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/atomic.h"
  25 #include "qemu/option.h"
  26 #include "qemu/config-file.h"
  27 #include "sysemu/sysemu.h"
  28 #include "hw/hw.h"
  29 #include "hw/pci/msi.h"
  30 #include "exec/gdbstub.h"
  31 #include "sysemu/kvm.h"
  32 #include "qemu/bswap.h"
  33 #include "exec/memory.h"
  34 #include "exec/address-spaces.h"
  35 #include "qemu/event_notifier.h"
  36
  37 /* This check must be after config-host.h is included */
  38 #ifdef CONFIG_EVENTFD
  39 #include <sys/eventfd.h>
  40 #endif
  41
  42 #ifdef CONFIG_VALGRIND_H
  43 #include <valgrind/memcheck.h>
  44 #endif
  45
  46 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
  47 #define PAGE_SIZE TARGET_PAGE_SIZE
  48
  49 //#define DEBUG_KVM
  50
  51 #ifdef DEBUG_KVM
  52 #define DPRINTF(fmt, ...) \
  53     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  54 #else
  55 #define DPRINTF(fmt, ...) \
  56     do { } while (0)
  57 #endif
  58
  59 #define KVM_MSI_HASHTAB_SIZE    256
  60
  61 typedef struct KVMSlot
  62 {
  63     hwaddr start_addr;
  64     ram_addr_t memory_size;
  65     void *ram;
  66     int slot;
  67     int flags;
  68 } KVMSlot;
  69
  70 typedef struct kvm_dirty_log KVMDirtyLog;
  71
  72 struct KVMState
  73 {
  74     KVMSlot slots[32];
  75     int fd;
  76     int vmfd;
  77     int coalesced_mmio;
  78     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  79     bool coalesced_flush_in_progress;
  80     int broken_set_mem_region;
  81     int migration_log;
  82     int vcpu_events;
  83     int robust_singlestep;
  84     int debugregs;
  85 #ifdef KVM_CAP_SET_GUEST_DEBUG
  86     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  87 #endif
  88     int pit_state2;
  89     int xsave, xcrs;
  90     int many_ioeventfds;
  91     int intx_set_mask;
  92     /* The man page (and posix) say ioctl numbers are signed int, but
  93      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
  94      * unsigned, and treating them as signed here can break things */
  95     unsigned irq_set_ioctl;
  96 #ifdef KVM_CAP_IRQ_ROUTING
  97     struct kvm_irq_routing *irq_routes;
  98     int nr_allocated_irq_routes;
  99     uint32_t *used_gsi_bitmap;
 100     unsigned int gsi_count;
 101     QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
 102     bool direct_msi;
 103 #endif
 104 };
 105
 106 KVMState *kvm_state;
 107 bool kvm_kernel_irqchip;
 108 bool kvm_async_interrupts_allowed;
 109 bool kvm_irqfds_allowed;
 110 bool kvm_msi_via_irqfd_allowed;
 111 bool kvm_gsi_routing_allowed;
 112
 113 static const KVMCapabilityInfo kvm_required_capabilites[] = {
 114     KVM_CAP_INFO(USER_MEMORY),
 115     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
 116     KVM_CAP_LAST_INFO
 117 };
 118
 119 static KVMSlot *kvm_alloc_slot(KVMState *s)
 120 {
 121     int i;
 122
 123     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 124         if (s->slots[i].memory_size == 0) {
 125             return &s->slots[i];
 126         }
 127     }
 128
 129     fprintf(stderr, "%s: no free slot available\n", __func__);
 130     abort();
 131 }
 132
 133 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 134                                          hwaddr start_addr,
 135                                          hwaddr end_addr)
 136 {
 137     int i;
 138
 139     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 140         KVMSlot *mem = &s->slots[i];
 141
 142         if (start_addr == mem->start_addr &&
 143             end_addr == mem->start_addr + mem->memory_size) {
 144             return mem;
 145         }
 146     }
 147
 148     return NULL;
 149 }
 150
 151 /*
 152  * Find overlapping slot with lowest start address
 153  */
 154 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 155                                             hwaddr start_addr,
 156                                             hwaddr end_addr)
 157 {
 158     KVMSlot *found = NULL;
 159     int i;
 160
 161     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 162         KVMSlot *mem = &s->slots[i];
 163
 164         if (mem->memory_size == 0 ||
 165             (found && found->start_addr < mem->start_addr)) {
 166             continue;
 167         }
 168
 169         if (end_addr > mem->start_addr &&
 170             start_addr < mem->start_addr + mem->memory_size) {
 171             found = mem;
 172         }
 173     }
 174
 175     return found;
 176 }
 177
 178 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
 179                                        hwaddr *phys_addr)
 180 {
 181     int i;
 182
 183     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 184         KVMSlot *mem = &s->slots[i];
 185
 186         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
 187             *phys_addr = mem->start_addr + (ram - mem->ram);
 188             return 1;
 189         }
 190     }
 191
 192     return 0;
 193 }
 194
 195 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 196 {
 197     struct kvm_userspace_memory_region mem;
 198
 199     mem.slot = slot->slot;
 200     mem.guest_phys_addr = slot->start_addr;
 201     mem.memory_size = slot->memory_size;
 202     mem.userspace_addr = (unsigned long)slot->ram;
 203     mem.flags = slot->flags;
 204     if (s->migration_log) {
 205         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 206     }
 207     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 208 }
 209
 210 static void kvm_reset_vcpu(void *opaque)
 211 {
 212     CPUState *cpu = opaque;
 213
 214     kvm_arch_reset_vcpu(cpu);
 215 }
 216
 217 int kvm_init_vcpu(CPUState *cpu)
 218 {
 219     KVMState *s = kvm_state;
 220     long mmap_size;
 221     int ret;
 222
 223     DPRINTF("kvm_init_vcpu\n");
 224
 225     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
 226     if (ret < 0) {
 227         DPRINTF("kvm_create_vcpu failed\n");
 228         goto err;
 229     }
 230
 231     cpu->kvm_fd = ret;
 232     cpu->kvm_state = s;
 233     cpu->kvm_vcpu_dirty = true;
 234
 235     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 236     if (mmap_size < 0) {
 237         ret = mmap_size;
 238         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 239         goto err;
 240     }
 241
 242     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 243                         cpu->kvm_fd, 0);
 244     if (cpu->kvm_run == MAP_FAILED) {
 245         ret = -errno;
 246         DPRINTF("mmap'ing vcpu state failed\n");
 247         goto err;
 248     }
 249
 250     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 251         s->coalesced_mmio_ring =
 252             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 253     }
 254
 255     ret = kvm_arch_init_vcpu(cpu);
 256     if (ret == 0) {
 257         qemu_register_reset(kvm_reset_vcpu, cpu);
 258         kvm_arch_reset_vcpu(cpu);
 259     }
 260 err:
 261     return ret;
 262 }
 263
 264 /*
 265  * dirty pages logging control
 266  */
 267
 268 static int kvm_mem_flags(KVMState *s, bool log_dirty)
 269 {
 270     return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
 271 }
 272
 273 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
 274 {
 275     KVMState *s = kvm_state;
 276     int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
 277     int old_flags;
 278
 279     old_flags = mem->flags;
 280
 281     flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
 282     mem->flags = flags;
 283
 284     /* If nothing changed effectively, no need to issue ioctl */
 285     if (s->migration_log) {
 286         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 287     }
 288
 289     if (flags == old_flags) {
 290         return 0;
 291     }
 292
 293     return kvm_set_user_memory_region(s, mem);
 294 }
 295
 296 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
 297                                       ram_addr_t size, bool log_dirty)
 298 {
 299     KVMState *s = kvm_state;
 300     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 301
 302     if (mem == NULL)  {
 303         fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 304                 TARGET_FMT_plx "\n", __func__, phys_addr,
 305                 (hwaddr)(phys_addr + size - 1));
 306         return -EINVAL;
 307     }
 308     return kvm_slot_dirty_pages_log_change(mem, log_dirty);
 309 }
 310
 311 static void kvm_log_start(MemoryListener *listener,
 312                           MemoryRegionSection *section)
 313 {
 314     int r;
 315
 316     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 317                                    section->size, true);
 318     if (r < 0) {
 319         abort();
 320     }
 321 }
 322
 323 static void kvm_log_stop(MemoryListener *listener,
 324                           MemoryRegionSection *section)
 325 {
 326     int r;
 327
 328     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 329                                    section->size, false);
 330     if (r < 0) {
 331         abort();
 332     }
 333 }
 334
 335 static int kvm_set_migration_log(int enable)
 336 {
 337     KVMState *s = kvm_state;
 338     KVMSlot *mem;
 339     int i, err;
 340
 341     s->migration_log = enable;
 342
 343     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 344         mem = &s->slots[i];
 345
 346         if (!mem->memory_size) {
 347             continue;
 348         }
 349         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 350             continue;
 351         }
 352         err = kvm_set_user_memory_region(s, mem);
 353         if (err) {
 354             return err;
 355         }
 356     }
 357     return 0;
 358 }
 359
 360 /* get kvm's dirty pages bitmap and update qemu's */
 361 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
 362                                          unsigned long *bitmap)
 363 {
 364     unsigned int i, j;
 365     unsigned long page_number, c;
 366     hwaddr addr, addr1;
 367     unsigned int len = ((section->size / getpagesize()) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
 368     unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
 369
 370     /*
 371      * bitmap-traveling is faster than memory-traveling (for addr...)
 372      * especially when most of the memory is not dirty.
 373      */
 374     for (i = 0; i < len; i++) {
 375         if (bitmap[i] != 0) {
 376             c = leul_to_cpu(bitmap[i]);
 377             do {
 378                 j = ffsl(c) - 1;
 379                 c &= ~(1ul << j);
 380                 page_number = (i * HOST_LONG_BITS + j) * hpratio;
 381                 addr1 = page_number * TARGET_PAGE_SIZE;
 382                 addr = section->offset_within_region + addr1;
 383                 memory_region_set_dirty(section->mr, addr,
 384                                         TARGET_PAGE_SIZE * hpratio);
 385             } while (c != 0);
 386         }
 387     }
 388     return 0;
 389 }
 390
 391 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 392
 393 /**
 394  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 395  * This function updates qemu's dirty bitmap using
 396  * memory_region_set_dirty().  This means all bits are set
 397  * to dirty.
 398  *
 399  * @start_add: start of logged region.
 400  * @end_addr: end of logged region.
 401  */
 402 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
 403 {
 404     KVMState *s = kvm_state;
 405     unsigned long size, allocated_size = 0;
 406     KVMDirtyLog d;
 407     KVMSlot *mem;
 408     int ret = 0;
 409     hwaddr start_addr = section->offset_within_address_space;
 410     hwaddr end_addr = start_addr + section->size;
 411
 412     d.dirty_bitmap = NULL;
 413     while (start_addr < end_addr) {
 414         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 415         if (mem == NULL) {
 416             break;
 417         }
 418
 419         /* XXX bad kernel interface alert
 420          * For dirty bitmap, kernel allocates array of size aligned to
 421          * bits-per-long.  But for case when the kernel is 64bits and
 422          * the userspace is 32bits, userspace can't align to the same
 423          * bits-per-long, since sizeof(long) is different between kernel
 424          * and user space.  This way, userspace will provide buffer which
 425          * may be 4 bytes less than the kernel will use, resulting in
 426          * userspace memory corruption (which is not detectable by valgrind
 427          * too, in most cases).
 428          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
 429          * a hope that sizeof(long) wont become >8 any time soon.
 430          */
 431         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
 432                      /*HOST_LONG_BITS*/ 64) / 8;
 433         if (!d.dirty_bitmap) {
 434             d.dirty_bitmap = g_malloc(size);
 435         } else if (size > allocated_size) {
 436             d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
 437         }
 438         allocated_size = size;
 439         memset(d.dirty_bitmap, 0, allocated_size);
 440
 441         d.slot = mem->slot;
 442
 443         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 444             DPRINTF("ioctl failed %d\n", errno);
 445             ret = -1;
 446             break;
 447         }
 448
 449         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
 450         start_addr = mem->start_addr + mem->memory_size;
 451     }
 452     g_free(d.dirty_bitmap);
 453
 454     return ret;
 455 }
 456
 457 static void kvm_coalesce_mmio_region(MemoryListener *listener,
 458                                      MemoryRegionSection *secion,
 459                                      hwaddr start, hwaddr size)
 460 {
 461     KVMState *s = kvm_state;
 462
 463     if (s->coalesced_mmio) {
 464         struct kvm_coalesced_mmio_zone zone;
 465
 466         zone.addr = start;
 467         zone.size = size;
 468         zone.pad = 0;
 469
 470         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 471     }
 472 }
 473
 474 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
 475                                        MemoryRegionSection *secion,
 476                                        hwaddr start, hwaddr size)
 477 {
 478     KVMState *s = kvm_state;
 479
 480     if (s->coalesced_mmio) {
 481         struct kvm_coalesced_mmio_zone zone;
 482
 483         zone.addr = start;
 484         zone.size = size;
 485         zone.pad = 0;
 486
 487         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 488     }
 489 }
 490
 491 int kvm_check_extension(KVMState *s, unsigned int extension)
 492 {
 493     int ret;
 494
 495     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 496     if (ret < 0) {
 497         ret = 0;
 498     }
 499
 500     return ret;
 501 }
 502
 503 static int kvm_check_many_ioeventfds(void)
 504 {
 505     /* Userspace can use ioeventfd for io notification.  This requires a host
 506      * that supports eventfd(2) and an I/O thread; since eventfd does not
 507      * support SIGIO it cannot interrupt the vcpu.
 508      *
 509      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 510      * can avoid creating too many ioeventfds.
 511      */
 512 #if defined(CONFIG_EVENTFD)
 513     int ioeventfds[7];
 514     int i, ret = 0;
 515     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 516         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 517         if (ioeventfds[i] < 0) {
 518             break;
 519         }
 520         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 521         if (ret < 0) {
 522             close(ioeventfds[i]);
 523             break;
 524         }
 525     }
 526
 527     /* Decide whether many devices are supported or not */
 528     ret = i == ARRAY_SIZE(ioeventfds);
 529
 530     while (i-- > 0) {
 531         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 532         close(ioeventfds[i]);
 533     }
 534     return ret;
 535 #else
 536     return 0;
 537 #endif
 538 }
 539
 540 static const KVMCapabilityInfo *
 541 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 542 {
 543     while (list->name) {
 544         if (!kvm_check_extension(s, list->value)) {
 545             return list;
 546         }
 547         list++;
 548     }
 549     return NULL;
 550 }
 551
 552 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
 553 {
 554     KVMState *s = kvm_state;
 555     KVMSlot *mem, old;
 556     int err;
 557     MemoryRegion *mr = section->mr;
 558     bool log_dirty = memory_region_is_logging(mr);
 559     hwaddr start_addr = section->offset_within_address_space;
 560     ram_addr_t size = section->size;
 561     void *ram = NULL;
 562     unsigned delta;
 563
 564     /* kvm works in page size chunks, but the function may be called
 565        with sub-page size and unaligned start address. */
 566     delta = TARGET_PAGE_ALIGN(size) - size;
 567     if (delta > size) {
 568         return;
 569     }
 570     start_addr += delta;
 571     size -= delta;
 572     size &= TARGET_PAGE_MASK;
 573     if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
 574         return;
 575     }
 576
 577     if (!memory_region_is_ram(mr)) {
 578         return;
 579     }
 580
 581     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
 582
 583     while (1) {
 584         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 585         if (!mem) {
 586             break;
 587         }
 588
 589         if (add && start_addr >= mem->start_addr &&
 590             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 591             (ram - start_addr == mem->ram - mem->start_addr)) {
 592             /* The new slot fits into the existing one and comes with
 593              * identical parameters - update flags and done. */
 594             kvm_slot_dirty_pages_log_change(mem, log_dirty);
 595             return;
 596         }
 597
 598         old = *mem;
 599
 600         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 601             kvm_physical_sync_dirty_bitmap(section);
 602         }
 603
 604         /* unregister the overlapping slot */
 605         mem->memory_size = 0;
 606         err = kvm_set_user_memory_region(s, mem);
 607         if (err) {
 608             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 609                     __func__, strerror(-err));
 610             abort();
 611         }
 612
 613         /* Workaround for older KVM versions: we can't join slots, even not by
 614          * unregistering the previous ones and then registering the larger
 615          * slot. We have to maintain the existing fragmentation. Sigh.
 616          *
 617          * This workaround assumes that the new slot starts at the same
 618          * address as the first existing one. If not or if some overlapping
 619          * slot comes around later, we will fail (not seen in practice so far)
 620          * - and actually require a recent KVM version. */
 621         if (s->broken_set_mem_region &&
 622             old.start_addr == start_addr && old.memory_size < size && add) {
 623             mem = kvm_alloc_slot(s);
 624             mem->memory_size = old.memory_size;
 625             mem->start_addr = old.start_addr;
 626             mem->ram = old.ram;
 627             mem->flags = kvm_mem_flags(s, log_dirty);
 628
 629             err = kvm_set_user_memory_region(s, mem);
 630             if (err) {
 631                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 632                         strerror(-err));
 633                 abort();
 634             }
 635
 636             start_addr += old.memory_size;
 637             ram += old.memory_size;
 638             size -= old.memory_size;
 639             continue;
 640         }
 641
 642         /* register prefix slot */
 643         if (old.start_addr < start_addr) {
 644             mem = kvm_alloc_slot(s);
 645             mem->memory_size = start_addr - old.start_addr;
 646             mem->start_addr = old.start_addr;
 647             mem->ram = old.ram;
 648             mem->flags =  kvm_mem_flags(s, log_dirty);
 649
 650             err = kvm_set_user_memory_region(s, mem);
 651             if (err) {
 652                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 653                         __func__, strerror(-err));
 654 #ifdef TARGET_PPC
 655                 fprintf(stderr, "%s: This is probably because your kernel's " \
 656                                 "PAGE_SIZE is too big. Please try to use 4k " \
 657                                 "PAGE_SIZE!\n", __func__);
 658 #endif
 659                 abort();
 660             }
 661         }
 662
 663         /* register suffix slot */
 664         if (old.start_addr + old.memory_size > start_addr + size) {
 665             ram_addr_t size_delta;
 666
 667             mem = kvm_alloc_slot(s);
 668             mem->start_addr = start_addr + size;
 669             size_delta = mem->start_addr - old.start_addr;
 670             mem->memory_size = old.memory_size - size_delta;
 671             mem->ram = old.ram + size_delta;
 672             mem->flags = kvm_mem_flags(s, log_dirty);
 673
 674             err = kvm_set_user_memory_region(s, mem);
 675             if (err) {
 676                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 677                         __func__, strerror(-err));
 678                 abort();
 679             }
 680         }
 681     }
 682
 683     /* in case the KVM bug workaround already "consumed" the new slot */
 684     if (!size) {
 685         return;
 686     }
 687     if (!add) {
 688         return;
 689     }
 690     mem = kvm_alloc_slot(s);
 691     mem->memory_size = size;
 692     mem->start_addr = start_addr;
 693     mem->ram = ram;
 694     mem->flags = kvm_mem_flags(s, log_dirty);
 695
 696     err = kvm_set_user_memory_region(s, mem);
 697     if (err) {
 698         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 699                 strerror(-err));
 700         abort();
 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_log_sync(MemoryListener *listener,
 717                          MemoryRegionSection *section)
 718 {
 719     int r;
 720
 721     r = kvm_physical_sync_dirty_bitmap(section);
 722     if (r < 0) {
 723         abort();
 724     }
 725 }
 726
 727 static void kvm_log_global_start(struct MemoryListener *listener)
 728 {
 729     int r;
 730
 731     r = kvm_set_migration_log(1);
 732     assert(r >= 0);
 733 }
 734
 735 static void kvm_log_global_stop(struct MemoryListener *listener)
 736 {
 737     int r;
 738
 739     r = kvm_set_migration_log(0);
 740     assert(r >= 0);
 741 }
 742
 743 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
 744                                   MemoryRegionSection *section,
 745                                   bool match_data, uint64_t data,
 746                                   EventNotifier *e)
 747 {
 748     int fd = event_notifier_get_fd(e);
 749     int r;
 750
 751     assert(match_data && section->size <= 8);
 752
 753     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 754                                data, true, section->size);
 755     if (r < 0) {
 756         abort();
 757     }
 758 }
 759
 760 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
 761                                   MemoryRegionSection *section,
 762                                   bool match_data, uint64_t data,
 763                                   EventNotifier *e)
 764 {
 765     int fd = event_notifier_get_fd(e);
 766     int r;
 767
 768     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 769                                data, false, section->size);
 770     if (r < 0) {
 771         abort();
 772     }
 773 }
 774
 775 static void kvm_io_ioeventfd_add(MemoryListener *listener,
 776                                  MemoryRegionSection *section,
 777                                  bool match_data, uint64_t data,
 778                                  EventNotifier *e)
 779 {
 780     int fd = event_notifier_get_fd(e);
 781     int r;
 782
 783     assert(match_data && section->size == 2);
 784
 785     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 786                                    data, true);
 787     if (r < 0) {
 788         abort();
 789     }
 790 }
 791
 792 static void kvm_io_ioeventfd_del(MemoryListener *listener,
 793                                  MemoryRegionSection *section,
 794                                  bool match_data, uint64_t data,
 795                                  EventNotifier *e)
 796
 797 {
 798     int fd = event_notifier_get_fd(e);
 799     int r;
 800
 801     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 802                                    data, false);
 803     if (r < 0) {
 804         abort();
 805     }
 806 }
 807
 808 static MemoryListener kvm_memory_listener = {
 809     .region_add = kvm_region_add,
 810     .region_del = kvm_region_del,
 811     .log_start = kvm_log_start,
 812     .log_stop = kvm_log_stop,
 813     .log_sync = kvm_log_sync,
 814     .log_global_start = kvm_log_global_start,
 815     .log_global_stop = kvm_log_global_stop,
 816     .eventfd_add = kvm_mem_ioeventfd_add,
 817     .eventfd_del = kvm_mem_ioeventfd_del,
 818     .coalesced_mmio_add = kvm_coalesce_mmio_region,
 819     .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
 820     .priority = 10,
 821 };
 822
 823 static MemoryListener kvm_io_listener = {
 824     .eventfd_add = kvm_io_ioeventfd_add,
 825     .eventfd_del = kvm_io_ioeventfd_del,
 826     .priority = 10,
 827 };
 828
 829 static void kvm_handle_interrupt(CPUState *cpu, int mask)
 830 {
 831     cpu->interrupt_request |= mask;
 832
 833     if (!qemu_cpu_is_self(cpu)) {
 834         qemu_cpu_kick(cpu);
 835     }
 836 }
 837
 838 int kvm_set_irq(KVMState *s, int irq, int level)
 839 {
 840     struct kvm_irq_level event;
 841     int ret;
 842
 843     assert(kvm_async_interrupts_enabled());
 844
 845     event.level = level;
 846     event.irq = irq;
 847     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
 848     if (ret < 0) {
 849         perror("kvm_set_irq");
 850         abort();
 851     }
 852
 853     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
 854 }
 855
 856 #ifdef KVM_CAP_IRQ_ROUTING
 857 typedef struct KVMMSIRoute {
 858     struct kvm_irq_routing_entry kroute;
 859     QTAILQ_ENTRY(KVMMSIRoute) entry;
 860 } KVMMSIRoute;
 861
 862 static void set_gsi(KVMState *s, unsigned int gsi)
 863 {
 864     s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
 865 }
 866
 867 static void clear_gsi(KVMState *s, unsigned int gsi)
 868 {
 869     s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
 870 }
 871
 872 static void kvm_init_irq_routing(KVMState *s)
 873 {
 874     int gsi_count, i;
 875
 876     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
 877     if (gsi_count > 0) {
 878         unsigned int gsi_bits, i;
 879
 880         /* Round up so we can search ints using ffs */
 881         gsi_bits = ALIGN(gsi_count, 32);
 882         s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
 883         s->gsi_count = gsi_count;
 884
 885         /* Mark any over-allocated bits as already in use */
 886         for (i = gsi_count; i < gsi_bits; i++) {
 887             set_gsi(s, i);
 888         }
 889     }
 890
 891     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
 892     s->nr_allocated_irq_routes = 0;
 893
 894     if (!s->direct_msi) {
 895         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
 896             QTAILQ_INIT(&s->msi_hashtab[i]);
 897         }
 898     }
 899
 900     kvm_arch_init_irq_routing(s);
 901 }
 902
 903 static void kvm_irqchip_commit_routes(KVMState *s)
 904 {
 905     int ret;
 906
 907     s->irq_routes->flags = 0;
 908     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
 909     assert(ret == 0);
 910 }
 911
 912 static void kvm_add_routing_entry(KVMState *s,
 913                                   struct kvm_irq_routing_entry *entry)
 914 {
 915     struct kvm_irq_routing_entry *new;
 916     int n, size;
 917
 918     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
 919         n = s->nr_allocated_irq_routes * 2;
 920         if (n < 64) {
 921             n = 64;
 922         }
 923         size = sizeof(struct kvm_irq_routing);
 924         size += n * sizeof(*new);
 925         s->irq_routes = g_realloc(s->irq_routes, size);
 926         s->nr_allocated_irq_routes = n;
 927     }
 928     n = s->irq_routes->nr++;
 929     new = &s->irq_routes->entries[n];
 930     memset(new, 0, sizeof(*new));
 931     new->gsi = entry->gsi;
 932     new->type = entry->type;
 933     new->flags = entry->flags;
 934     new->u = entry->u;
 935
 936     set_gsi(s, entry->gsi);
 937
 938     kvm_irqchip_commit_routes(s);
 939 }
 940
 941 static int kvm_update_routing_entry(KVMState *s,
 942                                     struct kvm_irq_routing_entry *new_entry)
 943 {
 944     struct kvm_irq_routing_entry *entry;
 945     int n;
 946
 947     for (n = 0; n < s->irq_routes->nr; n++) {
 948         entry = &s->irq_routes->entries[n];
 949         if (entry->gsi != new_entry->gsi) {
 950             continue;
 951         }
 952
 953         entry->type = new_entry->type;
 954         entry->flags = new_entry->flags;
 955         entry->u = new_entry->u;
 956
 957         kvm_irqchip_commit_routes(s);
 958
 959         return 0;
 960     }
 961
 962     return -ESRCH;
 963 }
 964
 965 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
 966 {
 967     struct kvm_irq_routing_entry e;
 968
 969     assert(pin < s->gsi_count);
 970
 971     e.gsi = irq;
 972     e.type = KVM_IRQ_ROUTING_IRQCHIP;
 973     e.flags = 0;
 974     e.u.irqchip.irqchip = irqchip;
 975     e.u.irqchip.pin = pin;
 976     kvm_add_routing_entry(s, &e);
 977 }
 978
 979 void kvm_irqchip_release_virq(KVMState *s, int virq)
 980 {
 981     struct kvm_irq_routing_entry *e;
 982     int i;
 983
 984     for (i = 0; i < s->irq_routes->nr; i++) {
 985         e = &s->irq_routes->entries[i];
 986         if (e->gsi == virq) {
 987             s->irq_routes->nr--;
 988             *e = s->irq_routes->entries[s->irq_routes->nr];
 989         }
 990     }
 991     clear_gsi(s, virq);
 992 }
 993
 994 static unsigned int kvm_hash_msi(uint32_t data)
 995 {
 996     /* This is optimized for IA32 MSI layout. However, no other arch shall
 997      * repeat the mistake of not providing a direct MSI injection API. */
 998     return data & 0xff;
 999 }
1000
1001 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1002 {
1003     KVMMSIRoute *route, *next;
1004     unsigned int hash;
1005
1006     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1007         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1008             kvm_irqchip_release_virq(s, route->kroute.gsi);
1009             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1010             g_free(route);
1011         }
1012     }
1013 }
1014
1015 static int kvm_irqchip_get_virq(KVMState *s)
1016 {
1017     uint32_t *word = s->used_gsi_bitmap;
1018     int max_words = ALIGN(s->gsi_count, 32) / 32;
1019     int i, bit;
1020     bool retry = true;
1021
1022 again:
1023     /* Return the lowest unused GSI in the bitmap */
1024     for (i = 0; i < max_words; i++) {
1025         bit = ffs(~word[i]);
1026         if (!bit) {
1027             continue;
1028         }
1029
1030         return bit - 1 + i * 32;
1031     }
1032     if (!s->direct_msi && retry) {
1033         retry = false;
1034         kvm_flush_dynamic_msi_routes(s);
1035         goto again;
1036     }
1037     return -ENOSPC;
1038
1039 }
1040
1041 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1042 {
1043     unsigned int hash = kvm_hash_msi(msg.data);
1044     KVMMSIRoute *route;
1045
1046     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1047         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1048             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1049             route->kroute.u.msi.data == msg.data) {
1050             return route;
1051         }
1052     }
1053     return NULL;
1054 }
1055
1056 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1057 {
1058     struct kvm_msi msi;
1059     KVMMSIRoute *route;
1060
1061     if (s->direct_msi) {
1062         msi.address_lo = (uint32_t)msg.address;
1063         msi.address_hi = msg.address >> 32;
1064         msi.data = msg.data;
1065         msi.flags = 0;
1066         memset(msi.pad, 0, sizeof(msi.pad));
1067
1068         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1069     }
1070
1071     route = kvm_lookup_msi_route(s, msg);
1072     if (!route) {
1073         int virq;
1074
1075         virq = kvm_irqchip_get_virq(s);
1076         if (virq < 0) {
1077             return virq;
1078         }
1079
1080         route = g_malloc(sizeof(KVMMSIRoute));
1081         route->kroute.gsi = virq;
1082         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1083         route->kroute.flags = 0;
1084         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1085         route->kroute.u.msi.address_hi = msg.address >> 32;
1086         route->kroute.u.msi.data = msg.data;
1087
1088         kvm_add_routing_entry(s, &route->kroute);
1089
1090         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1091                            entry);
1092     }
1093
1094     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1095
1096     return kvm_set_irq(s, route->kroute.gsi, 1);
1097 }
1098
1099 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1100 {
1101     struct kvm_irq_routing_entry kroute;
1102     int virq;
1103
1104     if (!kvm_gsi_routing_enabled()) {
1105         return -ENOSYS;
1106     }
1107
1108     virq = kvm_irqchip_get_virq(s);
1109     if (virq < 0) {
1110         return virq;
1111     }
1112
1113     kroute.gsi = virq;
1114     kroute.type = KVM_IRQ_ROUTING_MSI;
1115     kroute.flags = 0;
1116     kroute.u.msi.address_lo = (uint32_t)msg.address;
1117     kroute.u.msi.address_hi = msg.address >> 32;
1118     kroute.u.msi.data = msg.data;
1119
1120     kvm_add_routing_entry(s, &kroute);
1121
1122     return virq;
1123 }
1124
1125 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1126 {
1127     struct kvm_irq_routing_entry kroute;
1128
1129     if (!kvm_irqchip_in_kernel()) {
1130         return -ENOSYS;
1131     }
1132
1133     kroute.gsi = virq;
1134     kroute.type = KVM_IRQ_ROUTING_MSI;
1135     kroute.flags = 0;
1136     kroute.u.msi.address_lo = (uint32_t)msg.address;
1137     kroute.u.msi.address_hi = msg.address >> 32;
1138     kroute.u.msi.data = msg.data;
1139
1140     return kvm_update_routing_entry(s, &kroute);
1141 }
1142
1143 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1144 {
1145     struct kvm_irqfd irqfd = {
1146         .fd = fd,
1147         .gsi = virq,
1148         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1149     };
1150
1151     if (!kvm_irqfds_enabled()) {
1152         return -ENOSYS;
1153     }
1154
1155     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1156 }
1157
1158 #else /* !KVM_CAP_IRQ_ROUTING */
1159
1160 static void kvm_init_irq_routing(KVMState *s)
1161 {
1162 }
1163
1164 void kvm_irqchip_release_virq(KVMState *s, int virq)
1165 {
1166 }
1167
1168 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1169 {
1170     abort();
1171 }
1172
1173 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1174 {
1175     return -ENOSYS;
1176 }
1177
1178 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1179 {
1180     abort();
1181 }
1182
1183 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1184 {
1185     return -ENOSYS;
1186 }
1187 #endif /* !KVM_CAP_IRQ_ROUTING */
1188
1189 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1190 {
1191     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, true);
1192 }
1193
1194 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1195 {
1196     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, false);
1197 }
1198
1199 static int kvm_irqchip_create(KVMState *s)
1200 {
1201     QemuOptsList *list = qemu_find_opts("machine");
1202     int ret;
1203
1204     if (QTAILQ_EMPTY(&list->head) ||
1205         !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1206                            "kernel_irqchip", true) ||
1207         !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1208         return 0;
1209     }
1210
1211     ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1212     if (ret < 0) {
1213         fprintf(stderr, "Create kernel irqchip failed\n");
1214         return ret;
1215     }
1216
1217     kvm_kernel_irqchip = true;
1218     /* If we have an in-kernel IRQ chip then we must have asynchronous
1219      * interrupt delivery (though the reverse is not necessarily true)
1220      */
1221     kvm_async_interrupts_allowed = true;
1222
1223     kvm_init_irq_routing(s);
1224
1225     return 0;
1226 }
1227
1228 static int kvm_max_vcpus(KVMState *s)
1229 {
1230     int ret;
1231
1232     /* Find number of supported CPUs using the recommended
1233      * procedure from the kernel API documentation to cope with
1234      * older kernels that may be missing capabilities.
1235      */
1236     ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1237     if (ret) {
1238         return ret;
1239     }
1240     ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1241     if (ret) {
1242         return ret;
1243     }
1244
1245     return 4;
1246 }
1247
1248 int kvm_init(void)
1249 {
1250     static const char upgrade_note[] =
1251         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1252         "(see http://sourceforge.net/projects/kvm).\n";
1253     KVMState *s;
1254     const KVMCapabilityInfo *missing_cap;
1255     int ret;
1256     int i;
1257     int max_vcpus;
1258
1259     s = g_malloc0(sizeof(KVMState));
1260
1261     /*
1262      * On systems where the kernel can support different base page
1263      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1264      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1265      * page size for the system though.
1266      */
1267     assert(TARGET_PAGE_SIZE <= getpagesize());
1268
1269 #ifdef KVM_CAP_SET_GUEST_DEBUG
1270     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1271 #endif
1272     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1273         s->slots[i].slot = i;
1274     }
1275     s->vmfd = -1;
1276     s->fd = qemu_open("/dev/kvm", O_RDWR);
1277     if (s->fd == -1) {
1278         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1279         ret = -errno;
1280         goto err;
1281     }
1282
1283     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1284     if (ret < KVM_API_VERSION) {
1285         if (ret > 0) {
1286             ret = -EINVAL;
1287         }
1288         fprintf(stderr, "kvm version too old\n");
1289         goto err;
1290     }
1291
1292     if (ret > KVM_API_VERSION) {
1293         ret = -EINVAL;
1294         fprintf(stderr, "kvm version not supported\n");
1295         goto err;
1296     }
1297
1298     max_vcpus = kvm_max_vcpus(s);
1299     if (smp_cpus > max_vcpus) {
1300         ret = -EINVAL;
1301         fprintf(stderr, "Number of SMP cpus requested (%d) exceeds max cpus "
1302                 "supported by KVM (%d)\n", smp_cpus, max_vcpus);
1303         goto err;
1304     }
1305
1306     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1307     if (s->vmfd < 0) {
1308 #ifdef TARGET_S390X
1309         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1310                         "your host kernel command line\n");
1311 #endif
1312         ret = s->vmfd;
1313         goto err;
1314     }
1315
1316     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1317     if (!missing_cap) {
1318         missing_cap =
1319             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1320     }
1321     if (missing_cap) {
1322         ret = -EINVAL;
1323         fprintf(stderr, "kvm does not support %s\n%s",
1324                 missing_cap->name, upgrade_note);
1325         goto err;
1326     }
1327
1328     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1329
1330     s->broken_set_mem_region = 1;
1331     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1332     if (ret > 0) {
1333         s->broken_set_mem_region = 0;
1334     }
1335
1336 #ifdef KVM_CAP_VCPU_EVENTS
1337     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1338 #endif
1339
1340     s->robust_singlestep =
1341         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1342
1343 #ifdef KVM_CAP_DEBUGREGS
1344     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1345 #endif
1346
1347 #ifdef KVM_CAP_XSAVE
1348     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1349 #endif
1350
1351 #ifdef KVM_CAP_XCRS
1352     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1353 #endif
1354
1355 #ifdef KVM_CAP_PIT_STATE2
1356     s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1357 #endif
1358
1359 #ifdef KVM_CAP_IRQ_ROUTING
1360     s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1361 #endif
1362
1363     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1364
1365     s->irq_set_ioctl = KVM_IRQ_LINE;
1366     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1367         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1368     }
1369
1370     ret = kvm_arch_init(s);
1371     if (ret < 0) {
1372         goto err;
1373     }
1374
1375     ret = kvm_irqchip_create(s);
1376     if (ret < 0) {
1377         goto err;
1378     }
1379
1380     kvm_state = s;
1381     memory_listener_register(&kvm_memory_listener, &address_space_memory);
1382     memory_listener_register(&kvm_io_listener, &address_space_io);
1383
1384     s->many_ioeventfds = kvm_check_many_ioeventfds();
1385
1386     cpu_interrupt_handler = kvm_handle_interrupt;
1387
1388     return 0;
1389
1390 err:
1391     if (s->vmfd >= 0) {
1392         close(s->vmfd);
1393     }
1394     if (s->fd != -1) {
1395         close(s->fd);
1396     }
1397     g_free(s);
1398
1399     return ret;
1400 }
1401
1402 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1403                           uint32_t count)
1404 {
1405     int i;
1406     uint8_t *ptr = data;
1407
1408     for (i = 0; i < count; i++) {
1409         if (direction == KVM_EXIT_IO_IN) {
1410             switch (size) {
1411             case 1:
1412                 stb_p(ptr, cpu_inb(port));
1413                 break;
1414             case 2:
1415                 stw_p(ptr, cpu_inw(port));
1416                 break;
1417             case 4:
1418                 stl_p(ptr, cpu_inl(port));
1419                 break;
1420             }
1421         } else {
1422             switch (size) {
1423             case 1:
1424                 cpu_outb(port, ldub_p(ptr));
1425                 break;
1426             case 2:
1427                 cpu_outw(port, lduw_p(ptr));
1428                 break;
1429             case 4:
1430                 cpu_outl(port, ldl_p(ptr));
1431                 break;
1432             }
1433         }
1434
1435         ptr += size;
1436     }
1437 }
1438
1439 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1440 {
1441     CPUState *cpu = ENV_GET_CPU(env);
1442
1443     fprintf(stderr, "KVM internal error.");
1444     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1445         int i;
1446
1447         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1448         for (i = 0; i < run->internal.ndata; ++i) {
1449             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1450                     i, (uint64_t)run->internal.data[i]);
1451         }
1452     } else {
1453         fprintf(stderr, "\n");
1454     }
1455     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1456         fprintf(stderr, "emulation failure\n");
1457         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1458             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1459             return EXCP_INTERRUPT;
1460         }
1461     }
1462     /* FIXME: Should trigger a qmp message to let management know
1463      * something went wrong.
1464      */
1465     return -1;
1466 }
1467
1468 void kvm_flush_coalesced_mmio_buffer(void)
1469 {
1470     KVMState *s = kvm_state;
1471
1472     if (s->coalesced_flush_in_progress) {
1473         return;
1474     }
1475
1476     s->coalesced_flush_in_progress = true;
1477
1478     if (s->coalesced_mmio_ring) {
1479         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1480         while (ring->first != ring->last) {
1481             struct kvm_coalesced_mmio *ent;
1482
1483             ent = &ring->coalesced_mmio[ring->first];
1484
1485             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1486             smp_wmb();
1487             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1488         }
1489     }
1490
1491     s->coalesced_flush_in_progress = false;
1492 }
1493
1494 static void do_kvm_cpu_synchronize_state(void *arg)
1495 {
1496     CPUState *cpu = arg;
1497
1498     if (!cpu->kvm_vcpu_dirty) {
1499         kvm_arch_get_registers(cpu);
1500         cpu->kvm_vcpu_dirty = true;
1501     }
1502 }
1503
1504 void kvm_cpu_synchronize_state(CPUArchState *env)
1505 {
1506     CPUState *cpu = ENV_GET_CPU(env);
1507
1508     if (!cpu->kvm_vcpu_dirty) {
1509         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1510     }
1511 }
1512
1513 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1514 {
1515     CPUState *cpu = ENV_GET_CPU(env);
1516
1517     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1518     cpu->kvm_vcpu_dirty = false;
1519 }
1520
1521 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1522 {
1523     CPUState *cpu = ENV_GET_CPU(env);
1524
1525     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1526     cpu->kvm_vcpu_dirty = false;
1527 }
1528
1529 int kvm_cpu_exec(CPUArchState *env)
1530 {
1531     CPUState *cpu = ENV_GET_CPU(env);
1532     struct kvm_run *run = cpu->kvm_run;
1533     int ret, run_ret;
1534
1535     DPRINTF("kvm_cpu_exec()\n");
1536
1537     if (kvm_arch_process_async_events(cpu)) {
1538         cpu->exit_request = 0;
1539         return EXCP_HLT;
1540     }
1541
1542     do {
1543         if (cpu->kvm_vcpu_dirty) {
1544             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1545             cpu->kvm_vcpu_dirty = false;
1546         }
1547
1548         kvm_arch_pre_run(cpu, run);
1549         if (cpu->exit_request) {
1550             DPRINTF("interrupt exit requested\n");
1551             /*
1552              * KVM requires us to reenter the kernel after IO exits to complete
1553              * instruction emulation. This self-signal will ensure that we
1554              * leave ASAP again.
1555              */
1556             qemu_cpu_kick_self();
1557         }
1558         qemu_mutex_unlock_iothread();
1559
1560         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1561
1562         qemu_mutex_lock_iothread();
1563         kvm_arch_post_run(cpu, run);
1564
1565         if (run_ret < 0) {
1566             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1567                 DPRINTF("io window exit\n");
1568                 ret = EXCP_INTERRUPT;
1569                 break;
1570             }
1571             fprintf(stderr, "error: kvm run failed %s\n",
1572                     strerror(-run_ret));
1573             abort();
1574         }
1575
1576         switch (run->exit_reason) {
1577         case KVM_EXIT_IO:
1578             DPRINTF("handle_io\n");
1579             kvm_handle_io(run->io.port,
1580                           (uint8_t *)run + run->io.data_offset,
1581                           run->io.direction,
1582                           run->io.size,
1583                           run->io.count);
1584             ret = 0;
1585             break;
1586         case KVM_EXIT_MMIO:
1587             DPRINTF("handle_mmio\n");
1588             cpu_physical_memory_rw(run->mmio.phys_addr,
1589                                    run->mmio.data,
1590                                    run->mmio.len,
1591                                    run->mmio.is_write);
1592             ret = 0;
1593             break;
1594         case KVM_EXIT_IRQ_WINDOW_OPEN:
1595             DPRINTF("irq_window_open\n");
1596             ret = EXCP_INTERRUPT;
1597             break;
1598         case KVM_EXIT_SHUTDOWN:
1599             DPRINTF("shutdown\n");
1600             qemu_system_reset_request();
1601             ret = EXCP_INTERRUPT;
1602             break;
1603         case KVM_EXIT_UNKNOWN:
1604             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1605                     (uint64_t)run->hw.hardware_exit_reason);
1606             ret = -1;
1607             break;
1608         case KVM_EXIT_INTERNAL_ERROR:
1609             ret = kvm_handle_internal_error(env, run);
1610             break;
1611         default:
1612             DPRINTF("kvm_arch_handle_exit\n");
1613             ret = kvm_arch_handle_exit(cpu, run);
1614             break;
1615         }
1616     } while (ret == 0);
1617
1618     if (ret < 0) {
1619         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1620         vm_stop(RUN_STATE_INTERNAL_ERROR);
1621     }
1622
1623     cpu->exit_request = 0;
1624     return ret;
1625 }
1626
1627 int kvm_ioctl(KVMState *s, int type, ...)
1628 {
1629     int ret;
1630     void *arg;
1631     va_list ap;
1632
1633     va_start(ap, type);
1634     arg = va_arg(ap, void *);
1635     va_end(ap);
1636
1637     ret = ioctl(s->fd, type, arg);
1638     if (ret == -1) {
1639         ret = -errno;
1640     }
1641     return ret;
1642 }
1643
1644 int kvm_vm_ioctl(KVMState *s, int type, ...)
1645 {
1646     int ret;
1647     void *arg;
1648     va_list ap;
1649
1650     va_start(ap, type);
1651     arg = va_arg(ap, void *);
1652     va_end(ap);
1653
1654     ret = ioctl(s->vmfd, type, arg);
1655     if (ret == -1) {
1656         ret = -errno;
1657     }
1658     return ret;
1659 }
1660
1661 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1662 {
1663     int ret;
1664     void *arg;
1665     va_list ap;
1666
1667     va_start(ap, type);
1668     arg = va_arg(ap, void *);
1669     va_end(ap);
1670
1671     ret = ioctl(cpu->kvm_fd, type, arg);
1672     if (ret == -1) {
1673         ret = -errno;
1674     }
1675     return ret;
1676 }
1677
1678 int kvm_has_sync_mmu(void)
1679 {
1680     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1681 }
1682
1683 int kvm_has_vcpu_events(void)
1684 {
1685     return kvm_state->vcpu_events;
1686 }
1687
1688 int kvm_has_robust_singlestep(void)
1689 {
1690     return kvm_state->robust_singlestep;
1691 }
1692
1693 int kvm_has_debugregs(void)
1694 {
1695     return kvm_state->debugregs;
1696 }
1697
1698 int kvm_has_xsave(void)
1699 {
1700     return kvm_state->xsave;
1701 }
1702
1703 int kvm_has_xcrs(void)
1704 {
1705     return kvm_state->xcrs;
1706 }
1707
1708 int kvm_has_pit_state2(void)
1709 {
1710     return kvm_state->pit_state2;
1711 }
1712
1713 int kvm_has_many_ioeventfds(void)
1714 {
1715     if (!kvm_enabled()) {
1716         return 0;
1717     }
1718     return kvm_state->many_ioeventfds;
1719 }
1720
1721 int kvm_has_gsi_routing(void)
1722 {
1723 #ifdef KVM_CAP_IRQ_ROUTING
1724     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1725 #else
1726     return false;
1727 #endif
1728 }
1729
1730 int kvm_has_intx_set_mask(void)
1731 {
1732     return kvm_state->intx_set_mask;
1733 }
1734
1735 void *kvm_vmalloc(ram_addr_t size)
1736 {
1737 #ifdef TARGET_S390X
1738     void *mem;
1739
1740     mem = kvm_arch_vmalloc(size);
1741     if (mem) {
1742         return mem;
1743     }
1744 #endif
1745     return qemu_vmalloc(size);
1746 }
1747
1748 void kvm_setup_guest_memory(void *start, size_t size)
1749 {
1750 #ifdef CONFIG_VALGRIND_H
1751     VALGRIND_MAKE_MEM_DEFINED(start, size);
1752 #endif
1753     if (!kvm_has_sync_mmu()) {
1754         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1755
1756         if (ret) {
1757             perror("qemu_madvise");
1758             fprintf(stderr,
1759                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1760             exit(1);
1761         }
1762     }
1763 }
1764
1765 #ifdef KVM_CAP_SET_GUEST_DEBUG
1766 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1767                                                  target_ulong pc)
1768 {
1769     struct kvm_sw_breakpoint *bp;
1770
1771     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1772         if (bp->pc == pc) {
1773             return bp;
1774         }
1775     }
1776     return NULL;
1777 }
1778
1779 int kvm_sw_breakpoints_active(CPUState *cpu)
1780 {
1781     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1782 }
1783
1784 struct kvm_set_guest_debug_data {
1785     struct kvm_guest_debug dbg;
1786     CPUState *cpu;
1787     int err;
1788 };
1789
1790 static void kvm_invoke_set_guest_debug(void *data)
1791 {
1792     struct kvm_set_guest_debug_data *dbg_data = data;
1793
1794     dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1795                                    &dbg_data->dbg);
1796 }
1797
1798 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1799 {
1800     CPUState *cpu = ENV_GET_CPU(env);
1801     struct kvm_set_guest_debug_data data;
1802
1803     data.dbg.control = reinject_trap;
1804
1805     if (env->singlestep_enabled) {
1806         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1807     }
1808     kvm_arch_update_guest_debug(cpu, &data.dbg);
1809     data.cpu = cpu;
1810
1811     run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1812     return data.err;
1813 }
1814
1815 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1816                           target_ulong len, int type)
1817 {
1818     CPUState *current_cpu = ENV_GET_CPU(current_env);
1819     struct kvm_sw_breakpoint *bp;
1820     CPUArchState *env;
1821     int err;
1822
1823     if (type == GDB_BREAKPOINT_SW) {
1824         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1825         if (bp) {
1826             bp->use_count++;
1827             return 0;
1828         }
1829
1830         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1831         if (!bp) {
1832             return -ENOMEM;
1833         }
1834
1835         bp->pc = addr;
1836         bp->use_count = 1;
1837         err = kvm_arch_insert_sw_breakpoint(current_cpu, bp);
1838         if (err) {
1839             g_free(bp);
1840             return err;
1841         }
1842
1843         QTAILQ_INSERT_HEAD(&current_cpu->kvm_state->kvm_sw_breakpoints,
1844                           bp, entry);
1845     } else {
1846         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1847         if (err) {
1848             return err;
1849         }
1850     }
1851
1852     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1853         err = kvm_update_guest_debug(env, 0);
1854         if (err) {
1855             return err;
1856         }
1857     }
1858     return 0;
1859 }
1860
1861 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1862                           target_ulong len, int type)
1863 {
1864     CPUState *current_cpu = ENV_GET_CPU(current_env);
1865     struct kvm_sw_breakpoint *bp;
1866     CPUArchState *env;
1867     int err;
1868
1869     if (type == GDB_BREAKPOINT_SW) {
1870         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1871         if (!bp) {
1872             return -ENOENT;
1873         }
1874
1875         if (bp->use_count > 1) {
1876             bp->use_count--;
1877             return 0;
1878         }
1879
1880         err = kvm_arch_remove_sw_breakpoint(current_cpu, bp);
1881         if (err) {
1882             return err;
1883         }
1884
1885         QTAILQ_REMOVE(&current_cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
1886         g_free(bp);
1887     } else {
1888         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1889         if (err) {
1890             return err;
1891         }
1892     }
1893
1894     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1895         err = kvm_update_guest_debug(env, 0);
1896         if (err) {
1897             return err;
1898         }
1899     }
1900     return 0;
1901 }
1902
1903 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1904 {
1905     CPUState *current_cpu = ENV_GET_CPU(current_env);
1906     struct kvm_sw_breakpoint *bp, *next;
1907     KVMState *s = current_cpu->kvm_state;
1908     CPUArchState *env;
1909     CPUState *cpu;
1910
1911     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1912         if (kvm_arch_remove_sw_breakpoint(current_cpu, bp) != 0) {
1913             /* Try harder to find a CPU that currently sees the breakpoint. */
1914             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1915                 cpu = ENV_GET_CPU(env);
1916                 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
1917                     break;
1918                 }
1919             }
1920         }
1921         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
1922         g_free(bp);
1923     }
1924     kvm_arch_remove_all_hw_breakpoints();
1925
1926     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1927         kvm_update_guest_debug(env, 0);
1928     }
1929 }
1930
1931 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1932
1933 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1934 {
1935     return -EINVAL;
1936 }
1937
1938 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1939                           target_ulong len, int type)
1940 {
1941     return -EINVAL;
1942 }
1943
1944 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1945                           target_ulong len, int type)
1946 {
1947     return -EINVAL;
1948 }
1949
1950 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1951 {
1952 }
1953 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1954
1955 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1956 {
1957     CPUState *cpu = ENV_GET_CPU(env);
1958     struct kvm_signal_mask *sigmask;
1959     int r;
1960
1961     if (!sigset) {
1962         return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
1963     }
1964
1965     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1966
1967     sigmask->len = 8;
1968     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1969     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
1970     g_free(sigmask);
1971
1972     return r;
1973 }
1974
1975 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1976                            uint32_t size)
1977 {
1978     int ret;
1979     struct kvm_ioeventfd iofd;
1980
1981     iofd.datamatch = val;
1982     iofd.addr = addr;
1983     iofd.len = size;
1984     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1985     iofd.fd = fd;
1986
1987     if (!kvm_enabled()) {
1988         return -ENOSYS;
1989     }
1990
1991     if (!assign) {
1992         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1993     }
1994
1995     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1996
1997     if (ret < 0) {
1998         return -errno;
1999     }
2000
2001     return 0;
2002 }
2003
2004 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
2005 {
2006     struct kvm_ioeventfd kick = {
2007         .datamatch = val,
2008         .addr = addr,
2009         .len = 2,
2010         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
2011         .fd = fd,
2012     };
2013     int r;
2014     if (!kvm_enabled()) {
2015         return -ENOSYS;
2016     }
2017     if (!assign) {
2018         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
2019     }
2020     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
2021     if (r < 0) {
2022         return r;
2023     }
2024     return 0;
2025 }
2026
2027 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2028 {
2029     return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2030 }
2031
2032 int kvm_on_sigbus(int code, void *addr)
2033 {
2034     return kvm_arch_on_sigbus(code, addr);
2035 }