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
 273     log = g_malloc0(size * sizeof *log);
 274     log_base = (uint64_t)(unsigned long)log;
 275     r = ioctl(dev->control, VHOST_SET_LOG_BASE, &log_base);
 276     assert(r >= 0);
 277     for (i = 0; i < dev->n_mem_sections; ++i) {
 278         /* Sync only the range covered by the old log */
 279         vhost_sync_dirty_bitmap(dev, &dev->mem_sections[i], 0,
 280                                 dev->log_size * VHOST_LOG_CHUNK - 1);
 281     }
 282     if (dev->log) {
 283         g_free(dev->log);
 284     }
 285     dev->log = log;
 286     dev->log_size = size;
 287 }
 288
 289 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
 290                                       uint64_t start_addr,
 291                                       uint64_t size)
 292 {
 293     int i;
 294     for (i = 0; i < dev->nvqs; ++i) {
 295         struct vhost_virtqueue *vq = dev->vqs + i;
 296         hwaddr l;
 297         void *p;
 298
 299         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
 300             continue;
 301         }
 302         l = vq->ring_size;
 303         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
 304         if (!p || l != vq->ring_size) {
 305             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
 306             return -ENOMEM;
 307         }
 308         if (p != vq->ring) {
 309             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
 310             return -EBUSY;
 311         }
 312         cpu_physical_memory_unmap(p, l, 0, 0);
 313     }
 314     return 0;
 315 }
 316
 317 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
 318                                                       uint64_t start_addr,
 319                                                       uint64_t size)
 320 {
 321     int i, n = dev->mem->nregions;
 322     for (i = 0; i < n; ++i) {
 323         struct vhost_memory_region *reg = dev->mem->regions + i;
 324         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 325                            start_addr, size)) {
 326             return reg;
 327         }
 328     }
 329     return NULL;
 330 }
 331
 332 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
 333                                  uint64_t start_addr,
 334                                  uint64_t size,
 335                                  uint64_t uaddr)
 336 {
 337     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
 338     uint64_t reglast;
 339     uint64_t memlast;
 340
 341     if (!reg) {
 342         return true;
 343     }
 344
 345     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 346     memlast = range_get_last(start_addr, size);
 347
 348     /* Need to extend region? */
 349     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
 350         return true;
 351     }
 352     /* userspace_addr changed? */
 353     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
 354 }
 355
 356 static void vhost_set_memory(MemoryListener *listener,
 357                              MemoryRegionSection *section,
 358                              bool add)
 359 {
 360     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 361                                          memory_listener);
 362     hwaddr start_addr = section->offset_within_address_space;
 363     ram_addr_t size = section->size;
 364     bool log_dirty = memory_region_is_logging(section->mr);
 365     int s = offsetof(struct vhost_memory, regions) +
 366         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
 367     uint64_t log_size;
 368     int r;
 369     void *ram;
 370
 371     dev->mem = g_realloc(dev->mem, s);
 372
 373     if (log_dirty) {
 374         add = false;
 375     }
 376
 377     assert(size);
 378
 379     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
 380     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
 381     if (add) {
 382         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
 383             /* Region exists with same address. Nothing to do. */
 384             return;
 385         }
 386     } else {
 387         if (!vhost_dev_find_reg(dev, start_addr, size)) {
 388             /* Removing region that we don't access. Nothing to do. */
 389             return;
 390         }
 391     }
 392
 393     vhost_dev_unassign_memory(dev, start_addr, size);
 394     if (add) {
 395         /* Add given mapping, merging adjacent regions if any */
 396         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
 397     } else {
 398         /* Remove old mapping for this memory, if any. */
 399         vhost_dev_unassign_memory(dev, start_addr, size);
 400     }
 401
 402     if (!dev->started) {
 403         return;
 404     }
 405
 406     if (dev->started) {
 407         r = vhost_verify_ring_mappings(dev, start_addr, size);
 408         assert(r >= 0);
 409     }
 410
 411     if (!dev->log_enabled) {
 412         r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 413         assert(r >= 0);
 414         return;
 415     }
 416     log_size = vhost_get_log_size(dev);
 417     /* We allocate an extra 4K bytes to log,
 418      * to reduce the * number of reallocations. */
 419 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
 420     /* To log more, must increase log size before table update. */
 421     if (dev->log_size < log_size) {
 422         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
 423     }
 424     r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 425     assert(r >= 0);
 426     /* To log less, can only decrease log size after table update. */
 427     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
 428         vhost_dev_log_resize(dev, log_size);
 429     }
 430 }
 431
 432 static bool vhost_section(MemoryRegionSection *section)
 433 {
 434     return memory_region_is_ram(section->mr);
 435 }
 436
 437 static void vhost_begin(MemoryListener *listener)
 438 {
 439 }
 440
 441 static void vhost_commit(MemoryListener *listener)
 442 {
 443 }
 444
 445 static void vhost_region_add(MemoryListener *listener,
 446                              MemoryRegionSection *section)
 447 {
 448     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 449                                          memory_listener);
 450
 451     if (!vhost_section(section)) {
 452         return;
 453     }
 454
 455     ++dev->n_mem_sections;
 456     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
 457                                 dev->n_mem_sections);
 458     dev->mem_sections[dev->n_mem_sections - 1] = *section;
 459     vhost_set_memory(listener, section, true);
 460 }
 461
 462 static void vhost_region_del(MemoryListener *listener,
 463                              MemoryRegionSection *section)
 464 {
 465     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 466                                          memory_listener);
 467     int i;
 468
 469     if (!vhost_section(section)) {
 470         return;
 471     }
 472
 473     vhost_set_memory(listener, section, false);
 474     for (i = 0; i < dev->n_mem_sections; ++i) {
 475         if (dev->mem_sections[i].offset_within_address_space
 476             == section->offset_within_address_space) {
 477             --dev->n_mem_sections;
 478             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
 479                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
 480             break;
 481         }
 482     }
 483 }
 484
 485 static void vhost_region_nop(MemoryListener *listener,
 486                              MemoryRegionSection *section)
 487 {
 488 }
 489
 490 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
 491                                     struct vhost_virtqueue *vq,
 492                                     unsigned idx, bool enable_log)
 493 {
 494     struct vhost_vring_addr addr = {
 495         .index = idx,
 496         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
 497         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
 498         .used_user_addr = (uint64_t)(unsigned long)vq->used,
 499         .log_guest_addr = vq->used_phys,
 500         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
 501     };
 502     int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
 503     if (r < 0) {
 504         return -errno;
 505     }
 506     return 0;
 507 }
 508
 509 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
 510 {
 511     uint64_t features = dev->acked_features;
 512     int r;
 513     if (enable_log) {
 514         features |= 0x1 << VHOST_F_LOG_ALL;
 515     }
 516     r = ioctl(dev->control, VHOST_SET_FEATURES, &features);
 517     return r < 0 ? -errno : 0;
 518 }
 519
 520 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
 521 {
 522     int r, t, i;
 523     r = vhost_dev_set_features(dev, enable_log);
 524     if (r < 0) {
 525         goto err_features;
 526     }
 527     for (i = 0; i < dev->nvqs; ++i) {
 528         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 529                                      enable_log);
 530         if (r < 0) {
 531             goto err_vq;
 532         }
 533     }
 534     return 0;
 535 err_vq:
 536     for (; i >= 0; --i) {
 537         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 538                                      dev->log_enabled);
 539         assert(t >= 0);
 540     }
 541     t = vhost_dev_set_features(dev, dev->log_enabled);
 542     assert(t >= 0);
 543 err_features:
 544     return r;
 545 }
 546
 547 static int vhost_migration_log(MemoryListener *listener, int enable)
 548 {
 549     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 550                                          memory_listener);
 551     int r;
 552     if (!!enable == dev->log_enabled) {
 553         return 0;
 554     }
 555     if (!dev->started) {
 556         dev->log_enabled = enable;
 557         return 0;
 558     }
 559     if (!enable) {
 560         r = vhost_dev_set_log(dev, false);
 561         if (r < 0) {
 562             return r;
 563         }
 564         if (dev->log) {
 565             g_free(dev->log);
 566         }
 567         dev->log = NULL;
 568         dev->log_size = 0;
 569     } else {
 570         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
 571         r = vhost_dev_set_log(dev, true);
 572         if (r < 0) {
 573             return r;
 574         }
 575     }
 576     dev->log_enabled = enable;
 577     return 0;
 578 }
 579
 580 static void vhost_log_global_start(MemoryListener *listener)
 581 {
 582     int r;
 583
 584     r = vhost_migration_log(listener, true);
 585     if (r < 0) {
 586         abort();
 587     }
 588 }
 589
 590 static void vhost_log_global_stop(MemoryListener *listener)
 591 {
 592     int r;
 593
 594     r = vhost_migration_log(listener, false);
 595     if (r < 0) {
 596         abort();
 597     }
 598 }
 599
 600 static void vhost_log_start(MemoryListener *listener,
 601                             MemoryRegionSection *section)
 602 {
 603     /* FIXME: implement */
 604 }
 605
 606 static void vhost_log_stop(MemoryListener *listener,
 607                            MemoryRegionSection *section)
 608 {
 609     /* FIXME: implement */
 610 }
 611
 612 static int vhost_virtqueue_start(struct vhost_dev *dev,
 613                                 struct VirtIODevice *vdev,
 614                                 struct vhost_virtqueue *vq,
 615                                 unsigned idx)
 616 {
 617     hwaddr s, l, a;
 618     int r;
 619     int vhost_vq_index = idx - dev->vq_index;
 620     struct vhost_vring_file file = {
 621         .index = vhost_vq_index
 622     };
 623     struct vhost_vring_state state = {
 624         .index = vhost_vq_index
 625     };
 626     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
 627
 628     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 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, vhost_vq_index, dev->log_enabled);
 673     if (r < 0) {
 674         r = -errno;
 675         goto fail_alloc;
 676     }
 677
 678     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
 679     r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file);
 680     if (r) {
 681         r = -errno;
 682         goto fail_kick;
 683     }
 684
 685     /* Clear and discard previous events if any. */
 686     event_notifier_test_and_clear(&vq->masked_notifier);
 687
 688     return 0;
 689
 690 fail_kick:
 691 fail_alloc:
 692     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 693                               0, 0);
 694 fail_alloc_ring:
 695     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 696                               0, 0);
 697 fail_alloc_used:
 698     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 699                               0, 0);
 700 fail_alloc_avail:
 701     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 702                               0, 0);
 703 fail_alloc_desc:
 704     return r;
 705 }
 706
 707 static void vhost_virtqueue_stop(struct vhost_dev *dev,
 708                                     struct VirtIODevice *vdev,
 709                                     struct vhost_virtqueue *vq,
 710                                     unsigned idx)
 711 {
 712     struct vhost_vring_state state = {
 713         .index = idx - dev->vq_index
 714     };
 715     int r;
 716     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 717     r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state);
 718     if (r < 0) {
 719         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
 720         fflush(stderr);
 721     }
 722     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
 723     assert (r >= 0);
 724     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 725                               0, virtio_queue_get_ring_size(vdev, idx));
 726     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 727                               1, virtio_queue_get_used_size(vdev, idx));
 728     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 729                               0, virtio_queue_get_avail_size(vdev, idx));
 730     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 731                               0, virtio_queue_get_desc_size(vdev, idx));
 732 }
 733
 734 static void vhost_eventfd_add(MemoryListener *listener,
 735                               MemoryRegionSection *section,
 736                               bool match_data, uint64_t data, EventNotifier *e)
 737 {
 738 }
 739
 740 static void vhost_eventfd_del(MemoryListener *listener,
 741                               MemoryRegionSection *section,
 742                               bool match_data, uint64_t data, EventNotifier *e)
 743 {
 744 }
 745
 746 static int vhost_virtqueue_init(struct vhost_dev *dev,
 747                                 struct vhost_virtqueue *vq, int n)
 748 {
 749     struct vhost_vring_file file = {
 750         .index = n,
 751     };
 752     int r = event_notifier_init(&vq->masked_notifier, 0);
 753     if (r < 0) {
 754         return r;
 755     }
 756
 757     file.fd = event_notifier_get_fd(&vq->masked_notifier);
 758     r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file);
 759     if (r) {
 760         r = -errno;
 761         goto fail_call;
 762     }
 763     return 0;
 764 fail_call:
 765     event_notifier_cleanup(&vq->masked_notifier);
 766     return r;
 767 }
 768
 769 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
 770 {
 771     event_notifier_cleanup(&vq->masked_notifier);
 772 }
 773
 774 int vhost_dev_init(struct vhost_dev *hdev, int devfd, const char *devpath,
 775                    bool force)
 776 {
 777     uint64_t features;
 778     int i, r;
 779     if (devfd >= 0) {
 780         hdev->control = devfd;
 781     } else {
 782         hdev->control = open(devpath, O_RDWR);
 783         if (hdev->control < 0) {
 784             return -errno;
 785         }
 786     }
 787     r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
 788     if (r < 0) {
 789         goto fail;
 790     }
 791
 792     r = ioctl(hdev->control, VHOST_GET_FEATURES, &features);
 793     if (r < 0) {
 794         goto fail;
 795     }
 796
 797     for (i = 0; i < hdev->nvqs; ++i) {
 798         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
 799         if (r < 0) {
 800             goto fail_vq;
 801         }
 802     }
 803     hdev->features = features;
 804
 805     hdev->memory_listener = (MemoryListener) {
 806         .begin = vhost_begin,
 807         .commit = vhost_commit,
 808         .region_add = vhost_region_add,
 809         .region_del = vhost_region_del,
 810         .region_nop = vhost_region_nop,
 811         .log_start = vhost_log_start,
 812         .log_stop = vhost_log_stop,
 813         .log_sync = vhost_log_sync,
 814         .log_global_start = vhost_log_global_start,
 815         .log_global_stop = vhost_log_global_stop,
 816         .eventfd_add = vhost_eventfd_add,
 817         .eventfd_del = vhost_eventfd_del,
 818         .priority = 10
 819     };
 820     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
 821     hdev->n_mem_sections = 0;
 822     hdev->mem_sections = NULL;
 823     hdev->log = NULL;
 824     hdev->log_size = 0;
 825     hdev->log_enabled = false;
 826     hdev->started = false;
 827     memory_listener_register(&hdev->memory_listener, &address_space_memory);
 828     hdev->force = force;
 829     return 0;
 830 fail_vq:
 831     while (--i >= 0) {
 832         vhost_virtqueue_cleanup(hdev->vqs + i);
 833     }
 834 fail:
 835     r = -errno;
 836     close(hdev->control);
 837     return r;
 838 }
 839
 840 void vhost_dev_cleanup(struct vhost_dev *hdev)
 841 {
 842     int i;
 843     for (i = 0; i < hdev->nvqs; ++i) {
 844         vhost_virtqueue_cleanup(hdev->vqs + i);
 845     }
 846     memory_listener_unregister(&hdev->memory_listener);
 847     g_free(hdev->mem);
 848     g_free(hdev->mem_sections);
 849     close(hdev->control);
 850 }
 851
 852 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
 853 {
 854     return !vdev->binding->query_guest_notifiers ||
 855         vdev->binding->query_guest_notifiers(vdev->binding_opaque) ||
 856         hdev->force;
 857 }
 858
 859 /* Stop processing guest IO notifications in qemu.
 860  * Start processing them in vhost in kernel.
 861  */
 862 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 863 {
 864     int i, r;
 865     if (!vdev->binding->set_host_notifier) {
 866         fprintf(stderr, "binding does not support host notifiers\n");
 867         r = -ENOSYS;
 868         goto fail;
 869     }
 870
 871     for (i = 0; i < hdev->nvqs; ++i) {
 872         r = vdev->binding->set_host_notifier(vdev->binding_opaque,
 873                                              hdev->vq_index + i,
 874                                              true);
 875         if (r < 0) {
 876             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
 877             goto fail_vq;
 878         }
 879     }
 880
 881     return 0;
 882 fail_vq:
 883     while (--i >= 0) {
 884         r = vdev->binding->set_host_notifier(vdev->binding_opaque,
 885                                              hdev->vq_index + i,
 886                                              false);
 887         if (r < 0) {
 888             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
 889             fflush(stderr);
 890         }
 891         assert (r >= 0);
 892     }
 893 fail:
 894     return r;
 895 }
 896
 897 /* Stop processing guest IO notifications in vhost.
 898  * Start processing them in qemu.
 899  * This might actually run the qemu handlers right away,
 900  * so virtio in qemu must be completely setup when this is called.
 901  */
 902 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 903 {
 904     int i, r;
 905
 906     for (i = 0; i < hdev->nvqs; ++i) {
 907         r = vdev->binding->set_host_notifier(vdev->binding_opaque,
 908                                              hdev->vq_index + i,
 909                                              false);
 910         if (r < 0) {
 911             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
 912             fflush(stderr);
 913         }
 914         assert (r >= 0);
 915     }
 916 }
 917
 918 /* Test and clear event pending status.
 919  * Should be called after unmask to avoid losing events.
 920  */
 921 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
 922 {
 923     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
 924     assert(hdev->started);
 925     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 926     return event_notifier_test_and_clear(&vq->masked_notifier);
 927 }
 928
 929 /* Mask/unmask events from this vq. */
 930 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
 931                          bool mask)
 932 {
 933     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
 934     int r, index = n - hdev->vq_index;
 935
 936     assert(hdev->started);
 937     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 938
 939     struct vhost_vring_file file = {
 940         .index = index
 941     };
 942     if (mask) {
 943         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
 944     } else {
 945         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
 946     }
 947     r = ioctl(hdev->control, VHOST_SET_VRING_CALL, &file);
 948     assert(r >= 0);
 949 }
 950
 951 /* Host notifiers must be enabled at this point. */
 952 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
 953 {
 954     int i, r;
 955
 956     hdev->started = true;
 957
 958     r = vhost_dev_set_features(hdev, hdev->log_enabled);
 959     if (r < 0) {
 960         goto fail_features;
 961     }
 962     r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem);
 963     if (r < 0) {
 964         r = -errno;
 965         goto fail_mem;
 966     }
 967     for (i = 0; i < hdev->nvqs; ++i) {
 968         r = vhost_virtqueue_start(hdev,
 969                                   vdev,
 970                                   hdev->vqs + i,
 971                                   hdev->vq_index + i);
 972         if (r < 0) {
 973             goto fail_vq;
 974         }
 975     }
 976
 977     if (hdev->log_enabled) {
 978         hdev->log_size = vhost_get_log_size(hdev);
 979         hdev->log = hdev->log_size ?
 980             g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL;
 981         r = ioctl(hdev->control, VHOST_SET_LOG_BASE,
 982                   (uint64_t)(unsigned long)hdev->log);
 983         if (r < 0) {
 984             r = -errno;
 985             goto fail_log;
 986         }
 987     }
 988
 989     return 0;
 990 fail_log:
 991 fail_vq:
 992     while (--i >= 0) {
 993         vhost_virtqueue_stop(hdev,
 994                              vdev,
 995                              hdev->vqs + i,
 996                              hdev->vq_index + i);
 997     }
 998     i = hdev->nvqs;
 999 fail_mem:
1000 fail_features:
1001
1002     hdev->started = false;
1003     return r;
1004 }
1005
1006 /* Host notifiers must be enabled at this point. */
1007 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1008 {
1009     int i;
1010
1011     for (i = 0; i < hdev->nvqs; ++i) {
1012         vhost_virtqueue_stop(hdev,
1013                              vdev,
1014                              hdev->vqs + i,
1015                              hdev->vq_index + i);
1016     }
1017     for (i = 0; i < hdev->n_mem_sections; ++i) {
1018         vhost_sync_dirty_bitmap(hdev, &hdev->mem_sections[i],
1019                                 0, (hwaddr)~0x0ull);
1020     }
1021
1022     hdev->started = false;
1023     g_free(hdev->log);
1024     hdev->log = NULL;
1025     hdev->log_size = 0;
1026 }
1027