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