xref: /linux/arch/x86/events/intel/ds.c (revision b0f84a84fff180718995b1269da2988e5b28be42)
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 	void *cea;
376 
377 	if (!x86_pmu.pebs)
378 		return;
379 
380 	kfree(per_cpu(insn_buffer, cpu));
381 	per_cpu(insn_buffer, cpu) = NULL;
382 
383 	/* Clear the fixmap */
384 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
385 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
386 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
387 	hwev->ds_pebs_vaddr = NULL;
388 }
389 
390 static int alloc_bts_buffer(int cpu)
391 {
392 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
393 	struct debug_store *ds = hwev->ds;
394 	void *buffer, *cea;
395 	int max;
396 
397 	if (!x86_pmu.bts)
398 		return 0;
399 
400 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
401 	if (unlikely(!buffer)) {
402 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
403 		return -ENOMEM;
404 	}
405 	hwev->ds_bts_vaddr = buffer;
406 	/* Update the fixmap */
407 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
408 	ds->bts_buffer_base = (unsigned long) cea;
409 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
410 	ds->bts_index = ds->bts_buffer_base;
411 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
412 	ds->bts_absolute_maximum = ds->bts_buffer_base +
413 					max * BTS_RECORD_SIZE;
414 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
415 					(max / 16) * BTS_RECORD_SIZE;
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 	void *cea;
423 
424 	if (!x86_pmu.bts)
425 		return;
426 
427 	/* Clear the fixmap */
428 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
429 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
430 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
431 	hwev->ds_bts_vaddr = NULL;
432 }
433 
434 static int alloc_ds_buffer(int cpu)
435 {
436 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
437 
438 	memset(ds, 0, sizeof(*ds));
439 	per_cpu(cpu_hw_events, cpu).ds = ds;
440 	return 0;
441 }
442 
443 static void release_ds_buffer(int cpu)
444 {
445 	per_cpu(cpu_hw_events, cpu).ds = NULL;
446 }
447 
448 void release_ds_buffers(void)
449 {
450 	int cpu;
451 
452 	if (!x86_pmu.bts && !x86_pmu.pebs)
453 		return;
454 
455 	for_each_possible_cpu(cpu)
456 		release_ds_buffer(cpu);
457 
458 	for_each_possible_cpu(cpu) {
459 		/*
460 		 * Again, ignore errors from offline CPUs, they will no longer
461 		 * observe cpu_hw_events.ds and not program the DS_AREA when
462 		 * they come up.
463 		 */
464 		fini_debug_store_on_cpu(cpu);
465 	}
466 
467 	for_each_possible_cpu(cpu) {
468 		release_pebs_buffer(cpu);
469 		release_bts_buffer(cpu);
470 	}
471 }
472 
473 void reserve_ds_buffers(void)
474 {
475 	int bts_err = 0, pebs_err = 0;
476 	int cpu;
477 
478 	x86_pmu.bts_active = 0;
479 	x86_pmu.pebs_active = 0;
480 
481 	if (!x86_pmu.bts && !x86_pmu.pebs)
482 		return;
483 
484 	if (!x86_pmu.bts)
485 		bts_err = 1;
486 
487 	if (!x86_pmu.pebs)
488 		pebs_err = 1;
489 
490 	for_each_possible_cpu(cpu) {
491 		if (alloc_ds_buffer(cpu)) {
492 			bts_err = 1;
493 			pebs_err = 1;
494 		}
495 
496 		if (!bts_err && alloc_bts_buffer(cpu))
497 			bts_err = 1;
498 
499 		if (!pebs_err && alloc_pebs_buffer(cpu))
500 			pebs_err = 1;
501 
502 		if (bts_err && pebs_err)
503 			break;
504 	}
505 
506 	if (bts_err) {
507 		for_each_possible_cpu(cpu)
508 			release_bts_buffer(cpu);
509 	}
510 
511 	if (pebs_err) {
512 		for_each_possible_cpu(cpu)
513 			release_pebs_buffer(cpu);
514 	}
515 
516 	if (bts_err && pebs_err) {
517 		for_each_possible_cpu(cpu)
518 			release_ds_buffer(cpu);
519 	} else {
520 		if (x86_pmu.bts && !bts_err)
521 			x86_pmu.bts_active = 1;
522 
523 		if (x86_pmu.pebs && !pebs_err)
524 			x86_pmu.pebs_active = 1;
525 
526 		for_each_possible_cpu(cpu) {
527 			/*
528 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
529 			 * will get this call through intel_pmu_cpu_starting().
530 			 */
531 			init_debug_store_on_cpu(cpu);
532 		}
533 	}
534 }
535 
536 /*
537  * BTS
538  */
539 
540 struct event_constraint bts_constraint =
541 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
542 
543 void intel_pmu_enable_bts(u64 config)
544 {
545 	unsigned long debugctlmsr;
546 
547 	debugctlmsr = get_debugctlmsr();
548 
549 	debugctlmsr |= DEBUGCTLMSR_TR;
550 	debugctlmsr |= DEBUGCTLMSR_BTS;
551 	if (config & ARCH_PERFMON_EVENTSEL_INT)
552 		debugctlmsr |= DEBUGCTLMSR_BTINT;
553 
554 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
555 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
556 
557 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
558 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
559 
560 	update_debugctlmsr(debugctlmsr);
561 }
562 
563 void intel_pmu_disable_bts(void)
564 {
565 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
566 	unsigned long debugctlmsr;
567 
568 	if (!cpuc->ds)
569 		return;
570 
571 	debugctlmsr = get_debugctlmsr();
572 
573 	debugctlmsr &=
574 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
575 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
576 
577 	update_debugctlmsr(debugctlmsr);
578 }
579 
580 int intel_pmu_drain_bts_buffer(void)
581 {
582 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
583 	struct debug_store *ds = cpuc->ds;
584 	struct bts_record {
585 		u64	from;
586 		u64	to;
587 		u64	flags;
588 	};
589 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
590 	struct bts_record *at, *base, *top;
591 	struct perf_output_handle handle;
592 	struct perf_event_header header;
593 	struct perf_sample_data data;
594 	unsigned long skip = 0;
595 	struct pt_regs regs;
596 
597 	if (!event)
598 		return 0;
599 
600 	if (!x86_pmu.bts_active)
601 		return 0;
602 
603 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
604 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
605 
606 	if (top <= base)
607 		return 0;
608 
609 	memset(&regs, 0, sizeof(regs));
610 
611 	ds->bts_index = ds->bts_buffer_base;
612 
613 	perf_sample_data_init(&data, 0, event->hw.last_period);
614 
615 	/*
616 	 * BTS leaks kernel addresses in branches across the cpl boundary,
617 	 * such as traps or system calls, so unless the user is asking for
618 	 * kernel tracing (and right now it's not possible), we'd need to
619 	 * filter them out. But first we need to count how many of those we
620 	 * have in the current batch. This is an extra O(n) pass, however,
621 	 * it's much faster than the other one especially considering that
622 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
623 	 * alloc_bts_buffer()).
624 	 */
625 	for (at = base; at < top; at++) {
626 		/*
627 		 * Note that right now *this* BTS code only works if
628 		 * attr::exclude_kernel is set, but let's keep this extra
629 		 * check here in case that changes.
630 		 */
631 		if (event->attr.exclude_kernel &&
632 		    (kernel_ip(at->from) || kernel_ip(at->to)))
633 			skip++;
634 	}
635 
636 	/*
637 	 * Prepare a generic sample, i.e. fill in the invariant fields.
638 	 * We will overwrite the from and to address before we output
639 	 * the sample.
640 	 */
641 	rcu_read_lock();
642 	perf_prepare_sample(&header, &data, event, &regs);
643 
644 	if (perf_output_begin(&handle, event, header.size *
645 			      (top - base - skip)))
646 		goto unlock;
647 
648 	for (at = base; at < top; at++) {
649 		/* Filter out any records that contain kernel addresses. */
650 		if (event->attr.exclude_kernel &&
651 		    (kernel_ip(at->from) || kernel_ip(at->to)))
652 			continue;
653 
654 		data.ip		= at->from;
655 		data.addr	= at->to;
656 
657 		perf_output_sample(&handle, &header, &data, event);
658 	}
659 
660 	perf_output_end(&handle);
661 
662 	/* There's new data available. */
663 	event->hw.interrupts++;
664 	event->pending_kill = POLL_IN;
665 unlock:
666 	rcu_read_unlock();
667 	return 1;
668 }
669 
670 static inline void intel_pmu_drain_pebs_buffer(void)
671 {
672 	struct pt_regs regs;
673 
674 	x86_pmu.drain_pebs(&regs);
675 }
676 
677 /*
678  * PEBS
679  */
680 struct event_constraint intel_core2_pebs_event_constraints[] = {
681 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
682 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
683 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
684 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
685 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
686 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
687 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
688 	EVENT_CONSTRAINT_END
689 };
690 
691 struct event_constraint intel_atom_pebs_event_constraints[] = {
692 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
693 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
694 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
695 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
696 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
697 	/* Allow all events as PEBS with no flags */
698 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
699 	EVENT_CONSTRAINT_END
700 };
701 
702 struct event_constraint intel_slm_pebs_event_constraints[] = {
703 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
704 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
705 	/* Allow all events as PEBS with no flags */
706 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
707 	EVENT_CONSTRAINT_END
708 };
709 
710 struct event_constraint intel_glm_pebs_event_constraints[] = {
711 	/* Allow all events as PEBS with no flags */
712 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
713 	EVENT_CONSTRAINT_END
714 };
715 
716 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
717 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
718 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
719 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
720 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
721 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
722 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
723 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
724 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
725 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
726 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
727 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
728 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
729 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
730 	EVENT_CONSTRAINT_END
731 };
732 
733 struct event_constraint intel_westmere_pebs_event_constraints[] = {
734 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
735 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
736 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
737 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
738 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
739 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
740 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
741 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
742 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
743 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
744 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
745 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
746 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
747 	EVENT_CONSTRAINT_END
748 };
749 
750 struct event_constraint intel_snb_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_UEVENT_CONSTRAINT(0x108001c2, 0xf),
756         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
757         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
758         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
759         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
760 	/* Allow all events as PEBS with no flags */
761 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
762 	EVENT_CONSTRAINT_END
763 };
764 
765 struct event_constraint intel_ivb_pebs_event_constraints[] = {
766         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
767         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
768 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
769 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
770 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
771 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
772 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
773 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
774 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
775 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
776 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
777 	/* Allow all events as PEBS with no flags */
778 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
779         EVENT_CONSTRAINT_END
780 };
781 
782 struct event_constraint intel_hsw_pebs_event_constraints[] = {
783 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
784 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
785 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
786 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
787 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
788 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
789 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
790 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
791 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
792 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
793 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
794 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
795 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
796 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
797 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
798 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
799 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
800 	/* Allow all events as PEBS with no flags */
801 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
802 	EVENT_CONSTRAINT_END
803 };
804 
805 struct event_constraint intel_bdw_pebs_event_constraints[] = {
806 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
807 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
808 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
809 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
810 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
811 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
812 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
813 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
814 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
815 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
816 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
817 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
818 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
819 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
820 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
821 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
822 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
823 	/* Allow all events as PEBS with no flags */
824 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
825 	EVENT_CONSTRAINT_END
826 };
827 
828 
829 struct event_constraint intel_skl_pebs_event_constraints[] = {
830 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
831 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
832 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
833 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
834 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
835 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
836 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
837 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
838 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
839 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
840 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
841 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
842 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
843 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
844 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
845 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
846 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
847 	/* Allow all events as PEBS with no flags */
848 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
849 	EVENT_CONSTRAINT_END
850 };
851 
852 struct event_constraint intel_icl_pebs_event_constraints[] = {
853 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
854 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x400000000ULL),	/* SLOTS */
855 
856 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
857 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),	/* MEM_INST_RETIRED.LOAD */
858 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),	/* MEM_INST_RETIRED.STORE */
859 
860 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
861 
862 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
863 
864 	/*
865 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
866 	 * need the full constraints from the main table.
867 	 */
868 
869 	EVENT_CONSTRAINT_END
870 };
871 
872 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
873 {
874 	struct event_constraint *c;
875 
876 	if (!event->attr.precise_ip)
877 		return NULL;
878 
879 	if (x86_pmu.pebs_constraints) {
880 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
881 			if (constraint_match(c, event->hw.config)) {
882 				event->hw.flags |= c->flags;
883 				return c;
884 			}
885 		}
886 	}
887 
888 	/*
889 	 * Extended PEBS support
890 	 * Makes the PEBS code search the normal constraints.
891 	 */
892 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
893 		return NULL;
894 
895 	return &emptyconstraint;
896 }
897 
898 /*
899  * We need the sched_task callback even for per-cpu events when we use
900  * the large interrupt threshold, such that we can provide PID and TID
901  * to PEBS samples.
902  */
903 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
904 {
905 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
906 }
907 
908 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
909 {
910 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
911 
912 	if (!sched_in && pebs_needs_sched_cb(cpuc))
913 		intel_pmu_drain_pebs_buffer();
914 }
915 
916 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
917 {
918 	struct debug_store *ds = cpuc->ds;
919 	u64 threshold;
920 	int reserved;
921 
922 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
923 		reserved = x86_pmu.max_pebs_events + x86_pmu.num_counters_fixed;
924 	else
925 		reserved = x86_pmu.max_pebs_events;
926 
927 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
928 		threshold = ds->pebs_absolute_maximum -
929 			reserved * cpuc->pebs_record_size;
930 	} else {
931 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
932 	}
933 
934 	ds->pebs_interrupt_threshold = threshold;
935 }
936 
937 static void adaptive_pebs_record_size_update(void)
938 {
939 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
940 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
941 	int sz = sizeof(struct pebs_basic);
942 
943 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
944 		sz += sizeof(struct pebs_meminfo);
945 	if (pebs_data_cfg & PEBS_DATACFG_GP)
946 		sz += sizeof(struct pebs_gprs);
947 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
948 		sz += sizeof(struct pebs_xmm);
949 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
950 		sz += x86_pmu.lbr_nr * sizeof(struct pebs_lbr_entry);
951 
952 	cpuc->pebs_record_size = sz;
953 }
954 
955 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
956 				PERF_SAMPLE_PHYS_ADDR | PERF_SAMPLE_WEIGHT | \
957 				PERF_SAMPLE_TRANSACTION)
958 
959 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
960 {
961 	struct perf_event_attr *attr = &event->attr;
962 	u64 sample_type = attr->sample_type;
963 	u64 pebs_data_cfg = 0;
964 	bool gprs, tsx_weight;
965 
966 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
967 	    attr->precise_ip > 1)
968 		return pebs_data_cfg;
969 
970 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
971 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
972 
973 	/*
974 	 * We need GPRs when:
975 	 * + user requested them
976 	 * + precise_ip < 2 for the non event IP
977 	 * + For RTM TSX weight we need GPRs for the abort code.
978 	 */
979 	gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
980 	       (attr->sample_regs_intr & PEBS_GP_REGS);
981 
982 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT) &&
983 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
984 		      x86_pmu.rtm_abort_event);
985 
986 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
987 		pebs_data_cfg |= PEBS_DATACFG_GP;
988 
989 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
990 	    (attr->sample_regs_intr & PEBS_XMM_REGS))
991 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
992 
993 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
994 		/*
995 		 * For now always log all LBRs. Could configure this
996 		 * later.
997 		 */
998 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
999 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1000 	}
1001 
1002 	return pebs_data_cfg;
1003 }
1004 
1005 static void
1006 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1007 		  struct perf_event *event, bool add)
1008 {
1009 	struct pmu *pmu = event->ctx->pmu;
1010 	/*
1011 	 * Make sure we get updated with the first PEBS
1012 	 * event. It will trigger also during removal, but
1013 	 * that does not hurt:
1014 	 */
1015 	bool update = cpuc->n_pebs == 1;
1016 
1017 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1018 		if (!needed_cb)
1019 			perf_sched_cb_inc(pmu);
1020 		else
1021 			perf_sched_cb_dec(pmu);
1022 
1023 		update = true;
1024 	}
1025 
1026 	/*
1027 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1028 	 * iterating all remaining PEBS events to reconstruct the config.
1029 	 */
1030 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1031 		u64 pebs_data_cfg;
1032 
1033 		/* Clear pebs_data_cfg and pebs_record_size for first PEBS. */
1034 		if (cpuc->n_pebs == 1) {
1035 			cpuc->pebs_data_cfg = 0;
1036 			cpuc->pebs_record_size = sizeof(struct pebs_basic);
1037 		}
1038 
1039 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1040 
1041 		/* Update pebs_record_size if new event requires more data. */
1042 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg) {
1043 			cpuc->pebs_data_cfg |= pebs_data_cfg;
1044 			adaptive_pebs_record_size_update();
1045 			update = true;
1046 		}
1047 	}
1048 
1049 	if (update)
1050 		pebs_update_threshold(cpuc);
1051 }
1052 
1053 void intel_pmu_pebs_add(struct perf_event *event)
1054 {
1055 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1056 	struct hw_perf_event *hwc = &event->hw;
1057 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1058 
1059 	cpuc->n_pebs++;
1060 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1061 		cpuc->n_large_pebs++;
1062 
1063 	pebs_update_state(needed_cb, cpuc, event, true);
1064 }
1065 
1066 void intel_pmu_pebs_enable(struct perf_event *event)
1067 {
1068 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1069 	struct hw_perf_event *hwc = &event->hw;
1070 	struct debug_store *ds = cpuc->ds;
1071 
1072 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1073 
1074 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1075 
1076 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1077 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1078 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1079 		cpuc->pebs_enabled |= 1ULL << 63;
1080 
1081 	if (x86_pmu.intel_cap.pebs_baseline) {
1082 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1083 		if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1084 			wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg);
1085 			cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg;
1086 		}
1087 	}
1088 
1089 	/*
1090 	 * Use auto-reload if possible to save a MSR write in the PMI.
1091 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1092 	 */
1093 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1094 		unsigned int idx = hwc->idx;
1095 
1096 		if (idx >= INTEL_PMC_IDX_FIXED)
1097 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1098 		ds->pebs_event_reset[idx] =
1099 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1100 	} else {
1101 		ds->pebs_event_reset[hwc->idx] = 0;
1102 	}
1103 }
1104 
1105 void intel_pmu_pebs_del(struct perf_event *event)
1106 {
1107 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1108 	struct hw_perf_event *hwc = &event->hw;
1109 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1110 
1111 	cpuc->n_pebs--;
1112 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1113 		cpuc->n_large_pebs--;
1114 
1115 	pebs_update_state(needed_cb, cpuc, event, false);
1116 }
1117 
1118 void intel_pmu_pebs_disable(struct perf_event *event)
1119 {
1120 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1121 	struct hw_perf_event *hwc = &event->hw;
1122 
1123 	if (cpuc->n_pebs == cpuc->n_large_pebs)
1124 		intel_pmu_drain_pebs_buffer();
1125 
1126 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1127 
1128 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1129 	    (x86_pmu.version < 5))
1130 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1131 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1132 		cpuc->pebs_enabled &= ~(1ULL << 63);
1133 
1134 	if (cpuc->enabled)
1135 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1136 
1137 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1138 }
1139 
1140 void intel_pmu_pebs_enable_all(void)
1141 {
1142 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1143 
1144 	if (cpuc->pebs_enabled)
1145 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1146 }
1147 
1148 void intel_pmu_pebs_disable_all(void)
1149 {
1150 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1151 
1152 	if (cpuc->pebs_enabled)
1153 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1154 }
1155 
1156 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1157 {
1158 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1159 	unsigned long from = cpuc->lbr_entries[0].from;
1160 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1161 	unsigned long ip = regs->ip;
1162 	int is_64bit = 0;
1163 	void *kaddr;
1164 	int size;
1165 
1166 	/*
1167 	 * We don't need to fixup if the PEBS assist is fault like
1168 	 */
1169 	if (!x86_pmu.intel_cap.pebs_trap)
1170 		return 1;
1171 
1172 	/*
1173 	 * No LBR entry, no basic block, no rewinding
1174 	 */
1175 	if (!cpuc->lbr_stack.nr || !from || !to)
1176 		return 0;
1177 
1178 	/*
1179 	 * Basic blocks should never cross user/kernel boundaries
1180 	 */
1181 	if (kernel_ip(ip) != kernel_ip(to))
1182 		return 0;
1183 
1184 	/*
1185 	 * unsigned math, either ip is before the start (impossible) or
1186 	 * the basic block is larger than 1 page (sanity)
1187 	 */
1188 	if ((ip - to) > PEBS_FIXUP_SIZE)
1189 		return 0;
1190 
1191 	/*
1192 	 * We sampled a branch insn, rewind using the LBR stack
1193 	 */
1194 	if (ip == to) {
1195 		set_linear_ip(regs, from);
1196 		return 1;
1197 	}
1198 
1199 	size = ip - to;
1200 	if (!kernel_ip(ip)) {
1201 		int bytes;
1202 		u8 *buf = this_cpu_read(insn_buffer);
1203 
1204 		/* 'size' must fit our buffer, see above */
1205 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1206 		if (bytes != 0)
1207 			return 0;
1208 
1209 		kaddr = buf;
1210 	} else {
1211 		kaddr = (void *)to;
1212 	}
1213 
1214 	do {
1215 		struct insn insn;
1216 
1217 		old_to = to;
1218 
1219 #ifdef CONFIG_X86_64
1220 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1221 #endif
1222 		insn_init(&insn, kaddr, size, is_64bit);
1223 		insn_get_length(&insn);
1224 		/*
1225 		 * Make sure there was not a problem decoding the
1226 		 * instruction and getting the length.  This is
1227 		 * doubly important because we have an infinite
1228 		 * loop if insn.length=0.
1229 		 */
1230 		if (!insn.length)
1231 			break;
1232 
1233 		to += insn.length;
1234 		kaddr += insn.length;
1235 		size -= insn.length;
1236 	} while (to < ip);
1237 
1238 	if (to == ip) {
1239 		set_linear_ip(regs, old_to);
1240 		return 1;
1241 	}
1242 
1243 	/*
1244 	 * Even though we decoded the basic block, the instruction stream
1245 	 * never matched the given IP, either the TO or the IP got corrupted.
1246 	 */
1247 	return 0;
1248 }
1249 
1250 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1251 {
1252 	if (tsx_tuning) {
1253 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1254 		return tsx.cycles_last_block;
1255 	}
1256 	return 0;
1257 }
1258 
1259 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1260 {
1261 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1262 
1263 	/* For RTM XABORTs also log the abort code from AX */
1264 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1265 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1266 	return txn;
1267 }
1268 
1269 static inline u64 get_pebs_status(void *n)
1270 {
1271 	if (x86_pmu.intel_cap.pebs_format < 4)
1272 		return ((struct pebs_record_nhm *)n)->status;
1273 	return ((struct pebs_basic *)n)->applicable_counters;
1274 }
1275 
1276 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1277 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1278 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1279 		 PERF_X86_EVENT_PEBS_NA_HSW)
1280 
1281 static u64 get_data_src(struct perf_event *event, u64 aux)
1282 {
1283 	u64 val = PERF_MEM_NA;
1284 	int fl = event->hw.flags;
1285 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1286 
1287 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1288 		val = load_latency_data(aux);
1289 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1290 		val = precise_datala_hsw(event, aux);
1291 	else if (fst)
1292 		val = precise_store_data(aux);
1293 	return val;
1294 }
1295 
1296 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1297 				   struct pt_regs *iregs, void *__pebs,
1298 				   struct perf_sample_data *data,
1299 				   struct pt_regs *regs)
1300 {
1301 	/*
1302 	 * We cast to the biggest pebs_record but are careful not to
1303 	 * unconditionally access the 'extra' entries.
1304 	 */
1305 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1306 	struct pebs_record_skl *pebs = __pebs;
1307 	u64 sample_type;
1308 	int fll;
1309 
1310 	if (pebs == NULL)
1311 		return;
1312 
1313 	sample_type = event->attr.sample_type;
1314 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1315 
1316 	perf_sample_data_init(data, 0, event->hw.last_period);
1317 
1318 	data->period = event->hw.last_period;
1319 
1320 	/*
1321 	 * Use latency for weight (only avail with PEBS-LL)
1322 	 */
1323 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1324 		data->weight = pebs->lat;
1325 
1326 	/*
1327 	 * data.data_src encodes the data source
1328 	 */
1329 	if (sample_type & PERF_SAMPLE_DATA_SRC)
1330 		data->data_src.val = get_data_src(event, pebs->dse);
1331 
1332 	/*
1333 	 * We must however always use iregs for the unwinder to stay sane; the
1334 	 * record BP,SP,IP can point into thin air when the record is from a
1335 	 * previous PMI context or an (I)RET happened between the record and
1336 	 * PMI.
1337 	 */
1338 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1339 		data->callchain = perf_callchain(event, iregs);
1340 
1341 	/*
1342 	 * We use the interrupt regs as a base because the PEBS record does not
1343 	 * contain a full regs set, specifically it seems to lack segment
1344 	 * descriptors, which get used by things like user_mode().
1345 	 *
1346 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1347 	 */
1348 	*regs = *iregs;
1349 
1350 	/*
1351 	 * Initialize regs_>flags from PEBS,
1352 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1353 	 * i.e., do not rely on it being zero:
1354 	 */
1355 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1356 
1357 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1358 		regs->ax = pebs->ax;
1359 		regs->bx = pebs->bx;
1360 		regs->cx = pebs->cx;
1361 		regs->dx = pebs->dx;
1362 		regs->si = pebs->si;
1363 		regs->di = pebs->di;
1364 
1365 		regs->bp = pebs->bp;
1366 		regs->sp = pebs->sp;
1367 
1368 #ifndef CONFIG_X86_32
1369 		regs->r8 = pebs->r8;
1370 		regs->r9 = pebs->r9;
1371 		regs->r10 = pebs->r10;
1372 		regs->r11 = pebs->r11;
1373 		regs->r12 = pebs->r12;
1374 		regs->r13 = pebs->r13;
1375 		regs->r14 = pebs->r14;
1376 		regs->r15 = pebs->r15;
1377 #endif
1378 	}
1379 
1380 	if (event->attr.precise_ip > 1) {
1381 		/*
1382 		 * Haswell and later processors have an 'eventing IP'
1383 		 * (real IP) which fixes the off-by-1 skid in hardware.
1384 		 * Use it when precise_ip >= 2 :
1385 		 */
1386 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1387 			set_linear_ip(regs, pebs->real_ip);
1388 			regs->flags |= PERF_EFLAGS_EXACT;
1389 		} else {
1390 			/* Otherwise, use PEBS off-by-1 IP: */
1391 			set_linear_ip(regs, pebs->ip);
1392 
1393 			/*
1394 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
1395 			 * using the LBR. If successful, the fixup function
1396 			 * corrects regs->ip and calls set_linear_ip() on regs:
1397 			 */
1398 			if (intel_pmu_pebs_fixup_ip(regs))
1399 				regs->flags |= PERF_EFLAGS_EXACT;
1400 		}
1401 	} else {
1402 		/*
1403 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
1404 		 * no fixup attempted:
1405 		 */
1406 		set_linear_ip(regs, pebs->ip);
1407 	}
1408 
1409 
1410 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1411 	    x86_pmu.intel_cap.pebs_format >= 1)
1412 		data->addr = pebs->dla;
1413 
1414 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1415 		/* Only set the TSX weight when no memory weight. */
1416 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1417 			data->weight = intel_get_tsx_weight(pebs->tsx_tuning);
1418 
1419 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1420 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1421 							      pebs->ax);
1422 	}
1423 
1424 	/*
1425 	 * v3 supplies an accurate time stamp, so we use that
1426 	 * for the time stamp.
1427 	 *
1428 	 * We can only do this for the default trace clock.
1429 	 */
1430 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1431 		event->attr.use_clockid == 0)
1432 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1433 
1434 	if (has_branch_stack(event))
1435 		data->br_stack = &cpuc->lbr_stack;
1436 }
1437 
1438 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1439 				    struct pebs_gprs *gprs)
1440 {
1441 	regs->ax = gprs->ax;
1442 	regs->bx = gprs->bx;
1443 	regs->cx = gprs->cx;
1444 	regs->dx = gprs->dx;
1445 	regs->si = gprs->si;
1446 	regs->di = gprs->di;
1447 	regs->bp = gprs->bp;
1448 	regs->sp = gprs->sp;
1449 #ifndef CONFIG_X86_32
1450 	regs->r8 = gprs->r8;
1451 	regs->r9 = gprs->r9;
1452 	regs->r10 = gprs->r10;
1453 	regs->r11 = gprs->r11;
1454 	regs->r12 = gprs->r12;
1455 	regs->r13 = gprs->r13;
1456 	regs->r14 = gprs->r14;
1457 	regs->r15 = gprs->r15;
1458 #endif
1459 }
1460 
1461 /*
1462  * With adaptive PEBS the layout depends on what fields are configured.
1463  */
1464 
1465 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1466 					    struct pt_regs *iregs, void *__pebs,
1467 					    struct perf_sample_data *data,
1468 					    struct pt_regs *regs)
1469 {
1470 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1471 	struct pebs_basic *basic = __pebs;
1472 	void *next_record = basic + 1;
1473 	u64 sample_type;
1474 	u64 format_size;
1475 	struct pebs_meminfo *meminfo = NULL;
1476 	struct pebs_gprs *gprs = NULL;
1477 	struct x86_perf_regs *perf_regs;
1478 
1479 	if (basic == NULL)
1480 		return;
1481 
1482 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
1483 	perf_regs->xmm_regs = NULL;
1484 
1485 	sample_type = event->attr.sample_type;
1486 	format_size = basic->format_size;
1487 	perf_sample_data_init(data, 0, event->hw.last_period);
1488 	data->period = event->hw.last_period;
1489 
1490 	if (event->attr.use_clockid == 0)
1491 		data->time = native_sched_clock_from_tsc(basic->tsc);
1492 
1493 	/*
1494 	 * We must however always use iregs for the unwinder to stay sane; the
1495 	 * record BP,SP,IP can point into thin air when the record is from a
1496 	 * previous PMI context or an (I)RET happened between the record and
1497 	 * PMI.
1498 	 */
1499 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1500 		data->callchain = perf_callchain(event, iregs);
1501 
1502 	*regs = *iregs;
1503 	/* The ip in basic is EventingIP */
1504 	set_linear_ip(regs, basic->ip);
1505 	regs->flags = PERF_EFLAGS_EXACT;
1506 
1507 	/*
1508 	 * The record for MEMINFO is in front of GP
1509 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1510 	 * Save the pointer here but process later.
1511 	 */
1512 	if (format_size & PEBS_DATACFG_MEMINFO) {
1513 		meminfo = next_record;
1514 		next_record = meminfo + 1;
1515 	}
1516 
1517 	if (format_size & PEBS_DATACFG_GP) {
1518 		gprs = next_record;
1519 		next_record = gprs + 1;
1520 
1521 		if (event->attr.precise_ip < 2) {
1522 			set_linear_ip(regs, gprs->ip);
1523 			regs->flags &= ~PERF_EFLAGS_EXACT;
1524 		}
1525 
1526 		if (sample_type & PERF_SAMPLE_REGS_INTR)
1527 			adaptive_pebs_save_regs(regs, gprs);
1528 	}
1529 
1530 	if (format_size & PEBS_DATACFG_MEMINFO) {
1531 		if (sample_type & PERF_SAMPLE_WEIGHT)
1532 			data->weight = meminfo->latency ?:
1533 				intel_get_tsx_weight(meminfo->tsx_tuning);
1534 
1535 		if (sample_type & PERF_SAMPLE_DATA_SRC)
1536 			data->data_src.val = get_data_src(event, meminfo->aux);
1537 
1538 		if (sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR))
1539 			data->addr = meminfo->address;
1540 
1541 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1542 			data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1543 							  gprs ? gprs->ax : 0);
1544 	}
1545 
1546 	if (format_size & PEBS_DATACFG_XMMS) {
1547 		struct pebs_xmm *xmm = next_record;
1548 
1549 		next_record = xmm + 1;
1550 		perf_regs->xmm_regs = xmm->xmm;
1551 	}
1552 
1553 	if (format_size & PEBS_DATACFG_LBRS) {
1554 		struct pebs_lbr *lbr = next_record;
1555 		int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1556 					& 0xff) + 1;
1557 		next_record = next_record + num_lbr*sizeof(struct pebs_lbr_entry);
1558 
1559 		if (has_branch_stack(event)) {
1560 			intel_pmu_store_pebs_lbrs(lbr);
1561 			data->br_stack = &cpuc->lbr_stack;
1562 		}
1563 	}
1564 
1565 	WARN_ONCE(next_record != __pebs + (format_size >> 48),
1566 			"PEBS record size %llu, expected %llu, config %llx\n",
1567 			format_size >> 48,
1568 			(u64)(next_record - __pebs),
1569 			basic->format_size);
1570 }
1571 
1572 static inline void *
1573 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1574 {
1575 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1576 	void *at;
1577 	u64 pebs_status;
1578 
1579 	/*
1580 	 * fmt0 does not have a status bitfield (does not use
1581 	 * perf_record_nhm format)
1582 	 */
1583 	if (x86_pmu.intel_cap.pebs_format < 1)
1584 		return base;
1585 
1586 	if (base == NULL)
1587 		return NULL;
1588 
1589 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1590 		unsigned long status = get_pebs_status(at);
1591 
1592 		if (test_bit(bit, (unsigned long *)&status)) {
1593 			/* PEBS v3 has accurate status bits */
1594 			if (x86_pmu.intel_cap.pebs_format >= 3)
1595 				return at;
1596 
1597 			if (status == (1 << bit))
1598 				return at;
1599 
1600 			/* clear non-PEBS bit and re-check */
1601 			pebs_status = status & cpuc->pebs_enabled;
1602 			pebs_status &= PEBS_COUNTER_MASK;
1603 			if (pebs_status == (1 << bit))
1604 				return at;
1605 		}
1606 	}
1607 	return NULL;
1608 }
1609 
1610 void intel_pmu_auto_reload_read(struct perf_event *event)
1611 {
1612 	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1613 
1614 	perf_pmu_disable(event->pmu);
1615 	intel_pmu_drain_pebs_buffer();
1616 	perf_pmu_enable(event->pmu);
1617 }
1618 
1619 /*
1620  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1621  */
1622 static int
1623 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1624 {
1625 	struct hw_perf_event *hwc = &event->hw;
1626 	int shift = 64 - x86_pmu.cntval_bits;
1627 	u64 period = hwc->sample_period;
1628 	u64 prev_raw_count, new_raw_count;
1629 	s64 new, old;
1630 
1631 	WARN_ON(!period);
1632 
1633 	/*
1634 	 * drain_pebs() only happens when the PMU is disabled.
1635 	 */
1636 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1637 
1638 	prev_raw_count = local64_read(&hwc->prev_count);
1639 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1640 	local64_set(&hwc->prev_count, new_raw_count);
1641 
1642 	/*
1643 	 * Since the counter increments a negative counter value and
1644 	 * overflows on the sign switch, giving the interval:
1645 	 *
1646 	 *   [-period, 0]
1647 	 *
1648 	 * the difference between two consequtive reads is:
1649 	 *
1650 	 *   A) value2 - value1;
1651 	 *      when no overflows have happened in between,
1652 	 *
1653 	 *   B) (0 - value1) + (value2 - (-period));
1654 	 *      when one overflow happened in between,
1655 	 *
1656 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1657 	 *      when @n overflows happened in between.
1658 	 *
1659 	 * Here A) is the obvious difference, B) is the extension to the
1660 	 * discrete interval, where the first term is to the top of the
1661 	 * interval and the second term is from the bottom of the next
1662 	 * interval and C) the extension to multiple intervals, where the
1663 	 * middle term is the whole intervals covered.
1664 	 *
1665 	 * An equivalent of C, by reduction, is:
1666 	 *
1667 	 *   value2 - value1 + n * period
1668 	 */
1669 	new = ((s64)(new_raw_count << shift) >> shift);
1670 	old = ((s64)(prev_raw_count << shift) >> shift);
1671 	local64_add(new - old + count * period, &event->count);
1672 
1673 	perf_event_update_userpage(event);
1674 
1675 	return 0;
1676 }
1677 
1678 static void __intel_pmu_pebs_event(struct perf_event *event,
1679 				   struct pt_regs *iregs,
1680 				   void *base, void *top,
1681 				   int bit, int count,
1682 				   void (*setup_sample)(struct perf_event *,
1683 						struct pt_regs *,
1684 						void *,
1685 						struct perf_sample_data *,
1686 						struct pt_regs *))
1687 {
1688 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1689 	struct hw_perf_event *hwc = &event->hw;
1690 	struct perf_sample_data data;
1691 	struct x86_perf_regs perf_regs;
1692 	struct pt_regs *regs = &perf_regs.regs;
1693 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1694 
1695 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1696 		/*
1697 		 * Now, auto-reload is only enabled in fixed period mode.
1698 		 * The reload value is always hwc->sample_period.
1699 		 * May need to change it, if auto-reload is enabled in
1700 		 * freq mode later.
1701 		 */
1702 		intel_pmu_save_and_restart_reload(event, count);
1703 	} else if (!intel_pmu_save_and_restart(event))
1704 		return;
1705 
1706 	while (count > 1) {
1707 		setup_sample(event, iregs, at, &data, regs);
1708 		perf_event_output(event, &data, regs);
1709 		at += cpuc->pebs_record_size;
1710 		at = get_next_pebs_record_by_bit(at, top, bit);
1711 		count--;
1712 	}
1713 
1714 	setup_sample(event, iregs, at, &data, regs);
1715 
1716 	/*
1717 	 * All but the last records are processed.
1718 	 * The last one is left to be able to call the overflow handler.
1719 	 */
1720 	if (perf_event_overflow(event, &data, regs)) {
1721 		x86_pmu_stop(event, 0);
1722 		return;
1723 	}
1724 
1725 }
1726 
1727 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1728 {
1729 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1730 	struct debug_store *ds = cpuc->ds;
1731 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1732 	struct pebs_record_core *at, *top;
1733 	int n;
1734 
1735 	if (!x86_pmu.pebs_active)
1736 		return;
1737 
1738 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1739 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1740 
1741 	/*
1742 	 * Whatever else happens, drain the thing
1743 	 */
1744 	ds->pebs_index = ds->pebs_buffer_base;
1745 
1746 	if (!test_bit(0, cpuc->active_mask))
1747 		return;
1748 
1749 	WARN_ON_ONCE(!event);
1750 
1751 	if (!event->attr.precise_ip)
1752 		return;
1753 
1754 	n = top - at;
1755 	if (n <= 0) {
1756 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1757 			intel_pmu_save_and_restart_reload(event, 0);
1758 		return;
1759 	}
1760 
1761 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n,
1762 			       setup_pebs_fixed_sample_data);
1763 }
1764 
1765 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
1766 {
1767 	struct perf_event *event;
1768 	int bit;
1769 
1770 	/*
1771 	 * The drain_pebs() could be called twice in a short period
1772 	 * for auto-reload event in pmu::read(). There are no
1773 	 * overflows have happened in between.
1774 	 * It needs to call intel_pmu_save_and_restart_reload() to
1775 	 * update the event->count for this case.
1776 	 */
1777 	for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
1778 		event = cpuc->events[bit];
1779 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1780 			intel_pmu_save_and_restart_reload(event, 0);
1781 	}
1782 }
1783 
1784 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1785 {
1786 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1787 	struct debug_store *ds = cpuc->ds;
1788 	struct perf_event *event;
1789 	void *base, *at, *top;
1790 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1791 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1792 	int bit, i, size;
1793 	u64 mask;
1794 
1795 	if (!x86_pmu.pebs_active)
1796 		return;
1797 
1798 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1799 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1800 
1801 	ds->pebs_index = ds->pebs_buffer_base;
1802 
1803 	mask = (1ULL << x86_pmu.max_pebs_events) - 1;
1804 	size = x86_pmu.max_pebs_events;
1805 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
1806 		mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
1807 		size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1808 	}
1809 
1810 	if (unlikely(base >= top)) {
1811 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
1812 		return;
1813 	}
1814 
1815 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1816 		struct pebs_record_nhm *p = at;
1817 		u64 pebs_status;
1818 
1819 		pebs_status = p->status & cpuc->pebs_enabled;
1820 		pebs_status &= mask;
1821 
1822 		/* PEBS v3 has more accurate status bits */
1823 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1824 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
1825 				counts[bit]++;
1826 
1827 			continue;
1828 		}
1829 
1830 		/*
1831 		 * On some CPUs the PEBS status can be zero when PEBS is
1832 		 * racing with clearing of GLOBAL_STATUS.
1833 		 *
1834 		 * Normally we would drop that record, but in the
1835 		 * case when there is only a single active PEBS event
1836 		 * we can assume it's for that event.
1837 		 */
1838 		if (!pebs_status && cpuc->pebs_enabled &&
1839 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1840 			pebs_status = cpuc->pebs_enabled;
1841 
1842 		bit = find_first_bit((unsigned long *)&pebs_status,
1843 					x86_pmu.max_pebs_events);
1844 		if (bit >= x86_pmu.max_pebs_events)
1845 			continue;
1846 
1847 		/*
1848 		 * The PEBS hardware does not deal well with the situation
1849 		 * when events happen near to each other and multiple bits
1850 		 * are set. But it should happen rarely.
1851 		 *
1852 		 * If these events include one PEBS and multiple non-PEBS
1853 		 * events, it doesn't impact PEBS record. The record will
1854 		 * be handled normally. (slow path)
1855 		 *
1856 		 * If these events include two or more PEBS events, the
1857 		 * records for the events can be collapsed into a single
1858 		 * one, and it's not possible to reconstruct all events
1859 		 * that caused the PEBS record. It's called collision.
1860 		 * If collision happened, the record will be dropped.
1861 		 */
1862 		if (p->status != (1ULL << bit)) {
1863 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
1864 				error[i]++;
1865 			continue;
1866 		}
1867 
1868 		counts[bit]++;
1869 	}
1870 
1871 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
1872 		if ((counts[bit] == 0) && (error[bit] == 0))
1873 			continue;
1874 
1875 		event = cpuc->events[bit];
1876 		if (WARN_ON_ONCE(!event))
1877 			continue;
1878 
1879 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1880 			continue;
1881 
1882 		/* log dropped samples number */
1883 		if (error[bit]) {
1884 			perf_log_lost_samples(event, error[bit]);
1885 
1886 			if (perf_event_account_interrupt(event))
1887 				x86_pmu_stop(event, 0);
1888 		}
1889 
1890 		if (counts[bit]) {
1891 			__intel_pmu_pebs_event(event, iregs, base,
1892 					       top, bit, counts[bit],
1893 					       setup_pebs_fixed_sample_data);
1894 		}
1895 	}
1896 }
1897 
1898 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs)
1899 {
1900 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1901 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1902 	struct debug_store *ds = cpuc->ds;
1903 	struct perf_event *event;
1904 	void *base, *at, *top;
1905 	int bit, size;
1906 	u64 mask;
1907 
1908 	if (!x86_pmu.pebs_active)
1909 		return;
1910 
1911 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
1912 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
1913 
1914 	ds->pebs_index = ds->pebs_buffer_base;
1915 
1916 	mask = ((1ULL << x86_pmu.max_pebs_events) - 1) |
1917 	       (((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
1918 	size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1919 
1920 	if (unlikely(base >= top)) {
1921 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
1922 		return;
1923 	}
1924 
1925 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1926 		u64 pebs_status;
1927 
1928 		pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
1929 		pebs_status &= mask;
1930 
1931 		for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
1932 			counts[bit]++;
1933 	}
1934 
1935 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
1936 		if (counts[bit] == 0)
1937 			continue;
1938 
1939 		event = cpuc->events[bit];
1940 		if (WARN_ON_ONCE(!event))
1941 			continue;
1942 
1943 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1944 			continue;
1945 
1946 		__intel_pmu_pebs_event(event, iregs, base,
1947 				       top, bit, counts[bit],
1948 				       setup_pebs_adaptive_sample_data);
1949 	}
1950 }
1951 
1952 /*
1953  * BTS, PEBS probe and setup
1954  */
1955 
1956 void __init intel_ds_init(void)
1957 {
1958 	/*
1959 	 * No support for 32bit formats
1960 	 */
1961 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1962 		return;
1963 
1964 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1965 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1966 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1967 	if (x86_pmu.version <= 4) {
1968 		x86_pmu.pebs_no_isolation = 1;
1969 		x86_pmu.pebs_no_xmm_regs = 1;
1970 	}
1971 	if (x86_pmu.pebs) {
1972 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1973 		char *pebs_qual = "";
1974 		int format = x86_pmu.intel_cap.pebs_format;
1975 
1976 		if (format < 4)
1977 			x86_pmu.intel_cap.pebs_baseline = 0;
1978 
1979 		switch (format) {
1980 		case 0:
1981 			pr_cont("PEBS fmt0%c, ", pebs_type);
1982 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1983 			/*
1984 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1985 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1986 			 * mysteriously hang on Core2.
1987 			 *
1988 			 * As a workaround, we don't do this.
1989 			 */
1990 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1991 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1992 			break;
1993 
1994 		case 1:
1995 			pr_cont("PEBS fmt1%c, ", pebs_type);
1996 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1997 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1998 			break;
1999 
2000 		case 2:
2001 			pr_cont("PEBS fmt2%c, ", pebs_type);
2002 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2003 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2004 			break;
2005 
2006 		case 3:
2007 			pr_cont("PEBS fmt3%c, ", pebs_type);
2008 			x86_pmu.pebs_record_size =
2009 						sizeof(struct pebs_record_skl);
2010 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2011 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2012 			break;
2013 
2014 		case 4:
2015 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2016 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2017 			if (x86_pmu.intel_cap.pebs_baseline) {
2018 				x86_pmu.large_pebs_flags |=
2019 					PERF_SAMPLE_BRANCH_STACK |
2020 					PERF_SAMPLE_TIME;
2021 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
2022 				pebs_qual = "-baseline";
2023 			} else {
2024 				/* Only basic record supported */
2025 				x86_pmu.pebs_no_xmm_regs = 1;
2026 				x86_pmu.large_pebs_flags &=
2027 					~(PERF_SAMPLE_ADDR |
2028 					  PERF_SAMPLE_TIME |
2029 					  PERF_SAMPLE_DATA_SRC |
2030 					  PERF_SAMPLE_TRANSACTION |
2031 					  PERF_SAMPLE_REGS_USER |
2032 					  PERF_SAMPLE_REGS_INTR);
2033 			}
2034 			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2035 			break;
2036 
2037 		default:
2038 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2039 			x86_pmu.pebs = 0;
2040 		}
2041 	}
2042 }
2043 
2044 void perf_restore_debug_store(void)
2045 {
2046 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2047 
2048 	if (!x86_pmu.bts && !x86_pmu.pebs)
2049 		return;
2050 
2051 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
2052 }
2053