klibc - Jan 2023 - mips64el stat/time/…? problem

On Tue, 2023-01-03 at 21:58 +0100, Arnd Bergmann wrote:
> On Tue, Jan 3, 2023, at 17:47, Thorsten Glaser wrote:
> > Hi,
> > 
> > I noticed a failure of mksh built with klibc on mips64el.
> > 
> > The failing test, on a high level, is this:
> > 
> > 	:>a
> > 	sleep 2
> > 	:>b
> > 	test a -nt b
> > 	echo $?
> > 
> > This is supposed to echo 1 (false) because a is not newer than b.
[...]
> > Does this suffice to begin debugging?
> 
> I don't yet see what the problem is, but it's worth noting that
> mips64 is the only architecture using 'unsigned int' for the
> st_mtime field, while everything else uses 'long'. There is
> a good chance that the different type causes this problem.
Right, and klibc's definition of struct stat for mips64 seems to have
been copied from mips and edited without fixing that difference.

Thorsten: I've attached a patch which I tested briefly in QEMU.  Let me
know if this works for you.

Arnd: I haven't looked at Y2038-proofing klibc yet, but when I do: is
there a simple upgrade path for stat() or will I need to wrap statx()?

Ben.

-- 
Ben Hutchings
The Peter principle: In a hierarchy, every employee tends to rise to
their level of incompetence.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-klibc-mips64-Fix-struct-stat-layout.patch
Type: text/x-patch
Size: 1249 bytes
Desc: not available
URL:
<https://lists.zytor.com/archives/klibc/attachments/20230104/16e12172/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: This is a digitally signed message part
URL:
<https://lists.zytor.com/archives/klibc/attachments/20230104/16e12172/attachment.sig>

On Wed, Jan 4, 2023, at 01:47, Ben Hutchings wrote:
>
> Right, and klibc's definition of struct stat for mips64 seems to have
> been copied from mips and edited without fixing that difference.
Ok, that makes sense.
> Arnd: I haven't looked at Y2038-proofing klibc yet, but when I do: is
> there a simple upgrade path for stat() or will I need to wrap statx()?
You have to use statx() here, but the upside is that this can be
done in an architecture-independent way.

I have recently started revisiting a larger cleanup of the way we
generate the system call tables. It should be possible to use this
to generate the libc-side wrappers as well, which may help with the
y2038 conversion.

       Arnd

Hi Ben,
>Thorsten: I've attached a patch which I tested briefly in QEMU.  Let me
>know if this works for you.
sorry for the delay. I was able to test this on eller today.
To reproduce the issue, this was needed, because the value
read was a nanosecond field and so the test didn?t always fail:

(sid_mips64el-dchroot)tg at eller:~$ while sleep  0.1; do :>b; ./mksh -c
'test a -nt b; echo $?'; done

Unpatched, this produced runs of (4 to 7) zeros and (3 to 4) ones;
patched, it produced all ones.

I cannot add this variant to the upstream testsuite as a subsecond
sleep is not guaranteed, but I can add it to the Debian tests.

Thanks for fixing this; I?ll upload mksh again once this is in sid.
Which probably should be soon considering the freeze?

bye,
//mirabilos
-- 
?It is inappropriate to require that a time represented as
 seconds since the Epoch precisely represent the number of
 seconds between the referenced time and the Epoch.?
	-- IEEE Std 1003.1b-1993 (POSIX) Section B.2.2.2