hw/virtio/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 "hw/virtio/vhost.h"
  17 #include "hw/hw.h"
  18 #include "qemu/atomic.h"
  19 #include "qemu/range.h"
  20 #include <linux/vhost.h>
  21 #include "exec/address-spaces.h"
  22 #include "hw/virtio/virtio-bus.h"
  23
  24 static void vhost_dev_sync_region(struct vhost_dev *dev,
  25                                   MemoryRegionSection *section,
  26                                   uint64_t mfirst, uint64_t mlast,
  27                                   uint64_t rfirst, uint64_t rlast)
  28 {
  29     uint64_t start = MAX(mfirst, rfirst);
  30     uint64_t end = MIN(mlast, rlast);
  31     vhost_log_chunk_t *from = dev->log + start / VHOST_LOG_CHUNK;
  32     vhost_log_chunk_t *to = dev->log + end / VHOST_LOG_CHUNK + 1;
  33     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
  34
  35     if (end < start) {
  36         return;
  37     }
  38     assert(end / VHOST_LOG_CHUNK < dev->log_size);
  39     assert(start / VHOST_LOG_CHUNK < dev->log_size);
  40
  41     for (;from < to; ++from) {
  42         vhost_log_chunk_t log;
  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 need barrier semantics
  50          * but it's easier to use atomic_* than roll our own. */
  51         log = atomic_xchg(from, 0);
  52         while (log) {
  53             int bit = ctzl(log);
  54             hwaddr page_addr;
  55             hwaddr section_offset;
  56             hwaddr mr_offset;
  57             page_addr = addr + bit * VHOST_LOG_PAGE;
  58             section_offset = page_addr - section->offset_within_address_space;
  59             mr_offset = section_offset + section->offset_within_region;
  60             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
  61             log &= ~(0x1ull << bit);
  62         }
  63         addr += VHOST_LOG_CHUNK;
  64     }
  65 }
  66
  67 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
  68                                    MemoryRegionSection *section,
  69                                    hwaddr first,
  70                                    hwaddr last)
  71 {
  72     int i;
  73     hwaddr start_addr;
  74     hwaddr end_addr;
  75
  76     if (!dev->log_enabled || !dev->started) {
  77         return 0;
  78     }
  79     start_addr = section->offset_within_address_space;
  80     end_addr = range_get_last(start_addr, int128_get64(section->size));
  81     start_addr = MAX(first, start_addr);
  82     end_addr = MIN(last, end_addr);
  83
  84     for (i = 0; i < dev->mem->nregions; ++i) {
  85         struct vhost_memory_region *reg = dev->mem->regions + i;
  86         vhost_dev_sync_region(dev, section, start_addr, end_addr,
  87                               reg->guest_phys_addr,
  88                               range_get_last(reg->guest_phys_addr,
  89                                              reg->memory_size));
  90     }
  91     for (i = 0; i < dev->nvqs; ++i) {
  92         struct vhost_virtqueue *vq = dev->vqs + i;
  93         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
  94                               range_get_last(vq->used_phys, vq->used_size));
  95     }
  96     return 0;
  97 }
  98
  99 static void vhost_log_sync(MemoryListener *listener,
 100                           MemoryRegionSection *section)
 101 {
 102     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 103                                          memory_listener);
 104     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
 105 }
 106
 107 static void vhost_log_sync_range(struct vhost_dev *dev,
 108                                  hwaddr first, hwaddr last)
 109 {
 110     int i;
 111     /* FIXME: this is N^2 in number of sections */
 112     for (i = 0; i < dev->n_mem_sections; ++i) {
 113         MemoryRegionSection *section = &dev->mem_sections[i];
 114         vhost_sync_dirty_bitmap(dev, section, first, last);
 115     }
 116 }
 117
 118 /* Assign/unassign. Keep an unsorted array of non-overlapping
 119  * memory regions in dev->mem. */
 120 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
 121                                       uint64_t start_addr,
 122                                       uint64_t size)
 123 {
 124     int from, to, n = dev->mem->nregions;
 125     /* Track overlapping/split regions for sanity checking. */
 126     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
 127
 128     for (from = 0, to = 0; from < n; ++from, ++to) {
 129         struct vhost_memory_region *reg = dev->mem->regions + to;
 130         uint64_t reglast;
 131         uint64_t memlast;
 132         uint64_t change;
 133
 134         /* clone old region */
 135         if (to != from) {
 136             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 137         }
 138
 139         /* No overlap is simple */
 140         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 141                             start_addr, size)) {
 142             continue;
 143         }
 144
 145         /* Split only happens if supplied region
 146          * is in the middle of an existing one. Thus it can not
 147          * overlap with any other existing region. */
 148         assert(!split);
 149
 150         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 151         memlast = range_get_last(start_addr, size);
 152
 153         /* Remove whole region */
 154         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
 155             --dev->mem->nregions;
 156             --to;
 157             ++overlap_middle;
 158             continue;
 159         }
 160
 161         /* Shrink region */
 162         if (memlast >= reglast) {
 163             reg->memory_size = start_addr - reg->guest_phys_addr;
 164             assert(reg->memory_size);
 165             assert(!overlap_end);
 166             ++overlap_end;
 167             continue;
 168         }
 169
 170         /* Shift region */
 171         if (start_addr <= reg->guest_phys_addr) {
 172             change = memlast + 1 - reg->guest_phys_addr;
 173             reg->memory_size -= change;
 174             reg->guest_phys_addr += change;
 175             reg->userspace_addr += change;
 176             assert(reg->memory_size);
 177             assert(!overlap_start);
 178             ++overlap_start;
 179             continue;
 180         }
 181
 182         /* This only happens if supplied region
 183          * is in the middle of an existing one. Thus it can not
 184          * overlap with any other existing region. */
 185         assert(!overlap_start);
 186         assert(!overlap_end);
 187         assert(!overlap_middle);
 188         /* Split region: shrink first part, shift second part. */
 189         memcpy(dev->mem->regions + n, reg, sizeof *reg);
 190         reg->memory_size = start_addr - reg->guest_phys_addr;
 191         assert(reg->memory_size);
 192         change = memlast + 1 - reg->guest_phys_addr;
 193         reg = dev->mem->regions + n;
 194         reg->memory_size -= change;
 195         assert(reg->memory_size);
 196         reg->guest_phys_addr += change;
 197         reg->userspace_addr += change;
 198         /* Never add more than 1 region */
 199         assert(dev->mem->nregions == n);
 200         ++dev->mem->nregions;
 201         ++split;
 202     }
 203 }
 204
 205 /* Called after unassign, so no regions overlap the given range. */
 206 static void vhost_dev_assign_memory(struct vhost_dev *dev,
 207                                     uint64_t start_addr,
 208                                     uint64_t size,
 209                                     uint64_t uaddr)
 210 {
 211     int from, to;
 212     struct vhost_memory_region *merged = NULL;
 213     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
 214         struct vhost_memory_region *reg = dev->mem->regions + to;
 215         uint64_t prlast, urlast;
 216         uint64_t pmlast, umlast;
 217         uint64_t s, e, u;
 218
 219         /* clone old region */
 220         if (to != from) {
 221             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 222         }
 223         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 224         pmlast = range_get_last(start_addr, size);
 225         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
 226         umlast = range_get_last(uaddr, size);
 227
 228         /* check for overlapping regions: should never happen. */
 229         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
 230         /* Not an adjacent or overlapping region - do not merge. */
 231         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
 232             (pmlast + 1 != reg->guest_phys_addr ||
 233              umlast + 1 != reg->userspace_addr)) {
 234             continue;
 235         }
 236
 237         if (merged) {
 238             --to;
 239             assert(to >= 0);
 240         } else {
 241             merged = reg;
 242         }
 243         u = MIN(uaddr, reg->userspace_addr);
 244         s = MIN(start_addr, reg->guest_phys_addr);
 245         e = MAX(pmlast, prlast);
 246         uaddr = merged->userspace_addr = u;
 247         start_addr = merged->guest_phys_addr = s;
 248         size = merged->memory_size = e - s + 1;
 249         assert(merged->memory_size);
 250     }
 251
 252     if (!merged) {
 253         struct vhost_memory_region *reg = dev->mem->regions + to;
 254         memset(reg, 0, sizeof *reg);
 255         reg->memory_size = size;
 256         assert(reg->memory_size);
 257         reg->guest_phys_addr = start_addr;
 258         reg->userspace_addr = uaddr;
 259         ++to;
 260     }
 261     assert(to <= dev->mem->nregions + 1);
 262     dev->mem->nregions = to;
 263 }
 264
 265 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
 266 {
 267     uint64_t log_size = 0;
 268     int i;
 269     for (i = 0; i < dev->mem->nregions; ++i) {
 270         struct vhost_memory_region *reg = dev->mem->regions + i;
 271         uint64_t last = range_get_last(reg->guest_phys_addr,
 272                                        reg->memory_size);
 273         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 274     }
 275     for (i = 0; i < dev->nvqs; ++i) {
 276         struct vhost_virtqueue *vq = dev->vqs + i;
 277         uint64_t last = vq->used_phys + vq->used_size - 1;
 278         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 279     }
 280     return log_size;
 281 }
 282
 283 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
 284 {
 285     vhost_log_chunk_t *log;
 286     uint64_t log_base;
 287     int r;
 288
 289     log = g_malloc0(size * sizeof *log);
 290     log_base = (uint64_t)(unsigned long)log;
 291     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_LOG_BASE, &log_base);
 292     assert(r >= 0);
 293     /* Sync only the range covered by the old log */
 294     if (dev->log_size) {
 295         vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
 296     }
 297     g_free(dev->log);
 298     dev->log = log;
 299     dev->log_size = size;
 300 }
 301
 302 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
 303                                       uint64_t start_addr,
 304                                       uint64_t size)
 305 {
 306     int i;
 307     for (i = 0; i < dev->nvqs; ++i) {
 308         struct vhost_virtqueue *vq = dev->vqs + i;
 309         hwaddr l;
 310         void *p;
 311
 312         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
 313             continue;
 314         }
 315         l = vq->ring_size;
 316         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
 317         if (!p || l != vq->ring_size) {
 318             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
 319             return -ENOMEM;
 320         }
 321         if (p != vq->ring) {
 322             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
 323             return -EBUSY;
 324         }
 325         cpu_physical_memory_unmap(p, l, 0, 0);
 326     }
 327     return 0;
 328 }
 329
 330 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
 331                                                       uint64_t start_addr,
 332                                                       uint64_t size)
 333 {
 334     int i, n = dev->mem->nregions;
 335     for (i = 0; i < n; ++i) {
 336         struct vhost_memory_region *reg = dev->mem->regions + i;
 337         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 338                            start_addr, size)) {
 339             return reg;
 340         }
 341     }
 342     return NULL;
 343 }
 344
 345 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
 346                                  uint64_t start_addr,
 347                                  uint64_t size,
 348                                  uint64_t uaddr)
 349 {
 350     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
 351     uint64_t reglast;
 352     uint64_t memlast;
 353
 354     if (!reg) {
 355         return true;
 356     }
 357
 358     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 359     memlast = range_get_last(start_addr, size);
 360
 361     /* Need to extend region? */
 362     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
 363         return true;
 364     }
 365     /* userspace_addr changed? */
 366     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
 367 }
 368
 369 static void vhost_set_memory(MemoryListener *listener,
 370                              MemoryRegionSection *section,
 371                              bool add)
 372 {
 373     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 374                                          memory_listener);
 375     hwaddr start_addr = section->offset_within_address_space;
 376     ram_addr_t size = int128_get64(section->size);
 377     bool log_dirty = memory_region_is_logging(section->mr);
 378     int s = offsetof(struct vhost_memory, regions) +
 379         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
 380     void *ram;
 381
 382     dev->mem = g_realloc(dev->mem, s);
 383
 384     if (log_dirty) {
 385         add = false;
 386     }
 387
 388     assert(size);
 389
 390     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
 391     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
 392     if (add) {
 393         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
 394             /* Region exists with same address. Nothing to do. */
 395             return;
 396         }
 397     } else {
 398         if (!vhost_dev_find_reg(dev, start_addr, size)) {
 399             /* Removing region that we don't access. Nothing to do. */
 400             return;
 401         }
 402     }
 403
 404     vhost_dev_unassign_memory(dev, start_addr, size);
 405     if (add) {
 406         /* Add given mapping, merging adjacent regions if any */
 407         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
 408     } else {
 409         /* Remove old mapping for this memory, if any. */
 410         vhost_dev_unassign_memory(dev, start_addr, size);
 411     }
 412     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
 413     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
 414     dev->memory_changed = true;
 415 }
 416
 417 static bool vhost_section(MemoryRegionSection *section)
 418 {
 419     return memory_region_is_ram(section->mr);
 420 }
 421
 422 static void vhost_begin(MemoryListener *listener)
 423 {
 424     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 425                                          memory_listener);
 426     dev->mem_changed_end_addr = 0;
 427     dev->mem_changed_start_addr = -1;
 428 }
 429
 430 static void vhost_commit(MemoryListener *listener)
 431 {
 432     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 433                                          memory_listener);
 434     hwaddr start_addr = 0;
 435     ram_addr_t size = 0;
 436     uint64_t log_size;
 437     int r;
 438
 439     if (!dev->memory_changed) {
 440         return;
 441     }
 442     if (!dev->started) {
 443         return;
 444     }
 445     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
 446         return;
 447     }
 448
 449     if (dev->started) {
 450         start_addr = dev->mem_changed_start_addr;
 451         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
 452
 453         r = vhost_verify_ring_mappings(dev, start_addr, size);
 454         assert(r >= 0);
 455     }
 456
 457     if (!dev->log_enabled) {
 458         r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
 459         assert(r >= 0);
 460         dev->memory_changed = false;
 461         return;
 462     }
 463     log_size = vhost_get_log_size(dev);
 464     /* We allocate an extra 4K bytes to log,
 465      * to reduce the * number of reallocations. */
 466 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
 467     /* To log more, must increase log size before table update. */
 468     if (dev->log_size < log_size) {
 469         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
 470     }
 471     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
 472     assert(r >= 0);
 473     /* To log less, can only decrease log size after table update. */
 474     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
 475         vhost_dev_log_resize(dev, log_size);
 476     }
 477     dev->memory_changed = false;
 478 }
 479
 480 static void vhost_region_add(MemoryListener *listener,
 481                              MemoryRegionSection *section)
 482 {
 483     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 484                                          memory_listener);
 485
 486     if (!vhost_section(section)) {
 487         return;
 488     }
 489
 490     ++dev->n_mem_sections;
 491     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
 492                                 dev->n_mem_sections);
 493     dev->mem_sections[dev->n_mem_sections - 1] = *section;
 494     memory_region_ref(section->mr);
 495     vhost_set_memory(listener, section, true);
 496 }
 497
 498 static void vhost_region_del(MemoryListener *listener,
 499                              MemoryRegionSection *section)
 500 {
 501     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 502                                          memory_listener);
 503     int i;
 504
 505     if (!vhost_section(section)) {
 506         return;
 507     }
 508
 509     vhost_set_memory(listener, section, false);
 510     memory_region_unref(section->mr);
 511     for (i = 0; i < dev->n_mem_sections; ++i) {
 512         if (dev->mem_sections[i].offset_within_address_space
 513             == section->offset_within_address_space) {
 514             --dev->n_mem_sections;
 515             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
 516                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
 517             break;
 518         }
 519     }
 520 }
 521
 522 static void vhost_region_nop(MemoryListener *listener,
 523                              MemoryRegionSection *section)
 524 {
 525 }
 526
 527 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
 528                                     struct vhost_virtqueue *vq,
 529                                     unsigned idx, bool enable_log)
 530 {
 531     struct vhost_vring_addr addr = {
 532         .index = idx,
 533         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
 534         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
 535         .used_user_addr = (uint64_t)(unsigned long)vq->used,
 536         .log_guest_addr = vq->used_phys,
 537         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
 538     };
 539     int r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ADDR, &addr);
 540     if (r < 0) {
 541         return -errno;
 542     }
 543     return 0;
 544 }
 545
 546 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
 547 {
 548     uint64_t features = dev->acked_features;
 549     int r;
 550     if (enable_log) {
 551         features |= 0x1 << VHOST_F_LOG_ALL;
 552     }
 553     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_FEATURES, &features);
 554     return r < 0 ? -errno : 0;
 555 }
 556
 557 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
 558 {
 559     int r, t, i;
 560     r = vhost_dev_set_features(dev, enable_log);
 561     if (r < 0) {
 562         goto err_features;
 563     }
 564     for (i = 0; i < dev->nvqs; ++i) {
 565         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 566                                      enable_log);
 567         if (r < 0) {
 568             goto err_vq;
 569         }
 570     }
 571     return 0;
 572 err_vq:
 573     for (; i >= 0; --i) {
 574         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 575                                      dev->log_enabled);
 576         assert(t >= 0);
 577     }
 578     t = vhost_dev_set_features(dev, dev->log_enabled);
 579     assert(t >= 0);
 580 err_features:
 581     return r;
 582 }
 583
 584 static int vhost_migration_log(MemoryListener *listener, int enable)
 585 {
 586     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 587                                          memory_listener);
 588     int r;
 589     if (!!enable == dev->log_enabled) {
 590         return 0;
 591     }
 592     if (!dev->started) {
 593         dev->log_enabled = enable;
 594         return 0;
 595     }
 596     if (!enable) {
 597         r = vhost_dev_set_log(dev, false);
 598         if (r < 0) {
 599             return r;
 600         }
 601         g_free(dev->log);
 602         dev->log = NULL;
 603         dev->log_size = 0;
 604     } else {
 605         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
 606         r = vhost_dev_set_log(dev, true);
 607         if (r < 0) {
 608             return r;
 609         }
 610     }
 611     dev->log_enabled = enable;
 612     return 0;
 613 }
 614
 615 static void vhost_log_global_start(MemoryListener *listener)
 616 {
 617     int r;
 618
 619     r = vhost_migration_log(listener, true);
 620     if (r < 0) {
 621         abort();
 622     }
 623 }
 624
 625 static void vhost_log_global_stop(MemoryListener *listener)
 626 {
 627     int r;
 628
 629     r = vhost_migration_log(listener, false);
 630     if (r < 0) {
 631         abort();
 632     }
 633 }
 634
 635 static void vhost_log_start(MemoryListener *listener,
 636                             MemoryRegionSection *section)
 637 {
 638     /* FIXME: implement */
 639 }
 640
 641 static void vhost_log_stop(MemoryListener *listener,
 642                            MemoryRegionSection *section)
 643 {
 644     /* FIXME: implement */
 645 }
 646
 647 static int vhost_virtqueue_start(struct vhost_dev *dev,
 648                                 struct VirtIODevice *vdev,
 649                                 struct vhost_virtqueue *vq,
 650                                 unsigned idx)
 651 {
 652     hwaddr s, l, a;
 653     int r;
 654     int vhost_vq_index = idx - dev->vq_index;
 655     struct vhost_vring_file file = {
 656         .index = vhost_vq_index
 657     };
 658     struct vhost_vring_state state = {
 659         .index = vhost_vq_index
 660     };
 661     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
 662
 663     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 664
 665     vq->num = state.num = virtio_queue_get_num(vdev, idx);
 666     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_NUM, &state);
 667     if (r) {
 668         return -errno;
 669     }
 670
 671     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
 672     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_BASE, &state);
 673     if (r) {
 674         return -errno;
 675     }
 676
 677     s = l = virtio_queue_get_desc_size(vdev, idx);
 678     a = virtio_queue_get_desc_addr(vdev, idx);
 679     vq->desc = cpu_physical_memory_map(a, &l, 0);
 680     if (!vq->desc || l != s) {
 681         r = -ENOMEM;
 682         goto fail_alloc_desc;
 683     }
 684     s = l = virtio_queue_get_avail_size(vdev, idx);
 685     a = virtio_queue_get_avail_addr(vdev, idx);
 686     vq->avail = cpu_physical_memory_map(a, &l, 0);
 687     if (!vq->avail || l != s) {
 688         r = -ENOMEM;
 689         goto fail_alloc_avail;
 690     }
 691     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
 692     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
 693     vq->used = cpu_physical_memory_map(a, &l, 1);
 694     if (!vq->used || l != s) {
 695         r = -ENOMEM;
 696         goto fail_alloc_used;
 697     }
 698
 699     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
 700     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
 701     vq->ring = cpu_physical_memory_map(a, &l, 1);
 702     if (!vq->ring || l != s) {
 703         r = -ENOMEM;
 704         goto fail_alloc_ring;
 705     }
 706
 707     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
 708     if (r < 0) {
 709         r = -errno;
 710         goto fail_alloc;
 711     }
 712
 713     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
 714     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_KICK, &file);
 715     if (r) {
 716         r = -errno;
 717         goto fail_kick;
 718     }
 719
 720     /* Clear and discard previous events if any. */
 721     event_notifier_test_and_clear(&vq->masked_notifier);
 722
 723     return 0;
 724
 725 fail_kick:
 726 fail_alloc:
 727     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 728                               0, 0);
 729 fail_alloc_ring:
 730     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 731                               0, 0);
 732 fail_alloc_used:
 733     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 734                               0, 0);
 735 fail_alloc_avail:
 736     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 737                               0, 0);
 738 fail_alloc_desc:
 739     return r;
 740 }
 741
 742 static void vhost_virtqueue_stop(struct vhost_dev *dev,
 743                                     struct VirtIODevice *vdev,
 744                                     struct vhost_virtqueue *vq,
 745                                     unsigned idx)
 746 {
 747     struct vhost_vring_state state = {
 748         .index = idx - dev->vq_index
 749     };
 750     int r;
 751     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 752     r = dev->vhost_ops->vhost_call(dev, VHOST_GET_VRING_BASE, &state);
 753     if (r < 0) {
 754         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
 755         fflush(stderr);
 756     }
 757     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
 758     virtio_queue_invalidate_signalled_used(vdev, idx);
 759     assert (r >= 0);
 760     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 761                               0, virtio_queue_get_ring_size(vdev, idx));
 762     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 763                               1, virtio_queue_get_used_size(vdev, idx));
 764     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 765                               0, virtio_queue_get_avail_size(vdev, idx));
 766     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 767                               0, virtio_queue_get_desc_size(vdev, idx));
 768 }
 769
 770 static void vhost_eventfd_add(MemoryListener *listener,
 771                               MemoryRegionSection *section,
 772                               bool match_data, uint64_t data, EventNotifier *e)
 773 {
 774 }
 775
 776 static void vhost_eventfd_del(MemoryListener *listener,
 777                               MemoryRegionSection *section,
 778                               bool match_data, uint64_t data, EventNotifier *e)
 779 {
 780 }
 781
 782 static int vhost_virtqueue_init(struct vhost_dev *dev,
 783                                 struct vhost_virtqueue *vq, int n)
 784 {
 785     struct vhost_vring_file file = {
 786         .index = n,
 787     };
 788     int r = event_notifier_init(&vq->masked_notifier, 0);
 789     if (r < 0) {
 790         return r;
 791     }
 792
 793     file.fd = event_notifier_get_fd(&vq->masked_notifier);
 794     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_CALL, &file);
 795     if (r) {
 796         r = -errno;
 797         goto fail_call;
 798     }
 799     return 0;
 800 fail_call:
 801     event_notifier_cleanup(&vq->masked_notifier);
 802     return r;
 803 }
 804
 805 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
 806 {
 807     event_notifier_cleanup(&vq->masked_notifier);
 808 }
 809
 810 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
 811                    bool force)
 812 {
 813     uint64_t features;
 814     int i, r;
 815
 816     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
 817         return -errno;
 818     }
 819
 820     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_OWNER, NULL);
 821     if (r < 0) {
 822         goto fail;
 823     }
 824
 825     r = hdev->vhost_ops->vhost_call(hdev, VHOST_GET_FEATURES, &features);
 826     if (r < 0) {
 827         goto fail;
 828     }
 829
 830     for (i = 0; i < hdev->nvqs; ++i) {
 831         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
 832         if (r < 0) {
 833             goto fail_vq;
 834         }
 835     }
 836     hdev->features = features;
 837
 838     hdev->memory_listener = (MemoryListener) {
 839         .begin = vhost_begin,
 840         .commit = vhost_commit,
 841         .region_add = vhost_region_add,
 842         .region_del = vhost_region_del,
 843         .region_nop = vhost_region_nop,
 844         .log_start = vhost_log_start,
 845         .log_stop = vhost_log_stop,
 846         .log_sync = vhost_log_sync,
 847         .log_global_start = vhost_log_global_start,
 848         .log_global_stop = vhost_log_global_stop,
 849         .eventfd_add = vhost_eventfd_add,
 850         .eventfd_del = vhost_eventfd_del,
 851         .priority = 10
 852     };
 853     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
 854     hdev->n_mem_sections = 0;
 855     hdev->mem_sections = NULL;
 856     hdev->log = NULL;
 857     hdev->log_size = 0;
 858     hdev->log_enabled = false;
 859     hdev->started = false;
 860     hdev->memory_changed = false;
 861     memory_listener_register(&hdev->memory_listener, &address_space_memory);
 862     hdev->force = force;
 863     return 0;
 864 fail_vq:
 865     while (--i >= 0) {
 866         vhost_virtqueue_cleanup(hdev->vqs + i);
 867     }
 868 fail:
 869     r = -errno;
 870     hdev->vhost_ops->vhost_backend_cleanup(hdev);
 871     return r;
 872 }
 873
 874 void vhost_dev_cleanup(struct vhost_dev *hdev)
 875 {
 876     int i;
 877     for (i = 0; i < hdev->nvqs; ++i) {
 878         vhost_virtqueue_cleanup(hdev->vqs + i);
 879     }
 880     memory_listener_unregister(&hdev->memory_listener);
 881     g_free(hdev->mem);
 882     g_free(hdev->mem_sections);
 883     hdev->vhost_ops->vhost_backend_cleanup(hdev);
 884 }
 885
 886 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
 887 {
 888     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 889     VirtioBusState *vbus = VIRTIO_BUS(qbus);
 890     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
 891
 892     return !k->query_guest_notifiers ||
 893            k->query_guest_notifiers(qbus->parent) ||
 894            hdev->force;
 895 }
 896
 897 /* Stop processing guest IO notifications in qemu.
 898  * Start processing them in vhost in kernel.
 899  */
 900 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 901 {
 902     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 903     VirtioBusState *vbus = VIRTIO_BUS(qbus);
 904     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
 905     int i, r;
 906     if (!k->set_host_notifier) {
 907         fprintf(stderr, "binding does not support host notifiers\n");
 908         r = -ENOSYS;
 909         goto fail;
 910     }
 911
 912     for (i = 0; i < hdev->nvqs; ++i) {
 913         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
 914         if (r < 0) {
 915             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
 916             goto fail_vq;
 917         }
 918     }
 919
 920     return 0;
 921 fail_vq:
 922     while (--i >= 0) {
 923         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
 924         if (r < 0) {
 925             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
 926             fflush(stderr);
 927         }
 928         assert (r >= 0);
 929     }
 930 fail:
 931     return r;
 932 }
 933
 934 /* Stop processing guest IO notifications in vhost.
 935  * Start processing them in qemu.
 936  * This might actually run the qemu handlers right away,
 937  * so virtio in qemu must be completely setup when this is called.
 938  */
 939 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 940 {
 941     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 942     VirtioBusState *vbus = VIRTIO_BUS(qbus);
 943     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
 944     int i, r;
 945
 946     for (i = 0; i < hdev->nvqs; ++i) {
 947         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
 948         if (r < 0) {
 949             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
 950             fflush(stderr);
 951         }
 952         assert (r >= 0);
 953     }
 954 }
 955
 956 /* Test and clear event pending status.
 957  * Should be called after unmask to avoid losing events.
 958  */
 959 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
 960 {
 961     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
 962     assert(hdev->started);
 963     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 964     return event_notifier_test_and_clear(&vq->masked_notifier);
 965 }
 966
 967 /* Mask/unmask events from this vq. */
 968 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
 969                          bool mask)
 970 {
 971     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
 972     int r, index = n - hdev->vq_index;
 973
 974     assert(hdev->started);
 975     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 976
 977     struct vhost_vring_file file = {
 978         .index = index
 979     };
 980     if (mask) {
 981         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
 982     } else {
 983         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
 984     }
 985     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_VRING_CALL, &file);
 986     assert(r >= 0);
 987 }
 988
 989 unsigned vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
 990         unsigned features)
 991 {
 992     const int *bit = feature_bits;
 993     while (*bit != VHOST_INVALID_FEATURE_BIT) {
 994         unsigned bit_mask = (1 << *bit);
 995         if (!(hdev->features & bit_mask)) {
 996             features &= ~bit_mask;
 997         }
 998         bit++;
 999     }
1000     return features;
1001 }
1002
1003 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1004         unsigned features)
1005 {
1006     const int *bit = feature_bits;
1007     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1008         unsigned bit_mask = (1 << *bit);
1009         if (features & bit_mask) {
1010             hdev->acked_features |= bit_mask;
1011         }
1012         bit++;
1013     }
1014 }
1015
1016 /* Host notifiers must be enabled at this point. */
1017 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1018 {
1019     int i, r;
1020
1021     hdev->started = true;
1022
1023     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1024     if (r < 0) {
1025         goto fail_features;
1026     }
1027     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_MEM_TABLE, hdev->mem);
1028     if (r < 0) {
1029         r = -errno;
1030         goto fail_mem;
1031     }
1032     for (i = 0; i < hdev->nvqs; ++i) {
1033         r = vhost_virtqueue_start(hdev,
1034                                   vdev,
1035                                   hdev->vqs + i,
1036                                   hdev->vq_index + i);
1037         if (r < 0) {
1038             goto fail_vq;
1039         }
1040     }
1041
1042     if (hdev->log_enabled) {
1043         hdev->log_size = vhost_get_log_size(hdev);
1044         hdev->log = hdev->log_size ?
1045             g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL;
1046         r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_LOG_BASE, hdev->log);
1047         if (r < 0) {
1048             r = -errno;
1049             goto fail_log;
1050         }
1051     }
1052
1053     return 0;
1054 fail_log:
1055 fail_vq:
1056     while (--i >= 0) {
1057         vhost_virtqueue_stop(hdev,
1058                              vdev,
1059                              hdev->vqs + i,
1060                              hdev->vq_index + i);
1061     }
1062     i = hdev->nvqs;
1063 fail_mem:
1064 fail_features:
1065
1066     hdev->started = false;
1067     return r;
1068 }
1069
1070 /* Host notifiers must be enabled at this point. */
1071 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1072 {
1073     int i;
1074
1075     for (i = 0; i < hdev->nvqs; ++i) {
1076         vhost_virtqueue_stop(hdev,
1077                              vdev,
1078                              hdev->vqs + i,
1079                              hdev->vq_index + i);
1080     }
1081     vhost_log_sync_range(hdev, 0, ~0x0ull);
1082
1083     hdev->started = false;
1084     g_free(hdev->log);
1085     hdev->log = NULL;
1086     hdev->log_size = 0;
1087 }
1088