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