Add a sanity check to madvise_dontneed_free() to address a corner case in madvise where a race condition causes the current vma being processed to be backed by a different page size.
During a madvise(MADV_DONTNEED) call on a memory region registered with a userfaultfd, there's a period of time where the process mm lock is temporarily released in order to send a UFFD_EVENT_REMOVE and let userspace handle the event. During this time, the vma covering the current address range may change due to an explicit mmap done concurrently by another thread.
If, after that change, the memory region, which was originally backed by 4KB pages, is now backed by hugepages, the end address is rounded down to a hugepage boundary to avoid data loss (see "Fixes" below). This rounding may cause the end address to be truncated to the same address as the start.
Make this corner case follow the same semantics as in other similar cases where the requested region has zero length (ie. return 0).
This will make madvise_walk_vmas() continue to the next vma in the range (this time holding the process mm lock) which, due to the prev pointer becoming stale because of the vma change, will be the same hugepage-backed vma that was just checked before. The next time madvise_dontneed_free() runs for this vma, if the start address isn't aligned to a hugepage boundary, it'll return -EINVAL, which is also in line with the madvise api.
From userspace perspective, madvise() will return EINVAL because the start address isn't aligned according to the new vma alignment requirements (hugepage), even though it was correctly page-aligned when the call was issued.
Fixes: 8ebe0a5eaaeb ("mm,madvise,hugetlb: fix unexpected data loss with MADV_DONTNEED on hugetlbfs") Cc: stable@vger.kernel.org Signed-off-by: Ricardo Cañuelo Navarro rcn@igalia.com --- Changes in v2: - Added documentation in the code to tell the user how this situation can happen. (Andrew) --- mm/madvise.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/mm/madvise.c b/mm/madvise.c index 49f3a75046f6..08b207f8e61e 100644 --- a/mm/madvise.c +++ b/mm/madvise.c @@ -933,7 +933,16 @@ static long madvise_dontneed_free(struct vm_area_struct *vma, */ end = vma->vm_end; } - VM_WARN_ON(start >= end); + /* + * If the memory region between start and end was + * originally backed by 4kB pages and then remapped to + * be backed by hugepages while mmap_lock was dropped, + * the adjustment for hugetlb vma above may have rounded + * end down to the start address. + */ + if (start == end) + return 0; + VM_WARN_ON(start > end); }
if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
On Mon, Feb 03, 2025 at 08:52:06AM +0100, Ricardo Cañuelo Navarro wrote:
Add a sanity check to madvise_dontneed_free() to address a corner case in madvise where a race condition causes the current vma being processed to be backed by a different page size.
During a madvise(MADV_DONTNEED) call on a memory region registered with a userfaultfd, there's a period of time where the process mm lock is temporarily released in order to send a UFFD_EVENT_REMOVE and let userspace handle the event. During this time, the vma covering the current address range may change due to an explicit mmap done concurrently by another thread.
If, after that change, the memory region, which was originally backed by 4KB pages, is now backed by hugepages, the end address is rounded down to a hugepage boundary to avoid data loss (see "Fixes" below). This rounding may cause the end address to be truncated to the same address as the start.
Make this corner case follow the same semantics as in other similar cases where the requested region has zero length (ie. return 0).
This will make madvise_walk_vmas() continue to the next vma in the range (this time holding the process mm lock) which, due to the prev pointer becoming stale because of the vma change, will be the same hugepage-backed vma that was just checked before. The next time madvise_dontneed_free() runs for this vma, if the start address isn't aligned to a hugepage boundary, it'll return -EINVAL, which is also in line with the madvise api.
From userspace perspective, madvise() will return EINVAL because the start address isn't aligned according to the new vma alignment requirements (hugepage), even though it was correctly page-aligned when the call was issued.
Fixes: 8ebe0a5eaaeb ("mm,madvise,hugetlb: fix unexpected data loss with MADV_DONTNEED on hugetlbfs") Cc: stable@vger.kernel.org Signed-off-by: Ricardo Cañuelo Navarro rcn@igalia.com
Reviewed-by: Oscar Salvador osalvador@suse.de
Changes in v2:
- Added documentation in the code to tell the user how this situation can happen. (Andrew)
mm/madvise.c | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/mm/madvise.c b/mm/madvise.c index 49f3a75046f6..08b207f8e61e 100644 --- a/mm/madvise.c +++ b/mm/madvise.c @@ -933,7 +933,16 @@ static long madvise_dontneed_free(struct vm_area_struct *vma, */ end = vma->vm_end; }
VM_WARN_ON(start >= end);
/** If the memory region between start and end was* originally backed by 4kB pages and then remapped to* be backed by hugepages while mmap_lock was dropped,* the adjustment for hugetlb vma above may have rounded* end down to the start address.*/if (start == end)return 0;VM_WARN_ON(start > end);
The change itself looks fine to me, although I am wondering whether it would make more sense to place the check right after the call to madvise_dontneed_free_valid_vma(). It looks kind of more logical to me, but not a big deal.
linux-stable-mirror@lists.linaro.org
-
Oscar Salvador -
Ricardo Cañuelo Navarro