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 #ifdef OBSOLETE_KVM_IMPL
  50
  51 typedef struct KVMSlot
  52 {
  53     target_phys_addr_t start_addr;
  54     ram_addr_t memory_size;
  55     ram_addr_t phys_offset;
  56     int slot;
  57     int flags;
  58 } KVMSlot;
  59
  60 typedef struct kvm_dirty_log KVMDirtyLog;
  61
  62 struct KVMState
  63 {
  64     KVMSlot slots[32];
  65     int fd;
  66     int vmfd;
  67     int coalesced_mmio;
  68 #ifdef KVM_CAP_COALESCED_MMIO
  69     struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
  70 #endif
  71     int broken_set_mem_region;
  72     int migration_log;
  73     int vcpu_events;
  74     int robust_singlestep;
  75     int debugregs;
  76 #ifdef KVM_CAP_SET_GUEST_DEBUG
  77     struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
  78 #endif
  79     int irqchip_in_kernel;
  80     int pit_in_kernel;
  81     int xsave, xcrs;
  82     int many_ioeventfds;
  83 };
  84
  85 static KVMState *kvm_state;
  86
  87 #endif
  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         if (s->slots[i].memory_size == 0)
  98             return &s->slots[i];
  99     }
 100
 101     fprintf(stderr, "%s: no free slot available\n", __func__);
 102     abort();
 103 }
 104
 105 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
 106                                          target_phys_addr_t start_addr,
 107                                          target_phys_addr_t end_addr)
 108 {
 109     int i;
 110
 111     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 112         KVMSlot *mem = &s->slots[i];
 113
 114         if (start_addr == mem->start_addr &&
 115             end_addr == mem->start_addr + mem->memory_size) {
 116             return mem;
 117         }
 118     }
 119
 120     return NULL;
 121 }
 122
 123 /*
 124  * Find overlapping slot with lowest start address
 125  */
 126 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
 127                                             target_phys_addr_t start_addr,
 128                                             target_phys_addr_t end_addr)
 129 {
 130     KVMSlot *found = NULL;
 131     int i;
 132
 133     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 134         KVMSlot *mem = &s->slots[i];
 135
 136         if (mem->memory_size == 0 ||
 137             (found && found->start_addr < mem->start_addr)) {
 138             continue;
 139         }
 140
 141         if (end_addr > mem->start_addr &&
 142             start_addr < mem->start_addr + mem->memory_size) {
 143             found = mem;
 144         }
 145     }
 146
 147     return found;
 148 }
 149
 150 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
 151                                       target_phys_addr_t *phys_addr)
 152 {
 153     int i;
 154
 155     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 156         KVMSlot *mem = &s->slots[i];
 157
 158         if (ram_addr >= mem->phys_offset &&
 159             ram_addr < mem->phys_offset + mem->memory_size) {
 160             *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
 161             return 1;
 162         }
 163     }
 164
 165     return 0;
 166 }
 167
 168 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
 169 {
 170     struct kvm_userspace_memory_region mem;
 171
 172     mem.slot = slot->slot;
 173     mem.guest_phys_addr = slot->start_addr;
 174     mem.memory_size = slot->memory_size;
 175     mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
 176     mem.flags = slot->flags;
 177     if (s->migration_log) {
 178         mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
 179     }
 180     return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
 181 }
 182
 183 #ifdef OBSOLETE_KVM_IMPL
 184 static void kvm_reset_vcpu(void *opaque)
 185 {
 186     CPUState *env = opaque;
 187
 188     kvm_arch_reset_vcpu(env);
 189 }
 190 #endif
 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 #ifdef OBSOLETE_KVM_IMPL
 204 int kvm_init_vcpu(CPUState *env)
 205 {
 206     KVMState *s = kvm_state;
 207     long mmap_size;
 208     int ret;
 209
 210     DPRINTF("kvm_init_vcpu\n");
 211
 212     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 213     if (ret < 0) {
 214         DPRINTF("kvm_create_vcpu failed\n");
 215         goto err;
 216     }
 217
 218     env->kvm_fd = ret;
 219     env->kvm_state = s;
 220
 221     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 222     if (mmap_size < 0) {
 223         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 224         goto err;
 225     }
 226
 227     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 228                         env->kvm_fd, 0);
 229     if (env->kvm_run == MAP_FAILED) {
 230         ret = -errno;
 231         DPRINTF("mmap'ing vcpu state failed\n");
 232         goto err;
 233     }
 234
 235 #ifdef KVM_CAP_COALESCED_MMIO
 236     if (s->coalesced_mmio && !s->coalesced_mmio_ring)
 237         s->coalesced_mmio_ring = (void *) env->kvm_run +
 238                 s->coalesced_mmio * PAGE_SIZE;
 239 #endif
 240
 241     ret = kvm_arch_init_vcpu(env);
 242     if (ret == 0) {
 243         qemu_register_reset(kvm_reset_vcpu, env);
 244         kvm_arch_reset_vcpu(env);
 245     }
 246 err:
 247     return ret;
 248 }
 249 #endif
 250
 251 /*
 252  * dirty pages logging control
 253  */
 254 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 255                                       ram_addr_t size, int flags, int mask)
 256 {
 257     KVMState *s = kvm_state;
 258     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 259     int old_flags;
 260
 261     if (mem == NULL)  {
 262             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 263                     TARGET_FMT_plx "\n", __func__, phys_addr,
 264                     (target_phys_addr_t)(phys_addr + size - 1));
 265             return -EINVAL;
 266     }
 267
 268     old_flags = mem->flags;
 269
 270     flags = (mem->flags & ~mask) | flags;
 271     mem->flags = flags;
 272
 273     /* If nothing changed effectively, no need to issue ioctl */
 274     if (s->migration_log) {
 275         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 276     }
 277     if (flags == old_flags) {
 278             return 0;
 279     }
 280
 281     return kvm_set_user_memory_region(s, mem);
 282 }
 283
 284 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
 285 {
 286         return kvm_dirty_pages_log_change(phys_addr, size,
 287                                           KVM_MEM_LOG_DIRTY_PAGES,
 288                                           KVM_MEM_LOG_DIRTY_PAGES);
 289 }
 290
 291 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
 292 {
 293         return kvm_dirty_pages_log_change(phys_addr, size,
 294                                           0,
 295                                           KVM_MEM_LOG_DIRTY_PAGES);
 296 }
 297
 298 static int kvm_set_migration_log(int enable)
 299 {
 300     KVMState *s = kvm_state;
 301     KVMSlot *mem;
 302     int i, err;
 303
 304     s->migration_log = enable;
 305
 306     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 307         mem = &s->slots[i];
 308
 309         if (!mem->memory_size) {
 310             continue;
 311         }
 312         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 313             continue;
 314         }
 315         err = kvm_set_user_memory_region(s, mem);
 316         if (err) {
 317             return err;
 318         }
 319     }
 320     return 0;
 321 }
 322
 323 /* get kvm's dirty pages bitmap and update qemu's */
 324 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 325                                          unsigned long *bitmap,
 326                                          unsigned long offset,
 327                                          unsigned long mem_size)
 328 {
 329     unsigned int i, j;
 330     unsigned long page_number, addr, addr1, c;
 331     ram_addr_t ram_addr;
 332     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 333         HOST_LONG_BITS;
 334
 335     /*
 336      * bitmap-traveling is faster than memory-traveling (for addr...)
 337      * especially when most of the memory is not dirty.
 338      */
 339     for (i = 0; i < len; i++) {
 340         if (bitmap[i] != 0) {
 341             c = leul_to_cpu(bitmap[i]);
 342             do {
 343                 j = ffsl(c) - 1;
 344                 c &= ~(1ul << j);
 345                 page_number = i * HOST_LONG_BITS + j;
 346                 addr1 = page_number * TARGET_PAGE_SIZE;
 347                 addr = offset + addr1;
 348                 ram_addr = cpu_get_physical_page_desc(addr);
 349                 cpu_physical_memory_set_dirty(ram_addr);
 350             } while (c != 0);
 351         }
 352     }
 353     return 0;
 354 }
 355
 356 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 357
 358 /**
 359  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 360  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 361  * This means all bits are set to dirty.
 362  *
 363  * @start_add: start of logged region.
 364  * @end_addr: end of logged region.
 365  */
 366 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 367                                           target_phys_addr_t end_addr)
 368 {
 369     KVMState *s = kvm_state;
 370     unsigned long size, allocated_size = 0;
 371     KVMDirtyLog d;
 372     KVMSlot *mem;
 373     int ret = 0;
 374
 375     d.dirty_bitmap = NULL;
 376     while (start_addr < end_addr) {
 377         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 378         if (mem == NULL) {
 379             break;
 380         }
 381
 382         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 383         if (!d.dirty_bitmap) {
 384             d.dirty_bitmap = qemu_malloc(size);
 385         } else if (size > allocated_size) {
 386             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 387         }
 388         allocated_size = size;
 389         memset(d.dirty_bitmap, 0, allocated_size);
 390
 391         d.slot = mem->slot;
 392
 393         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 394             DPRINTF("ioctl failed %d\n", errno);
 395             ret = -1;
 396             break;
 397         }
 398
 399         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 400                                       mem->start_addr, mem->memory_size);
 401         start_addr = mem->start_addr + mem->memory_size;
 402     }
 403     qemu_free(d.dirty_bitmap);
 404
 405     return ret;
 406 }
 407
 408 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 409 {
 410     int ret = -ENOSYS;
 411 #ifdef KVM_CAP_COALESCED_MMIO
 412     KVMState *s = kvm_state;
 413
 414     if (s->coalesced_mmio) {
 415         struct kvm_coalesced_mmio_zone zone;
 416
 417         zone.addr = start;
 418         zone.size = size;
 419
 420         ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
 421     }
 422 #endif
 423
 424     return ret;
 425 }
 426
 427 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 428 {
 429     int ret = -ENOSYS;
 430 #ifdef KVM_CAP_COALESCED_MMIO
 431     KVMState *s = kvm_state;
 432
 433     if (s->coalesced_mmio) {
 434         struct kvm_coalesced_mmio_zone zone;
 435
 436         zone.addr = start;
 437         zone.size = size;
 438
 439         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 440     }
 441 #endif
 442
 443     return ret;
 444 }
 445
 446 int kvm_check_extension(KVMState *s, unsigned int extension)
 447 {
 448     int ret;
 449
 450     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 451     if (ret < 0) {
 452         ret = 0;
 453     }
 454
 455     return ret;
 456 }
 457
 458 static int kvm_check_many_ioeventfds(void)
 459 {
 460     /* Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 461      * can avoid creating too many ioeventfds.
 462      */
 463 #ifdef CONFIG_EVENTFD
 464     int ioeventfds[7];
 465     int i, ret = 0;
 466     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 467         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 468         if (ioeventfds[i] < 0) {
 469             break;
 470         }
 471         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 472         if (ret < 0) {
 473             close(ioeventfds[i]);
 474             break;
 475         }
 476     }
 477
 478     /* Decide whether many devices are supported or not */
 479     ret = i == ARRAY_SIZE(ioeventfds);
 480
 481     while (i-- > 0) {
 482         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 483         close(ioeventfds[i]);
 484     }
 485     return ret;
 486 #else
 487     return 0;
 488 #endif
 489 }
 490
 491 static void kvm_set_phys_mem(target_phys_addr_t start_addr,
 492                              ram_addr_t size,
 493                              ram_addr_t phys_offset)
 494 {
 495     KVMState *s = kvm_state;
 496     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 497     KVMSlot *mem, old;
 498     int err;
 499
 500     /* kvm works in page size chunks, but the function may be called
 501        with sub-page size and unaligned start address. */
 502     size = TARGET_PAGE_ALIGN(size);
 503     start_addr = TARGET_PAGE_ALIGN(start_addr);
 504
 505     /* KVM does not support read-only slots */
 506     phys_offset &= ~IO_MEM_ROM;
 507
 508     while (1) {
 509         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 510         if (!mem) {
 511             break;
 512         }
 513
 514         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 515             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 516             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 517             /* The new slot fits into the existing one and comes with
 518              * identical parameters - nothing to be done. */
 519             return;
 520         }
 521
 522         old = *mem;
 523
 524         /* unregister the overlapping slot */
 525         mem->memory_size = 0;
 526         err = kvm_set_user_memory_region(s, mem);
 527         if (err) {
 528             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 529                     __func__, strerror(-err));
 530             abort();
 531         }
 532
 533         /* Workaround for older KVM versions: we can't join slots, even not by
 534          * unregistering the previous ones and then registering the larger
 535          * slot. We have to maintain the existing fragmentation. Sigh.
 536          *
 537          * This workaround assumes that the new slot starts at the same
 538          * address as the first existing one. If not or if some overlapping
 539          * slot comes around later, we will fail (not seen in practice so far)
 540          * - and actually require a recent KVM version. */
 541         if (s->broken_set_mem_region &&
 542             old.start_addr == start_addr && old.memory_size < size &&
 543             flags < IO_MEM_UNASSIGNED) {
 544             mem = kvm_alloc_slot(s);
 545             mem->memory_size = old.memory_size;
 546             mem->start_addr = old.start_addr;
 547             mem->phys_offset = old.phys_offset;
 548             mem->flags = 0;
 549
 550             err = kvm_set_user_memory_region(s, mem);
 551             if (err) {
 552                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 553                         strerror(-err));
 554                 abort();
 555             }
 556
 557             start_addr += old.memory_size;
 558             phys_offset += old.memory_size;
 559             size -= old.memory_size;
 560             continue;
 561         }
 562
 563         /* register prefix slot */
 564         if (old.start_addr < start_addr) {
 565             mem = kvm_alloc_slot(s);
 566             mem->memory_size = start_addr - old.start_addr;
 567             mem->start_addr = old.start_addr;
 568             mem->phys_offset = old.phys_offset;
 569             mem->flags = 0;
 570
 571             err = kvm_set_user_memory_region(s, mem);
 572             if (err) {
 573                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 574                         __func__, strerror(-err));
 575                 abort();
 576             }
 577         }
 578
 579         /* register suffix slot */
 580         if (old.start_addr + old.memory_size > start_addr + size) {
 581             ram_addr_t size_delta;
 582
 583             mem = kvm_alloc_slot(s);
 584             mem->start_addr = start_addr + size;
 585             size_delta = mem->start_addr - old.start_addr;
 586             mem->memory_size = old.memory_size - size_delta;
 587             mem->phys_offset = old.phys_offset + size_delta;
 588             mem->flags = 0;
 589
 590             err = kvm_set_user_memory_region(s, mem);
 591             if (err) {
 592                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 593                         __func__, strerror(-err));
 594                 abort();
 595             }
 596         }
 597     }
 598
 599     /* in case the KVM bug workaround already "consumed" the new slot */
 600     if (!size)
 601         return;
 602
 603     /* KVM does not need to know about this memory */
 604     if (flags >= IO_MEM_UNASSIGNED)
 605         return;
 606
 607     mem = kvm_alloc_slot(s);
 608     mem->memory_size = size;
 609     mem->start_addr = start_addr;
 610     mem->phys_offset = phys_offset;
 611     mem->flags = 0;
 612
 613     err = kvm_set_user_memory_region(s, mem);
 614     if (err) {
 615         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 616                 strerror(-err));
 617         abort();
 618     }
 619 }
 620
 621 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 622                                   target_phys_addr_t start_addr,
 623                                   ram_addr_t size,
 624                                   ram_addr_t phys_offset)
 625 {
 626         kvm_set_phys_mem(start_addr, size, phys_offset);
 627 }
 628
 629 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 630                                         target_phys_addr_t start_addr,
 631                                         target_phys_addr_t end_addr)
 632 {
 633         return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 634 }
 635
 636 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 637                                     int enable)
 638 {
 639         return kvm_set_migration_log(enable);
 640 }
 641
 642 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 643         .set_memory = kvm_client_set_memory,
 644         .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 645         .migration_log = kvm_client_migration_log,
 646 };
 647
 648
 649 void kvm_cpu_register_phys_memory_client(void)
 650 {
 651     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 652 }
 653
 654 #ifdef OBSOLETE_KVM_IMPL
 655
 656 int kvm_init(int smp_cpus)
 657 {
 658     static const char upgrade_note[] =
 659         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 660         "(see http://sourceforge.net/projects/kvm).\n";
 661     KVMState *s;
 662     int ret;
 663     int i;
 664
 665     s = qemu_mallocz(sizeof(KVMState));
 666
 667 #ifdef KVM_CAP_SET_GUEST_DEBUG
 668     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 669 #endif
 670     for (i = 0; i < ARRAY_SIZE(s->slots); i++)
 671         s->slots[i].slot = i;
 672
 673     s->vmfd = -1;
 674     s->fd = qemu_open("/dev/kvm", O_RDWR);
 675     if (s->fd == -1) {
 676         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 677         ret = -errno;
 678         goto err;
 679     }
 680
 681     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 682     if (ret < KVM_API_VERSION) {
 683         if (ret > 0)
 684             ret = -EINVAL;
 685         fprintf(stderr, "kvm version too old\n");
 686         goto err;
 687     }
 688
 689     if (ret > KVM_API_VERSION) {
 690         ret = -EINVAL;
 691         fprintf(stderr, "kvm version not supported\n");
 692         goto err;
 693     }
 694
 695     s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
 696     if (s->vmfd < 0) {
 697 #ifdef TARGET_S390X
 698         fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
 699                         "your host kernel command line\n");
 700 #endif
 701         goto err;
 702     }
 703
 704     /* initially, KVM allocated its own memory and we had to jump through
 705      * hooks to make phys_ram_base point to this.  Modern versions of KVM
 706      * just use a user allocated buffer so we can use regular pages
 707      * unmodified.  Make sure we have a sufficiently modern version of KVM.
 708      */
 709     if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
 710         ret = -EINVAL;
 711         fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
 712                 upgrade_note);
 713         goto err;
 714     }
 715
 716     /* There was a nasty bug in < kvm-80 that prevents memory slots from being
 717      * destroyed properly.  Since we rely on this capability, refuse to work
 718      * with any kernel without this capability. */
 719     if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
 720         ret = -EINVAL;
 721
 722         fprintf(stderr,
 723                 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
 724                 upgrade_note);
 725         goto err;
 726     }
 727
 728     s->coalesced_mmio = 0;
 729 #ifdef KVM_CAP_COALESCED_MMIO
 730     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 731     s->coalesced_mmio_ring = NULL;
 732 #endif
 733
 734     s->broken_set_mem_region = 1;
 735 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
 736     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
 737     if (ret > 0) {
 738         s->broken_set_mem_region = 0;
 739     }
 740 #endif
 741
 742     s->vcpu_events = 0;
 743 #ifdef KVM_CAP_VCPU_EVENTS
 744     s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
 745 #endif
 746
 747     s->robust_singlestep = 0;
 748 #ifdef KVM_CAP_X86_ROBUST_SINGLESTEP
 749     s->robust_singlestep =
 750         kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
 751 #endif
 752
 753     s->debugregs = 0;
 754 #ifdef KVM_CAP_DEBUGREGS
 755     s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
 756 #endif
 757
 758     s->xsave = 0;
 759 #ifdef KVM_CAP_XSAVE
 760     s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
 761 #endif
 762
 763     s->xcrs = 0;
 764 #ifdef KVM_CAP_XCRS
 765     s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
 766 #endif
 767
 768     ret = kvm_arch_init(s, smp_cpus);
 769     if (ret < 0)
 770         goto err;
 771
 772     kvm_state = s;
 773     cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
 774
 775     s->many_ioeventfds = kvm_check_many_ioeventfds();
 776
 777     return 0;
 778
 779 err:
 780     if (s) {
 781         if (s->vmfd != -1)
 782             close(s->vmfd);
 783         if (s->fd != -1)
 784             close(s->fd);
 785     }
 786     qemu_free(s);
 787
 788     return ret;
 789 }
 790 #endif
 791
 792 static int kvm_handle_io(uint16_t port, void *data, int direction, int size,
 793                          uint32_t count)
 794 {
 795     int i;
 796     uint8_t *ptr = data;
 797
 798     for (i = 0; i < count; i++) {
 799         if (direction == KVM_EXIT_IO_IN) {
 800             switch (size) {
 801             case 1:
 802                 stb_p(ptr, cpu_inb(port));
 803                 break;
 804             case 2:
 805                 stw_p(ptr, cpu_inw(port));
 806                 break;
 807             case 4:
 808                 stl_p(ptr, cpu_inl(port));
 809                 break;
 810             }
 811         } else {
 812             switch (size) {
 813             case 1:
 814                 cpu_outb(port, ldub_p(ptr));
 815                 break;
 816             case 2:
 817                 cpu_outw(port, lduw_p(ptr));
 818                 break;
 819             case 4:
 820                 cpu_outl(port, ldl_p(ptr));
 821                 break;
 822             }
 823         }
 824
 825         ptr += size;
 826     }
 827
 828     return 1;
 829 }
 830
 831 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 832 static void kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 833 {
 834
 835     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 836         int i;
 837
 838         fprintf(stderr, "KVM internal error. Suberror: %d\n",
 839                 run->internal.suberror);
 840
 841         for (i = 0; i < run->internal.ndata; ++i) {
 842             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 843                     i, (uint64_t)run->internal.data[i]);
 844         }
 845     }
 846     cpu_dump_state(env, stderr, fprintf, 0);
 847     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 848         fprintf(stderr, "emulation failure\n");
 849         if (!kvm_arch_stop_on_emulation_error(env))
 850                 return;
 851     }
 852     /* FIXME: Should trigger a qmp message to let management know
 853      * something went wrong.
 854      */
 855     vm_stop(0);
 856 }
 857 #endif
 858
 859 void kvm_flush_coalesced_mmio_buffer(void)
 860 {
 861 #ifdef KVM_CAP_COALESCED_MMIO
 862     KVMState *s = kvm_state;
 863     if (s->coalesced_mmio_ring) {
 864         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 865         while (ring->first != ring->last) {
 866             struct kvm_coalesced_mmio *ent;
 867
 868             ent = &ring->coalesced_mmio[ring->first];
 869
 870             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 871             smp_wmb();
 872             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 873         }
 874     }
 875 #endif
 876 }
 877
 878 #ifdef OBSOLETE_KVM_IMPL
 879
 880 static void do_kvm_cpu_synchronize_state(void *_env)
 881 {
 882     CPUState *env = _env;
 883
 884     if (!env->kvm_vcpu_dirty) {
 885         kvm_arch_get_registers(env);
 886         env->kvm_vcpu_dirty = 1;
 887     }
 888 }
 889
 890 void kvm_cpu_synchronize_state(CPUState *env)
 891 {
 892     if (!env->kvm_vcpu_dirty)
 893         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 894 }
 895
 896 void kvm_cpu_synchronize_post_reset(CPUState *env)
 897 {
 898     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 899     env->kvm_vcpu_dirty = 0;
 900 }
 901
 902 void kvm_cpu_synchronize_post_init(CPUState *env)
 903 {
 904     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 905     env->kvm_vcpu_dirty = 0;
 906 }
 907
 908 int kvm_cpu_exec(CPUState *env)
 909 {
 910     struct kvm_run *run = env->kvm_run;
 911     int ret;
 912
 913     DPRINTF("kvm_cpu_exec()\n");
 914
 915     do {
 916 #ifndef CONFIG_IOTHREAD
 917         if (env->exit_request) {
 918             DPRINTF("interrupt exit requested\n");
 919             ret = 0;
 920             break;
 921         }
 922 #endif
 923
 924         if (kvm_arch_process_irqchip_events(env)) {
 925             ret = 0;
 926             break;
 927         }
 928
 929         if (env->kvm_vcpu_dirty) {
 930             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 931             env->kvm_vcpu_dirty = 0;
 932         }
 933
 934         kvm_arch_pre_run(env, run);
 935         cpu_single_env = NULL;
 936         qemu_mutex_unlock_iothread();
 937         ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 938         qemu_mutex_lock_iothread();
 939         cpu_single_env = env;
 940         kvm_arch_post_run(env, run);
 941
 942         if (ret == -EINTR || ret == -EAGAIN) {
 943             cpu_exit(env);
 944             DPRINTF("io window exit\n");
 945             ret = 0;
 946             break;
 947         }
 948
 949         if (ret < 0) {
 950             DPRINTF("kvm run failed %s\n", strerror(-ret));
 951             abort();
 952         }
 953
 954         kvm_flush_coalesced_mmio_buffer();
 955
 956         ret = 0; /* exit loop */
 957         switch (run->exit_reason) {
 958         case KVM_EXIT_IO:
 959             DPRINTF("handle_io\n");
 960             ret = kvm_handle_io(run->io.port,
 961                                 (uint8_t *)run + run->io.data_offset,
 962                                 run->io.direction,
 963                                 run->io.size,
 964                                 run->io.count);
 965             break;
 966         case KVM_EXIT_MMIO:
 967             DPRINTF("handle_mmio\n");
 968             cpu_physical_memory_rw(run->mmio.phys_addr,
 969                                    run->mmio.data,
 970                                    run->mmio.len,
 971                                    run->mmio.is_write);
 972             ret = 1;
 973             break;
 974         case KVM_EXIT_IRQ_WINDOW_OPEN:
 975             DPRINTF("irq_window_open\n");
 976             break;
 977         case KVM_EXIT_SHUTDOWN:
 978             DPRINTF("shutdown\n");
 979             qemu_system_reset_request();
 980             ret = 1;
 981             break;
 982         case KVM_EXIT_UNKNOWN:
 983             DPRINTF("kvm_exit_unknown\n");
 984             break;
 985         case KVM_EXIT_FAIL_ENTRY:
 986             DPRINTF("kvm_exit_fail_entry\n");
 987             break;
 988         case KVM_EXIT_EXCEPTION:
 989             DPRINTF("kvm_exit_exception\n");
 990             break;
 991 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 992         case KVM_EXIT_INTERNAL_ERROR:
 993             kvm_handle_internal_error(env, run);
 994             break;
 995 #endif
 996         case KVM_EXIT_DEBUG:
 997             DPRINTF("kvm_exit_debug\n");
 998 #ifdef KVM_CAP_SET_GUEST_DEBUG
 999             if (kvm_arch_debug(&run->debug.arch)) {
1000                 env->exception_index = EXCP_DEBUG;
1001                 return 0;
1002             }
1003             /* re-enter, this exception was guest-internal */
1004             ret = 1;
1005 #endif /* KVM_CAP_SET_GUEST_DEBUG */
1006             break;
1007         default:
1008             DPRINTF("kvm_arch_handle_exit\n");
1009             ret = kvm_arch_handle_exit(env, run);
1010             break;
1011         }
1012     } while (ret > 0);
1013
1014     if (env->exit_request) {
1015         env->exit_request = 0;
1016         env->exception_index = EXCP_INTERRUPT;
1017     }
1018
1019     return ret;
1020 }
1021
1022 #endif
1023 int kvm_ioctl(KVMState *s, int type, ...)
1024 {
1025     int ret;
1026     void *arg;
1027     va_list ap;
1028
1029     va_start(ap, type);
1030     arg = va_arg(ap, void *);
1031     va_end(ap);
1032
1033     ret = ioctl(s->fd, type, arg);
1034     if (ret == -1)
1035         ret = -errno;
1036
1037     return ret;
1038 }
1039
1040 int kvm_vm_ioctl(KVMState *s, int type, ...)
1041 {
1042     int ret;
1043     void *arg;
1044     va_list ap;
1045
1046     va_start(ap, type);
1047     arg = va_arg(ap, void *);
1048     va_end(ap);
1049
1050     ret = ioctl(s->vmfd, type, arg);
1051     if (ret == -1)
1052         ret = -errno;
1053
1054     return ret;
1055 }
1056
1057 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1058 {
1059     int ret;
1060     void *arg;
1061     va_list ap;
1062
1063     va_start(ap, type);
1064     arg = va_arg(ap, void *);
1065     va_end(ap);
1066
1067     ret = ioctl(env->kvm_fd, type, arg);
1068     if (ret == -1)
1069         ret = -errno;
1070
1071     return ret;
1072 }
1073
1074 int kvm_has_sync_mmu(void)
1075 {
1076 #ifdef KVM_CAP_SYNC_MMU
1077     KVMState *s = kvm_state;
1078
1079     return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
1080 #else
1081     return 0;
1082 #endif
1083 }
1084
1085 int kvm_has_vcpu_events(void)
1086 {
1087     return kvm_state->vcpu_events;
1088 }
1089
1090 int kvm_has_robust_singlestep(void)
1091 {
1092     return kvm_state->robust_singlestep;
1093 }
1094
1095 int kvm_has_debugregs(void)
1096 {
1097     return kvm_state->debugregs;
1098 }
1099
1100 int kvm_has_xsave(void)
1101 {
1102     return kvm_state->xsave;
1103 }
1104
1105 int kvm_has_xcrs(void)
1106 {
1107     return kvm_state->xcrs;
1108 }
1109
1110 int kvm_has_many_ioeventfds(void)
1111 {
1112     if (!kvm_enabled()) {
1113         return 0;
1114     }
1115     return kvm_state->many_ioeventfds;
1116 }
1117
1118 void kvm_setup_guest_memory(void *start, size_t size)
1119 {
1120     if (!kvm_has_sync_mmu()) {
1121         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1122
1123         if (ret) {
1124             perror("qemu_madvise");
1125             fprintf(stderr,
1126                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1127             exit(1);
1128         }
1129     }
1130 }
1131
1132 #ifdef KVM_CAP_SET_GUEST_DEBUG
1133 #ifndef OBSOLETE_KVM_IMPL
1134 #define run_on_cpu on_vcpu
1135 #endif /* !OBSOLETE_KVM_IMPL */
1136
1137 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1138                                                  target_ulong pc)
1139 {
1140     struct kvm_sw_breakpoint *bp;
1141
1142     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1143         if (bp->pc == pc)
1144             return bp;
1145     }
1146     return NULL;
1147 }
1148
1149 int kvm_sw_breakpoints_active(CPUState *env)
1150 {
1151     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1152 }
1153
1154 struct kvm_set_guest_debug_data {
1155     struct kvm_guest_debug dbg;
1156     CPUState *env;
1157     int err;
1158 };
1159
1160 static void kvm_invoke_set_guest_debug(void *data)
1161 {
1162     struct kvm_set_guest_debug_data *dbg_data = data;
1163     CPUState *env = dbg_data->env;
1164
1165     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1166 }
1167
1168 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1169 {
1170     struct kvm_set_guest_debug_data data;
1171
1172     data.dbg.control = reinject_trap;
1173
1174     if (env->singlestep_enabled) {
1175         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1176     }
1177     kvm_arch_update_guest_debug(env, &data.dbg);
1178     data.env = env;
1179
1180     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1181     return data.err;
1182 }
1183
1184 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1185                           target_ulong len, int type)
1186 {
1187     struct kvm_sw_breakpoint *bp;
1188     CPUState *env;
1189     int err;
1190
1191     if (type == GDB_BREAKPOINT_SW) {
1192         bp = kvm_find_sw_breakpoint(current_env, addr);
1193         if (bp) {
1194             bp->use_count++;
1195             return 0;
1196         }
1197
1198         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1199         if (!bp)
1200             return -ENOMEM;
1201
1202         bp->pc = addr;
1203         bp->use_count = 1;
1204         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1205         if (err) {
1206             free(bp);
1207             return err;
1208         }
1209
1210         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1211                           bp, entry);
1212     } else {
1213         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1214         if (err)
1215             return err;
1216     }
1217
1218     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1219         err = kvm_update_guest_debug(env, 0);
1220         if (err)
1221             return err;
1222     }
1223     return 0;
1224 }
1225
1226 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1227                           target_ulong len, int type)
1228 {
1229     struct kvm_sw_breakpoint *bp;
1230     CPUState *env;
1231     int err;
1232
1233     if (type == GDB_BREAKPOINT_SW) {
1234         bp = kvm_find_sw_breakpoint(current_env, addr);
1235         if (!bp)
1236             return -ENOENT;
1237
1238         if (bp->use_count > 1) {
1239             bp->use_count--;
1240             return 0;
1241         }
1242
1243         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1244         if (err)
1245             return err;
1246
1247         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1248         qemu_free(bp);
1249     } else {
1250         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1251         if (err)
1252             return err;
1253     }
1254
1255     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1256         err = kvm_update_guest_debug(env, 0);
1257         if (err)
1258             return err;
1259     }
1260     return 0;
1261 }
1262
1263 void kvm_remove_all_breakpoints(CPUState *current_env)
1264 {
1265     struct kvm_sw_breakpoint *bp, *next;
1266     KVMState *s = current_env->kvm_state;
1267     CPUState *env;
1268
1269     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1270         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1271             /* Try harder to find a CPU that currently sees the breakpoint. */
1272             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1273                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1274                     break;
1275             }
1276         }
1277     }
1278     kvm_arch_remove_all_hw_breakpoints();
1279
1280     for (env = first_cpu; env != NULL; env = env->next_cpu)
1281         kvm_update_guest_debug(env, 0);
1282 }
1283
1284 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1285
1286 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1287 {
1288     return -EINVAL;
1289 }
1290
1291 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1292                           target_ulong len, int type)
1293 {
1294     return -EINVAL;
1295 }
1296
1297 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1298                           target_ulong len, int type)
1299 {
1300     return -EINVAL;
1301 }
1302
1303 void kvm_remove_all_breakpoints(CPUState *current_env)
1304 {
1305 }
1306 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1307
1308 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1309 {
1310     struct kvm_signal_mask *sigmask;
1311     int r;
1312
1313     if (!sigset)
1314         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1315
1316     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1317
1318     sigmask->len = 8;
1319     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1320     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1321     free(sigmask);
1322
1323     return r;
1324 }
1325
1326 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1327 {
1328 #ifdef KVM_IOEVENTFD
1329     int ret;
1330     struct kvm_ioeventfd iofd;
1331
1332     iofd.datamatch = val;
1333     iofd.addr = addr;
1334     iofd.len = 4;
1335     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1336     iofd.fd = fd;
1337
1338     if (!kvm_enabled()) {
1339         return -ENOSYS;
1340     }
1341
1342     if (!assign) {
1343         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1344     }
1345
1346     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1347
1348     if (ret < 0) {
1349         return -errno;
1350     }
1351
1352     return 0;
1353 #else
1354     return -ENOSYS;
1355 #endif
1356 }
1357
1358 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1359 {
1360 #ifdef KVM_IOEVENTFD
1361     struct kvm_ioeventfd kick = {
1362         .datamatch = val,
1363         .addr = addr,
1364         .len = 2,
1365         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1366         .fd = fd,
1367     };
1368     int r;
1369     if (!kvm_enabled())
1370         return -ENOSYS;
1371     if (!assign)
1372         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1373     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1374     if (r < 0)
1375         return r;
1376     return 0;
1377 #else
1378     return -ENOSYS;
1379 #endif
1380 }
1381
1382 #if defined(KVM_IRQFD)
1383 int kvm_set_irqfd(int gsi, int fd, bool assigned)
1384 {
1385     struct kvm_irqfd irqfd = {
1386         .fd = fd,
1387         .gsi = gsi,
1388         .flags = assigned ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1389     };
1390     int r;
1391     if (!kvm_enabled() || !kvm_irqchip_in_kernel())
1392         return -ENOSYS;
1393
1394     r = kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd);
1395     if (r < 0)
1396         return r;
1397     return 0;
1398 }
1399 #endif
1400
1401 #undef PAGE_SIZE
1402 #include "qemu-kvm.c"