net/xdp/xsk_buff_pool.c

   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         struct xsk_buff_pool *pool;
  59         struct xdp_buff_xsk *xskb;
  60         u32 i;
  61
  62         pool = kvzalloc(struct_size(pool, free_heads, umem->chunks),
  63                         GFP_KERNEL);
  64         if (!pool)
  65                 goto out;
  66
  67         pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
  68         if (!pool->heads)
  69                 goto out;
  70
  71         if (xs->tx)
  72                 if (xp_alloc_tx_descs(pool, xs))
  73                         goto out;
  74
  75         pool->chunk_mask = ~((u64)umem->chunk_size - 1);
  76         pool->addrs_cnt = umem->size;
  77         pool->heads_cnt = umem->chunks;
  78         pool->free_heads_cnt = umem->chunks;
  79         pool->headroom = umem->headroom;
  80         pool->chunk_size = umem->chunk_size;
  81         pool->unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
  82         pool->frame_len = umem->chunk_size - umem->headroom -
  83                 XDP_PACKET_HEADROOM;
  84         pool->umem = umem;
  85         pool->addrs = umem->addrs;
  86         INIT_LIST_HEAD(&pool->free_list);
  87         INIT_LIST_HEAD(&pool->xsk_tx_list);
  88         spin_lock_init(&pool->xsk_tx_list_lock);
  89         spin_lock_init(&pool->cq_lock);
  90         refcount_set(&pool->users, 1);
  91
  92         pool->fq = xs->fq_tmp;
  93         pool->cq = xs->cq_tmp;
  94
  95         for (i = 0; i < pool->free_heads_cnt; i++) {
  96                 xskb = &pool->heads[i];
  97                 xskb->pool = pool;
  98                 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
  99                 pool->free_heads[i] = xskb;
 100         }
 101
 102         return pool;
 103
 104 out:
 105         xp_destroy(pool);
 106         return NULL;
 107 }
 108
 109 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
 110 {
 111         u32 i;
 112
 113         for (i = 0; i < pool->heads_cnt; i++)
 114                 pool->heads[i].xdp.rxq = rxq;
 115 }
 116 EXPORT_SYMBOL(xp_set_rxq_info);
 117
 118 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
 119 {
 120         struct netdev_bpf bpf;
 121         int err;
 122
 123         ASSERT_RTNL();
 124
 125         if (pool->umem->zc) {
 126                 bpf.command = XDP_SETUP_XSK_POOL;
 127                 bpf.xsk.pool = NULL;
 128                 bpf.xsk.queue_id = pool->queue_id;
 129
 130                 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
 131
 132                 if (err)
 133                         WARN(1, "Failed to disable zero-copy!\n");
 134         }
 135 }
 136
 137 int xp_assign_dev(struct xsk_buff_pool *pool,
 138                   struct net_device *netdev, u16 queue_id, u16 flags)
 139 {
 140         bool force_zc, force_copy;
 141         struct netdev_bpf bpf;
 142         int err = 0;
 143
 144         ASSERT_RTNL();
 145
 146         force_zc = flags & XDP_ZEROCOPY;
 147         force_copy = flags & XDP_COPY;
 148
 149         if (force_zc && force_copy)
 150                 return -EINVAL;
 151
 152         if (xsk_get_pool_from_qid(netdev, queue_id))
 153                 return -EBUSY;
 154
 155         pool->netdev = netdev;
 156         pool->queue_id = queue_id;
 157         err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
 158         if (err)
 159                 return err;
 160
 161         if (flags & XDP_USE_NEED_WAKEUP)
 162                 pool->uses_need_wakeup = true;
 163         /* Tx needs to be explicitly woken up the first time.  Also
 164          * for supporting drivers that do not implement this
 165          * feature. They will always have to call sendto() or poll().
 166          */
 167         pool->cached_need_wakeup = XDP_WAKEUP_TX;
 168
 169         dev_hold(netdev);
 170
 171         if (force_copy)
 172                 /* For copy-mode, we are done. */
 173                 return 0;
 174
 175         if (!netdev->netdev_ops->ndo_bpf ||
 176             !netdev->netdev_ops->ndo_xsk_wakeup) {
 177                 err = -EOPNOTSUPP;
 178                 goto err_unreg_pool;
 179         }
 180
 181         bpf.command = XDP_SETUP_XSK_POOL;
 182         bpf.xsk.pool = pool;
 183         bpf.xsk.queue_id = queue_id;
 184
 185         err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
 186         if (err)
 187                 goto err_unreg_pool;
 188
 189         if (!pool->dma_pages) {
 190                 WARN(1, "Driver did not DMA map zero-copy buffers");
 191                 err = -EINVAL;
 192                 goto err_unreg_xsk;
 193         }
 194         pool->umem->zc = true;
 195         return 0;
 196
 197 err_unreg_xsk:
 198         xp_disable_drv_zc(pool);
 199 err_unreg_pool:
 200         if (!force_zc)
 201                 err = 0; /* fallback to copy mode */
 202         if (err) {
 203                 xsk_clear_pool_at_qid(netdev, queue_id);
 204                 dev_put(netdev);
 205         }
 206         return err;
 207 }
 208
 209 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
 210                          struct net_device *dev, u16 queue_id)
 211 {
 212         u16 flags;
 213         struct xdp_umem *umem = umem_xs->umem;
 214
 215         /* One fill and completion ring required for each queue id. */
 216         if (!pool->fq || !pool->cq)
 217                 return -EINVAL;
 218
 219         flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
 220         if (umem_xs->pool->uses_need_wakeup)
 221                 flags |= XDP_USE_NEED_WAKEUP;
 222
 223         return xp_assign_dev(pool, dev, queue_id, flags);
 224 }
 225
 226 void xp_clear_dev(struct xsk_buff_pool *pool)
 227 {
 228         if (!pool->netdev)
 229                 return;
 230
 231         xp_disable_drv_zc(pool);
 232         xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
 233         dev_put(pool->netdev);
 234         pool->netdev = NULL;
 235 }
 236
 237 static void xp_release_deferred(struct work_struct *work)
 238 {
 239         struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
 240                                                   work);
 241
 242         rtnl_lock();
 243         xp_clear_dev(pool);
 244         rtnl_unlock();
 245
 246         if (pool->fq) {
 247                 xskq_destroy(pool->fq);
 248                 pool->fq = NULL;
 249         }
 250
 251         if (pool->cq) {
 252                 xskq_destroy(pool->cq);
 253                 pool->cq = NULL;
 254         }
 255
 256         xdp_put_umem(pool->umem, false);
 257         xp_destroy(pool);
 258 }
 259
 260 void xp_get_pool(struct xsk_buff_pool *pool)
 261 {
 262         refcount_inc(&pool->users);
 263 }
 264
 265 bool xp_put_pool(struct xsk_buff_pool *pool)
 266 {
 267         if (!pool)
 268                 return false;
 269
 270         if (refcount_dec_and_test(&pool->users)) {
 271                 INIT_WORK(&pool->work, xp_release_deferred);
 272                 schedule_work(&pool->work);
 273                 return true;
 274         }
 275
 276         return false;
 277 }
 278
 279 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
 280 {
 281         struct xsk_dma_map *dma_map;
 282
 283         list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
 284                 if (dma_map->netdev == pool->netdev)
 285                         return dma_map;
 286         }
 287
 288         return NULL;
 289 }
 290
 291 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
 292                                              u32 nr_pages, struct xdp_umem *umem)
 293 {
 294         struct xsk_dma_map *dma_map;
 295
 296         dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
 297         if (!dma_map)
 298                 return NULL;
 299
 300         dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
 301         if (!dma_map->dma_pages) {
 302                 kfree(dma_map);
 303                 return NULL;
 304         }
 305
 306         dma_map->netdev = netdev;
 307         dma_map->dev = dev;
 308         dma_map->dma_need_sync = false;
 309         dma_map->dma_pages_cnt = nr_pages;
 310         refcount_set(&dma_map->users, 1);
 311         list_add(&dma_map->list, &umem->xsk_dma_list);
 312         return dma_map;
 313 }
 314
 315 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
 316 {
 317         list_del(&dma_map->list);
 318         kvfree(dma_map->dma_pages);
 319         kfree(dma_map);
 320 }
 321
 322 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
 323 {
 324         dma_addr_t *dma;
 325         u32 i;
 326
 327         for (i = 0; i < dma_map->dma_pages_cnt; i++) {
 328                 dma = &dma_map->dma_pages[i];
 329                 if (*dma) {
 330                         *dma &= ~XSK_NEXT_PG_CONTIG_MASK;
 331                         dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
 332                                              DMA_BIDIRECTIONAL, attrs);
 333                         *dma = 0;
 334                 }
 335         }
 336
 337         xp_destroy_dma_map(dma_map);
 338 }
 339
 340 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
 341 {
 342         struct xsk_dma_map *dma_map;
 343
 344         if (pool->dma_pages_cnt == 0)
 345                 return;
 346
 347         dma_map = xp_find_dma_map(pool);
 348         if (!dma_map) {
 349                 WARN(1, "Could not find dma_map for device");
 350                 return;
 351         }
 352
 353         if (!refcount_dec_and_test(&dma_map->users))
 354                 return;
 355
 356         __xp_dma_unmap(dma_map, attrs);
 357         kvfree(pool->dma_pages);
 358         pool->dma_pages_cnt = 0;
 359         pool->dev = NULL;
 360 }
 361 EXPORT_SYMBOL(xp_dma_unmap);
 362
 363 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
 364 {
 365         u32 i;
 366
 367         for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
 368                 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
 369                         dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
 370                 else
 371                         dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
 372         }
 373 }
 374
 375 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
 376 {
 377         pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
 378         if (!pool->dma_pages)
 379                 return -ENOMEM;
 380
 381         pool->dev = dma_map->dev;
 382         pool->dma_pages_cnt = dma_map->dma_pages_cnt;
 383         pool->dma_need_sync = dma_map->dma_need_sync;
 384         memcpy(pool->dma_pages, dma_map->dma_pages,
 385                pool->dma_pages_cnt * sizeof(*pool->dma_pages));
 386
 387         return 0;
 388 }
 389
 390 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
 391                unsigned long attrs, struct page **pages, u32 nr_pages)
 392 {
 393         struct xsk_dma_map *dma_map;
 394         dma_addr_t dma;
 395         int err;
 396         u32 i;
 397
 398         dma_map = xp_find_dma_map(pool);
 399         if (dma_map) {
 400                 err = xp_init_dma_info(pool, dma_map);
 401                 if (err)
 402                         return err;
 403
 404                 refcount_inc(&dma_map->users);
 405                 return 0;
 406         }
 407
 408         dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
 409         if (!dma_map)
 410                 return -ENOMEM;
 411
 412         for (i = 0; i < dma_map->dma_pages_cnt; i++) {
 413                 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
 414                                          DMA_BIDIRECTIONAL, attrs);
 415                 if (dma_mapping_error(dev, dma)) {
 416                         __xp_dma_unmap(dma_map, attrs);
 417                         return -ENOMEM;
 418                 }
 419                 if (dma_need_sync(dev, dma))
 420                         dma_map->dma_need_sync = true;
 421                 dma_map->dma_pages[i] = dma;
 422         }
 423
 424         if (pool->unaligned)
 425                 xp_check_dma_contiguity(dma_map);
 426
 427         err = xp_init_dma_info(pool, dma_map);
 428         if (err) {
 429                 __xp_dma_unmap(dma_map, attrs);
 430                 return err;
 431         }
 432
 433         return 0;
 434 }
 435 EXPORT_SYMBOL(xp_dma_map);
 436
 437 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
 438                                           u64 addr)
 439 {
 440         return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
 441 }
 442
 443 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
 444 {
 445         *addr = xp_unaligned_extract_addr(*addr);
 446         if (*addr >= pool->addrs_cnt ||
 447             *addr + pool->chunk_size > pool->addrs_cnt ||
 448             xp_addr_crosses_non_contig_pg(pool, *addr))
 449                 return false;
 450         return true;
 451 }
 452
 453 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
 454 {
 455         *addr = xp_aligned_extract_addr(pool, *addr);
 456         return *addr < pool->addrs_cnt;
 457 }
 458
 459 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
 460 {
 461         struct xdp_buff_xsk *xskb;
 462         u64 addr;
 463         bool ok;
 464
 465         if (pool->free_heads_cnt == 0)
 466                 return NULL;
 467
 468         xskb = pool->free_heads[--pool->free_heads_cnt];
 469
 470         for (;;) {
 471                 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
 472                         pool->fq->queue_empty_descs++;
 473                         xp_release(xskb);
 474                         return NULL;
 475                 }
 476
 477                 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
 478                      xp_check_aligned(pool, &addr);
 479                 if (!ok) {
 480                         pool->fq->invalid_descs++;
 481                         xskq_cons_release(pool->fq);
 482                         continue;
 483                 }
 484                 break;
 485         }
 486         xskq_cons_release(pool->fq);
 487
 488         xskb->orig_addr = addr;
 489         xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
 490         if (pool->dma_pages_cnt) {
 491                 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
 492                                    ~XSK_NEXT_PG_CONTIG_MASK) +
 493                                   (addr & ~PAGE_MASK);
 494                 xskb->dma = xskb->frame_dma + pool->headroom +
 495                             XDP_PACKET_HEADROOM;
 496         }
 497         return xskb;
 498 }
 499
 500 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
 501 {
 502         struct xdp_buff_xsk *xskb;
 503
 504         if (!pool->free_list_cnt) {
 505                 xskb = __xp_alloc(pool);
 506                 if (!xskb)
 507                         return NULL;
 508         } else {
 509                 pool->free_list_cnt--;
 510                 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
 511                                         free_list_node);
 512                 list_del(&xskb->free_list_node);
 513         }
 514
 515         xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
 516         xskb->xdp.data_meta = xskb->xdp.data;
 517
 518         if (pool->dma_need_sync) {
 519                 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
 520                                                  pool->frame_len,
 521                                                  DMA_BIDIRECTIONAL);
 522         }
 523         return &xskb->xdp;
 524 }
 525 EXPORT_SYMBOL(xp_alloc);
 526
 527 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
 528 {
 529         if (pool->free_list_cnt >= count)
 530                 return true;
 531         return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
 532 }
 533 EXPORT_SYMBOL(xp_can_alloc);
 534
 535 void xp_free(struct xdp_buff_xsk *xskb)
 536 {
 537         xskb->pool->free_list_cnt++;
 538         list_add(&xskb->free_list_node, &xskb->pool->free_list);
 539 }
 540 EXPORT_SYMBOL(xp_free);
 541
 542 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 543 {
 544         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 545         return pool->addrs + addr;
 546 }
 547 EXPORT_SYMBOL(xp_raw_get_data);
 548
 549 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
 550 {
 551         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 552         return (pool->dma_pages[addr >> PAGE_SHIFT] &
 553                 ~XSK_NEXT_PG_CONTIG_MASK) +
 554                 (addr & ~PAGE_MASK);
 555 }
 556 EXPORT_SYMBOL(xp_raw_get_dma);
 557
 558 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
 559 {
 560         dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
 561                                       xskb->pool->frame_len, DMA_BIDIRECTIONAL);
 562 }
 563 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
 564
 565 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
 566                                  size_t size)
 567 {
 568         dma_sync_single_range_for_device(pool->dev, dma, 0,
 569                                          size, DMA_BIDIRECTIONAL);
 570 }
 571 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);