[PATCH] mm: shmem: fix the shmem large folio allocation for the i915 driver

[Patryk Kowalczyk]

Hi,
This patch solves the performance issue very well.
best regards,
Patryk

śr., 30 lip 2025 o 09:46 Baolin Wang <baolin.wang at linux.alibaba.com> napisał(a):
>
>
>
> On 2025/7/30 14:54, Hugh Dickins wrote:
> > On Mon, 28 Jul 2025, Baolin Wang wrote:
> >
> >> After commit acd7ccb284b8 ("mm: shmem: add large folio support for tmpfs"),
> >> we extend the 'huge=' option to allow any sized large folios for tmpfs,
> >> which means tmpfs will allow getting a highest order hint based on the size
> >> of write() and fallocate() paths, and then will try each allowable large order.
> >>
> >> However, when the i915 driver allocates shmem memory, it doesn't provide hint
> >> information about the size of the large folio to be allocated, resulting in
> >> the inability to allocate PMD-sized shmem, which in turn affects GPU performance.
> >>
> >> To fix this issue, add the 'end' information for shmem_read_folio_gfp()  to help
> >> allocate PMD-sized large folios. Additionally, use the maximum allocation chunk
> >> (via mapping_max_folio_size()) to determine the size of the large folios to
> >> allocate in the i915 driver.
> >>
> >> Fixes: acd7ccb284b8 ("mm: shmem: add large folio support for tmpfs")
> >> Reported-by: Patryk Kowalczyk <patryk at kowalczyk.ws>
> >> Reported-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >> Tested-by: Patryk Kowalczyk <patryk at kowalczyk.ws>
> >> Signed-off-by: Baolin Wang <baolin.wang at linux.alibaba.com>
> >> ---
> >>   drivers/gpu/drm/drm_gem.c                 | 2 +-
> >>   drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 7 ++++++-
> >>   drivers/gpu/drm/ttm/ttm_backup.c          | 2 +-
> >>   include/linux/shmem_fs.h                  | 4 ++--
> >>   mm/shmem.c                                | 7 ++++---
> >>   5 files changed, 14 insertions(+), 8 deletions(-)
> >
> > I know I said "I shall not object to a temporary workaround to suit the
> > i915 driver", but really, I have to question this patch.  Why should any
> > change be required at the drivers/gpu/drm end?
> >
> > And in drivers/gpu/drm/{i915,v3d} I find they are using huge=within_size:
> > I had been complaining about the userspace regression in huge=always,
> > and thought it had been changed to behave like huge=within_size,
> > but apparently huge=within_size has itself regressed too.
>
> I'm preparing a RFC patch to discuss this.
>
> > Please explain why the below is not a better patch for i915 and v3d
> > (but still a temporary workaround, because the root of the within_size
> > regression must lie deeper, in the handling of write_end versus i_size).
>
> OK. This looks good to me. Patryk, could you try Hugh's simple patch?
> Thanks.
>
> > ---
> >   mm/shmem.c | 4 ++--
> >   1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/shmem.c b/mm/shmem.c
> > index 3a5a65b1f41a..c67dfc17a819 100644
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -5928,8 +5928,8 @@ struct folio *shmem_read_folio_gfp(struct address_space *mapping,
> >       struct folio *folio;
> >       int error;
> >
> > -     error = shmem_get_folio_gfp(inode, index, 0, &folio, SGP_CACHE,
> > -                                 gfp, NULL, NULL);
> > +     error = shmem_get_folio_gfp(inode, index, i_size_read(inode),
> > +                                 &folio, SGP_CACHE, gfp, NULL, NULL);
> >       if (error)
> >               return ERR_PTR(error);
> >
>