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 struct KVMState
  55 {
  56     KVMSlot slots[32];
  57     int fd;
  58     int vmfd;
  59     int coalesced_mmio;
  60 #ifdef KVM_CAP_COALESCED_MMIO
  61     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  62 #endif
  63     int broken_set_mem_region;
  64     int migration_log;
  65     int vcpu_events;
  66     int robust_singlestep;
  67     int debugregs;
  68 #ifdef KVM_CAP_SET_GUEST_DEBUG
  69     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  70 #endif
  71     int irqchip_in_kernel;
  72     int pit_in_kernel;
  73 };
  74
  75 static KVMState *kvm_state;
  76
  77 static KVMSlot *kvm_alloc_slot(KVMState *s)
  78 {
  79     int i;
  80
  81     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  82         /* KVM private memory slots */
  83         if (i >= 8 && i < 12)
  84             continue;
  85         if (s->slots[i].memory_size == 0)
  86             return &s->slots[i];
  87     }
  88
  89     fprintf(stderr, "%s: no free slot available\n", __func__);
  90     abort();
  91 }
  92
  93 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  94                                          target_phys_addr_t start_addr,
  95                                          target_phys_addr_t end_addr)
  96 {
  97     int i;
  98
  99     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 100         KVMSlot *mem = &s->slots[i];
 101
 102         if (start_addr == mem->start_addr &&
 103             end_addr == mem->start_addr + mem->memory_size) {
 104             return mem;
 105         }
 106     }
 107
 108     return NULL;
 109 }
 110
 111 /*
 112  * Find overlapping slot with lowest start address
 113  */
 114 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 115                                             target_phys_addr_t start_addr,
 116                                             target_phys_addr_t end_addr)
 117 {
 118     KVMSlot *found = NULL;
 119     int i;
 120
 121     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 122         KVMSlot *mem = &s->slots[i];
 123
 124         if (mem->memory_size == 0 ||
 125             (found && found->start_addr < mem->start_addr)) {
 126             continue;
 127         }
 128
 129         if (end_addr > mem->start_addr &&
 130             start_addr < mem->start_addr + mem->memory_size) {
 131             found = mem;
 132         }
 133     }
 134
 135     return found;
 136 }
 137
 138 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 139 {
 140     struct kvm_userspace_memory_region mem;
 141
 142     mem.slot = slot->slot;
 143     mem.guest_phys_addr = slot->start_addr;
 144     mem.memory_size = slot->memory_size;
 145     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 146     mem.flags = slot->flags;
 147     if (s->migration_log) {
 148         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 149     }
 150     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 151 }
 152
 153 static void kvm_reset_vcpu(void *opaque)
 154 {
 155     CPUState *env = opaque;
 156
 157     kvm_arch_reset_vcpu(env);
 158 }
 159
 160 int kvm_irqchip_in_kernel(void)
 161 {
 162     return kvm_state->irqchip_in_kernel;
 163 }
 164
 165 int kvm_pit_in_kernel(void)
 166 {
 167     return kvm_state->pit_in_kernel;
 168 }
 169
 170
 171 int kvm_init_vcpu(CPUState *env)
 172 {
 173     KVMState *s = kvm_state;
 174     long mmap_size;
 175     int ret;
 176
 177     DPRINTF("kvm_init_vcpu\n");
 178
 179     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 180     if (ret < 0) {
 181         DPRINTF("kvm_create_vcpu failed\n");
 182         goto err;
 183     }
 184
 185     env->kvm_fd = ret;
 186     env->kvm_state = s;
 187
 188     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 189     if (mmap_size < 0) {
 190         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 191         goto err;
 192     }
 193
 194     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 195                         env->kvm_fd, 0);
 196     if (env->kvm_run == MAP_FAILED) {
 197         ret = -errno;
 198         DPRINTF("mmap'ing vcpu state failed\n");
 199         goto err;
 200     }
 201
 202 #ifdef KVM_CAP_COALESCED_MMIO
 203     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 204         s->coalesced_mmio_ring = (void *) env->kvm_run +
 205                 s->coalesced_mmio * PAGE_SIZE;
 206 #endif
 207
 208     ret = kvm_arch_init_vcpu(env);
 209     if (ret == 0) {
 210         qemu_register_reset(kvm_reset_vcpu, env);
 211         kvm_arch_reset_vcpu(env);
 212     }
 213 err:
 214     return ret;
 215 }
 216
 217 /*
 218  * dirty pages logging control
 219  */
 220 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 221                                       ram_addr_t size, int flags, int mask)
 222 {
 223     KVMState *s = kvm_state;
 224     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 225     int old_flags;
 226
 227     if (mem == NULL)  {
 228             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 229                     TARGET_FMT_plx "\n", __func__, phys_addr,
 230                     (target_phys_addr_t)(phys_addr + size - 1));
 231             return -EINVAL;
 232     }
 233
 234     old_flags = mem->flags;
 235
 236     flags = (mem->flags & ~mask) | flags;
 237     mem->flags = flags;
 238
 239     /* If nothing changed effectively, no need to issue ioctl */
 240     if (s->migration_log) {
 241         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 242     }
 243     if (flags == old_flags) {
 244             return 0;
 245     }
 246
 247     return kvm_set_user_memory_region(s, mem);
 248 }
 249
 250 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 251 {
 252         return kvm_dirty_pages_log_change(phys_addr, size,
 253                                           KVM_MEM_LOG_DIRTY_PAGES,
 254                                           KVM_MEM_LOG_DIRTY_PAGES);
 255 }
 256
 257 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 258 {
 259         return kvm_dirty_pages_log_change(phys_addr, size,
 260                                           0,
 261                                           KVM_MEM_LOG_DIRTY_PAGES);
 262 }
 263
 264 static int kvm_set_migration_log(int enable)
 265 {
 266     KVMState *s = kvm_state;
 267     KVMSlot *mem;
 268     int i, err;
 269
 270     s->migration_log = enable;
 271
 272     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 273         mem = &s->slots[i];
 274
 275         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 276             continue;
 277         }
 278         err = kvm_set_user_memory_region(s, mem);
 279         if (err) {
 280             return err;
 281         }
 282     }
 283     return 0;
 284 }
 285
 286 static int test_le_bit(unsigned long nr, unsigned char *addr)
 287 {
 288     return (addr[nr >> 3] >> (nr & 7)) & 1;
 289 }
 290
 291 /**
 292  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 293  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 294  * This means all bits are set to dirty.
 295  *
 296  * @start_add: start of logged region.
 297  * @end_addr: end of logged region.
 298  */
 299 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 300                                           target_phys_addr_t end_addr)
 301 {
 302     KVMState *s = kvm_state;
 303     unsigned long size, allocated_size = 0;
 304     target_phys_addr_t phys_addr;
 305     ram_addr_t addr;
 306     KVMDirtyLog d;
 307     KVMSlot *mem;
 308     int ret = 0;
 309
 310     d.dirty_bitmap = NULL;
 311     while (start_addr < end_addr) {
 312         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 313         if (mem == NULL) {
 314             break;
 315         }
 316
 317         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 318         if (!d.dirty_bitmap) {
 319             d.dirty_bitmap = qemu_malloc(size);
 320         } else if (size > allocated_size) {
 321             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 322         }
 323         allocated_size = size;
 324         memset(d.dirty_bitmap, 0, allocated_size);
 325
 326         d.slot = mem->slot;
 327
 328         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 329             DPRINTF("ioctl failed %d\n", errno);
 330             ret = -1;
 331             break;
 332         }
 333
 334         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 335              phys_addr < mem->start_addr + mem->memory_size;
 336              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 337             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 338             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 339
 340             if (test_le_bit(nr, bitmap)) {
 341                 cpu_physical_memory_set_dirty(addr);
 342             }
 343         }
 344         start_addr = phys_addr;
 345     }
 346     qemu_free(d.dirty_bitmap);
 347
 348     return ret;
 349 }
 350
 351 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 352 {
 353     int ret = -ENOSYS;
 354 #ifdef KVM_CAP_COALESCED_MMIO
 355     KVMState *s = kvm_state;
 356
 357     if (s->coalesced_mmio) {
 358         struct kvm_coalesced_mmio_zone zone;
 359
 360         zone.addr = start;
 361         zone.size = size;
 362
 363         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 364     }
 365 #endif
 366
 367     return ret;
 368 }
 369
 370 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 371 {
 372     int ret = -ENOSYS;
 373 #ifdef KVM_CAP_COALESCED_MMIO
 374     KVMState *s = kvm_state;
 375
 376     if (s->coalesced_mmio) {
 377         struct kvm_coalesced_mmio_zone zone;
 378
 379         zone.addr = start;
 380         zone.size = size;
 381
 382         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 383     }
 384 #endif
 385
 386     return ret;
 387 }
 388
 389 int kvm_check_extension(KVMState *s, unsigned int extension)
 390 {
 391     int ret;
 392
 393     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 394     if (ret < 0) {
 395         ret = 0;
 396     }
 397
 398     return ret;
 399 }
 400
 401 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 402                              ram_addr_t size,
 403                              ram_addr_t phys_offset)
 404 {
 405     KVMState *s = kvm_state;
 406     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 407     KVMSlot *mem, old;
 408     int err;
 409
 410     if (start_addr & ~TARGET_PAGE_MASK) {
 411         if (flags >= IO_MEM_UNASSIGNED) {
 412             if (!kvm_lookup_overlapping_slot(s, start_addr,
 413                                              start_addr + size)) {
 414                 return;
 415             }
 416             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 417         } else {
 418             fprintf(stderr, "Only page-aligned memory slots supported\n");
 419         }
 420         abort();
 421     }
 422
 423     /* KVM does not support read-only slots */
 424     phys_offset &= ~IO_MEM_ROM;
 425
 426     while (1) {
 427         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 428         if (!mem) {
 429             break;
 430         }
 431
 432         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 433             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 434             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 435             /* The new slot fits into the existing one and comes with
 436              * identical parameters - nothing to be done. */
 437             return;
 438         }
 439
 440         old = *mem;
 441
 442         /* unregister the overlapping slot */
 443         mem->memory_size = 0;
 444         err = kvm_set_user_memory_region(s, mem);
 445         if (err) {
 446             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 447                     __func__, strerror(-err));
 448             abort();
 449         }
 450
 451         /* Workaround for older KVM versions: we can't join slots, even not by
 452          * unregistering the previous ones and then registering the larger
 453          * slot. We have to maintain the existing fragmentation. Sigh.
 454          *
 455          * This workaround assumes that the new slot starts at the same
 456          * address as the first existing one. If not or if some overlapping
 457          * slot comes around later, we will fail (not seen in practice so far)
 458          * - and actually require a recent KVM version. */
 459         if (s->broken_set_mem_region &&
 460             old.start_addr == start_addr && old.memory_size < size &&
 461             flags < IO_MEM_UNASSIGNED) {
 462             mem = kvm_alloc_slot(s);
 463             mem->memory_size = old.memory_size;
 464             mem->start_addr = old.start_addr;
 465             mem->phys_offset = old.phys_offset;
 466             mem->flags = 0;
 467
 468             err = kvm_set_user_memory_region(s, mem);
 469             if (err) {
 470                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 471                         strerror(-err));
 472                 abort();
 473             }
 474
 475             start_addr += old.memory_size;
 476             phys_offset += old.memory_size;
 477             size -= old.memory_size;
 478             continue;
 479         }
 480
 481         /* register prefix slot */
 482         if (old.start_addr < start_addr) {
 483             mem = kvm_alloc_slot(s);
 484             mem->memory_size = start_addr - old.start_addr;
 485             mem->start_addr = old.start_addr;
 486             mem->phys_offset = old.phys_offset;
 487             mem->flags = 0;
 488
 489             err = kvm_set_user_memory_region(s, mem);
 490             if (err) {
 491                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 492                         __func__, strerror(-err));
 493                 abort();
 494             }
 495         }
 496
 497         /* register suffix slot */
 498         if (old.start_addr + old.memory_size > start_addr + size) {
 499             ram_addr_t size_delta;
 500
 501             mem = kvm_alloc_slot(s);
 502             mem->start_addr = start_addr + size;
 503             size_delta = mem->start_addr - old.start_addr;
 504             mem->memory_size = old.memory_size - size_delta;
 505             mem->phys_offset = old.phys_offset + size_delta;
 506             mem->flags = 0;
 507
 508             err = kvm_set_user_memory_region(s, mem);
 509             if (err) {
 510                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 511                         __func__, strerror(-err));
 512                 abort();
 513             }
 514         }
 515     }
 516
 517     /* in case the KVM bug workaround already "consumed" the new slot */
 518     if (!size)
 519         return;
 520
 521     /* KVM does not need to know about this memory */
 522     if (flags >= IO_MEM_UNASSIGNED)
 523         return;
 524
 525     mem = kvm_alloc_slot(s);
 526     mem->memory_size = size;
 527     mem->start_addr = start_addr;
 528     mem->phys_offset = phys_offset;
 529     mem->flags = 0;
 530
 531     err = kvm_set_user_memory_region(s, mem);
 532     if (err) {
 533         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 534                 strerror(-err));
 535         abort();
 536     }
 537 }
 538
 539 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 540                                   target_phys_addr_t start_addr,
 541                                   ram_addr_t size,
 542                                   ram_addr_t phys_offset)
 543 {
 544         kvm_set_phys_mem(start_addr, size, phys_offset);
 545 }
 546
 547 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 548                                         target_phys_addr_t start_addr,
 549                                         target_phys_addr_t end_addr)
 550 {
 551         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 552 }
 553
 554 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 555                                     int enable)
 556 {
 557         return kvm_set_migration_log(enable);
 558 }
 559
 560 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 561         .set_memory = kvm_client_set_memory,
 562         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 563         .migration_log = kvm_client_migration_log,
 564 };
 565
 566 int kvm_init(int smp_cpus)
 567 {
 568     static const char upgrade_note[] =
 569         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 570         "(see http://sourceforge.net/projects/kvm).\n";
 571     KVMState *s;
 572     int ret;
 573     int i;
 574
 575     if (smp_cpus > 1) {
 576         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 577         return -EINVAL;
 578     }
 579
 580     s = qemu_mallocz(sizeof(KVMState));
 581
 582 #ifdef KVM_CAP_SET_GUEST_DEBUG
 583     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 584 #endif
 585     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 586         s->slots[i].slot = i;
 587
 588     s->vmfd = -1;
 589     s->fd = qemu_open("/dev/kvm", O_RDWR);
 590     if (s->fd == -1) {
 591         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 592         ret = -errno;
 593         goto err;
 594     }
 595
 596     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 597     if (ret < KVM_API_VERSION) {
 598         if (ret > 0)
 599             ret = -EINVAL;
 600         fprintf(stderr, "kvm version too old\n");
 601         goto err;
 602     }
 603
 604     if (ret > KVM_API_VERSION) {
 605         ret = -EINVAL;
 606         fprintf(stderr, "kvm version not supported\n");
 607         goto err;
 608     }
 609
 610     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 611     if (s->vmfd < 0) {
 612 #ifdef TARGET_S390X
 613         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 614                         "your host kernel command line\n");
 615 #endif
 616         goto err;
 617     }
 618
 619     /* initially, KVM allocated its own memory and we had to jump through
 620      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 621      * just use a user allocated buffer so we can use regular pages
 622      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 623      */
 624     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 625         ret = -EINVAL;
 626         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 627                 upgrade_note);
 628         goto err;
 629     }
 630
 631     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 632      * destroyed properly.  Since we rely on this capability, refuse to work
 633      * with any kernel without this capability. */
 634     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 635         ret = -EINVAL;
 636
 637         fprintf(stderr,
 638                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 639                 upgrade_note);
 640         goto err;
 641     }
 642
 643     s->coalesced_mmio = 0;
 644 #ifdef KVM_CAP_COALESCED_MMIO
 645     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 646     s->coalesced_mmio_ring = NULL;
 647 #endif
 648
 649     s->broken_set_mem_region = 1;
 650 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 651     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 652     if (ret > 0) {
 653         s->broken_set_mem_region = 0;
 654     }
 655 #endif
 656
 657     s->vcpu_events = 0;
 658 #ifdef KVM_CAP_VCPU_EVENTS
 659     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 660 #endif
 661
 662     s->robust_singlestep = 0;
 663 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 664     s->robust_singlestep =
 665         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 666 #endif
 667
 668     s->debugregs = 0;
 669 #ifdef KVM_CAP_DEBUGREGS
 670     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 671 #endif
 672
 673     ret = kvm_arch_init(s, smp_cpus);
 674     if (ret < 0)
 675         goto err;
 676
 677     kvm_state = s;
 678     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 679
 680     return 0;
 681
 682 err:
 683     if (s) {
 684         if (s->vmfd != -1)
 685             close(s->vmfd);
 686         if (s->fd != -1)
 687             close(s->fd);
 688     }
 689     qemu_free(s);
 690
 691     return ret;
 692 }
 693
 694 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 695                          uint32_t count)
 696 {
 697     int i;
 698     uint8_t *ptr = data;
 699
 700     for (i = 0; i < count; i++) {
 701         if (direction == KVM_EXIT_IO_IN) {
 702             switch (size) {
 703             case 1:
 704                 stb_p(ptr, cpu_inb(port));
 705                 break;
 706             case 2:
 707                 stw_p(ptr, cpu_inw(port));
 708                 break;
 709             case 4:
 710                 stl_p(ptr, cpu_inl(port));
 711                 break;
 712             }
 713         } else {
 714             switch (size) {
 715             case 1:
 716                 cpu_outb(port, ldub_p(ptr));
 717                 break;
 718             case 2:
 719                 cpu_outw(port, lduw_p(ptr));
 720                 break;
 721             case 4:
 722                 cpu_outl(port, ldl_p(ptr));
 723                 break;
 724             }
 725         }
 726
 727         ptr += size;
 728     }
 729
 730     return 1;
 731 }
 732
 733 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 734 static void kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 735 {
 736
 737     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 738         int i;
 739
 740         fprintf(stderr, "KVM internal error. Suberror: %d\n",
 741                 run->internal.suberror);
 742
 743         for (i = 0; i < run->internal.ndata; ++i) {
 744             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 745                     i, (uint64_t)run->internal.data[i]);
 746         }
 747     }
 748     cpu_dump_state(env, stderr, fprintf, 0);
 749     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 750         fprintf(stderr, "emulation failure\n");
 751     }
 752     /* FIXME: Should trigger a qmp message to let management know
 753      * something went wrong.
 754      */
 755     vm_stop(0);
 756 }
 757 #endif
 758
 759 void kvm_flush_coalesced_mmio_buffer(void)
 760 {
 761 #ifdef KVM_CAP_COALESCED_MMIO
 762     KVMState *s = kvm_state;
 763     if (s->coalesced_mmio_ring) {
 764         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 765         while (ring->first != ring->last) {
 766             struct kvm_coalesced_mmio *ent;
 767
 768             ent = &ring->coalesced_mmio[ring->first];
 769
 770             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 771             smp_wmb();
 772             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 773         }
 774     }
 775 #endif
 776 }
 777
 778 void kvm_cpu_synchronize_state(CPUState *env)
 779 {
 780     if (!env->kvm_vcpu_dirty) {
 781         kvm_arch_get_registers(env);
 782         env->kvm_vcpu_dirty = 1;
 783     }
 784 }
 785
 786 void kvm_cpu_synchronize_post_reset(CPUState *env)
 787 {
 788     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 789     env->kvm_vcpu_dirty = 0;
 790 }
 791
 792 void kvm_cpu_synchronize_post_init(CPUState *env)
 793 {
 794     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 795     env->kvm_vcpu_dirty = 0;
 796 }
 797
 798 int kvm_cpu_exec(CPUState *env)
 799 {
 800     struct kvm_run *run = env->kvm_run;
 801     int ret;
 802
 803     DPRINTF("kvm_cpu_exec()\n");
 804
 805     do {
 806 #ifndef CONFIG_IOTHREAD
 807         if (env->exit_request) {
 808             DPRINTF("interrupt exit requested\n");
 809             ret = 0;
 810             break;
 811         }
 812 #endif
 813
 814         if (env->kvm_vcpu_dirty) {
 815             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 816             env->kvm_vcpu_dirty = 0;
 817         }
 818
 819         kvm_arch_pre_run(env, run);
 820         qemu_mutex_unlock_iothread();
 821         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 822         qemu_mutex_lock_iothread();
 823         kvm_arch_post_run(env, run);
 824
 825         if (ret == -EINTR || ret == -EAGAIN) {
 826             cpu_exit(env);
 827             DPRINTF("io window exit\n");
 828             ret = 0;
 829             break;
 830         }
 831
 832         if (ret < 0) {
 833             DPRINTF("kvm run failed %s\n", strerror(-ret));
 834             abort();
 835         }
 836
 837         kvm_flush_coalesced_mmio_buffer();
 838
 839         ret = 0; /* exit loop */
 840         switch (run->exit_reason) {
 841         case KVM_EXIT_IO:
 842             DPRINTF("handle_io\n");
 843             ret = kvm_handle_io(run->io.port,
 844                                 (uint8_t *)run + run->io.data_offset,
 845                                 run->io.direction,
 846                                 run->io.size,
 847                                 run->io.count);
 848             break;
 849         case KVM_EXIT_MMIO:
 850             DPRINTF("handle_mmio\n");
 851             cpu_physical_memory_rw(run->mmio.phys_addr,
 852                                    run->mmio.data,
 853                                    run->mmio.len,
 854                                    run->mmio.is_write);
 855             ret = 1;
 856             break;
 857         case KVM_EXIT_IRQ_WINDOW_OPEN:
 858             DPRINTF("irq_window_open\n");
 859             break;
 860         case KVM_EXIT_SHUTDOWN:
 861             DPRINTF("shutdown\n");
 862             qemu_system_reset_request();
 863             ret = 1;
 864             break;
 865         case KVM_EXIT_UNKNOWN:
 866             DPRINTF("kvm_exit_unknown\n");
 867             break;
 868         case KVM_EXIT_FAIL_ENTRY:
 869             DPRINTF("kvm_exit_fail_entry\n");
 870             break;
 871         case KVM_EXIT_EXCEPTION:
 872             DPRINTF("kvm_exit_exception\n");
 873             break;
 874 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 875         case KVM_EXIT_INTERNAL_ERROR:
 876             kvm_handle_internal_error(env, run);
 877             break;
 878 #endif
 879         case KVM_EXIT_DEBUG:
 880             DPRINTF("kvm_exit_debug\n");
 881 #ifdef KVM_CAP_SET_GUEST_DEBUG
 882             if (kvm_arch_debug(&run->debug.arch)) {
 883                 gdb_set_stop_cpu(env);
 884                 vm_stop(EXCP_DEBUG);
 885                 env->exception_index = EXCP_DEBUG;
 886                 return 0;
 887             }
 888             /* re-enter, this exception was guest-internal */
 889             ret = 1;
 890 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 891             break;
 892         default:
 893             DPRINTF("kvm_arch_handle_exit\n");
 894             ret = kvm_arch_handle_exit(env, run);
 895             break;
 896         }
 897     } while (ret > 0);
 898
 899     if (env->exit_request) {
 900         env->exit_request = 0;
 901         env->exception_index = EXCP_INTERRUPT;
 902     }
 903
 904     return ret;
 905 }
 906
 907 int kvm_ioctl(KVMState *s, int type, ...)
 908 {
 909     int ret;
 910     void *arg;
 911     va_list ap;
 912
 913     va_start(ap, type);
 914     arg = va_arg(ap, void *);
 915     va_end(ap);
 916
 917     ret = ioctl(s->fd, type, arg);
 918     if (ret == -1)
 919         ret = -errno;
 920
 921     return ret;
 922 }
 923
 924 int kvm_vm_ioctl(KVMState *s, int type, ...)
 925 {
 926     int ret;
 927     void *arg;
 928     va_list ap;
 929
 930     va_start(ap, type);
 931     arg = va_arg(ap, void *);
 932     va_end(ap);
 933
 934     ret = ioctl(s->vmfd, type, arg);
 935     if (ret == -1)
 936         ret = -errno;
 937
 938     return ret;
 939 }
 940
 941 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 942 {
 943     int ret;
 944     void *arg;
 945     va_list ap;
 946
 947     va_start(ap, type);
 948     arg = va_arg(ap, void *);
 949     va_end(ap);
 950
 951     ret = ioctl(env->kvm_fd, type, arg);
 952     if (ret == -1)
 953         ret = -errno;
 954
 955     return ret;
 956 }
 957
 958 int kvm_has_sync_mmu(void)
 959 {
 960 #ifdef KVM_CAP_SYNC_MMU
 961     KVMState *s = kvm_state;
 962
 963     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 964 #else
 965     return 0;
 966 #endif
 967 }
 968
 969 int kvm_has_vcpu_events(void)
 970 {
 971     return kvm_state->vcpu_events;
 972 }
 973
 974 int kvm_has_robust_singlestep(void)
 975 {
 976     return kvm_state->robust_singlestep;
 977 }
 978
 979 int kvm_has_debugregs(void)
 980 {
 981     return kvm_state->debugregs;
 982 }
 983
 984 void kvm_setup_guest_memory(void *start, size_t size)
 985 {
 986     if (!kvm_has_sync_mmu()) {
 987 #ifdef MADV_DONTFORK
 988         int ret = madvise(start, size, MADV_DONTFORK);
 989
 990         if (ret) {
 991             perror("madvice");
 992             exit(1);
 993         }
 994 #else
 995         fprintf(stderr,
 996                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 997         exit(1);
 998 #endif
 999     }
1000 }
1001
1002 #ifdef KVM_CAP_SET_GUEST_DEBUG
1003 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
1004 {
1005 #ifdef CONFIG_IOTHREAD
1006     if (env != cpu_single_env) {
1007         abort();
1008     }
1009 #endif
1010     func(data);
1011 }
1012
1013 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1014                                                  target_ulong pc)
1015 {
1016     struct kvm_sw_breakpoint *bp;
1017
1018     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1019         if (bp->pc == pc)
1020             return bp;
1021     }
1022     return NULL;
1023 }
1024
1025 int kvm_sw_breakpoints_active(CPUState *env)
1026 {
1027     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1028 }
1029
1030 struct kvm_set_guest_debug_data {
1031     struct kvm_guest_debug dbg;
1032     CPUState *env;
1033     int err;
1034 };
1035
1036 static void kvm_invoke_set_guest_debug(void *data)
1037 {
1038     struct kvm_set_guest_debug_data *dbg_data = data;
1039     CPUState *env = dbg_data->env;
1040
1041     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1042 }
1043
1044 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1045 {
1046     struct kvm_set_guest_debug_data data;
1047
1048     data.dbg.control = reinject_trap;
1049
1050     if (env->singlestep_enabled) {
1051         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1052     }
1053     kvm_arch_update_guest_debug(env, &data.dbg);
1054     data.env = env;
1055
1056     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
1057     return data.err;
1058 }
1059
1060 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1061                           target_ulong len, int type)
1062 {
1063     struct kvm_sw_breakpoint *bp;
1064     CPUState *env;
1065     int err;
1066
1067     if (type == GDB_BREAKPOINT_SW) {
1068         bp = kvm_find_sw_breakpoint(current_env, addr);
1069         if (bp) {
1070             bp->use_count++;
1071             return 0;
1072         }
1073
1074         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1075         if (!bp)
1076             return -ENOMEM;
1077
1078         bp->pc = addr;
1079         bp->use_count = 1;
1080         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1081         if (err) {
1082             free(bp);
1083             return err;
1084         }
1085
1086         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1087                           bp, entry);
1088     } else {
1089         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1090         if (err)
1091             return err;
1092     }
1093
1094     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1095         err = kvm_update_guest_debug(env, 0);
1096         if (err)
1097             return err;
1098     }
1099     return 0;
1100 }
1101
1102 int kvm_remove_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             return -ENOENT;
1113
1114         if (bp->use_count > 1) {
1115             bp->use_count--;
1116             return 0;
1117         }
1118
1119         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1120         if (err)
1121             return err;
1122
1123         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1124         qemu_free(bp);
1125     } else {
1126         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1127         if (err)
1128             return err;
1129     }
1130
1131     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1132         err = kvm_update_guest_debug(env, 0);
1133         if (err)
1134             return err;
1135     }
1136     return 0;
1137 }
1138
1139 void kvm_remove_all_breakpoints(CPUState *current_env)
1140 {
1141     struct kvm_sw_breakpoint *bp, *next;
1142     KVMState *s = current_env->kvm_state;
1143     CPUState *env;
1144
1145     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1146         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1147             /* Try harder to find a CPU that currently sees the breakpoint. */
1148             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1149                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1150                     break;
1151             }
1152         }
1153     }
1154     kvm_arch_remove_all_hw_breakpoints();
1155
1156     for (env = first_cpu; env != NULL; env = env->next_cpu)
1157         kvm_update_guest_debug(env, 0);
1158 }
1159
1160 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1161
1162 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1163 {
1164     return -EINVAL;
1165 }
1166
1167 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1168                           target_ulong len, int type)
1169 {
1170     return -EINVAL;
1171 }
1172
1173 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1174                           target_ulong len, int type)
1175 {
1176     return -EINVAL;
1177 }
1178
1179 void kvm_remove_all_breakpoints(CPUState *current_env)
1180 {
1181 }
1182 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1183
1184 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1185 {
1186     struct kvm_signal_mask *sigmask;
1187     int r;
1188
1189     if (!sigset)
1190         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1191
1192     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1193
1194     sigmask->len = 8;
1195     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1196     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1197     free(sigmask);
1198
1199     return r;
1200 }
1201
1202 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1203 {
1204 #ifdef KVM_IOEVENTFD
1205     struct kvm_ioeventfd kick = {
1206         .datamatch = val,
1207         .addr = addr,
1208         .len = 2,
1209         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1210         .fd = fd,
1211     };
1212     int r;
1213     if (!kvm_enabled())
1214         return -ENOSYS;
1215     if (!assign)
1216         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1217     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1218     if (r < 0)
1219         return r;
1220     return 0;
1221 #else
1222     return -ENOSYS;
1223 #endif
1224 }