Re: [PATCH v1 0/7] mm: kasan: skip KASAN for GFP_HIGHUSER

From: Tomasz Figa

Date: Tue Nov 17 2026 - 09:19:59 EST


On Tue, Nov 17, 2026 at 7:10 PM David Hildenbrand (Arm)
<david@xxxxxxxxxx> wrote:
>
> On 11/17/26 00:06, Jeff Xu wrote:
> > Hi all,
> >
> > Currently, GFP_HIGHUSER_MOVABLE includes __GFP_SKIP_KASAN (commits
> > c275c5c6d50a, 70c248aca9e7, and 0a54864f8dfb), allowing movable
> > userspace page allocations to skip KASAN page unpoisoning on allocation
> > and poisoning on free under CONFIG_KASAN_HW_TAGS (ARM64 MTE). However,
> > GFP_HIGHUSER does not include __GFP_SKIP_KASAN, even though it shares
> > the same kernel-accessibility contract in include/linux/gfp_types.h
> > ("for userspace allocations that may be mapped to userspace, do not need
> > to be directly accessible by the kernel but that cannot move once in
> > use") and differs only by !__GFP_MOVABLE.
> >
> > Moving __GFP_SKIP_KASAN to GFP_HIGHUSER addresses the performance issue
> > below (following up on the earlier drm/panthor discussion [1, 2]). The
> > swap issue from that discussion is not addressed by this series:
> >
> > 1. Performance: For every 4 KB page without __GFP_SKIP_KASAN,
> > mte_set_mem_tag_range() tags all 256 16-byte MTE granules (via
> > STG/STZG and DC GVA/GZVA) on allocation and again on free. In
> > graphics and UI workloads, DMA-BUF (system_heap) and DRM/GPU
> > (shmem/TTM) buffers use GFP_HIGHUSER; while steady-state rendering
> > reuses buffers, window resizes, surface transitions, and app launches
> > synchronously allocate and free tens to hundreds of megabytes (~32 MB
> > per 4K RGBA frame) on UI/compositor/render threads. That takes a
> > significant share of the 8.3 ms (120 Hz) / 16.7 ms (60 Hz) frame
> > budget and thrashes CPU data/MTE tag caches (especially on
> > cold-cache free). On Armv9.2-A, skipping KASAN cuts ~32 MB DMA-BUF
> > alloc+free latency by ~80% (~98% faster close()) and improves the
> > OpenGL score of an app-loading benchmark by 15-23%.
> >
> > 2. False-positive KASAN report after shmem swap-in readahead [1] (not
> > addressed here): when a shmem page is swapped in,
> > swap_cluster_readahead() also reads the neighboring swap slots and
> > allocates their swap-cache folios with the gfp mask of that shmem
> > mapping. A neighboring slot can hold an unrelated page that
> > userspace maps with PROT_MTE. If the mask lacks __GFP_SKIP_KASAN,
> > that page's folio gets a KASAN tag in page->flags. When its owner
> > faults it in, arch_swap_restore() writes the saved userspace tags to
> > the memory, but page->flags keeps the KASAN tag. Later kernel
> > accesses through page_address() then use a tag that does not match
> > the memory and are reported by KASAN.
> >
> > This series does not fix the readahead path. It only removes one
> > way to hit it: shmem mappings that use GFP_HIGHUSER (e.g.
> > drm_gem_shmem_helper / panthor) now pass __GFP_SKIP_KASAN to
> > readahead, so the neighboring folios no longer get a KASAN tag.
> > Shmem mappings with other gfp masks that lack __GFP_SKIP_KASAN
> > (e.g. GFP_KERNEL) can still trigger the report. That issue should
> > be followed up separately, for example in [1].
> >
> > What changes for users of GFP_HIGHUSER pages under CONFIG_KASAN_HW_TAGS
> > (other KASAN modes are not affected):
> >
> > - Access through vmap(): nothing changes. Those pointers already carry
> > the match-all tag and are not checked today.
> > - Access through kmap_local_page() or page_address() (for example
> > ramfs): accesses are no longer checked by KASAN, as is already the
> > case for GFP_HIGHUSER_MOVABLE pages.
> > - Buffers that the kernel itself fills (pipe, FUSE, kexec, RDS) keep
> > KASAN checking by using the new GFP_HIGHUSER_KASAN (patches 1-4).
> >
> > GFP_KERNEL, GFP_USER, slab and vmalloc() allocations are not affected.
> >
> > Patch 1 adds GFP_HIGHUSER_KASAN
>
> I really don't want us to spread KASAN-related GFP flags in the kernel.
>
> There must be a proper way to abstract this.
>
> Why is GFP_HIGHUSER not sufficient? We were already told that this is memory to
> be mapped into user space. Placing data structure on it and handing out pointers
> to that would already be a bit questionable.
>
> So what are the ways to avoid GFP_HIGHUSER_KASAN?

I think the idea behind GFP_HIGHUSER_KASAN is to make it possible to
skip KASAN for GFP_HIGHUSER by default (for most of its users), but
keep it for some of the users which today use GFP_HIGHUSER for their
memory allocations, but could benefit from the added security. As per
the other patches in the series:

pipe, fuse: use GFP_HIGHUSER_KASAN for pipe buffer pages
kexec: use GFP_HIGHUSER_KASAN for segment pages
net/rds: use GFP_HIGHUSER_KASAN for message pages

That said, I'm not familiar with any of the code mentioned by those
patches, so I can't really tell how much they would benefit from KASAN
or if their choice of GFP_HIGHUSER to begin with is correct.

Best,
Tomasz