• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bitops.h>
3 #include <linux/types.h>
4 #include <linux/slab.h>
5 
6 #include <asm/cpu_entry_area.h>
7 #include <asm/perf_event.h>
8 #include <asm/tlbflush.h>
9 #include <asm/insn.h>
10 
11 #include "../perf_event.h"
12 
13 /* Waste a full page so it can be mapped into the cpu_entry_area */
14 DEFINE_PER_CPU_PAGE_ALIGNED(struct debug_store, cpu_debug_store);
15 
16 /* The size of a BTS record in bytes: */
17 #define BTS_RECORD_SIZE		24
18 
19 #define PEBS_FIXUP_SIZE		PAGE_SIZE
20 
21 /*
22  * pebs_record_32 for p4 and core not supported
23 
24 struct pebs_record_32 {
25 	u32 flags, ip;
26 	u32 ax, bc, cx, dx;
27 	u32 si, di, bp, sp;
28 };
29 
30  */
31 
32 union intel_x86_pebs_dse {
33 	u64 val;
34 	struct {
35 		unsigned int ld_dse:4;
36 		unsigned int ld_stlb_miss:1;
37 		unsigned int ld_locked:1;
38 		unsigned int ld_reserved:26;
39 	};
40 	struct {
41 		unsigned int st_l1d_hit:1;
42 		unsigned int st_reserved1:3;
43 		unsigned int st_stlb_miss:1;
44 		unsigned int st_locked:1;
45 		unsigned int st_reserved2:26;
46 	};
47 };
48 
49 
50 /*
51  * Map PEBS Load Latency Data Source encodings to generic
52  * memory data source information
53  */
54 #define P(a, b) PERF_MEM_S(a, b)
55 #define OP_LH (P(OP, LOAD) | P(LVL, HIT))
56 #define LEVEL(x) P(LVLNUM, x)
57 #define REM P(REMOTE, REMOTE)
58 #define SNOOP_NONE_MISS (P(SNOOP, NONE) | P(SNOOP, MISS))
59 
60 /* Version for Sandy Bridge and later */
61 static u64 pebs_data_source[] = {
62 	P(OP, LOAD) | P(LVL, MISS) | LEVEL(L3) | P(SNOOP, NA),/* 0x00:ukn L3 */
63 	OP_LH | P(LVL, L1)  | LEVEL(L1) | P(SNOOP, NONE),  /* 0x01: L1 local */
64 	OP_LH | P(LVL, LFB) | LEVEL(LFB) | P(SNOOP, NONE), /* 0x02: LFB hit */
65 	OP_LH | P(LVL, L2)  | LEVEL(L2) | P(SNOOP, NONE),  /* 0x03: L2 hit */
66 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, NONE),  /* 0x04: L3 hit */
67 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, MISS),  /* 0x05: L3 hit, snoop miss */
68 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HIT),   /* 0x06: L3 hit, snoop hit */
69 	OP_LH | P(LVL, L3)  | LEVEL(L3) | P(SNOOP, HITM),  /* 0x07: L3 hit, snoop hitm */
70 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x08: L3 miss snoop hit */
71 	OP_LH | P(LVL, REM_CCE1) | REM | LEVEL(L3) | P(SNOOP, HITM), /* 0x09: L3 miss snoop hitm*/
72 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | P(SNOOP, HIT),       /* 0x0a: L3 miss, shared */
73 	OP_LH | P(LVL, REM_RAM1) | REM | LEVEL(L3) | P(SNOOP, HIT),  /* 0x0b: L3 miss, shared */
74 	OP_LH | P(LVL, LOC_RAM)  | LEVEL(RAM) | SNOOP_NONE_MISS,     /* 0x0c: L3 miss, excl */
75 	OP_LH | P(LVL, REM_RAM1) | LEVEL(RAM) | REM | SNOOP_NONE_MISS, /* 0x0d: L3 miss, excl */
76 	OP_LH | P(LVL, IO)  | LEVEL(NA) | P(SNOOP, NONE), /* 0x0e: I/O */
77 	OP_LH | P(LVL, UNC) | LEVEL(NA) | P(SNOOP, NONE), /* 0x0f: uncached */
78 };
79 
80 /* Patch up minor differences in the bits */
intel_pmu_pebs_data_source_nhm(void)81 void __init intel_pmu_pebs_data_source_nhm(void)
82 {
83 	pebs_data_source[0x05] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HIT);
84 	pebs_data_source[0x06] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
85 	pebs_data_source[0x07] = OP_LH | P(LVL, L3) | LEVEL(L3) | P(SNOOP, HITM);
86 }
87 
intel_pmu_pebs_data_source_skl(bool pmem)88 void __init intel_pmu_pebs_data_source_skl(bool pmem)
89 {
90 	u64 pmem_or_l4 = pmem ? LEVEL(PMEM) : LEVEL(L4);
91 
92 	pebs_data_source[0x08] = OP_LH | pmem_or_l4 | P(SNOOP, HIT);
93 	pebs_data_source[0x09] = OP_LH | pmem_or_l4 | REM | P(SNOOP, HIT);
94 	pebs_data_source[0x0b] = OP_LH | LEVEL(RAM) | REM | P(SNOOP, NONE);
95 	pebs_data_source[0x0c] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOPX, FWD);
96 	pebs_data_source[0x0d] = OP_LH | LEVEL(ANY_CACHE) | REM | P(SNOOP, HITM);
97 }
98 
precise_store_data(u64 status)99 static u64 precise_store_data(u64 status)
100 {
101 	union intel_x86_pebs_dse dse;
102 	u64 val = P(OP, STORE) | P(SNOOP, NA) | P(LVL, L1) | P(TLB, L2);
103 
104 	dse.val = status;
105 
106 	/*
107 	 * bit 4: TLB access
108 	 * 1 = stored missed 2nd level TLB
109 	 *
110 	 * so it either hit the walker or the OS
111 	 * otherwise hit 2nd level TLB
112 	 */
113 	if (dse.st_stlb_miss)
114 		val |= P(TLB, MISS);
115 	else
116 		val |= P(TLB, HIT);
117 
118 	/*
119 	 * bit 0: hit L1 data cache
120 	 * if not set, then all we know is that
121 	 * it missed L1D
122 	 */
123 	if (dse.st_l1d_hit)
124 		val |= P(LVL, HIT);
125 	else
126 		val |= P(LVL, MISS);
127 
128 	/*
129 	 * bit 5: Locked prefix
130 	 */
131 	if (dse.st_locked)
132 		val |= P(LOCK, LOCKED);
133 
134 	return val;
135 }
136 
precise_datala_hsw(struct perf_event * event,u64 status)137 static u64 precise_datala_hsw(struct perf_event *event, u64 status)
138 {
139 	union perf_mem_data_src dse;
140 
141 	dse.val = PERF_MEM_NA;
142 
143 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW)
144 		dse.mem_op = PERF_MEM_OP_STORE;
145 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_LD_HSW)
146 		dse.mem_op = PERF_MEM_OP_LOAD;
147 
148 	/*
149 	 * L1 info only valid for following events:
150 	 *
151 	 * MEM_UOPS_RETIRED.STLB_MISS_STORES
152 	 * MEM_UOPS_RETIRED.LOCK_STORES
153 	 * MEM_UOPS_RETIRED.SPLIT_STORES
154 	 * MEM_UOPS_RETIRED.ALL_STORES
155 	 */
156 	if (event->hw.flags & PERF_X86_EVENT_PEBS_ST_HSW) {
157 		if (status & 1)
158 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_HIT;
159 		else
160 			dse.mem_lvl = PERF_MEM_LVL_L1 | PERF_MEM_LVL_MISS;
161 	}
162 	return dse.val;
163 }
164 
load_latency_data(u64 status)165 static u64 load_latency_data(u64 status)
166 {
167 	union intel_x86_pebs_dse dse;
168 	u64 val;
169 
170 	dse.val = status;
171 
172 	/*
173 	 * use the mapping table for bit 0-3
174 	 */
175 	val = pebs_data_source[dse.ld_dse];
176 
177 	/*
178 	 * Nehalem models do not support TLB, Lock infos
179 	 */
180 	if (x86_pmu.pebs_no_tlb) {
181 		val |= P(TLB, NA) | P(LOCK, NA);
182 		return val;
183 	}
184 	/*
185 	 * bit 4: TLB access
186 	 * 0 = did not miss 2nd level TLB
187 	 * 1 = missed 2nd level TLB
188 	 */
189 	if (dse.ld_stlb_miss)
190 		val |= P(TLB, MISS) | P(TLB, L2);
191 	else
192 		val |= P(TLB, HIT) | P(TLB, L1) | P(TLB, L2);
193 
194 	/*
195 	 * bit 5: locked prefix
196 	 */
197 	if (dse.ld_locked)
198 		val |= P(LOCK, LOCKED);
199 
200 	return val;
201 }
202 
203 struct pebs_record_core {
204 	u64 flags, ip;
205 	u64 ax, bx, cx, dx;
206 	u64 si, di, bp, sp;
207 	u64 r8,  r9,  r10, r11;
208 	u64 r12, r13, r14, r15;
209 };
210 
211 struct pebs_record_nhm {
212 	u64 flags, ip;
213 	u64 ax, bx, cx, dx;
214 	u64 si, di, bp, sp;
215 	u64 r8,  r9,  r10, r11;
216 	u64 r12, r13, r14, r15;
217 	u64 status, dla, dse, lat;
218 };
219 
220 /*
221  * Same as pebs_record_nhm, with two additional fields.
222  */
223 struct pebs_record_hsw {
224 	u64 flags, ip;
225 	u64 ax, bx, cx, dx;
226 	u64 si, di, bp, sp;
227 	u64 r8,  r9,  r10, r11;
228 	u64 r12, r13, r14, r15;
229 	u64 status, dla, dse, lat;
230 	u64 real_ip, tsx_tuning;
231 };
232 
233 union hsw_tsx_tuning {
234 	struct {
235 		u32 cycles_last_block     : 32,
236 		    hle_abort		  : 1,
237 		    rtm_abort		  : 1,
238 		    instruction_abort     : 1,
239 		    non_instruction_abort : 1,
240 		    retry		  : 1,
241 		    data_conflict	  : 1,
242 		    capacity_writes	  : 1,
243 		    capacity_reads	  : 1;
244 	};
245 	u64	    value;
246 };
247 
248 #define PEBS_HSW_TSX_FLAGS	0xff00000000ULL
249 
250 /* Same as HSW, plus TSC */
251 
252 struct pebs_record_skl {
253 	u64 flags, ip;
254 	u64 ax, bx, cx, dx;
255 	u64 si, di, bp, sp;
256 	u64 r8,  r9,  r10, r11;
257 	u64 r12, r13, r14, r15;
258 	u64 status, dla, dse, lat;
259 	u64 real_ip, tsx_tuning;
260 	u64 tsc;
261 };
262 
init_debug_store_on_cpu(int cpu)263 void init_debug_store_on_cpu(int cpu)
264 {
265 	struct debug_store *ds = per_cpu(cpu_hw_events, cpu).ds;
266 
267 	if (!ds)
268 		return;
269 
270 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA,
271 		     (u32)((u64)(unsigned long)ds),
272 		     (u32)((u64)(unsigned long)ds >> 32));
273 }
274 
fini_debug_store_on_cpu(int cpu)275 void fini_debug_store_on_cpu(int cpu)
276 {
277 	if (!per_cpu(cpu_hw_events, cpu).ds)
278 		return;
279 
280 	wrmsr_on_cpu(cpu, MSR_IA32_DS_AREA, 0, 0);
281 }
282 
283 static DEFINE_PER_CPU(void *, insn_buffer);
284 
ds_update_cea(void * cea,void * addr,size_t size,pgprot_t prot)285 static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
286 {
287 	unsigned long start = (unsigned long)cea;
288 	phys_addr_t pa;
289 	size_t msz = 0;
290 
291 	pa = virt_to_phys(addr);
292 
293 	preempt_disable();
294 	for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
295 		cea_set_pte(cea, pa, prot);
296 
297 	/*
298 	 * This is a cross-CPU update of the cpu_entry_area, we must shoot down
299 	 * all TLB entries for it.
300 	 */
301 	flush_tlb_kernel_range(start, start + size);
302 	preempt_enable();
303 }
304 
ds_clear_cea(void * cea,size_t size)305 static void ds_clear_cea(void *cea, size_t size)
306 {
307 	unsigned long start = (unsigned long)cea;
308 	size_t msz = 0;
309 
310 	preempt_disable();
311 	for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
312 		cea_set_pte(cea, 0, PAGE_NONE);
313 
314 	flush_tlb_kernel_range(start, start + size);
315 	preempt_enable();
316 }
317 
dsalloc_pages(size_t size,gfp_t flags,int cpu)318 static void *dsalloc_pages(size_t size, gfp_t flags, int cpu)
319 {
320 	unsigned int order = get_order(size);
321 	int node = cpu_to_node(cpu);
322 	struct page *page;
323 
324 	page = __alloc_pages_node(node, flags | __GFP_ZERO, order);
325 	return page ? page_address(page) : NULL;
326 }
327 
dsfree_pages(const void * buffer,size_t size)328 static void dsfree_pages(const void *buffer, size_t size)
329 {
330 	if (buffer)
331 		free_pages((unsigned long)buffer, get_order(size));
332 }
333 
alloc_pebs_buffer(int cpu)334 static int alloc_pebs_buffer(int cpu)
335 {
336 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
337 	struct debug_store *ds = hwev->ds;
338 	size_t bsiz = x86_pmu.pebs_buffer_size;
339 	int max, node = cpu_to_node(cpu);
340 	void *buffer, *ibuffer, *cea;
341 
342 	if (!x86_pmu.pebs)
343 		return 0;
344 
345 	buffer = dsalloc_pages(bsiz, GFP_KERNEL, cpu);
346 	if (unlikely(!buffer))
347 		return -ENOMEM;
348 
349 	/*
350 	 * HSW+ already provides us the eventing ip; no need to allocate this
351 	 * buffer then.
352 	 */
353 	if (x86_pmu.intel_cap.pebs_format < 2) {
354 		ibuffer = kzalloc_node(PEBS_FIXUP_SIZE, GFP_KERNEL, node);
355 		if (!ibuffer) {
356 			dsfree_pages(buffer, bsiz);
357 			return -ENOMEM;
358 		}
359 		per_cpu(insn_buffer, cpu) = ibuffer;
360 	}
361 	hwev->ds_pebs_vaddr = buffer;
362 	/* Update the cpu entry area mapping */
363 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
364 	ds->pebs_buffer_base = (unsigned long) cea;
365 	ds_update_cea(cea, buffer, bsiz, PAGE_KERNEL);
366 	ds->pebs_index = ds->pebs_buffer_base;
367 	max = x86_pmu.pebs_record_size * (bsiz / x86_pmu.pebs_record_size);
368 	ds->pebs_absolute_maximum = ds->pebs_buffer_base + max;
369 	return 0;
370 }
371 
release_pebs_buffer(int cpu)372 static void release_pebs_buffer(int cpu)
373 {
374 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
375 	struct debug_store *ds = hwev->ds;
376 	void *cea;
377 
378 	if (!ds || !x86_pmu.pebs)
379 		return;
380 
381 	kfree(per_cpu(insn_buffer, cpu));
382 	per_cpu(insn_buffer, cpu) = NULL;
383 
384 	/* Clear the fixmap */
385 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.pebs_buffer;
386 	ds_clear_cea(cea, x86_pmu.pebs_buffer_size);
387 	ds->pebs_buffer_base = 0;
388 	dsfree_pages(hwev->ds_pebs_vaddr, x86_pmu.pebs_buffer_size);
389 	hwev->ds_pebs_vaddr = NULL;
390 }
391 
alloc_bts_buffer(int cpu)392 static int alloc_bts_buffer(int cpu)
393 {
394 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
395 	struct debug_store *ds = hwev->ds;
396 	void *buffer, *cea;
397 	int max;
398 
399 	if (!x86_pmu.bts)
400 		return 0;
401 
402 	buffer = dsalloc_pages(BTS_BUFFER_SIZE, GFP_KERNEL | __GFP_NOWARN, cpu);
403 	if (unlikely(!buffer)) {
404 		WARN_ONCE(1, "%s: BTS buffer allocation failure\n", __func__);
405 		return -ENOMEM;
406 	}
407 	hwev->ds_bts_vaddr = buffer;
408 	/* Update the fixmap */
409 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
410 	ds->bts_buffer_base = (unsigned long) cea;
411 	ds_update_cea(cea, buffer, BTS_BUFFER_SIZE, PAGE_KERNEL);
412 	ds->bts_index = ds->bts_buffer_base;
413 	max = BTS_BUFFER_SIZE / BTS_RECORD_SIZE;
414 	ds->bts_absolute_maximum = ds->bts_buffer_base +
415 					max * BTS_RECORD_SIZE;
416 	ds->bts_interrupt_threshold = ds->bts_absolute_maximum -
417 					(max / 16) * BTS_RECORD_SIZE;
418 	return 0;
419 }
420 
release_bts_buffer(int cpu)421 static void release_bts_buffer(int cpu)
422 {
423 	struct cpu_hw_events *hwev = per_cpu_ptr(&cpu_hw_events, cpu);
424 	struct debug_store *ds = hwev->ds;
425 	void *cea;
426 
427 	if (!ds || !x86_pmu.bts)
428 		return;
429 
430 	/* Clear the fixmap */
431 	cea = &get_cpu_entry_area(cpu)->cpu_debug_buffers.bts_buffer;
432 	ds_clear_cea(cea, BTS_BUFFER_SIZE);
433 	ds->bts_buffer_base = 0;
434 	dsfree_pages(hwev->ds_bts_vaddr, BTS_BUFFER_SIZE);
435 	hwev->ds_bts_vaddr = NULL;
436 }
437 
alloc_ds_buffer(int cpu)438 static int alloc_ds_buffer(int cpu)
439 {
440 	struct debug_store *ds = &get_cpu_entry_area(cpu)->cpu_debug_store;
441 
442 	memset(ds, 0, sizeof(*ds));
443 	per_cpu(cpu_hw_events, cpu).ds = ds;
444 	return 0;
445 }
446 
release_ds_buffer(int cpu)447 static void release_ds_buffer(int cpu)
448 {
449 	per_cpu(cpu_hw_events, cpu).ds = NULL;
450 }
451 
release_ds_buffers(void)452 void release_ds_buffers(void)
453 {
454 	int cpu;
455 
456 	if (!x86_pmu.bts && !x86_pmu.pebs)
457 		return;
458 
459 	get_online_cpus();
460 	for_each_online_cpu(cpu)
461 		fini_debug_store_on_cpu(cpu);
462 
463 	for_each_possible_cpu(cpu) {
464 		release_pebs_buffer(cpu);
465 		release_bts_buffer(cpu);
466 		release_ds_buffer(cpu);
467 	}
468 	put_online_cpus();
469 }
470 
reserve_ds_buffers(void)471 void reserve_ds_buffers(void)
472 {
473 	int bts_err = 0, pebs_err = 0;
474 	int cpu;
475 
476 	x86_pmu.bts_active = 0;
477 	x86_pmu.pebs_active = 0;
478 
479 	if (!x86_pmu.bts && !x86_pmu.pebs)
480 		return;
481 
482 	if (!x86_pmu.bts)
483 		bts_err = 1;
484 
485 	if (!x86_pmu.pebs)
486 		pebs_err = 1;
487 
488 	get_online_cpus();
489 
490 	for_each_possible_cpu(cpu) {
491 		if (alloc_ds_buffer(cpu)) {
492 			bts_err = 1;
493 			pebs_err = 1;
494 		}
495 
496 		if (!bts_err && alloc_bts_buffer(cpu))
497 			bts_err = 1;
498 
499 		if (!pebs_err && alloc_pebs_buffer(cpu))
500 			pebs_err = 1;
501 
502 		if (bts_err && pebs_err)
503 			break;
504 	}
505 
506 	if (bts_err) {
507 		for_each_possible_cpu(cpu)
508 			release_bts_buffer(cpu);
509 	}
510 
511 	if (pebs_err) {
512 		for_each_possible_cpu(cpu)
513 			release_pebs_buffer(cpu);
514 	}
515 
516 	if (bts_err && pebs_err) {
517 		for_each_possible_cpu(cpu)
518 			release_ds_buffer(cpu);
519 	} else {
520 		if (x86_pmu.bts && !bts_err)
521 			x86_pmu.bts_active = 1;
522 
523 		if (x86_pmu.pebs && !pebs_err)
524 			x86_pmu.pebs_active = 1;
525 
526 		for_each_online_cpu(cpu)
527 			init_debug_store_on_cpu(cpu);
528 	}
529 
530 	put_online_cpus();
531 }
532 
533 /*
534  * BTS
535  */
536 
537 struct event_constraint bts_constraint =
538 	EVENT_CONSTRAINT(0, 1ULL << INTEL_PMC_IDX_FIXED_BTS, 0);
539 
intel_pmu_enable_bts(u64 config)540 void intel_pmu_enable_bts(u64 config)
541 {
542 	unsigned long debugctlmsr;
543 
544 	debugctlmsr = get_debugctlmsr();
545 
546 	debugctlmsr |= DEBUGCTLMSR_TR;
547 	debugctlmsr |= DEBUGCTLMSR_BTS;
548 	if (config & ARCH_PERFMON_EVENTSEL_INT)
549 		debugctlmsr |= DEBUGCTLMSR_BTINT;
550 
551 	if (!(config & ARCH_PERFMON_EVENTSEL_OS))
552 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_OS;
553 
554 	if (!(config & ARCH_PERFMON_EVENTSEL_USR))
555 		debugctlmsr |= DEBUGCTLMSR_BTS_OFF_USR;
556 
557 	update_debugctlmsr(debugctlmsr);
558 }
559 
intel_pmu_disable_bts(void)560 void intel_pmu_disable_bts(void)
561 {
562 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
563 	unsigned long debugctlmsr;
564 
565 	if (!cpuc->ds)
566 		return;
567 
568 	debugctlmsr = get_debugctlmsr();
569 
570 	debugctlmsr &=
571 		~(DEBUGCTLMSR_TR | DEBUGCTLMSR_BTS | DEBUGCTLMSR_BTINT |
572 		  DEBUGCTLMSR_BTS_OFF_OS | DEBUGCTLMSR_BTS_OFF_USR);
573 
574 	update_debugctlmsr(debugctlmsr);
575 }
576 
intel_pmu_drain_bts_buffer(void)577 int intel_pmu_drain_bts_buffer(void)
578 {
579 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
580 	struct debug_store *ds = cpuc->ds;
581 	struct bts_record {
582 		u64	from;
583 		u64	to;
584 		u64	flags;
585 	};
586 	struct perf_event *event = cpuc->events[INTEL_PMC_IDX_FIXED_BTS];
587 	struct bts_record *at, *base, *top;
588 	struct perf_output_handle handle;
589 	struct perf_event_header header;
590 	struct perf_sample_data data;
591 	unsigned long skip = 0;
592 	struct pt_regs regs;
593 
594 	if (!event)
595 		return 0;
596 
597 	if (!x86_pmu.bts_active)
598 		return 0;
599 
600 	base = (struct bts_record *)(unsigned long)ds->bts_buffer_base;
601 	top  = (struct bts_record *)(unsigned long)ds->bts_index;
602 
603 	if (top <= base)
604 		return 0;
605 
606 	memset(&regs, 0, sizeof(regs));
607 
608 	ds->bts_index = ds->bts_buffer_base;
609 
610 	perf_sample_data_init(&data, 0, event->hw.last_period);
611 
612 	/*
613 	 * BTS leaks kernel addresses in branches across the cpl boundary,
614 	 * such as traps or system calls, so unless the user is asking for
615 	 * kernel tracing (and right now it's not possible), we'd need to
616 	 * filter them out. But first we need to count how many of those we
617 	 * have in the current batch. This is an extra O(n) pass, however,
618 	 * it's much faster than the other one especially considering that
619 	 * n <= 2560 (BTS_BUFFER_SIZE / BTS_RECORD_SIZE * 15/16; see the
620 	 * alloc_bts_buffer()).
621 	 */
622 	for (at = base; at < top; at++) {
623 		/*
624 		 * Note that right now *this* BTS code only works if
625 		 * attr::exclude_kernel is set, but let's keep this extra
626 		 * check here in case that changes.
627 		 */
628 		if (event->attr.exclude_kernel &&
629 		    (kernel_ip(at->from) || kernel_ip(at->to)))
630 			skip++;
631 	}
632 
633 	/*
634 	 * Prepare a generic sample, i.e. fill in the invariant fields.
635 	 * We will overwrite the from and to address before we output
636 	 * the sample.
637 	 */
638 	rcu_read_lock();
639 	perf_prepare_sample(&header, &data, event, &regs);
640 
641 	if (perf_output_begin(&handle, event, header.size *
642 			      (top - base - skip)))
643 		goto unlock;
644 
645 	for (at = base; at < top; at++) {
646 		/* Filter out any records that contain kernel addresses. */
647 		if (event->attr.exclude_kernel &&
648 		    (kernel_ip(at->from) || kernel_ip(at->to)))
649 			continue;
650 
651 		data.ip		= at->from;
652 		data.addr	= at->to;
653 
654 		perf_output_sample(&handle, &header, &data, event);
655 	}
656 
657 	perf_output_end(&handle);
658 
659 	/* There's new data available. */
660 	event->hw.interrupts++;
661 	event->pending_kill = POLL_IN;
662 unlock:
663 	rcu_read_unlock();
664 	return 1;
665 }
666 
intel_pmu_drain_pebs_buffer(void)667 static inline void intel_pmu_drain_pebs_buffer(void)
668 {
669 	struct pt_regs regs;
670 
671 	x86_pmu.drain_pebs(&regs);
672 }
673 
674 /*
675  * PEBS
676  */
677 struct event_constraint intel_core2_pebs_event_constraints[] = {
678 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
679 	INTEL_FLAGS_UEVENT_CONSTRAINT(0xfec1, 0x1), /* X87_OPS_RETIRED.ANY */
680 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* BR_INST_RETIRED.MISPRED */
681 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1fc7, 0x1), /* SIMD_INST_RETURED.ANY */
682 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
683 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
684 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
685 	EVENT_CONSTRAINT_END
686 };
687 
688 struct event_constraint intel_atom_pebs_event_constraints[] = {
689 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c0, 0x1), /* INST_RETIRED.ANY */
690 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x00c5, 0x1), /* MISPREDICTED_BRANCH_RETIRED */
691 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0x1),    /* MEM_LOAD_RETIRED.* */
692 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
693 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x01),
694 	/* Allow all events as PEBS with no flags */
695 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
696 	EVENT_CONSTRAINT_END
697 };
698 
699 struct event_constraint intel_slm_pebs_event_constraints[] = {
700 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
701 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x1),
702 	/* Allow all events as PEBS with no flags */
703 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
704 	EVENT_CONSTRAINT_END
705 };
706 
707 struct event_constraint intel_glm_pebs_event_constraints[] = {
708 	/* Allow all events as PEBS with no flags */
709 	INTEL_ALL_EVENT_CONSTRAINT(0, 0x1),
710 	EVENT_CONSTRAINT_END
711 };
712 
713 struct event_constraint intel_glp_pebs_event_constraints[] = {
714 	/* Allow all events as PEBS with no flags */
715 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
716 	EVENT_CONSTRAINT_END
717 };
718 
719 struct event_constraint intel_nehalem_pebs_event_constraints[] = {
720 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
721 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
722 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
723 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INST_RETIRED.ANY */
724 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
725 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
726 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x02c5, 0xf), /* BR_MISP_RETIRED.NEAR_CALL */
727 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
728 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
729 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
730 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
731 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
732 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
733 	EVENT_CONSTRAINT_END
734 };
735 
736 struct event_constraint intel_westmere_pebs_event_constraints[] = {
737 	INTEL_PLD_CONSTRAINT(0x100b, 0xf),      /* MEM_INST_RETIRED.* */
738 	INTEL_FLAGS_EVENT_CONSTRAINT(0x0f, 0xf),    /* MEM_UNCORE_RETIRED.* */
739 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x010c, 0xf), /* MEM_STORE_RETIRED.DTLB_MISS */
740 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc0, 0xf),    /* INSTR_RETIRED.* */
741 	INTEL_EVENT_CONSTRAINT(0xc2, 0xf),    /* UOPS_RETIRED.* */
742 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc4, 0xf),    /* BR_INST_RETIRED.* */
743 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc5, 0xf),    /* BR_MISP_RETIRED.* */
744 	INTEL_FLAGS_EVENT_CONSTRAINT(0xc7, 0xf),    /* SSEX_UOPS_RETIRED.* */
745 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x20c8, 0xf), /* ITLB_MISS_RETIRED */
746 	INTEL_FLAGS_EVENT_CONSTRAINT(0xcb, 0xf),    /* MEM_LOAD_RETIRED.* */
747 	INTEL_FLAGS_EVENT_CONSTRAINT(0xf7, 0xf),    /* FP_ASSIST.* */
748 	/* INST_RETIRED.ANY_P, inv=1, cmask=16 (cycles:p). */
749 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
750 	EVENT_CONSTRAINT_END
751 };
752 
753 struct event_constraint intel_snb_pebs_event_constraints[] = {
754 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
755 	INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
756 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
757 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
758 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
759         INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
760         INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
761         INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
762         INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
763 	/* Allow all events as PEBS with no flags */
764 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
765 	EVENT_CONSTRAINT_END
766 };
767 
768 struct event_constraint intel_ivb_pebs_event_constraints[] = {
769         INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
770         INTEL_PLD_CONSTRAINT(0x01cd, 0x8),    /* MEM_TRANS_RETIRED.LAT_ABOVE_THR */
771 	INTEL_PST_CONSTRAINT(0x02cd, 0x8),    /* MEM_TRANS_RETIRED.PRECISE_STORES */
772 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
773 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
774 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
775 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
776 	INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf),    /* MEM_UOP_RETIRED.* */
777 	INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
778 	INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf),    /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */
779 	INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf),    /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */
780 	/* Allow all events as PEBS with no flags */
781 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
782         EVENT_CONSTRAINT_END
783 };
784 
785 struct event_constraint intel_hsw_pebs_event_constraints[] = {
786 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
787 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
788 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
789 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
790 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
791 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
792 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
793 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
794 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
795 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
796 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
797 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
798 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
799 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
800 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
801 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
802 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
803 	/* Allow all events as PEBS with no flags */
804 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
805 	EVENT_CONSTRAINT_END
806 };
807 
808 struct event_constraint intel_bdw_pebs_event_constraints[] = {
809 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PRECDIST */
810 	INTEL_PLD_CONSTRAINT(0x01cd, 0xf),    /* MEM_TRANS_RETIRED.* */
811 	/* UOPS_RETIRED.ALL, inv=1, cmask=16 (cycles:p). */
812 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c2, 0xf),
813 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
814 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
815 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(0x01c2, 0xf), /* UOPS_RETIRED.ALL */
816 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_LOADS */
817 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_UOPS_RETIRED.LOCK_LOADS */
818 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_LOADS */
819 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_UOPS_RETIRED.ALL_LOADS */
820 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_UOPS_RETIRED.STLB_MISS_STORES */
821 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_UOPS_RETIRED.SPLIT_STORES */
822 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_UOPS_RETIRED.ALL_STORES */
823 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_UOPS_RETIRED.* */
824 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_UOPS_L3_HIT_RETIRED.* */
825 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_UOPS_L3_MISS_RETIRED.* */
826 	/* Allow all events as PEBS with no flags */
827 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
828 	EVENT_CONSTRAINT_END
829 };
830 
831 
832 struct event_constraint intel_skl_pebs_event_constraints[] = {
833 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x1c0, 0x2),	/* INST_RETIRED.PREC_DIST */
834 	/* INST_RETIRED.PREC_DIST, inv=1, cmask=16 (cycles:ppp). */
835 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108001c0, 0x2),
836 	/* INST_RETIRED.TOTAL_CYCLES_PS (inv=1, cmask=16) (cycles:p). */
837 	INTEL_FLAGS_UEVENT_CONSTRAINT(0x108000c0, 0x0f),
838 	INTEL_PLD_CONSTRAINT(0x1cd, 0xf),		      /* MEM_TRANS_RETIRED.* */
839 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x11d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_LOADS */
840 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x12d0, 0xf), /* MEM_INST_RETIRED.STLB_MISS_STORES */
841 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x21d0, 0xf), /* MEM_INST_RETIRED.LOCK_LOADS */
842 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x22d0, 0xf), /* MEM_INST_RETIRED.LOCK_STORES */
843 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x41d0, 0xf), /* MEM_INST_RETIRED.SPLIT_LOADS */
844 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x42d0, 0xf), /* MEM_INST_RETIRED.SPLIT_STORES */
845 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(0x81d0, 0xf), /* MEM_INST_RETIRED.ALL_LOADS */
846 	INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(0x82d0, 0xf), /* MEM_INST_RETIRED.ALL_STORES */
847 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd1, 0xf),    /* MEM_LOAD_RETIRED.* */
848 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd2, 0xf),    /* MEM_LOAD_L3_HIT_RETIRED.* */
849 	INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(0xd3, 0xf),    /* MEM_LOAD_L3_MISS_RETIRED.* */
850 	/* Allow all events as PEBS with no flags */
851 	INTEL_ALL_EVENT_CONSTRAINT(0, 0xf),
852 	EVENT_CONSTRAINT_END
853 };
854 
intel_pebs_constraints(struct perf_event * event)855 struct event_constraint *intel_pebs_constraints(struct perf_event *event)
856 {
857 	struct event_constraint *c;
858 
859 	if (!event->attr.precise_ip)
860 		return NULL;
861 
862 	if (x86_pmu.pebs_constraints) {
863 		for_each_event_constraint(c, x86_pmu.pebs_constraints) {
864 			if ((event->hw.config & c->cmask) == c->code) {
865 				event->hw.flags |= c->flags;
866 				return c;
867 			}
868 		}
869 	}
870 
871 	return &emptyconstraint;
872 }
873 
874 /*
875  * We need the sched_task callback even for per-cpu events when we use
876  * the large interrupt threshold, such that we can provide PID and TID
877  * to PEBS samples.
878  */
pebs_needs_sched_cb(struct cpu_hw_events * cpuc)879 static inline bool pebs_needs_sched_cb(struct cpu_hw_events *cpuc)
880 {
881 	return cpuc->n_pebs && (cpuc->n_pebs == cpuc->n_large_pebs);
882 }
883 
intel_pmu_pebs_sched_task(struct perf_event_context * ctx,bool sched_in)884 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in)
885 {
886 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
887 
888 	if (!sched_in && pebs_needs_sched_cb(cpuc))
889 		intel_pmu_drain_pebs_buffer();
890 }
891 
pebs_update_threshold(struct cpu_hw_events * cpuc)892 static inline void pebs_update_threshold(struct cpu_hw_events *cpuc)
893 {
894 	struct debug_store *ds = cpuc->ds;
895 	u64 threshold;
896 
897 	if (cpuc->n_pebs == cpuc->n_large_pebs) {
898 		threshold = ds->pebs_absolute_maximum -
899 			x86_pmu.max_pebs_events * x86_pmu.pebs_record_size;
900 	} else {
901 		threshold = ds->pebs_buffer_base + x86_pmu.pebs_record_size;
902 	}
903 
904 	ds->pebs_interrupt_threshold = threshold;
905 }
906 
907 static void
pebs_update_state(bool needed_cb,struct cpu_hw_events * cpuc,struct pmu * pmu)908 pebs_update_state(bool needed_cb, struct cpu_hw_events *cpuc, struct pmu *pmu)
909 {
910 	/*
911 	 * Make sure we get updated with the first PEBS
912 	 * event. It will trigger also during removal, but
913 	 * that does not hurt:
914 	 */
915 	bool update = cpuc->n_pebs == 1;
916 
917 	if (needed_cb != pebs_needs_sched_cb(cpuc)) {
918 		if (!needed_cb)
919 			perf_sched_cb_inc(pmu);
920 		else
921 			perf_sched_cb_dec(pmu);
922 
923 		update = true;
924 	}
925 
926 	if (update)
927 		pebs_update_threshold(cpuc);
928 }
929 
intel_pmu_pebs_add(struct perf_event * event)930 void intel_pmu_pebs_add(struct perf_event *event)
931 {
932 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
933 	struct hw_perf_event *hwc = &event->hw;
934 	bool needed_cb = pebs_needs_sched_cb(cpuc);
935 
936 	cpuc->n_pebs++;
937 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
938 		cpuc->n_large_pebs++;
939 
940 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
941 }
942 
intel_pmu_pebs_enable(struct perf_event * event)943 void intel_pmu_pebs_enable(struct perf_event *event)
944 {
945 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
946 	struct hw_perf_event *hwc = &event->hw;
947 	struct debug_store *ds = cpuc->ds;
948 
949 	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
950 
951 	cpuc->pebs_enabled |= 1ULL << hwc->idx;
952 
953 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
954 		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
955 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
956 		cpuc->pebs_enabled |= 1ULL << 63;
957 
958 	/*
959 	 * Use auto-reload if possible to save a MSR write in the PMI.
960 	 * This must be done in pmu::start(), because PERF_EVENT_IOC_PERIOD.
961 	 */
962 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
963 		ds->pebs_event_reset[hwc->idx] =
964 			(u64)(-hwc->sample_period) & x86_pmu.cntval_mask;
965 	} else {
966 		ds->pebs_event_reset[hwc->idx] = 0;
967 	}
968 }
969 
intel_pmu_pebs_del(struct perf_event * event)970 void intel_pmu_pebs_del(struct perf_event *event)
971 {
972 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
973 	struct hw_perf_event *hwc = &event->hw;
974 	bool needed_cb = pebs_needs_sched_cb(cpuc);
975 
976 	cpuc->n_pebs--;
977 	if (hwc->flags & PERF_X86_EVENT_FREERUNNING)
978 		cpuc->n_large_pebs--;
979 
980 	pebs_update_state(needed_cb, cpuc, event->ctx->pmu);
981 }
982 
intel_pmu_pebs_disable(struct perf_event * event)983 void intel_pmu_pebs_disable(struct perf_event *event)
984 {
985 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
986 	struct hw_perf_event *hwc = &event->hw;
987 
988 	if (cpuc->n_pebs == cpuc->n_large_pebs)
989 		intel_pmu_drain_pebs_buffer();
990 
991 	cpuc->pebs_enabled &= ~(1ULL << hwc->idx);
992 
993 	if (event->hw.flags & PERF_X86_EVENT_PEBS_LDLAT)
994 		cpuc->pebs_enabled &= ~(1ULL << (hwc->idx + 32));
995 	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
996 		cpuc->pebs_enabled &= ~(1ULL << 63);
997 
998 	if (cpuc->enabled)
999 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1000 
1001 	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
1002 }
1003 
intel_pmu_pebs_enable_all(void)1004 void intel_pmu_pebs_enable_all(void)
1005 {
1006 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1007 
1008 	if (cpuc->pebs_enabled)
1009 		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
1010 }
1011 
intel_pmu_pebs_disable_all(void)1012 void intel_pmu_pebs_disable_all(void)
1013 {
1014 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1015 
1016 	if (cpuc->pebs_enabled)
1017 		wrmsrl(MSR_IA32_PEBS_ENABLE, 0);
1018 }
1019 
intel_pmu_pebs_fixup_ip(struct pt_regs * regs)1020 static int intel_pmu_pebs_fixup_ip(struct pt_regs *regs)
1021 {
1022 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1023 	unsigned long from = cpuc->lbr_entries[0].from;
1024 	unsigned long old_to, to = cpuc->lbr_entries[0].to;
1025 	unsigned long ip = regs->ip;
1026 	int is_64bit = 0;
1027 	void *kaddr;
1028 	int size;
1029 
1030 	/*
1031 	 * We don't need to fixup if the PEBS assist is fault like
1032 	 */
1033 	if (!x86_pmu.intel_cap.pebs_trap)
1034 		return 1;
1035 
1036 	/*
1037 	 * No LBR entry, no basic block, no rewinding
1038 	 */
1039 	if (!cpuc->lbr_stack.nr || !from || !to)
1040 		return 0;
1041 
1042 	/*
1043 	 * Basic blocks should never cross user/kernel boundaries
1044 	 */
1045 	if (kernel_ip(ip) != kernel_ip(to))
1046 		return 0;
1047 
1048 	/*
1049 	 * unsigned math, either ip is before the start (impossible) or
1050 	 * the basic block is larger than 1 page (sanity)
1051 	 */
1052 	if ((ip - to) > PEBS_FIXUP_SIZE)
1053 		return 0;
1054 
1055 	/*
1056 	 * We sampled a branch insn, rewind using the LBR stack
1057 	 */
1058 	if (ip == to) {
1059 		set_linear_ip(regs, from);
1060 		return 1;
1061 	}
1062 
1063 	size = ip - to;
1064 	if (!kernel_ip(ip)) {
1065 		int bytes;
1066 		u8 *buf = this_cpu_read(insn_buffer);
1067 
1068 		/* 'size' must fit our buffer, see above */
1069 		bytes = copy_from_user_nmi(buf, (void __user *)to, size);
1070 		if (bytes != 0)
1071 			return 0;
1072 
1073 		kaddr = buf;
1074 	} else {
1075 		kaddr = (void *)to;
1076 	}
1077 
1078 	do {
1079 		struct insn insn;
1080 
1081 		old_to = to;
1082 
1083 #ifdef CONFIG_X86_64
1084 		is_64bit = kernel_ip(to) || !test_thread_flag(TIF_IA32);
1085 #endif
1086 		insn_init(&insn, kaddr, size, is_64bit);
1087 		insn_get_length(&insn);
1088 		/*
1089 		 * Make sure there was not a problem decoding the
1090 		 * instruction and getting the length.  This is
1091 		 * doubly important because we have an infinite
1092 		 * loop if insn.length=0.
1093 		 */
1094 		if (!insn.length)
1095 			break;
1096 
1097 		to += insn.length;
1098 		kaddr += insn.length;
1099 		size -= insn.length;
1100 	} while (to < ip);
1101 
1102 	if (to == ip) {
1103 		set_linear_ip(regs, old_to);
1104 		return 1;
1105 	}
1106 
1107 	/*
1108 	 * Even though we decoded the basic block, the instruction stream
1109 	 * never matched the given IP, either the TO or the IP got corrupted.
1110 	 */
1111 	return 0;
1112 }
1113 
intel_hsw_weight(struct pebs_record_skl * pebs)1114 static inline u64 intel_hsw_weight(struct pebs_record_skl *pebs)
1115 {
1116 	if (pebs->tsx_tuning) {
1117 		union hsw_tsx_tuning tsx = { .value = pebs->tsx_tuning };
1118 		return tsx.cycles_last_block;
1119 	}
1120 	return 0;
1121 }
1122 
intel_hsw_transaction(struct pebs_record_skl * pebs)1123 static inline u64 intel_hsw_transaction(struct pebs_record_skl *pebs)
1124 {
1125 	u64 txn = (pebs->tsx_tuning & PEBS_HSW_TSX_FLAGS) >> 32;
1126 
1127 	/* For RTM XABORTs also log the abort code from AX */
1128 	if ((txn & PERF_TXN_TRANSACTION) && (pebs->ax & 1))
1129 		txn |= ((pebs->ax >> 24) & 0xff) << PERF_TXN_ABORT_SHIFT;
1130 	return txn;
1131 }
1132 
setup_pebs_sample_data(struct perf_event * event,struct pt_regs * iregs,void * __pebs,struct perf_sample_data * data,struct pt_regs * regs)1133 static void setup_pebs_sample_data(struct perf_event *event,
1134 				   struct pt_regs *iregs, void *__pebs,
1135 				   struct perf_sample_data *data,
1136 				   struct pt_regs *regs)
1137 {
1138 #define PERF_X86_EVENT_PEBS_HSW_PREC \
1139 		(PERF_X86_EVENT_PEBS_ST_HSW | \
1140 		 PERF_X86_EVENT_PEBS_LD_HSW | \
1141 		 PERF_X86_EVENT_PEBS_NA_HSW)
1142 	/*
1143 	 * We cast to the biggest pebs_record but are careful not to
1144 	 * unconditionally access the 'extra' entries.
1145 	 */
1146 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1147 	struct pebs_record_skl *pebs = __pebs;
1148 	u64 sample_type;
1149 	int fll, fst, dsrc;
1150 	int fl = event->hw.flags;
1151 
1152 	if (pebs == NULL)
1153 		return;
1154 
1155 	regs->flags &= ~PERF_EFLAGS_EXACT;
1156 	sample_type = event->attr.sample_type;
1157 	dsrc = sample_type & PERF_SAMPLE_DATA_SRC;
1158 
1159 	fll = fl & PERF_X86_EVENT_PEBS_LDLAT;
1160 	fst = fl & (PERF_X86_EVENT_PEBS_ST | PERF_X86_EVENT_PEBS_HSW_PREC);
1161 
1162 	perf_sample_data_init(data, 0, event->hw.last_period);
1163 
1164 	data->period = event->hw.last_period;
1165 
1166 	/*
1167 	 * Use latency for weight (only avail with PEBS-LL)
1168 	 */
1169 	if (fll && (sample_type & PERF_SAMPLE_WEIGHT))
1170 		data->weight = pebs->lat;
1171 
1172 	/*
1173 	 * data.data_src encodes the data source
1174 	 */
1175 	if (dsrc) {
1176 		u64 val = PERF_MEM_NA;
1177 		if (fll)
1178 			val = load_latency_data(pebs->dse);
1179 		else if (fst && (fl & PERF_X86_EVENT_PEBS_HSW_PREC))
1180 			val = precise_datala_hsw(event, pebs->dse);
1181 		else if (fst)
1182 			val = precise_store_data(pebs->dse);
1183 		data->data_src.val = val;
1184 	}
1185 
1186 	/*
1187 	 * We use the interrupt regs as a base because the PEBS record does not
1188 	 * contain a full regs set, specifically it seems to lack segment
1189 	 * descriptors, which get used by things like user_mode().
1190 	 *
1191 	 * In the simple case fix up only the IP for PERF_SAMPLE_IP.
1192 	 *
1193 	 * We must however always use BP,SP from iregs for the unwinder to stay
1194 	 * sane; the record BP,SP can point into thin air when the record is
1195 	 * from a previous PMI context or an (I)RET happend between the record
1196 	 * and PMI.
1197 	 */
1198 	*regs = *iregs;
1199 	regs->flags = pebs->flags;
1200 
1201 	if (sample_type & PERF_SAMPLE_REGS_INTR) {
1202 		regs->ax = pebs->ax;
1203 		regs->bx = pebs->bx;
1204 		regs->cx = pebs->cx;
1205 		regs->dx = pebs->dx;
1206 		regs->si = pebs->si;
1207 		regs->di = pebs->di;
1208 
1209 		/*
1210 		 * Per the above; only set BP,SP if we don't need callchains.
1211 		 *
1212 		 * XXX: does this make sense?
1213 		 */
1214 		if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
1215 			regs->bp = pebs->bp;
1216 			regs->sp = pebs->sp;
1217 		}
1218 
1219 		/*
1220 		 * Preserve PERF_EFLAGS_VM from set_linear_ip().
1221 		 */
1222 		regs->flags = pebs->flags | (regs->flags & PERF_EFLAGS_VM);
1223 #ifndef CONFIG_X86_32
1224 		regs->r8 = pebs->r8;
1225 		regs->r9 = pebs->r9;
1226 		regs->r10 = pebs->r10;
1227 		regs->r11 = pebs->r11;
1228 		regs->r12 = pebs->r12;
1229 		regs->r13 = pebs->r13;
1230 		regs->r14 = pebs->r14;
1231 		regs->r15 = pebs->r15;
1232 #endif
1233 	}
1234 
1235 	if (event->attr.precise_ip > 1) {
1236 		/* Haswell and later have the eventing IP, so use it: */
1237 		if (x86_pmu.intel_cap.pebs_format >= 2) {
1238 			set_linear_ip(regs, pebs->real_ip);
1239 			regs->flags |= PERF_EFLAGS_EXACT;
1240 		} else {
1241 			/* Otherwise use PEBS off-by-1 IP: */
1242 			set_linear_ip(regs, pebs->ip);
1243 
1244 			/* ... and try to fix it up using the LBR entries: */
1245 			if (intel_pmu_pebs_fixup_ip(regs))
1246 				regs->flags |= PERF_EFLAGS_EXACT;
1247 		}
1248 	} else
1249 		set_linear_ip(regs, pebs->ip);
1250 
1251 
1252 	if ((sample_type & (PERF_SAMPLE_ADDR | PERF_SAMPLE_PHYS_ADDR)) &&
1253 	    x86_pmu.intel_cap.pebs_format >= 1)
1254 		data->addr = pebs->dla;
1255 
1256 	if (x86_pmu.intel_cap.pebs_format >= 2) {
1257 		/* Only set the TSX weight when no memory weight. */
1258 		if ((sample_type & PERF_SAMPLE_WEIGHT) && !fll)
1259 			data->weight = intel_hsw_weight(pebs);
1260 
1261 		if (sample_type & PERF_SAMPLE_TRANSACTION)
1262 			data->txn = intel_hsw_transaction(pebs);
1263 	}
1264 
1265 	/*
1266 	 * v3 supplies an accurate time stamp, so we use that
1267 	 * for the time stamp.
1268 	 *
1269 	 * We can only do this for the default trace clock.
1270 	 */
1271 	if (x86_pmu.intel_cap.pebs_format >= 3 &&
1272 		event->attr.use_clockid == 0)
1273 		data->time = native_sched_clock_from_tsc(pebs->tsc);
1274 
1275 	if (has_branch_stack(event))
1276 		data->br_stack = &cpuc->lbr_stack;
1277 }
1278 
1279 static inline void *
get_next_pebs_record_by_bit(void * base,void * top,int bit)1280 get_next_pebs_record_by_bit(void *base, void *top, int bit)
1281 {
1282 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1283 	void *at;
1284 	u64 pebs_status;
1285 
1286 	/*
1287 	 * fmt0 does not have a status bitfield (does not use
1288 	 * perf_record_nhm format)
1289 	 */
1290 	if (x86_pmu.intel_cap.pebs_format < 1)
1291 		return base;
1292 
1293 	if (base == NULL)
1294 		return NULL;
1295 
1296 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1297 		struct pebs_record_nhm *p = at;
1298 
1299 		if (test_bit(bit, (unsigned long *)&p->status)) {
1300 			/* PEBS v3 has accurate status bits */
1301 			if (x86_pmu.intel_cap.pebs_format >= 3)
1302 				return at;
1303 
1304 			if (p->status == (1 << bit))
1305 				return at;
1306 
1307 			/* clear non-PEBS bit and re-check */
1308 			pebs_status = p->status & cpuc->pebs_enabled;
1309 			pebs_status &= PEBS_COUNTER_MASK;
1310 			if (pebs_status == (1 << bit))
1311 				return at;
1312 		}
1313 	}
1314 	return NULL;
1315 }
1316 
1317 /*
1318  * Special variant of intel_pmu_save_and_restart() for auto-reload.
1319  */
1320 static int
intel_pmu_save_and_restart_reload(struct perf_event * event,int count)1321 intel_pmu_save_and_restart_reload(struct perf_event *event, int count)
1322 {
1323 	struct hw_perf_event *hwc = &event->hw;
1324 	int shift = 64 - x86_pmu.cntval_bits;
1325 	u64 period = hwc->sample_period;
1326 	u64 prev_raw_count, new_raw_count;
1327 	s64 new, old;
1328 
1329 	WARN_ON(!period);
1330 
1331 	/*
1332 	 * drain_pebs() only happens when the PMU is disabled.
1333 	 */
1334 	WARN_ON(this_cpu_read(cpu_hw_events.enabled));
1335 
1336 	prev_raw_count = local64_read(&hwc->prev_count);
1337 	rdpmcl(hwc->event_base_rdpmc, new_raw_count);
1338 	local64_set(&hwc->prev_count, new_raw_count);
1339 
1340 	/*
1341 	 * Since the counter increments a negative counter value and
1342 	 * overflows on the sign switch, giving the interval:
1343 	 *
1344 	 *   [-period, 0]
1345 	 *
1346 	 * the difference between two consequtive reads is:
1347 	 *
1348 	 *   A) value2 - value1;
1349 	 *      when no overflows have happened in between,
1350 	 *
1351 	 *   B) (0 - value1) + (value2 - (-period));
1352 	 *      when one overflow happened in between,
1353 	 *
1354 	 *   C) (0 - value1) + (n - 1) * (period) + (value2 - (-period));
1355 	 *      when @n overflows happened in between.
1356 	 *
1357 	 * Here A) is the obvious difference, B) is the extension to the
1358 	 * discrete interval, where the first term is to the top of the
1359 	 * interval and the second term is from the bottom of the next
1360 	 * interval and C) the extension to multiple intervals, where the
1361 	 * middle term is the whole intervals covered.
1362 	 *
1363 	 * An equivalent of C, by reduction, is:
1364 	 *
1365 	 *   value2 - value1 + n * period
1366 	 */
1367 	new = ((s64)(new_raw_count << shift) >> shift);
1368 	old = ((s64)(prev_raw_count << shift) >> shift);
1369 	local64_add(new - old + count * period, &event->count);
1370 
1371 	local64_set(&hwc->period_left, -new);
1372 
1373 	perf_event_update_userpage(event);
1374 
1375 	return 0;
1376 }
1377 
__intel_pmu_pebs_event(struct perf_event * event,struct pt_regs * iregs,void * base,void * top,int bit,int count)1378 static void __intel_pmu_pebs_event(struct perf_event *event,
1379 				   struct pt_regs *iregs,
1380 				   void *base, void *top,
1381 				   int bit, int count)
1382 {
1383 	struct hw_perf_event *hwc = &event->hw;
1384 	struct perf_sample_data data;
1385 	struct pt_regs regs;
1386 	void *at = get_next_pebs_record_by_bit(base, top, bit);
1387 
1388 	if (hwc->flags & PERF_X86_EVENT_AUTO_RELOAD) {
1389 		/*
1390 		 * Now, auto-reload is only enabled in fixed period mode.
1391 		 * The reload value is always hwc->sample_period.
1392 		 * May need to change it, if auto-reload is enabled in
1393 		 * freq mode later.
1394 		 */
1395 		intel_pmu_save_and_restart_reload(event, count);
1396 	} else if (!intel_pmu_save_and_restart(event))
1397 		return;
1398 
1399 	while (count > 1) {
1400 		setup_pebs_sample_data(event, iregs, at, &data, &regs);
1401 		perf_event_output(event, &data, &regs);
1402 		at += x86_pmu.pebs_record_size;
1403 		at = get_next_pebs_record_by_bit(at, top, bit);
1404 		count--;
1405 	}
1406 
1407 	setup_pebs_sample_data(event, iregs, at, &data, &regs);
1408 
1409 	/*
1410 	 * All but the last records are processed.
1411 	 * The last one is left to be able to call the overflow handler.
1412 	 */
1413 	if (perf_event_overflow(event, &data, &regs)) {
1414 		x86_pmu_stop(event, 0);
1415 		return;
1416 	}
1417 
1418 }
1419 
intel_pmu_drain_pebs_core(struct pt_regs * iregs)1420 static void intel_pmu_drain_pebs_core(struct pt_regs *iregs)
1421 {
1422 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1423 	struct debug_store *ds = cpuc->ds;
1424 	struct perf_event *event = cpuc->events[0]; /* PMC0 only */
1425 	struct pebs_record_core *at, *top;
1426 	int n;
1427 
1428 	if (!x86_pmu.pebs_active)
1429 		return;
1430 
1431 	at  = (struct pebs_record_core *)(unsigned long)ds->pebs_buffer_base;
1432 	top = (struct pebs_record_core *)(unsigned long)ds->pebs_index;
1433 
1434 	/*
1435 	 * Whatever else happens, drain the thing
1436 	 */
1437 	ds->pebs_index = ds->pebs_buffer_base;
1438 
1439 	if (!test_bit(0, cpuc->active_mask))
1440 		return;
1441 
1442 	WARN_ON_ONCE(!event);
1443 
1444 	if (!event->attr.precise_ip)
1445 		return;
1446 
1447 	n = top - at;
1448 	if (n <= 0) {
1449 		if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1450 			intel_pmu_save_and_restart_reload(event, 0);
1451 		return;
1452 	}
1453 
1454 	__intel_pmu_pebs_event(event, iregs, at, top, 0, n);
1455 }
1456 
intel_pmu_drain_pebs_nhm(struct pt_regs * iregs)1457 static void intel_pmu_drain_pebs_nhm(struct pt_regs *iregs)
1458 {
1459 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1460 	struct debug_store *ds = cpuc->ds;
1461 	struct perf_event *event;
1462 	void *base, *at, *top;
1463 	short counts[MAX_PEBS_EVENTS] = {};
1464 	short error[MAX_PEBS_EVENTS] = {};
1465 	int bit, i;
1466 
1467 	if (!x86_pmu.pebs_active)
1468 		return;
1469 
1470 	base = (struct pebs_record_nhm *)(unsigned long)ds->pebs_buffer_base;
1471 	top = (struct pebs_record_nhm *)(unsigned long)ds->pebs_index;
1472 
1473 	ds->pebs_index = ds->pebs_buffer_base;
1474 
1475 	if (unlikely(base >= top)) {
1476 		/*
1477 		 * The drain_pebs() could be called twice in a short period
1478 		 * for auto-reload event in pmu::read(). There are no
1479 		 * overflows have happened in between.
1480 		 * It needs to call intel_pmu_save_and_restart_reload() to
1481 		 * update the event->count for this case.
1482 		 */
1483 		for_each_set_bit(bit, (unsigned long *)&cpuc->pebs_enabled,
1484 				 x86_pmu.max_pebs_events) {
1485 			event = cpuc->events[bit];
1486 			if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD)
1487 				intel_pmu_save_and_restart_reload(event, 0);
1488 		}
1489 		return;
1490 	}
1491 
1492 	for (at = base; at < top; at += x86_pmu.pebs_record_size) {
1493 		struct pebs_record_nhm *p = at;
1494 		u64 pebs_status;
1495 
1496 		pebs_status = p->status & cpuc->pebs_enabled;
1497 		pebs_status &= (1ULL << x86_pmu.max_pebs_events) - 1;
1498 
1499 		/* PEBS v3 has more accurate status bits */
1500 		if (x86_pmu.intel_cap.pebs_format >= 3) {
1501 			for_each_set_bit(bit, (unsigned long *)&pebs_status,
1502 					 x86_pmu.max_pebs_events)
1503 				counts[bit]++;
1504 
1505 			continue;
1506 		}
1507 
1508 		/*
1509 		 * On some CPUs the PEBS status can be zero when PEBS is
1510 		 * racing with clearing of GLOBAL_STATUS.
1511 		 *
1512 		 * Normally we would drop that record, but in the
1513 		 * case when there is only a single active PEBS event
1514 		 * we can assume it's for that event.
1515 		 */
1516 		if (!pebs_status && cpuc->pebs_enabled &&
1517 			!(cpuc->pebs_enabled & (cpuc->pebs_enabled-1)))
1518 			pebs_status = cpuc->pebs_enabled;
1519 
1520 		bit = find_first_bit((unsigned long *)&pebs_status,
1521 					x86_pmu.max_pebs_events);
1522 		if (bit >= x86_pmu.max_pebs_events)
1523 			continue;
1524 
1525 		/*
1526 		 * The PEBS hardware does not deal well with the situation
1527 		 * when events happen near to each other and multiple bits
1528 		 * are set. But it should happen rarely.
1529 		 *
1530 		 * If these events include one PEBS and multiple non-PEBS
1531 		 * events, it doesn't impact PEBS record. The record will
1532 		 * be handled normally. (slow path)
1533 		 *
1534 		 * If these events include two or more PEBS events, the
1535 		 * records for the events can be collapsed into a single
1536 		 * one, and it's not possible to reconstruct all events
1537 		 * that caused the PEBS record. It's called collision.
1538 		 * If collision happened, the record will be dropped.
1539 		 */
1540 		if (p->status != (1ULL << bit)) {
1541 			for_each_set_bit(i, (unsigned long *)&pebs_status,
1542 					 x86_pmu.max_pebs_events)
1543 				error[i]++;
1544 			continue;
1545 		}
1546 
1547 		counts[bit]++;
1548 	}
1549 
1550 	for (bit = 0; bit < x86_pmu.max_pebs_events; bit++) {
1551 		if ((counts[bit] == 0) && (error[bit] == 0))
1552 			continue;
1553 
1554 		event = cpuc->events[bit];
1555 		if (WARN_ON_ONCE(!event))
1556 			continue;
1557 
1558 		if (WARN_ON_ONCE(!event->attr.precise_ip))
1559 			continue;
1560 
1561 		/* log dropped samples number */
1562 		if (error[bit]) {
1563 			perf_log_lost_samples(event, error[bit]);
1564 
1565 			if (perf_event_account_interrupt(event))
1566 				x86_pmu_stop(event, 0);
1567 		}
1568 
1569 		if (counts[bit]) {
1570 			__intel_pmu_pebs_event(event, iregs, base,
1571 					       top, bit, counts[bit]);
1572 		}
1573 	}
1574 }
1575 
1576 /*
1577  * BTS, PEBS probe and setup
1578  */
1579 
intel_ds_init(void)1580 void __init intel_ds_init(void)
1581 {
1582 	/*
1583 	 * No support for 32bit formats
1584 	 */
1585 	if (!boot_cpu_has(X86_FEATURE_DTES64))
1586 		return;
1587 
1588 	x86_pmu.bts  = boot_cpu_has(X86_FEATURE_BTS);
1589 	x86_pmu.pebs = boot_cpu_has(X86_FEATURE_PEBS);
1590 	x86_pmu.pebs_buffer_size = PEBS_BUFFER_SIZE;
1591 	if (x86_pmu.pebs) {
1592 		char pebs_type = x86_pmu.intel_cap.pebs_trap ?  '+' : '-';
1593 		int format = x86_pmu.intel_cap.pebs_format;
1594 
1595 		switch (format) {
1596 		case 0:
1597 			pr_cont("PEBS fmt0%c, ", pebs_type);
1598 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_core);
1599 			/*
1600 			 * Using >PAGE_SIZE buffers makes the WRMSR to
1601 			 * PERF_GLOBAL_CTRL in intel_pmu_enable_all()
1602 			 * mysteriously hang on Core2.
1603 			 *
1604 			 * As a workaround, we don't do this.
1605 			 */
1606 			x86_pmu.pebs_buffer_size = PAGE_SIZE;
1607 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_core;
1608 			break;
1609 
1610 		case 1:
1611 			pr_cont("PEBS fmt1%c, ", pebs_type);
1612 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_nhm);
1613 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1614 			break;
1615 
1616 		case 2:
1617 			pr_cont("PEBS fmt2%c, ", pebs_type);
1618 			x86_pmu.pebs_record_size = sizeof(struct pebs_record_hsw);
1619 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1620 			break;
1621 
1622 		case 3:
1623 			pr_cont("PEBS fmt3%c, ", pebs_type);
1624 			x86_pmu.pebs_record_size =
1625 						sizeof(struct pebs_record_skl);
1626 			x86_pmu.drain_pebs = intel_pmu_drain_pebs_nhm;
1627 			x86_pmu.free_running_flags |= PERF_SAMPLE_TIME;
1628 			break;
1629 
1630 		default:
1631 			pr_cont("no PEBS fmt%d%c, ", format, pebs_type);
1632 			x86_pmu.pebs = 0;
1633 		}
1634 	}
1635 }
1636 
perf_restore_debug_store(void)1637 void perf_restore_debug_store(void)
1638 {
1639 	struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
1640 
1641 	if (!x86_pmu.bts && !x86_pmu.pebs)
1642 		return;
1643 
1644 	wrmsrl(MSR_IA32_DS_AREA, (unsigned long)ds);
1645 }
1646