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