xref: /openbmc/linux/arch/x86/events/intel/lbr.c (revision e82c878d)
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 #include <asm/insn.h>
8 
9 #include "../perf_event.h"
10 
11 enum {
12 	LBR_FORMAT_32		= 0x00,
13 	LBR_FORMAT_LIP		= 0x01,
14 	LBR_FORMAT_EIP		= 0x02,
15 	LBR_FORMAT_EIP_FLAGS	= 0x03,
16 	LBR_FORMAT_EIP_FLAGS2	= 0x04,
17 	LBR_FORMAT_INFO		= 0x05,
18 	LBR_FORMAT_TIME		= 0x06,
19 	LBR_FORMAT_MAX_KNOWN    = LBR_FORMAT_TIME,
20 };
21 
22 static const enum {
23 	LBR_EIP_FLAGS		= 1,
24 	LBR_TSX			= 2,
25 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
26 	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
27 	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
28 };
29 
30 /*
31  * Intel LBR_SELECT bits
32  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
33  *
34  * Hardware branch filter (not available on all CPUs)
35  */
36 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
37 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
38 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
39 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
40 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
41 #define LBR_RETURN_BIT		5 /* do not capture near returns */
42 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
43 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
44 #define LBR_FAR_BIT		8 /* do not capture far branches */
45 #define LBR_CALL_STACK_BIT	9 /* enable call stack */
46 
47 /*
48  * Following bit only exists in Linux; we mask it out before writing it to
49  * the actual MSR. But it helps the constraint perf code to understand
50  * that this is a separate configuration.
51  */
52 #define LBR_NO_INFO_BIT	       63 /* don't read LBR_INFO. */
53 
54 #define LBR_KERNEL	(1 << LBR_KERNEL_BIT)
55 #define LBR_USER	(1 << LBR_USER_BIT)
56 #define LBR_JCC		(1 << LBR_JCC_BIT)
57 #define LBR_REL_CALL	(1 << LBR_REL_CALL_BIT)
58 #define LBR_IND_CALL	(1 << LBR_IND_CALL_BIT)
59 #define LBR_RETURN	(1 << LBR_RETURN_BIT)
60 #define LBR_REL_JMP	(1 << LBR_REL_JMP_BIT)
61 #define LBR_IND_JMP	(1 << LBR_IND_JMP_BIT)
62 #define LBR_FAR		(1 << LBR_FAR_BIT)
63 #define LBR_CALL_STACK	(1 << LBR_CALL_STACK_BIT)
64 #define LBR_NO_INFO	(1ULL << LBR_NO_INFO_BIT)
65 
66 #define LBR_PLM (LBR_KERNEL | LBR_USER)
67 
68 #define LBR_SEL_MASK	0x3ff	/* valid bits in LBR_SELECT */
69 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
70 #define LBR_IGN		0	/* ignored */
71 
72 #define LBR_ANY		 \
73 	(LBR_JCC	|\
74 	 LBR_REL_CALL	|\
75 	 LBR_IND_CALL	|\
76 	 LBR_RETURN	|\
77 	 LBR_REL_JMP	|\
78 	 LBR_IND_JMP	|\
79 	 LBR_FAR)
80 
81 #define LBR_FROM_FLAG_MISPRED	BIT_ULL(63)
82 #define LBR_FROM_FLAG_IN_TX	BIT_ULL(62)
83 #define LBR_FROM_FLAG_ABORT	BIT_ULL(61)
84 
85 #define LBR_FROM_SIGNEXT_2MSB	(BIT_ULL(60) | BIT_ULL(59))
86 
87 /*
88  * x86control flow change classification
89  * x86control flow changes include branches, interrupts, traps, faults
90  */
91 enum {
92 	X86_BR_NONE		= 0,      /* unknown */
93 
94 	X86_BR_USER		= 1 << 0, /* branch target is user */
95 	X86_BR_KERNEL		= 1 << 1, /* branch target is kernel */
96 
97 	X86_BR_CALL		= 1 << 2, /* call */
98 	X86_BR_RET		= 1 << 3, /* return */
99 	X86_BR_SYSCALL		= 1 << 4, /* syscall */
100 	X86_BR_SYSRET		= 1 << 5, /* syscall return */
101 	X86_BR_INT		= 1 << 6, /* sw interrupt */
102 	X86_BR_IRET		= 1 << 7, /* return from interrupt */
103 	X86_BR_JCC		= 1 << 8, /* conditional */
104 	X86_BR_JMP		= 1 << 9, /* jump */
105 	X86_BR_IRQ		= 1 << 10,/* hw interrupt or trap or fault */
106 	X86_BR_IND_CALL		= 1 << 11,/* indirect calls */
107 	X86_BR_ABORT		= 1 << 12,/* transaction abort */
108 	X86_BR_IN_TX		= 1 << 13,/* in transaction */
109 	X86_BR_NO_TX		= 1 << 14,/* not in transaction */
110 	X86_BR_ZERO_CALL	= 1 << 15,/* zero length call */
111 	X86_BR_CALL_STACK	= 1 << 16,/* call stack */
112 	X86_BR_IND_JMP		= 1 << 17,/* indirect jump */
113 
114 	X86_BR_TYPE_SAVE	= 1 << 18,/* indicate to save branch type */
115 
116 };
117 
118 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
119 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
120 
121 #define X86_BR_ANY       \
122 	(X86_BR_CALL    |\
123 	 X86_BR_RET     |\
124 	 X86_BR_SYSCALL |\
125 	 X86_BR_SYSRET  |\
126 	 X86_BR_INT     |\
127 	 X86_BR_IRET    |\
128 	 X86_BR_JCC     |\
129 	 X86_BR_JMP	 |\
130 	 X86_BR_IRQ	 |\
131 	 X86_BR_ABORT	 |\
132 	 X86_BR_IND_CALL |\
133 	 X86_BR_IND_JMP  |\
134 	 X86_BR_ZERO_CALL)
135 
136 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
137 
138 #define X86_BR_ANY_CALL		 \
139 	(X86_BR_CALL		|\
140 	 X86_BR_IND_CALL	|\
141 	 X86_BR_ZERO_CALL	|\
142 	 X86_BR_SYSCALL		|\
143 	 X86_BR_IRQ		|\
144 	 X86_BR_INT)
145 
146 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
147 
148 /*
149  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
150  * otherwise it becomes near impossible to get a reliable stack.
151  */
152 
153 static void __intel_pmu_lbr_enable(bool pmi)
154 {
155 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
156 	u64 debugctl, lbr_select = 0, orig_debugctl;
157 
158 	/*
159 	 * No need to unfreeze manually, as v4 can do that as part
160 	 * of the GLOBAL_STATUS ack.
161 	 */
162 	if (pmi && x86_pmu.version >= 4)
163 		return;
164 
165 	/*
166 	 * No need to reprogram LBR_SELECT in a PMI, as it
167 	 * did not change.
168 	 */
169 	if (cpuc->lbr_sel)
170 		lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
171 	if (!pmi && cpuc->lbr_sel)
172 		wrmsrl(MSR_LBR_SELECT, lbr_select);
173 
174 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
175 	orig_debugctl = debugctl;
176 	debugctl |= DEBUGCTLMSR_LBR;
177 	/*
178 	 * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
179 	 * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
180 	 * may cause superfluous increase/decrease of LBR_TOS.
181 	 */
182 	if (!(lbr_select & LBR_CALL_STACK))
183 		debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
184 	if (orig_debugctl != debugctl)
185 		wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
186 }
187 
188 static void __intel_pmu_lbr_disable(void)
189 {
190 	u64 debugctl;
191 
192 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
193 	debugctl &= ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
194 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
195 }
196 
197 static void intel_pmu_lbr_reset_32(void)
198 {
199 	int i;
200 
201 	for (i = 0; i < x86_pmu.lbr_nr; i++)
202 		wrmsrl(x86_pmu.lbr_from + i, 0);
203 }
204 
205 static void intel_pmu_lbr_reset_64(void)
206 {
207 	int i;
208 
209 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
210 		wrmsrl(x86_pmu.lbr_from + i, 0);
211 		wrmsrl(x86_pmu.lbr_to   + i, 0);
212 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
213 			wrmsrl(MSR_LBR_INFO_0 + i, 0);
214 	}
215 }
216 
217 void intel_pmu_lbr_reset(void)
218 {
219 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
220 
221 	if (!x86_pmu.lbr_nr)
222 		return;
223 
224 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
225 		intel_pmu_lbr_reset_32();
226 	else
227 		intel_pmu_lbr_reset_64();
228 
229 	cpuc->last_task_ctx = NULL;
230 	cpuc->last_log_id = 0;
231 }
232 
233 /*
234  * TOS = most recently recorded branch
235  */
236 static inline u64 intel_pmu_lbr_tos(void)
237 {
238 	u64 tos;
239 
240 	rdmsrl(x86_pmu.lbr_tos, tos);
241 	return tos;
242 }
243 
244 enum {
245 	LBR_NONE,
246 	LBR_VALID,
247 };
248 
249 /*
250  * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
251  * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported, but when
252  * TSX is not supported they have no consistent behavior:
253  *
254  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
255  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
256  *     part of the sign extension.
257  *
258  * Therefore, if:
259  *
260  *   1) LBR has TSX format
261  *   2) CPU has no TSX support enabled
262  *
263  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
264  * value from rdmsr() must be converted to have a 61 bits sign extension,
265  * ignoring the TSX flags.
266  */
267 static inline bool lbr_from_signext_quirk_needed(void)
268 {
269 	int lbr_format = x86_pmu.intel_cap.lbr_format;
270 	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
271 			   boot_cpu_has(X86_FEATURE_RTM);
272 
273 	return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
274 }
275 
276 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
277 
278 /* If quirk is enabled, ensure sign extension is 63 bits: */
279 inline u64 lbr_from_signext_quirk_wr(u64 val)
280 {
281 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
282 		/*
283 		 * Sign extend into bits 61:62 while preserving bit 63.
284 		 *
285 		 * Quirk is enabled when TSX is disabled. Therefore TSX bits
286 		 * in val are always OFF and must be changed to be sign
287 		 * extension bits. Since bits 59:60 are guaranteed to be
288 		 * part of the sign extension bits, we can just copy them
289 		 * to 61:62.
290 		 */
291 		val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
292 	}
293 	return val;
294 }
295 
296 /*
297  * If quirk is needed, ensure sign extension is 61 bits:
298  */
299 static u64 lbr_from_signext_quirk_rd(u64 val)
300 {
301 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
302 		/*
303 		 * Quirk is on when TSX is not enabled. Therefore TSX
304 		 * flags must be read as OFF.
305 		 */
306 		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
307 	}
308 	return val;
309 }
310 
311 static inline void wrlbr_from(unsigned int idx, u64 val)
312 {
313 	val = lbr_from_signext_quirk_wr(val);
314 	wrmsrl(x86_pmu.lbr_from + idx, val);
315 }
316 
317 static inline void wrlbr_to(unsigned int idx, u64 val)
318 {
319 	wrmsrl(x86_pmu.lbr_to + idx, val);
320 }
321 
322 static inline u64 rdlbr_from(unsigned int idx)
323 {
324 	u64 val;
325 
326 	rdmsrl(x86_pmu.lbr_from + idx, val);
327 
328 	return lbr_from_signext_quirk_rd(val);
329 }
330 
331 static inline u64 rdlbr_to(unsigned int idx)
332 {
333 	u64 val;
334 
335 	rdmsrl(x86_pmu.lbr_to + idx, val);
336 
337 	return val;
338 }
339 
340 static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
341 {
342 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
343 	int i;
344 	unsigned lbr_idx, mask;
345 	u64 tos;
346 
347 	if (task_ctx->lbr_callstack_users == 0 ||
348 	    task_ctx->lbr_stack_state == LBR_NONE) {
349 		intel_pmu_lbr_reset();
350 		return;
351 	}
352 
353 	tos = task_ctx->tos;
354 	/*
355 	 * Does not restore the LBR registers, if
356 	 * - No one else touched them, and
357 	 * - Did not enter C6
358 	 */
359 	if ((task_ctx == cpuc->last_task_ctx) &&
360 	    (task_ctx->log_id == cpuc->last_log_id) &&
361 	    rdlbr_from(tos)) {
362 		task_ctx->lbr_stack_state = LBR_NONE;
363 		return;
364 	}
365 
366 	mask = x86_pmu.lbr_nr - 1;
367 	for (i = 0; i < task_ctx->valid_lbrs; i++) {
368 		lbr_idx = (tos - i) & mask;
369 		wrlbr_from(lbr_idx, task_ctx->lbr_from[i]);
370 		wrlbr_to  (lbr_idx, task_ctx->lbr_to[i]);
371 
372 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
373 			wrmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
374 	}
375 
376 	for (; i < x86_pmu.lbr_nr; i++) {
377 		lbr_idx = (tos - i) & mask;
378 		wrlbr_from(lbr_idx, 0);
379 		wrlbr_to(lbr_idx, 0);
380 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
381 			wrmsrl(MSR_LBR_INFO_0 + lbr_idx, 0);
382 	}
383 
384 	wrmsrl(x86_pmu.lbr_tos, tos);
385 	task_ctx->lbr_stack_state = LBR_NONE;
386 }
387 
388 static void __intel_pmu_lbr_save(struct x86_perf_task_context *task_ctx)
389 {
390 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
391 	unsigned lbr_idx, mask;
392 	u64 tos, from;
393 	int i;
394 
395 	if (task_ctx->lbr_callstack_users == 0) {
396 		task_ctx->lbr_stack_state = LBR_NONE;
397 		return;
398 	}
399 
400 	mask = x86_pmu.lbr_nr - 1;
401 	tos = intel_pmu_lbr_tos();
402 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
403 		lbr_idx = (tos - i) & mask;
404 		from = rdlbr_from(lbr_idx);
405 		if (!from)
406 			break;
407 		task_ctx->lbr_from[i] = from;
408 		task_ctx->lbr_to[i]   = rdlbr_to(lbr_idx);
409 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
410 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
411 	}
412 	task_ctx->valid_lbrs = i;
413 	task_ctx->tos = tos;
414 	task_ctx->lbr_stack_state = LBR_VALID;
415 
416 	cpuc->last_task_ctx = task_ctx;
417 	cpuc->last_log_id = ++task_ctx->log_id;
418 }
419 
420 void intel_pmu_lbr_swap_task_ctx(struct perf_event_context *prev,
421 				 struct perf_event_context *next)
422 {
423 	struct x86_perf_task_context *prev_ctx_data, *next_ctx_data;
424 
425 	swap(prev->task_ctx_data, next->task_ctx_data);
426 
427 	/*
428 	 * Architecture specific synchronization makes sense in
429 	 * case both prev->task_ctx_data and next->task_ctx_data
430 	 * pointers are allocated.
431 	 */
432 
433 	prev_ctx_data = next->task_ctx_data;
434 	next_ctx_data = prev->task_ctx_data;
435 
436 	if (!prev_ctx_data || !next_ctx_data)
437 		return;
438 
439 	swap(prev_ctx_data->lbr_callstack_users,
440 	     next_ctx_data->lbr_callstack_users);
441 }
442 
443 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
444 {
445 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
446 	struct x86_perf_task_context *task_ctx;
447 
448 	if (!cpuc->lbr_users)
449 		return;
450 
451 	/*
452 	 * If LBR callstack feature is enabled and the stack was saved when
453 	 * the task was scheduled out, restore the stack. Otherwise flush
454 	 * the LBR stack.
455 	 */
456 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
457 	if (task_ctx) {
458 		if (sched_in)
459 			__intel_pmu_lbr_restore(task_ctx);
460 		else
461 			__intel_pmu_lbr_save(task_ctx);
462 		return;
463 	}
464 
465 	/*
466 	 * Since a context switch can flip the address space and LBR entries
467 	 * are not tagged with an identifier, we need to wipe the LBR, even for
468 	 * per-cpu events. You simply cannot resolve the branches from the old
469 	 * address space.
470 	 */
471 	if (sched_in)
472 		intel_pmu_lbr_reset();
473 }
474 
475 static inline bool branch_user_callstack(unsigned br_sel)
476 {
477 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
478 }
479 
480 void intel_pmu_lbr_add(struct perf_event *event)
481 {
482 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
483 	struct x86_perf_task_context *task_ctx;
484 
485 	if (!x86_pmu.lbr_nr)
486 		return;
487 
488 	cpuc->br_sel = event->hw.branch_reg.reg;
489 
490 	if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data) {
491 		task_ctx = event->ctx->task_ctx_data;
492 		task_ctx->lbr_callstack_users++;
493 	}
494 
495 	/*
496 	 * Request pmu::sched_task() callback, which will fire inside the
497 	 * regular perf event scheduling, so that call will:
498 	 *
499 	 *  - restore or wipe; when LBR-callstack,
500 	 *  - wipe; otherwise,
501 	 *
502 	 * when this is from __perf_event_task_sched_in().
503 	 *
504 	 * However, if this is from perf_install_in_context(), no such callback
505 	 * will follow and we'll need to reset the LBR here if this is the
506 	 * first LBR event.
507 	 *
508 	 * The problem is, we cannot tell these cases apart... but we can
509 	 * exclude the biggest chunk of cases by looking at
510 	 * event->total_time_running. An event that has accrued runtime cannot
511 	 * be 'new'. Conversely, a new event can get installed through the
512 	 * context switch path for the first time.
513 	 */
514 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
515 		cpuc->lbr_pebs_users++;
516 	perf_sched_cb_inc(event->ctx->pmu);
517 	if (!cpuc->lbr_users++ && !event->total_time_running)
518 		intel_pmu_lbr_reset();
519 }
520 
521 void intel_pmu_lbr_del(struct perf_event *event)
522 {
523 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
524 	struct x86_perf_task_context *task_ctx;
525 
526 	if (!x86_pmu.lbr_nr)
527 		return;
528 
529 	if (branch_user_callstack(cpuc->br_sel) &&
530 	    event->ctx->task_ctx_data) {
531 		task_ctx = event->ctx->task_ctx_data;
532 		task_ctx->lbr_callstack_users--;
533 	}
534 
535 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
536 		cpuc->lbr_pebs_users--;
537 	cpuc->lbr_users--;
538 	WARN_ON_ONCE(cpuc->lbr_users < 0);
539 	WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
540 	perf_sched_cb_dec(event->ctx->pmu);
541 }
542 
543 void intel_pmu_lbr_enable_all(bool pmi)
544 {
545 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
546 
547 	if (cpuc->lbr_users)
548 		__intel_pmu_lbr_enable(pmi);
549 }
550 
551 void intel_pmu_lbr_disable_all(void)
552 {
553 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
554 
555 	if (cpuc->lbr_users)
556 		__intel_pmu_lbr_disable();
557 }
558 
559 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
560 {
561 	unsigned long mask = x86_pmu.lbr_nr - 1;
562 	u64 tos = intel_pmu_lbr_tos();
563 	int i;
564 
565 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
566 		unsigned long lbr_idx = (tos - i) & mask;
567 		union {
568 			struct {
569 				u32 from;
570 				u32 to;
571 			};
572 			u64     lbr;
573 		} msr_lastbranch;
574 
575 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
576 
577 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
578 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
579 		cpuc->lbr_entries[i].mispred	= 0;
580 		cpuc->lbr_entries[i].predicted	= 0;
581 		cpuc->lbr_entries[i].in_tx	= 0;
582 		cpuc->lbr_entries[i].abort	= 0;
583 		cpuc->lbr_entries[i].cycles	= 0;
584 		cpuc->lbr_entries[i].type	= 0;
585 		cpuc->lbr_entries[i].reserved	= 0;
586 	}
587 	cpuc->lbr_stack.nr = i;
588 	cpuc->lbr_stack.hw_idx = tos;
589 }
590 
591 /*
592  * Due to lack of segmentation in Linux the effective address (offset)
593  * is the same as the linear address, allowing us to merge the LIP and EIP
594  * LBR formats.
595  */
596 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
597 {
598 	bool need_info = false, call_stack = false;
599 	unsigned long mask = x86_pmu.lbr_nr - 1;
600 	int lbr_format = x86_pmu.intel_cap.lbr_format;
601 	u64 tos = intel_pmu_lbr_tos();
602 	int i;
603 	int out = 0;
604 	int num = x86_pmu.lbr_nr;
605 
606 	if (cpuc->lbr_sel) {
607 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
608 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
609 			call_stack = true;
610 	}
611 
612 	for (i = 0; i < num; i++) {
613 		unsigned long lbr_idx = (tos - i) & mask;
614 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
615 		int skip = 0;
616 		u16 cycles = 0;
617 		int lbr_flags = lbr_desc[lbr_format];
618 
619 		from = rdlbr_from(lbr_idx);
620 		to   = rdlbr_to(lbr_idx);
621 
622 		/*
623 		 * Read LBR call stack entries
624 		 * until invalid entry (0s) is detected.
625 		 */
626 		if (call_stack && !from)
627 			break;
628 
629 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
630 			u64 info;
631 
632 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, info);
633 			mis = !!(info & LBR_INFO_MISPRED);
634 			pred = !mis;
635 			in_tx = !!(info & LBR_INFO_IN_TX);
636 			abort = !!(info & LBR_INFO_ABORT);
637 			cycles = (info & LBR_INFO_CYCLES);
638 		}
639 
640 		if (lbr_format == LBR_FORMAT_TIME) {
641 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
642 			pred = !mis;
643 			skip = 1;
644 			cycles = ((to >> 48) & LBR_INFO_CYCLES);
645 
646 			to = (u64)((((s64)to) << 16) >> 16);
647 		}
648 
649 		if (lbr_flags & LBR_EIP_FLAGS) {
650 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
651 			pred = !mis;
652 			skip = 1;
653 		}
654 		if (lbr_flags & LBR_TSX) {
655 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
656 			abort = !!(from & LBR_FROM_FLAG_ABORT);
657 			skip = 3;
658 		}
659 		from = (u64)((((s64)from) << skip) >> skip);
660 
661 		/*
662 		 * Some CPUs report duplicated abort records,
663 		 * with the second entry not having an abort bit set.
664 		 * Skip them here. This loop runs backwards,
665 		 * so we need to undo the previous record.
666 		 * If the abort just happened outside the window
667 		 * the extra entry cannot be removed.
668 		 */
669 		if (abort && x86_pmu.lbr_double_abort && out > 0)
670 			out--;
671 
672 		cpuc->lbr_entries[out].from	 = from;
673 		cpuc->lbr_entries[out].to	 = to;
674 		cpuc->lbr_entries[out].mispred	 = mis;
675 		cpuc->lbr_entries[out].predicted = pred;
676 		cpuc->lbr_entries[out].in_tx	 = in_tx;
677 		cpuc->lbr_entries[out].abort	 = abort;
678 		cpuc->lbr_entries[out].cycles	 = cycles;
679 		cpuc->lbr_entries[out].type	 = 0;
680 		cpuc->lbr_entries[out].reserved	 = 0;
681 		out++;
682 	}
683 	cpuc->lbr_stack.nr = out;
684 	cpuc->lbr_stack.hw_idx = tos;
685 }
686 
687 void intel_pmu_lbr_read(void)
688 {
689 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
690 
691 	/*
692 	 * Don't read when all LBRs users are using adaptive PEBS.
693 	 *
694 	 * This could be smarter and actually check the event,
695 	 * but this simple approach seems to work for now.
696 	 */
697 	if (!cpuc->lbr_users || cpuc->lbr_users == cpuc->lbr_pebs_users)
698 		return;
699 
700 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
701 		intel_pmu_lbr_read_32(cpuc);
702 	else
703 		intel_pmu_lbr_read_64(cpuc);
704 
705 	intel_pmu_lbr_filter(cpuc);
706 }
707 
708 /*
709  * SW filter is used:
710  * - in case there is no HW filter
711  * - in case the HW filter has errata or limitations
712  */
713 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
714 {
715 	u64 br_type = event->attr.branch_sample_type;
716 	int mask = 0;
717 
718 	if (br_type & PERF_SAMPLE_BRANCH_USER)
719 		mask |= X86_BR_USER;
720 
721 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
722 		mask |= X86_BR_KERNEL;
723 
724 	/* we ignore BRANCH_HV here */
725 
726 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
727 		mask |= X86_BR_ANY;
728 
729 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
730 		mask |= X86_BR_ANY_CALL;
731 
732 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
733 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
734 
735 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
736 		mask |= X86_BR_IND_CALL;
737 
738 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
739 		mask |= X86_BR_ABORT;
740 
741 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
742 		mask |= X86_BR_IN_TX;
743 
744 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
745 		mask |= X86_BR_NO_TX;
746 
747 	if (br_type & PERF_SAMPLE_BRANCH_COND)
748 		mask |= X86_BR_JCC;
749 
750 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
751 		if (!x86_pmu_has_lbr_callstack())
752 			return -EOPNOTSUPP;
753 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
754 			return -EINVAL;
755 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
756 			X86_BR_CALL_STACK;
757 	}
758 
759 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
760 		mask |= X86_BR_IND_JMP;
761 
762 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
763 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
764 
765 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
766 		mask |= X86_BR_TYPE_SAVE;
767 
768 	/*
769 	 * stash actual user request into reg, it may
770 	 * be used by fixup code for some CPU
771 	 */
772 	event->hw.branch_reg.reg = mask;
773 	return 0;
774 }
775 
776 /*
777  * setup the HW LBR filter
778  * Used only when available, may not be enough to disambiguate
779  * all branches, may need the help of the SW filter
780  */
781 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
782 {
783 	struct hw_perf_event_extra *reg;
784 	u64 br_type = event->attr.branch_sample_type;
785 	u64 mask = 0, v;
786 	int i;
787 
788 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
789 		if (!(br_type & (1ULL << i)))
790 			continue;
791 
792 		v = x86_pmu.lbr_sel_map[i];
793 		if (v == LBR_NOT_SUPP)
794 			return -EOPNOTSUPP;
795 
796 		if (v != LBR_IGN)
797 			mask |= v;
798 	}
799 
800 	reg = &event->hw.branch_reg;
801 	reg->idx = EXTRA_REG_LBR;
802 
803 	/*
804 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
805 	 * in suppress mode. So LBR_SELECT should be set to
806 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
807 	 * But the 10th bit LBR_CALL_STACK does not operate
808 	 * in suppress mode.
809 	 */
810 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
811 
812 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
813 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
814 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
815 		reg->config |= LBR_NO_INFO;
816 
817 	return 0;
818 }
819 
820 int intel_pmu_setup_lbr_filter(struct perf_event *event)
821 {
822 	int ret = 0;
823 
824 	/*
825 	 * no LBR on this PMU
826 	 */
827 	if (!x86_pmu.lbr_nr)
828 		return -EOPNOTSUPP;
829 
830 	/*
831 	 * setup SW LBR filter
832 	 */
833 	ret = intel_pmu_setup_sw_lbr_filter(event);
834 	if (ret)
835 		return ret;
836 
837 	/*
838 	 * setup HW LBR filter, if any
839 	 */
840 	if (x86_pmu.lbr_sel_map)
841 		ret = intel_pmu_setup_hw_lbr_filter(event);
842 
843 	return ret;
844 }
845 
846 /*
847  * return the type of control flow change at address "from"
848  * instruction is not necessarily a branch (in case of interrupt).
849  *
850  * The branch type returned also includes the priv level of the
851  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
852  *
853  * If a branch type is unknown OR the instruction cannot be
854  * decoded (e.g., text page not present), then X86_BR_NONE is
855  * returned.
856  */
857 static int branch_type(unsigned long from, unsigned long to, int abort)
858 {
859 	struct insn insn;
860 	void *addr;
861 	int bytes_read, bytes_left;
862 	int ret = X86_BR_NONE;
863 	int ext, to_plm, from_plm;
864 	u8 buf[MAX_INSN_SIZE];
865 	int is64 = 0;
866 
867 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
868 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
869 
870 	/*
871 	 * maybe zero if lbr did not fill up after a reset by the time
872 	 * we get a PMU interrupt
873 	 */
874 	if (from == 0 || to == 0)
875 		return X86_BR_NONE;
876 
877 	if (abort)
878 		return X86_BR_ABORT | to_plm;
879 
880 	if (from_plm == X86_BR_USER) {
881 		/*
882 		 * can happen if measuring at the user level only
883 		 * and we interrupt in a kernel thread, e.g., idle.
884 		 */
885 		if (!current->mm)
886 			return X86_BR_NONE;
887 
888 		/* may fail if text not present */
889 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
890 						MAX_INSN_SIZE);
891 		bytes_read = MAX_INSN_SIZE - bytes_left;
892 		if (!bytes_read)
893 			return X86_BR_NONE;
894 
895 		addr = buf;
896 	} else {
897 		/*
898 		 * The LBR logs any address in the IP, even if the IP just
899 		 * faulted. This means userspace can control the from address.
900 		 * Ensure we don't blindy read any address by validating it is
901 		 * a known text address.
902 		 */
903 		if (kernel_text_address(from)) {
904 			addr = (void *)from;
905 			/*
906 			 * Assume we can get the maximum possible size
907 			 * when grabbing kernel data.  This is not
908 			 * _strictly_ true since we could possibly be
909 			 * executing up next to a memory hole, but
910 			 * it is very unlikely to be a problem.
911 			 */
912 			bytes_read = MAX_INSN_SIZE;
913 		} else {
914 			return X86_BR_NONE;
915 		}
916 	}
917 
918 	/*
919 	 * decoder needs to know the ABI especially
920 	 * on 64-bit systems running 32-bit apps
921 	 */
922 #ifdef CONFIG_X86_64
923 	is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
924 #endif
925 	insn_init(&insn, addr, bytes_read, is64);
926 	insn_get_opcode(&insn);
927 	if (!insn.opcode.got)
928 		return X86_BR_ABORT;
929 
930 	switch (insn.opcode.bytes[0]) {
931 	case 0xf:
932 		switch (insn.opcode.bytes[1]) {
933 		case 0x05: /* syscall */
934 		case 0x34: /* sysenter */
935 			ret = X86_BR_SYSCALL;
936 			break;
937 		case 0x07: /* sysret */
938 		case 0x35: /* sysexit */
939 			ret = X86_BR_SYSRET;
940 			break;
941 		case 0x80 ... 0x8f: /* conditional */
942 			ret = X86_BR_JCC;
943 			break;
944 		default:
945 			ret = X86_BR_NONE;
946 		}
947 		break;
948 	case 0x70 ... 0x7f: /* conditional */
949 		ret = X86_BR_JCC;
950 		break;
951 	case 0xc2: /* near ret */
952 	case 0xc3: /* near ret */
953 	case 0xca: /* far ret */
954 	case 0xcb: /* far ret */
955 		ret = X86_BR_RET;
956 		break;
957 	case 0xcf: /* iret */
958 		ret = X86_BR_IRET;
959 		break;
960 	case 0xcc ... 0xce: /* int */
961 		ret = X86_BR_INT;
962 		break;
963 	case 0xe8: /* call near rel */
964 		insn_get_immediate(&insn);
965 		if (insn.immediate1.value == 0) {
966 			/* zero length call */
967 			ret = X86_BR_ZERO_CALL;
968 			break;
969 		}
970 		/* fall through */
971 	case 0x9a: /* call far absolute */
972 		ret = X86_BR_CALL;
973 		break;
974 	case 0xe0 ... 0xe3: /* loop jmp */
975 		ret = X86_BR_JCC;
976 		break;
977 	case 0xe9 ... 0xeb: /* jmp */
978 		ret = X86_BR_JMP;
979 		break;
980 	case 0xff: /* call near absolute, call far absolute ind */
981 		insn_get_modrm(&insn);
982 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
983 		switch (ext) {
984 		case 2: /* near ind call */
985 		case 3: /* far ind call */
986 			ret = X86_BR_IND_CALL;
987 			break;
988 		case 4:
989 		case 5:
990 			ret = X86_BR_IND_JMP;
991 			break;
992 		}
993 		break;
994 	default:
995 		ret = X86_BR_NONE;
996 	}
997 	/*
998 	 * interrupts, traps, faults (and thus ring transition) may
999 	 * occur on any instructions. Thus, to classify them correctly,
1000 	 * we need to first look at the from and to priv levels. If they
1001 	 * are different and to is in the kernel, then it indicates
1002 	 * a ring transition. If the from instruction is not a ring
1003 	 * transition instr (syscall, systenter, int), then it means
1004 	 * it was a irq, trap or fault.
1005 	 *
1006 	 * we have no way of detecting kernel to kernel faults.
1007 	 */
1008 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
1009 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
1010 		ret = X86_BR_IRQ;
1011 
1012 	/*
1013 	 * branch priv level determined by target as
1014 	 * is done by HW when LBR_SELECT is implemented
1015 	 */
1016 	if (ret != X86_BR_NONE)
1017 		ret |= to_plm;
1018 
1019 	return ret;
1020 }
1021 
1022 #define X86_BR_TYPE_MAP_MAX	16
1023 
1024 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
1025 	PERF_BR_CALL,		/* X86_BR_CALL */
1026 	PERF_BR_RET,		/* X86_BR_RET */
1027 	PERF_BR_SYSCALL,	/* X86_BR_SYSCALL */
1028 	PERF_BR_SYSRET,		/* X86_BR_SYSRET */
1029 	PERF_BR_UNKNOWN,	/* X86_BR_INT */
1030 	PERF_BR_UNKNOWN,	/* X86_BR_IRET */
1031 	PERF_BR_COND,		/* X86_BR_JCC */
1032 	PERF_BR_UNCOND,		/* X86_BR_JMP */
1033 	PERF_BR_UNKNOWN,	/* X86_BR_IRQ */
1034 	PERF_BR_IND_CALL,	/* X86_BR_IND_CALL */
1035 	PERF_BR_UNKNOWN,	/* X86_BR_ABORT */
1036 	PERF_BR_UNKNOWN,	/* X86_BR_IN_TX */
1037 	PERF_BR_UNKNOWN,	/* X86_BR_NO_TX */
1038 	PERF_BR_CALL,		/* X86_BR_ZERO_CALL */
1039 	PERF_BR_UNKNOWN,	/* X86_BR_CALL_STACK */
1040 	PERF_BR_IND,		/* X86_BR_IND_JMP */
1041 };
1042 
1043 static int
1044 common_branch_type(int type)
1045 {
1046 	int i;
1047 
1048 	type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1049 
1050 	if (type) {
1051 		i = __ffs(type);
1052 		if (i < X86_BR_TYPE_MAP_MAX)
1053 			return branch_map[i];
1054 	}
1055 
1056 	return PERF_BR_UNKNOWN;
1057 }
1058 
1059 /*
1060  * implement actual branch filter based on user demand.
1061  * Hardware may not exactly satisfy that request, thus
1062  * we need to inspect opcodes. Mismatched branches are
1063  * discarded. Therefore, the number of branches returned
1064  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1065  */
1066 static void
1067 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1068 {
1069 	u64 from, to;
1070 	int br_sel = cpuc->br_sel;
1071 	int i, j, type;
1072 	bool compress = false;
1073 
1074 	/* if sampling all branches, then nothing to filter */
1075 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1076 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1077 		return;
1078 
1079 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1080 
1081 		from = cpuc->lbr_entries[i].from;
1082 		to = cpuc->lbr_entries[i].to;
1083 
1084 		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1085 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1086 			if (cpuc->lbr_entries[i].in_tx)
1087 				type |= X86_BR_IN_TX;
1088 			else
1089 				type |= X86_BR_NO_TX;
1090 		}
1091 
1092 		/* if type does not correspond, then discard */
1093 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1094 			cpuc->lbr_entries[i].from = 0;
1095 			compress = true;
1096 		}
1097 
1098 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1099 			cpuc->lbr_entries[i].type = common_branch_type(type);
1100 	}
1101 
1102 	if (!compress)
1103 		return;
1104 
1105 	/* remove all entries with from=0 */
1106 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1107 		if (!cpuc->lbr_entries[i].from) {
1108 			j = i;
1109 			while (++j < cpuc->lbr_stack.nr)
1110 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1111 			cpuc->lbr_stack.nr--;
1112 			if (!cpuc->lbr_entries[i].from)
1113 				continue;
1114 		}
1115 		i++;
1116 	}
1117 }
1118 
1119 void intel_pmu_store_pebs_lbrs(struct pebs_lbr *lbr)
1120 {
1121 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1122 	int i;
1123 
1124 	cpuc->lbr_stack.nr = x86_pmu.lbr_nr;
1125 
1126 	/* Cannot get TOS for large PEBS */
1127 	if (cpuc->n_pebs == cpuc->n_large_pebs)
1128 		cpuc->lbr_stack.hw_idx = -1ULL;
1129 	else
1130 		cpuc->lbr_stack.hw_idx = intel_pmu_lbr_tos();
1131 
1132 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
1133 		u64 info = lbr->lbr[i].info;
1134 		struct perf_branch_entry *e = &cpuc->lbr_entries[i];
1135 
1136 		e->from		= lbr->lbr[i].from;
1137 		e->to		= lbr->lbr[i].to;
1138 		e->mispred	= !!(info & LBR_INFO_MISPRED);
1139 		e->predicted	= !(info & LBR_INFO_MISPRED);
1140 		e->in_tx	= !!(info & LBR_INFO_IN_TX);
1141 		e->abort	= !!(info & LBR_INFO_ABORT);
1142 		e->cycles	= info & LBR_INFO_CYCLES;
1143 		e->reserved	= 0;
1144 	}
1145 	intel_pmu_lbr_filter(cpuc);
1146 }
1147 
1148 /*
1149  * Map interface branch filters onto LBR filters
1150  */
1151 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1152 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1153 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1154 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1155 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1156 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1157 						| LBR_IND_JMP | LBR_FAR,
1158 	/*
1159 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1160 	 */
1161 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1162 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1163 	/*
1164 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1165 	 */
1166 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1167 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1168 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1169 };
1170 
1171 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1172 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1173 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1174 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1175 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1176 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1177 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1178 						| LBR_FAR,
1179 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1180 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1181 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1182 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1183 };
1184 
1185 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1186 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1187 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1188 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1189 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1190 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1191 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1192 						| LBR_FAR,
1193 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1194 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1195 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1196 						| LBR_RETURN | LBR_CALL_STACK,
1197 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1198 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1199 };
1200 
1201 /* core */
1202 void __init intel_pmu_lbr_init_core(void)
1203 {
1204 	x86_pmu.lbr_nr     = 4;
1205 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1206 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1207 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1208 
1209 	/*
1210 	 * SW branch filter usage:
1211 	 * - compensate for lack of HW filter
1212 	 */
1213 }
1214 
1215 /* nehalem/westmere */
1216 void __init intel_pmu_lbr_init_nhm(void)
1217 {
1218 	x86_pmu.lbr_nr     = 16;
1219 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1220 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1221 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1222 
1223 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1224 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1225 
1226 	/*
1227 	 * SW branch filter usage:
1228 	 * - workaround LBR_SEL errata (see above)
1229 	 * - support syscall, sysret capture.
1230 	 *   That requires LBR_FAR but that means far
1231 	 *   jmp need to be filtered out
1232 	 */
1233 }
1234 
1235 /* sandy bridge */
1236 void __init intel_pmu_lbr_init_snb(void)
1237 {
1238 	x86_pmu.lbr_nr	 = 16;
1239 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1240 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1241 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1242 
1243 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1244 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1245 
1246 	/*
1247 	 * SW branch filter usage:
1248 	 * - support syscall, sysret capture.
1249 	 *   That requires LBR_FAR but that means far
1250 	 *   jmp need to be filtered out
1251 	 */
1252 }
1253 
1254 /* haswell */
1255 void intel_pmu_lbr_init_hsw(void)
1256 {
1257 	x86_pmu.lbr_nr	 = 16;
1258 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1259 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1260 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1261 
1262 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1263 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1264 
1265 	if (lbr_from_signext_quirk_needed())
1266 		static_branch_enable(&lbr_from_quirk_key);
1267 }
1268 
1269 /* skylake */
1270 __init void intel_pmu_lbr_init_skl(void)
1271 {
1272 	x86_pmu.lbr_nr	 = 32;
1273 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1274 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1275 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1276 
1277 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1278 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1279 
1280 	/*
1281 	 * SW branch filter usage:
1282 	 * - support syscall, sysret capture.
1283 	 *   That requires LBR_FAR but that means far
1284 	 *   jmp need to be filtered out
1285 	 */
1286 }
1287 
1288 /* atom */
1289 void __init intel_pmu_lbr_init_atom(void)
1290 {
1291 	/*
1292 	 * only models starting at stepping 10 seems
1293 	 * to have an operational LBR which can freeze
1294 	 * on PMU interrupt
1295 	 */
1296 	if (boot_cpu_data.x86_model == 28
1297 	    && boot_cpu_data.x86_stepping < 10) {
1298 		pr_cont("LBR disabled due to erratum");
1299 		return;
1300 	}
1301 
1302 	x86_pmu.lbr_nr	   = 8;
1303 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1304 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1305 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1306 
1307 	/*
1308 	 * SW branch filter usage:
1309 	 * - compensate for lack of HW filter
1310 	 */
1311 }
1312 
1313 /* slm */
1314 void __init intel_pmu_lbr_init_slm(void)
1315 {
1316 	x86_pmu.lbr_nr	   = 8;
1317 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1318 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1319 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1320 
1321 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1322 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1323 
1324 	/*
1325 	 * SW branch filter usage:
1326 	 * - compensate for lack of HW filter
1327 	 */
1328 	pr_cont("8-deep LBR, ");
1329 }
1330 
1331 /* Knights Landing */
1332 void intel_pmu_lbr_init_knl(void)
1333 {
1334 	x86_pmu.lbr_nr	   = 8;
1335 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1336 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1337 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1338 
1339 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1340 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1341 
1342 	/* Knights Landing does have MISPREDICT bit */
1343 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1344 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1345 }
1346