xref: /linux/arch/x86/events/intel/ds.c (revision 5e4e38446a62a4f50d77b0dd11d4b379dee08988)
1 #include <linux/bitops.h>
2 #include <linux/types.h>
3 #include <linux/slab.h>
4 
5 #include <asm/perf_event.h>
6 #include <asm/insn.h>
7 
8 #include "../perf_event.h"
9 
10 /* The size of a BTS record in bytes: */
11 #define BTS_RECORD_SIZE		24
12 
13 #define BTS_BUFFER_SIZE		(PAGE_SIZE << 4)
14 #define PEBS_BUFFER_SIZE	(PAGE_SIZE << 4)
15 #define PEBS_FIXUP_SIZE		PAGE_SIZE
16 
17 /*
18  * pebs_record_32 for p4 and core not supported
19 
20 struct pebs_record_32 {
21 	u32 flags, ip;
22 	u32 ax, bc, cx, dx;
23 	u32 si, di, bp, sp;
24 };
25 
26  */
27 
28 union intel_x86_pebs_dse {
29 	u64 val;
30 	struct {
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;
35 	};
36 	struct {
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;
42 	};
43 };
44 
45 
46 /*
47  * Map PEBS Load Latency Data Source encodings to generic
48  * memory data source information
49  */
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))
53 
54 /* Version for Sandy Bridge and later */
55 static u64 pebs_data_source[] = {
56 	P(OP, LOAD) | P(LVL, MISS) | P(LVL, L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
57 	OP_LH | P(LVL, L1)  | P(SNOOP, NONE),	/* 0x01: L1 local */
58 	OP_LH | P(LVL, LFB) | P(SNOOP, NONE),	/* 0x02: LFB hit */
59 	OP_LH | P(LVL, L2)  | P(SNOOP, NONE),	/* 0x03: L2 hit */
60 	OP_LH | P(LVL, L3)  | P(SNOOP, NONE),	/* 0x04: L3 hit */
61 	OP_LH | P(LVL, L3)  | P(SNOOP, MISS),	/* 0x05: L3 hit, snoop miss */
62 	OP_LH | P(LVL, L3)  | P(SNOOP, HIT),	/* 0x06: L3 hit, snoop hit */
63 	OP_LH | P(LVL, L3)  | P(SNOOP, HITM),	/* 0x07: L3 hit, snoop hitm */
64 	OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
65 	OP_LH | P(LVL, REM_CCE1) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
66 	OP_LH | P(LVL, LOC_RAM)  | P(SNOOP, HIT),  /* 0x0a: L3 miss, shared */
67 	OP_LH | P(LVL, REM_RAM1) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
68 	OP_LH | P(LVL, LOC_RAM)  | SNOOP_NONE_MISS,/* 0x0c: L3 miss, excl */
69 	OP_LH | P(LVL, REM_RAM1) | SNOOP_NONE_MISS,/* 0x0d: L3 miss, excl */
70 	OP_LH | P(LVL, IO)  | P(SNOOP, NONE), /* 0x0e: I/O */
71 	OP_LH | P(LVL, UNC) | P(SNOOP, NONE), /* 0x0f: uncached */
72 };
73 
74 /* Patch up minor differences in the bits */
75 void __init intel_pmu_pebs_data_source_nhm(void)
76 {
77 	pebs_data_source[0x05] = OP_LH | P(LVL, L3)  | P(SNOOP, HIT);
78 	pebs_data_source[0x06] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
79 	pebs_data_source[0x07] = OP_LH | P(LVL, L3)  | P(SNOOP, HITM);
80 }
81 
82 static u64 precise_store_data(u64 status)
83 {
84 	union intel_x86_pebs_dse dse;
85 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
86 
87 	dse.val = status;
88 
89 	/*
90 	 * bit 4: TLB access
91 	 * 1 = stored missed 2nd level TLB
92 	 *
93 	 * so it either hit the walker or the OS
94 	 * otherwise hit 2nd level TLB
95 	 */
96 	if (dse.st_stlb_miss)
97 		val |= P(TLB, MISS);
98 	else
99 		val |= P(TLB, HIT);
100 
101 	/*
102 	 * bit 0: hit L1 data cache
103 	 * if not set, then all we know is that
104 	 * it missed L1D
105 	 */
106 	if (dse.st_l1d_hit)
107 		val |= P(LVL, HIT);
108 	else
109 		val |= P(LVL, MISS);
110 
111 	/*
112 	 * bit 5: Locked prefix
113 	 */
114 	if (dse.st_locked)
115 		val |= P(LOCK, LOCKED);
116 
117 	return val;
118 }
119 
120 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
121 {
122 	union perf_mem_data_src dse;
123 
124 	dse.val = PERF_MEM_NA;
125 
126 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
127 		dse.mem_op = PERF_MEM_OP_STORE;
128 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
129 		dse.mem_op = PERF_MEM_OP_LOAD;
130 
131 	/*
132 	 * L1 info only valid for following events:
133 	 *
134 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
135 	 * MEM_UOPS_RETIRED.LOCK_STORES
136 	 * MEM_UOPS_RETIRED.SPLIT_STORES
137 	 * MEM_UOPS_RETIRED.ALL_STORES
138 	 */
139 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
140 		if (status & 1)
141 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
142 		else
143 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
144 	}
145 	return dse.val;
146 }
147 
148 static u64 load_latency_data(u64 status)
149 {
150 	union intel_x86_pebs_dse dse;
151 	u64 val;
152 	int model = boot_cpu_data.x86_model;
153 	int fam = boot_cpu_data.x86;
154 
155 	dse.val = status;
156 
157 	/*
158 	 * use the mapping table for bit 0-3
159 	 */
160 	val = pebs_data_source[dse.ld_dse];
161 
162 	/*
163 	 * Nehalem models do not support TLB, Lock infos
164 	 */
165 	if (fam == 0x6 && (model == 26 || model == 30
166 	    || model == 31 || model == 46)) {
167 		val |= P(TLB, NA) | P(LOCK, NA);
168 		return val;
169 	}
170 	/*
171 	 * bit 4: TLB access
172 	 * 0 = did not miss 2nd level TLB
173 	 * 1 = missed 2nd level TLB
174 	 */
175 	if (dse.ld_stlb_miss)
176 		val |= P(TLB, MISS) | P(TLB, L2);
177 	else
178 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
179 
180 	/*
181 	 * bit 5: locked prefix
182 	 */
183 	if (dse.ld_locked)
184 		val |= P(LOCK, LOCKED);
185 
186 	return val;
187 }
188 
189 struct pebs_record_core {
190 	u64 flags, ip;
191 	u64 ax, bx, cx, dx;
192 	u64 si, di, bp, sp;
193 	u64 r8,  r9,  r10, r11;
194 	u64 r12, r13, r14, r15;
195 };
196 
197 struct pebs_record_nhm {
198 	u64 flags, ip;
199 	u64 ax, bx, cx, dx;
200 	u64 si, di, bp, sp;
201 	u64 r8,  r9,  r10, r11;
202 	u64 r12, r13, r14, r15;
203 	u64 status, dla, dse, lat;
204 };
205 
206 /*
207  * Same as pebs_record_nhm, with two additional fields.
208  */
209 struct pebs_record_hsw {
210 	u64 flags, ip;
211 	u64 ax, bx, cx, dx;
212 	u64 si, di, bp, sp;
213 	u64 r8,  r9,  r10, r11;
214 	u64 r12, r13, r14, r15;
215 	u64 status, dla, dse, lat;
216 	u64 real_ip, tsx_tuning;
217 };
218 
219 union hsw_tsx_tuning {
220 	struct {
221 		u32 cycles_last_block     : 32,
222 		    hle_abort		  : 1,
223 		    rtm_abort		  : 1,
224 		    instruction_abort     : 1,
225 		    non_instruction_abort : 1,
226 		    retry		  : 1,
227 		    data_conflict	  : 1,
228 		    capacity_writes	  : 1,
229 		    capacity_reads	  : 1;
230 	};
231 	u64	    value;
232 };
233 
234 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
235 
236 /* Same as HSW, plus TSC */
237 
238 struct pebs_record_skl {
239 	u64 flags, ip;
240 	u64 ax, bx, cx, dx;
241 	u64 si, di, bp, sp;
242 	u64 r8,  r9,  r10, r11;
243 	u64 r12, r13, r14, r15;
244 	u64 status, dla, dse, lat;
245 	u64 real_ip, tsx_tuning;
246 	u64 tsc;
247 };
248 
249 void init_debug_store_on_cpu(int cpu)
250 {
251 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
252 
253 	if (!ds)
254 		return;
255 
256 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
257 		     (u32)((u64)(unsigned long)ds),
258 		     (u32)((u64)(unsigned long)ds >> 32));
259 }
260 
261 void fini_debug_store_on_cpu(int cpu)
262 {
263 	if (!per_cpu(cpu_hw_events, cpu).ds)
264 		return;
265 
266 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
267 }
268 
269 static DEFINE_PER_CPU(void *, insn_buffer);
270 
271 static int alloc_pebs_buffer(int cpu)
272 {
273 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
274 	int node = cpu_to_node(cpu);
275 	int max;
276 	void *buffer, *ibuffer;
277 
278 	if (!x86_pmu.pebs)
279 		return 0;
280 
281 	buffer = kzalloc_node(x86_pmu.pebs_buffer_size, GFP_KERNEL, node);
282 	if (unlikely(!buffer))
283 		return -ENOMEM;
284 
285 	/*
286 	 * HSW+ already provides us the eventing ip; no need to allocate this
287 	 * buffer then.
288 	 */
289 	if (x86_pmu.intel_cap.pebs_format < 2) {
290 		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
291 		if (!ibuffer) {
292 			kfree(buffer);
293 			return -ENOMEM;
294 		}
295 		per_cpu(insn_buffer, cpu) = ibuffer;
296 	}
297 
298 	max = x86_pmu.pebs_buffer_size / x86_pmu.pebs_record_size;
299 
300 	ds->pebs_buffer_base = (u64)(unsigned long)buffer;
301 	ds->pebs_index = ds->pebs_buffer_base;
302 	ds->pebs_absolute_maximum = ds->pebs_buffer_base +
303 		max * x86_pmu.pebs_record_size;
304 
305 	return 0;
306 }
307 
308 static void release_pebs_buffer(int cpu)
309 {
310 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
311 
312 	if (!ds || !x86_pmu.pebs)
313 		return;
314 
315 	kfree(per_cpu(insn_buffer, cpu));
316 	per_cpu(insn_buffer, cpu) = NULL;
317 
318 	kfree((void *)(unsigned long)ds->pebs_buffer_base);
319 	ds->pebs_buffer_base = 0;
320 }
321 
322 static int alloc_bts_buffer(int cpu)
323 {
324 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
325 	int node = cpu_to_node(cpu);
326 	int max, thresh;
327 	void *buffer;
328 
329 	if (!x86_pmu.bts)
330 		return 0;
331 
332 	buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, node);
333 	if (unlikely(!buffer)) {
334 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
335 		return -ENOMEM;
336 	}
337 
338 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
339 	thresh = max / 16;
340 
341 	ds->bts_buffer_base = (u64)(unsigned long)buffer;
342 	ds->bts_index = ds->bts_buffer_base;
343 	ds->bts_absolute_maximum = ds->bts_buffer_base +
344 		max * BTS_RECORD_SIZE;
345 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
346 		thresh * BTS_RECORD_SIZE;
347 
348 	return 0;
349 }
350 
351 static void release_bts_buffer(int cpu)
352 {
353 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
354 
355 	if (!ds || !x86_pmu.bts)
356 		return;
357 
358 	kfree((void *)(unsigned long)ds->bts_buffer_base);
359 	ds->bts_buffer_base = 0;
360 }
361 
362 static int alloc_ds_buffer(int cpu)
363 {
364 	int node = cpu_to_node(cpu);
365 	struct debug_store *ds;
366 
367 	ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
368 	if (unlikely(!ds))
369 		return -ENOMEM;
370 
371 	per_cpu(cpu_hw_events, cpu).ds = ds;
372 
373 	return 0;
374 }
375 
376 static void release_ds_buffer(int cpu)
377 {
378 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
379 
380 	if (!ds)
381 		return;
382 
383 	per_cpu(cpu_hw_events, cpu).ds = NULL;
384 	kfree(ds);
385 }
386 
387 void release_ds_buffers(void)
388 {
389 	int cpu;
390 
391 	if (!x86_pmu.bts && !x86_pmu.pebs)
392 		return;
393 
394 	get_online_cpus();
395 	for_each_online_cpu(cpu)
396 		fini_debug_store_on_cpu(cpu);
397 
398 	for_each_possible_cpu(cpu) {
399 		release_pebs_buffer(cpu);
400 		release_bts_buffer(cpu);
401 		release_ds_buffer(cpu);
402 	}
403 	put_online_cpus();
404 }
405 
406 void reserve_ds_buffers(void)
407 {
408 	int bts_err = 0, pebs_err = 0;
409 	int cpu;
410 
411 	x86_pmu.bts_active = 0;
412 	x86_pmu.pebs_active = 0;
413 
414 	if (!x86_pmu.bts && !x86_pmu.pebs)
415 		return;
416 
417 	if (!x86_pmu.bts)
418 		bts_err = 1;
419 
420 	if (!x86_pmu.pebs)
421 		pebs_err = 1;
422 
423 	get_online_cpus();
424 
425 	for_each_possible_cpu(cpu) {
426 		if (alloc_ds_buffer(cpu)) {
427 			bts_err = 1;
428 			pebs_err = 1;
429 		}
430 
431 		if (!bts_err && alloc_bts_buffer(cpu))
432 			bts_err = 1;
433 
434 		if (!pebs_err && alloc_pebs_buffer(cpu))
435 			pebs_err = 1;
436 
437 		if (bts_err && pebs_err)
438 			break;
439 	}
440 
441 	if (bts_err) {
442 		for_each_possible_cpu(cpu)
443 			release_bts_buffer(cpu);
444 	}
445 
446 	if (pebs_err) {
447 		for_each_possible_cpu(cpu)
448 			release_pebs_buffer(cpu);
449 	}
450 
451 	if (bts_err && pebs_err) {
452 		for_each_possible_cpu(cpu)
453 			release_ds_buffer(cpu);
454 	} else {
455 		if (x86_pmu.bts && !bts_err)
456 			x86_pmu.bts_active = 1;
457 
458 		if (x86_pmu.pebs && !pebs_err)
459 			x86_pmu.pebs_active = 1;
460 
461 		for_each_online_cpu(cpu)
462 			init_debug_store_on_cpu(cpu);
463 	}
464 
465 	put_online_cpus();
466 }
467 
468 /*
469  * BTS
470  */
471 
472 struct event_constraint bts_constraint =
473 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
474 
475 void intel_pmu_enable_bts(u64 config)
476 {
477 	unsigned long debugctlmsr;
478 
479 	debugctlmsr = get_debugctlmsr();
480 
481 	debugctlmsr |= DEBUGCTLMSR_TR;
482 	debugctlmsr |= DEBUGCTLMSR_BTS;
483 	if (config & ARCH_PERFMON_EVENTSEL_INT)
484 		debugctlmsr |= DEBUGCTLMSR_BTINT;
485 
486 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
487 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
488 
489 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
490 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
491 
492 	update_debugctlmsr(debugctlmsr);
493 }
494 
495 void intel_pmu_disable_bts(void)
496 {
497 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
498 	unsigned long debugctlmsr;
499 
500 	if (!cpuc->ds)
501 		return;
502 
503 	debugctlmsr = get_debugctlmsr();
504 
505 	debugctlmsr &=
506 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
507 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
508 
509 	update_debugctlmsr(debugctlmsr);
510 }
511 
512 int intel_pmu_drain_bts_buffer(void)
513 {
514 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
515 	struct debug_store *ds = cpuc->ds;
516 	struct bts_record {
517 		u64	from;
518 		u64	to;
519 		u64	flags;
520 	};
521 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
522 	struct bts_record *at, *base, *top;
523 	struct perf_output_handle handle;
524 	struct perf_event_header header;
525 	struct perf_sample_data data;
526 	unsigned long skip = 0;
527 	struct pt_regs regs;
528 
529 	if (!event)
530 		return 0;
531 
532 	if (!x86_pmu.bts_active)
533 		return 0;
534 
535 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
536 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
537 
538 	if (top <= base)
539 		return 0;
540 
541 	memset(&regs, 0, sizeof(regs));
542 
543 	ds->bts_index = ds->bts_buffer_base;
544 
545 	perf_sample_data_init(&data, 0, event->hw.last_period);
546 
547 	/*
548 	 * BTS leaks kernel addresses in branches across the cpl boundary,
549 	 * such as traps or system calls, so unless the user is asking for
550 	 * kernel tracing (and right now it's not possible), we'd need to
551 	 * filter them out. But first we need to count how many of those we
552 	 * have in the current batch. This is an extra O(n) pass, however,
553 	 * it's much faster than the other one especially considering that
554 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
555 	 * alloc_bts_buffer()).
556 	 */
557 	for (at = base; at < top; at++) {
558 		/*
559 		 * Note that right now *this* BTS code only works if
560 		 * attr::exclude_kernel is set, but let's keep this extra
561 		 * check here in case that changes.
562 		 */
563 		if (event->attr.exclude_kernel &&
564 		    (kernel_ip(at->from) || kernel_ip(at->to)))
565 			skip++;
566 	}
567 
568 	/*
569 	 * Prepare a generic sample, i.e. fill in the invariant fields.
570 	 * We will overwrite the from and to address before we output
571 	 * the sample.
572 	 */
573 	perf_prepare_sample(&header, &data, event, &regs);
574 
575 	if (perf_output_begin(&handle, event, header.size *
576 			      (top - base - skip)))
577 		return 1;
578 
579 	for (at = base; at < top; at++) {
580 		/* Filter out any records that contain kernel addresses. */
581 		if (event->attr.exclude_kernel &&
582 		    (kernel_ip(at->from) || kernel_ip(at->to)))
583 			continue;
584 
585 		data.ip		= at->from;
586 		data.addr	= at->to;
587 
588 		perf_output_sample(&handle, &header, &data, event);
589 	}
590 
591 	perf_output_end(&handle);
592 
593 	/* There's new data available. */
594 	event->hw.interrupts++;
595 	event->pending_kill = POLL_IN;
596 	return 1;
597 }
598 
599 static inline void intel_pmu_drain_pebs_buffer(void)
600 {
601 	struct pt_regs regs;
602 
603 	x86_pmu.drain_pebs(&regs);
604 }
605 
606 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
607 {
608 	if (!sched_in)
609 		intel_pmu_drain_pebs_buffer();
610 }
611 
612 /*
613  * PEBS
614  */
615 struct event_constraint intel_core2_pebs_event_constraints[] = {
616 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
617 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
618 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
619 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
620 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
621 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
622 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
623 	EVENT_CONSTRAINT_END
624 };
625 
626 struct event_constraint intel_atom_pebs_event_constraints[] = {
627 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
628 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
629 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
630 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
631 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
632 	/* Allow all events as PEBS with no flags */
633 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
634 	EVENT_CONSTRAINT_END
635 };
636 
637 struct event_constraint intel_slm_pebs_event_constraints[] = {
638 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
639 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
640 	/* Allow all events as PEBS with no flags */
641 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
642 	EVENT_CONSTRAINT_END
643 };
644 
645 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
646 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
647 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
648 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
649 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
650 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
651 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
652 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
653 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
654 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
655 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
656 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
657 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
658 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
659 	EVENT_CONSTRAINT_END
660 };
661 
662 struct event_constraint intel_westmere_pebs_event_constraints[] = {
663 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
664 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
665 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
666 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
667 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
668 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
669 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
670 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
671 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
672 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
673 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
674 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
675 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
676 	EVENT_CONSTRAINT_END
677 };
678 
679 struct event_constraint intel_snb_pebs_event_constraints[] = {
680 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
681 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
682 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
683 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
684 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
685         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
686         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
687         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
688         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
689 	/* Allow all events as PEBS with no flags */
690 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
691 	EVENT_CONSTRAINT_END
692 };
693 
694 struct event_constraint intel_ivb_pebs_event_constraints[] = {
695         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
696         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
697 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
698 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
699 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
700 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
701 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
702 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
703 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
704 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
705 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
706 	/* Allow all events as PEBS with no flags */
707 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
708         EVENT_CONSTRAINT_END
709 };
710 
711 struct event_constraint intel_hsw_pebs_event_constraints[] = {
712 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
713 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
714 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
715 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
716 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
717 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
718 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
719 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
720 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
721 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
722 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
723 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
724 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
725 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
726 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
727 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
728 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
729 	/* Allow all events as PEBS with no flags */
730 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
731 	EVENT_CONSTRAINT_END
732 };
733 
734 struct event_constraint intel_bdw_pebs_event_constraints[] = {
735 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
736 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
737 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
738 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
739 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
740 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
741 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
742 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
743 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
744 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
745 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
746 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
747 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
748 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
749 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
750 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
751 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
752 	/* Allow all events as PEBS with no flags */
753 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
754 	EVENT_CONSTRAINT_END
755 };
756 
757 
758 struct event_constraint intel_skl_pebs_event_constraints[] = {
759 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
760 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
761 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
762 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
763 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
764 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
765 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
766 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
767 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
768 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
769 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
770 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
771 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
772 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
773 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
774 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
775 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
776 	/* Allow all events as PEBS with no flags */
777 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
778 	EVENT_CONSTRAINT_END
779 };
780 
781 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
782 {
783 	struct event_constraint *c;
784 
785 	if (!event->attr.precise_ip)
786 		return NULL;
787 
788 	if (x86_pmu.pebs_constraints) {
789 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
790 			if ((event->hw.config & c->cmask) == c->code) {
791 				event->hw.flags |= c->flags;
792 				return c;
793 			}
794 		}
795 	}
796 
797 	return &emptyconstraint;
798 }
799 
800 static inline bool pebs_is_enabled(struct cpu_hw_events *cpuc)
801 {
802 	return (cpuc->pebs_enabled & ((1ULL << MAX_PEBS_EVENTS) - 1));
803 }
804 
805 void intel_pmu_pebs_enable(struct perf_event *event)
806 {
807 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
808 	struct hw_perf_event *hwc = &event->hw;
809 	struct debug_store *ds = cpuc->ds;
810 	bool first_pebs;
811 	u64 threshold;
812 
813 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
814 
815 	first_pebs = !pebs_is_enabled(cpuc);
816 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
817 
818 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
819 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
820 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
821 		cpuc->pebs_enabled |= 1ULL << 63;
822 
823 	/*
824 	 * When the event is constrained enough we can use a larger
825 	 * threshold and run the event with less frequent PMI.
826 	 */
827 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING) {
828 		threshold = ds->pebs_absolute_maximum -
829 			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
830 
831 		if (first_pebs)
832 			perf_sched_cb_inc(event->ctx->pmu);
833 	} else {
834 		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
835 
836 		/*
837 		 * If not all events can use larger buffer,
838 		 * roll back to threshold = 1
839 		 */
840 		if (!first_pebs &&
841 		    (ds->pebs_interrupt_threshold > threshold))
842 			perf_sched_cb_dec(event->ctx->pmu);
843 	}
844 
845 	/* Use auto-reload if possible to save a MSR write in the PMI */
846 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
847 		ds->pebs_event_reset[hwc->idx] =
848 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
849 	}
850 
851 	if (first_pebs || ds->pebs_interrupt_threshold > threshold)
852 		ds->pebs_interrupt_threshold = threshold;
853 }
854 
855 void intel_pmu_pebs_disable(struct perf_event *event)
856 {
857 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
858 	struct hw_perf_event *hwc = &event->hw;
859 	struct debug_store *ds = cpuc->ds;
860 	bool large_pebs = ds->pebs_interrupt_threshold >
861 		ds->pebs_buffer_base + x86_pmu.pebs_record_size;
862 
863 	if (large_pebs)
864 		intel_pmu_drain_pebs_buffer();
865 
866 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
867 
868 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
869 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
870 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
871 		cpuc->pebs_enabled &= ~(1ULL << 63);
872 
873 	if (large_pebs && !pebs_is_enabled(cpuc))
874 		perf_sched_cb_dec(event->ctx->pmu);
875 
876 	if (cpuc->enabled)
877 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
878 
879 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
880 }
881 
882 void intel_pmu_pebs_enable_all(void)
883 {
884 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
885 
886 	if (cpuc->pebs_enabled)
887 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
888 }
889 
890 void intel_pmu_pebs_disable_all(void)
891 {
892 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
893 
894 	if (cpuc->pebs_enabled)
895 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
896 }
897 
898 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
899 {
900 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
901 	unsigned long from = cpuc->lbr_entries[0].from;
902 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
903 	unsigned long ip = regs->ip;
904 	int is_64bit = 0;
905 	void *kaddr;
906 	int size;
907 
908 	/*
909 	 * We don't need to fixup if the PEBS assist is fault like
910 	 */
911 	if (!x86_pmu.intel_cap.pebs_trap)
912 		return 1;
913 
914 	/*
915 	 * No LBR entry, no basic block, no rewinding
916 	 */
917 	if (!cpuc->lbr_stack.nr || !from || !to)
918 		return 0;
919 
920 	/*
921 	 * Basic blocks should never cross user/kernel boundaries
922 	 */
923 	if (kernel_ip(ip) != kernel_ip(to))
924 		return 0;
925 
926 	/*
927 	 * unsigned math, either ip is before the start (impossible) or
928 	 * the basic block is larger than 1 page (sanity)
929 	 */
930 	if ((ip - to) > PEBS_FIXUP_SIZE)
931 		return 0;
932 
933 	/*
934 	 * We sampled a branch insn, rewind using the LBR stack
935 	 */
936 	if (ip == to) {
937 		set_linear_ip(regs, from);
938 		return 1;
939 	}
940 
941 	size = ip - to;
942 	if (!kernel_ip(ip)) {
943 		int bytes;
944 		u8 *buf = this_cpu_read(insn_buffer);
945 
946 		/* 'size' must fit our buffer, see above */
947 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
948 		if (bytes != 0)
949 			return 0;
950 
951 		kaddr = buf;
952 	} else {
953 		kaddr = (void *)to;
954 	}
955 
956 	do {
957 		struct insn insn;
958 
959 		old_to = to;
960 
961 #ifdef CONFIG_X86_64
962 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
963 #endif
964 		insn_init(&insn, kaddr, size, is_64bit);
965 		insn_get_length(&insn);
966 		/*
967 		 * Make sure there was not a problem decoding the
968 		 * instruction and getting the length.  This is
969 		 * doubly important because we have an infinite
970 		 * loop if insn.length=0.
971 		 */
972 		if (!insn.length)
973 			break;
974 
975 		to += insn.length;
976 		kaddr += insn.length;
977 		size -= insn.length;
978 	} while (to < ip);
979 
980 	if (to == ip) {
981 		set_linear_ip(regs, old_to);
982 		return 1;
983 	}
984 
985 	/*
986 	 * Even though we decoded the basic block, the instruction stream
987 	 * never matched the given IP, either the TO or the IP got corrupted.
988 	 */
989 	return 0;
990 }
991 
992 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
993 {
994 	if (pebs->tsx_tuning) {
995 		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
996 		return tsx.cycles_last_block;
997 	}
998 	return 0;
999 }
1000 
1001 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1002 {
1003 	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1004 
1005 	/* For RTM XABORTs also log the abort code from AX */
1006 	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1007 		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1008 	return txn;
1009 }
1010 
1011 static void setup_pebs_sample_data(struct perf_event *event,
1012 				   struct pt_regs *iregs, void *__pebs,
1013 				   struct perf_sample_data *data,
1014 				   struct pt_regs *regs)
1015 {
1016 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1017 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1018 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1019 		 PERF_X86_EVENT_PEBS_NA_HSW)
1020 	/*
1021 	 * We cast to the biggest pebs_record but are careful not to
1022 	 * unconditionally access the 'extra' entries.
1023 	 */
1024 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1025 	struct pebs_record_skl *pebs = __pebs;
1026 	u64 sample_type;
1027 	int fll, fst, dsrc;
1028 	int fl = event->hw.flags;
1029 
1030 	if (pebs == NULL)
1031 		return;
1032 
1033 	sample_type = event->attr.sample_type;
1034 	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1035 
1036 	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1037 	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1038 
1039 	perf_sample_data_init(data, 0, event->hw.last_period);
1040 
1041 	data->period = event->hw.last_period;
1042 
1043 	/*
1044 	 * Use latency for weight (only avail with PEBS-LL)
1045 	 */
1046 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1047 		data->weight = pebs->lat;
1048 
1049 	/*
1050 	 * data.data_src encodes the data source
1051 	 */
1052 	if (dsrc) {
1053 		u64 val = PERF_MEM_NA;
1054 		if (fll)
1055 			val = load_latency_data(pebs->dse);
1056 		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1057 			val = precise_datala_hsw(event, pebs->dse);
1058 		else if (fst)
1059 			val = precise_store_data(pebs->dse);
1060 		data->data_src.val = val;
1061 	}
1062 
1063 	/*
1064 	 * We use the interrupt regs as a base because the PEBS record
1065 	 * does not contain a full regs set, specifically it seems to
1066 	 * lack segment descriptors, which get used by things like
1067 	 * user_mode().
1068 	 *
1069 	 * In the simple case fix up only the IP and BP,SP regs, for
1070 	 * PERF_SAMPLE_IP and PERF_SAMPLE_CALLCHAIN to function properly.
1071 	 * A possible PERF_SAMPLE_REGS will have to transfer all regs.
1072 	 */
1073 	*regs = *iregs;
1074 	regs->flags = pebs->flags;
1075 	set_linear_ip(regs, pebs->ip);
1076 	regs->bp = pebs->bp;
1077 	regs->sp = pebs->sp;
1078 
1079 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1080 		regs->ax = pebs->ax;
1081 		regs->bx = pebs->bx;
1082 		regs->cx = pebs->cx;
1083 		regs->dx = pebs->dx;
1084 		regs->si = pebs->si;
1085 		regs->di = pebs->di;
1086 		regs->bp = pebs->bp;
1087 		regs->sp = pebs->sp;
1088 
1089 		regs->flags = pebs->flags;
1090 #ifndef CONFIG_X86_32
1091 		regs->r8 = pebs->r8;
1092 		regs->r9 = pebs->r9;
1093 		regs->r10 = pebs->r10;
1094 		regs->r11 = pebs->r11;
1095 		regs->r12 = pebs->r12;
1096 		regs->r13 = pebs->r13;
1097 		regs->r14 = pebs->r14;
1098 		regs->r15 = pebs->r15;
1099 #endif
1100 	}
1101 
1102 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1103 		regs->ip = pebs->real_ip;
1104 		regs->flags |= PERF_EFLAGS_EXACT;
1105 	} else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1106 		regs->flags |= PERF_EFLAGS_EXACT;
1107 	else
1108 		regs->flags &= ~PERF_EFLAGS_EXACT;
1109 
1110 	if ((sample_type & PERF_SAMPLE_ADDR) &&
1111 	    x86_pmu.intel_cap.pebs_format >= 1)
1112 		data->addr = pebs->dla;
1113 
1114 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1115 		/* Only set the TSX weight when no memory weight. */
1116 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1117 			data->weight = intel_hsw_weight(pebs);
1118 
1119 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1120 			data->txn = intel_hsw_transaction(pebs);
1121 	}
1122 
1123 	/*
1124 	 * v3 supplies an accurate time stamp, so we use that
1125 	 * for the time stamp.
1126 	 *
1127 	 * We can only do this for the default trace clock.
1128 	 */
1129 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1130 		event->attr.use_clockid == 0)
1131 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1132 
1133 	if (has_branch_stack(event))
1134 		data->br_stack = &cpuc->lbr_stack;
1135 }
1136 
1137 static inline void *
1138 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1139 {
1140 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1141 	void *at;
1142 	u64 pebs_status;
1143 
1144 	/*
1145 	 * fmt0 does not have a status bitfield (does not use
1146 	 * perf_record_nhm format)
1147 	 */
1148 	if (x86_pmu.intel_cap.pebs_format < 1)
1149 		return base;
1150 
1151 	if (base == NULL)
1152 		return NULL;
1153 
1154 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1155 		struct pebs_record_nhm *p = at;
1156 
1157 		if (test_bit(bit, (unsigned long *)&p->status)) {
1158 			/* PEBS v3 has accurate status bits */
1159 			if (x86_pmu.intel_cap.pebs_format >= 3)
1160 				return at;
1161 
1162 			if (p->status == (1 << bit))
1163 				return at;
1164 
1165 			/* clear non-PEBS bit and re-check */
1166 			pebs_status = p->status & cpuc->pebs_enabled;
1167 			pebs_status &= (1ULL << MAX_PEBS_EVENTS) - 1;
1168 			if (pebs_status == (1 << bit))
1169 				return at;
1170 		}
1171 	}
1172 	return NULL;
1173 }
1174 
1175 static void __intel_pmu_pebs_event(struct perf_event *event,
1176 				   struct pt_regs *iregs,
1177 				   void *base, void *top,
1178 				   int bit, int count)
1179 {
1180 	struct perf_sample_data data;
1181 	struct pt_regs regs;
1182 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1183 
1184 	if (!intel_pmu_save_and_restart(event) &&
1185 	    !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1186 		return;
1187 
1188 	while (count > 1) {
1189 		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1190 		perf_event_output(event, &data, &regs);
1191 		at += x86_pmu.pebs_record_size;
1192 		at = get_next_pebs_record_by_bit(at, top, bit);
1193 		count--;
1194 	}
1195 
1196 	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1197 
1198 	/*
1199 	 * All but the last records are processed.
1200 	 * The last one is left to be able to call the overflow handler.
1201 	 */
1202 	if (perf_event_overflow(event, &data, &regs)) {
1203 		x86_pmu_stop(event, 0);
1204 		return;
1205 	}
1206 
1207 }
1208 
1209 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1210 {
1211 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1212 	struct debug_store *ds = cpuc->ds;
1213 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1214 	struct pebs_record_core *at, *top;
1215 	int n;
1216 
1217 	if (!x86_pmu.pebs_active)
1218 		return;
1219 
1220 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1221 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1222 
1223 	/*
1224 	 * Whatever else happens, drain the thing
1225 	 */
1226 	ds->pebs_index = ds->pebs_buffer_base;
1227 
1228 	if (!test_bit(0, cpuc->active_mask))
1229 		return;
1230 
1231 	WARN_ON_ONCE(!event);
1232 
1233 	if (!event->attr.precise_ip)
1234 		return;
1235 
1236 	n = top - at;
1237 	if (n <= 0)
1238 		return;
1239 
1240 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1241 }
1242 
1243 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1244 {
1245 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1246 	struct debug_store *ds = cpuc->ds;
1247 	struct perf_event *event;
1248 	void *base, *at, *top;
1249 	short counts[MAX_PEBS_EVENTS] = {};
1250 	short error[MAX_PEBS_EVENTS] = {};
1251 	int bit, i;
1252 
1253 	if (!x86_pmu.pebs_active)
1254 		return;
1255 
1256 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1257 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1258 
1259 	ds->pebs_index = ds->pebs_buffer_base;
1260 
1261 	if (unlikely(base >= top))
1262 		return;
1263 
1264 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1265 		struct pebs_record_nhm *p = at;
1266 		u64 pebs_status;
1267 
1268 		/* PEBS v3 has accurate status bits */
1269 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1270 			for_each_set_bit(bit, (unsigned long *)&p->status,
1271 					 MAX_PEBS_EVENTS)
1272 				counts[bit]++;
1273 
1274 			continue;
1275 		}
1276 
1277 		pebs_status = p->status & cpuc->pebs_enabled;
1278 		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1279 
1280 		/*
1281 		 * On some CPUs the PEBS status can be zero when PEBS is
1282 		 * racing with clearing of GLOBAL_STATUS.
1283 		 *
1284 		 * Normally we would drop that record, but in the
1285 		 * case when there is only a single active PEBS event
1286 		 * we can assume it's for that event.
1287 		 */
1288 		if (!pebs_status && cpuc->pebs_enabled &&
1289 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1290 			pebs_status = cpuc->pebs_enabled;
1291 
1292 		bit = find_first_bit((unsigned long *)&pebs_status,
1293 					x86_pmu.max_pebs_events);
1294 		if (bit >= x86_pmu.max_pebs_events)
1295 			continue;
1296 
1297 		/*
1298 		 * The PEBS hardware does not deal well with the situation
1299 		 * when events happen near to each other and multiple bits
1300 		 * are set. But it should happen rarely.
1301 		 *
1302 		 * If these events include one PEBS and multiple non-PEBS
1303 		 * events, it doesn't impact PEBS record. The record will
1304 		 * be handled normally. (slow path)
1305 		 *
1306 		 * If these events include two or more PEBS events, the
1307 		 * records for the events can be collapsed into a single
1308 		 * one, and it's not possible to reconstruct all events
1309 		 * that caused the PEBS record. It's called collision.
1310 		 * If collision happened, the record will be dropped.
1311 		 */
1312 		if (p->status != (1ULL << bit)) {
1313 			for_each_set_bit(i, (unsigned long *)&pebs_status,
1314 					 x86_pmu.max_pebs_events)
1315 				error[i]++;
1316 			continue;
1317 		}
1318 
1319 		counts[bit]++;
1320 	}
1321 
1322 	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1323 		if ((counts[bit] == 0) && (error[bit] == 0))
1324 			continue;
1325 
1326 		event = cpuc->events[bit];
1327 		WARN_ON_ONCE(!event);
1328 		WARN_ON_ONCE(!event->attr.precise_ip);
1329 
1330 		/* log dropped samples number */
1331 		if (error[bit])
1332 			perf_log_lost_samples(event, error[bit]);
1333 
1334 		if (counts[bit]) {
1335 			__intel_pmu_pebs_event(event, iregs, base,
1336 					       top, bit, counts[bit]);
1337 		}
1338 	}
1339 }
1340 
1341 /*
1342  * BTS, PEBS probe and setup
1343  */
1344 
1345 void __init intel_ds_init(void)
1346 {
1347 	/*
1348 	 * No support for 32bit formats
1349 	 */
1350 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1351 		return;
1352 
1353 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1354 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1355 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1356 	if (x86_pmu.pebs) {
1357 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1358 		int format = x86_pmu.intel_cap.pebs_format;
1359 
1360 		switch (format) {
1361 		case 0:
1362 			pr_cont("PEBS fmt0%c, ", pebs_type);
1363 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1364 			/*
1365 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1366 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1367 			 * mysteriously hang on Core2.
1368 			 *
1369 			 * As a workaround, we don't do this.
1370 			 */
1371 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1372 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1373 			break;
1374 
1375 		case 1:
1376 			pr_cont("PEBS fmt1%c, ", pebs_type);
1377 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1378 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1379 			break;
1380 
1381 		case 2:
1382 			pr_cont("PEBS fmt2%c, ", pebs_type);
1383 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1384 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1385 			break;
1386 
1387 		case 3:
1388 			pr_cont("PEBS fmt3%c, ", pebs_type);
1389 			x86_pmu.pebs_record_size =
1390 						sizeof(struct pebs_record_skl);
1391 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1392 			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1393 			break;
1394 
1395 		default:
1396 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1397 			x86_pmu.pebs = 0;
1398 		}
1399 	}
1400 }
1401 
1402 void perf_restore_debug_store(void)
1403 {
1404 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1405 
1406 	if (!x86_pmu.bts && !x86_pmu.pebs)
1407 		return;
1408 
1409 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1410 }
1411