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 <pzijlstr@redhat.com>
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>
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 event_constraint 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 */
73 int nb_id; /* NorthBridge id */
74 int refcnt; /* reference count */
75 struct perf_event *owners[X86_PMC_IDX_MAX];
76 struct event_constraint event_constraints[X86_PMC_IDX_MAX];
79 /* The maximal number of PEBS events: */
80 #define MAX_PEBS_EVENTS 8
83 * A debug store configuration.
85 * We only support architectures that use 64bit fields.
90 u64 bts_absolute_maximum;
91 u64 bts_interrupt_threshold;
94 u64 pebs_absolute_maximum;
95 u64 pebs_interrupt_threshold;
96 u64 pebs_event_reset[MAX_PEBS_EVENTS];
100 * Per register state.
103 raw_spinlock_t lock; /* per-core: protect structure */
104 u64 config; /* extra MSR config */
105 u64 reg; /* extra MSR number */
106 atomic_t ref; /* reference count */
112 * Used to coordinate shared registers between HT threads or
113 * among events on a single PMU.
115 struct intel_shared_regs {
116 struct er_account regs[EXTRA_REG_MAX];
117 int refcnt; /* per-core: #HT threads */
118 unsigned core_id; /* per-core: core id */
121 #define MAX_LBR_ENTRIES 16
123 struct cpu_hw_events {
125 * Generic x86 PMC bits
127 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */
128 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
129 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
135 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
136 u64 tags[X86_PMC_IDX_MAX];
137 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
139 unsigned int group_flag;
143 * Intel DebugStore bits
145 struct debug_store *ds;
153 struct perf_branch_stack lbr_stack;
154 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES];
155 struct er_account *lbr_sel;
159 * Intel host/guest exclude bits
161 u64 intel_ctrl_guest_mask;
162 u64 intel_ctrl_host_mask;
163 struct perf_guest_switch_msr guest_switch_msrs[X86_PMC_IDX_MAX];
166 * manage shared (per-core, per-cpu) registers
167 * used on Intel NHM/WSM/SNB
169 struct intel_shared_regs *shared_regs;
174 struct amd_nb *amd_nb;
175 /* Inverted mask of bits to clear in the perf_ctr ctrl registers */
176 u64 perf_ctr_virt_mask;
178 void *kfree_on_online;
181 #define __EVENT_CONSTRAINT(c, n, m, w, o, f) {\
182 { .idxmsk64 = (n) }, \
190 #define EVENT_CONSTRAINT(c, n, m) \
191 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 0, 0)
194 * The overlap flag marks event constraints with overlapping counter
195 * masks. This is the case if the counter mask of such an event is not
196 * a subset of any other counter mask of a constraint with an equal or
197 * higher weight, e.g.:
199 * c_overlaps = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
200 * c_another1 = EVENT_CONSTRAINT(0, 0x07, 0);
201 * c_another2 = EVENT_CONSTRAINT(0, 0x38, 0);
203 * The event scheduler may not select the correct counter in the first
204 * cycle because it needs to know which subsequent events will be
205 * scheduled. It may fail to schedule the events then. So we set the
206 * overlap flag for such constraints to give the scheduler a hint which
207 * events to select for counter rescheduling.
209 * Care must be taken as the rescheduling algorithm is O(n!) which
210 * will increase scheduling cycles for an over-commited system
211 * dramatically. The number of such EVENT_CONSTRAINT_OVERLAP() macros
212 * and its counter masks must be kept at a minimum.
214 #define EVENT_CONSTRAINT_OVERLAP(c, n, m) \
215 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 1, 0)
218 * Constraint on the Event code.
220 #define INTEL_EVENT_CONSTRAINT(c, n) \
221 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT)
224 * Constraint on the Event code + UMask + fixed-mask
226 * filter mask to validate fixed counter events.
227 * the following filters disqualify for fixed counters:
232 * - in_tx_checkpointed
233 * The other filters are supported by fixed counters.
234 * The any-thread option is supported starting with v3.
236 #define FIXED_EVENT_FLAGS (X86_RAW_EVENT_MASK|HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)
237 #define FIXED_EVENT_CONSTRAINT(c, n) \
238 EVENT_CONSTRAINT(c, (1ULL << (32+n)), FIXED_EVENT_FLAGS)
241 * Constraint on the Event code + UMask
243 #define INTEL_UEVENT_CONSTRAINT(c, n) \
244 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK)
246 #define INTEL_PLD_CONSTRAINT(c, n) \
247 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
248 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LDLAT)
250 #define INTEL_PST_CONSTRAINT(c, n) \
251 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
252 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST)
254 /* DataLA version of store sampling without extra enable bit. */
255 #define INTEL_PST_HSW_CONSTRAINT(c, n) \
256 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \
257 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW)
259 #define EVENT_CONSTRAINT_END \
260 EVENT_CONSTRAINT(0, 0, 0)
262 #define for_each_event_constraint(e, c) \
263 for ((e) = (c); (e)->weight; (e)++)
266 * Extra registers for specific events.
268 * Some events need large masks and require external MSRs.
269 * Those extra MSRs end up being shared for all events on
270 * a PMU and sometimes between PMU of sibling HT threads.
271 * In either case, the kernel needs to handle conflicting
272 * accesses to those extra, shared, regs. The data structure
273 * to manage those registers is stored in cpu_hw_event.
280 int idx; /* per_xxx->regs[] reg index */
283 #define EVENT_EXTRA_REG(e, ms, m, vm, i) { \
286 .config_mask = (m), \
287 .valid_mask = (vm), \
288 .idx = EXTRA_REG_##i, \
291 #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx) \
292 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx)
294 #define INTEL_UEVENT_EXTRA_REG(event, msr, vm, idx) \
295 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \
296 ARCH_PERFMON_EVENTSEL_UMASK, vm, idx)
298 #define INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(c) \
299 INTEL_UEVENT_EXTRA_REG(c, \
300 MSR_PEBS_LD_LAT_THRESHOLD, \
304 #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0)
306 union perf_capabilities {
314 * PMU supports separate counter range for writing
317 u64 full_width_write:1;
322 struct x86_pmu_quirk {
323 struct x86_pmu_quirk *next;
327 union x86_pmu_config {
348 #define X86_CONFIG(args...) ((union x86_pmu_config){.bits = {args}}).value
351 * struct x86_pmu - generic x86 pmu
355 * Generic x86 PMC bits
359 int (*handle_irq)(struct pt_regs *);
360 void (*disable_all)(void);
361 void (*enable_all)(int added);
362 void (*enable)(struct perf_event *);
363 void (*disable)(struct perf_event *);
364 int (*hw_config)(struct perf_event *event);
365 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign);
368 int (*addr_offset)(int index, bool eventsel);
369 int (*rdpmc_index)(int index);
370 u64 (*event_map)(int);
373 int num_counters_fixed;
377 unsigned long events_maskl;
378 unsigned long events_mask[BITS_TO_LONGS(ARCH_PERFMON_EVENTS_COUNT)];
383 struct event_constraint *
384 (*get_event_constraints)(struct cpu_hw_events *cpuc,
385 struct perf_event *event);
387 void (*put_event_constraints)(struct cpu_hw_events *cpuc,
388 struct perf_event *event);
389 struct event_constraint *event_constraints;
390 struct x86_pmu_quirk *quirks;
391 int perfctr_second_write;
398 struct attribute **format_attrs;
399 struct attribute **event_attrs;
401 ssize_t (*events_sysfs_show)(char *page, u64 config);
402 struct attribute **cpu_events;
407 int (*cpu_prepare)(int cpu);
408 void (*cpu_starting)(int cpu);
409 void (*cpu_dying)(int cpu);
410 void (*cpu_dead)(int cpu);
412 void (*check_microcode)(void);
413 void (*flush_branch_stack)(void);
416 * Intel Arch Perfmon v2+
419 union perf_capabilities intel_cap;
422 * Intel DebugStore bits
429 int pebs_record_size;
430 void (*drain_pebs)(struct pt_regs *regs);
431 struct event_constraint *pebs_constraints;
432 void (*pebs_aliases)(struct perf_event *event);
438 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */
439 int lbr_nr; /* hardware stack size */
440 u64 lbr_sel_mask; /* LBR_SELECT valid bits */
441 const int *lbr_sel_map; /* lbr_select mappings */
444 * Extra registers for events
446 struct extra_reg *extra_regs;
447 unsigned int er_flags;
450 * Intel host/guest support (KVM)
452 struct perf_guest_switch_msr *(*guest_get_msrs)(int *nr);
455 #define x86_add_quirk(func_) \
457 static struct x86_pmu_quirk __quirk __initdata = { \
460 __quirk.next = x86_pmu.quirks; \
461 x86_pmu.quirks = &__quirk; \
464 #define ERF_NO_HT_SHARING 1
465 #define ERF_HAS_RSP_1 2
467 #define EVENT_VAR(_id) event_attr_##_id
468 #define EVENT_PTR(_id) &event_attr_##_id.attr.attr
470 #define EVENT_ATTR(_name, _id) \
471 static struct perf_pmu_events_attr EVENT_VAR(_id) = { \
472 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
473 .id = PERF_COUNT_HW_##_id, \
477 #define EVENT_ATTR_STR(_name, v, str) \
478 static struct perf_pmu_events_attr event_attr_##v = { \
479 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \
484 extern struct x86_pmu x86_pmu __read_mostly;
486 DECLARE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
488 int x86_perf_event_set_period(struct perf_event *event);
491 * Generalized hw caching related hw_event table, filled
492 * in on a per model basis. A value of 0 means
493 * 'not supported', -1 means 'hw_event makes no sense on
494 * this CPU', any other value means the raw hw_event
498 #define C(x) PERF_COUNT_HW_CACHE_##x
500 extern u64 __read_mostly hw_cache_event_ids
501 [PERF_COUNT_HW_CACHE_MAX]
502 [PERF_COUNT_HW_CACHE_OP_MAX]
503 [PERF_COUNT_HW_CACHE_RESULT_MAX];
504 extern u64 __read_mostly hw_cache_extra_regs
505 [PERF_COUNT_HW_CACHE_MAX]
506 [PERF_COUNT_HW_CACHE_OP_MAX]
507 [PERF_COUNT_HW_CACHE_RESULT_MAX];
509 u64 x86_perf_event_update(struct perf_event *event);
511 static inline unsigned int x86_pmu_config_addr(int index)
513 return x86_pmu.eventsel + (x86_pmu.addr_offset ?
514 x86_pmu.addr_offset(index, true) : index);
517 static inline unsigned int x86_pmu_event_addr(int index)
519 return x86_pmu.perfctr + (x86_pmu.addr_offset ?
520 x86_pmu.addr_offset(index, false) : index);
523 static inline int x86_pmu_rdpmc_index(int index)
525 return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index(index) : index;
528 int x86_setup_perfctr(struct perf_event *event);
530 int x86_pmu_hw_config(struct perf_event *event);
532 void x86_pmu_disable_all(void);
534 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc,
537 u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
539 if (hwc->extra_reg.reg)
540 wrmsrl(hwc->extra_reg.reg, hwc->extra_reg.config);
541 wrmsrl(hwc->config_base, (hwc->config | enable_mask) & ~disable_mask);
544 void x86_pmu_enable_all(int added);
546 int perf_assign_events(struct perf_event **events, int n,
547 int wmin, int wmax, int *assign);
548 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign);
550 void x86_pmu_stop(struct perf_event *event, int flags);
552 static inline void x86_pmu_disable_event(struct perf_event *event)
554 struct hw_perf_event *hwc = &event->hw;
556 wrmsrl(hwc->config_base, hwc->config);
559 void x86_pmu_enable_event(struct perf_event *event);
561 int x86_pmu_handle_irq(struct pt_regs *regs);
563 extern struct event_constraint emptyconstraint;
565 extern struct event_constraint unconstrained;
567 static inline bool kernel_ip(unsigned long ip)
570 return ip > PAGE_OFFSET;
577 * Not all PMUs provide the right context information to place the reported IP
578 * into full context. Specifically segment registers are typically not
581 * Assuming the address is a linear address (it is for IBS), we fake the CS and
582 * vm86 mode using the known zero-based code segment and 'fix up' the registers
585 * Intel PEBS/LBR appear to typically provide the effective address, nothing
586 * much we can do about that but pray and treat it like a linear address.
588 static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip)
590 regs->cs = kernel_ip(ip) ? __KERNEL_CS : __USER_CS;
591 if (regs->flags & X86_VM_MASK)
592 regs->flags ^= (PERF_EFLAGS_VM | X86_VM_MASK);
596 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event);
597 ssize_t intel_event_sysfs_show(char *page, u64 config);
599 #ifdef CONFIG_CPU_SUP_AMD
601 int amd_pmu_init(void);
603 #else /* CONFIG_CPU_SUP_AMD */
605 static inline int amd_pmu_init(void)
610 #endif /* CONFIG_CPU_SUP_AMD */
612 #ifdef CONFIG_CPU_SUP_INTEL
614 int intel_pmu_save_and_restart(struct perf_event *event);
616 struct event_constraint *
617 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event);
619 struct intel_shared_regs *allocate_shared_regs(int cpu);
621 int intel_pmu_init(void);
623 void init_debug_store_on_cpu(int cpu);
625 void fini_debug_store_on_cpu(int cpu);
627 void release_ds_buffers(void);
629 void reserve_ds_buffers(void);
631 extern struct event_constraint bts_constraint;
633 void intel_pmu_enable_bts(u64 config);
635 void intel_pmu_disable_bts(void);
637 int intel_pmu_drain_bts_buffer(void);
639 extern struct event_constraint intel_core2_pebs_event_constraints[];
641 extern struct event_constraint intel_atom_pebs_event_constraints[];
643 extern struct event_constraint intel_nehalem_pebs_event_constraints[];
645 extern struct event_constraint intel_westmere_pebs_event_constraints[];
647 extern struct event_constraint intel_snb_pebs_event_constraints[];
649 extern struct event_constraint intel_ivb_pebs_event_constraints[];
651 extern struct event_constraint intel_hsw_pebs_event_constraints[];
653 struct event_constraint *intel_pebs_constraints(struct perf_event *event);
655 void intel_pmu_pebs_enable(struct perf_event *event);
657 void intel_pmu_pebs_disable(struct perf_event *event);
659 void intel_pmu_pebs_enable_all(void);
661 void intel_pmu_pebs_disable_all(void);
663 void intel_ds_init(void);
665 void intel_pmu_lbr_reset(void);
667 void intel_pmu_lbr_enable(struct perf_event *event);
669 void intel_pmu_lbr_disable(struct perf_event *event);
671 void intel_pmu_lbr_enable_all(void);
673 void intel_pmu_lbr_disable_all(void);
675 void intel_pmu_lbr_read(void);
677 void intel_pmu_lbr_init_core(void);
679 void intel_pmu_lbr_init_nhm(void);
681 void intel_pmu_lbr_init_atom(void);
683 void intel_pmu_lbr_init_snb(void);
685 int intel_pmu_setup_lbr_filter(struct perf_event *event);
687 int p4_pmu_init(void);
689 int p6_pmu_init(void);
691 int knc_pmu_init(void);
693 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr,
696 #else /* CONFIG_CPU_SUP_INTEL */
698 static inline void reserve_ds_buffers(void)
702 static inline void release_ds_buffers(void)
706 static inline int intel_pmu_init(void)
711 static inline struct intel_shared_regs *allocate_shared_regs(int cpu)
716 #endif /* CONFIG_CPU_SUP_INTEL */