1 #include <linux/perf_event.h>
2 #include <linux/types.h>
3 #include <linux/init.h>
4 #include <linux/slab.h>
5 #include <asm/apicdef.h>
7 #include "perf_event.h"
9 static __initconst const u64 amd_hw_cache_event_ids
10 [PERF_COUNT_HW_CACHE_MAX]
11 [PERF_COUNT_HW_CACHE_OP_MAX]
12 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
16 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
17 [ C(RESULT_MISS) ] = 0x0141, /* Data Cache Misses */
20 [ C(RESULT_ACCESS) ] = 0x0142, /* Data Cache Refills :system */
21 [ C(RESULT_MISS) ] = 0,
23 [ C(OP_PREFETCH) ] = {
24 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts */
25 [ C(RESULT_MISS) ] = 0x0167, /* Data Prefetcher :cancelled */
30 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches */
31 [ C(RESULT_MISS) ] = 0x0081, /* Instruction cache misses */
34 [ C(RESULT_ACCESS) ] = -1,
35 [ C(RESULT_MISS) ] = -1,
37 [ C(OP_PREFETCH) ] = {
38 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
39 [ C(RESULT_MISS) ] = 0,
44 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
45 [ C(RESULT_MISS) ] = 0x037E, /* L2 Cache Misses : IC+DC */
48 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback */
49 [ C(RESULT_MISS) ] = 0,
51 [ C(OP_PREFETCH) ] = {
52 [ C(RESULT_ACCESS) ] = 0,
53 [ C(RESULT_MISS) ] = 0,
58 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
59 [ C(RESULT_MISS) ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
62 [ C(RESULT_ACCESS) ] = 0,
63 [ C(RESULT_MISS) ] = 0,
65 [ C(OP_PREFETCH) ] = {
66 [ C(RESULT_ACCESS) ] = 0,
67 [ C(RESULT_MISS) ] = 0,
72 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes */
73 [ C(RESULT_MISS) ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
76 [ C(RESULT_ACCESS) ] = -1,
77 [ C(RESULT_MISS) ] = -1,
79 [ C(OP_PREFETCH) ] = {
80 [ C(RESULT_ACCESS) ] = -1,
81 [ C(RESULT_MISS) ] = -1,
86 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr. */
87 [ C(RESULT_MISS) ] = 0x00c3, /* Retired Mispredicted BI */
90 [ C(RESULT_ACCESS) ] = -1,
91 [ C(RESULT_MISS) ] = -1,
93 [ C(OP_PREFETCH) ] = {
94 [ C(RESULT_ACCESS) ] = -1,
95 [ C(RESULT_MISS) ] = -1,
100 [ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
101 [ C(RESULT_MISS) ] = 0x98e9, /* CPU Request to Memory, r */
104 [ C(RESULT_ACCESS) ] = -1,
105 [ C(RESULT_MISS) ] = -1,
107 [ C(OP_PREFETCH) ] = {
108 [ C(RESULT_ACCESS) ] = -1,
109 [ C(RESULT_MISS) ] = -1,
115 * AMD Performance Monitor K7 and later.
117 static const u64 amd_perfmon_event_map[] =
119 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
120 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
121 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x0080,
122 [PERF_COUNT_HW_CACHE_MISSES] = 0x0081,
123 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
124 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
125 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00d0, /* "Decoder empty" event */
126 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x00d1, /* "Dispatch stalls" event */
129 static u64 amd_pmu_event_map(int hw_event)
131 return amd_perfmon_event_map[hw_event];
134 static int amd_pmu_hw_config(struct perf_event *event)
136 int ret = x86_pmu_hw_config(event);
141 if (event->attr.exclude_host && event->attr.exclude_guest)
143 * When HO == GO == 1 the hardware treats that as GO == HO == 0
144 * and will count in both modes. We don't want to count in that
145 * case so we emulate no-counting by setting US = OS = 0.
147 event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
148 ARCH_PERFMON_EVENTSEL_OS);
149 else if (event->attr.exclude_host)
150 event->hw.config |= AMD_PERFMON_EVENTSEL_GUESTONLY;
151 else if (event->attr.exclude_guest)
152 event->hw.config |= AMD_PERFMON_EVENTSEL_HOSTONLY;
154 if (event->attr.type != PERF_TYPE_RAW)
157 event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
163 * AMD64 events are detected based on their event codes.
165 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
167 return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
170 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
172 return (hwc->config & 0xe0) == 0xe0;
175 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
177 struct amd_nb *nb = cpuc->amd_nb;
179 return nb && nb->nb_id != -1;
182 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
183 struct perf_event *event)
185 struct hw_perf_event *hwc = &event->hw;
186 struct amd_nb *nb = cpuc->amd_nb;
190 * only care about NB events
192 if (!(amd_has_nb(cpuc) && amd_is_nb_event(hwc)))
196 * need to scan whole list because event may not have
197 * been assigned during scheduling
199 * no race condition possible because event can only
200 * be removed on one CPU at a time AND PMU is disabled
203 for (i = 0; i < x86_pmu.num_counters; i++) {
204 if (nb->owners[i] == event) {
205 cmpxchg(nb->owners+i, event, NULL);
212 * AMD64 NorthBridge events need special treatment because
213 * counter access needs to be synchronized across all cores
214 * of a package. Refer to BKDG section 3.12
216 * NB events are events measuring L3 cache, Hypertransport
217 * traffic. They are identified by an event code >= 0xe00.
218 * They measure events on the NorthBride which is shared
219 * by all cores on a package. NB events are counted on a
220 * shared set of counters. When a NB event is programmed
221 * in a counter, the data actually comes from a shared
222 * counter. Thus, access to those counters needs to be
225 * We implement the synchronization such that no two cores
226 * can be measuring NB events using the same counters. Thus,
227 * we maintain a per-NB allocation table. The available slot
228 * is propagated using the event_constraint structure.
230 * We provide only one choice for each NB event based on
231 * the fact that only NB events have restrictions. Consequently,
232 * if a counter is available, there is a guarantee the NB event
233 * will be assigned to it. If no slot is available, an empty
234 * constraint is returned and scheduling will eventually fail
237 * Note that all cores attached the same NB compete for the same
238 * counters to host NB events, this is why we use atomic ops. Some
239 * multi-chip CPUs may have more than one NB.
241 * Given that resources are allocated (cmpxchg), they must be
242 * eventually freed for others to use. This is accomplished by
243 * calling amd_put_event_constraints().
245 * Non NB events are not impacted by this restriction.
247 static struct event_constraint *
248 amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
250 struct hw_perf_event *hwc = &event->hw;
251 struct amd_nb *nb = cpuc->amd_nb;
252 struct perf_event *old = NULL;
253 int max = x86_pmu.num_counters;
257 * if not NB event or no NB, then no constraints
259 if (!(amd_has_nb(cpuc) && amd_is_nb_event(hwc)))
260 return &unconstrained;
263 * detect if already present, if so reuse
265 * cannot merge with actual allocation
266 * because of possible holes
268 * event can already be present yet not assigned (in hwc->idx)
269 * because of successive calls to x86_schedule_events() from
270 * hw_perf_group_sched_in() without hw_perf_enable()
272 for (i = 0; i < max; i++) {
274 * keep track of first free slot
276 if (k == -1 && !nb->owners[i])
279 /* already present, reuse */
280 if (nb->owners[i] == event)
284 * not present, so grab a new slot
285 * starting either at:
287 if (hwc->idx != -1) {
288 /* previous assignment */
290 } else if (k != -1) {
291 /* start from free slot found */
295 * event not found, no slot found in
296 * first pass, try again from the
303 old = cmpxchg(nb->owners+i, NULL, event);
311 return &nb->event_constraints[i];
313 return &emptyconstraint;
316 static struct amd_nb *amd_alloc_nb(int cpu)
321 nb = kmalloc_node(sizeof(struct amd_nb), GFP_KERNEL | __GFP_ZERO,
329 * initialize all possible NB constraints
331 for (i = 0; i < x86_pmu.num_counters; i++) {
332 __set_bit(i, nb->event_constraints[i].idxmsk);
333 nb->event_constraints[i].weight = 1;
338 static int amd_pmu_cpu_prepare(int cpu)
340 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
342 WARN_ON_ONCE(cpuc->amd_nb);
344 if (boot_cpu_data.x86_max_cores < 2)
347 cpuc->amd_nb = amd_alloc_nb(cpu);
354 static void amd_pmu_cpu_starting(int cpu)
356 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
360 if (boot_cpu_data.x86_max_cores < 2)
363 nb_id = amd_get_nb_id(cpu);
364 WARN_ON_ONCE(nb_id == BAD_APICID);
366 for_each_online_cpu(i) {
367 nb = per_cpu(cpu_hw_events, i).amd_nb;
368 if (WARN_ON_ONCE(!nb))
371 if (nb->nb_id == nb_id) {
372 cpuc->kfree_on_online = cpuc->amd_nb;
378 cpuc->amd_nb->nb_id = nb_id;
379 cpuc->amd_nb->refcnt++;
382 static void amd_pmu_cpu_dead(int cpu)
384 struct cpu_hw_events *cpuhw;
386 if (boot_cpu_data.x86_max_cores < 2)
389 cpuhw = &per_cpu(cpu_hw_events, cpu);
392 struct amd_nb *nb = cpuhw->amd_nb;
394 if (nb->nb_id == -1 || --nb->refcnt == 0)
397 cpuhw->amd_nb = NULL;
401 static __initconst const struct x86_pmu amd_pmu = {
403 .handle_irq = x86_pmu_handle_irq,
404 .disable_all = x86_pmu_disable_all,
405 .enable_all = x86_pmu_enable_all,
406 .enable = x86_pmu_enable_event,
407 .disable = x86_pmu_disable_event,
408 .hw_config = amd_pmu_hw_config,
409 .schedule_events = x86_schedule_events,
410 .eventsel = MSR_K7_EVNTSEL0,
411 .perfctr = MSR_K7_PERFCTR0,
412 .event_map = amd_pmu_event_map,
413 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
414 .num_counters = AMD64_NUM_COUNTERS,
416 .cntval_mask = (1ULL << 48) - 1,
418 /* use highest bit to detect overflow */
419 .max_period = (1ULL << 47) - 1,
420 .get_event_constraints = amd_get_event_constraints,
421 .put_event_constraints = amd_put_event_constraints,
423 .cpu_prepare = amd_pmu_cpu_prepare,
424 .cpu_starting = amd_pmu_cpu_starting,
425 .cpu_dead = amd_pmu_cpu_dead,
430 #define AMD_EVENT_TYPE_MASK 0x000000F0ULL
432 #define AMD_EVENT_FP 0x00000000ULL ... 0x00000010ULL
433 #define AMD_EVENT_LS 0x00000020ULL ... 0x00000030ULL
434 #define AMD_EVENT_DC 0x00000040ULL ... 0x00000050ULL
435 #define AMD_EVENT_CU 0x00000060ULL ... 0x00000070ULL
436 #define AMD_EVENT_IC_DE 0x00000080ULL ... 0x00000090ULL
437 #define AMD_EVENT_EX_LS 0x000000C0ULL
438 #define AMD_EVENT_DE 0x000000D0ULL
439 #define AMD_EVENT_NB 0x000000E0ULL ... 0x000000F0ULL
442 * AMD family 15h event code/PMC mappings:
444 * type = event_code & 0x0F0:
446 * 0x000 FP PERF_CTL[5:3]
447 * 0x010 FP PERF_CTL[5:3]
448 * 0x020 LS PERF_CTL[5:0]
449 * 0x030 LS PERF_CTL[5:0]
450 * 0x040 DC PERF_CTL[5:0]
451 * 0x050 DC PERF_CTL[5:0]
452 * 0x060 CU PERF_CTL[2:0]
453 * 0x070 CU PERF_CTL[2:0]
454 * 0x080 IC/DE PERF_CTL[2:0]
455 * 0x090 IC/DE PERF_CTL[2:0]
458 * 0x0C0 EX/LS PERF_CTL[5:0]
459 * 0x0D0 DE PERF_CTL[2:0]
460 * 0x0E0 NB NB_PERF_CTL[3:0]
461 * 0x0F0 NB NB_PERF_CTL[3:0]
465 * 0x000 FP PERF_CTL[3], PERF_CTL[5:3] (*)
466 * 0x003 FP PERF_CTL[3]
467 * 0x004 FP PERF_CTL[3], PERF_CTL[5:3] (*)
468 * 0x00B FP PERF_CTL[3]
469 * 0x00D FP PERF_CTL[3]
470 * 0x023 DE PERF_CTL[2:0]
471 * 0x02D LS PERF_CTL[3]
472 * 0x02E LS PERF_CTL[3,0]
473 * 0x043 CU PERF_CTL[2:0]
474 * 0x045 CU PERF_CTL[2:0]
475 * 0x046 CU PERF_CTL[2:0]
476 * 0x054 CU PERF_CTL[2:0]
477 * 0x055 CU PERF_CTL[2:0]
478 * 0x08F IC PERF_CTL[0]
479 * 0x187 DE PERF_CTL[0]
480 * 0x188 DE PERF_CTL[0]
481 * 0x0DB EX PERF_CTL[5:0]
482 * 0x0DC LS PERF_CTL[5:0]
483 * 0x0DD LS PERF_CTL[5:0]
484 * 0x0DE LS PERF_CTL[5:0]
485 * 0x0DF LS PERF_CTL[5:0]
486 * 0x1D6 EX PERF_CTL[5:0]
487 * 0x1D8 EX PERF_CTL[5:0]
489 * (*) depending on the umask all FPU counters may be used
492 static struct event_constraint amd_f15_PMC0 = EVENT_CONSTRAINT(0, 0x01, 0);
493 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
494 static struct event_constraint amd_f15_PMC3 = EVENT_CONSTRAINT(0, 0x08, 0);
495 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT(0, 0x09, 0);
496 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
497 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
499 static struct event_constraint *
500 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, struct perf_event *event)
502 struct hw_perf_event *hwc = &event->hw;
503 unsigned int event_code = amd_get_event_code(hwc);
505 switch (event_code & AMD_EVENT_TYPE_MASK) {
507 switch (event_code) {
509 if (!(hwc->config & 0x0000F000ULL))
511 if (!(hwc->config & 0x00000F00ULL))
513 return &amd_f15_PMC3;
515 if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
517 return &amd_f15_PMC3;
521 return &amd_f15_PMC3;
523 return &amd_f15_PMC53;
526 case AMD_EVENT_EX_LS:
527 switch (event_code) {
534 return &amd_f15_PMC20;
536 return &amd_f15_PMC3;
538 return &amd_f15_PMC30;
540 return &amd_f15_PMC50;
543 case AMD_EVENT_IC_DE:
545 switch (event_code) {
549 return &amd_f15_PMC0;
550 case 0x0DB ... 0x0DF:
553 return &amd_f15_PMC50;
555 return &amd_f15_PMC20;
558 /* not yet implemented */
559 return &emptyconstraint;
561 return &emptyconstraint;
565 static __initconst const struct x86_pmu amd_pmu_f15h = {
566 .name = "AMD Family 15h",
567 .handle_irq = x86_pmu_handle_irq,
568 .disable_all = x86_pmu_disable_all,
569 .enable_all = x86_pmu_enable_all,
570 .enable = x86_pmu_enable_event,
571 .disable = x86_pmu_disable_event,
572 .hw_config = amd_pmu_hw_config,
573 .schedule_events = x86_schedule_events,
574 .eventsel = MSR_F15H_PERF_CTL,
575 .perfctr = MSR_F15H_PERF_CTR,
576 .event_map = amd_pmu_event_map,
577 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
578 .num_counters = AMD64_NUM_COUNTERS_F15H,
580 .cntval_mask = (1ULL << 48) - 1,
582 /* use highest bit to detect overflow */
583 .max_period = (1ULL << 47) - 1,
584 .get_event_constraints = amd_get_event_constraints_f15h,
585 /* nortbridge counters not yet implemented: */
587 .put_event_constraints = amd_put_event_constraints,
589 .cpu_prepare = amd_pmu_cpu_prepare,
590 .cpu_starting = amd_pmu_cpu_starting,
591 .cpu_dead = amd_pmu_cpu_dead,
595 __init int amd_pmu_init(void)
597 /* Performance-monitoring supported from K7 and later: */
598 if (boot_cpu_data.x86 < 6)
602 * If core performance counter extensions exists, it must be
603 * family 15h, otherwise fail. See x86_pmu_addr_offset().
605 switch (boot_cpu_data.x86) {
607 if (!cpu_has_perfctr_core)
609 x86_pmu = amd_pmu_f15h;
612 if (cpu_has_perfctr_core)
618 /* Events are common for all AMDs */
619 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids,
620 sizeof(hw_cache_event_ids));