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         int 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 *dev)
 337 {
 338         struct xdp_txq_info txq = { .dev = dev };
 339         struct xdp_buff xdp;
 340         int i, nframes = 0;
 341
 342         for (i = 0; i < n; i++) {
 343                 struct xdp_frame *xdpf = frames[i];
 344                 u32 act;
 345                 int err;
 346
 347                 xdp_convert_frame_to_buff(xdpf, &xdp);
 348                 xdp.txq = &txq;
 349
 350                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
 351                 switch (act) {
 352                 case XDP_PASS:
 353                         err = xdp_update_frame_from_buff(&xdp, xdpf);
 354                         if (unlikely(err < 0))
 355                                 xdp_return_frame_rx_napi(xdpf);
 356                         else
 357                                 frames[nframes++] = xdpf;
 358                         break;
 359                 default:
 360                         bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
 361                         fallthrough;
 362                 case XDP_ABORTED:
 363                         trace_xdp_exception(dev, xdp_prog, act);
 364                         fallthrough;
 365                 case XDP_DROP:
 366                         xdp_return_frame_rx_napi(xdpf);
 367                         break;
 368                 }
 369         }
 370         return nframes; /* sent frames count */
 371 }
 372
 373 static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 374 {
 375         struct net_device *dev = bq->dev;
 376         unsigned int cnt = bq->count;
 377         int sent = 0, err = 0;
 378         int to_send = cnt;
 379         int i;
 380
 381         if (unlikely(!cnt))
 382                 return;
 383
 384         for (i = 0; i < cnt; i++) {
 385                 struct xdp_frame *xdpf = bq->q[i];
 386
 387                 prefetch(xdpf);
 388         }
 389
 390         if (bq->xdp_prog) {
 391                 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
 392                 if (!to_send)
 393                         goto out;
 394         }
 395
 396         sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
 397         if (sent < 0) {
 398                 /* If ndo_xdp_xmit fails with an errno, no frames have
 399                  * been xmit'ed.
 400                  */
 401                 err = sent;
 402                 sent = 0;
 403         }
 404
 405         /* If not all frames have been transmitted, it is our
 406          * responsibility to free them
 407          */
 408         for (i = sent; unlikely(i < to_send); i++)
 409                 xdp_return_frame_rx_napi(bq->q[i]);
 410
 411 out:
 412         bq->count = 0;
 413         trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
 414 }
 415
 416 /* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
 417  * driver before returning from its napi->poll() routine. See the comment above
 418  * xdp_do_flush() in filter.c.
 419  */
 420 void __dev_flush(struct list_head *flush_list)
 421 {
 422         struct xdp_dev_bulk_queue *bq, *tmp;
 423
 424         list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
 425                 bq_xmit_all(bq, XDP_XMIT_FLUSH);
 426                 bq->dev_rx = NULL;
 427                 bq->xdp_prog = NULL;
 428                 __list_del_clearprev(&bq->flush_node);
 429         }
 430 }
 431
 432 /* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
 433  * by local_bh_disable() (from XDP calls inside NAPI). The
 434  * rcu_read_lock_bh_held() below makes lockdep accept both.
 435  */
 436 static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 437 {
 438         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 439         struct bpf_dtab_netdev *obj;
 440
 441         if (key >= map->max_entries)
 442                 return NULL;
 443
 444         obj = rcu_dereference_check(dtab->netdev_map[key],
 445                                     rcu_read_lock_bh_held());
 446         return obj;
 447 }
 448
 449 /* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
 450  * variable access, and map elements stick around. See comment above
 451  * xdp_do_flush() in filter.c.
 452  */
 453 static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 454                        struct net_device *dev_rx, struct bpf_prog *xdp_prog)
 455 {
 456         struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 457
 458         if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 459                 bq_xmit_all(bq, 0);
 460
 461         /* Ingress dev_rx will be the same for all xdp_frame's in
 462          * bulk_queue, because bq stored per-CPU and must be flushed
 463          * from net_device drivers NAPI func end.
 464          *
 465          * Do the same with xdp_prog and flush_list since these fields
 466          * are only ever modified together.
 467          */
 468         if (!bq->dev_rx) {
 469                 struct list_head *flush_list = bpf_net_ctx_get_dev_flush_list();
 470
 471                 bq->dev_rx = dev_rx;
 472                 bq->xdp_prog = xdp_prog;
 473                 list_add(&bq->flush_node, flush_list);
 474         }
 475
 476         bq->q[bq->count++] = xdpf;
 477 }
 478
 479 static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 480                                 struct net_device *dev_rx,
 481                                 struct bpf_prog *xdp_prog)
 482 {
 483         int err;
 484
 485         if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 486                 return -EOPNOTSUPP;
 487
 488         if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 489                      xdp_frame_has_frags(xdpf)))
 490                 return -EOPNOTSUPP;
 491
 492         err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
 493         if (unlikely(err))
 494                 return err;
 495
 496         bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
 497         return 0;
 498 }
 499
 500 static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
 501 {
 502         struct xdp_txq_info txq = { .dev = dst->dev };
 503         struct xdp_buff xdp;
 504         u32 act;
 505
 506         if (!dst->xdp_prog)
 507                 return XDP_PASS;
 508
 509         __skb_pull(skb, skb->mac_len);
 510         xdp.txq = &txq;
 511
 512         act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
 513         switch (act) {
 514         case XDP_PASS:
 515                 __skb_push(skb, skb->mac_len);
 516                 break;
 517         default:
 518                 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
 519                 fallthrough;
 520         case XDP_ABORTED:
 521                 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
 522                 fallthrough;
 523         case XDP_DROP:
 524                 kfree_skb(skb);
 525                 break;
 526         }
 527
 528         return act;
 529 }
 530
 531 int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 532                     struct net_device *dev_rx)
 533 {
 534         return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
 535 }
 536
 537 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
 538                     struct net_device *dev_rx)
 539 {
 540         struct net_device *dev = dst->dev;
 541
 542         return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
 543 }
 544
 545 static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
 546 {
 547         if (!obj)
 548                 return false;
 549
 550         if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
 551                 return false;
 552
 553         if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
 554                      xdp_frame_has_frags(xdpf)))
 555                 return false;
 556
 557         if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
 558                 return false;
 559
 560         return true;
 561 }
 562
 563 static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
 564                                  struct net_device *dev_rx,
 565                                  struct xdp_frame *xdpf)
 566 {
 567         struct xdp_frame *nxdpf;
 568
 569         nxdpf = xdpf_clone(xdpf);
 570         if (!nxdpf)
 571                 return -ENOMEM;
 572
 573         bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
 574
 575         return 0;
 576 }
 577
 578 static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
 579 {
 580         while (num_excluded--) {
 581                 if (ifindex == excluded[num_excluded])
 582                         return true;
 583         }
 584         return false;
 585 }
 586
 587 /* Get ifindex of each upper device. 'indexes' must be able to hold at
 588  * least MAX_NEST_DEV elements.
 589  * Returns the number of ifindexes added.
 590  */
 591 static int get_upper_ifindexes(struct net_device *dev, int *indexes)
 592 {
 593         struct net_device *upper;
 594         struct list_head *iter;
 595         int n = 0;
 596
 597         netdev_for_each_upper_dev_rcu(dev, upper, iter) {
 598                 indexes[n++] = upper->ifindex;
 599         }
 600         return n;
 601 }
 602
 603 int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
 604                           struct bpf_map *map, bool exclude_ingress)
 605 {
 606         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 607         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 608         int excluded_devices[1+MAX_NEST_DEV];
 609         struct hlist_head *head;
 610         int num_excluded = 0;
 611         unsigned int i;
 612         int err;
 613
 614         if (exclude_ingress) {
 615                 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
 616                 excluded_devices[num_excluded++] = dev_rx->ifindex;
 617         }
 618
 619         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 620                 for (i = 0; i < map->max_entries; i++) {
 621                         dst = rcu_dereference_check(dtab->netdev_map[i],
 622                                                     rcu_read_lock_bh_held());
 623                         if (!is_valid_dst(dst, xdpf))
 624                                 continue;
 625
 626                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 627                                 continue;
 628
 629                         /* we only need n-1 clones; last_dst enqueued below */
 630                         if (!last_dst) {
 631                                 last_dst = dst;
 632                                 continue;
 633                         }
 634
 635                         err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 636                         if (err)
 637                                 return err;
 638
 639                         last_dst = dst;
 640                 }
 641         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 642                 for (i = 0; i < dtab->n_buckets; i++) {
 643                         head = dev_map_index_hash(dtab, i);
 644                         hlist_for_each_entry_rcu(dst, head, index_hlist,
 645                                                  lockdep_is_held(&dtab->index_lock)) {
 646                                 if (!is_valid_dst(dst, xdpf))
 647                                         continue;
 648
 649                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 650                                                         dst->dev->ifindex))
 651                                         continue;
 652
 653                                 /* we only need n-1 clones; last_dst enqueued below */
 654                                 if (!last_dst) {
 655                                         last_dst = dst;
 656                                         continue;
 657                                 }
 658
 659                                 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
 660                                 if (err)
 661                                         return err;
 662
 663                                 last_dst = dst;
 664                         }
 665                 }
 666         }
 667
 668         /* consume the last copy of the frame */
 669         if (last_dst)
 670                 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
 671         else
 672                 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
 673
 674         return 0;
 675 }
 676
 677 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 678                              struct bpf_prog *xdp_prog)
 679 {
 680         int err;
 681
 682         err = xdp_ok_fwd_dev(dst->dev, skb->len);
 683         if (unlikely(err))
 684                 return err;
 685
 686         /* Redirect has already succeeded semantically at this point, so we just
 687          * return 0 even if packet is dropped. Helper below takes care of
 688          * freeing skb.
 689          */
 690         if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
 691                 return 0;
 692
 693         skb->dev = dst->dev;
 694         generic_xdp_tx(skb, xdp_prog);
 695
 696         return 0;
 697 }
 698
 699 static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
 700                                   struct sk_buff *skb,
 701                                   struct bpf_prog *xdp_prog)
 702 {
 703         struct sk_buff *nskb;
 704         int err;
 705
 706         nskb = skb_clone(skb, GFP_ATOMIC);
 707         if (!nskb)
 708                 return -ENOMEM;
 709
 710         err = dev_map_generic_redirect(dst, nskb, xdp_prog);
 711         if (unlikely(err)) {
 712                 consume_skb(nskb);
 713                 return err;
 714         }
 715
 716         return 0;
 717 }
 718
 719 int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
 720                            struct bpf_prog *xdp_prog, struct bpf_map *map,
 721                            bool exclude_ingress)
 722 {
 723         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 724         struct bpf_dtab_netdev *dst, *last_dst = NULL;
 725         int excluded_devices[1+MAX_NEST_DEV];
 726         struct hlist_head *head;
 727         struct hlist_node *next;
 728         int num_excluded = 0;
 729         unsigned int i;
 730         int err;
 731
 732         if (exclude_ingress) {
 733                 num_excluded = get_upper_ifindexes(dev, excluded_devices);
 734                 excluded_devices[num_excluded++] = dev->ifindex;
 735         }
 736
 737         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
 738                 for (i = 0; i < map->max_entries; i++) {
 739                         dst = rcu_dereference_check(dtab->netdev_map[i],
 740                                                     rcu_read_lock_bh_held());
 741                         if (!dst)
 742                                 continue;
 743
 744                         if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
 745                                 continue;
 746
 747                         /* we only need n-1 clones; last_dst enqueued below */
 748                         if (!last_dst) {
 749                                 last_dst = dst;
 750                                 continue;
 751                         }
 752
 753                         err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 754                         if (err)
 755                                 return err;
 756
 757                         last_dst = dst;
 758
 759                 }
 760         } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
 761                 for (i = 0; i < dtab->n_buckets; i++) {
 762                         head = dev_map_index_hash(dtab, i);
 763                         hlist_for_each_entry_safe(dst, next, head, index_hlist) {
 764                                 if (is_ifindex_excluded(excluded_devices, num_excluded,
 765                                                         dst->dev->ifindex))
 766                                         continue;
 767
 768                                 /* we only need n-1 clones; last_dst enqueued below */
 769                                 if (!last_dst) {
 770                                         last_dst = dst;
 771                                         continue;
 772                                 }
 773
 774                                 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
 775                                 if (err)
 776                                         return err;
 777
 778                                 last_dst = dst;
 779                         }
 780                 }
 781         }
 782
 783         /* consume the first skb and return */
 784         if (last_dst)
 785                 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
 786
 787         /* dtab is empty */
 788         consume_skb(skb);
 789         return 0;
 790 }
 791
 792 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 793 {
 794         struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 795
 796         return obj ? &obj->val : NULL;
 797 }
 798
 799 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 800 {
 801         struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 802                                                                 *(u32 *)key);
 803         return obj ? &obj->val : NULL;
 804 }
 805
 806 static void __dev_map_entry_free(struct rcu_head *rcu)
 807 {
 808         struct bpf_dtab_netdev *dev;
 809
 810         dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 811         if (dev->xdp_prog)
 812                 bpf_prog_put(dev->xdp_prog);
 813         dev_put(dev->dev);
 814         kfree(dev);
 815 }
 816
 817 static long dev_map_delete_elem(struct bpf_map *map, void *key)
 818 {
 819         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 820         struct bpf_dtab_netdev *old_dev;
 821         int k = *(u32 *)key;
 822
 823         if (k >= map->max_entries)
 824                 return -EINVAL;
 825
 826         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
 827         if (old_dev) {
 828                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 829                 atomic_dec((atomic_t *)&dtab->items);
 830         }
 831         return 0;
 832 }
 833
 834 static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 835 {
 836         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 837         struct bpf_dtab_netdev *old_dev;
 838         int k = *(u32 *)key;
 839         unsigned long flags;
 840         int ret = -ENOENT;
 841
 842         spin_lock_irqsave(&dtab->index_lock, flags);
 843
 844         old_dev = __dev_map_hash_lookup_elem(map, k);
 845         if (old_dev) {
 846                 dtab->items--;
 847                 hlist_del_init_rcu(&old_dev->index_hlist);
 848                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 849                 ret = 0;
 850         }
 851         spin_unlock_irqrestore(&dtab->index_lock, flags);
 852
 853         return ret;
 854 }
 855
 856 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 857                                                     struct bpf_dtab *dtab,
 858                                                     struct bpf_devmap_val *val,
 859                                                     unsigned int idx)
 860 {
 861         struct bpf_prog *prog = NULL;
 862         struct bpf_dtab_netdev *dev;
 863
 864         dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 865                                    GFP_NOWAIT | __GFP_NOWARN,
 866                                    dtab->map.numa_node);
 867         if (!dev)
 868                 return ERR_PTR(-ENOMEM);
 869
 870         dev->dev = dev_get_by_index(net, val->ifindex);
 871         if (!dev->dev)
 872                 goto err_out;
 873
 874         if (val->bpf_prog.fd > 0) {
 875                 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 876                                              BPF_PROG_TYPE_XDP, false);
 877                 if (IS_ERR(prog))
 878                         goto err_put_dev;
 879                 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
 880                     !bpf_prog_map_compatible(&dtab->map, prog))
 881                         goto err_put_prog;
 882         }
 883
 884         dev->idx = idx;
 885         if (prog) {
 886                 dev->xdp_prog = prog;
 887                 dev->val.bpf_prog.id = prog->aux->id;
 888         } else {
 889                 dev->xdp_prog = NULL;
 890                 dev->val.bpf_prog.id = 0;
 891         }
 892         dev->val.ifindex = val->ifindex;
 893
 894         return dev;
 895 err_put_prog:
 896         bpf_prog_put(prog);
 897 err_put_dev:
 898         dev_put(dev->dev);
 899 err_out:
 900         kfree(dev);
 901         return ERR_PTR(-EINVAL);
 902 }
 903
 904 static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
 905                                   void *key, void *value, u64 map_flags)
 906 {
 907         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 908         struct bpf_dtab_netdev *dev, *old_dev;
 909         struct bpf_devmap_val val = {};
 910         u32 i = *(u32 *)key;
 911
 912         if (unlikely(map_flags > BPF_EXIST))
 913                 return -EINVAL;
 914         if (unlikely(i >= dtab->map.max_entries))
 915                 return -E2BIG;
 916         if (unlikely(map_flags == BPF_NOEXIST))
 917                 return -EEXIST;
 918
 919         /* already verified value_size <= sizeof val */
 920         memcpy(&val, value, map->value_size);
 921
 922         if (!val.ifindex) {
 923                 dev = NULL;
 924                 /* can not specify fd if ifindex is 0 */
 925                 if (val.bpf_prog.fd > 0)
 926                         return -EINVAL;
 927         } else {
 928                 dev = __dev_map_alloc_node(net, dtab, &val, i);
 929                 if (IS_ERR(dev))
 930                         return PTR_ERR(dev);
 931         }
 932
 933         /* Use call_rcu() here to ensure rcu critical sections have completed
 934          * Remembering the driver side flush operation will happen before the
 935          * net device is removed.
 936          */
 937         old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
 938         if (old_dev)
 939                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 940         else
 941                 atomic_inc((atomic_t *)&dtab->items);
 942
 943         return 0;
 944 }
 945
 946 static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 947                                 u64 map_flags)
 948 {
 949         return __dev_map_update_elem(current->nsproxy->net_ns,
 950                                      map, key, value, map_flags);
 951 }
 952
 953 static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 954                                        void *key, void *value, u64 map_flags)
 955 {
 956         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 957         struct bpf_dtab_netdev *dev, *old_dev;
 958         struct bpf_devmap_val val = {};
 959         u32 idx = *(u32 *)key;
 960         unsigned long flags;
 961         int err = -EEXIST;
 962
 963         /* already verified value_size <= sizeof val */
 964         memcpy(&val, value, map->value_size);
 965
 966         if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 967                 return -EINVAL;
 968
 969         spin_lock_irqsave(&dtab->index_lock, flags);
 970
 971         old_dev = __dev_map_hash_lookup_elem(map, idx);
 972         if (old_dev && (map_flags & BPF_NOEXIST))
 973                 goto out_err;
 974
 975         dev = __dev_map_alloc_node(net, dtab, &val, idx);
 976         if (IS_ERR(dev)) {
 977                 err = PTR_ERR(dev);
 978                 goto out_err;
 979         }
 980
 981         if (old_dev) {
 982                 hlist_del_rcu(&old_dev->index_hlist);
 983         } else {
 984                 if (dtab->items >= dtab->map.max_entries) {
 985                         spin_unlock_irqrestore(&dtab->index_lock, flags);
 986                         call_rcu(&dev->rcu, __dev_map_entry_free);
 987                         return -E2BIG;
 988                 }
 989                 dtab->items++;
 990         }
 991
 992         hlist_add_head_rcu(&dev->index_hlist,
 993                            dev_map_index_hash(dtab, idx));
 994         spin_unlock_irqrestore(&dtab->index_lock, flags);
 995
 996         if (old_dev)
 997                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 998
 999         return 0;
1000
1001 out_err:
1002         spin_unlock_irqrestore(&dtab->index_lock, flags);
1003         return err;
1004 }
1005
1006 static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1007                                      u64 map_flags)
1008 {
1009         return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1010                                          map, key, value, map_flags);
1011 }
1012
1013 static long dev_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_lookup_elem);
1018 }
1019
1020 static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1021 {
1022         return __bpf_xdp_redirect_map(map, ifindex, flags,
1023                                       BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1024                                       __dev_map_hash_lookup_elem);
1025 }
1026
1027 static u64 dev_map_mem_usage(const struct bpf_map *map)
1028 {
1029         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1030         u64 usage = sizeof(struct bpf_dtab);
1031
1032         if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1033                 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1034         else
1035                 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1036         usage += atomic_read((atomic_t *)&dtab->items) *
1037                          (u64)sizeof(struct bpf_dtab_netdev);
1038         return usage;
1039 }
1040
1041 BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1042 const struct bpf_map_ops dev_map_ops = {
1043         .map_meta_equal = bpf_map_meta_equal,
1044         .map_alloc_check = dev_map_alloc_check,
1045         .map_alloc = dev_map_alloc,
1046         .map_free = dev_map_free,
1047         .map_get_next_key = dev_map_get_next_key,
1048         .map_lookup_elem = dev_map_lookup_elem,
1049         .map_update_elem = dev_map_update_elem,
1050         .map_delete_elem = dev_map_delete_elem,
1051         .map_check_btf = map_check_no_btf,
1052         .map_mem_usage = dev_map_mem_usage,
1053         .map_btf_id = &dev_map_btf_ids[0],
1054         .map_redirect = dev_map_redirect,
1055 };
1056
1057 const struct bpf_map_ops dev_map_hash_ops = {
1058         .map_meta_equal = bpf_map_meta_equal,
1059         .map_alloc_check = dev_map_alloc_check,
1060         .map_alloc = dev_map_alloc,
1061         .map_free = dev_map_free,
1062         .map_get_next_key = dev_map_hash_get_next_key,
1063         .map_lookup_elem = dev_map_hash_lookup_elem,
1064         .map_update_elem = dev_map_hash_update_elem,
1065         .map_delete_elem = dev_map_hash_delete_elem,
1066         .map_check_btf = map_check_no_btf,
1067         .map_mem_usage = dev_map_mem_usage,
1068         .map_btf_id = &dev_map_btf_ids[0],
1069         .map_redirect = dev_hash_map_redirect,
1070 };
1071
1072 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1073                                        struct net_device *netdev)
1074 {
1075         unsigned long flags;
1076         u32 i;
1077
1078         spin_lock_irqsave(&dtab->index_lock, flags);
1079         for (i = 0; i < dtab->n_buckets; i++) {
1080                 struct bpf_dtab_netdev *dev;
1081                 struct hlist_head *head;
1082                 struct hlist_node *next;
1083
1084                 head = dev_map_index_hash(dtab, i);
1085
1086                 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1087                         if (netdev != dev->dev)
1088                                 continue;
1089
1090                         dtab->items--;
1091                         hlist_del_rcu(&dev->index_hlist);
1092                         call_rcu(&dev->rcu, __dev_map_entry_free);
1093                 }
1094         }
1095         spin_unlock_irqrestore(&dtab->index_lock, flags);
1096 }
1097
1098 static int dev_map_notification(struct notifier_block *notifier,
1099                                 ulong event, void *ptr)
1100 {
1101         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1102         struct bpf_dtab *dtab;
1103         int i, cpu;
1104
1105         switch (event) {
1106         case NETDEV_REGISTER:
1107                 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1108                         break;
1109
1110                 /* will be freed in free_netdev() */
1111                 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1112                 if (!netdev->xdp_bulkq)
1113                         return NOTIFY_BAD;
1114
1115                 for_each_possible_cpu(cpu)
1116                         per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1117                 break;
1118         case NETDEV_UNREGISTER:
1119                 /* This rcu_read_lock/unlock pair is needed because
1120                  * dev_map_list is an RCU list AND to ensure a delete
1121                  * operation does not free a netdev_map entry while we
1122                  * are comparing it against the netdev being unregistered.
1123                  */
1124                 rcu_read_lock();
1125                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1126                         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1127                                 dev_map_hash_remove_netdev(dtab, netdev);
1128                                 continue;
1129                         }
1130
1131                         for (i = 0; i < dtab->map.max_entries; i++) {
1132                                 struct bpf_dtab_netdev *dev, *odev;
1133
1134                                 dev = rcu_dereference(dtab->netdev_map[i]);
1135                                 if (!dev || netdev != dev->dev)
1136                                         continue;
1137                                 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1138                                 if (dev == odev) {
1139                                         call_rcu(&dev->rcu,
1140                                                  __dev_map_entry_free);
1141                                         atomic_dec((atomic_t *)&dtab->items);
1142                                 }
1143                         }
1144                 }
1145                 rcu_read_unlock();
1146                 break;
1147         default:
1148                 break;
1149         }
1150         return NOTIFY_OK;
1151 }
1152
1153 static struct notifier_block dev_map_notifier = {
1154         .notifier_call = dev_map_notification,
1155 };
1156
1157 static int __init dev_map_init(void)
1158 {
1159         /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1160         BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1161                      offsetof(struct _bpf_dtab_netdev, dev));
1162         register_netdevice_notifier(&dev_map_notifier);
1163
1164         return 0;
1165 }
1166
1167 subsys_initcall(dev_map_init);