Re: [RFC/PATCH] debug workqueue deadlocks with lockdep

[Johannes Berg <johannes@sipsolutions.net>]

[-- Attachment #1: Type: text/plain, Size: 1356 bytes --]

On Wed, 2007-07-04 at 16:52 +0400, Oleg Nesterov wrote:

> Yes. And no other work (except a barrier) can run before the caller of
> wait_on_work() is woken.

Alright, thanks. Yes, then what you proposed makes a lot of sense, I'll
implement it.

> Aha, now I see where I was confused. Yes, we can't avoid the false positives
> with flush_workqueue().
> 
> I hope this won't be a problem, because almost every usage of flush_workqueue()
> is pointless nowadays. So even if we have a false positive, it probably
> means the code needs cleanups anyway.
> 
> But see below,

[...]

> If you are going to do this, may I suggest you to make 2 separate patches?
> Exactly because we can't avoid the false positives with flush_workqueue(),
> it would be nice if we have an option to revert the 2-nd patch if there are
> too many false positives (I hope this won't happen).

I've run this patch on my system for a few days now and only seen
exactly one warning; however, it's *not* actually a *false* positive,
it's a positive but it's also perfectly possible to deadlock if the
system is loaded more than one work item is stuck on the workqueue for
some reason. Say A takes L1 and B runs without locks, and then you flush
the workqueue under L1; you'll get a real deadlock when both A and B are
actually scheduled to run!

johannes

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 190 bytes --]