Re: Race condition in SyncRepGetSyncStandbysPriority

At Wed, 15 Apr 2020 11:35:58 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> I'm looking this more closer.

It looks to be in the right direction to me.

As mentioned in the previous mail, I removed is_sync_standby from
SycnStandbyData. But just doing that breaks pg_stat_get_wal_senders.
It is an exsting issue but the logic for sync_state (values[10]) looks
odd. Fixed in the attached.

SyncRepInitConfig uses mutex instead of SyncRepLock. Since anyway the
integrity of sync_standby_priority is not guaranteed, it seems OK to
me. It seems fine to remove the assertion and requirement about
SyncRepLock from SyncRepGetSyncRecPtr for the same reason. (Actually
the lock is held, though.)

SyncRepGetSyncStandbysPriority doesn't seem worth existing as a
function. Removed in the attached.

+ num_standbys = SyncRepGetSyncStandbys(&sync_standbys);

The list is no longer consists only of synchronous standbys. I
changed the function name, variable name and tried to adjust related
comments.

It's not what the patch did, but I don't understand why
SyncRepGetNthLatestSyncRecPtr takes SyncRepConfig->num_sync but
SyncRepGetOldest.. accesses it directly. Changed the function
*Oldest* in the attached. I didn't do that but finally, the two
functions can be consolidated, just by moving the selection logic
currently in SyncRepGetSyncRecPtr into the new function.

The resulting patch is attached.

- removed is_sync_standby from SyncRepStandbyData
- Fixed the logic for values[10] in pg_stat_get_wal_senders
- Changed the signature of SyncRepGetOldestSyncRecPtr
- Adjusted some comments to the behavioral change of
SyncRepGet(Sync)Standbys.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center