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_prog *xdp_prog;
  69         struct rcu_head rcu;
  70         unsigned int idx;
  71         struct bpf_devmap_val val;
  72 };
  73
  74 struct bpf_dtab {
  75         struct bpf_map map;
  76         struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
  77         struct list_head list;
  78
  79         /* these are only used for DEVMAP_HASH type maps */
  80         struct hlist_head *dev_index_head;
  81         spinlock_t index_lock;
  82         unsigned int items;
  83         u32 n_buckets;
  84 };
  85
  86 static DEFINE_SPINLOCK(dev_map_lock);
  87 static LIST_HEAD(dev_map_list);
  88
  89 static struct hlist_head *dev_map_create_hash(unsigned int entries,
  90                                               int numa_node)
  91 {
  92         int i;
  93         struct hlist_head *hash;
  94
  95         hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
  96         if (hash != NULL)
  97                 for (i = 0; i < entries; i++)
  98                         INIT_HLIST_HEAD(&hash[i]);
  99
 100         return hash;
 101 }
 102
 103 static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
 104                                                     int idx)
 105 {
 106         return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
 107 }
 108
 109 static int dev_map_alloc_check(union bpf_attr *attr)
 110 {
 111         u32 valsize = attr->value_size;
 112
 113         /* check sanity of attributes. 2 value sizes supported:
 114          * 4 bytes: ifindex
 115          * 8 bytes: ifindex + prog fd
 116          */
 117         if (attr->max_entries == 0 || attr->key_size != 4 ||
 118             (valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
 119              valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
 120             attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 121                 return -EINVAL;
 122
 123         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 124                 /* Hash table size must be power of 2; roundup_pow_of_two()
 125                  * can overflow into UB on 32-bit arches
 126                  */
 127                 if (attr->max_entries > 1UL << 31)
 128                         return -EINVAL;
 129         }
 130
 131         return 0;
 132 }
 133
 134 static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
 135 {
 136         /* Lookup returns a pointer straight to dev->ifindex, so make sure the
 137          * verifier prevents writes from the BPF side
 138          */
 139         attr->map_flags |= BPF_F_RDONLY_PROG;
 140         bpf_map_init_from_attr(&dtab->map, attr);
 141
 142         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 143                 /* Hash table size must be power of 2 */
 144                 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
 145                 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
 146                                                            dtab->map.numa_node);
 147                 if (!dtab->dev_index_head)
 148                         return -ENOMEM;
 149
 150                 spin_lock_init(&dtab->index_lock);
 151         } else {
 152                 dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
 153                                                       sizeof(struct bpf_dtab_netdev *),
 154                                                       dtab->map.numa_node);
 155                 if (!dtab->netdev_map)
 156                         return -ENOMEM;
 157         }
 158
 159         return 0;
 160 }
 161
 162 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 163 {
 164         struct bpf_dtab *dtab;
 165         int err;
 166
 167         dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
 168         if (!dtab)
 169                 return ERR_PTR(-ENOMEM);
 170
 171         err = dev_map_init_map(dtab, attr);
 172         if (err) {
 173                 bpf_map_area_free(dtab);
 174                 return ERR_PTR(err);
 175         }
 176
 177         spin_lock(&dev_map_lock);
 178         list_add_tail_rcu(&dtab->list, &dev_map_list);
 179         spin_unlock(&dev_map_lock);
 180
 181         return &dtab->map;
 182 }
 183
 184 static void dev_map_free(struct bpf_map *map)
 185 {
 186         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 187         u32 i;
 188
 189         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 190          * so the programs (can be more than one that used this map) were
 191          * disconnected from events. The following synchronize_rcu() guarantees
 192          * both rcu read critical sections complete and waits for
 193          * preempt-disable regions (NAPI being the relevant context here) so we
 194          * are certain there will be no further reads against the netdev_map and
 195          * all flush operations are complete. Flush operations can only be done
 196          * from NAPI context for this reason.
 197          */
 198
 199         spin_lock(&dev_map_lock);
 200         list_del_rcu(&dtab->list);
 201         spin_unlock(&dev_map_lock);
 202
 203         /* bpf_redirect_info->map is assigned in __bpf_xdp_redirect_map()
 204          * during NAPI callback and cleared after the XDP redirect. There is no
 205          * explicit RCU read section which protects bpf_redirect_info->map but
 206          * local_bh_disable() also marks the beginning an RCU section. This
 207          * makes the complete softirq callback RCU protected. Thus after
 208          * following synchronize_rcu() there no bpf_redirect_info->map == map
 209          * assignment.
 210          */
 211         synchronize_rcu();
 212
 213         /* Make sure prior __dev_map_entry_free() have completed. */
 214         rcu_barrier();
 215
 216         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 217                 for (i = 0; i < dtab->n_buckets; i++) {
 218                         struct bpf_dtab_netdev *dev;
 219                         struct hlist_head *head;
 220                         struct hlist_node *next;
 221
 222                         head = dev_map_index_hash(dtab, i);
 223
 224                         hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 225                                 hlist_del_rcu(&dev->index_hlist);
 226                                 if (dev->xdp_prog)
 227                                         bpf_prog_put(dev->xdp_prog);
 228                                 dev_put(dev->dev);
 229                                 kfree(dev);
 230                         }
 231                 }
 232
 233                 bpf_map_area_free(dtab->dev_index_head);
 234         } else {
 235                 for (i = 0; i < dtab->map.max_entries; i++) {
 236                         struct bpf_dtab_netdev *dev;
 237
 238                         dev = rcu_dereference_raw(dtab->netdev_map[i]);
 239                         if (!dev)
 240                                 continue;
 241
 242                         if (dev->xdp_prog)
 243                                 bpf_prog_put(dev->xdp_prog);
 244                         dev_put(dev->dev);
 245                         kfree(dev);
 246                 }
 247
 248                 bpf_map_area_free(dtab->netdev_map);
 249         }
 250
 251         bpf_map_area_free(dtab);
 252 }
 253
 254 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 255 {
 256         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 257         u32 index = key ? *(u32 *)key : U32_MAX;
 258         u32 *next = next_key;
 259
 260         if (index >= dtab->map.max_entries) {
 261                 *next = 0;
 262                 return 0;
 263         }
 264
 265         if (index == dtab->map.max_entries - 1)
 266                 return -ENOENT;
 267         *next = index + 1;
 268         return 0;
 269 }
 270
 271 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 272  * by local_bh_disable() (from XDP calls inside NAPI). The
 273  * rcu_read_lock_bh_held() below makes lockdep accept both.
 274  */
 275 static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 276 {
 277         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 278         struct hlist_head *head = dev_map_index_hash(dtab, key);
 279         struct bpf_dtab_netdev *dev;
 280
 281         hlist_for_each_entry_rcu(dev, head, index_hlist,
 282                                  lockdep_is_held(&dtab->index_lock))
 283                 if (dev->idx == key)
 284                         return dev;
 285
 286         return NULL;
 287 }
 288
 289 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 290                                     void *next_key)
 291 {
 292         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 293         u32 idx, *next = next_key;
 294         struct bpf_dtab_netdev *dev, *next_dev;
 295         struct hlist_head *head;
 296         int i = 0;
 297
 298         if (!key)
 299                 goto find_first;
 300
 301         idx = *(u32 *)key;
 302
 303         dev = __dev_map_hash_lookup_elem(map, idx);
 304         if (!dev)
 305                 goto find_first;
 306
 307         next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 308                                     struct bpf_dtab_netdev, index_hlist);
 309
 310         if (next_dev) {
 311                 *next = next_dev->idx;
 312                 return 0;
 313         }
 314
 315         i = idx & (dtab->n_buckets - 1);
 316         i++;
 317
 318  find_first:
 319         for (; i < dtab->n_buckets; i++) {
 320                 head = dev_map_index_hash(dtab, i);
 321
 322                 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 323                                             struct bpf_dtab_netdev,
 324                                             index_hlist);
 325                 if (next_dev) {
 326                         *next = next_dev->idx;
 327                         return 0;
 328                 }
 329         }
 330
 331         return -ENOENT;
 332 }
 333
 334 static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
 335                                 struct xdp_frame **frames, int n,
 336                                 struct net_device *tx_dev,
 337                                 struct net_device *rx_dev)
 338 {
 339         struct xdp_txq_info txq = { .dev = tx_dev };
 340         struct xdp_rxq_info rxq = { .dev = rx_dev };
 341         struct xdp_buff xdp;
 342         int i, nframes = 0;
 343
 344         for (i = 0; i < n; i++) {
 345                 struct xdp_frame *xdpf = frames[i];
 346                 u32 act;
 347                 int err;
 348
 349                 xdp_convert_frame_to_buff(xdpf, &xdp);
 350                 xdp.txq = &txq;
 351                 xdp.rxq = &rxq;
 352
 353                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 354                 switch (act) {
 355                 case XDP_PASS:
 356                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 357                         if (unlikely(err < 0))
 358                                 xdp_return_frame_rx_napi(xdpf);
 359                         else
 360                                 frames[nframes++] = xdpf;
 361                         break;
 362                 default:
 363                         bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
 364                         fallthrough;
 365                 case XDP_ABORTED:
 366                         trace_xdp_exception(tx_dev, xdp_prog, act);
 367                         fallthrough;
 368                 case XDP_DROP:
 369                         xdp_return_frame_rx_napi(xdpf);
 370                         break;
 371                 }
 372         }
 373         return nframes; /* sent frames count */
 374 }
 375
 376 static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 377 {
 378         struct net_device *dev = bq->dev;
 379         unsigned int cnt = bq->count;
 380         int sent = 0, err = 0;
 381         int to_send = cnt;
 382         int i;
 383
 384         if (unlikely(!cnt))
 385                 return;
 386
 387         for (i = 0; i < cnt; i++) {
 388                 struct xdp_frame *xdpf = bq->q[i];
 389
 390                 prefetch(xdpf);
 391         }
 392
 393         if (bq->xdp_prog) {
 394                 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev, bq->dev_rx);
 395                 if (!to_send)
 396                         goto out;
 397         }
 398
 399         sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
 400         if (sent < 0) {
 401                 /* If ndo_xdp_xmit fails with an errno, no frames have
 402                  * been xmit'ed.
 403                  */
 404                 err = sent;
 405                 sent = 0;
 406         }
 407
 408         /* If not all frames have been transmitted, it is our
 409          * responsibility to free them
 410          */
 411         for (i = sent; unlikely(i < to_send); i++)
 412                 xdp_return_frame_rx_napi(bq->q[i]);
 413
 414 out:
 415         bq->count = 0;
 416         trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
 417 }
 418
 419 /* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
 420  * driver before returning from its napi->poll() routine. See the comment above
 421  * xdp_do_flush() in filter.c.
 422  */
 423 void __dev_flush(struct list_head *flush_list)
 424 {
 425         struct xdp_dev_bulk_queue *bq, *tmp;
 426
 427         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 428                 bq_xmit_all(bq, XDP_XMIT_FLUSH);
 429                 bq->dev_rx = NULL;
 430                 bq->xdp_prog = NULL;
 431                 __list_del_clearprev(&bq->flush_node);
 432         }
 433 }
 434
 435 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 436  * by local_bh_disable() (from XDP calls inside NAPI). The
 437  * rcu_read_lock_bh_held() below makes lockdep accept both.
 438  */
 439 static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 440 {
 441         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 442         struct bpf_dtab_netdev *obj;
 443
 444         if (key >= map->max_entries)
 445                 return NULL;
 446
 447         obj = rcu_dereference_check(dtab->netdev_map[key],
 448                                     rcu_read_lock_bh_held());
 449         return obj;
 450 }
 451
 452 /* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
 453  * variable access, and map elements stick around. See comment above
 454  * xdp_do_flush() in filter.c.
 455  */
 456 static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 457                        struct net_device *dev_rx, struct bpf_prog *xdp_prog)
 458 {
 459         struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 460
 461         if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 462                 bq_xmit_all(bq, 0);
 463
 464         /* Ingress dev_rx will be the same for all xdp_frame's in
 465          * bulk_queue, because bq stored per-CPU and must be flushed
 466          * from net_device drivers NAPI func end.
 467          *
 468          * Do the same with xdp_prog and flush_list since these fields
 469          * are only ever modified together.
 470          */
 471         if (!bq->dev_rx) {
 472                 struct list_head *flush_list = bpf_net_ctx_get_dev_flush_list();
 473
 474                 bq->dev_rx = dev_rx;
 475                 bq->xdp_prog = xdp_prog;
 476                 list_add(&bq->flush_node, flush_list);
 477         }
 478
 479         bq->q[bq->count++] = xdpf;
 480 }
 481
 482 static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 483                                 struct net_device *dev_rx,
 484                                 struct bpf_prog *xdp_prog)
 485 {
 486         int err;
 487
 488         if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 489                 return -EOPNOTSUPP;
 490
 491         if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 492                      xdp_frame_has_frags(xdpf)))
 493                 return -EOPNOTSUPP;
 494
 495         err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
 496         if (unlikely(err))
 497                 return err;
 498
 499         bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
 500         return 0;
 501 }
 502
 503 static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
 504 {
 505         struct xdp_txq_info txq = { .dev = dst->dev };
 506         struct xdp_buff xdp;
 507         u32 act;
 508
 509         if (!dst->xdp_prog)
 510                 return XDP_PASS;
 511
 512         __skb_pull(skb, skb->mac_len);
 513         xdp.txq = &txq;
 514
 515         act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
 516         switch (act) {
 517         case XDP_PASS:
 518                 __skb_push(skb, skb->mac_len);
 519                 break;
 520         default:
 521                 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
 522                 fallthrough;
 523         case XDP_ABORTED:
 524                 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
 525                 fallthrough;
 526         case XDP_DROP:
 527                 kfree_skb(skb);
 528                 break;
 529         }
 530
 531         return act;
 532 }
 533
 534 int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 535                     struct net_device *dev_rx)
 536 {
 537         return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
 538 }
 539
 540 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
 541                     struct net_device *dev_rx)
 542 {
 543         struct net_device *dev = dst->dev;
 544
 545         return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
 546 }
 547
 548 static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
 549 {
 550         if (!obj)
 551                 return false;
 552
 553         if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 554                 return false;
 555
 556         if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 557                      xdp_frame_has_frags(xdpf)))
 558                 return false;
 559
 560         if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
 561                 return false;
 562
 563         return true;
 564 }
 565
 566 static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
 567                                  struct net_device *dev_rx,
 568                                  struct xdp_frame *xdpf)
 569 {
 570         struct xdp_frame *nxdpf;
 571
 572         nxdpf = xdpf_clone(xdpf);
 573         if (!nxdpf)
 574                 return -ENOMEM;
 575
 576         bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
 577
 578         return 0;
 579 }
 580
 581 static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
 582 {
 583         while (num_excluded--) {
 584                 if (ifindex == excluded[num_excluded])
 585                         return true;
 586         }
 587         return false;
 588 }
 589
 590 /* Get ifindex of each upper device. 'indexes' must be able to hold at
 591  * least MAX_NEST_DEV elements.
 592  * Returns the number of ifindexes added.
 593  */
 594 static int get_upper_ifindexes(struct net_device *dev, int *indexes)
 595 {
 596         struct net_device *upper;
 597         struct list_head *iter;
 598         int n = 0;
 599
 600         netdev_for_each_upper_dev_rcu(dev, upper, iter) {
 601                 indexes[n++] = upper->ifindex;
 602         }
 603         return n;
 604 }
 605
 606 int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
 607                           struct bpf_map *map, bool exclude_ingress)
 608 {
 609         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 610         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 611         int excluded_devices[1+MAX_NEST_DEV];
 612         struct hlist_head *head;
 613         int num_excluded = 0;
 614         unsigned int i;
 615         int err;
 616
 617         if (exclude_ingress) {
 618                 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
 619                 excluded_devices[num_excluded++] = dev_rx->ifindex;
 620         }
 621
 622         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 623                 for (i = 0; i < map->max_entries; i++) {
 624                         dst = rcu_dereference_check(dtab->netdev_map[i],
 625                                                     rcu_read_lock_bh_held());
 626                         if (!is_valid_dst(dst, xdpf))
 627                                 continue;
 628
 629                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 630                                 continue;
 631
 632                         /* we only need n-1 clones; last_dst enqueued below */
 633                         if (!last_dst) {
 634                                 last_dst = dst;
 635                                 continue;
 636                         }
 637
 638                         err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 639                         if (err)
 640                                 return err;
 641
 642                         last_dst = dst;
 643                 }
 644         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 645                 for (i = 0; i < dtab->n_buckets; i++) {
 646                         head = dev_map_index_hash(dtab, i);
 647                         hlist_for_each_entry_rcu(dst, head, index_hlist,
 648                                                  lockdep_is_held(&dtab->index_lock)) {
 649                                 if (!is_valid_dst(dst, xdpf))
 650                                         continue;
 651
 652                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 653                                                         dst->dev->ifindex))
 654                                         continue;
 655
 656                                 /* we only need n-1 clones; last_dst enqueued below */
 657                                 if (!last_dst) {
 658                                         last_dst = dst;
 659                                         continue;
 660                                 }
 661
 662                                 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 663                                 if (err)
 664                                         return err;
 665
 666                                 last_dst = dst;
 667                         }
 668                 }
 669         }
 670
 671         /* consume the last copy of the frame */
 672         if (last_dst)
 673                 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
 674         else
 675                 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
 676
 677         return 0;
 678 }
 679
 680 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 681                              struct bpf_prog *xdp_prog)
 682 {
 683         int err;
 684
 685         err = xdp_ok_fwd_dev(dst->dev, skb->len);
 686         if (unlikely(err))
 687                 return err;
 688
 689         /* Redirect has already succeeded semantically at this point, so we just
 690          * return 0 even if packet is dropped. Helper below takes care of
 691          * freeing skb.
 692          */
 693         if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
 694                 return 0;
 695
 696         skb->dev = dst->dev;
 697         generic_xdp_tx(skb, xdp_prog);
 698
 699         return 0;
 700 }
 701
 702 static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
 703                                   struct sk_buff *skb,
 704                                   struct bpf_prog *xdp_prog)
 705 {
 706         struct sk_buff *nskb;
 707         int err;
 708
 709         nskb = skb_clone(skb, GFP_ATOMIC);
 710         if (!nskb)
 711                 return -ENOMEM;
 712
 713         err = dev_map_generic_redirect(dst, nskb, xdp_prog);
 714         if (unlikely(err)) {
 715                 consume_skb(nskb);
 716                 return err;
 717         }
 718
 719         return 0;
 720 }
 721
 722 int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
 723                            struct bpf_prog *xdp_prog, struct bpf_map *map,
 724                            bool exclude_ingress)
 725 {
 726         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 727         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 728         int excluded_devices[1+MAX_NEST_DEV];
 729         struct hlist_head *head;
 730         struct hlist_node *next;
 731         int num_excluded = 0;
 732         unsigned int i;
 733         int err;
 734
 735         if (exclude_ingress) {
 736                 num_excluded = get_upper_ifindexes(dev, excluded_devices);
 737                 excluded_devices[num_excluded++] = dev->ifindex;
 738         }
 739
 740         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 741                 for (i = 0; i < map->max_entries; i++) {
 742                         dst = rcu_dereference_check(dtab->netdev_map[i],
 743                                                     rcu_read_lock_bh_held());
 744                         if (!dst)
 745                                 continue;
 746
 747                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 748                                 continue;
 749
 750                         /* we only need n-1 clones; last_dst enqueued below */
 751                         if (!last_dst) {
 752                                 last_dst = dst;
 753                                 continue;
 754                         }
 755
 756                         err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 757                         if (err)
 758                                 return err;
 759
 760                         last_dst = dst;
 761
 762                 }
 763         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 764                 for (i = 0; i < dtab->n_buckets; i++) {
 765                         head = dev_map_index_hash(dtab, i);
 766                         hlist_for_each_entry_safe(dst, next, head, index_hlist) {
 767                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 768                                                         dst->dev->ifindex))
 769                                         continue;
 770
 771                                 /* we only need n-1 clones; last_dst enqueued below */
 772                                 if (!last_dst) {
 773                                         last_dst = dst;
 774                                         continue;
 775                                 }
 776
 777                                 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 778                                 if (err)
 779                                         return err;
 780
 781                                 last_dst = dst;
 782                         }
 783                 }
 784         }
 785
 786         /* consume the first skb and return */
 787         if (last_dst)
 788                 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
 789
 790         /* dtab is empty */
 791         consume_skb(skb);
 792         return 0;
 793 }
 794
 795 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 796 {
 797         struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 798
 799         return obj ? &obj->val : NULL;
 800 }
 801
 802 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 803 {
 804         struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 805                                                                 *(u32 *)key);
 806         return obj ? &obj->val : NULL;
 807 }
 808
 809 static void __dev_map_entry_free(struct rcu_head *rcu)
 810 {
 811         struct bpf_dtab_netdev *dev;
 812
 813         dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 814         if (dev->xdp_prog)
 815                 bpf_prog_put(dev->xdp_prog);
 816         dev_put(dev->dev);
 817         kfree(dev);
 818 }
 819
 820 static long dev_map_delete_elem(struct bpf_map *map, void *key)
 821 {
 822         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 823         struct bpf_dtab_netdev *old_dev;
 824         u32 k = *(u32 *)key;
 825
 826         if (k >= map->max_entries)
 827                 return -EINVAL;
 828
 829         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
 830         if (old_dev) {
 831                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 832                 atomic_dec((atomic_t *)&dtab->items);
 833         }
 834         return 0;
 835 }
 836
 837 static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 838 {
 839         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 840         struct bpf_dtab_netdev *old_dev;
 841         u32 k = *(u32 *)key;
 842         unsigned long flags;
 843         int ret = -ENOENT;
 844
 845         spin_lock_irqsave(&dtab->index_lock, flags);
 846
 847         old_dev = __dev_map_hash_lookup_elem(map, k);
 848         if (old_dev) {
 849                 dtab->items--;
 850                 hlist_del_init_rcu(&old_dev->index_hlist);
 851                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 852                 ret = 0;
 853         }
 854         spin_unlock_irqrestore(&dtab->index_lock, flags);
 855
 856         return ret;
 857 }
 858
 859 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 860                                                     struct bpf_dtab *dtab,
 861                                                     struct bpf_devmap_val *val,
 862                                                     unsigned int idx)
 863 {
 864         struct bpf_prog *prog = NULL;
 865         struct bpf_dtab_netdev *dev;
 866
 867         dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 868                                    GFP_NOWAIT | __GFP_NOWARN,
 869                                    dtab->map.numa_node);
 870         if (!dev)
 871                 return ERR_PTR(-ENOMEM);
 872
 873         dev->dev = dev_get_by_index(net, val->ifindex);
 874         if (!dev->dev)
 875                 goto err_out;
 876
 877         if (val->bpf_prog.fd > 0) {
 878                 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 879                                              BPF_PROG_TYPE_XDP, false);
 880                 if (IS_ERR(prog))
 881                         goto err_put_dev;
 882                 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
 883                     !bpf_prog_map_compatible(&dtab->map, prog))
 884                         goto err_put_prog;
 885         }
 886
 887         dev->idx = idx;
 888         if (prog) {
 889                 dev->xdp_prog = prog;
 890                 dev->val.bpf_prog.id = prog->aux->id;
 891         } else {
 892                 dev->xdp_prog = NULL;
 893                 dev->val.bpf_prog.id = 0;
 894         }
 895         dev->val.ifindex = val->ifindex;
 896
 897         return dev;
 898 err_put_prog:
 899         bpf_prog_put(prog);
 900 err_put_dev:
 901         dev_put(dev->dev);
 902 err_out:
 903         kfree(dev);
 904         return ERR_PTR(-EINVAL);
 905 }
 906
 907 static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
 908                                   void *key, void *value, u64 map_flags)
 909 {
 910         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 911         struct bpf_dtab_netdev *dev, *old_dev;
 912         struct bpf_devmap_val val = {};
 913         u32 i = *(u32 *)key;
 914
 915         if (unlikely(map_flags > BPF_EXIST))
 916                 return -EINVAL;
 917         if (unlikely(i >= dtab->map.max_entries))
 918                 return -E2BIG;
 919         if (unlikely(map_flags == BPF_NOEXIST))
 920                 return -EEXIST;
 921
 922         /* already verified value_size <= sizeof val */
 923         memcpy(&val, value, map->value_size);
 924
 925         if (!val.ifindex) {
 926                 dev = NULL;
 927                 /* can not specify fd if ifindex is 0 */
 928                 if (val.bpf_prog.fd > 0)
 929                         return -EINVAL;
 930         } else {
 931                 dev = __dev_map_alloc_node(net, dtab, &val, i);
 932                 if (IS_ERR(dev))
 933                         return PTR_ERR(dev);
 934         }
 935
 936         /* Use call_rcu() here to ensure rcu critical sections have completed
 937          * Remembering the driver side flush operation will happen before the
 938          * net device is removed.
 939          */
 940         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
 941         if (old_dev)
 942                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 943         else
 944                 atomic_inc((atomic_t *)&dtab->items);
 945
 946         return 0;
 947 }
 948
 949 static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 950                                 u64 map_flags)
 951 {
 952         return __dev_map_update_elem(current->nsproxy->net_ns,
 953                                      map, key, value, map_flags);
 954 }
 955
 956 static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 957                                        void *key, void *value, u64 map_flags)
 958 {
 959         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 960         struct bpf_dtab_netdev *dev, *old_dev;
 961         struct bpf_devmap_val val = {};
 962         u32 idx = *(u32 *)key;
 963         unsigned long flags;
 964         int err = -EEXIST;
 965
 966         /* already verified value_size <= sizeof val */
 967         memcpy(&val, value, map->value_size);
 968
 969         if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 970                 return -EINVAL;
 971
 972         spin_lock_irqsave(&dtab->index_lock, flags);
 973
 974         old_dev = __dev_map_hash_lookup_elem(map, idx);
 975         if (old_dev && (map_flags & BPF_NOEXIST))
 976                 goto out_err;
 977
 978         dev = __dev_map_alloc_node(net, dtab, &val, idx);
 979         if (IS_ERR(dev)) {
 980                 err = PTR_ERR(dev);
 981                 goto out_err;
 982         }
 983
 984         if (old_dev) {
 985                 hlist_del_rcu(&old_dev->index_hlist);
 986         } else {
 987                 if (dtab->items >= dtab->map.max_entries) {
 988                         spin_unlock_irqrestore(&dtab->index_lock, flags);
 989                         call_rcu(&dev->rcu, __dev_map_entry_free);
 990                         return -E2BIG;
 991                 }
 992                 dtab->items++;
 993         }
 994
 995         hlist_add_head_rcu(&dev->index_hlist,
 996                            dev_map_index_hash(dtab, idx));
 997         spin_unlock_irqrestore(&dtab->index_lock, flags);
 998
 999         if (old_dev)
1000                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
1001
1002         return 0;
1003
1004 out_err:
1005         spin_unlock_irqrestore(&dtab->index_lock, flags);
1006         return err;
1007 }
1008
1009 static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1010                                      u64 map_flags)
1011 {
1012         return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1013                                          map, key, value, map_flags);
1014 }
1015
1016 static long dev_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_lookup_elem);
1021 }
1022
1023 static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1024 {
1025         return __bpf_xdp_redirect_map(map, ifindex, flags,
1026                                       BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1027                                       __dev_map_hash_lookup_elem);
1028 }
1029
1030 static u64 dev_map_mem_usage(const struct bpf_map *map)
1031 {
1032         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1033         u64 usage = sizeof(struct bpf_dtab);
1034
1035         if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1036                 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1037         else
1038                 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1039         usage += atomic_read((atomic_t *)&dtab->items) *
1040                          (u64)sizeof(struct bpf_dtab_netdev);
1041         return usage;
1042 }
1043
1044 BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1045 const struct bpf_map_ops dev_map_ops = {
1046         .map_meta_equal = bpf_map_meta_equal,
1047         .map_alloc_check = dev_map_alloc_check,
1048         .map_alloc = dev_map_alloc,
1049         .map_free = dev_map_free,
1050         .map_get_next_key = dev_map_get_next_key,
1051         .map_lookup_elem = dev_map_lookup_elem,
1052         .map_update_elem = dev_map_update_elem,
1053         .map_delete_elem = dev_map_delete_elem,
1054         .map_check_btf = map_check_no_btf,
1055         .map_mem_usage = dev_map_mem_usage,
1056         .map_btf_id = &dev_map_btf_ids[0],
1057         .map_redirect = dev_map_redirect,
1058 };
1059
1060 const struct bpf_map_ops dev_map_hash_ops = {
1061         .map_meta_equal = bpf_map_meta_equal,
1062         .map_alloc_check = dev_map_alloc_check,
1063         .map_alloc = dev_map_alloc,
1064         .map_free = dev_map_free,
1065         .map_get_next_key = dev_map_hash_get_next_key,
1066         .map_lookup_elem = dev_map_hash_lookup_elem,
1067         .map_update_elem = dev_map_hash_update_elem,
1068         .map_delete_elem = dev_map_hash_delete_elem,
1069         .map_check_btf = map_check_no_btf,
1070         .map_mem_usage = dev_map_mem_usage,
1071         .map_btf_id = &dev_map_btf_ids[0],
1072         .map_redirect = dev_hash_map_redirect,
1073 };
1074
1075 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1076                                        struct net_device *netdev)
1077 {
1078         unsigned long flags;
1079         u32 i;
1080
1081         spin_lock_irqsave(&dtab->index_lock, flags);
1082         for (i = 0; i < dtab->n_buckets; i++) {
1083                 struct bpf_dtab_netdev *dev;
1084                 struct hlist_head *head;
1085                 struct hlist_node *next;
1086
1087                 head = dev_map_index_hash(dtab, i);
1088
1089                 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1090                         if (netdev != dev->dev)
1091                                 continue;
1092
1093                         dtab->items--;
1094                         hlist_del_rcu(&dev->index_hlist);
1095                         call_rcu(&dev->rcu, __dev_map_entry_free);
1096                 }
1097         }
1098         spin_unlock_irqrestore(&dtab->index_lock, flags);
1099 }
1100
1101 static int dev_map_notification(struct notifier_block *notifier,
1102                                 ulong event, void *ptr)
1103 {
1104         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1105         struct bpf_dtab *dtab;
1106         int i, cpu;
1107
1108         switch (event) {
1109         case NETDEV_REGISTER:
1110                 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1111                         break;
1112
1113                 /* will be freed in free_netdev() */
1114                 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1115                 if (!netdev->xdp_bulkq)
1116                         return NOTIFY_BAD;
1117
1118                 for_each_possible_cpu(cpu)
1119                         per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1120                 break;
1121         case NETDEV_UNREGISTER:
1122                 /* This rcu_read_lock/unlock pair is needed because
1123                  * dev_map_list is an RCU list AND to ensure a delete
1124                  * operation does not free a netdev_map entry while we
1125                  * are comparing it against the netdev being unregistered.
1126                  */
1127                 rcu_read_lock();
1128                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1129                         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1130                                 dev_map_hash_remove_netdev(dtab, netdev);
1131                                 continue;
1132                         }
1133
1134                         for (i = 0; i < dtab->map.max_entries; i++) {
1135                                 struct bpf_dtab_netdev *dev, *odev;
1136
1137                                 dev = rcu_dereference(dtab->netdev_map[i]);
1138                                 if (!dev || netdev != dev->dev)
1139                                         continue;
1140                                 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1141                                 if (dev == odev) {
1142                                         call_rcu(&dev->rcu,
1143                                                  __dev_map_entry_free);
1144                                         atomic_dec((atomic_t *)&dtab->items);
1145                                 }
1146                         }
1147                 }
1148                 rcu_read_unlock();
1149                 break;
1150         default:
1151                 break;
1152         }
1153         return NOTIFY_OK;
1154 }
1155
1156 static struct notifier_block dev_map_notifier = {
1157         .notifier_call = dev_map_notification,
1158 };
1159
1160 static int __init dev_map_init(void)
1161 {
1162         /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1163         BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1164                      offsetof(struct _bpf_dtab_netdev, dev));
1165         register_netdevice_notifier(&dev_map_notifier);
1166
1167         return 0;
1168 }
1169
1170 subsys_initcall(dev_map_init);