xref: /openbmc/linux/arch/x86/kernel/kvm.c (revision 2a9eb57e)
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 	ack_APIC_irq();
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 void kvm_guest_apic_eoi_write(u32 reg, u32 val)
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_write(APIC_EOI, APIC_EOI_ACK);
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 = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
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_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
438 		return;
439 
440 	for_each_possible_cpu(cpu) {
441 		__set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
442 		__set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
443 		__set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
444 	}
445 }
446 
447 static void kvm_guest_cpu_offline(bool shutdown)
448 {
449 	kvm_disable_steal_time();
450 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
451 		wrmsrl(MSR_KVM_PV_EOI_EN, 0);
452 	if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
453 		wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
454 	kvm_pv_disable_apf();
455 	if (!shutdown)
456 		apf_task_wake_all();
457 	kvmclock_disable();
458 }
459 
460 static int kvm_cpu_online(unsigned int cpu)
461 {
462 	unsigned long flags;
463 
464 	local_irq_save(flags);
465 	kvm_guest_cpu_init();
466 	local_irq_restore(flags);
467 	return 0;
468 }
469 
470 #ifdef CONFIG_SMP
471 
472 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
473 
474 static bool pv_tlb_flush_supported(void)
475 {
476 	return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
477 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
478 		kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
479 		!boot_cpu_has(X86_FEATURE_MWAIT) &&
480 		(num_possible_cpus() != 1));
481 }
482 
483 static bool pv_ipi_supported(void)
484 {
485 	return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
486 	       (num_possible_cpus() != 1));
487 }
488 
489 static bool pv_sched_yield_supported(void)
490 {
491 	return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
492 		!kvm_para_has_hint(KVM_HINTS_REALTIME) &&
493 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
494 	    !boot_cpu_has(X86_FEATURE_MWAIT) &&
495 	    (num_possible_cpus() != 1));
496 }
497 
498 #define KVM_IPI_CLUSTER_SIZE	(2 * BITS_PER_LONG)
499 
500 static void __send_ipi_mask(const struct cpumask *mask, int vector)
501 {
502 	unsigned long flags;
503 	int cpu, apic_id, icr;
504 	int 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 	long ret;
511 
512 	if (cpumask_empty(mask))
513 		return;
514 
515 	local_irq_save(flags);
516 
517 	switch (vector) {
518 	default:
519 		icr = APIC_DM_FIXED | vector;
520 		break;
521 	case NMI_VECTOR:
522 		icr = APIC_DM_NMI;
523 		break;
524 	}
525 
526 	for_each_cpu(cpu, mask) {
527 		apic_id = per_cpu(x86_cpu_to_apicid, cpu);
528 		if (!ipi_bitmap) {
529 			min = max = apic_id;
530 		} else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
531 			ipi_bitmap <<= min - apic_id;
532 			min = apic_id;
533 		} else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
534 			max = apic_id < max ? max : apic_id;
535 		} else {
536 			ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
537 				(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
538 			WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
539 				  ret);
540 			min = max = apic_id;
541 			ipi_bitmap = 0;
542 		}
543 		__set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
544 	}
545 
546 	if (ipi_bitmap) {
547 		ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
548 			(unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
549 		WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
550 			  ret);
551 	}
552 
553 	local_irq_restore(flags);
554 }
555 
556 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
557 {
558 	__send_ipi_mask(mask, vector);
559 }
560 
561 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
562 {
563 	unsigned int this_cpu = smp_processor_id();
564 	struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
565 	const struct cpumask *local_mask;
566 
567 	cpumask_copy(new_mask, mask);
568 	cpumask_clear_cpu(this_cpu, new_mask);
569 	local_mask = new_mask;
570 	__send_ipi_mask(local_mask, vector);
571 }
572 
573 static int __init setup_efi_kvm_sev_migration(void)
574 {
575 	efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
576 	efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
577 	efi_status_t status;
578 	unsigned long size;
579 	bool enabled;
580 
581 	if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
582 	    !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
583 		return 0;
584 
585 	if (!efi_enabled(EFI_BOOT))
586 		return 0;
587 
588 	if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
589 		pr_info("%s : EFI runtime services are not enabled\n", __func__);
590 		return 0;
591 	}
592 
593 	size = sizeof(enabled);
594 
595 	/* Get variable contents into buffer */
596 	status = efi.get_variable(efi_sev_live_migration_enabled,
597 				  &efi_variable_guid, NULL, &size, &enabled);
598 
599 	if (status == EFI_NOT_FOUND) {
600 		pr_info("%s : EFI live migration variable not found\n", __func__);
601 		return 0;
602 	}
603 
604 	if (status != EFI_SUCCESS) {
605 		pr_info("%s : EFI variable retrieval failed\n", __func__);
606 		return 0;
607 	}
608 
609 	if (enabled == 0) {
610 		pr_info("%s: live migration disabled in EFI\n", __func__);
611 		return 0;
612 	}
613 
614 	pr_info("%s : live migration enabled in EFI\n", __func__);
615 	wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
616 
617 	return 1;
618 }
619 
620 late_initcall(setup_efi_kvm_sev_migration);
621 
622 /*
623  * Set the IPI entry points
624  */
625 static void kvm_setup_pv_ipi(void)
626 {
627 	apic->send_IPI_mask = kvm_send_ipi_mask;
628 	apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
629 	pr_info("setup PV IPIs\n");
630 }
631 
632 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
633 {
634 	int cpu;
635 
636 	native_send_call_func_ipi(mask);
637 
638 	/* Make sure other vCPUs get a chance to run if they need to. */
639 	for_each_cpu(cpu, mask) {
640 		if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
641 			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
642 			break;
643 		}
644 	}
645 }
646 
647 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
648 			const struct flush_tlb_info *info)
649 {
650 	u8 state;
651 	int cpu;
652 	struct kvm_steal_time *src;
653 	struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
654 
655 	cpumask_copy(flushmask, cpumask);
656 	/*
657 	 * We have to call flush only on online vCPUs. And
658 	 * queue flush_on_enter for pre-empted vCPUs
659 	 */
660 	for_each_cpu(cpu, flushmask) {
661 		/*
662 		 * The local vCPU is never preempted, so we do not explicitly
663 		 * skip check for local vCPU - it will never be cleared from
664 		 * flushmask.
665 		 */
666 		src = &per_cpu(steal_time, cpu);
667 		state = READ_ONCE(src->preempted);
668 		if ((state & KVM_VCPU_PREEMPTED)) {
669 			if (try_cmpxchg(&src->preempted, &state,
670 					state | KVM_VCPU_FLUSH_TLB))
671 				__cpumask_clear_cpu(cpu, flushmask);
672 		}
673 	}
674 
675 	native_flush_tlb_multi(flushmask, info);
676 }
677 
678 static __init int kvm_alloc_cpumask(void)
679 {
680 	int cpu;
681 
682 	if (!kvm_para_available() || nopv)
683 		return 0;
684 
685 	if (pv_tlb_flush_supported() || pv_ipi_supported())
686 		for_each_possible_cpu(cpu) {
687 			zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
688 				GFP_KERNEL, cpu_to_node(cpu));
689 		}
690 
691 	return 0;
692 }
693 arch_initcall(kvm_alloc_cpumask);
694 
695 static void __init kvm_smp_prepare_boot_cpu(void)
696 {
697 	/*
698 	 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
699 	 * shares the guest physical address with the hypervisor.
700 	 */
701 	sev_map_percpu_data();
702 
703 	kvm_guest_cpu_init();
704 	native_smp_prepare_boot_cpu();
705 	kvm_spinlock_init();
706 }
707 
708 static int kvm_cpu_down_prepare(unsigned int cpu)
709 {
710 	unsigned long flags;
711 
712 	local_irq_save(flags);
713 	kvm_guest_cpu_offline(false);
714 	local_irq_restore(flags);
715 	return 0;
716 }
717 
718 #endif
719 
720 static int kvm_suspend(void)
721 {
722 	u64 val = 0;
723 
724 	kvm_guest_cpu_offline(false);
725 
726 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
727 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
728 		rdmsrl(MSR_KVM_POLL_CONTROL, val);
729 	has_guest_poll = !(val & 1);
730 #endif
731 	return 0;
732 }
733 
734 static void kvm_resume(void)
735 {
736 	kvm_cpu_online(raw_smp_processor_id());
737 
738 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
739 	if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
740 		wrmsrl(MSR_KVM_POLL_CONTROL, 0);
741 #endif
742 }
743 
744 static struct syscore_ops kvm_syscore_ops = {
745 	.suspend	= kvm_suspend,
746 	.resume		= kvm_resume,
747 };
748 
749 static void kvm_pv_guest_cpu_reboot(void *unused)
750 {
751 	kvm_guest_cpu_offline(true);
752 }
753 
754 static int kvm_pv_reboot_notify(struct notifier_block *nb,
755 				unsigned long code, void *unused)
756 {
757 	if (code == SYS_RESTART)
758 		on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
759 	return NOTIFY_DONE;
760 }
761 
762 static struct notifier_block kvm_pv_reboot_nb = {
763 	.notifier_call = kvm_pv_reboot_notify,
764 };
765 
766 /*
767  * After a PV feature is registered, the host will keep writing to the
768  * registered memory location. If the guest happens to shutdown, this memory
769  * won't be valid. In cases like kexec, in which you install a new kernel, this
770  * means a random memory location will be kept being written.
771  */
772 #ifdef CONFIG_KEXEC_CORE
773 static void kvm_crash_shutdown(struct pt_regs *regs)
774 {
775 	kvm_guest_cpu_offline(true);
776 	native_machine_crash_shutdown(regs);
777 }
778 #endif
779 
780 #if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
781 bool __kvm_vcpu_is_preempted(long cpu);
782 
783 __visible bool __kvm_vcpu_is_preempted(long cpu)
784 {
785 	struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
786 
787 	return !!(src->preempted & KVM_VCPU_PREEMPTED);
788 }
789 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
790 
791 #else
792 
793 #include <asm/asm-offsets.h>
794 
795 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
796 
797 /*
798  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
799  * restoring to/from the stack.
800  */
801 asm(
802 ".pushsection .text;"
803 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
804 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
805 "__raw_callee_save___kvm_vcpu_is_preempted:"
806 ASM_ENDBR
807 "movq	__per_cpu_offset(,%rdi,8), %rax;"
808 "cmpb	$0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
809 "setne	%al;"
810 ASM_RET
811 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
812 ".popsection");
813 
814 #endif
815 
816 static void __init kvm_guest_init(void)
817 {
818 	int i;
819 
820 	paravirt_ops_setup();
821 	register_reboot_notifier(&kvm_pv_reboot_nb);
822 	for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
823 		raw_spin_lock_init(&async_pf_sleepers[i].lock);
824 
825 	if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
826 		has_steal_clock = 1;
827 		static_call_update(pv_steal_clock, kvm_steal_clock);
828 
829 		pv_ops.lock.vcpu_is_preempted =
830 			PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
831 	}
832 
833 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
834 		apic_set_eoi_write(kvm_guest_apic_eoi_write);
835 
836 	if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
837 		static_branch_enable(&kvm_async_pf_enabled);
838 		alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
839 	}
840 
841 #ifdef CONFIG_SMP
842 	if (pv_tlb_flush_supported()) {
843 		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
844 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
845 		pr_info("KVM setup pv remote TLB flush\n");
846 	}
847 
848 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
849 	if (pv_sched_yield_supported()) {
850 		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
851 		pr_info("setup PV sched yield\n");
852 	}
853 	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
854 				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
855 		pr_err("failed to install cpu hotplug callbacks\n");
856 #else
857 	sev_map_percpu_data();
858 	kvm_guest_cpu_init();
859 #endif
860 
861 #ifdef CONFIG_KEXEC_CORE
862 	machine_ops.crash_shutdown = kvm_crash_shutdown;
863 #endif
864 
865 	register_syscore_ops(&kvm_syscore_ops);
866 
867 	/*
868 	 * Hard lockup detection is enabled by default. Disable it, as guests
869 	 * can get false positives too easily, for example if the host is
870 	 * overcommitted.
871 	 */
872 	hardlockup_detector_disable();
873 }
874 
875 static noinline uint32_t __kvm_cpuid_base(void)
876 {
877 	if (boot_cpu_data.cpuid_level < 0)
878 		return 0;	/* So we don't blow up on old processors */
879 
880 	if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
881 		return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
882 
883 	return 0;
884 }
885 
886 static inline uint32_t kvm_cpuid_base(void)
887 {
888 	static int kvm_cpuid_base = -1;
889 
890 	if (kvm_cpuid_base == -1)
891 		kvm_cpuid_base = __kvm_cpuid_base();
892 
893 	return kvm_cpuid_base;
894 }
895 
896 bool kvm_para_available(void)
897 {
898 	return kvm_cpuid_base() != 0;
899 }
900 EXPORT_SYMBOL_GPL(kvm_para_available);
901 
902 unsigned int kvm_arch_para_features(void)
903 {
904 	return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
905 }
906 
907 unsigned int kvm_arch_para_hints(void)
908 {
909 	return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
910 }
911 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
912 
913 static uint32_t __init kvm_detect(void)
914 {
915 	return kvm_cpuid_base();
916 }
917 
918 static void __init kvm_apic_init(void)
919 {
920 #ifdef CONFIG_SMP
921 	if (pv_ipi_supported())
922 		kvm_setup_pv_ipi();
923 #endif
924 }
925 
926 static bool __init kvm_msi_ext_dest_id(void)
927 {
928 	return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
929 }
930 
931 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
932 {
933 	kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
934 			   KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
935 }
936 
937 static void __init kvm_init_platform(void)
938 {
939 	if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
940 	    kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
941 		unsigned long nr_pages;
942 		int i;
943 
944 		pv_ops.mmu.notify_page_enc_status_changed =
945 			kvm_sev_hc_page_enc_status;
946 
947 		/*
948 		 * Reset the host's shared pages list related to kernel
949 		 * specific page encryption status settings before we load a
950 		 * new kernel by kexec. Reset the page encryption status
951 		 * during early boot intead of just before kexec to avoid SMP
952 		 * races during kvm_pv_guest_cpu_reboot().
953 		 * NOTE: We cannot reset the complete shared pages list
954 		 * here as we need to retain the UEFI/OVMF firmware
955 		 * specific settings.
956 		 */
957 
958 		for (i = 0; i < e820_table->nr_entries; i++) {
959 			struct e820_entry *entry = &e820_table->entries[i];
960 
961 			if (entry->type != E820_TYPE_RAM)
962 				continue;
963 
964 			nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
965 
966 			kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
967 				       nr_pages,
968 				       KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
969 		}
970 
971 		/*
972 		 * Ensure that _bss_decrypted section is marked as decrypted in the
973 		 * shared pages list.
974 		 */
975 		nr_pages = DIV_ROUND_UP(__end_bss_decrypted - __start_bss_decrypted,
976 					PAGE_SIZE);
977 		early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
978 						nr_pages, 0);
979 
980 		/*
981 		 * If not booted using EFI, enable Live migration support.
982 		 */
983 		if (!efi_enabled(EFI_BOOT))
984 			wrmsrl(MSR_KVM_MIGRATION_CONTROL,
985 			       KVM_MIGRATION_READY);
986 	}
987 	kvmclock_init();
988 	x86_platform.apic_post_init = kvm_apic_init;
989 }
990 
991 #if defined(CONFIG_AMD_MEM_ENCRYPT)
992 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
993 {
994 	/* RAX and CPL are already in the GHCB */
995 	ghcb_set_rbx(ghcb, regs->bx);
996 	ghcb_set_rcx(ghcb, regs->cx);
997 	ghcb_set_rdx(ghcb, regs->dx);
998 	ghcb_set_rsi(ghcb, regs->si);
999 }
1000 
1001 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
1002 {
1003 	/* No checking of the return state needed */
1004 	return true;
1005 }
1006 #endif
1007 
1008 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1009 	.name				= "KVM",
1010 	.detect				= kvm_detect,
1011 	.type				= X86_HYPER_KVM,
1012 	.init.guest_late_init		= kvm_guest_init,
1013 	.init.x2apic_available		= kvm_para_available,
1014 	.init.msi_ext_dest_id		= kvm_msi_ext_dest_id,
1015 	.init.init_platform		= kvm_init_platform,
1016 #if defined(CONFIG_AMD_MEM_ENCRYPT)
1017 	.runtime.sev_es_hcall_prepare	= kvm_sev_es_hcall_prepare,
1018 	.runtime.sev_es_hcall_finish	= kvm_sev_es_hcall_finish,
1019 #endif
1020 };
1021 
1022 static __init int activate_jump_labels(void)
1023 {
1024 	if (has_steal_clock) {
1025 		static_key_slow_inc(&paravirt_steal_enabled);
1026 		if (steal_acc)
1027 			static_key_slow_inc(&paravirt_steal_rq_enabled);
1028 	}
1029 
1030 	return 0;
1031 }
1032 arch_initcall(activate_jump_labels);
1033 
1034 #ifdef CONFIG_PARAVIRT_SPINLOCKS
1035 
1036 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1037 static void kvm_kick_cpu(int cpu)
1038 {
1039 	int apicid;
1040 	unsigned long flags = 0;
1041 
1042 	apicid = per_cpu(x86_cpu_to_apicid, cpu);
1043 	kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1044 }
1045 
1046 #include <asm/qspinlock.h>
1047 
1048 static void kvm_wait(u8 *ptr, u8 val)
1049 {
1050 	if (in_nmi())
1051 		return;
1052 
1053 	/*
1054 	 * halt until it's our turn and kicked. Note that we do safe halt
1055 	 * for irq enabled case to avoid hang when lock info is overwritten
1056 	 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1057 	 */
1058 	if (irqs_disabled()) {
1059 		if (READ_ONCE(*ptr) == val)
1060 			halt();
1061 	} else {
1062 		local_irq_disable();
1063 
1064 		/* safe_halt() will enable IRQ */
1065 		if (READ_ONCE(*ptr) == val)
1066 			safe_halt();
1067 		else
1068 			local_irq_enable();
1069 	}
1070 }
1071 
1072 /*
1073  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1074  */
1075 void __init kvm_spinlock_init(void)
1076 {
1077 	/*
1078 	 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1079 	 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1080 	 * preferred over native qspinlock when vCPU is preempted.
1081 	 */
1082 	if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1083 		pr_info("PV spinlocks disabled, no host support\n");
1084 		return;
1085 	}
1086 
1087 	/*
1088 	 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1089 	 * are available.
1090 	 */
1091 	if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1092 		pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1093 		goto out;
1094 	}
1095 
1096 	if (num_possible_cpus() == 1) {
1097 		pr_info("PV spinlocks disabled, single CPU\n");
1098 		goto out;
1099 	}
1100 
1101 	if (nopvspin) {
1102 		pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1103 		goto out;
1104 	}
1105 
1106 	pr_info("PV spinlocks enabled\n");
1107 
1108 	__pv_init_lock_hash();
1109 	pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1110 	pv_ops.lock.queued_spin_unlock =
1111 		PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1112 	pv_ops.lock.wait = kvm_wait;
1113 	pv_ops.lock.kick = kvm_kick_cpu;
1114 
1115 	/*
1116 	 * When PV spinlock is enabled which is preferred over
1117 	 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1118 	 * Just disable it anyway.
1119 	 */
1120 out:
1121 	static_branch_disable(&virt_spin_lock_key);
1122 }
1123 
1124 #endif	/* CONFIG_PARAVIRT_SPINLOCKS */
1125 
1126 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1127 
1128 static void kvm_disable_host_haltpoll(void *i)
1129 {
1130 	wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1131 }
1132 
1133 static void kvm_enable_host_haltpoll(void *i)
1134 {
1135 	wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1136 }
1137 
1138 void arch_haltpoll_enable(unsigned int cpu)
1139 {
1140 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1141 		pr_err_once("host does not support poll control\n");
1142 		pr_err_once("host upgrade recommended\n");
1143 		return;
1144 	}
1145 
1146 	/* Enable guest halt poll disables host halt poll */
1147 	smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1148 }
1149 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1150 
1151 void arch_haltpoll_disable(unsigned int cpu)
1152 {
1153 	if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1154 		return;
1155 
1156 	/* Disable guest halt poll enables host halt poll */
1157 	smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1158 }
1159 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1160 #endif
1161