Re: [PATCH] mtd/chips: add missing set_current_state() to cfi_{amdstd,staa}_sync()

["Dmitry Adamushko" <dmitry.adamushko@gmail.com>]

On 08/04/2008, Andrew Morton <akpm@linux-foundation.org> wrote:
> On Thu, 03 Apr 2008 21:38:23 +0200
>  Dmitry Adamushko <dmitry.adamushko@gmail.com> wrote:
>
>  > From: Dmitry Adamushko <dmitry.adamushko@gmail.com>
>  > Subject: [mtd/chips] add missing set_current_state() to cfi_{amdstd,staa}_sync()
>  >
>  > cfi_amdstd_sync() and cfi_staa_sync() call schedule() without changing
>  > task's state appropriately.
>  >
>  > In case of e.g. chip->state == FL_ERASING, cfi_*_sync() will be busy-looping
>  > either redundantly for a fixed interval of time (for SCHED_NORMAL tasks) or
>  > possibly endlessly (for RT tasks and UP).
>  >
>  > Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
>  >
>  > ---
>  >
>  > diff --git a/drivers/mtd/chips/cfi_cmdset_0002.c b/drivers/mtd/chips/cfi_cmdset_0002.c
>  > index d072e87..458d477 100644
>  > --- a/drivers/mtd/chips/cfi_cmdset_0002.c
>  > +++ b/drivers/mtd/chips/cfi_cmdset_0002.c
>  > @@ -1763,6 +1763,7 @@ static void cfi_amdstd_sync (struct mtd_info *mtd)
>  >
>  >               default:
>  >                       /* Not an idle state */
>  > +                     set_current_state(TASK_UNINTERRUPTIBLE);
>  >                       add_wait_queue(&chip->wq, &wait);
>  >
>  >                       spin_unlock(chip->mutex);
>  > diff --git a/drivers/mtd/chips/cfi_cmdset_0020.c b/drivers/mtd/chips/cfi_cmdset_0020.c
>  > index b344ff8..492e2ab 100644
>  > --- a/drivers/mtd/chips/cfi_cmdset_0020.c
>  > +++ b/drivers/mtd/chips/cfi_cmdset_0020.c
>  > @@ -1015,6 +1015,7 @@ static void cfi_staa_sync (struct mtd_info *mtd)
>  >
>  >               default:
>  >                       /* Not an idle state */
>  > +                     set_current_state(TASK_UNINTERRUPTIBLE);
>  >                       add_wait_queue(&chip->wq, &wait);
>  >
>  >                       spin_unlock_bh(chip->mutex);
>
>
> The change certainly looks correct.  Has it been runtime tested?

It has been tested with an oldish 2.6.8.1 where the problem initially
occured. It was a RT task that happened to close an mtd fd (and
resulting in ->sync() being called) and ran in the middle of the
->erase op... resulting in a "nice" endless loop.

The versions of cfi_{amdstd,staa}_sync() in the mainline look not that
much different from the respective versions in 2.6.8.1.

I guess, it went unnoticed for so long time due to :

(1) apps. don't often directly open/close() fd for mtd partitions;
(2) a race against ->erase() (or smth else) is rare;
(3) if (1) is not true, then an app. is still unlikely to be RT (in
which case, a task just loops until its timeslice is gone).


>
>  Thanks.
>

-- 
Best regards,
Dmitry Adamushko