]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - arch/x86/kernel/kvm.c
KVM guest: exit idleness when handling KVM_PV_REASON_PAGE_NOT_PRESENT
[~shefty/rdma-dev.git] / arch / x86 / kernel / kvm.c
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 #include <asm/idle.h>
42 #include <asm/apic.h>
43 #include <asm/apicdef.h>
44 #include <asm/hypervisor.h>
45
46 static int kvmapf = 1;
47
48 static int parse_no_kvmapf(char *arg)
49 {
50         kvmapf = 0;
51         return 0;
52 }
53
54 early_param("no-kvmapf", parse_no_kvmapf);
55
56 static int steal_acc = 1;
57 static int parse_no_stealacc(char *arg)
58 {
59         steal_acc = 0;
60         return 0;
61 }
62
63 early_param("no-steal-acc", parse_no_stealacc);
64
65 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
66 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
67 static int has_steal_clock = 0;
68
69 /*
70  * No need for any "IO delay" on KVM
71  */
72 static void kvm_io_delay(void)
73 {
74 }
75
76 #define KVM_TASK_SLEEP_HASHBITS 8
77 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
78
79 struct kvm_task_sleep_node {
80         struct hlist_node link;
81         wait_queue_head_t wq;
82         u32 token;
83         int cpu;
84         bool halted;
85 };
86
87 static struct kvm_task_sleep_head {
88         spinlock_t lock;
89         struct hlist_head list;
90 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
91
92 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
93                                                   u32 token)
94 {
95         struct hlist_node *p;
96
97         hlist_for_each(p, &b->list) {
98                 struct kvm_task_sleep_node *n =
99                         hlist_entry(p, typeof(*n), link);
100                 if (n->token == token)
101                         return n;
102         }
103
104         return NULL;
105 }
106
107 void kvm_async_pf_task_wait(u32 token)
108 {
109         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
110         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
111         struct kvm_task_sleep_node n, *e;
112         DEFINE_WAIT(wait);
113         int cpu, idle;
114
115         cpu = get_cpu();
116         idle = idle_cpu(cpu);
117         put_cpu();
118
119         spin_lock(&b->lock);
120         e = _find_apf_task(b, token);
121         if (e) {
122                 /* dummy entry exist -> wake up was delivered ahead of PF */
123                 hlist_del(&e->link);
124                 kfree(e);
125                 spin_unlock(&b->lock);
126                 return;
127         }
128
129         n.token = token;
130         n.cpu = smp_processor_id();
131         n.halted = idle || preempt_count() > 1;
132         init_waitqueue_head(&n.wq);
133         hlist_add_head(&n.link, &b->list);
134         spin_unlock(&b->lock);
135
136         for (;;) {
137                 if (!n.halted)
138                         prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
139                 if (hlist_unhashed(&n.link))
140                         break;
141
142                 if (!n.halted) {
143                         local_irq_enable();
144                         schedule();
145                         local_irq_disable();
146                 } else {
147                         /*
148                          * We cannot reschedule. So halt.
149                          */
150                         native_safe_halt();
151                         local_irq_disable();
152                 }
153         }
154         if (!n.halted)
155                 finish_wait(&n.wq, &wait);
156
157         return;
158 }
159 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
160
161 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
162 {
163         hlist_del_init(&n->link);
164         if (n->halted)
165                 smp_send_reschedule(n->cpu);
166         else if (waitqueue_active(&n->wq))
167                 wake_up(&n->wq);
168 }
169
170 static void apf_task_wake_all(void)
171 {
172         int i;
173
174         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
175                 struct hlist_node *p, *next;
176                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
177                 spin_lock(&b->lock);
178                 hlist_for_each_safe(p, next, &b->list) {
179                         struct kvm_task_sleep_node *n =
180                                 hlist_entry(p, typeof(*n), link);
181                         if (n->cpu == smp_processor_id())
182                                 apf_task_wake_one(n);
183                 }
184                 spin_unlock(&b->lock);
185         }
186 }
187
188 void kvm_async_pf_task_wake(u32 token)
189 {
190         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
191         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
192         struct kvm_task_sleep_node *n;
193
194         if (token == ~0) {
195                 apf_task_wake_all();
196                 return;
197         }
198
199 again:
200         spin_lock(&b->lock);
201         n = _find_apf_task(b, token);
202         if (!n) {
203                 /*
204                  * async PF was not yet handled.
205                  * Add dummy entry for the token.
206                  */
207                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
208                 if (!n) {
209                         /*
210                          * Allocation failed! Busy wait while other cpu
211                          * handles async PF.
212                          */
213                         spin_unlock(&b->lock);
214                         cpu_relax();
215                         goto again;
216                 }
217                 n->token = token;
218                 n->cpu = smp_processor_id();
219                 init_waitqueue_head(&n->wq);
220                 hlist_add_head(&n->link, &b->list);
221         } else
222                 apf_task_wake_one(n);
223         spin_unlock(&b->lock);
224         return;
225 }
226 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
227
228 u32 kvm_read_and_reset_pf_reason(void)
229 {
230         u32 reason = 0;
231
232         if (__get_cpu_var(apf_reason).enabled) {
233                 reason = __get_cpu_var(apf_reason).reason;
234                 __get_cpu_var(apf_reason).reason = 0;
235         }
236
237         return reason;
238 }
239 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
240
241 dotraplinkage void __kprobes
242 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
243 {
244         switch (kvm_read_and_reset_pf_reason()) {
245         default:
246                 do_page_fault(regs, error_code);
247                 break;
248         case KVM_PV_REASON_PAGE_NOT_PRESENT:
249                 /* page is swapped out by the host. */
250                 rcu_irq_enter();
251                 exit_idle();
252                 kvm_async_pf_task_wait((u32)read_cr2());
253                 rcu_irq_exit();
254                 break;
255         case KVM_PV_REASON_PAGE_READY:
256                 rcu_irq_enter();
257                 exit_idle();
258                 kvm_async_pf_task_wake((u32)read_cr2());
259                 rcu_irq_exit();
260                 break;
261         }
262 }
263
264 static void __init paravirt_ops_setup(void)
265 {
266         pv_info.name = "KVM";
267         pv_info.paravirt_enabled = 1;
268
269         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
270                 pv_cpu_ops.io_delay = kvm_io_delay;
271
272 #ifdef CONFIG_X86_IO_APIC
273         no_timer_check = 1;
274 #endif
275 }
276
277 static void kvm_register_steal_time(void)
278 {
279         int cpu = smp_processor_id();
280         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
281
282         if (!has_steal_clock)
283                 return;
284
285         memset(st, 0, sizeof(*st));
286
287         wrmsrl(MSR_KVM_STEAL_TIME, (__pa(st) | KVM_MSR_ENABLED));
288         printk(KERN_INFO "kvm-stealtime: cpu %d, msr %lx\n",
289                 cpu, __pa(st));
290 }
291
292 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
293
294 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
295 {
296         /**
297          * This relies on __test_and_clear_bit to modify the memory
298          * in a way that is atomic with respect to the local CPU.
299          * The hypervisor only accesses this memory from the local CPU so
300          * there's no need for lock or memory barriers.
301          * An optimization barrier is implied in apic write.
302          */
303         if (__test_and_clear_bit(KVM_PV_EOI_BIT, &__get_cpu_var(kvm_apic_eoi)))
304                 return;
305         apic_write(APIC_EOI, APIC_EOI_ACK);
306 }
307
308 void __cpuinit kvm_guest_cpu_init(void)
309 {
310         if (!kvm_para_available())
311                 return;
312
313         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
314                 u64 pa = __pa(&__get_cpu_var(apf_reason));
315
316 #ifdef CONFIG_PREEMPT
317                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
318 #endif
319                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
320                 __get_cpu_var(apf_reason).enabled = 1;
321                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
322                        smp_processor_id());
323         }
324
325         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
326                 unsigned long pa;
327                 /* Size alignment is implied but just to make it explicit. */
328                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
329                 __get_cpu_var(kvm_apic_eoi) = 0;
330                 pa = __pa(&__get_cpu_var(kvm_apic_eoi)) | KVM_MSR_ENABLED;
331                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
332         }
333
334         if (has_steal_clock)
335                 kvm_register_steal_time();
336 }
337
338 static void kvm_pv_disable_apf(void)
339 {
340         if (!__get_cpu_var(apf_reason).enabled)
341                 return;
342
343         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
344         __get_cpu_var(apf_reason).enabled = 0;
345
346         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
347                smp_processor_id());
348 }
349
350 static void kvm_pv_guest_cpu_reboot(void *unused)
351 {
352         /*
353          * We disable PV EOI before we load a new kernel by kexec,
354          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
355          * New kernel can re-enable when it boots.
356          */
357         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
358                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
359         kvm_pv_disable_apf();
360         kvm_disable_steal_time();
361 }
362
363 static int kvm_pv_reboot_notify(struct notifier_block *nb,
364                                 unsigned long code, void *unused)
365 {
366         if (code == SYS_RESTART)
367                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
368         return NOTIFY_DONE;
369 }
370
371 static struct notifier_block kvm_pv_reboot_nb = {
372         .notifier_call = kvm_pv_reboot_notify,
373 };
374
375 static u64 kvm_steal_clock(int cpu)
376 {
377         u64 steal;
378         struct kvm_steal_time *src;
379         int version;
380
381         src = &per_cpu(steal_time, cpu);
382         do {
383                 version = src->version;
384                 rmb();
385                 steal = src->steal;
386                 rmb();
387         } while ((version & 1) || (version != src->version));
388
389         return steal;
390 }
391
392 void kvm_disable_steal_time(void)
393 {
394         if (!has_steal_clock)
395                 return;
396
397         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
398 }
399
400 #ifdef CONFIG_SMP
401 static void __init kvm_smp_prepare_boot_cpu(void)
402 {
403         WARN_ON(kvm_register_clock("primary cpu clock"));
404         kvm_guest_cpu_init();
405         native_smp_prepare_boot_cpu();
406 }
407
408 static void __cpuinit kvm_guest_cpu_online(void *dummy)
409 {
410         kvm_guest_cpu_init();
411 }
412
413 static void kvm_guest_cpu_offline(void *dummy)
414 {
415         kvm_disable_steal_time();
416         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
417                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
418         kvm_pv_disable_apf();
419         apf_task_wake_all();
420 }
421
422 static int __cpuinit kvm_cpu_notify(struct notifier_block *self,
423                                     unsigned long action, void *hcpu)
424 {
425         int cpu = (unsigned long)hcpu;
426         switch (action) {
427         case CPU_ONLINE:
428         case CPU_DOWN_FAILED:
429         case CPU_ONLINE_FROZEN:
430                 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
431                 break;
432         case CPU_DOWN_PREPARE:
433         case CPU_DOWN_PREPARE_FROZEN:
434                 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
435                 break;
436         default:
437                 break;
438         }
439         return NOTIFY_OK;
440 }
441
442 static struct notifier_block __cpuinitdata kvm_cpu_notifier = {
443         .notifier_call  = kvm_cpu_notify,
444 };
445 #endif
446
447 static void __init kvm_apf_trap_init(void)
448 {
449         set_intr_gate(14, &async_page_fault);
450 }
451
452 void __init kvm_guest_init(void)
453 {
454         int i;
455
456         if (!kvm_para_available())
457                 return;
458
459         paravirt_ops_setup();
460         register_reboot_notifier(&kvm_pv_reboot_nb);
461         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
462                 spin_lock_init(&async_pf_sleepers[i].lock);
463         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
464                 x86_init.irqs.trap_init = kvm_apf_trap_init;
465
466         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
467                 has_steal_clock = 1;
468                 pv_time_ops.steal_clock = kvm_steal_clock;
469         }
470
471         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
472                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
473
474 #ifdef CONFIG_SMP
475         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
476         register_cpu_notifier(&kvm_cpu_notifier);
477 #else
478         kvm_guest_cpu_init();
479 #endif
480 }
481
482 static bool __init kvm_detect(void)
483 {
484         if (!kvm_para_available())
485                 return false;
486         return true;
487 }
488
489 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
490         .name                   = "KVM",
491         .detect                 = kvm_detect,
492 };
493 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
494
495 static __init int activate_jump_labels(void)
496 {
497         if (has_steal_clock) {
498                 static_key_slow_inc(&paravirt_steal_enabled);
499                 if (steal_acc)
500                         static_key_slow_inc(&paravirt_steal_rq_enabled);
501         }
502
503         return 0;
504 }
505 arch_initcall(activate_jump_labels);