tools/kvm/kvm.c

   1 #include "kvm/kvm.h"
   2
   3 #include "kvm/cpufeature.h"
   4 #include "kvm/interrupt.h"
   5 #include "kvm/boot-protocol.h"
   6 #include "kvm/util.h"
   7
   8 #include <linux/kvm.h>
   9
  10 #include <asm/bootparam.h>
  11
  12 #include <sys/ioctl.h>
  13 #include <inttypes.h>
  14 #include <sys/mman.h>
  15 #include <sys/stat.h>
  16 #include <stdbool.h>
  17 #include <assert.h>
  18 #include <limits.h>
  19 #include <signal.h>
  20 #include <stdarg.h>
  21 #include <stdlib.h>
  22 #include <string.h>
  23 #include <unistd.h>
  24 #include <stdio.h>
  25 #include <fcntl.h>
  26 #include <time.h>
  27
  28 #define DEFINE_KVM_EXIT_REASON(reason) [reason] = #reason
  29
  30 const char *kvm_exit_reasons[] = {
  31         DEFINE_KVM_EXIT_REASON(KVM_EXIT_UNKNOWN),
  32         DEFINE_KVM_EXIT_REASON(KVM_EXIT_EXCEPTION),
  33         DEFINE_KVM_EXIT_REASON(KVM_EXIT_IO),
  34         DEFINE_KVM_EXIT_REASON(KVM_EXIT_HYPERCALL),
  35         DEFINE_KVM_EXIT_REASON(KVM_EXIT_DEBUG),
  36         DEFINE_KVM_EXIT_REASON(KVM_EXIT_HLT),
  37         DEFINE_KVM_EXIT_REASON(KVM_EXIT_MMIO),
  38         DEFINE_KVM_EXIT_REASON(KVM_EXIT_IRQ_WINDOW_OPEN),
  39         DEFINE_KVM_EXIT_REASON(KVM_EXIT_SHUTDOWN),
  40         DEFINE_KVM_EXIT_REASON(KVM_EXIT_FAIL_ENTRY),
  41         DEFINE_KVM_EXIT_REASON(KVM_EXIT_INTR),
  42         DEFINE_KVM_EXIT_REASON(KVM_EXIT_SET_TPR),
  43         DEFINE_KVM_EXIT_REASON(KVM_EXIT_TPR_ACCESS),
  44         DEFINE_KVM_EXIT_REASON(KVM_EXIT_S390_SIEIC),
  45         DEFINE_KVM_EXIT_REASON(KVM_EXIT_S390_RESET),
  46         DEFINE_KVM_EXIT_REASON(KVM_EXIT_DCR),
  47         DEFINE_KVM_EXIT_REASON(KVM_EXIT_NMI),
  48         DEFINE_KVM_EXIT_REASON(KVM_EXIT_INTERNAL_ERROR),
  49 };
  50
  51 #define DEFINE_KVM_EXT(ext)             \
  52         .name = #ext,                   \
  53         .code = ext
  54
  55 struct {
  56         const char *name;
  57         int code;
  58 } kvm_req_ext[] = {
  59         { DEFINE_KVM_EXT(KVM_CAP_COALESCED_MMIO) },
  60         { DEFINE_KVM_EXT(KVM_CAP_SET_TSS_ADDR) },
  61         { DEFINE_KVM_EXT(KVM_CAP_PIT2) },
  62         { DEFINE_KVM_EXT(KVM_CAP_USER_MEMORY) },
  63         { DEFINE_KVM_EXT(KVM_CAP_IRQ_ROUTING) },
  64         { DEFINE_KVM_EXT(KVM_CAP_IRQCHIP) },
  65         { DEFINE_KVM_EXT(KVM_CAP_HLT) },
  66         { DEFINE_KVM_EXT(KVM_CAP_IRQ_INJECT_STATUS) },
  67         { DEFINE_KVM_EXT(KVM_CAP_EXT_CPUID) },
  68 };
  69
  70 static bool kvm__supports_extension(struct kvm *self, unsigned int extension)
  71 {
  72         int ret;
  73
  74         ret = ioctl(self->sys_fd, KVM_CHECK_EXTENSION, extension);
  75         if (ret < 0)
  76                 return false;
  77
  78         return ret;
  79 }
  80
  81 static int kvm__check_extensions(struct kvm *self)
  82 {
  83         unsigned int i;
  84
  85         for (i = 0; i < ARRAY_SIZE(kvm_req_ext); i++) {
  86                 if (!kvm__supports_extension(self, kvm_req_ext[i].code)) {
  87                         error("Unsuppored KVM extension detected: %s",
  88                                 kvm_req_ext[i].name);
  89                         return (int)-i;
  90                 }
  91         }
  92
  93         return 0;
  94 }
  95
  96 static struct kvm *kvm__new(void)
  97 {
  98         struct kvm *self = calloc(1, sizeof *self);
  99
 100         if (!self)
 101                 die("out of memory");
 102
 103         return self;
 104 }
 105
 106 void kvm__delete(struct kvm *self)
 107 {
 108         kvm__stop_timer(self);
 109
 110         free(self->ram_start);
 111         free(self);
 112 }
 113
 114 static bool kvm__cpu_supports_vm(void)
 115 {
 116         struct cpuid_regs regs;
 117         uint32_t eax_base;
 118         int feature;
 119
 120         regs    = (struct cpuid_regs) {
 121                 .eax            = 0x00,
 122         };
 123         host_cpuid(&regs);
 124
 125         switch (regs.ebx) {
 126         case CPUID_VENDOR_INTEL_1:
 127                 eax_base        = 0x00;
 128                 feature         = KVM__X86_FEATURE_VMX;
 129                 break;
 130
 131         case CPUID_VENDOR_AMD_1:
 132                 eax_base        = 0x80000000;
 133                 feature         = KVM__X86_FEATURE_SVM;
 134                 break;
 135
 136         default:
 137                 return false;
 138         }
 139
 140         regs    = (struct cpuid_regs) {
 141                 .eax            = eax_base,
 142         };
 143         host_cpuid(&regs);
 144
 145         if (regs.eax < eax_base + 0x01)
 146                 return false;
 147
 148         regs    = (struct cpuid_regs) {
 149                 .eax            = eax_base + 0x01
 150         };
 151         host_cpuid(&regs);
 152
 153         return regs.ecx & (1 << feature);
 154 }
 155
 156 struct kvm *kvm__init(const char *kvm_dev, unsigned long ram_size)
 157 {
 158         struct kvm_userspace_memory_region mem;
 159         struct kvm_pit_config pit_config = { .flags = 0, };
 160         struct kvm *self;
 161         long page_size;
 162         int ret;
 163
 164         if (!kvm__cpu_supports_vm())
 165                 die("Your CPU does not support hardware virtualization");
 166
 167         self = kvm__new();
 168
 169         self->sys_fd = open(kvm_dev, O_RDWR);
 170         if (self->sys_fd < 0) {
 171                 if (errno == ENOENT)
 172                         die("'%s' not found. Please make sure your kernel has CONFIG_KVM enabled and that the KVM modules are loaded.", kvm_dev);
 173                 if (errno == ENODEV)
 174                         die("'%s' KVM driver not available.\n  # (If the KVM module is loaded then 'dmesg' may offer further clues about the failure.)", kvm_dev);
 175
 176                 fprintf(stderr, "  Fatal, could not open %s: ", kvm_dev);
 177                 perror(NULL);
 178                 exit(1);
 179         }
 180
 181         ret = ioctl(self->sys_fd, KVM_GET_API_VERSION, 0);
 182         if (ret != KVM_API_VERSION)
 183                 die_perror("KVM_API_VERSION ioctl");
 184
 185         self->vm_fd = ioctl(self->sys_fd, KVM_CREATE_VM, 0);
 186         if (self->vm_fd < 0)
 187                 die_perror("KVM_CREATE_VM ioctl");
 188
 189         if (kvm__check_extensions(self))
 190                 die("A required KVM extention is not supported by OS");
 191
 192         ret = ioctl(self->vm_fd, KVM_SET_TSS_ADDR, 0xfffbd000);
 193         if (ret < 0)
 194                 die_perror("KVM_SET_TSS_ADDR ioctl");
 195
 196         ret = ioctl(self->vm_fd, KVM_CREATE_PIT2, &pit_config);
 197         if (ret < 0)
 198                 die_perror("KVM_CREATE_PIT2 ioctl");
 199
 200         self->ram_size          = ram_size;
 201
 202         page_size       = sysconf(_SC_PAGESIZE);
 203         if (posix_memalign(&self->ram_start, page_size, self->ram_size) != 0)
 204                 die("out of memory");
 205
 206         mem = (struct kvm_userspace_memory_region) {
 207                 .slot                   = 0,
 208                 .guest_phys_addr        = 0x0UL,
 209                 .memory_size            = self->ram_size,
 210                 .userspace_addr         = (unsigned long) self->ram_start,
 211         };
 212
 213         ret = ioctl(self->vm_fd, KVM_SET_USER_MEMORY_REGION, &mem);
 214         if (ret < 0)
 215                 die_perror("KVM_SET_USER_MEMORY_REGION ioctl");
 216
 217         ret = ioctl(self->vm_fd, KVM_CREATE_IRQCHIP);
 218         if (ret < 0)
 219                 die_perror("KVM_CREATE_IRQCHIP ioctl");
 220
 221         return self;
 222 }
 223
 224 #define BOOT_LOADER_SELECTOR    0x1000
 225 #define BOOT_LOADER_IP          0x0000
 226 #define BOOT_LOADER_SP          0x8000
 227 #define BOOT_CMDLINE_OFFSET     0x20000
 228
 229 #define BOOT_PROTOCOL_REQUIRED  0x206
 230 #define LOAD_HIGH               0x01
 231
 232 static int load_flat_binary(struct kvm *self, int fd)
 233 {
 234         void *p;
 235         int nr;
 236
 237         if (lseek(fd, 0, SEEK_SET) < 0)
 238                 die_perror("lseek");
 239
 240         p = guest_real_to_host(self, BOOT_LOADER_SELECTOR, BOOT_LOADER_IP);
 241
 242         while ((nr = read(fd, p, 65536)) > 0)
 243                 p += nr;
 244
 245         self->boot_selector     = BOOT_LOADER_SELECTOR;
 246         self->boot_ip           = BOOT_LOADER_IP;
 247         self->boot_sp           = BOOT_LOADER_SP;
 248
 249         return true;
 250 }
 251
 252 static const char *BZIMAGE_MAGIC        = "HdrS";
 253
 254 static bool load_bzimage(struct kvm *self, int fd_kernel,
 255                         int fd_initrd, const char *kernel_cmdline)
 256 {
 257         struct boot_params *kern_boot;
 258         unsigned long setup_sects;
 259         struct boot_params boot;
 260         size_t cmdline_size;
 261         ssize_t setup_size;
 262         void *p;
 263         int nr;
 264
 265         /*
 266          * See Documentation/x86/boot.txt for details no bzImage on-disk and
 267          * memory layout.
 268          */
 269
 270         if (lseek(fd_kernel, 0, SEEK_SET) < 0)
 271                 die_perror("lseek");
 272
 273         if (read(fd_kernel, &boot, sizeof(boot)) != sizeof(boot))
 274                 return false;
 275
 276         if (memcmp(&boot.hdr.header, BZIMAGE_MAGIC, strlen(BZIMAGE_MAGIC)))
 277                 return false;
 278
 279         if (boot.hdr.version < BOOT_PROTOCOL_REQUIRED)
 280                 die("Too old kernel");
 281
 282         if (lseek(fd_kernel, 0, SEEK_SET) < 0)
 283                 die_perror("lseek");
 284
 285         if (!boot.hdr.setup_sects)
 286                 boot.hdr.setup_sects = BZ_DEFAULT_SETUP_SECTS;
 287         setup_sects = boot.hdr.setup_sects + 1;
 288
 289         setup_size = setup_sects << 9;
 290         p = guest_real_to_host(self, BOOT_LOADER_SELECTOR, BOOT_LOADER_IP);
 291
 292         /* copy setup.bin to mem*/
 293         if (read(fd_kernel, p, setup_size) != setup_size)
 294                 die_perror("read");
 295
 296         /* copy vmlinux.bin to BZ_KERNEL_START*/
 297         p = guest_flat_to_host(self, BZ_KERNEL_START);
 298
 299         while ((nr = read(fd_kernel, p, 65536)) > 0)
 300                 p += nr;
 301
 302         p = guest_flat_to_host(self, BOOT_CMDLINE_OFFSET);
 303         if (kernel_cmdline) {
 304                 cmdline_size = strlen(kernel_cmdline) + 1;
 305                 if (cmdline_size > boot.hdr.cmdline_size)
 306                         cmdline_size = boot.hdr.cmdline_size;
 307
 308                 memset(p, 0, boot.hdr.cmdline_size);
 309                 memcpy(p, kernel_cmdline, cmdline_size - 1);
 310         }
 311
 312         kern_boot       = guest_real_to_host(self, BOOT_LOADER_SELECTOR, 0x00);
 313
 314         kern_boot->hdr.cmd_line_ptr     = BOOT_CMDLINE_OFFSET;
 315         kern_boot->hdr.type_of_loader   = 0xff;
 316         kern_boot->hdr.heap_end_ptr     = 0xfe00;
 317         kern_boot->hdr.loadflags        |= CAN_USE_HEAP;
 318
 319         /*
 320          * Read initrd image into guest memory
 321          */
 322         if (fd_initrd >= 0) {
 323                 struct stat initrd_stat;
 324                 unsigned long addr;
 325
 326                 if (fstat(fd_initrd, &initrd_stat))
 327                         die_perror("fstat");
 328
 329                 addr = boot.hdr.initrd_addr_max & ~0xfffff;
 330                 for (;;) {
 331                         if (addr < BZ_KERNEL_START)
 332                                 die("Not enough memory for initrd");
 333                         else if (addr < (self->ram_size - initrd_stat.st_size))
 334                                 break;
 335                         addr -= 0x100000;
 336                 }
 337
 338                 p = guest_flat_to_host(self, addr);
 339                 nr = read(fd_initrd, p, initrd_stat.st_size);
 340                 if (nr != initrd_stat.st_size)
 341                         die("Failed to read initrd");
 342
 343                 kern_boot->hdr.ramdisk_image    = addr;
 344                 kern_boot->hdr.ramdisk_size     = initrd_stat.st_size;
 345         }
 346
 347         self->boot_selector     = BOOT_LOADER_SELECTOR;
 348         /*
 349          * The real-mode setup code starts at offset 0x200 of a bzImage. See
 350          * Documentation/x86/boot.txt for details.
 351          */
 352         self->boot_ip           = BOOT_LOADER_IP + 0x200;
 353         self->boot_sp           = BOOT_LOADER_SP;
 354
 355         return true;
 356 }
 357
 358 bool kvm__load_kernel(struct kvm *kvm, const char *kernel_filename,
 359                 const char *initrd_filename, const char *kernel_cmdline)
 360 {
 361         bool ret;
 362         int fd_kernel = -1, fd_initrd = -1;
 363
 364         fd_kernel = open(kernel_filename, O_RDONLY);
 365         if (fd_kernel < 0)
 366                 die("Unable to open kernel %s", kernel_filename);
 367
 368         if (initrd_filename) {
 369                 fd_initrd = open(initrd_filename, O_RDONLY);
 370                 if (fd_initrd < 0)
 371                         die("Unable to open initrd %s", initrd_filename);
 372         }
 373
 374         ret = load_bzimage(kvm, fd_kernel, fd_initrd, kernel_cmdline);
 375
 376         if (initrd_filename)
 377                 close(fd_initrd);
 378
 379         if (ret)
 380                 goto found_kernel;
 381
 382         warning("%s is not a bzImage. Trying to load it as a flat binary...", kernel_filename);
 383
 384         ret = load_flat_binary(kvm, fd_kernel);
 385         if (ret)
 386                 goto found_kernel;
 387
 388         die("%s is not a valid bzImage or flat binary", kernel_filename);
 389
 390 found_kernel:
 391         return ret;
 392 }
 393
 394 /**
 395  * kvm__setup_bios - inject BIOS into guest system memory
 396  * @self - guest system descriptor
 397  *
 398  * This function is a main routine where we poke guest memory
 399  * and install BIOS there.
 400  */
 401 void kvm__setup_bios(struct kvm *self)
 402 {
 403         /* standart minimal configuration */
 404         setup_bios(self);
 405
 406         /* FIXME: SMP, ACPI and friends here */
 407 }
 408
 409 #define TIMER_INTERVAL_NS 1000000       /* 1 msec */
 410
 411 static void alarm_handler(int sig)
 412 {
 413 }
 414
 415 /*
 416  * This function sets up a timer that's used to inject interrupts from the
 417  * userspace hypervisor into the guest at periodical intervals. Please note
 418  * that clock interrupt, for example, is not handled here.
 419  */
 420 void kvm__start_timer(struct kvm *self)
 421 {
 422         struct itimerspec its;
 423         struct sigaction sa;
 424         struct sigevent sev;
 425
 426         sigfillset(&sa.sa_mask);
 427         sa.sa_flags                     = 0;
 428         sa.sa_handler                   = alarm_handler;
 429
 430         sigaction(SIGALRM, &sa, NULL);
 431
 432         memset(&sev, 0, sizeof(struct sigevent));
 433         sev.sigev_value.sival_int       = 0;
 434         sev.sigev_notify                = SIGEV_SIGNAL;
 435         sev.sigev_signo                 = SIGALRM;
 436
 437         if (timer_create(CLOCK_REALTIME, &sev, &self->timerid) < 0)
 438                 die("timer_create()");
 439
 440         its.it_value.tv_sec             = TIMER_INTERVAL_NS / 1000000000;
 441         its.it_value.tv_nsec            = TIMER_INTERVAL_NS % 1000000000;
 442         its.it_interval.tv_sec          = its.it_value.tv_sec;
 443         its.it_interval.tv_nsec         = its.it_value.tv_nsec;
 444
 445         if (timer_settime(self->timerid, 0, &its, NULL) < 0)
 446                 die("timer_settime()");
 447 }
 448
 449 void kvm__stop_timer(struct kvm *self)
 450 {
 451         if (self->timerid)
 452                 if (timer_delete(self->timerid) < 0)
 453                         die("timer_delete()");
 454
 455         self->timerid = 0;
 456 }
 457
 458 void kvm__irq_line(struct kvm *self, int irq, int level)
 459 {
 460         struct kvm_irq_level irq_level;
 461
 462         irq_level       = (struct kvm_irq_level) {
 463                 {
 464                         .irq            = irq,
 465                 },
 466                 .level          = level,
 467         };
 468
 469         if (ioctl(self->vm_fd, KVM_IRQ_LINE, &irq_level) < 0)
 470                 die_perror("KVM_IRQ_LINE failed");
 471 }
 472
 473 void kvm__dump_mem(struct kvm *self, unsigned long addr, unsigned long size)
 474 {
 475         unsigned char *p;
 476         unsigned long n;
 477
 478         size &= ~7; /* mod 8 */
 479         if (!size)
 480                 return;
 481
 482         p = guest_flat_to_host(self, addr);
 483
 484         for (n = 0; n < size; n += 8) {
 485                 if (!host_ptr_in_ram(self, p + n))
 486                         break;
 487
 488                 printf("  0x%08lx: %02x %02x %02x %02x  %02x %02x %02x %02x\n",
 489                         addr + n, p[n + 0], p[n + 1], p[n + 2], p[n + 3],
 490                                   p[n + 4], p[n + 5], p[n + 6], p[n + 7]);
 491         }
 492 }