blob: be451ee44249be80cdc4829df1f8a6ccc1f5d2f5 [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
Carsten Otte043405e2007-10-10 17:16:19 +02009 *
10 * Authors:
11 * Avi Kivity <avi@qumranet.com>
12 * Yaniv Kamay <yaniv@qumranet.com>
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +030013 * Amit Shah <amit.shah@qumranet.com>
14 * Ben-Ami Yassour <benami@il.ibm.com>
Carsten Otte043405e2007-10-10 17:16:19 +020015 *
16 * This work is licensed under the terms of the GNU GPL, version 2. See
17 * the COPYING file in the top-level directory.
18 *
19 */
20
Avi Kivityedf88412007-12-16 11:02:48 +020021#include <linux/kvm_host.h>
Carsten Otte313a3dc2007-10-11 19:16:52 +020022#include "irq.h"
Zhang Xiantao1d737c82007-12-14 09:35:10 +080023#include "mmu.h"
Sheng Yang78376992008-01-28 05:10:22 +080024#include "i8254.h"
Izik Eidus37817f22008-03-24 23:14:53 +020025#include "tss.h"
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -030026#include "kvm_cache_regs.h"
Avi Kivity26eef702008-07-03 14:59:22 +030027#include "x86.h"
Carsten Otte313a3dc2007-10-11 19:16:52 +020028
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -020029#include <linux/clocksource.h>
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +030030#include <linux/interrupt.h>
Carsten Otte313a3dc2007-10-11 19:16:52 +020031#include <linux/kvm.h>
32#include <linux/fs.h>
33#include <linux/vmalloc.h>
Carsten Otte5fb76f92007-10-29 16:08:51 +010034#include <linux/module.h>
Zhang Xiantao0de10342007-11-20 16:25:04 +080035#include <linux/mman.h>
Marcelo Tosatti2bacc552007-12-12 10:46:12 -050036#include <linux/highmem.h>
Joerg Roedel19de40a2008-12-03 14:43:34 +010037#include <linux/iommu.h>
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +030038#include <linux/intel-iommu.h>
Gerd Hoffmannc8076602009-02-04 17:52:04 +010039#include <linux/cpufreq.h>
Avi Kivityaec51dc2009-07-01 16:01:02 +030040#include <trace/events/kvm.h>
41#undef TRACE_INCLUDE_FILE
Marcelo Tosatti229456f2009-06-17 09:22:14 -030042#define CREATE_TRACE_POINTS
43#include "trace.h"
Carsten Otte043405e2007-10-10 17:16:19 +020044
45#include <asm/uaccess.h>
Zhang Xiantaod825ed02007-11-14 20:08:51 +080046#include <asm/msr.h>
Avi Kivitya5f61302008-02-20 17:57:21 +020047#include <asm/desc.h>
Sheng Yang0bed3b52008-10-09 16:01:54 +080048#include <asm/mtrr.h>
Huang Ying890ca9a2009-05-11 16:48:15 +080049#include <asm/mce.h>
Carsten Otte043405e2007-10-10 17:16:19 +020050
Carsten Otte313a3dc2007-10-11 19:16:52 +020051#define MAX_IO_MSRS 256
Carsten Ottea03490e2007-10-29 16:09:35 +010052#define CR0_RESERVED_BITS \
53 (~(unsigned long)(X86_CR0_PE | X86_CR0_MP | X86_CR0_EM | X86_CR0_TS \
54 | X86_CR0_ET | X86_CR0_NE | X86_CR0_WP | X86_CR0_AM \
55 | X86_CR0_NW | X86_CR0_CD | X86_CR0_PG))
56#define CR4_RESERVED_BITS \
57 (~(unsigned long)(X86_CR4_VME | X86_CR4_PVI | X86_CR4_TSD | X86_CR4_DE\
58 | X86_CR4_PSE | X86_CR4_PAE | X86_CR4_MCE \
59 | X86_CR4_PGE | X86_CR4_PCE | X86_CR4_OSFXSR \
60 | X86_CR4_OSXMMEXCPT | X86_CR4_VMXE))
61
62#define CR8_RESERVED_BITS (~(unsigned long)X86_CR8_TPR)
Huang Ying890ca9a2009-05-11 16:48:15 +080063
64#define KVM_MAX_MCE_BANKS 32
65#define KVM_MCE_CAP_SUPPORTED MCG_CTL_P
66
Joerg Roedel50a37eb2008-01-31 14:57:38 +010067/* EFER defaults:
68 * - enable syscall per default because its emulated by KVM
69 * - enable LME and LMA per default on 64 bit KVM
70 */
71#ifdef CONFIG_X86_64
72static u64 __read_mostly efer_reserved_bits = 0xfffffffffffffafeULL;
73#else
74static u64 __read_mostly efer_reserved_bits = 0xfffffffffffffffeULL;
75#endif
Carsten Otte313a3dc2007-10-11 19:16:52 +020076
Avi Kivityba1389b2007-11-18 16:24:12 +020077#define VM_STAT(x) offsetof(struct kvm, stat.x), KVM_STAT_VM
78#define VCPU_STAT(x) offsetof(struct kvm_vcpu, stat.x), KVM_STAT_VCPU
Hollis Blanchard417bc302007-10-31 17:24:23 -050079
Gleb Natapovcb142eb2009-08-09 15:17:40 +030080static void update_cr8_intercept(struct kvm_vcpu *vcpu);
Avi Kivity674eea02008-02-11 18:37:23 +020081static int kvm_dev_ioctl_get_supported_cpuid(struct kvm_cpuid2 *cpuid,
82 struct kvm_cpuid_entry2 __user *entries);
83
Zhang Xiantao97896d02007-11-14 20:09:30 +080084struct kvm_x86_ops *kvm_x86_ops;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -030085EXPORT_SYMBOL_GPL(kvm_x86_ops);
Zhang Xiantao97896d02007-11-14 20:09:30 +080086
Andre Przywaraed85c062009-06-25 12:36:49 +020087int ignore_msrs = 0;
88module_param_named(ignore_msrs, ignore_msrs, bool, S_IRUGO | S_IWUSR);
89
Hollis Blanchard417bc302007-10-31 17:24:23 -050090struct kvm_stats_debugfs_item debugfs_entries[] = {
Avi Kivityba1389b2007-11-18 16:24:12 +020091 { "pf_fixed", VCPU_STAT(pf_fixed) },
92 { "pf_guest", VCPU_STAT(pf_guest) },
93 { "tlb_flush", VCPU_STAT(tlb_flush) },
94 { "invlpg", VCPU_STAT(invlpg) },
95 { "exits", VCPU_STAT(exits) },
96 { "io_exits", VCPU_STAT(io_exits) },
97 { "mmio_exits", VCPU_STAT(mmio_exits) },
98 { "signal_exits", VCPU_STAT(signal_exits) },
99 { "irq_window", VCPU_STAT(irq_window_exits) },
Sheng Yangf08864b2008-05-15 18:23:25 +0800100 { "nmi_window", VCPU_STAT(nmi_window_exits) },
Avi Kivityba1389b2007-11-18 16:24:12 +0200101 { "halt_exits", VCPU_STAT(halt_exits) },
102 { "halt_wakeup", VCPU_STAT(halt_wakeup) },
Amit Shahf11c3a82008-02-21 01:00:30 +0530103 { "hypercalls", VCPU_STAT(hypercalls) },
Avi Kivityba1389b2007-11-18 16:24:12 +0200104 { "request_irq", VCPU_STAT(request_irq_exits) },
105 { "irq_exits", VCPU_STAT(irq_exits) },
106 { "host_state_reload", VCPU_STAT(host_state_reload) },
107 { "efer_reload", VCPU_STAT(efer_reload) },
108 { "fpu_reload", VCPU_STAT(fpu_reload) },
109 { "insn_emulation", VCPU_STAT(insn_emulation) },
110 { "insn_emulation_fail", VCPU_STAT(insn_emulation_fail) },
Avi Kivityfa89a812008-09-01 15:57:51 +0300111 { "irq_injections", VCPU_STAT(irq_injections) },
Jan Kiszkac4abb7c2008-09-26 09:30:55 +0200112 { "nmi_injections", VCPU_STAT(nmi_injections) },
Avi Kivity4cee5762007-11-18 16:37:07 +0200113 { "mmu_shadow_zapped", VM_STAT(mmu_shadow_zapped) },
114 { "mmu_pte_write", VM_STAT(mmu_pte_write) },
115 { "mmu_pte_updated", VM_STAT(mmu_pte_updated) },
116 { "mmu_pde_zapped", VM_STAT(mmu_pde_zapped) },
117 { "mmu_flooded", VM_STAT(mmu_flooded) },
118 { "mmu_recycled", VM_STAT(mmu_recycled) },
Avi Kivitydfc5aa02007-12-18 19:47:18 +0200119 { "mmu_cache_miss", VM_STAT(mmu_cache_miss) },
Marcelo Tosatti4731d4c2008-09-23 13:18:39 -0300120 { "mmu_unsync", VM_STAT(mmu_unsync) },
Avi Kivity0f74a242007-11-20 23:01:14 +0200121 { "remote_tlb_flush", VM_STAT(remote_tlb_flush) },
Marcelo Tosatti05da4552008-02-23 11:44:30 -0300122 { "largepages", VM_STAT(lpages) },
Hollis Blanchard417bc302007-10-31 17:24:23 -0500123 { NULL }
124};
125
Carsten Otte5fb76f92007-10-29 16:08:51 +0100126unsigned long segment_base(u16 selector)
127{
128 struct descriptor_table gdt;
Avi Kivitya5f61302008-02-20 17:57:21 +0200129 struct desc_struct *d;
Carsten Otte5fb76f92007-10-29 16:08:51 +0100130 unsigned long table_base;
131 unsigned long v;
132
133 if (selector == 0)
134 return 0;
135
Akinobu Mitab792c342009-07-19 00:00:01 +0900136 kvm_get_gdt(&gdt);
Carsten Otte5fb76f92007-10-29 16:08:51 +0100137 table_base = gdt.base;
138
139 if (selector & 4) { /* from ldt */
Akinobu Mitab792c342009-07-19 00:00:01 +0900140 u16 ldt_selector = kvm_read_ldt();
Carsten Otte5fb76f92007-10-29 16:08:51 +0100141
Carsten Otte5fb76f92007-10-29 16:08:51 +0100142 table_base = segment_base(ldt_selector);
143 }
Avi Kivitya5f61302008-02-20 17:57:21 +0200144 d = (struct desc_struct *)(table_base + (selector & ~7));
Akinobu Mita46a359e2009-07-18 23:58:32 +0900145 v = get_desc_base(d);
Carsten Otte5fb76f92007-10-29 16:08:51 +0100146#ifdef CONFIG_X86_64
Avi Kivitya5f61302008-02-20 17:57:21 +0200147 if (d->s == 0 && (d->type == 2 || d->type == 9 || d->type == 11))
148 v |= ((unsigned long)((struct ldttss_desc64 *)d)->base3) << 32;
Carsten Otte5fb76f92007-10-29 16:08:51 +0100149#endif
150 return v;
151}
152EXPORT_SYMBOL_GPL(segment_base);
153
Carsten Otte6866b832007-10-29 16:09:10 +0100154u64 kvm_get_apic_base(struct kvm_vcpu *vcpu)
155{
156 if (irqchip_in_kernel(vcpu->kvm))
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800157 return vcpu->arch.apic_base;
Carsten Otte6866b832007-10-29 16:09:10 +0100158 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800159 return vcpu->arch.apic_base;
Carsten Otte6866b832007-10-29 16:09:10 +0100160}
161EXPORT_SYMBOL_GPL(kvm_get_apic_base);
162
163void kvm_set_apic_base(struct kvm_vcpu *vcpu, u64 data)
164{
165 /* TODO: reserve bits check */
166 if (irqchip_in_kernel(vcpu->kvm))
167 kvm_lapic_set_base(vcpu, data);
168 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800169 vcpu->arch.apic_base = data;
Carsten Otte6866b832007-10-29 16:09:10 +0100170}
171EXPORT_SYMBOL_GPL(kvm_set_apic_base);
172
Avi Kivity298101d2007-11-25 13:41:11 +0200173void kvm_queue_exception(struct kvm_vcpu *vcpu, unsigned nr)
174{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800175 WARN_ON(vcpu->arch.exception.pending);
176 vcpu->arch.exception.pending = true;
177 vcpu->arch.exception.has_error_code = false;
178 vcpu->arch.exception.nr = nr;
Avi Kivity298101d2007-11-25 13:41:11 +0200179}
180EXPORT_SYMBOL_GPL(kvm_queue_exception);
181
Avi Kivityc3c91fe2007-11-25 14:04:58 +0200182void kvm_inject_page_fault(struct kvm_vcpu *vcpu, unsigned long addr,
183 u32 error_code)
184{
185 ++vcpu->stat.pf_guest;
Alexander Grafd8017472008-11-25 20:17:11 +0100186
Joerg Roedel71c4dfa2008-02-26 16:49:16 +0100187 if (vcpu->arch.exception.pending) {
Gleb Natapov6edf14d2009-06-11 11:26:11 +0300188 switch(vcpu->arch.exception.nr) {
189 case DF_VECTOR:
Joerg Roedel71c4dfa2008-02-26 16:49:16 +0100190 /* triple fault -> shutdown */
191 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
Gleb Natapov6edf14d2009-06-11 11:26:11 +0300192 return;
193 case PF_VECTOR:
194 vcpu->arch.exception.nr = DF_VECTOR;
195 vcpu->arch.exception.error_code = 0;
196 return;
197 default:
198 /* replace previous exception with a new one in a hope
199 that instruction re-execution will regenerate lost
200 exception */
201 vcpu->arch.exception.pending = false;
202 break;
Joerg Roedel71c4dfa2008-02-26 16:49:16 +0100203 }
Avi Kivityc3c91fe2007-11-25 14:04:58 +0200204 }
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800205 vcpu->arch.cr2 = addr;
Avi Kivityc3c91fe2007-11-25 14:04:58 +0200206 kvm_queue_exception_e(vcpu, PF_VECTOR, error_code);
207}
208
Sheng Yang3419ffc2008-05-15 09:52:48 +0800209void kvm_inject_nmi(struct kvm_vcpu *vcpu)
210{
211 vcpu->arch.nmi_pending = 1;
212}
213EXPORT_SYMBOL_GPL(kvm_inject_nmi);
214
Avi Kivity298101d2007-11-25 13:41:11 +0200215void kvm_queue_exception_e(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code)
216{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800217 WARN_ON(vcpu->arch.exception.pending);
218 vcpu->arch.exception.pending = true;
219 vcpu->arch.exception.has_error_code = true;
220 vcpu->arch.exception.nr = nr;
221 vcpu->arch.exception.error_code = error_code;
Avi Kivity298101d2007-11-25 13:41:11 +0200222}
223EXPORT_SYMBOL_GPL(kvm_queue_exception_e);
224
Carsten Ottea03490e2007-10-29 16:09:35 +0100225/*
Avi Kivity0a79b002009-09-01 12:03:25 +0300226 * Checks if cpl <= required_cpl; if true, return true. Otherwise queue
227 * a #GP and return false.
228 */
229bool kvm_require_cpl(struct kvm_vcpu *vcpu, int required_cpl)
Carsten Otte043405e2007-10-10 17:16:19 +0200230{
Avi Kivity0a79b002009-09-01 12:03:25 +0300231 if (kvm_x86_ops->get_cpl(vcpu) <= required_cpl)
232 return true;
233 kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
234 return false;
Carsten Ottea03490e2007-10-29 16:09:35 +0100235}
Avi Kivity0a79b002009-09-01 12:03:25 +0300236EXPORT_SYMBOL_GPL(kvm_require_cpl);
Carsten Ottea03490e2007-10-29 16:09:35 +0100237
238/*
239 * Load the pae pdptrs. Return true is they are all valid.
240 */
241int load_pdptrs(struct kvm_vcpu *vcpu, unsigned long cr3)
242{
243 gfn_t pdpt_gfn = cr3 >> PAGE_SHIFT;
244 unsigned offset = ((cr3 & (PAGE_SIZE-1)) >> 5) << 2;
245 int i;
246 int ret;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800247 u64 pdpte[ARRAY_SIZE(vcpu->arch.pdptrs)];
Carsten Ottea03490e2007-10-29 16:09:35 +0100248
Carsten Ottea03490e2007-10-29 16:09:35 +0100249 ret = kvm_read_guest_page(vcpu->kvm, pdpt_gfn, pdpte,
250 offset * sizeof(u64), sizeof(pdpte));
251 if (ret < 0) {
252 ret = 0;
253 goto out;
254 }
255 for (i = 0; i < ARRAY_SIZE(pdpte); ++i) {
Avi Kivity43a37952009-06-10 14:12:05 +0300256 if (is_present_gpte(pdpte[i]) &&
Dong, Eddie20c466b2009-03-31 23:03:45 +0800257 (pdpte[i] & vcpu->arch.mmu.rsvd_bits_mask[0][2])) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100258 ret = 0;
259 goto out;
260 }
261 }
262 ret = 1;
263
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800264 memcpy(vcpu->arch.pdptrs, pdpte, sizeof(vcpu->arch.pdptrs));
Avi Kivity6de4f3a2009-05-31 22:58:47 +0300265 __set_bit(VCPU_EXREG_PDPTR,
266 (unsigned long *)&vcpu->arch.regs_avail);
267 __set_bit(VCPU_EXREG_PDPTR,
268 (unsigned long *)&vcpu->arch.regs_dirty);
Carsten Ottea03490e2007-10-29 16:09:35 +0100269out:
Carsten Ottea03490e2007-10-29 16:09:35 +0100270
271 return ret;
272}
Joerg Roedelcc4b6872008-02-07 13:47:43 +0100273EXPORT_SYMBOL_GPL(load_pdptrs);
Carsten Ottea03490e2007-10-29 16:09:35 +0100274
Avi Kivityd835dfe2007-11-21 02:57:59 +0200275static bool pdptrs_changed(struct kvm_vcpu *vcpu)
276{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800277 u64 pdpte[ARRAY_SIZE(vcpu->arch.pdptrs)];
Avi Kivityd835dfe2007-11-21 02:57:59 +0200278 bool changed = true;
279 int r;
280
281 if (is_long_mode(vcpu) || !is_pae(vcpu))
282 return false;
283
Avi Kivity6de4f3a2009-05-31 22:58:47 +0300284 if (!test_bit(VCPU_EXREG_PDPTR,
285 (unsigned long *)&vcpu->arch.regs_avail))
286 return true;
287
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800288 r = kvm_read_guest(vcpu->kvm, vcpu->arch.cr3 & ~31u, pdpte, sizeof(pdpte));
Avi Kivityd835dfe2007-11-21 02:57:59 +0200289 if (r < 0)
290 goto out;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800291 changed = memcmp(pdpte, vcpu->arch.pdptrs, sizeof(pdpte)) != 0;
Avi Kivityd835dfe2007-11-21 02:57:59 +0200292out:
Avi Kivityd835dfe2007-11-21 02:57:59 +0200293
294 return changed;
295}
296
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200297void kvm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
Carsten Ottea03490e2007-10-29 16:09:35 +0100298{
299 if (cr0 & CR0_RESERVED_BITS) {
300 printk(KERN_DEBUG "set_cr0: 0x%lx #GP, reserved bits 0x%lx\n",
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800301 cr0, vcpu->arch.cr0);
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200302 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100303 return;
304 }
305
306 if ((cr0 & X86_CR0_NW) && !(cr0 & X86_CR0_CD)) {
307 printk(KERN_DEBUG "set_cr0: #GP, CD == 0 && NW == 1\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200308 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100309 return;
310 }
311
312 if ((cr0 & X86_CR0_PG) && !(cr0 & X86_CR0_PE)) {
313 printk(KERN_DEBUG "set_cr0: #GP, set PG flag "
314 "and a clear PE flag\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200315 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100316 return;
317 }
318
319 if (!is_paging(vcpu) && (cr0 & X86_CR0_PG)) {
320#ifdef CONFIG_X86_64
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800321 if ((vcpu->arch.shadow_efer & EFER_LME)) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100322 int cs_db, cs_l;
323
324 if (!is_pae(vcpu)) {
325 printk(KERN_DEBUG "set_cr0: #GP, start paging "
326 "in long mode while PAE is disabled\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200327 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100328 return;
329 }
330 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
331 if (cs_l) {
332 printk(KERN_DEBUG "set_cr0: #GP, start paging "
333 "in long mode while CS.L == 1\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200334 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100335 return;
336
337 }
338 } else
339#endif
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800340 if (is_pae(vcpu) && !load_pdptrs(vcpu, vcpu->arch.cr3)) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100341 printk(KERN_DEBUG "set_cr0: #GP, pdptrs "
342 "reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200343 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100344 return;
345 }
346
347 }
348
349 kvm_x86_ops->set_cr0(vcpu, cr0);
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800350 vcpu->arch.cr0 = cr0;
Carsten Ottea03490e2007-10-29 16:09:35 +0100351
Carsten Ottea03490e2007-10-29 16:09:35 +0100352 kvm_mmu_reset_context(vcpu);
Carsten Ottea03490e2007-10-29 16:09:35 +0100353 return;
354}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200355EXPORT_SYMBOL_GPL(kvm_set_cr0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100356
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200357void kvm_lmsw(struct kvm_vcpu *vcpu, unsigned long msw)
Carsten Ottea03490e2007-10-29 16:09:35 +0100358{
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200359 kvm_set_cr0(vcpu, (vcpu->arch.cr0 & ~0x0ful) | (msw & 0x0f));
Carsten Ottea03490e2007-10-29 16:09:35 +0100360}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200361EXPORT_SYMBOL_GPL(kvm_lmsw);
Carsten Ottea03490e2007-10-29 16:09:35 +0100362
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200363void kvm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
Carsten Ottea03490e2007-10-29 16:09:35 +0100364{
Avi Kivitya2edf572009-05-24 22:19:00 +0300365 unsigned long old_cr4 = vcpu->arch.cr4;
366 unsigned long pdptr_bits = X86_CR4_PGE | X86_CR4_PSE | X86_CR4_PAE;
367
Carsten Ottea03490e2007-10-29 16:09:35 +0100368 if (cr4 & CR4_RESERVED_BITS) {
369 printk(KERN_DEBUG "set_cr4: #GP, reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200370 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100371 return;
372 }
373
374 if (is_long_mode(vcpu)) {
375 if (!(cr4 & X86_CR4_PAE)) {
376 printk(KERN_DEBUG "set_cr4: #GP, clearing PAE while "
377 "in long mode\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200378 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100379 return;
380 }
Avi Kivitya2edf572009-05-24 22:19:00 +0300381 } else if (is_paging(vcpu) && (cr4 & X86_CR4_PAE)
382 && ((cr4 ^ old_cr4) & pdptr_bits)
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800383 && !load_pdptrs(vcpu, vcpu->arch.cr3)) {
Carsten Ottea03490e2007-10-29 16:09:35 +0100384 printk(KERN_DEBUG "set_cr4: #GP, pdptrs reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200385 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100386 return;
387 }
388
389 if (cr4 & X86_CR4_VMXE) {
390 printk(KERN_DEBUG "set_cr4: #GP, setting VMXE\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200391 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100392 return;
393 }
394 kvm_x86_ops->set_cr4(vcpu, cr4);
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800395 vcpu->arch.cr4 = cr4;
Avi Kivity5a41acc2009-01-11 17:19:35 +0200396 vcpu->arch.mmu.base_role.cr4_pge = (cr4 & X86_CR4_PGE) && !tdp_enabled;
Carsten Ottea03490e2007-10-29 16:09:35 +0100397 kvm_mmu_reset_context(vcpu);
Carsten Ottea03490e2007-10-29 16:09:35 +0100398}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200399EXPORT_SYMBOL_GPL(kvm_set_cr4);
Carsten Ottea03490e2007-10-29 16:09:35 +0100400
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200401void kvm_set_cr3(struct kvm_vcpu *vcpu, unsigned long cr3)
Carsten Ottea03490e2007-10-29 16:09:35 +0100402{
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800403 if (cr3 == vcpu->arch.cr3 && !pdptrs_changed(vcpu)) {
Marcelo Tosatti0ba73cd2008-09-23 13:18:34 -0300404 kvm_mmu_sync_roots(vcpu);
Avi Kivityd835dfe2007-11-21 02:57:59 +0200405 kvm_mmu_flush_tlb(vcpu);
406 return;
407 }
408
Carsten Ottea03490e2007-10-29 16:09:35 +0100409 if (is_long_mode(vcpu)) {
410 if (cr3 & CR3_L_MODE_RESERVED_BITS) {
411 printk(KERN_DEBUG "set_cr3: #GP, reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200412 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100413 return;
414 }
415 } else {
416 if (is_pae(vcpu)) {
417 if (cr3 & CR3_PAE_RESERVED_BITS) {
418 printk(KERN_DEBUG
419 "set_cr3: #GP, reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200420 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100421 return;
422 }
423 if (is_paging(vcpu) && !load_pdptrs(vcpu, cr3)) {
424 printk(KERN_DEBUG "set_cr3: #GP, pdptrs "
425 "reserved bits\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200426 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100427 return;
428 }
429 }
430 /*
431 * We don't check reserved bits in nonpae mode, because
432 * this isn't enforced, and VMware depends on this.
433 */
434 }
435
Carsten Ottea03490e2007-10-29 16:09:35 +0100436 /*
437 * Does the new cr3 value map to physical memory? (Note, we
438 * catch an invalid cr3 even in real-mode, because it would
439 * cause trouble later on when we turn on paging anyway.)
440 *
441 * A real CPU would silently accept an invalid cr3 and would
442 * attempt to use it - with largely undefined (and often hard
443 * to debug) behavior on the guest side.
444 */
445 if (unlikely(!gfn_to_memslot(vcpu->kvm, cr3 >> PAGE_SHIFT)))
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200446 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100447 else {
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800448 vcpu->arch.cr3 = cr3;
449 vcpu->arch.mmu.new_cr3(vcpu);
Carsten Ottea03490e2007-10-29 16:09:35 +0100450 }
Carsten Ottea03490e2007-10-29 16:09:35 +0100451}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200452EXPORT_SYMBOL_GPL(kvm_set_cr3);
Carsten Ottea03490e2007-10-29 16:09:35 +0100453
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200454void kvm_set_cr8(struct kvm_vcpu *vcpu, unsigned long cr8)
Carsten Ottea03490e2007-10-29 16:09:35 +0100455{
456 if (cr8 & CR8_RESERVED_BITS) {
457 printk(KERN_DEBUG "set_cr8: #GP, reserved bits 0x%lx\n", cr8);
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200458 kvm_inject_gp(vcpu, 0);
Carsten Ottea03490e2007-10-29 16:09:35 +0100459 return;
460 }
461 if (irqchip_in_kernel(vcpu->kvm))
462 kvm_lapic_set_tpr(vcpu, cr8);
463 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800464 vcpu->arch.cr8 = cr8;
Carsten Ottea03490e2007-10-29 16:09:35 +0100465}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200466EXPORT_SYMBOL_GPL(kvm_set_cr8);
Carsten Ottea03490e2007-10-29 16:09:35 +0100467
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200468unsigned long kvm_get_cr8(struct kvm_vcpu *vcpu)
Carsten Ottea03490e2007-10-29 16:09:35 +0100469{
470 if (irqchip_in_kernel(vcpu->kvm))
471 return kvm_lapic_get_cr8(vcpu);
472 else
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800473 return vcpu->arch.cr8;
Carsten Ottea03490e2007-10-29 16:09:35 +0100474}
Avi Kivity2d3ad1f2008-02-24 11:20:43 +0200475EXPORT_SYMBOL_GPL(kvm_get_cr8);
Carsten Ottea03490e2007-10-29 16:09:35 +0100476
Alexander Grafd8017472008-11-25 20:17:11 +0100477static inline u32 bit(int bitno)
478{
479 return 1 << (bitno & 31);
480}
481
Carsten Otte043405e2007-10-10 17:16:19 +0200482/*
483 * List of msr numbers which we expose to userspace through KVM_GET_MSRS
484 * and KVM_SET_MSRS, and KVM_GET_MSR_INDEX_LIST.
485 *
486 * This list is modified at module load time to reflect the
487 * capabilities of the host cpu.
488 */
489static u32 msrs_to_save[] = {
490 MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP,
491 MSR_K6_STAR,
492#ifdef CONFIG_X86_64
493 MSR_CSTAR, MSR_KERNEL_GS_BASE, MSR_SYSCALL_MASK, MSR_LSTAR,
494#endif
Jaswinder Singh Rajputaf24a4e2009-05-15 18:42:05 +0530495 MSR_IA32_TSC, MSR_KVM_SYSTEM_TIME, MSR_KVM_WALL_CLOCK,
Alexander Grafb286d5d2008-11-25 20:17:05 +0100496 MSR_IA32_PERF_STATUS, MSR_IA32_CR_PAT, MSR_VM_HSAVE_PA
Carsten Otte043405e2007-10-10 17:16:19 +0200497};
498
499static unsigned num_msrs_to_save;
500
501static u32 emulated_msrs[] = {
502 MSR_IA32_MISC_ENABLE,
503};
504
Carsten Otte15c4a642007-10-30 18:44:17 +0100505static void set_efer(struct kvm_vcpu *vcpu, u64 efer)
506{
Joerg Roedelf2b4b7d2008-01-31 14:57:37 +0100507 if (efer & efer_reserved_bits) {
Carsten Otte15c4a642007-10-30 18:44:17 +0100508 printk(KERN_DEBUG "set_efer: 0x%llx #GP, reserved bits\n",
509 efer);
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200510 kvm_inject_gp(vcpu, 0);
Carsten Otte15c4a642007-10-30 18:44:17 +0100511 return;
512 }
513
514 if (is_paging(vcpu)
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800515 && (vcpu->arch.shadow_efer & EFER_LME) != (efer & EFER_LME)) {
Carsten Otte15c4a642007-10-30 18:44:17 +0100516 printk(KERN_DEBUG "set_efer: #GP, change LME while paging\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +0200517 kvm_inject_gp(vcpu, 0);
Carsten Otte15c4a642007-10-30 18:44:17 +0100518 return;
519 }
520
Alexander Graf1b2fd702009-02-02 16:23:51 +0100521 if (efer & EFER_FFXSR) {
522 struct kvm_cpuid_entry2 *feat;
523
524 feat = kvm_find_cpuid_entry(vcpu, 0x80000001, 0);
525 if (!feat || !(feat->edx & bit(X86_FEATURE_FXSR_OPT))) {
526 printk(KERN_DEBUG "set_efer: #GP, enable FFXSR w/o CPUID capability\n");
527 kvm_inject_gp(vcpu, 0);
528 return;
529 }
530 }
531
Alexander Grafd8017472008-11-25 20:17:11 +0100532 if (efer & EFER_SVME) {
533 struct kvm_cpuid_entry2 *feat;
534
535 feat = kvm_find_cpuid_entry(vcpu, 0x80000001, 0);
536 if (!feat || !(feat->ecx & bit(X86_FEATURE_SVM))) {
537 printk(KERN_DEBUG "set_efer: #GP, enable SVM w/o SVM\n");
538 kvm_inject_gp(vcpu, 0);
539 return;
540 }
541 }
542
Carsten Otte15c4a642007-10-30 18:44:17 +0100543 kvm_x86_ops->set_efer(vcpu, efer);
544
545 efer &= ~EFER_LMA;
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800546 efer |= vcpu->arch.shadow_efer & EFER_LMA;
Carsten Otte15c4a642007-10-30 18:44:17 +0100547
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800548 vcpu->arch.shadow_efer = efer;
Avi Kivity9645bb562009-03-31 11:31:54 +0300549
550 vcpu->arch.mmu.base_role.nxe = (efer & EFER_NX) && !tdp_enabled;
551 kvm_mmu_reset_context(vcpu);
Carsten Otte15c4a642007-10-30 18:44:17 +0100552}
553
Joerg Roedelf2b4b7d2008-01-31 14:57:37 +0100554void kvm_enable_efer_bits(u64 mask)
555{
556 efer_reserved_bits &= ~mask;
557}
558EXPORT_SYMBOL_GPL(kvm_enable_efer_bits);
559
560
Carsten Otte15c4a642007-10-30 18:44:17 +0100561/*
562 * Writes msr value into into the appropriate "register".
563 * Returns 0 on success, non-0 otherwise.
564 * Assumes vcpu_load() was already called.
565 */
566int kvm_set_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 data)
567{
568 return kvm_x86_ops->set_msr(vcpu, msr_index, data);
569}
570
Carsten Otte313a3dc2007-10-11 19:16:52 +0200571/*
572 * Adapt set_msr() to msr_io()'s calling convention
573 */
574static int do_set_msr(struct kvm_vcpu *vcpu, unsigned index, u64 *data)
575{
576 return kvm_set_msr(vcpu, index, *data);
577}
578
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200579static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock)
580{
581 static int version;
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200582 struct pvclock_wall_clock wc;
583 struct timespec now, sys, boot;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200584
585 if (!wall_clock)
586 return;
587
588 version++;
589
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200590 kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
591
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200592 /*
593 * The guest calculates current wall clock time by adding
594 * system time (updated by kvm_write_guest_time below) to the
595 * wall clock specified here. guest system time equals host
596 * system time for us, thus we must fill in host boot time here.
597 */
598 now = current_kernel_time();
599 ktime_get_ts(&sys);
600 boot = ns_to_timespec(timespec_to_ns(&now) - timespec_to_ns(&sys));
601
602 wc.sec = boot.tv_sec;
603 wc.nsec = boot.tv_nsec;
604 wc.version = version;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200605
606 kvm_write_guest(kvm, wall_clock, &wc, sizeof(wc));
607
608 version++;
609 kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200610}
611
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200612static uint32_t div_frac(uint32_t dividend, uint32_t divisor)
613{
614 uint32_t quotient, remainder;
615
616 /* Don't try to replace with do_div(), this one calculates
617 * "(dividend << 32) / divisor" */
618 __asm__ ( "divl %4"
619 : "=a" (quotient), "=d" (remainder)
620 : "0" (0), "1" (dividend), "r" (divisor) );
621 return quotient;
622}
623
624static void kvm_set_time_scale(uint32_t tsc_khz, struct pvclock_vcpu_time_info *hv_clock)
625{
626 uint64_t nsecs = 1000000000LL;
627 int32_t shift = 0;
628 uint64_t tps64;
629 uint32_t tps32;
630
631 tps64 = tsc_khz * 1000LL;
632 while (tps64 > nsecs*2) {
633 tps64 >>= 1;
634 shift--;
635 }
636
637 tps32 = (uint32_t)tps64;
638 while (tps32 <= (uint32_t)nsecs) {
639 tps32 <<= 1;
640 shift++;
641 }
642
643 hv_clock->tsc_shift = shift;
644 hv_clock->tsc_to_system_mul = div_frac(nsecs, tps32);
645
646 pr_debug("%s: tsc_khz %u, tsc_shift %d, tsc_mul %u\n",
Harvey Harrison80a914d2008-10-15 22:01:25 -0700647 __func__, tsc_khz, hv_clock->tsc_shift,
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200648 hv_clock->tsc_to_system_mul);
649}
650
Gerd Hoffmannc8076602009-02-04 17:52:04 +0100651static DEFINE_PER_CPU(unsigned long, cpu_tsc_khz);
652
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200653static void kvm_write_guest_time(struct kvm_vcpu *v)
654{
655 struct timespec ts;
656 unsigned long flags;
657 struct kvm_vcpu_arch *vcpu = &v->arch;
658 void *shared_kaddr;
Avi Kivity463656c2009-04-12 15:49:07 +0300659 unsigned long this_tsc_khz;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200660
661 if ((!vcpu->time_page))
662 return;
663
Avi Kivity463656c2009-04-12 15:49:07 +0300664 this_tsc_khz = get_cpu_var(cpu_tsc_khz);
665 if (unlikely(vcpu->hv_clock_tsc_khz != this_tsc_khz)) {
666 kvm_set_time_scale(this_tsc_khz, &vcpu->hv_clock);
667 vcpu->hv_clock_tsc_khz = this_tsc_khz;
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200668 }
Avi Kivity463656c2009-04-12 15:49:07 +0300669 put_cpu_var(cpu_tsc_khz);
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200670
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200671 /* Keep irq disabled to prevent changes to the clock */
672 local_irq_save(flags);
Jaswinder Singh Rajputaf24a4e2009-05-15 18:42:05 +0530673 kvm_get_msr(v, MSR_IA32_TSC, &vcpu->hv_clock.tsc_timestamp);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200674 ktime_get_ts(&ts);
675 local_irq_restore(flags);
676
677 /* With all the info we got, fill in the values */
678
679 vcpu->hv_clock.system_time = ts.tv_nsec +
680 (NSEC_PER_SEC * (u64)ts.tv_sec);
681 /*
682 * The interface expects us to write an even number signaling that the
683 * update is finished. Since the guest won't see the intermediate
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200684 * state, we just increase by 2 at the end.
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200685 */
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200686 vcpu->hv_clock.version += 2;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200687
688 shared_kaddr = kmap_atomic(vcpu->time_page, KM_USER0);
689
690 memcpy(shared_kaddr + vcpu->time_offset, &vcpu->hv_clock,
Gerd Hoffmann50d0a0f2008-06-03 16:17:31 +0200691 sizeof(vcpu->hv_clock));
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200692
693 kunmap_atomic(shared_kaddr, KM_USER0);
694
695 mark_page_dirty(v->kvm, vcpu->time >> PAGE_SHIFT);
696}
697
Gerd Hoffmannc8076602009-02-04 17:52:04 +0100698static int kvm_request_guest_time_update(struct kvm_vcpu *v)
699{
700 struct kvm_vcpu_arch *vcpu = &v->arch;
701
702 if (!vcpu->time_page)
703 return 0;
704 set_bit(KVM_REQ_KVMCLOCK_UPDATE, &v->requests);
705 return 1;
706}
707
Avi Kivity9ba075a2008-05-26 20:06:35 +0300708static bool msr_mtrr_valid(unsigned msr)
709{
710 switch (msr) {
711 case 0x200 ... 0x200 + 2 * KVM_NR_VAR_MTRR - 1:
712 case MSR_MTRRfix64K_00000:
713 case MSR_MTRRfix16K_80000:
714 case MSR_MTRRfix16K_A0000:
715 case MSR_MTRRfix4K_C0000:
716 case MSR_MTRRfix4K_C8000:
717 case MSR_MTRRfix4K_D0000:
718 case MSR_MTRRfix4K_D8000:
719 case MSR_MTRRfix4K_E0000:
720 case MSR_MTRRfix4K_E8000:
721 case MSR_MTRRfix4K_F0000:
722 case MSR_MTRRfix4K_F8000:
723 case MSR_MTRRdefType:
724 case MSR_IA32_CR_PAT:
725 return true;
726 case 0x2f8:
727 return true;
728 }
729 return false;
730}
731
Marcelo Tosattid6289b92009-06-22 15:27:56 -0300732static bool valid_pat_type(unsigned t)
733{
734 return t < 8 && (1 << t) & 0xf3; /* 0, 1, 4, 5, 6, 7 */
735}
736
737static bool valid_mtrr_type(unsigned t)
738{
739 return t < 8 && (1 << t) & 0x73; /* 0, 1, 4, 5, 6 */
740}
741
742static bool mtrr_valid(struct kvm_vcpu *vcpu, u32 msr, u64 data)
743{
744 int i;
745
746 if (!msr_mtrr_valid(msr))
747 return false;
748
749 if (msr == MSR_IA32_CR_PAT) {
750 for (i = 0; i < 8; i++)
751 if (!valid_pat_type((data >> (i * 8)) & 0xff))
752 return false;
753 return true;
754 } else if (msr == MSR_MTRRdefType) {
755 if (data & ~0xcff)
756 return false;
757 return valid_mtrr_type(data & 0xff);
758 } else if (msr >= MSR_MTRRfix64K_00000 && msr <= MSR_MTRRfix4K_F8000) {
759 for (i = 0; i < 8 ; i++)
760 if (!valid_mtrr_type((data >> (i * 8)) & 0xff))
761 return false;
762 return true;
763 }
764
765 /* variable MTRRs */
766 return valid_mtrr_type(data & 0xff);
767}
768
Avi Kivity9ba075a2008-05-26 20:06:35 +0300769static int set_msr_mtrr(struct kvm_vcpu *vcpu, u32 msr, u64 data)
770{
Sheng Yang0bed3b52008-10-09 16:01:54 +0800771 u64 *p = (u64 *)&vcpu->arch.mtrr_state.fixed_ranges;
772
Marcelo Tosattid6289b92009-06-22 15:27:56 -0300773 if (!mtrr_valid(vcpu, msr, data))
Avi Kivity9ba075a2008-05-26 20:06:35 +0300774 return 1;
775
Sheng Yang0bed3b52008-10-09 16:01:54 +0800776 if (msr == MSR_MTRRdefType) {
777 vcpu->arch.mtrr_state.def_type = data;
778 vcpu->arch.mtrr_state.enabled = (data & 0xc00) >> 10;
779 } else if (msr == MSR_MTRRfix64K_00000)
780 p[0] = data;
781 else if (msr == MSR_MTRRfix16K_80000 || msr == MSR_MTRRfix16K_A0000)
782 p[1 + msr - MSR_MTRRfix16K_80000] = data;
783 else if (msr >= MSR_MTRRfix4K_C0000 && msr <= MSR_MTRRfix4K_F8000)
784 p[3 + msr - MSR_MTRRfix4K_C0000] = data;
785 else if (msr == MSR_IA32_CR_PAT)
786 vcpu->arch.pat = data;
787 else { /* Variable MTRRs */
788 int idx, is_mtrr_mask;
789 u64 *pt;
790
791 idx = (msr - 0x200) / 2;
792 is_mtrr_mask = msr - 0x200 - 2 * idx;
793 if (!is_mtrr_mask)
794 pt =
795 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].base_lo;
796 else
797 pt =
798 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].mask_lo;
799 *pt = data;
800 }
801
802 kvm_mmu_reset_context(vcpu);
Avi Kivity9ba075a2008-05-26 20:06:35 +0300803 return 0;
804}
Carsten Otte15c4a642007-10-30 18:44:17 +0100805
Huang Ying890ca9a2009-05-11 16:48:15 +0800806static int set_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 data)
807{
808 u64 mcg_cap = vcpu->arch.mcg_cap;
809 unsigned bank_num = mcg_cap & 0xff;
810
811 switch (msr) {
812 case MSR_IA32_MCG_STATUS:
813 vcpu->arch.mcg_status = data;
814 break;
815 case MSR_IA32_MCG_CTL:
816 if (!(mcg_cap & MCG_CTL_P))
817 return 1;
818 if (data != 0 && data != ~(u64)0)
819 return -1;
820 vcpu->arch.mcg_ctl = data;
821 break;
822 default:
823 if (msr >= MSR_IA32_MC0_CTL &&
824 msr < MSR_IA32_MC0_CTL + 4 * bank_num) {
825 u32 offset = msr - MSR_IA32_MC0_CTL;
826 /* only 0 or all 1s can be written to IA32_MCi_CTL */
827 if ((offset & 0x3) == 0 &&
828 data != 0 && data != ~(u64)0)
829 return -1;
830 vcpu->arch.mce_banks[offset] = data;
831 break;
832 }
833 return 1;
834 }
835 return 0;
836}
837
Carsten Otte15c4a642007-10-30 18:44:17 +0100838int kvm_set_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 data)
839{
840 switch (msr) {
Carsten Otte15c4a642007-10-30 18:44:17 +0100841 case MSR_EFER:
842 set_efer(vcpu, data);
843 break;
Andre Przywara8f1589d2009-06-24 12:44:33 +0200844 case MSR_K7_HWCR:
845 data &= ~(u64)0x40; /* ignore flush filter disable */
846 if (data != 0) {
847 pr_unimpl(vcpu, "unimplemented HWCR wrmsr: 0x%llx\n",
848 data);
849 return 1;
850 }
Carsten Otte15c4a642007-10-30 18:44:17 +0100851 break;
Andre Przywaraf7c6d142009-07-02 15:04:14 +0200852 case MSR_FAM10H_MMIO_CONF_BASE:
853 if (data != 0) {
854 pr_unimpl(vcpu, "unimplemented MMIO_CONF_BASE wrmsr: "
855 "0x%llx\n", data);
856 return 1;
857 }
Carsten Otte15c4a642007-10-30 18:44:17 +0100858 break;
Andre Przywarac323c0e2009-06-24 15:37:05 +0200859 case MSR_AMD64_NB_CFG:
Joerg Roedelc7ac6792008-02-11 20:28:27 +0100860 break;
Alexander Grafb5e2fec2008-07-22 08:00:45 +0200861 case MSR_IA32_DEBUGCTLMSR:
862 if (!data) {
863 /* We support the non-activated case already */
864 break;
865 } else if (data & ~(DEBUGCTLMSR_LBR | DEBUGCTLMSR_BTF)) {
866 /* Values other than LBR and BTF are vendor-specific,
867 thus reserved and should throw a #GP */
868 return 1;
869 }
870 pr_unimpl(vcpu, "%s: MSR_IA32_DEBUGCTLMSR 0x%llx, nop\n",
871 __func__, data);
872 break;
Carsten Otte15c4a642007-10-30 18:44:17 +0100873 case MSR_IA32_UCODE_REV:
874 case MSR_IA32_UCODE_WRITE:
Avi Kivity61a6bd62008-12-29 17:32:28 +0200875 case MSR_VM_HSAVE_PA:
Andre Przywara6098ca92009-07-03 16:00:14 +0200876 case MSR_AMD64_PATCH_LOADER:
Carsten Otte15c4a642007-10-30 18:44:17 +0100877 break;
Avi Kivity9ba075a2008-05-26 20:06:35 +0300878 case 0x200 ... 0x2ff:
879 return set_msr_mtrr(vcpu, msr, data);
Carsten Otte15c4a642007-10-30 18:44:17 +0100880 case MSR_IA32_APICBASE:
881 kvm_set_apic_base(vcpu, data);
882 break;
Gleb Natapov0105d1a2009-07-05 17:39:36 +0300883 case APIC_BASE_MSR ... APIC_BASE_MSR + 0x3ff:
884 return kvm_x2apic_msr_write(vcpu, msr, data);
Carsten Otte15c4a642007-10-30 18:44:17 +0100885 case MSR_IA32_MISC_ENABLE:
Zhang Xiantaoad312c72007-12-13 23:50:52 +0800886 vcpu->arch.ia32_misc_enable_msr = data;
Carsten Otte15c4a642007-10-30 18:44:17 +0100887 break;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200888 case MSR_KVM_WALL_CLOCK:
889 vcpu->kvm->arch.wall_clock = data;
890 kvm_write_wall_clock(vcpu->kvm, data);
891 break;
892 case MSR_KVM_SYSTEM_TIME: {
893 if (vcpu->arch.time_page) {
894 kvm_release_page_dirty(vcpu->arch.time_page);
895 vcpu->arch.time_page = NULL;
896 }
897
898 vcpu->arch.time = data;
899
900 /* we verify if the enable bit is set... */
901 if (!(data & 1))
902 break;
903
904 /* ...but clean it before doing the actual write */
905 vcpu->arch.time_offset = data & ~(PAGE_MASK | 1);
906
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200907 vcpu->arch.time_page =
908 gfn_to_page(vcpu->kvm, data >> PAGE_SHIFT);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200909
910 if (is_error_page(vcpu->arch.time_page)) {
911 kvm_release_page_clean(vcpu->arch.time_page);
912 vcpu->arch.time_page = NULL;
913 }
914
Gerd Hoffmannc8076602009-02-04 17:52:04 +0100915 kvm_request_guest_time_update(vcpu);
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -0200916 break;
917 }
Huang Ying890ca9a2009-05-11 16:48:15 +0800918 case MSR_IA32_MCG_CTL:
919 case MSR_IA32_MCG_STATUS:
920 case MSR_IA32_MC0_CTL ... MSR_IA32_MC0_CTL + 4 * KVM_MAX_MCE_BANKS - 1:
921 return set_msr_mce(vcpu, msr, data);
Andre Przywara71db6022009-06-12 22:01:29 +0200922
923 /* Performance counters are not protected by a CPUID bit,
924 * so we should check all of them in the generic path for the sake of
925 * cross vendor migration.
926 * Writing a zero into the event select MSRs disables them,
927 * which we perfectly emulate ;-). Any other value should be at least
928 * reported, some guests depend on them.
929 */
930 case MSR_P6_EVNTSEL0:
931 case MSR_P6_EVNTSEL1:
932 case MSR_K7_EVNTSEL0:
933 case MSR_K7_EVNTSEL1:
934 case MSR_K7_EVNTSEL2:
935 case MSR_K7_EVNTSEL3:
936 if (data != 0)
937 pr_unimpl(vcpu, "unimplemented perfctr wrmsr: "
938 "0x%x data 0x%llx\n", msr, data);
939 break;
940 /* at least RHEL 4 unconditionally writes to the perfctr registers,
941 * so we ignore writes to make it happy.
942 */
943 case MSR_P6_PERFCTR0:
944 case MSR_P6_PERFCTR1:
945 case MSR_K7_PERFCTR0:
946 case MSR_K7_PERFCTR1:
947 case MSR_K7_PERFCTR2:
948 case MSR_K7_PERFCTR3:
949 pr_unimpl(vcpu, "unimplemented perfctr wrmsr: "
950 "0x%x data 0x%llx\n", msr, data);
951 break;
Carsten Otte15c4a642007-10-30 18:44:17 +0100952 default:
Andre Przywaraed85c062009-06-25 12:36:49 +0200953 if (!ignore_msrs) {
954 pr_unimpl(vcpu, "unhandled wrmsr: 0x%x data %llx\n",
955 msr, data);
956 return 1;
957 } else {
958 pr_unimpl(vcpu, "ignored wrmsr: 0x%x data %llx\n",
959 msr, data);
960 break;
961 }
Carsten Otte15c4a642007-10-30 18:44:17 +0100962 }
963 return 0;
964}
965EXPORT_SYMBOL_GPL(kvm_set_msr_common);
966
967
968/*
969 * Reads an msr value (of 'msr_index') into 'pdata'.
970 * Returns 0 on success, non-0 otherwise.
971 * Assumes vcpu_load() was already called.
972 */
973int kvm_get_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 *pdata)
974{
975 return kvm_x86_ops->get_msr(vcpu, msr_index, pdata);
976}
977
Avi Kivity9ba075a2008-05-26 20:06:35 +0300978static int get_msr_mtrr(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
979{
Sheng Yang0bed3b52008-10-09 16:01:54 +0800980 u64 *p = (u64 *)&vcpu->arch.mtrr_state.fixed_ranges;
981
Avi Kivity9ba075a2008-05-26 20:06:35 +0300982 if (!msr_mtrr_valid(msr))
983 return 1;
984
Sheng Yang0bed3b52008-10-09 16:01:54 +0800985 if (msr == MSR_MTRRdefType)
986 *pdata = vcpu->arch.mtrr_state.def_type +
987 (vcpu->arch.mtrr_state.enabled << 10);
988 else if (msr == MSR_MTRRfix64K_00000)
989 *pdata = p[0];
990 else if (msr == MSR_MTRRfix16K_80000 || msr == MSR_MTRRfix16K_A0000)
991 *pdata = p[1 + msr - MSR_MTRRfix16K_80000];
992 else if (msr >= MSR_MTRRfix4K_C0000 && msr <= MSR_MTRRfix4K_F8000)
993 *pdata = p[3 + msr - MSR_MTRRfix4K_C0000];
994 else if (msr == MSR_IA32_CR_PAT)
995 *pdata = vcpu->arch.pat;
996 else { /* Variable MTRRs */
997 int idx, is_mtrr_mask;
998 u64 *pt;
999
1000 idx = (msr - 0x200) / 2;
1001 is_mtrr_mask = msr - 0x200 - 2 * idx;
1002 if (!is_mtrr_mask)
1003 pt =
1004 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].base_lo;
1005 else
1006 pt =
1007 (u64 *)&vcpu->arch.mtrr_state.var_ranges[idx].mask_lo;
1008 *pdata = *pt;
1009 }
1010
Avi Kivity9ba075a2008-05-26 20:06:35 +03001011 return 0;
1012}
1013
Huang Ying890ca9a2009-05-11 16:48:15 +08001014static int get_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1015{
1016 u64 data;
1017 u64 mcg_cap = vcpu->arch.mcg_cap;
1018 unsigned bank_num = mcg_cap & 0xff;
1019
1020 switch (msr) {
1021 case MSR_IA32_P5_MC_ADDR:
1022 case MSR_IA32_P5_MC_TYPE:
1023 data = 0;
1024 break;
1025 case MSR_IA32_MCG_CAP:
1026 data = vcpu->arch.mcg_cap;
1027 break;
1028 case MSR_IA32_MCG_CTL:
1029 if (!(mcg_cap & MCG_CTL_P))
1030 return 1;
1031 data = vcpu->arch.mcg_ctl;
1032 break;
1033 case MSR_IA32_MCG_STATUS:
1034 data = vcpu->arch.mcg_status;
1035 break;
1036 default:
1037 if (msr >= MSR_IA32_MC0_CTL &&
1038 msr < MSR_IA32_MC0_CTL + 4 * bank_num) {
1039 u32 offset = msr - MSR_IA32_MC0_CTL;
1040 data = vcpu->arch.mce_banks[offset];
1041 break;
1042 }
1043 return 1;
1044 }
1045 *pdata = data;
1046 return 0;
1047}
1048
Carsten Otte15c4a642007-10-30 18:44:17 +01001049int kvm_get_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata)
1050{
1051 u64 data;
1052
1053 switch (msr) {
Carsten Otte15c4a642007-10-30 18:44:17 +01001054 case MSR_IA32_PLATFORM_ID:
Carsten Otte15c4a642007-10-30 18:44:17 +01001055 case MSR_IA32_UCODE_REV:
Carsten Otte15c4a642007-10-30 18:44:17 +01001056 case MSR_IA32_EBL_CR_POWERON:
Alexander Grafb5e2fec2008-07-22 08:00:45 +02001057 case MSR_IA32_DEBUGCTLMSR:
1058 case MSR_IA32_LASTBRANCHFROMIP:
1059 case MSR_IA32_LASTBRANCHTOIP:
1060 case MSR_IA32_LASTINTFROMIP:
1061 case MSR_IA32_LASTINTTOIP:
Jaswinder Singh Rajput60af2ec2009-05-14 11:00:10 +05301062 case MSR_K8_SYSCFG:
1063 case MSR_K7_HWCR:
Avi Kivity61a6bd62008-12-29 17:32:28 +02001064 case MSR_VM_HSAVE_PA:
Amit Shah1f3ee612009-06-30 16:24:28 +05301065 case MSR_P6_PERFCTR0:
1066 case MSR_P6_PERFCTR1:
Amit Shah7fe29e02009-03-20 12:39:00 +05301067 case MSR_P6_EVNTSEL0:
1068 case MSR_P6_EVNTSEL1:
Amit Shah9e699622009-06-15 13:25:34 +05301069 case MSR_K7_EVNTSEL0:
Amit Shah1f3ee612009-06-30 16:24:28 +05301070 case MSR_K7_PERFCTR0:
Andre Przywara1fdbd482009-06-24 12:44:34 +02001071 case MSR_K8_INT_PENDING_MSG:
Andre Przywarac323c0e2009-06-24 15:37:05 +02001072 case MSR_AMD64_NB_CFG:
Andre Przywaraf7c6d142009-07-02 15:04:14 +02001073 case MSR_FAM10H_MMIO_CONF_BASE:
Carsten Otte15c4a642007-10-30 18:44:17 +01001074 data = 0;
1075 break;
Avi Kivity9ba075a2008-05-26 20:06:35 +03001076 case MSR_MTRRcap:
1077 data = 0x500 | KVM_NR_VAR_MTRR;
1078 break;
1079 case 0x200 ... 0x2ff:
1080 return get_msr_mtrr(vcpu, msr, pdata);
Carsten Otte15c4a642007-10-30 18:44:17 +01001081 case 0xcd: /* fsb frequency */
1082 data = 3;
1083 break;
1084 case MSR_IA32_APICBASE:
1085 data = kvm_get_apic_base(vcpu);
1086 break;
Gleb Natapov0105d1a2009-07-05 17:39:36 +03001087 case APIC_BASE_MSR ... APIC_BASE_MSR + 0x3ff:
1088 return kvm_x2apic_msr_read(vcpu, msr, pdata);
1089 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001090 case MSR_IA32_MISC_ENABLE:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001091 data = vcpu->arch.ia32_misc_enable_msr;
Carsten Otte15c4a642007-10-30 18:44:17 +01001092 break;
Alexander Graf847f0ad2008-02-21 12:11:01 +01001093 case MSR_IA32_PERF_STATUS:
1094 /* TSC increment by tick */
1095 data = 1000ULL;
1096 /* CPU multiplier */
1097 data |= (((uint64_t)4ULL) << 40);
1098 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001099 case MSR_EFER:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001100 data = vcpu->arch.shadow_efer;
Carsten Otte15c4a642007-10-30 18:44:17 +01001101 break;
Glauber de Oliveira Costa18068522008-02-15 17:52:47 -02001102 case MSR_KVM_WALL_CLOCK:
1103 data = vcpu->kvm->arch.wall_clock;
1104 break;
1105 case MSR_KVM_SYSTEM_TIME:
1106 data = vcpu->arch.time;
1107 break;
Huang Ying890ca9a2009-05-11 16:48:15 +08001108 case MSR_IA32_P5_MC_ADDR:
1109 case MSR_IA32_P5_MC_TYPE:
1110 case MSR_IA32_MCG_CAP:
1111 case MSR_IA32_MCG_CTL:
1112 case MSR_IA32_MCG_STATUS:
1113 case MSR_IA32_MC0_CTL ... MSR_IA32_MC0_CTL + 4 * KVM_MAX_MCE_BANKS - 1:
1114 return get_msr_mce(vcpu, msr, pdata);
Carsten Otte15c4a642007-10-30 18:44:17 +01001115 default:
Andre Przywaraed85c062009-06-25 12:36:49 +02001116 if (!ignore_msrs) {
1117 pr_unimpl(vcpu, "unhandled rdmsr: 0x%x\n", msr);
1118 return 1;
1119 } else {
1120 pr_unimpl(vcpu, "ignored rdmsr: 0x%x\n", msr);
1121 data = 0;
1122 }
1123 break;
Carsten Otte15c4a642007-10-30 18:44:17 +01001124 }
1125 *pdata = data;
1126 return 0;
1127}
1128EXPORT_SYMBOL_GPL(kvm_get_msr_common);
1129
Carsten Otte313a3dc2007-10-11 19:16:52 +02001130/*
1131 * Read or write a bunch of msrs. All parameters are kernel addresses.
1132 *
1133 * @return number of msrs set successfully.
1134 */
1135static int __msr_io(struct kvm_vcpu *vcpu, struct kvm_msrs *msrs,
1136 struct kvm_msr_entry *entries,
1137 int (*do_msr)(struct kvm_vcpu *vcpu,
1138 unsigned index, u64 *data))
1139{
1140 int i;
1141
1142 vcpu_load(vcpu);
1143
Marcelo Tosatti3200f402008-03-29 20:17:59 -03001144 down_read(&vcpu->kvm->slots_lock);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001145 for (i = 0; i < msrs->nmsrs; ++i)
1146 if (do_msr(vcpu, entries[i].index, &entries[i].data))
1147 break;
Marcelo Tosatti3200f402008-03-29 20:17:59 -03001148 up_read(&vcpu->kvm->slots_lock);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001149
1150 vcpu_put(vcpu);
1151
1152 return i;
1153}
1154
1155/*
1156 * Read or write a bunch of msrs. Parameters are user addresses.
1157 *
1158 * @return number of msrs set successfully.
1159 */
1160static int msr_io(struct kvm_vcpu *vcpu, struct kvm_msrs __user *user_msrs,
1161 int (*do_msr)(struct kvm_vcpu *vcpu,
1162 unsigned index, u64 *data),
1163 int writeback)
1164{
1165 struct kvm_msrs msrs;
1166 struct kvm_msr_entry *entries;
1167 int r, n;
1168 unsigned size;
1169
1170 r = -EFAULT;
1171 if (copy_from_user(&msrs, user_msrs, sizeof msrs))
1172 goto out;
1173
1174 r = -E2BIG;
1175 if (msrs.nmsrs >= MAX_IO_MSRS)
1176 goto out;
1177
1178 r = -ENOMEM;
1179 size = sizeof(struct kvm_msr_entry) * msrs.nmsrs;
1180 entries = vmalloc(size);
1181 if (!entries)
1182 goto out;
1183
1184 r = -EFAULT;
1185 if (copy_from_user(entries, user_msrs->entries, size))
1186 goto out_free;
1187
1188 r = n = __msr_io(vcpu, &msrs, entries, do_msr);
1189 if (r < 0)
1190 goto out_free;
1191
1192 r = -EFAULT;
1193 if (writeback && copy_to_user(user_msrs->entries, entries, size))
1194 goto out_free;
1195
1196 r = n;
1197
1198out_free:
1199 vfree(entries);
1200out:
1201 return r;
1202}
1203
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001204int kvm_dev_ioctl_check_extension(long ext)
1205{
1206 int r;
1207
1208 switch (ext) {
1209 case KVM_CAP_IRQCHIP:
1210 case KVM_CAP_HLT:
1211 case KVM_CAP_MMU_SHADOW_CACHE_CONTROL:
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001212 case KVM_CAP_SET_TSS_ADDR:
Dan Kenigsberg07716712007-11-21 17:10:04 +02001213 case KVM_CAP_EXT_CPUID:
Gerd Hoffmannc8076602009-02-04 17:52:04 +01001214 case KVM_CAP_CLOCKSOURCE:
Sheng Yang78376992008-01-28 05:10:22 +08001215 case KVM_CAP_PIT:
Marcelo Tosattia28e4f52008-02-22 12:21:36 -05001216 case KVM_CAP_NOP_IO_DELAY:
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03001217 case KVM_CAP_MP_STATE:
Avi Kivityed848622008-07-29 11:30:57 +03001218 case KVM_CAP_SYNC_MMU:
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02001219 case KVM_CAP_REINJECT_CONTROL:
Gleb Natapov49256632009-02-04 17:28:14 +02001220 case KVM_CAP_IRQ_INJECT_STATUS:
Sheng Yange56d5322009-03-12 21:45:39 +08001221 case KVM_CAP_ASSIGN_DEV_IRQ:
Gregory Haskins721eecbf2009-05-20 10:30:49 -04001222 case KVM_CAP_IRQFD:
Gregory Haskinsd34e6b12009-07-07 17:08:49 -04001223 case KVM_CAP_IOEVENTFD:
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02001224 case KVM_CAP_PIT2:
Beth Kone9f42752009-07-07 11:50:38 -04001225 case KVM_CAP_PIT_STATE2:
Sheng Yangb927a3c2009-07-21 10:42:48 +08001226 case KVM_CAP_SET_IDENTITY_MAP_ADDR:
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001227 r = 1;
1228 break;
Laurent Vivier542472b2008-05-30 16:05:55 +02001229 case KVM_CAP_COALESCED_MMIO:
1230 r = KVM_COALESCED_MMIO_PAGE_OFFSET;
1231 break;
Avi Kivity774ead32007-12-26 13:57:04 +02001232 case KVM_CAP_VAPIC:
1233 r = !kvm_x86_ops->cpu_has_accelerated_tpr();
1234 break;
Avi Kivityf7252302008-02-20 11:53:16 +02001235 case KVM_CAP_NR_VCPUS:
1236 r = KVM_MAX_VCPUS;
1237 break;
Avi Kivitya988b912008-02-20 11:59:20 +02001238 case KVM_CAP_NR_MEMSLOTS:
1239 r = KVM_MEMORY_SLOTS;
1240 break;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05001241 case KVM_CAP_PV_MMU:
1242 r = !tdp_enabled;
1243 break;
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +03001244 case KVM_CAP_IOMMU:
Joerg Roedel19de40a2008-12-03 14:43:34 +01001245 r = iommu_found();
Ben-Ami Yassour62c476c2008-09-14 03:48:28 +03001246 break;
Huang Ying890ca9a2009-05-11 16:48:15 +08001247 case KVM_CAP_MCE:
1248 r = KVM_MAX_MCE_BANKS;
1249 break;
Zhang Xiantao018d00d2007-11-15 23:07:47 +08001250 default:
1251 r = 0;
1252 break;
1253 }
1254 return r;
1255
1256}
1257
Carsten Otte043405e2007-10-10 17:16:19 +02001258long kvm_arch_dev_ioctl(struct file *filp,
1259 unsigned int ioctl, unsigned long arg)
1260{
1261 void __user *argp = (void __user *)arg;
1262 long r;
1263
1264 switch (ioctl) {
1265 case KVM_GET_MSR_INDEX_LIST: {
1266 struct kvm_msr_list __user *user_msr_list = argp;
1267 struct kvm_msr_list msr_list;
1268 unsigned n;
1269
1270 r = -EFAULT;
1271 if (copy_from_user(&msr_list, user_msr_list, sizeof msr_list))
1272 goto out;
1273 n = msr_list.nmsrs;
1274 msr_list.nmsrs = num_msrs_to_save + ARRAY_SIZE(emulated_msrs);
1275 if (copy_to_user(user_msr_list, &msr_list, sizeof msr_list))
1276 goto out;
1277 r = -E2BIG;
Jan Kiszkae125e7b2009-07-02 21:45:47 +02001278 if (n < msr_list.nmsrs)
Carsten Otte043405e2007-10-10 17:16:19 +02001279 goto out;
1280 r = -EFAULT;
1281 if (copy_to_user(user_msr_list->indices, &msrs_to_save,
1282 num_msrs_to_save * sizeof(u32)))
1283 goto out;
Jan Kiszkae125e7b2009-07-02 21:45:47 +02001284 if (copy_to_user(user_msr_list->indices + num_msrs_to_save,
Carsten Otte043405e2007-10-10 17:16:19 +02001285 &emulated_msrs,
1286 ARRAY_SIZE(emulated_msrs) * sizeof(u32)))
1287 goto out;
1288 r = 0;
1289 break;
1290 }
Avi Kivity674eea02008-02-11 18:37:23 +02001291 case KVM_GET_SUPPORTED_CPUID: {
1292 struct kvm_cpuid2 __user *cpuid_arg = argp;
1293 struct kvm_cpuid2 cpuid;
1294
1295 r = -EFAULT;
1296 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
1297 goto out;
1298 r = kvm_dev_ioctl_get_supported_cpuid(&cpuid,
Amit Shah19355472009-01-14 16:56:00 +00001299 cpuid_arg->entries);
Avi Kivity674eea02008-02-11 18:37:23 +02001300 if (r)
1301 goto out;
1302
1303 r = -EFAULT;
1304 if (copy_to_user(cpuid_arg, &cpuid, sizeof cpuid))
1305 goto out;
1306 r = 0;
1307 break;
1308 }
Huang Ying890ca9a2009-05-11 16:48:15 +08001309 case KVM_X86_GET_MCE_CAP_SUPPORTED: {
1310 u64 mce_cap;
1311
1312 mce_cap = KVM_MCE_CAP_SUPPORTED;
1313 r = -EFAULT;
1314 if (copy_to_user(argp, &mce_cap, sizeof mce_cap))
1315 goto out;
1316 r = 0;
1317 break;
1318 }
Carsten Otte043405e2007-10-10 17:16:19 +02001319 default:
1320 r = -EINVAL;
1321 }
1322out:
1323 return r;
1324}
1325
Carsten Otte313a3dc2007-10-11 19:16:52 +02001326void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
1327{
1328 kvm_x86_ops->vcpu_load(vcpu, cpu);
Gerd Hoffmannc8076602009-02-04 17:52:04 +01001329 kvm_request_guest_time_update(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001330}
1331
1332void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
1333{
1334 kvm_x86_ops->vcpu_put(vcpu);
Amit Shah9327fd12007-11-15 18:38:46 +02001335 kvm_put_guest_fpu(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001336}
1337
Dan Kenigsberg07716712007-11-21 17:10:04 +02001338static int is_efer_nx(void)
Carsten Otte313a3dc2007-10-11 19:16:52 +02001339{
Avi Kivitye286e862009-05-03 18:50:55 +03001340 unsigned long long efer = 0;
Carsten Otte313a3dc2007-10-11 19:16:52 +02001341
Avi Kivitye286e862009-05-03 18:50:55 +03001342 rdmsrl_safe(MSR_EFER, &efer);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001343 return efer & EFER_NX;
1344}
1345
1346static void cpuid_fix_nx_cap(struct kvm_vcpu *vcpu)
1347{
1348 int i;
1349 struct kvm_cpuid_entry2 *e, *entry;
1350
Carsten Otte313a3dc2007-10-11 19:16:52 +02001351 entry = NULL;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001352 for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
1353 e = &vcpu->arch.cpuid_entries[i];
Carsten Otte313a3dc2007-10-11 19:16:52 +02001354 if (e->function == 0x80000001) {
1355 entry = e;
1356 break;
1357 }
1358 }
Dan Kenigsberg07716712007-11-21 17:10:04 +02001359 if (entry && (entry->edx & (1 << 20)) && !is_efer_nx()) {
Carsten Otte313a3dc2007-10-11 19:16:52 +02001360 entry->edx &= ~(1 << 20);
1361 printk(KERN_INFO "kvm: guest NX capability removed\n");
1362 }
1363}
1364
Dan Kenigsberg07716712007-11-21 17:10:04 +02001365/* when an old userspace process fills a new kernel module */
Carsten Otte313a3dc2007-10-11 19:16:52 +02001366static int kvm_vcpu_ioctl_set_cpuid(struct kvm_vcpu *vcpu,
1367 struct kvm_cpuid *cpuid,
1368 struct kvm_cpuid_entry __user *entries)
1369{
Dan Kenigsberg07716712007-11-21 17:10:04 +02001370 int r, i;
1371 struct kvm_cpuid_entry *cpuid_entries;
1372
1373 r = -E2BIG;
1374 if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1375 goto out;
1376 r = -ENOMEM;
1377 cpuid_entries = vmalloc(sizeof(struct kvm_cpuid_entry) * cpuid->nent);
1378 if (!cpuid_entries)
1379 goto out;
1380 r = -EFAULT;
1381 if (copy_from_user(cpuid_entries, entries,
1382 cpuid->nent * sizeof(struct kvm_cpuid_entry)))
1383 goto out_free;
1384 for (i = 0; i < cpuid->nent; i++) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001385 vcpu->arch.cpuid_entries[i].function = cpuid_entries[i].function;
1386 vcpu->arch.cpuid_entries[i].eax = cpuid_entries[i].eax;
1387 vcpu->arch.cpuid_entries[i].ebx = cpuid_entries[i].ebx;
1388 vcpu->arch.cpuid_entries[i].ecx = cpuid_entries[i].ecx;
1389 vcpu->arch.cpuid_entries[i].edx = cpuid_entries[i].edx;
1390 vcpu->arch.cpuid_entries[i].index = 0;
1391 vcpu->arch.cpuid_entries[i].flags = 0;
1392 vcpu->arch.cpuid_entries[i].padding[0] = 0;
1393 vcpu->arch.cpuid_entries[i].padding[1] = 0;
1394 vcpu->arch.cpuid_entries[i].padding[2] = 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001395 }
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001396 vcpu->arch.cpuid_nent = cpuid->nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001397 cpuid_fix_nx_cap(vcpu);
1398 r = 0;
Gleb Natapovfc61b802009-07-05 17:39:35 +03001399 kvm_apic_set_version(vcpu);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001400
1401out_free:
1402 vfree(cpuid_entries);
1403out:
1404 return r;
1405}
1406
1407static int kvm_vcpu_ioctl_set_cpuid2(struct kvm_vcpu *vcpu,
Amit Shah19355472009-01-14 16:56:00 +00001408 struct kvm_cpuid2 *cpuid,
1409 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001410{
Carsten Otte313a3dc2007-10-11 19:16:52 +02001411 int r;
1412
1413 r = -E2BIG;
1414 if (cpuid->nent > KVM_MAX_CPUID_ENTRIES)
1415 goto out;
1416 r = -EFAULT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001417 if (copy_from_user(&vcpu->arch.cpuid_entries, entries,
Dan Kenigsberg07716712007-11-21 17:10:04 +02001418 cpuid->nent * sizeof(struct kvm_cpuid_entry2)))
Carsten Otte313a3dc2007-10-11 19:16:52 +02001419 goto out;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001420 vcpu->arch.cpuid_nent = cpuid->nent;
Gleb Natapovfc61b802009-07-05 17:39:35 +03001421 kvm_apic_set_version(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001422 return 0;
1423
1424out:
1425 return r;
1426}
1427
Dan Kenigsberg07716712007-11-21 17:10:04 +02001428static int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,
Amit Shah19355472009-01-14 16:56:00 +00001429 struct kvm_cpuid2 *cpuid,
1430 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001431{
1432 int r;
1433
1434 r = -E2BIG;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001435 if (cpuid->nent < vcpu->arch.cpuid_nent)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001436 goto out;
1437 r = -EFAULT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001438 if (copy_to_user(entries, &vcpu->arch.cpuid_entries,
Amit Shah19355472009-01-14 16:56:00 +00001439 vcpu->arch.cpuid_nent * sizeof(struct kvm_cpuid_entry2)))
Dan Kenigsberg07716712007-11-21 17:10:04 +02001440 goto out;
1441 return 0;
1442
1443out:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001444 cpuid->nent = vcpu->arch.cpuid_nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001445 return r;
1446}
1447
Dan Kenigsberg07716712007-11-21 17:10:04 +02001448static void do_cpuid_1_ent(struct kvm_cpuid_entry2 *entry, u32 function,
Amit Shah19355472009-01-14 16:56:00 +00001449 u32 index)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001450{
1451 entry->function = function;
1452 entry->index = index;
1453 cpuid_count(entry->function, entry->index,
Amit Shah19355472009-01-14 16:56:00 +00001454 &entry->eax, &entry->ebx, &entry->ecx, &entry->edx);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001455 entry->flags = 0;
1456}
1457
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001458#define F(x) bit(X86_FEATURE_##x)
1459
Dan Kenigsberg07716712007-11-21 17:10:04 +02001460static void do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
1461 u32 index, int *nent, int maxnent)
1462{
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001463 unsigned f_nx = is_efer_nx() ? F(NX) : 0;
Joerg Roedel344f4142009-07-27 16:30:48 +02001464 unsigned f_gbpages = kvm_x86_ops->gb_page_enable() ? F(GBPAGES) : 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001465#ifdef CONFIG_X86_64
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001466 unsigned f_lm = F(LM);
1467#else
1468 unsigned f_lm = 0;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001469#endif
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001470
1471 /* cpuid 1.edx */
1472 const u32 kvm_supported_word0_x86_features =
1473 F(FPU) | F(VME) | F(DE) | F(PSE) |
1474 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
1475 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SEP) |
1476 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
1477 F(PAT) | F(PSE36) | 0 /* PSN */ | F(CLFLSH) |
1478 0 /* Reserved, DS, ACPI */ | F(MMX) |
1479 F(FXSR) | F(XMM) | F(XMM2) | F(SELFSNOOP) |
1480 0 /* HTT, TM, Reserved, PBE */;
1481 /* cpuid 0x80000001.edx */
1482 const u32 kvm_supported_word1_x86_features =
1483 F(FPU) | F(VME) | F(DE) | F(PSE) |
1484 F(TSC) | F(MSR) | F(PAE) | F(MCE) |
1485 F(CX8) | F(APIC) | 0 /* Reserved */ | F(SYSCALL) |
1486 F(MTRR) | F(PGE) | F(MCA) | F(CMOV) |
1487 F(PAT) | F(PSE36) | 0 /* Reserved */ |
1488 f_nx | 0 /* Reserved */ | F(MMXEXT) | F(MMX) |
Joerg Roedel344f4142009-07-27 16:30:48 +02001489 F(FXSR) | F(FXSR_OPT) | f_gbpages | 0 /* RDTSCP */ |
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001490 0 /* Reserved */ | f_lm | F(3DNOWEXT) | F(3DNOW);
1491 /* cpuid 1.ecx */
1492 const u32 kvm_supported_word4_x86_features =
Avi Kivityd149c732009-05-10 14:41:56 +03001493 F(XMM3) | 0 /* Reserved, DTES64, MONITOR */ |
1494 0 /* DS-CPL, VMX, SMX, EST */ |
1495 0 /* TM2 */ | F(SSSE3) | 0 /* CNXT-ID */ | 0 /* Reserved */ |
1496 0 /* Reserved */ | F(CX16) | 0 /* xTPR Update, PDCM */ |
1497 0 /* Reserved, DCA */ | F(XMM4_1) |
Gleb Natapov0105d1a2009-07-05 17:39:36 +03001498 F(XMM4_2) | F(X2APIC) | F(MOVBE) | F(POPCNT) |
Avi Kivityd149c732009-05-10 14:41:56 +03001499 0 /* Reserved, XSAVE, OSXSAVE */;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001500 /* cpuid 0x80000001.ecx */
Dan Kenigsberg07716712007-11-21 17:10:04 +02001501 const u32 kvm_supported_word6_x86_features =
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001502 F(LAHF_LM) | F(CMP_LEGACY) | F(SVM) | 0 /* ExtApicSpace */ |
1503 F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
1504 F(3DNOWPREFETCH) | 0 /* OSVW */ | 0 /* IBS */ | F(SSE5) |
1505 0 /* SKINIT */ | 0 /* WDT */;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001506
Amit Shah19355472009-01-14 16:56:00 +00001507 /* all calls to cpuid_count() should be made on the same cpu */
Dan Kenigsberg07716712007-11-21 17:10:04 +02001508 get_cpu();
1509 do_cpuid_1_ent(entry, function, index);
1510 ++*nent;
1511
1512 switch (function) {
1513 case 0:
1514 entry->eax = min(entry->eax, (u32)0xb);
1515 break;
1516 case 1:
1517 entry->edx &= kvm_supported_word0_x86_features;
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001518 entry->ecx &= kvm_supported_word4_x86_features;
Gleb Natapov0d1de2d92009-07-12 16:10:55 +03001519 /* we support x2apic emulation even if host does not support
1520 * it since we emulate x2apic in software */
1521 entry->ecx |= F(X2APIC);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001522 break;
1523 /* function 2 entries are STATEFUL. That is, repeated cpuid commands
1524 * may return different values. This forces us to get_cpu() before
1525 * issuing the first command, and also to emulate this annoying behavior
1526 * in kvm_emulate_cpuid() using KVM_CPUID_FLAG_STATE_READ_NEXT */
1527 case 2: {
1528 int t, times = entry->eax & 0xff;
1529
1530 entry->flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
Nitin A Kamble0fdf8e52008-11-05 15:56:21 -08001531 entry->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001532 for (t = 1; t < times && *nent < maxnent; ++t) {
1533 do_cpuid_1_ent(&entry[t], function, 0);
1534 entry[t].flags |= KVM_CPUID_FLAG_STATEFUL_FUNC;
1535 ++*nent;
1536 }
1537 break;
1538 }
1539 /* function 4 and 0xb have additional index. */
1540 case 4: {
Harvey Harrison14af3f32008-02-19 10:25:50 -08001541 int i, cache_type;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001542
1543 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1544 /* read more entries until cache_type is zero */
Harvey Harrison14af3f32008-02-19 10:25:50 -08001545 for (i = 1; *nent < maxnent; ++i) {
1546 cache_type = entry[i - 1].eax & 0x1f;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001547 if (!cache_type)
1548 break;
Harvey Harrison14af3f32008-02-19 10:25:50 -08001549 do_cpuid_1_ent(&entry[i], function, i);
1550 entry[i].flags |=
Dan Kenigsberg07716712007-11-21 17:10:04 +02001551 KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1552 ++*nent;
1553 }
1554 break;
1555 }
1556 case 0xb: {
Harvey Harrison14af3f32008-02-19 10:25:50 -08001557 int i, level_type;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001558
1559 entry->flags |= KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1560 /* read more entries until level_type is zero */
Harvey Harrison14af3f32008-02-19 10:25:50 -08001561 for (i = 1; *nent < maxnent; ++i) {
Nitin A Kamble0853d2c2008-11-05 15:37:36 -08001562 level_type = entry[i - 1].ecx & 0xff00;
Dan Kenigsberg07716712007-11-21 17:10:04 +02001563 if (!level_type)
1564 break;
Harvey Harrison14af3f32008-02-19 10:25:50 -08001565 do_cpuid_1_ent(&entry[i], function, i);
1566 entry[i].flags |=
Dan Kenigsberg07716712007-11-21 17:10:04 +02001567 KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
1568 ++*nent;
1569 }
1570 break;
1571 }
1572 case 0x80000000:
1573 entry->eax = min(entry->eax, 0x8000001a);
1574 break;
1575 case 0x80000001:
1576 entry->edx &= kvm_supported_word1_x86_features;
1577 entry->ecx &= kvm_supported_word6_x86_features;
1578 break;
1579 }
1580 put_cpu();
1581}
1582
Avi Kivity7faa4ee2009-05-10 13:55:35 +03001583#undef F
1584
Avi Kivity674eea02008-02-11 18:37:23 +02001585static int kvm_dev_ioctl_get_supported_cpuid(struct kvm_cpuid2 *cpuid,
Amit Shah19355472009-01-14 16:56:00 +00001586 struct kvm_cpuid_entry2 __user *entries)
Dan Kenigsberg07716712007-11-21 17:10:04 +02001587{
1588 struct kvm_cpuid_entry2 *cpuid_entries;
1589 int limit, nent = 0, r = -E2BIG;
1590 u32 func;
1591
1592 if (cpuid->nent < 1)
1593 goto out;
1594 r = -ENOMEM;
1595 cpuid_entries = vmalloc(sizeof(struct kvm_cpuid_entry2) * cpuid->nent);
1596 if (!cpuid_entries)
1597 goto out;
1598
1599 do_cpuid_ent(&cpuid_entries[0], 0, 0, &nent, cpuid->nent);
1600 limit = cpuid_entries[0].eax;
1601 for (func = 1; func <= limit && nent < cpuid->nent; ++func)
1602 do_cpuid_ent(&cpuid_entries[nent], func, 0,
Amit Shah19355472009-01-14 16:56:00 +00001603 &nent, cpuid->nent);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001604 r = -E2BIG;
1605 if (nent >= cpuid->nent)
1606 goto out_free;
1607
1608 do_cpuid_ent(&cpuid_entries[nent], 0x80000000, 0, &nent, cpuid->nent);
1609 limit = cpuid_entries[nent - 1].eax;
1610 for (func = 0x80000001; func <= limit && nent < cpuid->nent; ++func)
1611 do_cpuid_ent(&cpuid_entries[nent], func, 0,
Amit Shah19355472009-01-14 16:56:00 +00001612 &nent, cpuid->nent);
Mark McLoughlincb007642009-05-12 12:36:44 +01001613 r = -E2BIG;
1614 if (nent >= cpuid->nent)
1615 goto out_free;
1616
Dan Kenigsberg07716712007-11-21 17:10:04 +02001617 r = -EFAULT;
1618 if (copy_to_user(entries, cpuid_entries,
Amit Shah19355472009-01-14 16:56:00 +00001619 nent * sizeof(struct kvm_cpuid_entry2)))
Dan Kenigsberg07716712007-11-21 17:10:04 +02001620 goto out_free;
1621 cpuid->nent = nent;
1622 r = 0;
1623
1624out_free:
1625 vfree(cpuid_entries);
1626out:
1627 return r;
1628}
1629
Carsten Otte313a3dc2007-10-11 19:16:52 +02001630static int kvm_vcpu_ioctl_get_lapic(struct kvm_vcpu *vcpu,
1631 struct kvm_lapic_state *s)
1632{
1633 vcpu_load(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001634 memcpy(s->regs, vcpu->arch.apic->regs, sizeof *s);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001635 vcpu_put(vcpu);
1636
1637 return 0;
1638}
1639
1640static int kvm_vcpu_ioctl_set_lapic(struct kvm_vcpu *vcpu,
1641 struct kvm_lapic_state *s)
1642{
1643 vcpu_load(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08001644 memcpy(vcpu->arch.apic->regs, s->regs, sizeof *s);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001645 kvm_apic_post_state_restore(vcpu);
Gleb Natapovcb142eb2009-08-09 15:17:40 +03001646 update_cr8_intercept(vcpu);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001647 vcpu_put(vcpu);
1648
1649 return 0;
1650}
1651
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08001652static int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu,
1653 struct kvm_interrupt *irq)
1654{
1655 if (irq->irq < 0 || irq->irq >= 256)
1656 return -EINVAL;
1657 if (irqchip_in_kernel(vcpu->kvm))
1658 return -ENXIO;
1659 vcpu_load(vcpu);
1660
Gleb Natapov66fd3f72009-05-11 13:35:50 +03001661 kvm_queue_interrupt(vcpu, irq->irq, false);
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08001662
1663 vcpu_put(vcpu);
1664
1665 return 0;
1666}
1667
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02001668static int kvm_vcpu_ioctl_nmi(struct kvm_vcpu *vcpu)
1669{
1670 vcpu_load(vcpu);
1671 kvm_inject_nmi(vcpu);
1672 vcpu_put(vcpu);
1673
1674 return 0;
1675}
1676
Avi Kivityb209749f2007-10-22 16:50:39 +02001677static int vcpu_ioctl_tpr_access_reporting(struct kvm_vcpu *vcpu,
1678 struct kvm_tpr_access_ctl *tac)
1679{
1680 if (tac->flags)
1681 return -EINVAL;
1682 vcpu->arch.tpr_access_reporting = !!tac->enabled;
1683 return 0;
1684}
1685
Huang Ying890ca9a2009-05-11 16:48:15 +08001686static int kvm_vcpu_ioctl_x86_setup_mce(struct kvm_vcpu *vcpu,
1687 u64 mcg_cap)
1688{
1689 int r;
1690 unsigned bank_num = mcg_cap & 0xff, bank;
1691
1692 r = -EINVAL;
1693 if (!bank_num)
1694 goto out;
1695 if (mcg_cap & ~(KVM_MCE_CAP_SUPPORTED | 0xff | 0xff0000))
1696 goto out;
1697 r = 0;
1698 vcpu->arch.mcg_cap = mcg_cap;
1699 /* Init IA32_MCG_CTL to all 1s */
1700 if (mcg_cap & MCG_CTL_P)
1701 vcpu->arch.mcg_ctl = ~(u64)0;
1702 /* Init IA32_MCi_CTL to all 1s */
1703 for (bank = 0; bank < bank_num; bank++)
1704 vcpu->arch.mce_banks[bank*4] = ~(u64)0;
1705out:
1706 return r;
1707}
1708
1709static int kvm_vcpu_ioctl_x86_set_mce(struct kvm_vcpu *vcpu,
1710 struct kvm_x86_mce *mce)
1711{
1712 u64 mcg_cap = vcpu->arch.mcg_cap;
1713 unsigned bank_num = mcg_cap & 0xff;
1714 u64 *banks = vcpu->arch.mce_banks;
1715
1716 if (mce->bank >= bank_num || !(mce->status & MCI_STATUS_VAL))
1717 return -EINVAL;
1718 /*
1719 * if IA32_MCG_CTL is not all 1s, the uncorrected error
1720 * reporting is disabled
1721 */
1722 if ((mce->status & MCI_STATUS_UC) && (mcg_cap & MCG_CTL_P) &&
1723 vcpu->arch.mcg_ctl != ~(u64)0)
1724 return 0;
1725 banks += 4 * mce->bank;
1726 /*
1727 * if IA32_MCi_CTL is not all 1s, the uncorrected error
1728 * reporting is disabled for the bank
1729 */
1730 if ((mce->status & MCI_STATUS_UC) && banks[0] != ~(u64)0)
1731 return 0;
1732 if (mce->status & MCI_STATUS_UC) {
1733 if ((vcpu->arch.mcg_status & MCG_STATUS_MCIP) ||
1734 !(vcpu->arch.cr4 & X86_CR4_MCE)) {
1735 printk(KERN_DEBUG "kvm: set_mce: "
1736 "injects mce exception while "
1737 "previous one is in progress!\n");
1738 set_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests);
1739 return 0;
1740 }
1741 if (banks[1] & MCI_STATUS_VAL)
1742 mce->status |= MCI_STATUS_OVER;
1743 banks[2] = mce->addr;
1744 banks[3] = mce->misc;
1745 vcpu->arch.mcg_status = mce->mcg_status;
1746 banks[1] = mce->status;
1747 kvm_queue_exception(vcpu, MC_VECTOR);
1748 } else if (!(banks[1] & MCI_STATUS_VAL)
1749 || !(banks[1] & MCI_STATUS_UC)) {
1750 if (banks[1] & MCI_STATUS_VAL)
1751 mce->status |= MCI_STATUS_OVER;
1752 banks[2] = mce->addr;
1753 banks[3] = mce->misc;
1754 banks[1] = mce->status;
1755 } else
1756 banks[1] |= MCI_STATUS_OVER;
1757 return 0;
1758}
1759
Carsten Otte313a3dc2007-10-11 19:16:52 +02001760long kvm_arch_vcpu_ioctl(struct file *filp,
1761 unsigned int ioctl, unsigned long arg)
1762{
1763 struct kvm_vcpu *vcpu = filp->private_data;
1764 void __user *argp = (void __user *)arg;
1765 int r;
Dave Hansenb772ff32008-08-11 10:01:47 -07001766 struct kvm_lapic_state *lapic = NULL;
Carsten Otte313a3dc2007-10-11 19:16:52 +02001767
1768 switch (ioctl) {
1769 case KVM_GET_LAPIC: {
Dave Hansenb772ff32008-08-11 10:01:47 -07001770 lapic = kzalloc(sizeof(struct kvm_lapic_state), GFP_KERNEL);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001771
Dave Hansenb772ff32008-08-11 10:01:47 -07001772 r = -ENOMEM;
1773 if (!lapic)
1774 goto out;
1775 r = kvm_vcpu_ioctl_get_lapic(vcpu, lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001776 if (r)
1777 goto out;
1778 r = -EFAULT;
Dave Hansenb772ff32008-08-11 10:01:47 -07001779 if (copy_to_user(argp, lapic, sizeof(struct kvm_lapic_state)))
Carsten Otte313a3dc2007-10-11 19:16:52 +02001780 goto out;
1781 r = 0;
1782 break;
1783 }
1784 case KVM_SET_LAPIC: {
Dave Hansenb772ff32008-08-11 10:01:47 -07001785 lapic = kmalloc(sizeof(struct kvm_lapic_state), GFP_KERNEL);
1786 r = -ENOMEM;
1787 if (!lapic)
Carsten Otte313a3dc2007-10-11 19:16:52 +02001788 goto out;
Dave Hansenb772ff32008-08-11 10:01:47 -07001789 r = -EFAULT;
1790 if (copy_from_user(lapic, argp, sizeof(struct kvm_lapic_state)))
1791 goto out;
1792 r = kvm_vcpu_ioctl_set_lapic(vcpu, lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001793 if (r)
1794 goto out;
1795 r = 0;
1796 break;
1797 }
Zhang Xiantaof77bc6a2007-11-21 04:36:41 +08001798 case KVM_INTERRUPT: {
1799 struct kvm_interrupt irq;
1800
1801 r = -EFAULT;
1802 if (copy_from_user(&irq, argp, sizeof irq))
1803 goto out;
1804 r = kvm_vcpu_ioctl_interrupt(vcpu, &irq);
1805 if (r)
1806 goto out;
1807 r = 0;
1808 break;
1809 }
Jan Kiszkac4abb7c2008-09-26 09:30:55 +02001810 case KVM_NMI: {
1811 r = kvm_vcpu_ioctl_nmi(vcpu);
1812 if (r)
1813 goto out;
1814 r = 0;
1815 break;
1816 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02001817 case KVM_SET_CPUID: {
1818 struct kvm_cpuid __user *cpuid_arg = argp;
1819 struct kvm_cpuid cpuid;
1820
1821 r = -EFAULT;
1822 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
1823 goto out;
1824 r = kvm_vcpu_ioctl_set_cpuid(vcpu, &cpuid, cpuid_arg->entries);
1825 if (r)
1826 goto out;
1827 break;
1828 }
Dan Kenigsberg07716712007-11-21 17:10:04 +02001829 case KVM_SET_CPUID2: {
1830 struct kvm_cpuid2 __user *cpuid_arg = argp;
1831 struct kvm_cpuid2 cpuid;
1832
1833 r = -EFAULT;
1834 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
1835 goto out;
1836 r = kvm_vcpu_ioctl_set_cpuid2(vcpu, &cpuid,
Amit Shah19355472009-01-14 16:56:00 +00001837 cpuid_arg->entries);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001838 if (r)
1839 goto out;
1840 break;
1841 }
1842 case KVM_GET_CPUID2: {
1843 struct kvm_cpuid2 __user *cpuid_arg = argp;
1844 struct kvm_cpuid2 cpuid;
1845
1846 r = -EFAULT;
1847 if (copy_from_user(&cpuid, cpuid_arg, sizeof cpuid))
1848 goto out;
1849 r = kvm_vcpu_ioctl_get_cpuid2(vcpu, &cpuid,
Amit Shah19355472009-01-14 16:56:00 +00001850 cpuid_arg->entries);
Dan Kenigsberg07716712007-11-21 17:10:04 +02001851 if (r)
1852 goto out;
1853 r = -EFAULT;
1854 if (copy_to_user(cpuid_arg, &cpuid, sizeof cpuid))
1855 goto out;
1856 r = 0;
1857 break;
1858 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02001859 case KVM_GET_MSRS:
1860 r = msr_io(vcpu, argp, kvm_get_msr, 1);
1861 break;
1862 case KVM_SET_MSRS:
1863 r = msr_io(vcpu, argp, do_set_msr, 0);
1864 break;
Avi Kivityb209749f2007-10-22 16:50:39 +02001865 case KVM_TPR_ACCESS_REPORTING: {
1866 struct kvm_tpr_access_ctl tac;
1867
1868 r = -EFAULT;
1869 if (copy_from_user(&tac, argp, sizeof tac))
1870 goto out;
1871 r = vcpu_ioctl_tpr_access_reporting(vcpu, &tac);
1872 if (r)
1873 goto out;
1874 r = -EFAULT;
1875 if (copy_to_user(argp, &tac, sizeof tac))
1876 goto out;
1877 r = 0;
1878 break;
1879 };
Avi Kivityb93463a2007-10-25 16:52:32 +02001880 case KVM_SET_VAPIC_ADDR: {
1881 struct kvm_vapic_addr va;
1882
1883 r = -EINVAL;
1884 if (!irqchip_in_kernel(vcpu->kvm))
1885 goto out;
1886 r = -EFAULT;
1887 if (copy_from_user(&va, argp, sizeof va))
1888 goto out;
1889 r = 0;
1890 kvm_lapic_set_vapic_addr(vcpu, va.vapic_addr);
1891 break;
1892 }
Huang Ying890ca9a2009-05-11 16:48:15 +08001893 case KVM_X86_SETUP_MCE: {
1894 u64 mcg_cap;
1895
1896 r = -EFAULT;
1897 if (copy_from_user(&mcg_cap, argp, sizeof mcg_cap))
1898 goto out;
1899 r = kvm_vcpu_ioctl_x86_setup_mce(vcpu, mcg_cap);
1900 break;
1901 }
1902 case KVM_X86_SET_MCE: {
1903 struct kvm_x86_mce mce;
1904
1905 r = -EFAULT;
1906 if (copy_from_user(&mce, argp, sizeof mce))
1907 goto out;
1908 r = kvm_vcpu_ioctl_x86_set_mce(vcpu, &mce);
1909 break;
1910 }
Carsten Otte313a3dc2007-10-11 19:16:52 +02001911 default:
1912 r = -EINVAL;
1913 }
1914out:
Wei Yongjun7a6ce842009-03-31 16:47:44 +08001915 kfree(lapic);
Carsten Otte313a3dc2007-10-11 19:16:52 +02001916 return r;
1917}
1918
Carsten Otte1fe779f2007-10-29 16:08:35 +01001919static int kvm_vm_ioctl_set_tss_addr(struct kvm *kvm, unsigned long addr)
1920{
1921 int ret;
1922
1923 if (addr > (unsigned int)(-3 * PAGE_SIZE))
1924 return -1;
1925 ret = kvm_x86_ops->set_tss_addr(kvm, addr);
1926 return ret;
1927}
1928
Sheng Yangb927a3c2009-07-21 10:42:48 +08001929static int kvm_vm_ioctl_set_identity_map_addr(struct kvm *kvm,
1930 u64 ident_addr)
1931{
1932 kvm->arch.ept_identity_map_addr = ident_addr;
1933 return 0;
1934}
1935
Carsten Otte1fe779f2007-10-29 16:08:35 +01001936static int kvm_vm_ioctl_set_nr_mmu_pages(struct kvm *kvm,
1937 u32 kvm_nr_mmu_pages)
1938{
1939 if (kvm_nr_mmu_pages < KVM_MIN_ALLOC_MMU_PAGES)
1940 return -EINVAL;
1941
Izik Eidus72dc67a2008-02-10 18:04:15 +02001942 down_write(&kvm->slots_lock);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03001943 spin_lock(&kvm->mmu_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01001944
1945 kvm_mmu_change_mmu_pages(kvm, kvm_nr_mmu_pages);
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08001946 kvm->arch.n_requested_mmu_pages = kvm_nr_mmu_pages;
Carsten Otte1fe779f2007-10-29 16:08:35 +01001947
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03001948 spin_unlock(&kvm->mmu_lock);
Izik Eidus72dc67a2008-02-10 18:04:15 +02001949 up_write(&kvm->slots_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01001950 return 0;
1951}
1952
1953static int kvm_vm_ioctl_get_nr_mmu_pages(struct kvm *kvm)
1954{
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08001955 return kvm->arch.n_alloc_mmu_pages;
Carsten Otte1fe779f2007-10-29 16:08:35 +01001956}
1957
Zhang Xiantaoe9f85cd2007-11-22 11:20:33 +08001958gfn_t unalias_gfn(struct kvm *kvm, gfn_t gfn)
1959{
1960 int i;
1961 struct kvm_mem_alias *alias;
1962
Zhang Xiantaod69fb812007-12-14 09:54:20 +08001963 for (i = 0; i < kvm->arch.naliases; ++i) {
1964 alias = &kvm->arch.aliases[i];
Zhang Xiantaoe9f85cd2007-11-22 11:20:33 +08001965 if (gfn >= alias->base_gfn
1966 && gfn < alias->base_gfn + alias->npages)
1967 return alias->target_gfn + gfn - alias->base_gfn;
1968 }
1969 return gfn;
1970}
1971
Carsten Otte1fe779f2007-10-29 16:08:35 +01001972/*
1973 * Set a new alias region. Aliases map a portion of physical memory into
1974 * another portion. This is useful for memory windows, for example the PC
1975 * VGA region.
1976 */
1977static int kvm_vm_ioctl_set_memory_alias(struct kvm *kvm,
1978 struct kvm_memory_alias *alias)
1979{
1980 int r, n;
1981 struct kvm_mem_alias *p;
1982
1983 r = -EINVAL;
1984 /* General sanity checks */
1985 if (alias->memory_size & (PAGE_SIZE - 1))
1986 goto out;
1987 if (alias->guest_phys_addr & (PAGE_SIZE - 1))
1988 goto out;
1989 if (alias->slot >= KVM_ALIAS_SLOTS)
1990 goto out;
1991 if (alias->guest_phys_addr + alias->memory_size
1992 < alias->guest_phys_addr)
1993 goto out;
1994 if (alias->target_phys_addr + alias->memory_size
1995 < alias->target_phys_addr)
1996 goto out;
1997
Izik Eidus72dc67a2008-02-10 18:04:15 +02001998 down_write(&kvm->slots_lock);
Andrea Arcangelia1708ce2008-07-25 16:26:39 +02001999 spin_lock(&kvm->mmu_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002000
Zhang Xiantaod69fb812007-12-14 09:54:20 +08002001 p = &kvm->arch.aliases[alias->slot];
Carsten Otte1fe779f2007-10-29 16:08:35 +01002002 p->base_gfn = alias->guest_phys_addr >> PAGE_SHIFT;
2003 p->npages = alias->memory_size >> PAGE_SHIFT;
2004 p->target_gfn = alias->target_phys_addr >> PAGE_SHIFT;
2005
2006 for (n = KVM_ALIAS_SLOTS; n > 0; --n)
Zhang Xiantaod69fb812007-12-14 09:54:20 +08002007 if (kvm->arch.aliases[n - 1].npages)
Carsten Otte1fe779f2007-10-29 16:08:35 +01002008 break;
Zhang Xiantaod69fb812007-12-14 09:54:20 +08002009 kvm->arch.naliases = n;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002010
Andrea Arcangelia1708ce2008-07-25 16:26:39 +02002011 spin_unlock(&kvm->mmu_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002012 kvm_mmu_zap_all(kvm);
2013
Izik Eidus72dc67a2008-02-10 18:04:15 +02002014 up_write(&kvm->slots_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002015
2016 return 0;
2017
2018out:
2019 return r;
2020}
2021
2022static int kvm_vm_ioctl_get_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
2023{
2024 int r;
2025
2026 r = 0;
2027 switch (chip->chip_id) {
2028 case KVM_IRQCHIP_PIC_MASTER:
2029 memcpy(&chip->chip.pic,
2030 &pic_irqchip(kvm)->pics[0],
2031 sizeof(struct kvm_pic_state));
2032 break;
2033 case KVM_IRQCHIP_PIC_SLAVE:
2034 memcpy(&chip->chip.pic,
2035 &pic_irqchip(kvm)->pics[1],
2036 sizeof(struct kvm_pic_state));
2037 break;
2038 case KVM_IRQCHIP_IOAPIC:
2039 memcpy(&chip->chip.ioapic,
2040 ioapic_irqchip(kvm),
2041 sizeof(struct kvm_ioapic_state));
2042 break;
2043 default:
2044 r = -EINVAL;
2045 break;
2046 }
2047 return r;
2048}
2049
2050static int kvm_vm_ioctl_set_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
2051{
2052 int r;
2053
2054 r = 0;
2055 switch (chip->chip_id) {
2056 case KVM_IRQCHIP_PIC_MASTER:
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002057 spin_lock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002058 memcpy(&pic_irqchip(kvm)->pics[0],
2059 &chip->chip.pic,
2060 sizeof(struct kvm_pic_state));
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002061 spin_unlock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002062 break;
2063 case KVM_IRQCHIP_PIC_SLAVE:
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002064 spin_lock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002065 memcpy(&pic_irqchip(kvm)->pics[1],
2066 &chip->chip.pic,
2067 sizeof(struct kvm_pic_state));
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002068 spin_unlock(&pic_irqchip(kvm)->lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002069 break;
2070 case KVM_IRQCHIP_IOAPIC:
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002071 mutex_lock(&kvm->irq_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002072 memcpy(ioapic_irqchip(kvm),
2073 &chip->chip.ioapic,
2074 sizeof(struct kvm_ioapic_state));
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002075 mutex_unlock(&kvm->irq_lock);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002076 break;
2077 default:
2078 r = -EINVAL;
2079 break;
2080 }
2081 kvm_pic_update_irq(pic_irqchip(kvm));
2082 return r;
2083}
2084
Sheng Yange0f63cb2008-03-04 00:50:59 +08002085static int kvm_vm_ioctl_get_pit(struct kvm *kvm, struct kvm_pit_state *ps)
2086{
2087 int r = 0;
2088
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002089 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002090 memcpy(ps, &kvm->arch.vpit->pit_state, sizeof(struct kvm_pit_state));
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002091 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002092 return r;
2093}
2094
2095static int kvm_vm_ioctl_set_pit(struct kvm *kvm, struct kvm_pit_state *ps)
2096{
2097 int r = 0;
2098
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002099 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002100 memcpy(&kvm->arch.vpit->pit_state, ps, sizeof(struct kvm_pit_state));
Beth Kone9f42752009-07-07 11:50:38 -04002101 kvm_pit_load_count(kvm, 0, ps->channels[0].count, 0);
2102 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
2103 return r;
2104}
2105
2106static int kvm_vm_ioctl_get_pit2(struct kvm *kvm, struct kvm_pit_state2 *ps)
2107{
2108 int r = 0;
2109
2110 mutex_lock(&kvm->arch.vpit->pit_state.lock);
2111 memcpy(ps->channels, &kvm->arch.vpit->pit_state.channels,
2112 sizeof(ps->channels));
2113 ps->flags = kvm->arch.vpit->pit_state.flags;
2114 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
2115 return r;
2116}
2117
2118static int kvm_vm_ioctl_set_pit2(struct kvm *kvm, struct kvm_pit_state2 *ps)
2119{
2120 int r = 0, start = 0;
2121 u32 prev_legacy, cur_legacy;
2122 mutex_lock(&kvm->arch.vpit->pit_state.lock);
2123 prev_legacy = kvm->arch.vpit->pit_state.flags & KVM_PIT_FLAGS_HPET_LEGACY;
2124 cur_legacy = ps->flags & KVM_PIT_FLAGS_HPET_LEGACY;
2125 if (!prev_legacy && cur_legacy)
2126 start = 1;
2127 memcpy(&kvm->arch.vpit->pit_state.channels, &ps->channels,
2128 sizeof(kvm->arch.vpit->pit_state.channels));
2129 kvm->arch.vpit->pit_state.flags = ps->flags;
2130 kvm_pit_load_count(kvm, 0, kvm->arch.vpit->pit_state.channels[0].count, start);
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002131 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002132 return r;
2133}
2134
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002135static int kvm_vm_ioctl_reinject(struct kvm *kvm,
2136 struct kvm_reinject_control *control)
2137{
2138 if (!kvm->arch.vpit)
2139 return -ENXIO;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002140 mutex_lock(&kvm->arch.vpit->pit_state.lock);
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002141 kvm->arch.vpit->pit_state.pit_timer.reinject = control->pit_reinject;
Marcelo Tosatti894a9c52009-06-23 15:05:14 -03002142 mutex_unlock(&kvm->arch.vpit->pit_state.lock);
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002143 return 0;
2144}
2145
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002146/*
2147 * Get (and clear) the dirty memory log for a memory slot.
2148 */
2149int kvm_vm_ioctl_get_dirty_log(struct kvm *kvm,
2150 struct kvm_dirty_log *log)
2151{
2152 int r;
2153 int n;
2154 struct kvm_memory_slot *memslot;
2155 int is_dirty = 0;
2156
Izik Eidus72dc67a2008-02-10 18:04:15 +02002157 down_write(&kvm->slots_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002158
2159 r = kvm_get_dirty_log(kvm, log, &is_dirty);
2160 if (r)
2161 goto out;
2162
2163 /* If nothing is dirty, don't bother messing with page tables. */
2164 if (is_dirty) {
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03002165 spin_lock(&kvm->mmu_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002166 kvm_mmu_slot_remove_write_access(kvm, log->slot);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03002167 spin_unlock(&kvm->mmu_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002168 memslot = &kvm->memslots[log->slot];
2169 n = ALIGN(memslot->npages, BITS_PER_LONG) / 8;
2170 memset(memslot->dirty_bitmap, 0, n);
2171 }
2172 r = 0;
2173out:
Izik Eidus72dc67a2008-02-10 18:04:15 +02002174 up_write(&kvm->slots_lock);
Zhang Xiantao5bb064d2007-11-18 20:29:43 +08002175 return r;
2176}
2177
Carsten Otte1fe779f2007-10-29 16:08:35 +01002178long kvm_arch_vm_ioctl(struct file *filp,
2179 unsigned int ioctl, unsigned long arg)
2180{
2181 struct kvm *kvm = filp->private_data;
2182 void __user *argp = (void __user *)arg;
2183 int r = -EINVAL;
Dave Hansenf0d66272008-08-11 10:01:45 -07002184 /*
2185 * This union makes it completely explicit to gcc-3.x
2186 * that these two variables' stack usage should be
2187 * combined, not added together.
2188 */
2189 union {
2190 struct kvm_pit_state ps;
Beth Kone9f42752009-07-07 11:50:38 -04002191 struct kvm_pit_state2 ps2;
Dave Hansenf0d66272008-08-11 10:01:45 -07002192 struct kvm_memory_alias alias;
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02002193 struct kvm_pit_config pit_config;
Dave Hansenf0d66272008-08-11 10:01:45 -07002194 } u;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002195
2196 switch (ioctl) {
2197 case KVM_SET_TSS_ADDR:
2198 r = kvm_vm_ioctl_set_tss_addr(kvm, arg);
2199 if (r < 0)
2200 goto out;
2201 break;
Sheng Yangb927a3c2009-07-21 10:42:48 +08002202 case KVM_SET_IDENTITY_MAP_ADDR: {
2203 u64 ident_addr;
2204
2205 r = -EFAULT;
2206 if (copy_from_user(&ident_addr, argp, sizeof ident_addr))
2207 goto out;
2208 r = kvm_vm_ioctl_set_identity_map_addr(kvm, ident_addr);
2209 if (r < 0)
2210 goto out;
2211 break;
2212 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01002213 case KVM_SET_MEMORY_REGION: {
2214 struct kvm_memory_region kvm_mem;
2215 struct kvm_userspace_memory_region kvm_userspace_mem;
2216
2217 r = -EFAULT;
2218 if (copy_from_user(&kvm_mem, argp, sizeof kvm_mem))
2219 goto out;
2220 kvm_userspace_mem.slot = kvm_mem.slot;
2221 kvm_userspace_mem.flags = kvm_mem.flags;
2222 kvm_userspace_mem.guest_phys_addr = kvm_mem.guest_phys_addr;
2223 kvm_userspace_mem.memory_size = kvm_mem.memory_size;
2224 r = kvm_vm_ioctl_set_memory_region(kvm, &kvm_userspace_mem, 0);
2225 if (r)
2226 goto out;
2227 break;
2228 }
2229 case KVM_SET_NR_MMU_PAGES:
2230 r = kvm_vm_ioctl_set_nr_mmu_pages(kvm, arg);
2231 if (r)
2232 goto out;
2233 break;
2234 case KVM_GET_NR_MMU_PAGES:
2235 r = kvm_vm_ioctl_get_nr_mmu_pages(kvm);
2236 break;
Dave Hansenf0d66272008-08-11 10:01:45 -07002237 case KVM_SET_MEMORY_ALIAS:
Carsten Otte1fe779f2007-10-29 16:08:35 +01002238 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07002239 if (copy_from_user(&u.alias, argp, sizeof(struct kvm_memory_alias)))
Carsten Otte1fe779f2007-10-29 16:08:35 +01002240 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07002241 r = kvm_vm_ioctl_set_memory_alias(kvm, &u.alias);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002242 if (r)
2243 goto out;
2244 break;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002245 case KVM_CREATE_IRQCHIP:
2246 r = -ENOMEM;
Zhang Xiantaod7deeeb02007-12-14 10:17:34 +08002247 kvm->arch.vpic = kvm_create_pic(kvm);
2248 if (kvm->arch.vpic) {
Carsten Otte1fe779f2007-10-29 16:08:35 +01002249 r = kvm_ioapic_init(kvm);
2250 if (r) {
Zhang Xiantaod7deeeb02007-12-14 10:17:34 +08002251 kfree(kvm->arch.vpic);
2252 kvm->arch.vpic = NULL;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002253 goto out;
2254 }
2255 } else
2256 goto out;
Avi Kivity399ec802008-11-19 13:58:46 +02002257 r = kvm_setup_default_irq_routing(kvm);
2258 if (r) {
2259 kfree(kvm->arch.vpic);
2260 kfree(kvm->arch.vioapic);
2261 goto out;
2262 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01002263 break;
Sheng Yang78376992008-01-28 05:10:22 +08002264 case KVM_CREATE_PIT:
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02002265 u.pit_config.flags = KVM_PIT_SPEAKER_DUMMY;
2266 goto create_pit;
2267 case KVM_CREATE_PIT2:
2268 r = -EFAULT;
2269 if (copy_from_user(&u.pit_config, argp,
2270 sizeof(struct kvm_pit_config)))
2271 goto out;
2272 create_pit:
Michael S. Tsirkin108b5662009-06-29 22:24:21 +03002273 down_write(&kvm->slots_lock);
Avi Kivity269e05e2009-01-05 15:21:42 +02002274 r = -EEXIST;
2275 if (kvm->arch.vpit)
2276 goto create_pit_unlock;
Sheng Yang78376992008-01-28 05:10:22 +08002277 r = -ENOMEM;
Jan Kiszkac5ff41c2009-05-14 22:42:53 +02002278 kvm->arch.vpit = kvm_create_pit(kvm, u.pit_config.flags);
Sheng Yang78376992008-01-28 05:10:22 +08002279 if (kvm->arch.vpit)
2280 r = 0;
Avi Kivity269e05e2009-01-05 15:21:42 +02002281 create_pit_unlock:
Michael S. Tsirkin108b5662009-06-29 22:24:21 +03002282 up_write(&kvm->slots_lock);
Sheng Yang78376992008-01-28 05:10:22 +08002283 break;
Gleb Natapov49256632009-02-04 17:28:14 +02002284 case KVM_IRQ_LINE_STATUS:
Carsten Otte1fe779f2007-10-29 16:08:35 +01002285 case KVM_IRQ_LINE: {
2286 struct kvm_irq_level irq_event;
2287
2288 r = -EFAULT;
2289 if (copy_from_user(&irq_event, argp, sizeof irq_event))
2290 goto out;
2291 if (irqchip_in_kernel(kvm)) {
Gleb Natapov49256632009-02-04 17:28:14 +02002292 __s32 status;
Marcelo Tosattifa40a822009-06-04 15:08:24 -03002293 mutex_lock(&kvm->irq_lock);
Gleb Natapov49256632009-02-04 17:28:14 +02002294 status = kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID,
2295 irq_event.irq, irq_event.level);
Marcelo Tosattifa40a822009-06-04 15:08:24 -03002296 mutex_unlock(&kvm->irq_lock);
Gleb Natapov49256632009-02-04 17:28:14 +02002297 if (ioctl == KVM_IRQ_LINE_STATUS) {
2298 irq_event.status = status;
2299 if (copy_to_user(argp, &irq_event,
2300 sizeof irq_event))
2301 goto out;
2302 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01002303 r = 0;
2304 }
2305 break;
2306 }
2307 case KVM_GET_IRQCHIP: {
2308 /* 0: PIC master, 1: PIC slave, 2: IOAPIC */
Dave Hansenf0d66272008-08-11 10:01:45 -07002309 struct kvm_irqchip *chip = kmalloc(sizeof(*chip), GFP_KERNEL);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002310
Dave Hansenf0d66272008-08-11 10:01:45 -07002311 r = -ENOMEM;
2312 if (!chip)
Carsten Otte1fe779f2007-10-29 16:08:35 +01002313 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07002314 r = -EFAULT;
2315 if (copy_from_user(chip, argp, sizeof *chip))
2316 goto get_irqchip_out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002317 r = -ENXIO;
2318 if (!irqchip_in_kernel(kvm))
Dave Hansenf0d66272008-08-11 10:01:45 -07002319 goto get_irqchip_out;
2320 r = kvm_vm_ioctl_get_irqchip(kvm, chip);
2321 if (r)
2322 goto get_irqchip_out;
2323 r = -EFAULT;
2324 if (copy_to_user(argp, chip, sizeof *chip))
2325 goto get_irqchip_out;
2326 r = 0;
2327 get_irqchip_out:
2328 kfree(chip);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002329 if (r)
2330 goto out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002331 break;
2332 }
2333 case KVM_SET_IRQCHIP: {
2334 /* 0: PIC master, 1: PIC slave, 2: IOAPIC */
Dave Hansenf0d66272008-08-11 10:01:45 -07002335 struct kvm_irqchip *chip = kmalloc(sizeof(*chip), GFP_KERNEL);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002336
Dave Hansenf0d66272008-08-11 10:01:45 -07002337 r = -ENOMEM;
2338 if (!chip)
Carsten Otte1fe779f2007-10-29 16:08:35 +01002339 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07002340 r = -EFAULT;
2341 if (copy_from_user(chip, argp, sizeof *chip))
2342 goto set_irqchip_out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002343 r = -ENXIO;
2344 if (!irqchip_in_kernel(kvm))
Dave Hansenf0d66272008-08-11 10:01:45 -07002345 goto set_irqchip_out;
2346 r = kvm_vm_ioctl_set_irqchip(kvm, chip);
2347 if (r)
2348 goto set_irqchip_out;
2349 r = 0;
2350 set_irqchip_out:
2351 kfree(chip);
Carsten Otte1fe779f2007-10-29 16:08:35 +01002352 if (r)
2353 goto out;
Carsten Otte1fe779f2007-10-29 16:08:35 +01002354 break;
2355 }
Sheng Yange0f63cb2008-03-04 00:50:59 +08002356 case KVM_GET_PIT: {
Sheng Yange0f63cb2008-03-04 00:50:59 +08002357 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07002358 if (copy_from_user(&u.ps, argp, sizeof(struct kvm_pit_state)))
Sheng Yange0f63cb2008-03-04 00:50:59 +08002359 goto out;
2360 r = -ENXIO;
2361 if (!kvm->arch.vpit)
2362 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07002363 r = kvm_vm_ioctl_get_pit(kvm, &u.ps);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002364 if (r)
2365 goto out;
2366 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07002367 if (copy_to_user(argp, &u.ps, sizeof(struct kvm_pit_state)))
Sheng Yange0f63cb2008-03-04 00:50:59 +08002368 goto out;
2369 r = 0;
2370 break;
2371 }
2372 case KVM_SET_PIT: {
Sheng Yange0f63cb2008-03-04 00:50:59 +08002373 r = -EFAULT;
Dave Hansenf0d66272008-08-11 10:01:45 -07002374 if (copy_from_user(&u.ps, argp, sizeof u.ps))
Sheng Yange0f63cb2008-03-04 00:50:59 +08002375 goto out;
2376 r = -ENXIO;
2377 if (!kvm->arch.vpit)
2378 goto out;
Dave Hansenf0d66272008-08-11 10:01:45 -07002379 r = kvm_vm_ioctl_set_pit(kvm, &u.ps);
Sheng Yange0f63cb2008-03-04 00:50:59 +08002380 if (r)
2381 goto out;
2382 r = 0;
2383 break;
2384 }
Beth Kone9f42752009-07-07 11:50:38 -04002385 case KVM_GET_PIT2: {
2386 r = -ENXIO;
2387 if (!kvm->arch.vpit)
2388 goto out;
2389 r = kvm_vm_ioctl_get_pit2(kvm, &u.ps2);
2390 if (r)
2391 goto out;
2392 r = -EFAULT;
2393 if (copy_to_user(argp, &u.ps2, sizeof(u.ps2)))
2394 goto out;
2395 r = 0;
2396 break;
2397 }
2398 case KVM_SET_PIT2: {
2399 r = -EFAULT;
2400 if (copy_from_user(&u.ps2, argp, sizeof(u.ps2)))
2401 goto out;
2402 r = -ENXIO;
2403 if (!kvm->arch.vpit)
2404 goto out;
2405 r = kvm_vm_ioctl_set_pit2(kvm, &u.ps2);
2406 if (r)
2407 goto out;
2408 r = 0;
2409 break;
2410 }
Marcelo Tosatti52d939a2008-12-30 15:55:06 -02002411 case KVM_REINJECT_CONTROL: {
2412 struct kvm_reinject_control control;
2413 r = -EFAULT;
2414 if (copy_from_user(&control, argp, sizeof(control)))
2415 goto out;
2416 r = kvm_vm_ioctl_reinject(kvm, &control);
2417 if (r)
2418 goto out;
2419 r = 0;
2420 break;
2421 }
Carsten Otte1fe779f2007-10-29 16:08:35 +01002422 default:
2423 ;
2424 }
2425out:
2426 return r;
2427}
2428
Zhang Xiantaoa16b0432007-11-16 14:38:21 +08002429static void kvm_init_msr_list(void)
Carsten Otte043405e2007-10-10 17:16:19 +02002430{
2431 u32 dummy[2];
2432 unsigned i, j;
2433
2434 for (i = j = 0; i < ARRAY_SIZE(msrs_to_save); i++) {
2435 if (rdmsr_safe(msrs_to_save[i], &dummy[0], &dummy[1]) < 0)
2436 continue;
2437 if (j < i)
2438 msrs_to_save[j] = msrs_to_save[i];
2439 j++;
2440 }
2441 num_msrs_to_save = j;
2442}
2443
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002444static int vcpu_mmio_write(struct kvm_vcpu *vcpu, gpa_t addr, int len,
2445 const void *v)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002446{
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002447 if (vcpu->arch.apic &&
2448 !kvm_iodevice_write(&vcpu->arch.apic->dev, addr, len, v))
2449 return 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002450
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002451 return kvm_io_bus_write(&vcpu->kvm->mmio_bus, addr, len, v);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002452}
2453
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002454static int vcpu_mmio_read(struct kvm_vcpu *vcpu, gpa_t addr, int len, void *v)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002455{
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002456 if (vcpu->arch.apic &&
2457 !kvm_iodevice_read(&vcpu->arch.apic->dev, addr, len, v))
2458 return 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002459
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002460 return kvm_io_bus_read(&vcpu->kvm->mmio_bus, addr, len, v);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002461}
2462
Hannes Edercded19f2009-02-21 02:19:13 +01002463static int kvm_read_guest_virt(gva_t addr, void *val, unsigned int bytes,
2464 struct kvm_vcpu *vcpu)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002465{
2466 void *data = val;
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002467 int r = X86EMUL_CONTINUE;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002468
2469 while (bytes) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002470 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002471 unsigned offset = addr & (PAGE_SIZE-1);
Izik Eidus77c20022008-12-29 01:42:19 +02002472 unsigned toread = min(bytes, (unsigned)PAGE_SIZE - offset);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002473 int ret;
2474
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002475 if (gpa == UNMAPPED_GVA) {
2476 r = X86EMUL_PROPAGATE_FAULT;
2477 goto out;
2478 }
Izik Eidus77c20022008-12-29 01:42:19 +02002479 ret = kvm_read_guest(vcpu->kvm, gpa, data, toread);
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002480 if (ret < 0) {
2481 r = X86EMUL_UNHANDLEABLE;
2482 goto out;
2483 }
Carsten Ottebbd9b642007-10-30 18:44:21 +01002484
Izik Eidus77c20022008-12-29 01:42:19 +02002485 bytes -= toread;
2486 data += toread;
2487 addr += toread;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002488 }
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002489out:
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002490 return r;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002491}
Izik Eidus77c20022008-12-29 01:42:19 +02002492
Hannes Edercded19f2009-02-21 02:19:13 +01002493static int kvm_write_guest_virt(gva_t addr, void *val, unsigned int bytes,
2494 struct kvm_vcpu *vcpu)
Izik Eidus77c20022008-12-29 01:42:19 +02002495{
2496 void *data = val;
2497 int r = X86EMUL_CONTINUE;
2498
2499 while (bytes) {
2500 gpa_t gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr);
2501 unsigned offset = addr & (PAGE_SIZE-1);
2502 unsigned towrite = min(bytes, (unsigned)PAGE_SIZE - offset);
2503 int ret;
2504
2505 if (gpa == UNMAPPED_GVA) {
2506 r = X86EMUL_PROPAGATE_FAULT;
2507 goto out;
2508 }
2509 ret = kvm_write_guest(vcpu->kvm, gpa, data, towrite);
2510 if (ret < 0) {
2511 r = X86EMUL_UNHANDLEABLE;
2512 goto out;
2513 }
2514
2515 bytes -= towrite;
2516 data += towrite;
2517 addr += towrite;
2518 }
2519out:
2520 return r;
2521}
2522
Carsten Ottebbd9b642007-10-30 18:44:21 +01002523
Carsten Ottebbd9b642007-10-30 18:44:21 +01002524static int emulator_read_emulated(unsigned long addr,
2525 void *val,
2526 unsigned int bytes,
2527 struct kvm_vcpu *vcpu)
2528{
Carsten Ottebbd9b642007-10-30 18:44:21 +01002529 gpa_t gpa;
2530
2531 if (vcpu->mmio_read_completed) {
2532 memcpy(val, vcpu->mmio_data, bytes);
Avi Kivityaec51dc2009-07-01 16:01:02 +03002533 trace_kvm_mmio(KVM_TRACE_MMIO_READ, bytes,
2534 vcpu->mmio_phys_addr, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002535 vcpu->mmio_read_completed = 0;
2536 return X86EMUL_CONTINUE;
2537 }
2538
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002539 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002540
2541 /* For APIC access vmexit */
2542 if ((gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
2543 goto mmio;
2544
Izik Eidus77c20022008-12-29 01:42:19 +02002545 if (kvm_read_guest_virt(addr, val, bytes, vcpu)
2546 == X86EMUL_CONTINUE)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002547 return X86EMUL_CONTINUE;
2548 if (gpa == UNMAPPED_GVA)
2549 return X86EMUL_PROPAGATE_FAULT;
2550
2551mmio:
2552 /*
2553 * Is this MMIO handled locally?
2554 */
Avi Kivityaec51dc2009-07-01 16:01:02 +03002555 if (!vcpu_mmio_read(vcpu, gpa, bytes, val)) {
2556 trace_kvm_mmio(KVM_TRACE_MMIO_READ, bytes, gpa, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002557 return X86EMUL_CONTINUE;
2558 }
Avi Kivityaec51dc2009-07-01 16:01:02 +03002559
2560 trace_kvm_mmio(KVM_TRACE_MMIO_READ_UNSATISFIED, bytes, gpa, 0);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002561
2562 vcpu->mmio_needed = 1;
2563 vcpu->mmio_phys_addr = gpa;
2564 vcpu->mmio_size = bytes;
2565 vcpu->mmio_is_write = 0;
2566
2567 return X86EMUL_UNHANDLEABLE;
2568}
2569
Marcelo Tosatti3200f402008-03-29 20:17:59 -03002570int emulator_write_phys(struct kvm_vcpu *vcpu, gpa_t gpa,
Avi Kivity9f811282008-03-02 14:06:05 +02002571 const void *val, int bytes)
2572{
2573 int ret;
2574
2575 ret = kvm_write_guest(vcpu->kvm, gpa, val, bytes);
2576 if (ret < 0)
2577 return 0;
Marcelo Tosattiad218f82008-12-01 22:32:05 -02002578 kvm_mmu_pte_write(vcpu, gpa, val, bytes, 1);
Avi Kivity9f811282008-03-02 14:06:05 +02002579 return 1;
2580}
2581
Carsten Ottebbd9b642007-10-30 18:44:21 +01002582static int emulator_write_emulated_onepage(unsigned long addr,
2583 const void *val,
2584 unsigned int bytes,
2585 struct kvm_vcpu *vcpu)
2586{
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002587 gpa_t gpa;
2588
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002589 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002590
2591 if (gpa == UNMAPPED_GVA) {
Avi Kivityc3c91fe2007-11-25 14:04:58 +02002592 kvm_inject_page_fault(vcpu, addr, 2);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002593 return X86EMUL_PROPAGATE_FAULT;
2594 }
2595
2596 /* For APIC access vmexit */
2597 if ((gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
2598 goto mmio;
2599
2600 if (emulator_write_phys(vcpu, gpa, val, bytes))
2601 return X86EMUL_CONTINUE;
2602
2603mmio:
Avi Kivityaec51dc2009-07-01 16:01:02 +03002604 trace_kvm_mmio(KVM_TRACE_MMIO_WRITE, bytes, gpa, *(u64 *)val);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002605 /*
2606 * Is this MMIO handled locally?
2607 */
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002608 if (!vcpu_mmio_write(vcpu, gpa, bytes, val))
Carsten Ottebbd9b642007-10-30 18:44:21 +01002609 return X86EMUL_CONTINUE;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002610
2611 vcpu->mmio_needed = 1;
2612 vcpu->mmio_phys_addr = gpa;
2613 vcpu->mmio_size = bytes;
2614 vcpu->mmio_is_write = 1;
2615 memcpy(vcpu->mmio_data, val, bytes);
2616
2617 return X86EMUL_CONTINUE;
2618}
2619
2620int emulator_write_emulated(unsigned long addr,
2621 const void *val,
2622 unsigned int bytes,
2623 struct kvm_vcpu *vcpu)
2624{
2625 /* Crossing a page boundary? */
2626 if (((addr + bytes - 1) ^ addr) & PAGE_MASK) {
2627 int rc, now;
2628
2629 now = -addr & ~PAGE_MASK;
2630 rc = emulator_write_emulated_onepage(addr, val, now, vcpu);
2631 if (rc != X86EMUL_CONTINUE)
2632 return rc;
2633 addr += now;
2634 val += now;
2635 bytes -= now;
2636 }
2637 return emulator_write_emulated_onepage(addr, val, bytes, vcpu);
2638}
2639EXPORT_SYMBOL_GPL(emulator_write_emulated);
2640
2641static int emulator_cmpxchg_emulated(unsigned long addr,
2642 const void *old,
2643 const void *new,
2644 unsigned int bytes,
2645 struct kvm_vcpu *vcpu)
2646{
Marcin Slusarz9f51e242009-08-09 21:54:00 +02002647 printk_once(KERN_WARNING "kvm: emulating exchange as write\n");
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002648#ifndef CONFIG_X86_64
2649 /* guests cmpxchg8b have to be emulated atomically */
2650 if (bytes == 8) {
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002651 gpa_t gpa;
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002652 struct page *page;
Andrew Mortonc0b49b02008-02-04 22:27:18 -08002653 char *kaddr;
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002654 u64 val;
2655
Marcelo Tosatti10589a42007-12-20 19:18:22 -05002656 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, addr);
2657
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002658 if (gpa == UNMAPPED_GVA ||
2659 (gpa & PAGE_MASK) == APIC_DEFAULT_PHYS_BASE)
2660 goto emul_write;
2661
2662 if (((gpa + bytes - 1) & PAGE_MASK) != (gpa & PAGE_MASK))
2663 goto emul_write;
2664
2665 val = *(u64 *)new;
Izik Eidus72dc67a2008-02-10 18:04:15 +02002666
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002667 page = gfn_to_page(vcpu->kvm, gpa >> PAGE_SHIFT);
Izik Eidus72dc67a2008-02-10 18:04:15 +02002668
Andrew Mortonc0b49b02008-02-04 22:27:18 -08002669 kaddr = kmap_atomic(page, KM_USER0);
2670 set_64bit((u64 *)(kaddr + offset_in_page(gpa)), val);
2671 kunmap_atomic(kaddr, KM_USER0);
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002672 kvm_release_page_dirty(page);
2673 }
Marcelo Tosatti3200f402008-03-29 20:17:59 -03002674emul_write:
Marcelo Tosatti2bacc552007-12-12 10:46:12 -05002675#endif
2676
Carsten Ottebbd9b642007-10-30 18:44:21 +01002677 return emulator_write_emulated(addr, new, bytes, vcpu);
2678}
2679
2680static unsigned long get_segment_base(struct kvm_vcpu *vcpu, int seg)
2681{
2682 return kvm_x86_ops->get_segment_base(vcpu, seg);
2683}
2684
2685int emulate_invlpg(struct kvm_vcpu *vcpu, gva_t address)
2686{
Marcelo Tosattia7052892008-09-23 13:18:35 -03002687 kvm_mmu_invlpg(vcpu, address);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002688 return X86EMUL_CONTINUE;
2689}
2690
2691int emulate_clts(struct kvm_vcpu *vcpu)
2692{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002693 kvm_x86_ops->set_cr0(vcpu, vcpu->arch.cr0 & ~X86_CR0_TS);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002694 return X86EMUL_CONTINUE;
2695}
2696
2697int emulator_get_dr(struct x86_emulate_ctxt *ctxt, int dr, unsigned long *dest)
2698{
2699 struct kvm_vcpu *vcpu = ctxt->vcpu;
2700
2701 switch (dr) {
2702 case 0 ... 3:
2703 *dest = kvm_x86_ops->get_dr(vcpu, dr);
2704 return X86EMUL_CONTINUE;
2705 default:
Harvey Harrisonb8688d52008-03-03 12:59:56 -08002706 pr_unimpl(vcpu, "%s: unexpected dr %u\n", __func__, dr);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002707 return X86EMUL_UNHANDLEABLE;
2708 }
2709}
2710
2711int emulator_set_dr(struct x86_emulate_ctxt *ctxt, int dr, unsigned long value)
2712{
2713 unsigned long mask = (ctxt->mode == X86EMUL_MODE_PROT64) ? ~0ULL : ~0U;
2714 int exception;
2715
2716 kvm_x86_ops->set_dr(ctxt->vcpu, dr, value & mask, &exception);
2717 if (exception) {
2718 /* FIXME: better handling */
2719 return X86EMUL_UNHANDLEABLE;
2720 }
2721 return X86EMUL_CONTINUE;
2722}
2723
2724void kvm_report_emulation_failure(struct kvm_vcpu *vcpu, const char *context)
2725{
Carsten Ottebbd9b642007-10-30 18:44:21 +01002726 u8 opcodes[4];
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002727 unsigned long rip = kvm_rip_read(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002728 unsigned long rip_linear;
2729
Avi Kivityf76c7102008-06-13 22:45:42 +03002730 if (!printk_ratelimit())
Carsten Ottebbd9b642007-10-30 18:44:21 +01002731 return;
2732
Glauber Costa25be4602008-06-10 10:46:53 -03002733 rip_linear = rip + get_segment_base(vcpu, VCPU_SREG_CS);
2734
Izik Eidus77c20022008-12-29 01:42:19 +02002735 kvm_read_guest_virt(rip_linear, (void *)opcodes, 4, vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002736
2737 printk(KERN_ERR "emulation failed (%s) rip %lx %02x %02x %02x %02x\n",
2738 context, rip, opcodes[0], opcodes[1], opcodes[2], opcodes[3]);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002739}
2740EXPORT_SYMBOL_GPL(kvm_report_emulation_failure);
2741
Harvey Harrison14af3f32008-02-19 10:25:50 -08002742static struct x86_emulate_ops emulate_ops = {
Izik Eidus77c20022008-12-29 01:42:19 +02002743 .read_std = kvm_read_guest_virt,
Carsten Ottebbd9b642007-10-30 18:44:21 +01002744 .read_emulated = emulator_read_emulated,
2745 .write_emulated = emulator_write_emulated,
2746 .cmpxchg_emulated = emulator_cmpxchg_emulated,
2747};
2748
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002749static void cache_all_regs(struct kvm_vcpu *vcpu)
2750{
2751 kvm_register_read(vcpu, VCPU_REGS_RAX);
2752 kvm_register_read(vcpu, VCPU_REGS_RSP);
2753 kvm_register_read(vcpu, VCPU_REGS_RIP);
2754 vcpu->arch.regs_dirty = ~0;
2755}
2756
Carsten Ottebbd9b642007-10-30 18:44:21 +01002757int emulate_instruction(struct kvm_vcpu *vcpu,
2758 struct kvm_run *run,
2759 unsigned long cr2,
2760 u16 error_code,
Sheng Yang571008d2008-01-02 14:49:22 +08002761 int emulation_type)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002762{
Glauber Costa310b5d32009-05-12 16:21:06 -04002763 int r, shadow_mask;
Sheng Yang571008d2008-01-02 14:49:22 +08002764 struct decode_cache *c;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002765
Avi Kivity26eef702008-07-03 14:59:22 +03002766 kvm_clear_exception_queue(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002767 vcpu->arch.mmio_fault_cr2 = cr2;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002768 /*
Avi Kivity56e82312009-08-12 15:04:37 +03002769 * TODO: fix emulate.c to use guest_read/write_register
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002770 * instead of direct ->regs accesses, can save hundred cycles
2771 * on Intel for instructions that don't read/change RSP, for
2772 * for example.
2773 */
2774 cache_all_regs(vcpu);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002775
2776 vcpu->mmio_is_write = 0;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002777 vcpu->arch.pio.string = 0;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002778
Sheng Yang571008d2008-01-02 14:49:22 +08002779 if (!(emulation_type & EMULTYPE_NO_DECODE)) {
Carsten Ottebbd9b642007-10-30 18:44:21 +01002780 int cs_db, cs_l;
2781 kvm_x86_ops->get_cs_db_l_bits(vcpu, &cs_db, &cs_l);
2782
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002783 vcpu->arch.emulate_ctxt.vcpu = vcpu;
2784 vcpu->arch.emulate_ctxt.eflags = kvm_x86_ops->get_rflags(vcpu);
2785 vcpu->arch.emulate_ctxt.mode =
2786 (vcpu->arch.emulate_ctxt.eflags & X86_EFLAGS_VM)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002787 ? X86EMUL_MODE_REAL : cs_l
2788 ? X86EMUL_MODE_PROT64 : cs_db
2789 ? X86EMUL_MODE_PROT32 : X86EMUL_MODE_PROT16;
2790
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002791 r = x86_decode_insn(&vcpu->arch.emulate_ctxt, &emulate_ops);
Sheng Yang571008d2008-01-02 14:49:22 +08002792
Andre Przywara0cb57622009-06-17 15:50:31 +02002793 /* Only allow emulation of specific instructions on #UD
2794 * (namely VMMCALL, sysenter, sysexit, syscall)*/
Sheng Yang571008d2008-01-02 14:49:22 +08002795 c = &vcpu->arch.emulate_ctxt.decode;
Andre Przywara0cb57622009-06-17 15:50:31 +02002796 if (emulation_type & EMULTYPE_TRAP_UD) {
2797 if (!c->twobyte)
2798 return EMULATE_FAIL;
2799 switch (c->b) {
2800 case 0x01: /* VMMCALL */
2801 if (c->modrm_mod != 3 || c->modrm_rm != 1)
2802 return EMULATE_FAIL;
2803 break;
2804 case 0x34: /* sysenter */
2805 case 0x35: /* sysexit */
2806 if (c->modrm_mod != 0 || c->modrm_rm != 0)
2807 return EMULATE_FAIL;
2808 break;
2809 case 0x05: /* syscall */
2810 if (c->modrm_mod != 0 || c->modrm_rm != 0)
2811 return EMULATE_FAIL;
2812 break;
2813 default:
2814 return EMULATE_FAIL;
2815 }
2816
2817 if (!(c->modrm_reg == 0 || c->modrm_reg == 3))
2818 return EMULATE_FAIL;
2819 }
Sheng Yang571008d2008-01-02 14:49:22 +08002820
Avi Kivityf2b57562007-11-18 15:17:51 +02002821 ++vcpu->stat.insn_emulation;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002822 if (r) {
Avi Kivityf2b57562007-11-18 15:17:51 +02002823 ++vcpu->stat.insn_emulation_fail;
Carsten Ottebbd9b642007-10-30 18:44:21 +01002824 if (kvm_mmu_unprotect_page_virt(vcpu, cr2))
2825 return EMULATE_DONE;
2826 return EMULATE_FAIL;
2827 }
2828 }
2829
Gleb Natapovba8afb62009-04-12 13:36:57 +03002830 if (emulation_type & EMULTYPE_SKIP) {
2831 kvm_rip_write(vcpu, vcpu->arch.emulate_ctxt.decode.eip);
2832 return EMULATE_DONE;
2833 }
2834
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002835 r = x86_emulate_insn(&vcpu->arch.emulate_ctxt, &emulate_ops);
Glauber Costa310b5d32009-05-12 16:21:06 -04002836 shadow_mask = vcpu->arch.emulate_ctxt.interruptibility;
2837
2838 if (r == 0)
2839 kvm_x86_ops->set_interrupt_shadow(vcpu, shadow_mask);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002840
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002841 if (vcpu->arch.pio.string)
Carsten Ottebbd9b642007-10-30 18:44:21 +01002842 return EMULATE_DO_MMIO;
2843
2844 if ((r || vcpu->mmio_is_write) && run) {
2845 run->exit_reason = KVM_EXIT_MMIO;
2846 run->mmio.phys_addr = vcpu->mmio_phys_addr;
2847 memcpy(run->mmio.data, vcpu->mmio_data, 8);
2848 run->mmio.len = vcpu->mmio_size;
2849 run->mmio.is_write = vcpu->mmio_is_write;
2850 }
2851
2852 if (r) {
2853 if (kvm_mmu_unprotect_page_virt(vcpu, cr2))
2854 return EMULATE_DONE;
2855 if (!vcpu->mmio_needed) {
2856 kvm_report_emulation_failure(vcpu, "mmio");
2857 return EMULATE_FAIL;
2858 }
2859 return EMULATE_DO_MMIO;
2860 }
2861
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002862 kvm_x86_ops->set_rflags(vcpu, vcpu->arch.emulate_ctxt.eflags);
Carsten Ottebbd9b642007-10-30 18:44:21 +01002863
2864 if (vcpu->mmio_is_write) {
2865 vcpu->mmio_needed = 0;
2866 return EMULATE_DO_MMIO;
2867 }
2868
2869 return EMULATE_DONE;
2870}
2871EXPORT_SYMBOL_GPL(emulate_instruction);
2872
Carsten Ottede7d7892007-10-30 18:44:25 +01002873static int pio_copy_data(struct kvm_vcpu *vcpu)
2874{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002875 void *p = vcpu->arch.pio_data;
Izik Eidus0f346072008-12-29 01:42:20 +02002876 gva_t q = vcpu->arch.pio.guest_gva;
Carsten Ottede7d7892007-10-30 18:44:25 +01002877 unsigned bytes;
Izik Eidus0f346072008-12-29 01:42:20 +02002878 int ret;
Carsten Ottede7d7892007-10-30 18:44:25 +01002879
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002880 bytes = vcpu->arch.pio.size * vcpu->arch.pio.cur_count;
2881 if (vcpu->arch.pio.in)
Izik Eidus0f346072008-12-29 01:42:20 +02002882 ret = kvm_write_guest_virt(q, p, bytes, vcpu);
Carsten Ottede7d7892007-10-30 18:44:25 +01002883 else
Izik Eidus0f346072008-12-29 01:42:20 +02002884 ret = kvm_read_guest_virt(q, p, bytes, vcpu);
2885 return ret;
Carsten Ottede7d7892007-10-30 18:44:25 +01002886}
2887
2888int complete_pio(struct kvm_vcpu *vcpu)
2889{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002890 struct kvm_pio_request *io = &vcpu->arch.pio;
Carsten Ottede7d7892007-10-30 18:44:25 +01002891 long delta;
2892 int r;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002893 unsigned long val;
Carsten Ottede7d7892007-10-30 18:44:25 +01002894
2895 if (!io->string) {
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002896 if (io->in) {
2897 val = kvm_register_read(vcpu, VCPU_REGS_RAX);
2898 memcpy(&val, vcpu->arch.pio_data, io->size);
2899 kvm_register_write(vcpu, VCPU_REGS_RAX, val);
2900 }
Carsten Ottede7d7892007-10-30 18:44:25 +01002901 } else {
2902 if (io->in) {
2903 r = pio_copy_data(vcpu);
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002904 if (r)
Carsten Ottede7d7892007-10-30 18:44:25 +01002905 return r;
Carsten Ottede7d7892007-10-30 18:44:25 +01002906 }
2907
2908 delta = 1;
2909 if (io->rep) {
2910 delta *= io->cur_count;
2911 /*
2912 * The size of the register should really depend on
2913 * current address size.
2914 */
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002915 val = kvm_register_read(vcpu, VCPU_REGS_RCX);
2916 val -= delta;
2917 kvm_register_write(vcpu, VCPU_REGS_RCX, val);
Carsten Ottede7d7892007-10-30 18:44:25 +01002918 }
2919 if (io->down)
2920 delta = -delta;
2921 delta *= io->size;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002922 if (io->in) {
2923 val = kvm_register_read(vcpu, VCPU_REGS_RDI);
2924 val += delta;
2925 kvm_register_write(vcpu, VCPU_REGS_RDI, val);
2926 } else {
2927 val = kvm_register_read(vcpu, VCPU_REGS_RSI);
2928 val += delta;
2929 kvm_register_write(vcpu, VCPU_REGS_RSI, val);
2930 }
Carsten Ottede7d7892007-10-30 18:44:25 +01002931 }
2932
Carsten Ottede7d7892007-10-30 18:44:25 +01002933 io->count -= io->cur_count;
2934 io->cur_count = 0;
2935
2936 return 0;
2937}
2938
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002939static int kernel_pio(struct kvm_vcpu *vcpu, void *pd)
Carsten Ottede7d7892007-10-30 18:44:25 +01002940{
2941 /* TODO: String I/O for in kernel device */
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002942 int r;
Carsten Ottede7d7892007-10-30 18:44:25 +01002943
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002944 if (vcpu->arch.pio.in)
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002945 r = kvm_io_bus_read(&vcpu->kvm->pio_bus, vcpu->arch.pio.port,
2946 vcpu->arch.pio.size, pd);
Carsten Ottede7d7892007-10-30 18:44:25 +01002947 else
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002948 r = kvm_io_bus_write(&vcpu->kvm->pio_bus, vcpu->arch.pio.port,
2949 vcpu->arch.pio.size, pd);
2950 return r;
Carsten Ottede7d7892007-10-30 18:44:25 +01002951}
2952
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002953static int pio_string_write(struct kvm_vcpu *vcpu)
Carsten Ottede7d7892007-10-30 18:44:25 +01002954{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002955 struct kvm_pio_request *io = &vcpu->arch.pio;
2956 void *pd = vcpu->arch.pio_data;
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002957 int i, r = 0;
Carsten Ottede7d7892007-10-30 18:44:25 +01002958
Carsten Ottede7d7892007-10-30 18:44:25 +01002959 for (i = 0; i < io->cur_count; i++) {
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002960 if (kvm_io_bus_write(&vcpu->kvm->pio_bus,
2961 io->port, io->size, pd)) {
2962 r = -EOPNOTSUPP;
2963 break;
2964 }
Carsten Ottede7d7892007-10-30 18:44:25 +01002965 pd += io->size;
2966 }
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002967 return r;
Carsten Ottede7d7892007-10-30 18:44:25 +01002968}
2969
2970int kvm_emulate_pio(struct kvm_vcpu *vcpu, struct kvm_run *run, int in,
2971 int size, unsigned port)
2972{
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002973 unsigned long val;
Carsten Ottede7d7892007-10-30 18:44:25 +01002974
2975 vcpu->run->exit_reason = KVM_EXIT_IO;
2976 vcpu->run->io.direction = in ? KVM_EXIT_IO_IN : KVM_EXIT_IO_OUT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002977 vcpu->run->io.size = vcpu->arch.pio.size = size;
Carsten Ottede7d7892007-10-30 18:44:25 +01002978 vcpu->run->io.data_offset = KVM_PIO_PAGE_OFFSET * PAGE_SIZE;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002979 vcpu->run->io.count = vcpu->arch.pio.count = vcpu->arch.pio.cur_count = 1;
2980 vcpu->run->io.port = vcpu->arch.pio.port = port;
2981 vcpu->arch.pio.in = in;
2982 vcpu->arch.pio.string = 0;
2983 vcpu->arch.pio.down = 0;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08002984 vcpu->arch.pio.rep = 0;
Carsten Ottede7d7892007-10-30 18:44:25 +01002985
Marcelo Tosatti229456f2009-06-17 09:22:14 -03002986 trace_kvm_pio(vcpu->run->io.direction == KVM_EXIT_IO_OUT, port,
2987 size, 1);
Feng (Eric) Liu2714d1d2008-04-10 15:31:10 -04002988
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03002989 val = kvm_register_read(vcpu, VCPU_REGS_RAX);
2990 memcpy(vcpu->arch.pio_data, &val, 4);
Carsten Ottede7d7892007-10-30 18:44:25 +01002991
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03002992 if (!kernel_pio(vcpu, vcpu->arch.pio_data)) {
Carsten Ottede7d7892007-10-30 18:44:25 +01002993 complete_pio(vcpu);
2994 return 1;
2995 }
2996 return 0;
2997}
2998EXPORT_SYMBOL_GPL(kvm_emulate_pio);
2999
3000int kvm_emulate_pio_string(struct kvm_vcpu *vcpu, struct kvm_run *run, int in,
3001 int size, unsigned long count, int down,
3002 gva_t address, int rep, unsigned port)
3003{
3004 unsigned now, in_page;
Izik Eidus0f346072008-12-29 01:42:20 +02003005 int ret = 0;
Carsten Ottede7d7892007-10-30 18:44:25 +01003006
3007 vcpu->run->exit_reason = KVM_EXIT_IO;
3008 vcpu->run->io.direction = in ? KVM_EXIT_IO_IN : KVM_EXIT_IO_OUT;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003009 vcpu->run->io.size = vcpu->arch.pio.size = size;
Carsten Ottede7d7892007-10-30 18:44:25 +01003010 vcpu->run->io.data_offset = KVM_PIO_PAGE_OFFSET * PAGE_SIZE;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003011 vcpu->run->io.count = vcpu->arch.pio.count = vcpu->arch.pio.cur_count = count;
3012 vcpu->run->io.port = vcpu->arch.pio.port = port;
3013 vcpu->arch.pio.in = in;
3014 vcpu->arch.pio.string = 1;
3015 vcpu->arch.pio.down = down;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003016 vcpu->arch.pio.rep = rep;
Carsten Ottede7d7892007-10-30 18:44:25 +01003017
Marcelo Tosatti229456f2009-06-17 09:22:14 -03003018 trace_kvm_pio(vcpu->run->io.direction == KVM_EXIT_IO_OUT, port,
3019 size, count);
Feng (Eric) Liu2714d1d2008-04-10 15:31:10 -04003020
Carsten Ottede7d7892007-10-30 18:44:25 +01003021 if (!count) {
3022 kvm_x86_ops->skip_emulated_instruction(vcpu);
3023 return 1;
3024 }
3025
3026 if (!down)
3027 in_page = PAGE_SIZE - offset_in_page(address);
3028 else
3029 in_page = offset_in_page(address) + size;
3030 now = min(count, (unsigned long)in_page / size);
Izik Eidus0f346072008-12-29 01:42:20 +02003031 if (!now)
Carsten Ottede7d7892007-10-30 18:44:25 +01003032 now = 1;
Carsten Ottede7d7892007-10-30 18:44:25 +01003033 if (down) {
3034 /*
3035 * String I/O in reverse. Yuck. Kill the guest, fix later.
3036 */
3037 pr_unimpl(vcpu, "guest string pio down\n");
Avi Kivityc1a5d4f2007-11-25 14:12:03 +02003038 kvm_inject_gp(vcpu, 0);
Carsten Ottede7d7892007-10-30 18:44:25 +01003039 return 1;
3040 }
3041 vcpu->run->io.count = now;
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003042 vcpu->arch.pio.cur_count = now;
Carsten Ottede7d7892007-10-30 18:44:25 +01003043
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003044 if (vcpu->arch.pio.cur_count == vcpu->arch.pio.count)
Carsten Ottede7d7892007-10-30 18:44:25 +01003045 kvm_x86_ops->skip_emulated_instruction(vcpu);
3046
Izik Eidus0f346072008-12-29 01:42:20 +02003047 vcpu->arch.pio.guest_gva = address;
Carsten Ottede7d7892007-10-30 18:44:25 +01003048
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003049 if (!vcpu->arch.pio.in) {
Carsten Ottede7d7892007-10-30 18:44:25 +01003050 /* string PIO write */
3051 ret = pio_copy_data(vcpu);
Izik Eidus0f346072008-12-29 01:42:20 +02003052 if (ret == X86EMUL_PROPAGATE_FAULT) {
3053 kvm_inject_gp(vcpu, 0);
3054 return 1;
3055 }
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003056 if (ret == 0 && !pio_string_write(vcpu)) {
Carsten Ottede7d7892007-10-30 18:44:25 +01003057 complete_pio(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003058 if (vcpu->arch.pio.count == 0)
Carsten Ottede7d7892007-10-30 18:44:25 +01003059 ret = 1;
3060 }
Michael S. Tsirkinbda90202009-06-29 22:24:32 +03003061 }
3062 /* no string PIO read support yet */
Carsten Ottede7d7892007-10-30 18:44:25 +01003063
3064 return ret;
3065}
3066EXPORT_SYMBOL_GPL(kvm_emulate_pio_string);
3067
Gerd Hoffmannc8076602009-02-04 17:52:04 +01003068static void bounce_off(void *info)
3069{
3070 /* nothing */
3071}
3072
3073static unsigned int ref_freq;
3074static unsigned long tsc_khz_ref;
3075
3076static int kvmclock_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
3077 void *data)
3078{
3079 struct cpufreq_freqs *freq = data;
3080 struct kvm *kvm;
3081 struct kvm_vcpu *vcpu;
3082 int i, send_ipi = 0;
3083
3084 if (!ref_freq)
3085 ref_freq = freq->old;
3086
3087 if (val == CPUFREQ_PRECHANGE && freq->old > freq->new)
3088 return 0;
3089 if (val == CPUFREQ_POSTCHANGE && freq->old < freq->new)
3090 return 0;
3091 per_cpu(cpu_tsc_khz, freq->cpu) = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
3092
3093 spin_lock(&kvm_lock);
3094 list_for_each_entry(kvm, &vm_list, vm_list) {
Gleb Natapov988a2ca2009-06-09 15:56:29 +03003095 kvm_for_each_vcpu(i, vcpu, kvm) {
Gerd Hoffmannc8076602009-02-04 17:52:04 +01003096 if (vcpu->cpu != freq->cpu)
3097 continue;
3098 if (!kvm_request_guest_time_update(vcpu))
3099 continue;
3100 if (vcpu->cpu != smp_processor_id())
3101 send_ipi++;
3102 }
3103 }
3104 spin_unlock(&kvm_lock);
3105
3106 if (freq->old < freq->new && send_ipi) {
3107 /*
3108 * We upscale the frequency. Must make the guest
3109 * doesn't see old kvmclock values while running with
3110 * the new frequency, otherwise we risk the guest sees
3111 * time go backwards.
3112 *
3113 * In case we update the frequency for another cpu
3114 * (which might be in guest context) send an interrupt
3115 * to kick the cpu out of guest context. Next time
3116 * guest context is entered kvmclock will be updated,
3117 * so the guest will not see stale values.
3118 */
3119 smp_call_function_single(freq->cpu, bounce_off, NULL, 1);
3120 }
3121 return 0;
3122}
3123
3124static struct notifier_block kvmclock_cpufreq_notifier_block = {
3125 .notifier_call = kvmclock_cpufreq_notifier
3126};
3127
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003128int kvm_arch_init(void *opaque)
Carsten Otte043405e2007-10-10 17:16:19 +02003129{
Gerd Hoffmannc8076602009-02-04 17:52:04 +01003130 int r, cpu;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003131 struct kvm_x86_ops *ops = (struct kvm_x86_ops *)opaque;
3132
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003133 if (kvm_x86_ops) {
3134 printk(KERN_ERR "kvm: already loaded the other module\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003135 r = -EEXIST;
3136 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003137 }
3138
3139 if (!ops->cpu_has_kvm_support()) {
3140 printk(KERN_ERR "kvm: no hardware support\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003141 r = -EOPNOTSUPP;
3142 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003143 }
3144 if (ops->disabled_by_bios()) {
3145 printk(KERN_ERR "kvm: disabled by bios\n");
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003146 r = -EOPNOTSUPP;
3147 goto out;
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003148 }
3149
Avi Kivity97db56c2008-01-13 13:23:56 +02003150 r = kvm_mmu_module_init();
3151 if (r)
3152 goto out;
3153
3154 kvm_init_msr_list();
3155
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003156 kvm_x86_ops = ops;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003157 kvm_mmu_set_nonpresent_ptes(0ull, 0ull);
Sheng Yang7b523452008-04-25 21:13:50 +08003158 kvm_mmu_set_base_ptes(PT_PRESENT_MASK);
3159 kvm_mmu_set_mask_ptes(PT_USER_MASK, PT_ACCESSED_MASK,
Sheng Yang4b12f0d2009-04-27 20:35:42 +08003160 PT_DIRTY_MASK, PT64_NX_MASK, 0);
Gerd Hoffmannc8076602009-02-04 17:52:04 +01003161
3162 for_each_possible_cpu(cpu)
3163 per_cpu(cpu_tsc_khz, cpu) = tsc_khz;
3164 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) {
3165 tsc_khz_ref = tsc_khz;
3166 cpufreq_register_notifier(&kvmclock_cpufreq_notifier_block,
3167 CPUFREQ_TRANSITION_NOTIFIER);
3168 }
3169
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003170 return 0;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003171
3172out:
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003173 return r;
Carsten Otte043405e2007-10-10 17:16:19 +02003174}
Hollis Blanchard8776e512007-10-31 17:24:24 -05003175
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003176void kvm_arch_exit(void)
3177{
Jan Kiszka888d2562009-04-17 19:24:58 +02003178 if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
3179 cpufreq_unregister_notifier(&kvmclock_cpufreq_notifier_block,
3180 CPUFREQ_TRANSITION_NOTIFIER);
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003181 kvm_x86_ops = NULL;
Zhang Xiantao56c6d282007-11-18 20:43:21 +08003182 kvm_mmu_module_exit();
3183}
Zhang Xiantaof8c16bb2007-11-14 20:40:21 +08003184
Hollis Blanchard8776e512007-10-31 17:24:24 -05003185int kvm_emulate_halt(struct kvm_vcpu *vcpu)
3186{
3187 ++vcpu->stat.halt_exits;
3188 if (irqchip_in_kernel(vcpu->kvm)) {
Avi Kivitya4535292008-04-13 17:54:35 +03003189 vcpu->arch.mp_state = KVM_MP_STATE_HALTED;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003190 return 1;
3191 } else {
3192 vcpu->run->exit_reason = KVM_EXIT_HLT;
3193 return 0;
3194 }
3195}
3196EXPORT_SYMBOL_GPL(kvm_emulate_halt);
3197
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05003198static inline gpa_t hc_gpa(struct kvm_vcpu *vcpu, unsigned long a0,
3199 unsigned long a1)
3200{
3201 if (is_long_mode(vcpu))
3202 return a0;
3203 else
3204 return a0 | ((gpa_t)a1 << 32);
3205}
3206
Hollis Blanchard8776e512007-10-31 17:24:24 -05003207int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
3208{
3209 unsigned long nr, a0, a1, a2, a3, ret;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05003210 int r = 1;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003211
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003212 nr = kvm_register_read(vcpu, VCPU_REGS_RAX);
3213 a0 = kvm_register_read(vcpu, VCPU_REGS_RBX);
3214 a1 = kvm_register_read(vcpu, VCPU_REGS_RCX);
3215 a2 = kvm_register_read(vcpu, VCPU_REGS_RDX);
3216 a3 = kvm_register_read(vcpu, VCPU_REGS_RSI);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003217
Marcelo Tosatti229456f2009-06-17 09:22:14 -03003218 trace_kvm_hypercall(nr, a0, a1, a2, a3);
Feng (Eric) Liu2714d1d2008-04-10 15:31:10 -04003219
Hollis Blanchard8776e512007-10-31 17:24:24 -05003220 if (!is_long_mode(vcpu)) {
3221 nr &= 0xFFFFFFFF;
3222 a0 &= 0xFFFFFFFF;
3223 a1 &= 0xFFFFFFFF;
3224 a2 &= 0xFFFFFFFF;
3225 a3 &= 0xFFFFFFFF;
3226 }
3227
Jan Kiszka07708c42009-08-03 18:43:28 +02003228 if (kvm_x86_ops->get_cpl(vcpu) != 0) {
3229 ret = -KVM_EPERM;
3230 goto out;
3231 }
3232
Hollis Blanchard8776e512007-10-31 17:24:24 -05003233 switch (nr) {
Avi Kivityb93463a2007-10-25 16:52:32 +02003234 case KVM_HC_VAPIC_POLL_IRQ:
3235 ret = 0;
3236 break;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05003237 case KVM_HC_MMU_OP:
3238 r = kvm_pv_mmu_op(vcpu, a0, hc_gpa(vcpu, a1, a2), &ret);
3239 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003240 default:
3241 ret = -KVM_ENOSYS;
3242 break;
3243 }
Jan Kiszka07708c42009-08-03 18:43:28 +02003244out:
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003245 kvm_register_write(vcpu, VCPU_REGS_RAX, ret);
Amit Shahf11c3a82008-02-21 01:00:30 +05303246 ++vcpu->stat.hypercalls;
Marcelo Tosatti2f333bc2008-02-22 12:21:37 -05003247 return r;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003248}
3249EXPORT_SYMBOL_GPL(kvm_emulate_hypercall);
3250
3251int kvm_fix_hypercall(struct kvm_vcpu *vcpu)
3252{
3253 char instruction[3];
3254 int ret = 0;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003255 unsigned long rip = kvm_rip_read(vcpu);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003256
Hollis Blanchard8776e512007-10-31 17:24:24 -05003257
3258 /*
3259 * Blow out the MMU to ensure that no other VCPU has an active mapping
3260 * to ensure that the updated hypercall appears atomically across all
3261 * VCPUs.
3262 */
3263 kvm_mmu_zap_all(vcpu->kvm);
3264
Hollis Blanchard8776e512007-10-31 17:24:24 -05003265 kvm_x86_ops->patch_hypercall(vcpu, instruction);
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003266 if (emulator_write_emulated(rip, instruction, 3, vcpu)
Hollis Blanchard8776e512007-10-31 17:24:24 -05003267 != X86EMUL_CONTINUE)
3268 ret = -EFAULT;
3269
Hollis Blanchard8776e512007-10-31 17:24:24 -05003270 return ret;
3271}
3272
3273static u64 mk_cr_64(u64 curr_cr, u32 new_val)
3274{
3275 return (curr_cr & ~((1ULL << 32) - 1)) | new_val;
3276}
3277
3278void realmode_lgdt(struct kvm_vcpu *vcpu, u16 limit, unsigned long base)
3279{
3280 struct descriptor_table dt = { limit, base };
3281
3282 kvm_x86_ops->set_gdt(vcpu, &dt);
3283}
3284
3285void realmode_lidt(struct kvm_vcpu *vcpu, u16 limit, unsigned long base)
3286{
3287 struct descriptor_table dt = { limit, base };
3288
3289 kvm_x86_ops->set_idt(vcpu, &dt);
3290}
3291
3292void realmode_lmsw(struct kvm_vcpu *vcpu, unsigned long msw,
3293 unsigned long *rflags)
3294{
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003295 kvm_lmsw(vcpu, msw);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003296 *rflags = kvm_x86_ops->get_rflags(vcpu);
3297}
3298
3299unsigned long realmode_get_cr(struct kvm_vcpu *vcpu, int cr)
3300{
Joerg Roedel54e445c2008-04-30 17:56:02 +02003301 unsigned long value;
3302
Hollis Blanchard8776e512007-10-31 17:24:24 -05003303 kvm_x86_ops->decache_cr4_guest_bits(vcpu);
3304 switch (cr) {
3305 case 0:
Joerg Roedel54e445c2008-04-30 17:56:02 +02003306 value = vcpu->arch.cr0;
3307 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003308 case 2:
Joerg Roedel54e445c2008-04-30 17:56:02 +02003309 value = vcpu->arch.cr2;
3310 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003311 case 3:
Joerg Roedel54e445c2008-04-30 17:56:02 +02003312 value = vcpu->arch.cr3;
3313 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003314 case 4:
Joerg Roedel54e445c2008-04-30 17:56:02 +02003315 value = vcpu->arch.cr4;
3316 break;
Joerg Roedel152ff9b2007-12-06 15:46:52 +01003317 case 8:
Joerg Roedel54e445c2008-04-30 17:56:02 +02003318 value = kvm_get_cr8(vcpu);
3319 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003320 default:
Harvey Harrisonb8688d52008-03-03 12:59:56 -08003321 vcpu_printf(vcpu, "%s: unexpected cr %u\n", __func__, cr);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003322 return 0;
3323 }
Joerg Roedel54e445c2008-04-30 17:56:02 +02003324
3325 return value;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003326}
3327
3328void realmode_set_cr(struct kvm_vcpu *vcpu, int cr, unsigned long val,
3329 unsigned long *rflags)
3330{
3331 switch (cr) {
3332 case 0:
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003333 kvm_set_cr0(vcpu, mk_cr_64(vcpu->arch.cr0, val));
Hollis Blanchard8776e512007-10-31 17:24:24 -05003334 *rflags = kvm_x86_ops->get_rflags(vcpu);
3335 break;
3336 case 2:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003337 vcpu->arch.cr2 = val;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003338 break;
3339 case 3:
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003340 kvm_set_cr3(vcpu, val);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003341 break;
3342 case 4:
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003343 kvm_set_cr4(vcpu, mk_cr_64(vcpu->arch.cr4, val));
Hollis Blanchard8776e512007-10-31 17:24:24 -05003344 break;
Joerg Roedel152ff9b2007-12-06 15:46:52 +01003345 case 8:
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003346 kvm_set_cr8(vcpu, val & 0xfUL);
Joerg Roedel152ff9b2007-12-06 15:46:52 +01003347 break;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003348 default:
Harvey Harrisonb8688d52008-03-03 12:59:56 -08003349 vcpu_printf(vcpu, "%s: unexpected cr %u\n", __func__, cr);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003350 }
3351}
3352
Dan Kenigsberg07716712007-11-21 17:10:04 +02003353static int move_to_next_stateful_cpuid_entry(struct kvm_vcpu *vcpu, int i)
3354{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003355 struct kvm_cpuid_entry2 *e = &vcpu->arch.cpuid_entries[i];
3356 int j, nent = vcpu->arch.cpuid_nent;
Dan Kenigsberg07716712007-11-21 17:10:04 +02003357
3358 e->flags &= ~KVM_CPUID_FLAG_STATE_READ_NEXT;
3359 /* when no next entry is found, the current entry[i] is reselected */
Nitin A Kamble0fdf8e52008-11-05 15:56:21 -08003360 for (j = i + 1; ; j = (j + 1) % nent) {
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003361 struct kvm_cpuid_entry2 *ej = &vcpu->arch.cpuid_entries[j];
Dan Kenigsberg07716712007-11-21 17:10:04 +02003362 if (ej->function == e->function) {
3363 ej->flags |= KVM_CPUID_FLAG_STATE_READ_NEXT;
3364 return j;
3365 }
3366 }
3367 return 0; /* silence gcc, even though control never reaches here */
3368}
3369
3370/* find an entry with matching function, matching index (if needed), and that
3371 * should be read next (if it's stateful) */
3372static int is_matching_cpuid_entry(struct kvm_cpuid_entry2 *e,
3373 u32 function, u32 index)
3374{
3375 if (e->function != function)
3376 return 0;
3377 if ((e->flags & KVM_CPUID_FLAG_SIGNIFCANT_INDEX) && e->index != index)
3378 return 0;
3379 if ((e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC) &&
Amit Shah19355472009-01-14 16:56:00 +00003380 !(e->flags & KVM_CPUID_FLAG_STATE_READ_NEXT))
Dan Kenigsberg07716712007-11-21 17:10:04 +02003381 return 0;
3382 return 1;
3383}
3384
Alexander Grafd8017472008-11-25 20:17:11 +01003385struct kvm_cpuid_entry2 *kvm_find_cpuid_entry(struct kvm_vcpu *vcpu,
3386 u32 function, u32 index)
Hollis Blanchard8776e512007-10-31 17:24:24 -05003387{
3388 int i;
Alexander Grafd8017472008-11-25 20:17:11 +01003389 struct kvm_cpuid_entry2 *best = NULL;
Hollis Blanchard8776e512007-10-31 17:24:24 -05003390
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003391 for (i = 0; i < vcpu->arch.cpuid_nent; ++i) {
Alexander Grafd8017472008-11-25 20:17:11 +01003392 struct kvm_cpuid_entry2 *e;
3393
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003394 e = &vcpu->arch.cpuid_entries[i];
Dan Kenigsberg07716712007-11-21 17:10:04 +02003395 if (is_matching_cpuid_entry(e, function, index)) {
3396 if (e->flags & KVM_CPUID_FLAG_STATEFUL_FUNC)
3397 move_to_next_stateful_cpuid_entry(vcpu, i);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003398 best = e;
3399 break;
3400 }
3401 /*
3402 * Both basic or both extended?
3403 */
3404 if (((e->function ^ function) & 0x80000000) == 0)
3405 if (!best || e->function > best->function)
3406 best = e;
3407 }
Alexander Grafd8017472008-11-25 20:17:11 +01003408 return best;
3409}
3410
Dong, Eddie82725b22009-03-30 16:21:08 +08003411int cpuid_maxphyaddr(struct kvm_vcpu *vcpu)
3412{
3413 struct kvm_cpuid_entry2 *best;
3414
3415 best = kvm_find_cpuid_entry(vcpu, 0x80000008, 0);
3416 if (best)
3417 return best->eax & 0xff;
3418 return 36;
3419}
3420
Alexander Grafd8017472008-11-25 20:17:11 +01003421void kvm_emulate_cpuid(struct kvm_vcpu *vcpu)
3422{
3423 u32 function, index;
3424 struct kvm_cpuid_entry2 *best;
3425
3426 function = kvm_register_read(vcpu, VCPU_REGS_RAX);
3427 index = kvm_register_read(vcpu, VCPU_REGS_RCX);
3428 kvm_register_write(vcpu, VCPU_REGS_RAX, 0);
3429 kvm_register_write(vcpu, VCPU_REGS_RBX, 0);
3430 kvm_register_write(vcpu, VCPU_REGS_RCX, 0);
3431 kvm_register_write(vcpu, VCPU_REGS_RDX, 0);
3432 best = kvm_find_cpuid_entry(vcpu, function, index);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003433 if (best) {
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003434 kvm_register_write(vcpu, VCPU_REGS_RAX, best->eax);
3435 kvm_register_write(vcpu, VCPU_REGS_RBX, best->ebx);
3436 kvm_register_write(vcpu, VCPU_REGS_RCX, best->ecx);
3437 kvm_register_write(vcpu, VCPU_REGS_RDX, best->edx);
Hollis Blanchard8776e512007-10-31 17:24:24 -05003438 }
Hollis Blanchard8776e512007-10-31 17:24:24 -05003439 kvm_x86_ops->skip_emulated_instruction(vcpu);
Marcelo Tosatti229456f2009-06-17 09:22:14 -03003440 trace_kvm_cpuid(function,
3441 kvm_register_read(vcpu, VCPU_REGS_RAX),
3442 kvm_register_read(vcpu, VCPU_REGS_RBX),
3443 kvm_register_read(vcpu, VCPU_REGS_RCX),
3444 kvm_register_read(vcpu, VCPU_REGS_RDX));
Hollis Blanchard8776e512007-10-31 17:24:24 -05003445}
3446EXPORT_SYMBOL_GPL(kvm_emulate_cpuid);
Hollis Blanchardd0752062007-10-31 17:24:25 -05003447
3448/*
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003449 * Check if userspace requested an interrupt window, and that the
3450 * interrupt window is open.
3451 *
3452 * No need to exit to userspace if we already have an interrupt queued.
3453 */
3454static int dm_request_for_irq_injection(struct kvm_vcpu *vcpu,
3455 struct kvm_run *kvm_run)
3456{
Gleb Natapov80618232009-04-21 17:44:56 +03003457 return (!irqchip_in_kernel(vcpu->kvm) && !kvm_cpu_has_interrupt(vcpu) &&
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003458 kvm_run->request_interrupt_window &&
Gleb Natapov5df56642009-04-21 17:44:59 +03003459 kvm_arch_interrupt_allowed(vcpu));
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003460}
3461
3462static void post_kvm_run_save(struct kvm_vcpu *vcpu,
3463 struct kvm_run *kvm_run)
3464{
3465 kvm_run->if_flag = (kvm_x86_ops->get_rflags(vcpu) & X86_EFLAGS_IF) != 0;
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003466 kvm_run->cr8 = kvm_get_cr8(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003467 kvm_run->apic_base = kvm_get_apic_base(vcpu);
Jan Kiszka45312202008-12-11 16:54:54 +01003468 if (irqchip_in_kernel(vcpu->kvm))
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003469 kvm_run->ready_for_interrupt_injection = 1;
Jan Kiszka45312202008-12-11 16:54:54 +01003470 else
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003471 kvm_run->ready_for_interrupt_injection =
Gleb Natapovfa9726b2009-05-11 13:35:47 +03003472 kvm_arch_interrupt_allowed(vcpu) &&
3473 !kvm_cpu_has_interrupt(vcpu) &&
3474 !kvm_event_needs_reinjection(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003475}
3476
Avi Kivityb93463a2007-10-25 16:52:32 +02003477static void vapic_enter(struct kvm_vcpu *vcpu)
3478{
3479 struct kvm_lapic *apic = vcpu->arch.apic;
3480 struct page *page;
3481
3482 if (!apic || !apic->vapic_addr)
3483 return;
3484
3485 page = gfn_to_page(vcpu->kvm, apic->vapic_addr >> PAGE_SHIFT);
Izik Eidus72dc67a2008-02-10 18:04:15 +02003486
3487 vcpu->arch.apic->vapic_page = page;
Avi Kivityb93463a2007-10-25 16:52:32 +02003488}
3489
3490static void vapic_exit(struct kvm_vcpu *vcpu)
3491{
3492 struct kvm_lapic *apic = vcpu->arch.apic;
3493
3494 if (!apic || !apic->vapic_addr)
3495 return;
3496
Marcelo Tosattif8b78fa2008-06-23 12:04:25 -03003497 down_read(&vcpu->kvm->slots_lock);
Avi Kivityb93463a2007-10-25 16:52:32 +02003498 kvm_release_page_dirty(apic->vapic_page);
3499 mark_page_dirty(vcpu->kvm, apic->vapic_addr >> PAGE_SHIFT);
Marcelo Tosattif8b78fa2008-06-23 12:04:25 -03003500 up_read(&vcpu->kvm->slots_lock);
Avi Kivityb93463a2007-10-25 16:52:32 +02003501}
3502
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003503static void update_cr8_intercept(struct kvm_vcpu *vcpu)
3504{
3505 int max_irr, tpr;
3506
3507 if (!kvm_x86_ops->update_cr8_intercept)
3508 return;
3509
Avi Kivity88c808f2009-08-17 22:49:40 +03003510 if (!vcpu->arch.apic)
3511 return;
3512
Gleb Natapov8db3baa2009-05-11 13:35:54 +03003513 if (!vcpu->arch.apic->vapic_addr)
3514 max_irr = kvm_lapic_find_highest_irr(vcpu);
3515 else
3516 max_irr = -1;
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003517
3518 if (max_irr != -1)
3519 max_irr >>= 4;
3520
3521 tpr = kvm_lapic_get_cr8(vcpu);
3522
3523 kvm_x86_ops->update_cr8_intercept(vcpu, tpr, max_irr);
3524}
3525
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03003526static void inject_pending_event(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003527{
3528 /* try to reinject previous events if any */
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03003529 if (vcpu->arch.exception.pending) {
3530 kvm_x86_ops->queue_exception(vcpu, vcpu->arch.exception.nr,
3531 vcpu->arch.exception.has_error_code,
3532 vcpu->arch.exception.error_code);
3533 return;
3534 }
3535
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003536 if (vcpu->arch.nmi_injected) {
3537 kvm_x86_ops->set_nmi(vcpu);
3538 return;
3539 }
3540
3541 if (vcpu->arch.interrupt.pending) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03003542 kvm_x86_ops->set_irq(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003543 return;
3544 }
3545
3546 /* try to inject new event if pending */
3547 if (vcpu->arch.nmi_pending) {
3548 if (kvm_x86_ops->nmi_allowed(vcpu)) {
3549 vcpu->arch.nmi_pending = false;
3550 vcpu->arch.nmi_injected = true;
3551 kvm_x86_ops->set_nmi(vcpu);
3552 }
3553 } else if (kvm_cpu_has_interrupt(vcpu)) {
3554 if (kvm_x86_ops->interrupt_allowed(vcpu)) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03003555 kvm_queue_interrupt(vcpu, kvm_cpu_get_interrupt(vcpu),
3556 false);
3557 kvm_x86_ops->set_irq(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003558 }
3559 }
3560}
3561
Marcelo Tosattid7690172008-09-08 15:23:48 -03003562static int vcpu_enter_guest(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003563{
3564 int r;
Gleb Natapov6a8b1d12009-05-11 13:35:51 +03003565 bool req_int_win = !irqchip_in_kernel(vcpu->kvm) &&
3566 kvm_run->request_interrupt_window;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003567
Marcelo Tosatti2e53d632008-02-20 14:47:24 -05003568 if (vcpu->requests)
3569 if (test_and_clear_bit(KVM_REQ_MMU_RELOAD, &vcpu->requests))
3570 kvm_mmu_unload(vcpu);
3571
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003572 r = kvm_mmu_reload(vcpu);
3573 if (unlikely(r))
3574 goto out;
3575
Avi Kivity2f52d582008-01-16 12:49:30 +02003576 if (vcpu->requests) {
3577 if (test_and_clear_bit(KVM_REQ_MIGRATE_TIMER, &vcpu->requests))
Marcelo Tosatti2f599712008-05-27 12:10:20 -03003578 __kvm_migrate_timers(vcpu);
Gerd Hoffmannc8076602009-02-04 17:52:04 +01003579 if (test_and_clear_bit(KVM_REQ_KVMCLOCK_UPDATE, &vcpu->requests))
3580 kvm_write_guest_time(vcpu);
Marcelo Tosatti4731d4c2008-09-23 13:18:39 -03003581 if (test_and_clear_bit(KVM_REQ_MMU_SYNC, &vcpu->requests))
3582 kvm_mmu_sync_roots(vcpu);
Marcelo Tosattid4acf7e2008-06-06 16:37:35 -03003583 if (test_and_clear_bit(KVM_REQ_TLB_FLUSH, &vcpu->requests))
3584 kvm_x86_ops->tlb_flush(vcpu);
Avi Kivityb93463a2007-10-25 16:52:32 +02003585 if (test_and_clear_bit(KVM_REQ_REPORT_TPR_ACCESS,
3586 &vcpu->requests)) {
3587 kvm_run->exit_reason = KVM_EXIT_TPR_ACCESS;
3588 r = 0;
3589 goto out;
3590 }
Joerg Roedel71c4dfa2008-02-26 16:49:16 +01003591 if (test_and_clear_bit(KVM_REQ_TRIPLE_FAULT, &vcpu->requests)) {
3592 kvm_run->exit_reason = KVM_EXIT_SHUTDOWN;
3593 r = 0;
3594 goto out;
3595 }
Avi Kivity2f52d582008-01-16 12:49:30 +02003596 }
Avi Kivityb93463a2007-10-25 16:52:32 +02003597
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003598 preempt_disable();
3599
3600 kvm_x86_ops->prepare_guest_switch(vcpu);
3601 kvm_load_guest_fpu(vcpu);
3602
3603 local_irq_disable();
3604
Marcelo Tosatti32f88402009-05-07 17:55:12 -03003605 clear_bit(KVM_REQ_KICK, &vcpu->requests);
3606 smp_mb__after_clear_bit();
3607
Marcelo Tosattid7690172008-09-08 15:23:48 -03003608 if (vcpu->requests || need_resched() || signal_pending(current)) {
Gleb Natapovc7f0f242009-07-07 15:27:32 +03003609 set_bit(KVM_REQ_KICK, &vcpu->requests);
Avi Kivity6c1428012008-01-15 18:27:32 +02003610 local_irq_enable();
3611 preempt_enable();
3612 r = 1;
3613 goto out;
3614 }
3615
Gleb Natapovb59bb7b2009-07-09 15:33:51 +03003616 inject_pending_event(vcpu, kvm_run);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003617
Gleb Natapov6a8b1d12009-05-11 13:35:51 +03003618 /* enable NMI/IRQ window open exits if needed */
3619 if (vcpu->arch.nmi_pending)
3620 kvm_x86_ops->enable_nmi_window(vcpu);
3621 else if (kvm_cpu_has_interrupt(vcpu) || req_int_win)
3622 kvm_x86_ops->enable_irq_window(vcpu);
3623
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003624 if (kvm_lapic_enabled(vcpu)) {
Gleb Natapov8db3baa2009-05-11 13:35:54 +03003625 update_cr8_intercept(vcpu);
3626 kvm_lapic_sync_to_vapic(vcpu);
Gleb Natapov95ba8273132009-04-21 17:45:08 +03003627 }
Avi Kivityb93463a2007-10-25 16:52:32 +02003628
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003629 up_read(&vcpu->kvm->slots_lock);
3630
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003631 kvm_guest_enter();
3632
Jan Kiszka42dbaa52008-12-15 13:52:10 +01003633 if (unlikely(vcpu->arch.switch_db_regs)) {
Jan Kiszka42dbaa52008-12-15 13:52:10 +01003634 set_debugreg(0, 7);
3635 set_debugreg(vcpu->arch.eff_db[0], 0);
3636 set_debugreg(vcpu->arch.eff_db[1], 1);
3637 set_debugreg(vcpu->arch.eff_db[2], 2);
3638 set_debugreg(vcpu->arch.eff_db[3], 3);
3639 }
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003640
Marcelo Tosatti229456f2009-06-17 09:22:14 -03003641 trace_kvm_entry(vcpu->vcpu_id);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003642 kvm_x86_ops->run(vcpu, kvm_run);
3643
Avi Kivity3d53c272009-09-01 12:34:07 +03003644 if (unlikely(vcpu->arch.switch_db_regs || test_thread_flag(TIF_DEBUG))) {
3645 set_debugreg(current->thread.debugreg0, 0);
3646 set_debugreg(current->thread.debugreg1, 1);
3647 set_debugreg(current->thread.debugreg2, 2);
3648 set_debugreg(current->thread.debugreg3, 3);
3649 set_debugreg(current->thread.debugreg6, 6);
3650 set_debugreg(current->thread.debugreg7, 7);
Jan Kiszka42dbaa52008-12-15 13:52:10 +01003651 }
Jan Kiszka42dbaa52008-12-15 13:52:10 +01003652
Marcelo Tosatti32f88402009-05-07 17:55:12 -03003653 set_bit(KVM_REQ_KICK, &vcpu->requests);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003654 local_irq_enable();
3655
3656 ++vcpu->stat.exits;
3657
3658 /*
3659 * We must have an instruction between local_irq_enable() and
3660 * kvm_guest_exit(), so the timer interrupt isn't delayed by
3661 * the interrupt shadow. The stat.exits increment will do nicely.
3662 * But we need to prevent reordering, hence this barrier():
3663 */
3664 barrier();
3665
3666 kvm_guest_exit();
3667
3668 preempt_enable();
3669
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003670 down_read(&vcpu->kvm->slots_lock);
3671
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003672 /*
3673 * Profile KVM exit RIPs:
3674 */
3675 if (unlikely(prof_on == KVM_PROFILING)) {
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003676 unsigned long rip = kvm_rip_read(vcpu);
3677 profile_hit(KVM_PROFILING, (void *)rip);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003678 }
3679
Avi Kivity298101d2007-11-25 13:41:11 +02003680
Avi Kivityb93463a2007-10-25 16:52:32 +02003681 kvm_lapic_sync_from_vapic(vcpu);
3682
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003683 r = kvm_x86_ops->handle_exit(kvm_run, vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003684out:
Marcelo Tosattid7690172008-09-08 15:23:48 -03003685 return r;
3686}
3687
Gleb Natapov09cec752009-03-23 15:11:44 +02003688
Marcelo Tosattid7690172008-09-08 15:23:48 -03003689static int __vcpu_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
3690{
3691 int r;
3692
3693 if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_SIPI_RECEIVED)) {
Jan Kiszka1b10bf32008-09-30 10:41:06 +02003694 pr_debug("vcpu %d received sipi with vector # %x\n",
3695 vcpu->vcpu_id, vcpu->arch.sipi_vector);
Marcelo Tosattid7690172008-09-08 15:23:48 -03003696 kvm_lapic_reset(vcpu);
Gleb Natapov5f179282008-10-07 15:42:33 +02003697 r = kvm_arch_vcpu_reset(vcpu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03003698 if (r)
3699 return r;
3700 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003701 }
3702
Marcelo Tosattid7690172008-09-08 15:23:48 -03003703 down_read(&vcpu->kvm->slots_lock);
3704 vapic_enter(vcpu);
3705
3706 r = 1;
3707 while (r > 0) {
Gleb Natapovaf2152f2008-09-22 14:28:53 +03003708 if (vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE)
Marcelo Tosattid7690172008-09-08 15:23:48 -03003709 r = vcpu_enter_guest(vcpu, kvm_run);
3710 else {
3711 up_read(&vcpu->kvm->slots_lock);
3712 kvm_vcpu_block(vcpu);
3713 down_read(&vcpu->kvm->slots_lock);
3714 if (test_and_clear_bit(KVM_REQ_UNHALT, &vcpu->requests))
Gleb Natapov09cec752009-03-23 15:11:44 +02003715 {
3716 switch(vcpu->arch.mp_state) {
3717 case KVM_MP_STATE_HALTED:
Marcelo Tosattid7690172008-09-08 15:23:48 -03003718 vcpu->arch.mp_state =
Gleb Natapov09cec752009-03-23 15:11:44 +02003719 KVM_MP_STATE_RUNNABLE;
3720 case KVM_MP_STATE_RUNNABLE:
3721 break;
3722 case KVM_MP_STATE_SIPI_RECEIVED:
3723 default:
3724 r = -EINTR;
3725 break;
3726 }
3727 }
Marcelo Tosattid7690172008-09-08 15:23:48 -03003728 }
3729
Gleb Natapov09cec752009-03-23 15:11:44 +02003730 if (r <= 0)
3731 break;
3732
3733 clear_bit(KVM_REQ_PENDING_TIMER, &vcpu->requests);
3734 if (kvm_cpu_has_pending_timer(vcpu))
3735 kvm_inject_pending_timer_irqs(vcpu);
3736
3737 if (dm_request_for_irq_injection(vcpu, kvm_run)) {
3738 r = -EINTR;
3739 kvm_run->exit_reason = KVM_EXIT_INTR;
3740 ++vcpu->stat.request_irq_exits;
3741 }
3742 if (signal_pending(current)) {
3743 r = -EINTR;
3744 kvm_run->exit_reason = KVM_EXIT_INTR;
3745 ++vcpu->stat.signal_exits;
3746 }
3747 if (need_resched()) {
3748 up_read(&vcpu->kvm->slots_lock);
3749 kvm_resched(vcpu);
3750 down_read(&vcpu->kvm->slots_lock);
Marcelo Tosattid7690172008-09-08 15:23:48 -03003751 }
3752 }
3753
3754 up_read(&vcpu->kvm->slots_lock);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003755 post_kvm_run_save(vcpu, kvm_run);
3756
Avi Kivityb93463a2007-10-25 16:52:32 +02003757 vapic_exit(vcpu);
3758
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003759 return r;
3760}
3761
3762int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
3763{
3764 int r;
3765 sigset_t sigsaved;
3766
3767 vcpu_load(vcpu);
3768
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003769 if (vcpu->sigset_active)
3770 sigprocmask(SIG_SETMASK, &vcpu->sigset, &sigsaved);
3771
Avi Kivityac9f6dc2008-07-06 15:48:31 +03003772 if (unlikely(vcpu->arch.mp_state == KVM_MP_STATE_UNINITIALIZED)) {
3773 kvm_vcpu_block(vcpu);
Marcelo Tosattid7690172008-09-08 15:23:48 -03003774 clear_bit(KVM_REQ_UNHALT, &vcpu->requests);
Avi Kivityac9f6dc2008-07-06 15:48:31 +03003775 r = -EAGAIN;
3776 goto out;
3777 }
3778
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003779 /* re-sync apic's tpr */
3780 if (!irqchip_in_kernel(vcpu->kvm))
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003781 kvm_set_cr8(vcpu, kvm_run->cr8);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003782
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003783 if (vcpu->arch.pio.cur_count) {
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003784 r = complete_pio(vcpu);
3785 if (r)
3786 goto out;
3787 }
3788#if CONFIG_HAS_IOMEM
3789 if (vcpu->mmio_needed) {
3790 memcpy(vcpu->mmio_data, kvm_run->mmio.data, 8);
3791 vcpu->mmio_read_completed = 1;
3792 vcpu->mmio_needed = 0;
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003793
3794 down_read(&vcpu->kvm->slots_lock);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003795 r = emulate_instruction(vcpu, kvm_run,
Sheng Yang571008d2008-01-02 14:49:22 +08003796 vcpu->arch.mmio_fault_cr2, 0,
3797 EMULTYPE_NO_DECODE);
Marcelo Tosatti3200f402008-03-29 20:17:59 -03003798 up_read(&vcpu->kvm->slots_lock);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003799 if (r == EMULATE_DO_MMIO) {
3800 /*
3801 * Read-modify-write. Back to userspace.
3802 */
3803 r = 0;
3804 goto out;
3805 }
3806 }
3807#endif
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003808 if (kvm_run->exit_reason == KVM_EXIT_HYPERCALL)
3809 kvm_register_write(vcpu, VCPU_REGS_RAX,
3810 kvm_run->hypercall.ret);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003811
3812 r = __vcpu_run(vcpu, kvm_run);
3813
3814out:
3815 if (vcpu->sigset_active)
3816 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
3817
3818 vcpu_put(vcpu);
3819 return r;
3820}
3821
3822int kvm_arch_vcpu_ioctl_get_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs)
3823{
3824 vcpu_load(vcpu);
3825
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003826 regs->rax = kvm_register_read(vcpu, VCPU_REGS_RAX);
3827 regs->rbx = kvm_register_read(vcpu, VCPU_REGS_RBX);
3828 regs->rcx = kvm_register_read(vcpu, VCPU_REGS_RCX);
3829 regs->rdx = kvm_register_read(vcpu, VCPU_REGS_RDX);
3830 regs->rsi = kvm_register_read(vcpu, VCPU_REGS_RSI);
3831 regs->rdi = kvm_register_read(vcpu, VCPU_REGS_RDI);
3832 regs->rsp = kvm_register_read(vcpu, VCPU_REGS_RSP);
3833 regs->rbp = kvm_register_read(vcpu, VCPU_REGS_RBP);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003834#ifdef CONFIG_X86_64
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003835 regs->r8 = kvm_register_read(vcpu, VCPU_REGS_R8);
3836 regs->r9 = kvm_register_read(vcpu, VCPU_REGS_R9);
3837 regs->r10 = kvm_register_read(vcpu, VCPU_REGS_R10);
3838 regs->r11 = kvm_register_read(vcpu, VCPU_REGS_R11);
3839 regs->r12 = kvm_register_read(vcpu, VCPU_REGS_R12);
3840 regs->r13 = kvm_register_read(vcpu, VCPU_REGS_R13);
3841 regs->r14 = kvm_register_read(vcpu, VCPU_REGS_R14);
3842 regs->r15 = kvm_register_read(vcpu, VCPU_REGS_R15);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003843#endif
3844
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003845 regs->rip = kvm_rip_read(vcpu);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003846 regs->rflags = kvm_x86_ops->get_rflags(vcpu);
3847
3848 /*
3849 * Don't leak debug flags in case they were set for guest debugging
3850 */
Jan Kiszkad0bfb942008-12-15 13:52:10 +01003851 if (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003852 regs->rflags &= ~(X86_EFLAGS_TF | X86_EFLAGS_RF);
3853
3854 vcpu_put(vcpu);
3855
3856 return 0;
3857}
3858
3859int kvm_arch_vcpu_ioctl_set_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs)
3860{
3861 vcpu_load(vcpu);
3862
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003863 kvm_register_write(vcpu, VCPU_REGS_RAX, regs->rax);
3864 kvm_register_write(vcpu, VCPU_REGS_RBX, regs->rbx);
3865 kvm_register_write(vcpu, VCPU_REGS_RCX, regs->rcx);
3866 kvm_register_write(vcpu, VCPU_REGS_RDX, regs->rdx);
3867 kvm_register_write(vcpu, VCPU_REGS_RSI, regs->rsi);
3868 kvm_register_write(vcpu, VCPU_REGS_RDI, regs->rdi);
3869 kvm_register_write(vcpu, VCPU_REGS_RSP, regs->rsp);
3870 kvm_register_write(vcpu, VCPU_REGS_RBP, regs->rbp);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003871#ifdef CONFIG_X86_64
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003872 kvm_register_write(vcpu, VCPU_REGS_R8, regs->r8);
3873 kvm_register_write(vcpu, VCPU_REGS_R9, regs->r9);
3874 kvm_register_write(vcpu, VCPU_REGS_R10, regs->r10);
3875 kvm_register_write(vcpu, VCPU_REGS_R11, regs->r11);
3876 kvm_register_write(vcpu, VCPU_REGS_R12, regs->r12);
3877 kvm_register_write(vcpu, VCPU_REGS_R13, regs->r13);
3878 kvm_register_write(vcpu, VCPU_REGS_R14, regs->r14);
3879 kvm_register_write(vcpu, VCPU_REGS_R15, regs->r15);
3880
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003881#endif
3882
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03003883 kvm_rip_write(vcpu, regs->rip);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003884 kvm_x86_ops->set_rflags(vcpu, regs->rflags);
3885
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003886
Jan Kiszkab4f14ab2008-04-30 17:59:04 +02003887 vcpu->arch.exception.pending = false;
3888
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003889 vcpu_put(vcpu);
3890
3891 return 0;
3892}
3893
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02003894void kvm_get_segment(struct kvm_vcpu *vcpu,
3895 struct kvm_segment *var, int seg)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003896{
Harvey Harrison14af3f32008-02-19 10:25:50 -08003897 kvm_x86_ops->get_segment(vcpu, var, seg);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003898}
3899
3900void kvm_get_cs_db_l_bits(struct kvm_vcpu *vcpu, int *db, int *l)
3901{
3902 struct kvm_segment cs;
3903
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02003904 kvm_get_segment(vcpu, &cs, VCPU_SREG_CS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003905 *db = cs.db;
3906 *l = cs.l;
3907}
3908EXPORT_SYMBOL_GPL(kvm_get_cs_db_l_bits);
3909
3910int kvm_arch_vcpu_ioctl_get_sregs(struct kvm_vcpu *vcpu,
3911 struct kvm_sregs *sregs)
3912{
3913 struct descriptor_table dt;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003914
3915 vcpu_load(vcpu);
3916
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02003917 kvm_get_segment(vcpu, &sregs->cs, VCPU_SREG_CS);
3918 kvm_get_segment(vcpu, &sregs->ds, VCPU_SREG_DS);
3919 kvm_get_segment(vcpu, &sregs->es, VCPU_SREG_ES);
3920 kvm_get_segment(vcpu, &sregs->fs, VCPU_SREG_FS);
3921 kvm_get_segment(vcpu, &sregs->gs, VCPU_SREG_GS);
3922 kvm_get_segment(vcpu, &sregs->ss, VCPU_SREG_SS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003923
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02003924 kvm_get_segment(vcpu, &sregs->tr, VCPU_SREG_TR);
3925 kvm_get_segment(vcpu, &sregs->ldt, VCPU_SREG_LDTR);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003926
3927 kvm_x86_ops->get_idt(vcpu, &dt);
3928 sregs->idt.limit = dt.limit;
3929 sregs->idt.base = dt.base;
3930 kvm_x86_ops->get_gdt(vcpu, &dt);
3931 sregs->gdt.limit = dt.limit;
3932 sregs->gdt.base = dt.base;
3933
3934 kvm_x86_ops->decache_cr4_guest_bits(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003935 sregs->cr0 = vcpu->arch.cr0;
3936 sregs->cr2 = vcpu->arch.cr2;
3937 sregs->cr3 = vcpu->arch.cr3;
3938 sregs->cr4 = vcpu->arch.cr4;
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02003939 sregs->cr8 = kvm_get_cr8(vcpu);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08003940 sregs->efer = vcpu->arch.shadow_efer;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003941 sregs->apic_base = kvm_get_apic_base(vcpu);
3942
Gleb Natapov923c61b2009-05-11 13:35:48 +03003943 memset(sregs->interrupt_bitmap, 0, sizeof sregs->interrupt_bitmap);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003944
Gleb Natapov36752c92009-05-11 13:35:53 +03003945 if (vcpu->arch.interrupt.pending && !vcpu->arch.interrupt.soft)
Gleb Natapov14d0bc12009-04-21 17:45:11 +03003946 set_bit(vcpu->arch.interrupt.nr,
3947 (unsigned long *)sregs->interrupt_bitmap);
Gleb Natapov16d7a192009-04-21 17:45:10 +03003948
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003949 vcpu_put(vcpu);
3950
3951 return 0;
3952}
3953
Marcelo Tosatti62d9f0d2008-04-11 13:24:45 -03003954int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
3955 struct kvm_mp_state *mp_state)
3956{
3957 vcpu_load(vcpu);
3958 mp_state->mp_state = vcpu->arch.mp_state;
3959 vcpu_put(vcpu);
3960 return 0;
3961}
3962
3963int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
3964 struct kvm_mp_state *mp_state)
3965{
3966 vcpu_load(vcpu);
3967 vcpu->arch.mp_state = mp_state->mp_state;
3968 vcpu_put(vcpu);
3969 return 0;
3970}
3971
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02003972static void kvm_set_segment(struct kvm_vcpu *vcpu,
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003973 struct kvm_segment *var, int seg)
3974{
Harvey Harrison14af3f32008-02-19 10:25:50 -08003975 kvm_x86_ops->set_segment(vcpu, var, seg);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05003976}
3977
Izik Eidus37817f22008-03-24 23:14:53 +02003978static void seg_desct_to_kvm_desct(struct desc_struct *seg_desc, u16 selector,
3979 struct kvm_segment *kvm_desct)
3980{
Akinobu Mita46a359e2009-07-18 23:58:32 +09003981 kvm_desct->base = get_desc_base(seg_desc);
3982 kvm_desct->limit = get_desc_limit(seg_desc);
Marcelo Tosattic93cd3a2008-07-19 19:08:07 -03003983 if (seg_desc->g) {
3984 kvm_desct->limit <<= 12;
3985 kvm_desct->limit |= 0xfff;
3986 }
Izik Eidus37817f22008-03-24 23:14:53 +02003987 kvm_desct->selector = selector;
3988 kvm_desct->type = seg_desc->type;
3989 kvm_desct->present = seg_desc->p;
3990 kvm_desct->dpl = seg_desc->dpl;
3991 kvm_desct->db = seg_desc->d;
3992 kvm_desct->s = seg_desc->s;
3993 kvm_desct->l = seg_desc->l;
3994 kvm_desct->g = seg_desc->g;
3995 kvm_desct->avl = seg_desc->avl;
3996 if (!selector)
3997 kvm_desct->unusable = 1;
3998 else
3999 kvm_desct->unusable = 0;
4000 kvm_desct->padding = 0;
4001}
4002
Amit Shahb8222ad2008-10-22 16:39:47 +05304003static void get_segment_descriptor_dtable(struct kvm_vcpu *vcpu,
4004 u16 selector,
4005 struct descriptor_table *dtable)
Izik Eidus37817f22008-03-24 23:14:53 +02004006{
4007 if (selector & 1 << 2) {
4008 struct kvm_segment kvm_seg;
4009
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004010 kvm_get_segment(vcpu, &kvm_seg, VCPU_SREG_LDTR);
Izik Eidus37817f22008-03-24 23:14:53 +02004011
4012 if (kvm_seg.unusable)
4013 dtable->limit = 0;
4014 else
4015 dtable->limit = kvm_seg.limit;
4016 dtable->base = kvm_seg.base;
4017 }
4018 else
4019 kvm_x86_ops->get_gdt(vcpu, dtable);
4020}
4021
4022/* allowed just for 8 bytes segments */
4023static int load_guest_segment_descriptor(struct kvm_vcpu *vcpu, u16 selector,
4024 struct desc_struct *seg_desc)
4025{
4026 struct descriptor_table dtable;
4027 u16 index = selector >> 3;
4028
Amit Shahb8222ad2008-10-22 16:39:47 +05304029 get_segment_descriptor_dtable(vcpu, selector, &dtable);
Izik Eidus37817f22008-03-24 23:14:53 +02004030
4031 if (dtable.limit < index * 8 + 7) {
4032 kvm_queue_exception_e(vcpu, GP_VECTOR, selector & 0xfffc);
4033 return 1;
4034 }
Mikhail Ershovd9048d32009-08-19 14:08:07 +04004035 return kvm_read_guest_virt(dtable.base + index*8, seg_desc, sizeof(*seg_desc), vcpu);
Izik Eidus37817f22008-03-24 23:14:53 +02004036}
4037
4038/* allowed just for 8 bytes segments */
4039static int save_guest_segment_descriptor(struct kvm_vcpu *vcpu, u16 selector,
4040 struct desc_struct *seg_desc)
4041{
4042 struct descriptor_table dtable;
4043 u16 index = selector >> 3;
4044
Amit Shahb8222ad2008-10-22 16:39:47 +05304045 get_segment_descriptor_dtable(vcpu, selector, &dtable);
Izik Eidus37817f22008-03-24 23:14:53 +02004046
4047 if (dtable.limit < index * 8 + 7)
4048 return 1;
Mikhail Ershovd9048d32009-08-19 14:08:07 +04004049 return kvm_write_guest_virt(dtable.base + index*8, seg_desc, sizeof(*seg_desc), vcpu);
Izik Eidus37817f22008-03-24 23:14:53 +02004050}
4051
4052static u32 get_tss_base_addr(struct kvm_vcpu *vcpu,
4053 struct desc_struct *seg_desc)
4054{
Akinobu Mita46a359e2009-07-18 23:58:32 +09004055 u32 base_addr = get_desc_base(seg_desc);
Izik Eidus37817f22008-03-24 23:14:53 +02004056
Marcelo Tosatti98899aa2008-07-16 19:07:10 -03004057 return vcpu->arch.mmu.gva_to_gpa(vcpu, base_addr);
Izik Eidus37817f22008-03-24 23:14:53 +02004058}
4059
Izik Eidus37817f22008-03-24 23:14:53 +02004060static u16 get_segment_selector(struct kvm_vcpu *vcpu, int seg)
4061{
4062 struct kvm_segment kvm_seg;
4063
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004064 kvm_get_segment(vcpu, &kvm_seg, seg);
Izik Eidus37817f22008-03-24 23:14:53 +02004065 return kvm_seg.selector;
4066}
4067
4068static int load_segment_descriptor_to_kvm_desct(struct kvm_vcpu *vcpu,
4069 u16 selector,
4070 struct kvm_segment *kvm_seg)
4071{
4072 struct desc_struct seg_desc;
4073
4074 if (load_guest_segment_descriptor(vcpu, selector, &seg_desc))
4075 return 1;
4076 seg_desct_to_kvm_desct(&seg_desc, selector, kvm_seg);
4077 return 0;
4078}
4079
Harvey Harrison2259e3a2008-08-22 13:29:17 -07004080static int kvm_load_realmode_segment(struct kvm_vcpu *vcpu, u16 selector, int seg)
Avi Kivityf4bbd9a2008-08-20 15:51:42 +03004081{
4082 struct kvm_segment segvar = {
4083 .base = selector << 4,
4084 .limit = 0xffff,
4085 .selector = selector,
4086 .type = 3,
4087 .present = 1,
4088 .dpl = 3,
4089 .db = 0,
4090 .s = 1,
4091 .l = 0,
4092 .g = 0,
4093 .avl = 0,
4094 .unusable = 0,
4095 };
4096 kvm_x86_ops->set_segment(vcpu, &segvar, seg);
4097 return 0;
4098}
4099
Anthony Liguoric0c7c042009-08-11 15:57:59 -05004100static int is_vm86_segment(struct kvm_vcpu *vcpu, int seg)
4101{
4102 return (seg != VCPU_SREG_LDTR) &&
4103 (seg != VCPU_SREG_TR) &&
4104 (kvm_x86_ops->get_rflags(vcpu) & X86_EFLAGS_VM);
4105}
4106
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004107int kvm_load_segment_descriptor(struct kvm_vcpu *vcpu, u16 selector,
4108 int type_bits, int seg)
Izik Eidus37817f22008-03-24 23:14:53 +02004109{
4110 struct kvm_segment kvm_seg;
4111
Anthony Liguoric0c7c042009-08-11 15:57:59 -05004112 if (is_vm86_segment(vcpu, seg) || !(vcpu->arch.cr0 & X86_CR0_PE))
Avi Kivityf4bbd9a2008-08-20 15:51:42 +03004113 return kvm_load_realmode_segment(vcpu, selector, seg);
Izik Eidus37817f22008-03-24 23:14:53 +02004114 if (load_segment_descriptor_to_kvm_desct(vcpu, selector, &kvm_seg))
4115 return 1;
4116 kvm_seg.type |= type_bits;
4117
4118 if (seg != VCPU_SREG_SS && seg != VCPU_SREG_CS &&
4119 seg != VCPU_SREG_LDTR)
4120 if (!kvm_seg.s)
4121 kvm_seg.unusable = 1;
4122
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004123 kvm_set_segment(vcpu, &kvm_seg, seg);
Izik Eidus37817f22008-03-24 23:14:53 +02004124 return 0;
4125}
4126
4127static void save_state_to_tss32(struct kvm_vcpu *vcpu,
4128 struct tss_segment_32 *tss)
4129{
4130 tss->cr3 = vcpu->arch.cr3;
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004131 tss->eip = kvm_rip_read(vcpu);
Izik Eidus37817f22008-03-24 23:14:53 +02004132 tss->eflags = kvm_x86_ops->get_rflags(vcpu);
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004133 tss->eax = kvm_register_read(vcpu, VCPU_REGS_RAX);
4134 tss->ecx = kvm_register_read(vcpu, VCPU_REGS_RCX);
4135 tss->edx = kvm_register_read(vcpu, VCPU_REGS_RDX);
4136 tss->ebx = kvm_register_read(vcpu, VCPU_REGS_RBX);
4137 tss->esp = kvm_register_read(vcpu, VCPU_REGS_RSP);
4138 tss->ebp = kvm_register_read(vcpu, VCPU_REGS_RBP);
4139 tss->esi = kvm_register_read(vcpu, VCPU_REGS_RSI);
4140 tss->edi = kvm_register_read(vcpu, VCPU_REGS_RDI);
Izik Eidus37817f22008-03-24 23:14:53 +02004141 tss->es = get_segment_selector(vcpu, VCPU_SREG_ES);
4142 tss->cs = get_segment_selector(vcpu, VCPU_SREG_CS);
4143 tss->ss = get_segment_selector(vcpu, VCPU_SREG_SS);
4144 tss->ds = get_segment_selector(vcpu, VCPU_SREG_DS);
4145 tss->fs = get_segment_selector(vcpu, VCPU_SREG_FS);
4146 tss->gs = get_segment_selector(vcpu, VCPU_SREG_GS);
4147 tss->ldt_selector = get_segment_selector(vcpu, VCPU_SREG_LDTR);
Izik Eidus37817f22008-03-24 23:14:53 +02004148}
4149
4150static int load_state_from_tss32(struct kvm_vcpu *vcpu,
4151 struct tss_segment_32 *tss)
4152{
4153 kvm_set_cr3(vcpu, tss->cr3);
4154
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004155 kvm_rip_write(vcpu, tss->eip);
Izik Eidus37817f22008-03-24 23:14:53 +02004156 kvm_x86_ops->set_rflags(vcpu, tss->eflags | 2);
4157
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004158 kvm_register_write(vcpu, VCPU_REGS_RAX, tss->eax);
4159 kvm_register_write(vcpu, VCPU_REGS_RCX, tss->ecx);
4160 kvm_register_write(vcpu, VCPU_REGS_RDX, tss->edx);
4161 kvm_register_write(vcpu, VCPU_REGS_RBX, tss->ebx);
4162 kvm_register_write(vcpu, VCPU_REGS_RSP, tss->esp);
4163 kvm_register_write(vcpu, VCPU_REGS_RBP, tss->ebp);
4164 kvm_register_write(vcpu, VCPU_REGS_RSI, tss->esi);
4165 kvm_register_write(vcpu, VCPU_REGS_RDI, tss->edi);
Izik Eidus37817f22008-03-24 23:14:53 +02004166
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004167 if (kvm_load_segment_descriptor(vcpu, tss->ldt_selector, 0, VCPU_SREG_LDTR))
Izik Eidus37817f22008-03-24 23:14:53 +02004168 return 1;
4169
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004170 if (kvm_load_segment_descriptor(vcpu, tss->es, 1, VCPU_SREG_ES))
Izik Eidus37817f22008-03-24 23:14:53 +02004171 return 1;
4172
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004173 if (kvm_load_segment_descriptor(vcpu, tss->cs, 9, VCPU_SREG_CS))
Izik Eidus37817f22008-03-24 23:14:53 +02004174 return 1;
4175
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004176 if (kvm_load_segment_descriptor(vcpu, tss->ss, 1, VCPU_SREG_SS))
Izik Eidus37817f22008-03-24 23:14:53 +02004177 return 1;
4178
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004179 if (kvm_load_segment_descriptor(vcpu, tss->ds, 1, VCPU_SREG_DS))
Izik Eidus37817f22008-03-24 23:14:53 +02004180 return 1;
4181
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004182 if (kvm_load_segment_descriptor(vcpu, tss->fs, 1, VCPU_SREG_FS))
Izik Eidus37817f22008-03-24 23:14:53 +02004183 return 1;
4184
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004185 if (kvm_load_segment_descriptor(vcpu, tss->gs, 1, VCPU_SREG_GS))
Izik Eidus37817f22008-03-24 23:14:53 +02004186 return 1;
4187 return 0;
4188}
4189
4190static void save_state_to_tss16(struct kvm_vcpu *vcpu,
4191 struct tss_segment_16 *tss)
4192{
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004193 tss->ip = kvm_rip_read(vcpu);
Izik Eidus37817f22008-03-24 23:14:53 +02004194 tss->flag = kvm_x86_ops->get_rflags(vcpu);
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004195 tss->ax = kvm_register_read(vcpu, VCPU_REGS_RAX);
4196 tss->cx = kvm_register_read(vcpu, VCPU_REGS_RCX);
4197 tss->dx = kvm_register_read(vcpu, VCPU_REGS_RDX);
4198 tss->bx = kvm_register_read(vcpu, VCPU_REGS_RBX);
4199 tss->sp = kvm_register_read(vcpu, VCPU_REGS_RSP);
4200 tss->bp = kvm_register_read(vcpu, VCPU_REGS_RBP);
4201 tss->si = kvm_register_read(vcpu, VCPU_REGS_RSI);
4202 tss->di = kvm_register_read(vcpu, VCPU_REGS_RDI);
Izik Eidus37817f22008-03-24 23:14:53 +02004203
4204 tss->es = get_segment_selector(vcpu, VCPU_SREG_ES);
4205 tss->cs = get_segment_selector(vcpu, VCPU_SREG_CS);
4206 tss->ss = get_segment_selector(vcpu, VCPU_SREG_SS);
4207 tss->ds = get_segment_selector(vcpu, VCPU_SREG_DS);
4208 tss->ldt = get_segment_selector(vcpu, VCPU_SREG_LDTR);
4209 tss->prev_task_link = get_segment_selector(vcpu, VCPU_SREG_TR);
4210}
4211
4212static int load_state_from_tss16(struct kvm_vcpu *vcpu,
4213 struct tss_segment_16 *tss)
4214{
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004215 kvm_rip_write(vcpu, tss->ip);
Izik Eidus37817f22008-03-24 23:14:53 +02004216 kvm_x86_ops->set_rflags(vcpu, tss->flag | 2);
Marcelo Tosatti5fdbf972008-06-27 14:58:02 -03004217 kvm_register_write(vcpu, VCPU_REGS_RAX, tss->ax);
4218 kvm_register_write(vcpu, VCPU_REGS_RCX, tss->cx);
4219 kvm_register_write(vcpu, VCPU_REGS_RDX, tss->dx);
4220 kvm_register_write(vcpu, VCPU_REGS_RBX, tss->bx);
4221 kvm_register_write(vcpu, VCPU_REGS_RSP, tss->sp);
4222 kvm_register_write(vcpu, VCPU_REGS_RBP, tss->bp);
4223 kvm_register_write(vcpu, VCPU_REGS_RSI, tss->si);
4224 kvm_register_write(vcpu, VCPU_REGS_RDI, tss->di);
Izik Eidus37817f22008-03-24 23:14:53 +02004225
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004226 if (kvm_load_segment_descriptor(vcpu, tss->ldt, 0, VCPU_SREG_LDTR))
Izik Eidus37817f22008-03-24 23:14:53 +02004227 return 1;
4228
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004229 if (kvm_load_segment_descriptor(vcpu, tss->es, 1, VCPU_SREG_ES))
Izik Eidus37817f22008-03-24 23:14:53 +02004230 return 1;
4231
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004232 if (kvm_load_segment_descriptor(vcpu, tss->cs, 9, VCPU_SREG_CS))
Izik Eidus37817f22008-03-24 23:14:53 +02004233 return 1;
4234
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004235 if (kvm_load_segment_descriptor(vcpu, tss->ss, 1, VCPU_SREG_SS))
Izik Eidus37817f22008-03-24 23:14:53 +02004236 return 1;
4237
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004238 if (kvm_load_segment_descriptor(vcpu, tss->ds, 1, VCPU_SREG_DS))
Izik Eidus37817f22008-03-24 23:14:53 +02004239 return 1;
4240 return 0;
4241}
4242
Harvey Harrison8b2cf732008-04-27 12:14:13 -07004243static int kvm_task_switch_16(struct kvm_vcpu *vcpu, u16 tss_selector,
Gleb Natapovb237ac32009-03-30 16:03:24 +03004244 u16 old_tss_sel, u32 old_tss_base,
4245 struct desc_struct *nseg_desc)
Izik Eidus37817f22008-03-24 23:14:53 +02004246{
4247 struct tss_segment_16 tss_segment_16;
4248 int ret = 0;
4249
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004250 if (kvm_read_guest(vcpu->kvm, old_tss_base, &tss_segment_16,
4251 sizeof tss_segment_16))
Izik Eidus37817f22008-03-24 23:14:53 +02004252 goto out;
4253
4254 save_state_to_tss16(vcpu, &tss_segment_16);
Izik Eidus37817f22008-03-24 23:14:53 +02004255
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004256 if (kvm_write_guest(vcpu->kvm, old_tss_base, &tss_segment_16,
4257 sizeof tss_segment_16))
Izik Eidus37817f22008-03-24 23:14:53 +02004258 goto out;
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004259
4260 if (kvm_read_guest(vcpu->kvm, get_tss_base_addr(vcpu, nseg_desc),
4261 &tss_segment_16, sizeof tss_segment_16))
4262 goto out;
4263
Gleb Natapovb237ac32009-03-30 16:03:24 +03004264 if (old_tss_sel != 0xffff) {
4265 tss_segment_16.prev_task_link = old_tss_sel;
4266
4267 if (kvm_write_guest(vcpu->kvm,
4268 get_tss_base_addr(vcpu, nseg_desc),
4269 &tss_segment_16.prev_task_link,
4270 sizeof tss_segment_16.prev_task_link))
4271 goto out;
4272 }
4273
Izik Eidus37817f22008-03-24 23:14:53 +02004274 if (load_state_from_tss16(vcpu, &tss_segment_16))
4275 goto out;
4276
4277 ret = 1;
4278out:
4279 return ret;
4280}
4281
Harvey Harrison8b2cf732008-04-27 12:14:13 -07004282static int kvm_task_switch_32(struct kvm_vcpu *vcpu, u16 tss_selector,
Gleb Natapovb237ac32009-03-30 16:03:24 +03004283 u16 old_tss_sel, u32 old_tss_base,
Izik Eidus37817f22008-03-24 23:14:53 +02004284 struct desc_struct *nseg_desc)
4285{
4286 struct tss_segment_32 tss_segment_32;
4287 int ret = 0;
4288
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004289 if (kvm_read_guest(vcpu->kvm, old_tss_base, &tss_segment_32,
4290 sizeof tss_segment_32))
Izik Eidus37817f22008-03-24 23:14:53 +02004291 goto out;
4292
4293 save_state_to_tss32(vcpu, &tss_segment_32);
Izik Eidus37817f22008-03-24 23:14:53 +02004294
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004295 if (kvm_write_guest(vcpu->kvm, old_tss_base, &tss_segment_32,
4296 sizeof tss_segment_32))
Izik Eidus37817f22008-03-24 23:14:53 +02004297 goto out;
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004298
4299 if (kvm_read_guest(vcpu->kvm, get_tss_base_addr(vcpu, nseg_desc),
4300 &tss_segment_32, sizeof tss_segment_32))
4301 goto out;
4302
Gleb Natapovb237ac32009-03-30 16:03:24 +03004303 if (old_tss_sel != 0xffff) {
4304 tss_segment_32.prev_task_link = old_tss_sel;
4305
4306 if (kvm_write_guest(vcpu->kvm,
4307 get_tss_base_addr(vcpu, nseg_desc),
4308 &tss_segment_32.prev_task_link,
4309 sizeof tss_segment_32.prev_task_link))
4310 goto out;
4311 }
4312
Izik Eidus37817f22008-03-24 23:14:53 +02004313 if (load_state_from_tss32(vcpu, &tss_segment_32))
4314 goto out;
4315
4316 ret = 1;
4317out:
4318 return ret;
4319}
4320
4321int kvm_task_switch(struct kvm_vcpu *vcpu, u16 tss_selector, int reason)
4322{
4323 struct kvm_segment tr_seg;
4324 struct desc_struct cseg_desc;
4325 struct desc_struct nseg_desc;
4326 int ret = 0;
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004327 u32 old_tss_base = get_segment_base(vcpu, VCPU_SREG_TR);
4328 u16 old_tss_sel = get_segment_selector(vcpu, VCPU_SREG_TR);
Izik Eidus37817f22008-03-24 23:14:53 +02004329
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004330 old_tss_base = vcpu->arch.mmu.gva_to_gpa(vcpu, old_tss_base);
Izik Eidus37817f22008-03-24 23:14:53 +02004331
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004332 /* FIXME: Handle errors. Failure to read either TSS or their
4333 * descriptors should generate a pagefault.
4334 */
Izik Eidus37817f22008-03-24 23:14:53 +02004335 if (load_guest_segment_descriptor(vcpu, tss_selector, &nseg_desc))
4336 goto out;
4337
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004338 if (load_guest_segment_descriptor(vcpu, old_tss_sel, &cseg_desc))
Izik Eidus37817f22008-03-24 23:14:53 +02004339 goto out;
4340
Izik Eidus37817f22008-03-24 23:14:53 +02004341 if (reason != TASK_SWITCH_IRET) {
4342 int cpl;
4343
4344 cpl = kvm_x86_ops->get_cpl(vcpu);
4345 if ((tss_selector & 3) > nseg_desc.dpl || cpl > nseg_desc.dpl) {
4346 kvm_queue_exception_e(vcpu, GP_VECTOR, 0);
4347 return 1;
4348 }
4349 }
4350
Akinobu Mita46a359e2009-07-18 23:58:32 +09004351 if (!nseg_desc.p || get_desc_limit(&nseg_desc) < 0x67) {
Izik Eidus37817f22008-03-24 23:14:53 +02004352 kvm_queue_exception_e(vcpu, TS_VECTOR, tss_selector & 0xfffc);
4353 return 1;
4354 }
4355
4356 if (reason == TASK_SWITCH_IRET || reason == TASK_SWITCH_JMP) {
Izik Eidus3fe913e2008-04-28 18:23:52 +03004357 cseg_desc.type &= ~(1 << 1); //clear the B flag
Marcelo Tosatti34198bf82008-07-16 19:07:11 -03004358 save_guest_segment_descriptor(vcpu, old_tss_sel, &cseg_desc);
Izik Eidus37817f22008-03-24 23:14:53 +02004359 }
4360
4361 if (reason == TASK_SWITCH_IRET) {
4362 u32 eflags = kvm_x86_ops->get_rflags(vcpu);
4363 kvm_x86_ops->set_rflags(vcpu, eflags & ~X86_EFLAGS_NT);
4364 }
4365
Gleb Natapov64a7ec02009-03-30 16:03:29 +03004366 /* set back link to prev task only if NT bit is set in eflags
4367 note that old_tss_sel is not used afetr this point */
4368 if (reason != TASK_SWITCH_CALL && reason != TASK_SWITCH_GATE)
4369 old_tss_sel = 0xffff;
Izik Eidus37817f22008-03-24 23:14:53 +02004370
Gleb Natapovb237ac32009-03-30 16:03:24 +03004371 /* set back link to prev task only if NT bit is set in eflags
4372 note that old_tss_sel is not used afetr this point */
4373 if (reason != TASK_SWITCH_CALL && reason != TASK_SWITCH_GATE)
4374 old_tss_sel = 0xffff;
4375
Izik Eidus37817f22008-03-24 23:14:53 +02004376 if (nseg_desc.type & 8)
Gleb Natapovb237ac32009-03-30 16:03:24 +03004377 ret = kvm_task_switch_32(vcpu, tss_selector, old_tss_sel,
4378 old_tss_base, &nseg_desc);
Izik Eidus37817f22008-03-24 23:14:53 +02004379 else
Gleb Natapovb237ac32009-03-30 16:03:24 +03004380 ret = kvm_task_switch_16(vcpu, tss_selector, old_tss_sel,
4381 old_tss_base, &nseg_desc);
Izik Eidus37817f22008-03-24 23:14:53 +02004382
4383 if (reason == TASK_SWITCH_CALL || reason == TASK_SWITCH_GATE) {
4384 u32 eflags = kvm_x86_ops->get_rflags(vcpu);
4385 kvm_x86_ops->set_rflags(vcpu, eflags | X86_EFLAGS_NT);
4386 }
4387
4388 if (reason != TASK_SWITCH_IRET) {
Izik Eidus3fe913e2008-04-28 18:23:52 +03004389 nseg_desc.type |= (1 << 1);
Izik Eidus37817f22008-03-24 23:14:53 +02004390 save_guest_segment_descriptor(vcpu, tss_selector,
4391 &nseg_desc);
4392 }
4393
4394 kvm_x86_ops->set_cr0(vcpu, vcpu->arch.cr0 | X86_CR0_TS);
4395 seg_desct_to_kvm_desct(&nseg_desc, tss_selector, &tr_seg);
4396 tr_seg.type = 11;
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004397 kvm_set_segment(vcpu, &tr_seg, VCPU_SREG_TR);
Izik Eidus37817f22008-03-24 23:14:53 +02004398out:
Izik Eidus37817f22008-03-24 23:14:53 +02004399 return ret;
4400}
4401EXPORT_SYMBOL_GPL(kvm_task_switch);
4402
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004403int kvm_arch_vcpu_ioctl_set_sregs(struct kvm_vcpu *vcpu,
4404 struct kvm_sregs *sregs)
4405{
4406 int mmu_reset_needed = 0;
Gleb Natapov923c61b2009-05-11 13:35:48 +03004407 int pending_vec, max_bits;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004408 struct descriptor_table dt;
4409
4410 vcpu_load(vcpu);
4411
4412 dt.limit = sregs->idt.limit;
4413 dt.base = sregs->idt.base;
4414 kvm_x86_ops->set_idt(vcpu, &dt);
4415 dt.limit = sregs->gdt.limit;
4416 dt.base = sregs->gdt.base;
4417 kvm_x86_ops->set_gdt(vcpu, &dt);
4418
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004419 vcpu->arch.cr2 = sregs->cr2;
4420 mmu_reset_needed |= vcpu->arch.cr3 != sregs->cr3;
Jan Kiszkadc7e7952009-07-01 20:52:03 +02004421 vcpu->arch.cr3 = sregs->cr3;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004422
Avi Kivity2d3ad1f2008-02-24 11:20:43 +02004423 kvm_set_cr8(vcpu, sregs->cr8);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004424
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004425 mmu_reset_needed |= vcpu->arch.shadow_efer != sregs->efer;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004426 kvm_x86_ops->set_efer(vcpu, sregs->efer);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004427 kvm_set_apic_base(vcpu, sregs->apic_base);
4428
4429 kvm_x86_ops->decache_cr4_guest_bits(vcpu);
4430
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004431 mmu_reset_needed |= vcpu->arch.cr0 != sregs->cr0;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004432 kvm_x86_ops->set_cr0(vcpu, sregs->cr0);
Paul Knowlesd7306162008-02-06 11:02:35 +00004433 vcpu->arch.cr0 = sregs->cr0;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004434
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004435 mmu_reset_needed |= vcpu->arch.cr4 != sregs->cr4;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004436 kvm_x86_ops->set_cr4(vcpu, sregs->cr4);
4437 if (!is_long_mode(vcpu) && is_pae(vcpu))
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004438 load_pdptrs(vcpu, vcpu->arch.cr3);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004439
4440 if (mmu_reset_needed)
4441 kvm_mmu_reset_context(vcpu);
4442
Gleb Natapov923c61b2009-05-11 13:35:48 +03004443 max_bits = (sizeof sregs->interrupt_bitmap) << 3;
4444 pending_vec = find_first_bit(
4445 (const unsigned long *)sregs->interrupt_bitmap, max_bits);
4446 if (pending_vec < max_bits) {
Gleb Natapov66fd3f72009-05-11 13:35:50 +03004447 kvm_queue_interrupt(vcpu, pending_vec, false);
Gleb Natapov923c61b2009-05-11 13:35:48 +03004448 pr_debug("Set back pending irq %d\n", pending_vec);
4449 if (irqchip_in_kernel(vcpu->kvm))
4450 kvm_pic_clear_isr_ack(vcpu->kvm);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004451 }
4452
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004453 kvm_set_segment(vcpu, &sregs->cs, VCPU_SREG_CS);
4454 kvm_set_segment(vcpu, &sregs->ds, VCPU_SREG_DS);
4455 kvm_set_segment(vcpu, &sregs->es, VCPU_SREG_ES);
4456 kvm_set_segment(vcpu, &sregs->fs, VCPU_SREG_FS);
4457 kvm_set_segment(vcpu, &sregs->gs, VCPU_SREG_GS);
4458 kvm_set_segment(vcpu, &sregs->ss, VCPU_SREG_SS);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004459
Guillaume Thouvenin3e6e0aa2008-05-27 10:18:46 +02004460 kvm_set_segment(vcpu, &sregs->tr, VCPU_SREG_TR);
4461 kvm_set_segment(vcpu, &sregs->ldt, VCPU_SREG_LDTR);
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004462
Mikhail Ershov5f0269f2009-08-03 14:58:25 +03004463 update_cr8_intercept(vcpu);
4464
Marcelo Tosatti9c3e4aa2008-09-10 16:40:55 -03004465 /* Older userspace won't unhalt the vcpu on reset. */
Gleb Natapovc5af89b2009-06-09 15:56:26 +03004466 if (kvm_vcpu_is_bsp(vcpu) && kvm_rip_read(vcpu) == 0xfff0 &&
Marcelo Tosatti9c3e4aa2008-09-10 16:40:55 -03004467 sregs->cs.selector == 0xf000 && sregs->cs.base == 0xffff0000 &&
4468 !(vcpu->arch.cr0 & X86_CR0_PE))
4469 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
4470
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004471 vcpu_put(vcpu);
4472
4473 return 0;
4474}
4475
Jan Kiszkad0bfb942008-12-15 13:52:10 +01004476int kvm_arch_vcpu_ioctl_set_guest_debug(struct kvm_vcpu *vcpu,
4477 struct kvm_guest_debug *dbg)
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004478{
Jan Kiszkaae675ef2008-12-15 13:52:10 +01004479 int i, r;
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004480
4481 vcpu_load(vcpu);
4482
Jan Kiszkaae675ef2008-12-15 13:52:10 +01004483 if ((dbg->control & (KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW_BP)) ==
4484 (KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW_BP)) {
4485 for (i = 0; i < KVM_NR_DB_REGS; ++i)
4486 vcpu->arch.eff_db[i] = dbg->arch.debugreg[i];
4487 vcpu->arch.switch_db_regs =
4488 (dbg->arch.debugreg[7] & DR7_BP_EN_MASK);
4489 } else {
4490 for (i = 0; i < KVM_NR_DB_REGS; i++)
4491 vcpu->arch.eff_db[i] = vcpu->arch.db[i];
4492 vcpu->arch.switch_db_regs = (vcpu->arch.dr7 & DR7_BP_EN_MASK);
4493 }
4494
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004495 r = kvm_x86_ops->set_guest_debug(vcpu, dbg);
4496
Jan Kiszkad0bfb942008-12-15 13:52:10 +01004497 if (dbg->control & KVM_GUESTDBG_INJECT_DB)
4498 kvm_queue_exception(vcpu, DB_VECTOR);
4499 else if (dbg->control & KVM_GUESTDBG_INJECT_BP)
4500 kvm_queue_exception(vcpu, BP_VECTOR);
4501
Hollis Blanchardb6c7a5d2007-11-01 14:16:10 -05004502 vcpu_put(vcpu);
4503
4504 return r;
4505}
4506
4507/*
Hollis Blanchardd0752062007-10-31 17:24:25 -05004508 * fxsave fpu state. Taken from x86_64/processor.h. To be killed when
4509 * we have asm/x86/processor.h
4510 */
4511struct fxsave {
4512 u16 cwd;
4513 u16 swd;
4514 u16 twd;
4515 u16 fop;
4516 u64 rip;
4517 u64 rdp;
4518 u32 mxcsr;
4519 u32 mxcsr_mask;
4520 u32 st_space[32]; /* 8*16 bytes for each FP-reg = 128 bytes */
4521#ifdef CONFIG_X86_64
4522 u32 xmm_space[64]; /* 16*16 bytes for each XMM-reg = 256 bytes */
4523#else
4524 u32 xmm_space[32]; /* 8*16 bytes for each XMM-reg = 128 bytes */
4525#endif
4526};
4527
Zhang Xiantao8b006792007-11-16 13:05:55 +08004528/*
4529 * Translate a guest virtual address to a guest physical address.
4530 */
4531int kvm_arch_vcpu_ioctl_translate(struct kvm_vcpu *vcpu,
4532 struct kvm_translation *tr)
4533{
4534 unsigned long vaddr = tr->linear_address;
4535 gpa_t gpa;
4536
4537 vcpu_load(vcpu);
Izik Eidus72dc67a2008-02-10 18:04:15 +02004538 down_read(&vcpu->kvm->slots_lock);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004539 gpa = vcpu->arch.mmu.gva_to_gpa(vcpu, vaddr);
Izik Eidus72dc67a2008-02-10 18:04:15 +02004540 up_read(&vcpu->kvm->slots_lock);
Zhang Xiantao8b006792007-11-16 13:05:55 +08004541 tr->physical_address = gpa;
4542 tr->valid = gpa != UNMAPPED_GVA;
4543 tr->writeable = 1;
4544 tr->usermode = 0;
Zhang Xiantao8b006792007-11-16 13:05:55 +08004545 vcpu_put(vcpu);
4546
4547 return 0;
4548}
4549
Hollis Blanchardd0752062007-10-31 17:24:25 -05004550int kvm_arch_vcpu_ioctl_get_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
4551{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004552 struct fxsave *fxsave = (struct fxsave *)&vcpu->arch.guest_fx_image;
Hollis Blanchardd0752062007-10-31 17:24:25 -05004553
4554 vcpu_load(vcpu);
4555
4556 memcpy(fpu->fpr, fxsave->st_space, 128);
4557 fpu->fcw = fxsave->cwd;
4558 fpu->fsw = fxsave->swd;
4559 fpu->ftwx = fxsave->twd;
4560 fpu->last_opcode = fxsave->fop;
4561 fpu->last_ip = fxsave->rip;
4562 fpu->last_dp = fxsave->rdp;
4563 memcpy(fpu->xmm, fxsave->xmm_space, sizeof fxsave->xmm_space);
4564
4565 vcpu_put(vcpu);
4566
4567 return 0;
4568}
4569
4570int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu)
4571{
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004572 struct fxsave *fxsave = (struct fxsave *)&vcpu->arch.guest_fx_image;
Hollis Blanchardd0752062007-10-31 17:24:25 -05004573
4574 vcpu_load(vcpu);
4575
4576 memcpy(fxsave->st_space, fpu->fpr, 128);
4577 fxsave->cwd = fpu->fcw;
4578 fxsave->swd = fpu->fsw;
4579 fxsave->twd = fpu->ftwx;
4580 fxsave->fop = fpu->last_opcode;
4581 fxsave->rip = fpu->last_ip;
4582 fxsave->rdp = fpu->last_dp;
4583 memcpy(fxsave->xmm_space, fpu->xmm, sizeof fxsave->xmm_space);
4584
4585 vcpu_put(vcpu);
4586
4587 return 0;
4588}
4589
4590void fx_init(struct kvm_vcpu *vcpu)
4591{
4592 unsigned after_mxcsr_mask;
4593
Andrea Arcangelibc1a34f2008-05-01 18:43:33 +02004594 /*
4595 * Touch the fpu the first time in non atomic context as if
4596 * this is the first fpu instruction the exception handler
4597 * will fire before the instruction returns and it'll have to
4598 * allocate ram with GFP_KERNEL.
4599 */
4600 if (!used_math())
Avi Kivityd6e88ae2008-07-10 16:53:33 +03004601 kvm_fx_save(&vcpu->arch.host_fx_image);
Andrea Arcangelibc1a34f2008-05-01 18:43:33 +02004602
Hollis Blanchardd0752062007-10-31 17:24:25 -05004603 /* Initialize guest FPU by resetting ours and saving into guest's */
4604 preempt_disable();
Avi Kivityd6e88ae2008-07-10 16:53:33 +03004605 kvm_fx_save(&vcpu->arch.host_fx_image);
4606 kvm_fx_finit();
4607 kvm_fx_save(&vcpu->arch.guest_fx_image);
4608 kvm_fx_restore(&vcpu->arch.host_fx_image);
Hollis Blanchardd0752062007-10-31 17:24:25 -05004609 preempt_enable();
4610
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004611 vcpu->arch.cr0 |= X86_CR0_ET;
Hollis Blanchardd0752062007-10-31 17:24:25 -05004612 after_mxcsr_mask = offsetof(struct i387_fxsave_struct, st_space);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004613 vcpu->arch.guest_fx_image.mxcsr = 0x1f80;
4614 memset((void *)&vcpu->arch.guest_fx_image + after_mxcsr_mask,
Hollis Blanchardd0752062007-10-31 17:24:25 -05004615 0, sizeof(struct i387_fxsave_struct) - after_mxcsr_mask);
4616}
4617EXPORT_SYMBOL_GPL(fx_init);
4618
4619void kvm_load_guest_fpu(struct kvm_vcpu *vcpu)
4620{
4621 if (!vcpu->fpu_active || vcpu->guest_fpu_loaded)
4622 return;
4623
4624 vcpu->guest_fpu_loaded = 1;
Avi Kivityd6e88ae2008-07-10 16:53:33 +03004625 kvm_fx_save(&vcpu->arch.host_fx_image);
4626 kvm_fx_restore(&vcpu->arch.guest_fx_image);
Hollis Blanchardd0752062007-10-31 17:24:25 -05004627}
4628EXPORT_SYMBOL_GPL(kvm_load_guest_fpu);
4629
4630void kvm_put_guest_fpu(struct kvm_vcpu *vcpu)
4631{
4632 if (!vcpu->guest_fpu_loaded)
4633 return;
4634
4635 vcpu->guest_fpu_loaded = 0;
Avi Kivityd6e88ae2008-07-10 16:53:33 +03004636 kvm_fx_save(&vcpu->arch.guest_fx_image);
4637 kvm_fx_restore(&vcpu->arch.host_fx_image);
Avi Kivityf096ed82007-11-18 13:54:33 +02004638 ++vcpu->stat.fpu_reload;
Hollis Blanchardd0752062007-10-31 17:24:25 -05004639}
4640EXPORT_SYMBOL_GPL(kvm_put_guest_fpu);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004641
4642void kvm_arch_vcpu_free(struct kvm_vcpu *vcpu)
4643{
Joerg Roedel7f1ea202009-02-25 16:08:31 +01004644 if (vcpu->arch.time_page) {
4645 kvm_release_page_dirty(vcpu->arch.time_page);
4646 vcpu->arch.time_page = NULL;
4647 }
4648
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004649 kvm_x86_ops->vcpu_free(vcpu);
4650}
4651
4652struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm,
4653 unsigned int id)
4654{
Avi Kivity26e52152007-11-20 15:30:24 +02004655 return kvm_x86_ops->vcpu_create(kvm, id);
4656}
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004657
Avi Kivity26e52152007-11-20 15:30:24 +02004658int kvm_arch_vcpu_setup(struct kvm_vcpu *vcpu)
4659{
4660 int r;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004661
4662 /* We do fxsave: this must be aligned. */
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004663 BUG_ON((unsigned long)&vcpu->arch.host_fx_image & 0xF);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004664
Sheng Yang0bed3b52008-10-09 16:01:54 +08004665 vcpu->arch.mtrr_state.have_fixed = 1;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004666 vcpu_load(vcpu);
4667 r = kvm_arch_vcpu_reset(vcpu);
4668 if (r == 0)
4669 r = kvm_mmu_setup(vcpu);
4670 vcpu_put(vcpu);
4671 if (r < 0)
4672 goto free_vcpu;
4673
Avi Kivity26e52152007-11-20 15:30:24 +02004674 return 0;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004675free_vcpu:
4676 kvm_x86_ops->vcpu_free(vcpu);
Avi Kivity26e52152007-11-20 15:30:24 +02004677 return r;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004678}
4679
Hollis Blanchardd40ccc62007-11-19 14:04:43 -06004680void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004681{
4682 vcpu_load(vcpu);
4683 kvm_mmu_unload(vcpu);
4684 vcpu_put(vcpu);
4685
4686 kvm_x86_ops->vcpu_free(vcpu);
4687}
4688
4689int kvm_arch_vcpu_reset(struct kvm_vcpu *vcpu)
4690{
Jan Kiszka448fa4a2008-09-26 09:30:48 +02004691 vcpu->arch.nmi_pending = false;
4692 vcpu->arch.nmi_injected = false;
4693
Jan Kiszka42dbaa52008-12-15 13:52:10 +01004694 vcpu->arch.switch_db_regs = 0;
4695 memset(vcpu->arch.db, 0, sizeof(vcpu->arch.db));
4696 vcpu->arch.dr6 = DR6_FIXED_1;
4697 vcpu->arch.dr7 = DR7_FIXED_1;
4698
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004699 return kvm_x86_ops->vcpu_reset(vcpu);
4700}
4701
4702void kvm_arch_hardware_enable(void *garbage)
4703{
4704 kvm_x86_ops->hardware_enable(garbage);
4705}
4706
4707void kvm_arch_hardware_disable(void *garbage)
4708{
4709 kvm_x86_ops->hardware_disable(garbage);
4710}
4711
4712int kvm_arch_hardware_setup(void)
4713{
4714 return kvm_x86_ops->hardware_setup();
4715}
4716
4717void kvm_arch_hardware_unsetup(void)
4718{
4719 kvm_x86_ops->hardware_unsetup();
4720}
4721
4722void kvm_arch_check_processor_compat(void *rtn)
4723{
4724 kvm_x86_ops->check_processor_compatibility(rtn);
4725}
4726
4727int kvm_arch_vcpu_init(struct kvm_vcpu *vcpu)
4728{
4729 struct page *page;
4730 struct kvm *kvm;
4731 int r;
4732
4733 BUG_ON(vcpu->kvm == NULL);
4734 kvm = vcpu->kvm;
4735
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004736 vcpu->arch.mmu.root_hpa = INVALID_PAGE;
Gleb Natapovc5af89b2009-06-09 15:56:26 +03004737 if (!irqchip_in_kernel(kvm) || kvm_vcpu_is_bsp(vcpu))
Avi Kivitya4535292008-04-13 17:54:35 +03004738 vcpu->arch.mp_state = KVM_MP_STATE_RUNNABLE;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004739 else
Avi Kivitya4535292008-04-13 17:54:35 +03004740 vcpu->arch.mp_state = KVM_MP_STATE_UNINITIALIZED;
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004741
4742 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
4743 if (!page) {
4744 r = -ENOMEM;
4745 goto fail;
4746 }
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004747 vcpu->arch.pio_data = page_address(page);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004748
4749 r = kvm_mmu_create(vcpu);
4750 if (r < 0)
4751 goto fail_free_pio_data;
4752
4753 if (irqchip_in_kernel(kvm)) {
4754 r = kvm_create_lapic(vcpu);
4755 if (r < 0)
4756 goto fail_mmu_destroy;
4757 }
4758
Huang Ying890ca9a2009-05-11 16:48:15 +08004759 vcpu->arch.mce_banks = kzalloc(KVM_MAX_MCE_BANKS * sizeof(u64) * 4,
4760 GFP_KERNEL);
4761 if (!vcpu->arch.mce_banks) {
4762 r = -ENOMEM;
4763 goto fail_mmu_destroy;
4764 }
4765 vcpu->arch.mcg_cap = KVM_MAX_MCE_BANKS;
4766
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004767 return 0;
4768
4769fail_mmu_destroy:
4770 kvm_mmu_destroy(vcpu);
4771fail_free_pio_data:
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004772 free_page((unsigned long)vcpu->arch.pio_data);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004773fail:
4774 return r;
4775}
4776
4777void kvm_arch_vcpu_uninit(struct kvm_vcpu *vcpu)
4778{
4779 kvm_free_lapic(vcpu);
Marcelo Tosatti3200f402008-03-29 20:17:59 -03004780 down_read(&vcpu->kvm->slots_lock);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004781 kvm_mmu_destroy(vcpu);
Marcelo Tosatti3200f402008-03-29 20:17:59 -03004782 up_read(&vcpu->kvm->slots_lock);
Zhang Xiantaoad312c72007-12-13 23:50:52 +08004783 free_page((unsigned long)vcpu->arch.pio_data);
Zhang Xiantaoe9b11c12007-11-14 20:38:21 +08004784}
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004785
4786struct kvm *kvm_arch_create_vm(void)
4787{
4788 struct kvm *kvm = kzalloc(sizeof(struct kvm), GFP_KERNEL);
4789
4790 if (!kvm)
4791 return ERR_PTR(-ENOMEM);
4792
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08004793 INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
Ben-Ami Yassour4d5c5d02008-07-28 19:26:26 +03004794 INIT_LIST_HEAD(&kvm->arch.assigned_dev_head);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004795
Sheng Yang5550af42008-10-15 20:15:06 +08004796 /* Reserve bit 0 of irq_sources_bitmap for userspace irq source */
4797 set_bit(KVM_USERSPACE_IRQ_SOURCE_ID, &kvm->arch.irq_sources_bitmap);
4798
Marcelo Tosatti53f658b32008-12-11 20:45:05 +01004799 rdtscll(kvm->arch.vm_init_tsc);
4800
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004801 return kvm;
4802}
4803
4804static void kvm_unload_vcpu_mmu(struct kvm_vcpu *vcpu)
4805{
4806 vcpu_load(vcpu);
4807 kvm_mmu_unload(vcpu);
4808 vcpu_put(vcpu);
4809}
4810
4811static void kvm_free_vcpus(struct kvm *kvm)
4812{
4813 unsigned int i;
Gleb Natapov988a2ca2009-06-09 15:56:29 +03004814 struct kvm_vcpu *vcpu;
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004815
4816 /*
4817 * Unpin any mmu pages first.
4818 */
Gleb Natapov988a2ca2009-06-09 15:56:29 +03004819 kvm_for_each_vcpu(i, vcpu, kvm)
4820 kvm_unload_vcpu_mmu(vcpu);
4821 kvm_for_each_vcpu(i, vcpu, kvm)
4822 kvm_arch_vcpu_free(vcpu);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004823
Gleb Natapov988a2ca2009-06-09 15:56:29 +03004824 mutex_lock(&kvm->lock);
4825 for (i = 0; i < atomic_read(&kvm->online_vcpus); i++)
4826 kvm->vcpus[i] = NULL;
4827
4828 atomic_set(&kvm->online_vcpus, 0);
4829 mutex_unlock(&kvm->lock);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004830}
4831
Sheng Yangad8ba2c2009-01-06 10:03:02 +08004832void kvm_arch_sync_events(struct kvm *kvm)
4833{
Sheng Yangba4cef32009-01-06 10:03:03 +08004834 kvm_free_all_assigned_devices(kvm);
Sheng Yangad8ba2c2009-01-06 10:03:02 +08004835}
4836
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004837void kvm_arch_destroy_vm(struct kvm *kvm)
4838{
Sheng Yang6eb55812008-10-31 12:37:41 +08004839 kvm_iommu_unmap_guest(kvm);
Sheng Yang78376992008-01-28 05:10:22 +08004840 kvm_free_pit(kvm);
Zhang Xiantaod7deeeb02007-12-14 10:17:34 +08004841 kfree(kvm->arch.vpic);
4842 kfree(kvm->arch.vioapic);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004843 kvm_free_vcpus(kvm);
4844 kvm_free_physmem(kvm);
Avi Kivity3d458302008-03-25 11:26:13 +02004845 if (kvm->arch.apic_access_page)
4846 put_page(kvm->arch.apic_access_page);
Sheng Yangb7ebfb02008-04-25 21:44:52 +08004847 if (kvm->arch.ept_identity_pagetable)
4848 put_page(kvm->arch.ept_identity_pagetable);
Zhang Xiantaod19a9cd2007-11-18 18:43:45 +08004849 kfree(kvm);
4850}
Zhang Xiantao0de10342007-11-20 16:25:04 +08004851
4852int kvm_arch_set_memory_region(struct kvm *kvm,
4853 struct kvm_userspace_memory_region *mem,
4854 struct kvm_memory_slot old,
4855 int user_alloc)
4856{
4857 int npages = mem->memory_size >> PAGE_SHIFT;
4858 struct kvm_memory_slot *memslot = &kvm->memslots[mem->slot];
4859
4860 /*To keep backward compatibility with older userspace,
4861 *x86 needs to hanlde !user_alloc case.
4862 */
4863 if (!user_alloc) {
4864 if (npages && !old.rmap) {
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02004865 unsigned long userspace_addr;
4866
Izik Eidus72dc67a2008-02-10 18:04:15 +02004867 down_write(&current->mm->mmap_sem);
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02004868 userspace_addr = do_mmap(NULL, 0,
4869 npages * PAGE_SIZE,
4870 PROT_READ | PROT_WRITE,
Avi Kivityacee3c02008-08-26 17:22:47 +03004871 MAP_PRIVATE | MAP_ANONYMOUS,
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02004872 0);
Izik Eidus72dc67a2008-02-10 18:04:15 +02004873 up_write(&current->mm->mmap_sem);
Zhang Xiantao0de10342007-11-20 16:25:04 +08004874
Andrea Arcangeli604b38a2008-07-25 16:32:03 +02004875 if (IS_ERR((void *)userspace_addr))
4876 return PTR_ERR((void *)userspace_addr);
4877
4878 /* set userspace_addr atomically for kvm_hva_to_rmapp */
4879 spin_lock(&kvm->mmu_lock);
4880 memslot->userspace_addr = userspace_addr;
4881 spin_unlock(&kvm->mmu_lock);
Zhang Xiantao0de10342007-11-20 16:25:04 +08004882 } else {
4883 if (!old.user_alloc && old.rmap) {
4884 int ret;
4885
Izik Eidus72dc67a2008-02-10 18:04:15 +02004886 down_write(&current->mm->mmap_sem);
Zhang Xiantao0de10342007-11-20 16:25:04 +08004887 ret = do_munmap(current->mm, old.userspace_addr,
4888 old.npages * PAGE_SIZE);
Izik Eidus72dc67a2008-02-10 18:04:15 +02004889 up_write(&current->mm->mmap_sem);
Zhang Xiantao0de10342007-11-20 16:25:04 +08004890 if (ret < 0)
4891 printk(KERN_WARNING
4892 "kvm_vm_ioctl_set_memory_region: "
4893 "failed to munmap memory\n");
4894 }
4895 }
4896 }
4897
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03004898 spin_lock(&kvm->mmu_lock);
Zhang Xiantaof05e70a2007-12-14 10:01:48 +08004899 if (!kvm->arch.n_requested_mmu_pages) {
Zhang Xiantao0de10342007-11-20 16:25:04 +08004900 unsigned int nr_mmu_pages = kvm_mmu_calculate_mmu_pages(kvm);
4901 kvm_mmu_change_mmu_pages(kvm, nr_mmu_pages);
4902 }
4903
4904 kvm_mmu_slot_remove_write_access(kvm, mem->slot);
Marcelo Tosatti7c8a83b2009-05-12 18:55:43 -03004905 spin_unlock(&kvm->mmu_lock);
Zhang Xiantao0de10342007-11-20 16:25:04 +08004906
4907 return 0;
4908}
Zhang Xiantao1d737c82007-12-14 09:35:10 +08004909
Marcelo Tosatti34d4cb82008-07-10 20:49:31 -03004910void kvm_arch_flush_shadow(struct kvm *kvm)
4911{
4912 kvm_mmu_zap_all(kvm);
Marcelo Tosatti8986ecc2009-05-12 18:55:45 -03004913 kvm_reload_remote_mmus(kvm);
Marcelo Tosatti34d4cb82008-07-10 20:49:31 -03004914}
4915
Zhang Xiantao1d737c82007-12-14 09:35:10 +08004916int kvm_arch_vcpu_runnable(struct kvm_vcpu *vcpu)
4917{
Avi Kivitya4535292008-04-13 17:54:35 +03004918 return vcpu->arch.mp_state == KVM_MP_STATE_RUNNABLE
Gleb Natapova1b37102009-07-09 15:33:52 +03004919 || vcpu->arch.mp_state == KVM_MP_STATE_SIPI_RECEIVED
4920 || vcpu->arch.nmi_pending ||
4921 (kvm_arch_interrupt_allowed(vcpu) &&
4922 kvm_cpu_has_interrupt(vcpu));
Zhang Xiantao1d737c82007-12-14 09:35:10 +08004923}
Zhang Xiantao57361992007-12-17 14:21:40 +08004924
Zhang Xiantao57361992007-12-17 14:21:40 +08004925void kvm_vcpu_kick(struct kvm_vcpu *vcpu)
4926{
Marcelo Tosatti32f88402009-05-07 17:55:12 -03004927 int me;
4928 int cpu = vcpu->cpu;
Zhang Xiantao57361992007-12-17 14:21:40 +08004929
4930 if (waitqueue_active(&vcpu->wq)) {
4931 wake_up_interruptible(&vcpu->wq);
4932 ++vcpu->stat.halt_wakeup;
4933 }
Marcelo Tosatti32f88402009-05-07 17:55:12 -03004934
4935 me = get_cpu();
4936 if (cpu != me && (unsigned)cpu < nr_cpu_ids && cpu_online(cpu))
4937 if (!test_and_set_bit(KVM_REQ_KICK, &vcpu->requests))
4938 smp_send_reschedule(cpu);
Marcelo Tosattie9571ed2008-04-11 15:01:22 -03004939 put_cpu();
Zhang Xiantao57361992007-12-17 14:21:40 +08004940}
Gleb Natapov78646122009-03-23 12:12:11 +02004941
4942int kvm_arch_interrupt_allowed(struct kvm_vcpu *vcpu)
4943{
4944 return kvm_x86_ops->interrupt_allowed(vcpu);
4945}
Marcelo Tosatti229456f2009-06-17 09:22:14 -03004946
4947EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_exit);
4948EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_inj_virq);
4949EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_page_fault);
4950EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_msr);
4951EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_cr);