Re: [PATCH 2/2] drop vmerge accounting

[Jens Axboe <jens.axboe@oracle.com>]

On Fri, Aug 08 2008, Mikulas Patocka wrote:
> On Fri, 8 Aug 2008, Mikulas Patocka wrote:
> 
> > On Fri, 8 Aug 2008, Jens Axboe wrote:
> > 
> > > On Fri, Aug 08 2008, FUJITA Tomonori wrote:
> > > > On Fri, 8 Aug 2008 08:15:37 +0200
> > > > Jens Axboe <jens.axboe@oracle.com> wrote:
> > > > 
> > > > > > BTW, we also wait for your verdict on:
> > > > > > 
> > > > > > http://marc.info/?t=121611935500002&r=1&w=2
> > > > > 
> > > > > I've always hated the iommu virtual merging complexity. My plan is to
> > > > > rip it out.
> > > > 
> > > > No complaint from me. I'm just happy to see the verdict at length.
> > > 
> > > Mikulas, you had a patch for this. Can you strip out the arch bits and
> > > just send me the block bits? The arch bits should go in via the arch
> > > maintainers.
> > > 
> > > The iommu code may still do virtual merging, it would be silly not to do
> > > that if possible. Now that later kernels expose the necessary parameters
> > > at that level as well, it's perfectly feasible.
> > 
> > Here I'm sending the first one, it removes virtual merge accounting from 
> > blk-merge.c (it acts as if all architectures undefined 
> > BIO_VMERGE_BOUNDARY). I also created second patch that removes 
> > bi_hw_segments field from struct bio and struct request. The next task 
> > would be to remove nr_hw_segments from the request_queue, but it will 
> > require changing the drivers.
> > 
> > Mikulas
> 
> (note: raid5 uses nr_hw_segments field for some other purpose, so this 
> patch will break it. It should be fixed by raid5 maintainers)
> 
> Remove hw_segments field from struct bio and struct request. Without virtual
> merge accounting they have no purpose.

I have applied this and the previous, with a fixup to raid5 to overload
both active stripe and process count in bi_phys_segments.

-- 
Jens Axboe