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/mtrr.h>
41#include <asm/tlb.h>
42#include <asm/cpuidle_haltpoll.h>
43#include <asm/msr.h>
44#include <asm/ptrace.h>
45#include <asm/reboot.h>
46#include <asm/svm.h>
47#include <asm/e820/api.h>
48
49DEFINE_STATIC_KEY_FALSE_RO(kvm_async_pf_enabled);
50
51static int kvmapf = 1;
52
53static int __init parse_no_kvmapf(char *arg)
54{
55 kvmapf = 0;
56 return 0;
57}
58
59early_param("no-kvmapf", parse_no_kvmapf);
60
61static int steal_acc = 1;
62static int __init parse_no_stealacc(char *arg)
63{
64 steal_acc = 0;
65 return 0;
66}
67
68early_param("no-steal-acc", parse_no_stealacc);
69
70static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
71static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
72DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
73static int has_steal_clock = 0;
74
75static int has_guest_poll = 0;
76/*
77 * No need for any "IO delay" on KVM
78 */
79static void kvm_io_delay(void)
80{
81}
82
83#define KVM_TASK_SLEEP_HASHBITS 8
84#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
85
86struct kvm_task_sleep_node {
87 struct hlist_node link;
88 struct swait_queue_head wq;
89 u32 token;
90 int cpu;
91};
92
93static struct kvm_task_sleep_head {
94 raw_spinlock_t lock;
95 struct hlist_head list;
96} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
97
98static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
99 u32 token)
100{
101 struct hlist_node *p;
102
103 hlist_for_each(p, &b->list) {
104 struct kvm_task_sleep_node *n =
105 hlist_entry(p, typeof(*n), link);
106 if (n->token == token)
107 return n;
108 }
109
110 return NULL;
111}
112
113static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
114{
115 u32 key = hash_32(val: token, KVM_TASK_SLEEP_HASHBITS);
116 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
117 struct kvm_task_sleep_node *e;
118
119 raw_spin_lock(&b->lock);
120 e = _find_apf_task(b, token);
121 if (e) {
122 /* dummy entry exist -> wake up was delivered ahead of PF */
123 hlist_del(n: &e->link);
124 raw_spin_unlock(&b->lock);
125 kfree(objp: e);
126 return false;
127 }
128
129 n->token = token;
130 n->cpu = smp_processor_id();
131 init_swait_queue_head(&n->wq);
132 hlist_add_head(n: &n->link, h: &b->list);
133 raw_spin_unlock(&b->lock);
134 return true;
135}
136
137/*
138 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
139 * @token: Token to identify the sleep node entry
140 *
141 * Invoked from the async pagefault handling code or from the VM exit page
142 * fault handler. In both cases RCU is watching.
143 */
144void kvm_async_pf_task_wait_schedule(u32 token)
145{
146 struct kvm_task_sleep_node n;
147 DECLARE_SWAITQUEUE(wait);
148
149 lockdep_assert_irqs_disabled();
150
151 if (!kvm_async_pf_queue_task(token, n: &n))
152 return;
153
154 for (;;) {
155 prepare_to_swait_exclusive(q: &n.wq, wait: &wait, TASK_UNINTERRUPTIBLE);
156 if (hlist_unhashed(h: &n.link))
157 break;
158
159 local_irq_enable();
160 schedule();
161 local_irq_disable();
162 }
163 finish_swait(q: &n.wq, wait: &wait);
164}
165EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
166
167static void apf_task_wake_one(struct kvm_task_sleep_node *n)
168{
169 hlist_del_init(n: &n->link);
170 if (swq_has_sleeper(wq: &n->wq))
171 swake_up_one(q: &n->wq);
172}
173
174static void apf_task_wake_all(void)
175{
176 int i;
177
178 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
179 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
180 struct kvm_task_sleep_node *n;
181 struct hlist_node *p, *next;
182
183 raw_spin_lock(&b->lock);
184 hlist_for_each_safe(p, next, &b->list) {
185 n = hlist_entry(p, typeof(*n), link);
186 if (n->cpu == smp_processor_id())
187 apf_task_wake_one(n);
188 }
189 raw_spin_unlock(&b->lock);
190 }
191}
192
193static void kvm_async_pf_task_wake(u32 token)
194{
195 u32 key = hash_32(val: token, KVM_TASK_SLEEP_HASHBITS);
196 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
197 struct kvm_task_sleep_node *n, *dummy = NULL;
198
199 if (token == ~0) {
200 apf_task_wake_all();
201 return;
202 }
203
204again:
205 raw_spin_lock(&b->lock);
206 n = _find_apf_task(b, token);
207 if (!n) {
208 /*
209 * Async #PF not yet handled, add a dummy entry for the token.
210 * Allocating the token must be down outside of the raw lock
211 * as the allocator is preemptible on PREEMPT_RT kernels.
212 */
213 if (!dummy) {
214 raw_spin_unlock(&b->lock);
215 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
216
217 /*
218 * Continue looping on allocation failure, eventually
219 * the async #PF will be handled and allocating a new
220 * node will be unnecessary.
221 */
222 if (!dummy)
223 cpu_relax();
224
225 /*
226 * Recheck for async #PF completion before enqueueing
227 * the dummy token to avoid duplicate list entries.
228 */
229 goto again;
230 }
231 dummy->token = token;
232 dummy->cpu = smp_processor_id();
233 init_swait_queue_head(&dummy->wq);
234 hlist_add_head(n: &dummy->link, h: &b->list);
235 dummy = NULL;
236 } else {
237 apf_task_wake_one(n);
238 }
239 raw_spin_unlock(&b->lock);
240
241 /* A dummy token might be allocated and ultimately not used. */
242 kfree(objp: dummy);
243}
244
245noinstr u32 kvm_read_and_reset_apf_flags(void)
246{
247 u32 flags = 0;
248
249 if (__this_cpu_read(async_pf_enabled)) {
250 flags = __this_cpu_read(apf_reason.flags);
251 __this_cpu_write(apf_reason.flags, 0);
252 }
253
254 return flags;
255}
256EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
257
258noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
259{
260 u32 flags = kvm_read_and_reset_apf_flags();
261 irqentry_state_t state;
262
263 if (!flags)
264 return false;
265
266 state = irqentry_enter(regs);
267 instrumentation_begin();
268
269 /*
270 * If the host managed to inject an async #PF into an interrupt
271 * disabled region, then die hard as this is not going to end well
272 * and the host side is seriously broken.
273 */
274 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
275 panic(fmt: "Host injected async #PF in interrupt disabled region\n");
276
277 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
278 if (unlikely(!(user_mode(regs))))
279 panic(fmt: "Host injected async #PF in kernel mode\n");
280 /* Page is swapped out by the host. */
281 kvm_async_pf_task_wait_schedule(token);
282 } else {
283 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
284 }
285
286 instrumentation_end();
287 irqentry_exit(regs, state);
288 return true;
289}
290
291DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
292{
293 struct pt_regs *old_regs = set_irq_regs(regs);
294 u32 token;
295
296 apic_eoi();
297
298 inc_irq_stat(irq_hv_callback_count);
299
300 if (__this_cpu_read(async_pf_enabled)) {
301 token = __this_cpu_read(apf_reason.token);
302 kvm_async_pf_task_wake(token);
303 __this_cpu_write(apf_reason.token, 0);
304 wrmsrq(MSR_KVM_ASYNC_PF_ACK, val: 1);
305 }
306
307 set_irq_regs(old_regs);
308}
309
310static void __init paravirt_ops_setup(void)
311{
312 pv_info.name = "KVM";
313
314 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
315 pv_ops.cpu.io_delay = kvm_io_delay;
316
317#ifdef CONFIG_X86_IO_APIC
318 no_timer_check = 1;
319#endif
320}
321
322static void kvm_register_steal_time(void)
323{
324 int cpu = smp_processor_id();
325 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
326
327 if (!has_steal_clock)
328 return;
329
330 wrmsrq(MSR_KVM_STEAL_TIME, val: (slow_virt_to_phys(address: st) | KVM_MSR_ENABLED));
331 pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
332 (unsigned long long) slow_virt_to_phys(st));
333}
334
335static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
336
337static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
338{
339 /**
340 * This relies on __test_and_clear_bit to modify the memory
341 * in a way that is atomic with respect to the local CPU.
342 * The hypervisor only accesses this memory from the local CPU so
343 * there's no need for lock or memory barriers.
344 * An optimization barrier is implied in apic write.
345 */
346 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
347 return;
348 apic_native_eoi();
349}
350
351static void kvm_guest_cpu_init(void)
352{
353 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
354 u64 pa;
355
356 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
357
358 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
359 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
360
361 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
362 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
363
364 wrmsrq(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
365
366 wrmsrq(MSR_KVM_ASYNC_PF_EN, val: pa);
367 __this_cpu_write(async_pf_enabled, true);
368 pr_debug("setup async PF for cpu %d\n", smp_processor_id());
369 }
370
371 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
372 unsigned long pa;
373
374 /* Size alignment is implied but just to make it explicit. */
375 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
376 __this_cpu_write(kvm_apic_eoi, 0);
377 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
378 | KVM_MSR_ENABLED;
379 wrmsrq(MSR_KVM_PV_EOI_EN, val: pa);
380 }
381
382 if (has_steal_clock)
383 kvm_register_steal_time();
384}
385
386static void kvm_pv_disable_apf(void)
387{
388 if (!__this_cpu_read(async_pf_enabled))
389 return;
390
391 wrmsrq(MSR_KVM_ASYNC_PF_EN, val: 0);
392 __this_cpu_write(async_pf_enabled, false);
393
394 pr_debug("disable async PF for cpu %d\n", smp_processor_id());
395}
396
397static void kvm_disable_steal_time(void)
398{
399 if (!has_steal_clock)
400 return;
401
402 wrmsrq(MSR_KVM_STEAL_TIME, val: 0);
403}
404
405static u64 kvm_steal_clock(int cpu)
406{
407 u64 steal;
408 struct kvm_steal_time *src;
409 int version;
410
411 src = &per_cpu(steal_time, cpu);
412 do {
413 version = src->version;
414 virt_rmb();
415 steal = src->steal;
416 virt_rmb();
417 } while ((version & 1) || (version != src->version));
418
419 return steal;
420}
421
422static inline __init void __set_percpu_decrypted(void *ptr, unsigned long size)
423{
424 early_set_memory_decrypted(vaddr: (unsigned long) ptr, size);
425}
426
427/*
428 * Iterate through all possible CPUs and map the memory region pointed
429 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
430 *
431 * Note: we iterate through all possible CPUs to ensure that CPUs
432 * hotplugged will have their per-cpu variable already mapped as
433 * decrypted.
434 */
435static void __init sev_map_percpu_data(void)
436{
437 int cpu;
438
439 if (cc_vendor != CC_VENDOR_AMD ||
440 !cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT))
441 return;
442
443 for_each_possible_cpu(cpu) {
444 __set_percpu_decrypted(ptr: &per_cpu(apf_reason, cpu), size: sizeof(apf_reason));
445 __set_percpu_decrypted(ptr: &per_cpu(steal_time, cpu), size: sizeof(steal_time));
446 __set_percpu_decrypted(ptr: &per_cpu(kvm_apic_eoi, cpu), size: sizeof(kvm_apic_eoi));
447 }
448}
449
450static void kvm_guest_cpu_offline(bool shutdown)
451{
452 kvm_disable_steal_time();
453 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
454 wrmsrq(MSR_KVM_PV_EOI_EN, val: 0);
455 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
456 wrmsrq(MSR_KVM_MIGRATION_CONTROL, val: 0);
457 kvm_pv_disable_apf();
458 if (!shutdown)
459 apf_task_wake_all();
460 kvmclock_disable();
461}
462
463static int kvm_cpu_online(unsigned int cpu)
464{
465 unsigned long flags;
466
467 local_irq_save(flags);
468 kvm_guest_cpu_init();
469 local_irq_restore(flags);
470 return 0;
471}
472
473#ifdef CONFIG_SMP
474
475static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
476
477static bool pv_tlb_flush_supported(void)
478{
479 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
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
486static bool pv_ipi_supported(void)
487{
488 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
489 (num_possible_cpus() != 1));
490}
491
492static bool pv_sched_yield_supported(void)
493{
494 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
495 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
496 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
497 !boot_cpu_has(X86_FEATURE_MWAIT) &&
498 (num_possible_cpus() != 1));
499}
500
501#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
502
503static void __send_ipi_mask(const struct cpumask *mask, int vector)
504{
505 unsigned long flags;
506 int cpu, min = 0, max = 0;
507#ifdef CONFIG_X86_64
508 __uint128_t ipi_bitmap = 0;
509#else
510 u64 ipi_bitmap = 0;
511#endif
512 u32 apic_id, icr;
513 long ret;
514
515 if (cpumask_empty(srcp: mask))
516 return;
517
518 local_irq_save(flags);
519
520 switch (vector) {
521 default:
522 icr = APIC_DM_FIXED | vector;
523 break;
524 case NMI_VECTOR:
525 icr = APIC_DM_NMI;
526 break;
527 }
528
529 for_each_cpu(cpu, mask) {
530 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
531 if (!ipi_bitmap) {
532 min = max = apic_id;
533 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
534 ipi_bitmap <<= min - apic_id;
535 min = apic_id;
536 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
537 max = apic_id < max ? max : apic_id;
538 } else {
539 ret = kvm_hypercall4(KVM_HC_SEND_IPI, p1: (unsigned long)ipi_bitmap,
540 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: icr);
541 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
542 ret);
543 min = max = apic_id;
544 ipi_bitmap = 0;
545 }
546 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
547 }
548
549 if (ipi_bitmap) {
550 ret = kvm_hypercall4(KVM_HC_SEND_IPI, p1: (unsigned long)ipi_bitmap,
551 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: icr);
552 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
553 ret);
554 }
555
556 local_irq_restore(flags);
557}
558
559static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
560{
561 __send_ipi_mask(mask, vector);
562}
563
564static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
565{
566 unsigned int this_cpu = smp_processor_id();
567 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
568 const struct cpumask *local_mask;
569
570 cpumask_copy(dstp: new_mask, srcp: mask);
571 cpumask_clear_cpu(cpu: this_cpu, dstp: new_mask);
572 local_mask = new_mask;
573 __send_ipi_mask(mask: local_mask, vector);
574}
575
576static int __init setup_efi_kvm_sev_migration(void)
577{
578 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
579 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
580 efi_status_t status;
581 unsigned long size;
582 bool enabled;
583
584 if (!cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT) ||
585 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
586 return 0;
587
588 if (!efi_enabled(EFI_BOOT))
589 return 0;
590
591 if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
592 pr_info("%s : EFI runtime services are not enabled\n", __func__);
593 return 0;
594 }
595
596 size = sizeof(enabled);
597
598 /* Get variable contents into buffer */
599 status = efi.get_variable(efi_sev_live_migration_enabled,
600 &efi_variable_guid, NULL, &size, &enabled);
601
602 if (status == EFI_NOT_FOUND) {
603 pr_info("%s : EFI live migration variable not found\n", __func__);
604 return 0;
605 }
606
607 if (status != EFI_SUCCESS) {
608 pr_info("%s : EFI variable retrieval failed\n", __func__);
609 return 0;
610 }
611
612 if (enabled == 0) {
613 pr_info("%s: live migration disabled in EFI\n", __func__);
614 return 0;
615 }
616
617 pr_info("%s : live migration enabled in EFI\n", __func__);
618 wrmsrq(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
619
620 return 1;
621}
622
623late_initcall(setup_efi_kvm_sev_migration);
624
625/*
626 * Set the IPI entry points
627 */
628static __init void kvm_setup_pv_ipi(void)
629{
630 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
631 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
632 pr_info("setup PV IPIs\n");
633}
634
635static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
636{
637 int cpu;
638
639 native_send_call_func_ipi(mask);
640
641 /* Make sure other vCPUs get a chance to run if they need to. */
642 for_each_cpu(cpu, mask) {
643 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
644 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
645 break;
646 }
647 }
648}
649
650static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
651 const struct flush_tlb_info *info)
652{
653 u8 state;
654 int cpu;
655 struct kvm_steal_time *src;
656 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
657
658 cpumask_copy(dstp: flushmask, srcp: cpumask);
659 /*
660 * We have to call flush only on online vCPUs. And
661 * queue flush_on_enter for pre-empted vCPUs
662 */
663 for_each_cpu(cpu, flushmask) {
664 /*
665 * The local vCPU is never preempted, so we do not explicitly
666 * skip check for local vCPU - it will never be cleared from
667 * flushmask.
668 */
669 src = &per_cpu(steal_time, cpu);
670 state = READ_ONCE(src->preempted);
671 if ((state & KVM_VCPU_PREEMPTED)) {
672 if (try_cmpxchg(&src->preempted, &state,
673 state | KVM_VCPU_FLUSH_TLB))
674 __cpumask_clear_cpu(cpu, dstp: flushmask);
675 }
676 }
677
678 native_flush_tlb_multi(cpumask: flushmask, info);
679}
680
681static __init int kvm_alloc_cpumask(void)
682{
683 int cpu;
684
685 if (!kvm_para_available() || nopv)
686 return 0;
687
688 if (pv_tlb_flush_supported() || pv_ipi_supported())
689 for_each_possible_cpu(cpu) {
690 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
691 GFP_KERNEL, node: cpu_to_node(cpu));
692 }
693
694 return 0;
695}
696arch_initcall(kvm_alloc_cpumask);
697
698static void __init kvm_smp_prepare_boot_cpu(void)
699{
700 /*
701 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
702 * shares the guest physical address with the hypervisor.
703 */
704 sev_map_percpu_data();
705
706 kvm_guest_cpu_init();
707 native_smp_prepare_boot_cpu();
708 kvm_spinlock_init();
709}
710
711static int kvm_cpu_down_prepare(unsigned int cpu)
712{
713 unsigned long flags;
714
715 local_irq_save(flags);
716 kvm_guest_cpu_offline(shutdown: false);
717 local_irq_restore(flags);
718 return 0;
719}
720
721#endif
722
723static int kvm_suspend(void)
724{
725 u64 val = 0;
726
727 kvm_guest_cpu_offline(shutdown: false);
728
729#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
730 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
731 rdmsrq(MSR_KVM_POLL_CONTROL, val);
732 has_guest_poll = !(val & 1);
733#endif
734 return 0;
735}
736
737static void kvm_resume(void)
738{
739 kvm_cpu_online(raw_smp_processor_id());
740
741#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
742 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
743 wrmsrq(MSR_KVM_POLL_CONTROL, val: 0);
744#endif
745}
746
747static struct syscore_ops kvm_syscore_ops = {
748 .suspend = kvm_suspend,
749 .resume = kvm_resume,
750};
751
752static void kvm_pv_guest_cpu_reboot(void *unused)
753{
754 kvm_guest_cpu_offline(shutdown: true);
755}
756
757static int kvm_pv_reboot_notify(struct notifier_block *nb,
758 unsigned long code, void *unused)
759{
760 if (code == SYS_RESTART)
761 on_each_cpu(func: kvm_pv_guest_cpu_reboot, NULL, wait: 1);
762 return NOTIFY_DONE;
763}
764
765static struct notifier_block kvm_pv_reboot_nb = {
766 .notifier_call = kvm_pv_reboot_notify,
767};
768
769/*
770 * After a PV feature is registered, the host will keep writing to the
771 * registered memory location. If the guest happens to shutdown, this memory
772 * won't be valid. In cases like kexec, in which you install a new kernel, this
773 * means a random memory location will be kept being written.
774 */
775#ifdef CONFIG_CRASH_DUMP
776static void kvm_crash_shutdown(struct pt_regs *regs)
777{
778 kvm_guest_cpu_offline(shutdown: true);
779 native_machine_crash_shutdown(regs);
780}
781#endif
782
783#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
784bool __kvm_vcpu_is_preempted(long cpu);
785
786__visible bool __kvm_vcpu_is_preempted(long cpu)
787{
788 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
789
790 return !!(src->preempted & KVM_VCPU_PREEMPTED);
791}
792PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
793
794#else
795
796#include <asm/asm-offsets.h>
797
798extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
799
800/*
801 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
802 * restoring to/from the stack.
803 */
804#define PV_VCPU_PREEMPTED_ASM \
805 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
806 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
807 "setne %al\n\t"
808
809DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
810 PV_VCPU_PREEMPTED_ASM, .text);
811#endif
812
813static void __init kvm_guest_init(void)
814{
815 int i;
816
817 paravirt_ops_setup();
818 register_reboot_notifier(&kvm_pv_reboot_nb);
819 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
820 raw_spin_lock_init(&async_pf_sleepers[i].lock);
821
822 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
823 has_steal_clock = 1;
824 static_call_update(pv_steal_clock, kvm_steal_clock);
825
826 pv_ops.lock.vcpu_is_preempted =
827 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
828 }
829
830 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
831 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
832
833 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
834 static_branch_enable(&kvm_async_pf_enabled);
835 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt);
836 }
837
838#ifdef CONFIG_SMP
839 if (pv_tlb_flush_supported()) {
840 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
841 pr_info("KVM setup pv remote TLB flush\n");
842 }
843
844 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
845 if (pv_sched_yield_supported()) {
846 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
847 pr_info("setup PV sched yield\n");
848 }
849 if (cpuhp_setup_state_nocalls(state: CPUHP_AP_ONLINE_DYN, name: "x86/kvm:online",
850 startup: kvm_cpu_online, teardown: kvm_cpu_down_prepare) < 0)
851 pr_err("failed to install cpu hotplug callbacks\n");
852#else
853 sev_map_percpu_data();
854 kvm_guest_cpu_init();
855#endif
856
857#ifdef CONFIG_CRASH_DUMP
858 machine_ops.crash_shutdown = kvm_crash_shutdown;
859#endif
860
861 register_syscore_ops(ops: &kvm_syscore_ops);
862
863 /*
864 * Hard lockup detection is enabled by default. Disable it, as guests
865 * can get false positives too easily, for example if the host is
866 * overcommitted.
867 */
868 hardlockup_detector_disable();
869}
870
871static noinline uint32_t __kvm_cpuid_base(void)
872{
873 if (boot_cpu_data.cpuid_level < 0)
874 return 0; /* So we don't blow up on old processors */
875
876 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
877 return cpuid_base_hypervisor(KVM_SIGNATURE, leaves: 0);
878
879 return 0;
880}
881
882static inline uint32_t kvm_cpuid_base(void)
883{
884 static int kvm_cpuid_base = -1;
885
886 if (kvm_cpuid_base == -1)
887 kvm_cpuid_base = __kvm_cpuid_base();
888
889 return kvm_cpuid_base;
890}
891
892bool kvm_para_available(void)
893{
894 return kvm_cpuid_base() != 0;
895}
896EXPORT_SYMBOL_GPL(kvm_para_available);
897
898unsigned int kvm_arch_para_features(void)
899{
900 return cpuid_eax(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
901}
902
903unsigned int kvm_arch_para_hints(void)
904{
905 return cpuid_edx(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
906}
907EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
908
909static uint32_t __init kvm_detect(void)
910{
911 return kvm_cpuid_base();
912}
913
914static void __init kvm_apic_init(void)
915{
916#ifdef CONFIG_SMP
917 if (pv_ipi_supported())
918 kvm_setup_pv_ipi();
919#endif
920}
921
922static bool __init kvm_msi_ext_dest_id(void)
923{
924 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
925}
926
927static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
928{
929 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: pfn << PAGE_SHIFT, p2: npages,
930 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
931}
932
933static void __init kvm_init_platform(void)
934{
935 u64 tolud = PFN_PHYS(e820__end_of_low_ram_pfn());
936 /*
937 * Note, hardware requires variable MTRR ranges to be power-of-2 sized
938 * and naturally aligned. But when forcing guest MTRR state, Linux
939 * doesn't program the forced ranges into hardware. Don't bother doing
940 * the math to generate a technically-legal range.
941 */
942 struct mtrr_var_range pci_hole = {
943 .base_lo = tolud | X86_MEMTYPE_UC,
944 .mask_lo = (u32)(~(SZ_4G - tolud - 1)) | MTRR_PHYSMASK_V,
945 .mask_hi = (BIT_ULL(boot_cpu_data.x86_phys_bits) - 1) >> 32,
946 };
947
948 if (cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT) &&
949 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
950 unsigned long nr_pages;
951 int i;
952
953 pv_ops.mmu.notify_page_enc_status_changed =
954 kvm_sev_hc_page_enc_status;
955
956 /*
957 * Reset the host's shared pages list related to kernel
958 * specific page encryption status settings before we load a
959 * new kernel by kexec. Reset the page encryption status
960 * during early boot instead of just before kexec to avoid SMP
961 * races during kvm_pv_guest_cpu_reboot().
962 * NOTE: We cannot reset the complete shared pages list
963 * here as we need to retain the UEFI/OVMF firmware
964 * specific settings.
965 */
966
967 for (i = 0; i < e820_table->nr_entries; i++) {
968 struct e820_entry *entry = &e820_table->entries[i];
969
970 if (entry->type != E820_TYPE_RAM)
971 continue;
972
973 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
974
975 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: entry->addr,
976 p2: nr_pages,
977 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
978 }
979
980 /*
981 * Ensure that _bss_decrypted section is marked as decrypted in the
982 * shared pages list.
983 */
984 early_set_mem_enc_dec_hypercall(vaddr: (unsigned long)__start_bss_decrypted,
985 size: __end_bss_decrypted - __start_bss_decrypted, enc: 0);
986
987 /*
988 * If not booted using EFI, enable Live migration support.
989 */
990 if (!efi_enabled(EFI_BOOT))
991 wrmsrq(MSR_KVM_MIGRATION_CONTROL,
992 KVM_MIGRATION_READY);
993 }
994 kvmclock_init();
995 x86_platform.apic_post_init = kvm_apic_init;
996
997 /*
998 * Set WB as the default cache mode for SEV-SNP and TDX, with a single
999 * UC range for the legacy PCI hole, e.g. so that devices that expect
1000 * to get UC/WC mappings don't get surprised with WB.
1001 */
1002 guest_force_mtrr_state(var: &pci_hole, num_var: 1, MTRR_TYPE_WRBACK);
1003}
1004
1005#if defined(CONFIG_AMD_MEM_ENCRYPT)
1006static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
1007{
1008 /* RAX and CPL are already in the GHCB */
1009 ghcb_set_rbx(ghcb, regs->bx);
1010 ghcb_set_rcx(ghcb, regs->cx);
1011 ghcb_set_rdx(ghcb, regs->dx);
1012 ghcb_set_rsi(ghcb, regs->si);
1013}
1014
1015static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
1016{
1017 /* No checking of the return state needed */
1018 return true;
1019}
1020#endif
1021
1022const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1023 .name = "KVM",
1024 .detect = kvm_detect,
1025 .type = X86_HYPER_KVM,
1026 .init.guest_late_init = kvm_guest_init,
1027 .init.x2apic_available = kvm_para_available,
1028 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1029 .init.init_platform = kvm_init_platform,
1030#if defined(CONFIG_AMD_MEM_ENCRYPT)
1031 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1032 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1033#endif
1034};
1035
1036static __init int activate_jump_labels(void)
1037{
1038 if (has_steal_clock) {
1039 static_key_slow_inc(key: &paravirt_steal_enabled);
1040 if (steal_acc)
1041 static_key_slow_inc(key: &paravirt_steal_rq_enabled);
1042 }
1043
1044 return 0;
1045}
1046arch_initcall(activate_jump_labels);
1047
1048#ifdef CONFIG_PARAVIRT_SPINLOCKS
1049
1050/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1051static void kvm_kick_cpu(int cpu)
1052{
1053 unsigned long flags = 0;
1054 u32 apicid;
1055
1056 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1057 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1058}
1059
1060#include <asm/qspinlock.h>
1061
1062static void kvm_wait(u8 *ptr, u8 val)
1063{
1064 if (in_nmi())
1065 return;
1066
1067 /*
1068 * halt until it's our turn and kicked. Note that we do safe halt
1069 * for irq enabled case to avoid hang when lock info is overwritten
1070 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1071 */
1072 if (irqs_disabled()) {
1073 if (READ_ONCE(*ptr) == val)
1074 halt();
1075 } else {
1076 local_irq_disable();
1077
1078 /* safe_halt() will enable IRQ */
1079 if (READ_ONCE(*ptr) == val)
1080 safe_halt();
1081 else
1082 local_irq_enable();
1083 }
1084}
1085
1086/*
1087 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1088 */
1089void __init kvm_spinlock_init(void)
1090{
1091 /*
1092 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1093 * are available.
1094 */
1095 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1096 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1097 goto out;
1098 }
1099
1100 if (num_possible_cpus() == 1) {
1101 pr_info("PV spinlocks disabled, single CPU\n");
1102 goto out;
1103 }
1104
1105 if (nopvspin) {
1106 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1107 goto out;
1108 }
1109
1110 /*
1111 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1112 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1113 * preferred over native qspinlock when vCPU is preempted.
1114 */
1115 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1116 pr_info("PV spinlocks disabled, no host support\n");
1117 return;
1118 }
1119
1120 pr_info("PV spinlocks enabled\n");
1121
1122 __pv_init_lock_hash();
1123 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1124 pv_ops.lock.queued_spin_unlock =
1125 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1126 pv_ops.lock.wait = kvm_wait;
1127 pv_ops.lock.kick = kvm_kick_cpu;
1128
1129 /*
1130 * When PV spinlock is enabled which is preferred over
1131 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1132 * Just disable it anyway.
1133 */
1134out:
1135 static_branch_disable(&virt_spin_lock_key);
1136}
1137
1138#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1139
1140#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1141
1142static void kvm_disable_host_haltpoll(void *i)
1143{
1144 wrmsrq(MSR_KVM_POLL_CONTROL, val: 0);
1145}
1146
1147static void kvm_enable_host_haltpoll(void *i)
1148{
1149 wrmsrq(MSR_KVM_POLL_CONTROL, val: 1);
1150}
1151
1152void arch_haltpoll_enable(unsigned int cpu)
1153{
1154 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1155 pr_err_once("host does not support poll control\n");
1156 pr_err_once("host upgrade recommended\n");
1157 return;
1158 }
1159
1160 /* Enable guest halt poll disables host halt poll */
1161 smp_call_function_single(cpuid: cpu, func: kvm_disable_host_haltpoll, NULL, wait: 1);
1162}
1163EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1164
1165void arch_haltpoll_disable(unsigned int cpu)
1166{
1167 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1168 return;
1169
1170 /* Disable guest halt poll enables host halt poll */
1171 smp_call_function_single(cpuid: cpu, func: kvm_enable_host_haltpoll, NULL, wait: 1);
1172}
1173EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1174#endif
1175