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 regs_modified;
  61     int coalesced_mmio;
  62     int broken_set_mem_region;
  63     int migration_log;
  64 #ifdef KVM_CAP_SET_GUEST_DEBUG
  65     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  66 #endif
  67     int irqchip_in_kernel;
  68     int pit_in_kernel;
  69 };
  70
  71 static KVMState *kvm_state;
  72
  73 static KVMSlot *kvm_alloc_slot(KVMState *s)
  74 {
  75     int i;
  76
  77     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  78         /* KVM private memory slots */
  79         if (i >= 8 && i < 12)
  80             continue;
  81         if (s->slots[i].memory_size == 0)
  82             return &s->slots[i];
  83     }
  84
  85     fprintf(stderr, "%s: no free slot available\n", __func__);
  86     abort();
  87 }
  88
  89 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  90                                          target_phys_addr_t start_addr,
  91                                          target_phys_addr_t end_addr)
  92 {
  93     int i;
  94
  95     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  96         KVMSlot *mem = &s->slots[i];
  97
  98         if (start_addr == mem->start_addr &&
  99             end_addr == mem->start_addr + mem->memory_size) {
 100             return mem;
 101         }
 102     }
 103
 104     return NULL;
 105 }
 106
 107 /*
 108  * Find overlapping slot with lowest start address
 109  */
 110 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 111                                             target_phys_addr_t start_addr,
 112                                             target_phys_addr_t end_addr)
 113 {
 114     KVMSlot *found = NULL;
 115     int i;
 116
 117     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 118         KVMSlot *mem = &s->slots[i];
 119
 120         if (mem->memory_size == 0 ||
 121             (found && found->start_addr < mem->start_addr)) {
 122             continue;
 123         }
 124
 125         if (end_addr > mem->start_addr &&
 126             start_addr < mem->start_addr + mem->memory_size) {
 127             found = mem;
 128         }
 129     }
 130
 131     return found;
 132 }
 133
 134 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 135 {
 136     struct kvm_userspace_memory_region mem;
 137
 138     mem.slot = slot->slot;
 139     mem.guest_phys_addr = slot->start_addr;
 140     mem.memory_size = slot->memory_size;
 141     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 142     mem.flags = slot->flags;
 143     if (s->migration_log) {
 144         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 145     }
 146     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 147 }
 148
 149 static void kvm_reset_vcpu(void *opaque)
 150 {
 151     CPUState *env = opaque;
 152
 153     kvm_arch_reset_vcpu(env);
 154     if (kvm_arch_put_registers(env)) {
 155         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 156         abort();
 157     }
 158 }
 159
 160 int kvm_irqchip_in_kernel(void)
 161 {
 162     return kvm_state->irqchip_in_kernel;
 163 }
 164
 165 int kvm_pit_in_kernel(void)
 166 {
 167     return kvm_state->pit_in_kernel;
 168 }
 169
 170
 171 int kvm_init_vcpu(CPUState *env)
 172 {
 173     KVMState *s = kvm_state;
 174     long mmap_size;
 175     int ret;
 176
 177     dprintf("kvm_init_vcpu\n");
 178
 179     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 180     if (ret < 0) {
 181         dprintf("kvm_create_vcpu failed\n");
 182         goto err;
 183     }
 184
 185     env->kvm_fd = ret;
 186     env->kvm_state = s;
 187
 188     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 189     if (mmap_size < 0) {
 190         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 191         goto err;
 192     }
 193
 194     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 195                         env->kvm_fd, 0);
 196     if (env->kvm_run == MAP_FAILED) {
 197         ret = -errno;
 198         dprintf("mmap'ing vcpu state failed\n");
 199         goto err;
 200     }
 201
 202     ret = kvm_arch_init_vcpu(env);
 203     if (ret == 0) {
 204         qemu_register_reset(kvm_reset_vcpu, env);
 205         kvm_arch_reset_vcpu(env);
 206         ret = kvm_arch_put_registers(env);
 207     }
 208 err:
 209     return ret;
 210 }
 211
 212 /*
 213  * dirty pages logging control
 214  */
 215 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 216                                       ram_addr_t size, int flags, int mask)
 217 {
 218     KVMState *s = kvm_state;
 219     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 220     int old_flags;
 221
 222     if (mem == NULL)  {
 223             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 224                     TARGET_FMT_plx "\n", __func__, phys_addr,
 225                     (target_phys_addr_t)(phys_addr + size - 1));
 226             return -EINVAL;
 227     }
 228
 229     old_flags = mem->flags;
 230
 231     flags = (mem->flags & ~mask) | flags;
 232     mem->flags = flags;
 233
 234     /* If nothing changed effectively, no need to issue ioctl */
 235     if (s->migration_log) {
 236         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 237     }
 238     if (flags == old_flags) {
 239             return 0;
 240     }
 241
 242     return kvm_set_user_memory_region(s, mem);
 243 }
 244
 245 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 246 {
 247         return kvm_dirty_pages_log_change(phys_addr, size,
 248                                           KVM_MEM_LOG_DIRTY_PAGES,
 249                                           KVM_MEM_LOG_DIRTY_PAGES);
 250 }
 251
 252 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 253 {
 254         return kvm_dirty_pages_log_change(phys_addr, size,
 255                                           0,
 256                                           KVM_MEM_LOG_DIRTY_PAGES);
 257 }
 258
 259 int kvm_set_migration_log(int enable)
 260 {
 261     KVMState *s = kvm_state;
 262     KVMSlot *mem;
 263     int i, err;
 264
 265     s->migration_log = enable;
 266
 267     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 268         mem = &s->slots[i];
 269
 270         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 271             continue;
 272         }
 273         err = kvm_set_user_memory_region(s, mem);
 274         if (err) {
 275             return err;
 276         }
 277     }
 278     return 0;
 279 }
 280
 281 static int test_le_bit(unsigned long nr, unsigned char *addr)
 282 {
 283     return (addr[nr >> 3] >> (nr & 7)) & 1;
 284 }
 285
 286 /**
 287  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 288  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 289  * This means all bits are set to dirty.
 290  *
 291  * @start_add: start of logged region.
 292  * @end_addr: end of logged region.
 293  */
 294 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 295                                    target_phys_addr_t end_addr)
 296 {
 297     KVMState *s = kvm_state;
 298     unsigned long size, allocated_size = 0;
 299     target_phys_addr_t phys_addr;
 300     ram_addr_t addr;
 301     KVMDirtyLog d;
 302     KVMSlot *mem;
 303     int ret = 0;
 304
 305     d.dirty_bitmap = NULL;
 306     while (start_addr < end_addr) {
 307         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 308         if (mem == NULL) {
 309             break;
 310         }
 311
 312         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 313         if (!d.dirty_bitmap) {
 314             d.dirty_bitmap = qemu_malloc(size);
 315         } else if (size > allocated_size) {
 316             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 317         }
 318         allocated_size = size;
 319         memset(d.dirty_bitmap, 0, allocated_size);
 320
 321         d.slot = mem->slot;
 322
 323         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 324             dprintf("ioctl failed %d\n", errno);
 325             ret = -1;
 326             break;
 327         }
 328
 329         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 330              phys_addr < mem->start_addr + mem->memory_size;
 331              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 332             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 333             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 334
 335             if (test_le_bit(nr, bitmap)) {
 336                 cpu_physical_memory_set_dirty(addr);
 337             }
 338         }
 339         start_addr = phys_addr;
 340     }
 341     qemu_free(d.dirty_bitmap);
 342
 343     return ret;
 344 }
 345
 346 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 347 {
 348     int ret = -ENOSYS;
 349 #ifdef KVM_CAP_COALESCED_MMIO
 350     KVMState *s = kvm_state;
 351
 352     if (s->coalesced_mmio) {
 353         struct kvm_coalesced_mmio_zone zone;
 354
 355         zone.addr = start;
 356         zone.size = size;
 357
 358         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 359     }
 360 #endif
 361
 362     return ret;
 363 }
 364
 365 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 366 {
 367     int ret = -ENOSYS;
 368 #ifdef KVM_CAP_COALESCED_MMIO
 369     KVMState *s = kvm_state;
 370
 371     if (s->coalesced_mmio) {
 372         struct kvm_coalesced_mmio_zone zone;
 373
 374         zone.addr = start;
 375         zone.size = size;
 376
 377         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 378     }
 379 #endif
 380
 381     return ret;
 382 }
 383
 384 int kvm_check_extension(KVMState *s, unsigned int extension)
 385 {
 386     int ret;
 387
 388     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 389     if (ret < 0) {
 390         ret = 0;
 391     }
 392
 393     return ret;
 394 }
 395
 396 int kvm_init(int smp_cpus)
 397 {
 398     static const char upgrade_note[] =
 399         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 400         "(see http://sourceforge.net/projects/kvm).\n";
 401     KVMState *s;
 402     int ret;
 403     int i;
 404
 405     if (smp_cpus > 1) {
 406         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 407         return -EINVAL;
 408     }
 409
 410     s = qemu_mallocz(sizeof(KVMState));
 411
 412 #ifdef KVM_CAP_SET_GUEST_DEBUG
 413     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 414 #endif
 415     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 416         s->slots[i].slot = i;
 417
 418     s->vmfd = -1;
 419     s->fd = open("/dev/kvm", O_RDWR);
 420     if (s->fd == -1) {
 421         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 422         ret = -errno;
 423         goto err;
 424     }
 425
 426     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 427     if (ret < KVM_API_VERSION) {
 428         if (ret > 0)
 429             ret = -EINVAL;
 430         fprintf(stderr, "kvm version too old\n");
 431         goto err;
 432     }
 433
 434     if (ret > KVM_API_VERSION) {
 435         ret = -EINVAL;
 436         fprintf(stderr, "kvm version not supported\n");
 437         goto err;
 438     }
 439
 440     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 441     if (s->vmfd < 0)
 442         goto err;
 443
 444     /* initially, KVM allocated its own memory and we had to jump through
 445      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 446      * just use a user allocated buffer so we can use regular pages
 447      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 448      */
 449     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 450         ret = -EINVAL;
 451         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 452                 upgrade_note);
 453         goto err;
 454     }
 455
 456     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 457      * destroyed properly.  Since we rely on this capability, refuse to work
 458      * with any kernel without this capability. */
 459     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 460         ret = -EINVAL;
 461
 462         fprintf(stderr,
 463                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 464                 upgrade_note);
 465         goto err;
 466     }
 467
 468 #ifdef KVM_CAP_COALESCED_MMIO
 469     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 470 #else
 471     s->coalesced_mmio = 0;
 472 #endif
 473
 474     s->broken_set_mem_region = 1;
 475 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 476     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 477     if (ret > 0) {
 478         s->broken_set_mem_region = 0;
 479     }
 480 #endif
 481
 482     ret = kvm_arch_init(s, smp_cpus);
 483     if (ret < 0)
 484         goto err;
 485
 486     kvm_state = s;
 487
 488     return 0;
 489
 490 err:
 491     if (s) {
 492         if (s->vmfd != -1)
 493             close(s->vmfd);
 494         if (s->fd != -1)
 495             close(s->fd);
 496     }
 497     qemu_free(s);
 498
 499     return ret;
 500 }
 501
 502 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 503                          uint32_t count)
 504 {
 505     int i;
 506     uint8_t *ptr = data;
 507
 508     for (i = 0; i < count; i++) {
 509         if (direction == KVM_EXIT_IO_IN) {
 510             switch (size) {
 511             case 1:
 512                 stb_p(ptr, cpu_inb(port));
 513                 break;
 514             case 2:
 515                 stw_p(ptr, cpu_inw(port));
 516                 break;
 517             case 4:
 518                 stl_p(ptr, cpu_inl(port));
 519                 break;
 520             }
 521         } else {
 522             switch (size) {
 523             case 1:
 524                 cpu_outb(port, ldub_p(ptr));
 525                 break;
 526             case 2:
 527                 cpu_outw(port, lduw_p(ptr));
 528                 break;
 529             case 4:
 530                 cpu_outl(port, ldl_p(ptr));
 531                 break;
 532             }
 533         }
 534
 535         ptr += size;
 536     }
 537
 538     return 1;
 539 }
 540
 541 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 542 {
 543 #ifdef KVM_CAP_COALESCED_MMIO
 544     KVMState *s = kvm_state;
 545     if (s->coalesced_mmio) {
 546         struct kvm_coalesced_mmio_ring *ring;
 547
 548         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 549         while (ring->first != ring->last) {
 550             struct kvm_coalesced_mmio *ent;
 551
 552             ent = &ring->coalesced_mmio[ring->first];
 553
 554             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 555             /* FIXME smp_wmb() */
 556             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 557         }
 558     }
 559 #endif
 560 }
 561
 562 void kvm_cpu_synchronize_state(CPUState *env)
 563 {
 564     if (!env->kvm_state->regs_modified) {
 565         kvm_arch_get_registers(env);
 566         env->kvm_state->regs_modified = 1;
 567     }
 568 }
 569
 570 int kvm_cpu_exec(CPUState *env)
 571 {
 572     struct kvm_run *run = env->kvm_run;
 573     int ret;
 574
 575     dprintf("kvm_cpu_exec()\n");
 576
 577     do {
 578         if (env->exit_request) {
 579             dprintf("interrupt exit requested\n");
 580             ret = 0;
 581             break;
 582         }
 583
 584         if (env->kvm_state->regs_modified) {
 585             kvm_arch_put_registers(env);
 586             env->kvm_state->regs_modified = 0;
 587         }
 588
 589         kvm_arch_pre_run(env, run);
 590         qemu_mutex_unlock_iothread();
 591         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 592         qemu_mutex_lock_iothread();
 593         kvm_arch_post_run(env, run);
 594
 595         if (ret == -EINTR || ret == -EAGAIN) {
 596             dprintf("io window exit\n");
 597             ret = 0;
 598             break;
 599         }
 600
 601         if (ret < 0) {
 602             dprintf("kvm run failed %s\n", strerror(-ret));
 603             abort();
 604         }
 605
 606         kvm_run_coalesced_mmio(env, run);
 607
 608         ret = 0; /* exit loop */
 609         switch (run->exit_reason) {
 610         case KVM_EXIT_IO:
 611             dprintf("handle_io\n");
 612             ret = kvm_handle_io(run->io.port,
 613                                 (uint8_t *)run + run->io.data_offset,
 614                                 run->io.direction,
 615                                 run->io.size,
 616                                 run->io.count);
 617             break;
 618         case KVM_EXIT_MMIO:
 619             dprintf("handle_mmio\n");
 620             cpu_physical_memory_rw(run->mmio.phys_addr,
 621                                    run->mmio.data,
 622                                    run->mmio.len,
 623                                    run->mmio.is_write);
 624             ret = 1;
 625             break;
 626         case KVM_EXIT_IRQ_WINDOW_OPEN:
 627             dprintf("irq_window_open\n");
 628             break;
 629         case KVM_EXIT_SHUTDOWN:
 630             dprintf("shutdown\n");
 631             qemu_system_reset_request();
 632             ret = 1;
 633             break;
 634         case KVM_EXIT_UNKNOWN:
 635             dprintf("kvm_exit_unknown\n");
 636             break;
 637         case KVM_EXIT_FAIL_ENTRY:
 638             dprintf("kvm_exit_fail_entry\n");
 639             break;
 640         case KVM_EXIT_EXCEPTION:
 641             dprintf("kvm_exit_exception\n");
 642             break;
 643         case KVM_EXIT_DEBUG:
 644             dprintf("kvm_exit_debug\n");
 645 #ifdef KVM_CAP_SET_GUEST_DEBUG
 646             if (kvm_arch_debug(&run->debug.arch)) {
 647                 gdb_set_stop_cpu(env);
 648                 vm_stop(EXCP_DEBUG);
 649                 env->exception_index = EXCP_DEBUG;
 650                 return 0;
 651             }
 652             /* re-enter, this exception was guest-internal */
 653             ret = 1;
 654 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 655             break;
 656         default:
 657             dprintf("kvm_arch_handle_exit\n");
 658             ret = kvm_arch_handle_exit(env, run);
 659             break;
 660         }
 661     } while (ret > 0);
 662
 663     if (env->exit_request) {
 664         env->exit_request = 0;
 665         env->exception_index = EXCP_INTERRUPT;
 666     }
 667
 668     return ret;
 669 }
 670
 671 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 672                       ram_addr_t size,
 673                       ram_addr_t phys_offset)
 674 {
 675     KVMState *s = kvm_state;
 676     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 677     KVMSlot *mem, old;
 678     int err;
 679
 680     if (start_addr & ~TARGET_PAGE_MASK) {
 681         if (flags >= IO_MEM_UNASSIGNED) {
 682             if (!kvm_lookup_overlapping_slot(s, start_addr,
 683                                              start_addr + size)) {
 684                 return;
 685             }
 686             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 687         } else {
 688             fprintf(stderr, "Only page-aligned memory slots supported\n");
 689         }
 690         abort();
 691     }
 692
 693     /* KVM does not support read-only slots */
 694     phys_offset &= ~IO_MEM_ROM;
 695
 696     while (1) {
 697         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 698         if (!mem) {
 699             break;
 700         }
 701
 702         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 703             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 704             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 705             /* The new slot fits into the existing one and comes with
 706              * identical parameters - nothing to be done. */
 707             return;
 708         }
 709
 710         old = *mem;
 711
 712         /* unregister the overlapping slot */
 713         mem->memory_size = 0;
 714         err = kvm_set_user_memory_region(s, mem);
 715         if (err) {
 716             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 717                     __func__, strerror(-err));
 718             abort();
 719         }
 720
 721         /* Workaround for older KVM versions: we can't join slots, even not by
 722          * unregistering the previous ones and then registering the larger
 723          * slot. We have to maintain the existing fragmentation. Sigh.
 724          *
 725          * This workaround assumes that the new slot starts at the same
 726          * address as the first existing one. If not or if some overlapping
 727          * slot comes around later, we will fail (not seen in practice so far)
 728          * - and actually require a recent KVM version. */
 729         if (s->broken_set_mem_region &&
 730             old.start_addr == start_addr && old.memory_size < size &&
 731             flags < IO_MEM_UNASSIGNED) {
 732             mem = kvm_alloc_slot(s);
 733             mem->memory_size = old.memory_size;
 734             mem->start_addr = old.start_addr;
 735             mem->phys_offset = old.phys_offset;
 736             mem->flags = 0;
 737
 738             err = kvm_set_user_memory_region(s, mem);
 739             if (err) {
 740                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 741                         strerror(-err));
 742                 abort();
 743             }
 744
 745             start_addr += old.memory_size;
 746             phys_offset += old.memory_size;
 747             size -= old.memory_size;
 748             continue;
 749         }
 750
 751         /* register prefix slot */
 752         if (old.start_addr < start_addr) {
 753             mem = kvm_alloc_slot(s);
 754             mem->memory_size = start_addr - old.start_addr;
 755             mem->start_addr = old.start_addr;
 756             mem->phys_offset = old.phys_offset;
 757             mem->flags = 0;
 758
 759             err = kvm_set_user_memory_region(s, mem);
 760             if (err) {
 761                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 762                         __func__, strerror(-err));
 763                 abort();
 764             }
 765         }
 766
 767         /* register suffix slot */
 768         if (old.start_addr + old.memory_size > start_addr + size) {
 769             ram_addr_t size_delta;
 770
 771             mem = kvm_alloc_slot(s);
 772             mem->start_addr = start_addr + size;
 773             size_delta = mem->start_addr - old.start_addr;
 774             mem->memory_size = old.memory_size - size_delta;
 775             mem->phys_offset = old.phys_offset + size_delta;
 776             mem->flags = 0;
 777
 778             err = kvm_set_user_memory_region(s, mem);
 779             if (err) {
 780                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 781                         __func__, strerror(-err));
 782                 abort();
 783             }
 784         }
 785     }
 786
 787     /* in case the KVM bug workaround already "consumed" the new slot */
 788     if (!size)
 789         return;
 790
 791     /* KVM does not need to know about this memory */
 792     if (flags >= IO_MEM_UNASSIGNED)
 793         return;
 794
 795     mem = kvm_alloc_slot(s);
 796     mem->memory_size = size;
 797     mem->start_addr = start_addr;
 798     mem->phys_offset = phys_offset;
 799     mem->flags = 0;
 800
 801     err = kvm_set_user_memory_region(s, mem);
 802     if (err) {
 803         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 804                 strerror(-err));
 805         abort();
 806     }
 807 }
 808
 809 int kvm_ioctl(KVMState *s, int type, ...)
 810 {
 811     int ret;
 812     void *arg;
 813     va_list ap;
 814
 815     va_start(ap, type);
 816     arg = va_arg(ap, void *);
 817     va_end(ap);
 818
 819     ret = ioctl(s->fd, type, arg);
 820     if (ret == -1)
 821         ret = -errno;
 822
 823     return ret;
 824 }
 825
 826 int kvm_vm_ioctl(KVMState *s, int type, ...)
 827 {
 828     int ret;
 829     void *arg;
 830     va_list ap;
 831
 832     va_start(ap, type);
 833     arg = va_arg(ap, void *);
 834     va_end(ap);
 835
 836     ret = ioctl(s->vmfd, type, arg);
 837     if (ret == -1)
 838         ret = -errno;
 839
 840     return ret;
 841 }
 842
 843 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 844 {
 845     int ret;
 846     void *arg;
 847     va_list ap;
 848
 849     va_start(ap, type);
 850     arg = va_arg(ap, void *);
 851     va_end(ap);
 852
 853     ret = ioctl(env->kvm_fd, type, arg);
 854     if (ret == -1)
 855         ret = -errno;
 856
 857     return ret;
 858 }
 859
 860 int kvm_has_sync_mmu(void)
 861 {
 862 #ifdef KVM_CAP_SYNC_MMU
 863     KVMState *s = kvm_state;
 864
 865     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 866 #else
 867     return 0;
 868 #endif
 869 }
 870
 871 void kvm_setup_guest_memory(void *start, size_t size)
 872 {
 873     if (!kvm_has_sync_mmu()) {
 874 #ifdef MADV_DONTFORK
 875         int ret = madvise(start, size, MADV_DONTFORK);
 876
 877         if (ret) {
 878             perror("madvice");
 879             exit(1);
 880         }
 881 #else
 882         fprintf(stderr,
 883                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 884         exit(1);
 885 #endif
 886     }
 887 }
 888
 889 #ifdef KVM_CAP_SET_GUEST_DEBUG
 890 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 891 {
 892 #ifdef CONFIG_IOTHREAD
 893     if (env == cpu_single_env) {
 894         func(data);
 895         return;
 896     }
 897     abort();
 898 #else
 899     func(data);
 900 #endif
 901 }
 902
 903 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 904                                                  target_ulong pc)
 905 {
 906     struct kvm_sw_breakpoint *bp;
 907
 908     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 909         if (bp->pc == pc)
 910             return bp;
 911     }
 912     return NULL;
 913 }
 914
 915 int kvm_sw_breakpoints_active(CPUState *env)
 916 {
 917     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 918 }
 919
 920 struct kvm_set_guest_debug_data {
 921     struct kvm_guest_debug dbg;
 922     CPUState *env;
 923     int err;
 924 };
 925
 926 static void kvm_invoke_set_guest_debug(void *data)
 927 {
 928     struct kvm_set_guest_debug_data *dbg_data = data;
 929     CPUState *env = dbg_data->env;
 930
 931     if (env->kvm_state->regs_modified) {
 932         kvm_arch_put_registers(env);
 933         env->kvm_state->regs_modified = 0;
 934     }
 935     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 936 }
 937
 938 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 939 {
 940     struct kvm_set_guest_debug_data data;
 941
 942     data.dbg.control = 0;
 943     if (env->singlestep_enabled)
 944         data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 945
 946     kvm_arch_update_guest_debug(env, &data.dbg);
 947     data.dbg.control |= reinject_trap;
 948     data.env = env;
 949
 950     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
 951     return data.err;
 952 }
 953
 954 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 955                           target_ulong len, int type)
 956 {
 957     struct kvm_sw_breakpoint *bp;
 958     CPUState *env;
 959     int err;
 960
 961     if (type == GDB_BREAKPOINT_SW) {
 962         bp = kvm_find_sw_breakpoint(current_env, addr);
 963         if (bp) {
 964             bp->use_count++;
 965             return 0;
 966         }
 967
 968         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 969         if (!bp)
 970             return -ENOMEM;
 971
 972         bp->pc = addr;
 973         bp->use_count = 1;
 974         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 975         if (err) {
 976             free(bp);
 977             return err;
 978         }
 979
 980         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 981                           bp, entry);
 982     } else {
 983         err = kvm_arch_insert_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 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
 997                           target_ulong len, int type)
 998 {
 999     struct kvm_sw_breakpoint *bp;
1000     CPUState *env;
1001     int err;
1002
1003     if (type == GDB_BREAKPOINT_SW) {
1004         bp = kvm_find_sw_breakpoint(current_env, addr);
1005         if (!bp)
1006             return -ENOENT;
1007
1008         if (bp->use_count > 1) {
1009             bp->use_count--;
1010             return 0;
1011         }
1012
1013         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1014         if (err)
1015             return err;
1016
1017         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1018         qemu_free(bp);
1019     } else {
1020         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1021         if (err)
1022             return err;
1023     }
1024
1025     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1026         err = kvm_update_guest_debug(env, 0);
1027         if (err)
1028             return err;
1029     }
1030     return 0;
1031 }
1032
1033 void kvm_remove_all_breakpoints(CPUState *current_env)
1034 {
1035     struct kvm_sw_breakpoint *bp, *next;
1036     KVMState *s = current_env->kvm_state;
1037     CPUState *env;
1038
1039     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1040         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1041             /* Try harder to find a CPU that currently sees the breakpoint. */
1042             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1043                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1044                     break;
1045             }
1046         }
1047     }
1048     kvm_arch_remove_all_hw_breakpoints();
1049
1050     for (env = first_cpu; env != NULL; env = env->next_cpu)
1051         kvm_update_guest_debug(env, 0);
1052 }
1053
1054 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1055
1056 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1057 {
1058     return -EINVAL;
1059 }
1060
1061 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1062                           target_ulong len, int type)
1063 {
1064     return -EINVAL;
1065 }
1066
1067 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1068                           target_ulong len, int type)
1069 {
1070     return -EINVAL;
1071 }
1072
1073 void kvm_remove_all_breakpoints(CPUState *current_env)
1074 {
1075 }
1076 #endif /* !KVM_CAP_SET_GUEST_DEBUG */