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 #include "bswap.h"
  30
  31 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  32 #define PAGE_SIZE TARGET_PAGE_SIZE
  33
  34 //#define DEBUG_KVM
  35
  36 #ifdef DEBUG_KVM
  37 #define DPRINTF(fmt, ...) \
  38     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  39 #else
  40 #define DPRINTF(fmt, ...) \
  41     do { } while (0)
  42 #endif
  43
  44 typedef struct KVMSlot
  45 {
  46     target_phys_addr_t start_addr;
  47     ram_addr_t memory_size;
  48     ram_addr_t phys_offset;
  49     int slot;
  50     int flags;
  51 } KVMSlot;
  52
  53 typedef struct kvm_dirty_log KVMDirtyLog;
  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     int robust_singlestep;
  68     int debugregs;
  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     int xsave, xcrs;
  75 };
  76
  77 static KVMState *kvm_state;
  78
  79 static KVMSlot *kvm_alloc_slot(KVMState *s)
  80 {
  81     int i;
  82
  83     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  84         /* KVM private memory slots */
  85         if (i >= 8 && i < 12)
  86             continue;
  87         if (s->slots[i].memory_size == 0)
  88             return &s->slots[i];
  89     }
  90
  91     fprintf(stderr, "%s: no free slot available\n", __func__);
  92     abort();
  93 }
  94
  95 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  96                                          target_phys_addr_t start_addr,
  97                                          target_phys_addr_t end_addr)
  98 {
  99     int i;
 100
 101     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 102         KVMSlot *mem = &s->slots[i];
 103
 104         if (start_addr == mem->start_addr &&
 105             end_addr == mem->start_addr + mem->memory_size) {
 106             return mem;
 107         }
 108     }
 109
 110     return NULL;
 111 }
 112
 113 /*
 114  * Find overlapping slot with lowest start address
 115  */
 116 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 117                                             target_phys_addr_t start_addr,
 118                                             target_phys_addr_t end_addr)
 119 {
 120     KVMSlot *found = NULL;
 121     int i;
 122
 123     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 124         KVMSlot *mem = &s->slots[i];
 125
 126         if (mem->memory_size == 0 ||
 127             (found && found->start_addr < mem->start_addr)) {
 128             continue;
 129         }
 130
 131         if (end_addr > mem->start_addr &&
 132             start_addr < mem->start_addr + mem->memory_size) {
 133             found = mem;
 134         }
 135     }
 136
 137     return found;
 138 }
 139
 140 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 141 {
 142     struct kvm_userspace_memory_region mem;
 143
 144     mem.slot = slot->slot;
 145     mem.guest_phys_addr = slot->start_addr;
 146     mem.memory_size = slot->memory_size;
 147     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 148     mem.flags = slot->flags;
 149     if (s->migration_log) {
 150         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 151     }
 152     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 153 }
 154
 155 static void kvm_reset_vcpu(void *opaque)
 156 {
 157     CPUState *env = opaque;
 158
 159     kvm_arch_reset_vcpu(env);
 160 }
 161
 162 int kvm_irqchip_in_kernel(void)
 163 {
 164     return kvm_state->irqchip_in_kernel;
 165 }
 166
 167 int kvm_pit_in_kernel(void)
 168 {
 169     return kvm_state->pit_in_kernel;
 170 }
 171
 172
 173 int kvm_init_vcpu(CPUState *env)
 174 {
 175     KVMState *s = kvm_state;
 176     long mmap_size;
 177     int ret;
 178
 179     DPRINTF("kvm_init_vcpu\n");
 180
 181     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 182     if (ret < 0) {
 183         DPRINTF("kvm_create_vcpu failed\n");
 184         goto err;
 185     }
 186
 187     env->kvm_fd = ret;
 188     env->kvm_state = s;
 189
 190     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 191     if (mmap_size < 0) {
 192         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 193         goto err;
 194     }
 195
 196     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 197                         env->kvm_fd, 0);
 198     if (env->kvm_run == MAP_FAILED) {
 199         ret = -errno;
 200         DPRINTF("mmap'ing vcpu state failed\n");
 201         goto err;
 202     }
 203
 204 #ifdef KVM_CAP_COALESCED_MMIO
 205     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 206         s->coalesced_mmio_ring = (void *) env->kvm_run +
 207                 s->coalesced_mmio * PAGE_SIZE;
 208 #endif
 209
 210     ret = kvm_arch_init_vcpu(env);
 211     if (ret == 0) {
 212         qemu_register_reset(kvm_reset_vcpu, env);
 213         kvm_arch_reset_vcpu(env);
 214     }
 215 err:
 216     return ret;
 217 }
 218
 219 /*
 220  * dirty pages logging control
 221  */
 222 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 223                                       ram_addr_t size, int flags, int mask)
 224 {
 225     KVMState *s = kvm_state;
 226     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 227     int old_flags;
 228
 229     if (mem == NULL)  {
 230             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 231                     TARGET_FMT_plx "\n", __func__, phys_addr,
 232                     (target_phys_addr_t)(phys_addr + size - 1));
 233             return -EINVAL;
 234     }
 235
 236     old_flags = mem->flags;
 237
 238     flags = (mem->flags & ~mask) | flags;
 239     mem->flags = flags;
 240
 241     /* If nothing changed effectively, no need to issue ioctl */
 242     if (s->migration_log) {
 243         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 244     }
 245     if (flags == old_flags) {
 246             return 0;
 247     }
 248
 249     return kvm_set_user_memory_region(s, mem);
 250 }
 251
 252 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 253 {
 254         return kvm_dirty_pages_log_change(phys_addr, size,
 255                                           KVM_MEM_LOG_DIRTY_PAGES,
 256                                           KVM_MEM_LOG_DIRTY_PAGES);
 257 }
 258
 259 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 260 {
 261         return kvm_dirty_pages_log_change(phys_addr, size,
 262                                           0,
 263                                           KVM_MEM_LOG_DIRTY_PAGES);
 264 }
 265
 266 static int kvm_set_migration_log(int enable)
 267 {
 268     KVMState *s = kvm_state;
 269     KVMSlot *mem;
 270     int i, err;
 271
 272     s->migration_log = enable;
 273
 274     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 275         mem = &s->slots[i];
 276
 277         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 278             continue;
 279         }
 280         err = kvm_set_user_memory_region(s, mem);
 281         if (err) {
 282             return err;
 283         }
 284     }
 285     return 0;
 286 }
 287
 288 /* get kvm's dirty pages bitmap and update qemu's */
 289 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 290                                          unsigned long *bitmap,
 291                                          unsigned long offset,
 292                                          unsigned long mem_size)
 293 {
 294     unsigned int i, j;
 295     unsigned long page_number, addr, addr1, c;
 296     ram_addr_t ram_addr;
 297     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 298         HOST_LONG_BITS;
 299
 300     /*
 301      * bitmap-traveling is faster than memory-traveling (for addr...)
 302      * especially when most of the memory is not dirty.
 303      */
 304     for (i = 0; i < len; i++) {
 305         if (bitmap[i] != 0) {
 306             c = leul_to_cpu(bitmap[i]);
 307             do {
 308                 j = ffsl(c) - 1;
 309                 c &= ~(1ul << j);
 310                 page_number = i * HOST_LONG_BITS + j;
 311                 addr1 = page_number * TARGET_PAGE_SIZE;
 312                 addr = offset + addr1;
 313                 ram_addr = cpu_get_physical_page_desc(addr);
 314                 cpu_physical_memory_set_dirty(ram_addr);
 315             } while (c != 0);
 316         }
 317     }
 318     return 0;
 319 }
 320
 321 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 322
 323 /**
 324  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 325  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 326  * This means all bits are set to dirty.
 327  *
 328  * @start_add: start of logged region.
 329  * @end_addr: end of logged region.
 330  */
 331 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 332                                           target_phys_addr_t end_addr)
 333 {
 334     KVMState *s = kvm_state;
 335     unsigned long size, allocated_size = 0;
 336     KVMDirtyLog d;
 337     KVMSlot *mem;
 338     int ret = 0;
 339
 340     d.dirty_bitmap = NULL;
 341     while (start_addr < end_addr) {
 342         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 343         if (mem == NULL) {
 344             break;
 345         }
 346
 347         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 348         if (!d.dirty_bitmap) {
 349             d.dirty_bitmap = qemu_malloc(size);
 350         } else if (size > allocated_size) {
 351             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 352         }
 353         allocated_size = size;
 354         memset(d.dirty_bitmap, 0, allocated_size);
 355
 356         d.slot = mem->slot;
 357
 358         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 359             DPRINTF("ioctl failed %d\n", errno);
 360             ret = -1;
 361             break;
 362         }
 363
 364         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 365                                       mem->start_addr, mem->memory_size);
 366         start_addr = mem->start_addr + mem->memory_size;
 367     }
 368     qemu_free(d.dirty_bitmap);
 369
 370     return ret;
 371 }
 372
 373 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 374 {
 375     int ret = -ENOSYS;
 376 #ifdef KVM_CAP_COALESCED_MMIO
 377     KVMState *s = kvm_state;
 378
 379     if (s->coalesced_mmio) {
 380         struct kvm_coalesced_mmio_zone zone;
 381
 382         zone.addr = start;
 383         zone.size = size;
 384
 385         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 386     }
 387 #endif
 388
 389     return ret;
 390 }
 391
 392 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 393 {
 394     int ret = -ENOSYS;
 395 #ifdef KVM_CAP_COALESCED_MMIO
 396     KVMState *s = kvm_state;
 397
 398     if (s->coalesced_mmio) {
 399         struct kvm_coalesced_mmio_zone zone;
 400
 401         zone.addr = start;
 402         zone.size = size;
 403
 404         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 405     }
 406 #endif
 407
 408     return ret;
 409 }
 410
 411 int kvm_check_extension(KVMState *s, unsigned int extension)
 412 {
 413     int ret;
 414
 415     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 416     if (ret < 0) {
 417         ret = 0;
 418     }
 419
 420     return ret;
 421 }
 422
 423 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 424                              ram_addr_t size,
 425                              ram_addr_t phys_offset)
 426 {
 427     KVMState *s = kvm_state;
 428     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 429     KVMSlot *mem, old;
 430     int err;
 431
 432     if (start_addr & ~TARGET_PAGE_MASK) {
 433         if (flags >= IO_MEM_UNASSIGNED) {
 434             if (!kvm_lookup_overlapping_slot(s, start_addr,
 435                                              start_addr + size)) {
 436                 return;
 437             }
 438             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 439         } else {
 440             fprintf(stderr, "Only page-aligned memory slots supported\n");
 441         }
 442         abort();
 443     }
 444
 445     /* KVM does not support read-only slots */
 446     phys_offset &= ~IO_MEM_ROM;
 447
 448     while (1) {
 449         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 450         if (!mem) {
 451             break;
 452         }
 453
 454         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 455             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 456             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 457             /* The new slot fits into the existing one and comes with
 458              * identical parameters - nothing to be done. */
 459             return;
 460         }
 461
 462         old = *mem;
 463
 464         /* unregister the overlapping slot */
 465         mem->memory_size = 0;
 466         err = kvm_set_user_memory_region(s, mem);
 467         if (err) {
 468             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 469                     __func__, strerror(-err));
 470             abort();
 471         }
 472
 473         /* Workaround for older KVM versions: we can't join slots, even not by
 474          * unregistering the previous ones and then registering the larger
 475          * slot. We have to maintain the existing fragmentation. Sigh.
 476          *
 477          * This workaround assumes that the new slot starts at the same
 478          * address as the first existing one. If not or if some overlapping
 479          * slot comes around later, we will fail (not seen in practice so far)
 480          * - and actually require a recent KVM version. */
 481         if (s->broken_set_mem_region &&
 482             old.start_addr == start_addr && old.memory_size < size &&
 483             flags < IO_MEM_UNASSIGNED) {
 484             mem = kvm_alloc_slot(s);
 485             mem->memory_size = old.memory_size;
 486             mem->start_addr = old.start_addr;
 487             mem->phys_offset = old.phys_offset;
 488             mem->flags = 0;
 489
 490             err = kvm_set_user_memory_region(s, mem);
 491             if (err) {
 492                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 493                         strerror(-err));
 494                 abort();
 495             }
 496
 497             start_addr += old.memory_size;
 498             phys_offset += old.memory_size;
 499             size -= old.memory_size;
 500             continue;
 501         }
 502
 503         /* register prefix slot */
 504         if (old.start_addr < start_addr) {
 505             mem = kvm_alloc_slot(s);
 506             mem->memory_size = start_addr - old.start_addr;
 507             mem->start_addr = old.start_addr;
 508             mem->phys_offset = old.phys_offset;
 509             mem->flags = 0;
 510
 511             err = kvm_set_user_memory_region(s, mem);
 512             if (err) {
 513                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 514                         __func__, strerror(-err));
 515                 abort();
 516             }
 517         }
 518
 519         /* register suffix slot */
 520         if (old.start_addr + old.memory_size > start_addr + size) {
 521             ram_addr_t size_delta;
 522
 523             mem = kvm_alloc_slot(s);
 524             mem->start_addr = start_addr + size;
 525             size_delta = mem->start_addr - old.start_addr;
 526             mem->memory_size = old.memory_size - size_delta;
 527             mem->phys_offset = old.phys_offset + size_delta;
 528             mem->flags = 0;
 529
 530             err = kvm_set_user_memory_region(s, mem);
 531             if (err) {
 532                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 533                         __func__, strerror(-err));
 534                 abort();
 535             }
 536         }
 537     }
 538
 539     /* in case the KVM bug workaround already "consumed" the new slot */
 540     if (!size)
 541         return;
 542
 543     /* KVM does not need to know about this memory */
 544     if (flags >= IO_MEM_UNASSIGNED)
 545         return;
 546
 547     mem = kvm_alloc_slot(s);
 548     mem->memory_size = size;
 549     mem->start_addr = start_addr;
 550     mem->phys_offset = phys_offset;
 551     mem->flags = 0;
 552
 553     err = kvm_set_user_memory_region(s, mem);
 554     if (err) {
 555         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 556                 strerror(-err));
 557         abort();
 558     }
 559 }
 560
 561 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 562                                   target_phys_addr_t start_addr,
 563                                   ram_addr_t size,
 564                                   ram_addr_t phys_offset)
 565 {
 566         kvm_set_phys_mem(start_addr, size, phys_offset);
 567 }
 568
 569 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 570                                         target_phys_addr_t start_addr,
 571                                         target_phys_addr_t end_addr)
 572 {
 573         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 574 }
 575
 576 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 577                                     int enable)
 578 {
 579         return kvm_set_migration_log(enable);
 580 }
 581
 582 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 583         .set_memory = kvm_client_set_memory,
 584         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 585         .migration_log = kvm_client_migration_log,
 586 };
 587
 588 int kvm_init(int smp_cpus)
 589 {
 590     static const char upgrade_note[] =
 591         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 592         "(see http://sourceforge.net/projects/kvm).\n";
 593     KVMState *s;
 594     int ret;
 595     int i;
 596
 597     s = qemu_mallocz(sizeof(KVMState));
 598
 599 #ifdef KVM_CAP_SET_GUEST_DEBUG
 600     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 601 #endif
 602     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 603         s->slots[i].slot = i;
 604
 605     s->vmfd = -1;
 606     s->fd = qemu_open("/dev/kvm", O_RDWR);
 607     if (s->fd == -1) {
 608         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 609         ret = -errno;
 610         goto err;
 611     }
 612
 613     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 614     if (ret < KVM_API_VERSION) {
 615         if (ret > 0)
 616             ret = -EINVAL;
 617         fprintf(stderr, "kvm version too old\n");
 618         goto err;
 619     }
 620
 621     if (ret > KVM_API_VERSION) {
 622         ret = -EINVAL;
 623         fprintf(stderr, "kvm version not supported\n");
 624         goto err;
 625     }
 626
 627     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 628     if (s->vmfd < 0) {
 629 #ifdef TARGET_S390X
 630         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 631                         "your host kernel command line\n");
 632 #endif
 633         goto err;
 634     }
 635
 636     /* initially, KVM allocated its own memory and we had to jump through
 637      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 638      * just use a user allocated buffer so we can use regular pages
 639      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 640      */
 641     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 642         ret = -EINVAL;
 643         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 644                 upgrade_note);
 645         goto err;
 646     }
 647
 648     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 649      * destroyed properly.  Since we rely on this capability, refuse to work
 650      * with any kernel without this capability. */
 651     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 652         ret = -EINVAL;
 653
 654         fprintf(stderr,
 655                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 656                 upgrade_note);
 657         goto err;
 658     }
 659
 660     s->coalesced_mmio = 0;
 661 #ifdef KVM_CAP_COALESCED_MMIO
 662     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 663     s->coalesced_mmio_ring = NULL;
 664 #endif
 665
 666     s->broken_set_mem_region = 1;
 667 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 668     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 669     if (ret > 0) {
 670         s->broken_set_mem_region = 0;
 671     }
 672 #endif
 673
 674     s->vcpu_events = 0;
 675 #ifdef KVM_CAP_VCPU_EVENTS
 676     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 677 #endif
 678
 679     s->robust_singlestep = 0;
 680 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 681     s->robust_singlestep =
 682         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 683 #endif
 684
 685     s->debugregs = 0;
 686 #ifdef KVM_CAP_DEBUGREGS
 687     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 688 #endif
 689
 690     s->xsave = 0;
 691 #ifdef KVM_CAP_XSAVE
 692     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 693 #endif
 694
 695     s->xcrs = 0;
 696 #ifdef KVM_CAP_XCRS
 697     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 698 #endif
 699
 700     ret = kvm_arch_init(s, smp_cpus);
 701     if (ret < 0)
 702         goto err;
 703
 704     kvm_state = s;
 705     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 706
 707     return 0;
 708
 709 err:
 710     if (s) {
 711         if (s->vmfd != -1)
 712             close(s->vmfd);
 713         if (s->fd != -1)
 714             close(s->fd);
 715     }
 716     qemu_free(s);
 717
 718     return ret;
 719 }
 720
 721 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 722                          uint32_t count)
 723 {
 724     int i;
 725     uint8_t *ptr = data;
 726
 727     for (i = 0; i < count; i++) {
 728         if (direction == KVM_EXIT_IO_IN) {
 729             switch (size) {
 730             case 1:
 731                 stb_p(ptr, cpu_inb(port));
 732                 break;
 733             case 2:
 734                 stw_p(ptr, cpu_inw(port));
 735                 break;
 736             case 4:
 737                 stl_p(ptr, cpu_inl(port));
 738                 break;
 739             }
 740         } else {
 741             switch (size) {
 742             case 1:
 743                 cpu_outb(port, ldub_p(ptr));
 744                 break;
 745             case 2:
 746                 cpu_outw(port, lduw_p(ptr));
 747                 break;
 748             case 4:
 749                 cpu_outl(port, ldl_p(ptr));
 750                 break;
 751             }
 752         }
 753
 754         ptr += size;
 755     }
 756
 757     return 1;
 758 }
 759
 760 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 761 static void kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 762 {
 763
 764     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 765         int i;
 766
 767         fprintf(stderr, "KVM internal error. Suberror: %d\n",
 768                 run->internal.suberror);
 769
 770         for (i = 0; i < run->internal.ndata; ++i) {
 771             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 772                     i, (uint64_t)run->internal.data[i]);
 773         }
 774     }
 775     cpu_dump_state(env, stderr, fprintf, 0);
 776     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 777         fprintf(stderr, "emulation failure\n");
 778         if (!kvm_arch_stop_on_emulation_error(env))
 779                 return;
 780     }
 781     /* FIXME: Should trigger a qmp message to let management know
 782      * something went wrong.
 783      */
 784     vm_stop(0);
 785 }
 786 #endif
 787
 788 void kvm_flush_coalesced_mmio_buffer(void)
 789 {
 790 #ifdef KVM_CAP_COALESCED_MMIO
 791     KVMState *s = kvm_state;
 792     if (s->coalesced_mmio_ring) {
 793         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 794         while (ring->first != ring->last) {
 795             struct kvm_coalesced_mmio *ent;
 796
 797             ent = &ring->coalesced_mmio[ring->first];
 798
 799             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 800             smp_wmb();
 801             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 802         }
 803     }
 804 #endif
 805 }
 806
 807 static void do_kvm_cpu_synchronize_state(void *_env)
 808 {
 809     CPUState *env = _env;
 810
 811     if (!env->kvm_vcpu_dirty) {
 812         kvm_arch_get_registers(env);
 813         env->kvm_vcpu_dirty = 1;
 814     }
 815 }
 816
 817 void kvm_cpu_synchronize_state(CPUState *env)
 818 {
 819     if (!env->kvm_vcpu_dirty)
 820         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 821 }
 822
 823 void kvm_cpu_synchronize_post_reset(CPUState *env)
 824 {
 825     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 826     env->kvm_vcpu_dirty = 0;
 827 }
 828
 829 void kvm_cpu_synchronize_post_init(CPUState *env)
 830 {
 831     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 832     env->kvm_vcpu_dirty = 0;
 833 }
 834
 835 int kvm_cpu_exec(CPUState *env)
 836 {
 837     struct kvm_run *run = env->kvm_run;
 838     int ret;
 839
 840     DPRINTF("kvm_cpu_exec()\n");
 841
 842     do {
 843 #ifndef CONFIG_IOTHREAD
 844         if (env->exit_request) {
 845             DPRINTF("interrupt exit requested\n");
 846             ret = 0;
 847             break;
 848         }
 849 #endif
 850
 851         if (kvm_arch_process_irqchip_events(env)) {
 852             ret = 0;
 853             break;
 854         }
 855
 856         if (env->kvm_vcpu_dirty) {
 857             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 858             env->kvm_vcpu_dirty = 0;
 859         }
 860
 861         kvm_arch_pre_run(env, run);
 862         cpu_single_env = NULL;
 863         qemu_mutex_unlock_iothread();
 864         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 865         qemu_mutex_lock_iothread();
 866         cpu_single_env = env;
 867         kvm_arch_post_run(env, run);
 868
 869         if (ret == -EINTR || ret == -EAGAIN) {
 870             cpu_exit(env);
 871             DPRINTF("io window exit\n");
 872             ret = 0;
 873             break;
 874         }
 875
 876         if (ret < 0) {
 877             DPRINTF("kvm run failed %s\n", strerror(-ret));
 878             abort();
 879         }
 880
 881         kvm_flush_coalesced_mmio_buffer();
 882
 883         ret = 0; /* exit loop */
 884         switch (run->exit_reason) {
 885         case KVM_EXIT_IO:
 886             DPRINTF("handle_io\n");
 887             ret = kvm_handle_io(run->io.port,
 888                                 (uint8_t *)run + run->io.data_offset,
 889                                 run->io.direction,
 890                                 run->io.size,
 891                                 run->io.count);
 892             break;
 893         case KVM_EXIT_MMIO:
 894             DPRINTF("handle_mmio\n");
 895             cpu_physical_memory_rw(run->mmio.phys_addr,
 896                                    run->mmio.data,
 897                                    run->mmio.len,
 898                                    run->mmio.is_write);
 899             ret = 1;
 900             break;
 901         case KVM_EXIT_IRQ_WINDOW_OPEN:
 902             DPRINTF("irq_window_open\n");
 903             break;
 904         case KVM_EXIT_SHUTDOWN:
 905             DPRINTF("shutdown\n");
 906             qemu_system_reset_request();
 907             ret = 1;
 908             break;
 909         case KVM_EXIT_UNKNOWN:
 910             DPRINTF("kvm_exit_unknown\n");
 911             break;
 912         case KVM_EXIT_FAIL_ENTRY:
 913             DPRINTF("kvm_exit_fail_entry\n");
 914             break;
 915         case KVM_EXIT_EXCEPTION:
 916             DPRINTF("kvm_exit_exception\n");
 917             break;
 918 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 919         case KVM_EXIT_INTERNAL_ERROR:
 920             kvm_handle_internal_error(env, run);
 921             break;
 922 #endif
 923         case KVM_EXIT_DEBUG:
 924             DPRINTF("kvm_exit_debug\n");
 925 #ifdef KVM_CAP_SET_GUEST_DEBUG
 926             if (kvm_arch_debug(&run->debug.arch)) {
 927                 env->exception_index = EXCP_DEBUG;
 928                 return 0;
 929             }
 930             /* re-enter, this exception was guest-internal */
 931             ret = 1;
 932 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 933             break;
 934         default:
 935             DPRINTF("kvm_arch_handle_exit\n");
 936             ret = kvm_arch_handle_exit(env, run);
 937             break;
 938         }
 939     } while (ret > 0);
 940
 941     if (env->exit_request) {
 942         env->exit_request = 0;
 943         env->exception_index = EXCP_INTERRUPT;
 944     }
 945
 946     return ret;
 947 }
 948
 949 int kvm_ioctl(KVMState *s, int type, ...)
 950 {
 951     int ret;
 952     void *arg;
 953     va_list ap;
 954
 955     va_start(ap, type);
 956     arg = va_arg(ap, void *);
 957     va_end(ap);
 958
 959     ret = ioctl(s->fd, type, arg);
 960     if (ret == -1)
 961         ret = -errno;
 962
 963     return ret;
 964 }
 965
 966 int kvm_vm_ioctl(KVMState *s, int type, ...)
 967 {
 968     int ret;
 969     void *arg;
 970     va_list ap;
 971
 972     va_start(ap, type);
 973     arg = va_arg(ap, void *);
 974     va_end(ap);
 975
 976     ret = ioctl(s->vmfd, type, arg);
 977     if (ret == -1)
 978         ret = -errno;
 979
 980     return ret;
 981 }
 982
 983 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 984 {
 985     int ret;
 986     void *arg;
 987     va_list ap;
 988
 989     va_start(ap, type);
 990     arg = va_arg(ap, void *);
 991     va_end(ap);
 992
 993     ret = ioctl(env->kvm_fd, type, arg);
 994     if (ret == -1)
 995         ret = -errno;
 996
 997     return ret;
 998 }
 999
1000 int kvm_has_sync_mmu(void)
1001 {
1002 #ifdef KVM_CAP_SYNC_MMU
1003     KVMState *s = kvm_state;
1004
1005     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
1006 #else
1007     return 0;
1008 #endif
1009 }
1010
1011 int kvm_has_vcpu_events(void)
1012 {
1013     return kvm_state->vcpu_events;
1014 }
1015
1016 int kvm_has_robust_singlestep(void)
1017 {
1018     return kvm_state->robust_singlestep;
1019 }
1020
1021 int kvm_has_debugregs(void)
1022 {
1023     return kvm_state->debugregs;
1024 }
1025
1026 int kvm_has_xsave(void)
1027 {
1028     return kvm_state->xsave;
1029 }
1030
1031 int kvm_has_xcrs(void)
1032 {
1033     return kvm_state->xcrs;
1034 }
1035
1036 void kvm_setup_guest_memory(void *start, size_t size)
1037 {
1038     if (!kvm_has_sync_mmu()) {
1039 #ifdef MADV_DONTFORK
1040         int ret = madvise(start, size, MADV_DONTFORK);
1041
1042         if (ret) {
1043             perror("madvice");
1044             exit(1);
1045         }
1046 #else
1047         fprintf(stderr,
1048                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1049         exit(1);
1050 #endif
1051     }
1052 }
1053
1054 #ifdef KVM_CAP_SET_GUEST_DEBUG
1055 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1056                                                  target_ulong pc)
1057 {
1058     struct kvm_sw_breakpoint *bp;
1059
1060     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1061         if (bp->pc == pc)
1062             return bp;
1063     }
1064     return NULL;
1065 }
1066
1067 int kvm_sw_breakpoints_active(CPUState *env)
1068 {
1069     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1070 }
1071
1072 struct kvm_set_guest_debug_data {
1073     struct kvm_guest_debug dbg;
1074     CPUState *env;
1075     int err;
1076 };
1077
1078 static void kvm_invoke_set_guest_debug(void *data)
1079 {
1080     struct kvm_set_guest_debug_data *dbg_data = data;
1081     CPUState *env = dbg_data->env;
1082
1083     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1084 }
1085
1086 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1087 {
1088     struct kvm_set_guest_debug_data data;
1089
1090     data.dbg.control = reinject_trap;
1091
1092     if (env->singlestep_enabled) {
1093         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1094     }
1095     kvm_arch_update_guest_debug(env, &data.dbg);
1096     data.env = env;
1097
1098     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1099     return data.err;
1100 }
1101
1102 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1103                           target_ulong len, int type)
1104 {
1105     struct kvm_sw_breakpoint *bp;
1106     CPUState *env;
1107     int err;
1108
1109     if (type == GDB_BREAKPOINT_SW) {
1110         bp = kvm_find_sw_breakpoint(current_env, addr);
1111         if (bp) {
1112             bp->use_count++;
1113             return 0;
1114         }
1115
1116         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1117         if (!bp)
1118             return -ENOMEM;
1119
1120         bp->pc = addr;
1121         bp->use_count = 1;
1122         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1123         if (err) {
1124             free(bp);
1125             return err;
1126         }
1127
1128         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1129                           bp, entry);
1130     } else {
1131         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1132         if (err)
1133             return err;
1134     }
1135
1136     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1137         err = kvm_update_guest_debug(env, 0);
1138         if (err)
1139             return err;
1140     }
1141     return 0;
1142 }
1143
1144 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1145                           target_ulong len, int type)
1146 {
1147     struct kvm_sw_breakpoint *bp;
1148     CPUState *env;
1149     int err;
1150
1151     if (type == GDB_BREAKPOINT_SW) {
1152         bp = kvm_find_sw_breakpoint(current_env, addr);
1153         if (!bp)
1154             return -ENOENT;
1155
1156         if (bp->use_count > 1) {
1157             bp->use_count--;
1158             return 0;
1159         }
1160
1161         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1162         if (err)
1163             return err;
1164
1165         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1166         qemu_free(bp);
1167     } else {
1168         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1169         if (err)
1170             return err;
1171     }
1172
1173     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1174         err = kvm_update_guest_debug(env, 0);
1175         if (err)
1176             return err;
1177     }
1178     return 0;
1179 }
1180
1181 void kvm_remove_all_breakpoints(CPUState *current_env)
1182 {
1183     struct kvm_sw_breakpoint *bp, *next;
1184     KVMState *s = current_env->kvm_state;
1185     CPUState *env;
1186
1187     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1188         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1189             /* Try harder to find a CPU that currently sees the breakpoint. */
1190             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1191                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1192                     break;
1193             }
1194         }
1195     }
1196     kvm_arch_remove_all_hw_breakpoints();
1197
1198     for (env = first_cpu; env != NULL; env = env->next_cpu)
1199         kvm_update_guest_debug(env, 0);
1200 }
1201
1202 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1203
1204 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1205 {
1206     return -EINVAL;
1207 }
1208
1209 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1210                           target_ulong len, int type)
1211 {
1212     return -EINVAL;
1213 }
1214
1215 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1216                           target_ulong len, int type)
1217 {
1218     return -EINVAL;
1219 }
1220
1221 void kvm_remove_all_breakpoints(CPUState *current_env)
1222 {
1223 }
1224 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1225
1226 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1227 {
1228     struct kvm_signal_mask *sigmask;
1229     int r;
1230
1231     if (!sigset)
1232         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1233
1234     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1235
1236     sigmask->len = 8;
1237     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1238     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1239     free(sigmask);
1240
1241     return r;
1242 }
1243
1244 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1245 {
1246 #ifdef KVM_IOEVENTFD
1247     struct kvm_ioeventfd kick = {
1248         .datamatch = val,
1249         .addr = addr,
1250         .len = 2,
1251         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1252         .fd = fd,
1253     };
1254     int r;
1255     if (!kvm_enabled())
1256         return -ENOSYS;
1257     if (!assign)
1258         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1259     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1260     if (r < 0)
1261         return r;
1262     return 0;
1263 #else
1264     return -ENOSYS;
1265 #endif
1266 }