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