Re: [PATCH] Reset file->f_op in snd_card_file_remove(). Take 2

[Karsten Wiese <annabellesgarden@yahoo.de>]

[-- Attachment #1: Type: text/plain, Size: 3114 bytes --]

Am Freitag, 29. September 2006 14:45 schrieb Takashi Iwai:
> At Fri, 29 Sep 2006 14:29:20 +0200,
> Karsten Wiese wrote:
> > 
> > Am Freitag, 29. September 2006 12:48 schrieb Takashi Iwai:
> > > At Thu, 28 Sep 2006 22:28:02 +0200,
> > > Karsten Wiese wrote:
> > > > 
> > > > Hi
> > > > 
> > > > It oopses with 2.6.18-rt4 + alsa-kernel-1.0.13rc3 now.
> > > > I wrote before, 2.6.18-rt3 + alsa-driver-1.0.13rc3 would be ok,
> > > > but its not. bug showed again reliably under memory-pressure.
> > > > 
> > > >       Karsten
> > > > 
> > > > ===
> > > > 
> > > > Reset file->f_op in snd_card_file_remove(). Take 2
> > > > 
> > > > 
> > > > i think what happens here is:
> > > > 
> > > >   us428control runs, kernel has allocated a struct file for /dev/hwC1D0.
> > > > 
> > > >   usb disconnect
> > > > 
> > > >   snd_usb_usx2y calls snd_card_disconnect,
> > > >   tells us428control to exit.
> > > > 
> > > >   snd_card_disconnect replaces /dev/hwC1D0's file->f_op
> > > >   with a kmalloc()ed version, that would only allow releases.
> > > > 
> > > >   us428control starts exiting
> > > > 
> > > >   __fput is called with struct file for /dev/hwC1D0.
> > > > 
> > > >   snd_card_file_remove() is called, alsa notices struct file
> > > >   for /dev/hwC1D0 is about to be closed.
> > > >   with patch below, file->f_op would be set NULL now.
> > > > 
> > > >   snd_usb_usx2y's free()s snd_card instance and /dev/hwC1D0's
> > > >   file->f_ops, those that would only allow releases.
> > > > 
> > > >   for reason I would like to know,
> > > >   __fput is called again with struct file for /dev/hwC1D0
> > > >   from us428control's do_exit().
> > > >   __fput see's file->f_op is still set.
> > > >   Without patch and under memory pressure, file->f_op can
> > > >   point to anything now.
> > > > 
> > > > 
> > > > Signed-off-by: Karsten Wiese <annabellesgarden@yahoo.de>
> > > 
> > > I guess this bug is fixed by Florin's patch below, juding from your
> > > explanation.  Could you check it?
> > > 
> > Florin's patch fixes it.
> > 
> > This one for immediate consumation by mainline, mm, rt,
> > and the stable teams, hmm?
> 
> It'll be pushed to mainline soon together with other ALSA fixes.  Then
> I'll forward to stable, too.
> 
bug is back. Florin's patch _is_ right.
And installed, unless im getting confuzed.
To help me proove,
please consider this controll of flow:

	usb disconnect

	usX2Y_usb_disconnect() calls snd_card_free(card);

	snd_card_free waits:
		wait_event(card->shutdown_sleep, card->files == NULL);

	us428control starts exiting, closes /dev/hwC1D0, calls __fput.
	__fput calls snd_hwdep_release,
	snd_hwdep_release calls snd_card_file_remove.
	snd_card_file_remove sees lastclose is set, does
		wake_up(&card->shutdown_sleep);

lets assume, snd_card_free's thread prio is FIFO, we are on UP
and us428control' prio is not FIFO:

What keeps snd_card_free from waking up now, deleting /dev/hwC1D0's
file->f_op _before_  __fput is rescheduled,
seeing a set but freeed file->f_op?

Attached crash dmesg also has "BUG: time warp detected!", but that BUG
is propably unrelated.


      Karsten






[-- Attachment #2: crash9.bz2 --]
[-- Type: application/x-bzip2, Size: 8395 bytes --]