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 #ifdef KVM_CAP_SET_GUEST_DEBUG
  68     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  69 #endif
  70     int irqchip_in_kernel;
  71     int pit_in_kernel;
  72 };
  73
  74 static KVMState *kvm_state;
  75
  76 static KVMSlot *kvm_alloc_slot(KVMState *s)
  77 {
  78     int i;
  79
  80     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  81         /* KVM private memory slots */
  82         if (i >= 8 && i < 12)
  83             continue;
  84         if (s->slots[i].memory_size == 0)
  85             return &s->slots[i];
  86     }
  87
  88     fprintf(stderr, "%s: no free slot available\n", __func__);
  89     abort();
  90 }
  91
  92 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  93                                          target_phys_addr_t start_addr,
  94                                          target_phys_addr_t end_addr)
  95 {
  96     int i;
  97
  98     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  99         KVMSlot *mem = &s->slots[i];
 100
 101         if (start_addr == mem->start_addr &&
 102             end_addr == mem->start_addr + mem->memory_size) {
 103             return mem;
 104         }
 105     }
 106
 107     return NULL;
 108 }
 109
 110 /*
 111  * Find overlapping slot with lowest start address
 112  */
 113 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 114                                             target_phys_addr_t start_addr,
 115                                             target_phys_addr_t end_addr)
 116 {
 117     KVMSlot *found = NULL;
 118     int i;
 119
 120     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 121         KVMSlot *mem = &s->slots[i];
 122
 123         if (mem->memory_size == 0 ||
 124             (found && found->start_addr < mem->start_addr)) {
 125             continue;
 126         }
 127
 128         if (end_addr > mem->start_addr &&
 129             start_addr < mem->start_addr + mem->memory_size) {
 130             found = mem;
 131         }
 132     }
 133
 134     return found;
 135 }
 136
 137 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 138 {
 139     struct kvm_userspace_memory_region mem;
 140
 141     mem.slot = slot->slot;
 142     mem.guest_phys_addr = slot->start_addr;
 143     mem.memory_size = slot->memory_size;
 144     mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
 145     mem.flags = slot->flags;
 146     if (s->migration_log) {
 147         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 148     }
 149     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 150 }
 151
 152 static void kvm_reset_vcpu(void *opaque)
 153 {
 154     CPUState *env = opaque;
 155
 156     kvm_arch_reset_vcpu(env);
 157 }
 158
 159 int kvm_irqchip_in_kernel(void)
 160 {
 161     return kvm_state->irqchip_in_kernel;
 162 }
 163
 164 int kvm_pit_in_kernel(void)
 165 {
 166     return kvm_state->pit_in_kernel;
 167 }
 168
 169
 170 int kvm_init_vcpu(CPUState *env)
 171 {
 172     KVMState *s = kvm_state;
 173     long mmap_size;
 174     int ret;
 175
 176     dprintf("kvm_init_vcpu\n");
 177
 178     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 179     if (ret < 0) {
 180         dprintf("kvm_create_vcpu failed\n");
 181         goto err;
 182     }
 183
 184     env->kvm_fd = ret;
 185     env->kvm_state = s;
 186
 187     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 188     if (mmap_size < 0) {
 189         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 190         goto err;
 191     }
 192
 193     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 194                         env->kvm_fd, 0);
 195     if (env->kvm_run == MAP_FAILED) {
 196         ret = -errno;
 197         dprintf("mmap'ing vcpu state failed\n");
 198         goto err;
 199     }
 200
 201 #ifdef KVM_CAP_COALESCED_MMIO
 202     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 203         s->coalesced_mmio_ring = (void *) env->kvm_run +
 204                 s->coalesced_mmio * PAGE_SIZE;
 205 #endif
 206
 207     ret = kvm_arch_init_vcpu(env);
 208     if (ret == 0) {
 209         qemu_register_reset(kvm_reset_vcpu, env);
 210         kvm_arch_reset_vcpu(env);
 211     }
 212 err:
 213     return ret;
 214 }
 215
 216 /*
 217  * dirty pages logging control
 218  */
 219 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 220                                       ram_addr_t size, int flags, int mask)
 221 {
 222     KVMState *s = kvm_state;
 223     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 224     int old_flags;
 225
 226     if (mem == NULL)  {
 227             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 228                     TARGET_FMT_plx "\n", __func__, phys_addr,
 229                     (target_phys_addr_t)(phys_addr + size - 1));
 230             return -EINVAL;
 231     }
 232
 233     old_flags = mem->flags;
 234
 235     flags = (mem->flags & ~mask) | flags;
 236     mem->flags = flags;
 237
 238     /* If nothing changed effectively, no need to issue ioctl */
 239     if (s->migration_log) {
 240         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 241     }
 242     if (flags == old_flags) {
 243             return 0;
 244     }
 245
 246     return kvm_set_user_memory_region(s, mem);
 247 }
 248
 249 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 250 {
 251         return kvm_dirty_pages_log_change(phys_addr, size,
 252                                           KVM_MEM_LOG_DIRTY_PAGES,
 253                                           KVM_MEM_LOG_DIRTY_PAGES);
 254 }
 255
 256 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 257 {
 258         return kvm_dirty_pages_log_change(phys_addr, size,
 259                                           0,
 260                                           KVM_MEM_LOG_DIRTY_PAGES);
 261 }
 262
 263 static int kvm_set_migration_log(int enable)
 264 {
 265     KVMState *s = kvm_state;
 266     KVMSlot *mem;
 267     int i, err;
 268
 269     s->migration_log = enable;
 270
 271     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 272         mem = &s->slots[i];
 273
 274         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 275             continue;
 276         }
 277         err = kvm_set_user_memory_region(s, mem);
 278         if (err) {
 279             return err;
 280         }
 281     }
 282     return 0;
 283 }
 284
 285 static int test_le_bit(unsigned long nr, unsigned char *addr)
 286 {
 287     return (addr[nr >> 3] >> (nr & 7)) & 1;
 288 }
 289
 290 /**
 291  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 292  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 293  * This means all bits are set to dirty.
 294  *
 295  * @start_add: start of logged region.
 296  * @end_addr: end of logged region.
 297  */
 298 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 299                                           target_phys_addr_t end_addr)
 300 {
 301     KVMState *s = kvm_state;
 302     unsigned long size, allocated_size = 0;
 303     target_phys_addr_t phys_addr;
 304     ram_addr_t addr;
 305     KVMDirtyLog d;
 306     KVMSlot *mem;
 307     int ret = 0;
 308
 309     d.dirty_bitmap = NULL;
 310     while (start_addr < end_addr) {
 311         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 312         if (mem == NULL) {
 313             break;
 314         }
 315
 316         size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
 317         if (!d.dirty_bitmap) {
 318             d.dirty_bitmap = qemu_malloc(size);
 319         } else if (size > allocated_size) {
 320             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 321         }
 322         allocated_size = size;
 323         memset(d.dirty_bitmap, 0, allocated_size);
 324
 325         d.slot = mem->slot;
 326
 327         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 328             dprintf("ioctl failed %d\n", errno);
 329             ret = -1;
 330             break;
 331         }
 332
 333         for (phys_addr = mem->start_addr, addr = mem->phys_offset;
 334              phys_addr < mem->start_addr + mem->memory_size;
 335              phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 336             unsigned char *bitmap = (unsigned char *)d.dirty_bitmap;
 337             unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
 338
 339             if (test_le_bit(nr, bitmap)) {
 340                 cpu_physical_memory_set_dirty(addr);
 341             }
 342         }
 343         start_addr = phys_addr;
 344     }
 345     qemu_free(d.dirty_bitmap);
 346
 347     return ret;
 348 }
 349
 350 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 351 {
 352     int ret = -ENOSYS;
 353 #ifdef KVM_CAP_COALESCED_MMIO
 354     KVMState *s = kvm_state;
 355
 356     if (s->coalesced_mmio) {
 357         struct kvm_coalesced_mmio_zone zone;
 358
 359         zone.addr = start;
 360         zone.size = size;
 361
 362         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 363     }
 364 #endif
 365
 366     return ret;
 367 }
 368
 369 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 370 {
 371     int ret = -ENOSYS;
 372 #ifdef KVM_CAP_COALESCED_MMIO
 373     KVMState *s = kvm_state;
 374
 375     if (s->coalesced_mmio) {
 376         struct kvm_coalesced_mmio_zone zone;
 377
 378         zone.addr = start;
 379         zone.size = size;
 380
 381         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 382     }
 383 #endif
 384
 385     return ret;
 386 }
 387
 388 int kvm_check_extension(KVMState *s, unsigned int extension)
 389 {
 390     int ret;
 391
 392     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 393     if (ret < 0) {
 394         ret = 0;
 395     }
 396
 397     return ret;
 398 }
 399
 400 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 401                              ram_addr_t size,
 402                              ram_addr_t phys_offset)
 403 {
 404     KVMState *s = kvm_state;
 405     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 406     KVMSlot *mem, old;
 407     int err;
 408
 409     if (start_addr & ~TARGET_PAGE_MASK) {
 410         if (flags >= IO_MEM_UNASSIGNED) {
 411             if (!kvm_lookup_overlapping_slot(s, start_addr,
 412                                              start_addr + size)) {
 413                 return;
 414             }
 415             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 416         } else {
 417             fprintf(stderr, "Only page-aligned memory slots supported\n");
 418         }
 419         abort();
 420     }
 421
 422     /* KVM does not support read-only slots */
 423     phys_offset &= ~IO_MEM_ROM;
 424
 425     while (1) {
 426         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 427         if (!mem) {
 428             break;
 429         }
 430
 431         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 432             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 433             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 434             /* The new slot fits into the existing one and comes with
 435              * identical parameters - nothing to be done. */
 436             return;
 437         }
 438
 439         old = *mem;
 440
 441         /* unregister the overlapping slot */
 442         mem->memory_size = 0;
 443         err = kvm_set_user_memory_region(s, mem);
 444         if (err) {
 445             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 446                     __func__, strerror(-err));
 447             abort();
 448         }
 449
 450         /* Workaround for older KVM versions: we can't join slots, even not by
 451          * unregistering the previous ones and then registering the larger
 452          * slot. We have to maintain the existing fragmentation. Sigh.
 453          *
 454          * This workaround assumes that the new slot starts at the same
 455          * address as the first existing one. If not or if some overlapping
 456          * slot comes around later, we will fail (not seen in practice so far)
 457          * - and actually require a recent KVM version. */
 458         if (s->broken_set_mem_region &&
 459             old.start_addr == start_addr && old.memory_size < size &&
 460             flags < IO_MEM_UNASSIGNED) {
 461             mem = kvm_alloc_slot(s);
 462             mem->memory_size = old.memory_size;
 463             mem->start_addr = old.start_addr;
 464             mem->phys_offset = old.phys_offset;
 465             mem->flags = 0;
 466
 467             err = kvm_set_user_memory_region(s, mem);
 468             if (err) {
 469                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 470                         strerror(-err));
 471                 abort();
 472             }
 473
 474             start_addr += old.memory_size;
 475             phys_offset += old.memory_size;
 476             size -= old.memory_size;
 477             continue;
 478         }
 479
 480         /* register prefix slot */
 481         if (old.start_addr < start_addr) {
 482             mem = kvm_alloc_slot(s);
 483             mem->memory_size = start_addr - old.start_addr;
 484             mem->start_addr = old.start_addr;
 485             mem->phys_offset = old.phys_offset;
 486             mem->flags = 0;
 487
 488             err = kvm_set_user_memory_region(s, mem);
 489             if (err) {
 490                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 491                         __func__, strerror(-err));
 492                 abort();
 493             }
 494         }
 495
 496         /* register suffix slot */
 497         if (old.start_addr + old.memory_size > start_addr + size) {
 498             ram_addr_t size_delta;
 499
 500             mem = kvm_alloc_slot(s);
 501             mem->start_addr = start_addr + size;
 502             size_delta = mem->start_addr - old.start_addr;
 503             mem->memory_size = old.memory_size - size_delta;
 504             mem->phys_offset = old.phys_offset + size_delta;
 505             mem->flags = 0;
 506
 507             err = kvm_set_user_memory_region(s, mem);
 508             if (err) {
 509                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 510                         __func__, strerror(-err));
 511                 abort();
 512             }
 513         }
 514     }
 515
 516     /* in case the KVM bug workaround already "consumed" the new slot */
 517     if (!size)
 518         return;
 519
 520     /* KVM does not need to know about this memory */
 521     if (flags >= IO_MEM_UNASSIGNED)
 522         return;
 523
 524     mem = kvm_alloc_slot(s);
 525     mem->memory_size = size;
 526     mem->start_addr = start_addr;
 527     mem->phys_offset = phys_offset;
 528     mem->flags = 0;
 529
 530     err = kvm_set_user_memory_region(s, mem);
 531     if (err) {
 532         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 533                 strerror(-err));
 534         abort();
 535     }
 536 }
 537
 538 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 539                                   target_phys_addr_t start_addr,
 540                                   ram_addr_t size,
 541                                   ram_addr_t phys_offset)
 542 {
 543         kvm_set_phys_mem(start_addr, size, phys_offset);
 544 }
 545
 546 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 547                                         target_phys_addr_t start_addr,
 548                                         target_phys_addr_t end_addr)
 549 {
 550         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 551 }
 552
 553 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 554                                     int enable)
 555 {
 556         return kvm_set_migration_log(enable);
 557 }
 558
 559 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 560         .set_memory = kvm_client_set_memory,
 561         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 562         .migration_log = kvm_client_migration_log,
 563 };
 564
 565 int kvm_init(int smp_cpus)
 566 {
 567     static const char upgrade_note[] =
 568         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 569         "(see http://sourceforge.net/projects/kvm).\n";
 570     KVMState *s;
 571     int ret;
 572     int i;
 573
 574     if (smp_cpus > 1) {
 575         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 576         return -EINVAL;
 577     }
 578
 579     s = qemu_mallocz(sizeof(KVMState));
 580
 581 #ifdef KVM_CAP_SET_GUEST_DEBUG
 582     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 583 #endif
 584     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 585         s->slots[i].slot = i;
 586
 587     s->vmfd = -1;
 588     s->fd = qemu_open("/dev/kvm", O_RDWR);
 589     if (s->fd == -1) {
 590         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 591         ret = -errno;
 592         goto err;
 593     }
 594
 595     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 596     if (ret < KVM_API_VERSION) {
 597         if (ret > 0)
 598             ret = -EINVAL;
 599         fprintf(stderr, "kvm version too old\n");
 600         goto err;
 601     }
 602
 603     if (ret > KVM_API_VERSION) {
 604         ret = -EINVAL;
 605         fprintf(stderr, "kvm version not supported\n");
 606         goto err;
 607     }
 608
 609     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 610     if (s->vmfd < 0) {
 611 #ifdef TARGET_S390X
 612         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 613                         "your host kernel command line\n");
 614 #endif
 615         goto err;
 616     }
 617
 618     /* initially, KVM allocated its own memory and we had to jump through
 619      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 620      * just use a user allocated buffer so we can use regular pages
 621      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 622      */
 623     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 624         ret = -EINVAL;
 625         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 626                 upgrade_note);
 627         goto err;
 628     }
 629
 630     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 631      * destroyed properly.  Since we rely on this capability, refuse to work
 632      * with any kernel without this capability. */
 633     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 634         ret = -EINVAL;
 635
 636         fprintf(stderr,
 637                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 638                 upgrade_note);
 639         goto err;
 640     }
 641
 642     s->coalesced_mmio = 0;
 643 #ifdef KVM_CAP_COALESCED_MMIO
 644     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 645     s->coalesced_mmio_ring = NULL;
 646 #endif
 647
 648     s->broken_set_mem_region = 1;
 649 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 650     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 651     if (ret > 0) {
 652         s->broken_set_mem_region = 0;
 653     }
 654 #endif
 655
 656     s->vcpu_events = 0;
 657 #ifdef KVM_CAP_VCPU_EVENTS
 658     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 659 #endif
 660
 661     s->robust_singlestep = 0;
 662 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 663     s->robust_singlestep =
 664         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 665 #endif
 666
 667     ret = kvm_arch_init(s, smp_cpus);
 668     if (ret < 0)
 669         goto err;
 670
 671     kvm_state = s;
 672     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 673
 674     return 0;
 675
 676 err:
 677     if (s) {
 678         if (s->vmfd != -1)
 679             close(s->vmfd);
 680         if (s->fd != -1)
 681             close(s->fd);
 682     }
 683     qemu_free(s);
 684
 685     return ret;
 686 }
 687
 688 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 689                          uint32_t count)
 690 {
 691     int i;
 692     uint8_t *ptr = data;
 693
 694     for (i = 0; i < count; i++) {
 695         if (direction == KVM_EXIT_IO_IN) {
 696             switch (size) {
 697             case 1:
 698                 stb_p(ptr, cpu_inb(port));
 699                 break;
 700             case 2:
 701                 stw_p(ptr, cpu_inw(port));
 702                 break;
 703             case 4:
 704                 stl_p(ptr, cpu_inl(port));
 705                 break;
 706             }
 707         } else {
 708             switch (size) {
 709             case 1:
 710                 cpu_outb(port, ldub_p(ptr));
 711                 break;
 712             case 2:
 713                 cpu_outw(port, lduw_p(ptr));
 714                 break;
 715             case 4:
 716                 cpu_outl(port, ldl_p(ptr));
 717                 break;
 718             }
 719         }
 720
 721         ptr += size;
 722     }
 723
 724     return 1;
 725 }
 726
 727 void kvm_flush_coalesced_mmio_buffer(void)
 728 {
 729 #ifdef KVM_CAP_COALESCED_MMIO
 730     KVMState *s = kvm_state;
 731     if (s->coalesced_mmio_ring) {
 732         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 733         while (ring->first != ring->last) {
 734             struct kvm_coalesced_mmio *ent;
 735
 736             ent = &ring->coalesced_mmio[ring->first];
 737
 738             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 739             smp_wmb();
 740             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 741         }
 742     }
 743 #endif
 744 }
 745
 746 void kvm_cpu_synchronize_state(CPUState *env)
 747 {
 748     if (!env->kvm_vcpu_dirty) {
 749         kvm_arch_get_registers(env);
 750         env->kvm_vcpu_dirty = 1;
 751     }
 752 }
 753
 754 void kvm_cpu_synchronize_post_reset(CPUState *env)
 755 {
 756     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 757     env->kvm_vcpu_dirty = 0;
 758 }
 759
 760 void kvm_cpu_synchronize_post_init(CPUState *env)
 761 {
 762     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 763     env->kvm_vcpu_dirty = 0;
 764 }
 765
 766 int kvm_cpu_exec(CPUState *env)
 767 {
 768     struct kvm_run *run = env->kvm_run;
 769     int ret;
 770
 771     dprintf("kvm_cpu_exec()\n");
 772
 773     do {
 774 #ifndef CONFIG_IOTHREAD
 775         if (env->exit_request) {
 776             dprintf("interrupt exit requested\n");
 777             ret = 0;
 778             break;
 779         }
 780 #endif
 781
 782         if (env->kvm_vcpu_dirty) {
 783             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 784             env->kvm_vcpu_dirty = 0;
 785         }
 786
 787         kvm_arch_pre_run(env, run);
 788         qemu_mutex_unlock_iothread();
 789         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 790         qemu_mutex_lock_iothread();
 791         kvm_arch_post_run(env, run);
 792
 793         if (ret == -EINTR || ret == -EAGAIN) {
 794             cpu_exit(env);
 795             dprintf("io window exit\n");
 796             ret = 0;
 797             break;
 798         }
 799
 800         if (ret < 0) {
 801             dprintf("kvm run failed %s\n", strerror(-ret));
 802             abort();
 803         }
 804
 805         kvm_flush_coalesced_mmio_buffer();
 806
 807         ret = 0; /* exit loop */
 808         switch (run->exit_reason) {
 809         case KVM_EXIT_IO:
 810             dprintf("handle_io\n");
 811             ret = kvm_handle_io(run->io.port,
 812                                 (uint8_t *)run + run->io.data_offset,
 813                                 run->io.direction,
 814                                 run->io.size,
 815                                 run->io.count);
 816             break;
 817         case KVM_EXIT_MMIO:
 818             dprintf("handle_mmio\n");
 819             cpu_physical_memory_rw(run->mmio.phys_addr,
 820                                    run->mmio.data,
 821                                    run->mmio.len,
 822                                    run->mmio.is_write);
 823             ret = 1;
 824             break;
 825         case KVM_EXIT_IRQ_WINDOW_OPEN:
 826             dprintf("irq_window_open\n");
 827             break;
 828         case KVM_EXIT_SHUTDOWN:
 829             dprintf("shutdown\n");
 830             qemu_system_reset_request();
 831             ret = 1;
 832             break;
 833         case KVM_EXIT_UNKNOWN:
 834             dprintf("kvm_exit_unknown\n");
 835             break;
 836         case KVM_EXIT_FAIL_ENTRY:
 837             dprintf("kvm_exit_fail_entry\n");
 838             break;
 839         case KVM_EXIT_EXCEPTION:
 840             dprintf("kvm_exit_exception\n");
 841             break;
 842         case KVM_EXIT_DEBUG:
 843             dprintf("kvm_exit_debug\n");
 844 #ifdef KVM_CAP_SET_GUEST_DEBUG
 845             if (kvm_arch_debug(&run->debug.arch)) {
 846                 gdb_set_stop_cpu(env);
 847                 vm_stop(EXCP_DEBUG);
 848                 env->exception_index = EXCP_DEBUG;
 849                 return 0;
 850             }
 851             /* re-enter, this exception was guest-internal */
 852             ret = 1;
 853 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 854             break;
 855         default:
 856             dprintf("kvm_arch_handle_exit\n");
 857             ret = kvm_arch_handle_exit(env, run);
 858             break;
 859         }
 860     } while (ret > 0);
 861
 862     if (env->exit_request) {
 863         env->exit_request = 0;
 864         env->exception_index = EXCP_INTERRUPT;
 865     }
 866
 867     return ret;
 868 }
 869
 870 int kvm_ioctl(KVMState *s, int type, ...)
 871 {
 872     int ret;
 873     void *arg;
 874     va_list ap;
 875
 876     va_start(ap, type);
 877     arg = va_arg(ap, void *);
 878     va_end(ap);
 879
 880     ret = ioctl(s->fd, type, arg);
 881     if (ret == -1)
 882         ret = -errno;
 883
 884     return ret;
 885 }
 886
 887 int kvm_vm_ioctl(KVMState *s, int type, ...)
 888 {
 889     int ret;
 890     void *arg;
 891     va_list ap;
 892
 893     va_start(ap, type);
 894     arg = va_arg(ap, void *);
 895     va_end(ap);
 896
 897     ret = ioctl(s->vmfd, type, arg);
 898     if (ret == -1)
 899         ret = -errno;
 900
 901     return ret;
 902 }
 903
 904 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 905 {
 906     int ret;
 907     void *arg;
 908     va_list ap;
 909
 910     va_start(ap, type);
 911     arg = va_arg(ap, void *);
 912     va_end(ap);
 913
 914     ret = ioctl(env->kvm_fd, type, arg);
 915     if (ret == -1)
 916         ret = -errno;
 917
 918     return ret;
 919 }
 920
 921 int kvm_has_sync_mmu(void)
 922 {
 923 #ifdef KVM_CAP_SYNC_MMU
 924     KVMState *s = kvm_state;
 925
 926     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 927 #else
 928     return 0;
 929 #endif
 930 }
 931
 932 int kvm_has_vcpu_events(void)
 933 {
 934     return kvm_state->vcpu_events;
 935 }
 936
 937 int kvm_has_robust_singlestep(void)
 938 {
 939     return kvm_state->robust_singlestep;
 940 }
 941
 942 void kvm_setup_guest_memory(void *start, size_t size)
 943 {
 944     if (!kvm_has_sync_mmu()) {
 945 #ifdef MADV_DONTFORK
 946         int ret = madvise(start, size, MADV_DONTFORK);
 947
 948         if (ret) {
 949             perror("madvice");
 950             exit(1);
 951         }
 952 #else
 953         fprintf(stderr,
 954                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 955         exit(1);
 956 #endif
 957     }
 958 }
 959
 960 #ifdef KVM_CAP_SET_GUEST_DEBUG
 961 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
 962 {
 963 #ifdef CONFIG_IOTHREAD
 964     if (env != cpu_single_env) {
 965         abort();
 966     }
 967 #endif
 968     func(data);
 969 }
 970
 971 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
 972                                                  target_ulong pc)
 973 {
 974     struct kvm_sw_breakpoint *bp;
 975
 976     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
 977         if (bp->pc == pc)
 978             return bp;
 979     }
 980     return NULL;
 981 }
 982
 983 int kvm_sw_breakpoints_active(CPUState *env)
 984 {
 985     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
 986 }
 987
 988 struct kvm_set_guest_debug_data {
 989     struct kvm_guest_debug dbg;
 990     CPUState *env;
 991     int err;
 992 };
 993
 994 static void kvm_invoke_set_guest_debug(void *data)
 995 {
 996     struct kvm_set_guest_debug_data *dbg_data = data;
 997     CPUState *env = dbg_data->env;
 998
 999     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1000 }
1001
1002 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1003 {
1004     struct kvm_set_guest_debug_data data;
1005
1006     data.dbg.control = reinject_trap;
1007
1008     if (env->singlestep_enabled) {
1009         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1010     }
1011     kvm_arch_update_guest_debug(env, &data.dbg);
1012     data.env = env;
1013
1014     on_vcpu(env, kvm_invoke_set_guest_debug, &data);
1015     return data.err;
1016 }
1017
1018 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1019                           target_ulong len, int type)
1020 {
1021     struct kvm_sw_breakpoint *bp;
1022     CPUState *env;
1023     int err;
1024
1025     if (type == GDB_BREAKPOINT_SW) {
1026         bp = kvm_find_sw_breakpoint(current_env, addr);
1027         if (bp) {
1028             bp->use_count++;
1029             return 0;
1030         }
1031
1032         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1033         if (!bp)
1034             return -ENOMEM;
1035
1036         bp->pc = addr;
1037         bp->use_count = 1;
1038         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1039         if (err) {
1040             free(bp);
1041             return err;
1042         }
1043
1044         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1045                           bp, entry);
1046     } else {
1047         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1048         if (err)
1049             return err;
1050     }
1051
1052     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1053         err = kvm_update_guest_debug(env, 0);
1054         if (err)
1055             return err;
1056     }
1057     return 0;
1058 }
1059
1060 int kvm_remove_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             return -ENOENT;
1071
1072         if (bp->use_count > 1) {
1073             bp->use_count--;
1074             return 0;
1075         }
1076
1077         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1078         if (err)
1079             return err;
1080
1081         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1082         qemu_free(bp);
1083     } else {
1084         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1085         if (err)
1086             return err;
1087     }
1088
1089     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1090         err = kvm_update_guest_debug(env, 0);
1091         if (err)
1092             return err;
1093     }
1094     return 0;
1095 }
1096
1097 void kvm_remove_all_breakpoints(CPUState *current_env)
1098 {
1099     struct kvm_sw_breakpoint *bp, *next;
1100     KVMState *s = current_env->kvm_state;
1101     CPUState *env;
1102
1103     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1104         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1105             /* Try harder to find a CPU that currently sees the breakpoint. */
1106             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1107                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1108                     break;
1109             }
1110         }
1111     }
1112     kvm_arch_remove_all_hw_breakpoints();
1113
1114     for (env = first_cpu; env != NULL; env = env->next_cpu)
1115         kvm_update_guest_debug(env, 0);
1116 }
1117
1118 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1119
1120 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1121 {
1122     return -EINVAL;
1123 }
1124
1125 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1126                           target_ulong len, int type)
1127 {
1128     return -EINVAL;
1129 }
1130
1131 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1132                           target_ulong len, int type)
1133 {
1134     return -EINVAL;
1135 }
1136
1137 void kvm_remove_all_breakpoints(CPUState *current_env)
1138 {
1139 }
1140 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1141
1142 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1143 {
1144     struct kvm_signal_mask *sigmask;
1145     int r;
1146
1147     if (!sigset)
1148         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1149
1150     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1151
1152     sigmask->len = 8;
1153     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1154     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1155     free(sigmask);
1156
1157     return r;
1158 }
1159
1160 #ifdef KVM_IOEVENTFD
1161 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1162 {
1163     struct kvm_ioeventfd kick = {
1164         .datamatch = val,
1165         .addr = addr,
1166         .len = 2,
1167         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1168         .fd = fd,
1169     };
1170     int r;
1171     if (!kvm_enabled())
1172         return -ENOSYS;
1173     if (!assign)
1174         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1175     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1176     if (r < 0)
1177         return r;
1178     return 0;
1179 }
1180 #endif