xref: /openbmc/linux/arch/x86/events/intel/ds.c (revision 77423a62)
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 static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source)
98 {
99 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
100 
101 	data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
102 	data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
103 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
104 	data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
105 	data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
106 }
107 
108 void __init intel_pmu_pebs_data_source_skl(bool pmem)
109 {
110 	__intel_pmu_pebs_data_source_skl(pmem, pebs_data_source);
111 }
112 
113 static void __init intel_pmu_pebs_data_source_grt(u64 *data_source)
114 {
115 	data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
116 	data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
117 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
118 }
119 
120 void __init intel_pmu_pebs_data_source_adl(void)
121 {
122 	u64 *data_source;
123 
124 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
125 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
126 	__intel_pmu_pebs_data_source_skl(false, data_source);
127 
128 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
129 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
130 	intel_pmu_pebs_data_source_grt(data_source);
131 }
132 
133 static u64 precise_store_data(u64 status)
134 {
135 	union intel_x86_pebs_dse dse;
136 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
137 
138 	dse.val = status;
139 
140 	/*
141 	 * bit 4: TLB access
142 	 * 1 = stored missed 2nd level TLB
143 	 *
144 	 * so it either hit the walker or the OS
145 	 * otherwise hit 2nd level TLB
146 	 */
147 	if (dse.st_stlb_miss)
148 		val |= P(TLB, MISS);
149 	else
150 		val |= P(TLB, HIT);
151 
152 	/*
153 	 * bit 0: hit L1 data cache
154 	 * if not set, then all we know is that
155 	 * it missed L1D
156 	 */
157 	if (dse.st_l1d_hit)
158 		val |= P(LVL, HIT);
159 	else
160 		val |= P(LVL, MISS);
161 
162 	/*
163 	 * bit 5: Locked prefix
164 	 */
165 	if (dse.st_locked)
166 		val |= P(LOCK, LOCKED);
167 
168 	return val;
169 }
170 
171 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
172 {
173 	union perf_mem_data_src dse;
174 
175 	dse.val = PERF_MEM_NA;
176 
177 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
178 		dse.mem_op = PERF_MEM_OP_STORE;
179 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
180 		dse.mem_op = PERF_MEM_OP_LOAD;
181 
182 	/*
183 	 * L1 info only valid for following events:
184 	 *
185 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
186 	 * MEM_UOPS_RETIRED.LOCK_STORES
187 	 * MEM_UOPS_RETIRED.SPLIT_STORES
188 	 * MEM_UOPS_RETIRED.ALL_STORES
189 	 */
190 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
191 		if (status & 1)
192 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
193 		else
194 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
195 	}
196 	return dse.val;
197 }
198 
199 static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock)
200 {
201 	/*
202 	 * TLB access
203 	 * 0 = did not miss 2nd level TLB
204 	 * 1 = missed 2nd level TLB
205 	 */
206 	if (tlb)
207 		*val |= P(TLB, MISS) | P(TLB, L2);
208 	else
209 		*val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
210 
211 	/* locked prefix */
212 	if (lock)
213 		*val |= P(LOCK, LOCKED);
214 }
215 
216 /* Retrieve the latency data for e-core of ADL */
217 u64 adl_latency_data_small(struct perf_event *event, u64 status)
218 {
219 	union intel_x86_pebs_dse dse;
220 	u64 val;
221 
222 	WARN_ON_ONCE(hybrid_pmu(event->pmu)->cpu_type == hybrid_big);
223 
224 	dse.val = status;
225 
226 	val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
227 
228 	/*
229 	 * For the atom core on ADL,
230 	 * bit 4: lock, bit 5: TLB access.
231 	 */
232 	pebs_set_tlb_lock(&val, dse.ld_locked, dse.ld_stlb_miss);
233 
234 	if (dse.ld_data_blk)
235 		val |= P(BLK, DATA);
236 	else
237 		val |= P(BLK, NA);
238 
239 	return val;
240 }
241 
242 static u64 load_latency_data(struct perf_event *event, u64 status)
243 {
244 	union intel_x86_pebs_dse dse;
245 	u64 val;
246 
247 	dse.val = status;
248 
249 	/*
250 	 * use the mapping table for bit 0-3
251 	 */
252 	val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
253 
254 	/*
255 	 * Nehalem models do not support TLB, Lock infos
256 	 */
257 	if (x86_pmu.pebs_no_tlb) {
258 		val |= P(TLB, NA) | P(LOCK, NA);
259 		return val;
260 	}
261 
262 	pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked);
263 
264 	/*
265 	 * Ice Lake and earlier models do not support block infos.
266 	 */
267 	if (!x86_pmu.pebs_block) {
268 		val |= P(BLK, NA);
269 		return val;
270 	}
271 	/*
272 	 * bit 6: load was blocked since its data could not be forwarded
273 	 *        from a preceding store
274 	 */
275 	if (dse.ld_data_blk)
276 		val |= P(BLK, DATA);
277 
278 	/*
279 	 * bit 7: load was blocked due to potential address conflict with
280 	 *        a preceding store
281 	 */
282 	if (dse.ld_addr_blk)
283 		val |= P(BLK, ADDR);
284 
285 	if (!dse.ld_data_blk && !dse.ld_addr_blk)
286 		val |= P(BLK, NA);
287 
288 	return val;
289 }
290 
291 static u64 store_latency_data(struct perf_event *event, u64 status)
292 {
293 	union intel_x86_pebs_dse dse;
294 	u64 val;
295 
296 	dse.val = status;
297 
298 	/*
299 	 * use the mapping table for bit 0-3
300 	 */
301 	val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse];
302 
303 	pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked);
304 
305 	val |= P(BLK, NA);
306 
307 	return val;
308 }
309 
310 struct pebs_record_core {
311 	u64 flags, ip;
312 	u64 ax, bx, cx, dx;
313 	u64 si, di, bp, sp;
314 	u64 r8,  r9,  r10, r11;
315 	u64 r12, r13, r14, r15;
316 };
317 
318 struct pebs_record_nhm {
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 };
326 
327 /*
328  * Same as pebs_record_nhm, with two additional fields.
329  */
330 struct pebs_record_hsw {
331 	u64 flags, ip;
332 	u64 ax, bx, cx, dx;
333 	u64 si, di, bp, sp;
334 	u64 r8,  r9,  r10, r11;
335 	u64 r12, r13, r14, r15;
336 	u64 status, dla, dse, lat;
337 	u64 real_ip, tsx_tuning;
338 };
339 
340 union hsw_tsx_tuning {
341 	struct {
342 		u32 cycles_last_block     : 32,
343 		    hle_abort		  : 1,
344 		    rtm_abort		  : 1,
345 		    instruction_abort     : 1,
346 		    non_instruction_abort : 1,
347 		    retry		  : 1,
348 		    data_conflict	  : 1,
349 		    capacity_writes	  : 1,
350 		    capacity_reads	  : 1;
351 	};
352 	u64	    value;
353 };
354 
355 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
356 
357 /* Same as HSW, plus TSC */
358 
359 struct pebs_record_skl {
360 	u64 flags, ip;
361 	u64 ax, bx, cx, dx;
362 	u64 si, di, bp, sp;
363 	u64 r8,  r9,  r10, r11;
364 	u64 r12, r13, r14, r15;
365 	u64 status, dla, dse, lat;
366 	u64 real_ip, tsx_tuning;
367 	u64 tsc;
368 };
369 
370 void init_debug_store_on_cpu(int cpu)
371 {
372 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
373 
374 	if (!ds)
375 		return;
376 
377 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
378 		     (u32)((u64)(unsigned long)ds),
379 		     (u32)((u64)(unsigned long)ds >> 32));
380 }
381 
382 void fini_debug_store_on_cpu(int cpu)
383 {
384 	if (!per_cpu(cpu_hw_events, cpu).ds)
385 		return;
386 
387 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
388 }
389 
390 static DEFINE_PER_CPU(void *, insn_buffer);
391 
392 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
393 {
394 	unsigned long start = (unsigned long)cea;
395 	phys_addr_t pa;
396 	size_t msz = 0;
397 
398 	pa = virt_to_phys(addr);
399 
400 	preempt_disable();
401 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
402 		cea_set_pte(cea, pa, prot);
403 
404 	/*
405 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
406 	 * all TLB entries for it.
407 	 */
408 	flush_tlb_kernel_range(start, start + size);
409 	preempt_enable();
410 }
411 
412 static void ds_clear_cea(void *cea, size_t size)
413 {
414 	unsigned long start = (unsigned long)cea;
415 	size_t msz = 0;
416 
417 	preempt_disable();
418 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
419 		cea_set_pte(cea, 0, PAGE_NONE);
420 
421 	flush_tlb_kernel_range(start, start + size);
422 	preempt_enable();
423 }
424 
425 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
426 {
427 	unsigned int order = get_order(size);
428 	int node = cpu_to_node(cpu);
429 	struct page *page;
430 
431 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
432 	return page ? page_address(page) : NULL;
433 }
434 
435 static void dsfree_pages(const void *buffer, size_t size)
436 {
437 	if (buffer)
438 		free_pages((unsigned long)buffer, get_order(size));
439 }
440 
441 static int alloc_pebs_buffer(int cpu)
442 {
443 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
444 	struct debug_store *ds = hwev->ds;
445 	size_t bsiz = x86_pmu.pebs_buffer_size;
446 	int max, node = cpu_to_node(cpu);
447 	void *buffer, *insn_buff, *cea;
448 
449 	if (!x86_pmu.pebs)
450 		return 0;
451 
452 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
453 	if (unlikely(!buffer))
454 		return -ENOMEM;
455 
456 	/*
457 	 * HSW+ already provides us the eventing ip; no need to allocate this
458 	 * buffer then.
459 	 */
460 	if (x86_pmu.intel_cap.pebs_format < 2) {
461 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
462 		if (!insn_buff) {
463 			dsfree_pages(buffer, bsiz);
464 			return -ENOMEM;
465 		}
466 		per_cpu(insn_buffer, cpu) = insn_buff;
467 	}
468 	hwev->ds_pebs_vaddr = buffer;
469 	/* Update the cpu entry area mapping */
470 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
471 	ds->pebs_buffer_base = (unsigned long) cea;
472 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
473 	ds->pebs_index = ds->pebs_buffer_base;
474 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
475 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
476 	return 0;
477 }
478 
479 static void release_pebs_buffer(int cpu)
480 {
481 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
482 	void *cea;
483 
484 	if (!x86_pmu.pebs)
485 		return;
486 
487 	kfree(per_cpu(insn_buffer, cpu));
488 	per_cpu(insn_buffer, cpu) = NULL;
489 
490 	/* Clear the fixmap */
491 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
492 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
493 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
494 	hwev->ds_pebs_vaddr = NULL;
495 }
496 
497 static int alloc_bts_buffer(int cpu)
498 {
499 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
500 	struct debug_store *ds = hwev->ds;
501 	void *buffer, *cea;
502 	int max;
503 
504 	if (!x86_pmu.bts)
505 		return 0;
506 
507 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
508 	if (unlikely(!buffer)) {
509 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
510 		return -ENOMEM;
511 	}
512 	hwev->ds_bts_vaddr = buffer;
513 	/* Update the fixmap */
514 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
515 	ds->bts_buffer_base = (unsigned long) cea;
516 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
517 	ds->bts_index = ds->bts_buffer_base;
518 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
519 	ds->bts_absolute_maximum = ds->bts_buffer_base +
520 					max * BTS_RECORD_SIZE;
521 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
522 					(max / 16) * BTS_RECORD_SIZE;
523 	return 0;
524 }
525 
526 static void release_bts_buffer(int cpu)
527 {
528 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
529 	void *cea;
530 
531 	if (!x86_pmu.bts)
532 		return;
533 
534 	/* Clear the fixmap */
535 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
536 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
537 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
538 	hwev->ds_bts_vaddr = NULL;
539 }
540 
541 static int alloc_ds_buffer(int cpu)
542 {
543 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
544 
545 	memset(ds, 0, sizeof(*ds));
546 	per_cpu(cpu_hw_events, cpu).ds = ds;
547 	return 0;
548 }
549 
550 static void release_ds_buffer(int cpu)
551 {
552 	per_cpu(cpu_hw_events, cpu).ds = NULL;
553 }
554 
555 void release_ds_buffers(void)
556 {
557 	int cpu;
558 
559 	if (!x86_pmu.bts && !x86_pmu.pebs)
560 		return;
561 
562 	for_each_possible_cpu(cpu)
563 		release_ds_buffer(cpu);
564 
565 	for_each_possible_cpu(cpu) {
566 		/*
567 		 * Again, ignore errors from offline CPUs, they will no longer
568 		 * observe cpu_hw_events.ds and not program the DS_AREA when
569 		 * they come up.
570 		 */
571 		fini_debug_store_on_cpu(cpu);
572 	}
573 
574 	for_each_possible_cpu(cpu) {
575 		release_pebs_buffer(cpu);
576 		release_bts_buffer(cpu);
577 	}
578 }
579 
580 void reserve_ds_buffers(void)
581 {
582 	int bts_err = 0, pebs_err = 0;
583 	int cpu;
584 
585 	x86_pmu.bts_active = 0;
586 	x86_pmu.pebs_active = 0;
587 
588 	if (!x86_pmu.bts && !x86_pmu.pebs)
589 		return;
590 
591 	if (!x86_pmu.bts)
592 		bts_err = 1;
593 
594 	if (!x86_pmu.pebs)
595 		pebs_err = 1;
596 
597 	for_each_possible_cpu(cpu) {
598 		if (alloc_ds_buffer(cpu)) {
599 			bts_err = 1;
600 			pebs_err = 1;
601 		}
602 
603 		if (!bts_err && alloc_bts_buffer(cpu))
604 			bts_err = 1;
605 
606 		if (!pebs_err && alloc_pebs_buffer(cpu))
607 			pebs_err = 1;
608 
609 		if (bts_err && pebs_err)
610 			break;
611 	}
612 
613 	if (bts_err) {
614 		for_each_possible_cpu(cpu)
615 			release_bts_buffer(cpu);
616 	}
617 
618 	if (pebs_err) {
619 		for_each_possible_cpu(cpu)
620 			release_pebs_buffer(cpu);
621 	}
622 
623 	if (bts_err && pebs_err) {
624 		for_each_possible_cpu(cpu)
625 			release_ds_buffer(cpu);
626 	} else {
627 		if (x86_pmu.bts && !bts_err)
628 			x86_pmu.bts_active = 1;
629 
630 		if (x86_pmu.pebs && !pebs_err)
631 			x86_pmu.pebs_active = 1;
632 
633 		for_each_possible_cpu(cpu) {
634 			/*
635 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
636 			 * will get this call through intel_pmu_cpu_starting().
637 			 */
638 			init_debug_store_on_cpu(cpu);
639 		}
640 	}
641 }
642 
643 /*
644  * BTS
645  */
646 
647 struct event_constraint bts_constraint =
648 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
649 
650 void intel_pmu_enable_bts(u64 config)
651 {
652 	unsigned long debugctlmsr;
653 
654 	debugctlmsr = get_debugctlmsr();
655 
656 	debugctlmsr |= DEBUGCTLMSR_TR;
657 	debugctlmsr |= DEBUGCTLMSR_BTS;
658 	if (config & ARCH_PERFMON_EVENTSEL_INT)
659 		debugctlmsr |= DEBUGCTLMSR_BTINT;
660 
661 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
662 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
663 
664 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
665 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
666 
667 	update_debugctlmsr(debugctlmsr);
668 }
669 
670 void intel_pmu_disable_bts(void)
671 {
672 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
673 	unsigned long debugctlmsr;
674 
675 	if (!cpuc->ds)
676 		return;
677 
678 	debugctlmsr = get_debugctlmsr();
679 
680 	debugctlmsr &=
681 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
682 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
683 
684 	update_debugctlmsr(debugctlmsr);
685 }
686 
687 int intel_pmu_drain_bts_buffer(void)
688 {
689 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
690 	struct debug_store *ds = cpuc->ds;
691 	struct bts_record {
692 		u64	from;
693 		u64	to;
694 		u64	flags;
695 	};
696 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
697 	struct bts_record *at, *base, *top;
698 	struct perf_output_handle handle;
699 	struct perf_event_header header;
700 	struct perf_sample_data data;
701 	unsigned long skip = 0;
702 	struct pt_regs regs;
703 
704 	if (!event)
705 		return 0;
706 
707 	if (!x86_pmu.bts_active)
708 		return 0;
709 
710 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
711 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
712 
713 	if (top <= base)
714 		return 0;
715 
716 	memset(&regs, 0, sizeof(regs));
717 
718 	ds->bts_index = ds->bts_buffer_base;
719 
720 	perf_sample_data_init(&data, 0, event->hw.last_period);
721 
722 	/*
723 	 * BTS leaks kernel addresses in branches across the cpl boundary,
724 	 * such as traps or system calls, so unless the user is asking for
725 	 * kernel tracing (and right now it's not possible), we'd need to
726 	 * filter them out. But first we need to count how many of those we
727 	 * have in the current batch. This is an extra O(n) pass, however,
728 	 * it's much faster than the other one especially considering that
729 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
730 	 * alloc_bts_buffer()).
731 	 */
732 	for (at = base; at < top; at++) {
733 		/*
734 		 * Note that right now *this* BTS code only works if
735 		 * attr::exclude_kernel is set, but let's keep this extra
736 		 * check here in case that changes.
737 		 */
738 		if (event->attr.exclude_kernel &&
739 		    (kernel_ip(at->from) || kernel_ip(at->to)))
740 			skip++;
741 	}
742 
743 	/*
744 	 * Prepare a generic sample, i.e. fill in the invariant fields.
745 	 * We will overwrite the from and to address before we output
746 	 * the sample.
747 	 */
748 	rcu_read_lock();
749 	perf_prepare_sample(&header, &data, event, &regs);
750 
751 	if (perf_output_begin(&handle, &data, event,
752 			      header.size * (top - base - skip)))
753 		goto unlock;
754 
755 	for (at = base; at < top; at++) {
756 		/* Filter out any records that contain kernel addresses. */
757 		if (event->attr.exclude_kernel &&
758 		    (kernel_ip(at->from) || kernel_ip(at->to)))
759 			continue;
760 
761 		data.ip		= at->from;
762 		data.addr	= at->to;
763 
764 		perf_output_sample(&handle, &header, &data, event);
765 	}
766 
767 	perf_output_end(&handle);
768 
769 	/* There's new data available. */
770 	event->hw.interrupts++;
771 	event->pending_kill = POLL_IN;
772 unlock:
773 	rcu_read_unlock();
774 	return 1;
775 }
776 
777 static inline void intel_pmu_drain_pebs_buffer(void)
778 {
779 	struct perf_sample_data data;
780 
781 	x86_pmu.drain_pebs(NULL, &data);
782 }
783 
784 /*
785  * PEBS
786  */
787 struct event_constraint intel_core2_pebs_event_constraints[] = {
788 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
789 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
790 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
791 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
792 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
793 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
794 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
795 	EVENT_CONSTRAINT_END
796 };
797 
798 struct event_constraint intel_atom_pebs_event_constraints[] = {
799 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
800 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
801 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
802 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
803 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
804 	/* Allow all events as PEBS with no flags */
805 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
806 	EVENT_CONSTRAINT_END
807 };
808 
809 struct event_constraint intel_slm_pebs_event_constraints[] = {
810 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
811 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
812 	/* Allow all events as PEBS with no flags */
813 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
814 	EVENT_CONSTRAINT_END
815 };
816 
817 struct event_constraint intel_glm_pebs_event_constraints[] = {
818 	/* Allow all events as PEBS with no flags */
819 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
820 	EVENT_CONSTRAINT_END
821 };
822 
823 struct event_constraint intel_grt_pebs_event_constraints[] = {
824 	/* Allow all events as PEBS with no flags */
825 	INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0xf),
826 	INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf),
827 	EVENT_CONSTRAINT_END
828 };
829 
830 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
831 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
832 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
833 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
834 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
835 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
836 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
837 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
838 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
839 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
840 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
841 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
842 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
843 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
844 	EVENT_CONSTRAINT_END
845 };
846 
847 struct event_constraint intel_westmere_pebs_event_constraints[] = {
848 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
849 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
850 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
851 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
852 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
853 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
854 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
855 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
856 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
857 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
858 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
859 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
860 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
861 	EVENT_CONSTRAINT_END
862 };
863 
864 struct event_constraint intel_snb_pebs_event_constraints[] = {
865 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
866 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
867 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
868 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
869 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
870         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
871         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
872         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
873         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
874 	/* Allow all events as PEBS with no flags */
875 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
876 	EVENT_CONSTRAINT_END
877 };
878 
879 struct event_constraint intel_ivb_pebs_event_constraints[] = {
880         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
881         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
882 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
883 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
884 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
885 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
886 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
887 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
888 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
889 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
890 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
891 	/* Allow all events as PEBS with no flags */
892 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
893         EVENT_CONSTRAINT_END
894 };
895 
896 struct event_constraint intel_hsw_pebs_event_constraints[] = {
897 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
898 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
899 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
900 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
901 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
902 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
903 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
904 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
905 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
906 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
907 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
908 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
909 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
910 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
911 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
912 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
913 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
914 	/* Allow all events as PEBS with no flags */
915 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
916 	EVENT_CONSTRAINT_END
917 };
918 
919 struct event_constraint intel_bdw_pebs_event_constraints[] = {
920 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
921 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
922 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
923 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
924 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
925 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
926 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
927 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
928 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
929 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
930 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
931 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
932 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
933 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
934 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
935 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
936 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
937 	/* Allow all events as PEBS with no flags */
938 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
939 	EVENT_CONSTRAINT_END
940 };
941 
942 
943 struct event_constraint intel_skl_pebs_event_constraints[] = {
944 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
945 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
946 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
947 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
948 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
949 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
950 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
951 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
952 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
953 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
954 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
955 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
956 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
957 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
958 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
959 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
960 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
961 	/* Allow all events as PEBS with no flags */
962 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
963 	EVENT_CONSTRAINT_END
964 };
965 
966 struct event_constraint intel_icl_pebs_event_constraints[] = {
967 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL),	/* old INST_RETIRED.PREC_DIST */
968 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
969 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
970 
971 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
972 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),	/* MEM_INST_RETIRED.LOAD */
973 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),	/* MEM_INST_RETIRED.STORE */
974 
975 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
976 
977 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
978 
979 	/*
980 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
981 	 * need the full constraints from the main table.
982 	 */
983 
984 	EVENT_CONSTRAINT_END
985 };
986 
987 struct event_constraint intel_spr_pebs_event_constraints[] = {
988 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
989 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
990 
991 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
992 	INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
993 	INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
994 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x1d0, 0xf),
995 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x2d0, 0xf),
996 
997 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
998 
999 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1000 
1001 	/*
1002 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1003 	 * need the full constraints from the main table.
1004 	 */
1005 
1006 	EVENT_CONSTRAINT_END
1007 };
1008 
1009 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
1010 {
1011 	struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
1012 	struct event_constraint *c;
1013 
1014 	if (!event->attr.precise_ip)
1015 		return NULL;
1016 
1017 	if (pebs_constraints) {
1018 		for_each_event_constraint(c, pebs_constraints) {
1019 			if (constraint_match(c, event->hw.config)) {
1020 				event->hw.flags |= c->flags;
1021 				return c;
1022 			}
1023 		}
1024 	}
1025 
1026 	/*
1027 	 * Extended PEBS support
1028 	 * Makes the PEBS code search the normal constraints.
1029 	 */
1030 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1031 		return NULL;
1032 
1033 	return &emptyconstraint;
1034 }
1035 
1036 /*
1037  * We need the sched_task callback even for per-cpu events when we use
1038  * the large interrupt threshold, such that we can provide PID and TID
1039  * to PEBS samples.
1040  */
1041 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
1042 {
1043 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
1044 		return false;
1045 
1046 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
1047 }
1048 
1049 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
1050 {
1051 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1052 
1053 	if (!sched_in && pebs_needs_sched_cb(cpuc))
1054 		intel_pmu_drain_pebs_buffer();
1055 }
1056 
1057 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1058 {
1059 	struct debug_store *ds = cpuc->ds;
1060 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
1061 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
1062 	u64 threshold;
1063 	int reserved;
1064 
1065 	if (cpuc->n_pebs_via_pt)
1066 		return;
1067 
1068 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1069 		reserved = max_pebs_events + num_counters_fixed;
1070 	else
1071 		reserved = max_pebs_events;
1072 
1073 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
1074 		threshold = ds->pebs_absolute_maximum -
1075 			reserved * cpuc->pebs_record_size;
1076 	} else {
1077 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1078 	}
1079 
1080 	ds->pebs_interrupt_threshold = threshold;
1081 }
1082 
1083 static void adaptive_pebs_record_size_update(void)
1084 {
1085 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1086 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1087 	int sz = sizeof(struct pebs_basic);
1088 
1089 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1090 		sz += sizeof(struct pebs_meminfo);
1091 	if (pebs_data_cfg & PEBS_DATACFG_GP)
1092 		sz += sizeof(struct pebs_gprs);
1093 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1094 		sz += sizeof(struct pebs_xmm);
1095 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1096 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1097 
1098 	cpuc->pebs_record_size = sz;
1099 }
1100 
1101 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
1102 				PERF_SAMPLE_PHYS_ADDR |			     \
1103 				PERF_SAMPLE_WEIGHT_TYPE |		     \
1104 				PERF_SAMPLE_TRANSACTION |		     \
1105 				PERF_SAMPLE_DATA_PAGE_SIZE)
1106 
1107 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1108 {
1109 	struct perf_event_attr *attr = &event->attr;
1110 	u64 sample_type = attr->sample_type;
1111 	u64 pebs_data_cfg = 0;
1112 	bool gprs, tsx_weight;
1113 
1114 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1115 	    attr->precise_ip > 1)
1116 		return pebs_data_cfg;
1117 
1118 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1119 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1120 
1121 	/*
1122 	 * We need GPRs when:
1123 	 * + user requested them
1124 	 * + precise_ip < 2 for the non event IP
1125 	 * + For RTM TSX weight we need GPRs for the abort code.
1126 	 */
1127 	gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
1128 	       (attr->sample_regs_intr & PEBS_GP_REGS);
1129 
1130 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1131 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1132 		      x86_pmu.rtm_abort_event);
1133 
1134 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
1135 		pebs_data_cfg |= PEBS_DATACFG_GP;
1136 
1137 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1138 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1139 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
1140 
1141 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1142 		/*
1143 		 * For now always log all LBRs. Could configure this
1144 		 * later.
1145 		 */
1146 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1147 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1148 	}
1149 
1150 	return pebs_data_cfg;
1151 }
1152 
1153 static void
1154 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1155 		  struct perf_event *event, bool add)
1156 {
1157 	struct pmu *pmu = event->ctx->pmu;
1158 	/*
1159 	 * Make sure we get updated with the first PEBS
1160 	 * event. It will trigger also during removal, but
1161 	 * that does not hurt:
1162 	 */
1163 	bool update = cpuc->n_pebs == 1;
1164 
1165 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1166 		if (!needed_cb)
1167 			perf_sched_cb_inc(pmu);
1168 		else
1169 			perf_sched_cb_dec(pmu);
1170 
1171 		update = true;
1172 	}
1173 
1174 	/*
1175 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1176 	 * iterating all remaining PEBS events to reconstruct the config.
1177 	 */
1178 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1179 		u64 pebs_data_cfg;
1180 
1181 		/* Clear pebs_data_cfg and pebs_record_size for first PEBS. */
1182 		if (cpuc->n_pebs == 1) {
1183 			cpuc->pebs_data_cfg = 0;
1184 			cpuc->pebs_record_size = sizeof(struct pebs_basic);
1185 		}
1186 
1187 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1188 
1189 		/* Update pebs_record_size if new event requires more data. */
1190 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg) {
1191 			cpuc->pebs_data_cfg |= pebs_data_cfg;
1192 			adaptive_pebs_record_size_update();
1193 			update = true;
1194 		}
1195 	}
1196 
1197 	if (update)
1198 		pebs_update_threshold(cpuc);
1199 }
1200 
1201 void intel_pmu_pebs_add(struct perf_event *event)
1202 {
1203 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1204 	struct hw_perf_event *hwc = &event->hw;
1205 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1206 
1207 	cpuc->n_pebs++;
1208 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1209 		cpuc->n_large_pebs++;
1210 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1211 		cpuc->n_pebs_via_pt++;
1212 
1213 	pebs_update_state(needed_cb, cpuc, event, true);
1214 }
1215 
1216 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1217 {
1218 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1219 
1220 	if (!is_pebs_pt(event))
1221 		return;
1222 
1223 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1224 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1225 }
1226 
1227 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1228 {
1229 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1230 	struct hw_perf_event *hwc = &event->hw;
1231 	struct debug_store *ds = cpuc->ds;
1232 	u64 value = ds->pebs_event_reset[hwc->idx];
1233 	u32 base = MSR_RELOAD_PMC0;
1234 	unsigned int idx = hwc->idx;
1235 
1236 	if (!is_pebs_pt(event))
1237 		return;
1238 
1239 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1240 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1241 
1242 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1243 
1244 	if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1245 		base = MSR_RELOAD_FIXED_CTR0;
1246 		idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1247 		if (x86_pmu.intel_cap.pebs_format < 5)
1248 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
1249 		else
1250 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
1251 	}
1252 	wrmsrl(base + idx, value);
1253 }
1254 
1255 void intel_pmu_pebs_enable(struct perf_event *event)
1256 {
1257 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1258 	struct hw_perf_event *hwc = &event->hw;
1259 	struct debug_store *ds = cpuc->ds;
1260 	unsigned int idx = hwc->idx;
1261 
1262 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1263 
1264 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1265 
1266 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1267 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1268 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1269 		cpuc->pebs_enabled |= 1ULL << 63;
1270 
1271 	if (x86_pmu.intel_cap.pebs_baseline) {
1272 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1273 		if (cpuc->pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1274 			wrmsrl(MSR_PEBS_DATA_CFG, cpuc->pebs_data_cfg);
1275 			cpuc->active_pebs_data_cfg = cpuc->pebs_data_cfg;
1276 		}
1277 	}
1278 
1279 	if (idx >= INTEL_PMC_IDX_FIXED) {
1280 		if (x86_pmu.intel_cap.pebs_format < 5)
1281 			idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
1282 		else
1283 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1284 	}
1285 
1286 	/*
1287 	 * Use auto-reload if possible to save a MSR write in the PMI.
1288 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1289 	 */
1290 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1291 		ds->pebs_event_reset[idx] =
1292 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1293 	} else {
1294 		ds->pebs_event_reset[idx] = 0;
1295 	}
1296 
1297 	intel_pmu_pebs_via_pt_enable(event);
1298 }
1299 
1300 void intel_pmu_pebs_del(struct perf_event *event)
1301 {
1302 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1303 	struct hw_perf_event *hwc = &event->hw;
1304 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1305 
1306 	cpuc->n_pebs--;
1307 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1308 		cpuc->n_large_pebs--;
1309 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1310 		cpuc->n_pebs_via_pt--;
1311 
1312 	pebs_update_state(needed_cb, cpuc, event, false);
1313 }
1314 
1315 void intel_pmu_pebs_disable(struct perf_event *event)
1316 {
1317 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1318 	struct hw_perf_event *hwc = &event->hw;
1319 
1320 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1321 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1322 		intel_pmu_drain_pebs_buffer();
1323 
1324 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1325 
1326 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1327 	    (x86_pmu.version < 5))
1328 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1329 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1330 		cpuc->pebs_enabled &= ~(1ULL << 63);
1331 
1332 	intel_pmu_pebs_via_pt_disable(event);
1333 
1334 	if (cpuc->enabled)
1335 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1336 
1337 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1338 }
1339 
1340 void intel_pmu_pebs_enable_all(void)
1341 {
1342 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1343 
1344 	if (cpuc->pebs_enabled)
1345 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1346 }
1347 
1348 void intel_pmu_pebs_disable_all(void)
1349 {
1350 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1351 
1352 	if (cpuc->pebs_enabled)
1353 		__intel_pmu_pebs_disable_all();
1354 }
1355 
1356 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1357 {
1358 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1359 	unsigned long from = cpuc->lbr_entries[0].from;
1360 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1361 	unsigned long ip = regs->ip;
1362 	int is_64bit = 0;
1363 	void *kaddr;
1364 	int size;
1365 
1366 	/*
1367 	 * We don't need to fixup if the PEBS assist is fault like
1368 	 */
1369 	if (!x86_pmu.intel_cap.pebs_trap)
1370 		return 1;
1371 
1372 	/*
1373 	 * No LBR entry, no basic block, no rewinding
1374 	 */
1375 	if (!cpuc->lbr_stack.nr || !from || !to)
1376 		return 0;
1377 
1378 	/*
1379 	 * Basic blocks should never cross user/kernel boundaries
1380 	 */
1381 	if (kernel_ip(ip) != kernel_ip(to))
1382 		return 0;
1383 
1384 	/*
1385 	 * unsigned math, either ip is before the start (impossible) or
1386 	 * the basic block is larger than 1 page (sanity)
1387 	 */
1388 	if ((ip - to) > PEBS_FIXUP_SIZE)
1389 		return 0;
1390 
1391 	/*
1392 	 * We sampled a branch insn, rewind using the LBR stack
1393 	 */
1394 	if (ip == to) {
1395 		set_linear_ip(regs, from);
1396 		return 1;
1397 	}
1398 
1399 	size = ip - to;
1400 	if (!kernel_ip(ip)) {
1401 		int bytes;
1402 		u8 *buf = this_cpu_read(insn_buffer);
1403 
1404 		/* 'size' must fit our buffer, see above */
1405 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1406 		if (bytes != 0)
1407 			return 0;
1408 
1409 		kaddr = buf;
1410 	} else {
1411 		kaddr = (void *)to;
1412 	}
1413 
1414 	do {
1415 		struct insn insn;
1416 
1417 		old_to = to;
1418 
1419 #ifdef CONFIG_X86_64
1420 		is_64bit = kernel_ip(to) || any_64bit_mode(regs);
1421 #endif
1422 		insn_init(&insn, kaddr, size, is_64bit);
1423 
1424 		/*
1425 		 * Make sure there was not a problem decoding the instruction.
1426 		 * This is doubly important because we have an infinite loop if
1427 		 * insn.length=0.
1428 		 */
1429 		if (insn_get_length(&insn))
1430 			break;
1431 
1432 		to += insn.length;
1433 		kaddr += insn.length;
1434 		size -= insn.length;
1435 	} while (to < ip);
1436 
1437 	if (to == ip) {
1438 		set_linear_ip(regs, old_to);
1439 		return 1;
1440 	}
1441 
1442 	/*
1443 	 * Even though we decoded the basic block, the instruction stream
1444 	 * never matched the given IP, either the TO or the IP got corrupted.
1445 	 */
1446 	return 0;
1447 }
1448 
1449 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1450 {
1451 	if (tsx_tuning) {
1452 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1453 		return tsx.cycles_last_block;
1454 	}
1455 	return 0;
1456 }
1457 
1458 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1459 {
1460 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1461 
1462 	/* For RTM XABORTs also log the abort code from AX */
1463 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1464 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1465 	return txn;
1466 }
1467 
1468 static inline u64 get_pebs_status(void *n)
1469 {
1470 	if (x86_pmu.intel_cap.pebs_format < 4)
1471 		return ((struct pebs_record_nhm *)n)->status;
1472 	return ((struct pebs_basic *)n)->applicable_counters;
1473 }
1474 
1475 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1476 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1477 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1478 		 PERF_X86_EVENT_PEBS_NA_HSW)
1479 
1480 static u64 get_data_src(struct perf_event *event, u64 aux)
1481 {
1482 	u64 val = PERF_MEM_NA;
1483 	int fl = event->hw.flags;
1484 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1485 
1486 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1487 		val = load_latency_data(event, aux);
1488 	else if (fl & PERF_X86_EVENT_PEBS_STLAT)
1489 		val = store_latency_data(event, aux);
1490 	else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID)
1491 		val = x86_pmu.pebs_latency_data(event, aux);
1492 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1493 		val = precise_datala_hsw(event, aux);
1494 	else if (fst)
1495 		val = precise_store_data(aux);
1496 	return val;
1497 }
1498 
1499 #define PERF_SAMPLE_ADDR_TYPE	(PERF_SAMPLE_ADDR |		\
1500 				 PERF_SAMPLE_PHYS_ADDR |	\
1501 				 PERF_SAMPLE_DATA_PAGE_SIZE)
1502 
1503 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1504 				   struct pt_regs *iregs, void *__pebs,
1505 				   struct perf_sample_data *data,
1506 				   struct pt_regs *regs)
1507 {
1508 	/*
1509 	 * We cast to the biggest pebs_record but are careful not to
1510 	 * unconditionally access the 'extra' entries.
1511 	 */
1512 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1513 	struct pebs_record_skl *pebs = __pebs;
1514 	u64 sample_type;
1515 	int fll;
1516 
1517 	if (pebs == NULL)
1518 		return;
1519 
1520 	sample_type = event->attr.sample_type;
1521 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1522 
1523 	perf_sample_data_init(data, 0, event->hw.last_period);
1524 
1525 	data->period = event->hw.last_period;
1526 
1527 	/*
1528 	 * Use latency for weight (only avail with PEBS-LL)
1529 	 */
1530 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE))
1531 		data->weight.full = pebs->lat;
1532 
1533 	/*
1534 	 * data.data_src encodes the data source
1535 	 */
1536 	if (sample_type & PERF_SAMPLE_DATA_SRC)
1537 		data->data_src.val = get_data_src(event, pebs->dse);
1538 
1539 	/*
1540 	 * We must however always use iregs for the unwinder to stay sane; the
1541 	 * record BP,SP,IP can point into thin air when the record is from a
1542 	 * previous PMI context or an (I)RET happened between the record and
1543 	 * PMI.
1544 	 */
1545 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1546 		data->callchain = perf_callchain(event, iregs);
1547 
1548 	/*
1549 	 * We use the interrupt regs as a base because the PEBS record does not
1550 	 * contain a full regs set, specifically it seems to lack segment
1551 	 * descriptors, which get used by things like user_mode().
1552 	 *
1553 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1554 	 */
1555 	*regs = *iregs;
1556 
1557 	/*
1558 	 * Initialize regs_>flags from PEBS,
1559 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1560 	 * i.e., do not rely on it being zero:
1561 	 */
1562 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1563 
1564 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1565 		regs->ax = pebs->ax;
1566 		regs->bx = pebs->bx;
1567 		regs->cx = pebs->cx;
1568 		regs->dx = pebs->dx;
1569 		regs->si = pebs->si;
1570 		regs->di = pebs->di;
1571 
1572 		regs->bp = pebs->bp;
1573 		regs->sp = pebs->sp;
1574 
1575 #ifndef CONFIG_X86_32
1576 		regs->r8 = pebs->r8;
1577 		regs->r9 = pebs->r9;
1578 		regs->r10 = pebs->r10;
1579 		regs->r11 = pebs->r11;
1580 		regs->r12 = pebs->r12;
1581 		regs->r13 = pebs->r13;
1582 		regs->r14 = pebs->r14;
1583 		regs->r15 = pebs->r15;
1584 #endif
1585 	}
1586 
1587 	if (event->attr.precise_ip > 1) {
1588 		/*
1589 		 * Haswell and later processors have an 'eventing IP'
1590 		 * (real IP) which fixes the off-by-1 skid in hardware.
1591 		 * Use it when precise_ip >= 2 :
1592 		 */
1593 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1594 			set_linear_ip(regs, pebs->real_ip);
1595 			regs->flags |= PERF_EFLAGS_EXACT;
1596 		} else {
1597 			/* Otherwise, use PEBS off-by-1 IP: */
1598 			set_linear_ip(regs, pebs->ip);
1599 
1600 			/*
1601 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
1602 			 * using the LBR. If successful, the fixup function
1603 			 * corrects regs->ip and calls set_linear_ip() on regs:
1604 			 */
1605 			if (intel_pmu_pebs_fixup_ip(regs))
1606 				regs->flags |= PERF_EFLAGS_EXACT;
1607 		}
1608 	} else {
1609 		/*
1610 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
1611 		 * no fixup attempted:
1612 		 */
1613 		set_linear_ip(regs, pebs->ip);
1614 	}
1615 
1616 
1617 	if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
1618 	    x86_pmu.intel_cap.pebs_format >= 1)
1619 		data->addr = pebs->dla;
1620 
1621 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1622 		/* Only set the TSX weight when no memory weight. */
1623 		if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll)
1624 			data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
1625 
1626 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1627 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1628 							      pebs->ax);
1629 	}
1630 
1631 	/*
1632 	 * v3 supplies an accurate time stamp, so we use that
1633 	 * for the time stamp.
1634 	 *
1635 	 * We can only do this for the default trace clock.
1636 	 */
1637 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1638 		event->attr.use_clockid == 0)
1639 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1640 
1641 	if (has_branch_stack(event))
1642 		data->br_stack = &cpuc->lbr_stack;
1643 }
1644 
1645 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1646 				    struct pebs_gprs *gprs)
1647 {
1648 	regs->ax = gprs->ax;
1649 	regs->bx = gprs->bx;
1650 	regs->cx = gprs->cx;
1651 	regs->dx = gprs->dx;
1652 	regs->si = gprs->si;
1653 	regs->di = gprs->di;
1654 	regs->bp = gprs->bp;
1655 	regs->sp = gprs->sp;
1656 #ifndef CONFIG_X86_32
1657 	regs->r8 = gprs->r8;
1658 	regs->r9 = gprs->r9;
1659 	regs->r10 = gprs->r10;
1660 	regs->r11 = gprs->r11;
1661 	regs->r12 = gprs->r12;
1662 	regs->r13 = gprs->r13;
1663 	regs->r14 = gprs->r14;
1664 	regs->r15 = gprs->r15;
1665 #endif
1666 }
1667 
1668 #define PEBS_LATENCY_MASK			0xffff
1669 #define PEBS_CACHE_LATENCY_OFFSET		32
1670 
1671 /*
1672  * With adaptive PEBS the layout depends on what fields are configured.
1673  */
1674 
1675 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1676 					    struct pt_regs *iregs, void *__pebs,
1677 					    struct perf_sample_data *data,
1678 					    struct pt_regs *regs)
1679 {
1680 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1681 	struct pebs_basic *basic = __pebs;
1682 	void *next_record = basic + 1;
1683 	u64 sample_type;
1684 	u64 format_size;
1685 	struct pebs_meminfo *meminfo = NULL;
1686 	struct pebs_gprs *gprs = NULL;
1687 	struct x86_perf_regs *perf_regs;
1688 
1689 	if (basic == NULL)
1690 		return;
1691 
1692 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
1693 	perf_regs->xmm_regs = NULL;
1694 
1695 	sample_type = event->attr.sample_type;
1696 	format_size = basic->format_size;
1697 	perf_sample_data_init(data, 0, event->hw.last_period);
1698 	data->period = event->hw.last_period;
1699 
1700 	if (event->attr.use_clockid == 0)
1701 		data->time = native_sched_clock_from_tsc(basic->tsc);
1702 
1703 	/*
1704 	 * We must however always use iregs for the unwinder to stay sane; the
1705 	 * record BP,SP,IP can point into thin air when the record is from a
1706 	 * previous PMI context or an (I)RET happened between the record and
1707 	 * PMI.
1708 	 */
1709 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1710 		data->callchain = perf_callchain(event, iregs);
1711 
1712 	*regs = *iregs;
1713 	/* The ip in basic is EventingIP */
1714 	set_linear_ip(regs, basic->ip);
1715 	regs->flags = PERF_EFLAGS_EXACT;
1716 
1717 	/*
1718 	 * The record for MEMINFO is in front of GP
1719 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1720 	 * Save the pointer here but process later.
1721 	 */
1722 	if (format_size & PEBS_DATACFG_MEMINFO) {
1723 		meminfo = next_record;
1724 		next_record = meminfo + 1;
1725 	}
1726 
1727 	if (format_size & PEBS_DATACFG_GP) {
1728 		gprs = next_record;
1729 		next_record = gprs + 1;
1730 
1731 		if (event->attr.precise_ip < 2) {
1732 			set_linear_ip(regs, gprs->ip);
1733 			regs->flags &= ~PERF_EFLAGS_EXACT;
1734 		}
1735 
1736 		if (sample_type & PERF_SAMPLE_REGS_INTR)
1737 			adaptive_pebs_save_regs(regs, gprs);
1738 	}
1739 
1740 	if (format_size & PEBS_DATACFG_MEMINFO) {
1741 		if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
1742 			u64 weight = meminfo->latency;
1743 
1744 			if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) {
1745 				data->weight.var2_w = weight & PEBS_LATENCY_MASK;
1746 				weight >>= PEBS_CACHE_LATENCY_OFFSET;
1747 			}
1748 
1749 			/*
1750 			 * Although meminfo::latency is defined as a u64,
1751 			 * only the lower 32 bits include the valid data
1752 			 * in practice on Ice Lake and earlier platforms.
1753 			 */
1754 			if (sample_type & PERF_SAMPLE_WEIGHT) {
1755 				data->weight.full = weight ?:
1756 					intel_get_tsx_weight(meminfo->tsx_tuning);
1757 			} else {
1758 				data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?:
1759 					intel_get_tsx_weight(meminfo->tsx_tuning);
1760 			}
1761 		}
1762 
1763 		if (sample_type & PERF_SAMPLE_DATA_SRC)
1764 			data->data_src.val = get_data_src(event, meminfo->aux);
1765 
1766 		if (sample_type & PERF_SAMPLE_ADDR_TYPE)
1767 			data->addr = meminfo->address;
1768 
1769 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1770 			data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1771 							  gprs ? gprs->ax : 0);
1772 	}
1773 
1774 	if (format_size & PEBS_DATACFG_XMMS) {
1775 		struct pebs_xmm *xmm = next_record;
1776 
1777 		next_record = xmm + 1;
1778 		perf_regs->xmm_regs = xmm->xmm;
1779 	}
1780 
1781 	if (format_size & PEBS_DATACFG_LBRS) {
1782 		struct lbr_entry *lbr = next_record;
1783 		int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1784 					& 0xff) + 1;
1785 		next_record = next_record + num_lbr * sizeof(struct lbr_entry);
1786 
1787 		if (has_branch_stack(event)) {
1788 			intel_pmu_store_pebs_lbrs(lbr);
1789 			data->br_stack = &cpuc->lbr_stack;
1790 		}
1791 	}
1792 
1793 	WARN_ONCE(next_record != __pebs + (format_size >> 48),
1794 			"PEBS record size %llu, expected %llu, config %llx\n",
1795 			format_size >> 48,
1796 			(u64)(next_record - __pebs),
1797 			basic->format_size);
1798 }
1799 
1800 static inline void *
1801 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1802 {
1803 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1804 	void *at;
1805 	u64 pebs_status;
1806 
1807 	/*
1808 	 * fmt0 does not have a status bitfield (does not use
1809 	 * perf_record_nhm format)
1810 	 */
1811 	if (x86_pmu.intel_cap.pebs_format < 1)
1812 		return base;
1813 
1814 	if (base == NULL)
1815 		return NULL;
1816 
1817 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1818 		unsigned long status = get_pebs_status(at);
1819 
1820 		if (test_bit(bit, (unsigned long *)&status)) {
1821 			/* PEBS v3 has accurate status bits */
1822 			if (x86_pmu.intel_cap.pebs_format >= 3)
1823 				return at;
1824 
1825 			if (status == (1 << bit))
1826 				return at;
1827 
1828 			/* clear non-PEBS bit and re-check */
1829 			pebs_status = status & cpuc->pebs_enabled;
1830 			pebs_status &= PEBS_COUNTER_MASK;
1831 			if (pebs_status == (1 << bit))
1832 				return at;
1833 		}
1834 	}
1835 	return NULL;
1836 }
1837 
1838 void intel_pmu_auto_reload_read(struct perf_event *event)
1839 {
1840 	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1841 
1842 	perf_pmu_disable(event->pmu);
1843 	intel_pmu_drain_pebs_buffer();
1844 	perf_pmu_enable(event->pmu);
1845 }
1846 
1847 /*
1848  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1849  */
1850 static int
1851 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1852 {
1853 	struct hw_perf_event *hwc = &event->hw;
1854 	int shift = 64 - x86_pmu.cntval_bits;
1855 	u64 period = hwc->sample_period;
1856 	u64 prev_raw_count, new_raw_count;
1857 	s64 new, old;
1858 
1859 	WARN_ON(!period);
1860 
1861 	/*
1862 	 * drain_pebs() only happens when the PMU is disabled.
1863 	 */
1864 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1865 
1866 	prev_raw_count = local64_read(&hwc->prev_count);
1867 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1868 	local64_set(&hwc->prev_count, new_raw_count);
1869 
1870 	/*
1871 	 * Since the counter increments a negative counter value and
1872 	 * overflows on the sign switch, giving the interval:
1873 	 *
1874 	 *   [-period, 0]
1875 	 *
1876 	 * the difference between two consecutive reads is:
1877 	 *
1878 	 *   A) value2 - value1;
1879 	 *      when no overflows have happened in between,
1880 	 *
1881 	 *   B) (0 - value1) + (value2 - (-period));
1882 	 *      when one overflow happened in between,
1883 	 *
1884 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1885 	 *      when @n overflows happened in between.
1886 	 *
1887 	 * Here A) is the obvious difference, B) is the extension to the
1888 	 * discrete interval, where the first term is to the top of the
1889 	 * interval and the second term is from the bottom of the next
1890 	 * interval and C) the extension to multiple intervals, where the
1891 	 * middle term is the whole intervals covered.
1892 	 *
1893 	 * An equivalent of C, by reduction, is:
1894 	 *
1895 	 *   value2 - value1 + n * period
1896 	 */
1897 	new = ((s64)(new_raw_count << shift) >> shift);
1898 	old = ((s64)(prev_raw_count << shift) >> shift);
1899 	local64_add(new - old + count * period, &event->count);
1900 
1901 	local64_set(&hwc->period_left, -new);
1902 
1903 	perf_event_update_userpage(event);
1904 
1905 	return 0;
1906 }
1907 
1908 static __always_inline void
1909 __intel_pmu_pebs_event(struct perf_event *event,
1910 		       struct pt_regs *iregs,
1911 		       struct perf_sample_data *data,
1912 		       void *base, void *top,
1913 		       int bit, int count,
1914 		       void (*setup_sample)(struct perf_event *,
1915 					    struct pt_regs *,
1916 					    void *,
1917 					    struct perf_sample_data *,
1918 					    struct pt_regs *))
1919 {
1920 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1921 	struct hw_perf_event *hwc = &event->hw;
1922 	struct x86_perf_regs perf_regs;
1923 	struct pt_regs *regs = &perf_regs.regs;
1924 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1925 	static struct pt_regs dummy_iregs;
1926 
1927 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1928 		/*
1929 		 * Now, auto-reload is only enabled in fixed period mode.
1930 		 * The reload value is always hwc->sample_period.
1931 		 * May need to change it, if auto-reload is enabled in
1932 		 * freq mode later.
1933 		 */
1934 		intel_pmu_save_and_restart_reload(event, count);
1935 	} else if (!intel_pmu_save_and_restart(event))
1936 		return;
1937 
1938 	if (!iregs)
1939 		iregs = &dummy_iregs;
1940 
1941 	while (count > 1) {
1942 		setup_sample(event, iregs, at, data, regs);
1943 		perf_event_output(event, data, regs);
1944 		at += cpuc->pebs_record_size;
1945 		at = get_next_pebs_record_by_bit(at, top, bit);
1946 		count--;
1947 	}
1948 
1949 	setup_sample(event, iregs, at, data, regs);
1950 	if (iregs == &dummy_iregs) {
1951 		/*
1952 		 * The PEBS records may be drained in the non-overflow context,
1953 		 * e.g., large PEBS + context switch. Perf should treat the
1954 		 * last record the same as other PEBS records, and doesn't
1955 		 * invoke the generic overflow handler.
1956 		 */
1957 		perf_event_output(event, data, regs);
1958 	} else {
1959 		/*
1960 		 * All but the last records are processed.
1961 		 * The last one is left to be able to call the overflow handler.
1962 		 */
1963 		if (perf_event_overflow(event, data, regs))
1964 			x86_pmu_stop(event, 0);
1965 	}
1966 }
1967 
1968 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
1969 {
1970 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1971 	struct debug_store *ds = cpuc->ds;
1972 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1973 	struct pebs_record_core *at, *top;
1974 	int n;
1975 
1976 	if (!x86_pmu.pebs_active)
1977 		return;
1978 
1979 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1980 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1981 
1982 	/*
1983 	 * Whatever else happens, drain the thing
1984 	 */
1985 	ds->pebs_index = ds->pebs_buffer_base;
1986 
1987 	if (!test_bit(0, cpuc->active_mask))
1988 		return;
1989 
1990 	WARN_ON_ONCE(!event);
1991 
1992 	if (!event->attr.precise_ip)
1993 		return;
1994 
1995 	n = top - at;
1996 	if (n <= 0) {
1997 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1998 			intel_pmu_save_and_restart_reload(event, 0);
1999 		return;
2000 	}
2001 
2002 	__intel_pmu_pebs_event(event, iregs, data, at, top, 0, n,
2003 			       setup_pebs_fixed_sample_data);
2004 }
2005 
2006 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
2007 {
2008 	struct perf_event *event;
2009 	int bit;
2010 
2011 	/*
2012 	 * The drain_pebs() could be called twice in a short period
2013 	 * for auto-reload event in pmu::read(). There are no
2014 	 * overflows have happened in between.
2015 	 * It needs to call intel_pmu_save_and_restart_reload() to
2016 	 * update the event->count for this case.
2017 	 */
2018 	for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
2019 		event = cpuc->events[bit];
2020 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2021 			intel_pmu_save_and_restart_reload(event, 0);
2022 	}
2023 }
2024 
2025 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
2026 {
2027 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2028 	struct debug_store *ds = cpuc->ds;
2029 	struct perf_event *event;
2030 	void *base, *at, *top;
2031 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2032 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2033 	int bit, i, size;
2034 	u64 mask;
2035 
2036 	if (!x86_pmu.pebs_active)
2037 		return;
2038 
2039 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
2040 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
2041 
2042 	ds->pebs_index = ds->pebs_buffer_base;
2043 
2044 	mask = (1ULL << x86_pmu.max_pebs_events) - 1;
2045 	size = x86_pmu.max_pebs_events;
2046 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
2047 		mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
2048 		size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
2049 	}
2050 
2051 	if (unlikely(base >= top)) {
2052 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2053 		return;
2054 	}
2055 
2056 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
2057 		struct pebs_record_nhm *p = at;
2058 		u64 pebs_status;
2059 
2060 		pebs_status = p->status & cpuc->pebs_enabled;
2061 		pebs_status &= mask;
2062 
2063 		/* PEBS v3 has more accurate status bits */
2064 		if (x86_pmu.intel_cap.pebs_format >= 3) {
2065 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2066 				counts[bit]++;
2067 
2068 			continue;
2069 		}
2070 
2071 		/*
2072 		 * On some CPUs the PEBS status can be zero when PEBS is
2073 		 * racing with clearing of GLOBAL_STATUS.
2074 		 *
2075 		 * Normally we would drop that record, but in the
2076 		 * case when there is only a single active PEBS event
2077 		 * we can assume it's for that event.
2078 		 */
2079 		if (!pebs_status && cpuc->pebs_enabled &&
2080 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
2081 			pebs_status = p->status = cpuc->pebs_enabled;
2082 
2083 		bit = find_first_bit((unsigned long *)&pebs_status,
2084 					x86_pmu.max_pebs_events);
2085 		if (bit >= x86_pmu.max_pebs_events)
2086 			continue;
2087 
2088 		/*
2089 		 * The PEBS hardware does not deal well with the situation
2090 		 * when events happen near to each other and multiple bits
2091 		 * are set. But it should happen rarely.
2092 		 *
2093 		 * If these events include one PEBS and multiple non-PEBS
2094 		 * events, it doesn't impact PEBS record. The record will
2095 		 * be handled normally. (slow path)
2096 		 *
2097 		 * If these events include two or more PEBS events, the
2098 		 * records for the events can be collapsed into a single
2099 		 * one, and it's not possible to reconstruct all events
2100 		 * that caused the PEBS record. It's called collision.
2101 		 * If collision happened, the record will be dropped.
2102 		 */
2103 		if (pebs_status != (1ULL << bit)) {
2104 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
2105 				error[i]++;
2106 			continue;
2107 		}
2108 
2109 		counts[bit]++;
2110 	}
2111 
2112 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2113 		if ((counts[bit] == 0) && (error[bit] == 0))
2114 			continue;
2115 
2116 		event = cpuc->events[bit];
2117 		if (WARN_ON_ONCE(!event))
2118 			continue;
2119 
2120 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2121 			continue;
2122 
2123 		/* log dropped samples number */
2124 		if (error[bit]) {
2125 			perf_log_lost_samples(event, error[bit]);
2126 
2127 			if (iregs && perf_event_account_interrupt(event))
2128 				x86_pmu_stop(event, 0);
2129 		}
2130 
2131 		if (counts[bit]) {
2132 			__intel_pmu_pebs_event(event, iregs, data, base,
2133 					       top, bit, counts[bit],
2134 					       setup_pebs_fixed_sample_data);
2135 		}
2136 	}
2137 }
2138 
2139 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
2140 {
2141 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2142 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2143 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
2144 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
2145 	struct debug_store *ds = cpuc->ds;
2146 	struct perf_event *event;
2147 	void *base, *at, *top;
2148 	int bit, size;
2149 	u64 mask;
2150 
2151 	if (!x86_pmu.pebs_active)
2152 		return;
2153 
2154 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
2155 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
2156 
2157 	ds->pebs_index = ds->pebs_buffer_base;
2158 
2159 	mask = ((1ULL << max_pebs_events) - 1) |
2160 	       (((1ULL << num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
2161 	size = INTEL_PMC_IDX_FIXED + num_counters_fixed;
2162 
2163 	if (unlikely(base >= top)) {
2164 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2165 		return;
2166 	}
2167 
2168 	for (at = base; at < top; at += cpuc->pebs_record_size) {
2169 		u64 pebs_status;
2170 
2171 		pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
2172 		pebs_status &= mask;
2173 
2174 		for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2175 			counts[bit]++;
2176 	}
2177 
2178 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2179 		if (counts[bit] == 0)
2180 			continue;
2181 
2182 		event = cpuc->events[bit];
2183 		if (WARN_ON_ONCE(!event))
2184 			continue;
2185 
2186 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2187 			continue;
2188 
2189 		__intel_pmu_pebs_event(event, iregs, data, base,
2190 				       top, bit, counts[bit],
2191 				       setup_pebs_adaptive_sample_data);
2192 	}
2193 }
2194 
2195 /*
2196  * BTS, PEBS probe and setup
2197  */
2198 
2199 void __init intel_ds_init(void)
2200 {
2201 	/*
2202 	 * No support for 32bit formats
2203 	 */
2204 	if (!boot_cpu_has(X86_FEATURE_DTES64))
2205 		return;
2206 
2207 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
2208 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
2209 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
2210 	if (x86_pmu.version <= 4)
2211 		x86_pmu.pebs_no_isolation = 1;
2212 
2213 	if (x86_pmu.pebs) {
2214 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
2215 		char *pebs_qual = "";
2216 		int format = x86_pmu.intel_cap.pebs_format;
2217 
2218 		if (format < 4)
2219 			x86_pmu.intel_cap.pebs_baseline = 0;
2220 
2221 		switch (format) {
2222 		case 0:
2223 			pr_cont("PEBS fmt0%c, ", pebs_type);
2224 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
2225 			/*
2226 			 * Using >PAGE_SIZE buffers makes the WRMSR to
2227 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
2228 			 * mysteriously hang on Core2.
2229 			 *
2230 			 * As a workaround, we don't do this.
2231 			 */
2232 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
2233 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
2234 			break;
2235 
2236 		case 1:
2237 			pr_cont("PEBS fmt1%c, ", pebs_type);
2238 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
2239 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2240 			break;
2241 
2242 		case 2:
2243 			pr_cont("PEBS fmt2%c, ", pebs_type);
2244 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2245 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2246 			break;
2247 
2248 		case 3:
2249 			pr_cont("PEBS fmt3%c, ", pebs_type);
2250 			x86_pmu.pebs_record_size =
2251 						sizeof(struct pebs_record_skl);
2252 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2253 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2254 			break;
2255 
2256 		case 4:
2257 		case 5:
2258 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2259 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2260 			if (x86_pmu.intel_cap.pebs_baseline) {
2261 				x86_pmu.large_pebs_flags |=
2262 					PERF_SAMPLE_BRANCH_STACK |
2263 					PERF_SAMPLE_TIME;
2264 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
2265 				pebs_qual = "-baseline";
2266 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
2267 			} else {
2268 				/* Only basic record supported */
2269 				x86_pmu.large_pebs_flags &=
2270 					~(PERF_SAMPLE_ADDR |
2271 					  PERF_SAMPLE_TIME |
2272 					  PERF_SAMPLE_DATA_SRC |
2273 					  PERF_SAMPLE_TRANSACTION |
2274 					  PERF_SAMPLE_REGS_USER |
2275 					  PERF_SAMPLE_REGS_INTR);
2276 			}
2277 			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2278 
2279 			if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) {
2280 				pr_cont("PEBS-via-PT, ");
2281 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
2282 			}
2283 
2284 			break;
2285 
2286 		default:
2287 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2288 			x86_pmu.pebs = 0;
2289 		}
2290 	}
2291 }
2292 
2293 void perf_restore_debug_store(void)
2294 {
2295 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2296 
2297 	if (!x86_pmu.bts && !x86_pmu.pebs)
2298 		return;
2299 
2300 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
2301 }
2302