[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 <syos...@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

[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 <j...@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

[Marek Szyprowski]

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