projects/linux·blob v5.5 · refs

820 lines · 21.8 KiB

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
56struct bpf_dtab_netdev;
57
58struct 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
66struct 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
75struct 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
88static DEFINE_SPINLOCK(dev_map_lock);
89static LIST_HEAD(dev_map_list);
90
91static 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
104static 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
110static 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
169free_percpu:
170 free_percpu(dtab->flush_list);
171free_charge:
172 bpf_map_charge_finish(&dtab->map.memory);
173 return -ENOMEM;
174}
175
176static 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
201static 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
273static 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
290struct 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 if (dev->idx == key)
298 return dev;
299
300 return NULL;
301}
302
303static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
304 void *next_key)
305{
306 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
307 u32 idx, *next = next_key;
308 struct bpf_dtab_netdev *dev, *next_dev;
309 struct hlist_head *head;
310 int i = 0;
311
312 if (!key)
313 goto find_first;
314
315 idx = *(u32 *)key;
316
317 dev = __dev_map_hash_lookup_elem(map, idx);
318 if (!dev)
319 goto find_first;
320
321 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
322 struct bpf_dtab_netdev, index_hlist);
323
324 if (next_dev) {
325 *next = next_dev->idx;
326 return 0;
327 }
328
329 i = idx & (dtab->n_buckets - 1);
330 i++;
331
332 find_first:
333 for (; i < dtab->n_buckets; i++) {
334 head = dev_map_index_hash(dtab, i);
335
336 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
337 struct bpf_dtab_netdev,
338 index_hlist);
339 if (next_dev) {
340 *next = next_dev->idx;
341 return 0;
342 }
343 }
344
345 return -ENOENT;
346}
347
348static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags,
349 bool in_napi_ctx)
350{
351 struct bpf_dtab_netdev *obj = bq->obj;
352 struct net_device *dev = obj->dev;
353 int sent = 0, drops = 0, err = 0;
354 int i;
355
356 if (unlikely(!bq->count))
357 return 0;
358
359 for (i = 0; i < bq->count; i++) {
360 struct xdp_frame *xdpf = bq->q[i];
361
362 prefetch(xdpf);
363 }
364
365 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
366 if (sent < 0) {
367 err = sent;
368 sent = 0;
369 goto error;
370 }
371 drops = bq->count - sent;
372out:
373 bq->count = 0;
374
375 trace_xdp_devmap_xmit(&obj->dtab->map, obj->idx,
376 sent, drops, bq->dev_rx, dev, err);
377 bq->dev_rx = NULL;
378 __list_del_clearprev(&bq->flush_node);
379 return 0;
380error:
381 /* If ndo_xdp_xmit fails with an errno, no frames have been
382 * xmit'ed and it's our responsibility to them free all.
383 */
384 for (i = 0; i < bq->count; i++) {
385 struct xdp_frame *xdpf = bq->q[i];
386
387 /* RX path under NAPI protection, can return frames faster */
388 if (likely(in_napi_ctx))
389 xdp_return_frame_rx_napi(xdpf);
390 else
391 xdp_return_frame(xdpf);
392 drops++;
393 }
394 goto out;
395}
396
397/* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
398 * from the driver before returning from its napi->poll() routine. The poll()
399 * routine is called either from busy_poll context or net_rx_action signaled
400 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
401 * net device can be torn down. On devmap tear down we ensure the flush list
402 * is empty before completing to ensure all flush operations have completed.
403 */
404void __dev_map_flush(struct bpf_map *map)
405{
406 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
407 struct list_head *flush_list = this_cpu_ptr(dtab->flush_list);
408 struct xdp_bulk_queue *bq, *tmp;
409
410 rcu_read_lock();
411 list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
412 bq_xmit_all(bq, XDP_XMIT_FLUSH, true);
413 rcu_read_unlock();
414}
415
416/* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
417 * update happens in parallel here a dev_put wont happen until after reading the
418 * ifindex.
419 */
420struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
421{
422 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
423 struct bpf_dtab_netdev *obj;
424
425 if (key >= map->max_entries)
426 return NULL;
427
428 obj = READ_ONCE(dtab->netdev_map[key]);
429 return obj;
430}
431
432/* Runs under RCU-read-side, plus in softirq under NAPI protection.
433 * Thus, safe percpu variable access.
434 */
435static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
436 struct net_device *dev_rx)
437
438{
439 struct list_head *flush_list = this_cpu_ptr(obj->dtab->flush_list);
440 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
441
442 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
443 bq_xmit_all(bq, 0, true);
444
445 /* Ingress dev_rx will be the same for all xdp_frame's in
446 * bulk_queue, because bq stored per-CPU and must be flushed
447 * from net_device drivers NAPI func end.
448 */
449 if (!bq->dev_rx)
450 bq->dev_rx = dev_rx;
451
452 bq->q[bq->count++] = xdpf;
453
454 if (!bq->flush_node.prev)
455 list_add(&bq->flush_node, flush_list);
456
457 return 0;
458}
459
460int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
461 struct net_device *dev_rx)
462{
463 struct net_device *dev = dst->dev;
464 struct xdp_frame *xdpf;
465 int err;
466
467 if (!dev->netdev_ops->ndo_xdp_xmit)
468 return -EOPNOTSUPP;
469
470 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
471 if (unlikely(err))
472 return err;
473
474 xdpf = convert_to_xdp_frame(xdp);
475 if (unlikely(!xdpf))
476 return -EOVERFLOW;
477
478 return bq_enqueue(dst, xdpf, dev_rx);
479}
480
481int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
482 struct bpf_prog *xdp_prog)
483{
484 int err;
485
486 err = xdp_ok_fwd_dev(dst->dev, skb->len);
487 if (unlikely(err))
488 return err;
489 skb->dev = dst->dev;
490 generic_xdp_tx(skb, xdp_prog);
491
492 return 0;
493}
494
495static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
496{
497 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
498 struct net_device *dev = obj ? obj->dev : NULL;
499
500 return dev ? &dev->ifindex : NULL;
501}
502
503static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
504{
505 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
506 *(u32 *)key);
507 struct net_device *dev = obj ? obj->dev : NULL;
508
509 return dev ? &dev->ifindex : NULL;
510}
511
512static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
513{
514 if (dev->dev->netdev_ops->ndo_xdp_xmit) {
515 struct xdp_bulk_queue *bq;
516 int cpu;
517
518 rcu_read_lock();
519 for_each_online_cpu(cpu) {
520 bq = per_cpu_ptr(dev->bulkq, cpu);
521 bq_xmit_all(bq, XDP_XMIT_FLUSH, false);
522 }
523 rcu_read_unlock();
524 }
525}
526
527static void __dev_map_entry_free(struct rcu_head *rcu)
528{
529 struct bpf_dtab_netdev *dev;
530
531 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
532 dev_map_flush_old(dev);
533 free_percpu(dev->bulkq);
534 dev_put(dev->dev);
535 kfree(dev);
536}
537
538static int dev_map_delete_elem(struct bpf_map *map, void *key)
539{
540 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
541 struct bpf_dtab_netdev *old_dev;
542 int k = *(u32 *)key;
543
544 if (k >= map->max_entries)
545 return -EINVAL;
546
547 /* Use call_rcu() here to ensure any rcu critical sections have
548 * completed, but this does not guarantee a flush has happened
549 * yet. Because driver side rcu_read_lock/unlock only protects the
550 * running XDP program. However, for pending flush operations the
551 * dev and ctx are stored in another per cpu map. And additionally,
552 * the driver tear down ensures all soft irqs are complete before
553 * removing the net device in the case of dev_put equals zero.
554 */
555 old_dev = xchg(&dtab->netdev_map[k], NULL);
556 if (old_dev)
557 call_rcu(&old_dev->rcu, __dev_map_entry_free);
558 return 0;
559}
560
561static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
562{
563 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
564 struct bpf_dtab_netdev *old_dev;
565 int k = *(u32 *)key;
566 unsigned long flags;
567 int ret = -ENOENT;
568
569 spin_lock_irqsave(&dtab->index_lock, flags);
570
571 old_dev = __dev_map_hash_lookup_elem(map, k);
572 if (old_dev) {
573 dtab->items--;
574 hlist_del_init_rcu(&old_dev->index_hlist);
575 call_rcu(&old_dev->rcu, __dev_map_entry_free);
576 ret = 0;
577 }
578 spin_unlock_irqrestore(&dtab->index_lock, flags);
579
580 return ret;
581}
582
583static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
584 struct bpf_dtab *dtab,
585 u32 ifindex,
586 unsigned int idx)
587{
588 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
589 struct bpf_dtab_netdev *dev;
590 struct xdp_bulk_queue *bq;
591 int cpu;
592
593 dev = kmalloc_node(sizeof(*dev), gfp, dtab->map.numa_node);
594 if (!dev)
595 return ERR_PTR(-ENOMEM);
596
597 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
598 sizeof(void *), gfp);
599 if (!dev->bulkq) {
600 kfree(dev);
601 return ERR_PTR(-ENOMEM);
602 }
603
604 for_each_possible_cpu(cpu) {
605 bq = per_cpu_ptr(dev->bulkq, cpu);
606 bq->obj = dev;
607 }
608
609 dev->dev = dev_get_by_index(net, ifindex);
610 if (!dev->dev) {
611 free_percpu(dev->bulkq);
612 kfree(dev);
613 return ERR_PTR(-EINVAL);
614 }
615
616 dev->idx = idx;
617 dev->dtab = dtab;
618
619 return dev;
620}
621
622static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
623 void *key, void *value, u64 map_flags)
624{
625 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
626 struct bpf_dtab_netdev *dev, *old_dev;
627 u32 ifindex = *(u32 *)value;
628 u32 i = *(u32 *)key;
629
630 if (unlikely(map_flags > BPF_EXIST))
631 return -EINVAL;
632 if (unlikely(i >= dtab->map.max_entries))
633 return -E2BIG;
634 if (unlikely(map_flags == BPF_NOEXIST))
635 return -EEXIST;
636
637 if (!ifindex) {
638 dev = NULL;
639 } else {
640 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
641 if (IS_ERR(dev))
642 return PTR_ERR(dev);
643 }
644
645 /* Use call_rcu() here to ensure rcu critical sections have completed
646 * Remembering the driver side flush operation will happen before the
647 * net device is removed.
648 */
649 old_dev = xchg(&dtab->netdev_map[i], dev);
650 if (old_dev)
651 call_rcu(&old_dev->rcu, __dev_map_entry_free);
652
653 return 0;
654}
655
656static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
657 u64 map_flags)
658{
659 return __dev_map_update_elem(current->nsproxy->net_ns,
660 map, key, value, map_flags);
661}
662
663static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
664 void *key, void *value, u64 map_flags)
665{
666 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
667 struct bpf_dtab_netdev *dev, *old_dev;
668 u32 ifindex = *(u32 *)value;
669 u32 idx = *(u32 *)key;
670 unsigned long flags;
671 int err = -EEXIST;
672
673 if (unlikely(map_flags > BPF_EXIST || !ifindex))
674 return -EINVAL;
675
676 spin_lock_irqsave(&dtab->index_lock, flags);
677
678 old_dev = __dev_map_hash_lookup_elem(map, idx);
679 if (old_dev && (map_flags & BPF_NOEXIST))
680 goto out_err;
681
682 dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
683 if (IS_ERR(dev)) {
684 err = PTR_ERR(dev);
685 goto out_err;
686 }
687
688 if (old_dev) {
689 hlist_del_rcu(&old_dev->index_hlist);
690 } else {
691 if (dtab->items >= dtab->map.max_entries) {
692 spin_unlock_irqrestore(&dtab->index_lock, flags);
693 call_rcu(&dev->rcu, __dev_map_entry_free);
694 return -E2BIG;
695 }
696 dtab->items++;
697 }
698
699 hlist_add_head_rcu(&dev->index_hlist,
700 dev_map_index_hash(dtab, idx));
701 spin_unlock_irqrestore(&dtab->index_lock, flags);
702
703 if (old_dev)
704 call_rcu(&old_dev->rcu, __dev_map_entry_free);
705
706 return 0;
707
708out_err:
709 spin_unlock_irqrestore(&dtab->index_lock, flags);
710 return err;
711}
712
713static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
714 u64 map_flags)
715{
716 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
717 map, key, value, map_flags);
718}
719
720const struct bpf_map_ops dev_map_ops = {
721 .map_alloc = dev_map_alloc,
722 .map_free = dev_map_free,
723 .map_get_next_key = dev_map_get_next_key,
724 .map_lookup_elem = dev_map_lookup_elem,
725 .map_update_elem = dev_map_update_elem,
726 .map_delete_elem = dev_map_delete_elem,
727 .map_check_btf = map_check_no_btf,
728};
729
730const struct bpf_map_ops dev_map_hash_ops = {
731 .map_alloc = dev_map_alloc,
732 .map_free = dev_map_free,
733 .map_get_next_key = dev_map_hash_get_next_key,
734 .map_lookup_elem = dev_map_hash_lookup_elem,
735 .map_update_elem = dev_map_hash_update_elem,
736 .map_delete_elem = dev_map_hash_delete_elem,
737 .map_check_btf = map_check_no_btf,
738};
739
740static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
741 struct net_device *netdev)
742{
743 unsigned long flags;
744 u32 i;
745
746 spin_lock_irqsave(&dtab->index_lock, flags);
747 for (i = 0; i < dtab->n_buckets; i++) {
748 struct bpf_dtab_netdev *dev;
749 struct hlist_head *head;
750 struct hlist_node *next;
751
752 head = dev_map_index_hash(dtab, i);
753
754 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
755 if (netdev != dev->dev)
756 continue;
757
758 dtab->items--;
759 hlist_del_rcu(&dev->index_hlist);
760 call_rcu(&dev->rcu, __dev_map_entry_free);
761 }
762 }
763 spin_unlock_irqrestore(&dtab->index_lock, flags);
764}
765
766static int dev_map_notification(struct notifier_block *notifier,
767 ulong event, void *ptr)
768{
769 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
770 struct bpf_dtab *dtab;
771 int i;
772
773 switch (event) {
774 case NETDEV_UNREGISTER:
775 /* This rcu_read_lock/unlock pair is needed because
776 * dev_map_list is an RCU list AND to ensure a delete
777 * operation does not free a netdev_map entry while we
778 * are comparing it against the netdev being unregistered.
779 */
780 rcu_read_lock();
781 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
782 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
783 dev_map_hash_remove_netdev(dtab, netdev);
784 continue;
785 }
786
787 for (i = 0; i < dtab->map.max_entries; i++) {
788 struct bpf_dtab_netdev *dev, *odev;
789
790 dev = READ_ONCE(dtab->netdev_map[i]);
791 if (!dev || netdev != dev->dev)
792 continue;
793 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
794 if (dev == odev)
795 call_rcu(&dev->rcu,
796 __dev_map_entry_free);
797 }
798 }
799 rcu_read_unlock();
800 break;
801 default:
802 break;
803 }
804 return NOTIFY_OK;
805}
806
807static struct notifier_block dev_map_notifier = {
808 .notifier_call = dev_map_notification,
809};
810
811static int __init dev_map_init(void)
812{
813 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
814 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
815 offsetof(struct _bpf_dtab_netdev, dev));
816 register_netdevice_notifier(&dev_map_notifier);
817 return 0;
818}
819
820subsys_initcall(dev_map_init);