]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - virt/kvm/arm/pmu.c
Merge branches 'intel_pstate' and 'pm-domains'
[karo-tx-linux.git] / virt / kvm / arm / pmu.c
1 /*
2  * Copyright (C) 2015 Linaro Ltd.
3  * Author: Shannon Zhao <shannon.zhao@linaro.org>
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17
18 #include <linux/cpu.h>
19 #include <linux/kvm.h>
20 #include <linux/kvm_host.h>
21 #include <linux/perf_event.h>
22 #include <linux/uaccess.h>
23 #include <asm/kvm_emulate.h>
24 #include <kvm/arm_pmu.h>
25 #include <kvm/arm_vgic.h>
26
27 /**
28  * kvm_pmu_get_counter_value - get PMU counter value
29  * @vcpu: The vcpu pointer
30  * @select_idx: The counter index
31  */
32 u64 kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u64 select_idx)
33 {
34         u64 counter, reg, enabled, running;
35         struct kvm_pmu *pmu = &vcpu->arch.pmu;
36         struct kvm_pmc *pmc = &pmu->pmc[select_idx];
37
38         reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
39               ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + select_idx;
40         counter = vcpu_sys_reg(vcpu, reg);
41
42         /* The real counter value is equal to the value of counter register plus
43          * the value perf event counts.
44          */
45         if (pmc->perf_event)
46                 counter += perf_event_read_value(pmc->perf_event, &enabled,
47                                                  &running);
48
49         return counter & pmc->bitmask;
50 }
51
52 /**
53  * kvm_pmu_set_counter_value - set PMU counter value
54  * @vcpu: The vcpu pointer
55  * @select_idx: The counter index
56  * @val: The counter value
57  */
58 void kvm_pmu_set_counter_value(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
59 {
60         u64 reg;
61
62         reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
63               ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + select_idx;
64         vcpu_sys_reg(vcpu, reg) += (s64)val - kvm_pmu_get_counter_value(vcpu, select_idx);
65 }
66
67 /**
68  * kvm_pmu_stop_counter - stop PMU counter
69  * @pmc: The PMU counter pointer
70  *
71  * If this counter has been configured to monitor some event, release it here.
72  */
73 static void kvm_pmu_stop_counter(struct kvm_vcpu *vcpu, struct kvm_pmc *pmc)
74 {
75         u64 counter, reg;
76
77         if (pmc->perf_event) {
78                 counter = kvm_pmu_get_counter_value(vcpu, pmc->idx);
79                 reg = (pmc->idx == ARMV8_PMU_CYCLE_IDX)
80                        ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + pmc->idx;
81                 vcpu_sys_reg(vcpu, reg) = counter;
82                 perf_event_disable(pmc->perf_event);
83                 perf_event_release_kernel(pmc->perf_event);
84                 pmc->perf_event = NULL;
85         }
86 }
87
88 /**
89  * kvm_pmu_vcpu_reset - reset pmu state for cpu
90  * @vcpu: The vcpu pointer
91  *
92  */
93 void kvm_pmu_vcpu_reset(struct kvm_vcpu *vcpu)
94 {
95         int i;
96         struct kvm_pmu *pmu = &vcpu->arch.pmu;
97
98         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
99                 kvm_pmu_stop_counter(vcpu, &pmu->pmc[i]);
100                 pmu->pmc[i].idx = i;
101                 pmu->pmc[i].bitmask = 0xffffffffUL;
102         }
103 }
104
105 /**
106  * kvm_pmu_vcpu_destroy - free perf event of PMU for cpu
107  * @vcpu: The vcpu pointer
108  *
109  */
110 void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
111 {
112         int i;
113         struct kvm_pmu *pmu = &vcpu->arch.pmu;
114
115         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
116                 struct kvm_pmc *pmc = &pmu->pmc[i];
117
118                 if (pmc->perf_event) {
119                         perf_event_disable(pmc->perf_event);
120                         perf_event_release_kernel(pmc->perf_event);
121                         pmc->perf_event = NULL;
122                 }
123         }
124 }
125
126 u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu)
127 {
128         u64 val = vcpu_sys_reg(vcpu, PMCR_EL0) >> ARMV8_PMU_PMCR_N_SHIFT;
129
130         val &= ARMV8_PMU_PMCR_N_MASK;
131         if (val == 0)
132                 return BIT(ARMV8_PMU_CYCLE_IDX);
133         else
134                 return GENMASK(val - 1, 0) | BIT(ARMV8_PMU_CYCLE_IDX);
135 }
136
137 /**
138  * kvm_pmu_enable_counter - enable selected PMU counter
139  * @vcpu: The vcpu pointer
140  * @val: the value guest writes to PMCNTENSET register
141  *
142  * Call perf_event_enable to start counting the perf event
143  */
144 void kvm_pmu_enable_counter(struct kvm_vcpu *vcpu, u64 val)
145 {
146         int i;
147         struct kvm_pmu *pmu = &vcpu->arch.pmu;
148         struct kvm_pmc *pmc;
149
150         if (!(vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) || !val)
151                 return;
152
153         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
154                 if (!(val & BIT(i)))
155                         continue;
156
157                 pmc = &pmu->pmc[i];
158                 if (pmc->perf_event) {
159                         perf_event_enable(pmc->perf_event);
160                         if (pmc->perf_event->state != PERF_EVENT_STATE_ACTIVE)
161                                 kvm_debug("fail to enable perf event\n");
162                 }
163         }
164 }
165
166 /**
167  * kvm_pmu_disable_counter - disable selected PMU counter
168  * @vcpu: The vcpu pointer
169  * @val: the value guest writes to PMCNTENCLR register
170  *
171  * Call perf_event_disable to stop counting the perf event
172  */
173 void kvm_pmu_disable_counter(struct kvm_vcpu *vcpu, u64 val)
174 {
175         int i;
176         struct kvm_pmu *pmu = &vcpu->arch.pmu;
177         struct kvm_pmc *pmc;
178
179         if (!val)
180                 return;
181
182         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
183                 if (!(val & BIT(i)))
184                         continue;
185
186                 pmc = &pmu->pmc[i];
187                 if (pmc->perf_event)
188                         perf_event_disable(pmc->perf_event);
189         }
190 }
191
192 static u64 kvm_pmu_overflow_status(struct kvm_vcpu *vcpu)
193 {
194         u64 reg = 0;
195
196         if ((vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E)) {
197                 reg = vcpu_sys_reg(vcpu, PMOVSSET_EL0);
198                 reg &= vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
199                 reg &= vcpu_sys_reg(vcpu, PMINTENSET_EL1);
200                 reg &= kvm_pmu_valid_counter_mask(vcpu);
201         }
202
203         return reg;
204 }
205
206 static void kvm_pmu_check_overflow(struct kvm_vcpu *vcpu)
207 {
208         struct kvm_pmu *pmu = &vcpu->arch.pmu;
209         bool overflow = !!kvm_pmu_overflow_status(vcpu);
210
211         if (pmu->irq_level == overflow)
212                 return;
213
214         pmu->irq_level = overflow;
215
216         if (likely(irqchip_in_kernel(vcpu->kvm))) {
217                 int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
218                                               pmu->irq_num, overflow,
219                                               &vcpu->arch.pmu);
220                 WARN_ON(ret);
221         }
222 }
223
224 /**
225  * kvm_pmu_overflow_set - set PMU overflow interrupt
226  * @vcpu: The vcpu pointer
227  * @val: the value guest writes to PMOVSSET register
228  */
229 void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val)
230 {
231         if (val == 0)
232                 return;
233
234         vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= val;
235         kvm_pmu_check_overflow(vcpu);
236 }
237
238 static void kvm_pmu_update_state(struct kvm_vcpu *vcpu)
239 {
240         if (!kvm_arm_pmu_v3_ready(vcpu))
241                 return;
242         kvm_pmu_check_overflow(vcpu);
243 }
244
245 bool kvm_pmu_should_notify_user(struct kvm_vcpu *vcpu)
246 {
247         struct kvm_pmu *pmu = &vcpu->arch.pmu;
248         struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
249         bool run_level = sregs->device_irq_level & KVM_ARM_DEV_PMU;
250
251         if (likely(irqchip_in_kernel(vcpu->kvm)))
252                 return false;
253
254         return pmu->irq_level != run_level;
255 }
256
257 /*
258  * Reflect the PMU overflow interrupt output level into the kvm_run structure
259  */
260 void kvm_pmu_update_run(struct kvm_vcpu *vcpu)
261 {
262         struct kvm_sync_regs *regs = &vcpu->run->s.regs;
263
264         /* Populate the timer bitmap for user space */
265         regs->device_irq_level &= ~KVM_ARM_DEV_PMU;
266         if (vcpu->arch.pmu.irq_level)
267                 regs->device_irq_level |= KVM_ARM_DEV_PMU;
268 }
269
270 /**
271  * kvm_pmu_flush_hwstate - flush pmu state to cpu
272  * @vcpu: The vcpu pointer
273  *
274  * Check if the PMU has overflowed while we were running in the host, and inject
275  * an interrupt if that was the case.
276  */
277 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
278 {
279         kvm_pmu_update_state(vcpu);
280 }
281
282 /**
283  * kvm_pmu_sync_hwstate - sync pmu state from cpu
284  * @vcpu: The vcpu pointer
285  *
286  * Check if the PMU has overflowed while we were running in the guest, and
287  * inject an interrupt if that was the case.
288  */
289 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
290 {
291         kvm_pmu_update_state(vcpu);
292 }
293
294 static inline struct kvm_vcpu *kvm_pmc_to_vcpu(struct kvm_pmc *pmc)
295 {
296         struct kvm_pmu *pmu;
297         struct kvm_vcpu_arch *vcpu_arch;
298
299         pmc -= pmc->idx;
300         pmu = container_of(pmc, struct kvm_pmu, pmc[0]);
301         vcpu_arch = container_of(pmu, struct kvm_vcpu_arch, pmu);
302         return container_of(vcpu_arch, struct kvm_vcpu, arch);
303 }
304
305 /**
306  * When perf event overflows, call kvm_pmu_overflow_set to set overflow status.
307  */
308 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
309                                   struct perf_sample_data *data,
310                                   struct pt_regs *regs)
311 {
312         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
313         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
314         int idx = pmc->idx;
315
316         kvm_pmu_overflow_set(vcpu, BIT(idx));
317 }
318
319 /**
320  * kvm_pmu_software_increment - do software increment
321  * @vcpu: The vcpu pointer
322  * @val: the value guest writes to PMSWINC register
323  */
324 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
325 {
326         int i;
327         u64 type, enable, reg;
328
329         if (val == 0)
330                 return;
331
332         enable = vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
333         for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++) {
334                 if (!(val & BIT(i)))
335                         continue;
336                 type = vcpu_sys_reg(vcpu, PMEVTYPER0_EL0 + i)
337                        & ARMV8_PMU_EVTYPE_EVENT;
338                 if ((type == ARMV8_PMUV3_PERFCTR_SW_INCR)
339                     && (enable & BIT(i))) {
340                         reg = vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) + 1;
341                         reg = lower_32_bits(reg);
342                         vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) = reg;
343                         if (!reg)
344                                 kvm_pmu_overflow_set(vcpu, BIT(i));
345                 }
346         }
347 }
348
349 /**
350  * kvm_pmu_handle_pmcr - handle PMCR register
351  * @vcpu: The vcpu pointer
352  * @val: the value guest writes to PMCR register
353  */
354 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
355 {
356         struct kvm_pmu *pmu = &vcpu->arch.pmu;
357         struct kvm_pmc *pmc;
358         u64 mask;
359         int i;
360
361         mask = kvm_pmu_valid_counter_mask(vcpu);
362         if (val & ARMV8_PMU_PMCR_E) {
363                 kvm_pmu_enable_counter(vcpu,
364                                 vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & mask);
365         } else {
366                 kvm_pmu_disable_counter(vcpu, mask);
367         }
368
369         if (val & ARMV8_PMU_PMCR_C)
370                 kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
371
372         if (val & ARMV8_PMU_PMCR_P) {
373                 for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++)
374                         kvm_pmu_set_counter_value(vcpu, i, 0);
375         }
376
377         if (val & ARMV8_PMU_PMCR_LC) {
378                 pmc = &pmu->pmc[ARMV8_PMU_CYCLE_IDX];
379                 pmc->bitmask = 0xffffffffffffffffUL;
380         }
381 }
382
383 static bool kvm_pmu_counter_is_enabled(struct kvm_vcpu *vcpu, u64 select_idx)
384 {
385         return (vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) &&
386                (vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(select_idx));
387 }
388
389 /**
390  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
391  * @vcpu: The vcpu pointer
392  * @data: The data guest writes to PMXEVTYPER_EL0
393  * @select_idx: The number of selected counter
394  *
395  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
396  * event with given hardware event number. Here we call perf_event API to
397  * emulate this action and create a kernel perf event for it.
398  */
399 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
400                                     u64 select_idx)
401 {
402         struct kvm_pmu *pmu = &vcpu->arch.pmu;
403         struct kvm_pmc *pmc = &pmu->pmc[select_idx];
404         struct perf_event *event;
405         struct perf_event_attr attr;
406         u64 eventsel, counter;
407
408         kvm_pmu_stop_counter(vcpu, pmc);
409         eventsel = data & ARMV8_PMU_EVTYPE_EVENT;
410
411         /* Software increment event does't need to be backed by a perf event */
412         if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR &&
413             select_idx != ARMV8_PMU_CYCLE_IDX)
414                 return;
415
416         memset(&attr, 0, sizeof(struct perf_event_attr));
417         attr.type = PERF_TYPE_RAW;
418         attr.size = sizeof(attr);
419         attr.pinned = 1;
420         attr.disabled = !kvm_pmu_counter_is_enabled(vcpu, select_idx);
421         attr.exclude_user = data & ARMV8_PMU_EXCLUDE_EL0 ? 1 : 0;
422         attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0;
423         attr.exclude_hv = 1; /* Don't count EL2 events */
424         attr.exclude_host = 1; /* Don't count host events */
425         attr.config = (select_idx == ARMV8_PMU_CYCLE_IDX) ?
426                 ARMV8_PMUV3_PERFCTR_CPU_CYCLES : eventsel;
427
428         counter = kvm_pmu_get_counter_value(vcpu, select_idx);
429         /* The initial sample period (overflow count) of an event. */
430         attr.sample_period = (-counter) & pmc->bitmask;
431
432         event = perf_event_create_kernel_counter(&attr, -1, current,
433                                                  kvm_pmu_perf_overflow, pmc);
434         if (IS_ERR(event)) {
435                 pr_err_once("kvm: pmu event creation failed %ld\n",
436                             PTR_ERR(event));
437                 return;
438         }
439
440         pmc->perf_event = event;
441 }
442
443 bool kvm_arm_support_pmu_v3(void)
444 {
445         /*
446          * Check if HW_PERF_EVENTS are supported by checking the number of
447          * hardware performance counters. This could ensure the presence of
448          * a physical PMU and CONFIG_PERF_EVENT is selected.
449          */
450         return (perf_num_counters() > 0);
451 }
452
453 int kvm_arm_pmu_v3_enable(struct kvm_vcpu *vcpu)
454 {
455         if (!vcpu->arch.pmu.created)
456                 return 0;
457
458         /*
459          * A valid interrupt configuration for the PMU is either to have a
460          * properly configured interrupt number and using an in-kernel
461          * irqchip, or to not have an in-kernel GIC and not set an IRQ.
462          */
463         if (irqchip_in_kernel(vcpu->kvm)) {
464                 int irq = vcpu->arch.pmu.irq_num;
465                 if (!kvm_arm_pmu_irq_initialized(vcpu))
466                         return -EINVAL;
467
468                 /*
469                  * If we are using an in-kernel vgic, at this point we know
470                  * the vgic will be initialized, so we can check the PMU irq
471                  * number against the dimensions of the vgic and make sure
472                  * it's valid.
473                  */
474                 if (!irq_is_ppi(irq) && !vgic_valid_spi(vcpu->kvm, irq))
475                         return -EINVAL;
476         } else if (kvm_arm_pmu_irq_initialized(vcpu)) {
477                    return -EINVAL;
478         }
479
480         kvm_pmu_vcpu_reset(vcpu);
481         vcpu->arch.pmu.ready = true;
482
483         return 0;
484 }
485
486 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
487 {
488         if (!kvm_arm_support_pmu_v3())
489                 return -ENODEV;
490
491         if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
492                 return -ENXIO;
493
494         if (vcpu->arch.pmu.created)
495                 return -EBUSY;
496
497         if (irqchip_in_kernel(vcpu->kvm)) {
498                 int ret;
499
500                 /*
501                  * If using the PMU with an in-kernel virtual GIC
502                  * implementation, we require the GIC to be already
503                  * initialized when initializing the PMU.
504                  */
505                 if (!vgic_initialized(vcpu->kvm))
506                         return -ENODEV;
507
508                 if (!kvm_arm_pmu_irq_initialized(vcpu))
509                         return -ENXIO;
510
511                 ret = kvm_vgic_set_owner(vcpu, vcpu->arch.pmu.irq_num,
512                                          &vcpu->arch.pmu);
513                 if (ret)
514                         return ret;
515         }
516
517         vcpu->arch.pmu.created = true;
518         return 0;
519 }
520
521 /*
522  * For one VM the interrupt type must be same for each vcpu.
523  * As a PPI, the interrupt number is the same for all vcpus,
524  * while as an SPI it must be a separate number per vcpu.
525  */
526 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
527 {
528         int i;
529         struct kvm_vcpu *vcpu;
530
531         kvm_for_each_vcpu(i, vcpu, kvm) {
532                 if (!kvm_arm_pmu_irq_initialized(vcpu))
533                         continue;
534
535                 if (irq_is_ppi(irq)) {
536                         if (vcpu->arch.pmu.irq_num != irq)
537                                 return false;
538                 } else {
539                         if (vcpu->arch.pmu.irq_num == irq)
540                                 return false;
541                 }
542         }
543
544         return true;
545 }
546
547 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
548 {
549         switch (attr->attr) {
550         case KVM_ARM_VCPU_PMU_V3_IRQ: {
551                 int __user *uaddr = (int __user *)(long)attr->addr;
552                 int irq;
553
554                 if (!irqchip_in_kernel(vcpu->kvm))
555                         return -EINVAL;
556
557                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
558                         return -ENODEV;
559
560                 if (get_user(irq, uaddr))
561                         return -EFAULT;
562
563                 /* The PMU overflow interrupt can be a PPI or a valid SPI. */
564                 if (!(irq_is_ppi(irq) || irq_is_spi(irq)))
565                         return -EINVAL;
566
567                 if (!pmu_irq_is_valid(vcpu->kvm, irq))
568                         return -EINVAL;
569
570                 if (kvm_arm_pmu_irq_initialized(vcpu))
571                         return -EBUSY;
572
573                 kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
574                 vcpu->arch.pmu.irq_num = irq;
575                 return 0;
576         }
577         case KVM_ARM_VCPU_PMU_V3_INIT:
578                 return kvm_arm_pmu_v3_init(vcpu);
579         }
580
581         return -ENXIO;
582 }
583
584 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
585 {
586         switch (attr->attr) {
587         case KVM_ARM_VCPU_PMU_V3_IRQ: {
588                 int __user *uaddr = (int __user *)(long)attr->addr;
589                 int irq;
590
591                 if (!irqchip_in_kernel(vcpu->kvm))
592                         return -EINVAL;
593
594                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
595                         return -ENODEV;
596
597                 if (!kvm_arm_pmu_irq_initialized(vcpu))
598                         return -ENXIO;
599
600                 irq = vcpu->arch.pmu.irq_num;
601                 return put_user(irq, uaddr);
602         }
603         }
604
605         return -ENXIO;
606 }
607
608 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
609 {
610         switch (attr->attr) {
611         case KVM_ARM_VCPU_PMU_V3_IRQ:
612         case KVM_ARM_VCPU_PMU_V3_INIT:
613                 if (kvm_arm_support_pmu_v3() &&
614                     test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
615                         return 0;
616         }
617
618         return -ENXIO;
619 }