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