perf, amd: Enable AMD family 15h northbridge counters
Subsystems (2)
v5 cover perf, amd: Enable AMD family 15h northbridge counters
The following patchset enables 4 additional performance counters in AMD family 15h processors that count northbridge events -- such as number of DRAM accesses.
This patchset is based on previous work done by Robert Richter <[email protected]> :
https://lkml.org/lkml/2012/6/19/324
The main differences are:
* The northbridge counters are indexed contiguously right above the core performance counters.
* MSR address offset calculations are moved to architecture specific files.
* Interrups are set up to be delivered only to a single core.
V5:
Rebased against latest tip
V4:
* Moved interrupt core select set up back to event constraints function, sicne during ->hw_config time we do not yet know on which CPU the the event will run on.
* Tested on and made minor revisions to make sure that the patchset is compatible with upcoming AMD Family 16h processors, and will support core and NB counters without any further patches.
V3:
Addressed the following feedback/comments from Robert's review
* https://lkml.org/lkml/2012/11/16/484
* https://lkml.org/lkml/2012/11/26/162
V2:
Separate out Robert's patches, and add properly ordered certificate of origins.
Jacob Shin (4):
perf, amd: Use proper naming scheme for AMD bit field definitions perf, x86: Move MSR address offset calculation to architecture
specific files
perf, x86: Allow for architecture specific RDPMC indexes perf, amd: Enable northbridge performance counters on AMD family 15h
Robert Richter (2):
perf, amd: Rework northbridge event constraints handler perf, amd: Generalize northbridge constraints code for family 15h
arch/x86/include/asm/cpufeature.h | 2 + arch/x86/include/asm/perf_event.h | 13 +- arch/x86/include/uapi/asm/msr-index.h | 2 + arch/x86/kernel/cpu/perf_event.c | 2 +- arch/x86/kernel/cpu/perf_event.h | 25 ++- arch/x86/kernel/cpu/perf_event_amd.c | 318 +++++++++++++++++++++++++-------- 6 files changed, 268 insertions(+), 94 deletions(-)
Signature · 1 line
1.7.9.5
PATCH RESEND v5 1/6 perf, amd: Rework northbridge event constraints handler
From: Robert Richter <[email protected]>
Code simplification. No functional changes.
--- arch/x86/kernel/cpu/perf_event_amd.c | 68 +++++++++++++--------------------- 1 file changed, 26 insertions(+), 42 deletions(-)
arch/x86/kernel/cpu/perf_event_amd.c +26 -42
@@ -256,9 +256,8 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) {struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb; -struct perf_event *old =NULL ; -int max = x86_pmu.num_counters; -int i, j, k =-1 ; +struct perf_event *old; +int idx, new =-1 ;/* * if not NB event or no NB, then no constraints
@@ -276,48 +275,33 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)* because of successive calls to x86_schedule_events() from * hw_perf_group_sched_in() without hw_perf_enable() */ -for (i =0 ; i < max; i++) { -/* -* keep track of first free slot -*/ -if (k ==-1 && !nb->owners[i]) - k = i; +for (idx =0 ; idx < x86_pmu.num_counters; idx++) { +if (new ==-1 || hwc->idx == idx) +/* assign free slot, prefer hwc->idx */ + old =cmpxchg (nb->owners + idx,NULL , event); +else if (nb->owners[idx] == event) +/* event already present */ + old = event; +else +continue ; + +if (old && old != event) +continue ; + +/* reassign to this slot */ +if (new !=-1 ) +cmpxchg (nb->owners + new, event,NULL ); + new = idx;/* already present, reuse */ -if (nb->owners[i] == event) - goto done; - } -/* -* not present, so grab a new slot -* starting either at: -*/ -if (hwc->idx !=-1 ) { -/* previous assignment */ - i = hwc->idx; - }else if (k !=-1 ) { -/* start from free slot found */ - i = k; - }else { -/* -* event not found, no slot found in -* first pass, try again from the -* beginning -*/ - i =0 ; - } - j = i; -do { - old =cmpxchg (nb->owners+i,NULL , event); -if (!old) +if (old == event)break ; -if (++i == max) - i =0 ; - }while (i != j); -done: -if (!old) -return &nb->event_constraints[i]; - -return &emptyconstraint; + } + +if (new ==-1 ) +return &emptyconstraint; + +return &nb->event_constraints[new]; }static struct amd_nb *amd_alloc_nb (int cpu)
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 1/6] perf, amd: Rework northbridge event constraints handler
Quoted 6 lines · Jacob Shin
From: Robert Richter <[email protected]>
Code simplification. No functional changes.
Signed-off-by Robert Richter <[email protected]>Signed-off-by Jacob Shin <[email protected]>
Quoted 97 lines
--- arch/x86/kernel/cpu/perf_event_amd.c | 68 +++++++++++++--------------------- 1 file changed, 26 insertions(+), 42 deletions(-)arch/x86/kernel/cpu/perf_event_amd.c +26
-43@@ -256,9 +256,8 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) {struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb; -struct perf_event *old =NULL ; -int max = x86_pmu.num_counters; -int i, j, k =-1 ; +struct perf_event *old; +int idx, new =-1 ;/* * if not NB event or no NB, then no constraints @@ -276,48 +275,33 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)* because of successive calls to x86_schedule_events() from * hw_perf_group_sched_in() without hw_perf_enable() */ -for (i =0 ; i < max; i++) { -/* -* keep track of first free slot -*/ -if (k ==-1 && !nb->owners[i]) - k = i; +for (idx =0 ; idx < x86_pmu.num_counters; idx++) { +if (new ==-1 || hwc->idx == idx) +/* assign free slot, prefer hwc->idx */ + old =cmpxchg (nb->owners + idx,NULL , event); +else if (nb->owners[idx] == event) +/* event already present */ + old = event; +else +continue ; + +if (old && old != event) +continue ; + +/* reassign to this slot */ +if (new !=-1 ) +cmpxchg (nb->owners + new, event,NULL ); + new = idx;/* already present, reuse */ -if (nb->owners[i] == event) - goto done; - } -/* -* not present, so grab a new slot -* starting either at: -*/ -if (hwc->idx !=-1 ) { -/* previous assignment */ - i = hwc->idx; - }else if (k !=-1 ) { -/* start from free slot found */ - i = k; - }else { -/* -* event not found, no slot found in -* first pass, try again from the -* beginning -*/ - i =0 ; - } - j = i; -do { - old =cmpxchg (nb->owners+i,NULL , event); -if (!old) +if (old == event)break ; -if (++i == max) - i =0 ; - }while (i != j); -done: -if (!old) -return &nb->event_constraints[i]; - -return &emptyconstraint; + } + +if (new ==-1 ) +return &emptyconstraint; + +return &nb->event_constraints[new]; }static struct amd_nb *amd_alloc_nb (int cpu) --1.7.9.5
PATCH RESEND v5 2/6 perf, amd: Generalize northbridge constraints code for family 15h
In reply to Jacob ShinFrom: Robert Richter <[email protected]>
Generalize northbridge constraints code for family 10h so that later we can reuse the same code path with other AMD processor families that have the same northbridge event constraints.
--- arch/x86/kernel/cpu/perf_event_amd.c | 43 ++++++++++++++++++++-------------- 1 file changed, 25 insertions(+), 18 deletions(-)
arch/x86/kernel/cpu/perf_event_amd.c +25 -18
@@ -188,20 +188,13 @@ static inline int amd_has_nb(struct cpu_hw_events *cpuc)return nb && nb->nb_id !=-1 ; } -static void amd_put_event_constraints (struct cpu_hw_events *cpuc, -struct perf_event *event) +static void __amd_put_nb_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) { -struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;int i;/* -* only care about NB events -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return ; - -/* * need to scan whole list because event may not have * been assigned during scheduling *
@@ -247,12 +240,13 @@ static void amd_put_event_constraints(struct cpu_hw_events *cpuc,* * Given that resources are allocated (cmpxchg), they must be * eventually freed for others to use. This is accomplished by -* calling amd_put_event_constraints(). +* calling __amd_put_nb_event_constraints() * * Non NB events are not impacted by this restriction. */ static struct event_constraint * -amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +__amd_get_nb_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event, +struct event_constraint *c) {struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;
@@ -260,12 +254,6 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)int idx, new =-1 ;/* -* if not NB event or no NB, then no constraints -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return &unconstrained; - -/* * detect if already present, if so reuse * * cannot merge with actual allocation
@@ -275,7 +263,7 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)* because of successive calls to x86_schedule_events() from * hw_perf_group_sched_in() without hw_perf_enable() */ -for (idx =0 ; idx < x86_pmu.num_counters; idx++) { +for_each_set_bit (idx, c->idxmsk,X86_PMC_IDX_MAX ) {if (new ==-1 || hwc->idx == idx)/* assign free slot, prefer hwc->idx */ old =cmpxchg (nb->owners + idx,NULL , event);
@@ -391,6 +379,25 @@ static void amd_pmu_cpu_dead(int cpu) } } +static struct event_constraint * +amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +{ +/* +* if not NB event or no NB, then no constraints +*/ +if (!(amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw))) +return &unconstrained; + +return __amd_get_nb_event_constraints (cpuc, event, &unconstrained); +} + +static void amd_put_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) +{ +if (amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw)) +__amd_put_nb_event_constraints (cpuc, event); +} +PMU_FORMAT_ATTR (event,"config:0-7,32-35" );PMU_FORMAT_ATTR (umask,"config:8-15" );PMU_FORMAT_ATTR (edge,"config:18" );
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 2/6] perf, amd: Generalize northbridge constraints code for family 15h
Quoted 69 lines · Jacob Shin
From: Robert Richter <[email protected]>
Generalize northbridge constraints code for family 10h so that later we can reuse the same code path with other AMD processor families that have the same northbridge event constraints.
Signed-off-by Robert Richter <[email protected]>Signed-off-by Jacob Shin <[email protected]>--- arch/x86/kernel/cpu/perf_event_amd.c | 43 ++++++++++++++++++++-------------- 1 file changed, 25 insertions(+), 18 deletions(-)arch/x86/kernel/cpu/perf_event_amd.c +5
-17@@ -188,20 +188,13 @@ static inline int amd_has_nb(struct cpu_hw_events *cpuc)return nb && nb->nb_id !=-1 ; } -static void amd_put_event_constraints (struct cpu_hw_events *cpuc, -struct perf_event *event) +static void __amd_put_nb_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) { -struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;int i;/* -* only care about NB events -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return ; - -/* * need to scan whole list because event may not have * been assigned during scheduling * @@ -247,12 +240,13 @@ static void amd_put_event_constraints(struct cpu_hw_events *cpuc,* * Given that resources are allocated (cmpxchg), they must be * eventually freed for others to use. This is accomplished by -* calling amd_put_event_constraints(). +* calling __amd_put_nb_event_constraints() * * Non NB events are not impacted by this restriction. */ static struct event_constraint * -amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +__amd_get_nb_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event, +struct event_constraint *c) {struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;@@ -260,12 +254,6 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)int idx, new =-1 ;/* -* if not NB event or no NB, then no constraints -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return &unconstrained; - -/* * detect if already present, if so reuse * * cannot merge with actual allocation
arch/x86/kernel/cpu/perf_event_amd.c amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) · line 266
Quoted 5 lines
arch/x86/kernel/cpu/perf_event_amd.c +1
-1@@ -275,7 +263,7 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)* because of successive calls to x86_schedule_events() from * hw_perf_group_sched_in() without hw_perf_enable() */ -for (idx =0 ; idx < x86_pmu.num_counters; idx++) { +for_each_set_bit (idx, c->idxmsk,X86_PMC_IDX_MAX ) {
So here you're using X86_PMC_IDX_MAX but in__amd_put_nb_event_constraints() you're using
x86_pmu.num_counters.
There is implicit assumption in the AMD code the counters index namespace is contiguous. That
means the uncore counters show up right after the core counters. On Fam15h, that would be NB
counters start at index 6, on Fam10h at index 4. In that case, the constraint mask cannot have bits set
beyond num_counters, so why not use that limit in amd_get_event_constraints()? It would significantly
cut down on the number of iterations in the loop from 64 down to 10 on Fam15h.
Quoted 33 lines
arch/x86/kernel/cpu/perf_event_amd.c +19
-1if (new ==-1 || hwc->idx == idx)/* assign free slot, prefer hwc->idx */ old =cmpxchg (nb->owners + idx,NULL , event);@@ -391,6 +379,25 @@ static void amd_pmu_cpu_dead(int cpu) } } +static struct event_constraint * +amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +{ +/* +* if not NB event or no NB, then no constraints +*/ +if (!(amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw))) +return &unconstrained; + +return __amd_get_nb_event_constraints (cpuc, event, &unconstrained); +} + +static void amd_put_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) +{ +if (amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw)) +__amd_put_nb_event_constraints (cpuc, event); +} +PMU_FORMAT_ATTR (event,"config:0-7,32-35" );PMU_FORMAT_ATTR (umask,"config:8-15" );PMU_FORMAT_ATTR (edge,"config:18" ); --1.7.9.5
Re: [PATCH RESEND V5 2/6] perf, amd: Generalize northbridge constraints code for family 15h
Quoted 80 lines · Stephane Eranian
Quoted 69 lines · Jacob Shin
From: Robert Richter <[email protected]>
Generalize northbridge constraints code for family 10h so that later we can reuse the same code path with other AMD processor families that have the same northbridge event constraints.
Signed-off-by Robert Richter <[email protected]>Signed-off-by Jacob Shin <[email protected]>--- arch/x86/kernel/cpu/perf_event_amd.c | 43 ++++++++++++++++++++-------------- 1 file changed, 25 insertions(+), 18 deletions(-)arch/x86/kernel/cpu/perf_event_amd.c +5
-17@@ -188,20 +188,13 @@ static inline int amd_has_nb(struct cpu_hw_events *cpuc)return nb && nb->nb_id !=-1 ; } -static void amd_put_event_constraints (struct cpu_hw_events *cpuc, -struct perf_event *event) +static void __amd_put_nb_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) { -struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;int i;/* -* only care about NB events -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return ; - -/* * need to scan whole list because event may not have * been assigned during scheduling * @@ -247,12 +240,13 @@ static void amd_put_event_constraints(struct cpu_hw_events *cpuc,* * Given that resources are allocated (cmpxchg), they must be * eventually freed for others to use. This is accomplished by -* calling amd_put_event_constraints(). +* calling __amd_put_nb_event_constraints() * * Non NB events are not impacted by this restriction. */ static struct event_constraint * -amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +__amd_get_nb_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event, +struct event_constraint *c) {struct hw_perf_event *hwc = &event->hw;struct amd_nb *nb = cpuc->amd_nb;@@ -260,12 +254,6 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)int idx, new =-1 ;/* -* if not NB event or no NB, then no constraints -*/ -if (!(amd_has_nb (cpuc) &&amd_is_nb_event (hwc))) -return &unconstrained; - -/* * detect if already present, if so reuse * * cannot merge with actual allocation arch/x86/kernel/cpu/perf_event_amd.c amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event) · line 266
Quoted 5 lines
arch/x86/kernel/cpu/perf_event_amd.c +1
-1@@ -275,7 +263,7 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)* because of successive calls to x86_schedule_events() from * hw_perf_group_sched_in() without hw_perf_enable() */ -for (idx =0 ; idx < x86_pmu.num_counters; idx++) { +for_each_set_bit (idx, c->idxmsk,X86_PMC_IDX_MAX ) {So here you're using
X86_PMC_IDX_MAXbut in__amd_put_nb_event_constraints()you're using
x86_pmu.num_counters.
Quoted 9 lines
There is implicit assumption in the AMD code the counters index namespace is contiguous. That
means the uncore counters show up right after the core counters. On Fam15h, that would be NB
counters start at index 6, on Fam10h at index 4. In that case, the constraint mask cannot have bits set
beyondnum_counters, so why not use that limit inamd_get_event_constraints()? It would significantly
cut down on the number of iterations in the loop from 64 down to 10 on Fam15h.
Yes, you are right, I will change that in V6.
Thanks,
Quoted 36 lines
Quoted 35 lines
arch/x86/kernel/cpu/perf_event_amd.c +19
-1if (new ==-1 || hwc->idx == idx)/* assign free slot, prefer hwc->idx */ old =cmpxchg (nb->owners + idx,NULL , event);@@ -391,6 +379,25 @@ static void amd_pmu_cpu_dead(int cpu) } } +static struct event_constraint * +amd_get_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) +{ +/* +* if not NB event or no NB, then no constraints +*/ +if (!(amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw))) +return &unconstrained; + +return __amd_get_nb_event_constraints (cpuc, event, &unconstrained); +} + +static void amd_put_event_constraints (struct cpu_hw_events *cpuc, +struct perf_event *event) +{ +if (amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw)) +__amd_put_nb_event_constraints (cpuc, event); +} +PMU_FORMAT_ATTR (event,"config:0-7,32-35" );PMU_FORMAT_ATTR (umask,"config:8-15" );PMU_FORMAT_ATTR (edge,"config:18" ); --1.7.9.5
PATCH RESEND v5 3/6 perf, amd: Use proper naming scheme for AMD bit field definitions
In reply to Jacob ShinUpdate these AMD bit field names to be consistent with naming convention followed by the rest of the file.
--- arch/x86/include/asm/perf_event.h | 4 ++-- arch/x86/kernel/cpu/perf_event_amd.c | 8 ++++---- 2 files changed, 6 insertions(+), 6 deletions(-)
arch/x86/include/asm/perf_event.h +2 -2
@@ -29,8 +29,8 @@#define ARCH_PERFMON_EVENTSEL_INV (1ULL << 23)#define ARCH_PERFMON_EVENTSEL_CMASK 0xFF000000ULL -#define AMD_PERFMON_EVENTSEL_GUESTONLY (1ULL << 40) -#define AMD_PERFMON_EVENTSEL_HOSTONLY (1ULL << 41) +#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40) +#define AMD64_EVENTSEL_HOSTONLY (1ULL << 41)#define AMD64_EVENTSEL_EVENT \ (ARCH_PERFMON_EVENTSEL_EVENT | (0x0FULL << 32))
arch/x86/kernel/cpu/perf_event_amd.c +4 -4
@@ -156,9 +156,9 @@ static int amd_pmu_hw_config(struct perf_event *event) event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |ARCH_PERFMON_EVENTSEL_OS ); else if (event->attr.exclude_host) - event->hw.config |= AMD_PERFMON_EVENTSEL_GUESTONLY; + event->hw.config |= AMD64_EVENTSEL_GUESTONLY; else if (event->attr.exclude_guest) - event->hw.config |= AMD_PERFMON_EVENTSEL_HOSTONLY; + event->hw.config |= AMD64_EVENTSEL_HOSTONLY;if (event->attr.type !=PERF_TYPE_RAW )return 0 ;
@@ -336,7 +336,7 @@ static void amd_pmu_cpu_starting(int cpu)struct amd_nb *nb;int i, nb_id; - cpuc->perf_ctr_virt_mask =AMD_PERFMON_EVENTSEL_HOSTONLY ; + cpuc->perf_ctr_virt_mask =AMD64_EVENTSEL_HOSTONLY ;if (boot_cpu_data.x86_max_cores <2 )return ;
@@ -669,7 +669,7 @@ void amd_pmu_disable_virt(void)* SVM is disabled the Guest-only bits still gets set and the counter * will not count anything. */ - cpuc->perf_ctr_virt_mask =AMD_PERFMON_EVENTSEL_HOSTONLY ; + cpuc->perf_ctr_virt_mask =AMD64_EVENTSEL_HOSTONLY ;/* Reload all events */ x86_pmu_disable_all ();
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 3/6] perf, amd: Use proper naming scheme for AMD bit field definitions
Quoted 4 lines · Jacob Shin
Update these AMD bit field names to be consistent with naming convention followed by the rest of the file.
Signed-off-by Jacob Shin <[email protected]>
Quoted 58 lines
--- arch/x86/include/asm/perf_event.h | 4 ++-- arch/x86/kernel/cpu/perf_event_amd.c | 8 ++++---- 2 files changed, 6 insertions(+), 6 deletions(-)arch/x86/include/asm/perf_event.h +2
-2@@ -29,8 +29,8 @@#define ARCH_PERFMON_EVENTSEL_INV (1ULL << 23)#define ARCH_PERFMON_EVENTSEL_CMASK 0xFF000000ULL -#define AMD_PERFMON_EVENTSEL_GUESTONLY (1ULL << 40) -#define AMD_PERFMON_EVENTSEL_HOSTONLY (1ULL << 41) +#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40) +#define AMD64_EVENTSEL_HOSTONLY (1ULL << 41)#define AMD64_EVENTSEL_EVENT \ (ARCH_PERFMON_EVENTSEL_EVENT | (0x0FULL << 32))arch/x86/kernel/cpu/perf_event_amd.c +4
-5@@ -156,9 +156,9 @@ static int amd_pmu_hw_config(struct perf_event *event) event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |ARCH_PERFMON_EVENTSEL_OS ); else if (event->attr.exclude_host) - event->hw.config |= AMD_PERFMON_EVENTSEL_GUESTONLY; + event->hw.config |= AMD64_EVENTSEL_GUESTONLY; else if (event->attr.exclude_guest) - event->hw.config |= AMD_PERFMON_EVENTSEL_HOSTONLY; + event->hw.config |= AMD64_EVENTSEL_HOSTONLY;if (event->attr.type !=PERF_TYPE_RAW )return 0 ;@@ -336,7 +336,7 @@ static void amd_pmu_cpu_starting(int cpu)struct amd_nb *nb;int i, nb_id; - cpuc->perf_ctr_virt_mask =AMD_PERFMON_EVENTSEL_HOSTONLY ; + cpuc->perf_ctr_virt_mask =AMD64_EVENTSEL_HOSTONLY ;if (boot_cpu_data.x86_max_cores <2 )return ;@@ -669,7 +669,7 @@ void amd_pmu_disable_virt(void)* SVM is disabled the Guest-only bits still gets set and the counter * will not count anything. */ - cpuc->perf_ctr_virt_mask =AMD_PERFMON_EVENTSEL_HOSTONLY ; + cpuc->perf_ctr_virt_mask =AMD64_EVENTSEL_HOSTONLY ;/* Reload all events */ x86_pmu_disable_all (); --1.7.9.5
PATCH RESEND v5 4/6 perf, x86: Move MSR address offset calculation to architecture specific files
In reply to Jacob ShinMove counter index to MSR address offset calculation to architecture specific files. This prepares the way for perf_event_amd to enable counter addresses that are not contiguous -- for example AMD Family 15h processors have 6 core performance counters starting at 0xc0010200 and 4 northbridge performance counters starting at 0xc0010240.
--- arch/x86/kernel/cpu/perf_event.h | 21 ++++------------- arch/x86/kernel/cpu/perf_event_amd.c | 42 ++++++++++++++++++++++++++++++++++ 2 files changed, 47 insertions(+), 16 deletions(-)
arch/x86/kernel/cpu/perf_event.h +5 -16
@@ -325,6 +325,7 @@ struct x86_pmu {int (*schedule_events)(struct cpu_hw_events *cpuc,int n,int *assign);unsigned eventsel;unsigned perfctr; +int (*addr_offset)(int index,int eventsel);u64 (*event_map)(int);int max_events;int num_counters;
@@ -446,28 +447,16 @@ extern u64 __read_mostly hw_cache_extra_regsu64 x86_perf_event_update (struct perf_event *event); -static inline int x86_pmu_addr_offset (int index) -{ -int offset; - -/* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */ -alternative_io (ASM_NOP2 , -"shll $1, %%eax" , -X86_FEATURE_PERFCTR_CORE , -"=a" (offset), -"a" (index)); - -return offset; -} -static inline unsigned int x86_pmu_config_addr (int index) { -return x86_pmu.eventsel +x86_pmu_addr_offset (index); +return x86_pmu.eventsel + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,1 ) : index); }static inline unsigned int x86_pmu_event_addr (int index) { -return x86_pmu.perfctr +x86_pmu_addr_offset (index); +return x86_pmu.perfctr + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,0 ) : index); }int x86_setup_perfctr (struct perf_event *event);
arch/x86/kernel/cpu/perf_event_amd.c +42 -0
54 lines not on this page · raw message
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 4/6] perf, x86: Move MSR address offset calculation to architecture specific files
Quoted 45 lines · Jacob Shin
Move counter index to MSR address offset calculation to architecture specific files. This prepares the way for
perf_event_amdto enable counter addresses that are not contiguous -- for example AMD Family 15h processors have 6 core performance counters starting at0xc0010200and 4 northbridge performance counters starting at0xc0010240.Signed-off-by Jacob Shin <[email protected]>--- arch/x86/kernel/cpu/perf_event.h | 21 ++++------------- arch/x86/kernel/cpu/perf_event_amd.c | 42 ++++++++++++++++++++++++++++++++++ 2 files changed, 47 insertions(+), 16 deletions(-)arch/x86/kernel/cpu/perf_event.h +1
-15@@ -325,6 +325,7 @@ struct x86_pmu {int (*schedule_events)(struct cpu_hw_events *cpuc,int n,int *assign);unsigned eventsel;unsigned perfctr; +int (*addr_offset)(int index,int eventsel);u64 (*event_map)(int);int max_events;int num_counters;@@ -446,28 +447,16 @@ extern u64 __read_mostly hw_cache_extra_regsu64 x86_perf_event_update (struct perf_event *event); -static inline int x86_pmu_addr_offset (int index) -{ -int offset; - -/* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */ -alternative_io (ASM_NOP2 , -"shll $1, %%eax" , -X86_FEATURE_PERFCTR_CORE , -"=a" (offset), -"a" (index)); - -return offset; -} -static inline unsigned int x86_pmu_config_addr (int index) { -return x86_pmu.eventsel +x86_pmu_addr_offset (index);
arch/x86/kernel/cpu/perf_event.h extern u64 __read_mostly hw_cache_extra_regs · line 460
Quoted 10 lines
arch/x86/kernel/cpu/perf_event.h +4
-1+return x86_pmu.eventsel + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,1 ) : index); }static inline unsigned int x86_pmu_event_addr (int index) { -return x86_pmu.perfctr +x86_pmu_addr_offset (index); +return x86_pmu.perfctr + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,0 ) : index); }
Would be better to use a constant name instead of 1 and 0 to name a event_sel vs. a counter. It would help the reader understand what this is about as that may
be useful for other processors as well.
Quoted 65 lines
arch/x86/kernel/cpu/perf_event.h +0
-0int x86_setup_perfctr (struct perf_event *event);arch/x86/kernel/cpu/perf_event_amd.c +42
-1@@ -132,6 +132,47 @@ static u64 amd_pmu_event_map(int hw_event)return amd_perfmon_event_map[hw_event]; } +/* +* Previously calculated offsets +*/ +static unsigned int event_offsets[X86_PMC_IDX_MAX ]__read_mostly ; +static unsigned int count_offsets[X86_PMC_IDX_MAX ]__read_mostly ; + +/* +* Legacy CPUs: +* 4 counters starting at 0xc0010000 each offset by 1 +* +* CPUs with core performance counter extensions: +* 6 counters starting at 0xc0010200 each offset by 2 +*/ +static inline int amd_pmu_addr_offset (int index,int eventsel) +{ +int offset; + +if (!index) +return index; + +if (eventsel) + offset = event_offsets[index]; +else + offset = count_offsets[index]; + +if (offset) +return offset; + +if (!cpu_has_perfctr_core) + offset = index; +else + offset = index <<1 ; + +if (eventsel) + event_offsets[index] = offset; +else + count_offsets[index] = offset; + +return offset; +} +static int amd_pmu_hw_config (struct perf_event *event) {int ret;@@ -578,6 +619,7 @@ static __initconst const struct x86_pmu amd_pmu = { .schedule_events = x86_schedule_events, .eventsel = MSR_K7_EVNTSEL0, .perfctr = MSR_K7_PERFCTR0, + .addr_offset = amd_pmu_addr_offset, .event_map = amd_pmu_event_map, .max_events = ARRAY_SIZE(amd_perfmon_event_map), .num_counters = AMD64_NUM_COUNTERS, --1.7.9.5
Re: [PATCH RESEND V5 4/6] perf, x86: Move MSR address offset calculation to architecture specific files
Quoted 50 lines · Stephane Eranian
Quoted 49 lines · Jacob Shin
Move counter index to MSR address offset calculation to architecture specific files. This prepares the way for
perf_event_amdto enable counter addresses that are not contiguous -- for example AMD Family 15h processors have 6 core performance counters starting at0xc0010200and 4 northbridge performance counters starting at0xc0010240.Signed-off-by Jacob Shin <[email protected]>--- arch/x86/kernel/cpu/perf_event.h | 21 ++++------------- arch/x86/kernel/cpu/perf_event_amd.c | 42 ++++++++++++++++++++++++++++++++++ 2 files changed, 47 insertions(+), 16 deletions(-)arch/x86/kernel/cpu/perf_event.h +3
-15@@ -325,6 +325,7 @@ struct x86_pmu {int (*schedule_events)(struct cpu_hw_events *cpuc,int n,int *assign);unsigned eventsel;unsigned perfctr; +int (*addr_offset)(int index,int eventsel);u64 (*event_map)(int);int max_events;int num_counters;@@ -446,28 +447,16 @@ extern u64 __read_mostly hw_cache_extra_regsu64 x86_perf_event_update (struct perf_event *event); -static inline int x86_pmu_addr_offset (int index) -{ -int offset; - -/* offset = X86_FEATURE_PERFCTR_CORE ? index << 1 : index */ -alternative_io (ASM_NOP2 , -"shll $1, %%eax" , -X86_FEATURE_PERFCTR_CORE , -"=a" (offset), -"a" (index)); - -return offset; -} -static inline unsigned int x86_pmu_config_addr (int index) { -return x86_pmu.eventsel +x86_pmu_addr_offset (index); +return x86_pmu.eventsel + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,1 ) : index); }
Quoted 10 lines
arch/x86/kernel/cpu/perf_event.h extern u64 __read_mostly hw_cache_extra_regs · line 460
Quoted 6 lines
arch/x86/kernel/cpu/perf_event.h +2
-1static inline unsigned int x86_pmu_event_addr (int index) { -return x86_pmu.perfctr +x86_pmu_addr_offset (index); +return x86_pmu.perfctr + + (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,0 ) : index); }Would be better to use a constant name instead of 1 and 0 to name a
event_selvs. a counter. It would help the reader understand what this is about as that may
be useful for other processors as well.
Yes will do .. or should I use bool instead? Which would be preferred?
Quoted 67 lines
Quoted 66 lines
arch/x86/kernel/cpu/perf_event.h +0
-0int x86_setup_perfctr (struct perf_event *event);arch/x86/kernel/cpu/perf_event_amd.c +42
-1@@ -132,6 +132,47 @@ static u64 amd_pmu_event_map(int hw_event)return amd_perfmon_event_map[hw_event]; } +/* +* Previously calculated offsets +*/ +static unsigned int event_offsets[X86_PMC_IDX_MAX ]__read_mostly ; +static unsigned int count_offsets[X86_PMC_IDX_MAX ]__read_mostly ; + +/* +* Legacy CPUs: +* 4 counters starting at 0xc0010000 each offset by 1 +* +* CPUs with core performance counter extensions: +* 6 counters starting at 0xc0010200 each offset by 2 +*/ +static inline int amd_pmu_addr_offset (int index,int eventsel) +{ +int offset; + +if (!index) +return index; + +if (eventsel) + offset = event_offsets[index]; +else + offset = count_offsets[index]; + +if (offset) +return offset; + +if (!cpu_has_perfctr_core) + offset = index; +else + offset = index <<1 ; + +if (eventsel) + event_offsets[index] = offset; +else + count_offsets[index] = offset; + +return offset; +} +static int amd_pmu_hw_config (struct perf_event *event) {int ret;@@ -578,6 +619,7 @@ static __initconst const struct x86_pmu amd_pmu = { .schedule_events = x86_schedule_events, .eventsel = MSR_K7_EVNTSEL0, .perfctr = MSR_K7_PERFCTR0, + .addr_offset = amd_pmu_addr_offset, .event_map = amd_pmu_event_map, .max_events = ARRAY_SIZE(amd_perfmon_event_map), .num_counters = AMD64_NUM_COUNTERS, --1.7.9.5
PATCH RESEND v5 5/6 perf, x86: Allow for architecture specific RDPMC indexes
In reply to Jacob ShinSimilar to config_base and event_base, allow architecture specific RDPMC ECX values.
--- arch/x86/kernel/cpu/perf_event.c | 2 +- arch/x86/kernel/cpu/perf_event.h | 6 ++++++ arch/x86/kernel/cpu/perf_event_amd.c | 6 ++++++ 3 files changed, 13 insertions(+), 1 deletion(-)
arch/x86/kernel/cpu/perf_event.c +1 -1
@@ -835,7 +835,7 @@ static inline void x86_assign_hw_event(struct perf_event *event, }else { hwc->config_base =x86_pmu_config_addr (hwc->idx); hwc->event_base =x86_pmu_event_addr (hwc->idx); - hwc->event_base_rdpmc = hwc->idx; + hwc->event_base_rdpmc =x86_pmu_rdpmc_index (hwc->idx); } }
arch/x86/kernel/cpu/perf_event.h +6 -0
@@ -326,6 +326,7 @@ struct x86_pmu {unsigned eventsel;unsigned perfctr;int (*addr_offset)(int index,int eventsel); +int (*rdpmc_index)(int index);u64 (*event_map)(int);int max_events;int num_counters;
@@ -459,6 +460,11 @@ static inline unsigned int x86_pmu_event_addr(int index) (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,0 ) : index); } +static inline int x86_pmu_rdpmc_index (int index) +{ +return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index (index) : index; +} +int x86_setup_perfctr (struct perf_event *event);int x86_pmu_hw_config (struct perf_event *event);
arch/x86/kernel/cpu/perf_event_amd.c +6 -0
@@ -173,6 +173,11 @@ static inline int amd_pmu_addr_offset(int index, int eventsel)return offset; } +static inline int amd_pmu_rdpmc_index (int index) +{ +return index; +} +static int amd_pmu_hw_config (struct perf_event *event) {int ret;
@@ -620,6 +625,7 @@ static __initconst const struct x86_pmu amd_pmu = { .eventsel = MSR_K7_EVNTSEL0, .perfctr = MSR_K7_PERFCTR0, .addr_offset = amd_pmu_addr_offset, + .rdpmc_index = amd_pmu_rdpmc_index, .event_map = amd_pmu_event_map, .max_events = ARRAY_SIZE(amd_perfmon_event_map), .num_counters = AMD64_NUM_COUNTERS,
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 5/6] perf, x86: Allow for architecture specific RDPMC indexes
Quoted 4 lines · Jacob Shin
Similar to
config_baseandevent_base, allow architecture specific RDPMC ECX values.Signed-off-by Jacob Shin <[email protected]>
Quoted 71 lines
--- arch/x86/kernel/cpu/perf_event.c | 2 +- arch/x86/kernel/cpu/perf_event.h | 6 ++++++ arch/x86/kernel/cpu/perf_event_amd.c | 6 ++++++ 3 files changed, 13 insertions(+), 1 deletion(-)arch/x86/kernel/cpu/perf_event.c +1
-1@@ -835,7 +835,7 @@ static inline void x86_assign_hw_event(struct perf_event *event, }else { hwc->config_base =x86_pmu_config_addr (hwc->idx); hwc->event_base =x86_pmu_event_addr (hwc->idx); - hwc->event_base_rdpmc = hwc->idx; + hwc->event_base_rdpmc =x86_pmu_rdpmc_index (hwc->idx); } }arch/x86/kernel/cpu/perf_event.h +6
-0@@ -326,6 +326,7 @@ struct x86_pmu {unsigned eventsel;unsigned perfctr;int (*addr_offset)(int index,int eventsel); +int (*rdpmc_index)(int index);u64 (*event_map)(int);int max_events;int num_counters;@@ -459,6 +460,11 @@ static inline unsigned int x86_pmu_event_addr(int index) (x86_pmu.addr_offset ? x86_pmu.addr_offset (index,0 ) : index); } +static inline int x86_pmu_rdpmc_index (int index) +{ +return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index (index) : index; +} +int x86_setup_perfctr (struct perf_event *event);int x86_pmu_hw_config (struct perf_event *event);arch/x86/kernel/cpu/perf_event_amd.c +6
-1@@ -173,6 +173,11 @@ static inline int amd_pmu_addr_offset(int index, int eventsel)return offset; } +static inline int amd_pmu_rdpmc_index (int index) +{ +return index; +} +static int amd_pmu_hw_config (struct perf_event *event) {int ret;@@ -620,6 +625,7 @@ static __initconst const struct x86_pmu amd_pmu = { .eventsel = MSR_K7_EVNTSEL0, .perfctr = MSR_K7_PERFCTR0, .addr_offset = amd_pmu_addr_offset, + .rdpmc_index = amd_pmu_rdpmc_index, .event_map = amd_pmu_event_map, .max_events = ARRAY_SIZE(amd_perfmon_event_map), .num_counters = AMD64_NUM_COUNTERS, --1.7.9.5
PATCH RESEND v5 6/6 perf, amd: Enable northbridge performance counters on AMD family 15h
In reply to Jacob ShinOn AMD family 15h processors, there are 4 new performance counters (in addition to 6 core performance counters) that can be used for counting northbridge events (i.e. DRAM accesses). Their bit fields are almost identical to the core performance counters. However, unlike the core performance counters, these MSRs are shared between multiple cores (that share the same northbridge). We will reuse the same code path as existing family 10h northbridge event constraints handler logic to enforce this sharing.
--- arch/x86/include/asm/cpufeature.h | 2 + arch/x86/include/asm/perf_event.h | 9 ++ arch/x86/include/uapi/asm/msr-index.h | 2 + arch/x86/kernel/cpu/perf_event_amd.c | 167 +++++++++++++++++++++++++++++---- 4 files changed, 160 insertions(+), 20 deletions(-)
arch/x86/include/asm/cpufeature.h +2 -0
@@ -167,6 +167,7 @@#define X86_FEATURE_TBM (6*32+21)/* trailing bit manipulations */ #define X86_FEATURE_TOPOEXT (6*32+22)/* topology extensions CPUID leafs */ #define X86_FEATURE_PERFCTR_CORE (6*32+23)/* core performance counter extensions */ +#define X86_FEATURE_PERFCTR_NB (6*32+24)/* NB performance counter extensions */ /* * Auxiliary flags: Linux defined - For features scattered in various
@@ -309,6 +310,7 @@ extern const char * const x86_power_flags[32];#define cpu_has_hypervisor boot_cpu_has(X86_FEATURE_HYPERVISOR)#define cpu_has_pclmulqdq boot_cpu_has(X86_FEATURE_PCLMULQDQ)#define cpu_has_perfctr_core boot_cpu_has(X86_FEATURE_PERFCTR_CORE) +#define cpu_has_perfctr_nb boot_cpu_has(X86_FEATURE_PERFCTR_NB)#define cpu_has_cx8 boot_cpu_has(X86_FEATURE_CX8)#define cpu_has_cx16 boot_cpu_has(X86_FEATURE_CX16)#define cpu_has_eager_fpu boot_cpu_has(X86_FEATURE_EAGER_FPU)
arch/x86/include/asm/perf_event.h +9 -0
@@ -29,9 +29,14 @@#define ARCH_PERFMON_EVENTSEL_INV (1ULL << 23)#define ARCH_PERFMON_EVENTSEL_CMASK 0xFF000000ULL +#define AMD64_EVENTSEL_INT_CORE_ENABLE (1ULL << 36)#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40)#define AMD64_EVENTSEL_HOSTONLY (1ULL << 41) +#define AMD64_EVENTSEL_INT_CORE_SEL_SHIFT 37 +#define AMD64_EVENTSEL_INT_CORE_SEL_MASK \ + (0xFULL << AMD64_EVENTSEL_INT_CORE_SEL_SHIFT) +#define AMD64_EVENTSEL_EVENT \ (ARCH_PERFMON_EVENTSEL_EVENT | (0x0FULL << 32))#define INTEL_ARCH_EVENT_MASK \
@@ -46,8 +51,12 @@#define AMD64_RAW_EVENT_MASK \ (X86_RAW_EVENT_MASK | \ AMD64_EVENTSEL_EVENT) +#define AMD64_RAW_EVENT_MASK_NB \ + (AMD64_EVENTSEL_EVENT | \ + ARCH_PERFMON_EVENTSEL_UMASK)#define AMD64_NUM_COUNTERS 4#define AMD64_NUM_COUNTERS_CORE 6 +#define AMD64_NUM_COUNTERS_NB 4#define ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL 0x3c#define ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK (0x00 << 8)
arch/x86/include/uapi/asm/msr-index.h +2 -0
@@ -194,6 +194,8 @@/* Fam 15h MSRs */ #define MSR_F15H_PERF_CTL 0xc0010200#define MSR_F15H_PERF_CTR 0xc0010201 +#define MSR_F15H_NB_PERF_CTL 0xc0010240 +#define MSR_F15H_NB_PERF_CTR 0xc0010241/* Fam 10h MSRs */ #define MSR_FAM10H_MMIO_CONF_BASE 0xc0010058
arch/x86/kernel/cpu/perf_event_amd.c +147 -20
279 lines not on this page · raw message
Signature · 1 line
1.7.9.5
Re: [PATCH RESEND V5 6/6] perf, amd: Enable northbridge performance counters on AMD family 15h
Quoted 43 lines · Jacob Shin
On AMD family 15h processors, there are 4 new performance counters (in addition to 6 core performance counters) that can be used for counting northbridge events (i.e. DRAM accesses). Their bit fields are almost identical to the core performance counters. However, unlike the core performance counters, these MSRs are shared between multiple cores (that share the same northbridge). We will reuse the same code path as existing family 10h northbridge event constraints handler logic to enforce this sharing.
Signed-off-by Jacob Shin <[email protected]>--- arch/x86/include/asm/cpufeature.h | 2 + arch/x86/include/asm/perf_event.h | 9 ++ arch/x86/include/uapi/asm/msr-index.h | 2 + arch/x86/kernel/cpu/perf_event_amd.c | 167 +++++++++++++++++++++++++++++---- 4 files changed, 160 insertions(+), 20 deletions(-)arch/x86/include/asm/cpufeature.h +2
-0@@ -167,6 +167,7 @@#define X86_FEATURE_TBM (6*32+21)/* trailing bit manipulations */ #define X86_FEATURE_TOPOEXT (6*32+22)/* topology extensions CPUID leafs */ #define X86_FEATURE_PERFCTR_CORE (6*32+23)/* core performance counter extensions */ +#define X86_FEATURE_PERFCTR_NB (6*32+24)/* NB performance counter extensions */ /* * Auxiliary flags: Linux defined - For features scattered in various @@ -309,6 +310,7 @@ extern const char * const x86_power_flags[32];#define cpu_has_hypervisor boot_cpu_has(X86_FEATURE_HYPERVISOR)#define cpu_has_pclmulqdq boot_cpu_has(X86_FEATURE_PCLMULQDQ)#define cpu_has_perfctr_core boot_cpu_has(X86_FEATURE_PERFCTR_CORE) +#define cpu_has_perfctr_nb boot_cpu_has(X86_FEATURE_PERFCTR_NB)#define cpu_has_cx8 boot_cpu_has(X86_FEATURE_CX8)#define cpu_has_cx16 boot_cpu_has(X86_FEATURE_CX16)#define cpu_has_eager_fpu boot_cpu_has(X86_FEATURE_EAGER_FPU)arch/x86/include/asm/perf_event.h +0
-0@@ -29,9 +29,14 @@#define ARCH_PERFMON_EVENTSEL_INV (1ULL << 23)
arch/x86/include/asm/perf_event.h line 39
Quoted 10 lines
arch/x86/include/asm/perf_event.h +5
-0#define ARCH_PERFMON_EVENTSEL_CMASK 0xFF000000ULL +#define AMD64_EVENTSEL_INT_CORE_ENABLE (1ULL << 36)#define AMD64_EVENTSEL_GUESTONLY (1ULL << 40)#define AMD64_EVENTSEL_HOSTONLY (1ULL << 41) +#define AMD64_EVENTSEL_INT_CORE_SEL_SHIFT 37 +#define AMD64_EVENTSEL_INT_CORE_SEL_MASK \ + (0xFULL << AMD64_EVENTSEL_INT_CORE_SEL_SHIFT) +
Interestingly enough, this bitfield is not yet documented in the public BKDG from March 2012.
I assume it will be in the next rev.
Quoted 220 lines
arch/x86/include/asm/perf_event.h +4
-0#define AMD64_EVENTSEL_EVENT \ (ARCH_PERFMON_EVENTSEL_EVENT | (0x0FULL << 32))#define INTEL_ARCH_EVENT_MASK \@@ -46,8 +51,12 @@#define AMD64_RAW_EVENT_MASK \ (X86_RAW_EVENT_MASK | \ AMD64_EVENTSEL_EVENT) +#define AMD64_RAW_EVENT_MASK_NB \ + (AMD64_EVENTSEL_EVENT | \ + ARCH_PERFMON_EVENTSEL_UMASK)#define AMD64_NUM_COUNTERS 4#define AMD64_NUM_COUNTERS_CORE 6 +#define AMD64_NUM_COUNTERS_NB 4#define ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL 0x3c#define ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK (0x00 << 8)arch/x86/include/uapi/asm/msr-index.h +2
-0@@ -194,6 +194,8 @@/* Fam 15h MSRs */ #define MSR_F15H_PERF_CTL 0xc0010200#define MSR_F15H_PERF_CTR 0xc0010201 +#define MSR_F15H_NB_PERF_CTL 0xc0010240 +#define MSR_F15H_NB_PERF_CTR 0xc0010241/* Fam 10h MSRs */ #define MSR_FAM10H_MMIO_CONF_BASE 0xc0010058arch/x86/kernel/cpu/perf_event_amd.c +102
-16@@ -132,11 +132,14 @@ static u64 amd_pmu_event_map(int hw_event)return amd_perfmon_event_map[hw_event]; } +static struct event_constraint *amd_nb_event_constraint; +/* * Previously calculated offsets */ static unsigned int event_offsets[X86_PMC_IDX_MAX ]__read_mostly ;static unsigned int count_offsets[X86_PMC_IDX_MAX ]__read_mostly ; +static unsigned int rdpmc_indexes[X86_PMC_IDX_MAX ]__read_mostly ;/* * Legacy CPUs: @@ -144,10 +147,14 @@ static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;* * CPUs with core performance counter extensions: * 6 counters starting at 0xc0010200 each offset by 2 +* +* CPUs with north bridge performance counter extensions: +* 4 additional counters starting at 0xc0010240 each offset by 2 +* (indexed right above either one of the above core counters) */ static inline int amd_pmu_addr_offset (int index,int eventsel) { -int offset; +int offset, first, base;if (!index)return index;@@ -160,7 +167,23 @@ static inline int amd_pmu_addr_offset(int index, int eventsel)if (offset)return offset; -if (!cpu_has_perfctr_core) +if (amd_nb_event_constraint && +test_bit (index, amd_nb_event_constraint->idxmsk)) { +/* +* calculate the offset of NB counters with respect to +* base eventsel or perfctr +*/ + + first =find_first_bit (amd_nb_event_constraint->idxmsk, +X86_PMC_IDX_MAX ); + +if (eventsel) + base =MSR_F15H_NB_PERF_CTL - x86_pmu.eventsel; +else + base =MSR_F15H_NB_PERF_CTR - x86_pmu.perfctr; + + offset = base + ((index - first) <<1 ); + }else if (!cpu_has_perfctr_core) offset = index;else offset = index <<1 ;@@ -175,24 +198,36 @@ static inline int amd_pmu_addr_offset(int index, int eventsel)static inline int amd_pmu_rdpmc_index (int index) { -return index; -} +int ret, first; -static int amd_pmu_hw_config (struct perf_event *event) -{ -int ret; +if (!index) +return index; -/* pass precise event sampling to ibs: */ -if (event->attr.precise_ip &&get_ibs_caps ()) -return -ENOENT ; + ret = rdpmc_indexes[index]; - ret =x86_pmu_hw_config (event);if (ret)return ret; -if (has_branch_stack (event)) -return -EOPNOTSUPP ; +if (amd_nb_event_constraint && +test_bit (index, amd_nb_event_constraint->idxmsk)) { +/* +* according to the mnual, ECX value of the NB counters is +* the index of the NB counter (0, 1, 2 or 3) plus 6 +*/ + + first =find_first_bit (amd_nb_event_constraint->idxmsk, +X86_PMC_IDX_MAX ); + ret = index - first +6 ; + }else + ret = index; + + rdpmc_indexes[index] = ret; +return ret; +} + +static int amd_core_hw_config (struct perf_event *event) +{if (event->attr.exclude_host && event->attr.exclude_guest)/* * When HO == GO == 1 the hardware treats that as GO == HO == 0 @@ -206,10 +241,29 @@ static int amd_pmu_hw_config(struct perf_event *event)else if (event->attr.exclude_guest) event->hw.config |=AMD64_EVENTSEL_HOSTONLY ; -if (event->attr.type !=PERF_TYPE_RAW ) -return 0 ; +return 0 ; +} - event->hw.config |= event->attr.config &AMD64_RAW_EVENT_MASK ; +/* +* NB counters do not support the following event select bits: +* Host/Guest only +* Counter mask +* Invert counter mask +* Edge detect +* OS/User mode +*/ +static int amd_nb_hw_config (struct perf_event *event) +{ +if (event->attr.exclude_user || event->attr.exclude_kernel || + event->attr.exclude_host || event->attr.exclude_guest) +return -EINVAL ; + + event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR | +ARCH_PERFMON_EVENTSEL_OS ); + +if (event->hw.config & ~(AMD64_RAW_EVENT_MASK_NB | +ARCH_PERFMON_EVENTSEL_INT )) +return -EINVAL ;return 0 ; }@@ -227,6 +281,11 @@ static inline int amd_is_nb_event(struct hw_perf_event *hwc)return (hwc->config &0xe0 ) ==0xe0 ; } +static inline int amd_is_perfctr_nb_event (struct hw_perf_event *hwc) +{ +return amd_nb_event_constraint &&amd_is_nb_event (hwc); +} +static inline int amd_has_nb (struct cpu_hw_events *cpuc) {struct amd_nb *nb = cpuc->amd_nb;@@ -234,6 +293,30 @@ static inline int amd_has_nb(struct cpu_hw_events *cpuc)return nb && nb->nb_id !=-1 ; } +static int amd_pmu_hw_config (struct perf_event *event) +{ +int ret; + +/* pass precise event sampling to ibs: */ +if (event->attr.precise_ip &&get_ibs_caps ()) +return -ENOENT ; + +if (has_branch_stack (event)) +return -EOPNOTSUPP ; + + ret =x86_pmu_hw_config (event); +if (ret) +return ret; + +if (event->attr.type ==PERF_TYPE_RAW ) + event->hw.config |= event->attr.config &AMD64_RAW_EVENT_MASK ; + +if (amd_is_perfctr_nb_event (&event->hw)) +return amd_nb_hw_config (event); + +return amd_core_hw_config (event); +} +static void __amd_put_nb_event_constraints (struct cpu_hw_events *cpuc,struct perf_event *event) {@@ -254,6 +337,19 @@ static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc, } } +static void amd_nb_interrupt_hw_config (struct hw_perf_event *hwc) +{ +int core_id =cpu_data (smp_processor_id ()).cpu_core_id;
arch/x86/kernel/cpu/perf_event_amd.c static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc, · line 352
Quoted 10 lines
arch/x86/kernel/cpu/perf_event_amd.c +10
-0+ +/* deliver interrupts only to this core */ +if (hwc->config &ARCH_PERFMON_EVENTSEL_INT ) { + hwc->config |=AMD64_EVENTSEL_INT_CORE_ENABLE ; + hwc->config &= ~AMD64_EVENTSEL_INT_CORE_SEL_MASK ; + hwc->config |= (u64)(core_id) << +AMD64_EVENTSEL_INT_CORE_SEL_SHIFT ; + } +} +
Well, given the model that you are using, i.e., fused with the core PMU, then you can
supposedly measure NB events in per-thread mode. If the thread migrates from one CPU to another. Then the uncore interrupt has to follow. Otherwise, you may get an interrupt on CPU0 where there is no active events if the thread is now
running on CPU1, for instance. So this does not work well. I think for NB events,
you may want to disable per-thread support. This is how it's done for Intel uncore.
The NB interrupt is useful for:
- 64-bit virtualization of the HW counter
- sampling.
Now, I admit given the list NB events, it does not really make sense to sample on
those and hope to correlate to a meaningful address in a program. For all I know,
the actual cause of an event may come from a program that is not even running on the measured core (assuming you were doing: perf record -a -C 0 -e r01e0 sleep 10).
Next, I tried running a simple example:
$ perf record -a -C 0 -e r01e0 sleep 10
But, there is nothing running on CPU0, so it's idle and even goes low-power. Yet, if on CPU1 there is load causing memory traffic and therefore firing the event.
First, in your setup, you're using the current CPU and NOT the target CPU of the event. Even though, you want to measure CPU0, you can be invoking perf_event_open()
from CPU3. Unless that setup phase is guaranteed to ALWAYS run on the target CPU for the event, here CPU0, that setup won't work.
I am using a simple memory benchmark called triad for which I can pin where the memory init (-i) and execution (-r) occur. So here I run everything on CPU0 and I sample only one NB event via CPU0. First, I count the actual number of events in per-process mode. This gives me a ballpark figure for a sampling period.
$ perf stat --no-big-num --pfm-event dram_accesses:dct0_page_miss
triad -i 0 -r 0
62302388 dram_accesses:dct0_page_miss
6.099754492 seconds time elapsedWant about 1000 samples, so I use -c 62302388/1000
$ perf record -a -C 0 -c 62302 --pfm-event
dram_accesses:dct0_page_miss triad -i 0 -r 0
$ perf report -D | tail -6
dram_accesses:dct0_page_miss stats:
TOTAL events: 3451
MMAP events: 2197
COMM events: 254
EXIT events: 2
SAMPLE events: 998Got about a 1000 samples. All is good.
Now, I modify the setup by forcing triad to init and run on CPU3 but I am still measuring
from CPU0:
$ perf record -a -C 0 -c 62302 --pfm-event
dram_accesses:dct0_page_miss triad -i 3 -r 3 -l 3
$ perf report -D | tail -6
TOTAL events: 2792
MMAP events: 2201
COMM events: 255
THROTTLE events: 3
UNTHROTTLE events: 3
SAMPLE events: 330I have lost 2/3rd of the samples. Why?
NB counters are shared across all 4 cores (of my Fam15h CPU), so when the counter
overflows it will interrupt CPU0. But how come samples disappear?
At first, I thought it could be due to the same problem that exists on Intel uncore:
uncore interrupt does not wake up a core in low-power state. But that is not quite
the problem here. Instead, you have to look at throttling. Why is this run throttled?
Capturing about 1000 samples over 6s. We are very far from the default 100000 intrs/event/cpu/s. So something is wrong here. Looks like CPU0 may get a burst of NB interrupts and decides the throttle for a timer tick or so. It does not
happen a lot but enough to cause 2/3rd of the sample to disappear. If I raise the threshold to 1000000 (million), then the problem goes away:
$ perf report -D | tail -6
dram_accesses:dct0_page_miss stats:
TOTAL events: 3296
MMAP events: 2037
COMM events: 249
SAMPLE events: 1010So something is broken with NB interrupts and I don't quite know what it is.
But in general, I remain convinced that unless you have broadcast interrupt and perf_event is modified to handle this case, sampling on the NB is pointless. So we might as well disable this and keep counting. But that does still require working NB interrupts. Worst case, we can do like on Intel and use hrtimer-based polling.
Quoted 102 lines
arch/x86/kernel/cpu/perf_event_amd.c +35
-5/* * AMD64 NorthBridge events need special treatment because * counter access needs to be synchronized across all cores @@ -299,6 +395,12 @@ __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *evstruct perf_event *old;int idx, new =-1 ; +if (!c) + c = &unconstrained; + +if (cpuc->is_fake) +return c; +/* * detect if already present, if so reuse * @@ -335,6 +437,9 @@ __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *evif (new ==-1 )return &emptyconstraint; +if (amd_is_perfctr_nb_event (hwc)) +amd_nb_interrupt_hw_config (hwc); +return &nb->event_constraints[new]; }@@ -434,7 +539,8 @@ amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)if (!(amd_has_nb (cpuc) &&amd_is_nb_event (&event->hw)))return &unconstrained; -return __amd_get_nb_event_constraints (cpuc, event, &unconstrained); +return __amd_get_nb_event_constraints (cpuc, event, + amd_nb_event_constraint); } static void amd_put_event_constraints(struct cpu_hw_events *cpuc,@@ -533,6 +639,9 @@ static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09,static struct event_constraint amd_f15_PMC50 =EVENT_CONSTRAINT (0 ,0x3F ,0 );static struct event_constraint amd_f15_PMC53 =EVENT_CONSTRAINT (0 ,0x38 ,0 ); +static struct event_constraint amd_NBPMC96 =EVENT_CONSTRAINT (0 ,0x3C0 ,0 ); +static struct event_constraint amd_NBPMC74 =EVENT_CONSTRAINT (0 ,0xF0 ,0 ); +static struct event_constraint *amd_get_event_constraints_f15h (struct cpu_hw_events *cpuc,struct perf_event *event) {@@ -598,8 +707,8 @@ amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, struct perf_event *evreturn &amd_f15_PMC20; }case AMD_EVENT_NB : -/* not yet implemented */ -return &emptyconstraint; +return __amd_get_nb_event_constraints (cpuc, event, + amd_nb_event_constraint);default :return &emptyconstraint; }@@ -647,7 +756,7 @@ static __initconst const struct x86_pmu amd_pmu = {static int setup_event_constraints (void ) { -if (boot_cpu_data.x86 >=0x15 ) +if (boot_cpu_data.x86 ==0x15 ) x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;return 0 ; }@@ -677,6 +786,23 @@ static int setup_perfctr_core(void)return 0 ; } +static int setup_perfctr_nb (void ) +{ +if (!cpu_has_perfctr_nb) +return -ENODEV ; + + x86_pmu.num_counters +=AMD64_NUM_COUNTERS_NB ; + +if (cpu_has_perfctr_core) + amd_nb_event_constraint = &amd_NBPMC96; +else + amd_nb_event_constraint = &amd_NBPMC74; + +printk (KERN_INFO "perf: AMD northbridge performance counters detected\n" ); + +return 0 ; +} +__init int amd_pmu_init (void ) {/* Performance-monitoring supported from K7 and later: */ @@ -687,6 +813,7 @@ __init int amd_pmu_init(void)setup_event_constraints ();setup_perfctr_core (); +setup_perfctr_nb ();/* Events are common for all AMDs */ memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, --1.7.9.5
Quoted 4 lines · Jacob Shin
The following patchset enables 4 additional performance counters in AMD family 15h processors that count northbridge events -- such as number of DRAM accesses.
In order for me to test this patch set more thoroughly it would help if you could also provide me a patch to add the Fam15h uncore events to libpfm4. In the past, Robert Richter took care of this. I hope you can fill his role for this. So please, if you could send me the patch quickly, that would help the review of your patch.
Thanks.
Quoted 58 lines
This patchset is based on previous work done by Robert Richter <[email protected]> :
https://lkml.org/lkml/2012/6/19/324
The main differences are:
* The northbridge counters are indexed contiguously right above the core performance counters.
* MSR address offset calculations are moved to architecture specific files.
* Interrups are set up to be delivered only to a single core.
V5:
Rebased against latest tipV4:
* Moved interrupt core select set up back to event constraints function, sicne during ->hw_config time we do not yet know on which CPU the the event will run on.
* Tested on and made minor revisions to make sure that the patchset is compatible with upcoming AMD Family 16h processors, and will support core and NB counters without any further patches.V3:
Addressed the following feedback/comments from Robert's review
* https://lkml.org/lkml/2012/11/16/484
* https://lkml.org/lkml/2012/11/26/162V2:
Separate out Robert's patches, and add properly ordered certificate of origins.Jacob Shin (4):
perf, amd: Use proper naming scheme for AMD bit field definitions perf, x86: Move MSR address offset calculation to architecture
specific files
perf, x86: Allow for architecture specific RDPMC indexes perf, amd: Enable northbridge performance counters on AMD family 15hRobert Richter (2):
perf, amd: Rework northbridge event constraints handler perf, amd: Generalize northbridge constraints code for family 15harch/x86/include/asm/cpufeature.h | 2 + arch/x86/include/asm/perf_event.h | 13 +- arch/x86/include/uapi/asm/msr-index.h | 2 + arch/x86/kernel/cpu/perf_event.c | 2 +- arch/x86/kernel/cpu/perf_event.h | 25 ++- arch/x86/kernel/cpu/perf_event_amd.c | 318 +++++++++++++++++++++++++-------- 6 files changed, 268 insertions(+), 94 deletions(-)--
1.7.9.5
Quoted 10 lines · Stephane Eranian
Quoted 4 lines · Jacob Shin
The following patchset enables 4 additional performance counters in AMD family 15h processors that count northbridge events -- such as number of DRAM accesses.
In order for me to test this patch set more thoroughly it would help if you could also provide me a patch to add the Fam15h uncore events to libpfm4. In the past, Robert Richter took care of this. I hope you can fill his role for this. So please, if you could send me the patch quickly, that would help the review of your patch.
Hi Stephane,
Here is the corresponding libpfm4 patch. Thank you for taking the time to review the patchset. I hope this helps .. If we can get AMD related perf kernel side patchsets to upstream, I will be more than happy to support AMD related libpfm4 efforts going forward.
Thanks!
Quoted 1 line
From
47d3267dfa24b9071c76f4a22bd059b0e4032002Mon Sep 17 00:00:00 2001
From: Jacob Shin <[email protected]>
Date: Thu, 24 Jan 2013 15:37:37 -0600
Subject: [PATCH 1/1] Add AMD Family 15h northbridge performance events
libpfm4 side support for the following Linux kernel patchset:
http://lkml.org/lkml/2013/1/10/450
Reference -- BIOS and Kernel Developer Guide (BKDG) for AMD Family 15h
Models 00h-0Fh Processors:
http://support.amd.com/us/Processor_TechDocs/42301_15h_Mod_00h-0Fh_BKDG.pdf
--- lib/events/amd64_events_fam15h.h | 155 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 155 insertions(+)
lib/events/amd64_events_fam15h.h +155 -0
@@ -752,6 +752,126 @@ static const amd64_umask_t amd64_fam15h_l2_prefetcher_trigger_events[]={ }, }; +static const amd64_umask_t amd64_fam15h_dram_accesses[]={ + { .uname ="DCT0_PAGE_HIT" , + .udesc ="DCT0 Page hit" , + .ucode =0x1 , + }, + { .uname ="DCT0_PAGE_MISS" , + .udesc ="DCT0 Page Miss" , + .ucode =0x2 , + }, + { .uname ="DCT0_PAGE_CONFLICT" , + .udesc ="DCT0 Page Conflict" , + .ucode =0x4 , + }, + { .uname ="DCT1_PAGE_HIT" , + .udesc ="DCT1 Page hit" , + .ucode =0x8 , + }, + { .uname ="DCT1_PAGE_MISS" , + .udesc ="DCT1 Page Miss" , + .ucode =0x10 , + }, + { .uname ="DCT1_PAGE_CONFLICT" , + .udesc ="DCT1 Page Conflict" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_dram_controller_page_table_overflows[]={ + { .uname ="DCT0_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT0 Page Table Overflow" , + .ucode =0x1 , + }, + { .uname ="DCT1_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT1 Page Table Overflow" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_dram_command_slots_missed[]={ + { .uname ="DCT0_COMMAND_SLOTS_MISSED" , + .udesc ="DCT0 Command Slots Missed (in MemClks)" , + .ucode =0x1 , + }, + { .uname ="DCT1_COMMAND_SLOTS_MISSED" , + .udesc ="DCT1 Command Slots Missed (in MemClks)" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_turnarounds[]={ + { .uname ="DCT0_DIMM_TURNAROUND" , + .udesc ="DCT0 DIMM (chip select) turnaround" , + .ucode =0x1 , + }, + { .uname ="DCT0_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT0 Read to write turnaround" , + .ucode =0x2 , + }, + { .uname ="DCT0_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT0 Write to read turnaround" , + .ucode =0x4 , + }, + { .uname ="DCT1_DIMM_TURNAROUND" , + .udesc ="DCT1 DIMM (chip select) turnaround" , + .ucode =0x8 , + }, + { .uname ="DCT1_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT1 Read to write turnaround" , + .ucode =0x10 , + }, + { .uname ="DCT1_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT1 Write to read turnaround" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_bypass_counter_saturation[]={ + { .uname ="MEMORY_CONTROLLER_HIGH_PRIORITY_BYPASS" , + .udesc ="Memory controller high priority bypass" , + .ucode =0x1 , + }, + { .uname ="MEMORY_CONTROLLER_MEDIUM_PRIORITY_BYPASS" , + .udesc ="Memory controller medium priority bypass" , + .ucode =0x2 , + }, + { .uname ="DCT0_DCQ_BYPASS" , + .udesc ="DCT0 DCQ bypass" , + .ucode =0x4 , + }, + { .uname ="DCT1_DCQ_BYPASS" , + .udesc ="DCT1 DCQ bypass" , + .ucode =0x8 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0xf , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; +static const amd64_entry_t amd64_fam15h_pe[]={ { .name ="DISPATCHED_FPU_OPS" , .desc ="FPU Pipe Assignment" ,
@@ -1256,4 +1376,39 @@ static const amd64_entry_t amd64_fam15h_pe[]={ .modmsk = AMD64_FAM15H_ATTRS, .code = 0x1d8, }, +{ .name = "DRAM_ACCESSES", + .desc = "DRAM Accesses", + .code = 0xe0, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_accesses), + .ngrp = 1, + .umasks = amd64_fam15h_dram_accesses, +}, +{ .name = "DRAM_CONTROLLER_PAGE_TABLE_OVERFLOWS", + .desc = "DRAM Controller Page Table Overflows", + .code = 0xe1, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_controller_page_table_overflows), + .ngrp = 1, + .umasks = amd64_fam15h_dram_controller_page_table_overflows, +}, +{ .name = "MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED", + .desc = "Memory Controller DRAM Command Slots Missed", + .code = 0xe2, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_dram_command_slots_missed), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_dram_command_slots_missed, +}, +{ .name = "MEMORY_CONTROLLER_TURNAROUNDS", + .desc = "Memory Controller Turnarounds", + .code = 0xe3, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_turnarounds), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_turnarounds, +}, +{ .name = "MEMORY_CONTROLLER_BYPASS_COUNTER_SATURATION", + .desc = "Memory Controller Bypass Counter Saturation", + .code = 0xe4, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_bypass_counter_saturation), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_bypass_counter_saturation, +}, };
Signature · 1 line
1.7.9.5
Hi Jacob,
I will apply this patch to libpfm4.
But I have a question. Why aren't the other uncore events included here as well? I am talking about the events listed in BKDG sections 3.16.2 to 3.16.6? Are those NOT supported by your kernel patchset?
Quoted 224 lines · Jacob Shin
Quoted 10 lines · Stephane Eranian
Quoted 4 lines · Jacob Shin
The following patchset enables 4 additional performance counters in AMD family 15h processors that count northbridge events -- such as number of DRAM accesses.
In order for me to test this patch set more thoroughly it would help if you could also provide me a patch to add the Fam15h uncore events to libpfm4. In the past, Robert Richter took care of this. I hope you can fill his role for this. So please, if you could send me the patch quickly, that would help the review of your patch.
Hi Stephane,
Here is the corresponding libpfm4 patch. Thank you for taking the time to review the patchset. I hope this helps .. If we can get AMD related perf kernel side patchsets to upstream, I will be more than happy to support AMD related libpfm4 efforts going forward.
Thanks!
Quoted 1 line
From
47d3267dfa24b9071c76f4a22bd059b0e4032002Mon Sep 17 00:00:00 2001From: Jacob Shin <[email protected]>
Date: Thu, 24 Jan 2013 15:37:37 -0600
Subject: [PATCH 1/1] Add AMD Family 15h northbridge performance eventslibpfm4 side support for the following Linux kernel patchset:
http://lkml.org/lkml/2013/1/10/450Reference -- BIOS and Kernel Developer Guide (BKDG) for AMD Family 15h
Models 00h-0Fh Processors:
http://support.amd.com/us/Processor_TechDocs/42301_15h_Mod_00h-0Fh_BKDG.pdfSigned-off-by Jacob Shin <[email protected]>--- lib/events/amd64_events_fam15h.h | 155 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 155 insertions(+)lib/events/amd64_events_fam15h.h +155
-1@@ -752,6 +752,126 @@ static const amd64_umask_t amd64_fam15h_l2_prefetcher_trigger_events[]={ }, }; +static const amd64_umask_t amd64_fam15h_dram_accesses[]={ + { .uname ="DCT0_PAGE_HIT" , + .udesc ="DCT0 Page hit" , + .ucode =0x1 , + }, + { .uname ="DCT0_PAGE_MISS" , + .udesc ="DCT0 Page Miss" , + .ucode =0x2 , + }, + { .uname ="DCT0_PAGE_CONFLICT" , + .udesc ="DCT0 Page Conflict" , + .ucode =0x4 , + }, + { .uname ="DCT1_PAGE_HIT" , + .udesc ="DCT1 Page hit" , + .ucode =0x8 , + }, + { .uname ="DCT1_PAGE_MISS" , + .udesc ="DCT1 Page Miss" , + .ucode =0x10 , + }, + { .uname ="DCT1_PAGE_CONFLICT" , + .udesc ="DCT1 Page Conflict" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_dram_controller_page_table_overflows[]={ + { .uname ="DCT0_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT0 Page Table Overflow" , + .ucode =0x1 , + }, + { .uname ="DCT1_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT1 Page Table Overflow" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_dram_command_slots_missed[]={ + { .uname ="DCT0_COMMAND_SLOTS_MISSED" , + .udesc ="DCT0 Command Slots Missed (in MemClks)" , + .ucode =0x1 , + }, + { .uname ="DCT1_COMMAND_SLOTS_MISSED" , + .udesc ="DCT1 Command Slots Missed (in MemClks)" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_turnarounds[]={ + { .uname ="DCT0_DIMM_TURNAROUND" , + .udesc ="DCT0 DIMM (chip select) turnaround" , + .ucode =0x1 , + }, + { .uname ="DCT0_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT0 Read to write turnaround" , + .ucode =0x2 , + }, + { .uname ="DCT0_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT0 Write to read turnaround" , + .ucode =0x4 , + }, + { .uname ="DCT1_DIMM_TURNAROUND" , + .udesc ="DCT1 DIMM (chip select) turnaround" , + .ucode =0x8 , + }, + { .uname ="DCT1_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT1 Read to write turnaround" , + .ucode =0x10 , + }, + { .uname ="DCT1_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT1 Write to read turnaround" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_bypass_counter_saturation[]={ + { .uname ="MEMORY_CONTROLLER_HIGH_PRIORITY_BYPASS" , + .udesc ="Memory controller high priority bypass" , + .ucode =0x1 , + }, + { .uname ="MEMORY_CONTROLLER_MEDIUM_PRIORITY_BYPASS" , + .udesc ="Memory controller medium priority bypass" , + .ucode =0x2 , + }, + { .uname ="DCT0_DCQ_BYPASS" , + .udesc ="DCT0 DCQ bypass" , + .ucode =0x4 , + }, + { .uname ="DCT1_DCQ_BYPASS" , + .udesc ="DCT1 DCQ bypass" , + .ucode =0x8 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0xf , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; +static const amd64_entry_t amd64_fam15h_pe[]={ { .name ="DISPATCHED_FPU_OPS" , .desc ="FPU Pipe Assignment" ,@@ -1256,4 +1376,39 @@ static const amd64_entry_t amd64_fam15h_pe[]={ .modmsk = AMD64_FAM15H_ATTRS, .code = 0x1d8, }, +{ .name = "DRAM_ACCESSES", + .desc = "DRAM Accesses", + .code = 0xe0, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_accesses), + .ngrp = 1, + .umasks = amd64_fam15h_dram_accesses, +}, +{ .name = "DRAM_CONTROLLER_PAGE_TABLE_OVERFLOWS", + .desc = "DRAM Controller Page Table Overflows", + .code = 0xe1, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_controller_page_table_overflows), + .ngrp = 1, + .umasks = amd64_fam15h_dram_controller_page_table_overflows, +}, +{ .name = "MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED", + .desc = "Memory Controller DRAM Command Slots Missed", + .code = 0xe2, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_dram_command_slots_missed), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_dram_command_slots_missed, +}, +{ .name = "MEMORY_CONTROLLER_TURNAROUNDS", + .desc = "Memory Controller Turnarounds", + .code = 0xe3, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_turnarounds), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_turnarounds, +}, +{ .name = "MEMORY_CONTROLLER_BYPASS_COUNTER_SATURATION", + .desc = "Memory Controller Bypass Counter Saturation", + .code = 0xe4, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_bypass_counter_saturation), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_bypass_counter_saturation, +}, }; --1.7.9.5
------------------------------------------------------------------------------
Master Visual Studio, SharePoint, SQL, ASP.NET, C# 2012, HTML5, CSS, MVC, Windows 8 Apps, JavaScript and much more. Keep your skills current with LearnDevNow - 3,200 step-by-step video tutorials by Microsoft MVPs and experts. ON SALE this month only -- learn more at:
http://p.sf.net/sfu/learnnow-d2d
_______________________________________________
perfmon2-devel mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
Quoted 7 lines · Stephane Eranian
Hi Jacob,
I will apply this patch to libpfm4.
But I have a question. Why aren't the other uncore events included here as well? I am talking about the events listed in BKDG sections 3.16.2 to 3.16.6? Are those NOT supported by your kernel patchset?
Oh, you are right, they are supported. I'm not sure why I overlooked them. I will send out a V2 that includes all of those events as well.
Sorry about that.
-Jacob
Quoted 228 lines · Jacob Shin
Quoted 224 lines · Stephane Eranian
Quoted 10 lines · Jacob Shin
Quoted 4 lines
The following patchset enables 4 additional performance counters in AMD family 15h processors that count northbridge events -- such as number of DRAM accesses.
In order for me to test this patch set more thoroughly it would help if you could also provide me a patch to add the Fam15h uncore events to libpfm4. In the past, Robert Richter took care of this. I hope you can fill his role for this. So please, if you could send me the patch quickly, that would help the review of your patch.
Hi Stephane,
Here is the corresponding libpfm4 patch. Thank you for taking the time to review the patchset. I hope this helps .. If we can get AMD related perf kernel side patchsets to upstream, I will be more than happy to support AMD related libpfm4 efforts going forward.
Thanks!
Quoted 1 line
From
47d3267dfa24b9071c76f4a22bd059b0e4032002Mon Sep 17 00:00:00 2001From: Jacob Shin <[email protected]>
Date: Thu, 24 Jan 2013 15:37:37 -0600
Subject: [PATCH 1/1] Add AMD Family 15h northbridge performance eventslibpfm4 side support for the following Linux kernel patchset:
http://lkml.org/lkml/2013/1/10/450Reference -- BIOS and Kernel Developer Guide (BKDG) for AMD Family 15h
Models 00h-0Fh Processors:
http://support.amd.com/us/Processor_TechDocs/42301_15h_Mod_00h-0Fh_BKDG.pdfSigned-off-by Jacob Shin <[email protected]>--- lib/events/amd64_events_fam15h.h | 155 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 155 insertions(+)lib/events/amd64_events_fam15h.h +155
-1@@ -752,6 +752,126 @@ static const amd64_umask_t amd64_fam15h_l2_prefetcher_trigger_events[]={ }, }; +static const amd64_umask_t amd64_fam15h_dram_accesses[]={ + { .uname ="DCT0_PAGE_HIT" , + .udesc ="DCT0 Page hit" , + .ucode =0x1 , + }, + { .uname ="DCT0_PAGE_MISS" , + .udesc ="DCT0 Page Miss" , + .ucode =0x2 , + }, + { .uname ="DCT0_PAGE_CONFLICT" , + .udesc ="DCT0 Page Conflict" , + .ucode =0x4 , + }, + { .uname ="DCT1_PAGE_HIT" , + .udesc ="DCT1 Page hit" , + .ucode =0x8 , + }, + { .uname ="DCT1_PAGE_MISS" , + .udesc ="DCT1 Page Miss" , + .ucode =0x10 , + }, + { .uname ="DCT1_PAGE_CONFLICT" , + .udesc ="DCT1 Page Conflict" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_dram_controller_page_table_overflows[]={ + { .uname ="DCT0_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT0 Page Table Overflow" , + .ucode =0x1 , + }, + { .uname ="DCT1_PAGE_TABLE_OVERFLOW" , + .udesc ="DCT1 Page Table Overflow" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_dram_command_slots_missed[]={ + { .uname ="DCT0_COMMAND_SLOTS_MISSED" , + .udesc ="DCT0 Command Slots Missed (in MemClks)" , + .ucode =0x1 , + }, + { .uname ="DCT1_COMMAND_SLOTS_MISSED" , + .udesc ="DCT1 Command Slots Missed (in MemClks)" , + .ucode =0x2 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3 , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_turnarounds[]={ + { .uname ="DCT0_DIMM_TURNAROUND" , + .udesc ="DCT0 DIMM (chip select) turnaround" , + .ucode =0x1 , + }, + { .uname ="DCT0_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT0 Read to write turnaround" , + .ucode =0x2 , + }, + { .uname ="DCT0_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT0 Write to read turnaround" , + .ucode =0x4 , + }, + { .uname ="DCT1_DIMM_TURNAROUND" , + .udesc ="DCT1 DIMM (chip select) turnaround" , + .ucode =0x8 , + }, + { .uname ="DCT1_READ_TO_WRITE_TURNAROUND" , + .udesc ="DCT1 Read to write turnaround" , + .ucode =0x10 , + }, + { .uname ="DCT1_WRITE_TO_READ_TURNAROUND" , + .udesc ="DCT1 Write to read turnaround" , + .ucode =0x20 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0x3f , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; + +static const amd64_umask_t amd64_fam15h_memory_controller_bypass_counter_saturation[]={ + { .uname ="MEMORY_CONTROLLER_HIGH_PRIORITY_BYPASS" , + .udesc ="Memory controller high priority bypass" , + .ucode =0x1 , + }, + { .uname ="MEMORY_CONTROLLER_MEDIUM_PRIORITY_BYPASS" , + .udesc ="Memory controller medium priority bypass" , + .ucode =0x2 , + }, + { .uname ="DCT0_DCQ_BYPASS" , + .udesc ="DCT0 DCQ bypass" , + .ucode =0x4 , + }, + { .uname ="DCT1_DCQ_BYPASS" , + .udesc ="DCT1 DCQ bypass" , + .ucode =0x8 , + }, + { .uname ="ALL" , + .udesc ="All sub-events selected" , + .ucode =0xf , + .uflags =AMD64_FL_NCOMBO |AMD64_FL_DFL , + }, +}; +static const amd64_entry_t amd64_fam15h_pe[]={ { .name ="DISPATCHED_FPU_OPS" , .desc ="FPU Pipe Assignment" ,@@ -1256,4 +1376,39 @@ static const amd64_entry_t amd64_fam15h_pe[]={ .modmsk = AMD64_FAM15H_ATTRS, .code = 0x1d8, }, +{ .name = "DRAM_ACCESSES", + .desc = "DRAM Accesses", + .code = 0xe0, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_accesses), + .ngrp = 1, + .umasks = amd64_fam15h_dram_accesses, +}, +{ .name = "DRAM_CONTROLLER_PAGE_TABLE_OVERFLOWS", + .desc = "DRAM Controller Page Table Overflows", + .code = 0xe1, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_dram_controller_page_table_overflows), + .ngrp = 1, + .umasks = amd64_fam15h_dram_controller_page_table_overflows, +}, +{ .name = "MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED", + .desc = "Memory Controller DRAM Command Slots Missed", + .code = 0xe2, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_dram_command_slots_missed), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_dram_command_slots_missed, +}, +{ .name = "MEMORY_CONTROLLER_TURNAROUNDS", + .desc = "Memory Controller Turnarounds", + .code = 0xe3, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_turnarounds), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_turnarounds, +}, +{ .name = "MEMORY_CONTROLLER_BYPASS_COUNTER_SATURATION", + .desc = "Memory Controller Bypass Counter Saturation", + .code = 0xe4, + .numasks = LIBPFM_ARRAY_SIZE(amd64_fam15h_memory_controller_bypass_counter_saturation), + .ngrp = 1, + .umasks = amd64_fam15h_memory_controller_bypass_counter_saturation, +}, }; --1.7.9.5
------------------------------------------------------------------------------
Master Visual Studio, SharePoint, SQL, ASP.NET, C# 2012, HTML5, CSS, MVC, Windows 8 Apps, JavaScript and much more. Keep your skills current with LearnDevNow - 3,200 step-by-step video tutorials by Microsoft MVPs and experts. ON SALE this month only -- learn more at:
http://p.sf.net/sfu/learnnow-d2d
_______________________________________________
perfmon2-devel mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel