Re: [bug-report] task info hung problem in fb_deferred_io_work()

From: Nam Cao
Date: Fri Apr 19 2024 - 11:20:59 EST


On 2024-04-18 Patrik Jakobsson wrote:
> On Thu, Apr 18, 2024 at 4:05 PM Nam Cao <namcao@xxxxxxxxxxxxx> wrote:
> >
> > On 2024-04-18 Patrik Jakobsson wrote:
> > > This sounds similar to the SUSE bug [1]. We fixed it by reverting [2]
> > > in the SUSE kernel. The problem seems to be that flush_delayed_work()
> > > kills the timer and re-queues the work but doesn't guarantee that it
> > > is finished when returning. So when the device is closed, the
> > > fb_deferred_io_work() function can still be queued and tries to access
> > > memory that's been freed.
> >
> > flush_delayed_work() *does* guarantee the work is finished before
> > returning.
>
> Right, flush_work() does guarantee that the work is finished, but it
> doesn't guarantee that the queue is idle if work has been requeued
> since flush started. So fb_deferred_io_work() should be able to happen
> after the device is closed. Or am I missing something?

I'm confused: how is it possible for fb_deferred_io_work() to run after
the device is closed?

The workqueue is flushed during closing of device. If more work is
queued during/after the flushing, then someone must be writing to the
device during/after the closing of the device. How can that happen?

Best regards,
Nam

> >
> > > Patch [2] tries to solve the problem of not throwing away pending data
> > > when closing the device. Perhaps calling cancel_delayed_work_sync()
> > > and then follow up with a manual call to fb_deferred_io_work() would
> > > be enough to flush the remaining data?
> > >
> > > -Patrik
> > >
> > > [1] https://bugzilla.suse.com/show_bug.cgi?id=1221814
> > > [2] 33cd6ea9c067 fbdev: flush deferred IO before closing
> > >
> > > >
> > > >
> > > > Thanks,
> > > > Harshit
> >