Re: Repeatable, raid1+O_DIRECT, hang/warn

["Dr. David Alan Gilbert" <dave@treblig.org>]

* Dr. David Alan Gilbert (dave@treblig.org) wrote:
> * Keith Busch (kbusch@kernel.org) wrote:
> > On Mon, Jun 15, 2026 at 04:16:12PM -0700, Vjaceslavs Klimovs wrote:
> > > Your trace looks like what the two earlier reports hit: a read reaching
> > > a leaf device with sectors > 0 but phys_seg 0 (an empty bio). One aside
> > > that may help read the trace: blk_io_trace.error is a __u16, so the
> > > bracketed values on your C lines are errnos as u16 (65514 = -EINVAL,
> > > 65531 = -EIO).
> > > 
> > > The WARN itself is new, the bad bio isn't. bio_add_page() only started
> > > rejecting len == 0 in 643893647cac ("block: reject zero length in
> > > bio_add_page()", v7.1-rc1); on 7.0.8 the same empty bio tripped
> > > scsi_alloc_sgtables()'s !nr_segs instead, which matches what you saw.
> > > That fits your "not a recent regression": the condition is older, v7.1
> > > just made it loud.
> > > 
> > > For Tomas's and my reports (QEMU O_DIRECT to the LV block device) the
> > > origin looks like 5ff3f74e145a ("block: simplify direct io validity
> > > check", v6.18): blkdev_dio_invalid() now checks only aggregate
> > > ki_pos | count alignment and dropped the per-segment
> > > bdev_iter_is_aligned() walk, so a degenerate or misaligned O_DIRECT no
> > > longer gets -EINVAL at the fops boundary. But your reproducer reads a
> > > file, which goes through the filesystem O_DIRECT path and never calls
> > > blkdev_dio_invalid(), and still makes the empty bio. So it isn't only
> > > that one entry point.
> > > 
> > > dm-mirror then hangs because Keith's f7b24c7b41f2 only covers md
> > > raid1/raid10; legacy dm-mirror (dm-raid1.c) has no equivalent and
> > > rebuilds the empty read onto the other leg. Note the leg's status isn't
> > > even consistent (your SATA path returns BLK_STS_IOERR, not
> > > BLK_STS_INVAL), so copying that status check into dm-mirror probably
> > > wouldn't catch every case.
> > > 
> > > For what it's worth, that points me toward rejecting the empty or
> > > misaligned bio once, at submission, with -EINVAL, rather than teaching
> > > each consumer to tolerate it. But you'll know the tradeoffs far better
> > > than I do.
> > > 
> > > I have a small QEMU + LVM raid1/mirror setup that reproduces the
> > > block-device variant and bisects to 5ff3f74e. Happy to run your file
> > > reproducer with some instrumentation at the dm-mirror read entry
> > > (bi_size vs bio_sectors vs bvec lengths) to see whether the bio is
> > > already empty on arrival or built that way on the retry, and to test
> > > any patch.
> > 
> > Thanks for following up here. I didn't initially see your follow-up
> > until Thorsten linked it. I apologize for missing that, this feature is
> > important so I don't want to see anything regress for it.
> > 
> > There is a known bug fix I think future tests should include:
> > 
> >   https://lore.kernel.org/linux-block/20260612223205.465913-1-kbusch@meta.com/
> 
> > This likely isn't the fix you're looking for, but including it rules out
> > conditions that are not important here.
> > 
> > After that, can we try this suggestion and see if the hang goes away?
> > 
> >   https://lore.kernel.org/linux-block/ajBb8tK-0aJBpIgF@kbusch-mbp/
> 
> With just that one in, the machine survives - thanks!
> 
> It does give:
> 
> [  505.208354] device-mapper: raid1: Mirror read failed from 252:24. Trying alternative device.
> [  505.239376] device-mapper: raid1: All sides of mirror have failed.
> [  505.239389] device-mapper: raid1: Read failure on mirror device 252:25.  Failing I/O.
> [  505.239394] device-mapper: raid1: Mirror read failed.
> 
> Although as far as I can tell the RAID hasn't errored and is still in sync.
> 
> If I turn the test case into a write (just s/pread/pwrite/ ) - the machine
> still survives but then it does lose raid sync, and the raid resync
> seems to stick until I do a 'lvchange --refresh main/lvol0'
> which recovers after having spat out a:
> 
> [  865.319527] Buffer I/O error on dev dm-26, logical block 262128, async page read
> 
> > I expect the original test case to still return an error (and I think it
> > was designed to), but it shouldn't produce the warn or bug splats with a
> > stuck uninterruptable task.
> 
> It's not clear to me if it was designed to fail or not; I've not had
> a chance to rerun the original qemu block tests yet, and I don't know
> if old kernels succesfully used O_DIRECT in this case.
> 
> It still feels that my pwrite case above shouldn't cause a raid de-sync
> (especially since a normal user can do it).

Just to follow up on that;  if I use the modern lvm mode 
( lvcreate  -m 1 -L 1G main /dev/sda2 /dev/sdb2 ) rather than
the old mirror with the same patch, then:

  a) I get no log errors with either read or write
  b) read still gives EIO
  c) write apparently succeeds ?!

Dave

> Dave
> -- 
>  -----Open up your eyes, open up your mind, open up your code -------   
> / Dr. David Alan Gilbert    |       Running GNU/Linux       | Happy  \ 
> \        dave @ treblig.org |                               | In Hex /
>  \ _________________________|_____ http://www.treblig.org   |_______/
-- 
 -----Open up your eyes, open up your mind, open up your code -------   
/ Dr. David Alan Gilbert    |       Running GNU/Linux       | Happy  \ 
\        dave @ treblig.org |                               | In Hex /
 \ _________________________|_____ http://www.treblig.org   |_______/