linux-fsdevel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* Significant brokenness in DIO loopback path
@ 2022-03-10  1:44 Kent Overstreet
  2022-03-10  3:16 ` Ming Lei
  0 siblings, 1 reply; 3+ messages in thread
From: Kent Overstreet @ 2022-03-10  1:44 UTC (permalink / raw)
  To: linux-fsdevel, linux-kernel, Pavel Begunkov
  Cc: Christoph Hellwig, Ming Lei, Jens Axboe, Al Viro

So I'm testing bcachefs with the loopback driver in dio mode, and noticing
_significant_ brokenness in the bio_iov_iter_get_pages() path and elsewhere.

1) We don't check that we're not asking for more pages than we're in the
original bio

Noticed this because of another bug:

2) the loopback driver appears to never look at the underlying filesystem's
block size, meaning if the filesystem advertises a block size of 4k the loopback
device's blocksize will still be 512, and we'll end up issuing IOs the DIO path
shouldn't allow due to alignment.

3) iov_iter_bvec_advance() looks like utter nonsense. We're synthesizing a fake
bvec_iter and never using or even looking at one from the original bio, looking
at the construction in iov_iter_bvec().

This is broken; you're assuming you're never going to see bios with partially
completed bvec_iters, or things are going to explode.

Try putting a md raid0 on top of two loopback devices with a sub page block
size, things are just going to explode.

iov_iter_bvec() needs to be changed to take a bio, not a bvec array, and
iov_iter_bvec_advance() should probably just call bio_advance() - and
bio_iov_bvec_set() needs to be changed to just copy bi_iter from the original
bio into the dest bio. You guys made this way more complicated than it needed to
be.

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2022-03-10  5:17 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2022-03-10  1:44 Significant brokenness in DIO loopback path Kent Overstreet
2022-03-10  3:16 ` Ming Lei
2022-03-10  5:17   ` Kent Overstreet

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).