Re: [PATCH v2 2/2] perf bench: Add support for 32-bit systems with 64-bit time_t

From: Arnd Bergmann
Date: Fri Sep 10 2021 - 03:18:03 EST


On Fri, Sep 10, 2021 at 6:31 AM Alistair Francis
<alistair.francis@xxxxxxxxxxxxxxxxxx> wrote:
>
> From: Alistair Francis <alistair.francis@xxxxxxx>
>
> Some 32-bit architectures (such are 32-bit RISC-V) only have a 64-bit
> time_t and as such don't have the SYS_futex syscall. This patch will
> allow us to use the SYS_futex_time64 syscall on those platforms.
>
> Signed-off-by: Alistair Francis <alistair.francis@xxxxxxx>
> ---
> tools/perf/bench/futex.h | 43 +++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 42 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/bench/futex.h b/tools/perf/bench/futex.h
> index f80a4759ee79b..09c5596726c60 100644
> --- a/tools/perf/bench/futex.h
> +++ b/tools/perf/bench/futex.h
> @@ -27,6 +27,12 @@ struct bench_futex_parameters {
> unsigned int nrequeue;
> };
>
> +/* A version of 'struct timespec' with 32-bit time_t and nanoseconds. */
> +struct __timespec32 {
> + __kernel_long_t tv_sec;
> + __kernel_long_t tv_nsec;
> +};

Why not use the existing __kernel_old_timespec structure definition
from the header?

> +#if defined(SYS_futex)
> + if (sizeof(*timeout) == sizeof(struct __timespec32))
> + return syscall(SYS_futex, uaddr, op | opflags, val, timeout, uaddr2, val3);
> +
> + if (timeout && timeout->tv_sec == (long)timeout->tv_sec) {
> + struct __timespec32 ts32;
> +
> + ts32.tv_sec = (__kernel_long_t) timeout->tv_sec;
> + ts32.tv_nsec = (__kernel_long_t) timeout->tv_nsec;
> +
> + return syscall(SYS_futex, uaddr, op | opflags, val, ts32, uaddr2, val3);
> + } else if (!timeout) {
> + return syscall(SYS_futex, uaddr, op | opflags, val, NULL, uaddr2, val3);
> + }
> +#endif
> +
> + errno = ENOSYS;
> + return -1;

Similarly, I think just using the kernel's constants like __NR_futex instead of
SYS_futex would be more appropriate here, this way you know if futex_time64
is available and don't have to do the emulation for running the binary
on pre-5.1
kernels. Perf already uses __NR_perf_event_open, __NR_getcpu, and
__NR_setns instead of the SYS_* versions.

Arnd