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 void kvm_flush_coalesced_mmio_buffer(void)
 734 {
 735 #ifdef KVM_CAP_COALESCED_MMIO
 736     KVMState *s = kvm_state;
 737     if (s->coalesced_mmio_ring) {
 738         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 739         while (ring->first != ring->last) {
 740             struct kvm_coalesced_mmio *ent;
 741
 742             ent = &ring->coalesced_mmio[ring->first];
 743
 744             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 745             smp_wmb();
 746             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 747         }
 748     }
 749 #endif
 750 }
 751
 752 void kvm_cpu_synchronize_state(CPUState *env)
 753 {
 754     if (!env->kvm_vcpu_dirty) {
 755         kvm_arch_get_registers(env);
 756         env->kvm_vcpu_dirty = 1;
 757     }
 758 }
 759
 760 void kvm_cpu_synchronize_post_reset(CPUState *env)
 761 {
 762     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 763     env->kvm_vcpu_dirty = 0;
 764 }
 765
 766 void kvm_cpu_synchronize_post_init(CPUState *env)
 767 {
 768     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 769     env->kvm_vcpu_dirty = 0;
 770 }
 771
 772 int kvm_cpu_exec(CPUState *env)
 773 {
 774     struct kvm_run *run = env->kvm_run;
 775     int ret;
 776
 777     DPRINTF("kvm_cpu_exec()\n");
 778
 779     do {
 780 #ifndef CONFIG_IOTHREAD
 781         if (env->exit_request) {
 782             DPRINTF("interrupt exit requested\n");
 783             ret = 0;
 784             break;
 785         }
 786 #endif
 787
 788         if (env->kvm_vcpu_dirty) {
 789             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 790             env->kvm_vcpu_dirty = 0;
 791         }
 792
 793         kvm_arch_pre_run(env, run);
 794         qemu_mutex_unlock_iothread();
 795         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 796         qemu_mutex_lock_iothread();
 797         kvm_arch_post_run(env, run);
 798
 799         if (ret == -EINTR || ret == -EAGAIN) {
 800             cpu_exit(env);
 801             DPRINTF("io window exit\n");
 802             ret = 0;
 803             break;
 804         }
 805
 806         if (ret < 0) {
 807             DPRINTF("kvm run failed %s\n", strerror(-ret));
 808             abort();
 809         }
 810
 811         kvm_flush_coalesced_mmio_buffer();
 812
 813         ret = 0; /* exit loop */
 814         switch (run->exit_reason) {
 815         case KVM_EXIT_IO:
 816             DPRINTF("handle_io\n");
 817             ret = kvm_handle_io(run->io.port,
 818                                 (uint8_t *)run + run->io.data_offset,
 819                                 run->io.direction,
 820                                 run->io.size,
 821                                 run->io.count);
 822             break;
 823         case KVM_EXIT_MMIO:
 824             DPRINTF("handle_mmio\n");
 825             cpu_physical_memory_rw(run->mmio.phys_addr,
 826                                    run->mmio.data,
 827                                    run->mmio.len,
 828                                    run->mmio.is_write);
 829             ret = 1;
 830             break;
 831         case KVM_EXIT_IRQ_WINDOW_OPEN:
 832             DPRINTF("irq_window_open\n");
 833             break;
 834         case KVM_EXIT_SHUTDOWN:
 835             DPRINTF("shutdown\n");
 836             qemu_system_reset_request();
 837             ret = 1;
 838             break;
 839         case KVM_EXIT_UNKNOWN:
 840             DPRINTF("kvm_exit_unknown\n");
 841             break;
 842         case KVM_EXIT_FAIL_ENTRY:
 843             DPRINTF("kvm_exit_fail_entry\n");
 844             break;
 845         case KVM_EXIT_EXCEPTION:
 846             DPRINTF("kvm_exit_exception\n");
 847             break;
 848         case KVM_EXIT_DEBUG:
 849             DPRINTF("kvm_exit_debug\n");
 850 #ifdef KVM_CAP_SET_GUEST_DEBUG
 851             if (kvm_arch_debug(&run->debug.arch)) {
 852                 gdb_set_stop_cpu(env);
 853                 vm_stop(EXCP_DEBUG);
 854                 env->exception_index = EXCP_DEBUG;
 855                 return 0;
 856             }
 857             /* re-enter, this exception was guest-internal */
 858             ret = 1;
 859 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 860             break;
 861         default:
 862             DPRINTF("kvm_arch_handle_exit\n");
 863             ret = kvm_arch_handle_exit(env, run);
 864             break;
 865         }
 866     } while (ret > 0);
 867
 868     if (env->exit_request) {
 869         env->exit_request = 0;
 870         env->exception_index = EXCP_INTERRUPT;
 871     }
 872
 873     return ret;
 874 }
 875
 876 int kvm_ioctl(KVMState *s, int type, ...)
 877 {
 878     int ret;
 879     void *arg;
 880     va_list ap;
 881
 882     va_start(ap, type);
 883     arg = va_arg(ap, void *);
 884     va_end(ap);
 885
 886     ret = ioctl(s->fd, type, arg);
 887     if (ret == -1)
 888         ret = -errno;
 889
 890     return ret;
 891 }
 892
 893 int kvm_vm_ioctl(KVMState *s, int type, ...)
 894 {
 895     int ret;
 896     void *arg;
 897     va_list ap;
 898
 899     va_start(ap, type);
 900     arg = va_arg(ap, void *);
 901     va_end(ap);
 902
 903     ret = ioctl(s->vmfd, type, arg);
 904     if (ret == -1)
 905         ret = -errno;
 906
 907     return ret;
 908 }
 909
 910 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 911 {
 912     int ret;
 913     void *arg;
 914     va_list ap;
 915
 916     va_start(ap, type);
 917     arg = va_arg(ap, void *);
 918     va_end(ap);
 919
 920     ret = ioctl(env->kvm_fd, type, arg);
 921     if (ret == -1)
 922         ret = -errno;
 923
 924     return ret;
 925 }
 926
 927 int kvm_has_sync_mmu(void)
 928 {
 929 #ifdef KVM_CAP_SYNC_MMU
 930     KVMState *s = kvm_state;
 931
 932     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 933 #else
 934     return 0;
 935 #endif
 936 }
 937
 938 int kvm_has_vcpu_events(void)
 939 {
 940     return kvm_state->vcpu_events;
 941 }
 942
 943 int kvm_has_robust_singlestep(void)
 944 {
 945     return kvm_state->robust_singlestep;
 946 }
 947
 948 int kvm_has_debugregs(void)
 949 {
 950     return kvm_state->debugregs;
 951 }
 952
 953 void kvm_setup_guest_memory(void *start, size_t size)
 954 {
 955     if (!kvm_has_sync_mmu()) {
 956 #ifdef MADV_DONTFORK
 957         int ret = madvise(start, size, MADV_DONTFORK);
 958
 959         if (ret) {
 960             perror("madvice");
 961             exit(1);
 962         }
 963 #else
 964         fprintf(stderr,
 965                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 966         exit(1);
 967 #endif
 968     }
 969 }
 970
 971 #ifdef KVM_CAP_SET_GUEST_DEBUG
 972 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 973 {
 974 #ifdef CONFIG_IOTHREAD
 975     if (env != cpu_single_env) {
 976         abort();
 977     }
 978 #endif
 979     func(data);
 980 }
 981
 982 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 983                                                  target_ulong pc)
 984 {
 985     struct kvm_sw_breakpoint *bp;
 986
 987     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 988         if (bp->pc == pc)
 989             return bp;
 990     }
 991     return NULL;
 992 }
 993
 994 int kvm_sw_breakpoints_active(CPUState *env)
 995 {
 996     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 997 }
 998
 999 struct kvm_set_guest_debug_data {
1000     struct kvm_guest_debug dbg;
1001     CPUState *env;
1002     int err;
1003 };
1004
1005 static void kvm_invoke_set_guest_debug(void *data)
1006 {
1007     struct kvm_set_guest_debug_data *dbg_data = data;
1008     CPUState *env = dbg_data->env;
1009
1010     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1011 }
1012
1013 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1014 {
1015     struct kvm_set_guest_debug_data data;
1016
1017     data.dbg.control = reinject_trap;
1018
1019     if (env->singlestep_enabled) {
1020         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1021     }
1022     kvm_arch_update_guest_debug(env, &data.dbg);
1023     data.env = env;
1024
1025     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
1026     return data.err;
1027 }
1028
1029 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1030                           target_ulong len, int type)
1031 {
1032     struct kvm_sw_breakpoint *bp;
1033     CPUState *env;
1034     int err;
1035
1036     if (type == GDB_BREAKPOINT_SW) {
1037         bp = kvm_find_sw_breakpoint(current_env, addr);
1038         if (bp) {
1039             bp->use_count++;
1040             return 0;
1041         }
1042
1043         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1044         if (!bp)
1045             return -ENOMEM;
1046
1047         bp->pc = addr;
1048         bp->use_count = 1;
1049         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1050         if (err) {
1051             free(bp);
1052             return err;
1053         }
1054
1055         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1056                           bp, entry);
1057     } else {
1058         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1059         if (err)
1060             return err;
1061     }
1062
1063     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1064         err = kvm_update_guest_debug(env, 0);
1065         if (err)
1066             return err;
1067     }
1068     return 0;
1069 }
1070
1071 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1072                           target_ulong len, int type)
1073 {
1074     struct kvm_sw_breakpoint *bp;
1075     CPUState *env;
1076     int err;
1077
1078     if (type == GDB_BREAKPOINT_SW) {
1079         bp = kvm_find_sw_breakpoint(current_env, addr);
1080         if (!bp)
1081             return -ENOENT;
1082
1083         if (bp->use_count > 1) {
1084             bp->use_count--;
1085             return 0;
1086         }
1087
1088         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1089         if (err)
1090             return err;
1091
1092         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1093         qemu_free(bp);
1094     } else {
1095         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1096         if (err)
1097             return err;
1098     }
1099
1100     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1101         err = kvm_update_guest_debug(env, 0);
1102         if (err)
1103             return err;
1104     }
1105     return 0;
1106 }
1107
1108 void kvm_remove_all_breakpoints(CPUState *current_env)
1109 {
1110     struct kvm_sw_breakpoint *bp, *next;
1111     KVMState *s = current_env->kvm_state;
1112     CPUState *env;
1113
1114     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1115         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1116             /* Try harder to find a CPU that currently sees the breakpoint. */
1117             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1118                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1119                     break;
1120             }
1121         }
1122     }
1123     kvm_arch_remove_all_hw_breakpoints();
1124
1125     for (env = first_cpu; env != NULL; env = env->next_cpu)
1126         kvm_update_guest_debug(env, 0);
1127 }
1128
1129 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1130
1131 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1132 {
1133     return -EINVAL;
1134 }
1135
1136 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1137                           target_ulong len, int type)
1138 {
1139     return -EINVAL;
1140 }
1141
1142 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1143                           target_ulong len, int type)
1144 {
1145     return -EINVAL;
1146 }
1147
1148 void kvm_remove_all_breakpoints(CPUState *current_env)
1149 {
1150 }
1151 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1152
1153 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1154 {
1155     struct kvm_signal_mask *sigmask;
1156     int r;
1157
1158     if (!sigset)
1159         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1160
1161     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1162
1163     sigmask->len = 8;
1164     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1165     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1166     free(sigmask);
1167
1168     return r;
1169 }
1170
1171 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1172 {
1173 #ifdef KVM_IOEVENTFD
1174     struct kvm_ioeventfd kick = {
1175         .datamatch = val,
1176         .addr = addr,
1177         .len = 2,
1178         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1179         .fd = fd,
1180     };
1181     int r;
1182     if (!kvm_enabled())
1183         return -ENOSYS;
1184     if (!assign)
1185         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1186     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1187     if (r < 0)
1188         return r;
1189     return 0;
1190 #else
1191     return -ENOSYS;
1192 #endif
1193 }