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>
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;
94 struct swait_queue_head wq;
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;
125 DECLARE_SWAITQUEUE(wait);
129 raw_spin_lock(&b->lock);
130 e = _find_apf_task(b, token);
132 /* dummy entry exist -> wake up was delivered ahead of PF */
135 raw_spin_unlock(&b->lock);
142 n.cpu = smp_processor_id();
143 n.halted = is_idle_task(current) || preempt_count() > 1;
144 init_swait_queue_head(&n.wq);
145 hlist_add_head(&n.link, &b->list);
146 raw_spin_unlock(&b->lock);
150 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
151 if (hlist_unhashed(&n.link))
160 * We cannot reschedule. So halt.
169 finish_swait(&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 (swait_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];
192 raw_spin_lock(&b->lock);
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 raw_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;
215 raw_spin_lock(&b->lock);
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 raw_spin_unlock(&b->lock);
233 n->cpu = smp_processor_id();
234 init_swait_queue_head(&n->wq);
235 hlist_add_head(&n->link, &b->list);
237 apf_task_wake_one(n);
238 raw_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();
269 kvm_async_pf_task_wait((u32)read_cr2());
270 exception_exit(prev_state);
272 case KVM_PV_REASON_PAGE_READY:
274 kvm_async_pf_task_wake((u32)read_cr2());
279 NOKPROBE_SYMBOL(do_async_page_fault);
281 static void __init paravirt_ops_setup(void)
283 pv_info.name = "KVM";
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 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
302 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
303 cpu, (unsigned long long) slow_virt_to_phys(st));
306 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
308 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
311 * This relies on __test_and_clear_bit to modify the memory
312 * in a way that is atomic with respect to the local CPU.
313 * The hypervisor only accesses this memory from the local CPU so
314 * there's no need for lock or memory barriers.
315 * An optimization barrier is implied in apic write.
317 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
319 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
322 static void kvm_guest_cpu_init(void)
324 if (!kvm_para_available())
327 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
328 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
330 #ifdef CONFIG_PREEMPT
331 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
333 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
334 __this_cpu_write(apf_reason.enabled, 1);
335 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
339 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
341 /* Size alignment is implied but just to make it explicit. */
342 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
343 __this_cpu_write(kvm_apic_eoi, 0);
344 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
346 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
350 kvm_register_steal_time();
353 static void kvm_pv_disable_apf(void)
355 if (!__this_cpu_read(apf_reason.enabled))
358 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
359 __this_cpu_write(apf_reason.enabled, 0);
361 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
365 static void kvm_pv_guest_cpu_reboot(void *unused)
368 * We disable PV EOI before we load a new kernel by kexec,
369 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
370 * New kernel can re-enable when it boots.
372 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
373 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
374 kvm_pv_disable_apf();
375 kvm_disable_steal_time();
378 static int kvm_pv_reboot_notify(struct notifier_block *nb,
379 unsigned long code, void *unused)
381 if (code == SYS_RESTART)
382 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
386 static struct notifier_block kvm_pv_reboot_nb = {
387 .notifier_call = kvm_pv_reboot_notify,
390 static u64 kvm_steal_clock(int cpu)
393 struct kvm_steal_time *src;
396 src = &per_cpu(steal_time, cpu);
398 version = src->version;
402 } while ((version & 1) || (version != src->version));
407 void kvm_disable_steal_time(void)
409 if (!has_steal_clock)
412 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
416 static void __init kvm_smp_prepare_boot_cpu(void)
418 kvm_guest_cpu_init();
419 native_smp_prepare_boot_cpu();
423 static void kvm_guest_cpu_offline(void)
425 kvm_disable_steal_time();
426 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
427 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
428 kvm_pv_disable_apf();
432 static int kvm_cpu_online(unsigned int cpu)
435 kvm_guest_cpu_init();
440 static int kvm_cpu_down_prepare(unsigned int cpu)
443 kvm_guest_cpu_offline();
449 static void __init kvm_apf_trap_init(void)
451 set_intr_gate(14, async_page_fault);
454 void __init kvm_guest_init(void)
458 if (!kvm_para_available())
461 paravirt_ops_setup();
462 register_reboot_notifier(&kvm_pv_reboot_nb);
463 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
464 raw_spin_lock_init(&async_pf_sleepers[i].lock);
465 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
466 x86_init.irqs.trap_init = kvm_apf_trap_init;
468 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
470 pv_time_ops.steal_clock = kvm_steal_clock;
473 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
474 apic_set_eoi_write(kvm_guest_apic_eoi_write);
476 if (kvmclock_vsyscall)
477 kvm_setup_vsyscall_timeinfo();
480 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
481 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
482 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
483 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
485 kvm_guest_cpu_init();
489 * Hard lockup detection is enabled by default. Disable it, as guests
490 * can get false positives too easily, for example if the host is
493 hardlockup_detector_disable();
496 static noinline uint32_t __kvm_cpuid_base(void)
498 if (boot_cpu_data.cpuid_level < 0)
499 return 0; /* So we don't blow up on old processors */
501 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
502 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
507 static inline uint32_t kvm_cpuid_base(void)
509 static int kvm_cpuid_base = -1;
511 if (kvm_cpuid_base == -1)
512 kvm_cpuid_base = __kvm_cpuid_base();
514 return kvm_cpuid_base;
517 bool kvm_para_available(void)
519 return kvm_cpuid_base() != 0;
521 EXPORT_SYMBOL_GPL(kvm_para_available);
523 unsigned int kvm_arch_para_features(void)
525 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
528 static uint32_t __init kvm_detect(void)
530 return kvm_cpuid_base();
533 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
535 .detect = kvm_detect,
536 .x2apic_available = kvm_para_available,
538 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
540 static __init int activate_jump_labels(void)
542 if (has_steal_clock) {
543 static_key_slow_inc(¶virt_steal_enabled);
545 static_key_slow_inc(¶virt_steal_rq_enabled);
550 arch_initcall(activate_jump_labels);
552 #ifdef CONFIG_PARAVIRT_SPINLOCKS
554 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
555 static void kvm_kick_cpu(int cpu)
558 unsigned long flags = 0;
560 apicid = per_cpu(x86_cpu_to_apicid, cpu);
561 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
564 #include <asm/qspinlock.h>
566 static void kvm_wait(u8 *ptr, u8 val)
573 local_irq_save(flags);
575 if (READ_ONCE(*ptr) != val)
579 * halt until it's our turn and kicked. Note that we do safe halt
580 * for irq enabled case to avoid hang when lock info is overwritten
581 * in irq spinlock slowpath and no spurious interrupt occur to save us.
583 if (arch_irqs_disabled_flags(flags))
589 local_irq_restore(flags);
592 __visible bool __kvm_vcpu_is_preempted(int cpu)
594 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
596 return !!src->preempted;
598 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
601 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
603 void __init kvm_spinlock_init(void)
605 if (!kvm_para_available())
607 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
608 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
611 __pv_init_lock_hash();
612 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
613 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
614 pv_lock_ops.wait = kvm_wait;
615 pv_lock_ops.kick = kvm_kick_cpu;
617 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
618 pv_lock_ops.vcpu_is_preempted =
619 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
623 #endif /* CONFIG_PARAVIRT_SPINLOCKS */