Haswell microarchitecture don't have Stalled-cycles-backend in perf

Question

I installed perf on Haswell CPU( Intel Core i7-4790 ). But the "perf list" does not include "stalled-cycles-frontend" nor "stalled-cycles-backend". I checked the http://www.intel.com/content/www/us/en/processors/architectures-software-developer-manuals.html and not found the performance events relevant to stalled-cycles-backend from the Table 19-7( Non-Architectural Performance Events In the Processor Core of 4th Generation Intel Core Processors).

So my question is: how can I measure stalled-cycles-backend using perf or other tools in Haswell CPU cores. The kernel is 3.19 and perf version is also 3.19.

Thanks

Auroraauroral

Did you try ocperf.py? It's a wrapper with more hw-specific event names. Does perf stat work normally? It includes those counters by default. – Witchcraft

@PeterCordes Thanks to your reply, I tried ocperf and I found the so called "cycle_activity.cycles_no_execute", "uops_executed.stall_cycles" and "uops_retired.stall_cycles". But I'm not sure those are "stalled-cycles-backend". – Auroraauroral

hank wing, probably, authors of perf_events subsystem mapping were unsure too and just did not map this (badly designed and named) "stalled-cycles-backend" into some single event on haswell? (The code, there is mapping of PERF_COUNT_HW_STALLED_CYCLES_BACKEND for Westmere WSM, Sandy SNB; not for Ivy, Haswell, Broadwell, and newer). Authors of toplev.py or PAPI may map complex events, they may use formulas to compute them from several. – Bodhisattva

Answer 1

Yes, there is no mapping of "stalled-cycles-frontend" and "stalled-cycles-backend" synthetic events in perf_events subsystem in kernel for newer processors like Ivy Bridge or Haswell. And no mapping on older Core 2. Probably, this name/concept/idea is not good for changed and complex microarchitectures of modern Out-of-order CPUs without simple scalar measurement of global "Stall".

The code is in arch/x86/events/intel/core.c, and the synthetic event names are PERF_COUNT_HW_STALLED_CYCLES_FRONTEND and PERF_COUNT_HW_STALLED_CYCLES_BACKEND:

__init int intel_pmu_init(void)
{...

Both are defined since Nehalem, for Westmere, Sandy Bridge:

    case INTEL_FAM6_NEHALEM:
    case INTEL_FAM6_NEHALEM_EP:
    case INTEL_FAM6_NEHALEM_EX:

        /* UOPS_ISSUED.STALLED_CYCLES */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
            X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
        /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
            X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);

    case INTEL_FAM6_WESTMERE:
    case INTEL_FAM6_WESTMERE_EP:
    case INTEL_FAM6_WESTMERE_EX:

        /* UOPS_ISSUED.STALLED_CYCLES */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
            X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
        /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
            X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1);


    case INTEL_FAM6_SANDYBRIDGE:
    case INTEL_FAM6_SANDYBRIDGE_X:


        /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
            X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);
        /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =
            X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1);

Only frontend stall is defined for Ivy Bridge

    case INTEL_FAM6_IVYBRIDGE:
    case INTEL_FAM6_IVYBRIDGE_X:

        /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
        intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =
            X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1);

No mapping for frontend and for backend stalls for more recent CPUs desktop (Haswell, Broadwell, Skylake, Kaby Lake) and Phi (KNL, KNM):

    case INTEL_FAM6_HASWELL_CORE:
    case INTEL_FAM6_HASWELL_X:
    case INTEL_FAM6_HASWELL_ULT:
    case INTEL_FAM6_HASWELL_GT3E:

    case INTEL_FAM6_BROADWELL_CORE:
    case INTEL_FAM6_BROADWELL_XEON_D:
    case INTEL_FAM6_BROADWELL_GT3E:
    case INTEL_FAM6_BROADWELL_X:


    case INTEL_FAM6_XEON_PHI_KNL:
    case INTEL_FAM6_XEON_PHI_KNM:


    case INTEL_FAM6_SKYLAKE_MOBILE:
    case INTEL_FAM6_SKYLAKE_DESKTOP:
    case INTEL_FAM6_SKYLAKE_X:
    case INTEL_FAM6_KABYLAKE_MOBILE:
    case INTEL_FAM6_KABYLAKE_DESKTOP:

Not defined for old Core2 too (did not check Atoms):

http://elixir.free-electrons.com/linux/v4.11/source/arch/x86/events/intel/core.c#L27

static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
{
    [PERF_COUNT_HW_CPU_CYCLES]      = 0x003c,
    [PERF_COUNT_HW_INSTRUCTIONS]        = 0x00c0,
    [PERF_COUNT_HW_CACHE_REFERENCES]    = 0x4f2e,
    [PERF_COUNT_HW_CACHE_MISSES]        = 0x412e,
    [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4,
    [PERF_COUNT_HW_BRANCH_MISSES]       = 0x00c5,
    [PERF_COUNT_HW_BUS_CYCLES]      = 0x013c,
    [PERF_COUNT_HW_REF_CPU_CYCLES]      = 0x0300, /* pseudo-encoding */
};

Bodhisattva