kernel/bpf/devmap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
   3  */
   4
   5 /* Devmaps primary use is as a backend map for XDP BPF helper call
   6  * bpf_redirect_map(). Because XDP is mostly concerned with performance we
   7  * spent some effort to ensure the datapath with redirect maps does not use
   8  * any locking. This is a quick note on the details.
   9  *
  10  * We have three possible paths to get into the devmap control plane bpf
  11  * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
  12  * will invoke an update, delete, or lookup operation. To ensure updates and
  13  * deletes appear atomic from the datapath side xchg() is used to modify the
  14  * netdev_map array. Then because the datapath does a lookup into the netdev_map
  15  * array (read-only) from an RCU critical section we use call_rcu() to wait for
  16  * an rcu grace period before free'ing the old data structures. This ensures the
  17  * datapath always has a valid copy. However, the datapath does a "flush"
  18  * operation that pushes any pending packets in the driver outside the RCU
  19  * critical section. Each bpf_dtab_netdev tracks these pending operations using
  20  * a per-cpu flush list. The bpf_dtab_netdev object will not be destroyed  until
  21  * this list is empty, indicating outstanding flush operations have completed.
  22  *
  23  * BPF syscalls may race with BPF program calls on any of the update, delete
  24  * or lookup operations. As noted above the xchg() operation also keep the
  25  * netdev_map consistent in this case. From the devmap side BPF programs
  26  * calling into these operations are the same as multiple user space threads
  27  * making system calls.
  28  *
  29  * Finally, any of the above may race with a netdev_unregister notifier. The
  30  * unregister notifier must search for net devices in the map structure that
  31  * contain a reference to the net device and remove them. This is a two step
  32  * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
  33  * check to see if the ifindex is the same as the net_device being removed.
  34  * When removing the dev a cmpxchg() is used to ensure the correct dev is
  35  * removed, in the case of a concurrent update or delete operation it is
  36  * possible that the initially referenced dev is no longer in the map. As the
  37  * notifier hook walks the map we know that new dev references can not be
  38  * added by the user because core infrastructure ensures dev_get_by_index()
  39  * calls will fail at this point.
  40  *
  41  * The devmap_hash type is a map type which interprets keys as ifindexes and
  42  * indexes these using a hashmap. This allows maps that use ifindex as key to be
  43  * densely packed instead of having holes in the lookup array for unused
  44  * ifindexes. The setup and packet enqueue/send code is shared between the two
  45  * types of devmap; only the lookup and insertion is different.
  46  */
  47 #include <linux/bpf.h>
  48 #include <net/xdp.h>
  49 #include <linux/filter.h>
  50 #include <trace/events/xdp.h>
  51 #include <linux/btf_ids.h>
  52
  53 #define DEV_CREATE_FLAG_MASK \
  54         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  55
  56 struct xdp_dev_bulk_queue {
  57         struct xdp_frame *q[DEV_MAP_BULK_SIZE];
  58         struct list_head flush_node;
  59         struct net_device *dev;
  60         struct net_device *dev_rx;
  61         struct bpf_prog *xdp_prog;
  62         unsigned int count;
  63 };
  64
  65 struct bpf_dtab_netdev {
  66         struct net_device *dev; /* must be first member, due to tracepoint */
  67         struct hlist_node index_hlist;
  68         struct bpf_dtab *dtab;
  69         struct bpf_prog *xdp_prog;
  70         struct rcu_head rcu;
  71         unsigned int idx;
  72         struct bpf_devmap_val val;
  73 };
  74
  75 struct bpf_dtab {
  76         struct bpf_map map;
  77         struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
  78         struct list_head list;
  79
  80         /* these are only used for DEVMAP_HASH type maps */
  81         struct hlist_head *dev_index_head;
  82         spinlock_t index_lock;
  83         unsigned int items;
  84         u32 n_buckets;
  85 };
  86
  87 static DEFINE_PER_CPU(struct list_head, dev_flush_list);
  88 static DEFINE_SPINLOCK(dev_map_lock);
  89 static LIST_HEAD(dev_map_list);
  90
  91 static struct hlist_head *dev_map_create_hash(unsigned int entries,
  92                                               int numa_node)
  93 {
  94         int i;
  95         struct hlist_head *hash;
  96
  97         hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
  98         if (hash != NULL)
  99                 for (i = 0; i < entries; i++)
 100                         INIT_HLIST_HEAD(&hash[i]);
 101
 102         return hash;
 103 }
 104
 105 static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
 106                                                     int idx)
 107 {
 108         return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
 109 }
 110
 111 static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
 112 {
 113         u32 valsize = attr->value_size;
 114
 115         /* check sanity of attributes. 2 value sizes supported:
 116          * 4 bytes: ifindex
 117          * 8 bytes: ifindex + prog fd
 118          */
 119         if (attr->max_entries == 0 || attr->key_size != 4 ||
 120             (valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
 121              valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
 122             attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 123                 return -EINVAL;
 124
 125         /* Lookup returns a pointer straight to dev->ifindex, so make sure the
 126          * verifier prevents writes from the BPF side
 127          */
 128         attr->map_flags |= BPF_F_RDONLY_PROG;
 129
 130
 131         bpf_map_init_from_attr(&dtab->map, attr);
 132
 133         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 134                 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
 135
 136                 if (!dtab->n_buckets) /* Overflow check */
 137                         return -EINVAL;
 138         }
 139
 140         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 141                 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
 142                                                            dtab->map.numa_node);
 143                 if (!dtab->dev_index_head)
 144                         return -ENOMEM;
 145
 146                 spin_lock_init(&dtab->index_lock);
 147         } else {
 148                 dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
 149                                                       sizeof(struct bpf_dtab_netdev *),
 150                                                       dtab->map.numa_node);
 151                 if (!dtab->netdev_map)
 152                         return -ENOMEM;
 153         }
 154
 155         return 0;
 156 }
 157
 158 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 159 {
 160         struct bpf_dtab *dtab;
 161         int err;
 162
 163         if (!capable(CAP_NET_ADMIN))
 164                 return ERR_PTR(-EPERM);
 165
 166         dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
 167         if (!dtab)
 168                 return ERR_PTR(-ENOMEM);
 169
 170         err = dev_map_init_map(dtab, attr);
 171         if (err) {
 172                 bpf_map_area_free(dtab);
 173                 return ERR_PTR(err);
 174         }
 175
 176         spin_lock(&dev_map_lock);
 177         list_add_tail_rcu(&dtab->list, &dev_map_list);
 178         spin_unlock(&dev_map_lock);
 179
 180         return &dtab->map;
 181 }
 182
 183 static void dev_map_free(struct bpf_map *map)
 184 {
 185         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 186         int i;
 187
 188         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 189          * so the programs (can be more than one that used this map) were
 190          * disconnected from events. The following synchronize_rcu() guarantees
 191          * both rcu read critical sections complete and waits for
 192          * preempt-disable regions (NAPI being the relevant context here) so we
 193          * are certain there will be no further reads against the netdev_map and
 194          * all flush operations are complete. Flush operations can only be done
 195          * from NAPI context for this reason.
 196          */
 197
 198         spin_lock(&dev_map_lock);
 199         list_del_rcu(&dtab->list);
 200         spin_unlock(&dev_map_lock);
 201
 202         bpf_clear_redirect_map(map);
 203         synchronize_rcu();
 204
 205         /* Make sure prior __dev_map_entry_free() have completed. */
 206         rcu_barrier();
 207
 208         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 209                 for (i = 0; i < dtab->n_buckets; i++) {
 210                         struct bpf_dtab_netdev *dev;
 211                         struct hlist_head *head;
 212                         struct hlist_node *next;
 213
 214                         head = dev_map_index_hash(dtab, i);
 215
 216                         hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 217                                 hlist_del_rcu(&dev->index_hlist);
 218                                 if (dev->xdp_prog)
 219                                         bpf_prog_put(dev->xdp_prog);
 220                                 dev_put(dev->dev);
 221                                 kfree(dev);
 222                         }
 223                 }
 224
 225                 bpf_map_area_free(dtab->dev_index_head);
 226         } else {
 227                 for (i = 0; i < dtab->map.max_entries; i++) {
 228                         struct bpf_dtab_netdev *dev;
 229
 230                         dev = rcu_dereference_raw(dtab->netdev_map[i]);
 231                         if (!dev)
 232                                 continue;
 233
 234                         if (dev->xdp_prog)
 235                                 bpf_prog_put(dev->xdp_prog);
 236                         dev_put(dev->dev);
 237                         kfree(dev);
 238                 }
 239
 240                 bpf_map_area_free(dtab->netdev_map);
 241         }
 242
 243         bpf_map_area_free(dtab);
 244 }
 245
 246 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 247 {
 248         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 249         u32 index = key ? *(u32 *)key : U32_MAX;
 250         u32 *next = next_key;
 251
 252         if (index >= dtab->map.max_entries) {
 253                 *next = 0;
 254                 return 0;
 255         }
 256
 257         if (index == dtab->map.max_entries - 1)
 258                 return -ENOENT;
 259         *next = index + 1;
 260         return 0;
 261 }
 262
 263 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 264  * by local_bh_disable() (from XDP calls inside NAPI). The
 265  * rcu_read_lock_bh_held() below makes lockdep accept both.
 266  */
 267 static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 268 {
 269         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 270         struct hlist_head *head = dev_map_index_hash(dtab, key);
 271         struct bpf_dtab_netdev *dev;
 272
 273         hlist_for_each_entry_rcu(dev, head, index_hlist,
 274                                  lockdep_is_held(&dtab->index_lock))
 275                 if (dev->idx == key)
 276                         return dev;
 277
 278         return NULL;
 279 }
 280
 281 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 282                                     void *next_key)
 283 {
 284         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 285         u32 idx, *next = next_key;
 286         struct bpf_dtab_netdev *dev, *next_dev;
 287         struct hlist_head *head;
 288         int i = 0;
 289
 290         if (!key)
 291                 goto find_first;
 292
 293         idx = *(u32 *)key;
 294
 295         dev = __dev_map_hash_lookup_elem(map, idx);
 296         if (!dev)
 297                 goto find_first;
 298
 299         next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 300                                     struct bpf_dtab_netdev, index_hlist);
 301
 302         if (next_dev) {
 303                 *next = next_dev->idx;
 304                 return 0;
 305         }
 306
 307         i = idx & (dtab->n_buckets - 1);
 308         i++;
 309
 310  find_first:
 311         for (; i < dtab->n_buckets; i++) {
 312                 head = dev_map_index_hash(dtab, i);
 313
 314                 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 315                                             struct bpf_dtab_netdev,
 316                                             index_hlist);
 317                 if (next_dev) {
 318                         *next = next_dev->idx;
 319                         return 0;
 320                 }
 321         }
 322
 323         return -ENOENT;
 324 }
 325
 326 static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
 327                                 struct xdp_frame **frames, int n,
 328                                 struct net_device *dev)
 329 {
 330         struct xdp_txq_info txq = { .dev = dev };
 331         struct xdp_buff xdp;
 332         int i, nframes = 0;
 333
 334         for (i = 0; i < n; i++) {
 335                 struct xdp_frame *xdpf = frames[i];
 336                 u32 act;
 337                 int err;
 338
 339                 xdp_convert_frame_to_buff(xdpf, &xdp);
 340                 xdp.txq = &txq;
 341
 342                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 343                 switch (act) {
 344                 case XDP_PASS:
 345                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 346                         if (unlikely(err < 0))
 347                                 xdp_return_frame_rx_napi(xdpf);
 348                         else
 349                                 frames[nframes++] = xdpf;
 350                         break;
 351                 default:
 352                         bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
 353                         fallthrough;
 354                 case XDP_ABORTED:
 355                         trace_xdp_exception(dev, xdp_prog, act);
 356                         fallthrough;
 357                 case XDP_DROP:
 358                         xdp_return_frame_rx_napi(xdpf);
 359                         break;
 360                 }
 361         }
 362         return nframes; /* sent frames count */
 363 }
 364
 365 static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 366 {
 367         struct net_device *dev = bq->dev;
 368         unsigned int cnt = bq->count;
 369         int sent = 0, err = 0;
 370         int to_send = cnt;
 371         int i;
 372
 373         if (unlikely(!cnt))
 374                 return;
 375
 376         for (i = 0; i < cnt; i++) {
 377                 struct xdp_frame *xdpf = bq->q[i];
 378
 379                 prefetch(xdpf);
 380         }
 381
 382         if (bq->xdp_prog) {
 383                 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
 384                 if (!to_send)
 385                         goto out;
 386         }
 387
 388         sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
 389         if (sent < 0) {
 390                 /* If ndo_xdp_xmit fails with an errno, no frames have
 391                  * been xmit'ed.
 392                  */
 393                 err = sent;
 394                 sent = 0;
 395         }
 396
 397         /* If not all frames have been transmitted, it is our
 398          * responsibility to free them
 399          */
 400         for (i = sent; unlikely(i < to_send); i++)
 401                 xdp_return_frame_rx_napi(bq->q[i]);
 402
 403 out:
 404         bq->count = 0;
 405         trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
 406 }
 407
 408 /* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
 409  * driver before returning from its napi->poll() routine. See the comment above
 410  * xdp_do_flush() in filter.c.
 411  */
 412 void __dev_flush(void)
 413 {
 414         struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 415         struct xdp_dev_bulk_queue *bq, *tmp;
 416
 417         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 418                 bq_xmit_all(bq, XDP_XMIT_FLUSH);
 419                 bq->dev_rx = NULL;
 420                 bq->xdp_prog = NULL;
 421                 __list_del_clearprev(&bq->flush_node);
 422         }
 423 }
 424
 425 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 426  * by local_bh_disable() (from XDP calls inside NAPI). The
 427  * rcu_read_lock_bh_held() below makes lockdep accept both.
 428  */
 429 static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 430 {
 431         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 432         struct bpf_dtab_netdev *obj;
 433
 434         if (key >= map->max_entries)
 435                 return NULL;
 436
 437         obj = rcu_dereference_check(dtab->netdev_map[key],
 438                                     rcu_read_lock_bh_held());
 439         return obj;
 440 }
 441
 442 /* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
 443  * variable access, and map elements stick around. See comment above
 444  * xdp_do_flush() in filter.c.
 445  */
 446 static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 447                        struct net_device *dev_rx, struct bpf_prog *xdp_prog)
 448 {
 449         struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 450         struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 451
 452         if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 453                 bq_xmit_all(bq, 0);
 454
 455         /* Ingress dev_rx will be the same for all xdp_frame's in
 456          * bulk_queue, because bq stored per-CPU and must be flushed
 457          * from net_device drivers NAPI func end.
 458          *
 459          * Do the same with xdp_prog and flush_list since these fields
 460          * are only ever modified together.
 461          */
 462         if (!bq->dev_rx) {
 463                 bq->dev_rx = dev_rx;
 464                 bq->xdp_prog = xdp_prog;
 465                 list_add(&bq->flush_node, flush_list);
 466         }
 467
 468         bq->q[bq->count++] = xdpf;
 469 }
 470
 471 static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 472                                 struct net_device *dev_rx,
 473                                 struct bpf_prog *xdp_prog)
 474 {
 475         int err;
 476
 477         if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 478                 return -EOPNOTSUPP;
 479
 480         if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 481                      xdp_frame_has_frags(xdpf)))
 482                 return -EOPNOTSUPP;
 483
 484         err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
 485         if (unlikely(err))
 486                 return err;
 487
 488         bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
 489         return 0;
 490 }
 491
 492 static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
 493 {
 494         struct xdp_txq_info txq = { .dev = dst->dev };
 495         struct xdp_buff xdp;
 496         u32 act;
 497
 498         if (!dst->xdp_prog)
 499                 return XDP_PASS;
 500
 501         __skb_pull(skb, skb->mac_len);
 502         xdp.txq = &txq;
 503
 504         act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
 505         switch (act) {
 506         case XDP_PASS:
 507                 __skb_push(skb, skb->mac_len);
 508                 break;
 509         default:
 510                 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
 511                 fallthrough;
 512         case XDP_ABORTED:
 513                 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
 514                 fallthrough;
 515         case XDP_DROP:
 516                 kfree_skb(skb);
 517                 break;
 518         }
 519
 520         return act;
 521 }
 522
 523 int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 524                     struct net_device *dev_rx)
 525 {
 526         return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
 527 }
 528
 529 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
 530                     struct net_device *dev_rx)
 531 {
 532         struct net_device *dev = dst->dev;
 533
 534         return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
 535 }
 536
 537 static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
 538 {
 539         if (!obj)
 540                 return false;
 541
 542         if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 543                 return false;
 544
 545         if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 546                      xdp_frame_has_frags(xdpf)))
 547                 return false;
 548
 549         if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
 550                 return false;
 551
 552         return true;
 553 }
 554
 555 static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
 556                                  struct net_device *dev_rx,
 557                                  struct xdp_frame *xdpf)
 558 {
 559         struct xdp_frame *nxdpf;
 560
 561         nxdpf = xdpf_clone(xdpf);
 562         if (!nxdpf)
 563                 return -ENOMEM;
 564
 565         bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
 566
 567         return 0;
 568 }
 569
 570 static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
 571 {
 572         while (num_excluded--) {
 573                 if (ifindex == excluded[num_excluded])
 574                         return true;
 575         }
 576         return false;
 577 }
 578
 579 /* Get ifindex of each upper device. 'indexes' must be able to hold at
 580  * least MAX_NEST_DEV elements.
 581  * Returns the number of ifindexes added.
 582  */
 583 static int get_upper_ifindexes(struct net_device *dev, int *indexes)
 584 {
 585         struct net_device *upper;
 586         struct list_head *iter;
 587         int n = 0;
 588
 589         netdev_for_each_upper_dev_rcu(dev, upper, iter) {
 590                 indexes[n++] = upper->ifindex;
 591         }
 592         return n;
 593 }
 594
 595 int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
 596                           struct bpf_map *map, bool exclude_ingress)
 597 {
 598         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 599         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 600         int excluded_devices[1+MAX_NEST_DEV];
 601         struct hlist_head *head;
 602         int num_excluded = 0;
 603         unsigned int i;
 604         int err;
 605
 606         if (exclude_ingress) {
 607                 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
 608                 excluded_devices[num_excluded++] = dev_rx->ifindex;
 609         }
 610
 611         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 612                 for (i = 0; i < map->max_entries; i++) {
 613                         dst = rcu_dereference_check(dtab->netdev_map[i],
 614                                                     rcu_read_lock_bh_held());
 615                         if (!is_valid_dst(dst, xdpf))
 616                                 continue;
 617
 618                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 619                                 continue;
 620
 621                         /* we only need n-1 clones; last_dst enqueued below */
 622                         if (!last_dst) {
 623                                 last_dst = dst;
 624                                 continue;
 625                         }
 626
 627                         err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 628                         if (err)
 629                                 return err;
 630
 631                         last_dst = dst;
 632                 }
 633         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 634                 for (i = 0; i < dtab->n_buckets; i++) {
 635                         head = dev_map_index_hash(dtab, i);
 636                         hlist_for_each_entry_rcu(dst, head, index_hlist,
 637                                                  lockdep_is_held(&dtab->index_lock)) {
 638                                 if (!is_valid_dst(dst, xdpf))
 639                                         continue;
 640
 641                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 642                                                         dst->dev->ifindex))
 643                                         continue;
 644
 645                                 /* we only need n-1 clones; last_dst enqueued below */
 646                                 if (!last_dst) {
 647                                         last_dst = dst;
 648                                         continue;
 649                                 }
 650
 651                                 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 652                                 if (err)
 653                                         return err;
 654
 655                                 last_dst = dst;
 656                         }
 657                 }
 658         }
 659
 660         /* consume the last copy of the frame */
 661         if (last_dst)
 662                 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
 663         else
 664                 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
 665
 666         return 0;
 667 }
 668
 669 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 670                              struct bpf_prog *xdp_prog)
 671 {
 672         int err;
 673
 674         err = xdp_ok_fwd_dev(dst->dev, skb->len);
 675         if (unlikely(err))
 676                 return err;
 677
 678         /* Redirect has already succeeded semantically at this point, so we just
 679          * return 0 even if packet is dropped. Helper below takes care of
 680          * freeing skb.
 681          */
 682         if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
 683                 return 0;
 684
 685         skb->dev = dst->dev;
 686         generic_xdp_tx(skb, xdp_prog);
 687
 688         return 0;
 689 }
 690
 691 static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
 692                                   struct sk_buff *skb,
 693                                   struct bpf_prog *xdp_prog)
 694 {
 695         struct sk_buff *nskb;
 696         int err;
 697
 698         nskb = skb_clone(skb, GFP_ATOMIC);
 699         if (!nskb)
 700                 return -ENOMEM;
 701
 702         err = dev_map_generic_redirect(dst, nskb, xdp_prog);
 703         if (unlikely(err)) {
 704                 consume_skb(nskb);
 705                 return err;
 706         }
 707
 708         return 0;
 709 }
 710
 711 int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
 712                            struct bpf_prog *xdp_prog, struct bpf_map *map,
 713                            bool exclude_ingress)
 714 {
 715         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 716         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 717         int excluded_devices[1+MAX_NEST_DEV];
 718         struct hlist_head *head;
 719         struct hlist_node *next;
 720         int num_excluded = 0;
 721         unsigned int i;
 722         int err;
 723
 724         if (exclude_ingress) {
 725                 num_excluded = get_upper_ifindexes(dev, excluded_devices);
 726                 excluded_devices[num_excluded++] = dev->ifindex;
 727         }
 728
 729         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 730                 for (i = 0; i < map->max_entries; i++) {
 731                         dst = rcu_dereference_check(dtab->netdev_map[i],
 732                                                     rcu_read_lock_bh_held());
 733                         if (!dst)
 734                                 continue;
 735
 736                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 737                                 continue;
 738
 739                         /* we only need n-1 clones; last_dst enqueued below */
 740                         if (!last_dst) {
 741                                 last_dst = dst;
 742                                 continue;
 743                         }
 744
 745                         err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 746                         if (err)
 747                                 return err;
 748
 749                         last_dst = dst;
 750
 751                 }
 752         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 753                 for (i = 0; i < dtab->n_buckets; i++) {
 754                         head = dev_map_index_hash(dtab, i);
 755                         hlist_for_each_entry_safe(dst, next, head, index_hlist) {
 756                                 if (!dst)
 757                                         continue;
 758
 759                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 760                                                         dst->dev->ifindex))
 761                                         continue;
 762
 763                                 /* we only need n-1 clones; last_dst enqueued below */
 764                                 if (!last_dst) {
 765                                         last_dst = dst;
 766                                         continue;
 767                                 }
 768
 769                                 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 770                                 if (err)
 771                                         return err;
 772
 773                                 last_dst = dst;
 774                         }
 775                 }
 776         }
 777
 778         /* consume the first skb and return */
 779         if (last_dst)
 780                 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
 781
 782         /* dtab is empty */
 783         consume_skb(skb);
 784         return 0;
 785 }
 786
 787 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 788 {
 789         struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 790
 791         return obj ? &obj->val : NULL;
 792 }
 793
 794 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 795 {
 796         struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 797                                                                 *(u32 *)key);
 798         return obj ? &obj->val : NULL;
 799 }
 800
 801 static void __dev_map_entry_free(struct rcu_head *rcu)
 802 {
 803         struct bpf_dtab_netdev *dev;
 804
 805         dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 806         if (dev->xdp_prog)
 807                 bpf_prog_put(dev->xdp_prog);
 808         dev_put(dev->dev);
 809         kfree(dev);
 810 }
 811
 812 static int dev_map_delete_elem(struct bpf_map *map, void *key)
 813 {
 814         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 815         struct bpf_dtab_netdev *old_dev;
 816         int k = *(u32 *)key;
 817
 818         if (k >= map->max_entries)
 819                 return -EINVAL;
 820
 821         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
 822         if (old_dev) {
 823                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 824                 atomic_dec((atomic_t *)&dtab->items);
 825         }
 826         return 0;
 827 }
 828
 829 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 830 {
 831         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 832         struct bpf_dtab_netdev *old_dev;
 833         int k = *(u32 *)key;
 834         unsigned long flags;
 835         int ret = -ENOENT;
 836
 837         spin_lock_irqsave(&dtab->index_lock, flags);
 838
 839         old_dev = __dev_map_hash_lookup_elem(map, k);
 840         if (old_dev) {
 841                 dtab->items--;
 842                 hlist_del_init_rcu(&old_dev->index_hlist);
 843                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 844                 ret = 0;
 845         }
 846         spin_unlock_irqrestore(&dtab->index_lock, flags);
 847
 848         return ret;
 849 }
 850
 851 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 852                                                     struct bpf_dtab *dtab,
 853                                                     struct bpf_devmap_val *val,
 854                                                     unsigned int idx)
 855 {
 856         struct bpf_prog *prog = NULL;
 857         struct bpf_dtab_netdev *dev;
 858
 859         dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 860                                    GFP_NOWAIT | __GFP_NOWARN,
 861                                    dtab->map.numa_node);
 862         if (!dev)
 863                 return ERR_PTR(-ENOMEM);
 864
 865         dev->dev = dev_get_by_index(net, val->ifindex);
 866         if (!dev->dev)
 867                 goto err_out;
 868
 869         if (val->bpf_prog.fd > 0) {
 870                 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 871                                              BPF_PROG_TYPE_XDP, false);
 872                 if (IS_ERR(prog))
 873                         goto err_put_dev;
 874                 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
 875                     !bpf_prog_map_compatible(&dtab->map, prog))
 876                         goto err_put_prog;
 877         }
 878
 879         dev->idx = idx;
 880         dev->dtab = dtab;
 881         if (prog) {
 882                 dev->xdp_prog = prog;
 883                 dev->val.bpf_prog.id = prog->aux->id;
 884         } else {
 885                 dev->xdp_prog = NULL;
 886                 dev->val.bpf_prog.id = 0;
 887         }
 888         dev->val.ifindex = val->ifindex;
 889
 890         return dev;
 891 err_put_prog:
 892         bpf_prog_put(prog);
 893 err_put_dev:
 894         dev_put(dev->dev);
 895 err_out:
 896         kfree(dev);
 897         return ERR_PTR(-EINVAL);
 898 }
 899
 900 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
 901                                  void *key, void *value, u64 map_flags)
 902 {
 903         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 904         struct bpf_dtab_netdev *dev, *old_dev;
 905         struct bpf_devmap_val val = {};
 906         u32 i = *(u32 *)key;
 907
 908         if (unlikely(map_flags > BPF_EXIST))
 909                 return -EINVAL;
 910         if (unlikely(i >= dtab->map.max_entries))
 911                 return -E2BIG;
 912         if (unlikely(map_flags == BPF_NOEXIST))
 913                 return -EEXIST;
 914
 915         /* already verified value_size <= sizeof val */
 916         memcpy(&val, value, map->value_size);
 917
 918         if (!val.ifindex) {
 919                 dev = NULL;
 920                 /* can not specify fd if ifindex is 0 */
 921                 if (val.bpf_prog.fd > 0)
 922                         return -EINVAL;
 923         } else {
 924                 dev = __dev_map_alloc_node(net, dtab, &val, i);
 925                 if (IS_ERR(dev))
 926                         return PTR_ERR(dev);
 927         }
 928
 929         /* Use call_rcu() here to ensure rcu critical sections have completed
 930          * Remembering the driver side flush operation will happen before the
 931          * net device is removed.
 932          */
 933         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
 934         if (old_dev)
 935                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 936         else
 937                 atomic_inc((atomic_t *)&dtab->items);
 938
 939         return 0;
 940 }
 941
 942 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 943                                u64 map_flags)
 944 {
 945         return __dev_map_update_elem(current->nsproxy->net_ns,
 946                                      map, key, value, map_flags);
 947 }
 948
 949 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 950                                      void *key, void *value, u64 map_flags)
 951 {
 952         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 953         struct bpf_dtab_netdev *dev, *old_dev;
 954         struct bpf_devmap_val val = {};
 955         u32 idx = *(u32 *)key;
 956         unsigned long flags;
 957         int err = -EEXIST;
 958
 959         /* already verified value_size <= sizeof val */
 960         memcpy(&val, value, map->value_size);
 961
 962         if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 963                 return -EINVAL;
 964
 965         spin_lock_irqsave(&dtab->index_lock, flags);
 966
 967         old_dev = __dev_map_hash_lookup_elem(map, idx);
 968         if (old_dev && (map_flags & BPF_NOEXIST))
 969                 goto out_err;
 970
 971         dev = __dev_map_alloc_node(net, dtab, &val, idx);
 972         if (IS_ERR(dev)) {
 973                 err = PTR_ERR(dev);
 974                 goto out_err;
 975         }
 976
 977         if (old_dev) {
 978                 hlist_del_rcu(&old_dev->index_hlist);
 979         } else {
 980                 if (dtab->items >= dtab->map.max_entries) {
 981                         spin_unlock_irqrestore(&dtab->index_lock, flags);
 982                         call_rcu(&dev->rcu, __dev_map_entry_free);
 983                         return -E2BIG;
 984                 }
 985                 dtab->items++;
 986         }
 987
 988         hlist_add_head_rcu(&dev->index_hlist,
 989                            dev_map_index_hash(dtab, idx));
 990         spin_unlock_irqrestore(&dtab->index_lock, flags);
 991
 992         if (old_dev)
 993                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 994
 995         return 0;
 996
 997 out_err:
 998         spin_unlock_irqrestore(&dtab->index_lock, flags);
 999         return err;
1000 }
1001
1002 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1003                                    u64 map_flags)
1004 {
1005         return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1006                                          map, key, value, map_flags);
1007 }
1008
1009 static int dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1010 {
1011         return __bpf_xdp_redirect_map(map, ifindex, flags,
1012                                       BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1013                                       __dev_map_lookup_elem);
1014 }
1015
1016 static int dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1017 {
1018         return __bpf_xdp_redirect_map(map, ifindex, flags,
1019                                       BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1020                                       __dev_map_hash_lookup_elem);
1021 }
1022
1023 static u64 dev_map_mem_usage(const struct bpf_map *map)
1024 {
1025         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1026         u64 usage = sizeof(struct bpf_dtab);
1027
1028         if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1029                 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1030         else
1031                 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1032         usage += atomic_read((atomic_t *)&dtab->items) *
1033                          (u64)sizeof(struct bpf_dtab_netdev);
1034         return usage;
1035 }
1036
1037 BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1038 const struct bpf_map_ops dev_map_ops = {
1039         .map_meta_equal = bpf_map_meta_equal,
1040         .map_alloc = dev_map_alloc,
1041         .map_free = dev_map_free,
1042         .map_get_next_key = dev_map_get_next_key,
1043         .map_lookup_elem = dev_map_lookup_elem,
1044         .map_update_elem = dev_map_update_elem,
1045         .map_delete_elem = dev_map_delete_elem,
1046         .map_check_btf = map_check_no_btf,
1047         .map_mem_usage = dev_map_mem_usage,
1048         .map_btf_id = &dev_map_btf_ids[0],
1049         .map_redirect = dev_map_redirect,
1050 };
1051
1052 const struct bpf_map_ops dev_map_hash_ops = {
1053         .map_meta_equal = bpf_map_meta_equal,
1054         .map_alloc = dev_map_alloc,
1055         .map_free = dev_map_free,
1056         .map_get_next_key = dev_map_hash_get_next_key,
1057         .map_lookup_elem = dev_map_hash_lookup_elem,
1058         .map_update_elem = dev_map_hash_update_elem,
1059         .map_delete_elem = dev_map_hash_delete_elem,
1060         .map_check_btf = map_check_no_btf,
1061         .map_mem_usage = dev_map_mem_usage,
1062         .map_btf_id = &dev_map_btf_ids[0],
1063         .map_redirect = dev_hash_map_redirect,
1064 };
1065
1066 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1067                                        struct net_device *netdev)
1068 {
1069         unsigned long flags;
1070         u32 i;
1071
1072         spin_lock_irqsave(&dtab->index_lock, flags);
1073         for (i = 0; i < dtab->n_buckets; i++) {
1074                 struct bpf_dtab_netdev *dev;
1075                 struct hlist_head *head;
1076                 struct hlist_node *next;
1077
1078                 head = dev_map_index_hash(dtab, i);
1079
1080                 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1081                         if (netdev != dev->dev)
1082                                 continue;
1083
1084                         dtab->items--;
1085                         hlist_del_rcu(&dev->index_hlist);
1086                         call_rcu(&dev->rcu, __dev_map_entry_free);
1087                 }
1088         }
1089         spin_unlock_irqrestore(&dtab->index_lock, flags);
1090 }
1091
1092 static int dev_map_notification(struct notifier_block *notifier,
1093                                 ulong event, void *ptr)
1094 {
1095         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1096         struct bpf_dtab *dtab;
1097         int i, cpu;
1098
1099         switch (event) {
1100         case NETDEV_REGISTER:
1101                 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1102                         break;
1103
1104                 /* will be freed in free_netdev() */
1105                 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1106                 if (!netdev->xdp_bulkq)
1107                         return NOTIFY_BAD;
1108
1109                 for_each_possible_cpu(cpu)
1110                         per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1111                 break;
1112         case NETDEV_UNREGISTER:
1113                 /* This rcu_read_lock/unlock pair is needed because
1114                  * dev_map_list is an RCU list AND to ensure a delete
1115                  * operation does not free a netdev_map entry while we
1116                  * are comparing it against the netdev being unregistered.
1117                  */
1118                 rcu_read_lock();
1119                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1120                         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1121                                 dev_map_hash_remove_netdev(dtab, netdev);
1122                                 continue;
1123                         }
1124
1125                         for (i = 0; i < dtab->map.max_entries; i++) {
1126                                 struct bpf_dtab_netdev *dev, *odev;
1127
1128                                 dev = rcu_dereference(dtab->netdev_map[i]);
1129                                 if (!dev || netdev != dev->dev)
1130                                         continue;
1131                                 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1132                                 if (dev == odev) {
1133                                         call_rcu(&dev->rcu,
1134                                                  __dev_map_entry_free);
1135                                         atomic_dec((atomic_t *)&dtab->items);
1136                                 }
1137                         }
1138                 }
1139                 rcu_read_unlock();
1140                 break;
1141         default:
1142                 break;
1143         }
1144         return NOTIFY_OK;
1145 }
1146
1147 static struct notifier_block dev_map_notifier = {
1148         .notifier_call = dev_map_notification,
1149 };
1150
1151 static int __init dev_map_init(void)
1152 {
1153         int cpu;
1154
1155         /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1156         BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1157                      offsetof(struct _bpf_dtab_netdev, dev));
1158         register_netdevice_notifier(&dev_map_notifier);
1159
1160         for_each_possible_cpu(cpu)
1161                 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1162         return 0;
1163 }
1164
1165 subsys_initcall(dev_map_init);