On Sun, Aug 01, 2021 at 08:53:36PM +0800, Eryu Guan wrote:

quoted

So the underlying disk is 1TB in size, and we ended up using this 1T
filesystem when _scratch_mkfs_sized failed?

But we have done _try_wipe_scratch_devs before each test to make sure we
don't use previous scratch dev accidently (just like this case), and the
subsesquent _scratch_mount will fail and fail the whole test. So it's
not clear to me what caused the failure you hit.

The call to _try_wipe_scratch_devs was added in 2019.  My commit to
add:

	|| _notrun "mkfs.${FSTYP} failed"

dates from 2017.  So the reason I was seeing the problem was because
it was before we started running wipefs between tests.

That being said, I've checked a recent test run, and the _notrun
hasn't triggered recently.  Looking at the git history, it looks like
a large number of tests had their arguments to _scratch_mkfs_sized
adjusted upwards to avoid failures when running with 64k block sizes
on powerpc.

Going back to generic/204, I see why Josef ran into issues, however.
even though we are running wipefs before each test.  In the case of
generic/204, it runs _scratch_mkfs to determine the blocksize, and
then it runs _scratch_mkfs_sized --- and if it fails, the file system
is left at the full size of the scratch file system, and then
generic/204 takes a vey long time.

So even if we can rely on wipefs causing the tests to fail, maybe we
should just add a check for mkfs failure to _scratch_mkfs_sized?  I
think that's a better fix than Josef's proposed patch to generic/204.
One benefit of adding the check to _scratch_mkfs_sized is we can
supply a clearer explanation of the failure since the failure would be
"mkfs failed" as opposed to "mount: /vdc: wrong fs type, bad option,
bad superblock on /dev/vdc, missing codepage or helper program, or
other error."

It might also make sense to adjust the size passed to
_scratch_mkfs_sized in generic/204 to be a something slightly larger,
since otherwise it's pretty much guaranteed that generic/204 will
start failing on PowerPC when testing with a 64k block size.

Cheers,

						- Ted