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         spin_lock_init(&pool->cq_lock);
  75         refcount_set(&pool->users, 1);
  76
  77         pool->fq = xs->fq_tmp;
  78         pool->cq = xs->cq_tmp;
  79
  80         for (i = 0; i < pool->free_heads_cnt; i++) {
  81                 xskb = &pool->heads[i];
  82                 xskb->pool = pool;
  83                 xskb->xdp.frame_sz = umem->chunk_size - umem->headroom;
  84                 pool->free_heads[i] = xskb;
  85         }
  86
  87         return pool;
  88
  89 out:
  90         xp_destroy(pool);
  91         return NULL;
  92 }
  93
  94 void xp_set_rxq_info(struct xsk_buff_pool *pool, struct xdp_rxq_info *rxq)
  95 {
  96         u32 i;
  97
  98         for (i = 0; i < pool->heads_cnt; i++)
  99                 pool->heads[i].xdp.rxq = rxq;
 100 }
 101 EXPORT_SYMBOL(xp_set_rxq_info);
 102
 103 static void xp_disable_drv_zc(struct xsk_buff_pool *pool)
 104 {
 105         struct netdev_bpf bpf;
 106         int err;
 107
 108         ASSERT_RTNL();
 109
 110         if (pool->umem->zc) {
 111                 bpf.command = XDP_SETUP_XSK_POOL;
 112                 bpf.xsk.pool = NULL;
 113                 bpf.xsk.queue_id = pool->queue_id;
 114
 115                 err = pool->netdev->netdev_ops->ndo_bpf(pool->netdev, &bpf);
 116
 117                 if (err)
 118                         WARN(1, "Failed to disable zero-copy!\n");
 119         }
 120 }
 121
 122 static int __xp_assign_dev(struct xsk_buff_pool *pool,
 123                            struct net_device *netdev, u16 queue_id, u16 flags)
 124 {
 125         bool force_zc, force_copy;
 126         struct netdev_bpf bpf;
 127         int err = 0;
 128
 129         ASSERT_RTNL();
 130
 131         force_zc = flags & XDP_ZEROCOPY;
 132         force_copy = flags & XDP_COPY;
 133
 134         if (force_zc && force_copy)
 135                 return -EINVAL;
 136
 137         if (xsk_get_pool_from_qid(netdev, queue_id))
 138                 return -EBUSY;
 139
 140         pool->netdev = netdev;
 141         pool->queue_id = queue_id;
 142         err = xsk_reg_pool_at_qid(netdev, pool, queue_id);
 143         if (err)
 144                 return err;
 145
 146         if (flags & XDP_USE_NEED_WAKEUP) {
 147                 pool->uses_need_wakeup = true;
 148                 /* Tx needs to be explicitly woken up the first time.
 149                  * Also for supporting drivers that do not implement this
 150                  * feature. They will always have to call sendto().
 151                  */
 152                 pool->cached_need_wakeup = XDP_WAKEUP_TX;
 153         }
 154
 155         dev_hold(netdev);
 156
 157         if (force_copy)
 158                 /* For copy-mode, we are done. */
 159                 return 0;
 160
 161         if (!netdev->netdev_ops->ndo_bpf ||
 162             !netdev->netdev_ops->ndo_xsk_wakeup) {
 163                 err = -EOPNOTSUPP;
 164                 goto err_unreg_pool;
 165         }
 166
 167         bpf.command = XDP_SETUP_XSK_POOL;
 168         bpf.xsk.pool = pool;
 169         bpf.xsk.queue_id = queue_id;
 170
 171         err = netdev->netdev_ops->ndo_bpf(netdev, &bpf);
 172         if (err)
 173                 goto err_unreg_pool;
 174
 175         if (!pool->dma_pages) {
 176                 WARN(1, "Driver did not DMA map zero-copy buffers");
 177                 err = -EINVAL;
 178                 goto err_unreg_xsk;
 179         }
 180         pool->umem->zc = true;
 181         return 0;
 182
 183 err_unreg_xsk:
 184         xp_disable_drv_zc(pool);
 185 err_unreg_pool:
 186         if (!force_zc)
 187                 err = 0; /* fallback to copy mode */
 188         if (err) {
 189                 xsk_clear_pool_at_qid(netdev, queue_id);
 190                 dev_put(netdev);
 191         }
 192         return err;
 193 }
 194
 195 int xp_assign_dev(struct xsk_buff_pool *pool, struct net_device *dev,
 196                   u16 queue_id, u16 flags)
 197 {
 198         return __xp_assign_dev(pool, dev, queue_id, flags);
 199 }
 200
 201 int xp_assign_dev_shared(struct xsk_buff_pool *pool, struct xdp_sock *umem_xs,
 202                          struct net_device *dev, u16 queue_id)
 203 {
 204         u16 flags;
 205         struct xdp_umem *umem = umem_xs->umem;
 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 (umem_xs->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 &= ~XSK_NEXT_PG_CONTIG_MASK;
 323                         dma_unmap_page_attrs(dma_map->dev, *dma, PAGE_SIZE,
 324                                              DMA_BIDIRECTIONAL, attrs);
 325                         *dma = 0;
 326                 }
 327         }
 328
 329         xp_destroy_dma_map(dma_map);
 330 }
 331
 332 void xp_dma_unmap(struct xsk_buff_pool *pool, unsigned long attrs)
 333 {
 334         struct xsk_dma_map *dma_map;
 335
 336         if (pool->dma_pages_cnt == 0)
 337                 return;
 338
 339         dma_map = xp_find_dma_map(pool);
 340         if (!dma_map) {
 341                 WARN(1, "Could not find dma_map for device");
 342                 return;
 343         }
 344
 345         if (!refcount_dec_and_test(&dma_map->users))
 346                 return;
 347
 348         __xp_dma_unmap(dma_map, attrs);
 349         kvfree(pool->dma_pages);
 350         pool->dma_pages_cnt = 0;
 351         pool->dev = NULL;
 352 }
 353 EXPORT_SYMBOL(xp_dma_unmap);
 354
 355 static void xp_check_dma_contiguity(struct xsk_dma_map *dma_map)
 356 {
 357         u32 i;
 358
 359         for (i = 0; i < dma_map->dma_pages_cnt - 1; i++) {
 360                 if (dma_map->dma_pages[i] + PAGE_SIZE == dma_map->dma_pages[i + 1])
 361                         dma_map->dma_pages[i] |= XSK_NEXT_PG_CONTIG_MASK;
 362                 else
 363                         dma_map->dma_pages[i] &= ~XSK_NEXT_PG_CONTIG_MASK;
 364         }
 365 }
 366
 367 static int xp_init_dma_info(struct xsk_buff_pool *pool, struct xsk_dma_map *dma_map)
 368 {
 369         pool->dma_pages = kvcalloc(dma_map->dma_pages_cnt, sizeof(*pool->dma_pages), GFP_KERNEL);
 370         if (!pool->dma_pages)
 371                 return -ENOMEM;
 372
 373         pool->dev = dma_map->dev;
 374         pool->dma_pages_cnt = dma_map->dma_pages_cnt;
 375         pool->dma_need_sync = dma_map->dma_need_sync;
 376         memcpy(pool->dma_pages, dma_map->dma_pages,
 377                pool->dma_pages_cnt * sizeof(*pool->dma_pages));
 378
 379         return 0;
 380 }
 381
 382 int xp_dma_map(struct xsk_buff_pool *pool, struct device *dev,
 383                unsigned long attrs, struct page **pages, u32 nr_pages)
 384 {
 385         struct xsk_dma_map *dma_map;
 386         dma_addr_t dma;
 387         int err;
 388         u32 i;
 389
 390         dma_map = xp_find_dma_map(pool);
 391         if (dma_map) {
 392                 err = xp_init_dma_info(pool, dma_map);
 393                 if (err)
 394                         return err;
 395
 396                 refcount_inc(&dma_map->users);
 397                 return 0;
 398         }
 399
 400         dma_map = xp_create_dma_map(dev, pool->netdev, nr_pages, pool->umem);
 401         if (!dma_map)
 402                 return -ENOMEM;
 403
 404         for (i = 0; i < dma_map->dma_pages_cnt; i++) {
 405                 dma = dma_map_page_attrs(dev, pages[i], 0, PAGE_SIZE,
 406                                          DMA_BIDIRECTIONAL, attrs);
 407                 if (dma_mapping_error(dev, dma)) {
 408                         __xp_dma_unmap(dma_map, attrs);
 409                         return -ENOMEM;
 410                 }
 411                 if (dma_need_sync(dev, dma))
 412                         dma_map->dma_need_sync = true;
 413                 dma_map->dma_pages[i] = dma;
 414         }
 415
 416         if (pool->unaligned)
 417                 xp_check_dma_contiguity(dma_map);
 418
 419         err = xp_init_dma_info(pool, dma_map);
 420         if (err) {
 421                 __xp_dma_unmap(dma_map, attrs);
 422                 return err;
 423         }
 424
 425         return 0;
 426 }
 427 EXPORT_SYMBOL(xp_dma_map);
 428
 429 static bool xp_addr_crosses_non_contig_pg(struct xsk_buff_pool *pool,
 430                                           u64 addr)
 431 {
 432         return xp_desc_crosses_non_contig_pg(pool, addr, pool->chunk_size);
 433 }
 434
 435 static bool xp_check_unaligned(struct xsk_buff_pool *pool, u64 *addr)
 436 {
 437         *addr = xp_unaligned_extract_addr(*addr);
 438         if (*addr >= pool->addrs_cnt ||
 439             *addr + pool->chunk_size > pool->addrs_cnt ||
 440             xp_addr_crosses_non_contig_pg(pool, *addr))
 441                 return false;
 442         return true;
 443 }
 444
 445 static bool xp_check_aligned(struct xsk_buff_pool *pool, u64 *addr)
 446 {
 447         *addr = xp_aligned_extract_addr(pool, *addr);
 448         return *addr < pool->addrs_cnt;
 449 }
 450
 451 static struct xdp_buff_xsk *__xp_alloc(struct xsk_buff_pool *pool)
 452 {
 453         struct xdp_buff_xsk *xskb;
 454         u64 addr;
 455         bool ok;
 456
 457         if (pool->free_heads_cnt == 0)
 458                 return NULL;
 459
 460         xskb = pool->free_heads[--pool->free_heads_cnt];
 461
 462         for (;;) {
 463                 if (!xskq_cons_peek_addr_unchecked(pool->fq, &addr)) {
 464                         pool->fq->queue_empty_descs++;
 465                         xp_release(xskb);
 466                         return NULL;
 467                 }
 468
 469                 ok = pool->unaligned ? xp_check_unaligned(pool, &addr) :
 470                      xp_check_aligned(pool, &addr);
 471                 if (!ok) {
 472                         pool->fq->invalid_descs++;
 473                         xskq_cons_release(pool->fq);
 474                         continue;
 475                 }
 476                 break;
 477         }
 478         xskq_cons_release(pool->fq);
 479
 480         xskb->orig_addr = addr;
 481         xskb->xdp.data_hard_start = pool->addrs + addr + pool->headroom;
 482         if (pool->dma_pages_cnt) {
 483                 xskb->frame_dma = (pool->dma_pages[addr >> PAGE_SHIFT] &
 484                                    ~XSK_NEXT_PG_CONTIG_MASK) +
 485                                   (addr & ~PAGE_MASK);
 486                 xskb->dma = xskb->frame_dma + pool->headroom +
 487                             XDP_PACKET_HEADROOM;
 488         }
 489         return xskb;
 490 }
 491
 492 struct xdp_buff *xp_alloc(struct xsk_buff_pool *pool)
 493 {
 494         struct xdp_buff_xsk *xskb;
 495
 496         if (!pool->free_list_cnt) {
 497                 xskb = __xp_alloc(pool);
 498                 if (!xskb)
 499                         return NULL;
 500         } else {
 501                 pool->free_list_cnt--;
 502                 xskb = list_first_entry(&pool->free_list, struct xdp_buff_xsk,
 503                                         free_list_node);
 504                 list_del(&xskb->free_list_node);
 505         }
 506
 507         xskb->xdp.data = xskb->xdp.data_hard_start + XDP_PACKET_HEADROOM;
 508         xskb->xdp.data_meta = xskb->xdp.data;
 509
 510         if (pool->dma_need_sync) {
 511                 dma_sync_single_range_for_device(pool->dev, xskb->dma, 0,
 512                                                  pool->frame_len,
 513                                                  DMA_BIDIRECTIONAL);
 514         }
 515         return &xskb->xdp;
 516 }
 517 EXPORT_SYMBOL(xp_alloc);
 518
 519 bool xp_can_alloc(struct xsk_buff_pool *pool, u32 count)
 520 {
 521         if (pool->free_list_cnt >= count)
 522                 return true;
 523         return xskq_cons_has_entries(pool->fq, count - pool->free_list_cnt);
 524 }
 525 EXPORT_SYMBOL(xp_can_alloc);
 526
 527 void xp_free(struct xdp_buff_xsk *xskb)
 528 {
 529         xskb->pool->free_list_cnt++;
 530         list_add(&xskb->free_list_node, &xskb->pool->free_list);
 531 }
 532 EXPORT_SYMBOL(xp_free);
 533
 534 void *xp_raw_get_data(struct xsk_buff_pool *pool, u64 addr)
 535 {
 536         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 537         return pool->addrs + addr;
 538 }
 539 EXPORT_SYMBOL(xp_raw_get_data);
 540
 541 dma_addr_t xp_raw_get_dma(struct xsk_buff_pool *pool, u64 addr)
 542 {
 543         addr = pool->unaligned ? xp_unaligned_add_offset_to_addr(addr) : addr;
 544         return (pool->dma_pages[addr >> PAGE_SHIFT] &
 545                 ~XSK_NEXT_PG_CONTIG_MASK) +
 546                 (addr & ~PAGE_MASK);
 547 }
 548 EXPORT_SYMBOL(xp_raw_get_dma);
 549
 550 void xp_dma_sync_for_cpu_slow(struct xdp_buff_xsk *xskb)
 551 {
 552         dma_sync_single_range_for_cpu(xskb->pool->dev, xskb->dma, 0,
 553                                       xskb->pool->frame_len, DMA_BIDIRECTIONAL);
 554 }
 555 EXPORT_SYMBOL(xp_dma_sync_for_cpu_slow);
 556
 557 void xp_dma_sync_for_device_slow(struct xsk_buff_pool *pool, dma_addr_t dma,
 558                                  size_t size)
 559 {
 560         dma_sync_single_range_for_device(pool->dev, dma, 0,
 561                                          size, DMA_BIDIRECTIONAL);
 562 }
 563 EXPORT_SYMBOL(xp_dma_sync_for_device_slow);