xref: /openbmc/linux/arch/x86/events/intel/ds.c (revision 7a2f6f61)
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_data_blk:1;
40 		unsigned int ld_addr_blk:1;
41 		unsigned int ld_reserved:24;
42 	};
43 	struct {
44 		unsigned int st_l1d_hit:1;
45 		unsigned int st_reserved1:3;
46 		unsigned int st_stlb_miss:1;
47 		unsigned int st_locked:1;
48 		unsigned int st_reserved2:26;
49 	};
50 	struct {
51 		unsigned int st_lat_dse:4;
52 		unsigned int st_lat_stlb_miss:1;
53 		unsigned int st_lat_locked:1;
54 		unsigned int ld_reserved3:26;
55 	};
56 };
57 
58 
59 /*
60  * Map PEBS Load Latency Data Source encodings to generic
61  * memory data source information
62  */
63 #define P(a, b) PERF_MEM_S(a, b)
64 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
65 #define LEVEL(x) P(LVLNUM, x)
66 #define REM P(REMOTE, REMOTE)
67 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
68 
69 /* Version for Sandy Bridge and later */
70 static u64 pebs_data_source[] = {
71 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
72 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
73 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
74 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
75 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
76 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
77 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
78 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
79 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
80 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
81 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
82 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
83 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
84 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
85 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
86 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
87 };
88 
89 /* Patch up minor differences in the bits */
90 void __init intel_pmu_pebs_data_source_nhm(void)
91 {
92 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
93 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
94 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
95 }
96 
97 void __init intel_pmu_pebs_data_source_skl(bool pmem)
98 {
99 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
100 
101 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
102 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
103 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
104 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
105 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
106 }
107 
108 static u64 precise_store_data(u64 status)
109 {
110 	union intel_x86_pebs_dse dse;
111 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
112 
113 	dse.val = status;
114 
115 	/*
116 	 * bit 4: TLB access
117 	 * 1 = stored missed 2nd level TLB
118 	 *
119 	 * so it either hit the walker or the OS
120 	 * otherwise hit 2nd level TLB
121 	 */
122 	if (dse.st_stlb_miss)
123 		val |= P(TLB, MISS);
124 	else
125 		val |= P(TLB, HIT);
126 
127 	/*
128 	 * bit 0: hit L1 data cache
129 	 * if not set, then all we know is that
130 	 * it missed L1D
131 	 */
132 	if (dse.st_l1d_hit)
133 		val |= P(LVL, HIT);
134 	else
135 		val |= P(LVL, MISS);
136 
137 	/*
138 	 * bit 5: Locked prefix
139 	 */
140 	if (dse.st_locked)
141 		val |= P(LOCK, LOCKED);
142 
143 	return val;
144 }
145 
146 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
147 {
148 	union perf_mem_data_src dse;
149 
150 	dse.val = PERF_MEM_NA;
151 
152 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
153 		dse.mem_op = PERF_MEM_OP_STORE;
154 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
155 		dse.mem_op = PERF_MEM_OP_LOAD;
156 
157 	/*
158 	 * L1 info only valid for following events:
159 	 *
160 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
161 	 * MEM_UOPS_RETIRED.LOCK_STORES
162 	 * MEM_UOPS_RETIRED.SPLIT_STORES
163 	 * MEM_UOPS_RETIRED.ALL_STORES
164 	 */
165 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
166 		if (status & 1)
167 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
168 		else
169 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
170 	}
171 	return dse.val;
172 }
173 
174 static u64 load_latency_data(u64 status)
175 {
176 	union intel_x86_pebs_dse dse;
177 	u64 val;
178 
179 	dse.val = status;
180 
181 	/*
182 	 * use the mapping table for bit 0-3
183 	 */
184 	val = pebs_data_source[dse.ld_dse];
185 
186 	/*
187 	 * Nehalem models do not support TLB, Lock infos
188 	 */
189 	if (x86_pmu.pebs_no_tlb) {
190 		val |= P(TLB, NA) | P(LOCK, NA);
191 		return val;
192 	}
193 	/*
194 	 * bit 4: TLB access
195 	 * 0 = did not miss 2nd level TLB
196 	 * 1 = missed 2nd level TLB
197 	 */
198 	if (dse.ld_stlb_miss)
199 		val |= P(TLB, MISS) | P(TLB, L2);
200 	else
201 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
202 
203 	/*
204 	 * bit 5: locked prefix
205 	 */
206 	if (dse.ld_locked)
207 		val |= P(LOCK, LOCKED);
208 
209 	/*
210 	 * Ice Lake and earlier models do not support block infos.
211 	 */
212 	if (!x86_pmu.pebs_block) {
213 		val |= P(BLK, NA);
214 		return val;
215 	}
216 	/*
217 	 * bit 6: load was blocked since its data could not be forwarded
218 	 *        from a preceding store
219 	 */
220 	if (dse.ld_data_blk)
221 		val |= P(BLK, DATA);
222 
223 	/*
224 	 * bit 7: load was blocked due to potential address conflict with
225 	 *        a preceding store
226 	 */
227 	if (dse.ld_addr_blk)
228 		val |= P(BLK, ADDR);
229 
230 	if (!dse.ld_data_blk && !dse.ld_addr_blk)
231 		val |= P(BLK, NA);
232 
233 	return val;
234 }
235 
236 static u64 store_latency_data(u64 status)
237 {
238 	union intel_x86_pebs_dse dse;
239 	u64 val;
240 
241 	dse.val = status;
242 
243 	/*
244 	 * use the mapping table for bit 0-3
245 	 */
246 	val = pebs_data_source[dse.st_lat_dse];
247 
248 	/*
249 	 * bit 4: TLB access
250 	 * 0 = did not miss 2nd level TLB
251 	 * 1 = missed 2nd level TLB
252 	 */
253 	if (dse.st_lat_stlb_miss)
254 		val |= P(TLB, MISS) | P(TLB, L2);
255 	else
256 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
257 
258 	/*
259 	 * bit 5: locked prefix
260 	 */
261 	if (dse.st_lat_locked)
262 		val |= P(LOCK, LOCKED);
263 
264 	val |= P(BLK, NA);
265 
266 	return val;
267 }
268 
269 struct pebs_record_core {
270 	u64 flags, ip;
271 	u64 ax, bx, cx, dx;
272 	u64 si, di, bp, sp;
273 	u64 r8,  r9,  r10, r11;
274 	u64 r12, r13, r14, r15;
275 };
276 
277 struct pebs_record_nhm {
278 	u64 flags, ip;
279 	u64 ax, bx, cx, dx;
280 	u64 si, di, bp, sp;
281 	u64 r8,  r9,  r10, r11;
282 	u64 r12, r13, r14, r15;
283 	u64 status, dla, dse, lat;
284 };
285 
286 /*
287  * Same as pebs_record_nhm, with two additional fields.
288  */
289 struct pebs_record_hsw {
290 	u64 flags, ip;
291 	u64 ax, bx, cx, dx;
292 	u64 si, di, bp, sp;
293 	u64 r8,  r9,  r10, r11;
294 	u64 r12, r13, r14, r15;
295 	u64 status, dla, dse, lat;
296 	u64 real_ip, tsx_tuning;
297 };
298 
299 union hsw_tsx_tuning {
300 	struct {
301 		u32 cycles_last_block     : 32,
302 		    hle_abort		  : 1,
303 		    rtm_abort		  : 1,
304 		    instruction_abort     : 1,
305 		    non_instruction_abort : 1,
306 		    retry		  : 1,
307 		    data_conflict	  : 1,
308 		    capacity_writes	  : 1,
309 		    capacity_reads	  : 1;
310 	};
311 	u64	    value;
312 };
313 
314 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
315 
316 /* Same as HSW, plus TSC */
317 
318 struct pebs_record_skl {
319 	u64 flags, ip;
320 	u64 ax, bx, cx, dx;
321 	u64 si, di, bp, sp;
322 	u64 r8,  r9,  r10, r11;
323 	u64 r12, r13, r14, r15;
324 	u64 status, dla, dse, lat;
325 	u64 real_ip, tsx_tuning;
326 	u64 tsc;
327 };
328 
329 void init_debug_store_on_cpu(int cpu)
330 {
331 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
332 
333 	if (!ds)
334 		return;
335 
336 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
337 		     (u32)((u64)(unsigned long)ds),
338 		     (u32)((u64)(unsigned long)ds >> 32));
339 }
340 
341 void fini_debug_store_on_cpu(int cpu)
342 {
343 	if (!per_cpu(cpu_hw_events, cpu).ds)
344 		return;
345 
346 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
347 }
348 
349 static DEFINE_PER_CPU(void *, insn_buffer);
350 
351 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
352 {
353 	unsigned long start = (unsigned long)cea;
354 	phys_addr_t pa;
355 	size_t msz = 0;
356 
357 	pa = virt_to_phys(addr);
358 
359 	preempt_disable();
360 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
361 		cea_set_pte(cea, pa, prot);
362 
363 	/*
364 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
365 	 * all TLB entries for it.
366 	 */
367 	flush_tlb_kernel_range(start, start + size);
368 	preempt_enable();
369 }
370 
371 static void ds_clear_cea(void *cea, size_t size)
372 {
373 	unsigned long start = (unsigned long)cea;
374 	size_t msz = 0;
375 
376 	preempt_disable();
377 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
378 		cea_set_pte(cea, 0, PAGE_NONE);
379 
380 	flush_tlb_kernel_range(start, start + size);
381 	preempt_enable();
382 }
383 
384 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
385 {
386 	unsigned int order = get_order(size);
387 	int node = cpu_to_node(cpu);
388 	struct page *page;
389 
390 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
391 	return page ? page_address(page) : NULL;
392 }
393 
394 static void dsfree_pages(const void *buffer, size_t size)
395 {
396 	if (buffer)
397 		free_pages((unsigned long)buffer, get_order(size));
398 }
399 
400 static int alloc_pebs_buffer(int cpu)
401 {
402 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
403 	struct debug_store *ds = hwev->ds;
404 	size_t bsiz = x86_pmu.pebs_buffer_size;
405 	int max, node = cpu_to_node(cpu);
406 	void *buffer, *insn_buff, *cea;
407 
408 	if (!x86_pmu.pebs)
409 		return 0;
410 
411 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
412 	if (unlikely(!buffer))
413 		return -ENOMEM;
414 
415 	/*
416 	 * HSW+ already provides us the eventing ip; no need to allocate this
417 	 * buffer then.
418 	 */
419 	if (x86_pmu.intel_cap.pebs_format < 2) {
420 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
421 		if (!insn_buff) {
422 			dsfree_pages(buffer, bsiz);
423 			return -ENOMEM;
424 		}
425 		per_cpu(insn_buffer, cpu) = insn_buff;
426 	}
427 	hwev->ds_pebs_vaddr = buffer;
428 	/* Update the cpu entry area mapping */
429 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
430 	ds->pebs_buffer_base = (unsigned long) cea;
431 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
432 	ds->pebs_index = ds->pebs_buffer_base;
433 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
434 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
435 	return 0;
436 }
437 
438 static void release_pebs_buffer(int cpu)
439 {
440 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
441 	void *cea;
442 
443 	if (!x86_pmu.pebs)
444 		return;
445 
446 	kfree(per_cpu(insn_buffer, cpu));
447 	per_cpu(insn_buffer, cpu) = NULL;
448 
449 	/* Clear the fixmap */
450 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
451 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
452 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
453 	hwev->ds_pebs_vaddr = NULL;
454 }
455 
456 static int alloc_bts_buffer(int cpu)
457 {
458 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
459 	struct debug_store *ds = hwev->ds;
460 	void *buffer, *cea;
461 	int max;
462 
463 	if (!x86_pmu.bts)
464 		return 0;
465 
466 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
467 	if (unlikely(!buffer)) {
468 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
469 		return -ENOMEM;
470 	}
471 	hwev->ds_bts_vaddr = buffer;
472 	/* Update the fixmap */
473 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
474 	ds->bts_buffer_base = (unsigned long) cea;
475 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
476 	ds->bts_index = ds->bts_buffer_base;
477 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
478 	ds->bts_absolute_maximum = ds->bts_buffer_base +
479 					max * BTS_RECORD_SIZE;
480 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
481 					(max / 16) * BTS_RECORD_SIZE;
482 	return 0;
483 }
484 
485 static void release_bts_buffer(int cpu)
486 {
487 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
488 	void *cea;
489 
490 	if (!x86_pmu.bts)
491 		return;
492 
493 	/* Clear the fixmap */
494 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
495 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
496 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
497 	hwev->ds_bts_vaddr = NULL;
498 }
499 
500 static int alloc_ds_buffer(int cpu)
501 {
502 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
503 
504 	memset(ds, 0, sizeof(*ds));
505 	per_cpu(cpu_hw_events, cpu).ds = ds;
506 	return 0;
507 }
508 
509 static void release_ds_buffer(int cpu)
510 {
511 	per_cpu(cpu_hw_events, cpu).ds = NULL;
512 }
513 
514 void release_ds_buffers(void)
515 {
516 	int cpu;
517 
518 	if (!x86_pmu.bts && !x86_pmu.pebs)
519 		return;
520 
521 	for_each_possible_cpu(cpu)
522 		release_ds_buffer(cpu);
523 
524 	for_each_possible_cpu(cpu) {
525 		/*
526 		 * Again, ignore errors from offline CPUs, they will no longer
527 		 * observe cpu_hw_events.ds and not program the DS_AREA when
528 		 * they come up.
529 		 */
530 		fini_debug_store_on_cpu(cpu);
531 	}
532 
533 	for_each_possible_cpu(cpu) {
534 		release_pebs_buffer(cpu);
535 		release_bts_buffer(cpu);
536 	}
537 }
538 
539 void reserve_ds_buffers(void)
540 {
541 	int bts_err = 0, pebs_err = 0;
542 	int cpu;
543 
544 	x86_pmu.bts_active = 0;
545 	x86_pmu.pebs_active = 0;
546 
547 	if (!x86_pmu.bts && !x86_pmu.pebs)
548 		return;
549 
550 	if (!x86_pmu.bts)
551 		bts_err = 1;
552 
553 	if (!x86_pmu.pebs)
554 		pebs_err = 1;
555 
556 	for_each_possible_cpu(cpu) {
557 		if (alloc_ds_buffer(cpu)) {
558 			bts_err = 1;
559 			pebs_err = 1;
560 		}
561 
562 		if (!bts_err && alloc_bts_buffer(cpu))
563 			bts_err = 1;
564 
565 		if (!pebs_err && alloc_pebs_buffer(cpu))
566 			pebs_err = 1;
567 
568 		if (bts_err && pebs_err)
569 			break;
570 	}
571 
572 	if (bts_err) {
573 		for_each_possible_cpu(cpu)
574 			release_bts_buffer(cpu);
575 	}
576 
577 	if (pebs_err) {
578 		for_each_possible_cpu(cpu)
579 			release_pebs_buffer(cpu);
580 	}
581 
582 	if (bts_err && pebs_err) {
583 		for_each_possible_cpu(cpu)
584 			release_ds_buffer(cpu);
585 	} else {
586 		if (x86_pmu.bts && !bts_err)
587 			x86_pmu.bts_active = 1;
588 
589 		if (x86_pmu.pebs && !pebs_err)
590 			x86_pmu.pebs_active = 1;
591 
592 		for_each_possible_cpu(cpu) {
593 			/*
594 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
595 			 * will get this call through intel_pmu_cpu_starting().
596 			 */
597 			init_debug_store_on_cpu(cpu);
598 		}
599 	}
600 }
601 
602 /*
603  * BTS
604  */
605 
606 struct event_constraint bts_constraint =
607 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
608 
609 void intel_pmu_enable_bts(u64 config)
610 {
611 	unsigned long debugctlmsr;
612 
613 	debugctlmsr = get_debugctlmsr();
614 
615 	debugctlmsr |= DEBUGCTLMSR_TR;
616 	debugctlmsr |= DEBUGCTLMSR_BTS;
617 	if (config & ARCH_PERFMON_EVENTSEL_INT)
618 		debugctlmsr |= DEBUGCTLMSR_BTINT;
619 
620 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
621 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
622 
623 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
624 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
625 
626 	update_debugctlmsr(debugctlmsr);
627 }
628 
629 void intel_pmu_disable_bts(void)
630 {
631 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
632 	unsigned long debugctlmsr;
633 
634 	if (!cpuc->ds)
635 		return;
636 
637 	debugctlmsr = get_debugctlmsr();
638 
639 	debugctlmsr &=
640 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
641 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
642 
643 	update_debugctlmsr(debugctlmsr);
644 }
645 
646 int intel_pmu_drain_bts_buffer(void)
647 {
648 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
649 	struct debug_store *ds = cpuc->ds;
650 	struct bts_record {
651 		u64	from;
652 		u64	to;
653 		u64	flags;
654 	};
655 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
656 	struct bts_record *at, *base, *top;
657 	struct perf_output_handle handle;
658 	struct perf_event_header header;
659 	struct perf_sample_data data;
660 	unsigned long skip = 0;
661 	struct pt_regs regs;
662 
663 	if (!event)
664 		return 0;
665 
666 	if (!x86_pmu.bts_active)
667 		return 0;
668 
669 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
670 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
671 
672 	if (top <= base)
673 		return 0;
674 
675 	memset(&regs, 0, sizeof(regs));
676 
677 	ds->bts_index = ds->bts_buffer_base;
678 
679 	perf_sample_data_init(&data, 0, event->hw.last_period);
680 
681 	/*
682 	 * BTS leaks kernel addresses in branches across the cpl boundary,
683 	 * such as traps or system calls, so unless the user is asking for
684 	 * kernel tracing (and right now it's not possible), we'd need to
685 	 * filter them out. But first we need to count how many of those we
686 	 * have in the current batch. This is an extra O(n) pass, however,
687 	 * it's much faster than the other one especially considering that
688 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
689 	 * alloc_bts_buffer()).
690 	 */
691 	for (at = base; at < top; at++) {
692 		/*
693 		 * Note that right now *this* BTS code only works if
694 		 * attr::exclude_kernel is set, but let's keep this extra
695 		 * check here in case that changes.
696 		 */
697 		if (event->attr.exclude_kernel &&
698 		    (kernel_ip(at->from) || kernel_ip(at->to)))
699 			skip++;
700 	}
701 
702 	/*
703 	 * Prepare a generic sample, i.e. fill in the invariant fields.
704 	 * We will overwrite the from and to address before we output
705 	 * the sample.
706 	 */
707 	rcu_read_lock();
708 	perf_prepare_sample(&header, &data, event, &regs);
709 
710 	if (perf_output_begin(&handle, &data, event,
711 			      header.size * (top - base - skip)))
712 		goto unlock;
713 
714 	for (at = base; at < top; at++) {
715 		/* Filter out any records that contain kernel addresses. */
716 		if (event->attr.exclude_kernel &&
717 		    (kernel_ip(at->from) || kernel_ip(at->to)))
718 			continue;
719 
720 		data.ip		= at->from;
721 		data.addr	= at->to;
722 
723 		perf_output_sample(&handle, &header, &data, event);
724 	}
725 
726 	perf_output_end(&handle);
727 
728 	/* There's new data available. */
729 	event->hw.interrupts++;
730 	event->pending_kill = POLL_IN;
731 unlock:
732 	rcu_read_unlock();
733 	return 1;
734 }
735 
736 static inline void intel_pmu_drain_pebs_buffer(void)
737 {
738 	struct perf_sample_data data;
739 
740 	x86_pmu.drain_pebs(NULL, &data);
741 }
742 
743 /*
744  * PEBS
745  */
746 struct event_constraint intel_core2_pebs_event_constraints[] = {
747 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
748 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
749 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
750 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
751 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
752 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
753 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
754 	EVENT_CONSTRAINT_END
755 };
756 
757 struct event_constraint intel_atom_pebs_event_constraints[] = {
758 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
759 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
760 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
761 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
762 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
763 	/* Allow all events as PEBS with no flags */
764 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
765 	EVENT_CONSTRAINT_END
766 };
767 
768 struct event_constraint intel_slm_pebs_event_constraints[] = {
769 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
770 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
771 	/* Allow all events as PEBS with no flags */
772 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
773 	EVENT_CONSTRAINT_END
774 };
775 
776 struct event_constraint intel_glm_pebs_event_constraints[] = {
777 	/* Allow all events as PEBS with no flags */
778 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
779 	EVENT_CONSTRAINT_END
780 };
781 
782 struct event_constraint intel_grt_pebs_event_constraints[] = {
783 	/* Allow all events as PEBS with no flags */
784 	INTEL_PLD_CONSTRAINT(0x5d0, 0xf),
785 	INTEL_PSD_CONSTRAINT(0x6d0, 0xf),
786 	EVENT_CONSTRAINT_END
787 };
788 
789 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
790 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
791 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
792 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
793 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
794 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
795 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
796 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
797 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
798 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
799 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
800 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
801 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
802 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
803 	EVENT_CONSTRAINT_END
804 };
805 
806 struct event_constraint intel_westmere_pebs_event_constraints[] = {
807 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
808 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
809 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
810 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
811 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
812 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
813 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
814 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
815 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
816 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
817 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
818 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
819 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
820 	EVENT_CONSTRAINT_END
821 };
822 
823 struct event_constraint intel_snb_pebs_event_constraints[] = {
824 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
825 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
826 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
827 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
828 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
829         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
830         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
831         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
832         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
833 	/* Allow all events as PEBS with no flags */
834 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
835 	EVENT_CONSTRAINT_END
836 };
837 
838 struct event_constraint intel_ivb_pebs_event_constraints[] = {
839         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
840         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
841 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
842 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
843 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
844 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
845 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
846 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
847 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
848 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
849 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
850 	/* Allow all events as PEBS with no flags */
851 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
852         EVENT_CONSTRAINT_END
853 };
854 
855 struct event_constraint intel_hsw_pebs_event_constraints[] = {
856 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
857 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
858 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
859 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
860 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
861 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
862 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
863 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
864 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
865 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
866 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
867 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
868 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
869 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
870 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
871 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
872 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
873 	/* Allow all events as PEBS with no flags */
874 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
875 	EVENT_CONSTRAINT_END
876 };
877 
878 struct event_constraint intel_bdw_pebs_event_constraints[] = {
879 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
880 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
881 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
882 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
883 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
884 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
885 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
886 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
887 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
888 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
889 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
890 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
891 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
892 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
893 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
894 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
895 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
896 	/* Allow all events as PEBS with no flags */
897 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
898 	EVENT_CONSTRAINT_END
899 };
900 
901 
902 struct event_constraint intel_skl_pebs_event_constraints[] = {
903 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
904 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
905 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
906 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
907 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
908 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
909 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
910 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
911 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
912 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
913 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
914 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
915 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
916 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
917 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
918 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
919 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
920 	/* Allow all events as PEBS with no flags */
921 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
922 	EVENT_CONSTRAINT_END
923 };
924 
925 struct event_constraint intel_icl_pebs_event_constraints[] = {
926 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL),	/* old INST_RETIRED.PREC_DIST */
927 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
928 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
929 
930 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
931 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),	/* MEM_INST_RETIRED.LOAD */
932 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),	/* MEM_INST_RETIRED.STORE */
933 
934 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
935 
936 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
937 
938 	/*
939 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
940 	 * need the full constraints from the main table.
941 	 */
942 
943 	EVENT_CONSTRAINT_END
944 };
945 
946 struct event_constraint intel_spr_pebs_event_constraints[] = {
947 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
948 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
949 
950 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
951 	INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
952 	INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
953 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),
954 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),
955 
956 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
957 
958 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
959 
960 	/*
961 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
962 	 * need the full constraints from the main table.
963 	 */
964 
965 	EVENT_CONSTRAINT_END
966 };
967 
968 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
969 {
970 	struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
971 	struct event_constraint *c;
972 
973 	if (!event->attr.precise_ip)
974 		return NULL;
975 
976 	if (pebs_constraints) {
977 		for_each_event_constraint(c, pebs_constraints) {
978 			if (constraint_match(c, event->hw.config)) {
979 				event->hw.flags |= c->flags;
980 				return c;
981 			}
982 		}
983 	}
984 
985 	/*
986 	 * Extended PEBS support
987 	 * Makes the PEBS code search the normal constraints.
988 	 */
989 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
990 		return NULL;
991 
992 	return &emptyconstraint;
993 }
994 
995 /*
996  * We need the sched_task callback even for per-cpu events when we use
997  * the large interrupt threshold, such that we can provide PID and TID
998  * to PEBS samples.
999  */
1000 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
1001 {
1002 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
1003 		return false;
1004 
1005 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
1006 }
1007 
1008 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
1009 {
1010 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1011 
1012 	if (!sched_in && pebs_needs_sched_cb(cpuc))
1013 		intel_pmu_drain_pebs_buffer();
1014 }
1015 
1016 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1017 {
1018 	struct debug_store *ds = cpuc->ds;
1019 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
1020 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
1021 	u64 threshold;
1022 	int reserved;
1023 
1024 	if (cpuc->n_pebs_via_pt)
1025 		return;
1026 
1027 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1028 		reserved = max_pebs_events + num_counters_fixed;
1029 	else
1030 		reserved = max_pebs_events;
1031 
1032 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
1033 		threshold = ds->pebs_absolute_maximum -
1034 			reserved * cpuc->pebs_record_size;
1035 	} else {
1036 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1037 	}
1038 
1039 	ds->pebs_interrupt_threshold = threshold;
1040 }
1041 
1042 static void adaptive_pebs_record_size_update(void)
1043 {
1044 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1045 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1046 	int sz = sizeof(struct pebs_basic);
1047 
1048 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1049 		sz += sizeof(struct pebs_meminfo);
1050 	if (pebs_data_cfg & PEBS_DATACFG_GP)
1051 		sz += sizeof(struct pebs_gprs);
1052 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1053 		sz += sizeof(struct pebs_xmm);
1054 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1055 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1056 
1057 	cpuc->pebs_record_size = sz;
1058 }
1059 
1060 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
1061 				PERF_SAMPLE_PHYS_ADDR |			     \
1062 				PERF_SAMPLE_WEIGHT_TYPE |		     \
1063 				PERF_SAMPLE_TRANSACTION |		     \
1064 				PERF_SAMPLE_DATA_PAGE_SIZE)
1065 
1066 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1067 {
1068 	struct perf_event_attr *attr = &event->attr;
1069 	u64 sample_type = attr->sample_type;
1070 	u64 pebs_data_cfg = 0;
1071 	bool gprs, tsx_weight;
1072 
1073 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1074 	    attr->precise_ip > 1)
1075 		return pebs_data_cfg;
1076 
1077 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1078 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1079 
1080 	/*
1081 	 * We need GPRs when:
1082 	 * + user requested them
1083 	 * + precise_ip < 2 for the non event IP
1084 	 * + For RTM TSX weight we need GPRs for the abort code.
1085 	 */
1086 	gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
1087 	       (attr->sample_regs_intr & PEBS_GP_REGS);
1088 
1089 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1090 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1091 		      x86_pmu.rtm_abort_event);
1092 
1093 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
1094 		pebs_data_cfg |= PEBS_DATACFG_GP;
1095 
1096 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1097 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1098 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
1099 
1100 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1101 		/*
1102 		 * For now always log all LBRs. Could configure this
1103 		 * later.
1104 		 */
1105 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1106 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1107 	}
1108 
1109 	return pebs_data_cfg;
1110 }
1111 
1112 static void
1113 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1114 		  struct perf_event *event, bool add)
1115 {
1116 	struct pmu *pmu = event->ctx->pmu;
1117 	/*
1118 	 * Make sure we get updated with the first PEBS
1119 	 * event. It will trigger also during removal, but
1120 	 * that does not hurt:
1121 	 */
1122 	bool update = cpuc->n_pebs == 1;
1123 
1124 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1125 		if (!needed_cb)
1126 			perf_sched_cb_inc(pmu);
1127 		else
1128 			perf_sched_cb_dec(pmu);
1129 
1130 		update = true;
1131 	}
1132 
1133 	/*
1134 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1135 	 * iterating all remaining PEBS events to reconstruct the config.
1136 	 */
1137 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1138 		u64 pebs_data_cfg;
1139 
1140 		/* Clear pebs_data_cfg and pebs_record_size for first PEBS. */
1141 		if (cpuc->n_pebs == 1) {
1142 			cpuc->pebs_data_cfg = 0;
1143 			cpuc->pebs_record_size = sizeof(struct pebs_basic);
1144 		}
1145 
1146 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1147 
1148 		/* Update pebs_record_size if new event requires more data. */
1149 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg) {
1150 			cpuc->pebs_data_cfg |= pebs_data_cfg;
1151 			adaptive_pebs_record_size_update();
1152 			update = true;
1153 		}
1154 	}
1155 
1156 	if (update)
1157 		pebs_update_threshold(cpuc);
1158 }
1159 
1160 void intel_pmu_pebs_add(struct perf_event *event)
1161 {
1162 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1163 	struct hw_perf_event *hwc = &event->hw;
1164 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1165 
1166 	cpuc->n_pebs++;
1167 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1168 		cpuc->n_large_pebs++;
1169 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1170 		cpuc->n_pebs_via_pt++;
1171 
1172 	pebs_update_state(needed_cb, cpuc, event, true);
1173 }
1174 
1175 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1176 {
1177 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1178 
1179 	if (!is_pebs_pt(event))
1180 		return;
1181 
1182 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1183 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1184 }
1185 
1186 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1187 {
1188 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1189 	struct hw_perf_event *hwc = &event->hw;
1190 	struct debug_store *ds = cpuc->ds;
1191 	u64 value = ds->pebs_event_reset[hwc->idx];
1192 	u32 base = MSR_RELOAD_PMC0;
1193 	unsigned int idx = hwc->idx;
1194 
1195 	if (!is_pebs_pt(event))
1196 		return;
1197 
1198 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1199 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1200 
1201 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1202 
1203 	if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1204 		base = MSR_RELOAD_FIXED_CTR0;
1205 		idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1206 		if (x86_pmu.intel_cap.pebs_format < 5)
1207 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
1208 		else
1209 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
1210 	}
1211 	wrmsrl(base + idx, value);
1212 }
1213 
1214 void intel_pmu_pebs_enable(struct perf_event *event)
1215 {
1216 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1217 	struct hw_perf_event *hwc = &event->hw;
1218 	struct debug_store *ds = cpuc->ds;
1219 	unsigned int idx = hwc->idx;
1220 
1221 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1222 
1223 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1224 
1225 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1226 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1227 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1228 		cpuc->pebs_enabled |= 1ULL << 63;
1229 
1230 	if (x86_pmu.intel_cap.pebs_baseline) {
1231 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1232 		if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1233 			wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg);
1234 			cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg;
1235 		}
1236 	}
1237 
1238 	if (idx >= INTEL_PMC_IDX_FIXED) {
1239 		if (x86_pmu.intel_cap.pebs_format < 5)
1240 			idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
1241 		else
1242 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1243 	}
1244 
1245 	/*
1246 	 * Use auto-reload if possible to save a MSR write in the PMI.
1247 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1248 	 */
1249 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1250 		ds->pebs_event_reset[idx] =
1251 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1252 	} else {
1253 		ds->pebs_event_reset[idx] = 0;
1254 	}
1255 
1256 	intel_pmu_pebs_via_pt_enable(event);
1257 }
1258 
1259 void intel_pmu_pebs_del(struct perf_event *event)
1260 {
1261 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1262 	struct hw_perf_event *hwc = &event->hw;
1263 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1264 
1265 	cpuc->n_pebs--;
1266 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1267 		cpuc->n_large_pebs--;
1268 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1269 		cpuc->n_pebs_via_pt--;
1270 
1271 	pebs_update_state(needed_cb, cpuc, event, false);
1272 }
1273 
1274 void intel_pmu_pebs_disable(struct perf_event *event)
1275 {
1276 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1277 	struct hw_perf_event *hwc = &event->hw;
1278 
1279 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1280 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1281 		intel_pmu_drain_pebs_buffer();
1282 
1283 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1284 
1285 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1286 	    (x86_pmu.version < 5))
1287 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1288 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1289 		cpuc->pebs_enabled &= ~(1ULL << 63);
1290 
1291 	intel_pmu_pebs_via_pt_disable(event);
1292 
1293 	if (cpuc->enabled)
1294 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1295 
1296 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1297 }
1298 
1299 void intel_pmu_pebs_enable_all(void)
1300 {
1301 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1302 
1303 	if (cpuc->pebs_enabled)
1304 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1305 }
1306 
1307 void intel_pmu_pebs_disable_all(void)
1308 {
1309 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1310 
1311 	if (cpuc->pebs_enabled)
1312 		__intel_pmu_pebs_disable_all();
1313 }
1314 
1315 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1316 {
1317 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1318 	unsigned long from = cpuc->lbr_entries[0].from;
1319 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1320 	unsigned long ip = regs->ip;
1321 	int is_64bit = 0;
1322 	void *kaddr;
1323 	int size;
1324 
1325 	/*
1326 	 * We don't need to fixup if the PEBS assist is fault like
1327 	 */
1328 	if (!x86_pmu.intel_cap.pebs_trap)
1329 		return 1;
1330 
1331 	/*
1332 	 * No LBR entry, no basic block, no rewinding
1333 	 */
1334 	if (!cpuc->lbr_stack.nr || !from || !to)
1335 		return 0;
1336 
1337 	/*
1338 	 * Basic blocks should never cross user/kernel boundaries
1339 	 */
1340 	if (kernel_ip(ip) != kernel_ip(to))
1341 		return 0;
1342 
1343 	/*
1344 	 * unsigned math, either ip is before the start (impossible) or
1345 	 * the basic block is larger than 1 page (sanity)
1346 	 */
1347 	if ((ip - to) > PEBS_FIXUP_SIZE)
1348 		return 0;
1349 
1350 	/*
1351 	 * We sampled a branch insn, rewind using the LBR stack
1352 	 */
1353 	if (ip == to) {
1354 		set_linear_ip(regs, from);
1355 		return 1;
1356 	}
1357 
1358 	size = ip - to;
1359 	if (!kernel_ip(ip)) {
1360 		int bytes;
1361 		u8 *buf = this_cpu_read(insn_buffer);
1362 
1363 		/* 'size' must fit our buffer, see above */
1364 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1365 		if (bytes != 0)
1366 			return 0;
1367 
1368 		kaddr = buf;
1369 	} else {
1370 		kaddr = (void *)to;
1371 	}
1372 
1373 	do {
1374 		struct insn insn;
1375 
1376 		old_to = to;
1377 
1378 #ifdef CONFIG_X86_64
1379 		is_64bit = kernel_ip(to) || any_64bit_mode(regs);
1380 #endif
1381 		insn_init(&insn, kaddr, size, is_64bit);
1382 
1383 		/*
1384 		 * Make sure there was not a problem decoding the instruction.
1385 		 * This is doubly important because we have an infinite loop if
1386 		 * insn.length=0.
1387 		 */
1388 		if (insn_get_length(&insn))
1389 			break;
1390 
1391 		to += insn.length;
1392 		kaddr += insn.length;
1393 		size -= insn.length;
1394 	} while (to < ip);
1395 
1396 	if (to == ip) {
1397 		set_linear_ip(regs, old_to);
1398 		return 1;
1399 	}
1400 
1401 	/*
1402 	 * Even though we decoded the basic block, the instruction stream
1403 	 * never matched the given IP, either the TO or the IP got corrupted.
1404 	 */
1405 	return 0;
1406 }
1407 
1408 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1409 {
1410 	if (tsx_tuning) {
1411 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1412 		return tsx.cycles_last_block;
1413 	}
1414 	return 0;
1415 }
1416 
1417 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1418 {
1419 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1420 
1421 	/* For RTM XABORTs also log the abort code from AX */
1422 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1423 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1424 	return txn;
1425 }
1426 
1427 static inline u64 get_pebs_status(void *n)
1428 {
1429 	if (x86_pmu.intel_cap.pebs_format < 4)
1430 		return ((struct pebs_record_nhm *)n)->status;
1431 	return ((struct pebs_basic *)n)->applicable_counters;
1432 }
1433 
1434 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1435 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1436 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1437 		 PERF_X86_EVENT_PEBS_NA_HSW)
1438 
1439 static u64 get_data_src(struct perf_event *event, u64 aux)
1440 {
1441 	u64 val = PERF_MEM_NA;
1442 	int fl = event->hw.flags;
1443 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1444 
1445 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1446 		val = load_latency_data(aux);
1447 	else if (fl & PERF_X86_EVENT_PEBS_STLAT)
1448 		val = store_latency_data(aux);
1449 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1450 		val = precise_datala_hsw(event, aux);
1451 	else if (fst)
1452 		val = precise_store_data(aux);
1453 	return val;
1454 }
1455 
1456 #define PERF_SAMPLE_ADDR_TYPE	(PERF_SAMPLE_ADDR |		\
1457 				 PERF_SAMPLE_PHYS_ADDR |	\
1458 				 PERF_SAMPLE_DATA_PAGE_SIZE)
1459 
1460 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1461 				   struct pt_regs *iregs, void *__pebs,
1462 				   struct perf_sample_data *data,
1463 				   struct pt_regs *regs)
1464 {
1465 	/*
1466 	 * We cast to the biggest pebs_record but are careful not to
1467 	 * unconditionally access the 'extra' entries.
1468 	 */
1469 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1470 	struct pebs_record_skl *pebs = __pebs;
1471 	u64 sample_type;
1472 	int fll;
1473 
1474 	if (pebs == NULL)
1475 		return;
1476 
1477 	sample_type = event->attr.sample_type;
1478 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1479 
1480 	perf_sample_data_init(data, 0, event->hw.last_period);
1481 
1482 	data->period = event->hw.last_period;
1483 
1484 	/*
1485 	 * Use latency for weight (only avail with PEBS-LL)
1486 	 */
1487 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE))
1488 		data->weight.full = pebs->lat;
1489 
1490 	/*
1491 	 * data.data_src encodes the data source
1492 	 */
1493 	if (sample_type & PERF_SAMPLE_DATA_SRC)
1494 		data->data_src.val = get_data_src(event, pebs->dse);
1495 
1496 	/*
1497 	 * We must however always use iregs for the unwinder to stay sane; the
1498 	 * record BP,SP,IP can point into thin air when the record is from a
1499 	 * previous PMI context or an (I)RET happened between the record and
1500 	 * PMI.
1501 	 */
1502 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1503 		data->callchain = perf_callchain(event, iregs);
1504 
1505 	/*
1506 	 * We use the interrupt regs as a base because the PEBS record does not
1507 	 * contain a full regs set, specifically it seems to lack segment
1508 	 * descriptors, which get used by things like user_mode().
1509 	 *
1510 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1511 	 */
1512 	*regs = *iregs;
1513 
1514 	/*
1515 	 * Initialize regs_>flags from PEBS,
1516 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1517 	 * i.e., do not rely on it being zero:
1518 	 */
1519 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1520 
1521 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1522 		regs->ax = pebs->ax;
1523 		regs->bx = pebs->bx;
1524 		regs->cx = pebs->cx;
1525 		regs->dx = pebs->dx;
1526 		regs->si = pebs->si;
1527 		regs->di = pebs->di;
1528 
1529 		regs->bp = pebs->bp;
1530 		regs->sp = pebs->sp;
1531 
1532 #ifndef CONFIG_X86_32
1533 		regs->r8 = pebs->r8;
1534 		regs->r9 = pebs->r9;
1535 		regs->r10 = pebs->r10;
1536 		regs->r11 = pebs->r11;
1537 		regs->r12 = pebs->r12;
1538 		regs->r13 = pebs->r13;
1539 		regs->r14 = pebs->r14;
1540 		regs->r15 = pebs->r15;
1541 #endif
1542 	}
1543 
1544 	if (event->attr.precise_ip > 1) {
1545 		/*
1546 		 * Haswell and later processors have an 'eventing IP'
1547 		 * (real IP) which fixes the off-by-1 skid in hardware.
1548 		 * Use it when precise_ip >= 2 :
1549 		 */
1550 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1551 			set_linear_ip(regs, pebs->real_ip);
1552 			regs->flags |= PERF_EFLAGS_EXACT;
1553 		} else {
1554 			/* Otherwise, use PEBS off-by-1 IP: */
1555 			set_linear_ip(regs, pebs->ip);
1556 
1557 			/*
1558 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
1559 			 * using the LBR. If successful, the fixup function
1560 			 * corrects regs->ip and calls set_linear_ip() on regs:
1561 			 */
1562 			if (intel_pmu_pebs_fixup_ip(regs))
1563 				regs->flags |= PERF_EFLAGS_EXACT;
1564 		}
1565 	} else {
1566 		/*
1567 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
1568 		 * no fixup attempted:
1569 		 */
1570 		set_linear_ip(regs, pebs->ip);
1571 	}
1572 
1573 
1574 	if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
1575 	    x86_pmu.intel_cap.pebs_format >= 1)
1576 		data->addr = pebs->dla;
1577 
1578 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1579 		/* Only set the TSX weight when no memory weight. */
1580 		if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll)
1581 			data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
1582 
1583 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1584 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1585 							      pebs->ax);
1586 	}
1587 
1588 	/*
1589 	 * v3 supplies an accurate time stamp, so we use that
1590 	 * for the time stamp.
1591 	 *
1592 	 * We can only do this for the default trace clock.
1593 	 */
1594 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1595 		event->attr.use_clockid == 0)
1596 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1597 
1598 	if (has_branch_stack(event))
1599 		data->br_stack = &cpuc->lbr_stack;
1600 }
1601 
1602 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1603 				    struct pebs_gprs *gprs)
1604 {
1605 	regs->ax = gprs->ax;
1606 	regs->bx = gprs->bx;
1607 	regs->cx = gprs->cx;
1608 	regs->dx = gprs->dx;
1609 	regs->si = gprs->si;
1610 	regs->di = gprs->di;
1611 	regs->bp = gprs->bp;
1612 	regs->sp = gprs->sp;
1613 #ifndef CONFIG_X86_32
1614 	regs->r8 = gprs->r8;
1615 	regs->r9 = gprs->r9;
1616 	regs->r10 = gprs->r10;
1617 	regs->r11 = gprs->r11;
1618 	regs->r12 = gprs->r12;
1619 	regs->r13 = gprs->r13;
1620 	regs->r14 = gprs->r14;
1621 	regs->r15 = gprs->r15;
1622 #endif
1623 }
1624 
1625 #define PEBS_LATENCY_MASK			0xffff
1626 #define PEBS_CACHE_LATENCY_OFFSET		32
1627 
1628 /*
1629  * With adaptive PEBS the layout depends on what fields are configured.
1630  */
1631 
1632 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1633 					    struct pt_regs *iregs, void *__pebs,
1634 					    struct perf_sample_data *data,
1635 					    struct pt_regs *regs)
1636 {
1637 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1638 	struct pebs_basic *basic = __pebs;
1639 	void *next_record = basic + 1;
1640 	u64 sample_type;
1641 	u64 format_size;
1642 	struct pebs_meminfo *meminfo = NULL;
1643 	struct pebs_gprs *gprs = NULL;
1644 	struct x86_perf_regs *perf_regs;
1645 
1646 	if (basic == NULL)
1647 		return;
1648 
1649 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
1650 	perf_regs->xmm_regs = NULL;
1651 
1652 	sample_type = event->attr.sample_type;
1653 	format_size = basic->format_size;
1654 	perf_sample_data_init(data, 0, event->hw.last_period);
1655 	data->period = event->hw.last_period;
1656 
1657 	if (event->attr.use_clockid == 0)
1658 		data->time = native_sched_clock_from_tsc(basic->tsc);
1659 
1660 	/*
1661 	 * We must however always use iregs for the unwinder to stay sane; the
1662 	 * record BP,SP,IP can point into thin air when the record is from a
1663 	 * previous PMI context or an (I)RET happened between the record and
1664 	 * PMI.
1665 	 */
1666 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1667 		data->callchain = perf_callchain(event, iregs);
1668 
1669 	*regs = *iregs;
1670 	/* The ip in basic is EventingIP */
1671 	set_linear_ip(regs, basic->ip);
1672 	regs->flags = PERF_EFLAGS_EXACT;
1673 
1674 	/*
1675 	 * The record for MEMINFO is in front of GP
1676 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1677 	 * Save the pointer here but process later.
1678 	 */
1679 	if (format_size & PEBS_DATACFG_MEMINFO) {
1680 		meminfo = next_record;
1681 		next_record = meminfo + 1;
1682 	}
1683 
1684 	if (format_size & PEBS_DATACFG_GP) {
1685 		gprs = next_record;
1686 		next_record = gprs + 1;
1687 
1688 		if (event->attr.precise_ip < 2) {
1689 			set_linear_ip(regs, gprs->ip);
1690 			regs->flags &= ~PERF_EFLAGS_EXACT;
1691 		}
1692 
1693 		if (sample_type & PERF_SAMPLE_REGS_INTR)
1694 			adaptive_pebs_save_regs(regs, gprs);
1695 	}
1696 
1697 	if (format_size & PEBS_DATACFG_MEMINFO) {
1698 		if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
1699 			u64 weight = meminfo->latency;
1700 
1701 			if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) {
1702 				data->weight.var2_w = weight & PEBS_LATENCY_MASK;
1703 				weight >>= PEBS_CACHE_LATENCY_OFFSET;
1704 			}
1705 
1706 			/*
1707 			 * Although meminfo::latency is defined as a u64,
1708 			 * only the lower 32 bits include the valid data
1709 			 * in practice on Ice Lake and earlier platforms.
1710 			 */
1711 			if (sample_type & PERF_SAMPLE_WEIGHT) {
1712 				data->weight.full = weight ?:
1713 					intel_get_tsx_weight(meminfo->tsx_tuning);
1714 			} else {
1715 				data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?:
1716 					intel_get_tsx_weight(meminfo->tsx_tuning);
1717 			}
1718 		}
1719 
1720 		if (sample_type & PERF_SAMPLE_DATA_SRC)
1721 			data->data_src.val = get_data_src(event, meminfo->aux);
1722 
1723 		if (sample_type & PERF_SAMPLE_ADDR_TYPE)
1724 			data->addr = meminfo->address;
1725 
1726 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1727 			data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1728 							  gprs ? gprs->ax : 0);
1729 	}
1730 
1731 	if (format_size & PEBS_DATACFG_XMMS) {
1732 		struct pebs_xmm *xmm = next_record;
1733 
1734 		next_record = xmm + 1;
1735 		perf_regs->xmm_regs = xmm->xmm;
1736 	}
1737 
1738 	if (format_size & PEBS_DATACFG_LBRS) {
1739 		struct lbr_entry *lbr = next_record;
1740 		int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1741 					& 0xff) + 1;
1742 		next_record = next_record + num_lbr * sizeof(struct lbr_entry);
1743 
1744 		if (has_branch_stack(event)) {
1745 			intel_pmu_store_pebs_lbrs(lbr);
1746 			data->br_stack = &cpuc->lbr_stack;
1747 		}
1748 	}
1749 
1750 	WARN_ONCE(next_record != __pebs + (format_size >> 48),
1751 			"PEBS record size %llu, expected %llu, config %llx\n",
1752 			format_size >> 48,
1753 			(u64)(next_record - __pebs),
1754 			basic->format_size);
1755 }
1756 
1757 static inline void *
1758 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1759 {
1760 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1761 	void *at;
1762 	u64 pebs_status;
1763 
1764 	/*
1765 	 * fmt0 does not have a status bitfield (does not use
1766 	 * perf_record_nhm format)
1767 	 */
1768 	if (x86_pmu.intel_cap.pebs_format < 1)
1769 		return base;
1770 
1771 	if (base == NULL)
1772 		return NULL;
1773 
1774 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1775 		unsigned long status = get_pebs_status(at);
1776 
1777 		if (test_bit(bit, (unsigned long *)&status)) {
1778 			/* PEBS v3 has accurate status bits */
1779 			if (x86_pmu.intel_cap.pebs_format >= 3)
1780 				return at;
1781 
1782 			if (status == (1 << bit))
1783 				return at;
1784 
1785 			/* clear non-PEBS bit and re-check */
1786 			pebs_status = status & cpuc->pebs_enabled;
1787 			pebs_status &= PEBS_COUNTER_MASK;
1788 			if (pebs_status == (1 << bit))
1789 				return at;
1790 		}
1791 	}
1792 	return NULL;
1793 }
1794 
1795 void intel_pmu_auto_reload_read(struct perf_event *event)
1796 {
1797 	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1798 
1799 	perf_pmu_disable(event->pmu);
1800 	intel_pmu_drain_pebs_buffer();
1801 	perf_pmu_enable(event->pmu);
1802 }
1803 
1804 /*
1805  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1806  */
1807 static int
1808 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1809 {
1810 	struct hw_perf_event *hwc = &event->hw;
1811 	int shift = 64 - x86_pmu.cntval_bits;
1812 	u64 period = hwc->sample_period;
1813 	u64 prev_raw_count, new_raw_count;
1814 	s64 new, old;
1815 
1816 	WARN_ON(!period);
1817 
1818 	/*
1819 	 * drain_pebs() only happens when the PMU is disabled.
1820 	 */
1821 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1822 
1823 	prev_raw_count = local64_read(&hwc->prev_count);
1824 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1825 	local64_set(&hwc->prev_count, new_raw_count);
1826 
1827 	/*
1828 	 * Since the counter increments a negative counter value and
1829 	 * overflows on the sign switch, giving the interval:
1830 	 *
1831 	 *   [-period, 0]
1832 	 *
1833 	 * the difference between two consecutive reads is:
1834 	 *
1835 	 *   A) value2 - value1;
1836 	 *      when no overflows have happened in between,
1837 	 *
1838 	 *   B) (0 - value1) + (value2 - (-period));
1839 	 *      when one overflow happened in between,
1840 	 *
1841 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1842 	 *      when @n overflows happened in between.
1843 	 *
1844 	 * Here A) is the obvious difference, B) is the extension to the
1845 	 * discrete interval, where the first term is to the top of the
1846 	 * interval and the second term is from the bottom of the next
1847 	 * interval and C) the extension to multiple intervals, where the
1848 	 * middle term is the whole intervals covered.
1849 	 *
1850 	 * An equivalent of C, by reduction, is:
1851 	 *
1852 	 *   value2 - value1 + n * period
1853 	 */
1854 	new = ((s64)(new_raw_count << shift) >> shift);
1855 	old = ((s64)(prev_raw_count << shift) >> shift);
1856 	local64_add(new - old + count * period, &event->count);
1857 
1858 	local64_set(&hwc->period_left, -new);
1859 
1860 	perf_event_update_userpage(event);
1861 
1862 	return 0;
1863 }
1864 
1865 static __always_inline void
1866 __intel_pmu_pebs_event(struct perf_event *event,
1867 		       struct pt_regs *iregs,
1868 		       struct perf_sample_data *data,
1869 		       void *base, void *top,
1870 		       int bit, int count,
1871 		       void (*setup_sample)(struct perf_event *,
1872 					    struct pt_regs *,
1873 					    void *,
1874 					    struct perf_sample_data *,
1875 					    struct pt_regs *))
1876 {
1877 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1878 	struct hw_perf_event *hwc = &event->hw;
1879 	struct x86_perf_regs perf_regs;
1880 	struct pt_regs *regs = &perf_regs.regs;
1881 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1882 	static struct pt_regs dummy_iregs;
1883 
1884 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1885 		/*
1886 		 * Now, auto-reload is only enabled in fixed period mode.
1887 		 * The reload value is always hwc->sample_period.
1888 		 * May need to change it, if auto-reload is enabled in
1889 		 * freq mode later.
1890 		 */
1891 		intel_pmu_save_and_restart_reload(event, count);
1892 	} else if (!intel_pmu_save_and_restart(event))
1893 		return;
1894 
1895 	if (!iregs)
1896 		iregs = &dummy_iregs;
1897 
1898 	while (count > 1) {
1899 		setup_sample(event, iregs, at, data, regs);
1900 		perf_event_output(event, data, regs);
1901 		at += cpuc->pebs_record_size;
1902 		at = get_next_pebs_record_by_bit(at, top, bit);
1903 		count--;
1904 	}
1905 
1906 	setup_sample(event, iregs, at, data, regs);
1907 	if (iregs == &dummy_iregs) {
1908 		/*
1909 		 * The PEBS records may be drained in the non-overflow context,
1910 		 * e.g., large PEBS + context switch. Perf should treat the
1911 		 * last record the same as other PEBS records, and doesn't
1912 		 * invoke the generic overflow handler.
1913 		 */
1914 		perf_event_output(event, data, regs);
1915 	} else {
1916 		/*
1917 		 * All but the last records are processed.
1918 		 * The last one is left to be able to call the overflow handler.
1919 		 */
1920 		if (perf_event_overflow(event, data, regs))
1921 			x86_pmu_stop(event, 0);
1922 	}
1923 }
1924 
1925 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
1926 {
1927 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1928 	struct debug_store *ds = cpuc->ds;
1929 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1930 	struct pebs_record_core *at, *top;
1931 	int n;
1932 
1933 	if (!x86_pmu.pebs_active)
1934 		return;
1935 
1936 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1937 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1938 
1939 	/*
1940 	 * Whatever else happens, drain the thing
1941 	 */
1942 	ds->pebs_index = ds->pebs_buffer_base;
1943 
1944 	if (!test_bit(0, cpuc->active_mask))
1945 		return;
1946 
1947 	WARN_ON_ONCE(!event);
1948 
1949 	if (!event->attr.precise_ip)
1950 		return;
1951 
1952 	n = top - at;
1953 	if (n <= 0) {
1954 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1955 			intel_pmu_save_and_restart_reload(event, 0);
1956 		return;
1957 	}
1958 
1959 	__intel_pmu_pebs_event(event, iregs, data, at, top, 0, n,
1960 			       setup_pebs_fixed_sample_data);
1961 }
1962 
1963 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
1964 {
1965 	struct perf_event *event;
1966 	int bit;
1967 
1968 	/*
1969 	 * The drain_pebs() could be called twice in a short period
1970 	 * for auto-reload event in pmu::read(). There are no
1971 	 * overflows have happened in between.
1972 	 * It needs to call intel_pmu_save_and_restart_reload() to
1973 	 * update the event->count for this case.
1974 	 */
1975 	for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
1976 		event = cpuc->events[bit];
1977 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1978 			intel_pmu_save_and_restart_reload(event, 0);
1979 	}
1980 }
1981 
1982 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
1983 {
1984 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1985 	struct debug_store *ds = cpuc->ds;
1986 	struct perf_event *event;
1987 	void *base, *at, *top;
1988 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1989 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
1990 	int bit, i, size;
1991 	u64 mask;
1992 
1993 	if (!x86_pmu.pebs_active)
1994 		return;
1995 
1996 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1997 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1998 
1999 	ds->pebs_index = ds->pebs_buffer_base;
2000 
2001 	mask = (1ULL << x86_pmu.max_pebs_events) - 1;
2002 	size = x86_pmu.max_pebs_events;
2003 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
2004 		mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
2005 		size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
2006 	}
2007 
2008 	if (unlikely(base >= top)) {
2009 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2010 		return;
2011 	}
2012 
2013 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
2014 		struct pebs_record_nhm *p = at;
2015 		u64 pebs_status;
2016 
2017 		pebs_status = p->status & cpuc->pebs_enabled;
2018 		pebs_status &= mask;
2019 
2020 		/* PEBS v3 has more accurate status bits */
2021 		if (x86_pmu.intel_cap.pebs_format >= 3) {
2022 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2023 				counts[bit]++;
2024 
2025 			continue;
2026 		}
2027 
2028 		/*
2029 		 * On some CPUs the PEBS status can be zero when PEBS is
2030 		 * racing with clearing of GLOBAL_STATUS.
2031 		 *
2032 		 * Normally we would drop that record, but in the
2033 		 * case when there is only a single active PEBS event
2034 		 * we can assume it's for that event.
2035 		 */
2036 		if (!pebs_status && cpuc->pebs_enabled &&
2037 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
2038 			pebs_status = p->status = cpuc->pebs_enabled;
2039 
2040 		bit = find_first_bit((unsigned long *)&pebs_status,
2041 					x86_pmu.max_pebs_events);
2042 		if (bit >= x86_pmu.max_pebs_events)
2043 			continue;
2044 
2045 		/*
2046 		 * The PEBS hardware does not deal well with the situation
2047 		 * when events happen near to each other and multiple bits
2048 		 * are set. But it should happen rarely.
2049 		 *
2050 		 * If these events include one PEBS and multiple non-PEBS
2051 		 * events, it doesn't impact PEBS record. The record will
2052 		 * be handled normally. (slow path)
2053 		 *
2054 		 * If these events include two or more PEBS events, the
2055 		 * records for the events can be collapsed into a single
2056 		 * one, and it's not possible to reconstruct all events
2057 		 * that caused the PEBS record. It's called collision.
2058 		 * If collision happened, the record will be dropped.
2059 		 */
2060 		if (pebs_status != (1ULL << bit)) {
2061 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
2062 				error[i]++;
2063 			continue;
2064 		}
2065 
2066 		counts[bit]++;
2067 	}
2068 
2069 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2070 		if ((counts[bit] == 0) && (error[bit] == 0))
2071 			continue;
2072 
2073 		event = cpuc->events[bit];
2074 		if (WARN_ON_ONCE(!event))
2075 			continue;
2076 
2077 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2078 			continue;
2079 
2080 		/* log dropped samples number */
2081 		if (error[bit]) {
2082 			perf_log_lost_samples(event, error[bit]);
2083 
2084 			if (iregs && perf_event_account_interrupt(event))
2085 				x86_pmu_stop(event, 0);
2086 		}
2087 
2088 		if (counts[bit]) {
2089 			__intel_pmu_pebs_event(event, iregs, data, base,
2090 					       top, bit, counts[bit],
2091 					       setup_pebs_fixed_sample_data);
2092 		}
2093 	}
2094 }
2095 
2096 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
2097 {
2098 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2099 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2100 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
2101 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
2102 	struct debug_store *ds = cpuc->ds;
2103 	struct perf_event *event;
2104 	void *base, *at, *top;
2105 	int bit, size;
2106 	u64 mask;
2107 
2108 	if (!x86_pmu.pebs_active)
2109 		return;
2110 
2111 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
2112 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
2113 
2114 	ds->pebs_index = ds->pebs_buffer_base;
2115 
2116 	mask = ((1ULL << max_pebs_events) - 1) |
2117 	       (((1ULL << num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
2118 	size = INTEL_PMC_IDX_FIXED + num_counters_fixed;
2119 
2120 	if (unlikely(base >= top)) {
2121 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2122 		return;
2123 	}
2124 
2125 	for (at = base; at < top; at += cpuc->pebs_record_size) {
2126 		u64 pebs_status;
2127 
2128 		pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
2129 		pebs_status &= mask;
2130 
2131 		for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2132 			counts[bit]++;
2133 	}
2134 
2135 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2136 		if (counts[bit] == 0)
2137 			continue;
2138 
2139 		event = cpuc->events[bit];
2140 		if (WARN_ON_ONCE(!event))
2141 			continue;
2142 
2143 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2144 			continue;
2145 
2146 		__intel_pmu_pebs_event(event, iregs, data, base,
2147 				       top, bit, counts[bit],
2148 				       setup_pebs_adaptive_sample_data);
2149 	}
2150 }
2151 
2152 /*
2153  * BTS, PEBS probe and setup
2154  */
2155 
2156 void __init intel_ds_init(void)
2157 {
2158 	/*
2159 	 * No support for 32bit formats
2160 	 */
2161 	if (!boot_cpu_has(X86_FEATURE_DTES64))
2162 		return;
2163 
2164 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
2165 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
2166 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
2167 	if (x86_pmu.version <= 4)
2168 		x86_pmu.pebs_no_isolation = 1;
2169 
2170 	if (x86_pmu.pebs) {
2171 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
2172 		char *pebs_qual = "";
2173 		int format = x86_pmu.intel_cap.pebs_format;
2174 
2175 		if (format < 4)
2176 			x86_pmu.intel_cap.pebs_baseline = 0;
2177 
2178 		switch (format) {
2179 		case 0:
2180 			pr_cont("PEBS fmt0%c, ", pebs_type);
2181 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
2182 			/*
2183 			 * Using >PAGE_SIZE buffers makes the WRMSR to
2184 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
2185 			 * mysteriously hang on Core2.
2186 			 *
2187 			 * As a workaround, we don't do this.
2188 			 */
2189 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
2190 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
2191 			break;
2192 
2193 		case 1:
2194 			pr_cont("PEBS fmt1%c, ", pebs_type);
2195 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
2196 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2197 			break;
2198 
2199 		case 2:
2200 			pr_cont("PEBS fmt2%c, ", pebs_type);
2201 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2202 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2203 			break;
2204 
2205 		case 3:
2206 			pr_cont("PEBS fmt3%c, ", pebs_type);
2207 			x86_pmu.pebs_record_size =
2208 						sizeof(struct pebs_record_skl);
2209 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2210 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2211 			break;
2212 
2213 		case 4:
2214 		case 5:
2215 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2216 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2217 			if (x86_pmu.intel_cap.pebs_baseline) {
2218 				x86_pmu.large_pebs_flags |=
2219 					PERF_SAMPLE_BRANCH_STACK |
2220 					PERF_SAMPLE_TIME;
2221 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
2222 				pebs_qual = "-baseline";
2223 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
2224 			} else {
2225 				/* Only basic record supported */
2226 				x86_pmu.large_pebs_flags &=
2227 					~(PERF_SAMPLE_ADDR |
2228 					  PERF_SAMPLE_TIME |
2229 					  PERF_SAMPLE_DATA_SRC |
2230 					  PERF_SAMPLE_TRANSACTION |
2231 					  PERF_SAMPLE_REGS_USER |
2232 					  PERF_SAMPLE_REGS_INTR);
2233 			}
2234 			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2235 
2236 			if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) {
2237 				pr_cont("PEBS-via-PT, ");
2238 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
2239 			}
2240 
2241 			break;
2242 
2243 		default:
2244 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2245 			x86_pmu.pebs = 0;
2246 		}
2247 	}
2248 }
2249 
2250 void perf_restore_debug_store(void)
2251 {
2252 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2253 
2254 	if (!x86_pmu.bts && !x86_pmu.pebs)
2255 		return;
2256 
2257 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
2258 }
2259