Re: work item still be scheduled to execute after destroy_workqueue?

[Lai Jiangshan]

On Thu, Dec 8, 2022 at 10:44 AM richard clark
<richard.xnu.clark@xxxxxxxxx> wrote:
>
> On Wed, Dec 7, 2022 at 10:38 AM Lai Jiangshan <jiangshanlai@xxxxxxxxx> wrote:
> >
> > On Tue, Dec 6, 2022 at 5:20 PM richard clark
> > <richard.xnu.clark@xxxxxxxxx> wrote:
> > >
> > > On Tue, Dec 6, 2022 at 2:23 PM Lai Jiangshan <jiangshanlai@xxxxxxxxx> wrote:
> > > >
> > > > On Tue, Dec 6, 2022 at 12:35 PM richard clark
> > > > <richard.xnu.clark@xxxxxxxxx> wrote:
> > > >
> > > > > >
> > > > > A WARN is definitely reasonable and has its benefits. Can I try to
> > > > > submit the patch and you're nice to review as maintainer?
> > > > >
> > > > > Thanks,
> > > > > Richard
> > > > > >
> > > >
> > > > Sure, go ahead.
> > > >
> > > > What's in my mind is that the following code is wrapped in a new function:
> > > >
> > > >         mutex_lock(&wq->mutex);
> > > >         if (!wq->nr_drainers++)
> > > >                 wq->flags |= __WQ_DRAINING;
> > > >         mutex_unlock(&wq->mutex);
> > > >
> > > >
> > > > and the new function replaces the open code drain_workqueue() and
> > > > is also called in destroy_workqueue() (before calling drain_workqueue()).
> > > >
> > > Except that, do we need to defer the __WQ_DRAINING clean to the
> > > rcu_call, thus we still have a close-loop of the drainer's count, like
> > > this?
> >
> > No, I don't think we need it. The wq is totally freed in rcu_free_wq.
> >
> > Or we can just introduce __WQ_DESTROYING.
> >
> > It seems using __WQ_DESTROYING is better.
>
> The wq->flags will be unreliable after kfree(wq), for example, in my
> machine, the wq->flags can be 0x7ec1e1a3, 0x37cff1a3 or 0x7fa23da3 ...
> after wq be kfreed, consequently the result of queueing a new work
> item to a kfreed wq is undetermined, sometimes it's ok because the
> queue_work will return directly(e.g, the wq->flags = 0x7ec1e1a3, a
> fake __WQ_DRAINING state), sometimes it will trigger a kernel NULL
> pointer dereference BUG when the wq->flags = 0x7fa23da3(fake
> !__WQ_DRAINING state).

The whole wq is unreliable after destroy_workqueue().

All we need is just adding something to help identify any
wrong usage while the wq is in RCU grace period.

>
> IMO, given the above condition,  we can handle this in 2 phases:
> before the rcu_call and after.
> a. before rcu_call. Using __WQ_DESTROYING to allow the chained work
> queued in or not in destroy_workqueue(...) level, __WQ_DRAINING is
> used to make the drain_workqueue(...) still can be standalone. The
> code snippet like this:
> destroy_workqueue(...)
> {
>         mutex_lock(&wq->mutex);
>         wq->flags |= __WQ_DESTROYING;
>         mutex_lock(&wq->mutex);

Ok, put it before calling drain_workqueue()

>         ...
> }
>
> __queue_work(...)
> {
>           if (unlikely((wq->flags & __WQ_DESTROYING) || (wq->flags &
> __WQ_DRAINING)) &&
>                    WARN_ON_ONCE(!is_chained_work(wq)))

Ok, combine __WQ_DESTROYING and __WQ_DRAINING together as:
           if (unlikely((wq->flags & (__WQ_DESTROYING | __WQ_DRAINING)) &&


>          return;
> }
>
> b. after rcu_call. What in my mind is:
> rcu_free_wq(struct rcu_head *rcu)
> {
>           ...
>           kfree(wq);
>           wq = NULL;

It is useless code.

> }
>
> __queue_work(...)
> {
>         if (!wq)
>                 return;

It is useless code.

>         ...
> }
>
> Any comments?
>
> >
> > >
> > > --- a/kernel/workqueue.c
> > > +++ b/kernel/workqueue.c
> > >
> > > @@ -3528,6 +3526,9 @@ static void rcu_free_wq(struct rcu_head *rcu)
> > >
> > >         else
> > >                 free_workqueue_attrs(wq->unbound_attrs);
> > >
> > > +       if (!--wq->nr_drainers)
> > > +               wq->flags &= ~__WQ_DRAINING;
> > > +
> > >         kfree(wq);
> > >
> > > >
> > > > __WQ_DRAINING will cause the needed WARN on illegally queuing items on
> > > > destroyed workqueue.
> > >
> > > I will re-test it if there are no concerns about the above fix...
> > >
> > > >
> > > > Thanks
> > > > Lai