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 <linux/cc_platform.h>
31#include <linux/efi.h>
32#include <asm/timer.h>
33#include <asm/cpu.h>
34#include <asm/traps.h>
35#include <asm/desc.h>
36#include <asm/tlbflush.h>
37#include <asm/apic.h>
38#include <asm/apicdef.h>
39#include <asm/hypervisor.h>
40#include <asm/tlb.h>
41#include <asm/cpuidle_haltpoll.h>
42#include <asm/ptrace.h>
43#include <asm/reboot.h>
44#include <asm/svm.h>
45#include <asm/e820/api.h>
46
47DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
48
49static int kvmapf = 1;
50
51static int __init parse_no_kvmapf(char *arg)
52{
53 kvmapf = 0;
54 return 0;
55}
56
57early_param("no-kvmapf", parse_no_kvmapf);
58
59static int steal_acc = 1;
60static int __init parse_no_stealacc(char *arg)
61{
62 steal_acc = 0;
63 return 0;
64}
65
66early_param("no-steal-acc", parse_no_stealacc);
67
68static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled);
69static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
70DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
71static int has_steal_clock = 0;
72
73static int has_guest_poll = 0;
74/*
75 * No need for any "IO delay" on KVM
76 */
77static void kvm_io_delay(void)
78{
79}
80
81#define KVM_TASK_SLEEP_HASHBITS 8
82#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
83
84struct kvm_task_sleep_node {
85 struct hlist_node link;
86 struct swait_queue_head wq;
87 u32 token;
88 int cpu;
89};
90
91static struct kvm_task_sleep_head {
92 raw_spinlock_t lock;
93 struct hlist_head list;
94} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
95
96static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
97 u32 token)
98{
99 struct hlist_node *p;
100
101 hlist_for_each(p, &b->list) {
102 struct kvm_task_sleep_node *n =
103 hlist_entry(p, typeof(*n), link);
104 if (n->token == token)
105 return n;
106 }
107
108 return NULL;
109}
110
111static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
112{
113 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
114 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
115 struct kvm_task_sleep_node *e;
116
117 raw_spin_lock(&b->lock);
118 e = _find_apf_task(b, token);
119 if (e) {
120 /* dummy entry exist -> wake up was delivered ahead of PF */
121 hlist_del(&e->link);
122 raw_spin_unlock(&b->lock);
123 kfree(e);
124 return false;
125 }
126
127 n->token = token;
128 n->cpu = smp_processor_id();
129 init_swait_queue_head(&n->wq);
130 hlist_add_head(&n->link, &b->list);
131 raw_spin_unlock(&b->lock);
132 return true;
133}
134
135/*
136 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
137 * @token: Token to identify the sleep node entry
138 *
139 * Invoked from the async pagefault handling code or from the VM exit page
140 * fault handler. In both cases RCU is watching.
141 */
142void kvm_async_pf_task_wait_schedule(u32 token)
143{
144 struct kvm_task_sleep_node n;
145 DECLARE_SWAITQUEUE(wait);
146
147 lockdep_assert_irqs_disabled();
148
149 if (!kvm_async_pf_queue_task(token, &n))
150 return;
151
152 for (;;) {
153 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
154 if (hlist_unhashed(&n.link))
155 break;
156
157 local_irq_enable();
158 schedule();
159 local_irq_disable();
160 }
161 finish_swait(&n.wq, &wait);
162}
163EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
164
165static void apf_task_wake_one(struct kvm_task_sleep_node *n)
166{
167 hlist_del_init(&n->link);
168 if (swq_has_sleeper(&n->wq))
169 swake_up_one(&n->wq);
170}
171
172static void apf_task_wake_all(void)
173{
174 int i;
175
176 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
177 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
178 struct kvm_task_sleep_node *n;
179 struct hlist_node *p, *next;
180
181 raw_spin_lock(&b->lock);
182 hlist_for_each_safe(p, next, &b->list) {
183 n = hlist_entry(p, typeof(*n), link);
184 if (n->cpu == smp_processor_id())
185 apf_task_wake_one(n);
186 }
187 raw_spin_unlock(&b->lock);
188 }
189}
190
191void kvm_async_pf_task_wake(u32 token)
192{
193 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
194 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
195 struct kvm_task_sleep_node *n, *dummy = NULL;
196
197 if (token == ~0) {
198 apf_task_wake_all();
199 return;
200 }
201
202again:
203 raw_spin_lock(&b->lock);
204 n = _find_apf_task(b, token);
205 if (!n) {
206 /*
207 * Async #PF not yet handled, add a dummy entry for the token.
208 * Allocating the token must be down outside of the raw lock
209 * as the allocator is preemptible on PREEMPT_RT kernels.
210 */
211 if (!dummy) {
212 raw_spin_unlock(&b->lock);
213 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
214
215 /*
216 * Continue looping on allocation failure, eventually
217 * the async #PF will be handled and allocating a new
218 * node will be unnecessary.
219 */
220 if (!dummy)
221 cpu_relax();
222
223 /*
224 * Recheck for async #PF completion before enqueueing
225 * the dummy token to avoid duplicate list entries.
226 */
227 goto again;
228 }
229 dummy->token = token;
230 dummy->cpu = smp_processor_id();
231 init_swait_queue_head(&dummy->wq);
232 hlist_add_head(&dummy->link, &b->list);
233 dummy = NULL;
234 } else {
235 apf_task_wake_one(n);
236 }
237 raw_spin_unlock(&b->lock);
238
239 /* A dummy token might be allocated and ultimately not used. */
240 kfree(dummy);
241}
242EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243
244noinstr u32 kvm_read_and_reset_apf_flags(void)
245{
246 u32 flags = 0;
247
248 if (__this_cpu_read(async_pf_enabled)) {
249 flags = __this_cpu_read(apf_reason.flags);
250 __this_cpu_write(apf_reason.flags, 0);
251 }
252
253 return flags;
254}
255EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
256
257noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
258{
259 u32 flags = kvm_read_and_reset_apf_flags();
260 irqentry_state_t state;
261
262 if (!flags)
263 return false;
264
265 state = irqentry_enter(regs);
266 instrumentation_begin();
267
268 /*
269 * If the host managed to inject an async #PF into an interrupt
270 * disabled region, then die hard as this is not going to end well
271 * and the host side is seriously broken.
272 */
273 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
274 panic("Host injected async #PF in interrupt disabled region\n");
275
276 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
277 if (unlikely(!(user_mode(regs))))
278 panic("Host injected async #PF in kernel mode\n");
279 /* Page is swapped out by the host. */
280 kvm_async_pf_task_wait_schedule(token);
281 } else {
282 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
283 }
284
285 instrumentation_end();
286 irqentry_exit(regs, state);
287 return true;
288}
289
290DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
291{
292 struct pt_regs *old_regs = set_irq_regs(regs);
293 u32 token;
294
295 apic_eoi();
296
297 inc_irq_stat(irq_hv_callback_count);
298
299 if (__this_cpu_read(async_pf_enabled)) {
300 token = __this_cpu_read(apf_reason.token);
301 kvm_async_pf_task_wake(token);
302 __this_cpu_write(apf_reason.token, 0);
303 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
304 }
305
306 set_irq_regs(old_regs);
307}
308
309static void __init paravirt_ops_setup(void)
310{
311 pv_info.name = "KVM";
312
313 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
314 pv_ops.cpu.io_delay = kvm_io_delay;
315
316#ifdef CONFIG_X86_IO_APIC
317 no_timer_check = 1;
318#endif
319}
320
321static void kvm_register_steal_time(void)
322{
323 int cpu = smp_processor_id();
324 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
325
326 if (!has_steal_clock)
327 return;
328
329 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
330 pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
331 (unsigned long long) slow_virt_to_phys(st));
332}
333
334static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
335
336static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
337{
338 /**
339 * This relies on __test_and_clear_bit to modify the memory
340 * in a way that is atomic with respect to the local CPU.
341 * The hypervisor only accesses this memory from the local CPU so
342 * there's no need for lock or memory barriers.
343 * An optimization barrier is implied in apic write.
344 */
345 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
346 return;
347 apic_native_eoi();
348}
349
350static void kvm_guest_cpu_init(void)
351{
352 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
353 u64 pa;
354
355 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
356
357 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
358 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
359
360 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
361 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
362
363 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
364
365 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
366 __this_cpu_write(async_pf_enabled, true);
367 pr_debug("setup async PF for cpu %d\n", smp_processor_id());
368 }
369
370 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
371 unsigned long pa;
372
373 /* Size alignment is implied but just to make it explicit. */
374 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
375 __this_cpu_write(kvm_apic_eoi, 0);
376 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
377 | KVM_MSR_ENABLED;
378 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
379 }
380
381 if (has_steal_clock)
382 kvm_register_steal_time();
383}
384
385static void kvm_pv_disable_apf(void)
386{
387 if (!__this_cpu_read(async_pf_enabled))
388 return;
389
390 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
391 __this_cpu_write(async_pf_enabled, false);
392
393 pr_debug("disable async PF for cpu %d\n", smp_processor_id());
394}
395
396static void kvm_disable_steal_time(void)
397{
398 if (!has_steal_clock)
399 return;
400
401 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
402}
403
404static u64 kvm_steal_clock(int cpu)
405{
406 u64 steal;
407 struct kvm_steal_time *src;
408 int version;
409
410 src = &per_cpu(steal_time, cpu);
411 do {
412 version = src->version;
413 virt_rmb();
414 steal = src->steal;
415 virt_rmb();
416 } while ((version & 1) || (version != src->version));
417
418 return steal;
419}
420
421static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
422{
423 early_set_memory_decrypted((unsigned long) ptr, size);
424}
425
426/*
427 * Iterate through all possible CPUs and map the memory region pointed
428 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
429 *
430 * Note: we iterate through all possible CPUs to ensure that CPUs
431 * hotplugged will have their per-cpu variable already mapped as
432 * decrypted.
433 */
434static void __init sev_map_percpu_data(void)
435{
436 int cpu;
437
438 if (cc_vendor != CC_VENDOR_AMD ||
439 !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
440 return;
441
442 for_each_possible_cpu(cpu) {
443 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
444 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
445 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
446 }
447}
448
449static void kvm_guest_cpu_offline(bool shutdown)
450{
451 kvm_disable_steal_time();
452 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
453 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
454 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
455 wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
456 kvm_pv_disable_apf();
457 if (!shutdown)
458 apf_task_wake_all();
459 kvmclock_disable();
460}
461
462static int kvm_cpu_online(unsigned int cpu)
463{
464 unsigned long flags;
465
466 local_irq_save(flags);
467 kvm_guest_cpu_init();
468 local_irq_restore(flags);
469 return 0;
470}
471
472#ifdef CONFIG_SMP
473
474static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
475
476static bool pv_tlb_flush_supported(void)
477{
478 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
479 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
480 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
481 !boot_cpu_has(X86_FEATURE_MWAIT) &&
482 (num_possible_cpus() != 1));
483}
484
485static bool pv_ipi_supported(void)
486{
487 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
488 (num_possible_cpus() != 1));
489}
490
491static bool pv_sched_yield_supported(void)
492{
493 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
494 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
495 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
496 !boot_cpu_has(X86_FEATURE_MWAIT) &&
497 (num_possible_cpus() != 1));
498}
499
500#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
501
502static void __send_ipi_mask(const struct cpumask *mask, int vector)
503{
504 unsigned long flags;
505 int cpu, min = 0, max = 0;
506#ifdef CONFIG_X86_64
507 __uint128_t ipi_bitmap = 0;
508#else
509 u64 ipi_bitmap = 0;
510#endif
511 u32 apic_id, icr;
512 long ret;
513
514 if (cpumask_empty(mask))
515 return;
516
517 local_irq_save(flags);
518
519 switch (vector) {
520 default:
521 icr = APIC_DM_FIXED | vector;
522 break;
523 case NMI_VECTOR:
524 icr = APIC_DM_NMI;
525 break;
526 }
527
528 for_each_cpu(cpu, mask) {
529 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
530 if (!ipi_bitmap) {
531 min = max = apic_id;
532 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
533 ipi_bitmap <<= min - apic_id;
534 min = apic_id;
535 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
536 max = apic_id < max ? max : apic_id;
537 } else {
538 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
539 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
540 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
541 ret);
542 min = max = apic_id;
543 ipi_bitmap = 0;
544 }
545 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
546 }
547
548 if (ipi_bitmap) {
549 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
550 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
551 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
552 ret);
553 }
554
555 local_irq_restore(flags);
556}
557
558static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
559{
560 __send_ipi_mask(mask, vector);
561}
562
563static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
564{
565 unsigned int this_cpu = smp_processor_id();
566 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
567 const struct cpumask *local_mask;
568
569 cpumask_copy(new_mask, mask);
570 cpumask_clear_cpu(this_cpu, new_mask);
571 local_mask = new_mask;
572 __send_ipi_mask(local_mask, vector);
573}
574
575static int __init setup_efi_kvm_sev_migration(void)
576{
577 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
578 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
579 efi_status_t status;
580 unsigned long size;
581 bool enabled;
582
583 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
584 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
585 return 0;
586
587 if (!efi_enabled(EFI_BOOT))
588 return 0;
589
590 if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
591 pr_info("%s : EFI runtime services are not enabled\n", __func__);
592 return 0;
593 }
594
595 size = sizeof(enabled);
596
597 /* Get variable contents into buffer */
598 status = efi.get_variable(efi_sev_live_migration_enabled,
599 &efi_variable_guid, NULL, &size, &enabled);
600
601 if (status == EFI_NOT_FOUND) {
602 pr_info("%s : EFI live migration variable not found\n", __func__);
603 return 0;
604 }
605
606 if (status != EFI_SUCCESS) {
607 pr_info("%s : EFI variable retrieval failed\n", __func__);
608 return 0;
609 }
610
611 if (enabled == 0) {
612 pr_info("%s: live migration disabled in EFI\n", __func__);
613 return 0;
614 }
615
616 pr_info("%s : live migration enabled in EFI\n", __func__);
617 wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
618
619 return 1;
620}
621
622late_initcall(setup_efi_kvm_sev_migration);
623
624/*
625 * Set the IPI entry points
626 */
627static __init void kvm_setup_pv_ipi(void)
628{
629 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
630 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
631 pr_info("setup PV IPIs\n");
632}
633
634static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
635{
636 int cpu;
637
638 native_send_call_func_ipi(mask);
639
640 /* Make sure other vCPUs get a chance to run if they need to. */
641 for_each_cpu(cpu, mask) {
642 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
643 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
644 break;
645 }
646 }
647}
648
649static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
650 const struct flush_tlb_info *info)
651{
652 u8 state;
653 int cpu;
654 struct kvm_steal_time *src;
655 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
656
657 cpumask_copy(flushmask, cpumask);
658 /*
659 * We have to call flush only on online vCPUs. And
660 * queue flush_on_enter for pre-empted vCPUs
661 */
662 for_each_cpu(cpu, flushmask) {
663 /*
664 * The local vCPU is never preempted, so we do not explicitly
665 * skip check for local vCPU - it will never be cleared from
666 * flushmask.
667 */
668 src = &per_cpu(steal_time, cpu);
669 state = READ_ONCE(src->preempted);
670 if ((state & KVM_VCPU_PREEMPTED)) {
671 if (try_cmpxchg(&src->preempted, &state,
672 state | KVM_VCPU_FLUSH_TLB))
673 __cpumask_clear_cpu(cpu, flushmask);
674 }
675 }
676
677 native_flush_tlb_multi(flushmask, info);
678}
679
680static __init int kvm_alloc_cpumask(void)
681{
682 int cpu;
683
684 if (!kvm_para_available() || nopv)
685 return 0;
686
687 if (pv_tlb_flush_supported() || pv_ipi_supported())
688 for_each_possible_cpu(cpu) {
689 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
690 GFP_KERNEL, cpu_to_node(cpu));
691 }
692
693 return 0;
694}
695arch_initcall(kvm_alloc_cpumask);
696
697static void __init kvm_smp_prepare_boot_cpu(void)
698{
699 /*
700 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
701 * shares the guest physical address with the hypervisor.
702 */
703 sev_map_percpu_data();
704
705 kvm_guest_cpu_init();
706 native_smp_prepare_boot_cpu();
707 kvm_spinlock_init();
708}
709
710static int kvm_cpu_down_prepare(unsigned int cpu)
711{
712 unsigned long flags;
713
714 local_irq_save(flags);
715 kvm_guest_cpu_offline(false);
716 local_irq_restore(flags);
717 return 0;
718}
719
720#endif
721
722static int kvm_suspend(void)
723{
724 u64 val = 0;
725
726 kvm_guest_cpu_offline(false);
727
728#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
729 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
730 rdmsrl(MSR_KVM_POLL_CONTROL, val);
731 has_guest_poll = !(val & 1);
732#endif
733 return 0;
734}
735
736static void kvm_resume(void)
737{
738 kvm_cpu_online(raw_smp_processor_id());
739
740#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
741 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
742 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
743#endif
744}
745
746static struct syscore_ops kvm_syscore_ops = {
747 .suspend = kvm_suspend,
748 .resume = kvm_resume,
749};
750
751static void kvm_pv_guest_cpu_reboot(void *unused)
752{
753 kvm_guest_cpu_offline(true);
754}
755
756static int kvm_pv_reboot_notify(struct notifier_block *nb,
757 unsigned long code, void *unused)
758{
759 if (code == SYS_RESTART)
760 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
761 return NOTIFY_DONE;
762}
763
764static struct notifier_block kvm_pv_reboot_nb = {
765 .notifier_call = kvm_pv_reboot_notify,
766};
767
768/*
769 * After a PV feature is registered, the host will keep writing to the
770 * registered memory location. If the guest happens to shutdown, this memory
771 * won't be valid. In cases like kexec, in which you install a new kernel, this
772 * means a random memory location will be kept being written.
773 */
774#ifdef CONFIG_CRASH_DUMP
775static void kvm_crash_shutdown(struct pt_regs *regs)
776{
777 kvm_guest_cpu_offline(true);
778 native_machine_crash_shutdown(regs);
779}
780#endif
781
782#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
783bool __kvm_vcpu_is_preempted(long cpu);
784
785__visible bool __kvm_vcpu_is_preempted(long cpu)
786{
787 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
788
789 return !!(src->preempted & KVM_VCPU_PREEMPTED);
790}
791PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
792
793#else
794
795#include <asm/asm-offsets.h>
796
797extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
798
799/*
800 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
801 * restoring to/from the stack.
802 */
803#define PV_VCPU_PREEMPTED_ASM \
804 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
805 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
806 "setne %al\n\t"
807
808DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
809 PV_VCPU_PREEMPTED_ASM, .text);
810#endif
811
812static void __init kvm_guest_init(void)
813{
814 int i;
815
816 paravirt_ops_setup();
817 register_reboot_notifier(&kvm_pv_reboot_nb);
818 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
819 raw_spin_lock_init(&async_pf_sleepers[i].lock);
820
821 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
822 has_steal_clock = 1;
823 static_call_update(pv_steal_clock, kvm_steal_clock);
824
825 pv_ops.lock.vcpu_is_preempted =
826 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
827 }
828
829 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
830 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
831
832 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
833 static_branch_enable(&kvm_async_pf_enabled);
834 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt);
835 }
836
837#ifdef CONFIG_SMP
838 if (pv_tlb_flush_supported()) {
839 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
840 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
841 pr_info("KVM setup pv remote TLB flush\n");
842 }
843
844 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
845 if (pv_sched_yield_supported()) {
846 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
847 pr_info("setup PV sched yield\n");
848 }
849 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
850 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
851 pr_err("failed to install cpu hotplug callbacks\n");
852#else
853 sev_map_percpu_data();
854 kvm_guest_cpu_init();
855#endif
856
857#ifdef CONFIG_CRASH_DUMP
858 machine_ops.crash_shutdown = kvm_crash_shutdown;
859#endif
860
861 register_syscore_ops(&kvm_syscore_ops);
862
863 /*
864 * Hard lockup detection is enabled by default. Disable it, as guests
865 * can get false positives too easily, for example if the host is
866 * overcommitted.
867 */
868 hardlockup_detector_disable();
869}
870
871static noinline uint32_t __kvm_cpuid_base(void)
872{
873 if (boot_cpu_data.cpuid_level < 0)
874 return 0; /* So we don't blow up on old processors */
875
876 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
877 return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
878
879 return 0;
880}
881
882static inline uint32_t kvm_cpuid_base(void)
883{
884 static int kvm_cpuid_base = -1;
885
886 if (kvm_cpuid_base == -1)
887 kvm_cpuid_base = __kvm_cpuid_base();
888
889 return kvm_cpuid_base;
890}
891
892bool kvm_para_available(void)
893{
894 return kvm_cpuid_base() != 0;
895}
896EXPORT_SYMBOL_GPL(kvm_para_available);
897
898unsigned int kvm_arch_para_features(void)
899{
900 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
901}
902
903unsigned int kvm_arch_para_hints(void)
904{
905 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
906}
907EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
908
909static uint32_t __init kvm_detect(void)
910{
911 return kvm_cpuid_base();
912}
913
914static void __init kvm_apic_init(void)
915{
916#ifdef CONFIG_SMP
917 if (pv_ipi_supported())
918 kvm_setup_pv_ipi();
919#endif
920}
921
922static bool __init kvm_msi_ext_dest_id(void)
923{
924 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
925}
926
927static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
928{
929 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
930 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
931}
932
933static void __init kvm_init_platform(void)
934{
935 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
936 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
937 unsigned long nr_pages;
938 int i;
939
940 pv_ops.mmu.notify_page_enc_status_changed =
941 kvm_sev_hc_page_enc_status;
942
943 /*
944 * Reset the host's shared pages list related to kernel
945 * specific page encryption status settings before we load a
946 * new kernel by kexec. Reset the page encryption status
947 * during early boot instead of just before kexec to avoid SMP
948 * races during kvm_pv_guest_cpu_reboot().
949 * NOTE: We cannot reset the complete shared pages list
950 * here as we need to retain the UEFI/OVMF firmware
951 * specific settings.
952 */
953
954 for (i = 0; i < e820_table->nr_entries; i++) {
955 struct e820_entry *entry = &e820_table->entries[i];
956
957 if (entry->type != E820_TYPE_RAM)
958 continue;
959
960 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
961
962 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
963 nr_pages,
964 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
965 }
966
967 /*
968 * Ensure that _bss_decrypted section is marked as decrypted in the
969 * shared pages list.
970 */
971 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
972 __end_bss_decrypted - __start_bss_decrypted, 0);
973
974 /*
975 * If not booted using EFI, enable Live migration support.
976 */
977 if (!efi_enabled(EFI_BOOT))
978 wrmsrl(MSR_KVM_MIGRATION_CONTROL,
979 KVM_MIGRATION_READY);
980 }
981 kvmclock_init();
982 x86_platform.apic_post_init = kvm_apic_init;
983}
984
985#if defined(CONFIG_AMD_MEM_ENCRYPT)
986static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
987{
988 /* RAX and CPL are already in the GHCB */
989 ghcb_set_rbx(ghcb, regs->bx);
990 ghcb_set_rcx(ghcb, regs->cx);
991 ghcb_set_rdx(ghcb, regs->dx);
992 ghcb_set_rsi(ghcb, regs->si);
993}
994
995static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
996{
997 /* No checking of the return state needed */
998 return true;
999}
1000#endif
1001
1002const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1003 .name = "KVM",
1004 .detect = kvm_detect,
1005 .type = X86_HYPER_KVM,
1006 .init.guest_late_init = kvm_guest_init,
1007 .init.x2apic_available = kvm_para_available,
1008 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1009 .init.init_platform = kvm_init_platform,
1010#if defined(CONFIG_AMD_MEM_ENCRYPT)
1011 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1012 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1013#endif
1014};
1015
1016static __init int activate_jump_labels(void)
1017{
1018 if (has_steal_clock) {
1019 static_key_slow_inc(¶virt_steal_enabled);
1020 if (steal_acc)
1021 static_key_slow_inc(¶virt_steal_rq_enabled);
1022 }
1023
1024 return 0;
1025}
1026arch_initcall(activate_jump_labels);
1027
1028#ifdef CONFIG_PARAVIRT_SPINLOCKS
1029
1030/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1031static void kvm_kick_cpu(int cpu)
1032{
1033 unsigned long flags = 0;
1034 u32 apicid;
1035
1036 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1037 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1038}
1039
1040#include <asm/qspinlock.h>
1041
1042static void kvm_wait(u8 *ptr, u8 val)
1043{
1044 if (in_nmi())
1045 return;
1046
1047 /*
1048 * halt until it's our turn and kicked. Note that we do safe halt
1049 * for irq enabled case to avoid hang when lock info is overwritten
1050 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1051 */
1052 if (irqs_disabled()) {
1053 if (READ_ONCE(*ptr) == val)
1054 halt();
1055 } else {
1056 local_irq_disable();
1057
1058 /* safe_halt() will enable IRQ */
1059 if (READ_ONCE(*ptr) == val)
1060 safe_halt();
1061 else
1062 local_irq_enable();
1063 }
1064}
1065
1066/*
1067 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1068 */
1069void __init kvm_spinlock_init(void)
1070{
1071 /*
1072 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1073 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1074 * preferred over native qspinlock when vCPU is preempted.
1075 */
1076 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1077 pr_info("PV spinlocks disabled, no host support\n");
1078 return;
1079 }
1080
1081 /*
1082 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1083 * are available.
1084 */
1085 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1086 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1087 goto out;
1088 }
1089
1090 if (num_possible_cpus() == 1) {
1091 pr_info("PV spinlocks disabled, single CPU\n");
1092 goto out;
1093 }
1094
1095 if (nopvspin) {
1096 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1097 goto out;
1098 }
1099
1100 pr_info("PV spinlocks enabled\n");
1101
1102 __pv_init_lock_hash();
1103 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1104 pv_ops.lock.queued_spin_unlock =
1105 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1106 pv_ops.lock.wait = kvm_wait;
1107 pv_ops.lock.kick = kvm_kick_cpu;
1108
1109 /*
1110 * When PV spinlock is enabled which is preferred over
1111 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1112 * Just disable it anyway.
1113 */
1114out:
1115 static_branch_disable(&virt_spin_lock_key);
1116}
1117
1118#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1119
1120#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1121
1122static void kvm_disable_host_haltpoll(void *i)
1123{
1124 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1125}
1126
1127static void kvm_enable_host_haltpoll(void *i)
1128{
1129 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1130}
1131
1132void arch_haltpoll_enable(unsigned int cpu)
1133{
1134 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1135 pr_err_once("host does not support poll control\n");
1136 pr_err_once("host upgrade recommended\n");
1137 return;
1138 }
1139
1140 /* Enable guest halt poll disables host halt poll */
1141 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1142}
1143EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1144
1145void arch_haltpoll_disable(unsigned int cpu)
1146{
1147 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1148 return;
1149
1150 /* Disable guest halt poll enables host halt poll */
1151 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1152}
1153EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1154#endif