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   return count;
 115 }
 116
 117
 118 // ------------------------------------
 119
 120 uint64_t JournalingObjectStore::ApplyManager::op_apply_start(uint64_t op)
 121 {
 122   Mutex::Locker l(apply_lock);
 123   while (blocked) {
 124     dout(10) << "op_apply_start blocked, waiting" << dendl;
 125     blocked_cond.Wait(apply_lock);
 126   }
 127   dout(10) << "op_apply_start " << op << " open_ops " << open_ops << " -> "
 128            << (open_ops+1) << dendl;
 129   assert(!blocked);
 130   assert(op > committed_seq);
 131   open_ops++;
 132   return op;
 133 }
 134
 135 void JournalingObjectStore::ApplyManager::op_apply_finish(uint64_t op)
 136 {
 137   Mutex::Locker l(apply_lock);
 138   dout(10) << "op_apply_finish " << op << " open_ops " << open_ops << " -> "
 139            << (open_ops-1) << ", max_applied_seq " << max_applied_seq << " -> "
 140            << MAX(op, max_applied_seq) << dendl;
 141   --open_ops;
 142   assert(open_ops >= 0);
 143
 144   // signal a blocked commit_start
 145   if (blocked) {
 146     blocked_cond.Signal();
 147   }
 148
 149   // there can be multiple applies in flight; track the max value we
 150   // note.  note that we can't _read_ this value and learn anything
 151   // meaningful unless/until we've quiesced all in-flight applies.
 152   if (op > max_applied_seq)
 153     max_applied_seq = op;
 154 }
 155
 156 uint64_t JournalingObjectStore::SubmitManager::op_submit_start()
 157 {
 158   lock.Lock();
 159   uint64_t op = ++op_seq;
 160   dout(10) << "op_submit_start " << op << dendl;
 161   return op;
 162 }
 163
 164 void JournalingObjectStore::SubmitManager::op_submit_finish(uint64_t op)
 165 {
 166   dout(10) << "op_submit_finish " << op << dendl;
 167   if (op != op_submitted + 1) {
 168     dout(0) << "op_submit_finish " << op << " expected " << (op_submitted + 1)
 169             << ", OUT OF ORDER" << dendl;
 170     assert(0 == "out of order op_submit_finish");
 171   }
 172   op_submitted = op;
 173   lock.Unlock();
 174 }
 175
 176
 177 // ------------------------------------------
 178
 179 void JournalingObjectStore::ApplyManager::add_waiter(uint64_t op, Context *c)
 180 {
 181   Mutex::Locker l(com_lock);
 182   assert(c);
 183   commit_waiters[op].push_back(c);
 184 }
 185
 186 bool JournalingObjectStore::ApplyManager::commit_start()
 187 {
 188   bool ret = false;
 189
 190   {
 191     Mutex::Locker l(apply_lock);
 192     dout(10) << "commit_start max_applied_seq " << max_applied_seq
 193              << ", open_ops " << open_ops << dendl;
 194     blocked = true;
 195     while (open_ops > 0) {
 196       dout(10) << "commit_start waiting for " << open_ops
 197                << " open ops to drain" << dendl;
 198       blocked_cond.Wait(apply_lock);
 199     }
 200     assert(open_ops == 0);
 201     dout(10) << "commit_start blocked, all open_ops have completed" << dendl;
 202     {
 203       Mutex::Locker l(com_lock);
 204       if (max_applied_seq == committed_seq) {
 205         dout(10) << "commit_start nothing to do" << dendl;
 206         blocked = false;
 207         assert(commit_waiters.empty());
 208         goto out;
 209       }
 210
 211       committing_seq = max_applied_seq;
 212
 213       dout(10) << "commit_start committing " << committing_seq
 214                << ", still blocked" << dendl;
 215     }
 216   }
 217   ret = true;
 218
 219   if (journal)
 220     journal->commit_start(committing_seq);  // tell the journal too
 221  out:
 222   return ret;
 223 }
 224
 225 void JournalingObjectStore::ApplyManager::commit_started()
 226 {
 227   Mutex::Locker l(apply_lock);
 228   // allow new ops. (underlying fs should now be committing all prior ops)
 229   dout(10) << "commit_started committing " << committing_seq << ", unblocking"
 230            << dendl;
 231   blocked = false;
 232   blocked_cond.Signal();
 233 }
 234
 235 void JournalingObjectStore::ApplyManager::commit_finish()
 236 {
 237   Mutex::Locker l(com_lock);
 238   dout(10) << "commit_finish thru " << committing_seq << dendl;
 239
 240   if (journal)
 241     journal->committed_thru(committing_seq);
 242
 243   committed_seq = committing_seq;
 244
 245   map<version_t, vector<Context*> >::iterator p = commit_waiters.begin();
 246   while (p != commit_waiters.end() &&
 247     p->first <= committing_seq) {
 248     finisher.queue(p->second);
 249     commit_waiters.erase(p++);
 250   }
 251 }
 252
 253 void JournalingObjectStore::_op_journal_transactions(
 254   bufferlist& tbl, uint32_t orig_len, uint64_t op,
 255   Context *onjournal, TrackedOpRef osd_op)
 256 {
 257   if (osd_op.get())
 258     dout(10) << "op_journal_transactions " << op << " reqid_t "
 259              << (static_cast<OpRequest *>(osd_op.get()))->get_reqid() << dendl;
 260   else
 261     dout(10) << "op_journal_transactions " << op  << dendl;
 262
 263   if (journal && journal->is_writeable()) {
 264     journal->submit_entry(op, tbl, orig_len, onjournal, osd_op);
 265   } else if (onjournal) {
 266     apply_manager.add_waiter(op, onjournal);
 267   }
 268 }