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 "qemu-option.h"
  26 #include "qemu-config.h"
  27 #include "sysemu.h"
  28 #include "hw/hw.h"
  29 #include "hw/msi.h"
  30 #include "gdbstub.h"
  31 #include "kvm.h"
  32 #include "bswap.h"
  33 #include "memory.h"
  34 #include "exec-memory.h"
  35
  36 /* This check must be after config-host.h is included */
  37 #ifdef CONFIG_EVENTFD
  38 #include <sys/eventfd.h>
  39 #endif
  40
  41 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
  42 #define PAGE_SIZE TARGET_PAGE_SIZE
  43
  44 //#define DEBUG_KVM
  45
  46 #ifdef DEBUG_KVM
  47 #define DPRINTF(fmt, ...) \
  48     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  49 #else
  50 #define DPRINTF(fmt, ...) \
  51     do { } while (0)
  52 #endif
  53
  54 #define KVM_MSI_HASHTAB_SIZE    256
  55
  56 typedef struct KVMSlot
  57 {
  58     target_phys_addr_t start_addr;
  59     ram_addr_t memory_size;
  60     void *ram;
  61     int slot;
  62     int flags;
  63 } KVMSlot;
  64
  65 typedef struct kvm_dirty_log KVMDirtyLog;
  66
  67 struct KVMState
  68 {
  69     KVMSlot slots[32];
  70     int fd;
  71     int vmfd;
  72     int coalesced_mmio;
  73     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  74     bool coalesced_flush_in_progress;
  75     int broken_set_mem_region;
  76     int migration_log;
  77     int vcpu_events;
  78     int robust_singlestep;
  79     int debugregs;
  80 #ifdef KVM_CAP_SET_GUEST_DEBUG
  81     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  82 #endif
  83     int pit_state2;
  84     int xsave, xcrs;
  85     int many_ioeventfds;
  86     /* The man page (and posix) say ioctl numbers are signed int, but
  87      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
  88      * unsigned, and treating them as signed here can break things */
  89     unsigned irqchip_inject_ioctl;
  90 #ifdef KVM_CAP_IRQ_ROUTING
  91     struct kvm_irq_routing *irq_routes;
  92     int nr_allocated_irq_routes;
  93     uint32_t *used_gsi_bitmap;
  94     unsigned int gsi_count;
  95     QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
  96     bool direct_msi;
  97 #endif
  98 };
  99
 100 KVMState *kvm_state;
 101 bool kvm_kernel_irqchip;
 102
 103 static const KVMCapabilityInfo kvm_required_capabilites[] = {
 104     KVM_CAP_INFO(USER_MEMORY),
 105     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
 106     KVM_CAP_LAST_INFO
 107 };
 108
 109 static KVMSlot *kvm_alloc_slot(KVMState *s)
 110 {
 111     int i;
 112
 113     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 114         if (s->slots[i].memory_size == 0) {
 115             return &s->slots[i];
 116         }
 117     }
 118
 119     fprintf(stderr, "%s: no free slot available\n", __func__);
 120     abort();
 121 }
 122
 123 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 124                                          target_phys_addr_t start_addr,
 125                                          target_phys_addr_t end_addr)
 126 {
 127     int i;
 128
 129     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 130         KVMSlot *mem = &s->slots[i];
 131
 132         if (start_addr == mem->start_addr &&
 133             end_addr == mem->start_addr + mem->memory_size) {
 134             return mem;
 135         }
 136     }
 137
 138     return NULL;
 139 }
 140
 141 /*
 142  * Find overlapping slot with lowest start address
 143  */
 144 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 145                                             target_phys_addr_t start_addr,
 146                                             target_phys_addr_t end_addr)
 147 {
 148     KVMSlot *found = NULL;
 149     int i;
 150
 151     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 152         KVMSlot *mem = &s->slots[i];
 153
 154         if (mem->memory_size == 0 ||
 155             (found && found->start_addr < mem->start_addr)) {
 156             continue;
 157         }
 158
 159         if (end_addr > mem->start_addr &&
 160             start_addr < mem->start_addr + mem->memory_size) {
 161             found = mem;
 162         }
 163     }
 164
 165     return found;
 166 }
 167
 168 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
 169                                        target_phys_addr_t *phys_addr)
 170 {
 171     int i;
 172
 173     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 174         KVMSlot *mem = &s->slots[i];
 175
 176         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
 177             *phys_addr = mem->start_addr + (ram - mem->ram);
 178             return 1;
 179         }
 180     }
 181
 182     return 0;
 183 }
 184
 185 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 186 {
 187     struct kvm_userspace_memory_region mem;
 188
 189     mem.slot = slot->slot;
 190     mem.guest_phys_addr = slot->start_addr;
 191     mem.memory_size = slot->memory_size;
 192     mem.userspace_addr = (unsigned long)slot->ram;
 193     mem.flags = slot->flags;
 194     if (s->migration_log) {
 195         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 196     }
 197     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 198 }
 199
 200 static void kvm_reset_vcpu(void *opaque)
 201 {
 202     CPUArchState *env = opaque;
 203
 204     kvm_arch_reset_vcpu(env);
 205 }
 206
 207 int kvm_init_vcpu(CPUArchState *env)
 208 {
 209     KVMState *s = kvm_state;
 210     long mmap_size;
 211     int ret;
 212
 213     DPRINTF("kvm_init_vcpu\n");
 214
 215     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 216     if (ret < 0) {
 217         DPRINTF("kvm_create_vcpu failed\n");
 218         goto err;
 219     }
 220
 221     env->kvm_fd = ret;
 222     env->kvm_state = s;
 223     env->kvm_vcpu_dirty = 1;
 224
 225     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 226     if (mmap_size < 0) {
 227         ret = mmap_size;
 228         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 229         goto err;
 230     }
 231
 232     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 233                         env->kvm_fd, 0);
 234     if (env->kvm_run == MAP_FAILED) {
 235         ret = -errno;
 236         DPRINTF("mmap'ing vcpu state failed\n");
 237         goto err;
 238     }
 239
 240     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 241         s->coalesced_mmio_ring =
 242             (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 243     }
 244
 245     ret = kvm_arch_init_vcpu(env);
 246     if (ret == 0) {
 247         qemu_register_reset(kvm_reset_vcpu, env);
 248         kvm_arch_reset_vcpu(env);
 249     }
 250 err:
 251     return ret;
 252 }
 253
 254 /*
 255  * dirty pages logging control
 256  */
 257
 258 static int kvm_mem_flags(KVMState *s, bool log_dirty)
 259 {
 260     return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
 261 }
 262
 263 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
 264 {
 265     KVMState *s = kvm_state;
 266     int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
 267     int old_flags;
 268
 269     old_flags = mem->flags;
 270
 271     flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
 272     mem->flags = flags;
 273
 274     /* If nothing changed effectively, no need to issue ioctl */
 275     if (s->migration_log) {
 276         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 277     }
 278
 279     if (flags == old_flags) {
 280         return 0;
 281     }
 282
 283     return kvm_set_user_memory_region(s, mem);
 284 }
 285
 286 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 287                                       ram_addr_t size, bool log_dirty)
 288 {
 289     KVMState *s = kvm_state;
 290     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 291
 292     if (mem == NULL)  {
 293         fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 294                 TARGET_FMT_plx "\n", __func__, phys_addr,
 295                 (target_phys_addr_t)(phys_addr + size - 1));
 296         return -EINVAL;
 297     }
 298     return kvm_slot_dirty_pages_log_change(mem, log_dirty);
 299 }
 300
 301 static void kvm_log_start(MemoryListener *listener,
 302                           MemoryRegionSection *section)
 303 {
 304     int r;
 305
 306     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 307                                    section->size, true);
 308     if (r < 0) {
 309         abort();
 310     }
 311 }
 312
 313 static void kvm_log_stop(MemoryListener *listener,
 314                           MemoryRegionSection *section)
 315 {
 316     int r;
 317
 318     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 319                                    section->size, false);
 320     if (r < 0) {
 321         abort();
 322     }
 323 }
 324
 325 static int kvm_set_migration_log(int enable)
 326 {
 327     KVMState *s = kvm_state;
 328     KVMSlot *mem;
 329     int i, err;
 330
 331     s->migration_log = enable;
 332
 333     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 334         mem = &s->slots[i];
 335
 336         if (!mem->memory_size) {
 337             continue;
 338         }
 339         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 340             continue;
 341         }
 342         err = kvm_set_user_memory_region(s, mem);
 343         if (err) {
 344             return err;
 345         }
 346     }
 347     return 0;
 348 }
 349
 350 /* get kvm's dirty pages bitmap and update qemu's */
 351 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
 352                                          unsigned long *bitmap)
 353 {
 354     unsigned int i, j;
 355     unsigned long page_number, c;
 356     target_phys_addr_t addr, addr1;
 357     unsigned int len = ((section->size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
 358     unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
 359
 360     /*
 361      * bitmap-traveling is faster than memory-traveling (for addr...)
 362      * especially when most of the memory is not dirty.
 363      */
 364     for (i = 0; i < len; i++) {
 365         if (bitmap[i] != 0) {
 366             c = leul_to_cpu(bitmap[i]);
 367             do {
 368                 j = ffsl(c) - 1;
 369                 c &= ~(1ul << j);
 370                 page_number = (i * HOST_LONG_BITS + j) * hpratio;
 371                 addr1 = page_number * TARGET_PAGE_SIZE;
 372                 addr = section->offset_within_region + addr1;
 373                 memory_region_set_dirty(section->mr, addr,
 374                                         TARGET_PAGE_SIZE * hpratio);
 375             } while (c != 0);
 376         }
 377     }
 378     return 0;
 379 }
 380
 381 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 382
 383 /**
 384  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 385  * This function updates qemu's dirty bitmap using
 386  * memory_region_set_dirty().  This means all bits are set
 387  * to dirty.
 388  *
 389  * @start_add: start of logged region.
 390  * @end_addr: end of logged region.
 391  */
 392 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
 393 {
 394     KVMState *s = kvm_state;
 395     unsigned long size, allocated_size = 0;
 396     KVMDirtyLog d;
 397     KVMSlot *mem;
 398     int ret = 0;
 399     target_phys_addr_t start_addr = section->offset_within_address_space;
 400     target_phys_addr_t end_addr = start_addr + section->size;
 401
 402     d.dirty_bitmap = NULL;
 403     while (start_addr < end_addr) {
 404         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 405         if (mem == NULL) {
 406             break;
 407         }
 408
 409         /* XXX bad kernel interface alert
 410          * For dirty bitmap, kernel allocates array of size aligned to
 411          * bits-per-long.  But for case when the kernel is 64bits and
 412          * the userspace is 32bits, userspace can't align to the same
 413          * bits-per-long, since sizeof(long) is different between kernel
 414          * and user space.  This way, userspace will provide buffer which
 415          * may be 4 bytes less than the kernel will use, resulting in
 416          * userspace memory corruption (which is not detectable by valgrind
 417          * too, in most cases).
 418          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
 419          * a hope that sizeof(long) wont become >8 any time soon.
 420          */
 421         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
 422                      /*HOST_LONG_BITS*/ 64) / 8;
 423         if (!d.dirty_bitmap) {
 424             d.dirty_bitmap = g_malloc(size);
 425         } else if (size > allocated_size) {
 426             d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
 427         }
 428         allocated_size = size;
 429         memset(d.dirty_bitmap, 0, allocated_size);
 430
 431         d.slot = mem->slot;
 432
 433         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 434             DPRINTF("ioctl failed %d\n", errno);
 435             ret = -1;
 436             break;
 437         }
 438
 439         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
 440         start_addr = mem->start_addr + mem->memory_size;
 441     }
 442     g_free(d.dirty_bitmap);
 443
 444     return ret;
 445 }
 446
 447 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 448 {
 449     int ret = -ENOSYS;
 450     KVMState *s = kvm_state;
 451
 452     if (s->coalesced_mmio) {
 453         struct kvm_coalesced_mmio_zone zone;
 454
 455         zone.addr = start;
 456         zone.size = size;
 457         zone.pad = 0;
 458
 459         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 460     }
 461
 462     return ret;
 463 }
 464
 465 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 466 {
 467     int ret = -ENOSYS;
 468     KVMState *s = kvm_state;
 469
 470     if (s->coalesced_mmio) {
 471         struct kvm_coalesced_mmio_zone zone;
 472
 473         zone.addr = start;
 474         zone.size = size;
 475         zone.pad = 0;
 476
 477         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 478     }
 479
 480     return ret;
 481 }
 482
 483 int kvm_check_extension(KVMState *s, unsigned int extension)
 484 {
 485     int ret;
 486
 487     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 488     if (ret < 0) {
 489         ret = 0;
 490     }
 491
 492     return ret;
 493 }
 494
 495 static int kvm_check_many_ioeventfds(void)
 496 {
 497     /* Userspace can use ioeventfd for io notification.  This requires a host
 498      * that supports eventfd(2) and an I/O thread; since eventfd does not
 499      * support SIGIO it cannot interrupt the vcpu.
 500      *
 501      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 502      * can avoid creating too many ioeventfds.
 503      */
 504 #if defined(CONFIG_EVENTFD)
 505     int ioeventfds[7];
 506     int i, ret = 0;
 507     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 508         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 509         if (ioeventfds[i] < 0) {
 510             break;
 511         }
 512         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 513         if (ret < 0) {
 514             close(ioeventfds[i]);
 515             break;
 516         }
 517     }
 518
 519     /* Decide whether many devices are supported or not */
 520     ret = i == ARRAY_SIZE(ioeventfds);
 521
 522     while (i-- > 0) {
 523         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 524         close(ioeventfds[i]);
 525     }
 526     return ret;
 527 #else
 528     return 0;
 529 #endif
 530 }
 531
 532 static const KVMCapabilityInfo *
 533 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 534 {
 535     while (list->name) {
 536         if (!kvm_check_extension(s, list->value)) {
 537             return list;
 538         }
 539         list++;
 540     }
 541     return NULL;
 542 }
 543
 544 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
 545 {
 546     KVMState *s = kvm_state;
 547     KVMSlot *mem, old;
 548     int err;
 549     MemoryRegion *mr = section->mr;
 550     bool log_dirty = memory_region_is_logging(mr);
 551     target_phys_addr_t start_addr = section->offset_within_address_space;
 552     ram_addr_t size = section->size;
 553     void *ram = NULL;
 554     unsigned delta;
 555
 556     /* kvm works in page size chunks, but the function may be called
 557        with sub-page size and unaligned start address. */
 558     delta = TARGET_PAGE_ALIGN(size) - size;
 559     if (delta > size) {
 560         return;
 561     }
 562     start_addr += delta;
 563     size -= delta;
 564     size &= TARGET_PAGE_MASK;
 565     if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
 566         return;
 567     }
 568
 569     if (!memory_region_is_ram(mr)) {
 570         return;
 571     }
 572
 573     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
 574
 575     while (1) {
 576         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 577         if (!mem) {
 578             break;
 579         }
 580
 581         if (add && start_addr >= mem->start_addr &&
 582             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 583             (ram - start_addr == mem->ram - mem->start_addr)) {
 584             /* The new slot fits into the existing one and comes with
 585              * identical parameters - update flags and done. */
 586             kvm_slot_dirty_pages_log_change(mem, log_dirty);
 587             return;
 588         }
 589
 590         old = *mem;
 591
 592         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 593             kvm_physical_sync_dirty_bitmap(section);
 594         }
 595
 596         /* unregister the overlapping slot */
 597         mem->memory_size = 0;
 598         err = kvm_set_user_memory_region(s, mem);
 599         if (err) {
 600             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 601                     __func__, strerror(-err));
 602             abort();
 603         }
 604
 605         /* Workaround for older KVM versions: we can't join slots, even not by
 606          * unregistering the previous ones and then registering the larger
 607          * slot. We have to maintain the existing fragmentation. Sigh.
 608          *
 609          * This workaround assumes that the new slot starts at the same
 610          * address as the first existing one. If not or if some overlapping
 611          * slot comes around later, we will fail (not seen in practice so far)
 612          * - and actually require a recent KVM version. */
 613         if (s->broken_set_mem_region &&
 614             old.start_addr == start_addr && old.memory_size < size && add) {
 615             mem = kvm_alloc_slot(s);
 616             mem->memory_size = old.memory_size;
 617             mem->start_addr = old.start_addr;
 618             mem->ram = old.ram;
 619             mem->flags = kvm_mem_flags(s, log_dirty);
 620
 621             err = kvm_set_user_memory_region(s, mem);
 622             if (err) {
 623                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 624                         strerror(-err));
 625                 abort();
 626             }
 627
 628             start_addr += old.memory_size;
 629             ram += old.memory_size;
 630             size -= old.memory_size;
 631             continue;
 632         }
 633
 634         /* register prefix slot */
 635         if (old.start_addr < start_addr) {
 636             mem = kvm_alloc_slot(s);
 637             mem->memory_size = start_addr - old.start_addr;
 638             mem->start_addr = old.start_addr;
 639             mem->ram = old.ram;
 640             mem->flags =  kvm_mem_flags(s, log_dirty);
 641
 642             err = kvm_set_user_memory_region(s, mem);
 643             if (err) {
 644                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 645                         __func__, strerror(-err));
 646 #ifdef TARGET_PPC
 647                 fprintf(stderr, "%s: This is probably because your kernel's " \
 648                                 "PAGE_SIZE is too big. Please try to use 4k " \
 649                                 "PAGE_SIZE!\n", __func__);
 650 #endif
 651                 abort();
 652             }
 653         }
 654
 655         /* register suffix slot */
 656         if (old.start_addr + old.memory_size > start_addr + size) {
 657             ram_addr_t size_delta;
 658
 659             mem = kvm_alloc_slot(s);
 660             mem->start_addr = start_addr + size;
 661             size_delta = mem->start_addr - old.start_addr;
 662             mem->memory_size = old.memory_size - size_delta;
 663             mem->ram = old.ram + size_delta;
 664             mem->flags = kvm_mem_flags(s, log_dirty);
 665
 666             err = kvm_set_user_memory_region(s, mem);
 667             if (err) {
 668                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 669                         __func__, strerror(-err));
 670                 abort();
 671             }
 672         }
 673     }
 674
 675     /* in case the KVM bug workaround already "consumed" the new slot */
 676     if (!size) {
 677         return;
 678     }
 679     if (!add) {
 680         return;
 681     }
 682     mem = kvm_alloc_slot(s);
 683     mem->memory_size = size;
 684     mem->start_addr = start_addr;
 685     mem->ram = ram;
 686     mem->flags = kvm_mem_flags(s, log_dirty);
 687
 688     err = kvm_set_user_memory_region(s, mem);
 689     if (err) {
 690         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 691                 strerror(-err));
 692         abort();
 693     }
 694 }
 695
 696 static void kvm_begin(MemoryListener *listener)
 697 {
 698 }
 699
 700 static void kvm_commit(MemoryListener *listener)
 701 {
 702 }
 703
 704 static void kvm_region_add(MemoryListener *listener,
 705                            MemoryRegionSection *section)
 706 {
 707     kvm_set_phys_mem(section, true);
 708 }
 709
 710 static void kvm_region_del(MemoryListener *listener,
 711                            MemoryRegionSection *section)
 712 {
 713     kvm_set_phys_mem(section, false);
 714 }
 715
 716 static void kvm_region_nop(MemoryListener *listener,
 717                            MemoryRegionSection *section)
 718 {
 719 }
 720
 721 static void kvm_log_sync(MemoryListener *listener,
 722                          MemoryRegionSection *section)
 723 {
 724     int r;
 725
 726     r = kvm_physical_sync_dirty_bitmap(section);
 727     if (r < 0) {
 728         abort();
 729     }
 730 }
 731
 732 static void kvm_log_global_start(struct MemoryListener *listener)
 733 {
 734     int r;
 735
 736     r = kvm_set_migration_log(1);
 737     assert(r >= 0);
 738 }
 739
 740 static void kvm_log_global_stop(struct MemoryListener *listener)
 741 {
 742     int r;
 743
 744     r = kvm_set_migration_log(0);
 745     assert(r >= 0);
 746 }
 747
 748 static void kvm_mem_ioeventfd_add(MemoryRegionSection *section,
 749                                   bool match_data, uint64_t data, int fd)
 750 {
 751     int r;
 752
 753     assert(match_data && section->size <= 8);
 754
 755     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 756                                data, true, section->size);
 757     if (r < 0) {
 758         abort();
 759     }
 760 }
 761
 762 static void kvm_mem_ioeventfd_del(MemoryRegionSection *section,
 763                                   bool match_data, uint64_t data, int fd)
 764 {
 765     int r;
 766
 767     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 768                                data, false, section->size);
 769     if (r < 0) {
 770         abort();
 771     }
 772 }
 773
 774 static void kvm_io_ioeventfd_add(MemoryRegionSection *section,
 775                                  bool match_data, uint64_t data, int fd)
 776 {
 777     int r;
 778
 779     assert(match_data && section->size == 2);
 780
 781     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 782                                    data, true);
 783     if (r < 0) {
 784         abort();
 785     }
 786 }
 787
 788 static void kvm_io_ioeventfd_del(MemoryRegionSection *section,
 789                                  bool match_data, uint64_t data, int fd)
 790
 791 {
 792     int r;
 793
 794     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 795                                    data, false);
 796     if (r < 0) {
 797         abort();
 798     }
 799 }
 800
 801 static void kvm_eventfd_add(MemoryListener *listener,
 802                             MemoryRegionSection *section,
 803                             bool match_data, uint64_t data, int fd)
 804 {
 805     if (section->address_space == get_system_memory()) {
 806         kvm_mem_ioeventfd_add(section, match_data, data, fd);
 807     } else {
 808         kvm_io_ioeventfd_add(section, match_data, data, fd);
 809     }
 810 }
 811
 812 static void kvm_eventfd_del(MemoryListener *listener,
 813                             MemoryRegionSection *section,
 814                             bool match_data, uint64_t data, int fd)
 815 {
 816     if (section->address_space == get_system_memory()) {
 817         kvm_mem_ioeventfd_del(section, match_data, data, fd);
 818     } else {
 819         kvm_io_ioeventfd_del(section, match_data, data, fd);
 820     }
 821 }
 822
 823 static MemoryListener kvm_memory_listener = {
 824     .begin = kvm_begin,
 825     .commit = kvm_commit,
 826     .region_add = kvm_region_add,
 827     .region_del = kvm_region_del,
 828     .region_nop = kvm_region_nop,
 829     .log_start = kvm_log_start,
 830     .log_stop = kvm_log_stop,
 831     .log_sync = kvm_log_sync,
 832     .log_global_start = kvm_log_global_start,
 833     .log_global_stop = kvm_log_global_stop,
 834     .eventfd_add = kvm_eventfd_add,
 835     .eventfd_del = kvm_eventfd_del,
 836     .priority = 10,
 837 };
 838
 839 static void kvm_handle_interrupt(CPUArchState *env, int mask)
 840 {
 841     env->interrupt_request |= mask;
 842
 843     if (!qemu_cpu_is_self(env)) {
 844         qemu_cpu_kick(env);
 845     }
 846 }
 847
 848 int kvm_irqchip_set_irq(KVMState *s, int irq, int level)
 849 {
 850     struct kvm_irq_level event;
 851     int ret;
 852
 853     assert(kvm_irqchip_in_kernel());
 854
 855     event.level = level;
 856     event.irq = irq;
 857     ret = kvm_vm_ioctl(s, s->irqchip_inject_ioctl, &event);
 858     if (ret < 0) {
 859         perror("kvm_set_irqchip_line");
 860         abort();
 861     }
 862
 863     return (s->irqchip_inject_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
 864 }
 865
 866 #ifdef KVM_CAP_IRQ_ROUTING
 867 typedef struct KVMMSIRoute {
 868     struct kvm_irq_routing_entry kroute;
 869     QTAILQ_ENTRY(KVMMSIRoute) entry;
 870 } KVMMSIRoute;
 871
 872 static void set_gsi(KVMState *s, unsigned int gsi)
 873 {
 874     s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
 875 }
 876
 877 static void clear_gsi(KVMState *s, unsigned int gsi)
 878 {
 879     s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
 880 }
 881
 882 static void kvm_init_irq_routing(KVMState *s)
 883 {
 884     int gsi_count, i;
 885
 886     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
 887     if (gsi_count > 0) {
 888         unsigned int gsi_bits, i;
 889
 890         /* Round up so we can search ints using ffs */
 891         gsi_bits = ALIGN(gsi_count, 32);
 892         s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
 893         s->gsi_count = gsi_count;
 894
 895         /* Mark any over-allocated bits as already in use */
 896         for (i = gsi_count; i < gsi_bits; i++) {
 897             set_gsi(s, i);
 898         }
 899     }
 900
 901     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
 902     s->nr_allocated_irq_routes = 0;
 903
 904     if (!s->direct_msi) {
 905         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
 906             QTAILQ_INIT(&s->msi_hashtab[i]);
 907         }
 908     }
 909
 910     kvm_arch_init_irq_routing(s);
 911 }
 912
 913 static void kvm_irqchip_commit_routes(KVMState *s)
 914 {
 915     int ret;
 916
 917     s->irq_routes->flags = 0;
 918     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
 919     assert(ret == 0);
 920 }
 921
 922 static void kvm_add_routing_entry(KVMState *s,
 923                                   struct kvm_irq_routing_entry *entry)
 924 {
 925     struct kvm_irq_routing_entry *new;
 926     int n, size;
 927
 928     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
 929         n = s->nr_allocated_irq_routes * 2;
 930         if (n < 64) {
 931             n = 64;
 932         }
 933         size = sizeof(struct kvm_irq_routing);
 934         size += n * sizeof(*new);
 935         s->irq_routes = g_realloc(s->irq_routes, size);
 936         s->nr_allocated_irq_routes = n;
 937     }
 938     n = s->irq_routes->nr++;
 939     new = &s->irq_routes->entries[n];
 940     memset(new, 0, sizeof(*new));
 941     new->gsi = entry->gsi;
 942     new->type = entry->type;
 943     new->flags = entry->flags;
 944     new->u = entry->u;
 945
 946     set_gsi(s, entry->gsi);
 947
 948     kvm_irqchip_commit_routes(s);
 949 }
 950
 951 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
 952 {
 953     struct kvm_irq_routing_entry e;
 954
 955     assert(pin < s->gsi_count);
 956
 957     e.gsi = irq;
 958     e.type = KVM_IRQ_ROUTING_IRQCHIP;
 959     e.flags = 0;
 960     e.u.irqchip.irqchip = irqchip;
 961     e.u.irqchip.pin = pin;
 962     kvm_add_routing_entry(s, &e);
 963 }
 964
 965 void kvm_irqchip_release_virq(KVMState *s, int virq)
 966 {
 967     struct kvm_irq_routing_entry *e;
 968     int i;
 969
 970     for (i = 0; i < s->irq_routes->nr; i++) {
 971         e = &s->irq_routes->entries[i];
 972         if (e->gsi == virq) {
 973             s->irq_routes->nr--;
 974             *e = s->irq_routes->entries[s->irq_routes->nr];
 975         }
 976     }
 977     clear_gsi(s, virq);
 978
 979     kvm_irqchip_commit_routes(s);
 980 }
 981
 982 static unsigned int kvm_hash_msi(uint32_t data)
 983 {
 984     /* This is optimized for IA32 MSI layout. However, no other arch shall
 985      * repeat the mistake of not providing a direct MSI injection API. */
 986     return data & 0xff;
 987 }
 988
 989 static void kvm_flush_dynamic_msi_routes(KVMState *s)
 990 {
 991     KVMMSIRoute *route, *next;
 992     unsigned int hash;
 993
 994     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
 995         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
 996             kvm_irqchip_release_virq(s, route->kroute.gsi);
 997             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
 998             g_free(route);
 999         }
1000     }
1001 }
1002
1003 static int kvm_irqchip_get_virq(KVMState *s)
1004 {
1005     uint32_t *word = s->used_gsi_bitmap;
1006     int max_words = ALIGN(s->gsi_count, 32) / 32;
1007     int i, bit;
1008     bool retry = true;
1009
1010 again:
1011     /* Return the lowest unused GSI in the bitmap */
1012     for (i = 0; i < max_words; i++) {
1013         bit = ffs(~word[i]);
1014         if (!bit) {
1015             continue;
1016         }
1017
1018         return bit - 1 + i * 32;
1019     }
1020     if (!s->direct_msi && retry) {
1021         retry = false;
1022         kvm_flush_dynamic_msi_routes(s);
1023         goto again;
1024     }
1025     return -ENOSPC;
1026
1027 }
1028
1029 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1030 {
1031     unsigned int hash = kvm_hash_msi(msg.data);
1032     KVMMSIRoute *route;
1033
1034     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1035         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1036             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1037             route->kroute.u.msi.data == msg.data) {
1038             return route;
1039         }
1040     }
1041     return NULL;
1042 }
1043
1044 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1045 {
1046     struct kvm_msi msi;
1047     KVMMSIRoute *route;
1048
1049     if (s->direct_msi) {
1050         msi.address_lo = (uint32_t)msg.address;
1051         msi.address_hi = msg.address >> 32;
1052         msi.data = msg.data;
1053         msi.flags = 0;
1054         memset(msi.pad, 0, sizeof(msi.pad));
1055
1056         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1057     }
1058
1059     route = kvm_lookup_msi_route(s, msg);
1060     if (!route) {
1061         int virq;
1062
1063         virq = kvm_irqchip_get_virq(s);
1064         if (virq < 0) {
1065             return virq;
1066         }
1067
1068         route = g_malloc(sizeof(KVMMSIRoute));
1069         route->kroute.gsi = virq;
1070         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1071         route->kroute.flags = 0;
1072         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1073         route->kroute.u.msi.address_hi = msg.address >> 32;
1074         route->kroute.u.msi.data = msg.data;
1075
1076         kvm_add_routing_entry(s, &route->kroute);
1077
1078         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1079                            entry);
1080     }
1081
1082     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1083
1084     return kvm_irqchip_set_irq(s, route->kroute.gsi, 1);
1085 }
1086
1087 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1088 {
1089     struct kvm_irq_routing_entry kroute;
1090     int virq;
1091
1092     if (!kvm_irqchip_in_kernel()) {
1093         return -ENOSYS;
1094     }
1095
1096     virq = kvm_irqchip_get_virq(s);
1097     if (virq < 0) {
1098         return virq;
1099     }
1100
1101     kroute.gsi = virq;
1102     kroute.type = KVM_IRQ_ROUTING_MSI;
1103     kroute.flags = 0;
1104     kroute.u.msi.address_lo = (uint32_t)msg.address;
1105     kroute.u.msi.address_hi = msg.address >> 32;
1106     kroute.u.msi.data = msg.data;
1107
1108     kvm_add_routing_entry(s, &kroute);
1109
1110     return virq;
1111 }
1112
1113 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1114 {
1115     struct kvm_irqfd irqfd = {
1116         .fd = fd,
1117         .gsi = virq,
1118         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1119     };
1120
1121     if (!kvm_irqchip_in_kernel()) {
1122         return -ENOSYS;
1123     }
1124
1125     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1126 }
1127
1128 #else /* !KVM_CAP_IRQ_ROUTING */
1129
1130 static void kvm_init_irq_routing(KVMState *s)
1131 {
1132 }
1133
1134 void kvm_irqchip_release_virq(KVMState *s, int virq)
1135 {
1136 }
1137
1138 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1139 {
1140     abort();
1141 }
1142
1143 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1144 {
1145     abort();
1146 }
1147
1148 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1149 {
1150     abort();
1151 }
1152 #endif /* !KVM_CAP_IRQ_ROUTING */
1153
1154 int kvm_irqchip_add_irqfd(KVMState *s, int fd, int virq)
1155 {
1156     return kvm_irqchip_assign_irqfd(s, fd, virq, true);
1157 }
1158
1159 int kvm_irqchip_remove_irqfd(KVMState *s, int fd, int virq)
1160 {
1161     return kvm_irqchip_assign_irqfd(s, fd, virq, false);
1162 }
1163
1164 static int kvm_irqchip_create(KVMState *s)
1165 {
1166     QemuOptsList *list = qemu_find_opts("machine");
1167     int ret;
1168
1169     if (QTAILQ_EMPTY(&list->head) ||
1170         !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1171                            "kernel_irqchip", true) ||
1172         !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1173         return 0;
1174     }
1175
1176     ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1177     if (ret < 0) {
1178         fprintf(stderr, "Create kernel irqchip failed\n");
1179         return ret;
1180     }
1181
1182     s->irqchip_inject_ioctl = KVM_IRQ_LINE;
1183     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1184         s->irqchip_inject_ioctl = KVM_IRQ_LINE_STATUS;
1185     }
1186     kvm_kernel_irqchip = true;
1187
1188     kvm_init_irq_routing(s);
1189
1190     return 0;
1191 }
1192
1193 int kvm_init(void)
1194 {
1195     static const char upgrade_note[] =
1196         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1197         "(see http://sourceforge.net/projects/kvm).\n";
1198     KVMState *s;
1199     const KVMCapabilityInfo *missing_cap;
1200     int ret;
1201     int i;
1202
1203     s = g_malloc0(sizeof(KVMState));
1204
1205     /*
1206      * On systems where the kernel can support different base page
1207      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1208      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1209      * page size for the system though.
1210      */
1211     assert(TARGET_PAGE_SIZE <= getpagesize());
1212
1213 #ifdef KVM_CAP_SET_GUEST_DEBUG
1214     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1215 #endif
1216     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1217         s->slots[i].slot = i;
1218     }
1219     s->vmfd = -1;
1220     s->fd = qemu_open("/dev/kvm", O_RDWR);
1221     if (s->fd == -1) {
1222         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1223         ret = -errno;
1224         goto err;
1225     }
1226
1227     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1228     if (ret < KVM_API_VERSION) {
1229         if (ret > 0) {
1230             ret = -EINVAL;
1231         }
1232         fprintf(stderr, "kvm version too old\n");
1233         goto err;
1234     }
1235
1236     if (ret > KVM_API_VERSION) {
1237         ret = -EINVAL;
1238         fprintf(stderr, "kvm version not supported\n");
1239         goto err;
1240     }
1241
1242     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1243     if (s->vmfd < 0) {
1244 #ifdef TARGET_S390X
1245         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1246                         "your host kernel command line\n");
1247 #endif
1248         ret = s->vmfd;
1249         goto err;
1250     }
1251
1252     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1253     if (!missing_cap) {
1254         missing_cap =
1255             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1256     }
1257     if (missing_cap) {
1258         ret = -EINVAL;
1259         fprintf(stderr, "kvm does not support %s\n%s",
1260                 missing_cap->name, upgrade_note);
1261         goto err;
1262     }
1263
1264     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1265
1266     s->broken_set_mem_region = 1;
1267     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1268     if (ret > 0) {
1269         s->broken_set_mem_region = 0;
1270     }
1271
1272 #ifdef KVM_CAP_VCPU_EVENTS
1273     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1274 #endif
1275
1276     s->robust_singlestep =
1277         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1278
1279 #ifdef KVM_CAP_DEBUGREGS
1280     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1281 #endif
1282
1283 #ifdef KVM_CAP_XSAVE
1284     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1285 #endif
1286
1287 #ifdef KVM_CAP_XCRS
1288     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1289 #endif
1290
1291 #ifdef KVM_CAP_PIT_STATE2
1292     s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1293 #endif
1294
1295 #ifdef KVM_CAP_IRQ_ROUTING
1296     s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1297 #endif
1298
1299     ret = kvm_arch_init(s);
1300     if (ret < 0) {
1301         goto err;
1302     }
1303
1304     ret = kvm_irqchip_create(s);
1305     if (ret < 0) {
1306         goto err;
1307     }
1308
1309     kvm_state = s;
1310     memory_listener_register(&kvm_memory_listener, NULL);
1311
1312     s->many_ioeventfds = kvm_check_many_ioeventfds();
1313
1314     cpu_interrupt_handler = kvm_handle_interrupt;
1315
1316     return 0;
1317
1318 err:
1319     if (s) {
1320         if (s->vmfd >= 0) {
1321             close(s->vmfd);
1322         }
1323         if (s->fd != -1) {
1324             close(s->fd);
1325         }
1326     }
1327     g_free(s);
1328
1329     return ret;
1330 }
1331
1332 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1333                           uint32_t count)
1334 {
1335     int i;
1336     uint8_t *ptr = data;
1337
1338     for (i = 0; i < count; i++) {
1339         if (direction == KVM_EXIT_IO_IN) {
1340             switch (size) {
1341             case 1:
1342                 stb_p(ptr, cpu_inb(port));
1343                 break;
1344             case 2:
1345                 stw_p(ptr, cpu_inw(port));
1346                 break;
1347             case 4:
1348                 stl_p(ptr, cpu_inl(port));
1349                 break;
1350             }
1351         } else {
1352             switch (size) {
1353             case 1:
1354                 cpu_outb(port, ldub_p(ptr));
1355                 break;
1356             case 2:
1357                 cpu_outw(port, lduw_p(ptr));
1358                 break;
1359             case 4:
1360                 cpu_outl(port, ldl_p(ptr));
1361                 break;
1362             }
1363         }
1364
1365         ptr += size;
1366     }
1367 }
1368
1369 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1370 {
1371     fprintf(stderr, "KVM internal error.");
1372     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1373         int i;
1374
1375         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1376         for (i = 0; i < run->internal.ndata; ++i) {
1377             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1378                     i, (uint64_t)run->internal.data[i]);
1379         }
1380     } else {
1381         fprintf(stderr, "\n");
1382     }
1383     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1384         fprintf(stderr, "emulation failure\n");
1385         if (!kvm_arch_stop_on_emulation_error(env)) {
1386             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1387             return EXCP_INTERRUPT;
1388         }
1389     }
1390     /* FIXME: Should trigger a qmp message to let management know
1391      * something went wrong.
1392      */
1393     return -1;
1394 }
1395
1396 void kvm_flush_coalesced_mmio_buffer(void)
1397 {
1398     KVMState *s = kvm_state;
1399
1400     if (s->coalesced_flush_in_progress) {
1401         return;
1402     }
1403
1404     s->coalesced_flush_in_progress = true;
1405
1406     if (s->coalesced_mmio_ring) {
1407         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1408         while (ring->first != ring->last) {
1409             struct kvm_coalesced_mmio *ent;
1410
1411             ent = &ring->coalesced_mmio[ring->first];
1412
1413             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1414             smp_wmb();
1415             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1416         }
1417     }
1418
1419     s->coalesced_flush_in_progress = false;
1420 }
1421
1422 static void do_kvm_cpu_synchronize_state(void *_env)
1423 {
1424     CPUArchState *env = _env;
1425
1426     if (!env->kvm_vcpu_dirty) {
1427         kvm_arch_get_registers(env);
1428         env->kvm_vcpu_dirty = 1;
1429     }
1430 }
1431
1432 void kvm_cpu_synchronize_state(CPUArchState *env)
1433 {
1434     if (!env->kvm_vcpu_dirty) {
1435         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
1436     }
1437 }
1438
1439 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1440 {
1441     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
1442     env->kvm_vcpu_dirty = 0;
1443 }
1444
1445 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1446 {
1447     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
1448     env->kvm_vcpu_dirty = 0;
1449 }
1450
1451 int kvm_cpu_exec(CPUArchState *env)
1452 {
1453     struct kvm_run *run = env->kvm_run;
1454     int ret, run_ret;
1455
1456     DPRINTF("kvm_cpu_exec()\n");
1457
1458     if (kvm_arch_process_async_events(env)) {
1459         env->exit_request = 0;
1460         return EXCP_HLT;
1461     }
1462
1463     do {
1464         if (env->kvm_vcpu_dirty) {
1465             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
1466             env->kvm_vcpu_dirty = 0;
1467         }
1468
1469         kvm_arch_pre_run(env, run);
1470         if (env->exit_request) {
1471             DPRINTF("interrupt exit requested\n");
1472             /*
1473              * KVM requires us to reenter the kernel after IO exits to complete
1474              * instruction emulation. This self-signal will ensure that we
1475              * leave ASAP again.
1476              */
1477             qemu_cpu_kick_self();
1478         }
1479         qemu_mutex_unlock_iothread();
1480
1481         run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
1482
1483         qemu_mutex_lock_iothread();
1484         kvm_arch_post_run(env, run);
1485
1486         kvm_flush_coalesced_mmio_buffer();
1487
1488         if (run_ret < 0) {
1489             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1490                 DPRINTF("io window exit\n");
1491                 ret = EXCP_INTERRUPT;
1492                 break;
1493             }
1494             fprintf(stderr, "error: kvm run failed %s\n",
1495                     strerror(-run_ret));
1496             abort();
1497         }
1498
1499         switch (run->exit_reason) {
1500         case KVM_EXIT_IO:
1501             DPRINTF("handle_io\n");
1502             kvm_handle_io(run->io.port,
1503                           (uint8_t *)run + run->io.data_offset,
1504                           run->io.direction,
1505                           run->io.size,
1506                           run->io.count);
1507             ret = 0;
1508             break;
1509         case KVM_EXIT_MMIO:
1510             DPRINTF("handle_mmio\n");
1511             cpu_physical_memory_rw(run->mmio.phys_addr,
1512                                    run->mmio.data,
1513                                    run->mmio.len,
1514                                    run->mmio.is_write);
1515             ret = 0;
1516             break;
1517         case KVM_EXIT_IRQ_WINDOW_OPEN:
1518             DPRINTF("irq_window_open\n");
1519             ret = EXCP_INTERRUPT;
1520             break;
1521         case KVM_EXIT_SHUTDOWN:
1522             DPRINTF("shutdown\n");
1523             qemu_system_reset_request();
1524             ret = EXCP_INTERRUPT;
1525             break;
1526         case KVM_EXIT_UNKNOWN:
1527             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1528                     (uint64_t)run->hw.hardware_exit_reason);
1529             ret = -1;
1530             break;
1531         case KVM_EXIT_INTERNAL_ERROR:
1532             ret = kvm_handle_internal_error(env, run);
1533             break;
1534         default:
1535             DPRINTF("kvm_arch_handle_exit\n");
1536             ret = kvm_arch_handle_exit(env, run);
1537             break;
1538         }
1539     } while (ret == 0);
1540
1541     if (ret < 0) {
1542         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1543         vm_stop(RUN_STATE_INTERNAL_ERROR);
1544     }
1545
1546     env->exit_request = 0;
1547     return ret;
1548 }
1549
1550 int kvm_ioctl(KVMState *s, int type, ...)
1551 {
1552     int ret;
1553     void *arg;
1554     va_list ap;
1555
1556     va_start(ap, type);
1557     arg = va_arg(ap, void *);
1558     va_end(ap);
1559
1560     ret = ioctl(s->fd, type, arg);
1561     if (ret == -1) {
1562         ret = -errno;
1563     }
1564     return ret;
1565 }
1566
1567 int kvm_vm_ioctl(KVMState *s, int type, ...)
1568 {
1569     int ret;
1570     void *arg;
1571     va_list ap;
1572
1573     va_start(ap, type);
1574     arg = va_arg(ap, void *);
1575     va_end(ap);
1576
1577     ret = ioctl(s->vmfd, type, arg);
1578     if (ret == -1) {
1579         ret = -errno;
1580     }
1581     return ret;
1582 }
1583
1584 int kvm_vcpu_ioctl(CPUArchState *env, int type, ...)
1585 {
1586     int ret;
1587     void *arg;
1588     va_list ap;
1589
1590     va_start(ap, type);
1591     arg = va_arg(ap, void *);
1592     va_end(ap);
1593
1594     ret = ioctl(env->kvm_fd, type, arg);
1595     if (ret == -1) {
1596         ret = -errno;
1597     }
1598     return ret;
1599 }
1600
1601 int kvm_has_sync_mmu(void)
1602 {
1603     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1604 }
1605
1606 int kvm_has_vcpu_events(void)
1607 {
1608     return kvm_state->vcpu_events;
1609 }
1610
1611 int kvm_has_robust_singlestep(void)
1612 {
1613     return kvm_state->robust_singlestep;
1614 }
1615
1616 int kvm_has_debugregs(void)
1617 {
1618     return kvm_state->debugregs;
1619 }
1620
1621 int kvm_has_xsave(void)
1622 {
1623     return kvm_state->xsave;
1624 }
1625
1626 int kvm_has_xcrs(void)
1627 {
1628     return kvm_state->xcrs;
1629 }
1630
1631 int kvm_has_pit_state2(void)
1632 {
1633     return kvm_state->pit_state2;
1634 }
1635
1636 int kvm_has_many_ioeventfds(void)
1637 {
1638     if (!kvm_enabled()) {
1639         return 0;
1640     }
1641     return kvm_state->many_ioeventfds;
1642 }
1643
1644 int kvm_has_gsi_routing(void)
1645 {
1646 #ifdef KVM_CAP_IRQ_ROUTING
1647     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1648 #else
1649     return false;
1650 #endif
1651 }
1652
1653 int kvm_allows_irq0_override(void)
1654 {
1655     return !kvm_irqchip_in_kernel() || kvm_has_gsi_routing();
1656 }
1657
1658 void *kvm_vmalloc(ram_addr_t size)
1659 {
1660 #ifdef TARGET_S390X
1661     void *mem;
1662
1663     mem = kvm_arch_vmalloc(size);
1664     if (mem) {
1665         return mem;
1666     }
1667 #endif
1668     return qemu_vmalloc(size);
1669 }
1670
1671 void kvm_setup_guest_memory(void *start, size_t size)
1672 {
1673     if (!kvm_has_sync_mmu()) {
1674         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1675
1676         if (ret) {
1677             perror("qemu_madvise");
1678             fprintf(stderr,
1679                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1680             exit(1);
1681         }
1682     }
1683 }
1684
1685 #ifdef KVM_CAP_SET_GUEST_DEBUG
1686 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUArchState *env,
1687                                                  target_ulong pc)
1688 {
1689     struct kvm_sw_breakpoint *bp;
1690
1691     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1692         if (bp->pc == pc) {
1693             return bp;
1694         }
1695     }
1696     return NULL;
1697 }
1698
1699 int kvm_sw_breakpoints_active(CPUArchState *env)
1700 {
1701     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1702 }
1703
1704 struct kvm_set_guest_debug_data {
1705     struct kvm_guest_debug dbg;
1706     CPUArchState *env;
1707     int err;
1708 };
1709
1710 static void kvm_invoke_set_guest_debug(void *data)
1711 {
1712     struct kvm_set_guest_debug_data *dbg_data = data;
1713     CPUArchState *env = dbg_data->env;
1714
1715     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1716 }
1717
1718 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1719 {
1720     struct kvm_set_guest_debug_data data;
1721
1722     data.dbg.control = reinject_trap;
1723
1724     if (env->singlestep_enabled) {
1725         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1726     }
1727     kvm_arch_update_guest_debug(env, &data.dbg);
1728     data.env = env;
1729
1730     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1731     return data.err;
1732 }
1733
1734 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1735                           target_ulong len, int type)
1736 {
1737     struct kvm_sw_breakpoint *bp;
1738     CPUArchState *env;
1739     int err;
1740
1741     if (type == GDB_BREAKPOINT_SW) {
1742         bp = kvm_find_sw_breakpoint(current_env, addr);
1743         if (bp) {
1744             bp->use_count++;
1745             return 0;
1746         }
1747
1748         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1749         if (!bp) {
1750             return -ENOMEM;
1751         }
1752
1753         bp->pc = addr;
1754         bp->use_count = 1;
1755         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1756         if (err) {
1757             g_free(bp);
1758             return err;
1759         }
1760
1761         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1762                           bp, entry);
1763     } else {
1764         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1765         if (err) {
1766             return err;
1767         }
1768     }
1769
1770     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1771         err = kvm_update_guest_debug(env, 0);
1772         if (err) {
1773             return err;
1774         }
1775     }
1776     return 0;
1777 }
1778
1779 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1780                           target_ulong len, int type)
1781 {
1782     struct kvm_sw_breakpoint *bp;
1783     CPUArchState *env;
1784     int err;
1785
1786     if (type == GDB_BREAKPOINT_SW) {
1787         bp = kvm_find_sw_breakpoint(current_env, addr);
1788         if (!bp) {
1789             return -ENOENT;
1790         }
1791
1792         if (bp->use_count > 1) {
1793             bp->use_count--;
1794             return 0;
1795         }
1796
1797         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1798         if (err) {
1799             return err;
1800         }
1801
1802         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1803         g_free(bp);
1804     } else {
1805         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1806         if (err) {
1807             return err;
1808         }
1809     }
1810
1811     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1812         err = kvm_update_guest_debug(env, 0);
1813         if (err) {
1814             return err;
1815         }
1816     }
1817     return 0;
1818 }
1819
1820 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1821 {
1822     struct kvm_sw_breakpoint *bp, *next;
1823     KVMState *s = current_env->kvm_state;
1824     CPUArchState *env;
1825
1826     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1827         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1828             /* Try harder to find a CPU that currently sees the breakpoint. */
1829             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1830                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1831                     break;
1832                 }
1833             }
1834         }
1835     }
1836     kvm_arch_remove_all_hw_breakpoints();
1837
1838     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1839         kvm_update_guest_debug(env, 0);
1840     }
1841 }
1842
1843 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1844
1845 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1846 {
1847     return -EINVAL;
1848 }
1849
1850 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1851                           target_ulong len, int type)
1852 {
1853     return -EINVAL;
1854 }
1855
1856 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1857                           target_ulong len, int type)
1858 {
1859     return -EINVAL;
1860 }
1861
1862 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1863 {
1864 }
1865 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1866
1867 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1868 {
1869     struct kvm_signal_mask *sigmask;
1870     int r;
1871
1872     if (!sigset) {
1873         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1874     }
1875
1876     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1877
1878     sigmask->len = 8;
1879     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1880     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1881     g_free(sigmask);
1882
1883     return r;
1884 }
1885
1886 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1887                            uint32_t size)
1888 {
1889     int ret;
1890     struct kvm_ioeventfd iofd;
1891
1892     iofd.datamatch = val;
1893     iofd.addr = addr;
1894     iofd.len = size;
1895     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1896     iofd.fd = fd;
1897
1898     if (!kvm_enabled()) {
1899         return -ENOSYS;
1900     }
1901
1902     if (!assign) {
1903         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1904     }
1905
1906     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1907
1908     if (ret < 0) {
1909         return -errno;
1910     }
1911
1912     return 0;
1913 }
1914
1915 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1916 {
1917     struct kvm_ioeventfd kick = {
1918         .datamatch = val,
1919         .addr = addr,
1920         .len = 2,
1921         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1922         .fd = fd,
1923     };
1924     int r;
1925     if (!kvm_enabled()) {
1926         return -ENOSYS;
1927     }
1928     if (!assign) {
1929         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1930     }
1931     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1932     if (r < 0) {
1933         return r;
1934     }
1935     return 0;
1936 }
1937
1938 int kvm_on_sigbus_vcpu(CPUArchState *env, int code, void *addr)
1939 {
1940     return kvm_arch_on_sigbus_vcpu(env, code, addr);
1941 }
1942
1943 int kvm_on_sigbus(int code, void *addr)
1944 {
1945     return kvm_arch_on_sigbus(code, addr);
1946 }