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     KVMState *s;
 408     int ret;
 409     int i;
 410
 411     if (smp_cpus > 1) {
 412         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 413         return -EINVAL;
 414     }
 415
 416     s = qemu_mallocz(sizeof(KVMState));
 417
 418 #ifdef KVM_CAP_SET_GUEST_DEBUG
 419     TAILQ_INIT(&s->kvm_sw_breakpoints);
 420 #endif
 421     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 422         s->slots[i].slot = i;
 423
 424     s->vmfd = -1;
 425     s->fd = open("/dev/kvm", O_RDWR);
 426     if (s->fd == -1) {
 427         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 428         ret = -errno;
 429         goto err;
 430     }
 431
 432     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 433     if (ret < KVM_API_VERSION) {
 434         if (ret > 0)
 435             ret = -EINVAL;
 436         fprintf(stderr, "kvm version too old\n");
 437         goto err;
 438     }
 439
 440     if (ret > KVM_API_VERSION) {
 441         ret = -EINVAL;
 442         fprintf(stderr, "kvm version not supported\n");
 443         goto err;
 444     }
 445
 446     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 447     if (s->vmfd < 0)
 448         goto err;
 449
 450     /* initially, KVM allocated its own memory and we had to jump through
 451      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 452      * just use a user allocated buffer so we can use regular pages
 453      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 454      */
 455     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 456         ret = -EINVAL;
 457         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
 458         goto err;
 459     }
 460
 461     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 462      * destroyed properly.  Since we rely on this capability, refuse to work
 463      * with any kernel without this capability. */
 464     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 465         ret = -EINVAL;
 466
 467         fprintf(stderr,
 468                 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
 469                 "Please upgrade to at least kvm-81.\n");
 470         goto err;
 471     }
 472
 473 #ifdef KVM_CAP_COALESCED_MMIO
 474     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 475 #else
 476     s->coalesced_mmio = 0;
 477 #endif
 478
 479     s->broken_set_mem_region = 1;
 480 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 481     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 482     if (ret > 0) {
 483         s->broken_set_mem_region = 0;
 484     }
 485 #endif
 486
 487     ret = kvm_arch_init(s, smp_cpus);
 488     if (ret < 0)
 489         goto err;
 490
 491     qemu_register_reset(kvm_reset_vcpus, INT_MAX, NULL);
 492
 493     kvm_state = s;
 494
 495     return 0;
 496
 497 err:
 498     if (s) {
 499         if (s->vmfd != -1)
 500             close(s->vmfd);
 501         if (s->fd != -1)
 502             close(s->fd);
 503     }
 504     qemu_free(s);
 505
 506     return ret;
 507 }
 508
 509 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 510                          int direction, int size, uint32_t count)
 511 {
 512     int i;
 513     uint8_t *ptr = data;
 514
 515     for (i = 0; i < count; i++) {
 516         if (direction == KVM_EXIT_IO_IN) {
 517             switch (size) {
 518             case 1:
 519                 stb_p(ptr, cpu_inb(env, port));
 520                 break;
 521             case 2:
 522                 stw_p(ptr, cpu_inw(env, port));
 523                 break;
 524             case 4:
 525                 stl_p(ptr, cpu_inl(env, port));
 526                 break;
 527             }
 528         } else {
 529             switch (size) {
 530             case 1:
 531                 cpu_outb(env, port, ldub_p(ptr));
 532                 break;
 533             case 2:
 534                 cpu_outw(env, port, lduw_p(ptr));
 535                 break;
 536             case 4:
 537                 cpu_outl(env, port, ldl_p(ptr));
 538                 break;
 539             }
 540         }
 541
 542         ptr += size;
 543     }
 544
 545     return 1;
 546 }
 547
 548 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 549 {
 550 #ifdef KVM_CAP_COALESCED_MMIO
 551     KVMState *s = kvm_state;
 552     if (s->coalesced_mmio) {
 553         struct kvm_coalesced_mmio_ring *ring;
 554
 555         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 556         while (ring->first != ring->last) {
 557             struct kvm_coalesced_mmio *ent;
 558
 559             ent = &ring->coalesced_mmio[ring->first];
 560
 561             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 562             /* FIXME smp_wmb() */
 563             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 564         }
 565     }
 566 #endif
 567 }
 568
 569 int kvm_cpu_exec(CPUState *env)
 570 {
 571     struct kvm_run *run = env->kvm_run;
 572     int ret;
 573
 574     dprintf("kvm_cpu_exec()\n");
 575
 576     do {
 577         kvm_arch_pre_run(env, run);
 578
 579         if (env->exit_request) {
 580             dprintf("interrupt exit requested\n");
 581             ret = 0;
 582             break;
 583         }
 584
 585         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 586         kvm_arch_post_run(env, run);
 587
 588         if (ret == -EINTR || ret == -EAGAIN) {
 589             dprintf("io window exit\n");
 590             ret = 0;
 591             break;
 592         }
 593
 594         if (ret < 0) {
 595             dprintf("kvm run failed %s\n", strerror(-ret));
 596             abort();
 597         }
 598
 599         kvm_run_coalesced_mmio(env, run);
 600
 601         ret = 0; /* exit loop */
 602         switch (run->exit_reason) {
 603         case KVM_EXIT_IO:
 604             dprintf("handle_io\n");
 605             ret = kvm_handle_io(env, run->io.port,
 606                                 (uint8_t *)run + run->io.data_offset,
 607                                 run->io.direction,
 608                                 run->io.size,
 609                                 run->io.count);
 610             break;
 611         case KVM_EXIT_MMIO:
 612             dprintf("handle_mmio\n");
 613             cpu_physical_memory_rw(run->mmio.phys_addr,
 614                                    run->mmio.data,
 615                                    run->mmio.len,
 616                                    run->mmio.is_write);
 617             ret = 1;
 618             break;
 619         case KVM_EXIT_IRQ_WINDOW_OPEN:
 620             dprintf("irq_window_open\n");
 621             break;
 622         case KVM_EXIT_SHUTDOWN:
 623             dprintf("shutdown\n");
 624             qemu_system_reset_request();
 625             ret = 1;
 626             break;
 627         case KVM_EXIT_UNKNOWN:
 628             dprintf("kvm_exit_unknown\n");
 629             break;
 630         case KVM_EXIT_FAIL_ENTRY:
 631             dprintf("kvm_exit_fail_entry\n");
 632             break;
 633         case KVM_EXIT_EXCEPTION:
 634             dprintf("kvm_exit_exception\n");
 635             break;
 636         case KVM_EXIT_DEBUG:
 637             dprintf("kvm_exit_debug\n");
 638 #ifdef KVM_CAP_SET_GUEST_DEBUG
 639             if (kvm_arch_debug(&run->debug.arch)) {
 640                 gdb_set_stop_cpu(env);
 641                 vm_stop(EXCP_DEBUG);
 642                 env->exception_index = EXCP_DEBUG;
 643                 return 0;
 644             }
 645             /* re-enter, this exception was guest-internal */
 646             ret = 1;
 647 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 648             break;
 649         default:
 650             dprintf("kvm_arch_handle_exit\n");
 651             ret = kvm_arch_handle_exit(env, run);
 652             break;
 653         }
 654     } while (ret > 0);
 655
 656     if (env->exit_request) {
 657         env->exit_request = 0;
 658         env->exception_index = EXCP_INTERRUPT;
 659     }
 660
 661     return ret;
 662 }
 663
 664 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 665                       ram_addr_t size,
 666                       ram_addr_t phys_offset)
 667 {
 668     KVMState *s = kvm_state;
 669     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 670     KVMSlot *mem, old;
 671     int err;
 672
 673     if (start_addr & ~TARGET_PAGE_MASK) {
 674         if (flags >= IO_MEM_UNASSIGNED) {
 675             if (!kvm_lookup_overlapping_slot(s, start_addr,
 676                                              start_addr + size)) {
 677                 return;
 678             }
 679             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 680         } else {
 681             fprintf(stderr, "Only page-aligned memory slots supported\n");
 682         }
 683         abort();
 684     }
 685
 686     /* KVM does not support read-only slots */
 687     phys_offset &= ~IO_MEM_ROM;
 688
 689     while (1) {
 690         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 691         if (!mem) {
 692             break;
 693         }
 694
 695         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 696             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 697             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 698             /* The new slot fits into the existing one and comes with
 699              * identical parameters - nothing to be done. */
 700             return;
 701         }
 702
 703         old = *mem;
 704
 705         /* unregister the overlapping slot */
 706         mem->memory_size = 0;
 707         err = kvm_set_user_memory_region(s, mem);
 708         if (err) {
 709             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 710                     __func__, strerror(-err));
 711             abort();
 712         }
 713
 714         /* Workaround for older KVM versions: we can't join slots, even not by
 715          * unregistering the previous ones and then registering the larger
 716          * slot. We have to maintain the existing fragmentation. Sigh.
 717          *
 718          * This workaround assumes that the new slot starts at the same
 719          * address as the first existing one. If not or if some overlapping
 720          * slot comes around later, we will fail (not seen in practice so far)
 721          * - and actually require a recent KVM version. */
 722         if (s->broken_set_mem_region &&
 723             old.start_addr == start_addr && old.memory_size < size &&
 724             flags < IO_MEM_UNASSIGNED) {
 725             mem = kvm_alloc_slot(s);
 726             mem->memory_size = old.memory_size;
 727             mem->start_addr = old.start_addr;
 728             mem->phys_offset = old.phys_offset;
 729             mem->flags = 0;
 730
 731             err = kvm_set_user_memory_region(s, mem);
 732             if (err) {
 733                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 734                         strerror(-err));
 735                 abort();
 736             }
 737
 738             start_addr += old.memory_size;
 739             phys_offset += old.memory_size;
 740             size -= old.memory_size;
 741             continue;
 742         }
 743
 744         /* register prefix slot */
 745         if (old.start_addr < start_addr) {
 746             mem = kvm_alloc_slot(s);
 747             mem->memory_size = start_addr - old.start_addr;
 748             mem->start_addr = old.start_addr;
 749             mem->phys_offset = old.phys_offset;
 750             mem->flags = 0;
 751
 752             err = kvm_set_user_memory_region(s, mem);
 753             if (err) {
 754                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 755                         __func__, strerror(-err));
 756                 abort();
 757             }
 758         }
 759
 760         /* register suffix slot */
 761         if (old.start_addr + old.memory_size > start_addr + size) {
 762             ram_addr_t size_delta;
 763
 764             mem = kvm_alloc_slot(s);
 765             mem->start_addr = start_addr + size;
 766             size_delta = mem->start_addr - old.start_addr;
 767             mem->memory_size = old.memory_size - size_delta;
 768             mem->phys_offset = old.phys_offset + size_delta;
 769             mem->flags = 0;
 770
 771             err = kvm_set_user_memory_region(s, mem);
 772             if (err) {
 773                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 774                         __func__, strerror(-err));
 775                 abort();
 776             }
 777         }
 778     }
 779
 780     /* in case the KVM bug workaround already "consumed" the new slot */
 781     if (!size)
 782         return;
 783
 784     /* KVM does not need to know about this memory */
 785     if (flags >= IO_MEM_UNASSIGNED)
 786         return;
 787
 788     mem = kvm_alloc_slot(s);
 789     mem->memory_size = size;
 790     mem->start_addr = start_addr;
 791     mem->phys_offset = phys_offset;
 792     mem->flags = 0;
 793
 794     err = kvm_set_user_memory_region(s, mem);
 795     if (err) {
 796         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 797                 strerror(-err));
 798         abort();
 799     }
 800 }
 801
 802 int kvm_ioctl(KVMState *s, int type, ...)
 803 {
 804     int ret;
 805     void *arg;
 806     va_list ap;
 807
 808     va_start(ap, type);
 809     arg = va_arg(ap, void *);
 810     va_end(ap);
 811
 812     ret = ioctl(s->fd, type, arg);
 813     if (ret == -1)
 814         ret = -errno;
 815
 816     return ret;
 817 }
 818
 819 int kvm_vm_ioctl(KVMState *s, int type, ...)
 820 {
 821     int ret;
 822     void *arg;
 823     va_list ap;
 824
 825     va_start(ap, type);
 826     arg = va_arg(ap, void *);
 827     va_end(ap);
 828
 829     ret = ioctl(s->vmfd, type, arg);
 830     if (ret == -1)
 831         ret = -errno;
 832
 833     return ret;
 834 }
 835
 836 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 837 {
 838     int ret;
 839     void *arg;
 840     va_list ap;
 841
 842     va_start(ap, type);
 843     arg = va_arg(ap, void *);
 844     va_end(ap);
 845
 846     ret = ioctl(env->kvm_fd, type, arg);
 847     if (ret == -1)
 848         ret = -errno;
 849
 850     return ret;
 851 }
 852
 853 int kvm_has_sync_mmu(void)
 854 {
 855 #ifdef KVM_CAP_SYNC_MMU
 856     KVMState *s = kvm_state;
 857
 858     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 859 #else
 860     return 0;
 861 #endif
 862 }
 863
 864 void kvm_setup_guest_memory(void *start, size_t size)
 865 {
 866     if (!kvm_has_sync_mmu()) {
 867 #ifdef MADV_DONTFORK
 868         int ret = madvise(start, size, MADV_DONTFORK);
 869
 870         if (ret) {
 871             perror("madvice");
 872             exit(1);
 873         }
 874 #else
 875         fprintf(stderr,
 876                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 877         exit(1);
 878 #endif
 879     }
 880 }
 881
 882 #ifdef KVM_CAP_SET_GUEST_DEBUG
 883 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 884                                                  target_ulong pc)
 885 {
 886     struct kvm_sw_breakpoint *bp;
 887
 888     TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 889         if (bp->pc == pc)
 890             return bp;
 891     }
 892     return NULL;
 893 }
 894
 895 int kvm_sw_breakpoints_active(CPUState *env)
 896 {
 897     return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 898 }
 899
 900 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 901 {
 902     struct kvm_guest_debug dbg;
 903
 904     dbg.control = 0;
 905     if (env->singlestep_enabled)
 906         dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 907
 908     kvm_arch_update_guest_debug(env, &dbg);
 909     dbg.control |= reinject_trap;
 910
 911     return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
 912 }
 913
 914 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 915                           target_ulong len, int type)
 916 {
 917     struct kvm_sw_breakpoint *bp;
 918     CPUState *env;
 919     int err;
 920
 921     if (type == GDB_BREAKPOINT_SW) {
 922         bp = kvm_find_sw_breakpoint(current_env, addr);
 923         if (bp) {
 924             bp->use_count++;
 925             return 0;
 926         }
 927
 928         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 929         if (!bp)
 930             return -ENOMEM;
 931
 932         bp->pc = addr;
 933         bp->use_count = 1;
 934         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 935         if (err) {
 936             free(bp);
 937             return err;
 938         }
 939
 940         TAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 941                           bp, entry);
 942     } else {
 943         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
 944         if (err)
 945             return err;
 946     }
 947
 948     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 949         err = kvm_update_guest_debug(env, 0);
 950         if (err)
 951             return err;
 952     }
 953     return 0;
 954 }
 955
 956 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
 957                           target_ulong len, int type)
 958 {
 959     struct kvm_sw_breakpoint *bp;
 960     CPUState *env;
 961     int err;
 962
 963     if (type == GDB_BREAKPOINT_SW) {
 964         bp = kvm_find_sw_breakpoint(current_env, addr);
 965         if (!bp)
 966             return -ENOENT;
 967
 968         if (bp->use_count > 1) {
 969             bp->use_count--;
 970             return 0;
 971         }
 972
 973         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
 974         if (err)
 975             return err;
 976
 977         TAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
 978         qemu_free(bp);
 979     } else {
 980         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
 981         if (err)
 982             return err;
 983     }
 984
 985     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 986         err = kvm_update_guest_debug(env, 0);
 987         if (err)
 988             return err;
 989     }
 990     return 0;
 991 }
 992
 993 void kvm_remove_all_breakpoints(CPUState *current_env)
 994 {
 995     struct kvm_sw_breakpoint *bp, *next;
 996     KVMState *s = current_env->kvm_state;
 997     CPUState *env;
 998
 999     TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1000         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1001             /* Try harder to find a CPU that currently sees the breakpoint. */
1002             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1003                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1004                     break;
1005             }
1006         }
1007     }
1008     kvm_arch_remove_all_hw_breakpoints();
1009
1010     for (env = first_cpu; env != NULL; env = env->next_cpu)
1011         kvm_update_guest_debug(env, 0);
1012 }
1013
1014 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1015
1016 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1017 {
1018     return -EINVAL;
1019 }
1020
1021 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1022                           target_ulong len, int type)
1023 {
1024     return -EINVAL;
1025 }
1026
1027 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1028                           target_ulong len, int type)
1029 {
1030     return -EINVAL;
1031 }
1032
1033 void kvm_remove_all_breakpoints(CPUState *current_env)
1034 {
1035 }
1036 #endif /* !KVM_CAP_SET_GUEST_DEBUG */