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 static 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 static 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 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 406                              ram_addr_t size,
 407                              ram_addr_t phys_offset)
 408 {
 409     KVMState *s = kvm_state;
 410     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 411     KVMSlot *mem, old;
 412     int err;
 413
 414     if (start_addr & ~TARGET_PAGE_MASK) {
 415         if (flags >= IO_MEM_UNASSIGNED) {
 416             if (!kvm_lookup_overlapping_slot(s, start_addr,
 417                                              start_addr + size)) {
 418                 return;
 419             }
 420             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 421         } else {
 422             fprintf(stderr, "Only page-aligned memory slots supported\n");
 423         }
 424         abort();
 425     }
 426
 427     /* KVM does not support read-only slots */
 428     phys_offset &= ~IO_MEM_ROM;
 429
 430     while (1) {
 431         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 432         if (!mem) {
 433             break;
 434         }
 435
 436         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 437             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 438             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 439             /* The new slot fits into the existing one and comes with
 440              * identical parameters - nothing to be done. */
 441             return;
 442         }
 443
 444         old = *mem;
 445
 446         /* unregister the overlapping slot */
 447         mem->memory_size = 0;
 448         err = kvm_set_user_memory_region(s, mem);
 449         if (err) {
 450             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 451                     __func__, strerror(-err));
 452             abort();
 453         }
 454
 455         /* Workaround for older KVM versions: we can't join slots, even not by
 456          * unregistering the previous ones and then registering the larger
 457          * slot. We have to maintain the existing fragmentation. Sigh.
 458          *
 459          * This workaround assumes that the new slot starts at the same
 460          * address as the first existing one. If not or if some overlapping
 461          * slot comes around later, we will fail (not seen in practice so far)
 462          * - and actually require a recent KVM version. */
 463         if (s->broken_set_mem_region &&
 464             old.start_addr == start_addr && old.memory_size < size &&
 465             flags < IO_MEM_UNASSIGNED) {
 466             mem = kvm_alloc_slot(s);
 467             mem->memory_size = old.memory_size;
 468             mem->start_addr = old.start_addr;
 469             mem->phys_offset = old.phys_offset;
 470             mem->flags = 0;
 471
 472             err = kvm_set_user_memory_region(s, mem);
 473             if (err) {
 474                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 475                         strerror(-err));
 476                 abort();
 477             }
 478
 479             start_addr += old.memory_size;
 480             phys_offset += old.memory_size;
 481             size -= old.memory_size;
 482             continue;
 483         }
 484
 485         /* register prefix slot */
 486         if (old.start_addr < start_addr) {
 487             mem = kvm_alloc_slot(s);
 488             mem->memory_size = start_addr - old.start_addr;
 489             mem->start_addr = old.start_addr;
 490             mem->phys_offset = old.phys_offset;
 491             mem->flags = 0;
 492
 493             err = kvm_set_user_memory_region(s, mem);
 494             if (err) {
 495                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 496                         __func__, strerror(-err));
 497                 abort();
 498             }
 499         }
 500
 501         /* register suffix slot */
 502         if (old.start_addr + old.memory_size > start_addr + size) {
 503             ram_addr_t size_delta;
 504
 505             mem = kvm_alloc_slot(s);
 506             mem->start_addr = start_addr + size;
 507             size_delta = mem->start_addr - old.start_addr;
 508             mem->memory_size = old.memory_size - size_delta;
 509             mem->phys_offset = old.phys_offset + size_delta;
 510             mem->flags = 0;
 511
 512             err = kvm_set_user_memory_region(s, mem);
 513             if (err) {
 514                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 515                         __func__, strerror(-err));
 516                 abort();
 517             }
 518         }
 519     }
 520
 521     /* in case the KVM bug workaround already "consumed" the new slot */
 522     if (!size)
 523         return;
 524
 525     /* KVM does not need to know about this memory */
 526     if (flags >= IO_MEM_UNASSIGNED)
 527         return;
 528
 529     mem = kvm_alloc_slot(s);
 530     mem->memory_size = size;
 531     mem->start_addr = start_addr;
 532     mem->phys_offset = phys_offset;
 533     mem->flags = 0;
 534
 535     err = kvm_set_user_memory_region(s, mem);
 536     if (err) {
 537         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 538                 strerror(-err));
 539         abort();
 540     }
 541 }
 542
 543 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 544                                   target_phys_addr_t start_addr,
 545                                   ram_addr_t size,
 546                                   ram_addr_t phys_offset)
 547 {
 548         kvm_set_phys_mem(start_addr, size, phys_offset);
 549 }
 550
 551 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 552                                         target_phys_addr_t start_addr,
 553                                         target_phys_addr_t end_addr)
 554 {
 555         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 556 }
 557
 558 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 559                                     int enable)
 560 {
 561         return kvm_set_migration_log(enable);
 562 }
 563
 564 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 565         .set_memory = kvm_client_set_memory,
 566         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 567         .migration_log = kvm_client_migration_log,
 568 };
 569
 570 int kvm_init(int smp_cpus)
 571 {
 572     static const char upgrade_note[] =
 573         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 574         "(see http://sourceforge.net/projects/kvm).\n";
 575     KVMState *s;
 576     int ret;
 577     int i;
 578
 579     if (smp_cpus > 1) {
 580         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 581         return -EINVAL;
 582     }
 583
 584     s = qemu_mallocz(sizeof(KVMState));
 585
 586 #ifdef KVM_CAP_SET_GUEST_DEBUG
 587     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 588 #endif
 589     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 590         s->slots[i].slot = i;
 591
 592     s->vmfd = -1;
 593     s->fd = qemu_open("/dev/kvm", O_RDWR);
 594     if (s->fd == -1) {
 595         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 596         ret = -errno;
 597         goto err;
 598     }
 599
 600     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 601     if (ret < KVM_API_VERSION) {
 602         if (ret > 0)
 603             ret = -EINVAL;
 604         fprintf(stderr, "kvm version too old\n");
 605         goto err;
 606     }
 607
 608     if (ret > KVM_API_VERSION) {
 609         ret = -EINVAL;
 610         fprintf(stderr, "kvm version not supported\n");
 611         goto err;
 612     }
 613
 614     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 615     if (s->vmfd < 0)
 616         goto err;
 617
 618     /* initially, KVM allocated its own memory and we had to jump through
 619      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 620      * just use a user allocated buffer so we can use regular pages
 621      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 622      */
 623     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 624         ret = -EINVAL;
 625         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 626                 upgrade_note);
 627         goto err;
 628     }
 629
 630     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 631      * destroyed properly.  Since we rely on this capability, refuse to work
 632      * with any kernel without this capability. */
 633     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 634         ret = -EINVAL;
 635
 636         fprintf(stderr,
 637                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 638                 upgrade_note);
 639         goto err;
 640     }
 641
 642     s->coalesced_mmio = 0;
 643 #ifdef KVM_CAP_COALESCED_MMIO
 644     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 645     s->coalesced_mmio_ring = NULL;
 646 #endif
 647
 648     s->broken_set_mem_region = 1;
 649 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 650     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 651     if (ret > 0) {
 652         s->broken_set_mem_region = 0;
 653     }
 654 #endif
 655
 656     s->vcpu_events = 0;
 657 #ifdef KVM_CAP_VCPU_EVENTS
 658     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 659 #endif
 660
 661     ret = kvm_arch_init(s, smp_cpus);
 662     if (ret < 0)
 663         goto err;
 664
 665     kvm_state = s;
 666     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 667
 668     return 0;
 669
 670 err:
 671     if (s) {
 672         if (s->vmfd != -1)
 673             close(s->vmfd);
 674         if (s->fd != -1)
 675             close(s->fd);
 676     }
 677     qemu_free(s);
 678
 679     return ret;
 680 }
 681
 682 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 683                          uint32_t count)
 684 {
 685     int i;
 686     uint8_t *ptr = data;
 687
 688     for (i = 0; i < count; i++) {
 689         if (direction == KVM_EXIT_IO_IN) {
 690             switch (size) {
 691             case 1:
 692                 stb_p(ptr, cpu_inb(port));
 693                 break;
 694             case 2:
 695                 stw_p(ptr, cpu_inw(port));
 696                 break;
 697             case 4:
 698                 stl_p(ptr, cpu_inl(port));
 699                 break;
 700             }
 701         } else {
 702             switch (size) {
 703             case 1:
 704                 cpu_outb(port, ldub_p(ptr));
 705                 break;
 706             case 2:
 707                 cpu_outw(port, lduw_p(ptr));
 708                 break;
 709             case 4:
 710                 cpu_outl(port, ldl_p(ptr));
 711                 break;
 712             }
 713         }
 714
 715         ptr += size;
 716     }
 717
 718     return 1;
 719 }
 720
 721 void kvm_flush_coalesced_mmio_buffer(void)
 722 {
 723 #ifdef KVM_CAP_COALESCED_MMIO
 724     KVMState *s = kvm_state;
 725     if (s->coalesced_mmio_ring) {
 726         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 727         while (ring->first != ring->last) {
 728             struct kvm_coalesced_mmio *ent;
 729
 730             ent = &ring->coalesced_mmio[ring->first];
 731
 732             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 733             /* FIXME smp_wmb() */
 734             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 735         }
 736     }
 737 #endif
 738 }
 739
 740 void kvm_cpu_synchronize_state(CPUState *env)
 741 {
 742     if (!env->kvm_vcpu_dirty) {
 743         kvm_arch_get_registers(env);
 744         env->kvm_vcpu_dirty = 1;
 745     }
 746 }
 747
 748 int kvm_cpu_exec(CPUState *env)
 749 {
 750     struct kvm_run *run = env->kvm_run;
 751     int ret;
 752
 753     dprintf("kvm_cpu_exec()\n");
 754
 755     do {
 756         if (env->exit_request) {
 757             dprintf("interrupt exit requested\n");
 758             ret = 0;
 759             break;
 760         }
 761
 762         if (env->kvm_vcpu_dirty) {
 763             kvm_arch_put_registers(env);
 764             env->kvm_vcpu_dirty = 0;
 765         }
 766
 767         kvm_arch_pre_run(env, run);
 768         qemu_mutex_unlock_iothread();
 769         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 770         qemu_mutex_lock_iothread();
 771         kvm_arch_post_run(env, run);
 772
 773         if (ret == -EINTR || ret == -EAGAIN) {
 774             cpu_exit(env);
 775             dprintf("io window exit\n");
 776             ret = 0;
 777             break;
 778         }
 779
 780         if (ret < 0) {
 781             dprintf("kvm run failed %s\n", strerror(-ret));
 782             abort();
 783         }
 784
 785         kvm_flush_coalesced_mmio_buffer();
 786
 787         ret = 0; /* exit loop */
 788         switch (run->exit_reason) {
 789         case KVM_EXIT_IO:
 790             dprintf("handle_io\n");
 791             ret = kvm_handle_io(run->io.port,
 792                                 (uint8_t *)run + run->io.data_offset,
 793                                 run->io.direction,
 794                                 run->io.size,
 795                                 run->io.count);
 796             break;
 797         case KVM_EXIT_MMIO:
 798             dprintf("handle_mmio\n");
 799             cpu_physical_memory_rw(run->mmio.phys_addr,
 800                                    run->mmio.data,
 801                                    run->mmio.len,
 802                                    run->mmio.is_write);
 803             ret = 1;
 804             break;
 805         case KVM_EXIT_IRQ_WINDOW_OPEN:
 806             dprintf("irq_window_open\n");
 807             break;
 808         case KVM_EXIT_SHUTDOWN:
 809             dprintf("shutdown\n");
 810             qemu_system_reset_request();
 811             ret = 1;
 812             break;
 813         case KVM_EXIT_UNKNOWN:
 814             dprintf("kvm_exit_unknown\n");
 815             break;
 816         case KVM_EXIT_FAIL_ENTRY:
 817             dprintf("kvm_exit_fail_entry\n");
 818             break;
 819         case KVM_EXIT_EXCEPTION:
 820             dprintf("kvm_exit_exception\n");
 821             break;
 822         case KVM_EXIT_DEBUG:
 823             dprintf("kvm_exit_debug\n");
 824 #ifdef KVM_CAP_SET_GUEST_DEBUG
 825             if (kvm_arch_debug(&run->debug.arch)) {
 826                 gdb_set_stop_cpu(env);
 827                 vm_stop(EXCP_DEBUG);
 828                 env->exception_index = EXCP_DEBUG;
 829                 return 0;
 830             }
 831             /* re-enter, this exception was guest-internal */
 832             ret = 1;
 833 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 834             break;
 835         default:
 836             dprintf("kvm_arch_handle_exit\n");
 837             ret = kvm_arch_handle_exit(env, run);
 838             break;
 839         }
 840     } while (ret > 0);
 841
 842     if (env->exit_request) {
 843         env->exit_request = 0;
 844         env->exception_index = EXCP_INTERRUPT;
 845     }
 846
 847     return ret;
 848 }
 849
 850 int kvm_ioctl(KVMState *s, int type, ...)
 851 {
 852     int ret;
 853     void *arg;
 854     va_list ap;
 855
 856     va_start(ap, type);
 857     arg = va_arg(ap, void *);
 858     va_end(ap);
 859
 860     ret = ioctl(s->fd, type, arg);
 861     if (ret == -1)
 862         ret = -errno;
 863
 864     return ret;
 865 }
 866
 867 int kvm_vm_ioctl(KVMState *s, int type, ...)
 868 {
 869     int ret;
 870     void *arg;
 871     va_list ap;
 872
 873     va_start(ap, type);
 874     arg = va_arg(ap, void *);
 875     va_end(ap);
 876
 877     ret = ioctl(s->vmfd, type, arg);
 878     if (ret == -1)
 879         ret = -errno;
 880
 881     return ret;
 882 }
 883
 884 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 885 {
 886     int ret;
 887     void *arg;
 888     va_list ap;
 889
 890     va_start(ap, type);
 891     arg = va_arg(ap, void *);
 892     va_end(ap);
 893
 894     ret = ioctl(env->kvm_fd, type, arg);
 895     if (ret == -1)
 896         ret = -errno;
 897
 898     return ret;
 899 }
 900
 901 int kvm_has_sync_mmu(void)
 902 {
 903 #ifdef KVM_CAP_SYNC_MMU
 904     KVMState *s = kvm_state;
 905
 906     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 907 #else
 908     return 0;
 909 #endif
 910 }
 911
 912 int kvm_has_vcpu_events(void)
 913 {
 914     return kvm_state->vcpu_events;
 915 }
 916
 917 void kvm_setup_guest_memory(void *start, size_t size)
 918 {
 919     if (!kvm_has_sync_mmu()) {
 920 #ifdef MADV_DONTFORK
 921         int ret = madvise(start, size, MADV_DONTFORK);
 922
 923         if (ret) {
 924             perror("madvice");
 925             exit(1);
 926         }
 927 #else
 928         fprintf(stderr,
 929                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 930         exit(1);
 931 #endif
 932     }
 933 }
 934
 935 #ifdef KVM_CAP_SET_GUEST_DEBUG
 936 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 937 {
 938 #ifdef CONFIG_IOTHREAD
 939     if (env != cpu_single_env) {
 940         abort();
 941     }
 942 #endif
 943     func(data);
 944 }
 945
 946 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 947                                                  target_ulong pc)
 948 {
 949     struct kvm_sw_breakpoint *bp;
 950
 951     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 952         if (bp->pc == pc)
 953             return bp;
 954     }
 955     return NULL;
 956 }
 957
 958 int kvm_sw_breakpoints_active(CPUState *env)
 959 {
 960     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 961 }
 962
 963 struct kvm_set_guest_debug_data {
 964     struct kvm_guest_debug dbg;
 965     CPUState *env;
 966     int err;
 967 };
 968
 969 static void kvm_invoke_set_guest_debug(void *data)
 970 {
 971     struct kvm_set_guest_debug_data *dbg_data = data;
 972     CPUState *env = dbg_data->env;
 973
 974     if (env->kvm_vcpu_dirty) {
 975         kvm_arch_put_registers(env);
 976         env->kvm_vcpu_dirty = 0;
 977     }
 978     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 979 }
 980
 981 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 982 {
 983     struct kvm_set_guest_debug_data data;
 984
 985     data.dbg.control = 0;
 986     if (env->singlestep_enabled)
 987         data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
 988
 989     kvm_arch_update_guest_debug(env, &data.dbg);
 990     data.dbg.control |= reinject_trap;
 991     data.env = env;
 992
 993     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
 994     return data.err;
 995 }
 996
 997 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
 998                           target_ulong len, int type)
 999 {
1000     struct kvm_sw_breakpoint *bp;
1001     CPUState *env;
1002     int err;
1003
1004     if (type == GDB_BREAKPOINT_SW) {
1005         bp = kvm_find_sw_breakpoint(current_env, addr);
1006         if (bp) {
1007             bp->use_count++;
1008             return 0;
1009         }
1010
1011         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1012         if (!bp)
1013             return -ENOMEM;
1014
1015         bp->pc = addr;
1016         bp->use_count = 1;
1017         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1018         if (err) {
1019             free(bp);
1020             return err;
1021         }
1022
1023         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1024                           bp, entry);
1025     } else {
1026         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1027         if (err)
1028             return err;
1029     }
1030
1031     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1032         err = kvm_update_guest_debug(env, 0);
1033         if (err)
1034             return err;
1035     }
1036     return 0;
1037 }
1038
1039 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1040                           target_ulong len, int type)
1041 {
1042     struct kvm_sw_breakpoint *bp;
1043     CPUState *env;
1044     int err;
1045
1046     if (type == GDB_BREAKPOINT_SW) {
1047         bp = kvm_find_sw_breakpoint(current_env, addr);
1048         if (!bp)
1049             return -ENOENT;
1050
1051         if (bp->use_count > 1) {
1052             bp->use_count--;
1053             return 0;
1054         }
1055
1056         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1057         if (err)
1058             return err;
1059
1060         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1061         qemu_free(bp);
1062     } else {
1063         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1064         if (err)
1065             return err;
1066     }
1067
1068     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1069         err = kvm_update_guest_debug(env, 0);
1070         if (err)
1071             return err;
1072     }
1073     return 0;
1074 }
1075
1076 void kvm_remove_all_breakpoints(CPUState *current_env)
1077 {
1078     struct kvm_sw_breakpoint *bp, *next;
1079     KVMState *s = current_env->kvm_state;
1080     CPUState *env;
1081
1082     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1083         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1084             /* Try harder to find a CPU that currently sees the breakpoint. */
1085             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1086                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1087                     break;
1088             }
1089         }
1090     }
1091     kvm_arch_remove_all_hw_breakpoints();
1092
1093     for (env = first_cpu; env != NULL; env = env->next_cpu)
1094         kvm_update_guest_debug(env, 0);
1095 }
1096
1097 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1098
1099 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1100 {
1101     return -EINVAL;
1102 }
1103
1104 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1105                           target_ulong len, int type)
1106 {
1107     return -EINVAL;
1108 }
1109
1110 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1111                           target_ulong len, int type)
1112 {
1113     return -EINVAL;
1114 }
1115
1116 void kvm_remove_all_breakpoints(CPUState *current_env)
1117 {
1118 }
1119 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1120
1121 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1122 {
1123     struct kvm_signal_mask *sigmask;
1124     int r;
1125
1126     if (!sigset)
1127         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1128
1129     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1130
1131     sigmask->len = 8;
1132     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1133     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1134     free(sigmask);
1135
1136     return r;
1137 }