xref: /linux/arch/x86/kvm/pmu.c (revision 856e7c4b619af622d56b3b454f7bec32a170ac99)
1 /*
2  * Kernel-based Virtual Machine -- Performance Monitoring Unit support
3  *
4  * Copyright 2015 Red Hat, Inc. and/or its affiliates.
5  *
6  * Authors:
7  *   Avi Kivity   <avi@redhat.com>
8  *   Gleb Natapov <gleb@redhat.com>
9  *   Wei Huang    <wei@redhat.com>
10  *
11  * This work is licensed under the terms of the GNU GPL, version 2.  See
12  * the COPYING file in the top-level directory.
13  *
14  */
15 
16 #include <linux/types.h>
17 #include <linux/kvm_host.h>
18 #include <linux/perf_event.h>
19 #include <asm/perf_event.h>
20 #include "x86.h"
21 #include "cpuid.h"
22 #include "lapic.h"
23 #include "pmu.h"
24 
25 /* NOTE:
26  * - Each perf counter is defined as "struct kvm_pmc";
27  * - There are two types of perf counters: general purpose (gp) and fixed.
28  *   gp counters are stored in gp_counters[] and fixed counters are stored
29  *   in fixed_counters[] respectively. Both of them are part of "struct
30  *   kvm_pmu";
31  * - pmu.c understands the difference between gp counters and fixed counters.
32  *   However AMD doesn't support fixed-counters;
33  * - There are three types of index to access perf counters (PMC):
34  *     1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
35  *        has MSR_K7_PERFCTRn.
36  *     2. MSR Index (named idx): This normally is used by RDPMC instruction.
37  *        For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
38  *        C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
39  *        that it also supports fixed counters. idx can be used to as index to
40  *        gp and fixed counters.
41  *     3. Global PMC Index (named pmc): pmc is an index specific to PMU
42  *        code. Each pmc, stored in kvm_pmc.idx field, is unique across
43  *        all perf counters (both gp and fixed). The mapping relationship
44  *        between pmc and perf counters is as the following:
45  *        * Intel: [0 .. INTEL_PMC_MAX_GENERIC-1] <=> gp counters
46  *                 [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
47  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
48  */
49 
50 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
51 {
52 	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
53 	struct kvm_vcpu *vcpu = pmu_to_vcpu(pmu);
54 
55 	kvm_pmu_deliver_pmi(vcpu);
56 }
57 
58 static void kvm_perf_overflow(struct perf_event *perf_event,
59 			      struct perf_sample_data *data,
60 			      struct pt_regs *regs)
61 {
62 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
63 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
64 
65 	if (!test_and_set_bit(pmc->idx,
66 			      (unsigned long *)&pmu->reprogram_pmi)) {
67 		__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
68 		kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
69 	}
70 }
71 
72 static void kvm_perf_overflow_intr(struct perf_event *perf_event,
73 				   struct perf_sample_data *data,
74 				   struct pt_regs *regs)
75 {
76 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
77 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
78 
79 	if (!test_and_set_bit(pmc->idx,
80 			      (unsigned long *)&pmu->reprogram_pmi)) {
81 		__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
82 		kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
83 
84 		/*
85 		 * Inject PMI. If vcpu was in a guest mode during NMI PMI
86 		 * can be ejected on a guest mode re-entry. Otherwise we can't
87 		 * be sure that vcpu wasn't executing hlt instruction at the
88 		 * time of vmexit and is not going to re-enter guest mode until
89 		 * woken up. So we should wake it, but this is impossible from
90 		 * NMI context. Do it from irq work instead.
91 		 */
92 		if (!kvm_is_in_guest())
93 			irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
94 		else
95 			kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
96 	}
97 }
98 
99 static void pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type,
100 				  unsigned config, bool exclude_user,
101 				  bool exclude_kernel, bool intr,
102 				  bool in_tx, bool in_tx_cp)
103 {
104 	struct perf_event *event;
105 	struct perf_event_attr attr = {
106 		.type = type,
107 		.size = sizeof(attr),
108 		.pinned = true,
109 		.exclude_idle = true,
110 		.exclude_host = 1,
111 		.exclude_user = exclude_user,
112 		.exclude_kernel = exclude_kernel,
113 		.config = config,
114 	};
115 
116 	attr.sample_period = (-pmc->counter) & pmc_bitmask(pmc);
117 
118 	if (in_tx)
119 		attr.config |= HSW_IN_TX;
120 	if (in_tx_cp) {
121 		/*
122 		 * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
123 		 * period. Just clear the sample period so at least
124 		 * allocating the counter doesn't fail.
125 		 */
126 		attr.sample_period = 0;
127 		attr.config |= HSW_IN_TX_CHECKPOINTED;
128 	}
129 
130 	event = perf_event_create_kernel_counter(&attr, -1, current,
131 						 intr ? kvm_perf_overflow_intr :
132 						 kvm_perf_overflow, pmc);
133 	if (IS_ERR(event)) {
134 		printk_once("kvm_pmu: event creation failed %ld\n",
135 			    PTR_ERR(event));
136 		return;
137 	}
138 
139 	pmc->perf_event = event;
140 	clear_bit(pmc->idx, (unsigned long*)&pmc_to_pmu(pmc)->reprogram_pmi);
141 }
142 
143 void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
144 {
145 	unsigned config, type = PERF_TYPE_RAW;
146 	u8 event_select, unit_mask;
147 
148 	if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
149 		printk_once("kvm pmu: pin control bit is ignored\n");
150 
151 	pmc->eventsel = eventsel;
152 
153 	pmc_stop_counter(pmc);
154 
155 	if (!(eventsel & ARCH_PERFMON_EVENTSEL_ENABLE) || !pmc_is_enabled(pmc))
156 		return;
157 
158 	event_select = eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
159 	unit_mask = (eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
160 
161 	if (!(eventsel & (ARCH_PERFMON_EVENTSEL_EDGE |
162 			  ARCH_PERFMON_EVENTSEL_INV |
163 			  ARCH_PERFMON_EVENTSEL_CMASK |
164 			  HSW_IN_TX |
165 			  HSW_IN_TX_CHECKPOINTED))) {
166 		config = kvm_x86_ops->pmu_ops->find_arch_event(pmc_to_pmu(pmc),
167 						      event_select,
168 						      unit_mask);
169 		if (config != PERF_COUNT_HW_MAX)
170 			type = PERF_TYPE_HARDWARE;
171 	}
172 
173 	if (type == PERF_TYPE_RAW)
174 		config = eventsel & X86_RAW_EVENT_MASK;
175 
176 	pmc_reprogram_counter(pmc, type, config,
177 			      !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
178 			      !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
179 			      eventsel & ARCH_PERFMON_EVENTSEL_INT,
180 			      (eventsel & HSW_IN_TX),
181 			      (eventsel & HSW_IN_TX_CHECKPOINTED));
182 }
183 EXPORT_SYMBOL_GPL(reprogram_gp_counter);
184 
185 void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
186 {
187 	unsigned en_field = ctrl & 0x3;
188 	bool pmi = ctrl & 0x8;
189 
190 	pmc_stop_counter(pmc);
191 
192 	if (!en_field || !pmc_is_enabled(pmc))
193 		return;
194 
195 	pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
196 			      kvm_x86_ops->pmu_ops->find_fixed_event(idx),
197 			      !(en_field & 0x2), /* exclude user */
198 			      !(en_field & 0x1), /* exclude kernel */
199 			      pmi, false, false);
200 }
201 EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
202 
203 void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
204 {
205 	struct kvm_pmc *pmc = kvm_x86_ops->pmu_ops->pmc_idx_to_pmc(pmu, pmc_idx);
206 
207 	if (!pmc)
208 		return;
209 
210 	if (pmc_is_gp(pmc))
211 		reprogram_gp_counter(pmc, pmc->eventsel);
212 	else {
213 		int idx = pmc_idx - INTEL_PMC_IDX_FIXED;
214 		u8 ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, idx);
215 
216 		reprogram_fixed_counter(pmc, ctrl, idx);
217 	}
218 }
219 EXPORT_SYMBOL_GPL(reprogram_counter);
220 
221 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
222 {
223 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
224 	u64 bitmask;
225 	int bit;
226 
227 	bitmask = pmu->reprogram_pmi;
228 
229 	for_each_set_bit(bit, (unsigned long *)&bitmask, X86_PMC_IDX_MAX) {
230 		struct kvm_pmc *pmc = kvm_x86_ops->pmu_ops->pmc_idx_to_pmc(pmu, bit);
231 
232 		if (unlikely(!pmc || !pmc->perf_event)) {
233 			clear_bit(bit, (unsigned long *)&pmu->reprogram_pmi);
234 			continue;
235 		}
236 
237 		reprogram_counter(pmu, bit);
238 	}
239 }
240 
241 /* check if idx is a valid index to access PMU */
242 int kvm_pmu_is_valid_msr_idx(struct kvm_vcpu *vcpu, unsigned idx)
243 {
244 	return kvm_x86_ops->pmu_ops->is_valid_msr_idx(vcpu, idx);
245 }
246 
247 bool is_vmware_backdoor_pmc(u32 pmc_idx)
248 {
249 	switch (pmc_idx) {
250 	case VMWARE_BACKDOOR_PMC_HOST_TSC:
251 	case VMWARE_BACKDOOR_PMC_REAL_TIME:
252 	case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
253 		return true;
254 	}
255 	return false;
256 }
257 
258 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
259 {
260 	u64 ctr_val;
261 
262 	switch (idx) {
263 	case VMWARE_BACKDOOR_PMC_HOST_TSC:
264 		ctr_val = rdtsc();
265 		break;
266 	case VMWARE_BACKDOOR_PMC_REAL_TIME:
267 		ctr_val = ktime_get_boot_ns();
268 		break;
269 	case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
270 		ctr_val = ktime_get_boot_ns() +
271 			vcpu->kvm->arch.kvmclock_offset;
272 		break;
273 	default:
274 		return 1;
275 	}
276 
277 	*data = ctr_val;
278 	return 0;
279 }
280 
281 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
282 {
283 	bool fast_mode = idx & (1u << 31);
284 	struct kvm_pmc *pmc;
285 	u64 ctr_val;
286 
287 	if (is_vmware_backdoor_pmc(idx))
288 		return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
289 
290 	pmc = kvm_x86_ops->pmu_ops->msr_idx_to_pmc(vcpu, idx);
291 	if (!pmc)
292 		return 1;
293 
294 	ctr_val = pmc_read_counter(pmc);
295 	if (fast_mode)
296 		ctr_val = (u32)ctr_val;
297 
298 	*data = ctr_val;
299 	return 0;
300 }
301 
302 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
303 {
304 	if (lapic_in_kernel(vcpu))
305 		kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
306 }
307 
308 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
309 {
310 	return kvm_x86_ops->pmu_ops->is_valid_msr(vcpu, msr);
311 }
312 
313 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, u32 msr, u64 *data)
314 {
315 	return kvm_x86_ops->pmu_ops->get_msr(vcpu, msr, data);
316 }
317 
318 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
319 {
320 	return kvm_x86_ops->pmu_ops->set_msr(vcpu, msr_info);
321 }
322 
323 /* refresh PMU settings. This function generally is called when underlying
324  * settings are changed (such as changes of PMU CPUID by guest VMs), which
325  * should rarely happen.
326  */
327 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
328 {
329 	kvm_x86_ops->pmu_ops->refresh(vcpu);
330 }
331 
332 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
333 {
334 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
335 
336 	irq_work_sync(&pmu->irq_work);
337 	kvm_x86_ops->pmu_ops->reset(vcpu);
338 }
339 
340 void kvm_pmu_init(struct kvm_vcpu *vcpu)
341 {
342 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
343 
344 	memset(pmu, 0, sizeof(*pmu));
345 	kvm_x86_ops->pmu_ops->init(vcpu);
346 	init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
347 	kvm_pmu_refresh(vcpu);
348 }
349 
350 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
351 {
352 	kvm_pmu_reset(vcpu);
353 }
354