Re: [PATCH] mm/execmem: only free ROX cache vm_struct when fully unused

[Mike Rapoport <rppt@kernel.org>]

On Wed, Oct 14, 2026 at 04:30:58PM -0700, Arnav Kansal wrote:
> execmem_cache_clean() currently uses:
> 
>     IS_ALIGNED(size, PMD_SIZE) && IS_ALIGNED(mas.index, PMD_SIZE)

This is not the case anymore:

https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/commit/mm/execmem.c?id=569db54d8b7a4ec463a62c444de0d6f5f10e9dc1
 
> as a proxy to determine whether a free range in execmem_cache.free_areas
> spans an entire unused vm_struct that can be returned to vmalloc.
> 
> This heuristic is broken in both directions:
> 
> 1. Premature vfree() of multi-PMD vm_structs (Use-After-Free):
>    When execmem_cache_populate_alloc() services a request larger than
>    PMD_SIZE (for example, a module with > 6 MiB of .text),
>    alloc_size = round_up(size, PMD_SIZE) creates a single multi-PMD
>    vm_struct (e.g., 8 MiB) whose remaining tail space is handed out to
>    subsequent module allocations (e.g., kvm.ko). When a module inside
>    that vm_struct unloads and leaves a PMD_SIZE-aligned free sub-range
>    (such as a 2 MiB or 4 MiB chunk at the start of the 8 MiB vm_struct),
>    execmem_cache_clean() passes the IS_ALIGNED() check and calls
>    vfree(area), which unmaps and frees the entire enclosing multi-PMD
>    vm_struct while sibling modules are still live:
> 
>      BUG: unable to handle page fault for address: ffffffffc07dc030
>      #PF: supervisor instruction fetch in kernel mode
>      #PF: error_code(0x0010) - not-present page
>      PGD 2c33067 P4D 2c33067 PUD 2c35067 PMD 0
>      Oops: Oops: 0010 [#1] SMP NOPTI
>      RIP: 0010:dummy_exit+0x0/0xfd0 [tailm]
> 
> 2. Failure to reclaim fallback vm_structs:
>    After commit 3bd4e0ac61b2 ("execmem: add fallback for failures in
>    vmalloc(VM_ALLOW_HUGE_VMAP)") and commit 4afb85f5e359 ("mm: remove
>    PMD alignment constraint in execmem_vmalloc()"), ROX cache vm_structs
>    allocated via the base-page or exact-size fallback paths are only
>    PAGE_SIZE-aligned in address and/or size, so they never satisfy the
>    PMD_SIZE alignment check even when completely unused.
> 
> Because vmalloc allocations include a trailing guard page, separate
> vm_struct allocations can never coalesce into a single contiguous entry
> in execmem_cache.free_areas. Replace the PMD_SIZE alignment check with
> an explicit check that the free range starts at vm->addr and spans the
> entire get_vm_area_size(vm).
> 
> Fixes: 2e45474ab14f ("execmem: add support for cache of large ROX pages")
> Cc: stable@vger.kernel.org
> Assisted-by: LLM
> Co-developed-by: Rohan Kumar <rohakum@google.com>
> Signed-off-by: Rohan Kumar <rohakum@google.com>
> Signed-off-by: Arnav Kansal <rnv@google.com>
> ---
>  mm/execmem.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/execmem.c b/mm/execmem.c
> index 74a178a87e75..3a1da0498fe2 100644
> --- a/mm/execmem.c
> +++ b/mm/execmem.c
> @@ -166,11 +166,9 @@ static void execmem_cache_clean(struct work_struct *work)
>  	mutex_lock(mutex);
>  	mas_for_each(&mas, area, ULONG_MAX) {
>  		size_t size = mas_range_len(&mas);
> +		struct vm_struct *vm = find_vm_area(area);
>  
> -		if (IS_ALIGNED(size, PMD_SIZE) &&
> -		    IS_ALIGNED(mas.index, PMD_SIZE)) {
> -			struct vm_struct *vm = find_vm_area(area);
> -
> +		if (vm->addr == area && get_vm_area_size(vm) == size) {
>  			execmem_set_direct_map_valid(vm, true);
>  			mas_store_gfp(&mas, NULL, GFP_KERNEL);
>  			vfree(area);
> -- 
> 2.47.0

-- 
Sincerely yours,
Mike.