[PATCH] vfio/mlx5: Fix an unwind issue in mlx5vf_add_pages()

From: Hungyu Lin

Date: Sun Nov 15 2026 - 22:01:52 EST


mlx5vf_add_pages() gets a pointer to the page list of its caller. If
alloc_pages_bulk() fails after some of the pages were allocated, the
unwind flow frees them with

__free_page(*page_list[i]);

The index is applied to page_list before it is dereferenced, so only
the first page is freed. For the other pages the code reads the memory
that follows the page list pointer of the caller and dereferences it.
In both callers this is a dma_iova_state that is still zeroed, which
ends in a NULL pointer dereference, and the pages are leaked.

The pages are allocated with GFP_KERNEL_ACCOUNT, so the allocation can
fail, for example when the memory cgroup of the task is out of memory.

Dereference page_list before indexing it.

Fixes: 089803c40193 ("vfio/mlx5: Enable the DMA link API")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Hungyu Lin <dennylin0707@xxxxxxxxx>
---
drivers/vfio/pci/mlx5/cmd.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/vfio/pci/mlx5/cmd.c b/drivers/vfio/pci/mlx5/cmd.c
index 04180035a7fe..8e8514f2660f 100644
--- a/drivers/vfio/pci/mlx5/cmd.c
+++ b/drivers/vfio/pci/mlx5/cmd.c
@@ -518,7 +518,7 @@ static int mlx5vf_add_pages(struct page ***page_list, unsigned int npages)

err:
for (i = 0; i < done; i++)
- __free_page(*page_list[i]);
+ __free_page((*page_list)[i]);

kvfree(*page_list);
*page_list = NULL;
--
2.43.0