xref: /linux/arch/x86/events/intel/ds.c (revision 6a4aee277740d04ac0fd54cfa17cc28261932ddc)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/types.h>
4 #include <linux/slab.h>
5 #include <linux/sched/clock.h>
6 
7 #include <asm/cpu_entry_area.h>
8 #include <asm/debugreg.h>
9 #include <asm/perf_event.h>
10 #include <asm/tlbflush.h>
11 #include <asm/insn.h>
12 #include <asm/io.h>
13 #include <asm/timer.h>
14 
15 #include "../perf_event.h"
16 
17 /* Waste a full page so it can be mapped into the cpu_entry_area */
18 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
19 
20 /* The size of a BTS record in bytes: */
21 #define BTS_RECORD_SIZE		24
22 
23 #define PEBS_FIXUP_SIZE		PAGE_SIZE
24 
25 /*
26  * pebs_record_32 for p4 and core not supported
27 
28 struct pebs_record_32 {
29 	u32 flags, ip;
30 	u32 ax, bc, cx, dx;
31 	u32 si, di, bp, sp;
32 };
33 
34  */
35 
36 union intel_x86_pebs_dse {
37 	u64 val;
38 	struct {
39 		unsigned int ld_dse:4;
40 		unsigned int ld_stlb_miss:1;
41 		unsigned int ld_locked:1;
42 		unsigned int ld_data_blk:1;
43 		unsigned int ld_addr_blk:1;
44 		unsigned int ld_reserved:24;
45 	};
46 	struct {
47 		unsigned int st_l1d_hit:1;
48 		unsigned int st_reserved1:3;
49 		unsigned int st_stlb_miss:1;
50 		unsigned int st_locked:1;
51 		unsigned int st_reserved2:26;
52 	};
53 	struct {
54 		unsigned int st_lat_dse:4;
55 		unsigned int st_lat_stlb_miss:1;
56 		unsigned int st_lat_locked:1;
57 		unsigned int ld_reserved3:26;
58 	};
59 	struct {
60 		unsigned int mtl_dse:5;
61 		unsigned int mtl_locked:1;
62 		unsigned int mtl_stlb_miss:1;
63 		unsigned int mtl_fwd_blk:1;
64 		unsigned int ld_reserved4:24;
65 	};
66 };
67 
68 
69 /*
70  * Map PEBS Load Latency Data Source encodings to generic
71  * memory data source information
72  */
73 #define P(a, b) PERF_MEM_S(a, b)
74 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
75 #define LEVEL(x) P(LVLNUM, x)
76 #define REM P(REMOTE, REMOTE)
77 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
78 
79 /* Version for Sandy Bridge and later */
80 static u64 pebs_data_source[] = {
81 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
82 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
83 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
84 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
85 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
86 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
87 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
88 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
89 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
90 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
91 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
92 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
93 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
94 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
95 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
96 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
97 };
98 
99 /* Patch up minor differences in the bits */
100 void __init intel_pmu_pebs_data_source_nhm(void)
101 {
102 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
103 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
104 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
105 }
106 
107 static void __init __intel_pmu_pebs_data_source_skl(bool pmem, u64 *data_source)
108 {
109 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
110 
111 	data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
112 	data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
113 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
114 	data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
115 	data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
116 }
117 
118 void __init intel_pmu_pebs_data_source_skl(bool pmem)
119 {
120 	__intel_pmu_pebs_data_source_skl(pmem, pebs_data_source);
121 }
122 
123 static void __init __intel_pmu_pebs_data_source_grt(u64 *data_source)
124 {
125 	data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
126 	data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
127 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
128 }
129 
130 void __init intel_pmu_pebs_data_source_grt(void)
131 {
132 	__intel_pmu_pebs_data_source_grt(pebs_data_source);
133 }
134 
135 void __init intel_pmu_pebs_data_source_adl(void)
136 {
137 	u64 *data_source;
138 
139 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
140 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
141 	__intel_pmu_pebs_data_source_skl(false, data_source);
142 
143 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
144 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
145 	__intel_pmu_pebs_data_source_grt(data_source);
146 }
147 
148 static void __init __intel_pmu_pebs_data_source_cmt(u64 *data_source)
149 {
150 	data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOPX, FWD);
151 	data_source[0x08] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
152 	data_source[0x0a] = OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, NONE);
153 	data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
154 	data_source[0x0c] = OP_LH | LEVEL(RAM) | REM | P(SNOOPX, FWD);
155 	data_source[0x0d] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, HITM);
156 }
157 
158 void __init intel_pmu_pebs_data_source_mtl(void)
159 {
160 	u64 *data_source;
161 
162 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX].pebs_data_source;
163 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
164 	__intel_pmu_pebs_data_source_skl(false, data_source);
165 
166 	data_source = x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX].pebs_data_source;
167 	memcpy(data_source, pebs_data_source, sizeof(pebs_data_source));
168 	__intel_pmu_pebs_data_source_cmt(data_source);
169 }
170 
171 void __init intel_pmu_pebs_data_source_cmt(void)
172 {
173 	__intel_pmu_pebs_data_source_cmt(pebs_data_source);
174 }
175 
176 static u64 precise_store_data(u64 status)
177 {
178 	union intel_x86_pebs_dse dse;
179 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
180 
181 	dse.val = status;
182 
183 	/*
184 	 * bit 4: TLB access
185 	 * 1 = stored missed 2nd level TLB
186 	 *
187 	 * so it either hit the walker or the OS
188 	 * otherwise hit 2nd level TLB
189 	 */
190 	if (dse.st_stlb_miss)
191 		val |= P(TLB, MISS);
192 	else
193 		val |= P(TLB, HIT);
194 
195 	/*
196 	 * bit 0: hit L1 data cache
197 	 * if not set, then all we know is that
198 	 * it missed L1D
199 	 */
200 	if (dse.st_l1d_hit)
201 		val |= P(LVL, HIT);
202 	else
203 		val |= P(LVL, MISS);
204 
205 	/*
206 	 * bit 5: Locked prefix
207 	 */
208 	if (dse.st_locked)
209 		val |= P(LOCK, LOCKED);
210 
211 	return val;
212 }
213 
214 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
215 {
216 	union perf_mem_data_src dse;
217 
218 	dse.val = PERF_MEM_NA;
219 
220 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
221 		dse.mem_op = PERF_MEM_OP_STORE;
222 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
223 		dse.mem_op = PERF_MEM_OP_LOAD;
224 
225 	/*
226 	 * L1 info only valid for following events:
227 	 *
228 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
229 	 * MEM_UOPS_RETIRED.LOCK_STORES
230 	 * MEM_UOPS_RETIRED.SPLIT_STORES
231 	 * MEM_UOPS_RETIRED.ALL_STORES
232 	 */
233 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
234 		if (status & 1)
235 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
236 		else
237 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
238 	}
239 	return dse.val;
240 }
241 
242 static inline void pebs_set_tlb_lock(u64 *val, bool tlb, bool lock)
243 {
244 	/*
245 	 * TLB access
246 	 * 0 = did not miss 2nd level TLB
247 	 * 1 = missed 2nd level TLB
248 	 */
249 	if (tlb)
250 		*val |= P(TLB, MISS) | P(TLB, L2);
251 	else
252 		*val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
253 
254 	/* locked prefix */
255 	if (lock)
256 		*val |= P(LOCK, LOCKED);
257 }
258 
259 /* Retrieve the latency data for e-core of ADL */
260 static u64 __adl_latency_data_small(struct perf_event *event, u64 status,
261 				     u8 dse, bool tlb, bool lock, bool blk)
262 {
263 	u64 val;
264 
265 	WARN_ON_ONCE(hybrid_pmu(event->pmu)->pmu_type == hybrid_big);
266 
267 	dse &= PERF_PEBS_DATA_SOURCE_MASK;
268 	val = hybrid_var(event->pmu, pebs_data_source)[dse];
269 
270 	pebs_set_tlb_lock(&val, tlb, lock);
271 
272 	if (blk)
273 		val |= P(BLK, DATA);
274 	else
275 		val |= P(BLK, NA);
276 
277 	return val;
278 }
279 
280 u64 adl_latency_data_small(struct perf_event *event, u64 status)
281 {
282 	union intel_x86_pebs_dse dse;
283 
284 	dse.val = status;
285 
286 	return __adl_latency_data_small(event, status, dse.ld_dse,
287 					dse.ld_locked, dse.ld_stlb_miss,
288 					dse.ld_data_blk);
289 }
290 
291 /* Retrieve the latency data for e-core of MTL */
292 u64 mtl_latency_data_small(struct perf_event *event, u64 status)
293 {
294 	union intel_x86_pebs_dse dse;
295 
296 	dse.val = status;
297 
298 	return __adl_latency_data_small(event, status, dse.mtl_dse,
299 					dse.mtl_stlb_miss, dse.mtl_locked,
300 					dse.mtl_fwd_blk);
301 }
302 
303 static u64 load_latency_data(struct perf_event *event, u64 status)
304 {
305 	union intel_x86_pebs_dse dse;
306 	u64 val;
307 
308 	dse.val = status;
309 
310 	/*
311 	 * use the mapping table for bit 0-3
312 	 */
313 	val = hybrid_var(event->pmu, pebs_data_source)[dse.ld_dse];
314 
315 	/*
316 	 * Nehalem models do not support TLB, Lock infos
317 	 */
318 	if (x86_pmu.pebs_no_tlb) {
319 		val |= P(TLB, NA) | P(LOCK, NA);
320 		return val;
321 	}
322 
323 	pebs_set_tlb_lock(&val, dse.ld_stlb_miss, dse.ld_locked);
324 
325 	/*
326 	 * Ice Lake and earlier models do not support block infos.
327 	 */
328 	if (!x86_pmu.pebs_block) {
329 		val |= P(BLK, NA);
330 		return val;
331 	}
332 	/*
333 	 * bit 6: load was blocked since its data could not be forwarded
334 	 *        from a preceding store
335 	 */
336 	if (dse.ld_data_blk)
337 		val |= P(BLK, DATA);
338 
339 	/*
340 	 * bit 7: load was blocked due to potential address conflict with
341 	 *        a preceding store
342 	 */
343 	if (dse.ld_addr_blk)
344 		val |= P(BLK, ADDR);
345 
346 	if (!dse.ld_data_blk && !dse.ld_addr_blk)
347 		val |= P(BLK, NA);
348 
349 	return val;
350 }
351 
352 static u64 store_latency_data(struct perf_event *event, u64 status)
353 {
354 	union intel_x86_pebs_dse dse;
355 	union perf_mem_data_src src;
356 	u64 val;
357 
358 	dse.val = status;
359 
360 	/*
361 	 * use the mapping table for bit 0-3
362 	 */
363 	val = hybrid_var(event->pmu, pebs_data_source)[dse.st_lat_dse];
364 
365 	pebs_set_tlb_lock(&val, dse.st_lat_stlb_miss, dse.st_lat_locked);
366 
367 	val |= P(BLK, NA);
368 
369 	/*
370 	 * the pebs_data_source table is only for loads
371 	 * so override the mem_op to say STORE instead
372 	 */
373 	src.val = val;
374 	src.mem_op = P(OP,STORE);
375 
376 	return src.val;
377 }
378 
379 struct pebs_record_core {
380 	u64 flags, ip;
381 	u64 ax, bx, cx, dx;
382 	u64 si, di, bp, sp;
383 	u64 r8,  r9,  r10, r11;
384 	u64 r12, r13, r14, r15;
385 };
386 
387 struct pebs_record_nhm {
388 	u64 flags, ip;
389 	u64 ax, bx, cx, dx;
390 	u64 si, di, bp, sp;
391 	u64 r8,  r9,  r10, r11;
392 	u64 r12, r13, r14, r15;
393 	u64 status, dla, dse, lat;
394 };
395 
396 /*
397  * Same as pebs_record_nhm, with two additional fields.
398  */
399 struct pebs_record_hsw {
400 	u64 flags, ip;
401 	u64 ax, bx, cx, dx;
402 	u64 si, di, bp, sp;
403 	u64 r8,  r9,  r10, r11;
404 	u64 r12, r13, r14, r15;
405 	u64 status, dla, dse, lat;
406 	u64 real_ip, tsx_tuning;
407 };
408 
409 union hsw_tsx_tuning {
410 	struct {
411 		u32 cycles_last_block     : 32,
412 		    hle_abort		  : 1,
413 		    rtm_abort		  : 1,
414 		    instruction_abort     : 1,
415 		    non_instruction_abort : 1,
416 		    retry		  : 1,
417 		    data_conflict	  : 1,
418 		    capacity_writes	  : 1,
419 		    capacity_reads	  : 1;
420 	};
421 	u64	    value;
422 };
423 
424 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
425 
426 /* Same as HSW, plus TSC */
427 
428 struct pebs_record_skl {
429 	u64 flags, ip;
430 	u64 ax, bx, cx, dx;
431 	u64 si, di, bp, sp;
432 	u64 r8,  r9,  r10, r11;
433 	u64 r12, r13, r14, r15;
434 	u64 status, dla, dse, lat;
435 	u64 real_ip, tsx_tuning;
436 	u64 tsc;
437 };
438 
439 void init_debug_store_on_cpu(int cpu)
440 {
441 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
442 
443 	if (!ds)
444 		return;
445 
446 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
447 		     (u32)((u64)(unsigned long)ds),
448 		     (u32)((u64)(unsigned long)ds >> 32));
449 }
450 
451 void fini_debug_store_on_cpu(int cpu)
452 {
453 	if (!per_cpu(cpu_hw_events, cpu).ds)
454 		return;
455 
456 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
457 }
458 
459 static DEFINE_PER_CPU(void *, insn_buffer);
460 
461 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
462 {
463 	unsigned long start = (unsigned long)cea;
464 	phys_addr_t pa;
465 	size_t msz = 0;
466 
467 	pa = virt_to_phys(addr);
468 
469 	preempt_disable();
470 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
471 		cea_set_pte(cea, pa, prot);
472 
473 	/*
474 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
475 	 * all TLB entries for it.
476 	 */
477 	flush_tlb_kernel_range(start, start + size);
478 	preempt_enable();
479 }
480 
481 static void ds_clear_cea(void *cea, size_t size)
482 {
483 	unsigned long start = (unsigned long)cea;
484 	size_t msz = 0;
485 
486 	preempt_disable();
487 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
488 		cea_set_pte(cea, 0, PAGE_NONE);
489 
490 	flush_tlb_kernel_range(start, start + size);
491 	preempt_enable();
492 }
493 
494 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
495 {
496 	unsigned int order = get_order(size);
497 	int node = cpu_to_node(cpu);
498 	struct page *page;
499 
500 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
501 	return page ? page_address(page) : NULL;
502 }
503 
504 static void dsfree_pages(const void *buffer, size_t size)
505 {
506 	if (buffer)
507 		free_pages((unsigned long)buffer, get_order(size));
508 }
509 
510 static int alloc_pebs_buffer(int cpu)
511 {
512 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
513 	struct debug_store *ds = hwev->ds;
514 	size_t bsiz = x86_pmu.pebs_buffer_size;
515 	int max, node = cpu_to_node(cpu);
516 	void *buffer, *insn_buff, *cea;
517 
518 	if (!x86_pmu.pebs)
519 		return 0;
520 
521 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
522 	if (unlikely(!buffer))
523 		return -ENOMEM;
524 
525 	/*
526 	 * HSW+ already provides us the eventing ip; no need to allocate this
527 	 * buffer then.
528 	 */
529 	if (x86_pmu.intel_cap.pebs_format < 2) {
530 		insn_buff = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
531 		if (!insn_buff) {
532 			dsfree_pages(buffer, bsiz);
533 			return -ENOMEM;
534 		}
535 		per_cpu(insn_buffer, cpu) = insn_buff;
536 	}
537 	hwev->ds_pebs_vaddr = buffer;
538 	/* Update the cpu entry area mapping */
539 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
540 	ds->pebs_buffer_base = (unsigned long) cea;
541 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
542 	ds->pebs_index = ds->pebs_buffer_base;
543 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
544 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
545 	return 0;
546 }
547 
548 static void release_pebs_buffer(int cpu)
549 {
550 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
551 	void *cea;
552 
553 	if (!x86_pmu.pebs)
554 		return;
555 
556 	kfree(per_cpu(insn_buffer, cpu));
557 	per_cpu(insn_buffer, cpu) = NULL;
558 
559 	/* Clear the fixmap */
560 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
561 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
562 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
563 	hwev->ds_pebs_vaddr = NULL;
564 }
565 
566 static int alloc_bts_buffer(int cpu)
567 {
568 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
569 	struct debug_store *ds = hwev->ds;
570 	void *buffer, *cea;
571 	int max;
572 
573 	if (!x86_pmu.bts)
574 		return 0;
575 
576 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
577 	if (unlikely(!buffer)) {
578 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
579 		return -ENOMEM;
580 	}
581 	hwev->ds_bts_vaddr = buffer;
582 	/* Update the fixmap */
583 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
584 	ds->bts_buffer_base = (unsigned long) cea;
585 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
586 	ds->bts_index = ds->bts_buffer_base;
587 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
588 	ds->bts_absolute_maximum = ds->bts_buffer_base +
589 					max * BTS_RECORD_SIZE;
590 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
591 					(max / 16) * BTS_RECORD_SIZE;
592 	return 0;
593 }
594 
595 static void release_bts_buffer(int cpu)
596 {
597 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
598 	void *cea;
599 
600 	if (!x86_pmu.bts)
601 		return;
602 
603 	/* Clear the fixmap */
604 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
605 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
606 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
607 	hwev->ds_bts_vaddr = NULL;
608 }
609 
610 static int alloc_ds_buffer(int cpu)
611 {
612 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
613 
614 	memset(ds, 0, sizeof(*ds));
615 	per_cpu(cpu_hw_events, cpu).ds = ds;
616 	return 0;
617 }
618 
619 static void release_ds_buffer(int cpu)
620 {
621 	per_cpu(cpu_hw_events, cpu).ds = NULL;
622 }
623 
624 void release_ds_buffers(void)
625 {
626 	int cpu;
627 
628 	if (!x86_pmu.bts && !x86_pmu.pebs)
629 		return;
630 
631 	for_each_possible_cpu(cpu)
632 		release_ds_buffer(cpu);
633 
634 	for_each_possible_cpu(cpu) {
635 		/*
636 		 * Again, ignore errors from offline CPUs, they will no longer
637 		 * observe cpu_hw_events.ds and not program the DS_AREA when
638 		 * they come up.
639 		 */
640 		fini_debug_store_on_cpu(cpu);
641 	}
642 
643 	for_each_possible_cpu(cpu) {
644 		release_pebs_buffer(cpu);
645 		release_bts_buffer(cpu);
646 	}
647 }
648 
649 void reserve_ds_buffers(void)
650 {
651 	int bts_err = 0, pebs_err = 0;
652 	int cpu;
653 
654 	x86_pmu.bts_active = 0;
655 	x86_pmu.pebs_active = 0;
656 
657 	if (!x86_pmu.bts && !x86_pmu.pebs)
658 		return;
659 
660 	if (!x86_pmu.bts)
661 		bts_err = 1;
662 
663 	if (!x86_pmu.pebs)
664 		pebs_err = 1;
665 
666 	for_each_possible_cpu(cpu) {
667 		if (alloc_ds_buffer(cpu)) {
668 			bts_err = 1;
669 			pebs_err = 1;
670 		}
671 
672 		if (!bts_err && alloc_bts_buffer(cpu))
673 			bts_err = 1;
674 
675 		if (!pebs_err && alloc_pebs_buffer(cpu))
676 			pebs_err = 1;
677 
678 		if (bts_err && pebs_err)
679 			break;
680 	}
681 
682 	if (bts_err) {
683 		for_each_possible_cpu(cpu)
684 			release_bts_buffer(cpu);
685 	}
686 
687 	if (pebs_err) {
688 		for_each_possible_cpu(cpu)
689 			release_pebs_buffer(cpu);
690 	}
691 
692 	if (bts_err && pebs_err) {
693 		for_each_possible_cpu(cpu)
694 			release_ds_buffer(cpu);
695 	} else {
696 		if (x86_pmu.bts && !bts_err)
697 			x86_pmu.bts_active = 1;
698 
699 		if (x86_pmu.pebs && !pebs_err)
700 			x86_pmu.pebs_active = 1;
701 
702 		for_each_possible_cpu(cpu) {
703 			/*
704 			 * Ignores wrmsr_on_cpu() errors for offline CPUs they
705 			 * will get this call through intel_pmu_cpu_starting().
706 			 */
707 			init_debug_store_on_cpu(cpu);
708 		}
709 	}
710 }
711 
712 /*
713  * BTS
714  */
715 
716 struct event_constraint bts_constraint =
717 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
718 
719 void intel_pmu_enable_bts(u64 config)
720 {
721 	unsigned long debugctlmsr;
722 
723 	debugctlmsr = get_debugctlmsr();
724 
725 	debugctlmsr |= DEBUGCTLMSR_TR;
726 	debugctlmsr |= DEBUGCTLMSR_BTS;
727 	if (config & ARCH_PERFMON_EVENTSEL_INT)
728 		debugctlmsr |= DEBUGCTLMSR_BTINT;
729 
730 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
731 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
732 
733 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
734 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
735 
736 	update_debugctlmsr(debugctlmsr);
737 }
738 
739 void intel_pmu_disable_bts(void)
740 {
741 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
742 	unsigned long debugctlmsr;
743 
744 	if (!cpuc->ds)
745 		return;
746 
747 	debugctlmsr = get_debugctlmsr();
748 
749 	debugctlmsr &=
750 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
751 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
752 
753 	update_debugctlmsr(debugctlmsr);
754 }
755 
756 int intel_pmu_drain_bts_buffer(void)
757 {
758 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
759 	struct debug_store *ds = cpuc->ds;
760 	struct bts_record {
761 		u64	from;
762 		u64	to;
763 		u64	flags;
764 	};
765 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
766 	struct bts_record *at, *base, *top;
767 	struct perf_output_handle handle;
768 	struct perf_event_header header;
769 	struct perf_sample_data data;
770 	unsigned long skip = 0;
771 	struct pt_regs regs;
772 
773 	if (!event)
774 		return 0;
775 
776 	if (!x86_pmu.bts_active)
777 		return 0;
778 
779 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
780 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
781 
782 	if (top <= base)
783 		return 0;
784 
785 	memset(&regs, 0, sizeof(regs));
786 
787 	ds->bts_index = ds->bts_buffer_base;
788 
789 	perf_sample_data_init(&data, 0, event->hw.last_period);
790 
791 	/*
792 	 * BTS leaks kernel addresses in branches across the cpl boundary,
793 	 * such as traps or system calls, so unless the user is asking for
794 	 * kernel tracing (and right now it's not possible), we'd need to
795 	 * filter them out. But first we need to count how many of those we
796 	 * have in the current batch. This is an extra O(n) pass, however,
797 	 * it's much faster than the other one especially considering that
798 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
799 	 * alloc_bts_buffer()).
800 	 */
801 	for (at = base; at < top; at++) {
802 		/*
803 		 * Note that right now *this* BTS code only works if
804 		 * attr::exclude_kernel is set, but let's keep this extra
805 		 * check here in case that changes.
806 		 */
807 		if (event->attr.exclude_kernel &&
808 		    (kernel_ip(at->from) || kernel_ip(at->to)))
809 			skip++;
810 	}
811 
812 	/*
813 	 * Prepare a generic sample, i.e. fill in the invariant fields.
814 	 * We will overwrite the from and to address before we output
815 	 * the sample.
816 	 */
817 	rcu_read_lock();
818 	perf_prepare_sample(&data, event, &regs);
819 	perf_prepare_header(&header, &data, event, &regs);
820 
821 	if (perf_output_begin(&handle, &data, event,
822 			      header.size * (top - base - skip)))
823 		goto unlock;
824 
825 	for (at = base; at < top; at++) {
826 		/* Filter out any records that contain kernel addresses. */
827 		if (event->attr.exclude_kernel &&
828 		    (kernel_ip(at->from) || kernel_ip(at->to)))
829 			continue;
830 
831 		data.ip		= at->from;
832 		data.addr	= at->to;
833 
834 		perf_output_sample(&handle, &header, &data, event);
835 	}
836 
837 	perf_output_end(&handle);
838 
839 	/* There's new data available. */
840 	event->hw.interrupts++;
841 	event->pending_kill = POLL_IN;
842 unlock:
843 	rcu_read_unlock();
844 	return 1;
845 }
846 
847 static inline void intel_pmu_drain_pebs_buffer(void)
848 {
849 	struct perf_sample_data data;
850 
851 	x86_pmu.drain_pebs(NULL, &data);
852 }
853 
854 /*
855  * PEBS
856  */
857 struct event_constraint intel_core2_pebs_event_constraints[] = {
858 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
859 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
860 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
861 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
862 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
863 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
864 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
865 	EVENT_CONSTRAINT_END
866 };
867 
868 struct event_constraint intel_atom_pebs_event_constraints[] = {
869 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
870 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
871 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
872 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
873 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
874 	/* Allow all events as PEBS with no flags */
875 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
876 	EVENT_CONSTRAINT_END
877 };
878 
879 struct event_constraint intel_slm_pebs_event_constraints[] = {
880 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
881 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
882 	/* Allow all events as PEBS with no flags */
883 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
884 	EVENT_CONSTRAINT_END
885 };
886 
887 struct event_constraint intel_glm_pebs_event_constraints[] = {
888 	/* Allow all events as PEBS with no flags */
889 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
890 	EVENT_CONSTRAINT_END
891 };
892 
893 struct event_constraint intel_grt_pebs_event_constraints[] = {
894 	/* Allow all events as PEBS with no flags */
895 	INTEL_HYBRID_LAT_CONSTRAINT(0x5d0, 0x3),
896 	INTEL_HYBRID_LAT_CONSTRAINT(0x6d0, 0xf),
897 	EVENT_CONSTRAINT_END
898 };
899 
900 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
901 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
902 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
903 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
904 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
905 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
906 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
907 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
908 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
909 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
910 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
911 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
912 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
913 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
914 	EVENT_CONSTRAINT_END
915 };
916 
917 struct event_constraint intel_westmere_pebs_event_constraints[] = {
918 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
919 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
920 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
921 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
922 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
923 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
924 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
925 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
926 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
927 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
928 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
929 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
930 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
931 	EVENT_CONSTRAINT_END
932 };
933 
934 struct event_constraint intel_snb_pebs_event_constraints[] = {
935 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
936 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
937 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
938 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
939 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
940         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
941         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
942         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
943         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
944 	/* Allow all events as PEBS with no flags */
945 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
946 	EVENT_CONSTRAINT_END
947 };
948 
949 struct event_constraint intel_ivb_pebs_event_constraints[] = {
950         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
951         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
952 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
953 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
954 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
955 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
956 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
957 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
958 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
959 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
960 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_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_hsw_pebs_event_constraints[] = {
967 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
968 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
969 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
970 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
971 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
972 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
973 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
974 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
975 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
976 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
977 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
978 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
979 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
980 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
981 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
982 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
983 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
984 	/* Allow all events as PEBS with no flags */
985 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
986 	EVENT_CONSTRAINT_END
987 };
988 
989 struct event_constraint intel_bdw_pebs_event_constraints[] = {
990 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
991 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
992 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
993 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
994 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
995 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
996 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
997 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
998 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
999 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
1000 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
1001 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
1002 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
1003 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
1004 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
1005 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
1006 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
1007 	/* Allow all events as PEBS with no flags */
1008 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1009 	EVENT_CONSTRAINT_END
1010 };
1011 
1012 
1013 struct event_constraint intel_skl_pebs_event_constraints[] = {
1014 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
1015 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
1016 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
1017 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
1018 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
1019 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
1020 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
1021 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
1022 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
1023 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
1024 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
1025 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
1026 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
1027 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
1028 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
1029 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
1030 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
1031 	/* Allow all events as PEBS with no flags */
1032 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
1033 	EVENT_CONSTRAINT_END
1034 };
1035 
1036 struct event_constraint intel_icl_pebs_event_constraints[] = {
1037 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x100000000ULL),	/* old INST_RETIRED.PREC_DIST */
1038 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1039 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),	/* SLOTS */
1040 
1041 	INTEL_PLD_CONSTRAINT(0x1cd, 0xff),			/* MEM_TRANS_RETIRED.LOAD_LATENCY */
1042 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1043 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1044 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1045 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1046 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1047 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1048 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1049 
1050 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf), /* MEM_LOAD_*_RETIRED.* */
1051 
1052 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),		/* MEM_INST_RETIRED.* */
1053 
1054 	/*
1055 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1056 	 * need the full constraints from the main table.
1057 	 */
1058 
1059 	EVENT_CONSTRAINT_END
1060 };
1061 
1062 struct event_constraint intel_glc_pebs_event_constraints[] = {
1063 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x100, 0x100000000ULL),	/* INST_RETIRED.PREC_DIST */
1064 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x0400, 0x800000000ULL),
1065 
1066 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xfe),
1067 	INTEL_PLD_CONSTRAINT(0x1cd, 0xfe),
1068 	INTEL_PSD_CONSTRAINT(0x2cd, 0x1),
1069 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_LOADS */
1070 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf),	/* MEM_INST_RETIRED.STLB_MISS_STORES */
1071 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf),	/* MEM_INST_RETIRED.LOCK_LOADS */
1072 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_LOADS */
1073 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf),	/* MEM_INST_RETIRED.SPLIT_STORES */
1074 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf),	/* MEM_INST_RETIRED.ALL_LOADS */
1075 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf),	/* MEM_INST_RETIRED.ALL_STORES */
1076 
1077 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD_RANGE(0xd1, 0xd4, 0xf),
1078 
1079 	INTEL_FLAGS_EVENT_CONSTRAINT(0xd0, 0xf),
1080 
1081 	/*
1082 	 * Everything else is handled by PMU_FL_PEBS_ALL, because we
1083 	 * need the full constraints from the main table.
1084 	 */
1085 
1086 	EVENT_CONSTRAINT_END
1087 };
1088 
1089 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
1090 {
1091 	struct event_constraint *pebs_constraints = hybrid(event->pmu, pebs_constraints);
1092 	struct event_constraint *c;
1093 
1094 	if (!event->attr.precise_ip)
1095 		return NULL;
1096 
1097 	if (pebs_constraints) {
1098 		for_each_event_constraint(c, pebs_constraints) {
1099 			if (constraint_match(c, event->hw.config)) {
1100 				event->hw.flags |= c->flags;
1101 				return c;
1102 			}
1103 		}
1104 	}
1105 
1106 	/*
1107 	 * Extended PEBS support
1108 	 * Makes the PEBS code search the normal constraints.
1109 	 */
1110 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1111 		return NULL;
1112 
1113 	return &emptyconstraint;
1114 }
1115 
1116 /*
1117  * We need the sched_task callback even for per-cpu events when we use
1118  * the large interrupt threshold, such that we can provide PID and TID
1119  * to PEBS samples.
1120  */
1121 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
1122 {
1123 	if (cpuc->n_pebs == cpuc->n_pebs_via_pt)
1124 		return false;
1125 
1126 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
1127 }
1128 
1129 void intel_pmu_pebs_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
1130 {
1131 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1132 
1133 	if (!sched_in && pebs_needs_sched_cb(cpuc))
1134 		intel_pmu_drain_pebs_buffer();
1135 }
1136 
1137 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
1138 {
1139 	struct debug_store *ds = cpuc->ds;
1140 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
1141 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
1142 	u64 threshold;
1143 	int reserved;
1144 
1145 	if (cpuc->n_pebs_via_pt)
1146 		return;
1147 
1148 	if (x86_pmu.flags & PMU_FL_PEBS_ALL)
1149 		reserved = max_pebs_events + num_counters_fixed;
1150 	else
1151 		reserved = max_pebs_events;
1152 
1153 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
1154 		threshold = ds->pebs_absolute_maximum -
1155 			reserved * cpuc->pebs_record_size;
1156 	} else {
1157 		threshold = ds->pebs_buffer_base + cpuc->pebs_record_size;
1158 	}
1159 
1160 	ds->pebs_interrupt_threshold = threshold;
1161 }
1162 
1163 static void adaptive_pebs_record_size_update(void)
1164 {
1165 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1166 	u64 pebs_data_cfg = cpuc->pebs_data_cfg;
1167 	int sz = sizeof(struct pebs_basic);
1168 
1169 	if (pebs_data_cfg & PEBS_DATACFG_MEMINFO)
1170 		sz += sizeof(struct pebs_meminfo);
1171 	if (pebs_data_cfg & PEBS_DATACFG_GP)
1172 		sz += sizeof(struct pebs_gprs);
1173 	if (pebs_data_cfg & PEBS_DATACFG_XMMS)
1174 		sz += sizeof(struct pebs_xmm);
1175 	if (pebs_data_cfg & PEBS_DATACFG_LBRS)
1176 		sz += x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1177 
1178 	cpuc->pebs_record_size = sz;
1179 }
1180 
1181 #define PERF_PEBS_MEMINFO_TYPE	(PERF_SAMPLE_ADDR | PERF_SAMPLE_DATA_SRC |   \
1182 				PERF_SAMPLE_PHYS_ADDR |			     \
1183 				PERF_SAMPLE_WEIGHT_TYPE |		     \
1184 				PERF_SAMPLE_TRANSACTION |		     \
1185 				PERF_SAMPLE_DATA_PAGE_SIZE)
1186 
1187 static u64 pebs_update_adaptive_cfg(struct perf_event *event)
1188 {
1189 	struct perf_event_attr *attr = &event->attr;
1190 	u64 sample_type = attr->sample_type;
1191 	u64 pebs_data_cfg = 0;
1192 	bool gprs, tsx_weight;
1193 
1194 	if (!(sample_type & ~(PERF_SAMPLE_IP|PERF_SAMPLE_TIME)) &&
1195 	    attr->precise_ip > 1)
1196 		return pebs_data_cfg;
1197 
1198 	if (sample_type & PERF_PEBS_MEMINFO_TYPE)
1199 		pebs_data_cfg |= PEBS_DATACFG_MEMINFO;
1200 
1201 	/*
1202 	 * We need GPRs when:
1203 	 * + user requested them
1204 	 * + precise_ip < 2 for the non event IP
1205 	 * + For RTM TSX weight we need GPRs for the abort code.
1206 	 */
1207 	gprs = (sample_type & PERF_SAMPLE_REGS_INTR) &&
1208 	       (attr->sample_regs_intr & PEBS_GP_REGS);
1209 
1210 	tsx_weight = (sample_type & PERF_SAMPLE_WEIGHT_TYPE) &&
1211 		     ((attr->config & INTEL_ARCH_EVENT_MASK) ==
1212 		      x86_pmu.rtm_abort_event);
1213 
1214 	if (gprs || (attr->precise_ip < 2) || tsx_weight)
1215 		pebs_data_cfg |= PEBS_DATACFG_GP;
1216 
1217 	if ((sample_type & PERF_SAMPLE_REGS_INTR) &&
1218 	    (attr->sample_regs_intr & PERF_REG_EXTENDED_MASK))
1219 		pebs_data_cfg |= PEBS_DATACFG_XMMS;
1220 
1221 	if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
1222 		/*
1223 		 * For now always log all LBRs. Could configure this
1224 		 * later.
1225 		 */
1226 		pebs_data_cfg |= PEBS_DATACFG_LBRS |
1227 			((x86_pmu.lbr_nr-1) << PEBS_DATACFG_LBR_SHIFT);
1228 	}
1229 
1230 	return pebs_data_cfg;
1231 }
1232 
1233 static void
1234 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc,
1235 		  struct perf_event *event, bool add)
1236 {
1237 	struct pmu *pmu = event->pmu;
1238 
1239 	/*
1240 	 * Make sure we get updated with the first PEBS
1241 	 * event. It will trigger also during removal, but
1242 	 * that does not hurt:
1243 	 */
1244 	if (cpuc->n_pebs == 1)
1245 		cpuc->pebs_data_cfg = PEBS_UPDATE_DS_SW;
1246 
1247 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
1248 		if (!needed_cb)
1249 			perf_sched_cb_inc(pmu);
1250 		else
1251 			perf_sched_cb_dec(pmu);
1252 
1253 		cpuc->pebs_data_cfg |= PEBS_UPDATE_DS_SW;
1254 	}
1255 
1256 	/*
1257 	 * The PEBS record doesn't shrink on pmu::del(). Doing so would require
1258 	 * iterating all remaining PEBS events to reconstruct the config.
1259 	 */
1260 	if (x86_pmu.intel_cap.pebs_baseline && add) {
1261 		u64 pebs_data_cfg;
1262 
1263 		pebs_data_cfg = pebs_update_adaptive_cfg(event);
1264 		/*
1265 		 * Be sure to update the thresholds when we change the record.
1266 		 */
1267 		if (pebs_data_cfg & ~cpuc->pebs_data_cfg)
1268 			cpuc->pebs_data_cfg |= pebs_data_cfg | PEBS_UPDATE_DS_SW;
1269 	}
1270 }
1271 
1272 void intel_pmu_pebs_add(struct perf_event *event)
1273 {
1274 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1275 	struct hw_perf_event *hwc = &event->hw;
1276 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1277 
1278 	cpuc->n_pebs++;
1279 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1280 		cpuc->n_large_pebs++;
1281 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1282 		cpuc->n_pebs_via_pt++;
1283 
1284 	pebs_update_state(needed_cb, cpuc, event, true);
1285 }
1286 
1287 static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
1288 {
1289 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1290 
1291 	if (!is_pebs_pt(event))
1292 		return;
1293 
1294 	if (!(cpuc->pebs_enabled & ~PEBS_VIA_PT_MASK))
1295 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
1296 }
1297 
1298 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
1299 {
1300 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1301 	struct hw_perf_event *hwc = &event->hw;
1302 	struct debug_store *ds = cpuc->ds;
1303 	u64 value = ds->pebs_event_reset[hwc->idx];
1304 	u32 base = MSR_RELOAD_PMC0;
1305 	unsigned int idx = hwc->idx;
1306 
1307 	if (!is_pebs_pt(event))
1308 		return;
1309 
1310 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
1311 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
1312 
1313 	cpuc->pebs_enabled |= PEBS_OUTPUT_PT;
1314 
1315 	if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
1316 		base = MSR_RELOAD_FIXED_CTR0;
1317 		idx = hwc->idx - INTEL_PMC_IDX_FIXED;
1318 		if (x86_pmu.intel_cap.pebs_format < 5)
1319 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS_FMT4 + idx];
1320 		else
1321 			value = ds->pebs_event_reset[MAX_PEBS_EVENTS + idx];
1322 	}
1323 	wrmsrl(base + idx, value);
1324 }
1325 
1326 static inline void intel_pmu_drain_large_pebs(struct cpu_hw_events *cpuc)
1327 {
1328 	if (cpuc->n_pebs == cpuc->n_large_pebs &&
1329 	    cpuc->n_pebs != cpuc->n_pebs_via_pt)
1330 		intel_pmu_drain_pebs_buffer();
1331 }
1332 
1333 void intel_pmu_pebs_enable(struct perf_event *event)
1334 {
1335 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1336 	u64 pebs_data_cfg = cpuc->pebs_data_cfg & ~PEBS_UPDATE_DS_SW;
1337 	struct hw_perf_event *hwc = &event->hw;
1338 	struct debug_store *ds = cpuc->ds;
1339 	unsigned int idx = hwc->idx;
1340 
1341 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
1342 
1343 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
1344 
1345 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) && (x86_pmu.version < 5))
1346 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
1347 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1348 		cpuc->pebs_enabled |= 1ULL << 63;
1349 
1350 	if (x86_pmu.intel_cap.pebs_baseline) {
1351 		hwc->config |= ICL_EVENTSEL_ADAPTIVE;
1352 		if (pebs_data_cfg != cpuc->active_pebs_data_cfg) {
1353 			/*
1354 			 * drain_pebs() assumes uniform record size;
1355 			 * hence we need to drain when changing said
1356 			 * size.
1357 			 */
1358 			intel_pmu_drain_large_pebs(cpuc);
1359 			adaptive_pebs_record_size_update();
1360 			wrmsrl(MSR_PEBS_DATA_CFG, pebs_data_cfg);
1361 			cpuc->active_pebs_data_cfg = pebs_data_cfg;
1362 		}
1363 	}
1364 	if (cpuc->pebs_data_cfg & PEBS_UPDATE_DS_SW) {
1365 		cpuc->pebs_data_cfg = pebs_data_cfg;
1366 		pebs_update_threshold(cpuc);
1367 	}
1368 
1369 	if (idx >= INTEL_PMC_IDX_FIXED) {
1370 		if (x86_pmu.intel_cap.pebs_format < 5)
1371 			idx = MAX_PEBS_EVENTS_FMT4 + (idx - INTEL_PMC_IDX_FIXED);
1372 		else
1373 			idx = MAX_PEBS_EVENTS + (idx - INTEL_PMC_IDX_FIXED);
1374 	}
1375 
1376 	/*
1377 	 * Use auto-reload if possible to save a MSR write in the PMI.
1378 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
1379 	 */
1380 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1381 		ds->pebs_event_reset[idx] =
1382 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
1383 	} else {
1384 		ds->pebs_event_reset[idx] = 0;
1385 	}
1386 
1387 	intel_pmu_pebs_via_pt_enable(event);
1388 }
1389 
1390 void intel_pmu_pebs_del(struct perf_event *event)
1391 {
1392 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1393 	struct hw_perf_event *hwc = &event->hw;
1394 	bool needed_cb = pebs_needs_sched_cb(cpuc);
1395 
1396 	cpuc->n_pebs--;
1397 	if (hwc->flags & PERF_X86_EVENT_LARGE_PEBS)
1398 		cpuc->n_large_pebs--;
1399 	if (hwc->flags & PERF_X86_EVENT_PEBS_VIA_PT)
1400 		cpuc->n_pebs_via_pt--;
1401 
1402 	pebs_update_state(needed_cb, cpuc, event, false);
1403 }
1404 
1405 void intel_pmu_pebs_disable(struct perf_event *event)
1406 {
1407 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1408 	struct hw_perf_event *hwc = &event->hw;
1409 
1410 	intel_pmu_drain_large_pebs(cpuc);
1411 
1412 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
1413 
1414 	if ((event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT) &&
1415 	    (x86_pmu.version < 5))
1416 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
1417 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
1418 		cpuc->pebs_enabled &= ~(1ULL << 63);
1419 
1420 	intel_pmu_pebs_via_pt_disable(event);
1421 
1422 	if (cpuc->enabled)
1423 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1424 
1425 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1426 }
1427 
1428 void intel_pmu_pebs_enable_all(void)
1429 {
1430 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1431 
1432 	if (cpuc->pebs_enabled)
1433 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1434 }
1435 
1436 void intel_pmu_pebs_disable_all(void)
1437 {
1438 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1439 
1440 	if (cpuc->pebs_enabled)
1441 		__intel_pmu_pebs_disable_all();
1442 }
1443 
1444 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1445 {
1446 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1447 	unsigned long from = cpuc->lbr_entries[0].from;
1448 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1449 	unsigned long ip = regs->ip;
1450 	int is_64bit = 0;
1451 	void *kaddr;
1452 	int size;
1453 
1454 	/*
1455 	 * We don't need to fixup if the PEBS assist is fault like
1456 	 */
1457 	if (!x86_pmu.intel_cap.pebs_trap)
1458 		return 1;
1459 
1460 	/*
1461 	 * No LBR entry, no basic block, no rewinding
1462 	 */
1463 	if (!cpuc->lbr_stack.nr || !from || !to)
1464 		return 0;
1465 
1466 	/*
1467 	 * Basic blocks should never cross user/kernel boundaries
1468 	 */
1469 	if (kernel_ip(ip) != kernel_ip(to))
1470 		return 0;
1471 
1472 	/*
1473 	 * unsigned math, either ip is before the start (impossible) or
1474 	 * the basic block is larger than 1 page (sanity)
1475 	 */
1476 	if ((ip - to) > PEBS_FIXUP_SIZE)
1477 		return 0;
1478 
1479 	/*
1480 	 * We sampled a branch insn, rewind using the LBR stack
1481 	 */
1482 	if (ip == to) {
1483 		set_linear_ip(regs, from);
1484 		return 1;
1485 	}
1486 
1487 	size = ip - to;
1488 	if (!kernel_ip(ip)) {
1489 		int bytes;
1490 		u8 *buf = this_cpu_read(insn_buffer);
1491 
1492 		/* 'size' must fit our buffer, see above */
1493 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1494 		if (bytes != 0)
1495 			return 0;
1496 
1497 		kaddr = buf;
1498 	} else {
1499 		kaddr = (void *)to;
1500 	}
1501 
1502 	do {
1503 		struct insn insn;
1504 
1505 		old_to = to;
1506 
1507 #ifdef CONFIG_X86_64
1508 		is_64bit = kernel_ip(to) || any_64bit_mode(regs);
1509 #endif
1510 		insn_init(&insn, kaddr, size, is_64bit);
1511 
1512 		/*
1513 		 * Make sure there was not a problem decoding the instruction.
1514 		 * This is doubly important because we have an infinite loop if
1515 		 * insn.length=0.
1516 		 */
1517 		if (insn_get_length(&insn))
1518 			break;
1519 
1520 		to += insn.length;
1521 		kaddr += insn.length;
1522 		size -= insn.length;
1523 	} while (to < ip);
1524 
1525 	if (to == ip) {
1526 		set_linear_ip(regs, old_to);
1527 		return 1;
1528 	}
1529 
1530 	/*
1531 	 * Even though we decoded the basic block, the instruction stream
1532 	 * never matched the given IP, either the TO or the IP got corrupted.
1533 	 */
1534 	return 0;
1535 }
1536 
1537 static inline u64 intel_get_tsx_weight(u64 tsx_tuning)
1538 {
1539 	if (tsx_tuning) {
1540 		union hsw_tsx_tuning tsx = { .value = tsx_tuning };
1541 		return tsx.cycles_last_block;
1542 	}
1543 	return 0;
1544 }
1545 
1546 static inline u64 intel_get_tsx_transaction(u64 tsx_tuning, u64 ax)
1547 {
1548 	u64 txn = (tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1549 
1550 	/* For RTM XABORTs also log the abort code from AX */
1551 	if ((txn & PERF_TXN_TRANSACTION) && (ax & 1))
1552 		txn |= ((ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1553 	return txn;
1554 }
1555 
1556 static inline u64 get_pebs_status(void *n)
1557 {
1558 	if (x86_pmu.intel_cap.pebs_format < 4)
1559 		return ((struct pebs_record_nhm *)n)->status;
1560 	return ((struct pebs_basic *)n)->applicable_counters;
1561 }
1562 
1563 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1564 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1565 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1566 		 PERF_X86_EVENT_PEBS_NA_HSW)
1567 
1568 static u64 get_data_src(struct perf_event *event, u64 aux)
1569 {
1570 	u64 val = PERF_MEM_NA;
1571 	int fl = event->hw.flags;
1572 	bool fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1573 
1574 	if (fl & PERF_X86_EVENT_PEBS_LDLAT)
1575 		val = load_latency_data(event, aux);
1576 	else if (fl & PERF_X86_EVENT_PEBS_STLAT)
1577 		val = store_latency_data(event, aux);
1578 	else if (fl & PERF_X86_EVENT_PEBS_LAT_HYBRID)
1579 		val = x86_pmu.pebs_latency_data(event, aux);
1580 	else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1581 		val = precise_datala_hsw(event, aux);
1582 	else if (fst)
1583 		val = precise_store_data(aux);
1584 	return val;
1585 }
1586 
1587 static void setup_pebs_time(struct perf_event *event,
1588 			    struct perf_sample_data *data,
1589 			    u64 tsc)
1590 {
1591 	/* Converting to a user-defined clock is not supported yet. */
1592 	if (event->attr.use_clockid != 0)
1593 		return;
1594 
1595 	/*
1596 	 * Doesn't support the conversion when the TSC is unstable.
1597 	 * The TSC unstable case is a corner case and very unlikely to
1598 	 * happen. If it happens, the TSC in a PEBS record will be
1599 	 * dropped and fall back to perf_event_clock().
1600 	 */
1601 	if (!using_native_sched_clock() || !sched_clock_stable())
1602 		return;
1603 
1604 	data->time = native_sched_clock_from_tsc(tsc) + __sched_clock_offset;
1605 	data->sample_flags |= PERF_SAMPLE_TIME;
1606 }
1607 
1608 #define PERF_SAMPLE_ADDR_TYPE	(PERF_SAMPLE_ADDR |		\
1609 				 PERF_SAMPLE_PHYS_ADDR |	\
1610 				 PERF_SAMPLE_DATA_PAGE_SIZE)
1611 
1612 static void setup_pebs_fixed_sample_data(struct perf_event *event,
1613 				   struct pt_regs *iregs, void *__pebs,
1614 				   struct perf_sample_data *data,
1615 				   struct pt_regs *regs)
1616 {
1617 	/*
1618 	 * We cast to the biggest pebs_record but are careful not to
1619 	 * unconditionally access the 'extra' entries.
1620 	 */
1621 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1622 	struct pebs_record_skl *pebs = __pebs;
1623 	u64 sample_type;
1624 	int fll;
1625 
1626 	if (pebs == NULL)
1627 		return;
1628 
1629 	sample_type = event->attr.sample_type;
1630 	fll = event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT;
1631 
1632 	perf_sample_data_init(data, 0, event->hw.last_period);
1633 
1634 	data->period = event->hw.last_period;
1635 
1636 	/*
1637 	 * Use latency for weight (only avail with PEBS-LL)
1638 	 */
1639 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT_TYPE)) {
1640 		data->weight.full = pebs->lat;
1641 		data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
1642 	}
1643 
1644 	/*
1645 	 * data.data_src encodes the data source
1646 	 */
1647 	if (sample_type & PERF_SAMPLE_DATA_SRC) {
1648 		data->data_src.val = get_data_src(event, pebs->dse);
1649 		data->sample_flags |= PERF_SAMPLE_DATA_SRC;
1650 	}
1651 
1652 	/*
1653 	 * We must however always use iregs for the unwinder to stay sane; the
1654 	 * record BP,SP,IP can point into thin air when the record is from a
1655 	 * previous PMI context or an (I)RET happened between the record and
1656 	 * PMI.
1657 	 */
1658 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1659 		perf_sample_save_callchain(data, event, iregs);
1660 
1661 	/*
1662 	 * We use the interrupt regs as a base because the PEBS record does not
1663 	 * contain a full regs set, specifically it seems to lack segment
1664 	 * descriptors, which get used by things like user_mode().
1665 	 *
1666 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1667 	 */
1668 	*regs = *iregs;
1669 
1670 	/*
1671 	 * Initialize regs_>flags from PEBS,
1672 	 * Clear exact bit (which uses x86 EFLAGS Reserved bit 3),
1673 	 * i.e., do not rely on it being zero:
1674 	 */
1675 	regs->flags = pebs->flags & ~PERF_EFLAGS_EXACT;
1676 
1677 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1678 		regs->ax = pebs->ax;
1679 		regs->bx = pebs->bx;
1680 		regs->cx = pebs->cx;
1681 		regs->dx = pebs->dx;
1682 		regs->si = pebs->si;
1683 		regs->di = pebs->di;
1684 
1685 		regs->bp = pebs->bp;
1686 		regs->sp = pebs->sp;
1687 
1688 #ifndef CONFIG_X86_32
1689 		regs->r8 = pebs->r8;
1690 		regs->r9 = pebs->r9;
1691 		regs->r10 = pebs->r10;
1692 		regs->r11 = pebs->r11;
1693 		regs->r12 = pebs->r12;
1694 		regs->r13 = pebs->r13;
1695 		regs->r14 = pebs->r14;
1696 		regs->r15 = pebs->r15;
1697 #endif
1698 	}
1699 
1700 	if (event->attr.precise_ip > 1) {
1701 		/*
1702 		 * Haswell and later processors have an 'eventing IP'
1703 		 * (real IP) which fixes the off-by-1 skid in hardware.
1704 		 * Use it when precise_ip >= 2 :
1705 		 */
1706 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1707 			set_linear_ip(regs, pebs->real_ip);
1708 			regs->flags |= PERF_EFLAGS_EXACT;
1709 		} else {
1710 			/* Otherwise, use PEBS off-by-1 IP: */
1711 			set_linear_ip(regs, pebs->ip);
1712 
1713 			/*
1714 			 * With precise_ip >= 2, try to fix up the off-by-1 IP
1715 			 * using the LBR. If successful, the fixup function
1716 			 * corrects regs->ip and calls set_linear_ip() on regs:
1717 			 */
1718 			if (intel_pmu_pebs_fixup_ip(regs))
1719 				regs->flags |= PERF_EFLAGS_EXACT;
1720 		}
1721 	} else {
1722 		/*
1723 		 * When precise_ip == 1, return the PEBS off-by-1 IP,
1724 		 * no fixup attempted:
1725 		 */
1726 		set_linear_ip(regs, pebs->ip);
1727 	}
1728 
1729 
1730 	if ((sample_type & PERF_SAMPLE_ADDR_TYPE) &&
1731 	    x86_pmu.intel_cap.pebs_format >= 1) {
1732 		data->addr = pebs->dla;
1733 		data->sample_flags |= PERF_SAMPLE_ADDR;
1734 	}
1735 
1736 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1737 		/* Only set the TSX weight when no memory weight. */
1738 		if ((sample_type & PERF_SAMPLE_WEIGHT_TYPE) && !fll) {
1739 			data->weight.full = intel_get_tsx_weight(pebs->tsx_tuning);
1740 			data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
1741 		}
1742 		if (sample_type & PERF_SAMPLE_TRANSACTION) {
1743 			data->txn = intel_get_tsx_transaction(pebs->tsx_tuning,
1744 							      pebs->ax);
1745 			data->sample_flags |= PERF_SAMPLE_TRANSACTION;
1746 		}
1747 	}
1748 
1749 	/*
1750 	 * v3 supplies an accurate time stamp, so we use that
1751 	 * for the time stamp.
1752 	 *
1753 	 * We can only do this for the default trace clock.
1754 	 */
1755 	if (x86_pmu.intel_cap.pebs_format >= 3)
1756 		setup_pebs_time(event, data, pebs->tsc);
1757 
1758 	if (has_branch_stack(event))
1759 		perf_sample_save_brstack(data, event, &cpuc->lbr_stack, NULL);
1760 }
1761 
1762 static void adaptive_pebs_save_regs(struct pt_regs *regs,
1763 				    struct pebs_gprs *gprs)
1764 {
1765 	regs->ax = gprs->ax;
1766 	regs->bx = gprs->bx;
1767 	regs->cx = gprs->cx;
1768 	regs->dx = gprs->dx;
1769 	regs->si = gprs->si;
1770 	regs->di = gprs->di;
1771 	regs->bp = gprs->bp;
1772 	regs->sp = gprs->sp;
1773 #ifndef CONFIG_X86_32
1774 	regs->r8 = gprs->r8;
1775 	regs->r9 = gprs->r9;
1776 	regs->r10 = gprs->r10;
1777 	regs->r11 = gprs->r11;
1778 	regs->r12 = gprs->r12;
1779 	regs->r13 = gprs->r13;
1780 	regs->r14 = gprs->r14;
1781 	regs->r15 = gprs->r15;
1782 #endif
1783 }
1784 
1785 #define PEBS_LATENCY_MASK			0xffff
1786 #define PEBS_CACHE_LATENCY_OFFSET		32
1787 #define PEBS_RETIRE_LATENCY_OFFSET		32
1788 
1789 /*
1790  * With adaptive PEBS the layout depends on what fields are configured.
1791  */
1792 
1793 static void setup_pebs_adaptive_sample_data(struct perf_event *event,
1794 					    struct pt_regs *iregs, void *__pebs,
1795 					    struct perf_sample_data *data,
1796 					    struct pt_regs *regs)
1797 {
1798 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1799 	struct pebs_basic *basic = __pebs;
1800 	void *next_record = basic + 1;
1801 	u64 sample_type;
1802 	u64 format_size;
1803 	struct pebs_meminfo *meminfo = NULL;
1804 	struct pebs_gprs *gprs = NULL;
1805 	struct x86_perf_regs *perf_regs;
1806 
1807 	if (basic == NULL)
1808 		return;
1809 
1810 	perf_regs = container_of(regs, struct x86_perf_regs, regs);
1811 	perf_regs->xmm_regs = NULL;
1812 
1813 	sample_type = event->attr.sample_type;
1814 	format_size = basic->format_size;
1815 	perf_sample_data_init(data, 0, event->hw.last_period);
1816 	data->period = event->hw.last_period;
1817 
1818 	setup_pebs_time(event, data, basic->tsc);
1819 
1820 	/*
1821 	 * We must however always use iregs for the unwinder to stay sane; the
1822 	 * record BP,SP,IP can point into thin air when the record is from a
1823 	 * previous PMI context or an (I)RET happened between the record and
1824 	 * PMI.
1825 	 */
1826 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
1827 		perf_sample_save_callchain(data, event, iregs);
1828 
1829 	*regs = *iregs;
1830 	/* The ip in basic is EventingIP */
1831 	set_linear_ip(regs, basic->ip);
1832 	regs->flags = PERF_EFLAGS_EXACT;
1833 
1834 	if ((sample_type & PERF_SAMPLE_WEIGHT_STRUCT) && (x86_pmu.flags & PMU_FL_RETIRE_LATENCY))
1835 		data->weight.var3_w = format_size >> PEBS_RETIRE_LATENCY_OFFSET & PEBS_LATENCY_MASK;
1836 
1837 	/*
1838 	 * The record for MEMINFO is in front of GP
1839 	 * But PERF_SAMPLE_TRANSACTION needs gprs->ax.
1840 	 * Save the pointer here but process later.
1841 	 */
1842 	if (format_size & PEBS_DATACFG_MEMINFO) {
1843 		meminfo = next_record;
1844 		next_record = meminfo + 1;
1845 	}
1846 
1847 	if (format_size & PEBS_DATACFG_GP) {
1848 		gprs = next_record;
1849 		next_record = gprs + 1;
1850 
1851 		if (event->attr.precise_ip < 2) {
1852 			set_linear_ip(regs, gprs->ip);
1853 			regs->flags &= ~PERF_EFLAGS_EXACT;
1854 		}
1855 
1856 		if (sample_type & PERF_SAMPLE_REGS_INTR)
1857 			adaptive_pebs_save_regs(regs, gprs);
1858 	}
1859 
1860 	if (format_size & PEBS_DATACFG_MEMINFO) {
1861 		if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
1862 			u64 weight = meminfo->latency;
1863 
1864 			if (x86_pmu.flags & PMU_FL_INSTR_LATENCY) {
1865 				data->weight.var2_w = weight & PEBS_LATENCY_MASK;
1866 				weight >>= PEBS_CACHE_LATENCY_OFFSET;
1867 			}
1868 
1869 			/*
1870 			 * Although meminfo::latency is defined as a u64,
1871 			 * only the lower 32 bits include the valid data
1872 			 * in practice on Ice Lake and earlier platforms.
1873 			 */
1874 			if (sample_type & PERF_SAMPLE_WEIGHT) {
1875 				data->weight.full = weight ?:
1876 					intel_get_tsx_weight(meminfo->tsx_tuning);
1877 			} else {
1878 				data->weight.var1_dw = (u32)(weight & PEBS_LATENCY_MASK) ?:
1879 					intel_get_tsx_weight(meminfo->tsx_tuning);
1880 			}
1881 			data->sample_flags |= PERF_SAMPLE_WEIGHT_TYPE;
1882 		}
1883 
1884 		if (sample_type & PERF_SAMPLE_DATA_SRC) {
1885 			data->data_src.val = get_data_src(event, meminfo->aux);
1886 			data->sample_flags |= PERF_SAMPLE_DATA_SRC;
1887 		}
1888 
1889 		if (sample_type & PERF_SAMPLE_ADDR_TYPE) {
1890 			data->addr = meminfo->address;
1891 			data->sample_flags |= PERF_SAMPLE_ADDR;
1892 		}
1893 
1894 		if (sample_type & PERF_SAMPLE_TRANSACTION) {
1895 			data->txn = intel_get_tsx_transaction(meminfo->tsx_tuning,
1896 							  gprs ? gprs->ax : 0);
1897 			data->sample_flags |= PERF_SAMPLE_TRANSACTION;
1898 		}
1899 	}
1900 
1901 	if (format_size & PEBS_DATACFG_XMMS) {
1902 		struct pebs_xmm *xmm = next_record;
1903 
1904 		next_record = xmm + 1;
1905 		perf_regs->xmm_regs = xmm->xmm;
1906 	}
1907 
1908 	if (format_size & PEBS_DATACFG_LBRS) {
1909 		struct lbr_entry *lbr = next_record;
1910 		int num_lbr = ((format_size >> PEBS_DATACFG_LBR_SHIFT)
1911 					& 0xff) + 1;
1912 		next_record = next_record + num_lbr * sizeof(struct lbr_entry);
1913 
1914 		if (has_branch_stack(event)) {
1915 			intel_pmu_store_pebs_lbrs(lbr);
1916 			intel_pmu_lbr_save_brstack(data, cpuc, event);
1917 		}
1918 	}
1919 
1920 	WARN_ONCE(next_record != __pebs + (format_size >> 48),
1921 			"PEBS record size %llu, expected %llu, config %llx\n",
1922 			format_size >> 48,
1923 			(u64)(next_record - __pebs),
1924 			basic->format_size);
1925 }
1926 
1927 static inline void *
1928 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1929 {
1930 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1931 	void *at;
1932 	u64 pebs_status;
1933 
1934 	/*
1935 	 * fmt0 does not have a status bitfield (does not use
1936 	 * perf_record_nhm format)
1937 	 */
1938 	if (x86_pmu.intel_cap.pebs_format < 1)
1939 		return base;
1940 
1941 	if (base == NULL)
1942 		return NULL;
1943 
1944 	for (at = base; at < top; at += cpuc->pebs_record_size) {
1945 		unsigned long status = get_pebs_status(at);
1946 
1947 		if (test_bit(bit, (unsigned long *)&status)) {
1948 			/* PEBS v3 has accurate status bits */
1949 			if (x86_pmu.intel_cap.pebs_format >= 3)
1950 				return at;
1951 
1952 			if (status == (1 << bit))
1953 				return at;
1954 
1955 			/* clear non-PEBS bit and re-check */
1956 			pebs_status = status & cpuc->pebs_enabled;
1957 			pebs_status &= PEBS_COUNTER_MASK;
1958 			if (pebs_status == (1 << bit))
1959 				return at;
1960 		}
1961 	}
1962 	return NULL;
1963 }
1964 
1965 void intel_pmu_auto_reload_read(struct perf_event *event)
1966 {
1967 	WARN_ON(!(event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD));
1968 
1969 	perf_pmu_disable(event->pmu);
1970 	intel_pmu_drain_pebs_buffer();
1971 	perf_pmu_enable(event->pmu);
1972 }
1973 
1974 /*
1975  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1976  */
1977 static int
1978 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1979 {
1980 	struct hw_perf_event *hwc = &event->hw;
1981 	int shift = 64 - x86_pmu.cntval_bits;
1982 	u64 period = hwc->sample_period;
1983 	u64 prev_raw_count, new_raw_count;
1984 	s64 new, old;
1985 
1986 	WARN_ON(!period);
1987 
1988 	/*
1989 	 * drain_pebs() only happens when the PMU is disabled.
1990 	 */
1991 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1992 
1993 	prev_raw_count = local64_read(&hwc->prev_count);
1994 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1995 	local64_set(&hwc->prev_count, new_raw_count);
1996 
1997 	/*
1998 	 * Since the counter increments a negative counter value and
1999 	 * overflows on the sign switch, giving the interval:
2000 	 *
2001 	 *   [-period, 0]
2002 	 *
2003 	 * the difference between two consecutive reads is:
2004 	 *
2005 	 *   A) value2 - value1;
2006 	 *      when no overflows have happened in between,
2007 	 *
2008 	 *   B) (0 - value1) + (value2 - (-period));
2009 	 *      when one overflow happened in between,
2010 	 *
2011 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
2012 	 *      when @n overflows happened in between.
2013 	 *
2014 	 * Here A) is the obvious difference, B) is the extension to the
2015 	 * discrete interval, where the first term is to the top of the
2016 	 * interval and the second term is from the bottom of the next
2017 	 * interval and C) the extension to multiple intervals, where the
2018 	 * middle term is the whole intervals covered.
2019 	 *
2020 	 * An equivalent of C, by reduction, is:
2021 	 *
2022 	 *   value2 - value1 + n * period
2023 	 */
2024 	new = ((s64)(new_raw_count << shift) >> shift);
2025 	old = ((s64)(prev_raw_count << shift) >> shift);
2026 	local64_add(new - old + count * period, &event->count);
2027 
2028 	local64_set(&hwc->period_left, -new);
2029 
2030 	perf_event_update_userpage(event);
2031 
2032 	return 0;
2033 }
2034 
2035 static __always_inline void
2036 __intel_pmu_pebs_event(struct perf_event *event,
2037 		       struct pt_regs *iregs,
2038 		       struct perf_sample_data *data,
2039 		       void *base, void *top,
2040 		       int bit, int count,
2041 		       void (*setup_sample)(struct perf_event *,
2042 					    struct pt_regs *,
2043 					    void *,
2044 					    struct perf_sample_data *,
2045 					    struct pt_regs *))
2046 {
2047 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2048 	struct hw_perf_event *hwc = &event->hw;
2049 	struct x86_perf_regs perf_regs;
2050 	struct pt_regs *regs = &perf_regs.regs;
2051 	void *at = get_next_pebs_record_by_bit(base, top, bit);
2052 	static struct pt_regs dummy_iregs;
2053 
2054 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
2055 		/*
2056 		 * Now, auto-reload is only enabled in fixed period mode.
2057 		 * The reload value is always hwc->sample_period.
2058 		 * May need to change it, if auto-reload is enabled in
2059 		 * freq mode later.
2060 		 */
2061 		intel_pmu_save_and_restart_reload(event, count);
2062 	} else if (!intel_pmu_save_and_restart(event))
2063 		return;
2064 
2065 	if (!iregs)
2066 		iregs = &dummy_iregs;
2067 
2068 	while (count > 1) {
2069 		setup_sample(event, iregs, at, data, regs);
2070 		perf_event_output(event, data, regs);
2071 		at += cpuc->pebs_record_size;
2072 		at = get_next_pebs_record_by_bit(at, top, bit);
2073 		count--;
2074 	}
2075 
2076 	setup_sample(event, iregs, at, data, regs);
2077 	if (iregs == &dummy_iregs) {
2078 		/*
2079 		 * The PEBS records may be drained in the non-overflow context,
2080 		 * e.g., large PEBS + context switch. Perf should treat the
2081 		 * last record the same as other PEBS records, and doesn't
2082 		 * invoke the generic overflow handler.
2083 		 */
2084 		perf_event_output(event, data, regs);
2085 	} else {
2086 		/*
2087 		 * All but the last records are processed.
2088 		 * The last one is left to be able to call the overflow handler.
2089 		 */
2090 		if (perf_event_overflow(event, data, regs))
2091 			x86_pmu_stop(event, 0);
2092 	}
2093 }
2094 
2095 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs, struct perf_sample_data *data)
2096 {
2097 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2098 	struct debug_store *ds = cpuc->ds;
2099 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
2100 	struct pebs_record_core *at, *top;
2101 	int n;
2102 
2103 	if (!x86_pmu.pebs_active)
2104 		return;
2105 
2106 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
2107 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
2108 
2109 	/*
2110 	 * Whatever else happens, drain the thing
2111 	 */
2112 	ds->pebs_index = ds->pebs_buffer_base;
2113 
2114 	if (!test_bit(0, cpuc->active_mask))
2115 		return;
2116 
2117 	WARN_ON_ONCE(!event);
2118 
2119 	if (!event->attr.precise_ip)
2120 		return;
2121 
2122 	n = top - at;
2123 	if (n <= 0) {
2124 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2125 			intel_pmu_save_and_restart_reload(event, 0);
2126 		return;
2127 	}
2128 
2129 	__intel_pmu_pebs_event(event, iregs, data, at, top, 0, n,
2130 			       setup_pebs_fixed_sample_data);
2131 }
2132 
2133 static void intel_pmu_pebs_event_update_no_drain(struct cpu_hw_events *cpuc, int size)
2134 {
2135 	struct perf_event *event;
2136 	int bit;
2137 
2138 	/*
2139 	 * The drain_pebs() could be called twice in a short period
2140 	 * for auto-reload event in pmu::read(). There are no
2141 	 * overflows have happened in between.
2142 	 * It needs to call intel_pmu_save_and_restart_reload() to
2143 	 * update the event->count for this case.
2144 	 */
2145 	for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled, size) {
2146 		event = cpuc->events[bit];
2147 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
2148 			intel_pmu_save_and_restart_reload(event, 0);
2149 	}
2150 }
2151 
2152 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs, struct perf_sample_data *data)
2153 {
2154 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2155 	struct debug_store *ds = cpuc->ds;
2156 	struct perf_event *event;
2157 	void *base, *at, *top;
2158 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2159 	short error[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2160 	int bit, i, size;
2161 	u64 mask;
2162 
2163 	if (!x86_pmu.pebs_active)
2164 		return;
2165 
2166 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
2167 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
2168 
2169 	ds->pebs_index = ds->pebs_buffer_base;
2170 
2171 	mask = (1ULL << x86_pmu.max_pebs_events) - 1;
2172 	size = x86_pmu.max_pebs_events;
2173 	if (x86_pmu.flags & PMU_FL_PEBS_ALL) {
2174 		mask |= ((1ULL << x86_pmu.num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED;
2175 		size = INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed;
2176 	}
2177 
2178 	if (unlikely(base >= top)) {
2179 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2180 		return;
2181 	}
2182 
2183 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
2184 		struct pebs_record_nhm *p = at;
2185 		u64 pebs_status;
2186 
2187 		pebs_status = p->status & cpuc->pebs_enabled;
2188 		pebs_status &= mask;
2189 
2190 		/* PEBS v3 has more accurate status bits */
2191 		if (x86_pmu.intel_cap.pebs_format >= 3) {
2192 			for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2193 				counts[bit]++;
2194 
2195 			continue;
2196 		}
2197 
2198 		/*
2199 		 * On some CPUs the PEBS status can be zero when PEBS is
2200 		 * racing with clearing of GLOBAL_STATUS.
2201 		 *
2202 		 * Normally we would drop that record, but in the
2203 		 * case when there is only a single active PEBS event
2204 		 * we can assume it's for that event.
2205 		 */
2206 		if (!pebs_status && cpuc->pebs_enabled &&
2207 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
2208 			pebs_status = p->status = cpuc->pebs_enabled;
2209 
2210 		bit = find_first_bit((unsigned long *)&pebs_status,
2211 					x86_pmu.max_pebs_events);
2212 		if (bit >= x86_pmu.max_pebs_events)
2213 			continue;
2214 
2215 		/*
2216 		 * The PEBS hardware does not deal well with the situation
2217 		 * when events happen near to each other and multiple bits
2218 		 * are set. But it should happen rarely.
2219 		 *
2220 		 * If these events include one PEBS and multiple non-PEBS
2221 		 * events, it doesn't impact PEBS record. The record will
2222 		 * be handled normally. (slow path)
2223 		 *
2224 		 * If these events include two or more PEBS events, the
2225 		 * records for the events can be collapsed into a single
2226 		 * one, and it's not possible to reconstruct all events
2227 		 * that caused the PEBS record. It's called collision.
2228 		 * If collision happened, the record will be dropped.
2229 		 */
2230 		if (pebs_status != (1ULL << bit)) {
2231 			for_each_set_bit(i, (unsigned long *)&pebs_status, size)
2232 				error[i]++;
2233 			continue;
2234 		}
2235 
2236 		counts[bit]++;
2237 	}
2238 
2239 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2240 		if ((counts[bit] == 0) && (error[bit] == 0))
2241 			continue;
2242 
2243 		event = cpuc->events[bit];
2244 		if (WARN_ON_ONCE(!event))
2245 			continue;
2246 
2247 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2248 			continue;
2249 
2250 		/* log dropped samples number */
2251 		if (error[bit]) {
2252 			perf_log_lost_samples(event, error[bit]);
2253 
2254 			if (iregs && perf_event_account_interrupt(event))
2255 				x86_pmu_stop(event, 0);
2256 		}
2257 
2258 		if (counts[bit]) {
2259 			__intel_pmu_pebs_event(event, iregs, data, base,
2260 					       top, bit, counts[bit],
2261 					       setup_pebs_fixed_sample_data);
2262 		}
2263 	}
2264 }
2265 
2266 static void intel_pmu_drain_pebs_icl(struct pt_regs *iregs, struct perf_sample_data *data)
2267 {
2268 	short counts[INTEL_PMC_IDX_FIXED + MAX_FIXED_PEBS_EVENTS] = {};
2269 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
2270 	int max_pebs_events = hybrid(cpuc->pmu, max_pebs_events);
2271 	int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed);
2272 	struct debug_store *ds = cpuc->ds;
2273 	struct perf_event *event;
2274 	void *base, *at, *top;
2275 	int bit, size;
2276 	u64 mask;
2277 
2278 	if (!x86_pmu.pebs_active)
2279 		return;
2280 
2281 	base = (struct pebs_basic *)(unsigned long)ds->pebs_buffer_base;
2282 	top = (struct pebs_basic *)(unsigned long)ds->pebs_index;
2283 
2284 	ds->pebs_index = ds->pebs_buffer_base;
2285 
2286 	mask = ((1ULL << max_pebs_events) - 1) |
2287 	       (((1ULL << num_counters_fixed) - 1) << INTEL_PMC_IDX_FIXED);
2288 	size = INTEL_PMC_IDX_FIXED + num_counters_fixed;
2289 
2290 	if (unlikely(base >= top)) {
2291 		intel_pmu_pebs_event_update_no_drain(cpuc, size);
2292 		return;
2293 	}
2294 
2295 	for (at = base; at < top; at += cpuc->pebs_record_size) {
2296 		u64 pebs_status;
2297 
2298 		pebs_status = get_pebs_status(at) & cpuc->pebs_enabled;
2299 		pebs_status &= mask;
2300 
2301 		for_each_set_bit(bit, (unsigned long *)&pebs_status, size)
2302 			counts[bit]++;
2303 	}
2304 
2305 	for_each_set_bit(bit, (unsigned long *)&mask, size) {
2306 		if (counts[bit] == 0)
2307 			continue;
2308 
2309 		event = cpuc->events[bit];
2310 		if (WARN_ON_ONCE(!event))
2311 			continue;
2312 
2313 		if (WARN_ON_ONCE(!event->attr.precise_ip))
2314 			continue;
2315 
2316 		__intel_pmu_pebs_event(event, iregs, data, base,
2317 				       top, bit, counts[bit],
2318 				       setup_pebs_adaptive_sample_data);
2319 	}
2320 }
2321 
2322 /*
2323  * BTS, PEBS probe and setup
2324  */
2325 
2326 void __init intel_ds_init(void)
2327 {
2328 	/*
2329 	 * No support for 32bit formats
2330 	 */
2331 	if (!boot_cpu_has(X86_FEATURE_DTES64))
2332 		return;
2333 
2334 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
2335 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
2336 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
2337 	if (x86_pmu.version <= 4)
2338 		x86_pmu.pebs_no_isolation = 1;
2339 
2340 	if (x86_pmu.pebs) {
2341 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
2342 		char *pebs_qual = "";
2343 		int format = x86_pmu.intel_cap.pebs_format;
2344 
2345 		if (format < 4)
2346 			x86_pmu.intel_cap.pebs_baseline = 0;
2347 
2348 		switch (format) {
2349 		case 0:
2350 			pr_cont("PEBS fmt0%c, ", pebs_type);
2351 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
2352 			/*
2353 			 * Using >PAGE_SIZE buffers makes the WRMSR to
2354 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
2355 			 * mysteriously hang on Core2.
2356 			 *
2357 			 * As a workaround, we don't do this.
2358 			 */
2359 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
2360 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
2361 			break;
2362 
2363 		case 1:
2364 			pr_cont("PEBS fmt1%c, ", pebs_type);
2365 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
2366 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2367 			break;
2368 
2369 		case 2:
2370 			pr_cont("PEBS fmt2%c, ", pebs_type);
2371 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
2372 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2373 			break;
2374 
2375 		case 3:
2376 			pr_cont("PEBS fmt3%c, ", pebs_type);
2377 			x86_pmu.pebs_record_size =
2378 						sizeof(struct pebs_record_skl);
2379 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
2380 			x86_pmu.large_pebs_flags |= PERF_SAMPLE_TIME;
2381 			break;
2382 
2383 		case 5:
2384 			x86_pmu.pebs_ept = 1;
2385 			fallthrough;
2386 		case 4:
2387 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_icl;
2388 			x86_pmu.pebs_record_size = sizeof(struct pebs_basic);
2389 			if (x86_pmu.intel_cap.pebs_baseline) {
2390 				x86_pmu.large_pebs_flags |=
2391 					PERF_SAMPLE_BRANCH_STACK |
2392 					PERF_SAMPLE_TIME;
2393 				x86_pmu.flags |= PMU_FL_PEBS_ALL;
2394 				x86_pmu.pebs_capable = ~0ULL;
2395 				pebs_qual = "-baseline";
2396 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_EXTENDED_REGS;
2397 			} else {
2398 				/* Only basic record supported */
2399 				x86_pmu.large_pebs_flags &=
2400 					~(PERF_SAMPLE_ADDR |
2401 					  PERF_SAMPLE_TIME |
2402 					  PERF_SAMPLE_DATA_SRC |
2403 					  PERF_SAMPLE_TRANSACTION |
2404 					  PERF_SAMPLE_REGS_USER |
2405 					  PERF_SAMPLE_REGS_INTR);
2406 			}
2407 			pr_cont("PEBS fmt4%c%s, ", pebs_type, pebs_qual);
2408 
2409 			if (!is_hybrid() && x86_pmu.intel_cap.pebs_output_pt_available) {
2410 				pr_cont("PEBS-via-PT, ");
2411 				x86_get_pmu(smp_processor_id())->capabilities |= PERF_PMU_CAP_AUX_OUTPUT;
2412 			}
2413 
2414 			break;
2415 
2416 		default:
2417 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
2418 			x86_pmu.pebs = 0;
2419 		}
2420 	}
2421 }
2422 
2423 void perf_restore_debug_store(void)
2424 {
2425 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
2426 
2427 	if (!x86_pmu.bts && !x86_pmu.pebs)
2428 		return;
2429 
2430 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
2431 }
2432