xref: /linux/arch/x86/kvm/pmu.c (revision c7fdfd2bee1cf1448e5244da1a734e680f634b02)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Kernel-based Virtual Machine -- Performance Monitoring Unit support
4  *
5  * Copyright 2015 Red Hat, Inc. and/or its affiliates.
6  *
7  * Authors:
8  *   Avi Kivity   <avi@redhat.com>
9  *   Gleb Natapov <gleb@redhat.com>
10  *   Wei Huang    <wei@redhat.com>
11  */
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 
14 #include <linux/types.h>
15 #include <linux/kvm_host.h>
16 #include <linux/perf_event.h>
17 #include <linux/bsearch.h>
18 #include <linux/sort.h>
19 #include <linux/moduleparam.h>
20 #include <asm/perf_event.h>
21 #include <asm/cpu_device_id.h>
22 #include "x86.h"
23 #include "cpuid.h"
24 #include "lapic.h"
25 #include "pmu.h"
26 
27 /* This is enough to filter the vast majority of currently defined events. */
28 #define KVM_PMU_EVENT_FILTER_MAX_EVENTS 300
29 
30 /* Unadultered PMU capabilities of the host, i.e. of hardware. */
31 static struct x86_pmu_capability __read_mostly kvm_host_pmu;
32 
33 /* KVM's PMU capabilities, i.e. the intersection of KVM and hardware support. */
34 struct x86_pmu_capability __read_mostly kvm_pmu_cap;
35 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_pmu_cap);
36 
37 /* Enable/disable PMU virtualization */
38 bool __read_mostly enable_pmu = true;
39 EXPORT_SYMBOL_FOR_KVM_INTERNAL(enable_pmu);
40 module_param(enable_pmu, bool, 0444);
41 
42 /* Enable/disabled mediated PMU virtualization. */
43 bool __read_mostly enable_mediated_pmu;
44 EXPORT_SYMBOL_FOR_KVM_INTERNAL(enable_mediated_pmu);
45 
46 struct kvm_pmu_emulated_event_selectors {
47 	u64 INSTRUCTIONS_RETIRED;
48 	u64 BRANCH_INSTRUCTIONS_RETIRED;
49 };
50 static struct kvm_pmu_emulated_event_selectors __read_mostly kvm_pmu_eventsel;
51 
52 /* Precise Distribution of Instructions Retired (PDIR) */
53 static const struct x86_cpu_id vmx_pebs_pdir_cpu[] = {
54 	X86_MATCH_VFM(INTEL_ICELAKE_D, NULL),
55 	X86_MATCH_VFM(INTEL_ICELAKE_X, NULL),
56 	/* Instruction-Accurate PDIR (PDIR++) */
57 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, NULL),
58 	{}
59 };
60 
61 /* Precise Distribution (PDist) */
62 static const struct x86_cpu_id vmx_pebs_pdist_cpu[] = {
63 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, NULL),
64 	{}
65 };
66 
67 /* NOTE:
68  * - Each perf counter is defined as "struct kvm_pmc";
69  * - There are two types of perf counters: general purpose (gp) and fixed.
70  *   gp counters are stored in gp_counters[] and fixed counters are stored
71  *   in fixed_counters[] respectively. Both of them are part of "struct
72  *   kvm_pmu";
73  * - pmu.c understands the difference between gp counters and fixed counters.
74  *   However AMD doesn't support fixed-counters;
75  * - There are three types of index to access perf counters (PMC):
76  *     1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
77  *        has MSR_K7_PERFCTRn and, for families 15H and later,
78  *        MSR_F15H_PERF_CTRn, where MSR_F15H_PERF_CTR[0-3] are
79  *        aliased to MSR_K7_PERFCTRn.
80  *     2. MSR Index (named idx): This normally is used by RDPMC instruction.
81  *        For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
82  *        C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
83  *        that it also supports fixed counters. idx can be used to as index to
84  *        gp and fixed counters.
85  *     3. Global PMC Index (named pmc): pmc is an index specific to PMU
86  *        code. Each pmc, stored in kvm_pmc.idx field, is unique across
87  *        all perf counters (both gp and fixed). The mapping relationship
88  *        between pmc and perf counters is as the following:
89  *        * Intel: [0 .. KVM_MAX_NR_INTEL_GP_COUNTERS-1] <=> gp counters
90  *                 [KVM_FIXED_PMC_BASE_IDX .. KVM_FIXED_PMC_BASE_IDX + 2] <=> fixed
91  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] and, for families 15H
92  *          and later, [0 .. AMD64_NUM_COUNTERS_CORE-1] <=> gp counters
93  */
94 
95 static struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
96 
97 #define KVM_X86_PMU_OP(func)					     \
98 	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,			     \
99 				*(((struct kvm_pmu_ops *)0)->func));
100 #define KVM_X86_PMU_OP_OPTIONAL KVM_X86_PMU_OP
101 #define KVM_X86_PMU_OP_OPTIONAL_RET0 KVM_X86_PMU_OP
102 #include <asm/kvm-x86-pmu-ops.h>
103 EXPORT_STATIC_CALL_GPL(kvm_x86_pmu_pmc_is_disabled_in_current_mode);
104 
105 void kvm_pmu_ops_update(const struct kvm_pmu_ops *pmu_ops)
106 {
107 	memcpy(&kvm_pmu_ops, pmu_ops, sizeof(kvm_pmu_ops));
108 
109 #define __KVM_X86_PMU_OP(func) \
110 	static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func);
111 #define KVM_X86_PMU_OP(func) \
112 	WARN_ON(!kvm_pmu_ops.func); __KVM_X86_PMU_OP(func)
113 #define KVM_X86_PMU_OP_OPTIONAL __KVM_X86_PMU_OP
114 #define KVM_X86_PMU_OP_OPTIONAL_RET0(func) \
115 	static_call_update(kvm_x86_pmu_##func, (void *)kvm_pmu_ops.func ? : \
116 					       (void *)__static_call_return0);
117 #include <asm/kvm-x86-pmu-ops.h>
118 #undef __KVM_X86_PMU_OP
119 }
120 
121 void kvm_init_pmu_capability(struct kvm_pmu_ops *pmu_ops)
122 {
123 	bool is_intel = boot_cpu_data.x86_vendor == X86_VENDOR_INTEL;
124 	int min_nr_gp_ctrs = pmu_ops->MIN_NR_GP_COUNTERS;
125 
126 	/*
127 	 * Hybrid PMUs don't play nice with virtualization without careful
128 	 * configuration by userspace, and KVM's APIs for reporting supported
129 	 * vPMU features do not account for hybrid PMUs.  Disable vPMU support
130 	 * for hybrid PMUs until KVM gains a way to let userspace opt-in.
131 	 */
132 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) {
133 		enable_pmu = false;
134 		memset(&kvm_host_pmu, 0, sizeof(kvm_host_pmu));
135 	} else {
136 		perf_get_x86_pmu_capability(&kvm_host_pmu);
137 	}
138 
139 	if (enable_pmu) {
140 		/*
141 		 * WARN if perf did NOT disable hardware PMU if the number of
142 		 * architecturally required GP counters aren't present, i.e. if
143 		 * there are a non-zero number of counters, but fewer than what
144 		 * is architecturally required.
145 		 */
146 		if (!kvm_host_pmu.num_counters_gp ||
147 		    WARN_ON_ONCE(kvm_host_pmu.num_counters_gp < min_nr_gp_ctrs))
148 			enable_pmu = false;
149 		else if (is_intel && !kvm_host_pmu.version)
150 			enable_pmu = false;
151 	}
152 
153 	if (!enable_pmu || !enable_mediated_pmu || !kvm_host_pmu.mediated ||
154 	    !pmu_ops->is_mediated_pmu_supported(&kvm_host_pmu))
155 		enable_mediated_pmu = false;
156 
157 	if (!enable_mediated_pmu)
158 		pmu_ops->write_global_ctrl = NULL;
159 
160 	if (!enable_pmu) {
161 		memset(&kvm_pmu_cap, 0, sizeof(kvm_pmu_cap));
162 		return;
163 	}
164 
165 	memcpy(&kvm_pmu_cap, &kvm_host_pmu, sizeof(kvm_host_pmu));
166 	kvm_pmu_cap.version = min(kvm_pmu_cap.version, 2);
167 	kvm_pmu_cap.num_counters_gp = min(kvm_pmu_cap.num_counters_gp,
168 					  pmu_ops->MAX_NR_GP_COUNTERS);
169 	kvm_pmu_cap.num_counters_fixed = min(kvm_pmu_cap.num_counters_fixed,
170 					     KVM_MAX_NR_FIXED_COUNTERS);
171 
172 	kvm_pmu_eventsel.INSTRUCTIONS_RETIRED =
173 		perf_get_hw_event_config(PERF_COUNT_HW_INSTRUCTIONS);
174 	kvm_pmu_eventsel.BRANCH_INSTRUCTIONS_RETIRED =
175 		perf_get_hw_event_config(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
176 }
177 
178 void kvm_handle_guest_mediated_pmi(void)
179 {
180 	struct kvm_vcpu *vcpu = kvm_get_running_vcpu();
181 
182 	if (WARN_ON_ONCE(!vcpu || !kvm_vcpu_has_mediated_pmu(vcpu)))
183 		return;
184 
185 	kvm_make_request(KVM_REQ_PMI, vcpu);
186 }
187 
188 static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
189 {
190 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
191 	bool skip_pmi = false;
192 
193 	if (pmc->perf_event && pmc->perf_event->attr.precise_ip) {
194 		if (!in_pmi) {
195 			/*
196 			 * TODO: KVM is currently _choosing_ to not generate records
197 			 * for emulated instructions, avoiding BUFFER_OVF PMI when
198 			 * there are no records. Strictly speaking, it should be done
199 			 * as well in the right context to improve sampling accuracy.
200 			 */
201 			skip_pmi = true;
202 		} else {
203 			/* Indicate PEBS overflow PMI to guest. */
204 			skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
205 						      (unsigned long *)&pmu->global_status);
206 		}
207 	} else {
208 		__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
209 	}
210 
211 	if (pmc->intr && !skip_pmi)
212 		kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
213 }
214 
215 static void kvm_perf_overflow(struct perf_event *perf_event,
216 			      struct perf_sample_data *data,
217 			      struct pt_regs *regs)
218 {
219 	struct kvm_pmc *pmc = perf_event->overflow_handler_context;
220 
221 	/*
222 	 * Ignore asynchronous overflow events for counters that are scheduled
223 	 * to be reprogrammed, e.g. if a PMI for the previous event races with
224 	 * KVM's handling of a related guest WRMSR.
225 	 */
226 	if (test_and_set_bit(pmc->idx, pmc_to_pmu(pmc)->reprogram_pmi))
227 		return;
228 
229 	__kvm_perf_overflow(pmc, true);
230 
231 	kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
232 }
233 
234 static u64 pmc_get_pebs_precise_level(struct kvm_pmc *pmc)
235 {
236 	/*
237 	 * For some model specific pebs counters with special capabilities
238 	 * (PDIR, PDIR++, PDIST), KVM needs to raise the event precise
239 	 * level to the maximum value (currently 3, backwards compatible)
240 	 * so that the perf subsystem would assign specific hardware counter
241 	 * with that capability for vPMC.
242 	 */
243 	if ((pmc->idx == 0 && x86_match_cpu(vmx_pebs_pdist_cpu)) ||
244 	    (pmc->idx == 32 && x86_match_cpu(vmx_pebs_pdir_cpu)))
245 		return 3;
246 
247 	/*
248 	 * The non-zero precision level of guest event makes the ordinary
249 	 * guest event becomes a guest PEBS event and triggers the host
250 	 * PEBS PMI handler to determine whether the PEBS overflow PMI
251 	 * comes from the host counters or the guest.
252 	 */
253 	return 1;
254 }
255 
256 static u64 get_sample_period(struct kvm_pmc *pmc, u64 counter_value)
257 {
258 	u64 sample_period = (-counter_value) & pmc_bitmask(pmc);
259 
260 	if (!sample_period)
261 		sample_period = pmc_bitmask(pmc) + 1;
262 	return sample_period;
263 }
264 
265 static int pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type, u64 config,
266 				 bool exclude_user, bool exclude_kernel,
267 				 bool intr)
268 {
269 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
270 	struct perf_event *event;
271 	struct perf_event_attr attr = {
272 		.type = type,
273 		.size = sizeof(attr),
274 		.pinned = true,
275 		.exclude_idle = true,
276 		.exclude_host = 1,
277 		.exclude_user = exclude_user,
278 		.exclude_kernel = exclude_kernel,
279 		.config = config,
280 	};
281 	bool pebs = test_bit(pmc->idx, (unsigned long *)&pmu->pebs_enable);
282 
283 	attr.sample_period = get_sample_period(pmc, pmc->counter);
284 
285 	if ((attr.config & HSW_IN_TX_CHECKPOINTED) &&
286 	    (boot_cpu_has(X86_FEATURE_RTM) || boot_cpu_has(X86_FEATURE_HLE))) {
287 		/*
288 		 * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
289 		 * period. Just clear the sample period so at least
290 		 * allocating the counter doesn't fail.
291 		 */
292 		attr.sample_period = 0;
293 	}
294 	if (pebs) {
295 		/*
296 		 * For most PEBS hardware events, the difference in the software
297 		 * precision levels of guest and host PEBS events will not affect
298 		 * the accuracy of the PEBS profiling result, because the "event IP"
299 		 * in the PEBS record is calibrated on the guest side.
300 		 */
301 		attr.precise_ip = pmc_get_pebs_precise_level(pmc);
302 	}
303 
304 	event = perf_event_create_kernel_counter(&attr, -1, current,
305 						 kvm_perf_overflow, pmc);
306 	if (IS_ERR(event)) {
307 		pr_debug_ratelimited("kvm_pmu: event creation failed %ld for pmc->idx = %d\n",
308 			    PTR_ERR(event), pmc->idx);
309 		return PTR_ERR(event);
310 	}
311 
312 	pmc->perf_event = event;
313 	pmc_to_pmu(pmc)->event_count++;
314 	pmc->is_paused = false;
315 	pmc->intr = intr || pebs;
316 	return 0;
317 }
318 
319 static bool pmc_pause_counter(struct kvm_pmc *pmc)
320 {
321 	u64 counter = pmc->counter;
322 	u64 prev_counter;
323 
324 	/* update counter, reset event value to avoid redundant accumulation */
325 	if (pmc->perf_event && !pmc->is_paused)
326 		counter += perf_event_pause(pmc->perf_event, true);
327 
328 	/*
329 	 * Snapshot the previous counter *after* accumulating state from perf.
330 	 * If overflow already happened, hardware (via perf) is responsible for
331 	 * generating a PMI.  KVM just needs to detect overflow on emulated
332 	 * counter events that haven't yet been processed.
333 	 */
334 	prev_counter = counter & pmc_bitmask(pmc);
335 
336 	counter += pmc->emulated_counter;
337 	pmc->counter = counter & pmc_bitmask(pmc);
338 
339 	pmc->emulated_counter = 0;
340 	pmc->is_paused = true;
341 
342 	return pmc->counter < prev_counter;
343 }
344 
345 static bool pmc_resume_counter(struct kvm_pmc *pmc)
346 {
347 	if (!pmc->perf_event)
348 		return false;
349 
350 	/* recalibrate sample period and check if it's accepted by perf core */
351 	if (is_sampling_event(pmc->perf_event) &&
352 	    perf_event_period(pmc->perf_event,
353 			      get_sample_period(pmc, pmc->counter)))
354 		return false;
355 
356 	if (test_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->pebs_enable) !=
357 	    (!!pmc->perf_event->attr.precise_ip))
358 		return false;
359 
360 	/* reuse perf_event to serve as pmc_reprogram_counter() does*/
361 	perf_event_enable(pmc->perf_event);
362 	pmc->is_paused = false;
363 
364 	return true;
365 }
366 
367 static void pmc_release_perf_event(struct kvm_pmc *pmc)
368 {
369 	if (pmc->perf_event) {
370 		perf_event_release_kernel(pmc->perf_event);
371 		pmc->perf_event = NULL;
372 		pmc->current_config = 0;
373 		pmc_to_pmu(pmc)->event_count--;
374 	}
375 }
376 
377 static void pmc_stop_counter(struct kvm_pmc *pmc)
378 {
379 	if (pmc->perf_event) {
380 		pmc->counter = pmc_read_counter(pmc);
381 		pmc_release_perf_event(pmc);
382 	}
383 }
384 
385 static void pmc_update_sample_period(struct kvm_pmc *pmc)
386 {
387 	if (!pmc->perf_event || pmc->is_paused ||
388 	    !is_sampling_event(pmc->perf_event))
389 		return;
390 
391 	perf_event_period(pmc->perf_event,
392 			  get_sample_period(pmc, pmc->counter));
393 }
394 
395 void pmc_write_counter(struct kvm_pmc *pmc, u64 val)
396 {
397 	if (kvm_vcpu_has_mediated_pmu(pmc->vcpu)) {
398 		pmc->counter = val & pmc_bitmask(pmc);
399 		return;
400 	}
401 
402 	/*
403 	 * Drop any unconsumed accumulated counts, the WRMSR is a write, not a
404 	 * read-modify-write.  Adjust the counter value so that its value is
405 	 * relative to the current count, as reading the current count from
406 	 * perf is faster than pausing and repgrogramming the event in order to
407 	 * reset it to '0'.  Note, this very sneakily offsets the accumulated
408 	 * emulated count too, by using pmc_read_counter()!
409 	 */
410 	pmc->emulated_counter = 0;
411 	pmc->counter += val - pmc_read_counter(pmc);
412 	pmc->counter &= pmc_bitmask(pmc);
413 	pmc_update_sample_period(pmc);
414 }
415 EXPORT_SYMBOL_FOR_KVM_INTERNAL(pmc_write_counter);
416 
417 static int filter_cmp(const void *pa, const void *pb, u64 mask)
418 {
419 	u64 a = *(u64 *)pa & mask;
420 	u64 b = *(u64 *)pb & mask;
421 
422 	return (a > b) - (a < b);
423 }
424 
425 
426 static int filter_sort_cmp(const void *pa, const void *pb)
427 {
428 	return filter_cmp(pa, pb, (KVM_PMU_MASKED_ENTRY_EVENT_SELECT |
429 				   KVM_PMU_MASKED_ENTRY_EXCLUDE));
430 }
431 
432 /*
433  * For the event filter, searching is done on the 'includes' list and
434  * 'excludes' list separately rather than on the 'events' list (which
435  * has both).  As a result the exclude bit can be ignored.
436  */
437 static int filter_event_cmp(const void *pa, const void *pb)
438 {
439 	return filter_cmp(pa, pb, (KVM_PMU_MASKED_ENTRY_EVENT_SELECT));
440 }
441 
442 static int find_filter_index(u64 *events, u64 nevents, u64 key)
443 {
444 	u64 *fe = bsearch(&key, events, nevents, sizeof(events[0]),
445 			  filter_event_cmp);
446 
447 	if (!fe)
448 		return -1;
449 
450 	return fe - events;
451 }
452 
453 static bool is_filter_entry_match(u64 filter_event, u64 umask)
454 {
455 	u64 mask = filter_event >> (KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT - 8);
456 	u64 match = filter_event & KVM_PMU_MASKED_ENTRY_UMASK_MATCH;
457 
458 	BUILD_BUG_ON((KVM_PMU_ENCODE_MASKED_ENTRY(0, 0xff, 0, false) >>
459 		     (KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT - 8)) !=
460 		     ARCH_PERFMON_EVENTSEL_UMASK);
461 
462 	return (umask & mask) == match;
463 }
464 
465 static bool filter_contains_match(u64 *events, u64 nevents, u64 eventsel)
466 {
467 	u64 event_select = eventsel & kvm_pmu_ops.EVENTSEL_EVENT;
468 	u64 umask = eventsel & ARCH_PERFMON_EVENTSEL_UMASK;
469 	int i, index;
470 
471 	index = find_filter_index(events, nevents, event_select);
472 	if (index < 0)
473 		return false;
474 
475 	/*
476 	 * Entries are sorted by the event select.  Walk the list in both
477 	 * directions to process all entries with the targeted event select.
478 	 */
479 	for (i = index; i < nevents; i++) {
480 		if (filter_event_cmp(&events[i], &event_select))
481 			break;
482 
483 		if (is_filter_entry_match(events[i], umask))
484 			return true;
485 	}
486 
487 	for (i = index - 1; i >= 0; i--) {
488 		if (filter_event_cmp(&events[i], &event_select))
489 			break;
490 
491 		if (is_filter_entry_match(events[i], umask))
492 			return true;
493 	}
494 
495 	return false;
496 }
497 
498 static bool is_gp_event_allowed(struct kvm_x86_pmu_event_filter *f,
499 				u64 eventsel)
500 {
501 	if (filter_contains_match(f->includes, f->nr_includes, eventsel) &&
502 	    !filter_contains_match(f->excludes, f->nr_excludes, eventsel))
503 		return f->action == KVM_PMU_EVENT_ALLOW;
504 
505 	return f->action == KVM_PMU_EVENT_DENY;
506 }
507 
508 static bool is_fixed_event_allowed(struct kvm_x86_pmu_event_filter *filter,
509 				   int idx)
510 {
511 	int fixed_idx = idx - KVM_FIXED_PMC_BASE_IDX;
512 
513 	if (filter->action == KVM_PMU_EVENT_DENY &&
514 	    test_bit(fixed_idx, (ulong *)&filter->fixed_counter_bitmap))
515 		return false;
516 	if (filter->action == KVM_PMU_EVENT_ALLOW &&
517 	    !test_bit(fixed_idx, (ulong *)&filter->fixed_counter_bitmap))
518 		return false;
519 
520 	return true;
521 }
522 
523 static bool pmc_is_event_allowed(struct kvm_pmc *pmc)
524 {
525 	struct kvm_x86_pmu_event_filter *filter;
526 	struct kvm *kvm = pmc->vcpu->kvm;
527 
528 	filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
529 	if (!filter)
530 		return true;
531 
532 	if (pmc_is_gp(pmc))
533 		return is_gp_event_allowed(filter, pmc->eventsel);
534 
535 	return is_fixed_event_allowed(filter, pmc->idx);
536 }
537 
538 static void kvm_mediated_pmu_refresh_event_filter(struct kvm_pmc *pmc)
539 {
540 	bool allowed = pmc_is_locally_enabled(pmc) && pmc_is_event_allowed(pmc);
541 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
542 
543 	if (pmc_is_gp(pmc)) {
544 		pmc->eventsel_hw &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
545 		if (allowed)
546 			pmc->eventsel_hw |= pmc->eventsel &
547 					    ARCH_PERFMON_EVENTSEL_ENABLE;
548 	} else {
549 		u64 mask = intel_fixed_bits_by_idx(pmc->idx - KVM_FIXED_PMC_BASE_IDX, 0xf);
550 
551 		pmu->fixed_ctr_ctrl_hw &= ~mask;
552 		if (allowed)
553 			pmu->fixed_ctr_ctrl_hw |= pmu->fixed_ctr_ctrl & mask;
554 	}
555 }
556 
557 static int reprogram_counter(struct kvm_pmc *pmc)
558 {
559 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
560 	u64 eventsel = pmc->eventsel;
561 	u64 new_config = eventsel;
562 	bool emulate_overflow;
563 	u8 fixed_ctr_ctrl;
564 
565 	if (kvm_vcpu_has_mediated_pmu(pmu_to_vcpu(pmu))) {
566 		kvm_mediated_pmu_refresh_event_filter(pmc);
567 		return 0;
568 	}
569 
570 	emulate_overflow = pmc_pause_counter(pmc);
571 
572 	if (!pmc_is_globally_enabled(pmc) || !pmc_is_locally_enabled(pmc) ||
573 	    !pmc_is_event_allowed(pmc))
574 		return 0;
575 
576 	if (emulate_overflow)
577 		__kvm_perf_overflow(pmc, false);
578 
579 	if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
580 		printk_once("kvm pmu: pin control bit is ignored\n");
581 
582 	if (pmc_is_fixed(pmc)) {
583 		fixed_ctr_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl,
584 						  pmc->idx - KVM_FIXED_PMC_BASE_IDX);
585 		if (fixed_ctr_ctrl & INTEL_FIXED_0_KERNEL)
586 			eventsel |= ARCH_PERFMON_EVENTSEL_OS;
587 		if (fixed_ctr_ctrl & INTEL_FIXED_0_USER)
588 			eventsel |= ARCH_PERFMON_EVENTSEL_USR;
589 		if (fixed_ctr_ctrl & INTEL_FIXED_0_ENABLE_PMI)
590 			eventsel |= ARCH_PERFMON_EVENTSEL_INT;
591 		new_config = (u64)fixed_ctr_ctrl;
592 	}
593 
594 	if (pmc->current_config == new_config && pmc_resume_counter(pmc))
595 		return 0;
596 
597 	pmc_release_perf_event(pmc);
598 
599 	pmc->current_config = new_config;
600 
601 	return pmc_reprogram_counter(pmc, PERF_TYPE_RAW,
602 				     (eventsel & pmu->raw_event_mask),
603 				     !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
604 				     !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
605 				     eventsel & ARCH_PERFMON_EVENTSEL_INT);
606 }
607 
608 static bool pmc_is_event_match(struct kvm_pmc *pmc, u64 eventsel)
609 {
610 	/*
611 	 * Ignore checks for edge detect (all events currently emulated by KVM
612 	 * are always rising edges), pin control (unsupported by modern CPUs),
613 	 * and counter mask and its invert flag (KVM doesn't emulate multiple
614 	 * events in a single clock cycle).
615 	 *
616 	 * Note, the uppermost nibble of AMD's mask overlaps Intel's IN_TX (bit
617 	 * 32) and IN_TXCP (bit 33), as well as two reserved bits (bits 35:34).
618 	 * Checking the "in HLE/RTM transaction" flags is correct as the vCPU
619 	 * can't be in a transaction if KVM is emulating an instruction.
620 	 *
621 	 * Checking the reserved bits might be wrong if they are defined in the
622 	 * future, but so could ignoring them, so do the simple thing for now.
623 	 */
624 	return !((pmc->eventsel ^ eventsel) & AMD64_RAW_EVENT_MASK_NB);
625 }
626 
627 void kvm_pmu_recalc_pmc_emulation(struct kvm_pmu *pmu, struct kvm_pmc *pmc)
628 {
629 	bitmap_clear(pmu->pmc_counting_instructions, pmc->idx, 1);
630 	bitmap_clear(pmu->pmc_counting_branches, pmc->idx, 1);
631 
632 	/*
633 	 * Do NOT consult the PMU event filters, as the filters must be checked
634 	 * at the time of emulation to ensure KVM uses fresh information, e.g.
635 	 * omitting a PMC from a bitmap could result in a missed event if the
636 	 * filter is changed to allow counting the event.
637 	 */
638 	if (!pmc_is_locally_enabled(pmc))
639 		return;
640 
641 	if (pmc_is_event_match(pmc, kvm_pmu_eventsel.INSTRUCTIONS_RETIRED))
642 		bitmap_set(pmu->pmc_counting_instructions, pmc->idx, 1);
643 
644 	if (pmc_is_event_match(pmc, kvm_pmu_eventsel.BRANCH_INSTRUCTIONS_RETIRED))
645 		bitmap_set(pmu->pmc_counting_branches, pmc->idx, 1);
646 }
647 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_pmu_recalc_pmc_emulation);
648 
649 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
650 {
651 	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);
652 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
653 	struct kvm_pmc *pmc;
654 	int bit;
655 
656 	bitmap_copy(bitmap, pmu->reprogram_pmi, X86_PMC_IDX_MAX);
657 
658 	/*
659 	 * The reprogramming bitmap can be written asynchronously by something
660 	 * other than the task that holds vcpu->mutex, take care to clear only
661 	 * the bits that will actually processed.
662 	 */
663 	BUILD_BUG_ON(sizeof(bitmap) != sizeof(atomic64_t));
664 	atomic64_andnot(*(s64 *)bitmap, &pmu->__reprogram_pmi);
665 
666 	kvm_for_each_pmc(pmu, pmc, bit, bitmap) {
667 		/*
668 		 * If reprogramming fails, e.g. due to contention, re-set the
669 		 * regprogram bit set, i.e. opportunistically try again on the
670 		 * next PMU refresh.  Don't make a new request as doing so can
671 		 * stall the guest if reprogramming repeatedly fails.
672 		 */
673 		if (reprogram_counter(pmc))
674 			set_bit(pmc->idx, pmu->reprogram_pmi);
675 	}
676 
677 	/*
678 	 * Release unused perf_events if the corresponding guest MSRs weren't
679 	 * accessed during the last vCPU time slice (need_cleanup is set when
680 	 * the vCPU is scheduled back in).
681 	 */
682 	if (unlikely(pmu->need_cleanup))
683 		kvm_pmu_cleanup(vcpu);
684 
685 	kvm_for_each_pmc(pmu, pmc, bit, bitmap)
686 		kvm_pmu_recalc_pmc_emulation(pmu, pmc);
687 }
688 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_pmu_handle_event);
689 
690 int kvm_pmu_check_rdpmc_early(struct kvm_vcpu *vcpu, unsigned int idx)
691 {
692 	/*
693 	 * On Intel, VMX interception has priority over RDPMC exceptions that
694 	 * aren't already handled by the emulator, i.e. there are no additional
695 	 * check needed for Intel PMUs.
696 	 *
697 	 * On AMD, _all_ exceptions on RDPMC have priority over SVM intercepts,
698 	 * i.e. an invalid PMC results in a #GP, not #VMEXIT.
699 	 */
700 	if (!kvm_pmu_ops.check_rdpmc_early)
701 		return 0;
702 
703 	return kvm_pmu_call(check_rdpmc_early)(vcpu, idx);
704 }
705 
706 bool is_vmware_backdoor_pmc(u32 pmc_idx)
707 {
708 	switch (pmc_idx) {
709 	case VMWARE_BACKDOOR_PMC_HOST_TSC:
710 	case VMWARE_BACKDOOR_PMC_REAL_TIME:
711 	case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
712 		return true;
713 	}
714 	return false;
715 }
716 
717 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
718 {
719 	u64 ctr_val;
720 
721 	switch (idx) {
722 	case VMWARE_BACKDOOR_PMC_HOST_TSC:
723 		ctr_val = rdtsc();
724 		break;
725 	case VMWARE_BACKDOOR_PMC_REAL_TIME:
726 		ctr_val = ktime_get_boottime_ns();
727 		break;
728 	case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
729 		ctr_val = ktime_get_boottime_ns() +
730 			vcpu->kvm->arch.kvmclock_offset;
731 		break;
732 	default:
733 		return 1;
734 	}
735 
736 	*data = ctr_val;
737 	return 0;
738 }
739 
740 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
741 {
742 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
743 	struct kvm_pmc *pmc;
744 	u64 mask = ~0ull;
745 
746 	if (!pmu->version)
747 		return 1;
748 
749 	if (is_vmware_backdoor_pmc(idx))
750 		return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
751 
752 	pmc = kvm_pmu_call(rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
753 	if (!pmc)
754 		return 1;
755 
756 	if (!kvm_is_cr4_bit_set(vcpu, X86_CR4_PCE) &&
757 	    (kvm_x86_call(get_cpl)(vcpu) != 0) &&
758 	    kvm_is_cr0_bit_set(vcpu, X86_CR0_PE))
759 		return 1;
760 
761 	*data = pmc_read_counter(pmc) & mask;
762 	return 0;
763 }
764 
765 static bool kvm_need_any_pmc_intercept(struct kvm_vcpu *vcpu)
766 {
767 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
768 
769 	if (!kvm_vcpu_has_mediated_pmu(vcpu))
770 		return true;
771 
772 	/*
773 	 * Note!  Check *host* PMU capabilities, not KVM's PMU capabilities, as
774 	 * KVM's capabilities are constrained based on KVM support, i.e. KVM's
775 	 * capabilities themselves may be a subset of hardware capabilities.
776 	 */
777 	return pmu->nr_arch_gp_counters != kvm_host_pmu.num_counters_gp ||
778 	       pmu->nr_arch_fixed_counters != kvm_host_pmu.num_counters_fixed;
779 }
780 
781 bool kvm_need_perf_global_ctrl_intercept(struct kvm_vcpu *vcpu)
782 {
783 	return kvm_need_any_pmc_intercept(vcpu) ||
784 	       !kvm_pmu_has_perf_global_ctrl(vcpu_to_pmu(vcpu));
785 }
786 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_need_perf_global_ctrl_intercept);
787 
788 bool kvm_need_rdpmc_intercept(struct kvm_vcpu *vcpu)
789 {
790 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
791 
792 	/*
793 	 * VMware allows access to these Pseduo-PMCs even when read via RDPMC
794 	 * in Ring3 when CR4.PCE=0.
795 	 */
796 	if (enable_vmware_backdoor)
797 		return true;
798 
799 	return kvm_need_any_pmc_intercept(vcpu) ||
800 	       pmu->counter_bitmask[KVM_PMC_GP] != (BIT_ULL(kvm_host_pmu.bit_width_gp) - 1) ||
801 	       pmu->counter_bitmask[KVM_PMC_FIXED] != (BIT_ULL(kvm_host_pmu.bit_width_fixed) - 1);
802 }
803 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_need_rdpmc_intercept);
804 
805 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
806 {
807 	if (lapic_in_kernel(vcpu)) {
808 		kvm_pmu_call(deliver_pmi)(vcpu);
809 		kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
810 	}
811 }
812 
813 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
814 {
815 	switch (msr) {
816 	case MSR_CORE_PERF_GLOBAL_STATUS:
817 	case MSR_CORE_PERF_GLOBAL_CTRL:
818 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
819 		return kvm_pmu_has_perf_global_ctrl(vcpu_to_pmu(vcpu));
820 	default:
821 		break;
822 	}
823 	return kvm_pmu_call(msr_idx_to_pmc)(vcpu, msr) ||
824 	       kvm_pmu_call(is_valid_msr)(vcpu, msr);
825 }
826 
827 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
828 {
829 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
830 	struct kvm_pmc *pmc = kvm_pmu_call(msr_idx_to_pmc)(vcpu, msr);
831 
832 	if (pmc)
833 		__set_bit(pmc->idx, pmu->pmc_in_use);
834 }
835 
836 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
837 {
838 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
839 	u32 msr = msr_info->index;
840 
841 	switch (msr) {
842 	case MSR_CORE_PERF_GLOBAL_STATUS:
843 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS:
844 		msr_info->data = pmu->global_status;
845 		break;
846 	case MSR_AMD64_PERF_CNTR_GLOBAL_CTL:
847 	case MSR_CORE_PERF_GLOBAL_CTRL:
848 		msr_info->data = pmu->global_ctrl;
849 		break;
850 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR:
851 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_SET:
852 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
853 		msr_info->data = 0;
854 		break;
855 	default:
856 		return kvm_pmu_call(get_msr)(vcpu, msr_info);
857 	}
858 
859 	return 0;
860 }
861 
862 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
863 {
864 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
865 	u32 msr = msr_info->index;
866 	u64 data = msr_info->data;
867 	u64 diff;
868 
869 	/*
870 	 * Note, AMD ignores writes to reserved bits and read-only PMU MSRs,
871 	 * whereas Intel generates #GP on attempts to write reserved/RO MSRs.
872 	 */
873 	switch (msr) {
874 	case MSR_CORE_PERF_GLOBAL_STATUS:
875 		if (!msr_info->host_initiated)
876 			return 1; /* RO MSR */
877 		fallthrough;
878 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS:
879 		/* Per PPR, Read-only MSR. Writes are ignored. */
880 		if (!msr_info->host_initiated)
881 			break;
882 
883 		if (data & pmu->global_status_rsvd)
884 			return 1;
885 
886 		pmu->global_status = data;
887 		break;
888 	case MSR_AMD64_PERF_CNTR_GLOBAL_CTL:
889 		data &= ~pmu->global_ctrl_rsvd;
890 		fallthrough;
891 	case MSR_CORE_PERF_GLOBAL_CTRL:
892 		if (!kvm_valid_perf_global_ctrl(pmu, data))
893 			return 1;
894 
895 		if (pmu->global_ctrl != data) {
896 			diff = pmu->global_ctrl ^ data;
897 			pmu->global_ctrl = data;
898 			kvm_pmu_request_counters_reprogram(pmu, diff);
899 		}
900 		/*
901 		 * Unconditionally forward writes to vendor code, i.e. to the
902 		 * VMC{B,S}, as pmu->global_ctrl is per-VCPU, not per-VMC{B,S}.
903 		 */
904 		if (kvm_vcpu_has_mediated_pmu(vcpu))
905 			kvm_pmu_call(write_global_ctrl)(data);
906 		break;
907 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
908 		/*
909 		 * GLOBAL_OVF_CTRL, a.k.a. GLOBAL STATUS_RESET, clears bits in
910 		 * GLOBAL_STATUS, and so the set of reserved bits is the same.
911 		 */
912 		if (data & pmu->global_status_rsvd)
913 			return 1;
914 		fallthrough;
915 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR:
916 		if (!msr_info->host_initiated)
917 			pmu->global_status &= ~data;
918 		break;
919 	case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_SET:
920 		if (!msr_info->host_initiated)
921 			pmu->global_status |= data & ~pmu->global_status_rsvd;
922 		break;
923 	default:
924 		kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
925 		return kvm_pmu_call(set_msr)(vcpu, msr_info);
926 	}
927 
928 	return 0;
929 }
930 
931 static void kvm_pmu_reset(struct kvm_vcpu *vcpu)
932 {
933 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
934 	struct kvm_pmc *pmc;
935 	int i;
936 
937 	pmu->need_cleanup = false;
938 
939 	bitmap_zero(pmu->reprogram_pmi, X86_PMC_IDX_MAX);
940 	bitmap_zero(pmu->pmc_has_mode_specific_enables, X86_PMC_IDX_MAX);
941 
942 	kvm_for_each_pmc(pmu, pmc, i, pmu->all_valid_pmc_idx) {
943 		pmc_stop_counter(pmc);
944 		pmc->counter = 0;
945 		pmc->emulated_counter = 0;
946 
947 		if (pmc_is_gp(pmc)) {
948 			pmc->eventsel = 0;
949 			pmc->eventsel_hw = 0;
950 		}
951 	}
952 
953 	pmu->fixed_ctr_ctrl = pmu->fixed_ctr_ctrl_hw = 0;
954 	pmu->global_ctrl = pmu->global_status = 0;
955 
956 	kvm_pmu_call(reset)(vcpu);
957 }
958 
959 
960 /*
961  * Refresh the PMU configuration for the vCPU, e.g. if userspace changes CPUID
962  * and/or PERF_CAPABILITIES.
963  */
964 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
965 {
966 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
967 
968 	if (KVM_BUG_ON(!kvm_can_set_cpuid_and_feature_msrs(vcpu), vcpu->kvm))
969 		return;
970 
971 	/*
972 	 * Stop/release all existing counters/events before realizing the new
973 	 * vPMU model.
974 	 */
975 	kvm_pmu_reset(vcpu);
976 
977 	pmu->version = 0;
978 	pmu->nr_arch_gp_counters = 0;
979 	pmu->nr_arch_fixed_counters = 0;
980 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
981 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
982 	pmu->reserved_bits = 0xffffffff00200000ull;
983 	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
984 	pmu->global_ctrl_rsvd = ~0ull;
985 	pmu->global_status_rsvd = ~0ull;
986 	pmu->fixed_ctr_ctrl_rsvd = ~0ull;
987 	pmu->pebs_enable_rsvd = ~0ull;
988 	pmu->pebs_data_cfg_rsvd = ~0ull;
989 	bitmap_zero(pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX);
990 
991 	if (!vcpu->kvm->arch.enable_pmu)
992 		return;
993 
994 	kvm_pmu_call(refresh)(vcpu);
995 
996 	/*
997 	 * At RESET, both Intel and AMD CPUs set all enable bits for general
998 	 * purpose counters in IA32_PERF_GLOBAL_CTRL (so that software that
999 	 * was written for v1 PMUs don't unknowingly leave GP counters disabled
1000 	 * in the global controls).  Emulate that behavior when refreshing the
1001 	 * PMU so that userspace doesn't need to manually set PERF_GLOBAL_CTRL.
1002 	 */
1003 	if (pmu->nr_arch_gp_counters &&
1004 	    (kvm_pmu_has_perf_global_ctrl(pmu) || kvm_vcpu_has_mediated_pmu(vcpu)))
1005 		pmu->global_ctrl = GENMASK_ULL(pmu->nr_arch_gp_counters - 1, 0);
1006 
1007 	if (kvm_vcpu_has_mediated_pmu(vcpu))
1008 		kvm_pmu_call(write_global_ctrl)(pmu->global_ctrl);
1009 
1010 	bitmap_set(pmu->all_valid_pmc_idx, 0, pmu->nr_arch_gp_counters);
1011 	bitmap_set(pmu->all_valid_pmc_idx, KVM_FIXED_PMC_BASE_IDX,
1012 		   pmu->nr_arch_fixed_counters);
1013 }
1014 
1015 void kvm_pmu_init(struct kvm_vcpu *vcpu)
1016 {
1017 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
1018 
1019 	memset(pmu, 0, sizeof(*pmu));
1020 	kvm_pmu_call(init)(vcpu);
1021 }
1022 
1023 /* Release perf_events for vPMCs that have been unused for a full time slice.  */
1024 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
1025 {
1026 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
1027 	struct kvm_pmc *pmc = NULL;
1028 	DECLARE_BITMAP(bitmask, X86_PMC_IDX_MAX);
1029 	int i;
1030 
1031 	pmu->need_cleanup = false;
1032 
1033 	bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
1034 		      pmu->pmc_in_use, X86_PMC_IDX_MAX);
1035 
1036 	kvm_for_each_pmc(pmu, pmc, i, bitmask) {
1037 		if (pmc->perf_event && !pmc_is_locally_enabled(pmc))
1038 			pmc_stop_counter(pmc);
1039 	}
1040 
1041 	kvm_pmu_call(cleanup)(vcpu);
1042 
1043 	bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
1044 }
1045 
1046 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
1047 {
1048 	kvm_pmu_reset(vcpu);
1049 }
1050 
1051 static bool pmc_is_pmi_enabled(struct kvm_pmc *pmc)
1052 {
1053 	u8 fixed_ctr_ctrl;
1054 
1055 	if (pmc_is_gp(pmc))
1056 		return pmc->eventsel & ARCH_PERFMON_EVENTSEL_INT;
1057 
1058 	fixed_ctr_ctrl = fixed_ctrl_field(pmc_to_pmu(pmc)->fixed_ctr_ctrl,
1059 					  pmc->idx - KVM_FIXED_PMC_BASE_IDX);
1060 	return fixed_ctr_ctrl & INTEL_FIXED_0_ENABLE_PMI;
1061 }
1062 
1063 static void kvm_pmu_incr_counter(struct kvm_pmc *pmc)
1064 {
1065 	struct kvm_vcpu *vcpu = pmc->vcpu;
1066 
1067 	/*
1068 	 * For perf-based PMUs, accumulate software-emulated events separately
1069 	 * from pmc->counter, as pmc->counter is offset by the count of the
1070 	 * associated perf event. Request reprogramming, which will consult
1071 	 * both emulated and hardware-generated events to detect overflow.
1072 	 */
1073 	if (!kvm_vcpu_has_mediated_pmu(vcpu)) {
1074 		pmc->emulated_counter++;
1075 		kvm_pmu_request_counter_reprogram(pmc);
1076 		return;
1077 	}
1078 
1079 	/*
1080 	 * For mediated PMUs, pmc->counter is updated when the vCPU's PMU is
1081 	 * put, and will be loaded into hardware when the PMU is loaded. Simply
1082 	 * increment the counter and signal overflow if it wraps to zero.
1083 	 */
1084 	pmc->counter = (pmc->counter + 1) & pmc_bitmask(pmc);
1085 	if (!pmc->counter) {
1086 		pmc_to_pmu(pmc)->global_status |= BIT_ULL(pmc->idx);
1087 		if (pmc_is_pmi_enabled(pmc))
1088 			kvm_make_request(KVM_REQ_PMI, vcpu);
1089 	}
1090 }
1091 
1092 static inline bool cpl_is_matched(struct kvm_pmc *pmc)
1093 {
1094 	bool select_os, select_user;
1095 	u64 config;
1096 
1097 	if (pmc_is_gp(pmc)) {
1098 		config = pmc->eventsel;
1099 		select_os = config & ARCH_PERFMON_EVENTSEL_OS;
1100 		select_user = config & ARCH_PERFMON_EVENTSEL_USR;
1101 	} else {
1102 		config = fixed_ctrl_field(pmc_to_pmu(pmc)->fixed_ctr_ctrl,
1103 					  pmc->idx - KVM_FIXED_PMC_BASE_IDX);
1104 		select_os = config & INTEL_FIXED_0_KERNEL;
1105 		select_user = config & INTEL_FIXED_0_USER;
1106 	}
1107 
1108 	/*
1109 	 * Skip the CPL lookup, which isn't free on Intel, if the result will
1110 	 * be the same regardless of the CPL.
1111 	 */
1112 	if (select_os == select_user)
1113 		return select_os;
1114 
1115 	return (kvm_x86_call(get_cpl)(pmc->vcpu) == 0) ? select_os :
1116 							 select_user;
1117 }
1118 
1119 static void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu,
1120 				  const unsigned long *event_pmcs)
1121 {
1122 	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);
1123 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
1124 	struct kvm_pmc *pmc;
1125 	int i, idx;
1126 
1127 	BUILD_BUG_ON(sizeof(pmu->global_ctrl) * BITS_PER_BYTE != X86_PMC_IDX_MAX);
1128 
1129 	if (bitmap_empty(event_pmcs, X86_PMC_IDX_MAX))
1130 		return;
1131 
1132 	if (!kvm_pmu_has_perf_global_ctrl(pmu))
1133 		bitmap_copy(bitmap, event_pmcs, X86_PMC_IDX_MAX);
1134 	else if (!bitmap_and(bitmap, event_pmcs,
1135 			     (unsigned long *)&pmu->global_ctrl, X86_PMC_IDX_MAX))
1136 		return;
1137 
1138 	idx = srcu_read_lock(&vcpu->kvm->srcu);
1139 	kvm_for_each_pmc(pmu, pmc, i, bitmap) {
1140 		if (!pmc_is_event_allowed(pmc) || !cpl_is_matched(pmc))
1141 			continue;
1142 
1143 		kvm_pmu_incr_counter(pmc);
1144 	}
1145 	srcu_read_unlock(&vcpu->kvm->srcu, idx);
1146 }
1147 
1148 void kvm_pmu_instruction_retired(struct kvm_vcpu *vcpu)
1149 {
1150 	kvm_pmu_trigger_event(vcpu, vcpu_to_pmu(vcpu)->pmc_counting_instructions);
1151 }
1152 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_pmu_instruction_retired);
1153 
1154 void kvm_pmu_branch_retired(struct kvm_vcpu *vcpu)
1155 {
1156 	kvm_pmu_trigger_event(vcpu, vcpu_to_pmu(vcpu)->pmc_counting_branches);
1157 }
1158 EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_pmu_branch_retired);
1159 
1160 static bool is_masked_filter_valid(const struct kvm_x86_pmu_event_filter *filter)
1161 {
1162 	u64 mask = kvm_pmu_ops.EVENTSEL_EVENT |
1163 		   KVM_PMU_MASKED_ENTRY_UMASK_MASK |
1164 		   KVM_PMU_MASKED_ENTRY_UMASK_MATCH |
1165 		   KVM_PMU_MASKED_ENTRY_EXCLUDE;
1166 	int i;
1167 
1168 	for (i = 0; i < filter->nevents; i++) {
1169 		if (filter->events[i] & ~mask)
1170 			return false;
1171 	}
1172 
1173 	return true;
1174 }
1175 
1176 static void convert_to_masked_filter(struct kvm_x86_pmu_event_filter *filter)
1177 {
1178 	int i, j;
1179 
1180 	for (i = 0, j = 0; i < filter->nevents; i++) {
1181 		/*
1182 		 * Skip events that are impossible to match against a guest
1183 		 * event.  When filtering, only the event select + unit mask
1184 		 * of the guest event is used.  To maintain backwards
1185 		 * compatibility, impossible filters can't be rejected :-(
1186 		 */
1187 		if (filter->events[i] & ~(kvm_pmu_ops.EVENTSEL_EVENT |
1188 					  ARCH_PERFMON_EVENTSEL_UMASK))
1189 			continue;
1190 		/*
1191 		 * Convert userspace events to a common in-kernel event so
1192 		 * only one code path is needed to support both events.  For
1193 		 * the in-kernel events use masked events because they are
1194 		 * flexible enough to handle both cases.  To convert to masked
1195 		 * events all that's needed is to add an "all ones" umask_mask,
1196 		 * (unmasked filter events don't support EXCLUDE).
1197 		 */
1198 		filter->events[j++] = filter->events[i] |
1199 				      (0xFFULL << KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT);
1200 	}
1201 
1202 	filter->nevents = j;
1203 }
1204 
1205 static int prepare_filter_lists(struct kvm_x86_pmu_event_filter *filter)
1206 {
1207 	int i;
1208 
1209 	if (!(filter->flags & KVM_PMU_EVENT_FLAG_MASKED_EVENTS))
1210 		convert_to_masked_filter(filter);
1211 	else if (!is_masked_filter_valid(filter))
1212 		return -EINVAL;
1213 
1214 	/*
1215 	 * Sort entries by event select and includes vs. excludes so that all
1216 	 * entries for a given event select can be processed efficiently during
1217 	 * filtering.  The EXCLUDE flag uses a more significant bit than the
1218 	 * event select, and so the sorted list is also effectively split into
1219 	 * includes and excludes sub-lists.
1220 	 */
1221 	sort(&filter->events, filter->nevents, sizeof(filter->events[0]),
1222 	     filter_sort_cmp, NULL);
1223 
1224 	i = filter->nevents;
1225 	/* Find the first EXCLUDE event (only supported for masked events). */
1226 	if (filter->flags & KVM_PMU_EVENT_FLAG_MASKED_EVENTS) {
1227 		for (i = 0; i < filter->nevents; i++) {
1228 			if (filter->events[i] & KVM_PMU_MASKED_ENTRY_EXCLUDE)
1229 				break;
1230 		}
1231 	}
1232 
1233 	filter->nr_includes = i;
1234 	filter->nr_excludes = filter->nevents - filter->nr_includes;
1235 	filter->includes = filter->events;
1236 	filter->excludes = filter->events + filter->nr_includes;
1237 
1238 	return 0;
1239 }
1240 
1241 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
1242 {
1243 	struct kvm_pmu_event_filter __user *user_filter = argp;
1244 	struct kvm_x86_pmu_event_filter *filter;
1245 	struct kvm_pmu_event_filter tmp;
1246 	struct kvm_vcpu *vcpu;
1247 	unsigned long i;
1248 	size_t size;
1249 	int r;
1250 
1251 	if (copy_from_user(&tmp, user_filter, sizeof(tmp)))
1252 		return -EFAULT;
1253 
1254 	if (tmp.action != KVM_PMU_EVENT_ALLOW &&
1255 	    tmp.action != KVM_PMU_EVENT_DENY)
1256 		return -EINVAL;
1257 
1258 	if (tmp.flags & ~KVM_PMU_EVENT_FLAGS_VALID_MASK)
1259 		return -EINVAL;
1260 
1261 	if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
1262 		return -E2BIG;
1263 
1264 	size = struct_size(filter, events, tmp.nevents);
1265 	filter = kzalloc(size, GFP_KERNEL_ACCOUNT);
1266 	if (!filter)
1267 		return -ENOMEM;
1268 
1269 	filter->action = tmp.action;
1270 	filter->nevents = tmp.nevents;
1271 	filter->fixed_counter_bitmap = tmp.fixed_counter_bitmap;
1272 	filter->flags = tmp.flags;
1273 
1274 	r = -EFAULT;
1275 	if (copy_from_user(filter->events, user_filter->events,
1276 			   flex_array_size(filter, events, filter->nevents)))
1277 		goto cleanup;
1278 
1279 	r = prepare_filter_lists(filter);
1280 	if (r)
1281 		goto cleanup;
1282 
1283 	mutex_lock(&kvm->lock);
1284 	filter = rcu_replace_pointer(kvm->arch.pmu_event_filter, filter,
1285 				     mutex_is_locked(&kvm->lock));
1286 	mutex_unlock(&kvm->lock);
1287 	synchronize_srcu_expedited(&kvm->srcu);
1288 
1289 	BUILD_BUG_ON(sizeof(((struct kvm_pmu *)0)->reprogram_pmi) >
1290 		     sizeof(((struct kvm_pmu *)0)->__reprogram_pmi));
1291 
1292 	kvm_for_each_vcpu(i, vcpu, kvm)
1293 		atomic64_set(&vcpu_to_pmu(vcpu)->__reprogram_pmi, -1ull);
1294 
1295 	kvm_make_all_cpus_request(kvm, KVM_REQ_PMU);
1296 
1297 	r = 0;
1298 cleanup:
1299 	kfree(filter);
1300 	return r;
1301 }
1302 
1303 static __always_inline u32 fixed_counter_msr(u32 idx)
1304 {
1305 	return kvm_pmu_ops.FIXED_COUNTER_BASE + idx * kvm_pmu_ops.MSR_STRIDE;
1306 }
1307 
1308 static __always_inline u32 gp_counter_msr(u32 idx)
1309 {
1310 	return kvm_pmu_ops.GP_COUNTER_BASE + idx * kvm_pmu_ops.MSR_STRIDE;
1311 }
1312 
1313 static __always_inline u32 gp_eventsel_msr(u32 idx)
1314 {
1315 	return kvm_pmu_ops.GP_EVENTSEL_BASE + idx * kvm_pmu_ops.MSR_STRIDE;
1316 }
1317 
1318 static void kvm_pmu_load_guest_pmcs(struct kvm_vcpu *vcpu)
1319 {
1320 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
1321 	struct kvm_pmc *pmc;
1322 	u32 i;
1323 
1324 	/*
1325 	 * No need to zero out unexposed GP/fixed counters/selectors since RDPMC
1326 	 * is intercepted if hardware has counters that aren't visible to the
1327 	 * guest (KVM will inject #GP as appropriate).
1328 	 */
1329 	for (i = 0; i < pmu->nr_arch_gp_counters; i++) {
1330 		pmc = &pmu->gp_counters[i];
1331 
1332 		if (pmc->counter != rdpmc(i))
1333 			wrmsrq(gp_counter_msr(i), pmc->counter);
1334 		wrmsrq(gp_eventsel_msr(i), pmc->eventsel_hw);
1335 	}
1336 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
1337 		pmc = &pmu->fixed_counters[i];
1338 
1339 		if (pmc->counter != rdpmc(INTEL_PMC_FIXED_RDPMC_BASE | i))
1340 			wrmsrq(fixed_counter_msr(i), pmc->counter);
1341 	}
1342 }
1343 
1344 void kvm_mediated_pmu_load(struct kvm_vcpu *vcpu)
1345 {
1346 	if (!kvm_vcpu_has_mediated_pmu(vcpu) ||
1347 	    KVM_BUG_ON(!lapic_in_kernel(vcpu), vcpu->kvm))
1348 		return;
1349 
1350 	lockdep_assert_irqs_disabled();
1351 
1352 	perf_load_guest_context();
1353 
1354 	/*
1355 	 * Explicitly clear PERF_GLOBAL_CTRL, as "loading" the guest's context
1356 	 * disables all individual counters (if any were enabled), but doesn't
1357 	 * globally disable the entire PMU.  Loading event selectors and PMCs
1358 	 * with guest values while PERF_GLOBAL_CTRL is non-zero will generate
1359 	 * unexpected events and PMIs.
1360 	 *
1361 	 * VMX will enable/disable counters at VM-Enter/VM-Exit by atomically
1362 	 * loading PERF_GLOBAL_CONTROL.  SVM effectively performs the switch by
1363 	 * configuring all events to be GUEST_ONLY.  Clear PERF_GLOBAL_CONTROL
1364 	 * even for SVM to minimize the damage if a perf event is left enabled,
1365 	 * and to ensure a consistent starting state.
1366 	 */
1367 	wrmsrq(kvm_pmu_ops.PERF_GLOBAL_CTRL, 0);
1368 
1369 	perf_load_guest_lvtpc(kvm_lapic_get_reg(vcpu->arch.apic, APIC_LVTPC));
1370 
1371 	kvm_pmu_load_guest_pmcs(vcpu);
1372 
1373 	kvm_pmu_call(mediated_load)(vcpu);
1374 }
1375 
1376 static void kvm_pmu_put_guest_pmcs(struct kvm_vcpu *vcpu)
1377 {
1378 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
1379 	struct kvm_pmc *pmc;
1380 	u32 i;
1381 
1382 	/*
1383 	 * Clear selectors and counters to ensure hardware doesn't count using
1384 	 * guest controls when the host (perf) restores its state.
1385 	 */
1386 	for (i = 0; i < pmu->nr_arch_gp_counters; i++) {
1387 		pmc = &pmu->gp_counters[i];
1388 
1389 		pmc->counter = rdpmc(i);
1390 		if (pmc->counter)
1391 			wrmsrq(gp_counter_msr(i), 0);
1392 		if (pmc->eventsel_hw)
1393 			wrmsrq(gp_eventsel_msr(i), 0);
1394 	}
1395 
1396 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
1397 		pmc = &pmu->fixed_counters[i];
1398 
1399 		pmc->counter = rdpmc(INTEL_PMC_FIXED_RDPMC_BASE | i);
1400 		if (pmc->counter)
1401 			wrmsrq(fixed_counter_msr(i), 0);
1402 	}
1403 }
1404 
1405 void kvm_mediated_pmu_put(struct kvm_vcpu *vcpu)
1406 {
1407 	if (!kvm_vcpu_has_mediated_pmu(vcpu) ||
1408 	    KVM_BUG_ON(!lapic_in_kernel(vcpu), vcpu->kvm))
1409 		return;
1410 
1411 	lockdep_assert_irqs_disabled();
1412 
1413 	/*
1414 	 * Defer handling of PERF_GLOBAL_CTRL to vendor code.  On Intel, it's
1415 	 * atomically cleared on VM-Exit, i.e. doesn't need to be clear here.
1416 	 */
1417 	kvm_pmu_call(mediated_put)(vcpu);
1418 
1419 	kvm_pmu_put_guest_pmcs(vcpu);
1420 
1421 	perf_put_guest_lvtpc();
1422 
1423 	perf_put_guest_context();
1424 }
1425