Re: [PATCH 01/16] perf: Unified API to record selective sets of arch registers

[Jiri Olsa <jolsa@redhat.com>]

On Thu, May 24, 2012 at 02:13:19PM +0200, Stephane Eranian wrote:
> 2012/5/24 Peter Zijlstra <a.p.zijlstra@chello.nl>:
> > On Wed, 2012-05-23 at 21:32 +0200, Jiri Olsa wrote:
> >> +enum perf_event_x86_32_regs {
> >> +       PERF_X86_32_REG_EAX,
> >> +       PERF_X86_32_REG_EBX,
> >> +       PERF_X86_32_REG_ECX,
> >> +       PERF_X86_32_REG_EDX,
> >> +       PERF_X86_32_REG_ESI,
> >> +       PERF_X86_32_REG_EDI,
> >> +       PERF_X86_32_REG_EBP,
> >> +       PERF_X86_32_REG_ESP,
> >> +       PERF_X86_32_REG_EIP,
> >> +       PERF_X86_32_REG_FLAGS,
> >> +       PERF_X86_32_REG_CS,
> >> +       PERF_X86_32_REG_DS,
> >> +       PERF_X86_32_REG_ES,
> >> +       PERF_X86_32_REG_FS,
> >> +       PERF_X86_32_REG_GS,
> >> +
> >> +       /* Non ABI */
> >> +       PERF_X86_32_REG_MAX,
> >> +};
> >
> >> +enum perf_event_x86_64_regs {
> >> +       PERF_X86_64_REG_RAX,
> >> +       PERF_X86_64_REG_RBX,
> >> +       PERF_X86_64_REG_RCX,
> >> +       PERF_X86_64_REG_RDX,
> >> +       PERF_X86_64_REG_RSI,
> >> +       PERF_X86_64_REG_RDI,
> >> +       PERF_X86_64_REG_R8,
> >> +       PERF_X86_64_REG_R9,
> >> +       PERF_X86_64_REG_R10,
> >> +       PERF_X86_64_REG_R11,
> >> +       PERF_X86_64_REG_R12,
> >> +       PERF_X86_64_REG_R13,
> >> +       PERF_X86_64_REG_R14,
> >> +       PERF_X86_64_REG_R15,
> >> +       PERF_X86_64_REG_RBP,
> >> +       PERF_X86_64_REG_RSP,
> >> +       PERF_X86_64_REG_RIP,
> >> +       PERF_X86_64_REG_FLAGS,
> >> +       PERF_X86_64_REG_CS,
> >> +       PERF_X86_64_REG_SS,
> >> +
> >> +       /* Non ABI */
> >> +       PERF_X86_64_REG_MAX,
> >> +};
> >
> > So why bother with two different APIs? Why not have only the x86_64 one
> > and maybe add a warning/error/fail whatever when x86_64 regs are
> > selected on i386/compat tasks.
> >
> >
> I agree with Peter here especially after our earlier discussion about how
> to handle 32-bit ABI threads on a 64-bit ABI kernel. You should simply
> fill in the record with zeroes when the register does not exists. You cannot
> predict what will be sampled in system-wide mode.
> 
> The only sanity check you can do is on 32-bit ABI kernel, reject any 64-bit
> ABI only regs but then that would require invoking arch specific code from
> perf_copy_attr()....

ok, looks like this could work.. and fix the confusion for compat tasks