On Thu, Jul 18, 2019 at 07:53:42PM -0700, Dan Williams wrote:
> [ add Sasha for -stable advice ]
>
> On Thu, Jul 18, 2019 at 5:13 PM Liu Bo <bo.liu@xxxxxxxxxxxxxxxxx> wrote:
> >
> > The livelock can be triggerred in the following pattern,
> >
> > while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
> > min(end - index, (pgoff_t)PAGEVEC_SIZE),
> > indices)) {
> > ...
> > for (i = 0; i < pagevec_count(&pvec); i++) {
> > index = indices[i];
> > ...
> > }
> > index++; /* BUG */
> > }
> >
> > multi order exceptional entry is not specially considered in
> > invalidate_inode_pages2_range() and it ended up with a livelock because
> > both index 0 and index 1 finds the same pmd, but this pmd is binded to
> > index 0, so index is set to 0 again.
> >
> > This introduces a helper to take the pmd entry's length into account when
> > deciding the next index.
> >
> > Note that there're other users of the above pattern which doesn't need to
> > fix,
> >
> > - dax_layout_busy_page
> > It's been fixed in commit d7782145e1ad
> > ("filesystem-dax: Fix dax_layout_busy_page() livelock")
> >
> > - truncate_inode_pages_range
> > This won't loop forever since the exceptional entries are immediately
> > removed from radix tree after the search.
> >
> > Fixes: 642261a ("dax: add struct iomap based DAX PMD support")
> > Cc: <stable@xxxxxxxxxxxxxxx> since 4.9 to 4.19
> > Signed-off-by: Liu Bo <bo.liu@xxxxxxxxxxxxxxxxx>
> > ---
> >
> > The problem is gone after commit f280bf092d48 ("page cache: Convert
> > find_get_entries to XArray"), but since xarray seems too new to backport
> > to 4.19, I made this fix based on radix tree implementation.
>
> I think in this situation, since mainline does not need this change
> and the bug has been buried under a major refactoring, is to send a
> backport directly against the v4.19 kernel. Include notes about how it
> replaces the fix that was inadvertently contained in f280bf092d48
> ("page cache: Convert find_get_entries to XArray"). Do you have a test
> case that you can include in the changelog?
The root cause behind the bug is exactly same as what commit
d7782145e1ad ("filesystem-dax: Fix dax_layout_busy_page() livelock")
does.
For test case, I have a not 100% reproducible one based on ltp's
rwtest[1] and virtiofs.
[1]:
$mount -t virtio_fs -o tag=alwaysdax -o rootmode=040000,user_id=0,group_id=0,dax,default_permissions,allow_other alwaysdax /mnt/virtio-fs/
$cat test.txt
rwtest01 export LTPROOT; rwtest -N rwtest01 -c -q -i 60s -f sync 10%25000:$TMPDIR/rw-sync-$$
$runltp -d /mnt/virtio-fs -f test.txt
thanks,
-liubo
