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
47DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
48
49static int kvmapf = 1;
50
51static int __init parse_no_kvmapf(char *arg)
52{
53 kvmapf = 0;
54 return 0;
55}
56
57early_param("no-kvmapf", parse_no_kvmapf);
58
59static int steal_acc = 1;
60static int __init parse_no_stealacc(char *arg)
61{
62 steal_acc = 0;
63 return 0;
64}
65
66early_param("no-steal-acc", parse_no_stealacc);
67
68static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
69DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
70static int has_steal_clock = 0;
71
72static int has_guest_poll = 0;
73/*
74 * No need for any "IO delay" on KVM
75 */
76static 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
83struct kvm_task_sleep_node {
84 struct hlist_node link;
85 struct swait_queue_head wq;
86 u32 token;
87 int cpu;
88};
89
90static struct kvm_task_sleep_head {
91 raw_spinlock_t lock;
92 struct hlist_head list;
93} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
94
95static 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
110static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
111{
112 u32 key = hash_32(val: 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(n: &e->link);
121 raw_spin_unlock(&b->lock);
122 kfree(objp: 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: &n->link, h: &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 */
141void 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: &n))
149 return;
150
151 for (;;) {
152 prepare_to_swait_exclusive(q: &n.wq, wait: &wait, TASK_UNINTERRUPTIBLE);
153 if (hlist_unhashed(h: &n.link))
154 break;
155
156 local_irq_enable();
157 schedule();
158 local_irq_disable();
159 }
160 finish_swait(q: &n.wq, wait: &wait);
161}
162EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
163
164static void apf_task_wake_one(struct kvm_task_sleep_node *n)
165{
166 hlist_del_init(n: &n->link);
167 if (swq_has_sleeper(wq: &n->wq))
168 swake_up_one(q: &n->wq);
169}
170
171static 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
190void kvm_async_pf_task_wake(u32 token)
191{
192 u32 key = hash_32(val: 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
201again:
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(size: 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(n: &dummy->link, h: &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(objp: dummy);
240}
241EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
242
243noinstr 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}
254EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
255
256noinstr 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(fmt: "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(fmt: "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
289DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
290{
291 struct pt_regs *old_regs = set_irq_regs(regs);
292 u32 token;
293
294 apic_eoi();
295
296 inc_irq_stat(irq_hv_callback_count);
297
298 if (__this_cpu_read(apf_reason.enabled)) {
299 token = __this_cpu_read(apf_reason.token);
300 kvm_async_pf_task_wake(token);
301 __this_cpu_write(apf_reason.token, 0);
302 wrmsrl(MSR_KVM_ASYNC_PF_ACK, val: 1);
303 }
304
305 set_irq_regs(old_regs);
306}
307
308static 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
320static 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, val: (slow_virt_to_phys(address: 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
333static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
334
335static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
336{
337 /**
338 * This relies on __test_and_clear_bit to modify the memory
339 * in a way that is atomic with respect to the local CPU.
340 * The hypervisor only accesses this memory from the local CPU so
341 * there's no need for lock or memory barriers.
342 * An optimization barrier is implied in apic write.
343 */
344 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
345 return;
346 apic_native_eoi();
347}
348
349static void kvm_guest_cpu_init(void)
350{
351 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
352 u64 pa;
353
354 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
355
356 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
357 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
358
359 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
360 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
361
362 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
363
364 wrmsrl(MSR_KVM_ASYNC_PF_EN, val: 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, val: pa);
378 }
379
380 if (has_steal_clock)
381 kvm_register_steal_time();
382}
383
384static 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, val: 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
395static 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
403static 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
420static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
421{
422 early_set_memory_decrypted(vaddr: (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 */
433static void __init sev_map_percpu_data(void)
434{
435 int cpu;
436
437 if (!cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT))
438 return;
439
440 for_each_possible_cpu(cpu) {
441 __set_percpu_decrypted(ptr: &per_cpu(apf_reason, cpu), size: sizeof(apf_reason));
442 __set_percpu_decrypted(ptr: &per_cpu(steal_time, cpu), size: sizeof(steal_time));
443 __set_percpu_decrypted(ptr: &per_cpu(kvm_apic_eoi, cpu), size: sizeof(kvm_apic_eoi));
444 }
445}
446
447static 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, val: 0);
452 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
453 wrmsrl(MSR_KVM_MIGRATION_CONTROL, val: 0);
454 kvm_pv_disable_apf();
455 if (!shutdown)
456 apf_task_wake_all();
457 kvmclock_disable();
458}
459
460static 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
472static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
473
474static 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
483static bool pv_ipi_supported(void)
484{
485 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
486 (num_possible_cpus() != 1));
487}
488
489static 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
500static void __send_ipi_mask(const struct cpumask *mask, int vector)
501{
502 unsigned long flags;
503 int cpu, min = 0, max = 0;
504#ifdef CONFIG_X86_64
505 __uint128_t ipi_bitmap = 0;
506#else
507 u64 ipi_bitmap = 0;
508#endif
509 u32 apic_id, icr;
510 long ret;
511
512 if (cpumask_empty(srcp: 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, p1: (unsigned long)ipi_bitmap,
537 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: 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, p1: (unsigned long)ipi_bitmap,
548 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: 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
556static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
557{
558 __send_ipi_mask(mask, vector);
559}
560
561static 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(dstp: new_mask, srcp: mask);
568 cpumask_clear_cpu(cpu: this_cpu, dstp: new_mask);
569 local_mask = new_mask;
570 __send_ipi_mask(mask: local_mask, vector);
571}
572
573static 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(attr: 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
620late_initcall(setup_efi_kvm_sev_migration);
621
622/*
623 * Set the IPI entry points
624 */
625static __init void kvm_setup_pv_ipi(void)
626{
627 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
628 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
629 pr_info("setup PV IPIs\n");
630}
631
632static 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
647static 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(dstp: flushmask, srcp: 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, dstp: flushmask);
672 }
673 }
674
675 native_flush_tlb_multi(cpumask: flushmask, info);
676}
677
678static __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}
693arch_initcall(kvm_alloc_cpumask);
694
695static 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
708static int kvm_cpu_down_prepare(unsigned int cpu)
709{
710 unsigned long flags;
711
712 local_irq_save(flags);
713 kvm_guest_cpu_offline(shutdown: false);
714 local_irq_restore(flags);
715 return 0;
716}
717
718#endif
719
720static int kvm_suspend(void)
721{
722 u64 val = 0;
723
724 kvm_guest_cpu_offline(shutdown: 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
734static 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, val: 0);
741#endif
742}
743
744static struct syscore_ops kvm_syscore_ops = {
745 .suspend = kvm_suspend,
746 .resume = kvm_resume,
747};
748
749static void kvm_pv_guest_cpu_reboot(void *unused)
750{
751 kvm_guest_cpu_offline(shutdown: true);
752}
753
754static 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(func: kvm_pv_guest_cpu_reboot, NULL, wait: 1);
759 return NOTIFY_DONE;
760}
761
762static 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
773static void kvm_crash_shutdown(struct pt_regs *regs)
774{
775 kvm_guest_cpu_offline(shutdown: true);
776 native_machine_crash_shutdown(regs);
777}
778#endif
779
780#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
781bool __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}
789PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
790
791#else
792
793#include <asm/asm-offsets.h>
794
795extern 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#define PV_VCPU_PREEMPTED_ASM \
802 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
803 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
804 "setne %al\n\t"
805
806DEFINE_PARAVIRT_ASM(__raw_callee_save___kvm_vcpu_is_preempted,
807 PV_VCPU_PREEMPTED_ASM, .text);
808#endif
809
810static void __init kvm_guest_init(void)
811{
812 int i;
813
814 paravirt_ops_setup();
815 register_reboot_notifier(&kvm_pv_reboot_nb);
816 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
817 raw_spin_lock_init(&async_pf_sleepers[i].lock);
818
819 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
820 has_steal_clock = 1;
821 static_call_update(pv_steal_clock, kvm_steal_clock);
822
823 pv_ops.lock.vcpu_is_preempted =
824 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
825 }
826
827 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
828 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
829
830 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
831 static_branch_enable(&kvm_async_pf_enabled);
832 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, addr: asm_sysvec_kvm_asyncpf_interrupt);
833 }
834
835#ifdef CONFIG_SMP
836 if (pv_tlb_flush_supported()) {
837 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
838 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
839 pr_info("KVM setup pv remote TLB flush\n");
840 }
841
842 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
843 if (pv_sched_yield_supported()) {
844 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
845 pr_info("setup PV sched yield\n");
846 }
847 if (cpuhp_setup_state_nocalls(state: CPUHP_AP_ONLINE_DYN, name: "x86/kvm:online",
848 startup: kvm_cpu_online, teardown: kvm_cpu_down_prepare) < 0)
849 pr_err("failed to install cpu hotplug callbacks\n");
850#else
851 sev_map_percpu_data();
852 kvm_guest_cpu_init();
853#endif
854
855#ifdef CONFIG_KEXEC_CORE
856 machine_ops.crash_shutdown = kvm_crash_shutdown;
857#endif
858
859 register_syscore_ops(ops: &kvm_syscore_ops);
860
861 /*
862 * Hard lockup detection is enabled by default. Disable it, as guests
863 * can get false positives too easily, for example if the host is
864 * overcommitted.
865 */
866 hardlockup_detector_disable();
867}
868
869static noinline uint32_t __kvm_cpuid_base(void)
870{
871 if (boot_cpu_data.cpuid_level < 0)
872 return 0; /* So we don't blow up on old processors */
873
874 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
875 return hypervisor_cpuid_base(KVM_SIGNATURE, leaves: 0);
876
877 return 0;
878}
879
880static inline uint32_t kvm_cpuid_base(void)
881{
882 static int kvm_cpuid_base = -1;
883
884 if (kvm_cpuid_base == -1)
885 kvm_cpuid_base = __kvm_cpuid_base();
886
887 return kvm_cpuid_base;
888}
889
890bool kvm_para_available(void)
891{
892 return kvm_cpuid_base() != 0;
893}
894EXPORT_SYMBOL_GPL(kvm_para_available);
895
896unsigned int kvm_arch_para_features(void)
897{
898 return cpuid_eax(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
899}
900
901unsigned int kvm_arch_para_hints(void)
902{
903 return cpuid_edx(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
904}
905EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
906
907static uint32_t __init kvm_detect(void)
908{
909 return kvm_cpuid_base();
910}
911
912static void __init kvm_apic_init(void)
913{
914#ifdef CONFIG_SMP
915 if (pv_ipi_supported())
916 kvm_setup_pv_ipi();
917#endif
918}
919
920static bool __init kvm_msi_ext_dest_id(void)
921{
922 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
923}
924
925static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
926{
927 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: pfn << PAGE_SHIFT, p2: npages,
928 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
929}
930
931static void __init kvm_init_platform(void)
932{
933 if (cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT) &&
934 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
935 unsigned long nr_pages;
936 int i;
937
938 pv_ops.mmu.notify_page_enc_status_changed =
939 kvm_sev_hc_page_enc_status;
940
941 /*
942 * Reset the host's shared pages list related to kernel
943 * specific page encryption status settings before we load a
944 * new kernel by kexec. Reset the page encryption status
945 * during early boot intead of just before kexec to avoid SMP
946 * races during kvm_pv_guest_cpu_reboot().
947 * NOTE: We cannot reset the complete shared pages list
948 * here as we need to retain the UEFI/OVMF firmware
949 * specific settings.
950 */
951
952 for (i = 0; i < e820_table->nr_entries; i++) {
953 struct e820_entry *entry = &e820_table->entries[i];
954
955 if (entry->type != E820_TYPE_RAM)
956 continue;
957
958 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
959
960 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: entry->addr,
961 p2: nr_pages,
962 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
963 }
964
965 /*
966 * Ensure that _bss_decrypted section is marked as decrypted in the
967 * shared pages list.
968 */
969 early_set_mem_enc_dec_hypercall(vaddr: (unsigned long)__start_bss_decrypted,
970 size: __end_bss_decrypted - __start_bss_decrypted, enc: 0);
971
972 /*
973 * If not booted using EFI, enable Live migration support.
974 */
975 if (!efi_enabled(EFI_BOOT))
976 wrmsrl(MSR_KVM_MIGRATION_CONTROL,
977 KVM_MIGRATION_READY);
978 }
979 kvmclock_init();
980 x86_platform.apic_post_init = kvm_apic_init;
981}
982
983#if defined(CONFIG_AMD_MEM_ENCRYPT)
984static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
985{
986 /* RAX and CPL are already in the GHCB */
987 ghcb_set_rbx(ghcb, value: regs->bx);
988 ghcb_set_rcx(ghcb, value: regs->cx);
989 ghcb_set_rdx(ghcb, value: regs->dx);
990 ghcb_set_rsi(ghcb, value: regs->si);
991}
992
993static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
994{
995 /* No checking of the return state needed */
996 return true;
997}
998#endif
999
1000const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1001 .name = "KVM",
1002 .detect = kvm_detect,
1003 .type = X86_HYPER_KVM,
1004 .init.guest_late_init = kvm_guest_init,
1005 .init.x2apic_available = kvm_para_available,
1006 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1007 .init.init_platform = kvm_init_platform,
1008#if defined(CONFIG_AMD_MEM_ENCRYPT)
1009 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1010 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1011#endif
1012};
1013
1014static __init int activate_jump_labels(void)
1015{
1016 if (has_steal_clock) {
1017 static_key_slow_inc(key: &paravirt_steal_enabled);
1018 if (steal_acc)
1019 static_key_slow_inc(key: &paravirt_steal_rq_enabled);
1020 }
1021
1022 return 0;
1023}
1024arch_initcall(activate_jump_labels);
1025
1026#ifdef CONFIG_PARAVIRT_SPINLOCKS
1027
1028/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1029static void kvm_kick_cpu(int cpu)
1030{
1031 unsigned long flags = 0;
1032 u32 apicid;
1033
1034 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1035 kvm_hypercall2(KVM_HC_KICK_CPU, p1: flags, p2: apicid);
1036}
1037
1038#include <asm/qspinlock.h>
1039
1040static void kvm_wait(u8 *ptr, u8 val)
1041{
1042 if (in_nmi())
1043 return;
1044
1045 /*
1046 * halt until it's our turn and kicked. Note that we do safe halt
1047 * for irq enabled case to avoid hang when lock info is overwritten
1048 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1049 */
1050 if (irqs_disabled()) {
1051 if (READ_ONCE(*ptr) == val)
1052 halt();
1053 } else {
1054 local_irq_disable();
1055
1056 /* safe_halt() will enable IRQ */
1057 if (READ_ONCE(*ptr) == val)
1058 safe_halt();
1059 else
1060 local_irq_enable();
1061 }
1062}
1063
1064/*
1065 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1066 */
1067void __init kvm_spinlock_init(void)
1068{
1069 /*
1070 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1071 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1072 * preferred over native qspinlock when vCPU is preempted.
1073 */
1074 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1075 pr_info("PV spinlocks disabled, no host support\n");
1076 return;
1077 }
1078
1079 /*
1080 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1081 * are available.
1082 */
1083 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1084 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1085 goto out;
1086 }
1087
1088 if (num_possible_cpus() == 1) {
1089 pr_info("PV spinlocks disabled, single CPU\n");
1090 goto out;
1091 }
1092
1093 if (nopvspin) {
1094 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1095 goto out;
1096 }
1097
1098 pr_info("PV spinlocks enabled\n");
1099
1100 __pv_init_lock_hash();
1101 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1102 pv_ops.lock.queued_spin_unlock =
1103 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1104 pv_ops.lock.wait = kvm_wait;
1105 pv_ops.lock.kick = kvm_kick_cpu;
1106
1107 /*
1108 * When PV spinlock is enabled which is preferred over
1109 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1110 * Just disable it anyway.
1111 */
1112out:
1113 static_branch_disable(&virt_spin_lock_key);
1114}
1115
1116#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1117
1118#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1119
1120static void kvm_disable_host_haltpoll(void *i)
1121{
1122 wrmsrl(MSR_KVM_POLL_CONTROL, val: 0);
1123}
1124
1125static void kvm_enable_host_haltpoll(void *i)
1126{
1127 wrmsrl(MSR_KVM_POLL_CONTROL, val: 1);
1128}
1129
1130void arch_haltpoll_enable(unsigned int cpu)
1131{
1132 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1133 pr_err_once("host does not support poll control\n");
1134 pr_err_once("host upgrade recommended\n");
1135 return;
1136 }
1137
1138 /* Enable guest halt poll disables host halt poll */
1139 smp_call_function_single(cpuid: cpu, func: kvm_disable_host_haltpoll, NULL, wait: 1);
1140}
1141EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1142
1143void arch_haltpoll_disable(unsigned int cpu)
1144{
1145 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1146 return;
1147
1148 /* Disable guest halt poll enables host halt poll */
1149 smp_call_function_single(cpuid: cpu, func: kvm_enable_host_haltpoll, NULL, wait: 1);
1150}
1151EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1152#endif
1153

source code of linux/arch/x86/kernel/kvm.c