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