2 * Performance events x86 architecture header
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2009 Jaswinder Singh Rajput
7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra
9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
17 #include <asm/intel_ds.h>
20 /* To enable MSR tracing please use the generic trace points. */
24 * register -------------------------------
25 * | HT | no HT | HT | no HT |
26 *-----------------------------------------
27 * offcore | core | core | cpu | core |
28 * lbr_sel | core | core | cpu | core |
29 * ld_lat | cpu | core | cpu | core |
30 *-----------------------------------------
32 * Given that there is a small number of shared regs,
33 * we can pre-allocate their slot in the per-cpu
34 * per-core reg tables.
37 EXTRA_REG_NONE = -1, /* not used */
39 EXTRA_REG_RSP_0 = 0, /* offcore_response_0 */
40 EXTRA_REG_RSP_1 = 1, /* offcore_response_1 */
41 EXTRA_REG_LBR = 2, /* lbr_select */
42 EXTRA_REG_LDLAT = 3, /* ld_lat_threshold */
43 EXTRA_REG_FE = 4, /* fe_* */
45 EXTRA_REG_MAX /* number of entries needed */
48 struct event_constraint {
50 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
61 static inline bool constraint_match(struct event_constraint *c, u64 ecode)
63 return ((ecode & c->cmask) - c->code) <= (u64)c->size;
67 * struct hw_perf_event.flags flags
69 #define PERF_X86_EVENT_PEBS_LDLAT 0x0001 /* ld+ldlat data address sampling */
70 #define PERF_X86_EVENT_PEBS_ST 0x0002 /* st data address sampling */
71 #define PERF_X86_EVENT_PEBS_ST_HSW 0x0004 /* haswell style datala, store */
72 #define PERF_X86_EVENT_PEBS_LD_HSW 0x0008 /* haswell style datala, load */
73 #define PERF_X86_EVENT_PEBS_NA_HSW 0x0010 /* haswell style datala, unknown */
74 #define PERF_X86_EVENT_EXCL 0x0020 /* HT exclusivity on counter */
75 #define PERF_X86_EVENT_DYNAMIC 0x0040 /* dynamic alloc'd constraint */
76 #define PERF_X86_EVENT_RDPMC_ALLOWED 0x0080 /* grant rdpmc permission */
77 #define PERF_X86_EVENT_EXCL_ACCT 0x0100 /* accounted EXCL event */
78 #define PERF_X86_EVENT_AUTO_RELOAD 0x0200 /* use PEBS auto-reload */
79 #define PERF_X86_EVENT_LARGE_PEBS 0x0400 /* use large PEBS */
80 #define PERF_X86_EVENT_PEBS_VIA_PT 0x0800 /* use PT buffer for PEBS */
81 #define PERF_X86_EVENT_PAIR 0x1000 /* Large Increment per Cycle */
82 #define PERF_X86_EVENT_LBR_SELECT 0x2000 /* Save/Restore MSR_LBR_SELECT */
83 #define PERF_X86_EVENT_TOPDOWN 0x4000 /* Count Topdown slots/metrics events */
84 #define PERF_X86_EVENT_PEBS_STLAT 0x8000 /* st+stlat data address sampling */
86 static inline bool is_topdown_count(struct perf_event *event)
88 return event->hw.flags & PERF_X86_EVENT_TOPDOWN;
91 static inline bool is_metric_event(struct perf_event *event)
93 u64 config = event->attr.config;
95 return ((config & ARCH_PERFMON_EVENTSEL_EVENT) == 0) &&
96 ((config & INTEL_ARCH_EVENT_MASK) >= INTEL_TD_METRIC_RETIRING) &&
97 ((config & INTEL_ARCH_EVENT_MASK) <= INTEL_TD_METRIC_MAX);
100 static inline bool is_slots_event(struct perf_event *event)
102 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_TD_SLOTS;
105 static inline bool is_topdown_event(struct perf_event *event)
107 return is_metric_event(event) || is_slots_event(event);
111 int nb_id; /* NorthBridge id */
112 int refcnt; /* reference count */
113 struct perf_event *owners[X86_PMC_IDX_MAX];
114 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
117 #define PEBS_COUNTER_MASK ((1ULL << MAX_PEBS_EVENTS) - 1)
118 #define PEBS_PMI_AFTER_EACH_RECORD BIT_ULL(60)
119 #define PEBS_OUTPUT_OFFSET 61
120 #define PEBS_OUTPUT_MASK (3ull << PEBS_OUTPUT_OFFSET)
121 #define PEBS_OUTPUT_PT (1ull << PEBS_OUTPUT_OFFSET)
122 #define PEBS_VIA_PT_MASK (PEBS_OUTPUT_PT | PEBS_PMI_AFTER_EACH_RECORD)
125 * Flags PEBS can handle without an PMI.
127 * TID can only be handled by flushing at context switch.
128 * REGS_USER can be handled for events limited to ring 3.
131 #define LARGE_PEBS_FLAGS \
132 (PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR | \
133 PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID | \
134 PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER | \
135 PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR | \
136 PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER | \
137 PERF_SAMPLE_PERIOD | PERF_SAMPLE_CODE_PAGE_SIZE)
139 #define PEBS_GP_REGS \
140 ((1ULL << PERF_REG_X86_AX) | \
141 (1ULL << PERF_REG_X86_BX) | \
142 (1ULL << PERF_REG_X86_CX) | \
143 (1ULL << PERF_REG_X86_DX) | \
144 (1ULL << PERF_REG_X86_DI) | \
145 (1ULL << PERF_REG_X86_SI) | \
146 (1ULL << PERF_REG_X86_SP) | \
147 (1ULL << PERF_REG_X86_BP) | \
148 (1ULL << PERF_REG_X86_IP) | \
149 (1ULL << PERF_REG_X86_FLAGS) | \
150 (1ULL << PERF_REG_X86_R8) | \
151 (1ULL << PERF_REG_X86_R9) | \
152 (1ULL << PERF_REG_X86_R10) | \
153 (1ULL << PERF_REG_X86_R11) | \
154 (1ULL << PERF_REG_X86_R12) | \
155 (1ULL << PERF_REG_X86_R13) | \
156 (1ULL << PERF_REG_X86_R14) | \
157 (1ULL << PERF_REG_X86_R15))
160 * Per register state.
163 raw_spinlock_t lock; /* per-core: protect structure */
164 u64 config; /* extra MSR config */
165 u64 reg; /* extra MSR number */
166 atomic_t ref; /* reference count */
172 * Used to coordinate shared registers between HT threads or
173 * among events on a single PMU.
175 struct intel_shared_regs {
176 struct er_account regs[EXTRA_REG_MAX];
177 int refcnt; /* per-core: #HT threads */
178 unsigned core_id; /* per-core: core id */
181 enum intel_excl_state_type {
182 INTEL_EXCL_UNUSED = 0, /* counter is unused */
183 INTEL_EXCL_SHARED = 1, /* counter can be used by both threads */
184 INTEL_EXCL_EXCLUSIVE = 2, /* counter can be used by one thread only */
187 struct intel_excl_states {
188 enum intel_excl_state_type state[X86_PMC_IDX_MAX];
189 bool sched_started; /* true if scheduling has started */
192 struct intel_excl_cntrs {
195 struct intel_excl_states states[2];
198 u16 has_exclusive[2];
199 u32 exclusive_present;
202 int refcnt; /* per-core: #HT threads */
203 unsigned core_id; /* per-core: core id */
206 struct x86_perf_task_context;
207 #define MAX_LBR_ENTRIES 32
210 LBR_FORMAT_32 = 0x00,
211 LBR_FORMAT_LIP = 0x01,
212 LBR_FORMAT_EIP = 0x02,
213 LBR_FORMAT_EIP_FLAGS = 0x03,
214 LBR_FORMAT_EIP_FLAGS2 = 0x04,
215 LBR_FORMAT_INFO = 0x05,
216 LBR_FORMAT_TIME = 0x06,
217 LBR_FORMAT_MAX_KNOWN = LBR_FORMAT_TIME,
221 X86_PERF_KFREE_SHARED = 0,
222 X86_PERF_KFREE_EXCL = 1,
226 struct cpu_hw_events {
228 * Generic x86 PMC bits
230 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
231 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
232 unsigned long dirty[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
235 int n_events; /* the # of events in the below arrays */
236 int n_added; /* the # last events in the below arrays;
237 they've never been enabled yet */
238 int n_txn; /* the # last events in the below arrays;
239 added in the current transaction */
242 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
243 u64 tags[X86_PMC_IDX_MAX];
245 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
246 struct event_constraint *event_constraint[X86_PMC_IDX_MAX];
248 int n_excl; /* the number of exclusive events */
250 unsigned int txn_flags;
254 * Intel DebugStore bits
256 struct debug_store *ds;
265 /* Current super set of events hardware configuration */
267 u64 active_pebs_data_cfg;
268 int pebs_record_size;
275 struct perf_branch_stack lbr_stack;
276 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
278 struct er_account *lbr_sel;
279 struct er_account *lbr_ctl;
288 * Intel host/guest exclude bits
290 u64 intel_ctrl_guest_mask;
291 u64 intel_ctrl_host_mask;
292 struct perf_guest_switch_msr guest_switch_msrs[X86_PMC_IDX_MAX];
295 * Intel checkpoint mask
300 * manage shared (per-core, per-cpu) registers
301 * used on Intel NHM/WSM/SNB
303 struct intel_shared_regs *shared_regs;
305 * manage exclusive counter access between hyperthread
307 struct event_constraint *constraint_list; /* in enable order */
308 struct intel_excl_cntrs *excl_cntrs;
309 int excl_thread_id; /* 0 or 1 */
312 * SKL TSX_FORCE_ABORT shadow
319 /* number of accepted metrics events */
325 struct amd_nb *amd_nb;
326 /* Inverted mask of bits to clear in the perf_ctr ctrl registers */
327 u64 perf_ctr_virt_mask;
328 int n_pair; /* Large increment events */
330 void *kfree_on_online[X86_PERF_KFREE_MAX];
335 #define __EVENT_CONSTRAINT_RANGE(c, e, n, m, w, o, f) { \
336 { .idxmsk64 = (n) }, \
345 #define __EVENT_CONSTRAINT(c, n, m, w, o, f) \
346 __EVENT_CONSTRAINT_RANGE(c, c, n, m, w, o, f)
348 #define EVENT_CONSTRAINT(c, n, m) \
349 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 0, 0)
352 * The constraint_match() function only works for 'simple' event codes
353 * and not for extended (AMD64_EVENTSEL_EVENT) events codes.
355 #define EVENT_CONSTRAINT_RANGE(c, e, n, m) \
356 __EVENT_CONSTRAINT_RANGE(c, e, n, m, HWEIGHT(n), 0, 0)
358 #define INTEL_EXCLEVT_CONSTRAINT(c, n) \
359 __EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT, HWEIGHT(n),\
360 0, PERF_X86_EVENT_EXCL)
363 * The overlap flag marks event constraints with overlapping counter
364 * masks. This is the case if the counter mask of such an event is not
365 * a subset of any other counter mask of a constraint with an equal or
366 * higher weight, e.g.:
368 * c_overlaps = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
369 * c_another1 = EVENT_CONSTRAINT(0, 0x07, 0);
370 * c_another2 = EVENT_CONSTRAINT(0, 0x38, 0);
372 * The event scheduler may not select the correct counter in the first
373 * cycle because it needs to know which subsequent events will be
374 * scheduled. It may fail to schedule the events then. So we set the
375 * overlap flag for such constraints to give the scheduler a hint which
376 * events to select for counter rescheduling.
378 * Care must be taken as the rescheduling algorithm is O(n!) which
379 * will increase scheduling cycles for an over-committed system
380 * dramatically. The number of such EVENT_CONSTRAINT_OVERLAP() macros
381 * and its counter masks must be kept at a minimum.
383 #define EVENT_CONSTRAINT_OVERLAP(c, n, m) \
384 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 1, 0)
387 * Constraint on the Event code.
389 #define INTEL_EVENT_CONSTRAINT(c, n) \
390 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
393 * Constraint on a range of Event codes
395 #define INTEL_EVENT_CONSTRAINT_RANGE(c, e, n) \
396 EVENT_CONSTRAINT_RANGE(c, e, n, ARCH_PERFMON_EVENTSEL_EVENT)
399 * Constraint on the Event code + UMask + fixed-mask
401 * filter mask to validate fixed counter events.
402 * the following filters disqualify for fixed counters:
407 * - in_tx_checkpointed
408 * The other filters are supported by fixed counters.
409 * The any-thread option is supported starting with v3.
411 #define FIXED_EVENT_FLAGS (X86_RAW_EVENT_MASK|HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)
412 #define FIXED_EVENT_CONSTRAINT(c, n) \
413 EVENT_CONSTRAINT(c, (1ULL << (32+n)), FIXED_EVENT_FLAGS)
416 * The special metric counters do not actually exist. They are calculated from
417 * the combination of the FxCtr3 + MSR_PERF_METRICS.
419 * The special metric counters are mapped to a dummy offset for the scheduler.
420 * The sharing between multiple users of the same metric without multiplexing
421 * is not allowed, even though the hardware supports that in principle.
424 #define METRIC_EVENT_CONSTRAINT(c, n) \
425 EVENT_CONSTRAINT(c, (1ULL << (INTEL_PMC_IDX_METRIC_BASE + n)), \
426 INTEL_ARCH_EVENT_MASK)
429 * Constraint on the Event code + UMask
431 #define INTEL_UEVENT_CONSTRAINT(c, n) \
432 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
434 /* Constraint on specific umask bit only + event */
435 #define INTEL_UBIT_EVENT_CONSTRAINT(c, n) \
436 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT|(c))
438 /* Like UEVENT_CONSTRAINT, but match flags too */
439 #define INTEL_FLAGS_UEVENT_CONSTRAINT(c, n) \
440 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS)
442 #define INTEL_EXCLUEVT_CONSTRAINT(c, n) \
443 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
444 HWEIGHT(n), 0, PERF_X86_EVENT_EXCL)
446 #define INTEL_PLD_CONSTRAINT(c, n) \
447 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
448 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LDLAT)
450 #define INTEL_PSD_CONSTRAINT(c, n) \
451 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
452 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_STLAT)
454 #define INTEL_PST_CONSTRAINT(c, n) \
455 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
456 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST)
458 /* Event constraint, but match on all event flags too. */
459 #define INTEL_FLAGS_EVENT_CONSTRAINT(c, n) \
460 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS)
462 #define INTEL_FLAGS_EVENT_CONSTRAINT_RANGE(c, e, n) \
463 EVENT_CONSTRAINT_RANGE(c, e, n, ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS)
465 /* Check only flags, but allow all event/umask */
466 #define INTEL_ALL_EVENT_CONSTRAINT(code, n) \
467 EVENT_CONSTRAINT(code, n, X86_ALL_EVENT_FLAGS)
469 /* Check flags and event code, and set the HSW store flag */
470 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_ST(code, n) \
471 __EVENT_CONSTRAINT(code, n, \
472 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \
473 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW)
475 /* Check flags and event code, and set the HSW load flag */
476 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(code, n) \
477 __EVENT_CONSTRAINT(code, n, \
478 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \
479 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LD_HSW)
481 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(code, end, n) \
482 __EVENT_CONSTRAINT_RANGE(code, end, n, \
483 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \
484 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LD_HSW)
486 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(code, n) \
487 __EVENT_CONSTRAINT(code, n, \
488 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \
490 PERF_X86_EVENT_PEBS_LD_HSW|PERF_X86_EVENT_EXCL)
492 /* Check flags and event code/umask, and set the HSW store flag */
493 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(code, n) \
494 __EVENT_CONSTRAINT(code, n, \
495 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
496 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW)
498 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(code, n) \
499 __EVENT_CONSTRAINT(code, n, \
500 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
502 PERF_X86_EVENT_PEBS_ST_HSW|PERF_X86_EVENT_EXCL)
504 /* Check flags and event code/umask, and set the HSW load flag */
505 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(code, n) \
506 __EVENT_CONSTRAINT(code, n, \
507 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
508 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LD_HSW)
510 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(code, n) \
511 __EVENT_CONSTRAINT(code, n, \
512 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
514 PERF_X86_EVENT_PEBS_LD_HSW|PERF_X86_EVENT_EXCL)
516 /* Check flags and event code/umask, and set the HSW N/A flag */
517 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(code, n) \
518 __EVENT_CONSTRAINT(code, n, \
519 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \
520 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_NA_HSW)
524 * We define the end marker as having a weight of -1
525 * to enable blacklisting of events using a counter bitmask
526 * of zero and thus a weight of zero.
527 * The end marker has a weight that cannot possibly be
528 * obtained from counting the bits in the bitmask.
530 #define EVENT_CONSTRAINT_END { .weight = -1 }
533 * Check for end marker with weight == -1
535 #define for_each_event_constraint(e, c) \
536 for ((e) = (c); (e)->weight != -1; (e)++)
539 * Extra registers for specific events.
541 * Some events need large masks and require external MSRs.
542 * Those extra MSRs end up being shared for all events on
543 * a PMU and sometimes between PMU of sibling HT threads.
544 * In either case, the kernel needs to handle conflicting
545 * accesses to those extra, shared, regs. The data structure
546 * to manage those registers is stored in cpu_hw_event.
553 int idx; /* per_xxx->regs[] reg index */
554 bool extra_msr_access;
557 #define EVENT_EXTRA_REG(e, ms, m, vm, i) { \
560 .config_mask = (m), \
561 .valid_mask = (vm), \
562 .idx = EXTRA_REG_##i, \
563 .extra_msr_access = true, \
566 #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx) \
567 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx)
569 #define INTEL_UEVENT_EXTRA_REG(event, msr, vm, idx) \
570 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
571 ARCH_PERFMON_EVENTSEL_UMASK, vm, idx)
573 #define INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(c) \
574 INTEL_UEVENT_EXTRA_REG(c, \
575 MSR_PEBS_LD_LAT_THRESHOLD, \
579 #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0)
581 union perf_capabilities {
589 * PMU supports separate counter range for writing
592 u64 full_width_write:1;
595 u64 pebs_output_pt_available:1;
596 u64 anythread_deprecated:1;
601 struct x86_pmu_quirk {
602 struct x86_pmu_quirk *next;
606 union x86_pmu_config {
627 #define X86_CONFIG(args...) ((union x86_pmu_config){.bits = {args}}).value
630 x86_lbr_exclusive_lbr,
631 x86_lbr_exclusive_bts,
632 x86_lbr_exclusive_pt,
633 x86_lbr_exclusive_max,
636 struct x86_hybrid_pmu {
640 cpumask_t supported_cpus;
641 union perf_capabilities intel_cap;
645 int num_counters_fixed;
646 struct event_constraint unconstrained;
648 u64 hw_cache_event_ids
649 [PERF_COUNT_HW_CACHE_MAX]
650 [PERF_COUNT_HW_CACHE_OP_MAX]
651 [PERF_COUNT_HW_CACHE_RESULT_MAX];
652 u64 hw_cache_extra_regs
653 [PERF_COUNT_HW_CACHE_MAX]
654 [PERF_COUNT_HW_CACHE_OP_MAX]
655 [PERF_COUNT_HW_CACHE_RESULT_MAX];
656 struct event_constraint *event_constraints;
657 struct event_constraint *pebs_constraints;
658 struct extra_reg *extra_regs;
660 unsigned int late_ack :1,
665 static __always_inline struct x86_hybrid_pmu *hybrid_pmu(struct pmu *pmu)
667 return container_of(pmu, struct x86_hybrid_pmu, pmu);
670 extern struct static_key_false perf_is_hybrid;
671 #define is_hybrid() static_branch_unlikely(&perf_is_hybrid)
673 #define hybrid(_pmu, _field) \
675 typeof(&x86_pmu._field) __Fp = &x86_pmu._field; \
677 if (is_hybrid() && (_pmu)) \
678 __Fp = &hybrid_pmu(_pmu)->_field; \
683 #define hybrid_var(_pmu, _var) \
685 typeof(&_var) __Fp = &_var; \
687 if (is_hybrid() && (_pmu)) \
688 __Fp = &hybrid_pmu(_pmu)->_var; \
693 #define hybrid_bit(_pmu, _field) \
695 bool __Fp = x86_pmu._field; \
697 if (is_hybrid() && (_pmu)) \
698 __Fp = hybrid_pmu(_pmu)->_field; \
703 enum hybrid_pmu_type {
707 hybrid_big_small = hybrid_big | hybrid_small,
710 #define X86_HYBRID_PMU_ATOM_IDX 0
711 #define X86_HYBRID_PMU_CORE_IDX 1
713 #define X86_HYBRID_NUM_PMUS 2
716 * struct x86_pmu - generic x86 pmu
720 * Generic x86 PMC bits
724 int (*handle_irq)(struct pt_regs *);
725 void (*disable_all)(void);
726 void (*enable_all)(int added);
727 void (*enable)(struct perf_event *);
728 void (*disable)(struct perf_event *);
729 void (*add)(struct perf_event *);
730 void (*del)(struct perf_event *);
731 void (*read)(struct perf_event *event);
732 int (*hw_config)(struct perf_event *event);
733 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
736 int (*addr_offset)(int index, bool eventsel);
737 int (*rdpmc_index)(int index);
738 u64 (*event_map)(int);
741 int num_counters_fixed;
745 unsigned long events_maskl;
746 unsigned long events_mask[BITS_TO_LONGS(ARCH_PERFMON_EVENTS_COUNT)];
751 struct event_constraint *
752 (*get_event_constraints)(struct cpu_hw_events *cpuc,
754 struct perf_event *event);
756 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
757 struct perf_event *event);
759 void (*start_scheduling)(struct cpu_hw_events *cpuc);
761 void (*commit_scheduling)(struct cpu_hw_events *cpuc, int idx, int cntr);
763 void (*stop_scheduling)(struct cpu_hw_events *cpuc);
765 struct event_constraint *event_constraints;
766 struct x86_pmu_quirk *quirks;
767 int perfctr_second_write;
768 u64 (*limit_period)(struct perf_event *event, u64 l);
770 /* PMI handler bits */
771 unsigned int late_ack :1,
777 int attr_rdpmc_broken;
779 struct attribute **format_attrs;
781 ssize_t (*events_sysfs_show)(char *page, u64 config);
782 const struct attribute_group **attr_update;
784 unsigned long attr_freeze_on_smi;
789 int (*cpu_prepare)(int cpu);
790 void (*cpu_starting)(int cpu);
791 void (*cpu_dying)(int cpu);
792 void (*cpu_dead)(int cpu);
794 void (*check_microcode)(void);
795 void (*sched_task)(struct perf_event_context *ctx,
799 * Intel Arch Perfmon v2+
802 union perf_capabilities intel_cap;
805 * Intel DebugStore bits
814 pebs_no_isolation :1,
816 int pebs_record_size;
817 int pebs_buffer_size;
819 void (*drain_pebs)(struct pt_regs *regs, struct perf_sample_data *data);
820 struct event_constraint *pebs_constraints;
821 void (*pebs_aliases)(struct perf_event *event);
822 unsigned long large_pebs_flags;
828 unsigned int lbr_tos, lbr_from, lbr_to,
829 lbr_info, lbr_nr; /* LBR base regs and size */
831 u64 lbr_sel_mask; /* LBR_SELECT valid bits */
832 u64 lbr_ctl_mask; /* LBR_CTL valid bits */
835 const int *lbr_sel_map; /* lbr_select mappings */
836 int *lbr_ctl_map; /* LBR_CTL mappings */
838 bool lbr_double_abort; /* duplicated lbr aborts */
839 bool lbr_pt_coexist; /* (LBR|BTS) may coexist with PT */
842 * Intel Architectural LBR CPUID Enumeration
844 unsigned int lbr_depth_mask:8;
845 unsigned int lbr_deep_c_reset:1;
846 unsigned int lbr_lip:1;
847 unsigned int lbr_cpl:1;
848 unsigned int lbr_filter:1;
849 unsigned int lbr_call_stack:1;
850 unsigned int lbr_mispred:1;
851 unsigned int lbr_timed_lbr:1;
852 unsigned int lbr_br_type:1;
854 void (*lbr_reset)(void);
855 void (*lbr_read)(struct cpu_hw_events *cpuc);
856 void (*lbr_save)(void *ctx);
857 void (*lbr_restore)(void *ctx);
860 * Intel PT/LBR/BTS are exclusive
862 atomic_t lbr_exclusive[x86_lbr_exclusive_max];
867 int num_topdown_events;
868 u64 (*update_topdown_event)(struct perf_event *event);
869 int (*set_topdown_event_period)(struct perf_event *event);
872 * perf task context (i.e. struct perf_event_context::task_ctx_data)
873 * switch helper to bridge calls from perf/core to perf/x86.
874 * See struct pmu::swap_task_ctx() usage for examples;
876 void (*swap_task_ctx)(struct perf_event_context *prev,
877 struct perf_event_context *next);
882 unsigned int amd_nb_constraints : 1;
883 u64 perf_ctr_pair_en;
886 * Extra registers for events
888 struct extra_reg *extra_regs;
892 * Intel host/guest support (KVM)
894 struct perf_guest_switch_msr *(*guest_get_msrs)(int *nr);
897 * Check period value for PERF_EVENT_IOC_PERIOD ioctl.
899 int (*check_period) (struct perf_event *event, u64 period);
901 int (*aux_output_match) (struct perf_event *event);
903 int (*filter_match)(struct perf_event *event);
907 * Most PMU capabilities are the same among different hybrid PMUs.
908 * The global x86_pmu saves the architecture capabilities, which
909 * are available for all PMUs. The hybrid_pmu only includes the
910 * unique capabilities.
913 struct x86_hybrid_pmu *hybrid_pmu;
914 u8 (*get_hybrid_cpu_type) (void);
917 struct x86_perf_task_context_opt {
918 int lbr_callstack_users;
923 struct x86_perf_task_context {
927 struct x86_perf_task_context_opt opt;
928 struct lbr_entry lbr[MAX_LBR_ENTRIES];
931 struct x86_perf_task_context_arch_lbr {
932 struct x86_perf_task_context_opt opt;
933 struct lbr_entry entries[];
937 * Add padding to guarantee the 64-byte alignment of the state buffer.
939 * The structure is dynamically allocated. The size of the LBR state may vary
940 * based on the number of LBR registers.
942 * Do not put anything after the LBR state.
944 struct x86_perf_task_context_arch_lbr_xsave {
945 struct x86_perf_task_context_opt opt;
948 struct xregs_state xsave;
950 struct fxregs_state i387;
951 struct xstate_header header;
952 struct arch_lbr_state lbr;
953 } __attribute__ ((packed, aligned (XSAVE_ALIGNMENT)));
957 #define x86_add_quirk(func_) \
959 static struct x86_pmu_quirk __quirk __initdata = { \
962 __quirk.next = x86_pmu.quirks; \
963 x86_pmu.quirks = &__quirk; \
969 #define PMU_FL_NO_HT_SHARING 0x1 /* no hyper-threading resource sharing */
970 #define PMU_FL_HAS_RSP_1 0x2 /* has 2 equivalent offcore_rsp regs */
971 #define PMU_FL_EXCL_CNTRS 0x4 /* has exclusive counter requirements */
972 #define PMU_FL_EXCL_ENABLED 0x8 /* exclusive counter active */
973 #define PMU_FL_PEBS_ALL 0x10 /* all events are valid PEBS events */
974 #define PMU_FL_TFA 0x20 /* deal with TSX force abort */
975 #define PMU_FL_PAIR 0x40 /* merge counters for large incr. events */
976 #define PMU_FL_INSTR_LATENCY 0x80 /* Support Instruction Latency in PEBS Memory Info Record */
977 #define PMU_FL_MEM_LOADS_AUX 0x100 /* Require an auxiliary event for the complete memory info */
979 #define EVENT_VAR(_id) event_attr_##_id
980 #define EVENT_PTR(_id) &event_attr_##_id.attr.attr
982 #define EVENT_ATTR(_name, _id) \
983 static struct perf_pmu_events_attr EVENT_VAR(_id) = { \
984 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
985 .id = PERF_COUNT_HW_##_id, \
989 #define EVENT_ATTR_STR(_name, v, str) \
990 static struct perf_pmu_events_attr event_attr_##v = { \
991 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
996 #define EVENT_ATTR_STR_HT(_name, v, noht, ht) \
997 static struct perf_pmu_events_ht_attr event_attr_##v = { \
998 .attr = __ATTR(_name, 0444, events_ht_sysfs_show, NULL),\
1000 .event_str_noht = noht, \
1001 .event_str_ht = ht, \
1004 #define EVENT_ATTR_STR_HYBRID(_name, v, str, _pmu) \
1005 static struct perf_pmu_events_hybrid_attr event_attr_##v = { \
1006 .attr = __ATTR(_name, 0444, events_hybrid_sysfs_show, NULL),\
1012 #define FORMAT_HYBRID_PTR(_id) (&format_attr_hybrid_##_id.attr.attr)
1014 #define FORMAT_ATTR_HYBRID(_name, _pmu) \
1015 static struct perf_pmu_format_hybrid_attr format_attr_hybrid_##_name = {\
1016 .attr = __ATTR_RO(_name), \
1020 struct pmu *x86_get_pmu(unsigned int cpu);
1021 extern struct x86_pmu x86_pmu __read_mostly;
1023 static __always_inline struct x86_perf_task_context_opt *task_context_opt(void *ctx)
1025 if (static_cpu_has(X86_FEATURE_ARCH_LBR))
1026 return &((struct x86_perf_task_context_arch_lbr *)ctx)->opt;
1028 return &((struct x86_perf_task_context *)ctx)->opt;
1031 static inline bool x86_pmu_has_lbr_callstack(void)
1033 return x86_pmu.lbr_sel_map &&
1034 x86_pmu.lbr_sel_map[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT] > 0;
1037 DECLARE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
1039 int x86_perf_event_set_period(struct perf_event *event);
1042 * Generalized hw caching related hw_event table, filled
1043 * in on a per model basis. A value of 0 means
1044 * 'not supported', -1 means 'hw_event makes no sense on
1045 * this CPU', any other value means the raw hw_event
1049 #define C(x) PERF_COUNT_HW_CACHE_##x
1051 extern u64 __read_mostly hw_cache_event_ids
1052 [PERF_COUNT_HW_CACHE_MAX]
1053 [PERF_COUNT_HW_CACHE_OP_MAX]
1054 [PERF_COUNT_HW_CACHE_RESULT_MAX];
1055 extern u64 __read_mostly hw_cache_extra_regs
1056 [PERF_COUNT_HW_CACHE_MAX]
1057 [PERF_COUNT_HW_CACHE_OP_MAX]
1058 [PERF_COUNT_HW_CACHE_RESULT_MAX];
1060 u64 x86_perf_event_update(struct perf_event *event);
1062 static inline unsigned int x86_pmu_config_addr(int index)
1064 return x86_pmu.eventsel + (x86_pmu.addr_offset ?
1065 x86_pmu.addr_offset(index, true) : index);
1068 static inline unsigned int x86_pmu_event_addr(int index)
1070 return x86_pmu.perfctr + (x86_pmu.addr_offset ?
1071 x86_pmu.addr_offset(index, false) : index);
1074 static inline int x86_pmu_rdpmc_index(int index)
1076 return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index(index) : index;
1079 bool check_hw_exists(struct pmu *pmu, int num_counters,
1080 int num_counters_fixed);
1082 int x86_add_exclusive(unsigned int what);
1084 void x86_del_exclusive(unsigned int what);
1086 int x86_reserve_hardware(void);
1088 void x86_release_hardware(void);
1090 int x86_pmu_max_precise(void);
1092 void hw_perf_lbr_event_destroy(struct perf_event *event);
1094 int x86_setup_perfctr(struct perf_event *event);
1096 int x86_pmu_hw_config(struct perf_event *event);
1098 void x86_pmu_disable_all(void);
1100 static inline bool is_counter_pair(struct hw_perf_event *hwc)
1102 return hwc->flags & PERF_X86_EVENT_PAIR;
1105 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
1108 u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
1110 if (hwc->extra_reg.reg)
1111 wrmsrl(hwc->extra_reg.reg, hwc->extra_reg.config);
1114 * Add enabled Merge event on next counter
1115 * if large increment event being enabled on this counter
1117 if (is_counter_pair(hwc))
1118 wrmsrl(x86_pmu_config_addr(hwc->idx + 1), x86_pmu.perf_ctr_pair_en);
1120 wrmsrl(hwc->config_base, (hwc->config | enable_mask) & ~disable_mask);
1123 void x86_pmu_enable_all(int added);
1125 int perf_assign_events(struct event_constraint **constraints, int n,
1126 int wmin, int wmax, int gpmax, int *assign);
1127 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign);
1129 void x86_pmu_stop(struct perf_event *event, int flags);
1131 static inline void x86_pmu_disable_event(struct perf_event *event)
1133 u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
1134 struct hw_perf_event *hwc = &event->hw;
1136 wrmsrl(hwc->config_base, hwc->config & ~disable_mask);
1138 if (is_counter_pair(hwc))
1139 wrmsrl(x86_pmu_config_addr(hwc->idx + 1), 0);
1142 void x86_pmu_enable_event(struct perf_event *event);
1144 int x86_pmu_handle_irq(struct pt_regs *regs);
1146 void x86_pmu_show_pmu_cap(int num_counters, int num_counters_fixed,
1149 void x86_pmu_update_cpu_context(struct pmu *pmu, int cpu);
1151 extern struct event_constraint emptyconstraint;
1153 extern struct event_constraint unconstrained;
1155 static inline bool kernel_ip(unsigned long ip)
1157 #ifdef CONFIG_X86_32
1158 return ip > PAGE_OFFSET;
1160 return (long)ip < 0;
1165 * Not all PMUs provide the right context information to place the reported IP
1166 * into full context. Specifically segment registers are typically not
1169 * Assuming the address is a linear address (it is for IBS), we fake the CS and
1170 * vm86 mode using the known zero-based code segment and 'fix up' the registers
1173 * Intel PEBS/LBR appear to typically provide the effective address, nothing
1174 * much we can do about that but pray and treat it like a linear address.
1176 static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip)
1178 regs->cs = kernel_ip(ip) ? __KERNEL_CS : __USER_CS;
1179 if (regs->flags & X86_VM_MASK)
1180 regs->flags ^= (PERF_EFLAGS_VM | X86_VM_MASK);
1184 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event);
1185 ssize_t intel_event_sysfs_show(char *page, u64 config);
1187 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
1189 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr,
1191 ssize_t events_hybrid_sysfs_show(struct device *dev,
1192 struct device_attribute *attr,
1195 static inline bool fixed_counter_disabled(int i, struct pmu *pmu)
1197 u64 intel_ctrl = hybrid(pmu, intel_ctrl);
1199 return !(intel_ctrl >> (i + INTEL_PMC_IDX_FIXED));
1202 #ifdef CONFIG_CPU_SUP_AMD
1204 int amd_pmu_init(void);
1206 #else /* CONFIG_CPU_SUP_AMD */
1208 static inline int amd_pmu_init(void)
1213 #endif /* CONFIG_CPU_SUP_AMD */
1215 static inline int is_pebs_pt(struct perf_event *event)
1217 return !!(event->hw.flags & PERF_X86_EVENT_PEBS_VIA_PT);
1220 #ifdef CONFIG_CPU_SUP_INTEL
1222 static inline bool intel_pmu_has_bts_period(struct perf_event *event, u64 period)
1224 struct hw_perf_event *hwc = &event->hw;
1225 unsigned int hw_event, bts_event;
1227 if (event->attr.freq)
1230 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1231 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1233 return hw_event == bts_event && period == 1;
1236 static inline bool intel_pmu_has_bts(struct perf_event *event)
1238 struct hw_perf_event *hwc = &event->hw;
1240 return intel_pmu_has_bts_period(event, hwc->sample_period);
1243 int intel_pmu_save_and_restart(struct perf_event *event);
1245 struct event_constraint *
1246 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
1247 struct perf_event *event);
1249 extern int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu);
1250 extern void intel_cpuc_finish(struct cpu_hw_events *cpuc);
1252 int intel_pmu_init(void);
1254 void init_debug_store_on_cpu(int cpu);
1256 void fini_debug_store_on_cpu(int cpu);
1258 void release_ds_buffers(void);
1260 void reserve_ds_buffers(void);
1262 void release_lbr_buffers(void);
1264 void reserve_lbr_buffers(void);
1266 extern struct event_constraint bts_constraint;
1267 extern struct event_constraint vlbr_constraint;
1269 void intel_pmu_enable_bts(u64 config);
1271 void intel_pmu_disable_bts(void);
1273 int intel_pmu_drain_bts_buffer(void);
1275 extern struct event_constraint intel_core2_pebs_event_constraints[];
1277 extern struct event_constraint intel_atom_pebs_event_constraints[];
1279 extern struct event_constraint intel_slm_pebs_event_constraints[];
1281 extern struct event_constraint intel_glm_pebs_event_constraints[];
1283 extern struct event_constraint intel_glp_pebs_event_constraints[];
1285 extern struct event_constraint intel_grt_pebs_event_constraints[];
1287 extern struct event_constraint intel_nehalem_pebs_event_constraints[];
1289 extern struct event_constraint intel_westmere_pebs_event_constraints[];
1291 extern struct event_constraint intel_snb_pebs_event_constraints[];
1293 extern struct event_constraint intel_ivb_pebs_event_constraints[];
1295 extern struct event_constraint intel_hsw_pebs_event_constraints[];
1297 extern struct event_constraint intel_bdw_pebs_event_constraints[];
1299 extern struct event_constraint intel_skl_pebs_event_constraints[];
1301 extern struct event_constraint intel_icl_pebs_event_constraints[];
1303 extern struct event_constraint intel_spr_pebs_event_constraints[];
1305 struct event_constraint *intel_pebs_constraints(struct perf_event *event);
1307 void intel_pmu_pebs_add(struct perf_event *event);
1309 void intel_pmu_pebs_del(struct perf_event *event);
1311 void intel_pmu_pebs_enable(struct perf_event *event);
1313 void intel_pmu_pebs_disable(struct perf_event *event);
1315 void intel_pmu_pebs_enable_all(void);
1317 void intel_pmu_pebs_disable_all(void);
1319 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in);
1321 void intel_pmu_auto_reload_read(struct perf_event *event);
1323 void intel_pmu_store_pebs_lbrs(struct lbr_entry *lbr);
1325 void intel_ds_init(void);
1327 void intel_pmu_lbr_swap_task_ctx(struct perf_event_context *prev,
1328 struct perf_event_context *next);
1330 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in);
1332 u64 lbr_from_signext_quirk_wr(u64 val);
1334 void intel_pmu_lbr_reset(void);
1336 void intel_pmu_lbr_reset_32(void);
1338 void intel_pmu_lbr_reset_64(void);
1340 void intel_pmu_lbr_add(struct perf_event *event);
1342 void intel_pmu_lbr_del(struct perf_event *event);
1344 void intel_pmu_lbr_enable_all(bool pmi);
1346 void intel_pmu_lbr_disable_all(void);
1348 void intel_pmu_lbr_read(void);
1350 void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc);
1352 void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc);
1354 void intel_pmu_lbr_save(void *ctx);
1356 void intel_pmu_lbr_restore(void *ctx);
1358 void intel_pmu_lbr_init_core(void);
1360 void intel_pmu_lbr_init_nhm(void);
1362 void intel_pmu_lbr_init_atom(void);
1364 void intel_pmu_lbr_init_slm(void);
1366 void intel_pmu_lbr_init_snb(void);
1368 void intel_pmu_lbr_init_hsw(void);
1370 void intel_pmu_lbr_init_skl(void);
1372 void intel_pmu_lbr_init_knl(void);
1374 void intel_pmu_arch_lbr_init(void);
1376 void intel_pmu_pebs_data_source_nhm(void);
1378 void intel_pmu_pebs_data_source_skl(bool pmem);
1380 int intel_pmu_setup_lbr_filter(struct perf_event *event);
1382 void intel_pt_interrupt(void);
1384 int intel_bts_interrupt(void);
1386 void intel_bts_enable_local(void);
1388 void intel_bts_disable_local(void);
1390 int p4_pmu_init(void);
1392 int p6_pmu_init(void);
1394 int knc_pmu_init(void);
1396 static inline int is_ht_workaround_enabled(void)
1398 return !!(x86_pmu.flags & PMU_FL_EXCL_ENABLED);
1401 #else /* CONFIG_CPU_SUP_INTEL */
1403 static inline void reserve_ds_buffers(void)
1407 static inline void release_ds_buffers(void)
1411 static inline void release_lbr_buffers(void)
1415 static inline void reserve_lbr_buffers(void)
1419 static inline int intel_pmu_init(void)
1424 static inline int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu)
1429 static inline void intel_cpuc_finish(struct cpu_hw_events *cpuc)
1433 static inline int is_ht_workaround_enabled(void)
1437 #endif /* CONFIG_CPU_SUP_INTEL */
1439 #if ((defined CONFIG_CPU_SUP_CENTAUR) || (defined CONFIG_CPU_SUP_ZHAOXIN))
1440 int zhaoxin_pmu_init(void);
1442 static inline int zhaoxin_pmu_init(void)
1446 #endif /*CONFIG_CPU_SUP_CENTAUR or CONFIG_CPU_SUP_ZHAOXIN*/