Loading...
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(token, KVM_TASK_SLEEP_HASHBITS);
113 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
114 struct kvm_task_sleep_node *e;
115
116 raw_spin_lock(&b->lock);
117 e = _find_apf_task(b, token);
118 if (e) {
119 /* dummy entry exist -> wake up was delivered ahead of PF */
120 hlist_del(&e->link);
121 raw_spin_unlock(&b->lock);
122 kfree(e);
123 return false;
124 }
125
126 n->token = token;
127 n->cpu = smp_processor_id();
128 init_swait_queue_head(&n->wq);
129 hlist_add_head(&n->link, &b->list);
130 raw_spin_unlock(&b->lock);
131 return true;
132}
133
134/*
135 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
136 * @token: Token to identify the sleep node entry
137 *
138 * Invoked from the async pagefault handling code or from the VM exit page
139 * fault handler. In both cases RCU is watching.
140 */
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))
149 return;
150
151 for (;;) {
152 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
153 if (hlist_unhashed(&n.link))
154 break;
155
156 local_irq_enable();
157 schedule();
158 local_irq_disable();
159 }
160 finish_swait(&n.wq, &wait);
161}
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->link);
167 if (swq_has_sleeper(&n->wq))
168 swake_up_one(&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(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(sizeof(*dummy), GFP_ATOMIC);
213
214 /*
215 * Continue looping on allocation failure, eventually
216 * the async #PF will be handled and allocating a new
217 * node will be unnecessary.
218 */
219 if (!dummy)
220 cpu_relax();
221
222 /*
223 * Recheck for async #PF completion before enqueueing
224 * the dummy token to avoid duplicate list entries.
225 */
226 goto again;
227 }
228 dummy->token = token;
229 dummy->cpu = smp_processor_id();
230 init_swait_queue_head(&dummy->wq);
231 hlist_add_head(&dummy->link, &b->list);
232 dummy = NULL;
233 } else {
234 apf_task_wake_one(n);
235 }
236 raw_spin_unlock(&b->lock);
237
238 /* A dummy token might be allocated and ultimately not used. */
239 kfree(dummy);
240}
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("Host injected async #PF in interrupt disabled region\n");
274
275 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
276 if (unlikely(!(user_mode(regs))))
277 panic("Host injected async #PF in kernel mode\n");
278 /* Page is swapped out by the host. */
279 kvm_async_pf_task_wait_schedule(token);
280 } else {
281 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
282 }
283
284 instrumentation_end();
285 irqentry_exit(regs, state);
286 return true;
287}
288
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, 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, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
329 pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
330 (unsigned long long) slow_virt_to_phys(st));
331}
332
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, pa);
365 __this_cpu_write(apf_reason.enabled, 1);
366 pr_debug("setup async PF for cpu %d\n", smp_processor_id());
367 }
368
369 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
370 unsigned long pa;
371
372 /* Size alignment is implied but just to make it explicit. */
373 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
374 __this_cpu_write(kvm_apic_eoi, 0);
375 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
376 | KVM_MSR_ENABLED;
377 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
378 }
379
380 if (has_steal_clock)
381 kvm_register_steal_time();
382}
383
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, 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((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_vendor != CC_VENDOR_AMD ||
438 !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
439 return;
440
441 for_each_possible_cpu(cpu) {
442 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
443 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
444 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
445 }
446}
447
448static void kvm_guest_cpu_offline(bool shutdown)
449{
450 kvm_disable_steal_time();
451 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
452 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
453 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
454 wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
455 kvm_pv_disable_apf();
456 if (!shutdown)
457 apf_task_wake_all();
458 kvmclock_disable();
459}
460
461static int kvm_cpu_online(unsigned int cpu)
462{
463 unsigned long flags;
464
465 local_irq_save(flags);
466 kvm_guest_cpu_init();
467 local_irq_restore(flags);
468 return 0;
469}
470
471#ifdef CONFIG_SMP
472
473static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
474
475static bool pv_tlb_flush_supported(void)
476{
477 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
478 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
479 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
480 !boot_cpu_has(X86_FEATURE_MWAIT) &&
481 (num_possible_cpus() != 1));
482}
483
484static bool pv_ipi_supported(void)
485{
486 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
487 (num_possible_cpus() != 1));
488}
489
490static bool pv_sched_yield_supported(void)
491{
492 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
493 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
494 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
495 !boot_cpu_has(X86_FEATURE_MWAIT) &&
496 (num_possible_cpus() != 1));
497}
498
499#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
500
501static void __send_ipi_mask(const struct cpumask *mask, int vector)
502{
503 unsigned long flags;
504 int cpu, min = 0, max = 0;
505#ifdef CONFIG_X86_64
506 __uint128_t ipi_bitmap = 0;
507#else
508 u64 ipi_bitmap = 0;
509#endif
510 u32 apic_id, icr;
511 long ret;
512
513 if (cpumask_empty(mask))
514 return;
515
516 local_irq_save(flags);
517
518 switch (vector) {
519 default:
520 icr = APIC_DM_FIXED | vector;
521 break;
522 case NMI_VECTOR:
523 icr = APIC_DM_NMI;
524 break;
525 }
526
527 for_each_cpu(cpu, mask) {
528 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
529 if (!ipi_bitmap) {
530 min = max = apic_id;
531 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
532 ipi_bitmap <<= min - apic_id;
533 min = apic_id;
534 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
535 max = apic_id < max ? max : apic_id;
536 } else {
537 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
538 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
539 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
540 ret);
541 min = max = apic_id;
542 ipi_bitmap = 0;
543 }
544 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
545 }
546
547 if (ipi_bitmap) {
548 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
549 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
550 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
551 ret);
552 }
553
554 local_irq_restore(flags);
555}
556
557static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
558{
559 __send_ipi_mask(mask, vector);
560}
561
562static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
563{
564 unsigned int this_cpu = smp_processor_id();
565 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
566 const struct cpumask *local_mask;
567
568 cpumask_copy(new_mask, mask);
569 cpumask_clear_cpu(this_cpu, new_mask);
570 local_mask = new_mask;
571 __send_ipi_mask(local_mask, vector);
572}
573
574static int __init setup_efi_kvm_sev_migration(void)
575{
576 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
577 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
578 efi_status_t status;
579 unsigned long size;
580 bool enabled;
581
582 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
583 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
584 return 0;
585
586 if (!efi_enabled(EFI_BOOT))
587 return 0;
588
589 if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
590 pr_info("%s : EFI runtime services are not enabled\n", __func__);
591 return 0;
592 }
593
594 size = sizeof(enabled);
595
596 /* Get variable contents into buffer */
597 status = efi.get_variable(efi_sev_live_migration_enabled,
598 &efi_variable_guid, NULL, &size, &enabled);
599
600 if (status == EFI_NOT_FOUND) {
601 pr_info("%s : EFI live migration variable not found\n", __func__);
602 return 0;
603 }
604
605 if (status != EFI_SUCCESS) {
606 pr_info("%s : EFI variable retrieval failed\n", __func__);
607 return 0;
608 }
609
610 if (enabled == 0) {
611 pr_info("%s: live migration disabled in EFI\n", __func__);
612 return 0;
613 }
614
615 pr_info("%s : live migration enabled in EFI\n", __func__);
616 wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
617
618 return 1;
619}
620
621late_initcall(setup_efi_kvm_sev_migration);
622
623/*
624 * Set the IPI entry points
625 */
626static __init void kvm_setup_pv_ipi(void)
627{
628 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
629 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
630 pr_info("setup PV IPIs\n");
631}
632
633static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
634{
635 int cpu;
636
637 native_send_call_func_ipi(mask);
638
639 /* Make sure other vCPUs get a chance to run if they need to. */
640 for_each_cpu(cpu, mask) {
641 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
642 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
643 break;
644 }
645 }
646}
647
648static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
649 const struct flush_tlb_info *info)
650{
651 u8 state;
652 int cpu;
653 struct kvm_steal_time *src;
654 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
655
656 cpumask_copy(flushmask, cpumask);
657 /*
658 * We have to call flush only on online vCPUs. And
659 * queue flush_on_enter for pre-empted vCPUs
660 */
661 for_each_cpu(cpu, flushmask) {
662 /*
663 * The local vCPU is never preempted, so we do not explicitly
664 * skip check for local vCPU - it will never be cleared from
665 * flushmask.
666 */
667 src = &per_cpu(steal_time, cpu);
668 state = READ_ONCE(src->preempted);
669 if ((state & KVM_VCPU_PREEMPTED)) {
670 if (try_cmpxchg(&src->preempted, &state,
671 state | KVM_VCPU_FLUSH_TLB))
672 __cpumask_clear_cpu(cpu, flushmask);
673 }
674 }
675
676 native_flush_tlb_multi(flushmask, info);
677}
678
679static __init int kvm_alloc_cpumask(void)
680{
681 int cpu;
682
683 if (!kvm_para_available() || nopv)
684 return 0;
685
686 if (pv_tlb_flush_supported() || pv_ipi_supported())
687 for_each_possible_cpu(cpu) {
688 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
689 GFP_KERNEL, cpu_to_node(cpu));
690 }
691
692 return 0;
693}
694arch_initcall(kvm_alloc_cpumask);
695
696static void __init kvm_smp_prepare_boot_cpu(void)
697{
698 /*
699 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
700 * shares the guest physical address with the hypervisor.
701 */
702 sev_map_percpu_data();
703
704 kvm_guest_cpu_init();
705 native_smp_prepare_boot_cpu();
706 kvm_spinlock_init();
707}
708
709static int kvm_cpu_down_prepare(unsigned int cpu)
710{
711 unsigned long flags;
712
713 local_irq_save(flags);
714 kvm_guest_cpu_offline(false);
715 local_irq_restore(flags);
716 return 0;
717}
718
719#endif
720
721static int kvm_suspend(void)
722{
723 u64 val = 0;
724
725 kvm_guest_cpu_offline(false);
726
727#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
728 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
729 rdmsrl(MSR_KVM_POLL_CONTROL, val);
730 has_guest_poll = !(val & 1);
731#endif
732 return 0;
733}
734
735static void kvm_resume(void)
736{
737 kvm_cpu_online(raw_smp_processor_id());
738
739#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
740 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
741 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
742#endif
743}
744
745static struct syscore_ops kvm_syscore_ops = {
746 .suspend = kvm_suspend,
747 .resume = kvm_resume,
748};
749
750static void kvm_pv_guest_cpu_reboot(void *unused)
751{
752 kvm_guest_cpu_offline(true);
753}
754
755static int kvm_pv_reboot_notify(struct notifier_block *nb,
756 unsigned long code, void *unused)
757{
758 if (code == SYS_RESTART)
759 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
760 return NOTIFY_DONE;
761}
762
763static struct notifier_block kvm_pv_reboot_nb = {
764 .notifier_call = kvm_pv_reboot_notify,
765};
766
767/*
768 * After a PV feature is registered, the host will keep writing to the
769 * registered memory location. If the guest happens to shutdown, this memory
770 * won't be valid. In cases like kexec, in which you install a new kernel, this
771 * means a random memory location will be kept being written.
772 */
773#ifdef CONFIG_KEXEC_CORE
774static void kvm_crash_shutdown(struct pt_regs *regs)
775{
776 kvm_guest_cpu_offline(true);
777 native_machine_crash_shutdown(regs);
778}
779#endif
780
781#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
782bool __kvm_vcpu_is_preempted(long cpu);
783
784__visible bool __kvm_vcpu_is_preempted(long cpu)
785{
786 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
787
788 return !!(src->preempted & KVM_VCPU_PREEMPTED);
789}
790PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
791
792#else
793
794#include <asm/asm-offsets.h>
795
796extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
797
798/*
799 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
800 * restoring to/from the stack.
801 */
802#define PV_VCPU_PREEMPTED_ASM \
803 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
804 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
805 "setne %al\n\t"
806
807DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
808 PV_VCPU_PREEMPTED_ASM, .text);
809#endif
810
811static void __init kvm_guest_init(void)
812{
813 int i;
814
815 paravirt_ops_setup();
816 register_reboot_notifier(&kvm_pv_reboot_nb);
817 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
818 raw_spin_lock_init(&async_pf_sleepers[i].lock);
819
820 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
821 has_steal_clock = 1;
822 static_call_update(pv_steal_clock, kvm_steal_clock);
823
824 pv_ops.lock.vcpu_is_preempted =
825 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
826 }
827
828 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
829 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
830
831 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
832 static_branch_enable(&kvm_async_pf_enabled);
833 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
834 }
835
836#ifdef CONFIG_SMP
837 if (pv_tlb_flush_supported()) {
838 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
839 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
840 pr_info("KVM setup pv remote TLB flush\n");
841 }
842
843 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
844 if (pv_sched_yield_supported()) {
845 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
846 pr_info("setup PV sched yield\n");
847 }
848 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
849 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
850 pr_err("failed to install cpu hotplug callbacks\n");
851#else
852 sev_map_percpu_data();
853 kvm_guest_cpu_init();
854#endif
855
856#ifdef CONFIG_KEXEC_CORE
857 machine_ops.crash_shutdown = kvm_crash_shutdown;
858#endif
859
860 register_syscore_ops(&kvm_syscore_ops);
861
862 /*
863 * Hard lockup detection is enabled by default. Disable it, as guests
864 * can get false positives too easily, for example if the host is
865 * overcommitted.
866 */
867 hardlockup_detector_disable();
868}
869
870static noinline uint32_t __kvm_cpuid_base(void)
871{
872 if (boot_cpu_data.cpuid_level < 0)
873 return 0; /* So we don't blow up on old processors */
874
875 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
876 return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
877
878 return 0;
879}
880
881static inline uint32_t kvm_cpuid_base(void)
882{
883 static int kvm_cpuid_base = -1;
884
885 if (kvm_cpuid_base == -1)
886 kvm_cpuid_base = __kvm_cpuid_base();
887
888 return kvm_cpuid_base;
889}
890
891bool kvm_para_available(void)
892{
893 return kvm_cpuid_base() != 0;
894}
895EXPORT_SYMBOL_GPL(kvm_para_available);
896
897unsigned int kvm_arch_para_features(void)
898{
899 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
900}
901
902unsigned int kvm_arch_para_hints(void)
903{
904 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
905}
906EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
907
908static uint32_t __init kvm_detect(void)
909{
910 return kvm_cpuid_base();
911}
912
913static void __init kvm_apic_init(void)
914{
915#ifdef CONFIG_SMP
916 if (pv_ipi_supported())
917 kvm_setup_pv_ipi();
918#endif
919}
920
921static bool __init kvm_msi_ext_dest_id(void)
922{
923 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
924}
925
926static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
927{
928 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
929 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
930}
931
932static void __init kvm_init_platform(void)
933{
934 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
935 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
936 unsigned long nr_pages;
937 int i;
938
939 pv_ops.mmu.notify_page_enc_status_changed =
940 kvm_sev_hc_page_enc_status;
941
942 /*
943 * Reset the host's shared pages list related to kernel
944 * specific page encryption status settings before we load a
945 * new kernel by kexec. Reset the page encryption status
946 * during early boot instead of just before kexec to avoid SMP
947 * races during kvm_pv_guest_cpu_reboot().
948 * NOTE: We cannot reset the complete shared pages list
949 * here as we need to retain the UEFI/OVMF firmware
950 * specific settings.
951 */
952
953 for (i = 0; i < e820_table->nr_entries; i++) {
954 struct e820_entry *entry = &e820_table->entries[i];
955
956 if (entry->type != E820_TYPE_RAM)
957 continue;
958
959 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
960
961 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
962 nr_pages,
963 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
964 }
965
966 /*
967 * Ensure that _bss_decrypted section is marked as decrypted in the
968 * shared pages list.
969 */
970 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
971 __end_bss_decrypted - __start_bss_decrypted, 0);
972
973 /*
974 * If not booted using EFI, enable Live migration support.
975 */
976 if (!efi_enabled(EFI_BOOT))
977 wrmsrl(MSR_KVM_MIGRATION_CONTROL,
978 KVM_MIGRATION_READY);
979 }
980 kvmclock_init();
981 x86_platform.apic_post_init = kvm_apic_init;
982}
983
984#if defined(CONFIG_AMD_MEM_ENCRYPT)
985static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
986{
987 /* RAX and CPL are already in the GHCB */
988 ghcb_set_rbx(ghcb, regs->bx);
989 ghcb_set_rcx(ghcb, regs->cx);
990 ghcb_set_rdx(ghcb, regs->dx);
991 ghcb_set_rsi(ghcb, regs->si);
992}
993
994static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
995{
996 /* No checking of the return state needed */
997 return true;
998}
999#endif
1000
1001const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1002 .name = "KVM",
1003 .detect = kvm_detect,
1004 .type = X86_HYPER_KVM,
1005 .init.guest_late_init = kvm_guest_init,
1006 .init.x2apic_available = kvm_para_available,
1007 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1008 .init.init_platform = kvm_init_platform,
1009#if defined(CONFIG_AMD_MEM_ENCRYPT)
1010 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1011 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1012#endif
1013};
1014
1015static __init int activate_jump_labels(void)
1016{
1017 if (has_steal_clock) {
1018 static_key_slow_inc(¶virt_steal_enabled);
1019 if (steal_acc)
1020 static_key_slow_inc(¶virt_steal_rq_enabled);
1021 }
1022
1023 return 0;
1024}
1025arch_initcall(activate_jump_labels);
1026
1027#ifdef CONFIG_PARAVIRT_SPINLOCKS
1028
1029/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1030static void kvm_kick_cpu(int cpu)
1031{
1032 unsigned long flags = 0;
1033 u32 apicid;
1034
1035 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1036 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1037}
1038
1039#include <asm/qspinlock.h>
1040
1041static void kvm_wait(u8 *ptr, u8 val)
1042{
1043 if (in_nmi())
1044 return;
1045
1046 /*
1047 * halt until it's our turn and kicked. Note that we do safe halt
1048 * for irq enabled case to avoid hang when lock info is overwritten
1049 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1050 */
1051 if (irqs_disabled()) {
1052 if (READ_ONCE(*ptr) == val)
1053 halt();
1054 } else {
1055 local_irq_disable();
1056
1057 /* safe_halt() will enable IRQ */
1058 if (READ_ONCE(*ptr) == val)
1059 safe_halt();
1060 else
1061 local_irq_enable();
1062 }
1063}
1064
1065/*
1066 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1067 */
1068void __init kvm_spinlock_init(void)
1069{
1070 /*
1071 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1072 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1073 * preferred over native qspinlock when vCPU is preempted.
1074 */
1075 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1076 pr_info("PV spinlocks disabled, no host support\n");
1077 return;
1078 }
1079
1080 /*
1081 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1082 * are available.
1083 */
1084 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1085 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1086 goto out;
1087 }
1088
1089 if (num_possible_cpus() == 1) {
1090 pr_info("PV spinlocks disabled, single CPU\n");
1091 goto out;
1092 }
1093
1094 if (nopvspin) {
1095 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1096 goto out;
1097 }
1098
1099 pr_info("PV spinlocks enabled\n");
1100
1101 __pv_init_lock_hash();
1102 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1103 pv_ops.lock.queued_spin_unlock =
1104 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1105 pv_ops.lock.wait = kvm_wait;
1106 pv_ops.lock.kick = kvm_kick_cpu;
1107
1108 /*
1109 * When PV spinlock is enabled which is preferred over
1110 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1111 * Just disable it anyway.
1112 */
1113out:
1114 static_branch_disable(&virt_spin_lock_key);
1115}
1116
1117#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1118
1119#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1120
1121static void kvm_disable_host_haltpoll(void *i)
1122{
1123 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1124}
1125
1126static void kvm_enable_host_haltpoll(void *i)
1127{
1128 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1129}
1130
1131void arch_haltpoll_enable(unsigned int cpu)
1132{
1133 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1134 pr_err_once("host does not support poll control\n");
1135 pr_err_once("host upgrade recommended\n");
1136 return;
1137 }
1138
1139 /* Enable guest halt poll disables host halt poll */
1140 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1141}
1142EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1143
1144void arch_haltpoll_disable(unsigned int cpu)
1145{
1146 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1147 return;
1148
1149 /* Disable guest halt poll enables host halt poll */
1150 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1151}
1152EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1153#endif
1/*
2 * KVM paravirt_ops implementation
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
17 *
18 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
19 * Copyright IBM Corporation, 2007
20 * Authors: Anthony Liguori <aliguori@us.ibm.com>
21 */
22
23#include <linux/context_tracking.h>
24#include <linux/module.h>
25#include <linux/kernel.h>
26#include <linux/kvm_para.h>
27#include <linux/cpu.h>
28#include <linux/mm.h>
29#include <linux/highmem.h>
30#include <linux/hardirq.h>
31#include <linux/notifier.h>
32#include <linux/reboot.h>
33#include <linux/hash.h>
34#include <linux/sched.h>
35#include <linux/slab.h>
36#include <linux/kprobes.h>
37#include <linux/debugfs.h>
38#include <linux/nmi.h>
39#include <linux/swait.h>
40#include <asm/timer.h>
41#include <asm/cpu.h>
42#include <asm/traps.h>
43#include <asm/desc.h>
44#include <asm/tlbflush.h>
45#include <asm/idle.h>
46#include <asm/apic.h>
47#include <asm/apicdef.h>
48#include <asm/hypervisor.h>
49#include <asm/kvm_guest.h>
50
51static int kvmapf = 1;
52
53static int 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 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 int kvmclock_vsyscall = 1;
71static int parse_no_kvmclock_vsyscall(char *arg)
72{
73 kvmclock_vsyscall = 0;
74 return 0;
75}
76
77early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
78
79static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
80static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
81static int has_steal_clock = 0;
82
83/*
84 * No need for any "IO delay" on KVM
85 */
86static void kvm_io_delay(void)
87{
88}
89
90#define KVM_TASK_SLEEP_HASHBITS 8
91#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
92
93struct kvm_task_sleep_node {
94 struct hlist_node link;
95 struct swait_queue_head wq;
96 u32 token;
97 int cpu;
98 bool halted;
99};
100
101static struct kvm_task_sleep_head {
102 raw_spinlock_t lock;
103 struct hlist_head list;
104} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
105
106static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
107 u32 token)
108{
109 struct hlist_node *p;
110
111 hlist_for_each(p, &b->list) {
112 struct kvm_task_sleep_node *n =
113 hlist_entry(p, typeof(*n), link);
114 if (n->token == token)
115 return n;
116 }
117
118 return NULL;
119}
120
121void kvm_async_pf_task_wait(u32 token)
122{
123 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
124 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
125 struct kvm_task_sleep_node n, *e;
126 DECLARE_SWAITQUEUE(wait);
127
128 rcu_irq_enter();
129
130 raw_spin_lock(&b->lock);
131 e = _find_apf_task(b, token);
132 if (e) {
133 /* dummy entry exist -> wake up was delivered ahead of PF */
134 hlist_del(&e->link);
135 kfree(e);
136 raw_spin_unlock(&b->lock);
137
138 rcu_irq_exit();
139 return;
140 }
141
142 n.token = token;
143 n.cpu = smp_processor_id();
144 n.halted = is_idle_task(current) || preempt_count() > 1;
145 init_swait_queue_head(&n.wq);
146 hlist_add_head(&n.link, &b->list);
147 raw_spin_unlock(&b->lock);
148
149 for (;;) {
150 if (!n.halted)
151 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
152 if (hlist_unhashed(&n.link))
153 break;
154
155 if (!n.halted) {
156 local_irq_enable();
157 schedule();
158 local_irq_disable();
159 } else {
160 /*
161 * We cannot reschedule. So halt.
162 */
163 rcu_irq_exit();
164 native_safe_halt();
165 rcu_irq_enter();
166 local_irq_disable();
167 }
168 }
169 if (!n.halted)
170 finish_swait(&n.wq, &wait);
171
172 rcu_irq_exit();
173 return;
174}
175EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
176
177static void apf_task_wake_one(struct kvm_task_sleep_node *n)
178{
179 hlist_del_init(&n->link);
180 if (n->halted)
181 smp_send_reschedule(n->cpu);
182 else if (swait_active(&n->wq))
183 swake_up(&n->wq);
184}
185
186static void apf_task_wake_all(void)
187{
188 int i;
189
190 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
191 struct hlist_node *p, *next;
192 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
193 raw_spin_lock(&b->lock);
194 hlist_for_each_safe(p, next, &b->list) {
195 struct kvm_task_sleep_node *n =
196 hlist_entry(p, typeof(*n), link);
197 if (n->cpu == smp_processor_id())
198 apf_task_wake_one(n);
199 }
200 raw_spin_unlock(&b->lock);
201 }
202}
203
204void kvm_async_pf_task_wake(u32 token)
205{
206 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
207 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
208 struct kvm_task_sleep_node *n;
209
210 if (token == ~0) {
211 apf_task_wake_all();
212 return;
213 }
214
215again:
216 raw_spin_lock(&b->lock);
217 n = _find_apf_task(b, token);
218 if (!n) {
219 /*
220 * async PF was not yet handled.
221 * Add dummy entry for the token.
222 */
223 n = kzalloc(sizeof(*n), GFP_ATOMIC);
224 if (!n) {
225 /*
226 * Allocation failed! Busy wait while other cpu
227 * handles async PF.
228 */
229 raw_spin_unlock(&b->lock);
230 cpu_relax();
231 goto again;
232 }
233 n->token = token;
234 n->cpu = smp_processor_id();
235 init_swait_queue_head(&n->wq);
236 hlist_add_head(&n->link, &b->list);
237 } else
238 apf_task_wake_one(n);
239 raw_spin_unlock(&b->lock);
240 return;
241}
242EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243
244u32 kvm_read_and_reset_pf_reason(void)
245{
246 u32 reason = 0;
247
248 if (__this_cpu_read(apf_reason.enabled)) {
249 reason = __this_cpu_read(apf_reason.reason);
250 __this_cpu_write(apf_reason.reason, 0);
251 }
252
253 return reason;
254}
255EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
256NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
257
258dotraplinkage void
259do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
260{
261 enum ctx_state prev_state;
262
263 switch (kvm_read_and_reset_pf_reason()) {
264 default:
265 trace_do_page_fault(regs, error_code);
266 break;
267 case KVM_PV_REASON_PAGE_NOT_PRESENT:
268 /* page is swapped out by the host. */
269 prev_state = exception_enter();
270 exit_idle();
271 kvm_async_pf_task_wait((u32)read_cr2());
272 exception_exit(prev_state);
273 break;
274 case KVM_PV_REASON_PAGE_READY:
275 rcu_irq_enter();
276 exit_idle();
277 kvm_async_pf_task_wake((u32)read_cr2());
278 rcu_irq_exit();
279 break;
280 }
281}
282NOKPROBE_SYMBOL(do_async_page_fault);
283
284static void __init paravirt_ops_setup(void)
285{
286 pv_info.name = "KVM";
287
288 /*
289 * KVM isn't paravirt in the sense of paravirt_enabled. A KVM
290 * guest kernel works like a bare metal kernel with additional
291 * features, and paravirt_enabled is about features that are
292 * missing.
293 */
294 pv_info.paravirt_enabled = 0;
295
296 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
297 pv_cpu_ops.io_delay = kvm_io_delay;
298
299#ifdef CONFIG_X86_IO_APIC
300 no_timer_check = 1;
301#endif
302}
303
304static void kvm_register_steal_time(void)
305{
306 int cpu = smp_processor_id();
307 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
308
309 if (!has_steal_clock)
310 return;
311
312 memset(st, 0, sizeof(*st));
313
314 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
315 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
316 cpu, (unsigned long long) slow_virt_to_phys(st));
317}
318
319static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
320
321static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
322{
323 /**
324 * This relies on __test_and_clear_bit to modify the memory
325 * in a way that is atomic with respect to the local CPU.
326 * The hypervisor only accesses this memory from the local CPU so
327 * there's no need for lock or memory barriers.
328 * An optimization barrier is implied in apic write.
329 */
330 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
331 return;
332 apic_write(APIC_EOI, APIC_EOI_ACK);
333}
334
335static void kvm_guest_cpu_init(void)
336{
337 if (!kvm_para_available())
338 return;
339
340 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
341 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
342
343#ifdef CONFIG_PREEMPT
344 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
345#endif
346 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
347 __this_cpu_write(apf_reason.enabled, 1);
348 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
349 smp_processor_id());
350 }
351
352 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
353 unsigned long pa;
354 /* Size alignment is implied but just to make it explicit. */
355 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
356 __this_cpu_write(kvm_apic_eoi, 0);
357 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
358 | KVM_MSR_ENABLED;
359 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
360 }
361
362 if (has_steal_clock)
363 kvm_register_steal_time();
364}
365
366static void kvm_pv_disable_apf(void)
367{
368 if (!__this_cpu_read(apf_reason.enabled))
369 return;
370
371 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
372 __this_cpu_write(apf_reason.enabled, 0);
373
374 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
375 smp_processor_id());
376}
377
378static void kvm_pv_guest_cpu_reboot(void *unused)
379{
380 /*
381 * We disable PV EOI before we load a new kernel by kexec,
382 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
383 * New kernel can re-enable when it boots.
384 */
385 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
386 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
387 kvm_pv_disable_apf();
388 kvm_disable_steal_time();
389}
390
391static int kvm_pv_reboot_notify(struct notifier_block *nb,
392 unsigned long code, void *unused)
393{
394 if (code == SYS_RESTART)
395 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
396 return NOTIFY_DONE;
397}
398
399static struct notifier_block kvm_pv_reboot_nb = {
400 .notifier_call = kvm_pv_reboot_notify,
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 rmb();
413 steal = src->steal;
414 rmb();
415 } while ((version & 1) || (version != src->version));
416
417 return steal;
418}
419
420void kvm_disable_steal_time(void)
421{
422 if (!has_steal_clock)
423 return;
424
425 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
426}
427
428#ifdef CONFIG_SMP
429static void __init kvm_smp_prepare_boot_cpu(void)
430{
431 kvm_guest_cpu_init();
432 native_smp_prepare_boot_cpu();
433 kvm_spinlock_init();
434}
435
436static void kvm_guest_cpu_online(void *dummy)
437{
438 kvm_guest_cpu_init();
439}
440
441static void kvm_guest_cpu_offline(void *dummy)
442{
443 kvm_disable_steal_time();
444 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
445 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
446 kvm_pv_disable_apf();
447 apf_task_wake_all();
448}
449
450static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
451 void *hcpu)
452{
453 int cpu = (unsigned long)hcpu;
454 switch (action) {
455 case CPU_ONLINE:
456 case CPU_DOWN_FAILED:
457 case CPU_ONLINE_FROZEN:
458 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
459 break;
460 case CPU_DOWN_PREPARE:
461 case CPU_DOWN_PREPARE_FROZEN:
462 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
463 break;
464 default:
465 break;
466 }
467 return NOTIFY_OK;
468}
469
470static struct notifier_block kvm_cpu_notifier = {
471 .notifier_call = kvm_cpu_notify,
472};
473#endif
474
475static void __init kvm_apf_trap_init(void)
476{
477 set_intr_gate(14, async_page_fault);
478}
479
480void __init kvm_guest_init(void)
481{
482 int i;
483
484 if (!kvm_para_available())
485 return;
486
487 paravirt_ops_setup();
488 register_reboot_notifier(&kvm_pv_reboot_nb);
489 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
490 raw_spin_lock_init(&async_pf_sleepers[i].lock);
491 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
492 x86_init.irqs.trap_init = kvm_apf_trap_init;
493
494 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
495 has_steal_clock = 1;
496 pv_time_ops.steal_clock = kvm_steal_clock;
497 }
498
499 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
500 apic_set_eoi_write(kvm_guest_apic_eoi_write);
501
502 if (kvmclock_vsyscall)
503 kvm_setup_vsyscall_timeinfo();
504
505#ifdef CONFIG_SMP
506 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
507 register_cpu_notifier(&kvm_cpu_notifier);
508#else
509 kvm_guest_cpu_init();
510#endif
511
512 /*
513 * Hard lockup detection is enabled by default. Disable it, as guests
514 * can get false positives too easily, for example if the host is
515 * overcommitted.
516 */
517 hardlockup_detector_disable();
518}
519
520static noinline uint32_t __kvm_cpuid_base(void)
521{
522 if (boot_cpu_data.cpuid_level < 0)
523 return 0; /* So we don't blow up on old processors */
524
525 if (cpu_has_hypervisor)
526 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
527
528 return 0;
529}
530
531static inline uint32_t kvm_cpuid_base(void)
532{
533 static int kvm_cpuid_base = -1;
534
535 if (kvm_cpuid_base == -1)
536 kvm_cpuid_base = __kvm_cpuid_base();
537
538 return kvm_cpuid_base;
539}
540
541bool kvm_para_available(void)
542{
543 return kvm_cpuid_base() != 0;
544}
545EXPORT_SYMBOL_GPL(kvm_para_available);
546
547unsigned int kvm_arch_para_features(void)
548{
549 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
550}
551
552static uint32_t __init kvm_detect(void)
553{
554 return kvm_cpuid_base();
555}
556
557const struct hypervisor_x86 x86_hyper_kvm __refconst = {
558 .name = "KVM",
559 .detect = kvm_detect,
560 .x2apic_available = kvm_para_available,
561};
562EXPORT_SYMBOL_GPL(x86_hyper_kvm);
563
564static __init int activate_jump_labels(void)
565{
566 if (has_steal_clock) {
567 static_key_slow_inc(¶virt_steal_enabled);
568 if (steal_acc)
569 static_key_slow_inc(¶virt_steal_rq_enabled);
570 }
571
572 return 0;
573}
574arch_initcall(activate_jump_labels);
575
576#ifdef CONFIG_PARAVIRT_SPINLOCKS
577
578/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
579static void kvm_kick_cpu(int cpu)
580{
581 int apicid;
582 unsigned long flags = 0;
583
584 apicid = per_cpu(x86_cpu_to_apicid, cpu);
585 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
586}
587
588
589#ifdef CONFIG_QUEUED_SPINLOCKS
590
591#include <asm/qspinlock.h>
592
593static void kvm_wait(u8 *ptr, u8 val)
594{
595 unsigned long flags;
596
597 if (in_nmi())
598 return;
599
600 local_irq_save(flags);
601
602 if (READ_ONCE(*ptr) != val)
603 goto out;
604
605 /*
606 * halt until it's our turn and kicked. Note that we do safe halt
607 * for irq enabled case to avoid hang when lock info is overwritten
608 * in irq spinlock slowpath and no spurious interrupt occur to save us.
609 */
610 if (arch_irqs_disabled_flags(flags))
611 halt();
612 else
613 safe_halt();
614
615out:
616 local_irq_restore(flags);
617}
618
619#else /* !CONFIG_QUEUED_SPINLOCKS */
620
621enum kvm_contention_stat {
622 TAKEN_SLOW,
623 TAKEN_SLOW_PICKUP,
624 RELEASED_SLOW,
625 RELEASED_SLOW_KICKED,
626 NR_CONTENTION_STATS
627};
628
629#ifdef CONFIG_KVM_DEBUG_FS
630#define HISTO_BUCKETS 30
631
632static struct kvm_spinlock_stats
633{
634 u32 contention_stats[NR_CONTENTION_STATS];
635 u32 histo_spin_blocked[HISTO_BUCKETS+1];
636 u64 time_blocked;
637} spinlock_stats;
638
639static u8 zero_stats;
640
641static inline void check_zero(void)
642{
643 u8 ret;
644 u8 old;
645
646 old = READ_ONCE(zero_stats);
647 if (unlikely(old)) {
648 ret = cmpxchg(&zero_stats, old, 0);
649 /* This ensures only one fellow resets the stat */
650 if (ret == old)
651 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
652 }
653}
654
655static inline void add_stats(enum kvm_contention_stat var, u32 val)
656{
657 check_zero();
658 spinlock_stats.contention_stats[var] += val;
659}
660
661
662static inline u64 spin_time_start(void)
663{
664 return sched_clock();
665}
666
667static void __spin_time_accum(u64 delta, u32 *array)
668{
669 unsigned index;
670
671 index = ilog2(delta);
672 check_zero();
673
674 if (index < HISTO_BUCKETS)
675 array[index]++;
676 else
677 array[HISTO_BUCKETS]++;
678}
679
680static inline void spin_time_accum_blocked(u64 start)
681{
682 u32 delta;
683
684 delta = sched_clock() - start;
685 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
686 spinlock_stats.time_blocked += delta;
687}
688
689static struct dentry *d_spin_debug;
690static struct dentry *d_kvm_debug;
691
692static struct dentry *kvm_init_debugfs(void)
693{
694 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
695 if (!d_kvm_debug)
696 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
697
698 return d_kvm_debug;
699}
700
701static int __init kvm_spinlock_debugfs(void)
702{
703 struct dentry *d_kvm;
704
705 d_kvm = kvm_init_debugfs();
706 if (d_kvm == NULL)
707 return -ENOMEM;
708
709 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
710
711 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
712
713 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
714 &spinlock_stats.contention_stats[TAKEN_SLOW]);
715 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
716 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
717
718 debugfs_create_u32("released_slow", 0444, d_spin_debug,
719 &spinlock_stats.contention_stats[RELEASED_SLOW]);
720 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
721 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
722
723 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
724 &spinlock_stats.time_blocked);
725
726 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
727 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
728
729 return 0;
730}
731fs_initcall(kvm_spinlock_debugfs);
732#else /* !CONFIG_KVM_DEBUG_FS */
733static inline void add_stats(enum kvm_contention_stat var, u32 val)
734{
735}
736
737static inline u64 spin_time_start(void)
738{
739 return 0;
740}
741
742static inline void spin_time_accum_blocked(u64 start)
743{
744}
745#endif /* CONFIG_KVM_DEBUG_FS */
746
747struct kvm_lock_waiting {
748 struct arch_spinlock *lock;
749 __ticket_t want;
750};
751
752/* cpus 'waiting' on a spinlock to become available */
753static cpumask_t waiting_cpus;
754
755/* Track spinlock on which a cpu is waiting */
756static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
757
758__visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
759{
760 struct kvm_lock_waiting *w;
761 int cpu;
762 u64 start;
763 unsigned long flags;
764 __ticket_t head;
765
766 if (in_nmi())
767 return;
768
769 w = this_cpu_ptr(&klock_waiting);
770 cpu = smp_processor_id();
771 start = spin_time_start();
772
773 /*
774 * Make sure an interrupt handler can't upset things in a
775 * partially setup state.
776 */
777 local_irq_save(flags);
778
779 /*
780 * The ordering protocol on this is that the "lock" pointer
781 * may only be set non-NULL if the "want" ticket is correct.
782 * If we're updating "want", we must first clear "lock".
783 */
784 w->lock = NULL;
785 smp_wmb();
786 w->want = want;
787 smp_wmb();
788 w->lock = lock;
789
790 add_stats(TAKEN_SLOW, 1);
791
792 /*
793 * This uses set_bit, which is atomic but we should not rely on its
794 * reordering gurantees. So barrier is needed after this call.
795 */
796 cpumask_set_cpu(cpu, &waiting_cpus);
797
798 barrier();
799
800 /*
801 * Mark entry to slowpath before doing the pickup test to make
802 * sure we don't deadlock with an unlocker.
803 */
804 __ticket_enter_slowpath(lock);
805
806 /* make sure enter_slowpath, which is atomic does not cross the read */
807 smp_mb__after_atomic();
808
809 /*
810 * check again make sure it didn't become free while
811 * we weren't looking.
812 */
813 head = READ_ONCE(lock->tickets.head);
814 if (__tickets_equal(head, want)) {
815 add_stats(TAKEN_SLOW_PICKUP, 1);
816 goto out;
817 }
818
819 /*
820 * halt until it's our turn and kicked. Note that we do safe halt
821 * for irq enabled case to avoid hang when lock info is overwritten
822 * in irq spinlock slowpath and no spurious interrupt occur to save us.
823 */
824 if (arch_irqs_disabled_flags(flags))
825 halt();
826 else
827 safe_halt();
828
829out:
830 cpumask_clear_cpu(cpu, &waiting_cpus);
831 w->lock = NULL;
832 local_irq_restore(flags);
833 spin_time_accum_blocked(start);
834}
835PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
836
837/* Kick vcpu waiting on @lock->head to reach value @ticket */
838static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
839{
840 int cpu;
841
842 add_stats(RELEASED_SLOW, 1);
843 for_each_cpu(cpu, &waiting_cpus) {
844 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
845 if (READ_ONCE(w->lock) == lock &&
846 READ_ONCE(w->want) == ticket) {
847 add_stats(RELEASED_SLOW_KICKED, 1);
848 kvm_kick_cpu(cpu);
849 break;
850 }
851 }
852}
853
854#endif /* !CONFIG_QUEUED_SPINLOCKS */
855
856/*
857 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
858 */
859void __init kvm_spinlock_init(void)
860{
861 if (!kvm_para_available())
862 return;
863 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
864 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
865 return;
866
867#ifdef CONFIG_QUEUED_SPINLOCKS
868 __pv_init_lock_hash();
869 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
870 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
871 pv_lock_ops.wait = kvm_wait;
872 pv_lock_ops.kick = kvm_kick_cpu;
873#else /* !CONFIG_QUEUED_SPINLOCKS */
874 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
875 pv_lock_ops.unlock_kick = kvm_unlock_kick;
876#endif
877}
878
879static __init int kvm_spinlock_init_jump(void)
880{
881 if (!kvm_para_available())
882 return 0;
883 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
884 return 0;
885
886 static_key_slow_inc(¶virt_ticketlocks_enabled);
887 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
888
889 return 0;
890}
891early_initcall(kvm_spinlock_init_jump);
892
893#endif /* CONFIG_PARAVIRT_SPINLOCKS */