Re: perf tools build broken for RISCV 32 bit

[Emiliano Ingrassia]

Hi Arnd,

thank you for the quick reply and support.

On Thu, Jan 14, 2021 at 08:16:28PM +0100, Arnd Bergmann wrote:
> On Thu, Jan 14, 2021 at 7:38 PM Emiliano Ingrassia
> <ingrassia@xxxxxxxxxxxxxx> wrote:
> >
> > Hi,
> >
> > When building perf for RISCV 32 bit (v5.10.7) I got the following
> >
> > | In file included from bench/futex-hash.c:29:
> > | bench/futex.h: In function ‘futex_wait’:
> > | bench/futex.h:37:10: error: ‘SYS_futex’ undeclared (first use in this function); did you mean ‘SYS_tee’?
> >
> > This issue is similar to the one reported in https://lkml.org/lkml/2019/4/19/631
> >
> > I found that patching tools/arch/riscv/include/uapi/asm/unistd.h as following:
> >
> >  #ifdef __LP64__
> >  #define __ARCH_WANT_NEW_STAT
> >  #define __ARCH_WANT_SET_GET_RLIMIT
> > +#else
> > +#define __ARCH_WANT_TIME32_SYSCALLS
> >  #endif /* __LP64__ */
> >
> > solved the problem.
> >
> > I also found that a similar patch for arch/riscv/include/uapi/asm/unistd.h
> > was removed in commit d4c08b9776b3, so probably this is not the right way(?).
>
> In short, it won't work, as rv32 does not provide the time32 syscalls.
> Your patch will make the application build, but it will not be able to
> call futex().
>
> You will in fact run into a related problem on any 32-bit architecture
> if CONFIG_COMPAT_32BIT_TIME is disabled, or if you pass a non-NULL
> timeout parameter and build with a time64-enabled libc.
>

I'm using glibc 2.32 which supports 64 bit time_t on RISCV 32.

In particular, searching for __NR_futex in glibc RISCV source code I got:

sysdeps/unix/sysv/linux/riscv/rv32/arch-syscall.h: #define __NR_futex_time64 422
sysdeps/unix/sysv/linux/riscv/sysdep.h: #define __NR_futex __NR_futex_time64
sysdeps/unix/sysv/linux/riscv/rv64/arch-syscall.h: #define __NR_futex 98

but in the generated bits/syscall.h, included in bench/futex.h, I found:

#ifdef __NR_futex
# define SYS_futex __NR_futex
#endif

#ifdef __NR_futex_time64
# define SYS_futex_time64 __NR_futex_time64
#endif

So the problem is that userspace applications can't see the definition
of __NR_futex which is in sysdep.h, but there are no problems in calling
futex() libc wrapper because glibc syscall.c includes that header.

A possible fix for the perf tool would be to use futex() libc wrapper
in tools/perf/bench/futex.h instead of syscall(), but, if I understand correctly,
there are some drawbacks that does not permit it.

So what should be the right solution?

> The fix in the application is to call either __NR_futex or __NR_futex64
> depending on the definition of time_t in the C library. I would recommend
> doing it like
>
> #ifdef __NR_futex
> #define do_futex (sizeof(time_t) == sizeof(__kernel_long_t)) ? \
>          __NR_futex : __NR_futex_time64
> #else
> #define do_futex __NR_futex
> #done
>
>        Arnd

Where should be this fix applied? To perf code?

Thank you,

Emiliano