blob: b5af88167613ded2aba4ebcdca2cb702d47c84eb [file] [log] [blame]
Avi Kivity6aa8b732006-12-10 02:21:36 -08001/*
2 * Kernel-based Virtual Machine driver for Linux
3 *
4 * This module enables machines with Intel VT-x extensions to run virtual
5 * machines without emulation or binary translation.
6 *
7 * Copyright (C) 2006 Qumranet, Inc.
8 *
9 * Authors:
10 * Avi Kivity <avi@qumranet.com>
11 * Yaniv Kamay <yaniv@qumranet.com>
12 *
13 * This work is licensed under the terms of the GNU GPL, version 2. See
14 * the COPYING file in the top-level directory.
15 *
16 */
17
Hollis Blancharde2174022007-12-03 15:30:24 -060018#include "iodev.h"
Avi Kivity6aa8b732006-12-10 02:21:36 -080019
Avi Kivityedf88412007-12-16 11:02:48 +020020#include <linux/kvm_host.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080021#include <linux/kvm.h>
22#include <linux/module.h>
23#include <linux/errno.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080024#include <linux/percpu.h>
25#include <linux/gfp.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080026#include <linux/mm.h>
27#include <linux/miscdevice.h>
28#include <linux/vmalloc.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080029#include <linux/reboot.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080030#include <linux/debugfs.h>
31#include <linux/highmem.h>
32#include <linux/file.h>
Avi Kivity59ae6c62007-02-12 00:54:48 -080033#include <linux/sysdev.h>
Avi Kivity774c47f2007-02-12 00:54:47 -080034#include <linux/cpu.h>
Alexey Dobriyane8edc6e2007-05-21 01:22:52 +040035#include <linux/sched.h>
Avi Kivityd9e368d2007-06-07 19:18:30 +030036#include <linux/cpumask.h>
37#include <linux/smp.h>
Avi Kivityd6d28162007-06-28 08:38:16 -040038#include <linux/anon_inodes.h>
Avi Kivity04d2cc72007-09-10 18:10:54 +030039#include <linux/profile.h>
Anthony Liguori7aa81cc2007-09-17 14:57:50 -050040#include <linux/kvm_para.h>
Izik Eidus6fc138d2007-10-09 19:20:39 +020041#include <linux/pagemap.h>
Anthony Liguori8d4e1282007-10-18 09:59:34 -050042#include <linux/mman.h>
Anthony Liguori35149e22008-04-02 14:46:56 -050043#include <linux/swap.h>
Sheng Yange56d5322009-03-12 21:45:39 +080044#include <linux/bitops.h>
Marcelo Tosatti547de292009-05-07 17:55:13 -030045#include <linux/spinlock.h>
Arnd Bergmann6ff58942009-10-22 14:19:27 +020046#include <linux/compat.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080047
Avi Kivitye4956062007-06-28 14:15:57 -040048#include <asm/processor.h>
Avi Kivitye4956062007-06-28 14:15:57 -040049#include <asm/io.h>
50#include <asm/uaccess.h>
Izik Eidus3e021bf2007-11-19 11:16:57 +020051#include <asm/pgtable.h>
Alexander Grafc8240bd2009-10-30 05:47:26 +000052#include <asm-generic/bitops/le.h>
Avi Kivity6aa8b732006-12-10 02:21:36 -080053
Laurent Vivier5f94c172008-05-30 16:05:54 +020054#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
55#include "coalesced_mmio.h"
56#endif
57
Marcelo Tosatti229456f2009-06-17 09:22:14 -030058#define CREATE_TRACE_POINTS
59#include <trace/events/kvm.h>
60
Avi Kivity6aa8b732006-12-10 02:21:36 -080061MODULE_AUTHOR("Qumranet");
62MODULE_LICENSE("GPL");
63
Marcelo Tosattifa40a822009-06-04 15:08:24 -030064/*
65 * Ordering of locks:
66 *
Michael S. Tsirkin22fc0292009-06-29 22:24:45 +030067 * kvm->slots_lock --> kvm->lock --> kvm->irq_lock
Marcelo Tosattifa40a822009-06-04 15:08:24 -030068 */
69
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +080070DEFINE_SPINLOCK(kvm_lock);
71LIST_HEAD(vm_list);
Avi Kivity133de902007-02-12 00:54:44 -080072
Rusty Russell7f59f492008-12-07 21:25:45 +103073static cpumask_var_t cpus_hardware_enabled;
Alexander Graf10474ae2009-09-15 11:37:46 +020074static int kvm_usage_count = 0;
75static atomic_t hardware_enable_failed;
Avi Kivity1b6c0162007-05-24 13:03:52 +030076
Rusty Russellc16f8622007-07-30 21:12:19 +100077struct kmem_cache *kvm_vcpu_cache;
78EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
Avi Kivity1165f5f2007-04-19 17:27:43 +030079
Avi Kivity15ad7142007-07-11 18:17:21 +030080static __read_mostly struct preempt_ops kvm_preempt_ops;
81
Hollis Blanchard76f7c872008-04-15 16:05:42 -050082struct dentry *kvm_debugfs_dir;
Avi Kivity6aa8b732006-12-10 02:21:36 -080083
Avi Kivitybccf2152007-02-21 18:04:26 +020084static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
85 unsigned long arg);
Alexander Graf10474ae2009-09-15 11:37:46 +020086static int hardware_enable_all(void);
87static void hardware_disable_all(void);
Avi Kivitybccf2152007-02-21 18:04:26 +020088
Hannes Edere8ba5d32008-11-28 17:02:42 +010089static bool kvm_rebooting;
Avi Kivity4ecac3f2008-05-13 13:23:38 +030090
Marcelo Tosatti54dee992009-06-11 12:07:44 -030091static bool largepages_enabled = true;
92
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +080093inline int kvm_is_mmio_pfn(pfn_t pfn)
Ben-Ami Yassourcbff90a2008-07-28 19:26:24 +030094{
Joerg Roedelfc5659c2009-02-18 14:08:58 +010095 if (pfn_valid(pfn)) {
96 struct page *page = compound_head(pfn_to_page(pfn));
97 return PageReserved(page);
98 }
Ben-Ami Yassourcbff90a2008-07-28 19:26:24 +030099
100 return true;
101}
102
Avi Kivity6aa8b732006-12-10 02:21:36 -0800103/*
104 * Switches to specified vcpu, until a matching vcpu_put()
105 */
Carsten Otte313a3dc2007-10-11 19:16:52 +0200106void vcpu_load(struct kvm_vcpu *vcpu)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800107{
Avi Kivity15ad7142007-07-11 18:17:21 +0300108 int cpu;
109
Avi Kivitybccf2152007-02-21 18:04:26 +0200110 mutex_lock(&vcpu->mutex);
Avi Kivity15ad7142007-07-11 18:17:21 +0300111 cpu = get_cpu();
112 preempt_notifier_register(&vcpu->preempt_notifier);
Carsten Otte313a3dc2007-10-11 19:16:52 +0200113 kvm_arch_vcpu_load(vcpu, cpu);
Avi Kivity15ad7142007-07-11 18:17:21 +0300114 put_cpu();
Avi Kivitybccf2152007-02-21 18:04:26 +0200115}
116
Carsten Otte313a3dc2007-10-11 19:16:52 +0200117void vcpu_put(struct kvm_vcpu *vcpu)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800118{
Avi Kivity15ad7142007-07-11 18:17:21 +0300119 preempt_disable();
Carsten Otte313a3dc2007-10-11 19:16:52 +0200120 kvm_arch_vcpu_put(vcpu);
Avi Kivity15ad7142007-07-11 18:17:21 +0300121 preempt_notifier_unregister(&vcpu->preempt_notifier);
122 preempt_enable();
Avi Kivity6aa8b732006-12-10 02:21:36 -0800123 mutex_unlock(&vcpu->mutex);
124}
125
Avi Kivityd9e368d2007-06-07 19:18:30 +0300126static void ack_flush(void *_completed)
127{
Avi Kivityd9e368d2007-06-07 19:18:30 +0300128}
129
Rusty Russell49846892008-12-08 20:26:24 +1030130static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
Avi Kivityd9e368d2007-06-07 19:18:30 +0300131{
Avi Kivity597a5f52008-07-20 14:24:22 +0300132 int i, cpu, me;
Rusty Russell6ef7a1b2008-12-08 20:28:04 +1030133 cpumask_var_t cpus;
134 bool called = true;
Avi Kivityd9e368d2007-06-07 19:18:30 +0300135 struct kvm_vcpu *vcpu;
Avi Kivityd9e368d2007-06-07 19:18:30 +0300136
Li Zefan79f55992009-06-15 14:58:26 +0800137 zalloc_cpumask_var(&cpus, GFP_ATOMIC);
Rusty Russell6ef7a1b2008-12-08 20:28:04 +1030138
Marcelo Tosatti84261922009-06-17 10:53:47 -0300139 spin_lock(&kvm->requests_lock);
Jan Kiszkae601e3b2009-07-20 11:30:12 +0200140 me = smp_processor_id();
Gleb Natapov988a2ca2009-06-09 15:56:29 +0300141 kvm_for_each_vcpu(i, vcpu, kvm) {
Rusty Russell49846892008-12-08 20:26:24 +1030142 if (test_and_set_bit(req, &vcpu->requests))
Avi Kivityd9e368d2007-06-07 19:18:30 +0300143 continue;
144 cpu = vcpu->cpu;
Rusty Russell6ef7a1b2008-12-08 20:28:04 +1030145 if (cpus != NULL && cpu != -1 && cpu != me)
146 cpumask_set_cpu(cpu, cpus);
Avi Kivityd9e368d2007-06-07 19:18:30 +0300147 }
Rusty Russell6ef7a1b2008-12-08 20:28:04 +1030148 if (unlikely(cpus == NULL))
149 smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1);
150 else if (!cpumask_empty(cpus))
151 smp_call_function_many(cpus, ack_flush, NULL, 1);
152 else
153 called = false;
Marcelo Tosatti84261922009-06-17 10:53:47 -0300154 spin_unlock(&kvm->requests_lock);
Rusty Russell6ef7a1b2008-12-08 20:28:04 +1030155 free_cpumask_var(cpus);
Rusty Russell49846892008-12-08 20:26:24 +1030156 return called;
157}
158
159void kvm_flush_remote_tlbs(struct kvm *kvm)
160{
161 if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
162 ++kvm->stat.remote_tlb_flush;
Avi Kivityd9e368d2007-06-07 19:18:30 +0300163}
164
Marcelo Tosatti2e53d632008-02-20 14:47:24 -0500165void kvm_reload_remote_mmus(struct kvm *kvm)
166{
Rusty Russell49846892008-12-08 20:26:24 +1030167 make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD);
Marcelo Tosatti2e53d632008-02-20 14:47:24 -0500168}
169
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000170int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
171{
172 struct page *page;
173 int r;
174
175 mutex_init(&vcpu->mutex);
176 vcpu->cpu = -1;
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000177 vcpu->kvm = kvm;
178 vcpu->vcpu_id = id;
Eddie Dongb6958ce2007-07-18 12:15:21 +0300179 init_waitqueue_head(&vcpu->wq);
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000180
181 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
182 if (!page) {
183 r = -ENOMEM;
184 goto fail;
185 }
186 vcpu->run = page_address(page);
187
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +0800188 r = kvm_arch_vcpu_init(vcpu);
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000189 if (r < 0)
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +0800190 goto fail_free_run;
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000191 return 0;
192
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000193fail_free_run:
194 free_page((unsigned long)vcpu->run);
195fail:
Rusty Russell76fafa52007-10-08 10:50:48 +1000196 return r;
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000197}
198EXPORT_SYMBOL_GPL(kvm_vcpu_init);
199
200void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
201{
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +0800202 kvm_arch_vcpu_uninit(vcpu);
Rusty Russellfb3f0f52007-07-27 17:16:56 +1000203 free_page((unsigned long)vcpu->run);
204}
205EXPORT_SYMBOL_GPL(kvm_vcpu_uninit);
206
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200207#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
208static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn)
209{
210 return container_of(mn, struct kvm, mmu_notifier);
211}
212
213static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
214 struct mm_struct *mm,
215 unsigned long address)
216{
217 struct kvm *kvm = mmu_notifier_to_kvm(mn);
218 int need_tlb_flush;
219
220 /*
221 * When ->invalidate_page runs, the linux pte has been zapped
222 * already but the page is still allocated until
223 * ->invalidate_page returns. So if we increase the sequence
224 * here the kvm page fault will notice if the spte can't be
225 * established because the page is going to be freed. If
226 * instead the kvm page fault establishes the spte before
227 * ->invalidate_page runs, kvm_unmap_hva will release it
228 * before returning.
229 *
230 * The sequence increase only need to be seen at spin_unlock
231 * time, and not at spin_lock time.
232 *
233 * Increasing the sequence after the spin_unlock would be
234 * unsafe because the kvm page fault could then establish the
235 * pte after kvm_unmap_hva returned, without noticing the page
236 * is going to be freed.
237 */
238 spin_lock(&kvm->mmu_lock);
239 kvm->mmu_notifier_seq++;
240 need_tlb_flush = kvm_unmap_hva(kvm, address);
241 spin_unlock(&kvm->mmu_lock);
242
243 /* we've to flush the tlb before the pages can be freed */
244 if (need_tlb_flush)
245 kvm_flush_remote_tlbs(kvm);
246
247}
248
Izik Eidus3da0dd42009-09-23 21:47:18 +0300249static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
250 struct mm_struct *mm,
251 unsigned long address,
252 pte_t pte)
253{
254 struct kvm *kvm = mmu_notifier_to_kvm(mn);
255
256 spin_lock(&kvm->mmu_lock);
257 kvm->mmu_notifier_seq++;
258 kvm_set_spte_hva(kvm, address, pte);
259 spin_unlock(&kvm->mmu_lock);
260}
261
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200262static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
263 struct mm_struct *mm,
264 unsigned long start,
265 unsigned long end)
266{
267 struct kvm *kvm = mmu_notifier_to_kvm(mn);
268 int need_tlb_flush = 0;
269
270 spin_lock(&kvm->mmu_lock);
271 /*
272 * The count increase must become visible at unlock time as no
273 * spte can be established without taking the mmu_lock and
274 * count is also read inside the mmu_lock critical section.
275 */
276 kvm->mmu_notifier_count++;
277 for (; start < end; start += PAGE_SIZE)
278 need_tlb_flush |= kvm_unmap_hva(kvm, start);
279 spin_unlock(&kvm->mmu_lock);
280
281 /* we've to flush the tlb before the pages can be freed */
282 if (need_tlb_flush)
283 kvm_flush_remote_tlbs(kvm);
284}
285
286static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,
287 struct mm_struct *mm,
288 unsigned long start,
289 unsigned long end)
290{
291 struct kvm *kvm = mmu_notifier_to_kvm(mn);
292
293 spin_lock(&kvm->mmu_lock);
294 /*
295 * This sequence increase will notify the kvm page fault that
296 * the page that is going to be mapped in the spte could have
297 * been freed.
298 */
299 kvm->mmu_notifier_seq++;
300 /*
301 * The above sequence increase must be visible before the
302 * below count decrease but both values are read by the kvm
303 * page fault under mmu_lock spinlock so we don't need to add
304 * a smb_wmb() here in between the two.
305 */
306 kvm->mmu_notifier_count--;
307 spin_unlock(&kvm->mmu_lock);
308
309 BUG_ON(kvm->mmu_notifier_count < 0);
310}
311
312static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
313 struct mm_struct *mm,
314 unsigned long address)
315{
316 struct kvm *kvm = mmu_notifier_to_kvm(mn);
317 int young;
318
319 spin_lock(&kvm->mmu_lock);
320 young = kvm_age_hva(kvm, address);
321 spin_unlock(&kvm->mmu_lock);
322
323 if (young)
324 kvm_flush_remote_tlbs(kvm);
325
326 return young;
327}
328
Marcelo Tosatti85db06e2008-12-10 21:23:26 +0100329static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
330 struct mm_struct *mm)
331{
332 struct kvm *kvm = mmu_notifier_to_kvm(mn);
333 kvm_arch_flush_shadow(kvm);
334}
335
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200336static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
337 .invalidate_page = kvm_mmu_notifier_invalidate_page,
338 .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
339 .invalidate_range_end = kvm_mmu_notifier_invalidate_range_end,
340 .clear_flush_young = kvm_mmu_notifier_clear_flush_young,
Izik Eidus3da0dd42009-09-23 21:47:18 +0300341 .change_pte = kvm_mmu_notifier_change_pte,
Marcelo Tosatti85db06e2008-12-10 21:23:26 +0100342 .release = kvm_mmu_notifier_release,
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200343};
344#endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
345
Avi Kivityf17abe92007-02-21 19:28:04 +0200346static struct kvm *kvm_create_vm(void)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800347{
Alexander Graf10474ae2009-09-15 11:37:46 +0200348 int r = 0;
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +0800349 struct kvm *kvm = kvm_arch_create_vm();
Laurent Vivier5f94c172008-05-30 16:05:54 +0200350#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
351 struct page *page;
352#endif
Avi Kivity6aa8b732006-12-10 02:21:36 -0800353
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +0800354 if (IS_ERR(kvm))
355 goto out;
Alexander Graf10474ae2009-09-15 11:37:46 +0200356
357 r = hardware_enable_all();
358 if (r)
359 goto out_err_nodisable;
360
Avi Kivity75858a82009-01-04 17:10:50 +0200361#ifdef CONFIG_HAVE_KVM_IRQCHIP
362 INIT_HLIST_HEAD(&kvm->mask_notifier_list);
Gleb Natapov136bdfe2009-08-24 11:54:23 +0300363 INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
Avi Kivity75858a82009-01-04 17:10:50 +0200364#endif
Avi Kivity6aa8b732006-12-10 02:21:36 -0800365
Laurent Vivier5f94c172008-05-30 16:05:54 +0200366#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
367 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
368 if (!page) {
Alexander Graf10474ae2009-09-15 11:37:46 +0200369 r = -ENOMEM;
370 goto out_err;
Laurent Vivier5f94c172008-05-30 16:05:54 +0200371 }
372 kvm->coalesced_mmio_ring =
373 (struct kvm_coalesced_mmio_ring *)page_address(page);
374#endif
375
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200376#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
377 {
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200378 kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
Alexander Graf10474ae2009-09-15 11:37:46 +0200379 r = mmu_notifier_register(&kvm->mmu_notifier, current->mm);
380 if (r) {
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200381#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
382 put_page(page);
383#endif
Alexander Graf10474ae2009-09-15 11:37:46 +0200384 goto out_err;
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200385 }
386 }
387#endif
388
Avi Kivity6d4e4c42007-11-21 16:41:05 +0200389 kvm->mm = current->mm;
390 atomic_inc(&kvm->mm->mm_count);
Marcelo Tosattiaaee2c92007-12-20 19:18:26 -0500391 spin_lock_init(&kvm->mmu_lock);
Marcelo Tosatti84261922009-06-17 10:53:47 -0300392 spin_lock_init(&kvm->requests_lock);
Eddie Dong74906342007-06-19 18:05:03 +0300393 kvm_io_bus_init(&kvm->pio_bus);
Gregory Haskinsd34e6b12009-07-07 17:08:49 -0400394 kvm_eventfd_init(kvm);
Shaohua Li11ec2802007-07-23 14:51:37 +0800395 mutex_init(&kvm->lock);
Marcelo Tosatti60eead72009-06-04 15:08:23 -0300396 mutex_init(&kvm->irq_lock);
Gregory Haskins2eeb2e92007-05-31 14:08:53 -0400397 kvm_io_bus_init(&kvm->mmio_bus);
Izik Eidus72dc67a2008-02-10 18:04:15 +0200398 init_rwsem(&kvm->slots_lock);
Izik Eidusd39f13b2008-03-30 16:01:25 +0300399 atomic_set(&kvm->users_count, 1);
Rusty Russell5e58cfe2007-07-23 17:08:21 +1000400 spin_lock(&kvm_lock);
401 list_add(&kvm->vm_list, &vm_list);
402 spin_unlock(&kvm_lock);
Laurent Vivier5f94c172008-05-30 16:05:54 +0200403#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
404 kvm_coalesced_mmio_init(kvm);
405#endif
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +0800406out:
Avi Kivityf17abe92007-02-21 19:28:04 +0200407 return kvm;
Alexander Graf10474ae2009-09-15 11:37:46 +0200408
409out_err:
410 hardware_disable_all();
411out_err_nodisable:
412 kfree(kvm);
413 return ERR_PTR(r);
Avi Kivityf17abe92007-02-21 19:28:04 +0200414}
415
Avi Kivity6aa8b732006-12-10 02:21:36 -0800416/*
417 * Free any memory in @free but not in @dont.
418 */
419static void kvm_free_physmem_slot(struct kvm_memory_slot *free,
420 struct kvm_memory_slot *dont)
421{
Joerg Roedelec04b262009-06-19 15:16:23 +0200422 int i;
423
Izik Eidus290fc382007-09-27 14:11:22 +0200424 if (!dont || free->rmap != dont->rmap)
425 vfree(free->rmap);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800426
427 if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
428 vfree(free->dirty_bitmap);
429
Joerg Roedelec04b262009-06-19 15:16:23 +0200430
431 for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) {
432 if (!dont || free->lpage_info[i] != dont->lpage_info[i]) {
433 vfree(free->lpage_info[i]);
434 free->lpage_info[i] = NULL;
435 }
436 }
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300437
Avi Kivity6aa8b732006-12-10 02:21:36 -0800438 free->npages = 0;
Al Viro8b6d44c2007-02-09 16:38:40 +0000439 free->dirty_bitmap = NULL;
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500440 free->rmap = NULL;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800441}
442
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +0800443void kvm_free_physmem(struct kvm *kvm)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800444{
445 int i;
446
447 for (i = 0; i < kvm->nmemslots; ++i)
Al Viro8b6d44c2007-02-09 16:38:40 +0000448 kvm_free_physmem_slot(&kvm->memslots[i], NULL);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800449}
450
Avi Kivityf17abe92007-02-21 19:28:04 +0200451static void kvm_destroy_vm(struct kvm *kvm)
452{
Avi Kivity6d4e4c42007-11-21 16:41:05 +0200453 struct mm_struct *mm = kvm->mm;
454
Sheng Yangad8ba2c2009-01-06 10:03:02 +0800455 kvm_arch_sync_events(kvm);
Avi Kivity133de902007-02-12 00:54:44 -0800456 spin_lock(&kvm_lock);
457 list_del(&kvm->vm_list);
458 spin_unlock(&kvm_lock);
Avi Kivity399ec802008-11-19 13:58:46 +0200459 kvm_free_irq_routing(kvm);
Eddie Dong74906342007-06-19 18:05:03 +0300460 kvm_io_bus_destroy(&kvm->pio_bus);
Gregory Haskins2eeb2e92007-05-31 14:08:53 -0400461 kvm_io_bus_destroy(&kvm->mmio_bus);
Laurent Vivier5f94c172008-05-30 16:05:54 +0200462#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
463 if (kvm->coalesced_mmio_ring != NULL)
464 free_page((unsigned long)kvm->coalesced_mmio_ring);
465#endif
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200466#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
467 mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
Gleb Natapovf00be0c2009-03-19 12:20:36 +0200468#else
469 kvm_arch_flush_shadow(kvm);
Andrea Arcangelie930bff2008-07-25 16:24:52 +0200470#endif
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +0800471 kvm_arch_destroy_vm(kvm);
Alexander Graf10474ae2009-09-15 11:37:46 +0200472 hardware_disable_all();
Avi Kivity6d4e4c42007-11-21 16:41:05 +0200473 mmdrop(mm);
Avi Kivityf17abe92007-02-21 19:28:04 +0200474}
475
Izik Eidusd39f13b2008-03-30 16:01:25 +0300476void kvm_get_kvm(struct kvm *kvm)
477{
478 atomic_inc(&kvm->users_count);
479}
480EXPORT_SYMBOL_GPL(kvm_get_kvm);
481
482void kvm_put_kvm(struct kvm *kvm)
483{
484 if (atomic_dec_and_test(&kvm->users_count))
485 kvm_destroy_vm(kvm);
486}
487EXPORT_SYMBOL_GPL(kvm_put_kvm);
488
489
Avi Kivityf17abe92007-02-21 19:28:04 +0200490static int kvm_vm_release(struct inode *inode, struct file *filp)
491{
492 struct kvm *kvm = filp->private_data;
493
Gregory Haskins721eecb2009-05-20 10:30:49 -0400494 kvm_irqfd_release(kvm);
495
Izik Eidusd39f13b2008-03-30 16:01:25 +0300496 kvm_put_kvm(kvm);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800497 return 0;
498}
499
Avi Kivity6aa8b732006-12-10 02:21:36 -0800500/*
Avi Kivity6aa8b732006-12-10 02:21:36 -0800501 * Allocate some memory and give it an address in the guest physical address
502 * space.
503 *
504 * Discontiguous memory is allowed, mostly for framebuffers.
Sheng Yangf78e0e22007-10-29 09:40:42 +0800505 *
Marcelo Tosatti10589a42007-12-20 19:18:22 -0500506 * Must be called holding mmap_sem for write.
Avi Kivity6aa8b732006-12-10 02:21:36 -0800507 */
Sheng Yangf78e0e22007-10-29 09:40:42 +0800508int __kvm_set_memory_region(struct kvm *kvm,
509 struct kvm_userspace_memory_region *mem,
510 int user_alloc)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800511{
512 int r;
513 gfn_t base_gfn;
Heiko Carstens28bcb112009-09-03 17:35:35 +0200514 unsigned long npages;
515 unsigned long i;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800516 struct kvm_memory_slot *memslot;
517 struct kvm_memory_slot old, new;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800518
519 r = -EINVAL;
520 /* General sanity checks */
521 if (mem->memory_size & (PAGE_SIZE - 1))
522 goto out;
523 if (mem->guest_phys_addr & (PAGE_SIZE - 1))
524 goto out;
Sheng Yange7cacd42008-11-11 15:30:40 +0800525 if (user_alloc && (mem->userspace_addr & (PAGE_SIZE - 1)))
Hollis Blanchard78749802008-11-07 13:32:12 -0600526 goto out;
Izik Eiduse0d62c72007-10-24 23:57:46 +0200527 if (mem->slot >= KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800528 goto out;
529 if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
530 goto out;
531
532 memslot = &kvm->memslots[mem->slot];
533 base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
534 npages = mem->memory_size >> PAGE_SHIFT;
535
536 if (!npages)
537 mem->flags &= ~KVM_MEM_LOG_DIRTY_PAGES;
538
Avi Kivity6aa8b732006-12-10 02:21:36 -0800539 new = old = *memslot;
540
541 new.base_gfn = base_gfn;
542 new.npages = npages;
543 new.flags = mem->flags;
544
545 /* Disallow changing a memory slot's size. */
546 r = -EINVAL;
547 if (npages && old.npages && npages != old.npages)
Sheng Yangf78e0e22007-10-29 09:40:42 +0800548 goto out_free;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800549
550 /* Check for overlaps */
551 r = -EEXIST;
552 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
553 struct kvm_memory_slot *s = &kvm->memslots[i];
554
Jan Kiszka4cd481f2009-04-13 11:59:32 +0200555 if (s == memslot || !s->npages)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800556 continue;
557 if (!((base_gfn + npages <= s->base_gfn) ||
558 (base_gfn >= s->base_gfn + s->npages)))
Sheng Yangf78e0e22007-10-29 09:40:42 +0800559 goto out_free;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800560 }
Avi Kivity6aa8b732006-12-10 02:21:36 -0800561
Avi Kivity6aa8b732006-12-10 02:21:36 -0800562 /* Free page dirty bitmap if unneeded */
563 if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
Al Viro8b6d44c2007-02-09 16:38:40 +0000564 new.dirty_bitmap = NULL;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800565
566 r = -ENOMEM;
567
568 /* Allocate if a slot is being created */
Carsten Otteeff01142008-06-27 15:05:31 +0200569#ifndef CONFIG_S390
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500570 if (npages && !new.rmap) {
Mike Dayd77c26f2007-10-08 09:02:08 -0400571 new.rmap = vmalloc(npages * sizeof(struct page *));
Izik Eidus290fc382007-09-27 14:11:22 +0200572
573 if (!new.rmap)
Sheng Yangf78e0e22007-10-29 09:40:42 +0800574 goto out_free;
Izik Eidus290fc382007-09-27 14:11:22 +0200575
Izik Eidus290fc382007-09-27 14:11:22 +0200576 memset(new.rmap, 0, npages * sizeof(*new.rmap));
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500577
Izik Eidus80b14b52007-10-25 11:54:04 +0200578 new.user_alloc = user_alloc;
Andrea Arcangeli604b38a2008-07-25 16:32:03 +0200579 /*
580 * hva_to_rmmap() serialzies with the mmu_lock and to be
581 * safe it has to ignore memslots with !user_alloc &&
582 * !userspace_addr.
583 */
584 if (user_alloc)
585 new.userspace_addr = mem->userspace_addr;
586 else
587 new.userspace_addr = 0;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800588 }
Joerg Roedelec04b262009-06-19 15:16:23 +0200589 if (!npages)
590 goto skip_lpage;
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300591
Joerg Roedelec04b262009-06-19 15:16:23 +0200592 for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) {
Heiko Carstens28bcb112009-09-03 17:35:35 +0200593 unsigned long ugfn;
594 unsigned long j;
595 int lpages;
Joerg Roedelec04b262009-06-19 15:16:23 +0200596 int level = i + 2;
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300597
Joerg Roedelec04b262009-06-19 15:16:23 +0200598 /* Avoid unused variable warning if no large pages */
599 (void)level;
600
601 if (new.lpage_info[i])
602 continue;
603
604 lpages = 1 + (base_gfn + npages - 1) /
605 KVM_PAGES_PER_HPAGE(level);
606 lpages -= base_gfn / KVM_PAGES_PER_HPAGE(level);
607
608 new.lpage_info[i] = vmalloc(lpages * sizeof(*new.lpage_info[i]));
609
610 if (!new.lpage_info[i])
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300611 goto out_free;
612
Joerg Roedelec04b262009-06-19 15:16:23 +0200613 memset(new.lpage_info[i], 0,
614 lpages * sizeof(*new.lpage_info[i]));
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300615
Joerg Roedelec04b262009-06-19 15:16:23 +0200616 if (base_gfn % KVM_PAGES_PER_HPAGE(level))
617 new.lpage_info[i][0].write_count = 1;
618 if ((base_gfn+npages) % KVM_PAGES_PER_HPAGE(level))
619 new.lpage_info[i][lpages - 1].write_count = 1;
Avi Kivityac045272009-06-08 15:52:39 +0300620 ugfn = new.userspace_addr >> PAGE_SHIFT;
621 /*
622 * If the gfn and userspace address are not aligned wrt each
Marcelo Tosatti54dee992009-06-11 12:07:44 -0300623 * other, or if explicitly asked to, disable large page
624 * support for this slot
Avi Kivityac045272009-06-08 15:52:39 +0300625 */
Joerg Roedelec04b262009-06-19 15:16:23 +0200626 if ((base_gfn ^ ugfn) & (KVM_PAGES_PER_HPAGE(level) - 1) ||
Marcelo Tosatti54dee992009-06-11 12:07:44 -0300627 !largepages_enabled)
Joerg Roedelec04b262009-06-19 15:16:23 +0200628 for (j = 0; j < lpages; ++j)
629 new.lpage_info[i][j].write_count = 1;
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300630 }
Avi Kivity6aa8b732006-12-10 02:21:36 -0800631
Joerg Roedelec04b262009-06-19 15:16:23 +0200632skip_lpage:
633
Avi Kivity6aa8b732006-12-10 02:21:36 -0800634 /* Allocate page dirty bitmap if needed */
635 if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
636 unsigned dirty_bytes = ALIGN(npages, BITS_PER_LONG) / 8;
637
638 new.dirty_bitmap = vmalloc(dirty_bytes);
639 if (!new.dirty_bitmap)
Sheng Yangf78e0e22007-10-29 09:40:42 +0800640 goto out_free;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800641 memset(new.dirty_bitmap, 0, dirty_bytes);
Izik Eiduse2445842009-06-10 19:23:24 +0300642 if (old.npages)
643 kvm_arch_flush_shadow(kvm);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800644 }
Christian Borntraeger3eea8432009-06-23 17:24:06 +0200645#else /* not defined CONFIG_S390 */
646 new.user_alloc = user_alloc;
647 if (user_alloc)
648 new.userspace_addr = mem->userspace_addr;
Carsten Otteeff01142008-06-27 15:05:31 +0200649#endif /* not defined CONFIG_S390 */
Avi Kivity6aa8b732006-12-10 02:21:36 -0800650
Marcelo Tosatti34d4cb82008-07-10 20:49:31 -0300651 if (!npages)
652 kvm_arch_flush_shadow(kvm);
653
Andrea Arcangeli604b38a2008-07-25 16:32:03 +0200654 spin_lock(&kvm->mmu_lock);
655 if (mem->slot >= kvm->nmemslots)
656 kvm->nmemslots = mem->slot + 1;
657
Avi Kivity6aa8b732006-12-10 02:21:36 -0800658 *memslot = new;
Andrea Arcangeli604b38a2008-07-25 16:32:03 +0200659 spin_unlock(&kvm->mmu_lock);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800660
Zhang Xiantao0de10342007-11-20 16:25:04 +0800661 r = kvm_arch_set_memory_region(kvm, mem, old, user_alloc);
662 if (r) {
Andrea Arcangeli604b38a2008-07-25 16:32:03 +0200663 spin_lock(&kvm->mmu_lock);
Zhang Xiantao0de10342007-11-20 16:25:04 +0800664 *memslot = old;
Andrea Arcangeli604b38a2008-07-25 16:32:03 +0200665 spin_unlock(&kvm->mmu_lock);
Zhang Xiantao0de10342007-11-20 16:25:04 +0800666 goto out_free;
Zhang Xiantao3ad82a72007-11-20 13:11:38 +0800667 }
668
Glauber Costa6f897242008-12-03 13:40:51 -0200669 kvm_free_physmem_slot(&old, npages ? &new : NULL);
670 /* Slot deletion case: we have to update the current slot */
Marcelo Tosattib43b1902009-05-12 18:55:44 -0300671 spin_lock(&kvm->mmu_lock);
Glauber Costa6f897242008-12-03 13:40:51 -0200672 if (!npages)
673 *memslot = old;
Marcelo Tosattib43b1902009-05-12 18:55:44 -0300674 spin_unlock(&kvm->mmu_lock);
Xiantao Zhang8a98f662008-10-06 13:47:38 +0800675#ifdef CONFIG_DMAR
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +0300676 /* map the pages in iommu page table */
677 r = kvm_iommu_map_pages(kvm, base_gfn, npages);
678 if (r)
679 goto out;
Xiantao Zhang8a98f662008-10-06 13:47:38 +0800680#endif
Avi Kivity6aa8b732006-12-10 02:21:36 -0800681 return 0;
682
Sheng Yangf78e0e22007-10-29 09:40:42 +0800683out_free:
Avi Kivity6aa8b732006-12-10 02:21:36 -0800684 kvm_free_physmem_slot(&new, &old);
685out:
686 return r;
Izik Eidus210c7c42007-10-24 23:52:57 +0200687
688}
Sheng Yangf78e0e22007-10-29 09:40:42 +0800689EXPORT_SYMBOL_GPL(__kvm_set_memory_region);
690
691int kvm_set_memory_region(struct kvm *kvm,
692 struct kvm_userspace_memory_region *mem,
693 int user_alloc)
694{
695 int r;
696
Izik Eidus72dc67a2008-02-10 18:04:15 +0200697 down_write(&kvm->slots_lock);
Sheng Yangf78e0e22007-10-29 09:40:42 +0800698 r = __kvm_set_memory_region(kvm, mem, user_alloc);
Izik Eidus72dc67a2008-02-10 18:04:15 +0200699 up_write(&kvm->slots_lock);
Sheng Yangf78e0e22007-10-29 09:40:42 +0800700 return r;
701}
Izik Eidus210c7c42007-10-24 23:52:57 +0200702EXPORT_SYMBOL_GPL(kvm_set_memory_region);
703
Carsten Otte1fe779f2007-10-29 16:08:35 +0100704int kvm_vm_ioctl_set_memory_region(struct kvm *kvm,
705 struct
706 kvm_userspace_memory_region *mem,
707 int user_alloc)
Izik Eidus210c7c42007-10-24 23:52:57 +0200708{
Izik Eiduse0d62c72007-10-24 23:57:46 +0200709 if (mem->slot >= KVM_MEMORY_SLOTS)
710 return -EINVAL;
Izik Eidus210c7c42007-10-24 23:52:57 +0200711 return kvm_set_memory_region(kvm, mem, user_alloc);
Avi Kivity6aa8b732006-12-10 02:21:36 -0800712}
713
Zhang Xiantao5bb064d2007-11-18 20:29:43 +0800714int kvm_get_dirty_log(struct kvm *kvm,
715 struct kvm_dirty_log *log, int *is_dirty)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800716{
717 struct kvm_memory_slot *memslot;
718 int r, i;
719 int n;
720 unsigned long any = 0;
721
Avi Kivity6aa8b732006-12-10 02:21:36 -0800722 r = -EINVAL;
723 if (log->slot >= KVM_MEMORY_SLOTS)
724 goto out;
725
726 memslot = &kvm->memslots[log->slot];
727 r = -ENOENT;
728 if (!memslot->dirty_bitmap)
729 goto out;
730
Uri Lublincd1a4a92007-02-22 16:43:09 +0200731 n = ALIGN(memslot->npages, BITS_PER_LONG) / 8;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800732
Uri Lublincd1a4a92007-02-22 16:43:09 +0200733 for (i = 0; !any && i < n/sizeof(long); ++i)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800734 any = memslot->dirty_bitmap[i];
735
736 r = -EFAULT;
737 if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n))
738 goto out;
739
Zhang Xiantao5bb064d2007-11-18 20:29:43 +0800740 if (any)
741 *is_dirty = 1;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800742
743 r = 0;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800744out:
Avi Kivity6aa8b732006-12-10 02:21:36 -0800745 return r;
746}
747
Marcelo Tosatti54dee992009-06-11 12:07:44 -0300748void kvm_disable_largepages(void)
749{
750 largepages_enabled = false;
751}
752EXPORT_SYMBOL_GPL(kvm_disable_largepages);
753
Izik Eiduscea7bb22007-10-17 19:17:48 +0200754int is_error_page(struct page *page)
755{
756 return page == bad_page;
757}
758EXPORT_SYMBOL_GPL(is_error_page);
759
Anthony Liguori35149e22008-04-02 14:46:56 -0500760int is_error_pfn(pfn_t pfn)
761{
762 return pfn == bad_pfn;
763}
764EXPORT_SYMBOL_GPL(is_error_pfn);
765
Izik Eidusf9d46eb2007-11-11 22:02:22 +0200766static inline unsigned long bad_hva(void)
767{
768 return PAGE_OFFSET;
769}
770
771int kvm_is_error_hva(unsigned long addr)
772{
773 return addr == bad_hva();
774}
775EXPORT_SYMBOL_GPL(kvm_is_error_hva);
776
Izik Eidus28430992008-10-03 17:40:32 +0300777struct kvm_memory_slot *gfn_to_memslot_unaliased(struct kvm *kvm, gfn_t gfn)
Avi Kivity6aa8b732006-12-10 02:21:36 -0800778{
779 int i;
780
781 for (i = 0; i < kvm->nmemslots; ++i) {
782 struct kvm_memory_slot *memslot = &kvm->memslots[i];
783
784 if (gfn >= memslot->base_gfn
785 && gfn < memslot->base_gfn + memslot->npages)
786 return memslot;
787 }
Al Viro8b6d44c2007-02-09 16:38:40 +0000788 return NULL;
Avi Kivity6aa8b732006-12-10 02:21:36 -0800789}
Izik Eidus28430992008-10-03 17:40:32 +0300790EXPORT_SYMBOL_GPL(gfn_to_memslot_unaliased);
Avi Kivitye8207542007-03-30 16:54:30 +0300791
792struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
793{
794 gfn = unalias_gfn(kvm, gfn);
Izik Eidus28430992008-10-03 17:40:32 +0300795 return gfn_to_memslot_unaliased(kvm, gfn);
Avi Kivitye8207542007-03-30 16:54:30 +0300796}
Avi Kivity6aa8b732006-12-10 02:21:36 -0800797
Izik Eiduse0d62c72007-10-24 23:57:46 +0200798int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
799{
800 int i;
801
802 gfn = unalias_gfn(kvm, gfn);
803 for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
804 struct kvm_memory_slot *memslot = &kvm->memslots[i];
805
806 if (gfn >= memslot->base_gfn
807 && gfn < memslot->base_gfn + memslot->npages)
808 return 1;
809 }
810 return 0;
811}
812EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
813
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300814unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
Izik Eidus539cb662007-11-11 22:05:04 +0200815{
816 struct kvm_memory_slot *slot;
817
818 gfn = unalias_gfn(kvm, gfn);
Izik Eidus28430992008-10-03 17:40:32 +0300819 slot = gfn_to_memslot_unaliased(kvm, gfn);
Izik Eidus539cb662007-11-11 22:05:04 +0200820 if (!slot)
821 return bad_hva();
822 return (slot->userspace_addr + (gfn - slot->base_gfn) * PAGE_SIZE);
823}
Sheng Yang0d150292008-04-25 21:44:50 +0800824EXPORT_SYMBOL_GPL(gfn_to_hva);
Izik Eidus539cb662007-11-11 22:05:04 +0200825
Anthony Liguori35149e22008-04-02 14:46:56 -0500826pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
Avi Kivity954bbbc22007-03-30 14:02:32 +0300827{
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500828 struct page *page[1];
Izik Eidus539cb662007-11-11 22:05:04 +0200829 unsigned long addr;
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500830 int npages;
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500831 pfn_t pfn;
Avi Kivity954bbbc22007-03-30 14:02:32 +0300832
Avi Kivity60395222007-10-21 11:03:36 +0200833 might_sleep();
834
Izik Eidus539cb662007-11-11 22:05:04 +0200835 addr = gfn_to_hva(kvm, gfn);
836 if (kvm_is_error_hva(addr)) {
Izik Eidus8a7ae052007-10-18 11:09:33 +0200837 get_page(bad_page);
Anthony Liguori35149e22008-04-02 14:46:56 -0500838 return page_to_pfn(bad_page);
Izik Eidus8a7ae052007-10-18 11:09:33 +0200839 }
Izik Eidus8a7ae052007-10-18 11:09:33 +0200840
Marcelo Tosatti4c2155c2008-09-16 20:54:47 -0300841 npages = get_user_pages_fast(addr, 1, 1, page);
Izik Eidus539cb662007-11-11 22:05:04 +0200842
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500843 if (unlikely(npages != 1)) {
844 struct vm_area_struct *vma;
Anthony Liguori8d4e1282007-10-18 09:59:34 -0500845
Marcelo Tosatti4c2155c2008-09-16 20:54:47 -0300846 down_read(&current->mm->mmap_sem);
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500847 vma = find_vma(current->mm, addr);
Marcelo Tosatti4c2155c2008-09-16 20:54:47 -0300848
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500849 if (vma == NULL || addr < vma->vm_start ||
850 !(vma->vm_flags & VM_PFNMAP)) {
Marcelo Tosatti4c2155c2008-09-16 20:54:47 -0300851 up_read(&current->mm->mmap_sem);
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500852 get_page(bad_page);
853 return page_to_pfn(bad_page);
854 }
855
856 pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Marcelo Tosatti4c2155c2008-09-16 20:54:47 -0300857 up_read(&current->mm->mmap_sem);
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800858 BUG_ON(!kvm_is_mmio_pfn(pfn));
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500859 } else
860 pfn = page_to_pfn(page[0]);
861
862 return pfn;
Anthony Liguori35149e22008-04-02 14:46:56 -0500863}
864
865EXPORT_SYMBOL_GPL(gfn_to_pfn);
866
867struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
868{
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500869 pfn_t pfn;
870
871 pfn = gfn_to_pfn(kvm, gfn);
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800872 if (!kvm_is_mmio_pfn(pfn))
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500873 return pfn_to_page(pfn);
874
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800875 WARN_ON(kvm_is_mmio_pfn(pfn));
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500876
877 get_page(bad_page);
878 return bad_page;
Avi Kivity954bbbc22007-03-30 14:02:32 +0300879}
Anthony Liguoriaab61cc2007-10-29 15:15:20 -0500880
Avi Kivity954bbbc22007-03-30 14:02:32 +0300881EXPORT_SYMBOL_GPL(gfn_to_page);
882
Izik Eidusb4231d62007-11-20 11:49:33 +0200883void kvm_release_page_clean(struct page *page)
884{
Anthony Liguori35149e22008-04-02 14:46:56 -0500885 kvm_release_pfn_clean(page_to_pfn(page));
Izik Eidusb4231d62007-11-20 11:49:33 +0200886}
887EXPORT_SYMBOL_GPL(kvm_release_page_clean);
888
Anthony Liguori35149e22008-04-02 14:46:56 -0500889void kvm_release_pfn_clean(pfn_t pfn)
890{
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800891 if (!kvm_is_mmio_pfn(pfn))
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500892 put_page(pfn_to_page(pfn));
Anthony Liguori35149e22008-04-02 14:46:56 -0500893}
894EXPORT_SYMBOL_GPL(kvm_release_pfn_clean);
895
Izik Eidusb4231d62007-11-20 11:49:33 +0200896void kvm_release_page_dirty(struct page *page)
Izik Eidus8a7ae052007-10-18 11:09:33 +0200897{
Anthony Liguori35149e22008-04-02 14:46:56 -0500898 kvm_release_pfn_dirty(page_to_pfn(page));
Izik Eidus8a7ae052007-10-18 11:09:33 +0200899}
Izik Eidusb4231d62007-11-20 11:49:33 +0200900EXPORT_SYMBOL_GPL(kvm_release_page_dirty);
Izik Eidus8a7ae052007-10-18 11:09:33 +0200901
Anthony Liguori35149e22008-04-02 14:46:56 -0500902void kvm_release_pfn_dirty(pfn_t pfn)
903{
904 kvm_set_pfn_dirty(pfn);
905 kvm_release_pfn_clean(pfn);
906}
907EXPORT_SYMBOL_GPL(kvm_release_pfn_dirty);
908
909void kvm_set_page_dirty(struct page *page)
910{
911 kvm_set_pfn_dirty(page_to_pfn(page));
912}
913EXPORT_SYMBOL_GPL(kvm_set_page_dirty);
914
915void kvm_set_pfn_dirty(pfn_t pfn)
916{
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800917 if (!kvm_is_mmio_pfn(pfn)) {
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500918 struct page *page = pfn_to_page(pfn);
919 if (!PageReserved(page))
920 SetPageDirty(page);
921 }
Anthony Liguori35149e22008-04-02 14:46:56 -0500922}
923EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty);
924
925void kvm_set_pfn_accessed(pfn_t pfn)
926{
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800927 if (!kvm_is_mmio_pfn(pfn))
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500928 mark_page_accessed(pfn_to_page(pfn));
Anthony Liguori35149e22008-04-02 14:46:56 -0500929}
930EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed);
931
932void kvm_get_pfn(pfn_t pfn)
933{
Xiantao Zhangc77fb9d2008-09-27 10:55:40 +0800934 if (!kvm_is_mmio_pfn(pfn))
Anthony Liguori2e2e3732008-04-30 15:37:07 -0500935 get_page(pfn_to_page(pfn));
Anthony Liguori35149e22008-04-02 14:46:56 -0500936}
937EXPORT_SYMBOL_GPL(kvm_get_pfn);
938
Izik Eidus195aefd2007-10-01 22:14:18 +0200939static int next_segment(unsigned long len, int offset)
940{
941 if (len > PAGE_SIZE - offset)
942 return PAGE_SIZE - offset;
943 else
944 return len;
945}
946
947int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
948 int len)
949{
Izik Eiduse0506bc2007-11-11 22:10:22 +0200950 int r;
951 unsigned long addr;
Izik Eidus195aefd2007-10-01 22:14:18 +0200952
Izik Eiduse0506bc2007-11-11 22:10:22 +0200953 addr = gfn_to_hva(kvm, gfn);
954 if (kvm_is_error_hva(addr))
Izik Eidus195aefd2007-10-01 22:14:18 +0200955 return -EFAULT;
Izik Eiduse0506bc2007-11-11 22:10:22 +0200956 r = copy_from_user(data, (void __user *)addr + offset, len);
957 if (r)
958 return -EFAULT;
Izik Eidus195aefd2007-10-01 22:14:18 +0200959 return 0;
960}
961EXPORT_SYMBOL_GPL(kvm_read_guest_page);
962
963int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len)
964{
965 gfn_t gfn = gpa >> PAGE_SHIFT;
966 int seg;
967 int offset = offset_in_page(gpa);
968 int ret;
969
970 while ((seg = next_segment(len, offset)) != 0) {
971 ret = kvm_read_guest_page(kvm, gfn, data, offset, seg);
972 if (ret < 0)
973 return ret;
974 offset = 0;
975 len -= seg;
976 data += seg;
977 ++gfn;
978 }
979 return 0;
980}
981EXPORT_SYMBOL_GPL(kvm_read_guest);
982
Marcelo Tosatti7ec54582007-12-20 19:18:23 -0500983int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data,
984 unsigned long len)
985{
986 int r;
987 unsigned long addr;
988 gfn_t gfn = gpa >> PAGE_SHIFT;
989 int offset = offset_in_page(gpa);
990
991 addr = gfn_to_hva(kvm, gfn);
992 if (kvm_is_error_hva(addr))
993 return -EFAULT;
Andrea Arcangeli0aac03f2008-01-30 19:57:35 +0100994 pagefault_disable();
Marcelo Tosatti7ec54582007-12-20 19:18:23 -0500995 r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len);
Andrea Arcangeli0aac03f2008-01-30 19:57:35 +0100996 pagefault_enable();
Marcelo Tosatti7ec54582007-12-20 19:18:23 -0500997 if (r)
998 return -EFAULT;
999 return 0;
1000}
1001EXPORT_SYMBOL(kvm_read_guest_atomic);
1002
Izik Eidus195aefd2007-10-01 22:14:18 +02001003int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, const void *data,
1004 int offset, int len)
1005{
Izik Eiduse0506bc2007-11-11 22:10:22 +02001006 int r;
1007 unsigned long addr;
Izik Eidus195aefd2007-10-01 22:14:18 +02001008
Izik Eiduse0506bc2007-11-11 22:10:22 +02001009 addr = gfn_to_hva(kvm, gfn);
1010 if (kvm_is_error_hva(addr))
Izik Eidus195aefd2007-10-01 22:14:18 +02001011 return -EFAULT;
Izik Eiduse0506bc2007-11-11 22:10:22 +02001012 r = copy_to_user((void __user *)addr + offset, data, len);
1013 if (r)
1014 return -EFAULT;
Izik Eidus195aefd2007-10-01 22:14:18 +02001015 mark_page_dirty(kvm, gfn);
1016 return 0;
1017}
1018EXPORT_SYMBOL_GPL(kvm_write_guest_page);
1019
1020int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
1021 unsigned long len)
1022{
1023 gfn_t gfn = gpa >> PAGE_SHIFT;
1024 int seg;
1025 int offset = offset_in_page(gpa);
1026 int ret;
1027
1028 while ((seg = next_segment(len, offset)) != 0) {
1029 ret = kvm_write_guest_page(kvm, gfn, data, offset, seg);
1030 if (ret < 0)
1031 return ret;
1032 offset = 0;
1033 len -= seg;
1034 data += seg;
1035 ++gfn;
1036 }
1037 return 0;
1038}
1039
1040int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
1041{
Izik Eidus3e021bf2007-11-19 11:16:57 +02001042 return kvm_write_guest_page(kvm, gfn, empty_zero_page, offset, len);
Izik Eidus195aefd2007-10-01 22:14:18 +02001043}
1044EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
1045
1046int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
1047{
1048 gfn_t gfn = gpa >> PAGE_SHIFT;
1049 int seg;
1050 int offset = offset_in_page(gpa);
1051 int ret;
1052
1053 while ((seg = next_segment(len, offset)) != 0) {
1054 ret = kvm_clear_guest_page(kvm, gfn, offset, seg);
1055 if (ret < 0)
1056 return ret;
1057 offset = 0;
1058 len -= seg;
1059 ++gfn;
1060 }
1061 return 0;
1062}
1063EXPORT_SYMBOL_GPL(kvm_clear_guest);
1064
Avi Kivity6aa8b732006-12-10 02:21:36 -08001065void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
1066{
Nguyen Anh Quynh31389942007-06-05 10:35:19 +03001067 struct kvm_memory_slot *memslot;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001068
Uri Lublin3b6fff12007-10-30 10:42:09 +02001069 gfn = unalias_gfn(kvm, gfn);
Izik Eidus28430992008-10-03 17:40:32 +03001070 memslot = gfn_to_memslot_unaliased(kvm, gfn);
Rusty Russell7e9d6192007-07-31 20:41:14 +10001071 if (memslot && memslot->dirty_bitmap) {
1072 unsigned long rel_gfn = gfn - memslot->base_gfn;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001073
Rusty Russell7e9d6192007-07-31 20:41:14 +10001074 /* avoid RMW */
Alexander Grafc8240bd2009-10-30 05:47:26 +00001075 if (!generic_test_le_bit(rel_gfn, memslot->dirty_bitmap))
1076 generic___set_le_bit(rel_gfn, memslot->dirty_bitmap);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001077 }
1078}
1079
Eddie Dongb6958ce2007-07-18 12:15:21 +03001080/*
1081 * The vCPU has executed a HLT instruction with in-kernel mode enabled.
1082 */
Hollis Blanchard8776e512007-10-31 17:24:24 -05001083void kvm_vcpu_block(struct kvm_vcpu *vcpu)
Eddie Dongb6958ce2007-07-18 12:15:21 +03001084{
Marcelo Tosattie5c239c2008-05-08 19:47:01 -03001085 DEFINE_WAIT(wait);
Eddie Dongb6958ce2007-07-18 12:15:21 +03001086
Marcelo Tosattie5c239c2008-05-08 19:47:01 -03001087 for (;;) {
1088 prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
Eddie Dongb6958ce2007-07-18 12:15:21 +03001089
Gleb Natapova1b37102009-07-09 15:33:52 +03001090 if (kvm_arch_vcpu_runnable(vcpu)) {
Marcelo Tosattid7690172008-09-08 15:23:48 -03001091 set_bit(KVM_REQ_UNHALT, &vcpu->requests);
Marcelo Tosattie5c239c2008-05-08 19:47:01 -03001092 break;
Marcelo Tosattid7690172008-09-08 15:23:48 -03001093 }
Gleb Natapov09cec752009-03-23 15:11:44 +02001094 if (kvm_cpu_has_pending_timer(vcpu))
1095 break;
Marcelo Tosattie5c239c2008-05-08 19:47:01 -03001096 if (signal_pending(current))
1097 break;
1098
Eddie Dongb6958ce2007-07-18 12:15:21 +03001099 schedule();
Eddie Dongb6958ce2007-07-18 12:15:21 +03001100 }
1101
Marcelo Tosattie5c239c2008-05-08 19:47:01 -03001102 finish_wait(&vcpu->wq, &wait);
Eddie Dongb6958ce2007-07-18 12:15:21 +03001103}
1104
Avi Kivity6aa8b732006-12-10 02:21:36 -08001105void kvm_resched(struct kvm_vcpu *vcpu)
1106{
Yaozu Dong3fca0362007-04-25 16:49:19 +03001107 if (!need_resched())
1108 return;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001109 cond_resched();
Avi Kivity6aa8b732006-12-10 02:21:36 -08001110}
1111EXPORT_SYMBOL_GPL(kvm_resched);
1112
Zhai, Edwind255f4f2009-10-09 18:03:20 +08001113void kvm_vcpu_on_spin(struct kvm_vcpu *vcpu)
1114{
1115 ktime_t expires;
1116 DEFINE_WAIT(wait);
1117
1118 prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
1119
1120 /* Sleep for 100 us, and hope lock-holder got scheduled */
1121 expires = ktime_add_ns(ktime_get(), 100000UL);
1122 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1123
1124 finish_wait(&vcpu->wq, &wait);
1125}
1126EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
1127
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001128static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001129{
1130 struct kvm_vcpu *vcpu = vma->vm_file->private_data;
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001131 struct page *page;
1132
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001133 if (vmf->pgoff == 0)
Avi Kivity039576c2007-03-20 12:46:50 +02001134 page = virt_to_page(vcpu->run);
Avi Kivity09566762008-01-23 18:14:23 +02001135#ifdef CONFIG_X86
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001136 else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET)
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001137 page = virt_to_page(vcpu->arch.pio_data);
Avi Kivity09566762008-01-23 18:14:23 +02001138#endif
Laurent Vivier5f94c172008-05-30 16:05:54 +02001139#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
1140 else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET)
1141 page = virt_to_page(vcpu->kvm->coalesced_mmio_ring);
1142#endif
Avi Kivity039576c2007-03-20 12:46:50 +02001143 else
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001144 return VM_FAULT_SIGBUS;
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001145 get_page(page);
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001146 vmf->page = page;
1147 return 0;
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001148}
1149
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04001150static const struct vm_operations_struct kvm_vcpu_vm_ops = {
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001151 .fault = kvm_vcpu_fault,
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001152};
1153
1154static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma)
1155{
1156 vma->vm_ops = &kvm_vcpu_vm_ops;
1157 return 0;
1158}
1159
Avi Kivitybccf2152007-02-21 18:04:26 +02001160static int kvm_vcpu_release(struct inode *inode, struct file *filp)
1161{
1162 struct kvm_vcpu *vcpu = filp->private_data;
1163
Al Viro66c0b392008-04-19 20:33:56 +01001164 kvm_put_kvm(vcpu->kvm);
Avi Kivitybccf2152007-02-21 18:04:26 +02001165 return 0;
1166}
1167
Christian Borntraeger3d3aab12008-12-02 11:17:32 +01001168static struct file_operations kvm_vcpu_fops = {
Avi Kivitybccf2152007-02-21 18:04:26 +02001169 .release = kvm_vcpu_release,
1170 .unlocked_ioctl = kvm_vcpu_ioctl,
1171 .compat_ioctl = kvm_vcpu_ioctl,
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001172 .mmap = kvm_vcpu_mmap,
Avi Kivitybccf2152007-02-21 18:04:26 +02001173};
1174
1175/*
1176 * Allocates an inode for the vcpu.
1177 */
1178static int create_vcpu_fd(struct kvm_vcpu *vcpu)
1179{
Roland Dreier628ff7c2009-12-18 09:41:24 -08001180 return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR);
Avi Kivitybccf2152007-02-21 18:04:26 +02001181}
1182
Avi Kivityc5ea7662007-02-20 18:41:05 +02001183/*
1184 * Creates some virtual cpus. Good luck creating more than one.
1185 */
Gleb Natapov73880c82009-06-09 15:56:28 +03001186static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
Avi Kivityc5ea7662007-02-20 18:41:05 +02001187{
1188 int r;
Gleb Natapov988a2ca2009-06-09 15:56:29 +03001189 struct kvm_vcpu *vcpu, *v;
Avi Kivityc5ea7662007-02-20 18:41:05 +02001190
Gleb Natapov73880c82009-06-09 15:56:28 +03001191 vcpu = kvm_arch_vcpu_create(kvm, id);
Rusty Russellfb3f0f52007-07-27 17:16:56 +10001192 if (IS_ERR(vcpu))
1193 return PTR_ERR(vcpu);
Avi Kivityc5ea7662007-02-20 18:41:05 +02001194
Avi Kivity15ad7142007-07-11 18:17:21 +03001195 preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops);
1196
Avi Kivity26e52152007-11-20 15:30:24 +02001197 r = kvm_arch_vcpu_setup(vcpu);
1198 if (r)
Glauber Costa7d8fece2008-09-17 23:16:59 -03001199 return r;
Avi Kivity26e52152007-11-20 15:30:24 +02001200
Shaohua Li11ec2802007-07-23 14:51:37 +08001201 mutex_lock(&kvm->lock);
Gleb Natapov73880c82009-06-09 15:56:28 +03001202 if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
1203 r = -EINVAL;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08001204 goto vcpu_destroy;
Rusty Russellfb3f0f52007-07-27 17:16:56 +10001205 }
Gleb Natapov73880c82009-06-09 15:56:28 +03001206
Gleb Natapov988a2ca2009-06-09 15:56:29 +03001207 kvm_for_each_vcpu(r, v, kvm)
1208 if (v->vcpu_id == id) {
Gleb Natapov73880c82009-06-09 15:56:28 +03001209 r = -EEXIST;
1210 goto vcpu_destroy;
1211 }
1212
1213 BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
Rusty Russellfb3f0f52007-07-27 17:16:56 +10001214
1215 /* Now it's all set up, let userspace reach it */
Al Viro66c0b392008-04-19 20:33:56 +01001216 kvm_get_kvm(kvm);
Avi Kivitybccf2152007-02-21 18:04:26 +02001217 r = create_vcpu_fd(vcpu);
Gleb Natapov73880c82009-06-09 15:56:28 +03001218 if (r < 0) {
1219 kvm_put_kvm(kvm);
1220 goto vcpu_destroy;
1221 }
1222
1223 kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
1224 smp_wmb();
1225 atomic_inc(&kvm->online_vcpus);
1226
1227#ifdef CONFIG_KVM_APIC_ARCHITECTURE
1228 if (kvm->bsp_vcpu_id == id)
1229 kvm->bsp_vcpu = vcpu;
1230#endif
1231 mutex_unlock(&kvm->lock);
Avi Kivitybccf2152007-02-21 18:04:26 +02001232 return r;
Avi Kivityc5ea7662007-02-20 18:41:05 +02001233
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08001234vcpu_destroy:
Glauber Costa7d8fece2008-09-17 23:16:59 -03001235 mutex_unlock(&kvm->lock);
Hollis Blanchardd40ccc62007-11-19 14:04:43 -06001236 kvm_arch_vcpu_destroy(vcpu);
Avi Kivityc5ea7662007-02-20 18:41:05 +02001237 return r;
1238}
1239
Avi Kivity1961d272007-03-05 19:46:05 +02001240static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
1241{
1242 if (sigset) {
1243 sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP));
1244 vcpu->sigset_active = 1;
1245 vcpu->sigset = *sigset;
1246 } else
1247 vcpu->sigset_active = 0;
1248 return 0;
1249}
1250
Avi Kivitybccf2152007-02-21 18:04:26 +02001251static long kvm_vcpu_ioctl(struct file *filp,
1252 unsigned int ioctl, unsigned long arg)
Avi Kivity6aa8b732006-12-10 02:21:36 -08001253{
Avi Kivitybccf2152007-02-21 18:04:26 +02001254 struct kvm_vcpu *vcpu = filp->private_data;
Al Viro2f366982007-02-09 16:38:35 +00001255 void __user *argp = (void __user *)arg;
Carsten Otte313a3dc2007-10-11 19:16:52 +02001256 int r;
Dave Hansenfa3795a2008-08-11 10:01:46 -07001257 struct kvm_fpu *fpu = NULL;
1258 struct kvm_sregs *kvm_sregs = NULL;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001259
Avi Kivity6d4e4c42007-11-21 16:41:05 +02001260 if (vcpu->kvm->mm != current->mm)
1261 return -EIO;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001262 switch (ioctl) {
Avi Kivity9a2bb7f2007-02-22 12:58:31 +02001263 case KVM_RUN:
Avi Kivityf0fe5102007-03-07 13:11:17 +02001264 r = -EINVAL;
1265 if (arg)
1266 goto out;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05001267 r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001268 break;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001269 case KVM_GET_REGS: {
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001270 struct kvm_regs *kvm_regs;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001271
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001272 r = -ENOMEM;
1273 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
1274 if (!kvm_regs)
1275 goto out;
1276 r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001277 if (r)
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001278 goto out_free1;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001279 r = -EFAULT;
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001280 if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs)))
1281 goto out_free1;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001282 r = 0;
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001283out_free1:
1284 kfree(kvm_regs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001285 break;
1286 }
1287 case KVM_SET_REGS: {
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001288 struct kvm_regs *kvm_regs;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001289
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001290 r = -ENOMEM;
1291 kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL);
1292 if (!kvm_regs)
1293 goto out;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001294 r = -EFAULT;
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001295 if (copy_from_user(kvm_regs, argp, sizeof(struct kvm_regs)))
1296 goto out_free2;
1297 r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001298 if (r)
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001299 goto out_free2;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001300 r = 0;
Xiantao Zhang3e4bb3a2008-02-25 18:52:20 +08001301out_free2:
1302 kfree(kvm_regs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001303 break;
1304 }
1305 case KVM_GET_SREGS: {
Dave Hansenfa3795a2008-08-11 10:01:46 -07001306 kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
1307 r = -ENOMEM;
1308 if (!kvm_sregs)
1309 goto out;
1310 r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001311 if (r)
1312 goto out;
1313 r = -EFAULT;
Dave Hansenfa3795a2008-08-11 10:01:46 -07001314 if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs)))
Avi Kivity6aa8b732006-12-10 02:21:36 -08001315 goto out;
1316 r = 0;
1317 break;
1318 }
1319 case KVM_SET_SREGS: {
Dave Hansenfa3795a2008-08-11 10:01:46 -07001320 kvm_sregs = kmalloc(sizeof(struct kvm_sregs), GFP_KERNEL);
1321 r = -ENOMEM;
1322 if (!kvm_sregs)
Avi Kivity6aa8b732006-12-10 02:21:36 -08001323 goto out;
Dave Hansenfa3795a2008-08-11 10:01:46 -07001324 r = -EFAULT;
1325 if (copy_from_user(kvm_sregs, argp, sizeof(struct kvm_sregs)))
1326 goto out;
1327 r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001328 if (r)
1329 goto out;
1330 r = 0;
1331 break;
1332 }
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03001333 case KVM_GET_MP_STATE: {
1334 struct kvm_mp_state mp_state;
1335
1336 r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state);
1337 if (r)
1338 goto out;
1339 r = -EFAULT;
1340 if (copy_to_user(argp, &mp_state, sizeof mp_state))
1341 goto out;
1342 r = 0;
1343 break;
1344 }
1345 case KVM_SET_MP_STATE: {
1346 struct kvm_mp_state mp_state;
1347
1348 r = -EFAULT;
1349 if (copy_from_user(&mp_state, argp, sizeof mp_state))
1350 goto out;
1351 r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state);
1352 if (r)
1353 goto out;
1354 r = 0;
1355 break;
1356 }
Avi Kivity6aa8b732006-12-10 02:21:36 -08001357 case KVM_TRANSLATE: {
1358 struct kvm_translation tr;
1359
1360 r = -EFAULT;
Al Viro2f366982007-02-09 16:38:35 +00001361 if (copy_from_user(&tr, argp, sizeof tr))
Avi Kivity6aa8b732006-12-10 02:21:36 -08001362 goto out;
Zhang Xiantao8b006792007-11-16 13:05:55 +08001363 r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001364 if (r)
1365 goto out;
1366 r = -EFAULT;
Al Viro2f366982007-02-09 16:38:35 +00001367 if (copy_to_user(argp, &tr, sizeof tr))
Avi Kivity6aa8b732006-12-10 02:21:36 -08001368 goto out;
1369 r = 0;
1370 break;
1371 }
Jan Kiszkad0bfb942008-12-15 13:52:10 +01001372 case KVM_SET_GUEST_DEBUG: {
1373 struct kvm_guest_debug dbg;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001374
1375 r = -EFAULT;
Al Viro2f366982007-02-09 16:38:35 +00001376 if (copy_from_user(&dbg, argp, sizeof dbg))
Avi Kivity6aa8b732006-12-10 02:21:36 -08001377 goto out;
Jan Kiszkad0bfb942008-12-15 13:52:10 +01001378 r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001379 if (r)
1380 goto out;
1381 r = 0;
1382 break;
1383 }
Avi Kivity1961d272007-03-05 19:46:05 +02001384 case KVM_SET_SIGNAL_MASK: {
1385 struct kvm_signal_mask __user *sigmask_arg = argp;
1386 struct kvm_signal_mask kvm_sigmask;
1387 sigset_t sigset, *p;
1388
1389 p = NULL;
1390 if (argp) {
1391 r = -EFAULT;
1392 if (copy_from_user(&kvm_sigmask, argp,
1393 sizeof kvm_sigmask))
1394 goto out;
1395 r = -EINVAL;
1396 if (kvm_sigmask.len != sizeof sigset)
1397 goto out;
1398 r = -EFAULT;
1399 if (copy_from_user(&sigset, sigmask_arg->sigset,
1400 sizeof sigset))
1401 goto out;
1402 p = &sigset;
1403 }
1404 r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
1405 break;
1406 }
Avi Kivityb8836732007-04-01 16:34:31 +03001407 case KVM_GET_FPU: {
Dave Hansenfa3795a2008-08-11 10:01:46 -07001408 fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
1409 r = -ENOMEM;
1410 if (!fpu)
1411 goto out;
1412 r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu);
Avi Kivityb8836732007-04-01 16:34:31 +03001413 if (r)
1414 goto out;
1415 r = -EFAULT;
Dave Hansenfa3795a2008-08-11 10:01:46 -07001416 if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu)))
Avi Kivityb8836732007-04-01 16:34:31 +03001417 goto out;
1418 r = 0;
1419 break;
1420 }
1421 case KVM_SET_FPU: {
Dave Hansenfa3795a2008-08-11 10:01:46 -07001422 fpu = kmalloc(sizeof(struct kvm_fpu), GFP_KERNEL);
1423 r = -ENOMEM;
1424 if (!fpu)
Avi Kivityb8836732007-04-01 16:34:31 +03001425 goto out;
Dave Hansenfa3795a2008-08-11 10:01:46 -07001426 r = -EFAULT;
1427 if (copy_from_user(fpu, argp, sizeof(struct kvm_fpu)))
1428 goto out;
1429 r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
Avi Kivityb8836732007-04-01 16:34:31 +03001430 if (r)
1431 goto out;
1432 r = 0;
1433 break;
1434 }
Avi Kivitybccf2152007-02-21 18:04:26 +02001435 default:
Carsten Otte313a3dc2007-10-11 19:16:52 +02001436 r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
Avi Kivitybccf2152007-02-21 18:04:26 +02001437 }
1438out:
Dave Hansenfa3795a2008-08-11 10:01:46 -07001439 kfree(fpu);
1440 kfree(kvm_sregs);
Avi Kivitybccf2152007-02-21 18:04:26 +02001441 return r;
1442}
1443
1444static long kvm_vm_ioctl(struct file *filp,
1445 unsigned int ioctl, unsigned long arg)
1446{
1447 struct kvm *kvm = filp->private_data;
1448 void __user *argp = (void __user *)arg;
Carsten Otte1fe779f2007-10-29 16:08:35 +01001449 int r;
Avi Kivitybccf2152007-02-21 18:04:26 +02001450
Avi Kivity6d4e4c42007-11-21 16:41:05 +02001451 if (kvm->mm != current->mm)
1452 return -EIO;
Avi Kivitybccf2152007-02-21 18:04:26 +02001453 switch (ioctl) {
1454 case KVM_CREATE_VCPU:
1455 r = kvm_vm_ioctl_create_vcpu(kvm, arg);
1456 if (r < 0)
1457 goto out;
1458 break;
Izik Eidus6fc138d2007-10-09 19:20:39 +02001459 case KVM_SET_USER_MEMORY_REGION: {
1460 struct kvm_userspace_memory_region kvm_userspace_mem;
1461
1462 r = -EFAULT;
1463 if (copy_from_user(&kvm_userspace_mem, argp,
1464 sizeof kvm_userspace_mem))
1465 goto out;
1466
1467 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem, 1);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001468 if (r)
1469 goto out;
1470 break;
1471 }
1472 case KVM_GET_DIRTY_LOG: {
1473 struct kvm_dirty_log log;
1474
1475 r = -EFAULT;
Al Viro2f366982007-02-09 16:38:35 +00001476 if (copy_from_user(&log, argp, sizeof log))
Avi Kivity6aa8b732006-12-10 02:21:36 -08001477 goto out;
Avi Kivity2c6f5df2007-02-20 18:27:58 +02001478 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001479 if (r)
1480 goto out;
1481 break;
1482 }
Laurent Vivier5f94c172008-05-30 16:05:54 +02001483#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
1484 case KVM_REGISTER_COALESCED_MMIO: {
1485 struct kvm_coalesced_mmio_zone zone;
1486 r = -EFAULT;
1487 if (copy_from_user(&zone, argp, sizeof zone))
1488 goto out;
1489 r = -ENXIO;
1490 r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
1491 if (r)
1492 goto out;
1493 r = 0;
1494 break;
1495 }
1496 case KVM_UNREGISTER_COALESCED_MMIO: {
1497 struct kvm_coalesced_mmio_zone zone;
1498 r = -EFAULT;
1499 if (copy_from_user(&zone, argp, sizeof zone))
1500 goto out;
1501 r = -ENXIO;
1502 r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
1503 if (r)
1504 goto out;
1505 r = 0;
1506 break;
1507 }
1508#endif
Gregory Haskins721eecb2009-05-20 10:30:49 -04001509 case KVM_IRQFD: {
1510 struct kvm_irqfd data;
1511
1512 r = -EFAULT;
1513 if (copy_from_user(&data, argp, sizeof data))
1514 goto out;
1515 r = kvm_irqfd(kvm, data.fd, data.gsi, data.flags);
1516 break;
1517 }
Gregory Haskinsd34e6b12009-07-07 17:08:49 -04001518 case KVM_IOEVENTFD: {
1519 struct kvm_ioeventfd data;
1520
1521 r = -EFAULT;
1522 if (copy_from_user(&data, argp, sizeof data))
1523 goto out;
1524 r = kvm_ioeventfd(kvm, &data);
1525 break;
1526 }
Gleb Natapov73880c82009-06-09 15:56:28 +03001527#ifdef CONFIG_KVM_APIC_ARCHITECTURE
1528 case KVM_SET_BOOT_CPU_ID:
1529 r = 0;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03001530 mutex_lock(&kvm->lock);
Gleb Natapov73880c82009-06-09 15:56:28 +03001531 if (atomic_read(&kvm->online_vcpus) != 0)
1532 r = -EBUSY;
1533 else
1534 kvm->bsp_vcpu_id = arg;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03001535 mutex_unlock(&kvm->lock);
Gleb Natapov73880c82009-06-09 15:56:28 +03001536 break;
1537#endif
Avi Kivityf17abe92007-02-21 19:28:04 +02001538 default:
Carsten Otte1fe779f2007-10-29 16:08:35 +01001539 r = kvm_arch_vm_ioctl(filp, ioctl, arg);
Avi Kivitybfd99ff2009-08-26 14:57:50 +03001540 if (r == -ENOTTY)
1541 r = kvm_vm_ioctl_assigned_device(kvm, ioctl, arg);
Avi Kivityf17abe92007-02-21 19:28:04 +02001542 }
1543out:
1544 return r;
1545}
1546
Arnd Bergmann6ff58942009-10-22 14:19:27 +02001547#ifdef CONFIG_COMPAT
1548struct compat_kvm_dirty_log {
1549 __u32 slot;
1550 __u32 padding1;
1551 union {
1552 compat_uptr_t dirty_bitmap; /* one bit per page */
1553 __u64 padding2;
1554 };
1555};
1556
1557static long kvm_vm_compat_ioctl(struct file *filp,
1558 unsigned int ioctl, unsigned long arg)
1559{
1560 struct kvm *kvm = filp->private_data;
1561 int r;
1562
1563 if (kvm->mm != current->mm)
1564 return -EIO;
1565 switch (ioctl) {
1566 case KVM_GET_DIRTY_LOG: {
1567 struct compat_kvm_dirty_log compat_log;
1568 struct kvm_dirty_log log;
1569
1570 r = -EFAULT;
1571 if (copy_from_user(&compat_log, (void __user *)arg,
1572 sizeof(compat_log)))
1573 goto out;
1574 log.slot = compat_log.slot;
1575 log.padding1 = compat_log.padding1;
1576 log.padding2 = compat_log.padding2;
1577 log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
1578
1579 r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
1580 if (r)
1581 goto out;
1582 break;
1583 }
1584 default:
1585 r = kvm_vm_ioctl(filp, ioctl, arg);
1586 }
1587
1588out:
1589 return r;
1590}
1591#endif
1592
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001593static int kvm_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
Avi Kivityf17abe92007-02-21 19:28:04 +02001594{
Marcelo Tosatti777b3f42008-09-16 20:54:46 -03001595 struct page *page[1];
1596 unsigned long addr;
1597 int npages;
1598 gfn_t gfn = vmf->pgoff;
Avi Kivityf17abe92007-02-21 19:28:04 +02001599 struct kvm *kvm = vma->vm_file->private_data;
Avi Kivityf17abe92007-02-21 19:28:04 +02001600
Marcelo Tosatti777b3f42008-09-16 20:54:46 -03001601 addr = gfn_to_hva(kvm, gfn);
1602 if (kvm_is_error_hva(addr))
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001603 return VM_FAULT_SIGBUS;
Marcelo Tosatti777b3f42008-09-16 20:54:46 -03001604
1605 npages = get_user_pages(current, current->mm, addr, 1, 1, 0, page,
1606 NULL);
1607 if (unlikely(npages != 1))
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001608 return VM_FAULT_SIGBUS;
Marcelo Tosatti777b3f42008-09-16 20:54:46 -03001609
1610 vmf->page = page[0];
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001611 return 0;
Avi Kivityf17abe92007-02-21 19:28:04 +02001612}
1613
Alexey Dobriyanf0f37e22009-09-27 22:29:37 +04001614static const struct vm_operations_struct kvm_vm_vm_ops = {
npiggin@suse.dee4a533a2007-12-05 18:15:52 +11001615 .fault = kvm_vm_fault,
Avi Kivityf17abe92007-02-21 19:28:04 +02001616};
1617
1618static int kvm_vm_mmap(struct file *file, struct vm_area_struct *vma)
1619{
1620 vma->vm_ops = &kvm_vm_vm_ops;
1621 return 0;
1622}
1623
Christian Borntraeger3d3aab12008-12-02 11:17:32 +01001624static struct file_operations kvm_vm_fops = {
Avi Kivityf17abe92007-02-21 19:28:04 +02001625 .release = kvm_vm_release,
1626 .unlocked_ioctl = kvm_vm_ioctl,
Arnd Bergmann6ff58942009-10-22 14:19:27 +02001627#ifdef CONFIG_COMPAT
1628 .compat_ioctl = kvm_vm_compat_ioctl,
1629#endif
Avi Kivityf17abe92007-02-21 19:28:04 +02001630 .mmap = kvm_vm_mmap,
1631};
1632
1633static int kvm_dev_ioctl_create_vm(void)
1634{
Al Viro2030a422008-02-23 06:46:49 -05001635 int fd;
Avi Kivityf17abe92007-02-21 19:28:04 +02001636 struct kvm *kvm;
1637
Avi Kivityf17abe92007-02-21 19:28:04 +02001638 kvm = kvm_create_vm();
Avi Kivityd6d28162007-06-28 08:38:16 -04001639 if (IS_ERR(kvm))
1640 return PTR_ERR(kvm);
Roland Dreier628ff7c2009-12-18 09:41:24 -08001641 fd = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR);
Al Viro2030a422008-02-23 06:46:49 -05001642 if (fd < 0)
Al Viro66c0b392008-04-19 20:33:56 +01001643 kvm_put_kvm(kvm);
Avi Kivityf17abe92007-02-21 19:28:04 +02001644
Avi Kivityf17abe92007-02-21 19:28:04 +02001645 return fd;
Avi Kivityf17abe92007-02-21 19:28:04 +02001646}
1647
Avi Kivity1a811b62008-12-08 18:25:27 +02001648static long kvm_dev_ioctl_check_extension_generic(long arg)
1649{
1650 switch (arg) {
Avi Kivityca9edae2008-12-08 18:29:29 +02001651 case KVM_CAP_USER_MEMORY:
Avi Kivity1a811b62008-12-08 18:25:27 +02001652 case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
Jan Kiszka4cd481f2009-04-13 11:59:32 +02001653 case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS:
Gleb Natapov73880c82009-06-09 15:56:28 +03001654#ifdef CONFIG_KVM_APIC_ARCHITECTURE
1655 case KVM_CAP_SET_BOOT_CPU_ID:
1656#endif
Avi Kivitya9c73992009-11-04 11:54:59 +02001657 case KVM_CAP_INTERNAL_ERROR_DATA:
Avi Kivity1a811b62008-12-08 18:25:27 +02001658 return 1;
Avi Kivity399ec802008-11-19 13:58:46 +02001659#ifdef CONFIG_HAVE_KVM_IRQCHIP
1660 case KVM_CAP_IRQ_ROUTING:
Sheng Yang36463142009-03-16 16:33:43 +08001661 return KVM_MAX_IRQ_ROUTES;
Avi Kivity399ec802008-11-19 13:58:46 +02001662#endif
Avi Kivity1a811b62008-12-08 18:25:27 +02001663 default:
1664 break;
1665 }
1666 return kvm_dev_ioctl_check_extension(arg);
1667}
1668
Avi Kivityf17abe92007-02-21 19:28:04 +02001669static long kvm_dev_ioctl(struct file *filp,
1670 unsigned int ioctl, unsigned long arg)
1671{
Avi Kivity07c45a32007-03-07 13:05:38 +02001672 long r = -EINVAL;
Avi Kivityf17abe92007-02-21 19:28:04 +02001673
1674 switch (ioctl) {
1675 case KVM_GET_API_VERSION:
Avi Kivityf0fe5102007-03-07 13:11:17 +02001676 r = -EINVAL;
1677 if (arg)
1678 goto out;
Avi Kivityf17abe92007-02-21 19:28:04 +02001679 r = KVM_API_VERSION;
1680 break;
1681 case KVM_CREATE_VM:
Avi Kivityf0fe5102007-03-07 13:11:17 +02001682 r = -EINVAL;
1683 if (arg)
1684 goto out;
Avi Kivityf17abe92007-02-21 19:28:04 +02001685 r = kvm_dev_ioctl_create_vm();
1686 break;
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001687 case KVM_CHECK_EXTENSION:
Avi Kivity1a811b62008-12-08 18:25:27 +02001688 r = kvm_dev_ioctl_check_extension_generic(arg);
Avi Kivity5d308f42007-03-01 17:56:20 +02001689 break;
Avi Kivity07c45a32007-03-07 13:05:38 +02001690 case KVM_GET_VCPU_MMAP_SIZE:
1691 r = -EINVAL;
1692 if (arg)
1693 goto out;
Avi Kivityadb1ff42008-01-24 15:13:08 +02001694 r = PAGE_SIZE; /* struct kvm_run */
1695#ifdef CONFIG_X86
1696 r += PAGE_SIZE; /* pio data page */
1697#endif
Laurent Vivier5f94c172008-05-30 16:05:54 +02001698#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
1699 r += PAGE_SIZE; /* coalesced mmio ring page */
1700#endif
Avi Kivity07c45a32007-03-07 13:05:38 +02001701 break;
Feng(Eric) Liud4c9ff22008-04-10 08:47:53 -04001702 case KVM_TRACE_ENABLE:
1703 case KVM_TRACE_PAUSE:
1704 case KVM_TRACE_DISABLE:
Marcelo Tosatti2023a292009-06-18 11:47:28 -03001705 r = -EOPNOTSUPP;
Feng(Eric) Liud4c9ff22008-04-10 08:47:53 -04001706 break;
Avi Kivity6aa8b732006-12-10 02:21:36 -08001707 default:
Carsten Otte043405e2007-10-10 17:16:19 +02001708 return kvm_arch_dev_ioctl(filp, ioctl, arg);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001709 }
1710out:
1711 return r;
1712}
1713
Avi Kivity6aa8b732006-12-10 02:21:36 -08001714static struct file_operations kvm_chardev_ops = {
Avi Kivity6aa8b732006-12-10 02:21:36 -08001715 .unlocked_ioctl = kvm_dev_ioctl,
1716 .compat_ioctl = kvm_dev_ioctl,
Avi Kivity6aa8b732006-12-10 02:21:36 -08001717};
1718
1719static struct miscdevice kvm_dev = {
Avi Kivitybbe44322007-03-04 13:27:36 +02001720 KVM_MINOR,
Avi Kivity6aa8b732006-12-10 02:21:36 -08001721 "kvm",
1722 &kvm_chardev_ops,
1723};
1724
Avi Kivity1b6c0162007-05-24 13:03:52 +03001725static void hardware_enable(void *junk)
1726{
1727 int cpu = raw_smp_processor_id();
Alexander Graf10474ae2009-09-15 11:37:46 +02001728 int r;
Avi Kivity1b6c0162007-05-24 13:03:52 +03001729
Rusty Russell7f59f492008-12-07 21:25:45 +10301730 if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
Avi Kivity1b6c0162007-05-24 13:03:52 +03001731 return;
Alexander Graf10474ae2009-09-15 11:37:46 +02001732
Rusty Russell7f59f492008-12-07 21:25:45 +10301733 cpumask_set_cpu(cpu, cpus_hardware_enabled);
Alexander Graf10474ae2009-09-15 11:37:46 +02001734
1735 r = kvm_arch_hardware_enable(NULL);
1736
1737 if (r) {
1738 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
1739 atomic_inc(&hardware_enable_failed);
1740 printk(KERN_INFO "kvm: enabling virtualization on "
1741 "CPU%d failed\n", cpu);
1742 }
Avi Kivity1b6c0162007-05-24 13:03:52 +03001743}
1744
1745static void hardware_disable(void *junk)
1746{
1747 int cpu = raw_smp_processor_id();
1748
Rusty Russell7f59f492008-12-07 21:25:45 +10301749 if (!cpumask_test_cpu(cpu, cpus_hardware_enabled))
Avi Kivity1b6c0162007-05-24 13:03:52 +03001750 return;
Rusty Russell7f59f492008-12-07 21:25:45 +10301751 cpumask_clear_cpu(cpu, cpus_hardware_enabled);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08001752 kvm_arch_hardware_disable(NULL);
Avi Kivity1b6c0162007-05-24 13:03:52 +03001753}
1754
Alexander Graf10474ae2009-09-15 11:37:46 +02001755static void hardware_disable_all_nolock(void)
1756{
1757 BUG_ON(!kvm_usage_count);
1758
1759 kvm_usage_count--;
1760 if (!kvm_usage_count)
1761 on_each_cpu(hardware_disable, NULL, 1);
1762}
1763
1764static void hardware_disable_all(void)
1765{
1766 spin_lock(&kvm_lock);
1767 hardware_disable_all_nolock();
1768 spin_unlock(&kvm_lock);
1769}
1770
1771static int hardware_enable_all(void)
1772{
1773 int r = 0;
1774
1775 spin_lock(&kvm_lock);
1776
1777 kvm_usage_count++;
1778 if (kvm_usage_count == 1) {
1779 atomic_set(&hardware_enable_failed, 0);
1780 on_each_cpu(hardware_enable, NULL, 1);
1781
1782 if (atomic_read(&hardware_enable_failed)) {
1783 hardware_disable_all_nolock();
1784 r = -EBUSY;
1785 }
1786 }
1787
1788 spin_unlock(&kvm_lock);
1789
1790 return r;
1791}
1792
Avi Kivity774c47f2007-02-12 00:54:47 -08001793static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
1794 void *v)
1795{
1796 int cpu = (long)v;
1797
Alexander Graf10474ae2009-09-15 11:37:46 +02001798 if (!kvm_usage_count)
1799 return NOTIFY_OK;
1800
Avi Kivity1a6f4d72007-11-11 18:37:32 +02001801 val &= ~CPU_TASKS_FROZEN;
Avi Kivity774c47f2007-02-12 00:54:47 -08001802 switch (val) {
Avi Kivitycec9ad22007-05-24 13:11:41 +03001803 case CPU_DYING:
Avi Kivity6ec8a852007-08-19 15:57:26 +03001804 printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n",
1805 cpu);
1806 hardware_disable(NULL);
1807 break;
Avi Kivity774c47f2007-02-12 00:54:47 -08001808 case CPU_UP_CANCELED:
Jeremy Katz43934a32007-02-19 14:37:46 +02001809 printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n",
1810 cpu);
Jens Axboe8691e5a2008-06-06 11:18:06 +02001811 smp_call_function_single(cpu, hardware_disable, NULL, 1);
Avi Kivity774c47f2007-02-12 00:54:47 -08001812 break;
Jeremy Katz43934a32007-02-19 14:37:46 +02001813 case CPU_ONLINE:
1814 printk(KERN_INFO "kvm: enabling virtualization on CPU%d\n",
1815 cpu);
Jens Axboe8691e5a2008-06-06 11:18:06 +02001816 smp_call_function_single(cpu, hardware_enable, NULL, 1);
Avi Kivity774c47f2007-02-12 00:54:47 -08001817 break;
1818 }
1819 return NOTIFY_OK;
1820}
1821
Avi Kivity4ecac3f2008-05-13 13:23:38 +03001822
1823asmlinkage void kvm_handle_fault_on_reboot(void)
1824{
1825 if (kvm_rebooting)
1826 /* spin while reset goes on */
1827 while (true)
1828 ;
1829 /* Fault while not rebooting. We want the trace. */
1830 BUG();
1831}
1832EXPORT_SYMBOL_GPL(kvm_handle_fault_on_reboot);
1833
Rusty Russell9a2b85c2007-07-17 23:17:55 +10001834static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
Mike Dayd77c26f2007-10-08 09:02:08 -04001835 void *v)
Rusty Russell9a2b85c2007-07-17 23:17:55 +10001836{
Sheng Yang8e1c1812009-04-29 11:09:04 +08001837 /*
1838 * Some (well, at least mine) BIOSes hang on reboot if
1839 * in vmx root mode.
1840 *
1841 * And Intel TXT required VMX off for all cpu when system shutdown.
1842 */
1843 printk(KERN_INFO "kvm: exiting hardware virtualization\n");
1844 kvm_rebooting = true;
1845 on_each_cpu(hardware_disable, NULL, 1);
Rusty Russell9a2b85c2007-07-17 23:17:55 +10001846 return NOTIFY_OK;
1847}
1848
1849static struct notifier_block kvm_reboot_notifier = {
1850 .notifier_call = kvm_reboot,
1851 .priority = 0,
1852};
1853
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001854void kvm_io_bus_init(struct kvm_io_bus *bus)
1855{
1856 memset(bus, 0, sizeof(*bus));
1857}
1858
1859void kvm_io_bus_destroy(struct kvm_io_bus *bus)
1860{
1861 int i;
1862
1863 for (i = 0; i < bus->dev_count; i++) {
1864 struct kvm_io_device *pos = bus->devs[i];
1865
1866 kvm_iodevice_destructor(pos);
1867 }
1868}
1869
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03001870/* kvm_io_bus_write - called under kvm->slots_lock */
1871int kvm_io_bus_write(struct kvm_io_bus *bus, gpa_t addr,
1872 int len, const void *val)
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001873{
1874 int i;
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03001875 for (i = 0; i < bus->dev_count; i++)
1876 if (!kvm_iodevice_write(bus->devs[i], addr, len, val))
1877 return 0;
1878 return -EOPNOTSUPP;
1879}
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001880
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03001881/* kvm_io_bus_read - called under kvm->slots_lock */
1882int kvm_io_bus_read(struct kvm_io_bus *bus, gpa_t addr, int len, void *val)
1883{
1884 int i;
1885 for (i = 0; i < bus->dev_count; i++)
1886 if (!kvm_iodevice_read(bus->devs[i], addr, len, val))
1887 return 0;
1888 return -EOPNOTSUPP;
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001889}
1890
Gregory Haskins090b7af2009-07-07 17:08:44 -04001891int kvm_io_bus_register_dev(struct kvm *kvm, struct kvm_io_bus *bus,
Michael S. Tsirkin6c474692009-06-29 22:24:26 +03001892 struct kvm_io_device *dev)
1893{
Gregory Haskins090b7af2009-07-07 17:08:44 -04001894 int ret;
1895
Michael S. Tsirkin6c474692009-06-29 22:24:26 +03001896 down_write(&kvm->slots_lock);
Gregory Haskins090b7af2009-07-07 17:08:44 -04001897 ret = __kvm_io_bus_register_dev(bus, dev);
1898 up_write(&kvm->slots_lock);
1899
1900 return ret;
1901}
1902
1903/* An unlocked version. Caller must have write lock on slots_lock. */
1904int __kvm_io_bus_register_dev(struct kvm_io_bus *bus,
1905 struct kvm_io_device *dev)
1906{
1907 if (bus->dev_count > NR_IOBUS_DEVS-1)
1908 return -ENOSPC;
1909
1910 bus->devs[bus->dev_count++] = dev;
1911
1912 return 0;
1913}
1914
1915void kvm_io_bus_unregister_dev(struct kvm *kvm,
1916 struct kvm_io_bus *bus,
1917 struct kvm_io_device *dev)
1918{
1919 down_write(&kvm->slots_lock);
1920 __kvm_io_bus_unregister_dev(bus, dev);
Michael S. Tsirkin6c474692009-06-29 22:24:26 +03001921 up_write(&kvm->slots_lock);
1922}
1923
1924/* An unlocked version. Caller must have write lock on slots_lock. */
Gregory Haskins090b7af2009-07-07 17:08:44 -04001925void __kvm_io_bus_unregister_dev(struct kvm_io_bus *bus,
1926 struct kvm_io_device *dev)
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001927{
Gregory Haskins090b7af2009-07-07 17:08:44 -04001928 int i;
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001929
Gregory Haskins090b7af2009-07-07 17:08:44 -04001930 for (i = 0; i < bus->dev_count; i++)
1931 if (bus->devs[i] == dev) {
1932 bus->devs[i] = bus->devs[--bus->dev_count];
1933 break;
1934 }
Gregory Haskins2eeb2e92007-05-31 14:08:53 -04001935}
1936
Avi Kivity774c47f2007-02-12 00:54:47 -08001937static struct notifier_block kvm_cpu_notifier = {
1938 .notifier_call = kvm_cpu_hotplug,
1939 .priority = 20, /* must be > scheduler priority */
1940};
1941
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001942static int vm_stat_get(void *_offset, u64 *val)
Avi Kivityba1389b2007-11-18 16:24:12 +02001943{
1944 unsigned offset = (long)_offset;
Avi Kivityba1389b2007-11-18 16:24:12 +02001945 struct kvm *kvm;
1946
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001947 *val = 0;
Avi Kivityba1389b2007-11-18 16:24:12 +02001948 spin_lock(&kvm_lock);
1949 list_for_each_entry(kvm, &vm_list, vm_list)
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001950 *val += *(u32 *)((void *)kvm + offset);
Avi Kivityba1389b2007-11-18 16:24:12 +02001951 spin_unlock(&kvm_lock);
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001952 return 0;
Avi Kivityba1389b2007-11-18 16:24:12 +02001953}
1954
1955DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n");
1956
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001957static int vcpu_stat_get(void *_offset, u64 *val)
Avi Kivity1165f5f2007-04-19 17:27:43 +03001958{
1959 unsigned offset = (long)_offset;
Avi Kivity1165f5f2007-04-19 17:27:43 +03001960 struct kvm *kvm;
1961 struct kvm_vcpu *vcpu;
1962 int i;
1963
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001964 *val = 0;
Avi Kivity1165f5f2007-04-19 17:27:43 +03001965 spin_lock(&kvm_lock);
1966 list_for_each_entry(kvm, &vm_list, vm_list)
Gleb Natapov988a2ca2009-06-09 15:56:29 +03001967 kvm_for_each_vcpu(i, vcpu, kvm)
1968 *val += *(u32 *)((void *)vcpu + offset);
1969
Avi Kivity1165f5f2007-04-19 17:27:43 +03001970 spin_unlock(&kvm_lock);
Christoph Hellwig8b88b092008-02-08 04:20:26 -08001971 return 0;
Avi Kivity1165f5f2007-04-19 17:27:43 +03001972}
1973
Avi Kivityba1389b2007-11-18 16:24:12 +02001974DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
1975
Alexey Dobriyan828c0952009-10-01 15:43:56 -07001976static const struct file_operations *stat_fops[] = {
Avi Kivityba1389b2007-11-18 16:24:12 +02001977 [KVM_STAT_VCPU] = &vcpu_stat_fops,
1978 [KVM_STAT_VM] = &vm_stat_fops,
1979};
Avi Kivity1165f5f2007-04-19 17:27:43 +03001980
Zhang Xiantaoa16b0432007-11-16 14:38:21 +08001981static void kvm_init_debug(void)
Avi Kivity6aa8b732006-12-10 02:21:36 -08001982{
1983 struct kvm_stats_debugfs_item *p;
1984
Hollis Blanchard76f7c872008-04-15 16:05:42 -05001985 kvm_debugfs_dir = debugfs_create_dir("kvm", NULL);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001986 for (p = debugfs_entries; p->name; ++p)
Hollis Blanchard76f7c872008-04-15 16:05:42 -05001987 p->dentry = debugfs_create_file(p->name, 0444, kvm_debugfs_dir,
Avi Kivity1165f5f2007-04-19 17:27:43 +03001988 (void *)(long)p->offset,
Avi Kivityba1389b2007-11-18 16:24:12 +02001989 stat_fops[p->kind]);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001990}
1991
1992static void kvm_exit_debug(void)
1993{
1994 struct kvm_stats_debugfs_item *p;
1995
1996 for (p = debugfs_entries; p->name; ++p)
1997 debugfs_remove(p->dentry);
Hollis Blanchard76f7c872008-04-15 16:05:42 -05001998 debugfs_remove(kvm_debugfs_dir);
Avi Kivity6aa8b732006-12-10 02:21:36 -08001999}
2000
Avi Kivity59ae6c62007-02-12 00:54:48 -08002001static int kvm_suspend(struct sys_device *dev, pm_message_t state)
2002{
Alexander Graf10474ae2009-09-15 11:37:46 +02002003 if (kvm_usage_count)
2004 hardware_disable(NULL);
Avi Kivity59ae6c62007-02-12 00:54:48 -08002005 return 0;
2006}
2007
2008static int kvm_resume(struct sys_device *dev)
2009{
Alexander Graf10474ae2009-09-15 11:37:46 +02002010 if (kvm_usage_count)
2011 hardware_enable(NULL);
Avi Kivity59ae6c62007-02-12 00:54:48 -08002012 return 0;
2013}
2014
2015static struct sysdev_class kvm_sysdev_class = {
Kay Sieversaf5ca3f2007-12-20 02:09:39 +01002016 .name = "kvm",
Avi Kivity59ae6c62007-02-12 00:54:48 -08002017 .suspend = kvm_suspend,
2018 .resume = kvm_resume,
2019};
2020
2021static struct sys_device kvm_sysdev = {
2022 .id = 0,
2023 .cls = &kvm_sysdev_class,
2024};
2025
Izik Eiduscea7bb22007-10-17 19:17:48 +02002026struct page *bad_page;
Anthony Liguori35149e22008-04-02 14:46:56 -05002027pfn_t bad_pfn;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002028
Avi Kivity15ad7142007-07-11 18:17:21 +03002029static inline
2030struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn)
2031{
2032 return container_of(pn, struct kvm_vcpu, preempt_notifier);
2033}
2034
2035static void kvm_sched_in(struct preempt_notifier *pn, int cpu)
2036{
2037 struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
2038
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002039 kvm_arch_vcpu_load(vcpu, cpu);
Avi Kivity15ad7142007-07-11 18:17:21 +03002040}
2041
2042static void kvm_sched_out(struct preempt_notifier *pn,
2043 struct task_struct *next)
2044{
2045 struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn);
2046
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002047 kvm_arch_vcpu_put(vcpu);
Avi Kivity15ad7142007-07-11 18:17:21 +03002048}
2049
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08002050int kvm_init(void *opaque, unsigned int vcpu_size,
Rusty Russellc16f8622007-07-30 21:12:19 +10002051 struct module *module)
Avi Kivity6aa8b732006-12-10 02:21:36 -08002052{
2053 int r;
Yang, Sheng002c7f72007-07-31 14:23:01 +03002054 int cpu;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002055
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08002056 r = kvm_arch_init(opaque);
2057 if (r)
Zhang Xiantaod23087842007-11-29 15:35:39 +08002058 goto out_fail;
Zhang Xiantaocb498ea2007-11-14 20:39:31 +08002059
2060 bad_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2061
2062 if (bad_page == NULL) {
2063 r = -ENOMEM;
2064 goto out;
2065 }
2066
Anthony Liguori35149e22008-04-02 14:46:56 -05002067 bad_pfn = page_to_pfn(bad_page);
2068
Avi Kivity8437a612009-06-06 14:52:35 -07002069 if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
Rusty Russell7f59f492008-12-07 21:25:45 +10302070 r = -ENOMEM;
2071 goto out_free_0;
2072 }
2073
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002074 r = kvm_arch_hardware_setup();
Avi Kivity6aa8b732006-12-10 02:21:36 -08002075 if (r < 0)
Rusty Russell7f59f492008-12-07 21:25:45 +10302076 goto out_free_0a;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002077
Yang, Sheng002c7f72007-07-31 14:23:01 +03002078 for_each_online_cpu(cpu) {
2079 smp_call_function_single(cpu,
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002080 kvm_arch_check_processor_compat,
Jens Axboe8691e5a2008-06-06 11:18:06 +02002081 &r, 1);
Yang, Sheng002c7f72007-07-31 14:23:01 +03002082 if (r < 0)
Zhang Xiantaod23087842007-11-29 15:35:39 +08002083 goto out_free_1;
Yang, Sheng002c7f72007-07-31 14:23:01 +03002084 }
2085
Avi Kivity774c47f2007-02-12 00:54:47 -08002086 r = register_cpu_notifier(&kvm_cpu_notifier);
2087 if (r)
Zhang Xiantaod23087842007-11-29 15:35:39 +08002088 goto out_free_2;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002089 register_reboot_notifier(&kvm_reboot_notifier);
2090
Avi Kivity59ae6c62007-02-12 00:54:48 -08002091 r = sysdev_class_register(&kvm_sysdev_class);
2092 if (r)
Zhang Xiantaod23087842007-11-29 15:35:39 +08002093 goto out_free_3;
Avi Kivity59ae6c62007-02-12 00:54:48 -08002094
2095 r = sysdev_register(&kvm_sysdev);
2096 if (r)
Zhang Xiantaod23087842007-11-29 15:35:39 +08002097 goto out_free_4;
Avi Kivity59ae6c62007-02-12 00:54:48 -08002098
Rusty Russellc16f8622007-07-30 21:12:19 +10002099 /* A kmem cache lets us meet the alignment requirements of fx_save. */
2100 kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size,
Joe Perches56919c52007-11-12 20:06:51 -08002101 __alignof__(struct kvm_vcpu),
2102 0, NULL);
Rusty Russellc16f8622007-07-30 21:12:19 +10002103 if (!kvm_vcpu_cache) {
2104 r = -ENOMEM;
Zhang Xiantaod23087842007-11-29 15:35:39 +08002105 goto out_free_5;
Rusty Russellc16f8622007-07-30 21:12:19 +10002106 }
2107
Avi Kivity6aa8b732006-12-10 02:21:36 -08002108 kvm_chardev_ops.owner = module;
Christian Borntraeger3d3aab12008-12-02 11:17:32 +01002109 kvm_vm_fops.owner = module;
2110 kvm_vcpu_fops.owner = module;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002111
2112 r = misc_register(&kvm_dev);
2113 if (r) {
Mike Dayd77c26f2007-10-08 09:02:08 -04002114 printk(KERN_ERR "kvm: misc device register failed\n");
Avi Kivity6aa8b732006-12-10 02:21:36 -08002115 goto out_free;
2116 }
2117
Avi Kivity15ad7142007-07-11 18:17:21 +03002118 kvm_preempt_ops.sched_in = kvm_sched_in;
2119 kvm_preempt_ops.sched_out = kvm_sched_out;
2120
Darrick J. Wong0ea4ed82009-10-14 16:21:00 -07002121 kvm_init_debug();
2122
Avi Kivityc7addb92007-09-16 18:58:32 +02002123 return 0;
Avi Kivity6aa8b732006-12-10 02:21:36 -08002124
2125out_free:
Rusty Russellc16f8622007-07-30 21:12:19 +10002126 kmem_cache_destroy(kvm_vcpu_cache);
Zhang Xiantaod23087842007-11-29 15:35:39 +08002127out_free_5:
Avi Kivity59ae6c62007-02-12 00:54:48 -08002128 sysdev_unregister(&kvm_sysdev);
Zhang Xiantaod23087842007-11-29 15:35:39 +08002129out_free_4:
Avi Kivity59ae6c62007-02-12 00:54:48 -08002130 sysdev_class_unregister(&kvm_sysdev_class);
Zhang Xiantaod23087842007-11-29 15:35:39 +08002131out_free_3:
Avi Kivity6aa8b732006-12-10 02:21:36 -08002132 unregister_reboot_notifier(&kvm_reboot_notifier);
Avi Kivity774c47f2007-02-12 00:54:47 -08002133 unregister_cpu_notifier(&kvm_cpu_notifier);
Zhang Xiantaod23087842007-11-29 15:35:39 +08002134out_free_2:
Zhang Xiantaod23087842007-11-29 15:35:39 +08002135out_free_1:
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002136 kvm_arch_hardware_unsetup();
Rusty Russell7f59f492008-12-07 21:25:45 +10302137out_free_0a:
2138 free_cpumask_var(cpus_hardware_enabled);
Zhang Xiantaod23087842007-11-29 15:35:39 +08002139out_free_0:
2140 __free_page(bad_page);
Avi Kivityca45aaa2007-03-01 19:21:03 +02002141out:
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08002142 kvm_arch_exit();
Zhang Xiantaod23087842007-11-29 15:35:39 +08002143out_fail:
Avi Kivity6aa8b732006-12-10 02:21:36 -08002144 return r;
2145}
Zhang Xiantaocb498ea2007-11-14 20:39:31 +08002146EXPORT_SYMBOL_GPL(kvm_init);
Avi Kivity6aa8b732006-12-10 02:21:36 -08002147
Zhang Xiantaocb498ea2007-11-14 20:39:31 +08002148void kvm_exit(void)
Avi Kivity6aa8b732006-12-10 02:21:36 -08002149{
Marcelo Tosatti229456f2009-06-17 09:22:14 -03002150 tracepoint_synchronize_unregister();
Darrick J. Wong0ea4ed82009-10-14 16:21:00 -07002151 kvm_exit_debug();
Avi Kivity6aa8b732006-12-10 02:21:36 -08002152 misc_deregister(&kvm_dev);
Rusty Russellc16f8622007-07-30 21:12:19 +10002153 kmem_cache_destroy(kvm_vcpu_cache);
Avi Kivity59ae6c62007-02-12 00:54:48 -08002154 sysdev_unregister(&kvm_sysdev);
2155 sysdev_class_unregister(&kvm_sysdev_class);
Avi Kivity6aa8b732006-12-10 02:21:36 -08002156 unregister_reboot_notifier(&kvm_reboot_notifier);
Avi Kivity59ae6c62007-02-12 00:54:48 -08002157 unregister_cpu_notifier(&kvm_cpu_notifier);
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002158 on_each_cpu(hardware_disable, NULL, 1);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08002159 kvm_arch_hardware_unsetup();
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08002160 kvm_arch_exit();
Rusty Russell7f59f492008-12-07 21:25:45 +10302161 free_cpumask_var(cpus_hardware_enabled);
Izik Eiduscea7bb22007-10-17 19:17:48 +02002162 __free_page(bad_page);
Avi Kivity6aa8b732006-12-10 02:21:36 -08002163}
Zhang Xiantaocb498ea2007-11-14 20:39:31 +08002164EXPORT_SYMBOL_GPL(kvm_exit);