Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1 | /* |
| 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. |
Nicolas Kaiser | 9611c18 | 2010-10-06 14:23:22 +0200 | [diff] [blame] | 8 | * Copyright 2010 Red Hat, Inc. and/or its affiliates. |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 9 | * |
| 10 | * Authors: |
| 11 | * Avi Kivity <avi@qumranet.com> |
| 12 | * Yaniv Kamay <yaniv@qumranet.com> |
| 13 | * |
| 14 | * This work is licensed under the terms of the GNU GPL, version 2. See |
| 15 | * the COPYING file in the top-level directory. |
| 16 | * |
| 17 | */ |
| 18 | |
Hollis Blanchard | e217402 | 2007-12-03 15:30:24 -0600 | [diff] [blame] | 19 | #include "iodev.h" |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 20 | |
Avi Kivity | edf8841 | 2007-12-16 11:02:48 +0200 | [diff] [blame] | 21 | #include <linux/kvm_host.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 22 | #include <linux/kvm.h> |
| 23 | #include <linux/module.h> |
| 24 | #include <linux/errno.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 25 | #include <linux/percpu.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 26 | #include <linux/mm.h> |
| 27 | #include <linux/miscdevice.h> |
| 28 | #include <linux/vmalloc.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 29 | #include <linux/reboot.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 30 | #include <linux/debugfs.h> |
| 31 | #include <linux/highmem.h> |
| 32 | #include <linux/file.h> |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 33 | #include <linux/sysdev.h> |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 34 | #include <linux/cpu.h> |
Alexey Dobriyan | e8edc6e | 2007-05-21 01:22:52 +0400 | [diff] [blame] | 35 | #include <linux/sched.h> |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 36 | #include <linux/cpumask.h> |
| 37 | #include <linux/smp.h> |
Avi Kivity | d6d2816 | 2007-06-28 08:38:16 -0400 | [diff] [blame] | 38 | #include <linux/anon_inodes.h> |
Avi Kivity | 04d2cc7 | 2007-09-10 18:10:54 +0300 | [diff] [blame] | 39 | #include <linux/profile.h> |
Anthony Liguori | 7aa81cc | 2007-09-17 14:57:50 -0500 | [diff] [blame] | 40 | #include <linux/kvm_para.h> |
Izik Eidus | 6fc138d | 2007-10-09 19:20:39 +0200 | [diff] [blame] | 41 | #include <linux/pagemap.h> |
Anthony Liguori | 8d4e128 | 2007-10-18 09:59:34 -0500 | [diff] [blame] | 42 | #include <linux/mman.h> |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 43 | #include <linux/swap.h> |
Sheng Yang | e56d532 | 2009-03-12 21:45:39 +0800 | [diff] [blame] | 44 | #include <linux/bitops.h> |
Marcelo Tosatti | 547de29 | 2009-05-07 17:55:13 -0300 | [diff] [blame] | 45 | #include <linux/spinlock.h> |
Arnd Bergmann | 6ff5894 | 2009-10-22 14:19:27 +0200 | [diff] [blame] | 46 | #include <linux/compat.h> |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 47 | #include <linux/srcu.h> |
Joerg Roedel | 8f0b1ab | 2010-01-28 12:37:56 +0100 | [diff] [blame] | 48 | #include <linux/hugetlb.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 49 | #include <linux/slab.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 50 | |
Avi Kivity | e495606 | 2007-06-28 14:15:57 -0400 | [diff] [blame] | 51 | #include <asm/processor.h> |
Avi Kivity | e495606 | 2007-06-28 14:15:57 -0400 | [diff] [blame] | 52 | #include <asm/io.h> |
| 53 | #include <asm/uaccess.h> |
Izik Eidus | 3e021bf | 2007-11-19 11:16:57 +0200 | [diff] [blame] | 54 | #include <asm/pgtable.h> |
Alexander Graf | c8240bd | 2009-10-30 05:47:26 +0000 | [diff] [blame] | 55 | #include <asm-generic/bitops/le.h> |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 56 | |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 57 | #include "coalesced_mmio.h" |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 58 | #include "async_pf.h" |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 59 | |
Marcelo Tosatti | 229456f | 2009-06-17 09:22:14 -0300 | [diff] [blame] | 60 | #define CREATE_TRACE_POINTS |
| 61 | #include <trace/events/kvm.h> |
| 62 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 63 | MODULE_AUTHOR("Qumranet"); |
| 64 | MODULE_LICENSE("GPL"); |
| 65 | |
Marcelo Tosatti | fa40a82 | 2009-06-04 15:08:24 -0300 | [diff] [blame] | 66 | /* |
| 67 | * Ordering of locks: |
| 68 | * |
Sheng Yang | fae3a35 | 2009-12-15 10:28:07 +0800 | [diff] [blame] | 69 | * kvm->lock --> kvm->slots_lock --> kvm->irq_lock |
Marcelo Tosatti | fa40a82 | 2009-06-04 15:08:24 -0300 | [diff] [blame] | 70 | */ |
| 71 | |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 72 | DEFINE_RAW_SPINLOCK(kvm_lock); |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 73 | LIST_HEAD(vm_list); |
Avi Kivity | 133de90 | 2007-02-12 00:54:44 -0800 | [diff] [blame] | 74 | |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 75 | static cpumask_var_t cpus_hardware_enabled; |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 76 | static int kvm_usage_count = 0; |
| 77 | static atomic_t hardware_enable_failed; |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 78 | |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 79 | struct kmem_cache *kvm_vcpu_cache; |
| 80 | EXPORT_SYMBOL_GPL(kvm_vcpu_cache); |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 81 | |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 82 | static __read_mostly struct preempt_ops kvm_preempt_ops; |
| 83 | |
Hollis Blanchard | 76f7c87 | 2008-04-15 16:05:42 -0500 | [diff] [blame] | 84 | struct dentry *kvm_debugfs_dir; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 85 | |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 86 | static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl, |
| 87 | unsigned long arg); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 88 | static int hardware_enable_all(void); |
| 89 | static void hardware_disable_all(void); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 90 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 91 | static void kvm_io_bus_destroy(struct kvm_io_bus *bus); |
| 92 | |
Avi Kivity | b7c4145 | 2010-12-02 17:52:50 +0200 | [diff] [blame] | 93 | bool kvm_rebooting; |
| 94 | EXPORT_SYMBOL_GPL(kvm_rebooting); |
Avi Kivity | 4ecac3f | 2008-05-13 13:23:38 +0300 | [diff] [blame] | 95 | |
Marcelo Tosatti | 54dee99 | 2009-06-11 12:07:44 -0300 | [diff] [blame] | 96 | static bool largepages_enabled = true; |
| 97 | |
Gleb Natapov | fa7bff8 | 2010-07-07 20:16:44 +0300 | [diff] [blame] | 98 | static struct page *hwpoison_page; |
| 99 | static pfn_t hwpoison_pfn; |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 100 | |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 101 | static struct page *fault_page; |
| 102 | static pfn_t fault_pfn; |
| 103 | |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 104 | inline int kvm_is_mmio_pfn(pfn_t pfn) |
Ben-Ami Yassour | cbff90a | 2008-07-28 19:26:24 +0300 | [diff] [blame] | 105 | { |
Joerg Roedel | fc5659c | 2009-02-18 14:08:58 +0100 | [diff] [blame] | 106 | if (pfn_valid(pfn)) { |
Andrea Arcangeli | 22e5c47 | 2011-01-13 15:47:20 -0800 | [diff] [blame] | 107 | int reserved; |
Andrea Arcangeli | 936a5fe | 2011-01-13 15:46:48 -0800 | [diff] [blame] | 108 | struct page *tail = pfn_to_page(pfn); |
Andrea Arcangeli | 22e5c47 | 2011-01-13 15:47:20 -0800 | [diff] [blame] | 109 | struct page *head = compound_trans_head(tail); |
| 110 | reserved = PageReserved(head); |
Andrea Arcangeli | 936a5fe | 2011-01-13 15:46:48 -0800 | [diff] [blame] | 111 | if (head != tail) { |
Andrea Arcangeli | 936a5fe | 2011-01-13 15:46:48 -0800 | [diff] [blame] | 112 | /* |
Andrea Arcangeli | 22e5c47 | 2011-01-13 15:47:20 -0800 | [diff] [blame] | 113 | * "head" is not a dangling pointer |
| 114 | * (compound_trans_head takes care of that) |
| 115 | * but the hugepage may have been splitted |
| 116 | * from under us (and we may not hold a |
| 117 | * reference count on the head page so it can |
| 118 | * be reused before we run PageReferenced), so |
| 119 | * we've to check PageTail before returning |
| 120 | * what we just read. |
Andrea Arcangeli | 936a5fe | 2011-01-13 15:46:48 -0800 | [diff] [blame] | 121 | */ |
Andrea Arcangeli | 22e5c47 | 2011-01-13 15:47:20 -0800 | [diff] [blame] | 122 | smp_rmb(); |
| 123 | if (PageTail(tail)) |
| 124 | return reserved; |
Andrea Arcangeli | 936a5fe | 2011-01-13 15:46:48 -0800 | [diff] [blame] | 125 | } |
| 126 | return PageReserved(tail); |
Joerg Roedel | fc5659c | 2009-02-18 14:08:58 +0100 | [diff] [blame] | 127 | } |
Ben-Ami Yassour | cbff90a | 2008-07-28 19:26:24 +0300 | [diff] [blame] | 128 | |
| 129 | return true; |
| 130 | } |
| 131 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 132 | /* |
| 133 | * Switches to specified vcpu, until a matching vcpu_put() |
| 134 | */ |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 135 | void vcpu_load(struct kvm_vcpu *vcpu) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 136 | { |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 137 | int cpu; |
| 138 | |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 139 | mutex_lock(&vcpu->mutex); |
Rik van Riel | 34bb10b | 2011-02-01 09:52:41 -0500 | [diff] [blame] | 140 | if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) { |
| 141 | /* The thread running this VCPU changed. */ |
| 142 | struct pid *oldpid = vcpu->pid; |
| 143 | struct pid *newpid = get_task_pid(current, PIDTYPE_PID); |
| 144 | rcu_assign_pointer(vcpu->pid, newpid); |
| 145 | synchronize_rcu(); |
| 146 | put_pid(oldpid); |
| 147 | } |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 148 | cpu = get_cpu(); |
| 149 | preempt_notifier_register(&vcpu->preempt_notifier); |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 150 | kvm_arch_vcpu_load(vcpu, cpu); |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 151 | put_cpu(); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 152 | } |
| 153 | |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 154 | void vcpu_put(struct kvm_vcpu *vcpu) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 155 | { |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 156 | preempt_disable(); |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 157 | kvm_arch_vcpu_put(vcpu); |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 158 | preempt_notifier_unregister(&vcpu->preempt_notifier); |
| 159 | preempt_enable(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 160 | mutex_unlock(&vcpu->mutex); |
| 161 | } |
| 162 | |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 163 | static void ack_flush(void *_completed) |
| 164 | { |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 165 | } |
| 166 | |
Rusty Russell | 4984689 | 2008-12-08 20:26:24 +1030 | [diff] [blame] | 167 | static bool make_all_cpus_request(struct kvm *kvm, unsigned int req) |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 168 | { |
Avi Kivity | 597a5f5 | 2008-07-20 14:24:22 +0300 | [diff] [blame] | 169 | int i, cpu, me; |
Rusty Russell | 6ef7a1b | 2008-12-08 20:28:04 +1030 | [diff] [blame] | 170 | cpumask_var_t cpus; |
| 171 | bool called = true; |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 172 | struct kvm_vcpu *vcpu; |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 173 | |
Li Zefan | 79f5599 | 2009-06-15 14:58:26 +0800 | [diff] [blame] | 174 | zalloc_cpumask_var(&cpus, GFP_ATOMIC); |
Rusty Russell | 6ef7a1b | 2008-12-08 20:28:04 +1030 | [diff] [blame] | 175 | |
Xiao Guangrong | 3cba413 | 2011-01-12 15:41:22 +0800 | [diff] [blame] | 176 | me = get_cpu(); |
Gleb Natapov | 988a2ca | 2009-06-09 15:56:29 +0300 | [diff] [blame] | 177 | kvm_for_each_vcpu(i, vcpu, kvm) { |
Xiao Guangrong | 3cba413 | 2011-01-12 15:41:22 +0800 | [diff] [blame] | 178 | kvm_make_request(req, vcpu); |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 179 | cpu = vcpu->cpu; |
Xiao Guangrong | 6b7e2d0 | 2011-01-12 15:40:31 +0800 | [diff] [blame] | 180 | |
| 181 | /* Set ->requests bit before we read ->mode */ |
| 182 | smp_mb(); |
| 183 | |
| 184 | if (cpus != NULL && cpu != -1 && cpu != me && |
| 185 | kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE) |
Rusty Russell | 6ef7a1b | 2008-12-08 20:28:04 +1030 | [diff] [blame] | 186 | cpumask_set_cpu(cpu, cpus); |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 187 | } |
Rusty Russell | 6ef7a1b | 2008-12-08 20:28:04 +1030 | [diff] [blame] | 188 | if (unlikely(cpus == NULL)) |
| 189 | smp_call_function_many(cpu_online_mask, ack_flush, NULL, 1); |
| 190 | else if (!cpumask_empty(cpus)) |
| 191 | smp_call_function_many(cpus, ack_flush, NULL, 1); |
| 192 | else |
| 193 | called = false; |
Xiao Guangrong | 3cba413 | 2011-01-12 15:41:22 +0800 | [diff] [blame] | 194 | put_cpu(); |
Rusty Russell | 6ef7a1b | 2008-12-08 20:28:04 +1030 | [diff] [blame] | 195 | free_cpumask_var(cpus); |
Rusty Russell | 4984689 | 2008-12-08 20:26:24 +1030 | [diff] [blame] | 196 | return called; |
| 197 | } |
| 198 | |
| 199 | void kvm_flush_remote_tlbs(struct kvm *kvm) |
| 200 | { |
Xiao Guangrong | a4ee1ca | 2010-11-23 11:13:00 +0800 | [diff] [blame] | 201 | int dirty_count = kvm->tlbs_dirty; |
| 202 | |
| 203 | smp_mb(); |
Rusty Russell | 4984689 | 2008-12-08 20:26:24 +1030 | [diff] [blame] | 204 | if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH)) |
| 205 | ++kvm->stat.remote_tlb_flush; |
Xiao Guangrong | a4ee1ca | 2010-11-23 11:13:00 +0800 | [diff] [blame] | 206 | cmpxchg(&kvm->tlbs_dirty, dirty_count, 0); |
Avi Kivity | d9e368d | 2007-06-07 19:18:30 +0300 | [diff] [blame] | 207 | } |
| 208 | |
Marcelo Tosatti | 2e53d63 | 2008-02-20 14:47:24 -0500 | [diff] [blame] | 209 | void kvm_reload_remote_mmus(struct kvm *kvm) |
| 210 | { |
Rusty Russell | 4984689 | 2008-12-08 20:26:24 +1030 | [diff] [blame] | 211 | make_all_cpus_request(kvm, KVM_REQ_MMU_RELOAD); |
Marcelo Tosatti | 2e53d63 | 2008-02-20 14:47:24 -0500 | [diff] [blame] | 212 | } |
| 213 | |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 214 | int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id) |
| 215 | { |
| 216 | struct page *page; |
| 217 | int r; |
| 218 | |
| 219 | mutex_init(&vcpu->mutex); |
| 220 | vcpu->cpu = -1; |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 221 | vcpu->kvm = kvm; |
| 222 | vcpu->vcpu_id = id; |
Rik van Riel | 34bb10b | 2011-02-01 09:52:41 -0500 | [diff] [blame] | 223 | vcpu->pid = NULL; |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 224 | init_waitqueue_head(&vcpu->wq); |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 225 | kvm_async_pf_vcpu_init(vcpu); |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 226 | |
| 227 | page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 228 | if (!page) { |
| 229 | r = -ENOMEM; |
| 230 | goto fail; |
| 231 | } |
| 232 | vcpu->run = page_address(page); |
| 233 | |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 234 | r = kvm_arch_vcpu_init(vcpu); |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 235 | if (r < 0) |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 236 | goto fail_free_run; |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 237 | return 0; |
| 238 | |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 239 | fail_free_run: |
| 240 | free_page((unsigned long)vcpu->run); |
| 241 | fail: |
Rusty Russell | 76fafa5 | 2007-10-08 10:50:48 +1000 | [diff] [blame] | 242 | return r; |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 243 | } |
| 244 | EXPORT_SYMBOL_GPL(kvm_vcpu_init); |
| 245 | |
| 246 | void kvm_vcpu_uninit(struct kvm_vcpu *vcpu) |
| 247 | { |
Rik van Riel | 34bb10b | 2011-02-01 09:52:41 -0500 | [diff] [blame] | 248 | put_pid(vcpu->pid); |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 249 | kvm_arch_vcpu_uninit(vcpu); |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 250 | free_page((unsigned long)vcpu->run); |
| 251 | } |
| 252 | EXPORT_SYMBOL_GPL(kvm_vcpu_uninit); |
| 253 | |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 254 | #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) |
| 255 | static inline struct kvm *mmu_notifier_to_kvm(struct mmu_notifier *mn) |
| 256 | { |
| 257 | return container_of(mn, struct kvm, mmu_notifier); |
| 258 | } |
| 259 | |
| 260 | static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn, |
| 261 | struct mm_struct *mm, |
| 262 | unsigned long address) |
| 263 | { |
| 264 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 265 | int need_tlb_flush, idx; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 266 | |
| 267 | /* |
| 268 | * When ->invalidate_page runs, the linux pte has been zapped |
| 269 | * already but the page is still allocated until |
| 270 | * ->invalidate_page returns. So if we increase the sequence |
| 271 | * here the kvm page fault will notice if the spte can't be |
| 272 | * established because the page is going to be freed. If |
| 273 | * instead the kvm page fault establishes the spte before |
| 274 | * ->invalidate_page runs, kvm_unmap_hva will release it |
| 275 | * before returning. |
| 276 | * |
| 277 | * The sequence increase only need to be seen at spin_unlock |
| 278 | * time, and not at spin_lock time. |
| 279 | * |
| 280 | * Increasing the sequence after the spin_unlock would be |
| 281 | * unsafe because the kvm page fault could then establish the |
| 282 | * pte after kvm_unmap_hva returned, without noticing the page |
| 283 | * is going to be freed. |
| 284 | */ |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 285 | idx = srcu_read_lock(&kvm->srcu); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 286 | spin_lock(&kvm->mmu_lock); |
| 287 | kvm->mmu_notifier_seq++; |
Xiao Guangrong | a4ee1ca | 2010-11-23 11:13:00 +0800 | [diff] [blame] | 288 | need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 289 | spin_unlock(&kvm->mmu_lock); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 290 | srcu_read_unlock(&kvm->srcu, idx); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 291 | |
| 292 | /* we've to flush the tlb before the pages can be freed */ |
| 293 | if (need_tlb_flush) |
| 294 | kvm_flush_remote_tlbs(kvm); |
| 295 | |
| 296 | } |
| 297 | |
Izik Eidus | 3da0dd4 | 2009-09-23 21:47:18 +0300 | [diff] [blame] | 298 | static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn, |
| 299 | struct mm_struct *mm, |
| 300 | unsigned long address, |
| 301 | pte_t pte) |
| 302 | { |
| 303 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 304 | int idx; |
Izik Eidus | 3da0dd4 | 2009-09-23 21:47:18 +0300 | [diff] [blame] | 305 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 306 | idx = srcu_read_lock(&kvm->srcu); |
Izik Eidus | 3da0dd4 | 2009-09-23 21:47:18 +0300 | [diff] [blame] | 307 | spin_lock(&kvm->mmu_lock); |
| 308 | kvm->mmu_notifier_seq++; |
| 309 | kvm_set_spte_hva(kvm, address, pte); |
| 310 | spin_unlock(&kvm->mmu_lock); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 311 | srcu_read_unlock(&kvm->srcu, idx); |
Izik Eidus | 3da0dd4 | 2009-09-23 21:47:18 +0300 | [diff] [blame] | 312 | } |
| 313 | |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 314 | static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, |
| 315 | struct mm_struct *mm, |
| 316 | unsigned long start, |
| 317 | unsigned long end) |
| 318 | { |
| 319 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 320 | int need_tlb_flush = 0, idx; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 321 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 322 | idx = srcu_read_lock(&kvm->srcu); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 323 | spin_lock(&kvm->mmu_lock); |
| 324 | /* |
| 325 | * The count increase must become visible at unlock time as no |
| 326 | * spte can be established without taking the mmu_lock and |
| 327 | * count is also read inside the mmu_lock critical section. |
| 328 | */ |
| 329 | kvm->mmu_notifier_count++; |
| 330 | for (; start < end; start += PAGE_SIZE) |
| 331 | need_tlb_flush |= kvm_unmap_hva(kvm, start); |
Xiao Guangrong | a4ee1ca | 2010-11-23 11:13:00 +0800 | [diff] [blame] | 332 | need_tlb_flush |= kvm->tlbs_dirty; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 333 | spin_unlock(&kvm->mmu_lock); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 334 | srcu_read_unlock(&kvm->srcu, idx); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 335 | |
| 336 | /* we've to flush the tlb before the pages can be freed */ |
| 337 | if (need_tlb_flush) |
| 338 | kvm_flush_remote_tlbs(kvm); |
| 339 | } |
| 340 | |
| 341 | static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, |
| 342 | struct mm_struct *mm, |
| 343 | unsigned long start, |
| 344 | unsigned long end) |
| 345 | { |
| 346 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
| 347 | |
| 348 | spin_lock(&kvm->mmu_lock); |
| 349 | /* |
| 350 | * This sequence increase will notify the kvm page fault that |
| 351 | * the page that is going to be mapped in the spte could have |
| 352 | * been freed. |
| 353 | */ |
| 354 | kvm->mmu_notifier_seq++; |
| 355 | /* |
| 356 | * The above sequence increase must be visible before the |
| 357 | * below count decrease but both values are read by the kvm |
| 358 | * page fault under mmu_lock spinlock so we don't need to add |
| 359 | * a smb_wmb() here in between the two. |
| 360 | */ |
| 361 | kvm->mmu_notifier_count--; |
| 362 | spin_unlock(&kvm->mmu_lock); |
| 363 | |
| 364 | BUG_ON(kvm->mmu_notifier_count < 0); |
| 365 | } |
| 366 | |
| 367 | static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn, |
| 368 | struct mm_struct *mm, |
| 369 | unsigned long address) |
| 370 | { |
| 371 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 372 | int young, idx; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 373 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 374 | idx = srcu_read_lock(&kvm->srcu); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 375 | spin_lock(&kvm->mmu_lock); |
| 376 | young = kvm_age_hva(kvm, address); |
| 377 | spin_unlock(&kvm->mmu_lock); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 378 | srcu_read_unlock(&kvm->srcu, idx); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 379 | |
| 380 | if (young) |
| 381 | kvm_flush_remote_tlbs(kvm); |
| 382 | |
| 383 | return young; |
| 384 | } |
| 385 | |
Andrea Arcangeli | 8ee5382 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 386 | static int kvm_mmu_notifier_test_young(struct mmu_notifier *mn, |
| 387 | struct mm_struct *mm, |
| 388 | unsigned long address) |
| 389 | { |
| 390 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
| 391 | int young, idx; |
| 392 | |
| 393 | idx = srcu_read_lock(&kvm->srcu); |
| 394 | spin_lock(&kvm->mmu_lock); |
| 395 | young = kvm_test_age_hva(kvm, address); |
| 396 | spin_unlock(&kvm->mmu_lock); |
| 397 | srcu_read_unlock(&kvm->srcu, idx); |
| 398 | |
| 399 | return young; |
| 400 | } |
| 401 | |
Marcelo Tosatti | 85db06e | 2008-12-10 21:23:26 +0100 | [diff] [blame] | 402 | static void kvm_mmu_notifier_release(struct mmu_notifier *mn, |
| 403 | struct mm_struct *mm) |
| 404 | { |
| 405 | struct kvm *kvm = mmu_notifier_to_kvm(mn); |
Lai Jiangshan | eda2bed | 2010-04-20 14:29:29 +0800 | [diff] [blame] | 406 | int idx; |
| 407 | |
| 408 | idx = srcu_read_lock(&kvm->srcu); |
Marcelo Tosatti | 85db06e | 2008-12-10 21:23:26 +0100 | [diff] [blame] | 409 | kvm_arch_flush_shadow(kvm); |
Lai Jiangshan | eda2bed | 2010-04-20 14:29:29 +0800 | [diff] [blame] | 410 | srcu_read_unlock(&kvm->srcu, idx); |
Marcelo Tosatti | 85db06e | 2008-12-10 21:23:26 +0100 | [diff] [blame] | 411 | } |
| 412 | |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 413 | static const struct mmu_notifier_ops kvm_mmu_notifier_ops = { |
| 414 | .invalidate_page = kvm_mmu_notifier_invalidate_page, |
| 415 | .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start, |
| 416 | .invalidate_range_end = kvm_mmu_notifier_invalidate_range_end, |
| 417 | .clear_flush_young = kvm_mmu_notifier_clear_flush_young, |
Andrea Arcangeli | 8ee5382 | 2011-01-13 15:47:10 -0800 | [diff] [blame] | 418 | .test_young = kvm_mmu_notifier_test_young, |
Izik Eidus | 3da0dd4 | 2009-09-23 21:47:18 +0300 | [diff] [blame] | 419 | .change_pte = kvm_mmu_notifier_change_pte, |
Marcelo Tosatti | 85db06e | 2008-12-10 21:23:26 +0100 | [diff] [blame] | 420 | .release = kvm_mmu_notifier_release, |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 421 | }; |
Avi Kivity | 4c07b0a | 2009-12-20 14:54:04 +0200 | [diff] [blame] | 422 | |
| 423 | static int kvm_init_mmu_notifier(struct kvm *kvm) |
| 424 | { |
| 425 | kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops; |
| 426 | return mmu_notifier_register(&kvm->mmu_notifier, current->mm); |
| 427 | } |
| 428 | |
| 429 | #else /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */ |
| 430 | |
| 431 | static int kvm_init_mmu_notifier(struct kvm *kvm) |
| 432 | { |
| 433 | return 0; |
| 434 | } |
| 435 | |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 436 | #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */ |
| 437 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 438 | static struct kvm *kvm_create_vm(void) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 439 | { |
Jan Kiszka | d89f5ef | 2010-11-09 17:02:49 +0100 | [diff] [blame] | 440 | int r, i; |
| 441 | struct kvm *kvm = kvm_arch_alloc_vm(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 442 | |
Jan Kiszka | d89f5ef | 2010-11-09 17:02:49 +0100 | [diff] [blame] | 443 | if (!kvm) |
| 444 | return ERR_PTR(-ENOMEM); |
| 445 | |
| 446 | r = kvm_arch_init_vm(kvm); |
| 447 | if (r) |
| 448 | goto out_err_nodisable; |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 449 | |
| 450 | r = hardware_enable_all(); |
| 451 | if (r) |
| 452 | goto out_err_nodisable; |
| 453 | |
Avi Kivity | 75858a8 | 2009-01-04 17:10:50 +0200 | [diff] [blame] | 454 | #ifdef CONFIG_HAVE_KVM_IRQCHIP |
| 455 | INIT_HLIST_HEAD(&kvm->mask_notifier_list); |
Gleb Natapov | 136bdfe | 2009-08-24 11:54:23 +0300 | [diff] [blame] | 456 | INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list); |
Avi Kivity | 75858a8 | 2009-01-04 17:10:50 +0200 | [diff] [blame] | 457 | #endif |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 458 | |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 459 | r = -ENOMEM; |
| 460 | kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL); |
| 461 | if (!kvm->memslots) |
Jan Kiszka | 57e7fbe | 2010-11-09 12:42:12 +0100 | [diff] [blame] | 462 | goto out_err_nosrcu; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 463 | if (init_srcu_struct(&kvm->srcu)) |
Jan Kiszka | 57e7fbe | 2010-11-09 12:42:12 +0100 | [diff] [blame] | 464 | goto out_err_nosrcu; |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 465 | for (i = 0; i < KVM_NR_BUSES; i++) { |
| 466 | kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus), |
| 467 | GFP_KERNEL); |
Jan Kiszka | 57e7fbe | 2010-11-09 12:42:12 +0100 | [diff] [blame] | 468 | if (!kvm->buses[i]) |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 469 | goto out_err; |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 470 | } |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 471 | |
Avi Kivity | 4c07b0a | 2009-12-20 14:54:04 +0200 | [diff] [blame] | 472 | r = kvm_init_mmu_notifier(kvm); |
Jan Kiszka | 57e7fbe | 2010-11-09 12:42:12 +0100 | [diff] [blame] | 473 | if (r) |
Avi Kivity | 283d0c6 | 2009-12-20 14:25:19 +0200 | [diff] [blame] | 474 | goto out_err; |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 475 | |
Avi Kivity | 6d4e4c4 | 2007-11-21 16:41:05 +0200 | [diff] [blame] | 476 | kvm->mm = current->mm; |
| 477 | atomic_inc(&kvm->mm->mm_count); |
Marcelo Tosatti | aaee2c9 | 2007-12-20 19:18:26 -0500 | [diff] [blame] | 478 | spin_lock_init(&kvm->mmu_lock); |
Gregory Haskins | d34e6b1 | 2009-07-07 17:08:49 -0400 | [diff] [blame] | 479 | kvm_eventfd_init(kvm); |
Shaohua Li | 11ec280 | 2007-07-23 14:51:37 +0800 | [diff] [blame] | 480 | mutex_init(&kvm->lock); |
Marcelo Tosatti | 60eead7 | 2009-06-04 15:08:23 -0300 | [diff] [blame] | 481 | mutex_init(&kvm->irq_lock); |
Marcelo Tosatti | 79fac95 | 2009-12-23 14:35:26 -0200 | [diff] [blame] | 482 | mutex_init(&kvm->slots_lock); |
Izik Eidus | d39f13b | 2008-03-30 16:01:25 +0300 | [diff] [blame] | 483 | atomic_set(&kvm->users_count, 1); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 484 | raw_spin_lock(&kvm_lock); |
Rusty Russell | 5e58cfe | 2007-07-23 17:08:21 +1000 | [diff] [blame] | 485 | list_add(&kvm->vm_list, &vm_list); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 486 | raw_spin_unlock(&kvm_lock); |
Jan Kiszka | d89f5ef | 2010-11-09 17:02:49 +0100 | [diff] [blame] | 487 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 488 | return kvm; |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 489 | |
| 490 | out_err: |
Jan Kiszka | 57e7fbe | 2010-11-09 12:42:12 +0100 | [diff] [blame] | 491 | cleanup_srcu_struct(&kvm->srcu); |
| 492 | out_err_nosrcu: |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 493 | hardware_disable_all(); |
| 494 | out_err_nodisable: |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 495 | for (i = 0; i < KVM_NR_BUSES; i++) |
| 496 | kfree(kvm->buses[i]); |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 497 | kfree(kvm->memslots); |
Jan Kiszka | d89f5ef | 2010-11-09 17:02:49 +0100 | [diff] [blame] | 498 | kvm_arch_free_vm(kvm); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 499 | return ERR_PTR(r); |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 500 | } |
| 501 | |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 502 | static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot) |
| 503 | { |
| 504 | if (!memslot->dirty_bitmap) |
| 505 | return; |
| 506 | |
Takuya Yoshikawa | 6f9e5c17 | 2010-11-01 14:36:09 +0900 | [diff] [blame] | 507 | if (2 * kvm_dirty_bitmap_bytes(memslot) > PAGE_SIZE) |
| 508 | vfree(memslot->dirty_bitmap_head); |
| 509 | else |
| 510 | kfree(memslot->dirty_bitmap_head); |
| 511 | |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 512 | memslot->dirty_bitmap = NULL; |
Takuya Yoshikawa | 515a012 | 2010-10-27 18:23:54 +0900 | [diff] [blame] | 513 | memslot->dirty_bitmap_head = NULL; |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 514 | } |
| 515 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 516 | /* |
| 517 | * Free any memory in @free but not in @dont. |
| 518 | */ |
| 519 | static void kvm_free_physmem_slot(struct kvm_memory_slot *free, |
| 520 | struct kvm_memory_slot *dont) |
| 521 | { |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 522 | int i; |
| 523 | |
Izik Eidus | 290fc38 | 2007-09-27 14:11:22 +0200 | [diff] [blame] | 524 | if (!dont || free->rmap != dont->rmap) |
| 525 | vfree(free->rmap); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 526 | |
| 527 | if (!dont || free->dirty_bitmap != dont->dirty_bitmap) |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 528 | kvm_destroy_dirty_bitmap(free); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 529 | |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 530 | |
| 531 | for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) { |
| 532 | if (!dont || free->lpage_info[i] != dont->lpage_info[i]) { |
| 533 | vfree(free->lpage_info[i]); |
| 534 | free->lpage_info[i] = NULL; |
| 535 | } |
| 536 | } |
Marcelo Tosatti | 05da455 | 2008-02-23 11:44:30 -0300 | [diff] [blame] | 537 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 538 | free->npages = 0; |
Anthony Liguori | 8d4e128 | 2007-10-18 09:59:34 -0500 | [diff] [blame] | 539 | free->rmap = NULL; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 540 | } |
| 541 | |
Zhang Xiantao | d19a9cd | 2007-11-18 18:43:45 +0800 | [diff] [blame] | 542 | void kvm_free_physmem(struct kvm *kvm) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 543 | { |
| 544 | int i; |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 545 | struct kvm_memslots *slots = kvm->memslots; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 546 | |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 547 | for (i = 0; i < slots->nmemslots; ++i) |
| 548 | kvm_free_physmem_slot(&slots->memslots[i], NULL); |
| 549 | |
| 550 | kfree(kvm->memslots); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 551 | } |
| 552 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 553 | static void kvm_destroy_vm(struct kvm *kvm) |
| 554 | { |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 555 | int i; |
Avi Kivity | 6d4e4c4 | 2007-11-21 16:41:05 +0200 | [diff] [blame] | 556 | struct mm_struct *mm = kvm->mm; |
| 557 | |
Sheng Yang | ad8ba2c | 2009-01-06 10:03:02 +0800 | [diff] [blame] | 558 | kvm_arch_sync_events(kvm); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 559 | raw_spin_lock(&kvm_lock); |
Avi Kivity | 133de90 | 2007-02-12 00:54:44 -0800 | [diff] [blame] | 560 | list_del(&kvm->vm_list); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 561 | raw_spin_unlock(&kvm_lock); |
Avi Kivity | 399ec80 | 2008-11-19 13:58:46 +0200 | [diff] [blame] | 562 | kvm_free_irq_routing(kvm); |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 563 | for (i = 0; i < KVM_NR_BUSES; i++) |
| 564 | kvm_io_bus_destroy(kvm->buses[i]); |
Avi Kivity | 980da6c | 2009-12-20 15:13:43 +0200 | [diff] [blame] | 565 | kvm_coalesced_mmio_free(kvm); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 566 | #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) |
| 567 | mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm); |
Gleb Natapov | f00be0c | 2009-03-19 12:20:36 +0200 | [diff] [blame] | 568 | #else |
| 569 | kvm_arch_flush_shadow(kvm); |
Andrea Arcangeli | e930bff | 2008-07-25 16:24:52 +0200 | [diff] [blame] | 570 | #endif |
Zhang Xiantao | d19a9cd | 2007-11-18 18:43:45 +0800 | [diff] [blame] | 571 | kvm_arch_destroy_vm(kvm); |
Jan Kiszka | d89f5ef | 2010-11-09 17:02:49 +0100 | [diff] [blame] | 572 | kvm_free_physmem(kvm); |
| 573 | cleanup_srcu_struct(&kvm->srcu); |
| 574 | kvm_arch_free_vm(kvm); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 575 | hardware_disable_all(); |
Avi Kivity | 6d4e4c4 | 2007-11-21 16:41:05 +0200 | [diff] [blame] | 576 | mmdrop(mm); |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 577 | } |
| 578 | |
Izik Eidus | d39f13b | 2008-03-30 16:01:25 +0300 | [diff] [blame] | 579 | void kvm_get_kvm(struct kvm *kvm) |
| 580 | { |
| 581 | atomic_inc(&kvm->users_count); |
| 582 | } |
| 583 | EXPORT_SYMBOL_GPL(kvm_get_kvm); |
| 584 | |
| 585 | void kvm_put_kvm(struct kvm *kvm) |
| 586 | { |
| 587 | if (atomic_dec_and_test(&kvm->users_count)) |
| 588 | kvm_destroy_vm(kvm); |
| 589 | } |
| 590 | EXPORT_SYMBOL_GPL(kvm_put_kvm); |
| 591 | |
| 592 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 593 | static int kvm_vm_release(struct inode *inode, struct file *filp) |
| 594 | { |
| 595 | struct kvm *kvm = filp->private_data; |
| 596 | |
Gregory Haskins | 721eecb | 2009-05-20 10:30:49 -0400 | [diff] [blame] | 597 | kvm_irqfd_release(kvm); |
| 598 | |
Izik Eidus | d39f13b | 2008-03-30 16:01:25 +0300 | [diff] [blame] | 599 | kvm_put_kvm(kvm); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 600 | return 0; |
| 601 | } |
| 602 | |
Heiko Carstens | d48ead8 | 2011-01-17 21:21:08 +0100 | [diff] [blame] | 603 | #ifndef CONFIG_S390 |
Takuya Yoshikawa | 515a012 | 2010-10-27 18:23:54 +0900 | [diff] [blame] | 604 | /* |
| 605 | * Allocation size is twice as large as the actual dirty bitmap size. |
| 606 | * This makes it possible to do double buffering: see x86's |
| 607 | * kvm_vm_ioctl_get_dirty_log(). |
| 608 | */ |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 609 | static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot) |
| 610 | { |
Takuya Yoshikawa | 515a012 | 2010-10-27 18:23:54 +0900 | [diff] [blame] | 611 | unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot); |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 612 | |
Takuya Yoshikawa | 6f9e5c17 | 2010-11-01 14:36:09 +0900 | [diff] [blame] | 613 | if (dirty_bytes > PAGE_SIZE) |
| 614 | memslot->dirty_bitmap = vzalloc(dirty_bytes); |
| 615 | else |
| 616 | memslot->dirty_bitmap = kzalloc(dirty_bytes, GFP_KERNEL); |
| 617 | |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 618 | if (!memslot->dirty_bitmap) |
| 619 | return -ENOMEM; |
| 620 | |
Takuya Yoshikawa | 515a012 | 2010-10-27 18:23:54 +0900 | [diff] [blame] | 621 | memslot->dirty_bitmap_head = memslot->dirty_bitmap; |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 622 | return 0; |
| 623 | } |
Heiko Carstens | d48ead8 | 2011-01-17 21:21:08 +0100 | [diff] [blame] | 624 | #endif /* !CONFIG_S390 */ |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 625 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 626 | /* |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 627 | * Allocate some memory and give it an address in the guest physical address |
| 628 | * space. |
| 629 | * |
| 630 | * Discontiguous memory is allowed, mostly for framebuffers. |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 631 | * |
Marcelo Tosatti | 10589a4 | 2007-12-20 19:18:22 -0500 | [diff] [blame] | 632 | * Must be called holding mmap_sem for write. |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 633 | */ |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 634 | int __kvm_set_memory_region(struct kvm *kvm, |
| 635 | struct kvm_userspace_memory_region *mem, |
| 636 | int user_alloc) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 637 | { |
Avi Kivity | 8234b22 | 2010-12-27 12:08:45 +0200 | [diff] [blame] | 638 | int r; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 639 | gfn_t base_gfn; |
Heiko Carstens | 28bcb11 | 2009-09-03 17:35:35 +0200 | [diff] [blame] | 640 | unsigned long npages; |
| 641 | unsigned long i; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 642 | struct kvm_memory_slot *memslot; |
| 643 | struct kvm_memory_slot old, new; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 644 | struct kvm_memslots *slots, *old_memslots; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 645 | |
| 646 | r = -EINVAL; |
| 647 | /* General sanity checks */ |
| 648 | if (mem->memory_size & (PAGE_SIZE - 1)) |
| 649 | goto out; |
| 650 | if (mem->guest_phys_addr & (PAGE_SIZE - 1)) |
| 651 | goto out; |
Sheng Yang | e7cacd4 | 2008-11-11 15:30:40 +0800 | [diff] [blame] | 652 | if (user_alloc && (mem->userspace_addr & (PAGE_SIZE - 1))) |
Hollis Blanchard | 7874980 | 2008-11-07 13:32:12 -0600 | [diff] [blame] | 653 | goto out; |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 654 | if (mem->slot >= KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 655 | goto out; |
| 656 | if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr) |
| 657 | goto out; |
| 658 | |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 659 | memslot = &kvm->memslots->memslots[mem->slot]; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 660 | base_gfn = mem->guest_phys_addr >> PAGE_SHIFT; |
| 661 | npages = mem->memory_size >> PAGE_SHIFT; |
| 662 | |
Takuya Yoshikawa | 660c22c | 2010-04-13 22:47:24 +0900 | [diff] [blame] | 663 | r = -EINVAL; |
| 664 | if (npages > KVM_MEM_MAX_NR_PAGES) |
| 665 | goto out; |
| 666 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 667 | if (!npages) |
| 668 | mem->flags &= ~KVM_MEM_LOG_DIRTY_PAGES; |
| 669 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 670 | new = old = *memslot; |
| 671 | |
Avi Kivity | e36d96f | 2010-06-21 10:56:36 +0300 | [diff] [blame] | 672 | new.id = mem->slot; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 673 | new.base_gfn = base_gfn; |
| 674 | new.npages = npages; |
| 675 | new.flags = mem->flags; |
| 676 | |
| 677 | /* Disallow changing a memory slot's size. */ |
| 678 | r = -EINVAL; |
| 679 | if (npages && old.npages && npages != old.npages) |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 680 | goto out_free; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 681 | |
| 682 | /* Check for overlaps */ |
| 683 | r = -EEXIST; |
| 684 | for (i = 0; i < KVM_MEMORY_SLOTS; ++i) { |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 685 | struct kvm_memory_slot *s = &kvm->memslots->memslots[i]; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 686 | |
Jan Kiszka | 4cd481f | 2009-04-13 11:59:32 +0200 | [diff] [blame] | 687 | if (s == memslot || !s->npages) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 688 | continue; |
| 689 | if (!((base_gfn + npages <= s->base_gfn) || |
| 690 | (base_gfn >= s->base_gfn + s->npages))) |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 691 | goto out_free; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 692 | } |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 693 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 694 | /* Free page dirty bitmap if unneeded */ |
| 695 | if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES)) |
Al Viro | 8b6d44c | 2007-02-09 16:38:40 +0000 | [diff] [blame] | 696 | new.dirty_bitmap = NULL; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 697 | |
| 698 | r = -ENOMEM; |
| 699 | |
| 700 | /* Allocate if a slot is being created */ |
Carsten Otte | eff0114 | 2008-06-27 15:05:31 +0200 | [diff] [blame] | 701 | #ifndef CONFIG_S390 |
Anthony Liguori | 8d4e128 | 2007-10-18 09:59:34 -0500 | [diff] [blame] | 702 | if (npages && !new.rmap) { |
Takuya Yoshikawa | 2653503 | 2010-11-02 10:49:34 +0900 | [diff] [blame] | 703 | new.rmap = vzalloc(npages * sizeof(*new.rmap)); |
Izik Eidus | 290fc38 | 2007-09-27 14:11:22 +0200 | [diff] [blame] | 704 | |
| 705 | if (!new.rmap) |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 706 | goto out_free; |
Izik Eidus | 290fc38 | 2007-09-27 14:11:22 +0200 | [diff] [blame] | 707 | |
Izik Eidus | 80b14b5 | 2007-10-25 11:54:04 +0200 | [diff] [blame] | 708 | new.user_alloc = user_alloc; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 709 | new.userspace_addr = mem->userspace_addr; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 710 | } |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 711 | if (!npages) |
| 712 | goto skip_lpage; |
Marcelo Tosatti | 05da455 | 2008-02-23 11:44:30 -0300 | [diff] [blame] | 713 | |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 714 | for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) { |
Heiko Carstens | 28bcb11 | 2009-09-03 17:35:35 +0200 | [diff] [blame] | 715 | unsigned long ugfn; |
| 716 | unsigned long j; |
| 717 | int lpages; |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 718 | int level = i + 2; |
Marcelo Tosatti | 05da455 | 2008-02-23 11:44:30 -0300 | [diff] [blame] | 719 | |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 720 | /* Avoid unused variable warning if no large pages */ |
| 721 | (void)level; |
| 722 | |
| 723 | if (new.lpage_info[i]) |
| 724 | continue; |
| 725 | |
Joerg Roedel | 8285541 | 2010-07-01 16:00:11 +0200 | [diff] [blame] | 726 | lpages = 1 + ((base_gfn + npages - 1) |
| 727 | >> KVM_HPAGE_GFN_SHIFT(level)); |
| 728 | lpages -= base_gfn >> KVM_HPAGE_GFN_SHIFT(level); |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 729 | |
Takuya Yoshikawa | 2653503 | 2010-11-02 10:49:34 +0900 | [diff] [blame] | 730 | new.lpage_info[i] = vzalloc(lpages * sizeof(*new.lpage_info[i])); |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 731 | |
| 732 | if (!new.lpage_info[i]) |
Marcelo Tosatti | 05da455 | 2008-02-23 11:44:30 -0300 | [diff] [blame] | 733 | goto out_free; |
| 734 | |
Joerg Roedel | 8285541 | 2010-07-01 16:00:11 +0200 | [diff] [blame] | 735 | if (base_gfn & (KVM_PAGES_PER_HPAGE(level) - 1)) |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 736 | new.lpage_info[i][0].write_count = 1; |
Joerg Roedel | 8285541 | 2010-07-01 16:00:11 +0200 | [diff] [blame] | 737 | if ((base_gfn+npages) & (KVM_PAGES_PER_HPAGE(level) - 1)) |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 738 | new.lpage_info[i][lpages - 1].write_count = 1; |
Avi Kivity | ac04527 | 2009-06-08 15:52:39 +0300 | [diff] [blame] | 739 | ugfn = new.userspace_addr >> PAGE_SHIFT; |
| 740 | /* |
| 741 | * If the gfn and userspace address are not aligned wrt each |
Marcelo Tosatti | 54dee99 | 2009-06-11 12:07:44 -0300 | [diff] [blame] | 742 | * other, or if explicitly asked to, disable large page |
| 743 | * support for this slot |
Avi Kivity | ac04527 | 2009-06-08 15:52:39 +0300 | [diff] [blame] | 744 | */ |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 745 | if ((base_gfn ^ ugfn) & (KVM_PAGES_PER_HPAGE(level) - 1) || |
Marcelo Tosatti | 54dee99 | 2009-06-11 12:07:44 -0300 | [diff] [blame] | 746 | !largepages_enabled) |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 747 | for (j = 0; j < lpages; ++j) |
| 748 | new.lpage_info[i][j].write_count = 1; |
Marcelo Tosatti | 05da455 | 2008-02-23 11:44:30 -0300 | [diff] [blame] | 749 | } |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 750 | |
Joerg Roedel | ec04b26 | 2009-06-19 15:16:23 +0200 | [diff] [blame] | 751 | skip_lpage: |
| 752 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 753 | /* Allocate page dirty bitmap if needed */ |
| 754 | if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) { |
Takuya Yoshikawa | a36a57b1 | 2010-10-27 18:22:19 +0900 | [diff] [blame] | 755 | if (kvm_create_dirty_bitmap(&new) < 0) |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 756 | goto out_free; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 757 | /* destroy any largepage mappings for dirty tracking */ |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 758 | } |
Christian Borntraeger | 3eea843 | 2009-06-23 17:24:06 +0200 | [diff] [blame] | 759 | #else /* not defined CONFIG_S390 */ |
| 760 | new.user_alloc = user_alloc; |
| 761 | if (user_alloc) |
| 762 | new.userspace_addr = mem->userspace_addr; |
Carsten Otte | eff0114 | 2008-06-27 15:05:31 +0200 | [diff] [blame] | 763 | #endif /* not defined CONFIG_S390 */ |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 764 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 765 | if (!npages) { |
| 766 | r = -ENOMEM; |
| 767 | slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL); |
| 768 | if (!slots) |
| 769 | goto out_free; |
| 770 | memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots)); |
| 771 | if (mem->slot >= slots->nmemslots) |
| 772 | slots->nmemslots = mem->slot + 1; |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 773 | slots->generation++; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 774 | slots->memslots[mem->slot].flags |= KVM_MEMSLOT_INVALID; |
| 775 | |
| 776 | old_memslots = kvm->memslots; |
| 777 | rcu_assign_pointer(kvm->memslots, slots); |
| 778 | synchronize_srcu_expedited(&kvm->srcu); |
| 779 | /* From this point no new shadow pages pointing to a deleted |
| 780 | * memslot will be created. |
| 781 | * |
| 782 | * validation of sp->gfn happens in: |
| 783 | * - gfn_to_hva (kvm_read_guest, gfn_to_pfn) |
| 784 | * - kvm_is_visible_gfn (mmu_check_roots) |
| 785 | */ |
Marcelo Tosatti | 34d4cb8 | 2008-07-10 20:49:31 -0300 | [diff] [blame] | 786 | kvm_arch_flush_shadow(kvm); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 787 | kfree(old_memslots); |
| 788 | } |
Marcelo Tosatti | 34d4cb8 | 2008-07-10 20:49:31 -0300 | [diff] [blame] | 789 | |
Marcelo Tosatti | f7784b8 | 2009-12-23 14:35:18 -0200 | [diff] [blame] | 790 | r = kvm_arch_prepare_memory_region(kvm, &new, old, mem, user_alloc); |
| 791 | if (r) |
| 792 | goto out_free; |
| 793 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 794 | /* map the pages in iommu page table */ |
| 795 | if (npages) { |
| 796 | r = kvm_iommu_map_pages(kvm, &new); |
| 797 | if (r) |
| 798 | goto out_free; |
| 799 | } |
Andrea Arcangeli | 604b38a | 2008-07-25 16:32:03 +0200 | [diff] [blame] | 800 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 801 | r = -ENOMEM; |
| 802 | slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL); |
| 803 | if (!slots) |
| 804 | goto out_free; |
| 805 | memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots)); |
| 806 | if (mem->slot >= slots->nmemslots) |
| 807 | slots->nmemslots = mem->slot + 1; |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 808 | slots->generation++; |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 809 | |
| 810 | /* actual memory is freed via old in kvm_free_physmem_slot below */ |
| 811 | if (!npages) { |
| 812 | new.rmap = NULL; |
| 813 | new.dirty_bitmap = NULL; |
| 814 | for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) |
| 815 | new.lpage_info[i] = NULL; |
| 816 | } |
| 817 | |
| 818 | slots->memslots[mem->slot] = new; |
| 819 | old_memslots = kvm->memslots; |
| 820 | rcu_assign_pointer(kvm->memslots, slots); |
| 821 | synchronize_srcu_expedited(&kvm->srcu); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 822 | |
Marcelo Tosatti | f7784b8 | 2009-12-23 14:35:18 -0200 | [diff] [blame] | 823 | kvm_arch_commit_memory_region(kvm, mem, old, user_alloc); |
Zhang Xiantao | 3ad82a7 | 2007-11-20 13:11:38 +0800 | [diff] [blame] | 824 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 825 | kvm_free_physmem_slot(&old, &new); |
| 826 | kfree(old_memslots); |
| 827 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 828 | return 0; |
| 829 | |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 830 | out_free: |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 831 | kvm_free_physmem_slot(&new, &old); |
| 832 | out: |
| 833 | return r; |
Izik Eidus | 210c7c4 | 2007-10-24 23:52:57 +0200 | [diff] [blame] | 834 | |
| 835 | } |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 836 | EXPORT_SYMBOL_GPL(__kvm_set_memory_region); |
| 837 | |
| 838 | int kvm_set_memory_region(struct kvm *kvm, |
| 839 | struct kvm_userspace_memory_region *mem, |
| 840 | int user_alloc) |
| 841 | { |
| 842 | int r; |
| 843 | |
Marcelo Tosatti | 79fac95 | 2009-12-23 14:35:26 -0200 | [diff] [blame] | 844 | mutex_lock(&kvm->slots_lock); |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 845 | r = __kvm_set_memory_region(kvm, mem, user_alloc); |
Marcelo Tosatti | 79fac95 | 2009-12-23 14:35:26 -0200 | [diff] [blame] | 846 | mutex_unlock(&kvm->slots_lock); |
Sheng Yang | f78e0e2 | 2007-10-29 09:40:42 +0800 | [diff] [blame] | 847 | return r; |
| 848 | } |
Izik Eidus | 210c7c4 | 2007-10-24 23:52:57 +0200 | [diff] [blame] | 849 | EXPORT_SYMBOL_GPL(kvm_set_memory_region); |
| 850 | |
Carsten Otte | 1fe779f | 2007-10-29 16:08:35 +0100 | [diff] [blame] | 851 | int kvm_vm_ioctl_set_memory_region(struct kvm *kvm, |
| 852 | struct |
| 853 | kvm_userspace_memory_region *mem, |
| 854 | int user_alloc) |
Izik Eidus | 210c7c4 | 2007-10-24 23:52:57 +0200 | [diff] [blame] | 855 | { |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 856 | if (mem->slot >= KVM_MEMORY_SLOTS) |
| 857 | return -EINVAL; |
Izik Eidus | 210c7c4 | 2007-10-24 23:52:57 +0200 | [diff] [blame] | 858 | return kvm_set_memory_region(kvm, mem, user_alloc); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 859 | } |
| 860 | |
Zhang Xiantao | 5bb064d | 2007-11-18 20:29:43 +0800 | [diff] [blame] | 861 | int kvm_get_dirty_log(struct kvm *kvm, |
| 862 | struct kvm_dirty_log *log, int *is_dirty) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 863 | { |
| 864 | struct kvm_memory_slot *memslot; |
| 865 | int r, i; |
Takuya Yoshikawa | 87bf6e7 | 2010-04-12 19:35:35 +0900 | [diff] [blame] | 866 | unsigned long n; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 867 | unsigned long any = 0; |
| 868 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 869 | r = -EINVAL; |
| 870 | if (log->slot >= KVM_MEMORY_SLOTS) |
| 871 | goto out; |
| 872 | |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 873 | memslot = &kvm->memslots->memslots[log->slot]; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 874 | r = -ENOENT; |
| 875 | if (!memslot->dirty_bitmap) |
| 876 | goto out; |
| 877 | |
Takuya Yoshikawa | 87bf6e7 | 2010-04-12 19:35:35 +0900 | [diff] [blame] | 878 | n = kvm_dirty_bitmap_bytes(memslot); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 879 | |
Uri Lublin | cd1a4a9 | 2007-02-22 16:43:09 +0200 | [diff] [blame] | 880 | for (i = 0; !any && i < n/sizeof(long); ++i) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 881 | any = memslot->dirty_bitmap[i]; |
| 882 | |
| 883 | r = -EFAULT; |
| 884 | if (copy_to_user(log->dirty_bitmap, memslot->dirty_bitmap, n)) |
| 885 | goto out; |
| 886 | |
Zhang Xiantao | 5bb064d | 2007-11-18 20:29:43 +0800 | [diff] [blame] | 887 | if (any) |
| 888 | *is_dirty = 1; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 889 | |
| 890 | r = 0; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 891 | out: |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 892 | return r; |
| 893 | } |
| 894 | |
Marcelo Tosatti | 54dee99 | 2009-06-11 12:07:44 -0300 | [diff] [blame] | 895 | void kvm_disable_largepages(void) |
| 896 | { |
| 897 | largepages_enabled = false; |
| 898 | } |
| 899 | EXPORT_SYMBOL_GPL(kvm_disable_largepages); |
| 900 | |
Izik Eidus | cea7bb2 | 2007-10-17 19:17:48 +0200 | [diff] [blame] | 901 | int is_error_page(struct page *page) |
| 902 | { |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 903 | return page == bad_page || page == hwpoison_page || page == fault_page; |
Izik Eidus | cea7bb2 | 2007-10-17 19:17:48 +0200 | [diff] [blame] | 904 | } |
| 905 | EXPORT_SYMBOL_GPL(is_error_page); |
| 906 | |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 907 | int is_error_pfn(pfn_t pfn) |
| 908 | { |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 909 | return pfn == bad_pfn || pfn == hwpoison_pfn || pfn == fault_pfn; |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 910 | } |
| 911 | EXPORT_SYMBOL_GPL(is_error_pfn); |
| 912 | |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 913 | int is_hwpoison_pfn(pfn_t pfn) |
| 914 | { |
| 915 | return pfn == hwpoison_pfn; |
| 916 | } |
| 917 | EXPORT_SYMBOL_GPL(is_hwpoison_pfn); |
| 918 | |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 919 | int is_fault_pfn(pfn_t pfn) |
| 920 | { |
| 921 | return pfn == fault_pfn; |
| 922 | } |
| 923 | EXPORT_SYMBOL_GPL(is_fault_pfn); |
| 924 | |
Izik Eidus | f9d46eb | 2007-11-11 22:02:22 +0200 | [diff] [blame] | 925 | static inline unsigned long bad_hva(void) |
| 926 | { |
| 927 | return PAGE_OFFSET; |
| 928 | } |
| 929 | |
| 930 | int kvm_is_error_hva(unsigned long addr) |
| 931 | { |
| 932 | return addr == bad_hva(); |
| 933 | } |
| 934 | EXPORT_SYMBOL_GPL(kvm_is_error_hva); |
| 935 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 936 | static struct kvm_memory_slot *__gfn_to_memslot(struct kvm_memslots *slots, |
| 937 | gfn_t gfn) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 938 | { |
| 939 | int i; |
| 940 | |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 941 | for (i = 0; i < slots->nmemslots; ++i) { |
| 942 | struct kvm_memory_slot *memslot = &slots->memslots[i]; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 943 | |
| 944 | if (gfn >= memslot->base_gfn |
| 945 | && gfn < memslot->base_gfn + memslot->npages) |
| 946 | return memslot; |
| 947 | } |
Al Viro | 8b6d44c | 2007-02-09 16:38:40 +0000 | [diff] [blame] | 948 | return NULL; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 949 | } |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 950 | |
| 951 | struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn) |
| 952 | { |
| 953 | return __gfn_to_memslot(kvm_memslots(kvm), gfn); |
| 954 | } |
Avi Kivity | a1f4d395 | 2010-06-21 11:44:20 +0300 | [diff] [blame] | 955 | EXPORT_SYMBOL_GPL(gfn_to_memslot); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 956 | |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 957 | int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn) |
| 958 | { |
| 959 | int i; |
Lai Jiangshan | 90d83dc | 2010-04-19 17:41:23 +0800 | [diff] [blame] | 960 | struct kvm_memslots *slots = kvm_memslots(kvm); |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 961 | |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 962 | for (i = 0; i < KVM_MEMORY_SLOTS; ++i) { |
Marcelo Tosatti | 46a26bf | 2009-12-23 14:35:16 -0200 | [diff] [blame] | 963 | struct kvm_memory_slot *memslot = &slots->memslots[i]; |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 964 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 965 | if (memslot->flags & KVM_MEMSLOT_INVALID) |
| 966 | continue; |
| 967 | |
Izik Eidus | e0d62c7 | 2007-10-24 23:57:46 +0200 | [diff] [blame] | 968 | if (gfn >= memslot->base_gfn |
| 969 | && gfn < memslot->base_gfn + memslot->npages) |
| 970 | return 1; |
| 971 | } |
| 972 | return 0; |
| 973 | } |
| 974 | EXPORT_SYMBOL_GPL(kvm_is_visible_gfn); |
| 975 | |
Joerg Roedel | 8f0b1ab | 2010-01-28 12:37:56 +0100 | [diff] [blame] | 976 | unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn) |
| 977 | { |
| 978 | struct vm_area_struct *vma; |
| 979 | unsigned long addr, size; |
| 980 | |
| 981 | size = PAGE_SIZE; |
| 982 | |
| 983 | addr = gfn_to_hva(kvm, gfn); |
| 984 | if (kvm_is_error_hva(addr)) |
| 985 | return PAGE_SIZE; |
| 986 | |
| 987 | down_read(¤t->mm->mmap_sem); |
| 988 | vma = find_vma(current->mm, addr); |
| 989 | if (!vma) |
| 990 | goto out; |
| 991 | |
| 992 | size = vma_kernel_pagesize(vma); |
| 993 | |
| 994 | out: |
| 995 | up_read(¤t->mm->mmap_sem); |
| 996 | |
| 997 | return size; |
| 998 | } |
| 999 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 1000 | int memslot_id(struct kvm *kvm, gfn_t gfn) |
| 1001 | { |
| 1002 | int i; |
Lai Jiangshan | 90d83dc | 2010-04-19 17:41:23 +0800 | [diff] [blame] | 1003 | struct kvm_memslots *slots = kvm_memslots(kvm); |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 1004 | struct kvm_memory_slot *memslot = NULL; |
| 1005 | |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 1006 | for (i = 0; i < slots->nmemslots; ++i) { |
| 1007 | memslot = &slots->memslots[i]; |
| 1008 | |
| 1009 | if (gfn >= memslot->base_gfn |
| 1010 | && gfn < memslot->base_gfn + memslot->npages) |
| 1011 | break; |
| 1012 | } |
| 1013 | |
| 1014 | return memslot - slots->memslots; |
| 1015 | } |
| 1016 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1017 | static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn, |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1018 | gfn_t *nr_pages) |
Izik Eidus | 539cb66 | 2007-11-11 22:05:04 +0200 | [diff] [blame] | 1019 | { |
Marcelo Tosatti | bc6678a | 2009-12-23 14:35:21 -0200 | [diff] [blame] | 1020 | if (!slot || slot->flags & KVM_MEMSLOT_INVALID) |
Izik Eidus | 539cb66 | 2007-11-11 22:05:04 +0200 | [diff] [blame] | 1021 | return bad_hva(); |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1022 | |
| 1023 | if (nr_pages) |
| 1024 | *nr_pages = slot->npages - (gfn - slot->base_gfn); |
| 1025 | |
Takuya Yoshikawa | f5c9803 | 2010-02-25 11:33:19 +0900 | [diff] [blame] | 1026 | return gfn_to_hva_memslot(slot, gfn); |
Izik Eidus | 539cb66 | 2007-11-11 22:05:04 +0200 | [diff] [blame] | 1027 | } |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1028 | |
| 1029 | unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn) |
| 1030 | { |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1031 | return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL); |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1032 | } |
Sheng Yang | 0d15029 | 2008-04-25 21:44:50 +0800 | [diff] [blame] | 1033 | EXPORT_SYMBOL_GPL(gfn_to_hva); |
Izik Eidus | 539cb66 | 2007-11-11 22:05:04 +0200 | [diff] [blame] | 1034 | |
Gleb Natapov | 8030089 | 2010-10-19 18:13:41 +0200 | [diff] [blame] | 1035 | static pfn_t get_fault_pfn(void) |
| 1036 | { |
| 1037 | get_page(fault_page); |
| 1038 | return fault_pfn; |
| 1039 | } |
| 1040 | |
Huang Ying | fafc3db | 2011-01-30 11:15:49 +0800 | [diff] [blame] | 1041 | static inline int check_user_page_hwpoison(unsigned long addr) |
| 1042 | { |
| 1043 | int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE; |
| 1044 | |
| 1045 | rc = __get_user_pages(current, current->mm, addr, 1, |
| 1046 | flags, NULL, NULL, NULL); |
| 1047 | return rc == -EHWPOISON; |
| 1048 | } |
| 1049 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1050 | static pfn_t hva_to_pfn(struct kvm *kvm, unsigned long addr, bool atomic, |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1051 | bool *async, bool write_fault, bool *writable) |
Avi Kivity | 954bbbc | 2007-03-30 14:02:32 +0300 | [diff] [blame] | 1052 | { |
Anthony Liguori | 8d4e128 | 2007-10-18 09:59:34 -0500 | [diff] [blame] | 1053 | struct page *page[1]; |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1054 | int npages = 0; |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1055 | pfn_t pfn; |
Avi Kivity | 954bbbc | 2007-03-30 14:02:32 +0300 | [diff] [blame] | 1056 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1057 | /* we can do it either atomically or asynchronously, not both */ |
| 1058 | BUG_ON(atomic && async); |
| 1059 | |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1060 | BUG_ON(!write_fault && !writable); |
| 1061 | |
| 1062 | if (writable) |
| 1063 | *writable = true; |
| 1064 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1065 | if (atomic || async) |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1066 | npages = __get_user_pages_fast(addr, 1, 1, page); |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1067 | |
| 1068 | if (unlikely(npages != 1) && !atomic) { |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1069 | might_sleep(); |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1070 | |
| 1071 | if (writable) |
| 1072 | *writable = write_fault; |
| 1073 | |
| 1074 | npages = get_user_pages_fast(addr, 1, write_fault, page); |
| 1075 | |
| 1076 | /* map read fault as writable if possible */ |
| 1077 | if (unlikely(!write_fault) && npages == 1) { |
| 1078 | struct page *wpage[1]; |
| 1079 | |
| 1080 | npages = __get_user_pages_fast(addr, 1, 1, wpage); |
| 1081 | if (npages == 1) { |
| 1082 | *writable = true; |
| 1083 | put_page(page[0]); |
| 1084 | page[0] = wpage[0]; |
| 1085 | } |
| 1086 | npages = 1; |
| 1087 | } |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1088 | } |
Izik Eidus | 539cb66 | 2007-11-11 22:05:04 +0200 | [diff] [blame] | 1089 | |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1090 | if (unlikely(npages != 1)) { |
| 1091 | struct vm_area_struct *vma; |
Anthony Liguori | 8d4e128 | 2007-10-18 09:59:34 -0500 | [diff] [blame] | 1092 | |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1093 | if (atomic) |
Gleb Natapov | 8030089 | 2010-10-19 18:13:41 +0200 | [diff] [blame] | 1094 | return get_fault_pfn(); |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1095 | |
Huang Ying | bbeb340 | 2010-06-22 14:23:11 +0800 | [diff] [blame] | 1096 | down_read(¤t->mm->mmap_sem); |
Huang Ying | fafc3db | 2011-01-30 11:15:49 +0800 | [diff] [blame] | 1097 | if (check_user_page_hwpoison(addr)) { |
Huang Ying | bbeb340 | 2010-06-22 14:23:11 +0800 | [diff] [blame] | 1098 | up_read(¤t->mm->mmap_sem); |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 1099 | get_page(hwpoison_page); |
| 1100 | return page_to_pfn(hwpoison_page); |
| 1101 | } |
| 1102 | |
Gleb Natapov | 8030089 | 2010-10-19 18:13:41 +0200 | [diff] [blame] | 1103 | vma = find_vma_intersection(current->mm, addr, addr+1); |
Marcelo Tosatti | 4c2155c | 2008-09-16 20:54:47 -0300 | [diff] [blame] | 1104 | |
Gleb Natapov | 8030089 | 2010-10-19 18:13:41 +0200 | [diff] [blame] | 1105 | if (vma == NULL) |
| 1106 | pfn = get_fault_pfn(); |
| 1107 | else if ((vma->vm_flags & VM_PFNMAP)) { |
| 1108 | pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) + |
| 1109 | vma->vm_pgoff; |
| 1110 | BUG_ON(!kvm_is_mmio_pfn(pfn)); |
| 1111 | } else { |
| 1112 | if (async && (vma->vm_flags & VM_WRITE)) |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1113 | *async = true; |
Gleb Natapov | 8030089 | 2010-10-19 18:13:41 +0200 | [diff] [blame] | 1114 | pfn = get_fault_pfn(); |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1115 | } |
Marcelo Tosatti | 4c2155c | 2008-09-16 20:54:47 -0300 | [diff] [blame] | 1116 | up_read(¤t->mm->mmap_sem); |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1117 | } else |
| 1118 | pfn = page_to_pfn(page[0]); |
| 1119 | |
| 1120 | return pfn; |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1121 | } |
| 1122 | |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1123 | pfn_t hva_to_pfn_atomic(struct kvm *kvm, unsigned long addr) |
| 1124 | { |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1125 | return hva_to_pfn(kvm, addr, true, NULL, true, NULL); |
Xiao Guangrong | 887c08a | 2010-08-22 19:10:28 +0800 | [diff] [blame] | 1126 | } |
| 1127 | EXPORT_SYMBOL_GPL(hva_to_pfn_atomic); |
| 1128 | |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1129 | static pfn_t __gfn_to_pfn(struct kvm *kvm, gfn_t gfn, bool atomic, bool *async, |
| 1130 | bool write_fault, bool *writable) |
Marcelo Tosatti | 506f0d6 | 2009-12-23 14:35:19 -0200 | [diff] [blame] | 1131 | { |
| 1132 | unsigned long addr; |
| 1133 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1134 | if (async) |
| 1135 | *async = false; |
| 1136 | |
Marcelo Tosatti | 506f0d6 | 2009-12-23 14:35:19 -0200 | [diff] [blame] | 1137 | addr = gfn_to_hva(kvm, gfn); |
| 1138 | if (kvm_is_error_hva(addr)) { |
| 1139 | get_page(bad_page); |
| 1140 | return page_to_pfn(bad_page); |
| 1141 | } |
| 1142 | |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1143 | return hva_to_pfn(kvm, addr, atomic, async, write_fault, writable); |
Xiao Guangrong | 365fb3f | 2010-08-28 19:24:13 +0800 | [diff] [blame] | 1144 | } |
| 1145 | |
| 1146 | pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn) |
| 1147 | { |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1148 | return __gfn_to_pfn(kvm, gfn, true, NULL, true, NULL); |
Xiao Guangrong | 365fb3f | 2010-08-28 19:24:13 +0800 | [diff] [blame] | 1149 | } |
| 1150 | EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic); |
| 1151 | |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1152 | pfn_t gfn_to_pfn_async(struct kvm *kvm, gfn_t gfn, bool *async, |
| 1153 | bool write_fault, bool *writable) |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1154 | { |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1155 | return __gfn_to_pfn(kvm, gfn, false, async, write_fault, writable); |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 1156 | } |
| 1157 | EXPORT_SYMBOL_GPL(gfn_to_pfn_async); |
| 1158 | |
Xiao Guangrong | 365fb3f | 2010-08-28 19:24:13 +0800 | [diff] [blame] | 1159 | pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn) |
| 1160 | { |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1161 | return __gfn_to_pfn(kvm, gfn, false, NULL, true, NULL); |
Marcelo Tosatti | 506f0d6 | 2009-12-23 14:35:19 -0200 | [diff] [blame] | 1162 | } |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1163 | EXPORT_SYMBOL_GPL(gfn_to_pfn); |
| 1164 | |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1165 | pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault, |
| 1166 | bool *writable) |
| 1167 | { |
| 1168 | return __gfn_to_pfn(kvm, gfn, false, NULL, write_fault, writable); |
| 1169 | } |
| 1170 | EXPORT_SYMBOL_GPL(gfn_to_pfn_prot); |
| 1171 | |
Marcelo Tosatti | 506f0d6 | 2009-12-23 14:35:19 -0200 | [diff] [blame] | 1172 | pfn_t gfn_to_pfn_memslot(struct kvm *kvm, |
| 1173 | struct kvm_memory_slot *slot, gfn_t gfn) |
| 1174 | { |
| 1175 | unsigned long addr = gfn_to_hva_memslot(slot, gfn); |
Marcelo Tosatti | 612819c | 2010-10-22 14:18:18 -0200 | [diff] [blame] | 1176 | return hva_to_pfn(kvm, addr, false, NULL, true, NULL); |
Marcelo Tosatti | 506f0d6 | 2009-12-23 14:35:19 -0200 | [diff] [blame] | 1177 | } |
| 1178 | |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1179 | int gfn_to_page_many_atomic(struct kvm *kvm, gfn_t gfn, struct page **pages, |
| 1180 | int nr_pages) |
| 1181 | { |
| 1182 | unsigned long addr; |
| 1183 | gfn_t entry; |
| 1184 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1185 | addr = gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, &entry); |
Xiao Guangrong | 4898778 | 2010-08-22 19:11:43 +0800 | [diff] [blame] | 1186 | if (kvm_is_error_hva(addr)) |
| 1187 | return -1; |
| 1188 | |
| 1189 | if (entry < nr_pages) |
| 1190 | return 0; |
| 1191 | |
| 1192 | return __get_user_pages_fast(addr, nr_pages, 1, pages); |
| 1193 | } |
| 1194 | EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic); |
| 1195 | |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1196 | struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn) |
| 1197 | { |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1198 | pfn_t pfn; |
| 1199 | |
| 1200 | pfn = gfn_to_pfn(kvm, gfn); |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1201 | if (!kvm_is_mmio_pfn(pfn)) |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1202 | return pfn_to_page(pfn); |
| 1203 | |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1204 | WARN_ON(kvm_is_mmio_pfn(pfn)); |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1205 | |
| 1206 | get_page(bad_page); |
| 1207 | return bad_page; |
Avi Kivity | 954bbbc | 2007-03-30 14:02:32 +0300 | [diff] [blame] | 1208 | } |
Anthony Liguori | aab61cc | 2007-10-29 15:15:20 -0500 | [diff] [blame] | 1209 | |
Avi Kivity | 954bbbc | 2007-03-30 14:02:32 +0300 | [diff] [blame] | 1210 | EXPORT_SYMBOL_GPL(gfn_to_page); |
| 1211 | |
Izik Eidus | b4231d6 | 2007-11-20 11:49:33 +0200 | [diff] [blame] | 1212 | void kvm_release_page_clean(struct page *page) |
| 1213 | { |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1214 | kvm_release_pfn_clean(page_to_pfn(page)); |
Izik Eidus | b4231d6 | 2007-11-20 11:49:33 +0200 | [diff] [blame] | 1215 | } |
| 1216 | EXPORT_SYMBOL_GPL(kvm_release_page_clean); |
| 1217 | |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1218 | void kvm_release_pfn_clean(pfn_t pfn) |
| 1219 | { |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1220 | if (!kvm_is_mmio_pfn(pfn)) |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1221 | put_page(pfn_to_page(pfn)); |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1222 | } |
| 1223 | EXPORT_SYMBOL_GPL(kvm_release_pfn_clean); |
| 1224 | |
Izik Eidus | b4231d6 | 2007-11-20 11:49:33 +0200 | [diff] [blame] | 1225 | void kvm_release_page_dirty(struct page *page) |
Izik Eidus | 8a7ae05 | 2007-10-18 11:09:33 +0200 | [diff] [blame] | 1226 | { |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1227 | kvm_release_pfn_dirty(page_to_pfn(page)); |
Izik Eidus | 8a7ae05 | 2007-10-18 11:09:33 +0200 | [diff] [blame] | 1228 | } |
Izik Eidus | b4231d6 | 2007-11-20 11:49:33 +0200 | [diff] [blame] | 1229 | EXPORT_SYMBOL_GPL(kvm_release_page_dirty); |
Izik Eidus | 8a7ae05 | 2007-10-18 11:09:33 +0200 | [diff] [blame] | 1230 | |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1231 | void kvm_release_pfn_dirty(pfn_t pfn) |
| 1232 | { |
| 1233 | kvm_set_pfn_dirty(pfn); |
| 1234 | kvm_release_pfn_clean(pfn); |
| 1235 | } |
| 1236 | EXPORT_SYMBOL_GPL(kvm_release_pfn_dirty); |
| 1237 | |
| 1238 | void kvm_set_page_dirty(struct page *page) |
| 1239 | { |
| 1240 | kvm_set_pfn_dirty(page_to_pfn(page)); |
| 1241 | } |
| 1242 | EXPORT_SYMBOL_GPL(kvm_set_page_dirty); |
| 1243 | |
| 1244 | void kvm_set_pfn_dirty(pfn_t pfn) |
| 1245 | { |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1246 | if (!kvm_is_mmio_pfn(pfn)) { |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1247 | struct page *page = pfn_to_page(pfn); |
| 1248 | if (!PageReserved(page)) |
| 1249 | SetPageDirty(page); |
| 1250 | } |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1251 | } |
| 1252 | EXPORT_SYMBOL_GPL(kvm_set_pfn_dirty); |
| 1253 | |
| 1254 | void kvm_set_pfn_accessed(pfn_t pfn) |
| 1255 | { |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1256 | if (!kvm_is_mmio_pfn(pfn)) |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1257 | mark_page_accessed(pfn_to_page(pfn)); |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1258 | } |
| 1259 | EXPORT_SYMBOL_GPL(kvm_set_pfn_accessed); |
| 1260 | |
| 1261 | void kvm_get_pfn(pfn_t pfn) |
| 1262 | { |
Xiantao Zhang | c77fb9d | 2008-09-27 10:55:40 +0800 | [diff] [blame] | 1263 | if (!kvm_is_mmio_pfn(pfn)) |
Anthony Liguori | 2e2e373 | 2008-04-30 15:37:07 -0500 | [diff] [blame] | 1264 | get_page(pfn_to_page(pfn)); |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 1265 | } |
| 1266 | EXPORT_SYMBOL_GPL(kvm_get_pfn); |
| 1267 | |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1268 | static int next_segment(unsigned long len, int offset) |
| 1269 | { |
| 1270 | if (len > PAGE_SIZE - offset) |
| 1271 | return PAGE_SIZE - offset; |
| 1272 | else |
| 1273 | return len; |
| 1274 | } |
| 1275 | |
| 1276 | int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset, |
| 1277 | int len) |
| 1278 | { |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1279 | int r; |
| 1280 | unsigned long addr; |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1281 | |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1282 | addr = gfn_to_hva(kvm, gfn); |
| 1283 | if (kvm_is_error_hva(addr)) |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1284 | return -EFAULT; |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1285 | r = copy_from_user(data, (void __user *)addr + offset, len); |
| 1286 | if (r) |
| 1287 | return -EFAULT; |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1288 | return 0; |
| 1289 | } |
| 1290 | EXPORT_SYMBOL_GPL(kvm_read_guest_page); |
| 1291 | |
| 1292 | int kvm_read_guest(struct kvm *kvm, gpa_t gpa, void *data, unsigned long len) |
| 1293 | { |
| 1294 | gfn_t gfn = gpa >> PAGE_SHIFT; |
| 1295 | int seg; |
| 1296 | int offset = offset_in_page(gpa); |
| 1297 | int ret; |
| 1298 | |
| 1299 | while ((seg = next_segment(len, offset)) != 0) { |
| 1300 | ret = kvm_read_guest_page(kvm, gfn, data, offset, seg); |
| 1301 | if (ret < 0) |
| 1302 | return ret; |
| 1303 | offset = 0; |
| 1304 | len -= seg; |
| 1305 | data += seg; |
| 1306 | ++gfn; |
| 1307 | } |
| 1308 | return 0; |
| 1309 | } |
| 1310 | EXPORT_SYMBOL_GPL(kvm_read_guest); |
| 1311 | |
Marcelo Tosatti | 7ec5458 | 2007-12-20 19:18:23 -0500 | [diff] [blame] | 1312 | int kvm_read_guest_atomic(struct kvm *kvm, gpa_t gpa, void *data, |
| 1313 | unsigned long len) |
| 1314 | { |
| 1315 | int r; |
| 1316 | unsigned long addr; |
| 1317 | gfn_t gfn = gpa >> PAGE_SHIFT; |
| 1318 | int offset = offset_in_page(gpa); |
| 1319 | |
| 1320 | addr = gfn_to_hva(kvm, gfn); |
| 1321 | if (kvm_is_error_hva(addr)) |
| 1322 | return -EFAULT; |
Andrea Arcangeli | 0aac03f | 2008-01-30 19:57:35 +0100 | [diff] [blame] | 1323 | pagefault_disable(); |
Marcelo Tosatti | 7ec5458 | 2007-12-20 19:18:23 -0500 | [diff] [blame] | 1324 | r = __copy_from_user_inatomic(data, (void __user *)addr + offset, len); |
Andrea Arcangeli | 0aac03f | 2008-01-30 19:57:35 +0100 | [diff] [blame] | 1325 | pagefault_enable(); |
Marcelo Tosatti | 7ec5458 | 2007-12-20 19:18:23 -0500 | [diff] [blame] | 1326 | if (r) |
| 1327 | return -EFAULT; |
| 1328 | return 0; |
| 1329 | } |
| 1330 | EXPORT_SYMBOL(kvm_read_guest_atomic); |
| 1331 | |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1332 | int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, const void *data, |
| 1333 | int offset, int len) |
| 1334 | { |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1335 | int r; |
| 1336 | unsigned long addr; |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1337 | |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1338 | addr = gfn_to_hva(kvm, gfn); |
| 1339 | if (kvm_is_error_hva(addr)) |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1340 | return -EFAULT; |
Izik Eidus | e0506bc | 2007-11-11 22:10:22 +0200 | [diff] [blame] | 1341 | r = copy_to_user((void __user *)addr + offset, data, len); |
| 1342 | if (r) |
| 1343 | return -EFAULT; |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1344 | mark_page_dirty(kvm, gfn); |
| 1345 | return 0; |
| 1346 | } |
| 1347 | EXPORT_SYMBOL_GPL(kvm_write_guest_page); |
| 1348 | |
| 1349 | int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data, |
| 1350 | unsigned long len) |
| 1351 | { |
| 1352 | gfn_t gfn = gpa >> PAGE_SHIFT; |
| 1353 | int seg; |
| 1354 | int offset = offset_in_page(gpa); |
| 1355 | int ret; |
| 1356 | |
| 1357 | while ((seg = next_segment(len, offset)) != 0) { |
| 1358 | ret = kvm_write_guest_page(kvm, gfn, data, offset, seg); |
| 1359 | if (ret < 0) |
| 1360 | return ret; |
| 1361 | offset = 0; |
| 1362 | len -= seg; |
| 1363 | data += seg; |
| 1364 | ++gfn; |
| 1365 | } |
| 1366 | return 0; |
| 1367 | } |
| 1368 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1369 | int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc, |
| 1370 | gpa_t gpa) |
| 1371 | { |
| 1372 | struct kvm_memslots *slots = kvm_memslots(kvm); |
| 1373 | int offset = offset_in_page(gpa); |
| 1374 | gfn_t gfn = gpa >> PAGE_SHIFT; |
| 1375 | |
| 1376 | ghc->gpa = gpa; |
| 1377 | ghc->generation = slots->generation; |
| 1378 | ghc->memslot = __gfn_to_memslot(slots, gfn); |
| 1379 | ghc->hva = gfn_to_hva_many(ghc->memslot, gfn, NULL); |
| 1380 | if (!kvm_is_error_hva(ghc->hva)) |
| 1381 | ghc->hva += offset; |
| 1382 | else |
| 1383 | return -EFAULT; |
| 1384 | |
| 1385 | return 0; |
| 1386 | } |
| 1387 | EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init); |
| 1388 | |
| 1389 | int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc, |
| 1390 | void *data, unsigned long len) |
| 1391 | { |
| 1392 | struct kvm_memslots *slots = kvm_memslots(kvm); |
| 1393 | int r; |
| 1394 | |
| 1395 | if (slots->generation != ghc->generation) |
| 1396 | kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa); |
| 1397 | |
| 1398 | if (kvm_is_error_hva(ghc->hva)) |
| 1399 | return -EFAULT; |
| 1400 | |
| 1401 | r = copy_to_user((void __user *)ghc->hva, data, len); |
| 1402 | if (r) |
| 1403 | return -EFAULT; |
| 1404 | mark_page_dirty_in_slot(kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT); |
| 1405 | |
| 1406 | return 0; |
| 1407 | } |
| 1408 | EXPORT_SYMBOL_GPL(kvm_write_guest_cached); |
| 1409 | |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1410 | int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len) |
| 1411 | { |
Heiko Carstens | 3bcc8a8 | 2010-10-27 17:21:21 +0200 | [diff] [blame] | 1412 | return kvm_write_guest_page(kvm, gfn, (const void *) empty_zero_page, |
| 1413 | offset, len); |
Izik Eidus | 195aefd | 2007-10-01 22:14:18 +0200 | [diff] [blame] | 1414 | } |
| 1415 | EXPORT_SYMBOL_GPL(kvm_clear_guest_page); |
| 1416 | |
| 1417 | int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len) |
| 1418 | { |
| 1419 | gfn_t gfn = gpa >> PAGE_SHIFT; |
| 1420 | int seg; |
| 1421 | int offset = offset_in_page(gpa); |
| 1422 | int ret; |
| 1423 | |
| 1424 | while ((seg = next_segment(len, offset)) != 0) { |
| 1425 | ret = kvm_clear_guest_page(kvm, gfn, offset, seg); |
| 1426 | if (ret < 0) |
| 1427 | return ret; |
| 1428 | offset = 0; |
| 1429 | len -= seg; |
| 1430 | ++gfn; |
| 1431 | } |
| 1432 | return 0; |
| 1433 | } |
| 1434 | EXPORT_SYMBOL_GPL(kvm_clear_guest); |
| 1435 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1436 | void mark_page_dirty_in_slot(struct kvm *kvm, struct kvm_memory_slot *memslot, |
| 1437 | gfn_t gfn) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1438 | { |
Rusty Russell | 7e9d619 | 2007-07-31 20:41:14 +1000 | [diff] [blame] | 1439 | if (memslot && memslot->dirty_bitmap) { |
| 1440 | unsigned long rel_gfn = gfn - memslot->base_gfn; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1441 | |
Takuya Yoshikawa | d147693 | 2010-04-23 17:48:35 +0900 | [diff] [blame] | 1442 | generic___set_le_bit(rel_gfn, memslot->dirty_bitmap); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1443 | } |
| 1444 | } |
| 1445 | |
Gleb Natapov | 49c7754 | 2010-10-18 15:22:23 +0200 | [diff] [blame] | 1446 | void mark_page_dirty(struct kvm *kvm, gfn_t gfn) |
| 1447 | { |
| 1448 | struct kvm_memory_slot *memslot; |
| 1449 | |
| 1450 | memslot = gfn_to_memslot(kvm, gfn); |
| 1451 | mark_page_dirty_in_slot(kvm, memslot, gfn); |
| 1452 | } |
| 1453 | |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1454 | /* |
| 1455 | * The vCPU has executed a HLT instruction with in-kernel mode enabled. |
| 1456 | */ |
Hollis Blanchard | 8776e51 | 2007-10-31 17:24:24 -0500 | [diff] [blame] | 1457 | void kvm_vcpu_block(struct kvm_vcpu *vcpu) |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1458 | { |
Marcelo Tosatti | e5c239c | 2008-05-08 19:47:01 -0300 | [diff] [blame] | 1459 | DEFINE_WAIT(wait); |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1460 | |
Marcelo Tosatti | e5c239c | 2008-05-08 19:47:01 -0300 | [diff] [blame] | 1461 | for (;;) { |
| 1462 | prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE); |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1463 | |
Gleb Natapov | a1b3710 | 2009-07-09 15:33:52 +0300 | [diff] [blame] | 1464 | if (kvm_arch_vcpu_runnable(vcpu)) { |
Avi Kivity | a8eeb04 | 2010-05-10 12:34:53 +0300 | [diff] [blame] | 1465 | kvm_make_request(KVM_REQ_UNHALT, vcpu); |
Marcelo Tosatti | e5c239c | 2008-05-08 19:47:01 -0300 | [diff] [blame] | 1466 | break; |
Marcelo Tosatti | d769017 | 2008-09-08 15:23:48 -0300 | [diff] [blame] | 1467 | } |
Gleb Natapov | 09cec75 | 2009-03-23 15:11:44 +0200 | [diff] [blame] | 1468 | if (kvm_cpu_has_pending_timer(vcpu)) |
| 1469 | break; |
Marcelo Tosatti | e5c239c | 2008-05-08 19:47:01 -0300 | [diff] [blame] | 1470 | if (signal_pending(current)) |
| 1471 | break; |
| 1472 | |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1473 | schedule(); |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1474 | } |
| 1475 | |
Marcelo Tosatti | e5c239c | 2008-05-08 19:47:01 -0300 | [diff] [blame] | 1476 | finish_wait(&vcpu->wq, &wait); |
Eddie Dong | b6958ce | 2007-07-18 12:15:21 +0300 | [diff] [blame] | 1477 | } |
| 1478 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1479 | void kvm_resched(struct kvm_vcpu *vcpu) |
| 1480 | { |
Yaozu Dong | 3fca036 | 2007-04-25 16:49:19 +0300 | [diff] [blame] | 1481 | if (!need_resched()) |
| 1482 | return; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1483 | cond_resched(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1484 | } |
| 1485 | EXPORT_SYMBOL_GPL(kvm_resched); |
| 1486 | |
Rik van Riel | 217ece6 | 2011-02-01 09:53:28 -0500 | [diff] [blame] | 1487 | void kvm_vcpu_on_spin(struct kvm_vcpu *me) |
Zhai, Edwin | d255f4f | 2009-10-09 18:03:20 +0800 | [diff] [blame] | 1488 | { |
Rik van Riel | 217ece6 | 2011-02-01 09:53:28 -0500 | [diff] [blame] | 1489 | struct kvm *kvm = me->kvm; |
| 1490 | struct kvm_vcpu *vcpu; |
| 1491 | int last_boosted_vcpu = me->kvm->last_boosted_vcpu; |
| 1492 | int yielded = 0; |
| 1493 | int pass; |
| 1494 | int i; |
Zhai, Edwin | d255f4f | 2009-10-09 18:03:20 +0800 | [diff] [blame] | 1495 | |
Rik van Riel | 217ece6 | 2011-02-01 09:53:28 -0500 | [diff] [blame] | 1496 | /* |
| 1497 | * We boost the priority of a VCPU that is runnable but not |
| 1498 | * currently running, because it got preempted by something |
| 1499 | * else and called schedule in __vcpu_run. Hopefully that |
| 1500 | * VCPU is holding the lock that we need and will release it. |
| 1501 | * We approximate round-robin by starting at the last boosted VCPU. |
| 1502 | */ |
| 1503 | for (pass = 0; pass < 2 && !yielded; pass++) { |
| 1504 | kvm_for_each_vcpu(i, vcpu, kvm) { |
| 1505 | struct task_struct *task = NULL; |
| 1506 | struct pid *pid; |
| 1507 | if (!pass && i < last_boosted_vcpu) { |
| 1508 | i = last_boosted_vcpu; |
| 1509 | continue; |
| 1510 | } else if (pass && i > last_boosted_vcpu) |
| 1511 | break; |
| 1512 | if (vcpu == me) |
| 1513 | continue; |
| 1514 | if (waitqueue_active(&vcpu->wq)) |
| 1515 | continue; |
| 1516 | rcu_read_lock(); |
| 1517 | pid = rcu_dereference(vcpu->pid); |
| 1518 | if (pid) |
| 1519 | task = get_pid_task(vcpu->pid, PIDTYPE_PID); |
| 1520 | rcu_read_unlock(); |
| 1521 | if (!task) |
| 1522 | continue; |
| 1523 | if (task->flags & PF_VCPU) { |
| 1524 | put_task_struct(task); |
| 1525 | continue; |
| 1526 | } |
| 1527 | if (yield_to(task, 1)) { |
| 1528 | put_task_struct(task); |
| 1529 | kvm->last_boosted_vcpu = i; |
| 1530 | yielded = 1; |
| 1531 | break; |
| 1532 | } |
| 1533 | put_task_struct(task); |
| 1534 | } |
| 1535 | } |
Zhai, Edwin | d255f4f | 2009-10-09 18:03:20 +0800 | [diff] [blame] | 1536 | } |
| 1537 | EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin); |
| 1538 | |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1539 | static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1540 | { |
| 1541 | struct kvm_vcpu *vcpu = vma->vm_file->private_data; |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1542 | struct page *page; |
| 1543 | |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1544 | if (vmf->pgoff == 0) |
Avi Kivity | 039576c | 2007-03-20 12:46:50 +0200 | [diff] [blame] | 1545 | page = virt_to_page(vcpu->run); |
Avi Kivity | 0956676 | 2008-01-23 18:14:23 +0200 | [diff] [blame] | 1546 | #ifdef CONFIG_X86 |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1547 | else if (vmf->pgoff == KVM_PIO_PAGE_OFFSET) |
Zhang Xiantao | ad312c7 | 2007-12-13 23:50:52 +0800 | [diff] [blame] | 1548 | page = virt_to_page(vcpu->arch.pio_data); |
Avi Kivity | 0956676 | 2008-01-23 18:14:23 +0200 | [diff] [blame] | 1549 | #endif |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 1550 | #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET |
| 1551 | else if (vmf->pgoff == KVM_COALESCED_MMIO_PAGE_OFFSET) |
| 1552 | page = virt_to_page(vcpu->kvm->coalesced_mmio_ring); |
| 1553 | #endif |
Avi Kivity | 039576c | 2007-03-20 12:46:50 +0200 | [diff] [blame] | 1554 | else |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1555 | return VM_FAULT_SIGBUS; |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1556 | get_page(page); |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1557 | vmf->page = page; |
| 1558 | return 0; |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1559 | } |
| 1560 | |
Alexey Dobriyan | f0f37e2 | 2009-09-27 22:29:37 +0400 | [diff] [blame] | 1561 | static const struct vm_operations_struct kvm_vcpu_vm_ops = { |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 1562 | .fault = kvm_vcpu_fault, |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1563 | }; |
| 1564 | |
| 1565 | static int kvm_vcpu_mmap(struct file *file, struct vm_area_struct *vma) |
| 1566 | { |
| 1567 | vma->vm_ops = &kvm_vcpu_vm_ops; |
| 1568 | return 0; |
| 1569 | } |
| 1570 | |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1571 | static int kvm_vcpu_release(struct inode *inode, struct file *filp) |
| 1572 | { |
| 1573 | struct kvm_vcpu *vcpu = filp->private_data; |
| 1574 | |
Al Viro | 66c0b39 | 2008-04-19 20:33:56 +0100 | [diff] [blame] | 1575 | kvm_put_kvm(vcpu->kvm); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1576 | return 0; |
| 1577 | } |
| 1578 | |
Christian Borntraeger | 3d3aab1 | 2008-12-02 11:17:32 +0100 | [diff] [blame] | 1579 | static struct file_operations kvm_vcpu_fops = { |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1580 | .release = kvm_vcpu_release, |
| 1581 | .unlocked_ioctl = kvm_vcpu_ioctl, |
| 1582 | .compat_ioctl = kvm_vcpu_ioctl, |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1583 | .mmap = kvm_vcpu_mmap, |
Arnd Bergmann | 6038f37 | 2010-08-15 18:52:59 +0200 | [diff] [blame] | 1584 | .llseek = noop_llseek, |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1585 | }; |
| 1586 | |
| 1587 | /* |
| 1588 | * Allocates an inode for the vcpu. |
| 1589 | */ |
| 1590 | static int create_vcpu_fd(struct kvm_vcpu *vcpu) |
| 1591 | { |
Roland Dreier | 628ff7c | 2009-12-18 09:41:24 -0800 | [diff] [blame] | 1592 | return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1593 | } |
| 1594 | |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1595 | /* |
| 1596 | * Creates some virtual cpus. Good luck creating more than one. |
| 1597 | */ |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1598 | static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id) |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1599 | { |
| 1600 | int r; |
Gleb Natapov | 988a2ca | 2009-06-09 15:56:29 +0300 | [diff] [blame] | 1601 | struct kvm_vcpu *vcpu, *v; |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1602 | |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1603 | vcpu = kvm_arch_vcpu_create(kvm, id); |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 1604 | if (IS_ERR(vcpu)) |
| 1605 | return PTR_ERR(vcpu); |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1606 | |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 1607 | preempt_notifier_init(&vcpu->preempt_notifier, &kvm_preempt_ops); |
| 1608 | |
Avi Kivity | 26e5215 | 2007-11-20 15:30:24 +0200 | [diff] [blame] | 1609 | r = kvm_arch_vcpu_setup(vcpu); |
| 1610 | if (r) |
Glauber Costa | 7d8fece | 2008-09-17 23:16:59 -0300 | [diff] [blame] | 1611 | return r; |
Avi Kivity | 26e5215 | 2007-11-20 15:30:24 +0200 | [diff] [blame] | 1612 | |
Shaohua Li | 11ec280 | 2007-07-23 14:51:37 +0800 | [diff] [blame] | 1613 | mutex_lock(&kvm->lock); |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1614 | if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) { |
| 1615 | r = -EINVAL; |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 1616 | goto vcpu_destroy; |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 1617 | } |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1618 | |
Gleb Natapov | 988a2ca | 2009-06-09 15:56:29 +0300 | [diff] [blame] | 1619 | kvm_for_each_vcpu(r, v, kvm) |
| 1620 | if (v->vcpu_id == id) { |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1621 | r = -EEXIST; |
| 1622 | goto vcpu_destroy; |
| 1623 | } |
| 1624 | |
| 1625 | BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]); |
Rusty Russell | fb3f0f5 | 2007-07-27 17:16:56 +1000 | [diff] [blame] | 1626 | |
| 1627 | /* Now it's all set up, let userspace reach it */ |
Al Viro | 66c0b39 | 2008-04-19 20:33:56 +0100 | [diff] [blame] | 1628 | kvm_get_kvm(kvm); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1629 | r = create_vcpu_fd(vcpu); |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1630 | if (r < 0) { |
| 1631 | kvm_put_kvm(kvm); |
| 1632 | goto vcpu_destroy; |
| 1633 | } |
| 1634 | |
| 1635 | kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu; |
| 1636 | smp_wmb(); |
| 1637 | atomic_inc(&kvm->online_vcpus); |
| 1638 | |
| 1639 | #ifdef CONFIG_KVM_APIC_ARCHITECTURE |
| 1640 | if (kvm->bsp_vcpu_id == id) |
| 1641 | kvm->bsp_vcpu = vcpu; |
| 1642 | #endif |
| 1643 | mutex_unlock(&kvm->lock); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1644 | return r; |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1645 | |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 1646 | vcpu_destroy: |
Glauber Costa | 7d8fece | 2008-09-17 23:16:59 -0300 | [diff] [blame] | 1647 | mutex_unlock(&kvm->lock); |
Hollis Blanchard | d40ccc6 | 2007-11-19 14:04:43 -0600 | [diff] [blame] | 1648 | kvm_arch_vcpu_destroy(vcpu); |
Avi Kivity | c5ea766 | 2007-02-20 18:41:05 +0200 | [diff] [blame] | 1649 | return r; |
| 1650 | } |
| 1651 | |
Avi Kivity | 1961d27 | 2007-03-05 19:46:05 +0200 | [diff] [blame] | 1652 | static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset) |
| 1653 | { |
| 1654 | if (sigset) { |
| 1655 | sigdelsetmask(sigset, sigmask(SIGKILL)|sigmask(SIGSTOP)); |
| 1656 | vcpu->sigset_active = 1; |
| 1657 | vcpu->sigset = *sigset; |
| 1658 | } else |
| 1659 | vcpu->sigset_active = 0; |
| 1660 | return 0; |
| 1661 | } |
| 1662 | |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1663 | static long kvm_vcpu_ioctl(struct file *filp, |
| 1664 | unsigned int ioctl, unsigned long arg) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1665 | { |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1666 | struct kvm_vcpu *vcpu = filp->private_data; |
Al Viro | 2f366987 | 2007-02-09 16:38:35 +0000 | [diff] [blame] | 1667 | void __user *argp = (void __user *)arg; |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 1668 | int r; |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1669 | struct kvm_fpu *fpu = NULL; |
| 1670 | struct kvm_sregs *kvm_sregs = NULL; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1671 | |
Avi Kivity | 6d4e4c4 | 2007-11-21 16:41:05 +0200 | [diff] [blame] | 1672 | if (vcpu->kvm->mm != current->mm) |
| 1673 | return -EIO; |
Avi Kivity | 2122ff5 | 2010-05-13 11:25:04 +0300 | [diff] [blame] | 1674 | |
| 1675 | #if defined(CONFIG_S390) || defined(CONFIG_PPC) |
| 1676 | /* |
| 1677 | * Special cases: vcpu ioctls that are asynchronous to vcpu execution, |
| 1678 | * so vcpu_load() would break it. |
| 1679 | */ |
| 1680 | if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_INTERRUPT) |
| 1681 | return kvm_arch_vcpu_ioctl(filp, ioctl, arg); |
| 1682 | #endif |
| 1683 | |
| 1684 | |
| 1685 | vcpu_load(vcpu); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1686 | switch (ioctl) { |
Avi Kivity | 9a2bb7f | 2007-02-22 12:58:31 +0200 | [diff] [blame] | 1687 | case KVM_RUN: |
Avi Kivity | f0fe510 | 2007-03-07 13:11:17 +0200 | [diff] [blame] | 1688 | r = -EINVAL; |
| 1689 | if (arg) |
| 1690 | goto out; |
Hollis Blanchard | b6c7a5d | 2007-11-01 14:16:10 -0500 | [diff] [blame] | 1691 | r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run); |
Gleb Natapov | 64be500 | 2010-10-24 16:49:08 +0200 | [diff] [blame] | 1692 | trace_kvm_userspace_exit(vcpu->run->exit_reason, r); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1693 | break; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1694 | case KVM_GET_REGS: { |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1695 | struct kvm_regs *kvm_regs; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1696 | |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1697 | r = -ENOMEM; |
| 1698 | kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL); |
| 1699 | if (!kvm_regs) |
| 1700 | goto out; |
| 1701 | r = kvm_arch_vcpu_ioctl_get_regs(vcpu, kvm_regs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1702 | if (r) |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1703 | goto out_free1; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1704 | r = -EFAULT; |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1705 | if (copy_to_user(argp, kvm_regs, sizeof(struct kvm_regs))) |
| 1706 | goto out_free1; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1707 | r = 0; |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1708 | out_free1: |
| 1709 | kfree(kvm_regs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1710 | break; |
| 1711 | } |
| 1712 | case KVM_SET_REGS: { |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1713 | struct kvm_regs *kvm_regs; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1714 | |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1715 | r = -ENOMEM; |
| 1716 | kvm_regs = kzalloc(sizeof(struct kvm_regs), GFP_KERNEL); |
| 1717 | if (!kvm_regs) |
| 1718 | goto out; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1719 | r = -EFAULT; |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1720 | if (copy_from_user(kvm_regs, argp, sizeof(struct kvm_regs))) |
| 1721 | goto out_free2; |
| 1722 | r = kvm_arch_vcpu_ioctl_set_regs(vcpu, kvm_regs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1723 | if (r) |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1724 | goto out_free2; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1725 | r = 0; |
Xiantao Zhang | 3e4bb3a | 2008-02-25 18:52:20 +0800 | [diff] [blame] | 1726 | out_free2: |
| 1727 | kfree(kvm_regs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1728 | break; |
| 1729 | } |
| 1730 | case KVM_GET_SREGS: { |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1731 | kvm_sregs = kzalloc(sizeof(struct kvm_sregs), GFP_KERNEL); |
| 1732 | r = -ENOMEM; |
| 1733 | if (!kvm_sregs) |
| 1734 | goto out; |
| 1735 | r = kvm_arch_vcpu_ioctl_get_sregs(vcpu, kvm_sregs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1736 | if (r) |
| 1737 | goto out; |
| 1738 | r = -EFAULT; |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1739 | if (copy_to_user(argp, kvm_sregs, sizeof(struct kvm_sregs))) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1740 | goto out; |
| 1741 | r = 0; |
| 1742 | break; |
| 1743 | } |
| 1744 | case KVM_SET_SREGS: { |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1745 | kvm_sregs = kmalloc(sizeof(struct kvm_sregs), GFP_KERNEL); |
| 1746 | r = -ENOMEM; |
| 1747 | if (!kvm_sregs) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1748 | goto out; |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1749 | r = -EFAULT; |
| 1750 | if (copy_from_user(kvm_sregs, argp, sizeof(struct kvm_sregs))) |
| 1751 | goto out; |
| 1752 | r = kvm_arch_vcpu_ioctl_set_sregs(vcpu, kvm_sregs); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1753 | if (r) |
| 1754 | goto out; |
| 1755 | r = 0; |
| 1756 | break; |
| 1757 | } |
Marcelo Tosatti | 62d9f0d | 2008-04-11 13:24:45 -0300 | [diff] [blame] | 1758 | case KVM_GET_MP_STATE: { |
| 1759 | struct kvm_mp_state mp_state; |
| 1760 | |
| 1761 | r = kvm_arch_vcpu_ioctl_get_mpstate(vcpu, &mp_state); |
| 1762 | if (r) |
| 1763 | goto out; |
| 1764 | r = -EFAULT; |
| 1765 | if (copy_to_user(argp, &mp_state, sizeof mp_state)) |
| 1766 | goto out; |
| 1767 | r = 0; |
| 1768 | break; |
| 1769 | } |
| 1770 | case KVM_SET_MP_STATE: { |
| 1771 | struct kvm_mp_state mp_state; |
| 1772 | |
| 1773 | r = -EFAULT; |
| 1774 | if (copy_from_user(&mp_state, argp, sizeof mp_state)) |
| 1775 | goto out; |
| 1776 | r = kvm_arch_vcpu_ioctl_set_mpstate(vcpu, &mp_state); |
| 1777 | if (r) |
| 1778 | goto out; |
| 1779 | r = 0; |
| 1780 | break; |
| 1781 | } |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1782 | case KVM_TRANSLATE: { |
| 1783 | struct kvm_translation tr; |
| 1784 | |
| 1785 | r = -EFAULT; |
Al Viro | 2f366987 | 2007-02-09 16:38:35 +0000 | [diff] [blame] | 1786 | if (copy_from_user(&tr, argp, sizeof tr)) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1787 | goto out; |
Zhang Xiantao | 8b00679 | 2007-11-16 13:05:55 +0800 | [diff] [blame] | 1788 | r = kvm_arch_vcpu_ioctl_translate(vcpu, &tr); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1789 | if (r) |
| 1790 | goto out; |
| 1791 | r = -EFAULT; |
Al Viro | 2f366987 | 2007-02-09 16:38:35 +0000 | [diff] [blame] | 1792 | if (copy_to_user(argp, &tr, sizeof tr)) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1793 | goto out; |
| 1794 | r = 0; |
| 1795 | break; |
| 1796 | } |
Jan Kiszka | d0bfb94 | 2008-12-15 13:52:10 +0100 | [diff] [blame] | 1797 | case KVM_SET_GUEST_DEBUG: { |
| 1798 | struct kvm_guest_debug dbg; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1799 | |
| 1800 | r = -EFAULT; |
Al Viro | 2f366987 | 2007-02-09 16:38:35 +0000 | [diff] [blame] | 1801 | if (copy_from_user(&dbg, argp, sizeof dbg)) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1802 | goto out; |
Jan Kiszka | d0bfb94 | 2008-12-15 13:52:10 +0100 | [diff] [blame] | 1803 | r = kvm_arch_vcpu_ioctl_set_guest_debug(vcpu, &dbg); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1804 | if (r) |
| 1805 | goto out; |
| 1806 | r = 0; |
| 1807 | break; |
| 1808 | } |
Avi Kivity | 1961d27 | 2007-03-05 19:46:05 +0200 | [diff] [blame] | 1809 | case KVM_SET_SIGNAL_MASK: { |
| 1810 | struct kvm_signal_mask __user *sigmask_arg = argp; |
| 1811 | struct kvm_signal_mask kvm_sigmask; |
| 1812 | sigset_t sigset, *p; |
| 1813 | |
| 1814 | p = NULL; |
| 1815 | if (argp) { |
| 1816 | r = -EFAULT; |
| 1817 | if (copy_from_user(&kvm_sigmask, argp, |
| 1818 | sizeof kvm_sigmask)) |
| 1819 | goto out; |
| 1820 | r = -EINVAL; |
| 1821 | if (kvm_sigmask.len != sizeof sigset) |
| 1822 | goto out; |
| 1823 | r = -EFAULT; |
| 1824 | if (copy_from_user(&sigset, sigmask_arg->sigset, |
| 1825 | sizeof sigset)) |
| 1826 | goto out; |
| 1827 | p = &sigset; |
| 1828 | } |
Andi Kleen | 376d41f | 2010-06-10 13:10:47 +0200 | [diff] [blame] | 1829 | r = kvm_vcpu_ioctl_set_sigmask(vcpu, p); |
Avi Kivity | 1961d27 | 2007-03-05 19:46:05 +0200 | [diff] [blame] | 1830 | break; |
| 1831 | } |
Avi Kivity | b883673 | 2007-04-01 16:34:31 +0300 | [diff] [blame] | 1832 | case KVM_GET_FPU: { |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1833 | fpu = kzalloc(sizeof(struct kvm_fpu), GFP_KERNEL); |
| 1834 | r = -ENOMEM; |
| 1835 | if (!fpu) |
| 1836 | goto out; |
| 1837 | r = kvm_arch_vcpu_ioctl_get_fpu(vcpu, fpu); |
Avi Kivity | b883673 | 2007-04-01 16:34:31 +0300 | [diff] [blame] | 1838 | if (r) |
| 1839 | goto out; |
| 1840 | r = -EFAULT; |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1841 | if (copy_to_user(argp, fpu, sizeof(struct kvm_fpu))) |
Avi Kivity | b883673 | 2007-04-01 16:34:31 +0300 | [diff] [blame] | 1842 | goto out; |
| 1843 | r = 0; |
| 1844 | break; |
| 1845 | } |
| 1846 | case KVM_SET_FPU: { |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1847 | fpu = kmalloc(sizeof(struct kvm_fpu), GFP_KERNEL); |
| 1848 | r = -ENOMEM; |
| 1849 | if (!fpu) |
Avi Kivity | b883673 | 2007-04-01 16:34:31 +0300 | [diff] [blame] | 1850 | goto out; |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1851 | r = -EFAULT; |
| 1852 | if (copy_from_user(fpu, argp, sizeof(struct kvm_fpu))) |
| 1853 | goto out; |
| 1854 | r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu); |
Avi Kivity | b883673 | 2007-04-01 16:34:31 +0300 | [diff] [blame] | 1855 | if (r) |
| 1856 | goto out; |
| 1857 | r = 0; |
| 1858 | break; |
| 1859 | } |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1860 | default: |
Carsten Otte | 313a3dc | 2007-10-11 19:16:52 +0200 | [diff] [blame] | 1861 | r = kvm_arch_vcpu_ioctl(filp, ioctl, arg); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1862 | } |
| 1863 | out: |
Avi Kivity | 2122ff5 | 2010-05-13 11:25:04 +0300 | [diff] [blame] | 1864 | vcpu_put(vcpu); |
Dave Hansen | fa3795a | 2008-08-11 10:01:46 -0700 | [diff] [blame] | 1865 | kfree(fpu); |
| 1866 | kfree(kvm_sregs); |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1867 | return r; |
| 1868 | } |
| 1869 | |
| 1870 | static long kvm_vm_ioctl(struct file *filp, |
| 1871 | unsigned int ioctl, unsigned long arg) |
| 1872 | { |
| 1873 | struct kvm *kvm = filp->private_data; |
| 1874 | void __user *argp = (void __user *)arg; |
Carsten Otte | 1fe779f | 2007-10-29 16:08:35 +0100 | [diff] [blame] | 1875 | int r; |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1876 | |
Avi Kivity | 6d4e4c4 | 2007-11-21 16:41:05 +0200 | [diff] [blame] | 1877 | if (kvm->mm != current->mm) |
| 1878 | return -EIO; |
Avi Kivity | bccf215 | 2007-02-21 18:04:26 +0200 | [diff] [blame] | 1879 | switch (ioctl) { |
| 1880 | case KVM_CREATE_VCPU: |
| 1881 | r = kvm_vm_ioctl_create_vcpu(kvm, arg); |
| 1882 | if (r < 0) |
| 1883 | goto out; |
| 1884 | break; |
Izik Eidus | 6fc138d | 2007-10-09 19:20:39 +0200 | [diff] [blame] | 1885 | case KVM_SET_USER_MEMORY_REGION: { |
| 1886 | struct kvm_userspace_memory_region kvm_userspace_mem; |
| 1887 | |
| 1888 | r = -EFAULT; |
| 1889 | if (copy_from_user(&kvm_userspace_mem, argp, |
| 1890 | sizeof kvm_userspace_mem)) |
| 1891 | goto out; |
| 1892 | |
| 1893 | r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem, 1); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1894 | if (r) |
| 1895 | goto out; |
| 1896 | break; |
| 1897 | } |
| 1898 | case KVM_GET_DIRTY_LOG: { |
| 1899 | struct kvm_dirty_log log; |
| 1900 | |
| 1901 | r = -EFAULT; |
Al Viro | 2f366987 | 2007-02-09 16:38:35 +0000 | [diff] [blame] | 1902 | if (copy_from_user(&log, argp, sizeof log)) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1903 | goto out; |
Avi Kivity | 2c6f5df | 2007-02-20 18:27:58 +0200 | [diff] [blame] | 1904 | r = kvm_vm_ioctl_get_dirty_log(kvm, &log); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 1905 | if (r) |
| 1906 | goto out; |
| 1907 | break; |
| 1908 | } |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 1909 | #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET |
| 1910 | case KVM_REGISTER_COALESCED_MMIO: { |
| 1911 | struct kvm_coalesced_mmio_zone zone; |
| 1912 | r = -EFAULT; |
| 1913 | if (copy_from_user(&zone, argp, sizeof zone)) |
| 1914 | goto out; |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 1915 | r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone); |
| 1916 | if (r) |
| 1917 | goto out; |
| 1918 | r = 0; |
| 1919 | break; |
| 1920 | } |
| 1921 | case KVM_UNREGISTER_COALESCED_MMIO: { |
| 1922 | struct kvm_coalesced_mmio_zone zone; |
| 1923 | r = -EFAULT; |
| 1924 | if (copy_from_user(&zone, argp, sizeof zone)) |
| 1925 | goto out; |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 1926 | r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone); |
| 1927 | if (r) |
| 1928 | goto out; |
| 1929 | r = 0; |
| 1930 | break; |
| 1931 | } |
| 1932 | #endif |
Gregory Haskins | 721eecb | 2009-05-20 10:30:49 -0400 | [diff] [blame] | 1933 | case KVM_IRQFD: { |
| 1934 | struct kvm_irqfd data; |
| 1935 | |
| 1936 | r = -EFAULT; |
| 1937 | if (copy_from_user(&data, argp, sizeof data)) |
| 1938 | goto out; |
| 1939 | r = kvm_irqfd(kvm, data.fd, data.gsi, data.flags); |
| 1940 | break; |
| 1941 | } |
Gregory Haskins | d34e6b1 | 2009-07-07 17:08:49 -0400 | [diff] [blame] | 1942 | case KVM_IOEVENTFD: { |
| 1943 | struct kvm_ioeventfd data; |
| 1944 | |
| 1945 | r = -EFAULT; |
| 1946 | if (copy_from_user(&data, argp, sizeof data)) |
| 1947 | goto out; |
| 1948 | r = kvm_ioeventfd(kvm, &data); |
| 1949 | break; |
| 1950 | } |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1951 | #ifdef CONFIG_KVM_APIC_ARCHITECTURE |
| 1952 | case KVM_SET_BOOT_CPU_ID: |
| 1953 | r = 0; |
Marcelo Tosatti | 894a9c5 | 2009-06-23 15:05:14 -0300 | [diff] [blame] | 1954 | mutex_lock(&kvm->lock); |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1955 | if (atomic_read(&kvm->online_vcpus) != 0) |
| 1956 | r = -EBUSY; |
| 1957 | else |
| 1958 | kvm->bsp_vcpu_id = arg; |
Marcelo Tosatti | 894a9c5 | 2009-06-23 15:05:14 -0300 | [diff] [blame] | 1959 | mutex_unlock(&kvm->lock); |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 1960 | break; |
| 1961 | #endif |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 1962 | default: |
Carsten Otte | 1fe779f | 2007-10-29 16:08:35 +0100 | [diff] [blame] | 1963 | r = kvm_arch_vm_ioctl(filp, ioctl, arg); |
Avi Kivity | bfd99ff | 2009-08-26 14:57:50 +0300 | [diff] [blame] | 1964 | if (r == -ENOTTY) |
| 1965 | r = kvm_vm_ioctl_assigned_device(kvm, ioctl, arg); |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 1966 | } |
| 1967 | out: |
| 1968 | return r; |
| 1969 | } |
| 1970 | |
Arnd Bergmann | 6ff5894 | 2009-10-22 14:19:27 +0200 | [diff] [blame] | 1971 | #ifdef CONFIG_COMPAT |
| 1972 | struct compat_kvm_dirty_log { |
| 1973 | __u32 slot; |
| 1974 | __u32 padding1; |
| 1975 | union { |
| 1976 | compat_uptr_t dirty_bitmap; /* one bit per page */ |
| 1977 | __u64 padding2; |
| 1978 | }; |
| 1979 | }; |
| 1980 | |
| 1981 | static long kvm_vm_compat_ioctl(struct file *filp, |
| 1982 | unsigned int ioctl, unsigned long arg) |
| 1983 | { |
| 1984 | struct kvm *kvm = filp->private_data; |
| 1985 | int r; |
| 1986 | |
| 1987 | if (kvm->mm != current->mm) |
| 1988 | return -EIO; |
| 1989 | switch (ioctl) { |
| 1990 | case KVM_GET_DIRTY_LOG: { |
| 1991 | struct compat_kvm_dirty_log compat_log; |
| 1992 | struct kvm_dirty_log log; |
| 1993 | |
| 1994 | r = -EFAULT; |
| 1995 | if (copy_from_user(&compat_log, (void __user *)arg, |
| 1996 | sizeof(compat_log))) |
| 1997 | goto out; |
| 1998 | log.slot = compat_log.slot; |
| 1999 | log.padding1 = compat_log.padding1; |
| 2000 | log.padding2 = compat_log.padding2; |
| 2001 | log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap); |
| 2002 | |
| 2003 | r = kvm_vm_ioctl_get_dirty_log(kvm, &log); |
| 2004 | if (r) |
| 2005 | goto out; |
| 2006 | break; |
| 2007 | } |
| 2008 | default: |
| 2009 | r = kvm_vm_ioctl(filp, ioctl, arg); |
| 2010 | } |
| 2011 | |
| 2012 | out: |
| 2013 | return r; |
| 2014 | } |
| 2015 | #endif |
| 2016 | |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 2017 | static int kvm_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2018 | { |
Marcelo Tosatti | 777b3f4 | 2008-09-16 20:54:46 -0300 | [diff] [blame] | 2019 | struct page *page[1]; |
| 2020 | unsigned long addr; |
| 2021 | int npages; |
| 2022 | gfn_t gfn = vmf->pgoff; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2023 | struct kvm *kvm = vma->vm_file->private_data; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2024 | |
Marcelo Tosatti | 777b3f4 | 2008-09-16 20:54:46 -0300 | [diff] [blame] | 2025 | addr = gfn_to_hva(kvm, gfn); |
| 2026 | if (kvm_is_error_hva(addr)) |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 2027 | return VM_FAULT_SIGBUS; |
Marcelo Tosatti | 777b3f4 | 2008-09-16 20:54:46 -0300 | [diff] [blame] | 2028 | |
| 2029 | npages = get_user_pages(current, current->mm, addr, 1, 1, 0, page, |
| 2030 | NULL); |
| 2031 | if (unlikely(npages != 1)) |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 2032 | return VM_FAULT_SIGBUS; |
Marcelo Tosatti | 777b3f4 | 2008-09-16 20:54:46 -0300 | [diff] [blame] | 2033 | |
| 2034 | vmf->page = page[0]; |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 2035 | return 0; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2036 | } |
| 2037 | |
Alexey Dobriyan | f0f37e2 | 2009-09-27 22:29:37 +0400 | [diff] [blame] | 2038 | static const struct vm_operations_struct kvm_vm_vm_ops = { |
npiggin@suse.de | e4a533a | 2007-12-05 18:15:52 +1100 | [diff] [blame] | 2039 | .fault = kvm_vm_fault, |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2040 | }; |
| 2041 | |
| 2042 | static int kvm_vm_mmap(struct file *file, struct vm_area_struct *vma) |
| 2043 | { |
| 2044 | vma->vm_ops = &kvm_vm_vm_ops; |
| 2045 | return 0; |
| 2046 | } |
| 2047 | |
Christian Borntraeger | 3d3aab1 | 2008-12-02 11:17:32 +0100 | [diff] [blame] | 2048 | static struct file_operations kvm_vm_fops = { |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2049 | .release = kvm_vm_release, |
| 2050 | .unlocked_ioctl = kvm_vm_ioctl, |
Arnd Bergmann | 6ff5894 | 2009-10-22 14:19:27 +0200 | [diff] [blame] | 2051 | #ifdef CONFIG_COMPAT |
| 2052 | .compat_ioctl = kvm_vm_compat_ioctl, |
| 2053 | #endif |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2054 | .mmap = kvm_vm_mmap, |
Arnd Bergmann | 6038f37 | 2010-08-15 18:52:59 +0200 | [diff] [blame] | 2055 | .llseek = noop_llseek, |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2056 | }; |
| 2057 | |
| 2058 | static int kvm_dev_ioctl_create_vm(void) |
| 2059 | { |
Heiko Carstens | aac8763 | 2010-10-27 17:22:10 +0200 | [diff] [blame] | 2060 | int r; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2061 | struct kvm *kvm; |
| 2062 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2063 | kvm = kvm_create_vm(); |
Avi Kivity | d6d2816 | 2007-06-28 08:38:16 -0400 | [diff] [blame] | 2064 | if (IS_ERR(kvm)) |
| 2065 | return PTR_ERR(kvm); |
Takuya Yoshikawa | 6ce5a09 | 2010-03-15 22:13:30 +0900 | [diff] [blame] | 2066 | #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET |
| 2067 | r = kvm_coalesced_mmio_init(kvm); |
| 2068 | if (r < 0) { |
| 2069 | kvm_put_kvm(kvm); |
| 2070 | return r; |
| 2071 | } |
| 2072 | #endif |
Heiko Carstens | aac8763 | 2010-10-27 17:22:10 +0200 | [diff] [blame] | 2073 | r = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR); |
| 2074 | if (r < 0) |
Al Viro | 66c0b39 | 2008-04-19 20:33:56 +0100 | [diff] [blame] | 2075 | kvm_put_kvm(kvm); |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2076 | |
Heiko Carstens | aac8763 | 2010-10-27 17:22:10 +0200 | [diff] [blame] | 2077 | return r; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2078 | } |
| 2079 | |
Avi Kivity | 1a811b6 | 2008-12-08 18:25:27 +0200 | [diff] [blame] | 2080 | static long kvm_dev_ioctl_check_extension_generic(long arg) |
| 2081 | { |
| 2082 | switch (arg) { |
Avi Kivity | ca9edae | 2008-12-08 18:29:29 +0200 | [diff] [blame] | 2083 | case KVM_CAP_USER_MEMORY: |
Avi Kivity | 1a811b6 | 2008-12-08 18:25:27 +0200 | [diff] [blame] | 2084 | case KVM_CAP_DESTROY_MEMORY_REGION_WORKS: |
Jan Kiszka | 4cd481f | 2009-04-13 11:59:32 +0200 | [diff] [blame] | 2085 | case KVM_CAP_JOIN_MEMORY_REGIONS_WORKS: |
Gleb Natapov | 73880c8 | 2009-06-09 15:56:28 +0300 | [diff] [blame] | 2086 | #ifdef CONFIG_KVM_APIC_ARCHITECTURE |
| 2087 | case KVM_CAP_SET_BOOT_CPU_ID: |
| 2088 | #endif |
Avi Kivity | a9c7399 | 2009-11-04 11:54:59 +0200 | [diff] [blame] | 2089 | case KVM_CAP_INTERNAL_ERROR_DATA: |
Avi Kivity | 1a811b6 | 2008-12-08 18:25:27 +0200 | [diff] [blame] | 2090 | return 1; |
Avi Kivity | 399ec80 | 2008-11-19 13:58:46 +0200 | [diff] [blame] | 2091 | #ifdef CONFIG_HAVE_KVM_IRQCHIP |
| 2092 | case KVM_CAP_IRQ_ROUTING: |
Sheng Yang | 3646314 | 2009-03-16 16:33:43 +0800 | [diff] [blame] | 2093 | return KVM_MAX_IRQ_ROUTES; |
Avi Kivity | 399ec80 | 2008-11-19 13:58:46 +0200 | [diff] [blame] | 2094 | #endif |
Avi Kivity | 1a811b6 | 2008-12-08 18:25:27 +0200 | [diff] [blame] | 2095 | default: |
| 2096 | break; |
| 2097 | } |
| 2098 | return kvm_dev_ioctl_check_extension(arg); |
| 2099 | } |
| 2100 | |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2101 | static long kvm_dev_ioctl(struct file *filp, |
| 2102 | unsigned int ioctl, unsigned long arg) |
| 2103 | { |
Avi Kivity | 07c45a3 | 2007-03-07 13:05:38 +0200 | [diff] [blame] | 2104 | long r = -EINVAL; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2105 | |
| 2106 | switch (ioctl) { |
| 2107 | case KVM_GET_API_VERSION: |
Avi Kivity | f0fe510 | 2007-03-07 13:11:17 +0200 | [diff] [blame] | 2108 | r = -EINVAL; |
| 2109 | if (arg) |
| 2110 | goto out; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2111 | r = KVM_API_VERSION; |
| 2112 | break; |
| 2113 | case KVM_CREATE_VM: |
Avi Kivity | f0fe510 | 2007-03-07 13:11:17 +0200 | [diff] [blame] | 2114 | r = -EINVAL; |
| 2115 | if (arg) |
| 2116 | goto out; |
Avi Kivity | f17abe9 | 2007-02-21 19:28:04 +0200 | [diff] [blame] | 2117 | r = kvm_dev_ioctl_create_vm(); |
| 2118 | break; |
Zhang Xiantao | 018d00d | 2007-11-15 23:07:47 +0800 | [diff] [blame] | 2119 | case KVM_CHECK_EXTENSION: |
Avi Kivity | 1a811b6 | 2008-12-08 18:25:27 +0200 | [diff] [blame] | 2120 | r = kvm_dev_ioctl_check_extension_generic(arg); |
Avi Kivity | 5d308f4 | 2007-03-01 17:56:20 +0200 | [diff] [blame] | 2121 | break; |
Avi Kivity | 07c45a3 | 2007-03-07 13:05:38 +0200 | [diff] [blame] | 2122 | case KVM_GET_VCPU_MMAP_SIZE: |
| 2123 | r = -EINVAL; |
| 2124 | if (arg) |
| 2125 | goto out; |
Avi Kivity | adb1ff4 | 2008-01-24 15:13:08 +0200 | [diff] [blame] | 2126 | r = PAGE_SIZE; /* struct kvm_run */ |
| 2127 | #ifdef CONFIG_X86 |
| 2128 | r += PAGE_SIZE; /* pio data page */ |
| 2129 | #endif |
Laurent Vivier | 5f94c17 | 2008-05-30 16:05:54 +0200 | [diff] [blame] | 2130 | #ifdef KVM_COALESCED_MMIO_PAGE_OFFSET |
| 2131 | r += PAGE_SIZE; /* coalesced mmio ring page */ |
| 2132 | #endif |
Avi Kivity | 07c45a3 | 2007-03-07 13:05:38 +0200 | [diff] [blame] | 2133 | break; |
Feng(Eric) Liu | d4c9ff2 | 2008-04-10 08:47:53 -0400 | [diff] [blame] | 2134 | case KVM_TRACE_ENABLE: |
| 2135 | case KVM_TRACE_PAUSE: |
| 2136 | case KVM_TRACE_DISABLE: |
Marcelo Tosatti | 2023a29 | 2009-06-18 11:47:28 -0300 | [diff] [blame] | 2137 | r = -EOPNOTSUPP; |
Feng(Eric) Liu | d4c9ff2 | 2008-04-10 08:47:53 -0400 | [diff] [blame] | 2138 | break; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2139 | default: |
Carsten Otte | 043405e | 2007-10-10 17:16:19 +0200 | [diff] [blame] | 2140 | return kvm_arch_dev_ioctl(filp, ioctl, arg); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2141 | } |
| 2142 | out: |
| 2143 | return r; |
| 2144 | } |
| 2145 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2146 | static struct file_operations kvm_chardev_ops = { |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2147 | .unlocked_ioctl = kvm_dev_ioctl, |
| 2148 | .compat_ioctl = kvm_dev_ioctl, |
Arnd Bergmann | 6038f37 | 2010-08-15 18:52:59 +0200 | [diff] [blame] | 2149 | .llseek = noop_llseek, |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2150 | }; |
| 2151 | |
| 2152 | static struct miscdevice kvm_dev = { |
Avi Kivity | bbe4432 | 2007-03-04 13:27:36 +0200 | [diff] [blame] | 2153 | KVM_MINOR, |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2154 | "kvm", |
| 2155 | &kvm_chardev_ops, |
| 2156 | }; |
| 2157 | |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2158 | static void hardware_enable_nolock(void *junk) |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2159 | { |
| 2160 | int cpu = raw_smp_processor_id(); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2161 | int r; |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2162 | |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2163 | if (cpumask_test_cpu(cpu, cpus_hardware_enabled)) |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2164 | return; |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2165 | |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2166 | cpumask_set_cpu(cpu, cpus_hardware_enabled); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2167 | |
| 2168 | r = kvm_arch_hardware_enable(NULL); |
| 2169 | |
| 2170 | if (r) { |
| 2171 | cpumask_clear_cpu(cpu, cpus_hardware_enabled); |
| 2172 | atomic_inc(&hardware_enable_failed); |
| 2173 | printk(KERN_INFO "kvm: enabling virtualization on " |
| 2174 | "CPU%d failed\n", cpu); |
| 2175 | } |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2176 | } |
| 2177 | |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2178 | static void hardware_enable(void *junk) |
| 2179 | { |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2180 | raw_spin_lock(&kvm_lock); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2181 | hardware_enable_nolock(junk); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2182 | raw_spin_unlock(&kvm_lock); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2183 | } |
| 2184 | |
| 2185 | static void hardware_disable_nolock(void *junk) |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2186 | { |
| 2187 | int cpu = raw_smp_processor_id(); |
| 2188 | |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2189 | if (!cpumask_test_cpu(cpu, cpus_hardware_enabled)) |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2190 | return; |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2191 | cpumask_clear_cpu(cpu, cpus_hardware_enabled); |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2192 | kvm_arch_hardware_disable(NULL); |
Avi Kivity | 1b6c016 | 2007-05-24 13:03:52 +0300 | [diff] [blame] | 2193 | } |
| 2194 | |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2195 | static void hardware_disable(void *junk) |
| 2196 | { |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2197 | raw_spin_lock(&kvm_lock); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2198 | hardware_disable_nolock(junk); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2199 | raw_spin_unlock(&kvm_lock); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2200 | } |
| 2201 | |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2202 | static void hardware_disable_all_nolock(void) |
| 2203 | { |
| 2204 | BUG_ON(!kvm_usage_count); |
| 2205 | |
| 2206 | kvm_usage_count--; |
| 2207 | if (!kvm_usage_count) |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2208 | on_each_cpu(hardware_disable_nolock, NULL, 1); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2209 | } |
| 2210 | |
| 2211 | static void hardware_disable_all(void) |
| 2212 | { |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2213 | raw_spin_lock(&kvm_lock); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2214 | hardware_disable_all_nolock(); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2215 | raw_spin_unlock(&kvm_lock); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2216 | } |
| 2217 | |
| 2218 | static int hardware_enable_all(void) |
| 2219 | { |
| 2220 | int r = 0; |
| 2221 | |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2222 | raw_spin_lock(&kvm_lock); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2223 | |
| 2224 | kvm_usage_count++; |
| 2225 | if (kvm_usage_count == 1) { |
| 2226 | atomic_set(&hardware_enable_failed, 0); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2227 | on_each_cpu(hardware_enable_nolock, NULL, 1); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2228 | |
| 2229 | if (atomic_read(&hardware_enable_failed)) { |
| 2230 | hardware_disable_all_nolock(); |
| 2231 | r = -EBUSY; |
| 2232 | } |
| 2233 | } |
| 2234 | |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2235 | raw_spin_unlock(&kvm_lock); |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2236 | |
| 2237 | return r; |
| 2238 | } |
| 2239 | |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2240 | static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val, |
| 2241 | void *v) |
| 2242 | { |
| 2243 | int cpu = (long)v; |
| 2244 | |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2245 | if (!kvm_usage_count) |
| 2246 | return NOTIFY_OK; |
| 2247 | |
Avi Kivity | 1a6f4d7 | 2007-11-11 18:37:32 +0200 | [diff] [blame] | 2248 | val &= ~CPU_TASKS_FROZEN; |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2249 | switch (val) { |
Avi Kivity | cec9ad2 | 2007-05-24 13:11:41 +0300 | [diff] [blame] | 2250 | case CPU_DYING: |
Avi Kivity | 6ec8a856 | 2007-08-19 15:57:26 +0300 | [diff] [blame] | 2251 | printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n", |
| 2252 | cpu); |
| 2253 | hardware_disable(NULL); |
| 2254 | break; |
Zachary Amsden | da908f2 | 2010-08-19 22:07:27 -1000 | [diff] [blame] | 2255 | case CPU_STARTING: |
Jeremy Katz | 43934a3 | 2007-02-19 14:37:46 +0200 | [diff] [blame] | 2256 | printk(KERN_INFO "kvm: enabling virtualization on CPU%d\n", |
| 2257 | cpu); |
Zachary Amsden | da908f2 | 2010-08-19 22:07:27 -1000 | [diff] [blame] | 2258 | hardware_enable(NULL); |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2259 | break; |
| 2260 | } |
| 2261 | return NOTIFY_OK; |
| 2262 | } |
| 2263 | |
Avi Kivity | 4ecac3f | 2008-05-13 13:23:38 +0300 | [diff] [blame] | 2264 | |
Avi Kivity | b7c4145 | 2010-12-02 17:52:50 +0200 | [diff] [blame] | 2265 | asmlinkage void kvm_spurious_fault(void) |
Avi Kivity | 4ecac3f | 2008-05-13 13:23:38 +0300 | [diff] [blame] | 2266 | { |
Avi Kivity | 4ecac3f | 2008-05-13 13:23:38 +0300 | [diff] [blame] | 2267 | /* Fault while not rebooting. We want the trace. */ |
| 2268 | BUG(); |
| 2269 | } |
Avi Kivity | b7c4145 | 2010-12-02 17:52:50 +0200 | [diff] [blame] | 2270 | EXPORT_SYMBOL_GPL(kvm_spurious_fault); |
Avi Kivity | 4ecac3f | 2008-05-13 13:23:38 +0300 | [diff] [blame] | 2271 | |
Rusty Russell | 9a2b85c | 2007-07-17 23:17:55 +1000 | [diff] [blame] | 2272 | static int kvm_reboot(struct notifier_block *notifier, unsigned long val, |
Mike Day | d77c26f | 2007-10-08 09:02:08 -0400 | [diff] [blame] | 2273 | void *v) |
Rusty Russell | 9a2b85c | 2007-07-17 23:17:55 +1000 | [diff] [blame] | 2274 | { |
Sheng Yang | 8e1c181 | 2009-04-29 11:09:04 +0800 | [diff] [blame] | 2275 | /* |
| 2276 | * Some (well, at least mine) BIOSes hang on reboot if |
| 2277 | * in vmx root mode. |
| 2278 | * |
| 2279 | * And Intel TXT required VMX off for all cpu when system shutdown. |
| 2280 | */ |
| 2281 | printk(KERN_INFO "kvm: exiting hardware virtualization\n"); |
| 2282 | kvm_rebooting = true; |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2283 | on_each_cpu(hardware_disable_nolock, NULL, 1); |
Rusty Russell | 9a2b85c | 2007-07-17 23:17:55 +1000 | [diff] [blame] | 2284 | return NOTIFY_OK; |
| 2285 | } |
| 2286 | |
| 2287 | static struct notifier_block kvm_reboot_notifier = { |
| 2288 | .notifier_call = kvm_reboot, |
| 2289 | .priority = 0, |
| 2290 | }; |
| 2291 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2292 | static void kvm_io_bus_destroy(struct kvm_io_bus *bus) |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2293 | { |
| 2294 | int i; |
| 2295 | |
| 2296 | for (i = 0; i < bus->dev_count; i++) { |
| 2297 | struct kvm_io_device *pos = bus->devs[i]; |
| 2298 | |
| 2299 | kvm_iodevice_destructor(pos); |
| 2300 | } |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2301 | kfree(bus); |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2302 | } |
| 2303 | |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2304 | /* kvm_io_bus_write - called under kvm->slots_lock */ |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2305 | int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2306 | int len, const void *val) |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2307 | { |
| 2308 | int i; |
Lai Jiangshan | 90d83dc | 2010-04-19 17:41:23 +0800 | [diff] [blame] | 2309 | struct kvm_io_bus *bus; |
| 2310 | |
| 2311 | bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu); |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2312 | for (i = 0; i < bus->dev_count; i++) |
| 2313 | if (!kvm_iodevice_write(bus->devs[i], addr, len, val)) |
| 2314 | return 0; |
| 2315 | return -EOPNOTSUPP; |
| 2316 | } |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2317 | |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2318 | /* kvm_io_bus_read - called under kvm->slots_lock */ |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2319 | int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr, |
| 2320 | int len, void *val) |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2321 | { |
| 2322 | int i; |
Lai Jiangshan | 90d83dc | 2010-04-19 17:41:23 +0800 | [diff] [blame] | 2323 | struct kvm_io_bus *bus; |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2324 | |
Lai Jiangshan | 90d83dc | 2010-04-19 17:41:23 +0800 | [diff] [blame] | 2325 | bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu); |
Michael S. Tsirkin | bda9020 | 2009-06-29 22:24:32 +0300 | [diff] [blame] | 2326 | for (i = 0; i < bus->dev_count; i++) |
| 2327 | if (!kvm_iodevice_read(bus->devs[i], addr, len, val)) |
| 2328 | return 0; |
| 2329 | return -EOPNOTSUPP; |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2330 | } |
| 2331 | |
Marcelo Tosatti | 79fac95 | 2009-12-23 14:35:26 -0200 | [diff] [blame] | 2332 | /* Caller must hold slots_lock. */ |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2333 | int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, |
| 2334 | struct kvm_io_device *dev) |
Michael S. Tsirkin | 6c47469 | 2009-06-29 22:24:26 +0300 | [diff] [blame] | 2335 | { |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2336 | struct kvm_io_bus *new_bus, *bus; |
Gregory Haskins | 090b7af | 2009-07-07 17:08:44 -0400 | [diff] [blame] | 2337 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2338 | bus = kvm->buses[bus_idx]; |
Gregory Haskins | 090b7af | 2009-07-07 17:08:44 -0400 | [diff] [blame] | 2339 | if (bus->dev_count > NR_IOBUS_DEVS-1) |
| 2340 | return -ENOSPC; |
| 2341 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2342 | new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL); |
| 2343 | if (!new_bus) |
| 2344 | return -ENOMEM; |
| 2345 | memcpy(new_bus, bus, sizeof(struct kvm_io_bus)); |
| 2346 | new_bus->devs[new_bus->dev_count++] = dev; |
| 2347 | rcu_assign_pointer(kvm->buses[bus_idx], new_bus); |
| 2348 | synchronize_srcu_expedited(&kvm->srcu); |
| 2349 | kfree(bus); |
Gregory Haskins | 090b7af | 2009-07-07 17:08:44 -0400 | [diff] [blame] | 2350 | |
| 2351 | return 0; |
| 2352 | } |
| 2353 | |
Marcelo Tosatti | 79fac95 | 2009-12-23 14:35:26 -0200 | [diff] [blame] | 2354 | /* Caller must hold slots_lock. */ |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2355 | int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx, |
| 2356 | struct kvm_io_device *dev) |
Gregory Haskins | 090b7af | 2009-07-07 17:08:44 -0400 | [diff] [blame] | 2357 | { |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2358 | int i, r; |
| 2359 | struct kvm_io_bus *new_bus, *bus; |
Michael S. Tsirkin | 6c47469 | 2009-06-29 22:24:26 +0300 | [diff] [blame] | 2360 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2361 | new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL); |
| 2362 | if (!new_bus) |
| 2363 | return -ENOMEM; |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2364 | |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2365 | bus = kvm->buses[bus_idx]; |
| 2366 | memcpy(new_bus, bus, sizeof(struct kvm_io_bus)); |
| 2367 | |
| 2368 | r = -ENOENT; |
| 2369 | for (i = 0; i < new_bus->dev_count; i++) |
| 2370 | if (new_bus->devs[i] == dev) { |
| 2371 | r = 0; |
| 2372 | new_bus->devs[i] = new_bus->devs[--new_bus->dev_count]; |
Gregory Haskins | 090b7af | 2009-07-07 17:08:44 -0400 | [diff] [blame] | 2373 | break; |
| 2374 | } |
Marcelo Tosatti | e93f8a0 | 2009-12-23 14:35:24 -0200 | [diff] [blame] | 2375 | |
| 2376 | if (r) { |
| 2377 | kfree(new_bus); |
| 2378 | return r; |
| 2379 | } |
| 2380 | |
| 2381 | rcu_assign_pointer(kvm->buses[bus_idx], new_bus); |
| 2382 | synchronize_srcu_expedited(&kvm->srcu); |
| 2383 | kfree(bus); |
| 2384 | return r; |
Gregory Haskins | 2eeb2e9 | 2007-05-31 14:08:53 -0400 | [diff] [blame] | 2385 | } |
| 2386 | |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2387 | static struct notifier_block kvm_cpu_notifier = { |
| 2388 | .notifier_call = kvm_cpu_hotplug, |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2389 | }; |
| 2390 | |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2391 | static int vm_stat_get(void *_offset, u64 *val) |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2392 | { |
| 2393 | unsigned offset = (long)_offset; |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2394 | struct kvm *kvm; |
| 2395 | |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2396 | *val = 0; |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2397 | raw_spin_lock(&kvm_lock); |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2398 | list_for_each_entry(kvm, &vm_list, vm_list) |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2399 | *val += *(u32 *)((void *)kvm + offset); |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2400 | raw_spin_unlock(&kvm_lock); |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2401 | return 0; |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2402 | } |
| 2403 | |
| 2404 | DEFINE_SIMPLE_ATTRIBUTE(vm_stat_fops, vm_stat_get, NULL, "%llu\n"); |
| 2405 | |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2406 | static int vcpu_stat_get(void *_offset, u64 *val) |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2407 | { |
| 2408 | unsigned offset = (long)_offset; |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2409 | struct kvm *kvm; |
| 2410 | struct kvm_vcpu *vcpu; |
| 2411 | int i; |
| 2412 | |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2413 | *val = 0; |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2414 | raw_spin_lock(&kvm_lock); |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2415 | list_for_each_entry(kvm, &vm_list, vm_list) |
Gleb Natapov | 988a2ca | 2009-06-09 15:56:29 +0300 | [diff] [blame] | 2416 | kvm_for_each_vcpu(i, vcpu, kvm) |
| 2417 | *val += *(u32 *)((void *)vcpu + offset); |
| 2418 | |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2419 | raw_spin_unlock(&kvm_lock); |
Christoph Hellwig | 8b88b09 | 2008-02-08 04:20:26 -0800 | [diff] [blame] | 2420 | return 0; |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2421 | } |
| 2422 | |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2423 | DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n"); |
| 2424 | |
Alexey Dobriyan | 828c095 | 2009-10-01 15:43:56 -0700 | [diff] [blame] | 2425 | static const struct file_operations *stat_fops[] = { |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2426 | [KVM_STAT_VCPU] = &vcpu_stat_fops, |
| 2427 | [KVM_STAT_VM] = &vm_stat_fops, |
| 2428 | }; |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2429 | |
Zhang Xiantao | a16b043 | 2007-11-16 14:38:21 +0800 | [diff] [blame] | 2430 | static void kvm_init_debug(void) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2431 | { |
| 2432 | struct kvm_stats_debugfs_item *p; |
| 2433 | |
Hollis Blanchard | 76f7c87 | 2008-04-15 16:05:42 -0500 | [diff] [blame] | 2434 | kvm_debugfs_dir = debugfs_create_dir("kvm", NULL); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2435 | for (p = debugfs_entries; p->name; ++p) |
Hollis Blanchard | 76f7c87 | 2008-04-15 16:05:42 -0500 | [diff] [blame] | 2436 | p->dentry = debugfs_create_file(p->name, 0444, kvm_debugfs_dir, |
Avi Kivity | 1165f5f | 2007-04-19 17:27:43 +0300 | [diff] [blame] | 2437 | (void *)(long)p->offset, |
Avi Kivity | ba1389b | 2007-11-18 16:24:12 +0200 | [diff] [blame] | 2438 | stat_fops[p->kind]); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2439 | } |
| 2440 | |
| 2441 | static void kvm_exit_debug(void) |
| 2442 | { |
| 2443 | struct kvm_stats_debugfs_item *p; |
| 2444 | |
| 2445 | for (p = debugfs_entries; p->name; ++p) |
| 2446 | debugfs_remove(p->dentry); |
Hollis Blanchard | 76f7c87 | 2008-04-15 16:05:42 -0500 | [diff] [blame] | 2447 | debugfs_remove(kvm_debugfs_dir); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2448 | } |
| 2449 | |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2450 | static int kvm_suspend(struct sys_device *dev, pm_message_t state) |
| 2451 | { |
Alexander Graf | 10474ae | 2009-09-15 11:37:46 +0200 | [diff] [blame] | 2452 | if (kvm_usage_count) |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2453 | hardware_disable_nolock(NULL); |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2454 | return 0; |
| 2455 | } |
| 2456 | |
| 2457 | static int kvm_resume(struct sys_device *dev) |
| 2458 | { |
Zachary Amsden | ca84d1a | 2010-08-19 22:07:28 -1000 | [diff] [blame] | 2459 | if (kvm_usage_count) { |
Jan Kiszka | e935b83 | 2011-02-08 12:55:33 +0100 | [diff] [blame] | 2460 | WARN_ON(raw_spin_is_locked(&kvm_lock)); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2461 | hardware_enable_nolock(NULL); |
Zachary Amsden | ca84d1a | 2010-08-19 22:07:28 -1000 | [diff] [blame] | 2462 | } |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2463 | return 0; |
| 2464 | } |
| 2465 | |
| 2466 | static struct sysdev_class kvm_sysdev_class = { |
Kay Sievers | af5ca3f | 2007-12-20 02:09:39 +0100 | [diff] [blame] | 2467 | .name = "kvm", |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2468 | .suspend = kvm_suspend, |
| 2469 | .resume = kvm_resume, |
| 2470 | }; |
| 2471 | |
| 2472 | static struct sys_device kvm_sysdev = { |
| 2473 | .id = 0, |
| 2474 | .cls = &kvm_sysdev_class, |
| 2475 | }; |
| 2476 | |
Izik Eidus | cea7bb2 | 2007-10-17 19:17:48 +0200 | [diff] [blame] | 2477 | struct page *bad_page; |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 2478 | pfn_t bad_pfn; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2479 | |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 2480 | static inline |
| 2481 | struct kvm_vcpu *preempt_notifier_to_vcpu(struct preempt_notifier *pn) |
| 2482 | { |
| 2483 | return container_of(pn, struct kvm_vcpu, preempt_notifier); |
| 2484 | } |
| 2485 | |
| 2486 | static void kvm_sched_in(struct preempt_notifier *pn, int cpu) |
| 2487 | { |
| 2488 | struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); |
| 2489 | |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2490 | kvm_arch_vcpu_load(vcpu, cpu); |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 2491 | } |
| 2492 | |
| 2493 | static void kvm_sched_out(struct preempt_notifier *pn, |
| 2494 | struct task_struct *next) |
| 2495 | { |
| 2496 | struct kvm_vcpu *vcpu = preempt_notifier_to_vcpu(pn); |
| 2497 | |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2498 | kvm_arch_vcpu_put(vcpu); |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 2499 | } |
| 2500 | |
Avi Kivity | 0ee75be | 2010-04-28 15:39:01 +0300 | [diff] [blame] | 2501 | int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align, |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2502 | struct module *module) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2503 | { |
| 2504 | int r; |
Yang, Sheng | 002c7f7 | 2007-07-31 14:23:01 +0300 | [diff] [blame] | 2505 | int cpu; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2506 | |
Zhang Xiantao | f8c16bb | 2007-11-14 20:40:21 +0800 | [diff] [blame] | 2507 | r = kvm_arch_init(opaque); |
| 2508 | if (r) |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2509 | goto out_fail; |
Zhang Xiantao | cb498ea | 2007-11-14 20:39:31 +0800 | [diff] [blame] | 2510 | |
| 2511 | bad_page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 2512 | |
| 2513 | if (bad_page == NULL) { |
| 2514 | r = -ENOMEM; |
| 2515 | goto out; |
| 2516 | } |
| 2517 | |
Anthony Liguori | 35149e2 | 2008-04-02 14:46:56 -0500 | [diff] [blame] | 2518 | bad_pfn = page_to_pfn(bad_page); |
| 2519 | |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 2520 | hwpoison_page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 2521 | |
| 2522 | if (hwpoison_page == NULL) { |
| 2523 | r = -ENOMEM; |
| 2524 | goto out_free_0; |
| 2525 | } |
| 2526 | |
| 2527 | hwpoison_pfn = page_to_pfn(hwpoison_page); |
| 2528 | |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 2529 | fault_page = alloc_page(GFP_KERNEL | __GFP_ZERO); |
| 2530 | |
| 2531 | if (fault_page == NULL) { |
| 2532 | r = -ENOMEM; |
| 2533 | goto out_free_0; |
| 2534 | } |
| 2535 | |
| 2536 | fault_pfn = page_to_pfn(fault_page); |
| 2537 | |
Avi Kivity | 8437a61 | 2009-06-06 14:52:35 -0700 | [diff] [blame] | 2538 | if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) { |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2539 | r = -ENOMEM; |
| 2540 | goto out_free_0; |
| 2541 | } |
| 2542 | |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2543 | r = kvm_arch_hardware_setup(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2544 | if (r < 0) |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2545 | goto out_free_0a; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2546 | |
Yang, Sheng | 002c7f7 | 2007-07-31 14:23:01 +0300 | [diff] [blame] | 2547 | for_each_online_cpu(cpu) { |
| 2548 | smp_call_function_single(cpu, |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2549 | kvm_arch_check_processor_compat, |
Jens Axboe | 8691e5a | 2008-06-06 11:18:06 +0200 | [diff] [blame] | 2550 | &r, 1); |
Yang, Sheng | 002c7f7 | 2007-07-31 14:23:01 +0300 | [diff] [blame] | 2551 | if (r < 0) |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2552 | goto out_free_1; |
Yang, Sheng | 002c7f7 | 2007-07-31 14:23:01 +0300 | [diff] [blame] | 2553 | } |
| 2554 | |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2555 | r = register_cpu_notifier(&kvm_cpu_notifier); |
| 2556 | if (r) |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2557 | goto out_free_2; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2558 | register_reboot_notifier(&kvm_reboot_notifier); |
| 2559 | |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2560 | r = sysdev_class_register(&kvm_sysdev_class); |
| 2561 | if (r) |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2562 | goto out_free_3; |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2563 | |
| 2564 | r = sysdev_register(&kvm_sysdev); |
| 2565 | if (r) |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2566 | goto out_free_4; |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2567 | |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2568 | /* A kmem cache lets us meet the alignment requirements of fx_save. */ |
Avi Kivity | 0ee75be | 2010-04-28 15:39:01 +0300 | [diff] [blame] | 2569 | if (!vcpu_align) |
| 2570 | vcpu_align = __alignof__(struct kvm_vcpu); |
| 2571 | kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align, |
Joe Perches | 56919c5 | 2007-11-12 20:06:51 -0800 | [diff] [blame] | 2572 | 0, NULL); |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2573 | if (!kvm_vcpu_cache) { |
| 2574 | r = -ENOMEM; |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2575 | goto out_free_5; |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2576 | } |
| 2577 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 2578 | r = kvm_async_pf_init(); |
| 2579 | if (r) |
| 2580 | goto out_free; |
| 2581 | |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2582 | kvm_chardev_ops.owner = module; |
Christian Borntraeger | 3d3aab1 | 2008-12-02 11:17:32 +0100 | [diff] [blame] | 2583 | kvm_vm_fops.owner = module; |
| 2584 | kvm_vcpu_fops.owner = module; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2585 | |
| 2586 | r = misc_register(&kvm_dev); |
| 2587 | if (r) { |
Mike Day | d77c26f | 2007-10-08 09:02:08 -0400 | [diff] [blame] | 2588 | printk(KERN_ERR "kvm: misc device register failed\n"); |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 2589 | goto out_unreg; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2590 | } |
| 2591 | |
Avi Kivity | 15ad714 | 2007-07-11 18:17:21 +0300 | [diff] [blame] | 2592 | kvm_preempt_ops.sched_in = kvm_sched_in; |
| 2593 | kvm_preempt_ops.sched_out = kvm_sched_out; |
| 2594 | |
Darrick J. Wong | 0ea4ed8 | 2009-10-14 16:21:00 -0700 | [diff] [blame] | 2595 | kvm_init_debug(); |
| 2596 | |
Avi Kivity | c7addb9 | 2007-09-16 18:58:32 +0200 | [diff] [blame] | 2597 | return 0; |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2598 | |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 2599 | out_unreg: |
| 2600 | kvm_async_pf_deinit(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2601 | out_free: |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2602 | kmem_cache_destroy(kvm_vcpu_cache); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2603 | out_free_5: |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2604 | sysdev_unregister(&kvm_sysdev); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2605 | out_free_4: |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2606 | sysdev_class_unregister(&kvm_sysdev_class); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2607 | out_free_3: |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2608 | unregister_reboot_notifier(&kvm_reboot_notifier); |
Avi Kivity | 774c47f | 2007-02-12 00:54:47 -0800 | [diff] [blame] | 2609 | unregister_cpu_notifier(&kvm_cpu_notifier); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2610 | out_free_2: |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2611 | out_free_1: |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2612 | kvm_arch_hardware_unsetup(); |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2613 | out_free_0a: |
| 2614 | free_cpumask_var(cpus_hardware_enabled); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2615 | out_free_0: |
Gleb Natapov | edba23e | 2010-07-07 20:16:45 +0300 | [diff] [blame] | 2616 | if (fault_page) |
| 2617 | __free_page(fault_page); |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 2618 | if (hwpoison_page) |
| 2619 | __free_page(hwpoison_page); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2620 | __free_page(bad_page); |
Avi Kivity | ca45aaa | 2007-03-01 19:21:03 +0200 | [diff] [blame] | 2621 | out: |
Zhang Xiantao | f8c16bb | 2007-11-14 20:40:21 +0800 | [diff] [blame] | 2622 | kvm_arch_exit(); |
Zhang Xiantao | d2308784 | 2007-11-29 15:35:39 +0800 | [diff] [blame] | 2623 | out_fail: |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2624 | return r; |
| 2625 | } |
Zhang Xiantao | cb498ea | 2007-11-14 20:39:31 +0800 | [diff] [blame] | 2626 | EXPORT_SYMBOL_GPL(kvm_init); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2627 | |
Zhang Xiantao | cb498ea | 2007-11-14 20:39:31 +0800 | [diff] [blame] | 2628 | void kvm_exit(void) |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2629 | { |
Darrick J. Wong | 0ea4ed8 | 2009-10-14 16:21:00 -0700 | [diff] [blame] | 2630 | kvm_exit_debug(); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2631 | misc_deregister(&kvm_dev); |
Rusty Russell | c16f862 | 2007-07-30 21:12:19 +1000 | [diff] [blame] | 2632 | kmem_cache_destroy(kvm_vcpu_cache); |
Gleb Natapov | af585b9 | 2010-10-14 11:22:46 +0200 | [diff] [blame] | 2633 | kvm_async_pf_deinit(); |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2634 | sysdev_unregister(&kvm_sysdev); |
| 2635 | sysdev_class_unregister(&kvm_sysdev_class); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2636 | unregister_reboot_notifier(&kvm_reboot_notifier); |
Avi Kivity | 59ae6c6 | 2007-02-12 00:54:48 -0800 | [diff] [blame] | 2637 | unregister_cpu_notifier(&kvm_cpu_notifier); |
Takuya Yoshikawa | 75b7127 | 2010-11-16 17:37:41 +0900 | [diff] [blame] | 2638 | on_each_cpu(hardware_disable_nolock, NULL, 1); |
Zhang Xiantao | e9b11c1 | 2007-11-14 20:38:21 +0800 | [diff] [blame] | 2639 | kvm_arch_hardware_unsetup(); |
Zhang Xiantao | f8c16bb | 2007-11-14 20:40:21 +0800 | [diff] [blame] | 2640 | kvm_arch_exit(); |
Rusty Russell | 7f59f49 | 2008-12-07 21:25:45 +1030 | [diff] [blame] | 2641 | free_cpumask_var(cpus_hardware_enabled); |
Huang Ying | bf99815 | 2010-05-31 14:28:19 +0800 | [diff] [blame] | 2642 | __free_page(hwpoison_page); |
Izik Eidus | cea7bb2 | 2007-10-17 19:17:48 +0200 | [diff] [blame] | 2643 | __free_page(bad_page); |
Avi Kivity | 6aa8b73 | 2006-12-10 02:21:36 -0800 | [diff] [blame] | 2644 | } |
Zhang Xiantao | cb498ea | 2007-11-14 20:39:31 +0800 | [diff] [blame] | 2645 | EXPORT_SYMBOL_GPL(kvm_exit); |