1 #include <linux/perf_event.h>
2 #include <linux/export.h>
3 #include <linux/types.h>
4 #include <linux/init.h>
5 #include <linux/slab.h>
6 #include <linux/delay.h>
7 #include <asm/apicdef.h>
10 #include "../perf_event.h"
12 static DEFINE_PER_CPU(unsigned int, perf_nmi_counter);
14 static __initconst const u64 amd_hw_cache_event_ids
15 [PERF_COUNT_HW_CACHE_MAX]
16 [PERF_COUNT_HW_CACHE_OP_MAX]
17 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
21 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
22 [ C(RESULT_MISS) ] = 0x0141, /* Data Cache Misses */
25 [ C(RESULT_ACCESS) ] = 0,
26 [ C(RESULT_MISS) ] = 0,
28 [ C(OP_PREFETCH) ] = {
29 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts */
30 [ C(RESULT_MISS) ] = 0x0167, /* Data Prefetcher :cancelled */
35 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches */
36 [ C(RESULT_MISS) ] = 0x0081, /* Instruction cache misses */
39 [ C(RESULT_ACCESS) ] = -1,
40 [ C(RESULT_MISS) ] = -1,
42 [ C(OP_PREFETCH) ] = {
43 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
44 [ C(RESULT_MISS) ] = 0,
49 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
50 [ C(RESULT_MISS) ] = 0x037E, /* L2 Cache Misses : IC+DC */
53 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback */
54 [ C(RESULT_MISS) ] = 0,
56 [ C(OP_PREFETCH) ] = {
57 [ C(RESULT_ACCESS) ] = 0,
58 [ C(RESULT_MISS) ] = 0,
63 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses */
64 [ C(RESULT_MISS) ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
67 [ C(RESULT_ACCESS) ] = 0,
68 [ C(RESULT_MISS) ] = 0,
70 [ C(OP_PREFETCH) ] = {
71 [ C(RESULT_ACCESS) ] = 0,
72 [ C(RESULT_MISS) ] = 0,
77 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes */
78 [ C(RESULT_MISS) ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
81 [ C(RESULT_ACCESS) ] = -1,
82 [ C(RESULT_MISS) ] = -1,
84 [ C(OP_PREFETCH) ] = {
85 [ C(RESULT_ACCESS) ] = -1,
86 [ C(RESULT_MISS) ] = -1,
91 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr. */
92 [ C(RESULT_MISS) ] = 0x00c3, /* Retired Mispredicted BI */
95 [ C(RESULT_ACCESS) ] = -1,
96 [ C(RESULT_MISS) ] = -1,
98 [ C(OP_PREFETCH) ] = {
99 [ C(RESULT_ACCESS) ] = -1,
100 [ C(RESULT_MISS) ] = -1,
105 [ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
106 [ C(RESULT_MISS) ] = 0x98e9, /* CPU Request to Memory, r */
109 [ C(RESULT_ACCESS) ] = -1,
110 [ C(RESULT_MISS) ] = -1,
112 [ C(OP_PREFETCH) ] = {
113 [ C(RESULT_ACCESS) ] = -1,
114 [ C(RESULT_MISS) ] = -1,
119 static __initconst const u64 amd_hw_cache_event_ids_f17h
120 [PERF_COUNT_HW_CACHE_MAX]
121 [PERF_COUNT_HW_CACHE_OP_MAX]
122 [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
125 [C(RESULT_ACCESS)] = 0x0040, /* Data Cache Accesses */
126 [C(RESULT_MISS)] = 0xc860, /* L2$ access from DC Miss */
129 [C(RESULT_ACCESS)] = 0,
130 [C(RESULT_MISS)] = 0,
133 [C(RESULT_ACCESS)] = 0xff5a, /* h/w prefetch DC Fills */
134 [C(RESULT_MISS)] = 0,
139 [C(RESULT_ACCESS)] = 0x0080, /* Instruction cache fetches */
140 [C(RESULT_MISS)] = 0x0081, /* Instruction cache misses */
143 [C(RESULT_ACCESS)] = -1,
144 [C(RESULT_MISS)] = -1,
147 [C(RESULT_ACCESS)] = 0,
148 [C(RESULT_MISS)] = 0,
153 [C(RESULT_ACCESS)] = 0,
154 [C(RESULT_MISS)] = 0,
157 [C(RESULT_ACCESS)] = 0,
158 [C(RESULT_MISS)] = 0,
161 [C(RESULT_ACCESS)] = 0,
162 [C(RESULT_MISS)] = 0,
167 [C(RESULT_ACCESS)] = 0xff45, /* All L2 DTLB accesses */
168 [C(RESULT_MISS)] = 0xf045, /* L2 DTLB misses (PT walks) */
171 [C(RESULT_ACCESS)] = 0,
172 [C(RESULT_MISS)] = 0,
175 [C(RESULT_ACCESS)] = 0,
176 [C(RESULT_MISS)] = 0,
181 [C(RESULT_ACCESS)] = 0x0084, /* L1 ITLB misses, L2 ITLB hits */
182 [C(RESULT_MISS)] = 0xff85, /* L1 ITLB misses, L2 misses */
185 [C(RESULT_ACCESS)] = -1,
186 [C(RESULT_MISS)] = -1,
189 [C(RESULT_ACCESS)] = -1,
190 [C(RESULT_MISS)] = -1,
195 [C(RESULT_ACCESS)] = 0x00c2, /* Retired Branch Instr. */
196 [C(RESULT_MISS)] = 0x00c3, /* Retired Mispredicted BI */
199 [C(RESULT_ACCESS)] = -1,
200 [C(RESULT_MISS)] = -1,
203 [C(RESULT_ACCESS)] = -1,
204 [C(RESULT_MISS)] = -1,
209 [C(RESULT_ACCESS)] = 0,
210 [C(RESULT_MISS)] = 0,
213 [C(RESULT_ACCESS)] = -1,
214 [C(RESULT_MISS)] = -1,
217 [C(RESULT_ACCESS)] = -1,
218 [C(RESULT_MISS)] = -1,
224 * AMD Performance Monitor K7 and later, up to and including Family 16h:
226 static const u64 amd_perfmon_event_map[PERF_COUNT_HW_MAX] =
228 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
229 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
230 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x077d,
231 [PERF_COUNT_HW_CACHE_MISSES] = 0x077e,
232 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
233 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
234 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x00d0, /* "Decoder empty" event */
235 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x00d1, /* "Dispatch stalls" event */
239 * AMD Performance Monitor Family 17h and later:
241 static const u64 amd_f17h_perfmon_event_map[PERF_COUNT_HW_MAX] =
243 [PERF_COUNT_HW_CPU_CYCLES] = 0x0076,
244 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0,
245 [PERF_COUNT_HW_CACHE_REFERENCES] = 0xff60,
246 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c2,
247 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c3,
248 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x0287,
249 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x0187,
252 static u64 amd_pmu_event_map(int hw_event)
254 if (boot_cpu_data.x86 >= 0x17)
255 return amd_f17h_perfmon_event_map[hw_event];
257 return amd_perfmon_event_map[hw_event];
261 * Previously calculated offsets
263 static unsigned int event_offsets[X86_PMC_IDX_MAX] __read_mostly;
264 static unsigned int count_offsets[X86_PMC_IDX_MAX] __read_mostly;
268 * 4 counters starting at 0xc0010000 each offset by 1
270 * CPUs with core performance counter extensions:
271 * 6 counters starting at 0xc0010200 each offset by 2
273 static inline int amd_pmu_addr_offset(int index, bool eventsel)
281 offset = event_offsets[index];
283 offset = count_offsets[index];
288 if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
294 event_offsets[index] = offset;
296 count_offsets[index] = offset;
301 static int amd_core_hw_config(struct perf_event *event)
303 if (event->attr.exclude_host && event->attr.exclude_guest)
305 * When HO == GO == 1 the hardware treats that as GO == HO == 0
306 * and will count in both modes. We don't want to count in that
307 * case so we emulate no-counting by setting US = OS = 0.
309 event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
310 ARCH_PERFMON_EVENTSEL_OS);
311 else if (event->attr.exclude_host)
312 event->hw.config |= AMD64_EVENTSEL_GUESTONLY;
313 else if (event->attr.exclude_guest)
314 event->hw.config |= AMD64_EVENTSEL_HOSTONLY;
320 * AMD64 events are detected based on their event codes.
322 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
324 return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
327 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
329 return (hwc->config & 0xe0) == 0xe0;
332 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
334 struct amd_nb *nb = cpuc->amd_nb;
336 return nb && nb->nb_id != -1;
339 static int amd_pmu_hw_config(struct perf_event *event)
343 /* pass precise event sampling to ibs: */
344 if (event->attr.precise_ip && get_ibs_caps())
347 if (has_branch_stack(event))
350 ret = x86_pmu_hw_config(event);
354 if (event->attr.type == PERF_TYPE_RAW)
355 event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
357 return amd_core_hw_config(event);
360 static void __amd_put_nb_event_constraints(struct cpu_hw_events *cpuc,
361 struct perf_event *event)
363 struct amd_nb *nb = cpuc->amd_nb;
367 * need to scan whole list because event may not have
368 * been assigned during scheduling
370 * no race condition possible because event can only
371 * be removed on one CPU at a time AND PMU is disabled
374 for (i = 0; i < x86_pmu.num_counters; i++) {
375 if (cmpxchg(nb->owners + i, event, NULL) == event)
381 * AMD64 NorthBridge events need special treatment because
382 * counter access needs to be synchronized across all cores
383 * of a package. Refer to BKDG section 3.12
385 * NB events are events measuring L3 cache, Hypertransport
386 * traffic. They are identified by an event code >= 0xe00.
387 * They measure events on the NorthBride which is shared
388 * by all cores on a package. NB events are counted on a
389 * shared set of counters. When a NB event is programmed
390 * in a counter, the data actually comes from a shared
391 * counter. Thus, access to those counters needs to be
394 * We implement the synchronization such that no two cores
395 * can be measuring NB events using the same counters. Thus,
396 * we maintain a per-NB allocation table. The available slot
397 * is propagated using the event_constraint structure.
399 * We provide only one choice for each NB event based on
400 * the fact that only NB events have restrictions. Consequently,
401 * if a counter is available, there is a guarantee the NB event
402 * will be assigned to it. If no slot is available, an empty
403 * constraint is returned and scheduling will eventually fail
406 * Note that all cores attached the same NB compete for the same
407 * counters to host NB events, this is why we use atomic ops. Some
408 * multi-chip CPUs may have more than one NB.
410 * Given that resources are allocated (cmpxchg), they must be
411 * eventually freed for others to use. This is accomplished by
412 * calling __amd_put_nb_event_constraints()
414 * Non NB events are not impacted by this restriction.
416 static struct event_constraint *
417 __amd_get_nb_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
418 struct event_constraint *c)
420 struct hw_perf_event *hwc = &event->hw;
421 struct amd_nb *nb = cpuc->amd_nb;
422 struct perf_event *old;
432 * detect if already present, if so reuse
434 * cannot merge with actual allocation
435 * because of possible holes
437 * event can already be present yet not assigned (in hwc->idx)
438 * because of successive calls to x86_schedule_events() from
439 * hw_perf_group_sched_in() without hw_perf_enable()
441 for_each_set_bit(idx, c->idxmsk, x86_pmu.num_counters) {
442 if (new == -1 || hwc->idx == idx)
443 /* assign free slot, prefer hwc->idx */
444 old = cmpxchg(nb->owners + idx, NULL, event);
445 else if (nb->owners[idx] == event)
446 /* event already present */
451 if (old && old != event)
454 /* reassign to this slot */
456 cmpxchg(nb->owners + new, event, NULL);
459 /* already present, reuse */
465 return &emptyconstraint;
467 return &nb->event_constraints[new];
470 static struct amd_nb *amd_alloc_nb(int cpu)
475 nb = kzalloc_node(sizeof(struct amd_nb), GFP_KERNEL, cpu_to_node(cpu));
482 * initialize all possible NB constraints
484 for (i = 0; i < x86_pmu.num_counters; i++) {
485 __set_bit(i, nb->event_constraints[i].idxmsk);
486 nb->event_constraints[i].weight = 1;
491 static int amd_pmu_cpu_prepare(int cpu)
493 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
495 WARN_ON_ONCE(cpuc->amd_nb);
497 if (!x86_pmu.amd_nb_constraints)
500 cpuc->amd_nb = amd_alloc_nb(cpu);
507 static void amd_pmu_cpu_starting(int cpu)
509 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
510 void **onln = &cpuc->kfree_on_online[X86_PERF_KFREE_SHARED];
514 cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
516 if (!x86_pmu.amd_nb_constraints)
519 nb_id = amd_get_nb_id(cpu);
520 WARN_ON_ONCE(nb_id == BAD_APICID);
522 for_each_online_cpu(i) {
523 nb = per_cpu(cpu_hw_events, i).amd_nb;
524 if (WARN_ON_ONCE(!nb))
527 if (nb->nb_id == nb_id) {
528 *onln = cpuc->amd_nb;
534 cpuc->amd_nb->nb_id = nb_id;
535 cpuc->amd_nb->refcnt++;
538 static void amd_pmu_cpu_dead(int cpu)
540 struct cpu_hw_events *cpuhw;
542 if (!x86_pmu.amd_nb_constraints)
545 cpuhw = &per_cpu(cpu_hw_events, cpu);
548 struct amd_nb *nb = cpuhw->amd_nb;
550 if (nb->nb_id == -1 || --nb->refcnt == 0)
553 cpuhw->amd_nb = NULL;
558 * When a PMC counter overflows, an NMI is used to process the event and
559 * reset the counter. NMI latency can result in the counter being updated
560 * before the NMI can run, which can result in what appear to be spurious
561 * NMIs. This function is intended to wait for the NMI to run and reset
562 * the counter to avoid possible unhandled NMI messages.
564 #define OVERFLOW_WAIT_COUNT 50
566 static void amd_pmu_wait_on_overflow(int idx)
572 * Wait for the counter to be reset if it has overflowed. This loop
573 * should exit very, very quickly, but just in case, don't wait
576 for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
577 rdmsrl(x86_pmu_event_addr(idx), counter);
578 if (counter & (1ULL << (x86_pmu.cntval_bits - 1)))
581 /* Might be in IRQ context, so can't sleep */
586 static void amd_pmu_disable_all(void)
588 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
591 x86_pmu_disable_all();
594 * This shouldn't be called from NMI context, but add a safeguard here
595 * to return, since if we're in NMI context we can't wait for an NMI
596 * to reset an overflowed counter value.
602 * Check each counter for overflow and wait for it to be reset by the
603 * NMI if it has overflowed. This relies on the fact that all active
604 * counters are always enabled when this function is caled and
605 * ARCH_PERFMON_EVENTSEL_INT is always set.
607 for (idx = 0; idx < x86_pmu.num_counters; idx++) {
608 if (!test_bit(idx, cpuc->active_mask))
611 amd_pmu_wait_on_overflow(idx);
615 static void amd_pmu_disable_event(struct perf_event *event)
617 x86_pmu_disable_event(event);
620 * This can be called from NMI context (via x86_pmu_stop). The counter
621 * may have overflowed, but either way, we'll never see it get reset
622 * by the NMI if we're already in the NMI. And the NMI latency support
623 * below will take care of any pending NMI that might have been
624 * generated by the overflow.
629 amd_pmu_wait_on_overflow(event->hw.idx);
633 * Because of NMI latency, if multiple PMC counters are active or other sources
634 * of NMIs are received, the perf NMI handler can handle one or more overflowed
635 * PMC counters outside of the NMI associated with the PMC overflow. If the NMI
636 * doesn't arrive at the LAPIC in time to become a pending NMI, then the kernel
637 * back-to-back NMI support won't be active. This PMC handler needs to take into
638 * account that this can occur, otherwise this could result in unknown NMI
639 * messages being issued. Examples of this is PMC overflow while in the NMI
640 * handler when multiple PMCs are active or PMC overflow while handling some
641 * other source of an NMI.
643 * Attempt to mitigate this by using the number of active PMCs to determine
644 * whether to return NMI_HANDLED if the perf NMI handler did not handle/reset
645 * any PMCs. The per-CPU perf_nmi_counter variable is set to a minimum of the
646 * number of active PMCs or 2. The value of 2 is used in case an NMI does not
647 * arrive at the LAPIC in time to be collapsed into an already pending NMI.
649 static int amd_pmu_handle_irq(struct pt_regs *regs)
651 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
655 * Obtain the active count before calling x86_pmu_handle_irq() since
656 * it is possible that x86_pmu_handle_irq() may make a counter
657 * inactive (through x86_pmu_stop).
659 active = __bitmap_weight(cpuc->active_mask, X86_PMC_IDX_MAX);
661 /* Process any counter overflows */
662 handled = x86_pmu_handle_irq(regs);
665 * If a counter was handled, record the number of possible remaining
666 * NMIs that can occur.
669 this_cpu_write(perf_nmi_counter,
670 min_t(unsigned int, 2, active));
675 if (!this_cpu_read(perf_nmi_counter))
678 this_cpu_dec(perf_nmi_counter);
683 static struct event_constraint *
684 amd_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
685 struct perf_event *event)
688 * if not NB event or no NB, then no constraints
690 if (!(amd_has_nb(cpuc) && amd_is_nb_event(&event->hw)))
691 return &unconstrained;
693 return __amd_get_nb_event_constraints(cpuc, event, NULL);
696 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
697 struct perf_event *event)
699 if (amd_has_nb(cpuc) && amd_is_nb_event(&event->hw))
700 __amd_put_nb_event_constraints(cpuc, event);
703 PMU_FORMAT_ATTR(event, "config:0-7,32-35");
704 PMU_FORMAT_ATTR(umask, "config:8-15" );
705 PMU_FORMAT_ATTR(edge, "config:18" );
706 PMU_FORMAT_ATTR(inv, "config:23" );
707 PMU_FORMAT_ATTR(cmask, "config:24-31" );
709 static struct attribute *amd_format_attr[] = {
710 &format_attr_event.attr,
711 &format_attr_umask.attr,
712 &format_attr_edge.attr,
713 &format_attr_inv.attr,
714 &format_attr_cmask.attr,
720 #define AMD_EVENT_TYPE_MASK 0x000000F0ULL
722 #define AMD_EVENT_FP 0x00000000ULL ... 0x00000010ULL
723 #define AMD_EVENT_LS 0x00000020ULL ... 0x00000030ULL
724 #define AMD_EVENT_DC 0x00000040ULL ... 0x00000050ULL
725 #define AMD_EVENT_CU 0x00000060ULL ... 0x00000070ULL
726 #define AMD_EVENT_IC_DE 0x00000080ULL ... 0x00000090ULL
727 #define AMD_EVENT_EX_LS 0x000000C0ULL
728 #define AMD_EVENT_DE 0x000000D0ULL
729 #define AMD_EVENT_NB 0x000000E0ULL ... 0x000000F0ULL
732 * AMD family 15h event code/PMC mappings:
734 * type = event_code & 0x0F0:
736 * 0x000 FP PERF_CTL[5:3]
737 * 0x010 FP PERF_CTL[5:3]
738 * 0x020 LS PERF_CTL[5:0]
739 * 0x030 LS PERF_CTL[5:0]
740 * 0x040 DC PERF_CTL[5:0]
741 * 0x050 DC PERF_CTL[5:0]
742 * 0x060 CU PERF_CTL[2:0]
743 * 0x070 CU PERF_CTL[2:0]
744 * 0x080 IC/DE PERF_CTL[2:0]
745 * 0x090 IC/DE PERF_CTL[2:0]
748 * 0x0C0 EX/LS PERF_CTL[5:0]
749 * 0x0D0 DE PERF_CTL[2:0]
750 * 0x0E0 NB NB_PERF_CTL[3:0]
751 * 0x0F0 NB NB_PERF_CTL[3:0]
755 * 0x000 FP PERF_CTL[3], PERF_CTL[5:3] (*)
756 * 0x003 FP PERF_CTL[3]
757 * 0x004 FP PERF_CTL[3], PERF_CTL[5:3] (*)
758 * 0x00B FP PERF_CTL[3]
759 * 0x00D FP PERF_CTL[3]
760 * 0x023 DE PERF_CTL[2:0]
761 * 0x02D LS PERF_CTL[3]
762 * 0x02E LS PERF_CTL[3,0]
763 * 0x031 LS PERF_CTL[2:0] (**)
764 * 0x043 CU PERF_CTL[2:0]
765 * 0x045 CU PERF_CTL[2:0]
766 * 0x046 CU PERF_CTL[2:0]
767 * 0x054 CU PERF_CTL[2:0]
768 * 0x055 CU PERF_CTL[2:0]
769 * 0x08F IC PERF_CTL[0]
770 * 0x187 DE PERF_CTL[0]
771 * 0x188 DE PERF_CTL[0]
772 * 0x0DB EX PERF_CTL[5:0]
773 * 0x0DC LS PERF_CTL[5:0]
774 * 0x0DD LS PERF_CTL[5:0]
775 * 0x0DE LS PERF_CTL[5:0]
776 * 0x0DF LS PERF_CTL[5:0]
777 * 0x1C0 EX PERF_CTL[5:3]
778 * 0x1D6 EX PERF_CTL[5:0]
779 * 0x1D8 EX PERF_CTL[5:0]
781 * (*) depending on the umask all FPU counters may be used
782 * (**) only one unitmask enabled at a time
785 static struct event_constraint amd_f15_PMC0 = EVENT_CONSTRAINT(0, 0x01, 0);
786 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
787 static struct event_constraint amd_f15_PMC3 = EVENT_CONSTRAINT(0, 0x08, 0);
788 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
789 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
790 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
792 static struct event_constraint *
793 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, int idx,
794 struct perf_event *event)
796 struct hw_perf_event *hwc = &event->hw;
797 unsigned int event_code = amd_get_event_code(hwc);
799 switch (event_code & AMD_EVENT_TYPE_MASK) {
801 switch (event_code) {
803 if (!(hwc->config & 0x0000F000ULL))
805 if (!(hwc->config & 0x00000F00ULL))
807 return &amd_f15_PMC3;
809 if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
811 return &amd_f15_PMC3;
815 return &amd_f15_PMC3;
817 return &amd_f15_PMC53;
820 case AMD_EVENT_EX_LS:
821 switch (event_code) {
828 return &amd_f15_PMC20;
830 return &amd_f15_PMC3;
832 return &amd_f15_PMC30;
834 if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
835 return &amd_f15_PMC20;
836 return &emptyconstraint;
838 return &amd_f15_PMC53;
840 return &amd_f15_PMC50;
843 case AMD_EVENT_IC_DE:
845 switch (event_code) {
849 return &amd_f15_PMC0;
850 case 0x0DB ... 0x0DF:
853 return &amd_f15_PMC50;
855 return &amd_f15_PMC20;
858 /* moved to uncore.c */
859 return &emptyconstraint;
861 return &emptyconstraint;
865 static ssize_t amd_event_sysfs_show(char *page, u64 config)
867 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
868 (config & AMD64_EVENTSEL_EVENT) >> 24;
870 return x86_event_sysfs_show(page, config, event);
873 static __initconst const struct x86_pmu amd_pmu = {
875 .handle_irq = amd_pmu_handle_irq,
876 .disable_all = amd_pmu_disable_all,
877 .enable_all = x86_pmu_enable_all,
878 .enable = x86_pmu_enable_event,
879 .disable = amd_pmu_disable_event,
880 .hw_config = amd_pmu_hw_config,
881 .schedule_events = x86_schedule_events,
882 .eventsel = MSR_K7_EVNTSEL0,
883 .perfctr = MSR_K7_PERFCTR0,
884 .addr_offset = amd_pmu_addr_offset,
885 .event_map = amd_pmu_event_map,
886 .max_events = ARRAY_SIZE(amd_perfmon_event_map),
887 .num_counters = AMD64_NUM_COUNTERS,
889 .cntval_mask = (1ULL << 48) - 1,
891 /* use highest bit to detect overflow */
892 .max_period = (1ULL << 47) - 1,
893 .get_event_constraints = amd_get_event_constraints,
894 .put_event_constraints = amd_put_event_constraints,
896 .format_attrs = amd_format_attr,
897 .events_sysfs_show = amd_event_sysfs_show,
899 .cpu_prepare = amd_pmu_cpu_prepare,
900 .cpu_starting = amd_pmu_cpu_starting,
901 .cpu_dead = amd_pmu_cpu_dead,
903 .amd_nb_constraints = 1,
906 static int __init amd_core_pmu_init(void)
908 if (!boot_cpu_has(X86_FEATURE_PERFCTR_CORE))
911 switch (boot_cpu_data.x86) {
914 x86_pmu.get_event_constraints = amd_get_event_constraints_f15h;
919 * In family 17h, there are no event constraints in the PMC hardware.
920 * We fallback to using default amd_get_event_constraints.
925 /* Using default amd_get_event_constraints. */
928 pr_err("core perfctr but no constraints; unknown hardware!\n");
933 * If core performance counter extensions exists, we must use
934 * MSR_F15H_PERF_CTL/MSR_F15H_PERF_CTR msrs. See also
935 * amd_pmu_addr_offset().
937 x86_pmu.eventsel = MSR_F15H_PERF_CTL;
938 x86_pmu.perfctr = MSR_F15H_PERF_CTR;
939 x86_pmu.num_counters = AMD64_NUM_COUNTERS_CORE;
941 * AMD Core perfctr has separate MSRs for the NB events, see
942 * the amd/uncore.c driver.
944 x86_pmu.amd_nb_constraints = 0;
946 pr_cont("core perfctr, ");
950 __init int amd_pmu_init(void)
954 /* Performance-monitoring supported from K7 and later: */
955 if (boot_cpu_data.x86 < 6)
960 ret = amd_core_pmu_init();
964 if (num_possible_cpus() == 1) {
966 * No point in allocating data structures to serialize
967 * against other CPUs, when there is only the one CPU.
969 x86_pmu.amd_nb_constraints = 0;
972 if (boot_cpu_data.x86 >= 0x17)
973 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids_f17h, sizeof(hw_cache_event_ids));
975 memcpy(hw_cache_event_ids, amd_hw_cache_event_ids, sizeof(hw_cache_event_ids));
980 void amd_pmu_enable_virt(void)
982 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
984 cpuc->perf_ctr_virt_mask = 0;
986 /* Reload all events */
987 amd_pmu_disable_all();
988 x86_pmu_enable_all(0);
990 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
992 void amd_pmu_disable_virt(void)
994 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
997 * We only mask out the Host-only bit so that host-only counting works
998 * when SVM is disabled. If someone sets up a guest-only counter when
999 * SVM is disabled the Guest-only bits still gets set and the counter
1000 * will not count anything.
1002 cpuc->perf_ctr_virt_mask = AMD64_EVENTSEL_HOSTONLY;
1004 /* Reload all events */
1005 amd_pmu_disable_all();
1006 x86_pmu_enable_all(0);
1008 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);