Omar Sandoval <osandov@xxxxxxxxxxx> wrote:
> From: Omar Sandoval <osandov@xxxxxx>
>
> iter_folioq_get_pages() decides to advance to the next folioq slot when
> it has reached the end of the current folio. However, it is checking
> offset, which is the beginning of the current part, instead of
> iov_offset, which is adjusted to the end of the current part, so it
> doesn't advance the slot when it's supposed to. As a result, on the next
> iteration, we'll use the same folio with an out-of-bounds offset and
> return an unrelated page.
>
> This manifested as various crashes and other failures in 9pfs in drgn's
> VM testing setup and BPF CI.
>
> Fixes: db0aa2e9566f ("mm: Define struct folio_queue and ITER_FOLIOQ to handle a sequence of folios")
> Link: https://lore.kernel.org/linux-fsdevel/20240923183432.1876750-1-chantr4@xxxxxxxxx/
> Tested-by: Manu Bretelle <chantr4@xxxxxxxxx>
> Signed-off-by: Omar Sandoval <osandov@xxxxxx>
Thanks for finding that! That would explain why I didn't see it with afs or
cifs - both of those pass the iterator directly to the socket rather than
pulling the pages out of it. I'm not sure how I managed to do things like run
xfstests to completion and git clone and build a kernel without encountering
the bug.
Christian: Can you add this to vfs.fixes and tag it:
Acked-by: David Howells <dhowells@xxxxxxxxxx>
Tested-by: Eduard Zingerman <eddyz87@xxxxxxxxx>
