osd: requeue replays we couldn't start in this tick
PATCH 1/1 osd: requeue replays we couldn't start in this tick
If a PG is in replay but it's not active yet by the time the OSD tick gets to running the replay queue, we'll drop the PG from the replay queue without clearing its replay state.
This patch arranges for PGs that weren't ready for replaying (or that weren't replayed for any other reason) to be requeued, so that they will be retried in the next tick.
I have confirmed, with this patch, that my memory-starved nodes would often trigger the new log message, indicating the PGs were not active yet. I have not observed instances of non-primary or replay_until mismatch, and I'm not sure the latter should justify a requeue, an assertion failure or whatever, but simply dropping the PG on the floor leaving it indefinitely in replay state is not something that should be done, since it prevents other operations from being processed.
This patch does NOT fix the other problem reported in Issue #8758, that restarting any of the osds holding a PG is enough to clear the replay state.
--- src/osd/OSD.cc | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
src/osd/OSD.cc +22 -0
@@ -7228,6 +7228,7 @@ void OSD::check_replay_queue() } replay_queue_lock.Unlock(); + list< pair<spg_t,utime_t> > requeue; for (list< pair<spg_t,utime_t> >::iterator p = pgids.begin(); p != pgids.end(); ++p) { spg_t pgid = p->first; if (pg_map.count(pgid)) {
@@ -7239,12 +7240,33 @@ void OSD::check_replay_queue() pg->replay_until == p->second) { pg->replay_queued_ops(); } + if (pg->is_replay()) { + dout(12) << "check_replay_queue requeueing " << pgid + << (pg->is_active() ? "" : ", not active") + << (pg->is_primary() ? "" : ", not primary") + << (pg->replay_until == p->second ? "" + : ", mismatched replay_until") + << ", pg->until " << pg->replay_until + << ", queued " << p->second << dendl; + requeue.push_back(*p); + } pg->unlock(); } else { dout(10) << "check_replay_queue pgid " << pgid << " (not found)" << dendl; } } + if (!requeue.empty()) { + replay_queue_lock.Lock(); + // reverse order to preserve it; push to the front of the queue, + // where these entries came from + while (!requeue.empty()) { + replay_queue.push_front(requeue.back()); + requeue.pop_back(); + } + replay_queue_lock.Unlock(); + } + // wake up _all_ pg waiters; raw pg -> actual pg mapping may have shifted wake_all_pg_waiters(); }
Signature · 4 lines
Alexandre Oliva, freedom fighter http://FSFLA.org/~lxoliva/ You must be the change you wish to see in the world. -- Gandhi Be Free! -- http://FSFLA.org/ FSF Latin America board member Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer
Ping?
If a PG is in replay but it's not active yet by the time the OSD tick gets to running the replay queue, we'll drop the PG from the replay queue without clearing its replay state.
This patch arranges for PGs that weren't ready for replaying (or that weren't replayed for any other reason) to be requeued, so that they will be retried in the next tick.
I have confirmed, with this patch, that my memory-starved nodes would often trigger the new log message, indicating the PGs were not active yet. I have not observed instances of non-primary or replay_until mismatch, and I'm not sure the latter should justify a requeue, an assertion failure or whatever, but simply dropping the PG on the floor leaving it indefinitely in replay state is not something that should be done, since it prevents other operations from being processed.
This patch does NOT fix the other problem reported in Issue #8758, that restarting any of the osds holding a PG is enough to clear the replay state.
--- src/osd/OSD.cc | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)
src/osd/OSD.cc +22 -0
@@ -7228,6 +7228,7 @@ void OSD::check_replay_queue() } replay_queue_lock.Unlock(); + list< pair<spg_t,utime_t> > requeue; for (list< pair<spg_t,utime_t> >::iterator p = pgids.begin(); p != pgids.end(); ++p) { spg_t pgid = p->first; if (pg_map.count(pgid)) {
@@ -7239,12 +7240,33 @@ void OSD::check_replay_queue() pg->replay_until == p->second) { pg->replay_queued_ops(); } + if (pg->is_replay()) { + dout(12) << "check_replay_queue requeueing " << pgid + << (pg->is_active() ? "" : ", not active") + << (pg->is_primary() ? "" : ", not primary") + << (pg->replay_until == p->second ? "" + : ", mismatched replay_until") + << ", pg->until " << pg->replay_until + << ", queued " << p->second << dendl; + requeue.push_back(*p); + } pg->unlock(); } else { dout(10) << "check_replay_queue pgid " << pgid << " (not found)" << dendl; } } + if (!requeue.empty()) { + replay_queue_lock.Lock(); + // reverse order to preserve it; push to the front of the queue, + // where these entries came from + while (!requeue.empty()) { + replay_queue.push_front(requeue.back()); + requeue.pop_back(); + } + replay_queue_lock.Unlock(); + } + // wake up _all_ pg waiters; raw pg -> actual pg mapping may have shifted wake_all_pg_waiters(); }
Signature · 4 lines
Alexandre Oliva, freedom fighter http://FSFLA.org/~lxoliva/ You must be the change you wish to see in the world. -- Gandhi Be Free! -- http://FSFLA.org/ FSF Latin America board member Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer
Updated to master and pushed as wip-8758 for the moment. -Sam
Quoted 85 lines · Alexandre Oliva
Ping?
If a PG is in replay but it's not active yet by the time the OSD tick gets to running the replay queue, we'll drop the PG from the replay queue without clearing its replay state.
This patch arranges for PGs that weren't ready for replaying (or that weren't replayed for any other reason) to be requeued, so that they will be retried in the next tick.
I have confirmed, with this patch, that my memory-starved nodes would often trigger the new log message, indicating the PGs were not active yet. I have not observed instances of non-primary or
replay_untilmismatch, and I'm not sure the latter should justify a requeue, an assertion failure or whatever, but simply dropping the PG on the floor leaving it indefinitely in replay state is not something that should be done, since it prevents other operations from being processed.This patch does NOT fix the other problem reported in Issue #8758, that restarting any of the osds holding a PG is enough to clear the replay state.
Signed-off-by Alexandre Oliva <[email protected]>--- src/osd/OSD.cc | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+)src/osd/OSD.cc +22
-1@@ -7228,6 +7228,7 @@ void OSD::check_replay_queue() } replay_queue_lock.Unlock(); + list< pair<spg_t,utime_t> > requeue; for (list< pair<spg_t,utime_t> >::iterator p = pgids.begin(); p != pgids.end(); ++p) { spg_t pgid = p->first; if (pg_map.count(pgid)) {@@ -7239,12 +7240,33 @@ void OSD::check_replay_queue() pg->replay_until == p->second) { pg->replay_queued_ops(); } + if (pg->is_replay()) { + dout(12) << "check_replay_queue requeueing " << pgid + << (pg->is_active() ? "" : ", not active") + << (pg->is_primary() ? "" : ", not primary") + << (pg->replay_until == p->second ? "" + : ", mismatched replay_until") + << ", pg->until " << pg->replay_until + << ", queued " << p->second << dendl; + requeue.push_back(*p); + } pg->unlock(); } else { dout(10) << "check_replay_queue pgid " << pgid << " (not found)" << dendl; } } + if (!requeue.empty()) { + replay_queue_lock.Lock(); + // reverse order to preserve it; push to the front of the queue, + // where these entries came from + while (!requeue.empty()) { + replay_queue.push_front(requeue.back()); + requeue.pop_back(); + } + replay_queue_lock.Unlock(); + } + // wake up _all_ pg waiters; raw pg -> actual pg mapping may have shifted wake_all_pg_waiters(); } --Alexandre Oliva, freedom fighter http://FSFLA.org/~lxoliva/
You must be the change you wish to see in the world. -- Gandhi
Be Free! -- http://FSFLA.org/ FSF Latin America board member
Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer --
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html