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
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_store_data_hsw(struct perf_event
*event
, u64 status
)
113 union perf_mem_data_src dse
;
114 u64 cfg
= event
->hw
.config
& INTEL_ARCH_EVENT_MASK
;
117 dse
.mem_op
= PERF_MEM_OP_NA
;
118 dse
.mem_lvl
= PERF_MEM_LVL_NA
;
121 * L1 info only valid for following events:
123 * MEM_UOPS_RETIRED.STLB_MISS_STORES
124 * MEM_UOPS_RETIRED.LOCK_STORES
125 * MEM_UOPS_RETIRED.SPLIT_STORES
126 * MEM_UOPS_RETIRED.ALL_STORES
128 if (cfg
!= 0x12d0 && cfg
!= 0x22d0 && cfg
!= 0x42d0 && cfg
!= 0x82d0)
132 dse
.mem_lvl
= PERF_MEM_LVL_L1
| PERF_MEM_LVL_HIT
;
134 dse
.mem_lvl
= PERF_MEM_LVL_L1
| PERF_MEM_LVL_MISS
;
136 /* Nothing else supported. Sorry. */
140 static u64
load_latency_data(u64 status
)
142 union intel_x86_pebs_dse dse
;
144 int model
= boot_cpu_data
.x86_model
;
145 int fam
= boot_cpu_data
.x86
;
150 * use the mapping table for bit 0-3
152 val
= pebs_data_source
[dse
.ld_dse
];
155 * Nehalem models do not support TLB, Lock infos
157 if (fam
== 0x6 && (model
== 26 || model
== 30
158 || model
== 31 || model
== 46)) {
159 val
|= P(TLB
, NA
) | P(LOCK
, NA
);
164 * 0 = did not miss 2nd level TLB
165 * 1 = missed 2nd level TLB
167 if (dse
.ld_stlb_miss
)
168 val
|= P(TLB
, MISS
) | P(TLB
, L2
);
170 val
|= P(TLB
, HIT
) | P(TLB
, L1
) | P(TLB
, L2
);
173 * bit 5: locked prefix
176 val
|= P(LOCK
, LOCKED
);
181 struct pebs_record_core
{
185 u64 r8
, r9
, r10
, r11
;
186 u64 r12
, r13
, r14
, r15
;
189 struct pebs_record_nhm
{
193 u64 r8
, r9
, r10
, r11
;
194 u64 r12
, r13
, r14
, r15
;
195 u64 status
, dla
, dse
, lat
;
199 * Same as pebs_record_nhm, with two additional fields.
201 struct pebs_record_hsw
{
205 u64 r8
, r9
, r10
, r11
;
206 u64 r12
, r13
, r14
, r15
;
207 u64 status
, dla
, dse
, lat
;
208 u64 real_ip
, tsx_tuning
;
211 union hsw_tsx_tuning
{
213 u32 cycles_last_block
: 32,
216 instruction_abort
: 1,
217 non_instruction_abort
: 1,
226 #define PEBS_HSW_TSX_FLAGS 0xff00000000ULL
228 void init_debug_store_on_cpu(int cpu
)
230 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
235 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
,
236 (u32
)((u64
)(unsigned long)ds
),
237 (u32
)((u64
)(unsigned long)ds
>> 32));
240 void fini_debug_store_on_cpu(int cpu
)
242 if (!per_cpu(cpu_hw_events
, cpu
).ds
)
245 wrmsr_on_cpu(cpu
, MSR_IA32_DS_AREA
, 0, 0);
248 static DEFINE_PER_CPU(void *, insn_buffer
);
250 static int alloc_pebs_buffer(int cpu
)
252 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
253 int node
= cpu_to_node(cpu
);
254 int max
, thresh
= 1; /* always use a single PEBS record */
255 void *buffer
, *ibuffer
;
260 buffer
= kzalloc_node(PEBS_BUFFER_SIZE
, GFP_KERNEL
, node
);
261 if (unlikely(!buffer
))
265 * HSW+ already provides us the eventing ip; no need to allocate this
268 if (x86_pmu
.intel_cap
.pebs_format
< 2) {
269 ibuffer
= kzalloc_node(PEBS_FIXUP_SIZE
, GFP_KERNEL
, node
);
274 per_cpu(insn_buffer
, cpu
) = ibuffer
;
277 max
= PEBS_BUFFER_SIZE
/ x86_pmu
.pebs_record_size
;
279 ds
->pebs_buffer_base
= (u64
)(unsigned long)buffer
;
280 ds
->pebs_index
= ds
->pebs_buffer_base
;
281 ds
->pebs_absolute_maximum
= ds
->pebs_buffer_base
+
282 max
* x86_pmu
.pebs_record_size
;
284 ds
->pebs_interrupt_threshold
= ds
->pebs_buffer_base
+
285 thresh
* x86_pmu
.pebs_record_size
;
290 static void release_pebs_buffer(int cpu
)
292 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
294 if (!ds
|| !x86_pmu
.pebs
)
297 kfree(per_cpu(insn_buffer
, cpu
));
298 per_cpu(insn_buffer
, cpu
) = NULL
;
300 kfree((void *)(unsigned long)ds
->pebs_buffer_base
);
301 ds
->pebs_buffer_base
= 0;
304 static int alloc_bts_buffer(int cpu
)
306 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
307 int node
= cpu_to_node(cpu
);
314 buffer
= kzalloc_node(BTS_BUFFER_SIZE
, GFP_KERNEL
| __GFP_NOWARN
, node
);
315 if (unlikely(!buffer
)) {
316 WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__
);
320 max
= BTS_BUFFER_SIZE
/ BTS_RECORD_SIZE
;
323 ds
->bts_buffer_base
= (u64
)(unsigned long)buffer
;
324 ds
->bts_index
= ds
->bts_buffer_base
;
325 ds
->bts_absolute_maximum
= ds
->bts_buffer_base
+
326 max
* BTS_RECORD_SIZE
;
327 ds
->bts_interrupt_threshold
= ds
->bts_absolute_maximum
-
328 thresh
* BTS_RECORD_SIZE
;
333 static void release_bts_buffer(int cpu
)
335 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
337 if (!ds
|| !x86_pmu
.bts
)
340 kfree((void *)(unsigned long)ds
->bts_buffer_base
);
341 ds
->bts_buffer_base
= 0;
344 static int alloc_ds_buffer(int cpu
)
346 int node
= cpu_to_node(cpu
);
347 struct debug_store
*ds
;
349 ds
= kzalloc_node(sizeof(*ds
), GFP_KERNEL
, node
);
353 per_cpu(cpu_hw_events
, cpu
).ds
= ds
;
358 static void release_ds_buffer(int cpu
)
360 struct debug_store
*ds
= per_cpu(cpu_hw_events
, cpu
).ds
;
365 per_cpu(cpu_hw_events
, cpu
).ds
= NULL
;
369 void release_ds_buffers(void)
373 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
377 for_each_online_cpu(cpu
)
378 fini_debug_store_on_cpu(cpu
);
380 for_each_possible_cpu(cpu
) {
381 release_pebs_buffer(cpu
);
382 release_bts_buffer(cpu
);
383 release_ds_buffer(cpu
);
388 void reserve_ds_buffers(void)
390 int bts_err
= 0, pebs_err
= 0;
393 x86_pmu
.bts_active
= 0;
394 x86_pmu
.pebs_active
= 0;
396 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
407 for_each_possible_cpu(cpu
) {
408 if (alloc_ds_buffer(cpu
)) {
413 if (!bts_err
&& alloc_bts_buffer(cpu
))
416 if (!pebs_err
&& alloc_pebs_buffer(cpu
))
419 if (bts_err
&& pebs_err
)
424 for_each_possible_cpu(cpu
)
425 release_bts_buffer(cpu
);
429 for_each_possible_cpu(cpu
)
430 release_pebs_buffer(cpu
);
433 if (bts_err
&& pebs_err
) {
434 for_each_possible_cpu(cpu
)
435 release_ds_buffer(cpu
);
437 if (x86_pmu
.bts
&& !bts_err
)
438 x86_pmu
.bts_active
= 1;
440 if (x86_pmu
.pebs
&& !pebs_err
)
441 x86_pmu
.pebs_active
= 1;
443 for_each_online_cpu(cpu
)
444 init_debug_store_on_cpu(cpu
);
454 struct event_constraint bts_constraint
=
455 EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS
, 0);
457 void intel_pmu_enable_bts(u64 config
)
459 unsigned long debugctlmsr
;
461 debugctlmsr
= get_debugctlmsr();
463 debugctlmsr
|= DEBUGCTLMSR_TR
;
464 debugctlmsr
|= DEBUGCTLMSR_BTS
;
465 debugctlmsr
|= DEBUGCTLMSR_BTINT
;
467 if (!(config
& ARCH_PERFMON_EVENTSEL_OS
))
468 debugctlmsr
|= DEBUGCTLMSR_BTS_OFF_OS
;
470 if (!(config
& ARCH_PERFMON_EVENTSEL_USR
))
471 debugctlmsr
|= DEBUGCTLMSR_BTS_OFF_USR
;
473 update_debugctlmsr(debugctlmsr
);
476 void intel_pmu_disable_bts(void)
478 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
479 unsigned long debugctlmsr
;
484 debugctlmsr
= get_debugctlmsr();
487 ~(DEBUGCTLMSR_TR
| DEBUGCTLMSR_BTS
| DEBUGCTLMSR_BTINT
|
488 DEBUGCTLMSR_BTS_OFF_OS
| DEBUGCTLMSR_BTS_OFF_USR
);
490 update_debugctlmsr(debugctlmsr
);
493 int intel_pmu_drain_bts_buffer(void)
495 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
496 struct debug_store
*ds
= cpuc
->ds
;
502 struct perf_event
*event
= cpuc
->events
[INTEL_PMC_IDX_FIXED_BTS
];
503 struct bts_record
*at
, *top
;
504 struct perf_output_handle handle
;
505 struct perf_event_header header
;
506 struct perf_sample_data data
;
512 if (!x86_pmu
.bts_active
)
515 at
= (struct bts_record
*)(unsigned long)ds
->bts_buffer_base
;
516 top
= (struct bts_record
*)(unsigned long)ds
->bts_index
;
521 memset(®s
, 0, sizeof(regs
));
523 ds
->bts_index
= ds
->bts_buffer_base
;
525 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
528 * Prepare a generic sample, i.e. fill in the invariant fields.
529 * We will overwrite the from and to address before we output
532 perf_prepare_sample(&header
, &data
, event
, ®s
);
534 if (perf_output_begin(&handle
, event
, header
.size
* (top
- at
)))
537 for (; at
< top
; at
++) {
541 perf_output_sample(&handle
, &header
, &data
, event
);
544 perf_output_end(&handle
);
546 /* There's new data available. */
547 event
->hw
.interrupts
++;
548 event
->pending_kill
= POLL_IN
;
555 struct event_constraint intel_core2_pebs_event_constraints
[] = {
556 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
557 INTEL_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
558 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
559 INTEL_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
560 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
564 struct event_constraint intel_atom_pebs_event_constraints
[] = {
565 INTEL_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
566 INTEL_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
567 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
571 struct event_constraint intel_slm_pebs_event_constraints
[] = {
572 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
573 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
574 /* Allow all events as PEBS with no flags */
575 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
579 struct event_constraint intel_nehalem_pebs_event_constraints
[] = {
580 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
581 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
582 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
583 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
584 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
585 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
586 INTEL_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
587 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
588 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
589 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
590 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
594 struct event_constraint intel_westmere_pebs_event_constraints
[] = {
595 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
596 INTEL_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
597 INTEL_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
598 INTEL_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
599 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
600 INTEL_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
601 INTEL_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
602 INTEL_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
603 INTEL_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
604 INTEL_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
605 INTEL_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
609 struct event_constraint intel_snb_pebs_event_constraints
[] = {
610 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
611 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
612 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
613 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
614 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
615 /* Allow all events as PEBS with no flags */
616 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
620 struct event_constraint intel_ivb_pebs_event_constraints
[] = {
621 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
622 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
623 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
624 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
625 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
626 /* Allow all events as PEBS with no flags */
627 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
631 struct event_constraint intel_hsw_pebs_event_constraints
[] = {
632 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
633 INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
634 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
635 INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
636 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
637 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
638 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
639 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
640 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
641 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
642 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
643 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
644 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
645 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
646 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
647 /* Allow all events as PEBS with no flags */
648 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
652 struct event_constraint
*intel_pebs_constraints(struct perf_event
*event
)
654 struct event_constraint
*c
;
656 if (!event
->attr
.precise_ip
)
659 if (x86_pmu
.pebs_constraints
) {
660 for_each_event_constraint(c
, x86_pmu
.pebs_constraints
) {
661 if ((event
->hw
.config
& c
->cmask
) == c
->code
) {
662 event
->hw
.flags
|= c
->flags
;
668 return &emptyconstraint
;
671 void intel_pmu_pebs_enable(struct perf_event
*event
)
673 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
674 struct hw_perf_event
*hwc
= &event
->hw
;
676 hwc
->config
&= ~ARCH_PERFMON_EVENTSEL_INT
;
678 cpuc
->pebs_enabled
|= 1ULL << hwc
->idx
;
680 if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_LDLAT
)
681 cpuc
->pebs_enabled
|= 1ULL << (hwc
->idx
+ 32);
682 else if (event
->hw
.flags
& PERF_X86_EVENT_PEBS_ST
)
683 cpuc
->pebs_enabled
|= 1ULL << 63;
686 void intel_pmu_pebs_disable(struct perf_event
*event
)
688 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
689 struct hw_perf_event
*hwc
= &event
->hw
;
691 cpuc
->pebs_enabled
&= ~(1ULL << hwc
->idx
);
693 if (event
->hw
.constraint
->flags
& PERF_X86_EVENT_PEBS_LDLAT
)
694 cpuc
->pebs_enabled
&= ~(1ULL << (hwc
->idx
+ 32));
695 else if (event
->hw
.constraint
->flags
& PERF_X86_EVENT_PEBS_ST
)
696 cpuc
->pebs_enabled
&= ~(1ULL << 63);
699 wrmsrl(MSR_IA32_PEBS_ENABLE
, cpuc
->pebs_enabled
);
701 hwc
->config
|= ARCH_PERFMON_EVENTSEL_INT
;
704 void intel_pmu_pebs_enable_all(void)
706 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
708 if (cpuc
->pebs_enabled
)
709 wrmsrl(MSR_IA32_PEBS_ENABLE
, cpuc
->pebs_enabled
);
712 void intel_pmu_pebs_disable_all(void)
714 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
716 if (cpuc
->pebs_enabled
)
717 wrmsrl(MSR_IA32_PEBS_ENABLE
, 0);
720 static int intel_pmu_pebs_fixup_ip(struct pt_regs
*regs
)
722 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
723 unsigned long from
= cpuc
->lbr_entries
[0].from
;
724 unsigned long old_to
, to
= cpuc
->lbr_entries
[0].to
;
725 unsigned long ip
= regs
->ip
;
730 * We don't need to fixup if the PEBS assist is fault like
732 if (!x86_pmu
.intel_cap
.pebs_trap
)
736 * No LBR entry, no basic block, no rewinding
738 if (!cpuc
->lbr_stack
.nr
|| !from
|| !to
)
742 * Basic blocks should never cross user/kernel boundaries
744 if (kernel_ip(ip
) != kernel_ip(to
))
748 * unsigned math, either ip is before the start (impossible) or
749 * the basic block is larger than 1 page (sanity)
751 if ((ip
- to
) > PEBS_FIXUP_SIZE
)
755 * We sampled a branch insn, rewind using the LBR stack
758 set_linear_ip(regs
, from
);
762 if (!kernel_ip(ip
)) {
764 u8
*buf
= this_cpu_read(insn_buffer
);
766 size
= ip
- to
; /* Must fit our buffer, see above */
767 bytes
= copy_from_user_nmi(buf
, (void __user
*)to
, size
);
782 is_64bit
= kernel_ip(to
) || !test_thread_flag(TIF_IA32
);
784 insn_init(&insn
, kaddr
, is_64bit
);
785 insn_get_length(&insn
);
788 kaddr
+= insn
.length
;
792 set_linear_ip(regs
, old_to
);
797 * Even though we decoded the basic block, the instruction stream
798 * never matched the given IP, either the TO or the IP got corrupted.
803 static inline u64
intel_hsw_weight(struct pebs_record_hsw
*pebs
)
805 if (pebs
->tsx_tuning
) {
806 union hsw_tsx_tuning tsx
= { .value
= pebs
->tsx_tuning
};
807 return tsx
.cycles_last_block
;
812 static inline u64
intel_hsw_transaction(struct pebs_record_hsw
*pebs
)
814 u64 txn
= (pebs
->tsx_tuning
& PEBS_HSW_TSX_FLAGS
) >> 32;
816 /* For RTM XABORTs also log the abort code from AX */
817 if ((txn
& PERF_TXN_TRANSACTION
) && (pebs
->ax
& 1))
818 txn
|= ((pebs
->ax
>> 24) & 0xff) << PERF_TXN_ABORT_SHIFT
;
822 static void __intel_pmu_pebs_event(struct perf_event
*event
,
823 struct pt_regs
*iregs
, void *__pebs
)
826 * We cast to the biggest pebs_record but are careful not to
827 * unconditionally access the 'extra' entries.
829 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
830 struct pebs_record_hsw
*pebs
= __pebs
;
831 struct perf_sample_data data
;
836 if (!intel_pmu_save_and_restart(event
))
839 fll
= event
->hw
.flags
& PERF_X86_EVENT_PEBS_LDLAT
;
840 fst
= event
->hw
.flags
& (PERF_X86_EVENT_PEBS_ST
|
841 PERF_X86_EVENT_PEBS_ST_HSW
|
842 PERF_X86_EVENT_PEBS_LD_HSW
|
843 PERF_X86_EVENT_PEBS_NA_HSW
);
845 perf_sample_data_init(&data
, 0, event
->hw
.last_period
);
847 data
.period
= event
->hw
.last_period
;
848 sample_type
= event
->attr
.sample_type
;
851 * if PEBS-LL or PreciseStore
855 * Use latency for weight (only avail with PEBS-LL)
857 if (fll
&& (sample_type
& PERF_SAMPLE_WEIGHT
))
858 data
.weight
= pebs
->lat
;
861 * data.data_src encodes the data source
863 if (sample_type
& PERF_SAMPLE_DATA_SRC
) {
865 data
.data_src
.val
= load_latency_data(pebs
->dse
);
866 else if (event
->hw
.flags
&
867 (PERF_X86_EVENT_PEBS_ST_HSW
|
868 PERF_X86_EVENT_PEBS_LD_HSW
|
869 PERF_X86_EVENT_PEBS_NA_HSW
))
871 precise_store_data_hsw(event
, pebs
->dse
);
873 data
.data_src
.val
= precise_store_data(pebs
->dse
);
878 * We use the interrupt regs as a base because the PEBS record
879 * does not contain a full regs set, specifically it seems to
880 * lack segment descriptors, which get used by things like
883 * In the simple case fix up only the IP and BP,SP regs, for
884 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
885 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
888 regs
.flags
= pebs
->flags
;
889 set_linear_ip(®s
, pebs
->ip
);
893 if (event
->attr
.precise_ip
> 1 && x86_pmu
.intel_cap
.pebs_format
>= 2) {
894 regs
.ip
= pebs
->real_ip
;
895 regs
.flags
|= PERF_EFLAGS_EXACT
;
896 } else if (event
->attr
.precise_ip
> 1 && intel_pmu_pebs_fixup_ip(®s
))
897 regs
.flags
|= PERF_EFLAGS_EXACT
;
899 regs
.flags
&= ~PERF_EFLAGS_EXACT
;
901 if ((event
->attr
.sample_type
& PERF_SAMPLE_ADDR
) &&
902 x86_pmu
.intel_cap
.pebs_format
>= 1)
903 data
.addr
= pebs
->dla
;
905 if (x86_pmu
.intel_cap
.pebs_format
>= 2) {
906 /* Only set the TSX weight when no memory weight. */
907 if ((event
->attr
.sample_type
& PERF_SAMPLE_WEIGHT
) && !fll
)
908 data
.weight
= intel_hsw_weight(pebs
);
910 if (event
->attr
.sample_type
& PERF_SAMPLE_TRANSACTION
)
911 data
.txn
= intel_hsw_transaction(pebs
);
914 if (has_branch_stack(event
))
915 data
.br_stack
= &cpuc
->lbr_stack
;
917 if (perf_event_overflow(event
, &data
, ®s
))
918 x86_pmu_stop(event
, 0);
921 static void intel_pmu_drain_pebs_core(struct pt_regs
*iregs
)
923 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
924 struct debug_store
*ds
= cpuc
->ds
;
925 struct perf_event
*event
= cpuc
->events
[0]; /* PMC0 only */
926 struct pebs_record_core
*at
, *top
;
929 if (!x86_pmu
.pebs_active
)
932 at
= (struct pebs_record_core
*)(unsigned long)ds
->pebs_buffer_base
;
933 top
= (struct pebs_record_core
*)(unsigned long)ds
->pebs_index
;
936 * Whatever else happens, drain the thing
938 ds
->pebs_index
= ds
->pebs_buffer_base
;
940 if (!test_bit(0, cpuc
->active_mask
))
943 WARN_ON_ONCE(!event
);
945 if (!event
->attr
.precise_ip
)
953 * Should not happen, we program the threshold at 1 and do not
956 WARN_ONCE(n
> 1, "bad leftover pebs %d\n", n
);
959 __intel_pmu_pebs_event(event
, iregs
, at
);
962 static void intel_pmu_drain_pebs_nhm(struct pt_regs
*iregs
)
964 struct cpu_hw_events
*cpuc
= &__get_cpu_var(cpu_hw_events
);
965 struct debug_store
*ds
= cpuc
->ds
;
966 struct perf_event
*event
= NULL
;
971 if (!x86_pmu
.pebs_active
)
974 at
= (struct pebs_record_nhm
*)(unsigned long)ds
->pebs_buffer_base
;
975 top
= (struct pebs_record_nhm
*)(unsigned long)ds
->pebs_index
;
977 ds
->pebs_index
= ds
->pebs_buffer_base
;
979 if (unlikely(at
> top
))
983 * Should not happen, we program the threshold at 1 and do not
986 WARN_ONCE(top
- at
> x86_pmu
.max_pebs_events
* x86_pmu
.pebs_record_size
,
987 "Unexpected number of pebs records %ld\n",
988 (long)(top
- at
) / x86_pmu
.pebs_record_size
);
990 for (; at
< top
; at
+= x86_pmu
.pebs_record_size
) {
991 struct pebs_record_nhm
*p
= at
;
993 for_each_set_bit(bit
, (unsigned long *)&p
->status
,
994 x86_pmu
.max_pebs_events
) {
995 event
= cpuc
->events
[bit
];
996 if (!test_bit(bit
, cpuc
->active_mask
))
999 WARN_ON_ONCE(!event
);
1001 if (!event
->attr
.precise_ip
)
1004 if (__test_and_set_bit(bit
, (unsigned long *)&status
))
1010 if (!event
|| bit
>= x86_pmu
.max_pebs_events
)
1013 __intel_pmu_pebs_event(event
, iregs
, at
);
1018 * BTS, PEBS probe and setup
1021 void intel_ds_init(void)
1024 * No support for 32bit formats
1026 if (!boot_cpu_has(X86_FEATURE_DTES64
))
1029 x86_pmu
.bts
= boot_cpu_has(X86_FEATURE_BTS
);
1030 x86_pmu
.pebs
= boot_cpu_has(X86_FEATURE_PEBS
);
1032 char pebs_type
= x86_pmu
.intel_cap
.pebs_trap
? '+' : '-';
1033 int format
= x86_pmu
.intel_cap
.pebs_format
;
1037 printk(KERN_CONT
"PEBS fmt0%c, ", pebs_type
);
1038 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_core
);
1039 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_core
;
1043 printk(KERN_CONT
"PEBS fmt1%c, ", pebs_type
);
1044 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_nhm
);
1045 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_nhm
;
1049 pr_cont("PEBS fmt2%c, ", pebs_type
);
1050 x86_pmu
.pebs_record_size
= sizeof(struct pebs_record_hsw
);
1051 x86_pmu
.drain_pebs
= intel_pmu_drain_pebs_nhm
;
1055 printk(KERN_CONT
"no PEBS fmt%d%c, ", format
, pebs_type
);
1061 void perf_restore_debug_store(void)
1063 struct debug_store
*ds
= __this_cpu_read(cpu_hw_events
.ds
);
1065 if (!x86_pmu
.bts
&& !x86_pmu
.pebs
)
1068 wrmsrl(MSR_IA32_DS_AREA
, (unsigned long)ds
);