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     int irqchip_in_kernel;
  67     int pit_in_kernel;
  68 };
  69
  70 static KVMState *kvm_state;
  71
  72 static KVMSlot *kvm_alloc_slot(KVMState *s)
  73 {
  74     int i;
  75
  76     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  77         /* KVM private memory slots */
  78         if (i >= 8 && i < 12)
  79             continue;
  80         if (s->slots[i].memory_size == 0)
  81             return &s->slots[i];
  82     }
  83
  84     fprintf(stderr, "%s: no free slot available\n", __func__);
  85     abort();
  86 }
  87
  88 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  89                                          target_phys_addr_t start_addr,
  90                                          target_phys_addr_t end_addr)
  91 {
  92     int i;
  93
  94     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  95         KVMSlot *mem = &s->slots[i];
  96
  97         if (start_addr == mem->start_addr &&
  98             end_addr == mem->start_addr + mem->memory_size) {
  99             return mem;
 100         }
 101     }
 102
 103     return NULL;
 104 }
 105
 106 /*
 107  * Find overlapping slot with lowest start address
 108  */
 109 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 110                                             target_phys_addr_t start_addr,
 111                                             target_phys_addr_t end_addr)
 112 {
 113     KVMSlot *found = NULL;
 114     int i;
 115
 116     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 117         KVMSlot *mem = &s->slots[i];
 118
 119         if (mem->memory_size == 0 ||
 120             (found && found->start_addr < mem->start_addr)) {
 121             continue;
 122         }
 123
 124         if (end_addr > mem->start_addr &&
 125             start_addr < mem->start_addr + mem->memory_size) {
 126             found = mem;
 127         }
 128     }
 129
 130     return found;
 131 }
 132
 133 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 134 {
 135     struct kvm_userspace_memory_region mem;
 136
 137     mem.slot = slot->slot;
 138     mem.guest_phys_addr = slot->start_addr;
 139     mem.memory_size = slot->memory_size;
 140     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 141     mem.flags = slot->flags;
 142     if (s->migration_log) {
 143         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 144     }
 145     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 146 }
 147
 148 static void kvm_reset_vcpu(void *opaque)
 149 {
 150     CPUState *env = opaque;
 151
 152     if (kvm_arch_put_registers(env)) {
 153         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 154         abort();
 155     }
 156 }
 157
 158 int kvm_irqchip_in_kernel(void)
 159 {
 160     return kvm_state->irqchip_in_kernel;
 161 }
 162
 163 int kvm_pit_in_kernel(void)
 164 {
 165     return kvm_state->pit_in_kernel;
 166 }
 167
 168
 169 int kvm_init_vcpu(CPUState *env)
 170 {
 171     KVMState *s = kvm_state;
 172     long mmap_size;
 173     int ret;
 174
 175     dprintf("kvm_init_vcpu\n");
 176
 177     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 178     if (ret < 0) {
 179         dprintf("kvm_create_vcpu failed\n");
 180         goto err;
 181     }
 182
 183     env->kvm_fd = ret;
 184     env->kvm_state = s;
 185
 186     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 187     if (mmap_size < 0) {
 188         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 189         goto err;
 190     }
 191
 192     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 193                         env->kvm_fd, 0);
 194     if (env->kvm_run == MAP_FAILED) {
 195         ret = -errno;
 196         dprintf("mmap'ing vcpu state failed\n");
 197         goto err;
 198     }
 199
 200     ret = kvm_arch_init_vcpu(env);
 201     if (ret == 0) {
 202         qemu_register_reset(kvm_reset_vcpu, env);
 203         ret = kvm_arch_put_registers(env);
 204     }
 205 err:
 206     return ret;
 207 }
 208
 209 int kvm_put_mp_state(CPUState *env)
 210 {
 211     struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
 212
 213     return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
 214 }
 215
 216 int kvm_get_mp_state(CPUState *env)
 217 {
 218     struct kvm_mp_state mp_state;
 219     int ret;
 220
 221     ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
 222     if (ret < 0) {
 223         return ret;
 224     }
 225     env->mp_state = mp_state.mp_state;
 226     return 0;
 227 }
 228
 229 /*
 230  * dirty pages logging control
 231  */
 232 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 233                                       ram_addr_t size, int flags, int mask)
 234 {
 235     KVMState *s = kvm_state;
 236     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 237     int old_flags;
 238
 239     if (mem == NULL)  {
 240             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 241                     TARGET_FMT_plx "\n", __func__, phys_addr,
 242                     (target_phys_addr_t)(phys_addr + size - 1));
 243             return -EINVAL;
 244     }
 245
 246     old_flags = mem->flags;
 247
 248     flags = (mem->flags & ~mask) | flags;
 249     mem->flags = flags;
 250
 251     /* If nothing changed effectively, no need to issue ioctl */
 252     if (s->migration_log) {
 253         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 254     }
 255     if (flags == old_flags) {
 256             return 0;
 257     }
 258
 259     return kvm_set_user_memory_region(s, mem);
 260 }
 261
 262 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 263 {
 264         return kvm_dirty_pages_log_change(phys_addr, size,
 265                                           KVM_MEM_LOG_DIRTY_PAGES,
 266                                           KVM_MEM_LOG_DIRTY_PAGES);
 267 }
 268
 269 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 270 {
 271         return kvm_dirty_pages_log_change(phys_addr, size,
 272                                           0,
 273                                           KVM_MEM_LOG_DIRTY_PAGES);
 274 }
 275
 276 int kvm_set_migration_log(int enable)
 277 {
 278     KVMState *s = kvm_state;
 279     KVMSlot *mem;
 280     int i, err;
 281
 282     s->migration_log = enable;
 283
 284     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 285         mem = &s->slots[i];
 286
 287         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 288             continue;
 289         }
 290         err = kvm_set_user_memory_region(s, mem);
 291         if (err) {
 292             return err;
 293         }
 294     }
 295     return 0;
 296 }
 297
 298 static int test_le_bit(unsigned long nr, unsigned char *addr)
 299 {
 300     return (addr[nr >> 3] >> (nr & 7)) & 1;
 301 }
 302
 303 /**
 304  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 305  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 306  * This means all bits are set to dirty.
 307  *
 308  * @start_add: start of logged region.
 309  * @end_addr: end of logged region.
 310  */
 311 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 312                                    target_phys_addr_t end_addr)
 313 {
 314     KVMState *s = kvm_state;
 315     unsigned long size, allocated_size = 0;
 316     target_phys_addr_t phys_addr;
 317     ram_addr_t addr;
 318     KVMDirtyLog d;
 319     KVMSlot *mem;
 320     int ret = 0;
 321     int r;
 322
 323     d.dirty_bitmap = NULL;
 324     while (start_addr < end_addr) {
 325         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 326         if (mem == NULL) {
 327             break;
 328         }
 329
 330         /* We didn't activate dirty logging? Don't care then. */
 331         if(!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES)) {
 332             continue;
 333         }
 334
 335         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 336         if (!d.dirty_bitmap) {
 337             d.dirty_bitmap = qemu_malloc(size);
 338         } else if (size > allocated_size) {
 339             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 340         }
 341         allocated_size = size;
 342         memset(d.dirty_bitmap, 0, allocated_size);
 343
 344         d.slot = mem->slot;
 345
 346         r = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
 347         if (r == -EINVAL) {
 348             dprintf("ioctl failed %d\n", errno);
 349             ret = -1;
 350             break;
 351         }
 352
 353         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 354              phys_addr < mem->start_addr + mem->memory_size;
 355              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 356             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 357             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 358
 359             if (test_le_bit(nr, bitmap)) {
 360                 cpu_physical_memory_set_dirty(addr);
 361             } else if (r < 0) {
 362                 /* When our KVM implementation doesn't know about dirty logging
 363                  * we can just assume it's always dirty and be fine. */
 364                 cpu_physical_memory_set_dirty(addr);
 365             }
 366         }
 367         start_addr = phys_addr;
 368     }
 369     qemu_free(d.dirty_bitmap);
 370
 371     return ret;
 372 }
 373
 374 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 375 {
 376     int ret = -ENOSYS;
 377 #ifdef KVM_CAP_COALESCED_MMIO
 378     KVMState *s = kvm_state;
 379
 380     if (s->coalesced_mmio) {
 381         struct kvm_coalesced_mmio_zone zone;
 382
 383         zone.addr = start;
 384         zone.size = size;
 385
 386         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 387     }
 388 #endif
 389
 390     return ret;
 391 }
 392
 393 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 394 {
 395     int ret = -ENOSYS;
 396 #ifdef KVM_CAP_COALESCED_MMIO
 397     KVMState *s = kvm_state;
 398
 399     if (s->coalesced_mmio) {
 400         struct kvm_coalesced_mmio_zone zone;
 401
 402         zone.addr = start;
 403         zone.size = size;
 404
 405         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 406     }
 407 #endif
 408
 409     return ret;
 410 }
 411
 412 int kvm_check_extension(KVMState *s, unsigned int extension)
 413 {
 414     int ret;
 415
 416     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 417     if (ret < 0) {
 418         ret = 0;
 419     }
 420
 421     return ret;
 422 }
 423
 424 int kvm_init(int smp_cpus)
 425 {
 426     static const char upgrade_note[] =
 427         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 428         "(see http://sourceforge.net/projects/kvm).\n";
 429     KVMState *s;
 430     int ret;
 431     int i;
 432
 433     if (smp_cpus > 1) {
 434         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 435         return -EINVAL;
 436     }
 437
 438     s = qemu_mallocz(sizeof(KVMState));
 439
 440 #ifdef KVM_CAP_SET_GUEST_DEBUG
 441     TAILQ_INIT(&s->kvm_sw_breakpoints);
 442 #endif
 443     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 444         s->slots[i].slot = i;
 445
 446     s->vmfd = -1;
 447     s->fd = open("/dev/kvm", O_RDWR);
 448     if (s->fd == -1) {
 449         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 450         ret = -errno;
 451         goto err;
 452     }
 453
 454     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 455     if (ret < KVM_API_VERSION) {
 456         if (ret > 0)
 457             ret = -EINVAL;
 458         fprintf(stderr, "kvm version too old\n");
 459         goto err;
 460     }
 461
 462     if (ret > KVM_API_VERSION) {
 463         ret = -EINVAL;
 464         fprintf(stderr, "kvm version not supported\n");
 465         goto err;
 466     }
 467
 468     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 469     if (s->vmfd < 0)
 470         goto err;
 471
 472     /* initially, KVM allocated its own memory and we had to jump through
 473      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 474      * just use a user allocated buffer so we can use regular pages
 475      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 476      */
 477     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 478         ret = -EINVAL;
 479         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 480                 upgrade_note);
 481         goto err;
 482     }
 483
 484     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 485      * destroyed properly.  Since we rely on this capability, refuse to work
 486      * with any kernel without this capability. */
 487     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 488         ret = -EINVAL;
 489
 490         fprintf(stderr,
 491                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 492                 upgrade_note);
 493         goto err;
 494     }
 495
 496 #ifdef KVM_CAP_COALESCED_MMIO
 497     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 498 #else
 499     s->coalesced_mmio = 0;
 500 #endif
 501
 502     s->broken_set_mem_region = 1;
 503 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 504     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 505     if (ret > 0) {
 506         s->broken_set_mem_region = 0;
 507     }
 508 #endif
 509
 510     ret = kvm_arch_init(s, smp_cpus);
 511     if (ret < 0)
 512         goto err;
 513
 514     kvm_state = s;
 515
 516     return 0;
 517
 518 err:
 519     if (s) {
 520         if (s->vmfd != -1)
 521             close(s->vmfd);
 522         if (s->fd != -1)
 523             close(s->fd);
 524     }
 525     qemu_free(s);
 526
 527     return ret;
 528 }
 529
 530 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 531                          int direction, int size, uint32_t count)
 532 {
 533     int i;
 534     uint8_t *ptr = data;
 535
 536     for (i = 0; i < count; i++) {
 537         if (direction == KVM_EXIT_IO_IN) {
 538             switch (size) {
 539             case 1:
 540                 stb_p(ptr, cpu_inb(env, port));
 541                 break;
 542             case 2:
 543                 stw_p(ptr, cpu_inw(env, port));
 544                 break;
 545             case 4:
 546                 stl_p(ptr, cpu_inl(env, port));
 547                 break;
 548             }
 549         } else {
 550             switch (size) {
 551             case 1:
 552                 cpu_outb(env, port, ldub_p(ptr));
 553                 break;
 554             case 2:
 555                 cpu_outw(env, port, lduw_p(ptr));
 556                 break;
 557             case 4:
 558                 cpu_outl(env, port, ldl_p(ptr));
 559                 break;
 560             }
 561         }
 562
 563         ptr += size;
 564     }
 565
 566     return 1;
 567 }
 568
 569 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 570 {
 571 #ifdef KVM_CAP_COALESCED_MMIO
 572     KVMState *s = kvm_state;
 573     if (s->coalesced_mmio) {
 574         struct kvm_coalesced_mmio_ring *ring;
 575
 576         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 577         while (ring->first != ring->last) {
 578             struct kvm_coalesced_mmio *ent;
 579
 580             ent = &ring->coalesced_mmio[ring->first];
 581
 582             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 583             /* FIXME smp_wmb() */
 584             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 585         }
 586     }
 587 #endif
 588 }
 589
 590 int kvm_cpu_exec(CPUState *env)
 591 {
 592     struct kvm_run *run = env->kvm_run;
 593     int ret;
 594
 595     dprintf("kvm_cpu_exec()\n");
 596
 597     do {
 598         if (env->exit_request) {
 599             dprintf("interrupt exit requested\n");
 600             ret = 0;
 601             break;
 602         }
 603
 604         kvm_arch_pre_run(env, run);
 605         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 606         kvm_arch_post_run(env, run);
 607
 608         if (ret == -EINTR || ret == -EAGAIN) {
 609             dprintf("io window exit\n");
 610             ret = 0;
 611             break;
 612         }
 613
 614         if (ret < 0) {
 615             dprintf("kvm run failed %s\n", strerror(-ret));
 616             abort();
 617         }
 618
 619         kvm_run_coalesced_mmio(env, run);
 620
 621         ret = 0; /* exit loop */
 622         switch (run->exit_reason) {
 623         case KVM_EXIT_IO:
 624             dprintf("handle_io\n");
 625             ret = kvm_handle_io(env, run->io.port,
 626                                 (uint8_t *)run + run->io.data_offset,
 627                                 run->io.direction,
 628                                 run->io.size,
 629                                 run->io.count);
 630             break;
 631         case KVM_EXIT_MMIO:
 632             dprintf("handle_mmio\n");
 633             cpu_physical_memory_rw(run->mmio.phys_addr,
 634                                    run->mmio.data,
 635                                    run->mmio.len,
 636                                    run->mmio.is_write);
 637             ret = 1;
 638             break;
 639         case KVM_EXIT_IRQ_WINDOW_OPEN:
 640             dprintf("irq_window_open\n");
 641             break;
 642         case KVM_EXIT_SHUTDOWN:
 643             dprintf("shutdown\n");
 644             qemu_system_reset_request();
 645             ret = 1;
 646             break;
 647         case KVM_EXIT_UNKNOWN:
 648             dprintf("kvm_exit_unknown\n");
 649             break;
 650         case KVM_EXIT_FAIL_ENTRY:
 651             dprintf("kvm_exit_fail_entry\n");
 652             break;
 653         case KVM_EXIT_EXCEPTION:
 654             dprintf("kvm_exit_exception\n");
 655             break;
 656         case KVM_EXIT_DEBUG:
 657             dprintf("kvm_exit_debug\n");
 658 #ifdef KVM_CAP_SET_GUEST_DEBUG
 659             if (kvm_arch_debug(&run->debug.arch)) {
 660                 gdb_set_stop_cpu(env);
 661                 vm_stop(EXCP_DEBUG);
 662                 env->exception_index = EXCP_DEBUG;
 663                 return 0;
 664             }
 665             /* re-enter, this exception was guest-internal */
 666             ret = 1;
 667 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 668             break;
 669         default:
 670             dprintf("kvm_arch_handle_exit\n");
 671             ret = kvm_arch_handle_exit(env, run);
 672             break;
 673         }
 674     } while (ret > 0);
 675
 676     if (env->exit_request) {
 677         env->exit_request = 0;
 678         env->exception_index = EXCP_INTERRUPT;
 679     }
 680
 681     return ret;
 682 }
 683
 684 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 685                       ram_addr_t size,
 686                       ram_addr_t phys_offset)
 687 {
 688     KVMState *s = kvm_state;
 689     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 690     KVMSlot *mem, old;
 691     int err;
 692
 693     if (start_addr & ~TARGET_PAGE_MASK) {
 694         if (flags >= IO_MEM_UNASSIGNED) {
 695             if (!kvm_lookup_overlapping_slot(s, start_addr,
 696                                              start_addr + size)) {
 697                 return;
 698             }
 699             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 700         } else {
 701             fprintf(stderr, "Only page-aligned memory slots supported\n");
 702         }
 703         abort();
 704     }
 705
 706     /* KVM does not support read-only slots */
 707     phys_offset &= ~IO_MEM_ROM;
 708
 709     while (1) {
 710         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 711         if (!mem) {
 712             break;
 713         }
 714
 715         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 716             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 717             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 718             /* The new slot fits into the existing one and comes with
 719              * identical parameters - nothing to be done. */
 720             return;
 721         }
 722
 723         old = *mem;
 724
 725         /* unregister the overlapping slot */
 726         mem->memory_size = 0;
 727         err = kvm_set_user_memory_region(s, mem);
 728         if (err) {
 729             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 730                     __func__, strerror(-err));
 731             abort();
 732         }
 733
 734         /* Workaround for older KVM versions: we can't join slots, even not by
 735          * unregistering the previous ones and then registering the larger
 736          * slot. We have to maintain the existing fragmentation. Sigh.
 737          *
 738          * This workaround assumes that the new slot starts at the same
 739          * address as the first existing one. If not or if some overlapping
 740          * slot comes around later, we will fail (not seen in practice so far)
 741          * - and actually require a recent KVM version. */
 742         if (s->broken_set_mem_region &&
 743             old.start_addr == start_addr && old.memory_size < size &&
 744             flags < IO_MEM_UNASSIGNED) {
 745             mem = kvm_alloc_slot(s);
 746             mem->memory_size = old.memory_size;
 747             mem->start_addr = old.start_addr;
 748             mem->phys_offset = old.phys_offset;
 749             mem->flags = 0;
 750
 751             err = kvm_set_user_memory_region(s, mem);
 752             if (err) {
 753                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 754                         strerror(-err));
 755                 abort();
 756             }
 757
 758             start_addr += old.memory_size;
 759             phys_offset += old.memory_size;
 760             size -= old.memory_size;
 761             continue;
 762         }
 763
 764         /* register prefix slot */
 765         if (old.start_addr < start_addr) {
 766             mem = kvm_alloc_slot(s);
 767             mem->memory_size = start_addr - old.start_addr;
 768             mem->start_addr = old.start_addr;
 769             mem->phys_offset = old.phys_offset;
 770             mem->flags = 0;
 771
 772             err = kvm_set_user_memory_region(s, mem);
 773             if (err) {
 774                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 775                         __func__, strerror(-err));
 776                 abort();
 777             }
 778         }
 779
 780         /* register suffix slot */
 781         if (old.start_addr + old.memory_size > start_addr + size) {
 782             ram_addr_t size_delta;
 783
 784             mem = kvm_alloc_slot(s);
 785             mem->start_addr = start_addr + size;
 786             size_delta = mem->start_addr - old.start_addr;
 787             mem->memory_size = old.memory_size - size_delta;
 788             mem->phys_offset = old.phys_offset + size_delta;
 789             mem->flags = 0;
 790
 791             err = kvm_set_user_memory_region(s, mem);
 792             if (err) {
 793                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 794                         __func__, strerror(-err));
 795                 abort();
 796             }
 797         }
 798     }
 799
 800     /* in case the KVM bug workaround already "consumed" the new slot */
 801     if (!size)
 802         return;
 803
 804     /* KVM does not need to know about this memory */
 805     if (flags >= IO_MEM_UNASSIGNED)
 806         return;
 807
 808     mem = kvm_alloc_slot(s);
 809     mem->memory_size = size;
 810     mem->start_addr = start_addr;
 811     mem->phys_offset = phys_offset;
 812     mem->flags = 0;
 813
 814     err = kvm_set_user_memory_region(s, mem);
 815     if (err) {
 816         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 817                 strerror(-err));
 818         abort();
 819     }
 820 }
 821
 822 int kvm_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->fd, type, arg);
 833     if (ret == -1)
 834         ret = -errno;
 835
 836     return ret;
 837 }
 838
 839 int kvm_vm_ioctl(KVMState *s, 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(s->vmfd, type, arg);
 850     if (ret == -1)
 851         ret = -errno;
 852
 853     return ret;
 854 }
 855
 856 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 857 {
 858     int ret;
 859     void *arg;
 860     va_list ap;
 861
 862     va_start(ap, type);
 863     arg = va_arg(ap, void *);
 864     va_end(ap);
 865
 866     ret = ioctl(env->kvm_fd, type, arg);
 867     if (ret == -1)
 868         ret = -errno;
 869
 870     return ret;
 871 }
 872
 873 int kvm_has_sync_mmu(void)
 874 {
 875 #ifdef KVM_CAP_SYNC_MMU
 876     KVMState *s = kvm_state;
 877
 878     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 879 #else
 880     return 0;
 881 #endif
 882 }
 883
 884 void kvm_setup_guest_memory(void *start, size_t size)
 885 {
 886     if (!kvm_has_sync_mmu()) {
 887 #ifdef MADV_DONTFORK
 888         int ret = madvise(start, size, MADV_DONTFORK);
 889
 890         if (ret) {
 891             perror("madvice");
 892             exit(1);
 893         }
 894 #else
 895         fprintf(stderr,
 896                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 897         exit(1);
 898 #endif
 899     }
 900 }
 901
 902 #ifdef KVM_CAP_SET_GUEST_DEBUG
 903 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 904 {
 905     if (env == cpu_single_env) {
 906         func(data);
 907         return;
 908     }
 909     abort();
 910 }
 911
 912 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 913                                                  target_ulong pc)
 914 {
 915     struct kvm_sw_breakpoint *bp;
 916
 917     TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 918         if (bp->pc == pc)
 919             return bp;
 920     }
 921     return NULL;
 922 }
 923
 924 int kvm_sw_breakpoints_active(CPUState *env)
 925 {
 926     return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 927 }
 928
 929 struct kvm_set_guest_debug_data {
 930     struct kvm_guest_debug dbg;
 931     CPUState *env;
 932     int err;
 933 };
 934
 935 static void kvm_invoke_set_guest_debug(void *data)
 936 {
 937     struct kvm_set_guest_debug_data *dbg_data = data;
 938     dbg_data->err = kvm_vcpu_ioctl(dbg_data->env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 939 }
 940
 941 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 942 {
 943     struct kvm_set_guest_debug_data data;
 944
 945     data.dbg.control = 0;
 946     if (env->singlestep_enabled)
 947         data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 948
 949     kvm_arch_update_guest_debug(env, &data.dbg);
 950     data.dbg.control |= reinject_trap;
 951     data.env = env;
 952
 953     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
 954     return data.err;
 955 }
 956
 957 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 958                           target_ulong len, int type)
 959 {
 960     struct kvm_sw_breakpoint *bp;
 961     CPUState *env;
 962     int err;
 963
 964     if (type == GDB_BREAKPOINT_SW) {
 965         bp = kvm_find_sw_breakpoint(current_env, addr);
 966         if (bp) {
 967             bp->use_count++;
 968             return 0;
 969         }
 970
 971         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 972         if (!bp)
 973             return -ENOMEM;
 974
 975         bp->pc = addr;
 976         bp->use_count = 1;
 977         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 978         if (err) {
 979             free(bp);
 980             return err;
 981         }
 982
 983         TAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 984                           bp, entry);
 985     } else {
 986         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
 987         if (err)
 988             return err;
 989     }
 990
 991     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 992         err = kvm_update_guest_debug(env, 0);
 993         if (err)
 994             return err;
 995     }
 996     return 0;
 997 }
 998
 999 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1000                           target_ulong len, int type)
1001 {
1002     struct kvm_sw_breakpoint *bp;
1003     CPUState *env;
1004     int err;
1005
1006     if (type == GDB_BREAKPOINT_SW) {
1007         bp = kvm_find_sw_breakpoint(current_env, addr);
1008         if (!bp)
1009             return -ENOENT;
1010
1011         if (bp->use_count > 1) {
1012             bp->use_count--;
1013             return 0;
1014         }
1015
1016         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1017         if (err)
1018             return err;
1019
1020         TAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1021         qemu_free(bp);
1022     } else {
1023         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1024         if (err)
1025             return err;
1026     }
1027
1028     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1029         err = kvm_update_guest_debug(env, 0);
1030         if (err)
1031             return err;
1032     }
1033     return 0;
1034 }
1035
1036 void kvm_remove_all_breakpoints(CPUState *current_env)
1037 {
1038     struct kvm_sw_breakpoint *bp, *next;
1039     KVMState *s = current_env->kvm_state;
1040     CPUState *env;
1041
1042     TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1043         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1044             /* Try harder to find a CPU that currently sees the breakpoint. */
1045             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1046                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1047                     break;
1048             }
1049         }
1050     }
1051     kvm_arch_remove_all_hw_breakpoints();
1052
1053     for (env = first_cpu; env != NULL; env = env->next_cpu)
1054         kvm_update_guest_debug(env, 0);
1055 }
1056
1057 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1058
1059 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1060 {
1061     return -EINVAL;
1062 }
1063
1064 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1065                           target_ulong len, int type)
1066 {
1067     return -EINVAL;
1068 }
1069
1070 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1071                           target_ulong len, int type)
1072 {
1073     return -EINVAL;
1074 }
1075
1076 void kvm_remove_all_breakpoints(CPUState *current_env)
1077 {
1078 }
1079 #endif /* !KVM_CAP_SET_GUEST_DEBUG */