kvm-all.c

   1 /*
   2  * QEMU KVM support
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *           Red Hat, Inc. 2008
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Glauber Costa     <gcosta@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include <sys/types.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/mman.h>
  19 #include <stdarg.h>
  20
  21 #include <linux/kvm.h>
  22
  23 #include "qemu-common.h"
  24 #include "sysemu.h"
  25 #include "hw/hw.h"
  26 #include "gdbstub.h"
  27 #include "kvm.h"
  28
  29 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  30 #define PAGE_SIZE TARGET_PAGE_SIZE
  31
  32 //#define DEBUG_KVM
  33
  34 #ifdef DEBUG_KVM
  35 #define dprintf(fmt, ...) \
  36     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  37 #else
  38 #define dprintf(fmt, ...) \
  39     do { } while (0)
  40 #endif
  41
  42 typedef struct KVMSlot
  43 {
  44     target_phys_addr_t start_addr;
  45     ram_addr_t memory_size;
  46     ram_addr_t phys_offset;
  47     int slot;
  48     int flags;
  49 } KVMSlot;
  50
  51 typedef struct kvm_dirty_log KVMDirtyLog;
  52
  53 int kvm_allowed = 0;
  54
  55 struct KVMState
  56 {
  57     KVMSlot slots[32];
  58     int fd;
  59     int vmfd;
  60     int coalesced_mmio;
  61     int broken_set_mem_region;
  62     int migration_log;
  63 #ifdef KVM_CAP_SET_GUEST_DEBUG
  64     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  65 #endif
  66     int irqchip_in_kernel;
  67     int pit_in_kernel;
  68 };
  69
  70 static KVMState *kvm_state;
  71
  72 static KVMSlot *kvm_alloc_slot(KVMState *s)
  73 {
  74     int i;
  75
  76     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  77         /* KVM private memory slots */
  78         if (i >= 8 && i < 12)
  79             continue;
  80         if (s->slots[i].memory_size == 0)
  81             return &s->slots[i];
  82     }
  83
  84     fprintf(stderr, "%s: no free slot available\n", __func__);
  85     abort();
  86 }
  87
  88 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  89                                          target_phys_addr_t start_addr,
  90                                          target_phys_addr_t end_addr)
  91 {
  92     int i;
  93
  94     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  95         KVMSlot *mem = &s->slots[i];
  96
  97         if (start_addr == mem->start_addr &&
  98             end_addr == mem->start_addr + mem->memory_size) {
  99             return mem;
 100         }
 101     }
 102
 103     return NULL;
 104 }
 105
 106 /*
 107  * Find overlapping slot with lowest start address
 108  */
 109 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 110                                             target_phys_addr_t start_addr,
 111                                             target_phys_addr_t end_addr)
 112 {
 113     KVMSlot *found = NULL;
 114     int i;
 115
 116     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 117         KVMSlot *mem = &s->slots[i];
 118
 119         if (mem->memory_size == 0 ||
 120             (found && found->start_addr < mem->start_addr)) {
 121             continue;
 122         }
 123
 124         if (end_addr > mem->start_addr &&
 125             start_addr < mem->start_addr + mem->memory_size) {
 126             found = mem;
 127         }
 128     }
 129
 130     return found;
 131 }
 132
 133 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 134 {
 135     struct kvm_userspace_memory_region mem;
 136
 137     mem.slot = slot->slot;
 138     mem.guest_phys_addr = slot->start_addr;
 139     mem.memory_size = slot->memory_size;
 140     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 141     mem.flags = slot->flags;
 142     if (s->migration_log) {
 143         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 144     }
 145     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 146 }
 147
 148 static void kvm_reset_vcpu(void *opaque)
 149 {
 150     CPUState *env = opaque;
 151
 152     if (kvm_arch_put_registers(env)) {
 153         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 154         abort();
 155     }
 156 }
 157
 158 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 159 {
 160     if (env == cpu_single_env) {
 161         func(data);
 162         return;
 163     }
 164     abort();
 165 }
 166
 167 int kvm_irqchip_in_kernel(void)
 168 {
 169     return kvm_state->irqchip_in_kernel;
 170 }
 171
 172 int kvm_pit_in_kernel(void)
 173 {
 174     return kvm_state->pit_in_kernel;
 175 }
 176
 177
 178 int kvm_init_vcpu(CPUState *env)
 179 {
 180     KVMState *s = kvm_state;
 181     long mmap_size;
 182     int ret;
 183
 184     dprintf("kvm_init_vcpu\n");
 185
 186     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 187     if (ret < 0) {
 188         dprintf("kvm_create_vcpu failed\n");
 189         goto err;
 190     }
 191
 192     env->kvm_fd = ret;
 193     env->kvm_state = s;
 194
 195     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 196     if (mmap_size < 0) {
 197         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 198         goto err;
 199     }
 200
 201     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 202                         env->kvm_fd, 0);
 203     if (env->kvm_run == MAP_FAILED) {
 204         ret = -errno;
 205         dprintf("mmap'ing vcpu state failed\n");
 206         goto err;
 207     }
 208
 209     ret = kvm_arch_init_vcpu(env);
 210     if (ret == 0) {
 211         qemu_register_reset(kvm_reset_vcpu, env);
 212         ret = kvm_arch_put_registers(env);
 213     }
 214 err:
 215     return ret;
 216 }
 217
 218 int kvm_put_mp_state(CPUState *env)
 219 {
 220     struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
 221
 222     return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
 223 }
 224
 225 int kvm_get_mp_state(CPUState *env)
 226 {
 227     struct kvm_mp_state mp_state;
 228     int ret;
 229
 230     ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
 231     if (ret < 0) {
 232         return ret;
 233     }
 234     env->mp_state = mp_state.mp_state;
 235     return 0;
 236 }
 237
 238 /*
 239  * dirty pages logging control
 240  */
 241 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 242                                       ram_addr_t size, int flags, int mask)
 243 {
 244     KVMState *s = kvm_state;
 245     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 246     int old_flags;
 247
 248     if (mem == NULL)  {
 249             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 250                     TARGET_FMT_plx "\n", __func__, phys_addr,
 251                     (target_phys_addr_t)(phys_addr + size - 1));
 252             return -EINVAL;
 253     }
 254
 255     old_flags = mem->flags;
 256
 257     flags = (mem->flags & ~mask) | flags;
 258     mem->flags = flags;
 259
 260     /* If nothing changed effectively, no need to issue ioctl */
 261     if (s->migration_log) {
 262         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 263     }
 264     if (flags == old_flags) {
 265             return 0;
 266     }
 267
 268     return kvm_set_user_memory_region(s, mem);
 269 }
 270
 271 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 272 {
 273         return kvm_dirty_pages_log_change(phys_addr, size,
 274                                           KVM_MEM_LOG_DIRTY_PAGES,
 275                                           KVM_MEM_LOG_DIRTY_PAGES);
 276 }
 277
 278 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 279 {
 280         return kvm_dirty_pages_log_change(phys_addr, size,
 281                                           0,
 282                                           KVM_MEM_LOG_DIRTY_PAGES);
 283 }
 284
 285 int kvm_set_migration_log(int enable)
 286 {
 287     KVMState *s = kvm_state;
 288     KVMSlot *mem;
 289     int i, err;
 290
 291     s->migration_log = enable;
 292
 293     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 294         mem = &s->slots[i];
 295
 296         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 297             continue;
 298         }
 299         err = kvm_set_user_memory_region(s, mem);
 300         if (err) {
 301             return err;
 302         }
 303     }
 304     return 0;
 305 }
 306
 307 /**
 308  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 309  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 310  * This means all bits are set to dirty.
 311  *
 312  * @start_add: start of logged region.
 313  * @end_addr: end of logged region.
 314  */
 315 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 316                                    target_phys_addr_t end_addr)
 317 {
 318     KVMState *s = kvm_state;
 319     unsigned long size, allocated_size = 0;
 320     target_phys_addr_t phys_addr;
 321     ram_addr_t addr;
 322     KVMDirtyLog d;
 323     KVMSlot *mem;
 324     int ret = 0;
 325     int r;
 326
 327     d.dirty_bitmap = NULL;
 328     while (start_addr < end_addr) {
 329         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 330         if (mem == NULL) {
 331             break;
 332         }
 333
 334         /* We didn't activate dirty logging? Don't care then. */
 335         if(!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES)) {
 336             continue;
 337         }
 338
 339         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 340         if (!d.dirty_bitmap) {
 341             d.dirty_bitmap = qemu_malloc(size);
 342         } else if (size > allocated_size) {
 343             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 344         }
 345         allocated_size = size;
 346         memset(d.dirty_bitmap, 0, allocated_size);
 347
 348         d.slot = mem->slot;
 349
 350         r = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
 351         if (r == -EINVAL) {
 352             dprintf("ioctl failed %d\n", errno);
 353             ret = -1;
 354             break;
 355         }
 356
 357         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 358              phys_addr < mem->start_addr + mem->memory_size;
 359              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 360             unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
 361             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 362             unsigned word = nr / (sizeof(*bitmap) * 8);
 363             unsigned bit = nr % (sizeof(*bitmap) * 8);
 364
 365             if ((bitmap[word] >> bit) & 1) {
 366                 cpu_physical_memory_set_dirty(addr);
 367             } else if (r < 0) {
 368                 /* When our KVM implementation doesn't know about dirty logging
 369                  * we can just assume it's always dirty and be fine. */
 370                 cpu_physical_memory_set_dirty(addr);
 371             }
 372         }
 373         start_addr = phys_addr;
 374     }
 375     qemu_free(d.dirty_bitmap);
 376
 377     return ret;
 378 }
 379
 380 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 381 {
 382     int ret = -ENOSYS;
 383 #ifdef KVM_CAP_COALESCED_MMIO
 384     KVMState *s = kvm_state;
 385
 386     if (s->coalesced_mmio) {
 387         struct kvm_coalesced_mmio_zone zone;
 388
 389         zone.addr = start;
 390         zone.size = size;
 391
 392         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 393     }
 394 #endif
 395
 396     return ret;
 397 }
 398
 399 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 400 {
 401     int ret = -ENOSYS;
 402 #ifdef KVM_CAP_COALESCED_MMIO
 403     KVMState *s = kvm_state;
 404
 405     if (s->coalesced_mmio) {
 406         struct kvm_coalesced_mmio_zone zone;
 407
 408         zone.addr = start;
 409         zone.size = size;
 410
 411         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 412     }
 413 #endif
 414
 415     return ret;
 416 }
 417
 418 int kvm_check_extension(KVMState *s, unsigned int extension)
 419 {
 420     int ret;
 421
 422     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 423     if (ret < 0) {
 424         ret = 0;
 425     }
 426
 427     return ret;
 428 }
 429
 430 int kvm_init(int smp_cpus)
 431 {
 432     static const char upgrade_note[] =
 433         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 434         "(see http://sourceforge.net/projects/kvm).\n";
 435     KVMState *s;
 436     int ret;
 437     int i;
 438
 439     if (smp_cpus > 1) {
 440         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 441         return -EINVAL;
 442     }
 443
 444     s = qemu_mallocz(sizeof(KVMState));
 445
 446 #ifdef KVM_CAP_SET_GUEST_DEBUG
 447     TAILQ_INIT(&s->kvm_sw_breakpoints);
 448 #endif
 449     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 450         s->slots[i].slot = i;
 451
 452     s->vmfd = -1;
 453     s->fd = open("/dev/kvm", O_RDWR);
 454     if (s->fd == -1) {
 455         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 456         ret = -errno;
 457         goto err;
 458     }
 459
 460     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 461     if (ret < KVM_API_VERSION) {
 462         if (ret > 0)
 463             ret = -EINVAL;
 464         fprintf(stderr, "kvm version too old\n");
 465         goto err;
 466     }
 467
 468     if (ret > KVM_API_VERSION) {
 469         ret = -EINVAL;
 470         fprintf(stderr, "kvm version not supported\n");
 471         goto err;
 472     }
 473
 474     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 475     if (s->vmfd < 0)
 476         goto err;
 477
 478     /* initially, KVM allocated its own memory and we had to jump through
 479      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 480      * just use a user allocated buffer so we can use regular pages
 481      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 482      */
 483     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 484         ret = -EINVAL;
 485         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 486                 upgrade_note);
 487         goto err;
 488     }
 489
 490     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 491      * destroyed properly.  Since we rely on this capability, refuse to work
 492      * with any kernel without this capability. */
 493     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 494         ret = -EINVAL;
 495
 496         fprintf(stderr,
 497                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 498                 upgrade_note);
 499         goto err;
 500     }
 501
 502 #ifdef KVM_CAP_COALESCED_MMIO
 503     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 504 #else
 505     s->coalesced_mmio = 0;
 506 #endif
 507
 508     s->broken_set_mem_region = 1;
 509 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 510     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 511     if (ret > 0) {
 512         s->broken_set_mem_region = 0;
 513     }
 514 #endif
 515
 516     ret = kvm_arch_init(s, smp_cpus);
 517     if (ret < 0)
 518         goto err;
 519
 520     kvm_state = s;
 521
 522     return 0;
 523
 524 err:
 525     if (s) {
 526         if (s->vmfd != -1)
 527             close(s->vmfd);
 528         if (s->fd != -1)
 529             close(s->fd);
 530     }
 531     qemu_free(s);
 532
 533     return ret;
 534 }
 535
 536 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 537                          int direction, int size, uint32_t count)
 538 {
 539     int i;
 540     uint8_t *ptr = data;
 541
 542     for (i = 0; i < count; i++) {
 543         if (direction == KVM_EXIT_IO_IN) {
 544             switch (size) {
 545             case 1:
 546                 stb_p(ptr, cpu_inb(env, port));
 547                 break;
 548             case 2:
 549                 stw_p(ptr, cpu_inw(env, port));
 550                 break;
 551             case 4:
 552                 stl_p(ptr, cpu_inl(env, port));
 553                 break;
 554             }
 555         } else {
 556             switch (size) {
 557             case 1:
 558                 cpu_outb(env, port, ldub_p(ptr));
 559                 break;
 560             case 2:
 561                 cpu_outw(env, port, lduw_p(ptr));
 562                 break;
 563             case 4:
 564                 cpu_outl(env, port, ldl_p(ptr));
 565                 break;
 566             }
 567         }
 568
 569         ptr += size;
 570     }
 571
 572     return 1;
 573 }
 574
 575 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 576 {
 577 #ifdef KVM_CAP_COALESCED_MMIO
 578     KVMState *s = kvm_state;
 579     if (s->coalesced_mmio) {
 580         struct kvm_coalesced_mmio_ring *ring;
 581
 582         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 583         while (ring->first != ring->last) {
 584             struct kvm_coalesced_mmio *ent;
 585
 586             ent = &ring->coalesced_mmio[ring->first];
 587
 588             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 589             /* FIXME smp_wmb() */
 590             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 591         }
 592     }
 593 #endif
 594 }
 595
 596 int kvm_cpu_exec(CPUState *env)
 597 {
 598     struct kvm_run *run = env->kvm_run;
 599     int ret;
 600
 601     dprintf("kvm_cpu_exec()\n");
 602
 603     do {
 604         if (env->exit_request) {
 605             dprintf("interrupt exit requested\n");
 606             ret = 0;
 607             break;
 608         }
 609
 610         kvm_arch_pre_run(env, run);
 611         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 612         kvm_arch_post_run(env, run);
 613
 614         if (ret == -EINTR || ret == -EAGAIN) {
 615             dprintf("io window exit\n");
 616             ret = 0;
 617             break;
 618         }
 619
 620         if (ret < 0) {
 621             dprintf("kvm run failed %s\n", strerror(-ret));
 622             abort();
 623         }
 624
 625         kvm_run_coalesced_mmio(env, run);
 626
 627         ret = 0; /* exit loop */
 628         switch (run->exit_reason) {
 629         case KVM_EXIT_IO:
 630             dprintf("handle_io\n");
 631             ret = kvm_handle_io(env, run->io.port,
 632                                 (uint8_t *)run + run->io.data_offset,
 633                                 run->io.direction,
 634                                 run->io.size,
 635                                 run->io.count);
 636             break;
 637         case KVM_EXIT_MMIO:
 638             dprintf("handle_mmio\n");
 639             cpu_physical_memory_rw(run->mmio.phys_addr,
 640                                    run->mmio.data,
 641                                    run->mmio.len,
 642                                    run->mmio.is_write);
 643             ret = 1;
 644             break;
 645         case KVM_EXIT_IRQ_WINDOW_OPEN:
 646             dprintf("irq_window_open\n");
 647             break;
 648         case KVM_EXIT_SHUTDOWN:
 649             dprintf("shutdown\n");
 650             qemu_system_reset_request();
 651             ret = 1;
 652             break;
 653         case KVM_EXIT_UNKNOWN:
 654             dprintf("kvm_exit_unknown\n");
 655             break;
 656         case KVM_EXIT_FAIL_ENTRY:
 657             dprintf("kvm_exit_fail_entry\n");
 658             break;
 659         case KVM_EXIT_EXCEPTION:
 660             dprintf("kvm_exit_exception\n");
 661             break;
 662         case KVM_EXIT_DEBUG:
 663             dprintf("kvm_exit_debug\n");
 664 #ifdef KVM_CAP_SET_GUEST_DEBUG
 665             if (kvm_arch_debug(&run->debug.arch)) {
 666                 gdb_set_stop_cpu(env);
 667                 vm_stop(EXCP_DEBUG);
 668                 env->exception_index = EXCP_DEBUG;
 669                 return 0;
 670             }
 671             /* re-enter, this exception was guest-internal */
 672             ret = 1;
 673 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 674             break;
 675         default:
 676             dprintf("kvm_arch_handle_exit\n");
 677             ret = kvm_arch_handle_exit(env, run);
 678             break;
 679         }
 680     } while (ret > 0);
 681
 682     if (env->exit_request) {
 683         env->exit_request = 0;
 684         env->exception_index = EXCP_INTERRUPT;
 685     }
 686
 687     return ret;
 688 }
 689
 690 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 691                       ram_addr_t size,
 692                       ram_addr_t phys_offset)
 693 {
 694     KVMState *s = kvm_state;
 695     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 696     KVMSlot *mem, old;
 697     int err;
 698
 699     if (start_addr & ~TARGET_PAGE_MASK) {
 700         if (flags >= IO_MEM_UNASSIGNED) {
 701             if (!kvm_lookup_overlapping_slot(s, start_addr,
 702                                              start_addr + size)) {
 703                 return;
 704             }
 705             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 706         } else {
 707             fprintf(stderr, "Only page-aligned memory slots supported\n");
 708         }
 709         abort();
 710     }
 711
 712     /* KVM does not support read-only slots */
 713     phys_offset &= ~IO_MEM_ROM;
 714
 715     while (1) {
 716         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 717         if (!mem) {
 718             break;
 719         }
 720
 721         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 722             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 723             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 724             /* The new slot fits into the existing one and comes with
 725              * identical parameters - nothing to be done. */
 726             return;
 727         }
 728
 729         old = *mem;
 730
 731         /* unregister the overlapping slot */
 732         mem->memory_size = 0;
 733         err = kvm_set_user_memory_region(s, mem);
 734         if (err) {
 735             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 736                     __func__, strerror(-err));
 737             abort();
 738         }
 739
 740         /* Workaround for older KVM versions: we can't join slots, even not by
 741          * unregistering the previous ones and then registering the larger
 742          * slot. We have to maintain the existing fragmentation. Sigh.
 743          *
 744          * This workaround assumes that the new slot starts at the same
 745          * address as the first existing one. If not or if some overlapping
 746          * slot comes around later, we will fail (not seen in practice so far)
 747          * - and actually require a recent KVM version. */
 748         if (s->broken_set_mem_region &&
 749             old.start_addr == start_addr && old.memory_size < size &&
 750             flags < IO_MEM_UNASSIGNED) {
 751             mem = kvm_alloc_slot(s);
 752             mem->memory_size = old.memory_size;
 753             mem->start_addr = old.start_addr;
 754             mem->phys_offset = old.phys_offset;
 755             mem->flags = 0;
 756
 757             err = kvm_set_user_memory_region(s, mem);
 758             if (err) {
 759                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 760                         strerror(-err));
 761                 abort();
 762             }
 763
 764             start_addr += old.memory_size;
 765             phys_offset += old.memory_size;
 766             size -= old.memory_size;
 767             continue;
 768         }
 769
 770         /* register prefix slot */
 771         if (old.start_addr < start_addr) {
 772             mem = kvm_alloc_slot(s);
 773             mem->memory_size = start_addr - old.start_addr;
 774             mem->start_addr = old.start_addr;
 775             mem->phys_offset = old.phys_offset;
 776             mem->flags = 0;
 777
 778             err = kvm_set_user_memory_region(s, mem);
 779             if (err) {
 780                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 781                         __func__, strerror(-err));
 782                 abort();
 783             }
 784         }
 785
 786         /* register suffix slot */
 787         if (old.start_addr + old.memory_size > start_addr + size) {
 788             ram_addr_t size_delta;
 789
 790             mem = kvm_alloc_slot(s);
 791             mem->start_addr = start_addr + size;
 792             size_delta = mem->start_addr - old.start_addr;
 793             mem->memory_size = old.memory_size - size_delta;
 794             mem->phys_offset = old.phys_offset + size_delta;
 795             mem->flags = 0;
 796
 797             err = kvm_set_user_memory_region(s, mem);
 798             if (err) {
 799                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 800                         __func__, strerror(-err));
 801                 abort();
 802             }
 803         }
 804     }
 805
 806     /* in case the KVM bug workaround already "consumed" the new slot */
 807     if (!size)
 808         return;
 809
 810     /* KVM does not need to know about this memory */
 811     if (flags >= IO_MEM_UNASSIGNED)
 812         return;
 813
 814     mem = kvm_alloc_slot(s);
 815     mem->memory_size = size;
 816     mem->start_addr = start_addr;
 817     mem->phys_offset = phys_offset;
 818     mem->flags = 0;
 819
 820     err = kvm_set_user_memory_region(s, mem);
 821     if (err) {
 822         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 823                 strerror(-err));
 824         abort();
 825     }
 826 }
 827
 828 int kvm_ioctl(KVMState *s, int type, ...)
 829 {
 830     int ret;
 831     void *arg;
 832     va_list ap;
 833
 834     va_start(ap, type);
 835     arg = va_arg(ap, void *);
 836     va_end(ap);
 837
 838     ret = ioctl(s->fd, type, arg);
 839     if (ret == -1)
 840         ret = -errno;
 841
 842     return ret;
 843 }
 844
 845 int kvm_vm_ioctl(KVMState *s, int type, ...)
 846 {
 847     int ret;
 848     void *arg;
 849     va_list ap;
 850
 851     va_start(ap, type);
 852     arg = va_arg(ap, void *);
 853     va_end(ap);
 854
 855     ret = ioctl(s->vmfd, type, arg);
 856     if (ret == -1)
 857         ret = -errno;
 858
 859     return ret;
 860 }
 861
 862 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 863 {
 864     int ret;
 865     void *arg;
 866     va_list ap;
 867
 868     va_start(ap, type);
 869     arg = va_arg(ap, void *);
 870     va_end(ap);
 871
 872     ret = ioctl(env->kvm_fd, type, arg);
 873     if (ret == -1)
 874         ret = -errno;
 875
 876     return ret;
 877 }
 878
 879 int kvm_has_sync_mmu(void)
 880 {
 881 #ifdef KVM_CAP_SYNC_MMU
 882     KVMState *s = kvm_state;
 883
 884     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 885 #else
 886     return 0;
 887 #endif
 888 }
 889
 890 void kvm_setup_guest_memory(void *start, size_t size)
 891 {
 892     if (!kvm_has_sync_mmu()) {
 893 #ifdef MADV_DONTFORK
 894         int ret = madvise(start, size, MADV_DONTFORK);
 895
 896         if (ret) {
 897             perror("madvice");
 898             exit(1);
 899         }
 900 #else
 901         fprintf(stderr,
 902                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 903         exit(1);
 904 #endif
 905     }
 906 }
 907
 908 #ifdef KVM_CAP_SET_GUEST_DEBUG
 909 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 910                                                  target_ulong pc)
 911 {
 912     struct kvm_sw_breakpoint *bp;
 913
 914     TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 915         if (bp->pc == pc)
 916             return bp;
 917     }
 918     return NULL;
 919 }
 920
 921 int kvm_sw_breakpoints_active(CPUState *env)
 922 {
 923     return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 924 }
 925
 926 struct kvm_set_guest_debug_data {
 927     struct kvm_guest_debug dbg;
 928     CPUState *env;
 929     int err;
 930 };
 931
 932 static void kvm_invoke_set_guest_debug(void *data)
 933 {
 934     struct kvm_set_guest_debug_data *dbg_data = data;
 935     dbg_data->err = kvm_vcpu_ioctl(dbg_data->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         TAILQ_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         TAILQ_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     TAILQ_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 */