xref: /linux/arch/arm64/kvm/pmu-emul.c (revision 7a9b709e7cc5ce1ffb84ce07bf6d157e1de758df)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 Linaro Ltd.
4  * Author: Shannon Zhao <shannon.zhao@linaro.org>
5  */
6 
7 #include <linux/cpu.h>
8 #include <linux/kvm.h>
9 #include <linux/kvm_host.h>
10 #include <linux/list.h>
11 #include <linux/perf_event.h>
12 #include <linux/perf/arm_pmu.h>
13 #include <linux/uaccess.h>
14 #include <asm/kvm_emulate.h>
15 #include <kvm/arm_pmu.h>
16 #include <kvm/arm_vgic.h>
17 
18 #define PERF_ATTR_CFG1_COUNTER_64BIT	BIT(0)
19 
20 static LIST_HEAD(arm_pmus);
21 static DEFINE_MUTEX(arm_pmus_lock);
22 
23 static void kvm_pmu_create_perf_event(struct kvm_pmc *pmc);
24 static void kvm_pmu_release_perf_event(struct kvm_pmc *pmc);
25 static bool kvm_pmu_counter_is_enabled(struct kvm_pmc *pmc);
26 
27 bool kvm_supports_guest_pmuv3(void)
28 {
29 	guard(mutex)(&arm_pmus_lock);
30 	return !list_empty(&arm_pmus);
31 }
32 
33 static struct kvm_vcpu *kvm_pmc_to_vcpu(const struct kvm_pmc *pmc)
34 {
35 	return container_of(pmc, struct kvm_vcpu, arch.pmu.pmc[pmc->idx]);
36 }
37 
38 static struct kvm_pmc *kvm_vcpu_idx_to_pmc(struct kvm_vcpu *vcpu, int cnt_idx)
39 {
40 	return &vcpu->arch.pmu.pmc[cnt_idx];
41 }
42 
43 static u32 __kvm_pmu_event_mask(unsigned int pmuver)
44 {
45 	switch (pmuver) {
46 	case ID_AA64DFR0_EL1_PMUVer_IMP:
47 		return GENMASK(9, 0);
48 	case ID_AA64DFR0_EL1_PMUVer_V3P1:
49 	case ID_AA64DFR0_EL1_PMUVer_V3P4:
50 	case ID_AA64DFR0_EL1_PMUVer_V3P5:
51 	case ID_AA64DFR0_EL1_PMUVer_V3P7:
52 		return GENMASK(15, 0);
53 	default:		/* Shouldn't be here, just for sanity */
54 		WARN_ONCE(1, "Unknown PMU version %d\n", pmuver);
55 		return 0;
56 	}
57 }
58 
59 static u32 kvm_pmu_event_mask(struct kvm *kvm)
60 {
61 	u64 dfr0 = kvm_read_vm_id_reg(kvm, SYS_ID_AA64DFR0_EL1);
62 	u8 pmuver = SYS_FIELD_GET(ID_AA64DFR0_EL1, PMUVer, dfr0);
63 
64 	return __kvm_pmu_event_mask(pmuver);
65 }
66 
67 u64 kvm_pmu_evtyper_mask(struct kvm *kvm)
68 {
69 	u64 mask = ARMV8_PMU_EXCLUDE_EL1 | ARMV8_PMU_EXCLUDE_EL0 |
70 		   kvm_pmu_event_mask(kvm);
71 
72 	if (kvm_has_feat(kvm, ID_AA64PFR0_EL1, EL2, IMP))
73 		mask |= ARMV8_PMU_INCLUDE_EL2;
74 
75 	if (kvm_has_feat(kvm, ID_AA64PFR0_EL1, EL3, IMP))
76 		mask |= ARMV8_PMU_EXCLUDE_NS_EL0 |
77 			ARMV8_PMU_EXCLUDE_NS_EL1 |
78 			ARMV8_PMU_EXCLUDE_EL3;
79 
80 	return mask;
81 }
82 
83 /**
84  * kvm_pmc_is_64bit - determine if counter is 64bit
85  * @pmc: counter context
86  */
87 static bool kvm_pmc_is_64bit(struct kvm_pmc *pmc)
88 {
89 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
90 
91 	return (pmc->idx == ARMV8_PMU_CYCLE_IDX ||
92 		kvm_has_feat(vcpu->kvm, ID_AA64DFR0_EL1, PMUVer, V3P5));
93 }
94 
95 static bool kvm_pmc_has_64bit_overflow(struct kvm_pmc *pmc)
96 {
97 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
98 	u64 val = kvm_vcpu_read_pmcr(vcpu);
99 
100 	if (kvm_pmu_counter_is_hyp(vcpu, pmc->idx))
101 		return __vcpu_sys_reg(vcpu, MDCR_EL2) & MDCR_EL2_HLP;
102 
103 	return (pmc->idx < ARMV8_PMU_CYCLE_IDX && (val & ARMV8_PMU_PMCR_LP)) ||
104 	       (pmc->idx == ARMV8_PMU_CYCLE_IDX && (val & ARMV8_PMU_PMCR_LC));
105 }
106 
107 static bool kvm_pmu_counter_can_chain(struct kvm_pmc *pmc)
108 {
109 	return (!(pmc->idx & 1) && (pmc->idx + 1) < ARMV8_PMU_CYCLE_IDX &&
110 		!kvm_pmc_has_64bit_overflow(pmc));
111 }
112 
113 static u32 counter_index_to_reg(u64 idx)
114 {
115 	return (idx == ARMV8_PMU_CYCLE_IDX) ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + idx;
116 }
117 
118 static u32 counter_index_to_evtreg(u64 idx)
119 {
120 	return (idx == ARMV8_PMU_CYCLE_IDX) ? PMCCFILTR_EL0 : PMEVTYPER0_EL0 + idx;
121 }
122 
123 static u64 kvm_pmc_read_evtreg(const struct kvm_pmc *pmc)
124 {
125 	return __vcpu_sys_reg(kvm_pmc_to_vcpu(pmc), counter_index_to_evtreg(pmc->idx));
126 }
127 
128 static u64 kvm_pmu_get_pmc_value(struct kvm_pmc *pmc)
129 {
130 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
131 	u64 counter, reg, enabled, running;
132 
133 	reg = counter_index_to_reg(pmc->idx);
134 	counter = __vcpu_sys_reg(vcpu, reg);
135 
136 	/*
137 	 * The real counter value is equal to the value of counter register plus
138 	 * the value perf event counts.
139 	 */
140 	if (pmc->perf_event)
141 		counter += perf_event_read_value(pmc->perf_event, &enabled,
142 						 &running);
143 
144 	if (!kvm_pmc_is_64bit(pmc))
145 		counter = lower_32_bits(counter);
146 
147 	return counter;
148 }
149 
150 /**
151  * kvm_pmu_get_counter_value - get PMU counter value
152  * @vcpu: The vcpu pointer
153  * @select_idx: The counter index
154  */
155 u64 kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u64 select_idx)
156 {
157 	return kvm_pmu_get_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, select_idx));
158 }
159 
160 static void kvm_pmu_set_pmc_value(struct kvm_pmc *pmc, u64 val, bool force)
161 {
162 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
163 	u64 reg;
164 
165 	kvm_pmu_release_perf_event(pmc);
166 
167 	reg = counter_index_to_reg(pmc->idx);
168 
169 	if (vcpu_mode_is_32bit(vcpu) && pmc->idx != ARMV8_PMU_CYCLE_IDX &&
170 	    !force) {
171 		/*
172 		 * Even with PMUv3p5, AArch32 cannot write to the top
173 		 * 32bit of the counters. The only possible course of
174 		 * action is to use PMCR.P, which will reset them to
175 		 * 0 (the only use of the 'force' parameter).
176 		 */
177 		val  = __vcpu_sys_reg(vcpu, reg) & GENMASK(63, 32);
178 		val |= lower_32_bits(val);
179 	}
180 
181 	__vcpu_sys_reg(vcpu, reg) = val;
182 
183 	/* Recreate the perf event to reflect the updated sample_period */
184 	kvm_pmu_create_perf_event(pmc);
185 }
186 
187 /**
188  * kvm_pmu_set_counter_value - set PMU counter value
189  * @vcpu: The vcpu pointer
190  * @select_idx: The counter index
191  * @val: The counter value
192  */
193 void kvm_pmu_set_counter_value(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
194 {
195 	kvm_pmu_set_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, select_idx), val, false);
196 }
197 
198 /**
199  * kvm_pmu_set_counter_value_user - set PMU counter value from user
200  * @vcpu: The vcpu pointer
201  * @select_idx: The counter index
202  * @val: The counter value
203  */
204 void kvm_pmu_set_counter_value_user(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
205 {
206 	kvm_pmu_release_perf_event(kvm_vcpu_idx_to_pmc(vcpu, select_idx));
207 	__vcpu_sys_reg(vcpu, counter_index_to_reg(select_idx)) = val;
208 	kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
209 }
210 
211 /**
212  * kvm_pmu_release_perf_event - remove the perf event
213  * @pmc: The PMU counter pointer
214  */
215 static void kvm_pmu_release_perf_event(struct kvm_pmc *pmc)
216 {
217 	if (pmc->perf_event) {
218 		perf_event_disable(pmc->perf_event);
219 		perf_event_release_kernel(pmc->perf_event);
220 		pmc->perf_event = NULL;
221 	}
222 }
223 
224 /**
225  * kvm_pmu_stop_counter - stop PMU counter
226  * @pmc: The PMU counter pointer
227  *
228  * If this counter has been configured to monitor some event, release it here.
229  */
230 static void kvm_pmu_stop_counter(struct kvm_pmc *pmc)
231 {
232 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
233 	u64 reg, val;
234 
235 	if (!pmc->perf_event)
236 		return;
237 
238 	val = kvm_pmu_get_pmc_value(pmc);
239 
240 	reg = counter_index_to_reg(pmc->idx);
241 
242 	__vcpu_sys_reg(vcpu, reg) = val;
243 
244 	kvm_pmu_release_perf_event(pmc);
245 }
246 
247 /**
248  * kvm_pmu_vcpu_init - assign pmu counter idx for cpu
249  * @vcpu: The vcpu pointer
250  *
251  */
252 void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)
253 {
254 	int i;
255 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
256 
257 	for (i = 0; i < KVM_ARMV8_PMU_MAX_COUNTERS; i++)
258 		pmu->pmc[i].idx = i;
259 }
260 
261 /**
262  * kvm_pmu_vcpu_destroy - free perf event of PMU for cpu
263  * @vcpu: The vcpu pointer
264  *
265  */
266 void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
267 {
268 	int i;
269 
270 	for (i = 0; i < KVM_ARMV8_PMU_MAX_COUNTERS; i++)
271 		kvm_pmu_release_perf_event(kvm_vcpu_idx_to_pmc(vcpu, i));
272 	irq_work_sync(&vcpu->arch.pmu.overflow_work);
273 }
274 
275 static u64 kvm_pmu_hyp_counter_mask(struct kvm_vcpu *vcpu)
276 {
277 	unsigned int hpmn, n;
278 
279 	if (!vcpu_has_nv(vcpu))
280 		return 0;
281 
282 	hpmn = SYS_FIELD_GET(MDCR_EL2, HPMN, __vcpu_sys_reg(vcpu, MDCR_EL2));
283 	n = vcpu->kvm->arch.pmcr_n;
284 
285 	/*
286 	 * Programming HPMN to a value greater than PMCR_EL0.N is
287 	 * CONSTRAINED UNPREDICTABLE. Make the implementation choice that an
288 	 * UNKNOWN number of counters (in our case, zero) are reserved for EL2.
289 	 */
290 	if (hpmn >= n)
291 		return 0;
292 
293 	/*
294 	 * Programming HPMN=0 is CONSTRAINED UNPREDICTABLE if FEAT_HPMN0 isn't
295 	 * implemented. Since KVM's ability to emulate HPMN=0 does not directly
296 	 * depend on hardware (all PMU registers are trapped), make the
297 	 * implementation choice that all counters are included in the second
298 	 * range reserved for EL2/EL3.
299 	 */
300 	return GENMASK(n - 1, hpmn);
301 }
302 
303 bool kvm_pmu_counter_is_hyp(struct kvm_vcpu *vcpu, unsigned int idx)
304 {
305 	return kvm_pmu_hyp_counter_mask(vcpu) & BIT(idx);
306 }
307 
308 u64 kvm_pmu_accessible_counter_mask(struct kvm_vcpu *vcpu)
309 {
310 	u64 mask = kvm_pmu_implemented_counter_mask(vcpu);
311 
312 	if (!vcpu_has_nv(vcpu) || vcpu_is_el2(vcpu))
313 		return mask;
314 
315 	return mask & ~kvm_pmu_hyp_counter_mask(vcpu);
316 }
317 
318 u64 kvm_pmu_implemented_counter_mask(struct kvm_vcpu *vcpu)
319 {
320 	u64 val = FIELD_GET(ARMV8_PMU_PMCR_N, kvm_vcpu_read_pmcr(vcpu));
321 
322 	if (val == 0)
323 		return BIT(ARMV8_PMU_CYCLE_IDX);
324 	else
325 		return GENMASK(val - 1, 0) | BIT(ARMV8_PMU_CYCLE_IDX);
326 }
327 
328 static void kvm_pmc_enable_perf_event(struct kvm_pmc *pmc)
329 {
330 	if (!pmc->perf_event) {
331 		kvm_pmu_create_perf_event(pmc);
332 		return;
333 	}
334 
335 	perf_event_enable(pmc->perf_event);
336 	if (pmc->perf_event->state != PERF_EVENT_STATE_ACTIVE)
337 		kvm_debug("fail to enable perf event\n");
338 }
339 
340 static void kvm_pmc_disable_perf_event(struct kvm_pmc *pmc)
341 {
342 	if (pmc->perf_event)
343 		perf_event_disable(pmc->perf_event);
344 }
345 
346 void kvm_pmu_reprogram_counter_mask(struct kvm_vcpu *vcpu, u64 val)
347 {
348 	int i;
349 
350 	if (!val)
351 		return;
352 
353 	for (i = 0; i < KVM_ARMV8_PMU_MAX_COUNTERS; i++) {
354 		struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
355 
356 		if (!(val & BIT(i)))
357 			continue;
358 
359 		if (kvm_pmu_counter_is_enabled(pmc))
360 			kvm_pmc_enable_perf_event(pmc);
361 		else
362 			kvm_pmc_disable_perf_event(pmc);
363 	}
364 
365 	kvm_vcpu_pmu_restore_guest(vcpu);
366 }
367 
368 /*
369  * Returns the PMU overflow state, which is true if there exists an event
370  * counter where the values of the global enable control, PMOVSSET_EL0[n], and
371  * PMINTENSET_EL1[n] are all 1.
372  */
373 static bool kvm_pmu_overflow_status(struct kvm_vcpu *vcpu)
374 {
375 	u64 reg = __vcpu_sys_reg(vcpu, PMOVSSET_EL0);
376 
377 	reg &= __vcpu_sys_reg(vcpu, PMINTENSET_EL1);
378 
379 	/*
380 	 * PMCR_EL0.E is the global enable control for event counters available
381 	 * to EL0 and EL1.
382 	 */
383 	if (!(kvm_vcpu_read_pmcr(vcpu) & ARMV8_PMU_PMCR_E))
384 		reg &= kvm_pmu_hyp_counter_mask(vcpu);
385 
386 	/*
387 	 * Otherwise, MDCR_EL2.HPME is the global enable control for event
388 	 * counters reserved for EL2.
389 	 */
390 	if (!(vcpu_read_sys_reg(vcpu, MDCR_EL2) & MDCR_EL2_HPME))
391 		reg &= ~kvm_pmu_hyp_counter_mask(vcpu);
392 
393 	return reg;
394 }
395 
396 static void kvm_pmu_update_state(struct kvm_vcpu *vcpu)
397 {
398 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
399 	bool overflow;
400 
401 	overflow = kvm_pmu_overflow_status(vcpu);
402 	if (pmu->irq_level == overflow)
403 		return;
404 
405 	pmu->irq_level = overflow;
406 
407 	if (likely(irqchip_in_kernel(vcpu->kvm))) {
408 		int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu,
409 					      pmu->irq_num, overflow, pmu);
410 		WARN_ON(ret);
411 	}
412 }
413 
414 bool kvm_pmu_should_notify_user(struct kvm_vcpu *vcpu)
415 {
416 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
417 	struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
418 	bool run_level = sregs->device_irq_level & KVM_ARM_DEV_PMU;
419 
420 	if (likely(irqchip_in_kernel(vcpu->kvm)))
421 		return false;
422 
423 	return pmu->irq_level != run_level;
424 }
425 
426 /*
427  * Reflect the PMU overflow interrupt output level into the kvm_run structure
428  */
429 void kvm_pmu_update_run(struct kvm_vcpu *vcpu)
430 {
431 	struct kvm_sync_regs *regs = &vcpu->run->s.regs;
432 
433 	/* Populate the timer bitmap for user space */
434 	regs->device_irq_level &= ~KVM_ARM_DEV_PMU;
435 	if (vcpu->arch.pmu.irq_level)
436 		regs->device_irq_level |= KVM_ARM_DEV_PMU;
437 }
438 
439 /**
440  * kvm_pmu_flush_hwstate - flush pmu state to cpu
441  * @vcpu: The vcpu pointer
442  *
443  * Check if the PMU has overflowed while we were running in the host, and inject
444  * an interrupt if that was the case.
445  */
446 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
447 {
448 	kvm_pmu_update_state(vcpu);
449 }
450 
451 /**
452  * kvm_pmu_sync_hwstate - sync pmu state from cpu
453  * @vcpu: The vcpu pointer
454  *
455  * Check if the PMU has overflowed while we were running in the guest, and
456  * inject an interrupt if that was the case.
457  */
458 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
459 {
460 	kvm_pmu_update_state(vcpu);
461 }
462 
463 /*
464  * When perf interrupt is an NMI, we cannot safely notify the vcpu corresponding
465  * to the event.
466  * This is why we need a callback to do it once outside of the NMI context.
467  */
468 static void kvm_pmu_perf_overflow_notify_vcpu(struct irq_work *work)
469 {
470 	struct kvm_vcpu *vcpu;
471 
472 	vcpu = container_of(work, struct kvm_vcpu, arch.pmu.overflow_work);
473 	kvm_vcpu_kick(vcpu);
474 }
475 
476 /*
477  * Perform an increment on any of the counters described in @mask,
478  * generating the overflow if required, and propagate it as a chained
479  * event if possible.
480  */
481 static void kvm_pmu_counter_increment(struct kvm_vcpu *vcpu,
482 				      unsigned long mask, u32 event)
483 {
484 	int i;
485 
486 	if (!(kvm_vcpu_read_pmcr(vcpu) & ARMV8_PMU_PMCR_E))
487 		return;
488 
489 	/* Weed out disabled counters */
490 	mask &= __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
491 
492 	for_each_set_bit(i, &mask, ARMV8_PMU_CYCLE_IDX) {
493 		struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
494 		u64 type, reg;
495 
496 		/* Filter on event type */
497 		type = __vcpu_sys_reg(vcpu, counter_index_to_evtreg(i));
498 		type &= kvm_pmu_event_mask(vcpu->kvm);
499 		if (type != event)
500 			continue;
501 
502 		/* Increment this counter */
503 		reg = __vcpu_sys_reg(vcpu, counter_index_to_reg(i)) + 1;
504 		if (!kvm_pmc_is_64bit(pmc))
505 			reg = lower_32_bits(reg);
506 		__vcpu_sys_reg(vcpu, counter_index_to_reg(i)) = reg;
507 
508 		/* No overflow? move on */
509 		if (kvm_pmc_has_64bit_overflow(pmc) ? reg : lower_32_bits(reg))
510 			continue;
511 
512 		/* Mark overflow */
513 		__vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(i);
514 
515 		if (kvm_pmu_counter_can_chain(pmc))
516 			kvm_pmu_counter_increment(vcpu, BIT(i + 1),
517 						  ARMV8_PMUV3_PERFCTR_CHAIN);
518 	}
519 }
520 
521 /* Compute the sample period for a given counter value */
522 static u64 compute_period(struct kvm_pmc *pmc, u64 counter)
523 {
524 	u64 val;
525 
526 	if (kvm_pmc_is_64bit(pmc) && kvm_pmc_has_64bit_overflow(pmc))
527 		val = (-counter) & GENMASK(63, 0);
528 	else
529 		val = (-counter) & GENMASK(31, 0);
530 
531 	return val;
532 }
533 
534 /*
535  * When the perf event overflows, set the overflow status and inform the vcpu.
536  */
537 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
538 				  struct perf_sample_data *data,
539 				  struct pt_regs *regs)
540 {
541 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
542 	struct arm_pmu *cpu_pmu = to_arm_pmu(perf_event->pmu);
543 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
544 	int idx = pmc->idx;
545 	u64 period;
546 
547 	cpu_pmu->pmu.stop(perf_event, PERF_EF_UPDATE);
548 
549 	/*
550 	 * Reset the sample period to the architectural limit,
551 	 * i.e. the point where the counter overflows.
552 	 */
553 	period = compute_period(pmc, local64_read(&perf_event->count));
554 
555 	local64_set(&perf_event->hw.period_left, 0);
556 	perf_event->attr.sample_period = period;
557 	perf_event->hw.sample_period = period;
558 
559 	__vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(idx);
560 
561 	if (kvm_pmu_counter_can_chain(pmc))
562 		kvm_pmu_counter_increment(vcpu, BIT(idx + 1),
563 					  ARMV8_PMUV3_PERFCTR_CHAIN);
564 
565 	if (kvm_pmu_overflow_status(vcpu)) {
566 		kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
567 
568 		if (!in_nmi())
569 			kvm_vcpu_kick(vcpu);
570 		else
571 			irq_work_queue(&vcpu->arch.pmu.overflow_work);
572 	}
573 
574 	cpu_pmu->pmu.start(perf_event, PERF_EF_RELOAD);
575 }
576 
577 /**
578  * kvm_pmu_software_increment - do software increment
579  * @vcpu: The vcpu pointer
580  * @val: the value guest writes to PMSWINC register
581  */
582 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
583 {
584 	kvm_pmu_counter_increment(vcpu, val, ARMV8_PMUV3_PERFCTR_SW_INCR);
585 }
586 
587 /**
588  * kvm_pmu_handle_pmcr - handle PMCR register
589  * @vcpu: The vcpu pointer
590  * @val: the value guest writes to PMCR register
591  */
592 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
593 {
594 	int i;
595 
596 	/* Fixup PMCR_EL0 to reconcile the PMU version and the LP bit */
597 	if (!kvm_has_feat(vcpu->kvm, ID_AA64DFR0_EL1, PMUVer, V3P5))
598 		val &= ~ARMV8_PMU_PMCR_LP;
599 
600 	/* Request a reload of the PMU to enable/disable affected counters */
601 	if ((__vcpu_sys_reg(vcpu, PMCR_EL0) ^ val) & ARMV8_PMU_PMCR_E)
602 		kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
603 
604 	/* The reset bits don't indicate any state, and shouldn't be saved. */
605 	__vcpu_sys_reg(vcpu, PMCR_EL0) = val & ~(ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_P);
606 
607 	if (val & ARMV8_PMU_PMCR_C)
608 		kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
609 
610 	if (val & ARMV8_PMU_PMCR_P) {
611 		/*
612 		 * Unlike other PMU sysregs, the controls in PMCR_EL0 always apply
613 		 * to the 'guest' range of counters and never the 'hyp' range.
614 		 */
615 		unsigned long mask = kvm_pmu_implemented_counter_mask(vcpu) &
616 				     ~kvm_pmu_hyp_counter_mask(vcpu) &
617 				     ~BIT(ARMV8_PMU_CYCLE_IDX);
618 
619 		for_each_set_bit(i, &mask, 32)
620 			kvm_pmu_set_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, i), 0, true);
621 	}
622 }
623 
624 static bool kvm_pmu_counter_is_enabled(struct kvm_pmc *pmc)
625 {
626 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
627 	unsigned int mdcr = __vcpu_sys_reg(vcpu, MDCR_EL2);
628 
629 	if (!(__vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(pmc->idx)))
630 		return false;
631 
632 	if (kvm_pmu_counter_is_hyp(vcpu, pmc->idx))
633 		return mdcr & MDCR_EL2_HPME;
634 
635 	return kvm_vcpu_read_pmcr(vcpu) & ARMV8_PMU_PMCR_E;
636 }
637 
638 static bool kvm_pmc_counts_at_el0(struct kvm_pmc *pmc)
639 {
640 	u64 evtreg = kvm_pmc_read_evtreg(pmc);
641 	bool nsu = evtreg & ARMV8_PMU_EXCLUDE_NS_EL0;
642 	bool u = evtreg & ARMV8_PMU_EXCLUDE_EL0;
643 
644 	return u == nsu;
645 }
646 
647 static bool kvm_pmc_counts_at_el1(struct kvm_pmc *pmc)
648 {
649 	u64 evtreg = kvm_pmc_read_evtreg(pmc);
650 	bool nsk = evtreg & ARMV8_PMU_EXCLUDE_NS_EL1;
651 	bool p = evtreg & ARMV8_PMU_EXCLUDE_EL1;
652 
653 	return p == nsk;
654 }
655 
656 static bool kvm_pmc_counts_at_el2(struct kvm_pmc *pmc)
657 {
658 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
659 	u64 mdcr = __vcpu_sys_reg(vcpu, MDCR_EL2);
660 
661 	if (!kvm_pmu_counter_is_hyp(vcpu, pmc->idx) && (mdcr & MDCR_EL2_HPMD))
662 		return false;
663 
664 	return kvm_pmc_read_evtreg(pmc) & ARMV8_PMU_INCLUDE_EL2;
665 }
666 
667 static int kvm_map_pmu_event(struct kvm *kvm, unsigned int eventsel)
668 {
669 	struct arm_pmu *pmu = kvm->arch.arm_pmu;
670 
671 	/*
672 	 * The CPU PMU likely isn't PMUv3; let the driver provide a mapping
673 	 * for the guest's PMUv3 event ID.
674 	 */
675 	if (unlikely(pmu->map_pmuv3_event))
676 		return pmu->map_pmuv3_event(eventsel);
677 
678 	return eventsel;
679 }
680 
681 /**
682  * kvm_pmu_create_perf_event - create a perf event for a counter
683  * @pmc: Counter context
684  */
685 static void kvm_pmu_create_perf_event(struct kvm_pmc *pmc)
686 {
687 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
688 	struct arm_pmu *arm_pmu = vcpu->kvm->arch.arm_pmu;
689 	struct perf_event *event;
690 	struct perf_event_attr attr;
691 	int eventsel;
692 	u64 evtreg;
693 
694 	evtreg = kvm_pmc_read_evtreg(pmc);
695 
696 	kvm_pmu_stop_counter(pmc);
697 	if (pmc->idx == ARMV8_PMU_CYCLE_IDX)
698 		eventsel = ARMV8_PMUV3_PERFCTR_CPU_CYCLES;
699 	else
700 		eventsel = evtreg & kvm_pmu_event_mask(vcpu->kvm);
701 
702 	/*
703 	 * Neither SW increment nor chained events need to be backed
704 	 * by a perf event.
705 	 */
706 	if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR ||
707 	    eventsel == ARMV8_PMUV3_PERFCTR_CHAIN)
708 		return;
709 
710 	/*
711 	 * If we have a filter in place and that the event isn't allowed, do
712 	 * not install a perf event either.
713 	 */
714 	if (vcpu->kvm->arch.pmu_filter &&
715 	    !test_bit(eventsel, vcpu->kvm->arch.pmu_filter))
716 		return;
717 
718 	/*
719 	 * Don't create an event if we're running on hardware that requires
720 	 * PMUv3 event translation and we couldn't find a valid mapping.
721 	 */
722 	eventsel = kvm_map_pmu_event(vcpu->kvm, eventsel);
723 	if (eventsel < 0)
724 		return;
725 
726 	memset(&attr, 0, sizeof(struct perf_event_attr));
727 	attr.type = arm_pmu->pmu.type;
728 	attr.size = sizeof(attr);
729 	attr.pinned = 1;
730 	attr.disabled = !kvm_pmu_counter_is_enabled(pmc);
731 	attr.exclude_user = !kvm_pmc_counts_at_el0(pmc);
732 	attr.exclude_hv = 1; /* Don't count EL2 events */
733 	attr.exclude_host = 1; /* Don't count host events */
734 	attr.config = eventsel;
735 
736 	/*
737 	 * Filter events at EL1 (i.e. vEL2) when in a hyp context based on the
738 	 * guest's EL2 filter.
739 	 */
740 	if (unlikely(is_hyp_ctxt(vcpu)))
741 		attr.exclude_kernel = !kvm_pmc_counts_at_el2(pmc);
742 	else
743 		attr.exclude_kernel = !kvm_pmc_counts_at_el1(pmc);
744 
745 	/*
746 	 * If counting with a 64bit counter, advertise it to the perf
747 	 * code, carefully dealing with the initial sample period
748 	 * which also depends on the overflow.
749 	 */
750 	if (kvm_pmc_is_64bit(pmc))
751 		attr.config1 |= PERF_ATTR_CFG1_COUNTER_64BIT;
752 
753 	attr.sample_period = compute_period(pmc, kvm_pmu_get_pmc_value(pmc));
754 
755 	event = perf_event_create_kernel_counter(&attr, -1, current,
756 						 kvm_pmu_perf_overflow, pmc);
757 
758 	if (IS_ERR(event)) {
759 		pr_err_once("kvm: pmu event creation failed %ld\n",
760 			    PTR_ERR(event));
761 		return;
762 	}
763 
764 	pmc->perf_event = event;
765 }
766 
767 /**
768  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
769  * @vcpu: The vcpu pointer
770  * @data: The data guest writes to PMXEVTYPER_EL0
771  * @select_idx: The number of selected counter
772  *
773  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
774  * event with given hardware event number. Here we call perf_event API to
775  * emulate this action and create a kernel perf event for it.
776  */
777 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
778 				    u64 select_idx)
779 {
780 	struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, select_idx);
781 	u64 reg;
782 
783 	reg = counter_index_to_evtreg(pmc->idx);
784 	__vcpu_sys_reg(vcpu, reg) = data & kvm_pmu_evtyper_mask(vcpu->kvm);
785 
786 	kvm_pmu_create_perf_event(pmc);
787 }
788 
789 void kvm_host_pmu_init(struct arm_pmu *pmu)
790 {
791 	struct arm_pmu_entry *entry;
792 
793 	/*
794 	 * Check the sanitised PMU version for the system, as KVM does not
795 	 * support implementations where PMUv3 exists on a subset of CPUs.
796 	 */
797 	if (!pmuv3_implemented(kvm_arm_pmu_get_pmuver_limit()))
798 		return;
799 
800 	guard(mutex)(&arm_pmus_lock);
801 
802 	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
803 	if (!entry)
804 		return;
805 
806 	entry->arm_pmu = pmu;
807 	list_add_tail(&entry->entry, &arm_pmus);
808 }
809 
810 static struct arm_pmu *kvm_pmu_probe_armpmu(void)
811 {
812 	struct arm_pmu_entry *entry;
813 	struct arm_pmu *pmu;
814 	int cpu;
815 
816 	guard(mutex)(&arm_pmus_lock);
817 
818 	/*
819 	 * It is safe to use a stale cpu to iterate the list of PMUs so long as
820 	 * the same value is used for the entirety of the loop. Given this, and
821 	 * the fact that no percpu data is used for the lookup there is no need
822 	 * to disable preemption.
823 	 *
824 	 * It is still necessary to get a valid cpu, though, to probe for the
825 	 * default PMU instance as userspace is not required to specify a PMU
826 	 * type. In order to uphold the preexisting behavior KVM selects the
827 	 * PMU instance for the core during vcpu init. A dependent use
828 	 * case would be a user with disdain of all things big.LITTLE that
829 	 * affines the VMM to a particular cluster of cores.
830 	 *
831 	 * In any case, userspace should just do the sane thing and use the UAPI
832 	 * to select a PMU type directly. But, be wary of the baggage being
833 	 * carried here.
834 	 */
835 	cpu = raw_smp_processor_id();
836 	list_for_each_entry(entry, &arm_pmus, entry) {
837 		pmu = entry->arm_pmu;
838 
839 		if (cpumask_test_cpu(cpu, &pmu->supported_cpus))
840 			return pmu;
841 	}
842 
843 	return NULL;
844 }
845 
846 static u64 __compute_pmceid(struct arm_pmu *pmu, bool pmceid1)
847 {
848 	u32 hi[2], lo[2];
849 
850 	bitmap_to_arr32(lo, pmu->pmceid_bitmap, ARMV8_PMUV3_MAX_COMMON_EVENTS);
851 	bitmap_to_arr32(hi, pmu->pmceid_ext_bitmap, ARMV8_PMUV3_MAX_COMMON_EVENTS);
852 
853 	return ((u64)hi[pmceid1] << 32) | lo[pmceid1];
854 }
855 
856 static u64 compute_pmceid0(struct arm_pmu *pmu)
857 {
858 	u64 val = __compute_pmceid(pmu, 0);
859 
860 	/* always support SW_INCR */
861 	val |= BIT(ARMV8_PMUV3_PERFCTR_SW_INCR);
862 	/* always support CHAIN */
863 	val |= BIT(ARMV8_PMUV3_PERFCTR_CHAIN);
864 	return val;
865 }
866 
867 static u64 compute_pmceid1(struct arm_pmu *pmu)
868 {
869 	u64 val = __compute_pmceid(pmu, 1);
870 
871 	/*
872 	 * Don't advertise STALL_SLOT*, as PMMIR_EL0 is handled
873 	 * as RAZ
874 	 */
875 	val &= ~(BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT - 32) |
876 		 BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT_FRONTEND - 32) |
877 		 BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT_BACKEND - 32));
878 	return val;
879 }
880 
881 u64 kvm_pmu_get_pmceid(struct kvm_vcpu *vcpu, bool pmceid1)
882 {
883 	struct arm_pmu *cpu_pmu = vcpu->kvm->arch.arm_pmu;
884 	unsigned long *bmap = vcpu->kvm->arch.pmu_filter;
885 	u64 val, mask = 0;
886 	int base, i, nr_events;
887 
888 	if (!pmceid1) {
889 		val = compute_pmceid0(cpu_pmu);
890 		base = 0;
891 	} else {
892 		val = compute_pmceid1(cpu_pmu);
893 		base = 32;
894 	}
895 
896 	if (!bmap)
897 		return val;
898 
899 	nr_events = kvm_pmu_event_mask(vcpu->kvm) + 1;
900 
901 	for (i = 0; i < 32; i += 8) {
902 		u64 byte;
903 
904 		byte = bitmap_get_value8(bmap, base + i);
905 		mask |= byte << i;
906 		if (nr_events >= (0x4000 + base + 32)) {
907 			byte = bitmap_get_value8(bmap, 0x4000 + base + i);
908 			mask |= byte << (32 + i);
909 		}
910 	}
911 
912 	return val & mask;
913 }
914 
915 void kvm_vcpu_reload_pmu(struct kvm_vcpu *vcpu)
916 {
917 	u64 mask = kvm_pmu_implemented_counter_mask(vcpu);
918 
919 	__vcpu_sys_reg(vcpu, PMOVSSET_EL0) &= mask;
920 	__vcpu_sys_reg(vcpu, PMINTENSET_EL1) &= mask;
921 	__vcpu_sys_reg(vcpu, PMCNTENSET_EL0) &= mask;
922 
923 	kvm_pmu_reprogram_counter_mask(vcpu, mask);
924 }
925 
926 int kvm_arm_pmu_v3_enable(struct kvm_vcpu *vcpu)
927 {
928 	if (!vcpu->arch.pmu.created)
929 		return -EINVAL;
930 
931 	/*
932 	 * A valid interrupt configuration for the PMU is either to have a
933 	 * properly configured interrupt number and using an in-kernel
934 	 * irqchip, or to not have an in-kernel GIC and not set an IRQ.
935 	 */
936 	if (irqchip_in_kernel(vcpu->kvm)) {
937 		int irq = vcpu->arch.pmu.irq_num;
938 		/*
939 		 * If we are using an in-kernel vgic, at this point we know
940 		 * the vgic will be initialized, so we can check the PMU irq
941 		 * number against the dimensions of the vgic and make sure
942 		 * it's valid.
943 		 */
944 		if (!irq_is_ppi(irq) && !vgic_valid_spi(vcpu->kvm, irq))
945 			return -EINVAL;
946 	} else if (kvm_arm_pmu_irq_initialized(vcpu)) {
947 		   return -EINVAL;
948 	}
949 
950 	return 0;
951 }
952 
953 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
954 {
955 	if (irqchip_in_kernel(vcpu->kvm)) {
956 		int ret;
957 
958 		/*
959 		 * If using the PMU with an in-kernel virtual GIC
960 		 * implementation, we require the GIC to be already
961 		 * initialized when initializing the PMU.
962 		 */
963 		if (!vgic_initialized(vcpu->kvm))
964 			return -ENODEV;
965 
966 		if (!kvm_arm_pmu_irq_initialized(vcpu))
967 			return -ENXIO;
968 
969 		ret = kvm_vgic_set_owner(vcpu, vcpu->arch.pmu.irq_num,
970 					 &vcpu->arch.pmu);
971 		if (ret)
972 			return ret;
973 	}
974 
975 	init_irq_work(&vcpu->arch.pmu.overflow_work,
976 		      kvm_pmu_perf_overflow_notify_vcpu);
977 
978 	vcpu->arch.pmu.created = true;
979 	return 0;
980 }
981 
982 /*
983  * For one VM the interrupt type must be same for each vcpu.
984  * As a PPI, the interrupt number is the same for all vcpus,
985  * while as an SPI it must be a separate number per vcpu.
986  */
987 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
988 {
989 	unsigned long i;
990 	struct kvm_vcpu *vcpu;
991 
992 	kvm_for_each_vcpu(i, vcpu, kvm) {
993 		if (!kvm_arm_pmu_irq_initialized(vcpu))
994 			continue;
995 
996 		if (irq_is_ppi(irq)) {
997 			if (vcpu->arch.pmu.irq_num != irq)
998 				return false;
999 		} else {
1000 			if (vcpu->arch.pmu.irq_num == irq)
1001 				return false;
1002 		}
1003 	}
1004 
1005 	return true;
1006 }
1007 
1008 /**
1009  * kvm_arm_pmu_get_max_counters - Return the max number of PMU counters.
1010  * @kvm: The kvm pointer
1011  */
1012 u8 kvm_arm_pmu_get_max_counters(struct kvm *kvm)
1013 {
1014 	struct arm_pmu *arm_pmu = kvm->arch.arm_pmu;
1015 
1016 	/*
1017 	 * PMUv3 requires that all event counters are capable of counting any
1018 	 * event, though the same may not be true of non-PMUv3 hardware.
1019 	 */
1020 	if (cpus_have_final_cap(ARM64_WORKAROUND_PMUV3_IMPDEF_TRAPS))
1021 		return 1;
1022 
1023 	/*
1024 	 * The arm_pmu->cntr_mask considers the fixed counter(s) as well.
1025 	 * Ignore those and return only the general-purpose counters.
1026 	 */
1027 	return bitmap_weight(arm_pmu->cntr_mask, ARMV8_PMU_MAX_GENERAL_COUNTERS);
1028 }
1029 
1030 static void kvm_arm_set_pmu(struct kvm *kvm, struct arm_pmu *arm_pmu)
1031 {
1032 	lockdep_assert_held(&kvm->arch.config_lock);
1033 
1034 	kvm->arch.arm_pmu = arm_pmu;
1035 	kvm->arch.pmcr_n = kvm_arm_pmu_get_max_counters(kvm);
1036 }
1037 
1038 /**
1039  * kvm_arm_set_default_pmu - No PMU set, get the default one.
1040  * @kvm: The kvm pointer
1041  *
1042  * The observant among you will notice that the supported_cpus
1043  * mask does not get updated for the default PMU even though it
1044  * is quite possible the selected instance supports only a
1045  * subset of cores in the system. This is intentional, and
1046  * upholds the preexisting behavior on heterogeneous systems
1047  * where vCPUs can be scheduled on any core but the guest
1048  * counters could stop working.
1049  */
1050 int kvm_arm_set_default_pmu(struct kvm *kvm)
1051 {
1052 	struct arm_pmu *arm_pmu = kvm_pmu_probe_armpmu();
1053 
1054 	if (!arm_pmu)
1055 		return -ENODEV;
1056 
1057 	kvm_arm_set_pmu(kvm, arm_pmu);
1058 	return 0;
1059 }
1060 
1061 static int kvm_arm_pmu_v3_set_pmu(struct kvm_vcpu *vcpu, int pmu_id)
1062 {
1063 	struct kvm *kvm = vcpu->kvm;
1064 	struct arm_pmu_entry *entry;
1065 	struct arm_pmu *arm_pmu;
1066 	int ret = -ENXIO;
1067 
1068 	lockdep_assert_held(&kvm->arch.config_lock);
1069 	mutex_lock(&arm_pmus_lock);
1070 
1071 	list_for_each_entry(entry, &arm_pmus, entry) {
1072 		arm_pmu = entry->arm_pmu;
1073 		if (arm_pmu->pmu.type == pmu_id) {
1074 			if (kvm_vm_has_ran_once(kvm) ||
1075 			    (kvm->arch.pmu_filter && kvm->arch.arm_pmu != arm_pmu)) {
1076 				ret = -EBUSY;
1077 				break;
1078 			}
1079 
1080 			kvm_arm_set_pmu(kvm, arm_pmu);
1081 			cpumask_copy(kvm->arch.supported_cpus, &arm_pmu->supported_cpus);
1082 			ret = 0;
1083 			break;
1084 		}
1085 	}
1086 
1087 	mutex_unlock(&arm_pmus_lock);
1088 	return ret;
1089 }
1090 
1091 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1092 {
1093 	struct kvm *kvm = vcpu->kvm;
1094 
1095 	lockdep_assert_held(&kvm->arch.config_lock);
1096 
1097 	if (!kvm_vcpu_has_pmu(vcpu))
1098 		return -ENODEV;
1099 
1100 	if (vcpu->arch.pmu.created)
1101 		return -EBUSY;
1102 
1103 	switch (attr->attr) {
1104 	case KVM_ARM_VCPU_PMU_V3_IRQ: {
1105 		int __user *uaddr = (int __user *)(long)attr->addr;
1106 		int irq;
1107 
1108 		if (!irqchip_in_kernel(kvm))
1109 			return -EINVAL;
1110 
1111 		if (get_user(irq, uaddr))
1112 			return -EFAULT;
1113 
1114 		/* The PMU overflow interrupt can be a PPI or a valid SPI. */
1115 		if (!(irq_is_ppi(irq) || irq_is_spi(irq)))
1116 			return -EINVAL;
1117 
1118 		if (!pmu_irq_is_valid(kvm, irq))
1119 			return -EINVAL;
1120 
1121 		if (kvm_arm_pmu_irq_initialized(vcpu))
1122 			return -EBUSY;
1123 
1124 		kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
1125 		vcpu->arch.pmu.irq_num = irq;
1126 		return 0;
1127 	}
1128 	case KVM_ARM_VCPU_PMU_V3_FILTER: {
1129 		u8 pmuver = kvm_arm_pmu_get_pmuver_limit();
1130 		struct kvm_pmu_event_filter __user *uaddr;
1131 		struct kvm_pmu_event_filter filter;
1132 		int nr_events;
1133 
1134 		/*
1135 		 * Allow userspace to specify an event filter for the entire
1136 		 * event range supported by PMUVer of the hardware, rather
1137 		 * than the guest's PMUVer for KVM backward compatibility.
1138 		 */
1139 		nr_events = __kvm_pmu_event_mask(pmuver) + 1;
1140 
1141 		uaddr = (struct kvm_pmu_event_filter __user *)(long)attr->addr;
1142 
1143 		if (copy_from_user(&filter, uaddr, sizeof(filter)))
1144 			return -EFAULT;
1145 
1146 		if (((u32)filter.base_event + filter.nevents) > nr_events ||
1147 		    (filter.action != KVM_PMU_EVENT_ALLOW &&
1148 		     filter.action != KVM_PMU_EVENT_DENY))
1149 			return -EINVAL;
1150 
1151 		if (kvm_vm_has_ran_once(kvm))
1152 			return -EBUSY;
1153 
1154 		if (!kvm->arch.pmu_filter) {
1155 			kvm->arch.pmu_filter = bitmap_alloc(nr_events, GFP_KERNEL_ACCOUNT);
1156 			if (!kvm->arch.pmu_filter)
1157 				return -ENOMEM;
1158 
1159 			/*
1160 			 * The default depends on the first applied filter.
1161 			 * If it allows events, the default is to deny.
1162 			 * Conversely, if the first filter denies a set of
1163 			 * events, the default is to allow.
1164 			 */
1165 			if (filter.action == KVM_PMU_EVENT_ALLOW)
1166 				bitmap_zero(kvm->arch.pmu_filter, nr_events);
1167 			else
1168 				bitmap_fill(kvm->arch.pmu_filter, nr_events);
1169 		}
1170 
1171 		if (filter.action == KVM_PMU_EVENT_ALLOW)
1172 			bitmap_set(kvm->arch.pmu_filter, filter.base_event, filter.nevents);
1173 		else
1174 			bitmap_clear(kvm->arch.pmu_filter, filter.base_event, filter.nevents);
1175 
1176 		return 0;
1177 	}
1178 	case KVM_ARM_VCPU_PMU_V3_SET_PMU: {
1179 		int __user *uaddr = (int __user *)(long)attr->addr;
1180 		int pmu_id;
1181 
1182 		if (get_user(pmu_id, uaddr))
1183 			return -EFAULT;
1184 
1185 		return kvm_arm_pmu_v3_set_pmu(vcpu, pmu_id);
1186 	}
1187 	case KVM_ARM_VCPU_PMU_V3_INIT:
1188 		return kvm_arm_pmu_v3_init(vcpu);
1189 	}
1190 
1191 	return -ENXIO;
1192 }
1193 
1194 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1195 {
1196 	switch (attr->attr) {
1197 	case KVM_ARM_VCPU_PMU_V3_IRQ: {
1198 		int __user *uaddr = (int __user *)(long)attr->addr;
1199 		int irq;
1200 
1201 		if (!irqchip_in_kernel(vcpu->kvm))
1202 			return -EINVAL;
1203 
1204 		if (!kvm_vcpu_has_pmu(vcpu))
1205 			return -ENODEV;
1206 
1207 		if (!kvm_arm_pmu_irq_initialized(vcpu))
1208 			return -ENXIO;
1209 
1210 		irq = vcpu->arch.pmu.irq_num;
1211 		return put_user(irq, uaddr);
1212 	}
1213 	}
1214 
1215 	return -ENXIO;
1216 }
1217 
1218 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1219 {
1220 	switch (attr->attr) {
1221 	case KVM_ARM_VCPU_PMU_V3_IRQ:
1222 	case KVM_ARM_VCPU_PMU_V3_INIT:
1223 	case KVM_ARM_VCPU_PMU_V3_FILTER:
1224 	case KVM_ARM_VCPU_PMU_V3_SET_PMU:
1225 		if (kvm_vcpu_has_pmu(vcpu))
1226 			return 0;
1227 	}
1228 
1229 	return -ENXIO;
1230 }
1231 
1232 u8 kvm_arm_pmu_get_pmuver_limit(void)
1233 {
1234 	unsigned int pmuver;
1235 
1236 	pmuver = SYS_FIELD_GET(ID_AA64DFR0_EL1, PMUVer,
1237 			       read_sanitised_ftr_reg(SYS_ID_AA64DFR0_EL1));
1238 
1239 	/*
1240 	 * Spoof a barebones PMUv3 implementation if the system supports IMPDEF
1241 	 * traps of the PMUv3 sysregs
1242 	 */
1243 	if (cpus_have_final_cap(ARM64_WORKAROUND_PMUV3_IMPDEF_TRAPS))
1244 		return ID_AA64DFR0_EL1_PMUVer_IMP;
1245 
1246 	/*
1247 	 * Otherwise, treat IMPLEMENTATION DEFINED functionality as
1248 	 * unimplemented
1249 	 */
1250 	if (pmuver == ID_AA64DFR0_EL1_PMUVer_IMP_DEF)
1251 		return 0;
1252 
1253 	return min(pmuver, ID_AA64DFR0_EL1_PMUVer_V3P5);
1254 }
1255 
1256 /**
1257  * kvm_vcpu_read_pmcr - Read PMCR_EL0 register for the vCPU
1258  * @vcpu: The vcpu pointer
1259  */
1260 u64 kvm_vcpu_read_pmcr(struct kvm_vcpu *vcpu)
1261 {
1262 	u64 pmcr = __vcpu_sys_reg(vcpu, PMCR_EL0);
1263 
1264 	return u64_replace_bits(pmcr, vcpu->kvm->arch.pmcr_n, ARMV8_PMU_PMCR_N);
1265 }
1266 
1267 void kvm_pmu_nested_transition(struct kvm_vcpu *vcpu)
1268 {
1269 	bool reprogrammed = false;
1270 	unsigned long mask;
1271 	int i;
1272 
1273 	mask = __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
1274 	for_each_set_bit(i, &mask, 32) {
1275 		struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
1276 
1277 		/*
1278 		 * We only need to reconfigure events where the filter is
1279 		 * different at EL1 vs. EL2, as we're multiplexing the true EL1
1280 		 * event filter bit for nested.
1281 		 */
1282 		if (kvm_pmc_counts_at_el1(pmc) == kvm_pmc_counts_at_el2(pmc))
1283 			continue;
1284 
1285 		kvm_pmu_create_perf_event(pmc);
1286 		reprogrammed = true;
1287 	}
1288 
1289 	if (reprogrammed)
1290 		kvm_vcpu_pmu_restore_guest(vcpu);
1291 }
1292