Subject: [PATCH] APM: fix deadlock in APM_IOC_SUSPEND ioctl

Subject: [PATCH] APM: fix deadlock in APM_IOC_SUSPEND ioctl

[NeilBrown]

[-- Attachment #1: Type: text/plain, Size: 1113 bytes --]



I found the Xorg server on my ARM device stuck in the 'msleep()' loop
in apm_ioctl.

I suspect it had attempted suspend immediately after resuming and lost
a race.
During that msleep(10);, a new suspend cycle must have started and
changed ->suspend_state to SUSPEND_PENDING, so it was never seen to
be SUSPEND_DONE and the loop could never exited.  It would have moved on
to SUSPEND_ACKTO but never been able to reach SUSPEND_DONE.

So change the loop to only run while SUSPEND_ACKED rather than until
SUSPEND_DONE.  This is much safer.

Signed-off-by: NeilBrown <neilb@suse.de>

diff --git a/drivers/char/apm-emulation.c b/drivers/char/apm-emulation.c
index f4837a8..6005c5c 100644
--- a/drivers/char/apm-emulation.c
+++ b/drivers/char/apm-emulation.c
@@ -302,7 +302,7 @@ apm_ioctl(struct file *filp, u_int cmd, u_long arg)
 			 * anything critical, chill a bit on each iteration.
 			 */
 			while (wait_event_freezable(apm_suspend_waitqueue,
-					as->suspend_state == SUSPEND_DONE))
+					as->suspend_state != SUSPEND_ACKED))
 				msleep(10);
 			break;
 		case SUSPEND_ACKTO:

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 828 bytes --]

Re: Subject: [PATCH] APM: fix deadlock in APM_IOC_SUSPEND ioctl

[Jiri Kosina]

On Sun, 1 Apr 2012, NeilBrown wrote:

> 
> 
> I found the Xorg server on my ARM device stuck in the 'msleep()' loop
> in apm_ioctl.
> 
> I suspect it had attempted suspend immediately after resuming and lost
> a race.
> During that msleep(10);, a new suspend cycle must have started and
> changed ->suspend_state to SUSPEND_PENDING, so it was never seen to
> be SUSPEND_DONE and the loop could never exited.  It would have moved on
> to SUSPEND_ACKTO but never been able to reach SUSPEND_DONE.
> 
> So change the loop to only run while SUSPEND_ACKED rather than until
> SUSPEND_DONE.  This is much safer.
> 
> Signed-off-by: NeilBrown <neilb@suse.de>
> 
> diff --git a/drivers/char/apm-emulation.c b/drivers/char/apm-emulation.c
> index f4837a8..6005c5c 100644
> --- a/drivers/char/apm-emulation.c
> +++ b/drivers/char/apm-emulation.c
> @@ -302,7 +302,7 @@ apm_ioctl(struct file *filp, u_int cmd, u_long arg)
>  			 * anything critical, chill a bit on each iteration.
>  			 */
>  			while (wait_event_freezable(apm_suspend_waitqueue,
> -					as->suspend_state == SUSPEND_DONE))
> +					as->suspend_state != SUSPEND_ACKED))
>  				msleep(10);
>  			break;
>  		case SUSPEND_ACKTO:

Good catch Neil! I have missed this race when we were moving away from 
freezer_*_count() to wait_event_freezable() here.

Applied, thanks.

-- 
Jiri Kosina
SUSE Labs