xref: /linux/arch/arm64/kvm/pmu-emul.c (revision fde668244d1d8d490b5b9daf53fe4f92a6751773)
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/perf_event.h>
11 #include <linux/perf/arm_pmu.h>
12 #include <linux/uaccess.h>
13 #include <asm/kvm_emulate.h>
14 #include <kvm/arm_pmu.h>
15 #include <kvm/arm_vgic.h>
16 
17 static void kvm_pmu_create_perf_event(struct kvm_vcpu *vcpu, u64 select_idx);
18 static void kvm_pmu_update_pmc_chained(struct kvm_vcpu *vcpu, u64 select_idx);
19 static void kvm_pmu_stop_counter(struct kvm_vcpu *vcpu, struct kvm_pmc *pmc);
20 
21 #define PERF_ATTR_CFG1_KVM_PMU_CHAINED 0x1
22 
23 static u32 kvm_pmu_event_mask(struct kvm *kvm)
24 {
25 	switch (kvm->arch.pmuver) {
26 	case 1:			/* ARMv8.0 */
27 		return GENMASK(9, 0);
28 	case 4:			/* ARMv8.1 */
29 	case 5:			/* ARMv8.4 */
30 	case 6:			/* ARMv8.5 */
31 		return GENMASK(15, 0);
32 	default:		/* Shouldn't be here, just for sanity */
33 		WARN_ONCE(1, "Unknown PMU version %d\n", kvm->arch.pmuver);
34 		return 0;
35 	}
36 }
37 
38 /**
39  * kvm_pmu_idx_is_64bit - determine if select_idx is a 64bit counter
40  * @vcpu: The vcpu pointer
41  * @select_idx: The counter index
42  */
43 static bool kvm_pmu_idx_is_64bit(struct kvm_vcpu *vcpu, u64 select_idx)
44 {
45 	return (select_idx == ARMV8_PMU_CYCLE_IDX &&
46 		__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_LC);
47 }
48 
49 static struct kvm_vcpu *kvm_pmc_to_vcpu(struct kvm_pmc *pmc)
50 {
51 	struct kvm_pmu *pmu;
52 	struct kvm_vcpu_arch *vcpu_arch;
53 
54 	pmc -= pmc->idx;
55 	pmu = container_of(pmc, struct kvm_pmu, pmc[0]);
56 	vcpu_arch = container_of(pmu, struct kvm_vcpu_arch, pmu);
57 	return container_of(vcpu_arch, struct kvm_vcpu, arch);
58 }
59 
60 /**
61  * kvm_pmu_pmc_is_chained - determine if the pmc is chained
62  * @pmc: The PMU counter pointer
63  */
64 static bool kvm_pmu_pmc_is_chained(struct kvm_pmc *pmc)
65 {
66 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
67 
68 	return test_bit(pmc->idx >> 1, vcpu->arch.pmu.chained);
69 }
70 
71 /**
72  * kvm_pmu_idx_is_high_counter - determine if select_idx is a high/low counter
73  * @select_idx: The counter index
74  */
75 static bool kvm_pmu_idx_is_high_counter(u64 select_idx)
76 {
77 	return select_idx & 0x1;
78 }
79 
80 /**
81  * kvm_pmu_get_canonical_pmc - obtain the canonical pmc
82  * @pmc: The PMU counter pointer
83  *
84  * When a pair of PMCs are chained together we use the low counter (canonical)
85  * to hold the underlying perf event.
86  */
87 static struct kvm_pmc *kvm_pmu_get_canonical_pmc(struct kvm_pmc *pmc)
88 {
89 	if (kvm_pmu_pmc_is_chained(pmc) &&
90 	    kvm_pmu_idx_is_high_counter(pmc->idx))
91 		return pmc - 1;
92 
93 	return pmc;
94 }
95 static struct kvm_pmc *kvm_pmu_get_alternate_pmc(struct kvm_pmc *pmc)
96 {
97 	if (kvm_pmu_idx_is_high_counter(pmc->idx))
98 		return pmc - 1;
99 	else
100 		return pmc + 1;
101 }
102 
103 /**
104  * kvm_pmu_idx_has_chain_evtype - determine if the event type is chain
105  * @vcpu: The vcpu pointer
106  * @select_idx: The counter index
107  */
108 static bool kvm_pmu_idx_has_chain_evtype(struct kvm_vcpu *vcpu, u64 select_idx)
109 {
110 	u64 eventsel, reg;
111 
112 	select_idx |= 0x1;
113 
114 	if (select_idx == ARMV8_PMU_CYCLE_IDX)
115 		return false;
116 
117 	reg = PMEVTYPER0_EL0 + select_idx;
118 	eventsel = __vcpu_sys_reg(vcpu, reg) & kvm_pmu_event_mask(vcpu->kvm);
119 
120 	return eventsel == ARMV8_PMUV3_PERFCTR_CHAIN;
121 }
122 
123 /**
124  * kvm_pmu_get_pair_counter_value - get PMU counter value
125  * @vcpu: The vcpu pointer
126  * @pmc: The PMU counter pointer
127  */
128 static u64 kvm_pmu_get_pair_counter_value(struct kvm_vcpu *vcpu,
129 					  struct kvm_pmc *pmc)
130 {
131 	u64 counter, counter_high, reg, enabled, running;
132 
133 	if (kvm_pmu_pmc_is_chained(pmc)) {
134 		pmc = kvm_pmu_get_canonical_pmc(pmc);
135 		reg = PMEVCNTR0_EL0 + pmc->idx;
136 
137 		counter = __vcpu_sys_reg(vcpu, reg);
138 		counter_high = __vcpu_sys_reg(vcpu, reg + 1);
139 
140 		counter = lower_32_bits(counter) | (counter_high << 32);
141 	} else {
142 		reg = (pmc->idx == ARMV8_PMU_CYCLE_IDX)
143 		      ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + pmc->idx;
144 		counter = __vcpu_sys_reg(vcpu, reg);
145 	}
146 
147 	/*
148 	 * The real counter value is equal to the value of counter register plus
149 	 * the value perf event counts.
150 	 */
151 	if (pmc->perf_event)
152 		counter += perf_event_read_value(pmc->perf_event, &enabled,
153 						 &running);
154 
155 	return counter;
156 }
157 
158 /**
159  * kvm_pmu_get_counter_value - get PMU counter value
160  * @vcpu: The vcpu pointer
161  * @select_idx: The counter index
162  */
163 u64 kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u64 select_idx)
164 {
165 	u64 counter;
166 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
167 	struct kvm_pmc *pmc = &pmu->pmc[select_idx];
168 
169 	counter = kvm_pmu_get_pair_counter_value(vcpu, pmc);
170 
171 	if (kvm_pmu_pmc_is_chained(pmc) &&
172 	    kvm_pmu_idx_is_high_counter(select_idx))
173 		counter = upper_32_bits(counter);
174 	else if (select_idx != ARMV8_PMU_CYCLE_IDX)
175 		counter = lower_32_bits(counter);
176 
177 	return counter;
178 }
179 
180 /**
181  * kvm_pmu_set_counter_value - set PMU counter value
182  * @vcpu: The vcpu pointer
183  * @select_idx: The counter index
184  * @val: The counter value
185  */
186 void kvm_pmu_set_counter_value(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
187 {
188 	u64 reg;
189 
190 	reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
191 	      ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + select_idx;
192 	__vcpu_sys_reg(vcpu, reg) += (s64)val - kvm_pmu_get_counter_value(vcpu, select_idx);
193 
194 	/* Recreate the perf event to reflect the updated sample_period */
195 	kvm_pmu_create_perf_event(vcpu, select_idx);
196 }
197 
198 /**
199  * kvm_pmu_release_perf_event - remove the perf event
200  * @pmc: The PMU counter pointer
201  */
202 static void kvm_pmu_release_perf_event(struct kvm_pmc *pmc)
203 {
204 	pmc = kvm_pmu_get_canonical_pmc(pmc);
205 	if (pmc->perf_event) {
206 		perf_event_disable(pmc->perf_event);
207 		perf_event_release_kernel(pmc->perf_event);
208 		pmc->perf_event = NULL;
209 	}
210 }
211 
212 /**
213  * kvm_pmu_stop_counter - stop PMU counter
214  * @pmc: The PMU counter pointer
215  *
216  * If this counter has been configured to monitor some event, release it here.
217  */
218 static void kvm_pmu_stop_counter(struct kvm_vcpu *vcpu, struct kvm_pmc *pmc)
219 {
220 	u64 counter, reg, val;
221 
222 	pmc = kvm_pmu_get_canonical_pmc(pmc);
223 	if (!pmc->perf_event)
224 		return;
225 
226 	counter = kvm_pmu_get_pair_counter_value(vcpu, pmc);
227 
228 	if (pmc->idx == ARMV8_PMU_CYCLE_IDX) {
229 		reg = PMCCNTR_EL0;
230 		val = counter;
231 	} else {
232 		reg = PMEVCNTR0_EL0 + pmc->idx;
233 		val = lower_32_bits(counter);
234 	}
235 
236 	__vcpu_sys_reg(vcpu, reg) = val;
237 
238 	if (kvm_pmu_pmc_is_chained(pmc))
239 		__vcpu_sys_reg(vcpu, reg + 1) = upper_32_bits(counter);
240 
241 	kvm_pmu_release_perf_event(pmc);
242 }
243 
244 /**
245  * kvm_pmu_vcpu_init - assign pmu counter idx for cpu
246  * @vcpu: The vcpu pointer
247  *
248  */
249 void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)
250 {
251 	int i;
252 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
253 
254 	for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
255 		pmu->pmc[i].idx = i;
256 }
257 
258 /**
259  * kvm_pmu_vcpu_reset - reset pmu state for cpu
260  * @vcpu: The vcpu pointer
261  *
262  */
263 void kvm_pmu_vcpu_reset(struct kvm_vcpu *vcpu)
264 {
265 	unsigned long mask = kvm_pmu_valid_counter_mask(vcpu);
266 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
267 	int i;
268 
269 	for_each_set_bit(i, &mask, 32)
270 		kvm_pmu_stop_counter(vcpu, &pmu->pmc[i]);
271 
272 	bitmap_zero(vcpu->arch.pmu.chained, ARMV8_PMU_MAX_COUNTER_PAIRS);
273 }
274 
275 /**
276  * kvm_pmu_vcpu_destroy - free perf event of PMU for cpu
277  * @vcpu: The vcpu pointer
278  *
279  */
280 void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
281 {
282 	int i;
283 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
284 
285 	for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
286 		kvm_pmu_release_perf_event(&pmu->pmc[i]);
287 	irq_work_sync(&vcpu->arch.pmu.overflow_work);
288 }
289 
290 u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu)
291 {
292 	u64 val = __vcpu_sys_reg(vcpu, PMCR_EL0) >> ARMV8_PMU_PMCR_N_SHIFT;
293 
294 	val &= ARMV8_PMU_PMCR_N_MASK;
295 	if (val == 0)
296 		return BIT(ARMV8_PMU_CYCLE_IDX);
297 	else
298 		return GENMASK(val - 1, 0) | BIT(ARMV8_PMU_CYCLE_IDX);
299 }
300 
301 /**
302  * kvm_pmu_enable_counter_mask - enable selected PMU counters
303  * @vcpu: The vcpu pointer
304  * @val: the value guest writes to PMCNTENSET register
305  *
306  * Call perf_event_enable to start counting the perf event
307  */
308 void kvm_pmu_enable_counter_mask(struct kvm_vcpu *vcpu, u64 val)
309 {
310 	int i;
311 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
312 	struct kvm_pmc *pmc;
313 
314 	if (!(__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) || !val)
315 		return;
316 
317 	for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
318 		if (!(val & BIT(i)))
319 			continue;
320 
321 		pmc = &pmu->pmc[i];
322 
323 		/* A change in the enable state may affect the chain state */
324 		kvm_pmu_update_pmc_chained(vcpu, i);
325 		kvm_pmu_create_perf_event(vcpu, i);
326 
327 		/* At this point, pmc must be the canonical */
328 		if (pmc->perf_event) {
329 			perf_event_enable(pmc->perf_event);
330 			if (pmc->perf_event->state != PERF_EVENT_STATE_ACTIVE)
331 				kvm_debug("fail to enable perf event\n");
332 		}
333 	}
334 }
335 
336 /**
337  * kvm_pmu_disable_counter_mask - disable selected PMU counters
338  * @vcpu: The vcpu pointer
339  * @val: the value guest writes to PMCNTENCLR register
340  *
341  * Call perf_event_disable to stop counting the perf event
342  */
343 void kvm_pmu_disable_counter_mask(struct kvm_vcpu *vcpu, u64 val)
344 {
345 	int i;
346 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
347 	struct kvm_pmc *pmc;
348 
349 	if (!val)
350 		return;
351 
352 	for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
353 		if (!(val & BIT(i)))
354 			continue;
355 
356 		pmc = &pmu->pmc[i];
357 
358 		/* A change in the enable state may affect the chain state */
359 		kvm_pmu_update_pmc_chained(vcpu, i);
360 		kvm_pmu_create_perf_event(vcpu, i);
361 
362 		/* At this point, pmc must be the canonical */
363 		if (pmc->perf_event)
364 			perf_event_disable(pmc->perf_event);
365 	}
366 }
367 
368 static u64 kvm_pmu_overflow_status(struct kvm_vcpu *vcpu)
369 {
370 	u64 reg = 0;
371 
372 	if ((__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E)) {
373 		reg = __vcpu_sys_reg(vcpu, PMOVSSET_EL0);
374 		reg &= __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
375 		reg &= __vcpu_sys_reg(vcpu, PMINTENSET_EL1);
376 		reg &= kvm_pmu_valid_counter_mask(vcpu);
377 	}
378 
379 	return reg;
380 }
381 
382 static void kvm_pmu_update_state(struct kvm_vcpu *vcpu)
383 {
384 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
385 	bool overflow;
386 
387 	if (!kvm_vcpu_has_pmu(vcpu))
388 		return;
389 
390 	overflow = !!kvm_pmu_overflow_status(vcpu);
391 	if (pmu->irq_level == overflow)
392 		return;
393 
394 	pmu->irq_level = overflow;
395 
396 	if (likely(irqchip_in_kernel(vcpu->kvm))) {
397 		int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
398 					      pmu->irq_num, overflow, pmu);
399 		WARN_ON(ret);
400 	}
401 }
402 
403 bool kvm_pmu_should_notify_user(struct kvm_vcpu *vcpu)
404 {
405 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
406 	struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
407 	bool run_level = sregs->device_irq_level & KVM_ARM_DEV_PMU;
408 
409 	if (likely(irqchip_in_kernel(vcpu->kvm)))
410 		return false;
411 
412 	return pmu->irq_level != run_level;
413 }
414 
415 /*
416  * Reflect the PMU overflow interrupt output level into the kvm_run structure
417  */
418 void kvm_pmu_update_run(struct kvm_vcpu *vcpu)
419 {
420 	struct kvm_sync_regs *regs = &vcpu->run->s.regs;
421 
422 	/* Populate the timer bitmap for user space */
423 	regs->device_irq_level &= ~KVM_ARM_DEV_PMU;
424 	if (vcpu->arch.pmu.irq_level)
425 		regs->device_irq_level |= KVM_ARM_DEV_PMU;
426 }
427 
428 /**
429  * kvm_pmu_flush_hwstate - flush pmu state to cpu
430  * @vcpu: The vcpu pointer
431  *
432  * Check if the PMU has overflowed while we were running in the host, and inject
433  * an interrupt if that was the case.
434  */
435 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
436 {
437 	kvm_pmu_update_state(vcpu);
438 }
439 
440 /**
441  * kvm_pmu_sync_hwstate - sync pmu state from cpu
442  * @vcpu: The vcpu pointer
443  *
444  * Check if the PMU has overflowed while we were running in the guest, and
445  * inject an interrupt if that was the case.
446  */
447 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
448 {
449 	kvm_pmu_update_state(vcpu);
450 }
451 
452 /**
453  * When perf interrupt is an NMI, we cannot safely notify the vcpu corresponding
454  * to the event.
455  * This is why we need a callback to do it once outside of the NMI context.
456  */
457 static void kvm_pmu_perf_overflow_notify_vcpu(struct irq_work *work)
458 {
459 	struct kvm_vcpu *vcpu;
460 	struct kvm_pmu *pmu;
461 
462 	pmu = container_of(work, struct kvm_pmu, overflow_work);
463 	vcpu = kvm_pmc_to_vcpu(pmu->pmc);
464 
465 	kvm_vcpu_kick(vcpu);
466 }
467 
468 /**
469  * When the perf event overflows, set the overflow status and inform the vcpu.
470  */
471 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
472 				  struct perf_sample_data *data,
473 				  struct pt_regs *regs)
474 {
475 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
476 	struct arm_pmu *cpu_pmu = to_arm_pmu(perf_event->pmu);
477 	struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
478 	int idx = pmc->idx;
479 	u64 period;
480 
481 	cpu_pmu->pmu.stop(perf_event, PERF_EF_UPDATE);
482 
483 	/*
484 	 * Reset the sample period to the architectural limit,
485 	 * i.e. the point where the counter overflows.
486 	 */
487 	period = -(local64_read(&perf_event->count));
488 
489 	if (!kvm_pmu_idx_is_64bit(vcpu, pmc->idx))
490 		period &= GENMASK(31, 0);
491 
492 	local64_set(&perf_event->hw.period_left, 0);
493 	perf_event->attr.sample_period = period;
494 	perf_event->hw.sample_period = period;
495 
496 	__vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(idx);
497 
498 	if (kvm_pmu_overflow_status(vcpu)) {
499 		kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
500 
501 		if (!in_nmi())
502 			kvm_vcpu_kick(vcpu);
503 		else
504 			irq_work_queue(&vcpu->arch.pmu.overflow_work);
505 	}
506 
507 	cpu_pmu->pmu.start(perf_event, PERF_EF_RELOAD);
508 }
509 
510 /**
511  * kvm_pmu_software_increment - do software increment
512  * @vcpu: The vcpu pointer
513  * @val: the value guest writes to PMSWINC register
514  */
515 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
516 {
517 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
518 	int i;
519 
520 	if (!(__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E))
521 		return;
522 
523 	/* Weed out disabled counters */
524 	val &= __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
525 
526 	for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++) {
527 		u64 type, reg;
528 
529 		if (!(val & BIT(i)))
530 			continue;
531 
532 		/* PMSWINC only applies to ... SW_INC! */
533 		type = __vcpu_sys_reg(vcpu, PMEVTYPER0_EL0 + i);
534 		type &= kvm_pmu_event_mask(vcpu->kvm);
535 		if (type != ARMV8_PMUV3_PERFCTR_SW_INCR)
536 			continue;
537 
538 		/* increment this even SW_INC counter */
539 		reg = __vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) + 1;
540 		reg = lower_32_bits(reg);
541 		__vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) = reg;
542 
543 		if (reg) /* no overflow on the low part */
544 			continue;
545 
546 		if (kvm_pmu_pmc_is_chained(&pmu->pmc[i])) {
547 			/* increment the high counter */
548 			reg = __vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i + 1) + 1;
549 			reg = lower_32_bits(reg);
550 			__vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i + 1) = reg;
551 			if (!reg) /* mark overflow on the high counter */
552 				__vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(i + 1);
553 		} else {
554 			/* mark overflow on low counter */
555 			__vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(i);
556 		}
557 	}
558 }
559 
560 /**
561  * kvm_pmu_handle_pmcr - handle PMCR register
562  * @vcpu: The vcpu pointer
563  * @val: the value guest writes to PMCR register
564  */
565 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
566 {
567 	unsigned long mask = kvm_pmu_valid_counter_mask(vcpu);
568 	int i;
569 
570 	if (val & ARMV8_PMU_PMCR_E) {
571 		kvm_pmu_enable_counter_mask(vcpu,
572 		       __vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & mask);
573 	} else {
574 		kvm_pmu_disable_counter_mask(vcpu, mask);
575 	}
576 
577 	if (val & ARMV8_PMU_PMCR_C)
578 		kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
579 
580 	if (val & ARMV8_PMU_PMCR_P) {
581 		for_each_set_bit(i, &mask, 32)
582 			kvm_pmu_set_counter_value(vcpu, i, 0);
583 	}
584 }
585 
586 static bool kvm_pmu_counter_is_enabled(struct kvm_vcpu *vcpu, u64 select_idx)
587 {
588 	return (__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) &&
589 	       (__vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(select_idx));
590 }
591 
592 /**
593  * kvm_pmu_create_perf_event - create a perf event for a counter
594  * @vcpu: The vcpu pointer
595  * @select_idx: The number of selected counter
596  */
597 static void kvm_pmu_create_perf_event(struct kvm_vcpu *vcpu, u64 select_idx)
598 {
599 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
600 	struct kvm_pmc *pmc;
601 	struct perf_event *event;
602 	struct perf_event_attr attr;
603 	u64 eventsel, counter, reg, data;
604 
605 	/*
606 	 * For chained counters the event type and filtering attributes are
607 	 * obtained from the low/even counter. We also use this counter to
608 	 * determine if the event is enabled/disabled.
609 	 */
610 	pmc = kvm_pmu_get_canonical_pmc(&pmu->pmc[select_idx]);
611 
612 	reg = (pmc->idx == ARMV8_PMU_CYCLE_IDX)
613 	      ? PMCCFILTR_EL0 : PMEVTYPER0_EL0 + pmc->idx;
614 	data = __vcpu_sys_reg(vcpu, reg);
615 
616 	kvm_pmu_stop_counter(vcpu, pmc);
617 	if (pmc->idx == ARMV8_PMU_CYCLE_IDX)
618 		eventsel = ARMV8_PMUV3_PERFCTR_CPU_CYCLES;
619 	else
620 		eventsel = data & kvm_pmu_event_mask(vcpu->kvm);
621 
622 	/* Software increment event doesn't need to be backed by a perf event */
623 	if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR)
624 		return;
625 
626 	/*
627 	 * If we have a filter in place and that the event isn't allowed, do
628 	 * not install a perf event either.
629 	 */
630 	if (vcpu->kvm->arch.pmu_filter &&
631 	    !test_bit(eventsel, vcpu->kvm->arch.pmu_filter))
632 		return;
633 
634 	memset(&attr, 0, sizeof(struct perf_event_attr));
635 	attr.type = PERF_TYPE_RAW;
636 	attr.size = sizeof(attr);
637 	attr.pinned = 1;
638 	attr.disabled = !kvm_pmu_counter_is_enabled(vcpu, pmc->idx);
639 	attr.exclude_user = data & ARMV8_PMU_EXCLUDE_EL0 ? 1 : 0;
640 	attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0;
641 	attr.exclude_hv = 1; /* Don't count EL2 events */
642 	attr.exclude_host = 1; /* Don't count host events */
643 	attr.config = eventsel;
644 
645 	counter = kvm_pmu_get_pair_counter_value(vcpu, pmc);
646 
647 	if (kvm_pmu_pmc_is_chained(pmc)) {
648 		/**
649 		 * The initial sample period (overflow count) of an event. For
650 		 * chained counters we only support overflow interrupts on the
651 		 * high counter.
652 		 */
653 		attr.sample_period = (-counter) & GENMASK(63, 0);
654 		attr.config1 |= PERF_ATTR_CFG1_KVM_PMU_CHAINED;
655 
656 		event = perf_event_create_kernel_counter(&attr, -1, current,
657 							 kvm_pmu_perf_overflow,
658 							 pmc + 1);
659 	} else {
660 		/* The initial sample period (overflow count) of an event. */
661 		if (kvm_pmu_idx_is_64bit(vcpu, pmc->idx))
662 			attr.sample_period = (-counter) & GENMASK(63, 0);
663 		else
664 			attr.sample_period = (-counter) & GENMASK(31, 0);
665 
666 		event = perf_event_create_kernel_counter(&attr, -1, current,
667 						 kvm_pmu_perf_overflow, pmc);
668 	}
669 
670 	if (IS_ERR(event)) {
671 		pr_err_once("kvm: pmu event creation failed %ld\n",
672 			    PTR_ERR(event));
673 		return;
674 	}
675 
676 	pmc->perf_event = event;
677 }
678 
679 /**
680  * kvm_pmu_update_pmc_chained - update chained bitmap
681  * @vcpu: The vcpu pointer
682  * @select_idx: The number of selected counter
683  *
684  * Update the chained bitmap based on the event type written in the
685  * typer register and the enable state of the odd register.
686  */
687 static void kvm_pmu_update_pmc_chained(struct kvm_vcpu *vcpu, u64 select_idx)
688 {
689 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
690 	struct kvm_pmc *pmc = &pmu->pmc[select_idx], *canonical_pmc;
691 	bool new_state, old_state;
692 
693 	old_state = kvm_pmu_pmc_is_chained(pmc);
694 	new_state = kvm_pmu_idx_has_chain_evtype(vcpu, pmc->idx) &&
695 		    kvm_pmu_counter_is_enabled(vcpu, pmc->idx | 0x1);
696 
697 	if (old_state == new_state)
698 		return;
699 
700 	canonical_pmc = kvm_pmu_get_canonical_pmc(pmc);
701 	kvm_pmu_stop_counter(vcpu, canonical_pmc);
702 	if (new_state) {
703 		/*
704 		 * During promotion from !chained to chained we must ensure
705 		 * the adjacent counter is stopped and its event destroyed
706 		 */
707 		kvm_pmu_stop_counter(vcpu, kvm_pmu_get_alternate_pmc(pmc));
708 		set_bit(pmc->idx >> 1, vcpu->arch.pmu.chained);
709 		return;
710 	}
711 	clear_bit(pmc->idx >> 1, vcpu->arch.pmu.chained);
712 }
713 
714 /**
715  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
716  * @vcpu: The vcpu pointer
717  * @data: The data guest writes to PMXEVTYPER_EL0
718  * @select_idx: The number of selected counter
719  *
720  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
721  * event with given hardware event number. Here we call perf_event API to
722  * emulate this action and create a kernel perf event for it.
723  */
724 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
725 				    u64 select_idx)
726 {
727 	u64 reg, mask;
728 
729 	mask  =  ARMV8_PMU_EVTYPE_MASK;
730 	mask &= ~ARMV8_PMU_EVTYPE_EVENT;
731 	mask |= kvm_pmu_event_mask(vcpu->kvm);
732 
733 	reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
734 	      ? PMCCFILTR_EL0 : PMEVTYPER0_EL0 + select_idx;
735 
736 	__vcpu_sys_reg(vcpu, reg) = data & mask;
737 
738 	kvm_pmu_update_pmc_chained(vcpu, select_idx);
739 	kvm_pmu_create_perf_event(vcpu, select_idx);
740 }
741 
742 static int kvm_pmu_probe_pmuver(void)
743 {
744 	struct perf_event_attr attr = { };
745 	struct perf_event *event;
746 	struct arm_pmu *pmu;
747 	int pmuver = 0xf;
748 
749 	/*
750 	 * Create a dummy event that only counts user cycles. As we'll never
751 	 * leave this function with the event being live, it will never
752 	 * count anything. But it allows us to probe some of the PMU
753 	 * details. Yes, this is terrible.
754 	 */
755 	attr.type = PERF_TYPE_RAW;
756 	attr.size = sizeof(attr);
757 	attr.pinned = 1;
758 	attr.disabled = 0;
759 	attr.exclude_user = 0;
760 	attr.exclude_kernel = 1;
761 	attr.exclude_hv = 1;
762 	attr.exclude_host = 1;
763 	attr.config = ARMV8_PMUV3_PERFCTR_CPU_CYCLES;
764 	attr.sample_period = GENMASK(63, 0);
765 
766 	event = perf_event_create_kernel_counter(&attr, -1, current,
767 						 kvm_pmu_perf_overflow, &attr);
768 
769 	if (IS_ERR(event)) {
770 		pr_err_once("kvm: pmu event creation failed %ld\n",
771 			    PTR_ERR(event));
772 		return 0xf;
773 	}
774 
775 	if (event->pmu) {
776 		pmu = to_arm_pmu(event->pmu);
777 		if (pmu->pmuver)
778 			pmuver = pmu->pmuver;
779 	}
780 
781 	perf_event_disable(event);
782 	perf_event_release_kernel(event);
783 
784 	return pmuver;
785 }
786 
787 u64 kvm_pmu_get_pmceid(struct kvm_vcpu *vcpu, bool pmceid1)
788 {
789 	unsigned long *bmap = vcpu->kvm->arch.pmu_filter;
790 	u64 val, mask = 0;
791 	int base, i;
792 
793 	if (!pmceid1) {
794 		val = read_sysreg(pmceid0_el0);
795 		base = 0;
796 	} else {
797 		val = read_sysreg(pmceid1_el0);
798 		base = 32;
799 	}
800 
801 	if (!bmap)
802 		return val;
803 
804 	for (i = 0; i < 32; i += 8) {
805 		u64 byte;
806 
807 		byte = bitmap_get_value8(bmap, base + i);
808 		mask |= byte << i;
809 		byte = bitmap_get_value8(bmap, 0x4000 + base + i);
810 		mask |= byte << (32 + i);
811 	}
812 
813 	return val & mask;
814 }
815 
816 bool kvm_arm_support_pmu_v3(void)
817 {
818 	/*
819 	 * Check if HW_PERF_EVENTS are supported by checking the number of
820 	 * hardware performance counters. This could ensure the presence of
821 	 * a physical PMU and CONFIG_PERF_EVENT is selected.
822 	 */
823 	return (perf_num_counters() > 0);
824 }
825 
826 int kvm_arm_pmu_v3_enable(struct kvm_vcpu *vcpu)
827 {
828 	if (!kvm_vcpu_has_pmu(vcpu))
829 		return 0;
830 
831 	if (!vcpu->arch.pmu.created)
832 		return -EINVAL;
833 
834 	/*
835 	 * A valid interrupt configuration for the PMU is either to have a
836 	 * properly configured interrupt number and using an in-kernel
837 	 * irqchip, or to not have an in-kernel GIC and not set an IRQ.
838 	 */
839 	if (irqchip_in_kernel(vcpu->kvm)) {
840 		int irq = vcpu->arch.pmu.irq_num;
841 		/*
842 		 * If we are using an in-kernel vgic, at this point we know
843 		 * the vgic will be initialized, so we can check the PMU irq
844 		 * number against the dimensions of the vgic and make sure
845 		 * it's valid.
846 		 */
847 		if (!irq_is_ppi(irq) && !vgic_valid_spi(vcpu->kvm, irq))
848 			return -EINVAL;
849 	} else if (kvm_arm_pmu_irq_initialized(vcpu)) {
850 		   return -EINVAL;
851 	}
852 
853 	kvm_pmu_vcpu_reset(vcpu);
854 
855 	return 0;
856 }
857 
858 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
859 {
860 	if (irqchip_in_kernel(vcpu->kvm)) {
861 		int ret;
862 
863 		/*
864 		 * If using the PMU with an in-kernel virtual GIC
865 		 * implementation, we require the GIC to be already
866 		 * initialized when initializing the PMU.
867 		 */
868 		if (!vgic_initialized(vcpu->kvm))
869 			return -ENODEV;
870 
871 		if (!kvm_arm_pmu_irq_initialized(vcpu))
872 			return -ENXIO;
873 
874 		ret = kvm_vgic_set_owner(vcpu, vcpu->arch.pmu.irq_num,
875 					 &vcpu->arch.pmu);
876 		if (ret)
877 			return ret;
878 	}
879 
880 	init_irq_work(&vcpu->arch.pmu.overflow_work,
881 		      kvm_pmu_perf_overflow_notify_vcpu);
882 
883 	vcpu->arch.pmu.created = true;
884 	return 0;
885 }
886 
887 /*
888  * For one VM the interrupt type must be same for each vcpu.
889  * As a PPI, the interrupt number is the same for all vcpus,
890  * while as an SPI it must be a separate number per vcpu.
891  */
892 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
893 {
894 	int i;
895 	struct kvm_vcpu *vcpu;
896 
897 	kvm_for_each_vcpu(i, vcpu, kvm) {
898 		if (!kvm_arm_pmu_irq_initialized(vcpu))
899 			continue;
900 
901 		if (irq_is_ppi(irq)) {
902 			if (vcpu->arch.pmu.irq_num != irq)
903 				return false;
904 		} else {
905 			if (vcpu->arch.pmu.irq_num == irq)
906 				return false;
907 		}
908 	}
909 
910 	return true;
911 }
912 
913 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
914 {
915 	if (!kvm_vcpu_has_pmu(vcpu))
916 		return -ENODEV;
917 
918 	if (vcpu->arch.pmu.created)
919 		return -EBUSY;
920 
921 	if (!vcpu->kvm->arch.pmuver)
922 		vcpu->kvm->arch.pmuver = kvm_pmu_probe_pmuver();
923 
924 	if (vcpu->kvm->arch.pmuver == 0xf)
925 		return -ENODEV;
926 
927 	switch (attr->attr) {
928 	case KVM_ARM_VCPU_PMU_V3_IRQ: {
929 		int __user *uaddr = (int __user *)(long)attr->addr;
930 		int irq;
931 
932 		if (!irqchip_in_kernel(vcpu->kvm))
933 			return -EINVAL;
934 
935 		if (get_user(irq, uaddr))
936 			return -EFAULT;
937 
938 		/* The PMU overflow interrupt can be a PPI or a valid SPI. */
939 		if (!(irq_is_ppi(irq) || irq_is_spi(irq)))
940 			return -EINVAL;
941 
942 		if (!pmu_irq_is_valid(vcpu->kvm, irq))
943 			return -EINVAL;
944 
945 		if (kvm_arm_pmu_irq_initialized(vcpu))
946 			return -EBUSY;
947 
948 		kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
949 		vcpu->arch.pmu.irq_num = irq;
950 		return 0;
951 	}
952 	case KVM_ARM_VCPU_PMU_V3_FILTER: {
953 		struct kvm_pmu_event_filter __user *uaddr;
954 		struct kvm_pmu_event_filter filter;
955 		int nr_events;
956 
957 		nr_events = kvm_pmu_event_mask(vcpu->kvm) + 1;
958 
959 		uaddr = (struct kvm_pmu_event_filter __user *)(long)attr->addr;
960 
961 		if (copy_from_user(&filter, uaddr, sizeof(filter)))
962 			return -EFAULT;
963 
964 		if (((u32)filter.base_event + filter.nevents) > nr_events ||
965 		    (filter.action != KVM_PMU_EVENT_ALLOW &&
966 		     filter.action != KVM_PMU_EVENT_DENY))
967 			return -EINVAL;
968 
969 		mutex_lock(&vcpu->kvm->lock);
970 
971 		if (!vcpu->kvm->arch.pmu_filter) {
972 			vcpu->kvm->arch.pmu_filter = bitmap_alloc(nr_events, GFP_KERNEL);
973 			if (!vcpu->kvm->arch.pmu_filter) {
974 				mutex_unlock(&vcpu->kvm->lock);
975 				return -ENOMEM;
976 			}
977 
978 			/*
979 			 * The default depends on the first applied filter.
980 			 * If it allows events, the default is to deny.
981 			 * Conversely, if the first filter denies a set of
982 			 * events, the default is to allow.
983 			 */
984 			if (filter.action == KVM_PMU_EVENT_ALLOW)
985 				bitmap_zero(vcpu->kvm->arch.pmu_filter, nr_events);
986 			else
987 				bitmap_fill(vcpu->kvm->arch.pmu_filter, nr_events);
988 		}
989 
990 		if (filter.action == KVM_PMU_EVENT_ALLOW)
991 			bitmap_set(vcpu->kvm->arch.pmu_filter, filter.base_event, filter.nevents);
992 		else
993 			bitmap_clear(vcpu->kvm->arch.pmu_filter, filter.base_event, filter.nevents);
994 
995 		mutex_unlock(&vcpu->kvm->lock);
996 
997 		return 0;
998 	}
999 	case KVM_ARM_VCPU_PMU_V3_INIT:
1000 		return kvm_arm_pmu_v3_init(vcpu);
1001 	}
1002 
1003 	return -ENXIO;
1004 }
1005 
1006 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1007 {
1008 	switch (attr->attr) {
1009 	case KVM_ARM_VCPU_PMU_V3_IRQ: {
1010 		int __user *uaddr = (int __user *)(long)attr->addr;
1011 		int irq;
1012 
1013 		if (!irqchip_in_kernel(vcpu->kvm))
1014 			return -EINVAL;
1015 
1016 		if (!kvm_vcpu_has_pmu(vcpu))
1017 			return -ENODEV;
1018 
1019 		if (!kvm_arm_pmu_irq_initialized(vcpu))
1020 			return -ENXIO;
1021 
1022 		irq = vcpu->arch.pmu.irq_num;
1023 		return put_user(irq, uaddr);
1024 	}
1025 	}
1026 
1027 	return -ENXIO;
1028 }
1029 
1030 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1031 {
1032 	switch (attr->attr) {
1033 	case KVM_ARM_VCPU_PMU_V3_IRQ:
1034 	case KVM_ARM_VCPU_PMU_V3_INIT:
1035 	case KVM_ARM_VCPU_PMU_V3_FILTER:
1036 		if (kvm_vcpu_has_pmu(vcpu))
1037 			return 0;
1038 	}
1039 
1040 	return -ENXIO;
1041 }
1042