Re: [PATCH 3/3] kernel/workqueue: Suppress a false positive lockdep complaint

[Johannes Berg]

On Thu, 2018-10-25 at 15:05 +0000, Bart Van Assche wrote:
> It can happen that the direct I/O queue creates and destroys an empty
> workqueue from inside a work function.

So, thinking about this more, can you guarantee (somehow) that the
workqueue is empty at this point? Perhaps then we should encode that
guarantee into the API, and actually make it WARN_ON() if something is
scheduled on it at that point? And then skip lockdep in this case.

You've actually done a superset of this here - you've basically said
this workqueue was never used, not just that it's empty right now.
However, if you could guarantee that it was never used at this point, I
guess I fail to see what you need it for anyway (or alternatively, why
you wouldn't allocate it later after this point.)

It does seem strange to me to make lockdep contingent on "has this
workqueue been used" because there may be things queued on it that only
happen in certain cases (hardware, phases of the moon ;-) ) and lockdep
sort of says "once you execute all the paths that _could_ lead to a
deadlock, I'll warn you about it, even if it never happened". It seems
that we'd want to extend that to workqueues.

Also, I think in a way what this does is try to differentiate the
classes of workqueues: used vs. not used per instance, but I'm not
really sure that's what you want - after all, could it really never have
been used? Surely it has a purpose?

But anyway, I'll need to take a better look at the lockdep splat later
(i.e. when the kids are sleeping.) Do you know how to reproduce this?

johannes