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/*
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 <asm/timer.h>
39#include <asm/cpu.h>
40#include <asm/traps.h>
41#include <asm/desc.h>
42#include <asm/tlbflush.h>
43#include <asm/idle.h>
44#include <asm/apic.h>
45#include <asm/apicdef.h>
46#include <asm/hypervisor.h>
47#include <asm/kvm_guest.h>
48
49static int kvmapf = 1;
50
51static int 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 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 int kvmclock_vsyscall = 1;
69static int parse_no_kvmclock_vsyscall(char *arg)
70{
71 kvmclock_vsyscall = 0;
72 return 0;
73}
74
75early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
76
77static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
78static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
79static int has_steal_clock = 0;
80
81/*
82 * No need for any "IO delay" on KVM
83 */
84static void kvm_io_delay(void)
85{
86}
87
88#define KVM_TASK_SLEEP_HASHBITS 8
89#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
90
91struct kvm_task_sleep_node {
92 struct hlist_node link;
93 wait_queue_head_t wq;
94 u32 token;
95 int cpu;
96 bool halted;
97};
98
99static struct kvm_task_sleep_head {
100 spinlock_t lock;
101 struct hlist_head list;
102} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
103
104static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
105 u32 token)
106{
107 struct hlist_node *p;
108
109 hlist_for_each(p, &b->list) {
110 struct kvm_task_sleep_node *n =
111 hlist_entry(p, typeof(*n), link);
112 if (n->token == token)
113 return n;
114 }
115
116 return NULL;
117}
118
119void kvm_async_pf_task_wait(u32 token)
120{
121 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
122 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
123 struct kvm_task_sleep_node n, *e;
124 DEFINE_WAIT(wait);
125
126 rcu_irq_enter();
127
128 spin_lock(&b->lock);
129 e = _find_apf_task(b, token);
130 if (e) {
131 /* dummy entry exist -> wake up was delivered ahead of PF */
132 hlist_del(&e->link);
133 kfree(e);
134 spin_unlock(&b->lock);
135
136 rcu_irq_exit();
137 return;
138 }
139
140 n.token = token;
141 n.cpu = smp_processor_id();
142 n.halted = is_idle_task(current) || preempt_count() > 1;
143 init_waitqueue_head(&n.wq);
144 hlist_add_head(&n.link, &b->list);
145 spin_unlock(&b->lock);
146
147 for (;;) {
148 if (!n.halted)
149 prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
150 if (hlist_unhashed(&n.link))
151 break;
152
153 if (!n.halted) {
154 local_irq_enable();
155 schedule();
156 local_irq_disable();
157 } else {
158 /*
159 * We cannot reschedule. So halt.
160 */
161 rcu_irq_exit();
162 native_safe_halt();
163 rcu_irq_enter();
164 local_irq_disable();
165 }
166 }
167 if (!n.halted)
168 finish_wait(&n.wq, &wait);
169
170 rcu_irq_exit();
171 return;
172}
173EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
174
175static void apf_task_wake_one(struct kvm_task_sleep_node *n)
176{
177 hlist_del_init(&n->link);
178 if (n->halted)
179 smp_send_reschedule(n->cpu);
180 else if (waitqueue_active(&n->wq))
181 wake_up(&n->wq);
182}
183
184static void apf_task_wake_all(void)
185{
186 int i;
187
188 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
189 struct hlist_node *p, *next;
190 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
191 spin_lock(&b->lock);
192 hlist_for_each_safe(p, next, &b->list) {
193 struct kvm_task_sleep_node *n =
194 hlist_entry(p, typeof(*n), link);
195 if (n->cpu == smp_processor_id())
196 apf_task_wake_one(n);
197 }
198 spin_unlock(&b->lock);
199 }
200}
201
202void kvm_async_pf_task_wake(u32 token)
203{
204 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
205 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
206 struct kvm_task_sleep_node *n;
207
208 if (token == ~0) {
209 apf_task_wake_all();
210 return;
211 }
212
213again:
214 spin_lock(&b->lock);
215 n = _find_apf_task(b, token);
216 if (!n) {
217 /*
218 * async PF was not yet handled.
219 * Add dummy entry for the token.
220 */
221 n = kzalloc(sizeof(*n), GFP_ATOMIC);
222 if (!n) {
223 /*
224 * Allocation failed! Busy wait while other cpu
225 * handles async PF.
226 */
227 spin_unlock(&b->lock);
228 cpu_relax();
229 goto again;
230 }
231 n->token = token;
232 n->cpu = smp_processor_id();
233 init_waitqueue_head(&n->wq);
234 hlist_add_head(&n->link, &b->list);
235 } else
236 apf_task_wake_one(n);
237 spin_unlock(&b->lock);
238 return;
239}
240EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
241
242u32 kvm_read_and_reset_pf_reason(void)
243{
244 u32 reason = 0;
245
246 if (__get_cpu_var(apf_reason).enabled) {
247 reason = __get_cpu_var(apf_reason).reason;
248 __get_cpu_var(apf_reason).reason = 0;
249 }
250
251 return reason;
252}
253EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
254
255dotraplinkage void __kprobes
256do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
257{
258 enum ctx_state prev_state;
259
260 switch (kvm_read_and_reset_pf_reason()) {
261 default:
262 do_page_fault(regs, error_code);
263 break;
264 case KVM_PV_REASON_PAGE_NOT_PRESENT:
265 /* page is swapped out by the host. */
266 prev_state = exception_enter();
267 exit_idle();
268 kvm_async_pf_task_wait((u32)read_cr2());
269 exception_exit(prev_state);
270 break;
271 case KVM_PV_REASON_PAGE_READY:
272 rcu_irq_enter();
273 exit_idle();
274 kvm_async_pf_task_wake((u32)read_cr2());
275 rcu_irq_exit();
276 break;
277 }
278}
279
280static void __init paravirt_ops_setup(void)
281{
282 pv_info.name = "KVM";
283 pv_info.paravirt_enabled = 1;
284
285 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
286 pv_cpu_ops.io_delay = kvm_io_delay;
287
288#ifdef CONFIG_X86_IO_APIC
289 no_timer_check = 1;
290#endif
291}
292
293static void kvm_register_steal_time(void)
294{
295 int cpu = smp_processor_id();
296 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
297
298 if (!has_steal_clock)
299 return;
300
301 memset(st, 0, sizeof(*st));
302
303 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
304 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
305 cpu, (unsigned long long) slow_virt_to_phys(st));
306}
307
308static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
309
310static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
311{
312 /**
313 * This relies on __test_and_clear_bit to modify the memory
314 * in a way that is atomic with respect to the local CPU.
315 * The hypervisor only accesses this memory from the local CPU so
316 * there's no need for lock or memory barriers.
317 * An optimization barrier is implied in apic write.
318 */
319 if (__test_and_clear_bit(KVM_PV_EOI_BIT, &__get_cpu_var(kvm_apic_eoi)))
320 return;
321 apic_write(APIC_EOI, APIC_EOI_ACK);
322}
323
324void kvm_guest_cpu_init(void)
325{
326 if (!kvm_para_available())
327 return;
328
329 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
330 u64 pa = slow_virt_to_phys(&__get_cpu_var(apf_reason));
331
332#ifdef CONFIG_PREEMPT
333 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
334#endif
335 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
336 __get_cpu_var(apf_reason).enabled = 1;
337 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
338 smp_processor_id());
339 }
340
341 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
342 unsigned long pa;
343 /* Size alignment is implied but just to make it explicit. */
344 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
345 __get_cpu_var(kvm_apic_eoi) = 0;
346 pa = slow_virt_to_phys(&__get_cpu_var(kvm_apic_eoi))
347 | KVM_MSR_ENABLED;
348 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
349 }
350
351 if (has_steal_clock)
352 kvm_register_steal_time();
353}
354
355static void kvm_pv_disable_apf(void)
356{
357 if (!__get_cpu_var(apf_reason).enabled)
358 return;
359
360 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
361 __get_cpu_var(apf_reason).enabled = 0;
362
363 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
364 smp_processor_id());
365}
366
367static void kvm_pv_guest_cpu_reboot(void *unused)
368{
369 /*
370 * We disable PV EOI before we load a new kernel by kexec,
371 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
372 * New kernel can re-enable when it boots.
373 */
374 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
375 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
376 kvm_pv_disable_apf();
377 kvm_disable_steal_time();
378}
379
380static int kvm_pv_reboot_notify(struct notifier_block *nb,
381 unsigned long code, void *unused)
382{
383 if (code == SYS_RESTART)
384 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
385 return NOTIFY_DONE;
386}
387
388static struct notifier_block kvm_pv_reboot_nb = {
389 .notifier_call = kvm_pv_reboot_notify,
390};
391
392static u64 kvm_steal_clock(int cpu)
393{
394 u64 steal;
395 struct kvm_steal_time *src;
396 int version;
397
398 src = &per_cpu(steal_time, cpu);
399 do {
400 version = src->version;
401 rmb();
402 steal = src->steal;
403 rmb();
404 } while ((version & 1) || (version != src->version));
405
406 return steal;
407}
408
409void kvm_disable_steal_time(void)
410{
411 if (!has_steal_clock)
412 return;
413
414 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
415}
416
417#ifdef CONFIG_SMP
418static void __init kvm_smp_prepare_boot_cpu(void)
419{
420 kvm_guest_cpu_init();
421 native_smp_prepare_boot_cpu();
422 kvm_spinlock_init();
423}
424
425static void kvm_guest_cpu_online(void *dummy)
426{
427 kvm_guest_cpu_init();
428}
429
430static void kvm_guest_cpu_offline(void *dummy)
431{
432 kvm_disable_steal_time();
433 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
434 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
435 kvm_pv_disable_apf();
436 apf_task_wake_all();
437}
438
439static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
440 void *hcpu)
441{
442 int cpu = (unsigned long)hcpu;
443 switch (action) {
444 case CPU_ONLINE:
445 case CPU_DOWN_FAILED:
446 case CPU_ONLINE_FROZEN:
447 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
448 break;
449 case CPU_DOWN_PREPARE:
450 case CPU_DOWN_PREPARE_FROZEN:
451 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
452 break;
453 default:
454 break;
455 }
456 return NOTIFY_OK;
457}
458
459static struct notifier_block kvm_cpu_notifier = {
460 .notifier_call = kvm_cpu_notify,
461};
462#endif
463
464static void __init kvm_apf_trap_init(void)
465{
466 set_intr_gate(14, async_page_fault);
467}
468
469void __init kvm_guest_init(void)
470{
471 int i;
472
473 if (!kvm_para_available())
474 return;
475
476 paravirt_ops_setup();
477 register_reboot_notifier(&kvm_pv_reboot_nb);
478 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
479 spin_lock_init(&async_pf_sleepers[i].lock);
480 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
481 x86_init.irqs.trap_init = kvm_apf_trap_init;
482
483 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
484 has_steal_clock = 1;
485 pv_time_ops.steal_clock = kvm_steal_clock;
486 }
487
488 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
489 apic_set_eoi_write(kvm_guest_apic_eoi_write);
490
491 if (kvmclock_vsyscall)
492 kvm_setup_vsyscall_timeinfo();
493
494#ifdef CONFIG_SMP
495 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
496 register_cpu_notifier(&kvm_cpu_notifier);
497#else
498 kvm_guest_cpu_init();
499#endif
500}
501
502static noinline uint32_t __kvm_cpuid_base(void)
503{
504 if (boot_cpu_data.cpuid_level < 0)
505 return 0; /* So we don't blow up on old processors */
506
507 if (cpu_has_hypervisor)
508 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
509
510 return 0;
511}
512
513static inline uint32_t kvm_cpuid_base(void)
514{
515 static int kvm_cpuid_base = -1;
516
517 if (kvm_cpuid_base == -1)
518 kvm_cpuid_base = __kvm_cpuid_base();
519
520 return kvm_cpuid_base;
521}
522
523bool kvm_para_available(void)
524{
525 return kvm_cpuid_base() != 0;
526}
527EXPORT_SYMBOL_GPL(kvm_para_available);
528
529unsigned int kvm_arch_para_features(void)
530{
531 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
532}
533
534static uint32_t __init kvm_detect(void)
535{
536 return kvm_cpuid_base();
537}
538
539const struct hypervisor_x86 x86_hyper_kvm __refconst = {
540 .name = "KVM",
541 .detect = kvm_detect,
542 .x2apic_available = kvm_para_available,
543};
544EXPORT_SYMBOL_GPL(x86_hyper_kvm);
545
546static __init int activate_jump_labels(void)
547{
548 if (has_steal_clock) {
549 static_key_slow_inc(¶virt_steal_enabled);
550 if (steal_acc)
551 static_key_slow_inc(¶virt_steal_rq_enabled);
552 }
553
554 return 0;
555}
556arch_initcall(activate_jump_labels);
557
558#ifdef CONFIG_PARAVIRT_SPINLOCKS
559
560/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
561static void kvm_kick_cpu(int cpu)
562{
563 int apicid;
564 unsigned long flags = 0;
565
566 apicid = per_cpu(x86_cpu_to_apicid, cpu);
567 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
568}
569
570enum kvm_contention_stat {
571 TAKEN_SLOW,
572 TAKEN_SLOW_PICKUP,
573 RELEASED_SLOW,
574 RELEASED_SLOW_KICKED,
575 NR_CONTENTION_STATS
576};
577
578#ifdef CONFIG_KVM_DEBUG_FS
579#define HISTO_BUCKETS 30
580
581static struct kvm_spinlock_stats
582{
583 u32 contention_stats[NR_CONTENTION_STATS];
584 u32 histo_spin_blocked[HISTO_BUCKETS+1];
585 u64 time_blocked;
586} spinlock_stats;
587
588static u8 zero_stats;
589
590static inline void check_zero(void)
591{
592 u8 ret;
593 u8 old;
594
595 old = ACCESS_ONCE(zero_stats);
596 if (unlikely(old)) {
597 ret = cmpxchg(&zero_stats, old, 0);
598 /* This ensures only one fellow resets the stat */
599 if (ret == old)
600 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
601 }
602}
603
604static inline void add_stats(enum kvm_contention_stat var, u32 val)
605{
606 check_zero();
607 spinlock_stats.contention_stats[var] += val;
608}
609
610
611static inline u64 spin_time_start(void)
612{
613 return sched_clock();
614}
615
616static void __spin_time_accum(u64 delta, u32 *array)
617{
618 unsigned index;
619
620 index = ilog2(delta);
621 check_zero();
622
623 if (index < HISTO_BUCKETS)
624 array[index]++;
625 else
626 array[HISTO_BUCKETS]++;
627}
628
629static inline void spin_time_accum_blocked(u64 start)
630{
631 u32 delta;
632
633 delta = sched_clock() - start;
634 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
635 spinlock_stats.time_blocked += delta;
636}
637
638static struct dentry *d_spin_debug;
639static struct dentry *d_kvm_debug;
640
641struct dentry *kvm_init_debugfs(void)
642{
643 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
644 if (!d_kvm_debug)
645 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
646
647 return d_kvm_debug;
648}
649
650static int __init kvm_spinlock_debugfs(void)
651{
652 struct dentry *d_kvm;
653
654 d_kvm = kvm_init_debugfs();
655 if (d_kvm == NULL)
656 return -ENOMEM;
657
658 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
659
660 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
661
662 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
663 &spinlock_stats.contention_stats[TAKEN_SLOW]);
664 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
665 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
666
667 debugfs_create_u32("released_slow", 0444, d_spin_debug,
668 &spinlock_stats.contention_stats[RELEASED_SLOW]);
669 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
670 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
671
672 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
673 &spinlock_stats.time_blocked);
674
675 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
676 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
677
678 return 0;
679}
680fs_initcall(kvm_spinlock_debugfs);
681#else /* !CONFIG_KVM_DEBUG_FS */
682static inline void add_stats(enum kvm_contention_stat var, u32 val)
683{
684}
685
686static inline u64 spin_time_start(void)
687{
688 return 0;
689}
690
691static inline void spin_time_accum_blocked(u64 start)
692{
693}
694#endif /* CONFIG_KVM_DEBUG_FS */
695
696struct kvm_lock_waiting {
697 struct arch_spinlock *lock;
698 __ticket_t want;
699};
700
701/* cpus 'waiting' on a spinlock to become available */
702static cpumask_t waiting_cpus;
703
704/* Track spinlock on which a cpu is waiting */
705static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
706
707__visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
708{
709 struct kvm_lock_waiting *w;
710 int cpu;
711 u64 start;
712 unsigned long flags;
713
714 if (in_nmi())
715 return;
716
717 w = &__get_cpu_var(klock_waiting);
718 cpu = smp_processor_id();
719 start = spin_time_start();
720
721 /*
722 * Make sure an interrupt handler can't upset things in a
723 * partially setup state.
724 */
725 local_irq_save(flags);
726
727 /*
728 * The ordering protocol on this is that the "lock" pointer
729 * may only be set non-NULL if the "want" ticket is correct.
730 * If we're updating "want", we must first clear "lock".
731 */
732 w->lock = NULL;
733 smp_wmb();
734 w->want = want;
735 smp_wmb();
736 w->lock = lock;
737
738 add_stats(TAKEN_SLOW, 1);
739
740 /*
741 * This uses set_bit, which is atomic but we should not rely on its
742 * reordering gurantees. So barrier is needed after this call.
743 */
744 cpumask_set_cpu(cpu, &waiting_cpus);
745
746 barrier();
747
748 /*
749 * Mark entry to slowpath before doing the pickup test to make
750 * sure we don't deadlock with an unlocker.
751 */
752 __ticket_enter_slowpath(lock);
753
754 /*
755 * check again make sure it didn't become free while
756 * we weren't looking.
757 */
758 if (ACCESS_ONCE(lock->tickets.head) == want) {
759 add_stats(TAKEN_SLOW_PICKUP, 1);
760 goto out;
761 }
762
763 /*
764 * halt until it's our turn and kicked. Note that we do safe halt
765 * for irq enabled case to avoid hang when lock info is overwritten
766 * in irq spinlock slowpath and no spurious interrupt occur to save us.
767 */
768 if (arch_irqs_disabled_flags(flags))
769 halt();
770 else
771 safe_halt();
772
773out:
774 cpumask_clear_cpu(cpu, &waiting_cpus);
775 w->lock = NULL;
776 local_irq_restore(flags);
777 spin_time_accum_blocked(start);
778}
779PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
780
781/* Kick vcpu waiting on @lock->head to reach value @ticket */
782static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
783{
784 int cpu;
785
786 add_stats(RELEASED_SLOW, 1);
787 for_each_cpu(cpu, &waiting_cpus) {
788 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
789 if (ACCESS_ONCE(w->lock) == lock &&
790 ACCESS_ONCE(w->want) == ticket) {
791 add_stats(RELEASED_SLOW_KICKED, 1);
792 kvm_kick_cpu(cpu);
793 break;
794 }
795 }
796}
797
798/*
799 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
800 */
801void __init kvm_spinlock_init(void)
802{
803 if (!kvm_para_available())
804 return;
805 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
806 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
807 return;
808
809 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
810 pv_lock_ops.unlock_kick = kvm_unlock_kick;
811}
812
813static __init int kvm_spinlock_init_jump(void)
814{
815 if (!kvm_para_available())
816 return 0;
817 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
818 return 0;
819
820 static_key_slow_inc(¶virt_ticketlocks_enabled);
821 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
822
823 return 0;
824}
825early_initcall(kvm_spinlock_init_jump);
826
827#endif /* CONFIG_PARAVIRT_SPINLOCKS */