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
);
255 dotraplinkage
void __kprobes
256 do_async_page_fault(struct pt_regs
*regs
, unsigned long error_code
)
258 enum ctx_state prev_state
;
260 switch (kvm_read_and_reset_pf_reason()) {
262 do_page_fault(regs
, error_code
);
264 case KVM_PV_REASON_PAGE_NOT_PRESENT
:
265 /* page is swapped out by the host. */
266 prev_state
= exception_enter();
268 kvm_async_pf_task_wait((u32
)read_cr2());
269 exception_exit(prev_state
);
271 case KVM_PV_REASON_PAGE_READY
:
274 kvm_async_pf_task_wake((u32
)read_cr2());
280 static void __init
paravirt_ops_setup(void)
282 pv_info
.name
= "KVM";
283 pv_info
.paravirt_enabled
= 1;
285 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY
))
286 pv_cpu_ops
.io_delay
= kvm_io_delay
;
288 #ifdef CONFIG_X86_IO_APIC
293 static void kvm_register_steal_time(void)
295 int cpu
= smp_processor_id();
296 struct kvm_steal_time
*st
= &per_cpu(steal_time
, cpu
);
298 if (!has_steal_clock
)
301 memset(st
, 0, sizeof(*st
));
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
));
308 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi
) = KVM_PV_EOI_DISABLED
;
310 static void kvm_guest_apic_eoi_write(u32 reg
, u32 val
)
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.
319 if (__test_and_clear_bit(KVM_PV_EOI_BIT
, &__get_cpu_var(kvm_apic_eoi
)))
321 apic_write(APIC_EOI
, APIC_EOI_ACK
);
324 void kvm_guest_cpu_init(void)
326 if (!kvm_para_available())
329 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF
) && kvmapf
) {
330 u64 pa
= slow_virt_to_phys(&__get_cpu_var(apf_reason
));
332 #ifdef CONFIG_PREEMPT
333 pa
|= KVM_ASYNC_PF_SEND_ALWAYS
;
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",
341 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI
)) {
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
))
348 wrmsrl(MSR_KVM_PV_EOI_EN
, pa
);
352 kvm_register_steal_time();
355 static void kvm_pv_disable_apf(void)
357 if (!__get_cpu_var(apf_reason
).enabled
)
360 wrmsrl(MSR_KVM_ASYNC_PF_EN
, 0);
361 __get_cpu_var(apf_reason
).enabled
= 0;
363 printk(KERN_INFO
"Unregister pv shared memory for cpu %d\n",
367 static void kvm_pv_guest_cpu_reboot(void *unused
)
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.
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();
380 static int kvm_pv_reboot_notify(struct notifier_block
*nb
,
381 unsigned long code
, void *unused
)
383 if (code
== SYS_RESTART
)
384 on_each_cpu(kvm_pv_guest_cpu_reboot
, NULL
, 1);
388 static struct notifier_block kvm_pv_reboot_nb
= {
389 .notifier_call
= kvm_pv_reboot_notify
,
392 static u64
kvm_steal_clock(int cpu
)
395 struct kvm_steal_time
*src
;
398 src
= &per_cpu(steal_time
, cpu
);
400 version
= src
->version
;
404 } while ((version
& 1) || (version
!= src
->version
));
409 void kvm_disable_steal_time(void)
411 if (!has_steal_clock
)
414 wrmsr(MSR_KVM_STEAL_TIME
, 0, 0);
418 static void __init
kvm_smp_prepare_boot_cpu(void)
420 WARN_ON(kvm_register_clock("primary cpu clock"));
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 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 static noinline
uint32_t __kvm_cpuid_base(void)
505 if (boot_cpu_data
.cpuid_level
< 0)
506 return 0; /* So we don't blow up on old processors */
508 if (cpu_has_hypervisor
)
509 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
514 static inline uint32_t kvm_cpuid_base(void)
516 static int kvm_cpuid_base
= -1;
518 if (kvm_cpuid_base
== -1)
519 kvm_cpuid_base
= __kvm_cpuid_base();
521 return kvm_cpuid_base
;
524 bool kvm_para_available(void)
526 return kvm_cpuid_base() != 0;
528 EXPORT_SYMBOL_GPL(kvm_para_available
);
530 unsigned int kvm_arch_para_features(void)
532 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES
);
535 static uint32_t __init
kvm_detect(void)
537 return kvm_cpuid_base();
540 const struct hypervisor_x86 x86_hyper_kvm __refconst
= {
542 .detect
= kvm_detect
,
543 .x2apic_available
= kvm_para_available
,
545 EXPORT_SYMBOL_GPL(x86_hyper_kvm
);
547 static __init
int activate_jump_labels(void)
549 if (has_steal_clock
) {
550 static_key_slow_inc(¶virt_steal_enabled
);
552 static_key_slow_inc(¶virt_steal_rq_enabled
);
557 arch_initcall(activate_jump_labels
);
559 #ifdef CONFIG_PARAVIRT_SPINLOCKS
561 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
562 static void kvm_kick_cpu(int cpu
)
565 unsigned long flags
= 0;
567 apicid
= per_cpu(x86_cpu_to_apicid
, cpu
);
568 kvm_hypercall2(KVM_HC_KICK_CPU
, flags
, apicid
);
571 enum kvm_contention_stat
{
575 RELEASED_SLOW_KICKED
,
579 #ifdef CONFIG_KVM_DEBUG_FS
580 #define HISTO_BUCKETS 30
582 static struct kvm_spinlock_stats
584 u32 contention_stats
[NR_CONTENTION_STATS
];
585 u32 histo_spin_blocked
[HISTO_BUCKETS
+1];
589 static u8 zero_stats
;
591 static inline void check_zero(void)
596 old
= ACCESS_ONCE(zero_stats
);
598 ret
= cmpxchg(&zero_stats
, old
, 0);
599 /* This ensures only one fellow resets the stat */
601 memset(&spinlock_stats
, 0, sizeof(spinlock_stats
));
605 static inline void add_stats(enum kvm_contention_stat var
, u32 val
)
608 spinlock_stats
.contention_stats
[var
] += val
;
612 static inline u64
spin_time_start(void)
614 return sched_clock();
617 static void __spin_time_accum(u64 delta
, u32
*array
)
621 index
= ilog2(delta
);
624 if (index
< HISTO_BUCKETS
)
627 array
[HISTO_BUCKETS
]++;
630 static inline void spin_time_accum_blocked(u64 start
)
634 delta
= sched_clock() - start
;
635 __spin_time_accum(delta
, spinlock_stats
.histo_spin_blocked
);
636 spinlock_stats
.time_blocked
+= delta
;
639 static struct dentry
*d_spin_debug
;
640 static struct dentry
*d_kvm_debug
;
642 struct dentry
*kvm_init_debugfs(void)
644 d_kvm_debug
= debugfs_create_dir("kvm-guest", NULL
);
646 printk(KERN_WARNING
"Could not create 'kvm' debugfs directory\n");
651 static int __init
kvm_spinlock_debugfs(void)
653 struct dentry
*d_kvm
;
655 d_kvm
= kvm_init_debugfs();
659 d_spin_debug
= debugfs_create_dir("spinlocks", d_kvm
);
661 debugfs_create_u8("zero_stats", 0644, d_spin_debug
, &zero_stats
);
663 debugfs_create_u32("taken_slow", 0444, d_spin_debug
,
664 &spinlock_stats
.contention_stats
[TAKEN_SLOW
]);
665 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug
,
666 &spinlock_stats
.contention_stats
[TAKEN_SLOW_PICKUP
]);
668 debugfs_create_u32("released_slow", 0444, d_spin_debug
,
669 &spinlock_stats
.contention_stats
[RELEASED_SLOW
]);
670 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug
,
671 &spinlock_stats
.contention_stats
[RELEASED_SLOW_KICKED
]);
673 debugfs_create_u64("time_blocked", 0444, d_spin_debug
,
674 &spinlock_stats
.time_blocked
);
676 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug
,
677 spinlock_stats
.histo_spin_blocked
, HISTO_BUCKETS
+ 1);
681 fs_initcall(kvm_spinlock_debugfs
);
682 #else /* !CONFIG_KVM_DEBUG_FS */
683 static inline void add_stats(enum kvm_contention_stat var
, u32 val
)
687 static inline u64
spin_time_start(void)
692 static inline void spin_time_accum_blocked(u64 start
)
695 #endif /* CONFIG_KVM_DEBUG_FS */
697 struct kvm_lock_waiting
{
698 struct arch_spinlock
*lock
;
702 /* cpus 'waiting' on a spinlock to become available */
703 static cpumask_t waiting_cpus
;
705 /* Track spinlock on which a cpu is waiting */
706 static DEFINE_PER_CPU(struct kvm_lock_waiting
, klock_waiting
);
708 static void kvm_lock_spinning(struct arch_spinlock
*lock
, __ticket_t want
)
710 struct kvm_lock_waiting
*w
;
718 w
= &__get_cpu_var(klock_waiting
);
719 cpu
= smp_processor_id();
720 start
= spin_time_start();
723 * Make sure an interrupt handler can't upset things in a
724 * partially setup state.
726 local_irq_save(flags
);
729 * The ordering protocol on this is that the "lock" pointer
730 * may only be set non-NULL if the "want" ticket is correct.
731 * If we're updating "want", we must first clear "lock".
739 add_stats(TAKEN_SLOW
, 1);
742 * This uses set_bit, which is atomic but we should not rely on its
743 * reordering gurantees. So barrier is needed after this call.
745 cpumask_set_cpu(cpu
, &waiting_cpus
);
750 * Mark entry to slowpath before doing the pickup test to make
751 * sure we don't deadlock with an unlocker.
753 __ticket_enter_slowpath(lock
);
756 * check again make sure it didn't become free while
757 * we weren't looking.
759 if (ACCESS_ONCE(lock
->tickets
.head
) == want
) {
760 add_stats(TAKEN_SLOW_PICKUP
, 1);
765 * halt until it's our turn and kicked. Note that we do safe halt
766 * for irq enabled case to avoid hang when lock info is overwritten
767 * in irq spinlock slowpath and no spurious interrupt occur to save us.
769 if (arch_irqs_disabled_flags(flags
))
775 cpumask_clear_cpu(cpu
, &waiting_cpus
);
777 local_irq_restore(flags
);
778 spin_time_accum_blocked(start
);
780 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning
);
782 /* Kick vcpu waiting on @lock->head to reach value @ticket */
783 static void kvm_unlock_kick(struct arch_spinlock
*lock
, __ticket_t ticket
)
787 add_stats(RELEASED_SLOW
, 1);
788 for_each_cpu(cpu
, &waiting_cpus
) {
789 const struct kvm_lock_waiting
*w
= &per_cpu(klock_waiting
, cpu
);
790 if (ACCESS_ONCE(w
->lock
) == lock
&&
791 ACCESS_ONCE(w
->want
) == ticket
) {
792 add_stats(RELEASED_SLOW_KICKED
, 1);
800 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
802 void __init
kvm_spinlock_init(void)
804 if (!kvm_para_available())
806 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
807 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT
))
810 pv_lock_ops
.lock_spinning
= PV_CALLEE_SAVE(kvm_lock_spinning
);
811 pv_lock_ops
.unlock_kick
= kvm_unlock_kick
;
814 static __init
int kvm_spinlock_init_jump(void)
816 if (!kvm_para_available())
818 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT
))
821 static_key_slow_inc(¶virt_ticketlocks_enabled
);
822 printk(KERN_INFO
"KVM setup paravirtual spinlock\n");
826 early_initcall(kvm_spinlock_init_jump
);
828 #endif /* CONFIG_PARAVIRT_SPINLOCKS */