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 "qemu/atomic.h" 20 #include "qemu/range.h" 21 #include "qemu/error-report.h" 22 #include "qemu/memfd.h" 23 #include "standard-headers/linux/vhost_types.h" 24 #include "exec/address-spaces.h" 25 #include "hw/virtio/virtio-bus.h" 26 #include "hw/virtio/virtio-access.h" 27 #include "migration/blocker.h" 28 #include "migration/qemu-file-types.h" 29 #include "sysemu/dma.h" 30 #include "sysemu/tcg.h" 31 #include "trace.h" 32 33 /* enabled until disconnected backend stabilizes */ 34 #define _VHOST_DEBUG 1 35 36 #ifdef _VHOST_DEBUG 37 #define VHOST_OPS_DEBUG(fmt, ...) \ 38 do { error_report(fmt ": %s (%d)", ## __VA_ARGS__, \ 39 strerror(errno), errno); } while (0) 40 #else 41 #define VHOST_OPS_DEBUG(fmt, ...) \ 42 do { } while (0) 43 #endif 44 45 static struct vhost_log *vhost_log; 46 static struct vhost_log *vhost_log_shm; 47 48 static unsigned int used_memslots; 49 static QLIST_HEAD(, vhost_dev) vhost_devices = 50 QLIST_HEAD_INITIALIZER(vhost_devices); 51 52 bool vhost_has_free_slot(void) 53 { 54 unsigned int slots_limit = ~0U; 55 struct vhost_dev *hdev; 56 57 QLIST_FOREACH(hdev, &vhost_devices, entry) { 58 unsigned int r = hdev->vhost_ops->vhost_backend_memslots_limit(hdev); 59 slots_limit = MIN(slots_limit, r); 60 } 61 return slots_limit > used_memslots; 62 } 63 64 static void vhost_dev_sync_region(struct vhost_dev *dev, 65 MemoryRegionSection *section, 66 uint64_t mfirst, uint64_t mlast, 67 uint64_t rfirst, uint64_t rlast) 68 { 69 vhost_log_chunk_t *log = dev->log->log; 70 71 uint64_t start = MAX(mfirst, rfirst); 72 uint64_t end = MIN(mlast, rlast); 73 vhost_log_chunk_t *from = log + start / VHOST_LOG_CHUNK; 74 vhost_log_chunk_t *to = log + end / VHOST_LOG_CHUNK + 1; 75 uint64_t addr = QEMU_ALIGN_DOWN(start, VHOST_LOG_CHUNK); 76 77 if (end < start) { 78 return; 79 } 80 assert(end / VHOST_LOG_CHUNK < dev->log_size); 81 assert(start / VHOST_LOG_CHUNK < dev->log_size); 82 83 for (;from < to; ++from) { 84 vhost_log_chunk_t log; 85 /* We first check with non-atomic: much cheaper, 86 * and we expect non-dirty to be the common case. */ 87 if (!*from) { 88 addr += VHOST_LOG_CHUNK; 89 continue; 90 } 91 /* Data must be read atomically. We don't really need barrier semantics 92 * but it's easier to use atomic_* than roll our own. */ 93 log = qatomic_xchg(from, 0); 94 while (log) { 95 int bit = ctzl(log); 96 hwaddr page_addr; 97 hwaddr section_offset; 98 hwaddr mr_offset; 99 page_addr = addr + bit * VHOST_LOG_PAGE; 100 section_offset = page_addr - section->offset_within_address_space; 101 mr_offset = section_offset + section->offset_within_region; 102 memory_region_set_dirty(section->mr, mr_offset, VHOST_LOG_PAGE); 103 log &= ~(0x1ull << bit); 104 } 105 addr += VHOST_LOG_CHUNK; 106 } 107 } 108 109 static int vhost_sync_dirty_bitmap(struct vhost_dev *dev, 110 MemoryRegionSection *section, 111 hwaddr first, 112 hwaddr last) 113 { 114 int i; 115 hwaddr start_addr; 116 hwaddr end_addr; 117 118 if (!dev->log_enabled || !dev->started) { 119 return 0; 120 } 121 start_addr = section->offset_within_address_space; 122 end_addr = range_get_last(start_addr, int128_get64(section->size)); 123 start_addr = MAX(first, start_addr); 124 end_addr = MIN(last, end_addr); 125 126 for (i = 0; i < dev->mem->nregions; ++i) { 127 struct vhost_memory_region *reg = dev->mem->regions + i; 128 vhost_dev_sync_region(dev, section, start_addr, end_addr, 129 reg->guest_phys_addr, 130 range_get_last(reg->guest_phys_addr, 131 reg->memory_size)); 132 } 133 for (i = 0; i < dev->nvqs; ++i) { 134 struct vhost_virtqueue *vq = dev->vqs + i; 135 136 if (!vq->used_phys && !vq->used_size) { 137 continue; 138 } 139 140 vhost_dev_sync_region(dev, section, start_addr, end_addr, vq->used_phys, 141 range_get_last(vq->used_phys, vq->used_size)); 142 } 143 return 0; 144 } 145 146 static void vhost_log_sync(MemoryListener *listener, 147 MemoryRegionSection *section) 148 { 149 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 150 memory_listener); 151 vhost_sync_dirty_bitmap(dev, section, 0x0, ~0x0ULL); 152 } 153 154 static void vhost_log_sync_range(struct vhost_dev *dev, 155 hwaddr first, hwaddr last) 156 { 157 int i; 158 /* FIXME: this is N^2 in number of sections */ 159 for (i = 0; i < dev->n_mem_sections; ++i) { 160 MemoryRegionSection *section = &dev->mem_sections[i]; 161 vhost_sync_dirty_bitmap(dev, section, first, last); 162 } 163 } 164 165 static uint64_t vhost_get_log_size(struct vhost_dev *dev) 166 { 167 uint64_t log_size = 0; 168 int i; 169 for (i = 0; i < dev->mem->nregions; ++i) { 170 struct vhost_memory_region *reg = dev->mem->regions + i; 171 uint64_t last = range_get_last(reg->guest_phys_addr, 172 reg->memory_size); 173 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 174 } 175 for (i = 0; i < dev->nvqs; ++i) { 176 struct vhost_virtqueue *vq = dev->vqs + i; 177 178 if (!vq->used_phys && !vq->used_size) { 179 continue; 180 } 181 182 uint64_t last = vq->used_phys + vq->used_size - 1; 183 log_size = MAX(log_size, last / VHOST_LOG_CHUNK + 1); 184 } 185 return log_size; 186 } 187 188 static struct vhost_log *vhost_log_alloc(uint64_t size, bool share) 189 { 190 Error *err = NULL; 191 struct vhost_log *log; 192 uint64_t logsize = size * sizeof(*(log->log)); 193 int fd = -1; 194 195 log = g_new0(struct vhost_log, 1); 196 if (share) { 197 log->log = qemu_memfd_alloc("vhost-log", logsize, 198 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL, 199 &fd, &err); 200 if (err) { 201 error_report_err(err); 202 g_free(log); 203 return NULL; 204 } 205 memset(log->log, 0, logsize); 206 } else { 207 log->log = g_malloc0(logsize); 208 } 209 210 log->size = size; 211 log->refcnt = 1; 212 log->fd = fd; 213 214 return log; 215 } 216 217 static struct vhost_log *vhost_log_get(uint64_t size, bool share) 218 { 219 struct vhost_log *log = share ? vhost_log_shm : vhost_log; 220 221 if (!log || log->size != size) { 222 log = vhost_log_alloc(size, share); 223 if (share) { 224 vhost_log_shm = log; 225 } else { 226 vhost_log = log; 227 } 228 } else { 229 ++log->refcnt; 230 } 231 232 return log; 233 } 234 235 static void vhost_log_put(struct vhost_dev *dev, bool sync) 236 { 237 struct vhost_log *log = dev->log; 238 239 if (!log) { 240 return; 241 } 242 243 --log->refcnt; 244 if (log->refcnt == 0) { 245 /* Sync only the range covered by the old log */ 246 if (dev->log_size && sync) { 247 vhost_log_sync_range(dev, 0, dev->log_size * VHOST_LOG_CHUNK - 1); 248 } 249 250 if (vhost_log == log) { 251 g_free(log->log); 252 vhost_log = NULL; 253 } else if (vhost_log_shm == log) { 254 qemu_memfd_free(log->log, log->size * sizeof(*(log->log)), 255 log->fd); 256 vhost_log_shm = NULL; 257 } 258 259 g_free(log); 260 } 261 262 dev->log = NULL; 263 dev->log_size = 0; 264 } 265 266 static bool vhost_dev_log_is_shared(struct vhost_dev *dev) 267 { 268 return dev->vhost_ops->vhost_requires_shm_log && 269 dev->vhost_ops->vhost_requires_shm_log(dev); 270 } 271 272 static inline void vhost_dev_log_resize(struct vhost_dev *dev, uint64_t size) 273 { 274 struct vhost_log *log = vhost_log_get(size, vhost_dev_log_is_shared(dev)); 275 uint64_t log_base = (uintptr_t)log->log; 276 int r; 277 278 /* inform backend of log switching, this must be done before 279 releasing the current log, to ensure no logging is lost */ 280 r = dev->vhost_ops->vhost_set_log_base(dev, log_base, log); 281 if (r < 0) { 282 VHOST_OPS_DEBUG("vhost_set_log_base failed"); 283 } 284 285 vhost_log_put(dev, true); 286 dev->log = log; 287 dev->log_size = size; 288 } 289 290 static int vhost_dev_has_iommu(struct vhost_dev *dev) 291 { 292 VirtIODevice *vdev = dev->vdev; 293 294 /* 295 * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support 296 * incremental memory mapping API via IOTLB API. For platform that 297 * does not have IOMMU, there's no need to enable this feature 298 * which may cause unnecessary IOTLB miss/update trnasactions. 299 */ 300 return vdev->dma_as != &address_space_memory && 301 virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM); 302 } 303 304 static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr, 305 hwaddr *plen, bool is_write) 306 { 307 if (!vhost_dev_has_iommu(dev)) { 308 return cpu_physical_memory_map(addr, plen, is_write); 309 } else { 310 return (void *)(uintptr_t)addr; 311 } 312 } 313 314 static void vhost_memory_unmap(struct vhost_dev *dev, void *buffer, 315 hwaddr len, int is_write, 316 hwaddr access_len) 317 { 318 if (!vhost_dev_has_iommu(dev)) { 319 cpu_physical_memory_unmap(buffer, len, is_write, access_len); 320 } 321 } 322 323 static int vhost_verify_ring_part_mapping(void *ring_hva, 324 uint64_t ring_gpa, 325 uint64_t ring_size, 326 void *reg_hva, 327 uint64_t reg_gpa, 328 uint64_t reg_size) 329 { 330 uint64_t hva_ring_offset; 331 uint64_t ring_last = range_get_last(ring_gpa, ring_size); 332 uint64_t reg_last = range_get_last(reg_gpa, reg_size); 333 334 if (ring_last < reg_gpa || ring_gpa > reg_last) { 335 return 0; 336 } 337 /* check that whole ring's is mapped */ 338 if (ring_last > reg_last) { 339 return -ENOMEM; 340 } 341 /* check that ring's MemoryRegion wasn't replaced */ 342 hva_ring_offset = ring_gpa - reg_gpa; 343 if (ring_hva != reg_hva + hva_ring_offset) { 344 return -EBUSY; 345 } 346 347 return 0; 348 } 349 350 static int vhost_verify_ring_mappings(struct vhost_dev *dev, 351 void *reg_hva, 352 uint64_t reg_gpa, 353 uint64_t reg_size) 354 { 355 int i, j; 356 int r = 0; 357 const char *part_name[] = { 358 "descriptor table", 359 "available ring", 360 "used ring" 361 }; 362 363 if (vhost_dev_has_iommu(dev)) { 364 return 0; 365 } 366 367 for (i = 0; i < dev->nvqs; ++i) { 368 struct vhost_virtqueue *vq = dev->vqs + i; 369 370 if (vq->desc_phys == 0) { 371 continue; 372 } 373 374 j = 0; 375 r = vhost_verify_ring_part_mapping( 376 vq->desc, vq->desc_phys, vq->desc_size, 377 reg_hva, reg_gpa, reg_size); 378 if (r) { 379 break; 380 } 381 382 j++; 383 r = vhost_verify_ring_part_mapping( 384 vq->avail, vq->avail_phys, vq->avail_size, 385 reg_hva, reg_gpa, reg_size); 386 if (r) { 387 break; 388 } 389 390 j++; 391 r = vhost_verify_ring_part_mapping( 392 vq->used, vq->used_phys, vq->used_size, 393 reg_hva, reg_gpa, reg_size); 394 if (r) { 395 break; 396 } 397 } 398 399 if (r == -ENOMEM) { 400 error_report("Unable to map %s for ring %d", part_name[j], i); 401 } else if (r == -EBUSY) { 402 error_report("%s relocated for ring %d", part_name[j], i); 403 } 404 return r; 405 } 406 407 /* 408 * vhost_section: identify sections needed for vhost access 409 * 410 * We only care about RAM sections here (where virtqueue and guest 411 * internals accessed by virtio might live). If we find one we still 412 * allow the backend to potentially filter it out of our list. 413 */ 414 static bool vhost_section(struct vhost_dev *dev, MemoryRegionSection *section) 415 { 416 MemoryRegion *mr = section->mr; 417 418 if (memory_region_is_ram(mr) && !memory_region_is_rom(mr)) { 419 uint8_t dirty_mask = memory_region_get_dirty_log_mask(mr); 420 uint8_t handled_dirty; 421 422 /* 423 * Kernel based vhost doesn't handle any block which is doing 424 * dirty-tracking other than migration for which it has 425 * specific logging support. However for TCG the kernel never 426 * gets involved anyway so we can also ignore it's 427 * self-modiying code detection flags. However a vhost-user 428 * client could still confuse a TCG guest if it re-writes 429 * executable memory that has already been translated. 430 */ 431 handled_dirty = (1 << DIRTY_MEMORY_MIGRATION) | 432 (1 << DIRTY_MEMORY_CODE); 433 434 if (dirty_mask & ~handled_dirty) { 435 trace_vhost_reject_section(mr->name, 1); 436 return false; 437 } 438 439 if (dev->vhost_ops->vhost_backend_mem_section_filter && 440 !dev->vhost_ops->vhost_backend_mem_section_filter(dev, section)) { 441 trace_vhost_reject_section(mr->name, 2); 442 return false; 443 } 444 445 trace_vhost_section(mr->name); 446 return true; 447 } else { 448 trace_vhost_reject_section(mr->name, 3); 449 return false; 450 } 451 } 452 453 static void vhost_begin(MemoryListener *listener) 454 { 455 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 456 memory_listener); 457 dev->tmp_sections = NULL; 458 dev->n_tmp_sections = 0; 459 } 460 461 static void vhost_commit(MemoryListener *listener) 462 { 463 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 464 memory_listener); 465 MemoryRegionSection *old_sections; 466 int n_old_sections; 467 uint64_t log_size; 468 size_t regions_size; 469 int r; 470 int i; 471 bool changed = false; 472 473 /* Note we can be called before the device is started, but then 474 * starting the device calls set_mem_table, so we need to have 475 * built the data structures. 476 */ 477 old_sections = dev->mem_sections; 478 n_old_sections = dev->n_mem_sections; 479 dev->mem_sections = dev->tmp_sections; 480 dev->n_mem_sections = dev->n_tmp_sections; 481 482 if (dev->n_mem_sections != n_old_sections) { 483 changed = true; 484 } else { 485 /* Same size, lets check the contents */ 486 for (int i = 0; i < n_old_sections; i++) { 487 if (!MemoryRegionSection_eq(&old_sections[i], 488 &dev->mem_sections[i])) { 489 changed = true; 490 break; 491 } 492 } 493 } 494 495 trace_vhost_commit(dev->started, changed); 496 if (!changed) { 497 goto out; 498 } 499 500 /* Rebuild the regions list from the new sections list */ 501 regions_size = offsetof(struct vhost_memory, regions) + 502 dev->n_mem_sections * sizeof dev->mem->regions[0]; 503 dev->mem = g_realloc(dev->mem, regions_size); 504 dev->mem->nregions = dev->n_mem_sections; 505 used_memslots = dev->mem->nregions; 506 for (i = 0; i < dev->n_mem_sections; i++) { 507 struct vhost_memory_region *cur_vmr = dev->mem->regions + i; 508 struct MemoryRegionSection *mrs = dev->mem_sections + i; 509 510 cur_vmr->guest_phys_addr = mrs->offset_within_address_space; 511 cur_vmr->memory_size = int128_get64(mrs->size); 512 cur_vmr->userspace_addr = 513 (uintptr_t)memory_region_get_ram_ptr(mrs->mr) + 514 mrs->offset_within_region; 515 cur_vmr->flags_padding = 0; 516 } 517 518 if (!dev->started) { 519 goto out; 520 } 521 522 for (i = 0; i < dev->mem->nregions; i++) { 523 if (vhost_verify_ring_mappings(dev, 524 (void *)(uintptr_t)dev->mem->regions[i].userspace_addr, 525 dev->mem->regions[i].guest_phys_addr, 526 dev->mem->regions[i].memory_size)) { 527 error_report("Verify ring failure on region %d", i); 528 abort(); 529 } 530 } 531 532 if (!dev->log_enabled) { 533 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); 534 if (r < 0) { 535 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 536 } 537 goto out; 538 } 539 log_size = vhost_get_log_size(dev); 540 /* We allocate an extra 4K bytes to log, 541 * to reduce the * number of reallocations. */ 542 #define VHOST_LOG_BUFFER (0x1000 / sizeof *dev->log) 543 /* To log more, must increase log size before table update. */ 544 if (dev->log_size < log_size) { 545 vhost_dev_log_resize(dev, log_size + VHOST_LOG_BUFFER); 546 } 547 r = dev->vhost_ops->vhost_set_mem_table(dev, dev->mem); 548 if (r < 0) { 549 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 550 } 551 /* To log less, can only decrease log size after table update. */ 552 if (dev->log_size > log_size + VHOST_LOG_BUFFER) { 553 vhost_dev_log_resize(dev, log_size); 554 } 555 556 out: 557 /* Deref the old list of sections, this must happen _after_ the 558 * vhost_set_mem_table to ensure the client isn't still using the 559 * section we're about to unref. 560 */ 561 while (n_old_sections--) { 562 memory_region_unref(old_sections[n_old_sections].mr); 563 } 564 g_free(old_sections); 565 return; 566 } 567 568 /* Adds the section data to the tmp_section structure. 569 * It relies on the listener calling us in memory address order 570 * and for each region (via the _add and _nop methods) to 571 * join neighbours. 572 */ 573 static void vhost_region_add_section(struct vhost_dev *dev, 574 MemoryRegionSection *section) 575 { 576 bool need_add = true; 577 uint64_t mrs_size = int128_get64(section->size); 578 uint64_t mrs_gpa = section->offset_within_address_space; 579 uintptr_t mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) + 580 section->offset_within_region; 581 RAMBlock *mrs_rb = section->mr->ram_block; 582 583 trace_vhost_region_add_section(section->mr->name, mrs_gpa, mrs_size, 584 mrs_host); 585 586 if (dev->vhost_ops->backend_type == VHOST_BACKEND_TYPE_USER) { 587 /* Round the section to it's page size */ 588 /* First align the start down to a page boundary */ 589 size_t mrs_page = qemu_ram_pagesize(mrs_rb); 590 uint64_t alignage = mrs_host & (mrs_page - 1); 591 if (alignage) { 592 mrs_host -= alignage; 593 mrs_size += alignage; 594 mrs_gpa -= alignage; 595 } 596 /* Now align the size up to a page boundary */ 597 alignage = mrs_size & (mrs_page - 1); 598 if (alignage) { 599 mrs_size += mrs_page - alignage; 600 } 601 trace_vhost_region_add_section_aligned(section->mr->name, mrs_gpa, 602 mrs_size, mrs_host); 603 } 604 605 if (dev->n_tmp_sections) { 606 /* Since we already have at least one section, lets see if 607 * this extends it; since we're scanning in order, we only 608 * have to look at the last one, and the FlatView that calls 609 * us shouldn't have overlaps. 610 */ 611 MemoryRegionSection *prev_sec = dev->tmp_sections + 612 (dev->n_tmp_sections - 1); 613 uint64_t prev_gpa_start = prev_sec->offset_within_address_space; 614 uint64_t prev_size = int128_get64(prev_sec->size); 615 uint64_t prev_gpa_end = range_get_last(prev_gpa_start, prev_size); 616 uint64_t prev_host_start = 617 (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) + 618 prev_sec->offset_within_region; 619 uint64_t prev_host_end = range_get_last(prev_host_start, prev_size); 620 621 if (mrs_gpa <= (prev_gpa_end + 1)) { 622 /* OK, looks like overlapping/intersecting - it's possible that 623 * the rounding to page sizes has made them overlap, but they should 624 * match up in the same RAMBlock if they do. 625 */ 626 if (mrs_gpa < prev_gpa_start) { 627 error_report("%s:Section '%s' rounded to %"PRIx64 628 " prior to previous '%s' %"PRIx64, 629 __func__, section->mr->name, mrs_gpa, 630 prev_sec->mr->name, prev_gpa_start); 631 /* A way to cleanly fail here would be better */ 632 return; 633 } 634 /* Offset from the start of the previous GPA to this GPA */ 635 size_t offset = mrs_gpa - prev_gpa_start; 636 637 if (prev_host_start + offset == mrs_host && 638 section->mr == prev_sec->mr && 639 (!dev->vhost_ops->vhost_backend_can_merge || 640 dev->vhost_ops->vhost_backend_can_merge(dev, 641 mrs_host, mrs_size, 642 prev_host_start, prev_size))) { 643 uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size); 644 need_add = false; 645 prev_sec->offset_within_address_space = 646 MIN(prev_gpa_start, mrs_gpa); 647 prev_sec->offset_within_region = 648 MIN(prev_host_start, mrs_host) - 649 (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr); 650 prev_sec->size = int128_make64(max_end - MIN(prev_host_start, 651 mrs_host)); 652 trace_vhost_region_add_section_merge(section->mr->name, 653 int128_get64(prev_sec->size), 654 prev_sec->offset_within_address_space, 655 prev_sec->offset_within_region); 656 } else { 657 /* adjoining regions are fine, but overlapping ones with 658 * different blocks/offsets shouldn't happen 659 */ 660 if (mrs_gpa != prev_gpa_end + 1) { 661 error_report("%s: Overlapping but not coherent sections " 662 "at %"PRIx64, 663 __func__, mrs_gpa); 664 return; 665 } 666 } 667 } 668 } 669 670 if (need_add) { 671 ++dev->n_tmp_sections; 672 dev->tmp_sections = g_renew(MemoryRegionSection, dev->tmp_sections, 673 dev->n_tmp_sections); 674 dev->tmp_sections[dev->n_tmp_sections - 1] = *section; 675 /* The flatview isn't stable and we don't use it, making it NULL 676 * means we can memcmp the list. 677 */ 678 dev->tmp_sections[dev->n_tmp_sections - 1].fv = NULL; 679 memory_region_ref(section->mr); 680 } 681 } 682 683 /* Used for both add and nop callbacks */ 684 static void vhost_region_addnop(MemoryListener *listener, 685 MemoryRegionSection *section) 686 { 687 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 688 memory_listener); 689 690 if (!vhost_section(dev, section)) { 691 return; 692 } 693 vhost_region_add_section(dev, section); 694 } 695 696 static void vhost_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) 697 { 698 struct vhost_iommu *iommu = container_of(n, struct vhost_iommu, n); 699 struct vhost_dev *hdev = iommu->hdev; 700 hwaddr iova = iotlb->iova + iommu->iommu_offset; 701 702 if (vhost_backend_invalidate_device_iotlb(hdev, iova, 703 iotlb->addr_mask + 1)) { 704 error_report("Fail to invalidate device iotlb"); 705 } 706 } 707 708 static void vhost_iommu_region_add(MemoryListener *listener, 709 MemoryRegionSection *section) 710 { 711 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 712 iommu_listener); 713 struct vhost_iommu *iommu; 714 Int128 end; 715 int iommu_idx, ret; 716 IOMMUMemoryRegion *iommu_mr; 717 Error *err = NULL; 718 719 if (!memory_region_is_iommu(section->mr)) { 720 return; 721 } 722 723 iommu_mr = IOMMU_MEMORY_REGION(section->mr); 724 725 iommu = g_malloc0(sizeof(*iommu)); 726 end = int128_add(int128_make64(section->offset_within_region), 727 section->size); 728 end = int128_sub(end, int128_one()); 729 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, 730 MEMTXATTRS_UNSPECIFIED); 731 iommu_notifier_init(&iommu->n, vhost_iommu_unmap_notify, 732 IOMMU_NOTIFIER_UNMAP, 733 section->offset_within_region, 734 int128_get64(end), 735 iommu_idx); 736 iommu->mr = section->mr; 737 iommu->iommu_offset = section->offset_within_address_space - 738 section->offset_within_region; 739 iommu->hdev = dev; 740 ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, &err); 741 if (ret) { 742 error_report_err(err); 743 exit(1); 744 } 745 QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next); 746 /* TODO: can replay help performance here? */ 747 } 748 749 static void vhost_iommu_region_del(MemoryListener *listener, 750 MemoryRegionSection *section) 751 { 752 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 753 iommu_listener); 754 struct vhost_iommu *iommu; 755 756 if (!memory_region_is_iommu(section->mr)) { 757 return; 758 } 759 760 QLIST_FOREACH(iommu, &dev->iommu_list, iommu_next) { 761 if (iommu->mr == section->mr && 762 iommu->n.start == section->offset_within_region) { 763 memory_region_unregister_iommu_notifier(iommu->mr, 764 &iommu->n); 765 QLIST_REMOVE(iommu, iommu_next); 766 g_free(iommu); 767 break; 768 } 769 } 770 } 771 772 static int vhost_virtqueue_set_addr(struct vhost_dev *dev, 773 struct vhost_virtqueue *vq, 774 unsigned idx, bool enable_log) 775 { 776 struct vhost_vring_addr addr; 777 int r; 778 memset(&addr, 0, sizeof(struct vhost_vring_addr)); 779 780 if (dev->vhost_ops->vhost_vq_get_addr) { 781 r = dev->vhost_ops->vhost_vq_get_addr(dev, &addr, vq); 782 if (r < 0) { 783 VHOST_OPS_DEBUG("vhost_vq_get_addr failed"); 784 return -errno; 785 } 786 } else { 787 addr.desc_user_addr = (uint64_t)(unsigned long)vq->desc; 788 addr.avail_user_addr = (uint64_t)(unsigned long)vq->avail; 789 addr.used_user_addr = (uint64_t)(unsigned long)vq->used; 790 } 791 addr.index = idx; 792 addr.log_guest_addr = vq->used_phys; 793 addr.flags = enable_log ? (1 << VHOST_VRING_F_LOG) : 0; 794 r = dev->vhost_ops->vhost_set_vring_addr(dev, &addr); 795 if (r < 0) { 796 VHOST_OPS_DEBUG("vhost_set_vring_addr failed"); 797 return -errno; 798 } 799 return 0; 800 } 801 802 static int vhost_dev_set_features(struct vhost_dev *dev, 803 bool enable_log) 804 { 805 uint64_t features = dev->acked_features; 806 int r; 807 if (enable_log) { 808 features |= 0x1ULL << VHOST_F_LOG_ALL; 809 } 810 if (!vhost_dev_has_iommu(dev)) { 811 features &= ~(0x1ULL << VIRTIO_F_IOMMU_PLATFORM); 812 } 813 if (dev->vhost_ops->vhost_force_iommu) { 814 if (dev->vhost_ops->vhost_force_iommu(dev) == true) { 815 features |= 0x1ULL << VIRTIO_F_IOMMU_PLATFORM; 816 } 817 } 818 r = dev->vhost_ops->vhost_set_features(dev, features); 819 if (r < 0) { 820 VHOST_OPS_DEBUG("vhost_set_features failed"); 821 goto out; 822 } 823 if (dev->vhost_ops->vhost_set_backend_cap) { 824 r = dev->vhost_ops->vhost_set_backend_cap(dev); 825 if (r < 0) { 826 VHOST_OPS_DEBUG("vhost_set_backend_cap failed"); 827 goto out; 828 } 829 } 830 831 out: 832 return r < 0 ? -errno : 0; 833 } 834 835 static int vhost_dev_set_log(struct vhost_dev *dev, bool enable_log) 836 { 837 int r, i, idx; 838 r = vhost_dev_set_features(dev, enable_log); 839 if (r < 0) { 840 goto err_features; 841 } 842 for (i = 0; i < dev->nvqs; ++i) { 843 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); 844 r = vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, 845 enable_log); 846 if (r < 0) { 847 goto err_vq; 848 } 849 } 850 return 0; 851 err_vq: 852 for (; i >= 0; --i) { 853 idx = dev->vhost_ops->vhost_get_vq_index(dev, dev->vq_index + i); 854 vhost_virtqueue_set_addr(dev, dev->vqs + i, idx, 855 dev->log_enabled); 856 } 857 vhost_dev_set_features(dev, dev->log_enabled); 858 err_features: 859 return r; 860 } 861 862 static int vhost_migration_log(MemoryListener *listener, bool enable) 863 { 864 struct vhost_dev *dev = container_of(listener, struct vhost_dev, 865 memory_listener); 866 int r; 867 if (enable == dev->log_enabled) { 868 return 0; 869 } 870 if (!dev->started) { 871 dev->log_enabled = enable; 872 return 0; 873 } 874 875 r = 0; 876 if (!enable) { 877 r = vhost_dev_set_log(dev, false); 878 if (r < 0) { 879 goto check_dev_state; 880 } 881 vhost_log_put(dev, false); 882 } else { 883 vhost_dev_log_resize(dev, vhost_get_log_size(dev)); 884 r = vhost_dev_set_log(dev, true); 885 if (r < 0) { 886 goto check_dev_state; 887 } 888 } 889 890 check_dev_state: 891 dev->log_enabled = enable; 892 /* 893 * vhost-user-* devices could change their state during log 894 * initialization due to disconnect. So check dev state after 895 * vhost communication. 896 */ 897 if (!dev->started) { 898 /* 899 * Since device is in the stopped state, it is okay for 900 * migration. Return success. 901 */ 902 r = 0; 903 } 904 if (r) { 905 /* An error is occured. */ 906 dev->log_enabled = false; 907 } 908 909 return r; 910 } 911 912 static void vhost_log_global_start(MemoryListener *listener) 913 { 914 int r; 915 916 r = vhost_migration_log(listener, true); 917 if (r < 0) { 918 abort(); 919 } 920 } 921 922 static void vhost_log_global_stop(MemoryListener *listener) 923 { 924 int r; 925 926 r = vhost_migration_log(listener, false); 927 if (r < 0) { 928 abort(); 929 } 930 } 931 932 static void vhost_log_start(MemoryListener *listener, 933 MemoryRegionSection *section, 934 int old, int new) 935 { 936 /* FIXME: implement */ 937 } 938 939 static void vhost_log_stop(MemoryListener *listener, 940 MemoryRegionSection *section, 941 int old, int new) 942 { 943 /* FIXME: implement */ 944 } 945 946 /* The vhost driver natively knows how to handle the vrings of non 947 * cross-endian legacy devices and modern devices. Only legacy devices 948 * exposed to a bi-endian guest may require the vhost driver to use a 949 * specific endianness. 950 */ 951 static inline bool vhost_needs_vring_endian(VirtIODevice *vdev) 952 { 953 if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) { 954 return false; 955 } 956 #ifdef HOST_WORDS_BIGENDIAN 957 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE; 958 #else 959 return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG; 960 #endif 961 } 962 963 static int vhost_virtqueue_set_vring_endian_legacy(struct vhost_dev *dev, 964 bool is_big_endian, 965 int vhost_vq_index) 966 { 967 struct vhost_vring_state s = { 968 .index = vhost_vq_index, 969 .num = is_big_endian 970 }; 971 972 if (!dev->vhost_ops->vhost_set_vring_endian(dev, &s)) { 973 return 0; 974 } 975 976 VHOST_OPS_DEBUG("vhost_set_vring_endian failed"); 977 if (errno == ENOTTY) { 978 error_report("vhost does not support cross-endian"); 979 return -ENOSYS; 980 } 981 982 return -errno; 983 } 984 985 static int vhost_memory_region_lookup(struct vhost_dev *hdev, 986 uint64_t gpa, uint64_t *uaddr, 987 uint64_t *len) 988 { 989 int i; 990 991 for (i = 0; i < hdev->mem->nregions; i++) { 992 struct vhost_memory_region *reg = hdev->mem->regions + i; 993 994 if (gpa >= reg->guest_phys_addr && 995 reg->guest_phys_addr + reg->memory_size > gpa) { 996 *uaddr = reg->userspace_addr + gpa - reg->guest_phys_addr; 997 *len = reg->guest_phys_addr + reg->memory_size - gpa; 998 return 0; 999 } 1000 } 1001 1002 return -EFAULT; 1003 } 1004 1005 int vhost_device_iotlb_miss(struct vhost_dev *dev, uint64_t iova, int write) 1006 { 1007 IOMMUTLBEntry iotlb; 1008 uint64_t uaddr, len; 1009 int ret = -EFAULT; 1010 1011 RCU_READ_LOCK_GUARD(); 1012 1013 trace_vhost_iotlb_miss(dev, 1); 1014 1015 iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as, 1016 iova, write, 1017 MEMTXATTRS_UNSPECIFIED); 1018 if (iotlb.target_as != NULL) { 1019 ret = vhost_memory_region_lookup(dev, iotlb.translated_addr, 1020 &uaddr, &len); 1021 if (ret) { 1022 trace_vhost_iotlb_miss(dev, 3); 1023 error_report("Fail to lookup the translated address " 1024 "%"PRIx64, iotlb.translated_addr); 1025 goto out; 1026 } 1027 1028 len = MIN(iotlb.addr_mask + 1, len); 1029 iova = iova & ~iotlb.addr_mask; 1030 1031 ret = vhost_backend_update_device_iotlb(dev, iova, uaddr, 1032 len, iotlb.perm); 1033 if (ret) { 1034 trace_vhost_iotlb_miss(dev, 4); 1035 error_report("Fail to update device iotlb"); 1036 goto out; 1037 } 1038 } 1039 1040 trace_vhost_iotlb_miss(dev, 2); 1041 1042 out: 1043 return ret; 1044 } 1045 1046 static int vhost_virtqueue_start(struct vhost_dev *dev, 1047 struct VirtIODevice *vdev, 1048 struct vhost_virtqueue *vq, 1049 unsigned idx) 1050 { 1051 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1052 VirtioBusState *vbus = VIRTIO_BUS(qbus); 1053 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(vbus); 1054 hwaddr s, l, a; 1055 int r; 1056 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); 1057 struct vhost_vring_file file = { 1058 .index = vhost_vq_index 1059 }; 1060 struct vhost_vring_state state = { 1061 .index = vhost_vq_index 1062 }; 1063 struct VirtQueue *vvq = virtio_get_queue(vdev, idx); 1064 1065 a = virtio_queue_get_desc_addr(vdev, idx); 1066 if (a == 0) { 1067 /* Queue might not be ready for start */ 1068 return 0; 1069 } 1070 1071 vq->num = state.num = virtio_queue_get_num(vdev, idx); 1072 r = dev->vhost_ops->vhost_set_vring_num(dev, &state); 1073 if (r) { 1074 VHOST_OPS_DEBUG("vhost_set_vring_num failed"); 1075 return -errno; 1076 } 1077 1078 state.num = virtio_queue_get_last_avail_idx(vdev, idx); 1079 r = dev->vhost_ops->vhost_set_vring_base(dev, &state); 1080 if (r) { 1081 VHOST_OPS_DEBUG("vhost_set_vring_base failed"); 1082 return -errno; 1083 } 1084 1085 if (vhost_needs_vring_endian(vdev)) { 1086 r = vhost_virtqueue_set_vring_endian_legacy(dev, 1087 virtio_is_big_endian(vdev), 1088 vhost_vq_index); 1089 if (r) { 1090 return -errno; 1091 } 1092 } 1093 1094 vq->desc_size = s = l = virtio_queue_get_desc_size(vdev, idx); 1095 vq->desc_phys = a; 1096 vq->desc = vhost_memory_map(dev, a, &l, false); 1097 if (!vq->desc || l != s) { 1098 r = -ENOMEM; 1099 goto fail_alloc_desc; 1100 } 1101 vq->avail_size = s = l = virtio_queue_get_avail_size(vdev, idx); 1102 vq->avail_phys = a = virtio_queue_get_avail_addr(vdev, idx); 1103 vq->avail = vhost_memory_map(dev, a, &l, false); 1104 if (!vq->avail || l != s) { 1105 r = -ENOMEM; 1106 goto fail_alloc_avail; 1107 } 1108 vq->used_size = s = l = virtio_queue_get_used_size(vdev, idx); 1109 vq->used_phys = a = virtio_queue_get_used_addr(vdev, idx); 1110 vq->used = vhost_memory_map(dev, a, &l, true); 1111 if (!vq->used || l != s) { 1112 r = -ENOMEM; 1113 goto fail_alloc_used; 1114 } 1115 1116 r = vhost_virtqueue_set_addr(dev, vq, vhost_vq_index, dev->log_enabled); 1117 if (r < 0) { 1118 r = -errno; 1119 goto fail_alloc; 1120 } 1121 1122 file.fd = event_notifier_get_fd(virtio_queue_get_host_notifier(vvq)); 1123 r = dev->vhost_ops->vhost_set_vring_kick(dev, &file); 1124 if (r) { 1125 VHOST_OPS_DEBUG("vhost_set_vring_kick failed"); 1126 r = -errno; 1127 goto fail_kick; 1128 } 1129 1130 /* Clear and discard previous events if any. */ 1131 event_notifier_test_and_clear(&vq->masked_notifier); 1132 1133 /* Init vring in unmasked state, unless guest_notifier_mask 1134 * will do it later. 1135 */ 1136 if (!vdev->use_guest_notifier_mask) { 1137 /* TODO: check and handle errors. */ 1138 vhost_virtqueue_mask(dev, vdev, idx, false); 1139 } 1140 1141 if (k->query_guest_notifiers && 1142 k->query_guest_notifiers(qbus->parent) && 1143 virtio_queue_vector(vdev, idx) == VIRTIO_NO_VECTOR) { 1144 file.fd = -1; 1145 r = dev->vhost_ops->vhost_set_vring_call(dev, &file); 1146 if (r) { 1147 goto fail_vector; 1148 } 1149 } 1150 1151 return 0; 1152 1153 fail_vector: 1154 fail_kick: 1155 fail_alloc: 1156 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx), 1157 0, 0); 1158 fail_alloc_used: 1159 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx), 1160 0, 0); 1161 fail_alloc_avail: 1162 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx), 1163 0, 0); 1164 fail_alloc_desc: 1165 return r; 1166 } 1167 1168 static void vhost_virtqueue_stop(struct vhost_dev *dev, 1169 struct VirtIODevice *vdev, 1170 struct vhost_virtqueue *vq, 1171 unsigned idx) 1172 { 1173 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, idx); 1174 struct vhost_vring_state state = { 1175 .index = vhost_vq_index, 1176 }; 1177 int r; 1178 1179 if (virtio_queue_get_desc_addr(vdev, idx) == 0) { 1180 /* Don't stop the virtqueue which might have not been started */ 1181 return; 1182 } 1183 1184 r = dev->vhost_ops->vhost_get_vring_base(dev, &state); 1185 if (r < 0) { 1186 VHOST_OPS_DEBUG("vhost VQ %u ring restore failed: %d", idx, r); 1187 /* Connection to the backend is broken, so let's sync internal 1188 * last avail idx to the device used idx. 1189 */ 1190 virtio_queue_restore_last_avail_idx(vdev, idx); 1191 } else { 1192 virtio_queue_set_last_avail_idx(vdev, idx, state.num); 1193 } 1194 virtio_queue_invalidate_signalled_used(vdev, idx); 1195 virtio_queue_update_used_idx(vdev, idx); 1196 1197 /* In the cross-endian case, we need to reset the vring endianness to 1198 * native as legacy devices expect so by default. 1199 */ 1200 if (vhost_needs_vring_endian(vdev)) { 1201 vhost_virtqueue_set_vring_endian_legacy(dev, 1202 !virtio_is_big_endian(vdev), 1203 vhost_vq_index); 1204 } 1205 1206 vhost_memory_unmap(dev, vq->used, virtio_queue_get_used_size(vdev, idx), 1207 1, virtio_queue_get_used_size(vdev, idx)); 1208 vhost_memory_unmap(dev, vq->avail, virtio_queue_get_avail_size(vdev, idx), 1209 0, virtio_queue_get_avail_size(vdev, idx)); 1210 vhost_memory_unmap(dev, vq->desc, virtio_queue_get_desc_size(vdev, idx), 1211 0, virtio_queue_get_desc_size(vdev, idx)); 1212 } 1213 1214 static void vhost_eventfd_add(MemoryListener *listener, 1215 MemoryRegionSection *section, 1216 bool match_data, uint64_t data, EventNotifier *e) 1217 { 1218 } 1219 1220 static void vhost_eventfd_del(MemoryListener *listener, 1221 MemoryRegionSection *section, 1222 bool match_data, uint64_t data, EventNotifier *e) 1223 { 1224 } 1225 1226 static int vhost_virtqueue_set_busyloop_timeout(struct vhost_dev *dev, 1227 int n, uint32_t timeout) 1228 { 1229 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); 1230 struct vhost_vring_state state = { 1231 .index = vhost_vq_index, 1232 .num = timeout, 1233 }; 1234 int r; 1235 1236 if (!dev->vhost_ops->vhost_set_vring_busyloop_timeout) { 1237 return -EINVAL; 1238 } 1239 1240 r = dev->vhost_ops->vhost_set_vring_busyloop_timeout(dev, &state); 1241 if (r) { 1242 VHOST_OPS_DEBUG("vhost_set_vring_busyloop_timeout failed"); 1243 return r; 1244 } 1245 1246 return 0; 1247 } 1248 1249 static int vhost_virtqueue_init(struct vhost_dev *dev, 1250 struct vhost_virtqueue *vq, int n) 1251 { 1252 int vhost_vq_index = dev->vhost_ops->vhost_get_vq_index(dev, n); 1253 struct vhost_vring_file file = { 1254 .index = vhost_vq_index, 1255 }; 1256 int r = event_notifier_init(&vq->masked_notifier, 0); 1257 if (r < 0) { 1258 return r; 1259 } 1260 1261 file.fd = event_notifier_get_fd(&vq->masked_notifier); 1262 r = dev->vhost_ops->vhost_set_vring_call(dev, &file); 1263 if (r) { 1264 VHOST_OPS_DEBUG("vhost_set_vring_call failed"); 1265 r = -errno; 1266 goto fail_call; 1267 } 1268 1269 vq->dev = dev; 1270 1271 return 0; 1272 fail_call: 1273 event_notifier_cleanup(&vq->masked_notifier); 1274 return r; 1275 } 1276 1277 static void vhost_virtqueue_cleanup(struct vhost_virtqueue *vq) 1278 { 1279 event_notifier_cleanup(&vq->masked_notifier); 1280 } 1281 1282 int vhost_dev_init(struct vhost_dev *hdev, void *opaque, 1283 VhostBackendType backend_type, uint32_t busyloop_timeout) 1284 { 1285 uint64_t features; 1286 int i, r, n_initialized_vqs = 0; 1287 Error *local_err = NULL; 1288 1289 hdev->vdev = NULL; 1290 hdev->migration_blocker = NULL; 1291 1292 r = vhost_set_backend_type(hdev, backend_type); 1293 assert(r >= 0); 1294 1295 r = hdev->vhost_ops->vhost_backend_init(hdev, opaque); 1296 if (r < 0) { 1297 goto fail; 1298 } 1299 1300 r = hdev->vhost_ops->vhost_set_owner(hdev); 1301 if (r < 0) { 1302 VHOST_OPS_DEBUG("vhost_set_owner failed"); 1303 goto fail; 1304 } 1305 1306 r = hdev->vhost_ops->vhost_get_features(hdev, &features); 1307 if (r < 0) { 1308 VHOST_OPS_DEBUG("vhost_get_features failed"); 1309 goto fail; 1310 } 1311 1312 for (i = 0; i < hdev->nvqs; ++i, ++n_initialized_vqs) { 1313 r = vhost_virtqueue_init(hdev, hdev->vqs + i, hdev->vq_index + i); 1314 if (r < 0) { 1315 goto fail; 1316 } 1317 } 1318 1319 if (busyloop_timeout) { 1320 for (i = 0; i < hdev->nvqs; ++i) { 1321 r = vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 1322 busyloop_timeout); 1323 if (r < 0) { 1324 goto fail_busyloop; 1325 } 1326 } 1327 } 1328 1329 hdev->features = features; 1330 1331 hdev->memory_listener = (MemoryListener) { 1332 .begin = vhost_begin, 1333 .commit = vhost_commit, 1334 .region_add = vhost_region_addnop, 1335 .region_nop = vhost_region_addnop, 1336 .log_start = vhost_log_start, 1337 .log_stop = vhost_log_stop, 1338 .log_sync = vhost_log_sync, 1339 .log_global_start = vhost_log_global_start, 1340 .log_global_stop = vhost_log_global_stop, 1341 .eventfd_add = vhost_eventfd_add, 1342 .eventfd_del = vhost_eventfd_del, 1343 .priority = 10 1344 }; 1345 1346 hdev->iommu_listener = (MemoryListener) { 1347 .region_add = vhost_iommu_region_add, 1348 .region_del = vhost_iommu_region_del, 1349 }; 1350 1351 if (hdev->migration_blocker == NULL) { 1352 if (!(hdev->features & (0x1ULL << VHOST_F_LOG_ALL))) { 1353 error_setg(&hdev->migration_blocker, 1354 "Migration disabled: vhost lacks VHOST_F_LOG_ALL feature."); 1355 } else if (vhost_dev_log_is_shared(hdev) && !qemu_memfd_alloc_check()) { 1356 error_setg(&hdev->migration_blocker, 1357 "Migration disabled: failed to allocate shared memory"); 1358 } 1359 } 1360 1361 if (hdev->migration_blocker != NULL) { 1362 r = migrate_add_blocker(hdev->migration_blocker, &local_err); 1363 if (local_err) { 1364 error_report_err(local_err); 1365 error_free(hdev->migration_blocker); 1366 goto fail_busyloop; 1367 } 1368 } 1369 1370 hdev->mem = g_malloc0(offsetof(struct vhost_memory, regions)); 1371 hdev->n_mem_sections = 0; 1372 hdev->mem_sections = NULL; 1373 hdev->log = NULL; 1374 hdev->log_size = 0; 1375 hdev->log_enabled = false; 1376 hdev->started = false; 1377 memory_listener_register(&hdev->memory_listener, &address_space_memory); 1378 QLIST_INSERT_HEAD(&vhost_devices, hdev, entry); 1379 1380 if (used_memslots > hdev->vhost_ops->vhost_backend_memslots_limit(hdev)) { 1381 error_report("vhost backend memory slots limit is less" 1382 " than current number of present memory slots"); 1383 r = -1; 1384 if (busyloop_timeout) { 1385 goto fail_busyloop; 1386 } else { 1387 goto fail; 1388 } 1389 } 1390 1391 return 0; 1392 1393 fail_busyloop: 1394 while (--i >= 0) { 1395 vhost_virtqueue_set_busyloop_timeout(hdev, hdev->vq_index + i, 0); 1396 } 1397 fail: 1398 hdev->nvqs = n_initialized_vqs; 1399 vhost_dev_cleanup(hdev); 1400 return r; 1401 } 1402 1403 void vhost_dev_cleanup(struct vhost_dev *hdev) 1404 { 1405 int i; 1406 1407 for (i = 0; i < hdev->nvqs; ++i) { 1408 vhost_virtqueue_cleanup(hdev->vqs + i); 1409 } 1410 if (hdev->mem) { 1411 /* those are only safe after successful init */ 1412 memory_listener_unregister(&hdev->memory_listener); 1413 QLIST_REMOVE(hdev, entry); 1414 } 1415 if (hdev->migration_blocker) { 1416 migrate_del_blocker(hdev->migration_blocker); 1417 error_free(hdev->migration_blocker); 1418 } 1419 g_free(hdev->mem); 1420 g_free(hdev->mem_sections); 1421 if (hdev->vhost_ops) { 1422 hdev->vhost_ops->vhost_backend_cleanup(hdev); 1423 } 1424 assert(!hdev->log); 1425 1426 memset(hdev, 0, sizeof(struct vhost_dev)); 1427 } 1428 1429 /* Stop processing guest IO notifications in qemu. 1430 * Start processing them in vhost in kernel. 1431 */ 1432 int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1433 { 1434 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1435 int i, r, e; 1436 1437 /* We will pass the notifiers to the kernel, make sure that QEMU 1438 * doesn't interfere. 1439 */ 1440 r = virtio_device_grab_ioeventfd(vdev); 1441 if (r < 0) { 1442 error_report("binding does not support host notifiers"); 1443 goto fail; 1444 } 1445 1446 for (i = 0; i < hdev->nvqs; ++i) { 1447 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1448 true); 1449 if (r < 0) { 1450 error_report("vhost VQ %d notifier binding failed: %d", i, -r); 1451 goto fail_vq; 1452 } 1453 } 1454 1455 return 0; 1456 fail_vq: 1457 while (--i >= 0) { 1458 e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1459 false); 1460 if (e < 0) { 1461 error_report("vhost VQ %d notifier cleanup error: %d", i, -r); 1462 } 1463 assert (e >= 0); 1464 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i); 1465 } 1466 virtio_device_release_ioeventfd(vdev); 1467 fail: 1468 return r; 1469 } 1470 1471 /* Stop processing guest IO notifications in vhost. 1472 * Start processing them in qemu. 1473 * This might actually run the qemu handlers right away, 1474 * so virtio in qemu must be completely setup when this is called. 1475 */ 1476 void vhost_dev_disable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev) 1477 { 1478 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1479 int i, r; 1480 1481 for (i = 0; i < hdev->nvqs; ++i) { 1482 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i, 1483 false); 1484 if (r < 0) { 1485 error_report("vhost VQ %d notifier cleanup failed: %d", i, -r); 1486 } 1487 assert (r >= 0); 1488 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i); 1489 } 1490 virtio_device_release_ioeventfd(vdev); 1491 } 1492 1493 /* Test and clear event pending status. 1494 * Should be called after unmask to avoid losing events. 1495 */ 1496 bool vhost_virtqueue_pending(struct vhost_dev *hdev, int n) 1497 { 1498 struct vhost_virtqueue *vq = hdev->vqs + n - hdev->vq_index; 1499 assert(n >= hdev->vq_index && n < hdev->vq_index + hdev->nvqs); 1500 return event_notifier_test_and_clear(&vq->masked_notifier); 1501 } 1502 1503 /* Mask/unmask events from this vq. */ 1504 void vhost_virtqueue_mask(struct vhost_dev *hdev, VirtIODevice *vdev, int n, 1505 bool mask) 1506 { 1507 struct VirtQueue *vvq = virtio_get_queue(vdev, n); 1508 int r, index = n - hdev->vq_index; 1509 struct vhost_vring_file file; 1510 1511 /* should only be called after backend is connected */ 1512 assert(hdev->vhost_ops); 1513 1514 if (mask) { 1515 assert(vdev->use_guest_notifier_mask); 1516 file.fd = event_notifier_get_fd(&hdev->vqs[index].masked_notifier); 1517 } else { 1518 file.fd = event_notifier_get_fd(virtio_queue_get_guest_notifier(vvq)); 1519 } 1520 1521 file.index = hdev->vhost_ops->vhost_get_vq_index(hdev, n); 1522 r = hdev->vhost_ops->vhost_set_vring_call(hdev, &file); 1523 if (r < 0) { 1524 VHOST_OPS_DEBUG("vhost_set_vring_call failed"); 1525 } 1526 } 1527 1528 uint64_t vhost_get_features(struct vhost_dev *hdev, const int *feature_bits, 1529 uint64_t features) 1530 { 1531 const int *bit = feature_bits; 1532 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1533 uint64_t bit_mask = (1ULL << *bit); 1534 if (!(hdev->features & bit_mask)) { 1535 features &= ~bit_mask; 1536 } 1537 bit++; 1538 } 1539 return features; 1540 } 1541 1542 void vhost_ack_features(struct vhost_dev *hdev, const int *feature_bits, 1543 uint64_t features) 1544 { 1545 const int *bit = feature_bits; 1546 while (*bit != VHOST_INVALID_FEATURE_BIT) { 1547 uint64_t bit_mask = (1ULL << *bit); 1548 if (features & bit_mask) { 1549 hdev->acked_features |= bit_mask; 1550 } 1551 bit++; 1552 } 1553 } 1554 1555 int vhost_dev_get_config(struct vhost_dev *hdev, uint8_t *config, 1556 uint32_t config_len) 1557 { 1558 assert(hdev->vhost_ops); 1559 1560 if (hdev->vhost_ops->vhost_get_config) { 1561 return hdev->vhost_ops->vhost_get_config(hdev, config, config_len); 1562 } 1563 1564 return -1; 1565 } 1566 1567 int vhost_dev_set_config(struct vhost_dev *hdev, const uint8_t *data, 1568 uint32_t offset, uint32_t size, uint32_t flags) 1569 { 1570 assert(hdev->vhost_ops); 1571 1572 if (hdev->vhost_ops->vhost_set_config) { 1573 return hdev->vhost_ops->vhost_set_config(hdev, data, offset, 1574 size, flags); 1575 } 1576 1577 return -1; 1578 } 1579 1580 void vhost_dev_set_config_notifier(struct vhost_dev *hdev, 1581 const VhostDevConfigOps *ops) 1582 { 1583 hdev->config_ops = ops; 1584 } 1585 1586 void vhost_dev_free_inflight(struct vhost_inflight *inflight) 1587 { 1588 if (inflight && inflight->addr) { 1589 qemu_memfd_free(inflight->addr, inflight->size, inflight->fd); 1590 inflight->addr = NULL; 1591 inflight->fd = -1; 1592 } 1593 } 1594 1595 static int vhost_dev_resize_inflight(struct vhost_inflight *inflight, 1596 uint64_t new_size) 1597 { 1598 Error *err = NULL; 1599 int fd = -1; 1600 void *addr = qemu_memfd_alloc("vhost-inflight", new_size, 1601 F_SEAL_GROW | F_SEAL_SHRINK | F_SEAL_SEAL, 1602 &fd, &err); 1603 1604 if (err) { 1605 error_report_err(err); 1606 return -1; 1607 } 1608 1609 vhost_dev_free_inflight(inflight); 1610 inflight->offset = 0; 1611 inflight->addr = addr; 1612 inflight->fd = fd; 1613 inflight->size = new_size; 1614 1615 return 0; 1616 } 1617 1618 void vhost_dev_save_inflight(struct vhost_inflight *inflight, QEMUFile *f) 1619 { 1620 if (inflight->addr) { 1621 qemu_put_be64(f, inflight->size); 1622 qemu_put_be16(f, inflight->queue_size); 1623 qemu_put_buffer(f, inflight->addr, inflight->size); 1624 } else { 1625 qemu_put_be64(f, 0); 1626 } 1627 } 1628 1629 int vhost_dev_load_inflight(struct vhost_inflight *inflight, QEMUFile *f) 1630 { 1631 uint64_t size; 1632 1633 size = qemu_get_be64(f); 1634 if (!size) { 1635 return 0; 1636 } 1637 1638 if (inflight->size != size) { 1639 if (vhost_dev_resize_inflight(inflight, size)) { 1640 return -1; 1641 } 1642 } 1643 inflight->queue_size = qemu_get_be16(f); 1644 1645 qemu_get_buffer(f, inflight->addr, size); 1646 1647 return 0; 1648 } 1649 1650 int vhost_dev_set_inflight(struct vhost_dev *dev, 1651 struct vhost_inflight *inflight) 1652 { 1653 int r; 1654 1655 if (dev->vhost_ops->vhost_set_inflight_fd && inflight->addr) { 1656 r = dev->vhost_ops->vhost_set_inflight_fd(dev, inflight); 1657 if (r) { 1658 VHOST_OPS_DEBUG("vhost_set_inflight_fd failed"); 1659 return -errno; 1660 } 1661 } 1662 1663 return 0; 1664 } 1665 1666 int vhost_dev_get_inflight(struct vhost_dev *dev, uint16_t queue_size, 1667 struct vhost_inflight *inflight) 1668 { 1669 int r; 1670 1671 if (dev->vhost_ops->vhost_get_inflight_fd) { 1672 r = dev->vhost_ops->vhost_get_inflight_fd(dev, queue_size, inflight); 1673 if (r) { 1674 VHOST_OPS_DEBUG("vhost_get_inflight_fd failed"); 1675 return -errno; 1676 } 1677 } 1678 1679 return 0; 1680 } 1681 1682 /* Host notifiers must be enabled at this point. */ 1683 int vhost_dev_start(struct vhost_dev *hdev, VirtIODevice *vdev) 1684 { 1685 int i, r; 1686 1687 /* should only be called after backend is connected */ 1688 assert(hdev->vhost_ops); 1689 1690 hdev->started = true; 1691 hdev->vdev = vdev; 1692 1693 r = vhost_dev_set_features(hdev, hdev->log_enabled); 1694 if (r < 0) { 1695 goto fail_features; 1696 } 1697 1698 if (vhost_dev_has_iommu(hdev)) { 1699 memory_listener_register(&hdev->iommu_listener, vdev->dma_as); 1700 } 1701 1702 r = hdev->vhost_ops->vhost_set_mem_table(hdev, hdev->mem); 1703 if (r < 0) { 1704 VHOST_OPS_DEBUG("vhost_set_mem_table failed"); 1705 r = -errno; 1706 goto fail_mem; 1707 } 1708 for (i = 0; i < hdev->nvqs; ++i) { 1709 r = vhost_virtqueue_start(hdev, 1710 vdev, 1711 hdev->vqs + i, 1712 hdev->vq_index + i); 1713 if (r < 0) { 1714 goto fail_vq; 1715 } 1716 } 1717 1718 if (hdev->log_enabled) { 1719 uint64_t log_base; 1720 1721 hdev->log_size = vhost_get_log_size(hdev); 1722 hdev->log = vhost_log_get(hdev->log_size, 1723 vhost_dev_log_is_shared(hdev)); 1724 log_base = (uintptr_t)hdev->log->log; 1725 r = hdev->vhost_ops->vhost_set_log_base(hdev, 1726 hdev->log_size ? log_base : 0, 1727 hdev->log); 1728 if (r < 0) { 1729 VHOST_OPS_DEBUG("vhost_set_log_base failed"); 1730 r = -errno; 1731 goto fail_log; 1732 } 1733 } 1734 if (hdev->vhost_ops->vhost_dev_start) { 1735 r = hdev->vhost_ops->vhost_dev_start(hdev, true); 1736 if (r) { 1737 goto fail_log; 1738 } 1739 } 1740 if (vhost_dev_has_iommu(hdev) && 1741 hdev->vhost_ops->vhost_set_iotlb_callback) { 1742 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, true); 1743 1744 /* Update used ring information for IOTLB to work correctly, 1745 * vhost-kernel code requires for this.*/ 1746 for (i = 0; i < hdev->nvqs; ++i) { 1747 struct vhost_virtqueue *vq = hdev->vqs + i; 1748 vhost_device_iotlb_miss(hdev, vq->used_phys, true); 1749 } 1750 } 1751 return 0; 1752 fail_log: 1753 vhost_log_put(hdev, false); 1754 fail_vq: 1755 while (--i >= 0) { 1756 vhost_virtqueue_stop(hdev, 1757 vdev, 1758 hdev->vqs + i, 1759 hdev->vq_index + i); 1760 } 1761 1762 fail_mem: 1763 fail_features: 1764 1765 hdev->started = false; 1766 return r; 1767 } 1768 1769 /* Host notifiers must be enabled at this point. */ 1770 void vhost_dev_stop(struct vhost_dev *hdev, VirtIODevice *vdev) 1771 { 1772 int i; 1773 1774 /* should only be called after backend is connected */ 1775 assert(hdev->vhost_ops); 1776 1777 if (hdev->vhost_ops->vhost_dev_start) { 1778 hdev->vhost_ops->vhost_dev_start(hdev, false); 1779 } 1780 for (i = 0; i < hdev->nvqs; ++i) { 1781 vhost_virtqueue_stop(hdev, 1782 vdev, 1783 hdev->vqs + i, 1784 hdev->vq_index + i); 1785 } 1786 1787 if (vhost_dev_has_iommu(hdev)) { 1788 if (hdev->vhost_ops->vhost_set_iotlb_callback) { 1789 hdev->vhost_ops->vhost_set_iotlb_callback(hdev, false); 1790 } 1791 memory_listener_unregister(&hdev->iommu_listener); 1792 } 1793 vhost_log_put(hdev, true); 1794 hdev->started = false; 1795 hdev->vdev = NULL; 1796 } 1797 1798 int vhost_net_set_backend(struct vhost_dev *hdev, 1799 struct vhost_vring_file *file) 1800 { 1801 if (hdev->vhost_ops->vhost_net_set_backend) { 1802 return hdev->vhost_ops->vhost_net_set_backend(hdev, file); 1803 } 1804 1805 return -1; 1806 } 1807