Loading...
1/*
2 * KVM paravirt_ops implementation
3 *
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.
8 *
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.
13 *
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.
17 *
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>
21 */
22
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>
28#include <linux/mm.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>
41#include <asm/cpu.h>
42#include <asm/traps.h>
43#include <asm/desc.h>
44#include <asm/tlbflush.h>
45#include <asm/apic.h>
46#include <asm/apicdef.h>
47#include <asm/hypervisor.h>
48#include <asm/kvm_guest.h>
49
50static int kvmapf = 1;
51
52static int __init parse_no_kvmapf(char *arg)
53{
54 kvmapf = 0;
55 return 0;
56}
57
58early_param("no-kvmapf", parse_no_kvmapf);
59
60static int steal_acc = 1;
61static int __init parse_no_stealacc(char *arg)
62{
63 steal_acc = 0;
64 return 0;
65}
66
67early_param("no-steal-acc", parse_no_stealacc);
68
69static int kvmclock_vsyscall = 1;
70static int __init parse_no_kvmclock_vsyscall(char *arg)
71{
72 kvmclock_vsyscall = 0;
73 return 0;
74}
75
76early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
77
78static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
79static DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64);
80static int has_steal_clock = 0;
81
82/*
83 * No need for any "IO delay" on KVM
84 */
85static void kvm_io_delay(void)
86{
87}
88
89#define KVM_TASK_SLEEP_HASHBITS 8
90#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
91
92struct kvm_task_sleep_node {
93 struct hlist_node link;
94 struct swait_queue_head wq;
95 u32 token;
96 int cpu;
97 bool halted;
98};
99
100static struct kvm_task_sleep_head {
101 raw_spinlock_t lock;
102 struct hlist_head list;
103} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
104
105static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
106 u32 token)
107{
108 struct hlist_node *p;
109
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)
114 return n;
115 }
116
117 return NULL;
118}
119
120/*
121 * @interrupt_kernel: Is this called from a routine which interrupts the kernel
122 * (other than user space)?
123 */
124void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
125{
126 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
127 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
128 struct kvm_task_sleep_node n, *e;
129 DECLARE_SWAITQUEUE(wait);
130
131 rcu_irq_enter();
132
133 raw_spin_lock(&b->lock);
134 e = _find_apf_task(b, token);
135 if (e) {
136 /* dummy entry exist -> wake up was delivered ahead of PF */
137 hlist_del(&e->link);
138 kfree(e);
139 raw_spin_unlock(&b->lock);
140
141 rcu_irq_exit();
142 return;
143 }
144
145 n.token = token;
146 n.cpu = smp_processor_id();
147 n.halted = is_idle_task(current) ||
148 (IS_ENABLED(CONFIG_PREEMPT_COUNT)
149 ? preempt_count() > 1 || rcu_preempt_depth()
150 : interrupt_kernel);
151 init_swait_queue_head(&n.wq);
152 hlist_add_head(&n.link, &b->list);
153 raw_spin_unlock(&b->lock);
154
155 for (;;) {
156 if (!n.halted)
157 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
158 if (hlist_unhashed(&n.link))
159 break;
160
161 rcu_irq_exit();
162
163 if (!n.halted) {
164 local_irq_enable();
165 schedule();
166 local_irq_disable();
167 } else {
168 /*
169 * We cannot reschedule. So halt.
170 */
171 native_safe_halt();
172 local_irq_disable();
173 }
174
175 rcu_irq_enter();
176 }
177 if (!n.halted)
178 finish_swait(&n.wq, &wait);
179
180 rcu_irq_exit();
181 return;
182}
183EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
184
185static void apf_task_wake_one(struct kvm_task_sleep_node *n)
186{
187 hlist_del_init(&n->link);
188 if (n->halted)
189 smp_send_reschedule(n->cpu);
190 else if (swq_has_sleeper(&n->wq))
191 swake_up(&n->wq);
192}
193
194static void apf_task_wake_all(void)
195{
196 int i;
197
198 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
199 struct hlist_node *p, *next;
200 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
201 raw_spin_lock(&b->lock);
202 hlist_for_each_safe(p, next, &b->list) {
203 struct kvm_task_sleep_node *n =
204 hlist_entry(p, typeof(*n), link);
205 if (n->cpu == smp_processor_id())
206 apf_task_wake_one(n);
207 }
208 raw_spin_unlock(&b->lock);
209 }
210}
211
212void kvm_async_pf_task_wake(u32 token)
213{
214 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
215 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
216 struct kvm_task_sleep_node *n;
217
218 if (token == ~0) {
219 apf_task_wake_all();
220 return;
221 }
222
223again:
224 raw_spin_lock(&b->lock);
225 n = _find_apf_task(b, token);
226 if (!n) {
227 /*
228 * async PF was not yet handled.
229 * Add dummy entry for the token.
230 */
231 n = kzalloc(sizeof(*n), GFP_ATOMIC);
232 if (!n) {
233 /*
234 * Allocation failed! Busy wait while other cpu
235 * handles async PF.
236 */
237 raw_spin_unlock(&b->lock);
238 cpu_relax();
239 goto again;
240 }
241 n->token = token;
242 n->cpu = smp_processor_id();
243 init_swait_queue_head(&n->wq);
244 hlist_add_head(&n->link, &b->list);
245 } else
246 apf_task_wake_one(n);
247 raw_spin_unlock(&b->lock);
248 return;
249}
250EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
251
252u32 kvm_read_and_reset_pf_reason(void)
253{
254 u32 reason = 0;
255
256 if (__this_cpu_read(apf_reason.enabled)) {
257 reason = __this_cpu_read(apf_reason.reason);
258 __this_cpu_write(apf_reason.reason, 0);
259 }
260
261 return reason;
262}
263EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
264NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
265
266dotraplinkage void
267do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
268{
269 enum ctx_state prev_state;
270
271 switch (kvm_read_and_reset_pf_reason()) {
272 default:
273 do_page_fault(regs, error_code);
274 break;
275 case KVM_PV_REASON_PAGE_NOT_PRESENT:
276 /* page is swapped out by the host. */
277 prev_state = exception_enter();
278 kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
279 exception_exit(prev_state);
280 break;
281 case KVM_PV_REASON_PAGE_READY:
282 rcu_irq_enter();
283 kvm_async_pf_task_wake((u32)read_cr2());
284 rcu_irq_exit();
285 break;
286 }
287}
288NOKPROBE_SYMBOL(do_async_page_fault);
289
290static void __init paravirt_ops_setup(void)
291{
292 pv_info.name = "KVM";
293
294 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
295 pv_cpu_ops.io_delay = kvm_io_delay;
296
297#ifdef CONFIG_X86_IO_APIC
298 no_timer_check = 1;
299#endif
300}
301
302static void kvm_register_steal_time(void)
303{
304 int cpu = smp_processor_id();
305 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
306
307 if (!has_steal_clock)
308 return;
309
310 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
311 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
312 cpu, (unsigned long long) slow_virt_to_phys(st));
313}
314
315static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
316
317static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
318{
319 /**
320 * This relies on __test_and_clear_bit to modify the memory
321 * in a way that is atomic with respect to the local CPU.
322 * The hypervisor only accesses this memory from the local CPU so
323 * there's no need for lock or memory barriers.
324 * An optimization barrier is implied in apic write.
325 */
326 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
327 return;
328 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
329}
330
331static void kvm_guest_cpu_init(void)
332{
333 if (!kvm_para_available())
334 return;
335
336 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
337 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
338
339#ifdef CONFIG_PREEMPT
340 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
341#endif
342 pa |= KVM_ASYNC_PF_ENABLED;
343
344 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
345 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
346
347 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
348 __this_cpu_write(apf_reason.enabled, 1);
349 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
350 smp_processor_id());
351 }
352
353 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
354 unsigned long pa;
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))
359 | KVM_MSR_ENABLED;
360 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
361 }
362
363 if (has_steal_clock)
364 kvm_register_steal_time();
365}
366
367static void kvm_pv_disable_apf(void)
368{
369 if (!__this_cpu_read(apf_reason.enabled))
370 return;
371
372 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
373 __this_cpu_write(apf_reason.enabled, 0);
374
375 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
376 smp_processor_id());
377}
378
379static void kvm_pv_guest_cpu_reboot(void *unused)
380{
381 /*
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.
385 */
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();
390}
391
392static int kvm_pv_reboot_notify(struct notifier_block *nb,
393 unsigned long code, void *unused)
394{
395 if (code == SYS_RESTART)
396 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
397 return NOTIFY_DONE;
398}
399
400static struct notifier_block kvm_pv_reboot_nb = {
401 .notifier_call = kvm_pv_reboot_notify,
402};
403
404static u64 kvm_steal_clock(int cpu)
405{
406 u64 steal;
407 struct kvm_steal_time *src;
408 int version;
409
410 src = &per_cpu(steal_time, cpu);
411 do {
412 version = src->version;
413 virt_rmb();
414 steal = src->steal;
415 virt_rmb();
416 } while ((version & 1) || (version != src->version));
417
418 return steal;
419}
420
421void kvm_disable_steal_time(void)
422{
423 if (!has_steal_clock)
424 return;
425
426 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
427}
428
429static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
430{
431 early_set_memory_decrypted((unsigned long) ptr, size);
432}
433
434/*
435 * Iterate through all possible CPUs and map the memory region pointed
436 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
437 *
438 * Note: we iterate through all possible CPUs to ensure that CPUs
439 * hotplugged will have their per-cpu variable already mapped as
440 * decrypted.
441 */
442static void __init sev_map_percpu_data(void)
443{
444 int cpu;
445
446 if (!sev_active())
447 return;
448
449 for_each_possible_cpu(cpu) {
450 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
451 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
452 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
453 }
454}
455
456#ifdef CONFIG_SMP
457static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
458{
459 native_smp_prepare_cpus(max_cpus);
460 if (kvm_para_has_hint(KVM_HINTS_REALTIME))
461 static_branch_disable(&virt_spin_lock_key);
462}
463
464static void __init kvm_smp_prepare_boot_cpu(void)
465{
466 /*
467 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
468 * shares the guest physical address with the hypervisor.
469 */
470 sev_map_percpu_data();
471
472 kvm_guest_cpu_init();
473 native_smp_prepare_boot_cpu();
474 kvm_spinlock_init();
475}
476
477static void kvm_guest_cpu_offline(void)
478{
479 kvm_disable_steal_time();
480 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
481 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
482 kvm_pv_disable_apf();
483 apf_task_wake_all();
484}
485
486static int kvm_cpu_online(unsigned int cpu)
487{
488 local_irq_disable();
489 kvm_guest_cpu_init();
490 local_irq_enable();
491 return 0;
492}
493
494static int kvm_cpu_down_prepare(unsigned int cpu)
495{
496 local_irq_disable();
497 kvm_guest_cpu_offline();
498 local_irq_enable();
499 return 0;
500}
501#endif
502
503static void __init kvm_apf_trap_init(void)
504{
505 update_intr_gate(X86_TRAP_PF, async_page_fault);
506}
507
508static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
509
510static void kvm_flush_tlb_others(const struct cpumask *cpumask,
511 const struct flush_tlb_info *info)
512{
513 u8 state;
514 int cpu;
515 struct kvm_steal_time *src;
516 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
517
518 cpumask_copy(flushmask, cpumask);
519 /*
520 * We have to call flush only on online vCPUs. And
521 * queue flush_on_enter for pre-empted vCPUs
522 */
523 for_each_cpu(cpu, flushmask) {
524 src = &per_cpu(steal_time, cpu);
525 state = READ_ONCE(src->preempted);
526 if ((state & KVM_VCPU_PREEMPTED)) {
527 if (try_cmpxchg(&src->preempted, &state,
528 state | KVM_VCPU_FLUSH_TLB))
529 __cpumask_clear_cpu(cpu, flushmask);
530 }
531 }
532
533 native_flush_tlb_others(flushmask, info);
534}
535
536static void __init kvm_guest_init(void)
537{
538 int i;
539
540 if (!kvm_para_available())
541 return;
542
543 paravirt_ops_setup();
544 register_reboot_notifier(&kvm_pv_reboot_nb);
545 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
546 raw_spin_lock_init(&async_pf_sleepers[i].lock);
547 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
548 x86_init.irqs.trap_init = kvm_apf_trap_init;
549
550 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
551 has_steal_clock = 1;
552 pv_time_ops.steal_clock = kvm_steal_clock;
553 }
554
555 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
556 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
557 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME))
558 pv_mmu_ops.flush_tlb_others = kvm_flush_tlb_others;
559
560 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
561 apic_set_eoi_write(kvm_guest_apic_eoi_write);
562
563 if (kvmclock_vsyscall)
564 kvm_setup_vsyscall_timeinfo();
565
566#ifdef CONFIG_SMP
567 smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
568 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
569 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
570 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
571 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
572#else
573 sev_map_percpu_data();
574 kvm_guest_cpu_init();
575#endif
576
577 /*
578 * Hard lockup detection is enabled by default. Disable it, as guests
579 * can get false positives too easily, for example if the host is
580 * overcommitted.
581 */
582 hardlockup_detector_disable();
583}
584
585static noinline uint32_t __kvm_cpuid_base(void)
586{
587 if (boot_cpu_data.cpuid_level < 0)
588 return 0; /* So we don't blow up on old processors */
589
590 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
591 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
592
593 return 0;
594}
595
596static inline uint32_t kvm_cpuid_base(void)
597{
598 static int kvm_cpuid_base = -1;
599
600 if (kvm_cpuid_base == -1)
601 kvm_cpuid_base = __kvm_cpuid_base();
602
603 return kvm_cpuid_base;
604}
605
606bool kvm_para_available(void)
607{
608 return kvm_cpuid_base() != 0;
609}
610EXPORT_SYMBOL_GPL(kvm_para_available);
611
612unsigned int kvm_arch_para_features(void)
613{
614 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
615}
616
617unsigned int kvm_arch_para_hints(void)
618{
619 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
620}
621
622static uint32_t __init kvm_detect(void)
623{
624 return kvm_cpuid_base();
625}
626
627const __initconst struct hypervisor_x86 x86_hyper_kvm = {
628 .name = "KVM",
629 .detect = kvm_detect,
630 .type = X86_HYPER_KVM,
631 .init.guest_late_init = kvm_guest_init,
632 .init.x2apic_available = kvm_para_available,
633};
634
635static __init int activate_jump_labels(void)
636{
637 if (has_steal_clock) {
638 static_key_slow_inc(¶virt_steal_enabled);
639 if (steal_acc)
640 static_key_slow_inc(¶virt_steal_rq_enabled);
641 }
642
643 return 0;
644}
645arch_initcall(activate_jump_labels);
646
647static __init int kvm_setup_pv_tlb_flush(void)
648{
649 int cpu;
650
651 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
652 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
653 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
654 for_each_possible_cpu(cpu) {
655 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
656 GFP_KERNEL, cpu_to_node(cpu));
657 }
658 pr_info("KVM setup pv remote TLB flush\n");
659 }
660
661 return 0;
662}
663arch_initcall(kvm_setup_pv_tlb_flush);
664
665#ifdef CONFIG_PARAVIRT_SPINLOCKS
666
667/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
668static void kvm_kick_cpu(int cpu)
669{
670 int apicid;
671 unsigned long flags = 0;
672
673 apicid = per_cpu(x86_cpu_to_apicid, cpu);
674 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
675}
676
677#include <asm/qspinlock.h>
678
679static void kvm_wait(u8 *ptr, u8 val)
680{
681 unsigned long flags;
682
683 if (in_nmi())
684 return;
685
686 local_irq_save(flags);
687
688 if (READ_ONCE(*ptr) != val)
689 goto out;
690
691 /*
692 * halt until it's our turn and kicked. Note that we do safe halt
693 * for irq enabled case to avoid hang when lock info is overwritten
694 * in irq spinlock slowpath and no spurious interrupt occur to save us.
695 */
696 if (arch_irqs_disabled_flags(flags))
697 halt();
698 else
699 safe_halt();
700
701out:
702 local_irq_restore(flags);
703}
704
705#ifdef CONFIG_X86_32
706__visible bool __kvm_vcpu_is_preempted(long cpu)
707{
708 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
709
710 return !!(src->preempted & KVM_VCPU_PREEMPTED);
711}
712PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
713
714#else
715
716#include <asm/asm-offsets.h>
717
718extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
719
720/*
721 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
722 * restoring to/from the stack.
723 */
724asm(
725".pushsection .text;"
726".global __raw_callee_save___kvm_vcpu_is_preempted;"
727".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
728"__raw_callee_save___kvm_vcpu_is_preempted:"
729"movq __per_cpu_offset(,%rdi,8), %rax;"
730"cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
731"setne %al;"
732"ret;"
733".popsection");
734
735#endif
736
737/*
738 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
739 */
740void __init kvm_spinlock_init(void)
741{
742 if (!kvm_para_available())
743 return;
744 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
745 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
746 return;
747
748 if (kvm_para_has_hint(KVM_HINTS_REALTIME))
749 return;
750
751 __pv_init_lock_hash();
752 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
753 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
754 pv_lock_ops.wait = kvm_wait;
755 pv_lock_ops.kick = kvm_kick_cpu;
756
757 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
758 pv_lock_ops.vcpu_is_preempted =
759 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
760 }
761}
762
763#endif /* CONFIG_PARAVIRT_SPINLOCKS */
1/*
2 * KVM paravirt_ops implementation
3 *
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.
8 *
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.
13 *
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.
17 *
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>
21 */
22
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>
28#include <linux/mm.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>
41#include <asm/cpu.h>
42#include <asm/traps.h>
43#include <asm/desc.h>
44#include <asm/tlbflush.h>
45#include <asm/apic.h>
46#include <asm/apicdef.h>
47#include <asm/hypervisor.h>
48#include <asm/kvm_guest.h>
49
50static int kvmapf = 1;
51
52static int parse_no_kvmapf(char *arg)
53{
54 kvmapf = 0;
55 return 0;
56}
57
58early_param("no-kvmapf", parse_no_kvmapf);
59
60static int steal_acc = 1;
61static int parse_no_stealacc(char *arg)
62{
63 steal_acc = 0;
64 return 0;
65}
66
67early_param("no-steal-acc", parse_no_stealacc);
68
69static int kvmclock_vsyscall = 1;
70static int parse_no_kvmclock_vsyscall(char *arg)
71{
72 kvmclock_vsyscall = 0;
73 return 0;
74}
75
76early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
77
78static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
79static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
80static int has_steal_clock = 0;
81
82/*
83 * No need for any "IO delay" on KVM
84 */
85static void kvm_io_delay(void)
86{
87}
88
89#define KVM_TASK_SLEEP_HASHBITS 8
90#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
91
92struct kvm_task_sleep_node {
93 struct hlist_node link;
94 struct swait_queue_head wq;
95 u32 token;
96 int cpu;
97 bool halted;
98};
99
100static struct kvm_task_sleep_head {
101 raw_spinlock_t lock;
102 struct hlist_head list;
103} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
104
105static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
106 u32 token)
107{
108 struct hlist_node *p;
109
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)
114 return n;
115 }
116
117 return NULL;
118}
119
120void kvm_async_pf_task_wait(u32 token)
121{
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);
126
127 rcu_irq_enter();
128
129 raw_spin_lock(&b->lock);
130 e = _find_apf_task(b, token);
131 if (e) {
132 /* dummy entry exist -> wake up was delivered ahead of PF */
133 hlist_del(&e->link);
134 kfree(e);
135 raw_spin_unlock(&b->lock);
136
137 rcu_irq_exit();
138 return;
139 }
140
141 n.token = token;
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);
147
148 for (;;) {
149 if (!n.halted)
150 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
151 if (hlist_unhashed(&n.link))
152 break;
153
154 if (!n.halted) {
155 local_irq_enable();
156 schedule();
157 local_irq_disable();
158 } else {
159 /*
160 * We cannot reschedule. So halt.
161 */
162 rcu_irq_exit();
163 native_safe_halt();
164 rcu_irq_enter();
165 local_irq_disable();
166 }
167 }
168 if (!n.halted)
169 finish_swait(&n.wq, &wait);
170
171 rcu_irq_exit();
172 return;
173}
174EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
175
176static void apf_task_wake_one(struct kvm_task_sleep_node *n)
177{
178 hlist_del_init(&n->link);
179 if (n->halted)
180 smp_send_reschedule(n->cpu);
181 else if (swait_active(&n->wq))
182 swake_up(&n->wq);
183}
184
185static void apf_task_wake_all(void)
186{
187 int i;
188
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);
198 }
199 raw_spin_unlock(&b->lock);
200 }
201}
202
203void kvm_async_pf_task_wake(u32 token)
204{
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;
208
209 if (token == ~0) {
210 apf_task_wake_all();
211 return;
212 }
213
214again:
215 raw_spin_lock(&b->lock);
216 n = _find_apf_task(b, token);
217 if (!n) {
218 /*
219 * async PF was not yet handled.
220 * Add dummy entry for the token.
221 */
222 n = kzalloc(sizeof(*n), GFP_ATOMIC);
223 if (!n) {
224 /*
225 * Allocation failed! Busy wait while other cpu
226 * handles async PF.
227 */
228 raw_spin_unlock(&b->lock);
229 cpu_relax();
230 goto again;
231 }
232 n->token = token;
233 n->cpu = smp_processor_id();
234 init_swait_queue_head(&n->wq);
235 hlist_add_head(&n->link, &b->list);
236 } else
237 apf_task_wake_one(n);
238 raw_spin_unlock(&b->lock);
239 return;
240}
241EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
242
243u32 kvm_read_and_reset_pf_reason(void)
244{
245 u32 reason = 0;
246
247 if (__this_cpu_read(apf_reason.enabled)) {
248 reason = __this_cpu_read(apf_reason.reason);
249 __this_cpu_write(apf_reason.reason, 0);
250 }
251
252 return reason;
253}
254EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
255NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
256
257dotraplinkage void
258do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
259{
260 enum ctx_state prev_state;
261
262 switch (kvm_read_and_reset_pf_reason()) {
263 default:
264 trace_do_page_fault(regs, error_code);
265 break;
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);
271 break;
272 case KVM_PV_REASON_PAGE_READY:
273 rcu_irq_enter();
274 kvm_async_pf_task_wake((u32)read_cr2());
275 rcu_irq_exit();
276 break;
277 }
278}
279NOKPROBE_SYMBOL(do_async_page_fault);
280
281static void __init paravirt_ops_setup(void)
282{
283 pv_info.name = "KVM";
284
285 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
286 pv_cpu_ops.io_delay = kvm_io_delay;
287
288#ifdef CONFIG_X86_IO_APIC
289 no_timer_check = 1;
290#endif
291}
292
293static void kvm_register_steal_time(void)
294{
295 int cpu = smp_processor_id();
296 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
297
298 if (!has_steal_clock)
299 return;
300
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));
304}
305
306static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
307
308static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
309{
310 /**
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.
316 */
317 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
318 return;
319 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
320}
321
322static void kvm_guest_cpu_init(void)
323{
324 if (!kvm_para_available())
325 return;
326
327 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
328 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
329
330#ifdef CONFIG_PREEMPT
331 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
332#endif
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",
336 smp_processor_id());
337 }
338
339 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
340 unsigned long pa;
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))
345 | KVM_MSR_ENABLED;
346 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
347 }
348
349 if (has_steal_clock)
350 kvm_register_steal_time();
351}
352
353static void kvm_pv_disable_apf(void)
354{
355 if (!__this_cpu_read(apf_reason.enabled))
356 return;
357
358 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
359 __this_cpu_write(apf_reason.enabled, 0);
360
361 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
362 smp_processor_id());
363}
364
365static void kvm_pv_guest_cpu_reboot(void *unused)
366{
367 /*
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.
371 */
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();
376}
377
378static int kvm_pv_reboot_notify(struct notifier_block *nb,
379 unsigned long code, void *unused)
380{
381 if (code == SYS_RESTART)
382 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
383 return NOTIFY_DONE;
384}
385
386static struct notifier_block kvm_pv_reboot_nb = {
387 .notifier_call = kvm_pv_reboot_notify,
388};
389
390static u64 kvm_steal_clock(int cpu)
391{
392 u64 steal;
393 struct kvm_steal_time *src;
394 int version;
395
396 src = &per_cpu(steal_time, cpu);
397 do {
398 version = src->version;
399 rmb();
400 steal = src->steal;
401 rmb();
402 } while ((version & 1) || (version != src->version));
403
404 return steal;
405}
406
407void kvm_disable_steal_time(void)
408{
409 if (!has_steal_clock)
410 return;
411
412 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
413}
414
415#ifdef CONFIG_SMP
416static void __init kvm_smp_prepare_boot_cpu(void)
417{
418 kvm_guest_cpu_init();
419 native_smp_prepare_boot_cpu();
420 kvm_spinlock_init();
421}
422
423static void kvm_guest_cpu_offline(void)
424{
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();
429 apf_task_wake_all();
430}
431
432static int kvm_cpu_online(unsigned int cpu)
433{
434 local_irq_disable();
435 kvm_guest_cpu_init();
436 local_irq_enable();
437 return 0;
438}
439
440static int kvm_cpu_down_prepare(unsigned int cpu)
441{
442 local_irq_disable();
443 kvm_guest_cpu_offline();
444 local_irq_enable();
445 return 0;
446}
447#endif
448
449static void __init kvm_apf_trap_init(void)
450{
451 set_intr_gate(14, async_page_fault);
452}
453
454void __init kvm_guest_init(void)
455{
456 int i;
457
458 if (!kvm_para_available())
459 return;
460
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;
467
468 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
469 has_steal_clock = 1;
470 pv_time_ops.steal_clock = kvm_steal_clock;
471 }
472
473 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
474 apic_set_eoi_write(kvm_guest_apic_eoi_write);
475
476 if (kvmclock_vsyscall)
477 kvm_setup_vsyscall_timeinfo();
478
479#ifdef CONFIG_SMP
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");
484#else
485 kvm_guest_cpu_init();
486#endif
487
488 /*
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
491 * overcommitted.
492 */
493 hardlockup_detector_disable();
494}
495
496static noinline uint32_t __kvm_cpuid_base(void)
497{
498 if (boot_cpu_data.cpuid_level < 0)
499 return 0; /* So we don't blow up on old processors */
500
501 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
502 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
503
504 return 0;
505}
506
507static inline uint32_t kvm_cpuid_base(void)
508{
509 static int kvm_cpuid_base = -1;
510
511 if (kvm_cpuid_base == -1)
512 kvm_cpuid_base = __kvm_cpuid_base();
513
514 return kvm_cpuid_base;
515}
516
517bool kvm_para_available(void)
518{
519 return kvm_cpuid_base() != 0;
520}
521EXPORT_SYMBOL_GPL(kvm_para_available);
522
523unsigned int kvm_arch_para_features(void)
524{
525 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
526}
527
528static uint32_t __init kvm_detect(void)
529{
530 return kvm_cpuid_base();
531}
532
533const struct hypervisor_x86 x86_hyper_kvm __refconst = {
534 .name = "KVM",
535 .detect = kvm_detect,
536 .x2apic_available = kvm_para_available,
537};
538EXPORT_SYMBOL_GPL(x86_hyper_kvm);
539
540static __init int activate_jump_labels(void)
541{
542 if (has_steal_clock) {
543 static_key_slow_inc(¶virt_steal_enabled);
544 if (steal_acc)
545 static_key_slow_inc(¶virt_steal_rq_enabled);
546 }
547
548 return 0;
549}
550arch_initcall(activate_jump_labels);
551
552#ifdef CONFIG_PARAVIRT_SPINLOCKS
553
554/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
555static void kvm_kick_cpu(int cpu)
556{
557 int apicid;
558 unsigned long flags = 0;
559
560 apicid = per_cpu(x86_cpu_to_apicid, cpu);
561 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
562}
563
564#include <asm/qspinlock.h>
565
566static void kvm_wait(u8 *ptr, u8 val)
567{
568 unsigned long flags;
569
570 if (in_nmi())
571 return;
572
573 local_irq_save(flags);
574
575 if (READ_ONCE(*ptr) != val)
576 goto out;
577
578 /*
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.
582 */
583 if (arch_irqs_disabled_flags(flags))
584 halt();
585 else
586 safe_halt();
587
588out:
589 local_irq_restore(flags);
590}
591
592__visible bool __kvm_vcpu_is_preempted(int cpu)
593{
594 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
595
596 return !!src->preempted;
597}
598PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
599
600/*
601 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
602 */
603void __init kvm_spinlock_init(void)
604{
605 if (!kvm_para_available())
606 return;
607 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
608 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
609 return;
610
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;
616
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);
620 }
621}
622
623static __init int kvm_spinlock_init_jump(void)
624{
625 if (!kvm_para_available())
626 return 0;
627 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
628 return 0;
629
630 static_key_slow_inc(¶virt_ticketlocks_enabled);
631 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
632
633 return 0;
634}
635early_initcall(kvm_spinlock_init_jump);
636
637#endif /* CONFIG_PARAVIRT_SPINLOCKS */