::C_EraseFinish : public Context {
1158   Journaler *journaler;
1159   C_OnFinisher *completion;
1160   public:
1161   C_EraseFinish(Journaler *j, C_OnFinisher *c) : journaler(j), completion(c) {}
1162   void finish(int r) override {
1163     journaler->_finish_erase(r, completion);
1164   }
1165 };
1166
1167 /**
1168  * Entirely erase the journal, including header.  For use when you
1169  * have already made a copy of the journal somewhere else.
1170  */
1171 void Journaler::erase(Context *completion)
1172 {
1173   lock_guard l(lock);
1174
1175   // Async delete the journal data
1176   uint64_t first = trimmed_pos / get_layout_period();
1177   uint64_t num = (write_pos - trimmed_pos) / get_layout_period() + 2;
1178   filer.purge_range(ino, &layout, SnapContext(), first, num,
1179                     ceph::real_clock::now(), 0,
1180                     wrap_finisher(new C_EraseFinish(
1181                                     this, wrap_finisher(completion))));
1182
1183   // We will not start the operation to delete the header until
1184   // _finish_erase has seen the data deletion succeed: otherwise if
1185   // there was an error deleting data we might prematurely delete the
1186   // header thereby lose our reference to the data.
1187 }
1188
1189 void Journaler::_finish_erase(int data_result, C_OnFinisher *completion)
1190 {
1191   lock_guard l(lock);
1192   if (is_stopping()) {
1193     completion->complete(-EAGAIN);
1194     return;
1195   }
1196
1197   if (data_result == 0) {
1198     // Async delete the journal header
1199     filer.purge_range(ino, &layout, SnapContext(), 0, 1,
1200                       ceph::real_clock::now(),
1201                       0, wrap_finisher(completion));
1202   } else {
1203     lderr(cct) << "Failed to delete journal " << ino << " data: "
1204                << cpp_strerror(data_result) << dendl;
1205     completion->complete(data_result);
1206   }
1207 }
1208
1209 /* try_read_entry(bl)
1210  *  read entry into bl if it's ready.
1211  *  otherwise, do nothing.
1212  */
1213 bool Journaler::try_read_entry(bufferlist& bl)
1214 {
1215   lock_guard l(lock);
1216
1217   if (!readable) {
1218     ldout(cct, 10) << "try_read_entry at " << read_pos << " not readable"
1219                    << dendl;
1220     return false;
1221   }
1222
1223   uint64_t start_ptr;
1224   size_t consumed;
1225   try {
1226     consumed = journal_stream.read(read_buf, &bl, &start_ptr);
1227     if (stream_format >= JOURNAL_FORMAT_RESILIENT) {
1228       assert(start_ptr == read_pos);
1229     }
1230   } catch (const buffer::error &e) {
1231     lderr(cct) << __func__ << ": decode error from journal_stream" << dendl;
1232     error = -EINVAL;
1233     return false;
1234   }
1235
1236   ldout(cct, 10) << "try_read_entry at " << read_pos << " read "
1237                  << read_pos << "~" << consumed << " (have "
1238                  << read_buf.length() << ")" << dendl;
1239
1240   read_pos += consumed;
1241   try {
1242     // We were readable, we might not be any more
1243     readable = _is_readable();
1244   } catch (const buffer::error &e) {
1245     lderr(cct) << __func__ << ": decode error from _is_readable" << dendl;
1246     error = -EINVAL;
1247     return false;
1248   }
1249
1250   // prefetch?
1251   _prefetch();
1252   return true;
1253 }
1254
1255 void Journaler::wait_for_readable(Context *onreadable)
1256 {
1257   lock_guard l(lock);
1258   if (is_stopping()) {
1259     finisher->queue(onreadable, -EAGAIN);
1260     return;
1261   }
1262
1263   assert(on_readable == 0);
1264   if (!readable) {
1265     ldout(cct, 10) << "wait_for_readable at " << read_pos << " onreadable "
1266                    << onreadable << dendl;
1267     on_readable = wrap_finisher(onreadable);
1268   } else {
1269     // race with OSD reply
1270     finisher->queue(onreadable, 0);
1271   }
1272 }
1273
1274 bool Journaler::have_waiter() const
1275 {
1276   return on_readable != nullptr;
1277 }
1278
1279
1280
1281
1282 /***************** TRIMMING *******************/
1283
1284
1285 class Journaler::C_Trim : public Context {
1286   Journaler *ls;
1287   uint64_t to;
1288 public:
1289   C_Trim(Journaler *l, int64_t t) : ls(l), to(t) {}
1290   void finish(int r) override {
1291     ls->_finish_trim(r, to);
1292   }
1293 };
1294
1295 void Journaler::trim()
1296 {
1297   lock_guard l(lock);
1298   _trim();
1299 }
1300
1301 void Journaler::_trim()
1302 {
1303   if (is_stopping())
1304     return;
1305
1306   assert(!readonly);
1307   uint64_t period = get_layout_period();
1308   uint64_t trim_to = last_committed.expire_pos;
1309   trim_to -= trim_to % period;
1310   ldout(cct, 10) << "trim last_commited head was " << last_committed
1311            << ", can trim to " << trim_to
1312            << dendl;
1313   if (trim_to == 0 || trim_to == trimming_pos) {
1314     ldout(cct, 10) << "trim already trimmed/trimming to "
1315                    << trimmed_pos << "/" << trimming_pos << dendl;
1316     return;
1317   }
1318
1319   if (trimming_pos > trimmed_pos) {
1320     ldout(cct, 10) << "trim already trimming atm, try again later.  "
1321       "trimmed/trimming is " << trimmed_pos << "/" << trimming_pos << dendl;
1322     return;
1323   }
1324
1325   // trim
1326   assert(trim_to <= write_pos);
1327   assert(trim_to <= expire_pos);
1328   assert(trim_to > trimming_pos);
1329   ldout(cct, 10) << "trim trimming to " << trim_to
1330                  << ", trimmed/trimming/expire are "
1331                  << trimmed_pos << "/" << trimming_pos << "/" << expire_pos
1332                  << dendl;
1333
1334   // delete range of objects
1335   uint64_t first = trimming_pos / period;
1336   uint64_t num = (trim_to - trimming_pos) / period;
1337   SnapContext snapc;
1338   filer.purge_range(ino, &layout, snapc, first, num,
1339                     ceph::real_clock::now(), 0,
1340                     wrap_finisher(new C_Trim(this, trim_to)));
1341   trimming_pos = trim_to;
1342 }
1343
1344 void Journaler::_finish_trim(int r, uint64_t to)
1345 {
1346   lock_guard l(lock);
1347
1348   assert(!readonly);
1349   ldout(cct, 10) << "_finish_trim trimmed_pos was " << trimmed_pos
1350            << ", trimmed/trimming/expire now "
1351            << to << "/" << trimming_pos << "/" << expire_pos
1352            << dendl;
1353   if (r < 0 && r != -ENOENT) {
1354     lderr(cct) << "_finish_trim got " << cpp_strerror(r) << dendl;
1355     handle_write_error(r);
1356     return;
1357   }
1358
1359   assert(r >= 0 || r == -ENOENT);
1360
1361   assert(to <= trimming_pos);
1362   assert(to > trimmed_pos);
1363   trimmed_pos = to;
1364 }
1365
1366 void Journaler::handle_write_error(int r)
1367 {
1368   // lock is locked
1369
1370   lderr(cct) << "handle_write_error " << cpp_strerror(r) << dendl;
1371   if (on_write_error) {
1372     on_write_error->complete(r);
1373     on_write_error = NULL;
1374     called_write_error = true;
1375   } else if (called_write_error) {
1376     /* We don't call error handler more than once, subsequent errors
1377      * are dropped -- this is okay as long as the error handler does
1378      * something dramatic like respawn */
1379     lderr(cct) << __func__ << ": multiple write errors, handler already called"
1380                << dendl;
1381   } else {
1382     assert(0 == "unhandled write error");
1383   }
1384 }
1385
1386
1387 /**
1388  * Test whether the 'read_buf' byte stream has enough data to read
1389  * an entry
1390  *
1391  * sets 'next_envelope_size' to the number of bytes needed to advance (enough
1392  * to get the next header if header was unavailable, or enough to get the whole
1393  * next entry if the header was available but the body wasn't).
1394  */
1395 bool JournalStream::readable(bufferlist &read_buf, uint64_t *need) const
1396 {
1397   assert(need != NULL);
1398
1399   uint32_t entry_size = 0;
1400   uint64_t entry_sentinel = 0;
1401   bufferlist::iterator p = read_buf.begin();
1402
1403   // Do we have enough data to decode an entry prefix?
1404   if (format >= JOURNAL_FORMAT_RESILIENT) {
1405     *need = sizeof(entry_size) + sizeof(entry_sentinel);
1406   } else {
1407     *need = sizeof(entry_size);
1408   }
1409   if (read_buf.length() >= *need) {
1410     if (format >= JOURNAL_FORMAT_RESILIENT) {
1411       ::decode(entry_sentinel, p);
1412       if (entry_sentinel != sentinel) {
1413         throw buffer::malformed_input("Invalid sentinel");
1414       }
1415     }
1416
1417     ::decode(entry_size, p);
1418   } else {
1419     return false;
1420   }
1421
1422   // Do we have enough data to decode an entry prefix, payload and suffix?
1423   if (format >= JOURNAL_FORMAT_RESILIENT) {
1424     *need = JOURNAL_ENVELOPE_RESILIENT + entry_size;
1425   } else {
1426     *need = JOURNAL_ENVELOPE_LEGACY + entry_size;
1427   }
1428   if (read_buf.length() >= *need) {
1429     return true;  // No more bytes needed
1430   }
1431
1432   return false;
1433 }
1434
1435
1436 /**
1437  * Consume one entry from a journal byte stream 'from', splicing a
1438  * serialized LogEvent blob into 'entry'.
1439  *
1440  * 'entry' must be non null and point to an empty bufferlist.
1441  *
1442  * 'from' must contain sufficient valid data (i.e. readable is true).
1443  *
1444  * 'start_ptr' will be set to the entry's start pointer, if the collection
1445  * format provides it.  It may not be null.
1446  *
1447  * @returns The number of bytes consumed from the `from` byte stream.  Note
1448  *          that this is not equal to the length of `entry`, which contains
1449  *          the inner serialized LogEvent and not the envelope.
1450  */
1451 size_t JournalStream::read(bufferlist &from, bufferlist *entry,
1452                            uint64_t *start_ptr)
1453 {
1454   assert(start_ptr != NULL);
1455   assert(entry != NULL);
1456   assert(entry->length() == 0);
1457
1458   uint32_t entry_size = 0;
1459
1460   // Consume envelope prefix: entry_size and entry_sentinel
1461   bufferlist::iterator from_ptr = from.begin();
1462   if (format >= JOURNAL_FORMAT_RESILIENT) {
1463     uint64_t entry_sentinel = 0;
1464     ::decode(entry_sentinel, from_ptr);
1465     // Assertion instead of clean check because of precondition of this
1466     // fn is that readable() already passed
1467     assert(entry_sentinel == sentinel);
1468   }
1469   ::decode(entry_size, from_ptr);
1470
1471   // Read out the payload
1472   from_ptr.copy(entry_size, *entry);
1473
1474   // Consume the envelope suffix (start_ptr)
1475   if (format >= JOURNAL_FORMAT_RESILIENT) {
1476     ::decode(*start_ptr, from_ptr);
1477   } else {
1478     *start_ptr = 0;
1479   }
1480
1481   // Trim the input buffer to discard the bytes we have consumed
1482   from.splice(0, from_ptr.get_off());
1483
1484   return from_ptr.get_off();
1485 }
1486
1487
1488 /**
1489  * Append one entry
1490  */
1491 size_t JournalStream::write(bufferlist &entry, bufferlist *to,
1492                             uint64_t const &start_ptr)
1493 {
1494   assert(to != NULL);
1495
1496   uint32_t const entry_size = entry.length();
1497   if (format >= JOURNAL_FORMAT_RESILIENT) {
1498     ::encode(sentinel, *to);
1499   }
1500   ::encode(entry_size, *to);
1501   to->claim_append(entry);
1502   if (format >= JOURNAL_FORMAT_RESILIENT) {
1503     ::encode(start_ptr, *to);
1504   }
1505
1506   if (format >= JOURNAL_FORMAT_RESILIENT) {
1507     return JOURNAL_ENVELOPE_RESILIENT + entry_size;
1508   } else {
1509     return JOURNAL_ENVELOPE_LEGACY + entry_size;
1510   }
1511 }
1512
1513 /**
1514  * set write error callback
1515  *
1516  * Set a callback/context to trigger if we get a write error from
1517  * the objecter.  This may be from an explicit request (e.g., flush)
1518  * or something async the journaler did on its own (e.g., journal
1519  * header update).
1520  *
1521  * It is only used once; if the caller continues to use the
1522  * Journaler and wants to hear about errors, it needs to reset the
1523  * error_handler.
1524  *
1525  * @param c callback/context to trigger on error
1526  */
1527 void Journaler::set_write_error_handler(Context *c) {
1528   lock_guard l(lock);
1529   assert(!on_write_error);
1530   on_write_error = wrap_finisher(c);
1531   called_write_error = false;
1532 }
1533
1534
1535 /**
1536  * Wrap a context in a C_OnFinisher, if it is non-NULL
1537  *
1538  * Utility function to avoid lots of error-prone and verbose
1539  * NULL checking on contexts passed in.
1540  */
1541 C_OnFinisher *Journaler::wrap_finisher(Context *c)
1542 {
1543   if (c != NULL) {
1544     return new C_OnFinisher(c, finisher);
1545   } else {
1546     return NULL;
1547   }
1548 }
1549
1550 void Journaler::shutdown()
1551 {
1552   lock_guard l(lock);
1553
1554   ldout(cct, 1) << __func__ << dendl;
1555
1556   state = STATE_STOPPING;
1557   readable = false;
1558
1559   // Kick out anyone reading from journal
1560   error = -EAGAIN;
1561   if (on_readable) {
1562     C_OnFinisher *f = on_readable;
1563     on_readable = 0;
1564     f->complete(-EAGAIN);
1565   }
1566
1567   list<Context*> ls;
1568   ls.swap(waitfor_recover);
1569   finish_contexts(cct, ls, -ESHUTDOWN);
1570
1571   std::map<uint64_t, std::list<Context*> >::iterator i;
1572   for (i = waitfor_safe.begin(); i != waitfor_safe.end(); ++i) {
1573     finish_contexts(cct, i->second, -EAGAIN);
1574   }
1575   waitfor_safe.clear();
1576 }
1577