xref: /linux/arch/x86/kernel/kvm.c (revision bc75dffadc063eb46200611cc41d1e2373219e11)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * KVM paravirt_ops implementation
4  *
5  * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6  * Copyright IBM Corporation, 2007
7  *   Authors: Anthony Liguori <aliguori@us.ibm.com>
8  */
9 
10 #define pr_fmt(fmt) "kvm-guest: " fmt
11 
12 #include <linux/context_tracking.h>
13 #include <linux/init.h>
14 #include <linux/irq.h>
15 #include <linux/kernel.h>
16 #include <linux/kvm_para.h>
17 #include <linux/cpu.h>
18 #include <linux/mm.h>
19 #include <linux/highmem.h>
20 #include <linux/hardirq.h>
21 #include <linux/notifier.h>
22 #include <linux/reboot.h>
23 #include <linux/hash.h>
24 #include <linux/sched.h>
25 #include <linux/slab.h>
26 #include <linux/kprobes.h>
27 #include <linux/nmi.h>
28 #include <linux/swait.h>
29 #include <linux/syscore_ops.h>
30 #include <linux/cc_platform.h>
31 #include <linux/efi.h>
32 #include <asm/timer.h>
33 #include <asm/cpu.h>
34 #include <asm/traps.h>
35 #include <asm/desc.h>
36 #include <asm/tlbflush.h>
37 #include <asm/apic.h>
38 #include <asm/apicdef.h>
39 #include <asm/hypervisor.h>
40 #include <asm/tlb.h>
41 #include <asm/cpuidle_haltpoll.h>
42 #include <asm/ptrace.h>
43 #include <asm/reboot.h>
44 #include <asm/svm.h>
45 #include <asm/e820/api.h>
46 
47 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
48 
49 static int kvmapf = 1;
50 
51 static int __init parse_no_kvmapf(char *arg)
52 {
53         kvmapf = 0;
54         return 0;
55 }
56 
57 early_param("no-kvmapf", parse_no_kvmapf);
58 
59 static int steal_acc = 1;
60 static int __init parse_no_stealacc(char *arg)
61 {
62         steal_acc = 0;
63         return 0;
64 }
65 
66 early_param("no-steal-acc", parse_no_stealacc);
67 
68 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
69 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
70 static int has_steal_clock = 0;
71 
72 static int has_guest_poll = 0;
73 /*
74  * No need for any "IO delay" on KVM
75  */
76 static void kvm_io_delay(void)
77 {
78 }
79 
80 #define KVM_TASK_SLEEP_HASHBITS 8
81 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
82 
83 struct kvm_task_sleep_node {
84 	struct hlist_node link;
85 	struct swait_queue_head wq;
86 	u32 token;
87 	int cpu;
88 };
89 
90 static struct kvm_task_sleep_head {
91 	raw_spinlock_t lock;
92 	struct hlist_head list;
93 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
94 
95 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
96 						  u32 token)
97 {
98 	struct hlist_node *p;
99 
100 	hlist_for_each(p, &b->list) {
101 		struct kvm_task_sleep_node *n =
102 			hlist_entry(p, typeof(*n), link);
103 		if (n->token == token)
104 			return n;
105 	}
106 
107 	return NULL;
108 }
109 
110 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
111 {
112 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
113 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
114 	struct kvm_task_sleep_node *e;
115 
116 	raw_spin_lock(&b->lock);
117 	e = _find_apf_task(b, token);
118 	if (e) {
119 		/* dummy entry exist -> wake up was delivered ahead of PF */
120 		hlist_del(&e->link);
121 		raw_spin_unlock(&b->lock);
122 		kfree(e);
123 		return false;
124 	}
125 
126 	n->token = token;
127 	n->cpu = smp_processor_id();
128 	init_swait_queue_head(&n->wq);
129 	hlist_add_head(&n->link, &b->list);
130 	raw_spin_unlock(&b->lock);
131 	return true;
132 }
133 
134 /*
135  * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
136  * @token:	Token to identify the sleep node entry
137  *
138  * Invoked from the async pagefault handling code or from the VM exit page
139  * fault handler. In both cases RCU is watching.
140  */
141 void kvm_async_pf_task_wait_schedule(u32 token)
142 {
143 	struct kvm_task_sleep_node n;
144 	DECLARE_SWAITQUEUE(wait);
145 
146 	lockdep_assert_irqs_disabled();
147 
148 	if (!kvm_async_pf_queue_task(token, &n))
149 		return;
150 
151 	for (;;) {
152 		prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
153 		if (hlist_unhashed(&n.link))
154 			break;
155 
156 		local_irq_enable();
157 		schedule();
158 		local_irq_disable();
159 	}
160 	finish_swait(&n.wq, &wait);
161 }
162 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
163 
164 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
165 {
166 	hlist_del_init(&n->link);
167 	if (swq_has_sleeper(&n->wq))
168 		swake_up_one(&n->wq);
169 }
170 
171 static void apf_task_wake_all(void)
172 {
173 	int i;
174 
175 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
176 		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
177 		struct kvm_task_sleep_node *n;
178 		struct hlist_node *p, *next;
179 
180 		raw_spin_lock(&b->lock);
181 		hlist_for_each_safe(p, next, &b->list) {
182 			n = hlist_entry(p, typeof(*n), link);
183 			if (n->cpu == smp_processor_id())
184 				apf_task_wake_one(n);
185 		}
186 		raw_spin_unlock(&b->lock);
187 	}
188 }
189 
190 void kvm_async_pf_task_wake(u32 token)
191 {
192 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
193 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
194 	struct kvm_task_sleep_node *n, *dummy = NULL;
195 
196 	if (token == ~0) {
197 		apf_task_wake_all();
198 		return;
199 	}
200 
201 again:
202 	raw_spin_lock(&b->lock);
203 	n = _find_apf_task(b, token);
204 	if (!n) {
205 		/*
206 		 * Async #PF not yet handled, add a dummy entry for the token.
207 		 * Allocating the token must be down outside of the raw lock
208 		 * as the allocator is preemptible on PREEMPT_RT kernels.
209 		 */
210 		if (!dummy) {
211 			raw_spin_unlock(&b->lock);
212 			dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
213 
214 			/*
215 			 * Continue looping on allocation failure, eventually
216 			 * the async #PF will be handled and allocating a new
217 			 * node will be unnecessary.
218 			 */
219 			if (!dummy)
220 				cpu_relax();
221 
222 			/*
223 			 * Recheck for async #PF completion before enqueueing
224 			 * the dummy token to avoid duplicate list entries.
225 			 */
226 			goto again;
227 		}
228 		dummy->token = token;
229 		dummy->cpu = smp_processor_id();
230 		init_swait_queue_head(&dummy->wq);
231 		hlist_add_head(&dummy->link, &b->list);
232 		dummy = NULL;
233 	} else {
234 		apf_task_wake_one(n);
235 	}
236 	raw_spin_unlock(&b->lock);
237 
238 	/* A dummy token might be allocated and ultimately not used.  */
239 	kfree(dummy);
240 }
241 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
242 
243 noinstr u32 kvm_read_and_reset_apf_flags(void)
244 {
245 	u32 flags = 0;
246 
247 	if (__this_cpu_read(apf_reason.enabled)) {
248 		flags = __this_cpu_read(apf_reason.flags);
249 		__this_cpu_write(apf_reason.flags, 0);
250 	}
251 
252 	return flags;
253 }
254 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
255 
256 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
257 {
258 	u32 flags = kvm_read_and_reset_apf_flags();
259 	irqentry_state_t state;
260 
261 	if (!flags)
262 		return false;
263 
264 	state = irqentry_enter(regs);
265 	instrumentation_begin();
266 
267 	/*
268 	 * If the host managed to inject an async #PF into an interrupt
269 	 * disabled region, then die hard as this is not going to end well
270 	 * and the host side is seriously broken.
271 	 */
272 	if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
273 		panic("Host injected async #PF in interrupt disabled region\n");
274 
275 	if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
276 		if (unlikely(!(user_mode(regs))))
277 			panic("Host injected async #PF in kernel mode\n");
278 		/* Page is swapped out by the host. */
279 		kvm_async_pf_task_wait_schedule(token);
280 	} else {
281 		WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
282 	}
283 
284 	instrumentation_end();
285 	irqentry_exit(regs, state);
286 	return true;
287 }
288 
289 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
290 {
291 	struct pt_regs *old_regs = set_irq_regs(regs);
292 	u32 token;
293 
294 	apic_eoi();
295 
296 	inc_irq_stat(irq_hv_callback_count);
297 
298 	if (__this_cpu_read(apf_reason.enabled)) {
299 		token = __this_cpu_read(apf_reason.token);
300 		kvm_async_pf_task_wake(token);
301 		__this_cpu_write(apf_reason.token, 0);
302 		wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
303 	}
304 
305 	set_irq_regs(old_regs);
306 }
307 
308 static void __init paravirt_ops_setup(void)
309 {
310 	pv_info.name = "KVM";
311 
312 	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
313 		pv_ops.cpu.io_delay = kvm_io_delay;
314 
315 #ifdef CONFIG_X86_IO_APIC
316 	no_timer_check = 1;
317 #endif
318 }
319 
320 static void kvm_register_steal_time(void)
321 {
322 	int cpu = smp_processor_id();
323 	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
324 
325 	if (!has_steal_clock)
326 		return;
327 
328 	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
329 	pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
330 		(unsigned long long) slow_virt_to_phys(st));
331 }
332 
333 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
334 
335 static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
336 {
337 	/**
338 	 * This relies on __test_and_clear_bit to modify the memory
339 	 * in a way that is atomic with respect to the local CPU.
340 	 * The hypervisor only accesses this memory from the local CPU so
341 	 * there's no need for lock or memory barriers.
342 	 * An optimization barrier is implied in apic write.
343 	 */
344 	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
345 		return;
346 	apic_native_eoi();
347 }
348 
349 static void kvm_guest_cpu_init(void)
350 {
351 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
352 		u64 pa;
353 
354 		WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
355 
356 		pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
357 		pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
358 
359 		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
360 			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
361 
362 		wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
363 
364 		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
365 		__this_cpu_write(apf_reason.enabled, 1);
366 		pr_debug("setup async PF for cpu %d\n", smp_processor_id());
367 	}
368 
369 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
370 		unsigned long pa;
371 
372 		/* Size alignment is implied but just to make it explicit. */
373 		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
374 		__this_cpu_write(kvm_apic_eoi, 0);
375 		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
376 			| KVM_MSR_ENABLED;
377 		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
378 	}
379 
380 	if (has_steal_clock)
381 		kvm_register_steal_time();
382 }
383 
384 static void kvm_pv_disable_apf(void)
385 {
386 	if (!__this_cpu_read(apf_reason.enabled))
387 		return;
388 
389 	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
390 	__this_cpu_write(apf_reason.enabled, 0);
391 
392 	pr_debug("disable async PF for cpu %d\n", smp_processor_id());
393 }
394 
395 static void kvm_disable_steal_time(void)
396 {
397 	if (!has_steal_clock)
398 		return;
399 
400 	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
401 }
402 
403 static u64 kvm_steal_clock(int cpu)
404 {
405 	u64 steal;
406 	struct kvm_steal_time *src;
407 	int version;
408 
409 	src = &per_cpu(steal_time, cpu);
410 	do {
411 		version = src->version;
412 		virt_rmb();
413 		steal = src->steal;
414 		virt_rmb();
415 	} while ((version & 1) || (version != src->version));
416 
417 	return steal;
418 }
419 
420 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
421 {
422 	early_set_memory_decrypted((unsigned long) ptr, size);
423 }
424 
425 /*
426  * Iterate through all possible CPUs and map the memory region pointed
427  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
428  *
429  * Note: we iterate through all possible CPUs to ensure that CPUs
430  * hotplugged will have their per-cpu variable already mapped as
431  * decrypted.
432  */
433 static void __init sev_map_percpu_data(void)
434 {
435 	int cpu;
436 
437 	if (cc_vendor != CC_VENDOR_AMD ||
438 	    !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
439 		return;
440 
441 	for_each_possible_cpu(cpu) {
442 		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
443 		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
444 		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
445 	}
446 }
447 
448 static void kvm_guest_cpu_offline(bool shutdown)
449 {
450 	kvm_disable_steal_time();
451 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
452 		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
453 	if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
454 		wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
455 	kvm_pv_disable_apf();
456 	if (!shutdown)
457 		apf_task_wake_all();
458 	kvmclock_disable();
459 }
460 
461 static int kvm_cpu_online(unsigned int cpu)
462 {
463 	unsigned long flags;
464 
465 	local_irq_save(flags);
466 	kvm_guest_cpu_init();
467 	local_irq_restore(flags);
468 	return 0;
469 }
470 
471 #ifdef CONFIG_SMP
472 
473 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
474 
475 static bool pv_tlb_flush_supported(void)
476 {
477 	return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
478 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
479 		kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
480 		!boot_cpu_has(X86_FEATURE_MWAIT) &&
481 		(num_possible_cpus() != 1));
482 }
483 
484 static bool pv_ipi_supported(void)
485 {
486 	return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
487 	       (num_possible_cpus() != 1));
488 }
489 
490 static bool pv_sched_yield_supported(void)
491 {
492 	return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
493 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
494 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
495 	    !boot_cpu_has(X86_FEATURE_MWAIT) &&
496 	    (num_possible_cpus() != 1));
497 }
498 
499 #define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
500 
501 static void __send_ipi_mask(const struct cpumask *mask, int vector)
502 {
503 	unsigned long flags;
504 	int cpu, min = 0, max = 0;
505 #ifdef CONFIG_X86_64
506 	__uint128_t ipi_bitmap = 0;
507 #else
508 	u64 ipi_bitmap = 0;
509 #endif
510 	u32 apic_id, icr;
511 	long ret;
512 
513 	if (cpumask_empty(mask))
514 		return;
515 
516 	local_irq_save(flags);
517 
518 	switch (vector) {
519 	default:
520 		icr = APIC_DM_FIXED | vector;
521 		break;
522 	case NMI_VECTOR:
523 		icr = APIC_DM_NMI;
524 		break;
525 	}
526 
527 	for_each_cpu(cpu, mask) {
528 		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
529 		if (!ipi_bitmap) {
530 			min = max = apic_id;
531 		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
532 			ipi_bitmap <<= min - apic_id;
533 			min = apic_id;
534 		} else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
535 			max = apic_id < max ? max : apic_id;
536 		} else {
537 			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
538 				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
539 			WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
540 				  ret);
541 			min = max = apic_id;
542 			ipi_bitmap = 0;
543 		}
544 		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
545 	}
546 
547 	if (ipi_bitmap) {
548 		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
549 			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
550 		WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
551 			  ret);
552 	}
553 
554 	local_irq_restore(flags);
555 }
556 
557 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
558 {
559 	__send_ipi_mask(mask, vector);
560 }
561 
562 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
563 {
564 	unsigned int this_cpu = smp_processor_id();
565 	struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
566 	const struct cpumask *local_mask;
567 
568 	cpumask_copy(new_mask, mask);
569 	cpumask_clear_cpu(this_cpu, new_mask);
570 	local_mask = new_mask;
571 	__send_ipi_mask(local_mask, vector);
572 }
573 
574 static int __init setup_efi_kvm_sev_migration(void)
575 {
576 	efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
577 	efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
578 	efi_status_t status;
579 	unsigned long size;
580 	bool enabled;
581 
582 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
583 	    !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
584 		return 0;
585 
586 	if (!efi_enabled(EFI_BOOT))
587 		return 0;
588 
589 	if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
590 		pr_info("%s : EFI runtime services are not enabled\n", __func__);
591 		return 0;
592 	}
593 
594 	size = sizeof(enabled);
595 
596 	/* Get variable contents into buffer */
597 	status = efi.get_variable(efi_sev_live_migration_enabled,
598 				  &efi_variable_guid, NULL, &size, &enabled);
599 
600 	if (status == EFI_NOT_FOUND) {
601 		pr_info("%s : EFI live migration variable not found\n", __func__);
602 		return 0;
603 	}
604 
605 	if (status != EFI_SUCCESS) {
606 		pr_info("%s : EFI variable retrieval failed\n", __func__);
607 		return 0;
608 	}
609 
610 	if (enabled == 0) {
611 		pr_info("%s: live migration disabled in EFI\n", __func__);
612 		return 0;
613 	}
614 
615 	pr_info("%s : live migration enabled in EFI\n", __func__);
616 	wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
617 
618 	return 1;
619 }
620 
621 late_initcall(setup_efi_kvm_sev_migration);
622 
623 /*
624  * Set the IPI entry points
625  */
626 static __init void kvm_setup_pv_ipi(void)
627 {
628 	apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
629 	apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
630 	pr_info("setup PV IPIs\n");
631 }
632 
633 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
634 {
635 	int cpu;
636 
637 	native_send_call_func_ipi(mask);
638 
639 	/* Make sure other vCPUs get a chance to run if they need to. */
640 	for_each_cpu(cpu, mask) {
641 		if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
642 			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
643 			break;
644 		}
645 	}
646 }
647 
648 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
649 			const struct flush_tlb_info *info)
650 {
651 	u8 state;
652 	int cpu;
653 	struct kvm_steal_time *src;
654 	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
655 
656 	cpumask_copy(flushmask, cpumask);
657 	/*
658 	 * We have to call flush only on online vCPUs. And
659 	 * queue flush_on_enter for pre-empted vCPUs
660 	 */
661 	for_each_cpu(cpu, flushmask) {
662 		/*
663 		 * The local vCPU is never preempted, so we do not explicitly
664 		 * skip check for local vCPU - it will never be cleared from
665 		 * flushmask.
666 		 */
667 		src = &per_cpu(steal_time, cpu);
668 		state = READ_ONCE(src->preempted);
669 		if ((state & KVM_VCPU_PREEMPTED)) {
670 			if (try_cmpxchg(&src->preempted, &state,
671 					state | KVM_VCPU_FLUSH_TLB))
672 				__cpumask_clear_cpu(cpu, flushmask);
673 		}
674 	}
675 
676 	native_flush_tlb_multi(flushmask, info);
677 }
678 
679 static __init int kvm_alloc_cpumask(void)
680 {
681 	int cpu;
682 
683 	if (!kvm_para_available() || nopv)
684 		return 0;
685 
686 	if (pv_tlb_flush_supported() || pv_ipi_supported())
687 		for_each_possible_cpu(cpu) {
688 			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
689 				GFP_KERNEL, cpu_to_node(cpu));
690 		}
691 
692 	return 0;
693 }
694 arch_initcall(kvm_alloc_cpumask);
695 
696 static void __init kvm_smp_prepare_boot_cpu(void)
697 {
698 	/*
699 	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
700 	 * shares the guest physical address with the hypervisor.
701 	 */
702 	sev_map_percpu_data();
703 
704 	kvm_guest_cpu_init();
705 	native_smp_prepare_boot_cpu();
706 	kvm_spinlock_init();
707 }
708 
709 static int kvm_cpu_down_prepare(unsigned int cpu)
710 {
711 	unsigned long flags;
712 
713 	local_irq_save(flags);
714 	kvm_guest_cpu_offline(false);
715 	local_irq_restore(flags);
716 	return 0;
717 }
718 
719 #endif
720 
721 static int kvm_suspend(void)
722 {
723 	u64 val = 0;
724 
725 	kvm_guest_cpu_offline(false);
726 
727 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
728 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
729 		rdmsrl(MSR_KVM_POLL_CONTROL, val);
730 	has_guest_poll = !(val & 1);
731 #endif
732 	return 0;
733 }
734 
735 static void kvm_resume(void)
736 {
737 	kvm_cpu_online(raw_smp_processor_id());
738 
739 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
740 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
741 		wrmsrl(MSR_KVM_POLL_CONTROL, 0);
742 #endif
743 }
744 
745 static struct syscore_ops kvm_syscore_ops = {
746 	.suspend	= kvm_suspend,
747 	.resume		= kvm_resume,
748 };
749 
750 static void kvm_pv_guest_cpu_reboot(void *unused)
751 {
752 	kvm_guest_cpu_offline(true);
753 }
754 
755 static int kvm_pv_reboot_notify(struct notifier_block *nb,
756 				unsigned long code, void *unused)
757 {
758 	if (code == SYS_RESTART)
759 		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
760 	return NOTIFY_DONE;
761 }
762 
763 static struct notifier_block kvm_pv_reboot_nb = {
764 	.notifier_call = kvm_pv_reboot_notify,
765 };
766 
767 /*
768  * After a PV feature is registered, the host will keep writing to the
769  * registered memory location. If the guest happens to shutdown, this memory
770  * won't be valid. In cases like kexec, in which you install a new kernel, this
771  * means a random memory location will be kept being written.
772  */
773 #ifdef CONFIG_KEXEC_CORE
774 static void kvm_crash_shutdown(struct pt_regs *regs)
775 {
776 	kvm_guest_cpu_offline(true);
777 	native_machine_crash_shutdown(regs);
778 }
779 #endif
780 
781 #if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
782 bool __kvm_vcpu_is_preempted(long cpu);
783 
784 __visible bool __kvm_vcpu_is_preempted(long cpu)
785 {
786 	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
787 
788 	return !!(src->preempted & KVM_VCPU_PREEMPTED);
789 }
790 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
791 
792 #else
793 
794 #include <asm/asm-offsets.h>
795 
796 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
797 
798 /*
799  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
800  * restoring to/from the stack.
801  */
802 #define PV_VCPU_PREEMPTED_ASM						     \
803  "movq   __per_cpu_offset(,%rdi,8), %rax\n\t"				     \
804  "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
805  "setne  %al\n\t"
806 
807 DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
808 		PV_VCPU_PREEMPTED_ASM, .text);
809 #endif
810 
811 static void __init kvm_guest_init(void)
812 {
813 	int i;
814 
815 	paravirt_ops_setup();
816 	register_reboot_notifier(&kvm_pv_reboot_nb);
817 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
818 		raw_spin_lock_init(&async_pf_sleepers[i].lock);
819 
820 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
821 		has_steal_clock = 1;
822 		static_call_update(pv_steal_clock, kvm_steal_clock);
823 
824 		pv_ops.lock.vcpu_is_preempted =
825 			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
826 	}
827 
828 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
829 		apic_update_callback(eoi, kvm_guest_apic_eoi_write);
830 
831 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
832 		static_branch_enable(&kvm_async_pf_enabled);
833 		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
834 	}
835 
836 #ifdef CONFIG_SMP
837 	if (pv_tlb_flush_supported()) {
838 		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
839 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
840 		pr_info("KVM setup pv remote TLB flush\n");
841 	}
842 
843 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
844 	if (pv_sched_yield_supported()) {
845 		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
846 		pr_info("setup PV sched yield\n");
847 	}
848 	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
849 				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
850 		pr_err("failed to install cpu hotplug callbacks\n");
851 #else
852 	sev_map_percpu_data();
853 	kvm_guest_cpu_init();
854 #endif
855 
856 #ifdef CONFIG_KEXEC_CORE
857 	machine_ops.crash_shutdown = kvm_crash_shutdown;
858 #endif
859 
860 	register_syscore_ops(&kvm_syscore_ops);
861 
862 	/*
863 	 * Hard lockup detection is enabled by default. Disable it, as guests
864 	 * can get false positives too easily, for example if the host is
865 	 * overcommitted.
866 	 */
867 	hardlockup_detector_disable();
868 }
869 
870 static noinline uint32_t __kvm_cpuid_base(void)
871 {
872 	if (boot_cpu_data.cpuid_level < 0)
873 		return 0;	/* So we don't blow up on old processors */
874 
875 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
876 		return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
877 
878 	return 0;
879 }
880 
881 static inline uint32_t kvm_cpuid_base(void)
882 {
883 	static int kvm_cpuid_base = -1;
884 
885 	if (kvm_cpuid_base == -1)
886 		kvm_cpuid_base = __kvm_cpuid_base();
887 
888 	return kvm_cpuid_base;
889 }
890 
891 bool kvm_para_available(void)
892 {
893 	return kvm_cpuid_base() != 0;
894 }
895 EXPORT_SYMBOL_GPL(kvm_para_available);
896 
897 unsigned int kvm_arch_para_features(void)
898 {
899 	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
900 }
901 
902 unsigned int kvm_arch_para_hints(void)
903 {
904 	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
905 }
906 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
907 
908 static uint32_t __init kvm_detect(void)
909 {
910 	return kvm_cpuid_base();
911 }
912 
913 static void __init kvm_apic_init(void)
914 {
915 #ifdef CONFIG_SMP
916 	if (pv_ipi_supported())
917 		kvm_setup_pv_ipi();
918 #endif
919 }
920 
921 static bool __init kvm_msi_ext_dest_id(void)
922 {
923 	return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
924 }
925 
926 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
927 {
928 	kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
929 			   KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
930 }
931 
932 static void __init kvm_init_platform(void)
933 {
934 	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
935 	    kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
936 		unsigned long nr_pages;
937 		int i;
938 
939 		pv_ops.mmu.notify_page_enc_status_changed =
940 			kvm_sev_hc_page_enc_status;
941 
942 		/*
943 		 * Reset the host's shared pages list related to kernel
944 		 * specific page encryption status settings before we load a
945 		 * new kernel by kexec. Reset the page encryption status
946 		 * during early boot instead of just before kexec to avoid SMP
947 		 * races during kvm_pv_guest_cpu_reboot().
948 		 * NOTE: We cannot reset the complete shared pages list
949 		 * here as we need to retain the UEFI/OVMF firmware
950 		 * specific settings.
951 		 */
952 
953 		for (i = 0; i < e820_table->nr_entries; i++) {
954 			struct e820_entry *entry = &e820_table->entries[i];
955 
956 			if (entry->type != E820_TYPE_RAM)
957 				continue;
958 
959 			nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
960 
961 			kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
962 				       nr_pages,
963 				       KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
964 		}
965 
966 		/*
967 		 * Ensure that _bss_decrypted section is marked as decrypted in the
968 		 * shared pages list.
969 		 */
970 		early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
971 						__end_bss_decrypted - __start_bss_decrypted, 0);
972 
973 		/*
974 		 * If not booted using EFI, enable Live migration support.
975 		 */
976 		if (!efi_enabled(EFI_BOOT))
977 			wrmsrl(MSR_KVM_MIGRATION_CONTROL,
978 			       KVM_MIGRATION_READY);
979 	}
980 	kvmclock_init();
981 	x86_platform.apic_post_init = kvm_apic_init;
982 }
983 
984 #if defined(CONFIG_AMD_MEM_ENCRYPT)
985 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
986 {
987 	/* RAX and CPL are already in the GHCB */
988 	ghcb_set_rbx(ghcb, regs->bx);
989 	ghcb_set_rcx(ghcb, regs->cx);
990 	ghcb_set_rdx(ghcb, regs->dx);
991 	ghcb_set_rsi(ghcb, regs->si);
992 }
993 
994 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
995 {
996 	/* No checking of the return state needed */
997 	return true;
998 }
999 #endif
1000 
1001 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1002 	.name				= "KVM",
1003 	.detect				= kvm_detect,
1004 	.type				= X86_HYPER_KVM,
1005 	.init.guest_late_init		= kvm_guest_init,
1006 	.init.x2apic_available		= kvm_para_available,
1007 	.init.msi_ext_dest_id		= kvm_msi_ext_dest_id,
1008 	.init.init_platform		= kvm_init_platform,
1009 #if defined(CONFIG_AMD_MEM_ENCRYPT)
1010 	.runtime.sev_es_hcall_prepare	= kvm_sev_es_hcall_prepare,
1011 	.runtime.sev_es_hcall_finish	= kvm_sev_es_hcall_finish,
1012 #endif
1013 };
1014 
1015 static __init int activate_jump_labels(void)
1016 {
1017 	if (has_steal_clock) {
1018 		static_key_slow_inc(&paravirt_steal_enabled);
1019 		if (steal_acc)
1020 			static_key_slow_inc(&paravirt_steal_rq_enabled);
1021 	}
1022 
1023 	return 0;
1024 }
1025 arch_initcall(activate_jump_labels);
1026 
1027 #ifdef CONFIG_PARAVIRT_SPINLOCKS
1028 
1029 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1030 static void kvm_kick_cpu(int cpu)
1031 {
1032 	unsigned long flags = 0;
1033 	u32 apicid;
1034 
1035 	apicid = per_cpu(x86_cpu_to_apicid, cpu);
1036 	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1037 }
1038 
1039 #include <asm/qspinlock.h>
1040 
1041 static void kvm_wait(u8 *ptr, u8 val)
1042 {
1043 	if (in_nmi())
1044 		return;
1045 
1046 	/*
1047 	 * halt until it's our turn and kicked. Note that we do safe halt
1048 	 * for irq enabled case to avoid hang when lock info is overwritten
1049 	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1050 	 */
1051 	if (irqs_disabled()) {
1052 		if (READ_ONCE(*ptr) == val)
1053 			halt();
1054 	} else {
1055 		local_irq_disable();
1056 
1057 		/* safe_halt() will enable IRQ */
1058 		if (READ_ONCE(*ptr) == val)
1059 			safe_halt();
1060 		else
1061 			local_irq_enable();
1062 	}
1063 }
1064 
1065 /*
1066  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1067  */
1068 void __init kvm_spinlock_init(void)
1069 {
1070 	/*
1071 	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1072 	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1073 	 * preferred over native qspinlock when vCPU is preempted.
1074 	 */
1075 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1076 		pr_info("PV spinlocks disabled, no host support\n");
1077 		return;
1078 	}
1079 
1080 	/*
1081 	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1082 	 * are available.
1083 	 */
1084 	if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1085 		pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1086 		goto out;
1087 	}
1088 
1089 	if (num_possible_cpus() == 1) {
1090 		pr_info("PV spinlocks disabled, single CPU\n");
1091 		goto out;
1092 	}
1093 
1094 	if (nopvspin) {
1095 		pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1096 		goto out;
1097 	}
1098 
1099 	pr_info("PV spinlocks enabled\n");
1100 
1101 	__pv_init_lock_hash();
1102 	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1103 	pv_ops.lock.queued_spin_unlock =
1104 		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1105 	pv_ops.lock.wait = kvm_wait;
1106 	pv_ops.lock.kick = kvm_kick_cpu;
1107 
1108 	/*
1109 	 * When PV spinlock is enabled which is preferred over
1110 	 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1111 	 * Just disable it anyway.
1112 	 */
1113 out:
1114 	static_branch_disable(&virt_spin_lock_key);
1115 }
1116 
1117 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
1118 
1119 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1120 
1121 static void kvm_disable_host_haltpoll(void *i)
1122 {
1123 	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1124 }
1125 
1126 static void kvm_enable_host_haltpoll(void *i)
1127 {
1128 	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1129 }
1130 
1131 void arch_haltpoll_enable(unsigned int cpu)
1132 {
1133 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1134 		pr_err_once("host does not support poll control\n");
1135 		pr_err_once("host upgrade recommended\n");
1136 		return;
1137 	}
1138 
1139 	/* Enable guest halt poll disables host halt poll */
1140 	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1141 }
1142 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1143 
1144 void arch_haltpoll_disable(unsigned int cpu)
1145 {
1146 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1147 		return;
1148 
1149 	/* Disable guest halt poll enables host halt poll */
1150 	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1151 }
1152 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1153 #endif
1154