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