RepoMirrors/ceph
1
0
Fork 0
mirror of https://github.com/ceph/ceph synced 2025-01-19 17:41:39 +00:00
Code Issues Packages Projects Releases Wiki Activity

os/bluestore/BlueFS: fix race with log flush during async log compaction

Browse Source 
During async log compaction we rely on _flush-and_sync_log to update the
log_writer to jump_to.  However, if racing threads are also trying to flush
the log and manage to flush our new log events for us, then our flush will
turn into a no-op, and we won't update jump_to correctly at all.  This
results in a corrupted log size a bit later one.

Fix by ensuring that there are no in-progress flushes before we add our
log entries.  Also, add asserts to _flush_and_sync_log to make sure we
never bail out early if jump_to is set (which would indicate this or
another similar bug is still present).

Fixes: http://tracker.ceph.com/issues/21878
Signed-off-by: Sage Weil <sage@redhat.com>
This commit is contained in:
Sage Weil 2017-10-20 08:51:17 -05:00
parent 34951266fe
commit 4324c8bc7e
1 changed files with 11 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
src/os/bluestore/BlueFS.cc
View File

@ -1195,6 +1195,14 @@ void BlueFS::_compact_log_async(std::unique_lock<std::mutex>& l)
new_log = new File;
new_log->fnode.ino = 0; // so that _flush_range won't try to log the fnode
// 0. wait for any racing flushes to complete. (We do not want to block
// in _flush_sync_log with jump_to set or else a racing thread might flush
// our entries and our jump_to update won't be correct.)
while (log_flushing) {
dout(10) << __func__ << " log is currently flushing, waiting" << dendl;
log_cond.wait(l);
}
// 1. allocate new log space and jump to it.
old_log_jump_to = log_file->fnode.get_allocated();
uint64_t need = old_log_jump_to + cct->_conf->bluefs_max_log_runway;
@ -1345,16 +1353,19 @@ int BlueFS::_flush_and_sync_log(std::unique_lock<std::mutex>& l,
while (log_flushing) {
dout(10) << __func__ << " want_seq " << want_seq
<< " log is currently flushing, waiting" << dendl;
assert(!jump_to);
log_cond.wait(l);
}
if (want_seq && want_seq <= log_seq_stable) {
dout(10) << __func__ << " want_seq " << want_seq << " <= log_seq_stable "
<< log_seq_stable << ", done" << dendl;
assert(!jump_to);
return 0;
}
if (log_t.empty() && dirty_files.empty()) {
dout(10) << __func__ << " want_seq " << want_seq
<< " " << log_t << " not dirty, dirty_files empty, no-op" << dendl;
assert(!jump_to);
return 0;
}