xref: /linux/arch/x86/events/intel/lbr.c (revision 1553a1c48281243359a9529a10ddb551f3b967ab)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/perf_event.h>
3 #include <linux/types.h>
4 
5 #include <asm/perf_event.h>
6 #include <asm/msr.h>
7 
8 #include "../perf_event.h"
9 
10 /*
11  * Intel LBR_SELECT bits
12  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
13  *
14  * Hardware branch filter (not available on all CPUs)
15  */
16 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
17 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
18 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
19 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
20 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
21 #define LBR_RETURN_BIT		5 /* do not capture near returns */
22 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
23 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
24 #define LBR_FAR_BIT		8 /* do not capture far branches */
25 #define LBR_CALL_STACK_BIT	9 /* enable call stack */
26 
27 /*
28  * Following bit only exists in Linux; we mask it out before writing it to
29  * the actual MSR. But it helps the constraint perf code to understand
30  * that this is a separate configuration.
31  */
32 #define LBR_NO_INFO_BIT	       63 /* don't read LBR_INFO. */
33 
34 #define LBR_KERNEL	(1 << LBR_KERNEL_BIT)
35 #define LBR_USER	(1 << LBR_USER_BIT)
36 #define LBR_JCC		(1 << LBR_JCC_BIT)
37 #define LBR_REL_CALL	(1 << LBR_REL_CALL_BIT)
38 #define LBR_IND_CALL	(1 << LBR_IND_CALL_BIT)
39 #define LBR_RETURN	(1 << LBR_RETURN_BIT)
40 #define LBR_REL_JMP	(1 << LBR_REL_JMP_BIT)
41 #define LBR_IND_JMP	(1 << LBR_IND_JMP_BIT)
42 #define LBR_FAR		(1 << LBR_FAR_BIT)
43 #define LBR_CALL_STACK	(1 << LBR_CALL_STACK_BIT)
44 #define LBR_NO_INFO	(1ULL << LBR_NO_INFO_BIT)
45 
46 #define LBR_PLM (LBR_KERNEL | LBR_USER)
47 
48 #define LBR_SEL_MASK	0x3ff	/* valid bits in LBR_SELECT */
49 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
50 #define LBR_IGN		0	/* ignored */
51 
52 #define LBR_ANY		 \
53 	(LBR_JCC	|\
54 	 LBR_REL_CALL	|\
55 	 LBR_IND_CALL	|\
56 	 LBR_RETURN	|\
57 	 LBR_REL_JMP	|\
58 	 LBR_IND_JMP	|\
59 	 LBR_FAR)
60 
61 #define LBR_FROM_FLAG_MISPRED	BIT_ULL(63)
62 #define LBR_FROM_FLAG_IN_TX	BIT_ULL(62)
63 #define LBR_FROM_FLAG_ABORT	BIT_ULL(61)
64 
65 #define LBR_FROM_SIGNEXT_2MSB	(BIT_ULL(60) | BIT_ULL(59))
66 
67 /*
68  * Intel LBR_CTL bits
69  *
70  * Hardware branch filter for Arch LBR
71  */
72 #define ARCH_LBR_KERNEL_BIT		1  /* capture at ring0 */
73 #define ARCH_LBR_USER_BIT		2  /* capture at ring > 0 */
74 #define ARCH_LBR_CALL_STACK_BIT		3  /* enable call stack */
75 #define ARCH_LBR_JCC_BIT		16 /* capture conditional branches */
76 #define ARCH_LBR_REL_JMP_BIT		17 /* capture relative jumps */
77 #define ARCH_LBR_IND_JMP_BIT		18 /* capture indirect jumps */
78 #define ARCH_LBR_REL_CALL_BIT		19 /* capture relative calls */
79 #define ARCH_LBR_IND_CALL_BIT		20 /* capture indirect calls */
80 #define ARCH_LBR_RETURN_BIT		21 /* capture near returns */
81 #define ARCH_LBR_OTHER_BRANCH_BIT	22 /* capture other branches */
82 
83 #define ARCH_LBR_KERNEL			(1ULL << ARCH_LBR_KERNEL_BIT)
84 #define ARCH_LBR_USER			(1ULL << ARCH_LBR_USER_BIT)
85 #define ARCH_LBR_CALL_STACK		(1ULL << ARCH_LBR_CALL_STACK_BIT)
86 #define ARCH_LBR_JCC			(1ULL << ARCH_LBR_JCC_BIT)
87 #define ARCH_LBR_REL_JMP		(1ULL << ARCH_LBR_REL_JMP_BIT)
88 #define ARCH_LBR_IND_JMP		(1ULL << ARCH_LBR_IND_JMP_BIT)
89 #define ARCH_LBR_REL_CALL		(1ULL << ARCH_LBR_REL_CALL_BIT)
90 #define ARCH_LBR_IND_CALL		(1ULL << ARCH_LBR_IND_CALL_BIT)
91 #define ARCH_LBR_RETURN			(1ULL << ARCH_LBR_RETURN_BIT)
92 #define ARCH_LBR_OTHER_BRANCH		(1ULL << ARCH_LBR_OTHER_BRANCH_BIT)
93 
94 #define ARCH_LBR_ANY			 \
95 	(ARCH_LBR_JCC			|\
96 	 ARCH_LBR_REL_JMP		|\
97 	 ARCH_LBR_IND_JMP		|\
98 	 ARCH_LBR_REL_CALL		|\
99 	 ARCH_LBR_IND_CALL		|\
100 	 ARCH_LBR_RETURN		|\
101 	 ARCH_LBR_OTHER_BRANCH)
102 
103 #define ARCH_LBR_CTL_MASK			0x7f000e
104 
105 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
106 
107 static __always_inline bool is_lbr_call_stack_bit_set(u64 config)
108 {
109 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
110 		return !!(config & ARCH_LBR_CALL_STACK);
111 
112 	return !!(config & LBR_CALL_STACK);
113 }
114 
115 /*
116  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
117  * otherwise it becomes near impossible to get a reliable stack.
118  */
119 
120 static void __intel_pmu_lbr_enable(bool pmi)
121 {
122 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
123 	u64 debugctl, lbr_select = 0, orig_debugctl;
124 
125 	/*
126 	 * No need to unfreeze manually, as v4 can do that as part
127 	 * of the GLOBAL_STATUS ack.
128 	 */
129 	if (pmi && x86_pmu.version >= 4)
130 		return;
131 
132 	/*
133 	 * No need to reprogram LBR_SELECT in a PMI, as it
134 	 * did not change.
135 	 */
136 	if (cpuc->lbr_sel)
137 		lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
138 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) && !pmi && cpuc->lbr_sel)
139 		wrmsrl(MSR_LBR_SELECT, lbr_select);
140 
141 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
142 	orig_debugctl = debugctl;
143 
144 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
145 		debugctl |= DEBUGCTLMSR_LBR;
146 	/*
147 	 * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
148 	 * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
149 	 * may cause superfluous increase/decrease of LBR_TOS.
150 	 */
151 	if (is_lbr_call_stack_bit_set(lbr_select))
152 		debugctl &= ~DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
153 	else
154 		debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
155 
156 	if (orig_debugctl != debugctl)
157 		wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
158 
159 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
160 		wrmsrl(MSR_ARCH_LBR_CTL, lbr_select | ARCH_LBR_CTL_LBREN);
161 }
162 
163 void intel_pmu_lbr_reset_32(void)
164 {
165 	int i;
166 
167 	for (i = 0; i < x86_pmu.lbr_nr; i++)
168 		wrmsrl(x86_pmu.lbr_from + i, 0);
169 }
170 
171 void intel_pmu_lbr_reset_64(void)
172 {
173 	int i;
174 
175 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
176 		wrmsrl(x86_pmu.lbr_from + i, 0);
177 		wrmsrl(x86_pmu.lbr_to   + i, 0);
178 		if (x86_pmu.lbr_has_info)
179 			wrmsrl(x86_pmu.lbr_info + i, 0);
180 	}
181 }
182 
183 static void intel_pmu_arch_lbr_reset(void)
184 {
185 	/* Write to ARCH_LBR_DEPTH MSR, all LBR entries are reset to 0 */
186 	wrmsrl(MSR_ARCH_LBR_DEPTH, x86_pmu.lbr_nr);
187 }
188 
189 void intel_pmu_lbr_reset(void)
190 {
191 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
192 
193 	if (!x86_pmu.lbr_nr)
194 		return;
195 
196 	x86_pmu.lbr_reset();
197 
198 	cpuc->last_task_ctx = NULL;
199 	cpuc->last_log_id = 0;
200 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) && cpuc->lbr_select)
201 		wrmsrl(MSR_LBR_SELECT, 0);
202 }
203 
204 /*
205  * TOS = most recently recorded branch
206  */
207 static inline u64 intel_pmu_lbr_tos(void)
208 {
209 	u64 tos;
210 
211 	rdmsrl(x86_pmu.lbr_tos, tos);
212 	return tos;
213 }
214 
215 enum {
216 	LBR_NONE,
217 	LBR_VALID,
218 };
219 
220 /*
221  * For format LBR_FORMAT_EIP_FLAGS2, bits 61:62 in MSR_LAST_BRANCH_FROM_x
222  * are the TSX flags when TSX is supported, but when TSX is not supported
223  * they have no consistent behavior:
224  *
225  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
226  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
227  *     part of the sign extension.
228  *
229  * Therefore, if:
230  *
231  *   1) LBR format LBR_FORMAT_EIP_FLAGS2
232  *   2) CPU has no TSX support enabled
233  *
234  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
235  * value from rdmsr() must be converted to have a 61 bits sign extension,
236  * ignoring the TSX flags.
237  */
238 static inline bool lbr_from_signext_quirk_needed(void)
239 {
240 	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
241 			   boot_cpu_has(X86_FEATURE_RTM);
242 
243 	return !tsx_support;
244 }
245 
246 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
247 
248 /* If quirk is enabled, ensure sign extension is 63 bits: */
249 inline u64 lbr_from_signext_quirk_wr(u64 val)
250 {
251 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
252 		/*
253 		 * Sign extend into bits 61:62 while preserving bit 63.
254 		 *
255 		 * Quirk is enabled when TSX is disabled. Therefore TSX bits
256 		 * in val are always OFF and must be changed to be sign
257 		 * extension bits. Since bits 59:60 are guaranteed to be
258 		 * part of the sign extension bits, we can just copy them
259 		 * to 61:62.
260 		 */
261 		val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
262 	}
263 	return val;
264 }
265 
266 /*
267  * If quirk is needed, ensure sign extension is 61 bits:
268  */
269 static u64 lbr_from_signext_quirk_rd(u64 val)
270 {
271 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
272 		/*
273 		 * Quirk is on when TSX is not enabled. Therefore TSX
274 		 * flags must be read as OFF.
275 		 */
276 		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
277 	}
278 	return val;
279 }
280 
281 static __always_inline void wrlbr_from(unsigned int idx, u64 val)
282 {
283 	val = lbr_from_signext_quirk_wr(val);
284 	wrmsrl(x86_pmu.lbr_from + idx, val);
285 }
286 
287 static __always_inline void wrlbr_to(unsigned int idx, u64 val)
288 {
289 	wrmsrl(x86_pmu.lbr_to + idx, val);
290 }
291 
292 static __always_inline void wrlbr_info(unsigned int idx, u64 val)
293 {
294 	wrmsrl(x86_pmu.lbr_info + idx, val);
295 }
296 
297 static __always_inline u64 rdlbr_from(unsigned int idx, struct lbr_entry *lbr)
298 {
299 	u64 val;
300 
301 	if (lbr)
302 		return lbr->from;
303 
304 	rdmsrl(x86_pmu.lbr_from + idx, val);
305 
306 	return lbr_from_signext_quirk_rd(val);
307 }
308 
309 static __always_inline u64 rdlbr_to(unsigned int idx, struct lbr_entry *lbr)
310 {
311 	u64 val;
312 
313 	if (lbr)
314 		return lbr->to;
315 
316 	rdmsrl(x86_pmu.lbr_to + idx, val);
317 
318 	return val;
319 }
320 
321 static __always_inline u64 rdlbr_info(unsigned int idx, struct lbr_entry *lbr)
322 {
323 	u64 val;
324 
325 	if (lbr)
326 		return lbr->info;
327 
328 	rdmsrl(x86_pmu.lbr_info + idx, val);
329 
330 	return val;
331 }
332 
333 static inline void
334 wrlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
335 {
336 	wrlbr_from(idx, lbr->from);
337 	wrlbr_to(idx, lbr->to);
338 	if (need_info)
339 		wrlbr_info(idx, lbr->info);
340 }
341 
342 static inline bool
343 rdlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
344 {
345 	u64 from = rdlbr_from(idx, NULL);
346 
347 	/* Don't read invalid entry */
348 	if (!from)
349 		return false;
350 
351 	lbr->from = from;
352 	lbr->to = rdlbr_to(idx, NULL);
353 	if (need_info)
354 		lbr->info = rdlbr_info(idx, NULL);
355 
356 	return true;
357 }
358 
359 void intel_pmu_lbr_restore(void *ctx)
360 {
361 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
362 	struct x86_perf_task_context *task_ctx = ctx;
363 	bool need_info = x86_pmu.lbr_has_info;
364 	u64 tos = task_ctx->tos;
365 	unsigned lbr_idx, mask;
366 	int i;
367 
368 	mask = x86_pmu.lbr_nr - 1;
369 	for (i = 0; i < task_ctx->valid_lbrs; i++) {
370 		lbr_idx = (tos - i) & mask;
371 		wrlbr_all(&task_ctx->lbr[i], lbr_idx, need_info);
372 	}
373 
374 	for (; i < x86_pmu.lbr_nr; i++) {
375 		lbr_idx = (tos - i) & mask;
376 		wrlbr_from(lbr_idx, 0);
377 		wrlbr_to(lbr_idx, 0);
378 		if (need_info)
379 			wrlbr_info(lbr_idx, 0);
380 	}
381 
382 	wrmsrl(x86_pmu.lbr_tos, tos);
383 
384 	if (cpuc->lbr_select)
385 		wrmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
386 }
387 
388 static void intel_pmu_arch_lbr_restore(void *ctx)
389 {
390 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
391 	struct lbr_entry *entries = task_ctx->entries;
392 	int i;
393 
394 	/* Fast reset the LBRs before restore if the call stack is not full. */
395 	if (!entries[x86_pmu.lbr_nr - 1].from)
396 		intel_pmu_arch_lbr_reset();
397 
398 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
399 		if (!entries[i].from)
400 			break;
401 		wrlbr_all(&entries[i], i, true);
402 	}
403 }
404 
405 /*
406  * Restore the Architecture LBR state from the xsave area in the perf
407  * context data for the task via the XRSTORS instruction.
408  */
409 static void intel_pmu_arch_lbr_xrstors(void *ctx)
410 {
411 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
412 
413 	xrstors(&task_ctx->xsave, XFEATURE_MASK_LBR);
414 }
415 
416 static __always_inline bool lbr_is_reset_in_cstate(void *ctx)
417 {
418 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
419 		return x86_pmu.lbr_deep_c_reset && !rdlbr_from(0, NULL);
420 
421 	return !rdlbr_from(((struct x86_perf_task_context *)ctx)->tos, NULL);
422 }
423 
424 static void __intel_pmu_lbr_restore(void *ctx)
425 {
426 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
427 
428 	if (task_context_opt(ctx)->lbr_callstack_users == 0 ||
429 	    task_context_opt(ctx)->lbr_stack_state == LBR_NONE) {
430 		intel_pmu_lbr_reset();
431 		return;
432 	}
433 
434 	/*
435 	 * Does not restore the LBR registers, if
436 	 * - No one else touched them, and
437 	 * - Was not cleared in Cstate
438 	 */
439 	if ((ctx == cpuc->last_task_ctx) &&
440 	    (task_context_opt(ctx)->log_id == cpuc->last_log_id) &&
441 	    !lbr_is_reset_in_cstate(ctx)) {
442 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
443 		return;
444 	}
445 
446 	x86_pmu.lbr_restore(ctx);
447 
448 	task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
449 }
450 
451 void intel_pmu_lbr_save(void *ctx)
452 {
453 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
454 	struct x86_perf_task_context *task_ctx = ctx;
455 	bool need_info = x86_pmu.lbr_has_info;
456 	unsigned lbr_idx, mask;
457 	u64 tos;
458 	int i;
459 
460 	mask = x86_pmu.lbr_nr - 1;
461 	tos = intel_pmu_lbr_tos();
462 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
463 		lbr_idx = (tos - i) & mask;
464 		if (!rdlbr_all(&task_ctx->lbr[i], lbr_idx, need_info))
465 			break;
466 	}
467 	task_ctx->valid_lbrs = i;
468 	task_ctx->tos = tos;
469 
470 	if (cpuc->lbr_select)
471 		rdmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
472 }
473 
474 static void intel_pmu_arch_lbr_save(void *ctx)
475 {
476 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
477 	struct lbr_entry *entries = task_ctx->entries;
478 	int i;
479 
480 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
481 		if (!rdlbr_all(&entries[i], i, true))
482 			break;
483 	}
484 
485 	/* LBR call stack is not full. Reset is required in restore. */
486 	if (i < x86_pmu.lbr_nr)
487 		entries[x86_pmu.lbr_nr - 1].from = 0;
488 }
489 
490 /*
491  * Save the Architecture LBR state to the xsave area in the perf
492  * context data for the task via the XSAVES instruction.
493  */
494 static void intel_pmu_arch_lbr_xsaves(void *ctx)
495 {
496 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
497 
498 	xsaves(&task_ctx->xsave, XFEATURE_MASK_LBR);
499 }
500 
501 static void __intel_pmu_lbr_save(void *ctx)
502 {
503 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
504 
505 	if (task_context_opt(ctx)->lbr_callstack_users == 0) {
506 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
507 		return;
508 	}
509 
510 	x86_pmu.lbr_save(ctx);
511 
512 	task_context_opt(ctx)->lbr_stack_state = LBR_VALID;
513 
514 	cpuc->last_task_ctx = ctx;
515 	cpuc->last_log_id = ++task_context_opt(ctx)->log_id;
516 }
517 
518 void intel_pmu_lbr_swap_task_ctx(struct perf_event_pmu_context *prev_epc,
519 				 struct perf_event_pmu_context *next_epc)
520 {
521 	void *prev_ctx_data, *next_ctx_data;
522 
523 	swap(prev_epc->task_ctx_data, next_epc->task_ctx_data);
524 
525 	/*
526 	 * Architecture specific synchronization makes sense in case
527 	 * both prev_epc->task_ctx_data and next_epc->task_ctx_data
528 	 * pointers are allocated.
529 	 */
530 
531 	prev_ctx_data = next_epc->task_ctx_data;
532 	next_ctx_data = prev_epc->task_ctx_data;
533 
534 	if (!prev_ctx_data || !next_ctx_data)
535 		return;
536 
537 	swap(task_context_opt(prev_ctx_data)->lbr_callstack_users,
538 	     task_context_opt(next_ctx_data)->lbr_callstack_users);
539 }
540 
541 void intel_pmu_lbr_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
542 {
543 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
544 	void *task_ctx;
545 
546 	if (!cpuc->lbr_users)
547 		return;
548 
549 	/*
550 	 * If LBR callstack feature is enabled and the stack was saved when
551 	 * the task was scheduled out, restore the stack. Otherwise flush
552 	 * the LBR stack.
553 	 */
554 	task_ctx = pmu_ctx ? pmu_ctx->task_ctx_data : NULL;
555 	if (task_ctx) {
556 		if (sched_in)
557 			__intel_pmu_lbr_restore(task_ctx);
558 		else
559 			__intel_pmu_lbr_save(task_ctx);
560 		return;
561 	}
562 
563 	/*
564 	 * Since a context switch can flip the address space and LBR entries
565 	 * are not tagged with an identifier, we need to wipe the LBR, even for
566 	 * per-cpu events. You simply cannot resolve the branches from the old
567 	 * address space.
568 	 */
569 	if (sched_in)
570 		intel_pmu_lbr_reset();
571 }
572 
573 static inline bool branch_user_callstack(unsigned br_sel)
574 {
575 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
576 }
577 
578 void intel_pmu_lbr_add(struct perf_event *event)
579 {
580 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
581 
582 	if (!x86_pmu.lbr_nr)
583 		return;
584 
585 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
586 		cpuc->lbr_select = 1;
587 
588 	cpuc->br_sel = event->hw.branch_reg.reg;
589 
590 	if (branch_user_callstack(cpuc->br_sel) && event->pmu_ctx->task_ctx_data)
591 		task_context_opt(event->pmu_ctx->task_ctx_data)->lbr_callstack_users++;
592 
593 	/*
594 	 * Request pmu::sched_task() callback, which will fire inside the
595 	 * regular perf event scheduling, so that call will:
596 	 *
597 	 *  - restore or wipe; when LBR-callstack,
598 	 *  - wipe; otherwise,
599 	 *
600 	 * when this is from __perf_event_task_sched_in().
601 	 *
602 	 * However, if this is from perf_install_in_context(), no such callback
603 	 * will follow and we'll need to reset the LBR here if this is the
604 	 * first LBR event.
605 	 *
606 	 * The problem is, we cannot tell these cases apart... but we can
607 	 * exclude the biggest chunk of cases by looking at
608 	 * event->total_time_running. An event that has accrued runtime cannot
609 	 * be 'new'. Conversely, a new event can get installed through the
610 	 * context switch path for the first time.
611 	 */
612 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
613 		cpuc->lbr_pebs_users++;
614 	perf_sched_cb_inc(event->pmu);
615 	if (!cpuc->lbr_users++ && !event->total_time_running)
616 		intel_pmu_lbr_reset();
617 }
618 
619 void release_lbr_buffers(void)
620 {
621 	struct kmem_cache *kmem_cache;
622 	struct cpu_hw_events *cpuc;
623 	int cpu;
624 
625 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
626 		return;
627 
628 	for_each_possible_cpu(cpu) {
629 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
630 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
631 		if (kmem_cache && cpuc->lbr_xsave) {
632 			kmem_cache_free(kmem_cache, cpuc->lbr_xsave);
633 			cpuc->lbr_xsave = NULL;
634 		}
635 	}
636 }
637 
638 void reserve_lbr_buffers(void)
639 {
640 	struct kmem_cache *kmem_cache;
641 	struct cpu_hw_events *cpuc;
642 	int cpu;
643 
644 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
645 		return;
646 
647 	for_each_possible_cpu(cpu) {
648 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
649 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
650 		if (!kmem_cache || cpuc->lbr_xsave)
651 			continue;
652 
653 		cpuc->lbr_xsave = kmem_cache_alloc_node(kmem_cache,
654 							GFP_KERNEL | __GFP_ZERO,
655 							cpu_to_node(cpu));
656 	}
657 }
658 
659 void intel_pmu_lbr_del(struct perf_event *event)
660 {
661 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
662 
663 	if (!x86_pmu.lbr_nr)
664 		return;
665 
666 	if (branch_user_callstack(cpuc->br_sel) &&
667 	    event->pmu_ctx->task_ctx_data)
668 		task_context_opt(event->pmu_ctx->task_ctx_data)->lbr_callstack_users--;
669 
670 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
671 		cpuc->lbr_select = 0;
672 
673 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
674 		cpuc->lbr_pebs_users--;
675 	cpuc->lbr_users--;
676 	WARN_ON_ONCE(cpuc->lbr_users < 0);
677 	WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
678 	perf_sched_cb_dec(event->pmu);
679 
680 	/*
681 	 * The logged occurrences information is only valid for the
682 	 * current LBR group. If another LBR group is scheduled in
683 	 * later, the information from the stale LBRs will be wrongly
684 	 * interpreted. Reset the LBRs here.
685 	 *
686 	 * Only clear once for a branch counter group with the leader
687 	 * event. Because
688 	 * - Cannot simply reset the LBRs with the !cpuc->lbr_users.
689 	 *   Because it's possible that the last LBR user is not in a
690 	 *   branch counter group, e.g., a branch_counters group +
691 	 *   several normal LBR events.
692 	 * - The LBR reset can be done with any one of the events in a
693 	 *   branch counter group, since they are always scheduled together.
694 	 *   It's easy to force the leader event an LBR event.
695 	 */
696 	if (is_branch_counters_group(event) && event == event->group_leader)
697 		intel_pmu_lbr_reset();
698 }
699 
700 static inline bool vlbr_exclude_host(void)
701 {
702 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
703 
704 	return test_bit(INTEL_PMC_IDX_FIXED_VLBR,
705 		(unsigned long *)&cpuc->intel_ctrl_guest_mask);
706 }
707 
708 void intel_pmu_lbr_enable_all(bool pmi)
709 {
710 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
711 
712 	if (cpuc->lbr_users && !vlbr_exclude_host())
713 		__intel_pmu_lbr_enable(pmi);
714 }
715 
716 void intel_pmu_lbr_disable_all(void)
717 {
718 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
719 
720 	if (cpuc->lbr_users && !vlbr_exclude_host()) {
721 		if (static_cpu_has(X86_FEATURE_ARCH_LBR))
722 			return __intel_pmu_arch_lbr_disable();
723 
724 		__intel_pmu_lbr_disable();
725 	}
726 }
727 
728 void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
729 {
730 	unsigned long mask = x86_pmu.lbr_nr - 1;
731 	struct perf_branch_entry *br = cpuc->lbr_entries;
732 	u64 tos = intel_pmu_lbr_tos();
733 	int i;
734 
735 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
736 		unsigned long lbr_idx = (tos - i) & mask;
737 		union {
738 			struct {
739 				u32 from;
740 				u32 to;
741 			};
742 			u64     lbr;
743 		} msr_lastbranch;
744 
745 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
746 
747 		perf_clear_branch_entry_bitfields(br);
748 
749 		br->from	= msr_lastbranch.from;
750 		br->to		= msr_lastbranch.to;
751 		br++;
752 	}
753 	cpuc->lbr_stack.nr = i;
754 	cpuc->lbr_stack.hw_idx = tos;
755 }
756 
757 /*
758  * Due to lack of segmentation in Linux the effective address (offset)
759  * is the same as the linear address, allowing us to merge the LIP and EIP
760  * LBR formats.
761  */
762 void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
763 {
764 	bool need_info = false, call_stack = false;
765 	unsigned long mask = x86_pmu.lbr_nr - 1;
766 	struct perf_branch_entry *br = cpuc->lbr_entries;
767 	u64 tos = intel_pmu_lbr_tos();
768 	int i;
769 	int out = 0;
770 	int num = x86_pmu.lbr_nr;
771 
772 	if (cpuc->lbr_sel) {
773 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
774 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
775 			call_stack = true;
776 	}
777 
778 	for (i = 0; i < num; i++) {
779 		unsigned long lbr_idx = (tos - i) & mask;
780 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
781 		u16 cycles = 0;
782 
783 		from = rdlbr_from(lbr_idx, NULL);
784 		to   = rdlbr_to(lbr_idx, NULL);
785 
786 		/*
787 		 * Read LBR call stack entries
788 		 * until invalid entry (0s) is detected.
789 		 */
790 		if (call_stack && !from)
791 			break;
792 
793 		if (x86_pmu.lbr_has_info) {
794 			if (need_info) {
795 				u64 info;
796 
797 				info = rdlbr_info(lbr_idx, NULL);
798 				mis = !!(info & LBR_INFO_MISPRED);
799 				pred = !mis;
800 				cycles = (info & LBR_INFO_CYCLES);
801 				if (x86_pmu.lbr_has_tsx) {
802 					in_tx = !!(info & LBR_INFO_IN_TX);
803 					abort = !!(info & LBR_INFO_ABORT);
804 				}
805 			}
806 		} else {
807 			int skip = 0;
808 
809 			if (x86_pmu.lbr_from_flags) {
810 				mis = !!(from & LBR_FROM_FLAG_MISPRED);
811 				pred = !mis;
812 				skip = 1;
813 			}
814 			if (x86_pmu.lbr_has_tsx) {
815 				in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
816 				abort = !!(from & LBR_FROM_FLAG_ABORT);
817 				skip = 3;
818 			}
819 			from = (u64)((((s64)from) << skip) >> skip);
820 
821 			if (x86_pmu.lbr_to_cycles) {
822 				cycles = ((to >> 48) & LBR_INFO_CYCLES);
823 				to = (u64)((((s64)to) << 16) >> 16);
824 			}
825 		}
826 
827 		/*
828 		 * Some CPUs report duplicated abort records,
829 		 * with the second entry not having an abort bit set.
830 		 * Skip them here. This loop runs backwards,
831 		 * so we need to undo the previous record.
832 		 * If the abort just happened outside the window
833 		 * the extra entry cannot be removed.
834 		 */
835 		if (abort && x86_pmu.lbr_double_abort && out > 0)
836 			out--;
837 
838 		perf_clear_branch_entry_bitfields(br+out);
839 		br[out].from	 = from;
840 		br[out].to	 = to;
841 		br[out].mispred	 = mis;
842 		br[out].predicted = pred;
843 		br[out].in_tx	 = in_tx;
844 		br[out].abort	 = abort;
845 		br[out].cycles	 = cycles;
846 		out++;
847 	}
848 	cpuc->lbr_stack.nr = out;
849 	cpuc->lbr_stack.hw_idx = tos;
850 }
851 
852 static DEFINE_STATIC_KEY_FALSE(x86_lbr_mispred);
853 static DEFINE_STATIC_KEY_FALSE(x86_lbr_cycles);
854 static DEFINE_STATIC_KEY_FALSE(x86_lbr_type);
855 
856 static __always_inline int get_lbr_br_type(u64 info)
857 {
858 	int type = 0;
859 
860 	if (static_branch_likely(&x86_lbr_type))
861 		type = (info & LBR_INFO_BR_TYPE) >> LBR_INFO_BR_TYPE_OFFSET;
862 
863 	return type;
864 }
865 
866 static __always_inline bool get_lbr_mispred(u64 info)
867 {
868 	bool mispred = 0;
869 
870 	if (static_branch_likely(&x86_lbr_mispred))
871 		mispred = !!(info & LBR_INFO_MISPRED);
872 
873 	return mispred;
874 }
875 
876 static __always_inline u16 get_lbr_cycles(u64 info)
877 {
878 	u16 cycles = info & LBR_INFO_CYCLES;
879 
880 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
881 	    (!static_branch_likely(&x86_lbr_cycles) ||
882 	     !(info & LBR_INFO_CYC_CNT_VALID)))
883 		cycles = 0;
884 
885 	return cycles;
886 }
887 
888 static_assert((64 - PERF_BRANCH_ENTRY_INFO_BITS_MAX) > LBR_INFO_BR_CNTR_NUM * LBR_INFO_BR_CNTR_BITS);
889 
890 static void intel_pmu_store_lbr(struct cpu_hw_events *cpuc,
891 				struct lbr_entry *entries)
892 {
893 	struct perf_branch_entry *e;
894 	struct lbr_entry *lbr;
895 	u64 from, to, info;
896 	int i;
897 
898 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
899 		lbr = entries ? &entries[i] : NULL;
900 		e = &cpuc->lbr_entries[i];
901 
902 		from = rdlbr_from(i, lbr);
903 		/*
904 		 * Read LBR entries until invalid entry (0s) is detected.
905 		 */
906 		if (!from)
907 			break;
908 
909 		to = rdlbr_to(i, lbr);
910 		info = rdlbr_info(i, lbr);
911 
912 		perf_clear_branch_entry_bitfields(e);
913 
914 		e->from		= from;
915 		e->to		= to;
916 		e->mispred	= get_lbr_mispred(info);
917 		e->predicted	= !e->mispred;
918 		e->in_tx	= !!(info & LBR_INFO_IN_TX);
919 		e->abort	= !!(info & LBR_INFO_ABORT);
920 		e->cycles	= get_lbr_cycles(info);
921 		e->type		= get_lbr_br_type(info);
922 
923 		/*
924 		 * Leverage the reserved field of cpuc->lbr_entries[i] to
925 		 * temporarily store the branch counters information.
926 		 * The later code will decide what content can be disclosed
927 		 * to the perf tool. Pleae see intel_pmu_lbr_counters_reorder().
928 		 */
929 		e->reserved	= (info >> LBR_INFO_BR_CNTR_OFFSET) & LBR_INFO_BR_CNTR_FULL_MASK;
930 	}
931 
932 	cpuc->lbr_stack.nr = i;
933 }
934 
935 /*
936  * The enabled order may be different from the counter order.
937  * Update the lbr_counters with the enabled order.
938  */
939 static void intel_pmu_lbr_counters_reorder(struct cpu_hw_events *cpuc,
940 					   struct perf_event *event)
941 {
942 	int i, j, pos = 0, order[X86_PMC_IDX_MAX];
943 	struct perf_event *leader, *sibling;
944 	u64 src, dst, cnt;
945 
946 	leader = event->group_leader;
947 	if (branch_sample_counters(leader))
948 		order[pos++] = leader->hw.idx;
949 
950 	for_each_sibling_event(sibling, leader) {
951 		if (!branch_sample_counters(sibling))
952 			continue;
953 		order[pos++] = sibling->hw.idx;
954 	}
955 
956 	WARN_ON_ONCE(!pos);
957 
958 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
959 		src = cpuc->lbr_entries[i].reserved;
960 		dst = 0;
961 		for (j = 0; j < pos; j++) {
962 			cnt = (src >> (order[j] * LBR_INFO_BR_CNTR_BITS)) & LBR_INFO_BR_CNTR_MASK;
963 			dst |= cnt << j * LBR_INFO_BR_CNTR_BITS;
964 		}
965 		cpuc->lbr_counters[i] = dst;
966 		cpuc->lbr_entries[i].reserved = 0;
967 	}
968 }
969 
970 void intel_pmu_lbr_save_brstack(struct perf_sample_data *data,
971 				struct cpu_hw_events *cpuc,
972 				struct perf_event *event)
973 {
974 	if (is_branch_counters_group(event)) {
975 		intel_pmu_lbr_counters_reorder(cpuc, event);
976 		perf_sample_save_brstack(data, event, &cpuc->lbr_stack, cpuc->lbr_counters);
977 		return;
978 	}
979 
980 	perf_sample_save_brstack(data, event, &cpuc->lbr_stack, NULL);
981 }
982 
983 static void intel_pmu_arch_lbr_read(struct cpu_hw_events *cpuc)
984 {
985 	intel_pmu_store_lbr(cpuc, NULL);
986 }
987 
988 static void intel_pmu_arch_lbr_read_xsave(struct cpu_hw_events *cpuc)
989 {
990 	struct x86_perf_task_context_arch_lbr_xsave *xsave = cpuc->lbr_xsave;
991 
992 	if (!xsave) {
993 		intel_pmu_store_lbr(cpuc, NULL);
994 		return;
995 	}
996 	xsaves(&xsave->xsave, XFEATURE_MASK_LBR);
997 
998 	intel_pmu_store_lbr(cpuc, xsave->lbr.entries);
999 }
1000 
1001 void intel_pmu_lbr_read(void)
1002 {
1003 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1004 
1005 	/*
1006 	 * Don't read when all LBRs users are using adaptive PEBS.
1007 	 *
1008 	 * This could be smarter and actually check the event,
1009 	 * but this simple approach seems to work for now.
1010 	 */
1011 	if (!cpuc->lbr_users || vlbr_exclude_host() ||
1012 	    cpuc->lbr_users == cpuc->lbr_pebs_users)
1013 		return;
1014 
1015 	x86_pmu.lbr_read(cpuc);
1016 
1017 	intel_pmu_lbr_filter(cpuc);
1018 }
1019 
1020 /*
1021  * SW filter is used:
1022  * - in case there is no HW filter
1023  * - in case the HW filter has errata or limitations
1024  */
1025 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
1026 {
1027 	u64 br_type = event->attr.branch_sample_type;
1028 	int mask = 0;
1029 
1030 	if (br_type & PERF_SAMPLE_BRANCH_USER)
1031 		mask |= X86_BR_USER;
1032 
1033 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
1034 		mask |= X86_BR_KERNEL;
1035 
1036 	/* we ignore BRANCH_HV here */
1037 
1038 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
1039 		mask |= X86_BR_ANY;
1040 
1041 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
1042 		mask |= X86_BR_ANY_CALL;
1043 
1044 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
1045 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
1046 
1047 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
1048 		mask |= X86_BR_IND_CALL;
1049 
1050 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
1051 		mask |= X86_BR_ABORT;
1052 
1053 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
1054 		mask |= X86_BR_IN_TX;
1055 
1056 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
1057 		mask |= X86_BR_NO_TX;
1058 
1059 	if (br_type & PERF_SAMPLE_BRANCH_COND)
1060 		mask |= X86_BR_JCC;
1061 
1062 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
1063 		if (!x86_pmu_has_lbr_callstack())
1064 			return -EOPNOTSUPP;
1065 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
1066 			return -EINVAL;
1067 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
1068 			X86_BR_CALL_STACK;
1069 	}
1070 
1071 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
1072 		mask |= X86_BR_IND_JMP;
1073 
1074 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
1075 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
1076 
1077 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
1078 		mask |= X86_BR_TYPE_SAVE;
1079 
1080 	/*
1081 	 * stash actual user request into reg, it may
1082 	 * be used by fixup code for some CPU
1083 	 */
1084 	event->hw.branch_reg.reg = mask;
1085 	return 0;
1086 }
1087 
1088 /*
1089  * setup the HW LBR filter
1090  * Used only when available, may not be enough to disambiguate
1091  * all branches, may need the help of the SW filter
1092  */
1093 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
1094 {
1095 	struct hw_perf_event_extra *reg;
1096 	u64 br_type = event->attr.branch_sample_type;
1097 	u64 mask = 0, v;
1098 	int i;
1099 
1100 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
1101 		if (!(br_type & (1ULL << i)))
1102 			continue;
1103 
1104 		v = x86_pmu.lbr_sel_map[i];
1105 		if (v == LBR_NOT_SUPP)
1106 			return -EOPNOTSUPP;
1107 
1108 		if (v != LBR_IGN)
1109 			mask |= v;
1110 	}
1111 
1112 	reg = &event->hw.branch_reg;
1113 	reg->idx = EXTRA_REG_LBR;
1114 
1115 	if (static_cpu_has(X86_FEATURE_ARCH_LBR)) {
1116 		reg->config = mask;
1117 
1118 		/*
1119 		 * The Arch LBR HW can retrieve the common branch types
1120 		 * from the LBR_INFO. It doesn't require the high overhead
1121 		 * SW disassemble.
1122 		 * Enable the branch type by default for the Arch LBR.
1123 		 */
1124 		reg->reg |= X86_BR_TYPE_SAVE;
1125 		return 0;
1126 	}
1127 
1128 	/*
1129 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
1130 	 * in suppress mode. So LBR_SELECT should be set to
1131 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
1132 	 * But the 10th bit LBR_CALL_STACK does not operate
1133 	 * in suppress mode.
1134 	 */
1135 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
1136 
1137 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
1138 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
1139 	    x86_pmu.lbr_has_info)
1140 		reg->config |= LBR_NO_INFO;
1141 
1142 	return 0;
1143 }
1144 
1145 int intel_pmu_setup_lbr_filter(struct perf_event *event)
1146 {
1147 	int ret = 0;
1148 
1149 	/*
1150 	 * no LBR on this PMU
1151 	 */
1152 	if (!x86_pmu.lbr_nr)
1153 		return -EOPNOTSUPP;
1154 
1155 	/*
1156 	 * setup SW LBR filter
1157 	 */
1158 	ret = intel_pmu_setup_sw_lbr_filter(event);
1159 	if (ret)
1160 		return ret;
1161 
1162 	/*
1163 	 * setup HW LBR filter, if any
1164 	 */
1165 	if (x86_pmu.lbr_sel_map)
1166 		ret = intel_pmu_setup_hw_lbr_filter(event);
1167 
1168 	return ret;
1169 }
1170 
1171 enum {
1172 	ARCH_LBR_BR_TYPE_JCC			= 0,
1173 	ARCH_LBR_BR_TYPE_NEAR_IND_JMP		= 1,
1174 	ARCH_LBR_BR_TYPE_NEAR_REL_JMP		= 2,
1175 	ARCH_LBR_BR_TYPE_NEAR_IND_CALL		= 3,
1176 	ARCH_LBR_BR_TYPE_NEAR_REL_CALL		= 4,
1177 	ARCH_LBR_BR_TYPE_NEAR_RET		= 5,
1178 	ARCH_LBR_BR_TYPE_KNOWN_MAX		= ARCH_LBR_BR_TYPE_NEAR_RET,
1179 
1180 	ARCH_LBR_BR_TYPE_MAP_MAX		= 16,
1181 };
1182 
1183 static const int arch_lbr_br_type_map[ARCH_LBR_BR_TYPE_MAP_MAX] = {
1184 	[ARCH_LBR_BR_TYPE_JCC]			= X86_BR_JCC,
1185 	[ARCH_LBR_BR_TYPE_NEAR_IND_JMP]		= X86_BR_IND_JMP,
1186 	[ARCH_LBR_BR_TYPE_NEAR_REL_JMP]		= X86_BR_JMP,
1187 	[ARCH_LBR_BR_TYPE_NEAR_IND_CALL]	= X86_BR_IND_CALL,
1188 	[ARCH_LBR_BR_TYPE_NEAR_REL_CALL]	= X86_BR_CALL,
1189 	[ARCH_LBR_BR_TYPE_NEAR_RET]		= X86_BR_RET,
1190 };
1191 
1192 /*
1193  * implement actual branch filter based on user demand.
1194  * Hardware may not exactly satisfy that request, thus
1195  * we need to inspect opcodes. Mismatched branches are
1196  * discarded. Therefore, the number of branches returned
1197  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1198  */
1199 static void
1200 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1201 {
1202 	u64 from, to;
1203 	int br_sel = cpuc->br_sel;
1204 	int i, j, type, to_plm;
1205 	bool compress = false;
1206 
1207 	/* if sampling all branches, then nothing to filter */
1208 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1209 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1210 		return;
1211 
1212 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1213 
1214 		from = cpuc->lbr_entries[i].from;
1215 		to = cpuc->lbr_entries[i].to;
1216 		type = cpuc->lbr_entries[i].type;
1217 
1218 		/*
1219 		 * Parse the branch type recorded in LBR_x_INFO MSR.
1220 		 * Doesn't support OTHER_BRANCH decoding for now.
1221 		 * OTHER_BRANCH branch type still rely on software decoding.
1222 		 */
1223 		if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
1224 		    type <= ARCH_LBR_BR_TYPE_KNOWN_MAX) {
1225 			to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
1226 			type = arch_lbr_br_type_map[type] | to_plm;
1227 		} else
1228 			type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1229 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1230 			if (cpuc->lbr_entries[i].in_tx)
1231 				type |= X86_BR_IN_TX;
1232 			else
1233 				type |= X86_BR_NO_TX;
1234 		}
1235 
1236 		/* if type does not correspond, then discard */
1237 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1238 			cpuc->lbr_entries[i].from = 0;
1239 			compress = true;
1240 		}
1241 
1242 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1243 			cpuc->lbr_entries[i].type = common_branch_type(type);
1244 	}
1245 
1246 	if (!compress)
1247 		return;
1248 
1249 	/* remove all entries with from=0 */
1250 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1251 		if (!cpuc->lbr_entries[i].from) {
1252 			j = i;
1253 			while (++j < cpuc->lbr_stack.nr) {
1254 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1255 				cpuc->lbr_counters[j-1] = cpuc->lbr_counters[j];
1256 			}
1257 			cpuc->lbr_stack.nr--;
1258 			if (!cpuc->lbr_entries[i].from)
1259 				continue;
1260 		}
1261 		i++;
1262 	}
1263 }
1264 
1265 void intel_pmu_store_pebs_lbrs(struct lbr_entry *lbr)
1266 {
1267 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1268 
1269 	/* Cannot get TOS for large PEBS and Arch LBR */
1270 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) ||
1271 	    (cpuc->n_pebs == cpuc->n_large_pebs))
1272 		cpuc->lbr_stack.hw_idx = -1ULL;
1273 	else
1274 		cpuc->lbr_stack.hw_idx = intel_pmu_lbr_tos();
1275 
1276 	intel_pmu_store_lbr(cpuc, lbr);
1277 	intel_pmu_lbr_filter(cpuc);
1278 }
1279 
1280 /*
1281  * Map interface branch filters onto LBR filters
1282  */
1283 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1284 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1285 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1286 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1287 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1288 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1289 						| LBR_IND_JMP | LBR_FAR,
1290 	/*
1291 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1292 	 */
1293 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1294 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1295 	/*
1296 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1297 	 */
1298 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1299 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1300 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1301 };
1302 
1303 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1304 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1305 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1306 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1307 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1308 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1309 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1310 						| LBR_FAR,
1311 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1312 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1313 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1314 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1315 };
1316 
1317 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1318 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1319 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1320 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1321 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1322 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1323 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1324 						| LBR_FAR,
1325 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1326 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1327 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1328 						| LBR_RETURN | LBR_CALL_STACK,
1329 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1330 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1331 };
1332 
1333 static int arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1334 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= ARCH_LBR_ANY,
1335 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= ARCH_LBR_USER,
1336 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= ARCH_LBR_KERNEL,
1337 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1338 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= ARCH_LBR_RETURN |
1339 						  ARCH_LBR_OTHER_BRANCH,
1340 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]     = ARCH_LBR_REL_CALL |
1341 						  ARCH_LBR_IND_CALL |
1342 						  ARCH_LBR_OTHER_BRANCH,
1343 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]     = ARCH_LBR_IND_CALL,
1344 	[PERF_SAMPLE_BRANCH_COND_SHIFT]         = ARCH_LBR_JCC,
1345 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]   = ARCH_LBR_REL_CALL |
1346 						  ARCH_LBR_IND_CALL |
1347 						  ARCH_LBR_RETURN |
1348 						  ARCH_LBR_CALL_STACK,
1349 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= ARCH_LBR_IND_JMP,
1350 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= ARCH_LBR_REL_CALL,
1351 };
1352 
1353 /* core */
1354 void __init intel_pmu_lbr_init_core(void)
1355 {
1356 	x86_pmu.lbr_nr     = 4;
1357 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1358 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1359 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1360 
1361 	/*
1362 	 * SW branch filter usage:
1363 	 * - compensate for lack of HW filter
1364 	 */
1365 }
1366 
1367 /* nehalem/westmere */
1368 void __init intel_pmu_lbr_init_nhm(void)
1369 {
1370 	x86_pmu.lbr_nr     = 16;
1371 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1372 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1373 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1374 
1375 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1376 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1377 
1378 	/*
1379 	 * SW branch filter usage:
1380 	 * - workaround LBR_SEL errata (see above)
1381 	 * - support syscall, sysret capture.
1382 	 *   That requires LBR_FAR but that means far
1383 	 *   jmp need to be filtered out
1384 	 */
1385 }
1386 
1387 /* sandy bridge */
1388 void __init intel_pmu_lbr_init_snb(void)
1389 {
1390 	x86_pmu.lbr_nr	 = 16;
1391 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1392 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1393 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1394 
1395 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1396 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1397 
1398 	/*
1399 	 * SW branch filter usage:
1400 	 * - support syscall, sysret capture.
1401 	 *   That requires LBR_FAR but that means far
1402 	 *   jmp need to be filtered out
1403 	 */
1404 }
1405 
1406 static inline struct kmem_cache *
1407 create_lbr_kmem_cache(size_t size, size_t align)
1408 {
1409 	return kmem_cache_create("x86_lbr", size, align, 0, NULL);
1410 }
1411 
1412 /* haswell */
1413 void intel_pmu_lbr_init_hsw(void)
1414 {
1415 	size_t size = sizeof(struct x86_perf_task_context);
1416 
1417 	x86_pmu.lbr_nr	 = 16;
1418 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1419 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1420 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1421 
1422 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1423 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1424 
1425 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1426 }
1427 
1428 /* skylake */
1429 __init void intel_pmu_lbr_init_skl(void)
1430 {
1431 	size_t size = sizeof(struct x86_perf_task_context);
1432 
1433 	x86_pmu.lbr_nr	 = 32;
1434 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1435 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1436 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1437 	x86_pmu.lbr_info = MSR_LBR_INFO_0;
1438 
1439 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1440 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1441 
1442 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1443 
1444 	/*
1445 	 * SW branch filter usage:
1446 	 * - support syscall, sysret capture.
1447 	 *   That requires LBR_FAR but that means far
1448 	 *   jmp need to be filtered out
1449 	 */
1450 }
1451 
1452 /* atom */
1453 void __init intel_pmu_lbr_init_atom(void)
1454 {
1455 	/*
1456 	 * only models starting at stepping 10 seems
1457 	 * to have an operational LBR which can freeze
1458 	 * on PMU interrupt
1459 	 */
1460 	if (boot_cpu_data.x86_model == 28
1461 	    && boot_cpu_data.x86_stepping < 10) {
1462 		pr_cont("LBR disabled due to erratum");
1463 		return;
1464 	}
1465 
1466 	x86_pmu.lbr_nr	   = 8;
1467 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1468 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1469 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1470 
1471 	/*
1472 	 * SW branch filter usage:
1473 	 * - compensate for lack of HW filter
1474 	 */
1475 }
1476 
1477 /* slm */
1478 void __init intel_pmu_lbr_init_slm(void)
1479 {
1480 	x86_pmu.lbr_nr	   = 8;
1481 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1482 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1483 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1484 
1485 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1486 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1487 
1488 	/*
1489 	 * SW branch filter usage:
1490 	 * - compensate for lack of HW filter
1491 	 */
1492 	pr_cont("8-deep LBR, ");
1493 }
1494 
1495 /* Knights Landing */
1496 void intel_pmu_lbr_init_knl(void)
1497 {
1498 	x86_pmu.lbr_nr	   = 8;
1499 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1500 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1501 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1502 
1503 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1504 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1505 
1506 	/* Knights Landing does have MISPREDICT bit */
1507 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1508 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1509 }
1510 
1511 void intel_pmu_lbr_init(void)
1512 {
1513 	switch (x86_pmu.intel_cap.lbr_format) {
1514 	case LBR_FORMAT_EIP_FLAGS2:
1515 		x86_pmu.lbr_has_tsx = 1;
1516 		x86_pmu.lbr_from_flags = 1;
1517 		if (lbr_from_signext_quirk_needed())
1518 			static_branch_enable(&lbr_from_quirk_key);
1519 		break;
1520 
1521 	case LBR_FORMAT_EIP_FLAGS:
1522 		x86_pmu.lbr_from_flags = 1;
1523 		break;
1524 
1525 	case LBR_FORMAT_INFO:
1526 		x86_pmu.lbr_has_tsx = 1;
1527 		fallthrough;
1528 	case LBR_FORMAT_INFO2:
1529 		x86_pmu.lbr_has_info = 1;
1530 		break;
1531 
1532 	case LBR_FORMAT_TIME:
1533 		x86_pmu.lbr_from_flags = 1;
1534 		x86_pmu.lbr_to_cycles = 1;
1535 		break;
1536 	}
1537 
1538 	if (x86_pmu.lbr_has_info) {
1539 		/*
1540 		 * Only used in combination with baseline pebs.
1541 		 */
1542 		static_branch_enable(&x86_lbr_mispred);
1543 		static_branch_enable(&x86_lbr_cycles);
1544 	}
1545 }
1546 
1547 /*
1548  * LBR state size is variable based on the max number of registers.
1549  * This calculates the expected state size, which should match
1550  * what the hardware enumerates for the size of XFEATURE_LBR.
1551  */
1552 static inline unsigned int get_lbr_state_size(void)
1553 {
1554 	return sizeof(struct arch_lbr_state) +
1555 	       x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1556 }
1557 
1558 static bool is_arch_lbr_xsave_available(void)
1559 {
1560 	if (!boot_cpu_has(X86_FEATURE_XSAVES))
1561 		return false;
1562 
1563 	/*
1564 	 * Check the LBR state with the corresponding software structure.
1565 	 * Disable LBR XSAVES support if the size doesn't match.
1566 	 */
1567 	if (xfeature_size(XFEATURE_LBR) == 0)
1568 		return false;
1569 
1570 	if (WARN_ON(xfeature_size(XFEATURE_LBR) != get_lbr_state_size()))
1571 		return false;
1572 
1573 	return true;
1574 }
1575 
1576 void __init intel_pmu_arch_lbr_init(void)
1577 {
1578 	struct pmu *pmu = x86_get_pmu(smp_processor_id());
1579 	union cpuid28_eax eax;
1580 	union cpuid28_ebx ebx;
1581 	union cpuid28_ecx ecx;
1582 	unsigned int unused_edx;
1583 	bool arch_lbr_xsave;
1584 	size_t size;
1585 	u64 lbr_nr;
1586 
1587 	/* Arch LBR Capabilities */
1588 	cpuid(28, &eax.full, &ebx.full, &ecx.full, &unused_edx);
1589 
1590 	lbr_nr = fls(eax.split.lbr_depth_mask) * 8;
1591 	if (!lbr_nr)
1592 		goto clear_arch_lbr;
1593 
1594 	/* Apply the max depth of Arch LBR */
1595 	if (wrmsrl_safe(MSR_ARCH_LBR_DEPTH, lbr_nr))
1596 		goto clear_arch_lbr;
1597 
1598 	x86_pmu.lbr_depth_mask = eax.split.lbr_depth_mask;
1599 	x86_pmu.lbr_deep_c_reset = eax.split.lbr_deep_c_reset;
1600 	x86_pmu.lbr_lip = eax.split.lbr_lip;
1601 	x86_pmu.lbr_cpl = ebx.split.lbr_cpl;
1602 	x86_pmu.lbr_filter = ebx.split.lbr_filter;
1603 	x86_pmu.lbr_call_stack = ebx.split.lbr_call_stack;
1604 	x86_pmu.lbr_mispred = ecx.split.lbr_mispred;
1605 	x86_pmu.lbr_timed_lbr = ecx.split.lbr_timed_lbr;
1606 	x86_pmu.lbr_br_type = ecx.split.lbr_br_type;
1607 	x86_pmu.lbr_counters = ecx.split.lbr_counters;
1608 	x86_pmu.lbr_nr = lbr_nr;
1609 
1610 	if (!!x86_pmu.lbr_counters)
1611 		x86_pmu.flags |= PMU_FL_BR_CNTR;
1612 
1613 	if (x86_pmu.lbr_mispred)
1614 		static_branch_enable(&x86_lbr_mispred);
1615 	if (x86_pmu.lbr_timed_lbr)
1616 		static_branch_enable(&x86_lbr_cycles);
1617 	if (x86_pmu.lbr_br_type)
1618 		static_branch_enable(&x86_lbr_type);
1619 
1620 	arch_lbr_xsave = is_arch_lbr_xsave_available();
1621 	if (arch_lbr_xsave) {
1622 		size = sizeof(struct x86_perf_task_context_arch_lbr_xsave) +
1623 		       get_lbr_state_size();
1624 		pmu->task_ctx_cache = create_lbr_kmem_cache(size,
1625 							    XSAVE_ALIGNMENT);
1626 	}
1627 
1628 	if (!pmu->task_ctx_cache) {
1629 		arch_lbr_xsave = false;
1630 
1631 		size = sizeof(struct x86_perf_task_context_arch_lbr) +
1632 		       lbr_nr * sizeof(struct lbr_entry);
1633 		pmu->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1634 	}
1635 
1636 	x86_pmu.lbr_from = MSR_ARCH_LBR_FROM_0;
1637 	x86_pmu.lbr_to = MSR_ARCH_LBR_TO_0;
1638 	x86_pmu.lbr_info = MSR_ARCH_LBR_INFO_0;
1639 
1640 	/* LBR callstack requires both CPL and Branch Filtering support */
1641 	if (!x86_pmu.lbr_cpl ||
1642 	    !x86_pmu.lbr_filter ||
1643 	    !x86_pmu.lbr_call_stack)
1644 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT] = LBR_NOT_SUPP;
1645 
1646 	if (!x86_pmu.lbr_cpl) {
1647 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_USER_SHIFT] = LBR_NOT_SUPP;
1648 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_KERNEL_SHIFT] = LBR_NOT_SUPP;
1649 	} else if (!x86_pmu.lbr_filter) {
1650 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_SHIFT] = LBR_NOT_SUPP;
1651 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT] = LBR_NOT_SUPP;
1652 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] = LBR_NOT_SUPP;
1653 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_NOT_SUPP;
1654 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_COND_SHIFT] = LBR_NOT_SUPP;
1655 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_NOT_SUPP;
1656 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_SHIFT] = LBR_NOT_SUPP;
1657 	}
1658 
1659 	x86_pmu.lbr_ctl_mask = ARCH_LBR_CTL_MASK;
1660 	x86_pmu.lbr_ctl_map  = arch_lbr_ctl_map;
1661 
1662 	if (!x86_pmu.lbr_cpl && !x86_pmu.lbr_filter)
1663 		x86_pmu.lbr_ctl_map = NULL;
1664 
1665 	x86_pmu.lbr_reset = intel_pmu_arch_lbr_reset;
1666 	if (arch_lbr_xsave) {
1667 		x86_pmu.lbr_save = intel_pmu_arch_lbr_xsaves;
1668 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_xrstors;
1669 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read_xsave;
1670 		pr_cont("XSAVE ");
1671 	} else {
1672 		x86_pmu.lbr_save = intel_pmu_arch_lbr_save;
1673 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_restore;
1674 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read;
1675 	}
1676 
1677 	pr_cont("Architectural LBR, ");
1678 
1679 	return;
1680 
1681 clear_arch_lbr:
1682 	setup_clear_cpu_cap(X86_FEATURE_ARCH_LBR);
1683 }
1684 
1685 /**
1686  * x86_perf_get_lbr - get the LBR records information
1687  *
1688  * @lbr: the caller's memory to store the LBR records information
1689  */
1690 void x86_perf_get_lbr(struct x86_pmu_lbr *lbr)
1691 {
1692 	lbr->nr = x86_pmu.lbr_nr;
1693 	lbr->from = x86_pmu.lbr_from;
1694 	lbr->to = x86_pmu.lbr_to;
1695 	lbr->info = x86_pmu.lbr_info;
1696 }
1697 EXPORT_SYMBOL_GPL(x86_perf_get_lbr);
1698 
1699 struct event_constraint vlbr_constraint =
1700 	__EVENT_CONSTRAINT(INTEL_FIXED_VLBR_EVENT, (1ULL << INTEL_PMC_IDX_FIXED_VLBR),
1701 			  FIXED_EVENT_FLAGS, 1, 0, PERF_X86_EVENT_LBR_SELECT);
1702