xref: /openbmc/linux/arch/x86/events/intel/lbr.c (revision f3956ebb)
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 static const enum {
12 	LBR_EIP_FLAGS		= 1,
13 	LBR_TSX			= 2,
14 } lbr_desc[LBR_FORMAT_MAX_KNOWN + 1] = {
15 	[LBR_FORMAT_EIP_FLAGS]  = LBR_EIP_FLAGS,
16 	[LBR_FORMAT_EIP_FLAGS2] = LBR_EIP_FLAGS | LBR_TSX,
17 };
18 
19 /*
20  * Intel LBR_SELECT bits
21  * Intel Vol3a, April 2011, Section 16.7 Table 16-10
22  *
23  * Hardware branch filter (not available on all CPUs)
24  */
25 #define LBR_KERNEL_BIT		0 /* do not capture at ring0 */
26 #define LBR_USER_BIT		1 /* do not capture at ring > 0 */
27 #define LBR_JCC_BIT		2 /* do not capture conditional branches */
28 #define LBR_REL_CALL_BIT	3 /* do not capture relative calls */
29 #define LBR_IND_CALL_BIT	4 /* do not capture indirect calls */
30 #define LBR_RETURN_BIT		5 /* do not capture near returns */
31 #define LBR_IND_JMP_BIT		6 /* do not capture indirect jumps */
32 #define LBR_REL_JMP_BIT		7 /* do not capture relative jumps */
33 #define LBR_FAR_BIT		8 /* do not capture far branches */
34 #define LBR_CALL_STACK_BIT	9 /* enable call stack */
35 
36 /*
37  * Following bit only exists in Linux; we mask it out before writing it to
38  * the actual MSR. But it helps the constraint perf code to understand
39  * that this is a separate configuration.
40  */
41 #define LBR_NO_INFO_BIT	       63 /* don't read LBR_INFO. */
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 #define LBR_CALL_STACK	(1 << LBR_CALL_STACK_BIT)
53 #define LBR_NO_INFO	(1ULL << LBR_NO_INFO_BIT)
54 
55 #define LBR_PLM (LBR_KERNEL | LBR_USER)
56 
57 #define LBR_SEL_MASK	0x3ff	/* valid bits in LBR_SELECT */
58 #define LBR_NOT_SUPP	-1	/* LBR filter not supported */
59 #define LBR_IGN		0	/* ignored */
60 
61 #define LBR_ANY		 \
62 	(LBR_JCC	|\
63 	 LBR_REL_CALL	|\
64 	 LBR_IND_CALL	|\
65 	 LBR_RETURN	|\
66 	 LBR_REL_JMP	|\
67 	 LBR_IND_JMP	|\
68 	 LBR_FAR)
69 
70 #define LBR_FROM_FLAG_MISPRED	BIT_ULL(63)
71 #define LBR_FROM_FLAG_IN_TX	BIT_ULL(62)
72 #define LBR_FROM_FLAG_ABORT	BIT_ULL(61)
73 
74 #define LBR_FROM_SIGNEXT_2MSB	(BIT_ULL(60) | BIT_ULL(59))
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 	X86_BR_ZERO_CALL	= 1 << 15,/* zero length call */
100 	X86_BR_CALL_STACK	= 1 << 16,/* call stack */
101 	X86_BR_IND_JMP		= 1 << 17,/* indirect jump */
102 
103 	X86_BR_TYPE_SAVE	= 1 << 18,/* indicate to save branch type */
104 
105 };
106 
107 #define X86_BR_PLM (X86_BR_USER | X86_BR_KERNEL)
108 #define X86_BR_ANYTX (X86_BR_NO_TX | X86_BR_IN_TX)
109 
110 #define X86_BR_ANY       \
111 	(X86_BR_CALL    |\
112 	 X86_BR_RET     |\
113 	 X86_BR_SYSCALL |\
114 	 X86_BR_SYSRET  |\
115 	 X86_BR_INT     |\
116 	 X86_BR_IRET    |\
117 	 X86_BR_JCC     |\
118 	 X86_BR_JMP	 |\
119 	 X86_BR_IRQ	 |\
120 	 X86_BR_ABORT	 |\
121 	 X86_BR_IND_CALL |\
122 	 X86_BR_IND_JMP  |\
123 	 X86_BR_ZERO_CALL)
124 
125 #define X86_BR_ALL (X86_BR_PLM | X86_BR_ANY)
126 
127 #define X86_BR_ANY_CALL		 \
128 	(X86_BR_CALL		|\
129 	 X86_BR_IND_CALL	|\
130 	 X86_BR_ZERO_CALL	|\
131 	 X86_BR_SYSCALL		|\
132 	 X86_BR_IRQ		|\
133 	 X86_BR_INT)
134 
135 /*
136  * Intel LBR_CTL bits
137  *
138  * Hardware branch filter for Arch LBR
139  */
140 #define ARCH_LBR_KERNEL_BIT		1  /* capture at ring0 */
141 #define ARCH_LBR_USER_BIT		2  /* capture at ring > 0 */
142 #define ARCH_LBR_CALL_STACK_BIT		3  /* enable call stack */
143 #define ARCH_LBR_JCC_BIT		16 /* capture conditional branches */
144 #define ARCH_LBR_REL_JMP_BIT		17 /* capture relative jumps */
145 #define ARCH_LBR_IND_JMP_BIT		18 /* capture indirect jumps */
146 #define ARCH_LBR_REL_CALL_BIT		19 /* capture relative calls */
147 #define ARCH_LBR_IND_CALL_BIT		20 /* capture indirect calls */
148 #define ARCH_LBR_RETURN_BIT		21 /* capture near returns */
149 #define ARCH_LBR_OTHER_BRANCH_BIT	22 /* capture other branches */
150 
151 #define ARCH_LBR_KERNEL			(1ULL << ARCH_LBR_KERNEL_BIT)
152 #define ARCH_LBR_USER			(1ULL << ARCH_LBR_USER_BIT)
153 #define ARCH_LBR_CALL_STACK		(1ULL << ARCH_LBR_CALL_STACK_BIT)
154 #define ARCH_LBR_JCC			(1ULL << ARCH_LBR_JCC_BIT)
155 #define ARCH_LBR_REL_JMP		(1ULL << ARCH_LBR_REL_JMP_BIT)
156 #define ARCH_LBR_IND_JMP		(1ULL << ARCH_LBR_IND_JMP_BIT)
157 #define ARCH_LBR_REL_CALL		(1ULL << ARCH_LBR_REL_CALL_BIT)
158 #define ARCH_LBR_IND_CALL		(1ULL << ARCH_LBR_IND_CALL_BIT)
159 #define ARCH_LBR_RETURN			(1ULL << ARCH_LBR_RETURN_BIT)
160 #define ARCH_LBR_OTHER_BRANCH		(1ULL << ARCH_LBR_OTHER_BRANCH_BIT)
161 
162 #define ARCH_LBR_ANY			 \
163 	(ARCH_LBR_JCC			|\
164 	 ARCH_LBR_REL_JMP		|\
165 	 ARCH_LBR_IND_JMP		|\
166 	 ARCH_LBR_REL_CALL		|\
167 	 ARCH_LBR_IND_CALL		|\
168 	 ARCH_LBR_RETURN		|\
169 	 ARCH_LBR_OTHER_BRANCH)
170 
171 #define ARCH_LBR_CTL_MASK			0x7f000e
172 
173 static void intel_pmu_lbr_filter(struct cpu_hw_events *cpuc);
174 
175 static __always_inline bool is_lbr_call_stack_bit_set(u64 config)
176 {
177 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
178 		return !!(config & ARCH_LBR_CALL_STACK);
179 
180 	return !!(config & LBR_CALL_STACK);
181 }
182 
183 /*
184  * We only support LBR implementations that have FREEZE_LBRS_ON_PMI
185  * otherwise it becomes near impossible to get a reliable stack.
186  */
187 
188 static void __intel_pmu_lbr_enable(bool pmi)
189 {
190 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
191 	u64 debugctl, lbr_select = 0, orig_debugctl;
192 
193 	/*
194 	 * No need to unfreeze manually, as v4 can do that as part
195 	 * of the GLOBAL_STATUS ack.
196 	 */
197 	if (pmi && x86_pmu.version >= 4)
198 		return;
199 
200 	/*
201 	 * No need to reprogram LBR_SELECT in a PMI, as it
202 	 * did not change.
203 	 */
204 	if (cpuc->lbr_sel)
205 		lbr_select = cpuc->lbr_sel->config & x86_pmu.lbr_sel_mask;
206 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) && !pmi && cpuc->lbr_sel)
207 		wrmsrl(MSR_LBR_SELECT, lbr_select);
208 
209 	rdmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
210 	orig_debugctl = debugctl;
211 
212 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
213 		debugctl |= DEBUGCTLMSR_LBR;
214 	/*
215 	 * LBR callstack does not work well with FREEZE_LBRS_ON_PMI.
216 	 * If FREEZE_LBRS_ON_PMI is set, PMI near call/return instructions
217 	 * may cause superfluous increase/decrease of LBR_TOS.
218 	 */
219 	if (is_lbr_call_stack_bit_set(lbr_select))
220 		debugctl &= ~DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
221 	else
222 		debugctl |= DEBUGCTLMSR_FREEZE_LBRS_ON_PMI;
223 
224 	if (orig_debugctl != debugctl)
225 		wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctl);
226 
227 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
228 		wrmsrl(MSR_ARCH_LBR_CTL, lbr_select | ARCH_LBR_CTL_LBREN);
229 }
230 
231 void intel_pmu_lbr_reset_32(void)
232 {
233 	int i;
234 
235 	for (i = 0; i < x86_pmu.lbr_nr; i++)
236 		wrmsrl(x86_pmu.lbr_from + i, 0);
237 }
238 
239 void intel_pmu_lbr_reset_64(void)
240 {
241 	int i;
242 
243 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
244 		wrmsrl(x86_pmu.lbr_from + i, 0);
245 		wrmsrl(x86_pmu.lbr_to   + i, 0);
246 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
247 			wrmsrl(x86_pmu.lbr_info + i, 0);
248 	}
249 }
250 
251 static void intel_pmu_arch_lbr_reset(void)
252 {
253 	/* Write to ARCH_LBR_DEPTH MSR, all LBR entries are reset to 0 */
254 	wrmsrl(MSR_ARCH_LBR_DEPTH, x86_pmu.lbr_nr);
255 }
256 
257 void intel_pmu_lbr_reset(void)
258 {
259 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
260 
261 	if (!x86_pmu.lbr_nr)
262 		return;
263 
264 	x86_pmu.lbr_reset();
265 
266 	cpuc->last_task_ctx = NULL;
267 	cpuc->last_log_id = 0;
268 }
269 
270 /*
271  * TOS = most recently recorded branch
272  */
273 static inline u64 intel_pmu_lbr_tos(void)
274 {
275 	u64 tos;
276 
277 	rdmsrl(x86_pmu.lbr_tos, tos);
278 	return tos;
279 }
280 
281 enum {
282 	LBR_NONE,
283 	LBR_VALID,
284 };
285 
286 /*
287  * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
288  * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported, but when
289  * TSX is not supported they have no consistent behavior:
290  *
291  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
292  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
293  *     part of the sign extension.
294  *
295  * Therefore, if:
296  *
297  *   1) LBR has TSX format
298  *   2) CPU has no TSX support enabled
299  *
300  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
301  * value from rdmsr() must be converted to have a 61 bits sign extension,
302  * ignoring the TSX flags.
303  */
304 static inline bool lbr_from_signext_quirk_needed(void)
305 {
306 	int lbr_format = x86_pmu.intel_cap.lbr_format;
307 	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
308 			   boot_cpu_has(X86_FEATURE_RTM);
309 
310 	return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
311 }
312 
313 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
314 
315 /* If quirk is enabled, ensure sign extension is 63 bits: */
316 inline u64 lbr_from_signext_quirk_wr(u64 val)
317 {
318 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
319 		/*
320 		 * Sign extend into bits 61:62 while preserving bit 63.
321 		 *
322 		 * Quirk is enabled when TSX is disabled. Therefore TSX bits
323 		 * in val are always OFF and must be changed to be sign
324 		 * extension bits. Since bits 59:60 are guaranteed to be
325 		 * part of the sign extension bits, we can just copy them
326 		 * to 61:62.
327 		 */
328 		val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
329 	}
330 	return val;
331 }
332 
333 /*
334  * If quirk is needed, ensure sign extension is 61 bits:
335  */
336 static u64 lbr_from_signext_quirk_rd(u64 val)
337 {
338 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
339 		/*
340 		 * Quirk is on when TSX is not enabled. Therefore TSX
341 		 * flags must be read as OFF.
342 		 */
343 		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
344 	}
345 	return val;
346 }
347 
348 static __always_inline void wrlbr_from(unsigned int idx, u64 val)
349 {
350 	val = lbr_from_signext_quirk_wr(val);
351 	wrmsrl(x86_pmu.lbr_from + idx, val);
352 }
353 
354 static __always_inline void wrlbr_to(unsigned int idx, u64 val)
355 {
356 	wrmsrl(x86_pmu.lbr_to + idx, val);
357 }
358 
359 static __always_inline void wrlbr_info(unsigned int idx, u64 val)
360 {
361 	wrmsrl(x86_pmu.lbr_info + idx, val);
362 }
363 
364 static __always_inline u64 rdlbr_from(unsigned int idx, struct lbr_entry *lbr)
365 {
366 	u64 val;
367 
368 	if (lbr)
369 		return lbr->from;
370 
371 	rdmsrl(x86_pmu.lbr_from + idx, val);
372 
373 	return lbr_from_signext_quirk_rd(val);
374 }
375 
376 static __always_inline u64 rdlbr_to(unsigned int idx, struct lbr_entry *lbr)
377 {
378 	u64 val;
379 
380 	if (lbr)
381 		return lbr->to;
382 
383 	rdmsrl(x86_pmu.lbr_to + idx, val);
384 
385 	return val;
386 }
387 
388 static __always_inline u64 rdlbr_info(unsigned int idx, struct lbr_entry *lbr)
389 {
390 	u64 val;
391 
392 	if (lbr)
393 		return lbr->info;
394 
395 	rdmsrl(x86_pmu.lbr_info + idx, val);
396 
397 	return val;
398 }
399 
400 static inline void
401 wrlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
402 {
403 	wrlbr_from(idx, lbr->from);
404 	wrlbr_to(idx, lbr->to);
405 	if (need_info)
406 		wrlbr_info(idx, lbr->info);
407 }
408 
409 static inline bool
410 rdlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
411 {
412 	u64 from = rdlbr_from(idx, NULL);
413 
414 	/* Don't read invalid entry */
415 	if (!from)
416 		return false;
417 
418 	lbr->from = from;
419 	lbr->to = rdlbr_to(idx, NULL);
420 	if (need_info)
421 		lbr->info = rdlbr_info(idx, NULL);
422 
423 	return true;
424 }
425 
426 void intel_pmu_lbr_restore(void *ctx)
427 {
428 	bool need_info = x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO;
429 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
430 	struct x86_perf_task_context *task_ctx = ctx;
431 	int i;
432 	unsigned lbr_idx, mask;
433 	u64 tos = task_ctx->tos;
434 
435 	mask = x86_pmu.lbr_nr - 1;
436 	for (i = 0; i < task_ctx->valid_lbrs; i++) {
437 		lbr_idx = (tos - i) & mask;
438 		wrlbr_all(&task_ctx->lbr[i], lbr_idx, need_info);
439 	}
440 
441 	for (; i < x86_pmu.lbr_nr; i++) {
442 		lbr_idx = (tos - i) & mask;
443 		wrlbr_from(lbr_idx, 0);
444 		wrlbr_to(lbr_idx, 0);
445 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
446 			wrlbr_info(lbr_idx, 0);
447 	}
448 
449 	wrmsrl(x86_pmu.lbr_tos, tos);
450 
451 	if (cpuc->lbr_select)
452 		wrmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
453 }
454 
455 static void intel_pmu_arch_lbr_restore(void *ctx)
456 {
457 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
458 	struct lbr_entry *entries = task_ctx->entries;
459 	int i;
460 
461 	/* Fast reset the LBRs before restore if the call stack is not full. */
462 	if (!entries[x86_pmu.lbr_nr - 1].from)
463 		intel_pmu_arch_lbr_reset();
464 
465 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
466 		if (!entries[i].from)
467 			break;
468 		wrlbr_all(&entries[i], i, true);
469 	}
470 }
471 
472 /*
473  * Restore the Architecture LBR state from the xsave area in the perf
474  * context data for the task via the XRSTORS instruction.
475  */
476 static void intel_pmu_arch_lbr_xrstors(void *ctx)
477 {
478 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
479 
480 	xrstors(&task_ctx->xsave, XFEATURE_MASK_LBR);
481 }
482 
483 static __always_inline bool lbr_is_reset_in_cstate(void *ctx)
484 {
485 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
486 		return x86_pmu.lbr_deep_c_reset && !rdlbr_from(0, NULL);
487 
488 	return !rdlbr_from(((struct x86_perf_task_context *)ctx)->tos, NULL);
489 }
490 
491 static void __intel_pmu_lbr_restore(void *ctx)
492 {
493 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
494 
495 	if (task_context_opt(ctx)->lbr_callstack_users == 0 ||
496 	    task_context_opt(ctx)->lbr_stack_state == LBR_NONE) {
497 		intel_pmu_lbr_reset();
498 		return;
499 	}
500 
501 	/*
502 	 * Does not restore the LBR registers, if
503 	 * - No one else touched them, and
504 	 * - Was not cleared in Cstate
505 	 */
506 	if ((ctx == cpuc->last_task_ctx) &&
507 	    (task_context_opt(ctx)->log_id == cpuc->last_log_id) &&
508 	    !lbr_is_reset_in_cstate(ctx)) {
509 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
510 		return;
511 	}
512 
513 	x86_pmu.lbr_restore(ctx);
514 
515 	task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
516 }
517 
518 void intel_pmu_lbr_save(void *ctx)
519 {
520 	bool need_info = x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO;
521 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
522 	struct x86_perf_task_context *task_ctx = ctx;
523 	unsigned lbr_idx, mask;
524 	u64 tos;
525 	int i;
526 
527 	mask = x86_pmu.lbr_nr - 1;
528 	tos = intel_pmu_lbr_tos();
529 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
530 		lbr_idx = (tos - i) & mask;
531 		if (!rdlbr_all(&task_ctx->lbr[i], lbr_idx, need_info))
532 			break;
533 	}
534 	task_ctx->valid_lbrs = i;
535 	task_ctx->tos = tos;
536 
537 	if (cpuc->lbr_select)
538 		rdmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
539 }
540 
541 static void intel_pmu_arch_lbr_save(void *ctx)
542 {
543 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
544 	struct lbr_entry *entries = task_ctx->entries;
545 	int i;
546 
547 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
548 		if (!rdlbr_all(&entries[i], i, true))
549 			break;
550 	}
551 
552 	/* LBR call stack is not full. Reset is required in restore. */
553 	if (i < x86_pmu.lbr_nr)
554 		entries[x86_pmu.lbr_nr - 1].from = 0;
555 }
556 
557 /*
558  * Save the Architecture LBR state to the xsave area in the perf
559  * context data for the task via the XSAVES instruction.
560  */
561 static void intel_pmu_arch_lbr_xsaves(void *ctx)
562 {
563 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
564 
565 	xsaves(&task_ctx->xsave, XFEATURE_MASK_LBR);
566 }
567 
568 static void __intel_pmu_lbr_save(void *ctx)
569 {
570 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
571 
572 	if (task_context_opt(ctx)->lbr_callstack_users == 0) {
573 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
574 		return;
575 	}
576 
577 	x86_pmu.lbr_save(ctx);
578 
579 	task_context_opt(ctx)->lbr_stack_state = LBR_VALID;
580 
581 	cpuc->last_task_ctx = ctx;
582 	cpuc->last_log_id = ++task_context_opt(ctx)->log_id;
583 }
584 
585 void intel_pmu_lbr_swap_task_ctx(struct perf_event_context *prev,
586 				 struct perf_event_context *next)
587 {
588 	void *prev_ctx_data, *next_ctx_data;
589 
590 	swap(prev->task_ctx_data, next->task_ctx_data);
591 
592 	/*
593 	 * Architecture specific synchronization makes sense in
594 	 * case both prev->task_ctx_data and next->task_ctx_data
595 	 * pointers are allocated.
596 	 */
597 
598 	prev_ctx_data = next->task_ctx_data;
599 	next_ctx_data = prev->task_ctx_data;
600 
601 	if (!prev_ctx_data || !next_ctx_data)
602 		return;
603 
604 	swap(task_context_opt(prev_ctx_data)->lbr_callstack_users,
605 	     task_context_opt(next_ctx_data)->lbr_callstack_users);
606 }
607 
608 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
609 {
610 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
611 	void *task_ctx;
612 
613 	if (!cpuc->lbr_users)
614 		return;
615 
616 	/*
617 	 * If LBR callstack feature is enabled and the stack was saved when
618 	 * the task was scheduled out, restore the stack. Otherwise flush
619 	 * the LBR stack.
620 	 */
621 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
622 	if (task_ctx) {
623 		if (sched_in)
624 			__intel_pmu_lbr_restore(task_ctx);
625 		else
626 			__intel_pmu_lbr_save(task_ctx);
627 		return;
628 	}
629 
630 	/*
631 	 * Since a context switch can flip the address space and LBR entries
632 	 * are not tagged with an identifier, we need to wipe the LBR, even for
633 	 * per-cpu events. You simply cannot resolve the branches from the old
634 	 * address space.
635 	 */
636 	if (sched_in)
637 		intel_pmu_lbr_reset();
638 }
639 
640 static inline bool branch_user_callstack(unsigned br_sel)
641 {
642 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
643 }
644 
645 void intel_pmu_lbr_add(struct perf_event *event)
646 {
647 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
648 
649 	if (!x86_pmu.lbr_nr)
650 		return;
651 
652 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
653 		cpuc->lbr_select = 1;
654 
655 	cpuc->br_sel = event->hw.branch_reg.reg;
656 
657 	if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data)
658 		task_context_opt(event->ctx->task_ctx_data)->lbr_callstack_users++;
659 
660 	/*
661 	 * Request pmu::sched_task() callback, which will fire inside the
662 	 * regular perf event scheduling, so that call will:
663 	 *
664 	 *  - restore or wipe; when LBR-callstack,
665 	 *  - wipe; otherwise,
666 	 *
667 	 * when this is from __perf_event_task_sched_in().
668 	 *
669 	 * However, if this is from perf_install_in_context(), no such callback
670 	 * will follow and we'll need to reset the LBR here if this is the
671 	 * first LBR event.
672 	 *
673 	 * The problem is, we cannot tell these cases apart... but we can
674 	 * exclude the biggest chunk of cases by looking at
675 	 * event->total_time_running. An event that has accrued runtime cannot
676 	 * be 'new'. Conversely, a new event can get installed through the
677 	 * context switch path for the first time.
678 	 */
679 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
680 		cpuc->lbr_pebs_users++;
681 	perf_sched_cb_inc(event->ctx->pmu);
682 	if (!cpuc->lbr_users++ && !event->total_time_running)
683 		intel_pmu_lbr_reset();
684 }
685 
686 void release_lbr_buffers(void)
687 {
688 	struct kmem_cache *kmem_cache;
689 	struct cpu_hw_events *cpuc;
690 	int cpu;
691 
692 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
693 		return;
694 
695 	for_each_possible_cpu(cpu) {
696 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
697 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
698 		if (kmem_cache && cpuc->lbr_xsave) {
699 			kmem_cache_free(kmem_cache, cpuc->lbr_xsave);
700 			cpuc->lbr_xsave = NULL;
701 		}
702 	}
703 }
704 
705 void reserve_lbr_buffers(void)
706 {
707 	struct kmem_cache *kmem_cache;
708 	struct cpu_hw_events *cpuc;
709 	int cpu;
710 
711 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
712 		return;
713 
714 	for_each_possible_cpu(cpu) {
715 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
716 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
717 		if (!kmem_cache || cpuc->lbr_xsave)
718 			continue;
719 
720 		cpuc->lbr_xsave = kmem_cache_alloc_node(kmem_cache,
721 							GFP_KERNEL | __GFP_ZERO,
722 							cpu_to_node(cpu));
723 	}
724 }
725 
726 void intel_pmu_lbr_del(struct perf_event *event)
727 {
728 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
729 
730 	if (!x86_pmu.lbr_nr)
731 		return;
732 
733 	if (branch_user_callstack(cpuc->br_sel) &&
734 	    event->ctx->task_ctx_data)
735 		task_context_opt(event->ctx->task_ctx_data)->lbr_callstack_users--;
736 
737 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
738 		cpuc->lbr_select = 0;
739 
740 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
741 		cpuc->lbr_pebs_users--;
742 	cpuc->lbr_users--;
743 	WARN_ON_ONCE(cpuc->lbr_users < 0);
744 	WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
745 	perf_sched_cb_dec(event->ctx->pmu);
746 }
747 
748 static inline bool vlbr_exclude_host(void)
749 {
750 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
751 
752 	return test_bit(INTEL_PMC_IDX_FIXED_VLBR,
753 		(unsigned long *)&cpuc->intel_ctrl_guest_mask);
754 }
755 
756 void intel_pmu_lbr_enable_all(bool pmi)
757 {
758 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
759 
760 	if (cpuc->lbr_users && !vlbr_exclude_host())
761 		__intel_pmu_lbr_enable(pmi);
762 }
763 
764 void intel_pmu_lbr_disable_all(void)
765 {
766 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
767 
768 	if (cpuc->lbr_users && !vlbr_exclude_host()) {
769 		if (static_cpu_has(X86_FEATURE_ARCH_LBR))
770 			return __intel_pmu_arch_lbr_disable();
771 
772 		__intel_pmu_lbr_disable();
773 	}
774 }
775 
776 void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
777 {
778 	unsigned long mask = x86_pmu.lbr_nr - 1;
779 	u64 tos = intel_pmu_lbr_tos();
780 	int i;
781 
782 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
783 		unsigned long lbr_idx = (tos - i) & mask;
784 		union {
785 			struct {
786 				u32 from;
787 				u32 to;
788 			};
789 			u64     lbr;
790 		} msr_lastbranch;
791 
792 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
793 
794 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
795 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
796 		cpuc->lbr_entries[i].mispred	= 0;
797 		cpuc->lbr_entries[i].predicted	= 0;
798 		cpuc->lbr_entries[i].in_tx	= 0;
799 		cpuc->lbr_entries[i].abort	= 0;
800 		cpuc->lbr_entries[i].cycles	= 0;
801 		cpuc->lbr_entries[i].type	= 0;
802 		cpuc->lbr_entries[i].reserved	= 0;
803 	}
804 	cpuc->lbr_stack.nr = i;
805 	cpuc->lbr_stack.hw_idx = tos;
806 }
807 
808 /*
809  * Due to lack of segmentation in Linux the effective address (offset)
810  * is the same as the linear address, allowing us to merge the LIP and EIP
811  * LBR formats.
812  */
813 void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
814 {
815 	bool need_info = false, call_stack = false;
816 	unsigned long mask = x86_pmu.lbr_nr - 1;
817 	int lbr_format = x86_pmu.intel_cap.lbr_format;
818 	u64 tos = intel_pmu_lbr_tos();
819 	int i;
820 	int out = 0;
821 	int num = x86_pmu.lbr_nr;
822 
823 	if (cpuc->lbr_sel) {
824 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
825 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
826 			call_stack = true;
827 	}
828 
829 	for (i = 0; i < num; i++) {
830 		unsigned long lbr_idx = (tos - i) & mask;
831 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
832 		int skip = 0;
833 		u16 cycles = 0;
834 		int lbr_flags = lbr_desc[lbr_format];
835 
836 		from = rdlbr_from(lbr_idx, NULL);
837 		to   = rdlbr_to(lbr_idx, NULL);
838 
839 		/*
840 		 * Read LBR call stack entries
841 		 * until invalid entry (0s) is detected.
842 		 */
843 		if (call_stack && !from)
844 			break;
845 
846 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
847 			u64 info;
848 
849 			info = rdlbr_info(lbr_idx, NULL);
850 			mis = !!(info & LBR_INFO_MISPRED);
851 			pred = !mis;
852 			in_tx = !!(info & LBR_INFO_IN_TX);
853 			abort = !!(info & LBR_INFO_ABORT);
854 			cycles = (info & LBR_INFO_CYCLES);
855 		}
856 
857 		if (lbr_format == LBR_FORMAT_TIME) {
858 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
859 			pred = !mis;
860 			skip = 1;
861 			cycles = ((to >> 48) & LBR_INFO_CYCLES);
862 
863 			to = (u64)((((s64)to) << 16) >> 16);
864 		}
865 
866 		if (lbr_flags & LBR_EIP_FLAGS) {
867 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
868 			pred = !mis;
869 			skip = 1;
870 		}
871 		if (lbr_flags & LBR_TSX) {
872 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
873 			abort = !!(from & LBR_FROM_FLAG_ABORT);
874 			skip = 3;
875 		}
876 		from = (u64)((((s64)from) << skip) >> skip);
877 
878 		/*
879 		 * Some CPUs report duplicated abort records,
880 		 * with the second entry not having an abort bit set.
881 		 * Skip them here. This loop runs backwards,
882 		 * so we need to undo the previous record.
883 		 * If the abort just happened outside the window
884 		 * the extra entry cannot be removed.
885 		 */
886 		if (abort && x86_pmu.lbr_double_abort && out > 0)
887 			out--;
888 
889 		cpuc->lbr_entries[out].from	 = from;
890 		cpuc->lbr_entries[out].to	 = to;
891 		cpuc->lbr_entries[out].mispred	 = mis;
892 		cpuc->lbr_entries[out].predicted = pred;
893 		cpuc->lbr_entries[out].in_tx	 = in_tx;
894 		cpuc->lbr_entries[out].abort	 = abort;
895 		cpuc->lbr_entries[out].cycles	 = cycles;
896 		cpuc->lbr_entries[out].type	 = 0;
897 		cpuc->lbr_entries[out].reserved	 = 0;
898 		out++;
899 	}
900 	cpuc->lbr_stack.nr = out;
901 	cpuc->lbr_stack.hw_idx = tos;
902 }
903 
904 static __always_inline int get_lbr_br_type(u64 info)
905 {
906 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) || !x86_pmu.lbr_br_type)
907 		return 0;
908 
909 	return (info & LBR_INFO_BR_TYPE) >> LBR_INFO_BR_TYPE_OFFSET;
910 }
911 
912 static __always_inline bool get_lbr_mispred(u64 info)
913 {
914 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) && !x86_pmu.lbr_mispred)
915 		return 0;
916 
917 	return !!(info & LBR_INFO_MISPRED);
918 }
919 
920 static __always_inline bool get_lbr_predicted(u64 info)
921 {
922 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) && !x86_pmu.lbr_mispred)
923 		return 0;
924 
925 	return !(info & LBR_INFO_MISPRED);
926 }
927 
928 static __always_inline u16 get_lbr_cycles(u64 info)
929 {
930 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
931 	    !(x86_pmu.lbr_timed_lbr && info & LBR_INFO_CYC_CNT_VALID))
932 		return 0;
933 
934 	return info & LBR_INFO_CYCLES;
935 }
936 
937 static void intel_pmu_store_lbr(struct cpu_hw_events *cpuc,
938 				struct lbr_entry *entries)
939 {
940 	struct perf_branch_entry *e;
941 	struct lbr_entry *lbr;
942 	u64 from, to, info;
943 	int i;
944 
945 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
946 		lbr = entries ? &entries[i] : NULL;
947 		e = &cpuc->lbr_entries[i];
948 
949 		from = rdlbr_from(i, lbr);
950 		/*
951 		 * Read LBR entries until invalid entry (0s) is detected.
952 		 */
953 		if (!from)
954 			break;
955 
956 		to = rdlbr_to(i, lbr);
957 		info = rdlbr_info(i, lbr);
958 
959 		e->from		= from;
960 		e->to		= to;
961 		e->mispred	= get_lbr_mispred(info);
962 		e->predicted	= get_lbr_predicted(info);
963 		e->in_tx	= !!(info & LBR_INFO_IN_TX);
964 		e->abort	= !!(info & LBR_INFO_ABORT);
965 		e->cycles	= get_lbr_cycles(info);
966 		e->type		= get_lbr_br_type(info);
967 		e->reserved	= 0;
968 	}
969 
970 	cpuc->lbr_stack.nr = i;
971 }
972 
973 static void intel_pmu_arch_lbr_read(struct cpu_hw_events *cpuc)
974 {
975 	intel_pmu_store_lbr(cpuc, NULL);
976 }
977 
978 static void intel_pmu_arch_lbr_read_xsave(struct cpu_hw_events *cpuc)
979 {
980 	struct x86_perf_task_context_arch_lbr_xsave *xsave = cpuc->lbr_xsave;
981 
982 	if (!xsave) {
983 		intel_pmu_store_lbr(cpuc, NULL);
984 		return;
985 	}
986 	xsaves(&xsave->xsave, XFEATURE_MASK_LBR);
987 
988 	intel_pmu_store_lbr(cpuc, xsave->lbr.entries);
989 }
990 
991 void intel_pmu_lbr_read(void)
992 {
993 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
994 
995 	/*
996 	 * Don't read when all LBRs users are using adaptive PEBS.
997 	 *
998 	 * This could be smarter and actually check the event,
999 	 * but this simple approach seems to work for now.
1000 	 */
1001 	if (!cpuc->lbr_users || vlbr_exclude_host() ||
1002 	    cpuc->lbr_users == cpuc->lbr_pebs_users)
1003 		return;
1004 
1005 	x86_pmu.lbr_read(cpuc);
1006 
1007 	intel_pmu_lbr_filter(cpuc);
1008 }
1009 
1010 /*
1011  * SW filter is used:
1012  * - in case there is no HW filter
1013  * - in case the HW filter has errata or limitations
1014  */
1015 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
1016 {
1017 	u64 br_type = event->attr.branch_sample_type;
1018 	int mask = 0;
1019 
1020 	if (br_type & PERF_SAMPLE_BRANCH_USER)
1021 		mask |= X86_BR_USER;
1022 
1023 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
1024 		mask |= X86_BR_KERNEL;
1025 
1026 	/* we ignore BRANCH_HV here */
1027 
1028 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
1029 		mask |= X86_BR_ANY;
1030 
1031 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
1032 		mask |= X86_BR_ANY_CALL;
1033 
1034 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
1035 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
1036 
1037 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
1038 		mask |= X86_BR_IND_CALL;
1039 
1040 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
1041 		mask |= X86_BR_ABORT;
1042 
1043 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
1044 		mask |= X86_BR_IN_TX;
1045 
1046 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
1047 		mask |= X86_BR_NO_TX;
1048 
1049 	if (br_type & PERF_SAMPLE_BRANCH_COND)
1050 		mask |= X86_BR_JCC;
1051 
1052 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
1053 		if (!x86_pmu_has_lbr_callstack())
1054 			return -EOPNOTSUPP;
1055 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
1056 			return -EINVAL;
1057 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
1058 			X86_BR_CALL_STACK;
1059 	}
1060 
1061 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
1062 		mask |= X86_BR_IND_JMP;
1063 
1064 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
1065 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
1066 
1067 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
1068 		mask |= X86_BR_TYPE_SAVE;
1069 
1070 	/*
1071 	 * stash actual user request into reg, it may
1072 	 * be used by fixup code for some CPU
1073 	 */
1074 	event->hw.branch_reg.reg = mask;
1075 	return 0;
1076 }
1077 
1078 /*
1079  * setup the HW LBR filter
1080  * Used only when available, may not be enough to disambiguate
1081  * all branches, may need the help of the SW filter
1082  */
1083 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
1084 {
1085 	struct hw_perf_event_extra *reg;
1086 	u64 br_type = event->attr.branch_sample_type;
1087 	u64 mask = 0, v;
1088 	int i;
1089 
1090 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
1091 		if (!(br_type & (1ULL << i)))
1092 			continue;
1093 
1094 		v = x86_pmu.lbr_sel_map[i];
1095 		if (v == LBR_NOT_SUPP)
1096 			return -EOPNOTSUPP;
1097 
1098 		if (v != LBR_IGN)
1099 			mask |= v;
1100 	}
1101 
1102 	reg = &event->hw.branch_reg;
1103 	reg->idx = EXTRA_REG_LBR;
1104 
1105 	if (static_cpu_has(X86_FEATURE_ARCH_LBR)) {
1106 		reg->config = mask;
1107 		return 0;
1108 	}
1109 
1110 	/*
1111 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
1112 	 * in suppress mode. So LBR_SELECT should be set to
1113 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
1114 	 * But the 10th bit LBR_CALL_STACK does not operate
1115 	 * in suppress mode.
1116 	 */
1117 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
1118 
1119 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
1120 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
1121 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
1122 		reg->config |= LBR_NO_INFO;
1123 
1124 	return 0;
1125 }
1126 
1127 int intel_pmu_setup_lbr_filter(struct perf_event *event)
1128 {
1129 	int ret = 0;
1130 
1131 	/*
1132 	 * no LBR on this PMU
1133 	 */
1134 	if (!x86_pmu.lbr_nr)
1135 		return -EOPNOTSUPP;
1136 
1137 	/*
1138 	 * setup SW LBR filter
1139 	 */
1140 	ret = intel_pmu_setup_sw_lbr_filter(event);
1141 	if (ret)
1142 		return ret;
1143 
1144 	/*
1145 	 * setup HW LBR filter, if any
1146 	 */
1147 	if (x86_pmu.lbr_sel_map)
1148 		ret = intel_pmu_setup_hw_lbr_filter(event);
1149 
1150 	return ret;
1151 }
1152 
1153 /*
1154  * return the type of control flow change at address "from"
1155  * instruction is not necessarily a branch (in case of interrupt).
1156  *
1157  * The branch type returned also includes the priv level of the
1158  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
1159  *
1160  * If a branch type is unknown OR the instruction cannot be
1161  * decoded (e.g., text page not present), then X86_BR_NONE is
1162  * returned.
1163  */
1164 static int branch_type(unsigned long from, unsigned long to, int abort)
1165 {
1166 	struct insn insn;
1167 	void *addr;
1168 	int bytes_read, bytes_left;
1169 	int ret = X86_BR_NONE;
1170 	int ext, to_plm, from_plm;
1171 	u8 buf[MAX_INSN_SIZE];
1172 	int is64 = 0;
1173 
1174 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
1175 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
1176 
1177 	/*
1178 	 * maybe zero if lbr did not fill up after a reset by the time
1179 	 * we get a PMU interrupt
1180 	 */
1181 	if (from == 0 || to == 0)
1182 		return X86_BR_NONE;
1183 
1184 	if (abort)
1185 		return X86_BR_ABORT | to_plm;
1186 
1187 	if (from_plm == X86_BR_USER) {
1188 		/*
1189 		 * can happen if measuring at the user level only
1190 		 * and we interrupt in a kernel thread, e.g., idle.
1191 		 */
1192 		if (!current->mm)
1193 			return X86_BR_NONE;
1194 
1195 		/* may fail if text not present */
1196 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
1197 						MAX_INSN_SIZE);
1198 		bytes_read = MAX_INSN_SIZE - bytes_left;
1199 		if (!bytes_read)
1200 			return X86_BR_NONE;
1201 
1202 		addr = buf;
1203 	} else {
1204 		/*
1205 		 * The LBR logs any address in the IP, even if the IP just
1206 		 * faulted. This means userspace can control the from address.
1207 		 * Ensure we don't blindly read any address by validating it is
1208 		 * a known text address.
1209 		 */
1210 		if (kernel_text_address(from)) {
1211 			addr = (void *)from;
1212 			/*
1213 			 * Assume we can get the maximum possible size
1214 			 * when grabbing kernel data.  This is not
1215 			 * _strictly_ true since we could possibly be
1216 			 * executing up next to a memory hole, but
1217 			 * it is very unlikely to be a problem.
1218 			 */
1219 			bytes_read = MAX_INSN_SIZE;
1220 		} else {
1221 			return X86_BR_NONE;
1222 		}
1223 	}
1224 
1225 	/*
1226 	 * decoder needs to know the ABI especially
1227 	 * on 64-bit systems running 32-bit apps
1228 	 */
1229 #ifdef CONFIG_X86_64
1230 	is64 = kernel_ip((unsigned long)addr) || any_64bit_mode(current_pt_regs());
1231 #endif
1232 	insn_init(&insn, addr, bytes_read, is64);
1233 	if (insn_get_opcode(&insn))
1234 		return X86_BR_ABORT;
1235 
1236 	switch (insn.opcode.bytes[0]) {
1237 	case 0xf:
1238 		switch (insn.opcode.bytes[1]) {
1239 		case 0x05: /* syscall */
1240 		case 0x34: /* sysenter */
1241 			ret = X86_BR_SYSCALL;
1242 			break;
1243 		case 0x07: /* sysret */
1244 		case 0x35: /* sysexit */
1245 			ret = X86_BR_SYSRET;
1246 			break;
1247 		case 0x80 ... 0x8f: /* conditional */
1248 			ret = X86_BR_JCC;
1249 			break;
1250 		default:
1251 			ret = X86_BR_NONE;
1252 		}
1253 		break;
1254 	case 0x70 ... 0x7f: /* conditional */
1255 		ret = X86_BR_JCC;
1256 		break;
1257 	case 0xc2: /* near ret */
1258 	case 0xc3: /* near ret */
1259 	case 0xca: /* far ret */
1260 	case 0xcb: /* far ret */
1261 		ret = X86_BR_RET;
1262 		break;
1263 	case 0xcf: /* iret */
1264 		ret = X86_BR_IRET;
1265 		break;
1266 	case 0xcc ... 0xce: /* int */
1267 		ret = X86_BR_INT;
1268 		break;
1269 	case 0xe8: /* call near rel */
1270 		if (insn_get_immediate(&insn) || insn.immediate1.value == 0) {
1271 			/* zero length call */
1272 			ret = X86_BR_ZERO_CALL;
1273 			break;
1274 		}
1275 		fallthrough;
1276 	case 0x9a: /* call far absolute */
1277 		ret = X86_BR_CALL;
1278 		break;
1279 	case 0xe0 ... 0xe3: /* loop jmp */
1280 		ret = X86_BR_JCC;
1281 		break;
1282 	case 0xe9 ... 0xeb: /* jmp */
1283 		ret = X86_BR_JMP;
1284 		break;
1285 	case 0xff: /* call near absolute, call far absolute ind */
1286 		if (insn_get_modrm(&insn))
1287 			return X86_BR_ABORT;
1288 
1289 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
1290 		switch (ext) {
1291 		case 2: /* near ind call */
1292 		case 3: /* far ind call */
1293 			ret = X86_BR_IND_CALL;
1294 			break;
1295 		case 4:
1296 		case 5:
1297 			ret = X86_BR_IND_JMP;
1298 			break;
1299 		}
1300 		break;
1301 	default:
1302 		ret = X86_BR_NONE;
1303 	}
1304 	/*
1305 	 * interrupts, traps, faults (and thus ring transition) may
1306 	 * occur on any instructions. Thus, to classify them correctly,
1307 	 * we need to first look at the from and to priv levels. If they
1308 	 * are different and to is in the kernel, then it indicates
1309 	 * a ring transition. If the from instruction is not a ring
1310 	 * transition instr (syscall, systenter, int), then it means
1311 	 * it was a irq, trap or fault.
1312 	 *
1313 	 * we have no way of detecting kernel to kernel faults.
1314 	 */
1315 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
1316 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
1317 		ret = X86_BR_IRQ;
1318 
1319 	/*
1320 	 * branch priv level determined by target as
1321 	 * is done by HW when LBR_SELECT is implemented
1322 	 */
1323 	if (ret != X86_BR_NONE)
1324 		ret |= to_plm;
1325 
1326 	return ret;
1327 }
1328 
1329 #define X86_BR_TYPE_MAP_MAX	16
1330 
1331 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
1332 	PERF_BR_CALL,		/* X86_BR_CALL */
1333 	PERF_BR_RET,		/* X86_BR_RET */
1334 	PERF_BR_SYSCALL,	/* X86_BR_SYSCALL */
1335 	PERF_BR_SYSRET,		/* X86_BR_SYSRET */
1336 	PERF_BR_UNKNOWN,	/* X86_BR_INT */
1337 	PERF_BR_UNKNOWN,	/* X86_BR_IRET */
1338 	PERF_BR_COND,		/* X86_BR_JCC */
1339 	PERF_BR_UNCOND,		/* X86_BR_JMP */
1340 	PERF_BR_UNKNOWN,	/* X86_BR_IRQ */
1341 	PERF_BR_IND_CALL,	/* X86_BR_IND_CALL */
1342 	PERF_BR_UNKNOWN,	/* X86_BR_ABORT */
1343 	PERF_BR_UNKNOWN,	/* X86_BR_IN_TX */
1344 	PERF_BR_UNKNOWN,	/* X86_BR_NO_TX */
1345 	PERF_BR_CALL,		/* X86_BR_ZERO_CALL */
1346 	PERF_BR_UNKNOWN,	/* X86_BR_CALL_STACK */
1347 	PERF_BR_IND,		/* X86_BR_IND_JMP */
1348 };
1349 
1350 static int
1351 common_branch_type(int type)
1352 {
1353 	int i;
1354 
1355 	type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1356 
1357 	if (type) {
1358 		i = __ffs(type);
1359 		if (i < X86_BR_TYPE_MAP_MAX)
1360 			return branch_map[i];
1361 	}
1362 
1363 	return PERF_BR_UNKNOWN;
1364 }
1365 
1366 enum {
1367 	ARCH_LBR_BR_TYPE_JCC			= 0,
1368 	ARCH_LBR_BR_TYPE_NEAR_IND_JMP		= 1,
1369 	ARCH_LBR_BR_TYPE_NEAR_REL_JMP		= 2,
1370 	ARCH_LBR_BR_TYPE_NEAR_IND_CALL		= 3,
1371 	ARCH_LBR_BR_TYPE_NEAR_REL_CALL		= 4,
1372 	ARCH_LBR_BR_TYPE_NEAR_RET		= 5,
1373 	ARCH_LBR_BR_TYPE_KNOWN_MAX		= ARCH_LBR_BR_TYPE_NEAR_RET,
1374 
1375 	ARCH_LBR_BR_TYPE_MAP_MAX		= 16,
1376 };
1377 
1378 static const int arch_lbr_br_type_map[ARCH_LBR_BR_TYPE_MAP_MAX] = {
1379 	[ARCH_LBR_BR_TYPE_JCC]			= X86_BR_JCC,
1380 	[ARCH_LBR_BR_TYPE_NEAR_IND_JMP]		= X86_BR_IND_JMP,
1381 	[ARCH_LBR_BR_TYPE_NEAR_REL_JMP]		= X86_BR_JMP,
1382 	[ARCH_LBR_BR_TYPE_NEAR_IND_CALL]	= X86_BR_IND_CALL,
1383 	[ARCH_LBR_BR_TYPE_NEAR_REL_CALL]	= X86_BR_CALL,
1384 	[ARCH_LBR_BR_TYPE_NEAR_RET]		= X86_BR_RET,
1385 };
1386 
1387 /*
1388  * implement actual branch filter based on user demand.
1389  * Hardware may not exactly satisfy that request, thus
1390  * we need to inspect opcodes. Mismatched branches are
1391  * discarded. Therefore, the number of branches returned
1392  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1393  */
1394 static void
1395 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1396 {
1397 	u64 from, to;
1398 	int br_sel = cpuc->br_sel;
1399 	int i, j, type, to_plm;
1400 	bool compress = false;
1401 
1402 	/* if sampling all branches, then nothing to filter */
1403 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1404 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1405 		return;
1406 
1407 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1408 
1409 		from = cpuc->lbr_entries[i].from;
1410 		to = cpuc->lbr_entries[i].to;
1411 		type = cpuc->lbr_entries[i].type;
1412 
1413 		/*
1414 		 * Parse the branch type recorded in LBR_x_INFO MSR.
1415 		 * Doesn't support OTHER_BRANCH decoding for now.
1416 		 * OTHER_BRANCH branch type still rely on software decoding.
1417 		 */
1418 		if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
1419 		    type <= ARCH_LBR_BR_TYPE_KNOWN_MAX) {
1420 			to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
1421 			type = arch_lbr_br_type_map[type] | to_plm;
1422 		} else
1423 			type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1424 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1425 			if (cpuc->lbr_entries[i].in_tx)
1426 				type |= X86_BR_IN_TX;
1427 			else
1428 				type |= X86_BR_NO_TX;
1429 		}
1430 
1431 		/* if type does not correspond, then discard */
1432 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1433 			cpuc->lbr_entries[i].from = 0;
1434 			compress = true;
1435 		}
1436 
1437 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1438 			cpuc->lbr_entries[i].type = common_branch_type(type);
1439 	}
1440 
1441 	if (!compress)
1442 		return;
1443 
1444 	/* remove all entries with from=0 */
1445 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1446 		if (!cpuc->lbr_entries[i].from) {
1447 			j = i;
1448 			while (++j < cpuc->lbr_stack.nr)
1449 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1450 			cpuc->lbr_stack.nr--;
1451 			if (!cpuc->lbr_entries[i].from)
1452 				continue;
1453 		}
1454 		i++;
1455 	}
1456 }
1457 
1458 void intel_pmu_store_pebs_lbrs(struct lbr_entry *lbr)
1459 {
1460 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1461 
1462 	/* Cannot get TOS for large PEBS and Arch LBR */
1463 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) ||
1464 	    (cpuc->n_pebs == cpuc->n_large_pebs))
1465 		cpuc->lbr_stack.hw_idx = -1ULL;
1466 	else
1467 		cpuc->lbr_stack.hw_idx = intel_pmu_lbr_tos();
1468 
1469 	intel_pmu_store_lbr(cpuc, lbr);
1470 	intel_pmu_lbr_filter(cpuc);
1471 }
1472 
1473 /*
1474  * Map interface branch filters onto LBR filters
1475  */
1476 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1477 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1478 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1479 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1480 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1481 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1482 						| LBR_IND_JMP | LBR_FAR,
1483 	/*
1484 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1485 	 */
1486 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1487 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1488 	/*
1489 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1490 	 */
1491 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1492 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1493 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1494 };
1495 
1496 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1497 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1498 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1499 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1500 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1501 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1502 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1503 						| LBR_FAR,
1504 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1505 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1506 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1507 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1508 };
1509 
1510 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1511 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1512 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1513 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1514 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1515 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1516 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1517 						| LBR_FAR,
1518 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1519 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1520 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1521 						| LBR_RETURN | LBR_CALL_STACK,
1522 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1523 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1524 };
1525 
1526 static int arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1527 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= ARCH_LBR_ANY,
1528 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= ARCH_LBR_USER,
1529 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= ARCH_LBR_KERNEL,
1530 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1531 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= ARCH_LBR_RETURN |
1532 						  ARCH_LBR_OTHER_BRANCH,
1533 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]     = ARCH_LBR_REL_CALL |
1534 						  ARCH_LBR_IND_CALL |
1535 						  ARCH_LBR_OTHER_BRANCH,
1536 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]     = ARCH_LBR_IND_CALL,
1537 	[PERF_SAMPLE_BRANCH_COND_SHIFT]         = ARCH_LBR_JCC,
1538 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]   = ARCH_LBR_REL_CALL |
1539 						  ARCH_LBR_IND_CALL |
1540 						  ARCH_LBR_RETURN |
1541 						  ARCH_LBR_CALL_STACK,
1542 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= ARCH_LBR_IND_JMP,
1543 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= ARCH_LBR_REL_CALL,
1544 };
1545 
1546 /* core */
1547 void __init intel_pmu_lbr_init_core(void)
1548 {
1549 	x86_pmu.lbr_nr     = 4;
1550 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1551 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1552 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1553 
1554 	/*
1555 	 * SW branch filter usage:
1556 	 * - compensate for lack of HW filter
1557 	 */
1558 }
1559 
1560 /* nehalem/westmere */
1561 void __init intel_pmu_lbr_init_nhm(void)
1562 {
1563 	x86_pmu.lbr_nr     = 16;
1564 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1565 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1566 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1567 
1568 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1569 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1570 
1571 	/*
1572 	 * SW branch filter usage:
1573 	 * - workaround LBR_SEL errata (see above)
1574 	 * - support syscall, sysret capture.
1575 	 *   That requires LBR_FAR but that means far
1576 	 *   jmp need to be filtered out
1577 	 */
1578 }
1579 
1580 /* sandy bridge */
1581 void __init intel_pmu_lbr_init_snb(void)
1582 {
1583 	x86_pmu.lbr_nr	 = 16;
1584 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1585 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1586 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1587 
1588 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1589 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1590 
1591 	/*
1592 	 * SW branch filter usage:
1593 	 * - support syscall, sysret capture.
1594 	 *   That requires LBR_FAR but that means far
1595 	 *   jmp need to be filtered out
1596 	 */
1597 }
1598 
1599 static inline struct kmem_cache *
1600 create_lbr_kmem_cache(size_t size, size_t align)
1601 {
1602 	return kmem_cache_create("x86_lbr", size, align, 0, NULL);
1603 }
1604 
1605 /* haswell */
1606 void intel_pmu_lbr_init_hsw(void)
1607 {
1608 	size_t size = sizeof(struct x86_perf_task_context);
1609 
1610 	x86_pmu.lbr_nr	 = 16;
1611 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1612 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1613 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1614 
1615 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1616 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1617 
1618 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1619 
1620 	if (lbr_from_signext_quirk_needed())
1621 		static_branch_enable(&lbr_from_quirk_key);
1622 }
1623 
1624 /* skylake */
1625 __init void intel_pmu_lbr_init_skl(void)
1626 {
1627 	size_t size = sizeof(struct x86_perf_task_context);
1628 
1629 	x86_pmu.lbr_nr	 = 32;
1630 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1631 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1632 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1633 	x86_pmu.lbr_info = MSR_LBR_INFO_0;
1634 
1635 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1636 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1637 
1638 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1639 
1640 	/*
1641 	 * SW branch filter usage:
1642 	 * - support syscall, sysret capture.
1643 	 *   That requires LBR_FAR but that means far
1644 	 *   jmp need to be filtered out
1645 	 */
1646 }
1647 
1648 /* atom */
1649 void __init intel_pmu_lbr_init_atom(void)
1650 {
1651 	/*
1652 	 * only models starting at stepping 10 seems
1653 	 * to have an operational LBR which can freeze
1654 	 * on PMU interrupt
1655 	 */
1656 	if (boot_cpu_data.x86_model == 28
1657 	    && boot_cpu_data.x86_stepping < 10) {
1658 		pr_cont("LBR disabled due to erratum");
1659 		return;
1660 	}
1661 
1662 	x86_pmu.lbr_nr	   = 8;
1663 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1664 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1665 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1666 
1667 	/*
1668 	 * SW branch filter usage:
1669 	 * - compensate for lack of HW filter
1670 	 */
1671 }
1672 
1673 /* slm */
1674 void __init intel_pmu_lbr_init_slm(void)
1675 {
1676 	x86_pmu.lbr_nr	   = 8;
1677 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1678 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1679 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1680 
1681 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1682 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1683 
1684 	/*
1685 	 * SW branch filter usage:
1686 	 * - compensate for lack of HW filter
1687 	 */
1688 	pr_cont("8-deep LBR, ");
1689 }
1690 
1691 /* Knights Landing */
1692 void intel_pmu_lbr_init_knl(void)
1693 {
1694 	x86_pmu.lbr_nr	   = 8;
1695 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1696 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1697 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1698 
1699 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1700 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1701 
1702 	/* Knights Landing does have MISPREDICT bit */
1703 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1704 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1705 }
1706 
1707 /*
1708  * LBR state size is variable based on the max number of registers.
1709  * This calculates the expected state size, which should match
1710  * what the hardware enumerates for the size of XFEATURE_LBR.
1711  */
1712 static inline unsigned int get_lbr_state_size(void)
1713 {
1714 	return sizeof(struct arch_lbr_state) +
1715 	       x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1716 }
1717 
1718 static bool is_arch_lbr_xsave_available(void)
1719 {
1720 	if (!boot_cpu_has(X86_FEATURE_XSAVES))
1721 		return false;
1722 
1723 	/*
1724 	 * Check the LBR state with the corresponding software structure.
1725 	 * Disable LBR XSAVES support if the size doesn't match.
1726 	 */
1727 	if (WARN_ON(xfeature_size(XFEATURE_LBR) != get_lbr_state_size()))
1728 		return false;
1729 
1730 	return true;
1731 }
1732 
1733 void __init intel_pmu_arch_lbr_init(void)
1734 {
1735 	struct pmu *pmu = x86_get_pmu(smp_processor_id());
1736 	union cpuid28_eax eax;
1737 	union cpuid28_ebx ebx;
1738 	union cpuid28_ecx ecx;
1739 	unsigned int unused_edx;
1740 	bool arch_lbr_xsave;
1741 	size_t size;
1742 	u64 lbr_nr;
1743 
1744 	/* Arch LBR Capabilities */
1745 	cpuid(28, &eax.full, &ebx.full, &ecx.full, &unused_edx);
1746 
1747 	lbr_nr = fls(eax.split.lbr_depth_mask) * 8;
1748 	if (!lbr_nr)
1749 		goto clear_arch_lbr;
1750 
1751 	/* Apply the max depth of Arch LBR */
1752 	if (wrmsrl_safe(MSR_ARCH_LBR_DEPTH, lbr_nr))
1753 		goto clear_arch_lbr;
1754 
1755 	x86_pmu.lbr_depth_mask = eax.split.lbr_depth_mask;
1756 	x86_pmu.lbr_deep_c_reset = eax.split.lbr_deep_c_reset;
1757 	x86_pmu.lbr_lip = eax.split.lbr_lip;
1758 	x86_pmu.lbr_cpl = ebx.split.lbr_cpl;
1759 	x86_pmu.lbr_filter = ebx.split.lbr_filter;
1760 	x86_pmu.lbr_call_stack = ebx.split.lbr_call_stack;
1761 	x86_pmu.lbr_mispred = ecx.split.lbr_mispred;
1762 	x86_pmu.lbr_timed_lbr = ecx.split.lbr_timed_lbr;
1763 	x86_pmu.lbr_br_type = ecx.split.lbr_br_type;
1764 	x86_pmu.lbr_nr = lbr_nr;
1765 
1766 
1767 	arch_lbr_xsave = is_arch_lbr_xsave_available();
1768 	if (arch_lbr_xsave) {
1769 		size = sizeof(struct x86_perf_task_context_arch_lbr_xsave) +
1770 		       get_lbr_state_size();
1771 		pmu->task_ctx_cache = create_lbr_kmem_cache(size,
1772 							    XSAVE_ALIGNMENT);
1773 	}
1774 
1775 	if (!pmu->task_ctx_cache) {
1776 		arch_lbr_xsave = false;
1777 
1778 		size = sizeof(struct x86_perf_task_context_arch_lbr) +
1779 		       lbr_nr * sizeof(struct lbr_entry);
1780 		pmu->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1781 	}
1782 
1783 	x86_pmu.lbr_from = MSR_ARCH_LBR_FROM_0;
1784 	x86_pmu.lbr_to = MSR_ARCH_LBR_TO_0;
1785 	x86_pmu.lbr_info = MSR_ARCH_LBR_INFO_0;
1786 
1787 	/* LBR callstack requires both CPL and Branch Filtering support */
1788 	if (!x86_pmu.lbr_cpl ||
1789 	    !x86_pmu.lbr_filter ||
1790 	    !x86_pmu.lbr_call_stack)
1791 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT] = LBR_NOT_SUPP;
1792 
1793 	if (!x86_pmu.lbr_cpl) {
1794 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_USER_SHIFT] = LBR_NOT_SUPP;
1795 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_KERNEL_SHIFT] = LBR_NOT_SUPP;
1796 	} else if (!x86_pmu.lbr_filter) {
1797 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_SHIFT] = LBR_NOT_SUPP;
1798 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT] = LBR_NOT_SUPP;
1799 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] = LBR_NOT_SUPP;
1800 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_NOT_SUPP;
1801 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_COND_SHIFT] = LBR_NOT_SUPP;
1802 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_NOT_SUPP;
1803 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_SHIFT] = LBR_NOT_SUPP;
1804 	}
1805 
1806 	x86_pmu.lbr_ctl_mask = ARCH_LBR_CTL_MASK;
1807 	x86_pmu.lbr_ctl_map  = arch_lbr_ctl_map;
1808 
1809 	if (!x86_pmu.lbr_cpl && !x86_pmu.lbr_filter)
1810 		x86_pmu.lbr_ctl_map = NULL;
1811 
1812 	x86_pmu.lbr_reset = intel_pmu_arch_lbr_reset;
1813 	if (arch_lbr_xsave) {
1814 		x86_pmu.lbr_save = intel_pmu_arch_lbr_xsaves;
1815 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_xrstors;
1816 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read_xsave;
1817 		pr_cont("XSAVE ");
1818 	} else {
1819 		x86_pmu.lbr_save = intel_pmu_arch_lbr_save;
1820 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_restore;
1821 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read;
1822 	}
1823 
1824 	pr_cont("Architectural LBR, ");
1825 
1826 	return;
1827 
1828 clear_arch_lbr:
1829 	clear_cpu_cap(&boot_cpu_data, X86_FEATURE_ARCH_LBR);
1830 }
1831 
1832 /**
1833  * x86_perf_get_lbr - get the LBR records information
1834  *
1835  * @lbr: the caller's memory to store the LBR records information
1836  *
1837  * Returns: 0 indicates the LBR info has been successfully obtained
1838  */
1839 int x86_perf_get_lbr(struct x86_pmu_lbr *lbr)
1840 {
1841 	int lbr_fmt = x86_pmu.intel_cap.lbr_format;
1842 
1843 	lbr->nr = x86_pmu.lbr_nr;
1844 	lbr->from = x86_pmu.lbr_from;
1845 	lbr->to = x86_pmu.lbr_to;
1846 	lbr->info = (lbr_fmt == LBR_FORMAT_INFO) ? x86_pmu.lbr_info : 0;
1847 
1848 	return 0;
1849 }
1850 EXPORT_SYMBOL_GPL(x86_perf_get_lbr);
1851 
1852 struct event_constraint vlbr_constraint =
1853 	__EVENT_CONSTRAINT(INTEL_FIXED_VLBR_EVENT, (1ULL << INTEL_PMC_IDX_FIXED_VLBR),
1854 			  FIXED_EVENT_FLAGS, 1, 0, PERF_X86_EVENT_LBR_SELECT);
1855