Hello Arnd,

Arnd Bergmann [off-list ref] writes:

On Tue, Sep 21, 2021 at 3:01 PM Richard Palethorpe [off-list ref] wrote:

quoted

The LTP test io_pgetevents02 fails in 32bit compat mode because an
nr_max of -1 appears to be treated as a large positive integer. This
causes pgetevents_time64 to return an event. The test expects the call
to fail and errno to be set to EINVAL.

Using the compat syscall fixes the issue.

Fixes: 7a35397f8c06 ("io_pgetevents: use __kernel_timespec")
Signed-off-by: Richard Palethorpe <redacted>

Thanks a lot for finding this, indeed there is definitely a mistake that
this function is defined and not used, but I don't yet see how it would
get to the specific failure you report.

Between the two implementations, I can see a difference in the
handling of the signal mask, but that should only affect architectures
with incompatible compat_sigset_t, i.e. big-endian or
_COMPAT_NSIG_WORDS!=_NSIG_WORDS, and the latter is
never true for currently supported architectures. On x86, there is
no difference in the sigset at all.

The negative 'nr' and 'min_nr' arguments that you list as causing
the problem /should/ be converted by the magic
SYSCALL_DEFINE6() definition. If this is currently broken, I would
expect other syscalls to be affected as well.

That is what I thought, but I couldn't think of another explanation for
it.

Have you tried reproducing this on non-x86 architectures? If I
misremembered how the compat conversion in SYSCALL_DEFINE6()
works, then all architectures that support CONFIG_COMPAT have
to be fixed.

         Arnd

No, but I suppose I can try it on ARM or PowerPC. I suppose printing the
arguments would be a good idea too.

-- 
Thank you,
Richard.