xref: /linux/arch/x86/kvm/pmu.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
1 /*
2  * Kernel-based Virtual Machine -- Performane Monitoring Unit support
3  *
4  * Copyright 2011 Red Hat, Inc. and/or its affiliates.
5  *
6  * Authors:
7  *   Avi Kivity   <avi@redhat.com>
8  *   Gleb Natapov <gleb@redhat.com>
9  *
10  * This work is licensed under the terms of the GNU GPL, version 2.  See
11  * the COPYING file in the top-level directory.
12  *
13  */
14 
15 #include <linux/types.h>
16 #include <linux/kvm_host.h>
17 #include <linux/perf_event.h>
18 #include "x86.h"
19 #include "cpuid.h"
20 #include "lapic.h"
21 
22 static struct kvm_arch_event_perf_mapping {
23 	u8 eventsel;
24 	u8 unit_mask;
25 	unsigned event_type;
26 	bool inexact;
27 } arch_events[] = {
28 	/* Index must match CPUID 0x0A.EBX bit vector */
29 	[0] = { 0x3c, 0x00, PERF_COUNT_HW_CPU_CYCLES },
30 	[1] = { 0xc0, 0x00, PERF_COUNT_HW_INSTRUCTIONS },
31 	[2] = { 0x3c, 0x01, PERF_COUNT_HW_BUS_CYCLES  },
32 	[3] = { 0x2e, 0x4f, PERF_COUNT_HW_CACHE_REFERENCES },
33 	[4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES },
34 	[5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
35 	[6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES },
36 };
37 
38 /* mapping between fixed pmc index and arch_events array */
39 int fixed_pmc_events[] = {1, 0, 2};
40 
41 static bool pmc_is_gp(struct kvm_pmc *pmc)
42 {
43 	return pmc->type == KVM_PMC_GP;
44 }
45 
46 static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
47 {
48 	struct kvm_pmu *pmu = &pmc->vcpu->arch.pmu;
49 
50 	return pmu->counter_bitmask[pmc->type];
51 }
52 
53 static inline bool pmc_enabled(struct kvm_pmc *pmc)
54 {
55 	struct kvm_pmu *pmu = &pmc->vcpu->arch.pmu;
56 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
57 }
58 
59 static inline struct kvm_pmc *get_gp_pmc(struct kvm_pmu *pmu, u32 msr,
60 					 u32 base)
61 {
62 	if (msr >= base && msr < base + pmu->nr_arch_gp_counters)
63 		return &pmu->gp_counters[msr - base];
64 	return NULL;
65 }
66 
67 static inline struct kvm_pmc *get_fixed_pmc(struct kvm_pmu *pmu, u32 msr)
68 {
69 	int base = MSR_CORE_PERF_FIXED_CTR0;
70 	if (msr >= base && msr < base + pmu->nr_arch_fixed_counters)
71 		return &pmu->fixed_counters[msr - base];
72 	return NULL;
73 }
74 
75 static inline struct kvm_pmc *get_fixed_pmc_idx(struct kvm_pmu *pmu, int idx)
76 {
77 	return get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + idx);
78 }
79 
80 static struct kvm_pmc *global_idx_to_pmc(struct kvm_pmu *pmu, int idx)
81 {
82 	if (idx < X86_PMC_IDX_FIXED)
83 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + idx, MSR_P6_EVNTSEL0);
84 	else
85 		return get_fixed_pmc_idx(pmu, idx - X86_PMC_IDX_FIXED);
86 }
87 
88 void kvm_deliver_pmi(struct kvm_vcpu *vcpu)
89 {
90 	if (vcpu->arch.apic)
91 		kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
92 }
93 
94 static void trigger_pmi(struct irq_work *irq_work)
95 {
96 	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu,
97 			irq_work);
98 	struct kvm_vcpu *vcpu = container_of(pmu, struct kvm_vcpu,
99 			arch.pmu);
100 
101 	kvm_deliver_pmi(vcpu);
102 }
103 
104 static void kvm_perf_overflow(struct perf_event *perf_event,
105 			      struct perf_sample_data *data,
106 			      struct pt_regs *regs)
107 {
108 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
109 	struct kvm_pmu *pmu = &pmc->vcpu->arch.pmu;
110 	__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
111 }
112 
113 static void kvm_perf_overflow_intr(struct perf_event *perf_event,
114 		struct perf_sample_data *data, struct pt_regs *regs)
115 {
116 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
117 	struct kvm_pmu *pmu = &pmc->vcpu->arch.pmu;
118 	if (!test_and_set_bit(pmc->idx, (unsigned long *)&pmu->reprogram_pmi)) {
119 		kvm_perf_overflow(perf_event, data, regs);
120 		kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
121 		/*
122 		 * Inject PMI. If vcpu was in a guest mode during NMI PMI
123 		 * can be ejected on a guest mode re-entry. Otherwise we can't
124 		 * be sure that vcpu wasn't executing hlt instruction at the
125 		 * time of vmexit and is not going to re-enter guest mode until,
126 		 * woken up. So we should wake it, but this is impossible from
127 		 * NMI context. Do it from irq work instead.
128 		 */
129 		if (!kvm_is_in_guest())
130 			irq_work_queue(&pmc->vcpu->arch.pmu.irq_work);
131 		else
132 			kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
133 	}
134 }
135 
136 static u64 read_pmc(struct kvm_pmc *pmc)
137 {
138 	u64 counter, enabled, running;
139 
140 	counter = pmc->counter;
141 
142 	if (pmc->perf_event)
143 		counter += perf_event_read_value(pmc->perf_event,
144 						 &enabled, &running);
145 
146 	/* FIXME: Scaling needed? */
147 
148 	return counter & pmc_bitmask(pmc);
149 }
150 
151 static void stop_counter(struct kvm_pmc *pmc)
152 {
153 	if (pmc->perf_event) {
154 		pmc->counter = read_pmc(pmc);
155 		perf_event_release_kernel(pmc->perf_event);
156 		pmc->perf_event = NULL;
157 	}
158 }
159 
160 static void reprogram_counter(struct kvm_pmc *pmc, u32 type,
161 		unsigned config, bool exclude_user, bool exclude_kernel,
162 		bool intr)
163 {
164 	struct perf_event *event;
165 	struct perf_event_attr attr = {
166 		.type = type,
167 		.size = sizeof(attr),
168 		.pinned = true,
169 		.exclude_idle = true,
170 		.exclude_host = 1,
171 		.exclude_user = exclude_user,
172 		.exclude_kernel = exclude_kernel,
173 		.config = config,
174 	};
175 
176 	attr.sample_period = (-pmc->counter) & pmc_bitmask(pmc);
177 
178 	event = perf_event_create_kernel_counter(&attr, -1, current,
179 						 intr ? kvm_perf_overflow_intr :
180 						 kvm_perf_overflow, pmc);
181 	if (IS_ERR(event)) {
182 		printk_once("kvm: pmu event creation failed %ld\n",
183 				PTR_ERR(event));
184 		return;
185 	}
186 
187 	pmc->perf_event = event;
188 	clear_bit(pmc->idx, (unsigned long*)&pmc->vcpu->arch.pmu.reprogram_pmi);
189 }
190 
191 static unsigned find_arch_event(struct kvm_pmu *pmu, u8 event_select,
192 		u8 unit_mask)
193 {
194 	int i;
195 
196 	for (i = 0; i < ARRAY_SIZE(arch_events); i++)
197 		if (arch_events[i].eventsel == event_select
198 				&& arch_events[i].unit_mask == unit_mask
199 				&& (pmu->available_event_types & (1 << i)))
200 			break;
201 
202 	if (i == ARRAY_SIZE(arch_events))
203 		return PERF_COUNT_HW_MAX;
204 
205 	return arch_events[i].event_type;
206 }
207 
208 static void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
209 {
210 	unsigned config, type = PERF_TYPE_RAW;
211 	u8 event_select, unit_mask;
212 
213 	pmc->eventsel = eventsel;
214 
215 	stop_counter(pmc);
216 
217 	if (!(eventsel & ARCH_PERFMON_EVENTSEL_ENABLE) || !pmc_enabled(pmc))
218 		return;
219 
220 	event_select = eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
221 	unit_mask = (eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
222 
223 	if (!(event_select & (ARCH_PERFMON_EVENTSEL_EDGE |
224 				ARCH_PERFMON_EVENTSEL_INV |
225 				ARCH_PERFMON_EVENTSEL_CMASK))) {
226 		config = find_arch_event(&pmc->vcpu->arch.pmu, event_select,
227 				unit_mask);
228 		if (config != PERF_COUNT_HW_MAX)
229 			type = PERF_TYPE_HARDWARE;
230 	}
231 
232 	if (type == PERF_TYPE_RAW)
233 		config = eventsel & X86_RAW_EVENT_MASK;
234 
235 	reprogram_counter(pmc, type, config,
236 			!(eventsel & ARCH_PERFMON_EVENTSEL_USR),
237 			!(eventsel & ARCH_PERFMON_EVENTSEL_OS),
238 			eventsel & ARCH_PERFMON_EVENTSEL_INT);
239 }
240 
241 static void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 en_pmi, int idx)
242 {
243 	unsigned en = en_pmi & 0x3;
244 	bool pmi = en_pmi & 0x8;
245 
246 	stop_counter(pmc);
247 
248 	if (!en || !pmc_enabled(pmc))
249 		return;
250 
251 	reprogram_counter(pmc, PERF_TYPE_HARDWARE,
252 			arch_events[fixed_pmc_events[idx]].event_type,
253 			!(en & 0x2), /* exclude user */
254 			!(en & 0x1), /* exclude kernel */
255 			pmi);
256 }
257 
258 static inline u8 fixed_en_pmi(u64 ctrl, int idx)
259 {
260 	return (ctrl >> (idx * 4)) & 0xf;
261 }
262 
263 static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
264 {
265 	int i;
266 
267 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
268 		u8 en_pmi = fixed_en_pmi(data, i);
269 		struct kvm_pmc *pmc = get_fixed_pmc_idx(pmu, i);
270 
271 		if (fixed_en_pmi(pmu->fixed_ctr_ctrl, i) == en_pmi)
272 			continue;
273 
274 		reprogram_fixed_counter(pmc, en_pmi, i);
275 	}
276 
277 	pmu->fixed_ctr_ctrl = data;
278 }
279 
280 static void reprogram_idx(struct kvm_pmu *pmu, int idx)
281 {
282 	struct kvm_pmc *pmc = global_idx_to_pmc(pmu, idx);
283 
284 	if (!pmc)
285 		return;
286 
287 	if (pmc_is_gp(pmc))
288 		reprogram_gp_counter(pmc, pmc->eventsel);
289 	else {
290 		int fidx = idx - X86_PMC_IDX_FIXED;
291 		reprogram_fixed_counter(pmc,
292 				fixed_en_pmi(pmu->fixed_ctr_ctrl, fidx), fidx);
293 	}
294 }
295 
296 static void global_ctrl_changed(struct kvm_pmu *pmu, u64 data)
297 {
298 	int bit;
299 	u64 diff = pmu->global_ctrl ^ data;
300 
301 	pmu->global_ctrl = data;
302 
303 	for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX)
304 		reprogram_idx(pmu, bit);
305 }
306 
307 bool kvm_pmu_msr(struct kvm_vcpu *vcpu, u32 msr)
308 {
309 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
310 	int ret;
311 
312 	switch (msr) {
313 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
314 	case MSR_CORE_PERF_GLOBAL_STATUS:
315 	case MSR_CORE_PERF_GLOBAL_CTRL:
316 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
317 		ret = pmu->version > 1;
318 		break;
319 	default:
320 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)
321 			|| get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0)
322 			|| get_fixed_pmc(pmu, msr);
323 		break;
324 	}
325 	return ret;
326 }
327 
328 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, u32 index, u64 *data)
329 {
330 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
331 	struct kvm_pmc *pmc;
332 
333 	switch (index) {
334 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
335 		*data = pmu->fixed_ctr_ctrl;
336 		return 0;
337 	case MSR_CORE_PERF_GLOBAL_STATUS:
338 		*data = pmu->global_status;
339 		return 0;
340 	case MSR_CORE_PERF_GLOBAL_CTRL:
341 		*data = pmu->global_ctrl;
342 		return 0;
343 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
344 		*data = pmu->global_ovf_ctrl;
345 		return 0;
346 	default:
347 		if ((pmc = get_gp_pmc(pmu, index, MSR_IA32_PERFCTR0)) ||
348 				(pmc = get_fixed_pmc(pmu, index))) {
349 			*data = read_pmc(pmc);
350 			return 0;
351 		} else if ((pmc = get_gp_pmc(pmu, index, MSR_P6_EVNTSEL0))) {
352 			*data = pmc->eventsel;
353 			return 0;
354 		}
355 	}
356 	return 1;
357 }
358 
359 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, u32 index, u64 data)
360 {
361 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
362 	struct kvm_pmc *pmc;
363 
364 	switch (index) {
365 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
366 		if (pmu->fixed_ctr_ctrl == data)
367 			return 0;
368 		if (!(data & 0xfffffffffffff444)) {
369 			reprogram_fixed_counters(pmu, data);
370 			return 0;
371 		}
372 		break;
373 	case MSR_CORE_PERF_GLOBAL_STATUS:
374 		break; /* RO MSR */
375 	case MSR_CORE_PERF_GLOBAL_CTRL:
376 		if (pmu->global_ctrl == data)
377 			return 0;
378 		if (!(data & pmu->global_ctrl_mask)) {
379 			global_ctrl_changed(pmu, data);
380 			return 0;
381 		}
382 		break;
383 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
384 		if (!(data & (pmu->global_ctrl_mask & ~(3ull<<62)))) {
385 			pmu->global_status &= ~data;
386 			pmu->global_ovf_ctrl = data;
387 			return 0;
388 		}
389 		break;
390 	default:
391 		if ((pmc = get_gp_pmc(pmu, index, MSR_IA32_PERFCTR0)) ||
392 				(pmc = get_fixed_pmc(pmu, index))) {
393 			data = (s64)(s32)data;
394 			pmc->counter += data - read_pmc(pmc);
395 			return 0;
396 		} else if ((pmc = get_gp_pmc(pmu, index, MSR_P6_EVNTSEL0))) {
397 			if (data == pmc->eventsel)
398 				return 0;
399 			if (!(data & 0xffffffff00200000ull)) {
400 				reprogram_gp_counter(pmc, data);
401 				return 0;
402 			}
403 		}
404 	}
405 	return 1;
406 }
407 
408 int kvm_pmu_read_pmc(struct kvm_vcpu *vcpu, unsigned pmc, u64 *data)
409 {
410 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
411 	bool fast_mode = pmc & (1u << 31);
412 	bool fixed = pmc & (1u << 30);
413 	struct kvm_pmc *counters;
414 	u64 ctr;
415 
416 	pmc &= (3u << 30) - 1;
417 	if (!fixed && pmc >= pmu->nr_arch_gp_counters)
418 		return 1;
419 	if (fixed && pmc >= pmu->nr_arch_fixed_counters)
420 		return 1;
421 	counters = fixed ? pmu->fixed_counters : pmu->gp_counters;
422 	ctr = read_pmc(&counters[pmc]);
423 	if (fast_mode)
424 		ctr = (u32)ctr;
425 	*data = ctr;
426 
427 	return 0;
428 }
429 
430 void kvm_pmu_cpuid_update(struct kvm_vcpu *vcpu)
431 {
432 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
433 	struct kvm_cpuid_entry2 *entry;
434 	unsigned bitmap_len;
435 
436 	pmu->nr_arch_gp_counters = 0;
437 	pmu->nr_arch_fixed_counters = 0;
438 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
439 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
440 	pmu->version = 0;
441 
442 	entry = kvm_find_cpuid_entry(vcpu, 0xa, 0);
443 	if (!entry)
444 		return;
445 
446 	pmu->version = entry->eax & 0xff;
447 	if (!pmu->version)
448 		return;
449 
450 	pmu->nr_arch_gp_counters = min((int)(entry->eax >> 8) & 0xff,
451 			X86_PMC_MAX_GENERIC);
452 	pmu->counter_bitmask[KVM_PMC_GP] =
453 		((u64)1 << ((entry->eax >> 16) & 0xff)) - 1;
454 	bitmap_len = (entry->eax >> 24) & 0xff;
455 	pmu->available_event_types = ~entry->ebx & ((1ull << bitmap_len) - 1);
456 
457 	if (pmu->version == 1) {
458 		pmu->global_ctrl = (1 << pmu->nr_arch_gp_counters) - 1;
459 		return;
460 	}
461 
462 	pmu->nr_arch_fixed_counters = min((int)(entry->edx & 0x1f),
463 			X86_PMC_MAX_FIXED);
464 	pmu->counter_bitmask[KVM_PMC_FIXED] =
465 		((u64)1 << ((entry->edx >> 5) & 0xff)) - 1;
466 	pmu->global_ctrl_mask = ~(((1 << pmu->nr_arch_gp_counters) - 1)
467 			| (((1ull << pmu->nr_arch_fixed_counters) - 1)
468 				<< X86_PMC_IDX_FIXED));
469 }
470 
471 void kvm_pmu_init(struct kvm_vcpu *vcpu)
472 {
473 	int i;
474 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
475 
476 	memset(pmu, 0, sizeof(*pmu));
477 	for (i = 0; i < X86_PMC_MAX_GENERIC; i++) {
478 		pmu->gp_counters[i].type = KVM_PMC_GP;
479 		pmu->gp_counters[i].vcpu = vcpu;
480 		pmu->gp_counters[i].idx = i;
481 	}
482 	for (i = 0; i < X86_PMC_MAX_FIXED; i++) {
483 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
484 		pmu->fixed_counters[i].vcpu = vcpu;
485 		pmu->fixed_counters[i].idx = i + X86_PMC_IDX_FIXED;
486 	}
487 	init_irq_work(&pmu->irq_work, trigger_pmi);
488 	kvm_pmu_cpuid_update(vcpu);
489 }
490 
491 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
492 {
493 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
494 	int i;
495 
496 	irq_work_sync(&pmu->irq_work);
497 	for (i = 0; i < X86_PMC_MAX_GENERIC; i++) {
498 		struct kvm_pmc *pmc = &pmu->gp_counters[i];
499 		stop_counter(pmc);
500 		pmc->counter = pmc->eventsel = 0;
501 	}
502 
503 	for (i = 0; i < X86_PMC_MAX_FIXED; i++)
504 		stop_counter(&pmu->fixed_counters[i]);
505 
506 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status =
507 		pmu->global_ovf_ctrl = 0;
508 }
509 
510 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
511 {
512 	kvm_pmu_reset(vcpu);
513 }
514 
515 void kvm_handle_pmu_event(struct kvm_vcpu *vcpu)
516 {
517 	struct kvm_pmu *pmu = &vcpu->arch.pmu;
518 	u64 bitmask;
519 	int bit;
520 
521 	bitmask = pmu->reprogram_pmi;
522 
523 	for_each_set_bit(bit, (unsigned long *)&bitmask, X86_PMC_IDX_MAX) {
524 		struct kvm_pmc *pmc = global_idx_to_pmc(pmu, bit);
525 
526 		if (unlikely(!pmc || !pmc->perf_event)) {
527 			clear_bit(bit, (unsigned long *)&pmu->reprogram_pmi);
528 			continue;
529 		}
530 
531 		reprogram_idx(pmu, bit);
532 	}
533 }
534