]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - virt/kvm/arm/pmu.c
Merge branch 'i2c/for-current' of git://git.kernel.org/pub/scm/linux/kernel/git/wsa...
[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_update_state(struct kvm_vcpu *vcpu)
207 {
208         struct kvm_pmu *pmu = &vcpu->arch.pmu;
209         bool overflow;
210
211         if (!kvm_arm_pmu_v3_ready(vcpu))
212                 return;
213
214         overflow = !!kvm_pmu_overflow_status(vcpu);
215         if (pmu->irq_level == overflow)
216                 return;
217
218         pmu->irq_level = overflow;
219
220         if (likely(irqchip_in_kernel(vcpu->kvm))) {
221                 int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
222                                               pmu->irq_num, overflow, pmu);
223                 WARN_ON(ret);
224         }
225 }
226
227 bool kvm_pmu_should_notify_user(struct kvm_vcpu *vcpu)
228 {
229         struct kvm_pmu *pmu = &vcpu->arch.pmu;
230         struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
231         bool run_level = sregs->device_irq_level & KVM_ARM_DEV_PMU;
232
233         if (likely(irqchip_in_kernel(vcpu->kvm)))
234                 return false;
235
236         return pmu->irq_level != run_level;
237 }
238
239 /*
240  * Reflect the PMU overflow interrupt output level into the kvm_run structure
241  */
242 void kvm_pmu_update_run(struct kvm_vcpu *vcpu)
243 {
244         struct kvm_sync_regs *regs = &vcpu->run->s.regs;
245
246         /* Populate the timer bitmap for user space */
247         regs->device_irq_level &= ~KVM_ARM_DEV_PMU;
248         if (vcpu->arch.pmu.irq_level)
249                 regs->device_irq_level |= KVM_ARM_DEV_PMU;
250 }
251
252 /**
253  * kvm_pmu_flush_hwstate - flush pmu state to cpu
254  * @vcpu: The vcpu pointer
255  *
256  * Check if the PMU has overflowed while we were running in the host, and inject
257  * an interrupt if that was the case.
258  */
259 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
260 {
261         kvm_pmu_update_state(vcpu);
262 }
263
264 /**
265  * kvm_pmu_sync_hwstate - sync pmu state from cpu
266  * @vcpu: The vcpu pointer
267  *
268  * Check if the PMU has overflowed while we were running in the guest, and
269  * inject an interrupt if that was the case.
270  */
271 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
272 {
273         kvm_pmu_update_state(vcpu);
274 }
275
276 static inline struct kvm_vcpu *kvm_pmc_to_vcpu(struct kvm_pmc *pmc)
277 {
278         struct kvm_pmu *pmu;
279         struct kvm_vcpu_arch *vcpu_arch;
280
281         pmc -= pmc->idx;
282         pmu = container_of(pmc, struct kvm_pmu, pmc[0]);
283         vcpu_arch = container_of(pmu, struct kvm_vcpu_arch, pmu);
284         return container_of(vcpu_arch, struct kvm_vcpu, arch);
285 }
286
287 /**
288  * When the perf event overflows, set the overflow status and inform the vcpu.
289  */
290 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
291                                   struct perf_sample_data *data,
292                                   struct pt_regs *regs)
293 {
294         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
295         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
296         int idx = pmc->idx;
297
298         vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(idx);
299
300         if (kvm_pmu_overflow_status(vcpu)) {
301                 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
302                 kvm_vcpu_kick(vcpu);
303         }
304 }
305
306 /**
307  * kvm_pmu_software_increment - do software increment
308  * @vcpu: The vcpu pointer
309  * @val: the value guest writes to PMSWINC register
310  */
311 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
312 {
313         int i;
314         u64 type, enable, reg;
315
316         if (val == 0)
317                 return;
318
319         enable = vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
320         for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++) {
321                 if (!(val & BIT(i)))
322                         continue;
323                 type = vcpu_sys_reg(vcpu, PMEVTYPER0_EL0 + i)
324                        & ARMV8_PMU_EVTYPE_EVENT;
325                 if ((type == ARMV8_PMUV3_PERFCTR_SW_INCR)
326                     && (enable & BIT(i))) {
327                         reg = vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) + 1;
328                         reg = lower_32_bits(reg);
329                         vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) = reg;
330                         if (!reg)
331                                 vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(i);
332                 }
333         }
334 }
335
336 /**
337  * kvm_pmu_handle_pmcr - handle PMCR register
338  * @vcpu: The vcpu pointer
339  * @val: the value guest writes to PMCR register
340  */
341 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
342 {
343         struct kvm_pmu *pmu = &vcpu->arch.pmu;
344         struct kvm_pmc *pmc;
345         u64 mask;
346         int i;
347
348         mask = kvm_pmu_valid_counter_mask(vcpu);
349         if (val & ARMV8_PMU_PMCR_E) {
350                 kvm_pmu_enable_counter(vcpu,
351                                 vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & mask);
352         } else {
353                 kvm_pmu_disable_counter(vcpu, mask);
354         }
355
356         if (val & ARMV8_PMU_PMCR_C)
357                 kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
358
359         if (val & ARMV8_PMU_PMCR_P) {
360                 for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++)
361                         kvm_pmu_set_counter_value(vcpu, i, 0);
362         }
363
364         if (val & ARMV8_PMU_PMCR_LC) {
365                 pmc = &pmu->pmc[ARMV8_PMU_CYCLE_IDX];
366                 pmc->bitmask = 0xffffffffffffffffUL;
367         }
368 }
369
370 static bool kvm_pmu_counter_is_enabled(struct kvm_vcpu *vcpu, u64 select_idx)
371 {
372         return (vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) &&
373                (vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(select_idx));
374 }
375
376 /**
377  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
378  * @vcpu: The vcpu pointer
379  * @data: The data guest writes to PMXEVTYPER_EL0
380  * @select_idx: The number of selected counter
381  *
382  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
383  * event with given hardware event number. Here we call perf_event API to
384  * emulate this action and create a kernel perf event for it.
385  */
386 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
387                                     u64 select_idx)
388 {
389         struct kvm_pmu *pmu = &vcpu->arch.pmu;
390         struct kvm_pmc *pmc = &pmu->pmc[select_idx];
391         struct perf_event *event;
392         struct perf_event_attr attr;
393         u64 eventsel, counter;
394
395         kvm_pmu_stop_counter(vcpu, pmc);
396         eventsel = data & ARMV8_PMU_EVTYPE_EVENT;
397
398         /* Software increment event does't need to be backed by a perf event */
399         if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR &&
400             select_idx != ARMV8_PMU_CYCLE_IDX)
401                 return;
402
403         memset(&attr, 0, sizeof(struct perf_event_attr));
404         attr.type = PERF_TYPE_RAW;
405         attr.size = sizeof(attr);
406         attr.pinned = 1;
407         attr.disabled = !kvm_pmu_counter_is_enabled(vcpu, select_idx);
408         attr.exclude_user = data & ARMV8_PMU_EXCLUDE_EL0 ? 1 : 0;
409         attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0;
410         attr.exclude_hv = 1; /* Don't count EL2 events */
411         attr.exclude_host = 1; /* Don't count host events */
412         attr.config = (select_idx == ARMV8_PMU_CYCLE_IDX) ?
413                 ARMV8_PMUV3_PERFCTR_CPU_CYCLES : eventsel;
414
415         counter = kvm_pmu_get_counter_value(vcpu, select_idx);
416         /* The initial sample period (overflow count) of an event. */
417         attr.sample_period = (-counter) & pmc->bitmask;
418
419         event = perf_event_create_kernel_counter(&attr, -1, current,
420                                                  kvm_pmu_perf_overflow, pmc);
421         if (IS_ERR(event)) {
422                 pr_err_once("kvm: pmu event creation failed %ld\n",
423                             PTR_ERR(event));
424                 return;
425         }
426
427         pmc->perf_event = event;
428 }
429
430 bool kvm_arm_support_pmu_v3(void)
431 {
432         /*
433          * Check if HW_PERF_EVENTS are supported by checking the number of
434          * hardware performance counters. This could ensure the presence of
435          * a physical PMU and CONFIG_PERF_EVENT is selected.
436          */
437         return (perf_num_counters() > 0);
438 }
439
440 int kvm_arm_pmu_v3_enable(struct kvm_vcpu *vcpu)
441 {
442         if (!vcpu->arch.pmu.created)
443                 return 0;
444
445         /*
446          * A valid interrupt configuration for the PMU is either to have a
447          * properly configured interrupt number and using an in-kernel
448          * irqchip, or to not have an in-kernel GIC and not set an IRQ.
449          */
450         if (irqchip_in_kernel(vcpu->kvm)) {
451                 int irq = vcpu->arch.pmu.irq_num;
452                 if (!kvm_arm_pmu_irq_initialized(vcpu))
453                         return -EINVAL;
454
455                 /*
456                  * If we are using an in-kernel vgic, at this point we know
457                  * the vgic will be initialized, so we can check the PMU irq
458                  * number against the dimensions of the vgic and make sure
459                  * it's valid.
460                  */
461                 if (!irq_is_ppi(irq) && !vgic_valid_spi(vcpu->kvm, irq))
462                         return -EINVAL;
463         } else if (kvm_arm_pmu_irq_initialized(vcpu)) {
464                    return -EINVAL;
465         }
466
467         kvm_pmu_vcpu_reset(vcpu);
468         vcpu->arch.pmu.ready = true;
469
470         return 0;
471 }
472
473 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
474 {
475         if (!kvm_arm_support_pmu_v3())
476                 return -ENODEV;
477
478         if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
479                 return -ENXIO;
480
481         if (vcpu->arch.pmu.created)
482                 return -EBUSY;
483
484         if (irqchip_in_kernel(vcpu->kvm)) {
485                 int ret;
486
487                 /*
488                  * If using the PMU with an in-kernel virtual GIC
489                  * implementation, we require the GIC to be already
490                  * initialized when initializing the PMU.
491                  */
492                 if (!vgic_initialized(vcpu->kvm))
493                         return -ENODEV;
494
495                 if (!kvm_arm_pmu_irq_initialized(vcpu))
496                         return -ENXIO;
497
498                 ret = kvm_vgic_set_owner(vcpu, vcpu->arch.pmu.irq_num,
499                                          &vcpu->arch.pmu);
500                 if (ret)
501                         return ret;
502         }
503
504         vcpu->arch.pmu.created = true;
505         return 0;
506 }
507
508 /*
509  * For one VM the interrupt type must be same for each vcpu.
510  * As a PPI, the interrupt number is the same for all vcpus,
511  * while as an SPI it must be a separate number per vcpu.
512  */
513 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
514 {
515         int i;
516         struct kvm_vcpu *vcpu;
517
518         kvm_for_each_vcpu(i, vcpu, kvm) {
519                 if (!kvm_arm_pmu_irq_initialized(vcpu))
520                         continue;
521
522                 if (irq_is_ppi(irq)) {
523                         if (vcpu->arch.pmu.irq_num != irq)
524                                 return false;
525                 } else {
526                         if (vcpu->arch.pmu.irq_num == irq)
527                                 return false;
528                 }
529         }
530
531         return true;
532 }
533
534 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
535 {
536         switch (attr->attr) {
537         case KVM_ARM_VCPU_PMU_V3_IRQ: {
538                 int __user *uaddr = (int __user *)(long)attr->addr;
539                 int irq;
540
541                 if (!irqchip_in_kernel(vcpu->kvm))
542                         return -EINVAL;
543
544                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
545                         return -ENODEV;
546
547                 if (get_user(irq, uaddr))
548                         return -EFAULT;
549
550                 /* The PMU overflow interrupt can be a PPI or a valid SPI. */
551                 if (!(irq_is_ppi(irq) || irq_is_spi(irq)))
552                         return -EINVAL;
553
554                 if (!pmu_irq_is_valid(vcpu->kvm, irq))
555                         return -EINVAL;
556
557                 if (kvm_arm_pmu_irq_initialized(vcpu))
558                         return -EBUSY;
559
560                 kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
561                 vcpu->arch.pmu.irq_num = irq;
562                 return 0;
563         }
564         case KVM_ARM_VCPU_PMU_V3_INIT:
565                 return kvm_arm_pmu_v3_init(vcpu);
566         }
567
568         return -ENXIO;
569 }
570
571 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
572 {
573         switch (attr->attr) {
574         case KVM_ARM_VCPU_PMU_V3_IRQ: {
575                 int __user *uaddr = (int __user *)(long)attr->addr;
576                 int irq;
577
578                 if (!irqchip_in_kernel(vcpu->kvm))
579                         return -EINVAL;
580
581                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
582                         return -ENODEV;
583
584                 if (!kvm_arm_pmu_irq_initialized(vcpu))
585                         return -ENXIO;
586
587                 irq = vcpu->arch.pmu.irq_num;
588                 return put_user(irq, uaddr);
589         }
590         }
591
592         return -ENXIO;
593 }
594
595 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
596 {
597         switch (attr->attr) {
598         case KVM_ARM_VCPU_PMU_V3_IRQ:
599         case KVM_ARM_VCPU_PMU_V3_INIT:
600                 if (kvm_arm_support_pmu_v3() &&
601                     test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
602                         return 0;
603         }
604
605         return -ENXIO;
606 }