Ceph » RADOS

We put code in cache tier eviction to check the scrub range, but that isn't sufficient. During scrub preemption restart we are selecting a smaller chunk. This allows eviction to evict something that had already been sent to the replica. The cleaned_meta_map can buffer the ScrubMap for a new evicted snapshot. The eviction will be received by the replica and the head object will be modified before _scan_snaps() is called with the removed snapshot still present.

snaps missing in mapper, should be: 188,18f,191,195,197,198,199,19d,19e,1a1,1a3 was r -2...repaired

http://pulpito.ceph.com/kchai-2018-05-05_14:56:43-rados-wip-kefu-testing-2018-05-05-1912-distro-basic-smithi/

$ git br --contains 83861a5b75ddb98366f1ec106487b88703f25cf7 | grep wip-kefu-testing-2018-05-05-1912

• wip-kefu-testing-2018-05-05-1912

Piecing together the sequence of events:

Chunk including smithi16216580-301...:2e but NOT its head object sent to replica
Replica stores smithi16216580-301...:2e in cleaned_meta_map
Preemption causes new chunk which does NOT include smithi16216580-301...:2e
Eviction is able to delete smithi16216580-301...:2e from cache tier
Replica sees delete and removes snapshot from head object
Scrub gets back to handling a chunk which includes smithi16216580-301...:head
Replica _scan_snaps() outputs error because snapshot 2e is present but head doesn't include it.