Re: [PATCH] xfs: allocate sector sized IO buffer via page_frag_alloc

[Ming Lei <ming.lei@redhat.com>]

On Tue, Feb 26, 2019 at 04:12:09AM -0800, Matthew Wilcox wrote:
> On Tue, Feb 26, 2019 at 07:12:49PM +0800, Ming Lei wrote:
> > On Tue, Feb 26, 2019 at 6:07 PM Vlastimil Babka <vbabka@suse.cz> wrote:
> > > On 2/26/19 10:33 AM, Ming Lei wrote:
> > > > On Tue, Feb 26, 2019 at 03:58:26PM +1100, Dave Chinner wrote:
> > > >> On Mon, Feb 25, 2019 at 07:27:37PM -0800, Matthew Wilcox wrote:
> > > >>> On Tue, Feb 26, 2019 at 02:02:14PM +1100, Dave Chinner wrote:
> > > >>>>> Or what is the exact size of sub-page IO in xfs most of time? For
> > > >>>>
> > > >>>> Determined by mkfs parameters. Any power of 2 between 512 bytes and
> > > >>>> 64kB needs to be supported. e.g:
> > > >>>>
> > > >>>> # mkfs.xfs -s size=512 -b size=1k -i size=2k -n size=8k ....
> > > >>>>
> > > >>>> will have metadata that is sector sized (512 bytes), filesystem
> > > >>>> block sized (1k), directory block sized (8k) and inode cluster sized
> > > >>>> (32k), and will use all of them in large quantities.
> > > >>>
> > > >>> If XFS is going to use each of these in large quantities, then it doesn't
> > > >>> seem unreasonable for XFS to create a slab for each type of metadata?
> > > >>
> > > >>
> > > >> Well, that is the question, isn't it? How many other filesystems
> > > >> will want to make similar "don't use entire pages just for 4k of
> > > >> metadata" optimisations as 64k page size machines become more
> > > >> common? There are others that have the same "use slab for sector
> > > >> aligned IO" which will fall foul of the same problem that has been
> > > >> reported for XFS....
> > > >>
> > > >> If nobody else cares/wants it, then it can be XFS only. But it's
> > > >> only fair we address the "will it be useful to others" question
> > > >> first.....
> > > >
> > > > This kind of slab cache should have been global, just like interface of
> > > > kmalloc(size).
> > > >
> > > > However, the alignment requirement depends on block device's block size,
> > > > then it becomes hard to implement as genera interface, for example:
> > > >
> > > >       block size: 512, 1024, 2048, 4096
> > > >       slab size: 512*N, 0 < N < PAGE_SIZE/512
> > > >
> > > > For 4k page size, 28(7*4) slabs need to be created, and 64k page size
> > > > needs to create 127*4 slabs.
> > > >
> > >
> > > Where does the '*4' multiplier come from?
> > 
> > The buffer needs to be device block size aligned for dio, and now the block
> > size can be 512, 1024, 2048 and 4096.
> 
> Why does the block size make a difference?  This requirement is due to
> some storage devices having shoddy DMA controllers.  Are you saying there
> are devices which can't even do 512-byte aligned I/O?

Direct IO requires that, see do_blockdev_direct_IO().

This issue can be triggered when running xfs over loop/dio. We could
fallback to buffered IO under this situation, but not sure it is the
only case.


Thanks,
Ming