kvm-all.c

   1 /*
   2  * QEMU KVM support
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *           Red Hat, Inc. 2008
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Glauber Costa     <gcosta@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include <sys/types.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/mman.h>
  19 #include <stdarg.h>
  20
  21 #include <linux/kvm.h>
  22
  23 #include "qemu-common.h"
  24 #include "sysemu.h"
  25 #include "kvm.h"
  26
  27 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  28 #define PAGE_SIZE TARGET_PAGE_SIZE
  29
  30 //#define DEBUG_KVM
  31
  32 #ifdef DEBUG_KVM
  33 #define dprintf(fmt, ...) \
  34     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  35 #else
  36 #define dprintf(fmt, ...) \
  37     do { } while (0)
  38 #endif
  39
  40 typedef struct KVMSlot
  41 {
  42     target_phys_addr_t start_addr;
  43     ram_addr_t memory_size;
  44     ram_addr_t phys_offset;
  45     int slot;
  46     int flags;
  47 } KVMSlot;
  48
  49 typedef struct kvm_dirty_log KVMDirtyLog;
  50
  51 int kvm_allowed = 0;
  52
  53 struct KVMState
  54 {
  55     KVMSlot slots[32];
  56     int fd;
  57     int vmfd;
  58     int coalesced_mmio;
  59 };
  60
  61 static KVMState *kvm_state;
  62
  63 static KVMSlot *kvm_alloc_slot(KVMState *s)
  64 {
  65     int i;
  66
  67     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  68         /* KVM private memory slots */
  69         if (i >= 8 && i < 12)
  70             continue;
  71         if (s->slots[i].memory_size == 0)
  72             return &s->slots[i];
  73     }
  74
  75     return NULL;
  76 }
  77
  78 static KVMSlot *kvm_lookup_slot(KVMState *s, target_phys_addr_t start_addr)
  79 {
  80     int i;
  81
  82     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  83         KVMSlot *mem = &s->slots[i];
  84
  85         if (start_addr >= mem->start_addr &&
  86             start_addr < (mem->start_addr + mem->memory_size))
  87             return mem;
  88     }
  89
  90     return NULL;
  91 }
  92
  93 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
  94 {
  95     struct kvm_userspace_memory_region mem;
  96
  97     mem.slot = slot->slot;
  98     mem.guest_phys_addr = slot->start_addr;
  99     mem.memory_size = slot->memory_size;
 100     mem.userspace_addr = (unsigned long)phys_ram_base + slot->phys_offset;
 101     mem.flags = slot->flags;
 102
 103     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 104 }
 105
 106
 107 int kvm_init_vcpu(CPUState *env)
 108 {
 109     KVMState *s = kvm_state;
 110     long mmap_size;
 111     int ret;
 112
 113     dprintf("kvm_init_vcpu\n");
 114
 115     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 116     if (ret < 0) {
 117         dprintf("kvm_create_vcpu failed\n");
 118         goto err;
 119     }
 120
 121     env->kvm_fd = ret;
 122     env->kvm_state = s;
 123
 124     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 125     if (mmap_size < 0) {
 126         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 127         goto err;
 128     }
 129
 130     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 131                         env->kvm_fd, 0);
 132     if (env->kvm_run == MAP_FAILED) {
 133         ret = -errno;
 134         dprintf("mmap'ing vcpu state failed\n");
 135         goto err;
 136     }
 137
 138     ret = kvm_arch_init_vcpu(env);
 139
 140 err:
 141     return ret;
 142 }
 143
 144 int kvm_sync_vcpus(void)
 145 {
 146     CPUState *env;
 147
 148     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 149         int ret;
 150
 151         ret = kvm_arch_put_registers(env);
 152         if (ret)
 153             return ret;
 154     }
 155
 156     return 0;
 157 }
 158
 159 /*
 160  * dirty pages logging control
 161  */
 162 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr, target_phys_addr_t end_addr,
 163                                       unsigned flags,
 164                                       unsigned mask)
 165 {
 166     KVMState *s = kvm_state;
 167     KVMSlot *mem = kvm_lookup_slot(s, phys_addr);
 168     if (mem == NULL)  {
 169             dprintf("invalid parameters %llx-%llx\n", phys_addr, end_addr);
 170             return -EINVAL;
 171     }
 172
 173     flags = (mem->flags & ~mask) | flags;
 174     /* Nothing changed, no need to issue ioctl */
 175     if (flags == mem->flags)
 176             return 0;
 177
 178     mem->flags = flags;
 179
 180     return kvm_set_user_memory_region(s, mem);
 181 }
 182
 183 int kvm_log_start(target_phys_addr_t phys_addr, target_phys_addr_t end_addr)
 184 {
 185         return kvm_dirty_pages_log_change(phys_addr, end_addr,
 186                                           KVM_MEM_LOG_DIRTY_PAGES,
 187                                           KVM_MEM_LOG_DIRTY_PAGES);
 188 }
 189
 190 int kvm_log_stop(target_phys_addr_t phys_addr, target_phys_addr_t end_addr)
 191 {
 192         return kvm_dirty_pages_log_change(phys_addr, end_addr,
 193                                           0,
 194                                           KVM_MEM_LOG_DIRTY_PAGES);
 195 }
 196
 197 /**
 198  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 199  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 200  * This means all bits are set to dirty.
 201  *
 202  * @start_add: start of logged region. This is what we use to search the memslot
 203  * @end_addr: end of logged region.
 204  */
 205 void kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr, target_phys_addr_t end_addr)
 206 {
 207     KVMState *s = kvm_state;
 208     KVMDirtyLog d;
 209     KVMSlot *mem = kvm_lookup_slot(s, start_addr);
 210     unsigned long alloc_size;
 211     ram_addr_t addr;
 212     target_phys_addr_t phys_addr = start_addr;
 213
 214     dprintf("sync addr: %llx into %lx\n", start_addr, mem->phys_offset);
 215     if (mem == NULL) {
 216             fprintf(stderr, "BUG: %s: invalid parameters\n", __func__);
 217             return;
 218     }
 219
 220     alloc_size = mem->memory_size >> TARGET_PAGE_BITS / sizeof(d.dirty_bitmap);
 221     d.dirty_bitmap = qemu_mallocz(alloc_size);
 222
 223     d.slot = mem->slot;
 224     dprintf("slot %d, phys_addr %llx, uaddr: %llx\n",
 225             d.slot, mem->start_addr, mem->phys_offset);
 226
 227     if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 228         dprintf("ioctl failed %d\n", errno);
 229         goto out;
 230     }
 231
 232     phys_addr = start_addr;
 233     for (addr = mem->phys_offset; phys_addr < end_addr; phys_addr+= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 234         unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
 235         unsigned nr = (phys_addr - start_addr) >> TARGET_PAGE_BITS;
 236         unsigned word = nr / (sizeof(*bitmap) * 8);
 237         unsigned bit = nr % (sizeof(*bitmap) * 8);
 238         if ((bitmap[word] >> bit) & 1)
 239             cpu_physical_memory_set_dirty(addr);
 240     }
 241 out:
 242     qemu_free(d.dirty_bitmap);
 243 }
 244
 245 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 246 {
 247     int ret = -ENOSYS;
 248 #ifdef KVM_CAP_COALESCED_MMIO
 249     KVMState *s = kvm_state;
 250
 251     if (s->coalesced_mmio) {
 252         struct kvm_coalesced_mmio_zone zone;
 253
 254         zone.addr = start;
 255         zone.size = size;
 256
 257         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 258     }
 259 #endif
 260
 261     return ret;
 262 }
 263
 264 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 265 {
 266     int ret = -ENOSYS;
 267 #ifdef KVM_CAP_COALESCED_MMIO
 268     KVMState *s = kvm_state;
 269
 270     if (s->coalesced_mmio) {
 271         struct kvm_coalesced_mmio_zone zone;
 272
 273         zone.addr = start;
 274         zone.size = size;
 275
 276         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 277     }
 278 #endif
 279
 280     return ret;
 281 }
 282
 283 int kvm_init(int smp_cpus)
 284 {
 285     KVMState *s;
 286     int ret;
 287     int i;
 288
 289     if (smp_cpus > 1)
 290         return -EINVAL;
 291
 292     s = qemu_mallocz(sizeof(KVMState));
 293
 294     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 295         s->slots[i].slot = i;
 296
 297     s->vmfd = -1;
 298     s->fd = open("/dev/kvm", O_RDWR);
 299     if (s->fd == -1) {
 300         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 301         ret = -errno;
 302         goto err;
 303     }
 304
 305     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 306     if (ret < KVM_API_VERSION) {
 307         if (ret > 0)
 308             ret = -EINVAL;
 309         fprintf(stderr, "kvm version too old\n");
 310         goto err;
 311     }
 312
 313     if (ret > KVM_API_VERSION) {
 314         ret = -EINVAL;
 315         fprintf(stderr, "kvm version not supported\n");
 316         goto err;
 317     }
 318
 319     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 320     if (s->vmfd < 0)
 321         goto err;
 322
 323     /* initially, KVM allocated its own memory and we had to jump through
 324      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 325      * just use a user allocated buffer so we can use phys_ram_base
 326      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 327      */
 328     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_USER_MEMORY);
 329     if (ret <= 0) {
 330         if (ret == 0)
 331             ret = -EINVAL;
 332         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
 333         goto err;
 334     }
 335
 336     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 337      * destroyed properly.  Since we rely on this capability, refuse to work
 338      * with any kernel without this capability. */
 339     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION,
 340                     KVM_CAP_DESTROY_MEMORY_REGION_WORKS);
 341     if (ret <= 0) {
 342         if (ret == 0)
 343             ret = -EINVAL;
 344
 345         fprintf(stderr,
 346                 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
 347                 "Please upgrade to at least kvm-81.\n");
 348         goto err;
 349     }
 350
 351     s->coalesced_mmio = 0;
 352 #ifdef KVM_CAP_COALESCED_MMIO
 353     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_COALESCED_MMIO);
 354     if (ret > 0)
 355         s->coalesced_mmio = ret;
 356 #endif
 357
 358     ret = kvm_arch_init(s, smp_cpus);
 359     if (ret < 0)
 360         goto err;
 361
 362     kvm_state = s;
 363
 364     return 0;
 365
 366 err:
 367     if (s) {
 368         if (s->vmfd != -1)
 369             close(s->vmfd);
 370         if (s->fd != -1)
 371             close(s->fd);
 372     }
 373     qemu_free(s);
 374
 375     return ret;
 376 }
 377
 378 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 379                          int direction, int size, uint32_t count)
 380 {
 381     int i;
 382     uint8_t *ptr = data;
 383
 384     for (i = 0; i < count; i++) {
 385         if (direction == KVM_EXIT_IO_IN) {
 386             switch (size) {
 387             case 1:
 388                 stb_p(ptr, cpu_inb(env, port));
 389                 break;
 390             case 2:
 391                 stw_p(ptr, cpu_inw(env, port));
 392                 break;
 393             case 4:
 394                 stl_p(ptr, cpu_inl(env, port));
 395                 break;
 396             }
 397         } else {
 398             switch (size) {
 399             case 1:
 400                 cpu_outb(env, port, ldub_p(ptr));
 401                 break;
 402             case 2:
 403                 cpu_outw(env, port, lduw_p(ptr));
 404                 break;
 405             case 4:
 406                 cpu_outl(env, port, ldl_p(ptr));
 407                 break;
 408             }
 409         }
 410
 411         ptr += size;
 412     }
 413
 414     return 1;
 415 }
 416
 417 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 418 {
 419 #ifdef KVM_CAP_COALESCED_MMIO
 420     KVMState *s = kvm_state;
 421     if (s->coalesced_mmio) {
 422         struct kvm_coalesced_mmio_ring *ring;
 423
 424         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 425         while (ring->first != ring->last) {
 426             struct kvm_coalesced_mmio *ent;
 427
 428             ent = &ring->coalesced_mmio[ring->first];
 429
 430             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 431             /* FIXME smp_wmb() */
 432             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 433         }
 434     }
 435 #endif
 436 }
 437
 438 int kvm_cpu_exec(CPUState *env)
 439 {
 440     struct kvm_run *run = env->kvm_run;
 441     int ret;
 442
 443     dprintf("kvm_cpu_exec()\n");
 444
 445     do {
 446         kvm_arch_pre_run(env, run);
 447
 448         if ((env->interrupt_request & CPU_INTERRUPT_EXIT)) {
 449             dprintf("interrupt exit requested\n");
 450             ret = 0;
 451             break;
 452         }
 453
 454         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 455         kvm_arch_post_run(env, run);
 456
 457         if (ret == -EINTR || ret == -EAGAIN) {
 458             dprintf("io window exit\n");
 459             ret = 0;
 460             break;
 461         }
 462
 463         if (ret < 0) {
 464             dprintf("kvm run failed %s\n", strerror(-ret));
 465             abort();
 466         }
 467
 468         kvm_run_coalesced_mmio(env, run);
 469
 470         ret = 0; /* exit loop */
 471         switch (run->exit_reason) {
 472         case KVM_EXIT_IO:
 473             dprintf("handle_io\n");
 474             ret = kvm_handle_io(env, run->io.port,
 475                                 (uint8_t *)run + run->io.data_offset,
 476                                 run->io.direction,
 477                                 run->io.size,
 478                                 run->io.count);
 479             break;
 480         case KVM_EXIT_MMIO:
 481             dprintf("handle_mmio\n");
 482             cpu_physical_memory_rw(run->mmio.phys_addr,
 483                                    run->mmio.data,
 484                                    run->mmio.len,
 485                                    run->mmio.is_write);
 486             ret = 1;
 487             break;
 488         case KVM_EXIT_IRQ_WINDOW_OPEN:
 489             dprintf("irq_window_open\n");
 490             break;
 491         case KVM_EXIT_SHUTDOWN:
 492             dprintf("shutdown\n");
 493             qemu_system_reset_request();
 494             ret = 1;
 495             break;
 496         case KVM_EXIT_UNKNOWN:
 497             dprintf("kvm_exit_unknown\n");
 498             break;
 499         case KVM_EXIT_FAIL_ENTRY:
 500             dprintf("kvm_exit_fail_entry\n");
 501             break;
 502         case KVM_EXIT_EXCEPTION:
 503             dprintf("kvm_exit_exception\n");
 504             break;
 505         case KVM_EXIT_DEBUG:
 506             dprintf("kvm_exit_debug\n");
 507             break;
 508         default:
 509             dprintf("kvm_arch_handle_exit\n");
 510             ret = kvm_arch_handle_exit(env, run);
 511             break;
 512         }
 513     } while (ret > 0);
 514
 515     if ((env->interrupt_request & CPU_INTERRUPT_EXIT)) {
 516         env->interrupt_request &= ~CPU_INTERRUPT_EXIT;
 517         env->exception_index = EXCP_INTERRUPT;
 518     }
 519
 520     return ret;
 521 }
 522
 523 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 524                       ram_addr_t size,
 525                       ram_addr_t phys_offset)
 526 {
 527     KVMState *s = kvm_state;
 528     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 529     KVMSlot *mem;
 530
 531     /* KVM does not support read-only slots */
 532     phys_offset &= ~IO_MEM_ROM;
 533
 534     mem = kvm_lookup_slot(s, start_addr);
 535     if (mem) {
 536         if ((flags == IO_MEM_UNASSIGNED) || (flags >= TLB_MMIO)) {
 537             mem->memory_size = 0;
 538             mem->start_addr = start_addr;
 539             mem->phys_offset = 0;
 540             mem->flags = 0;
 541
 542             kvm_set_user_memory_region(s, mem);
 543         } else if (start_addr >= mem->start_addr &&
 544                    (start_addr + size) <= (mem->start_addr +
 545                                            mem->memory_size)) {
 546             KVMSlot slot;
 547             target_phys_addr_t mem_start;
 548             ram_addr_t mem_size, mem_offset;
 549
 550             /* Not splitting */
 551             if ((phys_offset - (start_addr - mem->start_addr)) ==
 552                 mem->phys_offset)
 553                 return;
 554
 555             /* unregister whole slot */
 556             memcpy(&slot, mem, sizeof(slot));
 557             mem->memory_size = 0;
 558             kvm_set_user_memory_region(s, mem);
 559
 560             /* register prefix slot */
 561             mem_start = slot.start_addr;
 562             mem_size = start_addr - slot.start_addr;
 563             mem_offset = slot.phys_offset;
 564             if (mem_size)
 565                 kvm_set_phys_mem(mem_start, mem_size, mem_offset);
 566
 567             /* register new slot */
 568             kvm_set_phys_mem(start_addr, size, phys_offset);
 569
 570             /* register suffix slot */
 571             mem_start = start_addr + size;
 572             mem_offset += mem_size + size;
 573             mem_size = slot.memory_size - mem_size - size;
 574             if (mem_size)
 575                 kvm_set_phys_mem(mem_start, mem_size, mem_offset);
 576
 577             return;
 578         } else {
 579             printf("Registering overlapping slot\n");
 580             abort();
 581         }
 582     }
 583     /* KVM does not need to know about this memory */
 584     if (flags >= IO_MEM_UNASSIGNED)
 585         return;
 586
 587     mem = kvm_alloc_slot(s);
 588     mem->memory_size = size;
 589     mem->start_addr = start_addr;
 590     mem->phys_offset = phys_offset;
 591     mem->flags = 0;
 592
 593     kvm_set_user_memory_region(s, mem);
 594     /* FIXME deal with errors */
 595 }
 596
 597 int kvm_ioctl(KVMState *s, int type, ...)
 598 {
 599     int ret;
 600     void *arg;
 601     va_list ap;
 602
 603     va_start(ap, type);
 604     arg = va_arg(ap, void *);
 605     va_end(ap);
 606
 607     ret = ioctl(s->fd, type, arg);
 608     if (ret == -1)
 609         ret = -errno;
 610
 611     return ret;
 612 }
 613
 614 int kvm_vm_ioctl(KVMState *s, int type, ...)
 615 {
 616     int ret;
 617     void *arg;
 618     va_list ap;
 619
 620     va_start(ap, type);
 621     arg = va_arg(ap, void *);
 622     va_end(ap);
 623
 624     ret = ioctl(s->vmfd, type, arg);
 625     if (ret == -1)
 626         ret = -errno;
 627
 628     return ret;
 629 }
 630
 631 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 632 {
 633     int ret;
 634     void *arg;
 635     va_list ap;
 636
 637     va_start(ap, type);
 638     arg = va_arg(ap, void *);
 639     va_end(ap);
 640
 641     ret = ioctl(env->kvm_fd, type, arg);
 642     if (ret == -1)
 643         ret = -errno;
 644
 645     return ret;
 646 }
 647
 648 int kvm_has_sync_mmu(void)
 649 {
 650 #ifdef KVM_CAP_SYNC_MMU
 651     KVMState *s = kvm_state;
 652
 653     if (kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_SYNC_MMU) > 0)
 654         return 1;
 655 #endif
 656
 657     return 0;
 658 }