2009-06-16 20:45:56

by Mike Miller (OS Dev)

Subject: [PATCH 1/1] cciss: use timeout to ensure scan thread exits

Patch 1 of 1

This patch implements wait_for_completion_interruptible_timeout in the
scan_thread function to ensure it will exit cleanly during rmmod. Calling
complete in cciss_remove_one caused a race condition. Using the wait with a
timeout seems to work around that but it does fire the thread. The overhead
should be minimal.

Changelog:
Replace wait_for_completion_interruptible with
wait_for_completion_interruptible_timeout in scan_thread().
Use 5 second timeout value to avoid race.

Signed-off-by: Mike Miller

diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
index 4d4d5e0..76e7c10 100644
--- a/drivers/block/cciss.c
+++ b/drivers/block/cciss.c
@@ -3043,7 +3043,7 @@ static int scan_thread(void *data)
h->rescan_wait = &wait;

for (;;) {
- rc = wait_for_completion_interruptible(&wait);
+ rc = wait_for_completion_interruptible_timeout(&wait, 5);
if (kthread_should_stop())
break;
if (!rc)

2009-06-16 21:10:29

by Andrew Morton

Subject: Re: [PATCH 1/1] cciss: use timeout to ensure scan thread exits

On Tue, 16 Jun 2009 15:43:25 -0500
"Mike Miller (OS Dev)" <[email protected]> wrote:

> Patch 1 of 1
>
> This patch implements wait_for_completion_interruptible_timeout in the
> scan_thread function to ensure it will exit cleanly during rmmod. Calling
> complete in cciss_remove_one caused a race condition. Using the wait with a
> timeout seems to work around that but it does fire the thread. The overhead
> should be minimal.
>
> Changelog:
> Replace wait_for_completion_interruptible with
> wait_for_completion_interruptible_timeout in scan_thread().
> Use 5 second timeout value to avoid race.
>
> Signed-off-by: Mike Miller
>
> diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
> index 4d4d5e0..76e7c10 100644
> --- a/drivers/block/cciss.c
> +++ b/drivers/block/cciss.c
> @@ -3043,7 +3043,7 @@ static int scan_thread(void *data)
> h->rescan_wait = &wait;
>
> for (;;) {
> - rc = wait_for_completion_interruptible(&wait);
> + rc = wait_for_completion_interruptible_timeout(&wait, 5);
> if (kthread_should_stop())
> break;
> if (!rc)

c'mon Mike, that's a hack.

2009-06-16 21:56:18

by Mike Miller (OS Dev)

Subject: Re: [PATCH 1/1] cciss: use timeout to ensure scan thread exits

On Tue, Jun 16, 2009 at 02:10:07PM -0700, Andrew Morton wrote:
> On Tue, 16 Jun 2009 15:43:25 -0500
> "Mike Miller (OS Dev)" <[email protected]> wrote:
>
> > Patch 1 of 1
> >
> > This patch implements wait_for_completion_interruptible_timeout in the
> > scan_thread function to ensure it will exit cleanly during rmmod. Calling
> > complete in cciss_remove_one caused a race condition. Using the wait with a
> > timeout seems to work around that but it does fire the thread. The overhead
> > should be minimal.
> >
> > Changelog:
> > Replace wait_for_completion_interruptible with
> > wait_for_completion_interruptible_timeout in scan_thread().
> > Use 5 second timeout value to avoid race.
> >
> > Signed-off-by: Mike Miller
> >
> > diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
> > index 4d4d5e0..76e7c10 100644
> > --- a/drivers/block/cciss.c
> > +++ b/drivers/block/cciss.c
> > @@ -3043,7 +3043,7 @@ static int scan_thread(void *data)
> > h->rescan_wait = &wait;
> >
> > for (;;) {
> > - rc = wait_for_completion_interruptible(&wait);
> > + rc = wait_for_completion_interruptible_timeout(&wait, 5);
> > if (kthread_should_stop())
> > break;
> > if (!rc)
>
> c'mon Mike, that's a hack.

I'm open to suggestions. Calling complete in cciss_remove_one does not
resolve the issue. Other things I've tried may work once or twice but always
results in the hang.
Something I noticed in the trace was 2 calls to rebuild_lun_table when
trying to rmmod. Then it would get into the scheduler and just hang. During
rmmod I don't want to call rebuild_lun_table, just kill the thread and exit.

Ideas, suggestions, flames? Anyone????

-- mikem