On Wed, May 04, 2022 at 09:24:05PM -0700, Darrick J. Wong wrote:
> On Thu, May 05, 2022 at 05:18:27AM +0100, Matthew Wilcox wrote:
> > On Wed, May 04, 2022 at 07:40:12PM -0700, Darrick J. Wong wrote:
> > > On Tue, May 03, 2022 at 10:25:32AM -0700, Darrick J. Wong wrote:
> > > > On Tue, May 03, 2022 at 05:31:01AM +0100, Matthew Wilcox wrote:
> > > > > On Mon, May 02, 2022 at 08:25:34PM -0700, Darrick J. Wong wrote:
> > > > > > On Mon, May 02, 2022 at 08:20:00AM -0400, Brian Foster wrote:
> > > > > > > On Sat, Apr 30, 2022 at 10:40:31PM +0100, Matthew Wilcox wrote:
> > > > > > > > On Sat, Apr 30, 2022 at 04:44:07AM +0100, Matthew Wilcox wrote:
> > > > > > > > > (I do not love this, have not even compiled it; it's late. We may be
> > > > > > > > > better off just storing next_folio inside the folio_iter).
> > > > > > > >
> > > > > > > > Does anyone have a preference for fixing this between Option A:
> > > > > > > >
> > > > > > >
> > > > > > > After seeing the trace in my previous mail and several thousand
> > > > > > > successful iterations of the test hack, I had reworked it into this
> > > > > > > (which survived weekend testing until it ran into some other XFS problem
> > > > > > > that looks unrelated):
> > > > > > >
> > > > > > > diff --git a/include/linux/bio.h b/include/linux/bio.h
> > > > > > > index 278cc81cc1e7..aa820e09978e 100644
> > > > > > > --- a/include/linux/bio.h
> > > > > > > +++ b/include/linux/bio.h
> > > > > > > @@ -269,6 +269,7 @@ struct folio_iter {
> > > > > > > size_t offset;
> > > > > > > size_t length;
> > > > > > > /* private: for use by the iterator */
> > > > > > > + struct folio *_next;
> > > > > > > size_t _seg_count;
> > > > > > > int _i;
> > > > > > > };
> > > > > > > @@ -279,6 +280,7 @@ static inline void bio_first_folio(struct folio_iter *fi, struct bio *bio,
> > > > > > > struct bio_vec *bvec = bio_first_bvec_all(bio) + i;
> > > > > > >
> > > > > > > fi->folio = page_folio(bvec->bv_page);
> > > > > > > + fi->_next = folio_next(fi->folio);
> > > > > > > fi->offset = bvec->bv_offset +
> > > > > > > PAGE_SIZE * (bvec->bv_page - &fi->folio->page);
> > > > > > > fi->_seg_count = bvec->bv_len;
> > > > > > > @@ -290,13 +292,15 @@ static inline void bio_next_folio(struct folio_iter *fi, struct bio *bio)
> > > > > > > {
> > > > > > > fi->_seg_count -= fi->length;
> > > > > > > if (fi->_seg_count) {
> > > > > > > - fi->folio = folio_next(fi->folio);
> > > > > > > + fi->folio = fi->_next;
> > > > > > > + fi->_next = folio_next(fi->folio);
> > > > > > > fi->offset = 0;
> > > > > > > fi->length = min(folio_size(fi->folio), fi->_seg_count);
> > > > > > > } else if (fi->_i + 1 < bio->bi_vcnt) {
> > > > > > > bio_first_folio(fi, bio, fi->_i + 1);
> > > > > > > } else {
> > > > > > > fi->folio = NULL;
> > > > > > > + fi->_next = NULL;
> > > > > > > }
> > > > > > > }
> > > > > > >
> > > > > > > So FWIW, that is just to say that I find option A to be cleaner and more
> > > > > > > readable.
> > > > > >
> > > > > > Me too. I'll queue up the usual nightly tests with that patch added and
> > > > > > we'll see how that does.
> > > > >
> > > > > I've just pushed essentially that patch to my for-next tree in case
> > > > > anybody does any testing with that. I'll give it a couple of days
> > > > > before creating a folio-5.18f tag and asking Linus to pull the first two
> > > > > commits on
> > > > >
> > > > > git://git.infradead.org/users/willy/pagecache.git for-next
> > > > >
> > > > > That is, commits
> > > > >
> > > > > 1a4c97e2dd5b ("block: Do not call folio_next() on an unreferenced folio")
> > > > > 095099da208b ("mm/readahead: Fix readahead with large folios")
> > > >
> > > > Hmm. Well, I added 1a4c97 to my tree last night, and it seems to have
> > > > cleared up all but two of the problems I saw with the for-next branch.
> > > >
> > > > generic/388 still fails (40 minutes in) with:
> > > >
> > > > WARN_ON_ONCE(atomic_read(&iop->write_bytes_pending));
> > > > VM_BUG_ON_FOLIO(i_blocks_per_folio(inode, folio) > 1 && !iop, folio);
> > > >
> > > > Which I think is the same problem where the fs goes down, XFS throws an
> > > > error back to iomap_do_writepages, and it tries to discard a folio that
> > > > already had writeback running on it.
> > > >
> > > > There's also the same problem I reported a few days ago in xfs/501
> > > > on a 64k-page ARM64 VM where:
> > > >
> > > > run fstests xfs/501 at 2022-05-02 21:17:31
> > > > XFS: Assertion failed: IS_ALIGNED((unsigned long)lv->lv_buf, sizeof(uint64_t)), file: fs/xfs/xfs_log_cil.c, line: 430
> > > > XFS: Assertion failed: IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)), file: fs/xfs/xfs_log.c, line: 137
> > > > XFS: Assertion failed: IS_ALIGNED((unsigned long)buf, sizeof(uint64_t)), file: fs/xfs/xfs_log.c, line: 137
> > > >
> > > > But I think that's a new bug that came in with all the log buffer
> > > > alignment changes in the 5.19 branch.
> > > >
> > > > Oh. My tree still had the "disable large folios" patch in it. I guess
> > > > the "successful" results are mostly invalid then.
> > >
> > > Well... with large folios turned back on and those two patches added to
> > > the branch, *most* of the problems go away. The generic/388 problem
> > > persists, and last night's run showed that the weird xfs_dquot leak that
> > > I"ve occasionally seen on 5.18 with xfs/43[46] also exists in 5.17.
> >
> > OK, so let me just restate what I understand here ... these two patches
> > cure some, but not all of the currently observed problems with 5.18.
> > The problems that remain with 5.18 were also present in either 5.17
> > or in 5.18 with large folios disabled, so at this point we know of no
> > functional regressions that large folios can be blamed for?
>
> Not quite -- the generic/388 one is definitely new as of 5.18-rc1.
> Frustratingly, the problems "go away" if you enable KASAN, so I might
> try KFENCE (or whatever the new less molasses memory debugger is named)
> next.
I have some semi-good news for willy -- I've created an iomap-5.19-merge
candidate branch with Andreas' iomap fixes and willy's two folio
patches, and it tests cleanly on x64.
Unfortunately, I still see arm64 vms tripping over:
WARN_ON(i_blocks_per_folio(inode, folio) > 1 && !iop)
with the same weird pattern where we try to discard an entire folio
even though the (n>1)th block in the folio is under writeback. I added
even more debug info, and captured this:
ino 0x87 ibpf 0x40 mapping 0xfffffc0164c8c0e8 index 0x8
page:ffffffff004a0f00 refcount:4 mapcount:0 mapping:fffffc0164c8c0e8 index:0x8 pfn:0x1683c
head:ffffffff004a0f00 order:2 compound_mapcount:0 compound_pincount:0
memcg:fffffc0106c7c000
aops:xfs_address_space_operations [xfs] ino:87 dentry name:"file2"
flags: 0x1ffe000000018116(error|referenced|uptodate|lru|writeback|head|node=0|zone=1|lastcpupid=0x7ff)
raw: 1ffe000000018116 ffffffff00399808 ffffffff003bc408 fffffc0164c8c0e8
raw: 0000000000000008 0000000000000000 00000004ffffffff fffffc0106c7c000
page dumped because: VM_BUG_ON_FOLIO(i_blocks_per_folio(inode, folio) > 1 && !iop)
Looks like we have a single-page folio experiencing this problem. I'm
not sure if we've simply been tripping over this all along and I just
never noticed until I turned the WARN_ON into a VM_BUG_ON, which
actually takes down the system... but I suspect this is a problem with
writeback (albeit on a shut down filesystem) that has been around for a
while.
I also have some not so good news for Dave -- I think this implies that
something in the xfs 5.19 merge branch might be behind all of these
weird dquot leaks in xfs/43[46] and the blowups in generic/388. I'll
try to get to bisecting that next week.
--D
> I suspect the xfs/43[46] problems are probably more of the ongoing log
> abend whackamole, since every time we change something in the logging
> code, weird latent bugs from 20 years ago start pouring out like spiders
> fleeing the first winter rains.
>
> > I'll send these patches to Linus tomorrow then, since they fix problems
> > that have been observed, and I don't think there's any reason to keep
> > them from him.
>
> Right, I don't think there's a reason to hang on to those any longer.
>
> --D
