Ceph

I don't understand how this can happen. When we start backfilling, we set pi.last_update to info.last_update (as part of activate). I think this patch is wrong and the bug you observed was something else. Please explain in more detail what is going on here.

For pg 1.7d
In timestamp 2016-02-11 23:40:51.235507 osd.2 is auth_log_shard and it's log_tail is 191'185, the last_update of osd.0 is 205'198, so the osd.0 is not selected as backfil_targets

But in 2016-02-11 23:40:51.238756 aften proc_replica_log, the last_update of osd.0 is down to 173'179, the last_update of osd.0(173'179 ) < the log_tail of osd.2(191'185)

so in PG::activate, the osd.0 is should be backfill, but is not in backfill_targets. so needs_backfill would return false, pg is not go to backfill state.

I upload some logs of osd.2

Please point out my mistakes if exits.

Thanks

Hmm, I don't have 33ff281d60c825adf1889eb67c4fed280ba0ea6f. Can you push a branch corresponding to that commit?