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 "qemu-barrier.h"
  25 #include "sysemu.h"
  26 #include "hw/hw.h"
  27 #include "gdbstub.h"
  28 #include "kvm.h"
  29
  30 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  31 #define PAGE_SIZE TARGET_PAGE_SIZE
  32
  33 //#define DEBUG_KVM
  34
  35 #ifdef DEBUG_KVM
  36 #define dprintf(fmt, ...) \
  37     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  38 #else
  39 #define dprintf(fmt, ...) \
  40     do { } while (0)
  41 #endif
  42
  43 typedef struct KVMSlot
  44 {
  45     target_phys_addr_t start_addr;
  46     ram_addr_t memory_size;
  47     ram_addr_t phys_offset;
  48     int slot;
  49     int flags;
  50 } KVMSlot;
  51
  52 typedef struct kvm_dirty_log KVMDirtyLog;
  53
  54 int kvm_allowed = 0;
  55
  56 struct KVMState
  57 {
  58     KVMSlot slots[32];
  59     int fd;
  60     int vmfd;
  61     int coalesced_mmio;
  62 #ifdef KVM_CAP_COALESCED_MMIO
  63     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  64 #endif
  65     int broken_set_mem_region;
  66     int migration_log;
  67     int vcpu_events;
  68     int robust_singlestep;
  69 #ifdef KVM_CAP_SET_GUEST_DEBUG
  70     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  71 #endif
  72     int irqchip_in_kernel;
  73     int pit_in_kernel;
  74 };
  75
  76 static KVMState *kvm_state;
  77
  78 static KVMSlot *kvm_alloc_slot(KVMState *s)
  79 {
  80     int i;
  81
  82     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  83         /* KVM private memory slots */
  84         if (i >= 8 && i < 12)
  85             continue;
  86         if (s->slots[i].memory_size == 0)
  87             return &s->slots[i];
  88     }
  89
  90     fprintf(stderr, "%s: no free slot available\n", __func__);
  91     abort();
  92 }
  93
  94 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  95                                          target_phys_addr_t start_addr,
  96                                          target_phys_addr_t end_addr)
  97 {
  98     int i;
  99
 100     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 101         KVMSlot *mem = &s->slots[i];
 102
 103         if (start_addr == mem->start_addr &&
 104             end_addr == mem->start_addr + mem->memory_size) {
 105             return mem;
 106         }
 107     }
 108
 109     return NULL;
 110 }
 111
 112 /*
 113  * Find overlapping slot with lowest start address
 114  */
 115 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 116                                             target_phys_addr_t start_addr,
 117                                             target_phys_addr_t end_addr)
 118 {
 119     KVMSlot *found = NULL;
 120     int i;
 121
 122     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 123         KVMSlot *mem = &s->slots[i];
 124
 125         if (mem->memory_size == 0 ||
 126             (found && found->start_addr < mem->start_addr)) {
 127             continue;
 128         }
 129
 130         if (end_addr > mem->start_addr &&
 131             start_addr < mem->start_addr + mem->memory_size) {
 132             found = mem;
 133         }
 134     }
 135
 136     return found;
 137 }
 138
 139 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 140 {
 141     struct kvm_userspace_memory_region mem;
 142
 143     mem.slot = slot->slot;
 144     mem.guest_phys_addr = slot->start_addr;
 145     mem.memory_size = slot->memory_size;
 146     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 147     mem.flags = slot->flags;
 148     if (s->migration_log) {
 149         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 150     }
 151     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 152 }
 153
 154 static void kvm_reset_vcpu(void *opaque)
 155 {
 156     CPUState *env = opaque;
 157
 158     kvm_arch_reset_vcpu(env);
 159     if (kvm_arch_put_registers(env)) {
 160         fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
 161         abort();
 162     }
 163 }
 164
 165 int kvm_irqchip_in_kernel(void)
 166 {
 167     return kvm_state->irqchip_in_kernel;
 168 }
 169
 170 int kvm_pit_in_kernel(void)
 171 {
 172     return kvm_state->pit_in_kernel;
 173 }
 174
 175
 176 int kvm_init_vcpu(CPUState *env)
 177 {
 178     KVMState *s = kvm_state;
 179     long mmap_size;
 180     int ret;
 181
 182     dprintf("kvm_init_vcpu\n");
 183
 184     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 185     if (ret < 0) {
 186         dprintf("kvm_create_vcpu failed\n");
 187         goto err;
 188     }
 189
 190     env->kvm_fd = ret;
 191     env->kvm_state = s;
 192
 193     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 194     if (mmap_size < 0) {
 195         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 196         goto err;
 197     }
 198
 199     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 200                         env->kvm_fd, 0);
 201     if (env->kvm_run == MAP_FAILED) {
 202         ret = -errno;
 203         dprintf("mmap'ing vcpu state failed\n");
 204         goto err;
 205     }
 206
 207 #ifdef KVM_CAP_COALESCED_MMIO
 208     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 209         s->coalesced_mmio_ring = (void *) env->kvm_run +
 210                 s->coalesced_mmio * PAGE_SIZE;
 211 #endif
 212
 213     ret = kvm_arch_init_vcpu(env);
 214     if (ret == 0) {
 215         qemu_register_reset(kvm_reset_vcpu, env);
 216         kvm_arch_reset_vcpu(env);
 217         ret = kvm_arch_put_registers(env);
 218     }
 219 err:
 220     return ret;
 221 }
 222
 223 /*
 224  * dirty pages logging control
 225  */
 226 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 227                                       ram_addr_t size, int flags, int mask)
 228 {
 229     KVMState *s = kvm_state;
 230     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 231     int old_flags;
 232
 233     if (mem == NULL)  {
 234             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 235                     TARGET_FMT_plx "\n", __func__, phys_addr,
 236                     (target_phys_addr_t)(phys_addr + size - 1));
 237             return -EINVAL;
 238     }
 239
 240     old_flags = mem->flags;
 241
 242     flags = (mem->flags & ~mask) | flags;
 243     mem->flags = flags;
 244
 245     /* If nothing changed effectively, no need to issue ioctl */
 246     if (s->migration_log) {
 247         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 248     }
 249     if (flags == old_flags) {
 250             return 0;
 251     }
 252
 253     return kvm_set_user_memory_region(s, mem);
 254 }
 255
 256 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 257 {
 258         return kvm_dirty_pages_log_change(phys_addr, size,
 259                                           KVM_MEM_LOG_DIRTY_PAGES,
 260                                           KVM_MEM_LOG_DIRTY_PAGES);
 261 }
 262
 263 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 264 {
 265         return kvm_dirty_pages_log_change(phys_addr, size,
 266                                           0,
 267                                           KVM_MEM_LOG_DIRTY_PAGES);
 268 }
 269
 270 static int kvm_set_migration_log(int enable)
 271 {
 272     KVMState *s = kvm_state;
 273     KVMSlot *mem;
 274     int i, err;
 275
 276     s->migration_log = enable;
 277
 278     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 279         mem = &s->slots[i];
 280
 281         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 282             continue;
 283         }
 284         err = kvm_set_user_memory_region(s, mem);
 285         if (err) {
 286             return err;
 287         }
 288     }
 289     return 0;
 290 }
 291
 292 static int test_le_bit(unsigned long nr, unsigned char *addr)
 293 {
 294     return (addr[nr >> 3] >> (nr & 7)) & 1;
 295 }
 296
 297 /**
 298  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 299  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 300  * This means all bits are set to dirty.
 301  *
 302  * @start_add: start of logged region.
 303  * @end_addr: end of logged region.
 304  */
 305 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 306                                           target_phys_addr_t end_addr)
 307 {
 308     KVMState *s = kvm_state;
 309     unsigned long size, allocated_size = 0;
 310     target_phys_addr_t phys_addr;
 311     ram_addr_t addr;
 312     KVMDirtyLog d;
 313     KVMSlot *mem;
 314     int ret = 0;
 315
 316     d.dirty_bitmap = NULL;
 317     while (start_addr < end_addr) {
 318         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 319         if (mem == NULL) {
 320             break;
 321         }
 322
 323         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 324         if (!d.dirty_bitmap) {
 325             d.dirty_bitmap = qemu_malloc(size);
 326         } else if (size > allocated_size) {
 327             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 328         }
 329         allocated_size = size;
 330         memset(d.dirty_bitmap, 0, allocated_size);
 331
 332         d.slot = mem->slot;
 333
 334         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 335             dprintf("ioctl failed %d\n", errno);
 336             ret = -1;
 337             break;
 338         }
 339
 340         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 341              phys_addr < mem->start_addr + mem->memory_size;
 342              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 343             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 344             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 345
 346             if (test_le_bit(nr, bitmap)) {
 347                 cpu_physical_memory_set_dirty(addr);
 348             }
 349         }
 350         start_addr = phys_addr;
 351     }
 352     qemu_free(d.dirty_bitmap);
 353
 354     return ret;
 355 }
 356
 357 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 358 {
 359     int ret = -ENOSYS;
 360 #ifdef KVM_CAP_COALESCED_MMIO
 361     KVMState *s = kvm_state;
 362
 363     if (s->coalesced_mmio) {
 364         struct kvm_coalesced_mmio_zone zone;
 365
 366         zone.addr = start;
 367         zone.size = size;
 368
 369         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 370     }
 371 #endif
 372
 373     return ret;
 374 }
 375
 376 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 377 {
 378     int ret = -ENOSYS;
 379 #ifdef KVM_CAP_COALESCED_MMIO
 380     KVMState *s = kvm_state;
 381
 382     if (s->coalesced_mmio) {
 383         struct kvm_coalesced_mmio_zone zone;
 384
 385         zone.addr = start;
 386         zone.size = size;
 387
 388         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 389     }
 390 #endif
 391
 392     return ret;
 393 }
 394
 395 int kvm_check_extension(KVMState *s, unsigned int extension)
 396 {
 397     int ret;
 398
 399     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 400     if (ret < 0) {
 401         ret = 0;
 402     }
 403
 404     return ret;
 405 }
 406
 407 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 408                              ram_addr_t size,
 409                              ram_addr_t phys_offset)
 410 {
 411     KVMState *s = kvm_state;
 412     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 413     KVMSlot *mem, old;
 414     int err;
 415
 416     if (start_addr & ~TARGET_PAGE_MASK) {
 417         if (flags >= IO_MEM_UNASSIGNED) {
 418             if (!kvm_lookup_overlapping_slot(s, start_addr,
 419                                              start_addr + size)) {
 420                 return;
 421             }
 422             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 423         } else {
 424             fprintf(stderr, "Only page-aligned memory slots supported\n");
 425         }
 426         abort();
 427     }
 428
 429     /* KVM does not support read-only slots */
 430     phys_offset &= ~IO_MEM_ROM;
 431
 432     while (1) {
 433         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 434         if (!mem) {
 435             break;
 436         }
 437
 438         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 439             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 440             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 441             /* The new slot fits into the existing one and comes with
 442              * identical parameters - nothing to be done. */
 443             return;
 444         }
 445
 446         old = *mem;
 447
 448         /* unregister the overlapping slot */
 449         mem->memory_size = 0;
 450         err = kvm_set_user_memory_region(s, mem);
 451         if (err) {
 452             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 453                     __func__, strerror(-err));
 454             abort();
 455         }
 456
 457         /* Workaround for older KVM versions: we can't join slots, even not by
 458          * unregistering the previous ones and then registering the larger
 459          * slot. We have to maintain the existing fragmentation. Sigh.
 460          *
 461          * This workaround assumes that the new slot starts at the same
 462          * address as the first existing one. If not or if some overlapping
 463          * slot comes around later, we will fail (not seen in practice so far)
 464          * - and actually require a recent KVM version. */
 465         if (s->broken_set_mem_region &&
 466             old.start_addr == start_addr && old.memory_size < size &&
 467             flags < IO_MEM_UNASSIGNED) {
 468             mem = kvm_alloc_slot(s);
 469             mem->memory_size = old.memory_size;
 470             mem->start_addr = old.start_addr;
 471             mem->phys_offset = old.phys_offset;
 472             mem->flags = 0;
 473
 474             err = kvm_set_user_memory_region(s, mem);
 475             if (err) {
 476                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 477                         strerror(-err));
 478                 abort();
 479             }
 480
 481             start_addr += old.memory_size;
 482             phys_offset += old.memory_size;
 483             size -= old.memory_size;
 484             continue;
 485         }
 486
 487         /* register prefix slot */
 488         if (old.start_addr < start_addr) {
 489             mem = kvm_alloc_slot(s);
 490             mem->memory_size = start_addr - old.start_addr;
 491             mem->start_addr = old.start_addr;
 492             mem->phys_offset = old.phys_offset;
 493             mem->flags = 0;
 494
 495             err = kvm_set_user_memory_region(s, mem);
 496             if (err) {
 497                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 498                         __func__, strerror(-err));
 499                 abort();
 500             }
 501         }
 502
 503         /* register suffix slot */
 504         if (old.start_addr + old.memory_size > start_addr + size) {
 505             ram_addr_t size_delta;
 506
 507             mem = kvm_alloc_slot(s);
 508             mem->start_addr = start_addr + size;
 509             size_delta = mem->start_addr - old.start_addr;
 510             mem->memory_size = old.memory_size - size_delta;
 511             mem->phys_offset = old.phys_offset + size_delta;
 512             mem->flags = 0;
 513
 514             err = kvm_set_user_memory_region(s, mem);
 515             if (err) {
 516                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 517                         __func__, strerror(-err));
 518                 abort();
 519             }
 520         }
 521     }
 522
 523     /* in case the KVM bug workaround already "consumed" the new slot */
 524     if (!size)
 525         return;
 526
 527     /* KVM does not need to know about this memory */
 528     if (flags >= IO_MEM_UNASSIGNED)
 529         return;
 530
 531     mem = kvm_alloc_slot(s);
 532     mem->memory_size = size;
 533     mem->start_addr = start_addr;
 534     mem->phys_offset = phys_offset;
 535     mem->flags = 0;
 536
 537     err = kvm_set_user_memory_region(s, mem);
 538     if (err) {
 539         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 540                 strerror(-err));
 541         abort();
 542     }
 543 }
 544
 545 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 546                                   target_phys_addr_t start_addr,
 547                                   ram_addr_t size,
 548                                   ram_addr_t phys_offset)
 549 {
 550         kvm_set_phys_mem(start_addr, size, phys_offset);
 551 }
 552
 553 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 554                                         target_phys_addr_t start_addr,
 555                                         target_phys_addr_t end_addr)
 556 {
 557         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 558 }
 559
 560 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 561                                     int enable)
 562 {
 563         return kvm_set_migration_log(enable);
 564 }
 565
 566 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 567         .set_memory = kvm_client_set_memory,
 568         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 569         .migration_log = kvm_client_migration_log,
 570 };
 571
 572 int kvm_init(int smp_cpus)
 573 {
 574     static const char upgrade_note[] =
 575         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 576         "(see http://sourceforge.net/projects/kvm).\n";
 577     KVMState *s;
 578     int ret;
 579     int i;
 580
 581     if (smp_cpus > 1) {
 582         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 583         return -EINVAL;
 584     }
 585
 586     s = qemu_mallocz(sizeof(KVMState));
 587
 588 #ifdef KVM_CAP_SET_GUEST_DEBUG
 589     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 590 #endif
 591     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 592         s->slots[i].slot = i;
 593
 594     s->vmfd = -1;
 595     s->fd = qemu_open("/dev/kvm", O_RDWR);
 596     if (s->fd == -1) {
 597         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 598         ret = -errno;
 599         goto err;
 600     }
 601
 602     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 603     if (ret < KVM_API_VERSION) {
 604         if (ret > 0)
 605             ret = -EINVAL;
 606         fprintf(stderr, "kvm version too old\n");
 607         goto err;
 608     }
 609
 610     if (ret > KVM_API_VERSION) {
 611         ret = -EINVAL;
 612         fprintf(stderr, "kvm version not supported\n");
 613         goto err;
 614     }
 615
 616     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 617     if (s->vmfd < 0)
 618         goto err;
 619
 620     /* initially, KVM allocated its own memory and we had to jump through
 621      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 622      * just use a user allocated buffer so we can use regular pages
 623      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 624      */
 625     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 626         ret = -EINVAL;
 627         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 628                 upgrade_note);
 629         goto err;
 630     }
 631
 632     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 633      * destroyed properly.  Since we rely on this capability, refuse to work
 634      * with any kernel without this capability. */
 635     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 636         ret = -EINVAL;
 637
 638         fprintf(stderr,
 639                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 640                 upgrade_note);
 641         goto err;
 642     }
 643
 644     s->coalesced_mmio = 0;
 645 #ifdef KVM_CAP_COALESCED_MMIO
 646     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 647     s->coalesced_mmio_ring = NULL;
 648 #endif
 649
 650     s->broken_set_mem_region = 1;
 651 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 652     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 653     if (ret > 0) {
 654         s->broken_set_mem_region = 0;
 655     }
 656 #endif
 657
 658     s->vcpu_events = 0;
 659 #ifdef KVM_CAP_VCPU_EVENTS
 660     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 661 #endif
 662
 663     s->robust_singlestep = 0;
 664 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 665     s->robust_singlestep =
 666         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 667 #endif
 668
 669     ret = kvm_arch_init(s, smp_cpus);
 670     if (ret < 0)
 671         goto err;
 672
 673     kvm_state = s;
 674     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 675
 676     return 0;
 677
 678 err:
 679     if (s) {
 680         if (s->vmfd != -1)
 681             close(s->vmfd);
 682         if (s->fd != -1)
 683             close(s->fd);
 684     }
 685     qemu_free(s);
 686
 687     return ret;
 688 }
 689
 690 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 691                          uint32_t count)
 692 {
 693     int i;
 694     uint8_t *ptr = data;
 695
 696     for (i = 0; i < count; i++) {
 697         if (direction == KVM_EXIT_IO_IN) {
 698             switch (size) {
 699             case 1:
 700                 stb_p(ptr, cpu_inb(port));
 701                 break;
 702             case 2:
 703                 stw_p(ptr, cpu_inw(port));
 704                 break;
 705             case 4:
 706                 stl_p(ptr, cpu_inl(port));
 707                 break;
 708             }
 709         } else {
 710             switch (size) {
 711             case 1:
 712                 cpu_outb(port, ldub_p(ptr));
 713                 break;
 714             case 2:
 715                 cpu_outw(port, lduw_p(ptr));
 716                 break;
 717             case 4:
 718                 cpu_outl(port, ldl_p(ptr));
 719                 break;
 720             }
 721         }
 722
 723         ptr += size;
 724     }
 725
 726     return 1;
 727 }
 728
 729 void kvm_flush_coalesced_mmio_buffer(void)
 730 {
 731 #ifdef KVM_CAP_COALESCED_MMIO
 732     KVMState *s = kvm_state;
 733     if (s->coalesced_mmio_ring) {
 734         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 735         while (ring->first != ring->last) {
 736             struct kvm_coalesced_mmio *ent;
 737
 738             ent = &ring->coalesced_mmio[ring->first];
 739
 740             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 741             smp_wmb();
 742             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 743         }
 744     }
 745 #endif
 746 }
 747
 748 void kvm_cpu_synchronize_state(CPUState *env)
 749 {
 750     if (!env->kvm_vcpu_dirty) {
 751         kvm_arch_get_registers(env);
 752         env->kvm_vcpu_dirty = 1;
 753     }
 754 }
 755
 756 int kvm_cpu_exec(CPUState *env)
 757 {
 758     struct kvm_run *run = env->kvm_run;
 759     int ret;
 760
 761     dprintf("kvm_cpu_exec()\n");
 762
 763     do {
 764 #ifndef CONFIG_IOTHREAD
 765         if (env->exit_request) {
 766             dprintf("interrupt exit requested\n");
 767             ret = 0;
 768             break;
 769         }
 770 #endif
 771
 772         if (env->kvm_vcpu_dirty) {
 773             kvm_arch_put_registers(env);
 774             env->kvm_vcpu_dirty = 0;
 775         }
 776
 777         kvm_arch_pre_run(env, run);
 778         qemu_mutex_unlock_iothread();
 779         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 780         qemu_mutex_lock_iothread();
 781         kvm_arch_post_run(env, run);
 782
 783         if (ret == -EINTR || ret == -EAGAIN) {
 784             cpu_exit(env);
 785             dprintf("io window exit\n");
 786             ret = 0;
 787             break;
 788         }
 789
 790         if (ret < 0) {
 791             dprintf("kvm run failed %s\n", strerror(-ret));
 792             abort();
 793         }
 794
 795         kvm_flush_coalesced_mmio_buffer();
 796
 797         ret = 0; /* exit loop */
 798         switch (run->exit_reason) {
 799         case KVM_EXIT_IO:
 800             dprintf("handle_io\n");
 801             ret = kvm_handle_io(run->io.port,
 802                                 (uint8_t *)run + run->io.data_offset,
 803                                 run->io.direction,
 804                                 run->io.size,
 805                                 run->io.count);
 806             break;
 807         case KVM_EXIT_MMIO:
 808             dprintf("handle_mmio\n");
 809             cpu_physical_memory_rw(run->mmio.phys_addr,
 810                                    run->mmio.data,
 811                                    run->mmio.len,
 812                                    run->mmio.is_write);
 813             ret = 1;
 814             break;
 815         case KVM_EXIT_IRQ_WINDOW_OPEN:
 816             dprintf("irq_window_open\n");
 817             break;
 818         case KVM_EXIT_SHUTDOWN:
 819             dprintf("shutdown\n");
 820             qemu_system_reset_request();
 821             ret = 1;
 822             break;
 823         case KVM_EXIT_UNKNOWN:
 824             dprintf("kvm_exit_unknown\n");
 825             break;
 826         case KVM_EXIT_FAIL_ENTRY:
 827             dprintf("kvm_exit_fail_entry\n");
 828             break;
 829         case KVM_EXIT_EXCEPTION:
 830             dprintf("kvm_exit_exception\n");
 831             break;
 832         case KVM_EXIT_DEBUG:
 833             dprintf("kvm_exit_debug\n");
 834 #ifdef KVM_CAP_SET_GUEST_DEBUG
 835             if (kvm_arch_debug(&run->debug.arch)) {
 836                 gdb_set_stop_cpu(env);
 837                 vm_stop(EXCP_DEBUG);
 838                 env->exception_index = EXCP_DEBUG;
 839                 return 0;
 840             }
 841             /* re-enter, this exception was guest-internal */
 842             ret = 1;
 843 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 844             break;
 845         default:
 846             dprintf("kvm_arch_handle_exit\n");
 847             ret = kvm_arch_handle_exit(env, run);
 848             break;
 849         }
 850     } while (ret > 0);
 851
 852     if (env->exit_request) {
 853         env->exit_request = 0;
 854         env->exception_index = EXCP_INTERRUPT;
 855     }
 856
 857     return ret;
 858 }
 859
 860 int kvm_ioctl(KVMState *s, int type, ...)
 861 {
 862     int ret;
 863     void *arg;
 864     va_list ap;
 865
 866     va_start(ap, type);
 867     arg = va_arg(ap, void *);
 868     va_end(ap);
 869
 870     ret = ioctl(s->fd, type, arg);
 871     if (ret == -1)
 872         ret = -errno;
 873
 874     return ret;
 875 }
 876
 877 int kvm_vm_ioctl(KVMState *s, int type, ...)
 878 {
 879     int ret;
 880     void *arg;
 881     va_list ap;
 882
 883     va_start(ap, type);
 884     arg = va_arg(ap, void *);
 885     va_end(ap);
 886
 887     ret = ioctl(s->vmfd, type, arg);
 888     if (ret == -1)
 889         ret = -errno;
 890
 891     return ret;
 892 }
 893
 894 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 895 {
 896     int ret;
 897     void *arg;
 898     va_list ap;
 899
 900     va_start(ap, type);
 901     arg = va_arg(ap, void *);
 902     va_end(ap);
 903
 904     ret = ioctl(env->kvm_fd, type, arg);
 905     if (ret == -1)
 906         ret = -errno;
 907
 908     return ret;
 909 }
 910
 911 int kvm_has_sync_mmu(void)
 912 {
 913 #ifdef KVM_CAP_SYNC_MMU
 914     KVMState *s = kvm_state;
 915
 916     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 917 #else
 918     return 0;
 919 #endif
 920 }
 921
 922 int kvm_has_vcpu_events(void)
 923 {
 924     return kvm_state->vcpu_events;
 925 }
 926
 927 int kvm_has_robust_singlestep(void)
 928 {
 929     return kvm_state->robust_singlestep;
 930 }
 931
 932 void kvm_setup_guest_memory(void *start, size_t size)
 933 {
 934     if (!kvm_has_sync_mmu()) {
 935 #ifdef MADV_DONTFORK
 936         int ret = madvise(start, size, MADV_DONTFORK);
 937
 938         if (ret) {
 939             perror("madvice");
 940             exit(1);
 941         }
 942 #else
 943         fprintf(stderr,
 944                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 945         exit(1);
 946 #endif
 947     }
 948 }
 949
 950 #ifdef KVM_CAP_SET_GUEST_DEBUG
 951 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 952 {
 953 #ifdef CONFIG_IOTHREAD
 954     if (env != cpu_single_env) {
 955         abort();
 956     }
 957 #endif
 958     func(data);
 959 }
 960
 961 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 962                                                  target_ulong pc)
 963 {
 964     struct kvm_sw_breakpoint *bp;
 965
 966     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 967         if (bp->pc == pc)
 968             return bp;
 969     }
 970     return NULL;
 971 }
 972
 973 int kvm_sw_breakpoints_active(CPUState *env)
 974 {
 975     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 976 }
 977
 978 struct kvm_set_guest_debug_data {
 979     struct kvm_guest_debug dbg;
 980     CPUState *env;
 981     int err;
 982 };
 983
 984 static void kvm_invoke_set_guest_debug(void *data)
 985 {
 986     struct kvm_set_guest_debug_data *dbg_data = data;
 987     CPUState *env = dbg_data->env;
 988
 989     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
 990 }
 991
 992 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
 993 {
 994     struct kvm_set_guest_debug_data data;
 995
 996     data.dbg.control = reinject_trap;
 997
 998     if (env->singlestep_enabled) {
 999         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1000     }
1001     kvm_arch_update_guest_debug(env, &data.dbg);
1002     data.env = env;
1003
1004     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
1005     return data.err;
1006 }
1007
1008 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1009                           target_ulong len, int type)
1010 {
1011     struct kvm_sw_breakpoint *bp;
1012     CPUState *env;
1013     int err;
1014
1015     if (type == GDB_BREAKPOINT_SW) {
1016         bp = kvm_find_sw_breakpoint(current_env, addr);
1017         if (bp) {
1018             bp->use_count++;
1019             return 0;
1020         }
1021
1022         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1023         if (!bp)
1024             return -ENOMEM;
1025
1026         bp->pc = addr;
1027         bp->use_count = 1;
1028         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1029         if (err) {
1030             free(bp);
1031             return err;
1032         }
1033
1034         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1035                           bp, entry);
1036     } else {
1037         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1038         if (err)
1039             return err;
1040     }
1041
1042     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1043         err = kvm_update_guest_debug(env, 0);
1044         if (err)
1045             return err;
1046     }
1047     return 0;
1048 }
1049
1050 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1051                           target_ulong len, int type)
1052 {
1053     struct kvm_sw_breakpoint *bp;
1054     CPUState *env;
1055     int err;
1056
1057     if (type == GDB_BREAKPOINT_SW) {
1058         bp = kvm_find_sw_breakpoint(current_env, addr);
1059         if (!bp)
1060             return -ENOENT;
1061
1062         if (bp->use_count > 1) {
1063             bp->use_count--;
1064             return 0;
1065         }
1066
1067         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1068         if (err)
1069             return err;
1070
1071         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1072         qemu_free(bp);
1073     } else {
1074         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1075         if (err)
1076             return err;
1077     }
1078
1079     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1080         err = kvm_update_guest_debug(env, 0);
1081         if (err)
1082             return err;
1083     }
1084     return 0;
1085 }
1086
1087 void kvm_remove_all_breakpoints(CPUState *current_env)
1088 {
1089     struct kvm_sw_breakpoint *bp, *next;
1090     KVMState *s = current_env->kvm_state;
1091     CPUState *env;
1092
1093     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1094         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1095             /* Try harder to find a CPU that currently sees the breakpoint. */
1096             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1097                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1098                     break;
1099             }
1100         }
1101     }
1102     kvm_arch_remove_all_hw_breakpoints();
1103
1104     for (env = first_cpu; env != NULL; env = env->next_cpu)
1105         kvm_update_guest_debug(env, 0);
1106 }
1107
1108 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1109
1110 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1111 {
1112     return -EINVAL;
1113 }
1114
1115 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1116                           target_ulong len, int type)
1117 {
1118     return -EINVAL;
1119 }
1120
1121 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1122                           target_ulong len, int type)
1123 {
1124     return -EINVAL;
1125 }
1126
1127 void kvm_remove_all_breakpoints(CPUState *current_env)
1128 {
1129 }
1130 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1131
1132 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1133 {
1134     struct kvm_signal_mask *sigmask;
1135     int r;
1136
1137     if (!sigset)
1138         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1139
1140     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1141
1142     sigmask->len = 8;
1143     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1144     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1145     free(sigmask);
1146
1147     return r;
1148 }