Re: [PATCH] [9/20] Add blk_kmalloc/blk_alloc_pages

[FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>]

On Wed, 02 Apr 2008 11:43:21 +0300
Boaz Harrosh <bharrosh@panasas.com> wrote:

> On Wed, Apr 02 2008 at 6:37 +0300, FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp> wrote:
> > On Mon, 17 Mar 2008 22:34:22 +0100
> > Andi Kleen <andi@firstfloor.org> wrote:
> > 
> >>> Look further down in the email, queue_bounce_to_mask() or whatever you
> >>> would want to call it. As also written there, the PRINCIPLE is the same.
> >>> And that is that exporting a to_allocator_mask() helper is a lot saner
> >>> than exporting an allocator api tied to the queue.
> >>>
> >>> Can we get over this, please?
> >> Ok it would have helped if you had explained why it is saner, but I 
> >> bow to your superior experience on block layer issues.
> >>
> >> The only open issue is right now if it isn't better to go back
> >> for automatic bouncing for SCSI scan and the other users. Do you
> >> have an opinion on that too? If you have one can you please convince
> >> James of it too.
> > 
> > There are other things that don't want the automatic bouncing; sg, st,
> > and osst.
> > 
> > Your patches remove unchecked_isa_dma in them and Boaz said that they
> > are fine since they get bounced anyway, however, it's not correct.
> > 
> > As Doug said in another thread (about your patch to change GFP_ATOMIC
> > to GFP_KERNEL in sg), they try to avoid waiting for a long time and
> > want an early failure (though they are not complete; can't avoid non
> > unchecked_isa_dma bouncing) .
> > 
> > We can change the bounce path in that way, but I think it's better if
> > they can allocate memory that will not get bounced.
> 
> I was looking in blk_queue_bounce (in mm/bounce.c) and have not seen any
> case of q->bounce_gfp with __GFP_WAIT. Are you sure we are waiting for
> the bounce buffers? It will take few more CPU cycles, but I don't see
> where it will sleep.

Hmm, q->bounce_gfp is GFP_NOIO or (GFP_NOIO | __GFP_DMA). And
__blk_queue_bounce also allocates bio with GFP_NOIO. So the bounce
path always use __GFP_WAIT.