On 2024/11/12 03:34, Jann Horn wrote:
On 32-bit platforms, it is possible for the expression `len + old_addr < old_end` to be false-positive if `len + old_addr` wraps around. `old_addr` is the cursor in the old range up to which page table entries have been moved; so if the operation succeeded, `old_addr` is the *end* of the old region, and adding `len` to it can wrap.
The overflow causes mremap() to mistakenly believe that PTEs have been copied; the consequence is that mremap() bails out, but doesn't move the PTEs back before the new VMA is unmapped, causing anonymous pages in the region to be lost. So basically if userspace tries to mremap() a private-anon region and hits this bug, mremap() will return an error and the private-anon region's contents appear to have been zeroed.
The idea of this check is that `old_end - len` is the original start address, and writing the check that way also makes it easier to read; so fix the check by rearranging the comparison accordingly.
(An alternate fix would be to refactor this function by introducing an "orig_old_start" variable or such.)
Cc: stable@vger.kernel.org Fixes: af8ca1c14906 ("mm/mremap: optimize the start addresses in move_page_tables()") Signed-off-by: Jann Horn jannh@google.com
Acked-by: Qi Zheng zhengqi.arch@bytedance.com
Thanks!