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 niov->type = NET_IOV_IOURING; 430 } 431 432 area->free_count = nr_iovs; 433 area->ifq = ifq; 434 /* we're only supporting one area per ifq for now */ 435 area->area_id = 0; 436 area_reg->rq_area_token = (u64)area->area_id << IORING_ZCRX_AREA_SHIFT; 437 spin_lock_init(&area->freelist_lock); 438 *res = area; 439 return 0; 440 err: 441 if (area) 442 io_zcrx_free_area(area); 443 return ret; 444 } 445 446 static struct io_zcrx_ifq *io_zcrx_ifq_alloc(struct io_ring_ctx *ctx) 447 { 448 struct io_zcrx_ifq *ifq; 449 450 ifq = kzalloc(sizeof(*ifq), GFP_KERNEL); 451 if (!ifq) 452 return NULL; 453 454 ifq->if_rxq = -1; 455 ifq->ctx = ctx; 456 spin_lock_init(&ifq->lock); 457 spin_lock_init(&ifq->rq_lock); 458 mutex_init(&ifq->dma_lock); 459 return ifq; 460 } 461 462 static void io_zcrx_drop_netdev(struct io_zcrx_ifq *ifq) 463 { 464 spin_lock(&ifq->lock); 465 if (ifq->netdev) { 466 netdev_put(ifq->netdev, &ifq->netdev_tracker); 467 ifq->netdev = NULL; 468 } 469 spin_unlock(&ifq->lock); 470 } 471 472 static void io_close_queue(struct io_zcrx_ifq *ifq) 473 { 474 struct net_device *netdev; 475 netdevice_tracker netdev_tracker; 476 struct pp_memory_provider_params p = { 477 .mp_ops = &io_uring_pp_zc_ops, 478 .mp_priv = ifq, 479 }; 480 481 if (ifq->if_rxq == -1) 482 return; 483 484 spin_lock(&ifq->lock); 485 netdev = ifq->netdev; 486 netdev_tracker = ifq->netdev_tracker; 487 ifq->netdev = NULL; 488 spin_unlock(&ifq->lock); 489 490 if (netdev) { 491 net_mp_close_rxq(netdev, ifq->if_rxq, &p); 492 netdev_put(netdev, &netdev_tracker); 493 } 494 ifq->if_rxq = -1; 495 } 496 497 static void io_zcrx_ifq_free(struct io_zcrx_ifq *ifq) 498 { 499 io_close_queue(ifq); 500 io_zcrx_drop_netdev(ifq); 501 502 if (ifq->area) 503 io_zcrx_free_area(ifq->area); 504 if (ifq->dev) 505 put_device(ifq->dev); 506 507 io_free_rbuf_ring(ifq); 508 mutex_destroy(&ifq->dma_lock); 509 kfree(ifq); 510 } 511 512 struct io_mapped_region *io_zcrx_get_region(struct io_ring_ctx *ctx, 513 unsigned int id) 514 { 515 struct io_zcrx_ifq *ifq = xa_load(&ctx->zcrx_ctxs, id); 516 517 lockdep_assert_held(&ctx->mmap_lock); 518 519 return ifq ? &ifq->region : NULL; 520 } 521 522 int io_register_zcrx_ifq(struct io_ring_ctx *ctx, 523 struct io_uring_zcrx_ifq_reg __user *arg) 524 { 525 struct pp_memory_provider_params mp_param = {}; 526 struct io_uring_zcrx_area_reg area; 527 struct io_uring_zcrx_ifq_reg reg; 528 struct io_uring_region_desc rd; 529 struct io_zcrx_ifq *ifq; 530 int ret; 531 u32 id; 532 533 /* 534 * 1. Interface queue allocation. 535 * 2. It can observe data destined for sockets of other tasks. 536 */ 537 if (!capable(CAP_NET_ADMIN)) 538 return -EPERM; 539 540 /* mandatory io_uring features for zc rx */ 541 if (!(ctx->flags & IORING_SETUP_DEFER_TASKRUN && 542 ctx->flags & IORING_SETUP_CQE32)) 543 return -EINVAL; 544 if (copy_from_user(®, arg, sizeof(reg))) 545 return -EFAULT; 546 if (copy_from_user(&rd, u64_to_user_ptr(reg.region_ptr), sizeof(rd))) 547 return -EFAULT; 548 if (memchr_inv(®.__resv, 0, sizeof(reg.__resv)) || 549 reg.__resv2 || reg.zcrx_id) 550 return -EINVAL; 551 if (reg.if_rxq == -1 || !reg.rq_entries || reg.flags) 552 return -EINVAL; 553 if (reg.rq_entries > IO_RQ_MAX_ENTRIES) { 554 if (!(ctx->flags & IORING_SETUP_CLAMP)) 555 return -EINVAL; 556 reg.rq_entries = IO_RQ_MAX_ENTRIES; 557 } 558 reg.rq_entries = roundup_pow_of_two(reg.rq_entries); 559 560 if (copy_from_user(&area, u64_to_user_ptr(reg.area_ptr), sizeof(area))) 561 return -EFAULT; 562 563 ifq = io_zcrx_ifq_alloc(ctx); 564 if (!ifq) 565 return -ENOMEM; 566 ifq->rq_entries = reg.rq_entries; 567 568 scoped_guard(mutex, &ctx->mmap_lock) { 569 /* preallocate id */ 570 ret = xa_alloc(&ctx->zcrx_ctxs, &id, NULL, xa_limit_31b, GFP_KERNEL); 571 if (ret) 572 goto ifq_free; 573 } 574 575 ret = io_allocate_rbuf_ring(ifq, ®, &rd, id); 576 if (ret) 577 goto err; 578 579 ifq->netdev = netdev_get_by_index(current->nsproxy->net_ns, reg.if_idx, 580 &ifq->netdev_tracker, GFP_KERNEL); 581 if (!ifq->netdev) { 582 ret = -ENODEV; 583 goto err; 584 } 585 586 ifq->dev = ifq->netdev->dev.parent; 587 if (!ifq->dev) { 588 ret = -EOPNOTSUPP; 589 goto err; 590 } 591 get_device(ifq->dev); 592 593 ret = io_zcrx_create_area(ifq, &ifq->area, &area); 594 if (ret) 595 goto err; 596 597 mp_param.mp_ops = &io_uring_pp_zc_ops; 598 mp_param.mp_priv = ifq; 599 ret = net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param); 600 if (ret) 601 goto err; 602 ifq->if_rxq = reg.if_rxq; 603 604 reg.offsets.rqes = sizeof(struct io_uring); 605 reg.offsets.head = offsetof(struct io_uring, head); 606 reg.offsets.tail = offsetof(struct io_uring, tail); 607 reg.zcrx_id = id; 608 609 scoped_guard(mutex, &ctx->mmap_lock) { 610 /* publish ifq */ 611 ret = -ENOMEM; 612 if (xa_store(&ctx->zcrx_ctxs, id, ifq, GFP_KERNEL)) 613 goto err; 614 } 615 616 if (copy_to_user(arg, ®, sizeof(reg)) || 617 copy_to_user(u64_to_user_ptr(reg.region_ptr), &rd, sizeof(rd)) || 618 copy_to_user(u64_to_user_ptr(reg.area_ptr), &area, sizeof(area))) { 619 ret = -EFAULT; 620 goto err; 621 } 622 return 0; 623 err: 624 scoped_guard(mutex, &ctx->mmap_lock) 625 xa_erase(&ctx->zcrx_ctxs, id); 626 ifq_free: 627 io_zcrx_ifq_free(ifq); 628 return ret; 629 } 630 631 void io_unregister_zcrx_ifqs(struct io_ring_ctx *ctx) 632 { 633 struct io_zcrx_ifq *ifq; 634 unsigned long id; 635 636 lockdep_assert_held(&ctx->uring_lock); 637 638 while (1) { 639 scoped_guard(mutex, &ctx->mmap_lock) { 640 ifq = xa_find(&ctx->zcrx_ctxs, &id, ULONG_MAX, XA_PRESENT); 641 if (ifq) 642 xa_erase(&ctx->zcrx_ctxs, id); 643 } 644 if (!ifq) 645 break; 646 io_zcrx_ifq_free(ifq); 647 } 648 649 xa_destroy(&ctx->zcrx_ctxs); 650 } 651 652 static struct net_iov *__io_zcrx_get_free_niov(struct io_zcrx_area *area) 653 { 654 unsigned niov_idx; 655 656 lockdep_assert_held(&area->freelist_lock); 657 658 niov_idx = area->freelist[--area->free_count]; 659 return &area->nia.niovs[niov_idx]; 660 } 661 662 static void io_zcrx_return_niov_freelist(struct net_iov *niov) 663 { 664 struct io_zcrx_area *area = io_zcrx_iov_to_area(niov); 665 666 spin_lock_bh(&area->freelist_lock); 667 area->freelist[area->free_count++] = net_iov_idx(niov); 668 spin_unlock_bh(&area->freelist_lock); 669 } 670 671 static void io_zcrx_return_niov(struct net_iov *niov) 672 { 673 netmem_ref netmem = net_iov_to_netmem(niov); 674 675 if (!niov->pp) { 676 /* copy fallback allocated niovs */ 677 io_zcrx_return_niov_freelist(niov); 678 return; 679 } 680 page_pool_put_unrefed_netmem(niov->pp, netmem, -1, false); 681 } 682 683 static void io_zcrx_scrub(struct io_zcrx_ifq *ifq) 684 { 685 struct io_zcrx_area *area = ifq->area; 686 int i; 687 688 if (!area) 689 return; 690 691 /* Reclaim back all buffers given to the user space. */ 692 for (i = 0; i < area->nia.num_niovs; i++) { 693 struct net_iov *niov = &area->nia.niovs[i]; 694 int nr; 695 696 if (!atomic_read(io_get_user_counter(niov))) 697 continue; 698 nr = atomic_xchg(io_get_user_counter(niov), 0); 699 if (nr && !page_pool_unref_netmem(net_iov_to_netmem(niov), nr)) 700 io_zcrx_return_niov(niov); 701 } 702 } 703 704 void io_shutdown_zcrx_ifqs(struct io_ring_ctx *ctx) 705 { 706 struct io_zcrx_ifq *ifq; 707 unsigned long index; 708 709 lockdep_assert_held(&ctx->uring_lock); 710 711 xa_for_each(&ctx->zcrx_ctxs, index, ifq) { 712 io_zcrx_scrub(ifq); 713 io_close_queue(ifq); 714 } 715 } 716 717 static inline u32 io_zcrx_rqring_entries(struct io_zcrx_ifq *ifq) 718 { 719 u32 entries; 720 721 entries = smp_load_acquire(&ifq->rq_ring->tail) - ifq->cached_rq_head; 722 return min(entries, ifq->rq_entries); 723 } 724 725 static struct io_uring_zcrx_rqe *io_zcrx_get_rqe(struct io_zcrx_ifq *ifq, 726 unsigned mask) 727 { 728 unsigned int idx = ifq->cached_rq_head++ & mask; 729 730 return &ifq->rqes[idx]; 731 } 732 733 static void io_zcrx_ring_refill(struct page_pool *pp, 734 struct io_zcrx_ifq *ifq) 735 { 736 unsigned int mask = ifq->rq_entries - 1; 737 unsigned int entries; 738 netmem_ref netmem; 739 740 spin_lock_bh(&ifq->rq_lock); 741 742 entries = io_zcrx_rqring_entries(ifq); 743 entries = min_t(unsigned, entries, PP_ALLOC_CACHE_REFILL - pp->alloc.count); 744 if (unlikely(!entries)) { 745 spin_unlock_bh(&ifq->rq_lock); 746 return; 747 } 748 749 do { 750 struct io_uring_zcrx_rqe *rqe = io_zcrx_get_rqe(ifq, mask); 751 struct io_zcrx_area *area; 752 struct net_iov *niov; 753 unsigned niov_idx, area_idx; 754 755 area_idx = rqe->off >> IORING_ZCRX_AREA_SHIFT; 756 niov_idx = (rqe->off & ~IORING_ZCRX_AREA_MASK) >> PAGE_SHIFT; 757 758 if (unlikely(rqe->__pad || area_idx)) 759 continue; 760 area = ifq->area; 761 762 if (unlikely(niov_idx >= area->nia.num_niovs)) 763 continue; 764 niov_idx = array_index_nospec(niov_idx, area->nia.num_niovs); 765 766 niov = &area->nia.niovs[niov_idx]; 767 if (!io_zcrx_put_niov_uref(niov)) 768 continue; 769 770 netmem = net_iov_to_netmem(niov); 771 if (page_pool_unref_netmem(netmem, 1) != 0) 772 continue; 773 774 if (unlikely(niov->pp != pp)) { 775 io_zcrx_return_niov(niov); 776 continue; 777 } 778 779 io_zcrx_sync_for_device(pp, niov); 780 net_mp_netmem_place_in_cache(pp, netmem); 781 } while (--entries); 782 783 smp_store_release(&ifq->rq_ring->head, ifq->cached_rq_head); 784 spin_unlock_bh(&ifq->rq_lock); 785 } 786 787 static void io_zcrx_refill_slow(struct page_pool *pp, struct io_zcrx_ifq *ifq) 788 { 789 struct io_zcrx_area *area = ifq->area; 790 791 spin_lock_bh(&area->freelist_lock); 792 while (area->free_count && pp->alloc.count < PP_ALLOC_CACHE_REFILL) { 793 struct net_iov *niov = __io_zcrx_get_free_niov(area); 794 netmem_ref netmem = net_iov_to_netmem(niov); 795 796 net_mp_niov_set_page_pool(pp, niov); 797 io_zcrx_sync_for_device(pp, niov); 798 net_mp_netmem_place_in_cache(pp, netmem); 799 } 800 spin_unlock_bh(&area->freelist_lock); 801 } 802 803 static netmem_ref io_pp_zc_alloc_netmems(struct page_pool *pp, gfp_t gfp) 804 { 805 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 806 807 /* pp should already be ensuring that */ 808 if (unlikely(pp->alloc.count)) 809 goto out_return; 810 811 io_zcrx_ring_refill(pp, ifq); 812 if (likely(pp->alloc.count)) 813 goto out_return; 814 815 io_zcrx_refill_slow(pp, ifq); 816 if (!pp->alloc.count) 817 return 0; 818 out_return: 819 return pp->alloc.cache[--pp->alloc.count]; 820 } 821 822 static bool io_pp_zc_release_netmem(struct page_pool *pp, netmem_ref netmem) 823 { 824 struct net_iov *niov; 825 826 if (WARN_ON_ONCE(!netmem_is_net_iov(netmem))) 827 return false; 828 829 niov = netmem_to_net_iov(netmem); 830 net_mp_niov_clear_page_pool(niov); 831 io_zcrx_return_niov_freelist(niov); 832 return false; 833 } 834 835 static int io_pp_zc_init(struct page_pool *pp) 836 { 837 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 838 int ret; 839 840 if (WARN_ON_ONCE(!ifq)) 841 return -EINVAL; 842 if (WARN_ON_ONCE(ifq->dev != pp->p.dev)) 843 return -EINVAL; 844 if (WARN_ON_ONCE(!pp->dma_map)) 845 return -EOPNOTSUPP; 846 if (pp->p.order != 0) 847 return -EOPNOTSUPP; 848 if (pp->p.dma_dir != DMA_FROM_DEVICE) 849 return -EOPNOTSUPP; 850 851 ret = io_zcrx_map_area(ifq, ifq->area); 852 if (ret) 853 return ret; 854 855 percpu_ref_get(&ifq->ctx->refs); 856 return 0; 857 } 858 859 static void io_pp_zc_destroy(struct page_pool *pp) 860 { 861 struct io_zcrx_ifq *ifq = io_pp_to_ifq(pp); 862 struct io_zcrx_area *area = ifq->area; 863 864 if (WARN_ON_ONCE(area->free_count != area->nia.num_niovs)) 865 return; 866 percpu_ref_put(&ifq->ctx->refs); 867 } 868 869 static int io_pp_nl_fill(void *mp_priv, struct sk_buff *rsp, 870 struct netdev_rx_queue *rxq) 871 { 872 struct nlattr *nest; 873 int type; 874 875 type = rxq ? NETDEV_A_QUEUE_IO_URING : NETDEV_A_PAGE_POOL_IO_URING; 876 nest = nla_nest_start(rsp, type); 877 if (!nest) 878 return -EMSGSIZE; 879 nla_nest_end(rsp, nest); 880 881 return 0; 882 } 883 884 static void io_pp_uninstall(void *mp_priv, struct netdev_rx_queue *rxq) 885 { 886 struct pp_memory_provider_params *p = &rxq->mp_params; 887 struct io_zcrx_ifq *ifq = mp_priv; 888 889 io_zcrx_drop_netdev(ifq); 890 if (ifq->area) 891 io_zcrx_unmap_area(ifq, ifq->area); 892 893 p->mp_ops = NULL; 894 p->mp_priv = NULL; 895 } 896 897 static const struct memory_provider_ops io_uring_pp_zc_ops = { 898 .alloc_netmems = io_pp_zc_alloc_netmems, 899 .release_netmem = io_pp_zc_release_netmem, 900 .init = io_pp_zc_init, 901 .destroy = io_pp_zc_destroy, 902 .nl_fill = io_pp_nl_fill, 903 .uninstall = io_pp_uninstall, 904 }; 905 906 static bool io_zcrx_queue_cqe(struct io_kiocb *req, struct net_iov *niov, 907 struct io_zcrx_ifq *ifq, int off, int len) 908 { 909 struct io_uring_zcrx_cqe *rcqe; 910 struct io_zcrx_area *area; 911 struct io_uring_cqe *cqe; 912 u64 offset; 913 914 if (!io_defer_get_uncommited_cqe(req->ctx, &cqe)) 915 return false; 916 917 cqe->user_data = req->cqe.user_data; 918 cqe->res = len; 919 cqe->flags = IORING_CQE_F_MORE; 920 921 area = io_zcrx_iov_to_area(niov); 922 offset = off + (net_iov_idx(niov) << PAGE_SHIFT); 923 rcqe = (struct io_uring_zcrx_cqe *)(cqe + 1); 924 rcqe->off = offset + ((u64)area->area_id << IORING_ZCRX_AREA_SHIFT); 925 rcqe->__pad = 0; 926 return true; 927 } 928 929 static struct net_iov *io_zcrx_alloc_fallback(struct io_zcrx_area *area) 930 { 931 struct net_iov *niov = NULL; 932 933 spin_lock_bh(&area->freelist_lock); 934 if (area->free_count) 935 niov = __io_zcrx_get_free_niov(area); 936 spin_unlock_bh(&area->freelist_lock); 937 938 if (niov) 939 page_pool_fragment_netmem(net_iov_to_netmem(niov), 1); 940 return niov; 941 } 942 943 static ssize_t io_zcrx_copy_chunk(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 944 void *src_base, struct page *src_page, 945 unsigned int src_offset, size_t len) 946 { 947 struct io_zcrx_area *area = ifq->area; 948 size_t copied = 0; 949 int ret = 0; 950 951 if (area->mem.is_dmabuf) 952 return -EFAULT; 953 954 while (len) { 955 size_t copy_size = min_t(size_t, PAGE_SIZE, len); 956 const int dst_off = 0; 957 struct net_iov *niov; 958 struct page *dst_page; 959 void *dst_addr; 960 961 niov = io_zcrx_alloc_fallback(area); 962 if (!niov) { 963 ret = -ENOMEM; 964 break; 965 } 966 967 dst_page = io_zcrx_iov_page(niov); 968 dst_addr = kmap_local_page(dst_page); 969 if (src_page) 970 src_base = kmap_local_page(src_page); 971 972 memcpy(dst_addr, src_base + src_offset, copy_size); 973 974 if (src_page) 975 kunmap_local(src_base); 976 kunmap_local(dst_addr); 977 978 if (!io_zcrx_queue_cqe(req, niov, ifq, dst_off, copy_size)) { 979 io_zcrx_return_niov(niov); 980 ret = -ENOSPC; 981 break; 982 } 983 984 io_zcrx_get_niov_uref(niov); 985 src_offset += copy_size; 986 len -= copy_size; 987 copied += copy_size; 988 } 989 990 return copied ? copied : ret; 991 } 992 993 static int io_zcrx_copy_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 994 const skb_frag_t *frag, int off, int len) 995 { 996 struct page *page = skb_frag_page(frag); 997 u32 p_off, p_len, t, copied = 0; 998 int ret = 0; 999 1000 off += skb_frag_off(frag); 1001 1002 skb_frag_foreach_page(frag, off, len, 1003 page, p_off, p_len, t) { 1004 ret = io_zcrx_copy_chunk(req, ifq, NULL, page, p_off, p_len); 1005 if (ret < 0) 1006 return copied ? copied : ret; 1007 copied += ret; 1008 } 1009 return copied; 1010 } 1011 1012 static int io_zcrx_recv_frag(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1013 const skb_frag_t *frag, int off, int len) 1014 { 1015 struct net_iov *niov; 1016 1017 if (unlikely(!skb_frag_is_net_iov(frag))) 1018 return io_zcrx_copy_frag(req, ifq, frag, off, len); 1019 1020 niov = netmem_to_net_iov(frag->netmem); 1021 if (!niov->pp || niov->pp->mp_ops != &io_uring_pp_zc_ops || 1022 io_pp_to_ifq(niov->pp) != ifq) 1023 return -EFAULT; 1024 1025 if (!io_zcrx_queue_cqe(req, niov, ifq, off + skb_frag_off(frag), len)) 1026 return -ENOSPC; 1027 1028 /* 1029 * Prevent it from being recycled while user is accessing it. 1030 * It has to be done before grabbing a user reference. 1031 */ 1032 page_pool_ref_netmem(net_iov_to_netmem(niov)); 1033 io_zcrx_get_niov_uref(niov); 1034 return len; 1035 } 1036 1037 static int 1038 io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb, 1039 unsigned int offset, size_t len) 1040 { 1041 struct io_zcrx_args *args = desc->arg.data; 1042 struct io_zcrx_ifq *ifq = args->ifq; 1043 struct io_kiocb *req = args->req; 1044 struct sk_buff *frag_iter; 1045 unsigned start, start_off = offset; 1046 int i, copy, end, off; 1047 int ret = 0; 1048 1049 len = min_t(size_t, len, desc->count); 1050 /* 1051 * __tcp_read_sock() always calls io_zcrx_recv_skb one last time, even 1052 * if desc->count is already 0. This is caused by the if (offset + 1 != 1053 * skb->len) check. Return early in this case to break out of 1054 * __tcp_read_sock(). 1055 */ 1056 if (!len) 1057 return 0; 1058 if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT)) 1059 return -EAGAIN; 1060 1061 if (unlikely(offset < skb_headlen(skb))) { 1062 ssize_t copied; 1063 size_t to_copy; 1064 1065 to_copy = min_t(size_t, skb_headlen(skb) - offset, len); 1066 copied = io_zcrx_copy_chunk(req, ifq, skb->data, NULL, 1067 offset, to_copy); 1068 if (copied < 0) { 1069 ret = copied; 1070 goto out; 1071 } 1072 offset += copied; 1073 len -= copied; 1074 if (!len) 1075 goto out; 1076 if (offset != skb_headlen(skb)) 1077 goto out; 1078 } 1079 1080 start = skb_headlen(skb); 1081 1082 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 1083 const skb_frag_t *frag; 1084 1085 if (WARN_ON(start > offset + len)) 1086 return -EFAULT; 1087 1088 frag = &skb_shinfo(skb)->frags[i]; 1089 end = start + skb_frag_size(frag); 1090 1091 if (offset < end) { 1092 copy = end - offset; 1093 if (copy > len) 1094 copy = len; 1095 1096 off = offset - start; 1097 ret = io_zcrx_recv_frag(req, ifq, frag, off, copy); 1098 if (ret < 0) 1099 goto out; 1100 1101 offset += ret; 1102 len -= ret; 1103 if (len == 0 || ret != copy) 1104 goto out; 1105 } 1106 start = end; 1107 } 1108 1109 skb_walk_frags(skb, frag_iter) { 1110 if (WARN_ON(start > offset + len)) 1111 return -EFAULT; 1112 1113 end = start + frag_iter->len; 1114 if (offset < end) { 1115 copy = end - offset; 1116 if (copy > len) 1117 copy = len; 1118 1119 off = offset - start; 1120 ret = io_zcrx_recv_skb(desc, frag_iter, off, copy); 1121 if (ret < 0) 1122 goto out; 1123 1124 offset += ret; 1125 len -= ret; 1126 if (len == 0 || ret != copy) 1127 goto out; 1128 } 1129 start = end; 1130 } 1131 1132 out: 1133 if (offset == start_off) 1134 return ret; 1135 desc->count -= (offset - start_off); 1136 return offset - start_off; 1137 } 1138 1139 static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1140 struct sock *sk, int flags, 1141 unsigned issue_flags, unsigned int *outlen) 1142 { 1143 unsigned int len = *outlen; 1144 struct io_zcrx_args args = { 1145 .req = req, 1146 .ifq = ifq, 1147 .sock = sk->sk_socket, 1148 }; 1149 read_descriptor_t rd_desc = { 1150 .count = len ? len : UINT_MAX, 1151 .arg.data = &args, 1152 }; 1153 int ret; 1154 1155 lock_sock(sk); 1156 ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb); 1157 if (len && ret > 0) 1158 *outlen = len - ret; 1159 if (ret <= 0) { 1160 if (ret < 0 || sock_flag(sk, SOCK_DONE)) 1161 goto out; 1162 if (sk->sk_err) 1163 ret = sock_error(sk); 1164 else if (sk->sk_shutdown & RCV_SHUTDOWN) 1165 goto out; 1166 else if (sk->sk_state == TCP_CLOSE) 1167 ret = -ENOTCONN; 1168 else 1169 ret = -EAGAIN; 1170 } else if (unlikely(args.nr_skbs > IO_SKBS_PER_CALL_LIMIT) && 1171 (issue_flags & IO_URING_F_MULTISHOT)) { 1172 ret = IOU_REQUEUE; 1173 } else if (sock_flag(sk, SOCK_DONE)) { 1174 /* Make it to retry until it finally gets 0. */ 1175 if (issue_flags & IO_URING_F_MULTISHOT) 1176 ret = IOU_REQUEUE; 1177 else 1178 ret = -EAGAIN; 1179 } 1180 out: 1181 release_sock(sk); 1182 return ret; 1183 } 1184 1185 int io_zcrx_recv(struct io_kiocb *req, struct io_zcrx_ifq *ifq, 1186 struct socket *sock, unsigned int flags, 1187 unsigned issue_flags, unsigned int *len) 1188 { 1189 struct sock *sk = sock->sk; 1190 const struct proto *prot = READ_ONCE(sk->sk_prot); 1191 1192 if (prot->recvmsg != tcp_recvmsg) 1193 return -EPROTONOSUPPORT; 1194 1195 sock_rps_record_flow(sk); 1196 return io_zcrx_tcp_recvmsg(req, ifq, sk, flags, issue_flags, len); 1197 } 1198