Bug 1217010

The problem is very likely related to new O_DIRECT support on tmpfs [1] on 6.6.1-1-default. It needs to be checked, whether it's a test bug (likely) or really bug in the kernel implementation.

Previously, on 6.5.9-1, all dio tests were skipped (TCONF) due missing O_DIRECT support on tmpfs (on /tmp).

Also, because of this skipping, we set openQA variable LTP_ENV as TMPDIR=/var/tmp/ on ltp_dio job, which is used on kernel tests on o3. That's why we didn't notice. But somebody created jeos-ltp-dio without this setup.

IMHO we should

1) add in LTP upstream .all_filesystems = 1 to run at least on some filesystems (and after that remove TMPDIR=/var/tmp/ setup)

2) unify jeos-ltp-dio with ltp_dio (and generally all jeos-ltp-* with their ltp_*).

[1] https://kernelnewbies.org/Linux_6.6#TMPFS

aa5b9178c019 ("mm: invalidation check mapping before folio_contains")
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e88e0d366f9cfbb810b0c8509dc5d130d5a53e02

(In reply to Petr Vorel from comment #2)
> aa5b9178c019 ("mm: invalidation check mapping before folio_contains")
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/
> ?id=e88e0d366f9cfbb810b0c8509dc5d130d5a53e02

I overlook the

(In reply to Petr Vorel from comment #2)
> aa5b9178c019 ("mm: invalidation check mapping before folio_contains")
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/
> ?id=e88e0d366f9cfbb810b0c8509dc5d130d5a53e02

I overlook in the commit message, that author itself noticed tests failing on tmpfs:

	dio01 through dio30, except for dio04 and dio10, which fail because
	tmpfs dio read and write allow odd count: tmpfs could be made stricter,
	but would that be an improvement?

@Jan, reading your ack on v1 [1] and also on v2 [2] (where you maintainers discussed only amount of code duplication and not the fact that LTP tests failing, I suppose implementation is accepted and we need to make test on tmpfs less strict, right?

[1] https://lore.kernel.org/all/20230809095447.7iturpglkvbzyvmg@quack3/
[2] https://lore.kernel.org/all/20230811085632.hfmdzni2yzgmcy44@quack3/

Yes, tmpfs has no reason to reject direct IO of any size so it would be good if LTP test can be relaxed. As far as I'm checking ltp/testcases/kernel/io/direct_io/diotest4.c it seems there are already some filesystems that have exception (NFS, FUSE, BTRFS) so adding one more is likely no big deal?

(In reply to Jan Kara from comment #5)
> Yes, tmpfs has no reason to reject direct IO of any size so it would be good
> if LTP test can be relaxed. As far as I'm checking
> ltp/testcases/kernel/io/direct_io/diotest4.c it seems there are already some
> filesystems that have exception (NFS, FUSE, BTRFS) so adding one more is
> likely no big deal?

Thanks for confirmation that there is no plan to further fix tmpfs. And thanks for a hint, I'll send a patch shortly.

I guess we can close this as invalid (bug in the test).

(In reply to Petr Vorel from comment #2)
> aa5b9178c019 ("mm: invalidation check mapping before folio_contains")

The commit hash and message is actually:
e88e0d366f9c ("tmpfs: trivial support for direct IO")

> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/
> ?id=e88e0d366f9cfbb810b0c8509dc5d130d5a53e02

Patch fixing the test sent, it should be merged soon, thus closing as invalid:
https://lore.kernel.org/ltp/20231122211318.31358-1-pvorel@suse.cz/