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