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