kvm-all.c

   1 /*
   2  * QEMU KVM support
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *           Red Hat, Inc. 2008
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Glauber Costa     <gcosta@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include <sys/types.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/mman.h>
  19 #include <stdarg.h>
  20
  21 #include <linux/kvm.h>
  22
  23 #include "qemu-common.h"
  24 #include "qemu-barrier.h"
  25 #include "sysemu.h"
  26 #include "hw/hw.h"
  27 #include "gdbstub.h"
  28 #include "kvm.h"
  29 #include "bswap.h"
  30
  31 /* This check must be after config-host.h is included */
  32 #ifdef CONFIG_EVENTFD
  33 #include <sys/eventfd.h>
  34 #endif
  35
  36 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  37 #define PAGE_SIZE TARGET_PAGE_SIZE
  38
  39 //#define DEBUG_KVM
  40
  41 #ifdef DEBUG_KVM
  42 #define DPRINTF(fmt, ...) \
  43     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  44 #else
  45 #define DPRINTF(fmt, ...) \
  46     do { } while (0)
  47 #endif
  48
  49 #ifdef OBSOLETE_KVM_IMPL
  50
  51 typedef struct KVMSlot
  52 {
  53     target_phys_addr_t start_addr;
  54     ram_addr_t memory_size;
  55     ram_addr_t phys_offset;
  56     int slot;
  57     int flags;
  58 } KVMSlot;
  59
  60 typedef struct kvm_dirty_log KVMDirtyLog;
  61
  62 struct KVMState
  63 {
  64     KVMSlot slots[32];
  65     int fd;
  66     int vmfd;
  67     int coalesced_mmio;
  68     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  69     int broken_set_mem_region;
  70     int migration_log;
  71     int vcpu_events;
  72     int robust_singlestep;
  73     int debugregs;
  74 #ifdef KVM_CAP_SET_GUEST_DEBUG
  75     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  76 #endif
  77     int irqchip_in_kernel;
  78     int pit_in_kernel;
  79     int xsave, xcrs;
  80     int many_ioeventfds;
  81 };
  82
  83 KVMState *kvm_state;
  84
  85
  86 static const KVMCapabilityInfo kvm_required_capabilites[] = {
  87     KVM_CAP_INFO(USER_MEMORY),
  88     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
  89     KVM_CAP_LAST_INFO
  90 };
  91
  92 #endif
  93
  94 static KVMSlot *kvm_alloc_slot(KVMState *s)
  95 {
  96     int i;
  97
  98     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  99         if (s->slots[i].memory_size == 0) {
 100             return &s->slots[i];
 101         }
 102     }
 103
 104     fprintf(stderr, "%s: no free slot available\n", __func__);
 105     abort();
 106 }
 107
 108 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 109                                          target_phys_addr_t start_addr,
 110                                          target_phys_addr_t end_addr)
 111 {
 112     int i;
 113
 114     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 115         KVMSlot *mem = &s->slots[i];
 116
 117         if (start_addr == mem->start_addr &&
 118             end_addr == mem->start_addr + mem->memory_size) {
 119             return mem;
 120         }
 121     }
 122
 123     return NULL;
 124 }
 125
 126 /*
 127  * Find overlapping slot with lowest start address
 128  */
 129 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 130                                             target_phys_addr_t start_addr,
 131                                             target_phys_addr_t end_addr)
 132 {
 133     KVMSlot *found = NULL;
 134     int i;
 135
 136     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 137         KVMSlot *mem = &s->slots[i];
 138
 139         if (mem->memory_size == 0 ||
 140             (found && found->start_addr < mem->start_addr)) {
 141             continue;
 142         }
 143
 144         if (end_addr > mem->start_addr &&
 145             start_addr < mem->start_addr + mem->memory_size) {
 146             found = mem;
 147         }
 148     }
 149
 150     return found;
 151 }
 152
 153 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
 154                                       target_phys_addr_t *phys_addr)
 155 {
 156     int i;
 157
 158     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 159         KVMSlot *mem = &s->slots[i];
 160
 161         if (ram_addr >= mem->phys_offset &&
 162             ram_addr < mem->phys_offset + mem->memory_size) {
 163             *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
 164             return 1;
 165         }
 166     }
 167
 168     return 0;
 169 }
 170
 171 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 172 {
 173     struct kvm_userspace_memory_region mem;
 174
 175     mem.slot = slot->slot;
 176     mem.guest_phys_addr = slot->start_addr;
 177     mem.memory_size = slot->memory_size;
 178     mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
 179     mem.flags = slot->flags;
 180     if (s->migration_log) {
 181         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 182     }
 183     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 184 }
 185
 186 #ifdef OBSOLETE_KVM_IMPL
 187 static void kvm_reset_vcpu(void *opaque)
 188 {
 189     CPUState *env = opaque;
 190
 191     kvm_arch_reset_vcpu(env);
 192 }
 193 #endif
 194
 195 int kvm_irqchip_in_kernel(void)
 196 {
 197     return kvm_state->irqchip_in_kernel;
 198 }
 199
 200 int kvm_pit_in_kernel(void)
 201 {
 202     return kvm_state->pit_in_kernel;
 203 }
 204
 205 #ifdef OBSOLETE_KVM_IMPL
 206 int kvm_init_vcpu(CPUState *env)
 207 {
 208     KVMState *s = kvm_state;
 209     long mmap_size;
 210     int ret;
 211
 212     DPRINTF("kvm_init_vcpu\n");
 213
 214     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 215     if (ret < 0) {
 216         DPRINTF("kvm_create_vcpu failed\n");
 217         goto err;
 218     }
 219
 220     env->kvm_fd = ret;
 221     env->kvm_state = s;
 222     env->kvm_vcpu_dirty = 1;
 223
 224     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 225     if (mmap_size < 0) {
 226         ret = mmap_size;
 227         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 228         goto err;
 229     }
 230
 231     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 232                         env->kvm_fd, 0);
 233     if (env->kvm_run == MAP_FAILED) {
 234         ret = -errno;
 235         DPRINTF("mmap'ing vcpu state failed\n");
 236         goto err;
 237     }
 238
 239     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 240         s->coalesced_mmio_ring =
 241             (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 242     }
 243
 244     ret = kvm_arch_init_vcpu(env);
 245     if (ret == 0) {
 246         qemu_register_reset(kvm_reset_vcpu, env);
 247         kvm_arch_reset_vcpu(env);
 248     }
 249 err:
 250     return ret;
 251 }
 252 #endif
 253
 254 /*
 255  * dirty pages logging control
 256  */
 257 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 258                                       ram_addr_t size, int flags, int mask)
 259 {
 260     KVMState *s = kvm_state;
 261     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 262     int old_flags;
 263
 264     if (mem == NULL)  {
 265             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 266                     TARGET_FMT_plx "\n", __func__, phys_addr,
 267                     (target_phys_addr_t)(phys_addr + size - 1));
 268             return -EINVAL;
 269     }
 270
 271     old_flags = mem->flags;
 272
 273     flags = (mem->flags & ~mask) | flags;
 274     mem->flags = flags;
 275
 276     /* If nothing changed effectively, no need to issue ioctl */
 277     if (s->migration_log) {
 278         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 279     }
 280     if (flags == old_flags) {
 281             return 0;
 282     }
 283
 284     return kvm_set_user_memory_region(s, mem);
 285 }
 286
 287 static int kvm_log_start(CPUPhysMemoryClient *client,
 288                          target_phys_addr_t phys_addr, ram_addr_t size)
 289 {
 290     return kvm_dirty_pages_log_change(phys_addr, size, KVM_MEM_LOG_DIRTY_PAGES,
 291                                       KVM_MEM_LOG_DIRTY_PAGES);
 292 }
 293
 294 static int kvm_log_stop(CPUPhysMemoryClient *client,
 295                         target_phys_addr_t phys_addr, ram_addr_t size)
 296 {
 297     return kvm_dirty_pages_log_change(phys_addr, size, 0,
 298                                       KVM_MEM_LOG_DIRTY_PAGES);
 299 }
 300
 301 static int kvm_set_migration_log(int enable)
 302 {
 303     KVMState *s = kvm_state;
 304     KVMSlot *mem;
 305     int i, err;
 306
 307     s->migration_log = enable;
 308
 309     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 310         mem = &s->slots[i];
 311
 312         if (!mem->memory_size) {
 313             continue;
 314         }
 315         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 316             continue;
 317         }
 318         err = kvm_set_user_memory_region(s, mem);
 319         if (err) {
 320             return err;
 321         }
 322     }
 323     return 0;
 324 }
 325
 326 /* get kvm's dirty pages bitmap and update qemu's */
 327 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 328                                          unsigned long *bitmap,
 329                                          unsigned long offset,
 330                                          unsigned long mem_size)
 331 {
 332     unsigned int i, j;
 333     unsigned long page_number, addr, addr1, c;
 334     ram_addr_t ram_addr;
 335     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 336         HOST_LONG_BITS;
 337
 338     /*
 339      * bitmap-traveling is faster than memory-traveling (for addr...)
 340      * especially when most of the memory is not dirty.
 341      */
 342     for (i = 0; i < len; i++) {
 343         if (bitmap[i] != 0) {
 344             c = leul_to_cpu(bitmap[i]);
 345             do {
 346                 j = ffsl(c) - 1;
 347                 c &= ~(1ul << j);
 348                 page_number = i * HOST_LONG_BITS + j;
 349                 addr1 = page_number * TARGET_PAGE_SIZE;
 350                 addr = offset + addr1;
 351                 ram_addr = cpu_get_physical_page_desc(addr);
 352                 cpu_physical_memory_set_dirty(ram_addr);
 353             } while (c != 0);
 354         }
 355     }
 356     return 0;
 357 }
 358
 359 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 360
 361 /**
 362  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 363  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 364  * This means all bits are set to dirty.
 365  *
 366  * @start_add: start of logged region.
 367  * @end_addr: end of logged region.
 368  */
 369 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 370                                           target_phys_addr_t end_addr)
 371 {
 372     KVMState *s = kvm_state;
 373     unsigned long size, allocated_size = 0;
 374     KVMDirtyLog d;
 375     KVMSlot *mem;
 376     int ret = 0;
 377
 378     d.dirty_bitmap = NULL;
 379     while (start_addr < end_addr) {
 380         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 381         if (mem == NULL) {
 382             break;
 383         }
 384
 385         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 386         if (!d.dirty_bitmap) {
 387             d.dirty_bitmap = qemu_malloc(size);
 388         } else if (size > allocated_size) {
 389             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 390         }
 391         allocated_size = size;
 392         memset(d.dirty_bitmap, 0, allocated_size);
 393
 394         d.slot = mem->slot;
 395
 396         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 397             DPRINTF("ioctl failed %d\n", errno);
 398             ret = -1;
 399             break;
 400         }
 401
 402         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 403                                       mem->start_addr, mem->memory_size);
 404         start_addr = mem->start_addr + mem->memory_size;
 405     }
 406     qemu_free(d.dirty_bitmap);
 407
 408     return ret;
 409 }
 410
 411 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 412 {
 413     int ret = -ENOSYS;
 414     KVMState *s = kvm_state;
 415
 416     if (s->coalesced_mmio) {
 417         struct kvm_coalesced_mmio_zone zone;
 418
 419         zone.addr = start;
 420         zone.size = size;
 421
 422         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 423     }
 424
 425     return ret;
 426 }
 427
 428 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 429 {
 430     int ret = -ENOSYS;
 431     KVMState *s = kvm_state;
 432
 433     if (s->coalesced_mmio) {
 434         struct kvm_coalesced_mmio_zone zone;
 435
 436         zone.addr = start;
 437         zone.size = size;
 438
 439         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 440     }
 441
 442     return ret;
 443 }
 444
 445 int kvm_check_extension(KVMState *s, unsigned int extension)
 446 {
 447     int ret;
 448
 449     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 450     if (ret < 0) {
 451         ret = 0;
 452     }
 453
 454     return ret;
 455 }
 456
 457 static int kvm_check_many_ioeventfds(void)
 458 {
 459     /* Userspace can use ioeventfd for io notification.  This requires a host
 460      * that supports eventfd(2) and an I/O thread; since eventfd does not
 461      * support SIGIO it cannot interrupt the vcpu.
 462      *
 463      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 464      * can avoid creating too many ioeventfds.
 465      */
 466 #if defined(CONFIG_EVENTFD) /* && defined(CONFIG_IOTHREAD) */
 467     int ioeventfds[7];
 468     int i, ret = 0;
 469     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 470         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 471         if (ioeventfds[i] < 0) {
 472             break;
 473         }
 474         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 475         if (ret < 0) {
 476             close(ioeventfds[i]);
 477             break;
 478         }
 479     }
 480
 481     /* Decide whether many devices are supported or not */
 482     ret = i == ARRAY_SIZE(ioeventfds);
 483
 484     while (i-- > 0) {
 485         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 486         close(ioeventfds[i]);
 487     }
 488     return ret;
 489 #else
 490     return 0;
 491 #endif
 492 }
 493
 494 #ifdef OBSOLETE_KVM_IMPL
 495 static const KVMCapabilityInfo *
 496 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 497 {
 498     while (list->name) {
 499         if (!kvm_check_extension(s, list->value)) {
 500             return list;
 501         }
 502         list++;
 503     }
 504     return NULL;
 505 }
 506 #endif
 507
 508 static void kvm_set_phys_mem(target_phys_addr_t start_addr, ram_addr_t size,
 509                              ram_addr_t phys_offset)
 510 {
 511     KVMState *s = kvm_state;
 512     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 513     KVMSlot *mem, old;
 514     int err;
 515
 516     /* kvm works in page size chunks, but the function may be called
 517        with sub-page size and unaligned start address. */
 518     size = TARGET_PAGE_ALIGN(size);
 519     start_addr = TARGET_PAGE_ALIGN(start_addr);
 520
 521     /* KVM does not support read-only slots */
 522     phys_offset &= ~IO_MEM_ROM;
 523
 524     while (1) {
 525         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 526         if (!mem) {
 527             break;
 528         }
 529
 530         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 531             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 532             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 533             /* The new slot fits into the existing one and comes with
 534              * identical parameters - nothing to be done. */
 535             return;
 536         }
 537
 538         old = *mem;
 539
 540         /* unregister the overlapping slot */
 541         mem->memory_size = 0;
 542         err = kvm_set_user_memory_region(s, mem);
 543         if (err) {
 544             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 545                     __func__, strerror(-err));
 546             abort();
 547         }
 548
 549         /* Workaround for older KVM versions: we can't join slots, even not by
 550          * unregistering the previous ones and then registering the larger
 551          * slot. We have to maintain the existing fragmentation. Sigh.
 552          *
 553          * This workaround assumes that the new slot starts at the same
 554          * address as the first existing one. If not or if some overlapping
 555          * slot comes around later, we will fail (not seen in practice so far)
 556          * - and actually require a recent KVM version. */
 557         if (s->broken_set_mem_region &&
 558             old.start_addr == start_addr && old.memory_size < size &&
 559             flags < IO_MEM_UNASSIGNED) {
 560             mem = kvm_alloc_slot(s);
 561             mem->memory_size = old.memory_size;
 562             mem->start_addr = old.start_addr;
 563             mem->phys_offset = old.phys_offset;
 564             mem->flags = 0;
 565
 566             err = kvm_set_user_memory_region(s, mem);
 567             if (err) {
 568                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 569                         strerror(-err));
 570                 abort();
 571             }
 572
 573             start_addr += old.memory_size;
 574             phys_offset += old.memory_size;
 575             size -= old.memory_size;
 576             continue;
 577         }
 578
 579         /* register prefix slot */
 580         if (old.start_addr < start_addr) {
 581             mem = kvm_alloc_slot(s);
 582             mem->memory_size = start_addr - old.start_addr;
 583             mem->start_addr = old.start_addr;
 584             mem->phys_offset = old.phys_offset;
 585             mem->flags = 0;
 586
 587             err = kvm_set_user_memory_region(s, mem);
 588             if (err) {
 589                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 590                         __func__, strerror(-err));
 591                 abort();
 592             }
 593         }
 594
 595         /* register suffix slot */
 596         if (old.start_addr + old.memory_size > start_addr + size) {
 597             ram_addr_t size_delta;
 598
 599             mem = kvm_alloc_slot(s);
 600             mem->start_addr = start_addr + size;
 601             size_delta = mem->start_addr - old.start_addr;
 602             mem->memory_size = old.memory_size - size_delta;
 603             mem->phys_offset = old.phys_offset + size_delta;
 604             mem->flags = 0;
 605
 606             err = kvm_set_user_memory_region(s, mem);
 607             if (err) {
 608                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 609                         __func__, strerror(-err));
 610                 abort();
 611             }
 612         }
 613     }
 614
 615     /* in case the KVM bug workaround already "consumed" the new slot */
 616     if (!size) {
 617         return;
 618     }
 619     /* KVM does not need to know about this memory */
 620     if (flags >= IO_MEM_UNASSIGNED) {
 621         return;
 622     }
 623     mem = kvm_alloc_slot(s);
 624     mem->memory_size = size;
 625     mem->start_addr = start_addr;
 626     mem->phys_offset = phys_offset;
 627     mem->flags = 0;
 628
 629     err = kvm_set_user_memory_region(s, mem);
 630     if (err) {
 631         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 632                 strerror(-err));
 633         abort();
 634     }
 635 }
 636
 637 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 638                                   target_phys_addr_t start_addr,
 639                                   ram_addr_t size, ram_addr_t phys_offset)
 640 {
 641     kvm_set_phys_mem(start_addr, size, phys_offset);
 642 }
 643
 644 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 645                                         target_phys_addr_t start_addr,
 646                                         target_phys_addr_t end_addr)
 647 {
 648     return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 649 }
 650
 651 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 652                                     int enable)
 653 {
 654     return kvm_set_migration_log(enable);
 655 }
 656
 657 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 658     .set_memory = kvm_client_set_memory,
 659     .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 660     .migration_log = kvm_client_migration_log,
 661     .log_start = kvm_log_start,
 662     .log_stop = kvm_log_stop,
 663 };
 664
 665 void kvm_cpu_register_phys_memory_client(void)
 666 {
 667     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 668 }
 669
 670 #ifdef OBSOLETE_KVM_IMPL
 671
 672 int kvm_init(void)
 673 {
 674     static const char upgrade_note[] =
 675         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 676         "(see http://sourceforge.net/projects/kvm).\n";
 677     KVMState *s;
 678     const KVMCapabilityInfo *missing_cap;
 679     int ret;
 680     int i;
 681
 682     s = qemu_mallocz(sizeof(KVMState));
 683
 684 #ifdef KVM_CAP_SET_GUEST_DEBUG
 685     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 686 #endif
 687     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 688         s->slots[i].slot = i;
 689     }
 690     s->vmfd = -1;
 691     s->fd = qemu_open("/dev/kvm", O_RDWR);
 692     if (s->fd == -1) {
 693         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 694         ret = -errno;
 695         goto err;
 696     }
 697
 698     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 699     if (ret < KVM_API_VERSION) {
 700         if (ret > 0) {
 701             ret = -EINVAL;
 702         }
 703         fprintf(stderr, "kvm version too old\n");
 704         goto err;
 705     }
 706
 707     if (ret > KVM_API_VERSION) {
 708         ret = -EINVAL;
 709         fprintf(stderr, "kvm version not supported\n");
 710         goto err;
 711     }
 712
 713     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 714     if (s->vmfd < 0) {
 715 #ifdef TARGET_S390X
 716         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 717                         "your host kernel command line\n");
 718 #endif
 719         goto err;
 720     }
 721
 722     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
 723     if (!missing_cap) {
 724         missing_cap =
 725             kvm_check_extension_list(s, kvm_arch_required_capabilities);
 726     }
 727     if (missing_cap) {
 728         ret = -EINVAL;
 729         fprintf(stderr, "kvm does not support %s\n%s",
 730                 missing_cap->name, upgrade_note);
 731         goto err;
 732     }
 733
 734     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 735
 736     s->broken_set_mem_region = 1;
 737 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 738     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 739     if (ret > 0) {
 740         s->broken_set_mem_region = 0;
 741     }
 742 #endif
 743
 744     s->vcpu_events = 0;
 745 #ifdef KVM_CAP_VCPU_EVENTS
 746     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 747 #endif
 748
 749     s->robust_singlestep = 0;
 750 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 751     s->robust_singlestep =
 752         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 753 #endif
 754
 755     s->debugregs = 0;
 756 #ifdef KVM_CAP_DEBUGREGS
 757     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 758 #endif
 759
 760     s->xsave = 0;
 761 #ifdef KVM_CAP_XSAVE
 762     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 763 #endif
 764
 765     s->xcrs = 0;
 766 #ifdef KVM_CAP_XCRS
 767     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 768 #endif
 769
 770     ret = kvm_arch_init(s);
 771     if (ret < 0) {
 772         goto err;
 773     }
 774
 775     kvm_state = s;
 776     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 777
 778     s->many_ioeventfds = kvm_check_many_ioeventfds();
 779
 780     return 0;
 781
 782 err:
 783     if (s) {
 784         if (s->vmfd != -1) {
 785             close(s->vmfd);
 786         }
 787         if (s->fd != -1) {
 788             close(s->fd);
 789         }
 790     }
 791     qemu_free(s);
 792
 793     return ret;
 794 }
 795 #endif
 796
 797 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
 798                           uint32_t count)
 799 {
 800     int i;
 801     uint8_t *ptr = data;
 802
 803     for (i = 0; i < count; i++) {
 804         if (direction == KVM_EXIT_IO_IN) {
 805             switch (size) {
 806             case 1:
 807                 stb_p(ptr, cpu_inb(port));
 808                 break;
 809             case 2:
 810                 stw_p(ptr, cpu_inw(port));
 811                 break;
 812             case 4:
 813                 stl_p(ptr, cpu_inl(port));
 814                 break;
 815             }
 816         } else {
 817             switch (size) {
 818             case 1:
 819                 cpu_outb(port, ldub_p(ptr));
 820                 break;
 821             case 2:
 822                 cpu_outw(port, lduw_p(ptr));
 823                 break;
 824             case 4:
 825                 cpu_outl(port, ldl_p(ptr));
 826                 break;
 827             }
 828         }
 829
 830         ptr += size;
 831     }
 832 }
 833
 834 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 835 static int kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 836 {
 837     fprintf(stderr, "KVM internal error.");
 838     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 839         int i;
 840
 841         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
 842         for (i = 0; i < run->internal.ndata; ++i) {
 843             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 844                     i, (uint64_t)run->internal.data[i]);
 845         }
 846     } else {
 847         fprintf(stderr, "\n");
 848     }
 849     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 850         fprintf(stderr, "emulation failure\n");
 851         if (!kvm_arch_stop_on_emulation_error(env)) {
 852             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 853             return EXCP_INTERRUPT;
 854         }
 855     }
 856     /* FIXME: Should trigger a qmp message to let management know
 857      * something went wrong.
 858      */
 859     return -1;
 860 }
 861 #endif
 862
 863 void kvm_flush_coalesced_mmio_buffer(void)
 864 {
 865     KVMState *s = kvm_state;
 866     if (s->coalesced_mmio_ring) {
 867         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 868         while (ring->first != ring->last) {
 869             struct kvm_coalesced_mmio *ent;
 870
 871             ent = &ring->coalesced_mmio[ring->first];
 872
 873             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 874             smp_wmb();
 875             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 876         }
 877     }
 878 }
 879
 880 #ifdef OBSOLETE_KVM_IMPL
 881
 882 static void do_kvm_cpu_synchronize_state(void *_env)
 883 {
 884     CPUState *env = _env;
 885
 886     if (!env->kvm_vcpu_dirty) {
 887         kvm_arch_get_registers(env);
 888         env->kvm_vcpu_dirty = 1;
 889     }
 890 }
 891
 892 void kvm_cpu_synchronize_state(CPUState *env)
 893 {
 894     if (!env->kvm_vcpu_dirty) {
 895         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 896     }
 897 }
 898
 899 void kvm_cpu_synchronize_post_reset(CPUState *env)
 900 {
 901     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 902     env->kvm_vcpu_dirty = 0;
 903 }
 904
 905 void kvm_cpu_synchronize_post_init(CPUState *env)
 906 {
 907     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 908     env->kvm_vcpu_dirty = 0;
 909 }
 910
 911 int kvm_cpu_exec(CPUState *env)
 912 {
 913     struct kvm_run *run = env->kvm_run;
 914     int ret, run_ret;
 915
 916     DPRINTF("kvm_cpu_exec()\n");
 917
 918     if (kvm_arch_process_async_events(env)) {
 919         env->exit_request = 0;
 920         return EXCP_HLT;
 921     }
 922
 923     cpu_single_env = env;
 924
 925     do {
 926         if (env->kvm_vcpu_dirty) {
 927             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 928             env->kvm_vcpu_dirty = 0;
 929         }
 930
 931         kvm_arch_pre_run(env, run);
 932         if (env->exit_request) {
 933             DPRINTF("interrupt exit requested\n");
 934             /*
 935              * KVM requires us to reenter the kernel after IO exits to complete
 936              * instruction emulation. This self-signal will ensure that we
 937              * leave ASAP again.
 938              */
 939             qemu_cpu_kick_self();
 940         }
 941         cpu_single_env = NULL;
 942         qemu_mutex_unlock_iothread();
 943
 944         run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 945
 946         qemu_mutex_lock_iothread();
 947         cpu_single_env = env;
 948         kvm_arch_post_run(env, run);
 949
 950         kvm_flush_coalesced_mmio_buffer();
 951
 952         if (run_ret < 0) {
 953             if (run_ret == -EINTR || run_ret == -EAGAIN) {
 954                 DPRINTF("io window exit\n");
 955                 ret = EXCP_INTERRUPT;
 956                 break;
 957             }
 958             DPRINTF("kvm run failed %s\n", strerror(-run_ret));
 959             abort();
 960         }
 961
 962         switch (run->exit_reason) {
 963         case KVM_EXIT_IO:
 964             DPRINTF("handle_io\n");
 965             kvm_handle_io(run->io.port,
 966                           (uint8_t *)run + run->io.data_offset,
 967                           run->io.direction,
 968                           run->io.size,
 969                           run->io.count);
 970             ret = 0;
 971             break;
 972         case KVM_EXIT_MMIO:
 973             DPRINTF("handle_mmio\n");
 974             cpu_physical_memory_rw(run->mmio.phys_addr,
 975                                    run->mmio.data,
 976                                    run->mmio.len,
 977                                    run->mmio.is_write);
 978             ret = 0;
 979             break;
 980         case KVM_EXIT_IRQ_WINDOW_OPEN:
 981             DPRINTF("irq_window_open\n");
 982             ret = EXCP_INTERRUPT;
 983             break;
 984         case KVM_EXIT_SHUTDOWN:
 985             DPRINTF("shutdown\n");
 986             qemu_system_reset_request();
 987             ret = EXCP_INTERRUPT;
 988             break;
 989         case KVM_EXIT_UNKNOWN:
 990             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
 991                     (uint64_t)run->hw.hardware_exit_reason);
 992             ret = -1;
 993             break;
 994 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 995         case KVM_EXIT_INTERNAL_ERROR:
 996             ret = kvm_handle_internal_error(env, run);
 997             break;
 998 #endif
 999         default:
1000             DPRINTF("kvm_arch_handle_exit\n");
1001             ret = kvm_arch_handle_exit(env, run);
1002             break;
1003         }
1004     } while (ret == 0);
1005
1006     if (ret < 0) {
1007         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1008         vm_stop(VMSTOP_PANIC);
1009     }
1010
1011     env->exit_request = 0;
1012     cpu_single_env = NULL;
1013     return ret;
1014 }
1015
1016 #endif
1017 int kvm_ioctl(KVMState *s, int type, ...)
1018 {
1019     int ret;
1020     void *arg;
1021     va_list ap;
1022
1023     va_start(ap, type);
1024     arg = va_arg(ap, void *);
1025     va_end(ap);
1026
1027     ret = ioctl(s->fd, type, arg);
1028     if (ret == -1) {
1029         ret = -errno;
1030     }
1031     return ret;
1032 }
1033
1034 int kvm_vm_ioctl(KVMState *s, int type, ...)
1035 {
1036     int ret;
1037     void *arg;
1038     va_list ap;
1039
1040     va_start(ap, type);
1041     arg = va_arg(ap, void *);
1042     va_end(ap);
1043
1044     ret = ioctl(s->vmfd, type, arg);
1045     if (ret == -1) {
1046         ret = -errno;
1047     }
1048     return ret;
1049 }
1050
1051 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1052 {
1053     int ret;
1054     void *arg;
1055     va_list ap;
1056
1057     va_start(ap, type);
1058     arg = va_arg(ap, void *);
1059     va_end(ap);
1060
1061     ret = ioctl(env->kvm_fd, type, arg);
1062     if (ret == -1) {
1063         ret = -errno;
1064     }
1065     return ret;
1066 }
1067
1068 int kvm_has_sync_mmu(void)
1069 {
1070     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1071 }
1072
1073 int kvm_has_vcpu_events(void)
1074 {
1075     return kvm_state->vcpu_events;
1076 }
1077
1078 int kvm_has_robust_singlestep(void)
1079 {
1080     return kvm_state->robust_singlestep;
1081 }
1082
1083 int kvm_has_debugregs(void)
1084 {
1085     return kvm_state->debugregs;
1086 }
1087
1088 int kvm_has_xsave(void)
1089 {
1090     return kvm_state->xsave;
1091 }
1092
1093 int kvm_has_xcrs(void)
1094 {
1095     return kvm_state->xcrs;
1096 }
1097
1098 int kvm_has_many_ioeventfds(void)
1099 {
1100     if (!kvm_enabled()) {
1101         return 0;
1102     }
1103     return kvm_state->many_ioeventfds;
1104 }
1105
1106 void kvm_setup_guest_memory(void *start, size_t size)
1107 {
1108     if (!kvm_has_sync_mmu()) {
1109         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1110
1111         if (ret) {
1112             perror("qemu_madvise");
1113             fprintf(stderr,
1114                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1115             exit(1);
1116         }
1117     }
1118 }
1119
1120 #ifdef KVM_CAP_SET_GUEST_DEBUG
1121 #ifndef OBSOLETE_KVM_IMPL
1122 #define run_on_cpu on_vcpu
1123 #endif /* !OBSOLETE_KVM_IMPL */
1124
1125 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1126                                                  target_ulong pc)
1127 {
1128     struct kvm_sw_breakpoint *bp;
1129
1130     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1131         if (bp->pc == pc) {
1132             return bp;
1133         }
1134     }
1135     return NULL;
1136 }
1137
1138 int kvm_sw_breakpoints_active(CPUState *env)
1139 {
1140     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1141 }
1142
1143 struct kvm_set_guest_debug_data {
1144     struct kvm_guest_debug dbg;
1145     CPUState *env;
1146     int err;
1147 };
1148
1149 static void kvm_invoke_set_guest_debug(void *data)
1150 {
1151     struct kvm_set_guest_debug_data *dbg_data = data;
1152     CPUState *env = dbg_data->env;
1153
1154     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1155 }
1156
1157 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1158 {
1159     struct kvm_set_guest_debug_data data;
1160
1161     data.dbg.control = reinject_trap;
1162
1163     if (env->singlestep_enabled) {
1164         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1165     }
1166     kvm_arch_update_guest_debug(env, &data.dbg);
1167     data.env = env;
1168
1169     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1170     return data.err;
1171 }
1172
1173 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1174                           target_ulong len, int type)
1175 {
1176     struct kvm_sw_breakpoint *bp;
1177     CPUState *env;
1178     int err;
1179
1180     if (type == GDB_BREAKPOINT_SW) {
1181         bp = kvm_find_sw_breakpoint(current_env, addr);
1182         if (bp) {
1183             bp->use_count++;
1184             return 0;
1185         }
1186
1187         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1188         if (!bp) {
1189             return -ENOMEM;
1190         }
1191
1192         bp->pc = addr;
1193         bp->use_count = 1;
1194         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1195         if (err) {
1196             free(bp);
1197             return err;
1198         }
1199
1200         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1201                           bp, entry);
1202     } else {
1203         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1204         if (err) {
1205             return err;
1206         }
1207     }
1208
1209     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1210         err = kvm_update_guest_debug(env, 0);
1211         if (err) {
1212             return err;
1213         }
1214     }
1215     return 0;
1216 }
1217
1218 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1219                           target_ulong len, int type)
1220 {
1221     struct kvm_sw_breakpoint *bp;
1222     CPUState *env;
1223     int err;
1224
1225     if (type == GDB_BREAKPOINT_SW) {
1226         bp = kvm_find_sw_breakpoint(current_env, addr);
1227         if (!bp) {
1228             return -ENOENT;
1229         }
1230
1231         if (bp->use_count > 1) {
1232             bp->use_count--;
1233             return 0;
1234         }
1235
1236         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1237         if (err) {
1238             return err;
1239         }
1240
1241         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1242         qemu_free(bp);
1243     } else {
1244         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1245         if (err) {
1246             return err;
1247         }
1248     }
1249
1250     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1251         err = kvm_update_guest_debug(env, 0);
1252         if (err) {
1253             return err;
1254         }
1255     }
1256     return 0;
1257 }
1258
1259 void kvm_remove_all_breakpoints(CPUState *current_env)
1260 {
1261     struct kvm_sw_breakpoint *bp, *next;
1262     KVMState *s = current_env->kvm_state;
1263     CPUState *env;
1264
1265     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1266         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1267             /* Try harder to find a CPU that currently sees the breakpoint. */
1268             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1269                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1270                     break;
1271                 }
1272             }
1273         }
1274     }
1275     kvm_arch_remove_all_hw_breakpoints();
1276
1277     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1278         kvm_update_guest_debug(env, 0);
1279     }
1280 }
1281
1282 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1283
1284 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1285 {
1286     return -EINVAL;
1287 }
1288
1289 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1290                           target_ulong len, int type)
1291 {
1292     return -EINVAL;
1293 }
1294
1295 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1296                           target_ulong len, int type)
1297 {
1298     return -EINVAL;
1299 }
1300
1301 void kvm_remove_all_breakpoints(CPUState *current_env)
1302 {
1303 }
1304 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1305
1306 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1307 {
1308     struct kvm_signal_mask *sigmask;
1309     int r;
1310
1311     if (!sigset) {
1312         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1313     }
1314
1315     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1316
1317     sigmask->len = 8;
1318     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1319     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1320     free(sigmask);
1321
1322     return r;
1323 }
1324
1325 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1326 {
1327 #ifdef KVM_IOEVENTFD
1328     int ret;
1329     struct kvm_ioeventfd iofd;
1330
1331     iofd.datamatch = val;
1332     iofd.addr = addr;
1333     iofd.len = 4;
1334     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1335     iofd.fd = fd;
1336
1337     if (!kvm_enabled()) {
1338         return -ENOSYS;
1339     }
1340
1341     if (!assign) {
1342         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1343     }
1344
1345     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1346
1347     if (ret < 0) {
1348         return -errno;
1349     }
1350
1351     return 0;
1352 #else
1353     return -ENOSYS;
1354 #endif
1355 }
1356
1357 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1358 {
1359 #ifdef KVM_IOEVENTFD
1360     struct kvm_ioeventfd kick = {
1361         .datamatch = val,
1362         .addr = addr,
1363         .len = 2,
1364         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1365         .fd = fd,
1366     };
1367     int r;
1368     if (!kvm_enabled()) {
1369         return -ENOSYS;
1370     }
1371     if (!assign) {
1372         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1373     }
1374     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1375     if (r < 0) {
1376         return r;
1377     }
1378     return 0;
1379 #else
1380     return -ENOSYS;
1381 #endif
1382 }
1383
1384 #if defined(KVM_IRQFD)
1385 int kvm_set_irqfd(int gsi, int fd, bool assigned)
1386 {
1387     struct kvm_irqfd irqfd = {
1388         .fd = fd,
1389         .gsi = gsi,
1390         .flags = assigned ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1391     };
1392     int r;
1393     if (!kvm_enabled() || !kvm_irqchip_in_kernel())
1394         return -ENOSYS;
1395
1396     r = kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
1397     if (r < 0)
1398         return r;
1399     return 0;
1400 }
1401 #endif
1402
1403 int kvm_on_sigbus_vcpu(CPUState *env, int code, void *addr)
1404 {
1405     return kvm_arch_on_sigbus_vcpu(env, code, addr);
1406 }
1407
1408 int kvm_on_sigbus(int code, void *addr)
1409 {
1410     return kvm_arch_on_sigbus(code, addr);
1411 }
1412
1413 #undef PAGE_SIZE
1414 #include "qemu-kvm.c"