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 #ifdef KVM_CAP_COALESCED_MMIO
  67     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  68 #endif
  69     int broken_set_mem_region;
  70     int migration_log;
  71     int vcpu_events;
  72     int robust_singlestep;
  73     int debugregs;
  74 #ifdef KVM_CAP_SET_GUEST_DEBUG
  75     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  76 #endif
  77     int irqchip_in_kernel;
  78     int pit_in_kernel;
  79     int xsave, xcrs;
  80     int many_ioeventfds;
  81 };
  82
  83 static KVMState *kvm_state;
  84
  85 static KVMSlot *kvm_alloc_slot(KVMState *s)
  86 {
  87     int i;
  88
  89     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
  90         /* KVM private memory slots */
  91         if (i >= 8 && i < 12) {
  92             continue;
  93         }
  94         if (s->slots[i].memory_size == 0) {
  95             return &s->slots[i];
  96         }
  97     }
  98
  99     fprintf(stderr, "%s: no free slot available\n", __func__);
 100     abort();
 101 }
 102
 103 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 104                                          target_phys_addr_t start_addr,
 105                                          target_phys_addr_t end_addr)
 106 {
 107     int i;
 108
 109     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 110         KVMSlot *mem = &s->slots[i];
 111
 112         if (start_addr == mem->start_addr &&
 113             end_addr == mem->start_addr + mem->memory_size) {
 114             return mem;
 115         }
 116     }
 117
 118     return NULL;
 119 }
 120
 121 /*
 122  * Find overlapping slot with lowest start address
 123  */
 124 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 125                                             target_phys_addr_t start_addr,
 126                                             target_phys_addr_t end_addr)
 127 {
 128     KVMSlot *found = NULL;
 129     int i;
 130
 131     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 132         KVMSlot *mem = &s->slots[i];
 133
 134         if (mem->memory_size == 0 ||
 135             (found && found->start_addr < mem->start_addr)) {
 136             continue;
 137         }
 138
 139         if (end_addr > mem->start_addr &&
 140             start_addr < mem->start_addr + mem->memory_size) {
 141             found = mem;
 142         }
 143     }
 144
 145     return found;
 146 }
 147
 148 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
 149                                       target_phys_addr_t *phys_addr)
 150 {
 151     int i;
 152
 153     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 154         KVMSlot *mem = &s->slots[i];
 155
 156         if (ram_addr >= mem->phys_offset &&
 157             ram_addr < mem->phys_offset + mem->memory_size) {
 158             *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
 159             return 1;
 160         }
 161     }
 162
 163     return 0;
 164 }
 165
 166 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 167 {
 168     struct kvm_userspace_memory_region mem;
 169
 170     mem.slot = slot->slot;
 171     mem.guest_phys_addr = slot->start_addr;
 172     mem.memory_size = slot->memory_size;
 173     mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
 174     mem.flags = slot->flags;
 175     if (s->migration_log) {
 176         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 177     }
 178     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 179 }
 180
 181 static void kvm_reset_vcpu(void *opaque)
 182 {
 183     CPUState *env = opaque;
 184
 185     kvm_arch_reset_vcpu(env);
 186 }
 187
 188 int kvm_irqchip_in_kernel(void)
 189 {
 190     return kvm_state->irqchip_in_kernel;
 191 }
 192
 193 int kvm_pit_in_kernel(void)
 194 {
 195     return kvm_state->pit_in_kernel;
 196 }
 197
 198
 199 int kvm_init_vcpu(CPUState *env)
 200 {
 201     KVMState *s = kvm_state;
 202     long mmap_size;
 203     int ret;
 204
 205     DPRINTF("kvm_init_vcpu\n");
 206
 207     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 208     if (ret < 0) {
 209         DPRINTF("kvm_create_vcpu failed\n");
 210         goto err;
 211     }
 212
 213     env->kvm_fd = ret;
 214     env->kvm_state = s;
 215
 216     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 217     if (mmap_size < 0) {
 218         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 219         goto err;
 220     }
 221
 222     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 223                         env->kvm_fd, 0);
 224     if (env->kvm_run == MAP_FAILED) {
 225         ret = -errno;
 226         DPRINTF("mmap'ing vcpu state failed\n");
 227         goto err;
 228     }
 229
 230 #ifdef KVM_CAP_COALESCED_MMIO
 231     if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
 232         s->coalesced_mmio_ring =
 233             (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
 234     }
 235 #endif
 236
 237     ret = kvm_arch_init_vcpu(env);
 238     if (ret == 0) {
 239         qemu_register_reset(kvm_reset_vcpu, env);
 240         kvm_arch_reset_vcpu(env);
 241     }
 242 err:
 243     return ret;
 244 }
 245
 246 /*
 247  * dirty pages logging control
 248  */
 249 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 250                                       ram_addr_t size, int flags, int mask)
 251 {
 252     KVMState *s = kvm_state;
 253     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 254     int old_flags;
 255
 256     if (mem == NULL)  {
 257             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 258                     TARGET_FMT_plx "\n", __func__, phys_addr,
 259                     (target_phys_addr_t)(phys_addr + size - 1));
 260             return -EINVAL;
 261     }
 262
 263     old_flags = mem->flags;
 264
 265     flags = (mem->flags & ~mask) | flags;
 266     mem->flags = flags;
 267
 268     /* If nothing changed effectively, no need to issue ioctl */
 269     if (s->migration_log) {
 270         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 271     }
 272     if (flags == old_flags) {
 273             return 0;
 274     }
 275
 276     return kvm_set_user_memory_region(s, mem);
 277 }
 278
 279 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 280 {
 281     return kvm_dirty_pages_log_change(phys_addr, size, KVM_MEM_LOG_DIRTY_PAGES,
 282                                       KVM_MEM_LOG_DIRTY_PAGES);
 283 }
 284
 285 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 286 {
 287     return kvm_dirty_pages_log_change(phys_addr, size, 0,
 288                                       KVM_MEM_LOG_DIRTY_PAGES);
 289 }
 290
 291 static int kvm_set_migration_log(int enable)
 292 {
 293     KVMState *s = kvm_state;
 294     KVMSlot *mem;
 295     int i, err;
 296
 297     s->migration_log = enable;
 298
 299     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 300         mem = &s->slots[i];
 301
 302         if (!mem->memory_size) {
 303             continue;
 304         }
 305         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 306             continue;
 307         }
 308         err = kvm_set_user_memory_region(s, mem);
 309         if (err) {
 310             return err;
 311         }
 312     }
 313     return 0;
 314 }
 315
 316 /* get kvm's dirty pages bitmap and update qemu's */
 317 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 318                                          unsigned long *bitmap,
 319                                          unsigned long offset,
 320                                          unsigned long mem_size)
 321 {
 322     unsigned int i, j;
 323     unsigned long page_number, addr, addr1, c;
 324     ram_addr_t ram_addr;
 325     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 326         HOST_LONG_BITS;
 327
 328     /*
 329      * bitmap-traveling is faster than memory-traveling (for addr...)
 330      * especially when most of the memory is not dirty.
 331      */
 332     for (i = 0; i < len; i++) {
 333         if (bitmap[i] != 0) {
 334             c = leul_to_cpu(bitmap[i]);
 335             do {
 336                 j = ffsl(c) - 1;
 337                 c &= ~(1ul << j);
 338                 page_number = i * HOST_LONG_BITS + j;
 339                 addr1 = page_number * TARGET_PAGE_SIZE;
 340                 addr = offset + addr1;
 341                 ram_addr = cpu_get_physical_page_desc(addr);
 342                 cpu_physical_memory_set_dirty(ram_addr);
 343             } while (c != 0);
 344         }
 345     }
 346     return 0;
 347 }
 348
 349 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 350
 351 /**
 352  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 353  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 354  * This means all bits are set to dirty.
 355  *
 356  * @start_add: start of logged region.
 357  * @end_addr: end of logged region.
 358  */
 359 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 360                                           target_phys_addr_t end_addr)
 361 {
 362     KVMState *s = kvm_state;
 363     unsigned long size, allocated_size = 0;
 364     KVMDirtyLog d;
 365     KVMSlot *mem;
 366     int ret = 0;
 367
 368     d.dirty_bitmap = NULL;
 369     while (start_addr < end_addr) {
 370         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 371         if (mem == NULL) {
 372             break;
 373         }
 374
 375         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 376         if (!d.dirty_bitmap) {
 377             d.dirty_bitmap = qemu_malloc(size);
 378         } else if (size > allocated_size) {
 379             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 380         }
 381         allocated_size = size;
 382         memset(d.dirty_bitmap, 0, allocated_size);
 383
 384         d.slot = mem->slot;
 385
 386         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 387             DPRINTF("ioctl failed %d\n", errno);
 388             ret = -1;
 389             break;
 390         }
 391
 392         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 393                                       mem->start_addr, mem->memory_size);
 394         start_addr = mem->start_addr + mem->memory_size;
 395     }
 396     qemu_free(d.dirty_bitmap);
 397
 398     return ret;
 399 }
 400
 401 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 402 {
 403     int ret = -ENOSYS;
 404 #ifdef KVM_CAP_COALESCED_MMIO
 405     KVMState *s = kvm_state;
 406
 407     if (s->coalesced_mmio) {
 408         struct kvm_coalesced_mmio_zone zone;
 409
 410         zone.addr = start;
 411         zone.size = size;
 412
 413         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 414     }
 415 #endif
 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 #ifdef KVM_CAP_COALESCED_MMIO
 424     KVMState *s = kvm_state;
 425
 426     if (s->coalesced_mmio) {
 427         struct kvm_coalesced_mmio_zone zone;
 428
 429         zone.addr = start;
 430         zone.size = size;
 431
 432         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 433     }
 434 #endif
 435
 436     return ret;
 437 }
 438
 439 int kvm_check_extension(KVMState *s, unsigned int extension)
 440 {
 441     int ret;
 442
 443     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 444     if (ret < 0) {
 445         ret = 0;
 446     }
 447
 448     return ret;
 449 }
 450
 451 static int kvm_check_many_ioeventfds(void)
 452 {
 453     /* Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 454      * can avoid creating too many ioeventfds.
 455      */
 456 #ifdef CONFIG_EVENTFD
 457     int ioeventfds[7];
 458     int i, ret = 0;
 459     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 460         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 461         if (ioeventfds[i] < 0) {
 462             break;
 463         }
 464         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 465         if (ret < 0) {
 466             close(ioeventfds[i]);
 467             break;
 468         }
 469     }
 470
 471     /* Decide whether many devices are supported or not */
 472     ret = i == ARRAY_SIZE(ioeventfds);
 473
 474     while (i-- > 0) {
 475         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 476         close(ioeventfds[i]);
 477     }
 478     return ret;
 479 #else
 480     return 0;
 481 #endif
 482 }
 483
 484 static void kvm_set_phys_mem(target_phys_addr_t start_addr, ram_addr_t size,
 485                              ram_addr_t phys_offset)
 486 {
 487     KVMState *s = kvm_state;
 488     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 489     KVMSlot *mem, old;
 490     int err;
 491
 492     /* kvm works in page size chunks, but the function may be called
 493        with sub-page size and unaligned start address. */
 494     size = TARGET_PAGE_ALIGN(size);
 495     start_addr = TARGET_PAGE_ALIGN(start_addr);
 496
 497     /* KVM does not support read-only slots */
 498     phys_offset &= ~IO_MEM_ROM;
 499
 500     while (1) {
 501         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 502         if (!mem) {
 503             break;
 504         }
 505
 506         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 507             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 508             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 509             /* The new slot fits into the existing one and comes with
 510              * identical parameters - nothing to be done. */
 511             return;
 512         }
 513
 514         old = *mem;
 515
 516         /* unregister the overlapping slot */
 517         mem->memory_size = 0;
 518         err = kvm_set_user_memory_region(s, mem);
 519         if (err) {
 520             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 521                     __func__, strerror(-err));
 522             abort();
 523         }
 524
 525         /* Workaround for older KVM versions: we can't join slots, even not by
 526          * unregistering the previous ones and then registering the larger
 527          * slot. We have to maintain the existing fragmentation. Sigh.
 528          *
 529          * This workaround assumes that the new slot starts at the same
 530          * address as the first existing one. If not or if some overlapping
 531          * slot comes around later, we will fail (not seen in practice so far)
 532          * - and actually require a recent KVM version. */
 533         if (s->broken_set_mem_region &&
 534             old.start_addr == start_addr && old.memory_size < size &&
 535             flags < IO_MEM_UNASSIGNED) {
 536             mem = kvm_alloc_slot(s);
 537             mem->memory_size = old.memory_size;
 538             mem->start_addr = old.start_addr;
 539             mem->phys_offset = old.phys_offset;
 540             mem->flags = 0;
 541
 542             err = kvm_set_user_memory_region(s, mem);
 543             if (err) {
 544                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 545                         strerror(-err));
 546                 abort();
 547             }
 548
 549             start_addr += old.memory_size;
 550             phys_offset += old.memory_size;
 551             size -= old.memory_size;
 552             continue;
 553         }
 554
 555         /* register prefix slot */
 556         if (old.start_addr < start_addr) {
 557             mem = kvm_alloc_slot(s);
 558             mem->memory_size = start_addr - old.start_addr;
 559             mem->start_addr = old.start_addr;
 560             mem->phys_offset = old.phys_offset;
 561             mem->flags = 0;
 562
 563             err = kvm_set_user_memory_region(s, mem);
 564             if (err) {
 565                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 566                         __func__, strerror(-err));
 567                 abort();
 568             }
 569         }
 570
 571         /* register suffix slot */
 572         if (old.start_addr + old.memory_size > start_addr + size) {
 573             ram_addr_t size_delta;
 574
 575             mem = kvm_alloc_slot(s);
 576             mem->start_addr = start_addr + size;
 577             size_delta = mem->start_addr - old.start_addr;
 578             mem->memory_size = old.memory_size - size_delta;
 579             mem->phys_offset = old.phys_offset + size_delta;
 580             mem->flags = 0;
 581
 582             err = kvm_set_user_memory_region(s, mem);
 583             if (err) {
 584                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 585                         __func__, strerror(-err));
 586                 abort();
 587             }
 588         }
 589     }
 590
 591     /* in case the KVM bug workaround already "consumed" the new slot */
 592     if (!size) {
 593         return;
 594     }
 595     /* KVM does not need to know about this memory */
 596     if (flags >= IO_MEM_UNASSIGNED) {
 597         return;
 598     }
 599     mem = kvm_alloc_slot(s);
 600     mem->memory_size = size;
 601     mem->start_addr = start_addr;
 602     mem->phys_offset = phys_offset;
 603     mem->flags = 0;
 604
 605     err = kvm_set_user_memory_region(s, mem);
 606     if (err) {
 607         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 608                 strerror(-err));
 609         abort();
 610     }
 611 }
 612
 613 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 614                                   target_phys_addr_t start_addr,
 615                                   ram_addr_t size, ram_addr_t phys_offset)
 616 {
 617     kvm_set_phys_mem(start_addr, size, phys_offset);
 618 }
 619
 620 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 621                                         target_phys_addr_t start_addr,
 622                                         target_phys_addr_t end_addr)
 623 {
 624     return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 625 }
 626
 627 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 628                                     int enable)
 629 {
 630     return kvm_set_migration_log(enable);
 631 }
 632
 633 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 634     .set_memory = kvm_client_set_memory,
 635     .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 636     .migration_log = kvm_client_migration_log,
 637 };
 638
 639 int kvm_init(void)
 640 {
 641     static const char upgrade_note[] =
 642         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 643         "(see http://sourceforge.net/projects/kvm).\n";
 644     KVMState *s;
 645     int ret;
 646     int i;
 647
 648     s = qemu_mallocz(sizeof(KVMState));
 649
 650 #ifdef KVM_CAP_SET_GUEST_DEBUG
 651     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 652 #endif
 653     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 654         s->slots[i].slot = i;
 655     }
 656     s->vmfd = -1;
 657     s->fd = qemu_open("/dev/kvm", O_RDWR);
 658     if (s->fd == -1) {
 659         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 660         ret = -errno;
 661         goto err;
 662     }
 663
 664     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 665     if (ret < KVM_API_VERSION) {
 666         if (ret > 0) {
 667             ret = -EINVAL;
 668         }
 669         fprintf(stderr, "kvm version too old\n");
 670         goto err;
 671     }
 672
 673     if (ret > KVM_API_VERSION) {
 674         ret = -EINVAL;
 675         fprintf(stderr, "kvm version not supported\n");
 676         goto err;
 677     }
 678
 679     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 680     if (s->vmfd < 0) {
 681 #ifdef TARGET_S390X
 682         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 683                         "your host kernel command line\n");
 684 #endif
 685         goto err;
 686     }
 687
 688     /* initially, KVM allocated its own memory and we had to jump through
 689      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 690      * just use a user allocated buffer so we can use regular pages
 691      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 692      */
 693     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 694         ret = -EINVAL;
 695         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 696                 upgrade_note);
 697         goto err;
 698     }
 699
 700     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 701      * destroyed properly.  Since we rely on this capability, refuse to work
 702      * with any kernel without this capability. */
 703     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 704         ret = -EINVAL;
 705
 706         fprintf(stderr,
 707                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 708                 upgrade_note);
 709         goto err;
 710     }
 711
 712     s->coalesced_mmio = 0;
 713 #ifdef KVM_CAP_COALESCED_MMIO
 714     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 715     s->coalesced_mmio_ring = NULL;
 716 #endif
 717
 718     s->broken_set_mem_region = 1;
 719 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 720     ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 721     if (ret > 0) {
 722         s->broken_set_mem_region = 0;
 723     }
 724 #endif
 725
 726     s->vcpu_events = 0;
 727 #ifdef KVM_CAP_VCPU_EVENTS
 728     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 729 #endif
 730
 731     s->robust_singlestep = 0;
 732 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 733     s->robust_singlestep =
 734         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 735 #endif
 736
 737     s->debugregs = 0;
 738 #ifdef KVM_CAP_DEBUGREGS
 739     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 740 #endif
 741
 742     s->xsave = 0;
 743 #ifdef KVM_CAP_XSAVE
 744     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 745 #endif
 746
 747     s->xcrs = 0;
 748 #ifdef KVM_CAP_XCRS
 749     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 750 #endif
 751
 752     ret = kvm_arch_init(s);
 753     if (ret < 0) {
 754         goto err;
 755     }
 756
 757     kvm_state = s;
 758     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 759
 760     s->many_ioeventfds = kvm_check_many_ioeventfds();
 761
 762     return 0;
 763
 764 err:
 765     if (s) {
 766         if (s->vmfd != -1) {
 767             close(s->vmfd);
 768         }
 769         if (s->fd != -1) {
 770             close(s->fd);
 771         }
 772     }
 773     qemu_free(s);
 774
 775     return ret;
 776 }
 777
 778 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 779                          uint32_t count)
 780 {
 781     int i;
 782     uint8_t *ptr = data;
 783
 784     for (i = 0; i < count; i++) {
 785         if (direction == KVM_EXIT_IO_IN) {
 786             switch (size) {
 787             case 1:
 788                 stb_p(ptr, cpu_inb(port));
 789                 break;
 790             case 2:
 791                 stw_p(ptr, cpu_inw(port));
 792                 break;
 793             case 4:
 794                 stl_p(ptr, cpu_inl(port));
 795                 break;
 796             }
 797         } else {
 798             switch (size) {
 799             case 1:
 800                 cpu_outb(port, ldub_p(ptr));
 801                 break;
 802             case 2:
 803                 cpu_outw(port, lduw_p(ptr));
 804                 break;
 805             case 4:
 806                 cpu_outl(port, ldl_p(ptr));
 807                 break;
 808             }
 809         }
 810
 811         ptr += size;
 812     }
 813
 814     return 1;
 815 }
 816
 817 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 818 static int kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 819 {
 820     fprintf(stderr, "KVM internal error.");
 821     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 822         int i;
 823
 824         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
 825         for (i = 0; i < run->internal.ndata; ++i) {
 826             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 827                     i, (uint64_t)run->internal.data[i]);
 828         }
 829     } else {
 830         fprintf(stderr, "\n");
 831     }
 832     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 833         fprintf(stderr, "emulation failure\n");
 834         if (!kvm_arch_stop_on_emulation_error(env)) {
 835             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 836             return 0;
 837         }
 838     }
 839     /* FIXME: Should trigger a qmp message to let management know
 840      * something went wrong.
 841      */
 842     return -1;
 843 }
 844 #endif
 845
 846 void kvm_flush_coalesced_mmio_buffer(void)
 847 {
 848 #ifdef KVM_CAP_COALESCED_MMIO
 849     KVMState *s = kvm_state;
 850     if (s->coalesced_mmio_ring) {
 851         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 852         while (ring->first != ring->last) {
 853             struct kvm_coalesced_mmio *ent;
 854
 855             ent = &ring->coalesced_mmio[ring->first];
 856
 857             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 858             smp_wmb();
 859             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 860         }
 861     }
 862 #endif
 863 }
 864
 865 static void do_kvm_cpu_synchronize_state(void *_env)
 866 {
 867     CPUState *env = _env;
 868
 869     if (!env->kvm_vcpu_dirty) {
 870         kvm_arch_get_registers(env);
 871         env->kvm_vcpu_dirty = 1;
 872     }
 873 }
 874
 875 void kvm_cpu_synchronize_state(CPUState *env)
 876 {
 877     if (!env->kvm_vcpu_dirty) {
 878         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 879     }
 880 }
 881
 882 void kvm_cpu_synchronize_post_reset(CPUState *env)
 883 {
 884     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 885     env->kvm_vcpu_dirty = 0;
 886 }
 887
 888 void kvm_cpu_synchronize_post_init(CPUState *env)
 889 {
 890     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 891     env->kvm_vcpu_dirty = 0;
 892 }
 893
 894 int kvm_cpu_exec(CPUState *env)
 895 {
 896     struct kvm_run *run = env->kvm_run;
 897     int ret;
 898
 899     DPRINTF("kvm_cpu_exec()\n");
 900
 901     do {
 902 #ifndef CONFIG_IOTHREAD
 903         if (env->exit_request) {
 904             DPRINTF("interrupt exit requested\n");
 905             ret = 0;
 906             break;
 907         }
 908 #endif
 909
 910         if (kvm_arch_process_irqchip_events(env)) {
 911             ret = 0;
 912             break;
 913         }
 914
 915         if (env->kvm_vcpu_dirty) {
 916             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 917             env->kvm_vcpu_dirty = 0;
 918         }
 919
 920         kvm_arch_pre_run(env, run);
 921         cpu_single_env = NULL;
 922         qemu_mutex_unlock_iothread();
 923         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 924         qemu_mutex_lock_iothread();
 925         cpu_single_env = env;
 926         kvm_arch_post_run(env, run);
 927
 928         if (ret == -EINTR || ret == -EAGAIN) {
 929             cpu_exit(env);
 930             DPRINTF("io window exit\n");
 931             ret = 0;
 932             break;
 933         }
 934
 935         if (ret < 0) {
 936             DPRINTF("kvm run failed %s\n", strerror(-ret));
 937             abort();
 938         }
 939
 940         kvm_flush_coalesced_mmio_buffer();
 941
 942         ret = 0; /* exit loop */
 943         switch (run->exit_reason) {
 944         case KVM_EXIT_IO:
 945             DPRINTF("handle_io\n");
 946             ret = kvm_handle_io(run->io.port,
 947                                 (uint8_t *)run + run->io.data_offset,
 948                                 run->io.direction,
 949                                 run->io.size,
 950                                 run->io.count);
 951             break;
 952         case KVM_EXIT_MMIO:
 953             DPRINTF("handle_mmio\n");
 954             cpu_physical_memory_rw(run->mmio.phys_addr,
 955                                    run->mmio.data,
 956                                    run->mmio.len,
 957                                    run->mmio.is_write);
 958             ret = 1;
 959             break;
 960         case KVM_EXIT_IRQ_WINDOW_OPEN:
 961             DPRINTF("irq_window_open\n");
 962             break;
 963         case KVM_EXIT_SHUTDOWN:
 964             DPRINTF("shutdown\n");
 965             qemu_system_reset_request();
 966             ret = 1;
 967             break;
 968         case KVM_EXIT_UNKNOWN:
 969             fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
 970                     (uint64_t)run->hw.hardware_exit_reason);
 971             ret = -1;
 972             break;
 973 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 974         case KVM_EXIT_INTERNAL_ERROR:
 975             ret = kvm_handle_internal_error(env, run);
 976             break;
 977 #endif
 978         case KVM_EXIT_DEBUG:
 979             DPRINTF("kvm_exit_debug\n");
 980 #ifdef KVM_CAP_SET_GUEST_DEBUG
 981             if (kvm_arch_debug(&run->debug.arch)) {
 982                 env->exception_index = EXCP_DEBUG;
 983                 return 0;
 984             }
 985             /* re-enter, this exception was guest-internal */
 986             ret = 1;
 987 #endif /* KVM_CAP_SET_GUEST_DEBUG */
 988             break;
 989         default:
 990             DPRINTF("kvm_arch_handle_exit\n");
 991             ret = kvm_arch_handle_exit(env, run);
 992             break;
 993         }
 994     } while (ret > 0);
 995
 996     if (ret < 0) {
 997         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 998         vm_stop(0);
 999         env->exit_request = 1;
1000     }
1001     if (env->exit_request) {
1002         env->exit_request = 0;
1003         env->exception_index = EXCP_INTERRUPT;
1004     }
1005
1006     return ret;
1007 }
1008
1009 int kvm_ioctl(KVMState *s, int type, ...)
1010 {
1011     int ret;
1012     void *arg;
1013     va_list ap;
1014
1015     va_start(ap, type);
1016     arg = va_arg(ap, void *);
1017     va_end(ap);
1018
1019     ret = ioctl(s->fd, type, arg);
1020     if (ret == -1) {
1021         ret = -errno;
1022     }
1023     return ret;
1024 }
1025
1026 int kvm_vm_ioctl(KVMState *s, int type, ...)
1027 {
1028     int ret;
1029     void *arg;
1030     va_list ap;
1031
1032     va_start(ap, type);
1033     arg = va_arg(ap, void *);
1034     va_end(ap);
1035
1036     ret = ioctl(s->vmfd, type, arg);
1037     if (ret == -1) {
1038         ret = -errno;
1039     }
1040     return ret;
1041 }
1042
1043 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1044 {
1045     int ret;
1046     void *arg;
1047     va_list ap;
1048
1049     va_start(ap, type);
1050     arg = va_arg(ap, void *);
1051     va_end(ap);
1052
1053     ret = ioctl(env->kvm_fd, type, arg);
1054     if (ret == -1) {
1055         ret = -errno;
1056     }
1057     return ret;
1058 }
1059
1060 int kvm_has_sync_mmu(void)
1061 {
1062 #ifdef KVM_CAP_SYNC_MMU
1063     KVMState *s = kvm_state;
1064
1065     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
1066 #else
1067     return 0;
1068 #endif
1069 }
1070
1071 int kvm_has_vcpu_events(void)
1072 {
1073     return kvm_state->vcpu_events;
1074 }
1075
1076 int kvm_has_robust_singlestep(void)
1077 {
1078     return kvm_state->robust_singlestep;
1079 }
1080
1081 int kvm_has_debugregs(void)
1082 {
1083     return kvm_state->debugregs;
1084 }
1085
1086 int kvm_has_xsave(void)
1087 {
1088     return kvm_state->xsave;
1089 }
1090
1091 int kvm_has_xcrs(void)
1092 {
1093     return kvm_state->xcrs;
1094 }
1095
1096 int kvm_has_many_ioeventfds(void)
1097 {
1098     if (!kvm_enabled()) {
1099         return 0;
1100     }
1101     return kvm_state->many_ioeventfds;
1102 }
1103
1104 void kvm_setup_guest_memory(void *start, size_t size)
1105 {
1106     if (!kvm_has_sync_mmu()) {
1107         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1108
1109         if (ret) {
1110             perror("qemu_madvise");
1111             fprintf(stderr,
1112                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1113             exit(1);
1114         }
1115     }
1116 }
1117
1118 #ifdef KVM_CAP_SET_GUEST_DEBUG
1119 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1120                                                  target_ulong pc)
1121 {
1122     struct kvm_sw_breakpoint *bp;
1123
1124     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1125         if (bp->pc == pc) {
1126             return bp;
1127         }
1128     }
1129     return NULL;
1130 }
1131
1132 int kvm_sw_breakpoints_active(CPUState *env)
1133 {
1134     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1135 }
1136
1137 struct kvm_set_guest_debug_data {
1138     struct kvm_guest_debug dbg;
1139     CPUState *env;
1140     int err;
1141 };
1142
1143 static void kvm_invoke_set_guest_debug(void *data)
1144 {
1145     struct kvm_set_guest_debug_data *dbg_data = data;
1146     CPUState *env = dbg_data->env;
1147
1148     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1149 }
1150
1151 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1152 {
1153     struct kvm_set_guest_debug_data data;
1154
1155     data.dbg.control = reinject_trap;
1156
1157     if (env->singlestep_enabled) {
1158         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1159     }
1160     kvm_arch_update_guest_debug(env, &data.dbg);
1161     data.env = env;
1162
1163     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1164     return data.err;
1165 }
1166
1167 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1168                           target_ulong len, int type)
1169 {
1170     struct kvm_sw_breakpoint *bp;
1171     CPUState *env;
1172     int err;
1173
1174     if (type == GDB_BREAKPOINT_SW) {
1175         bp = kvm_find_sw_breakpoint(current_env, addr);
1176         if (bp) {
1177             bp->use_count++;
1178             return 0;
1179         }
1180
1181         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1182         if (!bp) {
1183             return -ENOMEM;
1184         }
1185
1186         bp->pc = addr;
1187         bp->use_count = 1;
1188         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1189         if (err) {
1190             free(bp);
1191             return err;
1192         }
1193
1194         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1195                           bp, entry);
1196     } else {
1197         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1198         if (err) {
1199             return err;
1200         }
1201     }
1202
1203     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1204         err = kvm_update_guest_debug(env, 0);
1205         if (err) {
1206             return err;
1207         }
1208     }
1209     return 0;
1210 }
1211
1212 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1213                           target_ulong len, int type)
1214 {
1215     struct kvm_sw_breakpoint *bp;
1216     CPUState *env;
1217     int err;
1218
1219     if (type == GDB_BREAKPOINT_SW) {
1220         bp = kvm_find_sw_breakpoint(current_env, addr);
1221         if (!bp) {
1222             return -ENOENT;
1223         }
1224
1225         if (bp->use_count > 1) {
1226             bp->use_count--;
1227             return 0;
1228         }
1229
1230         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1231         if (err) {
1232             return err;
1233         }
1234
1235         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1236         qemu_free(bp);
1237     } else {
1238         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1239         if (err) {
1240             return err;
1241         }
1242     }
1243
1244     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1245         err = kvm_update_guest_debug(env, 0);
1246         if (err) {
1247             return err;
1248         }
1249     }
1250     return 0;
1251 }
1252
1253 void kvm_remove_all_breakpoints(CPUState *current_env)
1254 {
1255     struct kvm_sw_breakpoint *bp, *next;
1256     KVMState *s = current_env->kvm_state;
1257     CPUState *env;
1258
1259     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1260         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1261             /* Try harder to find a CPU that currently sees the breakpoint. */
1262             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1263                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1264                     break;
1265                 }
1266             }
1267         }
1268     }
1269     kvm_arch_remove_all_hw_breakpoints();
1270
1271     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1272         kvm_update_guest_debug(env, 0);
1273     }
1274 }
1275
1276 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1277
1278 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1279 {
1280     return -EINVAL;
1281 }
1282
1283 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1284                           target_ulong len, int type)
1285 {
1286     return -EINVAL;
1287 }
1288
1289 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1290                           target_ulong len, int type)
1291 {
1292     return -EINVAL;
1293 }
1294
1295 void kvm_remove_all_breakpoints(CPUState *current_env)
1296 {
1297 }
1298 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1299
1300 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1301 {
1302     struct kvm_signal_mask *sigmask;
1303     int r;
1304
1305     if (!sigset) {
1306         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1307     }
1308
1309     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1310
1311     sigmask->len = 8;
1312     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1313     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1314     free(sigmask);
1315
1316     return r;
1317 }
1318
1319 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1320 {
1321 #ifdef KVM_IOEVENTFD
1322     int ret;
1323     struct kvm_ioeventfd iofd;
1324
1325     iofd.datamatch = val;
1326     iofd.addr = addr;
1327     iofd.len = 4;
1328     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1329     iofd.fd = fd;
1330
1331     if (!kvm_enabled()) {
1332         return -ENOSYS;
1333     }
1334
1335     if (!assign) {
1336         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1337     }
1338
1339     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1340
1341     if (ret < 0) {
1342         return -errno;
1343     }
1344
1345     return 0;
1346 #else
1347     return -ENOSYS;
1348 #endif
1349 }
1350
1351 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1352 {
1353 #ifdef KVM_IOEVENTFD
1354     struct kvm_ioeventfd kick = {
1355         .datamatch = val,
1356         .addr = addr,
1357         .len = 2,
1358         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1359         .fd = fd,
1360     };
1361     int r;
1362     if (!kvm_enabled()) {
1363         return -ENOSYS;
1364     }
1365     if (!assign) {
1366         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1367     }
1368     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1369     if (r < 0) {
1370         return r;
1371     }
1372     return 0;
1373 #else
1374     return -ENOSYS;
1375 #endif
1376 }