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