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/module.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 <asm/timer.h>
41 #include <asm/traps.h>
43 #include <asm/tlbflush.h>
46 #include <asm/apicdef.h>
47 #include <asm/hypervisor.h>
48 #include <asm/kvm_guest.h>
50 static int kvmapf = 1;
52 static int parse_no_kvmapf(char *arg)
58 early_param("no-kvmapf", parse_no_kvmapf);
60 static int steal_acc = 1;
61 static int parse_no_stealacc(char *arg)
67 early_param("no-steal-acc", parse_no_stealacc);
69 static int kvmclock_vsyscall = 1;
70 static int parse_no_kvmclock_vsyscall(char *arg)
72 kvmclock_vsyscall = 0;
76 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
78 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
79 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
80 static int has_steal_clock = 0;
83 * No need for any "IO delay" on KVM
85 static void kvm_io_delay(void)
89 #define KVM_TASK_SLEEP_HASHBITS 8
90 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
92 struct kvm_task_sleep_node {
93 struct hlist_node link;
100 static struct kvm_task_sleep_head {
102 struct hlist_head list;
103 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
105 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
108 struct hlist_node *p;
110 hlist_for_each(p, &b->list) {
111 struct kvm_task_sleep_node *n =
112 hlist_entry(p, typeof(*n), link);
113 if (n->token == token)
120 void kvm_async_pf_task_wait(u32 token)
122 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
123 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
124 struct kvm_task_sleep_node n, *e;
130 e = _find_apf_task(b, token);
132 /* dummy entry exist -> wake up was delivered ahead of PF */
135 spin_unlock(&b->lock);
142 n.cpu = smp_processor_id();
143 n.halted = is_idle_task(current) || preempt_count() > 1;
144 init_waitqueue_head(&n.wq);
145 hlist_add_head(&n.link, &b->list);
146 spin_unlock(&b->lock);
150 prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
151 if (hlist_unhashed(&n.link))
162 * We cannot reschedule. So halt.
171 finish_wait(&n.wq, &wait);
176 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
178 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
180 hlist_del_init(&n->link);
182 smp_send_reschedule(n->cpu);
183 else if (waitqueue_active(&n->wq))
187 static void apf_task_wake_all(void)
191 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
192 struct hlist_node *p, *next;
193 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
195 hlist_for_each_safe(p, next, &b->list) {
196 struct kvm_task_sleep_node *n =
197 hlist_entry(p, typeof(*n), link);
198 if (n->cpu == smp_processor_id())
199 apf_task_wake_one(n);
201 spin_unlock(&b->lock);
205 void kvm_async_pf_task_wake(u32 token)
207 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
208 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
209 struct kvm_task_sleep_node *n;
218 n = _find_apf_task(b, token);
221 * async PF was not yet handled.
222 * Add dummy entry for the token.
224 n = kzalloc(sizeof(*n), GFP_ATOMIC);
227 * Allocation failed! Busy wait while other cpu
230 spin_unlock(&b->lock);
235 n->cpu = smp_processor_id();
236 init_waitqueue_head(&n->wq);
237 hlist_add_head(&n->link, &b->list);
239 apf_task_wake_one(n);
240 spin_unlock(&b->lock);
243 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
245 u32 kvm_read_and_reset_pf_reason(void)
249 if (__this_cpu_read(apf_reason.enabled)) {
250 reason = __this_cpu_read(apf_reason.reason);
251 __this_cpu_write(apf_reason.reason, 0);
256 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
257 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
260 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
262 enum ctx_state prev_state;
264 switch (kvm_read_and_reset_pf_reason()) {
266 trace_do_page_fault(regs, error_code);
268 case KVM_PV_REASON_PAGE_NOT_PRESENT:
269 /* page is swapped out by the host. */
270 prev_state = exception_enter();
272 kvm_async_pf_task_wait((u32)read_cr2());
273 exception_exit(prev_state);
275 case KVM_PV_REASON_PAGE_READY:
278 kvm_async_pf_task_wake((u32)read_cr2());
283 NOKPROBE_SYMBOL(do_async_page_fault);
285 static void __init paravirt_ops_setup(void)
287 pv_info.name = "KVM";
290 * KVM isn't paravirt in the sense of paravirt_enabled. A KVM
291 * guest kernel works like a bare metal kernel with additional
292 * features, and paravirt_enabled is about features that are
295 pv_info.paravirt_enabled = 0;
297 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
298 pv_cpu_ops.io_delay = kvm_io_delay;
300 #ifdef CONFIG_X86_IO_APIC
305 static void kvm_register_steal_time(void)
307 int cpu = smp_processor_id();
308 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
310 if (!has_steal_clock)
313 memset(st, 0, sizeof(*st));
315 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
316 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
317 cpu, (unsigned long long) slow_virt_to_phys(st));
320 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
322 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
325 * This relies on __test_and_clear_bit to modify the memory
326 * in a way that is atomic with respect to the local CPU.
327 * The hypervisor only accesses this memory from the local CPU so
328 * there's no need for lock or memory barriers.
329 * An optimization barrier is implied in apic write.
331 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
333 apic_write(APIC_EOI, APIC_EOI_ACK);
336 static void kvm_guest_cpu_init(void)
338 if (!kvm_para_available())
341 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
342 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
344 #ifdef CONFIG_PREEMPT
345 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
347 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
348 __this_cpu_write(apf_reason.enabled, 1);
349 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
353 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
355 /* Size alignment is implied but just to make it explicit. */
356 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
357 __this_cpu_write(kvm_apic_eoi, 0);
358 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
360 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
364 kvm_register_steal_time();
367 static void kvm_pv_disable_apf(void)
369 if (!__this_cpu_read(apf_reason.enabled))
372 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
373 __this_cpu_write(apf_reason.enabled, 0);
375 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
379 static void kvm_pv_guest_cpu_reboot(void *unused)
382 * We disable PV EOI before we load a new kernel by kexec,
383 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
384 * New kernel can re-enable when it boots.
386 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
387 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
388 kvm_pv_disable_apf();
389 kvm_disable_steal_time();
392 static int kvm_pv_reboot_notify(struct notifier_block *nb,
393 unsigned long code, void *unused)
395 if (code == SYS_RESTART)
396 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
400 static struct notifier_block kvm_pv_reboot_nb = {
401 .notifier_call = kvm_pv_reboot_notify,
404 static u64 kvm_steal_clock(int cpu)
407 struct kvm_steal_time *src;
410 src = &per_cpu(steal_time, cpu);
412 version = src->version;
416 } while ((version & 1) || (version != src->version));
421 void kvm_disable_steal_time(void)
423 if (!has_steal_clock)
426 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
430 static void __init kvm_smp_prepare_boot_cpu(void)
432 kvm_guest_cpu_init();
433 native_smp_prepare_boot_cpu();
437 static void kvm_guest_cpu_online(void *dummy)
439 kvm_guest_cpu_init();
442 static void kvm_guest_cpu_offline(void *dummy)
444 kvm_disable_steal_time();
445 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
446 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
447 kvm_pv_disable_apf();
451 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
454 int cpu = (unsigned long)hcpu;
457 case CPU_DOWN_FAILED:
458 case CPU_ONLINE_FROZEN:
459 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
461 case CPU_DOWN_PREPARE:
462 case CPU_DOWN_PREPARE_FROZEN:
463 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
471 static struct notifier_block kvm_cpu_notifier = {
472 .notifier_call = kvm_cpu_notify,
476 static void __init kvm_apf_trap_init(void)
478 set_intr_gate(14, async_page_fault);
481 void __init kvm_guest_init(void)
485 if (!kvm_para_available())
488 paravirt_ops_setup();
489 register_reboot_notifier(&kvm_pv_reboot_nb);
490 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
491 spin_lock_init(&async_pf_sleepers[i].lock);
492 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
493 x86_init.irqs.trap_init = kvm_apf_trap_init;
495 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
497 pv_time_ops.steal_clock = kvm_steal_clock;
500 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
501 apic_set_eoi_write(kvm_guest_apic_eoi_write);
503 if (kvmclock_vsyscall)
504 kvm_setup_vsyscall_timeinfo();
507 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
508 register_cpu_notifier(&kvm_cpu_notifier);
510 kvm_guest_cpu_init();
514 * Hard lockup detection is enabled by default. Disable it, as guests
515 * can get false positives too easily, for example if the host is
518 hardlockup_detector_disable();
521 static noinline uint32_t __kvm_cpuid_base(void)
523 if (boot_cpu_data.cpuid_level < 0)
524 return 0; /* So we don't blow up on old processors */
526 if (cpu_has_hypervisor)
527 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
532 static inline uint32_t kvm_cpuid_base(void)
534 static int kvm_cpuid_base = -1;
536 if (kvm_cpuid_base == -1)
537 kvm_cpuid_base = __kvm_cpuid_base();
539 return kvm_cpuid_base;
542 bool kvm_para_available(void)
544 return kvm_cpuid_base() != 0;
546 EXPORT_SYMBOL_GPL(kvm_para_available);
548 unsigned int kvm_arch_para_features(void)
550 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
553 static uint32_t __init kvm_detect(void)
555 return kvm_cpuid_base();
558 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
560 .detect = kvm_detect,
561 .x2apic_available = kvm_para_available,
563 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
565 static __init int activate_jump_labels(void)
567 if (has_steal_clock) {
568 static_key_slow_inc(¶virt_steal_enabled);
570 static_key_slow_inc(¶virt_steal_rq_enabled);
575 arch_initcall(activate_jump_labels);
577 #ifdef CONFIG_PARAVIRT_SPINLOCKS
579 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
580 static void kvm_kick_cpu(int cpu)
583 unsigned long flags = 0;
585 apicid = per_cpu(x86_cpu_to_apicid, cpu);
586 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
590 #ifdef CONFIG_QUEUED_SPINLOCKS
592 #include <asm/qspinlock.h>
594 static void kvm_wait(u8 *ptr, u8 val)
601 local_irq_save(flags);
603 if (READ_ONCE(*ptr) != val)
607 * halt until it's our turn and kicked. Note that we do safe halt
608 * for irq enabled case to avoid hang when lock info is overwritten
609 * in irq spinlock slowpath and no spurious interrupt occur to save us.
611 if (arch_irqs_disabled_flags(flags))
617 local_irq_restore(flags);
620 #else /* !CONFIG_QUEUED_SPINLOCKS */
622 enum kvm_contention_stat {
626 RELEASED_SLOW_KICKED,
630 #ifdef CONFIG_KVM_DEBUG_FS
631 #define HISTO_BUCKETS 30
633 static struct kvm_spinlock_stats
635 u32 contention_stats[NR_CONTENTION_STATS];
636 u32 histo_spin_blocked[HISTO_BUCKETS+1];
640 static u8 zero_stats;
642 static inline void check_zero(void)
647 old = READ_ONCE(zero_stats);
649 ret = cmpxchg(&zero_stats, old, 0);
650 /* This ensures only one fellow resets the stat */
652 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
656 static inline void add_stats(enum kvm_contention_stat var, u32 val)
659 spinlock_stats.contention_stats[var] += val;
663 static inline u64 spin_time_start(void)
665 return sched_clock();
668 static void __spin_time_accum(u64 delta, u32 *array)
672 index = ilog2(delta);
675 if (index < HISTO_BUCKETS)
678 array[HISTO_BUCKETS]++;
681 static inline void spin_time_accum_blocked(u64 start)
685 delta = sched_clock() - start;
686 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
687 spinlock_stats.time_blocked += delta;
690 static struct dentry *d_spin_debug;
691 static struct dentry *d_kvm_debug;
693 static struct dentry *kvm_init_debugfs(void)
695 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
697 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
702 static int __init kvm_spinlock_debugfs(void)
704 struct dentry *d_kvm;
706 d_kvm = kvm_init_debugfs();
710 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
712 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
714 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
715 &spinlock_stats.contention_stats[TAKEN_SLOW]);
716 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
717 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
719 debugfs_create_u32("released_slow", 0444, d_spin_debug,
720 &spinlock_stats.contention_stats[RELEASED_SLOW]);
721 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
722 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
724 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
725 &spinlock_stats.time_blocked);
727 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
728 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
732 fs_initcall(kvm_spinlock_debugfs);
733 #else /* !CONFIG_KVM_DEBUG_FS */
734 static inline void add_stats(enum kvm_contention_stat var, u32 val)
738 static inline u64 spin_time_start(void)
743 static inline void spin_time_accum_blocked(u64 start)
746 #endif /* CONFIG_KVM_DEBUG_FS */
748 struct kvm_lock_waiting {
749 struct arch_spinlock *lock;
753 /* cpus 'waiting' on a spinlock to become available */
754 static cpumask_t waiting_cpus;
756 /* Track spinlock on which a cpu is waiting */
757 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
759 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
761 struct kvm_lock_waiting *w;
770 w = this_cpu_ptr(&klock_waiting);
771 cpu = smp_processor_id();
772 start = spin_time_start();
775 * Make sure an interrupt handler can't upset things in a
776 * partially setup state.
778 local_irq_save(flags);
781 * The ordering protocol on this is that the "lock" pointer
782 * may only be set non-NULL if the "want" ticket is correct.
783 * If we're updating "want", we must first clear "lock".
791 add_stats(TAKEN_SLOW, 1);
794 * This uses set_bit, which is atomic but we should not rely on its
795 * reordering gurantees. So barrier is needed after this call.
797 cpumask_set_cpu(cpu, &waiting_cpus);
802 * Mark entry to slowpath before doing the pickup test to make
803 * sure we don't deadlock with an unlocker.
805 __ticket_enter_slowpath(lock);
807 /* make sure enter_slowpath, which is atomic does not cross the read */
808 smp_mb__after_atomic();
811 * check again make sure it didn't become free while
812 * we weren't looking.
814 head = READ_ONCE(lock->tickets.head);
815 if (__tickets_equal(head, want)) {
816 add_stats(TAKEN_SLOW_PICKUP, 1);
821 * halt until it's our turn and kicked. Note that we do safe halt
822 * for irq enabled case to avoid hang when lock info is overwritten
823 * in irq spinlock slowpath and no spurious interrupt occur to save us.
825 if (arch_irqs_disabled_flags(flags))
831 cpumask_clear_cpu(cpu, &waiting_cpus);
833 local_irq_restore(flags);
834 spin_time_accum_blocked(start);
836 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
838 /* Kick vcpu waiting on @lock->head to reach value @ticket */
839 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
843 add_stats(RELEASED_SLOW, 1);
844 for_each_cpu(cpu, &waiting_cpus) {
845 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
846 if (READ_ONCE(w->lock) == lock &&
847 READ_ONCE(w->want) == ticket) {
848 add_stats(RELEASED_SLOW_KICKED, 1);
855 #endif /* !CONFIG_QUEUED_SPINLOCKS */
858 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
860 void __init kvm_spinlock_init(void)
862 if (!kvm_para_available())
864 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
865 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
868 #ifdef CONFIG_QUEUED_SPINLOCKS
869 __pv_init_lock_hash();
870 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
871 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
872 pv_lock_ops.wait = kvm_wait;
873 pv_lock_ops.kick = kvm_kick_cpu;
874 #else /* !CONFIG_QUEUED_SPINLOCKS */
875 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
876 pv_lock_ops.unlock_kick = kvm_unlock_kick;
880 static __init int kvm_spinlock_init_jump(void)
882 if (!kvm_para_available())
884 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
887 static_key_slow_inc(¶virt_ticketlocks_enabled);
888 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
892 early_initcall(kvm_spinlock_init_jump);
894 #endif /* CONFIG_PARAVIRT_SPINLOCKS */