1 #include <linux/bitops.h>
2 #include <linux/types.h>
3 #include <linux/slab.h>
5 #include <asm/perf_event.h>
8 #include "perf_event.h"
10 /* The size of a BTS record in bytes: */
11 #define BTS_RECORD_SIZE 24
13 #define BTS_BUFFER_SIZE (PAGE_SIZE << 4)
14 #define PEBS_BUFFER_SIZE (PAGE_SIZE << 4)
15 #define PEBS_FIXUP_SIZE PAGE_SIZE
18 * pebs_record_32 for p4 and core not supported
20 struct pebs_record_32 {
28 union intel_x86_pebs_dse
{
31 unsigned int ld_dse
:4;
32 unsigned int ld_stlb_miss
:1;
33 unsigned int ld_locked
:1;
34 unsigned int ld_reserved
:26;
37 unsigned int st_l1d_hit
:1;
38 unsigned int st_reserved1
:3;
39 unsigned int st_stlb_miss
:1;
40 unsigned int st_locked
:1;
41 unsigned int st_reserved2
:26;
47 * Map PEBS Load Latency Data Source encodings to generic
48 * memory data source information
50 #define P(a, b) PERF_MEM_S(a, b)
51 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
52 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
54 static const u64 pebs_data_source
[] = {
55 P(OP
, LOAD
) | P(LVL
, MISS
) | P(LVL
, L3
) | P(SNOOP
, NA
),/* 0x00:ukn L3 */
56 OP_LH
| P(LVL
, L1
) | P(SNOOP
, NONE
), /* 0x01: L1 local */
57 OP_LH
| P(LVL
, LFB
) | P(SNOOP
, NONE
), /* 0x02: LFB hit */
58 OP_LH
| P(LVL
, L2
) | P(SNOOP
, NONE
), /* 0x03: L2 hit */
59 OP_LH
| P(LVL
, L3
) | P(SNOOP
, NONE
), /* 0x04: L3 hit */
60 OP_LH
| P(LVL
, L3
) | P(SNOOP
, MISS
), /* 0x05: L3 hit, snoop miss */
61 OP_LH
| P(LVL
, L3
) | P(SNOOP
, HIT
), /* 0x06: L3 hit, snoop hit */
62 OP_LH
| P(LVL
, L3
) | P(SNOOP
, HITM
), /* 0x07: L3 hit, snoop hitm */
63 OP_LH
| P(LVL
, REM_CCE1
) | P(SNOOP
, HIT
), /* 0x08: L3 miss snoop hit */
64 OP_LH
| P(LVL
, REM_CCE1
) | P(SNOOP
, HITM
), /* 0x09: L3 miss snoop hitm*/
65 OP_LH
| P(LVL
, LOC_RAM
) | P(SNOOP
, HIT
), /* 0x0a: L3 miss, shared */
66 OP_LH
| P(LVL
, REM_RAM1
) | P(SNOOP
, HIT
), /* 0x0b: L3 miss, shared */
67 OP_LH
| P(LVL
, LOC_RAM
) | SNOOP_NONE_MISS
,/* 0x0c: L3 miss, excl */
68 OP_LH
| P(LVL
, REM_RAM1
) | SNOOP_NONE_MISS
,/* 0x0d: L3 miss, excl */
69 OP_LH
| P(LVL
, IO
) | P(SNOOP
, NONE
), /* 0x0e: I/O */
70 OP_LH
| P(LVL
, UNC
) | P(SNOOP
, NONE
), /* 0x0f: uncached */
73 static u64
precise_store_data(u64 status
)
75 union intel_x86_pebs_dse dse
;
76 u64 val
= P(OP
, STORE
) | P(SNOOP
, NA
) | P(LVL
, L1
) | P(TLB
, L2
);
82 * 1 = stored missed 2nd level TLB
84 * so it either hit the walker or the OS
85 * otherwise hit 2nd level TLB
93 * bit 0: hit L1 data cache
94 * if not set, then all we know is that
103 * bit 5: Locked prefix
106 val
|= P(LOCK
, LOCKED
);
111 static u64
precise_datala_hsw(struct perf_event
*event
, u64 status
)
113 union perf_mem_data_src dse
;
115 dse
.val
= PERF_MEM_NA
;
117 if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_ST_HSW
)
118 dse
.mem_op
= PERF_MEM_OP_STORE
;
119 else if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_LD_HSW
)
120 dse
.mem_op
= PERF_MEM_OP_LOAD
;
123 * L1 info only valid for following events:
125 * MEM_UOPS_RETIRED.STLB_MISS_STORES
126 * MEM_UOPS_RETIRED.LOCK_STORES
127 * MEM_UOPS_RETIRED.SPLIT_STORES
128 * MEM_UOPS_RETIRED.ALL_STORES
130 if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_ST_HSW
) {
132 dse
.mem_lvl
= PERF_MEM_LVL_L1
| PERF_MEM_LVL_HIT
;
134 dse
.mem_lvl
= PERF_MEM_LVL_L1
| PERF_MEM_LVL_MISS
;
139 static u64
load_latency_data(u64 status
)
141 union intel_x86_pebs_dse dse
;
143 int model
= boot_cpu_data
.x86_model
;
144 int fam
= boot_cpu_data
.x86
;
149 * use the mapping table for bit 0-3
151 val
= pebs_data_source
[dse
.ld_dse
];
154 * Nehalem models do not support TLB, Lock infos
156 if (fam
== 0x6 && (model
== 26 || model
== 30
157 || model
== 31 || model
== 46)) {
158 val
|= P(TLB
, NA
) | P(LOCK
, NA
);
163 * 0 = did not miss 2nd level TLB
164 * 1 = missed 2nd level TLB
166 if (dse
.ld_stlb_miss
)
167 val
|= P(TLB
, MISS
) | P(TLB
, L2
);
169 val
|= P(TLB
, HIT
) | P(TLB
, L1
) | P(TLB
, L2
);
172 * bit 5: locked prefix
175 val
|= P(LOCK
, LOCKED
);
180 struct pebs_record_core
{
184 u64 r8
, r9
, r10
, r11
;
185 u64 r12
, r13
, r14
, r15
;
188 struct pebs_record_nhm
{
192 u64 r8
, r9
, r10
, r11
;
193 u64 r12
, r13
, r14
, r15
;
194 u64 status
, dla
, dse
, lat
;
198 * Same as pebs_record_nhm, with two additional fields.
200 struct pebs_record_hsw
{
204 u64 r8
, r9
, r10
, r11
;
205 u64 r12
, r13
, r14
, r15
;
206 u64 status
, dla
, dse
, lat
;
207 u64 real_ip
, tsx_tuning
;
210 union hsw_tsx_tuning
{
212 u32 cycles_last_block
: 32,
215 instruction_abort
: 1,
216 non_instruction_abort
: 1,
225 #define PEBS_HSW_TSX_FLAGS 0xff00000000ULL
227 /* Same as HSW, plus TSC */
229 struct pebs_record_skl
{
233 u64 r8
, r9
, r10
, r11
;
234 u64 r12
, r13
, r14
, r15
;
235 u64 status
, dla
, dse
, lat
;
236 u64 real_ip
, tsx_tuning
;
240 void init_debug_store_on_cpu(int cpu
)
242 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
247 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
,
248 (u32
)((u64
)(unsigned long)ds
),
249 (u32
)((u64
)(unsigned long)ds
>> 32));
252 void fini_debug_store_on_cpu(int cpu
)
254 if (!per_cpu(cpu_hw_events
, cpu
).ds
)
257 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
, 0, 0);
260 static DEFINE_PER_CPU(void *, insn_buffer
);
262 static int alloc_pebs_buffer(int cpu
)
264 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
265 int node
= cpu_to_node(cpu
);
267 void *buffer
, *ibuffer
;
272 buffer
= kzalloc_node(PEBS_BUFFER_SIZE
, GFP_KERNEL
, node
);
273 if (unlikely(!buffer
))
277 * HSW+ already provides us the eventing ip; no need to allocate this
280 if (x86_pmu
.intel_cap
.pebs_format
< 2) {
281 ibuffer
= kzalloc_node(PEBS_FIXUP_SIZE
, GFP_KERNEL
, node
);
286 per_cpu(insn_buffer
, cpu
) = ibuffer
;
289 max
= PEBS_BUFFER_SIZE
/ x86_pmu
.pebs_record_size
;
291 ds
->pebs_buffer_base
= (u64
)(unsigned long)buffer
;
292 ds
->pebs_index
= ds
->pebs_buffer_base
;
293 ds
->pebs_absolute_maximum
= ds
->pebs_buffer_base
+
294 max
* x86_pmu
.pebs_record_size
;
299 static void release_pebs_buffer(int cpu
)
301 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
303 if (!ds
|| !x86_pmu
.pebs
)
306 kfree(per_cpu(insn_buffer
, cpu
));
307 per_cpu(insn_buffer
, cpu
) = NULL
;
309 kfree((void *)(unsigned long)ds
->pebs_buffer_base
);
310 ds
->pebs_buffer_base
= 0;
313 static int alloc_bts_buffer(int cpu
)
315 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
316 int node
= cpu_to_node(cpu
);
323 buffer
= kzalloc_node(BTS_BUFFER_SIZE
, GFP_KERNEL
| __GFP_NOWARN
, node
);
324 if (unlikely(!buffer
)) {
325 WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__
);
329 max
= BTS_BUFFER_SIZE
/ BTS_RECORD_SIZE
;
332 ds
->bts_buffer_base
= (u64
)(unsigned long)buffer
;
333 ds
->bts_index
= ds
->bts_buffer_base
;
334 ds
->bts_absolute_maximum
= ds
->bts_buffer_base
+
335 max
* BTS_RECORD_SIZE
;
336 ds
->bts_interrupt_threshold
= ds
->bts_absolute_maximum
-
337 thresh
* BTS_RECORD_SIZE
;
342 static void release_bts_buffer(int cpu
)
344 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
346 if (!ds
|| !x86_pmu
.bts
)
349 kfree((void *)(unsigned long)ds
->bts_buffer_base
);
350 ds
->bts_buffer_base
= 0;
353 static int alloc_ds_buffer(int cpu
)
355 int node
= cpu_to_node(cpu
);
356 struct debug_store
*ds
;
358 ds
= kzalloc_node(sizeof(*ds
), GFP_KERNEL
, node
);
362 per_cpu(cpu_hw_events
, cpu
).ds
= ds
;
367 static void release_ds_buffer(int cpu
)
369 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
374 per_cpu(cpu_hw_events
, cpu
).ds
= NULL
;
378 void release_ds_buffers(void)
382 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
386 for_each_online_cpu(cpu
)
387 fini_debug_store_on_cpu(cpu
);
389 for_each_possible_cpu(cpu
) {
390 release_pebs_buffer(cpu
);
391 release_bts_buffer(cpu
);
392 release_ds_buffer(cpu
);
397 void reserve_ds_buffers(void)
399 int bts_err
= 0, pebs_err
= 0;
402 x86_pmu
.bts_active
= 0;
403 x86_pmu
.pebs_active
= 0;
405 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
416 for_each_possible_cpu(cpu
) {
417 if (alloc_ds_buffer(cpu
)) {
422 if (!bts_err
&& alloc_bts_buffer(cpu
))
425 if (!pebs_err
&& alloc_pebs_buffer(cpu
))
428 if (bts_err
&& pebs_err
)
433 for_each_possible_cpu(cpu
)
434 release_bts_buffer(cpu
);
438 for_each_possible_cpu(cpu
)
439 release_pebs_buffer(cpu
);
442 if (bts_err
&& pebs_err
) {
443 for_each_possible_cpu(cpu
)
444 release_ds_buffer(cpu
);
446 if (x86_pmu
.bts
&& !bts_err
)
447 x86_pmu
.bts_active
= 1;
449 if (x86_pmu
.pebs
&& !pebs_err
)
450 x86_pmu
.pebs_active
= 1;
452 for_each_online_cpu(cpu
)
453 init_debug_store_on_cpu(cpu
);
463 struct event_constraint bts_constraint
=
464 EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS
, 0);
466 void intel_pmu_enable_bts(u64 config
)
468 unsigned long debugctlmsr
;
470 debugctlmsr
= get_debugctlmsr();
472 debugctlmsr
|= DEBUGCTLMSR_TR
;
473 debugctlmsr
|= DEBUGCTLMSR_BTS
;
474 if (config
& ARCH_PERFMON_EVENTSEL_INT
)
475 debugctlmsr
|= DEBUGCTLMSR_BTINT
;
477 if (!(config
& ARCH_PERFMON_EVENTSEL_OS
))
478 debugctlmsr
|= DEBUGCTLMSR_BTS_OFF_OS
;
480 if (!(config
& ARCH_PERFMON_EVENTSEL_USR
))
481 debugctlmsr
|= DEBUGCTLMSR_BTS_OFF_USR
;
483 update_debugctlmsr(debugctlmsr
);
486 void intel_pmu_disable_bts(void)
488 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
489 unsigned long debugctlmsr
;
494 debugctlmsr
= get_debugctlmsr();
497 ~(DEBUGCTLMSR_TR
| DEBUGCTLMSR_BTS
| DEBUGCTLMSR_BTINT
|
498 DEBUGCTLMSR_BTS_OFF_OS
| DEBUGCTLMSR_BTS_OFF_USR
);
500 update_debugctlmsr(debugctlmsr
);
503 int intel_pmu_drain_bts_buffer(void)
505 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
506 struct debug_store
*ds
= cpuc
->ds
;
512 struct perf_event
*event
= cpuc
->events
[INTEL_PMC_IDX_FIXED_BTS
];
513 struct bts_record
*at
, *top
;
514 struct perf_output_handle handle
;
515 struct perf_event_header header
;
516 struct perf_sample_data data
;
522 if (!x86_pmu
.bts_active
)
525 at
= (struct bts_record
*)(unsigned long)ds
->bts_buffer_base
;
526 top
= (struct bts_record
*)(unsigned long)ds
->bts_index
;
531 memset(®s
, 0, sizeof(regs
));
533 ds
->bts_index
= ds
->bts_buffer_base
;
535 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
538 * Prepare a generic sample, i.e. fill in the invariant fields.
539 * We will overwrite the from and to address before we output
542 perf_prepare_sample(&header
, &data
, event
, ®s
);
544 if (perf_output_begin(&handle
, event
, header
.size
* (top
- at
)))
547 for (; at
< top
; at
++) {
551 perf_output_sample(&handle
, &header
, &data
, event
);
554 perf_output_end(&handle
);
556 /* There's new data available. */
557 event
->hw
.interrupts
++;
558 event
->pending_kill
= POLL_IN
;
562 static inline void intel_pmu_drain_pebs_buffer(void)
566 x86_pmu
.drain_pebs(®s
);
569 void intel_pmu_pebs_sched_task(struct perf_event_context
*ctx
, bool sched_in
)
572 intel_pmu_drain_pebs_buffer();
578 struct event_constraint intel_core2_pebs_event_constraints
[] = {
579 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
580 INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
581 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
582 INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
583 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
584 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
585 INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
589 struct event_constraint intel_atom_pebs_event_constraints
[] = {
590 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
591 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
592 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
593 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
594 INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
598 struct event_constraint intel_slm_pebs_event_constraints
[] = {
599 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
600 INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
601 /* Allow all events as PEBS with no flags */
602 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
606 struct event_constraint intel_nehalem_pebs_event_constraints
[] = {
607 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
608 INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
609 INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
610 INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
611 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
612 INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
613 INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
614 INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
615 INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
616 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
617 INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
618 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
619 INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
623 struct event_constraint intel_westmere_pebs_event_constraints
[] = {
624 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
625 INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
626 INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
627 INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
628 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
629 INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
630 INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
631 INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
632 INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
633 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
634 INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
635 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
636 INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
640 struct event_constraint intel_snb_pebs_event_constraints
[] = {
641 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
642 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
643 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
644 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
645 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
646 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
647 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
648 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
649 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
650 /* Allow all events as PEBS with no flags */
651 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
655 struct event_constraint intel_ivb_pebs_event_constraints
[] = {
656 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
657 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
658 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
659 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
660 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
661 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
662 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
663 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
664 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
665 /* Allow all events as PEBS with no flags */
666 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
670 struct event_constraint intel_hsw_pebs_event_constraints
[] = {
671 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
672 INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
673 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
674 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
675 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
676 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
677 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
678 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
679 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
680 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
681 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
682 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
683 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
684 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
685 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
686 /* Allow all events as PEBS with no flags */
687 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
691 struct event_constraint
*intel_pebs_constraints(struct perf_event
*event
)
693 struct event_constraint
*c
;
695 if (!event
->attr
.precise_ip
)
698 if (x86_pmu
.pebs_constraints
) {
699 for_each_event_constraint(c
, x86_pmu
.pebs_constraints
) {
700 if ((event
->hw
.config
& c
->cmask
) == c
->code
) {
701 event
->hw
.flags
|= c
->flags
;
707 return &emptyconstraint
;
710 static inline bool pebs_is_enabled(struct cpu_hw_events
*cpuc
)
712 return (cpuc
->pebs_enabled
& ((1ULL << MAX_PEBS_EVENTS
) - 1));
715 void intel_pmu_pebs_enable(struct perf_event
*event
)
717 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
718 struct hw_perf_event
*hwc
= &event
->hw
;
719 struct debug_store
*ds
= cpuc
->ds
;
723 hwc
->config
&= ~ARCH_PERFMON_EVENTSEL_INT
;
725 first_pebs
= !pebs_is_enabled(cpuc
);
726 cpuc
->pebs_enabled
|= 1ULL << hwc
->idx
;
728 if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_LDLAT
)
729 cpuc
->pebs_enabled
|= 1ULL << (hwc
->idx
+ 32);
730 else if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_ST
)
731 cpuc
->pebs_enabled
|= 1ULL << 63;
734 * When the event is constrained enough we can use a larger
735 * threshold and run the event with less frequent PMI.
737 if (hwc
->flags
& PERF_X86_EVENT_FREERUNNING
) {
738 threshold
= ds
->pebs_absolute_maximum
-
739 x86_pmu
.max_pebs_events
* x86_pmu
.pebs_record_size
;
742 perf_sched_cb_inc(event
->ctx
->pmu
);
744 threshold
= ds
->pebs_buffer_base
+ x86_pmu
.pebs_record_size
;
747 * If not all events can use larger buffer,
748 * roll back to threshold = 1
751 (ds
->pebs_interrupt_threshold
> threshold
))
752 perf_sched_cb_dec(event
->ctx
->pmu
);
755 /* Use auto-reload if possible to save a MSR write in the PMI */
756 if (hwc
->flags
& PERF_X86_EVENT_AUTO_RELOAD
) {
757 ds
->pebs_event_reset
[hwc
->idx
] =
758 (u64
)(-hwc
->sample_period
) & x86_pmu
.cntval_mask
;
761 if (first_pebs
|| ds
->pebs_interrupt_threshold
> threshold
)
762 ds
->pebs_interrupt_threshold
= threshold
;
765 void intel_pmu_pebs_disable(struct perf_event
*event
)
767 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
768 struct hw_perf_event
*hwc
= &event
->hw
;
769 struct debug_store
*ds
= cpuc
->ds
;
771 cpuc
->pebs_enabled
&= ~(1ULL << hwc
->idx
);
773 if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_LDLAT
)
774 cpuc
->pebs_enabled
&= ~(1ULL << (hwc
->idx
+ 32));
775 else if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_ST
)
776 cpuc
->pebs_enabled
&= ~(1ULL << 63);
778 if (ds
->pebs_interrupt_threshold
>
779 ds
->pebs_buffer_base
+ x86_pmu
.pebs_record_size
) {
780 intel_pmu_drain_pebs_buffer();
781 if (!pebs_is_enabled(cpuc
))
782 perf_sched_cb_dec(event
->ctx
->pmu
);
786 wrmsrl(MSR_IA32_PEBS_ENABLE
, cpuc
->pebs_enabled
);
788 hwc
->config
|= ARCH_PERFMON_EVENTSEL_INT
;
791 void intel_pmu_pebs_enable_all(void)
793 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
795 if (cpuc
->pebs_enabled
)
796 wrmsrl(MSR_IA32_PEBS_ENABLE
, cpuc
->pebs_enabled
);
799 void intel_pmu_pebs_disable_all(void)
801 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
803 if (cpuc
->pebs_enabled
)
804 wrmsrl(MSR_IA32_PEBS_ENABLE
, 0);
807 static int intel_pmu_pebs_fixup_ip(struct pt_regs
*regs
)
809 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
810 unsigned long from
= cpuc
->lbr_entries
[0].from
;
811 unsigned long old_to
, to
= cpuc
->lbr_entries
[0].to
;
812 unsigned long ip
= regs
->ip
;
818 * We don't need to fixup if the PEBS assist is fault like
820 if (!x86_pmu
.intel_cap
.pebs_trap
)
824 * No LBR entry, no basic block, no rewinding
826 if (!cpuc
->lbr_stack
.nr
|| !from
|| !to
)
830 * Basic blocks should never cross user/kernel boundaries
832 if (kernel_ip(ip
) != kernel_ip(to
))
836 * unsigned math, either ip is before the start (impossible) or
837 * the basic block is larger than 1 page (sanity)
839 if ((ip
- to
) > PEBS_FIXUP_SIZE
)
843 * We sampled a branch insn, rewind using the LBR stack
846 set_linear_ip(regs
, from
);
851 if (!kernel_ip(ip
)) {
853 u8
*buf
= this_cpu_read(insn_buffer
);
855 /* 'size' must fit our buffer, see above */
856 bytes
= copy_from_user_nmi(buf
, (void __user
*)to
, size
);
871 is_64bit
= kernel_ip(to
) || !test_thread_flag(TIF_IA32
);
873 insn_init(&insn
, kaddr
, size
, is_64bit
);
874 insn_get_length(&insn
);
876 * Make sure there was not a problem decoding the
877 * instruction and getting the length. This is
878 * doubly important because we have an infinite
879 * loop if insn.length=0.
885 kaddr
+= insn
.length
;
890 set_linear_ip(regs
, old_to
);
895 * Even though we decoded the basic block, the instruction stream
896 * never matched the given IP, either the TO or the IP got corrupted.
901 static inline u64
intel_hsw_weight(struct pebs_record_skl
*pebs
)
903 if (pebs
->tsx_tuning
) {
904 union hsw_tsx_tuning tsx
= { .value
= pebs
->tsx_tuning
};
905 return tsx
.cycles_last_block
;
910 static inline u64
intel_hsw_transaction(struct pebs_record_skl
*pebs
)
912 u64 txn
= (pebs
->tsx_tuning
& PEBS_HSW_TSX_FLAGS
) >> 32;
914 /* For RTM XABORTs also log the abort code from AX */
915 if ((txn
& PERF_TXN_TRANSACTION
) && (pebs
->ax
& 1))
916 txn
|= ((pebs
->ax
>> 24) & 0xff) << PERF_TXN_ABORT_SHIFT
;
920 static void setup_pebs_sample_data(struct perf_event
*event
,
921 struct pt_regs
*iregs
, void *__pebs
,
922 struct perf_sample_data
*data
,
923 struct pt_regs
*regs
)
925 #define PERF_X86_EVENT_PEBS_HSW_PREC \
926 (PERF_X86_EVENT_PEBS_ST_HSW | \
927 PERF_X86_EVENT_PEBS_LD_HSW | \
928 PERF_X86_EVENT_PEBS_NA_HSW)
930 * We cast to the biggest pebs_record but are careful not to
931 * unconditionally access the 'extra' entries.
933 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
934 struct pebs_record_skl
*pebs
= __pebs
;
937 int fl
= event
->hw
.flags
;
942 sample_type
= event
->attr
.sample_type
;
943 dsrc
= sample_type
& PERF_SAMPLE_DATA_SRC
;
945 fll
= fl
& PERF_X86_EVENT_PEBS_LDLAT
;
946 fst
= fl
& (PERF_X86_EVENT_PEBS_ST
| PERF_X86_EVENT_PEBS_HSW_PREC
);
948 perf_sample_data_init(data
, 0, event
->hw
.last_period
);
950 data
->period
= event
->hw
.last_period
;
953 * Use latency for weight (only avail with PEBS-LL)
955 if (fll
&& (sample_type
& PERF_SAMPLE_WEIGHT
))
956 data
->weight
= pebs
->lat
;
959 * data.data_src encodes the data source
962 u64 val
= PERF_MEM_NA
;
964 val
= load_latency_data(pebs
->dse
);
965 else if (fst
&& (fl
& PERF_X86_EVENT_PEBS_HSW_PREC
))
966 val
= precise_datala_hsw(event
, pebs
->dse
);
968 val
= precise_store_data(pebs
->dse
);
969 data
->data_src
.val
= val
;
973 * We use the interrupt regs as a base because the PEBS record
974 * does not contain a full regs set, specifically it seems to
975 * lack segment descriptors, which get used by things like
978 * In the simple case fix up only the IP and BP,SP regs, for
979 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
980 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
983 regs
->flags
= pebs
->flags
;
984 set_linear_ip(regs
, pebs
->ip
);
988 if (sample_type
& PERF_SAMPLE_REGS_INTR
) {
998 regs
->flags
= pebs
->flags
;
999 #ifndef CONFIG_X86_32
1000 regs
->r8
= pebs
->r8
;
1001 regs
->r9
= pebs
->r9
;
1002 regs
->r10
= pebs
->r10
;
1003 regs
->r11
= pebs
->r11
;
1004 regs
->r12
= pebs
->r12
;
1005 regs
->r13
= pebs
->r13
;
1006 regs
->r14
= pebs
->r14
;
1007 regs
->r15
= pebs
->r15
;
1011 if (event
->attr
.precise_ip
> 1 && x86_pmu
.intel_cap
.pebs_format
>= 2) {
1012 regs
->ip
= pebs
->real_ip
;
1013 regs
->flags
|= PERF_EFLAGS_EXACT
;
1014 } else if (event
->attr
.precise_ip
> 1 && intel_pmu_pebs_fixup_ip(regs
))
1015 regs
->flags
|= PERF_EFLAGS_EXACT
;
1017 regs
->flags
&= ~PERF_EFLAGS_EXACT
;
1019 if ((sample_type
& PERF_SAMPLE_ADDR
) &&
1020 x86_pmu
.intel_cap
.pebs_format
>= 1)
1021 data
->addr
= pebs
->dla
;
1023 if (x86_pmu
.intel_cap
.pebs_format
>= 2) {
1024 /* Only set the TSX weight when no memory weight. */
1025 if ((sample_type
& PERF_SAMPLE_WEIGHT
) && !fll
)
1026 data
->weight
= intel_hsw_weight(pebs
);
1028 if (sample_type
& PERF_SAMPLE_TRANSACTION
)
1029 data
->txn
= intel_hsw_transaction(pebs
);
1033 * v3 supplies an accurate time stamp, so we use that
1034 * for the time stamp.
1036 * We can only do this for the default trace clock.
1038 if (x86_pmu
.intel_cap
.pebs_format
>= 3 &&
1039 event
->attr
.use_clockid
== 0)
1040 data
->time
= native_sched_clock_from_tsc(pebs
->tsc
);
1042 if (has_branch_stack(event
))
1043 data
->br_stack
= &cpuc
->lbr_stack
;
1046 static inline void *
1047 get_next_pebs_record_by_bit(void *base
, void *top
, int bit
)
1049 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1056 for (at
= base
; at
< top
; at
+= x86_pmu
.pebs_record_size
) {
1057 struct pebs_record_nhm
*p
= at
;
1059 if (test_bit(bit
, (unsigned long *)&p
->status
)) {
1060 /* PEBS v3 has accurate status bits */
1061 if (x86_pmu
.intel_cap
.pebs_format
>= 3)
1064 if (p
->status
== (1 << bit
))
1067 /* clear non-PEBS bit and re-check */
1068 pebs_status
= p
->status
& cpuc
->pebs_enabled
;
1069 pebs_status
&= (1ULL << MAX_PEBS_EVENTS
) - 1;
1070 if (pebs_status
== (1 << bit
))
1077 static void __intel_pmu_pebs_event(struct perf_event
*event
,
1078 struct pt_regs
*iregs
,
1079 void *base
, void *top
,
1082 struct perf_sample_data data
;
1083 struct pt_regs regs
;
1084 void *at
= get_next_pebs_record_by_bit(base
, top
, bit
);
1086 if (!intel_pmu_save_and_restart(event
) &&
1087 !(event
->hw
.flags
& PERF_X86_EVENT_AUTO_RELOAD
))
1091 setup_pebs_sample_data(event
, iregs
, at
, &data
, ®s
);
1092 perf_event_output(event
, &data
, ®s
);
1093 at
+= x86_pmu
.pebs_record_size
;
1094 at
= get_next_pebs_record_by_bit(at
, top
, bit
);
1098 setup_pebs_sample_data(event
, iregs
, at
, &data
, ®s
);
1101 * All but the last records are processed.
1102 * The last one is left to be able to call the overflow handler.
1104 if (perf_event_overflow(event
, &data
, ®s
)) {
1105 x86_pmu_stop(event
, 0);
1111 static void intel_pmu_drain_pebs_core(struct pt_regs
*iregs
)
1113 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1114 struct debug_store
*ds
= cpuc
->ds
;
1115 struct perf_event
*event
= cpuc
->events
[0]; /* PMC0 only */
1116 struct pebs_record_core
*at
, *top
;
1119 if (!x86_pmu
.pebs_active
)
1122 at
= (struct pebs_record_core
*)(unsigned long)ds
->pebs_buffer_base
;
1123 top
= (struct pebs_record_core
*)(unsigned long)ds
->pebs_index
;
1126 * Whatever else happens, drain the thing
1128 ds
->pebs_index
= ds
->pebs_buffer_base
;
1130 if (!test_bit(0, cpuc
->active_mask
))
1133 WARN_ON_ONCE(!event
);
1135 if (!event
->attr
.precise_ip
)
1138 n
= (top
- at
) / x86_pmu
.pebs_record_size
;
1142 __intel_pmu_pebs_event(event
, iregs
, at
, top
, 0, n
);
1145 static void intel_pmu_drain_pebs_nhm(struct pt_regs
*iregs
)
1147 struct cpu_hw_events
*cpuc
= this_cpu_ptr(&cpu_hw_events
);
1148 struct debug_store
*ds
= cpuc
->ds
;
1149 struct perf_event
*event
;
1150 void *base
, *at
, *top
;
1151 short counts
[MAX_PEBS_EVENTS
] = {};
1152 short error
[MAX_PEBS_EVENTS
] = {};
1155 if (!x86_pmu
.pebs_active
)
1158 base
= (struct pebs_record_nhm
*)(unsigned long)ds
->pebs_buffer_base
;
1159 top
= (struct pebs_record_nhm
*)(unsigned long)ds
->pebs_index
;
1161 ds
->pebs_index
= ds
->pebs_buffer_base
;
1163 if (unlikely(base
>= top
))
1166 for (at
= base
; at
< top
; at
+= x86_pmu
.pebs_record_size
) {
1167 struct pebs_record_nhm
*p
= at
;
1169 /* PEBS v3 has accurate status bits */
1170 if (x86_pmu
.intel_cap
.pebs_format
>= 3) {
1171 for_each_set_bit(bit
, (unsigned long *)&p
->status
,
1178 bit
= find_first_bit((unsigned long *)&p
->status
,
1179 x86_pmu
.max_pebs_events
);
1180 if (bit
>= x86_pmu
.max_pebs_events
)
1182 if (!test_bit(bit
, cpuc
->active_mask
))
1185 * The PEBS hardware does not deal well with the situation
1186 * when events happen near to each other and multiple bits
1187 * are set. But it should happen rarely.
1189 * If these events include one PEBS and multiple non-PEBS
1190 * events, it doesn't impact PEBS record. The record will
1191 * be handled normally. (slow path)
1193 * If these events include two or more PEBS events, the
1194 * records for the events can be collapsed into a single
1195 * one, and it's not possible to reconstruct all events
1196 * that caused the PEBS record. It's called collision.
1197 * If collision happened, the record will be dropped.
1200 if (p
->status
!= (1 << bit
)) {
1204 pebs_status
= p
->status
& cpuc
->pebs_enabled
;
1205 pebs_status
&= (1ULL << MAX_PEBS_EVENTS
) - 1;
1206 if (pebs_status
!= (1 << bit
)) {
1207 for_each_set_bit(i
, (unsigned long *)&pebs_status
,
1216 for (bit
= 0; bit
< x86_pmu
.max_pebs_events
; bit
++) {
1217 if ((counts
[bit
] == 0) && (error
[bit
] == 0))
1219 event
= cpuc
->events
[bit
];
1220 WARN_ON_ONCE(!event
);
1221 WARN_ON_ONCE(!event
->attr
.precise_ip
);
1223 /* log dropped samples number */
1225 perf_log_lost_samples(event
, error
[bit
]);
1228 __intel_pmu_pebs_event(event
, iregs
, base
,
1229 top
, bit
, counts
[bit
]);
1235 * BTS, PEBS probe and setup
1238 void __init
intel_ds_init(void)
1241 * No support for 32bit formats
1243 if (!boot_cpu_has(X86_FEATURE_DTES64
))
1246 x86_pmu
.bts
= boot_cpu_has(X86_FEATURE_BTS
);
1247 x86_pmu
.pebs
= boot_cpu_has(X86_FEATURE_PEBS
);
1249 char pebs_type
= x86_pmu
.intel_cap
.pebs_trap
? '+' : '-';
1250 int format
= x86_pmu
.intel_cap
.pebs_format
;
1254 printk(KERN_CONT
"PEBS fmt0%c, ", pebs_type
);
1255 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_core
);
1256 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_core
;
1260 printk(KERN_CONT
"PEBS fmt1%c, ", pebs_type
);
1261 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_nhm
);
1262 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_nhm
;
1266 pr_cont("PEBS fmt2%c, ", pebs_type
);
1267 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_hsw
);
1268 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_nhm
;
1272 pr_cont("PEBS fmt3%c, ", pebs_type
);
1273 x86_pmu
.pebs_record_size
=
1274 sizeof(struct pebs_record_skl
);
1275 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_nhm
;
1279 printk(KERN_CONT
"no PEBS fmt%d%c, ", format
, pebs_type
);
1285 void perf_restore_debug_store(void)
1287 struct debug_store
*ds
= __this_cpu_read(cpu_hw_events
.ds
);
1289 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
1292 wrmsrl(MSR_IA32_DS_AREA
, (unsigned long)ds
);