Re: [PATCH] xfs: always free inline data before resetting inode fork during ifree

[Sasha Levin]

On Fri, Mar 30, 2018 at 07:49:46PM +0000, Luis R. Rodriguez wrote:
>On Fri, Mar 30, 2018 at 02:47:05AM +0000, Sasha Levin wrote:
>> On Thu, Mar 29, 2018 at 10:05:35AM +1100, Dave Chinner wrote:
>> >On Wed, Mar 28, 2018 at 07:30:06PM +0000, Sasha Levin wrote:
>> >"./check -g auto" runs the full "expected to pass" regression test
>> >suite for all configured test configurations. (i.e. all config
>> >sections listed in the configs/<host>.config file)
>>
>> Great! With information from Darrick and yourself I've modified tests to
>> be more relevant. Right now I run 4 configs for each stable kernel, but
>> can add more or remove any - depends on what helps people analyse the
>> results.
>>
>> This brings me to the sad part of this mail: not a single stable kernel
>> survived a run. Most are paniced, some are hanging,
>
>I expected this. The semantics over -g auto yielding "expected to pass"
>are relative. Perhaps its better described as "should pass"?
>
>> and some were killed
>> because of KASan.
>>
>> All have hit various warnings in fs/iomap.c, and kernels accross several
>> versions hit the BUG at fs/xfs/xfs_message.c:113 (+-1 line)
>>
>> 4.15.12 is hitting a use-after-free in xfs_efi_release().
>> 4.14.29 and 4.9.89 seems to end up with corrupted memory (KASAN
>> warnings) at or before generic/027.
>> And finally, 3.18.101 is pretty unhappy with sleeping functions called
>> from atomic context.
>
>From my limited experience you have no option but to create an expunge list for
>each failure for now, and then pass the expunge lists -- that in essence would
>define the stable baseline and you should expect this to be different per
>kernel release. If you upgrade tooling, it can also change the results, and
>likewise if you upgrade fstests.
>
>If you define an expunge list you can then pass the list with the -E parameter,
>you can for instance categorize then failures by type and use a file for each
>type of failure, whether that's a triage list or a type of common failure.
>Format can be:
>
>test # comments are ignored
>
>Since you may want to database this somehow, perhaps a format that lists
>some tracking for it or other heuristics:
>
>generic/388 # bug#12345 - 1/300 run fails
>
>I'd recommend to just add all failures to one large expunge list for now,
>and later you can split / sort them them as needed.
>
>The idea later is that any failure later would be a regression. What would
>be good is to test a stable kernel prior to the auto-selection and use that
>as baseline, then bump the kernel and ensure no regressions were created.
>
>A dicey corner issue is that of tests which are supposed to "pass" but yet
>can fail every blue moon. For instance I've been running into one-off failures
>with generic/388 -- but only if I run it over 300 times.
>
>As such the baseline IMHO should also track these as just failures, however it
>will be often picked up as a regression first. The only way to rule this out
>is to loop test the same test prior to a kernel update and ensure it wasn't
>a regression -- ie, that it *was* still failing before.

Thanks for the pointers!

>This is why all this work is rather full time'ish. There is no way around it,
>it will take time to establish a baseline from fstests for a filesystem. There
>will also be a lot of odd ins and outs of each filesystem.

Right, but the way I see it, no one actually uses upstream. If anything,
it's a development branch, and the "real" users pick up one of the
stable trees to work with. So while there seems to be a lot of effort
dedicated to new features or fixing upstream bugs, not enough people
care that no one won't see those fixes for a few years.


--
Thanks,
Sasha