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