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