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