xref: /openbmc/linux/arch/x86/events/intel/lbr.c (revision f16fe2d3)
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 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) && cpuc->lbr_select)
269 		wrmsrl(MSR_LBR_SELECT, 0);
270 }
271 
272 /*
273  * TOS = most recently recorded branch
274  */
275 static inline u64 intel_pmu_lbr_tos(void)
276 {
277 	u64 tos;
278 
279 	rdmsrl(x86_pmu.lbr_tos, tos);
280 	return tos;
281 }
282 
283 enum {
284 	LBR_NONE,
285 	LBR_VALID,
286 };
287 
288 /*
289  * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
290  * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported, but when
291  * TSX is not supported they have no consistent behavior:
292  *
293  *   - For wrmsr(), bits 61:62 are considered part of the sign extension.
294  *   - For HW updates (branch captures) bits 61:62 are always OFF and are not
295  *     part of the sign extension.
296  *
297  * Therefore, if:
298  *
299  *   1) LBR has TSX format
300  *   2) CPU has no TSX support enabled
301  *
302  * ... then any value passed to wrmsr() must be sign extended to 63 bits and any
303  * value from rdmsr() must be converted to have a 61 bits sign extension,
304  * ignoring the TSX flags.
305  */
306 static inline bool lbr_from_signext_quirk_needed(void)
307 {
308 	int lbr_format = x86_pmu.intel_cap.lbr_format;
309 	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
310 			   boot_cpu_has(X86_FEATURE_RTM);
311 
312 	return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
313 }
314 
315 static DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
316 
317 /* If quirk is enabled, ensure sign extension is 63 bits: */
318 inline u64 lbr_from_signext_quirk_wr(u64 val)
319 {
320 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
321 		/*
322 		 * Sign extend into bits 61:62 while preserving bit 63.
323 		 *
324 		 * Quirk is enabled when TSX is disabled. Therefore TSX bits
325 		 * in val are always OFF and must be changed to be sign
326 		 * extension bits. Since bits 59:60 are guaranteed to be
327 		 * part of the sign extension bits, we can just copy them
328 		 * to 61:62.
329 		 */
330 		val |= (LBR_FROM_SIGNEXT_2MSB & val) << 2;
331 	}
332 	return val;
333 }
334 
335 /*
336  * If quirk is needed, ensure sign extension is 61 bits:
337  */
338 static u64 lbr_from_signext_quirk_rd(u64 val)
339 {
340 	if (static_branch_unlikely(&lbr_from_quirk_key)) {
341 		/*
342 		 * Quirk is on when TSX is not enabled. Therefore TSX
343 		 * flags must be read as OFF.
344 		 */
345 		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);
346 	}
347 	return val;
348 }
349 
350 static __always_inline void wrlbr_from(unsigned int idx, u64 val)
351 {
352 	val = lbr_from_signext_quirk_wr(val);
353 	wrmsrl(x86_pmu.lbr_from + idx, val);
354 }
355 
356 static __always_inline void wrlbr_to(unsigned int idx, u64 val)
357 {
358 	wrmsrl(x86_pmu.lbr_to + idx, val);
359 }
360 
361 static __always_inline void wrlbr_info(unsigned int idx, u64 val)
362 {
363 	wrmsrl(x86_pmu.lbr_info + idx, val);
364 }
365 
366 static __always_inline u64 rdlbr_from(unsigned int idx, struct lbr_entry *lbr)
367 {
368 	u64 val;
369 
370 	if (lbr)
371 		return lbr->from;
372 
373 	rdmsrl(x86_pmu.lbr_from + idx, val);
374 
375 	return lbr_from_signext_quirk_rd(val);
376 }
377 
378 static __always_inline u64 rdlbr_to(unsigned int idx, struct lbr_entry *lbr)
379 {
380 	u64 val;
381 
382 	if (lbr)
383 		return lbr->to;
384 
385 	rdmsrl(x86_pmu.lbr_to + idx, val);
386 
387 	return val;
388 }
389 
390 static __always_inline u64 rdlbr_info(unsigned int idx, struct lbr_entry *lbr)
391 {
392 	u64 val;
393 
394 	if (lbr)
395 		return lbr->info;
396 
397 	rdmsrl(x86_pmu.lbr_info + idx, val);
398 
399 	return val;
400 }
401 
402 static inline void
403 wrlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
404 {
405 	wrlbr_from(idx, lbr->from);
406 	wrlbr_to(idx, lbr->to);
407 	if (need_info)
408 		wrlbr_info(idx, lbr->info);
409 }
410 
411 static inline bool
412 rdlbr_all(struct lbr_entry *lbr, unsigned int idx, bool need_info)
413 {
414 	u64 from = rdlbr_from(idx, NULL);
415 
416 	/* Don't read invalid entry */
417 	if (!from)
418 		return false;
419 
420 	lbr->from = from;
421 	lbr->to = rdlbr_to(idx, NULL);
422 	if (need_info)
423 		lbr->info = rdlbr_info(idx, NULL);
424 
425 	return true;
426 }
427 
428 void intel_pmu_lbr_restore(void *ctx)
429 {
430 	bool need_info = x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO;
431 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
432 	struct x86_perf_task_context *task_ctx = ctx;
433 	int i;
434 	unsigned lbr_idx, mask;
435 	u64 tos = task_ctx->tos;
436 
437 	mask = x86_pmu.lbr_nr - 1;
438 	for (i = 0; i < task_ctx->valid_lbrs; i++) {
439 		lbr_idx = (tos - i) & mask;
440 		wrlbr_all(&task_ctx->lbr[i], lbr_idx, need_info);
441 	}
442 
443 	for (; i < x86_pmu.lbr_nr; i++) {
444 		lbr_idx = (tos - i) & mask;
445 		wrlbr_from(lbr_idx, 0);
446 		wrlbr_to(lbr_idx, 0);
447 		if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO)
448 			wrlbr_info(lbr_idx, 0);
449 	}
450 
451 	wrmsrl(x86_pmu.lbr_tos, tos);
452 
453 	if (cpuc->lbr_select)
454 		wrmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
455 }
456 
457 static void intel_pmu_arch_lbr_restore(void *ctx)
458 {
459 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
460 	struct lbr_entry *entries = task_ctx->entries;
461 	int i;
462 
463 	/* Fast reset the LBRs before restore if the call stack is not full. */
464 	if (!entries[x86_pmu.lbr_nr - 1].from)
465 		intel_pmu_arch_lbr_reset();
466 
467 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
468 		if (!entries[i].from)
469 			break;
470 		wrlbr_all(&entries[i], i, true);
471 	}
472 }
473 
474 /*
475  * Restore the Architecture LBR state from the xsave area in the perf
476  * context data for the task via the XRSTORS instruction.
477  */
478 static void intel_pmu_arch_lbr_xrstors(void *ctx)
479 {
480 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
481 
482 	xrstors(&task_ctx->xsave, XFEATURE_MASK_LBR);
483 }
484 
485 static __always_inline bool lbr_is_reset_in_cstate(void *ctx)
486 {
487 	if (static_cpu_has(X86_FEATURE_ARCH_LBR))
488 		return x86_pmu.lbr_deep_c_reset && !rdlbr_from(0, NULL);
489 
490 	return !rdlbr_from(((struct x86_perf_task_context *)ctx)->tos, NULL);
491 }
492 
493 static void __intel_pmu_lbr_restore(void *ctx)
494 {
495 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
496 
497 	if (task_context_opt(ctx)->lbr_callstack_users == 0 ||
498 	    task_context_opt(ctx)->lbr_stack_state == LBR_NONE) {
499 		intel_pmu_lbr_reset();
500 		return;
501 	}
502 
503 	/*
504 	 * Does not restore the LBR registers, if
505 	 * - No one else touched them, and
506 	 * - Was not cleared in Cstate
507 	 */
508 	if ((ctx == cpuc->last_task_ctx) &&
509 	    (task_context_opt(ctx)->log_id == cpuc->last_log_id) &&
510 	    !lbr_is_reset_in_cstate(ctx)) {
511 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
512 		return;
513 	}
514 
515 	x86_pmu.lbr_restore(ctx);
516 
517 	task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
518 }
519 
520 void intel_pmu_lbr_save(void *ctx)
521 {
522 	bool need_info = x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO;
523 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
524 	struct x86_perf_task_context *task_ctx = ctx;
525 	unsigned lbr_idx, mask;
526 	u64 tos;
527 	int i;
528 
529 	mask = x86_pmu.lbr_nr - 1;
530 	tos = intel_pmu_lbr_tos();
531 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
532 		lbr_idx = (tos - i) & mask;
533 		if (!rdlbr_all(&task_ctx->lbr[i], lbr_idx, need_info))
534 			break;
535 	}
536 	task_ctx->valid_lbrs = i;
537 	task_ctx->tos = tos;
538 
539 	if (cpuc->lbr_select)
540 		rdmsrl(MSR_LBR_SELECT, task_ctx->lbr_sel);
541 }
542 
543 static void intel_pmu_arch_lbr_save(void *ctx)
544 {
545 	struct x86_perf_task_context_arch_lbr *task_ctx = ctx;
546 	struct lbr_entry *entries = task_ctx->entries;
547 	int i;
548 
549 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
550 		if (!rdlbr_all(&entries[i], i, true))
551 			break;
552 	}
553 
554 	/* LBR call stack is not full. Reset is required in restore. */
555 	if (i < x86_pmu.lbr_nr)
556 		entries[x86_pmu.lbr_nr - 1].from = 0;
557 }
558 
559 /*
560  * Save the Architecture LBR state to the xsave area in the perf
561  * context data for the task via the XSAVES instruction.
562  */
563 static void intel_pmu_arch_lbr_xsaves(void *ctx)
564 {
565 	struct x86_perf_task_context_arch_lbr_xsave *task_ctx = ctx;
566 
567 	xsaves(&task_ctx->xsave, XFEATURE_MASK_LBR);
568 }
569 
570 static void __intel_pmu_lbr_save(void *ctx)
571 {
572 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
573 
574 	if (task_context_opt(ctx)->lbr_callstack_users == 0) {
575 		task_context_opt(ctx)->lbr_stack_state = LBR_NONE;
576 		return;
577 	}
578 
579 	x86_pmu.lbr_save(ctx);
580 
581 	task_context_opt(ctx)->lbr_stack_state = LBR_VALID;
582 
583 	cpuc->last_task_ctx = ctx;
584 	cpuc->last_log_id = ++task_context_opt(ctx)->log_id;
585 }
586 
587 void intel_pmu_lbr_swap_task_ctx(struct perf_event_context *prev,
588 				 struct perf_event_context *next)
589 {
590 	void *prev_ctx_data, *next_ctx_data;
591 
592 	swap(prev->task_ctx_data, next->task_ctx_data);
593 
594 	/*
595 	 * Architecture specific synchronization makes sense in
596 	 * case both prev->task_ctx_data and next->task_ctx_data
597 	 * pointers are allocated.
598 	 */
599 
600 	prev_ctx_data = next->task_ctx_data;
601 	next_ctx_data = prev->task_ctx_data;
602 
603 	if (!prev_ctx_data || !next_ctx_data)
604 		return;
605 
606 	swap(task_context_opt(prev_ctx_data)->lbr_callstack_users,
607 	     task_context_opt(next_ctx_data)->lbr_callstack_users);
608 }
609 
610 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
611 {
612 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
613 	void *task_ctx;
614 
615 	if (!cpuc->lbr_users)
616 		return;
617 
618 	/*
619 	 * If LBR callstack feature is enabled and the stack was saved when
620 	 * the task was scheduled out, restore the stack. Otherwise flush
621 	 * the LBR stack.
622 	 */
623 	task_ctx = ctx ? ctx->task_ctx_data : NULL;
624 	if (task_ctx) {
625 		if (sched_in)
626 			__intel_pmu_lbr_restore(task_ctx);
627 		else
628 			__intel_pmu_lbr_save(task_ctx);
629 		return;
630 	}
631 
632 	/*
633 	 * Since a context switch can flip the address space and LBR entries
634 	 * are not tagged with an identifier, we need to wipe the LBR, even for
635 	 * per-cpu events. You simply cannot resolve the branches from the old
636 	 * address space.
637 	 */
638 	if (sched_in)
639 		intel_pmu_lbr_reset();
640 }
641 
642 static inline bool branch_user_callstack(unsigned br_sel)
643 {
644 	return (br_sel & X86_BR_USER) && (br_sel & X86_BR_CALL_STACK);
645 }
646 
647 void intel_pmu_lbr_add(struct perf_event *event)
648 {
649 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
650 
651 	if (!x86_pmu.lbr_nr)
652 		return;
653 
654 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
655 		cpuc->lbr_select = 1;
656 
657 	cpuc->br_sel = event->hw.branch_reg.reg;
658 
659 	if (branch_user_callstack(cpuc->br_sel) && event->ctx->task_ctx_data)
660 		task_context_opt(event->ctx->task_ctx_data)->lbr_callstack_users++;
661 
662 	/*
663 	 * Request pmu::sched_task() callback, which will fire inside the
664 	 * regular perf event scheduling, so that call will:
665 	 *
666 	 *  - restore or wipe; when LBR-callstack,
667 	 *  - wipe; otherwise,
668 	 *
669 	 * when this is from __perf_event_task_sched_in().
670 	 *
671 	 * However, if this is from perf_install_in_context(), no such callback
672 	 * will follow and we'll need to reset the LBR here if this is the
673 	 * first LBR event.
674 	 *
675 	 * The problem is, we cannot tell these cases apart... but we can
676 	 * exclude the biggest chunk of cases by looking at
677 	 * event->total_time_running. An event that has accrued runtime cannot
678 	 * be 'new'. Conversely, a new event can get installed through the
679 	 * context switch path for the first time.
680 	 */
681 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
682 		cpuc->lbr_pebs_users++;
683 	perf_sched_cb_inc(event->ctx->pmu);
684 	if (!cpuc->lbr_users++ && !event->total_time_running)
685 		intel_pmu_lbr_reset();
686 }
687 
688 void release_lbr_buffers(void)
689 {
690 	struct kmem_cache *kmem_cache;
691 	struct cpu_hw_events *cpuc;
692 	int cpu;
693 
694 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
695 		return;
696 
697 	for_each_possible_cpu(cpu) {
698 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
699 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
700 		if (kmem_cache && cpuc->lbr_xsave) {
701 			kmem_cache_free(kmem_cache, cpuc->lbr_xsave);
702 			cpuc->lbr_xsave = NULL;
703 		}
704 	}
705 }
706 
707 void reserve_lbr_buffers(void)
708 {
709 	struct kmem_cache *kmem_cache;
710 	struct cpu_hw_events *cpuc;
711 	int cpu;
712 
713 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR))
714 		return;
715 
716 	for_each_possible_cpu(cpu) {
717 		cpuc = per_cpu_ptr(&cpu_hw_events, cpu);
718 		kmem_cache = x86_get_pmu(cpu)->task_ctx_cache;
719 		if (!kmem_cache || cpuc->lbr_xsave)
720 			continue;
721 
722 		cpuc->lbr_xsave = kmem_cache_alloc_node(kmem_cache,
723 							GFP_KERNEL | __GFP_ZERO,
724 							cpu_to_node(cpu));
725 	}
726 }
727 
728 void intel_pmu_lbr_del(struct perf_event *event)
729 {
730 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
731 
732 	if (!x86_pmu.lbr_nr)
733 		return;
734 
735 	if (branch_user_callstack(cpuc->br_sel) &&
736 	    event->ctx->task_ctx_data)
737 		task_context_opt(event->ctx->task_ctx_data)->lbr_callstack_users--;
738 
739 	if (event->hw.flags & PERF_X86_EVENT_LBR_SELECT)
740 		cpuc->lbr_select = 0;
741 
742 	if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
743 		cpuc->lbr_pebs_users--;
744 	cpuc->lbr_users--;
745 	WARN_ON_ONCE(cpuc->lbr_users < 0);
746 	WARN_ON_ONCE(cpuc->lbr_pebs_users < 0);
747 	perf_sched_cb_dec(event->ctx->pmu);
748 }
749 
750 static inline bool vlbr_exclude_host(void)
751 {
752 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
753 
754 	return test_bit(INTEL_PMC_IDX_FIXED_VLBR,
755 		(unsigned long *)&cpuc->intel_ctrl_guest_mask);
756 }
757 
758 void intel_pmu_lbr_enable_all(bool pmi)
759 {
760 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
761 
762 	if (cpuc->lbr_users && !vlbr_exclude_host())
763 		__intel_pmu_lbr_enable(pmi);
764 }
765 
766 void intel_pmu_lbr_disable_all(void)
767 {
768 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
769 
770 	if (cpuc->lbr_users && !vlbr_exclude_host()) {
771 		if (static_cpu_has(X86_FEATURE_ARCH_LBR))
772 			return __intel_pmu_arch_lbr_disable();
773 
774 		__intel_pmu_lbr_disable();
775 	}
776 }
777 
778 void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
779 {
780 	unsigned long mask = x86_pmu.lbr_nr - 1;
781 	u64 tos = intel_pmu_lbr_tos();
782 	int i;
783 
784 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
785 		unsigned long lbr_idx = (tos - i) & mask;
786 		union {
787 			struct {
788 				u32 from;
789 				u32 to;
790 			};
791 			u64     lbr;
792 		} msr_lastbranch;
793 
794 		rdmsrl(x86_pmu.lbr_from + lbr_idx, msr_lastbranch.lbr);
795 
796 		cpuc->lbr_entries[i].from	= msr_lastbranch.from;
797 		cpuc->lbr_entries[i].to		= msr_lastbranch.to;
798 		cpuc->lbr_entries[i].mispred	= 0;
799 		cpuc->lbr_entries[i].predicted	= 0;
800 		cpuc->lbr_entries[i].in_tx	= 0;
801 		cpuc->lbr_entries[i].abort	= 0;
802 		cpuc->lbr_entries[i].cycles	= 0;
803 		cpuc->lbr_entries[i].type	= 0;
804 		cpuc->lbr_entries[i].reserved	= 0;
805 	}
806 	cpuc->lbr_stack.nr = i;
807 	cpuc->lbr_stack.hw_idx = tos;
808 }
809 
810 /*
811  * Due to lack of segmentation in Linux the effective address (offset)
812  * is the same as the linear address, allowing us to merge the LIP and EIP
813  * LBR formats.
814  */
815 void intel_pmu_lbr_read_64(struct cpu_hw_events *cpuc)
816 {
817 	bool need_info = false, call_stack = false;
818 	unsigned long mask = x86_pmu.lbr_nr - 1;
819 	int lbr_format = x86_pmu.intel_cap.lbr_format;
820 	u64 tos = intel_pmu_lbr_tos();
821 	int i;
822 	int out = 0;
823 	int num = x86_pmu.lbr_nr;
824 
825 	if (cpuc->lbr_sel) {
826 		need_info = !(cpuc->lbr_sel->config & LBR_NO_INFO);
827 		if (cpuc->lbr_sel->config & LBR_CALL_STACK)
828 			call_stack = true;
829 	}
830 
831 	for (i = 0; i < num; i++) {
832 		unsigned long lbr_idx = (tos - i) & mask;
833 		u64 from, to, mis = 0, pred = 0, in_tx = 0, abort = 0;
834 		int skip = 0;
835 		u16 cycles = 0;
836 		int lbr_flags = lbr_desc[lbr_format];
837 
838 		from = rdlbr_from(lbr_idx, NULL);
839 		to   = rdlbr_to(lbr_idx, NULL);
840 
841 		/*
842 		 * Read LBR call stack entries
843 		 * until invalid entry (0s) is detected.
844 		 */
845 		if (call_stack && !from)
846 			break;
847 
848 		if (lbr_format == LBR_FORMAT_INFO && need_info) {
849 			u64 info;
850 
851 			info = rdlbr_info(lbr_idx, NULL);
852 			mis = !!(info & LBR_INFO_MISPRED);
853 			pred = !mis;
854 			in_tx = !!(info & LBR_INFO_IN_TX);
855 			abort = !!(info & LBR_INFO_ABORT);
856 			cycles = (info & LBR_INFO_CYCLES);
857 		}
858 
859 		if (lbr_format == LBR_FORMAT_TIME) {
860 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
861 			pred = !mis;
862 			skip = 1;
863 			cycles = ((to >> 48) & LBR_INFO_CYCLES);
864 
865 			to = (u64)((((s64)to) << 16) >> 16);
866 		}
867 
868 		if (lbr_flags & LBR_EIP_FLAGS) {
869 			mis = !!(from & LBR_FROM_FLAG_MISPRED);
870 			pred = !mis;
871 			skip = 1;
872 		}
873 		if (lbr_flags & LBR_TSX) {
874 			in_tx = !!(from & LBR_FROM_FLAG_IN_TX);
875 			abort = !!(from & LBR_FROM_FLAG_ABORT);
876 			skip = 3;
877 		}
878 		from = (u64)((((s64)from) << skip) >> skip);
879 
880 		/*
881 		 * Some CPUs report duplicated abort records,
882 		 * with the second entry not having an abort bit set.
883 		 * Skip them here. This loop runs backwards,
884 		 * so we need to undo the previous record.
885 		 * If the abort just happened outside the window
886 		 * the extra entry cannot be removed.
887 		 */
888 		if (abort && x86_pmu.lbr_double_abort && out > 0)
889 			out--;
890 
891 		cpuc->lbr_entries[out].from	 = from;
892 		cpuc->lbr_entries[out].to	 = to;
893 		cpuc->lbr_entries[out].mispred	 = mis;
894 		cpuc->lbr_entries[out].predicted = pred;
895 		cpuc->lbr_entries[out].in_tx	 = in_tx;
896 		cpuc->lbr_entries[out].abort	 = abort;
897 		cpuc->lbr_entries[out].cycles	 = cycles;
898 		cpuc->lbr_entries[out].type	 = 0;
899 		cpuc->lbr_entries[out].reserved	 = 0;
900 		out++;
901 	}
902 	cpuc->lbr_stack.nr = out;
903 	cpuc->lbr_stack.hw_idx = tos;
904 }
905 
906 static __always_inline int get_lbr_br_type(u64 info)
907 {
908 	if (!static_cpu_has(X86_FEATURE_ARCH_LBR) || !x86_pmu.lbr_br_type)
909 		return 0;
910 
911 	return (info & LBR_INFO_BR_TYPE) >> LBR_INFO_BR_TYPE_OFFSET;
912 }
913 
914 static __always_inline bool get_lbr_mispred(u64 info)
915 {
916 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) && !x86_pmu.lbr_mispred)
917 		return 0;
918 
919 	return !!(info & LBR_INFO_MISPRED);
920 }
921 
922 static __always_inline bool get_lbr_predicted(u64 info)
923 {
924 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) && !x86_pmu.lbr_mispred)
925 		return 0;
926 
927 	return !(info & LBR_INFO_MISPRED);
928 }
929 
930 static __always_inline u16 get_lbr_cycles(u64 info)
931 {
932 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
933 	    !(x86_pmu.lbr_timed_lbr && info & LBR_INFO_CYC_CNT_VALID))
934 		return 0;
935 
936 	return info & LBR_INFO_CYCLES;
937 }
938 
939 static void intel_pmu_store_lbr(struct cpu_hw_events *cpuc,
940 				struct lbr_entry *entries)
941 {
942 	struct perf_branch_entry *e;
943 	struct lbr_entry *lbr;
944 	u64 from, to, info;
945 	int i;
946 
947 	for (i = 0; i < x86_pmu.lbr_nr; i++) {
948 		lbr = entries ? &entries[i] : NULL;
949 		e = &cpuc->lbr_entries[i];
950 
951 		from = rdlbr_from(i, lbr);
952 		/*
953 		 * Read LBR entries until invalid entry (0s) is detected.
954 		 */
955 		if (!from)
956 			break;
957 
958 		to = rdlbr_to(i, lbr);
959 		info = rdlbr_info(i, lbr);
960 
961 		e->from		= from;
962 		e->to		= to;
963 		e->mispred	= get_lbr_mispred(info);
964 		e->predicted	= get_lbr_predicted(info);
965 		e->in_tx	= !!(info & LBR_INFO_IN_TX);
966 		e->abort	= !!(info & LBR_INFO_ABORT);
967 		e->cycles	= get_lbr_cycles(info);
968 		e->type		= get_lbr_br_type(info);
969 		e->reserved	= 0;
970 	}
971 
972 	cpuc->lbr_stack.nr = i;
973 }
974 
975 static void intel_pmu_arch_lbr_read(struct cpu_hw_events *cpuc)
976 {
977 	intel_pmu_store_lbr(cpuc, NULL);
978 }
979 
980 static void intel_pmu_arch_lbr_read_xsave(struct cpu_hw_events *cpuc)
981 {
982 	struct x86_perf_task_context_arch_lbr_xsave *xsave = cpuc->lbr_xsave;
983 
984 	if (!xsave) {
985 		intel_pmu_store_lbr(cpuc, NULL);
986 		return;
987 	}
988 	xsaves(&xsave->xsave, XFEATURE_MASK_LBR);
989 
990 	intel_pmu_store_lbr(cpuc, xsave->lbr.entries);
991 }
992 
993 void intel_pmu_lbr_read(void)
994 {
995 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
996 
997 	/*
998 	 * Don't read when all LBRs users are using adaptive PEBS.
999 	 *
1000 	 * This could be smarter and actually check the event,
1001 	 * but this simple approach seems to work for now.
1002 	 */
1003 	if (!cpuc->lbr_users || vlbr_exclude_host() ||
1004 	    cpuc->lbr_users == cpuc->lbr_pebs_users)
1005 		return;
1006 
1007 	x86_pmu.lbr_read(cpuc);
1008 
1009 	intel_pmu_lbr_filter(cpuc);
1010 }
1011 
1012 /*
1013  * SW filter is used:
1014  * - in case there is no HW filter
1015  * - in case the HW filter has errata or limitations
1016  */
1017 static int intel_pmu_setup_sw_lbr_filter(struct perf_event *event)
1018 {
1019 	u64 br_type = event->attr.branch_sample_type;
1020 	int mask = 0;
1021 
1022 	if (br_type & PERF_SAMPLE_BRANCH_USER)
1023 		mask |= X86_BR_USER;
1024 
1025 	if (br_type & PERF_SAMPLE_BRANCH_KERNEL)
1026 		mask |= X86_BR_KERNEL;
1027 
1028 	/* we ignore BRANCH_HV here */
1029 
1030 	if (br_type & PERF_SAMPLE_BRANCH_ANY)
1031 		mask |= X86_BR_ANY;
1032 
1033 	if (br_type & PERF_SAMPLE_BRANCH_ANY_CALL)
1034 		mask |= X86_BR_ANY_CALL;
1035 
1036 	if (br_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
1037 		mask |= X86_BR_RET | X86_BR_IRET | X86_BR_SYSRET;
1038 
1039 	if (br_type & PERF_SAMPLE_BRANCH_IND_CALL)
1040 		mask |= X86_BR_IND_CALL;
1041 
1042 	if (br_type & PERF_SAMPLE_BRANCH_ABORT_TX)
1043 		mask |= X86_BR_ABORT;
1044 
1045 	if (br_type & PERF_SAMPLE_BRANCH_IN_TX)
1046 		mask |= X86_BR_IN_TX;
1047 
1048 	if (br_type & PERF_SAMPLE_BRANCH_NO_TX)
1049 		mask |= X86_BR_NO_TX;
1050 
1051 	if (br_type & PERF_SAMPLE_BRANCH_COND)
1052 		mask |= X86_BR_JCC;
1053 
1054 	if (br_type & PERF_SAMPLE_BRANCH_CALL_STACK) {
1055 		if (!x86_pmu_has_lbr_callstack())
1056 			return -EOPNOTSUPP;
1057 		if (mask & ~(X86_BR_USER | X86_BR_KERNEL))
1058 			return -EINVAL;
1059 		mask |= X86_BR_CALL | X86_BR_IND_CALL | X86_BR_RET |
1060 			X86_BR_CALL_STACK;
1061 	}
1062 
1063 	if (br_type & PERF_SAMPLE_BRANCH_IND_JUMP)
1064 		mask |= X86_BR_IND_JMP;
1065 
1066 	if (br_type & PERF_SAMPLE_BRANCH_CALL)
1067 		mask |= X86_BR_CALL | X86_BR_ZERO_CALL;
1068 
1069 	if (br_type & PERF_SAMPLE_BRANCH_TYPE_SAVE)
1070 		mask |= X86_BR_TYPE_SAVE;
1071 
1072 	/*
1073 	 * stash actual user request into reg, it may
1074 	 * be used by fixup code for some CPU
1075 	 */
1076 	event->hw.branch_reg.reg = mask;
1077 	return 0;
1078 }
1079 
1080 /*
1081  * setup the HW LBR filter
1082  * Used only when available, may not be enough to disambiguate
1083  * all branches, may need the help of the SW filter
1084  */
1085 static int intel_pmu_setup_hw_lbr_filter(struct perf_event *event)
1086 {
1087 	struct hw_perf_event_extra *reg;
1088 	u64 br_type = event->attr.branch_sample_type;
1089 	u64 mask = 0, v;
1090 	int i;
1091 
1092 	for (i = 0; i < PERF_SAMPLE_BRANCH_MAX_SHIFT; i++) {
1093 		if (!(br_type & (1ULL << i)))
1094 			continue;
1095 
1096 		v = x86_pmu.lbr_sel_map[i];
1097 		if (v == LBR_NOT_SUPP)
1098 			return -EOPNOTSUPP;
1099 
1100 		if (v != LBR_IGN)
1101 			mask |= v;
1102 	}
1103 
1104 	reg = &event->hw.branch_reg;
1105 	reg->idx = EXTRA_REG_LBR;
1106 
1107 	if (static_cpu_has(X86_FEATURE_ARCH_LBR)) {
1108 		reg->config = mask;
1109 		return 0;
1110 	}
1111 
1112 	/*
1113 	 * The first 9 bits (LBR_SEL_MASK) in LBR_SELECT operate
1114 	 * in suppress mode. So LBR_SELECT should be set to
1115 	 * (~mask & LBR_SEL_MASK) | (mask & ~LBR_SEL_MASK)
1116 	 * But the 10th bit LBR_CALL_STACK does not operate
1117 	 * in suppress mode.
1118 	 */
1119 	reg->config = mask ^ (x86_pmu.lbr_sel_mask & ~LBR_CALL_STACK);
1120 
1121 	if ((br_type & PERF_SAMPLE_BRANCH_NO_CYCLES) &&
1122 	    (br_type & PERF_SAMPLE_BRANCH_NO_FLAGS) &&
1123 	    (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_INFO))
1124 		reg->config |= LBR_NO_INFO;
1125 
1126 	return 0;
1127 }
1128 
1129 int intel_pmu_setup_lbr_filter(struct perf_event *event)
1130 {
1131 	int ret = 0;
1132 
1133 	/*
1134 	 * no LBR on this PMU
1135 	 */
1136 	if (!x86_pmu.lbr_nr)
1137 		return -EOPNOTSUPP;
1138 
1139 	/*
1140 	 * setup SW LBR filter
1141 	 */
1142 	ret = intel_pmu_setup_sw_lbr_filter(event);
1143 	if (ret)
1144 		return ret;
1145 
1146 	/*
1147 	 * setup HW LBR filter, if any
1148 	 */
1149 	if (x86_pmu.lbr_sel_map)
1150 		ret = intel_pmu_setup_hw_lbr_filter(event);
1151 
1152 	return ret;
1153 }
1154 
1155 /*
1156  * return the type of control flow change at address "from"
1157  * instruction is not necessarily a branch (in case of interrupt).
1158  *
1159  * The branch type returned also includes the priv level of the
1160  * target of the control flow change (X86_BR_USER, X86_BR_KERNEL).
1161  *
1162  * If a branch type is unknown OR the instruction cannot be
1163  * decoded (e.g., text page not present), then X86_BR_NONE is
1164  * returned.
1165  */
1166 static int branch_type(unsigned long from, unsigned long to, int abort)
1167 {
1168 	struct insn insn;
1169 	void *addr;
1170 	int bytes_read, bytes_left;
1171 	int ret = X86_BR_NONE;
1172 	int ext, to_plm, from_plm;
1173 	u8 buf[MAX_INSN_SIZE];
1174 	int is64 = 0;
1175 
1176 	to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
1177 	from_plm = kernel_ip(from) ? X86_BR_KERNEL : X86_BR_USER;
1178 
1179 	/*
1180 	 * maybe zero if lbr did not fill up after a reset by the time
1181 	 * we get a PMU interrupt
1182 	 */
1183 	if (from == 0 || to == 0)
1184 		return X86_BR_NONE;
1185 
1186 	if (abort)
1187 		return X86_BR_ABORT | to_plm;
1188 
1189 	if (from_plm == X86_BR_USER) {
1190 		/*
1191 		 * can happen if measuring at the user level only
1192 		 * and we interrupt in a kernel thread, e.g., idle.
1193 		 */
1194 		if (!current->mm)
1195 			return X86_BR_NONE;
1196 
1197 		/* may fail if text not present */
1198 		bytes_left = copy_from_user_nmi(buf, (void __user *)from,
1199 						MAX_INSN_SIZE);
1200 		bytes_read = MAX_INSN_SIZE - bytes_left;
1201 		if (!bytes_read)
1202 			return X86_BR_NONE;
1203 
1204 		addr = buf;
1205 	} else {
1206 		/*
1207 		 * The LBR logs any address in the IP, even if the IP just
1208 		 * faulted. This means userspace can control the from address.
1209 		 * Ensure we don't blindly read any address by validating it is
1210 		 * a known text address.
1211 		 */
1212 		if (kernel_text_address(from)) {
1213 			addr = (void *)from;
1214 			/*
1215 			 * Assume we can get the maximum possible size
1216 			 * when grabbing kernel data.  This is not
1217 			 * _strictly_ true since we could possibly be
1218 			 * executing up next to a memory hole, but
1219 			 * it is very unlikely to be a problem.
1220 			 */
1221 			bytes_read = MAX_INSN_SIZE;
1222 		} else {
1223 			return X86_BR_NONE;
1224 		}
1225 	}
1226 
1227 	/*
1228 	 * decoder needs to know the ABI especially
1229 	 * on 64-bit systems running 32-bit apps
1230 	 */
1231 #ifdef CONFIG_X86_64
1232 	is64 = kernel_ip((unsigned long)addr) || any_64bit_mode(current_pt_regs());
1233 #endif
1234 	insn_init(&insn, addr, bytes_read, is64);
1235 	if (insn_get_opcode(&insn))
1236 		return X86_BR_ABORT;
1237 
1238 	switch (insn.opcode.bytes[0]) {
1239 	case 0xf:
1240 		switch (insn.opcode.bytes[1]) {
1241 		case 0x05: /* syscall */
1242 		case 0x34: /* sysenter */
1243 			ret = X86_BR_SYSCALL;
1244 			break;
1245 		case 0x07: /* sysret */
1246 		case 0x35: /* sysexit */
1247 			ret = X86_BR_SYSRET;
1248 			break;
1249 		case 0x80 ... 0x8f: /* conditional */
1250 			ret = X86_BR_JCC;
1251 			break;
1252 		default:
1253 			ret = X86_BR_NONE;
1254 		}
1255 		break;
1256 	case 0x70 ... 0x7f: /* conditional */
1257 		ret = X86_BR_JCC;
1258 		break;
1259 	case 0xc2: /* near ret */
1260 	case 0xc3: /* near ret */
1261 	case 0xca: /* far ret */
1262 	case 0xcb: /* far ret */
1263 		ret = X86_BR_RET;
1264 		break;
1265 	case 0xcf: /* iret */
1266 		ret = X86_BR_IRET;
1267 		break;
1268 	case 0xcc ... 0xce: /* int */
1269 		ret = X86_BR_INT;
1270 		break;
1271 	case 0xe8: /* call near rel */
1272 		if (insn_get_immediate(&insn) || insn.immediate1.value == 0) {
1273 			/* zero length call */
1274 			ret = X86_BR_ZERO_CALL;
1275 			break;
1276 		}
1277 		fallthrough;
1278 	case 0x9a: /* call far absolute */
1279 		ret = X86_BR_CALL;
1280 		break;
1281 	case 0xe0 ... 0xe3: /* loop jmp */
1282 		ret = X86_BR_JCC;
1283 		break;
1284 	case 0xe9 ... 0xeb: /* jmp */
1285 		ret = X86_BR_JMP;
1286 		break;
1287 	case 0xff: /* call near absolute, call far absolute ind */
1288 		if (insn_get_modrm(&insn))
1289 			return X86_BR_ABORT;
1290 
1291 		ext = (insn.modrm.bytes[0] >> 3) & 0x7;
1292 		switch (ext) {
1293 		case 2: /* near ind call */
1294 		case 3: /* far ind call */
1295 			ret = X86_BR_IND_CALL;
1296 			break;
1297 		case 4:
1298 		case 5:
1299 			ret = X86_BR_IND_JMP;
1300 			break;
1301 		}
1302 		break;
1303 	default:
1304 		ret = X86_BR_NONE;
1305 	}
1306 	/*
1307 	 * interrupts, traps, faults (and thus ring transition) may
1308 	 * occur on any instructions. Thus, to classify them correctly,
1309 	 * we need to first look at the from and to priv levels. If they
1310 	 * are different and to is in the kernel, then it indicates
1311 	 * a ring transition. If the from instruction is not a ring
1312 	 * transition instr (syscall, systenter, int), then it means
1313 	 * it was a irq, trap or fault.
1314 	 *
1315 	 * we have no way of detecting kernel to kernel faults.
1316 	 */
1317 	if (from_plm == X86_BR_USER && to_plm == X86_BR_KERNEL
1318 	    && ret != X86_BR_SYSCALL && ret != X86_BR_INT)
1319 		ret = X86_BR_IRQ;
1320 
1321 	/*
1322 	 * branch priv level determined by target as
1323 	 * is done by HW when LBR_SELECT is implemented
1324 	 */
1325 	if (ret != X86_BR_NONE)
1326 		ret |= to_plm;
1327 
1328 	return ret;
1329 }
1330 
1331 #define X86_BR_TYPE_MAP_MAX	16
1332 
1333 static int branch_map[X86_BR_TYPE_MAP_MAX] = {
1334 	PERF_BR_CALL,		/* X86_BR_CALL */
1335 	PERF_BR_RET,		/* X86_BR_RET */
1336 	PERF_BR_SYSCALL,	/* X86_BR_SYSCALL */
1337 	PERF_BR_SYSRET,		/* X86_BR_SYSRET */
1338 	PERF_BR_UNKNOWN,	/* X86_BR_INT */
1339 	PERF_BR_UNKNOWN,	/* X86_BR_IRET */
1340 	PERF_BR_COND,		/* X86_BR_JCC */
1341 	PERF_BR_UNCOND,		/* X86_BR_JMP */
1342 	PERF_BR_UNKNOWN,	/* X86_BR_IRQ */
1343 	PERF_BR_IND_CALL,	/* X86_BR_IND_CALL */
1344 	PERF_BR_UNKNOWN,	/* X86_BR_ABORT */
1345 	PERF_BR_UNKNOWN,	/* X86_BR_IN_TX */
1346 	PERF_BR_UNKNOWN,	/* X86_BR_NO_TX */
1347 	PERF_BR_CALL,		/* X86_BR_ZERO_CALL */
1348 	PERF_BR_UNKNOWN,	/* X86_BR_CALL_STACK */
1349 	PERF_BR_IND,		/* X86_BR_IND_JMP */
1350 };
1351 
1352 static int
1353 common_branch_type(int type)
1354 {
1355 	int i;
1356 
1357 	type >>= 2; /* skip X86_BR_USER and X86_BR_KERNEL */
1358 
1359 	if (type) {
1360 		i = __ffs(type);
1361 		if (i < X86_BR_TYPE_MAP_MAX)
1362 			return branch_map[i];
1363 	}
1364 
1365 	return PERF_BR_UNKNOWN;
1366 }
1367 
1368 enum {
1369 	ARCH_LBR_BR_TYPE_JCC			= 0,
1370 	ARCH_LBR_BR_TYPE_NEAR_IND_JMP		= 1,
1371 	ARCH_LBR_BR_TYPE_NEAR_REL_JMP		= 2,
1372 	ARCH_LBR_BR_TYPE_NEAR_IND_CALL		= 3,
1373 	ARCH_LBR_BR_TYPE_NEAR_REL_CALL		= 4,
1374 	ARCH_LBR_BR_TYPE_NEAR_RET		= 5,
1375 	ARCH_LBR_BR_TYPE_KNOWN_MAX		= ARCH_LBR_BR_TYPE_NEAR_RET,
1376 
1377 	ARCH_LBR_BR_TYPE_MAP_MAX		= 16,
1378 };
1379 
1380 static const int arch_lbr_br_type_map[ARCH_LBR_BR_TYPE_MAP_MAX] = {
1381 	[ARCH_LBR_BR_TYPE_JCC]			= X86_BR_JCC,
1382 	[ARCH_LBR_BR_TYPE_NEAR_IND_JMP]		= X86_BR_IND_JMP,
1383 	[ARCH_LBR_BR_TYPE_NEAR_REL_JMP]		= X86_BR_JMP,
1384 	[ARCH_LBR_BR_TYPE_NEAR_IND_CALL]	= X86_BR_IND_CALL,
1385 	[ARCH_LBR_BR_TYPE_NEAR_REL_CALL]	= X86_BR_CALL,
1386 	[ARCH_LBR_BR_TYPE_NEAR_RET]		= X86_BR_RET,
1387 };
1388 
1389 /*
1390  * implement actual branch filter based on user demand.
1391  * Hardware may not exactly satisfy that request, thus
1392  * we need to inspect opcodes. Mismatched branches are
1393  * discarded. Therefore, the number of branches returned
1394  * in PERF_SAMPLE_BRANCH_STACK sample may vary.
1395  */
1396 static void
1397 intel_pmu_lbr_filter(struct cpu_hw_events *cpuc)
1398 {
1399 	u64 from, to;
1400 	int br_sel = cpuc->br_sel;
1401 	int i, j, type, to_plm;
1402 	bool compress = false;
1403 
1404 	/* if sampling all branches, then nothing to filter */
1405 	if (((br_sel & X86_BR_ALL) == X86_BR_ALL) &&
1406 	    ((br_sel & X86_BR_TYPE_SAVE) != X86_BR_TYPE_SAVE))
1407 		return;
1408 
1409 	for (i = 0; i < cpuc->lbr_stack.nr; i++) {
1410 
1411 		from = cpuc->lbr_entries[i].from;
1412 		to = cpuc->lbr_entries[i].to;
1413 		type = cpuc->lbr_entries[i].type;
1414 
1415 		/*
1416 		 * Parse the branch type recorded in LBR_x_INFO MSR.
1417 		 * Doesn't support OTHER_BRANCH decoding for now.
1418 		 * OTHER_BRANCH branch type still rely on software decoding.
1419 		 */
1420 		if (static_cpu_has(X86_FEATURE_ARCH_LBR) &&
1421 		    type <= ARCH_LBR_BR_TYPE_KNOWN_MAX) {
1422 			to_plm = kernel_ip(to) ? X86_BR_KERNEL : X86_BR_USER;
1423 			type = arch_lbr_br_type_map[type] | to_plm;
1424 		} else
1425 			type = branch_type(from, to, cpuc->lbr_entries[i].abort);
1426 		if (type != X86_BR_NONE && (br_sel & X86_BR_ANYTX)) {
1427 			if (cpuc->lbr_entries[i].in_tx)
1428 				type |= X86_BR_IN_TX;
1429 			else
1430 				type |= X86_BR_NO_TX;
1431 		}
1432 
1433 		/* if type does not correspond, then discard */
1434 		if (type == X86_BR_NONE || (br_sel & type) != type) {
1435 			cpuc->lbr_entries[i].from = 0;
1436 			compress = true;
1437 		}
1438 
1439 		if ((br_sel & X86_BR_TYPE_SAVE) == X86_BR_TYPE_SAVE)
1440 			cpuc->lbr_entries[i].type = common_branch_type(type);
1441 	}
1442 
1443 	if (!compress)
1444 		return;
1445 
1446 	/* remove all entries with from=0 */
1447 	for (i = 0; i < cpuc->lbr_stack.nr; ) {
1448 		if (!cpuc->lbr_entries[i].from) {
1449 			j = i;
1450 			while (++j < cpuc->lbr_stack.nr)
1451 				cpuc->lbr_entries[j-1] = cpuc->lbr_entries[j];
1452 			cpuc->lbr_stack.nr--;
1453 			if (!cpuc->lbr_entries[i].from)
1454 				continue;
1455 		}
1456 		i++;
1457 	}
1458 }
1459 
1460 void intel_pmu_store_pebs_lbrs(struct lbr_entry *lbr)
1461 {
1462 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
1463 
1464 	/* Cannot get TOS for large PEBS and Arch LBR */
1465 	if (static_cpu_has(X86_FEATURE_ARCH_LBR) ||
1466 	    (cpuc->n_pebs == cpuc->n_large_pebs))
1467 		cpuc->lbr_stack.hw_idx = -1ULL;
1468 	else
1469 		cpuc->lbr_stack.hw_idx = intel_pmu_lbr_tos();
1470 
1471 	intel_pmu_store_lbr(cpuc, lbr);
1472 	intel_pmu_lbr_filter(cpuc);
1473 }
1474 
1475 /*
1476  * Map interface branch filters onto LBR filters
1477  */
1478 static const int nhm_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1479 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1480 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1481 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1482 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1483 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_REL_JMP
1484 						| LBR_IND_JMP | LBR_FAR,
1485 	/*
1486 	 * NHM/WSM erratum: must include REL_JMP+IND_JMP to get CALL branches
1487 	 */
1488 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] =
1489 	 LBR_REL_CALL | LBR_IND_CALL | LBR_REL_JMP | LBR_IND_JMP | LBR_FAR,
1490 	/*
1491 	 * NHM/WSM erratum: must include IND_JMP to capture IND_CALL
1492 	 */
1493 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_IND_CALL | LBR_IND_JMP,
1494 	[PERF_SAMPLE_BRANCH_COND_SHIFT]     = LBR_JCC,
1495 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_IND_JMP,
1496 };
1497 
1498 static const int snb_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1499 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1500 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1501 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1502 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1503 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1504 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1505 						| LBR_FAR,
1506 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1507 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1508 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1509 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1510 };
1511 
1512 static const int hsw_lbr_sel_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1513 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= LBR_ANY,
1514 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= LBR_USER,
1515 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= LBR_KERNEL,
1516 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1517 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= LBR_RETURN | LBR_FAR,
1518 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1519 						| LBR_FAR,
1520 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]	= LBR_IND_CALL,
1521 	[PERF_SAMPLE_BRANCH_COND_SHIFT]		= LBR_JCC,
1522 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]	= LBR_REL_CALL | LBR_IND_CALL
1523 						| LBR_RETURN | LBR_CALL_STACK,
1524 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= LBR_IND_JMP,
1525 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= LBR_REL_CALL,
1526 };
1527 
1528 static int arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_MAX_SHIFT] = {
1529 	[PERF_SAMPLE_BRANCH_ANY_SHIFT]		= ARCH_LBR_ANY,
1530 	[PERF_SAMPLE_BRANCH_USER_SHIFT]		= ARCH_LBR_USER,
1531 	[PERF_SAMPLE_BRANCH_KERNEL_SHIFT]	= ARCH_LBR_KERNEL,
1532 	[PERF_SAMPLE_BRANCH_HV_SHIFT]		= LBR_IGN,
1533 	[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT]	= ARCH_LBR_RETURN |
1534 						  ARCH_LBR_OTHER_BRANCH,
1535 	[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT]     = ARCH_LBR_REL_CALL |
1536 						  ARCH_LBR_IND_CALL |
1537 						  ARCH_LBR_OTHER_BRANCH,
1538 	[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT]     = ARCH_LBR_IND_CALL,
1539 	[PERF_SAMPLE_BRANCH_COND_SHIFT]         = ARCH_LBR_JCC,
1540 	[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT]   = ARCH_LBR_REL_CALL |
1541 						  ARCH_LBR_IND_CALL |
1542 						  ARCH_LBR_RETURN |
1543 						  ARCH_LBR_CALL_STACK,
1544 	[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT]	= ARCH_LBR_IND_JMP,
1545 	[PERF_SAMPLE_BRANCH_CALL_SHIFT]		= ARCH_LBR_REL_CALL,
1546 };
1547 
1548 /* core */
1549 void __init intel_pmu_lbr_init_core(void)
1550 {
1551 	x86_pmu.lbr_nr     = 4;
1552 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1553 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1554 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1555 
1556 	/*
1557 	 * SW branch filter usage:
1558 	 * - compensate for lack of HW filter
1559 	 */
1560 }
1561 
1562 /* nehalem/westmere */
1563 void __init intel_pmu_lbr_init_nhm(void)
1564 {
1565 	x86_pmu.lbr_nr     = 16;
1566 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1567 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1568 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1569 
1570 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1571 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1572 
1573 	/*
1574 	 * SW branch filter usage:
1575 	 * - workaround LBR_SEL errata (see above)
1576 	 * - support syscall, sysret capture.
1577 	 *   That requires LBR_FAR but that means far
1578 	 *   jmp need to be filtered out
1579 	 */
1580 }
1581 
1582 /* sandy bridge */
1583 void __init intel_pmu_lbr_init_snb(void)
1584 {
1585 	x86_pmu.lbr_nr	 = 16;
1586 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1587 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1588 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1589 
1590 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1591 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1592 
1593 	/*
1594 	 * SW branch filter usage:
1595 	 * - support syscall, sysret capture.
1596 	 *   That requires LBR_FAR but that means far
1597 	 *   jmp need to be filtered out
1598 	 */
1599 }
1600 
1601 static inline struct kmem_cache *
1602 create_lbr_kmem_cache(size_t size, size_t align)
1603 {
1604 	return kmem_cache_create("x86_lbr", size, align, 0, NULL);
1605 }
1606 
1607 /* haswell */
1608 void intel_pmu_lbr_init_hsw(void)
1609 {
1610 	size_t size = sizeof(struct x86_perf_task_context);
1611 
1612 	x86_pmu.lbr_nr	 = 16;
1613 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1614 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1615 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1616 
1617 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1618 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1619 
1620 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1621 
1622 	if (lbr_from_signext_quirk_needed())
1623 		static_branch_enable(&lbr_from_quirk_key);
1624 }
1625 
1626 /* skylake */
1627 __init void intel_pmu_lbr_init_skl(void)
1628 {
1629 	size_t size = sizeof(struct x86_perf_task_context);
1630 
1631 	x86_pmu.lbr_nr	 = 32;
1632 	x86_pmu.lbr_tos	 = MSR_LBR_TOS;
1633 	x86_pmu.lbr_from = MSR_LBR_NHM_FROM;
1634 	x86_pmu.lbr_to   = MSR_LBR_NHM_TO;
1635 	x86_pmu.lbr_info = MSR_LBR_INFO_0;
1636 
1637 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1638 	x86_pmu.lbr_sel_map  = hsw_lbr_sel_map;
1639 
1640 	x86_get_pmu(smp_processor_id())->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1641 
1642 	/*
1643 	 * SW branch filter usage:
1644 	 * - support syscall, sysret capture.
1645 	 *   That requires LBR_FAR but that means far
1646 	 *   jmp need to be filtered out
1647 	 */
1648 }
1649 
1650 /* atom */
1651 void __init intel_pmu_lbr_init_atom(void)
1652 {
1653 	/*
1654 	 * only models starting at stepping 10 seems
1655 	 * to have an operational LBR which can freeze
1656 	 * on PMU interrupt
1657 	 */
1658 	if (boot_cpu_data.x86_model == 28
1659 	    && boot_cpu_data.x86_stepping < 10) {
1660 		pr_cont("LBR disabled due to erratum");
1661 		return;
1662 	}
1663 
1664 	x86_pmu.lbr_nr	   = 8;
1665 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1666 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1667 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1668 
1669 	/*
1670 	 * SW branch filter usage:
1671 	 * - compensate for lack of HW filter
1672 	 */
1673 }
1674 
1675 /* slm */
1676 void __init intel_pmu_lbr_init_slm(void)
1677 {
1678 	x86_pmu.lbr_nr	   = 8;
1679 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1680 	x86_pmu.lbr_from   = MSR_LBR_CORE_FROM;
1681 	x86_pmu.lbr_to     = MSR_LBR_CORE_TO;
1682 
1683 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1684 	x86_pmu.lbr_sel_map  = nhm_lbr_sel_map;
1685 
1686 	/*
1687 	 * SW branch filter usage:
1688 	 * - compensate for lack of HW filter
1689 	 */
1690 	pr_cont("8-deep LBR, ");
1691 }
1692 
1693 /* Knights Landing */
1694 void intel_pmu_lbr_init_knl(void)
1695 {
1696 	x86_pmu.lbr_nr	   = 8;
1697 	x86_pmu.lbr_tos    = MSR_LBR_TOS;
1698 	x86_pmu.lbr_from   = MSR_LBR_NHM_FROM;
1699 	x86_pmu.lbr_to     = MSR_LBR_NHM_TO;
1700 
1701 	x86_pmu.lbr_sel_mask = LBR_SEL_MASK;
1702 	x86_pmu.lbr_sel_map  = snb_lbr_sel_map;
1703 
1704 	/* Knights Landing does have MISPREDICT bit */
1705 	if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_LIP)
1706 		x86_pmu.intel_cap.lbr_format = LBR_FORMAT_EIP_FLAGS;
1707 }
1708 
1709 /*
1710  * LBR state size is variable based on the max number of registers.
1711  * This calculates the expected state size, which should match
1712  * what the hardware enumerates for the size of XFEATURE_LBR.
1713  */
1714 static inline unsigned int get_lbr_state_size(void)
1715 {
1716 	return sizeof(struct arch_lbr_state) +
1717 	       x86_pmu.lbr_nr * sizeof(struct lbr_entry);
1718 }
1719 
1720 static bool is_arch_lbr_xsave_available(void)
1721 {
1722 	if (!boot_cpu_has(X86_FEATURE_XSAVES))
1723 		return false;
1724 
1725 	/*
1726 	 * Check the LBR state with the corresponding software structure.
1727 	 * Disable LBR XSAVES support if the size doesn't match.
1728 	 */
1729 	if (WARN_ON(xfeature_size(XFEATURE_LBR) != get_lbr_state_size()))
1730 		return false;
1731 
1732 	return true;
1733 }
1734 
1735 void __init intel_pmu_arch_lbr_init(void)
1736 {
1737 	struct pmu *pmu = x86_get_pmu(smp_processor_id());
1738 	union cpuid28_eax eax;
1739 	union cpuid28_ebx ebx;
1740 	union cpuid28_ecx ecx;
1741 	unsigned int unused_edx;
1742 	bool arch_lbr_xsave;
1743 	size_t size;
1744 	u64 lbr_nr;
1745 
1746 	/* Arch LBR Capabilities */
1747 	cpuid(28, &eax.full, &ebx.full, &ecx.full, &unused_edx);
1748 
1749 	lbr_nr = fls(eax.split.lbr_depth_mask) * 8;
1750 	if (!lbr_nr)
1751 		goto clear_arch_lbr;
1752 
1753 	/* Apply the max depth of Arch LBR */
1754 	if (wrmsrl_safe(MSR_ARCH_LBR_DEPTH, lbr_nr))
1755 		goto clear_arch_lbr;
1756 
1757 	x86_pmu.lbr_depth_mask = eax.split.lbr_depth_mask;
1758 	x86_pmu.lbr_deep_c_reset = eax.split.lbr_deep_c_reset;
1759 	x86_pmu.lbr_lip = eax.split.lbr_lip;
1760 	x86_pmu.lbr_cpl = ebx.split.lbr_cpl;
1761 	x86_pmu.lbr_filter = ebx.split.lbr_filter;
1762 	x86_pmu.lbr_call_stack = ebx.split.lbr_call_stack;
1763 	x86_pmu.lbr_mispred = ecx.split.lbr_mispred;
1764 	x86_pmu.lbr_timed_lbr = ecx.split.lbr_timed_lbr;
1765 	x86_pmu.lbr_br_type = ecx.split.lbr_br_type;
1766 	x86_pmu.lbr_nr = lbr_nr;
1767 
1768 
1769 	arch_lbr_xsave = is_arch_lbr_xsave_available();
1770 	if (arch_lbr_xsave) {
1771 		size = sizeof(struct x86_perf_task_context_arch_lbr_xsave) +
1772 		       get_lbr_state_size();
1773 		pmu->task_ctx_cache = create_lbr_kmem_cache(size,
1774 							    XSAVE_ALIGNMENT);
1775 	}
1776 
1777 	if (!pmu->task_ctx_cache) {
1778 		arch_lbr_xsave = false;
1779 
1780 		size = sizeof(struct x86_perf_task_context_arch_lbr) +
1781 		       lbr_nr * sizeof(struct lbr_entry);
1782 		pmu->task_ctx_cache = create_lbr_kmem_cache(size, 0);
1783 	}
1784 
1785 	x86_pmu.lbr_from = MSR_ARCH_LBR_FROM_0;
1786 	x86_pmu.lbr_to = MSR_ARCH_LBR_TO_0;
1787 	x86_pmu.lbr_info = MSR_ARCH_LBR_INFO_0;
1788 
1789 	/* LBR callstack requires both CPL and Branch Filtering support */
1790 	if (!x86_pmu.lbr_cpl ||
1791 	    !x86_pmu.lbr_filter ||
1792 	    !x86_pmu.lbr_call_stack)
1793 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT] = LBR_NOT_SUPP;
1794 
1795 	if (!x86_pmu.lbr_cpl) {
1796 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_USER_SHIFT] = LBR_NOT_SUPP;
1797 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_KERNEL_SHIFT] = LBR_NOT_SUPP;
1798 	} else if (!x86_pmu.lbr_filter) {
1799 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_SHIFT] = LBR_NOT_SUPP;
1800 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT] = LBR_NOT_SUPP;
1801 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT] = LBR_NOT_SUPP;
1802 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_CALL_SHIFT] = LBR_NOT_SUPP;
1803 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_COND_SHIFT] = LBR_NOT_SUPP;
1804 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT] = LBR_NOT_SUPP;
1805 		arch_lbr_ctl_map[PERF_SAMPLE_BRANCH_CALL_SHIFT] = LBR_NOT_SUPP;
1806 	}
1807 
1808 	x86_pmu.lbr_ctl_mask = ARCH_LBR_CTL_MASK;
1809 	x86_pmu.lbr_ctl_map  = arch_lbr_ctl_map;
1810 
1811 	if (!x86_pmu.lbr_cpl && !x86_pmu.lbr_filter)
1812 		x86_pmu.lbr_ctl_map = NULL;
1813 
1814 	x86_pmu.lbr_reset = intel_pmu_arch_lbr_reset;
1815 	if (arch_lbr_xsave) {
1816 		x86_pmu.lbr_save = intel_pmu_arch_lbr_xsaves;
1817 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_xrstors;
1818 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read_xsave;
1819 		pr_cont("XSAVE ");
1820 	} else {
1821 		x86_pmu.lbr_save = intel_pmu_arch_lbr_save;
1822 		x86_pmu.lbr_restore = intel_pmu_arch_lbr_restore;
1823 		x86_pmu.lbr_read = intel_pmu_arch_lbr_read;
1824 	}
1825 
1826 	pr_cont("Architectural LBR, ");
1827 
1828 	return;
1829 
1830 clear_arch_lbr:
1831 	clear_cpu_cap(&boot_cpu_data, X86_FEATURE_ARCH_LBR);
1832 }
1833 
1834 /**
1835  * x86_perf_get_lbr - get the LBR records information
1836  *
1837  * @lbr: the caller's memory to store the LBR records information
1838  *
1839  * Returns: 0 indicates the LBR info has been successfully obtained
1840  */
1841 int x86_perf_get_lbr(struct x86_pmu_lbr *lbr)
1842 {
1843 	int lbr_fmt = x86_pmu.intel_cap.lbr_format;
1844 
1845 	lbr->nr = x86_pmu.lbr_nr;
1846 	lbr->from = x86_pmu.lbr_from;
1847 	lbr->to = x86_pmu.lbr_to;
1848 	lbr->info = (lbr_fmt == LBR_FORMAT_INFO) ? x86_pmu.lbr_info : 0;
1849 
1850 	return 0;
1851 }
1852 EXPORT_SYMBOL_GPL(x86_perf_get_lbr);
1853 
1854 struct event_constraint vlbr_constraint =
1855 	__EVENT_CONSTRAINT(INTEL_FIXED_VLBR_EVENT, (1ULL << INTEL_PMC_IDX_FIXED_VLBR),
1856 			  FIXED_EVENT_FLAGS, 1, 0, PERF_X86_EVENT_LBR_SELECT);
1857