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 #ifdef KVM_CAP_COALESCED_MMIO
  62     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  63 #endif
  64     int broken_set_mem_region;
  65     int migration_log;
  66     int vcpu_events;
  67 #ifdef KVM_CAP_SET_GUEST_DEBUG
  68     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  69 #endif
  70     int irqchip_in_kernel;
  71     int pit_in_kernel;
  72 };
  73
  74 static KVMState *kvm_state;
  75
  76 static KVMSlot *kvm_alloc_slot(KVMState *s)
  77 {
  78     int i;
  79
  80     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  81         /* KVM private memory slots */
  82         if (i >= 8 && i < 12)
  83             continue;
  84         if (s->slots[i].memory_size == 0)
  85             return &s->slots[i];
  86     }
  87
  88     fprintf(stderr, "%s: no free slot available\n", __func__);
  89     abort();
  90 }
  91
  92 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  93                                          target_phys_addr_t start_addr,
  94                                          target_phys_addr_t end_addr)
  95 {
  96     int i;
  97
  98     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  99         KVMSlot *mem = &s->slots[i];
 100
 101         if (start_addr == mem->start_addr &&
 102             end_addr == mem->start_addr + mem->memory_size) {
 103             return mem;
 104         }
 105     }
 106
 107     return NULL;
 108 }
 109
 110 /*
 111  * Find overlapping slot with lowest start address
 112  */
 113 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 114                                             target_phys_addr_t start_addr,
 115                                             target_phys_addr_t end_addr)
 116 {
 117     KVMSlot *found = NULL;
 118     int i;
 119
 120     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 121         KVMSlot *mem = &s->slots[i];
 122
 123         if (mem->memory_size == 0 ||
 124             (found && found->start_addr < mem->start_addr)) {
 125             continue;
 126         }
 127
 128         if (end_addr > mem->start_addr &&
 129             start_addr < mem->start_addr + mem->memory_size) {
 130             found = mem;
 131         }
 132     }
 133
 134     return found;
 135 }
 136
 137 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 138 {
 139     struct kvm_userspace_memory_region mem;
 140
 141     mem.slot = slot->slot;
 142     mem.guest_phys_addr = slot->start_addr;
 143     mem.memory_size = slot->memory_size;
 144     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 145     mem.flags = slot->flags;
 146     if (s->migration_log) {
 147         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 148     }
 149     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 150 }
 151
 152 static void kvm_reset_vcpu(void *opaque)
 153 {
 154     CPUState *env = opaque;
 155
 156     kvm_arch_reset_vcpu(env);
 157     if (kvm_arch_put_registers(env)) {
 158         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 159         abort();
 160     }
 161 }
 162
 163 int kvm_irqchip_in_kernel(void)
 164 {
 165     return kvm_state->irqchip_in_kernel;
 166 }
 167
 168 int kvm_pit_in_kernel(void)
 169 {
 170     return kvm_state->pit_in_kernel;
 171 }
 172
 173
 174 int kvm_init_vcpu(CPUState *env)
 175 {
 176     KVMState *s = kvm_state;
 177     long mmap_size;
 178     int ret;
 179
 180     dprintf("kvm_init_vcpu\n");
 181
 182     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 183     if (ret < 0) {
 184         dprintf("kvm_create_vcpu failed\n");
 185         goto err;
 186     }
 187
 188     env->kvm_fd = ret;
 189     env->kvm_state = s;
 190
 191     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 192     if (mmap_size < 0) {
 193         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 194         goto err;
 195     }
 196
 197     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 198                         env->kvm_fd, 0);
 199     if (env->kvm_run == MAP_FAILED) {
 200         ret = -errno;
 201         dprintf("mmap'ing vcpu state failed\n");
 202         goto err;
 203     }
 204
 205 #ifdef KVM_CAP_COALESCED_MMIO
 206     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 207         s->coalesced_mmio_ring = (void *) env->kvm_run +
 208                 s->coalesced_mmio * PAGE_SIZE;
 209 #endif
 210
 211     ret = kvm_arch_init_vcpu(env);
 212     if (ret == 0) {
 213         qemu_register_reset(kvm_reset_vcpu, env);
 214         kvm_arch_reset_vcpu(env);
 215         ret = kvm_arch_put_registers(env);
 216     }
 217 err:
 218     return ret;
 219 }
 220
 221 /*
 222  * dirty pages logging control
 223  */
 224 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 225                                       ram_addr_t size, int flags, int mask)
 226 {
 227     KVMState *s = kvm_state;
 228     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 229     int old_flags;
 230
 231     if (mem == NULL)  {
 232             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 233                     TARGET_FMT_plx "\n", __func__, phys_addr,
 234                     (target_phys_addr_t)(phys_addr + size - 1));
 235             return -EINVAL;
 236     }
 237
 238     old_flags = mem->flags;
 239
 240     flags = (mem->flags & ~mask) | flags;
 241     mem->flags = flags;
 242
 243     /* If nothing changed effectively, no need to issue ioctl */
 244     if (s->migration_log) {
 245         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 246     }
 247     if (flags == old_flags) {
 248             return 0;
 249     }
 250
 251     return kvm_set_user_memory_region(s, mem);
 252 }
 253
 254 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 255 {
 256         return kvm_dirty_pages_log_change(phys_addr, size,
 257                                           KVM_MEM_LOG_DIRTY_PAGES,
 258                                           KVM_MEM_LOG_DIRTY_PAGES);
 259 }
 260
 261 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 262 {
 263         return kvm_dirty_pages_log_change(phys_addr, size,
 264                                           0,
 265                                           KVM_MEM_LOG_DIRTY_PAGES);
 266 }
 267
 268 int kvm_set_migration_log(int enable)
 269 {
 270     KVMState *s = kvm_state;
 271     KVMSlot *mem;
 272     int i, err;
 273
 274     s->migration_log = enable;
 275
 276     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 277         mem = &s->slots[i];
 278
 279         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 280             continue;
 281         }
 282         err = kvm_set_user_memory_region(s, mem);
 283         if (err) {
 284             return err;
 285         }
 286     }
 287     return 0;
 288 }
 289
 290 static int test_le_bit(unsigned long nr, unsigned char *addr)
 291 {
 292     return (addr[nr >> 3] >> (nr & 7)) & 1;
 293 }
 294
 295 /**
 296  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 297  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 298  * This means all bits are set to dirty.
 299  *
 300  * @start_add: start of logged region.
 301  * @end_addr: end of logged region.
 302  */
 303 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 304                                    target_phys_addr_t end_addr)
 305 {
 306     KVMState *s = kvm_state;
 307     unsigned long size, allocated_size = 0;
 308     target_phys_addr_t phys_addr;
 309     ram_addr_t addr;
 310     KVMDirtyLog d;
 311     KVMSlot *mem;
 312     int ret = 0;
 313
 314     d.dirty_bitmap = NULL;
 315     while (start_addr < end_addr) {
 316         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 317         if (mem == NULL) {
 318             break;
 319         }
 320
 321         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 322         if (!d.dirty_bitmap) {
 323             d.dirty_bitmap = qemu_malloc(size);
 324         } else if (size > allocated_size) {
 325             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 326         }
 327         allocated_size = size;
 328         memset(d.dirty_bitmap, 0, allocated_size);
 329
 330         d.slot = mem->slot;
 331
 332         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 333             dprintf("ioctl failed %d\n", errno);
 334             ret = -1;
 335             break;
 336         }
 337
 338         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 339              phys_addr < mem->start_addr + mem->memory_size;
 340              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 341             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 342             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 343
 344             if (test_le_bit(nr, bitmap)) {
 345                 cpu_physical_memory_set_dirty(addr);
 346             }
 347         }
 348         start_addr = phys_addr;
 349     }
 350     qemu_free(d.dirty_bitmap);
 351
 352     return ret;
 353 }
 354
 355 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 356 {
 357     int ret = -ENOSYS;
 358 #ifdef KVM_CAP_COALESCED_MMIO
 359     KVMState *s = kvm_state;
 360
 361     if (s->coalesced_mmio) {
 362         struct kvm_coalesced_mmio_zone zone;
 363
 364         zone.addr = start;
 365         zone.size = size;
 366
 367         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 368     }
 369 #endif
 370
 371     return ret;
 372 }
 373
 374 int kvm_uncoalesce_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_UNREGISTER_COALESCED_MMIO, &zone);
 387     }
 388 #endif
 389
 390     return ret;
 391 }
 392
 393 int kvm_check_extension(KVMState *s, unsigned int extension)
 394 {
 395     int ret;
 396
 397     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 398     if (ret < 0) {
 399         ret = 0;
 400     }
 401
 402     return ret;
 403 }
 404
 405 int kvm_init(int smp_cpus)
 406 {
 407     static const char upgrade_note[] =
 408         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 409         "(see http://sourceforge.net/projects/kvm).\n";
 410     KVMState *s;
 411     int ret;
 412     int i;
 413
 414     if (smp_cpus > 1) {
 415         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 416         return -EINVAL;
 417     }
 418
 419     s = qemu_mallocz(sizeof(KVMState));
 420
 421 #ifdef KVM_CAP_SET_GUEST_DEBUG
 422     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 423 #endif
 424     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 425         s->slots[i].slot = i;
 426
 427     s->vmfd = -1;
 428     s->fd = qemu_open("/dev/kvm", O_RDWR);
 429     if (s->fd == -1) {
 430         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 431         ret = -errno;
 432         goto err;
 433     }
 434
 435     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 436     if (ret < KVM_API_VERSION) {
 437         if (ret > 0)
 438             ret = -EINVAL;
 439         fprintf(stderr, "kvm version too old\n");
 440         goto err;
 441     }
 442
 443     if (ret > KVM_API_VERSION) {
 444         ret = -EINVAL;
 445         fprintf(stderr, "kvm version not supported\n");
 446         goto err;
 447     }
 448
 449     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 450     if (s->vmfd < 0)
 451         goto err;
 452
 453     /* initially, KVM allocated its own memory and we had to jump through
 454      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 455      * just use a user allocated buffer so we can use regular pages
 456      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 457      */
 458     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 459         ret = -EINVAL;
 460         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 461                 upgrade_note);
 462         goto err;
 463     }
 464
 465     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 466      * destroyed properly.  Since we rely on this capability, refuse to work
 467      * with any kernel without this capability. */
 468     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 469         ret = -EINVAL;
 470
 471         fprintf(stderr,
 472                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 473                 upgrade_note);
 474         goto err;
 475     }
 476
 477     s->coalesced_mmio = 0;
 478 #ifdef KVM_CAP_COALESCED_MMIO
 479     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 480     s->coalesced_mmio_ring = NULL;
 481 #endif
 482
 483     s->broken_set_mem_region = 1;
 484 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 485     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 486     if (ret > 0) {
 487         s->broken_set_mem_region = 0;
 488     }
 489 #endif
 490
 491     s->vcpu_events = 0;
 492 #ifdef KVM_CAP_VCPU_EVENTS
 493     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 494 #endif
 495
 496     ret = kvm_arch_init(s, smp_cpus);
 497     if (ret < 0)
 498         goto err;
 499
 500     kvm_state = s;
 501
 502     return 0;
 503
 504 err:
 505     if (s) {
 506         if (s->vmfd != -1)
 507             close(s->vmfd);
 508         if (s->fd != -1)
 509             close(s->fd);
 510     }
 511     qemu_free(s);
 512
 513     return ret;
 514 }
 515
 516 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 517                          uint32_t count)
 518 {
 519     int i;
 520     uint8_t *ptr = data;
 521
 522     for (i = 0; i < count; i++) {
 523         if (direction == KVM_EXIT_IO_IN) {
 524             switch (size) {
 525             case 1:
 526                 stb_p(ptr, cpu_inb(port));
 527                 break;
 528             case 2:
 529                 stw_p(ptr, cpu_inw(port));
 530                 break;
 531             case 4:
 532                 stl_p(ptr, cpu_inl(port));
 533                 break;
 534             }
 535         } else {
 536             switch (size) {
 537             case 1:
 538                 cpu_outb(port, ldub_p(ptr));
 539                 break;
 540             case 2:
 541                 cpu_outw(port, lduw_p(ptr));
 542                 break;
 543             case 4:
 544                 cpu_outl(port, ldl_p(ptr));
 545                 break;
 546             }
 547         }
 548
 549         ptr += size;
 550     }
 551
 552     return 1;
 553 }
 554
 555 void kvm_flush_coalesced_mmio_buffer(void)
 556 {
 557 #ifdef KVM_CAP_COALESCED_MMIO
 558     KVMState *s = kvm_state;
 559     if (s->coalesced_mmio_ring) {
 560         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 561         while (ring->first != ring->last) {
 562             struct kvm_coalesced_mmio *ent;
 563
 564             ent = &ring->coalesced_mmio[ring->first];
 565
 566             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 567             /* FIXME smp_wmb() */
 568             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 569         }
 570     }
 571 #endif
 572 }
 573
 574 void kvm_cpu_synchronize_state(CPUState *env)
 575 {
 576     if (!env->kvm_vcpu_dirty) {
 577         kvm_arch_get_registers(env);
 578         env->kvm_vcpu_dirty = 1;
 579     }
 580 }
 581
 582 int kvm_cpu_exec(CPUState *env)
 583 {
 584     struct kvm_run *run = env->kvm_run;
 585     int ret;
 586
 587     dprintf("kvm_cpu_exec()\n");
 588
 589     do {
 590         if (env->exit_request) {
 591             dprintf("interrupt exit requested\n");
 592             ret = 0;
 593             break;
 594         }
 595
 596         if (env->kvm_vcpu_dirty) {
 597             kvm_arch_put_registers(env);
 598             env->kvm_vcpu_dirty = 0;
 599         }
 600
 601         kvm_arch_pre_run(env, run);
 602         qemu_mutex_unlock_iothread();
 603         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 604         qemu_mutex_lock_iothread();
 605         kvm_arch_post_run(env, run);
 606
 607         if (ret == -EINTR || ret == -EAGAIN) {
 608             dprintf("io window exit\n");
 609             ret = 0;
 610             break;
 611         }
 612
 613         if (ret < 0) {
 614             dprintf("kvm run failed %s\n", strerror(-ret));
 615             abort();
 616         }
 617
 618         kvm_flush_coalesced_mmio_buffer();
 619
 620         ret = 0; /* exit loop */
 621         switch (run->exit_reason) {
 622         case KVM_EXIT_IO:
 623             dprintf("handle_io\n");
 624             ret = kvm_handle_io(run->io.port,
 625                                 (uint8_t *)run + run->io.data_offset,
 626                                 run->io.direction,
 627                                 run->io.size,
 628                                 run->io.count);
 629             break;
 630         case KVM_EXIT_MMIO:
 631             dprintf("handle_mmio\n");
 632             cpu_physical_memory_rw(run->mmio.phys_addr,
 633                                    run->mmio.data,
 634                                    run->mmio.len,
 635                                    run->mmio.is_write);
 636             ret = 1;
 637             break;
 638         case KVM_EXIT_IRQ_WINDOW_OPEN:
 639             dprintf("irq_window_open\n");
 640             break;
 641         case KVM_EXIT_SHUTDOWN:
 642             dprintf("shutdown\n");
 643             qemu_system_reset_request();
 644             ret = 1;
 645             break;
 646         case KVM_EXIT_UNKNOWN:
 647             dprintf("kvm_exit_unknown\n");
 648             break;
 649         case KVM_EXIT_FAIL_ENTRY:
 650             dprintf("kvm_exit_fail_entry\n");
 651             break;
 652         case KVM_EXIT_EXCEPTION:
 653             dprintf("kvm_exit_exception\n");
 654             break;
 655         case KVM_EXIT_DEBUG:
 656             dprintf("kvm_exit_debug\n");
 657 #ifdef KVM_CAP_SET_GUEST_DEBUG
 658             if (kvm_arch_debug(&run->debug.arch)) {
 659                 gdb_set_stop_cpu(env);
 660                 vm_stop(EXCP_DEBUG);
 661                 env->exception_index = EXCP_DEBUG;
 662                 return 0;
 663             }
 664             /* re-enter, this exception was guest-internal */
 665             ret = 1;
 666 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 667             break;
 668         default:
 669             dprintf("kvm_arch_handle_exit\n");
 670             ret = kvm_arch_handle_exit(env, run);
 671             break;
 672         }
 673     } while (ret > 0);
 674
 675     if (env->exit_request) {
 676         env->exit_request = 0;
 677         env->exception_index = EXCP_INTERRUPT;
 678     }
 679
 680     return ret;
 681 }
 682
 683 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 684                       ram_addr_t size,
 685                       ram_addr_t phys_offset)
 686 {
 687     KVMState *s = kvm_state;
 688     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 689     KVMSlot *mem, old;
 690     int err;
 691
 692     if (start_addr & ~TARGET_PAGE_MASK) {
 693         if (flags >= IO_MEM_UNASSIGNED) {
 694             if (!kvm_lookup_overlapping_slot(s, start_addr,
 695                                              start_addr + size)) {
 696                 return;
 697             }
 698             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 699         } else {
 700             fprintf(stderr, "Only page-aligned memory slots supported\n");
 701         }
 702         abort();
 703     }
 704
 705     /* KVM does not support read-only slots */
 706     phys_offset &= ~IO_MEM_ROM;
 707
 708     while (1) {
 709         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 710         if (!mem) {
 711             break;
 712         }
 713
 714         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 715             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 716             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 717             /* The new slot fits into the existing one and comes with
 718              * identical parameters - nothing to be done. */
 719             return;
 720         }
 721
 722         old = *mem;
 723
 724         /* unregister the overlapping slot */
 725         mem->memory_size = 0;
 726         err = kvm_set_user_memory_region(s, mem);
 727         if (err) {
 728             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 729                     __func__, strerror(-err));
 730             abort();
 731         }
 732
 733         /* Workaround for older KVM versions: we can't join slots, even not by
 734          * unregistering the previous ones and then registering the larger
 735          * slot. We have to maintain the existing fragmentation. Sigh.
 736          *
 737          * This workaround assumes that the new slot starts at the same
 738          * address as the first existing one. If not or if some overlapping
 739          * slot comes around later, we will fail (not seen in practice so far)
 740          * - and actually require a recent KVM version. */
 741         if (s->broken_set_mem_region &&
 742             old.start_addr == start_addr && old.memory_size < size &&
 743             flags < IO_MEM_UNASSIGNED) {
 744             mem = kvm_alloc_slot(s);
 745             mem->memory_size = old.memory_size;
 746             mem->start_addr = old.start_addr;
 747             mem->phys_offset = old.phys_offset;
 748             mem->flags = 0;
 749
 750             err = kvm_set_user_memory_region(s, mem);
 751             if (err) {
 752                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 753                         strerror(-err));
 754                 abort();
 755             }
 756
 757             start_addr += old.memory_size;
 758             phys_offset += old.memory_size;
 759             size -= old.memory_size;
 760             continue;
 761         }
 762
 763         /* register prefix slot */
 764         if (old.start_addr < start_addr) {
 765             mem = kvm_alloc_slot(s);
 766             mem->memory_size = start_addr - old.start_addr;
 767             mem->start_addr = old.start_addr;
 768             mem->phys_offset = old.phys_offset;
 769             mem->flags = 0;
 770
 771             err = kvm_set_user_memory_region(s, mem);
 772             if (err) {
 773                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 774                         __func__, strerror(-err));
 775                 abort();
 776             }
 777         }
 778
 779         /* register suffix slot */
 780         if (old.start_addr + old.memory_size > start_addr + size) {
 781             ram_addr_t size_delta;
 782
 783             mem = kvm_alloc_slot(s);
 784             mem->start_addr = start_addr + size;
 785             size_delta = mem->start_addr - old.start_addr;
 786             mem->memory_size = old.memory_size - size_delta;
 787             mem->phys_offset = old.phys_offset + size_delta;
 788             mem->flags = 0;
 789
 790             err = kvm_set_user_memory_region(s, mem);
 791             if (err) {
 792                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 793                         __func__, strerror(-err));
 794                 abort();
 795             }
 796         }
 797     }
 798
 799     /* in case the KVM bug workaround already "consumed" the new slot */
 800     if (!size)
 801         return;
 802
 803     /* KVM does not need to know about this memory */
 804     if (flags >= IO_MEM_UNASSIGNED)
 805         return;
 806
 807     mem = kvm_alloc_slot(s);
 808     mem->memory_size = size;
 809     mem->start_addr = start_addr;
 810     mem->phys_offset = phys_offset;
 811     mem->flags = 0;
 812
 813     err = kvm_set_user_memory_region(s, mem);
 814     if (err) {
 815         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 816                 strerror(-err));
 817         abort();
 818     }
 819 }
 820
 821 int kvm_ioctl(KVMState *s, int type, ...)
 822 {
 823     int ret;
 824     void *arg;
 825     va_list ap;
 826
 827     va_start(ap, type);
 828     arg = va_arg(ap, void *);
 829     va_end(ap);
 830
 831     ret = ioctl(s->fd, type, arg);
 832     if (ret == -1)
 833         ret = -errno;
 834
 835     return ret;
 836 }
 837
 838 int kvm_vm_ioctl(KVMState *s, int type, ...)
 839 {
 840     int ret;
 841     void *arg;
 842     va_list ap;
 843
 844     va_start(ap, type);
 845     arg = va_arg(ap, void *);
 846     va_end(ap);
 847
 848     ret = ioctl(s->vmfd, type, arg);
 849     if (ret == -1)
 850         ret = -errno;
 851
 852     return ret;
 853 }
 854
 855 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 856 {
 857     int ret;
 858     void *arg;
 859     va_list ap;
 860
 861     va_start(ap, type);
 862     arg = va_arg(ap, void *);
 863     va_end(ap);
 864
 865     ret = ioctl(env->kvm_fd, type, arg);
 866     if (ret == -1)
 867         ret = -errno;
 868
 869     return ret;
 870 }
 871
 872 int kvm_has_sync_mmu(void)
 873 {
 874 #ifdef KVM_CAP_SYNC_MMU
 875     KVMState *s = kvm_state;
 876
 877     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 878 #else
 879     return 0;
 880 #endif
 881 }
 882
 883 int kvm_has_vcpu_events(void)
 884 {
 885     return kvm_state->vcpu_events;
 886 }
 887
 888 void kvm_setup_guest_memory(void *start, size_t size)
 889 {
 890     if (!kvm_has_sync_mmu()) {
 891 #ifdef MADV_DONTFORK
 892         int ret = madvise(start, size, MADV_DONTFORK);
 893
 894         if (ret) {
 895             perror("madvice");
 896             exit(1);
 897         }
 898 #else
 899         fprintf(stderr,
 900                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 901         exit(1);
 902 #endif
 903     }
 904 }
 905
 906 #ifdef KVM_CAP_SET_GUEST_DEBUG
 907 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 908 {
 909 #ifdef CONFIG_IOTHREAD
 910     if (env == cpu_single_env) {
 911         func(data);
 912         return;
 913     }
 914     abort();
 915 #else
 916     func(data);
 917 #endif
 918 }
 919
 920 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 921                                                  target_ulong pc)
 922 {
 923     struct kvm_sw_breakpoint *bp;
 924
 925     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 926         if (bp->pc == pc)
 927             return bp;
 928     }
 929     return NULL;
 930 }
 931
 932 int kvm_sw_breakpoints_active(CPUState *env)
 933 {
 934     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 935 }
 936
 937 struct kvm_set_guest_debug_data {
 938     struct kvm_guest_debug dbg;
 939     CPUState *env;
 940     int err;
 941 };
 942
 943 static void kvm_invoke_set_guest_debug(void *data)
 944 {
 945     struct kvm_set_guest_debug_data *dbg_data = data;
 946     CPUState *env = dbg_data->env;
 947
 948     if (env->kvm_vcpu_dirty) {
 949         kvm_arch_put_registers(env);
 950         env->kvm_vcpu_dirty = 0;
 951     }
 952     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 953 }
 954
 955 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 956 {
 957     struct kvm_set_guest_debug_data data;
 958
 959     data.dbg.control = 0;
 960     if (env->singlestep_enabled)
 961         data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 962
 963     kvm_arch_update_guest_debug(env, &data.dbg);
 964     data.dbg.control |= reinject_trap;
 965     data.env = env;
 966
 967     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
 968     return data.err;
 969 }
 970
 971 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 972                           target_ulong len, int type)
 973 {
 974     struct kvm_sw_breakpoint *bp;
 975     CPUState *env;
 976     int err;
 977
 978     if (type == GDB_BREAKPOINT_SW) {
 979         bp = kvm_find_sw_breakpoint(current_env, addr);
 980         if (bp) {
 981             bp->use_count++;
 982             return 0;
 983         }
 984
 985         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 986         if (!bp)
 987             return -ENOMEM;
 988
 989         bp->pc = addr;
 990         bp->use_count = 1;
 991         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 992         if (err) {
 993             free(bp);
 994             return err;
 995         }
 996
 997         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 998                           bp, entry);
 999     } else {
1000         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1001         if (err)
1002             return err;
1003     }
1004
1005     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1006         err = kvm_update_guest_debug(env, 0);
1007         if (err)
1008             return err;
1009     }
1010     return 0;
1011 }
1012
1013 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1014                           target_ulong len, int type)
1015 {
1016     struct kvm_sw_breakpoint *bp;
1017     CPUState *env;
1018     int err;
1019
1020     if (type == GDB_BREAKPOINT_SW) {
1021         bp = kvm_find_sw_breakpoint(current_env, addr);
1022         if (!bp)
1023             return -ENOENT;
1024
1025         if (bp->use_count > 1) {
1026             bp->use_count--;
1027             return 0;
1028         }
1029
1030         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1031         if (err)
1032             return err;
1033
1034         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1035         qemu_free(bp);
1036     } else {
1037         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1038         if (err)
1039             return err;
1040     }
1041
1042     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1043         err = kvm_update_guest_debug(env, 0);
1044         if (err)
1045             return err;
1046     }
1047     return 0;
1048 }
1049
1050 void kvm_remove_all_breakpoints(CPUState *current_env)
1051 {
1052     struct kvm_sw_breakpoint *bp, *next;
1053     KVMState *s = current_env->kvm_state;
1054     CPUState *env;
1055
1056     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1057         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1058             /* Try harder to find a CPU that currently sees the breakpoint. */
1059             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1060                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1061                     break;
1062             }
1063         }
1064     }
1065     kvm_arch_remove_all_hw_breakpoints();
1066
1067     for (env = first_cpu; env != NULL; env = env->next_cpu)
1068         kvm_update_guest_debug(env, 0);
1069 }
1070
1071 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1072
1073 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1074 {
1075     return -EINVAL;
1076 }
1077
1078 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1079                           target_ulong len, int type)
1080 {
1081     return -EINVAL;
1082 }
1083
1084 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1085                           target_ulong len, int type)
1086 {
1087     return -EINVAL;
1088 }
1089
1090 void kvm_remove_all_breakpoints(CPUState *current_env)
1091 {
1092 }
1093 #endif /* !KVM_CAP_SET_GUEST_DEBUG */