Re: riscv32 EXT4 splat, 6.8 regression?

[Conor Dooley]

On Sat, Apr 13, 2024 at 12:35:42AM -0400, Theodore Ts'o wrote:
> On Fri, Apr 12, 2024 at 06:59:19PM +0200, Björn Töpel wrote:
> > 
> >   $ pipx install tuxrun
> > 
> > if you're on Debian.
> > 
> > Then you can get the splat by running:
> > 
> >   $ tuxrun  --runtime docker --device qemu-riscv32 --kernel https://storage.tuxsuite.com/public/linaro/lkft/builds/2esMBaAMQJpcmczj0aL94fp4QnP/Image.gz --parameters SKIPFILE=skipfile-lkft.yaml --parameters SHARD_NUMBER=10 --parameters SHARD_INDEX=1 --image docker.io/linaro/tuxrun-dispatcher:v0.66.1 --tests ltp-controllers
> 
> Yeah, what I was hoping for was a shell script or a .c file hich was
> the reproducer, because that way I can run the test in my test infrastructure [1]
> 
> [1] https://github.com/tytso/xfstests-bld/blob/master/Documentation/kvm-xfstests.md
> 
> I'm sure there are plenty of nice things about tuxrun, but with
> kvm-xfstests I can easily get a shell so I can run the test sccript by
> hand, perhaps with strace so I can see what is going on.  Or I attach
> gdb to the kernel via "gdb /path/to/vmlinux" and "target remote
> localhost:7499".
> 
> I'm guessing that "ltp-controllers" means that the test might be from
> the Linux Test Project?  If so, that's great because I've added ltp
> support to my test infrastructure (which also supports blktests,
> phoronix test suite, and can be run on gce and on android devices in
> addition to qemu, and on the arm64, i386, and x86_64 architectures).
> 
> > Build with "make ARCH=riscv CROSS_COMPILE=riscv64-linux-gnu-", and make
> > sure to have the riscv64 cross-compilation support (yes, same toolchain
> > for rv32!).
> > 
> > It's when the rootfs is mounted, and the kernel is looking an init.
> 
> Hmm, so this happening as soon as the VM starts, before actually
> starting to run any tests?  Is it possible for you to send me the
> rootfs as a downloading image, as opposed to my trying to paw through
> the docker image?
> 
> > I'll keep debugging -- it was more if anyone had seen it before. I'll
> > try to reproduce on some other 32b platform as well.
> 
> Well, it's not happening on my rootfs on i386 using my test infrastructure:
> 
> % cd /usr/projects/linux/ext4
> % git checkout v6.8
> % install-kconfig --arch i386
> % kbuild --arch i386
> % kvm-xfstests shell
>     ...
> root@kvm-xfstests:~# cd ltp
> root@kvm-xfstests:~# ./runltp
> 
> (I don't have ltp support fully automated the way I can run blktests
> using "kvm-xfstests --blktests" or run xfstests via "gce-xfstests -c
> ext4/all -g auto".  The main missing is teaching ltp to create an
> junit xml results file so that the test results can be summarized and
> so the test results can be more easily summarized and compared against
> past runs on different kernel versions.)
> 
> Anyway, if you can send me your rootfs, I can try to take a look at it.

I think this should be the rootfs here:
https://drive.google.com/file/d/1HIo8EkAKY0xpTIIlwd9fXjRzmIdD7BUA/view?usp=sharing

I also attempted to bisect this and ended up at a slightly different
commit to Bjorn: 8c9440fea774 ("Merge tag 'vfs-6.8.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs")
That's a merge of 3f6984e7301f & 5bd3cf8cbc8a, both of which booted for
me. I also tried to bisect in reverse to find the fix a la syzbot, since it
is not broken in 6.9, but that's pretty error prone and I ended up down
branches based on 6.7 and was not able to find the fix.

Attachment: signature.asc
Description: PGP signature