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 "hw/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     uint64_t log_size;
 386     int r;
 387     void *ram;
 388
 389     dev->mem = g_realloc(dev->mem, s);
 390
 391     if (log_dirty) {
 392         add = false;
 393     }
 394
 395     assert(size);
 396
 397     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
 398     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
 399     if (add) {
 400         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
 401             /* Region exists with same address. Nothing to do. */
 402             return;
 403         }
 404     } else {
 405         if (!vhost_dev_find_reg(dev, start_addr, size)) {
 406             /* Removing region that we don't access. Nothing to do. */
 407             return;
 408         }
 409     }
 410
 411     vhost_dev_unassign_memory(dev, start_addr, size);
 412     if (add) {
 413         /* Add given mapping, merging adjacent regions if any */
 414         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
 415     } else {
 416         /* Remove old mapping for this memory, if any. */
 417         vhost_dev_unassign_memory(dev, start_addr, size);
 418     }
 419
 420     if (!dev->started) {
 421         return;
 422     }
 423
 424     if (dev->started) {
 425         r = vhost_verify_ring_mappings(dev, start_addr, size);
 426         assert(r >= 0);
 427     }
 428
 429     if (!dev->log_enabled) {
 430         r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 431         assert(r >= 0);
 432         return;
 433     }
 434     log_size = vhost_get_log_size(dev);
 435     /* We allocate an extra 4K bytes to log,
 436      * to reduce the * number of reallocations. */
 437 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
 438     /* To log more, must increase log size before table update. */
 439     if (dev->log_size < log_size) {
 440         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
 441     }
 442     r = ioctl(dev->control, VHOST_SET_MEM_TABLE, dev->mem);
 443     assert(r >= 0);
 444     /* To log less, can only decrease log size after table update. */
 445     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
 446         vhost_dev_log_resize(dev, log_size);
 447     }
 448 }
 449
 450 static bool vhost_section(MemoryRegionSection *section)
 451 {
 452     return memory_region_is_ram(section->mr);
 453 }
 454
 455 static void vhost_begin(MemoryListener *listener)
 456 {
 457 }
 458
 459 static void vhost_commit(MemoryListener *listener)
 460 {
 461 }
 462
 463 static void vhost_region_add(MemoryListener *listener,
 464                              MemoryRegionSection *section)
 465 {
 466     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 467                                          memory_listener);
 468
 469     if (!vhost_section(section)) {
 470         return;
 471     }
 472
 473     ++dev->n_mem_sections;
 474     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
 475                                 dev->n_mem_sections);
 476     dev->mem_sections[dev->n_mem_sections - 1] = *section;
 477     vhost_set_memory(listener, section, true);
 478 }
 479
 480 static void vhost_region_del(MemoryListener *listener,
 481                              MemoryRegionSection *section)
 482 {
 483     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 484                                          memory_listener);
 485     int i;
 486
 487     if (!vhost_section(section)) {
 488         return;
 489     }
 490
 491     vhost_set_memory(listener, section, false);
 492     for (i = 0; i < dev->n_mem_sections; ++i) {
 493         if (dev->mem_sections[i].offset_within_address_space
 494             == section->offset_within_address_space) {
 495             --dev->n_mem_sections;
 496             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
 497                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
 498             break;
 499         }
 500     }
 501 }
 502
 503 static void vhost_region_nop(MemoryListener *listener,
 504                              MemoryRegionSection *section)
 505 {
 506 }
 507
 508 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
 509                                     struct vhost_virtqueue *vq,
 510                                     unsigned idx, bool enable_log)
 511 {
 512     struct vhost_vring_addr addr = {
 513         .index = idx,
 514         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
 515         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
 516         .used_user_addr = (uint64_t)(unsigned long)vq->used,
 517         .log_guest_addr = vq->used_phys,
 518         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
 519     };
 520     int r = ioctl(dev->control, VHOST_SET_VRING_ADDR, &addr);
 521     if (r < 0) {
 522         return -errno;
 523     }
 524     return 0;
 525 }
 526
 527 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
 528 {
 529     uint64_t features = dev->acked_features;
 530     int r;
 531     if (enable_log) {
 532         features |= 0x1 << VHOST_F_LOG_ALL;
 533     }
 534     r = ioctl(dev->control, VHOST_SET_FEATURES, &features);
 535     return r < 0 ? -errno : 0;
 536 }
 537
 538 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
 539 {
 540     int r, t, i;
 541     r = vhost_dev_set_features(dev, enable_log);
 542     if (r < 0) {
 543         goto err_features;
 544     }
 545     for (i = 0; i < dev->nvqs; ++i) {
 546         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 547                                      enable_log);
 548         if (r < 0) {
 549             goto err_vq;
 550         }
 551     }
 552     return 0;
 553 err_vq:
 554     for (; i >= 0; --i) {
 555         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 556                                      dev->log_enabled);
 557         assert(t >= 0);
 558     }
 559     t = vhost_dev_set_features(dev, dev->log_enabled);
 560     assert(t >= 0);
 561 err_features:
 562     return r;
 563 }
 564
 565 static int vhost_migration_log(MemoryListener *listener, int enable)
 566 {
 567     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 568                                          memory_listener);
 569     int r;
 570     if (!!enable == dev->log_enabled) {
 571         return 0;
 572     }
 573     if (!dev->started) {
 574         dev->log_enabled = enable;
 575         return 0;
 576     }
 577     if (!enable) {
 578         r = vhost_dev_set_log(dev, false);
 579         if (r < 0) {
 580             return r;
 581         }
 582         if (dev->log) {
 583             g_free(dev->log);
 584         }
 585         dev->log = NULL;
 586         dev->log_size = 0;
 587     } else {
 588         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
 589         r = vhost_dev_set_log(dev, true);
 590         if (r < 0) {
 591             return r;
 592         }
 593     }
 594     dev->log_enabled = enable;
 595     return 0;
 596 }
 597
 598 static void vhost_log_global_start(MemoryListener *listener)
 599 {
 600     int r;
 601
 602     r = vhost_migration_log(listener, true);
 603     if (r < 0) {
 604         abort();
 605     }
 606 }
 607
 608 static void vhost_log_global_stop(MemoryListener *listener)
 609 {
 610     int r;
 611
 612     r = vhost_migration_log(listener, false);
 613     if (r < 0) {
 614         abort();
 615     }
 616 }
 617
 618 static void vhost_log_start(MemoryListener *listener,
 619                             MemoryRegionSection *section)
 620 {
 621     /* FIXME: implement */
 622 }
 623
 624 static void vhost_log_stop(MemoryListener *listener,
 625                            MemoryRegionSection *section)
 626 {
 627     /* FIXME: implement */
 628 }
 629
 630 static int vhost_virtqueue_start(struct vhost_dev *dev,
 631                                 struct VirtIODevice *vdev,
 632                                 struct vhost_virtqueue *vq,
 633                                 unsigned idx)
 634 {
 635     hwaddr s, l, a;
 636     int r;
 637     int vhost_vq_index = idx - dev->vq_index;
 638     struct vhost_vring_file file = {
 639         .index = vhost_vq_index
 640     };
 641     struct vhost_vring_state state = {
 642         .index = vhost_vq_index
 643     };
 644     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
 645
 646     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 647
 648     vq->num = state.num = virtio_queue_get_num(vdev, idx);
 649     r = ioctl(dev->control, VHOST_SET_VRING_NUM, &state);
 650     if (r) {
 651         return -errno;
 652     }
 653
 654     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
 655     r = ioctl(dev->control, VHOST_SET_VRING_BASE, &state);
 656     if (r) {
 657         return -errno;
 658     }
 659
 660     s = l = virtio_queue_get_desc_size(vdev, idx);
 661     a = virtio_queue_get_desc_addr(vdev, idx);
 662     vq->desc = cpu_physical_memory_map(a, &l, 0);
 663     if (!vq->desc || l != s) {
 664         r = -ENOMEM;
 665         goto fail_alloc_desc;
 666     }
 667     s = l = virtio_queue_get_avail_size(vdev, idx);
 668     a = virtio_queue_get_avail_addr(vdev, idx);
 669     vq->avail = cpu_physical_memory_map(a, &l, 0);
 670     if (!vq->avail || l != s) {
 671         r = -ENOMEM;
 672         goto fail_alloc_avail;
 673     }
 674     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
 675     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
 676     vq->used = cpu_physical_memory_map(a, &l, 1);
 677     if (!vq->used || l != s) {
 678         r = -ENOMEM;
 679         goto fail_alloc_used;
 680     }
 681
 682     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
 683     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
 684     vq->ring = cpu_physical_memory_map(a, &l, 1);
 685     if (!vq->ring || l != s) {
 686         r = -ENOMEM;
 687         goto fail_alloc_ring;
 688     }
 689
 690     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
 691     if (r < 0) {
 692         r = -errno;
 693         goto fail_alloc;
 694     }
 695
 696     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
 697     r = ioctl(dev->control, VHOST_SET_VRING_KICK, &file);
 698     if (r) {
 699         r = -errno;
 700         goto fail_kick;
 701     }
 702
 703     /* Clear and discard previous events if any. */
 704     event_notifier_test_and_clear(&vq->masked_notifier);
 705
 706     return 0;
 707
 708 fail_kick:
 709 fail_alloc:
 710     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 711                               0, 0);
 712 fail_alloc_ring:
 713     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 714                               0, 0);
 715 fail_alloc_used:
 716     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 717                               0, 0);
 718 fail_alloc_avail:
 719     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 720                               0, 0);
 721 fail_alloc_desc:
 722     return r;
 723 }
 724
 725 static void vhost_virtqueue_stop(struct vhost_dev *dev,
 726                                     struct VirtIODevice *vdev,
 727                                     struct vhost_virtqueue *vq,
 728                                     unsigned idx)
 729 {
 730     struct vhost_vring_state state = {
 731         .index = idx - dev->vq_index
 732     };
 733     int r;
 734     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 735     r = ioctl(dev->control, VHOST_GET_VRING_BASE, &state);
 736     if (r < 0) {
 737         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
 738         fflush(stderr);
 739     }
 740     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
 741     assert (r >= 0);
 742     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 743                               0, virtio_queue_get_ring_size(vdev, idx));
 744     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 745                               1, virtio_queue_get_used_size(vdev, idx));
 746     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 747                               0, virtio_queue_get_avail_size(vdev, idx));
 748     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 749                               0, virtio_queue_get_desc_size(vdev, idx));
 750 }
 751
 752 static void vhost_eventfd_add(MemoryListener *listener,
 753                               MemoryRegionSection *section,
 754                               bool match_data, uint64_t data, EventNotifier *e)
 755 {
 756 }
 757
 758 static void vhost_eventfd_del(MemoryListener *listener,
 759                               MemoryRegionSection *section,
 760                               bool match_data, uint64_t data, EventNotifier *e)
 761 {
 762 }
 763
 764 static int vhost_virtqueue_init(struct vhost_dev *dev,
 765                                 struct vhost_virtqueue *vq, int n)
 766 {
 767     struct vhost_vring_file file = {
 768         .index = n,
 769     };
 770     int r = event_notifier_init(&vq->masked_notifier, 0);
 771     if (r < 0) {
 772         return r;
 773     }
 774
 775     file.fd = event_notifier_get_fd(&vq->masked_notifier);
 776     r = ioctl(dev->control, VHOST_SET_VRING_CALL, &file);
 777     if (r) {
 778         r = -errno;
 779         goto fail_call;
 780     }
 781     return 0;
 782 fail_call:
 783     event_notifier_cleanup(&vq->masked_notifier);
 784     return r;
 785 }
 786
 787 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
 788 {
 789     event_notifier_cleanup(&vq->masked_notifier);
 790 }
 791
 792 int vhost_dev_init(struct vhost_dev *hdev, int devfd, const char *devpath,
 793                    bool force)
 794 {
 795     uint64_t features;
 796     int i, r;
 797     if (devfd >= 0) {
 798         hdev->control = devfd;
 799     } else {
 800         hdev->control = open(devpath, O_RDWR);
 801         if (hdev->control < 0) {
 802             return -errno;
 803         }
 804     }
 805     r = ioctl(hdev->control, VHOST_SET_OWNER, NULL);
 806     if (r < 0) {
 807         goto fail;
 808     }
 809
 810     r = ioctl(hdev->control, VHOST_GET_FEATURES, &features);
 811     if (r < 0) {
 812         goto fail;
 813     }
 814
 815     for (i = 0; i < hdev->nvqs; ++i) {
 816         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
 817         if (r < 0) {
 818             goto fail_vq;
 819         }
 820     }
 821     hdev->features = features;
 822
 823     hdev->memory_listener = (MemoryListener) {
 824         .begin = vhost_begin,
 825         .commit = vhost_commit,
 826         .region_add = vhost_region_add,
 827         .region_del = vhost_region_del,
 828         .region_nop = vhost_region_nop,
 829         .log_start = vhost_log_start,
 830         .log_stop = vhost_log_stop,
 831         .log_sync = vhost_log_sync,
 832         .log_global_start = vhost_log_global_start,
 833         .log_global_stop = vhost_log_global_stop,
 834         .eventfd_add = vhost_eventfd_add,
 835         .eventfd_del = vhost_eventfd_del,
 836         .priority = 10
 837     };
 838     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
 839     hdev->n_mem_sections = 0;
 840     hdev->mem_sections = NULL;
 841     hdev->log = NULL;
 842     hdev->log_size = 0;
 843     hdev->log_enabled = false;
 844     hdev->started = false;
 845     memory_listener_register(&hdev->memory_listener, &address_space_memory);
 846     hdev->force = force;
 847     return 0;
 848 fail_vq:
 849     while (--i >= 0) {
 850         vhost_virtqueue_cleanup(hdev->vqs + i);
 851     }
 852 fail:
 853     r = -errno;
 854     close(hdev->control);
 855     return r;
 856 }
 857
 858 void vhost_dev_cleanup(struct vhost_dev *hdev)
 859 {
 860     int i;
 861     for (i = 0; i < hdev->nvqs; ++i) {
 862         vhost_virtqueue_cleanup(hdev->vqs + i);
 863     }
 864     memory_listener_unregister(&hdev->memory_listener);
 865     g_free(hdev->mem);
 866     g_free(hdev->mem_sections);
 867     close(hdev->control);
 868 }
 869
 870 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
 871 {
 872     return !vdev->binding->query_guest_notifiers ||
 873         vdev->binding->query_guest_notifiers(vdev->binding_opaque) ||
 874         hdev->force;
 875 }
 876
 877 /* Stop processing guest IO notifications in qemu.
 878  * Start processing them in vhost in kernel.
 879  */
 880 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 881 {
 882     int i, r;
 883     if (!vdev->binding->set_host_notifier) {
 884         fprintf(stderr, "binding does not support host notifiers\n");
 885         r = -ENOSYS;
 886         goto fail;
 887     }
 888
 889     for (i = 0; i < hdev->nvqs; ++i) {
 890         r = vdev->binding->set_host_notifier(vdev->binding_opaque,
 891                                              hdev->vq_index + i,
 892                                              true);
 893         if (r < 0) {
 894             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
 895             goto fail_vq;
 896         }
 897     }
 898
 899     return 0;
 900 fail_vq:
 901     while (--i >= 0) {
 902         r = vdev->binding->set_host_notifier(vdev->binding_opaque,
 903                                              hdev->vq_index + i,
 904                                              false);
 905         if (r < 0) {
 906             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
 907             fflush(stderr);
 908         }
 909         assert (r >= 0);
 910     }
 911 fail:
 912     return r;
 913 }
 914
 915 /* Stop processing guest IO notifications in vhost.
 916  * Start processing them in qemu.
 917  * This might actually run the qemu handlers right away,
 918  * so virtio in qemu must be completely setup when this is called.
 919  */
 920 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 921 {
 922     int i, r;
 923
 924     for (i = 0; i < hdev->nvqs; ++i) {
 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 failed: %d\n", i, -r);
 930             fflush(stderr);
 931         }
 932         assert (r >= 0);
 933     }
 934 }
 935
 936 /* Test and clear event pending status.
 937  * Should be called after unmask to avoid losing events.
 938  */
 939 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
 940 {
 941     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
 942     assert(hdev->started);
 943     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 944     return event_notifier_test_and_clear(&vq->masked_notifier);
 945 }
 946
 947 /* Mask/unmask events from this vq. */
 948 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
 949                          bool mask)
 950 {
 951     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
 952     int r, index = n - hdev->vq_index;
 953
 954     assert(hdev->started);
 955     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
 956
 957     struct vhost_vring_file file = {
 958         .index = index
 959     };
 960     if (mask) {
 961         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
 962     } else {
 963         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
 964     }
 965     r = ioctl(hdev->control, VHOST_SET_VRING_CALL, &file);
 966     assert(r >= 0);
 967 }
 968
 969 /* Host notifiers must be enabled at this point. */
 970 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
 971 {
 972     int i, r;
 973
 974     hdev->started = true;
 975
 976     r = vhost_dev_set_features(hdev, hdev->log_enabled);
 977     if (r < 0) {
 978         goto fail_features;
 979     }
 980     r = ioctl(hdev->control, VHOST_SET_MEM_TABLE, hdev->mem);
 981     if (r < 0) {
 982         r = -errno;
 983         goto fail_mem;
 984     }
 985     for (i = 0; i < hdev->nvqs; ++i) {
 986         r = vhost_virtqueue_start(hdev,
 987                                   vdev,
 988                                   hdev->vqs + i,
 989                                   hdev->vq_index + i);
 990         if (r < 0) {
 991             goto fail_vq;
 992         }
 993     }
 994
 995     if (hdev->log_enabled) {
 996         hdev->log_size = vhost_get_log_size(hdev);
 997         hdev->log = hdev->log_size ?
 998             g_malloc0(hdev->log_size * sizeof *hdev->log) : NULL;
 999         r = ioctl(hdev->control, VHOST_SET_LOG_BASE,
1000                   (uint64_t)(unsigned long)hdev->log);
1001         if (r < 0) {
1002             r = -errno;
1003             goto fail_log;
1004         }
1005     }
1006
1007     return 0;
1008 fail_log:
1009 fail_vq:
1010     while (--i >= 0) {
1011         vhost_virtqueue_stop(hdev,
1012                              vdev,
1013                              hdev->vqs + i,
1014                              hdev->vq_index + i);
1015     }
1016     i = hdev->nvqs;
1017 fail_mem:
1018 fail_features:
1019
1020     hdev->started = false;
1021     return r;
1022 }
1023
1024 /* Host notifiers must be enabled at this point. */
1025 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1026 {
1027     int i;
1028
1029     for (i = 0; i < hdev->nvqs; ++i) {
1030         vhost_virtqueue_stop(hdev,
1031                              vdev,
1032                              hdev->vqs + i,
1033                              hdev->vq_index + i);
1034     }
1035     vhost_log_sync_range(hdev, 0, ~0x0ull);
1036
1037     hdev->started = false;
1038     g_free(hdev->log);
1039     hdev->log = NULL;
1040     hdev->log_size = 0;
1041 }
1042