1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/dma-map-ops.h> 5 #include <linux/mm.h> 6 #include <linux/nospec.h> 7 #include <linux/io_uring.h> 8 #include <linux/netdevice.h> 9 #include <linux/rtnetlink.h> 10 #include <linux/skbuff_ref.h> 11 12 #include <net/page_pool/helpers.h> 13 #include <net/page_pool/memory_provider.h> 14 #include <net/netlink.h> 15 #include <net/netdev_rx_queue.h> 16 #include <net/tcp.h> 17 #include <net/rps.h> 18 19 #include <trace/events/page_pool.h> 20 21 #include <uapi/linux/io_uring.h> 22 23 #include "io_uring.h" 24 #include "kbuf.h" 25 #include "memmap.h" 26 #include "zcrx.h" 27 #include "rsrc.h" 28 29 #define IO_DMA_ATTR (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING) 30 31 static inline struct io_zcrx_ifq *io_pp_to_ifq(struct page_pool *pp) 32 { 33 return pp->mp_priv; 34 } 35 36 static inline struct io_zcrx_area *io_zcrx_iov_to_area(const struct net_iov *niov) 37 { 38 struct net_iov_area *owner = net_iov_owner(niov); 39 40 return container_of(owner, struct io_zcrx_area, nia); 41 } 42 43 static inline struct page *io_zcrx_iov_page(const struct net_iov *niov) 44 { 45 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 46 47 return area->mem.pages[net_iov_idx(niov)]; 48 } 49 50 static void io_release_dmabuf(struct io_zcrx_mem *mem) 51 { 52 if (!IS_ENABLED(CONFIG_DMA_SHARED_BUFFER)) 53 return; 54 55 if (mem->sgt) 56 dma_buf_unmap_attachment_unlocked(mem->attach, mem->sgt, 57 DMA_FROM_DEVICE); 58 if (mem->attach) 59 dma_buf_detach(mem->dmabuf, mem->attach); 60 if (mem->dmabuf) 61 dma_buf_put(mem->dmabuf); 62 63 mem->sgt = NULL; 64 mem->attach = NULL; 65 mem->dmabuf = NULL; 66 } 67 68 static int io_import_dmabuf(struct io_zcrx_ifq *ifq, 69 struct io_zcrx_mem *mem, 70 struct io_uring_zcrx_area_reg *area_reg) 71 { 72 unsigned long off = (unsigned long)area_reg->addr; 73 unsigned long len = (unsigned long)area_reg->len; 74 unsigned long total_size = 0; 75 struct scatterlist *sg; 76 int dmabuf_fd = area_reg->dmabuf_fd; 77 int i, ret; 78 79 if (WARN_ON_ONCE(!ifq->dev)) 80 return -EFAULT; 81 if (!IS_ENABLED(CONFIG_DMA_SHARED_BUFFER)) 82 return -EINVAL; 83 84 mem->is_dmabuf = true; 85 mem->dmabuf = dma_buf_get(dmabuf_fd); 86 if (IS_ERR(mem->dmabuf)) { 87 ret = PTR_ERR(mem->dmabuf); 88 mem->dmabuf = NULL; 89 goto err; 90 } 91 92 mem->attach = dma_buf_attach(mem->dmabuf, ifq->dev); 93 if (IS_ERR(mem->attach)) { 94 ret = PTR_ERR(mem->attach); 95 mem->attach = NULL; 96 goto err; 97 } 98 99 mem->sgt = dma_buf_map_attachment_unlocked(mem->attach, DMA_FROM_DEVICE); 100 if (IS_ERR(mem->sgt)) { 101 ret = PTR_ERR(mem->sgt); 102 mem->sgt = NULL; 103 goto err; 104 } 105 106 for_each_sgtable_dma_sg(mem->sgt, sg, i) 107 total_size += sg_dma_len(sg); 108 109 if (total_size < off + len) 110 return -EINVAL; 111 112 mem->dmabuf_offset = off; 113 mem->size = len; 114 return 0; 115 err: 116 io_release_dmabuf(mem); 117 return ret; 118 } 119 120 static int io_zcrx_map_area_dmabuf(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 121 { 122 unsigned long off = area->mem.dmabuf_offset; 123 struct scatterlist *sg; 124 unsigned i, niov_idx = 0; 125 126 if (!IS_ENABLED(CONFIG_DMA_SHARED_BUFFER)) 127 return -EINVAL; 128 129 for_each_sgtable_dma_sg(area->mem.sgt, sg, i) { 130 dma_addr_t dma = sg_dma_address(sg); 131 unsigned long sg_len = sg_dma_len(sg); 132 unsigned long sg_off = min(sg_len, off); 133 134 off -= sg_off; 135 sg_len -= sg_off; 136 dma += sg_off; 137 138 while (sg_len && niov_idx < area->nia.num_niovs) { 139 struct net_iov *niov = &area->nia.niovs[niov_idx]; 140 141 if (net_mp_niov_set_dma_addr(niov, dma)) 142 return 0; 143 sg_len -= PAGE_SIZE; 144 dma += PAGE_SIZE; 145 niov_idx++; 146 } 147 } 148 return niov_idx; 149 } 150 151 static int io_import_umem(struct io_zcrx_ifq *ifq, 152 struct io_zcrx_mem *mem, 153 struct io_uring_zcrx_area_reg *area_reg) 154 { 155 struct page **pages; 156 int nr_pages; 157 158 if (area_reg->dmabuf_fd) 159 return -EINVAL; 160 if (!area_reg->addr) 161 return -EFAULT; 162 pages = io_pin_pages((unsigned long)area_reg->addr, area_reg->len, 163 &nr_pages); 164 if (IS_ERR(pages)) 165 return PTR_ERR(pages); 166 167 mem->pages = pages; 168 mem->nr_folios = nr_pages; 169 mem->size = area_reg->len; 170 return 0; 171 } 172 173 static void io_release_area_mem(struct io_zcrx_mem *mem) 174 { 175 if (mem->is_dmabuf) { 176 io_release_dmabuf(mem); 177 return; 178 } 179 if (mem->pages) { 180 unpin_user_pages(mem->pages, mem->nr_folios); 181 kvfree(mem->pages); 182 } 183 } 184 185 static int io_import_area(struct io_zcrx_ifq *ifq, 186 struct io_zcrx_mem *mem, 187 struct io_uring_zcrx_area_reg *area_reg) 188 { 189 int ret; 190 191 ret = io_validate_user_buf_range(area_reg->addr, area_reg->len); 192 if (ret) 193 return ret; 194 if (area_reg->addr & ~PAGE_MASK || area_reg->len & ~PAGE_MASK) 195 return -EINVAL; 196 197 if (area_reg->flags & IORING_ZCRX_AREA_DMABUF) 198 return io_import_dmabuf(ifq, mem, area_reg); 199 return io_import_umem(ifq, mem, area_reg); 200 } 201 202 static void io_zcrx_unmap_umem(struct io_zcrx_ifq *ifq, 203 struct io_zcrx_area *area, int nr_mapped) 204 { 205 int i; 206 207 for (i = 0; i < nr_mapped; i++) { 208 netmem_ref netmem = net_iov_to_netmem(&area->nia.niovs[i]); 209 dma_addr_t dma = page_pool_get_dma_addr_netmem(netmem); 210 211 dma_unmap_page_attrs(ifq->dev, dma, PAGE_SIZE, 212 DMA_FROM_DEVICE, IO_DMA_ATTR); 213 } 214 } 215 216 static void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, 217 struct io_zcrx_area *area, int nr_mapped) 218 { 219 int i; 220 221 if (area->mem.is_dmabuf) 222 io_release_dmabuf(&area->mem); 223 else 224 io_zcrx_unmap_umem(ifq, area, nr_mapped); 225 226 for (i = 0; i < area->nia.num_niovs; i++) 227 net_mp_niov_set_dma_addr(&area->nia.niovs[i], 0); 228 } 229 230 static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 231 { 232 guard(mutex)(&ifq->dma_lock); 233 234 if (area->is_mapped) 235 __io_zcrx_unmap_area(ifq, area, area->nia.num_niovs); 236 area->is_mapped = false; 237 } 238 239 static int io_zcrx_map_area_umem(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 240 { 241 int i; 242 243 for (i = 0; i < area->nia.num_niovs; i++) { 244 struct net_iov *niov = &area->nia.niovs[i]; 245 dma_addr_t dma; 246 247 dma = dma_map_page_attrs(ifq->dev, area->mem.pages[i], 0, 248 PAGE_SIZE, DMA_FROM_DEVICE, IO_DMA_ATTR); 249 if (dma_mapping_error(ifq->dev, dma)) 250 break; 251 if (net_mp_niov_set_dma_addr(niov, dma)) { 252 dma_unmap_page_attrs(ifq->dev, dma, PAGE_SIZE, 253 DMA_FROM_DEVICE, IO_DMA_ATTR); 254 break; 255 } 256 } 257 return i; 258 } 259 260 static int io_zcrx_map_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 261 { 262 unsigned nr; 263 264 guard(mutex)(&ifq->dma_lock); 265 if (area->is_mapped) 266 return 0; 267 268 if (area->mem.is_dmabuf) 269 nr = io_zcrx_map_area_dmabuf(ifq, area); 270 else 271 nr = io_zcrx_map_area_umem(ifq, area); 272 273 if (nr != area->nia.num_niovs) { 274 __io_zcrx_unmap_area(ifq, area, nr); 275 return -EINVAL; 276 } 277 278 area->is_mapped = true; 279 return 0; 280 } 281 282 static void io_zcrx_sync_for_device(const struct page_pool *pool, 283 struct net_iov *niov) 284 { 285 #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC) 286 dma_addr_t dma_addr; 287 288 if (!dma_dev_need_sync(pool->p.dev)) 289 return; 290 291 dma_addr = page_pool_get_dma_addr_netmem(net_iov_to_netmem(niov)); 292 __dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset, 293 PAGE_SIZE, pool->p.dma_dir); 294 #endif 295 } 296 297 #define IO_RQ_MAX_ENTRIES 32768 298 299 #define IO_SKBS_PER_CALL_LIMIT 20 300 301 struct io_zcrx_args { 302 struct io_kiocb *req; 303 struct io_zcrx_ifq *ifq; 304 struct socket *sock; 305 unsigned nr_skbs; 306 }; 307 308 static const struct memory_provider_ops io_uring_pp_zc_ops; 309 310 static inline atomic_t *io_get_user_counter(struct net_iov *niov) 311 { 312 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 313 314 return &area->user_refs[net_iov_idx(niov)]; 315 } 316 317 static bool io_zcrx_put_niov_uref(struct net_iov *niov) 318 { 319 atomic_t *uref = io_get_user_counter(niov); 320 321 if (unlikely(!atomic_read(uref))) 322 return false; 323 atomic_dec(uref); 324 return true; 325 } 326 327 static void io_zcrx_get_niov_uref(struct net_iov *niov) 328 { 329 atomic_inc(io_get_user_counter(niov)); 330 } 331 332 static int io_allocate_rbuf_ring(struct io_zcrx_ifq *ifq, 333 struct io_uring_zcrx_ifq_reg *reg, 334 struct io_uring_region_desc *rd, 335 u32 id) 336 { 337 u64 mmap_offset; 338 size_t off, size; 339 void *ptr; 340 int ret; 341 342 off = sizeof(struct io_uring); 343 size = off + sizeof(struct io_uring_zcrx_rqe) * reg->rq_entries; 344 if (size > rd->size) 345 return -EINVAL; 346 347 mmap_offset = IORING_MAP_OFF_ZCRX_REGION; 348 mmap_offset += id << IORING_OFF_PBUF_SHIFT; 349 350 ret = io_create_region(ifq->ctx, &ifq->region, rd, mmap_offset); 351 if (ret < 0) 352 return ret; 353 354 ptr = io_region_get_ptr(&ifq->region); 355 ifq->rq_ring = (struct io_uring *)ptr; 356 ifq->rqes = (struct io_uring_zcrx_rqe *)(ptr + off); 357 return 0; 358 } 359 360 static void io_free_rbuf_ring(struct io_zcrx_ifq *ifq) 361 { 362 io_free_region(ifq->ctx, &ifq->region); 363 ifq->rq_ring = NULL; 364 ifq->rqes = NULL; 365 } 366 367 static void io_zcrx_free_area(struct io_zcrx_area *area) 368 { 369 io_zcrx_unmap_area(area->ifq, area); 370 io_release_area_mem(&area->mem); 371 372 kvfree(area->freelist); 373 kvfree(area->nia.niovs); 374 kvfree(area->user_refs); 375 kfree(area); 376 } 377 378 #define IO_ZCRX_AREA_SUPPORTED_FLAGS (IORING_ZCRX_AREA_DMABUF) 379 380 static int io_zcrx_create_area(struct io_zcrx_ifq *ifq, 381 struct io_zcrx_area **res, 382 struct io_uring_zcrx_area_reg *area_reg) 383 { 384 struct io_zcrx_area *area; 385 unsigned nr_iovs; 386 int i, ret; 387 388 if (area_reg->flags & ~IO_ZCRX_AREA_SUPPORTED_FLAGS) 389 return -EINVAL; 390 if (area_reg->rq_area_token) 391 return -EINVAL; 392 if (area_reg->__resv2[0] || area_reg->__resv2[1]) 393 return -EINVAL; 394 395 ret = -ENOMEM; 396 area = kzalloc(sizeof(*area), GFP_KERNEL); 397 if (!area) 398 goto err; 399 400 ret = io_import_area(ifq, &area->mem, area_reg); 401 if (ret) 402 goto err; 403 404 nr_iovs = area->mem.size >> PAGE_SHIFT; 405 area->nia.num_niovs = nr_iovs; 406 407 ret = -ENOMEM; 408 area->nia.niovs = kvmalloc_array(nr_iovs, sizeof(area->nia.niovs[0]), 409 GFP_KERNEL | __GFP_ZERO); 410 if (!area->nia.niovs) 411 goto err; 412 413 area->freelist = kvmalloc_array(nr_iovs, sizeof(area->freelist[0]), 414 GFP_KERNEL | __GFP_ZERO); 415 if (!area->freelist) 416 goto err; 417 418 area->user_refs = kvmalloc_array(nr_iovs, sizeof(area->user_refs[0]), 419 GFP_KERNEL | __GFP_ZERO); 420 if (!area->user_refs) 421 goto err; 422 423 for (i = 0; i < nr_iovs; i++) { 424 struct net_iov *niov = &area->nia.niovs[i]; 425 426 niov->owner = &area->nia; 427 area->freelist[i] = i; 428 atomic_set(&area->user_refs[i], 0); 429 } 430 431 area->free_count = nr_iovs; 432 area->ifq = ifq; 433 /* we're only supporting one area per ifq for now */ 434 area->area_id = 0; 435 area_reg->rq_area_token = (u64)area->area_id << IORING_ZCRX_AREA_SHIFT; 436 spin_lock_init(&area->freelist_lock); 437 *res = area; 438 return 0; 439 err: 440 if (area) 441 io_zcrx_free_area(area); 442 return ret; 443 } 444 445 static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx) 446 { 447 struct io_zcrx_ifq *ifq; 448 449 ifq = kzalloc(sizeof(*ifq), GFP_KERNEL); 450 if (!ifq) 451 return NULL; 452 453 ifq->if_rxq = -1; 454 ifq->ctx = ctx; 455 spin_lock_init(&ifq->lock); 456 spin_lock_init(&ifq->rq_lock); 457 mutex_init(&ifq->dma_lock); 458 return ifq; 459 } 460 461 static void io_zcrx_drop_netdev(struct io_zcrx_ifq *ifq) 462 { 463 spin_lock(&ifq->lock); 464 if (ifq->netdev) { 465 netdev_put(ifq->netdev, &ifq->netdev_tracker); 466 ifq->netdev = NULL; 467 } 468 spin_unlock(&ifq->lock); 469 } 470 471 static void io_close_queue(struct io_zcrx_ifq *ifq) 472 { 473 struct net_device *netdev; 474 netdevice_tracker netdev_tracker; 475 struct pp_memory_provider_params p = { 476 .mp_ops = &io_uring_pp_zc_ops, 477 .mp_priv = ifq, 478 }; 479 480 if (ifq->if_rxq == -1) 481 return; 482 483 spin_lock(&ifq->lock); 484 netdev = ifq->netdev; 485 netdev_tracker = ifq->netdev_tracker; 486 ifq->netdev = NULL; 487 spin_unlock(&ifq->lock); 488 489 if (netdev) { 490 net_mp_close_rxq(netdev, ifq->if_rxq, &p); 491 netdev_put(netdev, &netdev_tracker); 492 } 493 ifq->if_rxq = -1; 494 } 495 496 static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq) 497 { 498 io_close_queue(ifq); 499 io_zcrx_drop_netdev(ifq); 500 501 if (ifq->area) 502 io_zcrx_free_area(ifq->area); 503 if (ifq->dev) 504 put_device(ifq->dev); 505 506 io_free_rbuf_ring(ifq); 507 mutex_destroy(&ifq->dma_lock); 508 kfree(ifq); 509 } 510 511 struct io_mapped_region *io_zcrx_get_region(struct io_ring_ctx *ctx, 512 unsigned int id) 513 { 514 struct io_zcrx_ifq *ifq = xa_load(&ctx->zcrx_ctxs, id); 515 516 lockdep_assert_held(&ctx->mmap_lock); 517 518 return ifq ? &ifq->region : NULL; 519 } 520 521 int io_register_zcrx_ifq(struct io_ring_ctx *ctx, 522 struct io_uring_zcrx_ifq_reg __user *arg) 523 { 524 struct pp_memory_provider_params mp_param = {}; 525 struct io_uring_zcrx_area_reg area; 526 struct io_uring_zcrx_ifq_reg reg; 527 struct io_uring_region_desc rd; 528 struct io_zcrx_ifq *ifq; 529 int ret; 530 u32 id; 531 532 /* 533 * 1. Interface queue allocation. 534 * 2. It can observe data destined for sockets of other tasks. 535 */ 536 if (!capable(CAP_NET_ADMIN)) 537 return -EPERM; 538 539 /* mandatory io_uring features for zc rx */ 540 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN && 541 ctx->flags & IORING_SETUP_CQE32)) 542 return -EINVAL; 543 if (copy_from_user(®, arg, sizeof(reg))) 544 return -EFAULT; 545 if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd))) 546 return -EFAULT; 547 if (memchr_inv(®.__resv, 0, sizeof(reg.__resv)) || 548 reg.__resv2 || reg.zcrx_id) 549 return -EINVAL; 550 if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags) 551 return -EINVAL; 552 if (reg.rq_entries > IO_RQ_MAX_ENTRIES) { 553 if (!(ctx->flags & IORING_SETUP_CLAMP)) 554 return -EINVAL; 555 reg.rq_entries = IO_RQ_MAX_ENTRIES; 556 } 557 reg.rq_entries = roundup_pow_of_two(reg.rq_entries); 558 559 if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area))) 560 return -EFAULT; 561 562 ifq = io_zcrx_ifq_alloc(ctx); 563 if (!ifq) 564 return -ENOMEM; 565 ifq->rq_entries = reg.rq_entries; 566 567 scoped_guard(mutex, &ctx->mmap_lock) { 568 /* preallocate id */ 569 ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL); 570 if (ret) 571 goto ifq_free; 572 } 573 574 ret = io_allocate_rbuf_ring(ifq, ®, &rd, id); 575 if (ret) 576 goto err; 577 578 ifq->netdev = netdev_get_by_index(current->nsproxy->net_ns, reg.if_idx, 579 &ifq->netdev_tracker, GFP_KERNEL); 580 if (!ifq->netdev) { 581 ret = -ENODEV; 582 goto err; 583 } 584 585 ifq->dev = ifq->netdev->dev.parent; 586 if (!ifq->dev) { 587 ret = -EOPNOTSUPP; 588 goto err; 589 } 590 get_device(ifq->dev); 591 592 ret = io_zcrx_create_area(ifq, &ifq->area, &area); 593 if (ret) 594 goto err; 595 596 mp_param.mp_ops = &io_uring_pp_zc_ops; 597 mp_param.mp_priv = ifq; 598 ret = net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param); 599 if (ret) 600 goto err; 601 ifq->if_rxq = reg.if_rxq; 602 603 reg.offsets.rqes = sizeof(struct io_uring); 604 reg.offsets.head = offsetof(struct io_uring, head); 605 reg.offsets.tail = offsetof(struct io_uring, tail); 606 reg.zcrx_id = id; 607 608 scoped_guard(mutex, &ctx->mmap_lock) { 609 /* publish ifq */ 610 ret = -ENOMEM; 611 if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL)) 612 goto err; 613 } 614 615 if (copy_to_user(arg, ®, sizeof(reg)) || 616 copy_to_user(u64_to_user_ptr(reg.region_ptr), &rd, sizeof(rd)) || 617 copy_to_user(u64_to_user_ptr(reg.area_ptr), &area, sizeof(area))) { 618 ret = -EFAULT; 619 goto err; 620 } 621 return 0; 622 err: 623 scoped_guard(mutex, &ctx->mmap_lock) 624 xa_erase(&ctx->zcrx_ctxs, id); 625 ifq_free: 626 io_zcrx_ifq_free(ifq); 627 return ret; 628 } 629 630 void io_unregister_zcrx_ifqs(struct io_ring_ctx *ctx) 631 { 632 struct io_zcrx_ifq *ifq; 633 unsigned long id; 634 635 lockdep_assert_held(&ctx->uring_lock); 636 637 while (1) { 638 scoped_guard(mutex, &ctx->mmap_lock) { 639 ifq = xa_find(&ctx->zcrx_ctxs, &id, ULONG_MAX, XA_PRESENT); 640 if (ifq) 641 xa_erase(&ctx->zcrx_ctxs, id); 642 } 643 if (!ifq) 644 break; 645 io_zcrx_ifq_free(ifq); 646 } 647 648 xa_destroy(&ctx->zcrx_ctxs); 649 } 650 651 static struct net_iov *__io_zcrx_get_free_niov(struct io_zcrx_area *area) 652 { 653 unsigned niov_idx; 654 655 lockdep_assert_held(&area->freelist_lock); 656 657 niov_idx = area->freelist[--area->free_count]; 658 return &area->nia.niovs[niov_idx]; 659 } 660 661 static void io_zcrx_return_niov_freelist(struct net_iov *niov) 662 { 663 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 664 665 spin_lock_bh(&area->freelist_lock); 666 area->freelist[area->free_count++] = net_iov_idx(niov); 667 spin_unlock_bh(&area->freelist_lock); 668 } 669 670 static void io_zcrx_return_niov(struct net_iov *niov) 671 { 672 netmem_ref netmem = net_iov_to_netmem(niov); 673 674 if (!niov->pp) { 675 /* copy fallback allocated niovs */ 676 io_zcrx_return_niov_freelist(niov); 677 return; 678 } 679 page_pool_put_unrefed_netmem(niov->pp, netmem, -1, false); 680 } 681 682 static void io_zcrx_scrub(struct io_zcrx_ifq *ifq) 683 { 684 struct io_zcrx_area *area = ifq->area; 685 int i; 686 687 if (!area) 688 return; 689 690 /* Reclaim back all buffers given to the user space. */ 691 for (i = 0; i < area->nia.num_niovs; i++) { 692 struct net_iov *niov = &area->nia.niovs[i]; 693 int nr; 694 695 if (!atomic_read(io_get_user_counter(niov))) 696 continue; 697 nr = atomic_xchg(io_get_user_counter(niov), 0); 698 if (nr && !page_pool_unref_netmem(net_iov_to_netmem(niov), nr)) 699 io_zcrx_return_niov(niov); 700 } 701 } 702 703 void io_shutdown_zcrx_ifqs(struct io_ring_ctx *ctx) 704 { 705 struct io_zcrx_ifq *ifq; 706 unsigned long index; 707 708 lockdep_assert_held(&ctx->uring_lock); 709 710 xa_for_each(&ctx->zcrx_ctxs, index, ifq) { 711 io_zcrx_scrub(ifq); 712 io_close_queue(ifq); 713 } 714 } 715 716 static inline u32 io_zcrx_rqring_entries(struct io_zcrx_ifq *ifq) 717 { 718 u32 entries; 719 720 entries = smp_load_acquire(&ifq->rq_ring->tail) - ifq->cached_rq_head; 721 return min(entries, ifq->rq_entries); 722 } 723 724 static struct io_uring_zcrx_rqe *io_zcrx_get_rqe(struct io_zcrx_ifq *ifq, 725 unsigned mask) 726 { 727 unsigned int idx = ifq->cached_rq_head++ & mask; 728 729 return &ifq->rqes[idx]; 730 } 731 732 static void io_zcrx_ring_refill(struct page_pool *pp, 733 struct io_zcrx_ifq *ifq) 734 { 735 unsigned int mask = ifq->rq_entries - 1; 736 unsigned int entries; 737 netmem_ref netmem; 738 739 spin_lock_bh(&ifq->rq_lock); 740 741 entries = io_zcrx_rqring_entries(ifq); 742 entries = min_t(unsigned, entries, PP_ALLOC_CACHE_REFILL - pp->alloc.count); 743 if (unlikely(!entries)) { 744 spin_unlock_bh(&ifq->rq_lock); 745 return; 746 } 747 748 do { 749 struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(ifq, mask); 750 struct io_zcrx_area *area; 751 struct net_iov *niov; 752 unsigned niov_idx, area_idx; 753 754 area_idx = rqe->off >> IORING_ZCRX_AREA_SHIFT; 755 niov_idx = (rqe->off & ~IORING_ZCRX_AREA_MASK) >> PAGE_SHIFT; 756 757 if (unlikely(rqe->__pad || area_idx)) 758 continue; 759 area = ifq->area; 760 761 if (unlikely(niov_idx >= area->nia.num_niovs)) 762 continue; 763 niov_idx = array_index_nospec(niov_idx, area->nia.num_niovs); 764 765 niov = &area->nia.niovs[niov_idx]; 766 if (!io_zcrx_put_niov_uref(niov)) 767 continue; 768 769 netmem = net_iov_to_netmem(niov); 770 if (page_pool_unref_netmem(netmem, 1) != 0) 771 continue; 772 773 if (unlikely(niov->pp != pp)) { 774 io_zcrx_return_niov(niov); 775 continue; 776 } 777 778 io_zcrx_sync_for_device(pp, niov); 779 net_mp_netmem_place_in_cache(pp, netmem); 780 } while (--entries); 781 782 smp_store_release(&ifq->rq_ring->head, ifq->cached_rq_head); 783 spin_unlock_bh(&ifq->rq_lock); 784 } 785 786 static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq) 787 { 788 struct io_zcrx_area *area = ifq->area; 789 790 spin_lock_bh(&area->freelist_lock); 791 while (area->free_count && pp->alloc.count < PP_ALLOC_CACHE_REFILL) { 792 struct net_iov *niov = __io_zcrx_get_free_niov(area); 793 netmem_ref netmem = net_iov_to_netmem(niov); 794 795 net_mp_niov_set_page_pool(pp, niov); 796 io_zcrx_sync_for_device(pp, niov); 797 net_mp_netmem_place_in_cache(pp, netmem); 798 } 799 spin_unlock_bh(&area->freelist_lock); 800 } 801 802 static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp) 803 { 804 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 805 806 /* pp should already be ensuring that */ 807 if (unlikely(pp->alloc.count)) 808 goto out_return; 809 810 io_zcrx_ring_refill(pp, ifq); 811 if (likely(pp->alloc.count)) 812 goto out_return; 813 814 io_zcrx_refill_slow(pp, ifq); 815 if (!pp->alloc.count) 816 return 0; 817 out_return: 818 return pp->alloc.cache[--pp->alloc.count]; 819 } 820 821 static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem) 822 { 823 struct net_iov *niov; 824 825 if (WARN_ON_ONCE(!netmem_is_net_iov(netmem))) 826 return false; 827 828 niov = netmem_to_net_iov(netmem); 829 net_mp_niov_clear_page_pool(niov); 830 io_zcrx_return_niov_freelist(niov); 831 return false; 832 } 833 834 static int io_pp_zc_init(struct page_pool *pp) 835 { 836 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 837 int ret; 838 839 if (WARN_ON_ONCE(!ifq)) 840 return -EINVAL; 841 if (WARN_ON_ONCE(ifq->dev != pp->p.dev)) 842 return -EINVAL; 843 if (WARN_ON_ONCE(!pp->dma_map)) 844 return -EOPNOTSUPP; 845 if (pp->p.order != 0) 846 return -EOPNOTSUPP; 847 if (pp->p.dma_dir != DMA_FROM_DEVICE) 848 return -EOPNOTSUPP; 849 850 ret = io_zcrx_map_area(ifq, ifq->area); 851 if (ret) 852 return ret; 853 854 percpu_ref_get(&ifq->ctx->refs); 855 return 0; 856 } 857 858 static void io_pp_zc_destroy(struct page_pool *pp) 859 { 860 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 861 struct io_zcrx_area *area = ifq->area; 862 863 if (WARN_ON_ONCE(area->free_count != area->nia.num_niovs)) 864 return; 865 percpu_ref_put(&ifq->ctx->refs); 866 } 867 868 static int io_pp_nl_fill(void *mp_priv, struct sk_buff *rsp, 869 struct netdev_rx_queue *rxq) 870 { 871 struct nlattr *nest; 872 int type; 873 874 type = rxq ? NETDEV_A_QUEUE_IO_URING : NETDEV_A_PAGE_POOL_IO_URING; 875 nest = nla_nest_start(rsp, type); 876 if (!nest) 877 return -EMSGSIZE; 878 nla_nest_end(rsp, nest); 879 880 return 0; 881 } 882 883 static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq) 884 { 885 struct pp_memory_provider_params *p = &rxq->mp_params; 886 struct io_zcrx_ifq *ifq = mp_priv; 887 888 io_zcrx_drop_netdev(ifq); 889 if (ifq->area) 890 io_zcrx_unmap_area(ifq, ifq->area); 891 892 p->mp_ops = NULL; 893 p->mp_priv = NULL; 894 } 895 896 static const struct memory_provider_ops io_uring_pp_zc_ops = { 897 .alloc_netmems = io_pp_zc_alloc_netmems, 898 .release_netmem = io_pp_zc_release_netmem, 899 .init = io_pp_zc_init, 900 .destroy = io_pp_zc_destroy, 901 .nl_fill = io_pp_nl_fill, 902 .uninstall = io_pp_uninstall, 903 }; 904 905 static bool io_zcrx_queue_cqe(struct io_kiocb *req, struct net_iov *niov, 906 struct io_zcrx_ifq *ifq, int off, int len) 907 { 908 struct io_uring_zcrx_cqe *rcqe; 909 struct io_zcrx_area *area; 910 struct io_uring_cqe *cqe; 911 u64 offset; 912 913 if (!io_defer_get_uncommited_cqe(req->ctx, &cqe)) 914 return false; 915 916 cqe->user_data = req->cqe.user_data; 917 cqe->res = len; 918 cqe->flags = IORING_CQE_F_MORE; 919 920 area = io_zcrx_iov_to_area(niov); 921 offset = off + (net_iov_idx(niov) << PAGE_SHIFT); 922 rcqe = (struct io_uring_zcrx_cqe *)(cqe + 1); 923 rcqe->off = offset + ((u64)area->area_id << IORING_ZCRX_AREA_SHIFT); 924 rcqe->__pad = 0; 925 return true; 926 } 927 928 static struct net_iov *io_zcrx_alloc_fallback(struct io_zcrx_area *area) 929 { 930 struct net_iov *niov = NULL; 931 932 spin_lock_bh(&area->freelist_lock); 933 if (area->free_count) 934 niov = __io_zcrx_get_free_niov(area); 935 spin_unlock_bh(&area->freelist_lock); 936 937 if (niov) 938 page_pool_fragment_netmem(net_iov_to_netmem(niov), 1); 939 return niov; 940 } 941 942 static ssize_t io_zcrx_copy_chunk(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 943 void *src_base, struct page *src_page, 944 unsigned int src_offset, size_t len) 945 { 946 struct io_zcrx_area *area = ifq->area; 947 size_t copied = 0; 948 int ret = 0; 949 950 if (area->mem.is_dmabuf) 951 return -EFAULT; 952 953 while (len) { 954 size_t copy_size = min_t(size_t, PAGE_SIZE, len); 955 const int dst_off = 0; 956 struct net_iov *niov; 957 struct page *dst_page; 958 void *dst_addr; 959 960 niov = io_zcrx_alloc_fallback(area); 961 if (!niov) { 962 ret = -ENOMEM; 963 break; 964 } 965 966 dst_page = io_zcrx_iov_page(niov); 967 dst_addr = kmap_local_page(dst_page); 968 if (src_page) 969 src_base = kmap_local_page(src_page); 970 971 memcpy(dst_addr, src_base + src_offset, copy_size); 972 973 if (src_page) 974 kunmap_local(src_base); 975 kunmap_local(dst_addr); 976 977 if (!io_zcrx_queue_cqe(req, niov, ifq, dst_off, copy_size)) { 978 io_zcrx_return_niov(niov); 979 ret = -ENOSPC; 980 break; 981 } 982 983 io_zcrx_get_niov_uref(niov); 984 src_offset += copy_size; 985 len -= copy_size; 986 copied += copy_size; 987 } 988 989 return copied ? copied : ret; 990 } 991 992 static int io_zcrx_copy_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 993 const skb_frag_t *frag, int off, int len) 994 { 995 struct page *page = skb_frag_page(frag); 996 u32 p_off, p_len, t, copied = 0; 997 int ret = 0; 998 999 off += skb_frag_off(frag); 1000 1001 skb_frag_foreach_page(frag, off, len, 1002 page, p_off, p_len, t) { 1003 ret = io_zcrx_copy_chunk(req, ifq, NULL, page, p_off, p_len); 1004 if (ret < 0) 1005 return copied ? copied : ret; 1006 copied += ret; 1007 } 1008 return copied; 1009 } 1010 1011 static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1012 const skb_frag_t *frag, int off, int len) 1013 { 1014 struct net_iov *niov; 1015 1016 if (unlikely(!skb_frag_is_net_iov(frag))) 1017 return io_zcrx_copy_frag(req, ifq, frag, off, len); 1018 1019 niov = netmem_to_net_iov(frag->netmem); 1020 if (niov->pp->mp_ops != &io_uring_pp_zc_ops || 1021 io_pp_to_ifq(niov->pp) != ifq) 1022 return -EFAULT; 1023 1024 if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len)) 1025 return -ENOSPC; 1026 1027 /* 1028 * Prevent it from being recycled while user is accessing it. 1029 * It has to be done before grabbing a user reference. 1030 */ 1031 page_pool_ref_netmem(net_iov_to_netmem(niov)); 1032 io_zcrx_get_niov_uref(niov); 1033 return len; 1034 } 1035 1036 static int 1037 io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb, 1038 unsigned int offset, size_t len) 1039 { 1040 struct io_zcrx_args *args = desc->arg.data; 1041 struct io_zcrx_ifq *ifq = args->ifq; 1042 struct io_kiocb *req = args->req; 1043 struct sk_buff *frag_iter; 1044 unsigned start, start_off = offset; 1045 int i, copy, end, off; 1046 int ret = 0; 1047 1048 len = min_t(size_t, len, desc->count); 1049 /* 1050 * __tcp_read_sock() always calls io_zcrx_recv_skb one last time, even 1051 * if desc->count is already 0. This is caused by the if (offset + 1 != 1052 * skb->len) check. Return early in this case to break out of 1053 * __tcp_read_sock(). 1054 */ 1055 if (!len) 1056 return 0; 1057 if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT)) 1058 return -EAGAIN; 1059 1060 if (unlikely(offset < skb_headlen(skb))) { 1061 ssize_t copied; 1062 size_t to_copy; 1063 1064 to_copy = min_t(size_t, skb_headlen(skb) - offset, len); 1065 copied = io_zcrx_copy_chunk(req, ifq, skb->data, NULL, 1066 offset, to_copy); 1067 if (copied < 0) { 1068 ret = copied; 1069 goto out; 1070 } 1071 offset += copied; 1072 len -= copied; 1073 if (!len) 1074 goto out; 1075 if (offset != skb_headlen(skb)) 1076 goto out; 1077 } 1078 1079 start = skb_headlen(skb); 1080 1081 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 1082 const skb_frag_t *frag; 1083 1084 if (WARN_ON(start > offset + len)) 1085 return -EFAULT; 1086 1087 frag = &skb_shinfo(skb)->frags[i]; 1088 end = start + skb_frag_size(frag); 1089 1090 if (offset < end) { 1091 copy = end - offset; 1092 if (copy > len) 1093 copy = len; 1094 1095 off = offset - start; 1096 ret = io_zcrx_recv_frag(req, ifq, frag, off, copy); 1097 if (ret < 0) 1098 goto out; 1099 1100 offset += ret; 1101 len -= ret; 1102 if (len == 0 || ret != copy) 1103 goto out; 1104 } 1105 start = end; 1106 } 1107 1108 skb_walk_frags(skb, frag_iter) { 1109 if (WARN_ON(start > offset + len)) 1110 return -EFAULT; 1111 1112 end = start + frag_iter->len; 1113 if (offset < end) { 1114 copy = end - offset; 1115 if (copy > len) 1116 copy = len; 1117 1118 off = offset - start; 1119 ret = io_zcrx_recv_skb(desc, frag_iter, off, copy); 1120 if (ret < 0) 1121 goto out; 1122 1123 offset += ret; 1124 len -= ret; 1125 if (len == 0 || ret != copy) 1126 goto out; 1127 } 1128 start = end; 1129 } 1130 1131 out: 1132 if (offset == start_off) 1133 return ret; 1134 desc->count -= (offset - start_off); 1135 return offset - start_off; 1136 } 1137 1138 static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1139 struct sock *sk, int flags, 1140 unsigned issue_flags, unsigned int *outlen) 1141 { 1142 unsigned int len = *outlen; 1143 struct io_zcrx_args args = { 1144 .req = req, 1145 .ifq = ifq, 1146 .sock = sk->sk_socket, 1147 }; 1148 read_descriptor_t rd_desc = { 1149 .count = len ? len : UINT_MAX, 1150 .arg.data = &args, 1151 }; 1152 int ret; 1153 1154 lock_sock(sk); 1155 ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb); 1156 if (len && ret > 0) 1157 *outlen = len - ret; 1158 if (ret <= 0) { 1159 if (ret < 0 || sock_flag(sk, SOCK_DONE)) 1160 goto out; 1161 if (sk->sk_err) 1162 ret = sock_error(sk); 1163 else if (sk->sk_shutdown & RCV_SHUTDOWN) 1164 goto out; 1165 else if (sk->sk_state == TCP_CLOSE) 1166 ret = -ENOTCONN; 1167 else 1168 ret = -EAGAIN; 1169 } else if (unlikely(args.nr_skbs > IO_SKBS_PER_CALL_LIMIT) && 1170 (issue_flags & IO_URING_F_MULTISHOT)) { 1171 ret = IOU_REQUEUE; 1172 } else if (sock_flag(sk, SOCK_DONE)) { 1173 /* Make it to retry until it finally gets 0. */ 1174 if (issue_flags & IO_URING_F_MULTISHOT) 1175 ret = IOU_REQUEUE; 1176 else 1177 ret = -EAGAIN; 1178 } 1179 out: 1180 release_sock(sk); 1181 return ret; 1182 } 1183 1184 int io_zcrx_recv(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1185 struct socket *sock, unsigned int flags, 1186 unsigned issue_flags, unsigned int *len) 1187 { 1188 struct sock *sk = sock->sk; 1189 const struct proto *prot = READ_ONCE(sk->sk_prot); 1190 1191 if (prot->recvmsg != tcp_recvmsg) 1192 return -EPROTONOSUPPORT; 1193 1194 sock_rps_record_flow(sk); 1195 return io_zcrx_tcp_recvmsg(req, ifq, sk, flags, issue_flags, len); 1196 } 1197