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(CPUArchState *env)
 218 {
 219     CPUState *cpu = ENV_GET_CPU(env);
 220     KVMState *s = kvm_state;
 221     long mmap_size;
 222     int ret;
 223
 224     DPRINTF("kvm_init_vcpu\n");
 225
 226     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 227     if (ret < 0) {
 228         DPRINTF("kvm_create_vcpu failed\n");
 229         goto err;
 230     }
 231
 232     cpu->kvm_fd = ret;
 233     cpu->kvm_state = s;
 234     cpu->kvm_vcpu_dirty = true;
 235
 236     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 237     if (mmap_size < 0) {
 238         ret = mmap_size;
 239         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 240         goto err;
 241     }
 242
 243     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 244                         cpu->kvm_fd, 0);
 245     if (cpu->kvm_run == MAP_FAILED) {
 246         ret = -errno;
 247         DPRINTF("mmap'ing vcpu state failed\n");
 248         goto err;
 249     }
 250
 251     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 252         s->coalesced_mmio_ring =
 253             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 254     }
 255
 256     ret = kvm_arch_init_vcpu(cpu);
 257     if (ret == 0) {
 258         qemu_register_reset(kvm_reset_vcpu, cpu);
 259         kvm_arch_reset_vcpu(cpu);
 260     }
 261 err:
 262     return ret;
 263 }
 264
 265 /*
 266  * dirty pages logging control
 267  */
 268
 269 static int kvm_mem_flags(KVMState *s, bool log_dirty)
 270 {
 271     return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
 272 }
 273
 274 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
 275 {
 276     KVMState *s = kvm_state;
 277     int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
 278     int old_flags;
 279
 280     old_flags = mem->flags;
 281
 282     flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
 283     mem->flags = flags;
 284
 285     /* If nothing changed effectively, no need to issue ioctl */
 286     if (s->migration_log) {
 287         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 288     }
 289
 290     if (flags == old_flags) {
 291         return 0;
 292     }
 293
 294     return kvm_set_user_memory_region(s, mem);
 295 }
 296
 297 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
 298                                       ram_addr_t size, bool log_dirty)
 299 {
 300     KVMState *s = kvm_state;
 301     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 302
 303     if (mem == NULL)  {
 304         fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 305                 TARGET_FMT_plx "\n", __func__, phys_addr,
 306                 (hwaddr)(phys_addr + size - 1));
 307         return -EINVAL;
 308     }
 309     return kvm_slot_dirty_pages_log_change(mem, log_dirty);
 310 }
 311
 312 static void kvm_log_start(MemoryListener *listener,
 313                           MemoryRegionSection *section)
 314 {
 315     int r;
 316
 317     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 318                                    section->size, true);
 319     if (r < 0) {
 320         abort();
 321     }
 322 }
 323
 324 static void kvm_log_stop(MemoryListener *listener,
 325                           MemoryRegionSection *section)
 326 {
 327     int r;
 328
 329     r = kvm_dirty_pages_log_change(section->offset_within_address_space,
 330                                    section->size, false);
 331     if (r < 0) {
 332         abort();
 333     }
 334 }
 335
 336 static int kvm_set_migration_log(int enable)
 337 {
 338     KVMState *s = kvm_state;
 339     KVMSlot *mem;
 340     int i, err;
 341
 342     s->migration_log = enable;
 343
 344     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 345         mem = &s->slots[i];
 346
 347         if (!mem->memory_size) {
 348             continue;
 349         }
 350         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 351             continue;
 352         }
 353         err = kvm_set_user_memory_region(s, mem);
 354         if (err) {
 355             return err;
 356         }
 357     }
 358     return 0;
 359 }
 360
 361 /* get kvm's dirty pages bitmap and update qemu's */
 362 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
 363                                          unsigned long *bitmap)
 364 {
 365     unsigned int i, j;
 366     unsigned long page_number, c;
 367     hwaddr addr, addr1;
 368     unsigned int len = ((section->size / getpagesize()) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
 369     unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
 370
 371     /*
 372      * bitmap-traveling is faster than memory-traveling (for addr...)
 373      * especially when most of the memory is not dirty.
 374      */
 375     for (i = 0; i < len; i++) {
 376         if (bitmap[i] != 0) {
 377             c = leul_to_cpu(bitmap[i]);
 378             do {
 379                 j = ffsl(c) - 1;
 380                 c &= ~(1ul << j);
 381                 page_number = (i * HOST_LONG_BITS + j) * hpratio;
 382                 addr1 = page_number * TARGET_PAGE_SIZE;
 383                 addr = section->offset_within_region + addr1;
 384                 memory_region_set_dirty(section->mr, addr,
 385                                         TARGET_PAGE_SIZE * hpratio);
 386             } while (c != 0);
 387         }
 388     }
 389     return 0;
 390 }
 391
 392 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 393
 394 /**
 395  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 396  * This function updates qemu's dirty bitmap using
 397  * memory_region_set_dirty().  This means all bits are set
 398  * to dirty.
 399  *
 400  * @start_add: start of logged region.
 401  * @end_addr: end of logged region.
 402  */
 403 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
 404 {
 405     KVMState *s = kvm_state;
 406     unsigned long size, allocated_size = 0;
 407     KVMDirtyLog d;
 408     KVMSlot *mem;
 409     int ret = 0;
 410     hwaddr start_addr = section->offset_within_address_space;
 411     hwaddr end_addr = start_addr + section->size;
 412
 413     d.dirty_bitmap = NULL;
 414     while (start_addr < end_addr) {
 415         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 416         if (mem == NULL) {
 417             break;
 418         }
 419
 420         /* XXX bad kernel interface alert
 421          * For dirty bitmap, kernel allocates array of size aligned to
 422          * bits-per-long.  But for case when the kernel is 64bits and
 423          * the userspace is 32bits, userspace can't align to the same
 424          * bits-per-long, since sizeof(long) is different between kernel
 425          * and user space.  This way, userspace will provide buffer which
 426          * may be 4 bytes less than the kernel will use, resulting in
 427          * userspace memory corruption (which is not detectable by valgrind
 428          * too, in most cases).
 429          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
 430          * a hope that sizeof(long) wont become >8 any time soon.
 431          */
 432         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
 433                      /*HOST_LONG_BITS*/ 64) / 8;
 434         if (!d.dirty_bitmap) {
 435             d.dirty_bitmap = g_malloc(size);
 436         } else if (size > allocated_size) {
 437             d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
 438         }
 439         allocated_size = size;
 440         memset(d.dirty_bitmap, 0, allocated_size);
 441
 442         d.slot = mem->slot;
 443
 444         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 445             DPRINTF("ioctl failed %d\n", errno);
 446             ret = -1;
 447             break;
 448         }
 449
 450         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
 451         start_addr = mem->start_addr + mem->memory_size;
 452     }
 453     g_free(d.dirty_bitmap);
 454
 455     return ret;
 456 }
 457
 458 static void kvm_coalesce_mmio_region(MemoryListener *listener,
 459                                      MemoryRegionSection *secion,
 460                                      hwaddr start, hwaddr size)
 461 {
 462     KVMState *s = kvm_state;
 463
 464     if (s->coalesced_mmio) {
 465         struct kvm_coalesced_mmio_zone zone;
 466
 467         zone.addr = start;
 468         zone.size = size;
 469         zone.pad = 0;
 470
 471         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 472     }
 473 }
 474
 475 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
 476                                        MemoryRegionSection *secion,
 477                                        hwaddr start, hwaddr size)
 478 {
 479     KVMState *s = kvm_state;
 480
 481     if (s->coalesced_mmio) {
 482         struct kvm_coalesced_mmio_zone zone;
 483
 484         zone.addr = start;
 485         zone.size = size;
 486         zone.pad = 0;
 487
 488         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 489     }
 490 }
 491
 492 int kvm_check_extension(KVMState *s, unsigned int extension)
 493 {
 494     int ret;
 495
 496     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 497     if (ret < 0) {
 498         ret = 0;
 499     }
 500
 501     return ret;
 502 }
 503
 504 static int kvm_check_many_ioeventfds(void)
 505 {
 506     /* Userspace can use ioeventfd for io notification.  This requires a host
 507      * that supports eventfd(2) and an I/O thread; since eventfd does not
 508      * support SIGIO it cannot interrupt the vcpu.
 509      *
 510      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 511      * can avoid creating too many ioeventfds.
 512      */
 513 #if defined(CONFIG_EVENTFD)
 514     int ioeventfds[7];
 515     int i, ret = 0;
 516     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 517         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 518         if (ioeventfds[i] < 0) {
 519             break;
 520         }
 521         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 522         if (ret < 0) {
 523             close(ioeventfds[i]);
 524             break;
 525         }
 526     }
 527
 528     /* Decide whether many devices are supported or not */
 529     ret = i == ARRAY_SIZE(ioeventfds);
 530
 531     while (i-- > 0) {
 532         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 533         close(ioeventfds[i]);
 534     }
 535     return ret;
 536 #else
 537     return 0;
 538 #endif
 539 }
 540
 541 static const KVMCapabilityInfo *
 542 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 543 {
 544     while (list->name) {
 545         if (!kvm_check_extension(s, list->value)) {
 546             return list;
 547         }
 548         list++;
 549     }
 550     return NULL;
 551 }
 552
 553 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
 554 {
 555     KVMState *s = kvm_state;
 556     KVMSlot *mem, old;
 557     int err;
 558     MemoryRegion *mr = section->mr;
 559     bool log_dirty = memory_region_is_logging(mr);
 560     hwaddr start_addr = section->offset_within_address_space;
 561     ram_addr_t size = section->size;
 562     void *ram = NULL;
 563     unsigned delta;
 564
 565     /* kvm works in page size chunks, but the function may be called
 566        with sub-page size and unaligned start address. */
 567     delta = TARGET_PAGE_ALIGN(size) - size;
 568     if (delta > size) {
 569         return;
 570     }
 571     start_addr += delta;
 572     size -= delta;
 573     size &= TARGET_PAGE_MASK;
 574     if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
 575         return;
 576     }
 577
 578     if (!memory_region_is_ram(mr)) {
 579         return;
 580     }
 581
 582     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
 583
 584     while (1) {
 585         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 586         if (!mem) {
 587             break;
 588         }
 589
 590         if (add && start_addr >= mem->start_addr &&
 591             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 592             (ram - start_addr == mem->ram - mem->start_addr)) {
 593             /* The new slot fits into the existing one and comes with
 594              * identical parameters - update flags and done. */
 595             kvm_slot_dirty_pages_log_change(mem, log_dirty);
 596             return;
 597         }
 598
 599         old = *mem;
 600
 601         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 602             kvm_physical_sync_dirty_bitmap(section);
 603         }
 604
 605         /* unregister the overlapping slot */
 606         mem->memory_size = 0;
 607         err = kvm_set_user_memory_region(s, mem);
 608         if (err) {
 609             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 610                     __func__, strerror(-err));
 611             abort();
 612         }
 613
 614         /* Workaround for older KVM versions: we can't join slots, even not by
 615          * unregistering the previous ones and then registering the larger
 616          * slot. We have to maintain the existing fragmentation. Sigh.
 617          *
 618          * This workaround assumes that the new slot starts at the same
 619          * address as the first existing one. If not or if some overlapping
 620          * slot comes around later, we will fail (not seen in practice so far)
 621          * - and actually require a recent KVM version. */
 622         if (s->broken_set_mem_region &&
 623             old.start_addr == start_addr && old.memory_size < size && add) {
 624             mem = kvm_alloc_slot(s);
 625             mem->memory_size = old.memory_size;
 626             mem->start_addr = old.start_addr;
 627             mem->ram = old.ram;
 628             mem->flags = kvm_mem_flags(s, log_dirty);
 629
 630             err = kvm_set_user_memory_region(s, mem);
 631             if (err) {
 632                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 633                         strerror(-err));
 634                 abort();
 635             }
 636
 637             start_addr += old.memory_size;
 638             ram += old.memory_size;
 639             size -= old.memory_size;
 640             continue;
 641         }
 642
 643         /* register prefix slot */
 644         if (old.start_addr < start_addr) {
 645             mem = kvm_alloc_slot(s);
 646             mem->memory_size = start_addr - old.start_addr;
 647             mem->start_addr = old.start_addr;
 648             mem->ram = old.ram;
 649             mem->flags =  kvm_mem_flags(s, log_dirty);
 650
 651             err = kvm_set_user_memory_region(s, mem);
 652             if (err) {
 653                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 654                         __func__, strerror(-err));
 655 #ifdef TARGET_PPC
 656                 fprintf(stderr, "%s: This is probably because your kernel's " \
 657                                 "PAGE_SIZE is too big. Please try to use 4k " \
 658                                 "PAGE_SIZE!\n", __func__);
 659 #endif
 660                 abort();
 661             }
 662         }
 663
 664         /* register suffix slot */
 665         if (old.start_addr + old.memory_size > start_addr + size) {
 666             ram_addr_t size_delta;
 667
 668             mem = kvm_alloc_slot(s);
 669             mem->start_addr = start_addr + size;
 670             size_delta = mem->start_addr - old.start_addr;
 671             mem->memory_size = old.memory_size - size_delta;
 672             mem->ram = old.ram + size_delta;
 673             mem->flags = kvm_mem_flags(s, log_dirty);
 674
 675             err = kvm_set_user_memory_region(s, mem);
 676             if (err) {
 677                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 678                         __func__, strerror(-err));
 679                 abort();
 680             }
 681         }
 682     }
 683
 684     /* in case the KVM bug workaround already "consumed" the new slot */
 685     if (!size) {
 686         return;
 687     }
 688     if (!add) {
 689         return;
 690     }
 691     mem = kvm_alloc_slot(s);
 692     mem->memory_size = size;
 693     mem->start_addr = start_addr;
 694     mem->ram = ram;
 695     mem->flags = kvm_mem_flags(s, log_dirty);
 696
 697     err = kvm_set_user_memory_region(s, mem);
 698     if (err) {
 699         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 700                 strerror(-err));
 701         abort();
 702     }
 703 }
 704
 705 static void kvm_region_add(MemoryListener *listener,
 706                            MemoryRegionSection *section)
 707 {
 708     kvm_set_phys_mem(section, true);
 709 }
 710
 711 static void kvm_region_del(MemoryListener *listener,
 712                            MemoryRegionSection *section)
 713 {
 714     kvm_set_phys_mem(section, false);
 715 }
 716
 717 static void kvm_log_sync(MemoryListener *listener,
 718                          MemoryRegionSection *section)
 719 {
 720     int r;
 721
 722     r = kvm_physical_sync_dirty_bitmap(section);
 723     if (r < 0) {
 724         abort();
 725     }
 726 }
 727
 728 static void kvm_log_global_start(struct MemoryListener *listener)
 729 {
 730     int r;
 731
 732     r = kvm_set_migration_log(1);
 733     assert(r >= 0);
 734 }
 735
 736 static void kvm_log_global_stop(struct MemoryListener *listener)
 737 {
 738     int r;
 739
 740     r = kvm_set_migration_log(0);
 741     assert(r >= 0);
 742 }
 743
 744 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
 745                                   MemoryRegionSection *section,
 746                                   bool match_data, uint64_t data,
 747                                   EventNotifier *e)
 748 {
 749     int fd = event_notifier_get_fd(e);
 750     int r;
 751
 752     assert(match_data && section->size <= 8);
 753
 754     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 755                                data, true, section->size);
 756     if (r < 0) {
 757         abort();
 758     }
 759 }
 760
 761 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
 762                                   MemoryRegionSection *section,
 763                                   bool match_data, uint64_t data,
 764                                   EventNotifier *e)
 765 {
 766     int fd = event_notifier_get_fd(e);
 767     int r;
 768
 769     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 770                                data, false, section->size);
 771     if (r < 0) {
 772         abort();
 773     }
 774 }
 775
 776 static void kvm_io_ioeventfd_add(MemoryListener *listener,
 777                                  MemoryRegionSection *section,
 778                                  bool match_data, uint64_t data,
 779                                  EventNotifier *e)
 780 {
 781     int fd = event_notifier_get_fd(e);
 782     int r;
 783
 784     assert(match_data && section->size == 2);
 785
 786     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 787                                    data, true);
 788     if (r < 0) {
 789         abort();
 790     }
 791 }
 792
 793 static void kvm_io_ioeventfd_del(MemoryListener *listener,
 794                                  MemoryRegionSection *section,
 795                                  bool match_data, uint64_t data,
 796                                  EventNotifier *e)
 797
 798 {
 799     int fd = event_notifier_get_fd(e);
 800     int r;
 801
 802     r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
 803                                    data, false);
 804     if (r < 0) {
 805         abort();
 806     }
 807 }
 808
 809 static MemoryListener kvm_memory_listener = {
 810     .region_add = kvm_region_add,
 811     .region_del = kvm_region_del,
 812     .log_start = kvm_log_start,
 813     .log_stop = kvm_log_stop,
 814     .log_sync = kvm_log_sync,
 815     .log_global_start = kvm_log_global_start,
 816     .log_global_stop = kvm_log_global_stop,
 817     .eventfd_add = kvm_mem_ioeventfd_add,
 818     .eventfd_del = kvm_mem_ioeventfd_del,
 819     .coalesced_mmio_add = kvm_coalesce_mmio_region,
 820     .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
 821     .priority = 10,
 822 };
 823
 824 static MemoryListener kvm_io_listener = {
 825     .eventfd_add = kvm_io_ioeventfd_add,
 826     .eventfd_del = kvm_io_ioeventfd_del,
 827     .priority = 10,
 828 };
 829
 830 static void kvm_handle_interrupt(CPUArchState *env, int mask)
 831 {
 832     CPUState *cpu = ENV_GET_CPU(env);
 833
 834     env->interrupt_request |= mask;
 835
 836     if (!qemu_cpu_is_self(cpu)) {
 837         qemu_cpu_kick(cpu);
 838     }
 839 }
 840
 841 int kvm_set_irq(KVMState *s, int irq, int level)
 842 {
 843     struct kvm_irq_level event;
 844     int ret;
 845
 846     assert(kvm_async_interrupts_enabled());
 847
 848     event.level = level;
 849     event.irq = irq;
 850     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
 851     if (ret < 0) {
 852         perror("kvm_set_irq");
 853         abort();
 854     }
 855
 856     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
 857 }
 858
 859 #ifdef KVM_CAP_IRQ_ROUTING
 860 typedef struct KVMMSIRoute {
 861     struct kvm_irq_routing_entry kroute;
 862     QTAILQ_ENTRY(KVMMSIRoute) entry;
 863 } KVMMSIRoute;
 864
 865 static void set_gsi(KVMState *s, unsigned int gsi)
 866 {
 867     s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
 868 }
 869
 870 static void clear_gsi(KVMState *s, unsigned int gsi)
 871 {
 872     s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
 873 }
 874
 875 static void kvm_init_irq_routing(KVMState *s)
 876 {
 877     int gsi_count, i;
 878
 879     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
 880     if (gsi_count > 0) {
 881         unsigned int gsi_bits, i;
 882
 883         /* Round up so we can search ints using ffs */
 884         gsi_bits = ALIGN(gsi_count, 32);
 885         s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
 886         s->gsi_count = gsi_count;
 887
 888         /* Mark any over-allocated bits as already in use */
 889         for (i = gsi_count; i < gsi_bits; i++) {
 890             set_gsi(s, i);
 891         }
 892     }
 893
 894     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
 895     s->nr_allocated_irq_routes = 0;
 896
 897     if (!s->direct_msi) {
 898         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
 899             QTAILQ_INIT(&s->msi_hashtab[i]);
 900         }
 901     }
 902
 903     kvm_arch_init_irq_routing(s);
 904 }
 905
 906 static void kvm_irqchip_commit_routes(KVMState *s)
 907 {
 908     int ret;
 909
 910     s->irq_routes->flags = 0;
 911     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
 912     assert(ret == 0);
 913 }
 914
 915 static void kvm_add_routing_entry(KVMState *s,
 916                                   struct kvm_irq_routing_entry *entry)
 917 {
 918     struct kvm_irq_routing_entry *new;
 919     int n, size;
 920
 921     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
 922         n = s->nr_allocated_irq_routes * 2;
 923         if (n < 64) {
 924             n = 64;
 925         }
 926         size = sizeof(struct kvm_irq_routing);
 927         size += n * sizeof(*new);
 928         s->irq_routes = g_realloc(s->irq_routes, size);
 929         s->nr_allocated_irq_routes = n;
 930     }
 931     n = s->irq_routes->nr++;
 932     new = &s->irq_routes->entries[n];
 933     memset(new, 0, sizeof(*new));
 934     new->gsi = entry->gsi;
 935     new->type = entry->type;
 936     new->flags = entry->flags;
 937     new->u = entry->u;
 938
 939     set_gsi(s, entry->gsi);
 940
 941     kvm_irqchip_commit_routes(s);
 942 }
 943
 944 static int kvm_update_routing_entry(KVMState *s,
 945                                     struct kvm_irq_routing_entry *new_entry)
 946 {
 947     struct kvm_irq_routing_entry *entry;
 948     int n;
 949
 950     for (n = 0; n < s->irq_routes->nr; n++) {
 951         entry = &s->irq_routes->entries[n];
 952         if (entry->gsi != new_entry->gsi) {
 953             continue;
 954         }
 955
 956         entry->type = new_entry->type;
 957         entry->flags = new_entry->flags;
 958         entry->u = new_entry->u;
 959
 960         kvm_irqchip_commit_routes(s);
 961
 962         return 0;
 963     }
 964
 965     return -ESRCH;
 966 }
 967
 968 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
 969 {
 970     struct kvm_irq_routing_entry e;
 971
 972     assert(pin < s->gsi_count);
 973
 974     e.gsi = irq;
 975     e.type = KVM_IRQ_ROUTING_IRQCHIP;
 976     e.flags = 0;
 977     e.u.irqchip.irqchip = irqchip;
 978     e.u.irqchip.pin = pin;
 979     kvm_add_routing_entry(s, &e);
 980 }
 981
 982 void kvm_irqchip_release_virq(KVMState *s, int virq)
 983 {
 984     struct kvm_irq_routing_entry *e;
 985     int i;
 986
 987     for (i = 0; i < s->irq_routes->nr; i++) {
 988         e = &s->irq_routes->entries[i];
 989         if (e->gsi == virq) {
 990             s->irq_routes->nr--;
 991             *e = s->irq_routes->entries[s->irq_routes->nr];
 992         }
 993     }
 994     clear_gsi(s, virq);
 995 }
 996
 997 static unsigned int kvm_hash_msi(uint32_t data)
 998 {
 999     /* This is optimized for IA32 MSI layout. However, no other arch shall
1000      * repeat the mistake of not providing a direct MSI injection API. */
1001     return data & 0xff;
1002 }
1003
1004 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1005 {
1006     KVMMSIRoute *route, *next;
1007     unsigned int hash;
1008
1009     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1010         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1011             kvm_irqchip_release_virq(s, route->kroute.gsi);
1012             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1013             g_free(route);
1014         }
1015     }
1016 }
1017
1018 static int kvm_irqchip_get_virq(KVMState *s)
1019 {
1020     uint32_t *word = s->used_gsi_bitmap;
1021     int max_words = ALIGN(s->gsi_count, 32) / 32;
1022     int i, bit;
1023     bool retry = true;
1024
1025 again:
1026     /* Return the lowest unused GSI in the bitmap */
1027     for (i = 0; i < max_words; i++) {
1028         bit = ffs(~word[i]);
1029         if (!bit) {
1030             continue;
1031         }
1032
1033         return bit - 1 + i * 32;
1034     }
1035     if (!s->direct_msi && retry) {
1036         retry = false;
1037         kvm_flush_dynamic_msi_routes(s);
1038         goto again;
1039     }
1040     return -ENOSPC;
1041
1042 }
1043
1044 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1045 {
1046     unsigned int hash = kvm_hash_msi(msg.data);
1047     KVMMSIRoute *route;
1048
1049     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1050         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1051             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1052             route->kroute.u.msi.data == msg.data) {
1053             return route;
1054         }
1055     }
1056     return NULL;
1057 }
1058
1059 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1060 {
1061     struct kvm_msi msi;
1062     KVMMSIRoute *route;
1063
1064     if (s->direct_msi) {
1065         msi.address_lo = (uint32_t)msg.address;
1066         msi.address_hi = msg.address >> 32;
1067         msi.data = msg.data;
1068         msi.flags = 0;
1069         memset(msi.pad, 0, sizeof(msi.pad));
1070
1071         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1072     }
1073
1074     route = kvm_lookup_msi_route(s, msg);
1075     if (!route) {
1076         int virq;
1077
1078         virq = kvm_irqchip_get_virq(s);
1079         if (virq < 0) {
1080             return virq;
1081         }
1082
1083         route = g_malloc(sizeof(KVMMSIRoute));
1084         route->kroute.gsi = virq;
1085         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1086         route->kroute.flags = 0;
1087         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1088         route->kroute.u.msi.address_hi = msg.address >> 32;
1089         route->kroute.u.msi.data = msg.data;
1090
1091         kvm_add_routing_entry(s, &route->kroute);
1092
1093         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1094                            entry);
1095     }
1096
1097     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1098
1099     return kvm_set_irq(s, route->kroute.gsi, 1);
1100 }
1101
1102 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1103 {
1104     struct kvm_irq_routing_entry kroute;
1105     int virq;
1106
1107     if (!kvm_gsi_routing_enabled()) {
1108         return -ENOSYS;
1109     }
1110
1111     virq = kvm_irqchip_get_virq(s);
1112     if (virq < 0) {
1113         return virq;
1114     }
1115
1116     kroute.gsi = virq;
1117     kroute.type = KVM_IRQ_ROUTING_MSI;
1118     kroute.flags = 0;
1119     kroute.u.msi.address_lo = (uint32_t)msg.address;
1120     kroute.u.msi.address_hi = msg.address >> 32;
1121     kroute.u.msi.data = msg.data;
1122
1123     kvm_add_routing_entry(s, &kroute);
1124
1125     return virq;
1126 }
1127
1128 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1129 {
1130     struct kvm_irq_routing_entry kroute;
1131
1132     if (!kvm_irqchip_in_kernel()) {
1133         return -ENOSYS;
1134     }
1135
1136     kroute.gsi = virq;
1137     kroute.type = KVM_IRQ_ROUTING_MSI;
1138     kroute.flags = 0;
1139     kroute.u.msi.address_lo = (uint32_t)msg.address;
1140     kroute.u.msi.address_hi = msg.address >> 32;
1141     kroute.u.msi.data = msg.data;
1142
1143     return kvm_update_routing_entry(s, &kroute);
1144 }
1145
1146 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1147 {
1148     struct kvm_irqfd irqfd = {
1149         .fd = fd,
1150         .gsi = virq,
1151         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1152     };
1153
1154     if (!kvm_irqfds_enabled()) {
1155         return -ENOSYS;
1156     }
1157
1158     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1159 }
1160
1161 #else /* !KVM_CAP_IRQ_ROUTING */
1162
1163 static void kvm_init_irq_routing(KVMState *s)
1164 {
1165 }
1166
1167 void kvm_irqchip_release_virq(KVMState *s, int virq)
1168 {
1169 }
1170
1171 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1172 {
1173     abort();
1174 }
1175
1176 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1177 {
1178     return -ENOSYS;
1179 }
1180
1181 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1182 {
1183     abort();
1184 }
1185 #endif /* !KVM_CAP_IRQ_ROUTING */
1186
1187 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1188 {
1189     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, true);
1190 }
1191
1192 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1193 {
1194     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, false);
1195 }
1196
1197 static int kvm_irqchip_create(KVMState *s)
1198 {
1199     QemuOptsList *list = qemu_find_opts("machine");
1200     int ret;
1201
1202     if (QTAILQ_EMPTY(&list->head) ||
1203         !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1204                            "kernel_irqchip", true) ||
1205         !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1206         return 0;
1207     }
1208
1209     ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1210     if (ret < 0) {
1211         fprintf(stderr, "Create kernel irqchip failed\n");
1212         return ret;
1213     }
1214
1215     kvm_kernel_irqchip = true;
1216     /* If we have an in-kernel IRQ chip then we must have asynchronous
1217      * interrupt delivery (though the reverse is not necessarily true)
1218      */
1219     kvm_async_interrupts_allowed = true;
1220
1221     kvm_init_irq_routing(s);
1222
1223     return 0;
1224 }
1225
1226 static int kvm_max_vcpus(KVMState *s)
1227 {
1228     int ret;
1229
1230     /* Find number of supported CPUs using the recommended
1231      * procedure from the kernel API documentation to cope with
1232      * older kernels that may be missing capabilities.
1233      */
1234     ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1235     if (ret) {
1236         return ret;
1237     }
1238     ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1239     if (ret) {
1240         return ret;
1241     }
1242
1243     return 4;
1244 }
1245
1246 int kvm_init(void)
1247 {
1248     static const char upgrade_note[] =
1249         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1250         "(see http://sourceforge.net/projects/kvm).\n";
1251     KVMState *s;
1252     const KVMCapabilityInfo *missing_cap;
1253     int ret;
1254     int i;
1255     int max_vcpus;
1256
1257     s = g_malloc0(sizeof(KVMState));
1258
1259     /*
1260      * On systems where the kernel can support different base page
1261      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1262      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1263      * page size for the system though.
1264      */
1265     assert(TARGET_PAGE_SIZE <= getpagesize());
1266
1267 #ifdef KVM_CAP_SET_GUEST_DEBUG
1268     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1269 #endif
1270     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1271         s->slots[i].slot = i;
1272     }
1273     s->vmfd = -1;
1274     s->fd = qemu_open("/dev/kvm", O_RDWR);
1275     if (s->fd == -1) {
1276         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1277         ret = -errno;
1278         goto err;
1279     }
1280
1281     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1282     if (ret < KVM_API_VERSION) {
1283         if (ret > 0) {
1284             ret = -EINVAL;
1285         }
1286         fprintf(stderr, "kvm version too old\n");
1287         goto err;
1288     }
1289
1290     if (ret > KVM_API_VERSION) {
1291         ret = -EINVAL;
1292         fprintf(stderr, "kvm version not supported\n");
1293         goto err;
1294     }
1295
1296     max_vcpus = kvm_max_vcpus(s);
1297     if (smp_cpus > max_vcpus) {
1298         ret = -EINVAL;
1299         fprintf(stderr, "Number of SMP cpus requested (%d) exceeds max cpus "
1300                 "supported by KVM (%d)\n", smp_cpus, max_vcpus);
1301         goto err;
1302     }
1303
1304     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1305     if (s->vmfd < 0) {
1306 #ifdef TARGET_S390X
1307         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1308                         "your host kernel command line\n");
1309 #endif
1310         ret = s->vmfd;
1311         goto err;
1312     }
1313
1314     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1315     if (!missing_cap) {
1316         missing_cap =
1317             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1318     }
1319     if (missing_cap) {
1320         ret = -EINVAL;
1321         fprintf(stderr, "kvm does not support %s\n%s",
1322                 missing_cap->name, upgrade_note);
1323         goto err;
1324     }
1325
1326     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1327
1328     s->broken_set_mem_region = 1;
1329     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1330     if (ret > 0) {
1331         s->broken_set_mem_region = 0;
1332     }
1333
1334 #ifdef KVM_CAP_VCPU_EVENTS
1335     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1336 #endif
1337
1338     s->robust_singlestep =
1339         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1340
1341 #ifdef KVM_CAP_DEBUGREGS
1342     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1343 #endif
1344
1345 #ifdef KVM_CAP_XSAVE
1346     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1347 #endif
1348
1349 #ifdef KVM_CAP_XCRS
1350     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1351 #endif
1352
1353 #ifdef KVM_CAP_PIT_STATE2
1354     s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1355 #endif
1356
1357 #ifdef KVM_CAP_IRQ_ROUTING
1358     s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1359 #endif
1360
1361     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1362
1363     s->irq_set_ioctl = KVM_IRQ_LINE;
1364     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1365         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1366     }
1367
1368     ret = kvm_arch_init(s);
1369     if (ret < 0) {
1370         goto err;
1371     }
1372
1373     ret = kvm_irqchip_create(s);
1374     if (ret < 0) {
1375         goto err;
1376     }
1377
1378     kvm_state = s;
1379     memory_listener_register(&kvm_memory_listener, &address_space_memory);
1380     memory_listener_register(&kvm_io_listener, &address_space_io);
1381
1382     s->many_ioeventfds = kvm_check_many_ioeventfds();
1383
1384     cpu_interrupt_handler = kvm_handle_interrupt;
1385
1386     return 0;
1387
1388 err:
1389     if (s->vmfd >= 0) {
1390         close(s->vmfd);
1391     }
1392     if (s->fd != -1) {
1393         close(s->fd);
1394     }
1395     g_free(s);
1396
1397     return ret;
1398 }
1399
1400 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1401                           uint32_t count)
1402 {
1403     int i;
1404     uint8_t *ptr = data;
1405
1406     for (i = 0; i < count; i++) {
1407         if (direction == KVM_EXIT_IO_IN) {
1408             switch (size) {
1409             case 1:
1410                 stb_p(ptr, cpu_inb(port));
1411                 break;
1412             case 2:
1413                 stw_p(ptr, cpu_inw(port));
1414                 break;
1415             case 4:
1416                 stl_p(ptr, cpu_inl(port));
1417                 break;
1418             }
1419         } else {
1420             switch (size) {
1421             case 1:
1422                 cpu_outb(port, ldub_p(ptr));
1423                 break;
1424             case 2:
1425                 cpu_outw(port, lduw_p(ptr));
1426                 break;
1427             case 4:
1428                 cpu_outl(port, ldl_p(ptr));
1429                 break;
1430             }
1431         }
1432
1433         ptr += size;
1434     }
1435 }
1436
1437 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1438 {
1439     CPUState *cpu = ENV_GET_CPU(env);
1440
1441     fprintf(stderr, "KVM internal error.");
1442     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1443         int i;
1444
1445         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1446         for (i = 0; i < run->internal.ndata; ++i) {
1447             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1448                     i, (uint64_t)run->internal.data[i]);
1449         }
1450     } else {
1451         fprintf(stderr, "\n");
1452     }
1453     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1454         fprintf(stderr, "emulation failure\n");
1455         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1456             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1457             return EXCP_INTERRUPT;
1458         }
1459     }
1460     /* FIXME: Should trigger a qmp message to let management know
1461      * something went wrong.
1462      */
1463     return -1;
1464 }
1465
1466 void kvm_flush_coalesced_mmio_buffer(void)
1467 {
1468     KVMState *s = kvm_state;
1469
1470     if (s->coalesced_flush_in_progress) {
1471         return;
1472     }
1473
1474     s->coalesced_flush_in_progress = true;
1475
1476     if (s->coalesced_mmio_ring) {
1477         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1478         while (ring->first != ring->last) {
1479             struct kvm_coalesced_mmio *ent;
1480
1481             ent = &ring->coalesced_mmio[ring->first];
1482
1483             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1484             smp_wmb();
1485             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1486         }
1487     }
1488
1489     s->coalesced_flush_in_progress = false;
1490 }
1491
1492 static void do_kvm_cpu_synchronize_state(void *arg)
1493 {
1494     CPUState *cpu = arg;
1495
1496     if (!cpu->kvm_vcpu_dirty) {
1497         kvm_arch_get_registers(cpu);
1498         cpu->kvm_vcpu_dirty = true;
1499     }
1500 }
1501
1502 void kvm_cpu_synchronize_state(CPUArchState *env)
1503 {
1504     CPUState *cpu = ENV_GET_CPU(env);
1505
1506     if (!cpu->kvm_vcpu_dirty) {
1507         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1508     }
1509 }
1510
1511 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1512 {
1513     CPUState *cpu = ENV_GET_CPU(env);
1514
1515     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1516     cpu->kvm_vcpu_dirty = false;
1517 }
1518
1519 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1520 {
1521     CPUState *cpu = ENV_GET_CPU(env);
1522
1523     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1524     cpu->kvm_vcpu_dirty = false;
1525 }
1526
1527 int kvm_cpu_exec(CPUArchState *env)
1528 {
1529     CPUState *cpu = ENV_GET_CPU(env);
1530     struct kvm_run *run = cpu->kvm_run;
1531     int ret, run_ret;
1532
1533     DPRINTF("kvm_cpu_exec()\n");
1534
1535     if (kvm_arch_process_async_events(cpu)) {
1536         env->exit_request = 0;
1537         return EXCP_HLT;
1538     }
1539
1540     do {
1541         if (cpu->kvm_vcpu_dirty) {
1542             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1543             cpu->kvm_vcpu_dirty = false;
1544         }
1545
1546         kvm_arch_pre_run(cpu, run);
1547         if (env->exit_request) {
1548             DPRINTF("interrupt exit requested\n");
1549             /*
1550              * KVM requires us to reenter the kernel after IO exits to complete
1551              * instruction emulation. This self-signal will ensure that we
1552              * leave ASAP again.
1553              */
1554             qemu_cpu_kick_self();
1555         }
1556         qemu_mutex_unlock_iothread();
1557
1558         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1559
1560         qemu_mutex_lock_iothread();
1561         kvm_arch_post_run(cpu, run);
1562
1563         if (run_ret < 0) {
1564             if (run_ret == -EINTR || run_ret == -EAGAIN) {
1565                 DPRINTF("io window exit\n");
1566                 ret = EXCP_INTERRUPT;
1567                 break;
1568             }
1569             fprintf(stderr, "error: kvm run failed %s\n",
1570                     strerror(-run_ret));
1571             abort();
1572         }
1573
1574         switch (run->exit_reason) {
1575         case KVM_EXIT_IO:
1576             DPRINTF("handle_io\n");
1577             kvm_handle_io(run->io.port,
1578                           (uint8_t *)run + run->io.data_offset,
1579                           run->io.direction,
1580                           run->io.size,
1581                           run->io.count);
1582             ret = 0;
1583             break;
1584         case KVM_EXIT_MMIO:
1585             DPRINTF("handle_mmio\n");
1586             cpu_physical_memory_rw(run->mmio.phys_addr,
1587                                    run->mmio.data,
1588                                    run->mmio.len,
1589                                    run->mmio.is_write);
1590             ret = 0;
1591             break;
1592         case KVM_EXIT_IRQ_WINDOW_OPEN:
1593             DPRINTF("irq_window_open\n");
1594             ret = EXCP_INTERRUPT;
1595             break;
1596         case KVM_EXIT_SHUTDOWN:
1597             DPRINTF("shutdown\n");
1598             qemu_system_reset_request();
1599             ret = EXCP_INTERRUPT;
1600             break;
1601         case KVM_EXIT_UNKNOWN:
1602             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1603                     (uint64_t)run->hw.hardware_exit_reason);
1604             ret = -1;
1605             break;
1606         case KVM_EXIT_INTERNAL_ERROR:
1607             ret = kvm_handle_internal_error(env, run);
1608             break;
1609         default:
1610             DPRINTF("kvm_arch_handle_exit\n");
1611             ret = kvm_arch_handle_exit(cpu, run);
1612             break;
1613         }
1614     } while (ret == 0);
1615
1616     if (ret < 0) {
1617         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1618         vm_stop(RUN_STATE_INTERNAL_ERROR);
1619     }
1620
1621     env->exit_request = 0;
1622     return ret;
1623 }
1624
1625 int kvm_ioctl(KVMState *s, int type, ...)
1626 {
1627     int ret;
1628     void *arg;
1629     va_list ap;
1630
1631     va_start(ap, type);
1632     arg = va_arg(ap, void *);
1633     va_end(ap);
1634
1635     ret = ioctl(s->fd, type, arg);
1636     if (ret == -1) {
1637         ret = -errno;
1638     }
1639     return ret;
1640 }
1641
1642 int kvm_vm_ioctl(KVMState *s, int type, ...)
1643 {
1644     int ret;
1645     void *arg;
1646     va_list ap;
1647
1648     va_start(ap, type);
1649     arg = va_arg(ap, void *);
1650     va_end(ap);
1651
1652     ret = ioctl(s->vmfd, type, arg);
1653     if (ret == -1) {
1654         ret = -errno;
1655     }
1656     return ret;
1657 }
1658
1659 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1660 {
1661     int ret;
1662     void *arg;
1663     va_list ap;
1664
1665     va_start(ap, type);
1666     arg = va_arg(ap, void *);
1667     va_end(ap);
1668
1669     ret = ioctl(cpu->kvm_fd, type, arg);
1670     if (ret == -1) {
1671         ret = -errno;
1672     }
1673     return ret;
1674 }
1675
1676 int kvm_has_sync_mmu(void)
1677 {
1678     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1679 }
1680
1681 int kvm_has_vcpu_events(void)
1682 {
1683     return kvm_state->vcpu_events;
1684 }
1685
1686 int kvm_has_robust_singlestep(void)
1687 {
1688     return kvm_state->robust_singlestep;
1689 }
1690
1691 int kvm_has_debugregs(void)
1692 {
1693     return kvm_state->debugregs;
1694 }
1695
1696 int kvm_has_xsave(void)
1697 {
1698     return kvm_state->xsave;
1699 }
1700
1701 int kvm_has_xcrs(void)
1702 {
1703     return kvm_state->xcrs;
1704 }
1705
1706 int kvm_has_pit_state2(void)
1707 {
1708     return kvm_state->pit_state2;
1709 }
1710
1711 int kvm_has_many_ioeventfds(void)
1712 {
1713     if (!kvm_enabled()) {
1714         return 0;
1715     }
1716     return kvm_state->many_ioeventfds;
1717 }
1718
1719 int kvm_has_gsi_routing(void)
1720 {
1721 #ifdef KVM_CAP_IRQ_ROUTING
1722     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1723 #else
1724     return false;
1725 #endif
1726 }
1727
1728 int kvm_has_intx_set_mask(void)
1729 {
1730     return kvm_state->intx_set_mask;
1731 }
1732
1733 void *kvm_vmalloc(ram_addr_t size)
1734 {
1735 #ifdef TARGET_S390X
1736     void *mem;
1737
1738     mem = kvm_arch_vmalloc(size);
1739     if (mem) {
1740         return mem;
1741     }
1742 #endif
1743     return qemu_vmalloc(size);
1744 }
1745
1746 void kvm_setup_guest_memory(void *start, size_t size)
1747 {
1748 #ifdef CONFIG_VALGRIND_H
1749     VALGRIND_MAKE_MEM_DEFINED(start, size);
1750 #endif
1751     if (!kvm_has_sync_mmu()) {
1752         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1753
1754         if (ret) {
1755             perror("qemu_madvise");
1756             fprintf(stderr,
1757                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1758             exit(1);
1759         }
1760     }
1761 }
1762
1763 #ifdef KVM_CAP_SET_GUEST_DEBUG
1764 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1765                                                  target_ulong pc)
1766 {
1767     struct kvm_sw_breakpoint *bp;
1768
1769     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1770         if (bp->pc == pc) {
1771             return bp;
1772         }
1773     }
1774     return NULL;
1775 }
1776
1777 int kvm_sw_breakpoints_active(CPUState *cpu)
1778 {
1779     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1780 }
1781
1782 struct kvm_set_guest_debug_data {
1783     struct kvm_guest_debug dbg;
1784     CPUState *cpu;
1785     int err;
1786 };
1787
1788 static void kvm_invoke_set_guest_debug(void *data)
1789 {
1790     struct kvm_set_guest_debug_data *dbg_data = data;
1791
1792     dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1793                                    &dbg_data->dbg);
1794 }
1795
1796 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1797 {
1798     CPUState *cpu = ENV_GET_CPU(env);
1799     struct kvm_set_guest_debug_data data;
1800
1801     data.dbg.control = reinject_trap;
1802
1803     if (env->singlestep_enabled) {
1804         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1805     }
1806     kvm_arch_update_guest_debug(cpu, &data.dbg);
1807     data.cpu = cpu;
1808
1809     run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1810     return data.err;
1811 }
1812
1813 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1814                           target_ulong len, int type)
1815 {
1816     CPUState *current_cpu = ENV_GET_CPU(current_env);
1817     struct kvm_sw_breakpoint *bp;
1818     CPUArchState *env;
1819     int err;
1820
1821     if (type == GDB_BREAKPOINT_SW) {
1822         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1823         if (bp) {
1824             bp->use_count++;
1825             return 0;
1826         }
1827
1828         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1829         if (!bp) {
1830             return -ENOMEM;
1831         }
1832
1833         bp->pc = addr;
1834         bp->use_count = 1;
1835         err = kvm_arch_insert_sw_breakpoint(current_cpu, bp);
1836         if (err) {
1837             g_free(bp);
1838             return err;
1839         }
1840
1841         QTAILQ_INSERT_HEAD(&current_cpu->kvm_state->kvm_sw_breakpoints,
1842                           bp, entry);
1843     } else {
1844         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1845         if (err) {
1846             return err;
1847         }
1848     }
1849
1850     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1851         err = kvm_update_guest_debug(env, 0);
1852         if (err) {
1853             return err;
1854         }
1855     }
1856     return 0;
1857 }
1858
1859 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1860                           target_ulong len, int type)
1861 {
1862     CPUState *current_cpu = ENV_GET_CPU(current_env);
1863     struct kvm_sw_breakpoint *bp;
1864     CPUArchState *env;
1865     int err;
1866
1867     if (type == GDB_BREAKPOINT_SW) {
1868         bp = kvm_find_sw_breakpoint(current_cpu, addr);
1869         if (!bp) {
1870             return -ENOENT;
1871         }
1872
1873         if (bp->use_count > 1) {
1874             bp->use_count--;
1875             return 0;
1876         }
1877
1878         err = kvm_arch_remove_sw_breakpoint(current_cpu, bp);
1879         if (err) {
1880             return err;
1881         }
1882
1883         QTAILQ_REMOVE(&current_cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
1884         g_free(bp);
1885     } else {
1886         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1887         if (err) {
1888             return err;
1889         }
1890     }
1891
1892     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1893         err = kvm_update_guest_debug(env, 0);
1894         if (err) {
1895             return err;
1896         }
1897     }
1898     return 0;
1899 }
1900
1901 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1902 {
1903     CPUState *current_cpu = ENV_GET_CPU(current_env);
1904     struct kvm_sw_breakpoint *bp, *next;
1905     KVMState *s = current_cpu->kvm_state;
1906     CPUArchState *env;
1907     CPUState *cpu;
1908
1909     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1910         if (kvm_arch_remove_sw_breakpoint(current_cpu, bp) != 0) {
1911             /* Try harder to find a CPU that currently sees the breakpoint. */
1912             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1913                 cpu = ENV_GET_CPU(env);
1914                 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
1915                     break;
1916                 }
1917             }
1918         }
1919         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
1920         g_free(bp);
1921     }
1922     kvm_arch_remove_all_hw_breakpoints();
1923
1924     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1925         kvm_update_guest_debug(env, 0);
1926     }
1927 }
1928
1929 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1930
1931 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1932 {
1933     return -EINVAL;
1934 }
1935
1936 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1937                           target_ulong len, int type)
1938 {
1939     return -EINVAL;
1940 }
1941
1942 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1943                           target_ulong len, int type)
1944 {
1945     return -EINVAL;
1946 }
1947
1948 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1949 {
1950 }
1951 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1952
1953 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1954 {
1955     CPUState *cpu = ENV_GET_CPU(env);
1956     struct kvm_signal_mask *sigmask;
1957     int r;
1958
1959     if (!sigset) {
1960         return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
1961     }
1962
1963     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1964
1965     sigmask->len = 8;
1966     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1967     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
1968     g_free(sigmask);
1969
1970     return r;
1971 }
1972
1973 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1974                            uint32_t size)
1975 {
1976     int ret;
1977     struct kvm_ioeventfd iofd;
1978
1979     iofd.datamatch = val;
1980     iofd.addr = addr;
1981     iofd.len = size;
1982     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1983     iofd.fd = fd;
1984
1985     if (!kvm_enabled()) {
1986         return -ENOSYS;
1987     }
1988
1989     if (!assign) {
1990         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1991     }
1992
1993     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1994
1995     if (ret < 0) {
1996         return -errno;
1997     }
1998
1999     return 0;
2000 }
2001
2002 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
2003 {
2004     struct kvm_ioeventfd kick = {
2005         .datamatch = val,
2006         .addr = addr,
2007         .len = 2,
2008         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
2009         .fd = fd,
2010     };
2011     int r;
2012     if (!kvm_enabled()) {
2013         return -ENOSYS;
2014     }
2015     if (!assign) {
2016         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
2017     }
2018     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
2019     if (r < 0) {
2020         return r;
2021     }
2022     return 0;
2023 }
2024
2025 int kvm_on_sigbus_vcpu(CPUArchState *env, int code, void *addr)
2026 {
2027     CPUState *cpu = ENV_GET_CPU(env);
2028     return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2029 }
2030
2031 int kvm_on_sigbus(int code, void *addr)
2032 {
2033     return kvm_arch_on_sigbus(code, addr);
2034 }