drivers/kvm/kvm.h

   1 #ifndef __KVM_H
   2 #define __KVM_H
   3
   4 /*
   5  * This work is licensed under the terms of the GNU GPL, version 2.  See
   6  * the COPYING file in the top-level directory.
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/list.h>
  11 #include <linux/mutex.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/mm.h>
  14
  15 #include "vmx.h"
  16 #include <linux/kvm.h>
  17
  18 #define CR0_PE_MASK (1ULL << 0)
  19 #define CR0_TS_MASK (1ULL << 3)
  20 #define CR0_NE_MASK (1ULL << 5)
  21 #define CR0_WP_MASK (1ULL << 16)
  22 #define CR0_NW_MASK (1ULL << 29)
  23 #define CR0_CD_MASK (1ULL << 30)
  24 #define CR0_PG_MASK (1ULL << 31)
  25
  26 #define CR3_WPT_MASK (1ULL << 3)
  27 #define CR3_PCD_MASK (1ULL << 4)
  28
  29 #define CR3_RESEVED_BITS 0x07ULL
  30 #define CR3_L_MODE_RESEVED_BITS (~((1ULL << 40) - 1) | 0x0fe7ULL)
  31 #define CR3_FLAGS_MASK ((1ULL << 5) - 1)
  32
  33 #define CR4_VME_MASK (1ULL << 0)
  34 #define CR4_PSE_MASK (1ULL << 4)
  35 #define CR4_PAE_MASK (1ULL << 5)
  36 #define CR4_PGE_MASK (1ULL << 7)
  37 #define CR4_VMXE_MASK (1ULL << 13)
  38
  39 #define KVM_GUEST_CR0_MASK \
  40         (CR0_PG_MASK | CR0_PE_MASK | CR0_WP_MASK | CR0_NE_MASK \
  41          | CR0_NW_MASK | CR0_CD_MASK)
  42 #define KVM_VM_CR0_ALWAYS_ON \
  43         (CR0_PG_MASK | CR0_PE_MASK | CR0_WP_MASK | CR0_NE_MASK)
  44 #define KVM_GUEST_CR4_MASK \
  45         (CR4_PSE_MASK | CR4_PAE_MASK | CR4_PGE_MASK | CR4_VMXE_MASK | CR4_VME_MASK)
  46 #define KVM_PMODE_VM_CR4_ALWAYS_ON (CR4_VMXE_MASK | CR4_PAE_MASK)
  47 #define KVM_RMODE_VM_CR4_ALWAYS_ON (CR4_VMXE_MASK | CR4_PAE_MASK | CR4_VME_MASK)
  48
  49 #define INVALID_PAGE (~(hpa_t)0)
  50 #define UNMAPPED_GVA (~(gpa_t)0)
  51
  52 #define KVM_MAX_VCPUS 1
  53 #define KVM_MEMORY_SLOTS 4
  54 #define KVM_NUM_MMU_PAGES 256
  55
  56 #define FX_IMAGE_SIZE 512
  57 #define FX_IMAGE_ALIGN 16
  58 #define FX_BUF_SIZE (2 * FX_IMAGE_SIZE + FX_IMAGE_ALIGN)
  59
  60 #define DE_VECTOR 0
  61 #define DF_VECTOR 8
  62 #define TS_VECTOR 10
  63 #define NP_VECTOR 11
  64 #define SS_VECTOR 12
  65 #define GP_VECTOR 13
  66 #define PF_VECTOR 14
  67
  68 #define SELECTOR_TI_MASK (1 << 2)
  69 #define SELECTOR_RPL_MASK 0x03
  70
  71 #define IOPL_SHIFT 12
  72
  73 /*
  74  * Address types:
  75  *
  76  *  gva - guest virtual address
  77  *  gpa - guest physical address
  78  *  gfn - guest frame number
  79  *  hva - host virtual address
  80  *  hpa - host physical address
  81  *  hfn - host frame number
  82  */
  83
  84 typedef unsigned long  gva_t;
  85 typedef u64            gpa_t;
  86 typedef unsigned long  gfn_t;
  87
  88 typedef unsigned long  hva_t;
  89 typedef u64            hpa_t;
  90 typedef unsigned long  hfn_t;
  91
  92 struct kvm_mmu_page {
  93         struct list_head link;
  94         hpa_t page_hpa;
  95         unsigned long slot_bitmap; /* One bit set per slot which has memory
  96                                     * in this shadow page.
  97                                     */
  98         int global;              /* Set if all ptes in this page are global */
  99         u64 *parent_pte;
 100 };
 101
 102 struct vmcs {
 103         u32 revision_id;
 104         u32 abort;
 105         char data[0];
 106 };
 107
 108 #define vmx_msr_entry kvm_msr_entry
 109
 110 struct kvm_vcpu;
 111
 112 /*
 113  * x86 supports 3 paging modes (4-level 64-bit, 3-level 64-bit, and 2-level
 114  * 32-bit).  The kvm_mmu structure abstracts the details of the current mmu
 115  * mode.
 116  */
 117 struct kvm_mmu {
 118         void (*new_cr3)(struct kvm_vcpu *vcpu);
 119         int (*page_fault)(struct kvm_vcpu *vcpu, gva_t gva, u32 err);
 120         void (*inval_page)(struct kvm_vcpu *vcpu, gva_t gva);
 121         void (*free)(struct kvm_vcpu *vcpu);
 122         gpa_t (*gva_to_gpa)(struct kvm_vcpu *vcpu, gva_t gva);
 123         hpa_t root_hpa;
 124         int root_level;
 125         int shadow_root_level;
 126 };
 127
 128 struct kvm_guest_debug {
 129         int enabled;
 130         unsigned long bp[4];
 131         int singlestep;
 132 };
 133
 134 enum {
 135         VCPU_REGS_RAX = 0,
 136         VCPU_REGS_RCX = 1,
 137         VCPU_REGS_RDX = 2,
 138         VCPU_REGS_RBX = 3,
 139         VCPU_REGS_RSP = 4,
 140         VCPU_REGS_RBP = 5,
 141         VCPU_REGS_RSI = 6,
 142         VCPU_REGS_RDI = 7,
 143 #ifdef CONFIG_X86_64
 144         VCPU_REGS_R8 = 8,
 145         VCPU_REGS_R9 = 9,
 146         VCPU_REGS_R10 = 10,
 147         VCPU_REGS_R11 = 11,
 148         VCPU_REGS_R12 = 12,
 149         VCPU_REGS_R13 = 13,
 150         VCPU_REGS_R14 = 14,
 151         VCPU_REGS_R15 = 15,
 152 #endif
 153         NR_VCPU_REGS
 154 };
 155
 156 enum {
 157         VCPU_SREG_CS,
 158         VCPU_SREG_DS,
 159         VCPU_SREG_ES,
 160         VCPU_SREG_FS,
 161         VCPU_SREG_GS,
 162         VCPU_SREG_SS,
 163         VCPU_SREG_TR,
 164         VCPU_SREG_LDTR,
 165 };
 166
 167 struct kvm_vcpu {
 168         struct kvm *kvm;
 169         union {
 170                 struct vmcs *vmcs;
 171                 struct vcpu_svm *svm;
 172         };
 173         struct mutex mutex;
 174         int   cpu;
 175         int   launched;
 176         unsigned long irq_summary; /* bit vector: 1 per word in irq_pending */
 177 #define NR_IRQ_WORDS KVM_IRQ_BITMAP_SIZE(unsigned long)
 178         unsigned long irq_pending[NR_IRQ_WORDS];
 179         unsigned long regs[NR_VCPU_REGS]; /* for rsp: vcpu_load_rsp_rip() */
 180         unsigned long rip;      /* needs vcpu_load_rsp_rip() */
 181
 182         unsigned long cr0;
 183         unsigned long cr2;
 184         unsigned long cr3;
 185         unsigned long cr4;
 186         unsigned long cr8;
 187         u64 shadow_efer;
 188         u64 apic_base;
 189         int nmsrs;
 190         struct vmx_msr_entry *guest_msrs;
 191         struct vmx_msr_entry *host_msrs;
 192
 193         struct list_head free_pages;
 194         struct kvm_mmu_page page_header_buf[KVM_NUM_MMU_PAGES];
 195         struct kvm_mmu mmu;
 196
 197         struct kvm_guest_debug guest_debug;
 198
 199         char fx_buf[FX_BUF_SIZE];
 200         char *host_fx_image;
 201         char *guest_fx_image;
 202
 203         int mmio_needed;
 204         int mmio_read_completed;
 205         int mmio_is_write;
 206         int mmio_size;
 207         unsigned char mmio_data[8];
 208         gpa_t mmio_phys_addr;
 209
 210         struct {
 211                 int active;
 212                 u8 save_iopl;
 213                 struct kvm_save_segment {
 214                         u16 selector;
 215                         unsigned long base;
 216                         u32 limit;
 217                         u32 ar;
 218                 } tr, es, ds, fs, gs;
 219         } rmode;
 220 };
 221
 222 struct kvm_memory_slot {
 223         gfn_t base_gfn;
 224         unsigned long npages;
 225         unsigned long flags;
 226         struct page **phys_mem;
 227         unsigned long *dirty_bitmap;
 228 };
 229
 230 struct kvm {
 231         spinlock_t lock; /* protects everything except vcpus */
 232         int nmemslots;
 233         struct kvm_memory_slot memslots[KVM_MEMORY_SLOTS];
 234         struct list_head active_mmu_pages;
 235         struct kvm_vcpu vcpus[KVM_MAX_VCPUS];
 236         int memory_config_version;
 237         int busy;
 238 };
 239
 240 struct kvm_stat {
 241         u32 pf_fixed;
 242         u32 pf_guest;
 243         u32 tlb_flush;
 244         u32 invlpg;
 245
 246         u32 exits;
 247         u32 io_exits;
 248         u32 mmio_exits;
 249         u32 signal_exits;
 250         u32 irq_exits;
 251 };
 252
 253 struct descriptor_table {
 254         u16 limit;
 255         unsigned long base;
 256 } __attribute__((packed));
 257
 258 struct kvm_arch_ops {
 259         int (*cpu_has_kvm_support)(void);          /* __init */
 260         int (*disabled_by_bios)(void);             /* __init */
 261         void (*hardware_enable)(void *dummy);      /* __init */
 262         void (*hardware_disable)(void *dummy);
 263         int (*hardware_setup)(void);               /* __init */
 264         void (*hardware_unsetup)(void);            /* __exit */
 265
 266         int (*vcpu_create)(struct kvm_vcpu *vcpu);
 267         void (*vcpu_free)(struct kvm_vcpu *vcpu);
 268
 269         struct kvm_vcpu *(*vcpu_load)(struct kvm_vcpu *vcpu);
 270         void (*vcpu_put)(struct kvm_vcpu *vcpu);
 271
 272         int (*set_guest_debug)(struct kvm_vcpu *vcpu,
 273                                struct kvm_debug_guest *dbg);
 274         int (*get_msr)(struct kvm_vcpu *vcpu, u32 msr_index, u64 *pdata);
 275         int (*set_msr)(struct kvm_vcpu *vcpu, u32 msr_index, u64 data);
 276         u64 (*get_segment_base)(struct kvm_vcpu *vcpu, int seg);
 277         void (*get_segment)(struct kvm_vcpu *vcpu,
 278                             struct kvm_segment *var, int seg);
 279         void (*set_segment)(struct kvm_vcpu *vcpu,
 280                             struct kvm_segment *var, int seg);
 281         int (*is_long_mode)(struct kvm_vcpu *vcpu);
 282         void (*get_cs_db_l_bits)(struct kvm_vcpu *vcpu, int *db, int *l);
 283         void (*set_cr0)(struct kvm_vcpu *vcpu, unsigned long cr0);
 284         void (*set_cr0_no_modeswitch)(struct kvm_vcpu *vcpu,
 285                                       unsigned long cr0);
 286         void (*set_cr3)(struct kvm_vcpu *vcpu, unsigned long cr3);
 287         void (*set_cr4)(struct kvm_vcpu *vcpu, unsigned long cr4);
 288         void (*set_efer)(struct kvm_vcpu *vcpu, u64 efer);
 289         void (*get_idt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 290         void (*set_idt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 291         void (*get_gdt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 292         void (*set_gdt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 293         unsigned long (*get_dr)(struct kvm_vcpu *vcpu, int dr);
 294         void (*set_dr)(struct kvm_vcpu *vcpu, int dr, unsigned long value,
 295                        int *exception);
 296         void (*cache_regs)(struct kvm_vcpu *vcpu);
 297         void (*decache_regs)(struct kvm_vcpu *vcpu);
 298         unsigned long (*get_rflags)(struct kvm_vcpu *vcpu);
 299         void (*set_rflags)(struct kvm_vcpu *vcpu, unsigned long rflags);
 300
 301         void (*invlpg)(struct kvm_vcpu *vcpu, gva_t addr);
 302         void (*tlb_flush)(struct kvm_vcpu *vcpu);
 303         void (*inject_page_fault)(struct kvm_vcpu *vcpu,
 304                                   unsigned long addr, u32 err_code);
 305
 306         void (*inject_gp)(struct kvm_vcpu *vcpu, unsigned err_code);
 307
 308         int (*run)(struct kvm_vcpu *vcpu, struct kvm_run *run);
 309         int (*vcpu_setup)(struct kvm_vcpu *vcpu);
 310         void (*skip_emulated_instruction)(struct kvm_vcpu *vcpu);
 311 };
 312
 313 extern struct kvm_stat kvm_stat;
 314 extern struct kvm_arch_ops *kvm_arch_ops;
 315
 316 #define kvm_printf(kvm, fmt ...) printk(KERN_DEBUG fmt)
 317 #define vcpu_printf(vcpu, fmt...) kvm_printf(vcpu->kvm, fmt)
 318
 319 int kvm_init_arch(struct kvm_arch_ops *ops, struct module *module);
 320 void kvm_exit_arch(void);
 321
 322 void kvm_mmu_destroy(struct kvm_vcpu *vcpu);
 323 int kvm_mmu_init(struct kvm_vcpu *vcpu);
 324
 325 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu);
 326 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot);
 327
 328 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa);
 329 #define HPA_MSB ((sizeof(hpa_t) * 8) - 1)
 330 #define HPA_ERR_MASK ((hpa_t)1 << HPA_MSB)
 331 static inline int is_error_hpa(hpa_t hpa) { return hpa >> HPA_MSB; }
 332 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva);
 333
 334 void kvm_emulator_want_group7_invlpg(void);
 335
 336 extern hpa_t bad_page_address;
 337
 338 static inline struct page *gfn_to_page(struct kvm_memory_slot *slot, gfn_t gfn)
 339 {
 340         return slot->phys_mem[gfn - slot->base_gfn];
 341 }
 342
 343 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn);
 344 void mark_page_dirty(struct kvm *kvm, gfn_t gfn);
 345
 346 enum emulation_result {
 347         EMULATE_DONE,       /* no further processing */
 348         EMULATE_DO_MMIO,      /* kvm_run filled with mmio request */
 349         EMULATE_FAIL,         /* can't emulate this instruction */
 350 };
 351
 352 int emulate_instruction(struct kvm_vcpu *vcpu, struct kvm_run *run,
 353                         unsigned long cr2, u16 error_code);
 354 void realmode_lgdt(struct kvm_vcpu *vcpu, u16 size, unsigned long address);
 355 void realmode_lidt(struct kvm_vcpu *vcpu, u16 size, unsigned long address);
 356 void realmode_lmsw(struct kvm_vcpu *vcpu, unsigned long msw,
 357                    unsigned long *rflags);
 358
 359 unsigned long realmode_get_cr(struct kvm_vcpu *vcpu, int cr);
 360 void realmode_set_cr(struct kvm_vcpu *vcpu, int cr, unsigned long value,
 361                      unsigned long *rflags);
 362
 363 struct x86_emulate_ctxt;
 364
 365 int emulate_invlpg(struct kvm_vcpu *vcpu, gva_t address);
 366 int emulate_clts(struct kvm_vcpu *vcpu);
 367 int emulator_get_dr(struct x86_emulate_ctxt* ctxt, int dr,
 368                     unsigned long *dest);
 369 int emulator_set_dr(struct x86_emulate_ctxt *ctxt, int dr,
 370                     unsigned long value);
 371
 372 void set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0);
 373 void set_cr3(struct kvm_vcpu *vcpu, unsigned long cr0);
 374 void set_cr4(struct kvm_vcpu *vcpu, unsigned long cr0);
 375 void set_cr8(struct kvm_vcpu *vcpu, unsigned long cr0);
 376 void lmsw(struct kvm_vcpu *vcpu, unsigned long msw);
 377
 378 #ifdef CONFIG_X86_64
 379 void set_efer(struct kvm_vcpu *vcpu, u64 efer);
 380 #endif
 381
 382 void fx_init(struct kvm_vcpu *vcpu);
 383
 384 void load_msrs(struct vmx_msr_entry *e, int n);
 385 void save_msrs(struct vmx_msr_entry *e, int n);
 386 void kvm_resched(struct kvm_vcpu *vcpu);
 387
 388 int kvm_read_guest(struct kvm_vcpu *vcpu,
 389                gva_t addr,
 390                unsigned long size,
 391                void *dest);
 392
 393 int kvm_write_guest(struct kvm_vcpu *vcpu,
 394                 gva_t addr,
 395                 unsigned long size,
 396                 void *data);
 397
 398 unsigned long segment_base(u16 selector);
 399
 400 static inline struct page *_gfn_to_page(struct kvm *kvm, gfn_t gfn)
 401 {
 402         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 403         return (slot) ? slot->phys_mem[gfn - slot->base_gfn] : NULL;
 404 }
 405
 406 static inline int is_pae(struct kvm_vcpu *vcpu)
 407 {
 408         return vcpu->cr4 & CR4_PAE_MASK;
 409 }
 410
 411 static inline int is_pse(struct kvm_vcpu *vcpu)
 412 {
 413         return vcpu->cr4 & CR4_PSE_MASK;
 414 }
 415
 416 static inline int is_paging(struct kvm_vcpu *vcpu)
 417 {
 418         return vcpu->cr0 & CR0_PG_MASK;
 419 }
 420
 421 static inline int memslot_id(struct kvm *kvm, struct kvm_memory_slot *slot)
 422 {
 423         return slot - kvm->memslots;
 424 }
 425
 426 static inline struct kvm_mmu_page *page_header(hpa_t shadow_page)
 427 {
 428         struct page *page = pfn_to_page(shadow_page >> PAGE_SHIFT);
 429
 430         return (struct kvm_mmu_page *)page->private;
 431 }
 432
 433 static inline u16 read_fs(void)
 434 {
 435         u16 seg;
 436         asm ("mov %%fs, %0" : "=g"(seg));
 437         return seg;
 438 }
 439
 440 static inline u16 read_gs(void)
 441 {
 442         u16 seg;
 443         asm ("mov %%gs, %0" : "=g"(seg));
 444         return seg;
 445 }
 446
 447 static inline u16 read_ldt(void)
 448 {
 449         u16 ldt;
 450         asm ("sldt %0" : "=g"(ldt));
 451         return ldt;
 452 }
 453
 454 static inline void load_fs(u16 sel)
 455 {
 456         asm ("mov %0, %%fs" : : "rm"(sel));
 457 }
 458
 459 static inline void load_gs(u16 sel)
 460 {
 461         asm ("mov %0, %%gs" : : "rm"(sel));
 462 }
 463
 464 #ifndef load_ldt
 465 static inline void load_ldt(u16 sel)
 466 {
 467         asm ("lldt %0" : : "g"(sel));
 468 }
 469 #endif
 470
 471 static inline void get_idt(struct descriptor_table *table)
 472 {
 473         asm ("sidt %0" : "=m"(*table));
 474 }
 475
 476 static inline void get_gdt(struct descriptor_table *table)
 477 {
 478         asm ("sgdt %0" : "=m"(*table));
 479 }
 480
 481 static inline unsigned long read_tr_base(void)
 482 {
 483         u16 tr;
 484         asm ("str %0" : "=g"(tr));
 485         return segment_base(tr);
 486 }
 487
 488 #ifdef CONFIG_X86_64
 489 static inline unsigned long read_msr(unsigned long msr)
 490 {
 491         u64 value;
 492
 493         rdmsrl(msr, value);
 494         return value;
 495 }
 496 #endif
 497
 498 static inline void fx_save(void *image)
 499 {
 500         asm ("fxsave (%0)":: "r" (image));
 501 }
 502
 503 static inline void fx_restore(void *image)
 504 {
 505         asm ("fxrstor (%0)":: "r" (image));
 506 }
 507
 508 static inline void fpu_init(void)
 509 {
 510         asm ("finit");
 511 }
 512
 513 static inline u32 get_rdx_init_val(void)
 514 {
 515         return 0x600; /* P6 family */
 516 }
 517
 518 #define ASM_VMX_VMCLEAR_RAX       ".byte 0x66, 0x0f, 0xc7, 0x30"
 519 #define ASM_VMX_VMLAUNCH          ".byte 0x0f, 0x01, 0xc2"
 520 #define ASM_VMX_VMRESUME          ".byte 0x0f, 0x01, 0xc3"
 521 #define ASM_VMX_VMPTRLD_RAX       ".byte 0x0f, 0xc7, 0x30"
 522 #define ASM_VMX_VMREAD_RDX_RAX    ".byte 0x0f, 0x78, 0xd0"
 523 #define ASM_VMX_VMWRITE_RAX_RDX   ".byte 0x0f, 0x79, 0xd0"
 524 #define ASM_VMX_VMWRITE_RSP_RDX   ".byte 0x0f, 0x79, 0xd4"
 525 #define ASM_VMX_VMXOFF            ".byte 0x0f, 0x01, 0xc4"
 526 #define ASM_VMX_VMXON_RAX         ".byte 0xf3, 0x0f, 0xc7, 0x30"
 527
 528 #define MSR_IA32_TIME_STAMP_COUNTER             0x010
 529
 530 #define TSS_IOPB_BASE_OFFSET 0x66
 531 #define TSS_BASE_SIZE 0x68
 532 #define TSS_IOPB_SIZE (65536 / 8)
 533 #define TSS_REDIRECTION_SIZE (256 / 8)
 534 #define RMODE_TSS_SIZE (TSS_BASE_SIZE + TSS_REDIRECTION_SIZE + TSS_IOPB_SIZE + 1)
 535
 536 #ifdef CONFIG_X86_64
 537
 538 /*
 539  * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.  Therefore
 540  * we need to allocate shadow page tables in the first 4GB of memory, which
 541  * happens to fit the DMA32 zone.
 542  */
 543 #define GFP_KVM_MMU (GFP_KERNEL | __GFP_DMA32)
 544
 545 #else
 546
 547 #define GFP_KVM_MMU GFP_KERNEL
 548
 549 #endif
 550
 551 #endif