ceph/src/os/filestore/JournalingObjectStore.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2
   3 #include "JournalingObjectStore.h"
   4
   5 #include "common/errno.h"
   6 #include "common/debug.h"
   7
   8 #define dout_context cct
   9 #define dout_subsys ceph_subsys_journal
  10 #undef dout_prefix
  11 #define dout_prefix *_dout << "journal "
  12
  13
  14
  15 void JournalingObjectStore::journal_start()
  16 {
  17   dout(10) << "journal_start" << dendl;
  18   finisher.start();
  19 }
  20
  21 void JournalingObjectStore::journal_stop()
  22 {
  23   dout(10) << "journal_stop" << dendl;
  24   finisher.wait_for_empty();
  25   finisher.stop();
  26 }
  27
  28 // A journal_replay() makes journal writeable, this closes that out.
  29 void JournalingObjectStore::journal_write_close()
  30 {
  31   if (journal) {
  32     journal->close();
  33     delete journal;
  34     journal = 0;
  35   }
  36   apply_manager.reset();
  37 }
  38
  39 int JournalingObjectStore::journal_replay(uint64_t fs_op_seq)
  40 {
  41   dout(10) << "journal_replay fs op_seq " << fs_op_seq << dendl;
  42
  43   if (cct->_conf->journal_replay_from) {
  44     dout(0) << "journal_replay forcing replay from "
  45             << cct->_conf->journal_replay_from
  46             << " instead of " << fs_op_seq << dendl;
  47     // the previous op is the last one committed
  48     fs_op_seq = cct->_conf->journal_replay_from - 1;
  49   }
  50
  51   uint64_t op_seq = fs_op_seq;
  52   apply_manager.init_seq(fs_op_seq);
  53
  54   if (!journal) {
  55     submit_manager.set_op_seq(op_seq);
  56     return 0;
  57   }
  58
  59   int err = journal->open(op_seq);
  60   if (err < 0) {
  61     dout(3) << "journal_replay open failed with "
  62             << cpp_strerror(err) << dendl;
  63     delete journal;
  64     journal = 0;
  65     return err;
  66   }
  67
  68   replaying = true;
  69
  70   int count = 0;
  71   while (1) {
  72     bufferlist bl;
  73     uint64_t seq = op_seq + 1;
  74     if (!journal->read_entry(bl, seq)) {
  75       dout(3) << "journal_replay: end of journal, done." << dendl;
  76       break;
  77     }
  78
  79     if (seq <= op_seq) {
  80       dout(3) << "journal_replay: skipping old op seq " << seq << " <= " << op_seq << dendl;
  81       continue;
  82     }
  83     assert(op_seq == seq-1);
  84
  85     dout(3) << "journal_replay: applying op seq " << seq << dendl;
  86     bufferlist::iterator p = bl.begin();
  87     vector<ObjectStore::Transaction> tls;
  88     while (!p.end()) {
  89       tls.emplace_back(Transaction(p));
  90     }
  91
  92     apply_manager.op_apply_start(seq);
  93     int r = do_transactions(tls, seq);
  94     apply_manager.op_apply_finish(seq);
  95
  96     op_seq = seq;
  97     count++;
  98
  99     dout(3) << "journal_replay: r = " << r << ", op_seq now " << op_seq << dendl;
 100   }
 101
 102   if (count)
 103     dout(3) << "journal_replay: total = " << count << dendl;
 104
 105   replaying = false;
 106
 107   submit_manager.set_op_seq(op_seq);
 108
 109   // done reading, make writeable.
 110   err = journal->make_writeable();
 111   if (err < 0)
 112     return err;
 113
 114   if (!count)
 115     journal->committed_thru(fs_op_seq);
 116
 117   return count;
 118 }
 119
 120
 121 // ------------------------------------
 122
 123 uint64_t JournalingObjectStore::ApplyManager::op_apply_start(uint64_t op)
 124 {
 125   Mutex::Locker l(apply_lock);
 126   while (blocked) {
 127     dout(10) << "op_apply_start blocked, waiting" << dendl;
 128     blocked_cond.Wait(apply_lock);
 129   }
 130   dout(10) << "op_apply_start " << op << " open_ops " << open_ops << " -> "
 131            << (open_ops+1) << dendl;
 132   assert(!blocked);
 133   assert(op > committed_seq);
 134   open_ops++;
 135   return op;
 136 }
 137
 138 void JournalingObjectStore::ApplyManager::op_apply_finish(uint64_t op)
 139 {
 140   Mutex::Locker l(apply_lock);
 141   dout(10) << "op_apply_finish " << op << " open_ops " << open_ops << " -> "
 142            << (open_ops-1) << ", max_applied_seq " << max_applied_seq << " -> "
 143            << MAX(op, max_applied_seq) << dendl;
 144   --open_ops;
 145   assert(open_ops >= 0);
 146
 147   // signal a blocked commit_start
 148   if (blocked) {
 149     blocked_cond.Signal();
 150   }
 151
 152   // there can be multiple applies in flight; track the max value we
 153   // note.  note that we can't _read_ this value and learn anything
 154   // meaningful unless/until we've quiesced all in-flight applies.
 155   if (op > max_applied_seq)
 156     max_applied_seq = op;
 157 }
 158
 159 uint64_t JournalingObjectStore::SubmitManager::op_submit_start()
 160 {
 161   lock.Lock();
 162   uint64_t op = ++op_seq;
 163   dout(10) << "op_submit_start " << op << dendl;
 164   return op;
 165 }
 166
 167 void JournalingObjectStore::SubmitManager::op_submit_finish(uint64_t op)
 168 {
 169   dout(10) << "op_submit_finish " << op << dendl;
 170   if (op != op_submitted + 1) {
 171     dout(0) << "op_submit_finish " << op << " expected " << (op_submitted + 1)
 172             << ", OUT OF ORDER" << dendl;
 173     assert(0 == "out of order op_submit_finish");
 174   }
 175   op_submitted = op;
 176   lock.Unlock();
 177 }
 178
 179
 180 // ------------------------------------------
 181
 182 void JournalingObjectStore::ApplyManager::add_waiter(uint64_t op, Context *c)
 183 {
 184   Mutex::Locker l(com_lock);
 185   assert(c);
 186   commit_waiters[op].push_back(c);
 187 }
 188
 189 bool JournalingObjectStore::ApplyManager::commit_start()
 190 {
 191   bool ret = false;
 192
 193   {
 194     Mutex::Locker l(apply_lock);
 195     dout(10) << "commit_start max_applied_seq " << max_applied_seq
 196              << ", open_ops " << open_ops << dendl;
 197     blocked = true;
 198     while (open_ops > 0) {
 199       dout(10) << "commit_start waiting for " << open_ops
 200                << " open ops to drain" << dendl;
 201       blocked_cond.Wait(apply_lock);
 202     }
 203     assert(open_ops == 0);
 204     dout(10) << "commit_start blocked, all open_ops have completed" << dendl;
 205     {
 206       Mutex::Locker l(com_lock);
 207       if (max_applied_seq == committed_seq) {
 208         dout(10) << "commit_start nothing to do" << dendl;
 209         blocked = false;
 210         assert(commit_waiters.empty());
 211         goto out;
 212       }
 213
 214       committing_seq = max_applied_seq;
 215
 216       dout(10) << "commit_start committing " << committing_seq
 217                << ", still blocked" << dendl;
 218     }
 219   }
 220   ret = true;
 221
 222   if (journal)
 223     journal->commit_start(committing_seq);  // tell the journal too
 224  out:
 225   return ret;
 226 }
 227
 228 void JournalingObjectStore::ApplyManager::commit_started()
 229 {
 230   Mutex::Locker l(apply_lock);
 231   // allow new ops. (underlying fs should now be committing all prior ops)
 232   dout(10) << "commit_started committing " << committing_seq << ", unblocking"
 233            << dendl;
 234   blocked = false;
 235   blocked_cond.Signal();
 236 }
 237
 238 void JournalingObjectStore::ApplyManager::commit_finish()
 239 {
 240   Mutex::Locker l(com_lock);
 241   dout(10) << "commit_finish thru " << committing_seq << dendl;
 242
 243   if (journal)
 244     journal->committed_thru(committing_seq);
 245
 246   committed_seq = committing_seq;
 247
 248   map<version_t, vector<Context*> >::iterator p = commit_waiters.begin();
 249   while (p != commit_waiters.end() &&
 250     p->first <= committing_seq) {
 251     finisher.queue(p->second);
 252     commit_waiters.erase(p++);
 253   }
 254 }
 255
 256 void JournalingObjectStore::_op_journal_transactions(
 257   bufferlist& tbl, uint32_t orig_len, uint64_t op,
 258   Context *onjournal, TrackedOpRef osd_op)
 259 {
 260   if (osd_op.get())
 261     dout(10) << "op_journal_transactions " << op << " reqid_t "
 262              << (static_cast<OpRequest *>(osd_op.get()))->get_reqid() << dendl;
 263   else
 264     dout(10) << "op_journal_transactions " << op  << dendl;
 265
 266   if (journal && journal->is_writeable()) {
 267     journal->submit_entry(op, tbl, orig_len, onjournal, osd_op);
 268   } else if (onjournal) {
 269     apply_manager.add_waiter(op, onjournal);
 270   }
 271 }