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