1 // SPDX-License-Identifier: GPL-2.0-only
3 * Kernel-based Virtual Machine -- Performance Monitoring Unit support
5 * Copyright 2015 Red Hat, Inc. and/or its affiliates.
8 * Avi Kivity <avi@redhat.com>
9 * Gleb Natapov <gleb@redhat.com>
10 * Wei Huang <wei@redhat.com>
13 #include <linux/types.h>
14 #include <linux/kvm_host.h>
15 #include <linux/perf_event.h>
16 #include <linux/bsearch.h>
17 #include <linux/sort.h>
18 #include <asm/perf_event.h>
19 #include <asm/cpu_device_id.h>
25 /* This is enough to filter the vast majority of currently defined events. */
26 #define KVM_PMU_EVENT_FILTER_MAX_EVENTS 300
28 struct x86_pmu_capability __read_mostly kvm_pmu_cap;
29 EXPORT_SYMBOL_GPL(kvm_pmu_cap);
31 static const struct x86_cpu_id vmx_icl_pebs_cpu[] = {
32 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, NULL),
33 X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, NULL),
38 * - Each perf counter is defined as "struct kvm_pmc";
39 * - There are two types of perf counters: general purpose (gp) and fixed.
40 * gp counters are stored in gp_counters[] and fixed counters are stored
41 * in fixed_counters[] respectively. Both of them are part of "struct
43 * - pmu.c understands the difference between gp counters and fixed counters.
44 * However AMD doesn't support fixed-counters;
45 * - There are three types of index to access perf counters (PMC):
46 * 1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
47 * has MSR_K7_PERFCTRn and, for families 15H and later,
48 * MSR_F15H_PERF_CTRn, where MSR_F15H_PERF_CTR[0-3] are
49 * aliased to MSR_K7_PERFCTRn.
50 * 2. MSR Index (named idx): This normally is used by RDPMC instruction.
51 * For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
52 * C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
53 * that it also supports fixed counters. idx can be used to as index to
54 * gp and fixed counters.
55 * 3. Global PMC Index (named pmc): pmc is an index specific to PMU
56 * code. Each pmc, stored in kvm_pmc.idx field, is unique across
57 * all perf counters (both gp and fixed). The mapping relationship
58 * between pmc and perf counters is as the following:
59 * * Intel: [0 .. KVM_INTEL_PMC_MAX_GENERIC-1] <=> gp counters
60 * [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
61 * * AMD: [0 .. AMD64_NUM_COUNTERS-1] and, for families 15H
62 * and later, [0 .. AMD64_NUM_COUNTERS_CORE-1] <=> gp counters
65 static struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
67 #define KVM_X86_PMU_OP(func) \
68 DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func, \
69 *(((struct kvm_pmu_ops *)0)->func));
70 #define KVM_X86_PMU_OP_OPTIONAL KVM_X86_PMU_OP
71 #include <asm/kvm-x86-pmu-ops.h>
73 void kvm_pmu_ops_update(const struct kvm_pmu_ops *pmu_ops)
75 memcpy(&kvm_pmu_ops, pmu_ops, sizeof(kvm_pmu_ops));
77 #define __KVM_X86_PMU_OP(func) \
78 static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func);
79 #define KVM_X86_PMU_OP(func) \
80 WARN_ON(!kvm_pmu_ops.func); __KVM_X86_PMU_OP(func)
81 #define KVM_X86_PMU_OP_OPTIONAL __KVM_X86_PMU_OP
82 #include <asm/kvm-x86-pmu-ops.h>
83 #undef __KVM_X86_PMU_OP
86 static inline bool pmc_is_enabled(struct kvm_pmc *pmc)
88 return static_call(kvm_x86_pmu_pmc_is_enabled)(pmc);
91 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
93 struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
94 struct kvm_vcpu *vcpu = pmu_to_vcpu(pmu);
96 kvm_pmu_deliver_pmi(vcpu);
99 static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
101 struct kvm_pmu *pmu = pmc_to_pmu(pmc);
102 bool skip_pmi = false;
104 if (pmc->perf_event && pmc->perf_event->attr.precise_ip) {
107 * TODO: KVM is currently _choosing_ to not generate records
108 * for emulated instructions, avoiding BUFFER_OVF PMI when
109 * there are no records. Strictly speaking, it should be done
110 * as well in the right context to improve sampling accuracy.
114 /* Indicate PEBS overflow PMI to guest. */
115 skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
116 (unsigned long *)&pmu->global_status);
119 __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
122 if (!pmc->intr || skip_pmi)
126 * Inject PMI. If vcpu was in a guest mode during NMI PMI
127 * can be ejected on a guest mode re-entry. Otherwise we can't
128 * be sure that vcpu wasn't executing hlt instruction at the
129 * time of vmexit and is not going to re-enter guest mode until
130 * woken up. So we should wake it, but this is impossible from
131 * NMI context. Do it from irq work instead.
133 if (in_pmi && !kvm_handling_nmi_from_guest(pmc->vcpu))
134 irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
136 kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
139 static void kvm_perf_overflow(struct perf_event *perf_event,
140 struct perf_sample_data *data,
141 struct pt_regs *regs)
143 struct kvm_pmc *pmc = perf_event->overflow_handler_context;
146 * Ignore overflow events for counters that are scheduled to be
147 * reprogrammed, e.g. if a PMI for the previous event races with KVM's
148 * handling of a related guest WRMSR.
150 if (test_and_set_bit(pmc->idx, pmc_to_pmu(pmc)->reprogram_pmi))
153 __kvm_perf_overflow(pmc, true);
155 kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
158 static int pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type, u64 config,
159 bool exclude_user, bool exclude_kernel,
162 struct kvm_pmu *pmu = pmc_to_pmu(pmc);
163 struct perf_event *event;
164 struct perf_event_attr attr = {
166 .size = sizeof(attr),
168 .exclude_idle = true,
170 .exclude_user = exclude_user,
171 .exclude_kernel = exclude_kernel,
174 bool pebs = test_bit(pmc->idx, (unsigned long *)&pmu->pebs_enable);
176 attr.sample_period = get_sample_period(pmc, pmc->counter);
178 if ((attr.config & HSW_IN_TX_CHECKPOINTED) &&
179 guest_cpuid_is_intel(pmc->vcpu)) {
181 * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
182 * period. Just clear the sample period so at least
183 * allocating the counter doesn't fail.
185 attr.sample_period = 0;
189 * The non-zero precision level of guest event makes the ordinary
190 * guest event becomes a guest PEBS event and triggers the host
191 * PEBS PMI handler to determine whether the PEBS overflow PMI
192 * comes from the host counters or the guest.
194 * For most PEBS hardware events, the difference in the software
195 * precision levels of guest and host PEBS events will not affect
196 * the accuracy of the PEBS profiling result, because the "event IP"
197 * in the PEBS record is calibrated on the guest side.
199 * On Icelake everything is fine. Other hardware (GLC+, TNT+) that
200 * could possibly care here is unsupported and needs changes.
203 if (x86_match_cpu(vmx_icl_pebs_cpu) && pmc->idx == 32)
207 event = perf_event_create_kernel_counter(&attr, -1, current,
208 kvm_perf_overflow, pmc);
210 pr_debug_ratelimited("kvm_pmu: event creation failed %ld for pmc->idx = %d\n",
211 PTR_ERR(event), pmc->idx);
212 return PTR_ERR(event);
215 pmc->perf_event = event;
216 pmc_to_pmu(pmc)->event_count++;
217 pmc->is_paused = false;
218 pmc->intr = intr || pebs;
222 static void pmc_pause_counter(struct kvm_pmc *pmc)
224 u64 counter = pmc->counter;
226 if (!pmc->perf_event || pmc->is_paused)
229 /* update counter, reset event value to avoid redundant accumulation */
230 counter += perf_event_pause(pmc->perf_event, true);
231 pmc->counter = counter & pmc_bitmask(pmc);
232 pmc->is_paused = true;
235 static bool pmc_resume_counter(struct kvm_pmc *pmc)
237 if (!pmc->perf_event)
240 /* recalibrate sample period and check if it's accepted by perf core */
241 if (is_sampling_event(pmc->perf_event) &&
242 perf_event_period(pmc->perf_event,
243 get_sample_period(pmc, pmc->counter)))
246 if (test_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->pebs_enable) !=
247 (!!pmc->perf_event->attr.precise_ip))
250 /* reuse perf_event to serve as pmc_reprogram_counter() does*/
251 perf_event_enable(pmc->perf_event);
252 pmc->is_paused = false;
257 static int cmp_u64(const void *pa, const void *pb)
262 return (a > b) - (a < b);
265 static bool check_pmu_event_filter(struct kvm_pmc *pmc)
267 struct kvm_pmu_event_filter *filter;
268 struct kvm *kvm = pmc->vcpu->kvm;
269 bool allow_event = true;
273 if (!static_call(kvm_x86_pmu_hw_event_available)(pmc))
276 filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
280 if (pmc_is_gp(pmc)) {
281 key = pmc->eventsel & AMD64_RAW_EVENT_MASK_NB;
282 if (bsearch(&key, filter->events, filter->nevents,
283 sizeof(__u64), cmp_u64))
284 allow_event = filter->action == KVM_PMU_EVENT_ALLOW;
286 allow_event = filter->action == KVM_PMU_EVENT_DENY;
288 idx = pmc->idx - INTEL_PMC_IDX_FIXED;
289 if (filter->action == KVM_PMU_EVENT_DENY &&
290 test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
292 if (filter->action == KVM_PMU_EVENT_ALLOW &&
293 !test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
301 static void reprogram_counter(struct kvm_pmc *pmc)
303 struct kvm_pmu *pmu = pmc_to_pmu(pmc);
304 u64 eventsel = pmc->eventsel;
305 u64 new_config = eventsel;
308 pmc_pause_counter(pmc);
310 if (!pmc_speculative_in_use(pmc) || !pmc_is_enabled(pmc))
311 goto reprogram_complete;
313 if (!check_pmu_event_filter(pmc))
314 goto reprogram_complete;
316 if (pmc->counter < pmc->prev_counter)
317 __kvm_perf_overflow(pmc, false);
319 if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
320 printk_once("kvm pmu: pin control bit is ignored\n");
322 if (pmc_is_fixed(pmc)) {
323 fixed_ctr_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl,
324 pmc->idx - INTEL_PMC_IDX_FIXED);
325 if (fixed_ctr_ctrl & 0x1)
326 eventsel |= ARCH_PERFMON_EVENTSEL_OS;
327 if (fixed_ctr_ctrl & 0x2)
328 eventsel |= ARCH_PERFMON_EVENTSEL_USR;
329 if (fixed_ctr_ctrl & 0x8)
330 eventsel |= ARCH_PERFMON_EVENTSEL_INT;
331 new_config = (u64)fixed_ctr_ctrl;
334 if (pmc->current_config == new_config && pmc_resume_counter(pmc))
335 goto reprogram_complete;
337 pmc_release_perf_event(pmc);
339 pmc->current_config = new_config;
342 * If reprogramming fails, e.g. due to contention, leave the counter's
343 * regprogram bit set, i.e. opportunistically try again on the next PMU
344 * refresh. Don't make a new request as doing so can stall the guest
345 * if reprogramming repeatedly fails.
347 if (pmc_reprogram_counter(pmc, PERF_TYPE_RAW,
348 (eventsel & pmu->raw_event_mask),
349 !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
350 !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
351 eventsel & ARCH_PERFMON_EVENTSEL_INT))
355 clear_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->reprogram_pmi);
356 pmc->prev_counter = 0;
359 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
361 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
364 for_each_set_bit(bit, pmu->reprogram_pmi, X86_PMC_IDX_MAX) {
365 struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
367 if (unlikely(!pmc)) {
368 clear_bit(bit, pmu->reprogram_pmi);
372 reprogram_counter(pmc);
376 * Unused perf_events are only released if the corresponding MSRs
377 * weren't accessed during the last vCPU time slice. kvm_arch_sched_in
378 * triggers KVM_REQ_PMU if cleanup is needed.
380 if (unlikely(pmu->need_cleanup))
381 kvm_pmu_cleanup(vcpu);
384 /* check if idx is a valid index to access PMU */
385 bool kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
387 return static_call(kvm_x86_pmu_is_valid_rdpmc_ecx)(vcpu, idx);
390 bool is_vmware_backdoor_pmc(u32 pmc_idx)
393 case VMWARE_BACKDOOR_PMC_HOST_TSC:
394 case VMWARE_BACKDOOR_PMC_REAL_TIME:
395 case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
401 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
406 case VMWARE_BACKDOOR_PMC_HOST_TSC:
409 case VMWARE_BACKDOOR_PMC_REAL_TIME:
410 ctr_val = ktime_get_boottime_ns();
412 case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
413 ctr_val = ktime_get_boottime_ns() +
414 vcpu->kvm->arch.kvmclock_offset;
424 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
426 bool fast_mode = idx & (1u << 31);
427 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
429 u64 mask = fast_mode ? ~0u : ~0ull;
434 if (is_vmware_backdoor_pmc(idx))
435 return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
437 pmc = static_call(kvm_x86_pmu_rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
441 if (!(kvm_read_cr4(vcpu) & X86_CR4_PCE) &&
442 (static_call(kvm_x86_get_cpl)(vcpu) != 0) &&
443 (kvm_read_cr0(vcpu) & X86_CR0_PE))
446 *data = pmc_read_counter(pmc) & mask;
450 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
452 if (lapic_in_kernel(vcpu)) {
453 static_call_cond(kvm_x86_pmu_deliver_pmi)(vcpu);
454 kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
458 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
460 return static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr) ||
461 static_call(kvm_x86_pmu_is_valid_msr)(vcpu, msr);
464 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
466 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
467 struct kvm_pmc *pmc = static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr);
470 __set_bit(pmc->idx, pmu->pmc_in_use);
473 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
475 return static_call(kvm_x86_pmu_get_msr)(vcpu, msr_info);
478 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
480 kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
481 return static_call(kvm_x86_pmu_set_msr)(vcpu, msr_info);
484 /* refresh PMU settings. This function generally is called when underlying
485 * settings are changed (such as changes of PMU CPUID by guest VMs), which
486 * should rarely happen.
488 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
490 static_call(kvm_x86_pmu_refresh)(vcpu);
493 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
495 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
497 irq_work_sync(&pmu->irq_work);
498 static_call(kvm_x86_pmu_reset)(vcpu);
501 void kvm_pmu_init(struct kvm_vcpu *vcpu)
503 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
505 memset(pmu, 0, sizeof(*pmu));
506 static_call(kvm_x86_pmu_init)(vcpu);
507 init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
508 pmu->event_count = 0;
509 pmu->need_cleanup = false;
510 kvm_pmu_refresh(vcpu);
513 /* Release perf_events for vPMCs that have been unused for a full time slice. */
514 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
516 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
517 struct kvm_pmc *pmc = NULL;
518 DECLARE_BITMAP(bitmask, X86_PMC_IDX_MAX);
521 pmu->need_cleanup = false;
523 bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
524 pmu->pmc_in_use, X86_PMC_IDX_MAX);
526 for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
527 pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
529 if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
530 pmc_stop_counter(pmc);
533 static_call_cond(kvm_x86_pmu_cleanup)(vcpu);
535 bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
538 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
543 static void kvm_pmu_incr_counter(struct kvm_pmc *pmc)
545 pmc->prev_counter = pmc->counter;
546 pmc->counter = (pmc->counter + 1) & pmc_bitmask(pmc);
547 kvm_pmu_request_counter_reprogam(pmc);
550 static inline bool eventsel_match_perf_hw_id(struct kvm_pmc *pmc,
551 unsigned int perf_hw_id)
553 return !((pmc->eventsel ^ perf_get_hw_event_config(perf_hw_id)) &
554 AMD64_RAW_EVENT_MASK_NB);
557 static inline bool cpl_is_matched(struct kvm_pmc *pmc)
559 bool select_os, select_user;
562 if (pmc_is_gp(pmc)) {
563 config = pmc->eventsel;
564 select_os = config & ARCH_PERFMON_EVENTSEL_OS;
565 select_user = config & ARCH_PERFMON_EVENTSEL_USR;
567 config = fixed_ctrl_field(pmc_to_pmu(pmc)->fixed_ctr_ctrl,
568 pmc->idx - INTEL_PMC_IDX_FIXED);
569 select_os = config & 0x1;
570 select_user = config & 0x2;
573 return (static_call(kvm_x86_get_cpl)(pmc->vcpu) == 0) ? select_os : select_user;
576 void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
578 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
582 for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
583 pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
585 if (!pmc || !pmc_is_enabled(pmc) || !pmc_speculative_in_use(pmc))
588 /* Ignore checks for edge detect, pin control, invert and CMASK bits */
589 if (eventsel_match_perf_hw_id(pmc, perf_hw_id) && cpl_is_matched(pmc))
590 kvm_pmu_incr_counter(pmc);
593 EXPORT_SYMBOL_GPL(kvm_pmu_trigger_event);
595 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
597 struct kvm_pmu_event_filter tmp, *filter;
598 struct kvm_vcpu *vcpu;
603 if (copy_from_user(&tmp, argp, sizeof(tmp)))
606 if (tmp.action != KVM_PMU_EVENT_ALLOW &&
607 tmp.action != KVM_PMU_EVENT_DENY)
613 if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
616 size = struct_size(filter, events, tmp.nevents);
617 filter = kmalloc(size, GFP_KERNEL_ACCOUNT);
622 if (copy_from_user(filter, argp, size))
625 /* Ensure nevents can't be changed between the user copies. */
629 * Sort the in-kernel list so that we can search it with bsearch.
631 sort(&filter->events, filter->nevents, sizeof(__u64), cmp_u64, NULL);
633 mutex_lock(&kvm->lock);
634 filter = rcu_replace_pointer(kvm->arch.pmu_event_filter, filter,
635 mutex_is_locked(&kvm->lock));
636 synchronize_srcu_expedited(&kvm->srcu);
638 BUILD_BUG_ON(sizeof(((struct kvm_pmu *)0)->reprogram_pmi) >
639 sizeof(((struct kvm_pmu *)0)->__reprogram_pmi));
641 kvm_for_each_vcpu(i, vcpu, kvm)
642 atomic64_set(&vcpu_to_pmu(vcpu)->__reprogram_pmi, -1ull);
644 kvm_make_all_cpus_request(kvm, KVM_REQ_PMU);
646 mutex_unlock(&kvm->lock);