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))
160 * We cannot reschedule. So halt.
169 finish_wait(&n.wq, &wait);
174 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
176 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
178 hlist_del_init(&n->link);
180 smp_send_reschedule(n->cpu);
181 else if (waitqueue_active(&n->wq))
185 static void apf_task_wake_all(void)
189 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
190 struct hlist_node *p, *next;
191 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
193 hlist_for_each_safe(p, next, &b->list) {
194 struct kvm_task_sleep_node *n =
195 hlist_entry(p, typeof(*n), link);
196 if (n->cpu == smp_processor_id())
197 apf_task_wake_one(n);
199 spin_unlock(&b->lock);
203 void kvm_async_pf_task_wake(u32 token)
205 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
206 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
207 struct kvm_task_sleep_node *n;
216 n = _find_apf_task(b, token);
219 * async PF was not yet handled.
220 * Add dummy entry for the token.
222 n = kzalloc(sizeof(*n), GFP_ATOMIC);
225 * Allocation failed! Busy wait while other cpu
228 spin_unlock(&b->lock);
233 n->cpu = smp_processor_id();
234 init_waitqueue_head(&n->wq);
235 hlist_add_head(&n->link, &b->list);
237 apf_task_wake_one(n);
238 spin_unlock(&b->lock);
241 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243 u32 kvm_read_and_reset_pf_reason(void)
247 if (__this_cpu_read(apf_reason.enabled)) {
248 reason = __this_cpu_read(apf_reason.reason);
249 __this_cpu_write(apf_reason.reason, 0);
254 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
255 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
258 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
260 enum ctx_state prev_state;
262 switch (kvm_read_and_reset_pf_reason()) {
264 trace_do_page_fault(regs, error_code);
266 case KVM_PV_REASON_PAGE_NOT_PRESENT:
267 /* page is swapped out by the host. */
268 prev_state = exception_enter();
270 kvm_async_pf_task_wait((u32)read_cr2());
271 exception_exit(prev_state);
273 case KVM_PV_REASON_PAGE_READY:
276 kvm_async_pf_task_wake((u32)read_cr2());
281 NOKPROBE_SYMBOL(do_async_page_fault);
283 static void __init paravirt_ops_setup(void)
285 pv_info.name = "KVM";
288 * KVM isn't paravirt in the sense of paravirt_enabled. A KVM
289 * guest kernel works like a bare metal kernel with additional
290 * features, and paravirt_enabled is about features that are
293 pv_info.paravirt_enabled = 0;
295 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
296 pv_cpu_ops.io_delay = kvm_io_delay;
298 #ifdef CONFIG_X86_IO_APIC
303 static void kvm_register_steal_time(void)
305 int cpu = smp_processor_id();
306 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
308 if (!has_steal_clock)
311 memset(st, 0, sizeof(*st));
313 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
314 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
315 cpu, (unsigned long long) slow_virt_to_phys(st));
318 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
320 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
323 * This relies on __test_and_clear_bit to modify the memory
324 * in a way that is atomic with respect to the local CPU.
325 * The hypervisor only accesses this memory from the local CPU so
326 * there's no need for lock or memory barriers.
327 * An optimization barrier is implied in apic write.
329 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
331 apic_write(APIC_EOI, APIC_EOI_ACK);
334 void kvm_guest_cpu_init(void)
336 if (!kvm_para_available())
339 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
340 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
342 #ifdef CONFIG_PREEMPT
343 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
345 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
346 __this_cpu_write(apf_reason.enabled, 1);
347 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
351 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
353 /* Size alignment is implied but just to make it explicit. */
354 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
355 __this_cpu_write(kvm_apic_eoi, 0);
356 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
358 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
362 kvm_register_steal_time();
365 static void kvm_pv_disable_apf(void)
367 if (!__this_cpu_read(apf_reason.enabled))
370 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
371 __this_cpu_write(apf_reason.enabled, 0);
373 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
377 static void kvm_pv_guest_cpu_reboot(void *unused)
380 * We disable PV EOI before we load a new kernel by kexec,
381 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
382 * New kernel can re-enable when it boots.
384 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
385 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
386 kvm_pv_disable_apf();
387 kvm_disable_steal_time();
390 static int kvm_pv_reboot_notify(struct notifier_block *nb,
391 unsigned long code, void *unused)
393 if (code == SYS_RESTART)
394 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
398 static struct notifier_block kvm_pv_reboot_nb = {
399 .notifier_call = kvm_pv_reboot_notify,
402 static u64 kvm_steal_clock(int cpu)
405 struct kvm_steal_time *src;
408 src = &per_cpu(steal_time, cpu);
410 version = src->version;
414 } while ((version & 1) || (version != src->version));
419 void kvm_disable_steal_time(void)
421 if (!has_steal_clock)
424 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
428 static void __init kvm_smp_prepare_boot_cpu(void)
430 kvm_guest_cpu_init();
431 native_smp_prepare_boot_cpu();
435 static void kvm_guest_cpu_online(void *dummy)
437 kvm_guest_cpu_init();
440 static void kvm_guest_cpu_offline(void *dummy)
442 kvm_disable_steal_time();
443 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
444 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
445 kvm_pv_disable_apf();
449 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
452 int cpu = (unsigned long)hcpu;
455 case CPU_DOWN_FAILED:
456 case CPU_ONLINE_FROZEN:
457 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
459 case CPU_DOWN_PREPARE:
460 case CPU_DOWN_PREPARE_FROZEN:
461 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
469 static struct notifier_block kvm_cpu_notifier = {
470 .notifier_call = kvm_cpu_notify,
474 static void __init kvm_apf_trap_init(void)
476 set_intr_gate(14, async_page_fault);
479 void __init kvm_guest_init(void)
483 if (!kvm_para_available())
486 paravirt_ops_setup();
487 register_reboot_notifier(&kvm_pv_reboot_nb);
488 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
489 spin_lock_init(&async_pf_sleepers[i].lock);
490 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
491 x86_init.irqs.trap_init = kvm_apf_trap_init;
493 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
495 pv_time_ops.steal_clock = kvm_steal_clock;
498 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
499 apic_set_eoi_write(kvm_guest_apic_eoi_write);
501 if (kvmclock_vsyscall)
502 kvm_setup_vsyscall_timeinfo();
505 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
506 register_cpu_notifier(&kvm_cpu_notifier);
508 kvm_guest_cpu_init();
512 * Hard lockup detection is enabled by default. Disable it, as guests
513 * can get false positives too easily, for example if the host is
516 watchdog_enable_hardlockup_detector(false);
519 static noinline uint32_t __kvm_cpuid_base(void)
521 if (boot_cpu_data.cpuid_level < 0)
522 return 0; /* So we don't blow up on old processors */
524 if (cpu_has_hypervisor)
525 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
530 static inline uint32_t kvm_cpuid_base(void)
532 static int kvm_cpuid_base = -1;
534 if (kvm_cpuid_base == -1)
535 kvm_cpuid_base = __kvm_cpuid_base();
537 return kvm_cpuid_base;
540 bool kvm_para_available(void)
542 return kvm_cpuid_base() != 0;
544 EXPORT_SYMBOL_GPL(kvm_para_available);
546 unsigned int kvm_arch_para_features(void)
548 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
551 static uint32_t __init kvm_detect(void)
553 return kvm_cpuid_base();
556 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
558 .detect = kvm_detect,
559 .x2apic_available = kvm_para_available,
561 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
563 static __init int activate_jump_labels(void)
565 if (has_steal_clock) {
566 static_key_slow_inc(¶virt_steal_enabled);
568 static_key_slow_inc(¶virt_steal_rq_enabled);
573 arch_initcall(activate_jump_labels);
575 #ifdef CONFIG_PARAVIRT_SPINLOCKS
577 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
578 static void kvm_kick_cpu(int cpu)
581 unsigned long flags = 0;
583 apicid = per_cpu(x86_cpu_to_apicid, cpu);
584 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
587 enum kvm_contention_stat {
591 RELEASED_SLOW_KICKED,
595 #ifdef CONFIG_KVM_DEBUG_FS
596 #define HISTO_BUCKETS 30
598 static struct kvm_spinlock_stats
600 u32 contention_stats[NR_CONTENTION_STATS];
601 u32 histo_spin_blocked[HISTO_BUCKETS+1];
605 static u8 zero_stats;
607 static inline void check_zero(void)
612 old = ACCESS_ONCE(zero_stats);
614 ret = cmpxchg(&zero_stats, old, 0);
615 /* This ensures only one fellow resets the stat */
617 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
621 static inline void add_stats(enum kvm_contention_stat var, u32 val)
624 spinlock_stats.contention_stats[var] += val;
628 static inline u64 spin_time_start(void)
630 return sched_clock();
633 static void __spin_time_accum(u64 delta, u32 *array)
637 index = ilog2(delta);
640 if (index < HISTO_BUCKETS)
643 array[HISTO_BUCKETS]++;
646 static inline void spin_time_accum_blocked(u64 start)
650 delta = sched_clock() - start;
651 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
652 spinlock_stats.time_blocked += delta;
655 static struct dentry *d_spin_debug;
656 static struct dentry *d_kvm_debug;
658 struct dentry *kvm_init_debugfs(void)
660 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
662 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
667 static int __init kvm_spinlock_debugfs(void)
669 struct dentry *d_kvm;
671 d_kvm = kvm_init_debugfs();
675 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
677 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
679 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
680 &spinlock_stats.contention_stats[TAKEN_SLOW]);
681 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
682 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
684 debugfs_create_u32("released_slow", 0444, d_spin_debug,
685 &spinlock_stats.contention_stats[RELEASED_SLOW]);
686 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
687 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
689 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
690 &spinlock_stats.time_blocked);
692 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
693 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
697 fs_initcall(kvm_spinlock_debugfs);
698 #else /* !CONFIG_KVM_DEBUG_FS */
699 static inline void add_stats(enum kvm_contention_stat var, u32 val)
703 static inline u64 spin_time_start(void)
708 static inline void spin_time_accum_blocked(u64 start)
711 #endif /* CONFIG_KVM_DEBUG_FS */
713 struct kvm_lock_waiting {
714 struct arch_spinlock *lock;
718 /* cpus 'waiting' on a spinlock to become available */
719 static cpumask_t waiting_cpus;
721 /* Track spinlock on which a cpu is waiting */
722 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
724 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
726 struct kvm_lock_waiting *w;
734 w = this_cpu_ptr(&klock_waiting);
735 cpu = smp_processor_id();
736 start = spin_time_start();
739 * Make sure an interrupt handler can't upset things in a
740 * partially setup state.
742 local_irq_save(flags);
745 * The ordering protocol on this is that the "lock" pointer
746 * may only be set non-NULL if the "want" ticket is correct.
747 * If we're updating "want", we must first clear "lock".
755 add_stats(TAKEN_SLOW, 1);
758 * This uses set_bit, which is atomic but we should not rely on its
759 * reordering gurantees. So barrier is needed after this call.
761 cpumask_set_cpu(cpu, &waiting_cpus);
766 * Mark entry to slowpath before doing the pickup test to make
767 * sure we don't deadlock with an unlocker.
769 __ticket_enter_slowpath(lock);
772 * check again make sure it didn't become free while
773 * we weren't looking.
775 if (ACCESS_ONCE(lock->tickets.head) == want) {
776 add_stats(TAKEN_SLOW_PICKUP, 1);
781 * halt until it's our turn and kicked. Note that we do safe halt
782 * for irq enabled case to avoid hang when lock info is overwritten
783 * in irq spinlock slowpath and no spurious interrupt occur to save us.
785 if (arch_irqs_disabled_flags(flags))
791 cpumask_clear_cpu(cpu, &waiting_cpus);
793 local_irq_restore(flags);
794 spin_time_accum_blocked(start);
796 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
798 /* Kick vcpu waiting on @lock->head to reach value @ticket */
799 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
803 add_stats(RELEASED_SLOW, 1);
804 for_each_cpu(cpu, &waiting_cpus) {
805 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
806 if (ACCESS_ONCE(w->lock) == lock &&
807 ACCESS_ONCE(w->want) == ticket) {
808 add_stats(RELEASED_SLOW_KICKED, 1);
816 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
818 void __init kvm_spinlock_init(void)
820 if (!kvm_para_available())
822 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
823 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
826 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
827 pv_lock_ops.unlock_kick = kvm_unlock_kick;
830 static __init int kvm_spinlock_init_jump(void)
832 if (!kvm_para_available())
834 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
837 static_key_slow_inc(¶virt_ticketlocks_enabled);
838 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
842 early_initcall(kvm_spinlock_init_jump);
844 #endif /* CONFIG_PARAVIRT_SPINLOCKS */