xref: /openbmc/linux/arch/x86/events/intel/ds.c (revision 6724ed7f)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/types.h>
4 #include <linux/slab.h>
5 
6 #include <asm/cpu_entry_area.h>
7 #include <asm/perf_event.h>
8 #include <asm/insn.h>
9 
10 #include "../perf_event.h"
11 
12 /* Waste a full page so it can be mapped into the cpu_entry_area */
13 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
14 
15 /* The size of a BTS record in bytes: */
16 #define BTS_RECORD_SIZE		24
17 
18 #define PEBS_FIXUP_SIZE		PAGE_SIZE
19 
20 /*
21  * pebs_record_32 for p4 and core not supported
22 
23 struct pebs_record_32 {
24 	u32 flags, ip;
25 	u32 ax, bc, cx, dx;
26 	u32 si, di, bp, sp;
27 };
28 
29  */
30 
31 union intel_x86_pebs_dse {
32 	u64 val;
33 	struct {
34 		unsigned int ld_dse:4;
35 		unsigned int ld_stlb_miss:1;
36 		unsigned int ld_locked:1;
37 		unsigned int ld_reserved:26;
38 	};
39 	struct {
40 		unsigned int st_l1d_hit:1;
41 		unsigned int st_reserved1:3;
42 		unsigned int st_stlb_miss:1;
43 		unsigned int st_locked:1;
44 		unsigned int st_reserved2:26;
45 	};
46 };
47 
48 
49 /*
50  * Map PEBS Load Latency Data Source encodings to generic
51  * memory data source information
52  */
53 #define P(a, b) PERF_MEM_S(a, b)
54 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
55 #define LEVEL(x) P(LVLNUM, x)
56 #define REM P(REMOTE, REMOTE)
57 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
58 
59 /* Version for Sandy Bridge and later */
60 static u64 pebs_data_source[] = {
61 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
62 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
63 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
64 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
65 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
66 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
67 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
68 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
69 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
70 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
71 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
72 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
73 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
74 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
75 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
76 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
77 };
78 
79 /* Patch up minor differences in the bits */
80 void __init intel_pmu_pebs_data_source_nhm(void)
81 {
82 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
83 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
84 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
85 }
86 
87 void __init intel_pmu_pebs_data_source_skl(bool pmem)
88 {
89 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
90 
91 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
92 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
93 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
94 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
95 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
96 }
97 
98 static u64 precise_store_data(u64 status)
99 {
100 	union intel_x86_pebs_dse dse;
101 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
102 
103 	dse.val = status;
104 
105 	/*
106 	 * bit 4: TLB access
107 	 * 1 = stored missed 2nd level TLB
108 	 *
109 	 * so it either hit the walker or the OS
110 	 * otherwise hit 2nd level TLB
111 	 */
112 	if (dse.st_stlb_miss)
113 		val |= P(TLB, MISS);
114 	else
115 		val |= P(TLB, HIT);
116 
117 	/*
118 	 * bit 0: hit L1 data cache
119 	 * if not set, then all we know is that
120 	 * it missed L1D
121 	 */
122 	if (dse.st_l1d_hit)
123 		val |= P(LVL, HIT);
124 	else
125 		val |= P(LVL, MISS);
126 
127 	/*
128 	 * bit 5: Locked prefix
129 	 */
130 	if (dse.st_locked)
131 		val |= P(LOCK, LOCKED);
132 
133 	return val;
134 }
135 
136 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
137 {
138 	union perf_mem_data_src dse;
139 
140 	dse.val = PERF_MEM_NA;
141 
142 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
143 		dse.mem_op = PERF_MEM_OP_STORE;
144 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
145 		dse.mem_op = PERF_MEM_OP_LOAD;
146 
147 	/*
148 	 * L1 info only valid for following events:
149 	 *
150 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
151 	 * MEM_UOPS_RETIRED.LOCK_STORES
152 	 * MEM_UOPS_RETIRED.SPLIT_STORES
153 	 * MEM_UOPS_RETIRED.ALL_STORES
154 	 */
155 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
156 		if (status & 1)
157 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
158 		else
159 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
160 	}
161 	return dse.val;
162 }
163 
164 static u64 load_latency_data(u64 status)
165 {
166 	union intel_x86_pebs_dse dse;
167 	u64 val;
168 
169 	dse.val = status;
170 
171 	/*
172 	 * use the mapping table for bit 0-3
173 	 */
174 	val = pebs_data_source[dse.ld_dse];
175 
176 	/*
177 	 * Nehalem models do not support TLB, Lock infos
178 	 */
179 	if (x86_pmu.pebs_no_tlb) {
180 		val |= P(TLB, NA) | P(LOCK, NA);
181 		return val;
182 	}
183 	/*
184 	 * bit 4: TLB access
185 	 * 0 = did not miss 2nd level TLB
186 	 * 1 = missed 2nd level TLB
187 	 */
188 	if (dse.ld_stlb_miss)
189 		val |= P(TLB, MISS) | P(TLB, L2);
190 	else
191 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
192 
193 	/*
194 	 * bit 5: locked prefix
195 	 */
196 	if (dse.ld_locked)
197 		val |= P(LOCK, LOCKED);
198 
199 	return val;
200 }
201 
202 struct pebs_record_core {
203 	u64 flags, ip;
204 	u64 ax, bx, cx, dx;
205 	u64 si, di, bp, sp;
206 	u64 r8,  r9,  r10, r11;
207 	u64 r12, r13, r14, r15;
208 };
209 
210 struct pebs_record_nhm {
211 	u64 flags, ip;
212 	u64 ax, bx, cx, dx;
213 	u64 si, di, bp, sp;
214 	u64 r8,  r9,  r10, r11;
215 	u64 r12, r13, r14, r15;
216 	u64 status, dla, dse, lat;
217 };
218 
219 /*
220  * Same as pebs_record_nhm, with two additional fields.
221  */
222 struct pebs_record_hsw {
223 	u64 flags, ip;
224 	u64 ax, bx, cx, dx;
225 	u64 si, di, bp, sp;
226 	u64 r8,  r9,  r10, r11;
227 	u64 r12, r13, r14, r15;
228 	u64 status, dla, dse, lat;
229 	u64 real_ip, tsx_tuning;
230 };
231 
232 union hsw_tsx_tuning {
233 	struct {
234 		u32 cycles_last_block     : 32,
235 		    hle_abort		  : 1,
236 		    rtm_abort		  : 1,
237 		    instruction_abort     : 1,
238 		    non_instruction_abort : 1,
239 		    retry		  : 1,
240 		    data_conflict	  : 1,
241 		    capacity_writes	  : 1,
242 		    capacity_reads	  : 1;
243 	};
244 	u64	    value;
245 };
246 
247 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
248 
249 /* Same as HSW, plus TSC */
250 
251 struct pebs_record_skl {
252 	u64 flags, ip;
253 	u64 ax, bx, cx, dx;
254 	u64 si, di, bp, sp;
255 	u64 r8,  r9,  r10, r11;
256 	u64 r12, r13, r14, r15;
257 	u64 status, dla, dse, lat;
258 	u64 real_ip, tsx_tuning;
259 	u64 tsc;
260 };
261 
262 void init_debug_store_on_cpu(int cpu)
263 {
264 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
265 
266 	if (!ds)
267 		return;
268 
269 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
270 		     (u32)((u64)(unsigned long)ds),
271 		     (u32)((u64)(unsigned long)ds >> 32));
272 }
273 
274 void fini_debug_store_on_cpu(int cpu)
275 {
276 	if (!per_cpu(cpu_hw_events, cpu).ds)
277 		return;
278 
279 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
280 }
281 
282 static DEFINE_PER_CPU(void *, insn_buffer);
283 
284 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
285 {
286 	phys_addr_t pa;
287 	size_t msz = 0;
288 
289 	pa = virt_to_phys(addr);
290 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
291 		cea_set_pte(cea, pa, prot);
292 }
293 
294 static void ds_clear_cea(void *cea, size_t size)
295 {
296 	size_t msz = 0;
297 
298 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
299 		cea_set_pte(cea, 0, PAGE_NONE);
300 }
301 
302 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
303 {
304 	unsigned int order = get_order(size);
305 	int node = cpu_to_node(cpu);
306 	struct page *page;
307 
308 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
309 	return page ? page_address(page) : NULL;
310 }
311 
312 static void dsfree_pages(const void *buffer, size_t size)
313 {
314 	if (buffer)
315 		free_pages((unsigned long)buffer, get_order(size));
316 }
317 
318 static int alloc_pebs_buffer(int cpu)
319 {
320 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
321 	struct debug_store *ds = hwev->ds;
322 	size_t bsiz = x86_pmu.pebs_buffer_size;
323 	int max, node = cpu_to_node(cpu);
324 	void *buffer, *ibuffer, *cea;
325 
326 	if (!x86_pmu.pebs)
327 		return 0;
328 
329 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
330 	if (unlikely(!buffer))
331 		return -ENOMEM;
332 
333 	/*
334 	 * HSW+ already provides us the eventing ip; no need to allocate this
335 	 * buffer then.
336 	 */
337 	if (x86_pmu.intel_cap.pebs_format < 2) {
338 		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
339 		if (!ibuffer) {
340 			dsfree_pages(buffer, bsiz);
341 			return -ENOMEM;
342 		}
343 		per_cpu(insn_buffer, cpu) = ibuffer;
344 	}
345 	hwev->ds_pebs_vaddr = buffer;
346 	/* Update the cpu entry area mapping */
347 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
348 	ds->pebs_buffer_base = (unsigned long) cea;
349 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
350 	ds->pebs_index = ds->pebs_buffer_base;
351 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
352 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
353 	return 0;
354 }
355 
356 static void release_pebs_buffer(int cpu)
357 {
358 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
359 	struct debug_store *ds = hwev->ds;
360 	void *cea;
361 
362 	if (!ds || !x86_pmu.pebs)
363 		return;
364 
365 	kfree(per_cpu(insn_buffer, cpu));
366 	per_cpu(insn_buffer, cpu) = NULL;
367 
368 	/* Clear the fixmap */
369 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
370 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
371 	ds->pebs_buffer_base = 0;
372 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
373 	hwev->ds_pebs_vaddr = NULL;
374 }
375 
376 static int alloc_bts_buffer(int cpu)
377 {
378 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
379 	struct debug_store *ds = hwev->ds;
380 	void *buffer, *cea;
381 	int max;
382 
383 	if (!x86_pmu.bts)
384 		return 0;
385 
386 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
387 	if (unlikely(!buffer)) {
388 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
389 		return -ENOMEM;
390 	}
391 	hwev->ds_bts_vaddr = buffer;
392 	/* Update the fixmap */
393 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
394 	ds->bts_buffer_base = (unsigned long) cea;
395 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
396 	ds->bts_index = ds->bts_buffer_base;
397 	max = BTS_RECORD_SIZE * (BTS_BUFFER_SIZE / BTS_RECORD_SIZE);
398 	ds->bts_absolute_maximum = ds->bts_buffer_base + max;
399 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum - (max / 16);
400 	return 0;
401 }
402 
403 static void release_bts_buffer(int cpu)
404 {
405 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
406 	struct debug_store *ds = hwev->ds;
407 	void *cea;
408 
409 	if (!ds || !x86_pmu.bts)
410 		return;
411 
412 	/* Clear the fixmap */
413 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
414 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
415 	ds->bts_buffer_base = 0;
416 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
417 	hwev->ds_bts_vaddr = NULL;
418 }
419 
420 static int alloc_ds_buffer(int cpu)
421 {
422 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
423 
424 	memset(ds, 0, sizeof(*ds));
425 	per_cpu(cpu_hw_events, cpu).ds = ds;
426 	return 0;
427 }
428 
429 static void release_ds_buffer(int cpu)
430 {
431 	per_cpu(cpu_hw_events, cpu).ds = NULL;
432 }
433 
434 void release_ds_buffers(void)
435 {
436 	int cpu;
437 
438 	if (!x86_pmu.bts && !x86_pmu.pebs)
439 		return;
440 
441 	get_online_cpus();
442 	for_each_online_cpu(cpu)
443 		fini_debug_store_on_cpu(cpu);
444 
445 	for_each_possible_cpu(cpu) {
446 		release_pebs_buffer(cpu);
447 		release_bts_buffer(cpu);
448 		release_ds_buffer(cpu);
449 	}
450 	put_online_cpus();
451 }
452 
453 void reserve_ds_buffers(void)
454 {
455 	int bts_err = 0, pebs_err = 0;
456 	int cpu;
457 
458 	x86_pmu.bts_active = 0;
459 	x86_pmu.pebs_active = 0;
460 
461 	if (!x86_pmu.bts && !x86_pmu.pebs)
462 		return;
463 
464 	if (!x86_pmu.bts)
465 		bts_err = 1;
466 
467 	if (!x86_pmu.pebs)
468 		pebs_err = 1;
469 
470 	get_online_cpus();
471 
472 	for_each_possible_cpu(cpu) {
473 		if (alloc_ds_buffer(cpu)) {
474 			bts_err = 1;
475 			pebs_err = 1;
476 		}
477 
478 		if (!bts_err && alloc_bts_buffer(cpu))
479 			bts_err = 1;
480 
481 		if (!pebs_err && alloc_pebs_buffer(cpu))
482 			pebs_err = 1;
483 
484 		if (bts_err && pebs_err)
485 			break;
486 	}
487 
488 	if (bts_err) {
489 		for_each_possible_cpu(cpu)
490 			release_bts_buffer(cpu);
491 	}
492 
493 	if (pebs_err) {
494 		for_each_possible_cpu(cpu)
495 			release_pebs_buffer(cpu);
496 	}
497 
498 	if (bts_err && pebs_err) {
499 		for_each_possible_cpu(cpu)
500 			release_ds_buffer(cpu);
501 	} else {
502 		if (x86_pmu.bts && !bts_err)
503 			x86_pmu.bts_active = 1;
504 
505 		if (x86_pmu.pebs && !pebs_err)
506 			x86_pmu.pebs_active = 1;
507 
508 		for_each_online_cpu(cpu)
509 			init_debug_store_on_cpu(cpu);
510 	}
511 
512 	put_online_cpus();
513 }
514 
515 /*
516  * BTS
517  */
518 
519 struct event_constraint bts_constraint =
520 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
521 
522 void intel_pmu_enable_bts(u64 config)
523 {
524 	unsigned long debugctlmsr;
525 
526 	debugctlmsr = get_debugctlmsr();
527 
528 	debugctlmsr |= DEBUGCTLMSR_TR;
529 	debugctlmsr |= DEBUGCTLMSR_BTS;
530 	if (config & ARCH_PERFMON_EVENTSEL_INT)
531 		debugctlmsr |= DEBUGCTLMSR_BTINT;
532 
533 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
534 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
535 
536 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
537 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
538 
539 	update_debugctlmsr(debugctlmsr);
540 }
541 
542 void intel_pmu_disable_bts(void)
543 {
544 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
545 	unsigned long debugctlmsr;
546 
547 	if (!cpuc->ds)
548 		return;
549 
550 	debugctlmsr = get_debugctlmsr();
551 
552 	debugctlmsr &=
553 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
554 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
555 
556 	update_debugctlmsr(debugctlmsr);
557 }
558 
559 int intel_pmu_drain_bts_buffer(void)
560 {
561 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
562 	struct debug_store *ds = cpuc->ds;
563 	struct bts_record {
564 		u64	from;
565 		u64	to;
566 		u64	flags;
567 	};
568 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
569 	struct bts_record *at, *base, *top;
570 	struct perf_output_handle handle;
571 	struct perf_event_header header;
572 	struct perf_sample_data data;
573 	unsigned long skip = 0;
574 	struct pt_regs regs;
575 
576 	if (!event)
577 		return 0;
578 
579 	if (!x86_pmu.bts_active)
580 		return 0;
581 
582 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
583 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
584 
585 	if (top <= base)
586 		return 0;
587 
588 	memset(&regs, 0, sizeof(regs));
589 
590 	ds->bts_index = ds->bts_buffer_base;
591 
592 	perf_sample_data_init(&data, 0, event->hw.last_period);
593 
594 	/*
595 	 * BTS leaks kernel addresses in branches across the cpl boundary,
596 	 * such as traps or system calls, so unless the user is asking for
597 	 * kernel tracing (and right now it's not possible), we'd need to
598 	 * filter them out. But first we need to count how many of those we
599 	 * have in the current batch. This is an extra O(n) pass, however,
600 	 * it's much faster than the other one especially considering that
601 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
602 	 * alloc_bts_buffer()).
603 	 */
604 	for (at = base; at < top; at++) {
605 		/*
606 		 * Note that right now *this* BTS code only works if
607 		 * attr::exclude_kernel is set, but let's keep this extra
608 		 * check here in case that changes.
609 		 */
610 		if (event->attr.exclude_kernel &&
611 		    (kernel_ip(at->from) || kernel_ip(at->to)))
612 			skip++;
613 	}
614 
615 	/*
616 	 * Prepare a generic sample, i.e. fill in the invariant fields.
617 	 * We will overwrite the from and to address before we output
618 	 * the sample.
619 	 */
620 	rcu_read_lock();
621 	perf_prepare_sample(&header, &data, event, &regs);
622 
623 	if (perf_output_begin(&handle, event, header.size *
624 			      (top - base - skip)))
625 		goto unlock;
626 
627 	for (at = base; at < top; at++) {
628 		/* Filter out any records that contain kernel addresses. */
629 		if (event->attr.exclude_kernel &&
630 		    (kernel_ip(at->from) || kernel_ip(at->to)))
631 			continue;
632 
633 		data.ip		= at->from;
634 		data.addr	= at->to;
635 
636 		perf_output_sample(&handle, &header, &data, event);
637 	}
638 
639 	perf_output_end(&handle);
640 
641 	/* There's new data available. */
642 	event->hw.interrupts++;
643 	event->pending_kill = POLL_IN;
644 unlock:
645 	rcu_read_unlock();
646 	return 1;
647 }
648 
649 static inline void intel_pmu_drain_pebs_buffer(void)
650 {
651 	struct pt_regs regs;
652 
653 	x86_pmu.drain_pebs(&regs);
654 }
655 
656 /*
657  * PEBS
658  */
659 struct event_constraint intel_core2_pebs_event_constraints[] = {
660 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
661 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
662 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
663 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
664 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
665 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
666 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
667 	EVENT_CONSTRAINT_END
668 };
669 
670 struct event_constraint intel_atom_pebs_event_constraints[] = {
671 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
672 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
673 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
674 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
675 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
676 	/* Allow all events as PEBS with no flags */
677 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
678 	EVENT_CONSTRAINT_END
679 };
680 
681 struct event_constraint intel_slm_pebs_event_constraints[] = {
682 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
683 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
684 	/* Allow all events as PEBS with no flags */
685 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
686 	EVENT_CONSTRAINT_END
687 };
688 
689 struct event_constraint intel_glm_pebs_event_constraints[] = {
690 	/* Allow all events as PEBS with no flags */
691 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
692 	EVENT_CONSTRAINT_END
693 };
694 
695 struct event_constraint intel_glp_pebs_event_constraints[] = {
696 	/* Allow all events as PEBS with no flags */
697 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
698 	EVENT_CONSTRAINT_END
699 };
700 
701 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
702 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
703 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
704 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
705 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
706 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
707 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
708 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
709 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
710 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
711 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
712 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
713 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
714 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
715 	EVENT_CONSTRAINT_END
716 };
717 
718 struct event_constraint intel_westmere_pebs_event_constraints[] = {
719 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
720 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
721 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
722 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
723 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
724 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
725 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
726 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
727 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
728 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
729 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
730 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
731 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
732 	EVENT_CONSTRAINT_END
733 };
734 
735 struct event_constraint intel_snb_pebs_event_constraints[] = {
736 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
737 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
738 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
739 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
740 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
741         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
742         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
743         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
744         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
745 	/* Allow all events as PEBS with no flags */
746 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
747 	EVENT_CONSTRAINT_END
748 };
749 
750 struct event_constraint intel_ivb_pebs_event_constraints[] = {
751         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
752         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
753 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
754 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
755 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
756 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
757 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
758 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
759 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
760 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
761 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
762 	/* Allow all events as PEBS with no flags */
763 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
764         EVENT_CONSTRAINT_END
765 };
766 
767 struct event_constraint intel_hsw_pebs_event_constraints[] = {
768 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
769 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
770 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
771 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
772 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
773 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
774 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
775 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
776 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
777 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
778 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
779 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
780 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
781 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
782 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
783 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
784 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
785 	/* Allow all events as PEBS with no flags */
786 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
787 	EVENT_CONSTRAINT_END
788 };
789 
790 struct event_constraint intel_bdw_pebs_event_constraints[] = {
791 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
792 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
793 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
794 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
795 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
796 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
797 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
798 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
799 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
800 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
801 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
802 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
803 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
804 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
805 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
806 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
807 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
808 	/* Allow all events as PEBS with no flags */
809 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
810 	EVENT_CONSTRAINT_END
811 };
812 
813 
814 struct event_constraint intel_skl_pebs_event_constraints[] = {
815 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
816 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
817 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
818 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
819 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
820 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
821 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
822 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
823 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
824 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
825 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
826 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
827 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
828 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
829 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
830 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
831 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
832 	/* Allow all events as PEBS with no flags */
833 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
834 	EVENT_CONSTRAINT_END
835 };
836 
837 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
838 {
839 	struct event_constraint *c;
840 
841 	if (!event->attr.precise_ip)
842 		return NULL;
843 
844 	if (x86_pmu.pebs_constraints) {
845 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
846 			if ((event->hw.config & c->cmask) == c->code) {
847 				event->hw.flags |= c->flags;
848 				return c;
849 			}
850 		}
851 	}
852 
853 	return &emptyconstraint;
854 }
855 
856 /*
857  * We need the sched_task callback even for per-cpu events when we use
858  * the large interrupt threshold, such that we can provide PID and TID
859  * to PEBS samples.
860  */
861 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
862 {
863 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
864 }
865 
866 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
867 {
868 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
869 
870 	if (!sched_in && pebs_needs_sched_cb(cpuc))
871 		intel_pmu_drain_pebs_buffer();
872 }
873 
874 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
875 {
876 	struct debug_store *ds = cpuc->ds;
877 	u64 threshold;
878 
879 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
880 		threshold = ds->pebs_absolute_maximum -
881 			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
882 	} else {
883 		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
884 	}
885 
886 	ds->pebs_interrupt_threshold = threshold;
887 }
888 
889 static void
890 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
891 {
892 	/*
893 	 * Make sure we get updated with the first PEBS
894 	 * event. It will trigger also during removal, but
895 	 * that does not hurt:
896 	 */
897 	bool update = cpuc->n_pebs == 1;
898 
899 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
900 		if (!needed_cb)
901 			perf_sched_cb_inc(pmu);
902 		else
903 			perf_sched_cb_dec(pmu);
904 
905 		update = true;
906 	}
907 
908 	if (update)
909 		pebs_update_threshold(cpuc);
910 }
911 
912 void intel_pmu_pebs_add(struct perf_event *event)
913 {
914 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
915 	struct hw_perf_event *hwc = &event->hw;
916 	bool needed_cb = pebs_needs_sched_cb(cpuc);
917 
918 	cpuc->n_pebs++;
919 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
920 		cpuc->n_large_pebs++;
921 
922 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
923 }
924 
925 void intel_pmu_pebs_enable(struct perf_event *event)
926 {
927 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
928 	struct hw_perf_event *hwc = &event->hw;
929 	struct debug_store *ds = cpuc->ds;
930 
931 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
932 
933 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
934 
935 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
936 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
937 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
938 		cpuc->pebs_enabled |= 1ULL << 63;
939 
940 	/*
941 	 * Use auto-reload if possible to save a MSR write in the PMI.
942 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
943 	 */
944 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
945 		ds->pebs_event_reset[hwc->idx] =
946 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
947 	} else {
948 		ds->pebs_event_reset[hwc->idx] = 0;
949 	}
950 }
951 
952 void intel_pmu_pebs_del(struct perf_event *event)
953 {
954 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
955 	struct hw_perf_event *hwc = &event->hw;
956 	bool needed_cb = pebs_needs_sched_cb(cpuc);
957 
958 	cpuc->n_pebs--;
959 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
960 		cpuc->n_large_pebs--;
961 
962 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
963 }
964 
965 void intel_pmu_pebs_disable(struct perf_event *event)
966 {
967 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
968 	struct hw_perf_event *hwc = &event->hw;
969 
970 	if (cpuc->n_pebs == cpuc->n_large_pebs)
971 		intel_pmu_drain_pebs_buffer();
972 
973 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
974 
975 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
976 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
977 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
978 		cpuc->pebs_enabled &= ~(1ULL << 63);
979 
980 	if (cpuc->enabled)
981 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
982 
983 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
984 }
985 
986 void intel_pmu_pebs_enable_all(void)
987 {
988 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
989 
990 	if (cpuc->pebs_enabled)
991 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
992 }
993 
994 void intel_pmu_pebs_disable_all(void)
995 {
996 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
997 
998 	if (cpuc->pebs_enabled)
999 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1000 }
1001 
1002 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1003 {
1004 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1005 	unsigned long from = cpuc->lbr_entries[0].from;
1006 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1007 	unsigned long ip = regs->ip;
1008 	int is_64bit = 0;
1009 	void *kaddr;
1010 	int size;
1011 
1012 	/*
1013 	 * We don't need to fixup if the PEBS assist is fault like
1014 	 */
1015 	if (!x86_pmu.intel_cap.pebs_trap)
1016 		return 1;
1017 
1018 	/*
1019 	 * No LBR entry, no basic block, no rewinding
1020 	 */
1021 	if (!cpuc->lbr_stack.nr || !from || !to)
1022 		return 0;
1023 
1024 	/*
1025 	 * Basic blocks should never cross user/kernel boundaries
1026 	 */
1027 	if (kernel_ip(ip) != kernel_ip(to))
1028 		return 0;
1029 
1030 	/*
1031 	 * unsigned math, either ip is before the start (impossible) or
1032 	 * the basic block is larger than 1 page (sanity)
1033 	 */
1034 	if ((ip - to) > PEBS_FIXUP_SIZE)
1035 		return 0;
1036 
1037 	/*
1038 	 * We sampled a branch insn, rewind using the LBR stack
1039 	 */
1040 	if (ip == to) {
1041 		set_linear_ip(regs, from);
1042 		return 1;
1043 	}
1044 
1045 	size = ip - to;
1046 	if (!kernel_ip(ip)) {
1047 		int bytes;
1048 		u8 *buf = this_cpu_read(insn_buffer);
1049 
1050 		/* 'size' must fit our buffer, see above */
1051 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1052 		if (bytes != 0)
1053 			return 0;
1054 
1055 		kaddr = buf;
1056 	} else {
1057 		kaddr = (void *)to;
1058 	}
1059 
1060 	do {
1061 		struct insn insn;
1062 
1063 		old_to = to;
1064 
1065 #ifdef CONFIG_X86_64
1066 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1067 #endif
1068 		insn_init(&insn, kaddr, size, is_64bit);
1069 		insn_get_length(&insn);
1070 		/*
1071 		 * Make sure there was not a problem decoding the
1072 		 * instruction and getting the length.  This is
1073 		 * doubly important because we have an infinite
1074 		 * loop if insn.length=0.
1075 		 */
1076 		if (!insn.length)
1077 			break;
1078 
1079 		to += insn.length;
1080 		kaddr += insn.length;
1081 		size -= insn.length;
1082 	} while (to < ip);
1083 
1084 	if (to == ip) {
1085 		set_linear_ip(regs, old_to);
1086 		return 1;
1087 	}
1088 
1089 	/*
1090 	 * Even though we decoded the basic block, the instruction stream
1091 	 * never matched the given IP, either the TO or the IP got corrupted.
1092 	 */
1093 	return 0;
1094 }
1095 
1096 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1097 {
1098 	if (pebs->tsx_tuning) {
1099 		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1100 		return tsx.cycles_last_block;
1101 	}
1102 	return 0;
1103 }
1104 
1105 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1106 {
1107 	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1108 
1109 	/* For RTM XABORTs also log the abort code from AX */
1110 	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1111 		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1112 	return txn;
1113 }
1114 
1115 static void setup_pebs_sample_data(struct perf_event *event,
1116 				   struct pt_regs *iregs, void *__pebs,
1117 				   struct perf_sample_data *data,
1118 				   struct pt_regs *regs)
1119 {
1120 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1121 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1122 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1123 		 PERF_X86_EVENT_PEBS_NA_HSW)
1124 	/*
1125 	 * We cast to the biggest pebs_record but are careful not to
1126 	 * unconditionally access the 'extra' entries.
1127 	 */
1128 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1129 	struct pebs_record_skl *pebs = __pebs;
1130 	u64 sample_type;
1131 	int fll, fst, dsrc;
1132 	int fl = event->hw.flags;
1133 
1134 	if (pebs == NULL)
1135 		return;
1136 
1137 	sample_type = event->attr.sample_type;
1138 	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1139 
1140 	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1141 	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1142 
1143 	perf_sample_data_init(data, 0, event->hw.last_period);
1144 
1145 	data->period = event->hw.last_period;
1146 
1147 	/*
1148 	 * Use latency for weight (only avail with PEBS-LL)
1149 	 */
1150 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1151 		data->weight = pebs->lat;
1152 
1153 	/*
1154 	 * data.data_src encodes the data source
1155 	 */
1156 	if (dsrc) {
1157 		u64 val = PERF_MEM_NA;
1158 		if (fll)
1159 			val = load_latency_data(pebs->dse);
1160 		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1161 			val = precise_datala_hsw(event, pebs->dse);
1162 		else if (fst)
1163 			val = precise_store_data(pebs->dse);
1164 		data->data_src.val = val;
1165 	}
1166 
1167 	/*
1168 	 * We use the interrupt regs as a base because the PEBS record does not
1169 	 * contain a full regs set, specifically it seems to lack segment
1170 	 * descriptors, which get used by things like user_mode().
1171 	 *
1172 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1173 	 *
1174 	 * We must however always use BP,SP from iregs for the unwinder to stay
1175 	 * sane; the record BP,SP can point into thin air when the record is
1176 	 * from a previous PMI context or an (I)RET happend between the record
1177 	 * and PMI.
1178 	 */
1179 	*regs = *iregs;
1180 	regs->flags = pebs->flags;
1181 	set_linear_ip(regs, pebs->ip);
1182 
1183 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1184 		regs->ax = pebs->ax;
1185 		regs->bx = pebs->bx;
1186 		regs->cx = pebs->cx;
1187 		regs->dx = pebs->dx;
1188 		regs->si = pebs->si;
1189 		regs->di = pebs->di;
1190 
1191 		/*
1192 		 * Per the above; only set BP,SP if we don't need callchains.
1193 		 *
1194 		 * XXX: does this make sense?
1195 		 */
1196 		if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1197 			regs->bp = pebs->bp;
1198 			regs->sp = pebs->sp;
1199 		}
1200 
1201 		/*
1202 		 * Preserve PERF_EFLAGS_VM from set_linear_ip().
1203 		 */
1204 		regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1205 #ifndef CONFIG_X86_32
1206 		regs->r8 = pebs->r8;
1207 		regs->r9 = pebs->r9;
1208 		regs->r10 = pebs->r10;
1209 		regs->r11 = pebs->r11;
1210 		regs->r12 = pebs->r12;
1211 		regs->r13 = pebs->r13;
1212 		regs->r14 = pebs->r14;
1213 		regs->r15 = pebs->r15;
1214 #endif
1215 	}
1216 
1217 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1218 		regs->ip = pebs->real_ip;
1219 		regs->flags |= PERF_EFLAGS_EXACT;
1220 	} else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1221 		regs->flags |= PERF_EFLAGS_EXACT;
1222 	else
1223 		regs->flags &= ~PERF_EFLAGS_EXACT;
1224 
1225 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1226 	    x86_pmu.intel_cap.pebs_format >= 1)
1227 		data->addr = pebs->dla;
1228 
1229 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1230 		/* Only set the TSX weight when no memory weight. */
1231 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1232 			data->weight = intel_hsw_weight(pebs);
1233 
1234 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1235 			data->txn = intel_hsw_transaction(pebs);
1236 	}
1237 
1238 	/*
1239 	 * v3 supplies an accurate time stamp, so we use that
1240 	 * for the time stamp.
1241 	 *
1242 	 * We can only do this for the default trace clock.
1243 	 */
1244 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1245 		event->attr.use_clockid == 0)
1246 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1247 
1248 	if (has_branch_stack(event))
1249 		data->br_stack = &cpuc->lbr_stack;
1250 }
1251 
1252 static inline void *
1253 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1254 {
1255 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1256 	void *at;
1257 	u64 pebs_status;
1258 
1259 	/*
1260 	 * fmt0 does not have a status bitfield (does not use
1261 	 * perf_record_nhm format)
1262 	 */
1263 	if (x86_pmu.intel_cap.pebs_format < 1)
1264 		return base;
1265 
1266 	if (base == NULL)
1267 		return NULL;
1268 
1269 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1270 		struct pebs_record_nhm *p = at;
1271 
1272 		if (test_bit(bit, (unsigned long *)&p->status)) {
1273 			/* PEBS v3 has accurate status bits */
1274 			if (x86_pmu.intel_cap.pebs_format >= 3)
1275 				return at;
1276 
1277 			if (p->status == (1 << bit))
1278 				return at;
1279 
1280 			/* clear non-PEBS bit and re-check */
1281 			pebs_status = p->status & cpuc->pebs_enabled;
1282 			pebs_status &= PEBS_COUNTER_MASK;
1283 			if (pebs_status == (1 << bit))
1284 				return at;
1285 		}
1286 	}
1287 	return NULL;
1288 }
1289 
1290 static void __intel_pmu_pebs_event(struct perf_event *event,
1291 				   struct pt_regs *iregs,
1292 				   void *base, void *top,
1293 				   int bit, int count)
1294 {
1295 	struct perf_sample_data data;
1296 	struct pt_regs regs;
1297 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1298 
1299 	if (!intel_pmu_save_and_restart(event) &&
1300 	    !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1301 		return;
1302 
1303 	while (count > 1) {
1304 		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1305 		perf_event_output(event, &data, &regs);
1306 		at += x86_pmu.pebs_record_size;
1307 		at = get_next_pebs_record_by_bit(at, top, bit);
1308 		count--;
1309 	}
1310 
1311 	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1312 
1313 	/*
1314 	 * All but the last records are processed.
1315 	 * The last one is left to be able to call the overflow handler.
1316 	 */
1317 	if (perf_event_overflow(event, &data, &regs)) {
1318 		x86_pmu_stop(event, 0);
1319 		return;
1320 	}
1321 
1322 }
1323 
1324 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1325 {
1326 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1327 	struct debug_store *ds = cpuc->ds;
1328 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1329 	struct pebs_record_core *at, *top;
1330 	int n;
1331 
1332 	if (!x86_pmu.pebs_active)
1333 		return;
1334 
1335 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1336 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1337 
1338 	/*
1339 	 * Whatever else happens, drain the thing
1340 	 */
1341 	ds->pebs_index = ds->pebs_buffer_base;
1342 
1343 	if (!test_bit(0, cpuc->active_mask))
1344 		return;
1345 
1346 	WARN_ON_ONCE(!event);
1347 
1348 	if (!event->attr.precise_ip)
1349 		return;
1350 
1351 	n = top - at;
1352 	if (n <= 0)
1353 		return;
1354 
1355 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1356 }
1357 
1358 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1359 {
1360 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1361 	struct debug_store *ds = cpuc->ds;
1362 	struct perf_event *event;
1363 	void *base, *at, *top;
1364 	short counts[MAX_PEBS_EVENTS] = {};
1365 	short error[MAX_PEBS_EVENTS] = {};
1366 	int bit, i;
1367 
1368 	if (!x86_pmu.pebs_active)
1369 		return;
1370 
1371 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1372 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1373 
1374 	ds->pebs_index = ds->pebs_buffer_base;
1375 
1376 	if (unlikely(base >= top))
1377 		return;
1378 
1379 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1380 		struct pebs_record_nhm *p = at;
1381 		u64 pebs_status;
1382 
1383 		pebs_status = p->status & cpuc->pebs_enabled;
1384 		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1385 
1386 		/* PEBS v3 has more accurate status bits */
1387 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1388 			for_each_set_bit(bit, (unsigned long *)&pebs_status,
1389 					 x86_pmu.max_pebs_events)
1390 				counts[bit]++;
1391 
1392 			continue;
1393 		}
1394 
1395 		/*
1396 		 * On some CPUs the PEBS status can be zero when PEBS is
1397 		 * racing with clearing of GLOBAL_STATUS.
1398 		 *
1399 		 * Normally we would drop that record, but in the
1400 		 * case when there is only a single active PEBS event
1401 		 * we can assume it's for that event.
1402 		 */
1403 		if (!pebs_status && cpuc->pebs_enabled &&
1404 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1405 			pebs_status = cpuc->pebs_enabled;
1406 
1407 		bit = find_first_bit((unsigned long *)&pebs_status,
1408 					x86_pmu.max_pebs_events);
1409 		if (bit >= x86_pmu.max_pebs_events)
1410 			continue;
1411 
1412 		/*
1413 		 * The PEBS hardware does not deal well with the situation
1414 		 * when events happen near to each other and multiple bits
1415 		 * are set. But it should happen rarely.
1416 		 *
1417 		 * If these events include one PEBS and multiple non-PEBS
1418 		 * events, it doesn't impact PEBS record. The record will
1419 		 * be handled normally. (slow path)
1420 		 *
1421 		 * If these events include two or more PEBS events, the
1422 		 * records for the events can be collapsed into a single
1423 		 * one, and it's not possible to reconstruct all events
1424 		 * that caused the PEBS record. It's called collision.
1425 		 * If collision happened, the record will be dropped.
1426 		 */
1427 		if (p->status != (1ULL << bit)) {
1428 			for_each_set_bit(i, (unsigned long *)&pebs_status,
1429 					 x86_pmu.max_pebs_events)
1430 				error[i]++;
1431 			continue;
1432 		}
1433 
1434 		counts[bit]++;
1435 	}
1436 
1437 	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1438 		if ((counts[bit] == 0) && (error[bit] == 0))
1439 			continue;
1440 
1441 		event = cpuc->events[bit];
1442 		if (WARN_ON_ONCE(!event))
1443 			continue;
1444 
1445 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1446 			continue;
1447 
1448 		/* log dropped samples number */
1449 		if (error[bit]) {
1450 			perf_log_lost_samples(event, error[bit]);
1451 
1452 			if (perf_event_account_interrupt(event))
1453 				x86_pmu_stop(event, 0);
1454 		}
1455 
1456 		if (counts[bit]) {
1457 			__intel_pmu_pebs_event(event, iregs, base,
1458 					       top, bit, counts[bit]);
1459 		}
1460 	}
1461 }
1462 
1463 /*
1464  * BTS, PEBS probe and setup
1465  */
1466 
1467 void __init intel_ds_init(void)
1468 {
1469 	/*
1470 	 * No support for 32bit formats
1471 	 */
1472 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1473 		return;
1474 
1475 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1476 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1477 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1478 	if (x86_pmu.pebs) {
1479 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1480 		int format = x86_pmu.intel_cap.pebs_format;
1481 
1482 		switch (format) {
1483 		case 0:
1484 			pr_cont("PEBS fmt0%c, ", pebs_type);
1485 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1486 			/*
1487 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1488 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1489 			 * mysteriously hang on Core2.
1490 			 *
1491 			 * As a workaround, we don't do this.
1492 			 */
1493 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1494 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1495 			break;
1496 
1497 		case 1:
1498 			pr_cont("PEBS fmt1%c, ", pebs_type);
1499 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1500 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1501 			break;
1502 
1503 		case 2:
1504 			pr_cont("PEBS fmt2%c, ", pebs_type);
1505 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1506 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1507 			break;
1508 
1509 		case 3:
1510 			pr_cont("PEBS fmt3%c, ", pebs_type);
1511 			x86_pmu.pebs_record_size =
1512 						sizeof(struct pebs_record_skl);
1513 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1514 			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1515 			break;
1516 
1517 		default:
1518 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1519 			x86_pmu.pebs = 0;
1520 		}
1521 	}
1522 }
1523 
1524 void perf_restore_debug_store(void)
1525 {
1526 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1527 
1528 	if (!x86_pmu.bts && !x86_pmu.pebs)
1529 		return;
1530 
1531 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1532 }
1533