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
  52 #define DEV_CREATE_FLAG_MASK \
  53         (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  54
  55 struct xdp_dev_bulk_queue {
  56         struct xdp_frame *q[DEV_MAP_BULK_SIZE];
  57         struct list_head flush_node;
  58         struct net_device *dev;
  59         struct net_device *dev_rx;
  60         unsigned int count;
  61 };
  62
  63 struct bpf_dtab_netdev {
  64         struct net_device *dev; /* must be first member, due to tracepoint */
  65         struct hlist_node index_hlist;
  66         struct bpf_dtab *dtab;
  67         struct bpf_prog *xdp_prog;
  68         struct rcu_head rcu;
  69         unsigned int idx;
  70         struct bpf_devmap_val val;
  71 };
  72
  73 struct bpf_dtab {
  74         struct bpf_map map;
  75         struct bpf_dtab_netdev **netdev_map; /* DEVMAP type only */
  76         struct list_head list;
  77
  78         /* these are only used for DEVMAP_HASH type maps */
  79         struct hlist_head *dev_index_head;
  80         spinlock_t index_lock;
  81         unsigned int items;
  82         u32 n_buckets;
  83 };
  84
  85 static DEFINE_PER_CPU(struct list_head, dev_flush_list);
  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(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_init_map(struct bpf_dtab *dtab, 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         /* Lookup returns a pointer straight to dev->ifindex, so make sure the
 124          * verifier prevents writes from the BPF side
 125          */
 126         attr->map_flags |= BPF_F_RDONLY_PROG;
 127
 128
 129         bpf_map_init_from_attr(&dtab->map, attr);
 130
 131         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 132                 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
 133
 134                 if (!dtab->n_buckets) /* Overflow check */
 135                         return -EINVAL;
 136         }
 137
 138         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 139                 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
 140                                                            dtab->map.numa_node);
 141                 if (!dtab->dev_index_head)
 142                         return -ENOMEM;
 143
 144                 spin_lock_init(&dtab->index_lock);
 145         } else {
 146                 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
 147                                                       sizeof(struct bpf_dtab_netdev *),
 148                                                       dtab->map.numa_node);
 149                 if (!dtab->netdev_map)
 150                         return -ENOMEM;
 151         }
 152
 153         return 0;
 154 }
 155
 156 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 157 {
 158         struct bpf_dtab *dtab;
 159         int err;
 160
 161         if (!capable(CAP_NET_ADMIN))
 162                 return ERR_PTR(-EPERM);
 163
 164         dtab = kzalloc(sizeof(*dtab), GFP_USER | __GFP_ACCOUNT);
 165         if (!dtab)
 166                 return ERR_PTR(-ENOMEM);
 167
 168         err = dev_map_init_map(dtab, attr);
 169         if (err) {
 170                 kfree(dtab);
 171                 return ERR_PTR(err);
 172         }
 173
 174         spin_lock(&dev_map_lock);
 175         list_add_tail_rcu(&dtab->list, &dev_map_list);
 176         spin_unlock(&dev_map_lock);
 177
 178         return &dtab->map;
 179 }
 180
 181 static void dev_map_free(struct bpf_map *map)
 182 {
 183         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 184         int i;
 185
 186         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 187          * so the programs (can be more than one that used this map) were
 188          * disconnected from events. The following synchronize_rcu() guarantees
 189          * both rcu read critical sections complete and waits for
 190          * preempt-disable regions (NAPI being the relevant context here) so we
 191          * are certain there will be no further reads against the netdev_map and
 192          * all flush operations are complete. Flush operations can only be done
 193          * from NAPI context for this reason.
 194          */
 195
 196         spin_lock(&dev_map_lock);
 197         list_del_rcu(&dtab->list);
 198         spin_unlock(&dev_map_lock);
 199
 200         bpf_clear_redirect_map(map);
 201         synchronize_rcu();
 202
 203         /* Make sure prior __dev_map_entry_free() have completed. */
 204         rcu_barrier();
 205
 206         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 207                 for (i = 0; i < dtab->n_buckets; i++) {
 208                         struct bpf_dtab_netdev *dev;
 209                         struct hlist_head *head;
 210                         struct hlist_node *next;
 211
 212                         head = dev_map_index_hash(dtab, i);
 213
 214                         hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 215                                 hlist_del_rcu(&dev->index_hlist);
 216                                 if (dev->xdp_prog)
 217                                         bpf_prog_put(dev->xdp_prog);
 218                                 dev_put(dev->dev);
 219                                 kfree(dev);
 220                         }
 221                 }
 222
 223                 bpf_map_area_free(dtab->dev_index_head);
 224         } else {
 225                 for (i = 0; i < dtab->map.max_entries; i++) {
 226                         struct bpf_dtab_netdev *dev;
 227
 228                         dev = dtab->netdev_map[i];
 229                         if (!dev)
 230                                 continue;
 231
 232                         if (dev->xdp_prog)
 233                                 bpf_prog_put(dev->xdp_prog);
 234                         dev_put(dev->dev);
 235                         kfree(dev);
 236                 }
 237
 238                 bpf_map_area_free(dtab->netdev_map);
 239         }
 240
 241         kfree(dtab);
 242 }
 243
 244 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 245 {
 246         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 247         u32 index = key ? *(u32 *)key : U32_MAX;
 248         u32 *next = next_key;
 249
 250         if (index >= dtab->map.max_entries) {
 251                 *next = 0;
 252                 return 0;
 253         }
 254
 255         if (index == dtab->map.max_entries - 1)
 256                 return -ENOENT;
 257         *next = index + 1;
 258         return 0;
 259 }
 260
 261 struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 262 {
 263         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 264         struct hlist_head *head = dev_map_index_hash(dtab, key);
 265         struct bpf_dtab_netdev *dev;
 266
 267         hlist_for_each_entry_rcu(dev, head, index_hlist,
 268                                  lockdep_is_held(&dtab->index_lock))
 269                 if (dev->idx == key)
 270                         return dev;
 271
 272         return NULL;
 273 }
 274
 275 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 276                                     void *next_key)
 277 {
 278         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 279         u32 idx, *next = next_key;
 280         struct bpf_dtab_netdev *dev, *next_dev;
 281         struct hlist_head *head;
 282         int i = 0;
 283
 284         if (!key)
 285                 goto find_first;
 286
 287         idx = *(u32 *)key;
 288
 289         dev = __dev_map_hash_lookup_elem(map, idx);
 290         if (!dev)
 291                 goto find_first;
 292
 293         next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 294                                     struct bpf_dtab_netdev, index_hlist);
 295
 296         if (next_dev) {
 297                 *next = next_dev->idx;
 298                 return 0;
 299         }
 300
 301         i = idx & (dtab->n_buckets - 1);
 302         i++;
 303
 304  find_first:
 305         for (; i < dtab->n_buckets; i++) {
 306                 head = dev_map_index_hash(dtab, i);
 307
 308                 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 309                                             struct bpf_dtab_netdev,
 310                                             index_hlist);
 311                 if (next_dev) {
 312                         *next = next_dev->idx;
 313                         return 0;
 314                 }
 315         }
 316
 317         return -ENOENT;
 318 }
 319
 320 bool dev_map_can_have_prog(struct bpf_map *map)
 321 {
 322         if ((map->map_type == BPF_MAP_TYPE_DEVMAP ||
 323              map->map_type == BPF_MAP_TYPE_DEVMAP_HASH) &&
 324             map->value_size != offsetofend(struct bpf_devmap_val, ifindex))
 325                 return true;
 326
 327         return false;
 328 }
 329
 330 static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
 331 {
 332         struct net_device *dev = bq->dev;
 333         int sent = 0, drops = 0, err = 0;
 334         int i;
 335
 336         if (unlikely(!bq->count))
 337                 return;
 338
 339         for (i = 0; i < bq->count; i++) {
 340                 struct xdp_frame *xdpf = bq->q[i];
 341
 342                 prefetch(xdpf);
 343         }
 344
 345         sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
 346         if (sent < 0) {
 347                 err = sent;
 348                 sent = 0;
 349                 goto error;
 350         }
 351         drops = bq->count - sent;
 352 out:
 353         bq->count = 0;
 354
 355         trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, drops, err);
 356         bq->dev_rx = NULL;
 357         __list_del_clearprev(&bq->flush_node);
 358         return;
 359 error:
 360         /* If ndo_xdp_xmit fails with an errno, no frames have been
 361          * xmit'ed and it's our responsibility to them free all.
 362          */
 363         for (i = 0; i < bq->count; i++) {
 364                 struct xdp_frame *xdpf = bq->q[i];
 365
 366                 xdp_return_frame_rx_napi(xdpf);
 367                 drops++;
 368         }
 369         goto out;
 370 }
 371
 372 /* __dev_flush is called from xdp_do_flush() which _must_ be signaled
 373  * from the driver before returning from its napi->poll() routine. The poll()
 374  * routine is called either from busy_poll context or net_rx_action signaled
 375  * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
 376  * net device can be torn down. On devmap tear down we ensure the flush list
 377  * is empty before completing to ensure all flush operations have completed.
 378  * When drivers update the bpf program they may need to ensure any flush ops
 379  * are also complete. Using synchronize_rcu or call_rcu will suffice for this
 380  * because both wait for napi context to exit.
 381  */
 382 void __dev_flush(void)
 383 {
 384         struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 385         struct xdp_dev_bulk_queue *bq, *tmp;
 386
 387         list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
 388                 bq_xmit_all(bq, XDP_XMIT_FLUSH);
 389 }
 390
 391 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
 392  * update happens in parallel here a dev_put wont happen until after reading the
 393  * ifindex.
 394  */
 395 struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 396 {
 397         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 398         struct bpf_dtab_netdev *obj;
 399
 400         if (key >= map->max_entries)
 401                 return NULL;
 402
 403         obj = READ_ONCE(dtab->netdev_map[key]);
 404         return obj;
 405 }
 406
 407 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 408  * Thus, safe percpu variable access.
 409  */
 410 static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
 411                        struct net_device *dev_rx)
 412 {
 413         struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
 414         struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
 415
 416         if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 417                 bq_xmit_all(bq, 0);
 418
 419         /* Ingress dev_rx will be the same for all xdp_frame's in
 420          * bulk_queue, because bq stored per-CPU and must be flushed
 421          * from net_device drivers NAPI func end.
 422          */
 423         if (!bq->dev_rx)
 424                 bq->dev_rx = dev_rx;
 425
 426         bq->q[bq->count++] = xdpf;
 427
 428         if (!bq->flush_node.prev)
 429                 list_add(&bq->flush_node, flush_list);
 430 }
 431
 432 static inline int __xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
 433                                struct net_device *dev_rx)
 434 {
 435         struct xdp_frame *xdpf;
 436         int err;
 437
 438         if (!dev->netdev_ops->ndo_xdp_xmit)
 439                 return -EOPNOTSUPP;
 440
 441         err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
 442         if (unlikely(err))
 443                 return err;
 444
 445         xdpf = xdp_convert_buff_to_frame(xdp);
 446         if (unlikely(!xdpf))
 447                 return -EOVERFLOW;
 448
 449         bq_enqueue(dev, xdpf, dev_rx);
 450         return 0;
 451 }
 452
 453 static struct xdp_buff *dev_map_run_prog(struct net_device *dev,
 454                                          struct xdp_buff *xdp,
 455                                          struct bpf_prog *xdp_prog)
 456 {
 457         struct xdp_txq_info txq = { .dev = dev };
 458         u32 act;
 459
 460         xdp_set_data_meta_invalid(xdp);
 461         xdp->txq = &txq;
 462
 463         act = bpf_prog_run_xdp(xdp_prog, xdp);
 464         switch (act) {
 465         case XDP_PASS:
 466                 return xdp;
 467         case XDP_DROP:
 468                 break;
 469         default:
 470                 bpf_warn_invalid_xdp_action(act);
 471                 fallthrough;
 472         case XDP_ABORTED:
 473                 trace_xdp_exception(dev, xdp_prog, act);
 474                 break;
 475         }
 476
 477         xdp_return_buff(xdp);
 478         return NULL;
 479 }
 480
 481 int dev_xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
 482                     struct net_device *dev_rx)
 483 {
 484         return __xdp_enqueue(dev, xdp, dev_rx);
 485 }
 486
 487 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
 488                     struct net_device *dev_rx)
 489 {
 490         struct net_device *dev = dst->dev;
 491
 492         if (dst->xdp_prog) {
 493                 xdp = dev_map_run_prog(dev, xdp, dst->xdp_prog);
 494                 if (!xdp)
 495                         return 0;
 496         }
 497         return __xdp_enqueue(dev, xdp, dev_rx);
 498 }
 499
 500 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 501                              struct bpf_prog *xdp_prog)
 502 {
 503         int err;
 504
 505         err = xdp_ok_fwd_dev(dst->dev, skb->len);
 506         if (unlikely(err))
 507                 return err;
 508         skb->dev = dst->dev;
 509         generic_xdp_tx(skb, xdp_prog);
 510
 511         return 0;
 512 }
 513
 514 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 515 {
 516         struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 517
 518         return obj ? &obj->val : NULL;
 519 }
 520
 521 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 522 {
 523         struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 524                                                                 *(u32 *)key);
 525         return obj ? &obj->val : NULL;
 526 }
 527
 528 static void __dev_map_entry_free(struct rcu_head *rcu)
 529 {
 530         struct bpf_dtab_netdev *dev;
 531
 532         dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
 533         if (dev->xdp_prog)
 534                 bpf_prog_put(dev->xdp_prog);
 535         dev_put(dev->dev);
 536         kfree(dev);
 537 }
 538
 539 static int dev_map_delete_elem(struct bpf_map *map, void *key)
 540 {
 541         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 542         struct bpf_dtab_netdev *old_dev;
 543         int k = *(u32 *)key;
 544
 545         if (k >= map->max_entries)
 546                 return -EINVAL;
 547
 548         /* Use call_rcu() here to ensure any rcu critical sections have
 549          * completed as well as any flush operations because call_rcu
 550          * will wait for preempt-disable region to complete, NAPI in this
 551          * context.  And additionally, the driver tear down ensures all
 552          * soft irqs are complete before removing the net device in the
 553          * case of dev_put equals zero.
 554          */
 555         old_dev = xchg(&dtab->netdev_map[k], NULL);
 556         if (old_dev)
 557                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 558         return 0;
 559 }
 560
 561 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 562 {
 563         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 564         struct bpf_dtab_netdev *old_dev;
 565         int k = *(u32 *)key;
 566         unsigned long flags;
 567         int ret = -ENOENT;
 568
 569         spin_lock_irqsave(&dtab->index_lock, flags);
 570
 571         old_dev = __dev_map_hash_lookup_elem(map, k);
 572         if (old_dev) {
 573                 dtab->items--;
 574                 hlist_del_init_rcu(&old_dev->index_hlist);
 575                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 576                 ret = 0;
 577         }
 578         spin_unlock_irqrestore(&dtab->index_lock, flags);
 579
 580         return ret;
 581 }
 582
 583 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 584                                                     struct bpf_dtab *dtab,
 585                                                     struct bpf_devmap_val *val,
 586                                                     unsigned int idx)
 587 {
 588         struct bpf_prog *prog = NULL;
 589         struct bpf_dtab_netdev *dev;
 590
 591         dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
 592                                    GFP_ATOMIC | __GFP_NOWARN,
 593                                    dtab->map.numa_node);
 594         if (!dev)
 595                 return ERR_PTR(-ENOMEM);
 596
 597         dev->dev = dev_get_by_index(net, val->ifindex);
 598         if (!dev->dev)
 599                 goto err_out;
 600
 601         if (val->bpf_prog.fd > 0) {
 602                 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
 603                                              BPF_PROG_TYPE_XDP, false);
 604                 if (IS_ERR(prog))
 605                         goto err_put_dev;
 606                 if (prog->expected_attach_type != BPF_XDP_DEVMAP)
 607                         goto err_put_prog;
 608         }
 609
 610         dev->idx = idx;
 611         dev->dtab = dtab;
 612         if (prog) {
 613                 dev->xdp_prog = prog;
 614                 dev->val.bpf_prog.id = prog->aux->id;
 615         } else {
 616                 dev->xdp_prog = NULL;
 617                 dev->val.bpf_prog.id = 0;
 618         }
 619         dev->val.ifindex = val->ifindex;
 620
 621         return dev;
 622 err_put_prog:
 623         bpf_prog_put(prog);
 624 err_put_dev:
 625         dev_put(dev->dev);
 626 err_out:
 627         kfree(dev);
 628         return ERR_PTR(-EINVAL);
 629 }
 630
 631 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
 632                                  void *key, void *value, u64 map_flags)
 633 {
 634         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 635         struct bpf_dtab_netdev *dev, *old_dev;
 636         struct bpf_devmap_val val = {};
 637         u32 i = *(u32 *)key;
 638
 639         if (unlikely(map_flags > BPF_EXIST))
 640                 return -EINVAL;
 641         if (unlikely(i >= dtab->map.max_entries))
 642                 return -E2BIG;
 643         if (unlikely(map_flags == BPF_NOEXIST))
 644                 return -EEXIST;
 645
 646         /* already verified value_size <= sizeof val */
 647         memcpy(&val, value, map->value_size);
 648
 649         if (!val.ifindex) {
 650                 dev = NULL;
 651                 /* can not specify fd if ifindex is 0 */
 652                 if (val.bpf_prog.fd > 0)
 653                         return -EINVAL;
 654         } else {
 655                 dev = __dev_map_alloc_node(net, dtab, &val, i);
 656                 if (IS_ERR(dev))
 657                         return PTR_ERR(dev);
 658         }
 659
 660         /* Use call_rcu() here to ensure rcu critical sections have completed
 661          * Remembering the driver side flush operation will happen before the
 662          * net device is removed.
 663          */
 664         old_dev = xchg(&dtab->netdev_map[i], dev);
 665         if (old_dev)
 666                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 667
 668         return 0;
 669 }
 670
 671 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 672                                u64 map_flags)
 673 {
 674         return __dev_map_update_elem(current->nsproxy->net_ns,
 675                                      map, key, value, map_flags);
 676 }
 677
 678 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 679                                      void *key, void *value, u64 map_flags)
 680 {
 681         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 682         struct bpf_dtab_netdev *dev, *old_dev;
 683         struct bpf_devmap_val val = {};
 684         u32 idx = *(u32 *)key;
 685         unsigned long flags;
 686         int err = -EEXIST;
 687
 688         /* already verified value_size <= sizeof val */
 689         memcpy(&val, value, map->value_size);
 690
 691         if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
 692                 return -EINVAL;
 693
 694         spin_lock_irqsave(&dtab->index_lock, flags);
 695
 696         old_dev = __dev_map_hash_lookup_elem(map, idx);
 697         if (old_dev && (map_flags & BPF_NOEXIST))
 698                 goto out_err;
 699
 700         dev = __dev_map_alloc_node(net, dtab, &val, idx);
 701         if (IS_ERR(dev)) {
 702                 err = PTR_ERR(dev);
 703                 goto out_err;
 704         }
 705
 706         if (old_dev) {
 707                 hlist_del_rcu(&old_dev->index_hlist);
 708         } else {
 709                 if (dtab->items >= dtab->map.max_entries) {
 710                         spin_unlock_irqrestore(&dtab->index_lock, flags);
 711                         call_rcu(&dev->rcu, __dev_map_entry_free);
 712                         return -E2BIG;
 713                 }
 714                 dtab->items++;
 715         }
 716
 717         hlist_add_head_rcu(&dev->index_hlist,
 718                            dev_map_index_hash(dtab, idx));
 719         spin_unlock_irqrestore(&dtab->index_lock, flags);
 720
 721         if (old_dev)
 722                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 723
 724         return 0;
 725
 726 out_err:
 727         spin_unlock_irqrestore(&dtab->index_lock, flags);
 728         return err;
 729 }
 730
 731 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
 732                                    u64 map_flags)
 733 {
 734         return __dev_map_hash_update_elem(current->nsproxy->net_ns,
 735                                          map, key, value, map_flags);
 736 }
 737
 738 static int dev_map_btf_id;
 739 const struct bpf_map_ops dev_map_ops = {
 740         .map_meta_equal = bpf_map_meta_equal,
 741         .map_alloc = dev_map_alloc,
 742         .map_free = dev_map_free,
 743         .map_get_next_key = dev_map_get_next_key,
 744         .map_lookup_elem = dev_map_lookup_elem,
 745         .map_update_elem = dev_map_update_elem,
 746         .map_delete_elem = dev_map_delete_elem,
 747         .map_check_btf = map_check_no_btf,
 748         .map_btf_name = "bpf_dtab",
 749         .map_btf_id = &dev_map_btf_id,
 750 };
 751
 752 static int dev_map_hash_map_btf_id;
 753 const struct bpf_map_ops dev_map_hash_ops = {
 754         .map_meta_equal = bpf_map_meta_equal,
 755         .map_alloc = dev_map_alloc,
 756         .map_free = dev_map_free,
 757         .map_get_next_key = dev_map_hash_get_next_key,
 758         .map_lookup_elem = dev_map_hash_lookup_elem,
 759         .map_update_elem = dev_map_hash_update_elem,
 760         .map_delete_elem = dev_map_hash_delete_elem,
 761         .map_check_btf = map_check_no_btf,
 762         .map_btf_name = "bpf_dtab",
 763         .map_btf_id = &dev_map_hash_map_btf_id,
 764 };
 765
 766 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
 767                                        struct net_device *netdev)
 768 {
 769         unsigned long flags;
 770         u32 i;
 771
 772         spin_lock_irqsave(&dtab->index_lock, flags);
 773         for (i = 0; i < dtab->n_buckets; i++) {
 774                 struct bpf_dtab_netdev *dev;
 775                 struct hlist_head *head;
 776                 struct hlist_node *next;
 777
 778                 head = dev_map_index_hash(dtab, i);
 779
 780                 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 781                         if (netdev != dev->dev)
 782                                 continue;
 783
 784                         dtab->items--;
 785                         hlist_del_rcu(&dev->index_hlist);
 786                         call_rcu(&dev->rcu, __dev_map_entry_free);
 787                 }
 788         }
 789         spin_unlock_irqrestore(&dtab->index_lock, flags);
 790 }
 791
 792 static int dev_map_notification(struct notifier_block *notifier,
 793                                 ulong event, void *ptr)
 794 {
 795         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
 796         struct bpf_dtab *dtab;
 797         int i, cpu;
 798
 799         switch (event) {
 800         case NETDEV_REGISTER:
 801                 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
 802                         break;
 803
 804                 /* will be freed in free_netdev() */
 805                 netdev->xdp_bulkq =
 806                         __alloc_percpu_gfp(sizeof(struct xdp_dev_bulk_queue),
 807                                            sizeof(void *), GFP_ATOMIC);
 808                 if (!netdev->xdp_bulkq)
 809                         return NOTIFY_BAD;
 810
 811                 for_each_possible_cpu(cpu)
 812                         per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
 813                 break;
 814         case NETDEV_UNREGISTER:
 815                 /* This rcu_read_lock/unlock pair is needed because
 816                  * dev_map_list is an RCU list AND to ensure a delete
 817                  * operation does not free a netdev_map entry while we
 818                  * are comparing it against the netdev being unregistered.
 819                  */
 820                 rcu_read_lock();
 821                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
 822                         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 823                                 dev_map_hash_remove_netdev(dtab, netdev);
 824                                 continue;
 825                         }
 826
 827                         for (i = 0; i < dtab->map.max_entries; i++) {
 828                                 struct bpf_dtab_netdev *dev, *odev;
 829
 830                                 dev = READ_ONCE(dtab->netdev_map[i]);
 831                                 if (!dev || netdev != dev->dev)
 832                                         continue;
 833                                 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
 834                                 if (dev == odev)
 835                                         call_rcu(&dev->rcu,
 836                                                  __dev_map_entry_free);
 837                         }
 838                 }
 839                 rcu_read_unlock();
 840                 break;
 841         default:
 842                 break;
 843         }
 844         return NOTIFY_OK;
 845 }
 846
 847 static struct notifier_block dev_map_notifier = {
 848         .notifier_call = dev_map_notification,
 849 };
 850
 851 static int __init dev_map_init(void)
 852 {
 853         int cpu;
 854
 855         /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
 856         BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
 857                      offsetof(struct _bpf_dtab_netdev, dev));
 858         register_netdevice_notifier(&dev_map_notifier);
 859
 860         for_each_possible_cpu(cpu)
 861                 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
 862         return 0;
 863 }
 864
 865 subsys_initcall(dev_map_init);