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(CPUArchState *env, int mask)
 830 {
 831     CPUState *cpu = ENV_GET_CPU(env);
 832
 833     env->interrupt_request |= mask;
 834
 835     if (!qemu_cpu_is_self(cpu)) {
 836         qemu_cpu_kick(cpu);
 837     }
 838 }
 839
 840 int kvm_set_irq(KVMState *s, int irq, int level)
 841 {
 842     struct kvm_irq_level event;
 843     int ret;
 844
 845     assert(kvm_async_interrupts_enabled());
 846
 847     event.level = level;
 848     event.irq = irq;
 849     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
 850     if (ret < 0) {
 851         perror("kvm_set_irq");
 852         abort();
 853     }
 854
 855     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
 856 }
 857
 858 #ifdef KVM_CAP_IRQ_ROUTING
 859 typedef struct KVMMSIRoute {
 860     struct kvm_irq_routing_entry kroute;
 861     QTAILQ_ENTRY(KVMMSIRoute) entry;
 862 } KVMMSIRoute;
 863
 864 static void set_gsi(KVMState *s, unsigned int gsi)
 865 {
 866     s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
 867 }
 868
 869 static void clear_gsi(KVMState *s, unsigned int gsi)
 870 {
 871     s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
 872 }
 873
 874 static void kvm_init_irq_routing(KVMState *s)
 875 {
 876     int gsi_count, i;
 877
 878     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
 879     if (gsi_count > 0) {
 880         unsigned int gsi_bits, i;
 881
 882         /* Round up so we can search ints using ffs */
 883         gsi_bits = ALIGN(gsi_count, 32);
 884         s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
 885         s->gsi_count = gsi_count;
 886
 887         /* Mark any over-allocated bits as already in use */
 888         for (i = gsi_count; i < gsi_bits; i++) {
 889             set_gsi(s, i);
 890         }
 891     }
 892
 893     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
 894     s->nr_allocated_irq_routes = 0;
 895
 896     if (!s->direct_msi) {
 897         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
 898             QTAILQ_INIT(&s->msi_hashtab[i]);
 899         }
 900     }
 901
 902     kvm_arch_init_irq_routing(s);
 903 }
 904
 905 static void kvm_irqchip_commit_routes(KVMState *s)
 906 {
 907     int ret;
 908
 909     s->irq_routes->flags = 0;
 910     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
 911     assert(ret == 0);
 912 }
 913
 914 static void kvm_add_routing_entry(KVMState *s,
 915                                   struct kvm_irq_routing_entry *entry)
 916 {
 917     struct kvm_irq_routing_entry *new;
 918     int n, size;
 919
 920     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
 921         n = s->nr_allocated_irq_routes * 2;
 922         if (n < 64) {
 923             n = 64;
 924         }
 925         size = sizeof(struct kvm_irq_routing);
 926         size += n * sizeof(*new);
 927         s->irq_routes = g_realloc(s->irq_routes, size);
 928         s->nr_allocated_irq_routes = n;
 929     }
 930     n = s->irq_routes->nr++;
 931     new = &s->irq_routes->entries[n];
 932     memset(new, 0, sizeof(*new));
 933     new->gsi = entry->gsi;
 934     new->type = entry->type;
 935     new->flags = entry->flags;
 936     new->u = entry->u;
 937
 938     set_gsi(s, entry->gsi);
 939
 940     kvm_irqchip_commit_routes(s);
 941 }
 942
 943 static int kvm_update_routing_entry(KVMState *s,
 944                                     struct kvm_irq_routing_entry *new_entry)
 945 {
 946     struct kvm_irq_routing_entry *entry;
 947     int n;
 948
 949     for (n = 0; n < s->irq_routes->nr; n++) {
 950         entry = &s->irq_routes->entries[n];
 951         if (entry->gsi != new_entry->gsi) {
 952             continue;
 953         }
 954
 955         entry->type = new_entry->type;
 956         entry->flags = new_entry->flags;
 957         entry->u = new_entry->u;
 958
 959         kvm_irqchip_commit_routes(s);
 960
 961         return 0;
 962     }
 963
 964     return -ESRCH;
 965 }
 966
 967 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
 968 {
 969     struct kvm_irq_routing_entry e;
 970
 971     assert(pin < s->gsi_count);
 972
 973     e.gsi = irq;
 974     e.type = KVM_IRQ_ROUTING_IRQCHIP;
 975     e.flags = 0;
 976     e.u.irqchip.irqchip = irqchip;
 977     e.u.irqchip.pin = pin;
 978     kvm_add_routing_entry(s, &e);
 979 }
 980
 981 void kvm_irqchip_release_virq(KVMState *s, int virq)
 982 {
 983     struct kvm_irq_routing_entry *e;
 984     int i;
 985
 986     for (i = 0; i < s->irq_routes->nr; i++) {
 987         e = &s->irq_routes->entries[i];
 988         if (e->gsi == virq) {
 989             s->irq_routes->nr--;
 990             *e = s->irq_routes->entries[s->irq_routes->nr];
 991         }
 992     }
 993     clear_gsi(s, virq);
 994 }
 995
 996 static unsigned int kvm_hash_msi(uint32_t data)
 997 {
 998     /* This is optimized for IA32 MSI layout. However, no other arch shall
 999      * repeat the mistake of not providing a direct MSI injection API. */
1000     return data & 0xff;
1001 }
1002
1003 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1004 {
1005     KVMMSIRoute *route, *next;
1006     unsigned int hash;
1007
1008     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1009         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1010             kvm_irqchip_release_virq(s, route->kroute.gsi);
1011             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1012             g_free(route);
1013         }
1014     }
1015 }
1016
1017 static int kvm_irqchip_get_virq(KVMState *s)
1018 {
1019     uint32_t *word = s->used_gsi_bitmap;
1020     int max_words = ALIGN(s->gsi_count, 32) / 32;
1021     int i, bit;
1022     bool retry = true;
1023
1024 again:
1025     /* Return the lowest unused GSI in the bitmap */
1026     for (i = 0; i < max_words; i++) {
1027         bit = ffs(~word[i]);
1028         if (!bit) {
1029             continue;
1030         }
1031
1032         return bit - 1 + i * 32;
1033     }
1034     if (!s->direct_msi && retry) {
1035         retry = false;
1036         kvm_flush_dynamic_msi_routes(s);
1037         goto again;
1038     }
1039     return -ENOSPC;
1040
1041 }
1042
1043 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1044 {
1045     unsigned int hash = kvm_hash_msi(msg.data);
1046     KVMMSIRoute *route;
1047
1048     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1049         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1050             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1051             route->kroute.u.msi.data == msg.data) {
1052             return route;
1053         }
1054     }
1055     return NULL;
1056 }
1057
1058 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1059 {
1060     struct kvm_msi msi;
1061     KVMMSIRoute *route;
1062
1063     if (s->direct_msi) {
1064         msi.address_lo = (uint32_t)msg.address;
1065         msi.address_hi = msg.address >> 32;
1066         msi.data = msg.data;
1067         msi.flags = 0;
1068         memset(msi.pad, 0, sizeof(msi.pad));
1069
1070         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1071     }
1072
1073     route = kvm_lookup_msi_route(s, msg);
1074     if (!route) {
1075         int virq;
1076
1077         virq = kvm_irqchip_get_virq(s);
1078         if (virq < 0) {
1079             return virq;
1080         }
1081
1082         route = g_malloc(sizeof(KVMMSIRoute));
1083         route->kroute.gsi = virq;
1084         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1085         route->kroute.flags = 0;
1086         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1087         route->kroute.u.msi.address_hi = msg.address >> 32;
1088         route->kroute.u.msi.data = msg.data;
1089
1090         kvm_add_routing_entry(s, &route->kroute);
1091
1092         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1093                            entry);
1094     }
1095
1096     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1097
1098     return kvm_set_irq(s, route->kroute.gsi, 1);
1099 }
1100
1101 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1102 {
1103     struct kvm_irq_routing_entry kroute;
1104     int virq;
1105
1106     if (!kvm_gsi_routing_enabled()) {
1107         return -ENOSYS;
1108     }
1109
1110     virq = kvm_irqchip_get_virq(s);
1111     if (virq < 0) {
1112         return virq;
1113     }
1114
1115     kroute.gsi = virq;
1116     kroute.type = KVM_IRQ_ROUTING_MSI;
1117     kroute.flags = 0;
1118     kroute.u.msi.address_lo = (uint32_t)msg.address;
1119     kroute.u.msi.address_hi = msg.address >> 32;
1120     kroute.u.msi.data = msg.data;
1121
1122     kvm_add_routing_entry(s, &kroute);
1123
1124     return virq;
1125 }
1126
1127 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1128 {
1129     struct kvm_irq_routing_entry kroute;
1130
1131     if (!kvm_irqchip_in_kernel()) {
1132         return -ENOSYS;
1133     }
1134
1135     kroute.gsi = virq;
1136     kroute.type = KVM_IRQ_ROUTING_MSI;
1137     kroute.flags = 0;
1138     kroute.u.msi.address_lo = (uint32_t)msg.address;
1139     kroute.u.msi.address_hi = msg.address >> 32;
1140     kroute.u.msi.data = msg.data;
1141
1142     return kvm_update_routing_entry(s, &kroute);
1143 }
1144
1145 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1146 {
1147     struct kvm_irqfd irqfd = {
1148         .fd = fd,
1149         .gsi = virq,
1150         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1151     };
1152
1153     if (!kvm_irqfds_enabled()) {
1154         return -ENOSYS;
1155     }
1156
1157     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1158 }
1159
1160 #else /* !KVM_CAP_IRQ_ROUTING */
1161
1162 static void kvm_init_irq_routing(KVMState *s)
1163 {
1164 }
1165
1166 void kvm_irqchip_release_virq(KVMState *s, int virq)
1167 {
1168 }
1169
1170 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1171 {
1172     abort();
1173 }
1174
1175 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1176 {
1177     return -ENOSYS;
1178 }
1179
1180 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1181 {
1182     abort();
1183 }
1184
1185 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1186 {
1187     return -ENOSYS;
1188 }
1189 #endif /* !KVM_CAP_IRQ_ROUTING */
1190
1191 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1192 {
1193     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, true);
1194 }
1195
1196 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1197 {
1198     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, false);
1199 }
1200
1201 static int kvm_irqchip_create(KVMState *s)
1202 {
1203     QemuOptsList *list = qemu_find_opts("machine");
1204     int ret;
1205
1206     if (QTAILQ_EMPTY(&list->head) ||
1207         !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1208                            "kernel_irqchip", true) ||
1209         !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1210         return 0;
1211     }
1212
1213     ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1214     if (ret < 0) {
1215         fprintf(stderr, "Create kernel irqchip failed\n");
1216         return ret;
1217     }
1218
1219     kvm_kernel_irqchip = true;
1220     /* If we have an in-kernel IRQ chip then we must have asynchronous
1221      * interrupt delivery (though the reverse is not necessarily true)
1222      */
1223     kvm_async_interrupts_allowed = true;
1224
1225     kvm_init_irq_routing(s);
1226
1227     return 0;
1228 }
1229
1230 static int kvm_max_vcpus(KVMState *s)
1231 {
1232     int ret;
1233
1234     /* Find number of supported CPUs using the recommended
1235      * procedure from the kernel API documentation to cope with
1236      * older kernels that may be missing capabilities.
1237      */
1238     ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1239     if (ret) {
1240         return ret;
1241     }
1242     ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1243     if (ret) {
1244         return ret;
1245     }
1246
1247     return 4;
1248 }
1249
1250 int kvm_init(void)
1251 {
1252     static const char upgrade_note[] =
1253         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1254         "(see http://sourceforge.net/projects/kvm).\n";
1255     KVMState *s;
1256     const KVMCapabilityInfo *missing_cap;
1257     int ret;
1258     int i;
1259     int max_vcpus;
1260
1261     s = g_malloc0(sizeof(KVMState));
1262
1263     /*
1264      * On systems where the kernel can support different base page
1265      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1266      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1267      * page size for the system though.
1268      */
1269     assert(TARGET_PAGE_SIZE <= getpagesize());
1270
1271 #ifdef KVM_CAP_SET_GUEST_DEBUG
1272     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1273 #endif
1274     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1275         s->slots[i].slot = i;
1276     }
1277     s->vmfd = -1;
1278     s->fd = qemu_open("/dev/kvm", O_RDWR);
1279     if (s->fd == -1) {
1280         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1281         ret = -errno;
1282         goto err;
1283     }
1284
1285     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1286     if (ret < KVM_API_VERSION) {
1287         if (ret > 0) {
1288             ret = -EINVAL;
1289         }
1290         fprintf(stderr, "kvm version too old\n");
1291         goto err;
1292     }
1293
1294     if (ret > KVM_API_VERSION) {
1295         ret = -EINVAL;
1296         fprintf(stderr, "kvm version not supported\n");
1297         goto err;
1298     }
1299
1300     max_vcpus = kvm_max_vcpus(s);
1301     if (smp_cpus > max_vcpus) {
1302         ret = -EINVAL;
1303         fprintf(stderr, "Number of SMP cpus requested (%d) exceeds max cpus "
1304                 "supported by KVM (%d)\n", smp_cpus, max_vcpus);
1305         goto err;
1306     }
1307
1308     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1309     if (s->vmfd < 0) {
1310 #ifdef TARGET_S390X
1311         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1312                         "your host kernel command line\n");
1313 #endif
1314         ret = s->vmfd;
1315         goto err;
1316     }
1317
1318     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1319     if (!missing_cap) {
1320         missing_cap =
1321             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1322     }
1323     if (missing_cap) {
1324         ret = -EINVAL;
1325         fprintf(stderr, "kvm does not support %s\n%s",
1326                 missing_cap->name, upgrade_note);
1327         goto err;
1328     }
1329
1330     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1331
1332     s->broken_set_mem_region = 1;
1333     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1334     if (ret > 0) {
1335         s->broken_set_mem_region = 0;
1336     }
1337
1338 #ifdef KVM_CAP_VCPU_EVENTS
1339     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1340 #endif
1341
1342     s->robust_singlestep =
1343         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1344
1345 #ifdef KVM_CAP_DEBUGREGS
1346     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1347 #endif
1348
1349 #ifdef KVM_CAP_XSAVE
1350     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1351 #endif
1352
1353 #ifdef KVM_CAP_XCRS
1354     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1355 #endif
1356
1357 #ifdef KVM_CAP_PIT_STATE2
1358     s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1359 #endif
1360
1361 #ifdef KVM_CAP_IRQ_ROUTING
1362     s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1363 #endif
1364
1365     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1366
1367     s->irq_set_ioctl = KVM_IRQ_LINE;
1368     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1369         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1370     }
1371
1372     ret = kvm_arch_init(s);
1373     if (ret < 0) {
1374         goto err;
1375     }
1376
1377     ret = kvm_irqchip_create(s);
1378     if (ret < 0) {
1379         goto err;
1380     }
1381
1382     kvm_state = s;
1383     memory_listener_register(&kvm_memory_listener, &address_space_memory);
1384     memory_listener_register(&kvm_io_listener, &address_space_io);
1385
1386     s->many_ioeventfds = kvm_check_many_ioeventfds();
1387
1388     cpu_interrupt_handler = kvm_handle_interrupt;
1389
1390     return 0;
1391
1392 err:
1393     if (s->vmfd >= 0) {
1394         close(s->vmfd);
1395     }
1396     if (s->fd != -1) {
1397         close(s->fd);
1398     }
1399     g_free(s);
1400
1401     return ret;
1402 }
1403
1404 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1405                           uint32_t count)
1406 {
1407     int i;
1408     uint8_t *ptr = data;
1409
1410     for (i = 0; i < count; i++) {
1411         if (direction == KVM_EXIT_IO_IN) {
1412             switch (size) {
1413             case 1:
1414                 stb_p(ptr, cpu_inb(port));
1415                 break;
1416             case 2:
1417                 stw_p(ptr, cpu_inw(port));
1418                 break;
1419             case 4:
1420                 stl_p(ptr, cpu_inl(port));
1421                 break;
1422             }
1423         } else {
1424             switch (size) {
1425             case 1:
1426                 cpu_outb(port, ldub_p(ptr));
1427                 break;
1428             case 2:
1429                 cpu_outw(port, lduw_p(ptr));
1430                 break;
1431             case 4:
1432                 cpu_outl(port, ldl_p(ptr));
1433                 break;
1434             }
1435         }
1436
1437         ptr += size;
1438     }
1439 }
1440
1441 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1442 {
1443     CPUState *cpu = ENV_GET_CPU(env);
1444
1445     fprintf(stderr, "KVM internal error.");
1446     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1447         int i;
1448
1449         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1450         for (i = 0; i < run->internal.ndata; ++i) {
1451             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1452                     i, (uint64_t)run->internal.data[i]);
1453         }
1454     } else {
1455         fprintf(stderr, "\n");
1456     }
1457     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1458         fprintf(stderr, "emulation failure\n");
1459         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1460             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1461             return EXCP_INTERRUPT;
1462         }
1463     }
1464     /* FIXME: Should trigger a qmp message to let management know
1465      * something went wrong.
1466      */
1467     return -1;
1468 }
1469
1470 void kvm_flush_coalesced_mmio_buffer(void)
1471 {
1472     KVMState *s = kvm_state;
1473
1474     if (s->coalesced_flush_in_progress) {
1475         return;
1476     }
1477
1478     s->coalesced_flush_in_progress = true;
1479
1480     if (s->coalesced_mmio_ring) {
1481         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1482         while (ring->first != ring->last) {
1483             struct kvm_coalesced_mmio *ent;
1484
1485             ent = &ring->coalesced_mmio[ring->first];
1486
1487             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1488             smp_wmb();
1489             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1490         }
1491     }
1492
1493     s->coalesced_flush_in_progress = false;
1494 }
1495
1496 static void do_kvm_cpu_synchronize_state(void *arg)
1497 {
1498     CPUState *cpu = arg;
1499
1500     if (!cpu->kvm_vcpu_dirty) {
1501         kvm_arch_get_registers(cpu);
1502         cpu->kvm_vcpu_dirty = true;
1503     }
1504 }
1505
1506 void kvm_cpu_synchronize_state(CPUArchState *env)
1507 {
1508     CPUState *cpu = ENV_GET_CPU(env);
1509
1510     if (!cpu->kvm_vcpu_dirty) {
1511         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1512     }
1513 }
1514
1515 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1516 {
1517     CPUState *cpu = ENV_GET_CPU(env);
1518
1519     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1520     cpu->kvm_vcpu_dirty = false;
1521 }
1522
1523 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1524 {
1525     CPUState *cpu = ENV_GET_CPU(env);
1526
1527     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1528     cpu->kvm_vcpu_dirty = false;
1529 }
1530
1531 int kvm_cpu_exec(CPUArchState *env)
1532 {
1533     CPUState *cpu = ENV_GET_CPU(env);
1534     struct kvm_run *run = cpu->kvm_run;
1535     int ret, run_ret;
1536
1537     DPRINTF("kvm_cpu_exec()\n");
1538
1539     if (kvm_arch_process_async_events(cpu)) {
1540         env->exit_request = 0;
1541         return EXCP_HLT;
1542     }
1543
1544     do {
1545         if (cpu->kvm_vcpu_dirty) {
1546             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1547             cpu->kvm_vcpu_dirty = false;
1548         }
1549
1550         kvm_arch_pre_run(cpu, run);
1551         if (env->exit_request) {
1552             DPRINTF("interrupt exit requested\n");
1553             /*
1554              * KVM requires us to reenter the kernel after IO exits to complete
1555              * instruction emulation. This self-signal will ensure that we
1556              * leave ASAP again.
1557              */
1558             qemu_cpu_kick_self();
1559         }
1560         qemu_mutex_unlock_iothread();
1561
1562         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1563
1564         qemu_mutex_lock_iothread();
1565         kvm_arch_post_run(cpu, run);
1566
1567         if (run_ret < 0) {
1568             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1569                 DPRINTF("io window exit\n");
1570                 ret = EXCP_INTERRUPT;
1571                 break;
1572             }
1573             fprintf(stderr, "error: kvm run failed %s\n",
1574                     strerror(-run_ret));
1575             abort();
1576         }
1577
1578         switch (run->exit_reason) {
1579         case KVM_EXIT_IO:
1580             DPRINTF("handle_io\n");
1581             kvm_handle_io(run->io.port,
1582                           (uint8_t *)run + run->io.data_offset,
1583                           run->io.direction,
1584                           run->io.size,
1585                           run->io.count);
1586             ret = 0;
1587             break;
1588         case KVM_EXIT_MMIO:
1589             DPRINTF("handle_mmio\n");
1590             cpu_physical_memory_rw(run->mmio.phys_addr,
1591                                    run->mmio.data,
1592                                    run->mmio.len,
1593                                    run->mmio.is_write);
1594             ret = 0;
1595             break;
1596         case KVM_EXIT_IRQ_WINDOW_OPEN:
1597             DPRINTF("irq_window_open\n");
1598             ret = EXCP_INTERRUPT;
1599             break;
1600         case KVM_EXIT_SHUTDOWN:
1601             DPRINTF("shutdown\n");
1602             qemu_system_reset_request();
1603             ret = EXCP_INTERRUPT;
1604             break;
1605         case KVM_EXIT_UNKNOWN:
1606             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1607                     (uint64_t)run->hw.hardware_exit_reason);
1608             ret = -1;
1609             break;
1610         case KVM_EXIT_INTERNAL_ERROR:
1611             ret = kvm_handle_internal_error(env, run);
1612             break;
1613         default:
1614             DPRINTF("kvm_arch_handle_exit\n");
1615             ret = kvm_arch_handle_exit(cpu, run);
1616             break;
1617         }
1618     } while (ret == 0);
1619
1620     if (ret < 0) {
1621         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1622         vm_stop(RUN_STATE_INTERNAL_ERROR);
1623     }
1624
1625     env->exit_request = 0;
1626     return ret;
1627 }
1628
1629 int kvm_ioctl(KVMState *s, int type, ...)
1630 {
1631     int ret;
1632     void *arg;
1633     va_list ap;
1634
1635     va_start(ap, type);
1636     arg = va_arg(ap, void *);
1637     va_end(ap);
1638
1639     ret = ioctl(s->fd, type, arg);
1640     if (ret == -1) {
1641         ret = -errno;
1642     }
1643     return ret;
1644 }
1645
1646 int kvm_vm_ioctl(KVMState *s, int type, ...)
1647 {
1648     int ret;
1649     void *arg;
1650     va_list ap;
1651
1652     va_start(ap, type);
1653     arg = va_arg(ap, void *);
1654     va_end(ap);
1655
1656     ret = ioctl(s->vmfd, type, arg);
1657     if (ret == -1) {
1658         ret = -errno;
1659     }
1660     return ret;
1661 }
1662
1663 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1664 {
1665     int ret;
1666     void *arg;
1667     va_list ap;
1668
1669     va_start(ap, type);
1670     arg = va_arg(ap, void *);
1671     va_end(ap);
1672
1673     ret = ioctl(cpu->kvm_fd, type, arg);
1674     if (ret == -1) {
1675         ret = -errno;
1676     }
1677     return ret;
1678 }
1679
1680 int kvm_has_sync_mmu(void)
1681 {
1682     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1683 }
1684
1685 int kvm_has_vcpu_events(void)
1686 {
1687     return kvm_state->vcpu_events;
1688 }
1689
1690 int kvm_has_robust_singlestep(void)
1691 {
1692     return kvm_state->robust_singlestep;
1693 }
1694
1695 int kvm_has_debugregs(void)
1696 {
1697     return kvm_state->debugregs;
1698 }
1699
1700 int kvm_has_xsave(void)
1701 {
1702     return kvm_state->xsave;
1703 }
1704
1705 int kvm_has_xcrs(void)
1706 {
1707     return kvm_state->xcrs;
1708 }
1709
1710 int kvm_has_pit_state2(void)
1711 {
1712     return kvm_state->pit_state2;
1713 }
1714
1715 int kvm_has_many_ioeventfds(void)
1716 {
1717     if (!kvm_enabled()) {
1718         return 0;
1719     }
1720     return kvm_state->many_ioeventfds;
1721 }
1722
1723 int kvm_has_gsi_routing(void)
1724 {
1725 #ifdef KVM_CAP_IRQ_ROUTING
1726     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1727 #else
1728     return false;
1729 #endif
1730 }
1731
1732 int kvm_has_intx_set_mask(void)
1733 {
1734     return kvm_state->intx_set_mask;
1735 }
1736
1737 void *kvm_vmalloc(ram_addr_t size)
1738 {
1739 #ifdef TARGET_S390X
1740     void *mem;
1741
1742     mem = kvm_arch_vmalloc(size);
1743     if (mem) {
1744         return mem;
1745     }
1746 #endif
1747     return qemu_vmalloc(size);
1748 }
1749
1750 void kvm_setup_guest_memory(void *start, size_t size)
1751 {
1752 #ifdef CONFIG_VALGRIND_H
1753     VALGRIND_MAKE_MEM_DEFINED(start, size);
1754 #endif
1755     if (!kvm_has_sync_mmu()) {
1756         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1757
1758         if (ret) {
1759             perror("qemu_madvise");
1760             fprintf(stderr,
1761                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1762             exit(1);
1763         }
1764     }
1765 }
1766
1767 #ifdef KVM_CAP_SET_GUEST_DEBUG
1768 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1769                                                  target_ulong pc)
1770 {
1771     struct kvm_sw_breakpoint *bp;
1772
1773     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1774         if (bp->pc == pc) {
1775             return bp;
1776         }
1777     }
1778     return NULL;
1779 }
1780
1781 int kvm_sw_breakpoints_active(CPUState *cpu)
1782 {
1783     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1784 }
1785
1786 struct kvm_set_guest_debug_data {
1787     struct kvm_guest_debug dbg;
1788     CPUState *cpu;
1789     int err;
1790 };
1791
1792 static void kvm_invoke_set_guest_debug(void *data)
1793 {
1794     struct kvm_set_guest_debug_data *dbg_data = data;
1795
1796     dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1797                                    &dbg_data->dbg);
1798 }
1799
1800 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1801 {
1802     CPUState *cpu = ENV_GET_CPU(env);
1803     struct kvm_set_guest_debug_data data;
1804
1805     data.dbg.control = reinject_trap;
1806
1807     if (env->singlestep_enabled) {
1808         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1809     }
1810     kvm_arch_update_guest_debug(cpu, &data.dbg);
1811     data.cpu = cpu;
1812
1813     run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1814     return data.err;
1815 }
1816
1817 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1818                           target_ulong len, int type)
1819 {
1820     CPUState *current_cpu = ENV_GET_CPU(current_env);
1821     struct kvm_sw_breakpoint *bp;
1822     CPUArchState *env;
1823     int err;
1824
1825     if (type == GDB_BREAKPOINT_SW) {
1826         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1827         if (bp) {
1828             bp->use_count++;
1829             return 0;
1830         }
1831
1832         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1833         if (!bp) {
1834             return -ENOMEM;
1835         }
1836
1837         bp->pc = addr;
1838         bp->use_count = 1;
1839         err = kvm_arch_insert_sw_breakpoint(current_cpu, bp);
1840         if (err) {
1841             g_free(bp);
1842             return err;
1843         }
1844
1845         QTAILQ_INSERT_HEAD(&current_cpu->kvm_state->kvm_sw_breakpoints,
1846                           bp, entry);
1847     } else {
1848         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1849         if (err) {
1850             return err;
1851         }
1852     }
1853
1854     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1855         err = kvm_update_guest_debug(env, 0);
1856         if (err) {
1857             return err;
1858         }
1859     }
1860     return 0;
1861 }
1862
1863 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1864                           target_ulong len, int type)
1865 {
1866     CPUState *current_cpu = ENV_GET_CPU(current_env);
1867     struct kvm_sw_breakpoint *bp;
1868     CPUArchState *env;
1869     int err;
1870
1871     if (type == GDB_BREAKPOINT_SW) {
1872         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1873         if (!bp) {
1874             return -ENOENT;
1875         }
1876
1877         if (bp->use_count > 1) {
1878             bp->use_count--;
1879             return 0;
1880         }
1881
1882         err = kvm_arch_remove_sw_breakpoint(current_cpu, bp);
1883         if (err) {
1884             return err;
1885         }
1886
1887         QTAILQ_REMOVE(&current_cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
1888         g_free(bp);
1889     } else {
1890         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1891         if (err) {
1892             return err;
1893         }
1894     }
1895
1896     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1897         err = kvm_update_guest_debug(env, 0);
1898         if (err) {
1899             return err;
1900         }
1901     }
1902     return 0;
1903 }
1904
1905 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1906 {
1907     CPUState *current_cpu = ENV_GET_CPU(current_env);
1908     struct kvm_sw_breakpoint *bp, *next;
1909     KVMState *s = current_cpu->kvm_state;
1910     CPUArchState *env;
1911     CPUState *cpu;
1912
1913     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1914         if (kvm_arch_remove_sw_breakpoint(current_cpu, bp) != 0) {
1915             /* Try harder to find a CPU that currently sees the breakpoint. */
1916             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1917                 cpu = ENV_GET_CPU(env);
1918                 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
1919                     break;
1920                 }
1921             }
1922         }
1923         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
1924         g_free(bp);
1925     }
1926     kvm_arch_remove_all_hw_breakpoints();
1927
1928     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1929         kvm_update_guest_debug(env, 0);
1930     }
1931 }
1932
1933 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1934
1935 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1936 {
1937     return -EINVAL;
1938 }
1939
1940 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1941                           target_ulong len, int type)
1942 {
1943     return -EINVAL;
1944 }
1945
1946 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1947                           target_ulong len, int type)
1948 {
1949     return -EINVAL;
1950 }
1951
1952 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1953 {
1954 }
1955 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1956
1957 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1958 {
1959     CPUState *cpu = ENV_GET_CPU(env);
1960     struct kvm_signal_mask *sigmask;
1961     int r;
1962
1963     if (!sigset) {
1964         return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
1965     }
1966
1967     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1968
1969     sigmask->len = 8;
1970     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1971     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
1972     g_free(sigmask);
1973
1974     return r;
1975 }
1976
1977 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1978                            uint32_t size)
1979 {
1980     int ret;
1981     struct kvm_ioeventfd iofd;
1982
1983     iofd.datamatch = val;
1984     iofd.addr = addr;
1985     iofd.len = size;
1986     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1987     iofd.fd = fd;
1988
1989     if (!kvm_enabled()) {
1990         return -ENOSYS;
1991     }
1992
1993     if (!assign) {
1994         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1995     }
1996
1997     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1998
1999     if (ret < 0) {
2000         return -errno;
2001     }
2002
2003     return 0;
2004 }
2005
2006 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
2007 {
2008     struct kvm_ioeventfd kick = {
2009         .datamatch = val,
2010         .addr = addr,
2011         .len = 2,
2012         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
2013         .fd = fd,
2014     };
2015     int r;
2016     if (!kvm_enabled()) {
2017         return -ENOSYS;
2018     }
2019     if (!assign) {
2020         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
2021     }
2022     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
2023     if (r < 0) {
2024         return r;
2025     }
2026     return 0;
2027 }
2028
2029 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2030 {
2031     return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2032 }
2033
2034 int kvm_on_sigbus(int code, void *addr)
2035 {
2036     return kvm_arch_on_sigbus(code, addr);
2037 }