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 "hw/hw.h"
  26 #include "gdbstub.h"
  27 #include "kvm.h"
  28
  29 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  30 #define PAGE_SIZE TARGET_PAGE_SIZE
  31
  32 //#define DEBUG_KVM
  33
  34 #ifdef DEBUG_KVM
  35 #define dprintf(fmt, ...) \
  36     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  37 #else
  38 #define dprintf(fmt, ...) \
  39     do { } while (0)
  40 #endif
  41
  42 typedef struct KVMSlot
  43 {
  44     target_phys_addr_t start_addr;
  45     ram_addr_t memory_size;
  46     ram_addr_t phys_offset;
  47     int slot;
  48     int flags;
  49 } KVMSlot;
  50
  51 typedef struct kvm_dirty_log KVMDirtyLog;
  52
  53 int kvm_allowed = 0;
  54
  55 struct KVMState
  56 {
  57     KVMSlot slots[32];
  58     int fd;
  59     int vmfd;
  60     int coalesced_mmio;
  61     int broken_set_mem_region;
  62     int migration_log;
  63 #ifdef KVM_CAP_SET_GUEST_DEBUG
  64     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  65 #endif
  66 };
  67
  68 static KVMState *kvm_state;
  69
  70 static KVMSlot *kvm_alloc_slot(KVMState *s)
  71 {
  72     int i;
  73
  74     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  75         /* KVM private memory slots */
  76         if (i >= 8 && i < 12)
  77             continue;
  78         if (s->slots[i].memory_size == 0)
  79             return &s->slots[i];
  80     }
  81
  82     fprintf(stderr, "%s: no free slot available\n", __func__);
  83     abort();
  84 }
  85
  86 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  87                                          target_phys_addr_t start_addr,
  88                                          target_phys_addr_t end_addr)
  89 {
  90     int i;
  91
  92     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  93         KVMSlot *mem = &s->slots[i];
  94
  95         if (start_addr == mem->start_addr &&
  96             end_addr == mem->start_addr + mem->memory_size) {
  97             return mem;
  98         }
  99     }
 100
 101     return NULL;
 102 }
 103
 104 /*
 105  * Find overlapping slot with lowest start address
 106  */
 107 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 108                                             target_phys_addr_t start_addr,
 109                                             target_phys_addr_t end_addr)
 110 {
 111     KVMSlot *found = NULL;
 112     int i;
 113
 114     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 115         KVMSlot *mem = &s->slots[i];
 116
 117         if (mem->memory_size == 0 ||
 118             (found && found->start_addr < mem->start_addr)) {
 119             continue;
 120         }
 121
 122         if (end_addr > mem->start_addr &&
 123             start_addr < mem->start_addr + mem->memory_size) {
 124             found = mem;
 125         }
 126     }
 127
 128     return found;
 129 }
 130
 131 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 132 {
 133     struct kvm_userspace_memory_region mem;
 134
 135     mem.slot = slot->slot;
 136     mem.guest_phys_addr = slot->start_addr;
 137     mem.memory_size = slot->memory_size;
 138     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 139     mem.flags = slot->flags;
 140     if (s->migration_log) {
 141         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 142     }
 143     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 144 }
 145
 146 static void kvm_reset_vcpu(void *opaque)
 147 {
 148     CPUState *env = opaque;
 149
 150     if (kvm_arch_put_registers(env)) {
 151         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 152         abort();
 153     }
 154 }
 155
 156 int kvm_init_vcpu(CPUState *env)
 157 {
 158     KVMState *s = kvm_state;
 159     long mmap_size;
 160     int ret;
 161
 162     dprintf("kvm_init_vcpu\n");
 163
 164     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 165     if (ret < 0) {
 166         dprintf("kvm_create_vcpu failed\n");
 167         goto err;
 168     }
 169
 170     env->kvm_fd = ret;
 171     env->kvm_state = s;
 172
 173     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 174     if (mmap_size < 0) {
 175         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 176         goto err;
 177     }
 178
 179     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 180                         env->kvm_fd, 0);
 181     if (env->kvm_run == MAP_FAILED) {
 182         ret = -errno;
 183         dprintf("mmap'ing vcpu state failed\n");
 184         goto err;
 185     }
 186
 187     ret = kvm_arch_init_vcpu(env);
 188     if (ret == 0) {
 189         qemu_register_reset(kvm_reset_vcpu, env);
 190         ret = kvm_arch_put_registers(env);
 191     }
 192 err:
 193     return ret;
 194 }
 195
 196 int kvm_put_mp_state(CPUState *env)
 197 {
 198     struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
 199
 200     return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
 201 }
 202
 203 int kvm_get_mp_state(CPUState *env)
 204 {
 205     struct kvm_mp_state mp_state;
 206     int ret;
 207
 208     ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
 209     if (ret < 0) {
 210         return ret;
 211     }
 212     env->mp_state = mp_state.mp_state;
 213     return 0;
 214 }
 215
 216 /*
 217  * dirty pages logging control
 218  */
 219 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 220                                       ram_addr_t size, int flags, int mask)
 221 {
 222     KVMState *s = kvm_state;
 223     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 224     int old_flags;
 225
 226     if (mem == NULL)  {
 227             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 228                     TARGET_FMT_plx "\n", __func__, phys_addr,
 229                     phys_addr + size - 1);
 230             return -EINVAL;
 231     }
 232
 233     old_flags = mem->flags;
 234
 235     flags = (mem->flags & ~mask) | flags;
 236     mem->flags = flags;
 237
 238     /* If nothing changed effectively, no need to issue ioctl */
 239     if (s->migration_log) {
 240         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 241     }
 242     if (flags == old_flags) {
 243             return 0;
 244     }
 245
 246     return kvm_set_user_memory_region(s, mem);
 247 }
 248
 249 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 250 {
 251         return kvm_dirty_pages_log_change(phys_addr, size,
 252                                           KVM_MEM_LOG_DIRTY_PAGES,
 253                                           KVM_MEM_LOG_DIRTY_PAGES);
 254 }
 255
 256 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 257 {
 258         return kvm_dirty_pages_log_change(phys_addr, size,
 259                                           0,
 260                                           KVM_MEM_LOG_DIRTY_PAGES);
 261 }
 262
 263 int kvm_set_migration_log(int enable)
 264 {
 265     KVMState *s = kvm_state;
 266     KVMSlot *mem;
 267     int i, err;
 268
 269     s->migration_log = enable;
 270
 271     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 272         mem = &s->slots[i];
 273
 274         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 275             continue;
 276         }
 277         err = kvm_set_user_memory_region(s, mem);
 278         if (err) {
 279             return err;
 280         }
 281     }
 282     return 0;
 283 }
 284
 285 /**
 286  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 287  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 288  * This means all bits are set to dirty.
 289  *
 290  * @start_add: start of logged region.
 291  * @end_addr: end of logged region.
 292  */
 293 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 294                                    target_phys_addr_t end_addr)
 295 {
 296     KVMState *s = kvm_state;
 297     unsigned long size, allocated_size = 0;
 298     target_phys_addr_t phys_addr;
 299     ram_addr_t addr;
 300     KVMDirtyLog d;
 301     KVMSlot *mem;
 302     int ret = 0;
 303
 304     d.dirty_bitmap = NULL;
 305     while (start_addr < end_addr) {
 306         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 307         if (mem == NULL) {
 308             break;
 309         }
 310
 311         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 312         if (!d.dirty_bitmap) {
 313             d.dirty_bitmap = qemu_malloc(size);
 314         } else if (size > allocated_size) {
 315             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 316         }
 317         allocated_size = size;
 318         memset(d.dirty_bitmap, 0, allocated_size);
 319
 320         d.slot = mem->slot;
 321
 322         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 323             dprintf("ioctl failed %d\n", errno);
 324             ret = -1;
 325             break;
 326         }
 327
 328         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 329              phys_addr < mem->start_addr + mem->memory_size;
 330              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 331             unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
 332             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 333             unsigned word = nr / (sizeof(*bitmap) * 8);
 334             unsigned bit = nr % (sizeof(*bitmap) * 8);
 335
 336             if ((bitmap[word] >> bit) & 1) {
 337                 cpu_physical_memory_set_dirty(addr);
 338             }
 339         }
 340         start_addr = phys_addr;
 341     }
 342     qemu_free(d.dirty_bitmap);
 343
 344     return ret;
 345 }
 346
 347 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 348 {
 349     int ret = -ENOSYS;
 350 #ifdef KVM_CAP_COALESCED_MMIO
 351     KVMState *s = kvm_state;
 352
 353     if (s->coalesced_mmio) {
 354         struct kvm_coalesced_mmio_zone zone;
 355
 356         zone.addr = start;
 357         zone.size = size;
 358
 359         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 360     }
 361 #endif
 362
 363     return ret;
 364 }
 365
 366 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 367 {
 368     int ret = -ENOSYS;
 369 #ifdef KVM_CAP_COALESCED_MMIO
 370     KVMState *s = kvm_state;
 371
 372     if (s->coalesced_mmio) {
 373         struct kvm_coalesced_mmio_zone zone;
 374
 375         zone.addr = start;
 376         zone.size = size;
 377
 378         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 379     }
 380 #endif
 381
 382     return ret;
 383 }
 384
 385 int kvm_check_extension(KVMState *s, unsigned int extension)
 386 {
 387     int ret;
 388
 389     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 390     if (ret < 0) {
 391         ret = 0;
 392     }
 393
 394     return ret;
 395 }
 396
 397 int kvm_init(int smp_cpus)
 398 {
 399     static const char upgrade_note[] =
 400         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 401         "(see http://sourceforge.net/projects/kvm).\n";
 402     KVMState *s;
 403     int ret;
 404     int i;
 405
 406     if (smp_cpus > 1) {
 407         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 408         return -EINVAL;
 409     }
 410
 411     s = qemu_mallocz(sizeof(KVMState));
 412
 413 #ifdef KVM_CAP_SET_GUEST_DEBUG
 414     TAILQ_INIT(&s->kvm_sw_breakpoints);
 415 #endif
 416     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 417         s->slots[i].slot = i;
 418
 419     s->vmfd = -1;
 420     s->fd = open("/dev/kvm", O_RDWR);
 421     if (s->fd == -1) {
 422         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 423         ret = -errno;
 424         goto err;
 425     }
 426
 427     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 428     if (ret < KVM_API_VERSION) {
 429         if (ret > 0)
 430             ret = -EINVAL;
 431         fprintf(stderr, "kvm version too old\n");
 432         goto err;
 433     }
 434
 435     if (ret > KVM_API_VERSION) {
 436         ret = -EINVAL;
 437         fprintf(stderr, "kvm version not supported\n");
 438         goto err;
 439     }
 440
 441     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 442     if (s->vmfd < 0)
 443         goto err;
 444
 445     /* initially, KVM allocated its own memory and we had to jump through
 446      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 447      * just use a user allocated buffer so we can use regular pages
 448      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 449      */
 450     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 451         ret = -EINVAL;
 452         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 453                 upgrade_note);
 454         goto err;
 455     }
 456
 457     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 458      * destroyed properly.  Since we rely on this capability, refuse to work
 459      * with any kernel without this capability. */
 460     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 461         ret = -EINVAL;
 462
 463         fprintf(stderr,
 464                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 465                 upgrade_note);
 466         goto err;
 467     }
 468
 469 #ifdef KVM_CAP_COALESCED_MMIO
 470     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 471 #else
 472     s->coalesced_mmio = 0;
 473 #endif
 474
 475     s->broken_set_mem_region = 1;
 476 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 477     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 478     if (ret > 0) {
 479         s->broken_set_mem_region = 0;
 480     }
 481 #endif
 482
 483     ret = kvm_arch_init(s, smp_cpus);
 484     if (ret < 0)
 485         goto err;
 486
 487     kvm_state = s;
 488
 489     return 0;
 490
 491 err:
 492     if (s) {
 493         if (s->vmfd != -1)
 494             close(s->vmfd);
 495         if (s->fd != -1)
 496             close(s->fd);
 497     }
 498     qemu_free(s);
 499
 500     return ret;
 501 }
 502
 503 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 504                          int direction, int size, uint32_t count)
 505 {
 506     int i;
 507     uint8_t *ptr = data;
 508
 509     for (i = 0; i < count; i++) {
 510         if (direction == KVM_EXIT_IO_IN) {
 511             switch (size) {
 512             case 1:
 513                 stb_p(ptr, cpu_inb(env, port));
 514                 break;
 515             case 2:
 516                 stw_p(ptr, cpu_inw(env, port));
 517                 break;
 518             case 4:
 519                 stl_p(ptr, cpu_inl(env, port));
 520                 break;
 521             }
 522         } else {
 523             switch (size) {
 524             case 1:
 525                 cpu_outb(env, port, ldub_p(ptr));
 526                 break;
 527             case 2:
 528                 cpu_outw(env, port, lduw_p(ptr));
 529                 break;
 530             case 4:
 531                 cpu_outl(env, port, ldl_p(ptr));
 532                 break;
 533             }
 534         }
 535
 536         ptr += size;
 537     }
 538
 539     return 1;
 540 }
 541
 542 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 543 {
 544 #ifdef KVM_CAP_COALESCED_MMIO
 545     KVMState *s = kvm_state;
 546     if (s->coalesced_mmio) {
 547         struct kvm_coalesced_mmio_ring *ring;
 548
 549         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 550         while (ring->first != ring->last) {
 551             struct kvm_coalesced_mmio *ent;
 552
 553             ent = &ring->coalesced_mmio[ring->first];
 554
 555             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 556             /* FIXME smp_wmb() */
 557             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 558         }
 559     }
 560 #endif
 561 }
 562
 563 int kvm_cpu_exec(CPUState *env)
 564 {
 565     struct kvm_run *run = env->kvm_run;
 566     int ret;
 567
 568     dprintf("kvm_cpu_exec()\n");
 569
 570     do {
 571         if (env->exit_request) {
 572             dprintf("interrupt exit requested\n");
 573             ret = 0;
 574             break;
 575         }
 576
 577         kvm_arch_pre_run(env, run);
 578         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 579         kvm_arch_post_run(env, run);
 580
 581         if (ret == -EINTR || ret == -EAGAIN) {
 582             dprintf("io window exit\n");
 583             ret = 0;
 584             break;
 585         }
 586
 587         if (ret < 0) {
 588             dprintf("kvm run failed %s\n", strerror(-ret));
 589             abort();
 590         }
 591
 592         kvm_run_coalesced_mmio(env, run);
 593
 594         ret = 0; /* exit loop */
 595         switch (run->exit_reason) {
 596         case KVM_EXIT_IO:
 597             dprintf("handle_io\n");
 598             ret = kvm_handle_io(env, run->io.port,
 599                                 (uint8_t *)run + run->io.data_offset,
 600                                 run->io.direction,
 601                                 run->io.size,
 602                                 run->io.count);
 603             break;
 604         case KVM_EXIT_MMIO:
 605             dprintf("handle_mmio\n");
 606             cpu_physical_memory_rw(run->mmio.phys_addr,
 607                                    run->mmio.data,
 608                                    run->mmio.len,
 609                                    run->mmio.is_write);
 610             ret = 1;
 611             break;
 612         case KVM_EXIT_IRQ_WINDOW_OPEN:
 613             dprintf("irq_window_open\n");
 614             break;
 615         case KVM_EXIT_SHUTDOWN:
 616             dprintf("shutdown\n");
 617             qemu_system_reset_request();
 618             ret = 1;
 619             break;
 620         case KVM_EXIT_UNKNOWN:
 621             dprintf("kvm_exit_unknown\n");
 622             break;
 623         case KVM_EXIT_FAIL_ENTRY:
 624             dprintf("kvm_exit_fail_entry\n");
 625             break;
 626         case KVM_EXIT_EXCEPTION:
 627             dprintf("kvm_exit_exception\n");
 628             break;
 629         case KVM_EXIT_DEBUG:
 630             dprintf("kvm_exit_debug\n");
 631 #ifdef KVM_CAP_SET_GUEST_DEBUG
 632             if (kvm_arch_debug(&run->debug.arch)) {
 633                 gdb_set_stop_cpu(env);
 634                 vm_stop(EXCP_DEBUG);
 635                 env->exception_index = EXCP_DEBUG;
 636                 return 0;
 637             }
 638             /* re-enter, this exception was guest-internal */
 639             ret = 1;
 640 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 641             break;
 642         default:
 643             dprintf("kvm_arch_handle_exit\n");
 644             ret = kvm_arch_handle_exit(env, run);
 645             break;
 646         }
 647     } while (ret > 0);
 648
 649     if (env->exit_request) {
 650         env->exit_request = 0;
 651         env->exception_index = EXCP_INTERRUPT;
 652     }
 653
 654     return ret;
 655 }
 656
 657 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 658                       ram_addr_t size,
 659                       ram_addr_t phys_offset)
 660 {
 661     KVMState *s = kvm_state;
 662     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 663     KVMSlot *mem, old;
 664     int err;
 665
 666     if (start_addr & ~TARGET_PAGE_MASK) {
 667         if (flags >= IO_MEM_UNASSIGNED) {
 668             if (!kvm_lookup_overlapping_slot(s, start_addr,
 669                                              start_addr + size)) {
 670                 return;
 671             }
 672             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 673         } else {
 674             fprintf(stderr, "Only page-aligned memory slots supported\n");
 675         }
 676         abort();
 677     }
 678
 679     /* KVM does not support read-only slots */
 680     phys_offset &= ~IO_MEM_ROM;
 681
 682     while (1) {
 683         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 684         if (!mem) {
 685             break;
 686         }
 687
 688         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 689             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 690             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 691             /* The new slot fits into the existing one and comes with
 692              * identical parameters - nothing to be done. */
 693             return;
 694         }
 695
 696         old = *mem;
 697
 698         /* unregister the overlapping slot */
 699         mem->memory_size = 0;
 700         err = kvm_set_user_memory_region(s, mem);
 701         if (err) {
 702             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 703                     __func__, strerror(-err));
 704             abort();
 705         }
 706
 707         /* Workaround for older KVM versions: we can't join slots, even not by
 708          * unregistering the previous ones and then registering the larger
 709          * slot. We have to maintain the existing fragmentation. Sigh.
 710          *
 711          * This workaround assumes that the new slot starts at the same
 712          * address as the first existing one. If not or if some overlapping
 713          * slot comes around later, we will fail (not seen in practice so far)
 714          * - and actually require a recent KVM version. */
 715         if (s->broken_set_mem_region &&
 716             old.start_addr == start_addr && old.memory_size < size &&
 717             flags < IO_MEM_UNASSIGNED) {
 718             mem = kvm_alloc_slot(s);
 719             mem->memory_size = old.memory_size;
 720             mem->start_addr = old.start_addr;
 721             mem->phys_offset = old.phys_offset;
 722             mem->flags = 0;
 723
 724             err = kvm_set_user_memory_region(s, mem);
 725             if (err) {
 726                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 727                         strerror(-err));
 728                 abort();
 729             }
 730
 731             start_addr += old.memory_size;
 732             phys_offset += old.memory_size;
 733             size -= old.memory_size;
 734             continue;
 735         }
 736
 737         /* register prefix slot */
 738         if (old.start_addr < start_addr) {
 739             mem = kvm_alloc_slot(s);
 740             mem->memory_size = start_addr - old.start_addr;
 741             mem->start_addr = old.start_addr;
 742             mem->phys_offset = old.phys_offset;
 743             mem->flags = 0;
 744
 745             err = kvm_set_user_memory_region(s, mem);
 746             if (err) {
 747                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 748                         __func__, strerror(-err));
 749                 abort();
 750             }
 751         }
 752
 753         /* register suffix slot */
 754         if (old.start_addr + old.memory_size > start_addr + size) {
 755             ram_addr_t size_delta;
 756
 757             mem = kvm_alloc_slot(s);
 758             mem->start_addr = start_addr + size;
 759             size_delta = mem->start_addr - old.start_addr;
 760             mem->memory_size = old.memory_size - size_delta;
 761             mem->phys_offset = old.phys_offset + size_delta;
 762             mem->flags = 0;
 763
 764             err = kvm_set_user_memory_region(s, mem);
 765             if (err) {
 766                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 767                         __func__, strerror(-err));
 768                 abort();
 769             }
 770         }
 771     }
 772
 773     /* in case the KVM bug workaround already "consumed" the new slot */
 774     if (!size)
 775         return;
 776
 777     /* KVM does not need to know about this memory */
 778     if (flags >= IO_MEM_UNASSIGNED)
 779         return;
 780
 781     mem = kvm_alloc_slot(s);
 782     mem->memory_size = size;
 783     mem->start_addr = start_addr;
 784     mem->phys_offset = phys_offset;
 785     mem->flags = 0;
 786
 787     err = kvm_set_user_memory_region(s, mem);
 788     if (err) {
 789         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 790                 strerror(-err));
 791         abort();
 792     }
 793 }
 794
 795 int kvm_ioctl(KVMState *s, int type, ...)
 796 {
 797     int ret;
 798     void *arg;
 799     va_list ap;
 800
 801     va_start(ap, type);
 802     arg = va_arg(ap, void *);
 803     va_end(ap);
 804
 805     ret = ioctl(s->fd, type, arg);
 806     if (ret == -1)
 807         ret = -errno;
 808
 809     return ret;
 810 }
 811
 812 int kvm_vm_ioctl(KVMState *s, int type, ...)
 813 {
 814     int ret;
 815     void *arg;
 816     va_list ap;
 817
 818     va_start(ap, type);
 819     arg = va_arg(ap, void *);
 820     va_end(ap);
 821
 822     ret = ioctl(s->vmfd, type, arg);
 823     if (ret == -1)
 824         ret = -errno;
 825
 826     return ret;
 827 }
 828
 829 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 830 {
 831     int ret;
 832     void *arg;
 833     va_list ap;
 834
 835     va_start(ap, type);
 836     arg = va_arg(ap, void *);
 837     va_end(ap);
 838
 839     ret = ioctl(env->kvm_fd, type, arg);
 840     if (ret == -1)
 841         ret = -errno;
 842
 843     return ret;
 844 }
 845
 846 int kvm_has_sync_mmu(void)
 847 {
 848 #ifdef KVM_CAP_SYNC_MMU
 849     KVMState *s = kvm_state;
 850
 851     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 852 #else
 853     return 0;
 854 #endif
 855 }
 856
 857 void kvm_setup_guest_memory(void *start, size_t size)
 858 {
 859     if (!kvm_has_sync_mmu()) {
 860 #ifdef MADV_DONTFORK
 861         int ret = madvise(start, size, MADV_DONTFORK);
 862
 863         if (ret) {
 864             perror("madvice");
 865             exit(1);
 866         }
 867 #else
 868         fprintf(stderr,
 869                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 870         exit(1);
 871 #endif
 872     }
 873 }
 874
 875 #ifdef KVM_CAP_SET_GUEST_DEBUG
 876 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 877                                                  target_ulong pc)
 878 {
 879     struct kvm_sw_breakpoint *bp;
 880
 881     TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 882         if (bp->pc == pc)
 883             return bp;
 884     }
 885     return NULL;
 886 }
 887
 888 int kvm_sw_breakpoints_active(CPUState *env)
 889 {
 890     return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 891 }
 892
 893 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 894 {
 895     struct kvm_guest_debug dbg;
 896
 897     dbg.control = 0;
 898     if (env->singlestep_enabled)
 899         dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 900
 901     kvm_arch_update_guest_debug(env, &dbg);
 902     dbg.control |= reinject_trap;
 903
 904     return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
 905 }
 906
 907 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 908                           target_ulong len, int type)
 909 {
 910     struct kvm_sw_breakpoint *bp;
 911     CPUState *env;
 912     int err;
 913
 914     if (type == GDB_BREAKPOINT_SW) {
 915         bp = kvm_find_sw_breakpoint(current_env, addr);
 916         if (bp) {
 917             bp->use_count++;
 918             return 0;
 919         }
 920
 921         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 922         if (!bp)
 923             return -ENOMEM;
 924
 925         bp->pc = addr;
 926         bp->use_count = 1;
 927         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 928         if (err) {
 929             free(bp);
 930             return err;
 931         }
 932
 933         TAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 934                           bp, entry);
 935     } else {
 936         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
 937         if (err)
 938             return err;
 939     }
 940
 941     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 942         err = kvm_update_guest_debug(env, 0);
 943         if (err)
 944             return err;
 945     }
 946     return 0;
 947 }
 948
 949 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
 950                           target_ulong len, int type)
 951 {
 952     struct kvm_sw_breakpoint *bp;
 953     CPUState *env;
 954     int err;
 955
 956     if (type == GDB_BREAKPOINT_SW) {
 957         bp = kvm_find_sw_breakpoint(current_env, addr);
 958         if (!bp)
 959             return -ENOENT;
 960
 961         if (bp->use_count > 1) {
 962             bp->use_count--;
 963             return 0;
 964         }
 965
 966         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
 967         if (err)
 968             return err;
 969
 970         TAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
 971         qemu_free(bp);
 972     } else {
 973         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
 974         if (err)
 975             return err;
 976     }
 977
 978     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 979         err = kvm_update_guest_debug(env, 0);
 980         if (err)
 981             return err;
 982     }
 983     return 0;
 984 }
 985
 986 void kvm_remove_all_breakpoints(CPUState *current_env)
 987 {
 988     struct kvm_sw_breakpoint *bp, *next;
 989     KVMState *s = current_env->kvm_state;
 990     CPUState *env;
 991
 992     TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
 993         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
 994             /* Try harder to find a CPU that currently sees the breakpoint. */
 995             for (env = first_cpu; env != NULL; env = env->next_cpu) {
 996                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
 997                     break;
 998             }
 999         }
1000     }
1001     kvm_arch_remove_all_hw_breakpoints();
1002
1003     for (env = first_cpu; env != NULL; env = env->next_cpu)
1004         kvm_update_guest_debug(env, 0);
1005 }
1006
1007 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1008
1009 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1010 {
1011     return -EINVAL;
1012 }
1013
1014 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1015                           target_ulong len, int type)
1016 {
1017     return -EINVAL;
1018 }
1019
1020 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1021                           target_ulong len, int type)
1022 {
1023     return -EINVAL;
1024 }
1025
1026 void kvm_remove_all_breakpoints(CPUState *current_env)
1027 {
1028 }
1029 #endif /* !KVM_CAP_SET_GUEST_DEBUG */