librbd: optimize away unnecessary object map updates #10332
Conversation
|
@neurodrone Hopefully this should resolve the mkfs discard performance issue. It needs to run through the rbd teuthology suite to ensure a regression wasn't introduced. |
|
That's neat! Thanks for this fix! |
|
@neurodrone Thanks for the traces -- made it really easy to reproduce. |
|
woooot, good job guys! |
librbd: optimize away unnecessary object map updates #10332
|
@dillaman 'Fixes: http://tracker.ceph.com/issues/16707' to "rbd-replay: decode and replay discard IO operations" commit log? |
| imagectx, offset, length)); | ||
| thread->issued_io(io, &m_latest_ios); | ||
| ios->push_back(io); | ||
| m_pending_ios[completion] = io; |
There was a problem hiding this comment.
@dillaman This looks like wrong for not AIO case?
Fixes: http://tracker.ceph.com/issues/16707 Signed-off-by: Jason Dillaman <dillaman@redhat.com>
Fixes: http://tracker.ceph.com/issues/16689 Signed-off-by: Jason Dillaman <dillaman@redhat.com>
|
@trociny good catch -- update pushed |
|
@dillaman Do I understand the fix correctly, that the source of unnecessary object map updates was in send_pre() where we filtered out only updates that did not change the state, while now we also filter out updates changing state form OBJECT_NONEXISTENT to OBJECT_PENDING and from !OBJECT_PENDING to OBJECT_NONEXISTENT? Or was send_post() a source of unnecessary updates too? |
|
@trociny There wasn't an issue with the send_post state -- I just consolidated the logic to the same helper method. The issue was send_pre would only perform an exact check of current state against desired state. When deleting, the desired state is first the _PENDING state. Therefore, we can avoid updating the object map if our desired state is _PENDING and the current state is _NONEXISTENT. |
|
lgtm |
librbd: optimize away unnecessary object map updates #10332
librbd: optimize away unnecessary object map updates #10332
librbd: optimize away unnecessary object map updates #10332
No description provided.