linux-kernel - Re: [PATCH] drm/i915: Fix page cleanup on DMA remap failure

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CADyq12zbwJAGdZO1=SQvAGE9+cEeeWEAE=DXzDz_a_8n9mUrPw@mail.gmail.com>
Date: Thu, 16 Jan 2025 09:48:55 -0500
From: Brian Geffon <bgeffon@...gle.com>
To: Ville Syrjälä <ville.syrjala@...ux.intel.com>
Cc: intel-gfx@...ts.freedesktop.org, chris.p.wilson@...el.com, 
	jani.saarinen@...el.com, tomasz.mistat@...el.com, vidya.srinivas@...el.com, 
	jani.nikula@...ux.intel.com, linux-kernel@...r.kernel.org, 
	dri-devel@...ts.freedesktop.org, 
	Joonas Lahtinen <joonas.lahtinen@...ux.intel.com>, Tomasz Figa <tfiga@...gle.com>
Subject: Re: [PATCH] drm/i915: Fix page cleanup on DMA remap failure

On Thu, Jan 16, 2025 at 9:38 AM Ville Syrjälä
<ville.syrjala@...ux.intel.com> wrote:
>
> On Thu, Jan 16, 2025 at 04:24:26PM +0200, Ville Syrjälä wrote:
> > On Thu, Jan 16, 2025 at 08:56:36AM -0500, Brian Geffon wrote:
> > > When converting to folios the cleanup path of shmem_get_pages() was
> > > missed. When a DMA remap fails and the max segment size is greater than
> > > PAGE_SIZE it will attempt to retry the remap with a PAGE_SIZEd segment
> > > size. The cleanup code isn't properly using the folio apis and as a
> > > result isn't handling compound pages correctly.
> > >
> > > Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/13487
> > > Fixes: 0b62af28f249 ("i915: convert shmem_sg_free_table() to use a folio_batch")
> > > Signed-off-by: Brian Geffon <bgeffon@...gle.com>
> > > Suggested-by: Tomasz Figa <tfiga@...gle.com>
> > > ---
> > >  drivers/gpu/drm/i915/gem/i915_gem_shmem.c | 13 +++++--------
> > >  1 file changed, 5 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > index fe69f2c8527d..02ddab5bf5c0 100644
> > > --- a/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shmem.c
> > > @@ -37,8 +37,6 @@ void shmem_sg_free_table(struct sg_table *st, struct address_space *mapping,
> > >     struct folio *last = NULL;
> > >     struct page *page;
> > >
> > > -   mapping_clear_unevictable(mapping);
> > > -
> >
> > The assymmetry here between the alloc vs. free is a bit annoying.
> > Maybe we can just keep this here?
>
> Actually, I guess it's a bit more than just annoying since
> shmem_sg_free_table() is actually used from outside i915_gem_shmem.c
> as well.

You're correct, this was a bad oversight on my part. This patch is
going to require a v2 regardless, how do you feel about just changing
shmem_sg_free_table() to not accept a mapping given that managing the
mapping is really orthogonal to cleaning up the sg table.

>
> >
> > Or if avoiding the ping-pong actually mattes in the gtt prepare
> > error case, then maybe we should rename this guy into
> > __shmem_sg_free_table() without the mapping_clear_unevictable()
> > and wrap it in a higher level shmem_sg_free_table() that does
> > everything?
> >
> > >     folio_batch_init(&fbatch);
> > >     for_each_sgt_page(page, sgt_iter, st) {
> > >             struct folio *folio = page_folio(page);
> > > @@ -180,10 +178,10 @@ int shmem_sg_alloc_table(struct drm_i915_private *i915, struct sg_table *st,
> > >     return 0;
> > >  err_sg:
> > >     sg_mark_end(sg);
> > > +   mapping_clear_unevictable(mapping);
> > >     if (sg != st->sgl) {
> > >             shmem_sg_free_table(st, mapping, false, false);
> > >     } else {
> > > -           mapping_clear_unevictable(mapping);
> > >             sg_free_table(st);
> > >     }
> > >
> > > @@ -209,8 +207,6 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
> > >     struct address_space *mapping = obj->base.filp->f_mapping;
> > >     unsigned int max_segment = i915_sg_segment_size(i915->drm.dev);
> > >     struct sg_table *st;
> > > -   struct sgt_iter sgt_iter;
> > > -   struct page *page;
> > >     int ret;
> > >
> > >     /*
> > > @@ -239,9 +235,8 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
> > >              * for PAGE_SIZE chunks instead may be helpful.
> > >              */
> > >             if (max_segment > PAGE_SIZE) {
> > > -                   for_each_sgt_page(page, sgt_iter, st)
> > > -                           put_page(page);
> > > -                   sg_free_table(st);
> > > +                   /* Leave the mapping unevictable while we retry */
> > > +                   shmem_sg_free_table(st, mapping, false, false);
> > >                     kfree(st);
> > >
> > >                     max_segment = PAGE_SIZE;
> > > @@ -265,6 +260,7 @@ static int shmem_get_pages(struct drm_i915_gem_object *obj)
> > >     return 0;
> > >
> > >  err_pages:
> > > +   mapping_clear_unevictable(mapping);
> > >     shmem_sg_free_table(st, mapping, false, false);
> > >     /*
> > >      * shmemfs first checks if there is enough memory to allocate the page
> > > @@ -402,6 +398,7 @@ void i915_gem_object_put_pages_shmem(struct drm_i915_gem_object *obj, struct sg_
> > >     if (i915_gem_object_needs_bit17_swizzle(obj))
> > >             i915_gem_object_save_bit_17_swizzle(obj, pages);
> > >
> > > +   mapping_clear_unevictable(file_inode(obj->base.filp)->i_mapping);
> > >     shmem_sg_free_table(pages, file_inode(obj->base.filp)->i_mapping,
> > >                         obj->mm.dirty, obj->mm.madv == I915_MADV_WILLNEED);
> > >     kfree(pages);
> > > --
> > > 2.48.0.rc2.279.g1de40edade-goog
> >
> > --
> > Ville Syrjälä
> > Intel
>
> --
> Ville Syrjälä
> Intel