Re: [patch 12/14] drivers/dmapool: use TASK_UNINTERRUPTIBLE instead of TASK_INTERRUPTIBLE

[Nish Aravamudan <nish.aravamudan@gmail.com>]

On Sun, 6 Mar 2005 19:44:14 -0800, Andrew Morton <akpm@osdl.org> wrote:
> domen@coderock.org wrote:
> >
> >  use TASK_UNINTERRUPTIBLE  instead of TASK_INTERRUPTIBLE
> >
> >  Signed-off-by: Nishanth Aravamudan <nacc@us.ibm.com>
> >  Signed-off-by: Domen Puncer <domen@coderock.org>
> >  ---
> >
> >
> >   kj-domen/drivers/base/dmapool.c |    2 +-
> >   1 files changed, 1 insertion(+), 1 deletion(-)
> >
> >  diff -puN drivers/base/dmapool.c~task_unint-drivers_base_dmapool drivers/base/dmapool.c
> >  --- kj/drivers/base/dmapool.c~task_unint-drivers_base_dmapool        2005-03-05 16:11:21.000000000 +0100
> >  +++ kj-domen/drivers/base/dmapool.c  2005-03-05 16:11:21.000000000 +0100
> >  @@ -293,7 +293,7 @@ restart:
> >               if (mem_flags & __GFP_WAIT) {
> >                       DECLARE_WAITQUEUE (wait, current);
> >
> >  -                    current->state = TASK_INTERRUPTIBLE;
> >  +                    set_current_state(TASK_UNINTERRUPTIBLE);
> >                       add_wait_queue (&pool->waitq, &wait);
> >                       spin_unlock_irqrestore (&pool->lock, flags);
> 
> This code is alread a bit odd.  If we're prepared to sleep in there, then
> why use GFP_ATOMIC?
> 
> If it is so that we can dig a bit deeper into the free page pools then
> something like __GFP_WAIT|__GFP_HIGH would be preferable.
> 
> And why isn't mem_flags passed into pool_alloc_page() verbatim?

Sorry, far beyond my abilities :(
 
> I agree on the TASK_UNINTERRUPTIBLE change: if the calling task happens to
> have signal_pending() then the schedule_timeout() will fall right through.
> Why should we change kernel memory allocation strategy if the user hit ^C?

Yup, didn't make much sense to me.
 
> Also, __set_current_state() can be user here: the add_wait_queue() contains
> the necessary barriers.  (Grubby, but we do that in quite a few places with
> this particular code sequence (we should have an add_wait_queue() variant
> which does the add_wait_queue+__set_current_state all in one hit (but let's
> not, else I'll be buried in another 1000 cleanuplets))).

Ok, I will re-spin this patch. Or would you prefer an incremental one?

Thanks,
Nish