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 not enough 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.
v2: - Make sure there's at least 8 bits available and that the PAGE_OFFSET bit calculation doesn't wrap
Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/ Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool") Cc: stable@vger.kernel.org # 6.15+ Tested-by: Helge Deller deller@gmx.de Signed-off-by: Toke Høiland-Jørgensen toke@redhat.com --- include/linux/mm.h | 22 +++++++------ net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++-------------- 2 files changed, 66 insertions(+), 32 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h index 1ae97a0b8ec7..0905eb6b55ec 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 at least 8 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 +4174,13 @@ 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) +/* Use the lowest bit of PAGE_OFFSET if there's at least 8 bits available; see above */ +#define PP_DMA_INDEX_MIN_OFFSET (1 << (PP_DMA_INDEX_SHIFT + 8)) +#define PP_DMA_INDEX_BITS ((__builtin_constant_p(PAGE_OFFSET) && \ + PAGE_OFFSET >= PP_DMA_INDEX_MIN_OFFSET && \ + !(PAGE_OFFSET & (PP_DMA_INDEX_MIN_OFFSET - 1))) ? \ + MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0) + #endif
#define PP_DMA_INDEX_MASK GENMASK(PP_DMA_INDEX_BITS + PP_DMA_INDEX_SHIFT - 1, \ diff --git a/net/core/page_pool.c b/net/core/page_pool.c index 492728f9e021..1a5edec485f1 100644 --- a/net/core/page_pool.c +++ b/net/core/page_pool.c @@ -468,11 +468,60 @@ page_pool_dma_sync_for_device(const struct page_pool *pool, } }
+static int page_pool_register_dma_index(struct page_pool *pool, + netmem_ref netmem, gfp_t gfp) +{ + int err = 0; + u32 id; + + if (unlikely(!PP_DMA_INDEX_BITS)) + goto out; + + if (in_softirq()) + err = xa_alloc(&pool->dma_mapped, &id, netmem_to_page(netmem), + PP_DMA_INDEX_LIMIT, gfp); + else + err = xa_alloc_bh(&pool->dma_mapped, &id, netmem_to_page(netmem), + PP_DMA_INDEX_LIMIT, gfp); + if (err) { + WARN_ONCE(err != -ENOMEM, "couldn't track DMA mapping, please report to netdev@"); + goto out; + } + + netmem_set_dma_index(netmem, id); +out: + return err; +} + +static int page_pool_release_dma_index(struct page_pool *pool, + netmem_ref netmem) +{ + struct page *old, *page = netmem_to_page(netmem); + unsigned long id; + + if (unlikely(!PP_DMA_INDEX_BITS)) + return 0; + + id = netmem_get_dma_index(netmem); + if (!id) + return -1; + + if (in_softirq()) + old = xa_cmpxchg(&pool->dma_mapped, id, page, NULL, 0); + else + old = xa_cmpxchg_bh(&pool->dma_mapped, id, page, NULL, 0); + if (old != page) + return -1; + + netmem_set_dma_index(netmem, 0); + + return 0; +} + static bool page_pool_dma_map(struct page_pool *pool, netmem_ref netmem, gfp_t gfp) { dma_addr_t dma; int err; - u32 id;
/* Setup DMA mapping: use 'struct page' area for storing DMA-addr * since dma_addr_t can be either 32 or 64 bits and does not always fit @@ -491,18 +540,10 @@ static bool page_pool_dma_map(struct page_pool *pool, netmem_ref netmem, gfp_t g goto unmap_failed; }
- if (in_softirq()) - err = xa_alloc(&pool->dma_mapped, &id, netmem_to_page(netmem), - PP_DMA_INDEX_LIMIT, gfp); - else - err = xa_alloc_bh(&pool->dma_mapped, &id, netmem_to_page(netmem), - PP_DMA_INDEX_LIMIT, gfp); - if (err) { - WARN_ONCE(err != -ENOMEM, "couldn't track DMA mapping, please report to netdev@"); + err = page_pool_register_dma_index(pool, netmem, gfp); + if (err) goto unset_failed; - }
- netmem_set_dma_index(netmem, id); page_pool_dma_sync_for_device(pool, netmem, pool->p.max_len);
return true; @@ -680,8 +721,6 @@ void page_pool_clear_pp_info(netmem_ref netmem) static __always_inline void __page_pool_release_netmem_dma(struct page_pool *pool, netmem_ref netmem) { - struct page *old, *page = netmem_to_page(netmem); - unsigned long id; dma_addr_t dma;
if (!pool->dma_map) @@ -690,15 +729,7 @@ static __always_inline void __page_pool_release_netmem_dma(struct page_pool *poo */ return;
- id = netmem_get_dma_index(netmem); - if (!id) - return; - - if (in_softirq()) - old = xa_cmpxchg(&pool->dma_mapped, id, page, NULL, 0); - else - old = xa_cmpxchg_bh(&pool->dma_mapped, id, page, NULL, 0); - if (old != page) + if (page_pool_release_dma_index(pool, netmem)) return;
dma = page_pool_get_dma_addr_netmem(netmem); @@ -708,7 +739,6 @@ static __always_inline void __page_pool_release_netmem_dma(struct page_pool *poo PAGE_SIZE << pool->p.order, pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING); page_pool_set_dma_addr_netmem(netmem, 0); - netmem_set_dma_index(netmem, 0); }
/* Disconnects a page (from a page_pool). API users can have a need
On Tue, Sep 30, 2025 at 4:43 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 not enough 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.
v2:
- Make sure there's at least 8 bits available and that the PAGE_OFFSET bit calculation doesn't wrap
Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/ Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool") Cc: stable@vger.kernel.org # 6.15+ Tested-by: Helge Deller deller@gmx.de Signed-off-by: Toke Høiland-Jørgensen toke@redhat.com
include/linux/mm.h | 22 +++++++------ net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++-------------- 2 files changed, 66 insertions(+), 32 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h index 1ae97a0b8ec7..0905eb6b55ec 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 at least 8 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 +4174,13 @@ 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) +/* Use the lowest bit of PAGE_OFFSET if there's at least 8 bits available; see above */ +#define PP_DMA_INDEX_MIN_OFFSET (1 << (PP_DMA_INDEX_SHIFT + 8)) +#define PP_DMA_INDEX_BITS ((__builtin_constant_p(PAGE_OFFSET) && \
PAGE_OFFSET >= PP_DMA_INDEX_MIN_OFFSET && \!(PAGE_OFFSET & (PP_DMA_INDEX_MIN_OFFSET - 1))) ? \MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0)#endif
It took some staring at, but I think I understand this code and it is correct. This is the critical check, it's making sure that the bits used by PAGE_OFFSET are not shared with the bits used for the dma-index:
!(PAGE_OFFSET & (PP_DMA_INDEX_MIN_OFFSET - 1))) ? \
The following check confused me for a while, but I think I figured it out. It's checking that the bits used for PAGE_OFFSET are 'higher' than the bits used for PP_DMA_INDEX:
PAGE_OFFSET >= PP_DMA_INDEX_MIN_OFFSET && \
And finally this calculation should indeed be the bits we can use (the empty space between the lsb set by PAGE_OFFSET and the msb set by the pp magic:
MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0)
AFAIU we should not need the MIN anymore, since that subtraction is guaranteed to be positive, but that's a nit.
Reviewed-by: Mina Almasry almasrymina@google.com
Mina Almasry almasrymina@google.com writes:
On Tue, Sep 30, 2025 at 4:43 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 not enough 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.
v2:
- Make sure there's at least 8 bits available and that the PAGE_OFFSET bit calculation doesn't wrap
Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/ Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool") Cc: stable@vger.kernel.org # 6.15+ Tested-by: Helge Deller deller@gmx.de Signed-off-by: Toke Høiland-Jørgensen toke@redhat.com
include/linux/mm.h | 22 +++++++------ net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++-------------- 2 files changed, 66 insertions(+), 32 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h index 1ae97a0b8ec7..0905eb6b55ec 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 at least 8 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 +4174,13 @@ 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) +/* Use the lowest bit of PAGE_OFFSET if there's at least 8 bits available; see above */ +#define PP_DMA_INDEX_MIN_OFFSET (1 << (PP_DMA_INDEX_SHIFT + 8)) +#define PP_DMA_INDEX_BITS ((__builtin_constant_p(PAGE_OFFSET) && \
PAGE_OFFSET >= PP_DMA_INDEX_MIN_OFFSET && \!(PAGE_OFFSET & (PP_DMA_INDEX_MIN_OFFSET - 1))) ? \MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0)#endif
It took some staring at, but I think I understand this code and it is correct. This is the critical check, it's making sure that the bits used by PAGE_OFFSET are not shared with the bits used for the dma-index:
!(PAGE_OFFSET & (PP_DMA_INDEX_MIN_OFFSET - 1))) ? \The following check confused me for a while, but I think I figured it out. It's checking that the bits used for PAGE_OFFSET are 'higher' than the bits used for PP_DMA_INDEX:
PAGE_OFFSET >= PP_DMA_INDEX_MIN_OFFSET && \And finally this calculation should indeed be the bits we can use (the empty space between the lsb set by PAGE_OFFSET and the msb set by the pp magic:
MIN(32, __ffs(PAGE_OFFSET) - PP_DMA_INDEX_SHIFT) : 0)
Yup, exactly! Thanks for walking through it and confirming that the logic is sound :)
AFAIU we should not need the MIN anymore, since that subtraction is guaranteed to be positive, but that's a nit.
The MIN was originally there to limit how many bits we use for 64-bit systems that don't set POISON_POINTER_DELTA, since xarray uses a u32 for the size of the limits. Not sure if such a combination exists in the real world, but I figure that having it there doesn't hurt in any case.
Reviewed-by: Mina Almasry almasrymina@google.com
Thanks!
-Toke
On 9/30/25 13:43, Toke Høiland-Jørgensen 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 not enough 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.
v2:
- Make sure there's at least 8 bits available and that the PAGE_OFFSET bit calculation doesn't wrap
Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/ Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool") Cc: stable@vger.kernel.org # 6.15+ Tested-by: Helge Deller deller@gmx.de Signed-off-by: Toke Høiland-Jørgensen toke@redhat.com
include/linux/mm.h | 22 +++++++------ net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++-------------- 2 files changed, 66 insertions(+), 32 deletions(-)
I tested this v2 patch (the former tested-by was for v1), and v2 works too:
Tested-by: Helge Deller deller@gmx.de
Thanks! Helge
Helge Deller deller@gmx.de writes:
On 9/30/25 13:43, Toke Høiland-Jørgensen 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 not enough 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.
v2:
- Make sure there's at least 8 bits available and that the PAGE_OFFSET bit calculation doesn't wrap
Link: https://lore.kernel.org/all/aMNJMFa5fDalFmtn@p100/ Fixes: ee62ce7a1d90 ("page_pool: Track DMA-mapped pages and unmap them when destroying the pool") Cc: stable@vger.kernel.org # 6.15+ Tested-by: Helge Deller deller@gmx.de Signed-off-by: Toke Høiland-Jørgensen toke@redhat.com
include/linux/mm.h | 22 +++++++------ net/core/page_pool.c | 76 ++++++++++++++++++++++++++++++-------------- 2 files changed, 66 insertions(+), 32 deletions(-)
I tested this v2 patch (the former tested-by was for v1), and v2 works too:
Tested-by: Helge Deller deller@gmx.de
Great, thank you for re-testing! :)
-Toke
Hello:
This patch was applied to netdev/net.git (main) by Jakub Kicinski kuba@kernel.org:
On Tue, 30 Sep 2025 13:43:29 +0200 you 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.
[...]
Here is the summary with links: - [net,v2] page_pool: Fix PP_MAGIC_MASK to avoid crashing on some 32-bit arches https://git.kernel.org/netdev/net/c/95920c2ed02b
You are awesome, thank you!
linux-stable-mirror@lists.linaro.org
-
Helge Deller -
Mina Almasry -
patchwork-bot+netdevbpf＠kernel.org -
Toke Høiland-Jørgensen