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 "qemu/osdep.h"
  17 #include <sys/ioctl.h>
  18
  19 #include <linux/kvm.h>
  20
  21 #include "qemu-common.h"
  22 #include "qemu/atomic.h"
  23 #include "qemu/option.h"
  24 #include "qemu/config-file.h"
  25 #include "qemu/error-report.h"
  26 #include "hw/hw.h"
  27 #include "hw/pci/msi.h"
  28 #include "hw/pci/msix.h"
  29 #include "hw/s390x/adapter.h"
  30 #include "exec/gdbstub.h"
  31 #include "sysemu/kvm_int.h"
  32 #include "sysemu/cpus.h"
  33 #include "qemu/bswap.h"
  34 #include "exec/memory.h"
  35 #include "exec/ram_addr.h"
  36 #include "exec/address-spaces.h"
  37 #include "qemu/event_notifier.h"
  38 #include "trace-root.h"
  39 #include "hw/irq.h"
  40
  41 #include "hw/boards.h"
  42
  43 /* This check must be after config-host.h is included */
  44 #ifdef CONFIG_EVENTFD
  45 #include <sys/eventfd.h>
  46 #endif
  47
  48 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
  49  * need to use the real host PAGE_SIZE, as that's what KVM will use.
  50  */
  51 #define PAGE_SIZE getpagesize()
  52
  53 //#define DEBUG_KVM
  54
  55 #ifdef DEBUG_KVM
  56 #define DPRINTF(fmt, ...) \
  57     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  58 #else
  59 #define DPRINTF(fmt, ...) \
  60     do { } while (0)
  61 #endif
  62
  63 #define KVM_MSI_HASHTAB_SIZE    256
  64
  65 struct KVMParkedVcpu {
  66     unsigned long vcpu_id;
  67     int kvm_fd;
  68     QLIST_ENTRY(KVMParkedVcpu) node;
  69 };
  70
  71 struct KVMState
  72 {
  73     AccelState parent_obj;
  74
  75     int nr_slots;
  76     int fd;
  77     int vmfd;
  78     int coalesced_mmio;
  79     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  80     bool coalesced_flush_in_progress;
  81     int broken_set_mem_region;
  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 many_ioeventfds;
  89     int intx_set_mask;
  90     /* The man page (and posix) say ioctl numbers are signed int, but
  91      * they're not.  Linux, glibc and *BSD all treat ioctl numbers as
  92      * unsigned, and treating them as signed here can break things */
  93     unsigned irq_set_ioctl;
  94     unsigned int sigmask_len;
  95     GHashTable *gsimap;
  96 #ifdef KVM_CAP_IRQ_ROUTING
  97     struct kvm_irq_routing *irq_routes;
  98     int nr_allocated_irq_routes;
  99     unsigned long *used_gsi_bitmap;
 100     unsigned int gsi_count;
 101     QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
 102 #endif
 103     KVMMemoryListener memory_listener;
 104     QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
 105 };
 106
 107 KVMState *kvm_state;
 108 bool kvm_kernel_irqchip;
 109 bool kvm_split_irqchip;
 110 bool kvm_async_interrupts_allowed;
 111 bool kvm_halt_in_kernel_allowed;
 112 bool kvm_eventfds_allowed;
 113 bool kvm_irqfds_allowed;
 114 bool kvm_resamplefds_allowed;
 115 bool kvm_msi_via_irqfd_allowed;
 116 bool kvm_gsi_routing_allowed;
 117 bool kvm_gsi_direct_mapping;
 118 bool kvm_allowed;
 119 bool kvm_readonly_mem_allowed;
 120 bool kvm_vm_attributes_allowed;
 121 bool kvm_direct_msi_allowed;
 122 bool kvm_ioeventfd_any_length_allowed;
 123 bool kvm_msi_use_devid;
 124 static bool kvm_immediate_exit;
 125
 126 static const KVMCapabilityInfo kvm_required_capabilites[] = {
 127     KVM_CAP_INFO(USER_MEMORY),
 128     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
 129     KVM_CAP_LAST_INFO
 130 };
 131
 132 int kvm_get_max_memslots(void)
 133 {
 134     KVMState *s = KVM_STATE(current_machine->accelerator);
 135
 136     return s->nr_slots;
 137 }
 138
 139 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
 140 {
 141     KVMState *s = kvm_state;
 142     int i;
 143
 144     for (i = 0; i < s->nr_slots; i++) {
 145         if (kml->slots[i].memory_size == 0) {
 146             return &kml->slots[i];
 147         }
 148     }
 149
 150     return NULL;
 151 }
 152
 153 bool kvm_has_free_slot(MachineState *ms)
 154 {
 155     KVMState *s = KVM_STATE(ms->accelerator);
 156
 157     return kvm_get_free_slot(&s->memory_listener);
 158 }
 159
 160 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
 161 {
 162     KVMSlot *slot = kvm_get_free_slot(kml);
 163
 164     if (slot) {
 165         return slot;
 166     }
 167
 168     fprintf(stderr, "%s: no free slot available\n", __func__);
 169     abort();
 170 }
 171
 172 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
 173                                          hwaddr start_addr,
 174                                          hwaddr end_addr)
 175 {
 176     KVMState *s = kvm_state;
 177     int i;
 178
 179     for (i = 0; i < s->nr_slots; i++) {
 180         KVMSlot *mem = &kml->slots[i];
 181
 182         if (start_addr == mem->start_addr &&
 183             end_addr == mem->start_addr + mem->memory_size) {
 184             return mem;
 185         }
 186     }
 187
 188     return NULL;
 189 }
 190
 191 /*
 192  * Find overlapping slot with lowest start address
 193  */
 194 static KVMSlot *kvm_lookup_overlapping_slot(KVMMemoryListener *kml,
 195                                             hwaddr start_addr,
 196                                             hwaddr end_addr)
 197 {
 198     KVMState *s = kvm_state;
 199     KVMSlot *found = NULL;
 200     int i;
 201
 202     for (i = 0; i < s->nr_slots; i++) {
 203         KVMSlot *mem = &kml->slots[i];
 204
 205         if (mem->memory_size == 0 ||
 206             (found && found->start_addr < mem->start_addr)) {
 207             continue;
 208         }
 209
 210         if (end_addr > mem->start_addr &&
 211             start_addr < mem->start_addr + mem->memory_size) {
 212             found = mem;
 213         }
 214     }
 215
 216     return found;
 217 }
 218
 219 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
 220                                        hwaddr *phys_addr)
 221 {
 222     KVMMemoryListener *kml = &s->memory_listener;
 223     int i;
 224
 225     for (i = 0; i < s->nr_slots; i++) {
 226         KVMSlot *mem = &kml->slots[i];
 227
 228         if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
 229             *phys_addr = mem->start_addr + (ram - mem->ram);
 230             return 1;
 231         }
 232     }
 233
 234     return 0;
 235 }
 236
 237 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot)
 238 {
 239     KVMState *s = kvm_state;
 240     struct kvm_userspace_memory_region mem;
 241
 242     mem.slot = slot->slot | (kml->as_id << 16);
 243     mem.guest_phys_addr = slot->start_addr;
 244     mem.userspace_addr = (unsigned long)slot->ram;
 245     mem.flags = slot->flags;
 246
 247     if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
 248         /* Set the slot size to 0 before setting the slot to the desired
 249          * value. This is needed based on KVM commit 75d61fbc. */
 250         mem.memory_size = 0;
 251         kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 252     }
 253     mem.memory_size = slot->memory_size;
 254     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 255 }
 256
 257 int kvm_destroy_vcpu(CPUState *cpu)
 258 {
 259     KVMState *s = kvm_state;
 260     long mmap_size;
 261     struct KVMParkedVcpu *vcpu = NULL;
 262     int ret = 0;
 263
 264     DPRINTF("kvm_destroy_vcpu\n");
 265
 266     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 267     if (mmap_size < 0) {
 268         ret = mmap_size;
 269         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 270         goto err;
 271     }
 272
 273     ret = munmap(cpu->kvm_run, mmap_size);
 274     if (ret < 0) {
 275         goto err;
 276     }
 277
 278     vcpu = g_malloc0(sizeof(*vcpu));
 279     vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
 280     vcpu->kvm_fd = cpu->kvm_fd;
 281     QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
 282 err:
 283     return ret;
 284 }
 285
 286 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
 287 {
 288     struct KVMParkedVcpu *cpu;
 289
 290     QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
 291         if (cpu->vcpu_id == vcpu_id) {
 292             int kvm_fd;
 293
 294             QLIST_REMOVE(cpu, node);
 295             kvm_fd = cpu->kvm_fd;
 296             g_free(cpu);
 297             return kvm_fd;
 298         }
 299     }
 300
 301     return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
 302 }
 303
 304 int kvm_init_vcpu(CPUState *cpu)
 305 {
 306     KVMState *s = kvm_state;
 307     long mmap_size;
 308     int ret;
 309
 310     DPRINTF("kvm_init_vcpu\n");
 311
 312     ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
 313     if (ret < 0) {
 314         DPRINTF("kvm_create_vcpu failed\n");
 315         goto err;
 316     }
 317
 318     cpu->kvm_fd = ret;
 319     cpu->kvm_state = s;
 320     cpu->kvm_vcpu_dirty = true;
 321
 322     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 323     if (mmap_size < 0) {
 324         ret = mmap_size;
 325         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 326         goto err;
 327     }
 328
 329     cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 330                         cpu->kvm_fd, 0);
 331     if (cpu->kvm_run == MAP_FAILED) {
 332         ret = -errno;
 333         DPRINTF("mmap'ing vcpu state failed\n");
 334         goto err;
 335     }
 336
 337     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 338         s->coalesced_mmio_ring =
 339             (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 340     }
 341
 342     ret = kvm_arch_init_vcpu(cpu);
 343 err:
 344     return ret;
 345 }
 346
 347 /*
 348  * dirty pages logging control
 349  */
 350
 351 static int kvm_mem_flags(MemoryRegion *mr)
 352 {
 353     bool readonly = mr->readonly || memory_region_is_romd(mr);
 354     int flags = 0;
 355
 356     if (memory_region_get_dirty_log_mask(mr) != 0) {
 357         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 358     }
 359     if (readonly && kvm_readonly_mem_allowed) {
 360         flags |= KVM_MEM_READONLY;
 361     }
 362     return flags;
 363 }
 364
 365 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
 366                                  MemoryRegion *mr)
 367 {
 368     int old_flags;
 369
 370     old_flags = mem->flags;
 371     mem->flags = kvm_mem_flags(mr);
 372
 373     /* If nothing changed effectively, no need to issue ioctl */
 374     if (mem->flags == old_flags) {
 375         return 0;
 376     }
 377
 378     return kvm_set_user_memory_region(kml, mem);
 379 }
 380
 381 static int kvm_section_update_flags(KVMMemoryListener *kml,
 382                                     MemoryRegionSection *section)
 383 {
 384     hwaddr phys_addr = section->offset_within_address_space;
 385     ram_addr_t size = int128_get64(section->size);
 386     KVMSlot *mem = kvm_lookup_matching_slot(kml, phys_addr, phys_addr + size);
 387
 388     if (mem == NULL)  {
 389         return 0;
 390     } else {
 391         return kvm_slot_update_flags(kml, mem, section->mr);
 392     }
 393 }
 394
 395 static void kvm_log_start(MemoryListener *listener,
 396                           MemoryRegionSection *section,
 397                           int old, int new)
 398 {
 399     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 400     int r;
 401
 402     if (old != 0) {
 403         return;
 404     }
 405
 406     r = kvm_section_update_flags(kml, section);
 407     if (r < 0) {
 408         abort();
 409     }
 410 }
 411
 412 static void kvm_log_stop(MemoryListener *listener,
 413                           MemoryRegionSection *section,
 414                           int old, int new)
 415 {
 416     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 417     int r;
 418
 419     if (new != 0) {
 420         return;
 421     }
 422
 423     r = kvm_section_update_flags(kml, section);
 424     if (r < 0) {
 425         abort();
 426     }
 427 }
 428
 429 /* get kvm's dirty pages bitmap and update qemu's */
 430 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
 431                                          unsigned long *bitmap)
 432 {
 433     ram_addr_t start = section->offset_within_region +
 434                        memory_region_get_ram_addr(section->mr);
 435     ram_addr_t pages = int128_get64(section->size) / getpagesize();
 436
 437     cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
 438     return 0;
 439 }
 440
 441 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 442
 443 /**
 444  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 445  * This function updates qemu's dirty bitmap using
 446  * memory_region_set_dirty().  This means all bits are set
 447  * to dirty.
 448  *
 449  * @start_add: start of logged region.
 450  * @end_addr: end of logged region.
 451  */
 452 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
 453                                           MemoryRegionSection *section)
 454 {
 455     KVMState *s = kvm_state;
 456     unsigned long size, allocated_size = 0;
 457     struct kvm_dirty_log d = {};
 458     KVMSlot *mem;
 459     int ret = 0;
 460     hwaddr start_addr = section->offset_within_address_space;
 461     hwaddr end_addr = start_addr + int128_get64(section->size);
 462
 463     d.dirty_bitmap = NULL;
 464     while (start_addr < end_addr) {
 465         mem = kvm_lookup_overlapping_slot(kml, start_addr, end_addr);
 466         if (mem == NULL) {
 467             break;
 468         }
 469
 470         /* XXX bad kernel interface alert
 471          * For dirty bitmap, kernel allocates array of size aligned to
 472          * bits-per-long.  But for case when the kernel is 64bits and
 473          * the userspace is 32bits, userspace can't align to the same
 474          * bits-per-long, since sizeof(long) is different between kernel
 475          * and user space.  This way, userspace will provide buffer which
 476          * may be 4 bytes less than the kernel will use, resulting in
 477          * userspace memory corruption (which is not detectable by valgrind
 478          * too, in most cases).
 479          * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
 480          * a hope that sizeof(long) won't become >8 any time soon.
 481          */
 482         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
 483                      /*HOST_LONG_BITS*/ 64) / 8;
 484         if (!d.dirty_bitmap) {
 485             d.dirty_bitmap = g_malloc(size);
 486         } else if (size > allocated_size) {
 487             d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
 488         }
 489         allocated_size = size;
 490         memset(d.dirty_bitmap, 0, allocated_size);
 491
 492         d.slot = mem->slot | (kml->as_id << 16);
 493         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 494             DPRINTF("ioctl failed %d\n", errno);
 495             ret = -1;
 496             break;
 497         }
 498
 499         kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
 500         start_addr = mem->start_addr + mem->memory_size;
 501     }
 502     g_free(d.dirty_bitmap);
 503
 504     return ret;
 505 }
 506
 507 static void kvm_coalesce_mmio_region(MemoryListener *listener,
 508                                      MemoryRegionSection *secion,
 509                                      hwaddr start, hwaddr size)
 510 {
 511     KVMState *s = kvm_state;
 512
 513     if (s->coalesced_mmio) {
 514         struct kvm_coalesced_mmio_zone zone;
 515
 516         zone.addr = start;
 517         zone.size = size;
 518         zone.pad = 0;
 519
 520         (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 521     }
 522 }
 523
 524 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
 525                                        MemoryRegionSection *secion,
 526                                        hwaddr start, hwaddr size)
 527 {
 528     KVMState *s = kvm_state;
 529
 530     if (s->coalesced_mmio) {
 531         struct kvm_coalesced_mmio_zone zone;
 532
 533         zone.addr = start;
 534         zone.size = size;
 535         zone.pad = 0;
 536
 537         (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 538     }
 539 }
 540
 541 int kvm_check_extension(KVMState *s, unsigned int extension)
 542 {
 543     int ret;
 544
 545     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 546     if (ret < 0) {
 547         ret = 0;
 548     }
 549
 550     return ret;
 551 }
 552
 553 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
 554 {
 555     int ret;
 556
 557     ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 558     if (ret < 0) {
 559         /* VM wide version not implemented, use global one instead */
 560         ret = kvm_check_extension(s, extension);
 561     }
 562
 563     return ret;
 564 }
 565
 566 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
 567 {
 568 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
 569     /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
 570      * endianness, but the memory core hands them in target endianness.
 571      * For example, PPC is always treated as big-endian even if running
 572      * on KVM and on PPC64LE.  Correct here.
 573      */
 574     switch (size) {
 575     case 2:
 576         val = bswap16(val);
 577         break;
 578     case 4:
 579         val = bswap32(val);
 580         break;
 581     }
 582 #endif
 583     return val;
 584 }
 585
 586 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
 587                                   bool assign, uint32_t size, bool datamatch)
 588 {
 589     int ret;
 590     struct kvm_ioeventfd iofd = {
 591         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
 592         .addr = addr,
 593         .len = size,
 594         .flags = 0,
 595         .fd = fd,
 596     };
 597
 598     if (!kvm_enabled()) {
 599         return -ENOSYS;
 600     }
 601
 602     if (datamatch) {
 603         iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
 604     }
 605     if (!assign) {
 606         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
 607     }
 608
 609     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
 610
 611     if (ret < 0) {
 612         return -errno;
 613     }
 614
 615     return 0;
 616 }
 617
 618 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
 619                                  bool assign, uint32_t size, bool datamatch)
 620 {
 621     struct kvm_ioeventfd kick = {
 622         .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
 623         .addr = addr,
 624         .flags = KVM_IOEVENTFD_FLAG_PIO,
 625         .len = size,
 626         .fd = fd,
 627     };
 628     int r;
 629     if (!kvm_enabled()) {
 630         return -ENOSYS;
 631     }
 632     if (datamatch) {
 633         kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
 634     }
 635     if (!assign) {
 636         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
 637     }
 638     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
 639     if (r < 0) {
 640         return r;
 641     }
 642     return 0;
 643 }
 644
 645
 646 static int kvm_check_many_ioeventfds(void)
 647 {
 648     /* Userspace can use ioeventfd for io notification.  This requires a host
 649      * that supports eventfd(2) and an I/O thread; since eventfd does not
 650      * support SIGIO it cannot interrupt the vcpu.
 651      *
 652      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 653      * can avoid creating too many ioeventfds.
 654      */
 655 #if defined(CONFIG_EVENTFD)
 656     int ioeventfds[7];
 657     int i, ret = 0;
 658     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 659         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 660         if (ioeventfds[i] < 0) {
 661             break;
 662         }
 663         ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
 664         if (ret < 0) {
 665             close(ioeventfds[i]);
 666             break;
 667         }
 668     }
 669
 670     /* Decide whether many devices are supported or not */
 671     ret = i == ARRAY_SIZE(ioeventfds);
 672
 673     while (i-- > 0) {
 674         kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
 675         close(ioeventfds[i]);
 676     }
 677     return ret;
 678 #else
 679     return 0;
 680 #endif
 681 }
 682
 683 static const KVMCapabilityInfo *
 684 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 685 {
 686     while (list->name) {
 687         if (!kvm_check_extension(s, list->value)) {
 688             return list;
 689         }
 690         list++;
 691     }
 692     return NULL;
 693 }
 694
 695 static void kvm_set_phys_mem(KVMMemoryListener *kml,
 696                              MemoryRegionSection *section, bool add)
 697 {
 698     KVMState *s = kvm_state;
 699     KVMSlot *mem, old;
 700     int err;
 701     MemoryRegion *mr = section->mr;
 702     bool writeable = !mr->readonly && !mr->rom_device;
 703     hwaddr start_addr = section->offset_within_address_space;
 704     ram_addr_t size = int128_get64(section->size);
 705     void *ram = NULL;
 706     unsigned delta;
 707
 708     /* kvm works in page size chunks, but the function may be called
 709        with sub-page size and unaligned start address. Pad the start
 710        address to next and truncate size to previous page boundary. */
 711     delta = qemu_real_host_page_size - (start_addr & ~qemu_real_host_page_mask);
 712     delta &= ~qemu_real_host_page_mask;
 713     if (delta > size) {
 714         return;
 715     }
 716     start_addr += delta;
 717     size -= delta;
 718     size &= qemu_real_host_page_mask;
 719     if (!size || (start_addr & ~qemu_real_host_page_mask)) {
 720         return;
 721     }
 722
 723     if (!memory_region_is_ram(mr)) {
 724         if (writeable || !kvm_readonly_mem_allowed) {
 725             return;
 726         } else if (!mr->romd_mode) {
 727             /* If the memory device is not in romd_mode, then we actually want
 728              * to remove the kvm memory slot so all accesses will trap. */
 729             add = false;
 730         }
 731     }
 732
 733     ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
 734
 735     while (1) {
 736         mem = kvm_lookup_overlapping_slot(kml, start_addr, start_addr + size);
 737         if (!mem) {
 738             break;
 739         }
 740
 741         if (add && start_addr >= mem->start_addr &&
 742             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 743             (ram - start_addr == mem->ram - mem->start_addr)) {
 744             /* The new slot fits into the existing one and comes with
 745              * identical parameters - update flags and done. */
 746             kvm_slot_update_flags(kml, mem, mr);
 747             return;
 748         }
 749
 750         old = *mem;
 751
 752         if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
 753             kvm_physical_sync_dirty_bitmap(kml, section);
 754         }
 755
 756         /* unregister the overlapping slot */
 757         mem->memory_size = 0;
 758         err = kvm_set_user_memory_region(kml, mem);
 759         if (err) {
 760             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 761                     __func__, strerror(-err));
 762             abort();
 763         }
 764
 765         /* Workaround for older KVM versions: we can't join slots, even not by
 766          * unregistering the previous ones and then registering the larger
 767          * slot. We have to maintain the existing fragmentation. Sigh.
 768          *
 769          * This workaround assumes that the new slot starts at the same
 770          * address as the first existing one. If not or if some overlapping
 771          * slot comes around later, we will fail (not seen in practice so far)
 772          * - and actually require a recent KVM version. */
 773         if (s->broken_set_mem_region &&
 774             old.start_addr == start_addr && old.memory_size < size && add) {
 775             mem = kvm_alloc_slot(kml);
 776             mem->memory_size = old.memory_size;
 777             mem->start_addr = old.start_addr;
 778             mem->ram = old.ram;
 779             mem->flags = kvm_mem_flags(mr);
 780
 781             err = kvm_set_user_memory_region(kml, mem);
 782             if (err) {
 783                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 784                         strerror(-err));
 785                 abort();
 786             }
 787
 788             start_addr += old.memory_size;
 789             ram += old.memory_size;
 790             size -= old.memory_size;
 791             continue;
 792         }
 793
 794         /* register prefix slot */
 795         if (old.start_addr < start_addr) {
 796             mem = kvm_alloc_slot(kml);
 797             mem->memory_size = start_addr - old.start_addr;
 798             mem->start_addr = old.start_addr;
 799             mem->ram = old.ram;
 800             mem->flags =  kvm_mem_flags(mr);
 801
 802             err = kvm_set_user_memory_region(kml, mem);
 803             if (err) {
 804                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 805                         __func__, strerror(-err));
 806 #ifdef TARGET_PPC
 807                 fprintf(stderr, "%s: This is probably because your kernel's " \
 808                                 "PAGE_SIZE is too big. Please try to use 4k " \
 809                                 "PAGE_SIZE!\n", __func__);
 810 #endif
 811                 abort();
 812             }
 813         }
 814
 815         /* register suffix slot */
 816         if (old.start_addr + old.memory_size > start_addr + size) {
 817             ram_addr_t size_delta;
 818
 819             mem = kvm_alloc_slot(kml);
 820             mem->start_addr = start_addr + size;
 821             size_delta = mem->start_addr - old.start_addr;
 822             mem->memory_size = old.memory_size - size_delta;
 823             mem->ram = old.ram + size_delta;
 824             mem->flags = kvm_mem_flags(mr);
 825
 826             err = kvm_set_user_memory_region(kml, mem);
 827             if (err) {
 828                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 829                         __func__, strerror(-err));
 830                 abort();
 831             }
 832         }
 833     }
 834
 835     /* in case the KVM bug workaround already "consumed" the new slot */
 836     if (!size) {
 837         return;
 838     }
 839     if (!add) {
 840         return;
 841     }
 842     mem = kvm_alloc_slot(kml);
 843     mem->memory_size = size;
 844     mem->start_addr = start_addr;
 845     mem->ram = ram;
 846     mem->flags = kvm_mem_flags(mr);
 847
 848     err = kvm_set_user_memory_region(kml, mem);
 849     if (err) {
 850         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 851                 strerror(-err));
 852         abort();
 853     }
 854 }
 855
 856 static void kvm_region_add(MemoryListener *listener,
 857                            MemoryRegionSection *section)
 858 {
 859     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 860
 861     memory_region_ref(section->mr);
 862     kvm_set_phys_mem(kml, section, true);
 863 }
 864
 865 static void kvm_region_del(MemoryListener *listener,
 866                            MemoryRegionSection *section)
 867 {
 868     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 869
 870     kvm_set_phys_mem(kml, section, false);
 871     memory_region_unref(section->mr);
 872 }
 873
 874 static void kvm_log_sync(MemoryListener *listener,
 875                          MemoryRegionSection *section)
 876 {
 877     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
 878     int r;
 879
 880     r = kvm_physical_sync_dirty_bitmap(kml, section);
 881     if (r < 0) {
 882         abort();
 883     }
 884 }
 885
 886 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
 887                                   MemoryRegionSection *section,
 888                                   bool match_data, uint64_t data,
 889                                   EventNotifier *e)
 890 {
 891     int fd = event_notifier_get_fd(e);
 892     int r;
 893
 894     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 895                                data, true, int128_get64(section->size),
 896                                match_data);
 897     if (r < 0) {
 898         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
 899                 __func__, strerror(-r));
 900         abort();
 901     }
 902 }
 903
 904 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
 905                                   MemoryRegionSection *section,
 906                                   bool match_data, uint64_t data,
 907                                   EventNotifier *e)
 908 {
 909     int fd = event_notifier_get_fd(e);
 910     int r;
 911
 912     r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
 913                                data, false, int128_get64(section->size),
 914                                match_data);
 915     if (r < 0) {
 916         abort();
 917     }
 918 }
 919
 920 static void kvm_io_ioeventfd_add(MemoryListener *listener,
 921                                  MemoryRegionSection *section,
 922                                  bool match_data, uint64_t data,
 923                                  EventNotifier *e)
 924 {
 925     int fd = event_notifier_get_fd(e);
 926     int r;
 927
 928     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
 929                               data, true, int128_get64(section->size),
 930                               match_data);
 931     if (r < 0) {
 932         fprintf(stderr, "%s: error adding ioeventfd: %s\n",
 933                 __func__, strerror(-r));
 934         abort();
 935     }
 936 }
 937
 938 static void kvm_io_ioeventfd_del(MemoryListener *listener,
 939                                  MemoryRegionSection *section,
 940                                  bool match_data, uint64_t data,
 941                                  EventNotifier *e)
 942
 943 {
 944     int fd = event_notifier_get_fd(e);
 945     int r;
 946
 947     r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
 948                               data, false, int128_get64(section->size),
 949                               match_data);
 950     if (r < 0) {
 951         abort();
 952     }
 953 }
 954
 955 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
 956                                   AddressSpace *as, int as_id)
 957 {
 958     int i;
 959
 960     kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
 961     kml->as_id = as_id;
 962
 963     for (i = 0; i < s->nr_slots; i++) {
 964         kml->slots[i].slot = i;
 965     }
 966
 967     kml->listener.region_add = kvm_region_add;
 968     kml->listener.region_del = kvm_region_del;
 969     kml->listener.log_start = kvm_log_start;
 970     kml->listener.log_stop = kvm_log_stop;
 971     kml->listener.log_sync = kvm_log_sync;
 972     kml->listener.priority = 10;
 973
 974     memory_listener_register(&kml->listener, as);
 975 }
 976
 977 static MemoryListener kvm_io_listener = {
 978     .eventfd_add = kvm_io_ioeventfd_add,
 979     .eventfd_del = kvm_io_ioeventfd_del,
 980     .priority = 10,
 981 };
 982
 983 static void kvm_handle_interrupt(CPUState *cpu, int mask)
 984 {
 985     cpu->interrupt_request |= mask;
 986
 987     if (!qemu_cpu_is_self(cpu)) {
 988         qemu_cpu_kick(cpu);
 989     }
 990 }
 991
 992 int kvm_set_irq(KVMState *s, int irq, int level)
 993 {
 994     struct kvm_irq_level event;
 995     int ret;
 996
 997     assert(kvm_async_interrupts_enabled());
 998
 999     event.level = level;
1000     event.irq = irq;
1001     ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
1002     if (ret < 0) {
1003         perror("kvm_set_irq");
1004         abort();
1005     }
1006
1007     return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
1008 }
1009
1010 #ifdef KVM_CAP_IRQ_ROUTING
1011 typedef struct KVMMSIRoute {
1012     struct kvm_irq_routing_entry kroute;
1013     QTAILQ_ENTRY(KVMMSIRoute) entry;
1014 } KVMMSIRoute;
1015
1016 static void set_gsi(KVMState *s, unsigned int gsi)
1017 {
1018     set_bit(gsi, s->used_gsi_bitmap);
1019 }
1020
1021 static void clear_gsi(KVMState *s, unsigned int gsi)
1022 {
1023     clear_bit(gsi, s->used_gsi_bitmap);
1024 }
1025
1026 void kvm_init_irq_routing(KVMState *s)
1027 {
1028     int gsi_count, i;
1029
1030     gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1031     if (gsi_count > 0) {
1032         /* Round up so we can search ints using ffs */
1033         s->used_gsi_bitmap = bitmap_new(gsi_count);
1034         s->gsi_count = gsi_count;
1035     }
1036
1037     s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1038     s->nr_allocated_irq_routes = 0;
1039
1040     if (!kvm_direct_msi_allowed) {
1041         for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1042             QTAILQ_INIT(&s->msi_hashtab[i]);
1043         }
1044     }
1045
1046     kvm_arch_init_irq_routing(s);
1047 }
1048
1049 void kvm_irqchip_commit_routes(KVMState *s)
1050 {
1051     int ret;
1052
1053     if (kvm_gsi_direct_mapping()) {
1054         return;
1055     }
1056
1057     if (!kvm_gsi_routing_enabled()) {
1058         return;
1059     }
1060
1061     s->irq_routes->flags = 0;
1062     trace_kvm_irqchip_commit_routes();
1063     ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1064     assert(ret == 0);
1065 }
1066
1067 static void kvm_add_routing_entry(KVMState *s,
1068                                   struct kvm_irq_routing_entry *entry)
1069 {
1070     struct kvm_irq_routing_entry *new;
1071     int n, size;
1072
1073     if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1074         n = s->nr_allocated_irq_routes * 2;
1075         if (n < 64) {
1076             n = 64;
1077         }
1078         size = sizeof(struct kvm_irq_routing);
1079         size += n * sizeof(*new);
1080         s->irq_routes = g_realloc(s->irq_routes, size);
1081         s->nr_allocated_irq_routes = n;
1082     }
1083     n = s->irq_routes->nr++;
1084     new = &s->irq_routes->entries[n];
1085
1086     *new = *entry;
1087
1088     set_gsi(s, entry->gsi);
1089 }
1090
1091 static int kvm_update_routing_entry(KVMState *s,
1092                                     struct kvm_irq_routing_entry *new_entry)
1093 {
1094     struct kvm_irq_routing_entry *entry;
1095     int n;
1096
1097     for (n = 0; n < s->irq_routes->nr; n++) {
1098         entry = &s->irq_routes->entries[n];
1099         if (entry->gsi != new_entry->gsi) {
1100             continue;
1101         }
1102
1103         if(!memcmp(entry, new_entry, sizeof *entry)) {
1104             return 0;
1105         }
1106
1107         *entry = *new_entry;
1108
1109         return 0;
1110     }
1111
1112     return -ESRCH;
1113 }
1114
1115 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1116 {
1117     struct kvm_irq_routing_entry e = {};
1118
1119     assert(pin < s->gsi_count);
1120
1121     e.gsi = irq;
1122     e.type = KVM_IRQ_ROUTING_IRQCHIP;
1123     e.flags = 0;
1124     e.u.irqchip.irqchip = irqchip;
1125     e.u.irqchip.pin = pin;
1126     kvm_add_routing_entry(s, &e);
1127 }
1128
1129 void kvm_irqchip_release_virq(KVMState *s, int virq)
1130 {
1131     struct kvm_irq_routing_entry *e;
1132     int i;
1133
1134     if (kvm_gsi_direct_mapping()) {
1135         return;
1136     }
1137
1138     for (i = 0; i < s->irq_routes->nr; i++) {
1139         e = &s->irq_routes->entries[i];
1140         if (e->gsi == virq) {
1141             s->irq_routes->nr--;
1142             *e = s->irq_routes->entries[s->irq_routes->nr];
1143         }
1144     }
1145     clear_gsi(s, virq);
1146     kvm_arch_release_virq_post(virq);
1147 }
1148
1149 static unsigned int kvm_hash_msi(uint32_t data)
1150 {
1151     /* This is optimized for IA32 MSI layout. However, no other arch shall
1152      * repeat the mistake of not providing a direct MSI injection API. */
1153     return data & 0xff;
1154 }
1155
1156 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1157 {
1158     KVMMSIRoute *route, *next;
1159     unsigned int hash;
1160
1161     for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1162         QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1163             kvm_irqchip_release_virq(s, route->kroute.gsi);
1164             QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1165             g_free(route);
1166         }
1167     }
1168 }
1169
1170 static int kvm_irqchip_get_virq(KVMState *s)
1171 {
1172     int next_virq;
1173
1174     /*
1175      * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1176      * GSI numbers are more than the number of IRQ route. Allocating a GSI
1177      * number can succeed even though a new route entry cannot be added.
1178      * When this happens, flush dynamic MSI entries to free IRQ route entries.
1179      */
1180     if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1181         kvm_flush_dynamic_msi_routes(s);
1182     }
1183
1184     /* Return the lowest unused GSI in the bitmap */
1185     next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1186     if (next_virq >= s->gsi_count) {
1187         return -ENOSPC;
1188     } else {
1189         return next_virq;
1190     }
1191 }
1192
1193 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1194 {
1195     unsigned int hash = kvm_hash_msi(msg.data);
1196     KVMMSIRoute *route;
1197
1198     QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1199         if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1200             route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1201             route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1202             return route;
1203         }
1204     }
1205     return NULL;
1206 }
1207
1208 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1209 {
1210     struct kvm_msi msi;
1211     KVMMSIRoute *route;
1212
1213     if (kvm_direct_msi_allowed) {
1214         msi.address_lo = (uint32_t)msg.address;
1215         msi.address_hi = msg.address >> 32;
1216         msi.data = le32_to_cpu(msg.data);
1217         msi.flags = 0;
1218         memset(msi.pad, 0, sizeof(msi.pad));
1219
1220         return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1221     }
1222
1223     route = kvm_lookup_msi_route(s, msg);
1224     if (!route) {
1225         int virq;
1226
1227         virq = kvm_irqchip_get_virq(s);
1228         if (virq < 0) {
1229             return virq;
1230         }
1231
1232         route = g_malloc0(sizeof(KVMMSIRoute));
1233         route->kroute.gsi = virq;
1234         route->kroute.type = KVM_IRQ_ROUTING_MSI;
1235         route->kroute.flags = 0;
1236         route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1237         route->kroute.u.msi.address_hi = msg.address >> 32;
1238         route->kroute.u.msi.data = le32_to_cpu(msg.data);
1239
1240         kvm_add_routing_entry(s, &route->kroute);
1241         kvm_irqchip_commit_routes(s);
1242
1243         QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1244                            entry);
1245     }
1246
1247     assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1248
1249     return kvm_set_irq(s, route->kroute.gsi, 1);
1250 }
1251
1252 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1253 {
1254     struct kvm_irq_routing_entry kroute = {};
1255     int virq;
1256     MSIMessage msg = {0, 0};
1257
1258     if (dev) {
1259         msg = pci_get_msi_message(dev, vector);
1260     }
1261
1262     if (kvm_gsi_direct_mapping()) {
1263         return kvm_arch_msi_data_to_gsi(msg.data);
1264     }
1265
1266     if (!kvm_gsi_routing_enabled()) {
1267         return -ENOSYS;
1268     }
1269
1270     virq = kvm_irqchip_get_virq(s);
1271     if (virq < 0) {
1272         return virq;
1273     }
1274
1275     kroute.gsi = virq;
1276     kroute.type = KVM_IRQ_ROUTING_MSI;
1277     kroute.flags = 0;
1278     kroute.u.msi.address_lo = (uint32_t)msg.address;
1279     kroute.u.msi.address_hi = msg.address >> 32;
1280     kroute.u.msi.data = le32_to_cpu(msg.data);
1281     if (kvm_msi_devid_required()) {
1282         kroute.flags = KVM_MSI_VALID_DEVID;
1283         kroute.u.msi.devid = pci_requester_id(dev);
1284     }
1285     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1286         kvm_irqchip_release_virq(s, virq);
1287         return -EINVAL;
1288     }
1289
1290     trace_kvm_irqchip_add_msi_route(virq);
1291
1292     kvm_add_routing_entry(s, &kroute);
1293     kvm_arch_add_msi_route_post(&kroute, vector, dev);
1294     kvm_irqchip_commit_routes(s);
1295
1296     return virq;
1297 }
1298
1299 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1300                                  PCIDevice *dev)
1301 {
1302     struct kvm_irq_routing_entry kroute = {};
1303
1304     if (kvm_gsi_direct_mapping()) {
1305         return 0;
1306     }
1307
1308     if (!kvm_irqchip_in_kernel()) {
1309         return -ENOSYS;
1310     }
1311
1312     kroute.gsi = virq;
1313     kroute.type = KVM_IRQ_ROUTING_MSI;
1314     kroute.flags = 0;
1315     kroute.u.msi.address_lo = (uint32_t)msg.address;
1316     kroute.u.msi.address_hi = msg.address >> 32;
1317     kroute.u.msi.data = le32_to_cpu(msg.data);
1318     if (kvm_msi_devid_required()) {
1319         kroute.flags = KVM_MSI_VALID_DEVID;
1320         kroute.u.msi.devid = pci_requester_id(dev);
1321     }
1322     if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1323         return -EINVAL;
1324     }
1325
1326     trace_kvm_irqchip_update_msi_route(virq);
1327
1328     return kvm_update_routing_entry(s, &kroute);
1329 }
1330
1331 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1332                                     bool assign)
1333 {
1334     struct kvm_irqfd irqfd = {
1335         .fd = fd,
1336         .gsi = virq,
1337         .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1338     };
1339
1340     if (rfd != -1) {
1341         irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1342         irqfd.resamplefd = rfd;
1343     }
1344
1345     if (!kvm_irqfds_enabled()) {
1346         return -ENOSYS;
1347     }
1348
1349     return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1350 }
1351
1352 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1353 {
1354     struct kvm_irq_routing_entry kroute = {};
1355     int virq;
1356
1357     if (!kvm_gsi_routing_enabled()) {
1358         return -ENOSYS;
1359     }
1360
1361     virq = kvm_irqchip_get_virq(s);
1362     if (virq < 0) {
1363         return virq;
1364     }
1365
1366     kroute.gsi = virq;
1367     kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1368     kroute.flags = 0;
1369     kroute.u.adapter.summary_addr = adapter->summary_addr;
1370     kroute.u.adapter.ind_addr = adapter->ind_addr;
1371     kroute.u.adapter.summary_offset = adapter->summary_offset;
1372     kroute.u.adapter.ind_offset = adapter->ind_offset;
1373     kroute.u.adapter.adapter_id = adapter->adapter_id;
1374
1375     kvm_add_routing_entry(s, &kroute);
1376
1377     return virq;
1378 }
1379
1380 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1381 {
1382     struct kvm_irq_routing_entry kroute = {};
1383     int virq;
1384
1385     if (!kvm_gsi_routing_enabled()) {
1386         return -ENOSYS;
1387     }
1388     if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1389         return -ENOSYS;
1390     }
1391     virq = kvm_irqchip_get_virq(s);
1392     if (virq < 0) {
1393         return virq;
1394     }
1395
1396     kroute.gsi = virq;
1397     kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1398     kroute.flags = 0;
1399     kroute.u.hv_sint.vcpu = vcpu;
1400     kroute.u.hv_sint.sint = sint;
1401
1402     kvm_add_routing_entry(s, &kroute);
1403     kvm_irqchip_commit_routes(s);
1404
1405     return virq;
1406 }
1407
1408 #else /* !KVM_CAP_IRQ_ROUTING */
1409
1410 void kvm_init_irq_routing(KVMState *s)
1411 {
1412 }
1413
1414 void kvm_irqchip_release_virq(KVMState *s, int virq)
1415 {
1416 }
1417
1418 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1419 {
1420     abort();
1421 }
1422
1423 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1424 {
1425     return -ENOSYS;
1426 }
1427
1428 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1429 {
1430     return -ENOSYS;
1431 }
1432
1433 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1434 {
1435     return -ENOSYS;
1436 }
1437
1438 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1439 {
1440     abort();
1441 }
1442
1443 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1444 {
1445     return -ENOSYS;
1446 }
1447 #endif /* !KVM_CAP_IRQ_ROUTING */
1448
1449 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1450                                        EventNotifier *rn, int virq)
1451 {
1452     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1453            rn ? event_notifier_get_fd(rn) : -1, virq, true);
1454 }
1455
1456 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1457                                           int virq)
1458 {
1459     return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1460            false);
1461 }
1462
1463 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1464                                    EventNotifier *rn, qemu_irq irq)
1465 {
1466     gpointer key, gsi;
1467     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1468
1469     if (!found) {
1470         return -ENXIO;
1471     }
1472     return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1473 }
1474
1475 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1476                                       qemu_irq irq)
1477 {
1478     gpointer key, gsi;
1479     gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1480
1481     if (!found) {
1482         return -ENXIO;
1483     }
1484     return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1485 }
1486
1487 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1488 {
1489     g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1490 }
1491
1492 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1493 {
1494     int ret;
1495
1496     if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1497         ;
1498     } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1499         ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1500         if (ret < 0) {
1501             fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1502             exit(1);
1503         }
1504     } else {
1505         return;
1506     }
1507
1508     /* First probe and see if there's a arch-specific hook to create the
1509      * in-kernel irqchip for us */
1510     ret = kvm_arch_irqchip_create(machine, s);
1511     if (ret == 0) {
1512         if (machine_kernel_irqchip_split(machine)) {
1513             perror("Split IRQ chip mode not supported.");
1514             exit(1);
1515         } else {
1516             ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1517         }
1518     }
1519     if (ret < 0) {
1520         fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1521         exit(1);
1522     }
1523
1524     kvm_kernel_irqchip = true;
1525     /* If we have an in-kernel IRQ chip then we must have asynchronous
1526      * interrupt delivery (though the reverse is not necessarily true)
1527      */
1528     kvm_async_interrupts_allowed = true;
1529     kvm_halt_in_kernel_allowed = true;
1530
1531     kvm_init_irq_routing(s);
1532
1533     s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1534 }
1535
1536 /* Find number of supported CPUs using the recommended
1537  * procedure from the kernel API documentation to cope with
1538  * older kernels that may be missing capabilities.
1539  */
1540 static int kvm_recommended_vcpus(KVMState *s)
1541 {
1542     int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1543     return (ret) ? ret : 4;
1544 }
1545
1546 static int kvm_max_vcpus(KVMState *s)
1547 {
1548     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1549     return (ret) ? ret : kvm_recommended_vcpus(s);
1550 }
1551
1552 static int kvm_max_vcpu_id(KVMState *s)
1553 {
1554     int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1555     return (ret) ? ret : kvm_max_vcpus(s);
1556 }
1557
1558 bool kvm_vcpu_id_is_valid(int vcpu_id)
1559 {
1560     KVMState *s = KVM_STATE(current_machine->accelerator);
1561     return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1562 }
1563
1564 static int kvm_init(MachineState *ms)
1565 {
1566     MachineClass *mc = MACHINE_GET_CLASS(ms);
1567     static const char upgrade_note[] =
1568         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1569         "(see http://sourceforge.net/projects/kvm).\n";
1570     struct {
1571         const char *name;
1572         int num;
1573     } num_cpus[] = {
1574         { "SMP",          smp_cpus },
1575         { "hotpluggable", max_cpus },
1576         { NULL, }
1577     }, *nc = num_cpus;
1578     int soft_vcpus_limit, hard_vcpus_limit;
1579     KVMState *s;
1580     const KVMCapabilityInfo *missing_cap;
1581     int ret;
1582     int type = 0;
1583     const char *kvm_type;
1584
1585     s = KVM_STATE(ms->accelerator);
1586
1587     /*
1588      * On systems where the kernel can support different base page
1589      * sizes, host page size may be different from TARGET_PAGE_SIZE,
1590      * even with KVM.  TARGET_PAGE_SIZE is assumed to be the minimum
1591      * page size for the system though.
1592      */
1593     assert(TARGET_PAGE_SIZE <= getpagesize());
1594
1595     s->sigmask_len = 8;
1596
1597 #ifdef KVM_CAP_SET_GUEST_DEBUG
1598     QTAILQ_INIT(&s->kvm_sw_breakpoints);
1599 #endif
1600     QLIST_INIT(&s->kvm_parked_vcpus);
1601     s->vmfd = -1;
1602     s->fd = qemu_open("/dev/kvm", O_RDWR);
1603     if (s->fd == -1) {
1604         fprintf(stderr, "Could not access KVM kernel module: %m\n");
1605         ret = -errno;
1606         goto err;
1607     }
1608
1609     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1610     if (ret < KVM_API_VERSION) {
1611         if (ret >= 0) {
1612             ret = -EINVAL;
1613         }
1614         fprintf(stderr, "kvm version too old\n");
1615         goto err;
1616     }
1617
1618     if (ret > KVM_API_VERSION) {
1619         ret = -EINVAL;
1620         fprintf(stderr, "kvm version not supported\n");
1621         goto err;
1622     }
1623
1624     kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1625     s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1626
1627     /* If unspecified, use the default value */
1628     if (!s->nr_slots) {
1629         s->nr_slots = 32;
1630     }
1631
1632     /* check the vcpu limits */
1633     soft_vcpus_limit = kvm_recommended_vcpus(s);
1634     hard_vcpus_limit = kvm_max_vcpus(s);
1635
1636     while (nc->name) {
1637         if (nc->num > soft_vcpus_limit) {
1638             fprintf(stderr,
1639                     "Warning: Number of %s cpus requested (%d) exceeds "
1640                     "the recommended cpus supported by KVM (%d)\n",
1641                     nc->name, nc->num, soft_vcpus_limit);
1642
1643             if (nc->num > hard_vcpus_limit) {
1644                 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1645                         "the maximum cpus supported by KVM (%d)\n",
1646                         nc->name, nc->num, hard_vcpus_limit);
1647                 exit(1);
1648             }
1649         }
1650         nc++;
1651     }
1652
1653     kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1654     if (mc->kvm_type) {
1655         type = mc->kvm_type(kvm_type);
1656     } else if (kvm_type) {
1657         ret = -EINVAL;
1658         fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1659         goto err;
1660     }
1661
1662     do {
1663         ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1664     } while (ret == -EINTR);
1665
1666     if (ret < 0) {
1667         fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1668                 strerror(-ret));
1669
1670 #ifdef TARGET_S390X
1671         if (ret == -EINVAL) {
1672             fprintf(stderr,
1673                     "Host kernel setup problem detected. Please verify:\n");
1674             fprintf(stderr, "- for kernels supporting the switch_amode or"
1675                     " user_mode parameters, whether\n");
1676             fprintf(stderr,
1677                     "  user space is running in primary address space\n");
1678             fprintf(stderr,
1679                     "- for kernels supporting the vm.allocate_pgste sysctl, "
1680                     "whether it is enabled\n");
1681         }
1682 #endif
1683         goto err;
1684     }
1685
1686     s->vmfd = ret;
1687     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1688     if (!missing_cap) {
1689         missing_cap =
1690             kvm_check_extension_list(s, kvm_arch_required_capabilities);
1691     }
1692     if (missing_cap) {
1693         ret = -EINVAL;
1694         fprintf(stderr, "kvm does not support %s\n%s",
1695                 missing_cap->name, upgrade_note);
1696         goto err;
1697     }
1698
1699     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1700
1701     s->broken_set_mem_region = 1;
1702     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1703     if (ret > 0) {
1704         s->broken_set_mem_region = 0;
1705     }
1706
1707 #ifdef KVM_CAP_VCPU_EVENTS
1708     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1709 #endif
1710
1711     s->robust_singlestep =
1712         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1713
1714 #ifdef KVM_CAP_DEBUGREGS
1715     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1716 #endif
1717
1718 #ifdef KVM_CAP_IRQ_ROUTING
1719     kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1720 #endif
1721
1722     s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1723
1724     s->irq_set_ioctl = KVM_IRQ_LINE;
1725     if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1726         s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1727     }
1728
1729 #ifdef KVM_CAP_READONLY_MEM
1730     kvm_readonly_mem_allowed =
1731         (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1732 #endif
1733
1734     kvm_eventfds_allowed =
1735         (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1736
1737     kvm_irqfds_allowed =
1738         (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1739
1740     kvm_resamplefds_allowed =
1741         (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1742
1743     kvm_vm_attributes_allowed =
1744         (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1745
1746     kvm_ioeventfd_any_length_allowed =
1747         (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1748
1749     ret = kvm_arch_init(ms, s);
1750     if (ret < 0) {
1751         goto err;
1752     }
1753
1754     if (machine_kernel_irqchip_allowed(ms)) {
1755         kvm_irqchip_create(ms, s);
1756     }
1757
1758     kvm_state = s;
1759
1760     if (kvm_eventfds_allowed) {
1761         s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1762         s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1763     }
1764     s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1765     s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1766
1767     kvm_memory_listener_register(s, &s->memory_listener,
1768                                  &address_space_memory, 0);
1769     memory_listener_register(&kvm_io_listener,
1770                              &address_space_io);
1771
1772     s->many_ioeventfds = kvm_check_many_ioeventfds();
1773
1774     cpu_interrupt_handler = kvm_handle_interrupt;
1775
1776     return 0;
1777
1778 err:
1779     assert(ret < 0);
1780     if (s->vmfd >= 0) {
1781         close(s->vmfd);
1782     }
1783     if (s->fd != -1) {
1784         close(s->fd);
1785     }
1786     g_free(s->memory_listener.slots);
1787
1788     return ret;
1789 }
1790
1791 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1792 {
1793     s->sigmask_len = sigmask_len;
1794 }
1795
1796 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1797                           int size, uint32_t count)
1798 {
1799     int i;
1800     uint8_t *ptr = data;
1801
1802     for (i = 0; i < count; i++) {
1803         address_space_rw(&address_space_io, port, attrs,
1804                          ptr, size,
1805                          direction == KVM_EXIT_IO_OUT);
1806         ptr += size;
1807     }
1808 }
1809
1810 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1811 {
1812     fprintf(stderr, "KVM internal error. Suberror: %d\n",
1813             run->internal.suberror);
1814
1815     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1816         int i;
1817
1818         for (i = 0; i < run->internal.ndata; ++i) {
1819             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1820                     i, (uint64_t)run->internal.data[i]);
1821         }
1822     }
1823     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1824         fprintf(stderr, "emulation failure\n");
1825         if (!kvm_arch_stop_on_emulation_error(cpu)) {
1826             cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1827             return EXCP_INTERRUPT;
1828         }
1829     }
1830     /* FIXME: Should trigger a qmp message to let management know
1831      * something went wrong.
1832      */
1833     return -1;
1834 }
1835
1836 void kvm_flush_coalesced_mmio_buffer(void)
1837 {
1838     KVMState *s = kvm_state;
1839
1840     if (s->coalesced_flush_in_progress) {
1841         return;
1842     }
1843
1844     s->coalesced_flush_in_progress = true;
1845
1846     if (s->coalesced_mmio_ring) {
1847         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1848         while (ring->first != ring->last) {
1849             struct kvm_coalesced_mmio *ent;
1850
1851             ent = &ring->coalesced_mmio[ring->first];
1852
1853             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1854             smp_wmb();
1855             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1856         }
1857     }
1858
1859     s->coalesced_flush_in_progress = false;
1860 }
1861
1862 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1863 {
1864     if (!cpu->kvm_vcpu_dirty) {
1865         kvm_arch_get_registers(cpu);
1866         cpu->kvm_vcpu_dirty = true;
1867     }
1868 }
1869
1870 void kvm_cpu_synchronize_state(CPUState *cpu)
1871 {
1872     if (!cpu->kvm_vcpu_dirty) {
1873         run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1874     }
1875 }
1876
1877 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1878 {
1879     kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1880     cpu->kvm_vcpu_dirty = false;
1881 }
1882
1883 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1884 {
1885     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1886 }
1887
1888 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1889 {
1890     kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1891     cpu->kvm_vcpu_dirty = false;
1892 }
1893
1894 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1895 {
1896     run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1897 }
1898
1899 #ifdef KVM_HAVE_MCE_INJECTION
1900 static __thread void *pending_sigbus_addr;
1901 static __thread int pending_sigbus_code;
1902 static __thread bool have_sigbus_pending;
1903 #endif
1904
1905 static void kvm_cpu_kick(CPUState *cpu)
1906 {
1907     atomic_set(&cpu->kvm_run->immediate_exit, 1);
1908 }
1909
1910 static void kvm_cpu_kick_self(void)
1911 {
1912     if (kvm_immediate_exit) {
1913         kvm_cpu_kick(current_cpu);
1914     } else {
1915         qemu_cpu_kick_self();
1916     }
1917 }
1918
1919 static void kvm_eat_signals(CPUState *cpu)
1920 {
1921     struct timespec ts = { 0, 0 };
1922     siginfo_t siginfo;
1923     sigset_t waitset;
1924     sigset_t chkset;
1925     int r;
1926
1927     if (kvm_immediate_exit) {
1928         atomic_set(&cpu->kvm_run->immediate_exit, 0);
1929         /* Write kvm_run->immediate_exit before the cpu->exit_request
1930          * write in kvm_cpu_exec.
1931          */
1932         smp_wmb();
1933         return;
1934     }
1935
1936     sigemptyset(&waitset);
1937     sigaddset(&waitset, SIG_IPI);
1938
1939     do {
1940         r = sigtimedwait(&waitset, &siginfo, &ts);
1941         if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1942             perror("sigtimedwait");
1943             exit(1);
1944         }
1945
1946         r = sigpending(&chkset);
1947         if (r == -1) {
1948             perror("sigpending");
1949             exit(1);
1950         }
1951     } while (sigismember(&chkset, SIG_IPI));
1952 }
1953
1954 int kvm_cpu_exec(CPUState *cpu)
1955 {
1956     struct kvm_run *run = cpu->kvm_run;
1957     int ret, run_ret;
1958
1959     DPRINTF("kvm_cpu_exec()\n");
1960
1961     if (kvm_arch_process_async_events(cpu)) {
1962         atomic_set(&cpu->exit_request, 0);
1963         return EXCP_HLT;
1964     }
1965
1966     qemu_mutex_unlock_iothread();
1967
1968     do {
1969         MemTxAttrs attrs;
1970
1971         if (cpu->kvm_vcpu_dirty) {
1972             kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1973             cpu->kvm_vcpu_dirty = false;
1974         }
1975
1976         kvm_arch_pre_run(cpu, run);
1977         if (atomic_read(&cpu->exit_request)) {
1978             DPRINTF("interrupt exit requested\n");
1979             /*
1980              * KVM requires us to reenter the kernel after IO exits to complete
1981              * instruction emulation. This self-signal will ensure that we
1982              * leave ASAP again.
1983              */
1984             kvm_cpu_kick_self();
1985         }
1986
1987         /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1988          * Matching barrier in kvm_eat_signals.
1989          */
1990         smp_rmb();
1991
1992         run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1993
1994         attrs = kvm_arch_post_run(cpu, run);
1995
1996 #ifdef KVM_HAVE_MCE_INJECTION
1997         if (unlikely(have_sigbus_pending)) {
1998             qemu_mutex_lock_iothread();
1999             kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
2000                                     pending_sigbus_addr);
2001             have_sigbus_pending = false;
2002             qemu_mutex_unlock_iothread();
2003         }
2004 #endif
2005
2006         if (run_ret < 0) {
2007             if (run_ret == -EINTR || run_ret == -EAGAIN) {
2008                 DPRINTF("io window exit\n");
2009                 kvm_eat_signals(cpu);
2010                 ret = EXCP_INTERRUPT;
2011                 break;
2012             }
2013             fprintf(stderr, "error: kvm run failed %s\n",
2014                     strerror(-run_ret));
2015 #ifdef TARGET_PPC
2016             if (run_ret == -EBUSY) {
2017                 fprintf(stderr,
2018                         "This is probably because your SMT is enabled.\n"
2019                         "VCPU can only run on primary threads with all "
2020                         "secondary threads offline.\n");
2021             }
2022 #endif
2023             ret = -1;
2024             break;
2025         }
2026
2027         trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2028         switch (run->exit_reason) {
2029         case KVM_EXIT_IO:
2030             DPRINTF("handle_io\n");
2031             /* Called outside BQL */
2032             kvm_handle_io(run->io.port, attrs,
2033                           (uint8_t *)run + run->io.data_offset,
2034                           run->io.direction,
2035                           run->io.size,
2036                           run->io.count);
2037             ret = 0;
2038             break;
2039         case KVM_EXIT_MMIO:
2040             DPRINTF("handle_mmio\n");
2041             /* Called outside BQL */
2042             address_space_rw(&address_space_memory,
2043                              run->mmio.phys_addr, attrs,
2044                              run->mmio.data,
2045                              run->mmio.len,
2046                              run->mmio.is_write);
2047             ret = 0;
2048             break;
2049         case KVM_EXIT_IRQ_WINDOW_OPEN:
2050             DPRINTF("irq_window_open\n");
2051             ret = EXCP_INTERRUPT;
2052             break;
2053         case KVM_EXIT_SHUTDOWN:
2054             DPRINTF("shutdown\n");
2055             qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2056             ret = EXCP_INTERRUPT;
2057             break;
2058         case KVM_EXIT_UNKNOWN:
2059             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2060                     (uint64_t)run->hw.hardware_exit_reason);
2061             ret = -1;
2062             break;
2063         case KVM_EXIT_INTERNAL_ERROR:
2064             ret = kvm_handle_internal_error(cpu, run);
2065             break;
2066         case KVM_EXIT_SYSTEM_EVENT:
2067             switch (run->system_event.type) {
2068             case KVM_SYSTEM_EVENT_SHUTDOWN:
2069                 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2070                 ret = EXCP_INTERRUPT;
2071                 break;
2072             case KVM_SYSTEM_EVENT_RESET:
2073                 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2074                 ret = EXCP_INTERRUPT;
2075                 break;
2076             case KVM_SYSTEM_EVENT_CRASH:
2077                 kvm_cpu_synchronize_state(cpu);
2078                 qemu_mutex_lock_iothread();
2079                 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2080                 qemu_mutex_unlock_iothread();
2081                 ret = 0;
2082                 break;
2083             default:
2084                 DPRINTF("kvm_arch_handle_exit\n");
2085                 ret = kvm_arch_handle_exit(cpu, run);
2086                 break;
2087             }
2088             break;
2089         default:
2090             DPRINTF("kvm_arch_handle_exit\n");
2091             ret = kvm_arch_handle_exit(cpu, run);
2092             break;
2093         }
2094     } while (ret == 0);
2095
2096     qemu_mutex_lock_iothread();
2097
2098     if (ret < 0) {
2099         cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
2100         vm_stop(RUN_STATE_INTERNAL_ERROR);
2101     }
2102
2103     atomic_set(&cpu->exit_request, 0);
2104     return ret;
2105 }
2106
2107 int kvm_ioctl(KVMState *s, int type, ...)
2108 {
2109     int ret;
2110     void *arg;
2111     va_list ap;
2112
2113     va_start(ap, type);
2114     arg = va_arg(ap, void *);
2115     va_end(ap);
2116
2117     trace_kvm_ioctl(type, arg);
2118     ret = ioctl(s->fd, type, arg);
2119     if (ret == -1) {
2120         ret = -errno;
2121     }
2122     return ret;
2123 }
2124
2125 int kvm_vm_ioctl(KVMState *s, int type, ...)
2126 {
2127     int ret;
2128     void *arg;
2129     va_list ap;
2130
2131     va_start(ap, type);
2132     arg = va_arg(ap, void *);
2133     va_end(ap);
2134
2135     trace_kvm_vm_ioctl(type, arg);
2136     ret = ioctl(s->vmfd, type, arg);
2137     if (ret == -1) {
2138         ret = -errno;
2139     }
2140     return ret;
2141 }
2142
2143 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2144 {
2145     int ret;
2146     void *arg;
2147     va_list ap;
2148
2149     va_start(ap, type);
2150     arg = va_arg(ap, void *);
2151     va_end(ap);
2152
2153     trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2154     ret = ioctl(cpu->kvm_fd, type, arg);
2155     if (ret == -1) {
2156         ret = -errno;
2157     }
2158     return ret;
2159 }
2160
2161 int kvm_device_ioctl(int fd, int type, ...)
2162 {
2163     int ret;
2164     void *arg;
2165     va_list ap;
2166
2167     va_start(ap, type);
2168     arg = va_arg(ap, void *);
2169     va_end(ap);
2170
2171     trace_kvm_device_ioctl(fd, type, arg);
2172     ret = ioctl(fd, type, arg);
2173     if (ret == -1) {
2174         ret = -errno;
2175     }
2176     return ret;
2177 }
2178
2179 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2180 {
2181     int ret;
2182     struct kvm_device_attr attribute = {
2183         .group = group,
2184         .attr = attr,
2185     };
2186
2187     if (!kvm_vm_attributes_allowed) {
2188         return 0;
2189     }
2190
2191     ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2192     /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2193     return ret ? 0 : 1;
2194 }
2195
2196 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2197 {
2198     struct kvm_device_attr attribute = {
2199         .group = group,
2200         .attr = attr,
2201         .flags = 0,
2202     };
2203
2204     return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2205 }
2206
2207 void kvm_device_access(int fd, int group, uint64_t attr,
2208                        void *val, bool write)
2209 {
2210     struct kvm_device_attr kvmattr;
2211     int err;
2212
2213     kvmattr.flags = 0;
2214     kvmattr.group = group;
2215     kvmattr.attr = attr;
2216     kvmattr.addr = (uintptr_t)val;
2217
2218     err = kvm_device_ioctl(fd,
2219                            write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2220                            &kvmattr);
2221     if (err < 0) {
2222         error_report("KVM_%s_DEVICE_ATTR failed: %s",
2223                      write ? "SET" : "GET", strerror(-err));
2224         error_printf("Group %d attr 0x%016" PRIx64 "\n", group, attr);
2225         abort();
2226     }
2227 }
2228
2229 /* Return 1 on success, 0 on failure */
2230 int kvm_has_sync_mmu(void)
2231 {
2232     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2233 }
2234
2235 int kvm_has_vcpu_events(void)
2236 {
2237     return kvm_state->vcpu_events;
2238 }
2239
2240 int kvm_has_robust_singlestep(void)
2241 {
2242     return kvm_state->robust_singlestep;
2243 }
2244
2245 int kvm_has_debugregs(void)
2246 {
2247     return kvm_state->debugregs;
2248 }
2249
2250 int kvm_has_many_ioeventfds(void)
2251 {
2252     if (!kvm_enabled()) {
2253         return 0;
2254     }
2255     return kvm_state->many_ioeventfds;
2256 }
2257
2258 int kvm_has_gsi_routing(void)
2259 {
2260 #ifdef KVM_CAP_IRQ_ROUTING
2261     return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2262 #else
2263     return false;
2264 #endif
2265 }
2266
2267 int kvm_has_intx_set_mask(void)
2268 {
2269     return kvm_state->intx_set_mask;
2270 }
2271
2272 #ifdef KVM_CAP_SET_GUEST_DEBUG
2273 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2274                                                  target_ulong pc)
2275 {
2276     struct kvm_sw_breakpoint *bp;
2277
2278     QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2279         if (bp->pc == pc) {
2280             return bp;
2281         }
2282     }
2283     return NULL;
2284 }
2285
2286 int kvm_sw_breakpoints_active(CPUState *cpu)
2287 {
2288     return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2289 }
2290
2291 struct kvm_set_guest_debug_data {
2292     struct kvm_guest_debug dbg;
2293     int err;
2294 };
2295
2296 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2297 {
2298     struct kvm_set_guest_debug_data *dbg_data =
2299         (struct kvm_set_guest_debug_data *) data.host_ptr;
2300
2301     dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2302                                    &dbg_data->dbg);
2303 }
2304
2305 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2306 {
2307     struct kvm_set_guest_debug_data data;
2308
2309     data.dbg.control = reinject_trap;
2310
2311     if (cpu->singlestep_enabled) {
2312         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2313     }
2314     kvm_arch_update_guest_debug(cpu, &data.dbg);
2315
2316     run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2317                RUN_ON_CPU_HOST_PTR(&data));
2318     return data.err;
2319 }
2320
2321 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2322                           target_ulong len, int type)
2323 {
2324     struct kvm_sw_breakpoint *bp;
2325     int err;
2326
2327     if (type == GDB_BREAKPOINT_SW) {
2328         bp = kvm_find_sw_breakpoint(cpu, addr);
2329         if (bp) {
2330             bp->use_count++;
2331             return 0;
2332         }
2333
2334         bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2335         bp->pc = addr;
2336         bp->use_count = 1;
2337         err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2338         if (err) {
2339             g_free(bp);
2340             return err;
2341         }
2342
2343         QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2344     } else {
2345         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2346         if (err) {
2347             return err;
2348         }
2349     }
2350
2351     CPU_FOREACH(cpu) {
2352         err = kvm_update_guest_debug(cpu, 0);
2353         if (err) {
2354             return err;
2355         }
2356     }
2357     return 0;
2358 }
2359
2360 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2361                           target_ulong len, int type)
2362 {
2363     struct kvm_sw_breakpoint *bp;
2364     int err;
2365
2366     if (type == GDB_BREAKPOINT_SW) {
2367         bp = kvm_find_sw_breakpoint(cpu, addr);
2368         if (!bp) {
2369             return -ENOENT;
2370         }
2371
2372         if (bp->use_count > 1) {
2373             bp->use_count--;
2374             return 0;
2375         }
2376
2377         err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2378         if (err) {
2379             return err;
2380         }
2381
2382         QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2383         g_free(bp);
2384     } else {
2385         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2386         if (err) {
2387             return err;
2388         }
2389     }
2390
2391     CPU_FOREACH(cpu) {
2392         err = kvm_update_guest_debug(cpu, 0);
2393         if (err) {
2394             return err;
2395         }
2396     }
2397     return 0;
2398 }
2399
2400 void kvm_remove_all_breakpoints(CPUState *cpu)
2401 {
2402     struct kvm_sw_breakpoint *bp, *next;
2403     KVMState *s = cpu->kvm_state;
2404     CPUState *tmpcpu;
2405
2406     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2407         if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2408             /* Try harder to find a CPU that currently sees the breakpoint. */
2409             CPU_FOREACH(tmpcpu) {
2410                 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2411                     break;
2412                 }
2413             }
2414         }
2415         QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2416         g_free(bp);
2417     }
2418     kvm_arch_remove_all_hw_breakpoints();
2419
2420     CPU_FOREACH(cpu) {
2421         kvm_update_guest_debug(cpu, 0);
2422     }
2423 }
2424
2425 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2426
2427 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2428 {
2429     return -EINVAL;
2430 }
2431
2432 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2433                           target_ulong len, int type)
2434 {
2435     return -EINVAL;
2436 }
2437
2438 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2439                           target_ulong len, int type)
2440 {
2441     return -EINVAL;
2442 }
2443
2444 void kvm_remove_all_breakpoints(CPUState *cpu)
2445 {
2446 }
2447 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2448
2449 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2450 {
2451     KVMState *s = kvm_state;
2452     struct kvm_signal_mask *sigmask;
2453     int r;
2454
2455     sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2456
2457     sigmask->len = s->sigmask_len;
2458     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2459     r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2460     g_free(sigmask);
2461
2462     return r;
2463 }
2464
2465 static void kvm_ipi_signal(int sig)
2466 {
2467     if (current_cpu) {
2468         assert(kvm_immediate_exit);
2469         kvm_cpu_kick(current_cpu);
2470     }
2471 }
2472
2473 void kvm_init_cpu_signals(CPUState *cpu)
2474 {
2475     int r;
2476     sigset_t set;
2477     struct sigaction sigact;
2478
2479     memset(&sigact, 0, sizeof(sigact));
2480     sigact.sa_handler = kvm_ipi_signal;
2481     sigaction(SIG_IPI, &sigact, NULL);
2482
2483     pthread_sigmask(SIG_BLOCK, NULL, &set);
2484 #if defined KVM_HAVE_MCE_INJECTION
2485     sigdelset(&set, SIGBUS);
2486     pthread_sigmask(SIG_SETMASK, &set, NULL);
2487 #endif
2488     sigdelset(&set, SIG_IPI);
2489     if (kvm_immediate_exit) {
2490         r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2491     } else {
2492         r = kvm_set_signal_mask(cpu, &set);
2493     }
2494     if (r) {
2495         fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2496         exit(1);
2497     }
2498 }
2499
2500 /* Called asynchronously in VCPU thread.  */
2501 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2502 {
2503 #ifdef KVM_HAVE_MCE_INJECTION
2504     if (have_sigbus_pending) {
2505         return 1;
2506     }
2507     have_sigbus_pending = true;
2508     pending_sigbus_addr = addr;
2509     pending_sigbus_code = code;
2510     atomic_set(&cpu->exit_request, 1);
2511     return 0;
2512 #else
2513     return 1;
2514 #endif
2515 }
2516
2517 /* Called synchronously (via signalfd) in main thread.  */
2518 int kvm_on_sigbus(int code, void *addr)
2519 {
2520 #ifdef KVM_HAVE_MCE_INJECTION
2521     /* Action required MCE kills the process if SIGBUS is blocked.  Because
2522      * that's what happens in the I/O thread, where we handle MCE via signalfd,
2523      * we can only get action optional here.
2524      */
2525     assert(code != BUS_MCEERR_AR);
2526     kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2527     return 0;
2528 #else
2529     return 1;
2530 #endif
2531 }
2532
2533 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2534 {
2535     int ret;
2536     struct kvm_create_device create_dev;
2537
2538     create_dev.type = type;
2539     create_dev.fd = -1;
2540     create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2541
2542     if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2543         return -ENOTSUP;
2544     }
2545
2546     ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2547     if (ret) {
2548         return ret;
2549     }
2550
2551     return test ? 0 : create_dev.fd;
2552 }
2553
2554 bool kvm_device_supported(int vmfd, uint64_t type)
2555 {
2556     struct kvm_create_device create_dev = {
2557         .type = type,
2558         .fd = -1,
2559         .flags = KVM_CREATE_DEVICE_TEST,
2560     };
2561
2562     if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2563         return false;
2564     }
2565
2566     return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2567 }
2568
2569 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2570 {
2571     struct kvm_one_reg reg;
2572     int r;
2573
2574     reg.id = id;
2575     reg.addr = (uintptr_t) source;
2576     r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2577     if (r) {
2578         trace_kvm_failed_reg_set(id, strerror(-r));
2579     }
2580     return r;
2581 }
2582
2583 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2584 {
2585     struct kvm_one_reg reg;
2586     int r;
2587
2588     reg.id = id;
2589     reg.addr = (uintptr_t) target;
2590     r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2591     if (r) {
2592         trace_kvm_failed_reg_get(id, strerror(-r));
2593     }
2594     return r;
2595 }
2596
2597 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2598 {
2599     AccelClass *ac = ACCEL_CLASS(oc);
2600     ac->name = "KVM";
2601     ac->init_machine = kvm_init;
2602     ac->allowed = &kvm_allowed;
2603 }
2604
2605 static const TypeInfo kvm_accel_type = {
2606     .name = TYPE_KVM_ACCEL,
2607     .parent = TYPE_ACCEL,
2608     .class_init = kvm_accel_class_init,
2609     .instance_size = sizeof(KVMState),
2610 };
2611
2612 static void kvm_type_init(void)
2613 {
2614     type_register_static(&kvm_accel_type);
2615 }
2616
2617 type_init(kvm_type_init);