xref: /linux/arch/x86/events/intel/ds.c (revision ea8c64ace86647260ec4255f483e5844d62af2df)
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/tlbflush.h>
9 #include <asm/insn.h>
10 
11 #include "../perf_event.h"
12 
13 /* Waste a full page so it can be mapped into the cpu_entry_area */
14 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
15 
16 /* The size of a BTS record in bytes: */
17 #define BTS_RECORD_SIZE		24
18 
19 #define PEBS_FIXUP_SIZE		PAGE_SIZE
20 
21 /*
22  * pebs_record_32 for p4 and core not supported
23 
24 struct pebs_record_32 {
25 	u32 flags, ip;
26 	u32 ax, bc, cx, dx;
27 	u32 si, di, bp, sp;
28 };
29 
30  */
31 
32 union intel_x86_pebs_dse {
33 	u64 val;
34 	struct {
35 		unsigned int ld_dse:4;
36 		unsigned int ld_stlb_miss:1;
37 		unsigned int ld_locked:1;
38 		unsigned int ld_reserved:26;
39 	};
40 	struct {
41 		unsigned int st_l1d_hit:1;
42 		unsigned int st_reserved1:3;
43 		unsigned int st_stlb_miss:1;
44 		unsigned int st_locked:1;
45 		unsigned int st_reserved2:26;
46 	};
47 };
48 
49 
50 /*
51  * Map PEBS Load Latency Data Source encodings to generic
52  * memory data source information
53  */
54 #define P(a, b) PERF_MEM_S(a, b)
55 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
56 #define LEVEL(x) P(LVLNUM, x)
57 #define REM P(REMOTE, REMOTE)
58 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
59 
60 /* Version for Sandy Bridge and later */
61 static u64 pebs_data_source[] = {
62 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
63 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
64 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
65 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
66 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
67 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
68 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
69 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
70 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
71 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
72 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
73 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
74 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
75 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
76 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
77 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
78 };
79 
80 /* Patch up minor differences in the bits */
81 void __init intel_pmu_pebs_data_source_nhm(void)
82 {
83 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
84 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
85 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
86 }
87 
88 void __init intel_pmu_pebs_data_source_skl(bool pmem)
89 {
90 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
91 
92 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
93 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
94 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
95 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
96 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
97 }
98 
99 static u64 precise_store_data(u64 status)
100 {
101 	union intel_x86_pebs_dse dse;
102 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
103 
104 	dse.val = status;
105 
106 	/*
107 	 * bit 4: TLB access
108 	 * 1 = stored missed 2nd level TLB
109 	 *
110 	 * so it either hit the walker or the OS
111 	 * otherwise hit 2nd level TLB
112 	 */
113 	if (dse.st_stlb_miss)
114 		val |= P(TLB, MISS);
115 	else
116 		val |= P(TLB, HIT);
117 
118 	/*
119 	 * bit 0: hit L1 data cache
120 	 * if not set, then all we know is that
121 	 * it missed L1D
122 	 */
123 	if (dse.st_l1d_hit)
124 		val |= P(LVL, HIT);
125 	else
126 		val |= P(LVL, MISS);
127 
128 	/*
129 	 * bit 5: Locked prefix
130 	 */
131 	if (dse.st_locked)
132 		val |= P(LOCK, LOCKED);
133 
134 	return val;
135 }
136 
137 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
138 {
139 	union perf_mem_data_src dse;
140 
141 	dse.val = PERF_MEM_NA;
142 
143 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
144 		dse.mem_op = PERF_MEM_OP_STORE;
145 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
146 		dse.mem_op = PERF_MEM_OP_LOAD;
147 
148 	/*
149 	 * L1 info only valid for following events:
150 	 *
151 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
152 	 * MEM_UOPS_RETIRED.LOCK_STORES
153 	 * MEM_UOPS_RETIRED.SPLIT_STORES
154 	 * MEM_UOPS_RETIRED.ALL_STORES
155 	 */
156 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
157 		if (status & 1)
158 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
159 		else
160 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
161 	}
162 	return dse.val;
163 }
164 
165 static u64 load_latency_data(u64 status)
166 {
167 	union intel_x86_pebs_dse dse;
168 	u64 val;
169 
170 	dse.val = status;
171 
172 	/*
173 	 * use the mapping table for bit 0-3
174 	 */
175 	val = pebs_data_source[dse.ld_dse];
176 
177 	/*
178 	 * Nehalem models do not support TLB, Lock infos
179 	 */
180 	if (x86_pmu.pebs_no_tlb) {
181 		val |= P(TLB, NA) | P(LOCK, NA);
182 		return val;
183 	}
184 	/*
185 	 * bit 4: TLB access
186 	 * 0 = did not miss 2nd level TLB
187 	 * 1 = missed 2nd level TLB
188 	 */
189 	if (dse.ld_stlb_miss)
190 		val |= P(TLB, MISS) | P(TLB, L2);
191 	else
192 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
193 
194 	/*
195 	 * bit 5: locked prefix
196 	 */
197 	if (dse.ld_locked)
198 		val |= P(LOCK, LOCKED);
199 
200 	return val;
201 }
202 
203 struct pebs_record_core {
204 	u64 flags, ip;
205 	u64 ax, bx, cx, dx;
206 	u64 si, di, bp, sp;
207 	u64 r8,  r9,  r10, r11;
208 	u64 r12, r13, r14, r15;
209 };
210 
211 struct pebs_record_nhm {
212 	u64 flags, ip;
213 	u64 ax, bx, cx, dx;
214 	u64 si, di, bp, sp;
215 	u64 r8,  r9,  r10, r11;
216 	u64 r12, r13, r14, r15;
217 	u64 status, dla, dse, lat;
218 };
219 
220 /*
221  * Same as pebs_record_nhm, with two additional fields.
222  */
223 struct pebs_record_hsw {
224 	u64 flags, ip;
225 	u64 ax, bx, cx, dx;
226 	u64 si, di, bp, sp;
227 	u64 r8,  r9,  r10, r11;
228 	u64 r12, r13, r14, r15;
229 	u64 status, dla, dse, lat;
230 	u64 real_ip, tsx_tuning;
231 };
232 
233 union hsw_tsx_tuning {
234 	struct {
235 		u32 cycles_last_block     : 32,
236 		    hle_abort		  : 1,
237 		    rtm_abort		  : 1,
238 		    instruction_abort     : 1,
239 		    non_instruction_abort : 1,
240 		    retry		  : 1,
241 		    data_conflict	  : 1,
242 		    capacity_writes	  : 1,
243 		    capacity_reads	  : 1;
244 	};
245 	u64	    value;
246 };
247 
248 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
249 
250 /* Same as HSW, plus TSC */
251 
252 struct pebs_record_skl {
253 	u64 flags, ip;
254 	u64 ax, bx, cx, dx;
255 	u64 si, di, bp, sp;
256 	u64 r8,  r9,  r10, r11;
257 	u64 r12, r13, r14, r15;
258 	u64 status, dla, dse, lat;
259 	u64 real_ip, tsx_tuning;
260 	u64 tsc;
261 };
262 
263 void init_debug_store_on_cpu(int cpu)
264 {
265 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
266 
267 	if (!ds)
268 		return;
269 
270 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
271 		     (u32)((u64)(unsigned long)ds),
272 		     (u32)((u64)(unsigned long)ds >> 32));
273 }
274 
275 void fini_debug_store_on_cpu(int cpu)
276 {
277 	if (!per_cpu(cpu_hw_events, cpu).ds)
278 		return;
279 
280 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
281 }
282 
283 static DEFINE_PER_CPU(void *, insn_buffer);
284 
285 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
286 {
287 	unsigned long start = (unsigned long)cea;
288 	phys_addr_t pa;
289 	size_t msz = 0;
290 
291 	pa = virt_to_phys(addr);
292 
293 	preempt_disable();
294 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
295 		cea_set_pte(cea, pa, prot);
296 
297 	/*
298 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
299 	 * all TLB entries for it.
300 	 */
301 	flush_tlb_kernel_range(start, start + size);
302 	preempt_enable();
303 }
304 
305 static void ds_clear_cea(void *cea, size_t size)
306 {
307 	unsigned long start = (unsigned long)cea;
308 	size_t msz = 0;
309 
310 	preempt_disable();
311 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
312 		cea_set_pte(cea, 0, PAGE_NONE);
313 
314 	flush_tlb_kernel_range(start, start + size);
315 	preempt_enable();
316 }
317 
318 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
319 {
320 	unsigned int order = get_order(size);
321 	int node = cpu_to_node(cpu);
322 	struct page *page;
323 
324 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
325 	return page ? page_address(page) : NULL;
326 }
327 
328 static void dsfree_pages(const void *buffer, size_t size)
329 {
330 	if (buffer)
331 		free_pages((unsigned long)buffer, get_order(size));
332 }
333 
334 static int alloc_pebs_buffer(int cpu)
335 {
336 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
337 	struct debug_store *ds = hwev->ds;
338 	size_t bsiz = x86_pmu.pebs_buffer_size;
339 	int max, node = cpu_to_node(cpu);
340 	void *buffer, *ibuffer, *cea;
341 
342 	if (!x86_pmu.pebs)
343 		return 0;
344 
345 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
346 	if (unlikely(!buffer))
347 		return -ENOMEM;
348 
349 	/*
350 	 * HSW+ already provides us the eventing ip; no need to allocate this
351 	 * buffer then.
352 	 */
353 	if (x86_pmu.intel_cap.pebs_format < 2) {
354 		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
355 		if (!ibuffer) {
356 			dsfree_pages(buffer, bsiz);
357 			return -ENOMEM;
358 		}
359 		per_cpu(insn_buffer, cpu) = ibuffer;
360 	}
361 	hwev->ds_pebs_vaddr = buffer;
362 	/* Update the cpu entry area mapping */
363 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
364 	ds->pebs_buffer_base = (unsigned long) cea;
365 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
366 	ds->pebs_index = ds->pebs_buffer_base;
367 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
368 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
369 	return 0;
370 }
371 
372 static void release_pebs_buffer(int cpu)
373 {
374 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
375 	struct debug_store *ds = hwev->ds;
376 	void *cea;
377 
378 	if (!ds || !x86_pmu.pebs)
379 		return;
380 
381 	kfree(per_cpu(insn_buffer, cpu));
382 	per_cpu(insn_buffer, cpu) = NULL;
383 
384 	/* Clear the fixmap */
385 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
386 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
387 	ds->pebs_buffer_base = 0;
388 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
389 	hwev->ds_pebs_vaddr = NULL;
390 }
391 
392 static int alloc_bts_buffer(int cpu)
393 {
394 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
395 	struct debug_store *ds = hwev->ds;
396 	void *buffer, *cea;
397 	int max;
398 
399 	if (!x86_pmu.bts)
400 		return 0;
401 
402 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
403 	if (unlikely(!buffer)) {
404 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
405 		return -ENOMEM;
406 	}
407 	hwev->ds_bts_vaddr = buffer;
408 	/* Update the fixmap */
409 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
410 	ds->bts_buffer_base = (unsigned long) cea;
411 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
412 	ds->bts_index = ds->bts_buffer_base;
413 	max = BTS_RECORD_SIZE * (BTS_BUFFER_SIZE / BTS_RECORD_SIZE);
414 	ds->bts_absolute_maximum = ds->bts_buffer_base + max;
415 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum - (max / 16);
416 	return 0;
417 }
418 
419 static void release_bts_buffer(int cpu)
420 {
421 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
422 	struct debug_store *ds = hwev->ds;
423 	void *cea;
424 
425 	if (!ds || !x86_pmu.bts)
426 		return;
427 
428 	/* Clear the fixmap */
429 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
430 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
431 	ds->bts_buffer_base = 0;
432 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
433 	hwev->ds_bts_vaddr = NULL;
434 }
435 
436 static int alloc_ds_buffer(int cpu)
437 {
438 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
439 
440 	memset(ds, 0, sizeof(*ds));
441 	per_cpu(cpu_hw_events, cpu).ds = ds;
442 	return 0;
443 }
444 
445 static void release_ds_buffer(int cpu)
446 {
447 	per_cpu(cpu_hw_events, cpu).ds = NULL;
448 }
449 
450 void release_ds_buffers(void)
451 {
452 	int cpu;
453 
454 	if (!x86_pmu.bts && !x86_pmu.pebs)
455 		return;
456 
457 	get_online_cpus();
458 	for_each_online_cpu(cpu)
459 		fini_debug_store_on_cpu(cpu);
460 
461 	for_each_possible_cpu(cpu) {
462 		release_pebs_buffer(cpu);
463 		release_bts_buffer(cpu);
464 		release_ds_buffer(cpu);
465 	}
466 	put_online_cpus();
467 }
468 
469 void reserve_ds_buffers(void)
470 {
471 	int bts_err = 0, pebs_err = 0;
472 	int cpu;
473 
474 	x86_pmu.bts_active = 0;
475 	x86_pmu.pebs_active = 0;
476 
477 	if (!x86_pmu.bts && !x86_pmu.pebs)
478 		return;
479 
480 	if (!x86_pmu.bts)
481 		bts_err = 1;
482 
483 	if (!x86_pmu.pebs)
484 		pebs_err = 1;
485 
486 	get_online_cpus();
487 
488 	for_each_possible_cpu(cpu) {
489 		if (alloc_ds_buffer(cpu)) {
490 			bts_err = 1;
491 			pebs_err = 1;
492 		}
493 
494 		if (!bts_err && alloc_bts_buffer(cpu))
495 			bts_err = 1;
496 
497 		if (!pebs_err && alloc_pebs_buffer(cpu))
498 			pebs_err = 1;
499 
500 		if (bts_err && pebs_err)
501 			break;
502 	}
503 
504 	if (bts_err) {
505 		for_each_possible_cpu(cpu)
506 			release_bts_buffer(cpu);
507 	}
508 
509 	if (pebs_err) {
510 		for_each_possible_cpu(cpu)
511 			release_pebs_buffer(cpu);
512 	}
513 
514 	if (bts_err && pebs_err) {
515 		for_each_possible_cpu(cpu)
516 			release_ds_buffer(cpu);
517 	} else {
518 		if (x86_pmu.bts && !bts_err)
519 			x86_pmu.bts_active = 1;
520 
521 		if (x86_pmu.pebs && !pebs_err)
522 			x86_pmu.pebs_active = 1;
523 
524 		for_each_online_cpu(cpu)
525 			init_debug_store_on_cpu(cpu);
526 	}
527 
528 	put_online_cpus();
529 }
530 
531 /*
532  * BTS
533  */
534 
535 struct event_constraint bts_constraint =
536 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
537 
538 void intel_pmu_enable_bts(u64 config)
539 {
540 	unsigned long debugctlmsr;
541 
542 	debugctlmsr = get_debugctlmsr();
543 
544 	debugctlmsr |= DEBUGCTLMSR_TR;
545 	debugctlmsr |= DEBUGCTLMSR_BTS;
546 	if (config & ARCH_PERFMON_EVENTSEL_INT)
547 		debugctlmsr |= DEBUGCTLMSR_BTINT;
548 
549 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
550 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
551 
552 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
553 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
554 
555 	update_debugctlmsr(debugctlmsr);
556 }
557 
558 void intel_pmu_disable_bts(void)
559 {
560 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
561 	unsigned long debugctlmsr;
562 
563 	if (!cpuc->ds)
564 		return;
565 
566 	debugctlmsr = get_debugctlmsr();
567 
568 	debugctlmsr &=
569 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
570 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
571 
572 	update_debugctlmsr(debugctlmsr);
573 }
574 
575 int intel_pmu_drain_bts_buffer(void)
576 {
577 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
578 	struct debug_store *ds = cpuc->ds;
579 	struct bts_record {
580 		u64	from;
581 		u64	to;
582 		u64	flags;
583 	};
584 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
585 	struct bts_record *at, *base, *top;
586 	struct perf_output_handle handle;
587 	struct perf_event_header header;
588 	struct perf_sample_data data;
589 	unsigned long skip = 0;
590 	struct pt_regs regs;
591 
592 	if (!event)
593 		return 0;
594 
595 	if (!x86_pmu.bts_active)
596 		return 0;
597 
598 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
599 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
600 
601 	if (top <= base)
602 		return 0;
603 
604 	memset(&regs, 0, sizeof(regs));
605 
606 	ds->bts_index = ds->bts_buffer_base;
607 
608 	perf_sample_data_init(&data, 0, event->hw.last_period);
609 
610 	/*
611 	 * BTS leaks kernel addresses in branches across the cpl boundary,
612 	 * such as traps or system calls, so unless the user is asking for
613 	 * kernel tracing (and right now it's not possible), we'd need to
614 	 * filter them out. But first we need to count how many of those we
615 	 * have in the current batch. This is an extra O(n) pass, however,
616 	 * it's much faster than the other one especially considering that
617 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
618 	 * alloc_bts_buffer()).
619 	 */
620 	for (at = base; at < top; at++) {
621 		/*
622 		 * Note that right now *this* BTS code only works if
623 		 * attr::exclude_kernel is set, but let's keep this extra
624 		 * check here in case that changes.
625 		 */
626 		if (event->attr.exclude_kernel &&
627 		    (kernel_ip(at->from) || kernel_ip(at->to)))
628 			skip++;
629 	}
630 
631 	/*
632 	 * Prepare a generic sample, i.e. fill in the invariant fields.
633 	 * We will overwrite the from and to address before we output
634 	 * the sample.
635 	 */
636 	rcu_read_lock();
637 	perf_prepare_sample(&header, &data, event, &regs);
638 
639 	if (perf_output_begin(&handle, event, header.size *
640 			      (top - base - skip)))
641 		goto unlock;
642 
643 	for (at = base; at < top; at++) {
644 		/* Filter out any records that contain kernel addresses. */
645 		if (event->attr.exclude_kernel &&
646 		    (kernel_ip(at->from) || kernel_ip(at->to)))
647 			continue;
648 
649 		data.ip		= at->from;
650 		data.addr	= at->to;
651 
652 		perf_output_sample(&handle, &header, &data, event);
653 	}
654 
655 	perf_output_end(&handle);
656 
657 	/* There's new data available. */
658 	event->hw.interrupts++;
659 	event->pending_kill = POLL_IN;
660 unlock:
661 	rcu_read_unlock();
662 	return 1;
663 }
664 
665 static inline void intel_pmu_drain_pebs_buffer(void)
666 {
667 	struct pt_regs regs;
668 
669 	x86_pmu.drain_pebs(&regs);
670 }
671 
672 /*
673  * PEBS
674  */
675 struct event_constraint intel_core2_pebs_event_constraints[] = {
676 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
677 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
678 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
679 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
680 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
681 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
682 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
683 	EVENT_CONSTRAINT_END
684 };
685 
686 struct event_constraint intel_atom_pebs_event_constraints[] = {
687 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
688 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
689 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
690 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
691 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x01),
692 	/* Allow all events as PEBS with no flags */
693 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
694 	EVENT_CONSTRAINT_END
695 };
696 
697 struct event_constraint intel_slm_pebs_event_constraints[] = {
698 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
699 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x1),
700 	/* Allow all events as PEBS with no flags */
701 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
702 	EVENT_CONSTRAINT_END
703 };
704 
705 struct event_constraint intel_glm_pebs_event_constraints[] = {
706 	/* Allow all events as PEBS with no flags */
707 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
708 	EVENT_CONSTRAINT_END
709 };
710 
711 struct event_constraint intel_glp_pebs_event_constraints[] = {
712 	/* Allow all events as PEBS with no flags */
713 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
714 	EVENT_CONSTRAINT_END
715 };
716 
717 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
718 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
719 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
720 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
721 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
722 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
723 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
724 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
725 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
726 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
727 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
728 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
729 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
730 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
731 	EVENT_CONSTRAINT_END
732 };
733 
734 struct event_constraint intel_westmere_pebs_event_constraints[] = {
735 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
736 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
737 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
738 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
739 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
740 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
741 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
742 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
743 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
744 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
745 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
746 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
747 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
748 	EVENT_CONSTRAINT_END
749 };
750 
751 struct event_constraint intel_snb_pebs_event_constraints[] = {
752 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
753 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
754 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
755 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
756 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
757         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
758         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
759         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
760         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
761 	/* Allow all events as PEBS with no flags */
762 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
763 	EVENT_CONSTRAINT_END
764 };
765 
766 struct event_constraint intel_ivb_pebs_event_constraints[] = {
767         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
768         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
769 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
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_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
775 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
776 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
777 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
778 	/* Allow all events as PEBS with no flags */
779 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
780         EVENT_CONSTRAINT_END
781 };
782 
783 struct event_constraint intel_hsw_pebs_event_constraints[] = {
784 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
785 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
786 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
787 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
788 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
789 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
790 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
791 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
792 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
793 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
794 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
795 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
796 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
797 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
798 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
799 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
800 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
801 	/* Allow all events as PEBS with no flags */
802 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
803 	EVENT_CONSTRAINT_END
804 };
805 
806 struct event_constraint intel_bdw_pebs_event_constraints[] = {
807 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
808 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
809 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
810 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c2, 0xf),
811 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
812 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
813 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
814 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
815 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
816 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
817 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
818 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
819 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
820 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
821 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
822 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
823 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
824 	/* Allow all events as PEBS with no flags */
825 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
826 	EVENT_CONSTRAINT_END
827 };
828 
829 
830 struct event_constraint intel_skl_pebs_event_constraints[] = {
831 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
832 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
833 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108001c0, 0x2),
834 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
835 	INTEL_FLAGS_EVENT_CONSTRAINT(0x108000c0, 0x0f),
836 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
837 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
838 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
839 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
840 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
841 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
842 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
843 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
844 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
845 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
846 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
847 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
848 	/* Allow all events as PEBS with no flags */
849 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
850 	EVENT_CONSTRAINT_END
851 };
852 
853 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
854 {
855 	struct event_constraint *c;
856 
857 	if (!event->attr.precise_ip)
858 		return NULL;
859 
860 	if (x86_pmu.pebs_constraints) {
861 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
862 			if ((event->hw.config & c->cmask) == c->code) {
863 				event->hw.flags |= c->flags;
864 				return c;
865 			}
866 		}
867 	}
868 
869 	return &emptyconstraint;
870 }
871 
872 /*
873  * We need the sched_task callback even for per-cpu events when we use
874  * the large interrupt threshold, such that we can provide PID and TID
875  * to PEBS samples.
876  */
877 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
878 {
879 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
880 }
881 
882 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
883 {
884 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
885 
886 	if (!sched_in && pebs_needs_sched_cb(cpuc))
887 		intel_pmu_drain_pebs_buffer();
888 }
889 
890 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
891 {
892 	struct debug_store *ds = cpuc->ds;
893 	u64 threshold;
894 
895 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
896 		threshold = ds->pebs_absolute_maximum -
897 			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
898 	} else {
899 		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
900 	}
901 
902 	ds->pebs_interrupt_threshold = threshold;
903 }
904 
905 static void
906 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
907 {
908 	/*
909 	 * Make sure we get updated with the first PEBS
910 	 * event. It will trigger also during removal, but
911 	 * that does not hurt:
912 	 */
913 	bool update = cpuc->n_pebs == 1;
914 
915 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
916 		if (!needed_cb)
917 			perf_sched_cb_inc(pmu);
918 		else
919 			perf_sched_cb_dec(pmu);
920 
921 		update = true;
922 	}
923 
924 	if (update)
925 		pebs_update_threshold(cpuc);
926 }
927 
928 void intel_pmu_pebs_add(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 	bool needed_cb = pebs_needs_sched_cb(cpuc);
933 
934 	cpuc->n_pebs++;
935 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
936 		cpuc->n_large_pebs++;
937 
938 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
939 }
940 
941 void intel_pmu_pebs_enable(struct perf_event *event)
942 {
943 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
944 	struct hw_perf_event *hwc = &event->hw;
945 	struct debug_store *ds = cpuc->ds;
946 
947 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
948 
949 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
950 
951 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
952 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
953 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
954 		cpuc->pebs_enabled |= 1ULL << 63;
955 
956 	/*
957 	 * Use auto-reload if possible to save a MSR write in the PMI.
958 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
959 	 */
960 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
961 		ds->pebs_event_reset[hwc->idx] =
962 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
963 	} else {
964 		ds->pebs_event_reset[hwc->idx] = 0;
965 	}
966 }
967 
968 void intel_pmu_pebs_del(struct perf_event *event)
969 {
970 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
971 	struct hw_perf_event *hwc = &event->hw;
972 	bool needed_cb = pebs_needs_sched_cb(cpuc);
973 
974 	cpuc->n_pebs--;
975 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
976 		cpuc->n_large_pebs--;
977 
978 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
979 }
980 
981 void intel_pmu_pebs_disable(struct perf_event *event)
982 {
983 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
984 	struct hw_perf_event *hwc = &event->hw;
985 
986 	if (cpuc->n_pebs == cpuc->n_large_pebs)
987 		intel_pmu_drain_pebs_buffer();
988 
989 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
990 
991 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
992 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
993 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
994 		cpuc->pebs_enabled &= ~(1ULL << 63);
995 
996 	if (cpuc->enabled)
997 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
998 
999 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1000 }
1001 
1002 void intel_pmu_pebs_enable_all(void)
1003 {
1004 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1005 
1006 	if (cpuc->pebs_enabled)
1007 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1008 }
1009 
1010 void intel_pmu_pebs_disable_all(void)
1011 {
1012 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1013 
1014 	if (cpuc->pebs_enabled)
1015 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1016 }
1017 
1018 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1019 {
1020 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1021 	unsigned long from = cpuc->lbr_entries[0].from;
1022 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1023 	unsigned long ip = regs->ip;
1024 	int is_64bit = 0;
1025 	void *kaddr;
1026 	int size;
1027 
1028 	/*
1029 	 * We don't need to fixup if the PEBS assist is fault like
1030 	 */
1031 	if (!x86_pmu.intel_cap.pebs_trap)
1032 		return 1;
1033 
1034 	/*
1035 	 * No LBR entry, no basic block, no rewinding
1036 	 */
1037 	if (!cpuc->lbr_stack.nr || !from || !to)
1038 		return 0;
1039 
1040 	/*
1041 	 * Basic blocks should never cross user/kernel boundaries
1042 	 */
1043 	if (kernel_ip(ip) != kernel_ip(to))
1044 		return 0;
1045 
1046 	/*
1047 	 * unsigned math, either ip is before the start (impossible) or
1048 	 * the basic block is larger than 1 page (sanity)
1049 	 */
1050 	if ((ip - to) > PEBS_FIXUP_SIZE)
1051 		return 0;
1052 
1053 	/*
1054 	 * We sampled a branch insn, rewind using the LBR stack
1055 	 */
1056 	if (ip == to) {
1057 		set_linear_ip(regs, from);
1058 		return 1;
1059 	}
1060 
1061 	size = ip - to;
1062 	if (!kernel_ip(ip)) {
1063 		int bytes;
1064 		u8 *buf = this_cpu_read(insn_buffer);
1065 
1066 		/* 'size' must fit our buffer, see above */
1067 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1068 		if (bytes != 0)
1069 			return 0;
1070 
1071 		kaddr = buf;
1072 	} else {
1073 		kaddr = (void *)to;
1074 	}
1075 
1076 	do {
1077 		struct insn insn;
1078 
1079 		old_to = to;
1080 
1081 #ifdef CONFIG_X86_64
1082 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1083 #endif
1084 		insn_init(&insn, kaddr, size, is_64bit);
1085 		insn_get_length(&insn);
1086 		/*
1087 		 * Make sure there was not a problem decoding the
1088 		 * instruction and getting the length.  This is
1089 		 * doubly important because we have an infinite
1090 		 * loop if insn.length=0.
1091 		 */
1092 		if (!insn.length)
1093 			break;
1094 
1095 		to += insn.length;
1096 		kaddr += insn.length;
1097 		size -= insn.length;
1098 	} while (to < ip);
1099 
1100 	if (to == ip) {
1101 		set_linear_ip(regs, old_to);
1102 		return 1;
1103 	}
1104 
1105 	/*
1106 	 * Even though we decoded the basic block, the instruction stream
1107 	 * never matched the given IP, either the TO or the IP got corrupted.
1108 	 */
1109 	return 0;
1110 }
1111 
1112 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1113 {
1114 	if (pebs->tsx_tuning) {
1115 		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1116 		return tsx.cycles_last_block;
1117 	}
1118 	return 0;
1119 }
1120 
1121 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1122 {
1123 	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1124 
1125 	/* For RTM XABORTs also log the abort code from AX */
1126 	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1127 		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1128 	return txn;
1129 }
1130 
1131 static void setup_pebs_sample_data(struct perf_event *event,
1132 				   struct pt_regs *iregs, void *__pebs,
1133 				   struct perf_sample_data *data,
1134 				   struct pt_regs *regs)
1135 {
1136 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1137 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1138 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1139 		 PERF_X86_EVENT_PEBS_NA_HSW)
1140 	/*
1141 	 * We cast to the biggest pebs_record but are careful not to
1142 	 * unconditionally access the 'extra' entries.
1143 	 */
1144 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1145 	struct pebs_record_skl *pebs = __pebs;
1146 	u64 sample_type;
1147 	int fll, fst, dsrc;
1148 	int fl = event->hw.flags;
1149 
1150 	if (pebs == NULL)
1151 		return;
1152 
1153 	sample_type = event->attr.sample_type;
1154 	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1155 
1156 	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1157 	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1158 
1159 	perf_sample_data_init(data, 0, event->hw.last_period);
1160 
1161 	data->period = event->hw.last_period;
1162 
1163 	/*
1164 	 * Use latency for weight (only avail with PEBS-LL)
1165 	 */
1166 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1167 		data->weight = pebs->lat;
1168 
1169 	/*
1170 	 * data.data_src encodes the data source
1171 	 */
1172 	if (dsrc) {
1173 		u64 val = PERF_MEM_NA;
1174 		if (fll)
1175 			val = load_latency_data(pebs->dse);
1176 		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1177 			val = precise_datala_hsw(event, pebs->dse);
1178 		else if (fst)
1179 			val = precise_store_data(pebs->dse);
1180 		data->data_src.val = val;
1181 	}
1182 
1183 	/*
1184 	 * We use the interrupt regs as a base because the PEBS record does not
1185 	 * contain a full regs set, specifically it seems to lack segment
1186 	 * descriptors, which get used by things like user_mode().
1187 	 *
1188 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1189 	 *
1190 	 * We must however always use BP,SP from iregs for the unwinder to stay
1191 	 * sane; the record BP,SP can point into thin air when the record is
1192 	 * from a previous PMI context or an (I)RET happend between the record
1193 	 * and PMI.
1194 	 */
1195 	*regs = *iregs;
1196 	regs->flags = pebs->flags;
1197 	set_linear_ip(regs, pebs->ip);
1198 
1199 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1200 		regs->ax = pebs->ax;
1201 		regs->bx = pebs->bx;
1202 		regs->cx = pebs->cx;
1203 		regs->dx = pebs->dx;
1204 		regs->si = pebs->si;
1205 		regs->di = pebs->di;
1206 
1207 		/*
1208 		 * Per the above; only set BP,SP if we don't need callchains.
1209 		 *
1210 		 * XXX: does this make sense?
1211 		 */
1212 		if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1213 			regs->bp = pebs->bp;
1214 			regs->sp = pebs->sp;
1215 		}
1216 
1217 		/*
1218 		 * Preserve PERF_EFLAGS_VM from set_linear_ip().
1219 		 */
1220 		regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1221 #ifndef CONFIG_X86_32
1222 		regs->r8 = pebs->r8;
1223 		regs->r9 = pebs->r9;
1224 		regs->r10 = pebs->r10;
1225 		regs->r11 = pebs->r11;
1226 		regs->r12 = pebs->r12;
1227 		regs->r13 = pebs->r13;
1228 		regs->r14 = pebs->r14;
1229 		regs->r15 = pebs->r15;
1230 #endif
1231 	}
1232 
1233 	if (event->attr.precise_ip > 1 && x86_pmu.intel_cap.pebs_format >= 2) {
1234 		regs->ip = pebs->real_ip;
1235 		regs->flags |= PERF_EFLAGS_EXACT;
1236 	} else if (event->attr.precise_ip > 1 && intel_pmu_pebs_fixup_ip(regs))
1237 		regs->flags |= PERF_EFLAGS_EXACT;
1238 	else
1239 		regs->flags &= ~PERF_EFLAGS_EXACT;
1240 
1241 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1242 	    x86_pmu.intel_cap.pebs_format >= 1)
1243 		data->addr = pebs->dla;
1244 
1245 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1246 		/* Only set the TSX weight when no memory weight. */
1247 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1248 			data->weight = intel_hsw_weight(pebs);
1249 
1250 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1251 			data->txn = intel_hsw_transaction(pebs);
1252 	}
1253 
1254 	/*
1255 	 * v3 supplies an accurate time stamp, so we use that
1256 	 * for the time stamp.
1257 	 *
1258 	 * We can only do this for the default trace clock.
1259 	 */
1260 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1261 		event->attr.use_clockid == 0)
1262 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1263 
1264 	if (has_branch_stack(event))
1265 		data->br_stack = &cpuc->lbr_stack;
1266 }
1267 
1268 static inline void *
1269 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1270 {
1271 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1272 	void *at;
1273 	u64 pebs_status;
1274 
1275 	/*
1276 	 * fmt0 does not have a status bitfield (does not use
1277 	 * perf_record_nhm format)
1278 	 */
1279 	if (x86_pmu.intel_cap.pebs_format < 1)
1280 		return base;
1281 
1282 	if (base == NULL)
1283 		return NULL;
1284 
1285 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1286 		struct pebs_record_nhm *p = at;
1287 
1288 		if (test_bit(bit, (unsigned long *)&p->status)) {
1289 			/* PEBS v3 has accurate status bits */
1290 			if (x86_pmu.intel_cap.pebs_format >= 3)
1291 				return at;
1292 
1293 			if (p->status == (1 << bit))
1294 				return at;
1295 
1296 			/* clear non-PEBS bit and re-check */
1297 			pebs_status = p->status & cpuc->pebs_enabled;
1298 			pebs_status &= PEBS_COUNTER_MASK;
1299 			if (pebs_status == (1 << bit))
1300 				return at;
1301 		}
1302 	}
1303 	return NULL;
1304 }
1305 
1306 static void __intel_pmu_pebs_event(struct perf_event *event,
1307 				   struct pt_regs *iregs,
1308 				   void *base, void *top,
1309 				   int bit, int count)
1310 {
1311 	struct perf_sample_data data;
1312 	struct pt_regs regs;
1313 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1314 
1315 	if (!intel_pmu_save_and_restart(event) &&
1316 	    !(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD))
1317 		return;
1318 
1319 	while (count > 1) {
1320 		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1321 		perf_event_output(event, &data, &regs);
1322 		at += x86_pmu.pebs_record_size;
1323 		at = get_next_pebs_record_by_bit(at, top, bit);
1324 		count--;
1325 	}
1326 
1327 	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1328 
1329 	/*
1330 	 * All but the last records are processed.
1331 	 * The last one is left to be able to call the overflow handler.
1332 	 */
1333 	if (perf_event_overflow(event, &data, &regs)) {
1334 		x86_pmu_stop(event, 0);
1335 		return;
1336 	}
1337 
1338 }
1339 
1340 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1341 {
1342 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1343 	struct debug_store *ds = cpuc->ds;
1344 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1345 	struct pebs_record_core *at, *top;
1346 	int n;
1347 
1348 	if (!x86_pmu.pebs_active)
1349 		return;
1350 
1351 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1352 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1353 
1354 	/*
1355 	 * Whatever else happens, drain the thing
1356 	 */
1357 	ds->pebs_index = ds->pebs_buffer_base;
1358 
1359 	if (!test_bit(0, cpuc->active_mask))
1360 		return;
1361 
1362 	WARN_ON_ONCE(!event);
1363 
1364 	if (!event->attr.precise_ip)
1365 		return;
1366 
1367 	n = top - at;
1368 	if (n <= 0)
1369 		return;
1370 
1371 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1372 }
1373 
1374 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1375 {
1376 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1377 	struct debug_store *ds = cpuc->ds;
1378 	struct perf_event *event;
1379 	void *base, *at, *top;
1380 	short counts[MAX_PEBS_EVENTS] = {};
1381 	short error[MAX_PEBS_EVENTS] = {};
1382 	int bit, i;
1383 
1384 	if (!x86_pmu.pebs_active)
1385 		return;
1386 
1387 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1388 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1389 
1390 	ds->pebs_index = ds->pebs_buffer_base;
1391 
1392 	if (unlikely(base >= top))
1393 		return;
1394 
1395 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1396 		struct pebs_record_nhm *p = at;
1397 		u64 pebs_status;
1398 
1399 		pebs_status = p->status & cpuc->pebs_enabled;
1400 		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1401 
1402 		/* PEBS v3 has more accurate status bits */
1403 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1404 			for_each_set_bit(bit, (unsigned long *)&pebs_status,
1405 					 x86_pmu.max_pebs_events)
1406 				counts[bit]++;
1407 
1408 			continue;
1409 		}
1410 
1411 		/*
1412 		 * On some CPUs the PEBS status can be zero when PEBS is
1413 		 * racing with clearing of GLOBAL_STATUS.
1414 		 *
1415 		 * Normally we would drop that record, but in the
1416 		 * case when there is only a single active PEBS event
1417 		 * we can assume it's for that event.
1418 		 */
1419 		if (!pebs_status && cpuc->pebs_enabled &&
1420 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1421 			pebs_status = cpuc->pebs_enabled;
1422 
1423 		bit = find_first_bit((unsigned long *)&pebs_status,
1424 					x86_pmu.max_pebs_events);
1425 		if (bit >= x86_pmu.max_pebs_events)
1426 			continue;
1427 
1428 		/*
1429 		 * The PEBS hardware does not deal well with the situation
1430 		 * when events happen near to each other and multiple bits
1431 		 * are set. But it should happen rarely.
1432 		 *
1433 		 * If these events include one PEBS and multiple non-PEBS
1434 		 * events, it doesn't impact PEBS record. The record will
1435 		 * be handled normally. (slow path)
1436 		 *
1437 		 * If these events include two or more PEBS events, the
1438 		 * records for the events can be collapsed into a single
1439 		 * one, and it's not possible to reconstruct all events
1440 		 * that caused the PEBS record. It's called collision.
1441 		 * If collision happened, the record will be dropped.
1442 		 */
1443 		if (p->status != (1ULL << bit)) {
1444 			for_each_set_bit(i, (unsigned long *)&pebs_status,
1445 					 x86_pmu.max_pebs_events)
1446 				error[i]++;
1447 			continue;
1448 		}
1449 
1450 		counts[bit]++;
1451 	}
1452 
1453 	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1454 		if ((counts[bit] == 0) && (error[bit] == 0))
1455 			continue;
1456 
1457 		event = cpuc->events[bit];
1458 		if (WARN_ON_ONCE(!event))
1459 			continue;
1460 
1461 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1462 			continue;
1463 
1464 		/* log dropped samples number */
1465 		if (error[bit]) {
1466 			perf_log_lost_samples(event, error[bit]);
1467 
1468 			if (perf_event_account_interrupt(event))
1469 				x86_pmu_stop(event, 0);
1470 		}
1471 
1472 		if (counts[bit]) {
1473 			__intel_pmu_pebs_event(event, iregs, base,
1474 					       top, bit, counts[bit]);
1475 		}
1476 	}
1477 }
1478 
1479 /*
1480  * BTS, PEBS probe and setup
1481  */
1482 
1483 void __init intel_ds_init(void)
1484 {
1485 	/*
1486 	 * No support for 32bit formats
1487 	 */
1488 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1489 		return;
1490 
1491 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1492 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1493 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1494 	if (x86_pmu.pebs) {
1495 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1496 		int format = x86_pmu.intel_cap.pebs_format;
1497 
1498 		switch (format) {
1499 		case 0:
1500 			pr_cont("PEBS fmt0%c, ", pebs_type);
1501 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1502 			/*
1503 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1504 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1505 			 * mysteriously hang on Core2.
1506 			 *
1507 			 * As a workaround, we don't do this.
1508 			 */
1509 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1510 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1511 			break;
1512 
1513 		case 1:
1514 			pr_cont("PEBS fmt1%c, ", pebs_type);
1515 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1516 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1517 			break;
1518 
1519 		case 2:
1520 			pr_cont("PEBS fmt2%c, ", pebs_type);
1521 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1522 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1523 			break;
1524 
1525 		case 3:
1526 			pr_cont("PEBS fmt3%c, ", pebs_type);
1527 			x86_pmu.pebs_record_size =
1528 						sizeof(struct pebs_record_skl);
1529 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1530 			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1531 			break;
1532 
1533 		default:
1534 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1535 			x86_pmu.pebs = 0;
1536 		}
1537 	}
1538 }
1539 
1540 void perf_restore_debug_store(void)
1541 {
1542 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1543 
1544 	if (!x86_pmu.bts && !x86_pmu.pebs)
1545 		return;
1546 
1547 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1548 }
1549