xref: /openbmc/linux/arch/x86/events/intel/lbr.c (revision 58f9d806)
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 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_sched_task(struct perf_event_context *ctx, bool sched_in)
421 {
422 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
423 	struct x86_perf_task_context *task_ctx;
424 
425 	if (!cpuc->lbr_users)
426 		return;
427 
428 	/*
429 	 * If LBR callstack feature is enabled and the stack was saved when
430 	 * the task was scheduled out, restore the stack. Otherwise flush
431 	 * the LBR stack.
432 	 */
433 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
434 	if (task_ctx) {
435 		if (sched_in)
436 			__intel_pmu_lbr_restore(task_ctx);
437 		else
438 			__intel_pmu_lbr_save(task_ctx);
439 		return;
440 	}
441 
442 	/*
443 	 * Since a context switch can flip the address space and LBR entries
444 	 * are not tagged with an identifier, we need to wipe the LBR, even for
445 	 * per-cpu events. You simply cannot resolve the branches from the old
446 	 * address space.
447 	 */
448 	if (sched_in)
449 		intel_pmu_lbr_reset();
450 }
451 
452 static inline bool branch_user_callstack(unsigned br_sel)
453 {
454 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
455 }
456 
457 void intel_pmu_lbr_add(struct perf_event *event)
458 {
459 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
460 	struct x86_perf_task_context *task_ctx;
461 
462 	if (!x86_pmu.lbr_nr)
463 		return;
464 
465 	cpuc->br_sel = event->hw.branch_reg.reg;
466 
467 	if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data) {
468 		task_ctx = event->ctx->task_ctx_data;
469 		task_ctx->lbr_callstack_users++;
470 	}
471 
472 	/*
473 	 * Request pmu::sched_task() callback, which will fire inside the
474 	 * regular perf event scheduling, so that call will:
475 	 *
476 	 *  - restore or wipe; when LBR-callstack,
477 	 *  - wipe; otherwise,
478 	 *
479 	 * when this is from __perf_event_task_sched_in().
480 	 *
481 	 * However, if this is from perf_install_in_context(), no such callback
482 	 * will follow and we'll need to reset the LBR here if this is the
483 	 * first LBR event.
484 	 *
485 	 * The problem is, we cannot tell these cases apart... but we can
486 	 * exclude the biggest chunk of cases by looking at
487 	 * event->total_time_running. An event that has accrued runtime cannot
488 	 * be 'new'. Conversely, a new event can get installed through the
489 	 * context switch path for the first time.
490 	 */
491 	perf_sched_cb_inc(event->ctx->pmu);
492 	if (!cpuc->lbr_users++ && !event->total_time_running)
493 		intel_pmu_lbr_reset();
494 }
495 
496 void intel_pmu_lbr_del(struct perf_event *event)
497 {
498 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
499 	struct x86_perf_task_context *task_ctx;
500 
501 	if (!x86_pmu.lbr_nr)
502 		return;
503 
504 	if (branch_user_callstack(cpuc->br_sel) &&
505 	    event->ctx->task_ctx_data) {
506 		task_ctx = event->ctx->task_ctx_data;
507 		task_ctx->lbr_callstack_users--;
508 	}
509 
510 	cpuc->lbr_users--;
511 	WARN_ON_ONCE(cpuc->lbr_users < 0);
512 	perf_sched_cb_dec(event->ctx->pmu);
513 }
514 
515 void intel_pmu_lbr_enable_all(bool pmi)
516 {
517 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
518 
519 	if (cpuc->lbr_users)
520 		__intel_pmu_lbr_enable(pmi);
521 }
522 
523 void intel_pmu_lbr_disable_all(void)
524 {
525 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
526 
527 	if (cpuc->lbr_users)
528 		__intel_pmu_lbr_disable();
529 }
530 
531 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
532 {
533 	unsigned long mask = x86_pmu.lbr_nr - 1;
534 	u64 tos = intel_pmu_lbr_tos();
535 	int i;
536 
537 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
538 		unsigned long lbr_idx = (tos - i) & mask;
539 		union {
540 			struct {
541 				u32 from;
542 				u32 to;
543 			};
544 			u64     lbr;
545 		} msr_lastbranch;
546 
547 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
548 
549 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
550 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
551 		cpuc->lbr_entries[i].mispred	= 0;
552 		cpuc->lbr_entries[i].predicted	= 0;
553 		cpuc->lbr_entries[i].in_tx	= 0;
554 		cpuc->lbr_entries[i].abort	= 0;
555 		cpuc->lbr_entries[i].cycles	= 0;
556 		cpuc->lbr_entries[i].type	= 0;
557 		cpuc->lbr_entries[i].reserved	= 0;
558 	}
559 	cpuc->lbr_stack.nr = i;
560 }
561 
562 /*
563  * Due to lack of segmentation in Linux the effective address (offset)
564  * is the same as the linear address, allowing us to merge the LIP and EIP
565  * LBR formats.
566  */
567 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
568 {
569 	bool need_info = false, call_stack = false;
570 	unsigned long mask = x86_pmu.lbr_nr - 1;
571 	int lbr_format = x86_pmu.intel_cap.lbr_format;
572 	u64 tos = intel_pmu_lbr_tos();
573 	int i;
574 	int out = 0;
575 	int num = x86_pmu.lbr_nr;
576 
577 	if (cpuc->lbr_sel) {
578 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
579 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
580 			call_stack = true;
581 	}
582 
583 	for (i = 0; i < num; i++) {
584 		unsigned long lbr_idx = (tos - i) & mask;
585 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
586 		int skip = 0;
587 		u16 cycles = 0;
588 		int lbr_flags = lbr_desc[lbr_format];
589 
590 		from = rdlbr_from(lbr_idx);
591 		to   = rdlbr_to(lbr_idx);
592 
593 		/*
594 		 * Read LBR call stack entries
595 		 * until invalid entry (0s) is detected.
596 		 */
597 		if (call_stack && !from)
598 			break;
599 
600 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
601 			u64 info;
602 
603 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, info);
604 			mis = !!(info & LBR_INFO_MISPRED);
605 			pred = !mis;
606 			in_tx = !!(info & LBR_INFO_IN_TX);
607 			abort = !!(info & LBR_INFO_ABORT);
608 			cycles = (info & LBR_INFO_CYCLES);
609 		}
610 
611 		if (lbr_format == LBR_FORMAT_TIME) {
612 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
613 			pred = !mis;
614 			skip = 1;
615 			cycles = ((to >> 48) & LBR_INFO_CYCLES);
616 
617 			to = (u64)((((s64)to) << 16) >> 16);
618 		}
619 
620 		if (lbr_flags & LBR_EIP_FLAGS) {
621 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
622 			pred = !mis;
623 			skip = 1;
624 		}
625 		if (lbr_flags & LBR_TSX) {
626 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
627 			abort = !!(from & LBR_FROM_FLAG_ABORT);
628 			skip = 3;
629 		}
630 		from = (u64)((((s64)from) << skip) >> skip);
631 
632 		/*
633 		 * Some CPUs report duplicated abort records,
634 		 * with the second entry not having an abort bit set.
635 		 * Skip them here. This loop runs backwards,
636 		 * so we need to undo the previous record.
637 		 * If the abort just happened outside the window
638 		 * the extra entry cannot be removed.
639 		 */
640 		if (abort && x86_pmu.lbr_double_abort && out > 0)
641 			out--;
642 
643 		cpuc->lbr_entries[out].from	 = from;
644 		cpuc->lbr_entries[out].to	 = to;
645 		cpuc->lbr_entries[out].mispred	 = mis;
646 		cpuc->lbr_entries[out].predicted = pred;
647 		cpuc->lbr_entries[out].in_tx	 = in_tx;
648 		cpuc->lbr_entries[out].abort	 = abort;
649 		cpuc->lbr_entries[out].cycles	 = cycles;
650 		cpuc->lbr_entries[out].type	 = 0;
651 		cpuc->lbr_entries[out].reserved	 = 0;
652 		out++;
653 	}
654 	cpuc->lbr_stack.nr = out;
655 }
656 
657 void intel_pmu_lbr_read(void)
658 {
659 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
660 
661 	if (!cpuc->lbr_users)
662 		return;
663 
664 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
665 		intel_pmu_lbr_read_32(cpuc);
666 	else
667 		intel_pmu_lbr_read_64(cpuc);
668 
669 	intel_pmu_lbr_filter(cpuc);
670 }
671 
672 /*
673  * SW filter is used:
674  * - in case there is no HW filter
675  * - in case the HW filter has errata or limitations
676  */
677 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
678 {
679 	u64 br_type = event->attr.branch_sample_type;
680 	int mask = 0;
681 
682 	if (br_type & PERF_SAMPLE_BRANCH_USER)
683 		mask |= X86_BR_USER;
684 
685 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
686 		mask |= X86_BR_KERNEL;
687 
688 	/* we ignore BRANCH_HV here */
689 
690 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
691 		mask |= X86_BR_ANY;
692 
693 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
694 		mask |= X86_BR_ANY_CALL;
695 
696 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
697 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
698 
699 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
700 		mask |= X86_BR_IND_CALL;
701 
702 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
703 		mask |= X86_BR_ABORT;
704 
705 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
706 		mask |= X86_BR_IN_TX;
707 
708 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
709 		mask |= X86_BR_NO_TX;
710 
711 	if (br_type & PERF_SAMPLE_BRANCH_COND)
712 		mask |= X86_BR_JCC;
713 
714 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
715 		if (!x86_pmu_has_lbr_callstack())
716 			return -EOPNOTSUPP;
717 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
718 			return -EINVAL;
719 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
720 			X86_BR_CALL_STACK;
721 	}
722 
723 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
724 		mask |= X86_BR_IND_JMP;
725 
726 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
727 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
728 
729 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
730 		mask |= X86_BR_TYPE_SAVE;
731 
732 	/*
733 	 * stash actual user request into reg, it may
734 	 * be used by fixup code for some CPU
735 	 */
736 	event->hw.branch_reg.reg = mask;
737 	return 0;
738 }
739 
740 /*
741  * setup the HW LBR filter
742  * Used only when available, may not be enough to disambiguate
743  * all branches, may need the help of the SW filter
744  */
745 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
746 {
747 	struct hw_perf_event_extra *reg;
748 	u64 br_type = event->attr.branch_sample_type;
749 	u64 mask = 0, v;
750 	int i;
751 
752 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
753 		if (!(br_type & (1ULL << i)))
754 			continue;
755 
756 		v = x86_pmu.lbr_sel_map[i];
757 		if (v == LBR_NOT_SUPP)
758 			return -EOPNOTSUPP;
759 
760 		if (v != LBR_IGN)
761 			mask |= v;
762 	}
763 
764 	reg = &event->hw.branch_reg;
765 	reg->idx = EXTRA_REG_LBR;
766 
767 	/*
768 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
769 	 * in suppress mode. So LBR_SELECT should be set to
770 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
771 	 * But the 10th bit LBR_CALL_STACK does not operate
772 	 * in suppress mode.
773 	 */
774 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
775 
776 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
777 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
778 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
779 		reg->config |= LBR_NO_INFO;
780 
781 	return 0;
782 }
783 
784 int intel_pmu_setup_lbr_filter(struct perf_event *event)
785 {
786 	int ret = 0;
787 
788 	/*
789 	 * no LBR on this PMU
790 	 */
791 	if (!x86_pmu.lbr_nr)
792 		return -EOPNOTSUPP;
793 
794 	/*
795 	 * setup SW LBR filter
796 	 */
797 	ret = intel_pmu_setup_sw_lbr_filter(event);
798 	if (ret)
799 		return ret;
800 
801 	/*
802 	 * setup HW LBR filter, if any
803 	 */
804 	if (x86_pmu.lbr_sel_map)
805 		ret = intel_pmu_setup_hw_lbr_filter(event);
806 
807 	return ret;
808 }
809 
810 /*
811  * return the type of control flow change at address "from"
812  * instruction is not necessarily a branch (in case of interrupt).
813  *
814  * The branch type returned also includes the priv level of the
815  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
816  *
817  * If a branch type is unknown OR the instruction cannot be
818  * decoded (e.g., text page not present), then X86_BR_NONE is
819  * returned.
820  */
821 static int branch_type(unsigned long from, unsigned long to, int abort)
822 {
823 	struct insn insn;
824 	void *addr;
825 	int bytes_read, bytes_left;
826 	int ret = X86_BR_NONE;
827 	int ext, to_plm, from_plm;
828 	u8 buf[MAX_INSN_SIZE];
829 	int is64 = 0;
830 
831 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
832 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
833 
834 	/*
835 	 * maybe zero if lbr did not fill up after a reset by the time
836 	 * we get a PMU interrupt
837 	 */
838 	if (from == 0 || to == 0)
839 		return X86_BR_NONE;
840 
841 	if (abort)
842 		return X86_BR_ABORT | to_plm;
843 
844 	if (from_plm == X86_BR_USER) {
845 		/*
846 		 * can happen if measuring at the user level only
847 		 * and we interrupt in a kernel thread, e.g., idle.
848 		 */
849 		if (!current->mm)
850 			return X86_BR_NONE;
851 
852 		/* may fail if text not present */
853 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
854 						MAX_INSN_SIZE);
855 		bytes_read = MAX_INSN_SIZE - bytes_left;
856 		if (!bytes_read)
857 			return X86_BR_NONE;
858 
859 		addr = buf;
860 	} else {
861 		/*
862 		 * The LBR logs any address in the IP, even if the IP just
863 		 * faulted. This means userspace can control the from address.
864 		 * Ensure we don't blindy read any address by validating it is
865 		 * a known text address.
866 		 */
867 		if (kernel_text_address(from)) {
868 			addr = (void *)from;
869 			/*
870 			 * Assume we can get the maximum possible size
871 			 * when grabbing kernel data.  This is not
872 			 * _strictly_ true since we could possibly be
873 			 * executing up next to a memory hole, but
874 			 * it is very unlikely to be a problem.
875 			 */
876 			bytes_read = MAX_INSN_SIZE;
877 		} else {
878 			return X86_BR_NONE;
879 		}
880 	}
881 
882 	/*
883 	 * decoder needs to know the ABI especially
884 	 * on 64-bit systems running 32-bit apps
885 	 */
886 #ifdef CONFIG_X86_64
887 	is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
888 #endif
889 	insn_init(&insn, addr, bytes_read, is64);
890 	insn_get_opcode(&insn);
891 	if (!insn.opcode.got)
892 		return X86_BR_ABORT;
893 
894 	switch (insn.opcode.bytes[0]) {
895 	case 0xf:
896 		switch (insn.opcode.bytes[1]) {
897 		case 0x05: /* syscall */
898 		case 0x34: /* sysenter */
899 			ret = X86_BR_SYSCALL;
900 			break;
901 		case 0x07: /* sysret */
902 		case 0x35: /* sysexit */
903 			ret = X86_BR_SYSRET;
904 			break;
905 		case 0x80 ... 0x8f: /* conditional */
906 			ret = X86_BR_JCC;
907 			break;
908 		default:
909 			ret = X86_BR_NONE;
910 		}
911 		break;
912 	case 0x70 ... 0x7f: /* conditional */
913 		ret = X86_BR_JCC;
914 		break;
915 	case 0xc2: /* near ret */
916 	case 0xc3: /* near ret */
917 	case 0xca: /* far ret */
918 	case 0xcb: /* far ret */
919 		ret = X86_BR_RET;
920 		break;
921 	case 0xcf: /* iret */
922 		ret = X86_BR_IRET;
923 		break;
924 	case 0xcc ... 0xce: /* int */
925 		ret = X86_BR_INT;
926 		break;
927 	case 0xe8: /* call near rel */
928 		insn_get_immediate(&insn);
929 		if (insn.immediate1.value == 0) {
930 			/* zero length call */
931 			ret = X86_BR_ZERO_CALL;
932 			break;
933 		}
934 	case 0x9a: /* call far absolute */
935 		ret = X86_BR_CALL;
936 		break;
937 	case 0xe0 ... 0xe3: /* loop jmp */
938 		ret = X86_BR_JCC;
939 		break;
940 	case 0xe9 ... 0xeb: /* jmp */
941 		ret = X86_BR_JMP;
942 		break;
943 	case 0xff: /* call near absolute, call far absolute ind */
944 		insn_get_modrm(&insn);
945 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
946 		switch (ext) {
947 		case 2: /* near ind call */
948 		case 3: /* far ind call */
949 			ret = X86_BR_IND_CALL;
950 			break;
951 		case 4:
952 		case 5:
953 			ret = X86_BR_IND_JMP;
954 			break;
955 		}
956 		break;
957 	default:
958 		ret = X86_BR_NONE;
959 	}
960 	/*
961 	 * interrupts, traps, faults (and thus ring transition) may
962 	 * occur on any instructions. Thus, to classify them correctly,
963 	 * we need to first look at the from and to priv levels. If they
964 	 * are different and to is in the kernel, then it indicates
965 	 * a ring transition. If the from instruction is not a ring
966 	 * transition instr (syscall, systenter, int), then it means
967 	 * it was a irq, trap or fault.
968 	 *
969 	 * we have no way of detecting kernel to kernel faults.
970 	 */
971 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
972 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
973 		ret = X86_BR_IRQ;
974 
975 	/*
976 	 * branch priv level determined by target as
977 	 * is done by HW when LBR_SELECT is implemented
978 	 */
979 	if (ret != X86_BR_NONE)
980 		ret |= to_plm;
981 
982 	return ret;
983 }
984 
985 #define X86_BR_TYPE_MAP_MAX	16
986 
987 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
988 	PERF_BR_CALL,		/* X86_BR_CALL */
989 	PERF_BR_RET,		/* X86_BR_RET */
990 	PERF_BR_SYSCALL,	/* X86_BR_SYSCALL */
991 	PERF_BR_SYSRET,		/* X86_BR_SYSRET */
992 	PERF_BR_UNKNOWN,	/* X86_BR_INT */
993 	PERF_BR_UNKNOWN,	/* X86_BR_IRET */
994 	PERF_BR_COND,		/* X86_BR_JCC */
995 	PERF_BR_UNCOND,		/* X86_BR_JMP */
996 	PERF_BR_UNKNOWN,	/* X86_BR_IRQ */
997 	PERF_BR_IND_CALL,	/* X86_BR_IND_CALL */
998 	PERF_BR_UNKNOWN,	/* X86_BR_ABORT */
999 	PERF_BR_UNKNOWN,	/* X86_BR_IN_TX */
1000 	PERF_BR_UNKNOWN,	/* X86_BR_NO_TX */
1001 	PERF_BR_CALL,		/* X86_BR_ZERO_CALL */
1002 	PERF_BR_UNKNOWN,	/* X86_BR_CALL_STACK */
1003 	PERF_BR_IND,		/* X86_BR_IND_JMP */
1004 };
1005 
1006 static int
1007 common_branch_type(int type)
1008 {
1009 	int i;
1010 
1011 	type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1012 
1013 	if (type) {
1014 		i = __ffs(type);
1015 		if (i < X86_BR_TYPE_MAP_MAX)
1016 			return branch_map[i];
1017 	}
1018 
1019 	return PERF_BR_UNKNOWN;
1020 }
1021 
1022 /*
1023  * implement actual branch filter based on user demand.
1024  * Hardware may not exactly satisfy that request, thus
1025  * we need to inspect opcodes. Mismatched branches are
1026  * discarded. Therefore, the number of branches returned
1027  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1028  */
1029 static void
1030 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1031 {
1032 	u64 from, to;
1033 	int br_sel = cpuc->br_sel;
1034 	int i, j, type;
1035 	bool compress = false;
1036 
1037 	/* if sampling all branches, then nothing to filter */
1038 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1039 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1040 		return;
1041 
1042 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1043 
1044 		from = cpuc->lbr_entries[i].from;
1045 		to = cpuc->lbr_entries[i].to;
1046 
1047 		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1048 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1049 			if (cpuc->lbr_entries[i].in_tx)
1050 				type |= X86_BR_IN_TX;
1051 			else
1052 				type |= X86_BR_NO_TX;
1053 		}
1054 
1055 		/* if type does not correspond, then discard */
1056 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1057 			cpuc->lbr_entries[i].from = 0;
1058 			compress = true;
1059 		}
1060 
1061 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1062 			cpuc->lbr_entries[i].type = common_branch_type(type);
1063 	}
1064 
1065 	if (!compress)
1066 		return;
1067 
1068 	/* remove all entries with from=0 */
1069 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1070 		if (!cpuc->lbr_entries[i].from) {
1071 			j = i;
1072 			while (++j < cpuc->lbr_stack.nr)
1073 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1074 			cpuc->lbr_stack.nr--;
1075 			if (!cpuc->lbr_entries[i].from)
1076 				continue;
1077 		}
1078 		i++;
1079 	}
1080 }
1081 
1082 /*
1083  * Map interface branch filters onto LBR filters
1084  */
1085 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1086 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1087 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1088 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1089 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1090 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1091 						| LBR_IND_JMP | LBR_FAR,
1092 	/*
1093 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1094 	 */
1095 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1096 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1097 	/*
1098 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1099 	 */
1100 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1101 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1102 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1103 };
1104 
1105 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1106 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1107 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1108 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1109 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1110 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1111 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1112 						| LBR_FAR,
1113 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1114 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1115 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1116 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1117 };
1118 
1119 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1120 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1121 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1122 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1123 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1124 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1125 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1126 						| LBR_FAR,
1127 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1128 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1129 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1130 						| LBR_RETURN | LBR_CALL_STACK,
1131 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1132 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1133 };
1134 
1135 /* core */
1136 void __init intel_pmu_lbr_init_core(void)
1137 {
1138 	x86_pmu.lbr_nr     = 4;
1139 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1140 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1141 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1142 
1143 	/*
1144 	 * SW branch filter usage:
1145 	 * - compensate for lack of HW filter
1146 	 */
1147 }
1148 
1149 /* nehalem/westmere */
1150 void __init intel_pmu_lbr_init_nhm(void)
1151 {
1152 	x86_pmu.lbr_nr     = 16;
1153 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1154 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1155 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1156 
1157 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1158 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1159 
1160 	/*
1161 	 * SW branch filter usage:
1162 	 * - workaround LBR_SEL errata (see above)
1163 	 * - support syscall, sysret capture.
1164 	 *   That requires LBR_FAR but that means far
1165 	 *   jmp need to be filtered out
1166 	 */
1167 }
1168 
1169 /* sandy bridge */
1170 void __init intel_pmu_lbr_init_snb(void)
1171 {
1172 	x86_pmu.lbr_nr	 = 16;
1173 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1174 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1175 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1176 
1177 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1178 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1179 
1180 	/*
1181 	 * SW branch filter usage:
1182 	 * - support syscall, sysret capture.
1183 	 *   That requires LBR_FAR but that means far
1184 	 *   jmp need to be filtered out
1185 	 */
1186 }
1187 
1188 /* haswell */
1189 void intel_pmu_lbr_init_hsw(void)
1190 {
1191 	x86_pmu.lbr_nr	 = 16;
1192 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1193 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1194 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1195 
1196 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1197 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1198 
1199 	if (lbr_from_signext_quirk_needed())
1200 		static_branch_enable(&lbr_from_quirk_key);
1201 }
1202 
1203 /* skylake */
1204 __init void intel_pmu_lbr_init_skl(void)
1205 {
1206 	x86_pmu.lbr_nr	 = 32;
1207 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1208 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1209 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1210 
1211 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1212 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1213 
1214 	/*
1215 	 * SW branch filter usage:
1216 	 * - support syscall, sysret capture.
1217 	 *   That requires LBR_FAR but that means far
1218 	 *   jmp need to be filtered out
1219 	 */
1220 }
1221 
1222 /* atom */
1223 void __init intel_pmu_lbr_init_atom(void)
1224 {
1225 	/*
1226 	 * only models starting at stepping 10 seems
1227 	 * to have an operational LBR which can freeze
1228 	 * on PMU interrupt
1229 	 */
1230 	if (boot_cpu_data.x86_model == 28
1231 	    && boot_cpu_data.x86_stepping < 10) {
1232 		pr_cont("LBR disabled due to erratum");
1233 		return;
1234 	}
1235 
1236 	x86_pmu.lbr_nr	   = 8;
1237 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1238 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1239 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1240 
1241 	/*
1242 	 * SW branch filter usage:
1243 	 * - compensate for lack of HW filter
1244 	 */
1245 }
1246 
1247 /* slm */
1248 void __init intel_pmu_lbr_init_slm(void)
1249 {
1250 	x86_pmu.lbr_nr	   = 8;
1251 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1252 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1253 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1254 
1255 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1256 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1257 
1258 	/*
1259 	 * SW branch filter usage:
1260 	 * - compensate for lack of HW filter
1261 	 */
1262 	pr_cont("8-deep LBR, ");
1263 }
1264 
1265 /* Knights Landing */
1266 void intel_pmu_lbr_init_knl(void)
1267 {
1268 	x86_pmu.lbr_nr	   = 8;
1269 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1270 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1271 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1272 
1273 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1274 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1275 
1276 	/* Knights Landing does have MISPREDICT bit */
1277 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1278 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1279 }
1280