1 // SPDX-License-Identifier: GPL-2.0 2 3 #include <net/xsk_buff_pool.h> 4 #include <net/xdp_sock.h> 5 #include <net/xdp_sock_drv.h> 6 7 #include "xsk_queue.h" 8 #include "xdp_umem.h" 9 #include "xsk.h" 10 11 void xp_add_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs) 12 { 13 unsigned long flags; 14 15 if (!xs->tx) 16 return; 17 18 spin_lock_irqsave(&pool->xsk_tx_list_lock, flags); 19 list_add_rcu(&xs->tx_list, &pool->xsk_tx_list); 20 spin_unlock_irqrestore(&pool->xsk_tx_list_lock, flags); 21 } 22 23 void xp_del_xsk(struct xsk_buff_pool *pool, struct xdp_sock *xs) 24 { 25 unsigned long flags; 26 27 if (!xs->tx) 28 return; 29 30 spin_lock_irqsave(&pool->xsk_tx_list_lock, flags); 31 list_del_rcu(&xs->tx_list); 32 spin_unlock_irqrestore(&pool->xsk_tx_list_lock, flags); 33 } 34 35 void xp_destroy(struct xsk_buff_pool *pool) 36 { 37 if (!pool) 38 return; 39 40 kvfree(pool->tx_descs); 41 kvfree(pool->heads); 42 kvfree(pool); 43 } 44 45 int xp_alloc_tx_descs(struct xsk_buff_pool *pool, struct xdp_sock *xs) 46 { 47 pool->tx_descs = kvcalloc(xs->tx->nentries, sizeof(*pool->tx_descs), 48 GFP_KERNEL); 49 if (!pool->tx_descs) 50 return -ENOMEM; 51 52 return 0; 53 } 54 55 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs, 56 struct xdp_umem *umem) 57 { 58 bool unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG; 59 struct xsk_buff_pool *pool; 60 struct xdp_buff_xsk *xskb; 61 u32 i, entries; 62 63 entries = unaligned ? umem->chunks : 0; 64 pool = kvzalloc(struct_size(pool, free_heads, entries), GFP_KERNEL); 65 if (!pool) 66 goto out; 67 68 pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL); 69 if (!pool->heads) 70 goto out; 71 72 if (xs->tx) 73 if (xp_alloc_tx_descs(pool, xs)) 74 goto out; 75 76 pool->chunk_mask = ~((u64)umem->chunk_size - 1); 77 pool->addrs_cnt = umem->size; 78 pool->heads_cnt = umem->chunks; 79 pool->free_heads_cnt = umem->chunks; 80 pool->headroom = umem->headroom; 81 pool->chunk_size = umem->chunk_size; 82 pool->chunk_shift = ffs(umem->chunk_size) - 1; 83 pool->unaligned = unaligned; 84 pool->frame_len = umem->chunk_size - umem->headroom - 85 XDP_PACKET_HEADROOM; 86 pool->umem = umem; 87 pool->addrs = umem->addrs; 88 pool->tx_metadata_len = umem->tx_metadata_len; 89 pool->tx_sw_csum = umem->flags & XDP_UMEM_TX_SW_CSUM; 90 INIT_LIST_HEAD(&pool->free_list); 91 INIT_LIST_HEAD(&pool->xskb_list); 92 INIT_LIST_HEAD(&pool->xsk_tx_list); 93 spin_lock_init(&pool->xsk_tx_list_lock); 94 spin_lock_init(&pool->cq_lock); 95 refcount_set(&pool->users, 1); 96 97 pool->fq = xs->fq_tmp; 98 pool->cq = xs->cq_tmp; 99 100 for (i = 0; i < pool->free_heads_cnt; i++) { 101 xskb = &pool->heads[i]; 102 xskb->pool = pool; 103 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom; 104 INIT_LIST_HEAD(&xskb->free_list_node); 105 INIT_LIST_HEAD(&xskb->xskb_list_node); 106 if (pool->unaligned) 107 pool->free_heads[i] = xskb; 108 else 109 xp_init_xskb_addr(xskb, pool, i * pool->chunk_size); 110 } 111 112 return pool; 113 114 out: 115 xp_destroy(pool); 116 return NULL; 117 } 118 119 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq) 120 { 121 u32 i; 122 123 for (i = 0; i < pool->heads_cnt; i++) 124 pool->heads[i].xdp.rxq = rxq; 125 } 126 EXPORT_SYMBOL(xp_set_rxq_info); 127 128 void xp_fill_cb(struct xsk_buff_pool *pool, struct xsk_cb_desc *desc) 129 { 130 u32 i; 131 132 for (i = 0; i < pool->heads_cnt; i++) { 133 struct xdp_buff_xsk *xskb = &pool->heads[i]; 134 135 memcpy(xskb->cb + desc->off, desc->src, desc->bytes); 136 } 137 } 138 EXPORT_SYMBOL(xp_fill_cb); 139 140 static void xp_disable_drv_zc(struct xsk_buff_pool *pool) 141 { 142 struct netdev_bpf bpf; 143 int err; 144 145 ASSERT_RTNL(); 146 147 if (pool->umem->zc) { 148 bpf.command = XDP_SETUP_XSK_POOL; 149 bpf.xsk.pool = NULL; 150 bpf.xsk.queue_id = pool->queue_id; 151 152 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf); 153 154 if (err) 155 WARN(1, "Failed to disable zero-copy!\n"); 156 } 157 } 158 159 #define NETDEV_XDP_ACT_ZC (NETDEV_XDP_ACT_BASIC | \ 160 NETDEV_XDP_ACT_REDIRECT | \ 161 NETDEV_XDP_ACT_XSK_ZEROCOPY) 162 163 int xp_assign_dev(struct xsk_buff_pool *pool, 164 struct net_device *netdev, u16 queue_id, u16 flags) 165 { 166 bool force_zc, force_copy; 167 struct netdev_bpf bpf; 168 int err = 0; 169 170 ASSERT_RTNL(); 171 172 force_zc = flags & XDP_ZEROCOPY; 173 force_copy = flags & XDP_COPY; 174 175 if (force_zc && force_copy) 176 return -EINVAL; 177 178 if (xsk_get_pool_from_qid(netdev, queue_id)) 179 return -EBUSY; 180 181 pool->netdev = netdev; 182 pool->queue_id = queue_id; 183 err = xsk_reg_pool_at_qid(netdev, pool, queue_id); 184 if (err) 185 return err; 186 187 if (flags & XDP_USE_SG) 188 pool->umem->flags |= XDP_UMEM_SG_FLAG; 189 190 if (flags & XDP_USE_NEED_WAKEUP) 191 pool->uses_need_wakeup = true; 192 /* Tx needs to be explicitly woken up the first time. Also 193 * for supporting drivers that do not implement this 194 * feature. They will always have to call sendto() or poll(). 195 */ 196 pool->cached_need_wakeup = XDP_WAKEUP_TX; 197 198 dev_hold(netdev); 199 200 if (force_copy) 201 /* For copy-mode, we are done. */ 202 return 0; 203 204 if ((netdev->xdp_features & NETDEV_XDP_ACT_ZC) != NETDEV_XDP_ACT_ZC) { 205 err = -EOPNOTSUPP; 206 goto err_unreg_pool; 207 } 208 209 if (netdev->xdp_zc_max_segs == 1 && (flags & XDP_USE_SG)) { 210 err = -EOPNOTSUPP; 211 goto err_unreg_pool; 212 } 213 214 if (dev_get_min_mp_channel_count(netdev)) { 215 err = -EBUSY; 216 goto err_unreg_pool; 217 } 218 219 bpf.command = XDP_SETUP_XSK_POOL; 220 bpf.xsk.pool = pool; 221 bpf.xsk.queue_id = queue_id; 222 223 err = netdev->netdev_ops->ndo_bpf(netdev, &bpf); 224 if (err) 225 goto err_unreg_pool; 226 227 if (!pool->dma_pages) { 228 WARN(1, "Driver did not DMA map zero-copy buffers"); 229 err = -EINVAL; 230 goto err_unreg_xsk; 231 } 232 pool->umem->zc = true; 233 return 0; 234 235 err_unreg_xsk: 236 xp_disable_drv_zc(pool); 237 err_unreg_pool: 238 if (!force_zc) 239 err = 0; /* fallback to copy mode */ 240 if (err) { 241 xsk_clear_pool_at_qid(netdev, queue_id); 242 dev_put(netdev); 243 } 244 return err; 245 } 246 247 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs, 248 struct net_device *dev, u16 queue_id) 249 { 250 u16 flags; 251 struct xdp_umem *umem = umem_xs->umem; 252 253 /* One fill and completion ring required for each queue id. */ 254 if (!pool->fq || !pool->cq) 255 return -EINVAL; 256 257 flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY; 258 if (umem_xs->pool->uses_need_wakeup) 259 flags |= XDP_USE_NEED_WAKEUP; 260 261 return xp_assign_dev(pool, dev, queue_id, flags); 262 } 263 264 void xp_clear_dev(struct xsk_buff_pool *pool) 265 { 266 if (!pool->netdev) 267 return; 268 269 xp_disable_drv_zc(pool); 270 xsk_clear_pool_at_qid(pool->netdev, pool->queue_id); 271 dev_put(pool->netdev); 272 pool->netdev = NULL; 273 } 274 275 static void xp_release_deferred(struct work_struct *work) 276 { 277 struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool, 278 work); 279 280 rtnl_lock(); 281 xp_clear_dev(pool); 282 rtnl_unlock(); 283 284 if (pool->fq) { 285 xskq_destroy(pool->fq); 286 pool->fq = NULL; 287 } 288 289 if (pool->cq) { 290 xskq_destroy(pool->cq); 291 pool->cq = NULL; 292 } 293 294 xdp_put_umem(pool->umem, false); 295 xp_destroy(pool); 296 } 297 298 void xp_get_pool(struct xsk_buff_pool *pool) 299 { 300 refcount_inc(&pool->users); 301 } 302 303 bool xp_put_pool(struct xsk_buff_pool *pool) 304 { 305 if (!pool) 306 return false; 307 308 if (refcount_dec_and_test(&pool->users)) { 309 INIT_WORK(&pool->work, xp_release_deferred); 310 schedule_work(&pool->work); 311 return true; 312 } 313 314 return false; 315 } 316 317 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool) 318 { 319 struct xsk_dma_map *dma_map; 320 321 list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) { 322 if (dma_map->netdev == pool->netdev) 323 return dma_map; 324 } 325 326 return NULL; 327 } 328 329 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev, 330 u32 nr_pages, struct xdp_umem *umem) 331 { 332 struct xsk_dma_map *dma_map; 333 334 dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL); 335 if (!dma_map) 336 return NULL; 337 338 dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL); 339 if (!dma_map->dma_pages) { 340 kfree(dma_map); 341 return NULL; 342 } 343 344 dma_map->netdev = netdev; 345 dma_map->dev = dev; 346 dma_map->dma_pages_cnt = nr_pages; 347 refcount_set(&dma_map->users, 1); 348 list_add(&dma_map->list, &umem->xsk_dma_list); 349 return dma_map; 350 } 351 352 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map) 353 { 354 list_del(&dma_map->list); 355 kvfree(dma_map->dma_pages); 356 kfree(dma_map); 357 } 358 359 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs) 360 { 361 dma_addr_t *dma; 362 u32 i; 363 364 for (i = 0; i < dma_map->dma_pages_cnt; i++) { 365 dma = &dma_map->dma_pages[i]; 366 if (*dma) { 367 *dma &= ~XSK_NEXT_PG_CONTIG_MASK; 368 dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE, 369 DMA_BIDIRECTIONAL, attrs); 370 *dma = 0; 371 } 372 } 373 374 xp_destroy_dma_map(dma_map); 375 } 376 377 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs) 378 { 379 struct xsk_dma_map *dma_map; 380 381 if (!pool->dma_pages) 382 return; 383 384 dma_map = xp_find_dma_map(pool); 385 if (!dma_map) { 386 WARN(1, "Could not find dma_map for device"); 387 return; 388 } 389 390 if (!refcount_dec_and_test(&dma_map->users)) 391 return; 392 393 __xp_dma_unmap(dma_map, attrs); 394 kvfree(pool->dma_pages); 395 pool->dma_pages = NULL; 396 pool->dma_pages_cnt = 0; 397 pool->dev = NULL; 398 } 399 EXPORT_SYMBOL(xp_dma_unmap); 400 401 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map) 402 { 403 u32 i; 404 405 for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) { 406 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1]) 407 dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK; 408 else 409 dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK; 410 } 411 } 412 413 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map) 414 { 415 if (!pool->unaligned) { 416 u32 i; 417 418 for (i = 0; i < pool->heads_cnt; i++) { 419 struct xdp_buff_xsk *xskb = &pool->heads[i]; 420 421 xp_init_xskb_dma(xskb, pool, dma_map->dma_pages, xskb->orig_addr); 422 } 423 } 424 425 pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL); 426 if (!pool->dma_pages) 427 return -ENOMEM; 428 429 pool->dev = dma_map->dev; 430 pool->dma_pages_cnt = dma_map->dma_pages_cnt; 431 memcpy(pool->dma_pages, dma_map->dma_pages, 432 pool->dma_pages_cnt * sizeof(*pool->dma_pages)); 433 434 return 0; 435 } 436 437 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev, 438 unsigned long attrs, struct page **pages, u32 nr_pages) 439 { 440 struct xsk_dma_map *dma_map; 441 dma_addr_t dma; 442 int err; 443 u32 i; 444 445 dma_map = xp_find_dma_map(pool); 446 if (dma_map) { 447 err = xp_init_dma_info(pool, dma_map); 448 if (err) 449 return err; 450 451 refcount_inc(&dma_map->users); 452 return 0; 453 } 454 455 dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem); 456 if (!dma_map) 457 return -ENOMEM; 458 459 for (i = 0; i < dma_map->dma_pages_cnt; i++) { 460 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE, 461 DMA_BIDIRECTIONAL, attrs); 462 if (dma_mapping_error(dev, dma)) { 463 __xp_dma_unmap(dma_map, attrs); 464 return -ENOMEM; 465 } 466 dma_map->dma_pages[i] = dma; 467 } 468 469 if (pool->unaligned) 470 xp_check_dma_contiguity(dma_map); 471 472 err = xp_init_dma_info(pool, dma_map); 473 if (err) { 474 __xp_dma_unmap(dma_map, attrs); 475 return err; 476 } 477 478 return 0; 479 } 480 EXPORT_SYMBOL(xp_dma_map); 481 482 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool, 483 u64 addr) 484 { 485 return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size); 486 } 487 488 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr) 489 { 490 *addr = xp_unaligned_extract_addr(*addr); 491 if (*addr >= pool->addrs_cnt || 492 *addr + pool->chunk_size > pool->addrs_cnt || 493 xp_addr_crosses_non_contig_pg(pool, *addr)) 494 return false; 495 return true; 496 } 497 498 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr) 499 { 500 *addr = xp_aligned_extract_addr(pool, *addr); 501 return *addr < pool->addrs_cnt; 502 } 503 504 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool) 505 { 506 struct xdp_buff_xsk *xskb; 507 u64 addr; 508 bool ok; 509 510 if (pool->free_heads_cnt == 0) 511 return NULL; 512 513 for (;;) { 514 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) { 515 pool->fq->queue_empty_descs++; 516 return NULL; 517 } 518 519 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) : 520 xp_check_aligned(pool, &addr); 521 if (!ok) { 522 pool->fq->invalid_descs++; 523 xskq_cons_release(pool->fq); 524 continue; 525 } 526 break; 527 } 528 529 if (pool->unaligned) { 530 xskb = pool->free_heads[--pool->free_heads_cnt]; 531 xp_init_xskb_addr(xskb, pool, addr); 532 if (pool->dma_pages) 533 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr); 534 } else { 535 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)]; 536 } 537 538 xskq_cons_release(pool->fq); 539 return xskb; 540 } 541 542 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool) 543 { 544 struct xdp_buff_xsk *xskb; 545 546 if (!pool->free_list_cnt) { 547 xskb = __xp_alloc(pool); 548 if (!xskb) 549 return NULL; 550 } else { 551 pool->free_list_cnt--; 552 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk, 553 free_list_node); 554 list_del_init(&xskb->free_list_node); 555 } 556 557 xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM; 558 xskb->xdp.data_meta = xskb->xdp.data; 559 xskb->xdp.flags = 0; 560 561 if (pool->dev) 562 xp_dma_sync_for_device(pool, xskb->dma, pool->frame_len); 563 564 return &xskb->xdp; 565 } 566 EXPORT_SYMBOL(xp_alloc); 567 568 static u32 xp_alloc_new_from_fq(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max) 569 { 570 u32 i, cached_cons, nb_entries; 571 572 if (max > pool->free_heads_cnt) 573 max = pool->free_heads_cnt; 574 max = xskq_cons_nb_entries(pool->fq, max); 575 576 cached_cons = pool->fq->cached_cons; 577 nb_entries = max; 578 i = max; 579 while (i--) { 580 struct xdp_buff_xsk *xskb; 581 u64 addr; 582 bool ok; 583 584 __xskq_cons_read_addr_unchecked(pool->fq, cached_cons++, &addr); 585 586 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) : 587 xp_check_aligned(pool, &addr); 588 if (unlikely(!ok)) { 589 pool->fq->invalid_descs++; 590 nb_entries--; 591 continue; 592 } 593 594 if (pool->unaligned) { 595 xskb = pool->free_heads[--pool->free_heads_cnt]; 596 xp_init_xskb_addr(xskb, pool, addr); 597 if (pool->dma_pages) 598 xp_init_xskb_dma(xskb, pool, pool->dma_pages, addr); 599 } else { 600 xskb = &pool->heads[xp_aligned_extract_idx(pool, addr)]; 601 } 602 603 *xdp = &xskb->xdp; 604 xdp++; 605 } 606 607 xskq_cons_release_n(pool->fq, max); 608 return nb_entries; 609 } 610 611 static u32 xp_alloc_reused(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 nb_entries) 612 { 613 struct xdp_buff_xsk *xskb; 614 u32 i; 615 616 nb_entries = min_t(u32, nb_entries, pool->free_list_cnt); 617 618 i = nb_entries; 619 while (i--) { 620 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk, free_list_node); 621 list_del_init(&xskb->free_list_node); 622 623 *xdp = &xskb->xdp; 624 xdp++; 625 } 626 pool->free_list_cnt -= nb_entries; 627 628 return nb_entries; 629 } 630 631 static u32 xp_alloc_slow(struct xsk_buff_pool *pool, struct xdp_buff **xdp, 632 u32 max) 633 { 634 int i; 635 636 for (i = 0; i < max; i++) { 637 struct xdp_buff *buff; 638 639 buff = xp_alloc(pool); 640 if (unlikely(!buff)) 641 return i; 642 *xdp = buff; 643 xdp++; 644 } 645 646 return max; 647 } 648 649 u32 xp_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp, u32 max) 650 { 651 u32 nb_entries1 = 0, nb_entries2; 652 653 if (unlikely(pool->dev && dma_dev_need_sync(pool->dev))) 654 return xp_alloc_slow(pool, xdp, max); 655 656 if (unlikely(pool->free_list_cnt)) { 657 nb_entries1 = xp_alloc_reused(pool, xdp, max); 658 if (nb_entries1 == max) 659 return nb_entries1; 660 661 max -= nb_entries1; 662 xdp += nb_entries1; 663 } 664 665 nb_entries2 = xp_alloc_new_from_fq(pool, xdp, max); 666 if (!nb_entries2) 667 pool->fq->queue_empty_descs++; 668 669 return nb_entries1 + nb_entries2; 670 } 671 EXPORT_SYMBOL(xp_alloc_batch); 672 673 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count) 674 { 675 u32 req_count, avail_count; 676 677 if (pool->free_list_cnt >= count) 678 return true; 679 680 req_count = count - pool->free_list_cnt; 681 avail_count = xskq_cons_nb_entries(pool->fq, req_count); 682 if (!avail_count) 683 pool->fq->queue_empty_descs++; 684 685 return avail_count >= req_count; 686 } 687 EXPORT_SYMBOL(xp_can_alloc); 688 689 void xp_free(struct xdp_buff_xsk *xskb) 690 { 691 if (!list_empty(&xskb->free_list_node)) 692 return; 693 694 xskb->pool->free_list_cnt++; 695 list_add(&xskb->free_list_node, &xskb->pool->free_list); 696 } 697 EXPORT_SYMBOL(xp_free); 698 699 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr) 700 { 701 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr; 702 return pool->addrs + addr; 703 } 704 EXPORT_SYMBOL(xp_raw_get_data); 705 706 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr) 707 { 708 addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr; 709 return (pool->dma_pages[addr >> PAGE_SHIFT] & 710 ~XSK_NEXT_PG_CONTIG_MASK) + 711 (addr & ~PAGE_MASK); 712 } 713 EXPORT_SYMBOL(xp_raw_get_dma); 714