blob: 0c8dc9614e7d30790c3c5365541b0251d9f75452 [file] [log] [blame]
Carsten Otte043405e2007-10-10 17:16:19 +02001/*
2 * Kernel-based Virtual Machine driver for Linux
3 *
4 * derived from drivers/kvm/kvm_main.c
5 *
6 * Copyright (C) 2006 Qumranet, Inc.
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +03007 * Copyright (C) 2008 Qumranet, Inc.
8 * Copyright IBM Corporation, 2008
Avi Kivity221d0592010-05-23 18:37:00 +03009 * Copyright 2010 Red Hat, Inc. and/or its affilates.
Carsten Otte043405e2007-10-10 17:16:19 +020010 *
11 * Authors:
12 * Avi Kivity <avi@qumranet.com>
13 * Yaniv Kamay <yaniv@qumranet.com>
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +030014 * Amit Shah <amit.shah@qumranet.com>
15 * Ben-Ami Yassour <benami@il.ibm.com>
Carsten Otte043405e2007-10-10 17:16:19 +020016 *
17 * This work is licensed under the terms of the GNU GPL, version 2. See
18 * the COPYING file in the top-level directory.
19 *
20 */
21
Avi Kivityedf88412007-12-16 11:02:48 +020022#include <linux/kvm_host.h>
Carsten Otte313a3dc2007-10-11 19:16:52 +020023#include "irq.h"
Zhang Xiantao1d737c82007-12-14 09:35:10 +080024#include "mmu.h"
Sheng Yang78376992008-01-28 05:10:22 +080025#include "i8254.h"
Izik Eidus37817f22008-03-24 23:14:53 +020026#include "tss.h"
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -030027#include "kvm_cache_regs.h"
Avi Kivity26eef702008-07-03 14:59:22 +030028#include "x86.h"
Carsten Otte313a3dc2007-10-11 19:16:52 +020029
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -020030#include <linux/clocksource.h>
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +030031#include <linux/interrupt.h>
Carsten Otte313a3dc2007-10-11 19:16:52 +020032#include <linux/kvm.h>
33#include <linux/fs.h>
34#include <linux/vmalloc.h>
Carsten Otte5fb76f92007-10-29 16:08:51 +010035#include <linux/module.h>
Zhang Xiantao0de10342007-11-20 16:25:04 +080036#include <linux/mman.h>
Marcelo Tosatti2bacc552007-12-12 10:46:12 -050037#include <linux/highmem.h>
Joerg Roedel19de40a2008-12-03 14:43:34 +010038#include <linux/iommu.h>
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +030039#include <linux/intel-iommu.h>
Gerd Hoffmannc8076602009-02-04 17:52:04 +010040#include <linux/cpufreq.h>
Avi Kivity18863bd2009-09-07 11:12:18 +030041#include <linux/user-return-notifier.h>
Marcelo Tosattia983fb22009-12-23 14:35:23 -020042#include <linux/srcu.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090043#include <linux/slab.h>
Zhang, Yanminff9d07a2010-04-19 13:32:45 +080044#include <linux/perf_event.h>
Lai Jiangshan7bee3422010-06-02 17:06:03 +080045#include <linux/uaccess.h>
Avi Kivityaec51dc2009-07-01 16:01:02 +030046#include <trace/events/kvm.h>
Xiao Guangrong2ed152a2010-03-10 19:00:43 +080047
Marcelo Tosatti229456f2009-06-17 09:22:14 -030048#define CREATE_TRACE_POINTS
49#include "trace.h"
Carsten Otte043405e2007-10-10 17:16:19 +020050
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +020051#include <asm/debugreg.h>
Zhang Xiantaod825ed02007-11-14 20:08:51 +080052#include <asm/msr.h>
Avi Kivitya5f61302008-02-20 17:57:21 +020053#include <asm/desc.h>
Sheng Yang0bed3b52008-10-09 16:01:54 +080054#include <asm/mtrr.h>
Huang Ying890ca9a2009-05-11 16:48:15 +080055#include <asm/mce.h>
Sheng Yang7cf30852010-05-17 17:08:27 +080056#include <asm/i387.h>
Sheng Yang98918832010-05-17 17:08:28 +080057#include <asm/xcr.h>
Carsten Otte043405e2007-10-10 17:16:19 +020058
Carsten Otte313a3dc2007-10-11 19:16:52 +020059#define MAX_IO_MSRS 256
Carsten Ottea03490e2007-10-29 16:09:35 +010060#define CR0_RESERVED_BITS \
61 (~(unsigned long)(X86_CR0_PE | X86_CR0_MP | X86_CR0_EM | X86_CR0_TS \
62 | X86_CR0_ET | X86_CR0_NE | X86_CR0_WP | X86_CR0_AM \
63 | X86_CR0_NW | X86_CR0_CD | X86_CR0_PG))
64#define CR4_RESERVED_BITS \
65 (~(unsigned long)(X86_CR4_VME | X86_CR4_PVI | X86_CR4_TSD | X86_CR4_DE\
66 | X86_CR4_PSE | X86_CR4_PAE | X86_CR4_MCE \
67 | X86_CR4_PGE | X86_CR4_PCE | X86_CR4_OSFXSR \
Dexuan Cui2acf9232010-06-10 11:27:12 +080068 | X86_CR4_OSXSAVE \
Carsten Ottea03490e2007-10-29 16:09:35 +010069 | X86_CR4_OSXMMEXCPT | X86_CR4_VMXE))
70
71#define CR8_RESERVED_BITS (~(unsigned long)X86_CR8_TPR)
Huang Ying890ca9a2009-05-11 16:48:15 +080072
73#define KVM_MAX_MCE_BANKS 32
74#define KVM_MCE_CAP_SUPPORTED MCG_CTL_P
75
Joerg Roedel50a37eb2008-01-31 14:57:38 +010076/* EFER defaults:
77 * - enable syscall per default because its emulated by KVM
78 * - enable LME and LMA per default on 64 bit KVM
79 */
80#ifdef CONFIG_X86_64
81static u64 __read_mostly efer_reserved_bits = 0xfffffffffffffafeULL;
82#else
83static u64 __read_mostly efer_reserved_bits = 0xfffffffffffffffeULL;
84#endif
Carsten Otte313a3dc2007-10-11 19:16:52 +020085
Avi Kivityba1389b2007-11-18 16:24:12 +020086#define VM_STAT(x) offsetof(struct kvm, stat.x), KVM_STAT_VM
87#define VCPU_STAT(x) offsetof(struct kvm_vcpu, stat.x), KVM_STAT_VCPU
Hollis Blanchard417bc302007-10-31 17:24:23 -050088
Gleb Natapovcb142eb2009-08-09 15:17:40 +030089static void update_cr8_intercept(struct kvm_vcpu *vcpu);
Avi Kivity674eea02008-02-11 18:37:23 +020090static int kvm_dev_ioctl_get_supported_cpuid(struct kvm_cpuid2 *cpuid,
91 struct kvm_cpuid_entry2 __user *entries);
92
Zhang Xiantao97896d02007-11-14 20:09:30 +080093struct kvm_x86_ops *kvm_x86_ops;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -030094EXPORT_SYMBOL_GPL(kvm_x86_ops);
Zhang Xiantao97896d02007-11-14 20:09:30 +080095
Andre Przywaraed85c062009-06-25 12:36:49 +020096int ignore_msrs = 0;
97module_param_named(ignore_msrs, ignore_msrs, bool, S_IRUGO | S_IWUSR);
98
Avi Kivity18863bd2009-09-07 11:12:18 +030099#define KVM_NR_SHARED_MSRS 16
100
101struct kvm_shared_msrs_global {
102 int nr;
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800103 u32 msrs[KVM_NR_SHARED_MSRS];
Avi Kivity18863bd2009-09-07 11:12:18 +0300104};
105
106struct kvm_shared_msrs {
107 struct user_return_notifier urn;
108 bool registered;
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800109 struct kvm_shared_msr_values {
110 u64 host;
111 u64 curr;
112 } values[KVM_NR_SHARED_MSRS];
Avi Kivity18863bd2009-09-07 11:12:18 +0300113};
114
115static struct kvm_shared_msrs_global __read_mostly shared_msrs_global;
116static DEFINE_PER_CPU(struct kvm_shared_msrs, shared_msrs);
117
Hollis Blanchard417bc302007-10-31 17:24:23 -0500118struct kvm_stats_debugfs_item debugfs_entries[] = {
Avi Kivityba1389b2007-11-18 16:24:12 +0200119 { "pf_fixed", VCPU_STAT(pf_fixed) },
120 { "pf_guest", VCPU_STAT(pf_guest) },
121 { "tlb_flush", VCPU_STAT(tlb_flush) },
122 { "invlpg", VCPU_STAT(invlpg) },
123 { "exits", VCPU_STAT(exits) },
124 { "io_exits", VCPU_STAT(io_exits) },
125 { "mmio_exits", VCPU_STAT(mmio_exits) },
126 { "signal_exits", VCPU_STAT(signal_exits) },
127 { "irq_window", VCPU_STAT(irq_window_exits) },
Sheng Yangf08864b2008-05-15 18:23:25 +0800128 { "nmi_window", VCPU_STAT(nmi_window_exits) },
Avi Kivityba1389b2007-11-18 16:24:12 +0200129 { "halt_exits", VCPU_STAT(halt_exits) },
130 { "halt_wakeup", VCPU_STAT(halt_wakeup) },
Amit Shahf11c3a82008-02-21 01:00:30 +0530131 { "hypercalls", VCPU_STAT(hypercalls) },
Avi Kivityba1389b2007-11-18 16:24:12 +0200132 { "request_irq", VCPU_STAT(request_irq_exits) },
133 { "irq_exits", VCPU_STAT(irq_exits) },
134 { "host_state_reload", VCPU_STAT(host_state_reload) },
135 { "efer_reload", VCPU_STAT(efer_reload) },
136 { "fpu_reload", VCPU_STAT(fpu_reload) },
137 { "insn_emulation", VCPU_STAT(insn_emulation) },
138 { "insn_emulation_fail", VCPU_STAT(insn_emulation_fail) },
Avi Kivityfa89a812008-09-01 15:57:51 +0300139 { "irq_injections", VCPU_STAT(irq_injections) },
Jan Kiszkac4abb7c2008-09-26 09:30:55 +0200140 { "nmi_injections", VCPU_STAT(nmi_injections) },
Avi Kivity4cee5762007-11-18 16:37:07 +0200141 { "mmu_shadow_zapped", VM_STAT(mmu_shadow_zapped) },
142 { "mmu_pte_write", VM_STAT(mmu_pte_write) },
143 { "mmu_pte_updated", VM_STAT(mmu_pte_updated) },
144 { "mmu_pde_zapped", VM_STAT(mmu_pde_zapped) },
145 { "mmu_flooded", VM_STAT(mmu_flooded) },
146 { "mmu_recycled", VM_STAT(mmu_recycled) },
Avi Kivitydfc5aa02007-12-18 19:47:18 +0200147 { "mmu_cache_miss", VM_STAT(mmu_cache_miss) },
Marcelo Tosatti4731d4c2008-09-23 13:18:39 -0300148 { "mmu_unsync", VM_STAT(mmu_unsync) },
Avi Kivity0f74a242007-11-20 23:01:14 +0200149 { "remote_tlb_flush", VM_STAT(remote_tlb_flush) },
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300150 { "largepages", VM_STAT(lpages) },
Hollis Blanchard417bc302007-10-31 17:24:23 -0500151 { NULL }
152};
153
Dexuan Cui2acf9232010-06-10 11:27:12 +0800154u64 __read_mostly host_xcr0;
155
156static inline u32 bit(int bitno)
157{
158 return 1 << (bitno & 31);
159}
160
Avi Kivity18863bd2009-09-07 11:12:18 +0300161static void kvm_on_user_return(struct user_return_notifier *urn)
162{
163 unsigned slot;
Avi Kivity18863bd2009-09-07 11:12:18 +0300164 struct kvm_shared_msrs *locals
165 = container_of(urn, struct kvm_shared_msrs, urn);
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800166 struct kvm_shared_msr_values *values;
Avi Kivity18863bd2009-09-07 11:12:18 +0300167
168 for (slot = 0; slot < shared_msrs_global.nr; ++slot) {
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800169 values = &locals->values[slot];
170 if (values->host != values->curr) {
171 wrmsrl(shared_msrs_global.msrs[slot], values->host);
172 values->curr = values->host;
Avi Kivity18863bd2009-09-07 11:12:18 +0300173 }
174 }
175 locals->registered = false;
176 user_return_notifier_unregister(urn);
177}
178
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800179static void shared_msr_update(unsigned slot, u32 msr)
Avi Kivity18863bd2009-09-07 11:12:18 +0300180{
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800181 struct kvm_shared_msrs *smsr;
Avi Kivity18863bd2009-09-07 11:12:18 +0300182 u64 value;
183
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800184 smsr = &__get_cpu_var(shared_msrs);
185 /* only read, and nobody should modify it at this time,
186 * so don't need lock */
187 if (slot >= shared_msrs_global.nr) {
188 printk(KERN_ERR "kvm: invalid MSR slot!");
189 return;
190 }
191 rdmsrl_safe(msr, &value);
192 smsr->values[slot].host = value;
193 smsr->values[slot].curr = value;
194}
195
196void kvm_define_shared_msr(unsigned slot, u32 msr)
197{
Avi Kivity18863bd2009-09-07 11:12:18 +0300198 if (slot >= shared_msrs_global.nr)
199 shared_msrs_global.nr = slot + 1;
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800200 shared_msrs_global.msrs[slot] = msr;
201 /* we need ensured the shared_msr_global have been updated */
202 smp_wmb();
Avi Kivity18863bd2009-09-07 11:12:18 +0300203}
204EXPORT_SYMBOL_GPL(kvm_define_shared_msr);
205
206static void kvm_shared_msr_cpu_online(void)
207{
208 unsigned i;
Avi Kivity18863bd2009-09-07 11:12:18 +0300209
210 for (i = 0; i < shared_msrs_global.nr; ++i)
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800211 shared_msr_update(i, shared_msrs_global.msrs[i]);
Avi Kivity18863bd2009-09-07 11:12:18 +0300212}
213
Avi Kivityd5696722009-12-02 12:28:47 +0200214void kvm_set_shared_msr(unsigned slot, u64 value, u64 mask)
Avi Kivity18863bd2009-09-07 11:12:18 +0300215{
216 struct kvm_shared_msrs *smsr = &__get_cpu_var(shared_msrs);
217
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800218 if (((value ^ smsr->values[slot].curr) & mask) == 0)
Avi Kivity18863bd2009-09-07 11:12:18 +0300219 return;
Sheng Yang2bf78fa2009-12-18 16:48:44 +0800220 smsr->values[slot].curr = value;
221 wrmsrl(shared_msrs_global.msrs[slot], value);
Avi Kivity18863bd2009-09-07 11:12:18 +0300222 if (!smsr->registered) {
223 smsr->urn.on_user_return = kvm_on_user_return;
224 user_return_notifier_register(&smsr->urn);
225 smsr->registered = true;
226 }
227}
228EXPORT_SYMBOL_GPL(kvm_set_shared_msr);
229
Avi Kivity3548bab2009-11-28 14:18:47 +0200230static void drop_user_return_notifiers(void *ignore)
231{
232 struct kvm_shared_msrs *smsr = &__get_cpu_var(shared_msrs);
233
234 if (smsr->registered)
235 kvm_on_user_return(&smsr->urn);
236}
237
Carsten Otte6866b832007-10-29 16:09:10 +0100238u64 kvm_get_apic_base(struct kvm_vcpu *vcpu)
239{
240 if (irqchip_in_kernel(vcpu->kvm))
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800241 return vcpu->arch.apic_base;
Carsten Otte6866b832007-10-29 16:09:10 +0100242 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800243 return vcpu->arch.apic_base;
Carsten Otte6866b832007-10-29 16:09:10 +0100244}
245EXPORT_SYMBOL_GPL(kvm_get_apic_base);
246
247void kvm_set_apic_base(struct kvm_vcpu *vcpu, u64 data)
248{
249 /* TODO: reserve bits check */
250 if (irqchip_in_kernel(vcpu->kvm))
251 kvm_lapic_set_base(vcpu, data);
252 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800253 vcpu->arch.apic_base = data;
Carsten Otte6866b832007-10-29 16:09:10 +0100254}
255EXPORT_SYMBOL_GPL(kvm_set_apic_base);
256
Eddie Dong3fd28fc2009-11-19 17:54:07 +0200257#define EXCPT_BENIGN 0
258#define EXCPT_CONTRIBUTORY 1
259#define EXCPT_PF 2
260
261static int exception_class(int vector)
262{
263 switch (vector) {
264 case PF_VECTOR:
265 return EXCPT_PF;
266 case DE_VECTOR:
267 case TS_VECTOR:
268 case NP_VECTOR:
269 case SS_VECTOR:
270 case GP_VECTOR:
271 return EXCPT_CONTRIBUTORY;
272 default:
273 break;
274 }
275 return EXCPT_BENIGN;
276}
277
278static void kvm_multiple_exception(struct kvm_vcpu *vcpu,
Joerg Roedelce7ddec2010-04-22 12:33:13 +0200279 unsigned nr, bool has_error, u32 error_code,
280 bool reinject)
Eddie Dong3fd28fc2009-11-19 17:54:07 +0200281{
282 u32 prev_nr;
283 int class1, class2;
284
285 if (!vcpu->arch.exception.pending) {
286 queue:
287 vcpu->arch.exception.pending = true;
288 vcpu->arch.exception.has_error_code = has_error;
289 vcpu->arch.exception.nr = nr;
290 vcpu->arch.exception.error_code = error_code;
Joerg Roedel3f0fd292010-05-05 16:04:41 +0200291 vcpu->arch.exception.reinject = reinject;
Eddie Dong3fd28fc2009-11-19 17:54:07 +0200292 return;
293 }
294
295 /* to check exception */
296 prev_nr = vcpu->arch.exception.nr;
297 if (prev_nr == DF_VECTOR) {
298 /* triple fault -> shutdown */
299 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
300 return;
301 }
302 class1 = exception_class(prev_nr);
303 class2 = exception_class(nr);
304 if ((class1 == EXCPT_CONTRIBUTORY && class2 == EXCPT_CONTRIBUTORY)
305 || (class1 == EXCPT_PF && class2 != EXCPT_BENIGN)) {
306 /* generate double fault per SDM Table 5-5 */
307 vcpu->arch.exception.pending = true;
308 vcpu->arch.exception.has_error_code = true;
309 vcpu->arch.exception.nr = DF_VECTOR;
310 vcpu->arch.exception.error_code = 0;
311 } else
312 /* replace previous exception with a new one in a hope
313 that instruction re-execution will regenerate lost
314 exception */
315 goto queue;
316}
317
Avi Kivity298101d2007-11-25 13:41:11 +0200318void kvm_queue_exception(struct kvm_vcpu *vcpu, unsigned nr)
319{
Joerg Roedelce7ddec2010-04-22 12:33:13 +0200320 kvm_multiple_exception(vcpu, nr, false, 0, false);
Avi Kivity298101d2007-11-25 13:41:11 +0200321}
322EXPORT_SYMBOL_GPL(kvm_queue_exception);
323
Joerg Roedelce7ddec2010-04-22 12:33:13 +0200324void kvm_requeue_exception(struct kvm_vcpu *vcpu, unsigned nr)
325{
326 kvm_multiple_exception(vcpu, nr, false, 0, true);
327}
328EXPORT_SYMBOL_GPL(kvm_requeue_exception);
329
Avi Kivityc3c91fe2007-11-25 14:04:58 +0200330void kvm_inject_page_fault(struct kvm_vcpu *vcpu, unsigned long addr,
331 u32 error_code)
332{
333 ++vcpu->stat.pf_guest;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800334 vcpu->arch.cr2 = addr;
Avi Kivityc3c91fe2007-11-25 14:04:58 +0200335 kvm_queue_exception_e(vcpu, PF_VECTOR, error_code);
336}
337
Sheng Yang3419ffc2008-05-15 09:52:48 +0800338void kvm_inject_nmi(struct kvm_vcpu *vcpu)
339{
340 vcpu->arch.nmi_pending = 1;
341}
342EXPORT_SYMBOL_GPL(kvm_inject_nmi);
343
Avi Kivity298101d2007-11-25 13:41:11 +0200344void kvm_queue_exception_e(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code)
345{
Joerg Roedelce7ddec2010-04-22 12:33:13 +0200346 kvm_multiple_exception(vcpu, nr, true, error_code, false);
Avi Kivity298101d2007-11-25 13:41:11 +0200347}
348EXPORT_SYMBOL_GPL(kvm_queue_exception_e);
349
Joerg Roedelce7ddec2010-04-22 12:33:13 +0200350void kvm_requeue_exception_e(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code)
351{
352 kvm_multiple_exception(vcpu, nr, true, error_code, true);
353}
354EXPORT_SYMBOL_GPL(kvm_requeue_exception_e);
355
Carsten Ottea03490e2007-10-29 16:09:35 +0100356/*
Avi Kivity0a79b002009-09-01 12:03:25 +0300357 * Checks if cpl <= required_cpl; if true, return true. Otherwise queue
358 * a #GP and return false.
359 */
360bool kvm_require_cpl(struct kvm_vcpu *vcpu, int required_cpl)
Carsten Otte043405e2007-10-10 17:16:19 +0200361{
Avi Kivity0a79b002009-09-01 12:03:25 +0300362 if (kvm_x86_ops->get_cpl(vcpu) <= required_cpl)
363 return true;
364 kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
365 return false;
Carsten Ottea03490e2007-10-29 16:09:35 +0100366}
Avi Kivity0a79b002009-09-01 12:03:25 +0300367EXPORT_SYMBOL_GPL(kvm_require_cpl);
Carsten Ottea03490e2007-10-29 16:09:35 +0100368
369/*
370 * Load the pae pdptrs. Return true is they are all valid.
371 */
372int load_pdptrs(struct kvm_vcpu *vcpu, unsigned long cr3)
373{
374 gfn_t pdpt_gfn = cr3 >> PAGE_SHIFT;
375 unsigned offset = ((cr3 & (PAGE_SIZE-1)) >> 5) << 2;
376 int i;
377 int ret;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800378 u64 pdpte[ARRAY_SIZE(vcpu->arch.pdptrs)];
Carsten Ottea03490e2007-10-29 16:09:35 +0100379
Carsten Ottea03490e2007-10-29 16:09:35 +0100380 ret = kvm_read_guest_page(vcpu->kvm, pdpt_gfn, pdpte,
381 offset * sizeof(u64), sizeof(pdpte));
382 if (ret < 0) {
383 ret = 0;
384 goto out;
385 }
386 for (i = 0; i < ARRAY_SIZE(pdpte); ++i) {
Avi Kivity43a37952009-06-10 14:12:05 +0300387 if (is_present_gpte(pdpte[i]) &&
Dong, Eddie20c466b2009-03-31 23:03:45 +0800388 (pdpte[i] & vcpu->arch.mmu.rsvd_bits_mask[0][2])) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100389 ret = 0;
390 goto out;
391 }
392 }
393 ret = 1;
394
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800395 memcpy(vcpu->arch.pdptrs, pdpte, sizeof(vcpu->arch.pdptrs));
Avi Kivity6de4f3a2009-05-31 22:58:47 +0300396 __set_bit(VCPU_EXREG_PDPTR,
397 (unsigned long *)&vcpu->arch.regs_avail);
398 __set_bit(VCPU_EXREG_PDPTR,
399 (unsigned long *)&vcpu->arch.regs_dirty);
Carsten Ottea03490e2007-10-29 16:09:35 +0100400out:
Carsten Ottea03490e2007-10-29 16:09:35 +0100401
402 return ret;
403}
Joerg Roedelcc4b6872008-02-07 13:47:43 +0100404EXPORT_SYMBOL_GPL(load_pdptrs);
Carsten Ottea03490e2007-10-29 16:09:35 +0100405
Avi Kivityd835dfe2007-11-21 02:57:59 +0200406static bool pdptrs_changed(struct kvm_vcpu *vcpu)
407{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800408 u64 pdpte[ARRAY_SIZE(vcpu->arch.pdptrs)];
Avi Kivityd835dfe2007-11-21 02:57:59 +0200409 bool changed = true;
410 int r;
411
412 if (is_long_mode(vcpu) || !is_pae(vcpu))
413 return false;
414
Avi Kivity6de4f3a2009-05-31 22:58:47 +0300415 if (!test_bit(VCPU_EXREG_PDPTR,
416 (unsigned long *)&vcpu->arch.regs_avail))
417 return true;
418
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800419 r = kvm_read_guest(vcpu->kvm, vcpu->arch.cr3 & ~31u, pdpte, sizeof(pdpte));
Avi Kivityd835dfe2007-11-21 02:57:59 +0200420 if (r < 0)
421 goto out;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800422 changed = memcmp(pdpte, vcpu->arch.pdptrs, sizeof(pdpte)) != 0;
Avi Kivityd835dfe2007-11-21 02:57:59 +0200423out:
Avi Kivityd835dfe2007-11-21 02:57:59 +0200424
425 return changed;
426}
427
Avi Kivity49a9b072010-06-10 17:02:14 +0300428int kvm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
Carsten Ottea03490e2007-10-29 16:09:35 +0100429{
Sheng Yangaad82702010-05-12 16:40:42 +0800430 unsigned long old_cr0 = kvm_read_cr0(vcpu);
431 unsigned long update_bits = X86_CR0_PG | X86_CR0_WP |
432 X86_CR0_CD | X86_CR0_NW;
433
Avi Kivityf9a48e62010-01-06 19:10:22 +0200434 cr0 |= X86_CR0_ET;
435
Gleb Natapovab344822010-01-21 15:28:46 +0200436#ifdef CONFIG_X86_64
Gleb Natapov0f122442010-04-28 19:15:31 +0300437 if (cr0 & 0xffffffff00000000UL)
438 return 1;
Gleb Natapovab344822010-01-21 15:28:46 +0200439#endif
440
441 cr0 &= ~CR0_RESERVED_BITS;
Carsten Ottea03490e2007-10-29 16:09:35 +0100442
Gleb Natapov0f122442010-04-28 19:15:31 +0300443 if ((cr0 & X86_CR0_NW) && !(cr0 & X86_CR0_CD))
444 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100445
Gleb Natapov0f122442010-04-28 19:15:31 +0300446 if ((cr0 & X86_CR0_PG) && !(cr0 & X86_CR0_PE))
447 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100448
449 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
450#ifdef CONFIG_X86_64
Avi Kivityf6801df2010-01-21 15:31:50 +0200451 if ((vcpu->arch.efer & EFER_LME)) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100452 int cs_db, cs_l;
453
Gleb Natapov0f122442010-04-28 19:15:31 +0300454 if (!is_pae(vcpu))
455 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100456 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
Gleb Natapov0f122442010-04-28 19:15:31 +0300457 if (cs_l)
458 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100459 } else
460#endif
Gleb Natapov0f122442010-04-28 19:15:31 +0300461 if (is_pae(vcpu) && !load_pdptrs(vcpu, vcpu->arch.cr3))
462 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100463 }
464
465 kvm_x86_ops->set_cr0(vcpu, cr0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100466
Sheng Yangaad82702010-05-12 16:40:42 +0800467 if ((cr0 ^ old_cr0) & update_bits)
468 kvm_mmu_reset_context(vcpu);
Gleb Natapov0f122442010-04-28 19:15:31 +0300469 return 0;
470}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200471EXPORT_SYMBOL_GPL(kvm_set_cr0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100472
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200473void kvm_lmsw(struct kvm_vcpu *vcpu, unsigned long msw)
Carsten Ottea03490e2007-10-29 16:09:35 +0100474{
Avi Kivity49a9b072010-06-10 17:02:14 +0300475 (void)kvm_set_cr0(vcpu, kvm_read_cr0_bits(vcpu, ~0x0eul) | (msw & 0x0f));
Carsten Ottea03490e2007-10-29 16:09:35 +0100476}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200477EXPORT_SYMBOL_GPL(kvm_lmsw);
Carsten Ottea03490e2007-10-29 16:09:35 +0100478
Dexuan Cui2acf9232010-06-10 11:27:12 +0800479int __kvm_set_xcr(struct kvm_vcpu *vcpu, u32 index, u64 xcr)
480{
481 u64 xcr0;
482
483 /* Only support XCR_XFEATURE_ENABLED_MASK(xcr0) now */
484 if (index != XCR_XFEATURE_ENABLED_MASK)
485 return 1;
486 xcr0 = xcr;
487 if (kvm_x86_ops->get_cpl(vcpu) != 0)
488 return 1;
489 if (!(xcr0 & XSTATE_FP))
490 return 1;
491 if ((xcr0 & XSTATE_YMM) && !(xcr0 & XSTATE_SSE))
492 return 1;
493 if (xcr0 & ~host_xcr0)
494 return 1;
495 vcpu->arch.xcr0 = xcr0;
496 vcpu->guest_xcr0_loaded = 0;
497 return 0;
498}
499
500int kvm_set_xcr(struct kvm_vcpu *vcpu, u32 index, u64 xcr)
501{
502 if (__kvm_set_xcr(vcpu, index, xcr)) {
503 kvm_inject_gp(vcpu, 0);
504 return 1;
505 }
506 return 0;
507}
508EXPORT_SYMBOL_GPL(kvm_set_xcr);
509
510static bool guest_cpuid_has_xsave(struct kvm_vcpu *vcpu)
511{
512 struct kvm_cpuid_entry2 *best;
513
514 best = kvm_find_cpuid_entry(vcpu, 1, 0);
515 return best && (best->ecx & bit(X86_FEATURE_XSAVE));
516}
517
518static void update_cpuid(struct kvm_vcpu *vcpu)
519{
520 struct kvm_cpuid_entry2 *best;
521
522 best = kvm_find_cpuid_entry(vcpu, 1, 0);
523 if (!best)
524 return;
525
526 /* Update OSXSAVE bit */
527 if (cpu_has_xsave && best->function == 0x1) {
528 best->ecx &= ~(bit(X86_FEATURE_OSXSAVE));
529 if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE))
530 best->ecx |= bit(X86_FEATURE_OSXSAVE);
531 }
532}
533
Avi Kivitya83b29c2010-06-10 17:02:15 +0300534int kvm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
Carsten Ottea03490e2007-10-29 16:09:35 +0100535{
Avi Kivityfc78f512009-12-07 12:16:48 +0200536 unsigned long old_cr4 = kvm_read_cr4(vcpu);
Avi Kivitya2edf572009-05-24 22:19:00 +0300537 unsigned long pdptr_bits = X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PAE;
538
Gleb Natapov0f122442010-04-28 19:15:31 +0300539 if (cr4 & CR4_RESERVED_BITS)
540 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100541
Dexuan Cui2acf9232010-06-10 11:27:12 +0800542 if (!guest_cpuid_has_xsave(vcpu) && (cr4 & X86_CR4_OSXSAVE))
543 return 1;
544
Carsten Ottea03490e2007-10-29 16:09:35 +0100545 if (is_long_mode(vcpu)) {
Gleb Natapov0f122442010-04-28 19:15:31 +0300546 if (!(cr4 & X86_CR4_PAE))
547 return 1;
Avi Kivitya2edf572009-05-24 22:19:00 +0300548 } else if (is_paging(vcpu) && (cr4 & X86_CR4_PAE)
549 && ((cr4 ^ old_cr4) & pdptr_bits)
Gleb Natapov0f122442010-04-28 19:15:31 +0300550 && !load_pdptrs(vcpu, vcpu->arch.cr3))
551 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100552
Gleb Natapov0f122442010-04-28 19:15:31 +0300553 if (cr4 & X86_CR4_VMXE)
554 return 1;
555
Carsten Ottea03490e2007-10-29 16:09:35 +0100556 kvm_x86_ops->set_cr4(vcpu, cr4);
Sheng Yang62ad0752010-05-12 16:40:41 +0800557
Sheng Yangaad82702010-05-12 16:40:42 +0800558 if ((cr4 ^ old_cr4) & pdptr_bits)
559 kvm_mmu_reset_context(vcpu);
Gleb Natapov0f122442010-04-28 19:15:31 +0300560
Dexuan Cui2acf9232010-06-10 11:27:12 +0800561 if ((cr4 ^ old_cr4) & X86_CR4_OSXSAVE)
562 update_cpuid(vcpu);
563
Gleb Natapov0f122442010-04-28 19:15:31 +0300564 return 0;
565}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200566EXPORT_SYMBOL_GPL(kvm_set_cr4);
Carsten Ottea03490e2007-10-29 16:09:35 +0100567
Avi Kivity23902182010-06-10 17:02:16 +0300568int kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
Carsten Ottea03490e2007-10-29 16:09:35 +0100569{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800570 if (cr3 == vcpu->arch.cr3 && !pdptrs_changed(vcpu)) {
Marcelo Tosatti0ba73cd2008-09-23 13:18:34 -0300571 kvm_mmu_sync_roots(vcpu);
Avi Kivityd835dfe2007-11-21 02:57:59 +0200572 kvm_mmu_flush_tlb(vcpu);
Gleb Natapov0f122442010-04-28 19:15:31 +0300573 return 0;
Avi Kivityd835dfe2007-11-21 02:57:59 +0200574 }
575
Carsten Ottea03490e2007-10-29 16:09:35 +0100576 if (is_long_mode(vcpu)) {
Gleb Natapov0f122442010-04-28 19:15:31 +0300577 if (cr3 & CR3_L_MODE_RESERVED_BITS)
578 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100579 } else {
580 if (is_pae(vcpu)) {
Gleb Natapov0f122442010-04-28 19:15:31 +0300581 if (cr3 & CR3_PAE_RESERVED_BITS)
582 return 1;
583 if (is_paging(vcpu) && !load_pdptrs(vcpu, cr3))
584 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100585 }
586 /*
587 * We don't check reserved bits in nonpae mode, because
588 * this isn't enforced, and VMware depends on this.
589 */
590 }
591
Carsten Ottea03490e2007-10-29 16:09:35 +0100592 /*
593 * Does the new cr3 value map to physical memory? (Note, we
594 * catch an invalid cr3 even in real-mode, because it would
595 * cause trouble later on when we turn on paging anyway.)
596 *
597 * A real CPU would silently accept an invalid cr3 and would
598 * attempt to use it - with largely undefined (and often hard
599 * to debug) behavior on the guest side.
600 */
601 if (unlikely(!gfn_to_memslot(vcpu->kvm, cr3 >> PAGE_SHIFT)))
Gleb Natapov0f122442010-04-28 19:15:31 +0300602 return 1;
603 vcpu->arch.cr3 = cr3;
604 vcpu->arch.mmu.new_cr3(vcpu);
605 return 0;
606}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200607EXPORT_SYMBOL_GPL(kvm_set_cr3);
Carsten Ottea03490e2007-10-29 16:09:35 +0100608
Gleb Natapov0f122442010-04-28 19:15:31 +0300609int __kvm_set_cr8(struct kvm_vcpu *vcpu, unsigned long cr8)
Carsten Ottea03490e2007-10-29 16:09:35 +0100610{
Gleb Natapov0f122442010-04-28 19:15:31 +0300611 if (cr8 & CR8_RESERVED_BITS)
612 return 1;
Carsten Ottea03490e2007-10-29 16:09:35 +0100613 if (irqchip_in_kernel(vcpu->kvm))
614 kvm_lapic_set_tpr(vcpu, cr8);
615 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800616 vcpu->arch.cr8 = cr8;
Gleb Natapov0f122442010-04-28 19:15:31 +0300617 return 0;
618}
619
620void kvm_set_cr8(struct kvm_vcpu *vcpu, unsigned long cr8)
621{
622 if (__kvm_set_cr8(vcpu, cr8))
623 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100624}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200625EXPORT_SYMBOL_GPL(kvm_set_cr8);
Carsten Ottea03490e2007-10-29 16:09:35 +0100626
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200627unsigned long kvm_get_cr8(struct kvm_vcpu *vcpu)
Carsten Ottea03490e2007-10-29 16:09:35 +0100628{
629 if (irqchip_in_kernel(vcpu->kvm))
630 return kvm_lapic_get_cr8(vcpu);
631 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800632 return vcpu->arch.cr8;
Carsten Ottea03490e2007-10-29 16:09:35 +0100633}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200634EXPORT_SYMBOL_GPL(kvm_get_cr8);
Carsten Ottea03490e2007-10-29 16:09:35 +0100635
Gleb Natapov338dbc92010-04-28 19:15:32 +0300636static int __kvm_set_dr(struct kvm_vcpu *vcpu, int dr, unsigned long val)
Gleb Natapov020df072010-04-13 10:05:23 +0300637{
638 switch (dr) {
639 case 0 ... 3:
640 vcpu->arch.db[dr] = val;
641 if (!(vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP))
642 vcpu->arch.eff_db[dr] = val;
643 break;
644 case 4:
Gleb Natapov338dbc92010-04-28 19:15:32 +0300645 if (kvm_read_cr4_bits(vcpu, X86_CR4_DE))
646 return 1; /* #UD */
Gleb Natapov020df072010-04-13 10:05:23 +0300647 /* fall through */
648 case 6:
Gleb Natapov338dbc92010-04-28 19:15:32 +0300649 if (val & 0xffffffff00000000ULL)
650 return -1; /* #GP */
Gleb Natapov020df072010-04-13 10:05:23 +0300651 vcpu->arch.dr6 = (val & DR6_VOLATILE) | DR6_FIXED_1;
652 break;
653 case 5:
Gleb Natapov338dbc92010-04-28 19:15:32 +0300654 if (kvm_read_cr4_bits(vcpu, X86_CR4_DE))
655 return 1; /* #UD */
Gleb Natapov020df072010-04-13 10:05:23 +0300656 /* fall through */
657 default: /* 7 */
Gleb Natapov338dbc92010-04-28 19:15:32 +0300658 if (val & 0xffffffff00000000ULL)
659 return -1; /* #GP */
Gleb Natapov020df072010-04-13 10:05:23 +0300660 vcpu->arch.dr7 = (val & DR7_VOLATILE) | DR7_FIXED_1;
661 if (!(vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP)) {
662 kvm_x86_ops->set_dr7(vcpu, vcpu->arch.dr7);
663 vcpu->arch.switch_db_regs = (val & DR7_BP_EN_MASK);
664 }
665 break;
666 }
667
668 return 0;
669}
Gleb Natapov338dbc92010-04-28 19:15:32 +0300670
671int kvm_set_dr(struct kvm_vcpu *vcpu, int dr, unsigned long val)
672{
673 int res;
674
675 res = __kvm_set_dr(vcpu, dr, val);
676 if (res > 0)
677 kvm_queue_exception(vcpu, UD_VECTOR);
678 else if (res < 0)
679 kvm_inject_gp(vcpu, 0);
680
681 return res;
682}
Gleb Natapov020df072010-04-13 10:05:23 +0300683EXPORT_SYMBOL_GPL(kvm_set_dr);
684
Gleb Natapov338dbc92010-04-28 19:15:32 +0300685static int _kvm_get_dr(struct kvm_vcpu *vcpu, int dr, unsigned long *val)
Gleb Natapov020df072010-04-13 10:05:23 +0300686{
687 switch (dr) {
688 case 0 ... 3:
689 *val = vcpu->arch.db[dr];
690 break;
691 case 4:
Gleb Natapov338dbc92010-04-28 19:15:32 +0300692 if (kvm_read_cr4_bits(vcpu, X86_CR4_DE))
Gleb Natapov020df072010-04-13 10:05:23 +0300693 return 1;
Gleb Natapov020df072010-04-13 10:05:23 +0300694 /* fall through */
695 case 6:
696 *val = vcpu->arch.dr6;
697 break;
698 case 5:
Gleb Natapov338dbc92010-04-28 19:15:32 +0300699 if (kvm_read_cr4_bits(vcpu, X86_CR4_DE))
Gleb Natapov020df072010-04-13 10:05:23 +0300700 return 1;
Gleb Natapov020df072010-04-13 10:05:23 +0300701 /* fall through */
702 default: /* 7 */
703 *val = vcpu->arch.dr7;
704 break;
705 }
706
707 return 0;
708}
Gleb Natapov338dbc92010-04-28 19:15:32 +0300709
710int kvm_get_dr(struct kvm_vcpu *vcpu, int dr, unsigned long *val)
711{
712 if (_kvm_get_dr(vcpu, dr, val)) {
713 kvm_queue_exception(vcpu, UD_VECTOR);
714 return 1;
715 }
716 return 0;
717}
Gleb Natapov020df072010-04-13 10:05:23 +0300718EXPORT_SYMBOL_GPL(kvm_get_dr);
719
Carsten Otte043405e2007-10-10 17:16:19 +0200720/*
721 * List of msr numbers which we expose to userspace through KVM_GET_MSRS
722 * and KVM_SET_MSRS, and KVM_GET_MSR_INDEX_LIST.
723 *
724 * This list is modified at module load time to reflect the
Glauber Costae3267cb2009-10-06 13:24:50 -0400725 * capabilities of the host cpu. This capabilities test skips MSRs that are
726 * kvm-specific. Those are put in the beginning of the list.
Carsten Otte043405e2007-10-10 17:16:19 +0200727 */
Glauber Costae3267cb2009-10-06 13:24:50 -0400728
Glauber Costa11c6bff2010-05-11 12:17:41 -0400729#define KVM_SAVE_MSRS_BEGIN 7
Carsten Otte043405e2007-10-10 17:16:19 +0200730static u32 msrs_to_save[] = {
Glauber Costae3267cb2009-10-06 13:24:50 -0400731 MSR_KVM_SYSTEM_TIME, MSR_KVM_WALL_CLOCK,
Glauber Costa11c6bff2010-05-11 12:17:41 -0400732 MSR_KVM_SYSTEM_TIME_NEW, MSR_KVM_WALL_CLOCK_NEW,
Gleb Natapov55cd8e52010-01-17 15:51:22 +0200733 HV_X64_MSR_GUEST_OS_ID, HV_X64_MSR_HYPERCALL,
Gleb Natapov10388a02010-01-17 15:51:23 +0200734 HV_X64_MSR_APIC_ASSIST_PAGE,
Carsten Otte043405e2007-10-10 17:16:19 +0200735 MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP,
736 MSR_K6_STAR,
737#ifdef CONFIG_X86_64
738 MSR_CSTAR, MSR_KERNEL_GS_BASE, MSR_SYSCALL_MASK, MSR_LSTAR,
739#endif
Glauber Costae3267cb2009-10-06 13:24:50 -0400740 MSR_IA32_TSC, MSR_IA32_PERF_STATUS, MSR_IA32_CR_PAT, MSR_VM_HSAVE_PA
Carsten Otte043405e2007-10-10 17:16:19 +0200741};
742
743static unsigned num_msrs_to_save;
744
745static u32 emulated_msrs[] = {
746 MSR_IA32_MISC_ENABLE,
747};
748
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200749static int set_efer(struct kvm_vcpu *vcpu, u64 efer)
Carsten Otte15c4a642007-10-30 18:44:17 +0100750{
Sheng Yangaad82702010-05-12 16:40:42 +0800751 u64 old_efer = vcpu->arch.efer;
752
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200753 if (efer & efer_reserved_bits)
754 return 1;
Carsten Otte15c4a642007-10-30 18:44:17 +0100755
756 if (is_paging(vcpu)
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200757 && (vcpu->arch.efer & EFER_LME) != (efer & EFER_LME))
758 return 1;
Carsten Otte15c4a642007-10-30 18:44:17 +0100759
Alexander Graf1b2fd702009-02-02 16:23:51 +0100760 if (efer & EFER_FFXSR) {
761 struct kvm_cpuid_entry2 *feat;
762
763 feat = kvm_find_cpuid_entry(vcpu, 0x80000001, 0);
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200764 if (!feat || !(feat->edx & bit(X86_FEATURE_FXSR_OPT)))
765 return 1;
Alexander Graf1b2fd702009-02-02 16:23:51 +0100766 }
767
Alexander Grafd8017472008-11-25 20:17:11 +0100768 if (efer & EFER_SVME) {
769 struct kvm_cpuid_entry2 *feat;
770
771 feat = kvm_find_cpuid_entry(vcpu, 0x80000001, 0);
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200772 if (!feat || !(feat->ecx & bit(X86_FEATURE_SVM)))
773 return 1;
Alexander Grafd8017472008-11-25 20:17:11 +0100774 }
775
Carsten Otte15c4a642007-10-30 18:44:17 +0100776 efer &= ~EFER_LMA;
Avi Kivityf6801df2010-01-21 15:31:50 +0200777 efer |= vcpu->arch.efer & EFER_LMA;
Carsten Otte15c4a642007-10-30 18:44:17 +0100778
Sheng Yanga3d204e2010-05-12 16:40:40 +0800779 kvm_x86_ops->set_efer(vcpu, efer);
780
Avi Kivity9645bb562009-03-31 11:31:54 +0300781 vcpu->arch.mmu.base_role.nxe = (efer & EFER_NX) && !tdp_enabled;
782 kvm_mmu_reset_context(vcpu);
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200783
Sheng Yangaad82702010-05-12 16:40:42 +0800784 /* Update reserved bits */
785 if ((efer ^ old_efer) & EFER_NX)
786 kvm_mmu_reset_context(vcpu);
787
Roedel, Joergb69e8ca2010-05-06 11:38:43 +0200788 return 0;
Carsten Otte15c4a642007-10-30 18:44:17 +0100789}
790
Joerg Roedelf2b4b7d2008-01-31 14:57:37 +0100791void kvm_enable_efer_bits(u64 mask)
792{
793 efer_reserved_bits &= ~mask;
794}
795EXPORT_SYMBOL_GPL(kvm_enable_efer_bits);
796
797
Carsten Otte15c4a642007-10-30 18:44:17 +0100798/*
799 * Writes msr value into into the appropriate "register".
800 * Returns 0 on success, non-0 otherwise.
801 * Assumes vcpu_load() was already called.
802 */
803int kvm_set_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 data)
804{
805 return kvm_x86_ops->set_msr(vcpu, msr_index, data);
806}
807
Carsten Otte313a3dc2007-10-11 19:16:52 +0200808/*
809 * Adapt set_msr() to msr_io()'s calling convention
810 */
811static int do_set_msr(struct kvm_vcpu *vcpu, unsigned index, u64 *data)
812{
813 return kvm_set_msr(vcpu, index, *data);
814}
815
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200816static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock)
817{
Avi Kivity9ed3c442010-05-04 15:00:37 +0300818 int version;
819 int r;
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200820 struct pvclock_wall_clock wc;
Jason Wang923de3c2010-01-27 19:13:49 +0800821 struct timespec boot;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200822
823 if (!wall_clock)
824 return;
825
Avi Kivity9ed3c442010-05-04 15:00:37 +0300826 r = kvm_read_guest(kvm, wall_clock, &version, sizeof(version));
827 if (r)
828 return;
829
830 if (version & 1)
831 ++version; /* first time write, random junk */
832
833 ++version;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200834
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200835 kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
836
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200837 /*
838 * The guest calculates current wall clock time by adding
839 * system time (updated by kvm_write_guest_time below) to the
840 * wall clock specified here. guest system time equals host
841 * system time for us, thus we must fill in host boot time here.
842 */
Jason Wang923de3c2010-01-27 19:13:49 +0800843 getboottime(&boot);
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200844
845 wc.sec = boot.tv_sec;
846 wc.nsec = boot.tv_nsec;
847 wc.version = version;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200848
849 kvm_write_guest(kvm, wall_clock, &wc, sizeof(wc));
850
851 version++;
852 kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200853}
854
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200855static uint32_t div_frac(uint32_t dividend, uint32_t divisor)
856{
857 uint32_t quotient, remainder;
858
859 /* Don't try to replace with do_div(), this one calculates
860 * "(dividend << 32) / divisor" */
861 __asm__ ( "divl %4"
862 : "=a" (quotient), "=d" (remainder)
863 : "0" (0), "1" (dividend), "r" (divisor) );
864 return quotient;
865}
866
867static void kvm_set_time_scale(uint32_t tsc_khz, struct pvclock_vcpu_time_info *hv_clock)
868{
869 uint64_t nsecs = 1000000000LL;
870 int32_t shift = 0;
871 uint64_t tps64;
872 uint32_t tps32;
873
874 tps64 = tsc_khz * 1000LL;
875 while (tps64 > nsecs*2) {
876 tps64 >>= 1;
877 shift--;
878 }
879
880 tps32 = (uint32_t)tps64;
881 while (tps32 <= (uint32_t)nsecs) {
882 tps32 <<= 1;
883 shift++;
884 }
885
886 hv_clock->tsc_shift = shift;
887 hv_clock->tsc_to_system_mul = div_frac(nsecs, tps32);
888
889 pr_debug("%s: tsc_khz %u, tsc_shift %d, tsc_mul %u\n",
Harvey Harrison80a914d2008-10-15 22:01:25 -0700890 __func__, tsc_khz, hv_clock->tsc_shift,
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200891 hv_clock->tsc_to_system_mul);
892}
893
Gerd Hoffmannc8076602009-02-04 17:52:04 +0100894static DEFINE_PER_CPU(unsigned long, cpu_tsc_khz);
895
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200896static void kvm_write_guest_time(struct kvm_vcpu *v)
897{
898 struct timespec ts;
899 unsigned long flags;
900 struct kvm_vcpu_arch *vcpu = &v->arch;
901 void *shared_kaddr;
Avi Kivity463656c2009-04-12 15:49:07 +0300902 unsigned long this_tsc_khz;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200903
904 if ((!vcpu->time_page))
905 return;
906
Avi Kivity463656c2009-04-12 15:49:07 +0300907 this_tsc_khz = get_cpu_var(cpu_tsc_khz);
908 if (unlikely(vcpu->hv_clock_tsc_khz != this_tsc_khz)) {
909 kvm_set_time_scale(this_tsc_khz, &vcpu->hv_clock);
910 vcpu->hv_clock_tsc_khz = this_tsc_khz;
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200911 }
Avi Kivity463656c2009-04-12 15:49:07 +0300912 put_cpu_var(cpu_tsc_khz);
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200913
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200914 /* Keep irq disabled to prevent changes to the clock */
915 local_irq_save(flags);
Jaswinder Singh Rajputaf24a4e2009-05-15 18:42:05 +0530916 kvm_get_msr(v, MSR_IA32_TSC, &vcpu->hv_clock.tsc_timestamp);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200917 ktime_get_ts(&ts);
Jason Wang923de3c2010-01-27 19:13:49 +0800918 monotonic_to_bootbased(&ts);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200919 local_irq_restore(flags);
920
921 /* With all the info we got, fill in the values */
922
923 vcpu->hv_clock.system_time = ts.tv_nsec +
Glauber Costaafbcf7a2009-10-16 15:28:36 -0400924 (NSEC_PER_SEC * (u64)ts.tv_sec) + v->kvm->arch.kvmclock_offset;
925
Glauber Costa371bcf62010-05-11 12:17:46 -0400926 vcpu->hv_clock.flags = 0;
927
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200928 /*
929 * The interface expects us to write an even number signaling that the
930 * update is finished. Since the guest won't see the intermediate
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200931 * state, we just increase by 2 at the end.
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200932 */
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200933 vcpu->hv_clock.version += 2;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200934
935 shared_kaddr = kmap_atomic(vcpu->time_page, KM_USER0);
936
937 memcpy(shared_kaddr + vcpu->time_offset, &vcpu->hv_clock,
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200938 sizeof(vcpu->hv_clock));
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200939
940 kunmap_atomic(shared_kaddr, KM_USER0);
941
942 mark_page_dirty(v->kvm, vcpu->time >> PAGE_SHIFT);
943}
944
Gerd Hoffmannc8076602009-02-04 17:52:04 +0100945static int kvm_request_guest_time_update(struct kvm_vcpu *v)
946{
947 struct kvm_vcpu_arch *vcpu = &v->arch;
948
949 if (!vcpu->time_page)
950 return 0;
951 set_bit(KVM_REQ_KVMCLOCK_UPDATE, &v->requests);
952 return 1;
953}
954
Avi Kivity9ba075a2008-05-26 20:06:35 +0300955static bool msr_mtrr_valid(unsigned msr)
956{
957 switch (msr) {
958 case 0x200 ... 0x200 + 2 * KVM_NR_VAR_MTRR - 1:
959 case MSR_MTRRfix64K_00000:
960 case MSR_MTRRfix16K_80000:
961 case MSR_MTRRfix16K_A0000:
962 case MSR_MTRRfix4K_C0000:
963 case MSR_MTRRfix4K_C8000:
964 case MSR_MTRRfix4K_D0000:
965 case MSR_MTRRfix4K_D8000:
966 case MSR_MTRRfix4K_E0000:
967 case MSR_MTRRfix4K_E8000:
968 case MSR_MTRRfix4K_F0000:
969 case MSR_MTRRfix4K_F8000:
970 case MSR_MTRRdefType:
971 case MSR_IA32_CR_PAT:
972 return true;
973 case 0x2f8:
974 return true;
975 }
976 return false;
977}
978
Marcelo Tosattid6289b92009-06-22 15:27:56 -0300979static bool valid_pat_type(unsigned t)
980{
981 return t < 8 && (1 << t) & 0xf3; /* 0, 1, 4, 5, 6, 7 */
982}
983
984static bool valid_mtrr_type(unsigned t)
985{
986 return t < 8 && (1 << t) & 0x73; /* 0, 1, 4, 5, 6 */
987}
988
989static bool mtrr_valid(struct kvm_vcpu *vcpu, u32 msr, u64 data)
990{
991 int i;
992
993 if (!msr_mtrr_valid(msr))
994 return false;
995
996 if (msr == MSR_IA32_CR_PAT) {
997 for (i = 0; i < 8; i++)
998 if (!valid_pat_type((data >> (i * 8)) & 0xff))
999 return false;
1000 return true;
1001 } else if (msr == MSR_MTRRdefType) {
1002 if (data & ~0xcff)
1003 return false;
1004 return valid_mtrr_type(data & 0xff);
1005 } else if (msr >= MSR_MTRRfix64K_00000 && msr <= MSR_MTRRfix4K_F8000) {
1006 for (i = 0; i < 8 ; i++)
1007 if (!valid_mtrr_type((data >> (i * 8)) & 0xff))
1008 return false;
1009 return true;
1010 }
1011
1012 /* variable MTRRs */
1013 return valid_mtrr_type(data & 0xff);
1014}
1015
Avi Kivity9ba075a2008-05-26 20:06:35 +03001016static int set_msr_mtrr(struct kvm_vcpu *vcpu, u32 msr, u64 data)
1017{
Sheng Yang0bed3b52008-10-09 16:01:54 +08001018 u64 *p = (u64 *)&vcpu->arch.mtrr_state.fixed_ranges;
1019
Marcelo Tosattid6289b92009-06-22 15:27:56 -03001020 if (!mtrr_valid(vcpu, msr, data))
Avi Kivity9ba075a2008-05-26 20:06:35 +03001021 return 1;
1022
Sheng Yang0bed3b52008-10-09 16:01:54 +08001023 if (msr == MSR_MTRRdefType) {
1024 vcpu->arch.mtrr_state.def_type = data;
1025 vcpu->arch.mtrr_state.enabled = (data & 0xc00) >> 10;
1026 } else if (msr == MSR_MTRRfix64K_00000)
1027 p[0] = data;
1028 else if (msr == MSR_MTRRfix16K_80000 || msr == MSR_MTRRfix16K_A0000)
1029 p[1 + msr - MSR_MTRRfix16K_80000] = data;
1030 else if (msr >= MSR_MTRRfix4K_C0000 && msr <= MSR_MTRRfix4K_F8000)
1031 p[3 + msr - MSR_MTRRfix4K_C0000] = data;
1032 else if (msr == MSR_IA32_CR_PAT)
1033 vcpu->arch.pat = data;
1034 else { /* Variable MTRRs */
1035 int idx, is_mtrr_mask;
1036 u64 *pt;
1037
1038 idx = (msr - 0x200) / 2;
1039 is_mtrr_mask = msr - 0x200 - 2 * idx;
1040 if (!is_mtrr_mask)
1041 pt =
1042 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].base_lo;
1043 else
1044 pt =
1045 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].mask_lo;
1046 *pt = data;
1047 }
1048
1049 kvm_mmu_reset_context(vcpu);
Avi Kivity9ba075a2008-05-26 20:06:35 +03001050 return 0;
1051}
Carsten Otte15c4a642007-10-30 18:44:17 +01001052
Huang Ying890ca9a2009-05-11 16:48:15 +08001053static int set_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 data)
1054{
1055 u64 mcg_cap = vcpu->arch.mcg_cap;
1056 unsigned bank_num = mcg_cap & 0xff;
1057
1058 switch (msr) {
1059 case MSR_IA32_MCG_STATUS:
1060 vcpu->arch.mcg_status = data;
1061 break;
1062 case MSR_IA32_MCG_CTL:
1063 if (!(mcg_cap & MCG_CTL_P))
1064 return 1;
1065 if (data != 0 && data != ~(u64)0)
1066 return -1;
1067 vcpu->arch.mcg_ctl = data;
1068 break;
1069 default:
1070 if (msr >= MSR_IA32_MC0_CTL &&
1071 msr < MSR_IA32_MC0_CTL + 4 * bank_num) {
1072 u32 offset = msr - MSR_IA32_MC0_CTL;
Andre Przywara114be422010-03-24 17:46:42 +01001073 /* only 0 or all 1s can be written to IA32_MCi_CTL
1074 * some Linux kernels though clear bit 10 in bank 4 to
1075 * workaround a BIOS/GART TBL issue on AMD K8s, ignore
1076 * this to avoid an uncatched #GP in the guest
1077 */
Huang Ying890ca9a2009-05-11 16:48:15 +08001078 if ((offset & 0x3) == 0 &&
Andre Przywara114be422010-03-24 17:46:42 +01001079 data != 0 && (data | (1 << 10)) != ~(u64)0)
Huang Ying890ca9a2009-05-11 16:48:15 +08001080 return -1;
1081 vcpu->arch.mce_banks[offset] = data;
1082 break;
1083 }
1084 return 1;
1085 }
1086 return 0;
1087}
1088
Ed Swierkffde22a2009-10-15 15:21:43 -07001089static int xen_hvm_config(struct kvm_vcpu *vcpu, u64 data)
1090{
1091 struct kvm *kvm = vcpu->kvm;
1092 int lm = is_long_mode(vcpu);
1093 u8 *blob_addr = lm ? (u8 *)(long)kvm->arch.xen_hvm_config.blob_addr_64
1094 : (u8 *)(long)kvm->arch.xen_hvm_config.blob_addr_32;
1095 u8 blob_size = lm ? kvm->arch.xen_hvm_config.blob_size_64
1096 : kvm->arch.xen_hvm_config.blob_size_32;
1097 u32 page_num = data & ~PAGE_MASK;
1098 u64 page_addr = data & PAGE_MASK;
1099 u8 *page;
1100 int r;
1101
1102 r = -E2BIG;
1103 if (page_num >= blob_size)
1104 goto out;
1105 r = -ENOMEM;
1106 page = kzalloc(PAGE_SIZE, GFP_KERNEL);
1107 if (!page)
1108 goto out;
1109 r = -EFAULT;
1110 if (copy_from_user(page, blob_addr + (page_num * PAGE_SIZE), PAGE_SIZE))
1111 goto out_free;
1112 if (kvm_write_guest(kvm, page_addr, page, PAGE_SIZE))
1113 goto out_free;
1114 r = 0;
1115out_free:
1116 kfree(page);
1117out:
1118 return r;
1119}
1120
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001121static bool kvm_hv_hypercall_enabled(struct kvm *kvm)
1122{
1123 return kvm->arch.hv_hypercall & HV_X64_MSR_HYPERCALL_ENABLE;
1124}
1125
1126static bool kvm_hv_msr_partition_wide(u32 msr)
1127{
1128 bool r = false;
1129 switch (msr) {
1130 case HV_X64_MSR_GUEST_OS_ID:
1131 case HV_X64_MSR_HYPERCALL:
1132 r = true;
1133 break;
1134 }
1135
1136 return r;
1137}
1138
1139static int set_msr_hyperv_pw(struct kvm_vcpu *vcpu, u32 msr, u64 data)
1140{
1141 struct kvm *kvm = vcpu->kvm;
1142
1143 switch (msr) {
1144 case HV_X64_MSR_GUEST_OS_ID:
1145 kvm->arch.hv_guest_os_id = data;
1146 /* setting guest os id to zero disables hypercall page */
1147 if (!kvm->arch.hv_guest_os_id)
1148 kvm->arch.hv_hypercall &= ~HV_X64_MSR_HYPERCALL_ENABLE;
1149 break;
1150 case HV_X64_MSR_HYPERCALL: {
1151 u64 gfn;
1152 unsigned long addr;
1153 u8 instructions[4];
1154
1155 /* if guest os id is not set hypercall should remain disabled */
1156 if (!kvm->arch.hv_guest_os_id)
1157 break;
1158 if (!(data & HV_X64_MSR_HYPERCALL_ENABLE)) {
1159 kvm->arch.hv_hypercall = data;
1160 break;
1161 }
1162 gfn = data >> HV_X64_MSR_HYPERCALL_PAGE_ADDRESS_SHIFT;
1163 addr = gfn_to_hva(kvm, gfn);
1164 if (kvm_is_error_hva(addr))
1165 return 1;
1166 kvm_x86_ops->patch_hypercall(vcpu, instructions);
1167 ((unsigned char *)instructions)[3] = 0xc3; /* ret */
1168 if (copy_to_user((void __user *)addr, instructions, 4))
1169 return 1;
1170 kvm->arch.hv_hypercall = data;
1171 break;
1172 }
1173 default:
1174 pr_unimpl(vcpu, "HYPER-V unimplemented wrmsr: 0x%x "
1175 "data 0x%llx\n", msr, data);
1176 return 1;
1177 }
1178 return 0;
1179}
1180
1181static int set_msr_hyperv(struct kvm_vcpu *vcpu, u32 msr, u64 data)
1182{
Gleb Natapov10388a02010-01-17 15:51:23 +02001183 switch (msr) {
1184 case HV_X64_MSR_APIC_ASSIST_PAGE: {
1185 unsigned long addr;
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001186
Gleb Natapov10388a02010-01-17 15:51:23 +02001187 if (!(data & HV_X64_MSR_APIC_ASSIST_PAGE_ENABLE)) {
1188 vcpu->arch.hv_vapic = data;
1189 break;
1190 }
1191 addr = gfn_to_hva(vcpu->kvm, data >>
1192 HV_X64_MSR_APIC_ASSIST_PAGE_ADDRESS_SHIFT);
1193 if (kvm_is_error_hva(addr))
1194 return 1;
1195 if (clear_user((void __user *)addr, PAGE_SIZE))
1196 return 1;
1197 vcpu->arch.hv_vapic = data;
1198 break;
1199 }
1200 case HV_X64_MSR_EOI:
1201 return kvm_hv_vapic_msr_write(vcpu, APIC_EOI, data);
1202 case HV_X64_MSR_ICR:
1203 return kvm_hv_vapic_msr_write(vcpu, APIC_ICR, data);
1204 case HV_X64_MSR_TPR:
1205 return kvm_hv_vapic_msr_write(vcpu, APIC_TASKPRI, data);
1206 default:
1207 pr_unimpl(vcpu, "HYPER-V unimplemented wrmsr: 0x%x "
1208 "data 0x%llx\n", msr, data);
1209 return 1;
1210 }
1211
1212 return 0;
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001213}
1214
Carsten Otte15c4a642007-10-30 18:44:17 +01001215int kvm_set_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 data)
1216{
1217 switch (msr) {
Carsten Otte15c4a642007-10-30 18:44:17 +01001218 case MSR_EFER:
Roedel, Joergb69e8ca2010-05-06 11:38:43 +02001219 return set_efer(vcpu, data);
Andre Przywara8f1589d2009-06-24 12:44:33 +02001220 case MSR_K7_HWCR:
1221 data &= ~(u64)0x40; /* ignore flush filter disable */
Joerg Roedel82494022010-02-24 18:59:16 +01001222 data &= ~(u64)0x100; /* ignore ignne emulation enable */
Andre Przywara8f1589d2009-06-24 12:44:33 +02001223 if (data != 0) {
1224 pr_unimpl(vcpu, "unimplemented HWCR wrmsr: 0x%llx\n",
1225 data);
1226 return 1;
1227 }
Carsten Otte15c4a642007-10-30 18:44:17 +01001228 break;
Andre Przywaraf7c6d142009-07-02 15:04:14 +02001229 case MSR_FAM10H_MMIO_CONF_BASE:
1230 if (data != 0) {
1231 pr_unimpl(vcpu, "unimplemented MMIO_CONF_BASE wrmsr: "
1232 "0x%llx\n", data);
1233 return 1;
1234 }
Carsten Otte15c4a642007-10-30 18:44:17 +01001235 break;
Andre Przywarac323c0e2009-06-24 15:37:05 +02001236 case MSR_AMD64_NB_CFG:
Joerg Roedelc7ac6792008-02-11 20:28:27 +01001237 break;
Alexander Grafb5e2fec2008-07-22 08:00:45 +02001238 case MSR_IA32_DEBUGCTLMSR:
1239 if (!data) {
1240 /* We support the non-activated case already */
1241 break;
1242 } else if (data & ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_BTF)) {
1243 /* Values other than LBR and BTF are vendor-specific,
1244 thus reserved and should throw a #GP */
1245 return 1;
1246 }
1247 pr_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTLMSR 0x%llx, nop\n",
1248 __func__, data);
1249 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001250 case MSR_IA32_UCODE_REV:
1251 case MSR_IA32_UCODE_WRITE:
Avi Kivity61a6bd62008-12-29 17:32:28 +02001252 case MSR_VM_HSAVE_PA:
Andre Przywara6098ca92009-07-03 16:00:14 +02001253 case MSR_AMD64_PATCH_LOADER:
Carsten Otte15c4a642007-10-30 18:44:17 +01001254 break;
Avi Kivity9ba075a2008-05-26 20:06:35 +03001255 case 0x200 ... 0x2ff:
1256 return set_msr_mtrr(vcpu, msr, data);
Carsten Otte15c4a642007-10-30 18:44:17 +01001257 case MSR_IA32_APICBASE:
1258 kvm_set_apic_base(vcpu, data);
1259 break;
Gleb Natapov0105d1a2009-07-05 17:39:36 +03001260 case APIC_BASE_MSR ... APIC_BASE_MSR + 0x3ff:
1261 return kvm_x2apic_msr_write(vcpu, msr, data);
Carsten Otte15c4a642007-10-30 18:44:17 +01001262 case MSR_IA32_MISC_ENABLE:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001263 vcpu->arch.ia32_misc_enable_msr = data;
Carsten Otte15c4a642007-10-30 18:44:17 +01001264 break;
Glauber Costa11c6bff2010-05-11 12:17:41 -04001265 case MSR_KVM_WALL_CLOCK_NEW:
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001266 case MSR_KVM_WALL_CLOCK:
1267 vcpu->kvm->arch.wall_clock = data;
1268 kvm_write_wall_clock(vcpu->kvm, data);
1269 break;
Glauber Costa11c6bff2010-05-11 12:17:41 -04001270 case MSR_KVM_SYSTEM_TIME_NEW:
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001271 case MSR_KVM_SYSTEM_TIME: {
1272 if (vcpu->arch.time_page) {
1273 kvm_release_page_dirty(vcpu->arch.time_page);
1274 vcpu->arch.time_page = NULL;
1275 }
1276
1277 vcpu->arch.time = data;
1278
1279 /* we verify if the enable bit is set... */
1280 if (!(data & 1))
1281 break;
1282
1283 /* ...but clean it before doing the actual write */
1284 vcpu->arch.time_offset = data & ~(PAGE_MASK | 1);
1285
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001286 vcpu->arch.time_page =
1287 gfn_to_page(vcpu->kvm, data >> PAGE_SHIFT);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001288
1289 if (is_error_page(vcpu->arch.time_page)) {
1290 kvm_release_page_clean(vcpu->arch.time_page);
1291 vcpu->arch.time_page = NULL;
1292 }
1293
Gerd Hoffmannc8076602009-02-04 17:52:04 +01001294 kvm_request_guest_time_update(vcpu);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001295 break;
1296 }
Huang Ying890ca9a2009-05-11 16:48:15 +08001297 case MSR_IA32_MCG_CTL:
1298 case MSR_IA32_MCG_STATUS:
1299 case MSR_IA32_MC0_CTL ... MSR_IA32_MC0_CTL + 4 * KVM_MAX_MCE_BANKS - 1:
1300 return set_msr_mce(vcpu, msr, data);
Andre Przywara71db6022009-06-12 22:01:29 +02001301
1302 /* Performance counters are not protected by a CPUID bit,
1303 * so we should check all of them in the generic path for the sake of
1304 * cross vendor migration.
1305 * Writing a zero into the event select MSRs disables them,
1306 * which we perfectly emulate ;-). Any other value should be at least
1307 * reported, some guests depend on them.
1308 */
1309 case MSR_P6_EVNTSEL0:
1310 case MSR_P6_EVNTSEL1:
1311 case MSR_K7_EVNTSEL0:
1312 case MSR_K7_EVNTSEL1:
1313 case MSR_K7_EVNTSEL2:
1314 case MSR_K7_EVNTSEL3:
1315 if (data != 0)
1316 pr_unimpl(vcpu, "unimplemented perfctr wrmsr: "
1317 "0x%x data 0x%llx\n", msr, data);
1318 break;
1319 /* at least RHEL 4 unconditionally writes to the perfctr registers,
1320 * so we ignore writes to make it happy.
1321 */
1322 case MSR_P6_PERFCTR0:
1323 case MSR_P6_PERFCTR1:
1324 case MSR_K7_PERFCTR0:
1325 case MSR_K7_PERFCTR1:
1326 case MSR_K7_PERFCTR2:
1327 case MSR_K7_PERFCTR3:
1328 pr_unimpl(vcpu, "unimplemented perfctr wrmsr: "
1329 "0x%x data 0x%llx\n", msr, data);
1330 break;
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001331 case HV_X64_MSR_GUEST_OS_ID ... HV_X64_MSR_SINT15:
1332 if (kvm_hv_msr_partition_wide(msr)) {
1333 int r;
1334 mutex_lock(&vcpu->kvm->lock);
1335 r = set_msr_hyperv_pw(vcpu, msr, data);
1336 mutex_unlock(&vcpu->kvm->lock);
1337 return r;
1338 } else
1339 return set_msr_hyperv(vcpu, msr, data);
1340 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001341 default:
Ed Swierkffde22a2009-10-15 15:21:43 -07001342 if (msr && (msr == vcpu->kvm->arch.xen_hvm_config.msr))
1343 return xen_hvm_config(vcpu, data);
Andre Przywaraed85c062009-06-25 12:36:49 +02001344 if (!ignore_msrs) {
1345 pr_unimpl(vcpu, "unhandled wrmsr: 0x%x data %llx\n",
1346 msr, data);
1347 return 1;
1348 } else {
1349 pr_unimpl(vcpu, "ignored wrmsr: 0x%x data %llx\n",
1350 msr, data);
1351 break;
1352 }
Carsten Otte15c4a642007-10-30 18:44:17 +01001353 }
1354 return 0;
1355}
1356EXPORT_SYMBOL_GPL(kvm_set_msr_common);
1357
1358
1359/*
1360 * Reads an msr value (of 'msr_index') into 'pdata'.
1361 * Returns 0 on success, non-0 otherwise.
1362 * Assumes vcpu_load() was already called.
1363 */
1364int kvm_get_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 *pdata)
1365{
1366 return kvm_x86_ops->get_msr(vcpu, msr_index, pdata);
1367}
1368
Avi Kivity9ba075a2008-05-26 20:06:35 +03001369static int get_msr_mtrr(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1370{
Sheng Yang0bed3b52008-10-09 16:01:54 +08001371 u64 *p = (u64 *)&vcpu->arch.mtrr_state.fixed_ranges;
1372
Avi Kivity9ba075a2008-05-26 20:06:35 +03001373 if (!msr_mtrr_valid(msr))
1374 return 1;
1375
Sheng Yang0bed3b52008-10-09 16:01:54 +08001376 if (msr == MSR_MTRRdefType)
1377 *pdata = vcpu->arch.mtrr_state.def_type +
1378 (vcpu->arch.mtrr_state.enabled << 10);
1379 else if (msr == MSR_MTRRfix64K_00000)
1380 *pdata = p[0];
1381 else if (msr == MSR_MTRRfix16K_80000 || msr == MSR_MTRRfix16K_A0000)
1382 *pdata = p[1 + msr - MSR_MTRRfix16K_80000];
1383 else if (msr >= MSR_MTRRfix4K_C0000 && msr <= MSR_MTRRfix4K_F8000)
1384 *pdata = p[3 + msr - MSR_MTRRfix4K_C0000];
1385 else if (msr == MSR_IA32_CR_PAT)
1386 *pdata = vcpu->arch.pat;
1387 else { /* Variable MTRRs */
1388 int idx, is_mtrr_mask;
1389 u64 *pt;
1390
1391 idx = (msr - 0x200) / 2;
1392 is_mtrr_mask = msr - 0x200 - 2 * idx;
1393 if (!is_mtrr_mask)
1394 pt =
1395 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].base_lo;
1396 else
1397 pt =
1398 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].mask_lo;
1399 *pdata = *pt;
1400 }
1401
Avi Kivity9ba075a2008-05-26 20:06:35 +03001402 return 0;
1403}
1404
Huang Ying890ca9a2009-05-11 16:48:15 +08001405static int get_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1406{
1407 u64 data;
1408 u64 mcg_cap = vcpu->arch.mcg_cap;
1409 unsigned bank_num = mcg_cap & 0xff;
1410
1411 switch (msr) {
1412 case MSR_IA32_P5_MC_ADDR:
1413 case MSR_IA32_P5_MC_TYPE:
1414 data = 0;
1415 break;
1416 case MSR_IA32_MCG_CAP:
1417 data = vcpu->arch.mcg_cap;
1418 break;
1419 case MSR_IA32_MCG_CTL:
1420 if (!(mcg_cap & MCG_CTL_P))
1421 return 1;
1422 data = vcpu->arch.mcg_ctl;
1423 break;
1424 case MSR_IA32_MCG_STATUS:
1425 data = vcpu->arch.mcg_status;
1426 break;
1427 default:
1428 if (msr >= MSR_IA32_MC0_CTL &&
1429 msr < MSR_IA32_MC0_CTL + 4 * bank_num) {
1430 u32 offset = msr - MSR_IA32_MC0_CTL;
1431 data = vcpu->arch.mce_banks[offset];
1432 break;
1433 }
1434 return 1;
1435 }
1436 *pdata = data;
1437 return 0;
1438}
1439
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001440static int get_msr_hyperv_pw(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1441{
1442 u64 data = 0;
1443 struct kvm *kvm = vcpu->kvm;
1444
1445 switch (msr) {
1446 case HV_X64_MSR_GUEST_OS_ID:
1447 data = kvm->arch.hv_guest_os_id;
1448 break;
1449 case HV_X64_MSR_HYPERCALL:
1450 data = kvm->arch.hv_hypercall;
1451 break;
1452 default:
1453 pr_unimpl(vcpu, "Hyper-V unhandled rdmsr: 0x%x\n", msr);
1454 return 1;
1455 }
1456
1457 *pdata = data;
1458 return 0;
1459}
1460
1461static int get_msr_hyperv(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1462{
1463 u64 data = 0;
1464
1465 switch (msr) {
1466 case HV_X64_MSR_VP_INDEX: {
1467 int r;
1468 struct kvm_vcpu *v;
1469 kvm_for_each_vcpu(r, v, vcpu->kvm)
1470 if (v == vcpu)
1471 data = r;
1472 break;
1473 }
Gleb Natapov10388a02010-01-17 15:51:23 +02001474 case HV_X64_MSR_EOI:
1475 return kvm_hv_vapic_msr_read(vcpu, APIC_EOI, pdata);
1476 case HV_X64_MSR_ICR:
1477 return kvm_hv_vapic_msr_read(vcpu, APIC_ICR, pdata);
1478 case HV_X64_MSR_TPR:
1479 return kvm_hv_vapic_msr_read(vcpu, APIC_TASKPRI, pdata);
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001480 default:
1481 pr_unimpl(vcpu, "Hyper-V unhandled rdmsr: 0x%x\n", msr);
1482 return 1;
1483 }
1484 *pdata = data;
1485 return 0;
1486}
1487
Carsten Otte15c4a642007-10-30 18:44:17 +01001488int kvm_get_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1489{
1490 u64 data;
1491
1492 switch (msr) {
Carsten Otte15c4a642007-10-30 18:44:17 +01001493 case MSR_IA32_PLATFORM_ID:
Carsten Otte15c4a642007-10-30 18:44:17 +01001494 case MSR_IA32_UCODE_REV:
Carsten Otte15c4a642007-10-30 18:44:17 +01001495 case MSR_IA32_EBL_CR_POWERON:
Alexander Grafb5e2fec2008-07-22 08:00:45 +02001496 case MSR_IA32_DEBUGCTLMSR:
1497 case MSR_IA32_LASTBRANCHFROMIP:
1498 case MSR_IA32_LASTBRANCHTOIP:
1499 case MSR_IA32_LASTINTFROMIP:
1500 case MSR_IA32_LASTINTTOIP:
Jaswinder Singh Rajput60af2ec2009-05-14 11:00:10 +05301501 case MSR_K8_SYSCFG:
1502 case MSR_K7_HWCR:
Avi Kivity61a6bd62008-12-29 17:32:28 +02001503 case MSR_VM_HSAVE_PA:
Amit Shah1f3ee612009-06-30 16:24:28 +05301504 case MSR_P6_PERFCTR0:
1505 case MSR_P6_PERFCTR1:
Amit Shah7fe29e02009-03-20 12:39:00 +05301506 case MSR_P6_EVNTSEL0:
1507 case MSR_P6_EVNTSEL1:
Amit Shah9e699622009-06-15 13:25:34 +05301508 case MSR_K7_EVNTSEL0:
Amit Shah1f3ee612009-06-30 16:24:28 +05301509 case MSR_K7_PERFCTR0:
Andre Przywara1fdbd482009-06-24 12:44:34 +02001510 case MSR_K8_INT_PENDING_MSG:
Andre Przywarac323c0e2009-06-24 15:37:05 +02001511 case MSR_AMD64_NB_CFG:
Andre Przywaraf7c6d142009-07-02 15:04:14 +02001512 case MSR_FAM10H_MMIO_CONF_BASE:
Carsten Otte15c4a642007-10-30 18:44:17 +01001513 data = 0;
1514 break;
Avi Kivity9ba075a2008-05-26 20:06:35 +03001515 case MSR_MTRRcap:
1516 data = 0x500 | KVM_NR_VAR_MTRR;
1517 break;
1518 case 0x200 ... 0x2ff:
1519 return get_msr_mtrr(vcpu, msr, pdata);
Carsten Otte15c4a642007-10-30 18:44:17 +01001520 case 0xcd: /* fsb frequency */
1521 data = 3;
1522 break;
1523 case MSR_IA32_APICBASE:
1524 data = kvm_get_apic_base(vcpu);
1525 break;
Gleb Natapov0105d1a2009-07-05 17:39:36 +03001526 case APIC_BASE_MSR ... APIC_BASE_MSR + 0x3ff:
1527 return kvm_x2apic_msr_read(vcpu, msr, pdata);
1528 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001529 case MSR_IA32_MISC_ENABLE:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001530 data = vcpu->arch.ia32_misc_enable_msr;
Carsten Otte15c4a642007-10-30 18:44:17 +01001531 break;
Alexander Graf847f0ad2008-02-21 12:11:01 +01001532 case MSR_IA32_PERF_STATUS:
1533 /* TSC increment by tick */
1534 data = 1000ULL;
1535 /* CPU multiplier */
1536 data |= (((uint64_t)4ULL) << 40);
1537 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001538 case MSR_EFER:
Avi Kivityf6801df2010-01-21 15:31:50 +02001539 data = vcpu->arch.efer;
Carsten Otte15c4a642007-10-30 18:44:17 +01001540 break;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001541 case MSR_KVM_WALL_CLOCK:
Glauber Costa11c6bff2010-05-11 12:17:41 -04001542 case MSR_KVM_WALL_CLOCK_NEW:
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001543 data = vcpu->kvm->arch.wall_clock;
1544 break;
1545 case MSR_KVM_SYSTEM_TIME:
Glauber Costa11c6bff2010-05-11 12:17:41 -04001546 case MSR_KVM_SYSTEM_TIME_NEW:
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001547 data = vcpu->arch.time;
1548 break;
Huang Ying890ca9a2009-05-11 16:48:15 +08001549 case MSR_IA32_P5_MC_ADDR:
1550 case MSR_IA32_P5_MC_TYPE:
1551 case MSR_IA32_MCG_CAP:
1552 case MSR_IA32_MCG_CTL:
1553 case MSR_IA32_MCG_STATUS:
1554 case MSR_IA32_MC0_CTL ... MSR_IA32_MC0_CTL + 4 * KVM_MAX_MCE_BANKS - 1:
1555 return get_msr_mce(vcpu, msr, pdata);
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001556 case HV_X64_MSR_GUEST_OS_ID ... HV_X64_MSR_SINT15:
1557 if (kvm_hv_msr_partition_wide(msr)) {
1558 int r;
1559 mutex_lock(&vcpu->kvm->lock);
1560 r = get_msr_hyperv_pw(vcpu, msr, pdata);
1561 mutex_unlock(&vcpu->kvm->lock);
1562 return r;
1563 } else
1564 return get_msr_hyperv(vcpu, msr, pdata);
1565 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001566 default:
Andre Przywaraed85c062009-06-25 12:36:49 +02001567 if (!ignore_msrs) {
1568 pr_unimpl(vcpu, "unhandled rdmsr: 0x%x\n", msr);
1569 return 1;
1570 } else {
1571 pr_unimpl(vcpu, "ignored rdmsr: 0x%x\n", msr);
1572 data = 0;
1573 }
1574 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001575 }
1576 *pdata = data;
1577 return 0;
1578}
1579EXPORT_SYMBOL_GPL(kvm_get_msr_common);
1580
Carsten Otte313a3dc2007-10-11 19:16:52 +02001581/*
1582 * Read or write a bunch of msrs. All parameters are kernel addresses.
1583 *
1584 * @return number of msrs set successfully.
1585 */
1586static int __msr_io(struct kvm_vcpu *vcpu, struct kvm_msrs *msrs,
1587 struct kvm_msr_entry *entries,
1588 int (*do_msr)(struct kvm_vcpu *vcpu,
1589 unsigned index, u64 *data))
1590{
Marcelo Tosattif656ce02009-12-23 14:35:25 -02001591 int i, idx;
Carsten Otte313a3dc2007-10-11 19:16:52 +02001592
Marcelo Tosattif656ce02009-12-23 14:35:25 -02001593 idx = srcu_read_lock(&vcpu->kvm->srcu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001594 for (i = 0; i < msrs->nmsrs; ++i)
1595 if (do_msr(vcpu, entries[i].index, &entries[i].data))
1596 break;
Marcelo Tosattif656ce02009-12-23 14:35:25 -02001597 srcu_read_unlock(&vcpu->kvm->srcu, idx);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001598
Carsten Otte313a3dc2007-10-11 19:16:52 +02001599 return i;
1600}
1601
1602/*
1603 * Read or write a bunch of msrs. Parameters are user addresses.
1604 *
1605 * @return number of msrs set successfully.
1606 */
1607static int msr_io(struct kvm_vcpu *vcpu, struct kvm_msrs __user *user_msrs,
1608 int (*do_msr)(struct kvm_vcpu *vcpu,
1609 unsigned index, u64 *data),
1610 int writeback)
1611{
1612 struct kvm_msrs msrs;
1613 struct kvm_msr_entry *entries;
1614 int r, n;
1615 unsigned size;
1616
1617 r = -EFAULT;
1618 if (copy_from_user(&msrs, user_msrs, sizeof msrs))
1619 goto out;
1620
1621 r = -E2BIG;
1622 if (msrs.nmsrs >= MAX_IO_MSRS)
1623 goto out;
1624
1625 r = -ENOMEM;
1626 size = sizeof(struct kvm_msr_entry) * msrs.nmsrs;
Avi Kivity7a73c022010-07-22 23:24:52 +03001627 entries = kmalloc(size, GFP_KERNEL);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001628 if (!entries)
1629 goto out;
1630
1631 r = -EFAULT;
1632 if (copy_from_user(entries, user_msrs->entries, size))
1633 goto out_free;
1634
1635 r = n = __msr_io(vcpu, &msrs, entries, do_msr);
1636 if (r < 0)
1637 goto out_free;
1638
1639 r = -EFAULT;
1640 if (writeback && copy_to_user(user_msrs->entries, entries, size))
1641 goto out_free;
1642
1643 r = n;
1644
1645out_free:
Avi Kivity7a73c022010-07-22 23:24:52 +03001646 kfree(entries);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001647out:
1648 return r;
1649}
1650
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001651int kvm_dev_ioctl_check_extension(long ext)
1652{
1653 int r;
1654
1655 switch (ext) {
1656 case KVM_CAP_IRQCHIP:
1657 case KVM_CAP_HLT:
1658 case KVM_CAP_MMU_SHADOW_CACHE_CONTROL:
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001659 case KVM_CAP_SET_TSS_ADDR:
Dan Kenigsberg07716712007-11-21 17:10:04 +02001660 case KVM_CAP_EXT_CPUID:
Gerd Hoffmannc8076602009-02-04 17:52:04 +01001661 case KVM_CAP_CLOCKSOURCE:
Sheng Yang78376992008-01-28 05:10:22 +08001662 case KVM_CAP_PIT:
Marcelo Tosattia28e4f52008-02-22 12:21:36 -05001663 case KVM_CAP_NOP_IO_DELAY:
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03001664 case KVM_CAP_MP_STATE:
Avi Kivityed848622008-07-29 11:30:57 +03001665 case KVM_CAP_SYNC_MMU:
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02001666 case KVM_CAP_REINJECT_CONTROL:
Gleb Natapov49256632009-02-04 17:28:14 +02001667 case KVM_CAP_IRQ_INJECT_STATUS:
Sheng Yange56d5322009-03-12 21:45:39 +08001668 case KVM_CAP_ASSIGN_DEV_IRQ:
Gregory Haskins721eecbf2009-05-20 10:30:49 -04001669 case KVM_CAP_IRQFD:
Gregory Haskinsd34e6b12009-07-07 17:08:49 -04001670 case KVM_CAP_IOEVENTFD:
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02001671 case KVM_CAP_PIT2:
Beth Kone9f42752009-07-07 11:50:38 -04001672 case KVM_CAP_PIT_STATE2:
Sheng Yangb927a3c2009-07-21 10:42:48 +08001673 case KVM_CAP_SET_IDENTITY_MAP_ADDR:
Ed Swierkffde22a2009-10-15 15:21:43 -07001674 case KVM_CAP_XEN_HVM:
Glauber Costaafbcf7a2009-10-16 15:28:36 -04001675 case KVM_CAP_ADJUST_CLOCK:
Jan Kiszka3cfc3092009-11-12 01:04:25 +01001676 case KVM_CAP_VCPU_EVENTS:
Gleb Natapov55cd8e52010-01-17 15:51:22 +02001677 case KVM_CAP_HYPERV:
Gleb Natapov10388a02010-01-17 15:51:23 +02001678 case KVM_CAP_HYPERV_VAPIC:
Gleb Natapovc25bc162010-01-17 15:51:24 +02001679 case KVM_CAP_HYPERV_SPIN:
Zhai, Edwinab9f4ec2010-01-29 14:38:44 +08001680 case KVM_CAP_PCI_SEGMENT:
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01001681 case KVM_CAP_DEBUGREGS:
Jan Kiszkad2be1652010-02-23 17:47:57 +01001682 case KVM_CAP_X86_ROBUST_SINGLESTEP:
Sheng Yang2d5b5a62010-06-13 17:29:39 +08001683 case KVM_CAP_XSAVE:
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001684 r = 1;
1685 break;
Laurent Vivier542472b2008-05-30 16:05:55 +02001686 case KVM_CAP_COALESCED_MMIO:
1687 r = KVM_COALESCED_MMIO_PAGE_OFFSET;
1688 break;
Avi Kivity774ead32007-12-26 13:57:04 +02001689 case KVM_CAP_VAPIC:
1690 r = !kvm_x86_ops->cpu_has_accelerated_tpr();
1691 break;
Avi Kivityf7252302008-02-20 11:53:16 +02001692 case KVM_CAP_NR_VCPUS:
1693 r = KVM_MAX_VCPUS;
1694 break;
Avi Kivitya988b912008-02-20 11:59:20 +02001695 case KVM_CAP_NR_MEMSLOTS:
1696 r = KVM_MEMORY_SLOTS;
1697 break;
Marcelo Tosattia68a6a72009-10-01 19:28:39 -03001698 case KVM_CAP_PV_MMU: /* obsolete */
1699 r = 0;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05001700 break;
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +03001701 case KVM_CAP_IOMMU:
Joerg Roedel19de40a2008-12-03 14:43:34 +01001702 r = iommu_found();
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +03001703 break;
Huang Ying890ca9a2009-05-11 16:48:15 +08001704 case KVM_CAP_MCE:
1705 r = KVM_MAX_MCE_BANKS;
1706 break;
Sheng Yang2d5b5a62010-06-13 17:29:39 +08001707 case KVM_CAP_XCRS:
1708 r = cpu_has_xsave;
1709 break;
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001710 default:
1711 r = 0;
1712 break;
1713 }
1714 return r;
1715
1716}
1717
Carsten Otte043405e2007-10-10 17:16:19 +02001718long kvm_arch_dev_ioctl(struct file *filp,
1719 unsigned int ioctl, unsigned long arg)
1720{
1721 void __user *argp = (void __user *)arg;
1722 long r;
1723
1724 switch (ioctl) {
1725 case KVM_GET_MSR_INDEX_LIST: {
1726 struct kvm_msr_list __user *user_msr_list = argp;
1727 struct kvm_msr_list msr_list;
1728 unsigned n;
1729
1730 r = -EFAULT;
1731 if (copy_from_user(&msr_list, user_msr_list, sizeof msr_list))
1732 goto out;
1733 n = msr_list.nmsrs;
1734 msr_list.nmsrs = num_msrs_to_save + ARRAY_SIZE(emulated_msrs);
1735 if (copy_to_user(user_msr_list, &msr_list, sizeof msr_list))
1736 goto out;
1737 r = -E2BIG;
Jan Kiszkae125e7b2009-07-02 21:45:47 +02001738 if (n < msr_list.nmsrs)
Carsten Otte043405e2007-10-10 17:16:19 +02001739 goto out;
1740 r = -EFAULT;
1741 if (copy_to_user(user_msr_list->indices, &msrs_to_save,
1742 num_msrs_to_save * sizeof(u32)))
1743 goto out;
Jan Kiszkae125e7b2009-07-02 21:45:47 +02001744 if (copy_to_user(user_msr_list->indices + num_msrs_to_save,
Carsten Otte043405e2007-10-10 17:16:19 +02001745 &emulated_msrs,
1746 ARRAY_SIZE(emulated_msrs) * sizeof(u32)))
1747 goto out;
1748 r = 0;
1749 break;
1750 }
Avi Kivity674eea02008-02-11 18:37:23 +02001751 case KVM_GET_SUPPORTED_CPUID: {
1752 struct kvm_cpuid2 __user *cpuid_arg = argp;
1753 struct kvm_cpuid2 cpuid;
1754
1755 r = -EFAULT;
1756 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
1757 goto out;
1758 r = kvm_dev_ioctl_get_supported_cpuid(&cpuid,
Amit Shah19355472009-01-14 16:56:00 +00001759 cpuid_arg->entries);
Avi Kivity674eea02008-02-11 18:37:23 +02001760 if (r)
1761 goto out;
1762
1763 r = -EFAULT;
1764 if (copy_to_user(cpuid_arg, &cpuid, sizeof cpuid))
1765 goto out;
1766 r = 0;
1767 break;
1768 }
Huang Ying890ca9a2009-05-11 16:48:15 +08001769 case KVM_X86_GET_MCE_CAP_SUPPORTED: {
1770 u64 mce_cap;
1771
1772 mce_cap = KVM_MCE_CAP_SUPPORTED;
1773 r = -EFAULT;
1774 if (copy_to_user(argp, &mce_cap, sizeof mce_cap))
1775 goto out;
1776 r = 0;
1777 break;
1778 }
Carsten Otte043405e2007-10-10 17:16:19 +02001779 default:
1780 r = -EINVAL;
1781 }
1782out:
1783 return r;
1784}
1785
Carsten Otte313a3dc2007-10-11 19:16:52 +02001786void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1787{
1788 kvm_x86_ops->vcpu_load(vcpu, cpu);
Zachary Amsden6b7d7e72009-10-09 16:26:08 -10001789 if (unlikely(per_cpu(cpu_tsc_khz, cpu) == 0)) {
1790 unsigned long khz = cpufreq_quick_get(cpu);
1791 if (!khz)
1792 khz = tsc_khz;
1793 per_cpu(cpu_tsc_khz, cpu) = khz;
1794 }
Gerd Hoffmannc8076602009-02-04 17:52:04 +01001795 kvm_request_guest_time_update(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001796}
1797
1798void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
1799{
Avi Kivity02daab22009-12-30 12:40:26 +02001800 kvm_x86_ops->vcpu_put(vcpu);
Avi Kivity1c11e712010-05-03 16:05:44 +03001801 kvm_put_guest_fpu(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001802}
1803
Dan Kenigsberg07716712007-11-21 17:10:04 +02001804static int is_efer_nx(void)
Carsten Otte313a3dc2007-10-11 19:16:52 +02001805{
Avi Kivitye286e862009-05-03 18:50:55 +03001806 unsigned long long efer = 0;
Carsten Otte313a3dc2007-10-11 19:16:52 +02001807
Avi Kivitye286e862009-05-03 18:50:55 +03001808 rdmsrl_safe(MSR_EFER, &efer);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001809 return efer & EFER_NX;
1810}
1811
1812static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
1813{
1814 int i;
1815 struct kvm_cpuid_entry2 *e, *entry;
1816
Carsten Otte313a3dc2007-10-11 19:16:52 +02001817 entry = NULL;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001818 for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
1819 e = &vcpu->arch.cpuid_entries[i];
Carsten Otte313a3dc2007-10-11 19:16:52 +02001820 if (e->function == 0x80000001) {
1821 entry = e;
1822 break;
1823 }
1824 }
Dan Kenigsberg07716712007-11-21 17:10:04 +02001825 if (entry && (entry->edx & (1 << 20)) && !is_efer_nx()) {
Carsten Otte313a3dc2007-10-11 19:16:52 +02001826 entry->edx &= ~(1 << 20);
1827 printk(KERN_INFO "kvm: guest NX capability removed\n");
1828 }
1829}
1830
Dan Kenigsberg07716712007-11-21 17:10:04 +02001831/* when an old userspace process fills a new kernel module */
Carsten Otte313a3dc2007-10-11 19:16:52 +02001832static int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
1833 struct kvm_cpuid *cpuid,
1834 struct kvm_cpuid_entry __user *entries)
1835{
Dan Kenigsberg07716712007-11-21 17:10:04 +02001836 int r, i;
1837 struct kvm_cpuid_entry *cpuid_entries;
1838
1839 r = -E2BIG;
1840 if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1841 goto out;
1842 r = -ENOMEM;
1843 cpuid_entries = vmalloc(sizeof(struct kvm_cpuid_entry) * cpuid->nent);
1844 if (!cpuid_entries)
1845 goto out;
1846 r = -EFAULT;
1847 if (copy_from_user(cpuid_entries, entries,
1848 cpuid->nent * sizeof(struct kvm_cpuid_entry)))
1849 goto out_free;
1850 for (i = 0; i < cpuid->nent; i++) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001851 vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
1852 vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
1853 vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
1854 vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
1855 vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
1856 vcpu->arch.cpuid_entries[i].index = 0;
1857 vcpu->arch.cpuid_entries[i].flags = 0;
1858 vcpu->arch.cpuid_entries[i].padding[0] = 0;
1859 vcpu->arch.cpuid_entries[i].padding[1] = 0;
1860 vcpu->arch.cpuid_entries[i].padding[2] = 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001861 }
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001862 vcpu->arch.cpuid_nent = cpuid->nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001863 cpuid_fix_nx_cap(vcpu);
1864 r = 0;
Gleb Natapovfc61b802009-07-05 17:39:35 +03001865 kvm_apic_set_version(vcpu);
Sheng Yang0e851882009-12-18 16:48:46 +08001866 kvm_x86_ops->cpuid_update(vcpu);
Dexuan Cui2acf9232010-06-10 11:27:12 +08001867 update_cpuid(vcpu);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001868
1869out_free:
1870 vfree(cpuid_entries);
1871out:
1872 return r;
1873}
1874
1875static int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
Amit Shah19355472009-01-14 16:56:00 +00001876 struct kvm_cpuid2 *cpuid,
1877 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001878{
Carsten Otte313a3dc2007-10-11 19:16:52 +02001879 int r;
1880
1881 r = -E2BIG;
1882 if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1883 goto out;
1884 r = -EFAULT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001885 if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
Dan Kenigsberg07716712007-11-21 17:10:04 +02001886 cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
Carsten Otte313a3dc2007-10-11 19:16:52 +02001887 goto out;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001888 vcpu->arch.cpuid_nent = cpuid->nent;
Gleb Natapovfc61b802009-07-05 17:39:35 +03001889 kvm_apic_set_version(vcpu);
Sheng Yang0e851882009-12-18 16:48:46 +08001890 kvm_x86_ops->cpuid_update(vcpu);
Dexuan Cui2acf9232010-06-10 11:27:12 +08001891 update_cpuid(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001892 return 0;
1893
1894out:
1895 return r;
1896}
1897
Dan Kenigsberg07716712007-11-21 17:10:04 +02001898static int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
Amit Shah19355472009-01-14 16:56:00 +00001899 struct kvm_cpuid2 *cpuid,
1900 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001901{
1902 int r;
1903
1904 r = -E2BIG;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001905 if (cpuid->nent < vcpu->arch.cpuid_nent)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001906 goto out;
1907 r = -EFAULT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001908 if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
Amit Shah19355472009-01-14 16:56:00 +00001909 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
Dan Kenigsberg07716712007-11-21 17:10:04 +02001910 goto out;
1911 return 0;
1912
1913out:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001914 cpuid->nent = vcpu->arch.cpuid_nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001915 return r;
1916}
1917
Dan Kenigsberg07716712007-11-21 17:10:04 +02001918static void do_cpuid_1_ent(struct kvm_cpuid_entry2 *entry, u32 function,
Amit Shah19355472009-01-14 16:56:00 +00001919 u32 index)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001920{
1921 entry->function = function;
1922 entry->index = index;
1923 cpuid_count(entry->function, entry->index,
Amit Shah19355472009-01-14 16:56:00 +00001924 &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001925 entry->flags = 0;
1926}
1927
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001928#define F(x) bit(X86_FEATURE_##x)
1929
Dan Kenigsberg07716712007-11-21 17:10:04 +02001930static void do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
1931 u32 index, int *nent, int maxnent)
1932{
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001933 unsigned f_nx = is_efer_nx() ? F(NX) : 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001934#ifdef CONFIG_X86_64
Sheng Yang17cc3932010-01-05 19:02:27 +08001935 unsigned f_gbpages = (kvm_x86_ops->get_lpage_level() == PT_PDPE_LEVEL)
1936 ? F(GBPAGES) : 0;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001937 unsigned f_lm = F(LM);
1938#else
Sheng Yang17cc3932010-01-05 19:02:27 +08001939 unsigned f_gbpages = 0;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001940 unsigned f_lm = 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001941#endif
Sheng Yang4e47c7a2009-12-18 16:48:47 +08001942 unsigned f_rdtscp = kvm_x86_ops->rdtscp_supported() ? F(RDTSCP) : 0;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001943
1944 /* cpuid 1.edx */
1945 const u32 kvm_supported_word0_x86_features =
1946 F(FPU) | F(VME) | F(DE) | F(PSE) |
1947 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
1948 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
1949 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
1950 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLSH) |
1951 0 /* Reserved, DS, ACPI */ | F(MMX) |
1952 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
1953 0 /* HTT, TM, Reserved, PBE */;
1954 /* cpuid 0x80000001.edx */
1955 const u32 kvm_supported_word1_x86_features =
1956 F(FPU) | F(VME) | F(DE) | F(PSE) |
1957 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
1958 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
1959 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
1960 F(PAT) | F(PSE36) | 0 /* Reserved */ |
1961 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
Sheng Yang4e47c7a2009-12-18 16:48:47 +08001962 F(FXSR) | F(FXSR_OPT) | f_gbpages | f_rdtscp |
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001963 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
1964 /* cpuid 1.ecx */
1965 const u32 kvm_supported_word4_x86_features =
Avi Kivityd149c732009-05-10 14:41:56 +03001966 F(XMM3) | 0 /* Reserved, DTES64, MONITOR */ |
1967 0 /* DS-CPL, VMX, SMX, EST */ |
1968 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
1969 0 /* Reserved */ | F(CX16) | 0 /* xTPR Update, PDCM */ |
1970 0 /* Reserved, DCA */ | F(XMM4_1) |
Gleb Natapov0105d1a2009-07-05 17:39:36 +03001971 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
Dexuan Cui2acf9232010-06-10 11:27:12 +08001972 0 /* Reserved, AES */ | F(XSAVE) | 0 /* OSXSAVE */;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001973 /* cpuid 0x80000001.ecx */
Dan Kenigsberg07716712007-11-21 17:10:04 +02001974 const u32 kvm_supported_word6_x86_features =
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001975 F(LAHF_LM) | F(CMP_LEGACY) | F(SVM) | 0 /* ExtApicSpace */ |
1976 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
1977 F(3DNOWPREFETCH) | 0 /* OSVW */ | 0 /* IBS */ | F(SSE5) |
1978 0 /* SKINIT */ | 0 /* WDT */;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001979
Amit Shah19355472009-01-14 16:56:00 +00001980 /* all calls to cpuid_count() should be made on the same cpu */
Dan Kenigsberg07716712007-11-21 17:10:04 +02001981 get_cpu();
1982 do_cpuid_1_ent(entry, function, index);
1983 ++*nent;
1984
1985 switch (function) {
1986 case 0:
Dexuan Cui2acf9232010-06-10 11:27:12 +08001987 entry->eax = min(entry->eax, (u32)0xd);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001988 break;
1989 case 1:
1990 entry->edx &= kvm_supported_word0_x86_features;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001991 entry->ecx &= kvm_supported_word4_x86_features;
Gleb Natapov0d1de2d92009-07-12 16:10:55 +03001992 /* we support x2apic emulation even if host does not support
1993 * it since we emulate x2apic in software */
1994 entry->ecx |= F(X2APIC);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001995 break;
1996 /* function 2 entries are STATEFUL. That is, repeated cpuid commands
1997 * may return different values. This forces us to get_cpu() before
1998 * issuing the first command, and also to emulate this annoying behavior
1999 * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
2000 case 2: {
2001 int t, times = entry->eax & 0xff;
2002
2003 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
Nitin A Kamble0fdf8e52008-11-05 15:56:21 -08002004 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002005 for (t = 1; t < times && *nent < maxnent; ++t) {
2006 do_cpuid_1_ent(&entry[t], function, 0);
2007 entry[t].flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
2008 ++*nent;
2009 }
2010 break;
2011 }
2012 /* function 4 and 0xb have additional index. */
2013 case 4: {
Harvey Harrison14af3f32008-02-19 10:25:50 -08002014 int i, cache_type;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002015
2016 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2017 /* read more entries until cache_type is zero */
Harvey Harrison14af3f32008-02-19 10:25:50 -08002018 for (i = 1; *nent < maxnent; ++i) {
2019 cache_type = entry[i - 1].eax & 0x1f;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002020 if (!cache_type)
2021 break;
Harvey Harrison14af3f32008-02-19 10:25:50 -08002022 do_cpuid_1_ent(&entry[i], function, i);
2023 entry[i].flags |=
Dan Kenigsberg07716712007-11-21 17:10:04 +02002024 KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2025 ++*nent;
2026 }
2027 break;
2028 }
2029 case 0xb: {
Harvey Harrison14af3f32008-02-19 10:25:50 -08002030 int i, level_type;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002031
2032 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2033 /* read more entries until level_type is zero */
Harvey Harrison14af3f32008-02-19 10:25:50 -08002034 for (i = 1; *nent < maxnent; ++i) {
Nitin A Kamble0853d2c2008-11-05 15:37:36 -08002035 level_type = entry[i - 1].ecx & 0xff00;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002036 if (!level_type)
2037 break;
Harvey Harrison14af3f32008-02-19 10:25:50 -08002038 do_cpuid_1_ent(&entry[i], function, i);
2039 entry[i].flags |=
Dan Kenigsberg07716712007-11-21 17:10:04 +02002040 KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2041 ++*nent;
2042 }
2043 break;
2044 }
Dexuan Cui2acf9232010-06-10 11:27:12 +08002045 case 0xd: {
2046 int i;
2047
2048 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2049 for (i = 1; *nent < maxnent; ++i) {
2050 if (entry[i - 1].eax == 0 && i != 2)
2051 break;
2052 do_cpuid_1_ent(&entry[i], function, i);
2053 entry[i].flags |=
2054 KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
2055 ++*nent;
2056 }
2057 break;
2058 }
Glauber Costa84478c82010-05-11 12:17:43 -04002059 case KVM_CPUID_SIGNATURE: {
2060 char signature[12] = "KVMKVMKVM\0\0";
2061 u32 *sigptr = (u32 *)signature;
2062 entry->eax = 0;
2063 entry->ebx = sigptr[0];
2064 entry->ecx = sigptr[1];
2065 entry->edx = sigptr[2];
2066 break;
2067 }
2068 case KVM_CPUID_FEATURES:
2069 entry->eax = (1 << KVM_FEATURE_CLOCKSOURCE) |
2070 (1 << KVM_FEATURE_NOP_IO_DELAY) |
Glauber Costa371bcf62010-05-11 12:17:46 -04002071 (1 << KVM_FEATURE_CLOCKSOURCE2) |
2072 (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT);
Glauber Costa84478c82010-05-11 12:17:43 -04002073 entry->ebx = 0;
2074 entry->ecx = 0;
2075 entry->edx = 0;
2076 break;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002077 case 0x80000000:
2078 entry->eax = min(entry->eax, 0x8000001a);
2079 break;
2080 case 0x80000001:
2081 entry->edx &= kvm_supported_word1_x86_features;
2082 entry->ecx &= kvm_supported_word6_x86_features;
2083 break;
2084 }
Joerg Roedeld4330ef2010-04-22 12:33:11 +02002085
2086 kvm_x86_ops->set_supported_cpuid(function, entry);
2087
Dan Kenigsberg07716712007-11-21 17:10:04 +02002088 put_cpu();
2089}
2090
Avi Kivity7faa4ee2009-05-10 13:55:35 +03002091#undef F
2092
Avi Kivity674eea02008-02-11 18:37:23 +02002093static int kvm_dev_ioctl_get_supported_cpuid(struct kvm_cpuid2 *cpuid,
Amit Shah19355472009-01-14 16:56:00 +00002094 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02002095{
2096 struct kvm_cpuid_entry2 *cpuid_entries;
2097 int limit, nent = 0, r = -E2BIG;
2098 u32 func;
2099
2100 if (cpuid->nent < 1)
2101 goto out;
Avi Kivity6a544352009-10-04 16:45:13 +02002102 if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
2103 cpuid->nent = KVM_MAX_CPUID_ENTRIES;
Dan Kenigsberg07716712007-11-21 17:10:04 +02002104 r = -ENOMEM;
2105 cpuid_entries = vmalloc(sizeof(struct kvm_cpuid_entry2) * cpuid->nent);
2106 if (!cpuid_entries)
2107 goto out;
2108
2109 do_cpuid_ent(&cpuid_entries[0], 0, 0, &nent, cpuid->nent);
2110 limit = cpuid_entries[0].eax;
2111 for (func = 1; func <= limit && nent < cpuid->nent; ++func)
2112 do_cpuid_ent(&cpuid_entries[nent], func, 0,
Amit Shah19355472009-01-14 16:56:00 +00002113 &nent, cpuid->nent);
Dan Kenigsberg07716712007-11-21 17:10:04 +02002114 r = -E2BIG;
2115 if (nent >= cpuid->nent)
2116 goto out_free;
2117
2118 do_cpuid_ent(&cpuid_entries[nent], 0x80000000, 0, &nent, cpuid->nent);
2119 limit = cpuid_entries[nent - 1].eax;
2120 for (func = 0x80000001; func <= limit && nent < cpuid->nent; ++func)
2121 do_cpuid_ent(&cpuid_entries[nent], func, 0,
Amit Shah19355472009-01-14 16:56:00 +00002122 &nent, cpuid->nent);
Glauber Costa84478c82010-05-11 12:17:43 -04002123
2124
2125
2126 r = -E2BIG;
2127 if (nent >= cpuid->nent)
2128 goto out_free;
2129
2130 do_cpuid_ent(&cpuid_entries[nent], KVM_CPUID_SIGNATURE, 0, &nent,
2131 cpuid->nent);
2132
2133 r = -E2BIG;
2134 if (nent >= cpuid->nent)
2135 goto out_free;
2136
2137 do_cpuid_ent(&cpuid_entries[nent], KVM_CPUID_FEATURES, 0, &nent,
2138 cpuid->nent);
2139
Mark McLoughlincb007642009-05-12 12:36:44 +01002140 r = -E2BIG;
2141 if (nent >= cpuid->nent)
2142 goto out_free;
2143
Dan Kenigsberg07716712007-11-21 17:10:04 +02002144 r = -EFAULT;
2145 if (copy_to_user(entries, cpuid_entries,
Amit Shah19355472009-01-14 16:56:00 +00002146 nent * sizeof(struct kvm_cpuid_entry2)))
Dan Kenigsberg07716712007-11-21 17:10:04 +02002147 goto out_free;
2148 cpuid->nent = nent;
2149 r = 0;
2150
2151out_free:
2152 vfree(cpuid_entries);
2153out:
2154 return r;
2155}
2156
Carsten Otte313a3dc2007-10-11 19:16:52 +02002157static int kvm_vcpu_ioctl_get_lapic(struct kvm_vcpu *vcpu,
2158 struct kvm_lapic_state *s)
2159{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002160 memcpy(s->regs, vcpu->arch.apic->regs, sizeof *s);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002161
2162 return 0;
2163}
2164
2165static int kvm_vcpu_ioctl_set_lapic(struct kvm_vcpu *vcpu,
2166 struct kvm_lapic_state *s)
2167{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002168 memcpy(vcpu->arch.apic->regs, s->regs, sizeof *s);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002169 kvm_apic_post_state_restore(vcpu);
Gleb Natapovcb142eb2009-08-09 15:17:40 +03002170 update_cr8_intercept(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002171
2172 return 0;
2173}
2174
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08002175static int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu,
2176 struct kvm_interrupt *irq)
2177{
2178 if (irq->irq < 0 || irq->irq >= 256)
2179 return -EINVAL;
2180 if (irqchip_in_kernel(vcpu->kvm))
2181 return -ENXIO;
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08002182
Gleb Natapov66fd3f72009-05-11 13:35:50 +03002183 kvm_queue_interrupt(vcpu, irq->irq, false);
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08002184
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08002185 return 0;
2186}
2187
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02002188static int kvm_vcpu_ioctl_nmi(struct kvm_vcpu *vcpu)
2189{
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02002190 kvm_inject_nmi(vcpu);
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02002191
2192 return 0;
2193}
2194
Avi Kivityb209749f2007-10-22 16:50:39 +02002195static int vcpu_ioctl_tpr_access_reporting(struct kvm_vcpu *vcpu,
2196 struct kvm_tpr_access_ctl *tac)
2197{
2198 if (tac->flags)
2199 return -EINVAL;
2200 vcpu->arch.tpr_access_reporting = !!tac->enabled;
2201 return 0;
2202}
2203
Huang Ying890ca9a2009-05-11 16:48:15 +08002204static int kvm_vcpu_ioctl_x86_setup_mce(struct kvm_vcpu *vcpu,
2205 u64 mcg_cap)
2206{
2207 int r;
2208 unsigned bank_num = mcg_cap & 0xff, bank;
2209
2210 r = -EINVAL;
Jan Kiszkaa9e38c3e2009-10-23 09:37:00 +02002211 if (!bank_num || bank_num >= KVM_MAX_MCE_BANKS)
Huang Ying890ca9a2009-05-11 16:48:15 +08002212 goto out;
2213 if (mcg_cap & ~(KVM_MCE_CAP_SUPPORTED | 0xff | 0xff0000))
2214 goto out;
2215 r = 0;
2216 vcpu->arch.mcg_cap = mcg_cap;
2217 /* Init IA32_MCG_CTL to all 1s */
2218 if (mcg_cap & MCG_CTL_P)
2219 vcpu->arch.mcg_ctl = ~(u64)0;
2220 /* Init IA32_MCi_CTL to all 1s */
2221 for (bank = 0; bank < bank_num; bank++)
2222 vcpu->arch.mce_banks[bank*4] = ~(u64)0;
2223out:
2224 return r;
2225}
2226
2227static int kvm_vcpu_ioctl_x86_set_mce(struct kvm_vcpu *vcpu,
2228 struct kvm_x86_mce *mce)
2229{
2230 u64 mcg_cap = vcpu->arch.mcg_cap;
2231 unsigned bank_num = mcg_cap & 0xff;
2232 u64 *banks = vcpu->arch.mce_banks;
2233
2234 if (mce->bank >= bank_num || !(mce->status & MCI_STATUS_VAL))
2235 return -EINVAL;
2236 /*
2237 * if IA32_MCG_CTL is not all 1s, the uncorrected error
2238 * reporting is disabled
2239 */
2240 if ((mce->status & MCI_STATUS_UC) && (mcg_cap & MCG_CTL_P) &&
2241 vcpu->arch.mcg_ctl != ~(u64)0)
2242 return 0;
2243 banks += 4 * mce->bank;
2244 /*
2245 * if IA32_MCi_CTL is not all 1s, the uncorrected error
2246 * reporting is disabled for the bank
2247 */
2248 if ((mce->status & MCI_STATUS_UC) && banks[0] != ~(u64)0)
2249 return 0;
2250 if (mce->status & MCI_STATUS_UC) {
2251 if ((vcpu->arch.mcg_status & MCG_STATUS_MCIP) ||
Avi Kivityfc78f512009-12-07 12:16:48 +02002252 !kvm_read_cr4_bits(vcpu, X86_CR4_MCE)) {
Huang Ying890ca9a2009-05-11 16:48:15 +08002253 printk(KERN_DEBUG "kvm: set_mce: "
2254 "injects mce exception while "
2255 "previous one is in progress!\n");
2256 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
2257 return 0;
2258 }
2259 if (banks[1] & MCI_STATUS_VAL)
2260 mce->status |= MCI_STATUS_OVER;
2261 banks[2] = mce->addr;
2262 banks[3] = mce->misc;
2263 vcpu->arch.mcg_status = mce->mcg_status;
2264 banks[1] = mce->status;
2265 kvm_queue_exception(vcpu, MC_VECTOR);
2266 } else if (!(banks[1] & MCI_STATUS_VAL)
2267 || !(banks[1] & MCI_STATUS_UC)) {
2268 if (banks[1] & MCI_STATUS_VAL)
2269 mce->status |= MCI_STATUS_OVER;
2270 banks[2] = mce->addr;
2271 banks[3] = mce->misc;
2272 banks[1] = mce->status;
2273 } else
2274 banks[1] |= MCI_STATUS_OVER;
2275 return 0;
2276}
2277
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002278static void kvm_vcpu_ioctl_x86_get_vcpu_events(struct kvm_vcpu *vcpu,
2279 struct kvm_vcpu_events *events)
2280{
Jan Kiszka03b82a32010-02-15 10:45:41 +01002281 events->exception.injected =
2282 vcpu->arch.exception.pending &&
2283 !kvm_exception_is_soft(vcpu->arch.exception.nr);
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002284 events->exception.nr = vcpu->arch.exception.nr;
2285 events->exception.has_error_code = vcpu->arch.exception.has_error_code;
2286 events->exception.error_code = vcpu->arch.exception.error_code;
2287
Jan Kiszka03b82a32010-02-15 10:45:41 +01002288 events->interrupt.injected =
2289 vcpu->arch.interrupt.pending && !vcpu->arch.interrupt.soft;
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002290 events->interrupt.nr = vcpu->arch.interrupt.nr;
Jan Kiszka03b82a32010-02-15 10:45:41 +01002291 events->interrupt.soft = 0;
Jan Kiszka48005f62010-02-19 19:38:07 +01002292 events->interrupt.shadow =
2293 kvm_x86_ops->get_interrupt_shadow(vcpu,
2294 KVM_X86_SHADOW_INT_MOV_SS | KVM_X86_SHADOW_INT_STI);
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002295
2296 events->nmi.injected = vcpu->arch.nmi_injected;
2297 events->nmi.pending = vcpu->arch.nmi_pending;
2298 events->nmi.masked = kvm_x86_ops->get_nmi_mask(vcpu);
2299
2300 events->sipi_vector = vcpu->arch.sipi_vector;
2301
Jan Kiszkadab4b912009-12-06 18:24:15 +01002302 events->flags = (KVM_VCPUEVENT_VALID_NMI_PENDING
Jan Kiszka48005f62010-02-19 19:38:07 +01002303 | KVM_VCPUEVENT_VALID_SIPI_VECTOR
2304 | KVM_VCPUEVENT_VALID_SHADOW);
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002305}
2306
2307static int kvm_vcpu_ioctl_x86_set_vcpu_events(struct kvm_vcpu *vcpu,
2308 struct kvm_vcpu_events *events)
2309{
Jan Kiszkadab4b912009-12-06 18:24:15 +01002310 if (events->flags & ~(KVM_VCPUEVENT_VALID_NMI_PENDING
Jan Kiszka48005f62010-02-19 19:38:07 +01002311 | KVM_VCPUEVENT_VALID_SIPI_VECTOR
2312 | KVM_VCPUEVENT_VALID_SHADOW))
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002313 return -EINVAL;
2314
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002315 vcpu->arch.exception.pending = events->exception.injected;
2316 vcpu->arch.exception.nr = events->exception.nr;
2317 vcpu->arch.exception.has_error_code = events->exception.has_error_code;
2318 vcpu->arch.exception.error_code = events->exception.error_code;
2319
2320 vcpu->arch.interrupt.pending = events->interrupt.injected;
2321 vcpu->arch.interrupt.nr = events->interrupt.nr;
2322 vcpu->arch.interrupt.soft = events->interrupt.soft;
2323 if (vcpu->arch.interrupt.pending && irqchip_in_kernel(vcpu->kvm))
2324 kvm_pic_clear_isr_ack(vcpu->kvm);
Jan Kiszka48005f62010-02-19 19:38:07 +01002325 if (events->flags & KVM_VCPUEVENT_VALID_SHADOW)
2326 kvm_x86_ops->set_interrupt_shadow(vcpu,
2327 events->interrupt.shadow);
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002328
2329 vcpu->arch.nmi_injected = events->nmi.injected;
Jan Kiszkadab4b912009-12-06 18:24:15 +01002330 if (events->flags & KVM_VCPUEVENT_VALID_NMI_PENDING)
2331 vcpu->arch.nmi_pending = events->nmi.pending;
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002332 kvm_x86_ops->set_nmi_mask(vcpu, events->nmi.masked);
2333
Jan Kiszkadab4b912009-12-06 18:24:15 +01002334 if (events->flags & KVM_VCPUEVENT_VALID_SIPI_VECTOR)
2335 vcpu->arch.sipi_vector = events->sipi_vector;
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002336
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002337 return 0;
2338}
2339
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002340static void kvm_vcpu_ioctl_x86_get_debugregs(struct kvm_vcpu *vcpu,
2341 struct kvm_debugregs *dbgregs)
2342{
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002343 memcpy(dbgregs->db, vcpu->arch.db, sizeof(vcpu->arch.db));
2344 dbgregs->dr6 = vcpu->arch.dr6;
2345 dbgregs->dr7 = vcpu->arch.dr7;
2346 dbgregs->flags = 0;
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002347}
2348
2349static int kvm_vcpu_ioctl_x86_set_debugregs(struct kvm_vcpu *vcpu,
2350 struct kvm_debugregs *dbgregs)
2351{
2352 if (dbgregs->flags)
2353 return -EINVAL;
2354
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002355 memcpy(vcpu->arch.db, dbgregs->db, sizeof(vcpu->arch.db));
2356 vcpu->arch.dr6 = dbgregs->dr6;
2357 vcpu->arch.dr7 = dbgregs->dr7;
2358
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002359 return 0;
2360}
2361
Sheng Yang2d5b5a62010-06-13 17:29:39 +08002362static void kvm_vcpu_ioctl_x86_get_xsave(struct kvm_vcpu *vcpu,
2363 struct kvm_xsave *guest_xsave)
2364{
2365 if (cpu_has_xsave)
2366 memcpy(guest_xsave->region,
2367 &vcpu->arch.guest_fpu.state->xsave,
2368 sizeof(struct xsave_struct));
2369 else {
2370 memcpy(guest_xsave->region,
2371 &vcpu->arch.guest_fpu.state->fxsave,
2372 sizeof(struct i387_fxsave_struct));
2373 *(u64 *)&guest_xsave->region[XSAVE_HDR_OFFSET / sizeof(u32)] =
2374 XSTATE_FPSSE;
2375 }
2376}
2377
2378static int kvm_vcpu_ioctl_x86_set_xsave(struct kvm_vcpu *vcpu,
2379 struct kvm_xsave *guest_xsave)
2380{
2381 u64 xstate_bv =
2382 *(u64 *)&guest_xsave->region[XSAVE_HDR_OFFSET / sizeof(u32)];
2383
2384 if (cpu_has_xsave)
2385 memcpy(&vcpu->arch.guest_fpu.state->xsave,
2386 guest_xsave->region, sizeof(struct xsave_struct));
2387 else {
2388 if (xstate_bv & ~XSTATE_FPSSE)
2389 return -EINVAL;
2390 memcpy(&vcpu->arch.guest_fpu.state->fxsave,
2391 guest_xsave->region, sizeof(struct i387_fxsave_struct));
2392 }
2393 return 0;
2394}
2395
2396static void kvm_vcpu_ioctl_x86_get_xcrs(struct kvm_vcpu *vcpu,
2397 struct kvm_xcrs *guest_xcrs)
2398{
2399 if (!cpu_has_xsave) {
2400 guest_xcrs->nr_xcrs = 0;
2401 return;
2402 }
2403
2404 guest_xcrs->nr_xcrs = 1;
2405 guest_xcrs->flags = 0;
2406 guest_xcrs->xcrs[0].xcr = XCR_XFEATURE_ENABLED_MASK;
2407 guest_xcrs->xcrs[0].value = vcpu->arch.xcr0;
2408}
2409
2410static int kvm_vcpu_ioctl_x86_set_xcrs(struct kvm_vcpu *vcpu,
2411 struct kvm_xcrs *guest_xcrs)
2412{
2413 int i, r = 0;
2414
2415 if (!cpu_has_xsave)
2416 return -EINVAL;
2417
2418 if (guest_xcrs->nr_xcrs > KVM_MAX_XCRS || guest_xcrs->flags)
2419 return -EINVAL;
2420
2421 for (i = 0; i < guest_xcrs->nr_xcrs; i++)
2422 /* Only support XCR0 currently */
2423 if (guest_xcrs->xcrs[0].xcr == XCR_XFEATURE_ENABLED_MASK) {
2424 r = __kvm_set_xcr(vcpu, XCR_XFEATURE_ENABLED_MASK,
2425 guest_xcrs->xcrs[0].value);
2426 break;
2427 }
2428 if (r)
2429 r = -EINVAL;
2430 return r;
2431}
2432
Carsten Otte313a3dc2007-10-11 19:16:52 +02002433long kvm_arch_vcpu_ioctl(struct file *filp,
2434 unsigned int ioctl, unsigned long arg)
2435{
2436 struct kvm_vcpu *vcpu = filp->private_data;
2437 void __user *argp = (void __user *)arg;
2438 int r;
Dave Hansenb772ff32008-08-11 10:01:47 -07002439 struct kvm_lapic_state *lapic = NULL;
Carsten Otte313a3dc2007-10-11 19:16:52 +02002440
2441 switch (ioctl) {
2442 case KVM_GET_LAPIC: {
Marcelo Tosatti2204ae32009-10-29 13:44:16 -02002443 r = -EINVAL;
2444 if (!vcpu->arch.apic)
2445 goto out;
Dave Hansenb772ff32008-08-11 10:01:47 -07002446 lapic = kzalloc(sizeof(struct kvm_lapic_state), GFP_KERNEL);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002447
Dave Hansenb772ff32008-08-11 10:01:47 -07002448 r = -ENOMEM;
2449 if (!lapic)
2450 goto out;
2451 r = kvm_vcpu_ioctl_get_lapic(vcpu, lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002452 if (r)
2453 goto out;
2454 r = -EFAULT;
Dave Hansenb772ff32008-08-11 10:01:47 -07002455 if (copy_to_user(argp, lapic, sizeof(struct kvm_lapic_state)))
Carsten Otte313a3dc2007-10-11 19:16:52 +02002456 goto out;
2457 r = 0;
2458 break;
2459 }
2460 case KVM_SET_LAPIC: {
Marcelo Tosatti2204ae32009-10-29 13:44:16 -02002461 r = -EINVAL;
2462 if (!vcpu->arch.apic)
2463 goto out;
Dave Hansenb772ff32008-08-11 10:01:47 -07002464 lapic = kmalloc(sizeof(struct kvm_lapic_state), GFP_KERNEL);
2465 r = -ENOMEM;
2466 if (!lapic)
Carsten Otte313a3dc2007-10-11 19:16:52 +02002467 goto out;
Dave Hansenb772ff32008-08-11 10:01:47 -07002468 r = -EFAULT;
2469 if (copy_from_user(lapic, argp, sizeof(struct kvm_lapic_state)))
2470 goto out;
2471 r = kvm_vcpu_ioctl_set_lapic(vcpu, lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002472 if (r)
2473 goto out;
2474 r = 0;
2475 break;
2476 }
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08002477 case KVM_INTERRUPT: {
2478 struct kvm_interrupt irq;
2479
2480 r = -EFAULT;
2481 if (copy_from_user(&irq, argp, sizeof irq))
2482 goto out;
2483 r = kvm_vcpu_ioctl_interrupt(vcpu, &irq);
2484 if (r)
2485 goto out;
2486 r = 0;
2487 break;
2488 }
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02002489 case KVM_NMI: {
2490 r = kvm_vcpu_ioctl_nmi(vcpu);
2491 if (r)
2492 goto out;
2493 r = 0;
2494 break;
2495 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02002496 case KVM_SET_CPUID: {
2497 struct kvm_cpuid __user *cpuid_arg = argp;
2498 struct kvm_cpuid cpuid;
2499
2500 r = -EFAULT;
2501 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
2502 goto out;
2503 r = kvm_vcpu_ioctl_set_cpuid(vcpu, &cpuid, cpuid_arg->entries);
2504 if (r)
2505 goto out;
2506 break;
2507 }
Dan Kenigsberg07716712007-11-21 17:10:04 +02002508 case KVM_SET_CPUID2: {
2509 struct kvm_cpuid2 __user *cpuid_arg = argp;
2510 struct kvm_cpuid2 cpuid;
2511
2512 r = -EFAULT;
2513 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
2514 goto out;
2515 r = kvm_vcpu_ioctl_set_cpuid2(vcpu, &cpuid,
Amit Shah19355472009-01-14 16:56:00 +00002516 cpuid_arg->entries);
Dan Kenigsberg07716712007-11-21 17:10:04 +02002517 if (r)
2518 goto out;
2519 break;
2520 }
2521 case KVM_GET_CPUID2: {
2522 struct kvm_cpuid2 __user *cpuid_arg = argp;
2523 struct kvm_cpuid2 cpuid;
2524
2525 r = -EFAULT;
2526 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
2527 goto out;
2528 r = kvm_vcpu_ioctl_get_cpuid2(vcpu, &cpuid,
Amit Shah19355472009-01-14 16:56:00 +00002529 cpuid_arg->entries);
Dan Kenigsberg07716712007-11-21 17:10:04 +02002530 if (r)
2531 goto out;
2532 r = -EFAULT;
2533 if (copy_to_user(cpuid_arg, &cpuid, sizeof cpuid))
2534 goto out;
2535 r = 0;
2536 break;
2537 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02002538 case KVM_GET_MSRS:
2539 r = msr_io(vcpu, argp, kvm_get_msr, 1);
2540 break;
2541 case KVM_SET_MSRS:
2542 r = msr_io(vcpu, argp, do_set_msr, 0);
2543 break;
Avi Kivityb209749f2007-10-22 16:50:39 +02002544 case KVM_TPR_ACCESS_REPORTING: {
2545 struct kvm_tpr_access_ctl tac;
2546
2547 r = -EFAULT;
2548 if (copy_from_user(&tac, argp, sizeof tac))
2549 goto out;
2550 r = vcpu_ioctl_tpr_access_reporting(vcpu, &tac);
2551 if (r)
2552 goto out;
2553 r = -EFAULT;
2554 if (copy_to_user(argp, &tac, sizeof tac))
2555 goto out;
2556 r = 0;
2557 break;
2558 };
Avi Kivityb93463a2007-10-25 16:52:32 +02002559 case KVM_SET_VAPIC_ADDR: {
2560 struct kvm_vapic_addr va;
2561
2562 r = -EINVAL;
2563 if (!irqchip_in_kernel(vcpu->kvm))
2564 goto out;
2565 r = -EFAULT;
2566 if (copy_from_user(&va, argp, sizeof va))
2567 goto out;
2568 r = 0;
2569 kvm_lapic_set_vapic_addr(vcpu, va.vapic_addr);
2570 break;
2571 }
Huang Ying890ca9a2009-05-11 16:48:15 +08002572 case KVM_X86_SETUP_MCE: {
2573 u64 mcg_cap;
2574
2575 r = -EFAULT;
2576 if (copy_from_user(&mcg_cap, argp, sizeof mcg_cap))
2577 goto out;
2578 r = kvm_vcpu_ioctl_x86_setup_mce(vcpu, mcg_cap);
2579 break;
2580 }
2581 case KVM_X86_SET_MCE: {
2582 struct kvm_x86_mce mce;
2583
2584 r = -EFAULT;
2585 if (copy_from_user(&mce, argp, sizeof mce))
2586 goto out;
2587 r = kvm_vcpu_ioctl_x86_set_mce(vcpu, &mce);
2588 break;
2589 }
Jan Kiszka3cfc3092009-11-12 01:04:25 +01002590 case KVM_GET_VCPU_EVENTS: {
2591 struct kvm_vcpu_events events;
2592
2593 kvm_vcpu_ioctl_x86_get_vcpu_events(vcpu, &events);
2594
2595 r = -EFAULT;
2596 if (copy_to_user(argp, &events, sizeof(struct kvm_vcpu_events)))
2597 break;
2598 r = 0;
2599 break;
2600 }
2601 case KVM_SET_VCPU_EVENTS: {
2602 struct kvm_vcpu_events events;
2603
2604 r = -EFAULT;
2605 if (copy_from_user(&events, argp, sizeof(struct kvm_vcpu_events)))
2606 break;
2607
2608 r = kvm_vcpu_ioctl_x86_set_vcpu_events(vcpu, &events);
2609 break;
2610 }
Jan Kiszkaa1efbe72010-02-15 10:45:43 +01002611 case KVM_GET_DEBUGREGS: {
2612 struct kvm_debugregs dbgregs;
2613
2614 kvm_vcpu_ioctl_x86_get_debugregs(vcpu, &dbgregs);
2615
2616 r = -EFAULT;
2617 if (copy_to_user(argp, &dbgregs,
2618 sizeof(struct kvm_debugregs)))
2619 break;
2620 r = 0;
2621 break;
2622 }
2623 case KVM_SET_DEBUGREGS: {
2624 struct kvm_debugregs dbgregs;
2625
2626 r = -EFAULT;
2627 if (copy_from_user(&dbgregs, argp,
2628 sizeof(struct kvm_debugregs)))
2629 break;
2630
2631 r = kvm_vcpu_ioctl_x86_set_debugregs(vcpu, &dbgregs);
2632 break;
2633 }
Sheng Yang2d5b5a62010-06-13 17:29:39 +08002634 case KVM_GET_XSAVE: {
2635 struct kvm_xsave *xsave;
2636
2637 xsave = kzalloc(sizeof(struct kvm_xsave), GFP_KERNEL);
2638 r = -ENOMEM;
2639 if (!xsave)
2640 break;
2641
2642 kvm_vcpu_ioctl_x86_get_xsave(vcpu, xsave);
2643
2644 r = -EFAULT;
2645 if (copy_to_user(argp, xsave, sizeof(struct kvm_xsave)))
2646 break;
2647 r = 0;
2648 break;
2649 }
2650 case KVM_SET_XSAVE: {
2651 struct kvm_xsave *xsave;
2652
2653 xsave = kzalloc(sizeof(struct kvm_xsave), GFP_KERNEL);
2654 r = -ENOMEM;
2655 if (!xsave)
2656 break;
2657
2658 r = -EFAULT;
2659 if (copy_from_user(xsave, argp, sizeof(struct kvm_xsave)))
2660 break;
2661
2662 r = kvm_vcpu_ioctl_x86_set_xsave(vcpu, xsave);
2663 break;
2664 }
2665 case KVM_GET_XCRS: {
2666 struct kvm_xcrs *xcrs;
2667
2668 xcrs = kzalloc(sizeof(struct kvm_xcrs), GFP_KERNEL);
2669 r = -ENOMEM;
2670 if (!xcrs)
2671 break;
2672
2673 kvm_vcpu_ioctl_x86_get_xcrs(vcpu, xcrs);
2674
2675 r = -EFAULT;
2676 if (copy_to_user(argp, xcrs,
2677 sizeof(struct kvm_xcrs)))
2678 break;
2679 r = 0;
2680 break;
2681 }
2682 case KVM_SET_XCRS: {
2683 struct kvm_xcrs *xcrs;
2684
2685 xcrs = kzalloc(sizeof(struct kvm_xcrs), GFP_KERNEL);
2686 r = -ENOMEM;
2687 if (!xcrs)
2688 break;
2689
2690 r = -EFAULT;
2691 if (copy_from_user(xcrs, argp,
2692 sizeof(struct kvm_xcrs)))
2693 break;
2694
2695 r = kvm_vcpu_ioctl_x86_set_xcrs(vcpu, xcrs);
2696 break;
2697 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02002698 default:
2699 r = -EINVAL;
2700 }
2701out:
Wei Yongjun7a6ce842009-03-31 16:47:44 +08002702 kfree(lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02002703 return r;
2704}
2705
Carsten Otte1fe779f2007-10-29 16:08:35 +01002706static int kvm_vm_ioctl_set_tss_addr(struct kvm *kvm, unsigned long addr)
2707{
2708 int ret;
2709
2710 if (addr > (unsigned int)(-3 * PAGE_SIZE))
2711 return -1;
2712 ret = kvm_x86_ops->set_tss_addr(kvm, addr);
2713 return ret;
2714}
2715
Sheng Yangb927a3c2009-07-21 10:42:48 +08002716static int kvm_vm_ioctl_set_identity_map_addr(struct kvm *kvm,
2717 u64 ident_addr)
2718{
2719 kvm->arch.ept_identity_map_addr = ident_addr;
2720 return 0;
2721}
2722
Carsten Otte1fe779f2007-10-29 16:08:35 +01002723static int kvm_vm_ioctl_set_nr_mmu_pages(struct kvm *kvm,
2724 u32 kvm_nr_mmu_pages)
2725{
2726 if (kvm_nr_mmu_pages < KVM_MIN_ALLOC_MMU_PAGES)
2727 return -EINVAL;
2728
Marcelo Tosatti79fac952009-12-23 14:35:26 -02002729 mutex_lock(&kvm->slots_lock);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03002730 spin_lock(&kvm->mmu_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002731
2732 kvm_mmu_change_mmu_pages(kvm, kvm_nr_mmu_pages);
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08002733 kvm->arch.n_requested_mmu_pages = kvm_nr_mmu_pages;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002734
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03002735 spin_unlock(&kvm->mmu_lock);
Marcelo Tosatti79fac952009-12-23 14:35:26 -02002736 mutex_unlock(&kvm->slots_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002737 return 0;
2738}
2739
2740static int kvm_vm_ioctl_get_nr_mmu_pages(struct kvm *kvm)
2741{
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08002742 return kvm->arch.n_alloc_mmu_pages;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002743}
2744
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002745gfn_t unalias_gfn_instantiation(struct kvm *kvm, gfn_t gfn)
2746{
2747 int i;
2748 struct kvm_mem_alias *alias;
2749 struct kvm_mem_aliases *aliases;
2750
Lai Jiangshan90d83dc2010-04-19 17:41:23 +08002751 aliases = kvm_aliases(kvm);
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002752
2753 for (i = 0; i < aliases->naliases; ++i) {
2754 alias = &aliases->aliases[i];
2755 if (alias->flags & KVM_ALIAS_INVALID)
2756 continue;
2757 if (gfn >= alias->base_gfn
2758 && gfn < alias->base_gfn + alias->npages)
2759 return alias->target_gfn + gfn - alias->base_gfn;
2760 }
2761 return gfn;
2762}
2763
Zhang Xiantaoe9f85cd2007-11-22 11:20:33 +08002764gfn_t unalias_gfn(struct kvm *kvm, gfn_t gfn)
2765{
2766 int i;
2767 struct kvm_mem_alias *alias;
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002768 struct kvm_mem_aliases *aliases;
2769
Lai Jiangshan90d83dc2010-04-19 17:41:23 +08002770 aliases = kvm_aliases(kvm);
Zhang Xiantaoe9f85cd2007-11-22 11:20:33 +08002771
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02002772 for (i = 0; i < aliases->naliases; ++i) {
2773 alias = &aliases->aliases[i];
Zhang Xiantaoe9f85cd2007-11-22 11:20:33 +08002774 if (gfn >= alias->base_gfn
2775 && gfn < alias->base_gfn + alias->npages)
2776 return alias->target_gfn + gfn - alias->base_gfn;
2777 }
2778 return gfn;
2779}
2780
Carsten Otte1fe779f2007-10-29 16:08:35 +01002781/*
2782 * Set a new alias region. Aliases map a portion of physical memory into
2783 * another portion. This is useful for memory windows, for example the PC
2784 * VGA region.
2785 */
2786static int kvm_vm_ioctl_set_memory_alias(struct kvm *kvm,
2787 struct kvm_memory_alias *alias)
2788{
2789 int r, n;
2790 struct kvm_mem_alias *p;
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002791 struct kvm_mem_aliases *aliases, *old_aliases;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002792
2793 r = -EINVAL;
2794 /* General sanity checks */
2795 if (alias->memory_size & (PAGE_SIZE - 1))
2796 goto out;
2797 if (alias->guest_phys_addr & (PAGE_SIZE - 1))
2798 goto out;
2799 if (alias->slot >= KVM_ALIAS_SLOTS)
2800 goto out;
2801 if (alias->guest_phys_addr + alias->memory_size
2802 < alias->guest_phys_addr)
2803 goto out;
2804 if (alias->target_phys_addr + alias->memory_size
2805 < alias->target_phys_addr)
2806 goto out;
2807
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002808 r = -ENOMEM;
2809 aliases = kzalloc(sizeof(struct kvm_mem_aliases), GFP_KERNEL);
2810 if (!aliases)
2811 goto out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002812
Marcelo Tosatti79fac952009-12-23 14:35:26 -02002813 mutex_lock(&kvm->slots_lock);
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002814
2815 /* invalidate any gfn reference in case of deletion/shrinking */
2816 memcpy(aliases, kvm->arch.aliases, sizeof(struct kvm_mem_aliases));
2817 aliases->aliases[alias->slot].flags |= KVM_ALIAS_INVALID;
2818 old_aliases = kvm->arch.aliases;
2819 rcu_assign_pointer(kvm->arch.aliases, aliases);
2820 synchronize_srcu_expedited(&kvm->srcu);
2821 kvm_mmu_zap_all(kvm);
2822 kfree(old_aliases);
2823
2824 r = -ENOMEM;
2825 aliases = kzalloc(sizeof(struct kvm_mem_aliases), GFP_KERNEL);
2826 if (!aliases)
2827 goto out_unlock;
2828
2829 memcpy(aliases, kvm->arch.aliases, sizeof(struct kvm_mem_aliases));
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02002830
2831 p = &aliases->aliases[alias->slot];
Carsten Otte1fe779f2007-10-29 16:08:35 +01002832 p->base_gfn = alias->guest_phys_addr >> PAGE_SHIFT;
2833 p->npages = alias->memory_size >> PAGE_SHIFT;
2834 p->target_gfn = alias->target_phys_addr >> PAGE_SHIFT;
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002835 p->flags &= ~(KVM_ALIAS_INVALID);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002836
2837 for (n = KVM_ALIAS_SLOTS; n > 0; --n)
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02002838 if (aliases->aliases[n - 1].npages)
Carsten Otte1fe779f2007-10-29 16:08:35 +01002839 break;
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02002840 aliases->naliases = n;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002841
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002842 old_aliases = kvm->arch.aliases;
2843 rcu_assign_pointer(kvm->arch.aliases, aliases);
2844 synchronize_srcu_expedited(&kvm->srcu);
2845 kfree(old_aliases);
2846 r = 0;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002847
Marcelo Tosattia983fb22009-12-23 14:35:23 -02002848out_unlock:
Marcelo Tosatti79fac952009-12-23 14:35:26 -02002849 mutex_unlock(&kvm->slots_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002850out:
2851 return r;
2852}
2853
2854static int kvm_vm_ioctl_get_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
2855{
2856 int r;
2857
2858 r = 0;
2859 switch (chip->chip_id) {
2860 case KVM_IRQCHIP_PIC_MASTER:
2861 memcpy(&chip->chip.pic,
2862 &pic_irqchip(kvm)->pics[0],
2863 sizeof(struct kvm_pic_state));
2864 break;
2865 case KVM_IRQCHIP_PIC_SLAVE:
2866 memcpy(&chip->chip.pic,
2867 &pic_irqchip(kvm)->pics[1],
2868 sizeof(struct kvm_pic_state));
2869 break;
2870 case KVM_IRQCHIP_IOAPIC:
Gleb Natapoveba02262009-08-24 11:54:25 +03002871 r = kvm_get_ioapic(kvm, &chip->chip.ioapic);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002872 break;
2873 default:
2874 r = -EINVAL;
2875 break;
2876 }
2877 return r;
2878}
2879
2880static int kvm_vm_ioctl_set_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
2881{
2882 int r;
2883
2884 r = 0;
2885 switch (chip->chip_id) {
2886 case KVM_IRQCHIP_PIC_MASTER:
Thomas Gleixnerfa8273e2010-02-17 14:00:41 +00002887 raw_spin_lock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002888 memcpy(&pic_irqchip(kvm)->pics[0],
2889 &chip->chip.pic,
2890 sizeof(struct kvm_pic_state));
Thomas Gleixnerfa8273e2010-02-17 14:00:41 +00002891 raw_spin_unlock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002892 break;
2893 case KVM_IRQCHIP_PIC_SLAVE:
Thomas Gleixnerfa8273e2010-02-17 14:00:41 +00002894 raw_spin_lock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002895 memcpy(&pic_irqchip(kvm)->pics[1],
2896 &chip->chip.pic,
2897 sizeof(struct kvm_pic_state));
Thomas Gleixnerfa8273e2010-02-17 14:00:41 +00002898 raw_spin_unlock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002899 break;
2900 case KVM_IRQCHIP_IOAPIC:
Gleb Natapoveba02262009-08-24 11:54:25 +03002901 r = kvm_set_ioapic(kvm, &chip->chip.ioapic);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002902 break;
2903 default:
2904 r = -EINVAL;
2905 break;
2906 }
2907 kvm_pic_update_irq(pic_irqchip(kvm));
2908 return r;
2909}
2910
Sheng Yange0f63cb2008-03-04 00:50:59 +08002911static int kvm_vm_ioctl_get_pit(struct kvm *kvm, struct kvm_pit_state *ps)
2912{
2913 int r = 0;
2914
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002915 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002916 memcpy(ps, &kvm->arch.vpit->pit_state, sizeof(struct kvm_pit_state));
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002917 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002918 return r;
2919}
2920
2921static int kvm_vm_ioctl_set_pit(struct kvm *kvm, struct kvm_pit_state *ps)
2922{
2923 int r = 0;
2924
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002925 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002926 memcpy(&kvm->arch.vpit->pit_state, ps, sizeof(struct kvm_pit_state));
Beth Kone9f42752009-07-07 11:50:38 -04002927 kvm_pit_load_count(kvm, 0, ps->channels[0].count, 0);
2928 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
2929 return r;
2930}
2931
2932static int kvm_vm_ioctl_get_pit2(struct kvm *kvm, struct kvm_pit_state2 *ps)
2933{
2934 int r = 0;
2935
2936 mutex_lock(&kvm->arch.vpit->pit_state.lock);
2937 memcpy(ps->channels, &kvm->arch.vpit->pit_state.channels,
2938 sizeof(ps->channels));
2939 ps->flags = kvm->arch.vpit->pit_state.flags;
2940 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
2941 return r;
2942}
2943
2944static int kvm_vm_ioctl_set_pit2(struct kvm *kvm, struct kvm_pit_state2 *ps)
2945{
2946 int r = 0, start = 0;
2947 u32 prev_legacy, cur_legacy;
2948 mutex_lock(&kvm->arch.vpit->pit_state.lock);
2949 prev_legacy = kvm->arch.vpit->pit_state.flags & KVM_PIT_FLAGS_HPET_LEGACY;
2950 cur_legacy = ps->flags & KVM_PIT_FLAGS_HPET_LEGACY;
2951 if (!prev_legacy && cur_legacy)
2952 start = 1;
2953 memcpy(&kvm->arch.vpit->pit_state.channels, &ps->channels,
2954 sizeof(kvm->arch.vpit->pit_state.channels));
2955 kvm->arch.vpit->pit_state.flags = ps->flags;
2956 kvm_pit_load_count(kvm, 0, kvm->arch.vpit->pit_state.channels[0].count, start);
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002957 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002958 return r;
2959}
2960
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002961static int kvm_vm_ioctl_reinject(struct kvm *kvm,
2962 struct kvm_reinject_control *control)
2963{
2964 if (!kvm->arch.vpit)
2965 return -ENXIO;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002966 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002967 kvm->arch.vpit->pit_state.pit_timer.reinject = control->pit_reinject;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002968 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002969 return 0;
2970}
2971
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002972/*
2973 * Get (and clear) the dirty memory log for a memory slot.
2974 */
2975int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm,
2976 struct kvm_dirty_log *log)
2977{
Takuya Yoshikawa87bf6e72010-04-12 19:35:35 +09002978 int r, i;
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002979 struct kvm_memory_slot *memslot;
Takuya Yoshikawa87bf6e72010-04-12 19:35:35 +09002980 unsigned long n;
Marcelo Tosattib050b012009-12-23 14:35:22 -02002981 unsigned long is_dirty = 0;
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002982
Marcelo Tosatti79fac952009-12-23 14:35:26 -02002983 mutex_lock(&kvm->slots_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002984
Marcelo Tosattib050b012009-12-23 14:35:22 -02002985 r = -EINVAL;
2986 if (log->slot >= KVM_MEMORY_SLOTS)
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002987 goto out;
2988
Marcelo Tosattib050b012009-12-23 14:35:22 -02002989 memslot = &kvm->memslots->memslots[log->slot];
2990 r = -ENOENT;
2991 if (!memslot->dirty_bitmap)
2992 goto out;
2993
Takuya Yoshikawa87bf6e72010-04-12 19:35:35 +09002994 n = kvm_dirty_bitmap_bytes(memslot);
Marcelo Tosattib050b012009-12-23 14:35:22 -02002995
Marcelo Tosattib050b012009-12-23 14:35:22 -02002996 for (i = 0; !is_dirty && i < n/sizeof(long); i++)
2997 is_dirty = memslot->dirty_bitmap[i];
2998
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002999 /* If nothing is dirty, don't bother messing with page tables. */
3000 if (is_dirty) {
Marcelo Tosattib050b012009-12-23 14:35:22 -02003001 struct kvm_memslots *slots, *old_slots;
Takuya Yoshikawa914ebcc2010-04-28 18:50:36 +09003002 unsigned long *dirty_bitmap;
Marcelo Tosattib050b012009-12-23 14:35:22 -02003003
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03003004 spin_lock(&kvm->mmu_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08003005 kvm_mmu_slot_remove_write_access(kvm, log->slot);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03003006 spin_unlock(&kvm->mmu_lock);
Marcelo Tosattib050b012009-12-23 14:35:22 -02003007
Takuya Yoshikawa914ebcc2010-04-28 18:50:36 +09003008 r = -ENOMEM;
3009 dirty_bitmap = vmalloc(n);
3010 if (!dirty_bitmap)
3011 goto out;
3012 memset(dirty_bitmap, 0, n);
Marcelo Tosattib050b012009-12-23 14:35:22 -02003013
Takuya Yoshikawa914ebcc2010-04-28 18:50:36 +09003014 r = -ENOMEM;
3015 slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
3016 if (!slots) {
3017 vfree(dirty_bitmap);
3018 goto out;
3019 }
Marcelo Tosattib050b012009-12-23 14:35:22 -02003020 memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots));
3021 slots->memslots[log->slot].dirty_bitmap = dirty_bitmap;
3022
3023 old_slots = kvm->memslots;
3024 rcu_assign_pointer(kvm->memslots, slots);
3025 synchronize_srcu_expedited(&kvm->srcu);
3026 dirty_bitmap = old_slots->memslots[log->slot].dirty_bitmap;
3027 kfree(old_slots);
Takuya Yoshikawa914ebcc2010-04-28 18:50:36 +09003028
3029 r = -EFAULT;
3030 if (copy_to_user(log->dirty_bitmap, dirty_bitmap, n)) {
3031 vfree(dirty_bitmap);
3032 goto out;
3033 }
3034 vfree(dirty_bitmap);
3035 } else {
3036 r = -EFAULT;
3037 if (clear_user(log->dirty_bitmap, n))
3038 goto out;
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08003039 }
Marcelo Tosattib050b012009-12-23 14:35:22 -02003040
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08003041 r = 0;
3042out:
Marcelo Tosatti79fac952009-12-23 14:35:26 -02003043 mutex_unlock(&kvm->slots_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08003044 return r;
3045}
3046
Carsten Otte1fe779f2007-10-29 16:08:35 +01003047long kvm_arch_vm_ioctl(struct file *filp,
3048 unsigned int ioctl, unsigned long arg)
3049{
3050 struct kvm *kvm = filp->private_data;
3051 void __user *argp = (void __user *)arg;
Avi Kivity367e1312009-08-26 14:57:07 +03003052 int r = -ENOTTY;
Dave Hansenf0d66272008-08-11 10:01:45 -07003053 /*
3054 * This union makes it completely explicit to gcc-3.x
3055 * that these two variables' stack usage should be
3056 * combined, not added together.
3057 */
3058 union {
3059 struct kvm_pit_state ps;
Beth Kone9f42752009-07-07 11:50:38 -04003060 struct kvm_pit_state2 ps2;
Dave Hansenf0d66272008-08-11 10:01:45 -07003061 struct kvm_memory_alias alias;
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02003062 struct kvm_pit_config pit_config;
Dave Hansenf0d66272008-08-11 10:01:45 -07003063 } u;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003064
3065 switch (ioctl) {
3066 case KVM_SET_TSS_ADDR:
3067 r = kvm_vm_ioctl_set_tss_addr(kvm, arg);
3068 if (r < 0)
3069 goto out;
3070 break;
Sheng Yangb927a3c2009-07-21 10:42:48 +08003071 case KVM_SET_IDENTITY_MAP_ADDR: {
3072 u64 ident_addr;
3073
3074 r = -EFAULT;
3075 if (copy_from_user(&ident_addr, argp, sizeof ident_addr))
3076 goto out;
3077 r = kvm_vm_ioctl_set_identity_map_addr(kvm, ident_addr);
3078 if (r < 0)
3079 goto out;
3080 break;
3081 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01003082 case KVM_SET_MEMORY_REGION: {
3083 struct kvm_memory_region kvm_mem;
3084 struct kvm_userspace_memory_region kvm_userspace_mem;
3085
3086 r = -EFAULT;
3087 if (copy_from_user(&kvm_mem, argp, sizeof kvm_mem))
3088 goto out;
3089 kvm_userspace_mem.slot = kvm_mem.slot;
3090 kvm_userspace_mem.flags = kvm_mem.flags;
3091 kvm_userspace_mem.guest_phys_addr = kvm_mem.guest_phys_addr;
3092 kvm_userspace_mem.memory_size = kvm_mem.memory_size;
3093 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem, 0);
3094 if (r)
3095 goto out;
3096 break;
3097 }
3098 case KVM_SET_NR_MMU_PAGES:
3099 r = kvm_vm_ioctl_set_nr_mmu_pages(kvm, arg);
3100 if (r)
3101 goto out;
3102 break;
3103 case KVM_GET_NR_MMU_PAGES:
3104 r = kvm_vm_ioctl_get_nr_mmu_pages(kvm);
3105 break;
Dave Hansenf0d66272008-08-11 10:01:45 -07003106 case KVM_SET_MEMORY_ALIAS:
Carsten Otte1fe779f2007-10-29 16:08:35 +01003107 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07003108 if (copy_from_user(&u.alias, argp, sizeof(struct kvm_memory_alias)))
Carsten Otte1fe779f2007-10-29 16:08:35 +01003109 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07003110 r = kvm_vm_ioctl_set_memory_alias(kvm, &u.alias);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003111 if (r)
3112 goto out;
3113 break;
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003114 case KVM_CREATE_IRQCHIP: {
3115 struct kvm_pic *vpic;
3116
3117 mutex_lock(&kvm->lock);
3118 r = -EEXIST;
3119 if (kvm->arch.vpic)
3120 goto create_irqchip_unlock;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003121 r = -ENOMEM;
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003122 vpic = kvm_create_pic(kvm);
3123 if (vpic) {
Carsten Otte1fe779f2007-10-29 16:08:35 +01003124 r = kvm_ioapic_init(kvm);
3125 if (r) {
Wei Yongjun72bb2fc2010-02-09 10:33:03 +08003126 kvm_io_bus_unregister_dev(kvm, KVM_PIO_BUS,
3127 &vpic->dev);
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003128 kfree(vpic);
3129 goto create_irqchip_unlock;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003130 }
3131 } else
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003132 goto create_irqchip_unlock;
3133 smp_wmb();
3134 kvm->arch.vpic = vpic;
3135 smp_wmb();
Avi Kivity399ec802008-11-19 13:58:46 +02003136 r = kvm_setup_default_irq_routing(kvm);
3137 if (r) {
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003138 mutex_lock(&kvm->irq_lock);
Wei Yongjun72bb2fc2010-02-09 10:33:03 +08003139 kvm_ioapic_destroy(kvm);
3140 kvm_destroy_pic(kvm);
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003141 mutex_unlock(&kvm->irq_lock);
Avi Kivity399ec802008-11-19 13:58:46 +02003142 }
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003143 create_irqchip_unlock:
3144 mutex_unlock(&kvm->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003145 break;
Marcelo Tosatti3ddea122009-10-29 13:44:15 -02003146 }
Sheng Yang78376992008-01-28 05:10:22 +08003147 case KVM_CREATE_PIT:
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02003148 u.pit_config.flags = KVM_PIT_SPEAKER_DUMMY;
3149 goto create_pit;
3150 case KVM_CREATE_PIT2:
3151 r = -EFAULT;
3152 if (copy_from_user(&u.pit_config, argp,
3153 sizeof(struct kvm_pit_config)))
3154 goto out;
3155 create_pit:
Marcelo Tosatti79fac952009-12-23 14:35:26 -02003156 mutex_lock(&kvm->slots_lock);
Avi Kivity269e05e2009-01-05 15:21:42 +02003157 r = -EEXIST;
3158 if (kvm->arch.vpit)
3159 goto create_pit_unlock;
Sheng Yang78376992008-01-28 05:10:22 +08003160 r = -ENOMEM;
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02003161 kvm->arch.vpit = kvm_create_pit(kvm, u.pit_config.flags);
Sheng Yang78376992008-01-28 05:10:22 +08003162 if (kvm->arch.vpit)
3163 r = 0;
Avi Kivity269e05e2009-01-05 15:21:42 +02003164 create_pit_unlock:
Marcelo Tosatti79fac952009-12-23 14:35:26 -02003165 mutex_unlock(&kvm->slots_lock);
Sheng Yang78376992008-01-28 05:10:22 +08003166 break;
Gleb Natapov49256632009-02-04 17:28:14 +02003167 case KVM_IRQ_LINE_STATUS:
Carsten Otte1fe779f2007-10-29 16:08:35 +01003168 case KVM_IRQ_LINE: {
3169 struct kvm_irq_level irq_event;
3170
3171 r = -EFAULT;
3172 if (copy_from_user(&irq_event, argp, sizeof irq_event))
3173 goto out;
Wei Yongjun160d2f62010-03-12 10:09:45 +08003174 r = -ENXIO;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003175 if (irqchip_in_kernel(kvm)) {
Gleb Natapov49256632009-02-04 17:28:14 +02003176 __s32 status;
Gleb Natapov49256632009-02-04 17:28:14 +02003177 status = kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID,
3178 irq_event.irq, irq_event.level);
Gleb Natapov49256632009-02-04 17:28:14 +02003179 if (ioctl == KVM_IRQ_LINE_STATUS) {
Wei Yongjun160d2f62010-03-12 10:09:45 +08003180 r = -EFAULT;
Gleb Natapov49256632009-02-04 17:28:14 +02003181 irq_event.status = status;
3182 if (copy_to_user(argp, &irq_event,
3183 sizeof irq_event))
3184 goto out;
3185 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01003186 r = 0;
3187 }
3188 break;
3189 }
3190 case KVM_GET_IRQCHIP: {
3191 /* 0: PIC master, 1: PIC slave, 2: IOAPIC */
Dave Hansenf0d66272008-08-11 10:01:45 -07003192 struct kvm_irqchip *chip = kmalloc(sizeof(*chip), GFP_KERNEL);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003193
Dave Hansenf0d66272008-08-11 10:01:45 -07003194 r = -ENOMEM;
3195 if (!chip)
Carsten Otte1fe779f2007-10-29 16:08:35 +01003196 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07003197 r = -EFAULT;
3198 if (copy_from_user(chip, argp, sizeof *chip))
3199 goto get_irqchip_out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003200 r = -ENXIO;
3201 if (!irqchip_in_kernel(kvm))
Dave Hansenf0d66272008-08-11 10:01:45 -07003202 goto get_irqchip_out;
3203 r = kvm_vm_ioctl_get_irqchip(kvm, chip);
3204 if (r)
3205 goto get_irqchip_out;
3206 r = -EFAULT;
3207 if (copy_to_user(argp, chip, sizeof *chip))
3208 goto get_irqchip_out;
3209 r = 0;
3210 get_irqchip_out:
3211 kfree(chip);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003212 if (r)
3213 goto out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003214 break;
3215 }
3216 case KVM_SET_IRQCHIP: {
3217 /* 0: PIC master, 1: PIC slave, 2: IOAPIC */
Dave Hansenf0d66272008-08-11 10:01:45 -07003218 struct kvm_irqchip *chip = kmalloc(sizeof(*chip), GFP_KERNEL);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003219
Dave Hansenf0d66272008-08-11 10:01:45 -07003220 r = -ENOMEM;
3221 if (!chip)
Carsten Otte1fe779f2007-10-29 16:08:35 +01003222 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07003223 r = -EFAULT;
3224 if (copy_from_user(chip, argp, sizeof *chip))
3225 goto set_irqchip_out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003226 r = -ENXIO;
3227 if (!irqchip_in_kernel(kvm))
Dave Hansenf0d66272008-08-11 10:01:45 -07003228 goto set_irqchip_out;
3229 r = kvm_vm_ioctl_set_irqchip(kvm, chip);
3230 if (r)
3231 goto set_irqchip_out;
3232 r = 0;
3233 set_irqchip_out:
3234 kfree(chip);
Carsten Otte1fe779f2007-10-29 16:08:35 +01003235 if (r)
3236 goto out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01003237 break;
3238 }
Sheng Yange0f63cb2008-03-04 00:50:59 +08003239 case KVM_GET_PIT: {
Sheng Yange0f63cb2008-03-04 00:50:59 +08003240 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07003241 if (copy_from_user(&u.ps, argp, sizeof(struct kvm_pit_state)))
Sheng Yange0f63cb2008-03-04 00:50:59 +08003242 goto out;
3243 r = -ENXIO;
3244 if (!kvm->arch.vpit)
3245 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07003246 r = kvm_vm_ioctl_get_pit(kvm, &u.ps);
Sheng Yange0f63cb2008-03-04 00:50:59 +08003247 if (r)
3248 goto out;
3249 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07003250 if (copy_to_user(argp, &u.ps, sizeof(struct kvm_pit_state)))
Sheng Yange0f63cb2008-03-04 00:50:59 +08003251 goto out;
3252 r = 0;
3253 break;
3254 }
3255 case KVM_SET_PIT: {
Sheng Yange0f63cb2008-03-04 00:50:59 +08003256 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07003257 if (copy_from_user(&u.ps, argp, sizeof u.ps))
Sheng Yange0f63cb2008-03-04 00:50:59 +08003258 goto out;
3259 r = -ENXIO;
3260 if (!kvm->arch.vpit)
3261 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07003262 r = kvm_vm_ioctl_set_pit(kvm, &u.ps);
Sheng Yange0f63cb2008-03-04 00:50:59 +08003263 if (r)
3264 goto out;
3265 r = 0;
3266 break;
3267 }
Beth Kone9f42752009-07-07 11:50:38 -04003268 case KVM_GET_PIT2: {
3269 r = -ENXIO;
3270 if (!kvm->arch.vpit)
3271 goto out;
3272 r = kvm_vm_ioctl_get_pit2(kvm, &u.ps2);
3273 if (r)
3274 goto out;
3275 r = -EFAULT;
3276 if (copy_to_user(argp, &u.ps2, sizeof(u.ps2)))
3277 goto out;
3278 r = 0;
3279 break;
3280 }
3281 case KVM_SET_PIT2: {
3282 r = -EFAULT;
3283 if (copy_from_user(&u.ps2, argp, sizeof(u.ps2)))
3284 goto out;
3285 r = -ENXIO;
3286 if (!kvm->arch.vpit)
3287 goto out;
3288 r = kvm_vm_ioctl_set_pit2(kvm, &u.ps2);
3289 if (r)
3290 goto out;
3291 r = 0;
3292 break;
3293 }
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02003294 case KVM_REINJECT_CONTROL: {
3295 struct kvm_reinject_control control;
3296 r = -EFAULT;
3297 if (copy_from_user(&control, argp, sizeof(control)))
3298 goto out;
3299 r = kvm_vm_ioctl_reinject(kvm, &control);
3300 if (r)
3301 goto out;
3302 r = 0;
3303 break;
3304 }
Ed Swierkffde22a2009-10-15 15:21:43 -07003305 case KVM_XEN_HVM_CONFIG: {
3306 r = -EFAULT;
3307 if (copy_from_user(&kvm->arch.xen_hvm_config, argp,
3308 sizeof(struct kvm_xen_hvm_config)))
3309 goto out;
3310 r = -EINVAL;
3311 if (kvm->arch.xen_hvm_config.flags)
3312 goto out;
3313 r = 0;
3314 break;
3315 }
Glauber Costaafbcf7a2009-10-16 15:28:36 -04003316 case KVM_SET_CLOCK: {
3317 struct timespec now;
3318 struct kvm_clock_data user_ns;
3319 u64 now_ns;
3320 s64 delta;
3321
3322 r = -EFAULT;
3323 if (copy_from_user(&user_ns, argp, sizeof(user_ns)))
3324 goto out;
3325
3326 r = -EINVAL;
3327 if (user_ns.flags)
3328 goto out;
3329
3330 r = 0;
3331 ktime_get_ts(&now);
3332 now_ns = timespec_to_ns(&now);
3333 delta = user_ns.clock - now_ns;
3334 kvm->arch.kvmclock_offset = delta;
3335 break;
3336 }
3337 case KVM_GET_CLOCK: {
3338 struct timespec now;
3339 struct kvm_clock_data user_ns;
3340 u64 now_ns;
3341
3342 ktime_get_ts(&now);
3343 now_ns = timespec_to_ns(&now);
3344 user_ns.clock = kvm->arch.kvmclock_offset + now_ns;
3345 user_ns.flags = 0;
3346
3347 r = -EFAULT;
3348 if (copy_to_user(argp, &user_ns, sizeof(user_ns)))
3349 goto out;
3350 r = 0;
3351 break;
3352 }
3353
Carsten Otte1fe779f2007-10-29 16:08:35 +01003354 default:
3355 ;
3356 }
3357out:
3358 return r;
3359}
3360
Zhang Xiantaoa16b0432007-11-16 14:38:21 +08003361static void kvm_init_msr_list(void)
Carsten Otte043405e2007-10-10 17:16:19 +02003362{
3363 u32 dummy[2];
3364 unsigned i, j;
3365
Glauber Costae3267cb2009-10-06 13:24:50 -04003366 /* skip the first msrs in the list. KVM-specific */
3367 for (i = j = KVM_SAVE_MSRS_BEGIN; i < ARRAY_SIZE(msrs_to_save); i++) {
Carsten Otte043405e2007-10-10 17:16:19 +02003368 if (rdmsr_safe(msrs_to_save[i], &dummy[0], &dummy[1]) < 0)
3369 continue;
3370 if (j < i)
3371 msrs_to_save[j] = msrs_to_save[i];
3372 j++;
3373 }
3374 num_msrs_to_save = j;
3375}
3376
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003377static int vcpu_mmio_write(struct kvm_vcpu *vcpu, gpa_t addr, int len,
3378 const void *v)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003379{
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003380 if (vcpu->arch.apic &&
3381 !kvm_iodevice_write(&vcpu->arch.apic->dev, addr, len, v))
3382 return 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003383
Marcelo Tosattie93f8a02009-12-23 14:35:24 -02003384 return kvm_io_bus_write(vcpu->kvm, KVM_MMIO_BUS, addr, len, v);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003385}
3386
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003387static int vcpu_mmio_read(struct kvm_vcpu *vcpu, gpa_t addr, int len, void *v)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003388{
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003389 if (vcpu->arch.apic &&
3390 !kvm_iodevice_read(&vcpu->arch.apic->dev, addr, len, v))
3391 return 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003392
Marcelo Tosattie93f8a02009-12-23 14:35:24 -02003393 return kvm_io_bus_read(vcpu->kvm, KVM_MMIO_BUS, addr, len, v);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003394}
3395
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003396static void kvm_set_segment(struct kvm_vcpu *vcpu,
3397 struct kvm_segment *var, int seg)
3398{
3399 kvm_x86_ops->set_segment(vcpu, var, seg);
3400}
3401
3402void kvm_get_segment(struct kvm_vcpu *vcpu,
3403 struct kvm_segment *var, int seg)
3404{
3405 kvm_x86_ops->get_segment(vcpu, var, seg);
3406}
3407
Gleb Natapov1871c602010-02-10 14:21:32 +02003408gpa_t kvm_mmu_gva_to_gpa_read(struct kvm_vcpu *vcpu, gva_t gva, u32 *error)
3409{
3410 u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
3411 return vcpu->arch.mmu.gva_to_gpa(vcpu, gva, access, error);
3412}
3413
3414 gpa_t kvm_mmu_gva_to_gpa_fetch(struct kvm_vcpu *vcpu, gva_t gva, u32 *error)
3415{
3416 u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
3417 access |= PFERR_FETCH_MASK;
3418 return vcpu->arch.mmu.gva_to_gpa(vcpu, gva, access, error);
3419}
3420
3421gpa_t kvm_mmu_gva_to_gpa_write(struct kvm_vcpu *vcpu, gva_t gva, u32 *error)
3422{
3423 u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
3424 access |= PFERR_WRITE_MASK;
3425 return vcpu->arch.mmu.gva_to_gpa(vcpu, gva, access, error);
3426}
3427
3428/* uses this to access any guest's mapped memory without checking CPL */
3429gpa_t kvm_mmu_gva_to_gpa_system(struct kvm_vcpu *vcpu, gva_t gva, u32 *error)
3430{
3431 return vcpu->arch.mmu.gva_to_gpa(vcpu, gva, 0, error);
3432}
3433
3434static int kvm_read_guest_virt_helper(gva_t addr, void *val, unsigned int bytes,
3435 struct kvm_vcpu *vcpu, u32 access,
3436 u32 *error)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003437{
3438 void *data = val;
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003439 int r = X86EMUL_CONTINUE;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003440
3441 while (bytes) {
Gleb Natapov1871c602010-02-10 14:21:32 +02003442 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr, access, error);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003443 unsigned offset = addr & (PAGE_SIZE-1);
Izik Eidus77c20022008-12-29 01:42:19 +02003444 unsigned toread = min(bytes, (unsigned)PAGE_SIZE - offset);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003445 int ret;
3446
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003447 if (gpa == UNMAPPED_GVA) {
3448 r = X86EMUL_PROPAGATE_FAULT;
3449 goto out;
3450 }
Izik Eidus77c20022008-12-29 01:42:19 +02003451 ret = kvm_read_guest(vcpu->kvm, gpa, data, toread);
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003452 if (ret < 0) {
Gleb Natapovc3cd7ff2010-04-28 19:15:35 +03003453 r = X86EMUL_IO_NEEDED;
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003454 goto out;
3455 }
Carsten Ottebbd9b642007-10-30 18:44:21 +01003456
Izik Eidus77c20022008-12-29 01:42:19 +02003457 bytes -= toread;
3458 data += toread;
3459 addr += toread;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003460 }
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003461out:
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003462 return r;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003463}
Izik Eidus77c20022008-12-29 01:42:19 +02003464
Gleb Natapov1871c602010-02-10 14:21:32 +02003465/* used for instruction fetching */
3466static int kvm_fetch_guest_virt(gva_t addr, void *val, unsigned int bytes,
3467 struct kvm_vcpu *vcpu, u32 *error)
3468{
3469 u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
3470 return kvm_read_guest_virt_helper(addr, val, bytes, vcpu,
3471 access | PFERR_FETCH_MASK, error);
3472}
3473
3474static int kvm_read_guest_virt(gva_t addr, void *val, unsigned int bytes,
3475 struct kvm_vcpu *vcpu, u32 *error)
3476{
3477 u32 access = (kvm_x86_ops->get_cpl(vcpu) == 3) ? PFERR_USER_MASK : 0;
3478 return kvm_read_guest_virt_helper(addr, val, bytes, vcpu, access,
3479 error);
3480}
3481
3482static int kvm_read_guest_virt_system(gva_t addr, void *val, unsigned int bytes,
3483 struct kvm_vcpu *vcpu, u32 *error)
3484{
3485 return kvm_read_guest_virt_helper(addr, val, bytes, vcpu, 0, error);
3486}
3487
Gleb Natapov79729952010-03-18 15:20:24 +02003488static int kvm_write_guest_virt_system(gva_t addr, void *val,
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003489 unsigned int bytes,
Gleb Natapov79729952010-03-18 15:20:24 +02003490 struct kvm_vcpu *vcpu,
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003491 u32 *error)
Izik Eidus77c20022008-12-29 01:42:19 +02003492{
3493 void *data = val;
3494 int r = X86EMUL_CONTINUE;
3495
3496 while (bytes) {
Gleb Natapov79729952010-03-18 15:20:24 +02003497 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr,
3498 PFERR_WRITE_MASK, error);
Izik Eidus77c20022008-12-29 01:42:19 +02003499 unsigned offset = addr & (PAGE_SIZE-1);
3500 unsigned towrite = min(bytes, (unsigned)PAGE_SIZE - offset);
3501 int ret;
3502
3503 if (gpa == UNMAPPED_GVA) {
3504 r = X86EMUL_PROPAGATE_FAULT;
3505 goto out;
3506 }
3507 ret = kvm_write_guest(vcpu->kvm, gpa, data, towrite);
3508 if (ret < 0) {
Gleb Natapovc3cd7ff2010-04-28 19:15:35 +03003509 r = X86EMUL_IO_NEEDED;
Izik Eidus77c20022008-12-29 01:42:19 +02003510 goto out;
3511 }
3512
3513 bytes -= towrite;
3514 data += towrite;
3515 addr += towrite;
3516 }
3517out:
3518 return r;
3519}
3520
Carsten Ottebbd9b642007-10-30 18:44:21 +01003521static int emulator_read_emulated(unsigned long addr,
3522 void *val,
3523 unsigned int bytes,
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003524 unsigned int *error_code,
Carsten Ottebbd9b642007-10-30 18:44:21 +01003525 struct kvm_vcpu *vcpu)
3526{
Carsten Ottebbd9b642007-10-30 18:44:21 +01003527 gpa_t gpa;
3528
3529 if (vcpu->mmio_read_completed) {
3530 memcpy(val, vcpu->mmio_data, bytes);
Avi Kivityaec51dc2009-07-01 16:01:02 +03003531 trace_kvm_mmio(KVM_TRACE_MMIO_READ, bytes,
3532 vcpu->mmio_phys_addr, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003533 vcpu->mmio_read_completed = 0;
3534 return X86EMUL_CONTINUE;
3535 }
3536
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003537 gpa = kvm_mmu_gva_to_gpa_read(vcpu, addr, error_code);
Gleb Natapov1871c602010-02-10 14:21:32 +02003538
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003539 if (gpa == UNMAPPED_GVA)
Gleb Natapov1871c602010-02-10 14:21:32 +02003540 return X86EMUL_PROPAGATE_FAULT;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003541
3542 /* For APIC access vmexit */
3543 if ((gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
3544 goto mmio;
3545
Gleb Natapov1871c602010-02-10 14:21:32 +02003546 if (kvm_read_guest_virt(addr, val, bytes, vcpu, NULL)
Izik Eidus77c20022008-12-29 01:42:19 +02003547 == X86EMUL_CONTINUE)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003548 return X86EMUL_CONTINUE;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003549
3550mmio:
3551 /*
3552 * Is this MMIO handled locally?
3553 */
Avi Kivityaec51dc2009-07-01 16:01:02 +03003554 if (!vcpu_mmio_read(vcpu, gpa, bytes, val)) {
3555 trace_kvm_mmio(KVM_TRACE_MMIO_READ, bytes, gpa, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003556 return X86EMUL_CONTINUE;
3557 }
Avi Kivityaec51dc2009-07-01 16:01:02 +03003558
3559 trace_kvm_mmio(KVM_TRACE_MMIO_READ_UNSATISFIED, bytes, gpa, 0);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003560
3561 vcpu->mmio_needed = 1;
Gleb Natapov411c35b2010-04-28 19:15:34 +03003562 vcpu->run->exit_reason = KVM_EXIT_MMIO;
3563 vcpu->run->mmio.phys_addr = vcpu->mmio_phys_addr = gpa;
3564 vcpu->run->mmio.len = vcpu->mmio_size = bytes;
3565 vcpu->run->mmio.is_write = vcpu->mmio_is_write = 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003566
Gleb Natapovc3cd7ff2010-04-28 19:15:35 +03003567 return X86EMUL_IO_NEEDED;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003568}
3569
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003570int emulator_write_phys(struct kvm_vcpu *vcpu, gpa_t gpa,
Avi Kivity9f811282008-03-02 14:06:05 +02003571 const void *val, int bytes)
3572{
3573 int ret;
3574
3575 ret = kvm_write_guest(vcpu->kvm, gpa, val, bytes);
3576 if (ret < 0)
3577 return 0;
Marcelo Tosattiad218f82008-12-01 22:32:05 -02003578 kvm_mmu_pte_write(vcpu, gpa, val, bytes, 1);
Avi Kivity9f811282008-03-02 14:06:05 +02003579 return 1;
3580}
3581
Carsten Ottebbd9b642007-10-30 18:44:21 +01003582static int emulator_write_emulated_onepage(unsigned long addr,
3583 const void *val,
3584 unsigned int bytes,
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003585 unsigned int *error_code,
Carsten Ottebbd9b642007-10-30 18:44:21 +01003586 struct kvm_vcpu *vcpu)
3587{
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003588 gpa_t gpa;
3589
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003590 gpa = kvm_mmu_gva_to_gpa_write(vcpu, addr, error_code);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003591
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003592 if (gpa == UNMAPPED_GVA)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003593 return X86EMUL_PROPAGATE_FAULT;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003594
3595 /* For APIC access vmexit */
3596 if ((gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
3597 goto mmio;
3598
3599 if (emulator_write_phys(vcpu, gpa, val, bytes))
3600 return X86EMUL_CONTINUE;
3601
3602mmio:
Avi Kivityaec51dc2009-07-01 16:01:02 +03003603 trace_kvm_mmio(KVM_TRACE_MMIO_WRITE, bytes, gpa, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003604 /*
3605 * Is this MMIO handled locally?
3606 */
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003607 if (!vcpu_mmio_write(vcpu, gpa, bytes, val))
Carsten Ottebbd9b642007-10-30 18:44:21 +01003608 return X86EMUL_CONTINUE;
Carsten Ottebbd9b642007-10-30 18:44:21 +01003609
3610 vcpu->mmio_needed = 1;
Gleb Natapov411c35b2010-04-28 19:15:34 +03003611 vcpu->run->exit_reason = KVM_EXIT_MMIO;
3612 vcpu->run->mmio.phys_addr = vcpu->mmio_phys_addr = gpa;
3613 vcpu->run->mmio.len = vcpu->mmio_size = bytes;
3614 vcpu->run->mmio.is_write = vcpu->mmio_is_write = 1;
3615 memcpy(vcpu->run->mmio.data, val, bytes);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003616
3617 return X86EMUL_CONTINUE;
3618}
3619
3620int emulator_write_emulated(unsigned long addr,
Gleb Natapov8f6abd02010-04-13 10:21:56 +03003621 const void *val,
3622 unsigned int bytes,
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003623 unsigned int *error_code,
Gleb Natapov8f6abd02010-04-13 10:21:56 +03003624 struct kvm_vcpu *vcpu)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003625{
3626 /* Crossing a page boundary? */
3627 if (((addr + bytes - 1) ^ addr) & PAGE_MASK) {
3628 int rc, now;
3629
3630 now = -addr & ~PAGE_MASK;
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003631 rc = emulator_write_emulated_onepage(addr, val, now, error_code,
3632 vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003633 if (rc != X86EMUL_CONTINUE)
3634 return rc;
3635 addr += now;
3636 val += now;
3637 bytes -= now;
3638 }
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003639 return emulator_write_emulated_onepage(addr, val, bytes, error_code,
3640 vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003641}
Carsten Ottebbd9b642007-10-30 18:44:21 +01003642
Avi Kivitydaea3e72010-03-15 13:59:54 +02003643#define CMPXCHG_TYPE(t, ptr, old, new) \
3644 (cmpxchg((t *)(ptr), *(t *)(old), *(t *)(new)) == *(t *)(old))
3645
3646#ifdef CONFIG_X86_64
3647# define CMPXCHG64(ptr, old, new) CMPXCHG_TYPE(u64, ptr, old, new)
3648#else
3649# define CMPXCHG64(ptr, old, new) \
Jan Kiszka9749a6c2010-03-20 10:14:13 +01003650 (cmpxchg64((u64 *)(ptr), *(u64 *)(old), *(u64 *)(new)) == *(u64 *)(old))
Avi Kivitydaea3e72010-03-15 13:59:54 +02003651#endif
3652
Carsten Ottebbd9b642007-10-30 18:44:21 +01003653static int emulator_cmpxchg_emulated(unsigned long addr,
3654 const void *old,
3655 const void *new,
3656 unsigned int bytes,
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003657 unsigned int *error_code,
Carsten Ottebbd9b642007-10-30 18:44:21 +01003658 struct kvm_vcpu *vcpu)
3659{
Avi Kivitydaea3e72010-03-15 13:59:54 +02003660 gpa_t gpa;
3661 struct page *page;
3662 char *kaddr;
3663 bool exchanged;
3664
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003665 /* guests cmpxchg8b have to be emulated atomically */
Avi Kivitydaea3e72010-03-15 13:59:54 +02003666 if (bytes > 8 || (bytes & (bytes - 1)))
3667 goto emul_write;
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003668
Avi Kivitydaea3e72010-03-15 13:59:54 +02003669 gpa = kvm_mmu_gva_to_gpa_write(vcpu, addr, NULL);
Marcelo Tosatti10589a42007-12-20 19:18:22 -05003670
Avi Kivitydaea3e72010-03-15 13:59:54 +02003671 if (gpa == UNMAPPED_GVA ||
3672 (gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
3673 goto emul_write;
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003674
Avi Kivitydaea3e72010-03-15 13:59:54 +02003675 if (((gpa + bytes - 1) & PAGE_MASK) != (gpa & PAGE_MASK))
3676 goto emul_write;
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003677
Avi Kivitydaea3e72010-03-15 13:59:54 +02003678 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
Izik Eidus72dc67a2008-02-10 18:04:15 +02003679
Avi Kivitydaea3e72010-03-15 13:59:54 +02003680 kaddr = kmap_atomic(page, KM_USER0);
3681 kaddr += offset_in_page(gpa);
3682 switch (bytes) {
3683 case 1:
3684 exchanged = CMPXCHG_TYPE(u8, kaddr, old, new);
3685 break;
3686 case 2:
3687 exchanged = CMPXCHG_TYPE(u16, kaddr, old, new);
3688 break;
3689 case 4:
3690 exchanged = CMPXCHG_TYPE(u32, kaddr, old, new);
3691 break;
3692 case 8:
3693 exchanged = CMPXCHG64(kaddr, old, new);
3694 break;
3695 default:
3696 BUG();
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003697 }
Avi Kivitydaea3e72010-03-15 13:59:54 +02003698 kunmap_atomic(kaddr, KM_USER0);
3699 kvm_release_page_dirty(page);
3700
3701 if (!exchanged)
3702 return X86EMUL_CMPXCHG_FAILED;
3703
Gleb Natapov8f6abd02010-04-13 10:21:56 +03003704 kvm_mmu_pte_write(vcpu, gpa, new, bytes, 1);
3705
3706 return X86EMUL_CONTINUE;
Avi Kivity4a5f48f2010-03-15 13:59:55 +02003707
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003708emul_write:
Avi Kivitydaea3e72010-03-15 13:59:54 +02003709 printk_once(KERN_WARNING "kvm: emulating exchange as write\n");
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05003710
Gleb Natapov8fe681e2010-04-28 19:15:37 +03003711 return emulator_write_emulated(addr, new, bytes, error_code, vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003712}
3713
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003714static int kernel_pio(struct kvm_vcpu *vcpu, void *pd)
3715{
3716 /* TODO: String I/O for in kernel device */
3717 int r;
3718
3719 if (vcpu->arch.pio.in)
3720 r = kvm_io_bus_read(vcpu->kvm, KVM_PIO_BUS, vcpu->arch.pio.port,
3721 vcpu->arch.pio.size, pd);
3722 else
3723 r = kvm_io_bus_write(vcpu->kvm, KVM_PIO_BUS,
3724 vcpu->arch.pio.port, vcpu->arch.pio.size,
3725 pd);
3726 return r;
3727}
3728
3729
3730static int emulator_pio_in_emulated(int size, unsigned short port, void *val,
3731 unsigned int count, struct kvm_vcpu *vcpu)
3732{
Gleb Natapov79729952010-03-18 15:20:24 +02003733 if (vcpu->arch.pio.count)
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003734 goto data_avail;
3735
3736 trace_kvm_pio(1, port, size, 1);
3737
3738 vcpu->arch.pio.port = port;
3739 vcpu->arch.pio.in = 1;
Gleb Natapov79729952010-03-18 15:20:24 +02003740 vcpu->arch.pio.count = count;
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003741 vcpu->arch.pio.size = size;
3742
3743 if (!kernel_pio(vcpu, vcpu->arch.pio_data)) {
3744 data_avail:
3745 memcpy(val, vcpu->arch.pio_data, size * count);
Gleb Natapov79729952010-03-18 15:20:24 +02003746 vcpu->arch.pio.count = 0;
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003747 return 1;
3748 }
3749
3750 vcpu->run->exit_reason = KVM_EXIT_IO;
3751 vcpu->run->io.direction = KVM_EXIT_IO_IN;
3752 vcpu->run->io.size = size;
3753 vcpu->run->io.data_offset = KVM_PIO_PAGE_OFFSET * PAGE_SIZE;
3754 vcpu->run->io.count = count;
3755 vcpu->run->io.port = port;
3756
3757 return 0;
3758}
3759
3760static int emulator_pio_out_emulated(int size, unsigned short port,
3761 const void *val, unsigned int count,
3762 struct kvm_vcpu *vcpu)
3763{
3764 trace_kvm_pio(0, port, size, 1);
3765
3766 vcpu->arch.pio.port = port;
3767 vcpu->arch.pio.in = 0;
Gleb Natapov79729952010-03-18 15:20:24 +02003768 vcpu->arch.pio.count = count;
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003769 vcpu->arch.pio.size = size;
3770
3771 memcpy(vcpu->arch.pio_data, val, size * count);
3772
3773 if (!kernel_pio(vcpu, vcpu->arch.pio_data)) {
Gleb Natapov79729952010-03-18 15:20:24 +02003774 vcpu->arch.pio.count = 0;
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003775 return 1;
3776 }
3777
3778 vcpu->run->exit_reason = KVM_EXIT_IO;
3779 vcpu->run->io.direction = KVM_EXIT_IO_OUT;
3780 vcpu->run->io.size = size;
3781 vcpu->run->io.data_offset = KVM_PIO_PAGE_OFFSET * PAGE_SIZE;
3782 vcpu->run->io.count = count;
3783 vcpu->run->io.port = port;
3784
3785 return 0;
3786}
3787
Carsten Ottebbd9b642007-10-30 18:44:21 +01003788static unsigned long get_segment_base(struct kvm_vcpu *vcpu, int seg)
3789{
3790 return kvm_x86_ops->get_segment_base(vcpu, seg);
3791}
3792
3793int emulate_invlpg(struct kvm_vcpu *vcpu, gva_t address)
3794{
Marcelo Tosattia7052892008-09-23 13:18:35 -03003795 kvm_mmu_invlpg(vcpu, address);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003796 return X86EMUL_CONTINUE;
3797}
3798
3799int emulate_clts(struct kvm_vcpu *vcpu)
3800{
Avi Kivity4d4ec082009-12-29 18:07:30 +02003801 kvm_x86_ops->set_cr0(vcpu, kvm_read_cr0_bits(vcpu, ~X86_CR0_TS));
Avi Kivity6b52d182010-01-21 15:31:47 +02003802 kvm_x86_ops->fpu_activate(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003803 return X86EMUL_CONTINUE;
3804}
3805
Gleb Natapov35aa5372010-04-28 19:15:27 +03003806int emulator_get_dr(int dr, unsigned long *dest, struct kvm_vcpu *vcpu)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003807{
Gleb Natapov338dbc92010-04-28 19:15:32 +03003808 return _kvm_get_dr(vcpu, dr, dest);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003809}
3810
Gleb Natapov35aa5372010-04-28 19:15:27 +03003811int emulator_set_dr(int dr, unsigned long value, struct kvm_vcpu *vcpu)
Carsten Ottebbd9b642007-10-30 18:44:21 +01003812{
Gleb Natapov338dbc92010-04-28 19:15:32 +03003813
3814 return __kvm_set_dr(vcpu, dr, value);
Carsten Ottebbd9b642007-10-30 18:44:21 +01003815}
3816
Gleb Natapov52a46612010-03-18 15:20:03 +02003817static u64 mk_cr_64(u64 curr_cr, u32 new_val)
3818{
3819 return (curr_cr & ~((1ULL << 32) - 1)) | new_val;
3820}
3821
3822static unsigned long emulator_get_cr(int cr, struct kvm_vcpu *vcpu)
3823{
3824 unsigned long value;
3825
3826 switch (cr) {
3827 case 0:
3828 value = kvm_read_cr0(vcpu);
3829 break;
3830 case 2:
3831 value = vcpu->arch.cr2;
3832 break;
3833 case 3:
3834 value = vcpu->arch.cr3;
3835 break;
3836 case 4:
3837 value = kvm_read_cr4(vcpu);
3838 break;
3839 case 8:
3840 value = kvm_get_cr8(vcpu);
3841 break;
3842 default:
3843 vcpu_printf(vcpu, "%s: unexpected cr %u\n", __func__, cr);
3844 return 0;
3845 }
3846
3847 return value;
3848}
3849
Gleb Natapov0f122442010-04-28 19:15:31 +03003850static int emulator_set_cr(int cr, unsigned long val, struct kvm_vcpu *vcpu)
Gleb Natapov52a46612010-03-18 15:20:03 +02003851{
Gleb Natapov0f122442010-04-28 19:15:31 +03003852 int res = 0;
3853
Gleb Natapov52a46612010-03-18 15:20:03 +02003854 switch (cr) {
3855 case 0:
Avi Kivity49a9b072010-06-10 17:02:14 +03003856 res = kvm_set_cr0(vcpu, mk_cr_64(kvm_read_cr0(vcpu), val));
Gleb Natapov52a46612010-03-18 15:20:03 +02003857 break;
3858 case 2:
3859 vcpu->arch.cr2 = val;
3860 break;
3861 case 3:
Avi Kivity23902182010-06-10 17:02:16 +03003862 res = kvm_set_cr3(vcpu, val);
Gleb Natapov52a46612010-03-18 15:20:03 +02003863 break;
3864 case 4:
Avi Kivitya83b29c2010-06-10 17:02:15 +03003865 res = kvm_set_cr4(vcpu, mk_cr_64(kvm_read_cr4(vcpu), val));
Gleb Natapov52a46612010-03-18 15:20:03 +02003866 break;
3867 case 8:
Gleb Natapov0f122442010-04-28 19:15:31 +03003868 res = __kvm_set_cr8(vcpu, val & 0xfUL);
Gleb Natapov52a46612010-03-18 15:20:03 +02003869 break;
3870 default:
3871 vcpu_printf(vcpu, "%s: unexpected cr %u\n", __func__, cr);
Gleb Natapov0f122442010-04-28 19:15:31 +03003872 res = -1;
Gleb Natapov52a46612010-03-18 15:20:03 +02003873 }
Gleb Natapov0f122442010-04-28 19:15:31 +03003874
3875 return res;
Gleb Natapov52a46612010-03-18 15:20:03 +02003876}
3877
Gleb Natapov9c537242010-03-18 15:20:05 +02003878static int emulator_get_cpl(struct kvm_vcpu *vcpu)
3879{
3880 return kvm_x86_ops->get_cpl(vcpu);
3881}
3882
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003883static void emulator_get_gdt(struct desc_ptr *dt, struct kvm_vcpu *vcpu)
3884{
3885 kvm_x86_ops->get_gdt(vcpu, dt);
3886}
3887
Gleb Natapov5951c442010-04-28 19:15:29 +03003888static unsigned long emulator_get_cached_segment_base(int seg,
3889 struct kvm_vcpu *vcpu)
3890{
3891 return get_segment_base(vcpu, seg);
3892}
3893
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003894static bool emulator_get_cached_descriptor(struct desc_struct *desc, int seg,
3895 struct kvm_vcpu *vcpu)
3896{
3897 struct kvm_segment var;
3898
3899 kvm_get_segment(vcpu, &var, seg);
3900
3901 if (var.unusable)
3902 return false;
3903
3904 if (var.g)
3905 var.limit >>= 12;
3906 set_desc_limit(desc, var.limit);
3907 set_desc_base(desc, (unsigned long)var.base);
3908 desc->type = var.type;
3909 desc->s = var.s;
3910 desc->dpl = var.dpl;
3911 desc->p = var.present;
3912 desc->avl = var.avl;
3913 desc->l = var.l;
3914 desc->d = var.db;
3915 desc->g = var.g;
3916
3917 return true;
3918}
3919
3920static void emulator_set_cached_descriptor(struct desc_struct *desc, int seg,
3921 struct kvm_vcpu *vcpu)
3922{
3923 struct kvm_segment var;
3924
3925 /* needed to preserve selector */
3926 kvm_get_segment(vcpu, &var, seg);
3927
3928 var.base = get_desc_base(desc);
3929 var.limit = get_desc_limit(desc);
3930 if (desc->g)
3931 var.limit = (var.limit << 12) | 0xfff;
3932 var.type = desc->type;
3933 var.present = desc->p;
3934 var.dpl = desc->dpl;
3935 var.db = desc->d;
3936 var.s = desc->s;
3937 var.l = desc->l;
3938 var.g = desc->g;
3939 var.avl = desc->avl;
3940 var.present = desc->p;
3941 var.unusable = !var.present;
3942 var.padding = 0;
3943
3944 kvm_set_segment(vcpu, &var, seg);
3945 return;
3946}
3947
3948static u16 emulator_get_segment_selector(int seg, struct kvm_vcpu *vcpu)
3949{
3950 struct kvm_segment kvm_seg;
3951
3952 kvm_get_segment(vcpu, &kvm_seg, seg);
3953 return kvm_seg.selector;
3954}
3955
3956static void emulator_set_segment_selector(u16 sel, int seg,
3957 struct kvm_vcpu *vcpu)
3958{
3959 struct kvm_segment kvm_seg;
3960
3961 kvm_get_segment(vcpu, &kvm_seg, seg);
3962 kvm_seg.selector = sel;
3963 kvm_set_segment(vcpu, &kvm_seg, seg);
3964}
3965
Harvey Harrison14af3f32008-02-19 10:25:50 -08003966static struct x86_emulate_ops emulate_ops = {
Gleb Natapov1871c602010-02-10 14:21:32 +02003967 .read_std = kvm_read_guest_virt_system,
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003968 .write_std = kvm_write_guest_virt_system,
Gleb Natapov1871c602010-02-10 14:21:32 +02003969 .fetch = kvm_fetch_guest_virt,
Carsten Ottebbd9b642007-10-30 18:44:21 +01003970 .read_emulated = emulator_read_emulated,
3971 .write_emulated = emulator_write_emulated,
3972 .cmpxchg_emulated = emulator_cmpxchg_emulated,
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02003973 .pio_in_emulated = emulator_pio_in_emulated,
3974 .pio_out_emulated = emulator_pio_out_emulated,
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003975 .get_cached_descriptor = emulator_get_cached_descriptor,
3976 .set_cached_descriptor = emulator_set_cached_descriptor,
3977 .get_segment_selector = emulator_get_segment_selector,
3978 .set_segment_selector = emulator_set_segment_selector,
Gleb Natapov5951c442010-04-28 19:15:29 +03003979 .get_cached_segment_base = emulator_get_cached_segment_base,
Gleb Natapov2dafc6c2010-03-18 15:20:16 +02003980 .get_gdt = emulator_get_gdt,
Gleb Natapov52a46612010-03-18 15:20:03 +02003981 .get_cr = emulator_get_cr,
3982 .set_cr = emulator_set_cr,
Gleb Natapov9c537242010-03-18 15:20:05 +02003983 .cpl = emulator_get_cpl,
Gleb Natapov35aa5372010-04-28 19:15:27 +03003984 .get_dr = emulator_get_dr,
3985 .set_dr = emulator_set_dr,
Gleb Natapov3fb1b5d2010-04-28 19:15:28 +03003986 .set_msr = kvm_set_msr,
3987 .get_msr = kvm_get_msr,
Carsten Ottebbd9b642007-10-30 18:44:21 +01003988};
3989
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003990static void cache_all_regs(struct kvm_vcpu *vcpu)
3991{
3992 kvm_register_read(vcpu, VCPU_REGS_RAX);
3993 kvm_register_read(vcpu, VCPU_REGS_RSP);
3994 kvm_register_read(vcpu, VCPU_REGS_RIP);
3995 vcpu->arch.regs_dirty = ~0;
3996}
3997
Gleb Natapov95cb2292010-04-28 19:15:43 +03003998static void toggle_interruptibility(struct kvm_vcpu *vcpu, u32 mask)
3999{
4000 u32 int_shadow = kvm_x86_ops->get_interrupt_shadow(vcpu, mask);
4001 /*
4002 * an sti; sti; sequence only disable interrupts for the first
4003 * instruction. So, if the last instruction, be it emulated or
4004 * not, left the system with the INT_STI flag enabled, it
4005 * means that the last instruction is an sti. We should not
4006 * leave the flag on in this case. The same goes for mov ss
4007 */
4008 if (!(int_shadow & mask))
4009 kvm_x86_ops->set_interrupt_shadow(vcpu, mask);
4010}
4011
Gleb Natapov54b84862010-04-28 19:15:44 +03004012static void inject_emulated_exception(struct kvm_vcpu *vcpu)
4013{
4014 struct x86_emulate_ctxt *ctxt = &vcpu->arch.emulate_ctxt;
4015 if (ctxt->exception == PF_VECTOR)
4016 kvm_inject_page_fault(vcpu, ctxt->cr2, ctxt->error_code);
4017 else if (ctxt->error_code_valid)
4018 kvm_queue_exception_e(vcpu, ctxt->exception, ctxt->error_code);
4019 else
4020 kvm_queue_exception(vcpu, ctxt->exception);
4021}
4022
Gleb Natapov6d77dbf2010-05-10 11:16:56 +03004023static int handle_emulation_failure(struct kvm_vcpu *vcpu)
4024{
Gleb Natapov6d77dbf2010-05-10 11:16:56 +03004025 ++vcpu->stat.insn_emulation_fail;
4026 trace_kvm_emulate_insn_failed(vcpu);
4027 vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
4028 vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_EMULATION;
4029 vcpu->run->internal.ndata = 0;
4030 kvm_queue_exception(vcpu, UD_VECTOR);
4031 return EMULATE_FAIL;
4032}
4033
Carsten Ottebbd9b642007-10-30 18:44:21 +01004034int emulate_instruction(struct kvm_vcpu *vcpu,
Carsten Ottebbd9b642007-10-30 18:44:21 +01004035 unsigned long cr2,
4036 u16 error_code,
Sheng Yang571008d2008-01-02 14:49:22 +08004037 int emulation_type)
Carsten Ottebbd9b642007-10-30 18:44:21 +01004038{
Gleb Natapov95cb2292010-04-28 19:15:43 +03004039 int r;
Gleb Natapov4d2179e2010-04-28 19:15:42 +03004040 struct decode_cache *c = &vcpu->arch.emulate_ctxt.decode;
Carsten Ottebbd9b642007-10-30 18:44:21 +01004041
Avi Kivity26eef702008-07-03 14:59:22 +03004042 kvm_clear_exception_queue(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004043 vcpu->arch.mmio_fault_cr2 = cr2;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004044 /*
Avi Kivity56e82312009-08-12 15:04:37 +03004045 * TODO: fix emulate.c to use guest_read/write_register
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004046 * instead of direct ->regs accesses, can save hundred cycles
4047 * on Intel for instructions that don't read/change RSP, for
4048 * for example.
4049 */
4050 cache_all_regs(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01004051
Sheng Yang571008d2008-01-02 14:49:22 +08004052 if (!(emulation_type & EMULTYPE_NO_DECODE)) {
Carsten Ottebbd9b642007-10-30 18:44:21 +01004053 int cs_db, cs_l;
4054 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
4055
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004056 vcpu->arch.emulate_ctxt.vcpu = vcpu;
Jan Kiszka83bf0002010-02-23 17:47:59 +01004057 vcpu->arch.emulate_ctxt.eflags = kvm_x86_ops->get_rflags(vcpu);
Gleb Natapov063db062010-03-18 15:20:06 +02004058 vcpu->arch.emulate_ctxt.eip = kvm_rip_read(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004059 vcpu->arch.emulate_ctxt.mode =
Gleb Natapova0044752010-02-10 14:21:31 +02004060 (!is_protmode(vcpu)) ? X86EMUL_MODE_REAL :
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004061 (vcpu->arch.emulate_ctxt.eflags & X86_EFLAGS_VM)
Gleb Natapova0044752010-02-10 14:21:31 +02004062 ? X86EMUL_MODE_VM86 : cs_l
Carsten Ottebbd9b642007-10-30 18:44:21 +01004063 ? X86EMUL_MODE_PROT64 : cs_db
4064 ? X86EMUL_MODE_PROT32 : X86EMUL_MODE_PROT16;
Gleb Natapov4d2179e2010-04-28 19:15:42 +03004065 memset(c, 0, sizeof(struct decode_cache));
4066 memcpy(c->regs, vcpu->arch.regs, sizeof c->regs);
Gleb Natapov95cb2292010-04-28 19:15:43 +03004067 vcpu->arch.emulate_ctxt.interruptibility = 0;
Gleb Natapov54b84862010-04-28 19:15:44 +03004068 vcpu->arch.emulate_ctxt.exception = -1;
Carsten Ottebbd9b642007-10-30 18:44:21 +01004069
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004070 r = x86_decode_insn(&vcpu->arch.emulate_ctxt, &emulate_ops);
Avi Kivitye46479f2010-04-11 13:05:16 +03004071 trace_kvm_emulate_insn_start(vcpu);
Sheng Yang571008d2008-01-02 14:49:22 +08004072
Andre Przywara0cb57622009-06-17 15:50:31 +02004073 /* Only allow emulation of specific instructions on #UD
4074 * (namely VMMCALL, sysenter, sysexit, syscall)*/
Andre Przywara0cb57622009-06-17 15:50:31 +02004075 if (emulation_type & EMULTYPE_TRAP_UD) {
4076 if (!c->twobyte)
4077 return EMULATE_FAIL;
4078 switch (c->b) {
4079 case 0x01: /* VMMCALL */
4080 if (c->modrm_mod != 3 || c->modrm_rm != 1)
4081 return EMULATE_FAIL;
4082 break;
4083 case 0x34: /* sysenter */
4084 case 0x35: /* sysexit */
4085 if (c->modrm_mod != 0 || c->modrm_rm != 0)
4086 return EMULATE_FAIL;
4087 break;
4088 case 0x05: /* syscall */
4089 if (c->modrm_mod != 0 || c->modrm_rm != 0)
4090 return EMULATE_FAIL;
4091 break;
4092 default:
4093 return EMULATE_FAIL;
4094 }
4095
4096 if (!(c->modrm_reg == 0 || c->modrm_reg == 3))
4097 return EMULATE_FAIL;
4098 }
Sheng Yang571008d2008-01-02 14:49:22 +08004099
Avi Kivityf2b57562007-11-18 15:17:51 +02004100 ++vcpu->stat.insn_emulation;
Carsten Ottebbd9b642007-10-30 18:44:21 +01004101 if (r) {
4102 if (kvm_mmu_unprotect_page_virt(vcpu, cr2))
4103 return EMULATE_DONE;
Gleb Natapov6d77dbf2010-05-10 11:16:56 +03004104 if (emulation_type & EMULTYPE_SKIP)
4105 return EMULATE_FAIL;
4106 return handle_emulation_failure(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01004107 }
4108 }
4109
Gleb Natapovba8afb62009-04-12 13:36:57 +03004110 if (emulation_type & EMULTYPE_SKIP) {
4111 kvm_rip_write(vcpu, vcpu->arch.emulate_ctxt.decode.eip);
4112 return EMULATE_DONE;
4113 }
4114
Gleb Natapov4d2179e2010-04-28 19:15:42 +03004115 /* this is needed for vmware backdor interface to work since it
4116 changes registers values during IO operation */
4117 memcpy(c->regs, vcpu->arch.regs, sizeof c->regs);
4118
Gleb Natapov5cd21912010-03-18 15:20:26 +02004119restart:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004120 r = x86_emulate_insn(&vcpu->arch.emulate_ctxt, &emulate_ops);
Gleb Natapovc3cd7ff2010-04-28 19:15:35 +03004121
4122 if (r) { /* emulation failed */
4123 /*
4124 * if emulation was due to access to shadowed page table
4125 * and it failed try to unshadow page and re-entetr the
4126 * guest to let CPU execute the instruction.
4127 */
Carsten Ottebbd9b642007-10-30 18:44:21 +01004128 if (kvm_mmu_unprotect_page_virt(vcpu, cr2))
Gleb Natapovc3cd7ff2010-04-28 19:15:35 +03004129 return EMULATE_DONE;
4130
Gleb Natapov6d77dbf2010-05-10 11:16:56 +03004131 return handle_emulation_failure(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01004132 }
4133
Gleb Natapov95cb2292010-04-28 19:15:43 +03004134 toggle_interruptibility(vcpu, vcpu->arch.emulate_ctxt.interruptibility);
Gleb Natapovef050dc2010-04-28 19:15:40 +03004135 kvm_x86_ops->set_rflags(vcpu, vcpu->arch.emulate_ctxt.eflags);
Gleb Natapov4d2179e2010-04-28 19:15:42 +03004136 memcpy(vcpu->arch.regs, c->regs, sizeof c->regs);
Gleb Natapov95c55882010-04-28 19:15:39 +03004137 kvm_rip_write(vcpu, vcpu->arch.emulate_ctxt.eip);
Gleb Natapov3457e412010-04-28 19:15:38 +03004138
Gleb Natapov54b84862010-04-28 19:15:44 +03004139 if (vcpu->arch.emulate_ctxt.exception >= 0) {
4140 inject_emulated_exception(vcpu);
4141 return EMULATE_DONE;
4142 }
4143
Gleb Natapov3457e412010-04-28 19:15:38 +03004144 if (vcpu->arch.pio.count) {
4145 if (!vcpu->arch.pio.in)
4146 vcpu->arch.pio.count = 0;
4147 return EMULATE_DO_MMIO;
4148 }
4149
4150 if (vcpu->mmio_needed) {
4151 if (vcpu->mmio_is_write)
4152 vcpu->mmio_needed = 0;
4153 return EMULATE_DO_MMIO;
4154 }
4155
Gleb Natapov5cd21912010-03-18 15:20:26 +02004156 if (vcpu->arch.emulate_ctxt.restart)
4157 goto restart;
4158
Carsten Ottebbd9b642007-10-30 18:44:21 +01004159 return EMULATE_DONE;
4160}
4161EXPORT_SYMBOL_GPL(emulate_instruction);
4162
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02004163int kvm_fast_pio_out(struct kvm_vcpu *vcpu, int size, unsigned short port)
Carsten Ottede7d7892007-10-30 18:44:25 +01004164{
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02004165 unsigned long val = kvm_register_read(vcpu, VCPU_REGS_RAX);
4166 int ret = emulator_pio_out_emulated(size, port, &val, 1, vcpu);
4167 /* do not return to emulator after return from userspace */
Gleb Natapov79729952010-03-18 15:20:24 +02004168 vcpu->arch.pio.count = 0;
Izik Eidus0f346072008-12-29 01:42:20 +02004169 return ret;
Carsten Ottede7d7892007-10-30 18:44:25 +01004170}
Gleb Natapovcf8f70b2010-03-18 15:20:23 +02004171EXPORT_SYMBOL_GPL(kvm_fast_pio_out);
Carsten Ottede7d7892007-10-30 18:44:25 +01004172
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004173static void bounce_off(void *info)
4174{
4175 /* nothing */
4176}
4177
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004178static int kvmclock_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
4179 void *data)
4180{
4181 struct cpufreq_freqs *freq = data;
4182 struct kvm *kvm;
4183 struct kvm_vcpu *vcpu;
4184 int i, send_ipi = 0;
4185
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004186 if (val == CPUFREQ_PRECHANGE && freq->old > freq->new)
4187 return 0;
4188 if (val == CPUFREQ_POSTCHANGE && freq->old < freq->new)
4189 return 0;
Zachary Amsden0cca7902009-09-29 11:38:35 -10004190 per_cpu(cpu_tsc_khz, freq->cpu) = freq->new;
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004191
4192 spin_lock(&kvm_lock);
4193 list_for_each_entry(kvm, &vm_list, vm_list) {
Gleb Natapov988a2ca2009-06-09 15:56:29 +03004194 kvm_for_each_vcpu(i, vcpu, kvm) {
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004195 if (vcpu->cpu != freq->cpu)
4196 continue;
4197 if (!kvm_request_guest_time_update(vcpu))
4198 continue;
4199 if (vcpu->cpu != smp_processor_id())
4200 send_ipi++;
4201 }
4202 }
4203 spin_unlock(&kvm_lock);
4204
4205 if (freq->old < freq->new && send_ipi) {
4206 /*
4207 * We upscale the frequency. Must make the guest
4208 * doesn't see old kvmclock values while running with
4209 * the new frequency, otherwise we risk the guest sees
4210 * time go backwards.
4211 *
4212 * In case we update the frequency for another cpu
4213 * (which might be in guest context) send an interrupt
4214 * to kick the cpu out of guest context. Next time
4215 * guest context is entered kvmclock will be updated,
4216 * so the guest will not see stale values.
4217 */
4218 smp_call_function_single(freq->cpu, bounce_off, NULL, 1);
4219 }
4220 return 0;
4221}
4222
4223static struct notifier_block kvmclock_cpufreq_notifier_block = {
4224 .notifier_call = kvmclock_cpufreq_notifier
4225};
4226
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004227static void kvm_timer_init(void)
4228{
4229 int cpu;
4230
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004231 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004232 cpufreq_register_notifier(&kvmclock_cpufreq_notifier_block,
4233 CPUFREQ_TRANSITION_NOTIFIER);
Zachary Amsden6b7d7e72009-10-09 16:26:08 -10004234 for_each_online_cpu(cpu) {
4235 unsigned long khz = cpufreq_get(cpu);
4236 if (!khz)
4237 khz = tsc_khz;
4238 per_cpu(cpu_tsc_khz, cpu) = khz;
4239 }
Zachary Amsden0cca7902009-09-29 11:38:35 -10004240 } else {
4241 for_each_possible_cpu(cpu)
4242 per_cpu(cpu_tsc_khz, cpu) = tsc_khz;
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004243 }
4244}
4245
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004246static DEFINE_PER_CPU(struct kvm_vcpu *, current_vcpu);
4247
4248static int kvm_is_in_guest(void)
4249{
4250 return percpu_read(current_vcpu) != NULL;
4251}
4252
4253static int kvm_is_user_mode(void)
4254{
4255 int user_mode = 3;
Zhang, Yanmindcf46b92010-04-20 10:13:58 +08004256
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004257 if (percpu_read(current_vcpu))
4258 user_mode = kvm_x86_ops->get_cpl(percpu_read(current_vcpu));
Zhang, Yanmindcf46b92010-04-20 10:13:58 +08004259
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004260 return user_mode != 0;
4261}
4262
4263static unsigned long kvm_get_guest_ip(void)
4264{
4265 unsigned long ip = 0;
Zhang, Yanmindcf46b92010-04-20 10:13:58 +08004266
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004267 if (percpu_read(current_vcpu))
4268 ip = kvm_rip_read(percpu_read(current_vcpu));
Zhang, Yanmindcf46b92010-04-20 10:13:58 +08004269
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004270 return ip;
4271}
4272
4273static struct perf_guest_info_callbacks kvm_guest_cbs = {
4274 .is_in_guest = kvm_is_in_guest,
4275 .is_user_mode = kvm_is_user_mode,
4276 .get_guest_ip = kvm_get_guest_ip,
4277};
4278
4279void kvm_before_handle_nmi(struct kvm_vcpu *vcpu)
4280{
4281 percpu_write(current_vcpu, vcpu);
4282}
4283EXPORT_SYMBOL_GPL(kvm_before_handle_nmi);
4284
4285void kvm_after_handle_nmi(struct kvm_vcpu *vcpu)
4286{
4287 percpu_write(current_vcpu, NULL);
4288}
4289EXPORT_SYMBOL_GPL(kvm_after_handle_nmi);
4290
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004291int kvm_arch_init(void *opaque)
Carsten Otte043405e2007-10-10 17:16:19 +02004292{
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004293 int r;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004294 struct kvm_x86_ops *ops = (struct kvm_x86_ops *)opaque;
4295
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004296 if (kvm_x86_ops) {
4297 printk(KERN_ERR "kvm: already loaded the other module\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004298 r = -EEXIST;
4299 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004300 }
4301
4302 if (!ops->cpu_has_kvm_support()) {
4303 printk(KERN_ERR "kvm: no hardware support\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004304 r = -EOPNOTSUPP;
4305 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004306 }
4307 if (ops->disabled_by_bios()) {
4308 printk(KERN_ERR "kvm: disabled by bios\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004309 r = -EOPNOTSUPP;
4310 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004311 }
4312
Avi Kivity97db56c2008-01-13 13:23:56 +02004313 r = kvm_mmu_module_init();
4314 if (r)
4315 goto out;
4316
4317 kvm_init_msr_list();
4318
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004319 kvm_x86_ops = ops;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004320 kvm_mmu_set_nonpresent_ptes(0ull, 0ull);
Sheng Yang7b523452008-04-25 21:13:50 +08004321 kvm_mmu_set_base_ptes(PT_PRESENT_MASK);
4322 kvm_mmu_set_mask_ptes(PT_USER_MASK, PT_ACCESSED_MASK,
Sheng Yang4b12f0d2009-04-27 20:35:42 +08004323 PT_DIRTY_MASK, PT64_NX_MASK, 0);
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004324
Zachary Amsdenb820cc02009-09-29 11:38:34 -10004325 kvm_timer_init();
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004326
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004327 perf_register_guest_info_callbacks(&kvm_guest_cbs);
4328
Dexuan Cui2acf9232010-06-10 11:27:12 +08004329 if (cpu_has_xsave)
4330 host_xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
4331
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004332 return 0;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004333
4334out:
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004335 return r;
Carsten Otte043405e2007-10-10 17:16:19 +02004336}
Hollis Blanchard8776e512007-10-31 17:24:24 -05004337
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004338void kvm_arch_exit(void)
4339{
Zhang, Yanminff9d07a2010-04-19 13:32:45 +08004340 perf_unregister_guest_info_callbacks(&kvm_guest_cbs);
4341
Jan Kiszka888d2562009-04-17 19:24:58 +02004342 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
4343 cpufreq_unregister_notifier(&kvmclock_cpufreq_notifier_block,
4344 CPUFREQ_TRANSITION_NOTIFIER);
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004345 kvm_x86_ops = NULL;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08004346 kvm_mmu_module_exit();
4347}
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08004348
Hollis Blanchard8776e512007-10-31 17:24:24 -05004349int kvm_emulate_halt(struct kvm_vcpu *vcpu)
4350{
4351 ++vcpu->stat.halt_exits;
4352 if (irqchip_in_kernel(vcpu->kvm)) {
Avi Kivitya4535292008-04-13 17:54:35 +03004353 vcpu->arch.mp_state = KVM_MP_STATE_HALTED;
Hollis Blanchard8776e512007-10-31 17:24:24 -05004354 return 1;
4355 } else {
4356 vcpu->run->exit_reason = KVM_EXIT_HLT;
4357 return 0;
4358 }
4359}
4360EXPORT_SYMBOL_GPL(kvm_emulate_halt);
4361
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05004362static inline gpa_t hc_gpa(struct kvm_vcpu *vcpu, unsigned long a0,
4363 unsigned long a1)
4364{
4365 if (is_long_mode(vcpu))
4366 return a0;
4367 else
4368 return a0 | ((gpa_t)a1 << 32);
4369}
4370
Gleb Natapov55cd8e52010-01-17 15:51:22 +02004371int kvm_hv_hypercall(struct kvm_vcpu *vcpu)
4372{
4373 u64 param, ingpa, outgpa, ret;
4374 uint16_t code, rep_idx, rep_cnt, res = HV_STATUS_SUCCESS, rep_done = 0;
4375 bool fast, longmode;
4376 int cs_db, cs_l;
4377
4378 /*
4379 * hypercall generates UD from non zero cpl and real mode
4380 * per HYPER-V spec
4381 */
Avi Kivity3eeb3282010-01-21 15:31:48 +02004382 if (kvm_x86_ops->get_cpl(vcpu) != 0 || !is_protmode(vcpu)) {
Gleb Natapov55cd8e52010-01-17 15:51:22 +02004383 kvm_queue_exception(vcpu, UD_VECTOR);
4384 return 0;
4385 }
4386
4387 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
4388 longmode = is_long_mode(vcpu) && cs_l == 1;
4389
4390 if (!longmode) {
Gleb Natapovccd46932010-01-19 15:06:38 +02004391 param = ((u64)kvm_register_read(vcpu, VCPU_REGS_RDX) << 32) |
4392 (kvm_register_read(vcpu, VCPU_REGS_RAX) & 0xffffffff);
4393 ingpa = ((u64)kvm_register_read(vcpu, VCPU_REGS_RBX) << 32) |
4394 (kvm_register_read(vcpu, VCPU_REGS_RCX) & 0xffffffff);
4395 outgpa = ((u64)kvm_register_read(vcpu, VCPU_REGS_RDI) << 32) |
4396 (kvm_register_read(vcpu, VCPU_REGS_RSI) & 0xffffffff);
Gleb Natapov55cd8e52010-01-17 15:51:22 +02004397 }
4398#ifdef CONFIG_X86_64
4399 else {
4400 param = kvm_register_read(vcpu, VCPU_REGS_RCX);
4401 ingpa = kvm_register_read(vcpu, VCPU_REGS_RDX);
4402 outgpa = kvm_register_read(vcpu, VCPU_REGS_R8);
4403 }
4404#endif
4405
4406 code = param & 0xffff;
4407 fast = (param >> 16) & 0x1;
4408 rep_cnt = (param >> 32) & 0xfff;
4409 rep_idx = (param >> 48) & 0xfff;
4410
4411 trace_kvm_hv_hypercall(code, fast, rep_cnt, rep_idx, ingpa, outgpa);
4412
Gleb Natapovc25bc162010-01-17 15:51:24 +02004413 switch (code) {
4414 case HV_X64_HV_NOTIFY_LONG_SPIN_WAIT:
4415 kvm_vcpu_on_spin(vcpu);
4416 break;
4417 default:
4418 res = HV_STATUS_INVALID_HYPERCALL_CODE;
4419 break;
4420 }
Gleb Natapov55cd8e52010-01-17 15:51:22 +02004421
4422 ret = res | (((u64)rep_done & 0xfff) << 32);
4423 if (longmode) {
4424 kvm_register_write(vcpu, VCPU_REGS_RAX, ret);
4425 } else {
4426 kvm_register_write(vcpu, VCPU_REGS_RDX, ret >> 32);
4427 kvm_register_write(vcpu, VCPU_REGS_RAX, ret & 0xffffffff);
4428 }
4429
4430 return 1;
4431}
4432
Hollis Blanchard8776e512007-10-31 17:24:24 -05004433int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
4434{
4435 unsigned long nr, a0, a1, a2, a3, ret;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05004436 int r = 1;
Hollis Blanchard8776e512007-10-31 17:24:24 -05004437
Gleb Natapov55cd8e52010-01-17 15:51:22 +02004438 if (kvm_hv_hypercall_enabled(vcpu->kvm))
4439 return kvm_hv_hypercall(vcpu);
4440
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004441 nr = kvm_register_read(vcpu, VCPU_REGS_RAX);
4442 a0 = kvm_register_read(vcpu, VCPU_REGS_RBX);
4443 a1 = kvm_register_read(vcpu, VCPU_REGS_RCX);
4444 a2 = kvm_register_read(vcpu, VCPU_REGS_RDX);
4445 a3 = kvm_register_read(vcpu, VCPU_REGS_RSI);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004446
Marcelo Tosatti229456f2009-06-17 09:22:14 -03004447 trace_kvm_hypercall(nr, a0, a1, a2, a3);
Feng (Eric) Liu2714d1d2008-04-10 15:31:10 -04004448
Hollis Blanchard8776e512007-10-31 17:24:24 -05004449 if (!is_long_mode(vcpu)) {
4450 nr &= 0xFFFFFFFF;
4451 a0 &= 0xFFFFFFFF;
4452 a1 &= 0xFFFFFFFF;
4453 a2 &= 0xFFFFFFFF;
4454 a3 &= 0xFFFFFFFF;
4455 }
4456
Jan Kiszka07708c42009-08-03 18:43:28 +02004457 if (kvm_x86_ops->get_cpl(vcpu) != 0) {
4458 ret = -KVM_EPERM;
4459 goto out;
4460 }
4461
Hollis Blanchard8776e512007-10-31 17:24:24 -05004462 switch (nr) {
Avi Kivityb93463a2007-10-25 16:52:32 +02004463 case KVM_HC_VAPIC_POLL_IRQ:
4464 ret = 0;
4465 break;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05004466 case KVM_HC_MMU_OP:
4467 r = kvm_pv_mmu_op(vcpu, a0, hc_gpa(vcpu, a1, a2), &ret);
4468 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05004469 default:
4470 ret = -KVM_ENOSYS;
4471 break;
4472 }
Jan Kiszka07708c42009-08-03 18:43:28 +02004473out:
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004474 kvm_register_write(vcpu, VCPU_REGS_RAX, ret);
Amit Shahf11c3a82008-02-21 01:00:30 +05304475 ++vcpu->stat.hypercalls;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05004476 return r;
Hollis Blanchard8776e512007-10-31 17:24:24 -05004477}
4478EXPORT_SYMBOL_GPL(kvm_emulate_hypercall);
4479
4480int kvm_fix_hypercall(struct kvm_vcpu *vcpu)
4481{
4482 char instruction[3];
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004483 unsigned long rip = kvm_rip_read(vcpu);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004484
Hollis Blanchard8776e512007-10-31 17:24:24 -05004485 /*
4486 * Blow out the MMU to ensure that no other VCPU has an active mapping
4487 * to ensure that the updated hypercall appears atomically across all
4488 * VCPUs.
4489 */
4490 kvm_mmu_zap_all(vcpu->kvm);
4491
Hollis Blanchard8776e512007-10-31 17:24:24 -05004492 kvm_x86_ops->patch_hypercall(vcpu, instruction);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004493
Gleb Natapov8fe681e2010-04-28 19:15:37 +03004494 return emulator_write_emulated(rip, instruction, 3, NULL, vcpu);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004495}
4496
Hollis Blanchard8776e512007-10-31 17:24:24 -05004497void realmode_lgdt(struct kvm_vcpu *vcpu, u16 limit, unsigned long base)
4498{
Gleb Natapov89a27f42010-02-16 10:51:48 +02004499 struct desc_ptr dt = { limit, base };
Hollis Blanchard8776e512007-10-31 17:24:24 -05004500
4501 kvm_x86_ops->set_gdt(vcpu, &dt);
4502}
4503
4504void realmode_lidt(struct kvm_vcpu *vcpu, u16 limit, unsigned long base)
4505{
Gleb Natapov89a27f42010-02-16 10:51:48 +02004506 struct desc_ptr dt = { limit, base };
Hollis Blanchard8776e512007-10-31 17:24:24 -05004507
4508 kvm_x86_ops->set_idt(vcpu, &dt);
4509}
4510
Dan Kenigsberg07716712007-11-21 17:10:04 +02004511static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
4512{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004513 struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
4514 int j, nent = vcpu->arch.cpuid_nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02004515
4516 e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
4517 /* when no next entry is found, the current entry[i] is reselected */
Nitin A Kamble0fdf8e52008-11-05 15:56:21 -08004518 for (j = i + 1; ; j = (j + 1) % nent) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004519 struct kvm_cpuid_entry2 *ej = &vcpu->arch.cpuid_entries[j];
Dan Kenigsberg07716712007-11-21 17:10:04 +02004520 if (ej->function == e->function) {
4521 ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
4522 return j;
4523 }
4524 }
4525 return 0; /* silence gcc, even though control never reaches here */
4526}
4527
4528/* find an entry with matching function, matching index (if needed), and that
4529 * should be read next (if it's stateful) */
4530static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
4531 u32 function, u32 index)
4532{
4533 if (e->function != function)
4534 return 0;
4535 if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
4536 return 0;
4537 if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
Amit Shah19355472009-01-14 16:56:00 +00004538 !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
Dan Kenigsberg07716712007-11-21 17:10:04 +02004539 return 0;
4540 return 1;
4541}
4542
Alexander Grafd8017472008-11-25 20:17:11 +01004543struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
4544 u32 function, u32 index)
Hollis Blanchard8776e512007-10-31 17:24:24 -05004545{
4546 int i;
Alexander Grafd8017472008-11-25 20:17:11 +01004547 struct kvm_cpuid_entry2 *best = NULL;
Hollis Blanchard8776e512007-10-31 17:24:24 -05004548
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004549 for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
Alexander Grafd8017472008-11-25 20:17:11 +01004550 struct kvm_cpuid_entry2 *e;
4551
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004552 e = &vcpu->arch.cpuid_entries[i];
Dan Kenigsberg07716712007-11-21 17:10:04 +02004553 if (is_matching_cpuid_entry(e, function, index)) {
4554 if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
4555 move_to_next_stateful_cpuid_entry(vcpu, i);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004556 best = e;
4557 break;
4558 }
4559 /*
4560 * Both basic or both extended?
4561 */
4562 if (((e->function ^ function) & 0x80000000) == 0)
4563 if (!best || e->function > best->function)
4564 best = e;
4565 }
Alexander Grafd8017472008-11-25 20:17:11 +01004566 return best;
4567}
Sheng Yang0e851882009-12-18 16:48:46 +08004568EXPORT_SYMBOL_GPL(kvm_find_cpuid_entry);
Alexander Grafd8017472008-11-25 20:17:11 +01004569
Dong, Eddie82725b22009-03-30 16:21:08 +08004570int cpuid_maxphyaddr(struct kvm_vcpu *vcpu)
4571{
4572 struct kvm_cpuid_entry2 *best;
4573
Avi Kivityf7a71192010-04-11 15:33:32 +03004574 best = kvm_find_cpuid_entry(vcpu, 0x80000000, 0);
4575 if (!best || best->eax < 0x80000008)
4576 goto not_found;
Dong, Eddie82725b22009-03-30 16:21:08 +08004577 best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
4578 if (best)
4579 return best->eax & 0xff;
Avi Kivityf7a71192010-04-11 15:33:32 +03004580not_found:
Dong, Eddie82725b22009-03-30 16:21:08 +08004581 return 36;
4582}
4583
Alexander Grafd8017472008-11-25 20:17:11 +01004584void kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
4585{
4586 u32 function, index;
4587 struct kvm_cpuid_entry2 *best;
4588
4589 function = kvm_register_read(vcpu, VCPU_REGS_RAX);
4590 index = kvm_register_read(vcpu, VCPU_REGS_RCX);
4591 kvm_register_write(vcpu, VCPU_REGS_RAX, 0);
4592 kvm_register_write(vcpu, VCPU_REGS_RBX, 0);
4593 kvm_register_write(vcpu, VCPU_REGS_RCX, 0);
4594 kvm_register_write(vcpu, VCPU_REGS_RDX, 0);
4595 best = kvm_find_cpuid_entry(vcpu, function, index);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004596 if (best) {
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004597 kvm_register_write(vcpu, VCPU_REGS_RAX, best->eax);
4598 kvm_register_write(vcpu, VCPU_REGS_RBX, best->ebx);
4599 kvm_register_write(vcpu, VCPU_REGS_RCX, best->ecx);
4600 kvm_register_write(vcpu, VCPU_REGS_RDX, best->edx);
Hollis Blanchard8776e512007-10-31 17:24:24 -05004601 }
Hollis Blanchard8776e512007-10-31 17:24:24 -05004602 kvm_x86_ops->skip_emulated_instruction(vcpu);
Marcelo Tosatti229456f2009-06-17 09:22:14 -03004603 trace_kvm_cpuid(function,
4604 kvm_register_read(vcpu, VCPU_REGS_RAX),
4605 kvm_register_read(vcpu, VCPU_REGS_RBX),
4606 kvm_register_read(vcpu, VCPU_REGS_RCX),
4607 kvm_register_read(vcpu, VCPU_REGS_RDX));
Hollis Blanchard8776e512007-10-31 17:24:24 -05004608}
4609EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
Hollis Blanchardd0752062007-10-31 17:24:25 -05004610
4611/*
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004612 * Check if userspace requested an interrupt window, and that the
4613 * interrupt window is open.
4614 *
4615 * No need to exit to userspace if we already have an interrupt queued.
4616 */
Avi Kivity851ba692009-08-24 11:10:17 +03004617static int dm_request_for_irq_injection(struct kvm_vcpu *vcpu)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004618{
Gleb Natapov80618232009-04-21 17:44:56 +03004619 return (!irqchip_in_kernel(vcpu->kvm) && !kvm_cpu_has_interrupt(vcpu) &&
Avi Kivity851ba692009-08-24 11:10:17 +03004620 vcpu->run->request_interrupt_window &&
Gleb Natapov5df56642009-04-21 17:44:59 +03004621 kvm_arch_interrupt_allowed(vcpu));
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004622}
4623
Avi Kivity851ba692009-08-24 11:10:17 +03004624static void post_kvm_run_save(struct kvm_vcpu *vcpu)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004625{
Avi Kivity851ba692009-08-24 11:10:17 +03004626 struct kvm_run *kvm_run = vcpu->run;
4627
Jan Kiszka91586a32009-10-05 13:07:21 +02004628 kvm_run->if_flag = (kvm_get_rflags(vcpu) & X86_EFLAGS_IF) != 0;
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02004629 kvm_run->cr8 = kvm_get_cr8(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004630 kvm_run->apic_base = kvm_get_apic_base(vcpu);
Jan Kiszka45312202008-12-11 16:54:54 +01004631 if (irqchip_in_kernel(vcpu->kvm))
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004632 kvm_run->ready_for_interrupt_injection = 1;
Jan Kiszka45312202008-12-11 16:54:54 +01004633 else
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004634 kvm_run->ready_for_interrupt_injection =
Gleb Natapovfa9726b2009-05-11 13:35:47 +03004635 kvm_arch_interrupt_allowed(vcpu) &&
4636 !kvm_cpu_has_interrupt(vcpu) &&
4637 !kvm_event_needs_reinjection(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004638}
4639
Avi Kivityb93463a2007-10-25 16:52:32 +02004640static void vapic_enter(struct kvm_vcpu *vcpu)
4641{
4642 struct kvm_lapic *apic = vcpu->arch.apic;
4643 struct page *page;
4644
4645 if (!apic || !apic->vapic_addr)
4646 return;
4647
4648 page = gfn_to_page(vcpu->kvm, apic->vapic_addr >> PAGE_SHIFT);
Izik Eidus72dc67a2008-02-10 18:04:15 +02004649
4650 vcpu->arch.apic->vapic_page = page;
Avi Kivityb93463a2007-10-25 16:52:32 +02004651}
4652
4653static void vapic_exit(struct kvm_vcpu *vcpu)
4654{
4655 struct kvm_lapic *apic = vcpu->arch.apic;
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004656 int idx;
Avi Kivityb93463a2007-10-25 16:52:32 +02004657
4658 if (!apic || !apic->vapic_addr)
4659 return;
4660
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004661 idx = srcu_read_lock(&vcpu->kvm->srcu);
Avi Kivityb93463a2007-10-25 16:52:32 +02004662 kvm_release_page_dirty(apic->vapic_page);
4663 mark_page_dirty(vcpu->kvm, apic->vapic_addr >> PAGE_SHIFT);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004664 srcu_read_unlock(&vcpu->kvm->srcu, idx);
Avi Kivityb93463a2007-10-25 16:52:32 +02004665}
4666
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004667static void update_cr8_intercept(struct kvm_vcpu *vcpu)
4668{
4669 int max_irr, tpr;
4670
4671 if (!kvm_x86_ops->update_cr8_intercept)
4672 return;
4673
Avi Kivity88c808f2009-08-17 22:49:40 +03004674 if (!vcpu->arch.apic)
4675 return;
4676
Gleb Natapov8db3baa2009-05-11 13:35:54 +03004677 if (!vcpu->arch.apic->vapic_addr)
4678 max_irr = kvm_lapic_find_highest_irr(vcpu);
4679 else
4680 max_irr = -1;
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004681
4682 if (max_irr != -1)
4683 max_irr >>= 4;
4684
4685 tpr = kvm_lapic_get_cr8(vcpu);
4686
4687 kvm_x86_ops->update_cr8_intercept(vcpu, tpr, max_irr);
4688}
4689
Avi Kivity851ba692009-08-24 11:10:17 +03004690static void inject_pending_event(struct kvm_vcpu *vcpu)
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004691{
4692 /* try to reinject previous events if any */
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03004693 if (vcpu->arch.exception.pending) {
Avi Kivity5c1c85d02010-03-11 13:01:59 +02004694 trace_kvm_inj_exception(vcpu->arch.exception.nr,
4695 vcpu->arch.exception.has_error_code,
4696 vcpu->arch.exception.error_code);
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03004697 kvm_x86_ops->queue_exception(vcpu, vcpu->arch.exception.nr,
4698 vcpu->arch.exception.has_error_code,
Joerg Roedelce7ddec2010-04-22 12:33:13 +02004699 vcpu->arch.exception.error_code,
4700 vcpu->arch.exception.reinject);
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03004701 return;
4702 }
4703
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004704 if (vcpu->arch.nmi_injected) {
4705 kvm_x86_ops->set_nmi(vcpu);
4706 return;
4707 }
4708
4709 if (vcpu->arch.interrupt.pending) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03004710 kvm_x86_ops->set_irq(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004711 return;
4712 }
4713
4714 /* try to inject new event if pending */
4715 if (vcpu->arch.nmi_pending) {
4716 if (kvm_x86_ops->nmi_allowed(vcpu)) {
4717 vcpu->arch.nmi_pending = false;
4718 vcpu->arch.nmi_injected = true;
4719 kvm_x86_ops->set_nmi(vcpu);
4720 }
4721 } else if (kvm_cpu_has_interrupt(vcpu)) {
4722 if (kvm_x86_ops->interrupt_allowed(vcpu)) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03004723 kvm_queue_interrupt(vcpu, kvm_cpu_get_interrupt(vcpu),
4724 false);
4725 kvm_x86_ops->set_irq(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004726 }
4727 }
4728}
4729
Dexuan Cui2acf9232010-06-10 11:27:12 +08004730static void kvm_load_guest_xcr0(struct kvm_vcpu *vcpu)
4731{
4732 if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE) &&
4733 !vcpu->guest_xcr0_loaded) {
4734 /* kvm_set_xcr() also depends on this */
4735 xsetbv(XCR_XFEATURE_ENABLED_MASK, vcpu->arch.xcr0);
4736 vcpu->guest_xcr0_loaded = 1;
4737 }
4738}
4739
4740static void kvm_put_guest_xcr0(struct kvm_vcpu *vcpu)
4741{
4742 if (vcpu->guest_xcr0_loaded) {
4743 if (vcpu->arch.xcr0 != host_xcr0)
4744 xsetbv(XCR_XFEATURE_ENABLED_MASK, host_xcr0);
4745 vcpu->guest_xcr0_loaded = 0;
4746 }
4747}
4748
Avi Kivity851ba692009-08-24 11:10:17 +03004749static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004750{
4751 int r;
Gleb Natapov6a8b1d12009-05-11 13:35:51 +03004752 bool req_int_win = !irqchip_in_kernel(vcpu->kvm) &&
Avi Kivity851ba692009-08-24 11:10:17 +03004753 vcpu->run->request_interrupt_window;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004754
Marcelo Tosatti2e53d632008-02-20 14:47:24 -05004755 if (vcpu->requests)
4756 if (test_and_clear_bit(KVM_REQ_MMU_RELOAD, &vcpu->requests))
4757 kvm_mmu_unload(vcpu);
4758
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004759 r = kvm_mmu_reload(vcpu);
4760 if (unlikely(r))
4761 goto out;
4762
Avi Kivity2f52d582008-01-16 12:49:30 +02004763 if (vcpu->requests) {
4764 if (test_and_clear_bit(KVM_REQ_MIGRATE_TIMER, &vcpu->requests))
Marcelo Tosatti2f599712008-05-27 12:10:20 -03004765 __kvm_migrate_timers(vcpu);
Gerd Hoffmannc8076602009-02-04 17:52:04 +01004766 if (test_and_clear_bit(KVM_REQ_KVMCLOCK_UPDATE, &vcpu->requests))
4767 kvm_write_guest_time(vcpu);
Marcelo Tosatti4731d4c2008-09-23 13:18:39 -03004768 if (test_and_clear_bit(KVM_REQ_MMU_SYNC, &vcpu->requests))
4769 kvm_mmu_sync_roots(vcpu);
Marcelo Tosattid4acf7e2008-06-06 16:37:35 -03004770 if (test_and_clear_bit(KVM_REQ_TLB_FLUSH, &vcpu->requests))
4771 kvm_x86_ops->tlb_flush(vcpu);
Avi Kivityb93463a2007-10-25 16:52:32 +02004772 if (test_and_clear_bit(KVM_REQ_REPORT_TPR_ACCESS,
4773 &vcpu->requests)) {
Avi Kivity851ba692009-08-24 11:10:17 +03004774 vcpu->run->exit_reason = KVM_EXIT_TPR_ACCESS;
Avi Kivityb93463a2007-10-25 16:52:32 +02004775 r = 0;
4776 goto out;
4777 }
Joerg Roedel71c4dfa2008-02-26 16:49:16 +01004778 if (test_and_clear_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests)) {
Avi Kivity851ba692009-08-24 11:10:17 +03004779 vcpu->run->exit_reason = KVM_EXIT_SHUTDOWN;
Joerg Roedel71c4dfa2008-02-26 16:49:16 +01004780 r = 0;
4781 goto out;
4782 }
Avi Kivity02daab22009-12-30 12:40:26 +02004783 if (test_and_clear_bit(KVM_REQ_DEACTIVATE_FPU, &vcpu->requests)) {
4784 vcpu->fpu_active = 0;
4785 kvm_x86_ops->fpu_deactivate(vcpu);
4786 }
Avi Kivity2f52d582008-01-16 12:49:30 +02004787 }
Avi Kivityb93463a2007-10-25 16:52:32 +02004788
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004789 preempt_disable();
4790
4791 kvm_x86_ops->prepare_guest_switch(vcpu);
Avi Kivity2608d7a2010-01-21 15:31:45 +02004792 if (vcpu->fpu_active)
4793 kvm_load_guest_fpu(vcpu);
Dexuan Cui2acf9232010-06-10 11:27:12 +08004794 kvm_load_guest_xcr0(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004795
Avi Kivityd94e1dc2010-05-03 16:54:48 +03004796 atomic_set(&vcpu->guest_mode, 1);
4797 smp_wmb();
4798
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004799 local_irq_disable();
4800
Avi Kivityd94e1dc2010-05-03 16:54:48 +03004801 if (!atomic_read(&vcpu->guest_mode) || vcpu->requests
4802 || need_resched() || signal_pending(current)) {
4803 atomic_set(&vcpu->guest_mode, 0);
4804 smp_wmb();
Avi Kivity6c1428012008-01-15 18:27:32 +02004805 local_irq_enable();
4806 preempt_enable();
4807 r = 1;
4808 goto out;
4809 }
4810
Avi Kivity851ba692009-08-24 11:10:17 +03004811 inject_pending_event(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004812
Gleb Natapov6a8b1d12009-05-11 13:35:51 +03004813 /* enable NMI/IRQ window open exits if needed */
4814 if (vcpu->arch.nmi_pending)
4815 kvm_x86_ops->enable_nmi_window(vcpu);
4816 else if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
4817 kvm_x86_ops->enable_irq_window(vcpu);
4818
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004819 if (kvm_lapic_enabled(vcpu)) {
Gleb Natapov8db3baa2009-05-11 13:35:54 +03004820 update_cr8_intercept(vcpu);
4821 kvm_lapic_sync_to_vapic(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03004822 }
Avi Kivityb93463a2007-10-25 16:52:32 +02004823
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004824 srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
Marcelo Tosatti3200f402008-03-29 20:17:59 -03004825
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004826 kvm_guest_enter();
4827
Jan Kiszka42dbaa52008-12-15 13:52:10 +01004828 if (unlikely(vcpu->arch.switch_db_regs)) {
Jan Kiszka42dbaa52008-12-15 13:52:10 +01004829 set_debugreg(0, 7);
4830 set_debugreg(vcpu->arch.eff_db[0], 0);
4831 set_debugreg(vcpu->arch.eff_db[1], 1);
4832 set_debugreg(vcpu->arch.eff_db[2], 2);
4833 set_debugreg(vcpu->arch.eff_db[3], 3);
4834 }
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004835
Marcelo Tosatti229456f2009-06-17 09:22:14 -03004836 trace_kvm_entry(vcpu->vcpu_id);
Avi Kivity851ba692009-08-24 11:10:17 +03004837 kvm_x86_ops->run(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004838
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +02004839 /*
4840 * If the guest has used debug registers, at least dr7
4841 * will be disabled while returning to the host.
4842 * If we don't have active breakpoints in the host, we don't
4843 * care about the messed up debug address registers. But if
4844 * we have some of them active, restore the old state.
4845 */
Frederic Weisbecker59d8eb52009-11-10 11:03:12 +01004846 if (hw_breakpoint_active())
Frederic Weisbecker24f1e32c2009-09-09 19:22:48 +02004847 hw_breakpoint_restore();
Jan Kiszka42dbaa52008-12-15 13:52:10 +01004848
Avi Kivityd94e1dc2010-05-03 16:54:48 +03004849 atomic_set(&vcpu->guest_mode, 0);
4850 smp_wmb();
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004851 local_irq_enable();
4852
4853 ++vcpu->stat.exits;
4854
4855 /*
4856 * We must have an instruction between local_irq_enable() and
4857 * kvm_guest_exit(), so the timer interrupt isn't delayed by
4858 * the interrupt shadow. The stat.exits increment will do nicely.
4859 * But we need to prevent reordering, hence this barrier():
4860 */
4861 barrier();
4862
4863 kvm_guest_exit();
4864
4865 preempt_enable();
4866
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004867 vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
Marcelo Tosatti3200f402008-03-29 20:17:59 -03004868
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004869 /*
4870 * Profile KVM exit RIPs:
4871 */
4872 if (unlikely(prof_on == KVM_PROFILING)) {
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004873 unsigned long rip = kvm_rip_read(vcpu);
4874 profile_hit(KVM_PROFILING, (void *)rip);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004875 }
4876
Avi Kivity298101d2007-11-25 13:41:11 +02004877
Avi Kivityb93463a2007-10-25 16:52:32 +02004878 kvm_lapic_sync_from_vapic(vcpu);
4879
Avi Kivity851ba692009-08-24 11:10:17 +03004880 r = kvm_x86_ops->handle_exit(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004881out:
Marcelo Tosattid7690172008-09-08 15:23:48 -03004882 return r;
4883}
4884
Gleb Natapov09cec752009-03-23 15:11:44 +02004885
Avi Kivity851ba692009-08-24 11:10:17 +03004886static int __vcpu_run(struct kvm_vcpu *vcpu)
Marcelo Tosattid7690172008-09-08 15:23:48 -03004887{
4888 int r;
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004889 struct kvm *kvm = vcpu->kvm;
Marcelo Tosattid7690172008-09-08 15:23:48 -03004890
4891 if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_SIPI_RECEIVED)) {
Jan Kiszka1b10bf32008-09-30 10:41:06 +02004892 pr_debug("vcpu %d received sipi with vector # %x\n",
4893 vcpu->vcpu_id, vcpu->arch.sipi_vector);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004894 kvm_lapic_reset(vcpu);
Gleb Natapov5f179282008-10-07 15:42:33 +02004895 r = kvm_arch_vcpu_reset(vcpu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004896 if (r)
4897 return r;
4898 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004899 }
4900
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004901 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004902 vapic_enter(vcpu);
4903
4904 r = 1;
4905 while (r > 0) {
Gleb Natapovaf2152f2008-09-22 14:28:53 +03004906 if (vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE)
Avi Kivity851ba692009-08-24 11:10:17 +03004907 r = vcpu_enter_guest(vcpu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004908 else {
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004909 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004910 kvm_vcpu_block(vcpu);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004911 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004912 if (test_and_clear_bit(KVM_REQ_UNHALT, &vcpu->requests))
Gleb Natapov09cec752009-03-23 15:11:44 +02004913 {
4914 switch(vcpu->arch.mp_state) {
4915 case KVM_MP_STATE_HALTED:
Marcelo Tosattid7690172008-09-08 15:23:48 -03004916 vcpu->arch.mp_state =
Gleb Natapov09cec752009-03-23 15:11:44 +02004917 KVM_MP_STATE_RUNNABLE;
4918 case KVM_MP_STATE_RUNNABLE:
4919 break;
4920 case KVM_MP_STATE_SIPI_RECEIVED:
4921 default:
4922 r = -EINTR;
4923 break;
4924 }
4925 }
Marcelo Tosattid7690172008-09-08 15:23:48 -03004926 }
4927
Gleb Natapov09cec752009-03-23 15:11:44 +02004928 if (r <= 0)
4929 break;
4930
4931 clear_bit(KVM_REQ_PENDING_TIMER, &vcpu->requests);
4932 if (kvm_cpu_has_pending_timer(vcpu))
4933 kvm_inject_pending_timer_irqs(vcpu);
4934
Avi Kivity851ba692009-08-24 11:10:17 +03004935 if (dm_request_for_irq_injection(vcpu)) {
Gleb Natapov09cec752009-03-23 15:11:44 +02004936 r = -EINTR;
Avi Kivity851ba692009-08-24 11:10:17 +03004937 vcpu->run->exit_reason = KVM_EXIT_INTR;
Gleb Natapov09cec752009-03-23 15:11:44 +02004938 ++vcpu->stat.request_irq_exits;
4939 }
4940 if (signal_pending(current)) {
4941 r = -EINTR;
Avi Kivity851ba692009-08-24 11:10:17 +03004942 vcpu->run->exit_reason = KVM_EXIT_INTR;
Gleb Natapov09cec752009-03-23 15:11:44 +02004943 ++vcpu->stat.signal_exits;
4944 }
4945 if (need_resched()) {
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004946 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
Gleb Natapov09cec752009-03-23 15:11:44 +02004947 kvm_resched(vcpu);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004948 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004949 }
4950 }
4951
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004952 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004953
Avi Kivityb93463a2007-10-25 16:52:32 +02004954 vapic_exit(vcpu);
4955
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004956 return r;
4957}
4958
4959int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
4960{
4961 int r;
4962 sigset_t sigsaved;
4963
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004964 if (vcpu->sigset_active)
4965 sigprocmask(SIG_SETMASK, &vcpu->sigset, &sigsaved);
4966
Avi Kivityac9f6dc2008-07-06 15:48:31 +03004967 if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
4968 kvm_vcpu_block(vcpu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03004969 clear_bit(KVM_REQ_UNHALT, &vcpu->requests);
Avi Kivityac9f6dc2008-07-06 15:48:31 +03004970 r = -EAGAIN;
4971 goto out;
4972 }
4973
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004974 /* re-sync apic's tpr */
4975 if (!irqchip_in_kernel(vcpu->kvm))
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02004976 kvm_set_cr8(vcpu, kvm_run->cr8);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004977
Gleb Natapov92bf9742010-03-18 15:20:28 +02004978 if (vcpu->arch.pio.count || vcpu->mmio_needed ||
4979 vcpu->arch.emulate_ctxt.restart) {
4980 if (vcpu->mmio_needed) {
4981 memcpy(vcpu->mmio_data, kvm_run->mmio.data, 8);
4982 vcpu->mmio_read_completed = 1;
4983 vcpu->mmio_needed = 0;
Gleb Natapov79729952010-03-18 15:20:24 +02004984 }
Gleb Natapov7567cae2010-03-09 12:01:10 +02004985 vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
Gleb Natapov5cd21912010-03-18 15:20:26 +02004986 r = emulate_instruction(vcpu, 0, 0, EMULTYPE_NO_DECODE);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02004987 srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
Gleb Natapov6d77dbf2010-05-10 11:16:56 +03004988 if (r != EMULATE_DONE) {
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004989 r = 0;
4990 goto out;
4991 }
4992 }
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004993 if (kvm_run->exit_reason == KVM_EXIT_HYPERCALL)
4994 kvm_register_write(vcpu, VCPU_REGS_RAX,
4995 kvm_run->hypercall.ret);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004996
Avi Kivity851ba692009-08-24 11:10:17 +03004997 r = __vcpu_run(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004998
4999out:
Marcelo Tosattif1d86e42010-05-03 23:04:27 -03005000 post_kvm_run_save(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005001 if (vcpu->sigset_active)
5002 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
5003
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005004 return r;
5005}
5006
5007int kvm_arch_vcpu_ioctl_get_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs)
5008{
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005009 regs->rax = kvm_register_read(vcpu, VCPU_REGS_RAX);
5010 regs->rbx = kvm_register_read(vcpu, VCPU_REGS_RBX);
5011 regs->rcx = kvm_register_read(vcpu, VCPU_REGS_RCX);
5012 regs->rdx = kvm_register_read(vcpu, VCPU_REGS_RDX);
5013 regs->rsi = kvm_register_read(vcpu, VCPU_REGS_RSI);
5014 regs->rdi = kvm_register_read(vcpu, VCPU_REGS_RDI);
5015 regs->rsp = kvm_register_read(vcpu, VCPU_REGS_RSP);
5016 regs->rbp = kvm_register_read(vcpu, VCPU_REGS_RBP);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005017#ifdef CONFIG_X86_64
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005018 regs->r8 = kvm_register_read(vcpu, VCPU_REGS_R8);
5019 regs->r9 = kvm_register_read(vcpu, VCPU_REGS_R9);
5020 regs->r10 = kvm_register_read(vcpu, VCPU_REGS_R10);
5021 regs->r11 = kvm_register_read(vcpu, VCPU_REGS_R11);
5022 regs->r12 = kvm_register_read(vcpu, VCPU_REGS_R12);
5023 regs->r13 = kvm_register_read(vcpu, VCPU_REGS_R13);
5024 regs->r14 = kvm_register_read(vcpu, VCPU_REGS_R14);
5025 regs->r15 = kvm_register_read(vcpu, VCPU_REGS_R15);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005026#endif
5027
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005028 regs->rip = kvm_rip_read(vcpu);
Jan Kiszka91586a32009-10-05 13:07:21 +02005029 regs->rflags = kvm_get_rflags(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005030
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005031 return 0;
5032}
5033
5034int kvm_arch_vcpu_ioctl_set_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs)
5035{
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005036 kvm_register_write(vcpu, VCPU_REGS_RAX, regs->rax);
5037 kvm_register_write(vcpu, VCPU_REGS_RBX, regs->rbx);
5038 kvm_register_write(vcpu, VCPU_REGS_RCX, regs->rcx);
5039 kvm_register_write(vcpu, VCPU_REGS_RDX, regs->rdx);
5040 kvm_register_write(vcpu, VCPU_REGS_RSI, regs->rsi);
5041 kvm_register_write(vcpu, VCPU_REGS_RDI, regs->rdi);
5042 kvm_register_write(vcpu, VCPU_REGS_RSP, regs->rsp);
5043 kvm_register_write(vcpu, VCPU_REGS_RBP, regs->rbp);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005044#ifdef CONFIG_X86_64
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005045 kvm_register_write(vcpu, VCPU_REGS_R8, regs->r8);
5046 kvm_register_write(vcpu, VCPU_REGS_R9, regs->r9);
5047 kvm_register_write(vcpu, VCPU_REGS_R10, regs->r10);
5048 kvm_register_write(vcpu, VCPU_REGS_R11, regs->r11);
5049 kvm_register_write(vcpu, VCPU_REGS_R12, regs->r12);
5050 kvm_register_write(vcpu, VCPU_REGS_R13, regs->r13);
5051 kvm_register_write(vcpu, VCPU_REGS_R14, regs->r14);
5052 kvm_register_write(vcpu, VCPU_REGS_R15, regs->r15);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005053#endif
5054
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03005055 kvm_rip_write(vcpu, regs->rip);
Jan Kiszka91586a32009-10-05 13:07:21 +02005056 kvm_set_rflags(vcpu, regs->rflags);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005057
Jan Kiszkab4f14ab2008-04-30 17:59:04 +02005058 vcpu->arch.exception.pending = false;
5059
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005060 return 0;
5061}
5062
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005063void kvm_get_cs_db_l_bits(struct kvm_vcpu *vcpu, int *db, int *l)
5064{
5065 struct kvm_segment cs;
5066
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02005067 kvm_get_segment(vcpu, &cs, VCPU_SREG_CS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005068 *db = cs.db;
5069 *l = cs.l;
5070}
5071EXPORT_SYMBOL_GPL(kvm_get_cs_db_l_bits);
5072
5073int kvm_arch_vcpu_ioctl_get_sregs(struct kvm_vcpu *vcpu,
5074 struct kvm_sregs *sregs)
5075{
Gleb Natapov89a27f42010-02-16 10:51:48 +02005076 struct desc_ptr dt;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005077
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02005078 kvm_get_segment(vcpu, &sregs->cs, VCPU_SREG_CS);
5079 kvm_get_segment(vcpu, &sregs->ds, VCPU_SREG_DS);
5080 kvm_get_segment(vcpu, &sregs->es, VCPU_SREG_ES);
5081 kvm_get_segment(vcpu, &sregs->fs, VCPU_SREG_FS);
5082 kvm_get_segment(vcpu, &sregs->gs, VCPU_SREG_GS);
5083 kvm_get_segment(vcpu, &sregs->ss, VCPU_SREG_SS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005084
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02005085 kvm_get_segment(vcpu, &sregs->tr, VCPU_SREG_TR);
5086 kvm_get_segment(vcpu, &sregs->ldt, VCPU_SREG_LDTR);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005087
5088 kvm_x86_ops->get_idt(vcpu, &dt);
Gleb Natapov89a27f42010-02-16 10:51:48 +02005089 sregs->idt.limit = dt.size;
5090 sregs->idt.base = dt.address;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005091 kvm_x86_ops->get_gdt(vcpu, &dt);
Gleb Natapov89a27f42010-02-16 10:51:48 +02005092 sregs->gdt.limit = dt.size;
5093 sregs->gdt.base = dt.address;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005094
Avi Kivity4d4ec082009-12-29 18:07:30 +02005095 sregs->cr0 = kvm_read_cr0(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005096 sregs->cr2 = vcpu->arch.cr2;
5097 sregs->cr3 = vcpu->arch.cr3;
Avi Kivityfc78f512009-12-07 12:16:48 +02005098 sregs->cr4 = kvm_read_cr4(vcpu);
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02005099 sregs->cr8 = kvm_get_cr8(vcpu);
Avi Kivityf6801df2010-01-21 15:31:50 +02005100 sregs->efer = vcpu->arch.efer;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005101 sregs->apic_base = kvm_get_apic_base(vcpu);
5102
Gleb Natapov923c61b2009-05-11 13:35:48 +03005103 memset(sregs->interrupt_bitmap, 0, sizeof sregs->interrupt_bitmap);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005104
Gleb Natapov36752c92009-05-11 13:35:53 +03005105 if (vcpu->arch.interrupt.pending && !vcpu->arch.interrupt.soft)
Gleb Natapov14d0bc12009-04-21 17:45:11 +03005106 set_bit(vcpu->arch.interrupt.nr,
5107 (unsigned long *)sregs->interrupt_bitmap);
Gleb Natapov16d7a192009-04-21 17:45:10 +03005108
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005109 return 0;
5110}
5111
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03005112int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
5113 struct kvm_mp_state *mp_state)
5114{
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03005115 mp_state->mp_state = vcpu->arch.mp_state;
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03005116 return 0;
5117}
5118
5119int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
5120 struct kvm_mp_state *mp_state)
5121{
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03005122 vcpu->arch.mp_state = mp_state->mp_state;
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03005123 return 0;
5124}
5125
Jan Kiszkae269fb22010-04-14 15:51:09 +02005126int kvm_task_switch(struct kvm_vcpu *vcpu, u16 tss_selector, int reason,
5127 bool has_error_code, u32 error_code)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005128{
Gleb Natapov4d2179e2010-04-28 19:15:42 +03005129 struct decode_cache *c = &vcpu->arch.emulate_ctxt.decode;
Gleb Natapovceffb452010-03-18 15:20:19 +02005130 int cs_db, cs_l, ret;
5131 cache_all_regs(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005132
Gleb Natapovceffb452010-03-18 15:20:19 +02005133 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
Izik Eidus37817f22008-03-24 23:14:53 +02005134
Gleb Natapovceffb452010-03-18 15:20:19 +02005135 vcpu->arch.emulate_ctxt.vcpu = vcpu;
5136 vcpu->arch.emulate_ctxt.eflags = kvm_x86_ops->get_rflags(vcpu);
5137 vcpu->arch.emulate_ctxt.eip = kvm_rip_read(vcpu);
5138 vcpu->arch.emulate_ctxt.mode =
5139 (!is_protmode(vcpu)) ? X86EMUL_MODE_REAL :
5140 (vcpu->arch.emulate_ctxt.eflags & X86_EFLAGS_VM)
5141 ? X86EMUL_MODE_VM86 : cs_l
5142 ? X86EMUL_MODE_PROT64 : cs_db
5143 ? X86EMUL_MODE_PROT32 : X86EMUL_MODE_PROT16;
Gleb Natapov4d2179e2010-04-28 19:15:42 +03005144 memset(c, 0, sizeof(struct decode_cache));
5145 memcpy(c->regs, vcpu->arch.regs, sizeof c->regs);
Izik Eidus37817f22008-03-24 23:14:53 +02005146
Gleb Natapovceffb452010-03-18 15:20:19 +02005147 ret = emulator_task_switch(&vcpu->arch.emulate_ctxt, &emulate_ops,
Jan Kiszkae269fb22010-04-14 15:51:09 +02005148 tss_selector, reason, has_error_code,
5149 error_code);
Izik Eidus37817f22008-03-24 23:14:53 +02005150
Gleb Natapovc6975182010-02-18 12:15:01 +02005151 if (ret)
Gleb Natapov19d04432010-04-15 12:29:50 +03005152 return EMULATE_FAIL;
Gleb Natapovc6975182010-02-18 12:15:01 +02005153
Gleb Natapov4d2179e2010-04-28 19:15:42 +03005154 memcpy(vcpu->arch.regs, c->regs, sizeof c->regs);
Gleb Natapov95c55882010-04-28 19:15:39 +03005155 kvm_rip_write(vcpu, vcpu->arch.emulate_ctxt.eip);
Gleb Natapov19d04432010-04-15 12:29:50 +03005156 kvm_x86_ops->set_rflags(vcpu, vcpu->arch.emulate_ctxt.eflags);
5157 return EMULATE_DONE;
Izik Eidus37817f22008-03-24 23:14:53 +02005158}
5159EXPORT_SYMBOL_GPL(kvm_task_switch);
5160
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005161int kvm_arch_vcpu_ioctl_set_sregs(struct kvm_vcpu *vcpu,
5162 struct kvm_sregs *sregs)
5163{
5164 int mmu_reset_needed = 0;
Gleb Natapov923c61b2009-05-11 13:35:48 +03005165 int pending_vec, max_bits;
Gleb Natapov89a27f42010-02-16 10:51:48 +02005166 struct desc_ptr dt;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005167
Gleb Natapov89a27f42010-02-16 10:51:48 +02005168 dt.size = sregs->idt.limit;
5169 dt.address = sregs->idt.base;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005170 kvm_x86_ops->set_idt(vcpu, &dt);
Gleb Natapov89a27f42010-02-16 10:51:48 +02005171 dt.size = sregs->gdt.limit;
5172 dt.address = sregs->gdt.base;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005173 kvm_x86_ops->set_gdt(vcpu, &dt);
5174
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005175 vcpu->arch.cr2 = sregs->cr2;
5176 mmu_reset_needed |= vcpu->arch.cr3 != sregs->cr3;
Jan Kiszkadc7e7952009-07-01 20:52:03 +02005177 vcpu->arch.cr3 = sregs->cr3;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005178
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02005179 kvm_set_cr8(vcpu, sregs->cr8);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005180
Avi Kivityf6801df2010-01-21 15:31:50 +02005181 mmu_reset_needed |= vcpu->arch.efer != sregs->efer;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005182 kvm_x86_ops->set_efer(vcpu, sregs->efer);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005183 kvm_set_apic_base(vcpu, sregs->apic_base);
5184
Avi Kivity4d4ec082009-12-29 18:07:30 +02005185 mmu_reset_needed |= kvm_read_cr0(vcpu) != sregs->cr0;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005186 kvm_x86_ops->set_cr0(vcpu, sregs->cr0);
Paul Knowlesd7306162008-02-06 11:02:35 +00005187 vcpu->arch.cr0 = sregs->cr0;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005188
Avi Kivityfc78f512009-12-07 12:16:48 +02005189 mmu_reset_needed |= kvm_read_cr4(vcpu) != sregs->cr4;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005190 kvm_x86_ops->set_cr4(vcpu, sregs->cr4);
Marcelo Tosatti7c93be442009-10-26 16:48:33 -02005191 if (!is_long_mode(vcpu) && is_pae(vcpu)) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005192 load_pdptrs(vcpu, vcpu->arch.cr3);
Marcelo Tosatti7c93be442009-10-26 16:48:33 -02005193 mmu_reset_needed = 1;
5194 }
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005195
5196 if (mmu_reset_needed)
5197 kvm_mmu_reset_context(vcpu);
5198
Gleb Natapov923c61b2009-05-11 13:35:48 +03005199 max_bits = (sizeof sregs->interrupt_bitmap) << 3;
5200 pending_vec = find_first_bit(
5201 (const unsigned long *)sregs->interrupt_bitmap, max_bits);
5202 if (pending_vec < max_bits) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03005203 kvm_queue_interrupt(vcpu, pending_vec, false);
Gleb Natapov923c61b2009-05-11 13:35:48 +03005204 pr_debug("Set back pending irq %d\n", pending_vec);
5205 if (irqchip_in_kernel(vcpu->kvm))
5206 kvm_pic_clear_isr_ack(vcpu->kvm);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005207 }
5208
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02005209 kvm_set_segment(vcpu, &sregs->cs, VCPU_SREG_CS);
5210 kvm_set_segment(vcpu, &sregs->ds, VCPU_SREG_DS);
5211 kvm_set_segment(vcpu, &sregs->es, VCPU_SREG_ES);
5212 kvm_set_segment(vcpu, &sregs->fs, VCPU_SREG_FS);
5213 kvm_set_segment(vcpu, &sregs->gs, VCPU_SREG_GS);
5214 kvm_set_segment(vcpu, &sregs->ss, VCPU_SREG_SS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005215
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02005216 kvm_set_segment(vcpu, &sregs->tr, VCPU_SREG_TR);
5217 kvm_set_segment(vcpu, &sregs->ldt, VCPU_SREG_LDTR);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005218
Mikhail Ershov5f0269f2009-08-03 14:58:25 +03005219 update_cr8_intercept(vcpu);
5220
Marcelo Tosatti9c3e4aa2008-09-10 16:40:55 -03005221 /* Older userspace won't unhalt the vcpu on reset. */
Gleb Natapovc5af89b2009-06-09 15:56:26 +03005222 if (kvm_vcpu_is_bsp(vcpu) && kvm_rip_read(vcpu) == 0xfff0 &&
Marcelo Tosatti9c3e4aa2008-09-10 16:40:55 -03005223 sregs->cs.selector == 0xf000 && sregs->cs.base == 0xffff0000 &&
Avi Kivity3eeb3282010-01-21 15:31:48 +02005224 !is_protmode(vcpu))
Marcelo Tosatti9c3e4aa2008-09-10 16:40:55 -03005225 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
5226
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005227 return 0;
5228}
5229
Jan Kiszkad0bfb942008-12-15 13:52:10 +01005230int kvm_arch_vcpu_ioctl_set_guest_debug(struct kvm_vcpu *vcpu,
5231 struct kvm_guest_debug *dbg)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005232{
Jan Kiszka355be0b2009-10-03 00:31:21 +02005233 unsigned long rflags;
Jan Kiszkaae675ef2008-12-15 13:52:10 +01005234 int i, r;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005235
Jan Kiszka4f926bf22009-10-30 12:46:59 +01005236 if (dbg->control & (KVM_GUESTDBG_INJECT_DB | KVM_GUESTDBG_INJECT_BP)) {
5237 r = -EBUSY;
5238 if (vcpu->arch.exception.pending)
Avi Kivity2122ff52010-05-13 11:25:04 +03005239 goto out;
Jan Kiszka4f926bf22009-10-30 12:46:59 +01005240 if (dbg->control & KVM_GUESTDBG_INJECT_DB)
5241 kvm_queue_exception(vcpu, DB_VECTOR);
5242 else
5243 kvm_queue_exception(vcpu, BP_VECTOR);
5244 }
5245
Jan Kiszka91586a32009-10-05 13:07:21 +02005246 /*
5247 * Read rflags as long as potentially injected trace flags are still
5248 * filtered out.
5249 */
5250 rflags = kvm_get_rflags(vcpu);
Jan Kiszka355be0b2009-10-03 00:31:21 +02005251
5252 vcpu->guest_debug = dbg->control;
5253 if (!(vcpu->guest_debug & KVM_GUESTDBG_ENABLE))
5254 vcpu->guest_debug = 0;
5255
5256 if (vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP) {
Jan Kiszkaae675ef2008-12-15 13:52:10 +01005257 for (i = 0; i < KVM_NR_DB_REGS; ++i)
5258 vcpu->arch.eff_db[i] = dbg->arch.debugreg[i];
5259 vcpu->arch.switch_db_regs =
5260 (dbg->arch.debugreg[7] & DR7_BP_EN_MASK);
5261 } else {
5262 for (i = 0; i < KVM_NR_DB_REGS; i++)
5263 vcpu->arch.eff_db[i] = vcpu->arch.db[i];
5264 vcpu->arch.switch_db_regs = (vcpu->arch.dr7 & DR7_BP_EN_MASK);
5265 }
5266
Jan Kiszkaf92653e2010-02-23 17:47:55 +01005267 if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
5268 vcpu->arch.singlestep_rip = kvm_rip_read(vcpu) +
5269 get_segment_base(vcpu, VCPU_SREG_CS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005270
Jan Kiszka91586a32009-10-05 13:07:21 +02005271 /*
5272 * Trigger an rflags update that will inject or remove the trace
5273 * flags.
5274 */
5275 kvm_set_rflags(vcpu, rflags);
Jan Kiszkad0bfb942008-12-15 13:52:10 +01005276
Jan Kiszka355be0b2009-10-03 00:31:21 +02005277 kvm_x86_ops->set_guest_debug(vcpu, dbg);
5278
Jan Kiszka4f926bf22009-10-30 12:46:59 +01005279 r = 0;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005280
Avi Kivity2122ff52010-05-13 11:25:04 +03005281out:
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05005282
5283 return r;
5284}
5285
5286/*
Zhang Xiantao8b006792007-11-16 13:05:55 +08005287 * Translate a guest virtual address to a guest physical address.
5288 */
5289int kvm_arch_vcpu_ioctl_translate(struct kvm_vcpu *vcpu,
5290 struct kvm_translation *tr)
5291{
5292 unsigned long vaddr = tr->linear_address;
5293 gpa_t gpa;
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005294 int idx;
Zhang Xiantao8b006792007-11-16 13:05:55 +08005295
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005296 idx = srcu_read_lock(&vcpu->kvm->srcu);
Gleb Natapov1871c602010-02-10 14:21:32 +02005297 gpa = kvm_mmu_gva_to_gpa_system(vcpu, vaddr, NULL);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005298 srcu_read_unlock(&vcpu->kvm->srcu, idx);
Zhang Xiantao8b006792007-11-16 13:05:55 +08005299 tr->physical_address = gpa;
5300 tr->valid = gpa != UNMAPPED_GVA;
5301 tr->writeable = 1;
5302 tr->usermode = 0;
Zhang Xiantao8b006792007-11-16 13:05:55 +08005303
5304 return 0;
5305}
5306
Hollis Blanchardd0752062007-10-31 17:24:25 -05005307int kvm_arch_vcpu_ioctl_get_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
5308{
Sheng Yang98918832010-05-17 17:08:28 +08005309 struct i387_fxsave_struct *fxsave =
5310 &vcpu->arch.guest_fpu.state->fxsave;
Hollis Blanchardd0752062007-10-31 17:24:25 -05005311
Hollis Blanchardd0752062007-10-31 17:24:25 -05005312 memcpy(fpu->fpr, fxsave->st_space, 128);
5313 fpu->fcw = fxsave->cwd;
5314 fpu->fsw = fxsave->swd;
5315 fpu->ftwx = fxsave->twd;
5316 fpu->last_opcode = fxsave->fop;
5317 fpu->last_ip = fxsave->rip;
5318 fpu->last_dp = fxsave->rdp;
5319 memcpy(fpu->xmm, fxsave->xmm_space, sizeof fxsave->xmm_space);
5320
Hollis Blanchardd0752062007-10-31 17:24:25 -05005321 return 0;
5322}
5323
5324int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
5325{
Sheng Yang98918832010-05-17 17:08:28 +08005326 struct i387_fxsave_struct *fxsave =
5327 &vcpu->arch.guest_fpu.state->fxsave;
Hollis Blanchardd0752062007-10-31 17:24:25 -05005328
Hollis Blanchardd0752062007-10-31 17:24:25 -05005329 memcpy(fxsave->st_space, fpu->fpr, 128);
5330 fxsave->cwd = fpu->fcw;
5331 fxsave->swd = fpu->fsw;
5332 fxsave->twd = fpu->ftwx;
5333 fxsave->fop = fpu->last_opcode;
5334 fxsave->rip = fpu->last_ip;
5335 fxsave->rdp = fpu->last_dp;
5336 memcpy(fxsave->xmm_space, fpu->xmm, sizeof fxsave->xmm_space);
5337
Hollis Blanchardd0752062007-10-31 17:24:25 -05005338 return 0;
5339}
5340
Jan Kiszka10ab25c2010-05-25 16:01:50 +02005341int fx_init(struct kvm_vcpu *vcpu)
Hollis Blanchardd0752062007-10-31 17:24:25 -05005342{
Jan Kiszka10ab25c2010-05-25 16:01:50 +02005343 int err;
5344
5345 err = fpu_alloc(&vcpu->arch.guest_fpu);
5346 if (err)
5347 return err;
5348
Sheng Yang98918832010-05-17 17:08:28 +08005349 fpu_finit(&vcpu->arch.guest_fpu);
Hollis Blanchardd0752062007-10-31 17:24:25 -05005350
Dexuan Cui2acf9232010-06-10 11:27:12 +08005351 /*
5352 * Ensure guest xcr0 is valid for loading
5353 */
5354 vcpu->arch.xcr0 = XSTATE_FP;
5355
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005356 vcpu->arch.cr0 |= X86_CR0_ET;
Jan Kiszka10ab25c2010-05-25 16:01:50 +02005357
5358 return 0;
Hollis Blanchardd0752062007-10-31 17:24:25 -05005359}
5360EXPORT_SYMBOL_GPL(fx_init);
5361
Sheng Yang98918832010-05-17 17:08:28 +08005362static void fx_free(struct kvm_vcpu *vcpu)
5363{
5364 fpu_free(&vcpu->arch.guest_fpu);
5365}
5366
Hollis Blanchardd0752062007-10-31 17:24:25 -05005367void kvm_load_guest_fpu(struct kvm_vcpu *vcpu)
5368{
Avi Kivity2608d7a2010-01-21 15:31:45 +02005369 if (vcpu->guest_fpu_loaded)
Hollis Blanchardd0752062007-10-31 17:24:25 -05005370 return;
5371
Dexuan Cui2acf9232010-06-10 11:27:12 +08005372 /*
5373 * Restore all possible states in the guest,
5374 * and assume host would use all available bits.
5375 * Guest xcr0 would be loaded later.
5376 */
5377 kvm_put_guest_xcr0(vcpu);
Hollis Blanchardd0752062007-10-31 17:24:25 -05005378 vcpu->guest_fpu_loaded = 1;
Sheng Yang7cf30852010-05-17 17:08:27 +08005379 unlazy_fpu(current);
Sheng Yang98918832010-05-17 17:08:28 +08005380 fpu_restore_checking(&vcpu->arch.guest_fpu);
Avi Kivity0c048512010-01-21 15:31:52 +02005381 trace_kvm_fpu(1);
Hollis Blanchardd0752062007-10-31 17:24:25 -05005382}
Hollis Blanchardd0752062007-10-31 17:24:25 -05005383
5384void kvm_put_guest_fpu(struct kvm_vcpu *vcpu)
5385{
Dexuan Cui2acf9232010-06-10 11:27:12 +08005386 kvm_put_guest_xcr0(vcpu);
5387
Hollis Blanchardd0752062007-10-31 17:24:25 -05005388 if (!vcpu->guest_fpu_loaded)
5389 return;
5390
5391 vcpu->guest_fpu_loaded = 0;
Sheng Yang98918832010-05-17 17:08:28 +08005392 fpu_save_init(&vcpu->arch.guest_fpu);
Avi Kivityf096ed82007-11-18 13:54:33 +02005393 ++vcpu->stat.fpu_reload;
Avi Kivity02daab22009-12-30 12:40:26 +02005394 set_bit(KVM_REQ_DEACTIVATE_FPU, &vcpu->requests);
Avi Kivity0c048512010-01-21 15:31:52 +02005395 trace_kvm_fpu(0);
Hollis Blanchardd0752062007-10-31 17:24:25 -05005396}
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005397
5398void kvm_arch_vcpu_free(struct kvm_vcpu *vcpu)
5399{
Joerg Roedel7f1ea202009-02-25 16:08:31 +01005400 if (vcpu->arch.time_page) {
5401 kvm_release_page_dirty(vcpu->arch.time_page);
5402 vcpu->arch.time_page = NULL;
5403 }
5404
Sheng Yang98918832010-05-17 17:08:28 +08005405 fx_free(vcpu);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005406 kvm_x86_ops->vcpu_free(vcpu);
5407}
5408
5409struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm,
5410 unsigned int id)
5411{
Avi Kivity26e52152007-11-20 15:30:24 +02005412 return kvm_x86_ops->vcpu_create(kvm, id);
5413}
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005414
Avi Kivity26e52152007-11-20 15:30:24 +02005415int kvm_arch_vcpu_setup(struct kvm_vcpu *vcpu)
5416{
5417 int r;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005418
Sheng Yang0bed3b52008-10-09 16:01:54 +08005419 vcpu->arch.mtrr_state.have_fixed = 1;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005420 vcpu_load(vcpu);
5421 r = kvm_arch_vcpu_reset(vcpu);
5422 if (r == 0)
5423 r = kvm_mmu_setup(vcpu);
5424 vcpu_put(vcpu);
5425 if (r < 0)
5426 goto free_vcpu;
5427
Avi Kivity26e52152007-11-20 15:30:24 +02005428 return 0;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005429free_vcpu:
5430 kvm_x86_ops->vcpu_free(vcpu);
Avi Kivity26e52152007-11-20 15:30:24 +02005431 return r;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005432}
5433
Hollis Blanchardd40ccc62007-11-19 14:04:43 -06005434void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005435{
5436 vcpu_load(vcpu);
5437 kvm_mmu_unload(vcpu);
5438 vcpu_put(vcpu);
5439
Sheng Yang98918832010-05-17 17:08:28 +08005440 fx_free(vcpu);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005441 kvm_x86_ops->vcpu_free(vcpu);
5442}
5443
5444int kvm_arch_vcpu_reset(struct kvm_vcpu *vcpu)
5445{
Jan Kiszka448fa4a2008-09-26 09:30:48 +02005446 vcpu->arch.nmi_pending = false;
5447 vcpu->arch.nmi_injected = false;
5448
Jan Kiszka42dbaa52008-12-15 13:52:10 +01005449 vcpu->arch.switch_db_regs = 0;
5450 memset(vcpu->arch.db, 0, sizeof(vcpu->arch.db));
5451 vcpu->arch.dr6 = DR6_FIXED_1;
5452 vcpu->arch.dr7 = DR7_FIXED_1;
5453
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005454 return kvm_x86_ops->vcpu_reset(vcpu);
5455}
5456
Alexander Graf10474ae2009-09-15 11:37:46 +02005457int kvm_arch_hardware_enable(void *garbage)
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005458{
Zachary Amsden0cca7902009-09-29 11:38:35 -10005459 /*
5460 * Since this may be called from a hotplug notifcation,
5461 * we can't get the CPU frequency directly.
5462 */
5463 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
5464 int cpu = raw_smp_processor_id();
5465 per_cpu(cpu_tsc_khz, cpu) = 0;
5466 }
Avi Kivity18863bd2009-09-07 11:12:18 +03005467
5468 kvm_shared_msr_cpu_online();
5469
Alexander Graf10474ae2009-09-15 11:37:46 +02005470 return kvm_x86_ops->hardware_enable(garbage);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005471}
5472
5473void kvm_arch_hardware_disable(void *garbage)
5474{
5475 kvm_x86_ops->hardware_disable(garbage);
Avi Kivity3548bab2009-11-28 14:18:47 +02005476 drop_user_return_notifiers(garbage);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005477}
5478
5479int kvm_arch_hardware_setup(void)
5480{
5481 return kvm_x86_ops->hardware_setup();
5482}
5483
5484void kvm_arch_hardware_unsetup(void)
5485{
5486 kvm_x86_ops->hardware_unsetup();
5487}
5488
5489void kvm_arch_check_processor_compat(void *rtn)
5490{
5491 kvm_x86_ops->check_processor_compatibility(rtn);
5492}
5493
5494int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
5495{
5496 struct page *page;
5497 struct kvm *kvm;
5498 int r;
5499
5500 BUG_ON(vcpu->kvm == NULL);
5501 kvm = vcpu->kvm;
5502
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005503 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
Gleb Natapovc5af89b2009-06-09 15:56:26 +03005504 if (!irqchip_in_kernel(kvm) || kvm_vcpu_is_bsp(vcpu))
Avi Kivitya4535292008-04-13 17:54:35 +03005505 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005506 else
Avi Kivitya4535292008-04-13 17:54:35 +03005507 vcpu->arch.mp_state = KVM_MP_STATE_UNINITIALIZED;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005508
5509 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
5510 if (!page) {
5511 r = -ENOMEM;
5512 goto fail;
5513 }
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005514 vcpu->arch.pio_data = page_address(page);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005515
5516 r = kvm_mmu_create(vcpu);
5517 if (r < 0)
5518 goto fail_free_pio_data;
5519
5520 if (irqchip_in_kernel(kvm)) {
5521 r = kvm_create_lapic(vcpu);
5522 if (r < 0)
5523 goto fail_mmu_destroy;
5524 }
5525
Huang Ying890ca9a2009-05-11 16:48:15 +08005526 vcpu->arch.mce_banks = kzalloc(KVM_MAX_MCE_BANKS * sizeof(u64) * 4,
5527 GFP_KERNEL);
5528 if (!vcpu->arch.mce_banks) {
5529 r = -ENOMEM;
Wei Yongjun443c39b2010-01-22 14:21:29 +08005530 goto fail_free_lapic;
Huang Ying890ca9a2009-05-11 16:48:15 +08005531 }
5532 vcpu->arch.mcg_cap = KVM_MAX_MCE_BANKS;
5533
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005534 return 0;
Wei Yongjun443c39b2010-01-22 14:21:29 +08005535fail_free_lapic:
5536 kvm_free_lapic(vcpu);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005537fail_mmu_destroy:
5538 kvm_mmu_destroy(vcpu);
5539fail_free_pio_data:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005540 free_page((unsigned long)vcpu->arch.pio_data);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005541fail:
5542 return r;
5543}
5544
5545void kvm_arch_vcpu_uninit(struct kvm_vcpu *vcpu)
5546{
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005547 int idx;
5548
Wei Yongjun36cb93f2010-01-22 14:18:47 +08005549 kfree(vcpu->arch.mce_banks);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005550 kvm_free_lapic(vcpu);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005551 idx = srcu_read_lock(&vcpu->kvm->srcu);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005552 kvm_mmu_destroy(vcpu);
Marcelo Tosattif656ce02009-12-23 14:35:25 -02005553 srcu_read_unlock(&vcpu->kvm->srcu, idx);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08005554 free_page((unsigned long)vcpu->arch.pio_data);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08005555}
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005556
5557struct kvm *kvm_arch_create_vm(void)
5558{
5559 struct kvm *kvm = kzalloc(sizeof(struct kvm), GFP_KERNEL);
5560
5561 if (!kvm)
5562 return ERR_PTR(-ENOMEM);
5563
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02005564 kvm->arch.aliases = kzalloc(sizeof(struct kvm_mem_aliases), GFP_KERNEL);
5565 if (!kvm->arch.aliases) {
5566 kfree(kvm);
5567 return ERR_PTR(-ENOMEM);
5568 }
5569
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08005570 INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +03005571 INIT_LIST_HEAD(&kvm->arch.assigned_dev_head);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005572
Sheng Yang5550af42008-10-15 20:15:06 +08005573 /* Reserve bit 0 of irq_sources_bitmap for userspace irq source */
5574 set_bit(KVM_USERSPACE_IRQ_SOURCE_ID, &kvm->arch.irq_sources_bitmap);
5575
Marcelo Tosatti53f658b32008-12-11 20:45:05 +01005576 rdtscll(kvm->arch.vm_init_tsc);
5577
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005578 return kvm;
5579}
5580
5581static void kvm_unload_vcpu_mmu(struct kvm_vcpu *vcpu)
5582{
5583 vcpu_load(vcpu);
5584 kvm_mmu_unload(vcpu);
5585 vcpu_put(vcpu);
5586}
5587
5588static void kvm_free_vcpus(struct kvm *kvm)
5589{
5590 unsigned int i;
Gleb Natapov988a2ca2009-06-09 15:56:29 +03005591 struct kvm_vcpu *vcpu;
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005592
5593 /*
5594 * Unpin any mmu pages first.
5595 */
Gleb Natapov988a2ca2009-06-09 15:56:29 +03005596 kvm_for_each_vcpu(i, vcpu, kvm)
5597 kvm_unload_vcpu_mmu(vcpu);
5598 kvm_for_each_vcpu(i, vcpu, kvm)
5599 kvm_arch_vcpu_free(vcpu);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005600
Gleb Natapov988a2ca2009-06-09 15:56:29 +03005601 mutex_lock(&kvm->lock);
5602 for (i = 0; i < atomic_read(&kvm->online_vcpus); i++)
5603 kvm->vcpus[i] = NULL;
5604
5605 atomic_set(&kvm->online_vcpus, 0);
5606 mutex_unlock(&kvm->lock);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005607}
5608
Sheng Yangad8ba2c2009-01-06 10:03:02 +08005609void kvm_arch_sync_events(struct kvm *kvm)
5610{
Sheng Yangba4cef32009-01-06 10:03:03 +08005611 kvm_free_all_assigned_devices(kvm);
Sheng Yangad8ba2c2009-01-06 10:03:02 +08005612}
5613
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005614void kvm_arch_destroy_vm(struct kvm *kvm)
5615{
Sheng Yang6eb55812008-10-31 12:37:41 +08005616 kvm_iommu_unmap_guest(kvm);
Sheng Yang78376992008-01-28 05:10:22 +08005617 kvm_free_pit(kvm);
Zhang Xiantaod7deeeb02007-12-14 10:17:34 +08005618 kfree(kvm->arch.vpic);
5619 kfree(kvm->arch.vioapic);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005620 kvm_free_vcpus(kvm);
5621 kvm_free_physmem(kvm);
Avi Kivity3d458302008-03-25 11:26:13 +02005622 if (kvm->arch.apic_access_page)
5623 put_page(kvm->arch.apic_access_page);
Sheng Yangb7ebfb02008-04-25 21:44:52 +08005624 if (kvm->arch.ept_identity_pagetable)
5625 put_page(kvm->arch.ept_identity_pagetable);
Marcelo Tosatti64749202010-01-19 12:45:23 -02005626 cleanup_srcu_struct(&kvm->srcu);
Marcelo Tosattifef9cce2009-12-23 14:35:17 -02005627 kfree(kvm->arch.aliases);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08005628 kfree(kvm);
5629}
Zhang Xiantao0de10342007-11-20 16:25:04 +08005630
Marcelo Tosattif7784b82009-12-23 14:35:18 -02005631int kvm_arch_prepare_memory_region(struct kvm *kvm,
5632 struct kvm_memory_slot *memslot,
Zhang Xiantao0de10342007-11-20 16:25:04 +08005633 struct kvm_memory_slot old,
Marcelo Tosattif7784b82009-12-23 14:35:18 -02005634 struct kvm_userspace_memory_region *mem,
Zhang Xiantao0de10342007-11-20 16:25:04 +08005635 int user_alloc)
5636{
Marcelo Tosattif7784b82009-12-23 14:35:18 -02005637 int npages = memslot->npages;
Zhang Xiantao0de10342007-11-20 16:25:04 +08005638
5639 /*To keep backward compatibility with older userspace,
5640 *x86 needs to hanlde !user_alloc case.
5641 */
5642 if (!user_alloc) {
5643 if (npages && !old.rmap) {
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02005644 unsigned long userspace_addr;
5645
Izik Eidus72dc67a2008-02-10 18:04:15 +02005646 down_write(&current->mm->mmap_sem);
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02005647 userspace_addr = do_mmap(NULL, 0,
5648 npages * PAGE_SIZE,
5649 PROT_READ | PROT_WRITE,
Avi Kivityacee3c02008-08-26 17:22:47 +03005650 MAP_PRIVATE | MAP_ANONYMOUS,
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02005651 0);
Izik Eidus72dc67a2008-02-10 18:04:15 +02005652 up_write(&current->mm->mmap_sem);
Zhang Xiantao0de10342007-11-20 16:25:04 +08005653
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02005654 if (IS_ERR((void *)userspace_addr))
5655 return PTR_ERR((void *)userspace_addr);
5656
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02005657 memslot->userspace_addr = userspace_addr;
Zhang Xiantao0de10342007-11-20 16:25:04 +08005658 }
5659 }
5660
Marcelo Tosattif7784b82009-12-23 14:35:18 -02005661
5662 return 0;
5663}
5664
5665void kvm_arch_commit_memory_region(struct kvm *kvm,
5666 struct kvm_userspace_memory_region *mem,
5667 struct kvm_memory_slot old,
5668 int user_alloc)
5669{
5670
5671 int npages = mem->memory_size >> PAGE_SHIFT;
5672
5673 if (!user_alloc && !old.user_alloc && old.rmap && !npages) {
5674 int ret;
5675
5676 down_write(&current->mm->mmap_sem);
5677 ret = do_munmap(current->mm, old.userspace_addr,
5678 old.npages * PAGE_SIZE);
5679 up_write(&current->mm->mmap_sem);
5680 if (ret < 0)
5681 printk(KERN_WARNING
5682 "kvm_vm_ioctl_set_memory_region: "
5683 "failed to munmap memory\n");
5684 }
5685
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03005686 spin_lock(&kvm->mmu_lock);
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08005687 if (!kvm->arch.n_requested_mmu_pages) {
Zhang Xiantao0de10342007-11-20 16:25:04 +08005688 unsigned int nr_mmu_pages = kvm_mmu_calculate_mmu_pages(kvm);
5689 kvm_mmu_change_mmu_pages(kvm, nr_mmu_pages);
5690 }
5691
5692 kvm_mmu_slot_remove_write_access(kvm, mem->slot);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03005693 spin_unlock(&kvm->mmu_lock);
Zhang Xiantao0de10342007-11-20 16:25:04 +08005694}
Zhang Xiantao1d737c82007-12-14 09:35:10 +08005695
Marcelo Tosatti34d4cb82008-07-10 20:49:31 -03005696void kvm_arch_flush_shadow(struct kvm *kvm)
5697{
5698 kvm_mmu_zap_all(kvm);
Marcelo Tosatti8986ecc2009-05-12 18:55:45 -03005699 kvm_reload_remote_mmus(kvm);
Marcelo Tosatti34d4cb82008-07-10 20:49:31 -03005700}
5701
Zhang Xiantao1d737c82007-12-14 09:35:10 +08005702int kvm_arch_vcpu_runnable(struct kvm_vcpu *vcpu)
5703{
Avi Kivitya4535292008-04-13 17:54:35 +03005704 return vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE
Gleb Natapova1b37102009-07-09 15:33:52 +03005705 || vcpu->arch.mp_state == KVM_MP_STATE_SIPI_RECEIVED
5706 || vcpu->arch.nmi_pending ||
5707 (kvm_arch_interrupt_allowed(vcpu) &&
5708 kvm_cpu_has_interrupt(vcpu));
Zhang Xiantao1d737c82007-12-14 09:35:10 +08005709}
Zhang Xiantao57361992007-12-17 14:21:40 +08005710
Zhang Xiantao57361992007-12-17 14:21:40 +08005711void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
5712{
Marcelo Tosatti32f88402009-05-07 17:55:12 -03005713 int me;
5714 int cpu = vcpu->cpu;
Zhang Xiantao57361992007-12-17 14:21:40 +08005715
5716 if (waitqueue_active(&vcpu->wq)) {
5717 wake_up_interruptible(&vcpu->wq);
5718 ++vcpu->stat.halt_wakeup;
5719 }
Marcelo Tosatti32f88402009-05-07 17:55:12 -03005720
5721 me = get_cpu();
5722 if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
Avi Kivityd94e1dc2010-05-03 16:54:48 +03005723 if (atomic_xchg(&vcpu->guest_mode, 0))
Marcelo Tosatti32f88402009-05-07 17:55:12 -03005724 smp_send_reschedule(cpu);
Marcelo Tosattie9571ed2008-04-11 15:01:22 -03005725 put_cpu();
Zhang Xiantao57361992007-12-17 14:21:40 +08005726}
Gleb Natapov78646122009-03-23 12:12:11 +02005727
5728int kvm_arch_interrupt_allowed(struct kvm_vcpu *vcpu)
5729{
5730 return kvm_x86_ops->interrupt_allowed(vcpu);
5731}
Marcelo Tosatti229456f2009-06-17 09:22:14 -03005732
Jan Kiszkaf92653e2010-02-23 17:47:55 +01005733bool kvm_is_linear_rip(struct kvm_vcpu *vcpu, unsigned long linear_rip)
5734{
5735 unsigned long current_rip = kvm_rip_read(vcpu) +
5736 get_segment_base(vcpu, VCPU_SREG_CS);
5737
5738 return current_rip == linear_rip;
5739}
5740EXPORT_SYMBOL_GPL(kvm_is_linear_rip);
5741
Jan Kiszka94fe45d2009-10-18 13:24:44 +02005742unsigned long kvm_get_rflags(struct kvm_vcpu *vcpu)
5743{
5744 unsigned long rflags;
5745
5746 rflags = kvm_x86_ops->get_rflags(vcpu);
5747 if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
Jan Kiszkac310bac2010-02-23 17:47:58 +01005748 rflags &= ~X86_EFLAGS_TF;
Jan Kiszka94fe45d2009-10-18 13:24:44 +02005749 return rflags;
5750}
5751EXPORT_SYMBOL_GPL(kvm_get_rflags);
5752
5753void kvm_set_rflags(struct kvm_vcpu *vcpu, unsigned long rflags)
5754{
5755 if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP &&
Jan Kiszkaf92653e2010-02-23 17:47:55 +01005756 kvm_is_linear_rip(vcpu, vcpu->arch.singlestep_rip))
Jan Kiszkac310bac2010-02-23 17:47:58 +01005757 rflags |= X86_EFLAGS_TF;
Jan Kiszka94fe45d2009-10-18 13:24:44 +02005758 kvm_x86_ops->set_rflags(vcpu, rflags);
5759}
5760EXPORT_SYMBOL_GPL(kvm_set_rflags);
5761
Marcelo Tosatti229456f2009-06-17 09:22:14 -03005762EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_exit);
5763EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_inj_virq);
5764EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_page_fault);
5765EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_msr);
5766EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_cr);
Joerg Roedel0ac406d2009-10-09 16:08:27 +02005767EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_vmrun);
Joerg Roedeld8cabdd2009-10-09 16:08:28 +02005768EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_vmexit);
Joerg Roedel17897f32009-10-09 16:08:29 +02005769EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_vmexit_inject);
Joerg Roedel236649d2009-10-09 16:08:30 +02005770EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_intr_vmexit);
Joerg Roedelec1ff792009-10-09 16:08:31 +02005771EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_invlpga);
Joerg Roedel532a46b2009-10-09 16:08:32 +02005772EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_skinit);
Joerg Roedel2e554e82010-02-24 18:59:14 +01005773EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_intercepts);