Re: [PATCH net] page_pool: Fix PP_MAGIC_MASK to avoid crashing on some 32-bit arches

[Mina Almasry <almasrymina@google.com>]

On Fri, Sep 26, 2025 at 4:40 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:
>
> Helge reported that the introduction of PP_MAGIC_MASK let to crashes on
> boot on his 32-bit parisc machine. The cause of this is the mask is set
> too wide, so the page_pool_page_is_pp() incurs false positives which
> crashes the machine.
>
> Just disabling the check in page_pool_is_pp() will lead to the page_pool
> code itself malfunctioning; so instead of doing this, this patch changes
> the define for PP_DMA_INDEX_BITS to avoid mistaking arbitrary kernel
> pointers for page_pool-tagged pages.
>
> The fix relies on the kernel pointers that alias with the pp_magic field
> always being above PAGE_OFFSET. With this assumption, we can use the
> lowest bit of the value of PAGE_OFFSET as the upper bound of the
> PP_DMA_INDEX_MASK, which should avoid the false positives.
>
> Because we cannot rely on PAGE_OFFSET always being a compile-time
> constant, nor on it always being >0, we fall back to disabling the
> dma_index storage when there are no bits available. This leaves us in
> the situation we were in before the patch in the Fixes tag, but only on
> a subset of architecture configurations. This seems to be the best we
> can do until the transition to page types in complete for page_pool
> pages.
>
> Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/
> Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool")
> Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
> ---
> Sorry for the delay on getting this out. I have only compile-tested it,
> since I don't have any hardware that triggers the original bug. Helge, I'm
> hoping you can take it for a spin?
>
>  include/linux/mm.h   | 18 +++++------
>  net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++--------------
>  2 files changed, 62 insertions(+), 32 deletions(-)
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 1ae97a0b8ec7..28541cb40f69 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -4159,14 +4159,13 @@ int arch_lock_shadow_stack_status(struct task_struct *t, unsigned long status);
>   * since this value becomes part of PP_SIGNATURE; meaning we can just use the
>   * space between the PP_SIGNATURE value (without POISON_POINTER_DELTA), and the
>   * lowest bits of POISON_POINTER_DELTA. On arches where POISON_POINTER_DELTA is
> - * 0, we make sure that we leave the two topmost bits empty, as that guarantees
> - * we won't mistake a valid kernel pointer for a value we set, regardless of the
> - * VMSPLIT setting.
> + * 0, we use the lowest bit of PAGE_OFFSET as the boundary if that value is
> + * known at compile-time.
>   *
> - * Altogether, this means that the number of bits available is constrained by
> - * the size of an unsigned long (at the upper end, subtracting two bits per the
> - * above), and the definition of PP_SIGNATURE (with or without
> - * POISON_POINTER_DELTA).
> + * If the value of PAGE_OFFSET is not known at compile time, or if it is too
> + * small to leave some bits available above PP_SIGNATURE, we define the number
> + * of bits to be 0, which turns off the DMA index tracking altogether (see
> + * page_pool_register_dma_index()).
>   */
>  #define PP_DMA_INDEX_SHIFT (1 + __fls(PP_SIGNATURE - POISON_POINTER_DELTA))
>  #if POISON_POINTER_DELTA > 0
> @@ -4175,8 +4174,9 @@ int arch_lock_shadow_stack_status(struct task_struct *t, unsigned long status);
>   */
>  #define PP_DMA_INDEX_BITS MIN(32, __ffs(POISON_POINTER_DELTA) - PP_DMA_INDEX_SHIFT)
>  #else
> -/* Always leave out the topmost two; see above. */
> -#define PP_DMA_INDEX_BITS MIN(32, BITS_PER_LONG - PP_DMA_INDEX_SHIFT - 2)
> +/* Constrain to the lowest bit of PAGE_OFFSET if known; see above. */
> +#define PP_DMA_INDEX_BITS ((__builtin_constant_p(PAGE_OFFSET) && PAGE_OFFSET > PP_SIGNATURE) ? \
> +                             MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0)

Do you have to watch out for an underflow of __ffs(PAGE_OFFSET) -
PP_DMA_INDEX_SHIFT (at which point we'll presumably use 32 here
instead of the expected 0)? Or is that guaranteed to be positive for
some reason I'm not immediately grasping.

-- 
Thanks,
Mina