Re: [PATCH v3 2/2] perf bench: Add support for 32-bit systems with 64-bit time_t

[André Almeida]

Às 05:08 de 21/09/21, Arnd Bergmann escreveu:
> On Tue, Sep 21, 2021 at 12:47 AM André Almeida
> <andrealmeid@xxxxxxxxxxxxx> wrote:
>>
>> #if defined(__i386__) || __TIMESIZE == 32
>> # define NR_gettime64 __NR_clock_gettime64
>> #else
>> # define NR_gettime64 __NR_clock_gettime
>> #endif
>>
>> struct timespec64 {
>>         long long tv_sec;       /* seconds */
>>         long long tv_nsec;      /* nanoseconds */
>> };
>>
>> int gettime64(clock_t clockid, struct timespec64 *tv)
>> {
>>         return syscall(NR_gettime64, clockid, tv);
>> }
>>
>> Then we can just use &timeout at __NR_futex_time64 for 32bit arch and at
>> __NR_futex for 64bit arch.
> 
> This is still broken when you disable CONFIG_COMPAT_32BIT_TIME,
> which disables all system calls that take time32 arguments.
> 

Oh, I think my point was confusing then. My suggestion was to use only
the futex entry points that accepts time64, and to always use
clock_gettime that uses time64, for all platforms. Then it will work if
we disable CONFIG_COMPAT_32BIT_TIME.

>> This might be a simpler solution to the problem that you are facing but
>> I'm not entirely sure. Also, futex's selftests do use the timeout
>> argument and I think that they also won't compile in 32-bit RISC-V, so
>> maybe we can start from there so we can actually test the timeout
>> argument and check if it's working.
> 
> I would love to see the wrapper that Alistair wrote as part of some kernel
> uapi header provided to user space. futex is used by tons of applications,
> and we never had a library abstraction for it, so everyone has to do these
> by hand, and they all get them slightly wrong in different ways.

Why we don't have a futex() wrapper at glibc as we do have for others
syscalls?

> 
> We normally don't do this in kernel headers, but I think the benefits
> would be far greater compared to today's situation.
> 
>       Arnd
>