kvm-all.c

   1 /*
   2  * QEMU KVM support
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *           Red Hat, Inc. 2008
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Glauber Costa     <gcosta@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include <sys/types.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/mman.h>
  19 #include <stdarg.h>
  20
  21 #include <linux/kvm.h>
  22
  23 #include "qemu-common.h"
  24 #include "qemu-barrier.h"
  25 #include "sysemu.h"
  26 #include "hw/hw.h"
  27 #include "gdbstub.h"
  28 #include "kvm.h"
  29 #include "bswap.h"
  30
  31 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  32 #define PAGE_SIZE TARGET_PAGE_SIZE
  33
  34 //#define DEBUG_KVM
  35
  36 #ifdef DEBUG_KVM
  37 #define DPRINTF(fmt, ...) \
  38     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  39 #else
  40 #define DPRINTF(fmt, ...) \
  41     do { } while (0)
  42 #endif
  43
  44 typedef struct KVMSlot
  45 {
  46     target_phys_addr_t start_addr;
  47     ram_addr_t memory_size;
  48     ram_addr_t phys_offset;
  49     int slot;
  50     int flags;
  51 } KVMSlot;
  52
  53 typedef struct kvm_dirty_log KVMDirtyLog;
  54
  55 struct KVMState
  56 {
  57     KVMSlot slots[32];
  58     int fd;
  59     int vmfd;
  60     int coalesced_mmio;
  61 #ifdef KVM_CAP_COALESCED_MMIO
  62     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  63 #endif
  64     int broken_set_mem_region;
  65     int migration_log;
  66     int vcpu_events;
  67     int robust_singlestep;
  68     int debugregs;
  69 #ifdef KVM_CAP_SET_GUEST_DEBUG
  70     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  71 #endif
  72     int irqchip_in_kernel;
  73     int pit_in_kernel;
  74     int xsave, xcrs;
  75 };
  76
  77 static KVMState *kvm_state;
  78
  79 static KVMSlot *kvm_alloc_slot(KVMState *s)
  80 {
  81     int i;
  82
  83     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  84         /* KVM private memory slots */
  85         if (i >= 8 && i < 12)
  86             continue;
  87         if (s->slots[i].memory_size == 0)
  88             return &s->slots[i];
  89     }
  90
  91     fprintf(stderr, "%s: no free slot available\n", __func__);
  92     abort();
  93 }
  94
  95 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
  96                                          target_phys_addr_t start_addr,
  97                                          target_phys_addr_t end_addr)
  98 {
  99     int i;
 100
 101     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 102         KVMSlot *mem = &s->slots[i];
 103
 104         if (start_addr == mem->start_addr &&
 105             end_addr == mem->start_addr + mem->memory_size) {
 106             return mem;
 107         }
 108     }
 109
 110     return NULL;
 111 }
 112
 113 /*
 114  * Find overlapping slot with lowest start address
 115  */
 116 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 117                                             target_phys_addr_t start_addr,
 118                                             target_phys_addr_t end_addr)
 119 {
 120     KVMSlot *found = NULL;
 121     int i;
 122
 123     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 124         KVMSlot *mem = &s->slots[i];
 125
 126         if (mem->memory_size == 0 ||
 127             (found && found->start_addr < mem->start_addr)) {
 128             continue;
 129         }
 130
 131         if (end_addr > mem->start_addr &&
 132             start_addr < mem->start_addr + mem->memory_size) {
 133             found = mem;
 134         }
 135     }
 136
 137     return found;
 138 }
 139
 140 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
 141                                       target_phys_addr_t *phys_addr)
 142 {
 143     int i;
 144
 145     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 146         KVMSlot *mem = &s->slots[i];
 147
 148         if (ram_addr >= mem->phys_offset &&
 149             ram_addr < mem->phys_offset + mem->memory_size) {
 150             *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
 151             return 1;
 152         }
 153     }
 154
 155     return 0;
 156 }
 157
 158 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 159 {
 160     struct kvm_userspace_memory_region mem;
 161
 162     mem.slot = slot->slot;
 163     mem.guest_phys_addr = slot->start_addr;
 164     mem.memory_size = slot->memory_size;
 165     mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
 166     mem.flags = slot->flags;
 167     if (s->migration_log) {
 168         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 169     }
 170     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 171 }
 172
 173 static void kvm_reset_vcpu(void *opaque)
 174 {
 175     CPUState *env = opaque;
 176
 177     kvm_arch_reset_vcpu(env);
 178 }
 179
 180 int kvm_irqchip_in_kernel(void)
 181 {
 182     return kvm_state->irqchip_in_kernel;
 183 }
 184
 185 int kvm_pit_in_kernel(void)
 186 {
 187     return kvm_state->pit_in_kernel;
 188 }
 189
 190
 191 int kvm_init_vcpu(CPUState *env)
 192 {
 193     KVMState *s = kvm_state;
 194     long mmap_size;
 195     int ret;
 196
 197     DPRINTF("kvm_init_vcpu\n");
 198
 199     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 200     if (ret < 0) {
 201         DPRINTF("kvm_create_vcpu failed\n");
 202         goto err;
 203     }
 204
 205     env->kvm_fd = ret;
 206     env->kvm_state = s;
 207
 208     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 209     if (mmap_size < 0) {
 210         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 211         goto err;
 212     }
 213
 214     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 215                         env->kvm_fd, 0);
 216     if (env->kvm_run == MAP_FAILED) {
 217         ret = -errno;
 218         DPRINTF("mmap'ing vcpu state failed\n");
 219         goto err;
 220     }
 221
 222 #ifdef KVM_CAP_COALESCED_MMIO
 223     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 224         s->coalesced_mmio_ring = (void *) env->kvm_run +
 225                 s->coalesced_mmio * PAGE_SIZE;
 226 #endif
 227
 228     ret = kvm_arch_init_vcpu(env);
 229     if (ret == 0) {
 230         qemu_register_reset(kvm_reset_vcpu, env);
 231         kvm_arch_reset_vcpu(env);
 232     }
 233 err:
 234     return ret;
 235 }
 236
 237 /*
 238  * dirty pages logging control
 239  */
 240 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 241                                       ram_addr_t size, int flags, int mask)
 242 {
 243     KVMState *s = kvm_state;
 244     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 245     int old_flags;
 246
 247     if (mem == NULL)  {
 248             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 249                     TARGET_FMT_plx "\n", __func__, phys_addr,
 250                     (target_phys_addr_t)(phys_addr + size - 1));
 251             return -EINVAL;
 252     }
 253
 254     old_flags = mem->flags;
 255
 256     flags = (mem->flags & ~mask) | flags;
 257     mem->flags = flags;
 258
 259     /* If nothing changed effectively, no need to issue ioctl */
 260     if (s->migration_log) {
 261         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 262     }
 263     if (flags == old_flags) {
 264             return 0;
 265     }
 266
 267     return kvm_set_user_memory_region(s, mem);
 268 }
 269
 270 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 271 {
 272         return kvm_dirty_pages_log_change(phys_addr, size,
 273                                           KVM_MEM_LOG_DIRTY_PAGES,
 274                                           KVM_MEM_LOG_DIRTY_PAGES);
 275 }
 276
 277 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 278 {
 279         return kvm_dirty_pages_log_change(phys_addr, size,
 280                                           0,
 281                                           KVM_MEM_LOG_DIRTY_PAGES);
 282 }
 283
 284 static int kvm_set_migration_log(int enable)
 285 {
 286     KVMState *s = kvm_state;
 287     KVMSlot *mem;
 288     int i, err;
 289
 290     s->migration_log = enable;
 291
 292     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 293         mem = &s->slots[i];
 294
 295         if (!mem->memory_size) {
 296             continue;
 297         }
 298         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 299             continue;
 300         }
 301         err = kvm_set_user_memory_region(s, mem);
 302         if (err) {
 303             return err;
 304         }
 305     }
 306     return 0;
 307 }
 308
 309 /* get kvm's dirty pages bitmap and update qemu's */
 310 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 311                                          unsigned long *bitmap,
 312                                          unsigned long offset,
 313                                          unsigned long mem_size)
 314 {
 315     unsigned int i, j;
 316     unsigned long page_number, addr, addr1, c;
 317     ram_addr_t ram_addr;
 318     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 319         HOST_LONG_BITS;
 320
 321     /*
 322      * bitmap-traveling is faster than memory-traveling (for addr...)
 323      * especially when most of the memory is not dirty.
 324      */
 325     for (i = 0; i < len; i++) {
 326         if (bitmap[i] != 0) {
 327             c = leul_to_cpu(bitmap[i]);
 328             do {
 329                 j = ffsl(c) - 1;
 330                 c &= ~(1ul << j);
 331                 page_number = i * HOST_LONG_BITS + j;
 332                 addr1 = page_number * TARGET_PAGE_SIZE;
 333                 addr = offset + addr1;
 334                 ram_addr = cpu_get_physical_page_desc(addr);
 335                 cpu_physical_memory_set_dirty(ram_addr);
 336             } while (c != 0);
 337         }
 338     }
 339     return 0;
 340 }
 341
 342 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 343
 344 /**
 345  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 346  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 347  * This means all bits are set to dirty.
 348  *
 349  * @start_add: start of logged region.
 350  * @end_addr: end of logged region.
 351  */
 352 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 353                                           target_phys_addr_t end_addr)
 354 {
 355     KVMState *s = kvm_state;
 356     unsigned long size, allocated_size = 0;
 357     KVMDirtyLog d;
 358     KVMSlot *mem;
 359     int ret = 0;
 360
 361     d.dirty_bitmap = NULL;
 362     while (start_addr < end_addr) {
 363         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 364         if (mem == NULL) {
 365             break;
 366         }
 367
 368         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 369         if (!d.dirty_bitmap) {
 370             d.dirty_bitmap = qemu_malloc(size);
 371         } else if (size > allocated_size) {
 372             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 373         }
 374         allocated_size = size;
 375         memset(d.dirty_bitmap, 0, allocated_size);
 376
 377         d.slot = mem->slot;
 378
 379         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 380             DPRINTF("ioctl failed %d\n", errno);
 381             ret = -1;
 382             break;
 383         }
 384
 385         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 386                                       mem->start_addr, mem->memory_size);
 387         start_addr = mem->start_addr + mem->memory_size;
 388     }
 389     qemu_free(d.dirty_bitmap);
 390
 391     return ret;
 392 }
 393
 394 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 395 {
 396     int ret = -ENOSYS;
 397 #ifdef KVM_CAP_COALESCED_MMIO
 398     KVMState *s = kvm_state;
 399
 400     if (s->coalesced_mmio) {
 401         struct kvm_coalesced_mmio_zone zone;
 402
 403         zone.addr = start;
 404         zone.size = size;
 405
 406         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 407     }
 408 #endif
 409
 410     return ret;
 411 }
 412
 413 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 414 {
 415     int ret = -ENOSYS;
 416 #ifdef KVM_CAP_COALESCED_MMIO
 417     KVMState *s = kvm_state;
 418
 419     if (s->coalesced_mmio) {
 420         struct kvm_coalesced_mmio_zone zone;
 421
 422         zone.addr = start;
 423         zone.size = size;
 424
 425         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 426     }
 427 #endif
 428
 429     return ret;
 430 }
 431
 432 int kvm_check_extension(KVMState *s, unsigned int extension)
 433 {
 434     int ret;
 435
 436     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 437     if (ret < 0) {
 438         ret = 0;
 439     }
 440
 441     return ret;
 442 }
 443
 444 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 445                              ram_addr_t size,
 446                              ram_addr_t phys_offset)
 447 {
 448     KVMState *s = kvm_state;
 449     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 450     KVMSlot *mem, old;
 451     int err;
 452
 453     /* kvm works in page size chunks, but the function may be called
 454        with sub-page size and unaligned start address. */
 455     size = TARGET_PAGE_ALIGN(size);
 456     start_addr = TARGET_PAGE_ALIGN(start_addr);
 457
 458     /* KVM does not support read-only slots */
 459     phys_offset &= ~IO_MEM_ROM;
 460
 461     while (1) {
 462         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 463         if (!mem) {
 464             break;
 465         }
 466
 467         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 468             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 469             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 470             /* The new slot fits into the existing one and comes with
 471              * identical parameters - nothing to be done. */
 472             return;
 473         }
 474
 475         old = *mem;
 476
 477         /* unregister the overlapping slot */
 478         mem->memory_size = 0;
 479         err = kvm_set_user_memory_region(s, mem);
 480         if (err) {
 481             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 482                     __func__, strerror(-err));
 483             abort();
 484         }
 485
 486         /* Workaround for older KVM versions: we can't join slots, even not by
 487          * unregistering the previous ones and then registering the larger
 488          * slot. We have to maintain the existing fragmentation. Sigh.
 489          *
 490          * This workaround assumes that the new slot starts at the same
 491          * address as the first existing one. If not or if some overlapping
 492          * slot comes around later, we will fail (not seen in practice so far)
 493          * - and actually require a recent KVM version. */
 494         if (s->broken_set_mem_region &&
 495             old.start_addr == start_addr && old.memory_size < size &&
 496             flags < IO_MEM_UNASSIGNED) {
 497             mem = kvm_alloc_slot(s);
 498             mem->memory_size = old.memory_size;
 499             mem->start_addr = old.start_addr;
 500             mem->phys_offset = old.phys_offset;
 501             mem->flags = 0;
 502
 503             err = kvm_set_user_memory_region(s, mem);
 504             if (err) {
 505                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 506                         strerror(-err));
 507                 abort();
 508             }
 509
 510             start_addr += old.memory_size;
 511             phys_offset += old.memory_size;
 512             size -= old.memory_size;
 513             continue;
 514         }
 515
 516         /* register prefix slot */
 517         if (old.start_addr < start_addr) {
 518             mem = kvm_alloc_slot(s);
 519             mem->memory_size = start_addr - old.start_addr;
 520             mem->start_addr = old.start_addr;
 521             mem->phys_offset = old.phys_offset;
 522             mem->flags = 0;
 523
 524             err = kvm_set_user_memory_region(s, mem);
 525             if (err) {
 526                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 527                         __func__, strerror(-err));
 528                 abort();
 529             }
 530         }
 531
 532         /* register suffix slot */
 533         if (old.start_addr + old.memory_size > start_addr + size) {
 534             ram_addr_t size_delta;
 535
 536             mem = kvm_alloc_slot(s);
 537             mem->start_addr = start_addr + size;
 538             size_delta = mem->start_addr - old.start_addr;
 539             mem->memory_size = old.memory_size - size_delta;
 540             mem->phys_offset = old.phys_offset + size_delta;
 541             mem->flags = 0;
 542
 543             err = kvm_set_user_memory_region(s, mem);
 544             if (err) {
 545                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 546                         __func__, strerror(-err));
 547                 abort();
 548             }
 549         }
 550     }
 551
 552     /* in case the KVM bug workaround already "consumed" the new slot */
 553     if (!size)
 554         return;
 555
 556     /* KVM does not need to know about this memory */
 557     if (flags >= IO_MEM_UNASSIGNED)
 558         return;
 559
 560     mem = kvm_alloc_slot(s);
 561     mem->memory_size = size;
 562     mem->start_addr = start_addr;
 563     mem->phys_offset = phys_offset;
 564     mem->flags = 0;
 565
 566     err = kvm_set_user_memory_region(s, mem);
 567     if (err) {
 568         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 569                 strerror(-err));
 570         abort();
 571     }
 572 }
 573
 574 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 575                                   target_phys_addr_t start_addr,
 576                                   ram_addr_t size,
 577                                   ram_addr_t phys_offset)
 578 {
 579         kvm_set_phys_mem(start_addr, size, phys_offset);
 580 }
 581
 582 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 583                                         target_phys_addr_t start_addr,
 584                                         target_phys_addr_t end_addr)
 585 {
 586         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 587 }
 588
 589 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 590                                     int enable)
 591 {
 592         return kvm_set_migration_log(enable);
 593 }
 594
 595 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 596         .set_memory = kvm_client_set_memory,
 597         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 598         .migration_log = kvm_client_migration_log,
 599 };
 600
 601 int kvm_init(int smp_cpus)
 602 {
 603     static const char upgrade_note[] =
 604         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 605         "(see http://sourceforge.net/projects/kvm).\n";
 606     KVMState *s;
 607     int ret;
 608     int i;
 609
 610     s = qemu_mallocz(sizeof(KVMState));
 611
 612 #ifdef KVM_CAP_SET_GUEST_DEBUG
 613     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 614 #endif
 615     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 616         s->slots[i].slot = i;
 617
 618     s->vmfd = -1;
 619     s->fd = qemu_open("/dev/kvm", O_RDWR);
 620     if (s->fd == -1) {
 621         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 622         ret = -errno;
 623         goto err;
 624     }
 625
 626     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 627     if (ret < KVM_API_VERSION) {
 628         if (ret > 0)
 629             ret = -EINVAL;
 630         fprintf(stderr, "kvm version too old\n");
 631         goto err;
 632     }
 633
 634     if (ret > KVM_API_VERSION) {
 635         ret = -EINVAL;
 636         fprintf(stderr, "kvm version not supported\n");
 637         goto err;
 638     }
 639
 640     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 641     if (s->vmfd < 0) {
 642 #ifdef TARGET_S390X
 643         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 644                         "your host kernel command line\n");
 645 #endif
 646         goto err;
 647     }
 648
 649     /* initially, KVM allocated its own memory and we had to jump through
 650      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 651      * just use a user allocated buffer so we can use regular pages
 652      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 653      */
 654     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 655         ret = -EINVAL;
 656         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 657                 upgrade_note);
 658         goto err;
 659     }
 660
 661     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 662      * destroyed properly.  Since we rely on this capability, refuse to work
 663      * with any kernel without this capability. */
 664     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 665         ret = -EINVAL;
 666
 667         fprintf(stderr,
 668                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 669                 upgrade_note);
 670         goto err;
 671     }
 672
 673     s->coalesced_mmio = 0;
 674 #ifdef KVM_CAP_COALESCED_MMIO
 675     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 676     s->coalesced_mmio_ring = NULL;
 677 #endif
 678
 679     s->broken_set_mem_region = 1;
 680 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 681     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 682     if (ret > 0) {
 683         s->broken_set_mem_region = 0;
 684     }
 685 #endif
 686
 687     s->vcpu_events = 0;
 688 #ifdef KVM_CAP_VCPU_EVENTS
 689     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 690 #endif
 691
 692     s->robust_singlestep = 0;
 693 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 694     s->robust_singlestep =
 695         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 696 #endif
 697
 698     s->debugregs = 0;
 699 #ifdef KVM_CAP_DEBUGREGS
 700     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 701 #endif
 702
 703     s->xsave = 0;
 704 #ifdef KVM_CAP_XSAVE
 705     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 706 #endif
 707
 708     s->xcrs = 0;
 709 #ifdef KVM_CAP_XCRS
 710     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 711 #endif
 712
 713     ret = kvm_arch_init(s, smp_cpus);
 714     if (ret < 0)
 715         goto err;
 716
 717     kvm_state = s;
 718     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 719
 720     return 0;
 721
 722 err:
 723     if (s) {
 724         if (s->vmfd != -1)
 725             close(s->vmfd);
 726         if (s->fd != -1)
 727             close(s->fd);
 728     }
 729     qemu_free(s);
 730
 731     return ret;
 732 }
 733
 734 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 735                          uint32_t count)
 736 {
 737     int i;
 738     uint8_t *ptr = data;
 739
 740     for (i = 0; i < count; i++) {
 741         if (direction == KVM_EXIT_IO_IN) {
 742             switch (size) {
 743             case 1:
 744                 stb_p(ptr, cpu_inb(port));
 745                 break;
 746             case 2:
 747                 stw_p(ptr, cpu_inw(port));
 748                 break;
 749             case 4:
 750                 stl_p(ptr, cpu_inl(port));
 751                 break;
 752             }
 753         } else {
 754             switch (size) {
 755             case 1:
 756                 cpu_outb(port, ldub_p(ptr));
 757                 break;
 758             case 2:
 759                 cpu_outw(port, lduw_p(ptr));
 760                 break;
 761             case 4:
 762                 cpu_outl(port, ldl_p(ptr));
 763                 break;
 764             }
 765         }
 766
 767         ptr += size;
 768     }
 769
 770     return 1;
 771 }
 772
 773 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 774 static void kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 775 {
 776
 777     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 778         int i;
 779
 780         fprintf(stderr, "KVM internal error. Suberror: %d\n",
 781                 run->internal.suberror);
 782
 783         for (i = 0; i < run->internal.ndata; ++i) {
 784             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 785                     i, (uint64_t)run->internal.data[i]);
 786         }
 787     }
 788     cpu_dump_state(env, stderr, fprintf, 0);
 789     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 790         fprintf(stderr, "emulation failure\n");
 791         if (!kvm_arch_stop_on_emulation_error(env))
 792                 return;
 793     }
 794     /* FIXME: Should trigger a qmp message to let management know
 795      * something went wrong.
 796      */
 797     vm_stop(0);
 798 }
 799 #endif
 800
 801 void kvm_flush_coalesced_mmio_buffer(void)
 802 {
 803 #ifdef KVM_CAP_COALESCED_MMIO
 804     KVMState *s = kvm_state;
 805     if (s->coalesced_mmio_ring) {
 806         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 807         while (ring->first != ring->last) {
 808             struct kvm_coalesced_mmio *ent;
 809
 810             ent = &ring->coalesced_mmio[ring->first];
 811
 812             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 813             smp_wmb();
 814             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 815         }
 816     }
 817 #endif
 818 }
 819
 820 static void do_kvm_cpu_synchronize_state(void *_env)
 821 {
 822     CPUState *env = _env;
 823
 824     if (!env->kvm_vcpu_dirty) {
 825         kvm_arch_get_registers(env);
 826         env->kvm_vcpu_dirty = 1;
 827     }
 828 }
 829
 830 void kvm_cpu_synchronize_state(CPUState *env)
 831 {
 832     if (!env->kvm_vcpu_dirty)
 833         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 834 }
 835
 836 void kvm_cpu_synchronize_post_reset(CPUState *env)
 837 {
 838     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 839     env->kvm_vcpu_dirty = 0;
 840 }
 841
 842 void kvm_cpu_synchronize_post_init(CPUState *env)
 843 {
 844     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 845     env->kvm_vcpu_dirty = 0;
 846 }
 847
 848 int kvm_cpu_exec(CPUState *env)
 849 {
 850     struct kvm_run *run = env->kvm_run;
 851     int ret;
 852
 853     DPRINTF("kvm_cpu_exec()\n");
 854
 855     do {
 856 #ifndef CONFIG_IOTHREAD
 857         if (env->exit_request) {
 858             DPRINTF("interrupt exit requested\n");
 859             ret = 0;
 860             break;
 861         }
 862 #endif
 863
 864         if (kvm_arch_process_irqchip_events(env)) {
 865             ret = 0;
 866             break;
 867         }
 868
 869         if (env->kvm_vcpu_dirty) {
 870             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 871             env->kvm_vcpu_dirty = 0;
 872         }
 873
 874         kvm_arch_pre_run(env, run);
 875         cpu_single_env = NULL;
 876         qemu_mutex_unlock_iothread();
 877         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 878         qemu_mutex_lock_iothread();
 879         cpu_single_env = env;
 880         kvm_arch_post_run(env, run);
 881
 882         if (ret == -EINTR || ret == -EAGAIN) {
 883             cpu_exit(env);
 884             DPRINTF("io window exit\n");
 885             ret = 0;
 886             break;
 887         }
 888
 889         if (ret < 0) {
 890             DPRINTF("kvm run failed %s\n", strerror(-ret));
 891             abort();
 892         }
 893
 894         kvm_flush_coalesced_mmio_buffer();
 895
 896         ret = 0; /* exit loop */
 897         switch (run->exit_reason) {
 898         case KVM_EXIT_IO:
 899             DPRINTF("handle_io\n");
 900             ret = kvm_handle_io(run->io.port,
 901                                 (uint8_t *)run + run->io.data_offset,
 902                                 run->io.direction,
 903                                 run->io.size,
 904                                 run->io.count);
 905             break;
 906         case KVM_EXIT_MMIO:
 907             DPRINTF("handle_mmio\n");
 908             cpu_physical_memory_rw(run->mmio.phys_addr,
 909                                    run->mmio.data,
 910                                    run->mmio.len,
 911                                    run->mmio.is_write);
 912             ret = 1;
 913             break;
 914         case KVM_EXIT_IRQ_WINDOW_OPEN:
 915             DPRINTF("irq_window_open\n");
 916             break;
 917         case KVM_EXIT_SHUTDOWN:
 918             DPRINTF("shutdown\n");
 919             qemu_system_reset_request();
 920             ret = 1;
 921             break;
 922         case KVM_EXIT_UNKNOWN:
 923             DPRINTF("kvm_exit_unknown\n");
 924             break;
 925         case KVM_EXIT_FAIL_ENTRY:
 926             DPRINTF("kvm_exit_fail_entry\n");
 927             break;
 928         case KVM_EXIT_EXCEPTION:
 929             DPRINTF("kvm_exit_exception\n");
 930             break;
 931 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 932         case KVM_EXIT_INTERNAL_ERROR:
 933             kvm_handle_internal_error(env, run);
 934             break;
 935 #endif
 936         case KVM_EXIT_DEBUG:
 937             DPRINTF("kvm_exit_debug\n");
 938 #ifdef KVM_CAP_SET_GUEST_DEBUG
 939             if (kvm_arch_debug(&run->debug.arch)) {
 940                 env->exception_index = EXCP_DEBUG;
 941                 return 0;
 942             }
 943             /* re-enter, this exception was guest-internal */
 944             ret = 1;
 945 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 946             break;
 947         default:
 948             DPRINTF("kvm_arch_handle_exit\n");
 949             ret = kvm_arch_handle_exit(env, run);
 950             break;
 951         }
 952     } while (ret > 0);
 953
 954     if (env->exit_request) {
 955         env->exit_request = 0;
 956         env->exception_index = EXCP_INTERRUPT;
 957     }
 958
 959     return ret;
 960 }
 961
 962 int kvm_ioctl(KVMState *s, int type, ...)
 963 {
 964     int ret;
 965     void *arg;
 966     va_list ap;
 967
 968     va_start(ap, type);
 969     arg = va_arg(ap, void *);
 970     va_end(ap);
 971
 972     ret = ioctl(s->fd, type, arg);
 973     if (ret == -1)
 974         ret = -errno;
 975
 976     return ret;
 977 }
 978
 979 int kvm_vm_ioctl(KVMState *s, int type, ...)
 980 {
 981     int ret;
 982     void *arg;
 983     va_list ap;
 984
 985     va_start(ap, type);
 986     arg = va_arg(ap, void *);
 987     va_end(ap);
 988
 989     ret = ioctl(s->vmfd, type, arg);
 990     if (ret == -1)
 991         ret = -errno;
 992
 993     return ret;
 994 }
 995
 996 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 997 {
 998     int ret;
 999     void *arg;
1000     va_list ap;
1001
1002     va_start(ap, type);
1003     arg = va_arg(ap, void *);
1004     va_end(ap);
1005
1006     ret = ioctl(env->kvm_fd, type, arg);
1007     if (ret == -1)
1008         ret = -errno;
1009
1010     return ret;
1011 }
1012
1013 int kvm_has_sync_mmu(void)
1014 {
1015 #ifdef KVM_CAP_SYNC_MMU
1016     KVMState *s = kvm_state;
1017
1018     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
1019 #else
1020     return 0;
1021 #endif
1022 }
1023
1024 int kvm_has_vcpu_events(void)
1025 {
1026     return kvm_state->vcpu_events;
1027 }
1028
1029 int kvm_has_robust_singlestep(void)
1030 {
1031     return kvm_state->robust_singlestep;
1032 }
1033
1034 int kvm_has_debugregs(void)
1035 {
1036     return kvm_state->debugregs;
1037 }
1038
1039 int kvm_has_xsave(void)
1040 {
1041     return kvm_state->xsave;
1042 }
1043
1044 int kvm_has_xcrs(void)
1045 {
1046     return kvm_state->xcrs;
1047 }
1048
1049 void kvm_setup_guest_memory(void *start, size_t size)
1050 {
1051     if (!kvm_has_sync_mmu()) {
1052         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1053
1054         if (ret) {
1055             perror("qemu_madvise");
1056             fprintf(stderr,
1057                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1058             exit(1);
1059         }
1060     }
1061 }
1062
1063 #ifdef KVM_CAP_SET_GUEST_DEBUG
1064 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1065                                                  target_ulong pc)
1066 {
1067     struct kvm_sw_breakpoint *bp;
1068
1069     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1070         if (bp->pc == pc)
1071             return bp;
1072     }
1073     return NULL;
1074 }
1075
1076 int kvm_sw_breakpoints_active(CPUState *env)
1077 {
1078     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1079 }
1080
1081 struct kvm_set_guest_debug_data {
1082     struct kvm_guest_debug dbg;
1083     CPUState *env;
1084     int err;
1085 };
1086
1087 static void kvm_invoke_set_guest_debug(void *data)
1088 {
1089     struct kvm_set_guest_debug_data *dbg_data = data;
1090     CPUState *env = dbg_data->env;
1091
1092     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1093 }
1094
1095 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1096 {
1097     struct kvm_set_guest_debug_data data;
1098
1099     data.dbg.control = reinject_trap;
1100
1101     if (env->singlestep_enabled) {
1102         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1103     }
1104     kvm_arch_update_guest_debug(env, &data.dbg);
1105     data.env = env;
1106
1107     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1108     return data.err;
1109 }
1110
1111 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1112                           target_ulong len, int type)
1113 {
1114     struct kvm_sw_breakpoint *bp;
1115     CPUState *env;
1116     int err;
1117
1118     if (type == GDB_BREAKPOINT_SW) {
1119         bp = kvm_find_sw_breakpoint(current_env, addr);
1120         if (bp) {
1121             bp->use_count++;
1122             return 0;
1123         }
1124
1125         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1126         if (!bp)
1127             return -ENOMEM;
1128
1129         bp->pc = addr;
1130         bp->use_count = 1;
1131         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1132         if (err) {
1133             free(bp);
1134             return err;
1135         }
1136
1137         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1138                           bp, entry);
1139     } else {
1140         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1141         if (err)
1142             return err;
1143     }
1144
1145     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1146         err = kvm_update_guest_debug(env, 0);
1147         if (err)
1148             return err;
1149     }
1150     return 0;
1151 }
1152
1153 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1154                           target_ulong len, int type)
1155 {
1156     struct kvm_sw_breakpoint *bp;
1157     CPUState *env;
1158     int err;
1159
1160     if (type == GDB_BREAKPOINT_SW) {
1161         bp = kvm_find_sw_breakpoint(current_env, addr);
1162         if (!bp)
1163             return -ENOENT;
1164
1165         if (bp->use_count > 1) {
1166             bp->use_count--;
1167             return 0;
1168         }
1169
1170         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1171         if (err)
1172             return err;
1173
1174         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1175         qemu_free(bp);
1176     } else {
1177         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1178         if (err)
1179             return err;
1180     }
1181
1182     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1183         err = kvm_update_guest_debug(env, 0);
1184         if (err)
1185             return err;
1186     }
1187     return 0;
1188 }
1189
1190 void kvm_remove_all_breakpoints(CPUState *current_env)
1191 {
1192     struct kvm_sw_breakpoint *bp, *next;
1193     KVMState *s = current_env->kvm_state;
1194     CPUState *env;
1195
1196     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1197         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1198             /* Try harder to find a CPU that currently sees the breakpoint. */
1199             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1200                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1201                     break;
1202             }
1203         }
1204     }
1205     kvm_arch_remove_all_hw_breakpoints();
1206
1207     for (env = first_cpu; env != NULL; env = env->next_cpu)
1208         kvm_update_guest_debug(env, 0);
1209 }
1210
1211 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1212
1213 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1214 {
1215     return -EINVAL;
1216 }
1217
1218 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1219                           target_ulong len, int type)
1220 {
1221     return -EINVAL;
1222 }
1223
1224 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1225                           target_ulong len, int type)
1226 {
1227     return -EINVAL;
1228 }
1229
1230 void kvm_remove_all_breakpoints(CPUState *current_env)
1231 {
1232 }
1233 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1234
1235 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1236 {
1237     struct kvm_signal_mask *sigmask;
1238     int r;
1239
1240     if (!sigset)
1241         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1242
1243     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1244
1245     sigmask->len = 8;
1246     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1247     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1248     free(sigmask);
1249
1250     return r;
1251 }
1252
1253 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1254 {
1255 #ifdef KVM_IOEVENTFD
1256     int ret;
1257     struct kvm_ioeventfd iofd;
1258
1259     iofd.datamatch = val;
1260     iofd.addr = addr;
1261     iofd.len = 4;
1262     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1263     iofd.fd = fd;
1264
1265     if (!kvm_enabled()) {
1266         return -ENOSYS;
1267     }
1268
1269     if (!assign) {
1270         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1271     }
1272
1273     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1274
1275     if (ret < 0) {
1276         return -errno;
1277     }
1278
1279     return 0;
1280 #else
1281     return -ENOSYS;
1282 #endif
1283 }
1284
1285 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1286 {
1287 #ifdef KVM_IOEVENTFD
1288     struct kvm_ioeventfd kick = {
1289         .datamatch = val,
1290         .addr = addr,
1291         .len = 2,
1292         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1293         .fd = fd,
1294     };
1295     int r;
1296     if (!kvm_enabled())
1297         return -ENOSYS;
1298     if (!assign)
1299         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1300     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1301     if (r < 0)
1302         return r;
1303     return 0;
1304 #else
1305     return -ENOSYS;
1306 #endif
1307 }