[Tarantool-patches] [PATCH v27 2/3] qsync: order access to the limbo terms

Vladislav Shpilevoy v.shpilevoy at tarantool.org
Thu Jan 13 00:30:14 MSK 2022


Hi!

On 12.01.2022 15:01, Serge Petrenko wrote:
> 
> 
> 11.01.2022 23:39, Cyrill Gorcunov пишет:
>> On Mon, Jan 10, 2022 at 05:28:43PM +0300, Serge Petrenko wrote:
>>>     Hi! Thanks for the patch!
>>>          box_issue_promote() and box_issue_demote() need fine-grained locking
>>>     anyway.
>>>     Otherwise it’s possible that promote() is already issued, but not yet
>>>     written to WAL, and some
>>>     outdated request is applied by applier at that exact moment.
>> True. And in previous series Vlad has asked to not move in code which is
>> not covered by tests. So I think this is a task for the next part. Currently
>> we cover only the race between appliers.
> 
> Let's ask Vlad, then.
> 
> I feel like we should fix this now, not waiting for a full fine-grained locking
> patch.
> 
> First of all, this is a known bug (and fine-grained locking was meant to
> cover everything we don't know of, just in case).

I am not sure I understand what you both are talking about here. Sergey, do
you mean 'fine-grained locking' as big critical sections covering a lot of
code at once or as many small critical sections?

I am confused because of this sentence. "Cover everything we don't know" is
rather opposite to fine-grained locking. I voted for big locks because
apparently it was too hard to implement smaller more precise locks.

> Besides, simply locking issue_promote/issue_demote should be
> much easier than implementing the fine-grained locking patch.

Yes. I remember the proposal was to lock entire promote/demote and other
qsync/raft functions from beginning to end. Because it should be relatively
easy. I didn't look at the code in this patch though, can't comment it.


More information about the Tarantool-patches mailing list