Re: [PATCH] tst: Add test for utime

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 08/03/2021 10:46, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> On 05/03/2021 17:29, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>>> Maybe fix the kernel or/and use a non buggy FS instead? IMHO the test is
>>>> doing *exactly* is meant to do: trigger and expose an unexpected return
>>>> from the kernel or the libc. I don't think hiding it using a different
>>>> FS is the right thing to do on glibc testsuite.
>>>
>>> I disagree: We want to test glibc here, not file systems.  Test outputs
>>> are already difficult enough to interpret.  We should not add known
>>> FAILs that are very difficult to fix (reformat & reinstall with another
>>> file system is definitely in that category).
>>
>> My take is since we are the first step to drive the proper y2038 support
>> and it has time limit, it is better to fail hard and signal something
>> is wrong either with glibc or kernel *now* so we iron out the issues
>> instead of fail once users actually deploy glibc on system that are meant
>> to handle y2038 issues.
> 
> We know precisely what is wrong with XFS (and the other file systems
> which need to change their disk layout).  No one gains from testing this
> during glibc builds.

No, but indicates that your system need to be fixed if you intend to
deploy a y2038 binaries on it.

> 
> Again, I do not suggest not to test this, but rather use a path which is
> required to use a specific file system (tmpfs), so that the test outcome
> is more predictable.  The glibc side and the higher VFS layers in the
> kernel are still fully covered if we use tmpfs, so I do not think this
> reduces coverage as far as glibc is concerned.

What if kernel introduce a regression on tmpfs or any other filesystem,
should we gloss over to use a different filesystem instead? Or should we
add a routine to check for the filesystem support for y2038 and disable
or return unsupported instead?

Also, this issue will most likely get fixed way before y2038 will actually
be deployed. So I am not seeing the urgency here of avoid this kind of 
transient failures.