[PATCH v2 3/7] perf syscalltbl: Remove struct syscalltbl

[Ian Rogers]

On Tue, Feb 11, 2025 at 8:34 AM Arnd Bergmann <arnd at arndb.de> wrote:
>
> On Tue, Feb 11, 2025, at 17:18, Ian Rogers wrote:
> > On Mon, Feb 10, 2025 at 11:48 PM Arnd Bergmann <arnd at arndb.de> wrote:
>
> >> The syscall numbers on mips (and previously on ia64) are offset by
> >> a large number depending on the ABI (o32/n32/n64). I assume what
> >> we want here is to have the small numbers without the offset in
> >> syscall_num_to_name[], but that requires adding the offset during
> >> the lookup. Can you check if this is handled correctly?
> >
> > Thanks Arnd! I agree the tables are large and can be sparse, they'll
> > also be full of relocations. MIPS doesn't look like an outlier to me
> > here:
>
> Sorry, I should have been clearer what I meant, see
> arch/mips/include/uapi/asm/unistd.h:
>
> #if _MIPS_SIM == _MIPS_SIM_NABI32
> #define __NR_Linux      6000
> #include <asm/unistd_n32.h>
> #endif
>
> and
> arch/mips/include/generated/uapi/asm/unistd_n32.h
>
> #define __NR_read (__NR_Linux + 0)
> #define __NR_write (__NR_Linux + 1)
> #define __NR_open (__NR_Linux + 2)
>
> These offsets are 4000/5000/6000 respectively.
>
> > ```
> > #if defined(ALL_SYSCALLTBL) || defined(__mips__)
> > static const char *const syscall_num_to_name_EM_MIPS[] = {
> >        [0] = "read",
> >        [1] = "write",
> >        [2] = "open",
> > ...
> >         [465] = "listxattrat",
> >         [466] = "removexattrat",
> > };
>
> This means the array is not sparse, but the numbers
> here do not match the syscall number argument register.
>
> The question is whether tracing on mips adds the same
> per-ABI offset again, or if it tries and fails to look
> up index 6002 for 'open'.

Thanks for clarifying. I believe it will use 6002 and be broken. I
believe that'd be true without these changes too. I'm not testing on
MIPS so it'd be nice to have a fix targeted at making it work. Ideally
we wouldn't use __NR_Linux and instead fudge the id/system call number
based on e_machine/e_flags. I think it is follow on work, especially
because I don't find MIPS a very motivating use-case.

Thanks,
Ian