2 * Performance events x86 architecture header
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
10 * Copyright (C) 2009 Google, Inc., Stephane Eranian
12 * For licencing details see kernel-base/COPYING
15 #include <linux/perf_event.h>
19 #define wrmsrl(msr, val) \
21 unsigned int _msr = (msr); \
23 trace_printk("wrmsrl(%x, %Lx)\n", (unsigned int)(_msr), \
24 (unsigned long long)(_val)); \
25 native_write_msr((_msr), (u32)(_val), (u32)(_val >> 32)); \
31 * register -------------------------------
32 * | HT | no HT | HT | no HT |
33 *-----------------------------------------
34 * offcore | core | core | cpu | core |
35 * lbr_sel | core | core | cpu | core |
36 * ld_lat | cpu | core | cpu | core |
37 *-----------------------------------------
39 * Given that there is a small number of shared regs,
40 * we can pre-allocate their slot in the per-cpu
41 * per-core reg tables.
44 EXTRA_REG_NONE = -1, /* not used */
46 EXTRA_REG_RSP_0 = 0, /* offcore_response_0 */
47 EXTRA_REG_RSP_1 = 1, /* offcore_response_1 */
48 EXTRA_REG_LBR = 2, /* lbr_select */
49 EXTRA_REG_LDLAT = 3, /* ld_lat_threshold */
51 EXTRA_REG_MAX /* number of entries needed */
54 struct event_constraint {
56 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
66 * struct hw_perf_event.flags flags
68 #define PERF_X86_EVENT_PEBS_LDLAT 0x1 /* ld+ldlat data address sampling */
69 #define PERF_X86_EVENT_PEBS_ST 0x2 /* st data address sampling */
70 #define PERF_X86_EVENT_PEBS_ST_HSW 0x4 /* haswell style st data sampling */
71 #define PERF_X86_EVENT_COMMITTED 0x8 /* event passed commit_txn */
74 int nb_id; /* NorthBridge id */
75 int refcnt; /* reference count */
76 struct perf_event *owners[X86_PMC_IDX_MAX];
77 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
80 /* The maximal number of PEBS events: */
81 #define MAX_PEBS_EVENTS 8
84 * A debug store configuration.
86 * We only support architectures that use 64bit fields.
91 u64 bts_absolute_maximum;
92 u64 bts_interrupt_threshold;
95 u64 pebs_absolute_maximum;
96 u64 pebs_interrupt_threshold;
97 u64 pebs_event_reset[MAX_PEBS_EVENTS];
101 * Per register state.
104 raw_spinlock_t lock; /* per-core: protect structure */
105 u64 config; /* extra MSR config */
106 u64 reg; /* extra MSR number */
107 atomic_t ref; /* reference count */
113 * Used to coordinate shared registers between HT threads or
114 * among events on a single PMU.
116 struct intel_shared_regs {
117 struct er_account regs[EXTRA_REG_MAX];
118 int refcnt; /* per-core: #HT threads */
119 unsigned core_id; /* per-core: core id */
122 #define MAX_LBR_ENTRIES 16
124 struct cpu_hw_events {
126 * Generic x86 PMC bits
128 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
129 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
130 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
136 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
137 u64 tags[X86_PMC_IDX_MAX];
138 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
140 unsigned int group_flag;
144 * Intel DebugStore bits
146 struct debug_store *ds;
154 struct perf_branch_stack lbr_stack;
155 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
156 struct er_account *lbr_sel;
160 * Intel host/guest exclude bits
162 u64 intel_ctrl_guest_mask;
163 u64 intel_ctrl_host_mask;
164 struct perf_guest_switch_msr guest_switch_msrs[X86_PMC_IDX_MAX];
167 * Intel checkpoint mask
172 * manage shared (per-core, per-cpu) registers
173 * used on Intel NHM/WSM/SNB
175 struct intel_shared_regs *shared_regs;
180 struct amd_nb *amd_nb;
181 /* Inverted mask of bits to clear in the perf_ctr ctrl registers */
182 u64 perf_ctr_virt_mask;
184 void *kfree_on_online;
187 #define __EVENT_CONSTRAINT(c, n, m, w, o, f) {\
188 { .idxmsk64 = (n) }, \
196 #define EVENT_CONSTRAINT(c, n, m) \
197 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 0, 0)
200 * The overlap flag marks event constraints with overlapping counter
201 * masks. This is the case if the counter mask of such an event is not
202 * a subset of any other counter mask of a constraint with an equal or
203 * higher weight, e.g.:
205 * c_overlaps = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
206 * c_another1 = EVENT_CONSTRAINT(0, 0x07, 0);
207 * c_another2 = EVENT_CONSTRAINT(0, 0x38, 0);
209 * The event scheduler may not select the correct counter in the first
210 * cycle because it needs to know which subsequent events will be
211 * scheduled. It may fail to schedule the events then. So we set the
212 * overlap flag for such constraints to give the scheduler a hint which
213 * events to select for counter rescheduling.
215 * Care must be taken as the rescheduling algorithm is O(n!) which
216 * will increase scheduling cycles for an over-commited system
217 * dramatically. The number of such EVENT_CONSTRAINT_OVERLAP() macros
218 * and its counter masks must be kept at a minimum.
220 #define EVENT_CONSTRAINT_OVERLAP(c, n, m) \
221 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 1, 0)
224 * Constraint on the Event code.
226 #define INTEL_EVENT_CONSTRAINT(c, n) \
227 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
230 * Constraint on the Event code + UMask + fixed-mask
232 * filter mask to validate fixed counter events.
233 * the following filters disqualify for fixed counters:
238 * - in_tx_checkpointed
239 * The other filters are supported by fixed counters.
240 * The any-thread option is supported starting with v3.
242 #define FIXED_EVENT_FLAGS (X86_RAW_EVENT_MASK|HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)
243 #define FIXED_EVENT_CONSTRAINT(c, n) \
244 EVENT_CONSTRAINT(c, (1ULL << (32+n)), FIXED_EVENT_FLAGS)
247 * Constraint on the Event code + UMask
249 #define INTEL_UEVENT_CONSTRAINT(c, n) \
250 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
252 #define INTEL_PLD_CONSTRAINT(c, n) \
253 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
254 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LDLAT)
256 #define INTEL_PST_CONSTRAINT(c, n) \
257 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
258 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST)
260 /* DataLA version of store sampling without extra enable bit. */
261 #define INTEL_PST_HSW_CONSTRAINT(c, n) \
262 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
263 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW)
266 * We define the end marker as having a weight of -1
267 * to enable blacklisting of events using a counter bitmask
268 * of zero and thus a weight of zero.
269 * The end marker has a weight that cannot possibly be
270 * obtained from counting the bits in the bitmask.
272 #define EVENT_CONSTRAINT_END { .weight = -1 }
275 * Check for end marker with weight == -1
277 #define for_each_event_constraint(e, c) \
278 for ((e) = (c); (e)->weight != -1; (e)++)
281 * Extra registers for specific events.
283 * Some events need large masks and require external MSRs.
284 * Those extra MSRs end up being shared for all events on
285 * a PMU and sometimes between PMU of sibling HT threads.
286 * In either case, the kernel needs to handle conflicting
287 * accesses to those extra, shared, regs. The data structure
288 * to manage those registers is stored in cpu_hw_event.
295 int idx; /* per_xxx->regs[] reg index */
298 #define EVENT_EXTRA_REG(e, ms, m, vm, i) { \
301 .config_mask = (m), \
302 .valid_mask = (vm), \
303 .idx = EXTRA_REG_##i, \
306 #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx) \
307 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx)
309 #define INTEL_UEVENT_EXTRA_REG(event, msr, vm, idx) \
310 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
311 ARCH_PERFMON_EVENTSEL_UMASK, vm, idx)
313 #define INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(c) \
314 INTEL_UEVENT_EXTRA_REG(c, \
315 MSR_PEBS_LD_LAT_THRESHOLD, \
319 #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0)
321 union perf_capabilities {
329 * PMU supports separate counter range for writing
332 u64 full_width_write:1;
337 struct x86_pmu_quirk {
338 struct x86_pmu_quirk *next;
342 union x86_pmu_config {
363 #define X86_CONFIG(args...) ((union x86_pmu_config){.bits = {args}}).value
366 * struct x86_pmu - generic x86 pmu
370 * Generic x86 PMC bits
374 int (*handle_irq)(struct pt_regs *);
375 void (*disable_all)(void);
376 void (*enable_all)(int added);
377 void (*enable)(struct perf_event *);
378 void (*disable)(struct perf_event *);
379 int (*hw_config)(struct perf_event *event);
380 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
383 int (*addr_offset)(int index, bool eventsel);
384 int (*rdpmc_index)(int index);
385 u64 (*event_map)(int);
388 int num_counters_fixed;
392 unsigned long events_maskl;
393 unsigned long events_mask[BITS_TO_LONGS(ARCH_PERFMON_EVENTS_COUNT)];
398 struct event_constraint *
399 (*get_event_constraints)(struct cpu_hw_events *cpuc,
400 struct perf_event *event);
402 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
403 struct perf_event *event);
404 struct event_constraint *event_constraints;
405 struct x86_pmu_quirk *quirks;
406 int perfctr_second_write;
413 struct attribute **format_attrs;
414 struct attribute **event_attrs;
416 ssize_t (*events_sysfs_show)(char *page, u64 config);
417 struct attribute **cpu_events;
422 int (*cpu_prepare)(int cpu);
423 void (*cpu_starting)(int cpu);
424 void (*cpu_dying)(int cpu);
425 void (*cpu_dead)(int cpu);
427 void (*check_microcode)(void);
428 void (*flush_branch_stack)(void);
431 * Intel Arch Perfmon v2+
434 union perf_capabilities intel_cap;
437 * Intel DebugStore bits
444 int pebs_record_size;
445 void (*drain_pebs)(struct pt_regs *regs);
446 struct event_constraint *pebs_constraints;
447 void (*pebs_aliases)(struct perf_event *event);
453 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
454 int lbr_nr; /* hardware stack size */
455 u64 lbr_sel_mask; /* LBR_SELECT valid bits */
456 const int *lbr_sel_map; /* lbr_select mappings */
457 bool lbr_double_abort; /* duplicated lbr aborts */
460 * Extra registers for events
462 struct extra_reg *extra_regs;
463 unsigned int er_flags;
466 * Intel host/guest support (KVM)
468 struct perf_guest_switch_msr *(*guest_get_msrs)(int *nr);
471 #define x86_add_quirk(func_) \
473 static struct x86_pmu_quirk __quirk __initdata = { \
476 __quirk.next = x86_pmu.quirks; \
477 x86_pmu.quirks = &__quirk; \
480 #define ERF_NO_HT_SHARING 1
481 #define ERF_HAS_RSP_1 2
483 #define EVENT_VAR(_id) event_attr_##_id
484 #define EVENT_PTR(_id) &event_attr_##_id.attr.attr
486 #define EVENT_ATTR(_name, _id) \
487 static struct perf_pmu_events_attr EVENT_VAR(_id) = { \
488 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
489 .id = PERF_COUNT_HW_##_id, \
493 #define EVENT_ATTR_STR(_name, v, str) \
494 static struct perf_pmu_events_attr event_attr_##v = { \
495 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
500 extern struct x86_pmu x86_pmu __read_mostly;
502 DECLARE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
504 int x86_perf_event_set_period(struct perf_event *event);
507 * Generalized hw caching related hw_event table, filled
508 * in on a per model basis. A value of 0 means
509 * 'not supported', -1 means 'hw_event makes no sense on
510 * this CPU', any other value means the raw hw_event
514 #define C(x) PERF_COUNT_HW_CACHE_##x
516 extern u64 __read_mostly hw_cache_event_ids
517 [PERF_COUNT_HW_CACHE_MAX]
518 [PERF_COUNT_HW_CACHE_OP_MAX]
519 [PERF_COUNT_HW_CACHE_RESULT_MAX];
520 extern u64 __read_mostly hw_cache_extra_regs
521 [PERF_COUNT_HW_CACHE_MAX]
522 [PERF_COUNT_HW_CACHE_OP_MAX]
523 [PERF_COUNT_HW_CACHE_RESULT_MAX];
525 u64 x86_perf_event_update(struct perf_event *event);
527 static inline unsigned int x86_pmu_config_addr(int index)
529 return x86_pmu.eventsel + (x86_pmu.addr_offset ?
530 x86_pmu.addr_offset(index, true) : index);
533 static inline unsigned int x86_pmu_event_addr(int index)
535 return x86_pmu.perfctr + (x86_pmu.addr_offset ?
536 x86_pmu.addr_offset(index, false) : index);
539 static inline int x86_pmu_rdpmc_index(int index)
541 return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index(index) : index;
544 int x86_setup_perfctr(struct perf_event *event);
546 int x86_pmu_hw_config(struct perf_event *event);
548 void x86_pmu_disable_all(void);
550 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
553 u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
555 if (hwc->extra_reg.reg)
556 wrmsrl(hwc->extra_reg.reg, hwc->extra_reg.config);
557 wrmsrl(hwc->config_base, (hwc->config | enable_mask) & ~disable_mask);
560 void x86_pmu_enable_all(int added);
562 int perf_assign_events(struct perf_event **events, int n,
563 int wmin, int wmax, int *assign);
564 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign);
566 void x86_pmu_stop(struct perf_event *event, int flags);
568 static inline void x86_pmu_disable_event(struct perf_event *event)
570 struct hw_perf_event *hwc = &event->hw;
572 wrmsrl(hwc->config_base, hwc->config);
575 void x86_pmu_enable_event(struct perf_event *event);
577 int x86_pmu_handle_irq(struct pt_regs *regs);
579 extern struct event_constraint emptyconstraint;
581 extern struct event_constraint unconstrained;
583 static inline bool kernel_ip(unsigned long ip)
586 return ip > PAGE_OFFSET;
593 * Not all PMUs provide the right context information to place the reported IP
594 * into full context. Specifically segment registers are typically not
597 * Assuming the address is a linear address (it is for IBS), we fake the CS and
598 * vm86 mode using the known zero-based code segment and 'fix up' the registers
601 * Intel PEBS/LBR appear to typically provide the effective address, nothing
602 * much we can do about that but pray and treat it like a linear address.
604 static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip)
606 regs->cs = kernel_ip(ip) ? __KERNEL_CS : __USER_CS;
607 if (regs->flags & X86_VM_MASK)
608 regs->flags ^= (PERF_EFLAGS_VM | X86_VM_MASK);
612 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event);
613 ssize_t intel_event_sysfs_show(char *page, u64 config);
615 #ifdef CONFIG_CPU_SUP_AMD
617 int amd_pmu_init(void);
619 #else /* CONFIG_CPU_SUP_AMD */
621 static inline int amd_pmu_init(void)
626 #endif /* CONFIG_CPU_SUP_AMD */
628 #ifdef CONFIG_CPU_SUP_INTEL
630 int intel_pmu_save_and_restart(struct perf_event *event);
632 struct event_constraint *
633 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event);
635 struct intel_shared_regs *allocate_shared_regs(int cpu);
637 int intel_pmu_init(void);
639 void init_debug_store_on_cpu(int cpu);
641 void fini_debug_store_on_cpu(int cpu);
643 void release_ds_buffers(void);
645 void reserve_ds_buffers(void);
647 extern struct event_constraint bts_constraint;
649 void intel_pmu_enable_bts(u64 config);
651 void intel_pmu_disable_bts(void);
653 int intel_pmu_drain_bts_buffer(void);
655 extern struct event_constraint intel_core2_pebs_event_constraints[];
657 extern struct event_constraint intel_atom_pebs_event_constraints[];
659 extern struct event_constraint intel_slm_pebs_event_constraints[];
661 extern struct event_constraint intel_nehalem_pebs_event_constraints[];
663 extern struct event_constraint intel_westmere_pebs_event_constraints[];
665 extern struct event_constraint intel_snb_pebs_event_constraints[];
667 extern struct event_constraint intel_ivb_pebs_event_constraints[];
669 extern struct event_constraint intel_hsw_pebs_event_constraints[];
671 struct event_constraint *intel_pebs_constraints(struct perf_event *event);
673 void intel_pmu_pebs_enable(struct perf_event *event);
675 void intel_pmu_pebs_disable(struct perf_event *event);
677 void intel_pmu_pebs_enable_all(void);
679 void intel_pmu_pebs_disable_all(void);
681 void intel_ds_init(void);
683 void intel_pmu_lbr_reset(void);
685 void intel_pmu_lbr_enable(struct perf_event *event);
687 void intel_pmu_lbr_disable(struct perf_event *event);
689 void intel_pmu_lbr_enable_all(void);
691 void intel_pmu_lbr_disable_all(void);
693 void intel_pmu_lbr_read(void);
695 void intel_pmu_lbr_init_core(void);
697 void intel_pmu_lbr_init_nhm(void);
699 void intel_pmu_lbr_init_atom(void);
701 void intel_pmu_lbr_init_snb(void);
703 int intel_pmu_setup_lbr_filter(struct perf_event *event);
705 int p4_pmu_init(void);
707 int p6_pmu_init(void);
709 int knc_pmu_init(void);
711 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
714 #else /* CONFIG_CPU_SUP_INTEL */
716 static inline void reserve_ds_buffers(void)
720 static inline void release_ds_buffers(void)
724 static inline int intel_pmu_init(void)
729 static inline struct intel_shared_regs *allocate_shared_regs(int cpu)
734 #endif /* CONFIG_CPU_SUP_INTEL */