xref: /linux/arch/x86/events/intel/lbr.c (revision 9125f19bbb97e7444b5ea01f8262fa09679e1376)
1 #include <linux/perf_event.h>
2 #include <linux/types.h>
3 
4 #include <asm/perf_event.h>
5 #include <asm/msr.h>
6 #include <asm/insn.h>
7 
8 #include "../perf_event.h"
9 
10 enum {
11 	LBR_FORMAT_32		= 0x00,
12 	LBR_FORMAT_LIP		= 0x01,
13 	LBR_FORMAT_EIP		= 0x02,
14 	LBR_FORMAT_EIP_FLAGS	= 0x03,
15 	LBR_FORMAT_EIP_FLAGS2	= 0x04,
16 	LBR_FORMAT_INFO		= 0x05,
17 	LBR_FORMAT_MAX_KNOWN    = LBR_FORMAT_INFO,
18 };
19 
20 static enum {
21 	LBR_EIP_FLAGS		= 1,
22 	LBR_TSX			= 2,
23 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
24 	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
25 	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
26 };
27 
28 /*
29  * Intel LBR_SELECT bits
30  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
31  *
32  * Hardware branch filter (not available on all CPUs)
33  */
34 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
35 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
36 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
37 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
38 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
39 #define LBR_RETURN_BIT		5 /* do not capture near returns */
40 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
41 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
42 #define LBR_FAR_BIT		8 /* do not capture far branches */
43 #define LBR_CALL_STACK_BIT	9 /* enable call stack */
44 
45 /*
46  * Following bit only exists in Linux; we mask it out before writing it to
47  * the actual MSR. But it helps the constraint perf code to understand
48  * that this is a separate configuration.
49  */
50 #define LBR_NO_INFO_BIT	       63 /* don't read LBR_INFO. */
51 
52 #define LBR_KERNEL	(1 << LBR_KERNEL_BIT)
53 #define LBR_USER	(1 << LBR_USER_BIT)
54 #define LBR_JCC		(1 << LBR_JCC_BIT)
55 #define LBR_REL_CALL	(1 << LBR_REL_CALL_BIT)
56 #define LBR_IND_CALL	(1 << LBR_IND_CALL_BIT)
57 #define LBR_RETURN	(1 << LBR_RETURN_BIT)
58 #define LBR_REL_JMP	(1 << LBR_REL_JMP_BIT)
59 #define LBR_IND_JMP	(1 << LBR_IND_JMP_BIT)
60 #define LBR_FAR		(1 << LBR_FAR_BIT)
61 #define LBR_CALL_STACK	(1 << LBR_CALL_STACK_BIT)
62 #define LBR_NO_INFO	(1ULL << LBR_NO_INFO_BIT)
63 
64 #define LBR_PLM (LBR_KERNEL | LBR_USER)
65 
66 #define LBR_SEL_MASK	0x3ff	/* valid bits in LBR_SELECT */
67 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
68 #define LBR_IGN		0	/* ignored */
69 
70 #define LBR_ANY		 \
71 	(LBR_JCC	|\
72 	 LBR_REL_CALL	|\
73 	 LBR_IND_CALL	|\
74 	 LBR_RETURN	|\
75 	 LBR_REL_JMP	|\
76 	 LBR_IND_JMP	|\
77 	 LBR_FAR)
78 
79 #define LBR_FROM_FLAG_MISPRED  (1ULL << 63)
80 #define LBR_FROM_FLAG_IN_TX    (1ULL << 62)
81 #define LBR_FROM_FLAG_ABORT    (1ULL << 61)
82 
83 /*
84  * x86control flow change classification
85  * x86control flow changes include branches, interrupts, traps, faults
86  */
87 enum {
88 	X86_BR_NONE		= 0,      /* unknown */
89 
90 	X86_BR_USER		= 1 << 0, /* branch target is user */
91 	X86_BR_KERNEL		= 1 << 1, /* branch target is kernel */
92 
93 	X86_BR_CALL		= 1 << 2, /* call */
94 	X86_BR_RET		= 1 << 3, /* return */
95 	X86_BR_SYSCALL		= 1 << 4, /* syscall */
96 	X86_BR_SYSRET		= 1 << 5, /* syscall return */
97 	X86_BR_INT		= 1 << 6, /* sw interrupt */
98 	X86_BR_IRET		= 1 << 7, /* return from interrupt */
99 	X86_BR_JCC		= 1 << 8, /* conditional */
100 	X86_BR_JMP		= 1 << 9, /* jump */
101 	X86_BR_IRQ		= 1 << 10,/* hw interrupt or trap or fault */
102 	X86_BR_IND_CALL		= 1 << 11,/* indirect calls */
103 	X86_BR_ABORT		= 1 << 12,/* transaction abort */
104 	X86_BR_IN_TX		= 1 << 13,/* in transaction */
105 	X86_BR_NO_TX		= 1 << 14,/* not in transaction */
106 	X86_BR_ZERO_CALL	= 1 << 15,/* zero length call */
107 	X86_BR_CALL_STACK	= 1 << 16,/* call stack */
108 	X86_BR_IND_JMP		= 1 << 17,/* indirect jump */
109 };
110 
111 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
112 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
113 
114 #define X86_BR_ANY       \
115 	(X86_BR_CALL    |\
116 	 X86_BR_RET     |\
117 	 X86_BR_SYSCALL |\
118 	 X86_BR_SYSRET  |\
119 	 X86_BR_INT     |\
120 	 X86_BR_IRET    |\
121 	 X86_BR_JCC     |\
122 	 X86_BR_JMP	 |\
123 	 X86_BR_IRQ	 |\
124 	 X86_BR_ABORT	 |\
125 	 X86_BR_IND_CALL |\
126 	 X86_BR_IND_JMP  |\
127 	 X86_BR_ZERO_CALL)
128 
129 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
130 
131 #define X86_BR_ANY_CALL		 \
132 	(X86_BR_CALL		|\
133 	 X86_BR_IND_CALL	|\
134 	 X86_BR_ZERO_CALL	|\
135 	 X86_BR_SYSCALL		|\
136 	 X86_BR_IRQ		|\
137 	 X86_BR_INT)
138 
139 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
140 
141 /*
142  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
143  * otherwise it becomes near impossible to get a reliable stack.
144  */
145 
146 static void __intel_pmu_lbr_enable(bool pmi)
147 {
148 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
149 	u64 debugctl, lbr_select = 0, orig_debugctl;
150 
151 	/*
152 	 * No need to unfreeze manually, as v4 can do that as part
153 	 * of the GLOBAL_STATUS ack.
154 	 */
155 	if (pmi && x86_pmu.version >= 4)
156 		return;
157 
158 	/*
159 	 * No need to reprogram LBR_SELECT in a PMI, as it
160 	 * did not change.
161 	 */
162 	if (cpuc->lbr_sel)
163 		lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
164 	if (!pmi && cpuc->lbr_sel)
165 		wrmsrl(MSR_LBR_SELECT, lbr_select);
166 
167 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
168 	orig_debugctl = debugctl;
169 	debugctl |= DEBUGCTLMSR_LBR;
170 	/*
171 	 * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
172 	 * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
173 	 * may cause superfluous increase/decrease of LBR_TOS.
174 	 */
175 	if (!(lbr_select & LBR_CALL_STACK))
176 		debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
177 	if (orig_debugctl != debugctl)
178 		wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
179 }
180 
181 static void __intel_pmu_lbr_disable(void)
182 {
183 	u64 debugctl;
184 
185 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
186 	debugctl &= ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
187 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
188 }
189 
190 static void intel_pmu_lbr_reset_32(void)
191 {
192 	int i;
193 
194 	for (i = 0; i < x86_pmu.lbr_nr; i++)
195 		wrmsrl(x86_pmu.lbr_from + i, 0);
196 }
197 
198 static void intel_pmu_lbr_reset_64(void)
199 {
200 	int i;
201 
202 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
203 		wrmsrl(x86_pmu.lbr_from + i, 0);
204 		wrmsrl(x86_pmu.lbr_to   + i, 0);
205 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
206 			wrmsrl(MSR_LBR_INFO_0 + i, 0);
207 	}
208 }
209 
210 void intel_pmu_lbr_reset(void)
211 {
212 	if (!x86_pmu.lbr_nr)
213 		return;
214 
215 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
216 		intel_pmu_lbr_reset_32();
217 	else
218 		intel_pmu_lbr_reset_64();
219 }
220 
221 /*
222  * TOS = most recently recorded branch
223  */
224 static inline u64 intel_pmu_lbr_tos(void)
225 {
226 	u64 tos;
227 
228 	rdmsrl(x86_pmu.lbr_tos, tos);
229 	return tos;
230 }
231 
232 enum {
233 	LBR_NONE,
234 	LBR_VALID,
235 };
236 
237 static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
238 {
239 	int i;
240 	unsigned lbr_idx, mask;
241 	u64 tos;
242 
243 	if (task_ctx->lbr_callstack_users == 0 ||
244 	    task_ctx->lbr_stack_state == LBR_NONE) {
245 		intel_pmu_lbr_reset();
246 		return;
247 	}
248 
249 	mask = x86_pmu.lbr_nr - 1;
250 	tos = task_ctx->tos;
251 	for (i = 0; i < tos; i++) {
252 		lbr_idx = (tos - i) & mask;
253 		wrmsrl(x86_pmu.lbr_from + lbr_idx, task_ctx->lbr_from[i]);
254 		wrmsrl(x86_pmu.lbr_to + lbr_idx, task_ctx->lbr_to[i]);
255 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
256 			wrmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
257 	}
258 	wrmsrl(x86_pmu.lbr_tos, tos);
259 	task_ctx->lbr_stack_state = LBR_NONE;
260 }
261 
262 static void __intel_pmu_lbr_save(struct x86_perf_task_context *task_ctx)
263 {
264 	int i;
265 	unsigned lbr_idx, mask;
266 	u64 tos;
267 
268 	if (task_ctx->lbr_callstack_users == 0) {
269 		task_ctx->lbr_stack_state = LBR_NONE;
270 		return;
271 	}
272 
273 	mask = x86_pmu.lbr_nr - 1;
274 	tos = intel_pmu_lbr_tos();
275 	for (i = 0; i < tos; i++) {
276 		lbr_idx = (tos - i) & mask;
277 		rdmsrl(x86_pmu.lbr_from + lbr_idx, task_ctx->lbr_from[i]);
278 		rdmsrl(x86_pmu.lbr_to + lbr_idx, task_ctx->lbr_to[i]);
279 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
280 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, task_ctx->lbr_info[i]);
281 	}
282 	task_ctx->tos = tos;
283 	task_ctx->lbr_stack_state = LBR_VALID;
284 }
285 
286 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
287 {
288 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
289 	struct x86_perf_task_context *task_ctx;
290 
291 	/*
292 	 * If LBR callstack feature is enabled and the stack was saved when
293 	 * the task was scheduled out, restore the stack. Otherwise flush
294 	 * the LBR stack.
295 	 */
296 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
297 	if (task_ctx) {
298 		if (sched_in) {
299 			__intel_pmu_lbr_restore(task_ctx);
300 			cpuc->lbr_context = ctx;
301 		} else {
302 			__intel_pmu_lbr_save(task_ctx);
303 		}
304 		return;
305 	}
306 
307 	/*
308 	 * When sampling the branck stack in system-wide, it may be
309 	 * necessary to flush the stack on context switch. This happens
310 	 * when the branch stack does not tag its entries with the pid
311 	 * of the current task. Otherwise it becomes impossible to
312 	 * associate a branch entry with a task. This ambiguity is more
313 	 * likely to appear when the branch stack supports priv level
314 	 * filtering and the user sets it to monitor only at the user
315 	 * level (which could be a useful measurement in system-wide
316 	 * mode). In that case, the risk is high of having a branch
317 	 * stack with branch from multiple tasks.
318  	 */
319 	if (sched_in) {
320 		intel_pmu_lbr_reset();
321 		cpuc->lbr_context = ctx;
322 	}
323 }
324 
325 static inline bool branch_user_callstack(unsigned br_sel)
326 {
327 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
328 }
329 
330 void intel_pmu_lbr_enable(struct perf_event *event)
331 {
332 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
333 	struct x86_perf_task_context *task_ctx;
334 
335 	if (!x86_pmu.lbr_nr)
336 		return;
337 
338 	/*
339 	 * Reset the LBR stack if we changed task context to
340 	 * avoid data leaks.
341 	 */
342 	if (event->ctx->task && cpuc->lbr_context != event->ctx) {
343 		intel_pmu_lbr_reset();
344 		cpuc->lbr_context = event->ctx;
345 	}
346 	cpuc->br_sel = event->hw.branch_reg.reg;
347 
348 	if (branch_user_callstack(cpuc->br_sel) && event->ctx &&
349 					event->ctx->task_ctx_data) {
350 		task_ctx = event->ctx->task_ctx_data;
351 		task_ctx->lbr_callstack_users++;
352 	}
353 
354 	cpuc->lbr_users++;
355 	perf_sched_cb_inc(event->ctx->pmu);
356 }
357 
358 void intel_pmu_lbr_disable(struct perf_event *event)
359 {
360 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
361 	struct x86_perf_task_context *task_ctx;
362 
363 	if (!x86_pmu.lbr_nr)
364 		return;
365 
366 	if (branch_user_callstack(cpuc->br_sel) && event->ctx &&
367 					event->ctx->task_ctx_data) {
368 		task_ctx = event->ctx->task_ctx_data;
369 		task_ctx->lbr_callstack_users--;
370 	}
371 
372 	cpuc->lbr_users--;
373 	WARN_ON_ONCE(cpuc->lbr_users < 0);
374 	perf_sched_cb_dec(event->ctx->pmu);
375 
376 	if (cpuc->enabled && !cpuc->lbr_users) {
377 		__intel_pmu_lbr_disable();
378 		/* avoid stale pointer */
379 		cpuc->lbr_context = NULL;
380 	}
381 }
382 
383 void intel_pmu_lbr_enable_all(bool pmi)
384 {
385 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
386 
387 	if (cpuc->lbr_users)
388 		__intel_pmu_lbr_enable(pmi);
389 }
390 
391 void intel_pmu_lbr_disable_all(void)
392 {
393 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
394 
395 	if (cpuc->lbr_users)
396 		__intel_pmu_lbr_disable();
397 }
398 
399 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
400 {
401 	unsigned long mask = x86_pmu.lbr_nr - 1;
402 	u64 tos = intel_pmu_lbr_tos();
403 	int i;
404 
405 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
406 		unsigned long lbr_idx = (tos - i) & mask;
407 		union {
408 			struct {
409 				u32 from;
410 				u32 to;
411 			};
412 			u64     lbr;
413 		} msr_lastbranch;
414 
415 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
416 
417 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
418 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
419 		cpuc->lbr_entries[i].mispred	= 0;
420 		cpuc->lbr_entries[i].predicted	= 0;
421 		cpuc->lbr_entries[i].reserved	= 0;
422 	}
423 	cpuc->lbr_stack.nr = i;
424 }
425 
426 /*
427  * Due to lack of segmentation in Linux the effective address (offset)
428  * is the same as the linear address, allowing us to merge the LIP and EIP
429  * LBR formats.
430  */
431 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
432 {
433 	bool need_info = false;
434 	unsigned long mask = x86_pmu.lbr_nr - 1;
435 	int lbr_format = x86_pmu.intel_cap.lbr_format;
436 	u64 tos = intel_pmu_lbr_tos();
437 	int i;
438 	int out = 0;
439 	int num = x86_pmu.lbr_nr;
440 
441 	if (cpuc->lbr_sel) {
442 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
443 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
444 			num = tos;
445 	}
446 
447 	for (i = 0; i < num; i++) {
448 		unsigned long lbr_idx = (tos - i) & mask;
449 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
450 		int skip = 0;
451 		u16 cycles = 0;
452 		int lbr_flags = lbr_desc[lbr_format];
453 
454 		rdmsrl(x86_pmu.lbr_from + lbr_idx, from);
455 		rdmsrl(x86_pmu.lbr_to   + lbr_idx, to);
456 
457 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
458 			u64 info;
459 
460 			rdmsrl(MSR_LBR_INFO_0 + lbr_idx, info);
461 			mis = !!(info & LBR_INFO_MISPRED);
462 			pred = !mis;
463 			in_tx = !!(info & LBR_INFO_IN_TX);
464 			abort = !!(info & LBR_INFO_ABORT);
465 			cycles = (info & LBR_INFO_CYCLES);
466 		}
467 		if (lbr_flags & LBR_EIP_FLAGS) {
468 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
469 			pred = !mis;
470 			skip = 1;
471 		}
472 		if (lbr_flags & LBR_TSX) {
473 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
474 			abort = !!(from & LBR_FROM_FLAG_ABORT);
475 			skip = 3;
476 		}
477 		from = (u64)((((s64)from) << skip) >> skip);
478 
479 		/*
480 		 * Some CPUs report duplicated abort records,
481 		 * with the second entry not having an abort bit set.
482 		 * Skip them here. This loop runs backwards,
483 		 * so we need to undo the previous record.
484 		 * If the abort just happened outside the window
485 		 * the extra entry cannot be removed.
486 		 */
487 		if (abort && x86_pmu.lbr_double_abort && out > 0)
488 			out--;
489 
490 		cpuc->lbr_entries[out].from	 = from;
491 		cpuc->lbr_entries[out].to	 = to;
492 		cpuc->lbr_entries[out].mispred	 = mis;
493 		cpuc->lbr_entries[out].predicted = pred;
494 		cpuc->lbr_entries[out].in_tx	 = in_tx;
495 		cpuc->lbr_entries[out].abort	 = abort;
496 		cpuc->lbr_entries[out].cycles	 = cycles;
497 		cpuc->lbr_entries[out].reserved	 = 0;
498 		out++;
499 	}
500 	cpuc->lbr_stack.nr = out;
501 }
502 
503 void intel_pmu_lbr_read(void)
504 {
505 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
506 
507 	if (!cpuc->lbr_users)
508 		return;
509 
510 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
511 		intel_pmu_lbr_read_32(cpuc);
512 	else
513 		intel_pmu_lbr_read_64(cpuc);
514 
515 	intel_pmu_lbr_filter(cpuc);
516 }
517 
518 /*
519  * SW filter is used:
520  * - in case there is no HW filter
521  * - in case the HW filter has errata or limitations
522  */
523 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
524 {
525 	u64 br_type = event->attr.branch_sample_type;
526 	int mask = 0;
527 
528 	if (br_type & PERF_SAMPLE_BRANCH_USER)
529 		mask |= X86_BR_USER;
530 
531 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
532 		mask |= X86_BR_KERNEL;
533 
534 	/* we ignore BRANCH_HV here */
535 
536 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
537 		mask |= X86_BR_ANY;
538 
539 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
540 		mask |= X86_BR_ANY_CALL;
541 
542 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
543 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
544 
545 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
546 		mask |= X86_BR_IND_CALL;
547 
548 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
549 		mask |= X86_BR_ABORT;
550 
551 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
552 		mask |= X86_BR_IN_TX;
553 
554 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
555 		mask |= X86_BR_NO_TX;
556 
557 	if (br_type & PERF_SAMPLE_BRANCH_COND)
558 		mask |= X86_BR_JCC;
559 
560 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
561 		if (!x86_pmu_has_lbr_callstack())
562 			return -EOPNOTSUPP;
563 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
564 			return -EINVAL;
565 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
566 			X86_BR_CALL_STACK;
567 	}
568 
569 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
570 		mask |= X86_BR_IND_JMP;
571 
572 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
573 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
574 	/*
575 	 * stash actual user request into reg, it may
576 	 * be used by fixup code for some CPU
577 	 */
578 	event->hw.branch_reg.reg = mask;
579 	return 0;
580 }
581 
582 /*
583  * setup the HW LBR filter
584  * Used only when available, may not be enough to disambiguate
585  * all branches, may need the help of the SW filter
586  */
587 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
588 {
589 	struct hw_perf_event_extra *reg;
590 	u64 br_type = event->attr.branch_sample_type;
591 	u64 mask = 0, v;
592 	int i;
593 
594 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
595 		if (!(br_type & (1ULL << i)))
596 			continue;
597 
598 		v = x86_pmu.lbr_sel_map[i];
599 		if (v == LBR_NOT_SUPP)
600 			return -EOPNOTSUPP;
601 
602 		if (v != LBR_IGN)
603 			mask |= v;
604 	}
605 
606 	reg = &event->hw.branch_reg;
607 	reg->idx = EXTRA_REG_LBR;
608 
609 	/*
610 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
611 	 * in suppress mode. So LBR_SELECT should be set to
612 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
613 	 * But the 10th bit LBR_CALL_STACK does not operate
614 	 * in suppress mode.
615 	 */
616 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
617 
618 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
619 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
620 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
621 		reg->config |= LBR_NO_INFO;
622 
623 	return 0;
624 }
625 
626 int intel_pmu_setup_lbr_filter(struct perf_event *event)
627 {
628 	int ret = 0;
629 
630 	/*
631 	 * no LBR on this PMU
632 	 */
633 	if (!x86_pmu.lbr_nr)
634 		return -EOPNOTSUPP;
635 
636 	/*
637 	 * setup SW LBR filter
638 	 */
639 	ret = intel_pmu_setup_sw_lbr_filter(event);
640 	if (ret)
641 		return ret;
642 
643 	/*
644 	 * setup HW LBR filter, if any
645 	 */
646 	if (x86_pmu.lbr_sel_map)
647 		ret = intel_pmu_setup_hw_lbr_filter(event);
648 
649 	return ret;
650 }
651 
652 /*
653  * return the type of control flow change at address "from"
654  * instruction is not necessarily a branch (in case of interrupt).
655  *
656  * The branch type returned also includes the priv level of the
657  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
658  *
659  * If a branch type is unknown OR the instruction cannot be
660  * decoded (e.g., text page not present), then X86_BR_NONE is
661  * returned.
662  */
663 static int branch_type(unsigned long from, unsigned long to, int abort)
664 {
665 	struct insn insn;
666 	void *addr;
667 	int bytes_read, bytes_left;
668 	int ret = X86_BR_NONE;
669 	int ext, to_plm, from_plm;
670 	u8 buf[MAX_INSN_SIZE];
671 	int is64 = 0;
672 
673 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
674 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
675 
676 	/*
677 	 * maybe zero if lbr did not fill up after a reset by the time
678 	 * we get a PMU interrupt
679 	 */
680 	if (from == 0 || to == 0)
681 		return X86_BR_NONE;
682 
683 	if (abort)
684 		return X86_BR_ABORT | to_plm;
685 
686 	if (from_plm == X86_BR_USER) {
687 		/*
688 		 * can happen if measuring at the user level only
689 		 * and we interrupt in a kernel thread, e.g., idle.
690 		 */
691 		if (!current->mm)
692 			return X86_BR_NONE;
693 
694 		/* may fail if text not present */
695 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
696 						MAX_INSN_SIZE);
697 		bytes_read = MAX_INSN_SIZE - bytes_left;
698 		if (!bytes_read)
699 			return X86_BR_NONE;
700 
701 		addr = buf;
702 	} else {
703 		/*
704 		 * The LBR logs any address in the IP, even if the IP just
705 		 * faulted. This means userspace can control the from address.
706 		 * Ensure we don't blindy read any address by validating it is
707 		 * a known text address.
708 		 */
709 		if (kernel_text_address(from)) {
710 			addr = (void *)from;
711 			/*
712 			 * Assume we can get the maximum possible size
713 			 * when grabbing kernel data.  This is not
714 			 * _strictly_ true since we could possibly be
715 			 * executing up next to a memory hole, but
716 			 * it is very unlikely to be a problem.
717 			 */
718 			bytes_read = MAX_INSN_SIZE;
719 		} else {
720 			return X86_BR_NONE;
721 		}
722 	}
723 
724 	/*
725 	 * decoder needs to know the ABI especially
726 	 * on 64-bit systems running 32-bit apps
727 	 */
728 #ifdef CONFIG_X86_64
729 	is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
730 #endif
731 	insn_init(&insn, addr, bytes_read, is64);
732 	insn_get_opcode(&insn);
733 	if (!insn.opcode.got)
734 		return X86_BR_ABORT;
735 
736 	switch (insn.opcode.bytes[0]) {
737 	case 0xf:
738 		switch (insn.opcode.bytes[1]) {
739 		case 0x05: /* syscall */
740 		case 0x34: /* sysenter */
741 			ret = X86_BR_SYSCALL;
742 			break;
743 		case 0x07: /* sysret */
744 		case 0x35: /* sysexit */
745 			ret = X86_BR_SYSRET;
746 			break;
747 		case 0x80 ... 0x8f: /* conditional */
748 			ret = X86_BR_JCC;
749 			break;
750 		default:
751 			ret = X86_BR_NONE;
752 		}
753 		break;
754 	case 0x70 ... 0x7f: /* conditional */
755 		ret = X86_BR_JCC;
756 		break;
757 	case 0xc2: /* near ret */
758 	case 0xc3: /* near ret */
759 	case 0xca: /* far ret */
760 	case 0xcb: /* far ret */
761 		ret = X86_BR_RET;
762 		break;
763 	case 0xcf: /* iret */
764 		ret = X86_BR_IRET;
765 		break;
766 	case 0xcc ... 0xce: /* int */
767 		ret = X86_BR_INT;
768 		break;
769 	case 0xe8: /* call near rel */
770 		insn_get_immediate(&insn);
771 		if (insn.immediate1.value == 0) {
772 			/* zero length call */
773 			ret = X86_BR_ZERO_CALL;
774 			break;
775 		}
776 	case 0x9a: /* call far absolute */
777 		ret = X86_BR_CALL;
778 		break;
779 	case 0xe0 ... 0xe3: /* loop jmp */
780 		ret = X86_BR_JCC;
781 		break;
782 	case 0xe9 ... 0xeb: /* jmp */
783 		ret = X86_BR_JMP;
784 		break;
785 	case 0xff: /* call near absolute, call far absolute ind */
786 		insn_get_modrm(&insn);
787 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
788 		switch (ext) {
789 		case 2: /* near ind call */
790 		case 3: /* far ind call */
791 			ret = X86_BR_IND_CALL;
792 			break;
793 		case 4:
794 		case 5:
795 			ret = X86_BR_IND_JMP;
796 			break;
797 		}
798 		break;
799 	default:
800 		ret = X86_BR_NONE;
801 	}
802 	/*
803 	 * interrupts, traps, faults (and thus ring transition) may
804 	 * occur on any instructions. Thus, to classify them correctly,
805 	 * we need to first look at the from and to priv levels. If they
806 	 * are different and to is in the kernel, then it indicates
807 	 * a ring transition. If the from instruction is not a ring
808 	 * transition instr (syscall, systenter, int), then it means
809 	 * it was a irq, trap or fault.
810 	 *
811 	 * we have no way of detecting kernel to kernel faults.
812 	 */
813 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
814 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
815 		ret = X86_BR_IRQ;
816 
817 	/*
818 	 * branch priv level determined by target as
819 	 * is done by HW when LBR_SELECT is implemented
820 	 */
821 	if (ret != X86_BR_NONE)
822 		ret |= to_plm;
823 
824 	return ret;
825 }
826 
827 /*
828  * implement actual branch filter based on user demand.
829  * Hardware may not exactly satisfy that request, thus
830  * we need to inspect opcodes. Mismatched branches are
831  * discarded. Therefore, the number of branches returned
832  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
833  */
834 static void
835 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
836 {
837 	u64 from, to;
838 	int br_sel = cpuc->br_sel;
839 	int i, j, type;
840 	bool compress = false;
841 
842 	/* if sampling all branches, then nothing to filter */
843 	if ((br_sel & X86_BR_ALL) == X86_BR_ALL)
844 		return;
845 
846 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
847 
848 		from = cpuc->lbr_entries[i].from;
849 		to = cpuc->lbr_entries[i].to;
850 
851 		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
852 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
853 			if (cpuc->lbr_entries[i].in_tx)
854 				type |= X86_BR_IN_TX;
855 			else
856 				type |= X86_BR_NO_TX;
857 		}
858 
859 		/* if type does not correspond, then discard */
860 		if (type == X86_BR_NONE || (br_sel & type) != type) {
861 			cpuc->lbr_entries[i].from = 0;
862 			compress = true;
863 		}
864 	}
865 
866 	if (!compress)
867 		return;
868 
869 	/* remove all entries with from=0 */
870 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
871 		if (!cpuc->lbr_entries[i].from) {
872 			j = i;
873 			while (++j < cpuc->lbr_stack.nr)
874 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
875 			cpuc->lbr_stack.nr--;
876 			if (!cpuc->lbr_entries[i].from)
877 				continue;
878 		}
879 		i++;
880 	}
881 }
882 
883 /*
884  * Map interface branch filters onto LBR filters
885  */
886 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
887 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
888 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
889 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
890 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
891 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
892 						| LBR_IND_JMP | LBR_FAR,
893 	/*
894 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
895 	 */
896 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
897 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
898 	/*
899 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
900 	 */
901 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
902 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
903 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
904 };
905 
906 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
907 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
908 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
909 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
910 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
911 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
912 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
913 						| LBR_FAR,
914 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
915 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
916 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
917 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
918 };
919 
920 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
921 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
922 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
923 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
924 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
925 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
926 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
927 						| LBR_FAR,
928 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
929 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
930 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
931 						| LBR_RETURN | LBR_CALL_STACK,
932 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
933 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
934 };
935 
936 /* core */
937 void __init intel_pmu_lbr_init_core(void)
938 {
939 	x86_pmu.lbr_nr     = 4;
940 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
941 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
942 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
943 
944 	/*
945 	 * SW branch filter usage:
946 	 * - compensate for lack of HW filter
947 	 */
948 	pr_cont("4-deep LBR, ");
949 }
950 
951 /* nehalem/westmere */
952 void __init intel_pmu_lbr_init_nhm(void)
953 {
954 	x86_pmu.lbr_nr     = 16;
955 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
956 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
957 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
958 
959 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
960 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
961 
962 	/*
963 	 * SW branch filter usage:
964 	 * - workaround LBR_SEL errata (see above)
965 	 * - support syscall, sysret capture.
966 	 *   That requires LBR_FAR but that means far
967 	 *   jmp need to be filtered out
968 	 */
969 	pr_cont("16-deep LBR, ");
970 }
971 
972 /* sandy bridge */
973 void __init intel_pmu_lbr_init_snb(void)
974 {
975 	x86_pmu.lbr_nr	 = 16;
976 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
977 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
978 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
979 
980 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
981 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
982 
983 	/*
984 	 * SW branch filter usage:
985 	 * - support syscall, sysret capture.
986 	 *   That requires LBR_FAR but that means far
987 	 *   jmp need to be filtered out
988 	 */
989 	pr_cont("16-deep LBR, ");
990 }
991 
992 /* haswell */
993 void intel_pmu_lbr_init_hsw(void)
994 {
995 	x86_pmu.lbr_nr	 = 16;
996 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
997 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
998 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
999 
1000 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1001 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1002 
1003 	pr_cont("16-deep LBR, ");
1004 }
1005 
1006 /* skylake */
1007 __init void intel_pmu_lbr_init_skl(void)
1008 {
1009 	x86_pmu.lbr_nr	 = 32;
1010 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1011 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1012 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1013 
1014 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1015 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1016 
1017 	/*
1018 	 * SW branch filter usage:
1019 	 * - support syscall, sysret capture.
1020 	 *   That requires LBR_FAR but that means far
1021 	 *   jmp need to be filtered out
1022 	 */
1023 	pr_cont("32-deep LBR, ");
1024 }
1025 
1026 /* atom */
1027 void __init intel_pmu_lbr_init_atom(void)
1028 {
1029 	/*
1030 	 * only models starting at stepping 10 seems
1031 	 * to have an operational LBR which can freeze
1032 	 * on PMU interrupt
1033 	 */
1034 	if (boot_cpu_data.x86_model == 28
1035 	    && boot_cpu_data.x86_mask < 10) {
1036 		pr_cont("LBR disabled due to erratum");
1037 		return;
1038 	}
1039 
1040 	x86_pmu.lbr_nr	   = 8;
1041 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1042 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1043 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1044 
1045 	/*
1046 	 * SW branch filter usage:
1047 	 * - compensate for lack of HW filter
1048 	 */
1049 	pr_cont("8-deep LBR, ");
1050 }
1051 
1052 /* Knights Landing */
1053 void intel_pmu_lbr_init_knl(void)
1054 {
1055 	x86_pmu.lbr_nr	   = 8;
1056 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1057 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1058 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1059 
1060 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1061 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1062 
1063 	pr_cont("8-deep LBR, ");
1064 }
1065