os/bluestore/BlueFS: do not start racing async compaction #11010
Conversation
Compaction is triggred from sync_metadata. If one compaction is in progress and another thread also calls sync_metadata, do not trigger a second async compaction! Signed-off-by: Sage Weil <sage@redhat.com>
|
Will try.. |
| << " ratio " << ratio | ||
| << (log_flushing ? " (async compaction in progress)" : "") | ||
| << dendl; | ||
| if (log_flushing || |
There was a problem hiding this comment.
Looks good.
Should we wait till the compaction finishes? Anyways we will not have enough work in compaction if we trigger also. May be we can wait till the old extents are released.
|
I think we can just skip it in this call path. We're only doing compaction in sync_metadata() because it's a convenient time to check (and maybe block). It's probably a bad place to do it, though.. we probably want a separate thread doing the compaction so we don't affect calling threads. That's a bigger change... |
|
Agreed. we might have to get some throttling in place. We can have some naive algorithm to see how much we have to process or percentage to complete. So that we can make some decision recompact (if needed). |
Compaction is triggred from sync_metadata. If one compaction is
in progress and another thread also calls sync_metadata, do not
trigger a second async compaction!
Signed-off-by: Sage Weil sage@redhat.com