lkml.org 
[lkml]   [2011]   [Nov]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCHv2 6/9] perf: expose perf capability to other modules.
On Tue, Nov 08, 2011 at 02:26:51PM +0100, Peter Zijlstra wrote:
> On Tue, 2011-11-08 at 14:49 +0200, Gleb Natapov wrote:
> > > It might make sense to introduce cpuid10_ebx or so, also I think the
> > cpuid10_ebx will have only one field though (event_mask).
> >
> > > At the very least add a full ebx iteration to disable unsupported events
> > > in the intel-v1 case.
> > I do not understand what do you mean here, cpuid10_ebx was introduced by
> > intel v1 architectural PMU so it should already contain correct information.
>
> I meant something like the below
>
Isn't it better to introduce mapping between ebx bits and architectural
events and do for_each_set_bit loop? But I wouldn't want to introduce
patch as below as part of this series. I do not want to introduce
incidental regressions. For instance the patch below will introduce
regression on my Nehalem cpu. It reports value 0x44 in cpuid10.ebx which
means that unhalted_reference_cycles is not available (bit set means
event is not available), but event still works! Actually it is listed as
supported by the cpu in Table A-4 SDM 3B. Go figure.

> ---
> arch/x86/include/asm/perf_event.h | 13 +++++++++++++
> arch/x86/kernel/cpu/perf_event.c | 3 +++
> arch/x86/kernel/cpu/perf_event_intel.c | 21 ++++++++++++++++++---
> 3 files changed, 34 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
> index f61c62f..98e397a 100644
> --- a/arch/x86/include/asm/perf_event.h
> +++ b/arch/x86/include/asm/perf_event.h
> @@ -72,6 +72,19 @@ union cpuid10_eax {
> unsigned int full;
> };
>
> +union cpuid10_ebx {
> + struct {
> + unsigned int unhalted_core_cycles:1;
> + unsigned int instructions_retired:1;
> + unsigned int unhalted_reference_cycles:1;
> + unsigned int llc_reference:1;
> + unsigned int llc_misses:1;
> + unsigned int branch_instruction_retired:1;
> + unsigned int branch_misses_retired:1;
> + } split;
> + unsigned int full;
> +};
> +
> union cpuid10_edx {
> struct {
> unsigned int num_counters_fixed:5;
> diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
> index 6408910..e4fdb9d 100644
> --- a/arch/x86/kernel/cpu/perf_event.c
> +++ b/arch/x86/kernel/cpu/perf_event.c
> @@ -336,6 +336,9 @@ int x86_setup_perfctr(struct perf_event *event)
> if (config == -1LL)
> return -EINVAL;
>
> + if (config == -2LL)
> + return -EOPNOTSUPP;
> +
> /*
> * Branch tracing:
> */
> diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
> index e09ca20..aaaed9a 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel.c
> @@ -1547,9 +1547,9 @@ static void intel_clovertown_quirks(void)
> __init int intel_pmu_init(void)
> {
> union cpuid10_edx edx;
> + union cpuid10_ebx ebx;
> union cpuid10_eax eax;
> unsigned int unused;
> - unsigned int ebx;
> int version;
>
> if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
> @@ -1566,7 +1566,7 @@ __init int intel_pmu_init(void)
> * Check whether the Architectural PerfMon supports
> * Branch Misses Retired hw_event or not.
> */
> - cpuid(10, &eax.full, &ebx, &unused, &edx.full);
> + cpuid(10, &eax.full, &ebx.full, &unused, &edx.full);
> if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
> return -ENODEV;
>
> @@ -1598,6 +1598,21 @@ __init int intel_pmu_init(void)
> x86_pmu.intel_cap.capabilities = capabilities;
> }
>
> + if (!ebx.split.unhalted_core_cycles)
0 means event is available 1 it is no.

> + intel_perfmon_event_map[PERF_COUNT_HW_CPU_CYCLES] = -2;
> + if (!ebx.split.instructions_retired)
> + intel_perfmon_event_map[PERF_COUNT_HW_INSTRUCTIONS] = -2;
> + if (!ebx.split.unhalted_reference_cycles)
> + intel_perfmon_event_map[PERF_COUNT_HW_BUS_CYCLES] = -2;
> + if (!ebx.split.llc_reference)
> + intel_perfmon_event_map[PERF_COUNT_HW_CACHE_REFERENCES] = -2;
> + if (!ebx.split.llc_misses)
> + intel_perfmon_event_map[PERF_COUNT_HW_CACHE_MISSES] = -2;
> + if (!ebx.split.branch_instruction_retired)
> + intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = -2;
> + if (!ebx.split.branch_misses_retired)
> + intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = -2;
> +
> intel_ds_init();
>
> /*
> @@ -1643,7 +1658,7 @@ __init int intel_pmu_init(void)
> /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
> intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
>
> - if (ebx & 0x40) {
> + if (ebx.split.branch_misses_retired) {
> /*
> * Erratum AAJ80 detected, we work it around by using
> * the BR_MISP_EXEC.ANY event. This will over-count

--
Gleb.


\
 
 \ /
  Last update: 2011-11-08 14:57    [from the cache]
©2003-2011 Jasper Spaans