• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_MAX_KNOWN    = LBR_FORMAT_EIP_FLAGS2,
17 };
18 
19 static enum {
20 	LBR_EIP_FLAGS		= 1,
21 	LBR_TSX			= 2,
22 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
23 	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
24 	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
25 };
26 
27 /*
28  * Intel LBR_SELECT bits
29  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
30  *
31  * Hardware branch filter (not available on all CPUs)
32  */
33 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
34 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
35 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
36 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
37 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
38 #define LBR_RETURN_BIT		5 /* do not capture near returns */
39 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
40 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
41 #define LBR_FAR_BIT		8 /* do not capture far branches */
42 
43 #define LBR_KERNEL	(1 << LBR_KERNEL_BIT)
44 #define LBR_USER	(1 << LBR_USER_BIT)
45 #define LBR_JCC		(1 << LBR_JCC_BIT)
46 #define LBR_REL_CALL	(1 << LBR_REL_CALL_BIT)
47 #define LBR_IND_CALL	(1 << LBR_IND_CALL_BIT)
48 #define LBR_RETURN	(1 << LBR_RETURN_BIT)
49 #define LBR_REL_JMP	(1 << LBR_REL_JMP_BIT)
50 #define LBR_IND_JMP	(1 << LBR_IND_JMP_BIT)
51 #define LBR_FAR		(1 << LBR_FAR_BIT)
52 
53 #define LBR_PLM (LBR_KERNEL | LBR_USER)
54 
55 #define LBR_SEL_MASK	0x1ff	/* valid bits in LBR_SELECT */
56 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
57 #define LBR_IGN		0	/* ignored */
58 
59 #define LBR_ANY		 \
60 	(LBR_JCC	|\
61 	 LBR_REL_CALL	|\
62 	 LBR_IND_CALL	|\
63 	 LBR_RETURN	|\
64 	 LBR_REL_JMP	|\
65 	 LBR_IND_JMP	|\
66 	 LBR_FAR)
67 
68 #define LBR_FROM_FLAG_MISPRED  (1ULL << 63)
69 #define LBR_FROM_FLAG_IN_TX    (1ULL << 62)
70 #define LBR_FROM_FLAG_ABORT    (1ULL << 61)
71 
72 #define for_each_branch_sample_type(x) \
73 	for ((x) = PERF_SAMPLE_BRANCH_USER; \
74 	     (x) < PERF_SAMPLE_BRANCH_MAX; (x) <<= 1)
75 
76 /*
77  * x86control flow change classification
78  * x86control flow changes include branches, interrupts, traps, faults
79  */
80 enum {
81 	X86_BR_NONE     = 0,      /* unknown */
82 
83 	X86_BR_USER     = 1 << 0, /* branch target is user */
84 	X86_BR_KERNEL   = 1 << 1, /* branch target is kernel */
85 
86 	X86_BR_CALL     = 1 << 2, /* call */
87 	X86_BR_RET      = 1 << 3, /* return */
88 	X86_BR_SYSCALL  = 1 << 4, /* syscall */
89 	X86_BR_SYSRET   = 1 << 5, /* syscall return */
90 	X86_BR_INT      = 1 << 6, /* sw interrupt */
91 	X86_BR_IRET     = 1 << 7, /* return from interrupt */
92 	X86_BR_JCC      = 1 << 8, /* conditional */
93 	X86_BR_JMP      = 1 << 9, /* jump */
94 	X86_BR_IRQ      = 1 << 10,/* hw interrupt or trap or fault */
95 	X86_BR_IND_CALL = 1 << 11,/* indirect calls */
96 	X86_BR_ABORT    = 1 << 12,/* transaction abort */
97 	X86_BR_IN_TX    = 1 << 13,/* in transaction */
98 	X86_BR_NO_TX    = 1 << 14,/* not in transaction */
99 };
100 
101 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
102 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
103 
104 #define X86_BR_ANY       \
105 	(X86_BR_CALL    |\
106 	 X86_BR_RET     |\
107 	 X86_BR_SYSCALL |\
108 	 X86_BR_SYSRET  |\
109 	 X86_BR_INT     |\
110 	 X86_BR_IRET    |\
111 	 X86_BR_JCC     |\
112 	 X86_BR_JMP	 |\
113 	 X86_BR_IRQ	 |\
114 	 X86_BR_ABORT	 |\
115 	 X86_BR_IND_CALL)
116 
117 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
118 
119 #define X86_BR_ANY_CALL		 \
120 	(X86_BR_CALL		|\
121 	 X86_BR_IND_CALL	|\
122 	 X86_BR_SYSCALL		|\
123 	 X86_BR_IRQ		|\
124 	 X86_BR_INT)
125 
126 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
127 
128 /*
129  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
130  * otherwise it becomes near impossible to get a reliable stack.
131  */
132 
__intel_pmu_lbr_enable(void)133 static void __intel_pmu_lbr_enable(void)
134 {
135 	u64 debugctl;
136 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
137 
138 	if (cpuc->lbr_sel)
139 		wrmsrl(MSR_LBR_SELECT, cpuc->lbr_sel->config);
140 
141 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
142 	debugctl |= (DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
143 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
144 }
145 
__intel_pmu_lbr_disable(void)146 static void __intel_pmu_lbr_disable(void)
147 {
148 	u64 debugctl;
149 
150 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
151 	debugctl &= ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_FREEZE_LBRS_ON_PMI);
152 	wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
153 }
154 
intel_pmu_lbr_reset_32(void)155 static void intel_pmu_lbr_reset_32(void)
156 {
157 	int i;
158 
159 	for (i = 0; i < x86_pmu.lbr_nr; i++)
160 		wrmsrl(x86_pmu.lbr_from + i, 0);
161 }
162 
intel_pmu_lbr_reset_64(void)163 static void intel_pmu_lbr_reset_64(void)
164 {
165 	int i;
166 
167 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
168 		wrmsrl(x86_pmu.lbr_from + i, 0);
169 		wrmsrl(x86_pmu.lbr_to   + i, 0);
170 	}
171 }
172 
intel_pmu_lbr_reset(void)173 void intel_pmu_lbr_reset(void)
174 {
175 	if (!x86_pmu.lbr_nr)
176 		return;
177 
178 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
179 		intel_pmu_lbr_reset_32();
180 	else
181 		intel_pmu_lbr_reset_64();
182 }
183 
intel_pmu_lbr_enable(struct perf_event * event)184 void intel_pmu_lbr_enable(struct perf_event *event)
185 {
186 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
187 
188 	if (!x86_pmu.lbr_nr)
189 		return;
190 
191 	/*
192 	 * Reset the LBR stack if we changed task context to
193 	 * avoid data leaks.
194 	 */
195 	if (event->ctx->task && cpuc->lbr_context != event->ctx) {
196 		intel_pmu_lbr_reset();
197 		cpuc->lbr_context = event->ctx;
198 	}
199 	cpuc->br_sel = event->hw.branch_reg.reg;
200 
201 	cpuc->lbr_users++;
202 }
203 
intel_pmu_lbr_disable(struct perf_event * event)204 void intel_pmu_lbr_disable(struct perf_event *event)
205 {
206 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
207 
208 	if (!x86_pmu.lbr_nr)
209 		return;
210 
211 	cpuc->lbr_users--;
212 	WARN_ON_ONCE(cpuc->lbr_users < 0);
213 
214 	if (cpuc->enabled && !cpuc->lbr_users) {
215 		__intel_pmu_lbr_disable();
216 		/* avoid stale pointer */
217 		cpuc->lbr_context = NULL;
218 	}
219 }
220 
intel_pmu_lbr_enable_all(void)221 void intel_pmu_lbr_enable_all(void)
222 {
223 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
224 
225 	if (cpuc->lbr_users)
226 		__intel_pmu_lbr_enable();
227 }
228 
intel_pmu_lbr_disable_all(void)229 void intel_pmu_lbr_disable_all(void)
230 {
231 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
232 
233 	if (cpuc->lbr_users)
234 		__intel_pmu_lbr_disable();
235 }
236 
237 /*
238  * TOS = most recently recorded branch
239  */
intel_pmu_lbr_tos(void)240 static inline u64 intel_pmu_lbr_tos(void)
241 {
242 	u64 tos;
243 
244 	rdmsrl(x86_pmu.lbr_tos, tos);
245 
246 	return tos;
247 }
248 
intel_pmu_lbr_read_32(struct cpu_hw_events * cpuc)249 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
250 {
251 	unsigned long mask = x86_pmu.lbr_nr - 1;
252 	u64 tos = intel_pmu_lbr_tos();
253 	int i;
254 
255 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
256 		unsigned long lbr_idx = (tos - i) & mask;
257 		union {
258 			struct {
259 				u32 from;
260 				u32 to;
261 			};
262 			u64     lbr;
263 		} msr_lastbranch;
264 
265 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
266 
267 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
268 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
269 		cpuc->lbr_entries[i].mispred	= 0;
270 		cpuc->lbr_entries[i].predicted	= 0;
271 		cpuc->lbr_entries[i].in_tx	= 0;
272 		cpuc->lbr_entries[i].abort	= 0;
273 		cpuc->lbr_entries[i].reserved	= 0;
274 	}
275 	cpuc->lbr_stack.nr = i;
276 }
277 
278 /*
279  * Due to lack of segmentation in Linux the effective address (offset)
280  * is the same as the linear address, allowing us to merge the LIP and EIP
281  * LBR formats.
282  */
intel_pmu_lbr_read_64(struct cpu_hw_events * cpuc)283 static void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
284 {
285 	unsigned long mask = x86_pmu.lbr_nr - 1;
286 	int lbr_format = x86_pmu.intel_cap.lbr_format;
287 	u64 tos = intel_pmu_lbr_tos();
288 	int i;
289 	int out = 0;
290 
291 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
292 		unsigned long lbr_idx = (tos - i) & mask;
293 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
294 		int skip = 0;
295 		int lbr_flags = lbr_desc[lbr_format];
296 
297 		rdmsrl(x86_pmu.lbr_from + lbr_idx, from);
298 		rdmsrl(x86_pmu.lbr_to   + lbr_idx, to);
299 
300 		if (lbr_flags & LBR_EIP_FLAGS) {
301 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
302 			pred = !mis;
303 			skip = 1;
304 		}
305 		if (lbr_flags & LBR_TSX) {
306 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
307 			abort = !!(from & LBR_FROM_FLAG_ABORT);
308 			skip = 3;
309 		}
310 		from = (u64)((((s64)from) << skip) >> skip);
311 
312 		/*
313 		 * Some CPUs report duplicated abort records,
314 		 * with the second entry not having an abort bit set.
315 		 * Skip them here. This loop runs backwards,
316 		 * so we need to undo the previous record.
317 		 * If the abort just happened outside the window
318 		 * the extra entry cannot be removed.
319 		 */
320 		if (abort && x86_pmu.lbr_double_abort && out > 0)
321 			out--;
322 
323 		cpuc->lbr_entries[out].from	 = from;
324 		cpuc->lbr_entries[out].to	 = to;
325 		cpuc->lbr_entries[out].mispred	 = mis;
326 		cpuc->lbr_entries[out].predicted = pred;
327 		cpuc->lbr_entries[out].in_tx	 = in_tx;
328 		cpuc->lbr_entries[out].abort	 = abort;
329 		cpuc->lbr_entries[out].reserved	 = 0;
330 		out++;
331 	}
332 	cpuc->lbr_stack.nr = out;
333 }
334 
intel_pmu_lbr_read(void)335 void intel_pmu_lbr_read(void)
336 {
337 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
338 
339 	if (!cpuc->lbr_users)
340 		return;
341 
342 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32)
343 		intel_pmu_lbr_read_32(cpuc);
344 	else
345 		intel_pmu_lbr_read_64(cpuc);
346 
347 	intel_pmu_lbr_filter(cpuc);
348 }
349 
350 /*
351  * SW filter is used:
352  * - in case there is no HW filter
353  * - in case the HW filter has errata or limitations
354  */
intel_pmu_setup_sw_lbr_filter(struct perf_event * event)355 static void intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
356 {
357 	u64 br_type = event->attr.branch_sample_type;
358 	int mask = 0;
359 
360 	if (br_type & PERF_SAMPLE_BRANCH_USER)
361 		mask |= X86_BR_USER;
362 
363 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
364 		mask |= X86_BR_KERNEL;
365 
366 	/* we ignore BRANCH_HV here */
367 
368 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
369 		mask |= X86_BR_ANY;
370 
371 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
372 		mask |= X86_BR_ANY_CALL;
373 
374 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
375 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
376 
377 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
378 		mask |= X86_BR_IND_CALL;
379 
380 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
381 		mask |= X86_BR_ABORT;
382 
383 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
384 		mask |= X86_BR_IN_TX;
385 
386 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
387 		mask |= X86_BR_NO_TX;
388 
389 	if (br_type & PERF_SAMPLE_BRANCH_COND)
390 		mask |= X86_BR_JCC;
391 
392 	/*
393 	 * stash actual user request into reg, it may
394 	 * be used by fixup code for some CPU
395 	 */
396 	event->hw.branch_reg.reg = mask;
397 }
398 
399 /*
400  * setup the HW LBR filter
401  * Used only when available, may not be enough to disambiguate
402  * all branches, may need the help of the SW filter
403  */
intel_pmu_setup_hw_lbr_filter(struct perf_event * event)404 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
405 {
406 	struct hw_perf_event_extra *reg;
407 	u64 br_type = event->attr.branch_sample_type;
408 	u64 mask = 0, m;
409 	u64 v;
410 
411 	for_each_branch_sample_type(m) {
412 		if (!(br_type & m))
413 			continue;
414 
415 		v = x86_pmu.lbr_sel_map[m];
416 		if (v == LBR_NOT_SUPP)
417 			return -EOPNOTSUPP;
418 
419 		if (v != LBR_IGN)
420 			mask |= v;
421 	}
422 	reg = &event->hw.branch_reg;
423 	reg->idx = EXTRA_REG_LBR;
424 
425 	/* LBR_SELECT operates in suppress mode so invert mask */
426 	reg->config = ~mask & x86_pmu.lbr_sel_mask;
427 
428 	return 0;
429 }
430 
intel_pmu_setup_lbr_filter(struct perf_event * event)431 int intel_pmu_setup_lbr_filter(struct perf_event *event)
432 {
433 	int ret = 0;
434 
435 	/*
436 	 * no LBR on this PMU
437 	 */
438 	if (!x86_pmu.lbr_nr)
439 		return -EOPNOTSUPP;
440 
441 	/*
442 	 * setup SW LBR filter
443 	 */
444 	intel_pmu_setup_sw_lbr_filter(event);
445 
446 	/*
447 	 * setup HW LBR filter, if any
448 	 */
449 	if (x86_pmu.lbr_sel_map)
450 		ret = intel_pmu_setup_hw_lbr_filter(event);
451 
452 	return ret;
453 }
454 
455 /*
456  * return the type of control flow change at address "from"
457  * intruction is not necessarily a branch (in case of interrupt).
458  *
459  * The branch type returned also includes the priv level of the
460  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
461  *
462  * If a branch type is unknown OR the instruction cannot be
463  * decoded (e.g., text page not present), then X86_BR_NONE is
464  * returned.
465  */
branch_type(unsigned long from,unsigned long to,int abort)466 static int branch_type(unsigned long from, unsigned long to, int abort)
467 {
468 	struct insn insn;
469 	void *addr;
470 	int bytes, size = MAX_INSN_SIZE;
471 	int ret = X86_BR_NONE;
472 	int ext, to_plm, from_plm;
473 	u8 buf[MAX_INSN_SIZE];
474 	int is64 = 0;
475 
476 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
477 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
478 
479 	/*
480 	 * maybe zero if lbr did not fill up after a reset by the time
481 	 * we get a PMU interrupt
482 	 */
483 	if (from == 0 || to == 0)
484 		return X86_BR_NONE;
485 
486 	if (abort)
487 		return X86_BR_ABORT | to_plm;
488 
489 	if (from_plm == X86_BR_USER) {
490 		/*
491 		 * can happen if measuring at the user level only
492 		 * and we interrupt in a kernel thread, e.g., idle.
493 		 */
494 		if (!current->mm)
495 			return X86_BR_NONE;
496 
497 		/* may fail if text not present */
498 		bytes = copy_from_user_nmi(buf, (void __user *)from, size);
499 		if (bytes != 0)
500 			return X86_BR_NONE;
501 
502 		addr = buf;
503 	} else {
504 		/*
505 		 * The LBR logs any address in the IP, even if the IP just
506 		 * faulted. This means userspace can control the from address.
507 		 * Ensure we don't blindy read any address by validating it is
508 		 * a known text address.
509 		 */
510 		if (kernel_text_address(from))
511 			addr = (void *)from;
512 		else
513 			return X86_BR_NONE;
514 	}
515 
516 	/*
517 	 * decoder needs to know the ABI especially
518 	 * on 64-bit systems running 32-bit apps
519 	 */
520 #ifdef CONFIG_X86_64
521 	is64 = kernel_ip((unsigned long)addr) || !test_thread_flag(TIF_IA32);
522 #endif
523 	insn_init(&insn, addr, is64);
524 	insn_get_opcode(&insn);
525 
526 	switch (insn.opcode.bytes[0]) {
527 	case 0xf:
528 		switch (insn.opcode.bytes[1]) {
529 		case 0x05: /* syscall */
530 		case 0x34: /* sysenter */
531 			ret = X86_BR_SYSCALL;
532 			break;
533 		case 0x07: /* sysret */
534 		case 0x35: /* sysexit */
535 			ret = X86_BR_SYSRET;
536 			break;
537 		case 0x80 ... 0x8f: /* conditional */
538 			ret = X86_BR_JCC;
539 			break;
540 		default:
541 			ret = X86_BR_NONE;
542 		}
543 		break;
544 	case 0x70 ... 0x7f: /* conditional */
545 		ret = X86_BR_JCC;
546 		break;
547 	case 0xc2: /* near ret */
548 	case 0xc3: /* near ret */
549 	case 0xca: /* far ret */
550 	case 0xcb: /* far ret */
551 		ret = X86_BR_RET;
552 		break;
553 	case 0xcf: /* iret */
554 		ret = X86_BR_IRET;
555 		break;
556 	case 0xcc ... 0xce: /* int */
557 		ret = X86_BR_INT;
558 		break;
559 	case 0xe8: /* call near rel */
560 	case 0x9a: /* call far absolute */
561 		ret = X86_BR_CALL;
562 		break;
563 	case 0xe0 ... 0xe3: /* loop jmp */
564 		ret = X86_BR_JCC;
565 		break;
566 	case 0xe9 ... 0xeb: /* jmp */
567 		ret = X86_BR_JMP;
568 		break;
569 	case 0xff: /* call near absolute, call far absolute ind */
570 		insn_get_modrm(&insn);
571 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
572 		switch (ext) {
573 		case 2: /* near ind call */
574 		case 3: /* far ind call */
575 			ret = X86_BR_IND_CALL;
576 			break;
577 		case 4:
578 		case 5:
579 			ret = X86_BR_JMP;
580 			break;
581 		}
582 		break;
583 	default:
584 		ret = X86_BR_NONE;
585 	}
586 	/*
587 	 * interrupts, traps, faults (and thus ring transition) may
588 	 * occur on any instructions. Thus, to classify them correctly,
589 	 * we need to first look at the from and to priv levels. If they
590 	 * are different and to is in the kernel, then it indicates
591 	 * a ring transition. If the from instruction is not a ring
592 	 * transition instr (syscall, systenter, int), then it means
593 	 * it was a irq, trap or fault.
594 	 *
595 	 * we have no way of detecting kernel to kernel faults.
596 	 */
597 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
598 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
599 		ret = X86_BR_IRQ;
600 
601 	/*
602 	 * branch priv level determined by target as
603 	 * is done by HW when LBR_SELECT is implemented
604 	 */
605 	if (ret != X86_BR_NONE)
606 		ret |= to_plm;
607 
608 	return ret;
609 }
610 
611 /*
612  * implement actual branch filter based on user demand.
613  * Hardware may not exactly satisfy that request, thus
614  * we need to inspect opcodes. Mismatched branches are
615  * discarded. Therefore, the number of branches returned
616  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
617  */
618 static void
intel_pmu_lbr_filter(struct cpu_hw_events * cpuc)619 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
620 {
621 	u64 from, to;
622 	int br_sel = cpuc->br_sel;
623 	int i, j, type;
624 	bool compress = false;
625 
626 	/* if sampling all branches, then nothing to filter */
627 	if ((br_sel & X86_BR_ALL) == X86_BR_ALL)
628 		return;
629 
630 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
631 
632 		from = cpuc->lbr_entries[i].from;
633 		to = cpuc->lbr_entries[i].to;
634 
635 		type = branch_type(from, to, cpuc->lbr_entries[i].abort);
636 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
637 			if (cpuc->lbr_entries[i].in_tx)
638 				type |= X86_BR_IN_TX;
639 			else
640 				type |= X86_BR_NO_TX;
641 		}
642 
643 		/* if type does not correspond, then discard */
644 		if (type == X86_BR_NONE || (br_sel & type) != type) {
645 			cpuc->lbr_entries[i].from = 0;
646 			compress = true;
647 		}
648 	}
649 
650 	if (!compress)
651 		return;
652 
653 	/* remove all entries with from=0 */
654 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
655 		if (!cpuc->lbr_entries[i].from) {
656 			j = i;
657 			while (++j < cpuc->lbr_stack.nr)
658 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
659 			cpuc->lbr_stack.nr--;
660 			if (!cpuc->lbr_entries[i].from)
661 				continue;
662 		}
663 		i++;
664 	}
665 }
666 
667 /*
668  * Map interface branch filters onto LBR filters
669  */
670 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX] = {
671 	[PERF_SAMPLE_BRANCH_ANY]	= LBR_ANY,
672 	[PERF_SAMPLE_BRANCH_USER]	= LBR_USER,
673 	[PERF_SAMPLE_BRANCH_KERNEL]	= LBR_KERNEL,
674 	[PERF_SAMPLE_BRANCH_HV]		= LBR_IGN,
675 	[PERF_SAMPLE_BRANCH_ANY_RETURN]	= LBR_RETURN | LBR_REL_JMP
676 					| LBR_IND_JMP | LBR_FAR,
677 	/*
678 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
679 	 */
680 	[PERF_SAMPLE_BRANCH_ANY_CALL] =
681 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
682 	/*
683 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
684 	 */
685 	[PERF_SAMPLE_BRANCH_IND_CALL] = LBR_IND_CALL | LBR_IND_JMP,
686 	[PERF_SAMPLE_BRANCH_COND]     = LBR_JCC,
687 };
688 
689 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX] = {
690 	[PERF_SAMPLE_BRANCH_ANY]	= LBR_ANY,
691 	[PERF_SAMPLE_BRANCH_USER]	= LBR_USER,
692 	[PERF_SAMPLE_BRANCH_KERNEL]	= LBR_KERNEL,
693 	[PERF_SAMPLE_BRANCH_HV]		= LBR_IGN,
694 	[PERF_SAMPLE_BRANCH_ANY_RETURN]	= LBR_RETURN | LBR_FAR,
695 	[PERF_SAMPLE_BRANCH_ANY_CALL]	= LBR_REL_CALL | LBR_IND_CALL
696 					| LBR_FAR,
697 	[PERF_SAMPLE_BRANCH_IND_CALL]	= LBR_IND_CALL,
698 	[PERF_SAMPLE_BRANCH_COND]       = LBR_JCC,
699 };
700 
701 /* core */
intel_pmu_lbr_init_core(void)702 void __init intel_pmu_lbr_init_core(void)
703 {
704 	x86_pmu.lbr_nr     = 4;
705 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
706 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
707 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
708 
709 	/*
710 	 * SW branch filter usage:
711 	 * - compensate for lack of HW filter
712 	 */
713 	pr_cont("4-deep LBR, ");
714 }
715 
716 /* nehalem/westmere */
intel_pmu_lbr_init_nhm(void)717 void __init intel_pmu_lbr_init_nhm(void)
718 {
719 	x86_pmu.lbr_nr     = 16;
720 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
721 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
722 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
723 
724 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
725 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
726 
727 	/*
728 	 * SW branch filter usage:
729 	 * - workaround LBR_SEL errata (see above)
730 	 * - support syscall, sysret capture.
731 	 *   That requires LBR_FAR but that means far
732 	 *   jmp need to be filtered out
733 	 */
734 	pr_cont("16-deep LBR, ");
735 }
736 
737 /* sandy bridge */
intel_pmu_lbr_init_snb(void)738 void __init intel_pmu_lbr_init_snb(void)
739 {
740 	x86_pmu.lbr_nr	 = 16;
741 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
742 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
743 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
744 
745 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
746 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
747 
748 	/*
749 	 * SW branch filter usage:
750 	 * - support syscall, sysret capture.
751 	 *   That requires LBR_FAR but that means far
752 	 *   jmp need to be filtered out
753 	 */
754 	pr_cont("16-deep LBR, ");
755 }
756 
757 /* atom */
intel_pmu_lbr_init_atom(void)758 void __init intel_pmu_lbr_init_atom(void)
759 {
760 	/*
761 	 * only models starting at stepping 10 seems
762 	 * to have an operational LBR which can freeze
763 	 * on PMU interrupt
764 	 */
765 	if (boot_cpu_data.x86_model == 28
766 	    && boot_cpu_data.x86_mask < 10) {
767 		pr_cont("LBR disabled due to erratum");
768 		return;
769 	}
770 
771 	x86_pmu.lbr_nr	   = 8;
772 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
773 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
774 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
775 
776 	/*
777 	 * SW branch filter usage:
778 	 * - compensate for lack of HW filter
779 	 */
780 	pr_cont("8-deep LBR, ");
781 }
782