[PATCH] kmsan: fix kmsan_handle_dma() to avoid false positives

[PATCH] kmsan: fix kmsan_handle_dma() to avoid false positives

[Shigeru Yoshida]

KMSAN reports an uninitialized value issue in dma_map_phys()[1].  This
is a false positive caused by the way the virtual address is handled
in kmsan_handle_dma().  Fix it by translating the physical address to
a virtual address using phys_to_virt().

[1]
BUG: KMSAN: uninit-value in dma_map_phys+0xdc5/0x1060
 dma_map_phys+0xdc5/0x1060
 dma_map_page_attrs+0xcf/0x130
 e1000_xmit_frame+0x3c51/0x78f0
 dev_hard_start_xmit+0x22f/0xa30
 sch_direct_xmit+0x3b2/0xcf0
 __dev_queue_xmit+0x3588/0x5e60
 neigh_resolve_output+0x9c5/0xaf0
 ip6_finish_output2+0x24e0/0x2d30
 ip6_finish_output+0x903/0x10d0
 ip6_output+0x331/0x600
 mld_sendpack+0xb4a/0x1770
 mld_ifc_work+0x1328/0x19b0
 process_scheduled_works+0xb91/0x1d80
 worker_thread+0xedf/0x1590
 kthread+0xd5c/0xf00
 ret_from_fork+0x1f5/0x4c0
 ret_from_fork_asm+0x1a/0x30

Uninit was created at:
 __kmalloc_cache_noprof+0x8f5/0x16b0
 syslog_print+0x9a/0xef0
 do_syslog+0x849/0xfe0
 __x64_sys_syslog+0x97/0x100
 x64_sys_call+0x3cf8/0x3e30
 do_syscall_64+0xd9/0xfa0
 entry_SYSCALL_64_after_hwframe+0x77/0x7f

Bytes 0-89 of 90 are uninitialized
Memory access of size 90 starts at ffff8880367ed000

CPU: 1 UID: 0 PID: 1552 Comm: kworker/1:2 Not tainted 6.17.0-next-20250929 #26 PREEMPT(none)
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-5.fc42 04/01/2014
Workqueue: mld mld_ifc_work

Fixes: 6eb1e769b2c1 ("kmsan: convert kmsan_handle_dma to use physical addresses")
Signed-off-by: Shigeru Yoshida <syoshida@redhat.com>
---
The hash in the "Fixes" tag comes from the linux-next tree
(next-20250929), as it has not yet been included in the mainline tree.
---
 mm/kmsan/hooks.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/mm/kmsan/hooks.c b/mm/kmsan/hooks.c
index 90bee565b9bc..2cee59d89c80 100644
--- a/mm/kmsan/hooks.c
+++ b/mm/kmsan/hooks.c
@@ -339,13 +339,12 @@ static void kmsan_handle_dma_page(const void *addr, size_t size,
 void kmsan_handle_dma(phys_addr_t phys, size_t size,
 		      enum dma_data_direction dir)
 {
-	struct page *page = phys_to_page(phys);
 	u64 page_offset, to_go;
 	void *addr;
 
 	if (PhysHighMem(phys))
 		return;
-	addr = page_to_virt(page);
+	addr = phys_to_virt(phys);
 	/*
 	 * The kernel may occasionally give us adjacent DMA pages not belonging
 	 * to the same allocation. Process them separately to avoid triggering
-- 
2.51.0

Re: [PATCH] kmsan: fix kmsan_handle_dma() to avoid false positives

[Jason Gunthorpe]

On Thu, Oct 02, 2025 at 02:10:24PM +0900, Shigeru Yoshida wrote:
> KMSAN reports an uninitialized value issue in dma_map_phys()[1].  This
> is a false positive caused by the way the virtual address is handled
> in kmsan_handle_dma().  Fix it by translating the physical address to
> a virtual address using phys_to_virt().

This is the same sort of thinko as was found on the alpha patch, it is
tricky!

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

> @@ -339,13 +339,12 @@ static void kmsan_handle_dma_page(const void *addr, size_t size,
>  void kmsan_handle_dma(phys_addr_t phys, size_t size,
>  		      enum dma_data_direction dir)
>  {
> -	struct page *page = phys_to_page(phys);

This throws away the page_offset encoded in phys

>  	u64 page_offset, to_go;
>  	void *addr;
>  
>  	if (PhysHighMem(phys))
>  		return;
> -	addr = page_to_virt(page);

And this gives an addr that is now 0 page_offset, which is not right.

> +	addr = phys_to_virt(phys);

Make more sense anyhow when combined with PhysHighMem() and gives the
right page_offset.

Jason

Re: [PATCH] kmsan: fix kmsan_handle_dma() to avoid false positives

[Marek Szyprowski]

On 02.10.2025 07:10, Shigeru Yoshida wrote:
> KMSAN reports an uninitialized value issue in dma_map_phys()[1].  This
> is a false positive caused by the way the virtual address is handled
> in kmsan_handle_dma().  Fix it by translating the physical address to
> a virtual address using phys_to_virt().
>
> [1]
> BUG: KMSAN: uninit-value in dma_map_phys+0xdc5/0x1060
>   dma_map_phys+0xdc5/0x1060
>   dma_map_page_attrs+0xcf/0x130
>   e1000_xmit_frame+0x3c51/0x78f0
>   dev_hard_start_xmit+0x22f/0xa30
>   sch_direct_xmit+0x3b2/0xcf0
>   __dev_queue_xmit+0x3588/0x5e60
>   neigh_resolve_output+0x9c5/0xaf0
>   ip6_finish_output2+0x24e0/0x2d30
>   ip6_finish_output+0x903/0x10d0
>   ip6_output+0x331/0x600
>   mld_sendpack+0xb4a/0x1770
>   mld_ifc_work+0x1328/0x19b0
>   process_scheduled_works+0xb91/0x1d80
>   worker_thread+0xedf/0x1590
>   kthread+0xd5c/0xf00
>   ret_from_fork+0x1f5/0x4c0
>   ret_from_fork_asm+0x1a/0x30
>
> Uninit was created at:
>   __kmalloc_cache_noprof+0x8f5/0x16b0
>   syslog_print+0x9a/0xef0
>   do_syslog+0x849/0xfe0
>   __x64_sys_syslog+0x97/0x100
>   x64_sys_call+0x3cf8/0x3e30
>   do_syscall_64+0xd9/0xfa0
>   entry_SYSCALL_64_after_hwframe+0x77/0x7f
>
> Bytes 0-89 of 90 are uninitialized
> Memory access of size 90 starts at ffff8880367ed000
>
> CPU: 1 UID: 0 PID: 1552 Comm: kworker/1:2 Not tainted 6.17.0-next-20250929 #26 PREEMPT(none)
> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-5.fc42 04/01/2014
> Workqueue: mld mld_ifc_work
>
> Fixes: 6eb1e769b2c1 ("kmsan: convert kmsan_handle_dma to use physical addresses")
> Signed-off-by: Shigeru Yoshida <syoshida@redhat.com>

Applied to dma-mapping-for-next (for v6.18-rc1) branch. Thanks!

> ---
> The hash in the "Fixes" tag comes from the linux-next tree
> (next-20250929), as it has not yet been included in the mainline tree.
> ---
>   mm/kmsan/hooks.c | 3 +--
>   1 file changed, 1 insertion(+), 2 deletions(-)
>
> diff --git a/mm/kmsan/hooks.c b/mm/kmsan/hooks.c
> index 90bee565b9bc..2cee59d89c80 100644
> --- a/mm/kmsan/hooks.c
> +++ b/mm/kmsan/hooks.c
> @@ -339,13 +339,12 @@ static void kmsan_handle_dma_page(const void *addr, size_t size,
>   void kmsan_handle_dma(phys_addr_t phys, size_t size,
>   		      enum dma_data_direction dir)
>   {
> -	struct page *page = phys_to_page(phys);
>   	u64 page_offset, to_go;
>   	void *addr;
>   
>   	if (PhysHighMem(phys))
>   		return;
> -	addr = page_to_virt(page);
> +	addr = phys_to_virt(phys);
>   	/*
>   	 * The kernel may occasionally give us adjacent DMA pages not belonging
>   	 * to the same allocation. Process them separately to avoid triggering

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland