Re: [PATCH v3 0/2] Align mmap address for DAX pmd mappings
From: Matthew Wilcox
Date: Tue Apr 19 2016 - 14:23:40 EST
On Mon, Apr 18, 2016 at 10:26:10PM +0200, Jan Kara wrote:
> On Fri 15-04-16 22:05:31, Andrew Morton wrote:
> > On Thu, 14 Apr 2016 10:48:29 -0600 Toshi Kani <toshi.kani@xxxxxxx> wrote:
> > > When CONFIG_FS_DAX_PMD is set, DAX supports mmap() using pmd page
> > > size. This feature relies on both mmap virtual address and FS
> > > block (i.e. physical address) to be aligned by the pmd page size.
> > > Users can use mkfs options to specify FS to align block allocations.
> > > However, aligning mmap address requires code changes to existing
> > > applications for providing a pmd-aligned address to mmap().
> > >
> > > For instance, fio with "ioengine=mmap" performs I/Os with mmap() .
> > > It calls mmap() with a NULL address, which needs to be changed to
> > > provide a pmd-aligned address for testing with DAX pmd mappings.
> > > Changing all applications that call mmap() with NULL is undesirable.
> > >
> > > This patch-set extends filesystems to align an mmap address for
> > > a DAX file so that unmodified applications can use DAX pmd mappings.
> > Matthew sounded unconvinced about the need for this patchset, but I
> > must say that
> > : The point is that we do not need to modify existing applications for using
> > : DAX PMD mappings.
> > :
> > : For instance, fio with "ioengine=mmap" performs I/Os with mmap().
> > : https://github.com/caius/fio/blob/master/engines/mmap.c
> > :
> > : With this change, unmodified fio can be used for testing with DAX PMD
> > : mappings. There are many examples like this, and I do not think we want
> > : to modify all applications that we want to evaluate/test with.
> > sounds pretty convincing?
> > And if we go ahead with this, it looks like 4.7 material to me - it
> > affects ABI and we want to get that stabilized asap. What do people
> > think?
> So I think Mathew didn't question the patch set as a whole. I think we all
> agree that we should align the virtual address we map to so that PMD
> mappings can be used. What Mathew was questioning was whether we really
> need to play tricks when logical offset in the file where mmap is starting
> is not aligned (and similarly for map length). Whether allowing PMD
> mappings for unaligned file offsets is worth the complication is IMO a
> valid question.
I was questioning the approach as a whole ... since we have userspace
already doing this in the form of NVML, do we really need the kernel to
do this for us?
Now, a further wrinkle. We have two competing patch sets (from Kirill
and Hugh) which are going to give us THP for page cache filesystems.
I would suggest that this is not DAX functionality but rather VFS
functionality to opportunistically align all mmaps on files which are
reasonably likely to be able to use THP.
I hadn't thought about this until earlier today, and I'm sorry I didn't
raise it further. Perhaps we can do a lightning session on this later
today at LSFMM since all six (Toshi, Andrew, Jan, Hugh, Kirill and myself)