1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/types.h>
4 #include <linux/slab.h>
6 #include <asm/cpu_entry_area.h>
7 #include <asm/perf_event.h>
8 #include <asm/tlbflush.h>
12 #include "../perf_event.h"
14 /* Waste a full page so it can be mapped into the cpu_entry_area */
15 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
17 /* The size of a BTS record in bytes: */
18 #define BTS_RECORD_SIZE 24
20 #define PEBS_FIXUP_SIZE PAGE_SIZE
23 * pebs_record_32 for p4 and core not supported
25 struct pebs_record_32 {
33 union intel_x86_pebs_dse {
36 unsigned int ld_dse:4;
37 unsigned int ld_stlb_miss:1;
38 unsigned int ld_locked:1;
39 unsigned int ld_data_blk:1;
40 unsigned int ld_addr_blk:1;
41 unsigned int ld_reserved:24;
44 unsigned int st_l1d_hit:1;
45 unsigned int st_reserved1:3;
46 unsigned int st_stlb_miss:1;
47 unsigned int st_locked:1;
48 unsigned int st_reserved2:26;
51 unsigned int st_lat_dse:4;
52 unsigned int st_lat_stlb_miss:1;
53 unsigned int st_lat_locked:1;
54 unsigned int ld_reserved3:26;
60 * Map PEBS Load Latency Data Source encodings to generic
61 * memory data source information
63 #define P(a, b) PERF_MEM_S(a, b)
64 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
65 #define LEVEL(x) P(LVLNUM, x)
66 #define REM P(REMOTE, REMOTE)
67 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
69 /* Version for Sandy Bridge and later */
70 static u64 pebs_data_source[] = {
71 P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
72 OP_LH | P(LVL, L1) | LEVEL(L1) | P(SNOOP, NONE), /* 0x01: L1 local */
73 OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
74 OP_LH | P(LVL, L2) | LEVEL(L2) | P(SNOOP, NONE), /* 0x03: L2 hit */
75 OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, NONE), /* 0x04: L3 hit */
76 OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, MISS), /* 0x05: L3 hit, snoop miss */
77 OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT), /* 0x06: L3 hit, snoop hit */
78 OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM), /* 0x07: L3 hit, snoop hitm */
79 OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x08: L3 miss snoop hit */
80 OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
81 OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | P(SNOOP, HIT), /* 0x0a: L3 miss, shared */
82 OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT), /* 0x0b: L3 miss, shared */
83 OP_LH | P(LVL, LOC_RAM) | LEVEL(RAM) | SNOOP_NONE_MISS, /* 0x0c: L3 miss, excl */
84 OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
85 OP_LH | P(LVL, IO) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
86 OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
89 /* Patch up minor differences in the bits */
90 void __init intel_pmu_pebs_data_source_nhm(void)
92 pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
93 pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
94 pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
97 void __init intel_pmu_pebs_data_source_skl(bool pmem)
99 u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
101 pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
102 pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
103 pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
104 pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
105 pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
108 static u64 precise_store_data(u64 status)
110 union intel_x86_pebs_dse dse;
111 u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
117 * 1 = stored missed 2nd level TLB
119 * so it either hit the walker or the OS
120 * otherwise hit 2nd level TLB
122 if (dse.st_stlb_miss)
128 * bit 0: hit L1 data cache
129 * if not set, then all we know is that
138 * bit 5: Locked prefix
141 val |= P(LOCK, LOCKED);
146 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
148 union perf_mem_data_src dse;
150 dse.val = PERF_MEM_NA;
152 if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
153 dse.mem_op = PERF_MEM_OP_STORE;
154 else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
155 dse.mem_op = PERF_MEM_OP_LOAD;
158 * L1 info only valid for following events:
160 * MEM_UOPS_RETIRED.STLB_MISS_STORES
161 * MEM_UOPS_RETIRED.LOCK_STORES
162 * MEM_UOPS_RETIRED.SPLIT_STORES
163 * MEM_UOPS_RETIRED.ALL_STORES
165 if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
167 dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
169 dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
174 static u64 load_latency_data(u64 status)
176 union intel_x86_pebs_dse dse;
182 * use the mapping table for bit 0-3
184 val = pebs_data_source[dse.ld_dse];
187 * Nehalem models do not support TLB, Lock infos
189 if (x86_pmu.pebs_no_tlb) {
190 val |= P(TLB, NA) | P(LOCK, NA);
195 * 0 = did not miss 2nd level TLB
196 * 1 = missed 2nd level TLB
198 if (dse.ld_stlb_miss)
199 val |= P(TLB, MISS) | P(TLB, L2);
201 val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
204 * bit 5: locked prefix
207 val |= P(LOCK, LOCKED);
210 * Ice Lake and earlier models do not support block infos.
212 if (!x86_pmu.pebs_block) {
217 * bit 6: load was blocked since its data could not be forwarded
218 * from a preceding store
224 * bit 7: load was blocked due to potential address conflict with
230 if (!dse.ld_data_blk && !dse.ld_addr_blk)
236 static u64 store_latency_data(u64 status)
238 union intel_x86_pebs_dse dse;
244 * use the mapping table for bit 0-3
246 val = pebs_data_source[dse.st_lat_dse];
250 * 0 = did not miss 2nd level TLB
251 * 1 = missed 2nd level TLB
253 if (dse.st_lat_stlb_miss)
254 val |= P(TLB, MISS) | P(TLB, L2);
256 val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
259 * bit 5: locked prefix
261 if (dse.st_lat_locked)
262 val |= P(LOCK, LOCKED);
269 struct pebs_record_core {
273 u64 r8, r9, r10, r11;
274 u64 r12, r13, r14, r15;
277 struct pebs_record_nhm {
281 u64 r8, r9, r10, r11;
282 u64 r12, r13, r14, r15;
283 u64 status, dla, dse, lat;
287 * Same as pebs_record_nhm, with two additional fields.
289 struct pebs_record_hsw {
293 u64 r8, r9, r10, r11;
294 u64 r12, r13, r14, r15;
295 u64 status, dla, dse, lat;
296 u64 real_ip, tsx_tuning;
299 union hsw_tsx_tuning {
301 u32 cycles_last_block : 32,
304 instruction_abort : 1,
305 non_instruction_abort : 1,
314 #define PEBS_HSW_TSX_FLAGS 0xff00000000ULL
316 /* Same as HSW, plus TSC */
318 struct pebs_record_skl {
322 u64 r8, r9, r10, r11;
323 u64 r12, r13, r14, r15;
324 u64 status, dla, dse, lat;
325 u64 real_ip, tsx_tuning;
329 void init_debug_store_on_cpu(int cpu)
331 struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
336 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
337 (u32)((u64)(unsigned long)ds),
338 (u32)((u64)(unsigned long)ds >> 32));
341 void fini_debug_store_on_cpu(int cpu)
343 if (!per_cpu(cpu_hw_events, cpu).ds)
346 wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
349 static DEFINE_PER_CPU(void *, insn_buffer);
351 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
353 unsigned long start = (unsigned long)cea;
357 pa = virt_to_phys(addr);
360 for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
361 cea_set_pte(cea, pa, prot);
364 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
365 * all TLB entries for it.
367 flush_tlb_kernel_range(start, start + size);
371 static void ds_clear_cea(void *cea, size_t size)
373 unsigned long start = (unsigned long)cea;
377 for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
378 cea_set_pte(cea, 0, PAGE_NONE);
380 flush_tlb_kernel_range(start, start + size);
384 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
386 unsigned int order = get_order(size);
387 int node = cpu_to_node(cpu);
390 page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
391 return page ? page_address(page) : NULL;
394 static void dsfree_pages(const void *buffer, size_t size)
397 free_pages((unsigned long)buffer, get_order(size));
400 static int alloc_pebs_buffer(int cpu)
402 struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
403 struct debug_store *ds = hwev->ds;
404 size_t bsiz = x86_pmu.pebs_buffer_size;
405 int max, node = cpu_to_node(cpu);
406 void *buffer, *insn_buff, *cea;
411 buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
412 if (unlikely(!buffer))
416 * HSW+ already provides us the eventing ip; no need to allocate this
419 if (x86_pmu.intel_cap.pebs_format < 2) {
420 insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
422 dsfree_pages(buffer, bsiz);
425 per_cpu(insn_buffer, cpu) = insn_buff;
427 hwev->ds_pebs_vaddr = buffer;
428 /* Update the cpu entry area mapping */
429 cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
430 ds->pebs_buffer_base = (unsigned long) cea;
431 ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
432 ds->pebs_index = ds->pebs_buffer_base;
433 max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
434 ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
438 static void release_pebs_buffer(int cpu)
440 struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
446 kfree(per_cpu(insn_buffer, cpu));
447 per_cpu(insn_buffer, cpu) = NULL;
449 /* Clear the fixmap */
450 cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
451 ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
452 dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
453 hwev->ds_pebs_vaddr = NULL;
456 static int alloc_bts_buffer(int cpu)
458 struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
459 struct debug_store *ds = hwev->ds;
466 buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
467 if (unlikely(!buffer)) {
468 WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
471 hwev->ds_bts_vaddr = buffer;
472 /* Update the fixmap */
473 cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
474 ds->bts_buffer_base = (unsigned long) cea;
475 ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
476 ds->bts_index = ds->bts_buffer_base;
477 max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
478 ds->bts_absolute_maximum = ds->bts_buffer_base +
479 max * BTS_RECORD_SIZE;
480 ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
481 (max / 16) * BTS_RECORD_SIZE;
485 static void release_bts_buffer(int cpu)
487 struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
493 /* Clear the fixmap */
494 cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
495 ds_clear_cea(cea, BTS_BUFFER_SIZE);
496 dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
497 hwev->ds_bts_vaddr = NULL;
500 static int alloc_ds_buffer(int cpu)
502 struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
504 memset(ds, 0, sizeof(*ds));
505 per_cpu(cpu_hw_events, cpu).ds = ds;
509 static void release_ds_buffer(int cpu)
511 per_cpu(cpu_hw_events, cpu).ds = NULL;
514 void release_ds_buffers(void)
518 if (!x86_pmu.bts && !x86_pmu.pebs)
521 for_each_possible_cpu(cpu)
522 release_ds_buffer(cpu);
524 for_each_possible_cpu(cpu) {
526 * Again, ignore errors from offline CPUs, they will no longer
527 * observe cpu_hw_events.ds and not program the DS_AREA when
530 fini_debug_store_on_cpu(cpu);
533 for_each_possible_cpu(cpu) {
534 release_pebs_buffer(cpu);
535 release_bts_buffer(cpu);
539 void reserve_ds_buffers(void)
541 int bts_err = 0, pebs_err = 0;
544 x86_pmu.bts_active = 0;
545 x86_pmu.pebs_active = 0;
547 if (!x86_pmu.bts && !x86_pmu.pebs)
556 for_each_possible_cpu(cpu) {
557 if (alloc_ds_buffer(cpu)) {
562 if (!bts_err && alloc_bts_buffer(cpu))
565 if (!pebs_err && alloc_pebs_buffer(cpu))
568 if (bts_err && pebs_err)
573 for_each_possible_cpu(cpu)
574 release_bts_buffer(cpu);
578 for_each_possible_cpu(cpu)
579 release_pebs_buffer(cpu);
582 if (bts_err && pebs_err) {
583 for_each_possible_cpu(cpu)
584 release_ds_buffer(cpu);
586 if (x86_pmu.bts && !bts_err)
587 x86_pmu.bts_active = 1;
589 if (x86_pmu.pebs && !pebs_err)
590 x86_pmu.pebs_active = 1;
592 for_each_possible_cpu(cpu) {
594 * Ignores wrmsr_on_cpu() errors for offline CPUs they
595 * will get this call through intel_pmu_cpu_starting().
597 init_debug_store_on_cpu(cpu);
606 struct event_constraint bts_constraint =
607 EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
609 void intel_pmu_enable_bts(u64 config)
611 unsigned long debugctlmsr;
613 debugctlmsr = get_debugctlmsr();
615 debugctlmsr |= DEBUGCTLMSR_TR;
616 debugctlmsr |= DEBUGCTLMSR_BTS;
617 if (config & ARCH_PERFMON_EVENTSEL_INT)
618 debugctlmsr |= DEBUGCTLMSR_BTINT;
620 if (!(config & ARCH_PERFMON_EVENTSEL_OS))
621 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
623 if (!(config & ARCH_PERFMON_EVENTSEL_USR))
624 debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
626 update_debugctlmsr(debugctlmsr);
629 void intel_pmu_disable_bts(void)
631 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
632 unsigned long debugctlmsr;
637 debugctlmsr = get_debugctlmsr();
640 ~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
641 DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
643 update_debugctlmsr(debugctlmsr);
646 int intel_pmu_drain_bts_buffer(void)
648 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
649 struct debug_store *ds = cpuc->ds;
655 struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
656 struct bts_record *at, *base, *top;
657 struct perf_output_handle handle;
658 struct perf_event_header header;
659 struct perf_sample_data data;
660 unsigned long skip = 0;
666 if (!x86_pmu.bts_active)
669 base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
670 top = (struct bts_record *)(unsigned long)ds->bts_index;
675 memset(®s, 0, sizeof(regs));
677 ds->bts_index = ds->bts_buffer_base;
679 perf_sample_data_init(&data, 0, event->hw.last_period);
682 * BTS leaks kernel addresses in branches across the cpl boundary,
683 * such as traps or system calls, so unless the user is asking for
684 * kernel tracing (and right now it's not possible), we'd need to
685 * filter them out. But first we need to count how many of those we
686 * have in the current batch. This is an extra O(n) pass, however,
687 * it's much faster than the other one especially considering that
688 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
689 * alloc_bts_buffer()).
691 for (at = base; at < top; at++) {
693 * Note that right now *this* BTS code only works if
694 * attr::exclude_kernel is set, but let's keep this extra
695 * check here in case that changes.
697 if (event->attr.exclude_kernel &&
698 (kernel_ip(at->from) || kernel_ip(at->to)))
703 * Prepare a generic sample, i.e. fill in the invariant fields.
704 * We will overwrite the from and to address before we output
708 perf_prepare_sample(&header, &data, event, ®s);
710 if (perf_output_begin(&handle, &data, event,
711 header.size * (top - base - skip)))
714 for (at = base; at < top; at++) {
715 /* Filter out any records that contain kernel addresses. */
716 if (event->attr.exclude_kernel &&
717 (kernel_ip(at->from) || kernel_ip(at->to)))
723 perf_output_sample(&handle, &header, &data, event);
726 perf_output_end(&handle);
728 /* There's new data available. */
729 event->hw.interrupts++;
730 event->pending_kill = POLL_IN;
736 static inline void intel_pmu_drain_pebs_buffer(void)
738 struct perf_sample_data data;
740 x86_pmu.drain_pebs(NULL, &data);
746 struct event_constraint intel_core2_pebs_event_constraints[] = {
747 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
748 INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
749 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
750 INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
751 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
752 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
753 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
757 struct event_constraint intel_atom_pebs_event_constraints[] = {
758 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
759 INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
760 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED.* */
761 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
762 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
763 /* Allow all events as PEBS with no flags */
764 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
768 struct event_constraint intel_slm_pebs_event_constraints[] = {
769 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
770 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
771 /* Allow all events as PEBS with no flags */
772 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
776 struct event_constraint intel_glm_pebs_event_constraints[] = {
777 /* Allow all events as PEBS with no flags */
778 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
782 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
783 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
784 INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
785 INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
786 INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INST_RETIRED.ANY */
787 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
788 INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
789 INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
790 INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
791 INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
792 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
793 INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
794 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
795 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
799 struct event_constraint intel_westmere_pebs_event_constraints[] = {
800 INTEL_PLD_CONSTRAINT(0x100b, 0xf), /* MEM_INST_RETIRED.* */
801 INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf), /* MEM_UNCORE_RETIRED.* */
802 INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
803 INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf), /* INSTR_RETIRED.* */
804 INTEL_EVENT_CONSTRAINT(0xc2, 0xf), /* UOPS_RETIRED.* */
805 INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf), /* BR_INST_RETIRED.* */
806 INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf), /* BR_MISP_RETIRED.* */
807 INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf), /* SSEX_UOPS_RETIRED.* */
808 INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
809 INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf), /* MEM_LOAD_RETIRED.* */
810 INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf), /* FP_ASSIST.* */
811 /* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
812 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
816 struct event_constraint intel_snb_pebs_event_constraints[] = {
817 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
818 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
819 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
820 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
821 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
822 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
823 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
824 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
825 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
826 /* Allow all events as PEBS with no flags */
827 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
831 struct event_constraint intel_ivb_pebs_event_constraints[] = {
832 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
833 INTEL_PLD_CONSTRAINT(0x01cd, 0x8), /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
834 INTEL_PST_CONSTRAINT(0x02cd, 0x8), /* MEM_TRANS_RETIRED.PRECISE_STORES */
835 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
836 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
837 /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
838 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
839 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOP_RETIRED.* */
840 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
841 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
842 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
843 /* Allow all events as PEBS with no flags */
844 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
848 struct event_constraint intel_hsw_pebs_event_constraints[] = {
849 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
850 INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
851 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
852 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
853 /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
854 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
855 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
856 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
857 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
858 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
859 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
860 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
861 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
862 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
863 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
864 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
865 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
866 /* Allow all events as PEBS with no flags */
867 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
871 struct event_constraint intel_bdw_pebs_event_constraints[] = {
872 INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
873 INTEL_PLD_CONSTRAINT(0x01cd, 0xf), /* MEM_TRANS_RETIRED.* */
874 /* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
875 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
876 /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
877 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
878 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
879 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
880 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
881 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
882 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
883 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
884 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
885 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
886 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */
887 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
888 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
889 /* Allow all events as PEBS with no flags */
890 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
895 struct event_constraint intel_skl_pebs_event_constraints[] = {
896 INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */
897 /* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
898 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
899 /* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
900 INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
901 INTEL_PLD_CONSTRAINT(0x1cd, 0xf), /* MEM_TRANS_RETIRED.* */
902 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
903 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
904 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
905 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
906 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
907 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
908 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
909 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
910 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */
911 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */
912 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf), /* MEM_LOAD_L3_MISS_RETIRED.* */
913 /* Allow all events as PEBS with no flags */
914 INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
918 struct event_constraint intel_icl_pebs_event_constraints[] = {
919 INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x100000000ULL), /* INST_RETIRED.PREC_DIST */
920 INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL), /* SLOTS */
922 INTEL_PLD_CONSTRAINT(0x1cd, 0xff), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
923 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf), /* MEM_INST_RETIRED.LOAD */
924 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf), /* MEM_INST_RETIRED.STORE */
926 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
928 INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */
931 * Everything else is handled by PMU_FL_PEBS_ALL, because we
932 * need the full constraints from the main table.
938 struct event_constraint intel_spr_pebs_event_constraints[] = {
939 INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x100000000ULL),
940 INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
942 INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
943 INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
944 INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
945 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),
946 INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),
948 INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
950 INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
953 * Everything else is handled by PMU_FL_PEBS_ALL, because we
954 * need the full constraints from the main table.
960 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
962 struct event_constraint *c;
964 if (!event->attr.precise_ip)
967 if (x86_pmu.pebs_constraints) {
968 for_each_event_constraint(c, x86_pmu.pebs_constraints) {
969 if (constraint_match(c, event->hw.config)) {
970 event->hw.flags |= c->flags;
977 * Extended PEBS support
978 * Makes the PEBS code search the normal constraints.
980 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
983 return &emptyconstraint;
987 * We need the sched_task callback even for per-cpu events when we use
988 * the large interrupt threshold, such that we can provide PID and TID
991 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
993 if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
996 return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
999 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
1001 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1003 if (!sched_in && pebs_needs_sched_cb(cpuc))
1004 intel_pmu_drain_pebs_buffer();
1007 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1009 struct debug_store *ds = cpuc->ds;
1013 if (cpuc->n_pebs_via_pt)
1016 if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1017 reserved = x86_pmu.max_pebs_events + x86_pmu.num_counters_fixed;
1019 reserved = x86_pmu.max_pebs_events;
1021 if (cpuc->n_pebs == cpuc->n_large_pebs) {
1022 threshold = ds->pebs_absolute_maximum -
1023 reserved * cpuc->pebs_record_size;
1025 threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1028 ds->pebs_interrupt_threshold = threshold;
1031 static void adaptive_pebs_record_size_update(void)
1033 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1034 u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1035 int sz = sizeof(struct pebs_basic);
1037 if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1038 sz += sizeof(struct pebs_meminfo);
1039 if (pebs_data_cfg & PEBS_DATACFG_GP)
1040 sz += sizeof(struct pebs_gprs);
1041 if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1042 sz += sizeof(struct pebs_xmm);
1043 if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1044 sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1046 cpuc->pebs_record_size = sz;
1049 #define PERF_PEBS_MEMINFO_TYPE (PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC | \
1050 PERF_SAMPLE_PHYS_ADDR | \
1051 PERF_SAMPLE_WEIGHT_TYPE | \
1052 PERF_SAMPLE_TRANSACTION | \
1053 PERF_SAMPLE_DATA_PAGE_SIZE)
1055 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1057 struct perf_event_attr *attr = &event->attr;
1058 u64 sample_type = attr->sample_type;
1059 u64 pebs_data_cfg = 0;
1060 bool gprs, tsx_weight;
1062 if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1063 attr->precise_ip > 1)
1064 return pebs_data_cfg;
1066 if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1067 pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1070 * We need GPRs when:
1071 * + user requested them
1072 * + precise_ip < 2 for the non event IP
1073 * + For RTM TSX weight we need GPRs for the abort code.
1075 gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
1076 (attr->sample_regs_intr & PEBS_GP_REGS);
1078 tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1079 ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1080 x86_pmu.rtm_abort_event);
1082 if (gprs || (attr->precise_ip < 2) || tsx_weight)
1083 pebs_data_cfg |= PEBS_DATACFG_GP;
1085 if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1086 (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1087 pebs_data_cfg |= PEBS_DATACFG_XMMS;
1089 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1091 * For now always log all LBRs. Could configure this
1094 pebs_data_cfg |= PEBS_DATACFG_LBRS |
1095 ((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1098 return pebs_data_cfg;
1102 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1103 struct perf_event *event, bool add)
1105 struct pmu *pmu = event->ctx->pmu;
1107 * Make sure we get updated with the first PEBS
1108 * event. It will trigger also during removal, but
1109 * that does not hurt:
1111 bool update = cpuc->n_pebs == 1;
1113 if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1115 perf_sched_cb_inc(pmu);
1117 perf_sched_cb_dec(pmu);
1123 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1124 * iterating all remaining PEBS events to reconstruct the config.
1126 if (x86_pmu.intel_cap.pebs_baseline && add) {
1129 /* Clear pebs_data_cfg and pebs_record_size for first PEBS. */
1130 if (cpuc->n_pebs == 1) {
1131 cpuc->pebs_data_cfg = 0;
1132 cpuc->pebs_record_size = sizeof(struct pebs_basic);
1135 pebs_data_cfg = pebs_update_adaptive_cfg(event);
1137 /* Update pebs_record_size if new event requires more data. */
1138 if (pebs_data_cfg & ~cpuc->pebs_data_cfg) {
1139 cpuc->pebs_data_cfg |= pebs_data_cfg;
1140 adaptive_pebs_record_size_update();
1146 pebs_update_threshold(cpuc);
1149 void intel_pmu_pebs_add(struct perf_event *event)
1151 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1152 struct hw_perf_event *hwc = &event->hw;
1153 bool needed_cb = pebs_needs_sched_cb(cpuc);
1156 if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1157 cpuc->n_large_pebs++;
1158 if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1159 cpuc->n_pebs_via_pt++;
1161 pebs_update_state(needed_cb, cpuc, event, true);
1164 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1166 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1168 if (!is_pebs_pt(event))
1171 if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1172 cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1175 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1177 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1178 struct hw_perf_event *hwc = &event->hw;
1179 struct debug_store *ds = cpuc->ds;
1181 if (!is_pebs_pt(event))
1184 if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1185 cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1187 cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1189 wrmsrl(MSR_RELOAD_PMC0 + hwc->idx, ds->pebs_event_reset[hwc->idx]);
1192 void intel_pmu_pebs_enable(struct perf_event *event)
1194 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1195 struct hw_perf_event *hwc = &event->hw;
1196 struct debug_store *ds = cpuc->ds;
1198 hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1200 cpuc->pebs_enabled |= 1ULL << hwc->idx;
1202 if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1203 cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1204 else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1205 cpuc->pebs_enabled |= 1ULL << 63;
1207 if (x86_pmu.intel_cap.pebs_baseline) {
1208 hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1209 if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1210 wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg);
1211 cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg;
1216 * Use auto-reload if possible to save a MSR write in the PMI.
1217 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1219 if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1220 unsigned int idx = hwc->idx;
1222 if (idx >= INTEL_PMC_IDX_FIXED)
1223 idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1224 ds->pebs_event_reset[idx] =
1225 (u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1227 ds->pebs_event_reset[hwc->idx] = 0;
1230 intel_pmu_pebs_via_pt_enable(event);
1233 void intel_pmu_pebs_del(struct perf_event *event)
1235 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1236 struct hw_perf_event *hwc = &event->hw;
1237 bool needed_cb = pebs_needs_sched_cb(cpuc);
1240 if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1241 cpuc->n_large_pebs--;
1242 if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1243 cpuc->n_pebs_via_pt--;
1245 pebs_update_state(needed_cb, cpuc, event, false);
1248 void intel_pmu_pebs_disable(struct perf_event *event)
1250 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1251 struct hw_perf_event *hwc = &event->hw;
1253 if (cpuc->n_pebs == cpuc->n_large_pebs &&
1254 cpuc->n_pebs != cpuc->n_pebs_via_pt)
1255 intel_pmu_drain_pebs_buffer();
1257 cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1259 if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1260 (x86_pmu.version < 5))
1261 cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1262 else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1263 cpuc->pebs_enabled &= ~(1ULL << 63);
1265 intel_pmu_pebs_via_pt_disable(event);
1268 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1270 hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1273 void intel_pmu_pebs_enable_all(void)
1275 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1277 if (cpuc->pebs_enabled)
1278 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1281 void intel_pmu_pebs_disable_all(void)
1283 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1285 if (cpuc->pebs_enabled)
1286 wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1289 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1291 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1292 unsigned long from = cpuc->lbr_entries[0].from;
1293 unsigned long old_to, to = cpuc->lbr_entries[0].to;
1294 unsigned long ip = regs->ip;
1300 * We don't need to fixup if the PEBS assist is fault like
1302 if (!x86_pmu.intel_cap.pebs_trap)
1306 * No LBR entry, no basic block, no rewinding
1308 if (!cpuc->lbr_stack.nr || !from || !to)
1312 * Basic blocks should never cross user/kernel boundaries
1314 if (kernel_ip(ip) != kernel_ip(to))
1318 * unsigned math, either ip is before the start (impossible) or
1319 * the basic block is larger than 1 page (sanity)
1321 if ((ip - to) > PEBS_FIXUP_SIZE)
1325 * We sampled a branch insn, rewind using the LBR stack
1328 set_linear_ip(regs, from);
1333 if (!kernel_ip(ip)) {
1335 u8 *buf = this_cpu_read(insn_buffer);
1337 /* 'size' must fit our buffer, see above */
1338 bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1352 #ifdef CONFIG_X86_64
1353 is_64bit = kernel_ip(to) || any_64bit_mode(regs);
1355 insn_init(&insn, kaddr, size, is_64bit);
1356 insn_get_length(&insn);
1358 * Make sure there was not a problem decoding the
1359 * instruction and getting the length. This is
1360 * doubly important because we have an infinite
1361 * loop if insn.length=0.
1367 kaddr += insn.length;
1368 size -= insn.length;
1372 set_linear_ip(regs, old_to);
1377 * Even though we decoded the basic block, the instruction stream
1378 * never matched the given IP, either the TO or the IP got corrupted.
1383 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1386 union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1387 return tsx.cycles_last_block;
1392 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1394 u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1396 /* For RTM XABORTs also log the abort code from AX */
1397 if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1398 txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1402 static inline u64 get_pebs_status(void *n)
1404 if (x86_pmu.intel_cap.pebs_format < 4)
1405 return ((struct pebs_record_nhm *)n)->status;
1406 return ((struct pebs_basic *)n)->applicable_counters;
1409 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1410 (PERF_X86_EVENT_PEBS_ST_HSW | \
1411 PERF_X86_EVENT_PEBS_LD_HSW | \
1412 PERF_X86_EVENT_PEBS_NA_HSW)
1414 static u64 get_data_src(struct perf_event *event, u64 aux)
1416 u64 val = PERF_MEM_NA;
1417 int fl = event->hw.flags;
1418 bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1420 if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1421 val = load_latency_data(aux);
1422 else if (fl & PERF_X86_EVENT_PEBS_STLAT)
1423 val = store_latency_data(aux);
1424 else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1425 val = precise_datala_hsw(event, aux);
1427 val = precise_store_data(aux);
1431 #define PERF_SAMPLE_ADDR_TYPE (PERF_SAMPLE_ADDR | \
1432 PERF_SAMPLE_PHYS_ADDR | \
1433 PERF_SAMPLE_DATA_PAGE_SIZE)
1435 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1436 struct pt_regs *iregs, void *__pebs,
1437 struct perf_sample_data *data,
1438 struct pt_regs *regs)
1441 * We cast to the biggest pebs_record but are careful not to
1442 * unconditionally access the 'extra' entries.
1444 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1445 struct pebs_record_skl *pebs = __pebs;
1452 sample_type = event->attr.sample_type;
1453 fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1455 perf_sample_data_init(data, 0, event->hw.last_period);
1457 data->period = event->hw.last_period;
1460 * Use latency for weight (only avail with PEBS-LL)
1462 if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE))
1463 data->weight.full = pebs->lat;
1466 * data.data_src encodes the data source
1468 if (sample_type & PERF_SAMPLE_DATA_SRC)
1469 data->data_src.val = get_data_src(event, pebs->dse);
1472 * We must however always use iregs for the unwinder to stay sane; the
1473 * record BP,SP,IP can point into thin air when the record is from a
1474 * previous PMI context or an (I)RET happened between the record and
1477 if (sample_type & PERF_SAMPLE_CALLCHAIN)
1478 data->callchain = perf_callchain(event, iregs);
1481 * We use the interrupt regs as a base because the PEBS record does not
1482 * contain a full regs set, specifically it seems to lack segment
1483 * descriptors, which get used by things like user_mode().
1485 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1490 * Initialize regs_>flags from PEBS,
1491 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1492 * i.e., do not rely on it being zero:
1494 regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1496 if (sample_type & PERF_SAMPLE_REGS_INTR) {
1497 regs->ax = pebs->ax;
1498 regs->bx = pebs->bx;
1499 regs->cx = pebs->cx;
1500 regs->dx = pebs->dx;
1501 regs->si = pebs->si;
1502 regs->di = pebs->di;
1504 regs->bp = pebs->bp;
1505 regs->sp = pebs->sp;
1507 #ifndef CONFIG_X86_32
1508 regs->r8 = pebs->r8;
1509 regs->r9 = pebs->r9;
1510 regs->r10 = pebs->r10;
1511 regs->r11 = pebs->r11;
1512 regs->r12 = pebs->r12;
1513 regs->r13 = pebs->r13;
1514 regs->r14 = pebs->r14;
1515 regs->r15 = pebs->r15;
1519 if (event->attr.precise_ip > 1) {
1521 * Haswell and later processors have an 'eventing IP'
1522 * (real IP) which fixes the off-by-1 skid in hardware.
1523 * Use it when precise_ip >= 2 :
1525 if (x86_pmu.intel_cap.pebs_format >= 2) {
1526 set_linear_ip(regs, pebs->real_ip);
1527 regs->flags |= PERF_EFLAGS_EXACT;
1529 /* Otherwise, use PEBS off-by-1 IP: */
1530 set_linear_ip(regs, pebs->ip);
1533 * With precise_ip >= 2, try to fix up the off-by-1 IP
1534 * using the LBR. If successful, the fixup function
1535 * corrects regs->ip and calls set_linear_ip() on regs:
1537 if (intel_pmu_pebs_fixup_ip(regs))
1538 regs->flags |= PERF_EFLAGS_EXACT;
1542 * When precise_ip == 1, return the PEBS off-by-1 IP,
1543 * no fixup attempted:
1545 set_linear_ip(regs, pebs->ip);
1549 if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
1550 x86_pmu.intel_cap.pebs_format >= 1)
1551 data->addr = pebs->dla;
1553 if (x86_pmu.intel_cap.pebs_format >= 2) {
1554 /* Only set the TSX weight when no memory weight. */
1555 if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll)
1556 data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
1558 if (sample_type & PERF_SAMPLE_TRANSACTION)
1559 data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1564 * v3 supplies an accurate time stamp, so we use that
1565 * for the time stamp.
1567 * We can only do this for the default trace clock.
1569 if (x86_pmu.intel_cap.pebs_format >= 3 &&
1570 event->attr.use_clockid == 0)
1571 data->time = native_sched_clock_from_tsc(pebs->tsc);
1573 if (has_branch_stack(event))
1574 data->br_stack = &cpuc->lbr_stack;
1577 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1578 struct pebs_gprs *gprs)
1580 regs->ax = gprs->ax;
1581 regs->bx = gprs->bx;
1582 regs->cx = gprs->cx;
1583 regs->dx = gprs->dx;
1584 regs->si = gprs->si;
1585 regs->di = gprs->di;
1586 regs->bp = gprs->bp;
1587 regs->sp = gprs->sp;
1588 #ifndef CONFIG_X86_32
1589 regs->r8 = gprs->r8;
1590 regs->r9 = gprs->r9;
1591 regs->r10 = gprs->r10;
1592 regs->r11 = gprs->r11;
1593 regs->r12 = gprs->r12;
1594 regs->r13 = gprs->r13;
1595 regs->r14 = gprs->r14;
1596 regs->r15 = gprs->r15;
1600 #define PEBS_LATENCY_MASK 0xffff
1601 #define PEBS_CACHE_LATENCY_OFFSET 32
1604 * With adaptive PEBS the layout depends on what fields are configured.
1607 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1608 struct pt_regs *iregs, void *__pebs,
1609 struct perf_sample_data *data,
1610 struct pt_regs *regs)
1612 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1613 struct pebs_basic *basic = __pebs;
1614 void *next_record = basic + 1;
1617 struct pebs_meminfo *meminfo = NULL;
1618 struct pebs_gprs *gprs = NULL;
1619 struct x86_perf_regs *perf_regs;
1624 perf_regs = container_of(regs, struct x86_perf_regs, regs);
1625 perf_regs->xmm_regs = NULL;
1627 sample_type = event->attr.sample_type;
1628 format_size = basic->format_size;
1629 perf_sample_data_init(data, 0, event->hw.last_period);
1630 data->period = event->hw.last_period;
1632 if (event->attr.use_clockid == 0)
1633 data->time = native_sched_clock_from_tsc(basic->tsc);
1636 * We must however always use iregs for the unwinder to stay sane; the
1637 * record BP,SP,IP can point into thin air when the record is from a
1638 * previous PMI context or an (I)RET happened between the record and
1641 if (sample_type & PERF_SAMPLE_CALLCHAIN)
1642 data->callchain = perf_callchain(event, iregs);
1645 /* The ip in basic is EventingIP */
1646 set_linear_ip(regs, basic->ip);
1647 regs->flags = PERF_EFLAGS_EXACT;
1650 * The record for MEMINFO is in front of GP
1651 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1652 * Save the pointer here but process later.
1654 if (format_size & PEBS_DATACFG_MEMINFO) {
1655 meminfo = next_record;
1656 next_record = meminfo + 1;
1659 if (format_size & PEBS_DATACFG_GP) {
1661 next_record = gprs + 1;
1663 if (event->attr.precise_ip < 2) {
1664 set_linear_ip(regs, gprs->ip);
1665 regs->flags &= ~PERF_EFLAGS_EXACT;
1668 if (sample_type & PERF_SAMPLE_REGS_INTR)
1669 adaptive_pebs_save_regs(regs, gprs);
1672 if (format_size & PEBS_DATACFG_MEMINFO) {
1673 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
1674 u64 weight = meminfo->latency;
1676 if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) {
1677 data->weight.var2_w = weight & PEBS_LATENCY_MASK;
1678 weight >>= PEBS_CACHE_LATENCY_OFFSET;
1682 * Although meminfo::latency is defined as a u64,
1683 * only the lower 32 bits include the valid data
1684 * in practice on Ice Lake and earlier platforms.
1686 if (sample_type & PERF_SAMPLE_WEIGHT) {
1687 data->weight.full = weight ?:
1688 intel_get_tsx_weight(meminfo->tsx_tuning);
1690 data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?:
1691 intel_get_tsx_weight(meminfo->tsx_tuning);
1695 if (sample_type & PERF_SAMPLE_DATA_SRC)
1696 data->data_src.val = get_data_src(event, meminfo->aux);
1698 if (sample_type & PERF_SAMPLE_ADDR_TYPE)
1699 data->addr = meminfo->address;
1701 if (sample_type & PERF_SAMPLE_TRANSACTION)
1702 data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1703 gprs ? gprs->ax : 0);
1706 if (format_size & PEBS_DATACFG_XMMS) {
1707 struct pebs_xmm *xmm = next_record;
1709 next_record = xmm + 1;
1710 perf_regs->xmm_regs = xmm->xmm;
1713 if (format_size & PEBS_DATACFG_LBRS) {
1714 struct lbr_entry *lbr = next_record;
1715 int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1717 next_record = next_record + num_lbr * sizeof(struct lbr_entry);
1719 if (has_branch_stack(event)) {
1720 intel_pmu_store_pebs_lbrs(lbr);
1721 data->br_stack = &cpuc->lbr_stack;
1725 WARN_ONCE(next_record != __pebs + (format_size >> 48),
1726 "PEBS record size %llu, expected %llu, config %llx\n",
1728 (u64)(next_record - __pebs),
1729 basic->format_size);
1732 static inline void *
1733 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1735 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1740 * fmt0 does not have a status bitfield (does not use
1741 * perf_record_nhm format)
1743 if (x86_pmu.intel_cap.pebs_format < 1)
1749 for (at = base; at < top; at += cpuc->pebs_record_size) {
1750 unsigned long status = get_pebs_status(at);
1752 if (test_bit(bit, (unsigned long *)&status)) {
1753 /* PEBS v3 has accurate status bits */
1754 if (x86_pmu.intel_cap.pebs_format >= 3)
1757 if (status == (1 << bit))
1760 /* clear non-PEBS bit and re-check */
1761 pebs_status = status & cpuc->pebs_enabled;
1762 pebs_status &= PEBS_COUNTER_MASK;
1763 if (pebs_status == (1 << bit))
1770 void intel_pmu_auto_reload_read(struct perf_event *event)
1772 WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1774 perf_pmu_disable(event->pmu);
1775 intel_pmu_drain_pebs_buffer();
1776 perf_pmu_enable(event->pmu);
1780 * Special variant of intel_pmu_save_and_restart() for auto-reload.
1783 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1785 struct hw_perf_event *hwc = &event->hw;
1786 int shift = 64 - x86_pmu.cntval_bits;
1787 u64 period = hwc->sample_period;
1788 u64 prev_raw_count, new_raw_count;
1794 * drain_pebs() only happens when the PMU is disabled.
1796 WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1798 prev_raw_count = local64_read(&hwc->prev_count);
1799 rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1800 local64_set(&hwc->prev_count, new_raw_count);
1803 * Since the counter increments a negative counter value and
1804 * overflows on the sign switch, giving the interval:
1808 * the difference between two consequtive reads is:
1810 * A) value2 - value1;
1811 * when no overflows have happened in between,
1813 * B) (0 - value1) + (value2 - (-period));
1814 * when one overflow happened in between,
1816 * C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1817 * when @n overflows happened in between.
1819 * Here A) is the obvious difference, B) is the extension to the
1820 * discrete interval, where the first term is to the top of the
1821 * interval and the second term is from the bottom of the next
1822 * interval and C) the extension to multiple intervals, where the
1823 * middle term is the whole intervals covered.
1825 * An equivalent of C, by reduction, is:
1827 * value2 - value1 + n * period
1829 new = ((s64)(new_raw_count << shift) >> shift);
1830 old = ((s64)(prev_raw_count << shift) >> shift);
1831 local64_add(new - old + count * period, &event->count);
1833 local64_set(&hwc->period_left, -new);
1835 perf_event_update_userpage(event);
1840 static __always_inline void
1841 __intel_pmu_pebs_event(struct perf_event *event,
1842 struct pt_regs *iregs,
1843 struct perf_sample_data *data,
1844 void *base, void *top,
1846 void (*setup_sample)(struct perf_event *,
1849 struct perf_sample_data *,
1852 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1853 struct hw_perf_event *hwc = &event->hw;
1854 struct x86_perf_regs perf_regs;
1855 struct pt_regs *regs = &perf_regs.regs;
1856 void *at = get_next_pebs_record_by_bit(base, top, bit);
1857 static struct pt_regs dummy_iregs;
1859 if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1861 * Now, auto-reload is only enabled in fixed period mode.
1862 * The reload value is always hwc->sample_period.
1863 * May need to change it, if auto-reload is enabled in
1866 intel_pmu_save_and_restart_reload(event, count);
1867 } else if (!intel_pmu_save_and_restart(event))
1871 iregs = &dummy_iregs;
1874 setup_sample(event, iregs, at, data, regs);
1875 perf_event_output(event, data, regs);
1876 at += cpuc->pebs_record_size;
1877 at = get_next_pebs_record_by_bit(at, top, bit);
1881 setup_sample(event, iregs, at, data, regs);
1882 if (iregs == &dummy_iregs) {
1884 * The PEBS records may be drained in the non-overflow context,
1885 * e.g., large PEBS + context switch. Perf should treat the
1886 * last record the same as other PEBS records, and doesn't
1887 * invoke the generic overflow handler.
1889 perf_event_output(event, data, regs);
1892 * All but the last records are processed.
1893 * The last one is left to be able to call the overflow handler.
1895 if (perf_event_overflow(event, data, regs))
1896 x86_pmu_stop(event, 0);
1900 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
1902 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1903 struct debug_store *ds = cpuc->ds;
1904 struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1905 struct pebs_record_core *at, *top;
1908 if (!x86_pmu.pebs_active)
1911 at = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1912 top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1915 * Whatever else happens, drain the thing
1917 ds->pebs_index = ds->pebs_buffer_base;
1919 if (!test_bit(0, cpuc->active_mask))
1922 WARN_ON_ONCE(!event);
1924 if (!event->attr.precise_ip)
1929 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1930 intel_pmu_save_and_restart_reload(event, 0);
1934 __intel_pmu_pebs_event(event, iregs, data, at, top, 0, n,
1935 setup_pebs_fixed_sample_data);
1938 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
1940 struct perf_event *event;
1944 * The drain_pebs() could be called twice in a short period
1945 * for auto-reload event in pmu::read(). There are no
1946 * overflows have happened in between.
1947 * It needs to call intel_pmu_save_and_restart_reload() to
1948 * update the event->count for this case.
1950 for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
1951 event = cpuc->events[bit];
1952 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1953 intel_pmu_save_and_restart_reload(event, 0);
1957 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
1959 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1960 struct debug_store *ds = cpuc->ds;
1961 struct perf_event *event;
1962 void *base, *at, *top;
1963 short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1964 short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1968 if (!x86_pmu.pebs_active)
1971 base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1972 top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1974 ds->pebs_index = ds->pebs_buffer_base;
1976 mask = (1ULL << x86_pmu.max_pebs_events) - 1;
1977 size = x86_pmu.max_pebs_events;
1978 if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
1979 mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
1980 size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1983 if (unlikely(base >= top)) {
1984 intel_pmu_pebs_event_update_no_drain(cpuc, size);
1988 for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1989 struct pebs_record_nhm *p = at;
1992 pebs_status = p->status & cpuc->pebs_enabled;
1993 pebs_status &= mask;
1995 /* PEBS v3 has more accurate status bits */
1996 if (x86_pmu.intel_cap.pebs_format >= 3) {
1997 for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2004 * On some CPUs the PEBS status can be zero when PEBS is
2005 * racing with clearing of GLOBAL_STATUS.
2007 * Normally we would drop that record, but in the
2008 * case when there is only a single active PEBS event
2009 * we can assume it's for that event.
2011 if (!pebs_status && cpuc->pebs_enabled &&
2012 !(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
2013 pebs_status = cpuc->pebs_enabled;
2015 bit = find_first_bit((unsigned long *)&pebs_status,
2016 x86_pmu.max_pebs_events);
2017 if (bit >= x86_pmu.max_pebs_events)
2021 * The PEBS hardware does not deal well with the situation
2022 * when events happen near to each other and multiple bits
2023 * are set. But it should happen rarely.
2025 * If these events include one PEBS and multiple non-PEBS
2026 * events, it doesn't impact PEBS record. The record will
2027 * be handled normally. (slow path)
2029 * If these events include two or more PEBS events, the
2030 * records for the events can be collapsed into a single
2031 * one, and it's not possible to reconstruct all events
2032 * that caused the PEBS record. It's called collision.
2033 * If collision happened, the record will be dropped.
2035 if (pebs_status != (1ULL << bit)) {
2036 for_each_set_bit(i, (unsigned long *)&pebs_status, size)
2044 for_each_set_bit(bit, (unsigned long *)&mask, size) {
2045 if ((counts[bit] == 0) && (error[bit] == 0))
2048 event = cpuc->events[bit];
2049 if (WARN_ON_ONCE(!event))
2052 if (WARN_ON_ONCE(!event->attr.precise_ip))
2055 /* log dropped samples number */
2057 perf_log_lost_samples(event, error[bit]);
2059 if (iregs && perf_event_account_interrupt(event))
2060 x86_pmu_stop(event, 0);
2064 __intel_pmu_pebs_event(event, iregs, data, base,
2065 top, bit, counts[bit],
2066 setup_pebs_fixed_sample_data);
2071 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
2073 short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2074 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2075 struct debug_store *ds = cpuc->ds;
2076 struct perf_event *event;
2077 void *base, *at, *top;
2081 if (!x86_pmu.pebs_active)
2084 base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
2085 top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
2087 ds->pebs_index = ds->pebs_buffer_base;
2089 mask = ((1ULL << x86_pmu.max_pebs_events) - 1) |
2090 (((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
2091 size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
2093 if (unlikely(base >= top)) {
2094 intel_pmu_pebs_event_update_no_drain(cpuc, size);
2098 for (at = base; at < top; at += cpuc->pebs_record_size) {
2101 pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
2102 pebs_status &= mask;
2104 for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2108 for_each_set_bit(bit, (unsigned long *)&mask, size) {
2109 if (counts[bit] == 0)
2112 event = cpuc->events[bit];
2113 if (WARN_ON_ONCE(!event))
2116 if (WARN_ON_ONCE(!event->attr.precise_ip))
2119 __intel_pmu_pebs_event(event, iregs, data, base,
2120 top, bit, counts[bit],
2121 setup_pebs_adaptive_sample_data);
2126 * BTS, PEBS probe and setup
2129 void __init intel_ds_init(void)
2132 * No support for 32bit formats
2134 if (!boot_cpu_has(X86_FEATURE_DTES64))
2137 x86_pmu.bts = boot_cpu_has(X86_FEATURE_BTS);
2138 x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
2139 x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
2140 if (x86_pmu.version <= 4)
2141 x86_pmu.pebs_no_isolation = 1;
2144 char pebs_type = x86_pmu.intel_cap.pebs_trap ? '+' : '-';
2145 char *pebs_qual = "";
2146 int format = x86_pmu.intel_cap.pebs_format;
2149 x86_pmu.intel_cap.pebs_baseline = 0;
2153 pr_cont("PEBS fmt0%c, ", pebs_type);
2154 x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
2156 * Using >PAGE_SIZE buffers makes the WRMSR to
2157 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
2158 * mysteriously hang on Core2.
2160 * As a workaround, we don't do this.
2162 x86_pmu.pebs_buffer_size = PAGE_SIZE;
2163 x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
2167 pr_cont("PEBS fmt1%c, ", pebs_type);
2168 x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
2169 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2173 pr_cont("PEBS fmt2%c, ", pebs_type);
2174 x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2175 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2179 pr_cont("PEBS fmt3%c, ", pebs_type);
2180 x86_pmu.pebs_record_size =
2181 sizeof(struct pebs_record_skl);
2182 x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2183 x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2187 x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2188 x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2189 if (x86_pmu.intel_cap.pebs_baseline) {
2190 x86_pmu.large_pebs_flags |=
2191 PERF_SAMPLE_BRANCH_STACK |
2193 x86_pmu.flags |= PMU_FL_PEBS_ALL;
2194 pebs_qual = "-baseline";
2195 x86_get_pmu()->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
2197 /* Only basic record supported */
2198 x86_pmu.large_pebs_flags &=
2199 ~(PERF_SAMPLE_ADDR |
2201 PERF_SAMPLE_DATA_SRC |
2202 PERF_SAMPLE_TRANSACTION |
2203 PERF_SAMPLE_REGS_USER |
2204 PERF_SAMPLE_REGS_INTR);
2206 pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2208 if (x86_pmu.intel_cap.pebs_output_pt_available) {
2209 pr_cont("PEBS-via-PT, ");
2210 x86_get_pmu()->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
2216 pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2222 void perf_restore_debug_store(void)
2224 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2226 if (!x86_pmu.bts && !x86_pmu.pebs)
2229 wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);