xref: /openbmc/linux/arch/x86/kernel/kvm.c (revision 16c8d76a)
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;
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 was not yet handled.
207 		 * Add dummy entry for the token.
208 		 */
209 		n = kzalloc(sizeof(*n), GFP_ATOMIC);
210 		if (!n) {
211 			/*
212 			 * Allocation failed! Busy wait while other cpu
213 			 * handles async PF.
214 			 */
215 			raw_spin_unlock(&b->lock);
216 			cpu_relax();
217 			goto again;
218 		}
219 		n->token = token;
220 		n->cpu = smp_processor_id();
221 		init_swait_queue_head(&n->wq);
222 		hlist_add_head(&n->link, &b->list);
223 	} else {
224 		apf_task_wake_one(n);
225 	}
226 	raw_spin_unlock(&b->lock);
227 	return;
228 }
229 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
230 
231 noinstr u32 kvm_read_and_reset_apf_flags(void)
232 {
233 	u32 flags = 0;
234 
235 	if (__this_cpu_read(apf_reason.enabled)) {
236 		flags = __this_cpu_read(apf_reason.flags);
237 		__this_cpu_write(apf_reason.flags, 0);
238 	}
239 
240 	return flags;
241 }
242 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
243 
244 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
245 {
246 	u32 flags = kvm_read_and_reset_apf_flags();
247 	irqentry_state_t state;
248 
249 	if (!flags)
250 		return false;
251 
252 	state = irqentry_enter(regs);
253 	instrumentation_begin();
254 
255 	/*
256 	 * If the host managed to inject an async #PF into an interrupt
257 	 * disabled region, then die hard as this is not going to end well
258 	 * and the host side is seriously broken.
259 	 */
260 	if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
261 		panic("Host injected async #PF in interrupt disabled region\n");
262 
263 	if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
264 		if (unlikely(!(user_mode(regs))))
265 			panic("Host injected async #PF in kernel mode\n");
266 		/* Page is swapped out by the host. */
267 		kvm_async_pf_task_wait_schedule(token);
268 	} else {
269 		WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
270 	}
271 
272 	instrumentation_end();
273 	irqentry_exit(regs, state);
274 	return true;
275 }
276 
277 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
278 {
279 	struct pt_regs *old_regs = set_irq_regs(regs);
280 	u32 token;
281 
282 	ack_APIC_irq();
283 
284 	inc_irq_stat(irq_hv_callback_count);
285 
286 	if (__this_cpu_read(apf_reason.enabled)) {
287 		token = __this_cpu_read(apf_reason.token);
288 		kvm_async_pf_task_wake(token);
289 		__this_cpu_write(apf_reason.token, 0);
290 		wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
291 	}
292 
293 	set_irq_regs(old_regs);
294 }
295 
296 static void __init paravirt_ops_setup(void)
297 {
298 	pv_info.name = "KVM";
299 
300 	if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
301 		pv_ops.cpu.io_delay = kvm_io_delay;
302 
303 #ifdef CONFIG_X86_IO_APIC
304 	no_timer_check = 1;
305 #endif
306 }
307 
308 static void kvm_register_steal_time(void)
309 {
310 	int cpu = smp_processor_id();
311 	struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
312 
313 	if (!has_steal_clock)
314 		return;
315 
316 	wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
317 	pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
318 		(unsigned long long) slow_virt_to_phys(st));
319 }
320 
321 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
322 
323 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
324 {
325 	/**
326 	 * This relies on __test_and_clear_bit to modify the memory
327 	 * in a way that is atomic with respect to the local CPU.
328 	 * The hypervisor only accesses this memory from the local CPU so
329 	 * there's no need for lock or memory barriers.
330 	 * An optimization barrier is implied in apic write.
331 	 */
332 	if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
333 		return;
334 	apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
335 }
336 
337 static void kvm_guest_cpu_init(void)
338 {
339 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
340 		u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
341 
342 		WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
343 
344 		pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
345 		pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
346 
347 		if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
348 			pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
349 
350 		wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
351 
352 		wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
353 		__this_cpu_write(apf_reason.enabled, 1);
354 		pr_debug("setup async PF for cpu %d\n", smp_processor_id());
355 	}
356 
357 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
358 		unsigned long pa;
359 
360 		/* Size alignment is implied but just to make it explicit. */
361 		BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
362 		__this_cpu_write(kvm_apic_eoi, 0);
363 		pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
364 			| KVM_MSR_ENABLED;
365 		wrmsrl(MSR_KVM_PV_EOI_EN, pa);
366 	}
367 
368 	if (has_steal_clock)
369 		kvm_register_steal_time();
370 }
371 
372 static void kvm_pv_disable_apf(void)
373 {
374 	if (!__this_cpu_read(apf_reason.enabled))
375 		return;
376 
377 	wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
378 	__this_cpu_write(apf_reason.enabled, 0);
379 
380 	pr_debug("disable async PF for cpu %d\n", smp_processor_id());
381 }
382 
383 static void kvm_disable_steal_time(void)
384 {
385 	if (!has_steal_clock)
386 		return;
387 
388 	wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
389 }
390 
391 static u64 kvm_steal_clock(int cpu)
392 {
393 	u64 steal;
394 	struct kvm_steal_time *src;
395 	int version;
396 
397 	src = &per_cpu(steal_time, cpu);
398 	do {
399 		version = src->version;
400 		virt_rmb();
401 		steal = src->steal;
402 		virt_rmb();
403 	} while ((version & 1) || (version != src->version));
404 
405 	return steal;
406 }
407 
408 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
409 {
410 	early_set_memory_decrypted((unsigned long) ptr, size);
411 }
412 
413 /*
414  * Iterate through all possible CPUs and map the memory region pointed
415  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
416  *
417  * Note: we iterate through all possible CPUs to ensure that CPUs
418  * hotplugged will have their per-cpu variable already mapped as
419  * decrypted.
420  */
421 static void __init sev_map_percpu_data(void)
422 {
423 	int cpu;
424 
425 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
426 		return;
427 
428 	for_each_possible_cpu(cpu) {
429 		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
430 		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
431 		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
432 	}
433 }
434 
435 static void kvm_guest_cpu_offline(bool shutdown)
436 {
437 	kvm_disable_steal_time();
438 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
439 		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
440 	if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
441 		wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
442 	kvm_pv_disable_apf();
443 	if (!shutdown)
444 		apf_task_wake_all();
445 	kvmclock_disable();
446 }
447 
448 static int kvm_cpu_online(unsigned int cpu)
449 {
450 	unsigned long flags;
451 
452 	local_irq_save(flags);
453 	kvm_guest_cpu_init();
454 	local_irq_restore(flags);
455 	return 0;
456 }
457 
458 #ifdef CONFIG_SMP
459 
460 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
461 
462 static bool pv_tlb_flush_supported(void)
463 {
464 	return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
465 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
466 		kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
467 		!boot_cpu_has(X86_FEATURE_MWAIT) &&
468 		(num_possible_cpus() != 1));
469 }
470 
471 static bool pv_ipi_supported(void)
472 {
473 	return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
474 	       (num_possible_cpus() != 1));
475 }
476 
477 static bool pv_sched_yield_supported(void)
478 {
479 	return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
480 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
481 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
482 	    !boot_cpu_has(X86_FEATURE_MWAIT) &&
483 	    (num_possible_cpus() != 1));
484 }
485 
486 #define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
487 
488 static void __send_ipi_mask(const struct cpumask *mask, int vector)
489 {
490 	unsigned long flags;
491 	int cpu, apic_id, icr;
492 	int min = 0, max = 0;
493 #ifdef CONFIG_X86_64
494 	__uint128_t ipi_bitmap = 0;
495 #else
496 	u64 ipi_bitmap = 0;
497 #endif
498 	long ret;
499 
500 	if (cpumask_empty(mask))
501 		return;
502 
503 	local_irq_save(flags);
504 
505 	switch (vector) {
506 	default:
507 		icr = APIC_DM_FIXED | vector;
508 		break;
509 	case NMI_VECTOR:
510 		icr = APIC_DM_NMI;
511 		break;
512 	}
513 
514 	for_each_cpu(cpu, mask) {
515 		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
516 		if (!ipi_bitmap) {
517 			min = max = apic_id;
518 		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
519 			ipi_bitmap <<= min - apic_id;
520 			min = apic_id;
521 		} else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
522 			max = apic_id < max ? max : apic_id;
523 		} else {
524 			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
525 				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
526 			WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
527 				  ret);
528 			min = max = apic_id;
529 			ipi_bitmap = 0;
530 		}
531 		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
532 	}
533 
534 	if (ipi_bitmap) {
535 		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
536 			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
537 		WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
538 			  ret);
539 	}
540 
541 	local_irq_restore(flags);
542 }
543 
544 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
545 {
546 	__send_ipi_mask(mask, vector);
547 }
548 
549 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
550 {
551 	unsigned int this_cpu = smp_processor_id();
552 	struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
553 	const struct cpumask *local_mask;
554 
555 	cpumask_copy(new_mask, mask);
556 	cpumask_clear_cpu(this_cpu, new_mask);
557 	local_mask = new_mask;
558 	__send_ipi_mask(local_mask, vector);
559 }
560 
561 static int __init setup_efi_kvm_sev_migration(void)
562 {
563 	efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
564 	efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
565 	efi_status_t status;
566 	unsigned long size;
567 	bool enabled;
568 
569 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
570 	    !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
571 		return 0;
572 
573 	if (!efi_enabled(EFI_BOOT))
574 		return 0;
575 
576 	if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
577 		pr_info("%s : EFI runtime services are not enabled\n", __func__);
578 		return 0;
579 	}
580 
581 	size = sizeof(enabled);
582 
583 	/* Get variable contents into buffer */
584 	status = efi.get_variable(efi_sev_live_migration_enabled,
585 				  &efi_variable_guid, NULL, &size, &enabled);
586 
587 	if (status == EFI_NOT_FOUND) {
588 		pr_info("%s : EFI live migration variable not found\n", __func__);
589 		return 0;
590 	}
591 
592 	if (status != EFI_SUCCESS) {
593 		pr_info("%s : EFI variable retrieval failed\n", __func__);
594 		return 0;
595 	}
596 
597 	if (enabled == 0) {
598 		pr_info("%s: live migration disabled in EFI\n", __func__);
599 		return 0;
600 	}
601 
602 	pr_info("%s : live migration enabled in EFI\n", __func__);
603 	wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
604 
605 	return 1;
606 }
607 
608 late_initcall(setup_efi_kvm_sev_migration);
609 
610 /*
611  * Set the IPI entry points
612  */
613 static void kvm_setup_pv_ipi(void)
614 {
615 	apic->send_IPI_mask = kvm_send_ipi_mask;
616 	apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
617 	pr_info("setup PV IPIs\n");
618 }
619 
620 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
621 {
622 	int cpu;
623 
624 	native_send_call_func_ipi(mask);
625 
626 	/* Make sure other vCPUs get a chance to run if they need to. */
627 	for_each_cpu(cpu, mask) {
628 		if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
629 			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
630 			break;
631 		}
632 	}
633 }
634 
635 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
636 			const struct flush_tlb_info *info)
637 {
638 	u8 state;
639 	int cpu;
640 	struct kvm_steal_time *src;
641 	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
642 
643 	cpumask_copy(flushmask, cpumask);
644 	/*
645 	 * We have to call flush only on online vCPUs. And
646 	 * queue flush_on_enter for pre-empted vCPUs
647 	 */
648 	for_each_cpu(cpu, flushmask) {
649 		/*
650 		 * The local vCPU is never preempted, so we do not explicitly
651 		 * skip check for local vCPU - it will never be cleared from
652 		 * flushmask.
653 		 */
654 		src = &per_cpu(steal_time, cpu);
655 		state = READ_ONCE(src->preempted);
656 		if ((state & KVM_VCPU_PREEMPTED)) {
657 			if (try_cmpxchg(&src->preempted, &state,
658 					state | KVM_VCPU_FLUSH_TLB))
659 				__cpumask_clear_cpu(cpu, flushmask);
660 		}
661 	}
662 
663 	native_flush_tlb_multi(flushmask, info);
664 }
665 
666 static __init int kvm_alloc_cpumask(void)
667 {
668 	int cpu;
669 
670 	if (!kvm_para_available() || nopv)
671 		return 0;
672 
673 	if (pv_tlb_flush_supported() || pv_ipi_supported())
674 		for_each_possible_cpu(cpu) {
675 			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
676 				GFP_KERNEL, cpu_to_node(cpu));
677 		}
678 
679 	return 0;
680 }
681 arch_initcall(kvm_alloc_cpumask);
682 
683 static void __init kvm_smp_prepare_boot_cpu(void)
684 {
685 	/*
686 	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
687 	 * shares the guest physical address with the hypervisor.
688 	 */
689 	sev_map_percpu_data();
690 
691 	kvm_guest_cpu_init();
692 	native_smp_prepare_boot_cpu();
693 	kvm_spinlock_init();
694 }
695 
696 static int kvm_cpu_down_prepare(unsigned int cpu)
697 {
698 	unsigned long flags;
699 
700 	local_irq_save(flags);
701 	kvm_guest_cpu_offline(false);
702 	local_irq_restore(flags);
703 	return 0;
704 }
705 
706 #endif
707 
708 static int kvm_suspend(void)
709 {
710 	u64 val = 0;
711 
712 	kvm_guest_cpu_offline(false);
713 
714 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
715 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
716 		rdmsrl(MSR_KVM_POLL_CONTROL, val);
717 	has_guest_poll = !(val & 1);
718 #endif
719 	return 0;
720 }
721 
722 static void kvm_resume(void)
723 {
724 	kvm_cpu_online(raw_smp_processor_id());
725 
726 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
727 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
728 		wrmsrl(MSR_KVM_POLL_CONTROL, 0);
729 #endif
730 }
731 
732 static struct syscore_ops kvm_syscore_ops = {
733 	.suspend	= kvm_suspend,
734 	.resume		= kvm_resume,
735 };
736 
737 static void kvm_pv_guest_cpu_reboot(void *unused)
738 {
739 	kvm_guest_cpu_offline(true);
740 }
741 
742 static int kvm_pv_reboot_notify(struct notifier_block *nb,
743 				unsigned long code, void *unused)
744 {
745 	if (code == SYS_RESTART)
746 		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
747 	return NOTIFY_DONE;
748 }
749 
750 static struct notifier_block kvm_pv_reboot_nb = {
751 	.notifier_call = kvm_pv_reboot_notify,
752 };
753 
754 /*
755  * After a PV feature is registered, the host will keep writing to the
756  * registered memory location. If the guest happens to shutdown, this memory
757  * won't be valid. In cases like kexec, in which you install a new kernel, this
758  * means a random memory location will be kept being written.
759  */
760 #ifdef CONFIG_KEXEC_CORE
761 static void kvm_crash_shutdown(struct pt_regs *regs)
762 {
763 	kvm_guest_cpu_offline(true);
764 	native_machine_crash_shutdown(regs);
765 }
766 #endif
767 
768 static void __init kvm_guest_init(void)
769 {
770 	int i;
771 
772 	paravirt_ops_setup();
773 	register_reboot_notifier(&kvm_pv_reboot_nb);
774 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
775 		raw_spin_lock_init(&async_pf_sleepers[i].lock);
776 
777 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
778 		has_steal_clock = 1;
779 		static_call_update(pv_steal_clock, kvm_steal_clock);
780 	}
781 
782 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
783 		apic_set_eoi_write(kvm_guest_apic_eoi_write);
784 
785 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
786 		static_branch_enable(&kvm_async_pf_enabled);
787 		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
788 	}
789 
790 #ifdef CONFIG_SMP
791 	if (pv_tlb_flush_supported()) {
792 		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
793 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
794 		pr_info("KVM setup pv remote TLB flush\n");
795 	}
796 
797 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
798 	if (pv_sched_yield_supported()) {
799 		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
800 		pr_info("setup PV sched yield\n");
801 	}
802 	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
803 				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
804 		pr_err("failed to install cpu hotplug callbacks\n");
805 #else
806 	sev_map_percpu_data();
807 	kvm_guest_cpu_init();
808 #endif
809 
810 #ifdef CONFIG_KEXEC_CORE
811 	machine_ops.crash_shutdown = kvm_crash_shutdown;
812 #endif
813 
814 	register_syscore_ops(&kvm_syscore_ops);
815 
816 	/*
817 	 * Hard lockup detection is enabled by default. Disable it, as guests
818 	 * can get false positives too easily, for example if the host is
819 	 * overcommitted.
820 	 */
821 	hardlockup_detector_disable();
822 }
823 
824 static noinline uint32_t __kvm_cpuid_base(void)
825 {
826 	if (boot_cpu_data.cpuid_level < 0)
827 		return 0;	/* So we don't blow up on old processors */
828 
829 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
830 		return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
831 
832 	return 0;
833 }
834 
835 static inline uint32_t kvm_cpuid_base(void)
836 {
837 	static int kvm_cpuid_base = -1;
838 
839 	if (kvm_cpuid_base == -1)
840 		kvm_cpuid_base = __kvm_cpuid_base();
841 
842 	return kvm_cpuid_base;
843 }
844 
845 bool kvm_para_available(void)
846 {
847 	return kvm_cpuid_base() != 0;
848 }
849 EXPORT_SYMBOL_GPL(kvm_para_available);
850 
851 unsigned int kvm_arch_para_features(void)
852 {
853 	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
854 }
855 
856 unsigned int kvm_arch_para_hints(void)
857 {
858 	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
859 }
860 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
861 
862 static uint32_t __init kvm_detect(void)
863 {
864 	return kvm_cpuid_base();
865 }
866 
867 static void __init kvm_apic_init(void)
868 {
869 #ifdef CONFIG_SMP
870 	if (pv_ipi_supported())
871 		kvm_setup_pv_ipi();
872 #endif
873 }
874 
875 static bool __init kvm_msi_ext_dest_id(void)
876 {
877 	return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
878 }
879 
880 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
881 {
882 	kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
883 			   KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
884 }
885 
886 static void __init kvm_init_platform(void)
887 {
888 	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
889 	    kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
890 		unsigned long nr_pages;
891 		int i;
892 
893 		pv_ops.mmu.notify_page_enc_status_changed =
894 			kvm_sev_hc_page_enc_status;
895 
896 		/*
897 		 * Reset the host's shared pages list related to kernel
898 		 * specific page encryption status settings before we load a
899 		 * new kernel by kexec. Reset the page encryption status
900 		 * during early boot intead of just before kexec to avoid SMP
901 		 * races during kvm_pv_guest_cpu_reboot().
902 		 * NOTE: We cannot reset the complete shared pages list
903 		 * here as we need to retain the UEFI/OVMF firmware
904 		 * specific settings.
905 		 */
906 
907 		for (i = 0; i < e820_table->nr_entries; i++) {
908 			struct e820_entry *entry = &e820_table->entries[i];
909 
910 			if (entry->type != E820_TYPE_RAM)
911 				continue;
912 
913 			nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
914 
915 			kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
916 				       nr_pages,
917 				       KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
918 		}
919 
920 		/*
921 		 * Ensure that _bss_decrypted section is marked as decrypted in the
922 		 * shared pages list.
923 		 */
924 		nr_pages = DIV_ROUND_UP(__end_bss_decrypted - __start_bss_decrypted,
925 					PAGE_SIZE);
926 		early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
927 						nr_pages, 0);
928 
929 		/*
930 		 * If not booted using EFI, enable Live migration support.
931 		 */
932 		if (!efi_enabled(EFI_BOOT))
933 			wrmsrl(MSR_KVM_MIGRATION_CONTROL,
934 			       KVM_MIGRATION_READY);
935 	}
936 	kvmclock_init();
937 	x86_platform.apic_post_init = kvm_apic_init;
938 }
939 
940 #if defined(CONFIG_AMD_MEM_ENCRYPT)
941 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
942 {
943 	/* RAX and CPL are already in the GHCB */
944 	ghcb_set_rbx(ghcb, regs->bx);
945 	ghcb_set_rcx(ghcb, regs->cx);
946 	ghcb_set_rdx(ghcb, regs->dx);
947 	ghcb_set_rsi(ghcb, regs->si);
948 }
949 
950 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
951 {
952 	/* No checking of the return state needed */
953 	return true;
954 }
955 #endif
956 
957 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
958 	.name				= "KVM",
959 	.detect				= kvm_detect,
960 	.type				= X86_HYPER_KVM,
961 	.init.guest_late_init		= kvm_guest_init,
962 	.init.x2apic_available		= kvm_para_available,
963 	.init.msi_ext_dest_id		= kvm_msi_ext_dest_id,
964 	.init.init_platform		= kvm_init_platform,
965 #if defined(CONFIG_AMD_MEM_ENCRYPT)
966 	.runtime.sev_es_hcall_prepare	= kvm_sev_es_hcall_prepare,
967 	.runtime.sev_es_hcall_finish	= kvm_sev_es_hcall_finish,
968 #endif
969 };
970 
971 static __init int activate_jump_labels(void)
972 {
973 	if (has_steal_clock) {
974 		static_key_slow_inc(&paravirt_steal_enabled);
975 		if (steal_acc)
976 			static_key_slow_inc(&paravirt_steal_rq_enabled);
977 	}
978 
979 	return 0;
980 }
981 arch_initcall(activate_jump_labels);
982 
983 #ifdef CONFIG_PARAVIRT_SPINLOCKS
984 
985 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
986 static void kvm_kick_cpu(int cpu)
987 {
988 	int apicid;
989 	unsigned long flags = 0;
990 
991 	apicid = per_cpu(x86_cpu_to_apicid, cpu);
992 	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
993 }
994 
995 #include <asm/qspinlock.h>
996 
997 static void kvm_wait(u8 *ptr, u8 val)
998 {
999 	if (in_nmi())
1000 		return;
1001 
1002 	/*
1003 	 * halt until it's our turn and kicked. Note that we do safe halt
1004 	 * for irq enabled case to avoid hang when lock info is overwritten
1005 	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1006 	 */
1007 	if (irqs_disabled()) {
1008 		if (READ_ONCE(*ptr) == val)
1009 			halt();
1010 	} else {
1011 		local_irq_disable();
1012 
1013 		/* safe_halt() will enable IRQ */
1014 		if (READ_ONCE(*ptr) == val)
1015 			safe_halt();
1016 		else
1017 			local_irq_enable();
1018 	}
1019 }
1020 
1021 #ifdef CONFIG_X86_32
1022 __visible bool __kvm_vcpu_is_preempted(long cpu)
1023 {
1024 	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
1025 
1026 	return !!(src->preempted & KVM_VCPU_PREEMPTED);
1027 }
1028 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
1029 
1030 #else
1031 
1032 #include <asm/asm-offsets.h>
1033 
1034 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
1035 
1036 /*
1037  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
1038  * restoring to/from the stack.
1039  */
1040 asm(
1041 ".pushsection .text;"
1042 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
1043 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
1044 "__raw_callee_save___kvm_vcpu_is_preempted:"
1045 ASM_ENDBR
1046 "movq	__per_cpu_offset(,%rdi,8), %rax;"
1047 "cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
1048 "setne	%al;"
1049 ASM_RET
1050 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
1051 ".popsection");
1052 
1053 #endif
1054 
1055 /*
1056  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1057  */
1058 void __init kvm_spinlock_init(void)
1059 {
1060 	/*
1061 	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1062 	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1063 	 * preferred over native qspinlock when vCPU is preempted.
1064 	 */
1065 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1066 		pr_info("PV spinlocks disabled, no host support\n");
1067 		return;
1068 	}
1069 
1070 	/*
1071 	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1072 	 * are available.
1073 	 */
1074 	if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1075 		pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1076 		goto out;
1077 	}
1078 
1079 	if (num_possible_cpus() == 1) {
1080 		pr_info("PV spinlocks disabled, single CPU\n");
1081 		goto out;
1082 	}
1083 
1084 	if (nopvspin) {
1085 		pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1086 		goto out;
1087 	}
1088 
1089 	pr_info("PV spinlocks enabled\n");
1090 
1091 	__pv_init_lock_hash();
1092 	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1093 	pv_ops.lock.queued_spin_unlock =
1094 		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1095 	pv_ops.lock.wait = kvm_wait;
1096 	pv_ops.lock.kick = kvm_kick_cpu;
1097 
1098 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
1099 		pv_ops.lock.vcpu_is_preempted =
1100 			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
1101 	}
1102 	/*
1103 	 * When PV spinlock is enabled which is preferred over
1104 	 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1105 	 * Just disable it anyway.
1106 	 */
1107 out:
1108 	static_branch_disable(&virt_spin_lock_key);
1109 }
1110 
1111 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
1112 
1113 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1114 
1115 static void kvm_disable_host_haltpoll(void *i)
1116 {
1117 	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1118 }
1119 
1120 static void kvm_enable_host_haltpoll(void *i)
1121 {
1122 	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1123 }
1124 
1125 void arch_haltpoll_enable(unsigned int cpu)
1126 {
1127 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1128 		pr_err_once("host does not support poll control\n");
1129 		pr_err_once("host upgrade recommended\n");
1130 		return;
1131 	}
1132 
1133 	/* Enable guest halt poll disables host halt poll */
1134 	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1135 }
1136 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1137 
1138 void arch_haltpoll_disable(unsigned int cpu)
1139 {
1140 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1141 		return;
1142 
1143 	/* Disable guest halt poll enables host halt poll */
1144 	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1145 }
1146 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1147 #endif
1148