hw/vhost.c

   1 /*
   2  * vhost support
   3  *
   4  * Copyright Red Hat, Inc. 2010
   5  *
   6  * Authors:
   7  *  Michael S. Tsirkin <mst@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include <sys/ioctl.h>
  17 #include "vhost.h"
  18 #include "hw/hw.h"
  19 #include "qemu/range.h"
  20 #include <linux/vhost.h>
  21 #include "exec/address-spaces.h"
  22
  23 static void vhost_dev_sync_region(struct vhost_dev *dev,
  24                                   MemoryRegionSection *section,
  25                                   uint64_t mfirst, uint64_t mlast,
  26                                   uint64_t rfirst, uint64_t rlast)
  27 {
  28     uint64_t start = MAX(mfirst, rfirst);
  29     uint64_t end = MIN(mlast, rlast);
  30     vhost_log_chunk_t *from = dev->log + start / VHOST_LOG_CHUNK;
  31     vhost_log_chunk_t *to = dev->log + end / VHOST_LOG_CHUNK + 1;
  32     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
  33
  34     if (end < start) {
  35         return;
  36     }
  37     assert(end / VHOST_LOG_CHUNK < dev->log_size);
  38     assert(start / VHOST_LOG_CHUNK < dev->log_size);
  39
  40     for (;from < to; ++from) {
  41         vhost_log_chunk_t log;
  42         int bit;
  43         /* We first check with non-atomic: much cheaper,
  44          * and we expect non-dirty to be the common case. */
  45         if (!*from) {
  46             addr += VHOST_LOG_CHUNK;
  47             continue;
  48         }
  49         /* Data must be read atomically. We don't really
  50          * need the barrier semantics of __sync
  51          * builtins, but it's easier to use them than
  52          * roll our own. */
  53         log = __sync_fetch_and_and(from, 0);
  54         while ((bit = sizeof(log) > sizeof(int) ?
  55                 ffsll(log) : ffs(log))) {
  56             ram_addr_t ram_addr;
  57             bit -= 1;
  58             ram_addr = section->offset_within_region + bit * VHOST_LOG_PAGE;
  59             memory_region_set_dirty(section->mr, ram_addr, VHOST_LOG_PAGE);
  60             log &= ~(0x1ull << bit);
  61         }
  62         addr += VHOST_LOG_CHUNK;
  63     }
  64 }
  65
  66 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
  67                                    MemoryRegionSection *section,
  68                                    hwaddr start_addr,
  69                                    hwaddr end_addr)
  70 {
  71     int i;
  72
  73     if (!dev->log_enabled || !dev->started) {
  74         return 0;
  75     }
  76     for (i = 0; i < dev->mem->nregions; ++i) {
  77         struct vhost_memory_region *reg = dev->mem->regions + i;
  78         vhost_dev_sync_region(dev, section, start_addr, end_addr,
  79                               reg->guest_phys_addr,
  80                               range_get_last(reg->guest_phys_addr,
  81                                              reg->memory_size));
  82     }
  83     for (i = 0; i < dev->nvqs; ++i) {
  84         struct vhost_virtqueue *vq = dev->vqs + i;
  85         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
  86                               range_get_last(vq->used_phys, vq->used_size));
  87     }
  88     return 0;
  89 }
  90
  91 static void vhost_log_sync(MemoryListener *listener,
  92                           MemoryRegionSection *section)
  93 {
  94     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
  95                                          memory_listener);
  96     hwaddr start_addr = section->offset_within_address_space;
  97     hwaddr end_addr = start_addr + section->size;
  98
  99     vhost_sync_dirty_bitmap(dev, section, start_addr, end_addr);
 100 }
 101
 102 /* Assign/unassign. Keep an unsorted array of non-overlapping
 103  * memory regions in dev->mem. */
 104 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
 105                                       uint64_t start_addr,
 106                                       uint64_t size)
 107 {
 108     int from, to, n = dev->mem->nregions;
 109     /* Track overlapping/split regions for sanity checking. */
 110     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
 111
 112     for (from = 0, to = 0; from < n; ++from, ++to) {
 113         struct vhost_memory_region *reg = dev->mem->regions + to;
 114         uint64_t reglast;
 115         uint64_t memlast;
 116         uint64_t change;
 117
 118         /* clone old region */
 119         if (to != from) {
 120             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 121         }
 122
 123         /* No overlap is simple */
 124         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 125                             start_addr, size)) {
 126             continue;
 127         }
 128
 129         /* Split only happens if supplied region
 130          * is in the middle of an existing one. Thus it can not
 131          * overlap with any other existing region. */
 132         assert(!split);
 133
 134         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 135         memlast = range_get_last(start_addr, size);
 136
 137         /* Remove whole region */
 138         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
 139             --dev->mem->nregions;
 140             --to;
 141             ++overlap_middle;
 142             continue;
 143         }
 144
 145         /* Shrink region */
 146         if (memlast >= reglast) {
 147             reg->memory_size = start_addr - reg->guest_phys_addr;
 148             assert(reg->memory_size);
 149             assert(!overlap_end);
 150             ++overlap_end;
 151             continue;
 152         }
 153
 154         /* Shift region */
 155         if (start_addr <= reg->guest_phys_addr) {
 156             change = memlast + 1 - reg->guest_phys_addr;
 157             reg->memory_size -= change;
 158             reg->guest_phys_addr += change;
 159             reg->userspace_addr += change;
 160             assert(reg->memory_size);
 161             assert(!overlap_start);
 162             ++overlap_start;
 163             continue;
 164         }
 165
 166         /* This only happens if supplied region
 167          * is in the middle of an existing one. Thus it can not
 168          * overlap with any other existing region. */
 169         assert(!overlap_start);
 170         assert(!overlap_end);
 171         assert(!overlap_middle);
 172         /* Split region: shrink first part, shift second part. */
 173         memcpy(dev->mem->regions + n, reg, sizeof *reg);
 174         reg->memory_size = start_addr - reg->guest_phys_addr;
 175         assert(reg->memory_size);
 176         change = memlast + 1 - reg->guest_phys_addr;
 177         reg = dev->mem->regions + n;
 178         reg->memory_size -= change;
 179         assert(reg->memory_size);
 180         reg->guest_phys_addr += change;
 181         reg->userspace_addr += change;
 182         /* Never add more than 1 region */
 183         assert(dev->mem->nregions == n);
 184         ++dev->mem->nregions;
 185         ++split;
 186     }
 187 }
 188
 189 /* Called after unassign, so no regions overlap the given range. */
 190 static void vhost_dev_assign_memory(struct vhost_dev *dev,
 191                                     uint64_t start_addr,
 192                                     uint64_t size,
 193                                     uint64_t uaddr)
 194 {
 195     int from, to;
 196     struct vhost_memory_region *merged = NULL;
 197     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
 198         struct vhost_memory_region *reg = dev->mem->regions + to;
 199         uint64_t prlast, urlast;
 200         uint64_t pmlast, umlast;
 201         uint64_t s, e, u;
 202
 203         /* clone old region */
 204         if (to != from) {
 205             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 206         }
 207         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 208         pmlast = range_get_last(start_addr, size);
 209         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
 210         umlast = range_get_last(uaddr, size);
 211
 212         /* check for overlapping regions: should never happen. */
 213         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
 214         /* Not an adjacent or overlapping region - do not merge. */
 215         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
 216             (pmlast + 1 != reg->guest_phys_addr ||
 217              umlast + 1 != reg->userspace_addr)) {
 218             continue;
 219         }
 220
 221         if (merged) {
 222             --to;
 223             assert(to >= 0);
 224         } else {
 225             merged = reg;
 226         }
 227         u = MIN(uaddr, reg->userspace_addr);
 228         s = MIN(start_addr, reg->guest_phys_addr);
 229         e = MAX(pmlast, prlast);
 230         uaddr = merged->userspace_addr = u;
 231         start_addr = merged->guest_phys_addr = s;
 232         size = merged->memory_size = e - s + 1;
 233         assert(merged->memory_size);
 234     }
 235
 236     if (!merged) {
 237         struct vhost_memory_region *reg = dev->mem->regions + to;
 238         memset(reg, 0, sizeof *reg);
 239         reg->memory_size = size;
 240         assert(reg->memory_size);
 241         reg->guest_phys_addr = start_addr;
 242         reg->userspace_addr = uaddr;
 243         ++to;
 244     }
 245     assert(to <= dev->mem->nregions + 1);
 246     dev->mem->nregions = to;
 247 }
 248
 249 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
 250 {
 251     uint64_t log_size = 0;
 252     int i;
 253     for (i = 0; i < dev->mem->nregions; ++i) {
 254         struct vhost_memory_region *reg = dev->mem->regions + i;
 255         uint64_t last = range_get_last(reg->guest_phys_addr,
 256                                        reg->memory_size);
 257         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 258     }
 259     for (i = 0; i < dev->nvqs; ++i) {
 260         struct vhost_virtqueue *vq = dev->vqs + i;
 261         uint64_t last = vq->used_phys + vq->used_size - 1;
 262         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 263     }
 264     return log_size;
 265 }
 266
 267 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
 268 {
 269     vhost_log_chunk_t *log;
 270     uint64_t log_base;
 271     int r, i;
 272     if (size) {
 273         log = g_malloc0(size * sizeof *log);
 274     } else {
 275         log = NULL;
 276     }
 277     log_base = (uint64_t)(unsigned long)log;
 278     r = ioctl(dev->control, VHOST_SET_LOG_BASE, &log_base);
 279     assert(r >= 0);
 280     for (i = 0; i < dev->n_mem_sections; ++i) {
 281         /* Sync only the range covered by the old log */
 282         vhost_sync_dirty_bitmap(dev, &dev->mem_sections[i], 0,
 283                                 dev->log_size * VHOST_LOG_CHUNK - 1);
 284     }
 285     if (dev->log) {
 286         g_free(dev->log);
 287     }
 288     dev->log = log;
 289     dev->log_size = size;
 290 }
 291
 292 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
 293                                       uint64_t start_addr,
 294                                       uint64_t size)
 295 {
 296     int i;
 297     for (i = 0; i < dev->nvqs; ++i) {
 298         struct vhost_virtqueue *vq = dev->vqs + i;
 299         hwaddr l;
 300         void *p;
 301
 302         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
 303             continue;
 304         }
 305         l = vq->ring_size;
 306         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
 307         if (!p || l != vq->ring_size) {
 308             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
 309             return -ENOMEM;
 310         }
 311         if (p != vq->ring) {
 312             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
 313             return -EBUSY;
 314         }
 315         cpu_physical_memory_unmap(p, l, 0, 0);
 316     }
 317     return 0;
 318 }
 319
 320 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
 321                                                       uint64_t start_addr,
 322                                                       uint64_t size)
 323 {
 324     int i, n = dev->mem->nregions;
 325     for (i = 0; i < n; ++i) {
 326         struct vhost_memory_region *reg = dev->mem->regions + i;
 327         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 328                            start_addr, size)) {
 329             return reg;
 330         }
 331     }
 332     return NULL;
 333 }
 334
 335 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
 336                                  uint64_t start_addr,
 337                                  uint64_t size,
 338                                  uint64_t uaddr)
 339 {
 340     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
 341     uint64_t reglast;
 342     uint64_t memlast;
 343
 344     if (!reg) {
 345         return true;
 346     }
 347
 348     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 349     memlast = range_get_last(start_addr, size);
 350
 351     /* Need to extend region? */
 352     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
 353         return true;
 354     }
 355     /* userspace_addr changed? */
 356     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
 357 }
 358
 359 static void vhost_set_memory(MemoryListener *listener,
 360                              MemoryRegionSection *section,
 361                              bool add)
 362 {
 363     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 364                                          memory_listener);
 365     hwaddr start_addr = section->offset_within_address_space;
 366     ram_addr_t size = section->size;
 367     bool log_dirty = memory_region_is_logging(section->mr);
 368     int s = offsetof(struct vhost_memory, regions) +
 369         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
 370     uint64_t log_size;
 371     int r;
 372     void *ram;
 373
 374     dev->mem = g_realloc(dev->mem, s);
 375
 376     if (log_dirty) {
 377         add = false;
 378     }
 379
 380     assert(size);
 381
 382     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
 383     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
 384     if (add) {
 385         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
 386             /* Region exists with same address. Nothing to do. */
 387             return;
 388         }
 389     } else {
 390         if (!vhost_dev_find_reg(dev, start_addr, size)) {
 391             /* Removing region that we don't access. Nothing to do. */
 392             return;
 393         }
 394     }
 395
 396     vhost_dev_unassign_memory(dev, start_addr, size);
 397     if (add) {
 398         /* Add given mapping, merging adjacent regions if any */
 399         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
 400     } else {
 401         /* Remove old mapping for this memory, if any. */
 402         vhost_dev_unassign_memory(dev, start_addr, size);
 403     }
 404
 405     if (!dev->started) {
 406         return;
 407     }
 408
 409     if (dev->started) {
 410         r = vhost_verify_ring_mappings(dev, start_addr, size);
 411         assert(r >= 0);
 412     }
 413
 414     if (!dev->log_enabled) {
 415         r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 416         assert(r >= 0);
 417         return;
 418     }
 419     log_size = vhost_get_log_size(dev);
 420     /* We allocate an extra 4K bytes to log,
 421      * to reduce the * number of reallocations. */
 422 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
 423     /* To log more, must increase log size before table update. */
 424     if (dev->log_size < log_size) {
 425         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
 426     }
 427     r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 428     assert(r >= 0);
 429     /* To log less, can only decrease log size after table update. */
 430     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
 431         vhost_dev_log_resize(dev, log_size);
 432     }
 433 }
 434
 435 static bool vhost_section(MemoryRegionSection *section)
 436 {
 437     return memory_region_is_ram(section->mr);
 438 }
 439
 440 static void vhost_begin(MemoryListener *listener)
 441 {
 442 }
 443
 444 static void vhost_commit(MemoryListener *listener)
 445 {
 446 }
 447
 448 static void vhost_region_add(MemoryListener *listener,
 449                              MemoryRegionSection *section)
 450 {
 451     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 452                                          memory_listener);
 453
 454     if (!vhost_section(section)) {
 455         return;
 456     }
 457
 458     ++dev->n_mem_sections;
 459     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
 460                                 dev->n_mem_sections);
 461     dev->mem_sections[dev->n_mem_sections - 1] = *section;
 462     vhost_set_memory(listener, section, true);
 463 }
 464
 465 static void vhost_region_del(MemoryListener *listener,
 466                              MemoryRegionSection *section)
 467 {
 468     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 469                                          memory_listener);
 470     int i;
 471
 472     if (!vhost_section(section)) {
 473         return;
 474     }
 475
 476     vhost_set_memory(listener, section, false);
 477     for (i = 0; i < dev->n_mem_sections; ++i) {
 478         if (dev->mem_sections[i].offset_within_address_space
 479             == section->offset_within_address_space) {
 480             --dev->n_mem_sections;
 481             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
 482                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
 483             break;
 484         }
 485     }
 486 }
 487
 488 static void vhost_region_nop(MemoryListener *listener,
 489                              MemoryRegionSection *section)
 490 {
 491 }
 492
 493 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
 494                                     struct vhost_virtqueue *vq,
 495                                     unsigned idx, bool enable_log)
 496 {
 497     struct vhost_vring_addr addr = {
 498         .index = idx,
 499         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
 500         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
 501         .used_user_addr = (uint64_t)(unsigned long)vq->used,
 502         .log_guest_addr = vq->used_phys,
 503         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
 504     };
 505     int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
 506     if (r < 0) {
 507         return -errno;
 508     }
 509     return 0;
 510 }
 511
 512 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
 513 {
 514     uint64_t features = dev->acked_features;
 515     int r;
 516     if (enable_log) {
 517         features |= 0x1 << VHOST_F_LOG_ALL;
 518     }
 519     r = ioctl(dev->control, VHOST_SET_FEATURES, &features);
 520     return r < 0 ? -errno : 0;
 521 }
 522
 523 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
 524 {
 525     int r, t, i;
 526     r = vhost_dev_set_features(dev, enable_log);
 527     if (r < 0) {
 528         goto err_features;
 529     }
 530     for (i = 0; i < dev->nvqs; ++i) {
 531         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 532                                      enable_log);
 533         if (r < 0) {
 534             goto err_vq;
 535         }
 536     }
 537     return 0;
 538 err_vq:
 539     for (; i >= 0; --i) {
 540         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 541                                      dev->log_enabled);
 542         assert(t >= 0);
 543     }
 544     t = vhost_dev_set_features(dev, dev->log_enabled);
 545     assert(t >= 0);
 546 err_features:
 547     return r;
 548 }
 549
 550 static int vhost_migration_log(MemoryListener *listener, int enable)
 551 {
 552     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 553                                          memory_listener);
 554     int r;
 555     if (!!enable == dev->log_enabled) {
 556         return 0;
 557     }
 558     if (!dev->started) {
 559         dev->log_enabled = enable;
 560         return 0;
 561     }
 562     if (!enable) {
 563         r = vhost_dev_set_log(dev, false);
 564         if (r < 0) {
 565             return r;
 566         }
 567         if (dev->log) {
 568             g_free(dev->log);
 569         }
 570         dev->log = NULL;
 571         dev->log_size = 0;
 572     } else {
 573         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
 574         r = vhost_dev_set_log(dev, true);
 575         if (r < 0) {
 576             return r;
 577         }
 578     }
 579     dev->log_enabled = enable;
 580     return 0;
 581 }
 582
 583 static void vhost_log_global_start(MemoryListener *listener)
 584 {
 585     int r;
 586
 587     r = vhost_migration_log(listener, true);
 588     if (r < 0) {
 589         abort();
 590     }
 591 }
 592
 593 static void vhost_log_global_stop(MemoryListener *listener)
 594 {
 595     int r;
 596
 597     r = vhost_migration_log(listener, false);
 598     if (r < 0) {
 599         abort();
 600     }
 601 }
 602
 603 static void vhost_log_start(MemoryListener *listener,
 604                             MemoryRegionSection *section)
 605 {
 606     /* FIXME: implement */
 607 }
 608
 609 static void vhost_log_stop(MemoryListener *listener,
 610                            MemoryRegionSection *section)
 611 {
 612     /* FIXME: implement */
 613 }
 614
 615 static int vhost_virtqueue_start(struct vhost_dev *dev,
 616                                 struct VirtIODevice *vdev,
 617                                 struct vhost_virtqueue *vq,
 618                                 unsigned idx)
 619 {
 620     hwaddr s, l, a;
 621     int r;
 622     struct vhost_vring_file file = {
 623         .index = idx,
 624     };
 625     struct vhost_vring_state state = {
 626         .index = idx,
 627     };
 628     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
 629
 630     vq->num = state.num = virtio_queue_get_num(vdev, idx);
 631     r = ioctl(dev->control, VHOST_SET_VRING_NUM, &state);
 632     if (r) {
 633         return -errno;
 634     }
 635
 636     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
 637     r = ioctl(dev->control, VHOST_SET_VRING_BASE, &state);
 638     if (r) {
 639         return -errno;
 640     }
 641
 642     s = l = virtio_queue_get_desc_size(vdev, idx);
 643     a = virtio_queue_get_desc_addr(vdev, idx);
 644     vq->desc = cpu_physical_memory_map(a, &l, 0);
 645     if (!vq->desc || l != s) {
 646         r = -ENOMEM;
 647         goto fail_alloc_desc;
 648     }
 649     s = l = virtio_queue_get_avail_size(vdev, idx);
 650     a = virtio_queue_get_avail_addr(vdev, idx);
 651     vq->avail = cpu_physical_memory_map(a, &l, 0);
 652     if (!vq->avail || l != s) {
 653         r = -ENOMEM;
 654         goto fail_alloc_avail;
 655     }
 656     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
 657     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
 658     vq->used = cpu_physical_memory_map(a, &l, 1);
 659     if (!vq->used || l != s) {
 660         r = -ENOMEM;
 661         goto fail_alloc_used;
 662     }
 663
 664     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
 665     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
 666     vq->ring = cpu_physical_memory_map(a, &l, 1);
 667     if (!vq->ring || l != s) {
 668         r = -ENOMEM;
 669         goto fail_alloc_ring;
 670     }
 671
 672     r = vhost_virtqueue_set_addr(dev, vq, idx, dev->log_enabled);
 673     if (r < 0) {
 674         r = -errno;
 675         goto fail_alloc;
 676     }
 677     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
 678     r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file);
 679     if (r) {
 680         r = -errno;
 681         goto fail_kick;
 682     }
 683
 684     /* Clear and discard previous events if any. */
 685     event_notifier_test_and_clear(&vq->masked_notifier);
 686
 687     return 0;
 688
 689 fail_kick:
 690 fail_alloc:
 691     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 692                               0, 0);
 693 fail_alloc_ring:
 694     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 695                               0, 0);
 696 fail_alloc_used:
 697     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 698                               0, 0);
 699 fail_alloc_avail:
 700     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 701                               0, 0);
 702 fail_alloc_desc:
 703     return r;
 704 }
 705
 706 static void vhost_virtqueue_stop(struct vhost_dev *dev,
 707                                     struct VirtIODevice *vdev,
 708                                     struct vhost_virtqueue *vq,
 709                                     unsigned idx)
 710 {
 711     struct vhost_vring_state state = {
 712         .index = idx,
 713     };
 714     int r;
 715     r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state);
 716     if (r < 0) {
 717         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
 718         fflush(stderr);
 719     }
 720     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
 721     assert (r >= 0);
 722     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 723                               0, virtio_queue_get_ring_size(vdev, idx));
 724     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 725                               1, virtio_queue_get_used_size(vdev, idx));
 726     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 727                               0, virtio_queue_get_avail_size(vdev, idx));
 728     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 729                               0, virtio_queue_get_desc_size(vdev, idx));
 730 }
 731
 732 static void vhost_eventfd_add(MemoryListener *listener,
 733                               MemoryRegionSection *section,
 734                               bool match_data, uint64_t data, EventNotifier *e)
 735 {
 736 }
 737
 738 static void vhost_eventfd_del(MemoryListener *listener,
 739                               MemoryRegionSection *section,
 740                               bool match_data, uint64_t data, EventNotifier *e)
 741 {
 742 }
 743
 744 static int vhost_virtqueue_init(struct vhost_dev *dev,
 745                                 struct vhost_virtqueue *vq, int n)
 746 {
 747     struct vhost_vring_file file = {
 748         .index = n,
 749     };
 750     int r = event_notifier_init(&vq->masked_notifier, 0);
 751     if (r < 0) {
 752         return r;
 753     }
 754
 755     file.fd = event_notifier_get_fd(&vq->masked_notifier);
 756     r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file);
 757     if (r) {
 758         r = -errno;
 759         goto fail_call;
 760     }
 761     return 0;
 762 fail_call:
 763     event_notifier_cleanup(&vq->masked_notifier);
 764     return r;
 765 }
 766
 767 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
 768 {
 769     event_notifier_cleanup(&vq->masked_notifier);
 770 }
 771
 772 int vhost_dev_init(struct vhost_dev *hdev, int devfd, const char *devpath,
 773                    bool force)
 774 {
 775     uint64_t features;
 776     int i, r;
 777     if (devfd >= 0) {
 778         hdev->control = devfd;
 779     } else {
 780         hdev->control = open(devpath, O_RDWR);
 781         if (hdev->control < 0) {
 782             return -errno;
 783         }
 784     }
 785     r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
 786     if (r < 0) {
 787         goto fail;
 788     }
 789
 790     r = ioctl(hdev->control, VHOST_GET_FEATURES, &features);
 791     if (r < 0) {
 792         goto fail;
 793     }
 794
 795     for (i = 0; i < hdev->nvqs; ++i) {
 796         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
 797         if (r < 0) {
 798             goto fail_vq;
 799         }
 800     }
 801     hdev->features = features;
 802
 803     hdev->memory_listener = (MemoryListener) {
 804         .begin = vhost_begin,
 805         .commit = vhost_commit,
 806         .region_add = vhost_region_add,
 807         .region_del = vhost_region_del,
 808         .region_nop = vhost_region_nop,
 809         .log_start = vhost_log_start,
 810         .log_stop = vhost_log_stop,
 811         .log_sync = vhost_log_sync,
 812         .log_global_start = vhost_log_global_start,
 813         .log_global_stop = vhost_log_global_stop,
 814         .eventfd_add = vhost_eventfd_add,
 815         .eventfd_del = vhost_eventfd_del,
 816         .priority = 10
 817     };
 818     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
 819     hdev->n_mem_sections = 0;
 820     hdev->mem_sections = NULL;
 821     hdev->log = NULL;
 822     hdev->log_size = 0;
 823     hdev->log_enabled = false;
 824     hdev->started = false;
 825     memory_listener_register(&hdev->memory_listener, &address_space_memory);
 826     hdev->force = force;
 827     return 0;
 828 fail_vq:
 829     while (--i >= 0) {
 830         vhost_virtqueue_cleanup(hdev->vqs + i);
 831     }
 832 fail:
 833     r = -errno;
 834     close(hdev->control);
 835     return r;
 836 }
 837
 838 void vhost_dev_cleanup(struct vhost_dev *hdev)
 839 {
 840     int i;
 841     for (i = 0; i < hdev->nvqs; ++i) {
 842         vhost_virtqueue_cleanup(hdev->vqs + i);
 843     }
 844     memory_listener_unregister(&hdev->memory_listener);
 845     g_free(hdev->mem);
 846     g_free(hdev->mem_sections);
 847     close(hdev->control);
 848 }
 849
 850 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
 851 {
 852     return !vdev->binding->query_guest_notifiers ||
 853         vdev->binding->query_guest_notifiers(vdev->binding_opaque) ||
 854         hdev->force;
 855 }
 856
 857 /* Stop processing guest IO notifications in qemu.
 858  * Start processing them in vhost in kernel.
 859  */
 860 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 861 {
 862     int i, r;
 863     if (!vdev->binding->set_host_notifier) {
 864         fprintf(stderr, "binding does not support host notifiers\n");
 865         r = -ENOSYS;
 866         goto fail;
 867     }
 868
 869     for (i = 0; i < hdev->nvqs; ++i) {
 870         r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, true);
 871         if (r < 0) {
 872             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
 873             goto fail_vq;
 874         }
 875     }
 876
 877     return 0;
 878 fail_vq:
 879     while (--i >= 0) {
 880         r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, false);
 881         if (r < 0) {
 882             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
 883             fflush(stderr);
 884         }
 885         assert (r >= 0);
 886     }
 887 fail:
 888     return r;
 889 }
 890
 891 /* Stop processing guest IO notifications in vhost.
 892  * Start processing them in qemu.
 893  * This might actually run the qemu handlers right away,
 894  * so virtio in qemu must be completely setup when this is called.
 895  */
 896 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 897 {
 898     int i, r;
 899
 900     for (i = 0; i < hdev->nvqs; ++i) {
 901         r = vdev->binding->set_host_notifier(vdev->binding_opaque, i, false);
 902         if (r < 0) {
 903             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
 904             fflush(stderr);
 905         }
 906         assert (r >= 0);
 907     }
 908 }
 909
 910 /* Test and clear event pending status.
 911  * Should be called after unmask to avoid losing events.
 912  */
 913 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
 914 {
 915     struct vhost_virtqueue *vq = hdev->vqs + n;
 916     assert(hdev->started);
 917     return event_notifier_test_and_clear(&vq->masked_notifier);
 918 }
 919
 920 /* Mask/unmask events from this vq. */
 921 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
 922                          bool mask)
 923 {
 924     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
 925     int r;
 926
 927     assert(hdev->started);
 928
 929     struct vhost_vring_file file = {
 930         .index = n,
 931     };
 932     if (mask) {
 933         file.fd = event_notifier_get_fd(&hdev->vqs[n].masked_notifier);
 934     } else {
 935         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
 936     }
 937     r = ioctl(hdev->control, VHOST_SET_VRING_CALL, &file);
 938     assert(r >= 0);
 939 }
 940
 941 /* Host notifiers must be enabled at this point. */
 942 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
 943 {
 944     int i, r;
 945
 946     hdev->started = true;
 947
 948     if (!vdev->binding->set_guest_notifiers) {
 949         fprintf(stderr, "binding does not support guest notifiers\n");
 950         r = -ENOSYS;
 951         goto fail;
 952     }
 953
 954     r = vdev->binding->set_guest_notifiers(vdev->binding_opaque,
 955                                            hdev->nvqs,
 956                                            true);
 957     if (r < 0) {
 958         fprintf(stderr, "Error binding guest notifier: %d\n", -r);
 959         goto fail_notifiers;
 960     }
 961
 962     r = vhost_dev_set_features(hdev, hdev->log_enabled);
 963     if (r < 0) {
 964         goto fail_features;
 965     }
 966     r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem);
 967     if (r < 0) {
 968         r = -errno;
 969         goto fail_mem;
 970     }
 971     for (i = 0; i < hdev->nvqs; ++i) {
 972         r = vhost_virtqueue_start(hdev,
 973                                  vdev,
 974                                  hdev->vqs + i,
 975                                  i);
 976         if (r < 0) {
 977             goto fail_vq;
 978         }
 979     }
 980
 981     if (hdev->log_enabled) {
 982         hdev->log_size = vhost_get_log_size(hdev);
 983         hdev->log = hdev->log_size ?
 984             g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL;
 985         r = ioctl(hdev->control, VHOST_SET_LOG_BASE,
 986                   (uint64_t)(unsigned long)hdev->log);
 987         if (r < 0) {
 988             r = -errno;
 989             goto fail_log;
 990         }
 991     }
 992
 993     return 0;
 994 fail_log:
 995 fail_vq:
 996     while (--i >= 0) {
 997         vhost_virtqueue_stop(hdev,
 998                                 vdev,
 999                                 hdev->vqs + i,
1000                                 i);
1001     }
1002 fail_mem:
1003 fail_features:
1004     vdev->binding->set_guest_notifiers(vdev->binding_opaque, hdev->nvqs, false);
1005 fail_notifiers:
1006 fail:
1007
1008     hdev->started = false;
1009     return r;
1010 }
1011
1012 /* Host notifiers must be enabled at this point. */
1013 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1014 {
1015     int i, r;
1016
1017     for (i = 0; i < hdev->nvqs; ++i) {
1018         vhost_virtqueue_stop(hdev,
1019                                 vdev,
1020                                 hdev->vqs + i,
1021                                 i);
1022     }
1023     for (i = 0; i < hdev->n_mem_sections; ++i) {
1024         vhost_sync_dirty_bitmap(hdev, &hdev->mem_sections[i],
1025                                 0, (hwaddr)~0x0ull);
1026     }
1027     r = vdev->binding->set_guest_notifiers(vdev->binding_opaque,
1028                                            hdev->nvqs,
1029                                            false);
1030     if (r < 0) {
1031         fprintf(stderr, "vhost guest notifier cleanup failed: %d\n", r);
1032         fflush(stderr);
1033     }
1034     assert (r >= 0);
1035
1036     hdev->started = false;
1037     g_free(hdev->log);
1038     hdev->log = NULL;
1039     hdev->log_size = 0;
1040 }