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         int interrupt_window_open;
 177         unsigned long irq_summary; /* bit vector: 1 per word in irq_pending */
 178 #define NR_IRQ_WORDS KVM_IRQ_BITMAP_SIZE(unsigned long)
 179         unsigned long irq_pending[NR_IRQ_WORDS];
 180         unsigned long regs[NR_VCPU_REGS]; /* for rsp: vcpu_load_rsp_rip() */
 181         unsigned long rip;      /* needs vcpu_load_rsp_rip() */
 182
 183         unsigned long cr0;
 184         unsigned long cr2;
 185         unsigned long cr3;
 186         unsigned long cr4;
 187         unsigned long cr8;
 188         u64 shadow_efer;
 189         u64 apic_base;
 190         int nmsrs;
 191         struct vmx_msr_entry *guest_msrs;
 192         struct vmx_msr_entry *host_msrs;
 193
 194         struct list_head free_pages;
 195         struct kvm_mmu_page page_header_buf[KVM_NUM_MMU_PAGES];
 196         struct kvm_mmu mmu;
 197
 198         struct kvm_guest_debug guest_debug;
 199
 200         char fx_buf[FX_BUF_SIZE];
 201         char *host_fx_image;
 202         char *guest_fx_image;
 203
 204         int mmio_needed;
 205         int mmio_read_completed;
 206         int mmio_is_write;
 207         int mmio_size;
 208         unsigned char mmio_data[8];
 209         gpa_t mmio_phys_addr;
 210
 211         struct {
 212                 int active;
 213                 u8 save_iopl;
 214                 struct kvm_save_segment {
 215                         u16 selector;
 216                         unsigned long base;
 217                         u32 limit;
 218                         u32 ar;
 219                 } tr, es, ds, fs, gs;
 220         } rmode;
 221 };
 222
 223 struct kvm_memory_slot {
 224         gfn_t base_gfn;
 225         unsigned long npages;
 226         unsigned long flags;
 227         struct page **phys_mem;
 228         unsigned long *dirty_bitmap;
 229 };
 230
 231 struct kvm {
 232         spinlock_t lock; /* protects everything except vcpus */
 233         int nmemslots;
 234         struct kvm_memory_slot memslots[KVM_MEMORY_SLOTS];
 235         struct list_head active_mmu_pages;
 236         struct kvm_vcpu vcpus[KVM_MAX_VCPUS];
 237         int memory_config_version;
 238         int busy;
 239         unsigned long rmap_overflow;
 240 };
 241
 242 struct kvm_stat {
 243         u32 pf_fixed;
 244         u32 pf_guest;
 245         u32 tlb_flush;
 246         u32 invlpg;
 247
 248         u32 exits;
 249         u32 io_exits;
 250         u32 mmio_exits;
 251         u32 signal_exits;
 252         u32 irq_window_exits;
 253         u32 halt_exits;
 254         u32 request_irq_exits;
 255         u32 irq_exits;
 256 };
 257
 258 struct descriptor_table {
 259         u16 limit;
 260         unsigned long base;
 261 } __attribute__((packed));
 262
 263 struct kvm_arch_ops {
 264         int (*cpu_has_kvm_support)(void);          /* __init */
 265         int (*disabled_by_bios)(void);             /* __init */
 266         void (*hardware_enable)(void *dummy);      /* __init */
 267         void (*hardware_disable)(void *dummy);
 268         int (*hardware_setup)(void);               /* __init */
 269         void (*hardware_unsetup)(void);            /* __exit */
 270
 271         int (*vcpu_create)(struct kvm_vcpu *vcpu);
 272         void (*vcpu_free)(struct kvm_vcpu *vcpu);
 273
 274         struct kvm_vcpu *(*vcpu_load)(struct kvm_vcpu *vcpu);
 275         void (*vcpu_put)(struct kvm_vcpu *vcpu);
 276
 277         int (*set_guest_debug)(struct kvm_vcpu *vcpu,
 278                                struct kvm_debug_guest *dbg);
 279         int (*get_msr)(struct kvm_vcpu *vcpu, u32 msr_index, u64 *pdata);
 280         int (*set_msr)(struct kvm_vcpu *vcpu, u32 msr_index, u64 data);
 281         u64 (*get_segment_base)(struct kvm_vcpu *vcpu, int seg);
 282         void (*get_segment)(struct kvm_vcpu *vcpu,
 283                             struct kvm_segment *var, int seg);
 284         void (*set_segment)(struct kvm_vcpu *vcpu,
 285                             struct kvm_segment *var, int seg);
 286         void (*get_cs_db_l_bits)(struct kvm_vcpu *vcpu, int *db, int *l);
 287         void (*decache_cr0_cr4_guest_bits)(struct kvm_vcpu *vcpu);
 288         void (*set_cr0)(struct kvm_vcpu *vcpu, unsigned long cr0);
 289         void (*set_cr0_no_modeswitch)(struct kvm_vcpu *vcpu,
 290                                       unsigned long cr0);
 291         void (*set_cr3)(struct kvm_vcpu *vcpu, unsigned long cr3);
 292         void (*set_cr4)(struct kvm_vcpu *vcpu, unsigned long cr4);
 293         void (*set_efer)(struct kvm_vcpu *vcpu, u64 efer);
 294         void (*get_idt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 295         void (*set_idt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 296         void (*get_gdt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 297         void (*set_gdt)(struct kvm_vcpu *vcpu, struct descriptor_table *dt);
 298         unsigned long (*get_dr)(struct kvm_vcpu *vcpu, int dr);
 299         void (*set_dr)(struct kvm_vcpu *vcpu, int dr, unsigned long value,
 300                        int *exception);
 301         void (*cache_regs)(struct kvm_vcpu *vcpu);
 302         void (*decache_regs)(struct kvm_vcpu *vcpu);
 303         unsigned long (*get_rflags)(struct kvm_vcpu *vcpu);
 304         void (*set_rflags)(struct kvm_vcpu *vcpu, unsigned long rflags);
 305
 306         void (*invlpg)(struct kvm_vcpu *vcpu, gva_t addr);
 307         void (*tlb_flush)(struct kvm_vcpu *vcpu);
 308         void (*inject_page_fault)(struct kvm_vcpu *vcpu,
 309                                   unsigned long addr, u32 err_code);
 310
 311         void (*inject_gp)(struct kvm_vcpu *vcpu, unsigned err_code);
 312
 313         int (*run)(struct kvm_vcpu *vcpu, struct kvm_run *run);
 314         int (*vcpu_setup)(struct kvm_vcpu *vcpu);
 315         void (*skip_emulated_instruction)(struct kvm_vcpu *vcpu);
 316 };
 317
 318 extern struct kvm_stat kvm_stat;
 319 extern struct kvm_arch_ops *kvm_arch_ops;
 320
 321 #define kvm_printf(kvm, fmt ...) printk(KERN_DEBUG fmt)
 322 #define vcpu_printf(vcpu, fmt...) kvm_printf(vcpu->kvm, fmt)
 323
 324 int kvm_init_arch(struct kvm_arch_ops *ops, struct module *module);
 325 void kvm_exit_arch(void);
 326
 327 void kvm_mmu_destroy(struct kvm_vcpu *vcpu);
 328 int kvm_mmu_create(struct kvm_vcpu *vcpu);
 329 int kvm_mmu_setup(struct kvm_vcpu *vcpu);
 330
 331 int kvm_mmu_reset_context(struct kvm_vcpu *vcpu);
 332 void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot);
 333
 334 hpa_t gpa_to_hpa(struct kvm_vcpu *vcpu, gpa_t gpa);
 335 #define HPA_MSB ((sizeof(hpa_t) * 8) - 1)
 336 #define HPA_ERR_MASK ((hpa_t)1 << HPA_MSB)
 337 static inline int is_error_hpa(hpa_t hpa) { return hpa >> HPA_MSB; }
 338 hpa_t gva_to_hpa(struct kvm_vcpu *vcpu, gva_t gva);
 339
 340 void kvm_emulator_want_group7_invlpg(void);
 341
 342 extern hpa_t bad_page_address;
 343
 344 static inline struct page *gfn_to_page(struct kvm_memory_slot *slot, gfn_t gfn)
 345 {
 346         return slot->phys_mem[gfn - slot->base_gfn];
 347 }
 348
 349 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn);
 350 void mark_page_dirty(struct kvm *kvm, gfn_t gfn);
 351
 352 enum emulation_result {
 353         EMULATE_DONE,       /* no further processing */
 354         EMULATE_DO_MMIO,      /* kvm_run filled with mmio request */
 355         EMULATE_FAIL,         /* can't emulate this instruction */
 356 };
 357
 358 int emulate_instruction(struct kvm_vcpu *vcpu, struct kvm_run *run,
 359                         unsigned long cr2, u16 error_code);
 360 void realmode_lgdt(struct kvm_vcpu *vcpu, u16 size, unsigned long address);
 361 void realmode_lidt(struct kvm_vcpu *vcpu, u16 size, unsigned long address);
 362 void realmode_lmsw(struct kvm_vcpu *vcpu, unsigned long msw,
 363                    unsigned long *rflags);
 364
 365 unsigned long realmode_get_cr(struct kvm_vcpu *vcpu, int cr);
 366 void realmode_set_cr(struct kvm_vcpu *vcpu, int cr, unsigned long value,
 367                      unsigned long *rflags);
 368
 369 struct x86_emulate_ctxt;
 370
 371 int emulate_invlpg(struct kvm_vcpu *vcpu, gva_t address);
 372 int emulate_clts(struct kvm_vcpu *vcpu);
 373 int emulator_get_dr(struct x86_emulate_ctxt* ctxt, int dr,
 374                     unsigned long *dest);
 375 int emulator_set_dr(struct x86_emulate_ctxt *ctxt, int dr,
 376                     unsigned long value);
 377
 378 void set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0);
 379 void set_cr3(struct kvm_vcpu *vcpu, unsigned long cr0);
 380 void set_cr4(struct kvm_vcpu *vcpu, unsigned long cr0);
 381 void set_cr8(struct kvm_vcpu *vcpu, unsigned long cr0);
 382 void lmsw(struct kvm_vcpu *vcpu, unsigned long msw);
 383
 384 int kvm_get_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata);
 385 int kvm_set_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 data);
 386
 387 void fx_init(struct kvm_vcpu *vcpu);
 388
 389 void load_msrs(struct vmx_msr_entry *e, int n);
 390 void save_msrs(struct vmx_msr_entry *e, int n);
 391 void kvm_resched(struct kvm_vcpu *vcpu);
 392
 393 int kvm_read_guest(struct kvm_vcpu *vcpu,
 394                gva_t addr,
 395                unsigned long size,
 396                void *dest);
 397
 398 int kvm_write_guest(struct kvm_vcpu *vcpu,
 399                 gva_t addr,
 400                 unsigned long size,
 401                 void *data);
 402
 403 unsigned long segment_base(u16 selector);
 404
 405 static inline struct page *_gfn_to_page(struct kvm *kvm, gfn_t gfn)
 406 {
 407         struct kvm_memory_slot *slot = gfn_to_memslot(kvm, gfn);
 408         return (slot) ? slot->phys_mem[gfn - slot->base_gfn] : NULL;
 409 }
 410
 411 static inline int is_long_mode(struct kvm_vcpu *vcpu)
 412 {
 413 #ifdef CONFIG_X86_64
 414         return vcpu->shadow_efer & EFER_LME;
 415 #else
 416         return 0;
 417 #endif
 418 }
 419
 420 static inline int is_pae(struct kvm_vcpu *vcpu)
 421 {
 422         return vcpu->cr4 & CR4_PAE_MASK;
 423 }
 424
 425 static inline int is_pse(struct kvm_vcpu *vcpu)
 426 {
 427         return vcpu->cr4 & CR4_PSE_MASK;
 428 }
 429
 430 static inline int is_paging(struct kvm_vcpu *vcpu)
 431 {
 432         return vcpu->cr0 & CR0_PG_MASK;
 433 }
 434
 435 static inline int memslot_id(struct kvm *kvm, struct kvm_memory_slot *slot)
 436 {
 437         return slot - kvm->memslots;
 438 }
 439
 440 static inline struct kvm_mmu_page *page_header(hpa_t shadow_page)
 441 {
 442         struct page *page = pfn_to_page(shadow_page >> PAGE_SHIFT);
 443
 444         return (struct kvm_mmu_page *)page->private;
 445 }
 446
 447 static inline u16 read_fs(void)
 448 {
 449         u16 seg;
 450         asm ("mov %%fs, %0" : "=g"(seg));
 451         return seg;
 452 }
 453
 454 static inline u16 read_gs(void)
 455 {
 456         u16 seg;
 457         asm ("mov %%gs, %0" : "=g"(seg));
 458         return seg;
 459 }
 460
 461 static inline u16 read_ldt(void)
 462 {
 463         u16 ldt;
 464         asm ("sldt %0" : "=g"(ldt));
 465         return ldt;
 466 }
 467
 468 static inline void load_fs(u16 sel)
 469 {
 470         asm ("mov %0, %%fs" : : "rm"(sel));
 471 }
 472
 473 static inline void load_gs(u16 sel)
 474 {
 475         asm ("mov %0, %%gs" : : "rm"(sel));
 476 }
 477
 478 #ifndef load_ldt
 479 static inline void load_ldt(u16 sel)
 480 {
 481         asm ("lldt %0" : : "g"(sel));
 482 }
 483 #endif
 484
 485 static inline void get_idt(struct descriptor_table *table)
 486 {
 487         asm ("sidt %0" : "=m"(*table));
 488 }
 489
 490 static inline void get_gdt(struct descriptor_table *table)
 491 {
 492         asm ("sgdt %0" : "=m"(*table));
 493 }
 494
 495 static inline unsigned long read_tr_base(void)
 496 {
 497         u16 tr;
 498         asm ("str %0" : "=g"(tr));
 499         return segment_base(tr);
 500 }
 501
 502 #ifdef CONFIG_X86_64
 503 static inline unsigned long read_msr(unsigned long msr)
 504 {
 505         u64 value;
 506
 507         rdmsrl(msr, value);
 508         return value;
 509 }
 510 #endif
 511
 512 static inline void fx_save(void *image)
 513 {
 514         asm ("fxsave (%0)":: "r" (image));
 515 }
 516
 517 static inline void fx_restore(void *image)
 518 {
 519         asm ("fxrstor (%0)":: "r" (image));
 520 }
 521
 522 static inline void fpu_init(void)
 523 {
 524         asm ("finit");
 525 }
 526
 527 static inline u32 get_rdx_init_val(void)
 528 {
 529         return 0x600; /* P6 family */
 530 }
 531
 532 #define ASM_VMX_VMCLEAR_RAX       ".byte 0x66, 0x0f, 0xc7, 0x30"
 533 #define ASM_VMX_VMLAUNCH          ".byte 0x0f, 0x01, 0xc2"
 534 #define ASM_VMX_VMRESUME          ".byte 0x0f, 0x01, 0xc3"
 535 #define ASM_VMX_VMPTRLD_RAX       ".byte 0x0f, 0xc7, 0x30"
 536 #define ASM_VMX_VMREAD_RDX_RAX    ".byte 0x0f, 0x78, 0xd0"
 537 #define ASM_VMX_VMWRITE_RAX_RDX   ".byte 0x0f, 0x79, 0xd0"
 538 #define ASM_VMX_VMWRITE_RSP_RDX   ".byte 0x0f, 0x79, 0xd4"
 539 #define ASM_VMX_VMXOFF            ".byte 0x0f, 0x01, 0xc4"
 540 #define ASM_VMX_VMXON_RAX         ".byte 0xf3, 0x0f, 0xc7, 0x30"
 541
 542 #define MSR_IA32_TIME_STAMP_COUNTER             0x010
 543
 544 #define TSS_IOPB_BASE_OFFSET 0x66
 545 #define TSS_BASE_SIZE 0x68
 546 #define TSS_IOPB_SIZE (65536 / 8)
 547 #define TSS_REDIRECTION_SIZE (256 / 8)
 548 #define RMODE_TSS_SIZE (TSS_BASE_SIZE + TSS_REDIRECTION_SIZE + TSS_IOPB_SIZE + 1)
 549
 550 #ifdef CONFIG_X86_64
 551
 552 /*
 553  * When emulating 32-bit mode, cr3 is only 32 bits even on x86_64.  Therefore
 554  * we need to allocate shadow page tables in the first 4GB of memory, which
 555  * happens to fit the DMA32 zone.
 556  */
 557 #define GFP_KVM_MMU (GFP_KERNEL | __GFP_DMA32)
 558
 559 #else
 560
 561 #define GFP_KVM_MMU GFP_KERNEL
 562
 563 #endif
 564
 565 #endif