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 /* This check must be after config-host.h is included */
  32 #ifdef CONFIG_EVENTFD
  33 #include <sys/eventfd.h>
  34 #endif
  35
  36 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
  37 #define PAGE_SIZE TARGET_PAGE_SIZE
  38
  39 //#define DEBUG_KVM
  40
  41 #ifdef DEBUG_KVM
  42 #define DPRINTF(fmt, ...) \
  43     do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
  44 #else
  45 #define DPRINTF(fmt, ...) \
  46     do { } while (0)
  47 #endif
  48
  49 typedef struct KVMSlot
  50 {
  51     target_phys_addr_t start_addr;
  52     ram_addr_t memory_size;
  53     ram_addr_t phys_offset;
  54     int slot;
  55     int flags;
  56 } KVMSlot;
  57
  58 typedef struct kvm_dirty_log KVMDirtyLog;
  59
  60 struct KVMState
  61 {
  62     KVMSlot slots[32];
  63     int fd;
  64     int vmfd;
  65     int coalesced_mmio;
  66     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  67     int broken_set_mem_region;
  68     int migration_log;
  69     int vcpu_events;
  70     int robust_singlestep;
  71     int debugregs;
  72 #ifdef KVM_CAP_SET_GUEST_DEBUG
  73     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  74 #endif
  75     int irqchip_in_kernel;
  76     int pit_in_kernel;
  77     int xsave, xcrs;
  78     int many_ioeventfds;
  79 };
  80
  81 static KVMState *kvm_state;
  82
  83 static const KVMCapabilityInfo kvm_required_capabilites[] = {
  84     KVM_CAP_INFO(USER_MEMORY),
  85     KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
  86     KVM_CAP_LAST_INFO
  87 };
  88
  89 static KVMSlot *kvm_alloc_slot(KVMState *s)
  90 {
  91     int i;
  92
  93     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  94         /* KVM private memory slots */
  95         if (i >= 8 && i < 12) {
  96             continue;
  97         }
  98         if (s->slots[i].memory_size == 0) {
  99             return &s->slots[i];
 100         }
 101     }
 102
 103     fprintf(stderr, "%s: no free slot available\n", __func__);
 104     abort();
 105 }
 106
 107 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 108                                          target_phys_addr_t start_addr,
 109                                          target_phys_addr_t end_addr)
 110 {
 111     int i;
 112
 113     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 114         KVMSlot *mem = &s->slots[i];
 115
 116         if (start_addr == mem->start_addr &&
 117             end_addr == mem->start_addr + mem->memory_size) {
 118             return mem;
 119         }
 120     }
 121
 122     return NULL;
 123 }
 124
 125 /*
 126  * Find overlapping slot with lowest start address
 127  */
 128 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 129                                             target_phys_addr_t start_addr,
 130                                             target_phys_addr_t end_addr)
 131 {
 132     KVMSlot *found = NULL;
 133     int i;
 134
 135     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 136         KVMSlot *mem = &s->slots[i];
 137
 138         if (mem->memory_size == 0 ||
 139             (found && found->start_addr < mem->start_addr)) {
 140             continue;
 141         }
 142
 143         if (end_addr > mem->start_addr &&
 144             start_addr < mem->start_addr + mem->memory_size) {
 145             found = mem;
 146         }
 147     }
 148
 149     return found;
 150 }
 151
 152 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
 153                                       target_phys_addr_t *phys_addr)
 154 {
 155     int i;
 156
 157     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 158         KVMSlot *mem = &s->slots[i];
 159
 160         if (ram_addr >= mem->phys_offset &&
 161             ram_addr < mem->phys_offset + mem->memory_size) {
 162             *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
 163             return 1;
 164         }
 165     }
 166
 167     return 0;
 168 }
 169
 170 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 171 {
 172     struct kvm_userspace_memory_region mem;
 173
 174     mem.slot = slot->slot;
 175     mem.guest_phys_addr = slot->start_addr;
 176     mem.memory_size = slot->memory_size;
 177     mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
 178     mem.flags = slot->flags;
 179     if (s->migration_log) {
 180         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 181     }
 182     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 183 }
 184
 185 static void kvm_reset_vcpu(void *opaque)
 186 {
 187     CPUState *env = opaque;
 188
 189     kvm_arch_reset_vcpu(env);
 190 }
 191
 192 int kvm_irqchip_in_kernel(void)
 193 {
 194     return kvm_state->irqchip_in_kernel;
 195 }
 196
 197 int kvm_pit_in_kernel(void)
 198 {
 199     return kvm_state->pit_in_kernel;
 200 }
 201
 202
 203 int kvm_init_vcpu(CPUState *env)
 204 {
 205     KVMState *s = kvm_state;
 206     long mmap_size;
 207     int ret;
 208
 209     DPRINTF("kvm_init_vcpu\n");
 210
 211     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 212     if (ret < 0) {
 213         DPRINTF("kvm_create_vcpu failed\n");
 214         goto err;
 215     }
 216
 217     env->kvm_fd = ret;
 218     env->kvm_state = s;
 219
 220     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 221     if (mmap_size < 0) {
 222         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 223         goto err;
 224     }
 225
 226     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 227                         env->kvm_fd, 0);
 228     if (env->kvm_run == MAP_FAILED) {
 229         ret = -errno;
 230         DPRINTF("mmap'ing vcpu state failed\n");
 231         goto err;
 232     }
 233
 234     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 235         s->coalesced_mmio_ring =
 236             (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 237     }
 238
 239     ret = kvm_arch_init_vcpu(env);
 240     if (ret == 0) {
 241         qemu_register_reset(kvm_reset_vcpu, env);
 242         kvm_arch_reset_vcpu(env);
 243     }
 244 err:
 245     return ret;
 246 }
 247
 248 /*
 249  * dirty pages logging control
 250  */
 251 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 252                                       ram_addr_t size, int flags, int mask)
 253 {
 254     KVMState *s = kvm_state;
 255     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 256     int old_flags;
 257
 258     if (mem == NULL)  {
 259             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 260                     TARGET_FMT_plx "\n", __func__, phys_addr,
 261                     (target_phys_addr_t)(phys_addr + size - 1));
 262             return -EINVAL;
 263     }
 264
 265     old_flags = mem->flags;
 266
 267     flags = (mem->flags & ~mask) | flags;
 268     mem->flags = flags;
 269
 270     /* If nothing changed effectively, no need to issue ioctl */
 271     if (s->migration_log) {
 272         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 273     }
 274     if (flags == old_flags) {
 275             return 0;
 276     }
 277
 278     return kvm_set_user_memory_region(s, mem);
 279 }
 280
 281 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 282 {
 283     return kvm_dirty_pages_log_change(phys_addr, size, KVM_MEM_LOG_DIRTY_PAGES,
 284                                       KVM_MEM_LOG_DIRTY_PAGES);
 285 }
 286
 287 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 288 {
 289     return kvm_dirty_pages_log_change(phys_addr, size, 0,
 290                                       KVM_MEM_LOG_DIRTY_PAGES);
 291 }
 292
 293 static int kvm_set_migration_log(int enable)
 294 {
 295     KVMState *s = kvm_state;
 296     KVMSlot *mem;
 297     int i, err;
 298
 299     s->migration_log = enable;
 300
 301     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 302         mem = &s->slots[i];
 303
 304         if (!mem->memory_size) {
 305             continue;
 306         }
 307         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 308             continue;
 309         }
 310         err = kvm_set_user_memory_region(s, mem);
 311         if (err) {
 312             return err;
 313         }
 314     }
 315     return 0;
 316 }
 317
 318 /* get kvm's dirty pages bitmap and update qemu's */
 319 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 320                                          unsigned long *bitmap,
 321                                          unsigned long offset,
 322                                          unsigned long mem_size)
 323 {
 324     unsigned int i, j;
 325     unsigned long page_number, addr, addr1, c;
 326     ram_addr_t ram_addr;
 327     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 328         HOST_LONG_BITS;
 329
 330     /*
 331      * bitmap-traveling is faster than memory-traveling (for addr...)
 332      * especially when most of the memory is not dirty.
 333      */
 334     for (i = 0; i < len; i++) {
 335         if (bitmap[i] != 0) {
 336             c = leul_to_cpu(bitmap[i]);
 337             do {
 338                 j = ffsl(c) - 1;
 339                 c &= ~(1ul << j);
 340                 page_number = i * HOST_LONG_BITS + j;
 341                 addr1 = page_number * TARGET_PAGE_SIZE;
 342                 addr = offset + addr1;
 343                 ram_addr = cpu_get_physical_page_desc(addr);
 344                 cpu_physical_memory_set_dirty(ram_addr);
 345             } while (c != 0);
 346         }
 347     }
 348     return 0;
 349 }
 350
 351 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 352
 353 /**
 354  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 355  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 356  * This means all bits are set to dirty.
 357  *
 358  * @start_add: start of logged region.
 359  * @end_addr: end of logged region.
 360  */
 361 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 362                                           target_phys_addr_t end_addr)
 363 {
 364     KVMState *s = kvm_state;
 365     unsigned long size, allocated_size = 0;
 366     KVMDirtyLog d;
 367     KVMSlot *mem;
 368     int ret = 0;
 369
 370     d.dirty_bitmap = NULL;
 371     while (start_addr < end_addr) {
 372         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 373         if (mem == NULL) {
 374             break;
 375         }
 376
 377         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 378         if (!d.dirty_bitmap) {
 379             d.dirty_bitmap = qemu_malloc(size);
 380         } else if (size > allocated_size) {
 381             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 382         }
 383         allocated_size = size;
 384         memset(d.dirty_bitmap, 0, allocated_size);
 385
 386         d.slot = mem->slot;
 387
 388         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 389             DPRINTF("ioctl failed %d\n", errno);
 390             ret = -1;
 391             break;
 392         }
 393
 394         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 395                                       mem->start_addr, mem->memory_size);
 396         start_addr = mem->start_addr + mem->memory_size;
 397     }
 398     qemu_free(d.dirty_bitmap);
 399
 400     return ret;
 401 }
 402
 403 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 404 {
 405     int ret = -ENOSYS;
 406     KVMState *s = kvm_state;
 407
 408     if (s->coalesced_mmio) {
 409         struct kvm_coalesced_mmio_zone zone;
 410
 411         zone.addr = start;
 412         zone.size = size;
 413
 414         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 415     }
 416
 417     return ret;
 418 }
 419
 420 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 421 {
 422     int ret = -ENOSYS;
 423     KVMState *s = kvm_state;
 424
 425     if (s->coalesced_mmio) {
 426         struct kvm_coalesced_mmio_zone zone;
 427
 428         zone.addr = start;
 429         zone.size = size;
 430
 431         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 432     }
 433
 434     return ret;
 435 }
 436
 437 int kvm_check_extension(KVMState *s, unsigned int extension)
 438 {
 439     int ret;
 440
 441     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 442     if (ret < 0) {
 443         ret = 0;
 444     }
 445
 446     return ret;
 447 }
 448
 449 static int kvm_check_many_ioeventfds(void)
 450 {
 451     /* Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 452      * can avoid creating too many ioeventfds.
 453      */
 454 #ifdef CONFIG_EVENTFD
 455     int ioeventfds[7];
 456     int i, ret = 0;
 457     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 458         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 459         if (ioeventfds[i] < 0) {
 460             break;
 461         }
 462         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 463         if (ret < 0) {
 464             close(ioeventfds[i]);
 465             break;
 466         }
 467     }
 468
 469     /* Decide whether many devices are supported or not */
 470     ret = i == ARRAY_SIZE(ioeventfds);
 471
 472     while (i-- > 0) {
 473         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 474         close(ioeventfds[i]);
 475     }
 476     return ret;
 477 #else
 478     return 0;
 479 #endif
 480 }
 481
 482 static const KVMCapabilityInfo *
 483 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 484 {
 485     while (list->name) {
 486         if (!kvm_check_extension(s, list->value)) {
 487             return list;
 488         }
 489         list++;
 490     }
 491     return NULL;
 492 }
 493
 494 static void kvm_set_phys_mem(target_phys_addr_t start_addr, ram_addr_t size,
 495                              ram_addr_t phys_offset)
 496 {
 497     KVMState *s = kvm_state;
 498     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 499     KVMSlot *mem, old;
 500     int err;
 501
 502     /* kvm works in page size chunks, but the function may be called
 503        with sub-page size and unaligned start address. */
 504     size = TARGET_PAGE_ALIGN(size);
 505     start_addr = TARGET_PAGE_ALIGN(start_addr);
 506
 507     /* KVM does not support read-only slots */
 508     phys_offset &= ~IO_MEM_ROM;
 509
 510     while (1) {
 511         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 512         if (!mem) {
 513             break;
 514         }
 515
 516         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 517             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 518             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 519             /* The new slot fits into the existing one and comes with
 520              * identical parameters - nothing to be done. */
 521             return;
 522         }
 523
 524         old = *mem;
 525
 526         /* unregister the overlapping slot */
 527         mem->memory_size = 0;
 528         err = kvm_set_user_memory_region(s, mem);
 529         if (err) {
 530             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 531                     __func__, strerror(-err));
 532             abort();
 533         }
 534
 535         /* Workaround for older KVM versions: we can't join slots, even not by
 536          * unregistering the previous ones and then registering the larger
 537          * slot. We have to maintain the existing fragmentation. Sigh.
 538          *
 539          * This workaround assumes that the new slot starts at the same
 540          * address as the first existing one. If not or if some overlapping
 541          * slot comes around later, we will fail (not seen in practice so far)
 542          * - and actually require a recent KVM version. */
 543         if (s->broken_set_mem_region &&
 544             old.start_addr == start_addr && old.memory_size < size &&
 545             flags < IO_MEM_UNASSIGNED) {
 546             mem = kvm_alloc_slot(s);
 547             mem->memory_size = old.memory_size;
 548             mem->start_addr = old.start_addr;
 549             mem->phys_offset = old.phys_offset;
 550             mem->flags = 0;
 551
 552             err = kvm_set_user_memory_region(s, mem);
 553             if (err) {
 554                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 555                         strerror(-err));
 556                 abort();
 557             }
 558
 559             start_addr += old.memory_size;
 560             phys_offset += old.memory_size;
 561             size -= old.memory_size;
 562             continue;
 563         }
 564
 565         /* register prefix slot */
 566         if (old.start_addr < start_addr) {
 567             mem = kvm_alloc_slot(s);
 568             mem->memory_size = start_addr - old.start_addr;
 569             mem->start_addr = old.start_addr;
 570             mem->phys_offset = old.phys_offset;
 571             mem->flags = 0;
 572
 573             err = kvm_set_user_memory_region(s, mem);
 574             if (err) {
 575                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 576                         __func__, strerror(-err));
 577                 abort();
 578             }
 579         }
 580
 581         /* register suffix slot */
 582         if (old.start_addr + old.memory_size > start_addr + size) {
 583             ram_addr_t size_delta;
 584
 585             mem = kvm_alloc_slot(s);
 586             mem->start_addr = start_addr + size;
 587             size_delta = mem->start_addr - old.start_addr;
 588             mem->memory_size = old.memory_size - size_delta;
 589             mem->phys_offset = old.phys_offset + size_delta;
 590             mem->flags = 0;
 591
 592             err = kvm_set_user_memory_region(s, mem);
 593             if (err) {
 594                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 595                         __func__, strerror(-err));
 596                 abort();
 597             }
 598         }
 599     }
 600
 601     /* in case the KVM bug workaround already "consumed" the new slot */
 602     if (!size) {
 603         return;
 604     }
 605     /* KVM does not need to know about this memory */
 606     if (flags >= IO_MEM_UNASSIGNED) {
 607         return;
 608     }
 609     mem = kvm_alloc_slot(s);
 610     mem->memory_size = size;
 611     mem->start_addr = start_addr;
 612     mem->phys_offset = phys_offset;
 613     mem->flags = 0;
 614
 615     err = kvm_set_user_memory_region(s, mem);
 616     if (err) {
 617         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 618                 strerror(-err));
 619         abort();
 620     }
 621 }
 622
 623 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 624                                   target_phys_addr_t start_addr,
 625                                   ram_addr_t size, ram_addr_t phys_offset)
 626 {
 627     kvm_set_phys_mem(start_addr, size, phys_offset);
 628 }
 629
 630 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 631                                         target_phys_addr_t start_addr,
 632                                         target_phys_addr_t end_addr)
 633 {
 634     return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 635 }
 636
 637 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 638                                     int enable)
 639 {
 640     return kvm_set_migration_log(enable);
 641 }
 642
 643 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 644     .set_memory = kvm_client_set_memory,
 645     .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 646     .migration_log = kvm_client_migration_log,
 647 };
 648
 649 int kvm_init(void)
 650 {
 651     static const char upgrade_note[] =
 652         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 653         "(see http://sourceforge.net/projects/kvm).\n";
 654     KVMState *s;
 655     const KVMCapabilityInfo *missing_cap;
 656     int ret;
 657     int i;
 658
 659     s = qemu_mallocz(sizeof(KVMState));
 660
 661 #ifdef KVM_CAP_SET_GUEST_DEBUG
 662     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 663 #endif
 664     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 665         s->slots[i].slot = i;
 666     }
 667     s->vmfd = -1;
 668     s->fd = qemu_open("/dev/kvm", O_RDWR);
 669     if (s->fd == -1) {
 670         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 671         ret = -errno;
 672         goto err;
 673     }
 674
 675     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 676     if (ret < KVM_API_VERSION) {
 677         if (ret > 0) {
 678             ret = -EINVAL;
 679         }
 680         fprintf(stderr, "kvm version too old\n");
 681         goto err;
 682     }
 683
 684     if (ret > KVM_API_VERSION) {
 685         ret = -EINVAL;
 686         fprintf(stderr, "kvm version not supported\n");
 687         goto err;
 688     }
 689
 690     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 691     if (s->vmfd < 0) {
 692 #ifdef TARGET_S390X
 693         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 694                         "your host kernel command line\n");
 695 #endif
 696         goto err;
 697     }
 698
 699     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
 700     if (!missing_cap) {
 701         missing_cap =
 702             kvm_check_extension_list(s, kvm_arch_required_capabilities);
 703     }
 704     if (missing_cap) {
 705         ret = -EINVAL;
 706         fprintf(stderr, "kvm does not support %s\n%s",
 707                 missing_cap->name, upgrade_note);
 708         goto err;
 709     }
 710
 711     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 712
 713     s->broken_set_mem_region = 1;
 714 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 715     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 716     if (ret > 0) {
 717         s->broken_set_mem_region = 0;
 718     }
 719 #endif
 720
 721     s->vcpu_events = 0;
 722 #ifdef KVM_CAP_VCPU_EVENTS
 723     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 724 #endif
 725
 726     s->robust_singlestep = 0;
 727 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 728     s->robust_singlestep =
 729         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 730 #endif
 731
 732     s->debugregs = 0;
 733 #ifdef KVM_CAP_DEBUGREGS
 734     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 735 #endif
 736
 737     s->xsave = 0;
 738 #ifdef KVM_CAP_XSAVE
 739     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 740 #endif
 741
 742     s->xcrs = 0;
 743 #ifdef KVM_CAP_XCRS
 744     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 745 #endif
 746
 747     ret = kvm_arch_init(s);
 748     if (ret < 0) {
 749         goto err;
 750     }
 751
 752     kvm_state = s;
 753     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 754
 755     s->many_ioeventfds = kvm_check_many_ioeventfds();
 756
 757     return 0;
 758
 759 err:
 760     if (s) {
 761         if (s->vmfd != -1) {
 762             close(s->vmfd);
 763         }
 764         if (s->fd != -1) {
 765             close(s->fd);
 766         }
 767     }
 768     qemu_free(s);
 769
 770     return ret;
 771 }
 772
 773 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 774                          uint32_t count)
 775 {
 776     int i;
 777     uint8_t *ptr = data;
 778
 779     for (i = 0; i < count; i++) {
 780         if (direction == KVM_EXIT_IO_IN) {
 781             switch (size) {
 782             case 1:
 783                 stb_p(ptr, cpu_inb(port));
 784                 break;
 785             case 2:
 786                 stw_p(ptr, cpu_inw(port));
 787                 break;
 788             case 4:
 789                 stl_p(ptr, cpu_inl(port));
 790                 break;
 791             }
 792         } else {
 793             switch (size) {
 794             case 1:
 795                 cpu_outb(port, ldub_p(ptr));
 796                 break;
 797             case 2:
 798                 cpu_outw(port, lduw_p(ptr));
 799                 break;
 800             case 4:
 801                 cpu_outl(port, ldl_p(ptr));
 802                 break;
 803             }
 804         }
 805
 806         ptr += size;
 807     }
 808
 809     return 1;
 810 }
 811
 812 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 813 static int kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 814 {
 815     fprintf(stderr, "KVM internal error.");
 816     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 817         int i;
 818
 819         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
 820         for (i = 0; i < run->internal.ndata; ++i) {
 821             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 822                     i, (uint64_t)run->internal.data[i]);
 823         }
 824     } else {
 825         fprintf(stderr, "\n");
 826     }
 827     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 828         fprintf(stderr, "emulation failure\n");
 829         if (!kvm_arch_stop_on_emulation_error(env)) {
 830             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 831             return 0;
 832         }
 833     }
 834     /* FIXME: Should trigger a qmp message to let management know
 835      * something went wrong.
 836      */
 837     return -1;
 838 }
 839 #endif
 840
 841 void kvm_flush_coalesced_mmio_buffer(void)
 842 {
 843     KVMState *s = kvm_state;
 844     if (s->coalesced_mmio_ring) {
 845         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 846         while (ring->first != ring->last) {
 847             struct kvm_coalesced_mmio *ent;
 848
 849             ent = &ring->coalesced_mmio[ring->first];
 850
 851             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 852             smp_wmb();
 853             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 854         }
 855     }
 856 }
 857
 858 static void do_kvm_cpu_synchronize_state(void *_env)
 859 {
 860     CPUState *env = _env;
 861
 862     if (!env->kvm_vcpu_dirty) {
 863         kvm_arch_get_registers(env);
 864         env->kvm_vcpu_dirty = 1;
 865     }
 866 }
 867
 868 void kvm_cpu_synchronize_state(CPUState *env)
 869 {
 870     if (!env->kvm_vcpu_dirty) {
 871         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 872     }
 873 }
 874
 875 void kvm_cpu_synchronize_post_reset(CPUState *env)
 876 {
 877     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 878     env->kvm_vcpu_dirty = 0;
 879 }
 880
 881 void kvm_cpu_synchronize_post_init(CPUState *env)
 882 {
 883     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 884     env->kvm_vcpu_dirty = 0;
 885 }
 886
 887 int kvm_cpu_exec(CPUState *env)
 888 {
 889     struct kvm_run *run = env->kvm_run;
 890     int ret;
 891
 892     DPRINTF("kvm_cpu_exec()\n");
 893
 894     do {
 895 #ifndef CONFIG_IOTHREAD
 896         if (env->exit_request) {
 897             DPRINTF("interrupt exit requested\n");
 898             ret = 0;
 899             break;
 900         }
 901 #endif
 902
 903         if (kvm_arch_process_irqchip_events(env)) {
 904             ret = 0;
 905             break;
 906         }
 907
 908         if (env->kvm_vcpu_dirty) {
 909             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 910             env->kvm_vcpu_dirty = 0;
 911         }
 912
 913         kvm_arch_pre_run(env, run);
 914         cpu_single_env = NULL;
 915         qemu_mutex_unlock_iothread();
 916         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 917         qemu_mutex_lock_iothread();
 918         cpu_single_env = env;
 919         kvm_arch_post_run(env, run);
 920
 921         kvm_flush_coalesced_mmio_buffer();
 922
 923         if (ret == -EINTR || ret == -EAGAIN) {
 924             cpu_exit(env);
 925             DPRINTF("io window exit\n");
 926             ret = 0;
 927             break;
 928         }
 929
 930         if (ret < 0) {
 931             DPRINTF("kvm run failed %s\n", strerror(-ret));
 932             abort();
 933         }
 934
 935         ret = 0; /* exit loop */
 936         switch (run->exit_reason) {
 937         case KVM_EXIT_IO:
 938             DPRINTF("handle_io\n");
 939             ret = kvm_handle_io(run->io.port,
 940                                 (uint8_t *)run + run->io.data_offset,
 941                                 run->io.direction,
 942                                 run->io.size,
 943                                 run->io.count);
 944             break;
 945         case KVM_EXIT_MMIO:
 946             DPRINTF("handle_mmio\n");
 947             cpu_physical_memory_rw(run->mmio.phys_addr,
 948                                    run->mmio.data,
 949                                    run->mmio.len,
 950                                    run->mmio.is_write);
 951             ret = 1;
 952             break;
 953         case KVM_EXIT_IRQ_WINDOW_OPEN:
 954             DPRINTF("irq_window_open\n");
 955             break;
 956         case KVM_EXIT_SHUTDOWN:
 957             DPRINTF("shutdown\n");
 958             qemu_system_reset_request();
 959             ret = 1;
 960             break;
 961         case KVM_EXIT_UNKNOWN:
 962             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
 963                     (uint64_t)run->hw.hardware_exit_reason);
 964             ret = -1;
 965             break;
 966 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 967         case KVM_EXIT_INTERNAL_ERROR:
 968             ret = kvm_handle_internal_error(env, run);
 969             break;
 970 #endif
 971         case KVM_EXIT_DEBUG:
 972             DPRINTF("kvm_exit_debug\n");
 973 #ifdef KVM_CAP_SET_GUEST_DEBUG
 974             if (kvm_arch_debug(&run->debug.arch)) {
 975                 env->exception_index = EXCP_DEBUG;
 976                 return 0;
 977             }
 978             /* re-enter, this exception was guest-internal */
 979             ret = 1;
 980 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 981             break;
 982         default:
 983             DPRINTF("kvm_arch_handle_exit\n");
 984             ret = kvm_arch_handle_exit(env, run);
 985             break;
 986         }
 987     } while (ret > 0);
 988
 989     if (ret < 0) {
 990         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 991         vm_stop(0);
 992         env->exit_request = 1;
 993     }
 994     if (env->exit_request) {
 995         env->exit_request = 0;
 996         env->exception_index = EXCP_INTERRUPT;
 997     }
 998
 999     return ret;
1000 }
1001
1002 int kvm_ioctl(KVMState *s, int type, ...)
1003 {
1004     int ret;
1005     void *arg;
1006     va_list ap;
1007
1008     va_start(ap, type);
1009     arg = va_arg(ap, void *);
1010     va_end(ap);
1011
1012     ret = ioctl(s->fd, type, arg);
1013     if (ret == -1) {
1014         ret = -errno;
1015     }
1016     return ret;
1017 }
1018
1019 int kvm_vm_ioctl(KVMState *s, int type, ...)
1020 {
1021     int ret;
1022     void *arg;
1023     va_list ap;
1024
1025     va_start(ap, type);
1026     arg = va_arg(ap, void *);
1027     va_end(ap);
1028
1029     ret = ioctl(s->vmfd, type, arg);
1030     if (ret == -1) {
1031         ret = -errno;
1032     }
1033     return ret;
1034 }
1035
1036 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1037 {
1038     int ret;
1039     void *arg;
1040     va_list ap;
1041
1042     va_start(ap, type);
1043     arg = va_arg(ap, void *);
1044     va_end(ap);
1045
1046     ret = ioctl(env->kvm_fd, type, arg);
1047     if (ret == -1) {
1048         ret = -errno;
1049     }
1050     return ret;
1051 }
1052
1053 int kvm_has_sync_mmu(void)
1054 {
1055     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1056 }
1057
1058 int kvm_has_vcpu_events(void)
1059 {
1060     return kvm_state->vcpu_events;
1061 }
1062
1063 int kvm_has_robust_singlestep(void)
1064 {
1065     return kvm_state->robust_singlestep;
1066 }
1067
1068 int kvm_has_debugregs(void)
1069 {
1070     return kvm_state->debugregs;
1071 }
1072
1073 int kvm_has_xsave(void)
1074 {
1075     return kvm_state->xsave;
1076 }
1077
1078 int kvm_has_xcrs(void)
1079 {
1080     return kvm_state->xcrs;
1081 }
1082
1083 int kvm_has_many_ioeventfds(void)
1084 {
1085     if (!kvm_enabled()) {
1086         return 0;
1087     }
1088     return kvm_state->many_ioeventfds;
1089 }
1090
1091 void kvm_setup_guest_memory(void *start, size_t size)
1092 {
1093     if (!kvm_has_sync_mmu()) {
1094         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1095
1096         if (ret) {
1097             perror("qemu_madvise");
1098             fprintf(stderr,
1099                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1100             exit(1);
1101         }
1102     }
1103 }
1104
1105 #ifdef KVM_CAP_SET_GUEST_DEBUG
1106 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1107                                                  target_ulong pc)
1108 {
1109     struct kvm_sw_breakpoint *bp;
1110
1111     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1112         if (bp->pc == pc) {
1113             return bp;
1114         }
1115     }
1116     return NULL;
1117 }
1118
1119 int kvm_sw_breakpoints_active(CPUState *env)
1120 {
1121     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1122 }
1123
1124 struct kvm_set_guest_debug_data {
1125     struct kvm_guest_debug dbg;
1126     CPUState *env;
1127     int err;
1128 };
1129
1130 static void kvm_invoke_set_guest_debug(void *data)
1131 {
1132     struct kvm_set_guest_debug_data *dbg_data = data;
1133     CPUState *env = dbg_data->env;
1134
1135     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1136 }
1137
1138 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1139 {
1140     struct kvm_set_guest_debug_data data;
1141
1142     data.dbg.control = reinject_trap;
1143
1144     if (env->singlestep_enabled) {
1145         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1146     }
1147     kvm_arch_update_guest_debug(env, &data.dbg);
1148     data.env = env;
1149
1150     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1151     return data.err;
1152 }
1153
1154 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1155                           target_ulong len, int type)
1156 {
1157     struct kvm_sw_breakpoint *bp;
1158     CPUState *env;
1159     int err;
1160
1161     if (type == GDB_BREAKPOINT_SW) {
1162         bp = kvm_find_sw_breakpoint(current_env, addr);
1163         if (bp) {
1164             bp->use_count++;
1165             return 0;
1166         }
1167
1168         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1169         if (!bp) {
1170             return -ENOMEM;
1171         }
1172
1173         bp->pc = addr;
1174         bp->use_count = 1;
1175         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1176         if (err) {
1177             free(bp);
1178             return err;
1179         }
1180
1181         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1182                           bp, entry);
1183     } else {
1184         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1185         if (err) {
1186             return err;
1187         }
1188     }
1189
1190     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1191         err = kvm_update_guest_debug(env, 0);
1192         if (err) {
1193             return err;
1194         }
1195     }
1196     return 0;
1197 }
1198
1199 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1200                           target_ulong len, int type)
1201 {
1202     struct kvm_sw_breakpoint *bp;
1203     CPUState *env;
1204     int err;
1205
1206     if (type == GDB_BREAKPOINT_SW) {
1207         bp = kvm_find_sw_breakpoint(current_env, addr);
1208         if (!bp) {
1209             return -ENOENT;
1210         }
1211
1212         if (bp->use_count > 1) {
1213             bp->use_count--;
1214             return 0;
1215         }
1216
1217         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1218         if (err) {
1219             return err;
1220         }
1221
1222         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1223         qemu_free(bp);
1224     } else {
1225         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1226         if (err) {
1227             return err;
1228         }
1229     }
1230
1231     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1232         err = kvm_update_guest_debug(env, 0);
1233         if (err) {
1234             return err;
1235         }
1236     }
1237     return 0;
1238 }
1239
1240 void kvm_remove_all_breakpoints(CPUState *current_env)
1241 {
1242     struct kvm_sw_breakpoint *bp, *next;
1243     KVMState *s = current_env->kvm_state;
1244     CPUState *env;
1245
1246     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1247         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1248             /* Try harder to find a CPU that currently sees the breakpoint. */
1249             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1250                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1251                     break;
1252                 }
1253             }
1254         }
1255     }
1256     kvm_arch_remove_all_hw_breakpoints();
1257
1258     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1259         kvm_update_guest_debug(env, 0);
1260     }
1261 }
1262
1263 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1264
1265 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1266 {
1267     return -EINVAL;
1268 }
1269
1270 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1271                           target_ulong len, int type)
1272 {
1273     return -EINVAL;
1274 }
1275
1276 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1277                           target_ulong len, int type)
1278 {
1279     return -EINVAL;
1280 }
1281
1282 void kvm_remove_all_breakpoints(CPUState *current_env)
1283 {
1284 }
1285 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1286
1287 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1288 {
1289     struct kvm_signal_mask *sigmask;
1290     int r;
1291
1292     if (!sigset) {
1293         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1294     }
1295
1296     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1297
1298     sigmask->len = 8;
1299     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1300     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1301     free(sigmask);
1302
1303     return r;
1304 }
1305
1306 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1307 {
1308 #ifdef KVM_IOEVENTFD
1309     int ret;
1310     struct kvm_ioeventfd iofd;
1311
1312     iofd.datamatch = val;
1313     iofd.addr = addr;
1314     iofd.len = 4;
1315     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1316     iofd.fd = fd;
1317
1318     if (!kvm_enabled()) {
1319         return -ENOSYS;
1320     }
1321
1322     if (!assign) {
1323         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1324     }
1325
1326     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1327
1328     if (ret < 0) {
1329         return -errno;
1330     }
1331
1332     return 0;
1333 #else
1334     return -ENOSYS;
1335 #endif
1336 }
1337
1338 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1339 {
1340 #ifdef KVM_IOEVENTFD
1341     struct kvm_ioeventfd kick = {
1342         .datamatch = val,
1343         .addr = addr,
1344         .len = 2,
1345         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1346         .fd = fd,
1347     };
1348     int r;
1349     if (!kvm_enabled()) {
1350         return -ENOSYS;
1351     }
1352     if (!assign) {
1353         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1354     }
1355     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1356     if (r < 0) {
1357         return r;
1358     }
1359     return 0;
1360 #else
1361     return -ENOSYS;
1362 #endif
1363 }