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     if (start_addr & ~TARGET_PAGE_MASK) {
 436         if (flags >= IO_MEM_UNASSIGNED) {
 437             if (!kvm_lookup_overlapping_slot(s, start_addr,
 438                                              start_addr + size)) {
 439                 return;
 440             }
 441             fprintf(stderr, "Unaligned split of a KVM memory slot\n");
 442         } else {
 443             fprintf(stderr, "Only page-aligned memory slots supported\n");
 444         }
 445         abort();
 446     }
 447
 448     /* KVM does not support read-only slots */
 449     phys_offset &= ~IO_MEM_ROM;
 450
 451     while (1) {
 452         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 453         if (!mem) {
 454             break;
 455         }
 456
 457         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 458             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 459             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 460             /* The new slot fits into the existing one and comes with
 461              * identical parameters - nothing to be done. */
 462             return;
 463         }
 464
 465         old = *mem;
 466
 467         /* unregister the overlapping slot */
 468         mem->memory_size = 0;
 469         err = kvm_set_user_memory_region(s, mem);
 470         if (err) {
 471             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 472                     __func__, strerror(-err));
 473             abort();
 474         }
 475
 476         /* Workaround for older KVM versions: we can't join slots, even not by
 477          * unregistering the previous ones and then registering the larger
 478          * slot. We have to maintain the existing fragmentation. Sigh.
 479          *
 480          * This workaround assumes that the new slot starts at the same
 481          * address as the first existing one. If not or if some overlapping
 482          * slot comes around later, we will fail (not seen in practice so far)
 483          * - and actually require a recent KVM version. */
 484         if (s->broken_set_mem_region &&
 485             old.start_addr == start_addr && old.memory_size < size &&
 486             flags < IO_MEM_UNASSIGNED) {
 487             mem = kvm_alloc_slot(s);
 488             mem->memory_size = old.memory_size;
 489             mem->start_addr = old.start_addr;
 490             mem->phys_offset = old.phys_offset;
 491             mem->flags = 0;
 492
 493             err = kvm_set_user_memory_region(s, mem);
 494             if (err) {
 495                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 496                         strerror(-err));
 497                 abort();
 498             }
 499
 500             start_addr += old.memory_size;
 501             phys_offset += old.memory_size;
 502             size -= old.memory_size;
 503             continue;
 504         }
 505
 506         /* register prefix slot */
 507         if (old.start_addr < start_addr) {
 508             mem = kvm_alloc_slot(s);
 509             mem->memory_size = start_addr - old.start_addr;
 510             mem->start_addr = old.start_addr;
 511             mem->phys_offset = old.phys_offset;
 512             mem->flags = 0;
 513
 514             err = kvm_set_user_memory_region(s, mem);
 515             if (err) {
 516                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 517                         __func__, strerror(-err));
 518                 abort();
 519             }
 520         }
 521
 522         /* register suffix slot */
 523         if (old.start_addr + old.memory_size > start_addr + size) {
 524             ram_addr_t size_delta;
 525
 526             mem = kvm_alloc_slot(s);
 527             mem->start_addr = start_addr + size;
 528             size_delta = mem->start_addr - old.start_addr;
 529             mem->memory_size = old.memory_size - size_delta;
 530             mem->phys_offset = old.phys_offset + size_delta;
 531             mem->flags = 0;
 532
 533             err = kvm_set_user_memory_region(s, mem);
 534             if (err) {
 535                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 536                         __func__, strerror(-err));
 537                 abort();
 538             }
 539         }
 540     }
 541
 542     /* in case the KVM bug workaround already "consumed" the new slot */
 543     if (!size)
 544         return;
 545
 546     /* KVM does not need to know about this memory */
 547     if (flags >= IO_MEM_UNASSIGNED)
 548         return;
 549
 550     mem = kvm_alloc_slot(s);
 551     mem->memory_size = size;
 552     mem->start_addr = start_addr;
 553     mem->phys_offset = phys_offset;
 554     mem->flags = 0;
 555
 556     err = kvm_set_user_memory_region(s, mem);
 557     if (err) {
 558         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 559                 strerror(-err));
 560         abort();
 561     }
 562 }
 563
 564 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 565                                   target_phys_addr_t start_addr,
 566                                   ram_addr_t size,
 567                                   ram_addr_t phys_offset)
 568 {
 569         kvm_set_phys_mem(start_addr, size, phys_offset);
 570 }
 571
 572 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 573                                         target_phys_addr_t start_addr,
 574                                         target_phys_addr_t end_addr)
 575 {
 576         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 577 }
 578
 579 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 580                                     int enable)
 581 {
 582         return kvm_set_migration_log(enable);
 583 }
 584
 585 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 586         .set_memory = kvm_client_set_memory,
 587         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 588         .migration_log = kvm_client_migration_log,
 589 };
 590
 591 int kvm_init(int smp_cpus)
 592 {
 593     static const char upgrade_note[] =
 594         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 595         "(see http://sourceforge.net/projects/kvm).\n";
 596     KVMState *s;
 597     int ret;
 598     int i;
 599
 600     s = qemu_mallocz(sizeof(KVMState));
 601
 602 #ifdef KVM_CAP_SET_GUEST_DEBUG
 603     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 604 #endif
 605     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 606         s->slots[i].slot = i;
 607
 608     s->vmfd = -1;
 609     s->fd = qemu_open("/dev/kvm", O_RDWR);
 610     if (s->fd == -1) {
 611         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 612         ret = -errno;
 613         goto err;
 614     }
 615
 616     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 617     if (ret < KVM_API_VERSION) {
 618         if (ret > 0)
 619             ret = -EINVAL;
 620         fprintf(stderr, "kvm version too old\n");
 621         goto err;
 622     }
 623
 624     if (ret > KVM_API_VERSION) {
 625         ret = -EINVAL;
 626         fprintf(stderr, "kvm version not supported\n");
 627         goto err;
 628     }
 629
 630     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 631     if (s->vmfd < 0) {
 632 #ifdef TARGET_S390X
 633         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 634                         "your host kernel command line\n");
 635 #endif
 636         goto err;
 637     }
 638
 639     /* initially, KVM allocated its own memory and we had to jump through
 640      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 641      * just use a user allocated buffer so we can use regular pages
 642      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 643      */
 644     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 645         ret = -EINVAL;
 646         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 647                 upgrade_note);
 648         goto err;
 649     }
 650
 651     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 652      * destroyed properly.  Since we rely on this capability, refuse to work
 653      * with any kernel without this capability. */
 654     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 655         ret = -EINVAL;
 656
 657         fprintf(stderr,
 658                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 659                 upgrade_note);
 660         goto err;
 661     }
 662
 663     s->coalesced_mmio = 0;
 664 #ifdef KVM_CAP_COALESCED_MMIO
 665     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 666     s->coalesced_mmio_ring = NULL;
 667 #endif
 668
 669     s->broken_set_mem_region = 1;
 670 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 671     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 672     if (ret > 0) {
 673         s->broken_set_mem_region = 0;
 674     }
 675 #endif
 676
 677     s->vcpu_events = 0;
 678 #ifdef KVM_CAP_VCPU_EVENTS
 679     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 680 #endif
 681
 682     s->robust_singlestep = 0;
 683 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 684     s->robust_singlestep =
 685         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 686 #endif
 687
 688     s->debugregs = 0;
 689 #ifdef KVM_CAP_DEBUGREGS
 690     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 691 #endif
 692
 693     s->xsave = 0;
 694 #ifdef KVM_CAP_XSAVE
 695     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 696 #endif
 697
 698     s->xcrs = 0;
 699 #ifdef KVM_CAP_XCRS
 700     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 701 #endif
 702
 703     ret = kvm_arch_init(s, smp_cpus);
 704     if (ret < 0)
 705         goto err;
 706
 707     kvm_state = s;
 708     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 709
 710     return 0;
 711
 712 err:
 713     if (s) {
 714         if (s->vmfd != -1)
 715             close(s->vmfd);
 716         if (s->fd != -1)
 717             close(s->fd);
 718     }
 719     qemu_free(s);
 720
 721     return ret;
 722 }
 723
 724 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 725                          uint32_t count)
 726 {
 727     int i;
 728     uint8_t *ptr = data;
 729
 730     for (i = 0; i < count; i++) {
 731         if (direction == KVM_EXIT_IO_IN) {
 732             switch (size) {
 733             case 1:
 734                 stb_p(ptr, cpu_inb(port));
 735                 break;
 736             case 2:
 737                 stw_p(ptr, cpu_inw(port));
 738                 break;
 739             case 4:
 740                 stl_p(ptr, cpu_inl(port));
 741                 break;
 742             }
 743         } else {
 744             switch (size) {
 745             case 1:
 746                 cpu_outb(port, ldub_p(ptr));
 747                 break;
 748             case 2:
 749                 cpu_outw(port, lduw_p(ptr));
 750                 break;
 751             case 4:
 752                 cpu_outl(port, ldl_p(ptr));
 753                 break;
 754             }
 755         }
 756
 757         ptr += size;
 758     }
 759
 760     return 1;
 761 }
 762
 763 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 764 static void kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 765 {
 766
 767     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 768         int i;
 769
 770         fprintf(stderr, "KVM internal error. Suberror: %d\n",
 771                 run->internal.suberror);
 772
 773         for (i = 0; i < run->internal.ndata; ++i) {
 774             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 775                     i, (uint64_t)run->internal.data[i]);
 776         }
 777     }
 778     cpu_dump_state(env, stderr, fprintf, 0);
 779     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 780         fprintf(stderr, "emulation failure\n");
 781         if (!kvm_arch_stop_on_emulation_error(env))
 782                 return;
 783     }
 784     /* FIXME: Should trigger a qmp message to let management know
 785      * something went wrong.
 786      */
 787     vm_stop(0);
 788 }
 789 #endif
 790
 791 void kvm_flush_coalesced_mmio_buffer(void)
 792 {
 793 #ifdef KVM_CAP_COALESCED_MMIO
 794     KVMState *s = kvm_state;
 795     if (s->coalesced_mmio_ring) {
 796         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 797         while (ring->first != ring->last) {
 798             struct kvm_coalesced_mmio *ent;
 799
 800             ent = &ring->coalesced_mmio[ring->first];
 801
 802             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 803             smp_wmb();
 804             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 805         }
 806     }
 807 #endif
 808 }
 809
 810 static void do_kvm_cpu_synchronize_state(void *_env)
 811 {
 812     CPUState *env = _env;
 813
 814     if (!env->kvm_vcpu_dirty) {
 815         kvm_arch_get_registers(env);
 816         env->kvm_vcpu_dirty = 1;
 817     }
 818 }
 819
 820 void kvm_cpu_synchronize_state(CPUState *env)
 821 {
 822     if (!env->kvm_vcpu_dirty)
 823         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 824 }
 825
 826 void kvm_cpu_synchronize_post_reset(CPUState *env)
 827 {
 828     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 829     env->kvm_vcpu_dirty = 0;
 830 }
 831
 832 void kvm_cpu_synchronize_post_init(CPUState *env)
 833 {
 834     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 835     env->kvm_vcpu_dirty = 0;
 836 }
 837
 838 int kvm_cpu_exec(CPUState *env)
 839 {
 840     struct kvm_run *run = env->kvm_run;
 841     int ret;
 842
 843     DPRINTF("kvm_cpu_exec()\n");
 844
 845     do {
 846 #ifndef CONFIG_IOTHREAD
 847         if (env->exit_request) {
 848             DPRINTF("interrupt exit requested\n");
 849             ret = 0;
 850             break;
 851         }
 852 #endif
 853
 854         if (kvm_arch_process_irqchip_events(env)) {
 855             ret = 0;
 856             break;
 857         }
 858
 859         if (env->kvm_vcpu_dirty) {
 860             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 861             env->kvm_vcpu_dirty = 0;
 862         }
 863
 864         kvm_arch_pre_run(env, run);
 865         cpu_single_env = NULL;
 866         qemu_mutex_unlock_iothread();
 867         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 868         qemu_mutex_lock_iothread();
 869         cpu_single_env = env;
 870         kvm_arch_post_run(env, run);
 871
 872         if (ret == -EINTR || ret == -EAGAIN) {
 873             cpu_exit(env);
 874             DPRINTF("io window exit\n");
 875             ret = 0;
 876             break;
 877         }
 878
 879         if (ret < 0) {
 880             DPRINTF("kvm run failed %s\n", strerror(-ret));
 881             abort();
 882         }
 883
 884         kvm_flush_coalesced_mmio_buffer();
 885
 886         ret = 0; /* exit loop */
 887         switch (run->exit_reason) {
 888         case KVM_EXIT_IO:
 889             DPRINTF("handle_io\n");
 890             ret = kvm_handle_io(run->io.port,
 891                                 (uint8_t *)run + run->io.data_offset,
 892                                 run->io.direction,
 893                                 run->io.size,
 894                                 run->io.count);
 895             break;
 896         case KVM_EXIT_MMIO:
 897             DPRINTF("handle_mmio\n");
 898             cpu_physical_memory_rw(run->mmio.phys_addr,
 899                                    run->mmio.data,
 900                                    run->mmio.len,
 901                                    run->mmio.is_write);
 902             ret = 1;
 903             break;
 904         case KVM_EXIT_IRQ_WINDOW_OPEN:
 905             DPRINTF("irq_window_open\n");
 906             break;
 907         case KVM_EXIT_SHUTDOWN:
 908             DPRINTF("shutdown\n");
 909             qemu_system_reset_request();
 910             ret = 1;
 911             break;
 912         case KVM_EXIT_UNKNOWN:
 913             DPRINTF("kvm_exit_unknown\n");
 914             break;
 915         case KVM_EXIT_FAIL_ENTRY:
 916             DPRINTF("kvm_exit_fail_entry\n");
 917             break;
 918         case KVM_EXIT_EXCEPTION:
 919             DPRINTF("kvm_exit_exception\n");
 920             break;
 921 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 922         case KVM_EXIT_INTERNAL_ERROR:
 923             kvm_handle_internal_error(env, run);
 924             break;
 925 #endif
 926         case KVM_EXIT_DEBUG:
 927             DPRINTF("kvm_exit_debug\n");
 928 #ifdef KVM_CAP_SET_GUEST_DEBUG
 929             if (kvm_arch_debug(&run->debug.arch)) {
 930                 env->exception_index = EXCP_DEBUG;
 931                 return 0;
 932             }
 933             /* re-enter, this exception was guest-internal */
 934             ret = 1;
 935 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 936             break;
 937         default:
 938             DPRINTF("kvm_arch_handle_exit\n");
 939             ret = kvm_arch_handle_exit(env, run);
 940             break;
 941         }
 942     } while (ret > 0);
 943
 944     if (env->exit_request) {
 945         env->exit_request = 0;
 946         env->exception_index = EXCP_INTERRUPT;
 947     }
 948
 949     return ret;
 950 }
 951
 952 int kvm_ioctl(KVMState *s, int type, ...)
 953 {
 954     int ret;
 955     void *arg;
 956     va_list ap;
 957
 958     va_start(ap, type);
 959     arg = va_arg(ap, void *);
 960     va_end(ap);
 961
 962     ret = ioctl(s->fd, type, arg);
 963     if (ret == -1)
 964         ret = -errno;
 965
 966     return ret;
 967 }
 968
 969 int kvm_vm_ioctl(KVMState *s, int type, ...)
 970 {
 971     int ret;
 972     void *arg;
 973     va_list ap;
 974
 975     va_start(ap, type);
 976     arg = va_arg(ap, void *);
 977     va_end(ap);
 978
 979     ret = ioctl(s->vmfd, type, arg);
 980     if (ret == -1)
 981         ret = -errno;
 982
 983     return ret;
 984 }
 985
 986 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 987 {
 988     int ret;
 989     void *arg;
 990     va_list ap;
 991
 992     va_start(ap, type);
 993     arg = va_arg(ap, void *);
 994     va_end(ap);
 995
 996     ret = ioctl(env->kvm_fd, type, arg);
 997     if (ret == -1)
 998         ret = -errno;
 999
1000     return ret;
1001 }
1002
1003 int kvm_has_sync_mmu(void)
1004 {
1005 #ifdef KVM_CAP_SYNC_MMU
1006     KVMState *s = kvm_state;
1007
1008     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
1009 #else
1010     return 0;
1011 #endif
1012 }
1013
1014 int kvm_has_vcpu_events(void)
1015 {
1016     return kvm_state->vcpu_events;
1017 }
1018
1019 int kvm_has_robust_singlestep(void)
1020 {
1021     return kvm_state->robust_singlestep;
1022 }
1023
1024 int kvm_has_debugregs(void)
1025 {
1026     return kvm_state->debugregs;
1027 }
1028
1029 int kvm_has_xsave(void)
1030 {
1031     return kvm_state->xsave;
1032 }
1033
1034 int kvm_has_xcrs(void)
1035 {
1036     return kvm_state->xcrs;
1037 }
1038
1039 void kvm_setup_guest_memory(void *start, size_t size)
1040 {
1041     if (!kvm_has_sync_mmu()) {
1042 #ifdef MADV_DONTFORK
1043         int ret = madvise(start, size, MADV_DONTFORK);
1044
1045         if (ret) {
1046             perror("madvice");
1047             exit(1);
1048         }
1049 #else
1050         fprintf(stderr,
1051                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1052         exit(1);
1053 #endif
1054     }
1055 }
1056
1057 #ifdef KVM_CAP_SET_GUEST_DEBUG
1058 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1059                                                  target_ulong pc)
1060 {
1061     struct kvm_sw_breakpoint *bp;
1062
1063     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1064         if (bp->pc == pc)
1065             return bp;
1066     }
1067     return NULL;
1068 }
1069
1070 int kvm_sw_breakpoints_active(CPUState *env)
1071 {
1072     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1073 }
1074
1075 struct kvm_set_guest_debug_data {
1076     struct kvm_guest_debug dbg;
1077     CPUState *env;
1078     int err;
1079 };
1080
1081 static void kvm_invoke_set_guest_debug(void *data)
1082 {
1083     struct kvm_set_guest_debug_data *dbg_data = data;
1084     CPUState *env = dbg_data->env;
1085
1086     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1087 }
1088
1089 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1090 {
1091     struct kvm_set_guest_debug_data data;
1092
1093     data.dbg.control = reinject_trap;
1094
1095     if (env->singlestep_enabled) {
1096         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1097     }
1098     kvm_arch_update_guest_debug(env, &data.dbg);
1099     data.env = env;
1100
1101     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1102     return data.err;
1103 }
1104
1105 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1106                           target_ulong len, int type)
1107 {
1108     struct kvm_sw_breakpoint *bp;
1109     CPUState *env;
1110     int err;
1111
1112     if (type == GDB_BREAKPOINT_SW) {
1113         bp = kvm_find_sw_breakpoint(current_env, addr);
1114         if (bp) {
1115             bp->use_count++;
1116             return 0;
1117         }
1118
1119         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1120         if (!bp)
1121             return -ENOMEM;
1122
1123         bp->pc = addr;
1124         bp->use_count = 1;
1125         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1126         if (err) {
1127             free(bp);
1128             return err;
1129         }
1130
1131         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1132                           bp, entry);
1133     } else {
1134         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1135         if (err)
1136             return err;
1137     }
1138
1139     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1140         err = kvm_update_guest_debug(env, 0);
1141         if (err)
1142             return err;
1143     }
1144     return 0;
1145 }
1146
1147 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1148                           target_ulong len, int type)
1149 {
1150     struct kvm_sw_breakpoint *bp;
1151     CPUState *env;
1152     int err;
1153
1154     if (type == GDB_BREAKPOINT_SW) {
1155         bp = kvm_find_sw_breakpoint(current_env, addr);
1156         if (!bp)
1157             return -ENOENT;
1158
1159         if (bp->use_count > 1) {
1160             bp->use_count--;
1161             return 0;
1162         }
1163
1164         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1165         if (err)
1166             return err;
1167
1168         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1169         qemu_free(bp);
1170     } else {
1171         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1172         if (err)
1173             return err;
1174     }
1175
1176     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1177         err = kvm_update_guest_debug(env, 0);
1178         if (err)
1179             return err;
1180     }
1181     return 0;
1182 }
1183
1184 void kvm_remove_all_breakpoints(CPUState *current_env)
1185 {
1186     struct kvm_sw_breakpoint *bp, *next;
1187     KVMState *s = current_env->kvm_state;
1188     CPUState *env;
1189
1190     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1191         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1192             /* Try harder to find a CPU that currently sees the breakpoint. */
1193             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1194                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1195                     break;
1196             }
1197         }
1198     }
1199     kvm_arch_remove_all_hw_breakpoints();
1200
1201     for (env = first_cpu; env != NULL; env = env->next_cpu)
1202         kvm_update_guest_debug(env, 0);
1203 }
1204
1205 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1206
1207 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1208 {
1209     return -EINVAL;
1210 }
1211
1212 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1213                           target_ulong len, int type)
1214 {
1215     return -EINVAL;
1216 }
1217
1218 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1219                           target_ulong len, int type)
1220 {
1221     return -EINVAL;
1222 }
1223
1224 void kvm_remove_all_breakpoints(CPUState *current_env)
1225 {
1226 }
1227 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1228
1229 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1230 {
1231     struct kvm_signal_mask *sigmask;
1232     int r;
1233
1234     if (!sigset)
1235         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1236
1237     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1238
1239     sigmask->len = 8;
1240     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1241     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1242     free(sigmask);
1243
1244     return r;
1245 }
1246
1247 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1248 {
1249 #ifdef KVM_IOEVENTFD
1250     struct kvm_ioeventfd kick = {
1251         .datamatch = val,
1252         .addr = addr,
1253         .len = 2,
1254         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1255         .fd = fd,
1256     };
1257     int r;
1258     if (!kvm_enabled())
1259         return -ENOSYS;
1260     if (!assign)
1261         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1262     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1263     if (r < 0)
1264         return r;
1265     return 0;
1266 #else
1267     return -ENOSYS;
1268 #endif
1269 }