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 void __io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, 32 struct io_zcrx_area *area, int nr_mapped) 33 { 34 int i; 35 36 for (i = 0; i < nr_mapped; i++) { 37 struct net_iov *niov = &area->nia.niovs[i]; 38 dma_addr_t dma; 39 40 dma = page_pool_get_dma_addr_netmem(net_iov_to_netmem(niov)); 41 dma_unmap_page_attrs(ifq->dev, dma, PAGE_SIZE, 42 DMA_FROM_DEVICE, IO_DMA_ATTR); 43 net_mp_niov_set_dma_addr(niov, 0); 44 } 45 } 46 47 static void io_zcrx_unmap_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 48 { 49 if (area->is_mapped) 50 __io_zcrx_unmap_area(ifq, area, area->nia.num_niovs); 51 } 52 53 static int io_zcrx_map_area(struct io_zcrx_ifq *ifq, struct io_zcrx_area *area) 54 { 55 int i; 56 57 for (i = 0; i < area->nia.num_niovs; i++) { 58 struct net_iov *niov = &area->nia.niovs[i]; 59 dma_addr_t dma; 60 61 dma = dma_map_page_attrs(ifq->dev, area->pages[i], 0, PAGE_SIZE, 62 DMA_FROM_DEVICE, IO_DMA_ATTR); 63 if (dma_mapping_error(ifq->dev, dma)) 64 break; 65 if (net_mp_niov_set_dma_addr(niov, dma)) { 66 dma_unmap_page_attrs(ifq->dev, dma, PAGE_SIZE, 67 DMA_FROM_DEVICE, IO_DMA_ATTR); 68 break; 69 } 70 } 71 72 if (i != area->nia.num_niovs) { 73 __io_zcrx_unmap_area(ifq, area, i); 74 return -EINVAL; 75 } 76 77 area->is_mapped = true; 78 return 0; 79 } 80 81 static void io_zcrx_sync_for_device(const struct page_pool *pool, 82 struct net_iov *niov) 83 { 84 #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC) 85 dma_addr_t dma_addr; 86 87 if (!dma_dev_need_sync(pool->p.dev)) 88 return; 89 90 dma_addr = page_pool_get_dma_addr_netmem(net_iov_to_netmem(niov)); 91 __dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset, 92 PAGE_SIZE, pool->p.dma_dir); 93 #endif 94 } 95 96 #define IO_RQ_MAX_ENTRIES 32768 97 98 #define IO_SKBS_PER_CALL_LIMIT 20 99 100 struct io_zcrx_args { 101 struct io_kiocb *req; 102 struct io_zcrx_ifq *ifq; 103 struct socket *sock; 104 unsigned nr_skbs; 105 }; 106 107 static const struct memory_provider_ops io_uring_pp_zc_ops; 108 109 static inline struct io_zcrx_area *io_zcrx_iov_to_area(const struct net_iov *niov) 110 { 111 struct net_iov_area *owner = net_iov_owner(niov); 112 113 return container_of(owner, struct io_zcrx_area, nia); 114 } 115 116 static inline atomic_t *io_get_user_counter(struct net_iov *niov) 117 { 118 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 119 120 return &area->user_refs[net_iov_idx(niov)]; 121 } 122 123 static bool io_zcrx_put_niov_uref(struct net_iov *niov) 124 { 125 atomic_t *uref = io_get_user_counter(niov); 126 127 if (unlikely(!atomic_read(uref))) 128 return false; 129 atomic_dec(uref); 130 return true; 131 } 132 133 static void io_zcrx_get_niov_uref(struct net_iov *niov) 134 { 135 atomic_inc(io_get_user_counter(niov)); 136 } 137 138 static inline struct page *io_zcrx_iov_page(const struct net_iov *niov) 139 { 140 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 141 142 return area->pages[net_iov_idx(niov)]; 143 } 144 145 static int io_allocate_rbuf_ring(struct io_zcrx_ifq *ifq, 146 struct io_uring_zcrx_ifq_reg *reg, 147 struct io_uring_region_desc *rd) 148 { 149 size_t off, size; 150 void *ptr; 151 int ret; 152 153 off = sizeof(struct io_uring); 154 size = off + sizeof(struct io_uring_zcrx_rqe) * reg->rq_entries; 155 if (size > rd->size) 156 return -EINVAL; 157 158 ret = io_create_region_mmap_safe(ifq->ctx, &ifq->ctx->zcrx_region, rd, 159 IORING_MAP_OFF_ZCRX_REGION); 160 if (ret < 0) 161 return ret; 162 163 ptr = io_region_get_ptr(&ifq->ctx->zcrx_region); 164 ifq->rq_ring = (struct io_uring *)ptr; 165 ifq->rqes = (struct io_uring_zcrx_rqe *)(ptr + off); 166 return 0; 167 } 168 169 static void io_free_rbuf_ring(struct io_zcrx_ifq *ifq) 170 { 171 io_free_region(ifq->ctx, &ifq->ctx->zcrx_region); 172 ifq->rq_ring = NULL; 173 ifq->rqes = NULL; 174 } 175 176 static void io_zcrx_free_area(struct io_zcrx_area *area) 177 { 178 io_zcrx_unmap_area(area->ifq, area); 179 180 kvfree(area->freelist); 181 kvfree(area->nia.niovs); 182 kvfree(area->user_refs); 183 if (area->pages) { 184 unpin_user_pages(area->pages, area->nia.num_niovs); 185 kvfree(area->pages); 186 } 187 kfree(area); 188 } 189 190 static int io_zcrx_create_area(struct io_zcrx_ifq *ifq, 191 struct io_zcrx_area **res, 192 struct io_uring_zcrx_area_reg *area_reg) 193 { 194 struct io_zcrx_area *area; 195 int i, ret, nr_pages; 196 struct iovec iov; 197 198 if (area_reg->flags || area_reg->rq_area_token) 199 return -EINVAL; 200 if (area_reg->__resv1 || area_reg->__resv2[0] || area_reg->__resv2[1]) 201 return -EINVAL; 202 if (area_reg->addr & ~PAGE_MASK || area_reg->len & ~PAGE_MASK) 203 return -EINVAL; 204 205 iov.iov_base = u64_to_user_ptr(area_reg->addr); 206 iov.iov_len = area_reg->len; 207 ret = io_buffer_validate(&iov); 208 if (ret) 209 return ret; 210 211 ret = -ENOMEM; 212 area = kzalloc(sizeof(*area), GFP_KERNEL); 213 if (!area) 214 goto err; 215 216 area->pages = io_pin_pages((unsigned long)area_reg->addr, area_reg->len, 217 &nr_pages); 218 if (IS_ERR(area->pages)) { 219 ret = PTR_ERR(area->pages); 220 area->pages = NULL; 221 goto err; 222 } 223 area->nia.num_niovs = nr_pages; 224 225 area->nia.niovs = kvmalloc_array(nr_pages, sizeof(area->nia.niovs[0]), 226 GFP_KERNEL | __GFP_ZERO); 227 if (!area->nia.niovs) 228 goto err; 229 230 area->freelist = kvmalloc_array(nr_pages, sizeof(area->freelist[0]), 231 GFP_KERNEL | __GFP_ZERO); 232 if (!area->freelist) 233 goto err; 234 235 for (i = 0; i < nr_pages; i++) 236 area->freelist[i] = i; 237 238 area->user_refs = kvmalloc_array(nr_pages, sizeof(area->user_refs[0]), 239 GFP_KERNEL | __GFP_ZERO); 240 if (!area->user_refs) 241 goto err; 242 243 for (i = 0; i < nr_pages; i++) { 244 struct net_iov *niov = &area->nia.niovs[i]; 245 246 niov->owner = &area->nia; 247 area->freelist[i] = i; 248 atomic_set(&area->user_refs[i], 0); 249 } 250 251 area->free_count = nr_pages; 252 area->ifq = ifq; 253 /* we're only supporting one area per ifq for now */ 254 area->area_id = 0; 255 area_reg->rq_area_token = (u64)area->area_id << IORING_ZCRX_AREA_SHIFT; 256 spin_lock_init(&area->freelist_lock); 257 *res = area; 258 return 0; 259 err: 260 if (area) 261 io_zcrx_free_area(area); 262 return ret; 263 } 264 265 static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx) 266 { 267 struct io_zcrx_ifq *ifq; 268 269 ifq = kzalloc(sizeof(*ifq), GFP_KERNEL); 270 if (!ifq) 271 return NULL; 272 273 ifq->if_rxq = -1; 274 ifq->ctx = ctx; 275 spin_lock_init(&ifq->lock); 276 spin_lock_init(&ifq->rq_lock); 277 return ifq; 278 } 279 280 static void io_zcrx_drop_netdev(struct io_zcrx_ifq *ifq) 281 { 282 spin_lock(&ifq->lock); 283 if (ifq->netdev) { 284 netdev_put(ifq->netdev, &ifq->netdev_tracker); 285 ifq->netdev = NULL; 286 } 287 spin_unlock(&ifq->lock); 288 } 289 290 static void io_close_queue(struct io_zcrx_ifq *ifq) 291 { 292 struct net_device *netdev; 293 netdevice_tracker netdev_tracker; 294 struct pp_memory_provider_params p = { 295 .mp_ops = &io_uring_pp_zc_ops, 296 .mp_priv = ifq, 297 }; 298 299 if (ifq->if_rxq == -1) 300 return; 301 302 spin_lock(&ifq->lock); 303 netdev = ifq->netdev; 304 netdev_tracker = ifq->netdev_tracker; 305 ifq->netdev = NULL; 306 spin_unlock(&ifq->lock); 307 308 if (netdev) { 309 net_mp_close_rxq(netdev, ifq->if_rxq, &p); 310 netdev_put(netdev, &netdev_tracker); 311 } 312 ifq->if_rxq = -1; 313 } 314 315 static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq) 316 { 317 io_close_queue(ifq); 318 io_zcrx_drop_netdev(ifq); 319 320 if (ifq->area) 321 io_zcrx_free_area(ifq->area); 322 if (ifq->dev) 323 put_device(ifq->dev); 324 325 io_free_rbuf_ring(ifq); 326 kfree(ifq); 327 } 328 329 int io_register_zcrx_ifq(struct io_ring_ctx *ctx, 330 struct io_uring_zcrx_ifq_reg __user *arg) 331 { 332 struct pp_memory_provider_params mp_param = {}; 333 struct io_uring_zcrx_area_reg area; 334 struct io_uring_zcrx_ifq_reg reg; 335 struct io_uring_region_desc rd; 336 struct io_zcrx_ifq *ifq; 337 int ret; 338 339 /* 340 * 1. Interface queue allocation. 341 * 2. It can observe data destined for sockets of other tasks. 342 */ 343 if (!capable(CAP_NET_ADMIN)) 344 return -EPERM; 345 346 /* mandatory io_uring features for zc rx */ 347 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN && 348 ctx->flags & IORING_SETUP_CQE32)) 349 return -EINVAL; 350 if (ctx->ifq) 351 return -EBUSY; 352 if (copy_from_user(®, arg, sizeof(reg))) 353 return -EFAULT; 354 if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd))) 355 return -EFAULT; 356 if (memchr_inv(®.__resv, 0, sizeof(reg.__resv))) 357 return -EINVAL; 358 if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags) 359 return -EINVAL; 360 if (reg.rq_entries > IO_RQ_MAX_ENTRIES) { 361 if (!(ctx->flags & IORING_SETUP_CLAMP)) 362 return -EINVAL; 363 reg.rq_entries = IO_RQ_MAX_ENTRIES; 364 } 365 reg.rq_entries = roundup_pow_of_two(reg.rq_entries); 366 367 if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area))) 368 return -EFAULT; 369 370 ifq = io_zcrx_ifq_alloc(ctx); 371 if (!ifq) 372 return -ENOMEM; 373 374 ret = io_allocate_rbuf_ring(ifq, ®, &rd); 375 if (ret) 376 goto err; 377 378 ret = io_zcrx_create_area(ifq, &ifq->area, &area); 379 if (ret) 380 goto err; 381 382 ifq->rq_entries = reg.rq_entries; 383 384 ret = -ENODEV; 385 ifq->netdev = netdev_get_by_index(current->nsproxy->net_ns, reg.if_idx, 386 &ifq->netdev_tracker, GFP_KERNEL); 387 if (!ifq->netdev) 388 goto err; 389 390 ifq->dev = ifq->netdev->dev.parent; 391 ret = -EOPNOTSUPP; 392 if (!ifq->dev) 393 goto err; 394 get_device(ifq->dev); 395 396 ret = io_zcrx_map_area(ifq, ifq->area); 397 if (ret) 398 goto err; 399 400 mp_param.mp_ops = &io_uring_pp_zc_ops; 401 mp_param.mp_priv = ifq; 402 ret = net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param); 403 if (ret) 404 goto err; 405 ifq->if_rxq = reg.if_rxq; 406 407 reg.offsets.rqes = sizeof(struct io_uring); 408 reg.offsets.head = offsetof(struct io_uring, head); 409 reg.offsets.tail = offsetof(struct io_uring, tail); 410 411 if (copy_to_user(arg, ®, sizeof(reg)) || 412 copy_to_user(u64_to_user_ptr(reg.region_ptr), &rd, sizeof(rd)) || 413 copy_to_user(u64_to_user_ptr(reg.area_ptr), &area, sizeof(area))) { 414 ret = -EFAULT; 415 goto err; 416 } 417 ctx->ifq = ifq; 418 return 0; 419 err: 420 io_zcrx_ifq_free(ifq); 421 return ret; 422 } 423 424 void io_unregister_zcrx_ifqs(struct io_ring_ctx *ctx) 425 { 426 struct io_zcrx_ifq *ifq = ctx->ifq; 427 428 lockdep_assert_held(&ctx->uring_lock); 429 430 if (!ifq) 431 return; 432 433 ctx->ifq = NULL; 434 io_zcrx_ifq_free(ifq); 435 } 436 437 static struct net_iov *__io_zcrx_get_free_niov(struct io_zcrx_area *area) 438 { 439 unsigned niov_idx; 440 441 lockdep_assert_held(&area->freelist_lock); 442 443 niov_idx = area->freelist[--area->free_count]; 444 return &area->nia.niovs[niov_idx]; 445 } 446 447 static void io_zcrx_return_niov_freelist(struct net_iov *niov) 448 { 449 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 450 451 spin_lock_bh(&area->freelist_lock); 452 area->freelist[area->free_count++] = net_iov_idx(niov); 453 spin_unlock_bh(&area->freelist_lock); 454 } 455 456 static void io_zcrx_return_niov(struct net_iov *niov) 457 { 458 netmem_ref netmem = net_iov_to_netmem(niov); 459 460 if (!niov->pp) { 461 /* copy fallback allocated niovs */ 462 io_zcrx_return_niov_freelist(niov); 463 return; 464 } 465 page_pool_put_unrefed_netmem(niov->pp, netmem, -1, false); 466 } 467 468 static void io_zcrx_scrub(struct io_zcrx_ifq *ifq) 469 { 470 struct io_zcrx_area *area = ifq->area; 471 int i; 472 473 if (!area) 474 return; 475 476 /* Reclaim back all buffers given to the user space. */ 477 for (i = 0; i < area->nia.num_niovs; i++) { 478 struct net_iov *niov = &area->nia.niovs[i]; 479 int nr; 480 481 if (!atomic_read(io_get_user_counter(niov))) 482 continue; 483 nr = atomic_xchg(io_get_user_counter(niov), 0); 484 if (nr && !page_pool_unref_netmem(net_iov_to_netmem(niov), nr)) 485 io_zcrx_return_niov(niov); 486 } 487 } 488 489 void io_shutdown_zcrx_ifqs(struct io_ring_ctx *ctx) 490 { 491 lockdep_assert_held(&ctx->uring_lock); 492 493 if (!ctx->ifq) 494 return; 495 io_zcrx_scrub(ctx->ifq); 496 io_close_queue(ctx->ifq); 497 } 498 499 static inline u32 io_zcrx_rqring_entries(struct io_zcrx_ifq *ifq) 500 { 501 u32 entries; 502 503 entries = smp_load_acquire(&ifq->rq_ring->tail) - ifq->cached_rq_head; 504 return min(entries, ifq->rq_entries); 505 } 506 507 static struct io_uring_zcrx_rqe *io_zcrx_get_rqe(struct io_zcrx_ifq *ifq, 508 unsigned mask) 509 { 510 unsigned int idx = ifq->cached_rq_head++ & mask; 511 512 return &ifq->rqes[idx]; 513 } 514 515 static void io_zcrx_ring_refill(struct page_pool *pp, 516 struct io_zcrx_ifq *ifq) 517 { 518 unsigned int mask = ifq->rq_entries - 1; 519 unsigned int entries; 520 netmem_ref netmem; 521 522 spin_lock_bh(&ifq->rq_lock); 523 524 entries = io_zcrx_rqring_entries(ifq); 525 entries = min_t(unsigned, entries, PP_ALLOC_CACHE_REFILL - pp->alloc.count); 526 if (unlikely(!entries)) { 527 spin_unlock_bh(&ifq->rq_lock); 528 return; 529 } 530 531 do { 532 struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(ifq, mask); 533 struct io_zcrx_area *area; 534 struct net_iov *niov; 535 unsigned niov_idx, area_idx; 536 537 area_idx = rqe->off >> IORING_ZCRX_AREA_SHIFT; 538 niov_idx = (rqe->off & ~IORING_ZCRX_AREA_MASK) >> PAGE_SHIFT; 539 540 if (unlikely(rqe->__pad || area_idx)) 541 continue; 542 area = ifq->area; 543 544 if (unlikely(niov_idx >= area->nia.num_niovs)) 545 continue; 546 niov_idx = array_index_nospec(niov_idx, area->nia.num_niovs); 547 548 niov = &area->nia.niovs[niov_idx]; 549 if (!io_zcrx_put_niov_uref(niov)) 550 continue; 551 552 netmem = net_iov_to_netmem(niov); 553 if (page_pool_unref_netmem(netmem, 1) != 0) 554 continue; 555 556 if (unlikely(niov->pp != pp)) { 557 io_zcrx_return_niov(niov); 558 continue; 559 } 560 561 io_zcrx_sync_for_device(pp, niov); 562 net_mp_netmem_place_in_cache(pp, netmem); 563 } while (--entries); 564 565 smp_store_release(&ifq->rq_ring->head, ifq->cached_rq_head); 566 spin_unlock_bh(&ifq->rq_lock); 567 } 568 569 static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq) 570 { 571 struct io_zcrx_area *area = ifq->area; 572 573 spin_lock_bh(&area->freelist_lock); 574 while (area->free_count && pp->alloc.count < PP_ALLOC_CACHE_REFILL) { 575 struct net_iov *niov = __io_zcrx_get_free_niov(area); 576 netmem_ref netmem = net_iov_to_netmem(niov); 577 578 net_mp_niov_set_page_pool(pp, niov); 579 io_zcrx_sync_for_device(pp, niov); 580 net_mp_netmem_place_in_cache(pp, netmem); 581 } 582 spin_unlock_bh(&area->freelist_lock); 583 } 584 585 static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp) 586 { 587 struct io_zcrx_ifq *ifq = pp->mp_priv; 588 589 /* pp should already be ensuring that */ 590 if (unlikely(pp->alloc.count)) 591 goto out_return; 592 593 io_zcrx_ring_refill(pp, ifq); 594 if (likely(pp->alloc.count)) 595 goto out_return; 596 597 io_zcrx_refill_slow(pp, ifq); 598 if (!pp->alloc.count) 599 return 0; 600 out_return: 601 return pp->alloc.cache[--pp->alloc.count]; 602 } 603 604 static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem) 605 { 606 struct net_iov *niov; 607 608 if (WARN_ON_ONCE(!netmem_is_net_iov(netmem))) 609 return false; 610 611 niov = netmem_to_net_iov(netmem); 612 net_mp_niov_clear_page_pool(niov); 613 io_zcrx_return_niov_freelist(niov); 614 return false; 615 } 616 617 static int io_pp_zc_init(struct page_pool *pp) 618 { 619 struct io_zcrx_ifq *ifq = pp->mp_priv; 620 621 if (WARN_ON_ONCE(!ifq)) 622 return -EINVAL; 623 if (WARN_ON_ONCE(ifq->dev != pp->p.dev)) 624 return -EINVAL; 625 if (WARN_ON_ONCE(!pp->dma_map)) 626 return -EOPNOTSUPP; 627 if (pp->p.order != 0) 628 return -EOPNOTSUPP; 629 if (pp->p.dma_dir != DMA_FROM_DEVICE) 630 return -EOPNOTSUPP; 631 632 percpu_ref_get(&ifq->ctx->refs); 633 return 0; 634 } 635 636 static void io_pp_zc_destroy(struct page_pool *pp) 637 { 638 struct io_zcrx_ifq *ifq = pp->mp_priv; 639 struct io_zcrx_area *area = ifq->area; 640 641 if (WARN_ON_ONCE(area->free_count != area->nia.num_niovs)) 642 return; 643 percpu_ref_put(&ifq->ctx->refs); 644 } 645 646 static int io_pp_nl_fill(void *mp_priv, struct sk_buff *rsp, 647 struct netdev_rx_queue *rxq) 648 { 649 struct nlattr *nest; 650 int type; 651 652 type = rxq ? NETDEV_A_QUEUE_IO_URING : NETDEV_A_PAGE_POOL_IO_URING; 653 nest = nla_nest_start(rsp, type); 654 if (!nest) 655 return -EMSGSIZE; 656 nla_nest_end(rsp, nest); 657 658 return 0; 659 } 660 661 static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq) 662 { 663 struct pp_memory_provider_params *p = &rxq->mp_params; 664 struct io_zcrx_ifq *ifq = mp_priv; 665 666 io_zcrx_drop_netdev(ifq); 667 p->mp_ops = NULL; 668 p->mp_priv = NULL; 669 } 670 671 static const struct memory_provider_ops io_uring_pp_zc_ops = { 672 .alloc_netmems = io_pp_zc_alloc_netmems, 673 .release_netmem = io_pp_zc_release_netmem, 674 .init = io_pp_zc_init, 675 .destroy = io_pp_zc_destroy, 676 .nl_fill = io_pp_nl_fill, 677 .uninstall = io_pp_uninstall, 678 }; 679 680 static bool io_zcrx_queue_cqe(struct io_kiocb *req, struct net_iov *niov, 681 struct io_zcrx_ifq *ifq, int off, int len) 682 { 683 struct io_uring_zcrx_cqe *rcqe; 684 struct io_zcrx_area *area; 685 struct io_uring_cqe *cqe; 686 u64 offset; 687 688 if (!io_defer_get_uncommited_cqe(req->ctx, &cqe)) 689 return false; 690 691 cqe->user_data = req->cqe.user_data; 692 cqe->res = len; 693 cqe->flags = IORING_CQE_F_MORE; 694 695 area = io_zcrx_iov_to_area(niov); 696 offset = off + (net_iov_idx(niov) << PAGE_SHIFT); 697 rcqe = (struct io_uring_zcrx_cqe *)(cqe + 1); 698 rcqe->off = offset + ((u64)area->area_id << IORING_ZCRX_AREA_SHIFT); 699 rcqe->__pad = 0; 700 return true; 701 } 702 703 static struct net_iov *io_zcrx_alloc_fallback(struct io_zcrx_area *area) 704 { 705 struct net_iov *niov = NULL; 706 707 spin_lock_bh(&area->freelist_lock); 708 if (area->free_count) 709 niov = __io_zcrx_get_free_niov(area); 710 spin_unlock_bh(&area->freelist_lock); 711 712 if (niov) 713 page_pool_fragment_netmem(net_iov_to_netmem(niov), 1); 714 return niov; 715 } 716 717 static ssize_t io_zcrx_copy_chunk(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 718 void *src_base, struct page *src_page, 719 unsigned int src_offset, size_t len) 720 { 721 struct io_zcrx_area *area = ifq->area; 722 size_t copied = 0; 723 int ret = 0; 724 725 while (len) { 726 size_t copy_size = min_t(size_t, PAGE_SIZE, len); 727 const int dst_off = 0; 728 struct net_iov *niov; 729 struct page *dst_page; 730 void *dst_addr; 731 732 niov = io_zcrx_alloc_fallback(area); 733 if (!niov) { 734 ret = -ENOMEM; 735 break; 736 } 737 738 dst_page = io_zcrx_iov_page(niov); 739 dst_addr = kmap_local_page(dst_page); 740 if (src_page) 741 src_base = kmap_local_page(src_page); 742 743 memcpy(dst_addr, src_base + src_offset, copy_size); 744 745 if (src_page) 746 kunmap_local(src_base); 747 kunmap_local(dst_addr); 748 749 if (!io_zcrx_queue_cqe(req, niov, ifq, dst_off, copy_size)) { 750 io_zcrx_return_niov(niov); 751 ret = -ENOSPC; 752 break; 753 } 754 755 io_zcrx_get_niov_uref(niov); 756 src_offset += copy_size; 757 len -= copy_size; 758 copied += copy_size; 759 } 760 761 return copied ? copied : ret; 762 } 763 764 static int io_zcrx_copy_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 765 const skb_frag_t *frag, int off, int len) 766 { 767 struct page *page = skb_frag_page(frag); 768 u32 p_off, p_len, t, copied = 0; 769 int ret = 0; 770 771 off += skb_frag_off(frag); 772 773 skb_frag_foreach_page(frag, off, len, 774 page, p_off, p_len, t) { 775 ret = io_zcrx_copy_chunk(req, ifq, NULL, page, p_off, p_len); 776 if (ret < 0) 777 return copied ? copied : ret; 778 copied += ret; 779 } 780 return copied; 781 } 782 783 static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 784 const skb_frag_t *frag, int off, int len) 785 { 786 struct net_iov *niov; 787 788 if (unlikely(!skb_frag_is_net_iov(frag))) 789 return io_zcrx_copy_frag(req, ifq, frag, off, len); 790 791 niov = netmem_to_net_iov(frag->netmem); 792 if (niov->pp->mp_ops != &io_uring_pp_zc_ops || 793 niov->pp->mp_priv != ifq) 794 return -EFAULT; 795 796 if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len)) 797 return -ENOSPC; 798 799 /* 800 * Prevent it from being recycled while user is accessing it. 801 * It has to be done before grabbing a user reference. 802 */ 803 page_pool_ref_netmem(net_iov_to_netmem(niov)); 804 io_zcrx_get_niov_uref(niov); 805 return len; 806 } 807 808 static int 809 io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb, 810 unsigned int offset, size_t len) 811 { 812 struct io_zcrx_args *args = desc->arg.data; 813 struct io_zcrx_ifq *ifq = args->ifq; 814 struct io_kiocb *req = args->req; 815 struct sk_buff *frag_iter; 816 unsigned start, start_off = offset; 817 int i, copy, end, off; 818 int ret = 0; 819 820 len = min_t(size_t, len, desc->count); 821 /* 822 * __tcp_read_sock() always calls io_zcrx_recv_skb one last time, even 823 * if desc->count is already 0. This is caused by the if (offset + 1 != 824 * skb->len) check. Return early in this case to break out of 825 * __tcp_read_sock(). 826 */ 827 if (!len) 828 return 0; 829 if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT)) 830 return -EAGAIN; 831 832 if (unlikely(offset < skb_headlen(skb))) { 833 ssize_t copied; 834 size_t to_copy; 835 836 to_copy = min_t(size_t, skb_headlen(skb) - offset, len); 837 copied = io_zcrx_copy_chunk(req, ifq, skb->data, NULL, 838 offset, to_copy); 839 if (copied < 0) { 840 ret = copied; 841 goto out; 842 } 843 offset += copied; 844 len -= copied; 845 if (!len) 846 goto out; 847 if (offset != skb_headlen(skb)) 848 goto out; 849 } 850 851 start = skb_headlen(skb); 852 853 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 854 const skb_frag_t *frag; 855 856 if (WARN_ON(start > offset + len)) 857 return -EFAULT; 858 859 frag = &skb_shinfo(skb)->frags[i]; 860 end = start + skb_frag_size(frag); 861 862 if (offset < end) { 863 copy = end - offset; 864 if (copy > len) 865 copy = len; 866 867 off = offset - start; 868 ret = io_zcrx_recv_frag(req, ifq, frag, off, copy); 869 if (ret < 0) 870 goto out; 871 872 offset += ret; 873 len -= ret; 874 if (len == 0 || ret != copy) 875 goto out; 876 } 877 start = end; 878 } 879 880 skb_walk_frags(skb, frag_iter) { 881 if (WARN_ON(start > offset + len)) 882 return -EFAULT; 883 884 end = start + frag_iter->len; 885 if (offset < end) { 886 copy = end - offset; 887 if (copy > len) 888 copy = len; 889 890 off = offset - start; 891 ret = io_zcrx_recv_skb(desc, frag_iter, off, copy); 892 if (ret < 0) 893 goto out; 894 895 offset += ret; 896 len -= ret; 897 if (len == 0 || ret != copy) 898 goto out; 899 } 900 start = end; 901 } 902 903 out: 904 if (offset == start_off) 905 return ret; 906 desc->count -= (offset - start_off); 907 return offset - start_off; 908 } 909 910 static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 911 struct sock *sk, int flags, 912 unsigned issue_flags, unsigned int *outlen) 913 { 914 unsigned int len = *outlen; 915 struct io_zcrx_args args = { 916 .req = req, 917 .ifq = ifq, 918 .sock = sk->sk_socket, 919 }; 920 read_descriptor_t rd_desc = { 921 .count = len ? len : UINT_MAX, 922 .arg.data = &args, 923 }; 924 int ret; 925 926 lock_sock(sk); 927 ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb); 928 if (len && ret > 0) 929 *outlen = len - ret; 930 if (ret <= 0) { 931 if (ret < 0 || sock_flag(sk, SOCK_DONE)) 932 goto out; 933 if (sk->sk_err) 934 ret = sock_error(sk); 935 else if (sk->sk_shutdown & RCV_SHUTDOWN) 936 goto out; 937 else if (sk->sk_state == TCP_CLOSE) 938 ret = -ENOTCONN; 939 else 940 ret = -EAGAIN; 941 } else if (unlikely(args.nr_skbs > IO_SKBS_PER_CALL_LIMIT) && 942 (issue_flags & IO_URING_F_MULTISHOT)) { 943 ret = IOU_REQUEUE; 944 } else if (sock_flag(sk, SOCK_DONE)) { 945 /* Make it to retry until it finally gets 0. */ 946 if (issue_flags & IO_URING_F_MULTISHOT) 947 ret = IOU_REQUEUE; 948 else 949 ret = -EAGAIN; 950 } 951 out: 952 release_sock(sk); 953 return ret; 954 } 955 956 int io_zcrx_recv(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 957 struct socket *sock, unsigned int flags, 958 unsigned issue_flags, unsigned int *len) 959 { 960 struct sock *sk = sock->sk; 961 const struct proto *prot = READ_ONCE(sk->sk_prot); 962 963 if (prot->recvmsg != tcp_recvmsg) 964 return -EPROTONOSUPPORT; 965 966 sock_rps_record_flow(sk); 967 return io_zcrx_tcp_recvmsg(req, ifq, sk, flags, issue_flags, len); 968 } 969