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->heads);
  41         kvfree(pool);
  42 }
  43
  44 struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
  45                                                 struct xdp_umem *umem)
  46 {
  47         struct xsk_buff_pool *pool;
  48         struct xdp_buff_xsk *xskb;
  49         u32 i;
  50
  51         pool = kvzalloc(struct_size(pool, free_heads, umem->chunks),
  52                         GFP_KERNEL);
  53         if (!pool)
  54                 goto out;
  55
  56         pool->heads = kvcalloc(umem->chunks, sizeof(*pool->heads), GFP_KERNEL);
  57         if (!pool->heads)
  58                 goto out;
  59
  60         pool->chunk_mask = ~((u64)umem->chunk_size - 1);
  61         pool->addrs_cnt = umem->size;
  62         pool->heads_cnt = umem->chunks;
  63         pool->free_heads_cnt = umem->chunks;
  64         pool->headroom = umem->headroom;
  65         pool->chunk_size = umem->chunk_size;
  66         pool->unaligned = umem->flags & XDP_UMEM_UNALIGNED_CHUNK_FLAG;
  67         pool->frame_len = umem->chunk_size - umem->headroom -
  68                 XDP_PACKET_HEADROOM;
  69         pool->umem = umem;
  70         pool->addrs = umem->addrs;
  71         INIT_LIST_HEAD(&pool->free_list);
  72         INIT_LIST_HEAD(&pool->xsk_tx_list);
  73         spin_lock_init(&pool->xsk_tx_list_lock);
  74         refcount_set(&pool->users, 1);
  75
  76         pool->fq = xs->fq_tmp;
  77         pool->cq = xs->cq_tmp;
  78         xs->fq_tmp = NULL;
  79         xs->cq_tmp = NULL;
  80
  81         for (i = 0; i < pool->free_heads_cnt; i++) {
  82                 xskb = &pool->heads[i];
  83                 xskb->pool = pool;
  84                 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
  85                 pool->free_heads[i] = xskb;
  86         }
  87
  88         return pool;
  89
  90 out:
  91         xp_destroy(pool);
  92         return NULL;
  93 }
  94
  95 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
  96 {
  97         u32 i;
  98
  99         for (i = 0; i < pool->heads_cnt; i++)
 100                 pool->heads[i].xdp.rxq = rxq;
 101 }
 102 EXPORT_SYMBOL(xp_set_rxq_info);
 103
 104 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
 105 {
 106         struct netdev_bpf bpf;
 107         int err;
 108
 109         ASSERT_RTNL();
 110
 111         if (pool->umem->zc) {
 112                 bpf.command = XDP_SETUP_XSK_POOL;
 113                 bpf.xsk.pool = NULL;
 114                 bpf.xsk.queue_id = pool->queue_id;
 115
 116                 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
 117
 118                 if (err)
 119                         WARN(1, "Failed to disable zero-copy!\n");
 120         }
 121 }
 122
 123 static int __xp_assign_dev(struct xsk_buff_pool *pool,
 124                            struct net_device *netdev, u16 queue_id, u16 flags)
 125 {
 126         bool force_zc, force_copy;
 127         struct netdev_bpf bpf;
 128         int err = 0;
 129
 130         ASSERT_RTNL();
 131
 132         force_zc = flags & XDP_ZEROCOPY;
 133         force_copy = flags & XDP_COPY;
 134
 135         if (force_zc && force_copy)
 136                 return -EINVAL;
 137
 138         if (xsk_get_pool_from_qid(netdev, queue_id))
 139                 return -EBUSY;
 140
 141         pool->netdev = netdev;
 142         pool->queue_id = queue_id;
 143         err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
 144         if (err)
 145                 return err;
 146
 147         if (flags & XDP_USE_NEED_WAKEUP) {
 148                 pool->uses_need_wakeup = true;
 149                 /* Tx needs to be explicitly woken up the first time.
 150                  * Also for supporting drivers that do not implement this
 151                  * feature. They will always have to call sendto().
 152                  */
 153                 pool->cached_need_wakeup = XDP_WAKEUP_TX;
 154         }
 155
 156         dev_hold(netdev);
 157
 158         if (force_copy)
 159                 /* For copy-mode, we are done. */
 160                 return 0;
 161
 162         if (!netdev->netdev_ops->ndo_bpf ||
 163             !netdev->netdev_ops->ndo_xsk_wakeup) {
 164                 err = -EOPNOTSUPP;
 165                 goto err_unreg_pool;
 166         }
 167
 168         bpf.command = XDP_SETUP_XSK_POOL;
 169         bpf.xsk.pool = pool;
 170         bpf.xsk.queue_id = queue_id;
 171
 172         err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
 173         if (err)
 174                 goto err_unreg_pool;
 175
 176         if (!pool->dma_pages) {
 177                 WARN(1, "Driver did not DMA map zero-copy buffers");
 178                 err = -EINVAL;
 179                 goto err_unreg_xsk;
 180         }
 181         pool->umem->zc = true;
 182         return 0;
 183
 184 err_unreg_xsk:
 185         xp_disable_drv_zc(pool);
 186 err_unreg_pool:
 187         if (!force_zc)
 188                 err = 0; /* fallback to copy mode */
 189         if (err) {
 190                 xsk_clear_pool_at_qid(netdev, queue_id);
 191                 dev_put(netdev);
 192         }
 193         return err;
 194 }
 195
 196 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
 197                   u16 queue_id, u16 flags)
 198 {
 199         return __xp_assign_dev(pool, dev, queue_id, flags);
 200 }
 201
 202 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_umem *umem,
 203                          struct net_device *dev, u16 queue_id)
 204 {
 205         u16 flags;
 206
 207         /* One fill and completion ring required for each queue id. */
 208         if (!pool->fq || !pool->cq)
 209                 return -EINVAL;
 210
 211         flags = umem->zc ? XDP_ZEROCOPY : XDP_COPY;
 212         if (pool->uses_need_wakeup)
 213                 flags |= XDP_USE_NEED_WAKEUP;
 214
 215         return __xp_assign_dev(pool, dev, queue_id, flags);
 216 }
 217
 218 void xp_clear_dev(struct xsk_buff_pool *pool)
 219 {
 220         if (!pool->netdev)
 221                 return;
 222
 223         xp_disable_drv_zc(pool);
 224         xsk_clear_pool_at_qid(pool->netdev, pool->queue_id);
 225         dev_put(pool->netdev);
 226         pool->netdev = NULL;
 227 }
 228
 229 static void xp_release_deferred(struct work_struct *work)
 230 {
 231         struct xsk_buff_pool *pool = container_of(work, struct xsk_buff_pool,
 232                                                   work);
 233
 234         rtnl_lock();
 235         xp_clear_dev(pool);
 236         rtnl_unlock();
 237
 238         if (pool->fq) {
 239                 xskq_destroy(pool->fq);
 240                 pool->fq = NULL;
 241         }
 242
 243         if (pool->cq) {
 244                 xskq_destroy(pool->cq);
 245                 pool->cq = NULL;
 246         }
 247
 248         xdp_put_umem(pool->umem, false);
 249         xp_destroy(pool);
 250 }
 251
 252 void xp_get_pool(struct xsk_buff_pool *pool)
 253 {
 254         refcount_inc(&pool->users);
 255 }
 256
 257 bool xp_put_pool(struct xsk_buff_pool *pool)
 258 {
 259         if (!pool)
 260                 return false;
 261
 262         if (refcount_dec_and_test(&pool->users)) {
 263                 INIT_WORK(&pool->work, xp_release_deferred);
 264                 schedule_work(&pool->work);
 265                 return true;
 266         }
 267
 268         return false;
 269 }
 270
 271 static struct xsk_dma_map *xp_find_dma_map(struct xsk_buff_pool *pool)
 272 {
 273         struct xsk_dma_map *dma_map;
 274
 275         list_for_each_entry(dma_map, &pool->umem->xsk_dma_list, list) {
 276                 if (dma_map->netdev == pool->netdev)
 277                         return dma_map;
 278         }
 279
 280         return NULL;
 281 }
 282
 283 static struct xsk_dma_map *xp_create_dma_map(struct device *dev, struct net_device *netdev,
 284                                              u32 nr_pages, struct xdp_umem *umem)
 285 {
 286         struct xsk_dma_map *dma_map;
 287
 288         dma_map = kzalloc(sizeof(*dma_map), GFP_KERNEL);
 289         if (!dma_map)
 290                 return NULL;
 291
 292         dma_map->dma_pages = kvcalloc(nr_pages, sizeof(*dma_map->dma_pages), GFP_KERNEL);
 293         if (!dma_map->dma_pages) {
 294                 kfree(dma_map);
 295                 return NULL;
 296         }
 297
 298         dma_map->netdev = netdev;
 299         dma_map->dev = dev;
 300         dma_map->dma_need_sync = false;
 301         dma_map->dma_pages_cnt = nr_pages;
 302         refcount_set(&dma_map->users, 1);
 303         list_add(&dma_map->list, &umem->xsk_dma_list);
 304         return dma_map;
 305 }
 306
 307 static void xp_destroy_dma_map(struct xsk_dma_map *dma_map)
 308 {
 309         list_del(&dma_map->list);
 310         kvfree(dma_map->dma_pages);
 311         kfree(dma_map);
 312 }
 313
 314 static void __xp_dma_unmap(struct xsk_dma_map *dma_map, unsigned long attrs)
 315 {
 316         dma_addr_t *dma;
 317         u32 i;
 318
 319         for (i = 0; i < dma_map->dma_pages_cnt; i++) {
 320                 dma = &dma_map->dma_pages[i];
 321                 if (*dma) {
 322                         dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
 323                                              DMA_BIDIRECTIONAL, attrs);
 324                         *dma = 0;
 325                 }
 326         }
 327
 328         xp_destroy_dma_map(dma_map);
 329 }
 330
 331 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
 332 {
 333         struct xsk_dma_map *dma_map;
 334
 335         if (pool->dma_pages_cnt == 0)
 336                 return;
 337
 338         dma_map = xp_find_dma_map(pool);
 339         if (!dma_map) {
 340                 WARN(1, "Could not find dma_map for device");
 341                 return;
 342         }
 343
 344         if (!refcount_dec_and_test(&dma_map->users))
 345                 return;
 346
 347         __xp_dma_unmap(dma_map, attrs);
 348         kvfree(pool->dma_pages);
 349         pool->dma_pages_cnt = 0;
 350         pool->dev = NULL;
 351 }
 352 EXPORT_SYMBOL(xp_dma_unmap);
 353
 354 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
 355 {
 356         u32 i;
 357
 358         for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
 359                 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
 360                         dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
 361                 else
 362                         dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
 363         }
 364 }
 365
 366 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
 367 {
 368         pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
 369         if (!pool->dma_pages)
 370                 return -ENOMEM;
 371
 372         pool->dev = dma_map->dev;
 373         pool->dma_pages_cnt = dma_map->dma_pages_cnt;
 374         pool->dma_need_sync = dma_map->dma_need_sync;
 375         memcpy(pool->dma_pages, dma_map->dma_pages,
 376                pool->dma_pages_cnt * sizeof(*pool->dma_pages));
 377
 378         return 0;
 379 }
 380
 381 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
 382                unsigned long attrs, struct page **pages, u32 nr_pages)
 383 {
 384         struct xsk_dma_map *dma_map;
 385         dma_addr_t dma;
 386         int err;
 387         u32 i;
 388
 389         dma_map = xp_find_dma_map(pool);
 390         if (dma_map) {
 391                 err = xp_init_dma_info(pool, dma_map);
 392                 if (err)
 393                         return err;
 394
 395                 refcount_inc(&dma_map->users);
 396                 return 0;
 397         }
 398
 399         dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
 400         if (!dma_map)
 401                 return -ENOMEM;
 402
 403         for (i = 0; i < dma_map->dma_pages_cnt; i++) {
 404                 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
 405                                          DMA_BIDIRECTIONAL, attrs);
 406                 if (dma_mapping_error(dev, dma)) {
 407                         __xp_dma_unmap(dma_map, attrs);
 408                         return -ENOMEM;
 409                 }
 410                 if (dma_need_sync(dev, dma))
 411                         dma_map->dma_need_sync = true;
 412                 dma_map->dma_pages[i] = dma;
 413         }
 414
 415         if (pool->unaligned)
 416                 xp_check_dma_contiguity(dma_map);
 417
 418         err = xp_init_dma_info(pool, dma_map);
 419         if (err) {
 420                 __xp_dma_unmap(dma_map, attrs);
 421                 return err;
 422         }
 423
 424         return 0;
 425 }
 426 EXPORT_SYMBOL(xp_dma_map);
 427
 428 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
 429                                           u64 addr)
 430 {
 431         return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
 432 }
 433
 434 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
 435 {
 436         *addr = xp_unaligned_extract_addr(*addr);
 437         if (*addr >= pool->addrs_cnt ||
 438             *addr + pool->chunk_size > pool->addrs_cnt ||
 439             xp_addr_crosses_non_contig_pg(pool, *addr))
 440                 return false;
 441         return true;
 442 }
 443
 444 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
 445 {
 446         *addr = xp_aligned_extract_addr(pool, *addr);
 447         return *addr < pool->addrs_cnt;
 448 }
 449
 450 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
 451 {
 452         struct xdp_buff_xsk *xskb;
 453         u64 addr;
 454         bool ok;
 455
 456         if (pool->free_heads_cnt == 0)
 457                 return NULL;
 458
 459         xskb = pool->free_heads[--pool->free_heads_cnt];
 460
 461         for (;;) {
 462                 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
 463                         pool->fq->queue_empty_descs++;
 464                         xp_release(xskb);
 465                         return NULL;
 466                 }
 467
 468                 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
 469                      xp_check_aligned(pool, &addr);
 470                 if (!ok) {
 471                         pool->fq->invalid_descs++;
 472                         xskq_cons_release(pool->fq);
 473                         continue;
 474                 }
 475                 break;
 476         }
 477         xskq_cons_release(pool->fq);
 478
 479         xskb->orig_addr = addr;
 480         xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
 481         if (pool->dma_pages_cnt) {
 482                 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
 483                                    ~XSK_NEXT_PG_CONTIG_MASK) +
 484                                   (addr & ~PAGE_MASK);
 485                 xskb->dma = xskb->frame_dma + pool->headroom +
 486                             XDP_PACKET_HEADROOM;
 487         }
 488         return xskb;
 489 }
 490
 491 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
 492 {
 493         struct xdp_buff_xsk *xskb;
 494
 495         if (!pool->free_list_cnt) {
 496                 xskb = __xp_alloc(pool);
 497                 if (!xskb)
 498                         return NULL;
 499         } else {
 500                 pool->free_list_cnt--;
 501                 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
 502                                         free_list_node);
 503                 list_del(&xskb->free_list_node);
 504         }
 505
 506         xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
 507         xskb->xdp.data_meta = xskb->xdp.data;
 508
 509         if (pool->dma_need_sync) {
 510                 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
 511                                                  pool->frame_len,
 512                                                  DMA_BIDIRECTIONAL);
 513         }
 514         return &xskb->xdp;
 515 }
 516 EXPORT_SYMBOL(xp_alloc);
 517
 518 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
 519 {
 520         if (pool->free_list_cnt >= count)
 521                 return true;
 522         return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
 523 }
 524 EXPORT_SYMBOL(xp_can_alloc);
 525
 526 void xp_free(struct xdp_buff_xsk *xskb)
 527 {
 528         xskb->pool->free_list_cnt++;
 529         list_add(&xskb->free_list_node, &xskb->pool->free_list);
 530 }
 531 EXPORT_SYMBOL(xp_free);
 532
 533 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 534 {
 535         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 536         return pool->addrs + addr;
 537 }
 538 EXPORT_SYMBOL(xp_raw_get_data);
 539
 540 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
 541 {
 542         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 543         return (pool->dma_pages[addr >> PAGE_SHIFT] &
 544                 ~XSK_NEXT_PG_CONTIG_MASK) +
 545                 (addr & ~PAGE_MASK);
 546 }
 547 EXPORT_SYMBOL(xp_raw_get_dma);
 548
 549 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
 550 {
 551         dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
 552                                       xskb->pool->frame_len, DMA_BIDIRECTIONAL);
 553 }
 554 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
 555
 556 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
 557                                  size_t size)
 558 {
 559         dma_sync_single_range_for_device(pool->dev, dma, 0,
 560                                          size, DMA_BIDIRECTIONAL);
 561 }
 562 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);