xref: /openbmc/linux/arch/x86/events/intel/ds.c (revision 5a244f48)
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 LEVEL(x) P(LVLNUM, x)
53 #define REM P(REMOTE, REMOTE)
54 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
55 
56 /* Version for Sandy Bridge and later */
57 static u64 pebs_data_source[] = {
58 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
59 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
60 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
61 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
62 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
63 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
64 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
65 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
66 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
67 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
68 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
69 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
70 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
71 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
72 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
73 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
74 };
75 
76 /* Patch up minor differences in the bits */
77 void __init intel_pmu_pebs_data_source_nhm(void)
78 {
79 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
80 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
81 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
82 }
83 
84 void __init intel_pmu_pebs_data_source_skl(bool pmem)
85 {
86 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
87 
88 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
89 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
90 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
91 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
92 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
93 }
94 
95 static u64 precise_store_data(u64 status)
96 {
97 	union intel_x86_pebs_dse dse;
98 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
99 
100 	dse.val = status;
101 
102 	/*
103 	 * bit 4: TLB access
104 	 * 1 = stored missed 2nd level TLB
105 	 *
106 	 * so it either hit the walker or the OS
107 	 * otherwise hit 2nd level TLB
108 	 */
109 	if (dse.st_stlb_miss)
110 		val |= P(TLB, MISS);
111 	else
112 		val |= P(TLB, HIT);
113 
114 	/*
115 	 * bit 0: hit L1 data cache
116 	 * if not set, then all we know is that
117 	 * it missed L1D
118 	 */
119 	if (dse.st_l1d_hit)
120 		val |= P(LVL, HIT);
121 	else
122 		val |= P(LVL, MISS);
123 
124 	/*
125 	 * bit 5: Locked prefix
126 	 */
127 	if (dse.st_locked)
128 		val |= P(LOCK, LOCKED);
129 
130 	return val;
131 }
132 
133 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
134 {
135 	union perf_mem_data_src dse;
136 
137 	dse.val = PERF_MEM_NA;
138 
139 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
140 		dse.mem_op = PERF_MEM_OP_STORE;
141 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
142 		dse.mem_op = PERF_MEM_OP_LOAD;
143 
144 	/*
145 	 * L1 info only valid for following events:
146 	 *
147 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
148 	 * MEM_UOPS_RETIRED.LOCK_STORES
149 	 * MEM_UOPS_RETIRED.SPLIT_STORES
150 	 * MEM_UOPS_RETIRED.ALL_STORES
151 	 */
152 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
153 		if (status & 1)
154 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
155 		else
156 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
157 	}
158 	return dse.val;
159 }
160 
161 static u64 load_latency_data(u64 status)
162 {
163 	union intel_x86_pebs_dse dse;
164 	u64 val;
165 
166 	dse.val = status;
167 
168 	/*
169 	 * use the mapping table for bit 0-3
170 	 */
171 	val = pebs_data_source[dse.ld_dse];
172 
173 	/*
174 	 * Nehalem models do not support TLB, Lock infos
175 	 */
176 	if (x86_pmu.pebs_no_tlb) {
177 		val |= P(TLB, NA) | P(LOCK, NA);
178 		return val;
179 	}
180 	/*
181 	 * bit 4: TLB access
182 	 * 0 = did not miss 2nd level TLB
183 	 * 1 = missed 2nd level TLB
184 	 */
185 	if (dse.ld_stlb_miss)
186 		val |= P(TLB, MISS) | P(TLB, L2);
187 	else
188 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
189 
190 	/*
191 	 * bit 5: locked prefix
192 	 */
193 	if (dse.ld_locked)
194 		val |= P(LOCK, LOCKED);
195 
196 	return val;
197 }
198 
199 struct pebs_record_core {
200 	u64 flags, ip;
201 	u64 ax, bx, cx, dx;
202 	u64 si, di, bp, sp;
203 	u64 r8,  r9,  r10, r11;
204 	u64 r12, r13, r14, r15;
205 };
206 
207 struct pebs_record_nhm {
208 	u64 flags, ip;
209 	u64 ax, bx, cx, dx;
210 	u64 si, di, bp, sp;
211 	u64 r8,  r9,  r10, r11;
212 	u64 r12, r13, r14, r15;
213 	u64 status, dla, dse, lat;
214 };
215 
216 /*
217  * Same as pebs_record_nhm, with two additional fields.
218  */
219 struct pebs_record_hsw {
220 	u64 flags, ip;
221 	u64 ax, bx, cx, dx;
222 	u64 si, di, bp, sp;
223 	u64 r8,  r9,  r10, r11;
224 	u64 r12, r13, r14, r15;
225 	u64 status, dla, dse, lat;
226 	u64 real_ip, tsx_tuning;
227 };
228 
229 union hsw_tsx_tuning {
230 	struct {
231 		u32 cycles_last_block     : 32,
232 		    hle_abort		  : 1,
233 		    rtm_abort		  : 1,
234 		    instruction_abort     : 1,
235 		    non_instruction_abort : 1,
236 		    retry		  : 1,
237 		    data_conflict	  : 1,
238 		    capacity_writes	  : 1,
239 		    capacity_reads	  : 1;
240 	};
241 	u64	    value;
242 };
243 
244 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
245 
246 /* Same as HSW, plus TSC */
247 
248 struct pebs_record_skl {
249 	u64 flags, ip;
250 	u64 ax, bx, cx, dx;
251 	u64 si, di, bp, sp;
252 	u64 r8,  r9,  r10, r11;
253 	u64 r12, r13, r14, r15;
254 	u64 status, dla, dse, lat;
255 	u64 real_ip, tsx_tuning;
256 	u64 tsc;
257 };
258 
259 void init_debug_store_on_cpu(int cpu)
260 {
261 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
262 
263 	if (!ds)
264 		return;
265 
266 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
267 		     (u32)((u64)(unsigned long)ds),
268 		     (u32)((u64)(unsigned long)ds >> 32));
269 }
270 
271 void fini_debug_store_on_cpu(int cpu)
272 {
273 	if (!per_cpu(cpu_hw_events, cpu).ds)
274 		return;
275 
276 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
277 }
278 
279 static DEFINE_PER_CPU(void *, insn_buffer);
280 
281 static int alloc_pebs_buffer(int cpu)
282 {
283 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
284 	int node = cpu_to_node(cpu);
285 	int max;
286 	void *buffer, *ibuffer;
287 
288 	if (!x86_pmu.pebs)
289 		return 0;
290 
291 	buffer = kzalloc_node(x86_pmu.pebs_buffer_size, GFP_KERNEL, node);
292 	if (unlikely(!buffer))
293 		return -ENOMEM;
294 
295 	/*
296 	 * HSW+ already provides us the eventing ip; no need to allocate this
297 	 * buffer then.
298 	 */
299 	if (x86_pmu.intel_cap.pebs_format < 2) {
300 		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
301 		if (!ibuffer) {
302 			kfree(buffer);
303 			return -ENOMEM;
304 		}
305 		per_cpu(insn_buffer, cpu) = ibuffer;
306 	}
307 
308 	max = x86_pmu.pebs_buffer_size / x86_pmu.pebs_record_size;
309 
310 	ds->pebs_buffer_base = (u64)(unsigned long)buffer;
311 	ds->pebs_index = ds->pebs_buffer_base;
312 	ds->pebs_absolute_maximum = ds->pebs_buffer_base +
313 		max * x86_pmu.pebs_record_size;
314 
315 	return 0;
316 }
317 
318 static void release_pebs_buffer(int cpu)
319 {
320 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
321 
322 	if (!ds || !x86_pmu.pebs)
323 		return;
324 
325 	kfree(per_cpu(insn_buffer, cpu));
326 	per_cpu(insn_buffer, cpu) = NULL;
327 
328 	kfree((void *)(unsigned long)ds->pebs_buffer_base);
329 	ds->pebs_buffer_base = 0;
330 }
331 
332 static int alloc_bts_buffer(int cpu)
333 {
334 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
335 	int node = cpu_to_node(cpu);
336 	int max, thresh;
337 	void *buffer;
338 
339 	if (!x86_pmu.bts)
340 		return 0;
341 
342 	buffer = kzalloc_node(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, node);
343 	if (unlikely(!buffer)) {
344 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
345 		return -ENOMEM;
346 	}
347 
348 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
349 	thresh = max / 16;
350 
351 	ds->bts_buffer_base = (u64)(unsigned long)buffer;
352 	ds->bts_index = ds->bts_buffer_base;
353 	ds->bts_absolute_maximum = ds->bts_buffer_base +
354 		max * BTS_RECORD_SIZE;
355 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
356 		thresh * BTS_RECORD_SIZE;
357 
358 	return 0;
359 }
360 
361 static void release_bts_buffer(int cpu)
362 {
363 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
364 
365 	if (!ds || !x86_pmu.bts)
366 		return;
367 
368 	kfree((void *)(unsigned long)ds->bts_buffer_base);
369 	ds->bts_buffer_base = 0;
370 }
371 
372 static int alloc_ds_buffer(int cpu)
373 {
374 	int node = cpu_to_node(cpu);
375 	struct debug_store *ds;
376 
377 	ds = kzalloc_node(sizeof(*ds), GFP_KERNEL, node);
378 	if (unlikely(!ds))
379 		return -ENOMEM;
380 
381 	per_cpu(cpu_hw_events, cpu).ds = ds;
382 
383 	return 0;
384 }
385 
386 static void release_ds_buffer(int cpu)
387 {
388 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
389 
390 	if (!ds)
391 		return;
392 
393 	per_cpu(cpu_hw_events, cpu).ds = NULL;
394 	kfree(ds);
395 }
396 
397 void release_ds_buffers(void)
398 {
399 	int cpu;
400 
401 	if (!x86_pmu.bts && !x86_pmu.pebs)
402 		return;
403 
404 	get_online_cpus();
405 	for_each_online_cpu(cpu)
406 		fini_debug_store_on_cpu(cpu);
407 
408 	for_each_possible_cpu(cpu) {
409 		release_pebs_buffer(cpu);
410 		release_bts_buffer(cpu);
411 		release_ds_buffer(cpu);
412 	}
413 	put_online_cpus();
414 }
415 
416 void reserve_ds_buffers(void)
417 {
418 	int bts_err = 0, pebs_err = 0;
419 	int cpu;
420 
421 	x86_pmu.bts_active = 0;
422 	x86_pmu.pebs_active = 0;
423 
424 	if (!x86_pmu.bts && !x86_pmu.pebs)
425 		return;
426 
427 	if (!x86_pmu.bts)
428 		bts_err = 1;
429 
430 	if (!x86_pmu.pebs)
431 		pebs_err = 1;
432 
433 	get_online_cpus();
434 
435 	for_each_possible_cpu(cpu) {
436 		if (alloc_ds_buffer(cpu)) {
437 			bts_err = 1;
438 			pebs_err = 1;
439 		}
440 
441 		if (!bts_err && alloc_bts_buffer(cpu))
442 			bts_err = 1;
443 
444 		if (!pebs_err && alloc_pebs_buffer(cpu))
445 			pebs_err = 1;
446 
447 		if (bts_err && pebs_err)
448 			break;
449 	}
450 
451 	if (bts_err) {
452 		for_each_possible_cpu(cpu)
453 			release_bts_buffer(cpu);
454 	}
455 
456 	if (pebs_err) {
457 		for_each_possible_cpu(cpu)
458 			release_pebs_buffer(cpu);
459 	}
460 
461 	if (bts_err && pebs_err) {
462 		for_each_possible_cpu(cpu)
463 			release_ds_buffer(cpu);
464 	} else {
465 		if (x86_pmu.bts && !bts_err)
466 			x86_pmu.bts_active = 1;
467 
468 		if (x86_pmu.pebs && !pebs_err)
469 			x86_pmu.pebs_active = 1;
470 
471 		for_each_online_cpu(cpu)
472 			init_debug_store_on_cpu(cpu);
473 	}
474 
475 	put_online_cpus();
476 }
477 
478 /*
479  * BTS
480  */
481 
482 struct event_constraint bts_constraint =
483 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
484 
485 void intel_pmu_enable_bts(u64 config)
486 {
487 	unsigned long debugctlmsr;
488 
489 	debugctlmsr = get_debugctlmsr();
490 
491 	debugctlmsr |= DEBUGCTLMSR_TR;
492 	debugctlmsr |= DEBUGCTLMSR_BTS;
493 	if (config & ARCH_PERFMON_EVENTSEL_INT)
494 		debugctlmsr |= DEBUGCTLMSR_BTINT;
495 
496 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
497 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
498 
499 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
500 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
501 
502 	update_debugctlmsr(debugctlmsr);
503 }
504 
505 void intel_pmu_disable_bts(void)
506 {
507 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
508 	unsigned long debugctlmsr;
509 
510 	if (!cpuc->ds)
511 		return;
512 
513 	debugctlmsr = get_debugctlmsr();
514 
515 	debugctlmsr &=
516 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
517 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
518 
519 	update_debugctlmsr(debugctlmsr);
520 }
521 
522 int intel_pmu_drain_bts_buffer(void)
523 {
524 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
525 	struct debug_store *ds = cpuc->ds;
526 	struct bts_record {
527 		u64	from;
528 		u64	to;
529 		u64	flags;
530 	};
531 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
532 	struct bts_record *at, *base, *top;
533 	struct perf_output_handle handle;
534 	struct perf_event_header header;
535 	struct perf_sample_data data;
536 	unsigned long skip = 0;
537 	struct pt_regs regs;
538 
539 	if (!event)
540 		return 0;
541 
542 	if (!x86_pmu.bts_active)
543 		return 0;
544 
545 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
546 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
547 
548 	if (top <= base)
549 		return 0;
550 
551 	memset(&regs, 0, sizeof(regs));
552 
553 	ds->bts_index = ds->bts_buffer_base;
554 
555 	perf_sample_data_init(&data, 0, event->hw.last_period);
556 
557 	/*
558 	 * BTS leaks kernel addresses in branches across the cpl boundary,
559 	 * such as traps or system calls, so unless the user is asking for
560 	 * kernel tracing (and right now it's not possible), we'd need to
561 	 * filter them out. But first we need to count how many of those we
562 	 * have in the current batch. This is an extra O(n) pass, however,
563 	 * it's much faster than the other one especially considering that
564 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
565 	 * alloc_bts_buffer()).
566 	 */
567 	for (at = base; at < top; at++) {
568 		/*
569 		 * Note that right now *this* BTS code only works if
570 		 * attr::exclude_kernel is set, but let's keep this extra
571 		 * check here in case that changes.
572 		 */
573 		if (event->attr.exclude_kernel &&
574 		    (kernel_ip(at->from) || kernel_ip(at->to)))
575 			skip++;
576 	}
577 
578 	/*
579 	 * Prepare a generic sample, i.e. fill in the invariant fields.
580 	 * We will overwrite the from and to address before we output
581 	 * the sample.
582 	 */
583 	rcu_read_lock();
584 	perf_prepare_sample(&header, &data, event, &regs);
585 
586 	if (perf_output_begin(&handle, event, header.size *
587 			      (top - base - skip)))
588 		goto unlock;
589 
590 	for (at = base; at < top; at++) {
591 		/* Filter out any records that contain kernel addresses. */
592 		if (event->attr.exclude_kernel &&
593 		    (kernel_ip(at->from) || kernel_ip(at->to)))
594 			continue;
595 
596 		data.ip		= at->from;
597 		data.addr	= at->to;
598 
599 		perf_output_sample(&handle, &header, &data, event);
600 	}
601 
602 	perf_output_end(&handle);
603 
604 	/* There's new data available. */
605 	event->hw.interrupts++;
606 	event->pending_kill = POLL_IN;
607 unlock:
608 	rcu_read_unlock();
609 	return 1;
610 }
611 
612 static inline void intel_pmu_drain_pebs_buffer(void)
613 {
614 	struct pt_regs regs;
615 
616 	x86_pmu.drain_pebs(&regs);
617 }
618 
619 /*
620  * PEBS
621  */
622 struct event_constraint intel_core2_pebs_event_constraints[] = {
623 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
624 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
625 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
626 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
627 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
628 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
629 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
630 	EVENT_CONSTRAINT_END
631 };
632 
633 struct event_constraint intel_atom_pebs_event_constraints[] = {
634 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
635 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
636 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
637 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
638 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
639 	/* Allow all events as PEBS with no flags */
640 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
641 	EVENT_CONSTRAINT_END
642 };
643 
644 struct event_constraint intel_slm_pebs_event_constraints[] = {
645 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
646 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
647 	/* Allow all events as PEBS with no flags */
648 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
649 	EVENT_CONSTRAINT_END
650 };
651 
652 struct event_constraint intel_glm_pebs_event_constraints[] = {
653 	/* Allow all events as PEBS with no flags */
654 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
655 	EVENT_CONSTRAINT_END
656 };
657 
658 struct event_constraint intel_glp_pebs_event_constraints[] = {
659 	/* Allow all events as PEBS with no flags */
660 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
661 	EVENT_CONSTRAINT_END
662 };
663 
664 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
665 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
666 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
667 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
668 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
669 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
670 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
671 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
672 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
673 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
674 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
675 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
676 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
677 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
678 	EVENT_CONSTRAINT_END
679 };
680 
681 struct event_constraint intel_westmere_pebs_event_constraints[] = {
682 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
683 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
684 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
685 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
686 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
687 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
688 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
689 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
690 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
691 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
692 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
693 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
694 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
695 	EVENT_CONSTRAINT_END
696 };
697 
698 struct event_constraint intel_snb_pebs_event_constraints[] = {
699 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
700 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
701 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
702 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
703 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
704         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
705         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
706         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
707         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
708 	/* Allow all events as PEBS with no flags */
709 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
710 	EVENT_CONSTRAINT_END
711 };
712 
713 struct event_constraint intel_ivb_pebs_event_constraints[] = {
714         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
715         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
716 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
717 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
718 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
719 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
720 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
721 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
722 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
723 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
724 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
725 	/* Allow all events as PEBS with no flags */
726 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
727         EVENT_CONSTRAINT_END
728 };
729 
730 struct event_constraint intel_hsw_pebs_event_constraints[] = {
731 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
732 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
733 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
734 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
735 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
736 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
737 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
738 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
739 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
740 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
741 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
742 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
743 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
744 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
745 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
746 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
747 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
748 	/* Allow all events as PEBS with no flags */
749 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
750 	EVENT_CONSTRAINT_END
751 };
752 
753 struct event_constraint intel_bdw_pebs_event_constraints[] = {
754 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
755 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
756 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
757 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
758 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
759 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
760 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
761 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
762 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
763 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
764 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
765 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
766 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
767 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
768 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
769 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
770 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
771 	/* Allow all events as PEBS with no flags */
772 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
773 	EVENT_CONSTRAINT_END
774 };
775 
776 
777 struct event_constraint intel_skl_pebs_event_constraints[] = {
778 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
779 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
780 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
781 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
782 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
783 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
784 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
785 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
786 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
787 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
788 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
789 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
790 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
791 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
792 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
793 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
794 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
795 	/* Allow all events as PEBS with no flags */
796 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
797 	EVENT_CONSTRAINT_END
798 };
799 
800 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
801 {
802 	struct event_constraint *c;
803 
804 	if (!event->attr.precise_ip)
805 		return NULL;
806 
807 	if (x86_pmu.pebs_constraints) {
808 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
809 			if ((event->hw.config & c->cmask) == c->code) {
810 				event->hw.flags |= c->flags;
811 				return c;
812 			}
813 		}
814 	}
815 
816 	return &emptyconstraint;
817 }
818 
819 /*
820  * We need the sched_task callback even for per-cpu events when we use
821  * the large interrupt threshold, such that we can provide PID and TID
822  * to PEBS samples.
823  */
824 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
825 {
826 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
827 }
828 
829 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
830 {
831 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
832 
833 	if (!sched_in && pebs_needs_sched_cb(cpuc))
834 		intel_pmu_drain_pebs_buffer();
835 }
836 
837 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
838 {
839 	struct debug_store *ds = cpuc->ds;
840 	u64 threshold;
841 
842 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
843 		threshold = ds->pebs_absolute_maximum -
844 			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
845 	} else {
846 		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
847 	}
848 
849 	ds->pebs_interrupt_threshold = threshold;
850 }
851 
852 static void
853 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
854 {
855 	/*
856 	 * Make sure we get updated with the first PEBS
857 	 * event. It will trigger also during removal, but
858 	 * that does not hurt:
859 	 */
860 	bool update = cpuc->n_pebs == 1;
861 
862 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
863 		if (!needed_cb)
864 			perf_sched_cb_inc(pmu);
865 		else
866 			perf_sched_cb_dec(pmu);
867 
868 		update = true;
869 	}
870 
871 	if (update)
872 		pebs_update_threshold(cpuc);
873 }
874 
875 void intel_pmu_pebs_add(struct perf_event *event)
876 {
877 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
878 	struct hw_perf_event *hwc = &event->hw;
879 	bool needed_cb = pebs_needs_sched_cb(cpuc);
880 
881 	cpuc->n_pebs++;
882 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
883 		cpuc->n_large_pebs++;
884 
885 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
886 }
887 
888 void intel_pmu_pebs_enable(struct perf_event *event)
889 {
890 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
891 	struct hw_perf_event *hwc = &event->hw;
892 	struct debug_store *ds = cpuc->ds;
893 
894 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
895 
896 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
897 
898 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
899 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
900 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
901 		cpuc->pebs_enabled |= 1ULL << 63;
902 
903 	/*
904 	 * Use auto-reload if possible to save a MSR write in the PMI.
905 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
906 	 */
907 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
908 		ds->pebs_event_reset[hwc->idx] =
909 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
910 	} else {
911 		ds->pebs_event_reset[hwc->idx] = 0;
912 	}
913 }
914 
915 void intel_pmu_pebs_del(struct perf_event *event)
916 {
917 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
918 	struct hw_perf_event *hwc = &event->hw;
919 	bool needed_cb = pebs_needs_sched_cb(cpuc);
920 
921 	cpuc->n_pebs--;
922 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
923 		cpuc->n_large_pebs--;
924 
925 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
926 }
927 
928 void intel_pmu_pebs_disable(struct perf_event *event)
929 {
930 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
931 	struct hw_perf_event *hwc = &event->hw;
932 
933 	if (cpuc->n_pebs == cpuc->n_large_pebs)
934 		intel_pmu_drain_pebs_buffer();
935 
936 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
937 
938 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
939 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
940 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
941 		cpuc->pebs_enabled &= ~(1ULL << 63);
942 
943 	if (cpuc->enabled)
944 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
945 
946 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
947 }
948 
949 void intel_pmu_pebs_enable_all(void)
950 {
951 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
952 
953 	if (cpuc->pebs_enabled)
954 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
955 }
956 
957 void intel_pmu_pebs_disable_all(void)
958 {
959 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
960 
961 	if (cpuc->pebs_enabled)
962 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
963 }
964 
965 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
966 {
967 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
968 	unsigned long from = cpuc->lbr_entries[0].from;
969 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
970 	unsigned long ip = regs->ip;
971 	int is_64bit = 0;
972 	void *kaddr;
973 	int size;
974 
975 	/*
976 	 * We don't need to fixup if the PEBS assist is fault like
977 	 */
978 	if (!x86_pmu.intel_cap.pebs_trap)
979 		return 1;
980 
981 	/*
982 	 * No LBR entry, no basic block, no rewinding
983 	 */
984 	if (!cpuc->lbr_stack.nr || !from || !to)
985 		return 0;
986 
987 	/*
988 	 * Basic blocks should never cross user/kernel boundaries
989 	 */
990 	if (kernel_ip(ip) != kernel_ip(to))
991 		return 0;
992 
993 	/*
994 	 * unsigned math, either ip is before the start (impossible) or
995 	 * the basic block is larger than 1 page (sanity)
996 	 */
997 	if ((ip - to) > PEBS_FIXUP_SIZE)
998 		return 0;
999 
1000 	/*
1001 	 * We sampled a branch insn, rewind using the LBR stack
1002 	 */
1003 	if (ip == to) {
1004 		set_linear_ip(regs, from);
1005 		return 1;
1006 	}
1007 
1008 	size = ip - to;
1009 	if (!kernel_ip(ip)) {
1010 		int bytes;
1011 		u8 *buf = this_cpu_read(insn_buffer);
1012 
1013 		/* 'size' must fit our buffer, see above */
1014 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1015 		if (bytes != 0)
1016 			return 0;
1017 
1018 		kaddr = buf;
1019 	} else {
1020 		kaddr = (void *)to;
1021 	}
1022 
1023 	do {
1024 		struct insn insn;
1025 
1026 		old_to = to;
1027 
1028 #ifdef CONFIG_X86_64
1029 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1030 #endif
1031 		insn_init(&insn, kaddr, size, is_64bit);
1032 		insn_get_length(&insn);
1033 		/*
1034 		 * Make sure there was not a problem decoding the
1035 		 * instruction and getting the length.  This is
1036 		 * doubly important because we have an infinite
1037 		 * loop if insn.length=0.
1038 		 */
1039 		if (!insn.length)
1040 			break;
1041 
1042 		to += insn.length;
1043 		kaddr += insn.length;
1044 		size -= insn.length;
1045 	} while (to < ip);
1046 
1047 	if (to == ip) {
1048 		set_linear_ip(regs, old_to);
1049 		return 1;
1050 	}
1051 
1052 	/*
1053 	 * Even though we decoded the basic block, the instruction stream
1054 	 * never matched the given IP, either the TO or the IP got corrupted.
1055 	 */
1056 	return 0;
1057 }
1058 
1059 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1060 {
1061 	if (pebs->tsx_tuning) {
1062 		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1063 		return tsx.cycles_last_block;
1064 	}
1065 	return 0;
1066 }
1067 
1068 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1069 {
1070 	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1071 
1072 	/* For RTM XABORTs also log the abort code from AX */
1073 	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1074 		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1075 	return txn;
1076 }
1077 
1078 static void setup_pebs_sample_data(struct perf_event *event,
1079 				   struct pt_regs *iregs, void *__pebs,
1080 				   struct perf_sample_data *data,
1081 				   struct pt_regs *regs)
1082 {
1083 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1084 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1085 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1086 		 PERF_X86_EVENT_PEBS_NA_HSW)
1087 	/*
1088 	 * We cast to the biggest pebs_record but are careful not to
1089 	 * unconditionally access the 'extra' entries.
1090 	 */
1091 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1092 	struct pebs_record_skl *pebs = __pebs;
1093 	u64 sample_type;
1094 	int fll, fst, dsrc;
1095 	int fl = event->hw.flags;
1096 
1097 	if (pebs == NULL)
1098 		return;
1099 
1100 	sample_type = event->attr.sample_type;
1101 	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1102 
1103 	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1104 	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1105 
1106 	perf_sample_data_init(data, 0, event->hw.last_period);
1107 
1108 	data->period = event->hw.last_period;
1109 
1110 	/*
1111 	 * Use latency for weight (only avail with PEBS-LL)
1112 	 */
1113 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1114 		data->weight = pebs->lat;
1115 
1116 	/*
1117 	 * data.data_src encodes the data source
1118 	 */
1119 	if (dsrc) {
1120 		u64 val = PERF_MEM_NA;
1121 		if (fll)
1122 			val = load_latency_data(pebs->dse);
1123 		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1124 			val = precise_datala_hsw(event, pebs->dse);
1125 		else if (fst)
1126 			val = precise_store_data(pebs->dse);
1127 		data->data_src.val = val;
1128 	}
1129 
1130 	/*
1131 	 * We use the interrupt regs as a base because the PEBS record does not
1132 	 * contain a full regs set, specifically it seems to lack segment
1133 	 * descriptors, which get used by things like user_mode().
1134 	 *
1135 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1136 	 *
1137 	 * We must however always use BP,SP from iregs for the unwinder to stay
1138 	 * sane; the record BP,SP can point into thin air when the record is
1139 	 * from a previous PMI context or an (I)RET happend between the record
1140 	 * and PMI.
1141 	 */
1142 	*regs = *iregs;
1143 	regs->flags = pebs->flags;
1144 	set_linear_ip(regs, pebs->ip);
1145 
1146 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1147 		regs->ax = pebs->ax;
1148 		regs->bx = pebs->bx;
1149 		regs->cx = pebs->cx;
1150 		regs->dx = pebs->dx;
1151 		regs->si = pebs->si;
1152 		regs->di = pebs->di;
1153 
1154 		/*
1155 		 * Per the above; only set BP,SP if we don't need callchains.
1156 		 *
1157 		 * XXX: does this make sense?
1158 		 */
1159 		if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1160 			regs->bp = pebs->bp;
1161 			regs->sp = pebs->sp;
1162 		}
1163 
1164 		/*
1165 		 * Preserve PERF_EFLAGS_VM from set_linear_ip().
1166 		 */
1167 		regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1168 #ifndef CONFIG_X86_32
1169 		regs->r8 = pebs->r8;
1170 		regs->r9 = pebs->r9;
1171 		regs->r10 = pebs->r10;
1172 		regs->r11 = pebs->r11;
1173 		regs->r12 = pebs->r12;
1174 		regs->r13 = pebs->r13;
1175 		regs->r14 = pebs->r14;
1176 		regs->r15 = pebs->r15;
1177 #endif
1178 	}
1179 
1180 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1181 		regs->ip = pebs->real_ip;
1182 		regs->flags |= PERF_EFLAGS_EXACT;
1183 	} else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1184 		regs->flags |= PERF_EFLAGS_EXACT;
1185 	else
1186 		regs->flags &= ~PERF_EFLAGS_EXACT;
1187 
1188 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1189 	    x86_pmu.intel_cap.pebs_format >= 1)
1190 		data->addr = pebs->dla;
1191 
1192 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1193 		/* Only set the TSX weight when no memory weight. */
1194 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1195 			data->weight = intel_hsw_weight(pebs);
1196 
1197 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1198 			data->txn = intel_hsw_transaction(pebs);
1199 	}
1200 
1201 	/*
1202 	 * v3 supplies an accurate time stamp, so we use that
1203 	 * for the time stamp.
1204 	 *
1205 	 * We can only do this for the default trace clock.
1206 	 */
1207 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1208 		event->attr.use_clockid == 0)
1209 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1210 
1211 	if (has_branch_stack(event))
1212 		data->br_stack = &cpuc->lbr_stack;
1213 }
1214 
1215 static inline void *
1216 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1217 {
1218 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1219 	void *at;
1220 	u64 pebs_status;
1221 
1222 	/*
1223 	 * fmt0 does not have a status bitfield (does not use
1224 	 * perf_record_nhm format)
1225 	 */
1226 	if (x86_pmu.intel_cap.pebs_format < 1)
1227 		return base;
1228 
1229 	if (base == NULL)
1230 		return NULL;
1231 
1232 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1233 		struct pebs_record_nhm *p = at;
1234 
1235 		if (test_bit(bit, (unsigned long *)&p->status)) {
1236 			/* PEBS v3 has accurate status bits */
1237 			if (x86_pmu.intel_cap.pebs_format >= 3)
1238 				return at;
1239 
1240 			if (p->status == (1 << bit))
1241 				return at;
1242 
1243 			/* clear non-PEBS bit and re-check */
1244 			pebs_status = p->status & cpuc->pebs_enabled;
1245 			pebs_status &= PEBS_COUNTER_MASK;
1246 			if (pebs_status == (1 << bit))
1247 				return at;
1248 		}
1249 	}
1250 	return NULL;
1251 }
1252 
1253 static void __intel_pmu_pebs_event(struct perf_event *event,
1254 				   struct pt_regs *iregs,
1255 				   void *base, void *top,
1256 				   int bit, int count)
1257 {
1258 	struct perf_sample_data data;
1259 	struct pt_regs regs;
1260 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1261 
1262 	if (!intel_pmu_save_and_restart(event) &&
1263 	    !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1264 		return;
1265 
1266 	while (count > 1) {
1267 		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1268 		perf_event_output(event, &data, &regs);
1269 		at += x86_pmu.pebs_record_size;
1270 		at = get_next_pebs_record_by_bit(at, top, bit);
1271 		count--;
1272 	}
1273 
1274 	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1275 
1276 	/*
1277 	 * All but the last records are processed.
1278 	 * The last one is left to be able to call the overflow handler.
1279 	 */
1280 	if (perf_event_overflow(event, &data, &regs)) {
1281 		x86_pmu_stop(event, 0);
1282 		return;
1283 	}
1284 
1285 }
1286 
1287 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1288 {
1289 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1290 	struct debug_store *ds = cpuc->ds;
1291 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1292 	struct pebs_record_core *at, *top;
1293 	int n;
1294 
1295 	if (!x86_pmu.pebs_active)
1296 		return;
1297 
1298 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1299 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1300 
1301 	/*
1302 	 * Whatever else happens, drain the thing
1303 	 */
1304 	ds->pebs_index = ds->pebs_buffer_base;
1305 
1306 	if (!test_bit(0, cpuc->active_mask))
1307 		return;
1308 
1309 	WARN_ON_ONCE(!event);
1310 
1311 	if (!event->attr.precise_ip)
1312 		return;
1313 
1314 	n = top - at;
1315 	if (n <= 0)
1316 		return;
1317 
1318 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1319 }
1320 
1321 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1322 {
1323 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1324 	struct debug_store *ds = cpuc->ds;
1325 	struct perf_event *event;
1326 	void *base, *at, *top;
1327 	short counts[MAX_PEBS_EVENTS] = {};
1328 	short error[MAX_PEBS_EVENTS] = {};
1329 	int bit, i;
1330 
1331 	if (!x86_pmu.pebs_active)
1332 		return;
1333 
1334 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1335 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1336 
1337 	ds->pebs_index = ds->pebs_buffer_base;
1338 
1339 	if (unlikely(base >= top))
1340 		return;
1341 
1342 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1343 		struct pebs_record_nhm *p = at;
1344 		u64 pebs_status;
1345 
1346 		pebs_status = p->status & cpuc->pebs_enabled;
1347 		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1348 
1349 		/* PEBS v3 has more accurate status bits */
1350 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1351 			for_each_set_bit(bit, (unsigned long *)&pebs_status,
1352 					 x86_pmu.max_pebs_events)
1353 				counts[bit]++;
1354 
1355 			continue;
1356 		}
1357 
1358 		/*
1359 		 * On some CPUs the PEBS status can be zero when PEBS is
1360 		 * racing with clearing of GLOBAL_STATUS.
1361 		 *
1362 		 * Normally we would drop that record, but in the
1363 		 * case when there is only a single active PEBS event
1364 		 * we can assume it's for that event.
1365 		 */
1366 		if (!pebs_status && cpuc->pebs_enabled &&
1367 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1368 			pebs_status = cpuc->pebs_enabled;
1369 
1370 		bit = find_first_bit((unsigned long *)&pebs_status,
1371 					x86_pmu.max_pebs_events);
1372 		if (bit >= x86_pmu.max_pebs_events)
1373 			continue;
1374 
1375 		/*
1376 		 * The PEBS hardware does not deal well with the situation
1377 		 * when events happen near to each other and multiple bits
1378 		 * are set. But it should happen rarely.
1379 		 *
1380 		 * If these events include one PEBS and multiple non-PEBS
1381 		 * events, it doesn't impact PEBS record. The record will
1382 		 * be handled normally. (slow path)
1383 		 *
1384 		 * If these events include two or more PEBS events, the
1385 		 * records for the events can be collapsed into a single
1386 		 * one, and it's not possible to reconstruct all events
1387 		 * that caused the PEBS record. It's called collision.
1388 		 * If collision happened, the record will be dropped.
1389 		 */
1390 		if (p->status != (1ULL << bit)) {
1391 			for_each_set_bit(i, (unsigned long *)&pebs_status,
1392 					 x86_pmu.max_pebs_events)
1393 				error[i]++;
1394 			continue;
1395 		}
1396 
1397 		counts[bit]++;
1398 	}
1399 
1400 	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1401 		if ((counts[bit] == 0) && (error[bit] == 0))
1402 			continue;
1403 
1404 		event = cpuc->events[bit];
1405 		if (WARN_ON_ONCE(!event))
1406 			continue;
1407 
1408 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1409 			continue;
1410 
1411 		/* log dropped samples number */
1412 		if (error[bit]) {
1413 			perf_log_lost_samples(event, error[bit]);
1414 
1415 			if (perf_event_account_interrupt(event))
1416 				x86_pmu_stop(event, 0);
1417 		}
1418 
1419 		if (counts[bit]) {
1420 			__intel_pmu_pebs_event(event, iregs, base,
1421 					       top, bit, counts[bit]);
1422 		}
1423 	}
1424 }
1425 
1426 /*
1427  * BTS, PEBS probe and setup
1428  */
1429 
1430 void __init intel_ds_init(void)
1431 {
1432 	/*
1433 	 * No support for 32bit formats
1434 	 */
1435 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1436 		return;
1437 
1438 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1439 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1440 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1441 	if (x86_pmu.pebs) {
1442 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1443 		int format = x86_pmu.intel_cap.pebs_format;
1444 
1445 		switch (format) {
1446 		case 0:
1447 			pr_cont("PEBS fmt0%c, ", pebs_type);
1448 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1449 			/*
1450 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1451 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1452 			 * mysteriously hang on Core2.
1453 			 *
1454 			 * As a workaround, we don't do this.
1455 			 */
1456 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1457 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1458 			break;
1459 
1460 		case 1:
1461 			pr_cont("PEBS fmt1%c, ", pebs_type);
1462 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1463 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1464 			break;
1465 
1466 		case 2:
1467 			pr_cont("PEBS fmt2%c, ", pebs_type);
1468 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1469 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1470 			break;
1471 
1472 		case 3:
1473 			pr_cont("PEBS fmt3%c, ", pebs_type);
1474 			x86_pmu.pebs_record_size =
1475 						sizeof(struct pebs_record_skl);
1476 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1477 			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1478 			break;
1479 
1480 		default:
1481 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1482 			x86_pmu.pebs = 0;
1483 		}
1484 	}
1485 }
1486 
1487 void perf_restore_debug_store(void)
1488 {
1489 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1490 
1491 	if (!x86_pmu.bts && !x86_pmu.pebs)
1492 		return;
1493 
1494 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1495 }
1496