xref: /openbmc/linux/arch/x86/events/intel/ds.c (revision f125e2d4)
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 #include <asm/io.h>
11 
12 #include "../perf_event.h"
13 
14 /* Waste a full page so it can be mapped into the cpu_entry_area */
15 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
16 
17 /* The size of a BTS record in bytes: */
18 #define BTS_RECORD_SIZE		24
19 
20 #define PEBS_FIXUP_SIZE		PAGE_SIZE
21 
22 /*
23  * pebs_record_32 for p4 and core not supported
24 
25 struct pebs_record_32 {
26 	u32 flags, ip;
27 	u32 ax, bc, cx, dx;
28 	u32 si, di, bp, sp;
29 };
30 
31  */
32 
33 union intel_x86_pebs_dse {
34 	u64 val;
35 	struct {
36 		unsigned int ld_dse:4;
37 		unsigned int ld_stlb_miss:1;
38 		unsigned int ld_locked:1;
39 		unsigned int ld_reserved:26;
40 	};
41 	struct {
42 		unsigned int st_l1d_hit:1;
43 		unsigned int st_reserved1:3;
44 		unsigned int st_stlb_miss:1;
45 		unsigned int st_locked:1;
46 		unsigned int st_reserved2:26;
47 	};
48 };
49 
50 
51 /*
52  * Map PEBS Load Latency Data Source encodings to generic
53  * memory data source information
54  */
55 #define P(a, b) PERF_MEM_S(a, b)
56 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
57 #define LEVEL(x) P(LVLNUM, x)
58 #define REM P(REMOTE, REMOTE)
59 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
60 
61 /* Version for Sandy Bridge and later */
62 static u64 pebs_data_source[] = {
63 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
64 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
65 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
66 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
67 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
68 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
69 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
70 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
71 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
72 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
73 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
74 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
75 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
76 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
77 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
78 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
79 };
80 
81 /* Patch up minor differences in the bits */
82 void __init intel_pmu_pebs_data_source_nhm(void)
83 {
84 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
85 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
86 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
87 }
88 
89 void __init intel_pmu_pebs_data_source_skl(bool pmem)
90 {
91 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
92 
93 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
94 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
95 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
96 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
97 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
98 }
99 
100 static u64 precise_store_data(u64 status)
101 {
102 	union intel_x86_pebs_dse dse;
103 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
104 
105 	dse.val = status;
106 
107 	/*
108 	 * bit 4: TLB access
109 	 * 1 = stored missed 2nd level TLB
110 	 *
111 	 * so it either hit the walker or the OS
112 	 * otherwise hit 2nd level TLB
113 	 */
114 	if (dse.st_stlb_miss)
115 		val |= P(TLB, MISS);
116 	else
117 		val |= P(TLB, HIT);
118 
119 	/*
120 	 * bit 0: hit L1 data cache
121 	 * if not set, then all we know is that
122 	 * it missed L1D
123 	 */
124 	if (dse.st_l1d_hit)
125 		val |= P(LVL, HIT);
126 	else
127 		val |= P(LVL, MISS);
128 
129 	/*
130 	 * bit 5: Locked prefix
131 	 */
132 	if (dse.st_locked)
133 		val |= P(LOCK, LOCKED);
134 
135 	return val;
136 }
137 
138 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
139 {
140 	union perf_mem_data_src dse;
141 
142 	dse.val = PERF_MEM_NA;
143 
144 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
145 		dse.mem_op = PERF_MEM_OP_STORE;
146 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
147 		dse.mem_op = PERF_MEM_OP_LOAD;
148 
149 	/*
150 	 * L1 info only valid for following events:
151 	 *
152 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
153 	 * MEM_UOPS_RETIRED.LOCK_STORES
154 	 * MEM_UOPS_RETIRED.SPLIT_STORES
155 	 * MEM_UOPS_RETIRED.ALL_STORES
156 	 */
157 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
158 		if (status & 1)
159 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
160 		else
161 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
162 	}
163 	return dse.val;
164 }
165 
166 static u64 load_latency_data(u64 status)
167 {
168 	union intel_x86_pebs_dse dse;
169 	u64 val;
170 
171 	dse.val = status;
172 
173 	/*
174 	 * use the mapping table for bit 0-3
175 	 */
176 	val = pebs_data_source[dse.ld_dse];
177 
178 	/*
179 	 * Nehalem models do not support TLB, Lock infos
180 	 */
181 	if (x86_pmu.pebs_no_tlb) {
182 		val |= P(TLB, NA) | P(LOCK, NA);
183 		return val;
184 	}
185 	/*
186 	 * bit 4: TLB access
187 	 * 0 = did not miss 2nd level TLB
188 	 * 1 = missed 2nd level TLB
189 	 */
190 	if (dse.ld_stlb_miss)
191 		val |= P(TLB, MISS) | P(TLB, L2);
192 	else
193 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
194 
195 	/*
196 	 * bit 5: locked prefix
197 	 */
198 	if (dse.ld_locked)
199 		val |= P(LOCK, LOCKED);
200 
201 	return val;
202 }
203 
204 struct pebs_record_core {
205 	u64 flags, ip;
206 	u64 ax, bx, cx, dx;
207 	u64 si, di, bp, sp;
208 	u64 r8,  r9,  r10, r11;
209 	u64 r12, r13, r14, r15;
210 };
211 
212 struct pebs_record_nhm {
213 	u64 flags, ip;
214 	u64 ax, bx, cx, dx;
215 	u64 si, di, bp, sp;
216 	u64 r8,  r9,  r10, r11;
217 	u64 r12, r13, r14, r15;
218 	u64 status, dla, dse, lat;
219 };
220 
221 /*
222  * Same as pebs_record_nhm, with two additional fields.
223  */
224 struct pebs_record_hsw {
225 	u64 flags, ip;
226 	u64 ax, bx, cx, dx;
227 	u64 si, di, bp, sp;
228 	u64 r8,  r9,  r10, r11;
229 	u64 r12, r13, r14, r15;
230 	u64 status, dla, dse, lat;
231 	u64 real_ip, tsx_tuning;
232 };
233 
234 union hsw_tsx_tuning {
235 	struct {
236 		u32 cycles_last_block     : 32,
237 		    hle_abort		  : 1,
238 		    rtm_abort		  : 1,
239 		    instruction_abort     : 1,
240 		    non_instruction_abort : 1,
241 		    retry		  : 1,
242 		    data_conflict	  : 1,
243 		    capacity_writes	  : 1,
244 		    capacity_reads	  : 1;
245 	};
246 	u64	    value;
247 };
248 
249 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
250 
251 /* Same as HSW, plus TSC */
252 
253 struct pebs_record_skl {
254 	u64 flags, ip;
255 	u64 ax, bx, cx, dx;
256 	u64 si, di, bp, sp;
257 	u64 r8,  r9,  r10, r11;
258 	u64 r12, r13, r14, r15;
259 	u64 status, dla, dse, lat;
260 	u64 real_ip, tsx_tuning;
261 	u64 tsc;
262 };
263 
264 void init_debug_store_on_cpu(int cpu)
265 {
266 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
267 
268 	if (!ds)
269 		return;
270 
271 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
272 		     (u32)((u64)(unsigned long)ds),
273 		     (u32)((u64)(unsigned long)ds >> 32));
274 }
275 
276 void fini_debug_store_on_cpu(int cpu)
277 {
278 	if (!per_cpu(cpu_hw_events, cpu).ds)
279 		return;
280 
281 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
282 }
283 
284 static DEFINE_PER_CPU(void *, insn_buffer);
285 
286 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
287 {
288 	unsigned long start = (unsigned long)cea;
289 	phys_addr_t pa;
290 	size_t msz = 0;
291 
292 	pa = virt_to_phys(addr);
293 
294 	preempt_disable();
295 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
296 		cea_set_pte(cea, pa, prot);
297 
298 	/*
299 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
300 	 * all TLB entries for it.
301 	 */
302 	flush_tlb_kernel_range(start, start + size);
303 	preempt_enable();
304 }
305 
306 static void ds_clear_cea(void *cea, size_t size)
307 {
308 	unsigned long start = (unsigned long)cea;
309 	size_t msz = 0;
310 
311 	preempt_disable();
312 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
313 		cea_set_pte(cea, 0, PAGE_NONE);
314 
315 	flush_tlb_kernel_range(start, start + size);
316 	preempt_enable();
317 }
318 
319 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
320 {
321 	unsigned int order = get_order(size);
322 	int node = cpu_to_node(cpu);
323 	struct page *page;
324 
325 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
326 	return page ? page_address(page) : NULL;
327 }
328 
329 static void dsfree_pages(const void *buffer, size_t size)
330 {
331 	if (buffer)
332 		free_pages((unsigned long)buffer, get_order(size));
333 }
334 
335 static int alloc_pebs_buffer(int cpu)
336 {
337 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
338 	struct debug_store *ds = hwev->ds;
339 	size_t bsiz = x86_pmu.pebs_buffer_size;
340 	int max, node = cpu_to_node(cpu);
341 	void *buffer, *insn_buff, *cea;
342 
343 	if (!x86_pmu.pebs)
344 		return 0;
345 
346 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
347 	if (unlikely(!buffer))
348 		return -ENOMEM;
349 
350 	/*
351 	 * HSW+ already provides us the eventing ip; no need to allocate this
352 	 * buffer then.
353 	 */
354 	if (x86_pmu.intel_cap.pebs_format < 2) {
355 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
356 		if (!insn_buff) {
357 			dsfree_pages(buffer, bsiz);
358 			return -ENOMEM;
359 		}
360 		per_cpu(insn_buffer, cpu) = insn_buff;
361 	}
362 	hwev->ds_pebs_vaddr = buffer;
363 	/* Update the cpu entry area mapping */
364 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
365 	ds->pebs_buffer_base = (unsigned long) cea;
366 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
367 	ds->pebs_index = ds->pebs_buffer_base;
368 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
369 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
370 	return 0;
371 }
372 
373 static void release_pebs_buffer(int cpu)
374 {
375 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
376 	void *cea;
377 
378 	if (!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 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
388 	hwev->ds_pebs_vaddr = NULL;
389 }
390 
391 static int alloc_bts_buffer(int cpu)
392 {
393 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
394 	struct debug_store *ds = hwev->ds;
395 	void *buffer, *cea;
396 	int max;
397 
398 	if (!x86_pmu.bts)
399 		return 0;
400 
401 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
402 	if (unlikely(!buffer)) {
403 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
404 		return -ENOMEM;
405 	}
406 	hwev->ds_bts_vaddr = buffer;
407 	/* Update the fixmap */
408 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
409 	ds->bts_buffer_base = (unsigned long) cea;
410 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
411 	ds->bts_index = ds->bts_buffer_base;
412 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
413 	ds->bts_absolute_maximum = ds->bts_buffer_base +
414 					max * BTS_RECORD_SIZE;
415 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
416 					(max / 16) * BTS_RECORD_SIZE;
417 	return 0;
418 }
419 
420 static void release_bts_buffer(int cpu)
421 {
422 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
423 	void *cea;
424 
425 	if (!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 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
432 	hwev->ds_bts_vaddr = NULL;
433 }
434 
435 static int alloc_ds_buffer(int cpu)
436 {
437 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
438 
439 	memset(ds, 0, sizeof(*ds));
440 	per_cpu(cpu_hw_events, cpu).ds = ds;
441 	return 0;
442 }
443 
444 static void release_ds_buffer(int cpu)
445 {
446 	per_cpu(cpu_hw_events, cpu).ds = NULL;
447 }
448 
449 void release_ds_buffers(void)
450 {
451 	int cpu;
452 
453 	if (!x86_pmu.bts && !x86_pmu.pebs)
454 		return;
455 
456 	for_each_possible_cpu(cpu)
457 		release_ds_buffer(cpu);
458 
459 	for_each_possible_cpu(cpu) {
460 		/*
461 		 * Again, ignore errors from offline CPUs, they will no longer
462 		 * observe cpu_hw_events.ds and not program the DS_AREA when
463 		 * they come up.
464 		 */
465 		fini_debug_store_on_cpu(cpu);
466 	}
467 
468 	for_each_possible_cpu(cpu) {
469 		release_pebs_buffer(cpu);
470 		release_bts_buffer(cpu);
471 	}
472 }
473 
474 void reserve_ds_buffers(void)
475 {
476 	int bts_err = 0, pebs_err = 0;
477 	int cpu;
478 
479 	x86_pmu.bts_active = 0;
480 	x86_pmu.pebs_active = 0;
481 
482 	if (!x86_pmu.bts && !x86_pmu.pebs)
483 		return;
484 
485 	if (!x86_pmu.bts)
486 		bts_err = 1;
487 
488 	if (!x86_pmu.pebs)
489 		pebs_err = 1;
490 
491 	for_each_possible_cpu(cpu) {
492 		if (alloc_ds_buffer(cpu)) {
493 			bts_err = 1;
494 			pebs_err = 1;
495 		}
496 
497 		if (!bts_err && alloc_bts_buffer(cpu))
498 			bts_err = 1;
499 
500 		if (!pebs_err && alloc_pebs_buffer(cpu))
501 			pebs_err = 1;
502 
503 		if (bts_err && pebs_err)
504 			break;
505 	}
506 
507 	if (bts_err) {
508 		for_each_possible_cpu(cpu)
509 			release_bts_buffer(cpu);
510 	}
511 
512 	if (pebs_err) {
513 		for_each_possible_cpu(cpu)
514 			release_pebs_buffer(cpu);
515 	}
516 
517 	if (bts_err && pebs_err) {
518 		for_each_possible_cpu(cpu)
519 			release_ds_buffer(cpu);
520 	} else {
521 		if (x86_pmu.bts && !bts_err)
522 			x86_pmu.bts_active = 1;
523 
524 		if (x86_pmu.pebs && !pebs_err)
525 			x86_pmu.pebs_active = 1;
526 
527 		for_each_possible_cpu(cpu) {
528 			/*
529 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
530 			 * will get this call through intel_pmu_cpu_starting().
531 			 */
532 			init_debug_store_on_cpu(cpu);
533 		}
534 	}
535 }
536 
537 /*
538  * BTS
539  */
540 
541 struct event_constraint bts_constraint =
542 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
543 
544 void intel_pmu_enable_bts(u64 config)
545 {
546 	unsigned long debugctlmsr;
547 
548 	debugctlmsr = get_debugctlmsr();
549 
550 	debugctlmsr |= DEBUGCTLMSR_TR;
551 	debugctlmsr |= DEBUGCTLMSR_BTS;
552 	if (config & ARCH_PERFMON_EVENTSEL_INT)
553 		debugctlmsr |= DEBUGCTLMSR_BTINT;
554 
555 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
556 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
557 
558 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
559 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
560 
561 	update_debugctlmsr(debugctlmsr);
562 }
563 
564 void intel_pmu_disable_bts(void)
565 {
566 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
567 	unsigned long debugctlmsr;
568 
569 	if (!cpuc->ds)
570 		return;
571 
572 	debugctlmsr = get_debugctlmsr();
573 
574 	debugctlmsr &=
575 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
576 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
577 
578 	update_debugctlmsr(debugctlmsr);
579 }
580 
581 int intel_pmu_drain_bts_buffer(void)
582 {
583 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
584 	struct debug_store *ds = cpuc->ds;
585 	struct bts_record {
586 		u64	from;
587 		u64	to;
588 		u64	flags;
589 	};
590 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
591 	struct bts_record *at, *base, *top;
592 	struct perf_output_handle handle;
593 	struct perf_event_header header;
594 	struct perf_sample_data data;
595 	unsigned long skip = 0;
596 	struct pt_regs regs;
597 
598 	if (!event)
599 		return 0;
600 
601 	if (!x86_pmu.bts_active)
602 		return 0;
603 
604 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
605 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
606 
607 	if (top <= base)
608 		return 0;
609 
610 	memset(&regs, 0, sizeof(regs));
611 
612 	ds->bts_index = ds->bts_buffer_base;
613 
614 	perf_sample_data_init(&data, 0, event->hw.last_period);
615 
616 	/*
617 	 * BTS leaks kernel addresses in branches across the cpl boundary,
618 	 * such as traps or system calls, so unless the user is asking for
619 	 * kernel tracing (and right now it's not possible), we'd need to
620 	 * filter them out. But first we need to count how many of those we
621 	 * have in the current batch. This is an extra O(n) pass, however,
622 	 * it's much faster than the other one especially considering that
623 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
624 	 * alloc_bts_buffer()).
625 	 */
626 	for (at = base; at < top; at++) {
627 		/*
628 		 * Note that right now *this* BTS code only works if
629 		 * attr::exclude_kernel is set, but let's keep this extra
630 		 * check here in case that changes.
631 		 */
632 		if (event->attr.exclude_kernel &&
633 		    (kernel_ip(at->from) || kernel_ip(at->to)))
634 			skip++;
635 	}
636 
637 	/*
638 	 * Prepare a generic sample, i.e. fill in the invariant fields.
639 	 * We will overwrite the from and to address before we output
640 	 * the sample.
641 	 */
642 	rcu_read_lock();
643 	perf_prepare_sample(&header, &data, event, &regs);
644 
645 	if (perf_output_begin(&handle, event, header.size *
646 			      (top - base - skip)))
647 		goto unlock;
648 
649 	for (at = base; at < top; at++) {
650 		/* Filter out any records that contain kernel addresses. */
651 		if (event->attr.exclude_kernel &&
652 		    (kernel_ip(at->from) || kernel_ip(at->to)))
653 			continue;
654 
655 		data.ip		= at->from;
656 		data.addr	= at->to;
657 
658 		perf_output_sample(&handle, &header, &data, event);
659 	}
660 
661 	perf_output_end(&handle);
662 
663 	/* There's new data available. */
664 	event->hw.interrupts++;
665 	event->pending_kill = POLL_IN;
666 unlock:
667 	rcu_read_unlock();
668 	return 1;
669 }
670 
671 static inline void intel_pmu_drain_pebs_buffer(void)
672 {
673 	struct pt_regs regs;
674 
675 	x86_pmu.drain_pebs(&regs);
676 }
677 
678 /*
679  * PEBS
680  */
681 struct event_constraint intel_core2_pebs_event_constraints[] = {
682 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
683 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
684 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
685 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
686 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
687 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
688 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
689 	EVENT_CONSTRAINT_END
690 };
691 
692 struct event_constraint intel_atom_pebs_event_constraints[] = {
693 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
694 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
695 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
696 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
697 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
698 	/* Allow all events as PEBS with no flags */
699 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
700 	EVENT_CONSTRAINT_END
701 };
702 
703 struct event_constraint intel_slm_pebs_event_constraints[] = {
704 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
705 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
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_glm_pebs_event_constraints[] = {
712 	/* Allow all events as PEBS with no flags */
713 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
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_UEVENT_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_UEVENT_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_UEVENT_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_UEVENT_CONSTRAINT(0x108001c2, 0xf),
772 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
773 	INTEL_FLAGS_UEVENT_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_UEVENT_CONSTRAINT(0x108001c2, 0xf),
788 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
789 	INTEL_FLAGS_UEVENT_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_UEVENT_CONSTRAINT(0x108001c2, 0xf),
811 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
812 	INTEL_FLAGS_UEVENT_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_UEVENT_CONSTRAINT(0x108001c0, 0x2),
834 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
835 	INTEL_FLAGS_UEVENT_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_icl_pebs_event_constraints[] = {
854 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
855 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
856 
857 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
858 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),	/* MEM_INST_RETIRED.LOAD */
859 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),	/* MEM_INST_RETIRED.STORE */
860 
861 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
862 
863 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
864 
865 	/*
866 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
867 	 * need the full constraints from the main table.
868 	 */
869 
870 	EVENT_CONSTRAINT_END
871 };
872 
873 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
874 {
875 	struct event_constraint *c;
876 
877 	if (!event->attr.precise_ip)
878 		return NULL;
879 
880 	if (x86_pmu.pebs_constraints) {
881 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
882 			if (constraint_match(c, event->hw.config)) {
883 				event->hw.flags |= c->flags;
884 				return c;
885 			}
886 		}
887 	}
888 
889 	/*
890 	 * Extended PEBS support
891 	 * Makes the PEBS code search the normal constraints.
892 	 */
893 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
894 		return NULL;
895 
896 	return &emptyconstraint;
897 }
898 
899 /*
900  * We need the sched_task callback even for per-cpu events when we use
901  * the large interrupt threshold, such that we can provide PID and TID
902  * to PEBS samples.
903  */
904 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
905 {
906 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
907 		return false;
908 
909 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
910 }
911 
912 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
913 {
914 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
915 
916 	if (!sched_in && pebs_needs_sched_cb(cpuc))
917 		intel_pmu_drain_pebs_buffer();
918 }
919 
920 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
921 {
922 	struct debug_store *ds = cpuc->ds;
923 	u64 threshold;
924 	int reserved;
925 
926 	if (cpuc->n_pebs_via_pt)
927 		return;
928 
929 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
930 		reserved = x86_pmu.max_pebs_events + x86_pmu.num_counters_fixed;
931 	else
932 		reserved = x86_pmu.max_pebs_events;
933 
934 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
935 		threshold = ds->pebs_absolute_maximum -
936 			reserved * cpuc->pebs_record_size;
937 	} else {
938 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
939 	}
940 
941 	ds->pebs_interrupt_threshold = threshold;
942 }
943 
944 static void adaptive_pebs_record_size_update(void)
945 {
946 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
947 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
948 	int sz = sizeof(struct pebs_basic);
949 
950 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
951 		sz += sizeof(struct pebs_meminfo);
952 	if (pebs_data_cfg & PEBS_DATACFG_GP)
953 		sz += sizeof(struct pebs_gprs);
954 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
955 		sz += sizeof(struct pebs_xmm);
956 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
957 		sz += x86_pmu.lbr_nr * sizeof(struct pebs_lbr_entry);
958 
959 	cpuc->pebs_record_size = sz;
960 }
961 
962 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
963 				PERF_SAMPLE_PHYS_ADDR | PERF_SAMPLE_WEIGHT | \
964 				PERF_SAMPLE_TRANSACTION)
965 
966 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
967 {
968 	struct perf_event_attr *attr = &event->attr;
969 	u64 sample_type = attr->sample_type;
970 	u64 pebs_data_cfg = 0;
971 	bool gprs, tsx_weight;
972 
973 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
974 	    attr->precise_ip > 1)
975 		return pebs_data_cfg;
976 
977 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
978 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
979 
980 	/*
981 	 * We need GPRs when:
982 	 * + user requested them
983 	 * + precise_ip < 2 for the non event IP
984 	 * + For RTM TSX weight we need GPRs for the abort code.
985 	 */
986 	gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
987 	       (attr->sample_regs_intr & PEBS_GP_REGS);
988 
989 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT) &&
990 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
991 		      x86_pmu.rtm_abort_event);
992 
993 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
994 		pebs_data_cfg |= PEBS_DATACFG_GP;
995 
996 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
997 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
998 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
999 
1000 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1001 		/*
1002 		 * For now always log all LBRs. Could configure this
1003 		 * later.
1004 		 */
1005 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1006 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1007 	}
1008 
1009 	return pebs_data_cfg;
1010 }
1011 
1012 static void
1013 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1014 		  struct perf_event *event, bool add)
1015 {
1016 	struct pmu *pmu = event->ctx->pmu;
1017 	/*
1018 	 * Make sure we get updated with the first PEBS
1019 	 * event. It will trigger also during removal, but
1020 	 * that does not hurt:
1021 	 */
1022 	bool update = cpuc->n_pebs == 1;
1023 
1024 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1025 		if (!needed_cb)
1026 			perf_sched_cb_inc(pmu);
1027 		else
1028 			perf_sched_cb_dec(pmu);
1029 
1030 		update = true;
1031 	}
1032 
1033 	/*
1034 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1035 	 * iterating all remaining PEBS events to reconstruct the config.
1036 	 */
1037 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1038 		u64 pebs_data_cfg;
1039 
1040 		/* Clear pebs_data_cfg and pebs_record_size for first PEBS. */
1041 		if (cpuc->n_pebs == 1) {
1042 			cpuc->pebs_data_cfg = 0;
1043 			cpuc->pebs_record_size = sizeof(struct pebs_basic);
1044 		}
1045 
1046 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1047 
1048 		/* Update pebs_record_size if new event requires more data. */
1049 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg) {
1050 			cpuc->pebs_data_cfg |= pebs_data_cfg;
1051 			adaptive_pebs_record_size_update();
1052 			update = true;
1053 		}
1054 	}
1055 
1056 	if (update)
1057 		pebs_update_threshold(cpuc);
1058 }
1059 
1060 void intel_pmu_pebs_add(struct perf_event *event)
1061 {
1062 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1063 	struct hw_perf_event *hwc = &event->hw;
1064 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1065 
1066 	cpuc->n_pebs++;
1067 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1068 		cpuc->n_large_pebs++;
1069 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1070 		cpuc->n_pebs_via_pt++;
1071 
1072 	pebs_update_state(needed_cb, cpuc, event, true);
1073 }
1074 
1075 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1076 {
1077 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1078 
1079 	if (!is_pebs_pt(event))
1080 		return;
1081 
1082 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1083 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1084 }
1085 
1086 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1087 {
1088 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1089 	struct hw_perf_event *hwc = &event->hw;
1090 	struct debug_store *ds = cpuc->ds;
1091 
1092 	if (!is_pebs_pt(event))
1093 		return;
1094 
1095 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1096 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1097 
1098 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1099 
1100 	wrmsrl(MSR_RELOAD_PMC0 + hwc->idx, ds->pebs_event_reset[hwc->idx]);
1101 }
1102 
1103 void intel_pmu_pebs_enable(struct perf_event *event)
1104 {
1105 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1106 	struct hw_perf_event *hwc = &event->hw;
1107 	struct debug_store *ds = cpuc->ds;
1108 
1109 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1110 
1111 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1112 
1113 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1114 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1115 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1116 		cpuc->pebs_enabled |= 1ULL << 63;
1117 
1118 	if (x86_pmu.intel_cap.pebs_baseline) {
1119 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1120 		if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1121 			wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg);
1122 			cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg;
1123 		}
1124 	}
1125 
1126 	/*
1127 	 * Use auto-reload if possible to save a MSR write in the PMI.
1128 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1129 	 */
1130 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1131 		unsigned int idx = hwc->idx;
1132 
1133 		if (idx >= INTEL_PMC_IDX_FIXED)
1134 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1135 		ds->pebs_event_reset[idx] =
1136 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1137 	} else {
1138 		ds->pebs_event_reset[hwc->idx] = 0;
1139 	}
1140 
1141 	intel_pmu_pebs_via_pt_enable(event);
1142 }
1143 
1144 void intel_pmu_pebs_del(struct perf_event *event)
1145 {
1146 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1147 	struct hw_perf_event *hwc = &event->hw;
1148 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1149 
1150 	cpuc->n_pebs--;
1151 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1152 		cpuc->n_large_pebs--;
1153 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1154 		cpuc->n_pebs_via_pt--;
1155 
1156 	pebs_update_state(needed_cb, cpuc, event, false);
1157 }
1158 
1159 void intel_pmu_pebs_disable(struct perf_event *event)
1160 {
1161 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1162 	struct hw_perf_event *hwc = &event->hw;
1163 
1164 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1165 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1166 		intel_pmu_drain_pebs_buffer();
1167 
1168 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1169 
1170 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1171 	    (x86_pmu.version < 5))
1172 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1173 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1174 		cpuc->pebs_enabled &= ~(1ULL << 63);
1175 
1176 	intel_pmu_pebs_via_pt_disable(event);
1177 
1178 	if (cpuc->enabled)
1179 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1180 
1181 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1182 }
1183 
1184 void intel_pmu_pebs_enable_all(void)
1185 {
1186 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1187 
1188 	if (cpuc->pebs_enabled)
1189 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1190 }
1191 
1192 void intel_pmu_pebs_disable_all(void)
1193 {
1194 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1195 
1196 	if (cpuc->pebs_enabled)
1197 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1198 }
1199 
1200 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1201 {
1202 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1203 	unsigned long from = cpuc->lbr_entries[0].from;
1204 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1205 	unsigned long ip = regs->ip;
1206 	int is_64bit = 0;
1207 	void *kaddr;
1208 	int size;
1209 
1210 	/*
1211 	 * We don't need to fixup if the PEBS assist is fault like
1212 	 */
1213 	if (!x86_pmu.intel_cap.pebs_trap)
1214 		return 1;
1215 
1216 	/*
1217 	 * No LBR entry, no basic block, no rewinding
1218 	 */
1219 	if (!cpuc->lbr_stack.nr || !from || !to)
1220 		return 0;
1221 
1222 	/*
1223 	 * Basic blocks should never cross user/kernel boundaries
1224 	 */
1225 	if (kernel_ip(ip) != kernel_ip(to))
1226 		return 0;
1227 
1228 	/*
1229 	 * unsigned math, either ip is before the start (impossible) or
1230 	 * the basic block is larger than 1 page (sanity)
1231 	 */
1232 	if ((ip - to) > PEBS_FIXUP_SIZE)
1233 		return 0;
1234 
1235 	/*
1236 	 * We sampled a branch insn, rewind using the LBR stack
1237 	 */
1238 	if (ip == to) {
1239 		set_linear_ip(regs, from);
1240 		return 1;
1241 	}
1242 
1243 	size = ip - to;
1244 	if (!kernel_ip(ip)) {
1245 		int bytes;
1246 		u8 *buf = this_cpu_read(insn_buffer);
1247 
1248 		/* 'size' must fit our buffer, see above */
1249 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1250 		if (bytes != 0)
1251 			return 0;
1252 
1253 		kaddr = buf;
1254 	} else {
1255 		kaddr = (void *)to;
1256 	}
1257 
1258 	do {
1259 		struct insn insn;
1260 
1261 		old_to = to;
1262 
1263 #ifdef CONFIG_X86_64
1264 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1265 #endif
1266 		insn_init(&insn, kaddr, size, is_64bit);
1267 		insn_get_length(&insn);
1268 		/*
1269 		 * Make sure there was not a problem decoding the
1270 		 * instruction and getting the length.  This is
1271 		 * doubly important because we have an infinite
1272 		 * loop if insn.length=0.
1273 		 */
1274 		if (!insn.length)
1275 			break;
1276 
1277 		to += insn.length;
1278 		kaddr += insn.length;
1279 		size -= insn.length;
1280 	} while (to < ip);
1281 
1282 	if (to == ip) {
1283 		set_linear_ip(regs, old_to);
1284 		return 1;
1285 	}
1286 
1287 	/*
1288 	 * Even though we decoded the basic block, the instruction stream
1289 	 * never matched the given IP, either the TO or the IP got corrupted.
1290 	 */
1291 	return 0;
1292 }
1293 
1294 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1295 {
1296 	if (tsx_tuning) {
1297 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1298 		return tsx.cycles_last_block;
1299 	}
1300 	return 0;
1301 }
1302 
1303 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1304 {
1305 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1306 
1307 	/* For RTM XABORTs also log the abort code from AX */
1308 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1309 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1310 	return txn;
1311 }
1312 
1313 static inline u64 get_pebs_status(void *n)
1314 {
1315 	if (x86_pmu.intel_cap.pebs_format < 4)
1316 		return ((struct pebs_record_nhm *)n)->status;
1317 	return ((struct pebs_basic *)n)->applicable_counters;
1318 }
1319 
1320 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1321 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1322 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1323 		 PERF_X86_EVENT_PEBS_NA_HSW)
1324 
1325 static u64 get_data_src(struct perf_event *event, u64 aux)
1326 {
1327 	u64 val = PERF_MEM_NA;
1328 	int fl = event->hw.flags;
1329 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1330 
1331 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1332 		val = load_latency_data(aux);
1333 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1334 		val = precise_datala_hsw(event, aux);
1335 	else if (fst)
1336 		val = precise_store_data(aux);
1337 	return val;
1338 }
1339 
1340 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1341 				   struct pt_regs *iregs, void *__pebs,
1342 				   struct perf_sample_data *data,
1343 				   struct pt_regs *regs)
1344 {
1345 	/*
1346 	 * We cast to the biggest pebs_record but are careful not to
1347 	 * unconditionally access the 'extra' entries.
1348 	 */
1349 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1350 	struct pebs_record_skl *pebs = __pebs;
1351 	u64 sample_type;
1352 	int fll;
1353 
1354 	if (pebs == NULL)
1355 		return;
1356 
1357 	sample_type = event->attr.sample_type;
1358 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1359 
1360 	perf_sample_data_init(data, 0, event->hw.last_period);
1361 
1362 	data->period = event->hw.last_period;
1363 
1364 	/*
1365 	 * Use latency for weight (only avail with PEBS-LL)
1366 	 */
1367 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1368 		data->weight = pebs->lat;
1369 
1370 	/*
1371 	 * data.data_src encodes the data source
1372 	 */
1373 	if (sample_type & PERF_SAMPLE_DATA_SRC)
1374 		data->data_src.val = get_data_src(event, pebs->dse);
1375 
1376 	/*
1377 	 * We must however always use iregs for the unwinder to stay sane; the
1378 	 * record BP,SP,IP can point into thin air when the record is from a
1379 	 * previous PMI context or an (I)RET happened between the record and
1380 	 * PMI.
1381 	 */
1382 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1383 		data->callchain = perf_callchain(event, iregs);
1384 
1385 	/*
1386 	 * We use the interrupt regs as a base because the PEBS record does not
1387 	 * contain a full regs set, specifically it seems to lack segment
1388 	 * descriptors, which get used by things like user_mode().
1389 	 *
1390 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1391 	 */
1392 	*regs = *iregs;
1393 
1394 	/*
1395 	 * Initialize regs_>flags from PEBS,
1396 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1397 	 * i.e., do not rely on it being zero:
1398 	 */
1399 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1400 
1401 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1402 		regs->ax = pebs->ax;
1403 		regs->bx = pebs->bx;
1404 		regs->cx = pebs->cx;
1405 		regs->dx = pebs->dx;
1406 		regs->si = pebs->si;
1407 		regs->di = pebs->di;
1408 
1409 		regs->bp = pebs->bp;
1410 		regs->sp = pebs->sp;
1411 
1412 #ifndef CONFIG_X86_32
1413 		regs->r8 = pebs->r8;
1414 		regs->r9 = pebs->r9;
1415 		regs->r10 = pebs->r10;
1416 		regs->r11 = pebs->r11;
1417 		regs->r12 = pebs->r12;
1418 		regs->r13 = pebs->r13;
1419 		regs->r14 = pebs->r14;
1420 		regs->r15 = pebs->r15;
1421 #endif
1422 	}
1423 
1424 	if (event->attr.precise_ip > 1) {
1425 		/*
1426 		 * Haswell and later processors have an 'eventing IP'
1427 		 * (real IP) which fixes the off-by-1 skid in hardware.
1428 		 * Use it when precise_ip >= 2 :
1429 		 */
1430 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1431 			set_linear_ip(regs, pebs->real_ip);
1432 			regs->flags |= PERF_EFLAGS_EXACT;
1433 		} else {
1434 			/* Otherwise, use PEBS off-by-1 IP: */
1435 			set_linear_ip(regs, pebs->ip);
1436 
1437 			/*
1438 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
1439 			 * using the LBR. If successful, the fixup function
1440 			 * corrects regs->ip and calls set_linear_ip() on regs:
1441 			 */
1442 			if (intel_pmu_pebs_fixup_ip(regs))
1443 				regs->flags |= PERF_EFLAGS_EXACT;
1444 		}
1445 	} else {
1446 		/*
1447 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
1448 		 * no fixup attempted:
1449 		 */
1450 		set_linear_ip(regs, pebs->ip);
1451 	}
1452 
1453 
1454 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1455 	    x86_pmu.intel_cap.pebs_format >= 1)
1456 		data->addr = pebs->dla;
1457 
1458 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1459 		/* Only set the TSX weight when no memory weight. */
1460 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1461 			data->weight = intel_get_tsx_weight(pebs->tsx_tuning);
1462 
1463 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1464 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1465 							      pebs->ax);
1466 	}
1467 
1468 	/*
1469 	 * v3 supplies an accurate time stamp, so we use that
1470 	 * for the time stamp.
1471 	 *
1472 	 * We can only do this for the default trace clock.
1473 	 */
1474 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1475 		event->attr.use_clockid == 0)
1476 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1477 
1478 	if (has_branch_stack(event))
1479 		data->br_stack = &cpuc->lbr_stack;
1480 }
1481 
1482 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1483 				    struct pebs_gprs *gprs)
1484 {
1485 	regs->ax = gprs->ax;
1486 	regs->bx = gprs->bx;
1487 	regs->cx = gprs->cx;
1488 	regs->dx = gprs->dx;
1489 	regs->si = gprs->si;
1490 	regs->di = gprs->di;
1491 	regs->bp = gprs->bp;
1492 	regs->sp = gprs->sp;
1493 #ifndef CONFIG_X86_32
1494 	regs->r8 = gprs->r8;
1495 	regs->r9 = gprs->r9;
1496 	regs->r10 = gprs->r10;
1497 	regs->r11 = gprs->r11;
1498 	regs->r12 = gprs->r12;
1499 	regs->r13 = gprs->r13;
1500 	regs->r14 = gprs->r14;
1501 	regs->r15 = gprs->r15;
1502 #endif
1503 }
1504 
1505 /*
1506  * With adaptive PEBS the layout depends on what fields are configured.
1507  */
1508 
1509 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1510 					    struct pt_regs *iregs, void *__pebs,
1511 					    struct perf_sample_data *data,
1512 					    struct pt_regs *regs)
1513 {
1514 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1515 	struct pebs_basic *basic = __pebs;
1516 	void *next_record = basic + 1;
1517 	u64 sample_type;
1518 	u64 format_size;
1519 	struct pebs_meminfo *meminfo = NULL;
1520 	struct pebs_gprs *gprs = NULL;
1521 	struct x86_perf_regs *perf_regs;
1522 
1523 	if (basic == NULL)
1524 		return;
1525 
1526 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
1527 	perf_regs->xmm_regs = NULL;
1528 
1529 	sample_type = event->attr.sample_type;
1530 	format_size = basic->format_size;
1531 	perf_sample_data_init(data, 0, event->hw.last_period);
1532 	data->period = event->hw.last_period;
1533 
1534 	if (event->attr.use_clockid == 0)
1535 		data->time = native_sched_clock_from_tsc(basic->tsc);
1536 
1537 	/*
1538 	 * We must however always use iregs for the unwinder to stay sane; the
1539 	 * record BP,SP,IP can point into thin air when the record is from a
1540 	 * previous PMI context or an (I)RET happened between the record and
1541 	 * PMI.
1542 	 */
1543 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1544 		data->callchain = perf_callchain(event, iregs);
1545 
1546 	*regs = *iregs;
1547 	/* The ip in basic is EventingIP */
1548 	set_linear_ip(regs, basic->ip);
1549 	regs->flags = PERF_EFLAGS_EXACT;
1550 
1551 	/*
1552 	 * The record for MEMINFO is in front of GP
1553 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1554 	 * Save the pointer here but process later.
1555 	 */
1556 	if (format_size & PEBS_DATACFG_MEMINFO) {
1557 		meminfo = next_record;
1558 		next_record = meminfo + 1;
1559 	}
1560 
1561 	if (format_size & PEBS_DATACFG_GP) {
1562 		gprs = next_record;
1563 		next_record = gprs + 1;
1564 
1565 		if (event->attr.precise_ip < 2) {
1566 			set_linear_ip(regs, gprs->ip);
1567 			regs->flags &= ~PERF_EFLAGS_EXACT;
1568 		}
1569 
1570 		if (sample_type & PERF_SAMPLE_REGS_INTR)
1571 			adaptive_pebs_save_regs(regs, gprs);
1572 	}
1573 
1574 	if (format_size & PEBS_DATACFG_MEMINFO) {
1575 		if (sample_type & PERF_SAMPLE_WEIGHT)
1576 			data->weight = meminfo->latency ?:
1577 				intel_get_tsx_weight(meminfo->tsx_tuning);
1578 
1579 		if (sample_type & PERF_SAMPLE_DATA_SRC)
1580 			data->data_src.val = get_data_src(event, meminfo->aux);
1581 
1582 		if (sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR))
1583 			data->addr = meminfo->address;
1584 
1585 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1586 			data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1587 							  gprs ? gprs->ax : 0);
1588 	}
1589 
1590 	if (format_size & PEBS_DATACFG_XMMS) {
1591 		struct pebs_xmm *xmm = next_record;
1592 
1593 		next_record = xmm + 1;
1594 		perf_regs->xmm_regs = xmm->xmm;
1595 	}
1596 
1597 	if (format_size & PEBS_DATACFG_LBRS) {
1598 		struct pebs_lbr *lbr = next_record;
1599 		int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1600 					& 0xff) + 1;
1601 		next_record = next_record + num_lbr*sizeof(struct pebs_lbr_entry);
1602 
1603 		if (has_branch_stack(event)) {
1604 			intel_pmu_store_pebs_lbrs(lbr);
1605 			data->br_stack = &cpuc->lbr_stack;
1606 		}
1607 	}
1608 
1609 	WARN_ONCE(next_record != __pebs + (format_size >> 48),
1610 			"PEBS record size %llu, expected %llu, config %llx\n",
1611 			format_size >> 48,
1612 			(u64)(next_record - __pebs),
1613 			basic->format_size);
1614 }
1615 
1616 static inline void *
1617 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1618 {
1619 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1620 	void *at;
1621 	u64 pebs_status;
1622 
1623 	/*
1624 	 * fmt0 does not have a status bitfield (does not use
1625 	 * perf_record_nhm format)
1626 	 */
1627 	if (x86_pmu.intel_cap.pebs_format < 1)
1628 		return base;
1629 
1630 	if (base == NULL)
1631 		return NULL;
1632 
1633 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1634 		unsigned long status = get_pebs_status(at);
1635 
1636 		if (test_bit(bit, (unsigned long *)&status)) {
1637 			/* PEBS v3 has accurate status bits */
1638 			if (x86_pmu.intel_cap.pebs_format >= 3)
1639 				return at;
1640 
1641 			if (status == (1 << bit))
1642 				return at;
1643 
1644 			/* clear non-PEBS bit and re-check */
1645 			pebs_status = status & cpuc->pebs_enabled;
1646 			pebs_status &= PEBS_COUNTER_MASK;
1647 			if (pebs_status == (1 << bit))
1648 				return at;
1649 		}
1650 	}
1651 	return NULL;
1652 }
1653 
1654 void intel_pmu_auto_reload_read(struct perf_event *event)
1655 {
1656 	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1657 
1658 	perf_pmu_disable(event->pmu);
1659 	intel_pmu_drain_pebs_buffer();
1660 	perf_pmu_enable(event->pmu);
1661 }
1662 
1663 /*
1664  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1665  */
1666 static int
1667 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1668 {
1669 	struct hw_perf_event *hwc = &event->hw;
1670 	int shift = 64 - x86_pmu.cntval_bits;
1671 	u64 period = hwc->sample_period;
1672 	u64 prev_raw_count, new_raw_count;
1673 	s64 new, old;
1674 
1675 	WARN_ON(!period);
1676 
1677 	/*
1678 	 * drain_pebs() only happens when the PMU is disabled.
1679 	 */
1680 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1681 
1682 	prev_raw_count = local64_read(&hwc->prev_count);
1683 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1684 	local64_set(&hwc->prev_count, new_raw_count);
1685 
1686 	/*
1687 	 * Since the counter increments a negative counter value and
1688 	 * overflows on the sign switch, giving the interval:
1689 	 *
1690 	 *   [-period, 0]
1691 	 *
1692 	 * the difference between two consequtive reads is:
1693 	 *
1694 	 *   A) value2 - value1;
1695 	 *      when no overflows have happened in between,
1696 	 *
1697 	 *   B) (0 - value1) + (value2 - (-period));
1698 	 *      when one overflow happened in between,
1699 	 *
1700 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1701 	 *      when @n overflows happened in between.
1702 	 *
1703 	 * Here A) is the obvious difference, B) is the extension to the
1704 	 * discrete interval, where the first term is to the top of the
1705 	 * interval and the second term is from the bottom of the next
1706 	 * interval and C) the extension to multiple intervals, where the
1707 	 * middle term is the whole intervals covered.
1708 	 *
1709 	 * An equivalent of C, by reduction, is:
1710 	 *
1711 	 *   value2 - value1 + n * period
1712 	 */
1713 	new = ((s64)(new_raw_count << shift) >> shift);
1714 	old = ((s64)(prev_raw_count << shift) >> shift);
1715 	local64_add(new - old + count * period, &event->count);
1716 
1717 	local64_set(&hwc->period_left, -new);
1718 
1719 	perf_event_update_userpage(event);
1720 
1721 	return 0;
1722 }
1723 
1724 static void __intel_pmu_pebs_event(struct perf_event *event,
1725 				   struct pt_regs *iregs,
1726 				   void *base, void *top,
1727 				   int bit, int count,
1728 				   void (*setup_sample)(struct perf_event *,
1729 						struct pt_regs *,
1730 						void *,
1731 						struct perf_sample_data *,
1732 						struct pt_regs *))
1733 {
1734 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1735 	struct hw_perf_event *hwc = &event->hw;
1736 	struct perf_sample_data data;
1737 	struct x86_perf_regs perf_regs;
1738 	struct pt_regs *regs = &perf_regs.regs;
1739 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1740 
1741 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1742 		/*
1743 		 * Now, auto-reload is only enabled in fixed period mode.
1744 		 * The reload value is always hwc->sample_period.
1745 		 * May need to change it, if auto-reload is enabled in
1746 		 * freq mode later.
1747 		 */
1748 		intel_pmu_save_and_restart_reload(event, count);
1749 	} else if (!intel_pmu_save_and_restart(event))
1750 		return;
1751 
1752 	while (count > 1) {
1753 		setup_sample(event, iregs, at, &data, regs);
1754 		perf_event_output(event, &data, regs);
1755 		at += cpuc->pebs_record_size;
1756 		at = get_next_pebs_record_by_bit(at, top, bit);
1757 		count--;
1758 	}
1759 
1760 	setup_sample(event, iregs, at, &data, regs);
1761 
1762 	/*
1763 	 * All but the last records are processed.
1764 	 * The last one is left to be able to call the overflow handler.
1765 	 */
1766 	if (perf_event_overflow(event, &data, regs)) {
1767 		x86_pmu_stop(event, 0);
1768 		return;
1769 	}
1770 
1771 }
1772 
1773 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1774 {
1775 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1776 	struct debug_store *ds = cpuc->ds;
1777 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1778 	struct pebs_record_core *at, *top;
1779 	int n;
1780 
1781 	if (!x86_pmu.pebs_active)
1782 		return;
1783 
1784 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1785 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1786 
1787 	/*
1788 	 * Whatever else happens, drain the thing
1789 	 */
1790 	ds->pebs_index = ds->pebs_buffer_base;
1791 
1792 	if (!test_bit(0, cpuc->active_mask))
1793 		return;
1794 
1795 	WARN_ON_ONCE(!event);
1796 
1797 	if (!event->attr.precise_ip)
1798 		return;
1799 
1800 	n = top - at;
1801 	if (n <= 0) {
1802 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1803 			intel_pmu_save_and_restart_reload(event, 0);
1804 		return;
1805 	}
1806 
1807 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n,
1808 			       setup_pebs_fixed_sample_data);
1809 }
1810 
1811 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
1812 {
1813 	struct perf_event *event;
1814 	int bit;
1815 
1816 	/*
1817 	 * The drain_pebs() could be called twice in a short period
1818 	 * for auto-reload event in pmu::read(). There are no
1819 	 * overflows have happened in between.
1820 	 * It needs to call intel_pmu_save_and_restart_reload() to
1821 	 * update the event->count for this case.
1822 	 */
1823 	for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
1824 		event = cpuc->events[bit];
1825 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1826 			intel_pmu_save_and_restart_reload(event, 0);
1827 	}
1828 }
1829 
1830 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1831 {
1832 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1833 	struct debug_store *ds = cpuc->ds;
1834 	struct perf_event *event;
1835 	void *base, *at, *top;
1836 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1837 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1838 	int bit, i, size;
1839 	u64 mask;
1840 
1841 	if (!x86_pmu.pebs_active)
1842 		return;
1843 
1844 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1845 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1846 
1847 	ds->pebs_index = ds->pebs_buffer_base;
1848 
1849 	mask = (1ULL << x86_pmu.max_pebs_events) - 1;
1850 	size = x86_pmu.max_pebs_events;
1851 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
1852 		mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
1853 		size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1854 	}
1855 
1856 	if (unlikely(base >= top)) {
1857 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
1858 		return;
1859 	}
1860 
1861 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1862 		struct pebs_record_nhm *p = at;
1863 		u64 pebs_status;
1864 
1865 		pebs_status = p->status & cpuc->pebs_enabled;
1866 		pebs_status &= mask;
1867 
1868 		/* PEBS v3 has more accurate status bits */
1869 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1870 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
1871 				counts[bit]++;
1872 
1873 			continue;
1874 		}
1875 
1876 		/*
1877 		 * On some CPUs the PEBS status can be zero when PEBS is
1878 		 * racing with clearing of GLOBAL_STATUS.
1879 		 *
1880 		 * Normally we would drop that record, but in the
1881 		 * case when there is only a single active PEBS event
1882 		 * we can assume it's for that event.
1883 		 */
1884 		if (!pebs_status && cpuc->pebs_enabled &&
1885 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1886 			pebs_status = cpuc->pebs_enabled;
1887 
1888 		bit = find_first_bit((unsigned long *)&pebs_status,
1889 					x86_pmu.max_pebs_events);
1890 		if (bit >= x86_pmu.max_pebs_events)
1891 			continue;
1892 
1893 		/*
1894 		 * The PEBS hardware does not deal well with the situation
1895 		 * when events happen near to each other and multiple bits
1896 		 * are set. But it should happen rarely.
1897 		 *
1898 		 * If these events include one PEBS and multiple non-PEBS
1899 		 * events, it doesn't impact PEBS record. The record will
1900 		 * be handled normally. (slow path)
1901 		 *
1902 		 * If these events include two or more PEBS events, the
1903 		 * records for the events can be collapsed into a single
1904 		 * one, and it's not possible to reconstruct all events
1905 		 * that caused the PEBS record. It's called collision.
1906 		 * If collision happened, the record will be dropped.
1907 		 */
1908 		if (p->status != (1ULL << bit)) {
1909 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
1910 				error[i]++;
1911 			continue;
1912 		}
1913 
1914 		counts[bit]++;
1915 	}
1916 
1917 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
1918 		if ((counts[bit] == 0) && (error[bit] == 0))
1919 			continue;
1920 
1921 		event = cpuc->events[bit];
1922 		if (WARN_ON_ONCE(!event))
1923 			continue;
1924 
1925 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1926 			continue;
1927 
1928 		/* log dropped samples number */
1929 		if (error[bit]) {
1930 			perf_log_lost_samples(event, error[bit]);
1931 
1932 			if (perf_event_account_interrupt(event))
1933 				x86_pmu_stop(event, 0);
1934 		}
1935 
1936 		if (counts[bit]) {
1937 			__intel_pmu_pebs_event(event, iregs, base,
1938 					       top, bit, counts[bit],
1939 					       setup_pebs_fixed_sample_data);
1940 		}
1941 	}
1942 }
1943 
1944 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs)
1945 {
1946 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1947 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1948 	struct debug_store *ds = cpuc->ds;
1949 	struct perf_event *event;
1950 	void *base, *at, *top;
1951 	int bit, size;
1952 	u64 mask;
1953 
1954 	if (!x86_pmu.pebs_active)
1955 		return;
1956 
1957 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
1958 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
1959 
1960 	ds->pebs_index = ds->pebs_buffer_base;
1961 
1962 	mask = ((1ULL << x86_pmu.max_pebs_events) - 1) |
1963 	       (((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
1964 	size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
1965 
1966 	if (unlikely(base >= top)) {
1967 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
1968 		return;
1969 	}
1970 
1971 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1972 		u64 pebs_status;
1973 
1974 		pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
1975 		pebs_status &= mask;
1976 
1977 		for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
1978 			counts[bit]++;
1979 	}
1980 
1981 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
1982 		if (counts[bit] == 0)
1983 			continue;
1984 
1985 		event = cpuc->events[bit];
1986 		if (WARN_ON_ONCE(!event))
1987 			continue;
1988 
1989 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1990 			continue;
1991 
1992 		__intel_pmu_pebs_event(event, iregs, base,
1993 				       top, bit, counts[bit],
1994 				       setup_pebs_adaptive_sample_data);
1995 	}
1996 }
1997 
1998 /*
1999  * BTS, PEBS probe and setup
2000  */
2001 
2002 void __init intel_ds_init(void)
2003 {
2004 	/*
2005 	 * No support for 32bit formats
2006 	 */
2007 	if (!boot_cpu_has(X86_FEATURE_DTES64))
2008 		return;
2009 
2010 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
2011 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
2012 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
2013 	if (x86_pmu.version <= 4)
2014 		x86_pmu.pebs_no_isolation = 1;
2015 
2016 	if (x86_pmu.pebs) {
2017 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
2018 		char *pebs_qual = "";
2019 		int format = x86_pmu.intel_cap.pebs_format;
2020 
2021 		if (format < 4)
2022 			x86_pmu.intel_cap.pebs_baseline = 0;
2023 
2024 		switch (format) {
2025 		case 0:
2026 			pr_cont("PEBS fmt0%c, ", pebs_type);
2027 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
2028 			/*
2029 			 * Using >PAGE_SIZE buffers makes the WRMSR to
2030 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
2031 			 * mysteriously hang on Core2.
2032 			 *
2033 			 * As a workaround, we don't do this.
2034 			 */
2035 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
2036 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
2037 			break;
2038 
2039 		case 1:
2040 			pr_cont("PEBS fmt1%c, ", pebs_type);
2041 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
2042 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2043 			break;
2044 
2045 		case 2:
2046 			pr_cont("PEBS fmt2%c, ", pebs_type);
2047 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2048 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2049 			break;
2050 
2051 		case 3:
2052 			pr_cont("PEBS fmt3%c, ", pebs_type);
2053 			x86_pmu.pebs_record_size =
2054 						sizeof(struct pebs_record_skl);
2055 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2056 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2057 			break;
2058 
2059 		case 4:
2060 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2061 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2062 			if (x86_pmu.intel_cap.pebs_baseline) {
2063 				x86_pmu.large_pebs_flags |=
2064 					PERF_SAMPLE_BRANCH_STACK |
2065 					PERF_SAMPLE_TIME;
2066 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
2067 				pebs_qual = "-baseline";
2068 				x86_get_pmu()->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
2069 			} else {
2070 				/* Only basic record supported */
2071 				x86_pmu.large_pebs_flags &=
2072 					~(PERF_SAMPLE_ADDR |
2073 					  PERF_SAMPLE_TIME |
2074 					  PERF_SAMPLE_DATA_SRC |
2075 					  PERF_SAMPLE_TRANSACTION |
2076 					  PERF_SAMPLE_REGS_USER |
2077 					  PERF_SAMPLE_REGS_INTR);
2078 			}
2079 			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2080 
2081 			if (x86_pmu.intel_cap.pebs_output_pt_available) {
2082 				pr_cont("PEBS-via-PT, ");
2083 				x86_get_pmu()->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
2084 			}
2085 
2086 			break;
2087 
2088 		default:
2089 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2090 			x86_pmu.pebs = 0;
2091 		}
2092 	}
2093 }
2094 
2095 void perf_restore_debug_store(void)
2096 {
2097 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2098 
2099 	if (!x86_pmu.bts && !x86_pmu.pebs)
2100 		return;
2101 
2102 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
2103 }
2104