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