BUG: KCSAN: data-race in folio_batch_move_lru / mpage_read_end_io

[Matthew Wilcox]

On Mon, Aug 28, 2023 at 11:14:23PM +0200, Mirsad Todorovac wrote:
> In the vanilla torvalds tree 6.5 kernel on the Ubuntu 22.04 system, KCSAN found another data race:

KCSAN is wrong.

> [   34.102069] write (marked) to 0xffffef9a44978bc0 of 8 bytes by interrupt on cpu 28:
> [   34.108569] mpage_read_end_io (/home/marvin/linux/kernel/linux_torvalds/./arch/x86/include/asm/bitops.h:55 /home/marvin/linux/kernel/linux_torvalds/./include/asm-generic/bitops/instrumented-atomic.h:29 /home/marvin/linux/kernel/linux_torvalds/./include/linux/page-flags.h:739 /home/marvin/linux/kernel/linux_torvalds/fs/mpage.c:55)

        bio_for_each_folio_all(fi, bio) {
                if (err)
                        folio_set_error(fi.folio);
                else
                        folio_mark_uptodate(fi.folio);
                folio_unlock(fi.folio);
        }

It's noting the write to folio->flags in folio_mark_uptodate().  You can
see it's locked.  Also, the folio is under I/O.

> [   34.115221] read to 0xffffef9a44978bc0 of 8 bytes by task 348 on cpu 12:
> [   34.121702] folio_batch_move_lru (/home/marvin/linux/kernel/linux_torvalds/./include/linux/mm.h:1814 /home/marvin/linux/kernel/linux_torvalds/./include/linux/mm.h:1824 /home/marvin/linux/kernel/linux_torvalds/./include/linux/memcontrol.h:1636 /home/marvin/linux/kernel/linux_torvalds/./include/linux/memcontrol.h:1659 /home/marvin/linux/kernel/linux_torvalds/mm/swap.c:216)

Here, it's noting the read to folio->flags that's part of page_to_nid().

> [   34.121713] folio_batch_add_and_move (/home/marvin/linux/kernel/linux_torvalds/mm/swap.c:235)
> [   34.121724] folio_add_lru (/home/marvin/linux/kernel/linux_torvalds/./arch/x86/include/asm/preempt.h:95 /home/marvin/linux/kernel/linux_torvalds/mm/swap.c:518)
> [   34.121735] folio_add_lru_vma (/home/marvin/linux/kernel/linux_torvalds/mm/swap.c:538)
> [   34.121746] do_anonymous_page (/home/marvin/linux/kernel/linux_torvalds/mm/memory.c:4146)

Here we can see the page is freshly allocated.

So KCSAN has three things wrong here.  One is that the write to
folio_mark_uptodate() is setting a bit, that is nowhere near the bits
that are used for the node ID.  It can't know that; it doesn't track
writes at that granularity.

The second thing is that the node bits in folio->flags are immutable.
They're set at boot (or memory hotplug).  There is never a race risk when
reading them.  Presumably there needs to be some kind of annotation to
tell KCSAN that this is always safe.

The third thing is that these two accesses cannot race.  The write is
to a folio which is under I/O, so cannot be freed.  The read is to a
folio which has just been allocated, so cannot be under I/O.  This is
some kind of failure of KCSAN.