In the Linux kernel, the following vulnerability has been resolved:
mm: don't try to NUMA-migrate COW pages that have other uses
Oded Gabbay reports that enabling NUMA balancing causes corruption with his Gaudi accelerator test load:
"All the details are in the bug, but the bottom line is that somehow, this patch causes corruption when the numa balancing feature is enabled AND we don't use process affinity AND we use GUP to pin pages so our accelerator can DMA to/from system memory.
Either disabling numa balancing, using process affinity to bind to specific numa-node or reverting this patch causes the bug to disappear"
and Oded bisected the issue to commit 09854ba94c6a ("mm: do_wp_page() simplification").
Now, the NUMA balancing shouldn't actually be changing the writability of a page, and as such shouldn't matter for COW. But it appears it does. Suspicious.
However, regardless of that, the condition for enabling NUMA faults in change_pte_range() is nonsensical. It uses "page_mapcount(page)" to decide if a COW page should be NUMA-protected or not, and that makes absolutely no sense.
The number of mappings a page has is irrelevant: not only does GUP get a reference to a page as in Oded's case, but the other mappings migth be paged out and the only reference to them would be in the page count.
Since we should never try to NUMA-balance a page that we can't move anyway due to other references, just fix the code to use 'page_count()'. Oded confirms that that fixes his issue.
Now, this does imply that something in NUMA balancing ends up changing page protections (other than the obvious one of making the page inaccessible to get the NUMA faulting information). Otherwise the COW simplification wouldn't matter - since doing the GUP on the page would make sure it's writable.
The cause of that permission change would be good to figure out too, since it clearly results in spurious COW events - but fixing the nonsensical test that just happened to work before is obviously the CorrectThing(tm) to do regardless.