xref: /openbmc/linux/arch/x86/kernel/kvm.c (revision 979452fb)
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 /*
73  * No need for any "IO delay" on KVM
74  */
75 static void kvm_io_delay(void)
76 {
77 }
78 
79 #define KVM_TASK_SLEEP_HASHBITS 8
80 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
81 
82 struct kvm_task_sleep_node {
83 	struct hlist_node link;
84 	struct swait_queue_head wq;
85 	u32 token;
86 	int cpu;
87 };
88 
89 static struct kvm_task_sleep_head {
90 	raw_spinlock_t lock;
91 	struct hlist_head list;
92 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
93 
94 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
95 						  u32 token)
96 {
97 	struct hlist_node *p;
98 
99 	hlist_for_each(p, &b->list) {
100 		struct kvm_task_sleep_node *n =
101 			hlist_entry(p, typeof(*n), link);
102 		if (n->token == token)
103 			return n;
104 	}
105 
106 	return NULL;
107 }
108 
109 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
110 {
111 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
112 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
113 	struct kvm_task_sleep_node *e;
114 
115 	raw_spin_lock(&b->lock);
116 	e = _find_apf_task(b, token);
117 	if (e) {
118 		/* dummy entry exist -> wake up was delivered ahead of PF */
119 		hlist_del(&e->link);
120 		raw_spin_unlock(&b->lock);
121 		kfree(e);
122 		return false;
123 	}
124 
125 	n->token = token;
126 	n->cpu = smp_processor_id();
127 	init_swait_queue_head(&n->wq);
128 	hlist_add_head(&n->link, &b->list);
129 	raw_spin_unlock(&b->lock);
130 	return true;
131 }
132 
133 /*
134  * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
135  * @token:	Token to identify the sleep node entry
136  *
137  * Invoked from the async pagefault handling code or from the VM exit page
138  * fault handler. In both cases RCU is watching.
139  */
140 void kvm_async_pf_task_wait_schedule(u32 token)
141 {
142 	struct kvm_task_sleep_node n;
143 	DECLARE_SWAITQUEUE(wait);
144 
145 	lockdep_assert_irqs_disabled();
146 
147 	if (!kvm_async_pf_queue_task(token, &n))
148 		return;
149 
150 	for (;;) {
151 		prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
152 		if (hlist_unhashed(&n.link))
153 			break;
154 
155 		local_irq_enable();
156 		schedule();
157 		local_irq_disable();
158 	}
159 	finish_swait(&n.wq, &wait);
160 }
161 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
162 
163 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
164 {
165 	hlist_del_init(&n->link);
166 	if (swq_has_sleeper(&n->wq))
167 		swake_up_one(&n->wq);
168 }
169 
170 static void apf_task_wake_all(void)
171 {
172 	int i;
173 
174 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
175 		struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
176 		struct kvm_task_sleep_node *n;
177 		struct hlist_node *p, *next;
178 
179 		raw_spin_lock(&b->lock);
180 		hlist_for_each_safe(p, next, &b->list) {
181 			n = hlist_entry(p, typeof(*n), link);
182 			if (n->cpu == smp_processor_id())
183 				apf_task_wake_one(n);
184 		}
185 		raw_spin_unlock(&b->lock);
186 	}
187 }
188 
189 void kvm_async_pf_task_wake(u32 token)
190 {
191 	u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
192 	struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
193 	struct kvm_task_sleep_node *n;
194 
195 	if (token == ~0) {
196 		apf_task_wake_all();
197 		return;
198 	}
199 
200 again:
201 	raw_spin_lock(&b->lock);
202 	n = _find_apf_task(b, token);
203 	if (!n) {
204 		/*
205 		 * async PF was not yet handled.
206 		 * Add dummy entry for the token.
207 		 */
208 		n = kzalloc(sizeof(*n), GFP_ATOMIC);
209 		if (!n) {
210 			/*
211 			 * Allocation failed! Busy wait while other cpu
212 			 * handles async PF.
213 			 */
214 			raw_spin_unlock(&b->lock);
215 			cpu_relax();
216 			goto again;
217 		}
218 		n->token = token;
219 		n->cpu = smp_processor_id();
220 		init_swait_queue_head(&n->wq);
221 		hlist_add_head(&n->link, &b->list);
222 	} else {
223 		apf_task_wake_one(n);
224 	}
225 	raw_spin_unlock(&b->lock);
226 	return;
227 }
228 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
229 
230 noinstr u32 kvm_read_and_reset_apf_flags(void)
231 {
232 	u32 flags = 0;
233 
234 	if (__this_cpu_read(apf_reason.enabled)) {
235 		flags = __this_cpu_read(apf_reason.flags);
236 		__this_cpu_write(apf_reason.flags, 0);
237 	}
238 
239 	return flags;
240 }
241 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
242 
243 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
244 {
245 	u32 flags = kvm_read_and_reset_apf_flags();
246 	irqentry_state_t state;
247 
248 	if (!flags)
249 		return false;
250 
251 	state = irqentry_enter(regs);
252 	instrumentation_begin();
253 
254 	/*
255 	 * If the host managed to inject an async #PF into an interrupt
256 	 * disabled region, then die hard as this is not going to end well
257 	 * and the host side is seriously broken.
258 	 */
259 	if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
260 		panic("Host injected async #PF in interrupt disabled region\n");
261 
262 	if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
263 		if (unlikely(!(user_mode(regs))))
264 			panic("Host injected async #PF in kernel mode\n");
265 		/* Page is swapped out by the host. */
266 		kvm_async_pf_task_wait_schedule(token);
267 	} else {
268 		WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
269 	}
270 
271 	instrumentation_end();
272 	irqentry_exit(regs, state);
273 	return true;
274 }
275 
276 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
277 {
278 	struct pt_regs *old_regs = set_irq_regs(regs);
279 	u32 token;
280 
281 	ack_APIC_irq();
282 
283 	inc_irq_stat(irq_hv_callback_count);
284 
285 	if (__this_cpu_read(apf_reason.enabled)) {
286 		token = __this_cpu_read(apf_reason.token);
287 		kvm_async_pf_task_wake(token);
288 		__this_cpu_write(apf_reason.token, 0);
289 		wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
290 	}
291 
292 	set_irq_regs(old_regs);
293 }
294 
295 static void __init paravirt_ops_setup(void)
296 {
297 	pv_info.name = "KVM";
298 
299 	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
300 		pv_ops.cpu.io_delay = kvm_io_delay;
301 
302 #ifdef CONFIG_X86_IO_APIC
303 	no_timer_check = 1;
304 #endif
305 }
306 
307 static void kvm_register_steal_time(void)
308 {
309 	int cpu = smp_processor_id();
310 	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
311 
312 	if (!has_steal_clock)
313 		return;
314 
315 	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
316 	pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
317 		(unsigned long long) slow_virt_to_phys(st));
318 }
319 
320 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
321 
322 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
323 {
324 	/**
325 	 * This relies on __test_and_clear_bit to modify the memory
326 	 * in a way that is atomic with respect to the local CPU.
327 	 * The hypervisor only accesses this memory from the local CPU so
328 	 * there's no need for lock or memory barriers.
329 	 * An optimization barrier is implied in apic write.
330 	 */
331 	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
332 		return;
333 	apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
334 }
335 
336 static void kvm_guest_cpu_init(void)
337 {
338 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
339 		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
340 
341 		WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
342 
343 		pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
344 		pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
345 
346 		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
347 			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
348 
349 		wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
350 
351 		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
352 		__this_cpu_write(apf_reason.enabled, 1);
353 		pr_debug("setup async PF for cpu %d\n", smp_processor_id());
354 	}
355 
356 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
357 		unsigned long pa;
358 
359 		/* Size alignment is implied but just to make it explicit. */
360 		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
361 		__this_cpu_write(kvm_apic_eoi, 0);
362 		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
363 			| KVM_MSR_ENABLED;
364 		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
365 	}
366 
367 	if (has_steal_clock)
368 		kvm_register_steal_time();
369 }
370 
371 static void kvm_pv_disable_apf(void)
372 {
373 	if (!__this_cpu_read(apf_reason.enabled))
374 		return;
375 
376 	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
377 	__this_cpu_write(apf_reason.enabled, 0);
378 
379 	pr_debug("disable async PF for cpu %d\n", smp_processor_id());
380 }
381 
382 static void kvm_disable_steal_time(void)
383 {
384 	if (!has_steal_clock)
385 		return;
386 
387 	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
388 }
389 
390 static u64 kvm_steal_clock(int cpu)
391 {
392 	u64 steal;
393 	struct kvm_steal_time *src;
394 	int version;
395 
396 	src = &per_cpu(steal_time, cpu);
397 	do {
398 		version = src->version;
399 		virt_rmb();
400 		steal = src->steal;
401 		virt_rmb();
402 	} while ((version & 1) || (version != src->version));
403 
404 	return steal;
405 }
406 
407 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
408 {
409 	early_set_memory_decrypted((unsigned long) ptr, size);
410 }
411 
412 /*
413  * Iterate through all possible CPUs and map the memory region pointed
414  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
415  *
416  * Note: we iterate through all possible CPUs to ensure that CPUs
417  * hotplugged will have their per-cpu variable already mapped as
418  * decrypted.
419  */
420 static void __init sev_map_percpu_data(void)
421 {
422 	int cpu;
423 
424 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
425 		return;
426 
427 	for_each_possible_cpu(cpu) {
428 		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
429 		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
430 		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
431 	}
432 }
433 
434 static void kvm_guest_cpu_offline(bool shutdown)
435 {
436 	kvm_disable_steal_time();
437 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
438 		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
439 	if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
440 		wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
441 	kvm_pv_disable_apf();
442 	if (!shutdown)
443 		apf_task_wake_all();
444 	kvmclock_disable();
445 }
446 
447 static int kvm_cpu_online(unsigned int cpu)
448 {
449 	unsigned long flags;
450 
451 	local_irq_save(flags);
452 	kvm_guest_cpu_init();
453 	local_irq_restore(flags);
454 	return 0;
455 }
456 
457 #ifdef CONFIG_SMP
458 
459 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
460 
461 static bool pv_tlb_flush_supported(void)
462 {
463 	return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
464 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
465 		kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
466 		(num_possible_cpus() != 1));
467 }
468 
469 static bool pv_ipi_supported(void)
470 {
471 	return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
472 	       (num_possible_cpus() != 1));
473 }
474 
475 static bool pv_sched_yield_supported(void)
476 {
477 	return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
478 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
479 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
480 	    (num_possible_cpus() != 1));
481 }
482 
483 #define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
484 
485 static void __send_ipi_mask(const struct cpumask *mask, int vector)
486 {
487 	unsigned long flags;
488 	int cpu, apic_id, icr;
489 	int min = 0, max = 0;
490 #ifdef CONFIG_X86_64
491 	__uint128_t ipi_bitmap = 0;
492 #else
493 	u64 ipi_bitmap = 0;
494 #endif
495 	long ret;
496 
497 	if (cpumask_empty(mask))
498 		return;
499 
500 	local_irq_save(flags);
501 
502 	switch (vector) {
503 	default:
504 		icr = APIC_DM_FIXED | vector;
505 		break;
506 	case NMI_VECTOR:
507 		icr = APIC_DM_NMI;
508 		break;
509 	}
510 
511 	for_each_cpu(cpu, mask) {
512 		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
513 		if (!ipi_bitmap) {
514 			min = max = apic_id;
515 		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
516 			ipi_bitmap <<= min - apic_id;
517 			min = apic_id;
518 		} else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
519 			max = apic_id < max ? max : apic_id;
520 		} else {
521 			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
522 				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
523 			WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
524 				  ret);
525 			min = max = apic_id;
526 			ipi_bitmap = 0;
527 		}
528 		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
529 	}
530 
531 	if (ipi_bitmap) {
532 		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
533 			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
534 		WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
535 			  ret);
536 	}
537 
538 	local_irq_restore(flags);
539 }
540 
541 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
542 {
543 	__send_ipi_mask(mask, vector);
544 }
545 
546 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
547 {
548 	unsigned int this_cpu = smp_processor_id();
549 	struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
550 	const struct cpumask *local_mask;
551 
552 	cpumask_copy(new_mask, mask);
553 	cpumask_clear_cpu(this_cpu, new_mask);
554 	local_mask = new_mask;
555 	__send_ipi_mask(local_mask, vector);
556 }
557 
558 static int __init setup_efi_kvm_sev_migration(void)
559 {
560 	efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
561 	efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
562 	efi_status_t status;
563 	unsigned long size;
564 	bool enabled;
565 
566 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
567 	    !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
568 		return 0;
569 
570 	if (!efi_enabled(EFI_BOOT))
571 		return 0;
572 
573 	if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
574 		pr_info("%s : EFI runtime services are not enabled\n", __func__);
575 		return 0;
576 	}
577 
578 	size = sizeof(enabled);
579 
580 	/* Get variable contents into buffer */
581 	status = efi.get_variable(efi_sev_live_migration_enabled,
582 				  &efi_variable_guid, NULL, &size, &enabled);
583 
584 	if (status == EFI_NOT_FOUND) {
585 		pr_info("%s : EFI live migration variable not found\n", __func__);
586 		return 0;
587 	}
588 
589 	if (status != EFI_SUCCESS) {
590 		pr_info("%s : EFI variable retrieval failed\n", __func__);
591 		return 0;
592 	}
593 
594 	if (enabled == 0) {
595 		pr_info("%s: live migration disabled in EFI\n", __func__);
596 		return 0;
597 	}
598 
599 	pr_info("%s : live migration enabled in EFI\n", __func__);
600 	wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
601 
602 	return 1;
603 }
604 
605 late_initcall(setup_efi_kvm_sev_migration);
606 
607 /*
608  * Set the IPI entry points
609  */
610 static void kvm_setup_pv_ipi(void)
611 {
612 	apic->send_IPI_mask = kvm_send_ipi_mask;
613 	apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
614 	pr_info("setup PV IPIs\n");
615 }
616 
617 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
618 {
619 	int cpu;
620 
621 	native_send_call_func_ipi(mask);
622 
623 	/* Make sure other vCPUs get a chance to run if they need to. */
624 	for_each_cpu(cpu, mask) {
625 		if (vcpu_is_preempted(cpu)) {
626 			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
627 			break;
628 		}
629 	}
630 }
631 
632 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
633 			const struct flush_tlb_info *info)
634 {
635 	u8 state;
636 	int cpu;
637 	struct kvm_steal_time *src;
638 	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
639 
640 	cpumask_copy(flushmask, cpumask);
641 	/*
642 	 * We have to call flush only on online vCPUs. And
643 	 * queue flush_on_enter for pre-empted vCPUs
644 	 */
645 	for_each_cpu(cpu, flushmask) {
646 		/*
647 		 * The local vCPU is never preempted, so we do not explicitly
648 		 * skip check for local vCPU - it will never be cleared from
649 		 * flushmask.
650 		 */
651 		src = &per_cpu(steal_time, cpu);
652 		state = READ_ONCE(src->preempted);
653 		if ((state & KVM_VCPU_PREEMPTED)) {
654 			if (try_cmpxchg(&src->preempted, &state,
655 					state | KVM_VCPU_FLUSH_TLB))
656 				__cpumask_clear_cpu(cpu, flushmask);
657 		}
658 	}
659 
660 	native_flush_tlb_multi(flushmask, info);
661 }
662 
663 static __init int kvm_alloc_cpumask(void)
664 {
665 	int cpu;
666 
667 	if (!kvm_para_available() || nopv)
668 		return 0;
669 
670 	if (pv_tlb_flush_supported() || pv_ipi_supported())
671 		for_each_possible_cpu(cpu) {
672 			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
673 				GFP_KERNEL, cpu_to_node(cpu));
674 		}
675 
676 	return 0;
677 }
678 arch_initcall(kvm_alloc_cpumask);
679 
680 static void __init kvm_smp_prepare_boot_cpu(void)
681 {
682 	/*
683 	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
684 	 * shares the guest physical address with the hypervisor.
685 	 */
686 	sev_map_percpu_data();
687 
688 	kvm_guest_cpu_init();
689 	native_smp_prepare_boot_cpu();
690 	kvm_spinlock_init();
691 }
692 
693 static int kvm_cpu_down_prepare(unsigned int cpu)
694 {
695 	unsigned long flags;
696 
697 	local_irq_save(flags);
698 	kvm_guest_cpu_offline(false);
699 	local_irq_restore(flags);
700 	return 0;
701 }
702 
703 #endif
704 
705 static int kvm_suspend(void)
706 {
707 	kvm_guest_cpu_offline(false);
708 
709 	return 0;
710 }
711 
712 static void kvm_resume(void)
713 {
714 	kvm_cpu_online(raw_smp_processor_id());
715 }
716 
717 static struct syscore_ops kvm_syscore_ops = {
718 	.suspend	= kvm_suspend,
719 	.resume		= kvm_resume,
720 };
721 
722 static void kvm_pv_guest_cpu_reboot(void *unused)
723 {
724 	kvm_guest_cpu_offline(true);
725 }
726 
727 static int kvm_pv_reboot_notify(struct notifier_block *nb,
728 				unsigned long code, void *unused)
729 {
730 	if (code == SYS_RESTART)
731 		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
732 	return NOTIFY_DONE;
733 }
734 
735 static struct notifier_block kvm_pv_reboot_nb = {
736 	.notifier_call = kvm_pv_reboot_notify,
737 };
738 
739 /*
740  * After a PV feature is registered, the host will keep writing to the
741  * registered memory location. If the guest happens to shutdown, this memory
742  * won't be valid. In cases like kexec, in which you install a new kernel, this
743  * means a random memory location will be kept being written.
744  */
745 #ifdef CONFIG_KEXEC_CORE
746 static void kvm_crash_shutdown(struct pt_regs *regs)
747 {
748 	kvm_guest_cpu_offline(true);
749 	native_machine_crash_shutdown(regs);
750 }
751 #endif
752 
753 static void __init kvm_guest_init(void)
754 {
755 	int i;
756 
757 	paravirt_ops_setup();
758 	register_reboot_notifier(&kvm_pv_reboot_nb);
759 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
760 		raw_spin_lock_init(&async_pf_sleepers[i].lock);
761 
762 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
763 		has_steal_clock = 1;
764 		static_call_update(pv_steal_clock, kvm_steal_clock);
765 	}
766 
767 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
768 		apic_set_eoi_write(kvm_guest_apic_eoi_write);
769 
770 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
771 		static_branch_enable(&kvm_async_pf_enabled);
772 		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
773 	}
774 
775 #ifdef CONFIG_SMP
776 	if (pv_tlb_flush_supported()) {
777 		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
778 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
779 		pr_info("KVM setup pv remote TLB flush\n");
780 	}
781 
782 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
783 	if (pv_sched_yield_supported()) {
784 		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
785 		pr_info("setup PV sched yield\n");
786 	}
787 	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
788 				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
789 		pr_err("failed to install cpu hotplug callbacks\n");
790 #else
791 	sev_map_percpu_data();
792 	kvm_guest_cpu_init();
793 #endif
794 
795 #ifdef CONFIG_KEXEC_CORE
796 	machine_ops.crash_shutdown = kvm_crash_shutdown;
797 #endif
798 
799 	register_syscore_ops(&kvm_syscore_ops);
800 
801 	/*
802 	 * Hard lockup detection is enabled by default. Disable it, as guests
803 	 * can get false positives too easily, for example if the host is
804 	 * overcommitted.
805 	 */
806 	hardlockup_detector_disable();
807 }
808 
809 static noinline uint32_t __kvm_cpuid_base(void)
810 {
811 	if (boot_cpu_data.cpuid_level < 0)
812 		return 0;	/* So we don't blow up on old processors */
813 
814 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
815 		return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
816 
817 	return 0;
818 }
819 
820 static inline uint32_t kvm_cpuid_base(void)
821 {
822 	static int kvm_cpuid_base = -1;
823 
824 	if (kvm_cpuid_base == -1)
825 		kvm_cpuid_base = __kvm_cpuid_base();
826 
827 	return kvm_cpuid_base;
828 }
829 
830 bool kvm_para_available(void)
831 {
832 	return kvm_cpuid_base() != 0;
833 }
834 EXPORT_SYMBOL_GPL(kvm_para_available);
835 
836 unsigned int kvm_arch_para_features(void)
837 {
838 	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
839 }
840 
841 unsigned int kvm_arch_para_hints(void)
842 {
843 	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
844 }
845 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
846 
847 static uint32_t __init kvm_detect(void)
848 {
849 	return kvm_cpuid_base();
850 }
851 
852 static void __init kvm_apic_init(void)
853 {
854 #ifdef CONFIG_SMP
855 	if (pv_ipi_supported())
856 		kvm_setup_pv_ipi();
857 #endif
858 }
859 
860 static bool __init kvm_msi_ext_dest_id(void)
861 {
862 	return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
863 }
864 
865 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
866 {
867 	kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
868 			   KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
869 }
870 
871 static void __init kvm_init_platform(void)
872 {
873 	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
874 	    kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
875 		unsigned long nr_pages;
876 		int i;
877 
878 		pv_ops.mmu.notify_page_enc_status_changed =
879 			kvm_sev_hc_page_enc_status;
880 
881 		/*
882 		 * Reset the host's shared pages list related to kernel
883 		 * specific page encryption status settings before we load a
884 		 * new kernel by kexec. Reset the page encryption status
885 		 * during early boot intead of just before kexec to avoid SMP
886 		 * races during kvm_pv_guest_cpu_reboot().
887 		 * NOTE: We cannot reset the complete shared pages list
888 		 * here as we need to retain the UEFI/OVMF firmware
889 		 * specific settings.
890 		 */
891 
892 		for (i = 0; i < e820_table->nr_entries; i++) {
893 			struct e820_entry *entry = &e820_table->entries[i];
894 
895 			if (entry->type != E820_TYPE_RAM)
896 				continue;
897 
898 			nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
899 
900 			kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
901 				       nr_pages,
902 				       KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
903 		}
904 
905 		/*
906 		 * Ensure that _bss_decrypted section is marked as decrypted in the
907 		 * shared pages list.
908 		 */
909 		nr_pages = DIV_ROUND_UP(__end_bss_decrypted - __start_bss_decrypted,
910 					PAGE_SIZE);
911 		early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
912 						nr_pages, 0);
913 
914 		/*
915 		 * If not booted using EFI, enable Live migration support.
916 		 */
917 		if (!efi_enabled(EFI_BOOT))
918 			wrmsrl(MSR_KVM_MIGRATION_CONTROL,
919 			       KVM_MIGRATION_READY);
920 	}
921 	kvmclock_init();
922 	x86_platform.apic_post_init = kvm_apic_init;
923 }
924 
925 #if defined(CONFIG_AMD_MEM_ENCRYPT)
926 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
927 {
928 	/* RAX and CPL are already in the GHCB */
929 	ghcb_set_rbx(ghcb, regs->bx);
930 	ghcb_set_rcx(ghcb, regs->cx);
931 	ghcb_set_rdx(ghcb, regs->dx);
932 	ghcb_set_rsi(ghcb, regs->si);
933 }
934 
935 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
936 {
937 	/* No checking of the return state needed */
938 	return true;
939 }
940 #endif
941 
942 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
943 	.name				= "KVM",
944 	.detect				= kvm_detect,
945 	.type				= X86_HYPER_KVM,
946 	.init.guest_late_init		= kvm_guest_init,
947 	.init.x2apic_available		= kvm_para_available,
948 	.init.msi_ext_dest_id		= kvm_msi_ext_dest_id,
949 	.init.init_platform		= kvm_init_platform,
950 #if defined(CONFIG_AMD_MEM_ENCRYPT)
951 	.runtime.sev_es_hcall_prepare	= kvm_sev_es_hcall_prepare,
952 	.runtime.sev_es_hcall_finish	= kvm_sev_es_hcall_finish,
953 #endif
954 };
955 
956 static __init int activate_jump_labels(void)
957 {
958 	if (has_steal_clock) {
959 		static_key_slow_inc(&paravirt_steal_enabled);
960 		if (steal_acc)
961 			static_key_slow_inc(&paravirt_steal_rq_enabled);
962 	}
963 
964 	return 0;
965 }
966 arch_initcall(activate_jump_labels);
967 
968 #ifdef CONFIG_PARAVIRT_SPINLOCKS
969 
970 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
971 static void kvm_kick_cpu(int cpu)
972 {
973 	int apicid;
974 	unsigned long flags = 0;
975 
976 	apicid = per_cpu(x86_cpu_to_apicid, cpu);
977 	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
978 }
979 
980 #include <asm/qspinlock.h>
981 
982 static void kvm_wait(u8 *ptr, u8 val)
983 {
984 	if (in_nmi())
985 		return;
986 
987 	/*
988 	 * halt until it's our turn and kicked. Note that we do safe halt
989 	 * for irq enabled case to avoid hang when lock info is overwritten
990 	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
991 	 */
992 	if (irqs_disabled()) {
993 		if (READ_ONCE(*ptr) == val)
994 			halt();
995 	} else {
996 		local_irq_disable();
997 
998 		/* safe_halt() will enable IRQ */
999 		if (READ_ONCE(*ptr) == val)
1000 			safe_halt();
1001 		else
1002 			local_irq_enable();
1003 	}
1004 }
1005 
1006 #ifdef CONFIG_X86_32
1007 __visible bool __kvm_vcpu_is_preempted(long cpu)
1008 {
1009 	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
1010 
1011 	return !!(src->preempted & KVM_VCPU_PREEMPTED);
1012 }
1013 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
1014 
1015 #else
1016 
1017 #include <asm/asm-offsets.h>
1018 
1019 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
1020 
1021 /*
1022  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
1023  * restoring to/from the stack.
1024  */
1025 asm(
1026 ".pushsection .text;"
1027 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
1028 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
1029 "__raw_callee_save___kvm_vcpu_is_preempted:"
1030 "movq	__per_cpu_offset(,%rdi,8), %rax;"
1031 "cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
1032 "setne	%al;"
1033 "ret;"
1034 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
1035 ".popsection");
1036 
1037 #endif
1038 
1039 /*
1040  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1041  */
1042 void __init kvm_spinlock_init(void)
1043 {
1044 	/*
1045 	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1046 	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1047 	 * preferred over native qspinlock when vCPU is preempted.
1048 	 */
1049 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1050 		pr_info("PV spinlocks disabled, no host support\n");
1051 		return;
1052 	}
1053 
1054 	/*
1055 	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1056 	 * are available.
1057 	 */
1058 	if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1059 		pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1060 		goto out;
1061 	}
1062 
1063 	if (num_possible_cpus() == 1) {
1064 		pr_info("PV spinlocks disabled, single CPU\n");
1065 		goto out;
1066 	}
1067 
1068 	if (nopvspin) {
1069 		pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1070 		goto out;
1071 	}
1072 
1073 	pr_info("PV spinlocks enabled\n");
1074 
1075 	__pv_init_lock_hash();
1076 	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1077 	pv_ops.lock.queued_spin_unlock =
1078 		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1079 	pv_ops.lock.wait = kvm_wait;
1080 	pv_ops.lock.kick = kvm_kick_cpu;
1081 
1082 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
1083 		pv_ops.lock.vcpu_is_preempted =
1084 			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
1085 	}
1086 	/*
1087 	 * When PV spinlock is enabled which is preferred over
1088 	 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1089 	 * Just disable it anyway.
1090 	 */
1091 out:
1092 	static_branch_disable(&virt_spin_lock_key);
1093 }
1094 
1095 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
1096 
1097 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1098 
1099 static void kvm_disable_host_haltpoll(void *i)
1100 {
1101 	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1102 }
1103 
1104 static void kvm_enable_host_haltpoll(void *i)
1105 {
1106 	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1107 }
1108 
1109 void arch_haltpoll_enable(unsigned int cpu)
1110 {
1111 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1112 		pr_err_once("host does not support poll control\n");
1113 		pr_err_once("host upgrade recommended\n");
1114 		return;
1115 	}
1116 
1117 	/* Enable guest halt poll disables host halt poll */
1118 	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1119 }
1120 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1121 
1122 void arch_haltpoll_disable(unsigned int cpu)
1123 {
1124 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1125 		return;
1126 
1127 	/* Disable guest halt poll enables host halt poll */
1128 	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1129 }
1130 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1131 #endif
1132