osd: Use one timestamp for picking recovery and arming its retry timer #146
Reference in New Issue
Block a user
No description provided.
Delete Branch "koolkhel/vitastor:osd-recovery-retry-clock"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Description
Followup to #143.
pick_next_recovery()andschedule_recovery_retry()each read the clock. A PG pause that ended between the two reads was skipped by the pick but not counted by the timer, so no timer was armed and the PG's recovery waited for an unrelated event (a PG state change, another recovery operation finishing). Nowcontinue_recovery()takes the time once for both, and the timer is armed for at least 1 us in case the pause already ended.Tested:
tests/test_recovery_enospc.shpasses (25 ENOSPC recovery failures, recovery resumes by the timer afterrm-data). The window is microseconds wide, so there is no deterministic test for it.Contributor Name
Yury Luneff (MIND Software LLC, https://mindsw.io)
CLA
Rebase please
Rebased onto current master (
483f5818); only the context changed around the new tombstone loop inpick_next_recovery(), the patch lines are the same.tests/test_recovery_enospc.shpasses after the rebase.1b127186fato332c06989f