hw/virtio/vhost.c

   1 /*
   2  * vhost support
   3  *
   4  * Copyright Red Hat, Inc. 2010
   5  *
   6  * Authors:
   7  *  Michael S. Tsirkin <mst@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "hw/virtio/vhost.h"
  17 #include "hw/hw.h"
  18 #include "qemu/atomic.h"
  19 #include "qemu/range.h"
  20 #include <linux/vhost.h>
  21 #include "exec/address-spaces.h"
  22 #include "hw/virtio/virtio-bus.h"
  23 #include "migration/migration.h"
  24
  25 static struct vhost_log *vhost_log;
  26
  27 static void vhost_dev_sync_region(struct vhost_dev *dev,
  28                                   MemoryRegionSection *section,
  29                                   uint64_t mfirst, uint64_t mlast,
  30                                   uint64_t rfirst, uint64_t rlast)
  31 {
  32     vhost_log_chunk_t *log = dev->log->log;
  33
  34     uint64_t start = MAX(mfirst, rfirst);
  35     uint64_t end = MIN(mlast, rlast);
  36     vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK;
  37     vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1;
  38     uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK;
  39
  40     if (end < start) {
  41         return;
  42     }
  43     assert(end / VHOST_LOG_CHUNK < dev->log_size);
  44     assert(start / VHOST_LOG_CHUNK < dev->log_size);
  45
  46     for (;from < to; ++from) {
  47         vhost_log_chunk_t log;
  48         /* We first check with non-atomic: much cheaper,
  49          * and we expect non-dirty to be the common case. */
  50         if (!*from) {
  51             addr += VHOST_LOG_CHUNK;
  52             continue;
  53         }
  54         /* Data must be read atomically. We don't really need barrier semantics
  55          * but it's easier to use atomic_* than roll our own. */
  56         log = atomic_xchg(from, 0);
  57         while (log) {
  58             int bit = ctzl(log);
  59             hwaddr page_addr;
  60             hwaddr section_offset;
  61             hwaddr mr_offset;
  62             page_addr = addr + bit * VHOST_LOG_PAGE;
  63             section_offset = page_addr - section->offset_within_address_space;
  64             mr_offset = section_offset + section->offset_within_region;
  65             memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE);
  66             log &= ~(0x1ull << bit);
  67         }
  68         addr += VHOST_LOG_CHUNK;
  69     }
  70 }
  71
  72 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
  73                                    MemoryRegionSection *section,
  74                                    hwaddr first,
  75                                    hwaddr last)
  76 {
  77     int i;
  78     hwaddr start_addr;
  79     hwaddr end_addr;
  80
  81     if (!dev->log_enabled || !dev->started) {
  82         return 0;
  83     }
  84     start_addr = section->offset_within_address_space;
  85     end_addr = range_get_last(start_addr, int128_get64(section->size));
  86     start_addr = MAX(first, start_addr);
  87     end_addr = MIN(last, end_addr);
  88
  89     for (i = 0; i < dev->mem->nregions; ++i) {
  90         struct vhost_memory_region *reg = dev->mem->regions + i;
  91         vhost_dev_sync_region(dev, section, start_addr, end_addr,
  92                               reg->guest_phys_addr,
  93                               range_get_last(reg->guest_phys_addr,
  94                                              reg->memory_size));
  95     }
  96     for (i = 0; i < dev->nvqs; ++i) {
  97         struct vhost_virtqueue *vq = dev->vqs + i;
  98         vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys,
  99                               range_get_last(vq->used_phys, vq->used_size));
 100     }
 101     return 0;
 102 }
 103
 104 static void vhost_log_sync(MemoryListener *listener,
 105                           MemoryRegionSection *section)
 106 {
 107     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 108                                          memory_listener);
 109     vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL);
 110 }
 111
 112 static void vhost_log_sync_range(struct vhost_dev *dev,
 113                                  hwaddr first, hwaddr last)
 114 {
 115     int i;
 116     /* FIXME: this is N^2 in number of sections */
 117     for (i = 0; i < dev->n_mem_sections; ++i) {
 118         MemoryRegionSection *section = &dev->mem_sections[i];
 119         vhost_sync_dirty_bitmap(dev, section, first, last);
 120     }
 121 }
 122
 123 /* Assign/unassign. Keep an unsorted array of non-overlapping
 124  * memory regions in dev->mem. */
 125 static void vhost_dev_unassign_memory(struct vhost_dev *dev,
 126                                       uint64_t start_addr,
 127                                       uint64_t size)
 128 {
 129     int from, to, n = dev->mem->nregions;
 130     /* Track overlapping/split regions for sanity checking. */
 131     int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0;
 132
 133     for (from = 0, to = 0; from < n; ++from, ++to) {
 134         struct vhost_memory_region *reg = dev->mem->regions + to;
 135         uint64_t reglast;
 136         uint64_t memlast;
 137         uint64_t change;
 138
 139         /* clone old region */
 140         if (to != from) {
 141             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 142         }
 143
 144         /* No overlap is simple */
 145         if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 146                             start_addr, size)) {
 147             continue;
 148         }
 149
 150         /* Split only happens if supplied region
 151          * is in the middle of an existing one. Thus it can not
 152          * overlap with any other existing region. */
 153         assert(!split);
 154
 155         reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 156         memlast = range_get_last(start_addr, size);
 157
 158         /* Remove whole region */
 159         if (start_addr <= reg->guest_phys_addr && memlast >= reglast) {
 160             --dev->mem->nregions;
 161             --to;
 162             ++overlap_middle;
 163             continue;
 164         }
 165
 166         /* Shrink region */
 167         if (memlast >= reglast) {
 168             reg->memory_size = start_addr - reg->guest_phys_addr;
 169             assert(reg->memory_size);
 170             assert(!overlap_end);
 171             ++overlap_end;
 172             continue;
 173         }
 174
 175         /* Shift region */
 176         if (start_addr <= reg->guest_phys_addr) {
 177             change = memlast + 1 - reg->guest_phys_addr;
 178             reg->memory_size -= change;
 179             reg->guest_phys_addr += change;
 180             reg->userspace_addr += change;
 181             assert(reg->memory_size);
 182             assert(!overlap_start);
 183             ++overlap_start;
 184             continue;
 185         }
 186
 187         /* This only happens if supplied region
 188          * is in the middle of an existing one. Thus it can not
 189          * overlap with any other existing region. */
 190         assert(!overlap_start);
 191         assert(!overlap_end);
 192         assert(!overlap_middle);
 193         /* Split region: shrink first part, shift second part. */
 194         memcpy(dev->mem->regions + n, reg, sizeof *reg);
 195         reg->memory_size = start_addr - reg->guest_phys_addr;
 196         assert(reg->memory_size);
 197         change = memlast + 1 - reg->guest_phys_addr;
 198         reg = dev->mem->regions + n;
 199         reg->memory_size -= change;
 200         assert(reg->memory_size);
 201         reg->guest_phys_addr += change;
 202         reg->userspace_addr += change;
 203         /* Never add more than 1 region */
 204         assert(dev->mem->nregions == n);
 205         ++dev->mem->nregions;
 206         ++split;
 207     }
 208 }
 209
 210 /* Called after unassign, so no regions overlap the given range. */
 211 static void vhost_dev_assign_memory(struct vhost_dev *dev,
 212                                     uint64_t start_addr,
 213                                     uint64_t size,
 214                                     uint64_t uaddr)
 215 {
 216     int from, to;
 217     struct vhost_memory_region *merged = NULL;
 218     for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) {
 219         struct vhost_memory_region *reg = dev->mem->regions + to;
 220         uint64_t prlast, urlast;
 221         uint64_t pmlast, umlast;
 222         uint64_t s, e, u;
 223
 224         /* clone old region */
 225         if (to != from) {
 226             memcpy(reg, dev->mem->regions + from, sizeof *reg);
 227         }
 228         prlast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 229         pmlast = range_get_last(start_addr, size);
 230         urlast = range_get_last(reg->userspace_addr, reg->memory_size);
 231         umlast = range_get_last(uaddr, size);
 232
 233         /* check for overlapping regions: should never happen. */
 234         assert(prlast < start_addr || pmlast < reg->guest_phys_addr);
 235         /* Not an adjacent or overlapping region - do not merge. */
 236         if ((prlast + 1 != start_addr || urlast + 1 != uaddr) &&
 237             (pmlast + 1 != reg->guest_phys_addr ||
 238              umlast + 1 != reg->userspace_addr)) {
 239             continue;
 240         }
 241
 242         if (merged) {
 243             --to;
 244             assert(to >= 0);
 245         } else {
 246             merged = reg;
 247         }
 248         u = MIN(uaddr, reg->userspace_addr);
 249         s = MIN(start_addr, reg->guest_phys_addr);
 250         e = MAX(pmlast, prlast);
 251         uaddr = merged->userspace_addr = u;
 252         start_addr = merged->guest_phys_addr = s;
 253         size = merged->memory_size = e - s + 1;
 254         assert(merged->memory_size);
 255     }
 256
 257     if (!merged) {
 258         struct vhost_memory_region *reg = dev->mem->regions + to;
 259         memset(reg, 0, sizeof *reg);
 260         reg->memory_size = size;
 261         assert(reg->memory_size);
 262         reg->guest_phys_addr = start_addr;
 263         reg->userspace_addr = uaddr;
 264         ++to;
 265     }
 266     assert(to <= dev->mem->nregions + 1);
 267     dev->mem->nregions = to;
 268 }
 269
 270 static uint64_t vhost_get_log_size(struct vhost_dev *dev)
 271 {
 272     uint64_t log_size = 0;
 273     int i;
 274     for (i = 0; i < dev->mem->nregions; ++i) {
 275         struct vhost_memory_region *reg = dev->mem->regions + i;
 276         uint64_t last = range_get_last(reg->guest_phys_addr,
 277                                        reg->memory_size);
 278         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 279     }
 280     for (i = 0; i < dev->nvqs; ++i) {
 281         struct vhost_virtqueue *vq = dev->vqs + i;
 282         uint64_t last = vq->used_phys + vq->used_size - 1;
 283         log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1);
 284     }
 285     return log_size;
 286 }
 287 static struct vhost_log *vhost_log_alloc(uint64_t size)
 288 {
 289     struct vhost_log *log = g_malloc0(sizeof *log + size * sizeof(*(log->log)));
 290
 291     log->size = size;
 292     log->refcnt = 1;
 293
 294     return log;
 295 }
 296
 297 static struct vhost_log *vhost_log_get(uint64_t size)
 298 {
 299     if (!vhost_log || vhost_log->size != size) {
 300         vhost_log = vhost_log_alloc(size);
 301     } else {
 302         ++vhost_log->refcnt;
 303     }
 304
 305     return vhost_log;
 306 }
 307
 308 static void vhost_log_put(struct vhost_dev *dev, bool sync)
 309 {
 310     struct vhost_log *log = dev->log;
 311
 312     if (!log) {
 313         return;
 314     }
 315
 316     --log->refcnt;
 317     if (log->refcnt == 0) {
 318         /* Sync only the range covered by the old log */
 319         if (dev->log_size && sync) {
 320             vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1);
 321         }
 322         if (vhost_log == log) {
 323             vhost_log = NULL;
 324         }
 325         g_free(log);
 326     }
 327 }
 328
 329 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size)
 330 {
 331     struct vhost_log *log = vhost_log_get(size);
 332     uint64_t log_base = (uintptr_t)log->log;
 333     int r;
 334
 335     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_LOG_BASE, &log_base);
 336     assert(r >= 0);
 337     vhost_log_put(dev, true);
 338     dev->log = log;
 339     dev->log_size = size;
 340 }
 341
 342 static int vhost_verify_ring_mappings(struct vhost_dev *dev,
 343                                       uint64_t start_addr,
 344                                       uint64_t size)
 345 {
 346     int i;
 347     int r = 0;
 348
 349     for (i = 0; !r && i < dev->nvqs; ++i) {
 350         struct vhost_virtqueue *vq = dev->vqs + i;
 351         hwaddr l;
 352         void *p;
 353
 354         if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) {
 355             continue;
 356         }
 357         l = vq->ring_size;
 358         p = cpu_physical_memory_map(vq->ring_phys, &l, 1);
 359         if (!p || l != vq->ring_size) {
 360             fprintf(stderr, "Unable to map ring buffer for ring %d\n", i);
 361             r = -ENOMEM;
 362         }
 363         if (p != vq->ring) {
 364             fprintf(stderr, "Ring buffer relocated for ring %d\n", i);
 365             r = -EBUSY;
 366         }
 367         cpu_physical_memory_unmap(p, l, 0, 0);
 368     }
 369     return r;
 370 }
 371
 372 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev,
 373                                                       uint64_t start_addr,
 374                                                       uint64_t size)
 375 {
 376     int i, n = dev->mem->nregions;
 377     for (i = 0; i < n; ++i) {
 378         struct vhost_memory_region *reg = dev->mem->regions + i;
 379         if (ranges_overlap(reg->guest_phys_addr, reg->memory_size,
 380                            start_addr, size)) {
 381             return reg;
 382         }
 383     }
 384     return NULL;
 385 }
 386
 387 static bool vhost_dev_cmp_memory(struct vhost_dev *dev,
 388                                  uint64_t start_addr,
 389                                  uint64_t size,
 390                                  uint64_t uaddr)
 391 {
 392     struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size);
 393     uint64_t reglast;
 394     uint64_t memlast;
 395
 396     if (!reg) {
 397         return true;
 398     }
 399
 400     reglast = range_get_last(reg->guest_phys_addr, reg->memory_size);
 401     memlast = range_get_last(start_addr, size);
 402
 403     /* Need to extend region? */
 404     if (start_addr < reg->guest_phys_addr || memlast > reglast) {
 405         return true;
 406     }
 407     /* userspace_addr changed? */
 408     return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr;
 409 }
 410
 411 static void vhost_set_memory(MemoryListener *listener,
 412                              MemoryRegionSection *section,
 413                              bool add)
 414 {
 415     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 416                                          memory_listener);
 417     hwaddr start_addr = section->offset_within_address_space;
 418     ram_addr_t size = int128_get64(section->size);
 419     bool log_dirty =
 420         memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION);
 421     int s = offsetof(struct vhost_memory, regions) +
 422         (dev->mem->nregions + 1) * sizeof dev->mem->regions[0];
 423     void *ram;
 424
 425     dev->mem = g_realloc(dev->mem, s);
 426
 427     if (log_dirty) {
 428         add = false;
 429     }
 430
 431     assert(size);
 432
 433     /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */
 434     ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region;
 435     if (add) {
 436         if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) {
 437             /* Region exists with same address. Nothing to do. */
 438             return;
 439         }
 440     } else {
 441         if (!vhost_dev_find_reg(dev, start_addr, size)) {
 442             /* Removing region that we don't access. Nothing to do. */
 443             return;
 444         }
 445     }
 446
 447     vhost_dev_unassign_memory(dev, start_addr, size);
 448     if (add) {
 449         /* Add given mapping, merging adjacent regions if any */
 450         vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram);
 451     } else {
 452         /* Remove old mapping for this memory, if any. */
 453         vhost_dev_unassign_memory(dev, start_addr, size);
 454     }
 455     dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr);
 456     dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1);
 457     dev->memory_changed = true;
 458 }
 459
 460 static bool vhost_section(MemoryRegionSection *section)
 461 {
 462     return memory_region_is_ram(section->mr);
 463 }
 464
 465 static void vhost_begin(MemoryListener *listener)
 466 {
 467     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 468                                          memory_listener);
 469     dev->mem_changed_end_addr = 0;
 470     dev->mem_changed_start_addr = -1;
 471 }
 472
 473 static void vhost_commit(MemoryListener *listener)
 474 {
 475     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 476                                          memory_listener);
 477     hwaddr start_addr = 0;
 478     ram_addr_t size = 0;
 479     uint64_t log_size;
 480     int r;
 481
 482     if (!dev->memory_changed) {
 483         return;
 484     }
 485     if (!dev->started) {
 486         return;
 487     }
 488     if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) {
 489         return;
 490     }
 491
 492     if (dev->started) {
 493         start_addr = dev->mem_changed_start_addr;
 494         size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1;
 495
 496         r = vhost_verify_ring_mappings(dev, start_addr, size);
 497         assert(r >= 0);
 498     }
 499
 500     if (!dev->log_enabled) {
 501         r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
 502         assert(r >= 0);
 503         dev->memory_changed = false;
 504         return;
 505     }
 506     log_size = vhost_get_log_size(dev);
 507     /* We allocate an extra 4K bytes to log,
 508      * to reduce the * number of reallocations. */
 509 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log)
 510     /* To log more, must increase log size before table update. */
 511     if (dev->log_size < log_size) {
 512         vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER);
 513     }
 514     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem);
 515     assert(r >= 0);
 516     /* To log less, can only decrease log size after table update. */
 517     if (dev->log_size > log_size + VHOST_LOG_BUFFER) {
 518         vhost_dev_log_resize(dev, log_size);
 519     }
 520     dev->memory_changed = false;
 521 }
 522
 523 static void vhost_region_add(MemoryListener *listener,
 524                              MemoryRegionSection *section)
 525 {
 526     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 527                                          memory_listener);
 528
 529     if (!vhost_section(section)) {
 530         return;
 531     }
 532
 533     ++dev->n_mem_sections;
 534     dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections,
 535                                 dev->n_mem_sections);
 536     dev->mem_sections[dev->n_mem_sections - 1] = *section;
 537     memory_region_ref(section->mr);
 538     vhost_set_memory(listener, section, true);
 539 }
 540
 541 static void vhost_region_del(MemoryListener *listener,
 542                              MemoryRegionSection *section)
 543 {
 544     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 545                                          memory_listener);
 546     int i;
 547
 548     if (!vhost_section(section)) {
 549         return;
 550     }
 551
 552     vhost_set_memory(listener, section, false);
 553     memory_region_unref(section->mr);
 554     for (i = 0; i < dev->n_mem_sections; ++i) {
 555         if (dev->mem_sections[i].offset_within_address_space
 556             == section->offset_within_address_space) {
 557             --dev->n_mem_sections;
 558             memmove(&dev->mem_sections[i], &dev->mem_sections[i+1],
 559                     (dev->n_mem_sections - i) * sizeof(*dev->mem_sections));
 560             break;
 561         }
 562     }
 563 }
 564
 565 static void vhost_region_nop(MemoryListener *listener,
 566                              MemoryRegionSection *section)
 567 {
 568 }
 569
 570 static int vhost_virtqueue_set_addr(struct vhost_dev *dev,
 571                                     struct vhost_virtqueue *vq,
 572                                     unsigned idx, bool enable_log)
 573 {
 574     struct vhost_vring_addr addr = {
 575         .index = idx,
 576         .desc_user_addr = (uint64_t)(unsigned long)vq->desc,
 577         .avail_user_addr = (uint64_t)(unsigned long)vq->avail,
 578         .used_user_addr = (uint64_t)(unsigned long)vq->used,
 579         .log_guest_addr = vq->used_phys,
 580         .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0,
 581     };
 582     int r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ADDR, &addr);
 583     if (r < 0) {
 584         return -errno;
 585     }
 586     return 0;
 587 }
 588
 589 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log)
 590 {
 591     uint64_t features = dev->acked_features;
 592     int r;
 593     if (enable_log) {
 594         features |= 0x1ULL << VHOST_F_LOG_ALL;
 595     }
 596     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_FEATURES, &features);
 597     return r < 0 ? -errno : 0;
 598 }
 599
 600 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log)
 601 {
 602     int r, t, i;
 603     r = vhost_dev_set_features(dev, enable_log);
 604     if (r < 0) {
 605         goto err_features;
 606     }
 607     for (i = 0; i < dev->nvqs; ++i) {
 608         r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 609                                      enable_log);
 610         if (r < 0) {
 611             goto err_vq;
 612         }
 613     }
 614     return 0;
 615 err_vq:
 616     for (; i >= 0; --i) {
 617         t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i,
 618                                      dev->log_enabled);
 619         assert(t >= 0);
 620     }
 621     t = vhost_dev_set_features(dev, dev->log_enabled);
 622     assert(t >= 0);
 623 err_features:
 624     return r;
 625 }
 626
 627 static int vhost_migration_log(MemoryListener *listener, int enable)
 628 {
 629     struct vhost_dev *dev = container_of(listener, struct vhost_dev,
 630                                          memory_listener);
 631     int r;
 632     if (!!enable == dev->log_enabled) {
 633         return 0;
 634     }
 635     if (!dev->started) {
 636         dev->log_enabled = enable;
 637         return 0;
 638     }
 639     if (!enable) {
 640         r = vhost_dev_set_log(dev, false);
 641         if (r < 0) {
 642             return r;
 643         }
 644         vhost_log_put(dev, false);
 645         dev->log = NULL;
 646         dev->log_size = 0;
 647     } else {
 648         vhost_dev_log_resize(dev, vhost_get_log_size(dev));
 649         r = vhost_dev_set_log(dev, true);
 650         if (r < 0) {
 651             return r;
 652         }
 653     }
 654     dev->log_enabled = enable;
 655     return 0;
 656 }
 657
 658 static void vhost_log_global_start(MemoryListener *listener)
 659 {
 660     int r;
 661
 662     r = vhost_migration_log(listener, true);
 663     if (r < 0) {
 664         abort();
 665     }
 666 }
 667
 668 static void vhost_log_global_stop(MemoryListener *listener)
 669 {
 670     int r;
 671
 672     r = vhost_migration_log(listener, false);
 673     if (r < 0) {
 674         abort();
 675     }
 676 }
 677
 678 static void vhost_log_start(MemoryListener *listener,
 679                             MemoryRegionSection *section,
 680                             int old, int new)
 681 {
 682     /* FIXME: implement */
 683 }
 684
 685 static void vhost_log_stop(MemoryListener *listener,
 686                            MemoryRegionSection *section,
 687                            int old, int new)
 688 {
 689     /* FIXME: implement */
 690 }
 691
 692 static int vhost_virtqueue_start(struct vhost_dev *dev,
 693                                 struct VirtIODevice *vdev,
 694                                 struct vhost_virtqueue *vq,
 695                                 unsigned idx)
 696 {
 697     hwaddr s, l, a;
 698     int r;
 699     int vhost_vq_index = idx - dev->vq_index;
 700     struct vhost_vring_file file = {
 701         .index = vhost_vq_index
 702     };
 703     struct vhost_vring_state state = {
 704         .index = vhost_vq_index
 705     };
 706     struct VirtQueue *vvq = virtio_get_queue(vdev, idx);
 707
 708     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 709
 710     vq->num = state.num = virtio_queue_get_num(vdev, idx);
 711     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_NUM, &state);
 712     if (r) {
 713         return -errno;
 714     }
 715
 716     state.num = virtio_queue_get_last_avail_idx(vdev, idx);
 717     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_BASE, &state);
 718     if (r) {
 719         return -errno;
 720     }
 721
 722     s = l = virtio_queue_get_desc_size(vdev, idx);
 723     a = virtio_queue_get_desc_addr(vdev, idx);
 724     vq->desc = cpu_physical_memory_map(a, &l, 0);
 725     if (!vq->desc || l != s) {
 726         r = -ENOMEM;
 727         goto fail_alloc_desc;
 728     }
 729     s = l = virtio_queue_get_avail_size(vdev, idx);
 730     a = virtio_queue_get_avail_addr(vdev, idx);
 731     vq->avail = cpu_physical_memory_map(a, &l, 0);
 732     if (!vq->avail || l != s) {
 733         r = -ENOMEM;
 734         goto fail_alloc_avail;
 735     }
 736     vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx);
 737     vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx);
 738     vq->used = cpu_physical_memory_map(a, &l, 1);
 739     if (!vq->used || l != s) {
 740         r = -ENOMEM;
 741         goto fail_alloc_used;
 742     }
 743
 744     vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx);
 745     vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx);
 746     vq->ring = cpu_physical_memory_map(a, &l, 1);
 747     if (!vq->ring || l != s) {
 748         r = -ENOMEM;
 749         goto fail_alloc_ring;
 750     }
 751
 752     r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled);
 753     if (r < 0) {
 754         r = -errno;
 755         goto fail_alloc;
 756     }
 757
 758     file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq));
 759     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_KICK, &file);
 760     if (r) {
 761         r = -errno;
 762         goto fail_kick;
 763     }
 764
 765     /* Clear and discard previous events if any. */
 766     event_notifier_test_and_clear(&vq->masked_notifier);
 767
 768     return 0;
 769
 770 fail_kick:
 771 fail_alloc:
 772     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 773                               0, 0);
 774 fail_alloc_ring:
 775     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 776                               0, 0);
 777 fail_alloc_used:
 778     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 779                               0, 0);
 780 fail_alloc_avail:
 781     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 782                               0, 0);
 783 fail_alloc_desc:
 784     return r;
 785 }
 786
 787 static void vhost_virtqueue_stop(struct vhost_dev *dev,
 788                                     struct VirtIODevice *vdev,
 789                                     struct vhost_virtqueue *vq,
 790                                     unsigned idx)
 791 {
 792     struct vhost_vring_state state = {
 793         .index = idx - dev->vq_index
 794     };
 795     int r;
 796     assert(idx >= dev->vq_index && idx < dev->vq_index + dev->nvqs);
 797     r = dev->vhost_ops->vhost_call(dev, VHOST_GET_VRING_BASE, &state);
 798     if (r < 0) {
 799         fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r);
 800         fflush(stderr);
 801     }
 802     virtio_queue_set_last_avail_idx(vdev, idx, state.num);
 803     virtio_queue_invalidate_signalled_used(vdev, idx);
 804     assert (r >= 0);
 805     cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx),
 806                               0, virtio_queue_get_ring_size(vdev, idx));
 807     cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx),
 808                               1, virtio_queue_get_used_size(vdev, idx));
 809     cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx),
 810                               0, virtio_queue_get_avail_size(vdev, idx));
 811     cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx),
 812                               0, virtio_queue_get_desc_size(vdev, idx));
 813 }
 814
 815 static void vhost_eventfd_add(MemoryListener *listener,
 816                               MemoryRegionSection *section,
 817                               bool match_data, uint64_t data, EventNotifier *e)
 818 {
 819 }
 820
 821 static void vhost_eventfd_del(MemoryListener *listener,
 822                               MemoryRegionSection *section,
 823                               bool match_data, uint64_t data, EventNotifier *e)
 824 {
 825 }
 826
 827 static int vhost_virtqueue_init(struct vhost_dev *dev,
 828                                 struct vhost_virtqueue *vq, int n)
 829 {
 830     struct vhost_vring_file file = {
 831         .index = n,
 832     };
 833     int r = event_notifier_init(&vq->masked_notifier, 0);
 834     if (r < 0) {
 835         return r;
 836     }
 837
 838     file.fd = event_notifier_get_fd(&vq->masked_notifier);
 839     r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_CALL, &file);
 840     if (r) {
 841         r = -errno;
 842         goto fail_call;
 843     }
 844     return 0;
 845 fail_call:
 846     event_notifier_cleanup(&vq->masked_notifier);
 847     return r;
 848 }
 849
 850 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq)
 851 {
 852     event_notifier_cleanup(&vq->masked_notifier);
 853 }
 854
 855 int vhost_dev_init(struct vhost_dev *hdev, void *opaque,
 856                    VhostBackendType backend_type, bool force)
 857 {
 858     uint64_t features;
 859     int i, r;
 860
 861     if (vhost_set_backend_type(hdev, backend_type) < 0) {
 862         close((uintptr_t)opaque);
 863         return -1;
 864     }
 865
 866     if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) {
 867         close((uintptr_t)opaque);
 868         return -errno;
 869     }
 870
 871     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_OWNER, NULL);
 872     if (r < 0) {
 873         goto fail;
 874     }
 875
 876     r = hdev->vhost_ops->vhost_call(hdev, VHOST_GET_FEATURES, &features);
 877     if (r < 0) {
 878         goto fail;
 879     }
 880
 881     for (i = 0; i < hdev->nvqs; ++i) {
 882         r = vhost_virtqueue_init(hdev, hdev->vqs + i, i);
 883         if (r < 0) {
 884             goto fail_vq;
 885         }
 886     }
 887     hdev->features = features;
 888
 889     hdev->memory_listener = (MemoryListener) {
 890         .begin = vhost_begin,
 891         .commit = vhost_commit,
 892         .region_add = vhost_region_add,
 893         .region_del = vhost_region_del,
 894         .region_nop = vhost_region_nop,
 895         .log_start = vhost_log_start,
 896         .log_stop = vhost_log_stop,
 897         .log_sync = vhost_log_sync,
 898         .log_global_start = vhost_log_global_start,
 899         .log_global_stop = vhost_log_global_stop,
 900         .eventfd_add = vhost_eventfd_add,
 901         .eventfd_del = vhost_eventfd_del,
 902         .priority = 10
 903     };
 904     hdev->migration_blocker = NULL;
 905     if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) {
 906         error_setg(&hdev->migration_blocker,
 907                    "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature.");
 908         migrate_add_blocker(hdev->migration_blocker);
 909     }
 910     hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions));
 911     hdev->n_mem_sections = 0;
 912     hdev->mem_sections = NULL;
 913     hdev->log = NULL;
 914     hdev->log_size = 0;
 915     hdev->log_enabled = false;
 916     hdev->started = false;
 917     hdev->memory_changed = false;
 918     memory_listener_register(&hdev->memory_listener, &address_space_memory);
 919     hdev->force = force;
 920     return 0;
 921 fail_vq:
 922     while (--i >= 0) {
 923         vhost_virtqueue_cleanup(hdev->vqs + i);
 924     }
 925 fail:
 926     r = -errno;
 927     hdev->vhost_ops->vhost_backend_cleanup(hdev);
 928     return r;
 929 }
 930
 931 void vhost_dev_cleanup(struct vhost_dev *hdev)
 932 {
 933     int i;
 934     for (i = 0; i < hdev->nvqs; ++i) {
 935         vhost_virtqueue_cleanup(hdev->vqs + i);
 936     }
 937     memory_listener_unregister(&hdev->memory_listener);
 938     if (hdev->migration_blocker) {
 939         migrate_del_blocker(hdev->migration_blocker);
 940         error_free(hdev->migration_blocker);
 941     }
 942     g_free(hdev->mem);
 943     g_free(hdev->mem_sections);
 944     hdev->vhost_ops->vhost_backend_cleanup(hdev);
 945 }
 946
 947 bool vhost_dev_query(struct vhost_dev *hdev, VirtIODevice *vdev)
 948 {
 949     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 950     VirtioBusState *vbus = VIRTIO_BUS(qbus);
 951     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
 952
 953     return !k->query_guest_notifiers ||
 954            k->query_guest_notifiers(qbus->parent) ||
 955            hdev->force;
 956 }
 957
 958 /* Stop processing guest IO notifications in qemu.
 959  * Start processing them in vhost in kernel.
 960  */
 961 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
 962 {
 963     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 964     VirtioBusState *vbus = VIRTIO_BUS(qbus);
 965     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
 966     int i, r;
 967     if (!k->set_host_notifier) {
 968         fprintf(stderr, "binding does not support host notifiers\n");
 969         r = -ENOSYS;
 970         goto fail;
 971     }
 972
 973     for (i = 0; i < hdev->nvqs; ++i) {
 974         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true);
 975         if (r < 0) {
 976             fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r);
 977             goto fail_vq;
 978         }
 979     }
 980
 981     return 0;
 982 fail_vq:
 983     while (--i >= 0) {
 984         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
 985         if (r < 0) {
 986             fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r);
 987             fflush(stderr);
 988         }
 989         assert (r >= 0);
 990     }
 991 fail:
 992     return r;
 993 }
 994
 995 /* Stop processing guest IO notifications in vhost.
 996  * Start processing them in qemu.
 997  * This might actually run the qemu handlers right away,
 998  * so virtio in qemu must be completely setup when this is called.
 999  */
1000 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
1001 {
1002     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
1003     VirtioBusState *vbus = VIRTIO_BUS(qbus);
1004     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus);
1005     int i, r;
1006
1007     for (i = 0; i < hdev->nvqs; ++i) {
1008         r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false);
1009         if (r < 0) {
1010             fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r);
1011             fflush(stderr);
1012         }
1013         assert (r >= 0);
1014     }
1015 }
1016
1017 /* Test and clear event pending status.
1018  * Should be called after unmask to avoid losing events.
1019  */
1020 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n)
1021 {
1022     struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index;
1023     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1024     return event_notifier_test_and_clear(&vq->masked_notifier);
1025 }
1026
1027 /* Mask/unmask events from this vq. */
1028 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n,
1029                          bool mask)
1030 {
1031     struct VirtQueue *vvq = virtio_get_queue(vdev, n);
1032     int r, index = n - hdev->vq_index;
1033
1034     assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs);
1035
1036     struct vhost_vring_file file = {
1037         .index = index
1038     };
1039     if (mask) {
1040         file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier);
1041     } else {
1042         file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq));
1043     }
1044     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_VRING_CALL, &file);
1045     assert(r >= 0);
1046 }
1047
1048 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits,
1049                             uint64_t features)
1050 {
1051     const int *bit = feature_bits;
1052     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1053         uint64_t bit_mask = (1ULL << *bit);
1054         if (!(hdev->features & bit_mask)) {
1055             features &= ~bit_mask;
1056         }
1057         bit++;
1058     }
1059     return features;
1060 }
1061
1062 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits,
1063                         uint64_t features)
1064 {
1065     const int *bit = feature_bits;
1066     while (*bit != VHOST_INVALID_FEATURE_BIT) {
1067         uint64_t bit_mask = (1ULL << *bit);
1068         if (features & bit_mask) {
1069             hdev->acked_features |= bit_mask;
1070         }
1071         bit++;
1072     }
1073 }
1074
1075 /* Host notifiers must be enabled at this point. */
1076 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev)
1077 {
1078     int i, r;
1079
1080     hdev->started = true;
1081
1082     r = vhost_dev_set_features(hdev, hdev->log_enabled);
1083     if (r < 0) {
1084         goto fail_features;
1085     }
1086     r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_MEM_TABLE, hdev->mem);
1087     if (r < 0) {
1088         r = -errno;
1089         goto fail_mem;
1090     }
1091     for (i = 0; i < hdev->nvqs; ++i) {
1092         r = vhost_virtqueue_start(hdev,
1093                                   vdev,
1094                                   hdev->vqs + i,
1095                                   hdev->vq_index + i);
1096         if (r < 0) {
1097             goto fail_vq;
1098         }
1099     }
1100
1101     if (hdev->log_enabled) {
1102         uint64_t log_base;
1103
1104         hdev->log_size = vhost_get_log_size(hdev);
1105         hdev->log = vhost_log_get(hdev->log_size);
1106         log_base = (uintptr_t)hdev->log->log;
1107         r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_LOG_BASE,
1108                                         hdev->log_size ? &log_base : NULL);
1109         if (r < 0) {
1110             r = -errno;
1111             goto fail_log;
1112         }
1113     }
1114
1115     return 0;
1116 fail_log:
1117     vhost_log_put(hdev, false);
1118 fail_vq:
1119     while (--i >= 0) {
1120         vhost_virtqueue_stop(hdev,
1121                              vdev,
1122                              hdev->vqs + i,
1123                              hdev->vq_index + i);
1124     }
1125     i = hdev->nvqs;
1126 fail_mem:
1127 fail_features:
1128
1129     hdev->started = false;
1130     return r;
1131 }
1132
1133 /* Host notifiers must be enabled at this point. */
1134 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev)
1135 {
1136     int i;
1137
1138     for (i = 0; i < hdev->nvqs; ++i) {
1139         vhost_virtqueue_stop(hdev,
1140                              vdev,
1141                              hdev->vqs + i,
1142                              hdev->vq_index + i);
1143     }
1144
1145     vhost_log_put(hdev, true);
1146     hdev->started = false;
1147     hdev->log = NULL;
1148     hdev->log_size = 0;
1149 }
1150