In the Linux kernel, the following vulnerability has been resolved:
mm/rmap: initialize nr_pages to 1 at loop start in try_to_unmap_one
Initialize nr_pages to 1 at the start of each loop iteration, like folio_referenced_one() does.
Without this, nr_pages computed by a previous folio_unmap_pte_batch() call can be reused on a later iteration that does not run folio_unmap_pte_batch() again.
mmap a 64K large folio with MAP_ANONYMOUS | MAP_DROPPABLE, then call madvise(MADV_FREE), then make the last page device-exclusive via HMM_DMIRROR_EXCLUSIVE.
Trigger node reclaim through sysfs. Now, in try_to_unmap_one(), we will first clear the first 15 out of 16 entries mapping the lazyfree folio. This will set nr_pages to 15. In the next pvmw walk, this nr_pages gets reused on a device-exclusive pte, thus potentially corrupting folio refcount/mapcount.
At the moment, I have a userspace program which can make the kernel spit out a trace, but the blow up is in folio_referenced_one(), because there are existing bugs in the interaction between device-private and rmap (which too I am investigating). I did a one liner kernel change to avoid going into folio_referenced_one(), and the kernel blows up at folio_remove_rmap_ptes in try_to_unmap_one which is what I wanted.
Note that the bug is there not since file folio batching but lazyfree folio batching, since device-exclusive only works for anonymous folios.
Userspace visible effect is simply kernel crashing somewhere due to refcount/mapcount corruption.