perf: Add Tremont support
Subsystems (2)
cover perf: Add Tremont support
From: Kan Liang <[email protected]>
The patch series intends to add Tremont support for Linux perf.
The patch series is on top of Icelake V5 patch series (with Peter's cleanup patch).
https://lkml.org/lkml/2019/4/8/630
PATCH 1: The feature is for both Icelake and Tremont. It missed the
Icelake patch series.
PATCH 2: Tremont core PMU support.
Kan Liang (2):
perf/x86/intel: Support adaptive PEBS for fixed counters perf/x86/intel: Add Tremont core PMU support
arch/x86/events/intel/core.c | 93 +++++++++++++++++++++++++++++++++++++++ arch/x86/include/asm/perf_event.h | 1 + 2 files changed, 94 insertions(+)
Signature · 1 line
2.7.4
PATCH 1/2 perf/x86/intel: Support adaptive PEBS for fixed counters
From: Kan Liang <[email protected]>
Fixed counters can also generate adaptive PEBS record, if the corresponding bit in IA32_FIXED_CTR_CTRL is set. Otherwise, only basic record is generated.
Unconditionally set the bit when PEBS is enabled on fixed counters. Let MSR_PEBS_CFG decide which format of PEBS record should be generated. There is no harmful to leave the bit set.
--- arch/x86/events/intel/core.c | 5 +++++ arch/x86/include/asm/perf_event.h | 1 + 2 files changed, 6 insertions(+)
arch/x86/events/intel/core.c +5 -0
@@ -2174,6 +2174,11 @@ static void intel_pmu_enable_fixed(struct perf_event *event) bits <<= (idx *4 ); mask =0xfULL << (idx *4 ); +if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { + bits |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + mask |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + } +rdmsrl (hwc->config_base, ctrl_val); ctrl_val &= ~mask; ctrl_val |= bits;
arch/x86/include/asm/perf_event.h +1 -0
@@ -33,6 +33,7 @@#define HSW_IN_TX (1ULL << 32)#define HSW_IN_TX_CHECKPOINTED (1ULL << 33)#define ICL_EVENTSEL_ADAPTIVE (1ULL << 34) +#define ICL_FIXED_0_ADAPTIVE (1ULL << 32)#define AMD64_EVENTSEL_INT_CORE_ENABLE (1ULL << 36)#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40)
Signature · 1 line
2.7.4
Peter Zijlstra
maintainerRe: [PATCH 1/2] perf/x86/intel: Support adaptive PEBS for fixed counters
Quoted 9 lines · On Tue, Apr 09, 2019 at 06:09:59PM -0700, [email protected] wrote:
From: Kan Liang <[email protected]>
Fixed counters can also generate adaptive PEBS record, if the corresponding bit in
IA32_FIXED_CTR_CTRLis set. Otherwise, only basic record is generated.Unconditionally set the bit when PEBS is enabled on fixed counters. Let
MSR_PEBS_CFGdecide which format of PEBS record should be generated. There is no harmful to leave the bit set.
I'll merge this back into:
Subject: perf/x86/intel: Support adaptive PEBSv4
such that this bug never existed, ok?
Quoted 38 lines
Signed-off-by Kan Liang <[email protected]>--- arch/x86/events/intel/core.c | 5 +++++ arch/x86/include/asm/perf_event.h | 1 + 2 files changed, 6 insertions(+)arch/x86/events/intel/core.c +5
-0@@ -2174,6 +2174,11 @@ static void intel_pmu_enable_fixed(struct perf_event *event) bits <<= (idx *4 ); mask =0xfULL << (idx *4 ); +if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { + bits |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + mask |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + } +rdmsrl (hwc->config_base, ctrl_val); ctrl_val &= ~mask; ctrl_val |= bits;arch/x86/include/asm/perf_event.h +1
-0@@ -33,6 +33,7 @@#define HSW_IN_TX (1ULL << 32)#define HSW_IN_TX_CHECKPOINTED (1ULL << 33)#define ICL_EVENTSEL_ADAPTIVE (1ULL << 34) +#define ICL_FIXED_0_ADAPTIVE (1ULL << 32)#define AMD64_EVENTSEL_INT_CORE_ENABLE (1ULL << 36)#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40)Signature · 1 line
2.7.4
Re: [PATCH 1/2] perf/x86/intel: Support adaptive PEBS for fixed counters
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · On Tue, Apr 09, 2019 at 06:09:59PM -0700, [email protected] wrote:
From: Kan Liang <[email protected]>
Fixed counters can also generate adaptive PEBS record, if the corresponding bit in
IA32_FIXED_CTR_CTRLis set. Otherwise, only basic record is generated.Unconditionally set the bit when PEBS is enabled on fixed counters. Let
MSR_PEBS_CFGdecide which format of PEBS record should be generated. There is no harmful to leave the bit set.
Quoted 5 lines
I'll merge this back into:
Subject: perf/x86/intel: Support adaptive PEBSv4
such that this bug never existed, ok?
Yes, please.
Thanks,
Kan
Quoted 39 lines
Quoted 39 lines
Signed-off-by Kan Liang <[email protected]>--- arch/x86/events/intel/core.c | 5 +++++ arch/x86/include/asm/perf_event.h | 1 + 2 files changed, 6 insertions(+)arch/x86/events/intel/core.c +5
-0@@ -2174,6 +2174,11 @@ static void intel_pmu_enable_fixed(struct perf_event *event) bits <<= (idx *4 ); mask =0xfULL << (idx *4 ); +if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { + bits |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + mask |=ICL_FIXED_0_ADAPTIVE << (idx *4 ); + } +rdmsrl (hwc->config_base, ctrl_val); ctrl_val &= ~mask; ctrl_val |= bits;arch/x86/include/asm/perf_event.h +1
-0@@ -33,6 +33,7 @@#define HSW_IN_TX (1ULL << 32)#define HSW_IN_TX_CHECKPOINTED (1ULL << 33)#define ICL_EVENTSEL_ADAPTIVE (1ULL << 34) +#define ICL_FIXED_0_ADAPTIVE (1ULL << 32)#define AMD64_EVENTSEL_INT_CORE_ENABLE (1ULL << 36)#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40)Signature · 1 line
2.7.4
PATCH 2/2 perf/x86/intel: Add Tremont core PMU support
In reply to [email protected]From: Kan Liang <[email protected]>
Add perf core PMU support for Intel Tremont CPU.
The init code is based on Goldmont plus.
The generic purpose counter 0 and fixed counter 0 have less skid. Force :ppp events on generic purpose counter 0.
Force instruction:ppp always on fixed counter 0.
Updates LLC cache event table and OFFCORE_RESPONSE mask.
The adaptive PEBS, which is already enabled on ICL, is also supported on Tremont. No extra codes required.
--- arch/x86/events/intel/core.c | 88 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 88 insertions(+)
arch/x86/events/intel/core.c +88 -0
@@ -1856,6 +1856,45 @@ static __initconst const u64 glp_hw_cache_extra_regs }, }; +#define TNT_LOCAL_DRAM BIT_ULL(26) +#define TNT_DEMAND_READ GLM_DEMAND_DATA_RD +#define TNT_DEMAND_WRITE GLM_DEMAND_RFO +#define TNT_LLC_ACCESS GLM_ANY_RESPONSE +#define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \ + SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM) +#define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM) + +static __initconst const u64 tnt_hw_cache_extra_regs + [PERF_COUNT_HW_CACHE_MAX ] + [PERF_COUNT_HW_CACHE_OP_MAX ] + [PERF_COUNT_HW_CACHE_RESULT_MAX ] = { + [C (LL )] = { + [C (OP_READ )] = { + [C (RESULT_ACCESS )] =TNT_DEMAND_READ | +TNT_LLC_ACCESS , + [C (RESULT_MISS )] =TNT_DEMAND_READ | +TNT_LLC_MISS , + }, + [C (OP_WRITE )] = { + [C (RESULT_ACCESS )] =TNT_DEMAND_WRITE | +TNT_LLC_ACCESS , + [C (RESULT_MISS )] =TNT_DEMAND_WRITE | +TNT_LLC_MISS , + }, + [C (OP_PREFETCH )] = { + [C (RESULT_ACCESS )] =0x0 , + [C (RESULT_MISS )] =0x0 , + }, + }, +}; + +static struct extra_reg intel_tnt_extra_regs[]__read_mostly = { +/* must define OFFCORE_RSP_X first, see intel_fixup_er() */ +INTEL_UEVENT_EXTRA_REG (0x01b7 ,MSR_OFFCORE_RSP_0 ,0xffffff9fffull ,RSP_0 ), +INTEL_UEVENT_EXTRA_REG (0x02b7 ,MSR_OFFCORE_RSP_1 ,0xffffff9fffull ,RSP_1 ), +EVENT_EXTRA_END +}; +#define KNL_OT_L2_HITE BIT_ULL(19)/* Other Tile L2 Hit */ #define KNL_OT_L2_HITF BIT_ULL(20)/* Other Tile L2 Hit */ #define KNL_MCDRAM_LOCAL BIT_ULL(21)
@@ -3451,6 +3490,29 @@ glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx,return c; } +static struct event_constraint * +tnt_get_event_constraints (struct cpu_hw_events *cpuc,int idx, +struct perf_event *event) +{ +struct event_constraint *c; + +/* +* :ppp means to do reduced skid PEBS, +* which is available at PMC0 and fixed counter 0. +*/ +if (event->attr.precise_ip ==3 ) { +/* Force instruction:ppp in Fixed counter 0 */ + if (event->hw.config == X86_CONFIG(.event=0xc0)) +return &fixed_counter0_constraint; + +return &counter0_constraint; + } + + c =intel_get_event_constraints (cpuc, idx, event); + +return c; +} +static bool allow_tsx_force_abort = true; static struct event_constraint *
@@ -4530,6 +4592,32 @@ __init int intel_pmu_init(void) name ="goldmont_plus" ;break ; +case INTEL_FAM6_ATOM_TREMONT_X : + x86_pmu.late_ack = true; +memcpy (hw_cache_event_ids, glp_hw_cache_event_ids, +sizeof (hw_cache_event_ids)); +memcpy (hw_cache_extra_regs, tnt_hw_cache_extra_regs, +sizeof (hw_cache_extra_regs)); + hw_cache_event_ids[C (ITLB )][C (OP_READ )][C (RESULT_ACCESS )] =-1 ; + +intel_pmu_lbr_init_skl (); + + x86_pmu.event_constraints = intel_slm_event_constraints; + x86_pmu.extra_regs = intel_tnt_extra_regs; +/* +* It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS +* for precise cycles. +*/ + x86_pmu.pebs_aliases =NULL ; + x86_pmu.pebs_prec_dist = true; + x86_pmu.lbr_pt_coexist = true; + x86_pmu.flags |=PMU_FL_HAS_RSP_1 ; + x86_pmu.get_event_constraints = tnt_get_event_constraints; + extra_attr = slm_format_attr; +pr_cont ("Tremont events, " ); + name ="Tremont" ; +break ; +case INTEL_FAM6_WESTMERE :case INTEL_FAM6_WESTMERE_EP :case INTEL_FAM6_WESTMERE_EX :
Signature · 1 line
2.7.4
Peter Zijlstra
maintainerRe: [PATCH 2/2] perf/x86/intel: Add Tremont core PMU support
Quoted 10 lines · On Tue, Apr 09, 2019 at 06:10:00PM -0700, [email protected] wrote:
The generic purpose counter 0 and fixed counter 0 have less skid. Force :ppp events on generic purpose counter 0.
Force instruction:ppp always on fixed counter 0.arch/x86/events/intel/core.c +6
-0+static struct event_constraint * +tnt_get_event_constraints (struct cpu_hw_events *cpuc,int idx, +struct perf_event *event) +{ +struct event_constraint *c; +
arch/x86/events/intel/core.c glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, · line 3508
Quoted 10 lines
arch/x86/events/intel/core.c +10
-0+/* +* :ppp means to do reduced skid PEBS, +* which is available at PMC0 and fixed counter 0. +*/ +if (event->attr.precise_ip ==3 ) { +/* Force instruction:ppp in Fixed counter 0 */ + if (event->hw.config == X86_CONFIG(.event=0xc0)) +return &fixed_counter0_constraint; + +return &counter0_constraint;
I'm confused, 0xc0 is the architectural 'instructions' event, surely we can program that on pmc0 too?
Did we want a fixed0_counter0_constraint for that?
Quoted 6 lines
arch/x86/events/intel/core.c +6
-0+ } + + c =intel_get_event_constraints (cpuc, idx, event); + +return c; +}
Re: [PATCH 2/2] perf/x86/intel: Add Tremont core PMU support
Quoted 26 lines · Peter Zijlstra
Quoted 10 lines · On Tue, Apr 09, 2019 at 06:10:00PM -0700, [email protected] wrote:
The generic purpose counter 0 and fixed counter 0 have less skid. Force :ppp events on generic purpose counter 0.
Force instruction:ppp always on fixed counter 0.arch/x86/events/intel/core.c +6
-0+static struct event_constraint * +tnt_get_event_constraints (struct cpu_hw_events *cpuc,int idx, +struct perf_event *event) +{ +struct event_constraint *c; +arch/x86/events/intel/core.c glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, · line 3508
Quoted 10 lines
arch/x86/events/intel/core.c +10
-0+/* +* :ppp means to do reduced skid PEBS, +* which is available at PMC0 and fixed counter 0. +*/ +if (event->attr.precise_ip ==3 ) { +/* Force instruction:ppp in Fixed counter 0 */ + if (event->hw.config == X86_CONFIG(.event=0xc0)) +return &fixed_counter0_constraint; + +return &counter0_constraint;I'm confused,
0xc0is the architectural 'instructions' event, surely we can program that on pmc0 too?
Quoted 2 lines
Did we want a
fixed0_counter0_constraintfor that?
Yes, I will send out V2 shortly to fix it.
Thanks,
Kan