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 <linux/kvm_types.h>
33#include <asm/timer.h>
34#include <asm/cpu.h>
35#include <asm/traps.h>
36#include <asm/desc.h>
37#include <asm/tlbflush.h>
38#include <asm/apic.h>
39#include <asm/apicdef.h>
40#include <asm/hypervisor.h>
41#include <asm/mtrr.h>
42#include <asm/tlb.h>
43#include <asm/cpuidle_haltpoll.h>
44#include <asm/msr.h>
45#include <asm/ptrace.h>
46#include <asm/reboot.h>
47#include <asm/svm.h>
48#include <asm/e820/api.h>
49
50DEFINE_STATIC_KEY_FALSE_RO(kvm_async_pf_enabled);
51
52static int kvmapf = 1;
53
54static int __init parse_no_kvmapf(char *arg)
55{
56 kvmapf = 0;
57 return 0;
58}
59
60early_param("no-kvmapf", parse_no_kvmapf);
61
62static int steal_acc = 1;
63static int __init parse_no_stealacc(char *arg)
64{
65 steal_acc = 0;
66 return 0;
67}
68
69early_param("no-steal-acc", parse_no_stealacc);
70
71static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
72static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
73DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
74static int has_steal_clock = 0;
75
76static int has_guest_poll = 0;
77/*
78 * No need for any "IO delay" on KVM
79 */
80static void kvm_io_delay(void)
81{
82}
83
84#define KVM_TASK_SLEEP_HASHBITS 8
85#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
86
87struct kvm_task_sleep_node {
88 struct hlist_node link;
89 struct swait_queue_head wq;
90 u32 token;
91 int cpu;
92 bool dummy;
93};
94
95static struct kvm_task_sleep_head {
96 raw_spinlock_t lock;
97 struct hlist_head list;
98} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
99
100static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
101 u32 token)
102{
103 struct hlist_node *p;
104
105 hlist_for_each(p, &b->list) {
106 struct kvm_task_sleep_node *n =
107 hlist_entry(p, typeof(*n), link);
108 if (n->token == token)
109 return n;
110 }
111
112 return NULL;
113}
114
115static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
116{
117 u32 key = hash_32(val: token, KVM_TASK_SLEEP_HASHBITS);
118 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
119 struct kvm_task_sleep_node *e;
120
121 raw_spin_lock(&b->lock);
122 e = _find_apf_task(b, token);
123 if (e) {
124 struct kvm_task_sleep_node *dummy = NULL;
125
126 /*
127 * The entry can either be a 'dummy' entry (which is put on the
128 * list when wake-up happens ahead of APF handling completion)
129 * or a token from another task which should not be touched.
130 */
131 if (e->dummy) {
132 hlist_del(n: &e->link);
133 dummy = e;
134 }
135
136 raw_spin_unlock(&b->lock);
137 kfree(objp: dummy);
138 return false;
139 }
140
141 n->token = token;
142 n->cpu = smp_processor_id();
143 n->dummy = false;
144 init_swait_queue_head(&n->wq);
145 hlist_add_head(n: &n->link, h: &b->list);
146 raw_spin_unlock(&b->lock);
147 return true;
148}
149
150/*
151 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
152 * @token: Token to identify the sleep node entry
153 *
154 * Invoked from the async pagefault handling code or from the VM exit page
155 * fault handler. In both cases RCU is watching.
156 */
157void kvm_async_pf_task_wait_schedule(u32 token)
158{
159 struct kvm_task_sleep_node n;
160 DECLARE_SWAITQUEUE(wait);
161
162 lockdep_assert_irqs_disabled();
163
164 if (!kvm_async_pf_queue_task(token, n: &n))
165 return;
166
167 for (;;) {
168 prepare_to_swait_exclusive(q: &n.wq, wait: &wait, TASK_UNINTERRUPTIBLE);
169 if (hlist_unhashed(h: &n.link))
170 break;
171
172 local_irq_enable();
173 schedule();
174 local_irq_disable();
175 }
176 finish_swait(q: &n.wq, wait: &wait);
177}
178EXPORT_SYMBOL_FOR_KVM(kvm_async_pf_task_wait_schedule);
179
180static void apf_task_wake_one(struct kvm_task_sleep_node *n)
181{
182 hlist_del_init(n: &n->link);
183 if (swq_has_sleeper(wq: &n->wq))
184 swake_up_one(q: &n->wq);
185}
186
187static void apf_task_wake_all(void)
188{
189 int i;
190
191 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
192 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
193 struct kvm_task_sleep_node *n;
194 struct hlist_node *p, *next;
195
196 raw_spin_lock(&b->lock);
197 hlist_for_each_safe(p, next, &b->list) {
198 n = hlist_entry(p, typeof(*n), link);
199 if (n->cpu == smp_processor_id())
200 apf_task_wake_one(n);
201 }
202 raw_spin_unlock(&b->lock);
203 }
204}
205
206static void kvm_async_pf_task_wake(u32 token)
207{
208 u32 key = hash_32(val: token, KVM_TASK_SLEEP_HASHBITS);
209 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
210 struct kvm_task_sleep_node *n, *dummy = NULL;
211
212 if (token == ~0) {
213 apf_task_wake_all();
214 return;
215 }
216
217again:
218 raw_spin_lock(&b->lock);
219 n = _find_apf_task(b, token);
220 if (!n) {
221 /*
222 * Async #PF not yet handled, add a dummy entry for the token.
223 * Allocating the token must be down outside of the raw lock
224 * as the allocator is preemptible on PREEMPT_RT kernels.
225 */
226 if (!dummy) {
227 raw_spin_unlock(&b->lock);
228 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
229
230 /*
231 * Continue looping on allocation failure, eventually
232 * the async #PF will be handled and allocating a new
233 * node will be unnecessary.
234 */
235 if (!dummy)
236 cpu_relax();
237
238 /*
239 * Recheck for async #PF completion before enqueueing
240 * the dummy token to avoid duplicate list entries.
241 */
242 goto again;
243 }
244 dummy->token = token;
245 dummy->cpu = smp_processor_id();
246 dummy->dummy = true;
247 init_swait_queue_head(&dummy->wq);
248 hlist_add_head(n: &dummy->link, h: &b->list);
249 dummy = NULL;
250 } else {
251 apf_task_wake_one(n);
252 }
253 raw_spin_unlock(&b->lock);
254
255 /* A dummy token might be allocated and ultimately not used. */
256 kfree(objp: dummy);
257}
258
259noinstr u32 kvm_read_and_reset_apf_flags(void)
260{
261 u32 flags = 0;
262
263 if (__this_cpu_read(async_pf_enabled)) {
264 flags = __this_cpu_read(apf_reason.flags);
265 __this_cpu_write(apf_reason.flags, 0);
266 }
267
268 return flags;
269}
270EXPORT_SYMBOL_FOR_KVM(kvm_read_and_reset_apf_flags);
271
272noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
273{
274 u32 flags = kvm_read_and_reset_apf_flags();
275 irqentry_state_t state;
276
277 if (!flags)
278 return false;
279
280 state = irqentry_enter(regs);
281 instrumentation_begin();
282
283 /*
284 * If the host managed to inject an async #PF into an interrupt
285 * disabled region, then die hard as this is not going to end well
286 * and the host side is seriously broken.
287 */
288 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
289 panic(fmt: "Host injected async #PF in interrupt disabled region\n");
290
291 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
292 if (unlikely(!(user_mode(regs))))
293 panic(fmt: "Host injected async #PF in kernel mode\n");
294 /* Page is swapped out by the host. */
295 kvm_async_pf_task_wait_schedule(token);
296 } else {
297 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
298 }
299
300 instrumentation_end();
301 irqentry_exit(regs, state);
302 return true;
303}
304
305DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
306{
307 struct pt_regs *old_regs = set_irq_regs(regs);
308 u32 token;
309
310 apic_eoi();
311
312 inc_irq_stat(irq_hv_callback_count);
313
314 if (__this_cpu_read(async_pf_enabled)) {
315 token = __this_cpu_read(apf_reason.token);
316 kvm_async_pf_task_wake(token);
317 __this_cpu_write(apf_reason.token, 0);
318 wrmsrq(MSR_KVM_ASYNC_PF_ACK, val: 1);
319 }
320
321 set_irq_regs(old_regs);
322}
323
324static void __init paravirt_ops_setup(void)
325{
326 pv_info.name = "KVM";
327
328 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
329 pv_ops.cpu.io_delay = kvm_io_delay;
330
331#ifdef CONFIG_X86_IO_APIC
332 no_timer_check = 1;
333#endif
334}
335
336static void kvm_register_steal_time(void)
337{
338 int cpu = smp_processor_id();
339 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
340
341 if (!has_steal_clock)
342 return;
343
344 wrmsrq(MSR_KVM_STEAL_TIME, val: (slow_virt_to_phys(address: st) | KVM_MSR_ENABLED));
345 pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
346 (unsigned long long) slow_virt_to_phys(st));
347}
348
349static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
350
351static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
352{
353 /**
354 * This relies on __test_and_clear_bit to modify the memory
355 * in a way that is atomic with respect to the local CPU.
356 * The hypervisor only accesses this memory from the local CPU so
357 * there's no need for lock or memory barriers.
358 * An optimization barrier is implied in apic write.
359 */
360 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
361 return;
362 apic_native_eoi();
363}
364
365static void kvm_guest_cpu_init(void)
366{
367 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
368 u64 pa;
369
370 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
371
372 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
373 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
374
375 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
376 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
377
378 wrmsrq(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
379
380 wrmsrq(MSR_KVM_ASYNC_PF_EN, val: pa);
381 __this_cpu_write(async_pf_enabled, true);
382 pr_debug("setup async PF for cpu %d\n", smp_processor_id());
383 }
384
385 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
386 unsigned long pa;
387
388 /* Size alignment is implied but just to make it explicit. */
389 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
390 __this_cpu_write(kvm_apic_eoi, 0);
391 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
392 | KVM_MSR_ENABLED;
393 wrmsrq(MSR_KVM_PV_EOI_EN, val: pa);
394 }
395
396 if (has_steal_clock)
397 kvm_register_steal_time();
398}
399
400static void kvm_pv_disable_apf(void)
401{
402 if (!__this_cpu_read(async_pf_enabled))
403 return;
404
405 wrmsrq(MSR_KVM_ASYNC_PF_EN, val: 0);
406 __this_cpu_write(async_pf_enabled, false);
407
408 pr_debug("disable async PF for cpu %d\n", smp_processor_id());
409}
410
411static void kvm_disable_steal_time(void)
412{
413 if (!has_steal_clock)
414 return;
415
416 wrmsrq(MSR_KVM_STEAL_TIME, val: 0);
417}
418
419static u64 kvm_steal_clock(int cpu)
420{
421 u64 steal;
422 struct kvm_steal_time *src;
423 int version;
424
425 src = &per_cpu(steal_time, cpu);
426 do {
427 version = src->version;
428 virt_rmb();
429 steal = src->steal;
430 virt_rmb();
431 } while ((version & 1) || (version != src->version));
432
433 return steal;
434}
435
436static inline __init void __set_percpu_decrypted(void *ptr, unsigned long size)
437{
438 early_set_memory_decrypted(vaddr: (unsigned long) ptr, size);
439}
440
441/*
442 * Iterate through all possible CPUs and map the memory region pointed
443 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
444 *
445 * Note: we iterate through all possible CPUs to ensure that CPUs
446 * hotplugged will have their per-cpu variable already mapped as
447 * decrypted.
448 */
449static void __init sev_map_percpu_data(void)
450{
451 int cpu;
452
453 if (cc_vendor != CC_VENDOR_AMD ||
454 !cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT))
455 return;
456
457 for_each_possible_cpu(cpu) {
458 __set_percpu_decrypted(ptr: &per_cpu(apf_reason, cpu), size: sizeof(apf_reason));
459 __set_percpu_decrypted(ptr: &per_cpu(steal_time, cpu), size: sizeof(steal_time));
460 __set_percpu_decrypted(ptr: &per_cpu(kvm_apic_eoi, cpu), size: sizeof(kvm_apic_eoi));
461 }
462}
463
464static void kvm_guest_cpu_offline(bool shutdown)
465{
466 kvm_disable_steal_time();
467 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
468 wrmsrq(MSR_KVM_PV_EOI_EN, val: 0);
469 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
470 wrmsrq(MSR_KVM_MIGRATION_CONTROL, val: 0);
471 kvm_pv_disable_apf();
472 if (!shutdown)
473 apf_task_wake_all();
474 kvmclock_disable();
475}
476
477static int kvm_cpu_online(unsigned int cpu)
478{
479 unsigned long flags;
480
481 local_irq_save(flags);
482 kvm_guest_cpu_init();
483 local_irq_restore(flags);
484 return 0;
485}
486
487#ifdef CONFIG_SMP
488
489static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
490
491static bool pv_tlb_flush_supported(void)
492{
493 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
494 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
495 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
496 !boot_cpu_has(X86_FEATURE_MWAIT) &&
497 (num_possible_cpus() != 1));
498}
499
500static bool pv_ipi_supported(void)
501{
502 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
503 (num_possible_cpus() != 1));
504}
505
506static bool pv_sched_yield_supported(void)
507{
508 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
509 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
510 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
511 !boot_cpu_has(X86_FEATURE_MWAIT) &&
512 (num_possible_cpus() != 1));
513}
514
515#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
516
517static void __send_ipi_mask(const struct cpumask *mask, int vector)
518{
519 unsigned long flags;
520 int cpu, min = 0, max = 0;
521#ifdef CONFIG_X86_64
522 __uint128_t ipi_bitmap = 0;
523#else
524 u64 ipi_bitmap = 0;
525#endif
526 u32 apic_id, icr;
527 long ret;
528
529 if (cpumask_empty(srcp: mask))
530 return;
531
532 local_irq_save(flags);
533
534 switch (vector) {
535 default:
536 icr = APIC_DM_FIXED | vector;
537 break;
538 case NMI_VECTOR:
539 icr = APIC_DM_NMI;
540 break;
541 }
542
543 for_each_cpu(cpu, mask) {
544 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
545 if (!ipi_bitmap) {
546 min = max = apic_id;
547 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
548 ipi_bitmap <<= min - apic_id;
549 min = apic_id;
550 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
551 max = apic_id < max ? max : apic_id;
552 } else {
553 ret = kvm_hypercall4(KVM_HC_SEND_IPI, p1: (unsigned long)ipi_bitmap,
554 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: icr);
555 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
556 ret);
557 min = max = apic_id;
558 ipi_bitmap = 0;
559 }
560 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
561 }
562
563 if (ipi_bitmap) {
564 ret = kvm_hypercall4(KVM_HC_SEND_IPI, p1: (unsigned long)ipi_bitmap,
565 p2: (unsigned long)(ipi_bitmap >> BITS_PER_LONG), p3: min, p4: icr);
566 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
567 ret);
568 }
569
570 local_irq_restore(flags);
571}
572
573static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
574{
575 __send_ipi_mask(mask, vector);
576}
577
578static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
579{
580 unsigned int this_cpu = smp_processor_id();
581 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
582 const struct cpumask *local_mask;
583
584 cpumask_copy(dstp: new_mask, srcp: mask);
585 cpumask_clear_cpu(cpu: this_cpu, dstp: new_mask);
586 local_mask = new_mask;
587 __send_ipi_mask(mask: local_mask, vector);
588}
589
590static int __init setup_efi_kvm_sev_migration(void)
591{
592 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
593 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
594 efi_status_t status;
595 unsigned long size;
596 bool enabled;
597
598 if (!cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT) ||
599 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
600 return 0;
601
602 if (!efi_enabled(EFI_BOOT))
603 return 0;
604
605 if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
606 pr_info("%s : EFI runtime services are not enabled\n", __func__);
607 return 0;
608 }
609
610 size = sizeof(enabled);
611
612 /* Get variable contents into buffer */
613 status = efi.get_variable(efi_sev_live_migration_enabled,
614 &efi_variable_guid, NULL, &size, &enabled);
615
616 if (status == EFI_NOT_FOUND) {
617 pr_info("%s : EFI live migration variable not found\n", __func__);
618 return 0;
619 }
620
621 if (status != EFI_SUCCESS) {
622 pr_info("%s : EFI variable retrieval failed\n", __func__);
623 return 0;
624 }
625
626 if (enabled == 0) {
627 pr_info("%s: live migration disabled in EFI\n", __func__);
628 return 0;
629 }
630
631 pr_info("%s : live migration enabled in EFI\n", __func__);
632 wrmsrq(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
633
634 return 1;
635}
636
637late_initcall(setup_efi_kvm_sev_migration);
638
639/*
640 * Set the IPI entry points
641 */
642static __init void kvm_setup_pv_ipi(void)
643{
644 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
645 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
646 pr_info("setup PV IPIs\n");
647}
648
649static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
650{
651 int cpu;
652
653 native_send_call_func_ipi(mask);
654
655 /* Make sure other vCPUs get a chance to run if they need to. */
656 for_each_cpu(cpu, mask) {
657 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
658 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
659 break;
660 }
661 }
662}
663
664static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
665 const struct flush_tlb_info *info)
666{
667 u8 state;
668 int cpu;
669 struct kvm_steal_time *src;
670 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
671
672 cpumask_copy(dstp: flushmask, srcp: cpumask);
673 /*
674 * We have to call flush only on online vCPUs. And
675 * queue flush_on_enter for pre-empted vCPUs
676 */
677 for_each_cpu(cpu, flushmask) {
678 /*
679 * The local vCPU is never preempted, so we do not explicitly
680 * skip check for local vCPU - it will never be cleared from
681 * flushmask.
682 */
683 src = &per_cpu(steal_time, cpu);
684 state = READ_ONCE(src->preempted);
685 if ((state & KVM_VCPU_PREEMPTED)) {
686 if (try_cmpxchg(&src->preempted, &state,
687 state | KVM_VCPU_FLUSH_TLB))
688 __cpumask_clear_cpu(cpu, dstp: flushmask);
689 }
690 }
691
692 native_flush_tlb_multi(cpumask: flushmask, info);
693}
694
695static __init int kvm_alloc_cpumask(void)
696{
697 int cpu;
698
699 if (!kvm_para_available() || nopv)
700 return 0;
701
702 if (pv_tlb_flush_supported() || pv_ipi_supported())
703 for_each_possible_cpu(cpu) {
704 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
705 GFP_KERNEL, cpu_to_node(cpu));
706 }
707
708 return 0;
709}
710arch_initcall(kvm_alloc_cpumask);
711
712static void __init kvm_smp_prepare_boot_cpu(void)
713{
714 /*
715 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
716 * shares the guest physical address with the hypervisor.
717 */
718 sev_map_percpu_data();
719
720 kvm_guest_cpu_init();
721 native_smp_prepare_boot_cpu();
722 kvm_spinlock_init();
723}
724
725static int kvm_cpu_down_prepare(unsigned int cpu)
726{
727 unsigned long flags;
728
729 local_irq_save(flags);
730 kvm_guest_cpu_offline(shutdown: false);
731 local_irq_restore(flags);
732 return 0;
733}
734
735#endif
736
737static int kvm_suspend(void *data)
738{
739 u64 val = 0;
740
741 kvm_guest_cpu_offline(shutdown: false);
742
743#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
744 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
745 rdmsrq(MSR_KVM_POLL_CONTROL, val);
746 has_guest_poll = !(val & 1);
747#endif
748 return 0;
749}
750
751static void kvm_resume(void *data)
752{
753 kvm_cpu_online(raw_smp_processor_id());
754
755#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
756 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
757 wrmsrq(MSR_KVM_POLL_CONTROL, val: 0);
758#endif
759}
760
761static const struct syscore_ops kvm_syscore_ops = {
762 .suspend = kvm_suspend,
763 .resume = kvm_resume,
764};
765
766static struct syscore kvm_syscore = {
767 .ops = &kvm_syscore_ops,
768};
769
770static void kvm_pv_guest_cpu_reboot(void *unused)
771{
772 kvm_guest_cpu_offline(shutdown: true);
773}
774
775static int kvm_pv_reboot_notify(struct notifier_block *nb,
776 unsigned long code, void *unused)
777{
778 if (code == SYS_RESTART)
779 on_each_cpu(func: kvm_pv_guest_cpu_reboot, NULL, wait: 1);
780 return NOTIFY_DONE;
781}
782
783static struct notifier_block kvm_pv_reboot_nb = {
784 .notifier_call = kvm_pv_reboot_notify,
785};
786
787/*
788 * After a PV feature is registered, the host will keep writing to the
789 * registered memory location. If the guest happens to shutdown, this memory
790 * won't be valid. In cases like kexec, in which you install a new kernel, this
791 * means a random memory location will be kept being written.
792 */
793#ifdef CONFIG_CRASH_DUMP
794static void kvm_crash_shutdown(struct pt_regs *regs)
795{
796 kvm_guest_cpu_offline(shutdown: true);
797 native_machine_crash_shutdown(regs);
798}
799#endif
800
801#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
802bool __kvm_vcpu_is_preempted(long cpu);
803
804__visible bool __kvm_vcpu_is_preempted(long cpu)
805{
806 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
807
808 return !!(src->preempted & KVM_VCPU_PREEMPTED);
809}
810PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
811
812#else
813
814#include <asm/asm-offsets.h>
815
816extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
817
818/*
819 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
820 * restoring to/from the stack.
821 */
822#define PV_VCPU_PREEMPTED_ASM \
823 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
824 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
825 "setne %al\n\t"
826
827DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
828 PV_VCPU_PREEMPTED_ASM, .text);
829#endif
830
831static void __init kvm_guest_init(void)
832{
833 int i;
834
835 paravirt_ops_setup();
836 register_reboot_notifier(&kvm_pv_reboot_nb);
837 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
838 raw_spin_lock_init(&async_pf_sleepers[i].lock);
839
840 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
841 has_steal_clock = 1;
842 static_call_update(pv_steal_clock, kvm_steal_clock);
843
844 pv_ops.lock.vcpu_is_preempted =
845 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
846 }
847
848 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
849 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
850
851 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
852 static_branch_enable(&kvm_async_pf_enabled);
853 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt);
854 }
855
856#ifdef CONFIG_SMP
857 if (pv_tlb_flush_supported()) {
858 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
859 pr_info("KVM setup pv remote TLB flush\n");
860 }
861
862 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
863 if (pv_sched_yield_supported()) {
864 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
865 pr_info("setup PV sched yield\n");
866 }
867 if (cpuhp_setup_state_nocalls(state: CPUHP_AP_ONLINE_DYN, name: "x86/kvm:online",
868 startup: kvm_cpu_online, teardown: kvm_cpu_down_prepare) < 0)
869 pr_err("failed to install cpu hotplug callbacks\n");
870#else
871 sev_map_percpu_data();
872 kvm_guest_cpu_init();
873#endif
874
875#ifdef CONFIG_CRASH_DUMP
876 machine_ops.crash_shutdown = kvm_crash_shutdown;
877#endif
878
879 register_syscore(syscore: &kvm_syscore);
880
881 /*
882 * Hard lockup detection is enabled by default. Disable it, as guests
883 * can get false positives too easily, for example if the host is
884 * overcommitted.
885 */
886 hardlockup_detector_disable();
887}
888
889static noinline uint32_t __kvm_cpuid_base(void)
890{
891 if (boot_cpu_data.cpuid_level < 0)
892 return 0; /* So we don't blow up on old processors */
893
894 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
895 return cpuid_base_hypervisor(KVM_SIGNATURE, leaves: 0);
896
897 return 0;
898}
899
900static inline uint32_t kvm_cpuid_base(void)
901{
902 static int kvm_cpuid_base = -1;
903
904 if (kvm_cpuid_base == -1)
905 kvm_cpuid_base = __kvm_cpuid_base();
906
907 return kvm_cpuid_base;
908}
909
910bool kvm_para_available(void)
911{
912 return kvm_cpuid_base() != 0;
913}
914EXPORT_SYMBOL_GPL(kvm_para_available);
915
916unsigned int kvm_arch_para_features(void)
917{
918 return cpuid_eax(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
919}
920
921unsigned int kvm_arch_para_hints(void)
922{
923 return cpuid_edx(op: kvm_cpuid_base() | KVM_CPUID_FEATURES);
924}
925EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
926
927static uint32_t __init kvm_detect(void)
928{
929 return kvm_cpuid_base();
930}
931
932static void __init kvm_apic_init(void)
933{
934#ifdef CONFIG_SMP
935 if (pv_ipi_supported())
936 kvm_setup_pv_ipi();
937#endif
938}
939
940static bool __init kvm_msi_ext_dest_id(void)
941{
942 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
943}
944
945static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
946{
947 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: pfn << PAGE_SHIFT, p2: npages,
948 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
949}
950
951static void __init kvm_init_platform(void)
952{
953 u64 tolud = PFN_PHYS(e820__end_of_low_ram_pfn());
954 /*
955 * Note, hardware requires variable MTRR ranges to be power-of-2 sized
956 * and naturally aligned. But when forcing guest MTRR state, Linux
957 * doesn't program the forced ranges into hardware. Don't bother doing
958 * the math to generate a technically-legal range.
959 */
960 struct mtrr_var_range pci_hole = {
961 .base_lo = tolud | X86_MEMTYPE_UC,
962 .mask_lo = (u32)(~(SZ_4G - tolud - 1)) | MTRR_PHYSMASK_V,
963 .mask_hi = (BIT_ULL(boot_cpu_data.x86_phys_bits) - 1) >> 32,
964 };
965
966 if (cc_platform_has(attr: CC_ATTR_GUEST_MEM_ENCRYPT) &&
967 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
968 unsigned long nr_pages;
969 int i;
970
971 pv_ops.mmu.notify_page_enc_status_changed =
972 kvm_sev_hc_page_enc_status;
973
974 /*
975 * Reset the host's shared pages list related to kernel
976 * specific page encryption status settings before we load a
977 * new kernel by kexec. Reset the page encryption status
978 * during early boot instead of just before kexec to avoid SMP
979 * races during kvm_pv_guest_cpu_reboot().
980 * NOTE: We cannot reset the complete shared pages list
981 * here as we need to retain the UEFI/OVMF firmware
982 * specific settings.
983 */
984
985 for (i = 0; i < e820_table->nr_entries; i++) {
986 struct e820_entry *entry = &e820_table->entries[i];
987
988 if (entry->type != E820_TYPE_RAM)
989 continue;
990
991 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
992
993 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, p1: entry->addr,
994 p2: nr_pages,
995 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
996 }
997
998 /*
999 * Ensure that _bss_decrypted section is marked as decrypted in the
1000 * shared pages list.
1001 */
1002 early_set_mem_enc_dec_hypercall(vaddr: (unsigned long)__start_bss_decrypted,
1003 size: __end_bss_decrypted - __start_bss_decrypted, enc: 0);
1004
1005 /*
1006 * If not booted using EFI, enable Live migration support.
1007 */
1008 if (!efi_enabled(EFI_BOOT))
1009 wrmsrq(MSR_KVM_MIGRATION_CONTROL,
1010 KVM_MIGRATION_READY);
1011 }
1012 kvmclock_init();
1013 x86_platform.apic_post_init = kvm_apic_init;
1014
1015 /*
1016 * Set WB as the default cache mode for SEV-SNP and TDX, with a single
1017 * UC range for the legacy PCI hole, e.g. so that devices that expect
1018 * to get UC/WC mappings don't get surprised with WB.
1019 */
1020 guest_force_mtrr_state(var: &pci_hole, num_var: 1, MTRR_TYPE_WRBACK);
1021}
1022
1023#if defined(CONFIG_AMD_MEM_ENCRYPT)
1024static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
1025{
1026 /* RAX and CPL are already in the GHCB */
1027 ghcb_set_rbx(ghcb, value: regs->bx);
1028 ghcb_set_rcx(ghcb, value: regs->cx);
1029 ghcb_set_rdx(ghcb, value: regs->dx);
1030 ghcb_set_rsi(ghcb, value: regs->si);
1031}
1032
1033static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
1034{
1035 /* No checking of the return state needed */
1036 return true;
1037}
1038#endif
1039
1040const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1041 .name = "KVM",
1042 .detect = kvm_detect,
1043 .type = X86_HYPER_KVM,
1044 .init.guest_late_init = kvm_guest_init,
1045 .init.x2apic_available = kvm_para_available,
1046 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1047 .init.init_platform = kvm_init_platform,
1048#if defined(CONFIG_AMD_MEM_ENCRYPT)
1049 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1050 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1051#endif
1052};
1053
1054static __init int activate_jump_labels(void)
1055{
1056 if (has_steal_clock) {
1057 static_key_slow_inc(key: &paravirt_steal_enabled);
1058 if (steal_acc)
1059 static_key_slow_inc(key: &paravirt_steal_rq_enabled);
1060 }
1061
1062 return 0;
1063}
1064arch_initcall(activate_jump_labels);
1065
1066#ifdef CONFIG_PARAVIRT_SPINLOCKS
1067
1068/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1069static void kvm_kick_cpu(int cpu)
1070{
1071 unsigned long flags = 0;
1072 u32 apicid;
1073
1074 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1075 kvm_hypercall2(KVM_HC_KICK_CPU, p1: flags, p2: apicid);
1076}
1077
1078#include <asm/qspinlock.h>
1079
1080static void kvm_wait(u8 *ptr, u8 val)
1081{
1082 if (in_nmi())
1083 return;
1084
1085 /*
1086 * halt until it's our turn and kicked. Note that we do safe halt
1087 * for irq enabled case to avoid hang when lock info is overwritten
1088 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1089 */
1090 if (irqs_disabled()) {
1091 if (READ_ONCE(*ptr) == val)
1092 halt();
1093 } else {
1094 local_irq_disable();
1095
1096 /* safe_halt() will enable IRQ */
1097 if (READ_ONCE(*ptr) == val)
1098 safe_halt();
1099 else
1100 local_irq_enable();
1101 }
1102}
1103
1104/*
1105 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1106 */
1107void __init kvm_spinlock_init(void)
1108{
1109 /*
1110 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1111 * are available.
1112 */
1113 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1114 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1115 goto out;
1116 }
1117
1118 if (num_possible_cpus() == 1) {
1119 pr_info("PV spinlocks disabled, single CPU\n");
1120 goto out;
1121 }
1122
1123 if (nopvspin) {
1124 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1125 goto out;
1126 }
1127
1128 /*
1129 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1130 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1131 * preferred over native qspinlock when vCPU is preempted.
1132 */
1133 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1134 pr_info("PV spinlocks disabled, no host support\n");
1135 return;
1136 }
1137
1138 pr_info("PV spinlocks enabled\n");
1139
1140 __pv_init_lock_hash();
1141 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1142 pv_ops.lock.queued_spin_unlock =
1143 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1144 pv_ops.lock.wait = kvm_wait;
1145 pv_ops.lock.kick = kvm_kick_cpu;
1146
1147 /*
1148 * When PV spinlock is enabled which is preferred over
1149 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1150 * Just disable it anyway.
1151 */
1152out:
1153 static_branch_disable(&virt_spin_lock_key);
1154}
1155
1156#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1157
1158#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1159
1160static void kvm_disable_host_haltpoll(void *i)
1161{
1162 wrmsrq(MSR_KVM_POLL_CONTROL, val: 0);
1163}
1164
1165static void kvm_enable_host_haltpoll(void *i)
1166{
1167 wrmsrq(MSR_KVM_POLL_CONTROL, val: 1);
1168}
1169
1170void arch_haltpoll_enable(unsigned int cpu)
1171{
1172 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1173 pr_err_once("host does not support poll control\n");
1174 pr_err_once("host upgrade recommended\n");
1175 return;
1176 }
1177
1178 /* Enable guest halt poll disables host halt poll */
1179 smp_call_function_single(cpuid: cpu, func: kvm_disable_host_haltpoll, NULL, wait: 1);
1180}
1181EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1182
1183void arch_haltpoll_disable(unsigned int cpu)
1184{
1185 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1186 return;
1187
1188 /* Disable guest halt poll enables host halt poll */
1189 smp_call_function_single(cpuid: cpu, func: kvm_enable_host_haltpoll, NULL, wait: 1);
1190}
1191EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1192#endif
1193

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