kvm-all.c

   1 /*
   2  * QEMU KVM support
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *           Red Hat, Inc. 2008
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Glauber Costa     <gcosta@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include <sys/types.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/mman.h>
  19 #include <stdarg.h>
  20
  21 #include <linux/kvm.h>
  22
  23 #include "qemu-common.h"
  24 #include "sysemu.h"
  25 #include "hw/hw.h"
  26 #include "gdbstub.h"
  27 #include "kvm.h"
  28
  29 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  30 #define PAGE_SIZE TARGET_PAGE_SIZE
  31
  32 //#define DEBUG_KVM
  33
  34 #ifdef DEBUG_KVM
  35 #define dprintf(fmt, ...) \
  36     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  37 #else
  38 #define dprintf(fmt, ...) \
  39     do { } while (0)
  40 #endif
  41
  42 typedef struct KVMSlot
  43 {
  44     target_phys_addr_t start_addr;
  45     ram_addr_t memory_size;
  46     ram_addr_t phys_offset;
  47     int slot;
  48     int flags;
  49 } KVMSlot;
  50
  51 typedef struct kvm_dirty_log KVMDirtyLog;
  52
  53 int kvm_allowed = 0;
  54
  55 struct KVMState
  56 {
  57     KVMSlot slots[32];
  58     int fd;
  59     int vmfd;
  60     int regs_modified;
  61     int coalesced_mmio;
  62     int broken_set_mem_region;
  63     int migration_log;
  64 #ifdef KVM_CAP_SET_GUEST_DEBUG
  65     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  66 #endif
  67     int irqchip_in_kernel;
  68     int pit_in_kernel;
  69 };
  70
  71 static KVMState *kvm_state;
  72
  73 static KVMSlot *kvm_alloc_slot(KVMState *s)
  74 {
  75     int i;
  76
  77     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  78         /* KVM private memory slots */
  79         if (i >= 8 && i < 12)
  80             continue;
  81         if (s->slots[i].memory_size == 0)
  82             return &s->slots[i];
  83     }
  84
  85     fprintf(stderr, "%s: no free slot available\n", __func__);
  86     abort();
  87 }
  88
  89 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  90                                          target_phys_addr_t start_addr,
  91                                          target_phys_addr_t end_addr)
  92 {
  93     int i;
  94
  95     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  96         KVMSlot *mem = &s->slots[i];
  97
  98         if (start_addr == mem->start_addr &&
  99             end_addr == mem->start_addr + mem->memory_size) {
 100             return mem;
 101         }
 102     }
 103
 104     return NULL;
 105 }
 106
 107 /*
 108  * Find overlapping slot with lowest start address
 109  */
 110 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 111                                             target_phys_addr_t start_addr,
 112                                             target_phys_addr_t end_addr)
 113 {
 114     KVMSlot *found = NULL;
 115     int i;
 116
 117     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 118         KVMSlot *mem = &s->slots[i];
 119
 120         if (mem->memory_size == 0 ||
 121             (found && found->start_addr < mem->start_addr)) {
 122             continue;
 123         }
 124
 125         if (end_addr > mem->start_addr &&
 126             start_addr < mem->start_addr + mem->memory_size) {
 127             found = mem;
 128         }
 129     }
 130
 131     return found;
 132 }
 133
 134 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 135 {
 136     struct kvm_userspace_memory_region mem;
 137
 138     mem.slot = slot->slot;
 139     mem.guest_phys_addr = slot->start_addr;
 140     mem.memory_size = slot->memory_size;
 141     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 142     mem.flags = slot->flags;
 143     if (s->migration_log) {
 144         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 145     }
 146     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 147 }
 148
 149 static void kvm_reset_vcpu(void *opaque)
 150 {
 151     CPUState *env = opaque;
 152
 153     if (kvm_arch_put_registers(env)) {
 154         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 155         abort();
 156     }
 157 }
 158
 159 int kvm_irqchip_in_kernel(void)
 160 {
 161     return kvm_state->irqchip_in_kernel;
 162 }
 163
 164 int kvm_pit_in_kernel(void)
 165 {
 166     return kvm_state->pit_in_kernel;
 167 }
 168
 169
 170 int kvm_init_vcpu(CPUState *env)
 171 {
 172     KVMState *s = kvm_state;
 173     long mmap_size;
 174     int ret;
 175
 176     dprintf("kvm_init_vcpu\n");
 177
 178     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 179     if (ret < 0) {
 180         dprintf("kvm_create_vcpu failed\n");
 181         goto err;
 182     }
 183
 184     env->kvm_fd = ret;
 185     env->kvm_state = s;
 186
 187     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 188     if (mmap_size < 0) {
 189         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 190         goto err;
 191     }
 192
 193     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 194                         env->kvm_fd, 0);
 195     if (env->kvm_run == MAP_FAILED) {
 196         ret = -errno;
 197         dprintf("mmap'ing vcpu state failed\n");
 198         goto err;
 199     }
 200
 201     ret = kvm_arch_init_vcpu(env);
 202     if (ret == 0) {
 203         qemu_register_reset(kvm_reset_vcpu, env);
 204         ret = kvm_arch_put_registers(env);
 205     }
 206 err:
 207     return ret;
 208 }
 209
 210 /*
 211  * dirty pages logging control
 212  */
 213 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 214                                       ram_addr_t size, int flags, int mask)
 215 {
 216     KVMState *s = kvm_state;
 217     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 218     int old_flags;
 219
 220     if (mem == NULL)  {
 221             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 222                     TARGET_FMT_plx "\n", __func__, phys_addr,
 223                     (target_phys_addr_t)(phys_addr + size - 1));
 224             return -EINVAL;
 225     }
 226
 227     old_flags = mem->flags;
 228
 229     flags = (mem->flags & ~mask) | flags;
 230     mem->flags = flags;
 231
 232     /* If nothing changed effectively, no need to issue ioctl */
 233     if (s->migration_log) {
 234         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 235     }
 236     if (flags == old_flags) {
 237             return 0;
 238     }
 239
 240     return kvm_set_user_memory_region(s, mem);
 241 }
 242
 243 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 244 {
 245         return kvm_dirty_pages_log_change(phys_addr, size,
 246                                           KVM_MEM_LOG_DIRTY_PAGES,
 247                                           KVM_MEM_LOG_DIRTY_PAGES);
 248 }
 249
 250 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 251 {
 252         return kvm_dirty_pages_log_change(phys_addr, size,
 253                                           0,
 254                                           KVM_MEM_LOG_DIRTY_PAGES);
 255 }
 256
 257 int kvm_set_migration_log(int enable)
 258 {
 259     KVMState *s = kvm_state;
 260     KVMSlot *mem;
 261     int i, err;
 262
 263     s->migration_log = enable;
 264
 265     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 266         mem = &s->slots[i];
 267
 268         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 269             continue;
 270         }
 271         err = kvm_set_user_memory_region(s, mem);
 272         if (err) {
 273             return err;
 274         }
 275     }
 276     return 0;
 277 }
 278
 279 static int test_le_bit(unsigned long nr, unsigned char *addr)
 280 {
 281     return (addr[nr >> 3] >> (nr & 7)) & 1;
 282 }
 283
 284 /**
 285  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 286  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 287  * This means all bits are set to dirty.
 288  *
 289  * @start_add: start of logged region.
 290  * @end_addr: end of logged region.
 291  */
 292 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 293                                    target_phys_addr_t end_addr)
 294 {
 295     KVMState *s = kvm_state;
 296     unsigned long size, allocated_size = 0;
 297     target_phys_addr_t phys_addr;
 298     ram_addr_t addr;
 299     KVMDirtyLog d;
 300     KVMSlot *mem;
 301     int ret = 0;
 302
 303     d.dirty_bitmap = NULL;
 304     while (start_addr < end_addr) {
 305         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 306         if (mem == NULL) {
 307             break;
 308         }
 309
 310         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 311         if (!d.dirty_bitmap) {
 312             d.dirty_bitmap = qemu_malloc(size);
 313         } else if (size > allocated_size) {
 314             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 315         }
 316         allocated_size = size;
 317         memset(d.dirty_bitmap, 0, allocated_size);
 318
 319         d.slot = mem->slot;
 320
 321         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 322             dprintf("ioctl failed %d\n", errno);
 323             ret = -1;
 324             break;
 325         }
 326
 327         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 328              phys_addr < mem->start_addr + mem->memory_size;
 329              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 330             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 331             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 332
 333             if (test_le_bit(nr, bitmap)) {
 334                 cpu_physical_memory_set_dirty(addr);
 335             }
 336         }
 337         start_addr = phys_addr;
 338     }
 339     qemu_free(d.dirty_bitmap);
 340
 341     return ret;
 342 }
 343
 344 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 345 {
 346     int ret = -ENOSYS;
 347 #ifdef KVM_CAP_COALESCED_MMIO
 348     KVMState *s = kvm_state;
 349
 350     if (s->coalesced_mmio) {
 351         struct kvm_coalesced_mmio_zone zone;
 352
 353         zone.addr = start;
 354         zone.size = size;
 355
 356         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 357     }
 358 #endif
 359
 360     return ret;
 361 }
 362
 363 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 364 {
 365     int ret = -ENOSYS;
 366 #ifdef KVM_CAP_COALESCED_MMIO
 367     KVMState *s = kvm_state;
 368
 369     if (s->coalesced_mmio) {
 370         struct kvm_coalesced_mmio_zone zone;
 371
 372         zone.addr = start;
 373         zone.size = size;
 374
 375         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 376     }
 377 #endif
 378
 379     return ret;
 380 }
 381
 382 int kvm_check_extension(KVMState *s, unsigned int extension)
 383 {
 384     int ret;
 385
 386     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 387     if (ret < 0) {
 388         ret = 0;
 389     }
 390
 391     return ret;
 392 }
 393
 394 int kvm_init(int smp_cpus)
 395 {
 396     static const char upgrade_note[] =
 397         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 398         "(see http://sourceforge.net/projects/kvm).\n";
 399     KVMState *s;
 400     int ret;
 401     int i;
 402
 403     if (smp_cpus > 1) {
 404         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 405         return -EINVAL;
 406     }
 407
 408     s = qemu_mallocz(sizeof(KVMState));
 409
 410 #ifdef KVM_CAP_SET_GUEST_DEBUG
 411     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 412 #endif
 413     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 414         s->slots[i].slot = i;
 415
 416     s->vmfd = -1;
 417     s->fd = open("/dev/kvm", O_RDWR);
 418     if (s->fd == -1) {
 419         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 420         ret = -errno;
 421         goto err;
 422     }
 423
 424     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 425     if (ret < KVM_API_VERSION) {
 426         if (ret > 0)
 427             ret = -EINVAL;
 428         fprintf(stderr, "kvm version too old\n");
 429         goto err;
 430     }
 431
 432     if (ret > KVM_API_VERSION) {
 433         ret = -EINVAL;
 434         fprintf(stderr, "kvm version not supported\n");
 435         goto err;
 436     }
 437
 438     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 439     if (s->vmfd < 0)
 440         goto err;
 441
 442     /* initially, KVM allocated its own memory and we had to jump through
 443      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 444      * just use a user allocated buffer so we can use regular pages
 445      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 446      */
 447     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 448         ret = -EINVAL;
 449         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 450                 upgrade_note);
 451         goto err;
 452     }
 453
 454     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 455      * destroyed properly.  Since we rely on this capability, refuse to work
 456      * with any kernel without this capability. */
 457     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 458         ret = -EINVAL;
 459
 460         fprintf(stderr,
 461                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 462                 upgrade_note);
 463         goto err;
 464     }
 465
 466 #ifdef KVM_CAP_COALESCED_MMIO
 467     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 468 #else
 469     s->coalesced_mmio = 0;
 470 #endif
 471
 472     s->broken_set_mem_region = 1;
 473 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 474     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 475     if (ret > 0) {
 476         s->broken_set_mem_region = 0;
 477     }
 478 #endif
 479
 480     ret = kvm_arch_init(s, smp_cpus);
 481     if (ret < 0)
 482         goto err;
 483
 484     kvm_state = s;
 485
 486     return 0;
 487
 488 err:
 489     if (s) {
 490         if (s->vmfd != -1)
 491             close(s->vmfd);
 492         if (s->fd != -1)
 493             close(s->fd);
 494     }
 495     qemu_free(s);
 496
 497     return ret;
 498 }
 499
 500 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 501                          uint32_t count)
 502 {
 503     int i;
 504     uint8_t *ptr = data;
 505
 506     for (i = 0; i < count; i++) {
 507         if (direction == KVM_EXIT_IO_IN) {
 508             switch (size) {
 509             case 1:
 510                 stb_p(ptr, cpu_inb(port));
 511                 break;
 512             case 2:
 513                 stw_p(ptr, cpu_inw(port));
 514                 break;
 515             case 4:
 516                 stl_p(ptr, cpu_inl(port));
 517                 break;
 518             }
 519         } else {
 520             switch (size) {
 521             case 1:
 522                 cpu_outb(port, ldub_p(ptr));
 523                 break;
 524             case 2:
 525                 cpu_outw(port, lduw_p(ptr));
 526                 break;
 527             case 4:
 528                 cpu_outl(port, ldl_p(ptr));
 529                 break;
 530             }
 531         }
 532
 533         ptr += size;
 534     }
 535
 536     return 1;
 537 }
 538
 539 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 540 {
 541 #ifdef KVM_CAP_COALESCED_MMIO
 542     KVMState *s = kvm_state;
 543     if (s->coalesced_mmio) {
 544         struct kvm_coalesced_mmio_ring *ring;
 545
 546         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 547         while (ring->first != ring->last) {
 548             struct kvm_coalesced_mmio *ent;
 549
 550             ent = &ring->coalesced_mmio[ring->first];
 551
 552             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 553             /* FIXME smp_wmb() */
 554             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 555         }
 556     }
 557 #endif
 558 }
 559
 560 void kvm_cpu_synchronize_state(CPUState *env)
 561 {
 562     if (!env->kvm_state->regs_modified) {
 563         kvm_arch_get_registers(env);
 564         env->kvm_state->regs_modified = 1;
 565     }
 566 }
 567
 568 int kvm_cpu_exec(CPUState *env)
 569 {
 570     struct kvm_run *run = env->kvm_run;
 571     int ret;
 572
 573     dprintf("kvm_cpu_exec()\n");
 574
 575     do {
 576         if (env->exit_request) {
 577             dprintf("interrupt exit requested\n");
 578             ret = 0;
 579             break;
 580         }
 581
 582         if (env->kvm_state->regs_modified) {
 583             kvm_arch_put_registers(env);
 584             env->kvm_state->regs_modified = 0;
 585         }
 586
 587         kvm_arch_pre_run(env, run);
 588         qemu_mutex_unlock_iothread();
 589         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 590         qemu_mutex_lock_iothread();
 591         kvm_arch_post_run(env, run);
 592
 593         if (ret == -EINTR || ret == -EAGAIN) {
 594             dprintf("io window exit\n");
 595             ret = 0;
 596             break;
 597         }
 598
 599         if (ret < 0) {
 600             dprintf("kvm run failed %s\n", strerror(-ret));
 601             abort();
 602         }
 603
 604         kvm_run_coalesced_mmio(env, run);
 605
 606         ret = 0; /* exit loop */
 607         switch (run->exit_reason) {
 608         case KVM_EXIT_IO:
 609             dprintf("handle_io\n");
 610             ret = kvm_handle_io(run->io.port,
 611                                 (uint8_t *)run + run->io.data_offset,
 612                                 run->io.direction,
 613                                 run->io.size,
 614                                 run->io.count);
 615             break;
 616         case KVM_EXIT_MMIO:
 617             dprintf("handle_mmio\n");
 618             cpu_physical_memory_rw(run->mmio.phys_addr,
 619                                    run->mmio.data,
 620                                    run->mmio.len,
 621                                    run->mmio.is_write);
 622             ret = 1;
 623             break;
 624         case KVM_EXIT_IRQ_WINDOW_OPEN:
 625             dprintf("irq_window_open\n");
 626             break;
 627         case KVM_EXIT_SHUTDOWN:
 628             dprintf("shutdown\n");
 629             qemu_system_reset_request();
 630             ret = 1;
 631             break;
 632         case KVM_EXIT_UNKNOWN:
 633             dprintf("kvm_exit_unknown\n");
 634             break;
 635         case KVM_EXIT_FAIL_ENTRY:
 636             dprintf("kvm_exit_fail_entry\n");
 637             break;
 638         case KVM_EXIT_EXCEPTION:
 639             dprintf("kvm_exit_exception\n");
 640             break;
 641         case KVM_EXIT_DEBUG:
 642             dprintf("kvm_exit_debug\n");
 643 #ifdef KVM_CAP_SET_GUEST_DEBUG
 644             if (kvm_arch_debug(&run->debug.arch)) {
 645                 gdb_set_stop_cpu(env);
 646                 vm_stop(EXCP_DEBUG);
 647                 env->exception_index = EXCP_DEBUG;
 648                 return 0;
 649             }
 650             /* re-enter, this exception was guest-internal */
 651             ret = 1;
 652 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 653             break;
 654         default:
 655             dprintf("kvm_arch_handle_exit\n");
 656             ret = kvm_arch_handle_exit(env, run);
 657             break;
 658         }
 659     } while (ret > 0);
 660
 661     if (env->exit_request) {
 662         env->exit_request = 0;
 663         env->exception_index = EXCP_INTERRUPT;
 664     }
 665
 666     return ret;
 667 }
 668
 669 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 670                       ram_addr_t size,
 671                       ram_addr_t phys_offset)
 672 {
 673     KVMState *s = kvm_state;
 674     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 675     KVMSlot *mem, old;
 676     int err;
 677
 678     if (start_addr & ~TARGET_PAGE_MASK) {
 679         if (flags >= IO_MEM_UNASSIGNED) {
 680             if (!kvm_lookup_overlapping_slot(s, start_addr,
 681                                              start_addr + size)) {
 682                 return;
 683             }
 684             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 685         } else {
 686             fprintf(stderr, "Only page-aligned memory slots supported\n");
 687         }
 688         abort();
 689     }
 690
 691     /* KVM does not support read-only slots */
 692     phys_offset &= ~IO_MEM_ROM;
 693
 694     while (1) {
 695         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 696         if (!mem) {
 697             break;
 698         }
 699
 700         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 701             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 702             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 703             /* The new slot fits into the existing one and comes with
 704              * identical parameters - nothing to be done. */
 705             return;
 706         }
 707
 708         old = *mem;
 709
 710         /* unregister the overlapping slot */
 711         mem->memory_size = 0;
 712         err = kvm_set_user_memory_region(s, mem);
 713         if (err) {
 714             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 715                     __func__, strerror(-err));
 716             abort();
 717         }
 718
 719         /* Workaround for older KVM versions: we can't join slots, even not by
 720          * unregistering the previous ones and then registering the larger
 721          * slot. We have to maintain the existing fragmentation. Sigh.
 722          *
 723          * This workaround assumes that the new slot starts at the same
 724          * address as the first existing one. If not or if some overlapping
 725          * slot comes around later, we will fail (not seen in practice so far)
 726          * - and actually require a recent KVM version. */
 727         if (s->broken_set_mem_region &&
 728             old.start_addr == start_addr && old.memory_size < size &&
 729             flags < IO_MEM_UNASSIGNED) {
 730             mem = kvm_alloc_slot(s);
 731             mem->memory_size = old.memory_size;
 732             mem->start_addr = old.start_addr;
 733             mem->phys_offset = old.phys_offset;
 734             mem->flags = 0;
 735
 736             err = kvm_set_user_memory_region(s, mem);
 737             if (err) {
 738                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 739                         strerror(-err));
 740                 abort();
 741             }
 742
 743             start_addr += old.memory_size;
 744             phys_offset += old.memory_size;
 745             size -= old.memory_size;
 746             continue;
 747         }
 748
 749         /* register prefix slot */
 750         if (old.start_addr < start_addr) {
 751             mem = kvm_alloc_slot(s);
 752             mem->memory_size = start_addr - old.start_addr;
 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 registering prefix slot: %s\n",
 760                         __func__, strerror(-err));
 761                 abort();
 762             }
 763         }
 764
 765         /* register suffix slot */
 766         if (old.start_addr + old.memory_size > start_addr + size) {
 767             ram_addr_t size_delta;
 768
 769             mem = kvm_alloc_slot(s);
 770             mem->start_addr = start_addr + size;
 771             size_delta = mem->start_addr - old.start_addr;
 772             mem->memory_size = old.memory_size - size_delta;
 773             mem->phys_offset = old.phys_offset + size_delta;
 774             mem->flags = 0;
 775
 776             err = kvm_set_user_memory_region(s, mem);
 777             if (err) {
 778                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 779                         __func__, strerror(-err));
 780                 abort();
 781             }
 782         }
 783     }
 784
 785     /* in case the KVM bug workaround already "consumed" the new slot */
 786     if (!size)
 787         return;
 788
 789     /* KVM does not need to know about this memory */
 790     if (flags >= IO_MEM_UNASSIGNED)
 791         return;
 792
 793     mem = kvm_alloc_slot(s);
 794     mem->memory_size = size;
 795     mem->start_addr = start_addr;
 796     mem->phys_offset = phys_offset;
 797     mem->flags = 0;
 798
 799     err = kvm_set_user_memory_region(s, mem);
 800     if (err) {
 801         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 802                 strerror(-err));
 803         abort();
 804     }
 805 }
 806
 807 int kvm_ioctl(KVMState *s, int type, ...)
 808 {
 809     int ret;
 810     void *arg;
 811     va_list ap;
 812
 813     va_start(ap, type);
 814     arg = va_arg(ap, void *);
 815     va_end(ap);
 816
 817     ret = ioctl(s->fd, type, arg);
 818     if (ret == -1)
 819         ret = -errno;
 820
 821     return ret;
 822 }
 823
 824 int kvm_vm_ioctl(KVMState *s, int type, ...)
 825 {
 826     int ret;
 827     void *arg;
 828     va_list ap;
 829
 830     va_start(ap, type);
 831     arg = va_arg(ap, void *);
 832     va_end(ap);
 833
 834     ret = ioctl(s->vmfd, type, arg);
 835     if (ret == -1)
 836         ret = -errno;
 837
 838     return ret;
 839 }
 840
 841 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 842 {
 843     int ret;
 844     void *arg;
 845     va_list ap;
 846
 847     va_start(ap, type);
 848     arg = va_arg(ap, void *);
 849     va_end(ap);
 850
 851     ret = ioctl(env->kvm_fd, type, arg);
 852     if (ret == -1)
 853         ret = -errno;
 854
 855     return ret;
 856 }
 857
 858 int kvm_has_sync_mmu(void)
 859 {
 860 #ifdef KVM_CAP_SYNC_MMU
 861     KVMState *s = kvm_state;
 862
 863     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 864 #else
 865     return 0;
 866 #endif
 867 }
 868
 869 void kvm_setup_guest_memory(void *start, size_t size)
 870 {
 871     if (!kvm_has_sync_mmu()) {
 872 #ifdef MADV_DONTFORK
 873         int ret = madvise(start, size, MADV_DONTFORK);
 874
 875         if (ret) {
 876             perror("madvice");
 877             exit(1);
 878         }
 879 #else
 880         fprintf(stderr,
 881                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 882         exit(1);
 883 #endif
 884     }
 885 }
 886
 887 #ifdef KVM_CAP_SET_GUEST_DEBUG
 888 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 889 {
 890 #ifdef CONFIG_IOTHREAD
 891     if (env == cpu_single_env) {
 892         func(data);
 893         return;
 894     }
 895     abort();
 896 #else
 897     func(data);
 898 #endif
 899 }
 900
 901 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 902                                                  target_ulong pc)
 903 {
 904     struct kvm_sw_breakpoint *bp;
 905
 906     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 907         if (bp->pc == pc)
 908             return bp;
 909     }
 910     return NULL;
 911 }
 912
 913 int kvm_sw_breakpoints_active(CPUState *env)
 914 {
 915     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 916 }
 917
 918 struct kvm_set_guest_debug_data {
 919     struct kvm_guest_debug dbg;
 920     CPUState *env;
 921     int err;
 922 };
 923
 924 static void kvm_invoke_set_guest_debug(void *data)
 925 {
 926     struct kvm_set_guest_debug_data *dbg_data = data;
 927     CPUState *env = dbg_data->env;
 928
 929     if (env->kvm_state->regs_modified) {
 930         kvm_arch_put_registers(env);
 931         env->kvm_state->regs_modified = 0;
 932     }
 933     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 934 }
 935
 936 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 937 {
 938     struct kvm_set_guest_debug_data data;
 939
 940     data.dbg.control = 0;
 941     if (env->singlestep_enabled)
 942         data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 943
 944     kvm_arch_update_guest_debug(env, &data.dbg);
 945     data.dbg.control |= reinject_trap;
 946     data.env = env;
 947
 948     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
 949     return data.err;
 950 }
 951
 952 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 953                           target_ulong len, int type)
 954 {
 955     struct kvm_sw_breakpoint *bp;
 956     CPUState *env;
 957     int err;
 958
 959     if (type == GDB_BREAKPOINT_SW) {
 960         bp = kvm_find_sw_breakpoint(current_env, addr);
 961         if (bp) {
 962             bp->use_count++;
 963             return 0;
 964         }
 965
 966         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
 967         if (!bp)
 968             return -ENOMEM;
 969
 970         bp->pc = addr;
 971         bp->use_count = 1;
 972         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
 973         if (err) {
 974             free(bp);
 975             return err;
 976         }
 977
 978         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
 979                           bp, entry);
 980     } else {
 981         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
 982         if (err)
 983             return err;
 984     }
 985
 986     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 987         err = kvm_update_guest_debug(env, 0);
 988         if (err)
 989             return err;
 990     }
 991     return 0;
 992 }
 993
 994 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
 995                           target_ulong len, int type)
 996 {
 997     struct kvm_sw_breakpoint *bp;
 998     CPUState *env;
 999     int err;
1000
1001     if (type == GDB_BREAKPOINT_SW) {
1002         bp = kvm_find_sw_breakpoint(current_env, addr);
1003         if (!bp)
1004             return -ENOENT;
1005
1006         if (bp->use_count > 1) {
1007             bp->use_count--;
1008             return 0;
1009         }
1010
1011         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1012         if (err)
1013             return err;
1014
1015         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1016         qemu_free(bp);
1017     } else {
1018         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1019         if (err)
1020             return err;
1021     }
1022
1023     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1024         err = kvm_update_guest_debug(env, 0);
1025         if (err)
1026             return err;
1027     }
1028     return 0;
1029 }
1030
1031 void kvm_remove_all_breakpoints(CPUState *current_env)
1032 {
1033     struct kvm_sw_breakpoint *bp, *next;
1034     KVMState *s = current_env->kvm_state;
1035     CPUState *env;
1036
1037     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1038         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1039             /* Try harder to find a CPU that currently sees the breakpoint. */
1040             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1041                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1042                     break;
1043             }
1044         }
1045     }
1046     kvm_arch_remove_all_hw_breakpoints();
1047
1048     for (env = first_cpu; env != NULL; env = env->next_cpu)
1049         kvm_update_guest_debug(env, 0);
1050 }
1051
1052 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1053
1054 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1055 {
1056     return -EINVAL;
1057 }
1058
1059 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1060                           target_ulong len, int type)
1061 {
1062     return -EINVAL;
1063 }
1064
1065 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1066                           target_ulong len, int type)
1067 {
1068     return -EINVAL;
1069 }
1070
1071 void kvm_remove_all_breakpoints(CPUState *current_env)
1072 {
1073 }
1074 #endif /* !KVM_CAP_SET_GUEST_DEBUG */