Re: [PATCH] drm/panthor: set __GFP_SKIP_KASAN

[Chia-I Wu <olvaffe@gmail.com>]

+KASAN and shmem maintainers

On Wed, May 13, 2026 at 2:11 AM Boris Brezillon
<boris.brezillon@collabora.com> wrote:
>
> On Tue, 12 May 2026 10:36:28 -0700
> Chia-I Wu via B4 Relay <devnull+olvaffe.gmail.com@kernel.org> wrote:
>
> > From: Chia-I Wu <olvaffe@gmail.com>
> >
> > Pages that can be swapped out should be allocated with __GFP_SKIP_KASAN.
> > Rather than setting the flag directly, replace GFP_HIGHUSER by
> > (GFP_HIGHUSER_MOVABLE & ~__GFP_MOVABLE) instead, which should match the
> > preceding comment better.
>
> I'm not too sure GFP_HIGHUSER_MOVABLE & ~__GFP_MOVABLE is clearer than
> GFP_HIGHUSER (without MOVABLE in the name) to reflect we don't want the
> MOVABLE property.
>
> >
> > On a CONFIG_KASAN_HW_TAGS=y system, without __GFP_SKIP_KASAN, the page
> > allocator assigns a valid tag to both the kernel mapping and MTE,
> > instead of assigning the match-all KASAN_TAG_KERNEL tag to the kernel
> > mapping. If userspace also maps the page with PROT_MTE and modifies the
> > MTE tag, accessing the page via the kernel mapping results in KASAN
> > invalid-access, such as
> >
> >   BUG: KASAN: invalid-access in swap_writepage+0xb0/0x21c
> >   Read at addr f5ffff81aa71dff8 by task WM.task-4/6956
> >   Pointer tag: [f5], memory tag: [f9]
> >
> > While userspace cannot map drm gem objects with PROT_MTE, the problem is
> > shmem_swapin_cluster. When it swaps in a cluster of pages using our gfp
> > flags, some of the pages might belong to other mappings that have
> > PROT_MTE.
>
> Okay, let me see if I get this right. The gfp flags we set right now
> do what we want for out GEM mappings, it's only when swap readahead is
> involved that this becomes problematic:
>
> 1. swapin folio for GEM inode X
> 2. shmem layer optimistically does a readahead on the global swap, with
>    our GEM gfp flags
> 3. the next swap entry belongs to some other tmpfs/anonymous mapping
>    (not even something on the same GEM mountpoint actually)
> 4. because it's our gfp flags that get used for the folio allocation,
>   the KASAN poisoning takes place, thus messing up with the original
>   user-request PROT_MTE
Yes, that's right. At least that's what I observed.

A little more on the last point. shmem readahead allocates a folio for
an unrelated mapping using our gfp flags. It unpoisons both the page
tag and mte tag for the folio because our gfp flags do not have
__GFP_SKIP_KASAN.

When the unrelated mapping is an anonymous one, and when do_swap_page
maps the folio, arch_swap_restore restores the mte tag. This creates
mismatching page tag and mte tag. When, for example, swap_writepage
tries to access the folio, kasan triggers a false positive error.

>
> To me, this looks like a bad isolation of the readahead logic: we
> shouldn't cross the inode boundary, because gfp flags for one inode
> might be completely different from gfp flags for a different inode.
> Let alone the fact the boundary being crossed here is not just inode,
> it's basically the entire tmpfs mountpoint.
>
> TLDR; I strongly believe we're papering over some more fundamental
> issue in the shmem swap readahead logic: with MTE, we can no longer
> assume gfp flags are interchangeable, they have to be enforced and
> match exactly what the user of the swap entry expects, meaning
> readahead should be constrained by the inode boundary.
I totally agree. shmem should not readahead past its inode. Or perhaps
it should force __GFP_SKIP_KASAN when it does.

Adding shmem and kasan maintainers for suggestions/corrections.

>
> >
> > Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> > ---
> > The latest snapdragons appear to have MTE support. drm/msm might need
> > the same treatment.
> > ---
> >  drivers/gpu/drm/panthor/panthor_gem.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/gpu/drm/panthor/panthor_gem.c b/drivers/gpu/drm/panthor/panthor_gem.c
> > index 13295d7a593df..08c03aa0db2f7 100644
> > --- a/drivers/gpu/drm/panthor/panthor_gem.c
> > +++ b/drivers/gpu/drm/panthor/panthor_gem.c
> > @@ -1013,7 +1013,8 @@ panthor_gem_create(struct drm_device *dev, size_t size, uint32_t flags,
> >        * going to pin these pages.
> >        */
> >       mapping_set_gfp_mask(bo->base.filp->f_mapping,
> > -                          GFP_HIGHUSER | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
> > +                          (GFP_HIGHUSER_MOVABLE & ~__GFP_MOVABLE) |
> > +                                  __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
> >
> >       ret = drm_gem_create_mmap_offset(&bo->base);
> >       if (ret)
> >
> > ---
> > base-commit: 6101f78b684895d5860a96322e607e0f46f433ad
> > change-id: 20260512-panthor-kasan-10477239bad1
> >
> > Best regards,
> > --
> > Chia-I Wu <olvaffe@gmail.com>
> >
> >
>