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 "sysemu.h"
  25 #include "kvm.h"
  26
  27 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  28 #define PAGE_SIZE TARGET_PAGE_SIZE
  29
  30 //#define DEBUG_KVM
  31
  32 #ifdef DEBUG_KVM
  33 #define dprintf(fmt, ...) \
  34     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  35 #else
  36 #define dprintf(fmt, ...) \
  37     do { } while (0)
  38 #endif
  39
  40 typedef struct KVMSlot
  41 {
  42     target_phys_addr_t start_addr;
  43     ram_addr_t memory_size;
  44     ram_addr_t phys_offset;
  45     int slot;
  46     int flags;
  47 } KVMSlot;
  48
  49 typedef struct kvm_dirty_log KVMDirtyLog;
  50
  51 int kvm_allowed = 0;
  52
  53 struct KVMState
  54 {
  55     KVMSlot slots[32];
  56     int fd;
  57     int vmfd;
  58     int coalesced_mmio;
  59 };
  60
  61 static KVMState *kvm_state;
  62
  63 static KVMSlot *kvm_alloc_slot(KVMState *s)
  64 {
  65     int i;
  66
  67     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  68         /* KVM private memory slots */
  69         if (i >= 8 && i < 12)
  70             continue;
  71         if (s->slots[i].memory_size == 0)
  72             return &s->slots[i];
  73     }
  74
  75     return NULL;
  76 }
  77
  78 static KVMSlot *kvm_lookup_slot(KVMState *s, target_phys_addr_t start_addr)
  79 {
  80     int i;
  81
  82     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  83         KVMSlot *mem = &s->slots[i];
  84
  85         if (start_addr >= mem->start_addr &&
  86             start_addr < (mem->start_addr + mem->memory_size))
  87             return mem;
  88     }
  89
  90     return NULL;
  91 }
  92
  93 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
  94 {
  95     struct kvm_userspace_memory_region mem;
  96
  97     mem.slot = slot->slot;
  98     mem.guest_phys_addr = slot->start_addr;
  99     mem.memory_size = slot->memory_size;
 100     mem.userspace_addr = (unsigned long)phys_ram_base + slot->phys_offset;
 101     mem.flags = slot->flags;
 102
 103     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 104 }
 105
 106
 107 int kvm_init_vcpu(CPUState *env)
 108 {
 109     KVMState *s = kvm_state;
 110     long mmap_size;
 111     int ret;
 112
 113     dprintf("kvm_init_vcpu\n");
 114
 115     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 116     if (ret < 0) {
 117         dprintf("kvm_create_vcpu failed\n");
 118         goto err;
 119     }
 120
 121     env->kvm_fd = ret;
 122     env->kvm_state = s;
 123
 124     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 125     if (mmap_size < 0) {
 126         dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
 127         goto err;
 128     }
 129
 130     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 131                         env->kvm_fd, 0);
 132     if (env->kvm_run == MAP_FAILED) {
 133         ret = -errno;
 134         dprintf("mmap'ing vcpu state failed\n");
 135         goto err;
 136     }
 137
 138     ret = kvm_arch_init_vcpu(env);
 139
 140 err:
 141     return ret;
 142 }
 143
 144 int kvm_sync_vcpus(void)
 145 {
 146     CPUState *env;
 147
 148     for (env = first_cpu; env != NULL; env = env->next_cpu) {
 149         int ret;
 150
 151         ret = kvm_arch_put_registers(env);
 152         if (ret)
 153             return ret;
 154     }
 155
 156     return 0;
 157 }
 158
 159 /*
 160  * dirty pages logging control
 161  */
 162 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr, target_phys_addr_t end_addr,
 163                                       unsigned flags,
 164                                       unsigned mask)
 165 {
 166     KVMState *s = kvm_state;
 167     KVMSlot *mem = kvm_lookup_slot(s, phys_addr);
 168     if (mem == NULL)  {
 169             dprintf("invalid parameters %llx-%llx\n", phys_addr, end_addr);
 170             return -EINVAL;
 171     }
 172
 173     flags = (mem->flags & ~mask) | flags;
 174     /* Nothing changed, no need to issue ioctl */
 175     if (flags == mem->flags)
 176             return 0;
 177
 178     mem->flags = flags;
 179
 180     return kvm_set_user_memory_region(s, mem);
 181 }
 182
 183 int kvm_log_start(target_phys_addr_t phys_addr, target_phys_addr_t end_addr)
 184 {
 185         return kvm_dirty_pages_log_change(phys_addr, end_addr,
 186                                           KVM_MEM_LOG_DIRTY_PAGES,
 187                                           KVM_MEM_LOG_DIRTY_PAGES);
 188 }
 189
 190 int kvm_log_stop(target_phys_addr_t phys_addr, target_phys_addr_t end_addr)
 191 {
 192         return kvm_dirty_pages_log_change(phys_addr, end_addr,
 193                                           0,
 194                                           KVM_MEM_LOG_DIRTY_PAGES);
 195 }
 196
 197 /**
 198  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 199  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 200  * This means all bits are set to dirty.
 201  *
 202  * @start_add: start of logged region. This is what we use to search the memslot
 203  * @end_addr: end of logged region.
 204  */
 205 void kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr, target_phys_addr_t end_addr)
 206 {
 207     KVMState *s = kvm_state;
 208     KVMDirtyLog d;
 209     KVMSlot *mem = kvm_lookup_slot(s, start_addr);
 210     unsigned long alloc_size;
 211     ram_addr_t addr;
 212     target_phys_addr_t phys_addr = start_addr;
 213
 214     dprintf("sync addr: %llx into %lx\n", start_addr, mem->phys_offset);
 215     if (mem == NULL) {
 216             fprintf(stderr, "BUG: %s: invalid parameters\n", __func__);
 217             return;
 218     }
 219
 220     alloc_size = mem->memory_size >> TARGET_PAGE_BITS / sizeof(d.dirty_bitmap);
 221     d.dirty_bitmap = qemu_mallocz(alloc_size);
 222
 223     d.slot = mem->slot;
 224     dprintf("slot %d, phys_addr %llx, uaddr: %llx\n",
 225             d.slot, mem->start_addr, mem->phys_offset);
 226
 227     if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 228         dprintf("ioctl failed %d\n", errno);
 229         goto out;
 230     }
 231
 232     phys_addr = start_addr;
 233     for (addr = mem->phys_offset; phys_addr < end_addr; phys_addr+= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
 234         unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
 235         unsigned nr = (phys_addr - start_addr) >> TARGET_PAGE_BITS;
 236         unsigned word = nr / (sizeof(*bitmap) * 8);
 237         unsigned bit = nr % (sizeof(*bitmap) * 8);
 238         if ((bitmap[word] >> bit) & 1)
 239             cpu_physical_memory_set_dirty(addr);
 240     }
 241 out:
 242     qemu_free(d.dirty_bitmap);
 243 }
 244
 245 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 246 {
 247     int ret = -ENOSYS;
 248 #ifdef KVM_CAP_COALESCED_MMIO
 249     KVMState *s = kvm_state;
 250
 251     if (s->coalesced_mmio) {
 252         struct kvm_coalesced_mmio_zone zone;
 253
 254         zone.addr = start;
 255         zone.size = size;
 256
 257         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 258     }
 259 #endif
 260
 261     return ret;
 262 }
 263
 264 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 265 {
 266     int ret = -ENOSYS;
 267 #ifdef KVM_CAP_COALESCED_MMIO
 268     KVMState *s = kvm_state;
 269
 270     if (s->coalesced_mmio) {
 271         struct kvm_coalesced_mmio_zone zone;
 272
 273         zone.addr = start;
 274         zone.size = size;
 275
 276         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 277     }
 278 #endif
 279
 280     return ret;
 281 }
 282
 283 int kvm_check_extension(KVMState *s, unsigned int extension)
 284 {
 285     int ret;
 286
 287     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 288     if (ret < 0) {
 289         ret = 0;
 290     }
 291
 292     return ret;
 293 }
 294
 295 int kvm_init(int smp_cpus)
 296 {
 297     KVMState *s;
 298     int ret;
 299     int i;
 300
 301     if (smp_cpus > 1) {
 302         fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
 303         return -EINVAL;
 304     }
 305
 306     s = qemu_mallocz(sizeof(KVMState));
 307
 308     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 309         s->slots[i].slot = i;
 310
 311     s->vmfd = -1;
 312     s->fd = open("/dev/kvm", O_RDWR);
 313     if (s->fd == -1) {
 314         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 315         ret = -errno;
 316         goto err;
 317     }
 318
 319     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 320     if (ret < KVM_API_VERSION) {
 321         if (ret > 0)
 322             ret = -EINVAL;
 323         fprintf(stderr, "kvm version too old\n");
 324         goto err;
 325     }
 326
 327     if (ret > KVM_API_VERSION) {
 328         ret = -EINVAL;
 329         fprintf(stderr, "kvm version not supported\n");
 330         goto err;
 331     }
 332
 333     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 334     if (s->vmfd < 0)
 335         goto err;
 336
 337     /* initially, KVM allocated its own memory and we had to jump through
 338      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 339      * just use a user allocated buffer so we can use phys_ram_base
 340      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 341      */
 342     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 343         ret = -EINVAL;
 344         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
 345         goto err;
 346     }
 347
 348     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 349      * destroyed properly.  Since we rely on this capability, refuse to work
 350      * with any kernel without this capability. */
 351     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 352         ret = -EINVAL;
 353
 354         fprintf(stderr,
 355                 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
 356                 "Please upgrade to at least kvm-81.\n");
 357         goto err;
 358     }
 359
 360 #ifdef KVM_CAP_COALESCED_MMIO
 361     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 362 #else
 363     s->coalesced_mmio = 0;
 364 #endif
 365
 366     ret = kvm_arch_init(s, smp_cpus);
 367     if (ret < 0)
 368         goto err;
 369
 370     kvm_state = s;
 371
 372     return 0;
 373
 374 err:
 375     if (s) {
 376         if (s->vmfd != -1)
 377             close(s->vmfd);
 378         if (s->fd != -1)
 379             close(s->fd);
 380     }
 381     qemu_free(s);
 382
 383     return ret;
 384 }
 385
 386 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
 387                          int direction, int size, uint32_t count)
 388 {
 389     int i;
 390     uint8_t *ptr = data;
 391
 392     for (i = 0; i < count; i++) {
 393         if (direction == KVM_EXIT_IO_IN) {
 394             switch (size) {
 395             case 1:
 396                 stb_p(ptr, cpu_inb(env, port));
 397                 break;
 398             case 2:
 399                 stw_p(ptr, cpu_inw(env, port));
 400                 break;
 401             case 4:
 402                 stl_p(ptr, cpu_inl(env, port));
 403                 break;
 404             }
 405         } else {
 406             switch (size) {
 407             case 1:
 408                 cpu_outb(env, port, ldub_p(ptr));
 409                 break;
 410             case 2:
 411                 cpu_outw(env, port, lduw_p(ptr));
 412                 break;
 413             case 4:
 414                 cpu_outl(env, port, ldl_p(ptr));
 415                 break;
 416             }
 417         }
 418
 419         ptr += size;
 420     }
 421
 422     return 1;
 423 }
 424
 425 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
 426 {
 427 #ifdef KVM_CAP_COALESCED_MMIO
 428     KVMState *s = kvm_state;
 429     if (s->coalesced_mmio) {
 430         struct kvm_coalesced_mmio_ring *ring;
 431
 432         ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
 433         while (ring->first != ring->last) {
 434             struct kvm_coalesced_mmio *ent;
 435
 436             ent = &ring->coalesced_mmio[ring->first];
 437
 438             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 439             /* FIXME smp_wmb() */
 440             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 441         }
 442     }
 443 #endif
 444 }
 445
 446 int kvm_cpu_exec(CPUState *env)
 447 {
 448     struct kvm_run *run = env->kvm_run;
 449     int ret;
 450
 451     dprintf("kvm_cpu_exec()\n");
 452
 453     do {
 454         kvm_arch_pre_run(env, run);
 455
 456         if (env->exit_request) {
 457             dprintf("interrupt exit requested\n");
 458             ret = 0;
 459             break;
 460         }
 461
 462         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 463         kvm_arch_post_run(env, run);
 464
 465         if (ret == -EINTR || ret == -EAGAIN) {
 466             dprintf("io window exit\n");
 467             ret = 0;
 468             break;
 469         }
 470
 471         if (ret < 0) {
 472             dprintf("kvm run failed %s\n", strerror(-ret));
 473             abort();
 474         }
 475
 476         kvm_run_coalesced_mmio(env, run);
 477
 478         ret = 0; /* exit loop */
 479         switch (run->exit_reason) {
 480         case KVM_EXIT_IO:
 481             dprintf("handle_io\n");
 482             ret = kvm_handle_io(env, run->io.port,
 483                                 (uint8_t *)run + run->io.data_offset,
 484                                 run->io.direction,
 485                                 run->io.size,
 486                                 run->io.count);
 487             break;
 488         case KVM_EXIT_MMIO:
 489             dprintf("handle_mmio\n");
 490             cpu_physical_memory_rw(run->mmio.phys_addr,
 491                                    run->mmio.data,
 492                                    run->mmio.len,
 493                                    run->mmio.is_write);
 494             ret = 1;
 495             break;
 496         case KVM_EXIT_IRQ_WINDOW_OPEN:
 497             dprintf("irq_window_open\n");
 498             break;
 499         case KVM_EXIT_SHUTDOWN:
 500             dprintf("shutdown\n");
 501             qemu_system_reset_request();
 502             ret = 1;
 503             break;
 504         case KVM_EXIT_UNKNOWN:
 505             dprintf("kvm_exit_unknown\n");
 506             break;
 507         case KVM_EXIT_FAIL_ENTRY:
 508             dprintf("kvm_exit_fail_entry\n");
 509             break;
 510         case KVM_EXIT_EXCEPTION:
 511             dprintf("kvm_exit_exception\n");
 512             break;
 513         case KVM_EXIT_DEBUG:
 514             dprintf("kvm_exit_debug\n");
 515             break;
 516         default:
 517             dprintf("kvm_arch_handle_exit\n");
 518             ret = kvm_arch_handle_exit(env, run);
 519             break;
 520         }
 521     } while (ret > 0);
 522
 523     if (env->exit_request) {
 524         env->exit_request = 0;
 525         env->exception_index = EXCP_INTERRUPT;
 526     }
 527
 528     return ret;
 529 }
 530
 531 void kvm_set_phys_mem(target_phys_addr_t start_addr,
 532                       ram_addr_t size,
 533                       ram_addr_t phys_offset)
 534 {
 535     KVMState *s = kvm_state;
 536     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 537     KVMSlot *mem;
 538
 539     /* KVM does not support read-only slots */
 540     phys_offset &= ~IO_MEM_ROM;
 541
 542     mem = kvm_lookup_slot(s, start_addr);
 543     if (mem) {
 544         if ((flags == IO_MEM_UNASSIGNED) || (flags >= TLB_MMIO)) {
 545             mem->memory_size = 0;
 546             mem->start_addr = start_addr;
 547             mem->phys_offset = 0;
 548             mem->flags = 0;
 549
 550             kvm_set_user_memory_region(s, mem);
 551         } else if (start_addr >= mem->start_addr &&
 552                    (start_addr + size) <= (mem->start_addr +
 553                                            mem->memory_size)) {
 554             KVMSlot slot;
 555             target_phys_addr_t mem_start;
 556             ram_addr_t mem_size, mem_offset;
 557
 558             /* Not splitting */
 559             if ((phys_offset - (start_addr - mem->start_addr)) ==
 560                 mem->phys_offset)
 561                 return;
 562
 563             /* unregister whole slot */
 564             memcpy(&slot, mem, sizeof(slot));
 565             mem->memory_size = 0;
 566             kvm_set_user_memory_region(s, mem);
 567
 568             /* register prefix slot */
 569             mem_start = slot.start_addr;
 570             mem_size = start_addr - slot.start_addr;
 571             mem_offset = slot.phys_offset;
 572             if (mem_size)
 573                 kvm_set_phys_mem(mem_start, mem_size, mem_offset);
 574
 575             /* register new slot */
 576             kvm_set_phys_mem(start_addr, size, phys_offset);
 577
 578             /* register suffix slot */
 579             mem_start = start_addr + size;
 580             mem_offset += mem_size + size;
 581             mem_size = slot.memory_size - mem_size - size;
 582             if (mem_size)
 583                 kvm_set_phys_mem(mem_start, mem_size, mem_offset);
 584
 585             return;
 586         } else {
 587             printf("Registering overlapping slot\n");
 588             abort();
 589         }
 590     }
 591     /* KVM does not need to know about this memory */
 592     if (flags >= IO_MEM_UNASSIGNED)
 593         return;
 594
 595     mem = kvm_alloc_slot(s);
 596     mem->memory_size = size;
 597     mem->start_addr = start_addr;
 598     mem->phys_offset = phys_offset;
 599     mem->flags = 0;
 600
 601     kvm_set_user_memory_region(s, mem);
 602     /* FIXME deal with errors */
 603 }
 604
 605 int kvm_ioctl(KVMState *s, int type, ...)
 606 {
 607     int ret;
 608     void *arg;
 609     va_list ap;
 610
 611     va_start(ap, type);
 612     arg = va_arg(ap, void *);
 613     va_end(ap);
 614
 615     ret = ioctl(s->fd, type, arg);
 616     if (ret == -1)
 617         ret = -errno;
 618
 619     return ret;
 620 }
 621
 622 int kvm_vm_ioctl(KVMState *s, int type, ...)
 623 {
 624     int ret;
 625     void *arg;
 626     va_list ap;
 627
 628     va_start(ap, type);
 629     arg = va_arg(ap, void *);
 630     va_end(ap);
 631
 632     ret = ioctl(s->vmfd, type, arg);
 633     if (ret == -1)
 634         ret = -errno;
 635
 636     return ret;
 637 }
 638
 639 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
 640 {
 641     int ret;
 642     void *arg;
 643     va_list ap;
 644
 645     va_start(ap, type);
 646     arg = va_arg(ap, void *);
 647     va_end(ap);
 648
 649     ret = ioctl(env->kvm_fd, type, arg);
 650     if (ret == -1)
 651         ret = -errno;
 652
 653     return ret;
 654 }
 655
 656 int kvm_has_sync_mmu(void)
 657 {
 658 #ifdef KVM_CAP_SYNC_MMU
 659     KVMState *s = kvm_state;
 660
 661     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
 662 #else
 663     return 0;
 664 #endif
 665 }
 666
 667 void kvm_setup_guest_memory(void *start, size_t size)
 668 {
 669     if (!kvm_has_sync_mmu()) {
 670 #ifdef MADV_DONTFORK
 671         int ret = madvise(start, size, MADV_DONTFORK);
 672
 673         if (ret) {
 674             perror("madvice");
 675             exit(1);
 676         }
 677 #else
 678         fprintf(stderr,
 679                 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
 680         exit(1);
 681 #endif
 682     }
 683 }