Re: [PATCH v2 2/4] overlay: Document current outstanding shortcoming of volatile

[Amir Goldstein <amir73il@gmail.com>]

> I notice you maintain overlay tests outside of the kernel. Unfortunately, I
> think for this kind of test, it requires in kernel code to artificially bump the
> writeback error count on the upperdir, or it requires the failure injection
> infrastructure.
>
> Simulating this behaviour is non-trivial without in-kernel support:
>
> P1: Open(f) -> p1.fd
> P2: Open(f) -> p2.fd
> P1: syncfs(p1.fd) -> errrno
> P2: syncfs(p1.fd) -> 0 # This should return an error
>

failure injection is an option. xfstest generic/019 is an example.
generic/108 uses a different method and generic/361 uses a plain
loop image over commit without any fault injection to trigger writeback
errors.

With current xfstests, check -overlay run (runs all generic tests with
overlayfs over the configured base fs) all the 3 tests mentioned above
will be skipped because of _require_block_device, but the requirement
is there for a different reason for each of them.

At first look, the loop device approach is the most generic one and could
easily work also with overlayfs, so you could create an overlay specific
test (non generic) based on generic/361, but it is not easy to use the
helper _scratch_mkfs_sized, because check -overlay runs do not mkfs
the base scratch fs.

My recommendation would be to fix generic/019 in a similar manner
to the way that tests that _require_scratch_shutdown were fixed to run
with check -overlay:

* Instead of _require_block_device, add a specific helper
   _require_scratch_fail_make_request, which like _require_scratch_shutdown
   knows how to deal with overlay FSTYP correctly

* Instead of `_sysfs_dev $SCRATCH_DEV` use a helper _scratch_sysfs_dev
   that knows how to deal with overlay FSTYP correctly

This will add test coverage to overlayfs fsync/syncfs and then you can
do one or both of:
1. Run 'check -overlay generic/019' with  OVERLAY_MOUNT_OPTIONS="volatile"
2. Fork an overlay specific test from the generic test that will test the
    volatile error handling on every 'check -overlay -g quick' run

#2 will provide better coverage against regressions in volatile writeback error
handling and will be a good start for a test to reuse a volatile mount after
writeback errors.

Thanks,
Amir.