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
 147 int kvm_init_vcpu(CPUState *env)
 148 {
 149     KVMState *s = kvm_state;
 150     long mmap_size;
 151     int ret;
 152
 153     dprintf("kvm_init_vcpu\n");
 154
 155     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 156     if (ret < 0) {
 157         dprintf("kvm_create_vcpu failed\n");
 158         goto err;
 159     }
 160
 161     env->kvm_fd = ret;
 162     env->kvm_state = s;
 163
 164     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 165     if (mmap_size < 0) {
 166         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 167         goto err;
 168     }
 169
 170     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 171                         env->kvm_fd, 0);
 172     if (env->kvm_run == MAP_FAILED) {
 173         ret = -errno;
 174         dprintf("mmap'ing vcpu state failed\n");
 175         goto err;
 176     }
 177
 178     ret = kvm_arch_init_vcpu(env);
 179
 180 err:
 181     return ret;
 182 }
 183
 184 int kvm_put_mp_state(CPUState *env)
 185 {
 186     struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
 187
 188     return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
 189 }
 190
 191 int kvm_get_mp_state(CPUState *env)
 192 {
 193     struct kvm_mp_state mp_state;
 194     int ret;
 195
 196     ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
 197     if (ret < 0) {
 198         return ret;
 199     }
 200     env->mp_state = mp_state.mp_state;
 201     return 0;
 202 }
 203
 204 int kvm_sync_vcpus(void)
 205 {
 206     CPUState *env;
 207
 208     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 209         int ret;
 210
 211         ret = kvm_arch_put_registers(env);
 212         if (ret)
 213             return ret;
 214     }
 215
 216     return 0;
 217 }
 218
 219 /*
 220  * dirty pages logging control
 221  */
 222 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 223                                       ram_addr_t size, int flags, int mask)
 224 {
 225     KVMState *s = kvm_state;
 226     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 227     int old_flags;
 228
 229     if (mem == NULL)  {
 230             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 231                     TARGET_FMT_plx "\n", __func__, phys_addr,
 232                     phys_addr + size - 1);
 233             return -EINVAL;
 234     }
 235
 236     old_flags = mem->flags;
 237
 238     flags = (mem->flags & ~mask) | flags;
 239     mem->flags = flags;
 240
 241     /* If nothing changed effectively, no need to issue ioctl */
 242     if (s->migration_log) {
 243         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 244     }
 245     if (flags == old_flags) {
 246             return 0;
 247     }
 248
 249     return kvm_set_user_memory_region(s, mem);
 250 }
 251
 252 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 253 {
 254         return kvm_dirty_pages_log_change(phys_addr, size,
 255                                           KVM_MEM_LOG_DIRTY_PAGES,
 256                                           KVM_MEM_LOG_DIRTY_PAGES);
 257 }
 258
 259 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 260 {
 261         return kvm_dirty_pages_log_change(phys_addr, size,
 262                                           0,
 263                                           KVM_MEM_LOG_DIRTY_PAGES);
 264 }
 265
 266 int kvm_set_migration_log(int enable)
 267 {
 268     KVMState *s = kvm_state;
 269     KVMSlot *mem;
 270     int i, err;
 271
 272     s->migration_log = enable;
 273
 274     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 275         mem = &s->slots[i];
 276
 277         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 278             continue;
 279         }
 280         err = kvm_set_user_memory_region(s, mem);
 281         if (err) {
 282             return err;
 283         }
 284     }
 285     return 0;
 286 }
 287
 288 /**
 289  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 290  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 291  * This means all bits are set to dirty.
 292  *
 293  * @start_add: start of logged region.
 294  * @end_addr: end of logged region.
 295  */
 296 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 297                                    target_phys_addr_t end_addr)
 298 {
 299     KVMState *s = kvm_state;
 300     unsigned long size, allocated_size = 0;
 301     target_phys_addr_t phys_addr;
 302     ram_addr_t addr;
 303     KVMDirtyLog d;
 304     KVMSlot *mem;
 305     int ret = 0;
 306
 307     d.dirty_bitmap = NULL;
 308     while (start_addr < end_addr) {
 309         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 310         if (mem == NULL) {
 311             break;
 312         }
 313
 314         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 315         if (!d.dirty_bitmap) {
 316             d.dirty_bitmap = qemu_malloc(size);
 317         } else if (size > allocated_size) {
 318             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 319         }
 320         allocated_size = size;
 321         memset(d.dirty_bitmap, 0, allocated_size);
 322
 323         d.slot = mem->slot;
 324
 325         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 326             dprintf("ioctl failed %d\n", errno);
 327             ret = -1;
 328             break;
 329         }
 330
 331         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 332              phys_addr < mem->start_addr + mem->memory_size;
 333              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 334             unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
 335             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 336             unsigned word = nr / (sizeof(*bitmap) * 8);
 337             unsigned bit = nr % (sizeof(*bitmap) * 8);
 338
 339             if ((bitmap[word] >> bit) & 1) {
 340                 cpu_physical_memory_set_dirty(addr);
 341             }
 342         }
 343         start_addr = phys_addr;
 344     }
 345     qemu_free(d.dirty_bitmap);
 346
 347     return ret;
 348 }
 349
 350 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 351 {
 352     int ret = -ENOSYS;
 353 #ifdef KVM_CAP_COALESCED_MMIO
 354     KVMState *s = kvm_state;
 355
 356     if (s->coalesced_mmio) {
 357         struct kvm_coalesced_mmio_zone zone;
 358
 359         zone.addr = start;
 360         zone.size = size;
 361
 362         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 363     }
 364 #endif
 365
 366     return ret;
 367 }
 368
 369 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 370 {
 371     int ret = -ENOSYS;
 372 #ifdef KVM_CAP_COALESCED_MMIO
 373     KVMState *s = kvm_state;
 374
 375     if (s->coalesced_mmio) {
 376         struct kvm_coalesced_mmio_zone zone;
 377
 378         zone.addr = start;
 379         zone.size = size;
 380
 381         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 382     }
 383 #endif
 384
 385     return ret;
 386 }
 387
 388 int kvm_check_extension(KVMState *s, unsigned int extension)
 389 {
 390     int ret;
 391
 392     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 393     if (ret < 0) {
 394         ret = 0;
 395     }
 396
 397     return ret;
 398 }
 399
 400 static void kvm_reset_vcpus(void *opaque)
 401 {
 402     kvm_sync_vcpus();
 403 }
 404
 405 int kvm_init(int smp_cpus)
 406 {
 407     static const char upgrade_note[] =
 408         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 409         "(see http://sourceforge.net/projects/kvm).\n";
 410     KVMState *s;
 411     int ret;
 412     int i;
 413
 414     if (smp_cpus > 1) {
 415         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 416         return -EINVAL;
 417     }
 418
 419     s = qemu_mallocz(sizeof(KVMState));
 420
 421 #ifdef KVM_CAP_SET_GUEST_DEBUG
 422     TAILQ_INIT(&s->kvm_sw_breakpoints);
 423 #endif
 424     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 425         s->slots[i].slot = i;
 426
 427     s->vmfd = -1;
 428     s->fd = open("/dev/kvm", O_RDWR);
 429     if (s->fd == -1) {
 430         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 431         ret = -errno;
 432         goto err;
 433     }
 434
 435     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 436     if (ret < KVM_API_VERSION) {
 437         if (ret > 0)
 438             ret = -EINVAL;
 439         fprintf(stderr, "kvm version too old\n");
 440         goto err;
 441     }
 442
 443     if (ret > KVM_API_VERSION) {
 444         ret = -EINVAL;
 445         fprintf(stderr, "kvm version not supported\n");
 446         goto err;
 447     }
 448
 449     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 450     if (s->vmfd < 0)
 451         goto err;
 452
 453     /* initially, KVM allocated its own memory and we had to jump through
 454      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 455      * just use a user allocated buffer so we can use regular pages
 456      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 457      */
 458     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 459         ret = -EINVAL;
 460         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 461                 upgrade_note);
 462         goto err;
 463     }
 464
 465     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 466      * destroyed properly.  Since we rely on this capability, refuse to work
 467      * with any kernel without this capability. */
 468     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 469         ret = -EINVAL;
 470
 471         fprintf(stderr,
 472                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 473                 upgrade_note);
 474         goto err;
 475     }
 476
 477 #ifdef KVM_CAP_COALESCED_MMIO
 478     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 479 #else
 480     s->coalesced_mmio = 0;
 481 #endif
 482
 483     s->broken_set_mem_region = 1;
 484 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 485     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 486     if (ret > 0) {
 487         s->broken_set_mem_region = 0;
 488     }
 489 #endif
 490
 491     ret = kvm_arch_init(s, smp_cpus);
 492     if (ret < 0)
 493         goto err;
 494
 495     qemu_register_reset(kvm_reset_vcpus, INT_MAX, NULL);
 496
 497     kvm_state = s;
 498
 499     return 0;
 500
 501 err:
 502     if (s) {
 503         if (s->vmfd != -1)
 504             close(s->vmfd);
 505         if (s->fd != -1)
 506             close(s->fd);
 507     }
 508     qemu_free(s);
 509
 510     return ret;
 511 }
 512
 513 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 514                          int direction, int size, uint32_t count)
 515 {
 516     int i;
 517     uint8_t *ptr = data;
 518
 519     for (i = 0; i < count; i++) {
 520         if (direction == KVM_EXIT_IO_IN) {
 521             switch (size) {
 522             case 1:
 523                 stb_p(ptr, cpu_inb(env, port));
 524                 break;
 525             case 2:
 526                 stw_p(ptr, cpu_inw(env, port));
 527                 break;
 528             case 4:
 529                 stl_p(ptr, cpu_inl(env, port));
 530                 break;
 531             }
 532         } else {
 533             switch (size) {
 534             case 1:
 535                 cpu_outb(env, port, ldub_p(ptr));
 536                 break;
 537             case 2:
 538                 cpu_outw(env, port, lduw_p(ptr));
 539                 break;
 540             case 4:
 541                 cpu_outl(env, port, ldl_p(ptr));
 542                 break;
 543             }
 544         }
 545
 546         ptr += size;
 547     }
 548
 549     return 1;
 550 }
 551
 552 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 553 {
 554 #ifdef KVM_CAP_COALESCED_MMIO
 555     KVMState *s = kvm_state;
 556     if (s->coalesced_mmio) {
 557         struct kvm_coalesced_mmio_ring *ring;
 558
 559         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 560         while (ring->first != ring->last) {
 561             struct kvm_coalesced_mmio *ent;
 562
 563             ent = &ring->coalesced_mmio[ring->first];
 564
 565             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 566             /* FIXME smp_wmb() */
 567             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 568         }
 569     }
 570 #endif
 571 }
 572
 573 int kvm_cpu_exec(CPUState *env)
 574 {
 575     struct kvm_run *run = env->kvm_run;
 576     int ret;
 577
 578     dprintf("kvm_cpu_exec()\n");
 579
 580     do {
 581         if (env->exit_request) {
 582             dprintf("interrupt exit requested\n");
 583             ret = 0;
 584             break;
 585         }
 586
 587         kvm_arch_pre_run(env, run);
 588         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 589         kvm_arch_post_run(env, run);
 590
 591         if (ret == -EINTR || ret == -EAGAIN) {
 592             dprintf("io window exit\n");
 593             ret = 0;
 594             break;
 595         }
 596
 597         if (ret < 0) {
 598             dprintf("kvm run failed %s\n", strerror(-ret));
 599             abort();
 600         }
 601
 602         kvm_run_coalesced_mmio(env, run);
 603
 604         ret = 0; /* exit loop */
 605         switch (run->exit_reason) {
 606         case KVM_EXIT_IO:
 607             dprintf("handle_io\n");
 608             ret = kvm_handle_io(env, run->io.port,
 609                                 (uint8_t *)run + run->io.data_offset,
 610                                 run->io.direction,
 611                                 run->io.size,
 612                                 run->io.count);
 613             break;
 614         case KVM_EXIT_MMIO:
 615             dprintf("handle_mmio\n");
 616             cpu_physical_memory_rw(run->mmio.phys_addr,
 617                                    run->mmio.data,
 618                                    run->mmio.len,
 619                                    run->mmio.is_write);
 620             ret = 1;
 621             break;
 622         case KVM_EXIT_IRQ_WINDOW_OPEN:
 623             dprintf("irq_window_open\n");
 624             break;
 625         case KVM_EXIT_SHUTDOWN:
 626             dprintf("shutdown\n");
 627             qemu_system_reset_request();
 628             ret = 1;
 629             break;
 630         case KVM_EXIT_UNKNOWN:
 631             dprintf("kvm_exit_unknown\n");
 632             break;
 633         case KVM_EXIT_FAIL_ENTRY:
 634             dprintf("kvm_exit_fail_entry\n");
 635             break;
 636         case KVM_EXIT_EXCEPTION:
 637             dprintf("kvm_exit_exception\n");
 638             break;
 639         case KVM_EXIT_DEBUG:
 640             dprintf("kvm_exit_debug\n");
 641 #ifdef KVM_CAP_SET_GUEST_DEBUG
 642             if (kvm_arch_debug(&run->debug.arch)) {
 643                 gdb_set_stop_cpu(env);
 644                 vm_stop(EXCP_DEBUG);
 645                 env->exception_index = EXCP_DEBUG;
 646                 return 0;
 647             }
 648             /* re-enter, this exception was guest-internal */
 649             ret = 1;
 650 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 651             break;
 652         default:
 653             dprintf("kvm_arch_handle_exit\n");
 654             ret = kvm_arch_handle_exit(env, run);
 655             break;
 656         }
 657     } while (ret > 0);
 658
 659     if (env->exit_request) {
 660         env->exit_request = 0;
 661         env->exception_index = EXCP_INTERRUPT;
 662     }
 663
 664     return ret;
 665 }
 666
 667 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 668                       ram_addr_t size,
 669                       ram_addr_t phys_offset)
 670 {
 671     KVMState *s = kvm_state;
 672     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 673     KVMSlot *mem, old;
 674     int err;
 675
 676     if (start_addr & ~TARGET_PAGE_MASK) {
 677         if (flags >= IO_MEM_UNASSIGNED) {
 678             if (!kvm_lookup_overlapping_slot(s, start_addr,
 679                                              start_addr + size)) {
 680                 return;
 681             }
 682             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 683         } else {
 684             fprintf(stderr, "Only page-aligned memory slots supported\n");
 685         }
 686         abort();
 687     }
 688
 689     /* KVM does not support read-only slots */
 690     phys_offset &= ~IO_MEM_ROM;
 691
 692     while (1) {
 693         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 694         if (!mem) {
 695             break;
 696         }
 697
 698         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 699             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 700             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 701             /* The new slot fits into the existing one and comes with
 702              * identical parameters - nothing to be done. */
 703             return;
 704         }
 705
 706         old = *mem;
 707
 708         /* unregister the overlapping slot */
 709         mem->memory_size = 0;
 710         err = kvm_set_user_memory_region(s, mem);
 711         if (err) {
 712             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 713                     __func__, strerror(-err));
 714             abort();
 715         }
 716
 717         /* Workaround for older KVM versions: we can't join slots, even not by
 718          * unregistering the previous ones and then registering the larger
 719          * slot. We have to maintain the existing fragmentation. Sigh.
 720          *
 721          * This workaround assumes that the new slot starts at the same
 722          * address as the first existing one. If not or if some overlapping
 723          * slot comes around later, we will fail (not seen in practice so far)
 724          * - and actually require a recent KVM version. */
 725         if (s->broken_set_mem_region &&
 726             old.start_addr == start_addr && old.memory_size < size &&
 727             flags < IO_MEM_UNASSIGNED) {
 728             mem = kvm_alloc_slot(s);
 729             mem->memory_size = old.memory_size;
 730             mem->start_addr = old.start_addr;
 731             mem->phys_offset = old.phys_offset;
 732             mem->flags = 0;
 733
 734             err = kvm_set_user_memory_region(s, mem);
 735             if (err) {
 736                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 737                         strerror(-err));
 738                 abort();
 739             }
 740
 741             start_addr += old.memory_size;
 742             phys_offset += old.memory_size;
 743             size -= old.memory_size;
 744             continue;
 745         }
 746
 747         /* register prefix slot */
 748         if (old.start_addr < start_addr) {
 749             mem = kvm_alloc_slot(s);
 750             mem->memory_size = start_addr - old.start_addr;
 751             mem->start_addr = old.start_addr;
 752             mem->phys_offset = old.phys_offset;
 753             mem->flags = 0;
 754
 755             err = kvm_set_user_memory_region(s, mem);
 756             if (err) {
 757                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 758                         __func__, strerror(-err));
 759                 abort();
 760             }
 761         }
 762
 763         /* register suffix slot */
 764         if (old.start_addr + old.memory_size > start_addr + size) {
 765             ram_addr_t size_delta;
 766
 767             mem = kvm_alloc_slot(s);
 768             mem->start_addr = start_addr + size;
 769             size_delta = mem->start_addr - old.start_addr;
 770             mem->memory_size = old.memory_size - size_delta;
 771             mem->phys_offset = old.phys_offset + size_delta;
 772             mem->flags = 0;
 773
 774             err = kvm_set_user_memory_region(s, mem);
 775             if (err) {
 776                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 777                         __func__, strerror(-err));
 778                 abort();
 779             }
 780         }
 781     }
 782
 783     /* in case the KVM bug workaround already "consumed" the new slot */
 784     if (!size)
 785         return;
 786
 787     /* KVM does not need to know about this memory */
 788     if (flags >= IO_MEM_UNASSIGNED)
 789         return;
 790
 791     mem = kvm_alloc_slot(s);
 792     mem->memory_size = size;
 793     mem->start_addr = start_addr;
 794     mem->phys_offset = phys_offset;
 795     mem->flags = 0;
 796
 797     err = kvm_set_user_memory_region(s, mem);
 798     if (err) {
 799         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 800                 strerror(-err));
 801         abort();
 802     }
 803 }
 804
 805 int kvm_ioctl(KVMState *s, int type, ...)
 806 {
 807     int ret;
 808     void *arg;
 809     va_list ap;
 810
 811     va_start(ap, type);
 812     arg = va_arg(ap, void *);
 813     va_end(ap);
 814
 815     ret = ioctl(s->fd, type, arg);
 816     if (ret == -1)
 817         ret = -errno;
 818
 819     return ret;
 820 }
 821
 822 int kvm_vm_ioctl(KVMState *s, int type, ...)
 823 {
 824     int ret;
 825     void *arg;
 826     va_list ap;
 827
 828     va_start(ap, type);
 829     arg = va_arg(ap, void *);
 830     va_end(ap);
 831
 832     ret = ioctl(s->vmfd, type, arg);
 833     if (ret == -1)
 834         ret = -errno;
 835
 836     return ret;
 837 }
 838
 839 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 840 {
 841     int ret;
 842     void *arg;
 843     va_list ap;
 844
 845     va_start(ap, type);
 846     arg = va_arg(ap, void *);
 847     va_end(ap);
 848
 849     ret = ioctl(env->kvm_fd, type, arg);
 850     if (ret == -1)
 851         ret = -errno;
 852
 853     return ret;
 854 }
 855
 856 int kvm_has_sync_mmu(void)
 857 {
 858 #ifdef KVM_CAP_SYNC_MMU
 859     KVMState *s = kvm_state;
 860
 861     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 862 #else
 863     return 0;
 864 #endif
 865 }
 866
 867 void kvm_setup_guest_memory(void *start, size_t size)
 868 {
 869     if (!kvm_has_sync_mmu()) {
 870 #ifdef MADV_DONTFORK
 871         int ret = madvise(start, size, MADV_DONTFORK);
 872
 873         if (ret) {
 874             perror("madvice");
 875             exit(1);
 876         }
 877 #else
 878         fprintf(stderr,
 879                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 880         exit(1);
 881 #endif
 882     }
 883 }
 884
 885 #ifdef KVM_CAP_SET_GUEST_DEBUG
 886 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 887                                                  target_ulong pc)
 888 {
 889     struct kvm_sw_breakpoint *bp;
 890
 891     TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 892         if (bp->pc == pc)
 893             return bp;
 894     }
 895     return NULL;
 896 }
 897
 898 int kvm_sw_breakpoints_active(CPUState *env)
 899 {
 900     return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 901 }
 902
 903 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 904 {
 905     struct kvm_guest_debug dbg;
 906
 907     dbg.control = 0;
 908     if (env->singlestep_enabled)
 909         dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 910
 911     kvm_arch_update_guest_debug(env, &dbg);
 912     dbg.control |= reinject_trap;
 913
 914     return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
 915 }
 916
 917 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 918                           target_ulong len, int type)
 919 {
 920     struct kvm_sw_breakpoint *bp;
 921     CPUState *env;
 922     int err;
 923
 924     if (type == GDB_BREAKPOINT_SW) {
 925         bp = kvm_find_sw_breakpoint(current_env, addr);
 926         if (bp) {
 927             bp->use_count++;
 928             return 0;
 929         }
 930
 931         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 932         if (!bp)
 933             return -ENOMEM;
 934
 935         bp->pc = addr;
 936         bp->use_count = 1;
 937         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 938         if (err) {
 939             free(bp);
 940             return err;
 941         }
 942
 943         TAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 944                           bp, entry);
 945     } else {
 946         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
 947         if (err)
 948             return err;
 949     }
 950
 951     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 952         err = kvm_update_guest_debug(env, 0);
 953         if (err)
 954             return err;
 955     }
 956     return 0;
 957 }
 958
 959 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
 960                           target_ulong len, int type)
 961 {
 962     struct kvm_sw_breakpoint *bp;
 963     CPUState *env;
 964     int err;
 965
 966     if (type == GDB_BREAKPOINT_SW) {
 967         bp = kvm_find_sw_breakpoint(current_env, addr);
 968         if (!bp)
 969             return -ENOENT;
 970
 971         if (bp->use_count > 1) {
 972             bp->use_count--;
 973             return 0;
 974         }
 975
 976         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
 977         if (err)
 978             return err;
 979
 980         TAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
 981         qemu_free(bp);
 982     } else {
 983         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
 984         if (err)
 985             return err;
 986     }
 987
 988     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 989         err = kvm_update_guest_debug(env, 0);
 990         if (err)
 991             return err;
 992     }
 993     return 0;
 994 }
 995
 996 void kvm_remove_all_breakpoints(CPUState *current_env)
 997 {
 998     struct kvm_sw_breakpoint *bp, *next;
 999     KVMState *s = current_env->kvm_state;
1000     CPUState *env;
1001
1002     TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1003         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1004             /* Try harder to find a CPU that currently sees the breakpoint. */
1005             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1006                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1007                     break;
1008             }
1009         }
1010     }
1011     kvm_arch_remove_all_hw_breakpoints();
1012
1013     for (env = first_cpu; env != NULL; env = env->next_cpu)
1014         kvm_update_guest_debug(env, 0);
1015 }
1016
1017 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1018
1019 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1020 {
1021     return -EINVAL;
1022 }
1023
1024 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1025                           target_ulong len, int type)
1026 {
1027     return -EINVAL;
1028 }
1029
1030 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1031                           target_ulong len, int type)
1032 {
1033     return -EINVAL;
1034 }
1035
1036 void kvm_remove_all_breakpoints(CPUState *current_env)
1037 {
1038 }
1039 #endif /* !KVM_CAP_SET_GUEST_DEBUG */