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 "qemu/error-report.h" 21 #include <linux/vhost.h> 22 #include "exec/address-spaces.h" 23 #include "hw/virtio/virtio-bus.h" 24 #include "hw/virtio/virtio-access.h" 25 #include "migration/migration.h" 26 27 static struct vhost_log *vhost_log; 28 29 static unsigned int used_memslots; 30 static QLIST_HEAD(, vhost_dev) vhost_devices = 31 QLIST_HEAD_INITIALIZER(vhost_devices); 32 33 bool vhost_has_free_slot(void) 34 { 35 unsigned int slots_limit = ~0U; 36 struct vhost_dev *hdev; 37 38 QLIST_FOREACH(hdev, &vhost_devices, entry) { 39 unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev); 40 slots_limit = MIN(slots_limit, r); 41 } 42 return slots_limit > used_memslots; 43 } 44 45 static void vhost_dev_sync_region(struct vhost_dev *dev, 46 MemoryRegionSection *section, 47 uint64_t mfirst, uint64_t mlast, 48 uint64_t rfirst, uint64_t rlast) 49 { 50 vhost_log_chunk_t *log = dev->log->log; 51 52 uint64_t start = MAX(mfirst, rfirst); 53 uint64_t end = MIN(mlast, rlast); 54 vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK; 55 vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1; 56 uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK; 57 58 if (end < start) { 59 return; 60 } 61 assert(end / VHOST_LOG_CHUNK < dev->log_size); 62 assert(start / VHOST_LOG_CHUNK < dev->log_size); 63 64 for (;from < to; ++from) { 65 vhost_log_chunk_t log; 66 /* We first check with non-atomic: much cheaper, 67 * and we expect non-dirty to be the common case. */ 68 if (!*from) { 69 addr += VHOST_LOG_CHUNK; 70 continue; 71 } 72 /* Data must be read atomically. We don't really need barrier semantics 73 * but it's easier to use atomic_* than roll our own. */ 74 log = atomic_xchg(from, 0); 75 while (log) { 76 int bit = ctzl(log); 77 hwaddr page_addr; 78 hwaddr section_offset; 79 hwaddr mr_offset; 80 page_addr = addr + bit * VHOST_LOG_PAGE; 81 section_offset = page_addr - section->offset_within_address_space; 82 mr_offset = section_offset + section->offset_within_region; 83 memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE); 84 log &= ~(0x1ull << bit); 85 } 86 addr += VHOST_LOG_CHUNK; 87 } 88 } 89 90 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev, 91 MemoryRegionSection *section, 92 hwaddr first, 93 hwaddr last) 94 { 95 int i; 96 hwaddr start_addr; 97 hwaddr end_addr; 98 99 if (!dev->log_enabled || !dev->started) { 100 return 0; 101 } 102 start_addr = section->offset_within_address_space; 103 end_addr = range_get_last(start_addr, int128_get64(section->size)); 104 start_addr = MAX(first, start_addr); 105 end_addr = MIN(last, end_addr); 106 107 for (i = 0; i < dev->mem->nregions; ++i) { 108 struct vhost_memory_region *reg = dev->mem->regions + i; 109 vhost_dev_sync_region(dev, section, start_addr, end_addr, 110 reg->guest_phys_addr, 111 range_get_last(reg->guest_phys_addr, 112 reg->memory_size)); 113 } 114 for (i = 0; i < dev->nvqs; ++i) { 115 struct vhost_virtqueue *vq = dev->vqs + i; 116 vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys, 117 range_get_last(vq->used_phys, vq->used_size)); 118 } 119 return 0; 120 } 121 122 static void vhost_log_sync(MemoryListener *listener, 123 MemoryRegionSection *section) 124 { 125 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 126 memory_listener); 127 vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL); 128 } 129 130 static void vhost_log_sync_range(struct vhost_dev *dev, 131 hwaddr first, hwaddr last) 132 { 133 int i; 134 /* FIXME: this is N^2 in number of sections */ 135 for (i = 0; i < dev->n_mem_sections; ++i) { 136 MemoryRegionSection *section = &dev->mem_sections[i]; 137 vhost_sync_dirty_bitmap(dev, section, first, last); 138 } 139 } 140 141 /* Assign/unassign. Keep an unsorted array of non-overlapping 142 * memory regions in dev->mem. */ 143 static void vhost_dev_unassign_memory(struct vhost_dev *dev, 144 uint64_t start_addr, 145 uint64_t size) 146 { 147 int from, to, n = dev->mem->nregions; 148 /* Track overlapping/split regions for sanity checking. */ 149 int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0; 150 151 for (from = 0, to = 0; from < n; ++from, ++to) { 152 struct vhost_memory_region *reg = dev->mem->regions + to; 153 uint64_t reglast; 154 uint64_t memlast; 155 uint64_t change; 156 157 /* clone old region */ 158 if (to != from) { 159 memcpy(reg, dev->mem->regions + from, sizeof *reg); 160 } 161 162 /* No overlap is simple */ 163 if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size, 164 start_addr, size)) { 165 continue; 166 } 167 168 /* Split only happens if supplied region 169 * is in the middle of an existing one. Thus it can not 170 * overlap with any other existing region. */ 171 assert(!split); 172 173 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); 174 memlast = range_get_last(start_addr, size); 175 176 /* Remove whole region */ 177 if (start_addr <= reg->guest_phys_addr && memlast >= reglast) { 178 --dev->mem->nregions; 179 --to; 180 ++overlap_middle; 181 continue; 182 } 183 184 /* Shrink region */ 185 if (memlast >= reglast) { 186 reg->memory_size = start_addr - reg->guest_phys_addr; 187 assert(reg->memory_size); 188 assert(!overlap_end); 189 ++overlap_end; 190 continue; 191 } 192 193 /* Shift region */ 194 if (start_addr <= reg->guest_phys_addr) { 195 change = memlast + 1 - reg->guest_phys_addr; 196 reg->memory_size -= change; 197 reg->guest_phys_addr += change; 198 reg->userspace_addr += change; 199 assert(reg->memory_size); 200 assert(!overlap_start); 201 ++overlap_start; 202 continue; 203 } 204 205 /* This only happens if supplied region 206 * is in the middle of an existing one. Thus it can not 207 * overlap with any other existing region. */ 208 assert(!overlap_start); 209 assert(!overlap_end); 210 assert(!overlap_middle); 211 /* Split region: shrink first part, shift second part. */ 212 memcpy(dev->mem->regions + n, reg, sizeof *reg); 213 reg->memory_size = start_addr - reg->guest_phys_addr; 214 assert(reg->memory_size); 215 change = memlast + 1 - reg->guest_phys_addr; 216 reg = dev->mem->regions + n; 217 reg->memory_size -= change; 218 assert(reg->memory_size); 219 reg->guest_phys_addr += change; 220 reg->userspace_addr += change; 221 /* Never add more than 1 region */ 222 assert(dev->mem->nregions == n); 223 ++dev->mem->nregions; 224 ++split; 225 } 226 } 227 228 /* Called after unassign, so no regions overlap the given range. */ 229 static void vhost_dev_assign_memory(struct vhost_dev *dev, 230 uint64_t start_addr, 231 uint64_t size, 232 uint64_t uaddr) 233 { 234 int from, to; 235 struct vhost_memory_region *merged = NULL; 236 for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) { 237 struct vhost_memory_region *reg = dev->mem->regions + to; 238 uint64_t prlast, urlast; 239 uint64_t pmlast, umlast; 240 uint64_t s, e, u; 241 242 /* clone old region */ 243 if (to != from) { 244 memcpy(reg, dev->mem->regions + from, sizeof *reg); 245 } 246 prlast = range_get_last(reg->guest_phys_addr, reg->memory_size); 247 pmlast = range_get_last(start_addr, size); 248 urlast = range_get_last(reg->userspace_addr, reg->memory_size); 249 umlast = range_get_last(uaddr, size); 250 251 /* check for overlapping regions: should never happen. */ 252 assert(prlast < start_addr || pmlast < reg->guest_phys_addr); 253 /* Not an adjacent or overlapping region - do not merge. */ 254 if ((prlast + 1 != start_addr || urlast + 1 != uaddr) && 255 (pmlast + 1 != reg->guest_phys_addr || 256 umlast + 1 != reg->userspace_addr)) { 257 continue; 258 } 259 260 if (merged) { 261 --to; 262 assert(to >= 0); 263 } else { 264 merged = reg; 265 } 266 u = MIN(uaddr, reg->userspace_addr); 267 s = MIN(start_addr, reg->guest_phys_addr); 268 e = MAX(pmlast, prlast); 269 uaddr = merged->userspace_addr = u; 270 start_addr = merged->guest_phys_addr = s; 271 size = merged->memory_size = e - s + 1; 272 assert(merged->memory_size); 273 } 274 275 if (!merged) { 276 struct vhost_memory_region *reg = dev->mem->regions + to; 277 memset(reg, 0, sizeof *reg); 278 reg->memory_size = size; 279 assert(reg->memory_size); 280 reg->guest_phys_addr = start_addr; 281 reg->userspace_addr = uaddr; 282 ++to; 283 } 284 assert(to <= dev->mem->nregions + 1); 285 dev->mem->nregions = to; 286 } 287 288 static uint64_t vhost_get_log_size(struct vhost_dev *dev) 289 { 290 uint64_t log_size = 0; 291 int i; 292 for (i = 0; i < dev->mem->nregions; ++i) { 293 struct vhost_memory_region *reg = dev->mem->regions + i; 294 uint64_t last = range_get_last(reg->guest_phys_addr, 295 reg->memory_size); 296 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 297 } 298 for (i = 0; i < dev->nvqs; ++i) { 299 struct vhost_virtqueue *vq = dev->vqs + i; 300 uint64_t last = vq->used_phys + vq->used_size - 1; 301 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 302 } 303 return log_size; 304 } 305 static struct vhost_log *vhost_log_alloc(uint64_t size) 306 { 307 struct vhost_log *log = g_malloc0(sizeof *log + size * sizeof(*(log->log))); 308 309 log->size = size; 310 log->refcnt = 1; 311 312 return log; 313 } 314 315 static struct vhost_log *vhost_log_get(uint64_t size) 316 { 317 if (!vhost_log || vhost_log->size != size) { 318 vhost_log = vhost_log_alloc(size); 319 } else { 320 ++vhost_log->refcnt; 321 } 322 323 return vhost_log; 324 } 325 326 static void vhost_log_put(struct vhost_dev *dev, bool sync) 327 { 328 struct vhost_log *log = dev->log; 329 330 if (!log) { 331 return; 332 } 333 334 --log->refcnt; 335 if (log->refcnt == 0) { 336 /* Sync only the range covered by the old log */ 337 if (dev->log_size && sync) { 338 vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1); 339 } 340 if (vhost_log == log) { 341 vhost_log = NULL; 342 } 343 g_free(log); 344 } 345 } 346 347 static inline void vhost_dev_log_resize(struct vhost_dev* dev, uint64_t size) 348 { 349 struct vhost_log *log = vhost_log_get(size); 350 uint64_t log_base = (uintptr_t)log->log; 351 int r; 352 353 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_LOG_BASE, &log_base); 354 assert(r >= 0); 355 vhost_log_put(dev, true); 356 dev->log = log; 357 dev->log_size = size; 358 } 359 360 static int vhost_verify_ring_mappings(struct vhost_dev *dev, 361 uint64_t start_addr, 362 uint64_t size) 363 { 364 int i; 365 int r = 0; 366 367 for (i = 0; !r && i < dev->nvqs; ++i) { 368 struct vhost_virtqueue *vq = dev->vqs + i; 369 hwaddr l; 370 void *p; 371 372 if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) { 373 continue; 374 } 375 l = vq->ring_size; 376 p = cpu_physical_memory_map(vq->ring_phys, &l, 1); 377 if (!p || l != vq->ring_size) { 378 fprintf(stderr, "Unable to map ring buffer for ring %d\n", i); 379 r = -ENOMEM; 380 } 381 if (p != vq->ring) { 382 fprintf(stderr, "Ring buffer relocated for ring %d\n", i); 383 r = -EBUSY; 384 } 385 cpu_physical_memory_unmap(p, l, 0, 0); 386 } 387 return r; 388 } 389 390 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev, 391 uint64_t start_addr, 392 uint64_t size) 393 { 394 int i, n = dev->mem->nregions; 395 for (i = 0; i < n; ++i) { 396 struct vhost_memory_region *reg = dev->mem->regions + i; 397 if (ranges_overlap(reg->guest_phys_addr, reg->memory_size, 398 start_addr, size)) { 399 return reg; 400 } 401 } 402 return NULL; 403 } 404 405 static bool vhost_dev_cmp_memory(struct vhost_dev *dev, 406 uint64_t start_addr, 407 uint64_t size, 408 uint64_t uaddr) 409 { 410 struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size); 411 uint64_t reglast; 412 uint64_t memlast; 413 414 if (!reg) { 415 return true; 416 } 417 418 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); 419 memlast = range_get_last(start_addr, size); 420 421 /* Need to extend region? */ 422 if (start_addr < reg->guest_phys_addr || memlast > reglast) { 423 return true; 424 } 425 /* userspace_addr changed? */ 426 return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr; 427 } 428 429 static void vhost_set_memory(MemoryListener *listener, 430 MemoryRegionSection *section, 431 bool add) 432 { 433 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 434 memory_listener); 435 hwaddr start_addr = section->offset_within_address_space; 436 ram_addr_t size = int128_get64(section->size); 437 bool log_dirty = 438 memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION); 439 int s = offsetof(struct vhost_memory, regions) + 440 (dev->mem->nregions + 1) * sizeof dev->mem->regions[0]; 441 void *ram; 442 443 dev->mem = g_realloc(dev->mem, s); 444 445 if (log_dirty) { 446 add = false; 447 } 448 449 assert(size); 450 451 /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */ 452 ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region; 453 if (add) { 454 if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) { 455 /* Region exists with same address. Nothing to do. */ 456 return; 457 } 458 } else { 459 if (!vhost_dev_find_reg(dev, start_addr, size)) { 460 /* Removing region that we don't access. Nothing to do. */ 461 return; 462 } 463 } 464 465 vhost_dev_unassign_memory(dev, start_addr, size); 466 if (add) { 467 /* Add given mapping, merging adjacent regions if any */ 468 vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram); 469 } else { 470 /* Remove old mapping for this memory, if any. */ 471 vhost_dev_unassign_memory(dev, start_addr, size); 472 } 473 dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr); 474 dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1); 475 dev->memory_changed = true; 476 used_memslots = dev->mem->nregions; 477 } 478 479 static bool vhost_section(MemoryRegionSection *section) 480 { 481 return memory_region_is_ram(section->mr); 482 } 483 484 static void vhost_begin(MemoryListener *listener) 485 { 486 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 487 memory_listener); 488 dev->mem_changed_end_addr = 0; 489 dev->mem_changed_start_addr = -1; 490 } 491 492 static void vhost_commit(MemoryListener *listener) 493 { 494 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 495 memory_listener); 496 hwaddr start_addr = 0; 497 ram_addr_t size = 0; 498 uint64_t log_size; 499 int r; 500 501 if (!dev->memory_changed) { 502 return; 503 } 504 if (!dev->started) { 505 return; 506 } 507 if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) { 508 return; 509 } 510 511 if (dev->started) { 512 start_addr = dev->mem_changed_start_addr; 513 size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1; 514 515 r = vhost_verify_ring_mappings(dev, start_addr, size); 516 assert(r >= 0); 517 } 518 519 if (!dev->log_enabled) { 520 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem); 521 assert(r >= 0); 522 dev->memory_changed = false; 523 return; 524 } 525 log_size = vhost_get_log_size(dev); 526 /* We allocate an extra 4K bytes to log, 527 * to reduce the * number of reallocations. */ 528 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log) 529 /* To log more, must increase log size before table update. */ 530 if (dev->log_size < log_size) { 531 vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER); 532 } 533 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_MEM_TABLE, dev->mem); 534 assert(r >= 0); 535 /* To log less, can only decrease log size after table update. */ 536 if (dev->log_size > log_size + VHOST_LOG_BUFFER) { 537 vhost_dev_log_resize(dev, log_size); 538 } 539 dev->memory_changed = false; 540 } 541 542 static void vhost_region_add(MemoryListener *listener, 543 MemoryRegionSection *section) 544 { 545 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 546 memory_listener); 547 548 if (!vhost_section(section)) { 549 return; 550 } 551 552 ++dev->n_mem_sections; 553 dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections, 554 dev->n_mem_sections); 555 dev->mem_sections[dev->n_mem_sections - 1] = *section; 556 memory_region_ref(section->mr); 557 vhost_set_memory(listener, section, true); 558 } 559 560 static void vhost_region_del(MemoryListener *listener, 561 MemoryRegionSection *section) 562 { 563 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 564 memory_listener); 565 int i; 566 567 if (!vhost_section(section)) { 568 return; 569 } 570 571 vhost_set_memory(listener, section, false); 572 memory_region_unref(section->mr); 573 for (i = 0; i < dev->n_mem_sections; ++i) { 574 if (dev->mem_sections[i].offset_within_address_space 575 == section->offset_within_address_space) { 576 --dev->n_mem_sections; 577 memmove(&dev->mem_sections[i], &dev->mem_sections[i+1], 578 (dev->n_mem_sections - i) * sizeof(*dev->mem_sections)); 579 break; 580 } 581 } 582 } 583 584 static void vhost_region_nop(MemoryListener *listener, 585 MemoryRegionSection *section) 586 { 587 } 588 589 static int vhost_virtqueue_set_addr(struct vhost_dev *dev, 590 struct vhost_virtqueue *vq, 591 unsigned idx, bool enable_log) 592 { 593 struct vhost_vring_addr addr = { 594 .index = idx, 595 .desc_user_addr = (uint64_t)(unsigned long)vq->desc, 596 .avail_user_addr = (uint64_t)(unsigned long)vq->avail, 597 .used_user_addr = (uint64_t)(unsigned long)vq->used, 598 .log_guest_addr = vq->used_phys, 599 .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0, 600 }; 601 int r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ADDR, &addr); 602 if (r < 0) { 603 return -errno; 604 } 605 return 0; 606 } 607 608 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log) 609 { 610 uint64_t features = dev->acked_features; 611 int r; 612 if (enable_log) { 613 features |= 0x1ULL << VHOST_F_LOG_ALL; 614 } 615 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_FEATURES, &features); 616 return r < 0 ? -errno : 0; 617 } 618 619 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log) 620 { 621 int r, t, i; 622 r = vhost_dev_set_features(dev, enable_log); 623 if (r < 0) { 624 goto err_features; 625 } 626 for (i = 0; i < dev->nvqs; ++i) { 627 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, i, 628 enable_log); 629 if (r < 0) { 630 goto err_vq; 631 } 632 } 633 return 0; 634 err_vq: 635 for (; i >= 0; --i) { 636 t = vhost_virtqueue_set_addr(dev, dev->vqs + i, i, 637 dev->log_enabled); 638 assert(t >= 0); 639 } 640 t = vhost_dev_set_features(dev, dev->log_enabled); 641 assert(t >= 0); 642 err_features: 643 return r; 644 } 645 646 static int vhost_migration_log(MemoryListener *listener, int enable) 647 { 648 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 649 memory_listener); 650 int r; 651 if (!!enable == dev->log_enabled) { 652 return 0; 653 } 654 if (!dev->started) { 655 dev->log_enabled = enable; 656 return 0; 657 } 658 if (!enable) { 659 r = vhost_dev_set_log(dev, false); 660 if (r < 0) { 661 return r; 662 } 663 vhost_log_put(dev, false); 664 dev->log = NULL; 665 dev->log_size = 0; 666 } else { 667 vhost_dev_log_resize(dev, vhost_get_log_size(dev)); 668 r = vhost_dev_set_log(dev, true); 669 if (r < 0) { 670 return r; 671 } 672 } 673 dev->log_enabled = enable; 674 return 0; 675 } 676 677 static void vhost_log_global_start(MemoryListener *listener) 678 { 679 int r; 680 681 r = vhost_migration_log(listener, true); 682 if (r < 0) { 683 abort(); 684 } 685 } 686 687 static void vhost_log_global_stop(MemoryListener *listener) 688 { 689 int r; 690 691 r = vhost_migration_log(listener, false); 692 if (r < 0) { 693 abort(); 694 } 695 } 696 697 static void vhost_log_start(MemoryListener *listener, 698 MemoryRegionSection *section, 699 int old, int new) 700 { 701 /* FIXME: implement */ 702 } 703 704 static void vhost_log_stop(MemoryListener *listener, 705 MemoryRegionSection *section, 706 int old, int new) 707 { 708 /* FIXME: implement */ 709 } 710 711 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev, 712 bool is_big_endian, 713 int vhost_vq_index) 714 { 715 struct vhost_vring_state s = { 716 .index = vhost_vq_index, 717 .num = is_big_endian 718 }; 719 720 if (!dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_ENDIAN, &s)) { 721 return 0; 722 } 723 724 if (errno == ENOTTY) { 725 error_report("vhost does not support cross-endian"); 726 return -ENOSYS; 727 } 728 729 return -errno; 730 } 731 732 static int vhost_virtqueue_start(struct vhost_dev *dev, 733 struct VirtIODevice *vdev, 734 struct vhost_virtqueue *vq, 735 unsigned idx) 736 { 737 hwaddr s, l, a; 738 int r; 739 int vhost_vq_index = dev->vhost_ops->vhost_backend_get_vq_index(dev, idx); 740 struct vhost_vring_file file = { 741 .index = vhost_vq_index 742 }; 743 struct vhost_vring_state state = { 744 .index = vhost_vq_index 745 }; 746 struct VirtQueue *vvq = virtio_get_queue(vdev, idx); 747 748 749 vq->num = state.num = virtio_queue_get_num(vdev, idx); 750 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_NUM, &state); 751 if (r) { 752 return -errno; 753 } 754 755 state.num = virtio_queue_get_last_avail_idx(vdev, idx); 756 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_BASE, &state); 757 if (r) { 758 return -errno; 759 } 760 761 if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) && 762 virtio_legacy_is_cross_endian(vdev)) { 763 r = vhost_virtqueue_set_vring_endian_legacy(dev, 764 virtio_is_big_endian(vdev), 765 vhost_vq_index); 766 if (r) { 767 return -errno; 768 } 769 } 770 771 s = l = virtio_queue_get_desc_size(vdev, idx); 772 a = virtio_queue_get_desc_addr(vdev, idx); 773 vq->desc = cpu_physical_memory_map(a, &l, 0); 774 if (!vq->desc || l != s) { 775 r = -ENOMEM; 776 goto fail_alloc_desc; 777 } 778 s = l = virtio_queue_get_avail_size(vdev, idx); 779 a = virtio_queue_get_avail_addr(vdev, idx); 780 vq->avail = cpu_physical_memory_map(a, &l, 0); 781 if (!vq->avail || l != s) { 782 r = -ENOMEM; 783 goto fail_alloc_avail; 784 } 785 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx); 786 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx); 787 vq->used = cpu_physical_memory_map(a, &l, 1); 788 if (!vq->used || l != s) { 789 r = -ENOMEM; 790 goto fail_alloc_used; 791 } 792 793 vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx); 794 vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx); 795 vq->ring = cpu_physical_memory_map(a, &l, 1); 796 if (!vq->ring || l != s) { 797 r = -ENOMEM; 798 goto fail_alloc_ring; 799 } 800 801 r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled); 802 if (r < 0) { 803 r = -errno; 804 goto fail_alloc; 805 } 806 807 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq)); 808 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_KICK, &file); 809 if (r) { 810 r = -errno; 811 goto fail_kick; 812 } 813 814 /* Clear and discard previous events if any. */ 815 event_notifier_test_and_clear(&vq->masked_notifier); 816 817 return 0; 818 819 fail_kick: 820 fail_alloc: 821 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), 822 0, 0); 823 fail_alloc_ring: 824 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), 825 0, 0); 826 fail_alloc_used: 827 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), 828 0, 0); 829 fail_alloc_avail: 830 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), 831 0, 0); 832 fail_alloc_desc: 833 return r; 834 } 835 836 static void vhost_virtqueue_stop(struct vhost_dev *dev, 837 struct VirtIODevice *vdev, 838 struct vhost_virtqueue *vq, 839 unsigned idx) 840 { 841 int vhost_vq_index = dev->vhost_ops->vhost_backend_get_vq_index(dev, idx); 842 struct vhost_vring_state state = { 843 .index = vhost_vq_index, 844 }; 845 int r; 846 847 r = dev->vhost_ops->vhost_call(dev, VHOST_GET_VRING_BASE, &state); 848 if (r < 0) { 849 fprintf(stderr, "vhost VQ %d ring restore failed: %d\n", idx, r); 850 fflush(stderr); 851 } 852 virtio_queue_set_last_avail_idx(vdev, idx, state.num); 853 virtio_queue_invalidate_signalled_used(vdev, idx); 854 855 /* In the cross-endian case, we need to reset the vring endianness to 856 * native as legacy devices expect so by default. 857 */ 858 if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) && 859 virtio_legacy_is_cross_endian(vdev)) { 860 r = vhost_virtqueue_set_vring_endian_legacy(dev, 861 !virtio_is_big_endian(vdev), 862 vhost_vq_index); 863 if (r < 0) { 864 error_report("failed to reset vring endianness"); 865 } 866 } 867 868 assert (r >= 0); 869 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), 870 0, virtio_queue_get_ring_size(vdev, idx)); 871 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), 872 1, virtio_queue_get_used_size(vdev, idx)); 873 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), 874 0, virtio_queue_get_avail_size(vdev, idx)); 875 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), 876 0, virtio_queue_get_desc_size(vdev, idx)); 877 } 878 879 static void vhost_eventfd_add(MemoryListener *listener, 880 MemoryRegionSection *section, 881 bool match_data, uint64_t data, EventNotifier *e) 882 { 883 } 884 885 static void vhost_eventfd_del(MemoryListener *listener, 886 MemoryRegionSection *section, 887 bool match_data, uint64_t data, EventNotifier *e) 888 { 889 } 890 891 static int vhost_virtqueue_init(struct vhost_dev *dev, 892 struct vhost_virtqueue *vq, int n) 893 { 894 int vhost_vq_index = dev->vhost_ops->vhost_backend_get_vq_index(dev, n); 895 struct vhost_vring_file file = { 896 .index = vhost_vq_index, 897 }; 898 int r = event_notifier_init(&vq->masked_notifier, 0); 899 if (r < 0) { 900 return r; 901 } 902 903 file.fd = event_notifier_get_fd(&vq->masked_notifier); 904 r = dev->vhost_ops->vhost_call(dev, VHOST_SET_VRING_CALL, &file); 905 if (r) { 906 r = -errno; 907 goto fail_call; 908 } 909 return 0; 910 fail_call: 911 event_notifier_cleanup(&vq->masked_notifier); 912 return r; 913 } 914 915 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq) 916 { 917 event_notifier_cleanup(&vq->masked_notifier); 918 } 919 920 int vhost_dev_init(struct vhost_dev *hdev, void *opaque, 921 VhostBackendType backend_type) 922 { 923 uint64_t features; 924 int i, r; 925 926 if (vhost_set_backend_type(hdev, backend_type) < 0) { 927 close((uintptr_t)opaque); 928 return -1; 929 } 930 931 if (hdev->vhost_ops->vhost_backend_init(hdev, opaque) < 0) { 932 close((uintptr_t)opaque); 933 return -errno; 934 } 935 936 QLIST_INSERT_HEAD(&vhost_devices, hdev, entry); 937 938 r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_OWNER, NULL); 939 if (r < 0) { 940 goto fail; 941 } 942 943 r = hdev->vhost_ops->vhost_call(hdev, VHOST_GET_FEATURES, &features); 944 if (r < 0) { 945 goto fail; 946 } 947 948 for (i = 0; i < hdev->nvqs; ++i) { 949 r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i); 950 if (r < 0) { 951 goto fail_vq; 952 } 953 } 954 hdev->features = features; 955 956 hdev->memory_listener = (MemoryListener) { 957 .begin = vhost_begin, 958 .commit = vhost_commit, 959 .region_add = vhost_region_add, 960 .region_del = vhost_region_del, 961 .region_nop = vhost_region_nop, 962 .log_start = vhost_log_start, 963 .log_stop = vhost_log_stop, 964 .log_sync = vhost_log_sync, 965 .log_global_start = vhost_log_global_start, 966 .log_global_stop = vhost_log_global_stop, 967 .eventfd_add = vhost_eventfd_add, 968 .eventfd_del = vhost_eventfd_del, 969 .priority = 10 970 }; 971 hdev->migration_blocker = NULL; 972 if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) { 973 error_setg(&hdev->migration_blocker, 974 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature."); 975 migrate_add_blocker(hdev->migration_blocker); 976 } 977 hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions)); 978 hdev->n_mem_sections = 0; 979 hdev->mem_sections = NULL; 980 hdev->log = NULL; 981 hdev->log_size = 0; 982 hdev->log_enabled = false; 983 hdev->started = false; 984 hdev->memory_changed = false; 985 memory_listener_register(&hdev->memory_listener, &address_space_memory); 986 return 0; 987 fail_vq: 988 while (--i >= 0) { 989 vhost_virtqueue_cleanup(hdev->vqs + i); 990 } 991 fail: 992 r = -errno; 993 hdev->vhost_ops->vhost_backend_cleanup(hdev); 994 QLIST_REMOVE(hdev, entry); 995 return r; 996 } 997 998 void vhost_dev_cleanup(struct vhost_dev *hdev) 999 { 1000 int i; 1001 for (i = 0; i < hdev->nvqs; ++i) { 1002 vhost_virtqueue_cleanup(hdev->vqs + i); 1003 } 1004 memory_listener_unregister(&hdev->memory_listener); 1005 if (hdev->migration_blocker) { 1006 migrate_del_blocker(hdev->migration_blocker); 1007 error_free(hdev->migration_blocker); 1008 } 1009 g_free(hdev->mem); 1010 g_free(hdev->mem_sections); 1011 hdev->vhost_ops->vhost_backend_cleanup(hdev); 1012 QLIST_REMOVE(hdev, entry); 1013 } 1014 1015 /* Stop processing guest IO notifications in qemu. 1016 * Start processing them in vhost in kernel. 1017 */ 1018 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1019 { 1020 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1021 VirtioBusState *vbus = VIRTIO_BUS(qbus); 1022 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus); 1023 int i, r, e; 1024 if (!k->set_host_notifier) { 1025 fprintf(stderr, "binding does not support host notifiers\n"); 1026 r = -ENOSYS; 1027 goto fail; 1028 } 1029 1030 for (i = 0; i < hdev->nvqs; ++i) { 1031 r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, true); 1032 if (r < 0) { 1033 fprintf(stderr, "vhost VQ %d notifier binding failed: %d\n", i, -r); 1034 goto fail_vq; 1035 } 1036 } 1037 1038 return 0; 1039 fail_vq: 1040 while (--i >= 0) { 1041 e = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false); 1042 if (e < 0) { 1043 fprintf(stderr, "vhost VQ %d notifier cleanup error: %d\n", i, -r); 1044 fflush(stderr); 1045 } 1046 assert (e >= 0); 1047 } 1048 fail: 1049 return r; 1050 } 1051 1052 /* Stop processing guest IO notifications in vhost. 1053 * Start processing them in qemu. 1054 * This might actually run the qemu handlers right away, 1055 * so virtio in qemu must be completely setup when this is called. 1056 */ 1057 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1058 { 1059 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1060 VirtioBusState *vbus = VIRTIO_BUS(qbus); 1061 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus); 1062 int i, r; 1063 1064 for (i = 0; i < hdev->nvqs; ++i) { 1065 r = k->set_host_notifier(qbus->parent, hdev->vq_index + i, false); 1066 if (r < 0) { 1067 fprintf(stderr, "vhost VQ %d notifier cleanup failed: %d\n", i, -r); 1068 fflush(stderr); 1069 } 1070 assert (r >= 0); 1071 } 1072 } 1073 1074 /* Test and clear event pending status. 1075 * Should be called after unmask to avoid losing events. 1076 */ 1077 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n) 1078 { 1079 struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index; 1080 assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs); 1081 return event_notifier_test_and_clear(&vq->masked_notifier); 1082 } 1083 1084 /* Mask/unmask events from this vq. */ 1085 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n, 1086 bool mask) 1087 { 1088 struct VirtQueue *vvq = virtio_get_queue(vdev, n); 1089 int r, index = n - hdev->vq_index; 1090 struct vhost_vring_file file; 1091 1092 if (mask) { 1093 file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier); 1094 } else { 1095 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq)); 1096 } 1097 1098 file.index = hdev->vhost_ops->vhost_backend_get_vq_index(hdev, n); 1099 r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_VRING_CALL, &file); 1100 assert(r >= 0); 1101 } 1102 1103 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits, 1104 uint64_t features) 1105 { 1106 const int *bit = feature_bits; 1107 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1108 uint64_t bit_mask = (1ULL << *bit); 1109 if (!(hdev->features & bit_mask)) { 1110 features &= ~bit_mask; 1111 } 1112 bit++; 1113 } 1114 return features; 1115 } 1116 1117 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits, 1118 uint64_t features) 1119 { 1120 const int *bit = feature_bits; 1121 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1122 uint64_t bit_mask = (1ULL << *bit); 1123 if (features & bit_mask) { 1124 hdev->acked_features |= bit_mask; 1125 } 1126 bit++; 1127 } 1128 } 1129 1130 /* Host notifiers must be enabled at this point. */ 1131 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev) 1132 { 1133 int i, r; 1134 1135 hdev->started = true; 1136 1137 r = vhost_dev_set_features(hdev, hdev->log_enabled); 1138 if (r < 0) { 1139 goto fail_features; 1140 } 1141 r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_MEM_TABLE, hdev->mem); 1142 if (r < 0) { 1143 r = -errno; 1144 goto fail_mem; 1145 } 1146 for (i = 0; i < hdev->nvqs; ++i) { 1147 r = vhost_virtqueue_start(hdev, 1148 vdev, 1149 hdev->vqs + i, 1150 hdev->vq_index + i); 1151 if (r < 0) { 1152 goto fail_vq; 1153 } 1154 } 1155 1156 if (hdev->log_enabled) { 1157 uint64_t log_base; 1158 1159 hdev->log_size = vhost_get_log_size(hdev); 1160 hdev->log = vhost_log_get(hdev->log_size); 1161 log_base = (uintptr_t)hdev->log->log; 1162 r = hdev->vhost_ops->vhost_call(hdev, VHOST_SET_LOG_BASE, 1163 hdev->log_size ? &log_base : NULL); 1164 if (r < 0) { 1165 r = -errno; 1166 goto fail_log; 1167 } 1168 } 1169 1170 return 0; 1171 fail_log: 1172 vhost_log_put(hdev, false); 1173 fail_vq: 1174 while (--i >= 0) { 1175 vhost_virtqueue_stop(hdev, 1176 vdev, 1177 hdev->vqs + i, 1178 hdev->vq_index + i); 1179 } 1180 i = hdev->nvqs; 1181 fail_mem: 1182 fail_features: 1183 1184 hdev->started = false; 1185 return r; 1186 } 1187 1188 /* Host notifiers must be enabled at this point. */ 1189 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev) 1190 { 1191 int i; 1192 1193 for (i = 0; i < hdev->nvqs; ++i) { 1194 vhost_virtqueue_stop(hdev, 1195 vdev, 1196 hdev->vqs + i, 1197 hdev->vq_index + i); 1198 } 1199 1200 vhost_log_put(hdev, true); 1201 hdev->started = false; 1202 hdev->log = NULL; 1203 hdev->log_size = 0; 1204 } 1205 1206