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 <asm/timer.h>
40 #include <asm/traps.h>
42 #include <asm/tlbflush.h>
45 #include <asm/apicdef.h>
46 #include <asm/hypervisor.h>
47 #include <asm/kvm_guest.h>
49 static int kvmapf = 1;
51 static int parse_no_kvmapf(char *arg)
57 early_param("no-kvmapf", parse_no_kvmapf);
59 static int steal_acc = 1;
60 static int parse_no_stealacc(char *arg)
66 early_param("no-steal-acc", parse_no_stealacc);
68 static int kvmclock_vsyscall = 1;
69 static int parse_no_kvmclock_vsyscall(char *arg)
71 kvmclock_vsyscall = 0;
75 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
77 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
78 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
79 static int has_steal_clock = 0;
82 * No need for any "IO delay" on KVM
84 static void kvm_io_delay(void)
88 #define KVM_TASK_SLEEP_HASHBITS 8
89 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
91 struct kvm_task_sleep_node {
92 struct hlist_node link;
99 static struct kvm_task_sleep_head {
101 struct hlist_head list;
102 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
104 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
107 struct hlist_node *p;
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)
119 void kvm_async_pf_task_wait(u32 token)
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;
129 e = _find_apf_task(b, token);
131 /* dummy entry exist -> wake up was delivered ahead of PF */
134 spin_unlock(&b->lock);
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);
149 prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
150 if (hlist_unhashed(&n.link))
159 * We cannot reschedule. So halt.
168 finish_wait(&n.wq, &wait);
173 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
175 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
177 hlist_del_init(&n->link);
179 smp_send_reschedule(n->cpu);
180 else if (waitqueue_active(&n->wq))
184 static void apf_task_wake_all(void)
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];
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);
198 spin_unlock(&b->lock);
202 void kvm_async_pf_task_wake(u32 token)
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;
215 n = _find_apf_task(b, token);
218 * async PF was not yet handled.
219 * Add dummy entry for the token.
221 n = kzalloc(sizeof(*n), GFP_ATOMIC);
224 * Allocation failed! Busy wait while other cpu
227 spin_unlock(&b->lock);
232 n->cpu = smp_processor_id();
233 init_waitqueue_head(&n->wq);
234 hlist_add_head(&n->link, &b->list);
236 apf_task_wake_one(n);
237 spin_unlock(&b->lock);
240 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
242 u32 kvm_read_and_reset_pf_reason(void)
246 if (__get_cpu_var(apf_reason).enabled) {
247 reason = __get_cpu_var(apf_reason).reason;
248 __get_cpu_var(apf_reason).reason = 0;
253 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
254 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
257 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
259 enum ctx_state prev_state;
261 switch (kvm_read_and_reset_pf_reason()) {
263 trace_do_page_fault(regs, error_code);
265 case KVM_PV_REASON_PAGE_NOT_PRESENT:
266 /* page is swapped out by the host. */
267 prev_state = exception_enter();
269 kvm_async_pf_task_wait((u32)read_cr2());
270 exception_exit(prev_state);
272 case KVM_PV_REASON_PAGE_READY:
275 kvm_async_pf_task_wake((u32)read_cr2());
280 NOKPROBE_SYMBOL(do_async_page_fault);
282 static void __init paravirt_ops_setup(void)
284 pv_info.name = "KVM";
285 pv_info.paravirt_enabled = 1;
287 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
288 pv_cpu_ops.io_delay = kvm_io_delay;
290 #ifdef CONFIG_X86_IO_APIC
295 static void kvm_register_steal_time(void)
297 int cpu = smp_processor_id();
298 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
300 if (!has_steal_clock)
303 memset(st, 0, sizeof(*st));
305 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
306 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
307 cpu, (unsigned long long) slow_virt_to_phys(st));
310 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
312 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
315 * This relies on __test_and_clear_bit to modify the memory
316 * in a way that is atomic with respect to the local CPU.
317 * The hypervisor only accesses this memory from the local CPU so
318 * there's no need for lock or memory barriers.
319 * An optimization barrier is implied in apic write.
321 if (__test_and_clear_bit(KVM_PV_EOI_BIT, &__get_cpu_var(kvm_apic_eoi)))
323 apic_write(APIC_EOI, APIC_EOI_ACK);
326 void kvm_guest_cpu_init(void)
328 if (!kvm_para_available())
331 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
332 u64 pa = slow_virt_to_phys(&__get_cpu_var(apf_reason));
334 #ifdef CONFIG_PREEMPT
335 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
337 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
338 __get_cpu_var(apf_reason).enabled = 1;
339 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
343 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
345 /* Size alignment is implied but just to make it explicit. */
346 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
347 __get_cpu_var(kvm_apic_eoi) = 0;
348 pa = slow_virt_to_phys(&__get_cpu_var(kvm_apic_eoi))
350 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
354 kvm_register_steal_time();
357 static void kvm_pv_disable_apf(void)
359 if (!__get_cpu_var(apf_reason).enabled)
362 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
363 __get_cpu_var(apf_reason).enabled = 0;
365 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
369 static void kvm_pv_guest_cpu_reboot(void *unused)
372 * We disable PV EOI before we load a new kernel by kexec,
373 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
374 * New kernel can re-enable when it boots.
376 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
377 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
378 kvm_pv_disable_apf();
379 kvm_disable_steal_time();
382 static int kvm_pv_reboot_notify(struct notifier_block *nb,
383 unsigned long code, void *unused)
385 if (code == SYS_RESTART)
386 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
390 static struct notifier_block kvm_pv_reboot_nb = {
391 .notifier_call = kvm_pv_reboot_notify,
394 static u64 kvm_steal_clock(int cpu)
397 struct kvm_steal_time *src;
400 src = &per_cpu(steal_time, cpu);
402 version = src->version;
406 } while ((version & 1) || (version != src->version));
411 void kvm_disable_steal_time(void)
413 if (!has_steal_clock)
416 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
420 static void __init kvm_smp_prepare_boot_cpu(void)
422 kvm_guest_cpu_init();
423 native_smp_prepare_boot_cpu();
427 static void kvm_guest_cpu_online(void *dummy)
429 kvm_guest_cpu_init();
432 static void kvm_guest_cpu_offline(void *dummy)
434 kvm_disable_steal_time();
435 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
436 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
437 kvm_pv_disable_apf();
441 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
444 int cpu = (unsigned long)hcpu;
447 case CPU_DOWN_FAILED:
448 case CPU_ONLINE_FROZEN:
449 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
451 case CPU_DOWN_PREPARE:
452 case CPU_DOWN_PREPARE_FROZEN:
453 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
461 static struct notifier_block kvm_cpu_notifier = {
462 .notifier_call = kvm_cpu_notify,
466 static void __init kvm_apf_trap_init(void)
468 set_intr_gate(14, async_page_fault);
471 void __init kvm_guest_init(void)
475 if (!kvm_para_available())
478 paravirt_ops_setup();
479 register_reboot_notifier(&kvm_pv_reboot_nb);
480 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
481 spin_lock_init(&async_pf_sleepers[i].lock);
482 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
483 x86_init.irqs.trap_init = kvm_apf_trap_init;
485 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
487 pv_time_ops.steal_clock = kvm_steal_clock;
490 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
491 apic_set_eoi_write(kvm_guest_apic_eoi_write);
493 if (kvmclock_vsyscall)
494 kvm_setup_vsyscall_timeinfo();
497 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
498 register_cpu_notifier(&kvm_cpu_notifier);
500 kvm_guest_cpu_init();
504 static noinline uint32_t __kvm_cpuid_base(void)
506 if (boot_cpu_data.cpuid_level < 0)
507 return 0; /* So we don't blow up on old processors */
509 if (cpu_has_hypervisor)
510 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
515 static inline uint32_t kvm_cpuid_base(void)
517 static int kvm_cpuid_base = -1;
519 if (kvm_cpuid_base == -1)
520 kvm_cpuid_base = __kvm_cpuid_base();
522 return kvm_cpuid_base;
525 bool kvm_para_available(void)
527 return kvm_cpuid_base() != 0;
529 EXPORT_SYMBOL_GPL(kvm_para_available);
531 unsigned int kvm_arch_para_features(void)
533 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
536 static uint32_t __init kvm_detect(void)
538 return kvm_cpuid_base();
541 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
543 .detect = kvm_detect,
544 .x2apic_available = kvm_para_available,
546 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
548 static __init int activate_jump_labels(void)
550 if (has_steal_clock) {
551 static_key_slow_inc(¶virt_steal_enabled);
553 static_key_slow_inc(¶virt_steal_rq_enabled);
558 arch_initcall(activate_jump_labels);
560 #ifdef CONFIG_PARAVIRT_SPINLOCKS
562 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
563 static void kvm_kick_cpu(int cpu)
566 unsigned long flags = 0;
568 apicid = per_cpu(x86_cpu_to_apicid, cpu);
569 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
572 enum kvm_contention_stat {
576 RELEASED_SLOW_KICKED,
580 #ifdef CONFIG_KVM_DEBUG_FS
581 #define HISTO_BUCKETS 30
583 static struct kvm_spinlock_stats
585 u32 contention_stats[NR_CONTENTION_STATS];
586 u32 histo_spin_blocked[HISTO_BUCKETS+1];
590 static u8 zero_stats;
592 static inline void check_zero(void)
597 old = ACCESS_ONCE(zero_stats);
599 ret = cmpxchg(&zero_stats, old, 0);
600 /* This ensures only one fellow resets the stat */
602 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
606 static inline void add_stats(enum kvm_contention_stat var, u32 val)
609 spinlock_stats.contention_stats[var] += val;
613 static inline u64 spin_time_start(void)
615 return sched_clock();
618 static void __spin_time_accum(u64 delta, u32 *array)
622 index = ilog2(delta);
625 if (index < HISTO_BUCKETS)
628 array[HISTO_BUCKETS]++;
631 static inline void spin_time_accum_blocked(u64 start)
635 delta = sched_clock() - start;
636 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
637 spinlock_stats.time_blocked += delta;
640 static struct dentry *d_spin_debug;
641 static struct dentry *d_kvm_debug;
643 struct dentry *kvm_init_debugfs(void)
645 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
647 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
652 static int __init kvm_spinlock_debugfs(void)
654 struct dentry *d_kvm;
656 d_kvm = kvm_init_debugfs();
660 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
662 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
664 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
665 &spinlock_stats.contention_stats[TAKEN_SLOW]);
666 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
667 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
669 debugfs_create_u32("released_slow", 0444, d_spin_debug,
670 &spinlock_stats.contention_stats[RELEASED_SLOW]);
671 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
672 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
674 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
675 &spinlock_stats.time_blocked);
677 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
678 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
682 fs_initcall(kvm_spinlock_debugfs);
683 #else /* !CONFIG_KVM_DEBUG_FS */
684 static inline void add_stats(enum kvm_contention_stat var, u32 val)
688 static inline u64 spin_time_start(void)
693 static inline void spin_time_accum_blocked(u64 start)
696 #endif /* CONFIG_KVM_DEBUG_FS */
698 struct kvm_lock_waiting {
699 struct arch_spinlock *lock;
703 /* cpus 'waiting' on a spinlock to become available */
704 static cpumask_t waiting_cpus;
706 /* Track spinlock on which a cpu is waiting */
707 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
709 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
711 struct kvm_lock_waiting *w;
719 w = &__get_cpu_var(klock_waiting);
720 cpu = smp_processor_id();
721 start = spin_time_start();
724 * Make sure an interrupt handler can't upset things in a
725 * partially setup state.
727 local_irq_save(flags);
730 * The ordering protocol on this is that the "lock" pointer
731 * may only be set non-NULL if the "want" ticket is correct.
732 * If we're updating "want", we must first clear "lock".
740 add_stats(TAKEN_SLOW, 1);
743 * This uses set_bit, which is atomic but we should not rely on its
744 * reordering gurantees. So barrier is needed after this call.
746 cpumask_set_cpu(cpu, &waiting_cpus);
751 * Mark entry to slowpath before doing the pickup test to make
752 * sure we don't deadlock with an unlocker.
754 __ticket_enter_slowpath(lock);
757 * check again make sure it didn't become free while
758 * we weren't looking.
760 if (ACCESS_ONCE(lock->tickets.head) == want) {
761 add_stats(TAKEN_SLOW_PICKUP, 1);
766 * halt until it's our turn and kicked. Note that we do safe halt
767 * for irq enabled case to avoid hang when lock info is overwritten
768 * in irq spinlock slowpath and no spurious interrupt occur to save us.
770 if (arch_irqs_disabled_flags(flags))
776 cpumask_clear_cpu(cpu, &waiting_cpus);
778 local_irq_restore(flags);
779 spin_time_accum_blocked(start);
781 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
783 /* Kick vcpu waiting on @lock->head to reach value @ticket */
784 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
788 add_stats(RELEASED_SLOW, 1);
789 for_each_cpu(cpu, &waiting_cpus) {
790 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
791 if (ACCESS_ONCE(w->lock) == lock &&
792 ACCESS_ONCE(w->want) == ticket) {
793 add_stats(RELEASED_SLOW_KICKED, 1);
801 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
803 void __init kvm_spinlock_init(void)
805 if (!kvm_para_available())
807 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
808 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
811 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
812 pv_lock_ops.unlock_kick = kvm_unlock_kick;
815 static __init int kvm_spinlock_init_jump(void)
817 if (!kvm_para_available())
819 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
822 static_key_slow_inc(¶virt_ticketlocks_enabled);
823 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
827 early_initcall(kvm_spinlock_init_jump);
829 #endif /* CONFIG_PARAVIRT_SPINLOCKS */