Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[muen/linux.git] / arch / x86 / kernel / kvm.c
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/init.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/apic.h>
46 #include <asm/apicdef.h>
47 #include <asm/hypervisor.h>
48
49 static int kvmapf = 1;
50
51 static int __init parse_no_kvmapf(char *arg)
52 {
53         kvmapf = 0;
54         return 0;
55 }
56
57 early_param("no-kvmapf", parse_no_kvmapf);
58
59 static int steal_acc = 1;
60 static int __init parse_no_stealacc(char *arg)
61 {
62         steal_acc = 0;
63         return 0;
64 }
65
66 early_param("no-steal-acc", parse_no_stealacc);
67
68 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
69 static DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64);
70 static int has_steal_clock = 0;
71
72 /*
73  * No need for any "IO delay" on KVM
74  */
75 static void kvm_io_delay(void)
76 {
77 }
78
79 #define KVM_TASK_SLEEP_HASHBITS 8
80 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
81
82 struct kvm_task_sleep_node {
83         struct hlist_node link;
84         struct swait_queue_head wq;
85         u32 token;
86         int cpu;
87         bool halted;
88 };
89
90 static struct kvm_task_sleep_head {
91         raw_spinlock_t lock;
92         struct hlist_head list;
93 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
94
95 static 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
110 /*
111  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
112  *                    (other than user space)?
113  */
114 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
115 {
116         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
117         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
118         struct kvm_task_sleep_node n, *e;
119         DECLARE_SWAITQUEUE(wait);
120
121         rcu_irq_enter();
122
123         raw_spin_lock(&b->lock);
124         e = _find_apf_task(b, token);
125         if (e) {
126                 /* dummy entry exist -> wake up was delivered ahead of PF */
127                 hlist_del(&e->link);
128                 kfree(e);
129                 raw_spin_unlock(&b->lock);
130
131                 rcu_irq_exit();
132                 return;
133         }
134
135         n.token = token;
136         n.cpu = smp_processor_id();
137         n.halted = is_idle_task(current) ||
138                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
139                     ? preempt_count() > 1 || rcu_preempt_depth()
140                     : interrupt_kernel);
141         init_swait_queue_head(&n.wq);
142         hlist_add_head(&n.link, &b->list);
143         raw_spin_unlock(&b->lock);
144
145         for (;;) {
146                 if (!n.halted)
147                         prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
148                 if (hlist_unhashed(&n.link))
149                         break;
150
151                 rcu_irq_exit();
152
153                 if (!n.halted) {
154                         local_irq_enable();
155                         schedule();
156                         local_irq_disable();
157                 } else {
158                         /*
159                          * We cannot reschedule. So halt.
160                          */
161                         native_safe_halt();
162                         local_irq_disable();
163                 }
164
165                 rcu_irq_enter();
166         }
167         if (!n.halted)
168                 finish_swait(&n.wq, &wait);
169
170         rcu_irq_exit();
171         return;
172 }
173 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
174
175 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
176 {
177         hlist_del_init(&n->link);
178         if (n->halted)
179                 smp_send_reschedule(n->cpu);
180         else if (swq_has_sleeper(&n->wq))
181                 swake_up_one(&n->wq);
182 }
183
184 static void apf_task_wake_all(void)
185 {
186         int i;
187
188         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
189                 struct hlist_node *p, *next;
190                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
191                 raw_spin_lock(&b->lock);
192                 hlist_for_each_safe(p, next, &b->list) {
193                         struct kvm_task_sleep_node *n =
194                                 hlist_entry(p, typeof(*n), link);
195                         if (n->cpu == smp_processor_id())
196                                 apf_task_wake_one(n);
197                 }
198                 raw_spin_unlock(&b->lock);
199         }
200 }
201
202 void kvm_async_pf_task_wake(u32 token)
203 {
204         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
205         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
206         struct kvm_task_sleep_node *n;
207
208         if (token == ~0) {
209                 apf_task_wake_all();
210                 return;
211         }
212
213 again:
214         raw_spin_lock(&b->lock);
215         n = _find_apf_task(b, token);
216         if (!n) {
217                 /*
218                  * async PF was not yet handled.
219                  * Add dummy entry for the token.
220                  */
221                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
222                 if (!n) {
223                         /*
224                          * Allocation failed! Busy wait while other cpu
225                          * handles async PF.
226                          */
227                         raw_spin_unlock(&b->lock);
228                         cpu_relax();
229                         goto again;
230                 }
231                 n->token = token;
232                 n->cpu = smp_processor_id();
233                 init_swait_queue_head(&n->wq);
234                 hlist_add_head(&n->link, &b->list);
235         } else
236                 apf_task_wake_one(n);
237         raw_spin_unlock(&b->lock);
238         return;
239 }
240 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
241
242 u32 kvm_read_and_reset_pf_reason(void)
243 {
244         u32 reason = 0;
245
246         if (__this_cpu_read(apf_reason.enabled)) {
247                 reason = __this_cpu_read(apf_reason.reason);
248                 __this_cpu_write(apf_reason.reason, 0);
249         }
250
251         return reason;
252 }
253 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
254 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
255
256 dotraplinkage void
257 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
258 {
259         enum ctx_state prev_state;
260
261         switch (kvm_read_and_reset_pf_reason()) {
262         default:
263                 do_page_fault(regs, error_code);
264                 break;
265         case KVM_PV_REASON_PAGE_NOT_PRESENT:
266                 /* page is swapped out by the host. */
267                 prev_state = exception_enter();
268                 kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
269                 exception_exit(prev_state);
270                 break;
271         case KVM_PV_REASON_PAGE_READY:
272                 rcu_irq_enter();
273                 kvm_async_pf_task_wake((u32)read_cr2());
274                 rcu_irq_exit();
275                 break;
276         }
277 }
278 NOKPROBE_SYMBOL(do_async_page_fault);
279
280 static void __init paravirt_ops_setup(void)
281 {
282         pv_info.name = "KVM";
283
284         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
285                 pv_cpu_ops.io_delay = kvm_io_delay;
286
287 #ifdef CONFIG_X86_IO_APIC
288         no_timer_check = 1;
289 #endif
290 }
291
292 static void kvm_register_steal_time(void)
293 {
294         int cpu = smp_processor_id();
295         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
296
297         if (!has_steal_clock)
298                 return;
299
300         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
301         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
302                 cpu, (unsigned long long) slow_virt_to_phys(st));
303 }
304
305 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
306
307 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
308 {
309         /**
310          * This relies on __test_and_clear_bit to modify the memory
311          * in a way that is atomic with respect to the local CPU.
312          * The hypervisor only accesses this memory from the local CPU so
313          * there's no need for lock or memory barriers.
314          * An optimization barrier is implied in apic write.
315          */
316         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
317                 return;
318         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
319 }
320
321 static void kvm_guest_cpu_init(void)
322 {
323         if (!kvm_para_available())
324                 return;
325
326         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
327                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
328
329 #ifdef CONFIG_PREEMPT
330                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
331 #endif
332                 pa |= KVM_ASYNC_PF_ENABLED;
333
334                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
335                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
336
337                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
338                 __this_cpu_write(apf_reason.enabled, 1);
339                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
340                        smp_processor_id());
341         }
342
343         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
344                 unsigned long pa;
345                 /* Size alignment is implied but just to make it explicit. */
346                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
347                 __this_cpu_write(kvm_apic_eoi, 0);
348                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
349                         | KVM_MSR_ENABLED;
350                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
351         }
352
353         if (has_steal_clock)
354                 kvm_register_steal_time();
355 }
356
357 static void kvm_pv_disable_apf(void)
358 {
359         if (!__this_cpu_read(apf_reason.enabled))
360                 return;
361
362         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
363         __this_cpu_write(apf_reason.enabled, 0);
364
365         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
366                smp_processor_id());
367 }
368
369 static void kvm_pv_guest_cpu_reboot(void *unused)
370 {
371         /*
372          * We disable PV EOI before we load a new kernel by kexec,
373          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
374          * New kernel can re-enable when it boots.
375          */
376         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
377                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
378         kvm_pv_disable_apf();
379         kvm_disable_steal_time();
380 }
381
382 static int kvm_pv_reboot_notify(struct notifier_block *nb,
383                                 unsigned long code, void *unused)
384 {
385         if (code == SYS_RESTART)
386                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
387         return NOTIFY_DONE;
388 }
389
390 static struct notifier_block kvm_pv_reboot_nb = {
391         .notifier_call = kvm_pv_reboot_notify,
392 };
393
394 static u64 kvm_steal_clock(int cpu)
395 {
396         u64 steal;
397         struct kvm_steal_time *src;
398         int version;
399
400         src = &per_cpu(steal_time, cpu);
401         do {
402                 version = src->version;
403                 virt_rmb();
404                 steal = src->steal;
405                 virt_rmb();
406         } while ((version & 1) || (version != src->version));
407
408         return steal;
409 }
410
411 void kvm_disable_steal_time(void)
412 {
413         if (!has_steal_clock)
414                 return;
415
416         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
417 }
418
419 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
420 {
421         early_set_memory_decrypted((unsigned long) ptr, size);
422 }
423
424 /*
425  * Iterate through all possible CPUs and map the memory region pointed
426  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
427  *
428  * Note: we iterate through all possible CPUs to ensure that CPUs
429  * hotplugged will have their per-cpu variable already mapped as
430  * decrypted.
431  */
432 static void __init sev_map_percpu_data(void)
433 {
434         int cpu;
435
436         if (!sev_active())
437                 return;
438
439         for_each_possible_cpu(cpu) {
440                 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
441                 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
442                 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
443         }
444 }
445
446 #ifdef CONFIG_SMP
447 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
448
449 static void __send_ipi_mask(const struct cpumask *mask, int vector)
450 {
451         unsigned long flags;
452         int cpu, apic_id, icr;
453         int min = 0, max = 0;
454 #ifdef CONFIG_X86_64
455         __uint128_t ipi_bitmap = 0;
456 #else
457         u64 ipi_bitmap = 0;
458 #endif
459
460         if (cpumask_empty(mask))
461                 return;
462
463         local_irq_save(flags);
464
465         switch (vector) {
466         default:
467                 icr = APIC_DM_FIXED | vector;
468                 break;
469         case NMI_VECTOR:
470                 icr = APIC_DM_NMI;
471                 break;
472         }
473
474         for_each_cpu(cpu, mask) {
475                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
476                 if (!ipi_bitmap) {
477                         min = max = apic_id;
478                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
479                         ipi_bitmap <<= min - apic_id;
480                         min = apic_id;
481                 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
482                         max = apic_id < max ? max : apic_id;
483                 } else {
484                         kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
485                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
486                         min = max = apic_id;
487                         ipi_bitmap = 0;
488                 }
489                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
490         }
491
492         if (ipi_bitmap) {
493                 kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
494                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
495         }
496
497         local_irq_restore(flags);
498 }
499
500 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
501 {
502         __send_ipi_mask(mask, vector);
503 }
504
505 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
506 {
507         unsigned int this_cpu = smp_processor_id();
508         struct cpumask new_mask;
509         const struct cpumask *local_mask;
510
511         cpumask_copy(&new_mask, mask);
512         cpumask_clear_cpu(this_cpu, &new_mask);
513         local_mask = &new_mask;
514         __send_ipi_mask(local_mask, vector);
515 }
516
517 static void kvm_send_ipi_allbutself(int vector)
518 {
519         kvm_send_ipi_mask_allbutself(cpu_online_mask, vector);
520 }
521
522 static void kvm_send_ipi_all(int vector)
523 {
524         __send_ipi_mask(cpu_online_mask, vector);
525 }
526
527 /*
528  * Set the IPI entry points
529  */
530 static void kvm_setup_pv_ipi(void)
531 {
532         apic->send_IPI_mask = kvm_send_ipi_mask;
533         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
534         apic->send_IPI_allbutself = kvm_send_ipi_allbutself;
535         apic->send_IPI_all = kvm_send_ipi_all;
536         pr_info("KVM setup pv IPIs\n");
537 }
538
539 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
540 {
541         native_smp_prepare_cpus(max_cpus);
542         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
543                 static_branch_disable(&virt_spin_lock_key);
544 }
545
546 static void __init kvm_smp_prepare_boot_cpu(void)
547 {
548         /*
549          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
550          * shares the guest physical address with the hypervisor.
551          */
552         sev_map_percpu_data();
553
554         kvm_guest_cpu_init();
555         native_smp_prepare_boot_cpu();
556         kvm_spinlock_init();
557 }
558
559 static void kvm_guest_cpu_offline(void)
560 {
561         kvm_disable_steal_time();
562         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
563                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
564         kvm_pv_disable_apf();
565         apf_task_wake_all();
566 }
567
568 static int kvm_cpu_online(unsigned int cpu)
569 {
570         local_irq_disable();
571         kvm_guest_cpu_init();
572         local_irq_enable();
573         return 0;
574 }
575
576 static int kvm_cpu_down_prepare(unsigned int cpu)
577 {
578         local_irq_disable();
579         kvm_guest_cpu_offline();
580         local_irq_enable();
581         return 0;
582 }
583 #endif
584
585 static void __init kvm_apf_trap_init(void)
586 {
587         update_intr_gate(X86_TRAP_PF, async_page_fault);
588 }
589
590 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
591
592 static void kvm_flush_tlb_others(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_tlb_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                 src = &per_cpu(steal_time, cpu);
607                 state = READ_ONCE(src->preempted);
608                 if ((state & KVM_VCPU_PREEMPTED)) {
609                         if (try_cmpxchg(&src->preempted, &state,
610                                         state | KVM_VCPU_FLUSH_TLB))
611                                 __cpumask_clear_cpu(cpu, flushmask);
612                 }
613         }
614
615         native_flush_tlb_others(flushmask, info);
616 }
617
618 static void __init kvm_guest_init(void)
619 {
620         int i;
621
622         if (!kvm_para_available())
623                 return;
624
625         paravirt_ops_setup();
626         register_reboot_notifier(&kvm_pv_reboot_nb);
627         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
628                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
629         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
630                 x86_init.irqs.trap_init = kvm_apf_trap_init;
631
632         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
633                 has_steal_clock = 1;
634                 pv_time_ops.steal_clock = kvm_steal_clock;
635         }
636
637         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
638             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
639             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME))
640                 pv_mmu_ops.flush_tlb_others = kvm_flush_tlb_others;
641
642         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
643                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
644
645 #ifdef CONFIG_SMP
646         smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
647         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
648         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
649                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
650                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
651 #else
652         sev_map_percpu_data();
653         kvm_guest_cpu_init();
654 #endif
655
656         /*
657          * Hard lockup detection is enabled by default. Disable it, as guests
658          * can get false positives too easily, for example if the host is
659          * overcommitted.
660          */
661         hardlockup_detector_disable();
662 }
663
664 static noinline uint32_t __kvm_cpuid_base(void)
665 {
666         if (boot_cpu_data.cpuid_level < 0)
667                 return 0;       /* So we don't blow up on old processors */
668
669         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
670                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
671
672         return 0;
673 }
674
675 static inline uint32_t kvm_cpuid_base(void)
676 {
677         static int kvm_cpuid_base = -1;
678
679         if (kvm_cpuid_base == -1)
680                 kvm_cpuid_base = __kvm_cpuid_base();
681
682         return kvm_cpuid_base;
683 }
684
685 bool kvm_para_available(void)
686 {
687         return kvm_cpuid_base() != 0;
688 }
689 EXPORT_SYMBOL_GPL(kvm_para_available);
690
691 unsigned int kvm_arch_para_features(void)
692 {
693         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
694 }
695
696 unsigned int kvm_arch_para_hints(void)
697 {
698         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
699 }
700
701 static uint32_t __init kvm_detect(void)
702 {
703         return kvm_cpuid_base();
704 }
705
706 static void __init kvm_apic_init(void)
707 {
708 #if defined(CONFIG_SMP)
709         if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
710                 kvm_setup_pv_ipi();
711 #endif
712 }
713
714 static void __init kvm_init_platform(void)
715 {
716         kvmclock_init();
717         x86_platform.apic_post_init = kvm_apic_init;
718 }
719
720 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
721         .name                   = "KVM",
722         .detect                 = kvm_detect,
723         .type                   = X86_HYPER_KVM,
724         .init.guest_late_init   = kvm_guest_init,
725         .init.x2apic_available  = kvm_para_available,
726         .init.init_platform     = kvm_init_platform,
727 };
728
729 static __init int activate_jump_labels(void)
730 {
731         if (has_steal_clock) {
732                 static_key_slow_inc(&paravirt_steal_enabled);
733                 if (steal_acc)
734                         static_key_slow_inc(&paravirt_steal_rq_enabled);
735         }
736
737         return 0;
738 }
739 arch_initcall(activate_jump_labels);
740
741 static __init int kvm_setup_pv_tlb_flush(void)
742 {
743         int cpu;
744
745         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
746             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
747             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
748                 for_each_possible_cpu(cpu) {
749                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
750                                 GFP_KERNEL, cpu_to_node(cpu));
751                 }
752                 pr_info("KVM setup pv remote TLB flush\n");
753         }
754
755         return 0;
756 }
757 arch_initcall(kvm_setup_pv_tlb_flush);
758
759 #ifdef CONFIG_PARAVIRT_SPINLOCKS
760
761 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
762 static void kvm_kick_cpu(int cpu)
763 {
764         int apicid;
765         unsigned long flags = 0;
766
767         apicid = per_cpu(x86_cpu_to_apicid, cpu);
768         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
769 }
770
771 #include <asm/qspinlock.h>
772
773 static void kvm_wait(u8 *ptr, u8 val)
774 {
775         unsigned long flags;
776
777         if (in_nmi())
778                 return;
779
780         local_irq_save(flags);
781
782         if (READ_ONCE(*ptr) != val)
783                 goto out;
784
785         /*
786          * halt until it's our turn and kicked. Note that we do safe halt
787          * for irq enabled case to avoid hang when lock info is overwritten
788          * in irq spinlock slowpath and no spurious interrupt occur to save us.
789          */
790         if (arch_irqs_disabled_flags(flags))
791                 halt();
792         else
793                 safe_halt();
794
795 out:
796         local_irq_restore(flags);
797 }
798
799 #ifdef CONFIG_X86_32
800 __visible bool __kvm_vcpu_is_preempted(long cpu)
801 {
802         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
803
804         return !!(src->preempted & KVM_VCPU_PREEMPTED);
805 }
806 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
807
808 #else
809
810 #include <asm/asm-offsets.h>
811
812 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
813
814 /*
815  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
816  * restoring to/from the stack.
817  */
818 asm(
819 ".pushsection .text;"
820 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
821 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
822 "__raw_callee_save___kvm_vcpu_is_preempted:"
823 "movq   __per_cpu_offset(,%rdi,8), %rax;"
824 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
825 "setne  %al;"
826 "ret;"
827 ".popsection");
828
829 #endif
830
831 /*
832  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
833  */
834 void __init kvm_spinlock_init(void)
835 {
836         if (!kvm_para_available())
837                 return;
838         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
839         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
840                 return;
841
842         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
843                 return;
844
845         /* Don't use the pvqspinlock code if there is only 1 vCPU. */
846         if (num_possible_cpus() == 1)
847                 return;
848
849         __pv_init_lock_hash();
850         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
851         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
852         pv_lock_ops.wait = kvm_wait;
853         pv_lock_ops.kick = kvm_kick_cpu;
854
855         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
856                 pv_lock_ops.vcpu_is_preempted =
857                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
858         }
859 }
860
861 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */