2 * KVM paravirt_ops implementation
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.
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.
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.
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>
23 #include <linux/context_tracking.h>
24 #include <linux/init.h>
25 #include <linux/kernel.h>
26 #include <linux/kvm_para.h>
27 #include <linux/cpu.h>
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>
42 #include <asm/traps.h>
44 #include <asm/tlbflush.h>
47 #include <asm/apicdef.h>
48 #include <asm/hypervisor.h>
49 #include <asm/kvm_guest.h>
51 static int kvmapf = 1;
53 static int parse_no_kvmapf(char *arg)
59 early_param("no-kvmapf", parse_no_kvmapf);
61 static int steal_acc = 1;
62 static int parse_no_stealacc(char *arg)
68 early_param("no-steal-acc", parse_no_stealacc);
70 static int kvmclock_vsyscall = 1;
71 static int parse_no_kvmclock_vsyscall(char *arg)
73 kvmclock_vsyscall = 0;
77 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
79 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
80 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
81 static int has_steal_clock = 0;
84 * No need for any "IO delay" on KVM
86 static void kvm_io_delay(void)
90 #define KVM_TASK_SLEEP_HASHBITS 8
91 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
93 struct kvm_task_sleep_node {
94 struct hlist_node link;
95 struct swait_queue_head wq;
101 static struct kvm_task_sleep_head {
103 struct hlist_head list;
104 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
106 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
109 struct hlist_node *p;
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)
121 void kvm_async_pf_task_wait(u32 token)
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);
130 raw_spin_lock(&b->lock);
131 e = _find_apf_task(b, token);
133 /* dummy entry exist -> wake up was delivered ahead of PF */
136 raw_spin_unlock(&b->lock);
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);
151 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
152 if (hlist_unhashed(&n.link))
161 * We cannot reschedule. So halt.
170 finish_swait(&n.wq, &wait);
175 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
177 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
179 hlist_del_init(&n->link);
181 smp_send_reschedule(n->cpu);
182 else if (swait_active(&n->wq))
186 static void apf_task_wake_all(void)
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);
200 raw_spin_unlock(&b->lock);
204 void kvm_async_pf_task_wake(u32 token)
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;
216 raw_spin_lock(&b->lock);
217 n = _find_apf_task(b, token);
220 * async PF was not yet handled.
221 * Add dummy entry for the token.
223 n = kzalloc(sizeof(*n), GFP_ATOMIC);
226 * Allocation failed! Busy wait while other cpu
229 raw_spin_unlock(&b->lock);
234 n->cpu = smp_processor_id();
235 init_swait_queue_head(&n->wq);
236 hlist_add_head(&n->link, &b->list);
238 apf_task_wake_one(n);
239 raw_spin_unlock(&b->lock);
242 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
244 u32 kvm_read_and_reset_pf_reason(void)
248 if (__this_cpu_read(apf_reason.enabled)) {
249 reason = __this_cpu_read(apf_reason.reason);
250 __this_cpu_write(apf_reason.reason, 0);
255 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
256 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
259 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
261 enum ctx_state prev_state;
263 switch (kvm_read_and_reset_pf_reason()) {
265 trace_do_page_fault(regs, error_code);
267 case KVM_PV_REASON_PAGE_NOT_PRESENT:
268 /* page is swapped out by the host. */
269 prev_state = exception_enter();
271 kvm_async_pf_task_wait((u32)read_cr2());
272 exception_exit(prev_state);
274 case KVM_PV_REASON_PAGE_READY:
277 kvm_async_pf_task_wake((u32)read_cr2());
282 NOKPROBE_SYMBOL(do_async_page_fault);
284 static void __init paravirt_ops_setup(void)
286 pv_info.name = "KVM";
288 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
289 pv_cpu_ops.io_delay = kvm_io_delay;
291 #ifdef CONFIG_X86_IO_APIC
296 static void kvm_register_steal_time(void)
298 int cpu = smp_processor_id();
299 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
301 if (!has_steal_clock)
304 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
305 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
306 cpu, (unsigned long long) slow_virt_to_phys(st));
309 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
311 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
314 * This relies on __test_and_clear_bit to modify the memory
315 * in a way that is atomic with respect to the local CPU.
316 * The hypervisor only accesses this memory from the local CPU so
317 * there's no need for lock or memory barriers.
318 * An optimization barrier is implied in apic write.
320 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
322 apic_write(APIC_EOI, APIC_EOI_ACK);
325 static void kvm_guest_cpu_init(void)
327 if (!kvm_para_available())
330 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
331 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
333 #ifdef CONFIG_PREEMPT
334 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
336 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
337 __this_cpu_write(apf_reason.enabled, 1);
338 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
342 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
344 /* Size alignment is implied but just to make it explicit. */
345 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
346 __this_cpu_write(kvm_apic_eoi, 0);
347 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
349 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
353 kvm_register_steal_time();
356 static void kvm_pv_disable_apf(void)
358 if (!__this_cpu_read(apf_reason.enabled))
361 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
362 __this_cpu_write(apf_reason.enabled, 0);
364 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
368 static void kvm_pv_guest_cpu_reboot(void *unused)
371 * We disable PV EOI before we load a new kernel by kexec,
372 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
373 * New kernel can re-enable when it boots.
375 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
376 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
377 kvm_pv_disable_apf();
378 kvm_disable_steal_time();
381 static int kvm_pv_reboot_notify(struct notifier_block *nb,
382 unsigned long code, void *unused)
384 if (code == SYS_RESTART)
385 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
389 static struct notifier_block kvm_pv_reboot_nb = {
390 .notifier_call = kvm_pv_reboot_notify,
393 static u64 kvm_steal_clock(int cpu)
396 struct kvm_steal_time *src;
399 src = &per_cpu(steal_time, cpu);
401 version = src->version;
405 } while ((version & 1) || (version != src->version));
410 void kvm_disable_steal_time(void)
412 if (!has_steal_clock)
415 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
419 static void __init kvm_smp_prepare_boot_cpu(void)
421 kvm_guest_cpu_init();
422 native_smp_prepare_boot_cpu();
426 static void kvm_guest_cpu_online(void *dummy)
428 kvm_guest_cpu_init();
431 static void kvm_guest_cpu_offline(void *dummy)
433 kvm_disable_steal_time();
434 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
435 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
436 kvm_pv_disable_apf();
440 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
443 int cpu = (unsigned long)hcpu;
446 case CPU_DOWN_FAILED:
447 case CPU_ONLINE_FROZEN:
448 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
450 case CPU_DOWN_PREPARE:
451 case CPU_DOWN_PREPARE_FROZEN:
452 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
460 static struct notifier_block kvm_cpu_notifier = {
461 .notifier_call = kvm_cpu_notify,
465 static void __init kvm_apf_trap_init(void)
467 set_intr_gate(14, async_page_fault);
470 void __init kvm_guest_init(void)
474 if (!kvm_para_available())
477 paravirt_ops_setup();
478 register_reboot_notifier(&kvm_pv_reboot_nb);
479 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
480 raw_spin_lock_init(&async_pf_sleepers[i].lock);
481 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
482 x86_init.irqs.trap_init = kvm_apf_trap_init;
484 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
486 pv_time_ops.steal_clock = kvm_steal_clock;
489 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
490 apic_set_eoi_write(kvm_guest_apic_eoi_write);
492 if (kvmclock_vsyscall)
493 kvm_setup_vsyscall_timeinfo();
496 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
497 register_cpu_notifier(&kvm_cpu_notifier);
499 kvm_guest_cpu_init();
503 * Hard lockup detection is enabled by default. Disable it, as guests
504 * can get false positives too easily, for example if the host is
507 hardlockup_detector_disable();
510 static noinline uint32_t __kvm_cpuid_base(void)
512 if (boot_cpu_data.cpuid_level < 0)
513 return 0; /* So we don't blow up on old processors */
515 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
516 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
521 static inline uint32_t kvm_cpuid_base(void)
523 static int kvm_cpuid_base = -1;
525 if (kvm_cpuid_base == -1)
526 kvm_cpuid_base = __kvm_cpuid_base();
528 return kvm_cpuid_base;
531 bool kvm_para_available(void)
533 return kvm_cpuid_base() != 0;
535 EXPORT_SYMBOL_GPL(kvm_para_available);
537 unsigned int kvm_arch_para_features(void)
539 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
542 static uint32_t __init kvm_detect(void)
544 return kvm_cpuid_base();
547 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
549 .detect = kvm_detect,
550 .x2apic_available = kvm_para_available,
552 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
554 static __init int activate_jump_labels(void)
556 if (has_steal_clock) {
557 static_key_slow_inc(¶virt_steal_enabled);
559 static_key_slow_inc(¶virt_steal_rq_enabled);
564 arch_initcall(activate_jump_labels);
566 #ifdef CONFIG_PARAVIRT_SPINLOCKS
568 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
569 static void kvm_kick_cpu(int cpu)
572 unsigned long flags = 0;
574 apicid = per_cpu(x86_cpu_to_apicid, cpu);
575 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
579 #ifdef CONFIG_QUEUED_SPINLOCKS
581 #include <asm/qspinlock.h>
583 static void kvm_wait(u8 *ptr, u8 val)
590 local_irq_save(flags);
592 if (READ_ONCE(*ptr) != val)
596 * halt until it's our turn and kicked. Note that we do safe halt
597 * for irq enabled case to avoid hang when lock info is overwritten
598 * in irq spinlock slowpath and no spurious interrupt occur to save us.
600 if (arch_irqs_disabled_flags(flags))
606 local_irq_restore(flags);
609 #else /* !CONFIG_QUEUED_SPINLOCKS */
611 enum kvm_contention_stat {
615 RELEASED_SLOW_KICKED,
619 #ifdef CONFIG_KVM_DEBUG_FS
620 #define HISTO_BUCKETS 30
622 static struct kvm_spinlock_stats
624 u32 contention_stats[NR_CONTENTION_STATS];
625 u32 histo_spin_blocked[HISTO_BUCKETS+1];
629 static u8 zero_stats;
631 static inline void check_zero(void)
636 old = READ_ONCE(zero_stats);
638 ret = cmpxchg(&zero_stats, old, 0);
639 /* This ensures only one fellow resets the stat */
641 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
645 static inline void add_stats(enum kvm_contention_stat var, u32 val)
648 spinlock_stats.contention_stats[var] += val;
652 static inline u64 spin_time_start(void)
654 return sched_clock();
657 static void __spin_time_accum(u64 delta, u32 *array)
661 index = ilog2(delta);
664 if (index < HISTO_BUCKETS)
667 array[HISTO_BUCKETS]++;
670 static inline void spin_time_accum_blocked(u64 start)
674 delta = sched_clock() - start;
675 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
676 spinlock_stats.time_blocked += delta;
679 static struct dentry *d_spin_debug;
680 static struct dentry *d_kvm_debug;
682 static struct dentry *kvm_init_debugfs(void)
684 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
686 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
691 static int __init kvm_spinlock_debugfs(void)
693 struct dentry *d_kvm;
695 d_kvm = kvm_init_debugfs();
699 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
701 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
703 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
704 &spinlock_stats.contention_stats[TAKEN_SLOW]);
705 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
706 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
708 debugfs_create_u32("released_slow", 0444, d_spin_debug,
709 &spinlock_stats.contention_stats[RELEASED_SLOW]);
710 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
711 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
713 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
714 &spinlock_stats.time_blocked);
716 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
717 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
721 fs_initcall(kvm_spinlock_debugfs);
722 #else /* !CONFIG_KVM_DEBUG_FS */
723 static inline void add_stats(enum kvm_contention_stat var, u32 val)
727 static inline u64 spin_time_start(void)
732 static inline void spin_time_accum_blocked(u64 start)
735 #endif /* CONFIG_KVM_DEBUG_FS */
737 struct kvm_lock_waiting {
738 struct arch_spinlock *lock;
742 /* cpus 'waiting' on a spinlock to become available */
743 static cpumask_t waiting_cpus;
745 /* Track spinlock on which a cpu is waiting */
746 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
748 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
750 struct kvm_lock_waiting *w;
759 w = this_cpu_ptr(&klock_waiting);
760 cpu = smp_processor_id();
761 start = spin_time_start();
764 * Make sure an interrupt handler can't upset things in a
765 * partially setup state.
767 local_irq_save(flags);
770 * The ordering protocol on this is that the "lock" pointer
771 * may only be set non-NULL if the "want" ticket is correct.
772 * If we're updating "want", we must first clear "lock".
780 add_stats(TAKEN_SLOW, 1);
783 * This uses set_bit, which is atomic but we should not rely on its
784 * reordering gurantees. So barrier is needed after this call.
786 cpumask_set_cpu(cpu, &waiting_cpus);
791 * Mark entry to slowpath before doing the pickup test to make
792 * sure we don't deadlock with an unlocker.
794 __ticket_enter_slowpath(lock);
796 /* make sure enter_slowpath, which is atomic does not cross the read */
797 smp_mb__after_atomic();
800 * check again make sure it didn't become free while
801 * we weren't looking.
803 head = READ_ONCE(lock->tickets.head);
804 if (__tickets_equal(head, want)) {
805 add_stats(TAKEN_SLOW_PICKUP, 1);
810 * halt until it's our turn and kicked. Note that we do safe halt
811 * for irq enabled case to avoid hang when lock info is overwritten
812 * in irq spinlock slowpath and no spurious interrupt occur to save us.
814 if (arch_irqs_disabled_flags(flags))
820 cpumask_clear_cpu(cpu, &waiting_cpus);
822 local_irq_restore(flags);
823 spin_time_accum_blocked(start);
825 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
827 /* Kick vcpu waiting on @lock->head to reach value @ticket */
828 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
832 add_stats(RELEASED_SLOW, 1);
833 for_each_cpu(cpu, &waiting_cpus) {
834 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
835 if (READ_ONCE(w->lock) == lock &&
836 READ_ONCE(w->want) == ticket) {
837 add_stats(RELEASED_SLOW_KICKED, 1);
844 #endif /* !CONFIG_QUEUED_SPINLOCKS */
847 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
849 void __init kvm_spinlock_init(void)
851 if (!kvm_para_available())
853 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
854 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
857 #ifdef CONFIG_QUEUED_SPINLOCKS
858 __pv_init_lock_hash();
859 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
860 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
861 pv_lock_ops.wait = kvm_wait;
862 pv_lock_ops.kick = kvm_kick_cpu;
863 #else /* !CONFIG_QUEUED_SPINLOCKS */
864 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
865 pv_lock_ops.unlock_kick = kvm_unlock_kick;
869 static __init int kvm_spinlock_init_jump(void)
871 if (!kvm_para_available())
873 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
876 static_key_slow_inc(¶virt_ticketlocks_enabled);
877 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
881 early_initcall(kvm_spinlock_init_jump);
883 #endif /* CONFIG_PARAVIRT_SPINLOCKS */