On Tue, Nov 9, 2021 at 1:52 PM André Almeida <andrealmeid@xxxxxxxxxxxxx> wrote:
> Às 08:18 de 09/11/21, Vasily Gorbik escreveu:
> > On Thu, Sep 23, 2021 at 02:11:09PM -0300, André Almeida wrote:
> >> Test if the futex_waitv timeout is working as expected, using the
> >> supported clockid options.
> >
> >> + /* futex_waitv with CLOCK_MONOTONIC */
> >> + if (futex_get_abs_timeout(CLOCK_MONOTONIC, &to, timeout_ns))
> >> + return RET_FAIL;
> >> + res = futex_waitv(&waitv, 1, 0, &to, CLOCK_MONOTONIC);
> >> + test_timeout(res, &ret, "futex_waitv monotonic", ETIMEDOUT);
> >> +
> >> + /* futex_waitv with CLOCK_REALTIME */
> >> + if (futex_get_abs_timeout(CLOCK_REALTIME, &to, timeout_ns))
> >> + return RET_FAIL;
> >> + res = futex_waitv(&waitv, 1, 0, &to, CLOCK_REALTIME);
> >> + test_timeout(res, &ret, "futex_waitv realtime", ETIMEDOUT);
> >
> > Hi André,
> >
> > when built with -m32 and run as compat this two futex_waitv calls hang
> > on x86 and s390 (noticed while wiring up futex_waitv). The rest of the
> > futex selftests pass. This suggests some common compat issue? Any ideas?
>
> The issue is that futex_waitv() only accepts struct timespec that uses
> 64bit members. When using -m32, glibc will give you a 32bit timespec,
> thus the timeout won't wort. Someday glibc will provide 64bit timespec
> to 32bit userspace, given that this is affected by y2038 bug.
I think in the latest glibc you should be able to pass -D_TIME_BITS=64 to
the compiler to get the correct definition. Unfortunately, this only works
for simple test cases, but breaks if you call any interfaces from another
(non-glibc) library that depend on a particular time_t definition.
Alistair Francis also recently posted a set of helpers for the old futex()
call to make that easier to use from applications regardless of the libc
interface. I think it would be good to have this for futex_waitv() as well.
Arnd
