root/kernel/bpf/devmap.c

DEFINITIONS

1. dev_map_create_hash
2. dev_map_index_hash
3. dev_map_init_map
4. dev_map_alloc
5. dev_map_free
6. dev_map_get_next_key
7. __dev_map_hash_lookup_elem
8. dev_map_hash_get_next_key
9. bq_xmit_all
10. __dev_map_flush
11. __dev_map_lookup_elem
12. bq_enqueue
13. dev_map_enqueue
14. dev_map_generic_redirect
15. dev_map_lookup_elem
16. dev_map_hash_lookup_elem
17. dev_map_flush_old
18. __dev_map_entry_free
19. dev_map_delete_elem
20. dev_map_hash_delete_elem
21. __dev_map_alloc_node
22. __dev_map_update_elem
23. dev_map_update_elem
24. __dev_map_hash_update_elem
25. dev_map_hash_update_elem
26. dev_map_hash_remove_netdev
27. dev_map_notification
28. dev_map_init

   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 #define DEV_MAP_BULK_SIZE 16
  56 struct bpf_dtab_netdev;
  57 
  58 struct xdp_bulk_queue {
  59         struct xdp_frame *q[DEV_MAP_BULK_SIZE];
  60         struct list_head flush_node;
  61         struct net_device *dev_rx;
  62         struct bpf_dtab_netdev *obj;
  63         unsigned int count;
  64 };
  65 
  66 struct bpf_dtab_netdev {
  67         struct net_device *dev; /* must be first member, due to tracepoint */
  68         struct hlist_node index_hlist;
  69         struct bpf_dtab *dtab;
  70         struct xdp_bulk_queue __percpu *bulkq;
  71         struct rcu_head rcu;
  72         unsigned int idx; /* keep track of map index for tracepoint */
  73 };
  74 
  75 struct bpf_dtab {
  76         struct bpf_map map;
  77         struct bpf_dtab_netdev **netdev_map; /* DEVMAP type only */
  78         struct list_head __percpu *flush_list;
  79         struct list_head list;
  80 
  81         /* these are only used for DEVMAP_HASH type maps */
  82         struct hlist_head *dev_index_head;
  83         spinlock_t index_lock;
  84         unsigned int items;
  85         u32 n_buckets;
  86 };
  87 
  88 static DEFINE_SPINLOCK(dev_map_lock);
  89 static LIST_HEAD(dev_map_list);
  90 
  91 static struct hlist_head *dev_map_create_hash(unsigned int entries)
  92 {
  93         int i;
  94         struct hlist_head *hash;
  95 
  96         hash = kmalloc_array(entries, sizeof(*hash), GFP_KERNEL);
  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         int err, cpu;
 113         u64 cost;
 114 
 115         /* check sanity of attributes */
 116         if (attr->max_entries == 0 || attr->key_size != 4 ||
 117             attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
 118                 return -EINVAL;
 119 
 120         /* Lookup returns a pointer straight to dev->ifindex, so make sure the
 121          * verifier prevents writes from the BPF side
 122          */
 123         attr->map_flags |= BPF_F_RDONLY_PROG;
 124 
 125 
 126         bpf_map_init_from_attr(&dtab->map, attr);
 127 
 128         /* make sure page count doesn't overflow */
 129         cost = (u64) sizeof(struct list_head) * num_possible_cpus();
 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                 cost += (u64) sizeof(struct hlist_head) * dtab->n_buckets;
 137         } else {
 138                 cost += (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
 139         }
 140 
 141         /* if map size is larger than memlock limit, reject it */
 142         err = bpf_map_charge_init(&dtab->map.memory, cost);
 143         if (err)
 144                 return -EINVAL;
 145 
 146         dtab->flush_list = alloc_percpu(struct list_head);
 147         if (!dtab->flush_list)
 148                 goto free_charge;
 149 
 150         for_each_possible_cpu(cpu)
 151                 INIT_LIST_HEAD(per_cpu_ptr(dtab->flush_list, cpu));
 152 
 153         if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 154                 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets);
 155                 if (!dtab->dev_index_head)
 156                         goto free_percpu;
 157 
 158                 spin_lock_init(&dtab->index_lock);
 159         } else {
 160                 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
 161                                                       sizeof(struct bpf_dtab_netdev *),
 162                                                       dtab->map.numa_node);
 163                 if (!dtab->netdev_map)
 164                         goto free_percpu;
 165         }
 166 
 167         return 0;
 168 
 169 free_percpu:
 170         free_percpu(dtab->flush_list);
 171 free_charge:
 172         bpf_map_charge_finish(&dtab->map.memory);
 173         return -ENOMEM;
 174 }
 175 
 176 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
 177 {
 178         struct bpf_dtab *dtab;
 179         int err;
 180 
 181         if (!capable(CAP_NET_ADMIN))
 182                 return ERR_PTR(-EPERM);
 183 
 184         dtab = kzalloc(sizeof(*dtab), GFP_USER);
 185         if (!dtab)
 186                 return ERR_PTR(-ENOMEM);
 187 
 188         err = dev_map_init_map(dtab, attr);
 189         if (err) {
 190                 kfree(dtab);
 191                 return ERR_PTR(err);
 192         }
 193 
 194         spin_lock(&dev_map_lock);
 195         list_add_tail_rcu(&dtab->list, &dev_map_list);
 196         spin_unlock(&dev_map_lock);
 197 
 198         return &dtab->map;
 199 }
 200 
 201 static void dev_map_free(struct bpf_map *map)
 202 {
 203         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 204         int i, cpu;
 205 
 206         /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
 207          * so the programs (can be more than one that used this map) were
 208          * disconnected from events. Wait for outstanding critical sections in
 209          * these programs to complete. The rcu critical section only guarantees
 210          * no further reads against netdev_map. It does __not__ ensure pending
 211          * flush operations (if any) are complete.
 212          */
 213 
 214         spin_lock(&dev_map_lock);
 215         list_del_rcu(&dtab->list);
 216         spin_unlock(&dev_map_lock);
 217 
 218         bpf_clear_redirect_map(map);
 219         synchronize_rcu();
 220 
 221         /* Make sure prior __dev_map_entry_free() have completed. */
 222         rcu_barrier();
 223 
 224         /* To ensure all pending flush operations have completed wait for flush
 225          * list to empty on _all_ cpus.
 226          * Because the above synchronize_rcu() ensures the map is disconnected
 227          * from the program we can assume no new items will be added.
 228          */
 229         for_each_online_cpu(cpu) {
 230                 struct list_head *flush_list = per_cpu_ptr(dtab->flush_list, cpu);
 231 
 232                 while (!list_empty(flush_list))
 233                         cond_resched();
 234         }
 235 
 236         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 237                 for (i = 0; i < dtab->n_buckets; i++) {
 238                         struct bpf_dtab_netdev *dev;
 239                         struct hlist_head *head;
 240                         struct hlist_node *next;
 241 
 242                         head = dev_map_index_hash(dtab, i);
 243 
 244                         hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 245                                 hlist_del_rcu(&dev->index_hlist);
 246                                 free_percpu(dev->bulkq);
 247                                 dev_put(dev->dev);
 248                                 kfree(dev);
 249                         }
 250                 }
 251 
 252                 kfree(dtab->dev_index_head);
 253         } else {
 254                 for (i = 0; i < dtab->map.max_entries; i++) {
 255                         struct bpf_dtab_netdev *dev;
 256 
 257                         dev = dtab->netdev_map[i];
 258                         if (!dev)
 259                                 continue;
 260 
 261                         free_percpu(dev->bulkq);
 262                         dev_put(dev->dev);
 263                         kfree(dev);
 264                 }
 265 
 266                 bpf_map_area_free(dtab->netdev_map);
 267         }
 268 
 269         free_percpu(dtab->flush_list);
 270         kfree(dtab);
 271 }
 272 
 273 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 274 {
 275         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 276         u32 index = key ? *(u32 *)key : U32_MAX;
 277         u32 *next = next_key;
 278 
 279         if (index >= dtab->map.max_entries) {
 280                 *next = 0;
 281                 return 0;
 282         }
 283 
 284         if (index == dtab->map.max_entries - 1)
 285                 return -ENOENT;
 286         *next = index + 1;
 287         return 0;
 288 }
 289 
 290 struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
 291 {
 292         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 293         struct hlist_head *head = dev_map_index_hash(dtab, key);
 294         struct bpf_dtab_netdev *dev;
 295 
 296         hlist_for_each_entry_rcu(dev, head, index_hlist,
 297                                  lockdep_is_held(&dtab->index_lock))
 298                 if (dev->idx == key)
 299                         return dev;
 300 
 301         return NULL;
 302 }
 303 
 304 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
 305                                     void *next_key)
 306 {
 307         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 308         u32 idx, *next = next_key;
 309         struct bpf_dtab_netdev *dev, *next_dev;
 310         struct hlist_head *head;
 311         int i = 0;
 312 
 313         if (!key)
 314                 goto find_first;
 315 
 316         idx = *(u32 *)key;
 317 
 318         dev = __dev_map_hash_lookup_elem(map, idx);
 319         if (!dev)
 320                 goto find_first;
 321 
 322         next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
 323                                     struct bpf_dtab_netdev, index_hlist);
 324 
 325         if (next_dev) {
 326                 *next = next_dev->idx;
 327                 return 0;
 328         }
 329 
 330         i = idx & (dtab->n_buckets - 1);
 331         i++;
 332 
 333  find_first:
 334         for (; i < dtab->n_buckets; i++) {
 335                 head = dev_map_index_hash(dtab, i);
 336 
 337                 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 338                                             struct bpf_dtab_netdev,
 339                                             index_hlist);
 340                 if (next_dev) {
 341                         *next = next_dev->idx;
 342                         return 0;
 343                 }
 344         }
 345 
 346         return -ENOENT;
 347 }
 348 
 349 static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags,
 350                        bool in_napi_ctx)
 351 {
 352         struct bpf_dtab_netdev *obj = bq->obj;
 353         struct net_device *dev = obj->dev;
 354         int sent = 0, drops = 0, err = 0;
 355         int i;
 356 
 357         if (unlikely(!bq->count))
 358                 return 0;
 359 
 360         for (i = 0; i < bq->count; i++) {
 361                 struct xdp_frame *xdpf = bq->q[i];
 362 
 363                 prefetch(xdpf);
 364         }
 365 
 366         sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
 367         if (sent < 0) {
 368                 err = sent;
 369                 sent = 0;
 370                 goto error;
 371         }
 372         drops = bq->count - sent;
 373 out:
 374         bq->count = 0;
 375 
 376         trace_xdp_devmap_xmit(&obj->dtab->map, obj->idx,
 377                               sent, drops, bq->dev_rx, dev, err);
 378         bq->dev_rx = NULL;
 379         __list_del_clearprev(&bq->flush_node);
 380         return 0;
 381 error:
 382         /* If ndo_xdp_xmit fails with an errno, no frames have been
 383          * xmit'ed and it's our responsibility to them free all.
 384          */
 385         for (i = 0; i < bq->count; i++) {
 386                 struct xdp_frame *xdpf = bq->q[i];
 387 
 388                 /* RX path under NAPI protection, can return frames faster */
 389                 if (likely(in_napi_ctx))
 390                         xdp_return_frame_rx_napi(xdpf);
 391                 else
 392                         xdp_return_frame(xdpf);
 393                 drops++;
 394         }
 395         goto out;
 396 }
 397 
 398 /* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
 399  * from the driver before returning from its napi->poll() routine. The poll()
 400  * routine is called either from busy_poll context or net_rx_action signaled
 401  * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
 402  * net device can be torn down. On devmap tear down we ensure the flush list
 403  * is empty before completing to ensure all flush operations have completed.
 404  */
 405 void __dev_map_flush(struct bpf_map *map)
 406 {
 407         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 408         struct list_head *flush_list = this_cpu_ptr(dtab->flush_list);
 409         struct xdp_bulk_queue *bq, *tmp;
 410 
 411         rcu_read_lock();
 412         list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
 413                 bq_xmit_all(bq, XDP_XMIT_FLUSH, true);
 414         rcu_read_unlock();
 415 }
 416 
 417 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
 418  * update happens in parallel here a dev_put wont happen until after reading the
 419  * ifindex.
 420  */
 421 struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
 422 {
 423         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 424         struct bpf_dtab_netdev *obj;
 425 
 426         if (key >= map->max_entries)
 427                 return NULL;
 428 
 429         obj = READ_ONCE(dtab->netdev_map[key]);
 430         return obj;
 431 }
 432 
 433 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
 434  * Thus, safe percpu variable access.
 435  */
 436 static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
 437                       struct net_device *dev_rx)
 438 
 439 {
 440         struct list_head *flush_list = this_cpu_ptr(obj->dtab->flush_list);
 441         struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
 442 
 443         if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
 444                 bq_xmit_all(bq, 0, true);
 445 
 446         /* Ingress dev_rx will be the same for all xdp_frame's in
 447          * bulk_queue, because bq stored per-CPU and must be flushed
 448          * from net_device drivers NAPI func end.
 449          */
 450         if (!bq->dev_rx)
 451                 bq->dev_rx = dev_rx;
 452 
 453         bq->q[bq->count++] = xdpf;
 454 
 455         if (!bq->flush_node.prev)
 456                 list_add(&bq->flush_node, flush_list);
 457 
 458         return 0;
 459 }
 460 
 461 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
 462                     struct net_device *dev_rx)
 463 {
 464         struct net_device *dev = dst->dev;
 465         struct xdp_frame *xdpf;
 466         int err;
 467 
 468         if (!dev->netdev_ops->ndo_xdp_xmit)
 469                 return -EOPNOTSUPP;
 470 
 471         err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
 472         if (unlikely(err))
 473                 return err;
 474 
 475         xdpf = convert_to_xdp_frame(xdp);
 476         if (unlikely(!xdpf))
 477                 return -EOVERFLOW;
 478 
 479         return bq_enqueue(dst, xdpf, dev_rx);
 480 }
 481 
 482 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
 483                              struct bpf_prog *xdp_prog)
 484 {
 485         int err;
 486 
 487         err = xdp_ok_fwd_dev(dst->dev, skb->len);
 488         if (unlikely(err))
 489                 return err;
 490         skb->dev = dst->dev;
 491         generic_xdp_tx(skb, xdp_prog);
 492 
 493         return 0;
 494 }
 495 
 496 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
 497 {
 498         struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
 499         struct net_device *dev = obj ? obj->dev : NULL;
 500 
 501         return dev ? &dev->ifindex : NULL;
 502 }
 503 
 504 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
 505 {
 506         struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
 507                                                                 *(u32 *)key);
 508         struct net_device *dev = obj ? obj->dev : NULL;
 509 
 510         return dev ? &dev->ifindex : NULL;
 511 }
 512 
 513 static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
 514 {
 515         if (dev->dev->netdev_ops->ndo_xdp_xmit) {
 516                 struct xdp_bulk_queue *bq;
 517                 int cpu;
 518 
 519                 rcu_read_lock();
 520                 for_each_online_cpu(cpu) {
 521                         bq = per_cpu_ptr(dev->bulkq, cpu);
 522                         bq_xmit_all(bq, XDP_XMIT_FLUSH, false);
 523                 }
 524                 rcu_read_unlock();
 525         }
 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         dev_map_flush_old(dev);
 534         free_percpu(dev->bulkq);
 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, but this does not guarantee a flush has happened
 550          * yet. Because driver side rcu_read_lock/unlock only protects the
 551          * running XDP program. However, for pending flush operations the
 552          * dev and ctx are stored in another per cpu map. And additionally,
 553          * the driver tear down ensures all soft irqs are complete before
 554          * removing the net device in the case of dev_put equals zero.
 555          */
 556         old_dev = xchg(&dtab->netdev_map[k], NULL);
 557         if (old_dev)
 558                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 559         return 0;
 560 }
 561 
 562 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
 563 {
 564         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 565         struct bpf_dtab_netdev *old_dev;
 566         int k = *(u32 *)key;
 567         unsigned long flags;
 568         int ret = -ENOENT;
 569 
 570         spin_lock_irqsave(&dtab->index_lock, flags);
 571 
 572         old_dev = __dev_map_hash_lookup_elem(map, k);
 573         if (old_dev) {
 574                 dtab->items--;
 575                 hlist_del_init_rcu(&old_dev->index_hlist);
 576                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 577                 ret = 0;
 578         }
 579         spin_unlock_irqrestore(&dtab->index_lock, flags);
 580 
 581         return ret;
 582 }
 583 
 584 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
 585                                                     struct bpf_dtab *dtab,
 586                                                     u32 ifindex,
 587                                                     unsigned int idx)
 588 {
 589         gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
 590         struct bpf_dtab_netdev *dev;
 591         struct xdp_bulk_queue *bq;
 592         int cpu;
 593 
 594         dev = kmalloc_node(sizeof(*dev), gfp, dtab->map.numa_node);
 595         if (!dev)
 596                 return ERR_PTR(-ENOMEM);
 597 
 598         dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
 599                                         sizeof(void *), gfp);
 600         if (!dev->bulkq) {
 601                 kfree(dev);
 602                 return ERR_PTR(-ENOMEM);
 603         }
 604 
 605         for_each_possible_cpu(cpu) {
 606                 bq = per_cpu_ptr(dev->bulkq, cpu);
 607                 bq->obj = dev;
 608         }
 609 
 610         dev->dev = dev_get_by_index(net, ifindex);
 611         if (!dev->dev) {
 612                 free_percpu(dev->bulkq);
 613                 kfree(dev);
 614                 return ERR_PTR(-EINVAL);
 615         }
 616 
 617         dev->idx = idx;
 618         dev->dtab = dtab;
 619 
 620         return dev;
 621 }
 622 
 623 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
 624                                  void *key, void *value, u64 map_flags)
 625 {
 626         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 627         struct bpf_dtab_netdev *dev, *old_dev;
 628         u32 ifindex = *(u32 *)value;
 629         u32 i = *(u32 *)key;
 630 
 631         if (unlikely(map_flags > BPF_EXIST))
 632                 return -EINVAL;
 633         if (unlikely(i >= dtab->map.max_entries))
 634                 return -E2BIG;
 635         if (unlikely(map_flags == BPF_NOEXIST))
 636                 return -EEXIST;
 637 
 638         if (!ifindex) {
 639                 dev = NULL;
 640         } else {
 641                 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
 642                 if (IS_ERR(dev))
 643                         return PTR_ERR(dev);
 644         }
 645 
 646         /* Use call_rcu() here to ensure rcu critical sections have completed
 647          * Remembering the driver side flush operation will happen before the
 648          * net device is removed.
 649          */
 650         old_dev = xchg(&dtab->netdev_map[i], dev);
 651         if (old_dev)
 652                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 653 
 654         return 0;
 655 }
 656 
 657 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
 658                                u64 map_flags)
 659 {
 660         return __dev_map_update_elem(current->nsproxy->net_ns,
 661                                      map, key, value, map_flags);
 662 }
 663 
 664 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
 665                                      void *key, void *value, u64 map_flags)
 666 {
 667         struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
 668         struct bpf_dtab_netdev *dev, *old_dev;
 669         u32 ifindex = *(u32 *)value;
 670         u32 idx = *(u32 *)key;
 671         unsigned long flags;
 672         int err = -EEXIST;
 673 
 674         if (unlikely(map_flags > BPF_EXIST || !ifindex))
 675                 return -EINVAL;
 676 
 677         spin_lock_irqsave(&dtab->index_lock, flags);
 678 
 679         old_dev = __dev_map_hash_lookup_elem(map, idx);
 680         if (old_dev && (map_flags & BPF_NOEXIST))
 681                 goto out_err;
 682 
 683         dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
 684         if (IS_ERR(dev)) {
 685                 err = PTR_ERR(dev);
 686                 goto out_err;
 687         }
 688 
 689         if (old_dev) {
 690                 hlist_del_rcu(&old_dev->index_hlist);
 691         } else {
 692                 if (dtab->items >= dtab->map.max_entries) {
 693                         spin_unlock_irqrestore(&dtab->index_lock, flags);
 694                         call_rcu(&dev->rcu, __dev_map_entry_free);
 695                         return -E2BIG;
 696                 }
 697                 dtab->items++;
 698         }
 699 
 700         hlist_add_head_rcu(&dev->index_hlist,
 701                            dev_map_index_hash(dtab, idx));
 702         spin_unlock_irqrestore(&dtab->index_lock, flags);
 703 
 704         if (old_dev)
 705                 call_rcu(&old_dev->rcu, __dev_map_entry_free);
 706 
 707         return 0;
 708 
 709 out_err:
 710         spin_unlock_irqrestore(&dtab->index_lock, flags);
 711         return err;
 712 }
 713 
 714 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
 715                                    u64 map_flags)
 716 {
 717         return __dev_map_hash_update_elem(current->nsproxy->net_ns,
 718                                          map, key, value, map_flags);
 719 }
 720 
 721 const struct bpf_map_ops dev_map_ops = {
 722         .map_alloc = dev_map_alloc,
 723         .map_free = dev_map_free,
 724         .map_get_next_key = dev_map_get_next_key,
 725         .map_lookup_elem = dev_map_lookup_elem,
 726         .map_update_elem = dev_map_update_elem,
 727         .map_delete_elem = dev_map_delete_elem,
 728         .map_check_btf = map_check_no_btf,
 729 };
 730 
 731 const struct bpf_map_ops dev_map_hash_ops = {
 732         .map_alloc = dev_map_alloc,
 733         .map_free = dev_map_free,
 734         .map_get_next_key = dev_map_hash_get_next_key,
 735         .map_lookup_elem = dev_map_hash_lookup_elem,
 736         .map_update_elem = dev_map_hash_update_elem,
 737         .map_delete_elem = dev_map_hash_delete_elem,
 738         .map_check_btf = map_check_no_btf,
 739 };
 740 
 741 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
 742                                        struct net_device *netdev)
 743 {
 744         unsigned long flags;
 745         u32 i;
 746 
 747         spin_lock_irqsave(&dtab->index_lock, flags);
 748         for (i = 0; i < dtab->n_buckets; i++) {
 749                 struct bpf_dtab_netdev *dev;
 750                 struct hlist_head *head;
 751                 struct hlist_node *next;
 752 
 753                 head = dev_map_index_hash(dtab, i);
 754 
 755                 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
 756                         if (netdev != dev->dev)
 757                                 continue;
 758 
 759                         dtab->items--;
 760                         hlist_del_rcu(&dev->index_hlist);
 761                         call_rcu(&dev->rcu, __dev_map_entry_free);
 762                 }
 763         }
 764         spin_unlock_irqrestore(&dtab->index_lock, flags);
 765 }
 766 
 767 static int dev_map_notification(struct notifier_block *notifier,
 768                                 ulong event, void *ptr)
 769 {
 770         struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
 771         struct bpf_dtab *dtab;
 772         int i;
 773 
 774         switch (event) {
 775         case NETDEV_UNREGISTER:
 776                 /* This rcu_read_lock/unlock pair is needed because
 777                  * dev_map_list is an RCU list AND to ensure a delete
 778                  * operation does not free a netdev_map entry while we
 779                  * are comparing it against the netdev being unregistered.
 780                  */
 781                 rcu_read_lock();
 782                 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
 783                         if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
 784                                 dev_map_hash_remove_netdev(dtab, netdev);
 785                                 continue;
 786                         }
 787 
 788                         for (i = 0; i < dtab->map.max_entries; i++) {
 789                                 struct bpf_dtab_netdev *dev, *odev;
 790 
 791                                 dev = READ_ONCE(dtab->netdev_map[i]);
 792                                 if (!dev || netdev != dev->dev)
 793                                         continue;
 794                                 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
 795                                 if (dev == odev)
 796                                         call_rcu(&dev->rcu,
 797                                                  __dev_map_entry_free);
 798                         }
 799                 }
 800                 rcu_read_unlock();
 801                 break;
 802         default:
 803                 break;
 804         }
 805         return NOTIFY_OK;
 806 }
 807 
 808 static struct notifier_block dev_map_notifier = {
 809         .notifier_call = dev_map_notification,
 810 };
 811 
 812 static int __init dev_map_init(void)
 813 {
 814         /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
 815         BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
 816                      offsetof(struct _bpf_dtab_netdev, dev));
 817         register_netdevice_notifier(&dev_map_notifier);
 818         return 0;
 819 }
 820 
 821 subsys_initcall(dev_map_init);