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/module.h>
24#include <linux/kernel.h>
25#include <linux/kvm_para.h>
26#include <linux/cpu.h>
27#include <linux/mm.h>
28#include <linux/highmem.h>
29#include <linux/hardirq.h>
30#include <linux/notifier.h>
31#include <linux/reboot.h>
32#include <linux/hash.h>
33#include <linux/sched.h>
34#include <linux/slab.h>
35#include <linux/kprobes.h>
36#include <asm/timer.h>
37#include <asm/cpu.h>
38#include <asm/traps.h>
39#include <asm/desc.h>
40#include <asm/tlbflush.h>
41
42#define MMU_QUEUE_SIZE 1024
43
44static int kvmapf = 1;
45
46static int parse_no_kvmapf(char *arg)
47{
48 kvmapf = 0;
49 return 0;
50}
51
52early_param("no-kvmapf", parse_no_kvmapf);
53
54static int steal_acc = 1;
55static int parse_no_stealacc(char *arg)
56{
57 steal_acc = 0;
58 return 0;
59}
60
61early_param("no-steal-acc", parse_no_stealacc);
62
63struct kvm_para_state {
64 u8 mmu_queue[MMU_QUEUE_SIZE];
65 int mmu_queue_len;
66};
67
68static DEFINE_PER_CPU(struct kvm_para_state, para_state);
69static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
70static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
71static int has_steal_clock = 0;
72
73static struct kvm_para_state *kvm_para_state(void)
74{
75 return &per_cpu(para_state, raw_smp_processor_id());
76}
77
78/*
79 * No need for any "IO delay" on KVM
80 */
81static void kvm_io_delay(void)
82{
83}
84
85#define KVM_TASK_SLEEP_HASHBITS 8
86#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
87
88struct kvm_task_sleep_node {
89 struct hlist_node link;
90 wait_queue_head_t wq;
91 u32 token;
92 int cpu;
93 bool halted;
94 struct mm_struct *mm;
95};
96
97static struct kvm_task_sleep_head {
98 spinlock_t lock;
99 struct hlist_head list;
100} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
101
102static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
103 u32 token)
104{
105 struct hlist_node *p;
106
107 hlist_for_each(p, &b->list) {
108 struct kvm_task_sleep_node *n =
109 hlist_entry(p, typeof(*n), link);
110 if (n->token == token)
111 return n;
112 }
113
114 return NULL;
115}
116
117void kvm_async_pf_task_wait(u32 token)
118{
119 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
120 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
121 struct kvm_task_sleep_node n, *e;
122 DEFINE_WAIT(wait);
123 int cpu, idle;
124
125 cpu = get_cpu();
126 idle = idle_cpu(cpu);
127 put_cpu();
128
129 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 spin_unlock(&b->lock);
136 return;
137 }
138
139 n.token = token;
140 n.cpu = smp_processor_id();
141 n.mm = current->active_mm;
142 n.halted = idle || preempt_count() > 1;
143 atomic_inc(&n.mm->mm_count);
144 init_waitqueue_head(&n.wq);
145 hlist_add_head(&n.link, &b->list);
146 spin_unlock(&b->lock);
147
148 for (;;) {
149 if (!n.halted)
150 prepare_to_wait(&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 native_safe_halt();
163 local_irq_disable();
164 }
165 }
166 if (!n.halted)
167 finish_wait(&n.wq, &wait);
168
169 return;
170}
171EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
172
173static void apf_task_wake_one(struct kvm_task_sleep_node *n)
174{
175 hlist_del_init(&n->link);
176 if (!n->mm)
177 return;
178 mmdrop(n->mm);
179 if (n->halted)
180 smp_send_reschedule(n->cpu);
181 else if (waitqueue_active(&n->wq))
182 wake_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 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 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 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 = kmalloc(sizeof(*n), GFP_ATOMIC);
223 if (!n) {
224 /*
225 * Allocation failed! Busy wait while other cpu
226 * handles async PF.
227 */
228 spin_unlock(&b->lock);
229 cpu_relax();
230 goto again;
231 }
232 n->token = token;
233 n->cpu = smp_processor_id();
234 n->mm = NULL;
235 init_waitqueue_head(&n->wq);
236 hlist_add_head(&n->link, &b->list);
237 } else
238 apf_task_wake_one(n);
239 spin_unlock(&b->lock);
240 return;
241}
242EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243
244u32 kvm_read_and_reset_pf_reason(void)
245{
246 u32 reason = 0;
247
248 if (__get_cpu_var(apf_reason).enabled) {
249 reason = __get_cpu_var(apf_reason).reason;
250 __get_cpu_var(apf_reason).reason = 0;
251 }
252
253 return reason;
254}
255EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
256
257dotraplinkage void __kprobes
258do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
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 kvm_async_pf_task_wait((u32)read_cr2());
267 break;
268 case KVM_PV_REASON_PAGE_READY:
269 kvm_async_pf_task_wake((u32)read_cr2());
270 break;
271 }
272}
273
274static void kvm_mmu_op(void *buffer, unsigned len)
275{
276 int r;
277 unsigned long a1, a2;
278
279 do {
280 a1 = __pa(buffer);
281 a2 = 0; /* on i386 __pa() always returns <4G */
282 r = kvm_hypercall3(KVM_HC_MMU_OP, len, a1, a2);
283 buffer += r;
284 len -= r;
285 } while (len);
286}
287
288static void mmu_queue_flush(struct kvm_para_state *state)
289{
290 if (state->mmu_queue_len) {
291 kvm_mmu_op(state->mmu_queue, state->mmu_queue_len);
292 state->mmu_queue_len = 0;
293 }
294}
295
296static void kvm_deferred_mmu_op(void *buffer, int len)
297{
298 struct kvm_para_state *state = kvm_para_state();
299
300 if (paravirt_get_lazy_mode() != PARAVIRT_LAZY_MMU) {
301 kvm_mmu_op(buffer, len);
302 return;
303 }
304 if (state->mmu_queue_len + len > sizeof state->mmu_queue)
305 mmu_queue_flush(state);
306 memcpy(state->mmu_queue + state->mmu_queue_len, buffer, len);
307 state->mmu_queue_len += len;
308}
309
310static void kvm_mmu_write(void *dest, u64 val)
311{
312 __u64 pte_phys;
313 struct kvm_mmu_op_write_pte wpte;
314
315#ifdef CONFIG_HIGHPTE
316 struct page *page;
317 unsigned long dst = (unsigned long) dest;
318
319 page = kmap_atomic_to_page(dest);
320 pte_phys = page_to_pfn(page);
321 pte_phys <<= PAGE_SHIFT;
322 pte_phys += (dst & ~(PAGE_MASK));
323#else
324 pte_phys = (unsigned long)__pa(dest);
325#endif
326 wpte.header.op = KVM_MMU_OP_WRITE_PTE;
327 wpte.pte_val = val;
328 wpte.pte_phys = pte_phys;
329
330 kvm_deferred_mmu_op(&wpte, sizeof wpte);
331}
332
333/*
334 * We only need to hook operations that are MMU writes. We hook these so that
335 * we can use lazy MMU mode to batch these operations. We could probably
336 * improve the performance of the host code if we used some of the information
337 * here to simplify processing of batched writes.
338 */
339static void kvm_set_pte(pte_t *ptep, pte_t pte)
340{
341 kvm_mmu_write(ptep, pte_val(pte));
342}
343
344static void kvm_set_pte_at(struct mm_struct *mm, unsigned long addr,
345 pte_t *ptep, pte_t pte)
346{
347 kvm_mmu_write(ptep, pte_val(pte));
348}
349
350static void kvm_set_pmd(pmd_t *pmdp, pmd_t pmd)
351{
352 kvm_mmu_write(pmdp, pmd_val(pmd));
353}
354
355#if PAGETABLE_LEVELS >= 3
356#ifdef CONFIG_X86_PAE
357static void kvm_set_pte_atomic(pte_t *ptep, pte_t pte)
358{
359 kvm_mmu_write(ptep, pte_val(pte));
360}
361
362static void kvm_pte_clear(struct mm_struct *mm,
363 unsigned long addr, pte_t *ptep)
364{
365 kvm_mmu_write(ptep, 0);
366}
367
368static void kvm_pmd_clear(pmd_t *pmdp)
369{
370 kvm_mmu_write(pmdp, 0);
371}
372#endif
373
374static void kvm_set_pud(pud_t *pudp, pud_t pud)
375{
376 kvm_mmu_write(pudp, pud_val(pud));
377}
378
379#if PAGETABLE_LEVELS == 4
380static void kvm_set_pgd(pgd_t *pgdp, pgd_t pgd)
381{
382 kvm_mmu_write(pgdp, pgd_val(pgd));
383}
384#endif
385#endif /* PAGETABLE_LEVELS >= 3 */
386
387static void kvm_flush_tlb(void)
388{
389 struct kvm_mmu_op_flush_tlb ftlb = {
390 .header.op = KVM_MMU_OP_FLUSH_TLB,
391 };
392
393 kvm_deferred_mmu_op(&ftlb, sizeof ftlb);
394}
395
396static void kvm_release_pt(unsigned long pfn)
397{
398 struct kvm_mmu_op_release_pt rpt = {
399 .header.op = KVM_MMU_OP_RELEASE_PT,
400 .pt_phys = (u64)pfn << PAGE_SHIFT,
401 };
402
403 kvm_mmu_op(&rpt, sizeof rpt);
404}
405
406static void kvm_enter_lazy_mmu(void)
407{
408 paravirt_enter_lazy_mmu();
409}
410
411static void kvm_leave_lazy_mmu(void)
412{
413 struct kvm_para_state *state = kvm_para_state();
414
415 mmu_queue_flush(state);
416 paravirt_leave_lazy_mmu();
417}
418
419static void __init paravirt_ops_setup(void)
420{
421 pv_info.name = "KVM";
422 pv_info.paravirt_enabled = 1;
423
424 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
425 pv_cpu_ops.io_delay = kvm_io_delay;
426
427 if (kvm_para_has_feature(KVM_FEATURE_MMU_OP)) {
428 pv_mmu_ops.set_pte = kvm_set_pte;
429 pv_mmu_ops.set_pte_at = kvm_set_pte_at;
430 pv_mmu_ops.set_pmd = kvm_set_pmd;
431#if PAGETABLE_LEVELS >= 3
432#ifdef CONFIG_X86_PAE
433 pv_mmu_ops.set_pte_atomic = kvm_set_pte_atomic;
434 pv_mmu_ops.pte_clear = kvm_pte_clear;
435 pv_mmu_ops.pmd_clear = kvm_pmd_clear;
436#endif
437 pv_mmu_ops.set_pud = kvm_set_pud;
438#if PAGETABLE_LEVELS == 4
439 pv_mmu_ops.set_pgd = kvm_set_pgd;
440#endif
441#endif
442 pv_mmu_ops.flush_tlb_user = kvm_flush_tlb;
443 pv_mmu_ops.release_pte = kvm_release_pt;
444 pv_mmu_ops.release_pmd = kvm_release_pt;
445 pv_mmu_ops.release_pud = kvm_release_pt;
446
447 pv_mmu_ops.lazy_mode.enter = kvm_enter_lazy_mmu;
448 pv_mmu_ops.lazy_mode.leave = kvm_leave_lazy_mmu;
449 }
450#ifdef CONFIG_X86_IO_APIC
451 no_timer_check = 1;
452#endif
453}
454
455static void kvm_register_steal_time(void)
456{
457 int cpu = smp_processor_id();
458 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
459
460 if (!has_steal_clock)
461 return;
462
463 memset(st, 0, sizeof(*st));
464
465 wrmsrl(MSR_KVM_STEAL_TIME, (__pa(st) | KVM_MSR_ENABLED));
466 printk(KERN_INFO "kvm-stealtime: cpu %d, msr %lx\n",
467 cpu, __pa(st));
468}
469
470void __cpuinit kvm_guest_cpu_init(void)
471{
472 if (!kvm_para_available())
473 return;
474
475 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
476 u64 pa = __pa(&__get_cpu_var(apf_reason));
477
478#ifdef CONFIG_PREEMPT
479 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
480#endif
481 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
482 __get_cpu_var(apf_reason).enabled = 1;
483 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
484 smp_processor_id());
485 }
486
487 if (has_steal_clock)
488 kvm_register_steal_time();
489}
490
491static void kvm_pv_disable_apf(void *unused)
492{
493 if (!__get_cpu_var(apf_reason).enabled)
494 return;
495
496 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
497 __get_cpu_var(apf_reason).enabled = 0;
498
499 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
500 smp_processor_id());
501}
502
503static int kvm_pv_reboot_notify(struct notifier_block *nb,
504 unsigned long code, void *unused)
505{
506 if (code == SYS_RESTART)
507 on_each_cpu(kvm_pv_disable_apf, NULL, 1);
508 return NOTIFY_DONE;
509}
510
511static struct notifier_block kvm_pv_reboot_nb = {
512 .notifier_call = kvm_pv_reboot_notify,
513};
514
515static u64 kvm_steal_clock(int cpu)
516{
517 u64 steal;
518 struct kvm_steal_time *src;
519 int version;
520
521 src = &per_cpu(steal_time, cpu);
522 do {
523 version = src->version;
524 rmb();
525 steal = src->steal;
526 rmb();
527 } while ((version & 1) || (version != src->version));
528
529 return steal;
530}
531
532void kvm_disable_steal_time(void)
533{
534 if (!has_steal_clock)
535 return;
536
537 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
538}
539
540#ifdef CONFIG_SMP
541static void __init kvm_smp_prepare_boot_cpu(void)
542{
543#ifdef CONFIG_KVM_CLOCK
544 WARN_ON(kvm_register_clock("primary cpu clock"));
545#endif
546 kvm_guest_cpu_init();
547 native_smp_prepare_boot_cpu();
548}
549
550static void __cpuinit kvm_guest_cpu_online(void *dummy)
551{
552 kvm_guest_cpu_init();
553}
554
555static void kvm_guest_cpu_offline(void *dummy)
556{
557 kvm_disable_steal_time();
558 kvm_pv_disable_apf(NULL);
559 apf_task_wake_all();
560}
561
562static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
563 unsigned long action, void *hcpu)
564{
565 int cpu = (unsigned long)hcpu;
566 switch (action) {
567 case CPU_ONLINE:
568 case CPU_DOWN_FAILED:
569 case CPU_ONLINE_FROZEN:
570 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
571 break;
572 case CPU_DOWN_PREPARE:
573 case CPU_DOWN_PREPARE_FROZEN:
574 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
575 break;
576 default:
577 break;
578 }
579 return NOTIFY_OK;
580}
581
582static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
583 .notifier_call = kvm_cpu_notify,
584};
585#endif
586
587static void __init kvm_apf_trap_init(void)
588{
589 set_intr_gate(14, &async_page_fault);
590}
591
592void __init kvm_guest_init(void)
593{
594 int i;
595
596 if (!kvm_para_available())
597 return;
598
599 paravirt_ops_setup();
600 register_reboot_notifier(&kvm_pv_reboot_nb);
601 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
602 spin_lock_init(&async_pf_sleepers[i].lock);
603 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
604 x86_init.irqs.trap_init = kvm_apf_trap_init;
605
606 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
607 has_steal_clock = 1;
608 pv_time_ops.steal_clock = kvm_steal_clock;
609 }
610
611#ifdef CONFIG_SMP
612 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
613 register_cpu_notifier(&kvm_cpu_notifier);
614#else
615 kvm_guest_cpu_init();
616#endif
617}
618
619static __init int activate_jump_labels(void)
620{
621 if (has_steal_clock) {
622 jump_label_inc(¶virt_steal_enabled);
623 if (steal_acc)
624 jump_label_inc(¶virt_steal_rq_enabled);
625 }
626
627 return 0;
628}
629arch_initcall(activate_jump_labels);
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 */