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 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         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 int kvm_init_vcpu(CPUState *env)
 199 {
 200     KVMState *s = kvm_state;
 201     long mmap_size;
 202     int ret;
 203
 204     DPRINTF("kvm_init_vcpu\n");
 205
 206     ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
 207     if (ret < 0) {
 208         DPRINTF("kvm_create_vcpu failed\n");
 209         goto err;
 210     }
 211
 212     env->kvm_fd = ret;
 213     env->kvm_state = s;
 214     env->kvm_vcpu_dirty = 1;
 215
 216     mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
 217     if (mmap_size < 0) {
 218         ret = mmap_size;
 219         DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
 220         goto err;
 221     }
 222
 223     env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
 224                         env->kvm_fd, 0);
 225     if (env->kvm_run == MAP_FAILED) {
 226         ret = -errno;
 227         DPRINTF("mmap'ing vcpu state failed\n");
 228         goto err;
 229     }
 230
 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
 236     ret = kvm_arch_init_vcpu(env);
 237     if (ret == 0) {
 238         qemu_register_reset(kvm_reset_vcpu, env);
 239         kvm_arch_reset_vcpu(env);
 240     }
 241 err:
 242     return ret;
 243 }
 244
 245 /*
 246  * dirty pages logging control
 247  */
 248 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
 249                                       ram_addr_t size, int flags, int mask)
 250 {
 251     KVMState *s = kvm_state;
 252     KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
 253     int old_flags;
 254
 255     if (mem == NULL)  {
 256             fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
 257                     TARGET_FMT_plx "\n", __func__, phys_addr,
 258                     (target_phys_addr_t)(phys_addr + size - 1));
 259             return -EINVAL;
 260     }
 261
 262     old_flags = mem->flags;
 263
 264     flags = (mem->flags & ~mask) | flags;
 265     mem->flags = flags;
 266
 267     /* If nothing changed effectively, no need to issue ioctl */
 268     if (s->migration_log) {
 269         flags |= KVM_MEM_LOG_DIRTY_PAGES;
 270     }
 271     if (flags == old_flags) {
 272             return 0;
 273     }
 274
 275     return kvm_set_user_memory_region(s, mem);
 276 }
 277
 278 static int kvm_log_start(CPUPhysMemoryClient *client,
 279                          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 static int kvm_log_stop(CPUPhysMemoryClient *client,
 286                         target_phys_addr_t phys_addr, ram_addr_t size)
 287 {
 288     return kvm_dirty_pages_log_change(phys_addr, size, 0,
 289                                       KVM_MEM_LOG_DIRTY_PAGES);
 290 }
 291
 292 static int kvm_set_migration_log(int enable)
 293 {
 294     KVMState *s = kvm_state;
 295     KVMSlot *mem;
 296     int i, err;
 297
 298     s->migration_log = enable;
 299
 300     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 301         mem = &s->slots[i];
 302
 303         if (!mem->memory_size) {
 304             continue;
 305         }
 306         if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
 307             continue;
 308         }
 309         err = kvm_set_user_memory_region(s, mem);
 310         if (err) {
 311             return err;
 312         }
 313     }
 314     return 0;
 315 }
 316
 317 /* get kvm's dirty pages bitmap and update qemu's */
 318 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
 319                                          unsigned long *bitmap,
 320                                          unsigned long offset,
 321                                          unsigned long mem_size)
 322 {
 323     unsigned int i, j;
 324     unsigned long page_number, addr, addr1, c;
 325     ram_addr_t ram_addr;
 326     unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
 327         HOST_LONG_BITS;
 328
 329     /*
 330      * bitmap-traveling is faster than memory-traveling (for addr...)
 331      * especially when most of the memory is not dirty.
 332      */
 333     for (i = 0; i < len; i++) {
 334         if (bitmap[i] != 0) {
 335             c = leul_to_cpu(bitmap[i]);
 336             do {
 337                 j = ffsl(c) - 1;
 338                 c &= ~(1ul << j);
 339                 page_number = i * HOST_LONG_BITS + j;
 340                 addr1 = page_number * TARGET_PAGE_SIZE;
 341                 addr = offset + addr1;
 342                 ram_addr = cpu_get_physical_page_desc(addr);
 343                 cpu_physical_memory_set_dirty(ram_addr);
 344             } while (c != 0);
 345         }
 346     }
 347     return 0;
 348 }
 349
 350 #define ALIGN(x, y)  (((x)+(y)-1) & ~((y)-1))
 351
 352 /**
 353  * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
 354  * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
 355  * This means all bits are set to dirty.
 356  *
 357  * @start_add: start of logged region.
 358  * @end_addr: end of logged region.
 359  */
 360 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
 361                                           target_phys_addr_t end_addr)
 362 {
 363     KVMState *s = kvm_state;
 364     unsigned long size, allocated_size = 0;
 365     KVMDirtyLog d;
 366     KVMSlot *mem;
 367     int ret = 0;
 368
 369     d.dirty_bitmap = NULL;
 370     while (start_addr < end_addr) {
 371         mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
 372         if (mem == NULL) {
 373             break;
 374         }
 375
 376         size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS), HOST_LONG_BITS) / 8;
 377         if (!d.dirty_bitmap) {
 378             d.dirty_bitmap = qemu_malloc(size);
 379         } else if (size > allocated_size) {
 380             d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
 381         }
 382         allocated_size = size;
 383         memset(d.dirty_bitmap, 0, allocated_size);
 384
 385         d.slot = mem->slot;
 386
 387         if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
 388             DPRINTF("ioctl failed %d\n", errno);
 389             ret = -1;
 390             break;
 391         }
 392
 393         kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
 394                                       mem->start_addr, mem->memory_size);
 395         start_addr = mem->start_addr + mem->memory_size;
 396     }
 397     qemu_free(d.dirty_bitmap);
 398
 399     return ret;
 400 }
 401
 402 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 403 {
 404     int ret = -ENOSYS;
 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
 416     return ret;
 417 }
 418
 419 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
 420 {
 421     int ret = -ENOSYS;
 422     KVMState *s = kvm_state;
 423
 424     if (s->coalesced_mmio) {
 425         struct kvm_coalesced_mmio_zone zone;
 426
 427         zone.addr = start;
 428         zone.size = size;
 429
 430         ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
 431     }
 432
 433     return ret;
 434 }
 435
 436 int kvm_check_extension(KVMState *s, unsigned int extension)
 437 {
 438     int ret;
 439
 440     ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
 441     if (ret < 0) {
 442         ret = 0;
 443     }
 444
 445     return ret;
 446 }
 447
 448 static int kvm_check_many_ioeventfds(void)
 449 {
 450     /* Userspace can use ioeventfd for io notification.  This requires a host
 451      * that supports eventfd(2) and an I/O thread; since eventfd does not
 452      * support SIGIO it cannot interrupt the vcpu.
 453      *
 454      * Older kernels have a 6 device limit on the KVM io bus.  Find out so we
 455      * can avoid creating too many ioeventfds.
 456      */
 457 #if defined(CONFIG_EVENTFD) && defined(CONFIG_IOTHREAD)
 458     int ioeventfds[7];
 459     int i, ret = 0;
 460     for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
 461         ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
 462         if (ioeventfds[i] < 0) {
 463             break;
 464         }
 465         ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
 466         if (ret < 0) {
 467             close(ioeventfds[i]);
 468             break;
 469         }
 470     }
 471
 472     /* Decide whether many devices are supported or not */
 473     ret = i == ARRAY_SIZE(ioeventfds);
 474
 475     while (i-- > 0) {
 476         kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
 477         close(ioeventfds[i]);
 478     }
 479     return ret;
 480 #else
 481     return 0;
 482 #endif
 483 }
 484
 485 static const KVMCapabilityInfo *
 486 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
 487 {
 488     while (list->name) {
 489         if (!kvm_check_extension(s, list->value)) {
 490             return list;
 491         }
 492         list++;
 493     }
 494     return NULL;
 495 }
 496
 497 static void kvm_set_phys_mem(target_phys_addr_t start_addr, ram_addr_t size,
 498                              ram_addr_t phys_offset)
 499 {
 500     KVMState *s = kvm_state;
 501     ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
 502     KVMSlot *mem, old;
 503     int err;
 504
 505     /* kvm works in page size chunks, but the function may be called
 506        with sub-page size and unaligned start address. */
 507     size = TARGET_PAGE_ALIGN(size);
 508     start_addr = TARGET_PAGE_ALIGN(start_addr);
 509
 510     /* KVM does not support read-only slots */
 511     phys_offset &= ~IO_MEM_ROM;
 512
 513     while (1) {
 514         mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
 515         if (!mem) {
 516             break;
 517         }
 518
 519         if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
 520             (start_addr + size <= mem->start_addr + mem->memory_size) &&
 521             (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
 522             /* The new slot fits into the existing one and comes with
 523              * identical parameters - nothing to be done. */
 524             return;
 525         }
 526
 527         old = *mem;
 528
 529         /* unregister the overlapping slot */
 530         mem->memory_size = 0;
 531         err = kvm_set_user_memory_region(s, mem);
 532         if (err) {
 533             fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
 534                     __func__, strerror(-err));
 535             abort();
 536         }
 537
 538         /* Workaround for older KVM versions: we can't join slots, even not by
 539          * unregistering the previous ones and then registering the larger
 540          * slot. We have to maintain the existing fragmentation. Sigh.
 541          *
 542          * This workaround assumes that the new slot starts at the same
 543          * address as the first existing one. If not or if some overlapping
 544          * slot comes around later, we will fail (not seen in practice so far)
 545          * - and actually require a recent KVM version. */
 546         if (s->broken_set_mem_region &&
 547             old.start_addr == start_addr && old.memory_size < size &&
 548             flags < IO_MEM_UNASSIGNED) {
 549             mem = kvm_alloc_slot(s);
 550             mem->memory_size = old.memory_size;
 551             mem->start_addr = old.start_addr;
 552             mem->phys_offset = old.phys_offset;
 553             mem->flags = 0;
 554
 555             err = kvm_set_user_memory_region(s, mem);
 556             if (err) {
 557                 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
 558                         strerror(-err));
 559                 abort();
 560             }
 561
 562             start_addr += old.memory_size;
 563             phys_offset += old.memory_size;
 564             size -= old.memory_size;
 565             continue;
 566         }
 567
 568         /* register prefix slot */
 569         if (old.start_addr < start_addr) {
 570             mem = kvm_alloc_slot(s);
 571             mem->memory_size = start_addr - old.start_addr;
 572             mem->start_addr = old.start_addr;
 573             mem->phys_offset = old.phys_offset;
 574             mem->flags = 0;
 575
 576             err = kvm_set_user_memory_region(s, mem);
 577             if (err) {
 578                 fprintf(stderr, "%s: error registering prefix slot: %s\n",
 579                         __func__, strerror(-err));
 580                 abort();
 581             }
 582         }
 583
 584         /* register suffix slot */
 585         if (old.start_addr + old.memory_size > start_addr + size) {
 586             ram_addr_t size_delta;
 587
 588             mem = kvm_alloc_slot(s);
 589             mem->start_addr = start_addr + size;
 590             size_delta = mem->start_addr - old.start_addr;
 591             mem->memory_size = old.memory_size - size_delta;
 592             mem->phys_offset = old.phys_offset + size_delta;
 593             mem->flags = 0;
 594
 595             err = kvm_set_user_memory_region(s, mem);
 596             if (err) {
 597                 fprintf(stderr, "%s: error registering suffix slot: %s\n",
 598                         __func__, strerror(-err));
 599                 abort();
 600             }
 601         }
 602     }
 603
 604     /* in case the KVM bug workaround already "consumed" the new slot */
 605     if (!size) {
 606         return;
 607     }
 608     /* KVM does not need to know about this memory */
 609     if (flags >= IO_MEM_UNASSIGNED) {
 610         return;
 611     }
 612     mem = kvm_alloc_slot(s);
 613     mem->memory_size = size;
 614     mem->start_addr = start_addr;
 615     mem->phys_offset = phys_offset;
 616     mem->flags = 0;
 617
 618     err = kvm_set_user_memory_region(s, mem);
 619     if (err) {
 620         fprintf(stderr, "%s: error registering slot: %s\n", __func__,
 621                 strerror(-err));
 622         abort();
 623     }
 624 }
 625
 626 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
 627                                   target_phys_addr_t start_addr,
 628                                   ram_addr_t size, ram_addr_t phys_offset)
 629 {
 630     kvm_set_phys_mem(start_addr, size, phys_offset);
 631 }
 632
 633 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
 634                                         target_phys_addr_t start_addr,
 635                                         target_phys_addr_t end_addr)
 636 {
 637     return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
 638 }
 639
 640 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
 641                                     int enable)
 642 {
 643     return kvm_set_migration_log(enable);
 644 }
 645
 646 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
 647     .set_memory = kvm_client_set_memory,
 648     .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
 649     .migration_log = kvm_client_migration_log,
 650     .log_start = kvm_log_start,
 651     .log_stop = kvm_log_stop,
 652 };
 653
 654 int kvm_init(void)
 655 {
 656     static const char upgrade_note[] =
 657         "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
 658         "(see http://sourceforge.net/projects/kvm).\n";
 659     KVMState *s;
 660     const KVMCapabilityInfo *missing_cap;
 661     int ret;
 662     int i;
 663
 664     s = qemu_mallocz(sizeof(KVMState));
 665
 666 #ifdef KVM_CAP_SET_GUEST_DEBUG
 667     QTAILQ_INIT(&s->kvm_sw_breakpoints);
 668 #endif
 669     for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
 670         s->slots[i].slot = i;
 671     }
 672     s->vmfd = -1;
 673     s->fd = qemu_open("/dev/kvm", O_RDWR);
 674     if (s->fd == -1) {
 675         fprintf(stderr, "Could not access KVM kernel module: %m\n");
 676         ret = -errno;
 677         goto err;
 678     }
 679
 680     ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
 681     if (ret < KVM_API_VERSION) {
 682         if (ret > 0) {
 683             ret = -EINVAL;
 684         }
 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     missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
 705     if (!missing_cap) {
 706         missing_cap =
 707             kvm_check_extension_list(s, kvm_arch_required_capabilities);
 708     }
 709     if (missing_cap) {
 710         ret = -EINVAL;
 711         fprintf(stderr, "kvm does not support %s\n%s",
 712                 missing_cap->name, upgrade_note);
 713         goto err;
 714     }
 715
 716     s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
 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 void 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
 815 #ifdef KVM_CAP_INTERNAL_ERROR_DATA
 816 static int kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
 817 {
 818     fprintf(stderr, "KVM internal error.");
 819     if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
 820         int i;
 821
 822         fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
 823         for (i = 0; i < run->internal.ndata; ++i) {
 824             fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
 825                     i, (uint64_t)run->internal.data[i]);
 826         }
 827     } else {
 828         fprintf(stderr, "\n");
 829     }
 830     if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
 831         fprintf(stderr, "emulation failure\n");
 832         if (!kvm_arch_stop_on_emulation_error(env)) {
 833             cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 834             return EXCP_INTERRUPT;
 835         }
 836     }
 837     /* FIXME: Should trigger a qmp message to let management know
 838      * something went wrong.
 839      */
 840     return -1;
 841 }
 842 #endif
 843
 844 void kvm_flush_coalesced_mmio_buffer(void)
 845 {
 846     KVMState *s = kvm_state;
 847     if (s->coalesced_mmio_ring) {
 848         struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
 849         while (ring->first != ring->last) {
 850             struct kvm_coalesced_mmio *ent;
 851
 852             ent = &ring->coalesced_mmio[ring->first];
 853
 854             cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
 855             smp_wmb();
 856             ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
 857         }
 858     }
 859 }
 860
 861 static void do_kvm_cpu_synchronize_state(void *_env)
 862 {
 863     CPUState *env = _env;
 864
 865     if (!env->kvm_vcpu_dirty) {
 866         kvm_arch_get_registers(env);
 867         env->kvm_vcpu_dirty = 1;
 868     }
 869 }
 870
 871 void kvm_cpu_synchronize_state(CPUState *env)
 872 {
 873     if (!env->kvm_vcpu_dirty) {
 874         run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
 875     }
 876 }
 877
 878 void kvm_cpu_synchronize_post_reset(CPUState *env)
 879 {
 880     kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
 881     env->kvm_vcpu_dirty = 0;
 882 }
 883
 884 void kvm_cpu_synchronize_post_init(CPUState *env)
 885 {
 886     kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
 887     env->kvm_vcpu_dirty = 0;
 888 }
 889
 890 int kvm_cpu_exec(CPUState *env)
 891 {
 892     struct kvm_run *run = env->kvm_run;
 893     int ret, run_ret;
 894
 895     DPRINTF("kvm_cpu_exec()\n");
 896
 897     if (kvm_arch_process_async_events(env)) {
 898         env->exit_request = 0;
 899         return EXCP_HLT;
 900     }
 901
 902     cpu_single_env = env;
 903
 904     do {
 905         if (env->kvm_vcpu_dirty) {
 906             kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
 907             env->kvm_vcpu_dirty = 0;
 908         }
 909
 910         kvm_arch_pre_run(env, run);
 911         if (env->exit_request) {
 912             DPRINTF("interrupt exit requested\n");
 913             /*
 914              * KVM requires us to reenter the kernel after IO exits to complete
 915              * instruction emulation. This self-signal will ensure that we
 916              * leave ASAP again.
 917              */
 918             qemu_cpu_kick_self();
 919         }
 920         cpu_single_env = NULL;
 921         qemu_mutex_unlock_iothread();
 922
 923         run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
 924
 925         qemu_mutex_lock_iothread();
 926         cpu_single_env = env;
 927         kvm_arch_post_run(env, run);
 928
 929         kvm_flush_coalesced_mmio_buffer();
 930
 931         if (run_ret < 0) {
 932             if (run_ret == -EINTR || run_ret == -EAGAIN) {
 933                 DPRINTF("io window exit\n");
 934                 ret = EXCP_INTERRUPT;
 935                 break;
 936             }
 937             DPRINTF("kvm run failed %s\n", strerror(-run_ret));
 938             abort();
 939         }
 940
 941         switch (run->exit_reason) {
 942         case KVM_EXIT_IO:
 943             DPRINTF("handle_io\n");
 944             kvm_handle_io(run->io.port,
 945                           (uint8_t *)run + run->io.data_offset,
 946                           run->io.direction,
 947                           run->io.size,
 948                           run->io.count);
 949             ret = 0;
 950             break;
 951         case KVM_EXIT_MMIO:
 952             DPRINTF("handle_mmio\n");
 953             cpu_physical_memory_rw(run->mmio.phys_addr,
 954                                    run->mmio.data,
 955                                    run->mmio.len,
 956                                    run->mmio.is_write);
 957             ret = 0;
 958             break;
 959         case KVM_EXIT_IRQ_WINDOW_OPEN:
 960             DPRINTF("irq_window_open\n");
 961             ret = EXCP_INTERRUPT;
 962             break;
 963         case KVM_EXIT_SHUTDOWN:
 964             DPRINTF("shutdown\n");
 965             qemu_system_reset_request();
 966             ret = EXCP_INTERRUPT;
 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         default:
 979             DPRINTF("kvm_arch_handle_exit\n");
 980             ret = kvm_arch_handle_exit(env, run);
 981             break;
 982         }
 983     } while (ret == 0);
 984
 985     if (ret < 0) {
 986         cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
 987         vm_stop(VMSTOP_PANIC);
 988     }
 989
 990     env->exit_request = 0;
 991     cpu_single_env = NULL;
 992     return ret;
 993 }
 994
 995 int kvm_ioctl(KVMState *s, int type, ...)
 996 {
 997     int ret;
 998     void *arg;
 999     va_list ap;
1000
1001     va_start(ap, type);
1002     arg = va_arg(ap, void *);
1003     va_end(ap);
1004
1005     ret = ioctl(s->fd, type, arg);
1006     if (ret == -1) {
1007         ret = -errno;
1008     }
1009     return ret;
1010 }
1011
1012 int kvm_vm_ioctl(KVMState *s, int type, ...)
1013 {
1014     int ret;
1015     void *arg;
1016     va_list ap;
1017
1018     va_start(ap, type);
1019     arg = va_arg(ap, void *);
1020     va_end(ap);
1021
1022     ret = ioctl(s->vmfd, type, arg);
1023     if (ret == -1) {
1024         ret = -errno;
1025     }
1026     return ret;
1027 }
1028
1029 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1030 {
1031     int ret;
1032     void *arg;
1033     va_list ap;
1034
1035     va_start(ap, type);
1036     arg = va_arg(ap, void *);
1037     va_end(ap);
1038
1039     ret = ioctl(env->kvm_fd, type, arg);
1040     if (ret == -1) {
1041         ret = -errno;
1042     }
1043     return ret;
1044 }
1045
1046 int kvm_has_sync_mmu(void)
1047 {
1048     return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1049 }
1050
1051 int kvm_has_vcpu_events(void)
1052 {
1053     return kvm_state->vcpu_events;
1054 }
1055
1056 int kvm_has_robust_singlestep(void)
1057 {
1058     return kvm_state->robust_singlestep;
1059 }
1060
1061 int kvm_has_debugregs(void)
1062 {
1063     return kvm_state->debugregs;
1064 }
1065
1066 int kvm_has_xsave(void)
1067 {
1068     return kvm_state->xsave;
1069 }
1070
1071 int kvm_has_xcrs(void)
1072 {
1073     return kvm_state->xcrs;
1074 }
1075
1076 int kvm_has_many_ioeventfds(void)
1077 {
1078     if (!kvm_enabled()) {
1079         return 0;
1080     }
1081     return kvm_state->many_ioeventfds;
1082 }
1083
1084 void kvm_setup_guest_memory(void *start, size_t size)
1085 {
1086     if (!kvm_has_sync_mmu()) {
1087         int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1088
1089         if (ret) {
1090             perror("qemu_madvise");
1091             fprintf(stderr,
1092                     "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1093             exit(1);
1094         }
1095     }
1096 }
1097
1098 #ifdef KVM_CAP_SET_GUEST_DEBUG
1099 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1100                                                  target_ulong pc)
1101 {
1102     struct kvm_sw_breakpoint *bp;
1103
1104     QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1105         if (bp->pc == pc) {
1106             return bp;
1107         }
1108     }
1109     return NULL;
1110 }
1111
1112 int kvm_sw_breakpoints_active(CPUState *env)
1113 {
1114     return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1115 }
1116
1117 struct kvm_set_guest_debug_data {
1118     struct kvm_guest_debug dbg;
1119     CPUState *env;
1120     int err;
1121 };
1122
1123 static void kvm_invoke_set_guest_debug(void *data)
1124 {
1125     struct kvm_set_guest_debug_data *dbg_data = data;
1126     CPUState *env = dbg_data->env;
1127
1128     dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1129 }
1130
1131 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1132 {
1133     struct kvm_set_guest_debug_data data;
1134
1135     data.dbg.control = reinject_trap;
1136
1137     if (env->singlestep_enabled) {
1138         data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1139     }
1140     kvm_arch_update_guest_debug(env, &data.dbg);
1141     data.env = env;
1142
1143     run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1144     return data.err;
1145 }
1146
1147 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1148                           target_ulong len, int type)
1149 {
1150     struct kvm_sw_breakpoint *bp;
1151     CPUState *env;
1152     int err;
1153
1154     if (type == GDB_BREAKPOINT_SW) {
1155         bp = kvm_find_sw_breakpoint(current_env, addr);
1156         if (bp) {
1157             bp->use_count++;
1158             return 0;
1159         }
1160
1161         bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1162         if (!bp) {
1163             return -ENOMEM;
1164         }
1165
1166         bp->pc = addr;
1167         bp->use_count = 1;
1168         err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1169         if (err) {
1170             free(bp);
1171             return err;
1172         }
1173
1174         QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1175                           bp, entry);
1176     } else {
1177         err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1178         if (err) {
1179             return err;
1180         }
1181     }
1182
1183     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1184         err = kvm_update_guest_debug(env, 0);
1185         if (err) {
1186             return err;
1187         }
1188     }
1189     return 0;
1190 }
1191
1192 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1193                           target_ulong len, int type)
1194 {
1195     struct kvm_sw_breakpoint *bp;
1196     CPUState *env;
1197     int err;
1198
1199     if (type == GDB_BREAKPOINT_SW) {
1200         bp = kvm_find_sw_breakpoint(current_env, addr);
1201         if (!bp) {
1202             return -ENOENT;
1203         }
1204
1205         if (bp->use_count > 1) {
1206             bp->use_count--;
1207             return 0;
1208         }
1209
1210         err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1211         if (err) {
1212             return err;
1213         }
1214
1215         QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1216         qemu_free(bp);
1217     } else {
1218         err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1219         if (err) {
1220             return err;
1221         }
1222     }
1223
1224     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1225         err = kvm_update_guest_debug(env, 0);
1226         if (err) {
1227             return err;
1228         }
1229     }
1230     return 0;
1231 }
1232
1233 void kvm_remove_all_breakpoints(CPUState *current_env)
1234 {
1235     struct kvm_sw_breakpoint *bp, *next;
1236     KVMState *s = current_env->kvm_state;
1237     CPUState *env;
1238
1239     QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1240         if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1241             /* Try harder to find a CPU that currently sees the breakpoint. */
1242             for (env = first_cpu; env != NULL; env = env->next_cpu) {
1243                 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1244                     break;
1245                 }
1246             }
1247         }
1248     }
1249     kvm_arch_remove_all_hw_breakpoints();
1250
1251     for (env = first_cpu; env != NULL; env = env->next_cpu) {
1252         kvm_update_guest_debug(env, 0);
1253     }
1254 }
1255
1256 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1257
1258 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1259 {
1260     return -EINVAL;
1261 }
1262
1263 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1264                           target_ulong len, int type)
1265 {
1266     return -EINVAL;
1267 }
1268
1269 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1270                           target_ulong len, int type)
1271 {
1272     return -EINVAL;
1273 }
1274
1275 void kvm_remove_all_breakpoints(CPUState *current_env)
1276 {
1277 }
1278 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1279
1280 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1281 {
1282     struct kvm_signal_mask *sigmask;
1283     int r;
1284
1285     if (!sigset) {
1286         return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1287     }
1288
1289     sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1290
1291     sigmask->len = 8;
1292     memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1293     r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1294     free(sigmask);
1295
1296     return r;
1297 }
1298
1299 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1300 {
1301 #ifdef KVM_IOEVENTFD
1302     int ret;
1303     struct kvm_ioeventfd iofd;
1304
1305     iofd.datamatch = val;
1306     iofd.addr = addr;
1307     iofd.len = 4;
1308     iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1309     iofd.fd = fd;
1310
1311     if (!kvm_enabled()) {
1312         return -ENOSYS;
1313     }
1314
1315     if (!assign) {
1316         iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1317     }
1318
1319     ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1320
1321     if (ret < 0) {
1322         return -errno;
1323     }
1324
1325     return 0;
1326 #else
1327     return -ENOSYS;
1328 #endif
1329 }
1330
1331 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1332 {
1333 #ifdef KVM_IOEVENTFD
1334     struct kvm_ioeventfd kick = {
1335         .datamatch = val,
1336         .addr = addr,
1337         .len = 2,
1338         .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1339         .fd = fd,
1340     };
1341     int r;
1342     if (!kvm_enabled()) {
1343         return -ENOSYS;
1344     }
1345     if (!assign) {
1346         kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1347     }
1348     r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1349     if (r < 0) {
1350         return r;
1351     }
1352     return 0;
1353 #else
1354     return -ENOSYS;
1355 #endif
1356 }
1357
1358 int kvm_on_sigbus_vcpu(CPUState *env, int code, void *addr)
1359 {
1360     return kvm_arch_on_sigbus_vcpu(env, code, addr);
1361 }
1362
1363 int kvm_on_sigbus(int code, void *addr)
1364 {
1365     return kvm_arch_on_sigbus(code, addr);
1366 }