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 "qemu/osdep.h" 17 #include "qapi/error.h" 18 #include "hw/virtio/vhost.h" 19 #include "hw/hw.h" 20 #include "qemu/atomic.h" 21 #include "qemu/range.h" 22 #include "qemu/error-report.h" 23 #include "qemu/memfd.h" 24 #include <linux/vhost.h> 25 #include "exec/address-spaces.h" 26 #include "hw/virtio/virtio-bus.h" 27 #include "hw/virtio/virtio-access.h" 28 #include "migration/migration.h" 29 30 /* enabled until disconnected backend stabilizes */ 31 #define _VHOST_DEBUG 1 32 33 #ifdef _VHOST_DEBUG 34 #define VHOST_OPS_DEBUG(fmt, ...) \ 35 do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \ 36 strerror(errno), errno); } while (0) 37 #else 38 #define VHOST_OPS_DEBUG(fmt, ...) \ 39 do { } while (0) 40 #endif 41 42 static struct vhost_log *vhost_log; 43 static struct vhost_log *vhost_log_shm; 44 45 static unsigned int used_memslots; 46 static QLIST_HEAD(, vhost_dev) vhost_devices = 47 QLIST_HEAD_INITIALIZER(vhost_devices); 48 49 bool vhost_has_free_slot(void) 50 { 51 unsigned int slots_limit = ~0U; 52 struct vhost_dev *hdev; 53 54 QLIST_FOREACH(hdev, &vhost_devices, entry) { 55 unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev); 56 slots_limit = MIN(slots_limit, r); 57 } 58 return slots_limit > used_memslots; 59 } 60 61 static void vhost_dev_sync_region(struct vhost_dev *dev, 62 MemoryRegionSection *section, 63 uint64_t mfirst, uint64_t mlast, 64 uint64_t rfirst, uint64_t rlast) 65 { 66 vhost_log_chunk_t *log = dev->log->log; 67 68 uint64_t start = MAX(mfirst, rfirst); 69 uint64_t end = MIN(mlast, rlast); 70 vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK; 71 vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1; 72 uint64_t addr = (start / VHOST_LOG_CHUNK) * VHOST_LOG_CHUNK; 73 74 if (end < start) { 75 return; 76 } 77 assert(end / VHOST_LOG_CHUNK < dev->log_size); 78 assert(start / VHOST_LOG_CHUNK < dev->log_size); 79 80 for (;from < to; ++from) { 81 vhost_log_chunk_t log; 82 /* We first check with non-atomic: much cheaper, 83 * and we expect non-dirty to be the common case. */ 84 if (!*from) { 85 addr += VHOST_LOG_CHUNK; 86 continue; 87 } 88 /* Data must be read atomically. We don't really need barrier semantics 89 * but it's easier to use atomic_* than roll our own. */ 90 log = atomic_xchg(from, 0); 91 while (log) { 92 int bit = ctzl(log); 93 hwaddr page_addr; 94 hwaddr section_offset; 95 hwaddr mr_offset; 96 page_addr = addr + bit * VHOST_LOG_PAGE; 97 section_offset = page_addr - section->offset_within_address_space; 98 mr_offset = section_offset + section->offset_within_region; 99 memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE); 100 log &= ~(0x1ull << bit); 101 } 102 addr += VHOST_LOG_CHUNK; 103 } 104 } 105 106 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev, 107 MemoryRegionSection *section, 108 hwaddr first, 109 hwaddr last) 110 { 111 int i; 112 hwaddr start_addr; 113 hwaddr end_addr; 114 115 if (!dev->log_enabled || !dev->started) { 116 return 0; 117 } 118 start_addr = section->offset_within_address_space; 119 end_addr = range_get_last(start_addr, int128_get64(section->size)); 120 start_addr = MAX(first, start_addr); 121 end_addr = MIN(last, end_addr); 122 123 for (i = 0; i < dev->mem->nregions; ++i) { 124 struct vhost_memory_region *reg = dev->mem->regions + i; 125 vhost_dev_sync_region(dev, section, start_addr, end_addr, 126 reg->guest_phys_addr, 127 range_get_last(reg->guest_phys_addr, 128 reg->memory_size)); 129 } 130 for (i = 0; i < dev->nvqs; ++i) { 131 struct vhost_virtqueue *vq = dev->vqs + i; 132 vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys, 133 range_get_last(vq->used_phys, vq->used_size)); 134 } 135 return 0; 136 } 137 138 static void vhost_log_sync(MemoryListener *listener, 139 MemoryRegionSection *section) 140 { 141 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 142 memory_listener); 143 vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL); 144 } 145 146 static void vhost_log_sync_range(struct vhost_dev *dev, 147 hwaddr first, hwaddr last) 148 { 149 int i; 150 /* FIXME: this is N^2 in number of sections */ 151 for (i = 0; i < dev->n_mem_sections; ++i) { 152 MemoryRegionSection *section = &dev->mem_sections[i]; 153 vhost_sync_dirty_bitmap(dev, section, first, last); 154 } 155 } 156 157 /* Assign/unassign. Keep an unsorted array of non-overlapping 158 * memory regions in dev->mem. */ 159 static void vhost_dev_unassign_memory(struct vhost_dev *dev, 160 uint64_t start_addr, 161 uint64_t size) 162 { 163 int from, to, n = dev->mem->nregions; 164 /* Track overlapping/split regions for sanity checking. */ 165 int overlap_start = 0, overlap_end = 0, overlap_middle = 0, split = 0; 166 167 for (from = 0, to = 0; from < n; ++from, ++to) { 168 struct vhost_memory_region *reg = dev->mem->regions + to; 169 uint64_t reglast; 170 uint64_t memlast; 171 uint64_t change; 172 173 /* clone old region */ 174 if (to != from) { 175 memcpy(reg, dev->mem->regions + from, sizeof *reg); 176 } 177 178 /* No overlap is simple */ 179 if (!ranges_overlap(reg->guest_phys_addr, reg->memory_size, 180 start_addr, size)) { 181 continue; 182 } 183 184 /* Split only happens if supplied region 185 * is in the middle of an existing one. Thus it can not 186 * overlap with any other existing region. */ 187 assert(!split); 188 189 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); 190 memlast = range_get_last(start_addr, size); 191 192 /* Remove whole region */ 193 if (start_addr <= reg->guest_phys_addr && memlast >= reglast) { 194 --dev->mem->nregions; 195 --to; 196 ++overlap_middle; 197 continue; 198 } 199 200 /* Shrink region */ 201 if (memlast >= reglast) { 202 reg->memory_size = start_addr - reg->guest_phys_addr; 203 assert(reg->memory_size); 204 assert(!overlap_end); 205 ++overlap_end; 206 continue; 207 } 208 209 /* Shift region */ 210 if (start_addr <= reg->guest_phys_addr) { 211 change = memlast + 1 - reg->guest_phys_addr; 212 reg->memory_size -= change; 213 reg->guest_phys_addr += change; 214 reg->userspace_addr += change; 215 assert(reg->memory_size); 216 assert(!overlap_start); 217 ++overlap_start; 218 continue; 219 } 220 221 /* This only happens if supplied region 222 * is in the middle of an existing one. Thus it can not 223 * overlap with any other existing region. */ 224 assert(!overlap_start); 225 assert(!overlap_end); 226 assert(!overlap_middle); 227 /* Split region: shrink first part, shift second part. */ 228 memcpy(dev->mem->regions + n, reg, sizeof *reg); 229 reg->memory_size = start_addr - reg->guest_phys_addr; 230 assert(reg->memory_size); 231 change = memlast + 1 - reg->guest_phys_addr; 232 reg = dev->mem->regions + n; 233 reg->memory_size -= change; 234 assert(reg->memory_size); 235 reg->guest_phys_addr += change; 236 reg->userspace_addr += change; 237 /* Never add more than 1 region */ 238 assert(dev->mem->nregions == n); 239 ++dev->mem->nregions; 240 ++split; 241 } 242 } 243 244 /* Called after unassign, so no regions overlap the given range. */ 245 static void vhost_dev_assign_memory(struct vhost_dev *dev, 246 uint64_t start_addr, 247 uint64_t size, 248 uint64_t uaddr) 249 { 250 int from, to; 251 struct vhost_memory_region *merged = NULL; 252 for (from = 0, to = 0; from < dev->mem->nregions; ++from, ++to) { 253 struct vhost_memory_region *reg = dev->mem->regions + to; 254 uint64_t prlast, urlast; 255 uint64_t pmlast, umlast; 256 uint64_t s, e, u; 257 258 /* clone old region */ 259 if (to != from) { 260 memcpy(reg, dev->mem->regions + from, sizeof *reg); 261 } 262 prlast = range_get_last(reg->guest_phys_addr, reg->memory_size); 263 pmlast = range_get_last(start_addr, size); 264 urlast = range_get_last(reg->userspace_addr, reg->memory_size); 265 umlast = range_get_last(uaddr, size); 266 267 /* check for overlapping regions: should never happen. */ 268 assert(prlast < start_addr || pmlast < reg->guest_phys_addr); 269 /* Not an adjacent or overlapping region - do not merge. */ 270 if ((prlast + 1 != start_addr || urlast + 1 != uaddr) && 271 (pmlast + 1 != reg->guest_phys_addr || 272 umlast + 1 != reg->userspace_addr)) { 273 continue; 274 } 275 276 if (dev->vhost_ops->vhost_backend_can_merge && 277 !dev->vhost_ops->vhost_backend_can_merge(dev, uaddr, size, 278 reg->userspace_addr, 279 reg->memory_size)) { 280 continue; 281 } 282 283 if (merged) { 284 --to; 285 assert(to >= 0); 286 } else { 287 merged = reg; 288 } 289 u = MIN(uaddr, reg->userspace_addr); 290 s = MIN(start_addr, reg->guest_phys_addr); 291 e = MAX(pmlast, prlast); 292 uaddr = merged->userspace_addr = u; 293 start_addr = merged->guest_phys_addr = s; 294 size = merged->memory_size = e - s + 1; 295 assert(merged->memory_size); 296 } 297 298 if (!merged) { 299 struct vhost_memory_region *reg = dev->mem->regions + to; 300 memset(reg, 0, sizeof *reg); 301 reg->memory_size = size; 302 assert(reg->memory_size); 303 reg->guest_phys_addr = start_addr; 304 reg->userspace_addr = uaddr; 305 ++to; 306 } 307 assert(to <= dev->mem->nregions + 1); 308 dev->mem->nregions = to; 309 } 310 311 static uint64_t vhost_get_log_size(struct vhost_dev *dev) 312 { 313 uint64_t log_size = 0; 314 int i; 315 for (i = 0; i < dev->mem->nregions; ++i) { 316 struct vhost_memory_region *reg = dev->mem->regions + i; 317 uint64_t last = range_get_last(reg->guest_phys_addr, 318 reg->memory_size); 319 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 320 } 321 for (i = 0; i < dev->nvqs; ++i) { 322 struct vhost_virtqueue *vq = dev->vqs + i; 323 uint64_t last = vq->used_phys + vq->used_size - 1; 324 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 325 } 326 return log_size; 327 } 328 329 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share) 330 { 331 struct vhost_log *log; 332 uint64_t logsize = size * sizeof(*(log->log)); 333 int fd = -1; 334 335 log = g_new0(struct vhost_log, 1); 336 if (share) { 337 log->log = qemu_memfd_alloc("vhost-log", logsize, 338 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL, 339 &fd); 340 memset(log->log, 0, logsize); 341 } else { 342 log->log = g_malloc0(logsize); 343 } 344 345 log->size = size; 346 log->refcnt = 1; 347 log->fd = fd; 348 349 return log; 350 } 351 352 static struct vhost_log *vhost_log_get(uint64_t size, bool share) 353 { 354 struct vhost_log *log = share ? vhost_log_shm : vhost_log; 355 356 if (!log || log->size != size) { 357 log = vhost_log_alloc(size, share); 358 if (share) { 359 vhost_log_shm = log; 360 } else { 361 vhost_log = log; 362 } 363 } else { 364 ++log->refcnt; 365 } 366 367 return log; 368 } 369 370 static void vhost_log_put(struct vhost_dev *dev, bool sync) 371 { 372 struct vhost_log *log = dev->log; 373 374 if (!log) { 375 return; 376 } 377 dev->log = NULL; 378 dev->log_size = 0; 379 380 --log->refcnt; 381 if (log->refcnt == 0) { 382 /* Sync only the range covered by the old log */ 383 if (dev->log_size && sync) { 384 vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1); 385 } 386 387 if (vhost_log == log) { 388 g_free(log->log); 389 vhost_log = NULL; 390 } else if (vhost_log_shm == log) { 391 qemu_memfd_free(log->log, log->size * sizeof(*(log->log)), 392 log->fd); 393 vhost_log_shm = NULL; 394 } 395 396 g_free(log); 397 } 398 } 399 400 static bool vhost_dev_log_is_shared(struct vhost_dev *dev) 401 { 402 return dev->vhost_ops->vhost_requires_shm_log && 403 dev->vhost_ops->vhost_requires_shm_log(dev); 404 } 405 406 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size) 407 { 408 struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev)); 409 uint64_t log_base = (uintptr_t)log->log; 410 int r; 411 412 /* inform backend of log switching, this must be done before 413 releasing the current log, to ensure no logging is lost */ 414 r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log); 415 if (r < 0) { 416 VHOST_OPS_DEBUG("vhost_set_log_base failed"); 417 } 418 419 vhost_log_put(dev, true); 420 dev->log = log; 421 dev->log_size = size; 422 } 423 424 static int vhost_verify_ring_mappings(struct vhost_dev *dev, 425 uint64_t start_addr, 426 uint64_t size) 427 { 428 int i; 429 int r = 0; 430 431 for (i = 0; !r && i < dev->nvqs; ++i) { 432 struct vhost_virtqueue *vq = dev->vqs + i; 433 hwaddr l; 434 void *p; 435 436 if (!ranges_overlap(start_addr, size, vq->ring_phys, vq->ring_size)) { 437 continue; 438 } 439 l = vq->ring_size; 440 p = cpu_physical_memory_map(vq->ring_phys, &l, 1); 441 if (!p || l != vq->ring_size) { 442 error_report("Unable to map ring buffer for ring %d", i); 443 r = -ENOMEM; 444 } 445 if (p != vq->ring) { 446 error_report("Ring buffer relocated for ring %d", i); 447 r = -EBUSY; 448 } 449 cpu_physical_memory_unmap(p, l, 0, 0); 450 } 451 return r; 452 } 453 454 static struct vhost_memory_region *vhost_dev_find_reg(struct vhost_dev *dev, 455 uint64_t start_addr, 456 uint64_t size) 457 { 458 int i, n = dev->mem->nregions; 459 for (i = 0; i < n; ++i) { 460 struct vhost_memory_region *reg = dev->mem->regions + i; 461 if (ranges_overlap(reg->guest_phys_addr, reg->memory_size, 462 start_addr, size)) { 463 return reg; 464 } 465 } 466 return NULL; 467 } 468 469 static bool vhost_dev_cmp_memory(struct vhost_dev *dev, 470 uint64_t start_addr, 471 uint64_t size, 472 uint64_t uaddr) 473 { 474 struct vhost_memory_region *reg = vhost_dev_find_reg(dev, start_addr, size); 475 uint64_t reglast; 476 uint64_t memlast; 477 478 if (!reg) { 479 return true; 480 } 481 482 reglast = range_get_last(reg->guest_phys_addr, reg->memory_size); 483 memlast = range_get_last(start_addr, size); 484 485 /* Need to extend region? */ 486 if (start_addr < reg->guest_phys_addr || memlast > reglast) { 487 return true; 488 } 489 /* userspace_addr changed? */ 490 return uaddr != reg->userspace_addr + start_addr - reg->guest_phys_addr; 491 } 492 493 static void vhost_set_memory(MemoryListener *listener, 494 MemoryRegionSection *section, 495 bool add) 496 { 497 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 498 memory_listener); 499 hwaddr start_addr = section->offset_within_address_space; 500 ram_addr_t size = int128_get64(section->size); 501 bool log_dirty = 502 memory_region_get_dirty_log_mask(section->mr) & ~(1 << DIRTY_MEMORY_MIGRATION); 503 int s = offsetof(struct vhost_memory, regions) + 504 (dev->mem->nregions + 1) * sizeof dev->mem->regions[0]; 505 void *ram; 506 507 dev->mem = g_realloc(dev->mem, s); 508 509 if (log_dirty) { 510 add = false; 511 } 512 513 assert(size); 514 515 /* Optimize no-change case. At least cirrus_vga does this a lot at this time. */ 516 ram = memory_region_get_ram_ptr(section->mr) + section->offset_within_region; 517 if (add) { 518 if (!vhost_dev_cmp_memory(dev, start_addr, size, (uintptr_t)ram)) { 519 /* Region exists with same address. Nothing to do. */ 520 return; 521 } 522 } else { 523 if (!vhost_dev_find_reg(dev, start_addr, size)) { 524 /* Removing region that we don't access. Nothing to do. */ 525 return; 526 } 527 } 528 529 vhost_dev_unassign_memory(dev, start_addr, size); 530 if (add) { 531 /* Add given mapping, merging adjacent regions if any */ 532 vhost_dev_assign_memory(dev, start_addr, size, (uintptr_t)ram); 533 } else { 534 /* Remove old mapping for this memory, if any. */ 535 vhost_dev_unassign_memory(dev, start_addr, size); 536 } 537 dev->mem_changed_start_addr = MIN(dev->mem_changed_start_addr, start_addr); 538 dev->mem_changed_end_addr = MAX(dev->mem_changed_end_addr, start_addr + size - 1); 539 dev->memory_changed = true; 540 used_memslots = dev->mem->nregions; 541 } 542 543 static bool vhost_section(MemoryRegionSection *section) 544 { 545 return memory_region_is_ram(section->mr); 546 } 547 548 static void vhost_begin(MemoryListener *listener) 549 { 550 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 551 memory_listener); 552 dev->mem_changed_end_addr = 0; 553 dev->mem_changed_start_addr = -1; 554 } 555 556 static void vhost_commit(MemoryListener *listener) 557 { 558 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 559 memory_listener); 560 hwaddr start_addr = 0; 561 ram_addr_t size = 0; 562 uint64_t log_size; 563 int r; 564 565 if (!dev->memory_changed) { 566 return; 567 } 568 if (!dev->started) { 569 return; 570 } 571 if (dev->mem_changed_start_addr > dev->mem_changed_end_addr) { 572 return; 573 } 574 575 if (dev->started) { 576 start_addr = dev->mem_changed_start_addr; 577 size = dev->mem_changed_end_addr - dev->mem_changed_start_addr + 1; 578 579 r = vhost_verify_ring_mappings(dev, start_addr, size); 580 assert(r >= 0); 581 } 582 583 if (!dev->log_enabled) { 584 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); 585 if (r < 0) { 586 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 587 } 588 dev->memory_changed = false; 589 return; 590 } 591 log_size = vhost_get_log_size(dev); 592 /* We allocate an extra 4K bytes to log, 593 * to reduce the * number of reallocations. */ 594 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log) 595 /* To log more, must increase log size before table update. */ 596 if (dev->log_size < log_size) { 597 vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER); 598 } 599 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); 600 if (r < 0) { 601 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 602 } 603 /* To log less, can only decrease log size after table update. */ 604 if (dev->log_size > log_size + VHOST_LOG_BUFFER) { 605 vhost_dev_log_resize(dev, log_size); 606 } 607 dev->memory_changed = false; 608 } 609 610 static void vhost_region_add(MemoryListener *listener, 611 MemoryRegionSection *section) 612 { 613 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 614 memory_listener); 615 616 if (!vhost_section(section)) { 617 return; 618 } 619 620 ++dev->n_mem_sections; 621 dev->mem_sections = g_renew(MemoryRegionSection, dev->mem_sections, 622 dev->n_mem_sections); 623 dev->mem_sections[dev->n_mem_sections - 1] = *section; 624 memory_region_ref(section->mr); 625 vhost_set_memory(listener, section, true); 626 } 627 628 static void vhost_region_del(MemoryListener *listener, 629 MemoryRegionSection *section) 630 { 631 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 632 memory_listener); 633 int i; 634 635 if (!vhost_section(section)) { 636 return; 637 } 638 639 vhost_set_memory(listener, section, false); 640 memory_region_unref(section->mr); 641 for (i = 0; i < dev->n_mem_sections; ++i) { 642 if (dev->mem_sections[i].offset_within_address_space 643 == section->offset_within_address_space) { 644 --dev->n_mem_sections; 645 memmove(&dev->mem_sections[i], &dev->mem_sections[i+1], 646 (dev->n_mem_sections - i) * sizeof(*dev->mem_sections)); 647 break; 648 } 649 } 650 } 651 652 static void vhost_region_nop(MemoryListener *listener, 653 MemoryRegionSection *section) 654 { 655 } 656 657 static int vhost_virtqueue_set_addr(struct vhost_dev *dev, 658 struct vhost_virtqueue *vq, 659 unsigned idx, bool enable_log) 660 { 661 struct vhost_vring_addr addr = { 662 .index = idx, 663 .desc_user_addr = (uint64_t)(unsigned long)vq->desc, 664 .avail_user_addr = (uint64_t)(unsigned long)vq->avail, 665 .used_user_addr = (uint64_t)(unsigned long)vq->used, 666 .log_guest_addr = vq->used_phys, 667 .flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0, 668 }; 669 int r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr); 670 if (r < 0) { 671 VHOST_OPS_DEBUG("vhost_set_vring_addr failed"); 672 return -errno; 673 } 674 return 0; 675 } 676 677 static int vhost_dev_set_features(struct vhost_dev *dev, bool enable_log) 678 { 679 uint64_t features = dev->acked_features; 680 int r; 681 if (enable_log) { 682 features |= 0x1ULL << VHOST_F_LOG_ALL; 683 } 684 r = dev->vhost_ops->vhost_set_features(dev, features); 685 if (r < 0) { 686 VHOST_OPS_DEBUG("vhost_set_features failed"); 687 } 688 return r < 0 ? -errno : 0; 689 } 690 691 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log) 692 { 693 int r, i, idx; 694 r = vhost_dev_set_features(dev, enable_log); 695 if (r < 0) { 696 goto err_features; 697 } 698 for (i = 0; i < dev->nvqs; ++i) { 699 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); 700 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, 701 enable_log); 702 if (r < 0) { 703 goto err_vq; 704 } 705 } 706 return 0; 707 err_vq: 708 for (; i >= 0; --i) { 709 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); 710 vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, 711 dev->log_enabled); 712 } 713 vhost_dev_set_features(dev, dev->log_enabled); 714 err_features: 715 return r; 716 } 717 718 static int vhost_migration_log(MemoryListener *listener, int enable) 719 { 720 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 721 memory_listener); 722 int r; 723 if (!!enable == dev->log_enabled) { 724 return 0; 725 } 726 if (!dev->started) { 727 dev->log_enabled = enable; 728 return 0; 729 } 730 if (!enable) { 731 r = vhost_dev_set_log(dev, false); 732 if (r < 0) { 733 return r; 734 } 735 vhost_log_put(dev, false); 736 } else { 737 vhost_dev_log_resize(dev, vhost_get_log_size(dev)); 738 r = vhost_dev_set_log(dev, true); 739 if (r < 0) { 740 return r; 741 } 742 } 743 dev->log_enabled = enable; 744 return 0; 745 } 746 747 static void vhost_log_global_start(MemoryListener *listener) 748 { 749 int r; 750 751 r = vhost_migration_log(listener, true); 752 if (r < 0) { 753 abort(); 754 } 755 } 756 757 static void vhost_log_global_stop(MemoryListener *listener) 758 { 759 int r; 760 761 r = vhost_migration_log(listener, false); 762 if (r < 0) { 763 abort(); 764 } 765 } 766 767 static void vhost_log_start(MemoryListener *listener, 768 MemoryRegionSection *section, 769 int old, int new) 770 { 771 /* FIXME: implement */ 772 } 773 774 static void vhost_log_stop(MemoryListener *listener, 775 MemoryRegionSection *section, 776 int old, int new) 777 { 778 /* FIXME: implement */ 779 } 780 781 /* The vhost driver natively knows how to handle the vrings of non 782 * cross-endian legacy devices and modern devices. Only legacy devices 783 * exposed to a bi-endian guest may require the vhost driver to use a 784 * specific endianness. 785 */ 786 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev) 787 { 788 if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) { 789 return false; 790 } 791 #ifdef HOST_WORDS_BIGENDIAN 792 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE; 793 #else 794 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG; 795 #endif 796 } 797 798 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev, 799 bool is_big_endian, 800 int vhost_vq_index) 801 { 802 struct vhost_vring_state s = { 803 .index = vhost_vq_index, 804 .num = is_big_endian 805 }; 806 807 if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) { 808 return 0; 809 } 810 811 VHOST_OPS_DEBUG("vhost_set_vring_endian failed"); 812 if (errno == ENOTTY) { 813 error_report("vhost does not support cross-endian"); 814 return -ENOSYS; 815 } 816 817 return -errno; 818 } 819 820 static int vhost_virtqueue_start(struct vhost_dev *dev, 821 struct VirtIODevice *vdev, 822 struct vhost_virtqueue *vq, 823 unsigned idx) 824 { 825 hwaddr s, l, a; 826 int r; 827 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); 828 struct vhost_vring_file file = { 829 .index = vhost_vq_index 830 }; 831 struct vhost_vring_state state = { 832 .index = vhost_vq_index 833 }; 834 struct VirtQueue *vvq = virtio_get_queue(vdev, idx); 835 836 837 vq->num = state.num = virtio_queue_get_num(vdev, idx); 838 r = dev->vhost_ops->vhost_set_vring_num(dev, &state); 839 if (r) { 840 VHOST_OPS_DEBUG("vhost_set_vring_num failed"); 841 return -errno; 842 } 843 844 state.num = virtio_queue_get_last_avail_idx(vdev, idx); 845 r = dev->vhost_ops->vhost_set_vring_base(dev, &state); 846 if (r) { 847 VHOST_OPS_DEBUG("vhost_set_vring_base failed"); 848 return -errno; 849 } 850 851 if (vhost_needs_vring_endian(vdev)) { 852 r = vhost_virtqueue_set_vring_endian_legacy(dev, 853 virtio_is_big_endian(vdev), 854 vhost_vq_index); 855 if (r) { 856 return -errno; 857 } 858 } 859 860 s = l = virtio_queue_get_desc_size(vdev, idx); 861 a = virtio_queue_get_desc_addr(vdev, idx); 862 vq->desc = cpu_physical_memory_map(a, &l, 0); 863 if (!vq->desc || l != s) { 864 r = -ENOMEM; 865 goto fail_alloc_desc; 866 } 867 s = l = virtio_queue_get_avail_size(vdev, idx); 868 a = virtio_queue_get_avail_addr(vdev, idx); 869 vq->avail = cpu_physical_memory_map(a, &l, 0); 870 if (!vq->avail || l != s) { 871 r = -ENOMEM; 872 goto fail_alloc_avail; 873 } 874 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx); 875 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx); 876 vq->used = cpu_physical_memory_map(a, &l, 1); 877 if (!vq->used || l != s) { 878 r = -ENOMEM; 879 goto fail_alloc_used; 880 } 881 882 vq->ring_size = s = l = virtio_queue_get_ring_size(vdev, idx); 883 vq->ring_phys = a = virtio_queue_get_ring_addr(vdev, idx); 884 vq->ring = cpu_physical_memory_map(a, &l, 1); 885 if (!vq->ring || l != s) { 886 r = -ENOMEM; 887 goto fail_alloc_ring; 888 } 889 890 r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled); 891 if (r < 0) { 892 r = -errno; 893 goto fail_alloc; 894 } 895 896 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq)); 897 r = dev->vhost_ops->vhost_set_vring_kick(dev, &file); 898 if (r) { 899 VHOST_OPS_DEBUG("vhost_set_vring_kick failed"); 900 r = -errno; 901 goto fail_kick; 902 } 903 904 /* Clear and discard previous events if any. */ 905 event_notifier_test_and_clear(&vq->masked_notifier); 906 907 /* Init vring in unmasked state, unless guest_notifier_mask 908 * will do it later. 909 */ 910 if (!vdev->use_guest_notifier_mask) { 911 /* TODO: check and handle errors. */ 912 vhost_virtqueue_mask(dev, vdev, idx, false); 913 } 914 915 return 0; 916 917 fail_kick: 918 fail_alloc: 919 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), 920 0, 0); 921 fail_alloc_ring: 922 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), 923 0, 0); 924 fail_alloc_used: 925 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), 926 0, 0); 927 fail_alloc_avail: 928 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), 929 0, 0); 930 fail_alloc_desc: 931 return r; 932 } 933 934 static void vhost_virtqueue_stop(struct vhost_dev *dev, 935 struct VirtIODevice *vdev, 936 struct vhost_virtqueue *vq, 937 unsigned idx) 938 { 939 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); 940 struct vhost_vring_state state = { 941 .index = vhost_vq_index, 942 }; 943 int r; 944 945 r = dev->vhost_ops->vhost_get_vring_base(dev, &state); 946 if (r < 0) { 947 VHOST_OPS_DEBUG("vhost VQ %d ring restore failed: %d", idx, r); 948 } 949 virtio_queue_set_last_avail_idx(vdev, idx, state.num); 950 virtio_queue_invalidate_signalled_used(vdev, idx); 951 952 /* In the cross-endian case, we need to reset the vring endianness to 953 * native as legacy devices expect so by default. 954 */ 955 if (vhost_needs_vring_endian(vdev)) { 956 vhost_virtqueue_set_vring_endian_legacy(dev, 957 !virtio_is_big_endian(vdev), 958 vhost_vq_index); 959 } 960 961 cpu_physical_memory_unmap(vq->ring, virtio_queue_get_ring_size(vdev, idx), 962 0, virtio_queue_get_ring_size(vdev, idx)); 963 cpu_physical_memory_unmap(vq->used, virtio_queue_get_used_size(vdev, idx), 964 1, virtio_queue_get_used_size(vdev, idx)); 965 cpu_physical_memory_unmap(vq->avail, virtio_queue_get_avail_size(vdev, idx), 966 0, virtio_queue_get_avail_size(vdev, idx)); 967 cpu_physical_memory_unmap(vq->desc, virtio_queue_get_desc_size(vdev, idx), 968 0, virtio_queue_get_desc_size(vdev, idx)); 969 } 970 971 static void vhost_eventfd_add(MemoryListener *listener, 972 MemoryRegionSection *section, 973 bool match_data, uint64_t data, EventNotifier *e) 974 { 975 } 976 977 static void vhost_eventfd_del(MemoryListener *listener, 978 MemoryRegionSection *section, 979 bool match_data, uint64_t data, EventNotifier *e) 980 { 981 } 982 983 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev, 984 int n, uint32_t timeout) 985 { 986 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); 987 struct vhost_vring_state state = { 988 .index = vhost_vq_index, 989 .num = timeout, 990 }; 991 int r; 992 993 if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) { 994 return -EINVAL; 995 } 996 997 r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state); 998 if (r) { 999 VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed"); 1000 return r; 1001 } 1002 1003 return 0; 1004 } 1005 1006 static int vhost_virtqueue_init(struct vhost_dev *dev, 1007 struct vhost_virtqueue *vq, int n) 1008 { 1009 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); 1010 struct vhost_vring_file file = { 1011 .index = vhost_vq_index, 1012 }; 1013 int r = event_notifier_init(&vq->masked_notifier, 0); 1014 if (r < 0) { 1015 return r; 1016 } 1017 1018 file.fd = event_notifier_get_fd(&vq->masked_notifier); 1019 r = dev->vhost_ops->vhost_set_vring_call(dev, &file); 1020 if (r) { 1021 VHOST_OPS_DEBUG("vhost_set_vring_call failed"); 1022 r = -errno; 1023 goto fail_call; 1024 } 1025 return 0; 1026 fail_call: 1027 event_notifier_cleanup(&vq->masked_notifier); 1028 return r; 1029 } 1030 1031 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq) 1032 { 1033 event_notifier_cleanup(&vq->masked_notifier); 1034 } 1035 1036 int vhost_dev_init(struct vhost_dev *hdev, void *opaque, 1037 VhostBackendType backend_type, uint32_t busyloop_timeout) 1038 { 1039 uint64_t features; 1040 int i, r, n_initialized_vqs = 0; 1041 1042 hdev->migration_blocker = NULL; 1043 1044 r = vhost_set_backend_type(hdev, backend_type); 1045 assert(r >= 0); 1046 1047 r = hdev->vhost_ops->vhost_backend_init(hdev, opaque); 1048 if (r < 0) { 1049 goto fail; 1050 } 1051 1052 if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) { 1053 error_report("vhost backend memory slots limit is less" 1054 " than current number of present memory slots"); 1055 r = -1; 1056 goto fail; 1057 } 1058 1059 r = hdev->vhost_ops->vhost_set_owner(hdev); 1060 if (r < 0) { 1061 VHOST_OPS_DEBUG("vhost_set_owner failed"); 1062 goto fail; 1063 } 1064 1065 r = hdev->vhost_ops->vhost_get_features(hdev, &features); 1066 if (r < 0) { 1067 VHOST_OPS_DEBUG("vhost_get_features failed"); 1068 goto fail; 1069 } 1070 1071 for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) { 1072 r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i); 1073 if (r < 0) { 1074 goto fail; 1075 } 1076 } 1077 1078 if (busyloop_timeout) { 1079 for (i = 0; i < hdev->nvqs; ++i) { 1080 r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 1081 busyloop_timeout); 1082 if (r < 0) { 1083 goto fail_busyloop; 1084 } 1085 } 1086 } 1087 1088 hdev->features = features; 1089 1090 hdev->memory_listener = (MemoryListener) { 1091 .begin = vhost_begin, 1092 .commit = vhost_commit, 1093 .region_add = vhost_region_add, 1094 .region_del = vhost_region_del, 1095 .region_nop = vhost_region_nop, 1096 .log_start = vhost_log_start, 1097 .log_stop = vhost_log_stop, 1098 .log_sync = vhost_log_sync, 1099 .log_global_start = vhost_log_global_start, 1100 .log_global_stop = vhost_log_global_stop, 1101 .eventfd_add = vhost_eventfd_add, 1102 .eventfd_del = vhost_eventfd_del, 1103 .priority = 10 1104 }; 1105 1106 if (hdev->migration_blocker == NULL) { 1107 if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) { 1108 error_setg(&hdev->migration_blocker, 1109 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature."); 1110 } else if (!qemu_memfd_check()) { 1111 error_setg(&hdev->migration_blocker, 1112 "Migration disabled: failed to allocate shared memory"); 1113 } 1114 } 1115 1116 if (hdev->migration_blocker != NULL) { 1117 migrate_add_blocker(hdev->migration_blocker); 1118 } 1119 1120 hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions)); 1121 hdev->n_mem_sections = 0; 1122 hdev->mem_sections = NULL; 1123 hdev->log = NULL; 1124 hdev->log_size = 0; 1125 hdev->log_enabled = false; 1126 hdev->started = false; 1127 hdev->memory_changed = false; 1128 memory_listener_register(&hdev->memory_listener, &address_space_memory); 1129 QLIST_INSERT_HEAD(&vhost_devices, hdev, entry); 1130 return 0; 1131 1132 fail_busyloop: 1133 while (--i >= 0) { 1134 vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0); 1135 } 1136 fail: 1137 hdev->nvqs = n_initialized_vqs; 1138 vhost_dev_cleanup(hdev); 1139 return r; 1140 } 1141 1142 void vhost_dev_cleanup(struct vhost_dev *hdev) 1143 { 1144 int i; 1145 1146 for (i = 0; i < hdev->nvqs; ++i) { 1147 vhost_virtqueue_cleanup(hdev->vqs + i); 1148 } 1149 if (hdev->mem) { 1150 /* those are only safe after successful init */ 1151 memory_listener_unregister(&hdev->memory_listener); 1152 QLIST_REMOVE(hdev, entry); 1153 } 1154 if (hdev->migration_blocker) { 1155 migrate_del_blocker(hdev->migration_blocker); 1156 error_free(hdev->migration_blocker); 1157 } 1158 g_free(hdev->mem); 1159 g_free(hdev->mem_sections); 1160 if (hdev->vhost_ops) { 1161 hdev->vhost_ops->vhost_backend_cleanup(hdev); 1162 } 1163 assert(!hdev->log); 1164 1165 memset(hdev, 0, sizeof(struct vhost_dev)); 1166 } 1167 1168 /* Stop processing guest IO notifications in qemu. 1169 * Start processing them in vhost in kernel. 1170 */ 1171 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1172 { 1173 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1174 VirtioBusState *vbus = VIRTIO_BUS(qbus); 1175 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus); 1176 int i, r, e; 1177 1178 if (!k->ioeventfd_started) { 1179 error_report("binding does not support host notifiers"); 1180 r = -ENOSYS; 1181 goto fail; 1182 } 1183 1184 for (i = 0; i < hdev->nvqs; ++i) { 1185 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1186 true); 1187 if (r < 0) { 1188 error_report("vhost VQ %d notifier binding failed: %d", i, -r); 1189 goto fail_vq; 1190 } 1191 } 1192 1193 return 0; 1194 fail_vq: 1195 while (--i >= 0) { 1196 e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1197 false); 1198 if (e < 0) { 1199 error_report("vhost VQ %d notifier cleanup error: %d", i, -r); 1200 } 1201 assert (e >= 0); 1202 } 1203 fail: 1204 return r; 1205 } 1206 1207 /* Stop processing guest IO notifications in vhost. 1208 * Start processing them in qemu. 1209 * This might actually run the qemu handlers right away, 1210 * so virtio in qemu must be completely setup when this is called. 1211 */ 1212 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1213 { 1214 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1215 int i, r; 1216 1217 for (i = 0; i < hdev->nvqs; ++i) { 1218 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1219 false); 1220 if (r < 0) { 1221 error_report("vhost VQ %d notifier cleanup failed: %d", i, -r); 1222 } 1223 assert (r >= 0); 1224 } 1225 } 1226 1227 /* Test and clear event pending status. 1228 * Should be called after unmask to avoid losing events. 1229 */ 1230 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n) 1231 { 1232 struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index; 1233 assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs); 1234 return event_notifier_test_and_clear(&vq->masked_notifier); 1235 } 1236 1237 /* Mask/unmask events from this vq. */ 1238 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n, 1239 bool mask) 1240 { 1241 struct VirtQueue *vvq = virtio_get_queue(vdev, n); 1242 int r, index = n - hdev->vq_index; 1243 struct vhost_vring_file file; 1244 1245 /* should only be called after backend is connected */ 1246 assert(hdev->vhost_ops); 1247 1248 if (mask) { 1249 assert(vdev->use_guest_notifier_mask); 1250 file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier); 1251 } else { 1252 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq)); 1253 } 1254 1255 file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n); 1256 r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file); 1257 if (r < 0) { 1258 VHOST_OPS_DEBUG("vhost_set_vring_call failed"); 1259 } 1260 } 1261 1262 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits, 1263 uint64_t features) 1264 { 1265 const int *bit = feature_bits; 1266 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1267 uint64_t bit_mask = (1ULL << *bit); 1268 if (!(hdev->features & bit_mask)) { 1269 features &= ~bit_mask; 1270 } 1271 bit++; 1272 } 1273 return features; 1274 } 1275 1276 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits, 1277 uint64_t features) 1278 { 1279 const int *bit = feature_bits; 1280 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1281 uint64_t bit_mask = (1ULL << *bit); 1282 if (features & bit_mask) { 1283 hdev->acked_features |= bit_mask; 1284 } 1285 bit++; 1286 } 1287 } 1288 1289 /* Host notifiers must be enabled at this point. */ 1290 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev) 1291 { 1292 int i, r; 1293 1294 /* should only be called after backend is connected */ 1295 assert(hdev->vhost_ops); 1296 1297 hdev->started = true; 1298 1299 r = vhost_dev_set_features(hdev, hdev->log_enabled); 1300 if (r < 0) { 1301 goto fail_features; 1302 } 1303 r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem); 1304 if (r < 0) { 1305 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 1306 r = -errno; 1307 goto fail_mem; 1308 } 1309 for (i = 0; i < hdev->nvqs; ++i) { 1310 r = vhost_virtqueue_start(hdev, 1311 vdev, 1312 hdev->vqs + i, 1313 hdev->vq_index + i); 1314 if (r < 0) { 1315 goto fail_vq; 1316 } 1317 } 1318 1319 if (hdev->log_enabled) { 1320 uint64_t log_base; 1321 1322 hdev->log_size = vhost_get_log_size(hdev); 1323 hdev->log = vhost_log_get(hdev->log_size, 1324 vhost_dev_log_is_shared(hdev)); 1325 log_base = (uintptr_t)hdev->log->log; 1326 r = hdev->vhost_ops->vhost_set_log_base(hdev, 1327 hdev->log_size ? log_base : 0, 1328 hdev->log); 1329 if (r < 0) { 1330 VHOST_OPS_DEBUG("vhost_set_log_base failed"); 1331 r = -errno; 1332 goto fail_log; 1333 } 1334 } 1335 1336 return 0; 1337 fail_log: 1338 vhost_log_put(hdev, false); 1339 fail_vq: 1340 while (--i >= 0) { 1341 vhost_virtqueue_stop(hdev, 1342 vdev, 1343 hdev->vqs + i, 1344 hdev->vq_index + i); 1345 } 1346 i = hdev->nvqs; 1347 fail_mem: 1348 fail_features: 1349 1350 hdev->started = false; 1351 return r; 1352 } 1353 1354 /* Host notifiers must be enabled at this point. */ 1355 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev) 1356 { 1357 int i; 1358 1359 /* should only be called after backend is connected */ 1360 assert(hdev->vhost_ops); 1361 1362 for (i = 0; i < hdev->nvqs; ++i) { 1363 vhost_virtqueue_stop(hdev, 1364 vdev, 1365 hdev->vqs + i, 1366 hdev->vq_index + i); 1367 } 1368 1369 vhost_log_put(hdev, true); 1370 hdev->started = false; 1371 } 1372