Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * KVM paravirt_ops implementation
4 *
5 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Copyright IBM Corporation, 2007
7 * Authors: Anthony Liguori <aliguori@us.ibm.com>
8 */
9
10#include <linux/context_tracking.h>
11#include <linux/init.h>
12#include <linux/kernel.h>
13#include <linux/kvm_para.h>
14#include <linux/cpu.h>
15#include <linux/mm.h>
16#include <linux/highmem.h>
17#include <linux/hardirq.h>
18#include <linux/notifier.h>
19#include <linux/reboot.h>
20#include <linux/hash.h>
21#include <linux/sched.h>
22#include <linux/slab.h>
23#include <linux/kprobes.h>
24#include <linux/debugfs.h>
25#include <linux/nmi.h>
26#include <linux/swait.h>
27#include <asm/timer.h>
28#include <asm/cpu.h>
29#include <asm/traps.h>
30#include <asm/desc.h>
31#include <asm/tlbflush.h>
32#include <asm/apic.h>
33#include <asm/apicdef.h>
34#include <asm/hypervisor.h>
35#include <asm/tlb.h>
36
37static int kvmapf = 1;
38
39static int __init parse_no_kvmapf(char *arg)
40{
41 kvmapf = 0;
42 return 0;
43}
44
45early_param("no-kvmapf", parse_no_kvmapf);
46
47static int steal_acc = 1;
48static int __init parse_no_stealacc(char *arg)
49{
50 steal_acc = 0;
51 return 0;
52}
53
54early_param("no-steal-acc", parse_no_stealacc);
55
56static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
57DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
58static int has_steal_clock = 0;
59
60/*
61 * No need for any "IO delay" on KVM
62 */
63static void kvm_io_delay(void)
64{
65}
66
67#define KVM_TASK_SLEEP_HASHBITS 8
68#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
69
70struct kvm_task_sleep_node {
71 struct hlist_node link;
72 struct swait_queue_head wq;
73 u32 token;
74 int cpu;
75 bool halted;
76};
77
78static struct kvm_task_sleep_head {
79 raw_spinlock_t lock;
80 struct hlist_head list;
81} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
82
83static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
84 u32 token)
85{
86 struct hlist_node *p;
87
88 hlist_for_each(p, &b->list) {
89 struct kvm_task_sleep_node *n =
90 hlist_entry(p, typeof(*n), link);
91 if (n->token == token)
92 return n;
93 }
94
95 return NULL;
96}
97
98/*
99 * @interrupt_kernel: Is this called from a routine which interrupts the kernel
100 * (other than user space)?
101 */
102void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
103{
104 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
105 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
106 struct kvm_task_sleep_node n, *e;
107 DECLARE_SWAITQUEUE(wait);
108
109 rcu_irq_enter();
110
111 raw_spin_lock(&b->lock);
112 e = _find_apf_task(b, token);
113 if (e) {
114 /* dummy entry exist -> wake up was delivered ahead of PF */
115 hlist_del(&e->link);
116 kfree(e);
117 raw_spin_unlock(&b->lock);
118
119 rcu_irq_exit();
120 return;
121 }
122
123 n.token = token;
124 n.cpu = smp_processor_id();
125 n.halted = is_idle_task(current) ||
126 (IS_ENABLED(CONFIG_PREEMPT_COUNT)
127 ? preempt_count() > 1 || rcu_preempt_depth()
128 : interrupt_kernel);
129 init_swait_queue_head(&n.wq);
130 hlist_add_head(&n.link, &b->list);
131 raw_spin_unlock(&b->lock);
132
133 for (;;) {
134 if (!n.halted)
135 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
136 if (hlist_unhashed(&n.link))
137 break;
138
139 rcu_irq_exit();
140
141 if (!n.halted) {
142 local_irq_enable();
143 schedule();
144 local_irq_disable();
145 } else {
146 /*
147 * We cannot reschedule. So halt.
148 */
149 native_safe_halt();
150 local_irq_disable();
151 }
152
153 rcu_irq_enter();
154 }
155 if (!n.halted)
156 finish_swait(&n.wq, &wait);
157
158 rcu_irq_exit();
159 return;
160}
161EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
162
163static void apf_task_wake_one(struct kvm_task_sleep_node *n)
164{
165 hlist_del_init(&n->link);
166 if (n->halted)
167 smp_send_reschedule(n->cpu);
168 else if (swq_has_sleeper(&n->wq))
169 swake_up_one(&n->wq);
170}
171
172static void apf_task_wake_all(void)
173{
174 int i;
175
176 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
177 struct hlist_node *p, *next;
178 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
179 raw_spin_lock(&b->lock);
180 hlist_for_each_safe(p, next, &b->list) {
181 struct kvm_task_sleep_node *n =
182 hlist_entry(p, typeof(*n), link);
183 if (n->cpu == smp_processor_id())
184 apf_task_wake_one(n);
185 }
186 raw_spin_unlock(&b->lock);
187 }
188}
189
190void kvm_async_pf_task_wake(u32 token)
191{
192 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
193 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
194 struct kvm_task_sleep_node *n;
195
196 if (token == ~0) {
197 apf_task_wake_all();
198 return;
199 }
200
201again:
202 raw_spin_lock(&b->lock);
203 n = _find_apf_task(b, token);
204 if (!n) {
205 /*
206 * async PF was not yet handled.
207 * Add dummy entry for the token.
208 */
209 n = kzalloc(sizeof(*n), GFP_ATOMIC);
210 if (!n) {
211 /*
212 * Allocation failed! Busy wait while other cpu
213 * handles async PF.
214 */
215 raw_spin_unlock(&b->lock);
216 cpu_relax();
217 goto again;
218 }
219 n->token = token;
220 n->cpu = smp_processor_id();
221 init_swait_queue_head(&n->wq);
222 hlist_add_head(&n->link, &b->list);
223 } else
224 apf_task_wake_one(n);
225 raw_spin_unlock(&b->lock);
226 return;
227}
228EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
229
230u32 kvm_read_and_reset_pf_reason(void)
231{
232 u32 reason = 0;
233
234 if (__this_cpu_read(apf_reason.enabled)) {
235 reason = __this_cpu_read(apf_reason.reason);
236 __this_cpu_write(apf_reason.reason, 0);
237 }
238
239 return reason;
240}
241EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
242NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
243
244dotraplinkage void
245do_async_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
246{
247 enum ctx_state prev_state;
248
249 switch (kvm_read_and_reset_pf_reason()) {
250 default:
251 do_page_fault(regs, error_code, address);
252 break;
253 case KVM_PV_REASON_PAGE_NOT_PRESENT:
254 /* page is swapped out by the host. */
255 prev_state = exception_enter();
256 kvm_async_pf_task_wait((u32)address, !user_mode(regs));
257 exception_exit(prev_state);
258 break;
259 case KVM_PV_REASON_PAGE_READY:
260 rcu_irq_enter();
261 kvm_async_pf_task_wake((u32)address);
262 rcu_irq_exit();
263 break;
264 }
265}
266NOKPROBE_SYMBOL(do_async_page_fault);
267
268static void __init paravirt_ops_setup(void)
269{
270 pv_info.name = "KVM";
271
272 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
273 pv_ops.cpu.io_delay = kvm_io_delay;
274
275#ifdef CONFIG_X86_IO_APIC
276 no_timer_check = 1;
277#endif
278}
279
280static void kvm_register_steal_time(void)
281{
282 int cpu = smp_processor_id();
283 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
284
285 if (!has_steal_clock)
286 return;
287
288 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
289 pr_info("kvm-stealtime: cpu %d, msr %llx\n",
290 cpu, (unsigned long long) slow_virt_to_phys(st));
291}
292
293static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
294
295static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
296{
297 /**
298 * This relies on __test_and_clear_bit to modify the memory
299 * in a way that is atomic with respect to the local CPU.
300 * The hypervisor only accesses this memory from the local CPU so
301 * there's no need for lock or memory barriers.
302 * An optimization barrier is implied in apic write.
303 */
304 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
305 return;
306 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
307}
308
309static void kvm_guest_cpu_init(void)
310{
311 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
312 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
313
314#ifdef CONFIG_PREEMPTION
315 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
316#endif
317 pa |= KVM_ASYNC_PF_ENABLED;
318
319 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
320 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
321
322 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
323 __this_cpu_write(apf_reason.enabled, 1);
324 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
325 smp_processor_id());
326 }
327
328 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
329 unsigned long pa;
330 /* Size alignment is implied but just to make it explicit. */
331 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
332 __this_cpu_write(kvm_apic_eoi, 0);
333 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
334 | KVM_MSR_ENABLED;
335 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
336 }
337
338 if (has_steal_clock)
339 kvm_register_steal_time();
340}
341
342static void kvm_pv_disable_apf(void)
343{
344 if (!__this_cpu_read(apf_reason.enabled))
345 return;
346
347 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
348 __this_cpu_write(apf_reason.enabled, 0);
349
350 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
351 smp_processor_id());
352}
353
354static void kvm_pv_guest_cpu_reboot(void *unused)
355{
356 /*
357 * We disable PV EOI before we load a new kernel by kexec,
358 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
359 * New kernel can re-enable when it boots.
360 */
361 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
362 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
363 kvm_pv_disable_apf();
364 kvm_disable_steal_time();
365}
366
367static int kvm_pv_reboot_notify(struct notifier_block *nb,
368 unsigned long code, void *unused)
369{
370 if (code == SYS_RESTART)
371 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
372 return NOTIFY_DONE;
373}
374
375static struct notifier_block kvm_pv_reboot_nb = {
376 .notifier_call = kvm_pv_reboot_notify,
377};
378
379static u64 kvm_steal_clock(int cpu)
380{
381 u64 steal;
382 struct kvm_steal_time *src;
383 int version;
384
385 src = &per_cpu(steal_time, cpu);
386 do {
387 version = src->version;
388 virt_rmb();
389 steal = src->steal;
390 virt_rmb();
391 } while ((version & 1) || (version != src->version));
392
393 return steal;
394}
395
396void kvm_disable_steal_time(void)
397{
398 if (!has_steal_clock)
399 return;
400
401 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
402}
403
404static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
405{
406 early_set_memory_decrypted((unsigned long) ptr, size);
407}
408
409/*
410 * Iterate through all possible CPUs and map the memory region pointed
411 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
412 *
413 * Note: we iterate through all possible CPUs to ensure that CPUs
414 * hotplugged will have their per-cpu variable already mapped as
415 * decrypted.
416 */
417static void __init sev_map_percpu_data(void)
418{
419 int cpu;
420
421 if (!sev_active())
422 return;
423
424 for_each_possible_cpu(cpu) {
425 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
426 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
427 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
428 }
429}
430
431#ifdef CONFIG_SMP
432#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
433
434static void __send_ipi_mask(const struct cpumask *mask, int vector)
435{
436 unsigned long flags;
437 int cpu, apic_id, icr;
438 int min = 0, max = 0;
439#ifdef CONFIG_X86_64
440 __uint128_t ipi_bitmap = 0;
441#else
442 u64 ipi_bitmap = 0;
443#endif
444 long ret;
445
446 if (cpumask_empty(mask))
447 return;
448
449 local_irq_save(flags);
450
451 switch (vector) {
452 default:
453 icr = APIC_DM_FIXED | vector;
454 break;
455 case NMI_VECTOR:
456 icr = APIC_DM_NMI;
457 break;
458 }
459
460 for_each_cpu(cpu, mask) {
461 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
462 if (!ipi_bitmap) {
463 min = max = apic_id;
464 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
465 ipi_bitmap <<= min - apic_id;
466 min = apic_id;
467 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
468 max = apic_id < max ? max : apic_id;
469 } else {
470 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
471 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
472 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
473 min = max = apic_id;
474 ipi_bitmap = 0;
475 }
476 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
477 }
478
479 if (ipi_bitmap) {
480 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
481 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
482 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
483 }
484
485 local_irq_restore(flags);
486}
487
488static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
489{
490 __send_ipi_mask(mask, vector);
491}
492
493static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
494{
495 unsigned int this_cpu = smp_processor_id();
496 struct cpumask new_mask;
497 const struct cpumask *local_mask;
498
499 cpumask_copy(&new_mask, mask);
500 cpumask_clear_cpu(this_cpu, &new_mask);
501 local_mask = &new_mask;
502 __send_ipi_mask(local_mask, vector);
503}
504
505/*
506 * Set the IPI entry points
507 */
508static void kvm_setup_pv_ipi(void)
509{
510 apic->send_IPI_mask = kvm_send_ipi_mask;
511 apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
512 pr_info("KVM setup pv IPIs\n");
513}
514
515static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
516{
517 int cpu;
518
519 native_send_call_func_ipi(mask);
520
521 /* Make sure other vCPUs get a chance to run if they need to. */
522 for_each_cpu(cpu, mask) {
523 if (vcpu_is_preempted(cpu)) {
524 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
525 break;
526 }
527 }
528}
529
530static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
531{
532 native_smp_prepare_cpus(max_cpus);
533 if (kvm_para_has_hint(KVM_HINTS_REALTIME))
534 static_branch_disable(&virt_spin_lock_key);
535}
536
537static void __init kvm_smp_prepare_boot_cpu(void)
538{
539 /*
540 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
541 * shares the guest physical address with the hypervisor.
542 */
543 sev_map_percpu_data();
544
545 kvm_guest_cpu_init();
546 native_smp_prepare_boot_cpu();
547 kvm_spinlock_init();
548}
549
550static void kvm_guest_cpu_offline(void)
551{
552 kvm_disable_steal_time();
553 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
554 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
555 kvm_pv_disable_apf();
556 apf_task_wake_all();
557}
558
559static int kvm_cpu_online(unsigned int cpu)
560{
561 local_irq_disable();
562 kvm_guest_cpu_init();
563 local_irq_enable();
564 return 0;
565}
566
567static int kvm_cpu_down_prepare(unsigned int cpu)
568{
569 local_irq_disable();
570 kvm_guest_cpu_offline();
571 local_irq_enable();
572 return 0;
573}
574#endif
575
576static void __init kvm_apf_trap_init(void)
577{
578 update_intr_gate(X86_TRAP_PF, async_page_fault);
579}
580
581static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
582
583static void kvm_flush_tlb_others(const struct cpumask *cpumask,
584 const struct flush_tlb_info *info)
585{
586 u8 state;
587 int cpu;
588 struct kvm_steal_time *src;
589 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
590
591 cpumask_copy(flushmask, cpumask);
592 /*
593 * We have to call flush only on online vCPUs. And
594 * queue flush_on_enter for pre-empted vCPUs
595 */
596 for_each_cpu(cpu, flushmask) {
597 src = &per_cpu(steal_time, cpu);
598 state = READ_ONCE(src->preempted);
599 if ((state & KVM_VCPU_PREEMPTED)) {
600 if (try_cmpxchg(&src->preempted, &state,
601 state | KVM_VCPU_FLUSH_TLB))
602 __cpumask_clear_cpu(cpu, flushmask);
603 }
604 }
605
606 native_flush_tlb_others(flushmask, info);
607}
608
609static void __init kvm_guest_init(void)
610{
611 int i;
612
613 paravirt_ops_setup();
614 register_reboot_notifier(&kvm_pv_reboot_nb);
615 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
616 raw_spin_lock_init(&async_pf_sleepers[i].lock);
617 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
618 x86_init.irqs.trap_init = kvm_apf_trap_init;
619
620 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
621 has_steal_clock = 1;
622 pv_ops.time.steal_clock = kvm_steal_clock;
623 }
624
625 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
626 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
627 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
628 pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
629 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
630 }
631
632 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
633 apic_set_eoi_write(kvm_guest_apic_eoi_write);
634
635#ifdef CONFIG_SMP
636 smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
637 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
638 if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
639 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
640 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
641 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
642 pr_info("KVM setup pv sched yield\n");
643 }
644 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
645 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
646 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
647#else
648 sev_map_percpu_data();
649 kvm_guest_cpu_init();
650#endif
651
652 /*
653 * Hard lockup detection is enabled by default. Disable it, as guests
654 * can get false positives too easily, for example if the host is
655 * overcommitted.
656 */
657 hardlockup_detector_disable();
658}
659
660static noinline uint32_t __kvm_cpuid_base(void)
661{
662 if (boot_cpu_data.cpuid_level < 0)
663 return 0; /* So we don't blow up on old processors */
664
665 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
666 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
667
668 return 0;
669}
670
671static inline uint32_t kvm_cpuid_base(void)
672{
673 static int kvm_cpuid_base = -1;
674
675 if (kvm_cpuid_base == -1)
676 kvm_cpuid_base = __kvm_cpuid_base();
677
678 return kvm_cpuid_base;
679}
680
681bool kvm_para_available(void)
682{
683 return kvm_cpuid_base() != 0;
684}
685EXPORT_SYMBOL_GPL(kvm_para_available);
686
687unsigned int kvm_arch_para_features(void)
688{
689 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
690}
691
692unsigned int kvm_arch_para_hints(void)
693{
694 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
695}
696EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
697
698static uint32_t __init kvm_detect(void)
699{
700 return kvm_cpuid_base();
701}
702
703static void __init kvm_apic_init(void)
704{
705#if defined(CONFIG_SMP)
706 if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
707 kvm_setup_pv_ipi();
708#endif
709}
710
711static void __init kvm_init_platform(void)
712{
713 kvmclock_init();
714 x86_platform.apic_post_init = kvm_apic_init;
715}
716
717const __initconst struct hypervisor_x86 x86_hyper_kvm = {
718 .name = "KVM",
719 .detect = kvm_detect,
720 .type = X86_HYPER_KVM,
721 .init.guest_late_init = kvm_guest_init,
722 .init.x2apic_available = kvm_para_available,
723 .init.init_platform = kvm_init_platform,
724};
725
726static __init int activate_jump_labels(void)
727{
728 if (has_steal_clock) {
729 static_key_slow_inc(¶virt_steal_enabled);
730 if (steal_acc)
731 static_key_slow_inc(¶virt_steal_rq_enabled);
732 }
733
734 return 0;
735}
736arch_initcall(activate_jump_labels);
737
738static __init int kvm_setup_pv_tlb_flush(void)
739{
740 int cpu;
741
742 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
743 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
744 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
745 for_each_possible_cpu(cpu) {
746 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
747 GFP_KERNEL, cpu_to_node(cpu));
748 }
749 pr_info("KVM setup pv remote TLB flush\n");
750 }
751
752 return 0;
753}
754arch_initcall(kvm_setup_pv_tlb_flush);
755
756#ifdef CONFIG_PARAVIRT_SPINLOCKS
757
758/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
759static void kvm_kick_cpu(int cpu)
760{
761 int apicid;
762 unsigned long flags = 0;
763
764 apicid = per_cpu(x86_cpu_to_apicid, cpu);
765 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
766}
767
768#include <asm/qspinlock.h>
769
770static void kvm_wait(u8 *ptr, u8 val)
771{
772 unsigned long flags;
773
774 if (in_nmi())
775 return;
776
777 local_irq_save(flags);
778
779 if (READ_ONCE(*ptr) != val)
780 goto out;
781
782 /*
783 * halt until it's our turn and kicked. Note that we do safe halt
784 * for irq enabled case to avoid hang when lock info is overwritten
785 * in irq spinlock slowpath and no spurious interrupt occur to save us.
786 */
787 if (arch_irqs_disabled_flags(flags))
788 halt();
789 else
790 safe_halt();
791
792out:
793 local_irq_restore(flags);
794}
795
796#ifdef CONFIG_X86_32
797__visible bool __kvm_vcpu_is_preempted(long cpu)
798{
799 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
800
801 return !!(src->preempted & KVM_VCPU_PREEMPTED);
802}
803PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
804
805#else
806
807#include <asm/asm-offsets.h>
808
809extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
810
811/*
812 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
813 * restoring to/from the stack.
814 */
815asm(
816".pushsection .text;"
817".global __raw_callee_save___kvm_vcpu_is_preempted;"
818".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
819"__raw_callee_save___kvm_vcpu_is_preempted:"
820"movq __per_cpu_offset(,%rdi,8), %rax;"
821"cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
822"setne %al;"
823"ret;"
824".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
825".popsection");
826
827#endif
828
829/*
830 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
831 */
832void __init kvm_spinlock_init(void)
833{
834 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
835 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
836 return;
837
838 if (kvm_para_has_hint(KVM_HINTS_REALTIME))
839 return;
840
841 /* Don't use the pvqspinlock code if there is only 1 vCPU. */
842 if (num_possible_cpus() == 1)
843 return;
844
845 __pv_init_lock_hash();
846 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
847 pv_ops.lock.queued_spin_unlock =
848 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
849 pv_ops.lock.wait = kvm_wait;
850 pv_ops.lock.kick = kvm_kick_cpu;
851
852 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
853 pv_ops.lock.vcpu_is_preempted =
854 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
855 }
856}
857
858#endif /* CONFIG_PARAVIRT_SPINLOCKS */
859
860#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
861
862static void kvm_disable_host_haltpoll(void *i)
863{
864 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
865}
866
867static void kvm_enable_host_haltpoll(void *i)
868{
869 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
870}
871
872void arch_haltpoll_enable(unsigned int cpu)
873{
874 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
875 pr_err_once("kvm: host does not support poll control\n");
876 pr_err_once("kvm: host upgrade recommended\n");
877 return;
878 }
879
880 /* Enable guest halt poll disables host halt poll */
881 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
882}
883EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
884
885void arch_haltpoll_disable(unsigned int cpu)
886{
887 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
888 return;
889
890 /* Enable guest halt poll disables host halt poll */
891 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
892}
893EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
894#endif
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/module.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 <asm/timer.h>
39#include <asm/cpu.h>
40#include <asm/traps.h>
41#include <asm/desc.h>
42#include <asm/tlbflush.h>
43#include <asm/idle.h>
44#include <asm/apic.h>
45#include <asm/apicdef.h>
46#include <asm/hypervisor.h>
47#include <asm/kvm_guest.h>
48
49static int kvmapf = 1;
50
51static int parse_no_kvmapf(char *arg)
52{
53 kvmapf = 0;
54 return 0;
55}
56
57early_param("no-kvmapf", parse_no_kvmapf);
58
59static int steal_acc = 1;
60static int parse_no_stealacc(char *arg)
61{
62 steal_acc = 0;
63 return 0;
64}
65
66early_param("no-steal-acc", parse_no_stealacc);
67
68static int kvmclock_vsyscall = 1;
69static int parse_no_kvmclock_vsyscall(char *arg)
70{
71 kvmclock_vsyscall = 0;
72 return 0;
73}
74
75early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
76
77static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
78static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
79static int has_steal_clock = 0;
80
81/*
82 * No need for any "IO delay" on KVM
83 */
84static void kvm_io_delay(void)
85{
86}
87
88#define KVM_TASK_SLEEP_HASHBITS 8
89#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
90
91struct kvm_task_sleep_node {
92 struct hlist_node link;
93 wait_queue_head_t wq;
94 u32 token;
95 int cpu;
96 bool halted;
97};
98
99static struct kvm_task_sleep_head {
100 spinlock_t lock;
101 struct hlist_head list;
102} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
103
104static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
105 u32 token)
106{
107 struct hlist_node *p;
108
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)
113 return n;
114 }
115
116 return NULL;
117}
118
119void kvm_async_pf_task_wait(u32 token)
120{
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;
124 DEFINE_WAIT(wait);
125
126 rcu_irq_enter();
127
128 spin_lock(&b->lock);
129 e = _find_apf_task(b, token);
130 if (e) {
131 /* dummy entry exist -> wake up was delivered ahead of PF */
132 hlist_del(&e->link);
133 kfree(e);
134 spin_unlock(&b->lock);
135
136 rcu_irq_exit();
137 return;
138 }
139
140 n.token = token;
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);
146
147 for (;;) {
148 if (!n.halted)
149 prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
150 if (hlist_unhashed(&n.link))
151 break;
152
153 if (!n.halted) {
154 local_irq_enable();
155 schedule();
156 local_irq_disable();
157 } else {
158 /*
159 * We cannot reschedule. So halt.
160 */
161 rcu_irq_exit();
162 native_safe_halt();
163 rcu_irq_enter();
164 local_irq_disable();
165 }
166 }
167 if (!n.halted)
168 finish_wait(&n.wq, &wait);
169
170 rcu_irq_exit();
171 return;
172}
173EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
174
175static void apf_task_wake_one(struct kvm_task_sleep_node *n)
176{
177 hlist_del_init(&n->link);
178 if (n->halted)
179 smp_send_reschedule(n->cpu);
180 else if (waitqueue_active(&n->wq))
181 wake_up(&n->wq);
182}
183
184static void apf_task_wake_all(void)
185{
186 int i;
187
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];
191 spin_lock(&b->lock);
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);
197 }
198 spin_unlock(&b->lock);
199 }
200}
201
202void kvm_async_pf_task_wake(u32 token)
203{
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;
207
208 if (token == ~0) {
209 apf_task_wake_all();
210 return;
211 }
212
213again:
214 spin_lock(&b->lock);
215 n = _find_apf_task(b, token);
216 if (!n) {
217 /*
218 * async PF was not yet handled.
219 * Add dummy entry for the token.
220 */
221 n = kzalloc(sizeof(*n), GFP_ATOMIC);
222 if (!n) {
223 /*
224 * Allocation failed! Busy wait while other cpu
225 * handles async PF.
226 */
227 spin_unlock(&b->lock);
228 cpu_relax();
229 goto again;
230 }
231 n->token = token;
232 n->cpu = smp_processor_id();
233 init_waitqueue_head(&n->wq);
234 hlist_add_head(&n->link, &b->list);
235 } else
236 apf_task_wake_one(n);
237 spin_unlock(&b->lock);
238 return;
239}
240EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
241
242u32 kvm_read_and_reset_pf_reason(void)
243{
244 u32 reason = 0;
245
246 if (__get_cpu_var(apf_reason).enabled) {
247 reason = __get_cpu_var(apf_reason).reason;
248 __get_cpu_var(apf_reason).reason = 0;
249 }
250
251 return reason;
252}
253EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
254
255dotraplinkage void __kprobes
256do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
257{
258 enum ctx_state prev_state;
259
260 switch (kvm_read_and_reset_pf_reason()) {
261 default:
262 do_page_fault(regs, error_code);
263 break;
264 case KVM_PV_REASON_PAGE_NOT_PRESENT:
265 /* page is swapped out by the host. */
266 prev_state = exception_enter();
267 exit_idle();
268 kvm_async_pf_task_wait((u32)read_cr2());
269 exception_exit(prev_state);
270 break;
271 case KVM_PV_REASON_PAGE_READY:
272 rcu_irq_enter();
273 exit_idle();
274 kvm_async_pf_task_wake((u32)read_cr2());
275 rcu_irq_exit();
276 break;
277 }
278}
279
280static void __init paravirt_ops_setup(void)
281{
282 pv_info.name = "KVM";
283 pv_info.paravirt_enabled = 1;
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 memset(st, 0, sizeof(*st));
302
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));
306}
307
308static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
309
310static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
311{
312 /**
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.
318 */
319 if (__test_and_clear_bit(KVM_PV_EOI_BIT, &__get_cpu_var(kvm_apic_eoi)))
320 return;
321 apic_write(APIC_EOI, APIC_EOI_ACK);
322}
323
324void kvm_guest_cpu_init(void)
325{
326 if (!kvm_para_available())
327 return;
328
329 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
330 u64 pa = slow_virt_to_phys(&__get_cpu_var(apf_reason));
331
332#ifdef CONFIG_PREEMPT
333 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
334#endif
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",
338 smp_processor_id());
339 }
340
341 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
342 unsigned long pa;
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))
347 | KVM_MSR_ENABLED;
348 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
349 }
350
351 if (has_steal_clock)
352 kvm_register_steal_time();
353}
354
355static void kvm_pv_disable_apf(void)
356{
357 if (!__get_cpu_var(apf_reason).enabled)
358 return;
359
360 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
361 __get_cpu_var(apf_reason).enabled = 0;
362
363 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
364 smp_processor_id());
365}
366
367static void kvm_pv_guest_cpu_reboot(void *unused)
368{
369 /*
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.
373 */
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();
378}
379
380static int kvm_pv_reboot_notify(struct notifier_block *nb,
381 unsigned long code, void *unused)
382{
383 if (code == SYS_RESTART)
384 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
385 return NOTIFY_DONE;
386}
387
388static struct notifier_block kvm_pv_reboot_nb = {
389 .notifier_call = kvm_pv_reboot_notify,
390};
391
392static u64 kvm_steal_clock(int cpu)
393{
394 u64 steal;
395 struct kvm_steal_time *src;
396 int version;
397
398 src = &per_cpu(steal_time, cpu);
399 do {
400 version = src->version;
401 rmb();
402 steal = src->steal;
403 rmb();
404 } while ((version & 1) || (version != src->version));
405
406 return steal;
407}
408
409void kvm_disable_steal_time(void)
410{
411 if (!has_steal_clock)
412 return;
413
414 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
415}
416
417#ifdef CONFIG_SMP
418static void __init kvm_smp_prepare_boot_cpu(void)
419{
420 kvm_guest_cpu_init();
421 native_smp_prepare_boot_cpu();
422 kvm_spinlock_init();
423}
424
425static void kvm_guest_cpu_online(void *dummy)
426{
427 kvm_guest_cpu_init();
428}
429
430static void kvm_guest_cpu_offline(void *dummy)
431{
432 kvm_disable_steal_time();
433 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
434 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
435 kvm_pv_disable_apf();
436 apf_task_wake_all();
437}
438
439static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
440 void *hcpu)
441{
442 int cpu = (unsigned long)hcpu;
443 switch (action) {
444 case CPU_ONLINE:
445 case CPU_DOWN_FAILED:
446 case CPU_ONLINE_FROZEN:
447 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
448 break;
449 case CPU_DOWN_PREPARE:
450 case CPU_DOWN_PREPARE_FROZEN:
451 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
452 break;
453 default:
454 break;
455 }
456 return NOTIFY_OK;
457}
458
459static struct notifier_block kvm_cpu_notifier = {
460 .notifier_call = kvm_cpu_notify,
461};
462#endif
463
464static void __init kvm_apf_trap_init(void)
465{
466 set_intr_gate(14, async_page_fault);
467}
468
469void __init kvm_guest_init(void)
470{
471 int i;
472
473 if (!kvm_para_available())
474 return;
475
476 paravirt_ops_setup();
477 register_reboot_notifier(&kvm_pv_reboot_nb);
478 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
479 spin_lock_init(&async_pf_sleepers[i].lock);
480 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
481 x86_init.irqs.trap_init = kvm_apf_trap_init;
482
483 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
484 has_steal_clock = 1;
485 pv_time_ops.steal_clock = kvm_steal_clock;
486 }
487
488 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
489 apic_set_eoi_write(kvm_guest_apic_eoi_write);
490
491 if (kvmclock_vsyscall)
492 kvm_setup_vsyscall_timeinfo();
493
494#ifdef CONFIG_SMP
495 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
496 register_cpu_notifier(&kvm_cpu_notifier);
497#else
498 kvm_guest_cpu_init();
499#endif
500}
501
502static noinline uint32_t __kvm_cpuid_base(void)
503{
504 if (boot_cpu_data.cpuid_level < 0)
505 return 0; /* So we don't blow up on old processors */
506
507 if (cpu_has_hypervisor)
508 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
509
510 return 0;
511}
512
513static inline uint32_t kvm_cpuid_base(void)
514{
515 static int kvm_cpuid_base = -1;
516
517 if (kvm_cpuid_base == -1)
518 kvm_cpuid_base = __kvm_cpuid_base();
519
520 return kvm_cpuid_base;
521}
522
523bool kvm_para_available(void)
524{
525 return kvm_cpuid_base() != 0;
526}
527EXPORT_SYMBOL_GPL(kvm_para_available);
528
529unsigned int kvm_arch_para_features(void)
530{
531 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
532}
533
534static uint32_t __init kvm_detect(void)
535{
536 return kvm_cpuid_base();
537}
538
539const struct hypervisor_x86 x86_hyper_kvm __refconst = {
540 .name = "KVM",
541 .detect = kvm_detect,
542 .x2apic_available = kvm_para_available,
543};
544EXPORT_SYMBOL_GPL(x86_hyper_kvm);
545
546static __init int activate_jump_labels(void)
547{
548 if (has_steal_clock) {
549 static_key_slow_inc(¶virt_steal_enabled);
550 if (steal_acc)
551 static_key_slow_inc(¶virt_steal_rq_enabled);
552 }
553
554 return 0;
555}
556arch_initcall(activate_jump_labels);
557
558#ifdef CONFIG_PARAVIRT_SPINLOCKS
559
560/* Kick a cpu by its apicid. Used to wake up a halted vcpu */
561static void kvm_kick_cpu(int cpu)
562{
563 int apicid;
564 unsigned long flags = 0;
565
566 apicid = per_cpu(x86_cpu_to_apicid, cpu);
567 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
568}
569
570enum kvm_contention_stat {
571 TAKEN_SLOW,
572 TAKEN_SLOW_PICKUP,
573 RELEASED_SLOW,
574 RELEASED_SLOW_KICKED,
575 NR_CONTENTION_STATS
576};
577
578#ifdef CONFIG_KVM_DEBUG_FS
579#define HISTO_BUCKETS 30
580
581static struct kvm_spinlock_stats
582{
583 u32 contention_stats[NR_CONTENTION_STATS];
584 u32 histo_spin_blocked[HISTO_BUCKETS+1];
585 u64 time_blocked;
586} spinlock_stats;
587
588static u8 zero_stats;
589
590static inline void check_zero(void)
591{
592 u8 ret;
593 u8 old;
594
595 old = ACCESS_ONCE(zero_stats);
596 if (unlikely(old)) {
597 ret = cmpxchg(&zero_stats, old, 0);
598 /* This ensures only one fellow resets the stat */
599 if (ret == old)
600 memset(&spinlock_stats, 0, sizeof(spinlock_stats));
601 }
602}
603
604static inline void add_stats(enum kvm_contention_stat var, u32 val)
605{
606 check_zero();
607 spinlock_stats.contention_stats[var] += val;
608}
609
610
611static inline u64 spin_time_start(void)
612{
613 return sched_clock();
614}
615
616static void __spin_time_accum(u64 delta, u32 *array)
617{
618 unsigned index;
619
620 index = ilog2(delta);
621 check_zero();
622
623 if (index < HISTO_BUCKETS)
624 array[index]++;
625 else
626 array[HISTO_BUCKETS]++;
627}
628
629static inline void spin_time_accum_blocked(u64 start)
630{
631 u32 delta;
632
633 delta = sched_clock() - start;
634 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
635 spinlock_stats.time_blocked += delta;
636}
637
638static struct dentry *d_spin_debug;
639static struct dentry *d_kvm_debug;
640
641struct dentry *kvm_init_debugfs(void)
642{
643 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
644 if (!d_kvm_debug)
645 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
646
647 return d_kvm_debug;
648}
649
650static int __init kvm_spinlock_debugfs(void)
651{
652 struct dentry *d_kvm;
653
654 d_kvm = kvm_init_debugfs();
655 if (d_kvm == NULL)
656 return -ENOMEM;
657
658 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
659
660 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
661
662 debugfs_create_u32("taken_slow", 0444, d_spin_debug,
663 &spinlock_stats.contention_stats[TAKEN_SLOW]);
664 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
665 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
666
667 debugfs_create_u32("released_slow", 0444, d_spin_debug,
668 &spinlock_stats.contention_stats[RELEASED_SLOW]);
669 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
670 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
671
672 debugfs_create_u64("time_blocked", 0444, d_spin_debug,
673 &spinlock_stats.time_blocked);
674
675 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
676 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
677
678 return 0;
679}
680fs_initcall(kvm_spinlock_debugfs);
681#else /* !CONFIG_KVM_DEBUG_FS */
682static inline void add_stats(enum kvm_contention_stat var, u32 val)
683{
684}
685
686static inline u64 spin_time_start(void)
687{
688 return 0;
689}
690
691static inline void spin_time_accum_blocked(u64 start)
692{
693}
694#endif /* CONFIG_KVM_DEBUG_FS */
695
696struct kvm_lock_waiting {
697 struct arch_spinlock *lock;
698 __ticket_t want;
699};
700
701/* cpus 'waiting' on a spinlock to become available */
702static cpumask_t waiting_cpus;
703
704/* Track spinlock on which a cpu is waiting */
705static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
706
707__visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
708{
709 struct kvm_lock_waiting *w;
710 int cpu;
711 u64 start;
712 unsigned long flags;
713
714 if (in_nmi())
715 return;
716
717 w = &__get_cpu_var(klock_waiting);
718 cpu = smp_processor_id();
719 start = spin_time_start();
720
721 /*
722 * Make sure an interrupt handler can't upset things in a
723 * partially setup state.
724 */
725 local_irq_save(flags);
726
727 /*
728 * The ordering protocol on this is that the "lock" pointer
729 * may only be set non-NULL if the "want" ticket is correct.
730 * If we're updating "want", we must first clear "lock".
731 */
732 w->lock = NULL;
733 smp_wmb();
734 w->want = want;
735 smp_wmb();
736 w->lock = lock;
737
738 add_stats(TAKEN_SLOW, 1);
739
740 /*
741 * This uses set_bit, which is atomic but we should not rely on its
742 * reordering gurantees. So barrier is needed after this call.
743 */
744 cpumask_set_cpu(cpu, &waiting_cpus);
745
746 barrier();
747
748 /*
749 * Mark entry to slowpath before doing the pickup test to make
750 * sure we don't deadlock with an unlocker.
751 */
752 __ticket_enter_slowpath(lock);
753
754 /*
755 * check again make sure it didn't become free while
756 * we weren't looking.
757 */
758 if (ACCESS_ONCE(lock->tickets.head) == want) {
759 add_stats(TAKEN_SLOW_PICKUP, 1);
760 goto out;
761 }
762
763 /*
764 * halt until it's our turn and kicked. Note that we do safe halt
765 * for irq enabled case to avoid hang when lock info is overwritten
766 * in irq spinlock slowpath and no spurious interrupt occur to save us.
767 */
768 if (arch_irqs_disabled_flags(flags))
769 halt();
770 else
771 safe_halt();
772
773out:
774 cpumask_clear_cpu(cpu, &waiting_cpus);
775 w->lock = NULL;
776 local_irq_restore(flags);
777 spin_time_accum_blocked(start);
778}
779PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
780
781/* Kick vcpu waiting on @lock->head to reach value @ticket */
782static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
783{
784 int cpu;
785
786 add_stats(RELEASED_SLOW, 1);
787 for_each_cpu(cpu, &waiting_cpus) {
788 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
789 if (ACCESS_ONCE(w->lock) == lock &&
790 ACCESS_ONCE(w->want) == ticket) {
791 add_stats(RELEASED_SLOW_KICKED, 1);
792 kvm_kick_cpu(cpu);
793 break;
794 }
795 }
796}
797
798/*
799 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
800 */
801void __init kvm_spinlock_init(void)
802{
803 if (!kvm_para_available())
804 return;
805 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
806 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
807 return;
808
809 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
810 pv_lock_ops.unlock_kick = kvm_unlock_kick;
811}
812
813static __init int kvm_spinlock_init_jump(void)
814{
815 if (!kvm_para_available())
816 return 0;
817 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
818 return 0;
819
820 static_key_slow_inc(¶virt_ticketlocks_enabled);
821 printk(KERN_INFO "KVM setup paravirtual spinlock\n");
822
823 return 0;
824}
825early_initcall(kvm_spinlock_init_jump);
826
827#endif /* CONFIG_PARAVIRT_SPINLOCKS */