Re: [PATCH v2 -tip] perf: x86, add SandyBridge support

[Stephane Eranian]

On Mon, Feb 28, 2011 at 9:51 AM, Lin Ming <ming.m.lin@xxxxxxxxx> wrote:
> On Mon, 2011-02-28 at 16:20 +0800, Stephane Eranian wrote:
>> On Mon, Feb 28, 2011 at 8:22 AM, Lin Ming <ming.m.lin@xxxxxxxxx> wrote:
>> > This patch adds basic SandyBridge support, including hardware cache
>> > events and PEBS events support.
>> >
>> > LLC-* hareware cache events don't work for now, it depends on the
>> > offcore patches.
>> >
>> > All PEBS events are tested on my SandyBridge machine and work well.
>> > Note that SandyBridge does not support INSTR_RETIRED.ANY(0x00c0) PEBS
>> > event, instead it supports INST_RETIRED.PRECDIST(0x01c0) event and PMC1
>> > only.
>> >
>> > v1 -> v2:
>> > - add more raw and PEBS events constraints
>> > - use offcore events for LLC-* cache events
>> > - remove the call to Nehalem workaround enable_all function
>> >
>> > todo:
>> > - precise store
>> > - precise distribution of instructions retired
>> >
>> > Signed-off-by: Lin Ming <ming.m.lin@xxxxxxxxx>
>> > ---
>> > Âarch/x86/kernel/cpu/perf_event.c     Â|  Â2 +
>> > Âarch/x86/kernel/cpu/perf_event_intel.c  Â| Â123 +++++++++++++++++++++++++++++
>> > Âarch/x86/kernel/cpu/perf_event_intel_ds.c | Â 44 ++++++++++-
>> > Â3 files changed, 168 insertions(+), 1 deletions(-)
>> >
>> > diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
>> > index 10bfe24..49d51be 100644
>> > --- a/arch/x86/kernel/cpu/perf_event.c
>> > +++ b/arch/x86/kernel/cpu/perf_event.c
>> > @@ -148,6 +148,8 @@ struct cpu_hw_events {
>> > Â*/
>> > Â#define INTEL_EVENT_CONSTRAINT(c, n) Â \
>> > Â Â Â ÂEVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
>> > +#define INTEL_EVENT_CONSTRAINT2(c, n) Â\
>> > + Â Â Â EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
>> >
>> > Â/*
>> > Â* Constraint on the Event code + UMask + fixed-mask
>> > diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
>> > index 084b383..3085868 100644
>> > --- a/arch/x86/kernel/cpu/perf_event_intel.c
>> > +++ b/arch/x86/kernel/cpu/perf_event_intel.c
>> > @@ -76,6 +76,19 @@ static struct event_constraint intel_westmere_event_constraints[] =
>> > Â Â Â ÂEVENT_CONSTRAINT_END
>> > Â};
>> >
>> > +static struct event_constraint intel_snb_event_constraints[] =
>> > +{
>> > + Â Â Â FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
>> > + Â Â Â FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
>> > + Â Â Â /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
>> > + Â Â Â INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
>> > + Â Â Â INTEL_EVENT_CONSTRAINT(0xb7, 0x1), /* OFF_CORE_RESPONSE_0 */
>> > + Â Â Â INTEL_EVENT_CONSTRAINT(0xbb, 0x8), /* OFF_CORE_RESPONSE_1 */
>> > + Â Â Â INTEL_EVENT_CONSTRAINT2(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
>> > + Â Â Â INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
>> > + Â Â Â EVENT_CONSTRAINT_END
>> > +};
>> > +
>> > Âstatic struct event_constraint intel_gen_event_constraints[] =
>> > Â{
>> > Â Â Â ÂFIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
>> > @@ -89,6 +102,106 @@ static u64 intel_pmu_event_map(int hw_event)
>> > Â Â Â Âreturn intel_perfmon_event_map[hw_event];
>> > Â}
>> >
>> > +static __initconst const u64 snb_hw_cache_event_ids
>> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â [PERF_COUNT_HW_CACHE_MAX]
>> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â [PERF_COUNT_HW_CACHE_OP_MAX]
>> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â [PERF_COUNT_HW_CACHE_RESULT_MAX] =
>> > +{
>> > + [ C(L1D) ] = {
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS Â Â Â Â*/
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0151, /* L1D.REPLACEMENT Â Â Â Â Â Â Â*/
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES Â Â Â */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0851, /* L1D.ALL_M_REPLACEMENT Â Â Â Â*/
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x0,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x024e, /* HW_PRE_REQ.DL1_MISS Â Â Â Â Â*/
>> > + Â Â Â },
>> > + },
>> > + [ C(L1I ) ] = {
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x0,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0280, /* ICACHE.MISSES */
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = -1,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = -1,
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x0,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0,
>> > + Â Â Â },
>> > + },
>> > + [ C(LL Â) ] = {
>> > + Â Â Â /*
>> > + Â Â Â Â* TBD: Need Off-core Response Performance Monitoring support
>> > + Â Â Â Â*/
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_0.ANY_DATA.LOCAL_CACHE */
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x01b7,
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_1.ANY_DATA.ANY_LLC_MISS */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x01bb,
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_0.ANY_RFO.LOCAL_CACHE */
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x01b7,
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_1.ANY_RFO.ANY_LLC_MISS */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x01bb,
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_0.PREFETCH.LOCAL_CACHE */
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x01b7,
>> > + Â Â Â Â Â Â Â /* OFFCORE_RESPONSE_1.PREFETCH.ANY_LLC_MISS */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x01bb,
>> > + Â Â Â },
>> > + },
>> > + [ C(DTLB) ] = {
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x01d0, /* MEM_UOP_RETIRED.LOADS */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x02d0, /* MEM_UOP_RETIRED.STORES */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x0,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0,
>> > + Â Â Â },
>> > + },
>> > + [ C(ITLB) ] = {
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT Â Â Â Â */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK Â Â*/
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = -1,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = -1,
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = -1,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = -1,
>> > + Â Â Â },
>> > + },
>> > + [ C(BPU ) ] = {
>> > + Â Â Â [ C(OP_READ) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_WRITE) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = -1,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = -1,
>> > + Â Â Â },
>> > + Â Â Â [ C(OP_PREFETCH) ] = {
>> > + Â Â Â Â Â Â Â [ C(RESULT_ACCESS) ] = -1,
>> > + Â Â Â Â Â Â Â [ C(RESULT_MISS) Â ] = -1,
>> > + Â Â Â },
>> > + },
>> > +};
>> > +
>> > Âstatic __initconst const u64 westmere_hw_cache_event_ids
>> > Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â[PERF_COUNT_HW_CACHE_MAX]
>> > Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â[PERF_COUNT_HW_CACHE_OP_MAX]
>> > @@ -1062,6 +1175,16 @@ static __init int intel_pmu_init(void)
>> > Â Â Â Â Â Â Â Âpr_cont("Westmere events, ");
>> > Â Â Â Â Â Â Â Âbreak;
>> >
>> > + Â Â Â case 42: /* SandyBridge */
>> > + Â Â Â Â Â Â Â memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
>> > + Â Â Â Â Â Â Â Â Â Â Âsizeof(hw_cache_event_ids));
>> > +
>> > + Â Â Â Â Â Â Â intel_pmu_lbr_init_nhm();
>> > +
>> > + Â Â Â Â Â Â Â x86_pmu.event_constraints = intel_snb_event_constraints;
>> > + Â Â Â Â Â Â Â pr_cont("SandyBridge events, ");
>> > + Â Â Â Â Â Â Â break;
>> > +
>> > Â Â Â Âdefault:
>> > Â Â Â Â Â Â Â Â/*
>> > Â Â Â Â Â Â Â Â * default constraints for v2 and up
>> > diff --git a/arch/x86/kernel/cpu/perf_event_intel_ds.c b/arch/x86/kernel/cpu/perf_event_intel_ds.c
>> > index b7dcd9f..e60f91b 100644
>> > --- a/arch/x86/kernel/cpu/perf_event_intel_ds.c
>> > +++ b/arch/x86/kernel/cpu/perf_event_intel_ds.c
>> > @@ -388,6 +388,42 @@ static struct event_constraint intel_nehalem_pebs_events[] = {
>> > Â Â Â ÂEVENT_CONSTRAINT_END
>> > Â};
>> >
>> > +static struct event_constraint intel_snb_pebs_events[] = {
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x02c2, 0xf), /* UOPS_RETIRED.RETIRE_SLOTS */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01c4, 0xf), /* BR_INST_RETIRED.CONDITIONAL */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x02c4, 0xf), /* BR_INST_RETIRED.NEAR_CALL */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x04c4, 0xf), /* BR_INST_RETIRED.ALL_BRANCHES */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x08c4, 0xf), /* BR_INST_RETIRED.NEAR_RETURN */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x10c4, 0xf), /* BR_INST_RETIRED.NOT_TAKEN */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x20c4, 0xf), /* BR_INST_RETIRED.NEAR_TAKEN */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x40c4, 0xf), /* BR_INST_RETIRED.FAR_BRANCH */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01c5, 0xf), /* BR_MISP_RETIRED.CONDITIONAL */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x04c5, 0xf), /* BR_MISP_RETIRED.ALL_BRANCHES */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x10c5, 0xf), /* BR_MISP_RETIRED.NOT_TAKEN */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x20c5, 0xf), /* BR_MISP_RETIRED.TAKEN */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORE */
>>
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x01d0, 0xf), /* MEM_UOP_RETIRED.LOADS */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x02d0, 0xf), /* MEM_UOP_RETIRED.STORES */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x10d0, 0xf), /* MEM_UOP_RETIRED.STLB_MISS */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x20d0, 0xf), /* MEM_UOP_RETIRED.LOCK */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x40d0, 0xf), /* MEM_UOP_RETIRED.SPLIT */
>> > + Â Â Â PEBS_EVENT_CONSTRAINT(0x80d0, 0xf), /* MEM_UOP_RETIRED.ALL */
>>
>> Not quite. For event 0xd0, you are not listing the right umask combinations.
>> The following combinations are supported for event 0xd0:
>>
>> 0x5381d0 Â Â Âsnb::MEM_UOP_RETIRED:ANY_LOADS
>> 0x5382d0 Â Â Âsnb::MEM_UOP_RETIRED:ANY_STORES
>> 0x5321d0 Â Â Âsnb::MEM_UOP_RETIRED:LOCK_LOADS
>> 0x5322d0 Â Â Âsnb::MEM_UOP_RETIRED:LOCK_STORES
>> 0x5341d0 Â Â Âsnb::MEM_UOP_RETIRED:SPLIT_LOADS
>> 0x5342d0 Â Â Âsnb::MEM_UOP_RETIRED:SPLIT_STORES
>> 0x5311d0 Â Â Âsnb::MEM_UOP_RETIRED:STLB_MISS_LOADS
>> 0x5312d0 Â Â Âsnb::MEM_UOP_RETIRED:STLB_MISS_STORES
>>
>> In other words, bit 0-3 of the umask cannot be zero.
>
> I got the umask from "Table 30-20. PEBS Performance Events for Intel
> microarchitecture code name Sandy Bridge".
>
> But from "Table A-2. Non-Architectural Performance Events In the
> Processor Core for Intel Core Processor 2xxx Series", the combinations
> are needed as you show above.
>
> Which one is correct?
>
I think Table A-2 is correct. Umasks 10h, 20h, 40h, 80h MUST be combined
to collect something meaningful.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/