Loading...
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;
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 int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
105{
106 int err, cpu;
107 u64 cost;
108
109 /* check sanity of attributes */
110 if (attr->max_entries == 0 || attr->key_size != 4 ||
111 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
112 return -EINVAL;
113
114 /* Lookup returns a pointer straight to dev->ifindex, so make sure the
115 * verifier prevents writes from the BPF side
116 */
117 attr->map_flags |= BPF_F_RDONLY_PROG;
118
119
120 bpf_map_init_from_attr(&dtab->map, attr);
121
122 /* make sure page count doesn't overflow */
123 cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
124 cost += sizeof(struct list_head) * num_possible_cpus();
125
126 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
127 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
128
129 if (!dtab->n_buckets) /* Overflow check */
130 return -EINVAL;
131 cost += (u64) sizeof(struct hlist_head) * dtab->n_buckets;
132 }
133
134 /* if map size is larger than memlock limit, reject it */
135 err = bpf_map_charge_init(&dtab->map.memory, cost);
136 if (err)
137 return -EINVAL;
138
139 dtab->flush_list = alloc_percpu(struct list_head);
140 if (!dtab->flush_list)
141 goto free_charge;
142
143 for_each_possible_cpu(cpu)
144 INIT_LIST_HEAD(per_cpu_ptr(dtab->flush_list, cpu));
145
146 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
147 sizeof(struct bpf_dtab_netdev *),
148 dtab->map.numa_node);
149 if (!dtab->netdev_map)
150 goto free_percpu;
151
152 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
153 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets);
154 if (!dtab->dev_index_head)
155 goto free_map_area;
156
157 spin_lock_init(&dtab->index_lock);
158 }
159
160 return 0;
161
162free_map_area:
163 bpf_map_area_free(dtab->netdev_map);
164free_percpu:
165 free_percpu(dtab->flush_list);
166free_charge:
167 bpf_map_charge_finish(&dtab->map.memory);
168 return -ENOMEM;
169}
170
171static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
172{
173 struct bpf_dtab *dtab;
174 int err;
175
176 if (!capable(CAP_NET_ADMIN))
177 return ERR_PTR(-EPERM);
178
179 dtab = kzalloc(sizeof(*dtab), GFP_USER);
180 if (!dtab)
181 return ERR_PTR(-ENOMEM);
182
183 err = dev_map_init_map(dtab, attr);
184 if (err) {
185 kfree(dtab);
186 return ERR_PTR(err);
187 }
188
189 spin_lock(&dev_map_lock);
190 list_add_tail_rcu(&dtab->list, &dev_map_list);
191 spin_unlock(&dev_map_lock);
192
193 return &dtab->map;
194}
195
196static void dev_map_free(struct bpf_map *map)
197{
198 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
199 int i, cpu;
200
201 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
202 * so the programs (can be more than one that used this map) were
203 * disconnected from events. Wait for outstanding critical sections in
204 * these programs to complete. The rcu critical section only guarantees
205 * no further reads against netdev_map. It does __not__ ensure pending
206 * flush operations (if any) are complete.
207 */
208
209 spin_lock(&dev_map_lock);
210 list_del_rcu(&dtab->list);
211 spin_unlock(&dev_map_lock);
212
213 bpf_clear_redirect_map(map);
214 synchronize_rcu();
215
216 /* Make sure prior __dev_map_entry_free() have completed. */
217 rcu_barrier();
218
219 /* To ensure all pending flush operations have completed wait for flush
220 * list to empty on _all_ cpus.
221 * Because the above synchronize_rcu() ensures the map is disconnected
222 * from the program we can assume no new items will be added.
223 */
224 for_each_online_cpu(cpu) {
225 struct list_head *flush_list = per_cpu_ptr(dtab->flush_list, cpu);
226
227 while (!list_empty(flush_list))
228 cond_resched();
229 }
230
231 for (i = 0; i < dtab->map.max_entries; i++) {
232 struct bpf_dtab_netdev *dev;
233
234 dev = dtab->netdev_map[i];
235 if (!dev)
236 continue;
237
238 free_percpu(dev->bulkq);
239 dev_put(dev->dev);
240 kfree(dev);
241 }
242
243 free_percpu(dtab->flush_list);
244 bpf_map_area_free(dtab->netdev_map);
245 kfree(dtab->dev_index_head);
246 kfree(dtab);
247}
248
249static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
250{
251 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
252 u32 index = key ? *(u32 *)key : U32_MAX;
253 u32 *next = next_key;
254
255 if (index >= dtab->map.max_entries) {
256 *next = 0;
257 return 0;
258 }
259
260 if (index == dtab->map.max_entries - 1)
261 return -ENOENT;
262 *next = index + 1;
263 return 0;
264}
265
266static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
267 int idx)
268{
269 return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
270}
271
272struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
273{
274 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
275 struct hlist_head *head = dev_map_index_hash(dtab, key);
276 struct bpf_dtab_netdev *dev;
277
278 hlist_for_each_entry_rcu(dev, head, index_hlist)
279 if (dev->idx == key)
280 return dev;
281
282 return NULL;
283}
284
285static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
286 void *next_key)
287{
288 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
289 u32 idx, *next = next_key;
290 struct bpf_dtab_netdev *dev, *next_dev;
291 struct hlist_head *head;
292 int i = 0;
293
294 if (!key)
295 goto find_first;
296
297 idx = *(u32 *)key;
298
299 dev = __dev_map_hash_lookup_elem(map, idx);
300 if (!dev)
301 goto find_first;
302
303 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
304 struct bpf_dtab_netdev, index_hlist);
305
306 if (next_dev) {
307 *next = next_dev->idx;
308 return 0;
309 }
310
311 i = idx & (dtab->n_buckets - 1);
312 i++;
313
314 find_first:
315 for (; i < dtab->n_buckets; i++) {
316 head = dev_map_index_hash(dtab, i);
317
318 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
319 struct bpf_dtab_netdev,
320 index_hlist);
321 if (next_dev) {
322 *next = next_dev->idx;
323 return 0;
324 }
325 }
326
327 return -ENOENT;
328}
329
330static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags,
331 bool in_napi_ctx)
332{
333 struct bpf_dtab_netdev *obj = bq->obj;
334 struct net_device *dev = obj->dev;
335 int sent = 0, drops = 0, err = 0;
336 int i;
337
338 if (unlikely(!bq->count))
339 return 0;
340
341 for (i = 0; i < bq->count; i++) {
342 struct xdp_frame *xdpf = bq->q[i];
343
344 prefetch(xdpf);
345 }
346
347 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
348 if (sent < 0) {
349 err = sent;
350 sent = 0;
351 goto error;
352 }
353 drops = bq->count - sent;
354out:
355 bq->count = 0;
356
357 trace_xdp_devmap_xmit(&obj->dtab->map, obj->idx,
358 sent, drops, bq->dev_rx, dev, err);
359 bq->dev_rx = NULL;
360 __list_del_clearprev(&bq->flush_node);
361 return 0;
362error:
363 /* If ndo_xdp_xmit fails with an errno, no frames have been
364 * xmit'ed and it's our responsibility to them free all.
365 */
366 for (i = 0; i < bq->count; i++) {
367 struct xdp_frame *xdpf = bq->q[i];
368
369 /* RX path under NAPI protection, can return frames faster */
370 if (likely(in_napi_ctx))
371 xdp_return_frame_rx_napi(xdpf);
372 else
373 xdp_return_frame(xdpf);
374 drops++;
375 }
376 goto out;
377}
378
379/* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
380 * from the driver before returning from its napi->poll() routine. The poll()
381 * routine is called either from busy_poll context or net_rx_action signaled
382 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
383 * net device can be torn down. On devmap tear down we ensure the flush list
384 * is empty before completing to ensure all flush operations have completed.
385 */
386void __dev_map_flush(struct bpf_map *map)
387{
388 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
389 struct list_head *flush_list = this_cpu_ptr(dtab->flush_list);
390 struct xdp_bulk_queue *bq, *tmp;
391
392 rcu_read_lock();
393 list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
394 bq_xmit_all(bq, XDP_XMIT_FLUSH, true);
395 rcu_read_unlock();
396}
397
398/* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
399 * update happens in parallel here a dev_put wont happen until after reading the
400 * ifindex.
401 */
402struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
403{
404 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
405 struct bpf_dtab_netdev *obj;
406
407 if (key >= map->max_entries)
408 return NULL;
409
410 obj = READ_ONCE(dtab->netdev_map[key]);
411 return obj;
412}
413
414/* Runs under RCU-read-side, plus in softirq under NAPI protection.
415 * Thus, safe percpu variable access.
416 */
417static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
418 struct net_device *dev_rx)
419
420{
421 struct list_head *flush_list = this_cpu_ptr(obj->dtab->flush_list);
422 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
423
424 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
425 bq_xmit_all(bq, 0, true);
426
427 /* Ingress dev_rx will be the same for all xdp_frame's in
428 * bulk_queue, because bq stored per-CPU and must be flushed
429 * from net_device drivers NAPI func end.
430 */
431 if (!bq->dev_rx)
432 bq->dev_rx = dev_rx;
433
434 bq->q[bq->count++] = xdpf;
435
436 if (!bq->flush_node.prev)
437 list_add(&bq->flush_node, flush_list);
438
439 return 0;
440}
441
442int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
443 struct net_device *dev_rx)
444{
445 struct net_device *dev = dst->dev;
446 struct xdp_frame *xdpf;
447 int err;
448
449 if (!dev->netdev_ops->ndo_xdp_xmit)
450 return -EOPNOTSUPP;
451
452 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
453 if (unlikely(err))
454 return err;
455
456 xdpf = convert_to_xdp_frame(xdp);
457 if (unlikely(!xdpf))
458 return -EOVERFLOW;
459
460 return bq_enqueue(dst, xdpf, dev_rx);
461}
462
463int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
464 struct bpf_prog *xdp_prog)
465{
466 int err;
467
468 err = xdp_ok_fwd_dev(dst->dev, skb->len);
469 if (unlikely(err))
470 return err;
471 skb->dev = dst->dev;
472 generic_xdp_tx(skb, xdp_prog);
473
474 return 0;
475}
476
477static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
478{
479 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
480 struct net_device *dev = obj ? obj->dev : NULL;
481
482 return dev ? &dev->ifindex : NULL;
483}
484
485static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
486{
487 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
488 *(u32 *)key);
489 struct net_device *dev = obj ? obj->dev : NULL;
490
491 return dev ? &dev->ifindex : NULL;
492}
493
494static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
495{
496 if (dev->dev->netdev_ops->ndo_xdp_xmit) {
497 struct xdp_bulk_queue *bq;
498 int cpu;
499
500 rcu_read_lock();
501 for_each_online_cpu(cpu) {
502 bq = per_cpu_ptr(dev->bulkq, cpu);
503 bq_xmit_all(bq, XDP_XMIT_FLUSH, false);
504 }
505 rcu_read_unlock();
506 }
507}
508
509static void __dev_map_entry_free(struct rcu_head *rcu)
510{
511 struct bpf_dtab_netdev *dev;
512
513 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
514 dev_map_flush_old(dev);
515 free_percpu(dev->bulkq);
516 dev_put(dev->dev);
517 kfree(dev);
518}
519
520static int dev_map_delete_elem(struct bpf_map *map, void *key)
521{
522 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
523 struct bpf_dtab_netdev *old_dev;
524 int k = *(u32 *)key;
525
526 if (k >= map->max_entries)
527 return -EINVAL;
528
529 /* Use call_rcu() here to ensure any rcu critical sections have
530 * completed, but this does not guarantee a flush has happened
531 * yet. Because driver side rcu_read_lock/unlock only protects the
532 * running XDP program. However, for pending flush operations the
533 * dev and ctx are stored in another per cpu map. And additionally,
534 * the driver tear down ensures all soft irqs are complete before
535 * removing the net device in the case of dev_put equals zero.
536 */
537 old_dev = xchg(&dtab->netdev_map[k], NULL);
538 if (old_dev)
539 call_rcu(&old_dev->rcu, __dev_map_entry_free);
540 return 0;
541}
542
543static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
544{
545 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
546 struct bpf_dtab_netdev *old_dev;
547 int k = *(u32 *)key;
548 unsigned long flags;
549 int ret = -ENOENT;
550
551 spin_lock_irqsave(&dtab->index_lock, flags);
552
553 old_dev = __dev_map_hash_lookup_elem(map, k);
554 if (old_dev) {
555 dtab->items--;
556 hlist_del_init_rcu(&old_dev->index_hlist);
557 call_rcu(&old_dev->rcu, __dev_map_entry_free);
558 ret = 0;
559 }
560 spin_unlock_irqrestore(&dtab->index_lock, flags);
561
562 return ret;
563}
564
565static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
566 struct bpf_dtab *dtab,
567 u32 ifindex,
568 unsigned int idx)
569{
570 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
571 struct bpf_dtab_netdev *dev;
572 struct xdp_bulk_queue *bq;
573 int cpu;
574
575 dev = kmalloc_node(sizeof(*dev), gfp, dtab->map.numa_node);
576 if (!dev)
577 return ERR_PTR(-ENOMEM);
578
579 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
580 sizeof(void *), gfp);
581 if (!dev->bulkq) {
582 kfree(dev);
583 return ERR_PTR(-ENOMEM);
584 }
585
586 for_each_possible_cpu(cpu) {
587 bq = per_cpu_ptr(dev->bulkq, cpu);
588 bq->obj = dev;
589 }
590
591 dev->dev = dev_get_by_index(net, ifindex);
592 if (!dev->dev) {
593 free_percpu(dev->bulkq);
594 kfree(dev);
595 return ERR_PTR(-EINVAL);
596 }
597
598 dev->idx = idx;
599 dev->dtab = dtab;
600
601 return dev;
602}
603
604static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
605 void *key, void *value, u64 map_flags)
606{
607 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
608 struct bpf_dtab_netdev *dev, *old_dev;
609 u32 ifindex = *(u32 *)value;
610 u32 i = *(u32 *)key;
611
612 if (unlikely(map_flags > BPF_EXIST))
613 return -EINVAL;
614 if (unlikely(i >= dtab->map.max_entries))
615 return -E2BIG;
616 if (unlikely(map_flags == BPF_NOEXIST))
617 return -EEXIST;
618
619 if (!ifindex) {
620 dev = NULL;
621 } else {
622 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
623 if (IS_ERR(dev))
624 return PTR_ERR(dev);
625 }
626
627 /* Use call_rcu() here to ensure rcu critical sections have completed
628 * Remembering the driver side flush operation will happen before the
629 * net device is removed.
630 */
631 old_dev = xchg(&dtab->netdev_map[i], dev);
632 if (old_dev)
633 call_rcu(&old_dev->rcu, __dev_map_entry_free);
634
635 return 0;
636}
637
638static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
639 u64 map_flags)
640{
641 return __dev_map_update_elem(current->nsproxy->net_ns,
642 map, key, value, map_flags);
643}
644
645static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
646 void *key, void *value, u64 map_flags)
647{
648 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
649 struct bpf_dtab_netdev *dev, *old_dev;
650 u32 ifindex = *(u32 *)value;
651 u32 idx = *(u32 *)key;
652 unsigned long flags;
653 int err = -EEXIST;
654
655 if (unlikely(map_flags > BPF_EXIST || !ifindex))
656 return -EINVAL;
657
658 spin_lock_irqsave(&dtab->index_lock, flags);
659
660 old_dev = __dev_map_hash_lookup_elem(map, idx);
661 if (old_dev && (map_flags & BPF_NOEXIST))
662 goto out_err;
663
664 dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
665 if (IS_ERR(dev)) {
666 err = PTR_ERR(dev);
667 goto out_err;
668 }
669
670 if (old_dev) {
671 hlist_del_rcu(&old_dev->index_hlist);
672 } else {
673 if (dtab->items >= dtab->map.max_entries) {
674 spin_unlock_irqrestore(&dtab->index_lock, flags);
675 call_rcu(&dev->rcu, __dev_map_entry_free);
676 return -E2BIG;
677 }
678 dtab->items++;
679 }
680
681 hlist_add_head_rcu(&dev->index_hlist,
682 dev_map_index_hash(dtab, idx));
683 spin_unlock_irqrestore(&dtab->index_lock, flags);
684
685 if (old_dev)
686 call_rcu(&old_dev->rcu, __dev_map_entry_free);
687
688 return 0;
689
690out_err:
691 spin_unlock_irqrestore(&dtab->index_lock, flags);
692 return err;
693}
694
695static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
696 u64 map_flags)
697{
698 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
699 map, key, value, map_flags);
700}
701
702const struct bpf_map_ops dev_map_ops = {
703 .map_alloc = dev_map_alloc,
704 .map_free = dev_map_free,
705 .map_get_next_key = dev_map_get_next_key,
706 .map_lookup_elem = dev_map_lookup_elem,
707 .map_update_elem = dev_map_update_elem,
708 .map_delete_elem = dev_map_delete_elem,
709 .map_check_btf = map_check_no_btf,
710};
711
712const struct bpf_map_ops dev_map_hash_ops = {
713 .map_alloc = dev_map_alloc,
714 .map_free = dev_map_free,
715 .map_get_next_key = dev_map_hash_get_next_key,
716 .map_lookup_elem = dev_map_hash_lookup_elem,
717 .map_update_elem = dev_map_hash_update_elem,
718 .map_delete_elem = dev_map_hash_delete_elem,
719 .map_check_btf = map_check_no_btf,
720};
721
722static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
723 struct net_device *netdev)
724{
725 unsigned long flags;
726 u32 i;
727
728 spin_lock_irqsave(&dtab->index_lock, flags);
729 for (i = 0; i < dtab->n_buckets; i++) {
730 struct bpf_dtab_netdev *dev;
731 struct hlist_head *head;
732 struct hlist_node *next;
733
734 head = dev_map_index_hash(dtab, i);
735
736 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
737 if (netdev != dev->dev)
738 continue;
739
740 dtab->items--;
741 hlist_del_rcu(&dev->index_hlist);
742 call_rcu(&dev->rcu, __dev_map_entry_free);
743 }
744 }
745 spin_unlock_irqrestore(&dtab->index_lock, flags);
746}
747
748static int dev_map_notification(struct notifier_block *notifier,
749 ulong event, void *ptr)
750{
751 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
752 struct bpf_dtab *dtab;
753 int i;
754
755 switch (event) {
756 case NETDEV_UNREGISTER:
757 /* This rcu_read_lock/unlock pair is needed because
758 * dev_map_list is an RCU list AND to ensure a delete
759 * operation does not free a netdev_map entry while we
760 * are comparing it against the netdev being unregistered.
761 */
762 rcu_read_lock();
763 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
764 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
765 dev_map_hash_remove_netdev(dtab, netdev);
766 continue;
767 }
768
769 for (i = 0; i < dtab->map.max_entries; i++) {
770 struct bpf_dtab_netdev *dev, *odev;
771
772 dev = READ_ONCE(dtab->netdev_map[i]);
773 if (!dev || netdev != dev->dev)
774 continue;
775 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
776 if (dev == odev)
777 call_rcu(&dev->rcu,
778 __dev_map_entry_free);
779 }
780 }
781 rcu_read_unlock();
782 break;
783 default:
784 break;
785 }
786 return NOTIFY_OK;
787}
788
789static struct notifier_block dev_map_notifier = {
790 .notifier_call = dev_map_notification,
791};
792
793static int __init dev_map_init(void)
794{
795 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
796 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
797 offsetof(struct _bpf_dtab_netdev, dev));
798 register_netdevice_notifier(&dev_map_notifier);
799 return 0;
800}
801
802subsys_initcall(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
55struct xdp_dev_bulk_queue {
56 struct xdp_frame *q[DEV_MAP_BULK_SIZE];
57 struct list_head flush_node;
58 struct net_device *dev;
59 struct net_device *dev_rx;
60 struct bpf_prog *xdp_prog;
61 unsigned int count;
62};
63
64struct bpf_dtab_netdev {
65 struct net_device *dev; /* must be first member, due to tracepoint */
66 struct hlist_node index_hlist;
67 struct bpf_dtab *dtab;
68 struct bpf_prog *xdp_prog;
69 struct rcu_head rcu;
70 unsigned int idx;
71 struct bpf_devmap_val val;
72};
73
74struct 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
86static DEFINE_PER_CPU(struct list_head, dev_flush_list);
87static DEFINE_SPINLOCK(dev_map_lock);
88static LIST_HEAD(dev_map_list);
89
90static 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
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 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
157static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
158{
159 struct bpf_dtab *dtab;
160 int err;
161
162 if (!capable(CAP_NET_ADMIN))
163 return ERR_PTR(-EPERM);
164
165 dtab = kzalloc(sizeof(*dtab), GFP_USER | __GFP_ACCOUNT);
166 if (!dtab)
167 return ERR_PTR(-ENOMEM);
168
169 err = dev_map_init_map(dtab, attr);
170 if (err) {
171 kfree(dtab);
172 return ERR_PTR(err);
173 }
174
175 spin_lock(&dev_map_lock);
176 list_add_tail_rcu(&dtab->list, &dev_map_list);
177 spin_unlock(&dev_map_lock);
178
179 return &dtab->map;
180}
181
182static void dev_map_free(struct bpf_map *map)
183{
184 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
185 int i;
186
187 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
188 * so the programs (can be more than one that used this map) were
189 * disconnected from events. The following synchronize_rcu() guarantees
190 * both rcu read critical sections complete and waits for
191 * preempt-disable regions (NAPI being the relevant context here) so we
192 * are certain there will be no further reads against the netdev_map and
193 * all flush operations are complete. Flush operations can only be done
194 * from NAPI context for this reason.
195 */
196
197 spin_lock(&dev_map_lock);
198 list_del_rcu(&dtab->list);
199 spin_unlock(&dev_map_lock);
200
201 bpf_clear_redirect_map(map);
202 synchronize_rcu();
203
204 /* Make sure prior __dev_map_entry_free() have completed. */
205 rcu_barrier();
206
207 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
208 for (i = 0; i < dtab->n_buckets; i++) {
209 struct bpf_dtab_netdev *dev;
210 struct hlist_head *head;
211 struct hlist_node *next;
212
213 head = dev_map_index_hash(dtab, i);
214
215 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
216 hlist_del_rcu(&dev->index_hlist);
217 if (dev->xdp_prog)
218 bpf_prog_put(dev->xdp_prog);
219 dev_put(dev->dev);
220 kfree(dev);
221 }
222 }
223
224 bpf_map_area_free(dtab->dev_index_head);
225 } else {
226 for (i = 0; i < dtab->map.max_entries; i++) {
227 struct bpf_dtab_netdev *dev;
228
229 dev = rcu_dereference_raw(dtab->netdev_map[i]);
230 if (!dev)
231 continue;
232
233 if (dev->xdp_prog)
234 bpf_prog_put(dev->xdp_prog);
235 dev_put(dev->dev);
236 kfree(dev);
237 }
238
239 bpf_map_area_free(dtab->netdev_map);
240 }
241
242 kfree(dtab);
243}
244
245static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
246{
247 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
248 u32 index = key ? *(u32 *)key : U32_MAX;
249 u32 *next = next_key;
250
251 if (index >= dtab->map.max_entries) {
252 *next = 0;
253 return 0;
254 }
255
256 if (index == dtab->map.max_entries - 1)
257 return -ENOENT;
258 *next = index + 1;
259 return 0;
260}
261
262/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
263 * by local_bh_disable() (from XDP calls inside NAPI). The
264 * rcu_read_lock_bh_held() below makes lockdep accept both.
265 */
266static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
267{
268 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
269 struct hlist_head *head = dev_map_index_hash(dtab, key);
270 struct bpf_dtab_netdev *dev;
271
272 hlist_for_each_entry_rcu(dev, head, index_hlist,
273 lockdep_is_held(&dtab->index_lock))
274 if (dev->idx == key)
275 return dev;
276
277 return NULL;
278}
279
280static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
281 void *next_key)
282{
283 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
284 u32 idx, *next = next_key;
285 struct bpf_dtab_netdev *dev, *next_dev;
286 struct hlist_head *head;
287 int i = 0;
288
289 if (!key)
290 goto find_first;
291
292 idx = *(u32 *)key;
293
294 dev = __dev_map_hash_lookup_elem(map, idx);
295 if (!dev)
296 goto find_first;
297
298 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
299 struct bpf_dtab_netdev, index_hlist);
300
301 if (next_dev) {
302 *next = next_dev->idx;
303 return 0;
304 }
305
306 i = idx & (dtab->n_buckets - 1);
307 i++;
308
309 find_first:
310 for (; i < dtab->n_buckets; i++) {
311 head = dev_map_index_hash(dtab, i);
312
313 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
314 struct bpf_dtab_netdev,
315 index_hlist);
316 if (next_dev) {
317 *next = next_dev->idx;
318 return 0;
319 }
320 }
321
322 return -ENOENT;
323}
324
325bool dev_map_can_have_prog(struct bpf_map *map)
326{
327 if ((map->map_type == BPF_MAP_TYPE_DEVMAP ||
328 map->map_type == BPF_MAP_TYPE_DEVMAP_HASH) &&
329 map->value_size != offsetofend(struct bpf_devmap_val, ifindex))
330 return true;
331
332 return false;
333}
334
335static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
336 struct xdp_frame **frames, int n,
337 struct net_device *dev)
338{
339 struct xdp_txq_info txq = { .dev = dev };
340 struct xdp_buff xdp;
341 int i, nframes = 0;
342
343 for (i = 0; i < n; i++) {
344 struct xdp_frame *xdpf = frames[i];
345 u32 act;
346 int err;
347
348 xdp_convert_frame_to_buff(xdpf, &xdp);
349 xdp.txq = &txq;
350
351 act = bpf_prog_run_xdp(xdp_prog, &xdp);
352 switch (act) {
353 case XDP_PASS:
354 err = xdp_update_frame_from_buff(&xdp, xdpf);
355 if (unlikely(err < 0))
356 xdp_return_frame_rx_napi(xdpf);
357 else
358 frames[nframes++] = xdpf;
359 break;
360 default:
361 bpf_warn_invalid_xdp_action(act);
362 fallthrough;
363 case XDP_ABORTED:
364 trace_xdp_exception(dev, xdp_prog, act);
365 fallthrough;
366 case XDP_DROP:
367 xdp_return_frame_rx_napi(xdpf);
368 break;
369 }
370 }
371 return nframes; /* sent frames count */
372}
373
374static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
375{
376 struct net_device *dev = bq->dev;
377 unsigned int cnt = bq->count;
378 int sent = 0, err = 0;
379 int to_send = cnt;
380 int i;
381
382 if (unlikely(!cnt))
383 return;
384
385 for (i = 0; i < cnt; i++) {
386 struct xdp_frame *xdpf = bq->q[i];
387
388 prefetch(xdpf);
389 }
390
391 if (bq->xdp_prog) {
392 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
393 if (!to_send)
394 goto out;
395 }
396
397 sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
398 if (sent < 0) {
399 /* If ndo_xdp_xmit fails with an errno, no frames have
400 * been xmit'ed.
401 */
402 err = sent;
403 sent = 0;
404 }
405
406 /* If not all frames have been transmitted, it is our
407 * responsibility to free them
408 */
409 for (i = sent; unlikely(i < to_send); i++)
410 xdp_return_frame_rx_napi(bq->q[i]);
411
412out:
413 bq->count = 0;
414 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
415}
416
417/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
418 * driver before returning from its napi->poll() routine. See the comment above
419 * xdp_do_flush() in filter.c.
420 */
421void __dev_flush(void)
422{
423 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
424 struct xdp_dev_bulk_queue *bq, *tmp;
425
426 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
427 bq_xmit_all(bq, XDP_XMIT_FLUSH);
428 bq->dev_rx = NULL;
429 bq->xdp_prog = NULL;
430 __list_del_clearprev(&bq->flush_node);
431 }
432}
433
434/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
435 * by local_bh_disable() (from XDP calls inside NAPI). The
436 * rcu_read_lock_bh_held() below makes lockdep accept both.
437 */
438static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
439{
440 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
441 struct bpf_dtab_netdev *obj;
442
443 if (key >= map->max_entries)
444 return NULL;
445
446 obj = rcu_dereference_check(dtab->netdev_map[key],
447 rcu_read_lock_bh_held());
448 return obj;
449}
450
451/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
452 * variable access, and map elements stick around. See comment above
453 * xdp_do_flush() in filter.c.
454 */
455static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
456 struct net_device *dev_rx, struct bpf_prog *xdp_prog)
457{
458 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
459 struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
460
461 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
462 bq_xmit_all(bq, 0);
463
464 /* Ingress dev_rx will be the same for all xdp_frame's in
465 * bulk_queue, because bq stored per-CPU and must be flushed
466 * from net_device drivers NAPI func end.
467 *
468 * Do the same with xdp_prog and flush_list since these fields
469 * are only ever modified together.
470 */
471 if (!bq->dev_rx) {
472 bq->dev_rx = dev_rx;
473 bq->xdp_prog = xdp_prog;
474 list_add(&bq->flush_node, flush_list);
475 }
476
477 bq->q[bq->count++] = xdpf;
478}
479
480static inline int __xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
481 struct net_device *dev_rx,
482 struct bpf_prog *xdp_prog)
483{
484 struct xdp_frame *xdpf;
485 int err;
486
487 if (!dev->netdev_ops->ndo_xdp_xmit)
488 return -EOPNOTSUPP;
489
490 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
491 if (unlikely(err))
492 return err;
493
494 xdpf = xdp_convert_buff_to_frame(xdp);
495 if (unlikely(!xdpf))
496 return -EOVERFLOW;
497
498 bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
499 return 0;
500}
501
502int dev_xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
503 struct net_device *dev_rx)
504{
505 return __xdp_enqueue(dev, xdp, dev_rx, NULL);
506}
507
508int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
509 struct net_device *dev_rx)
510{
511 struct net_device *dev = dst->dev;
512
513 return __xdp_enqueue(dev, xdp, dev_rx, dst->xdp_prog);
514}
515
516static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_buff *xdp,
517 int exclude_ifindex)
518{
519 if (!obj || obj->dev->ifindex == exclude_ifindex ||
520 !obj->dev->netdev_ops->ndo_xdp_xmit)
521 return false;
522
523 if (xdp_ok_fwd_dev(obj->dev, xdp->data_end - xdp->data))
524 return false;
525
526 return true;
527}
528
529static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
530 struct net_device *dev_rx,
531 struct xdp_frame *xdpf)
532{
533 struct xdp_frame *nxdpf;
534
535 nxdpf = xdpf_clone(xdpf);
536 if (!nxdpf)
537 return -ENOMEM;
538
539 bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
540
541 return 0;
542}
543
544int dev_map_enqueue_multi(struct xdp_buff *xdp, struct net_device *dev_rx,
545 struct bpf_map *map, bool exclude_ingress)
546{
547 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
548 int exclude_ifindex = exclude_ingress ? dev_rx->ifindex : 0;
549 struct bpf_dtab_netdev *dst, *last_dst = NULL;
550 struct hlist_head *head;
551 struct xdp_frame *xdpf;
552 unsigned int i;
553 int err;
554
555 xdpf = xdp_convert_buff_to_frame(xdp);
556 if (unlikely(!xdpf))
557 return -EOVERFLOW;
558
559 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
560 for (i = 0; i < map->max_entries; i++) {
561 dst = rcu_dereference_check(dtab->netdev_map[i],
562 rcu_read_lock_bh_held());
563 if (!is_valid_dst(dst, xdp, exclude_ifindex))
564 continue;
565
566 /* we only need n-1 clones; last_dst enqueued below */
567 if (!last_dst) {
568 last_dst = dst;
569 continue;
570 }
571
572 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
573 if (err)
574 return err;
575
576 last_dst = dst;
577 }
578 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
579 for (i = 0; i < dtab->n_buckets; i++) {
580 head = dev_map_index_hash(dtab, i);
581 hlist_for_each_entry_rcu(dst, head, index_hlist,
582 lockdep_is_held(&dtab->index_lock)) {
583 if (!is_valid_dst(dst, xdp, exclude_ifindex))
584 continue;
585
586 /* we only need n-1 clones; last_dst enqueued below */
587 if (!last_dst) {
588 last_dst = dst;
589 continue;
590 }
591
592 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
593 if (err)
594 return err;
595
596 last_dst = dst;
597 }
598 }
599 }
600
601 /* consume the last copy of the frame */
602 if (last_dst)
603 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
604 else
605 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
606
607 return 0;
608}
609
610int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
611 struct bpf_prog *xdp_prog)
612{
613 int err;
614
615 err = xdp_ok_fwd_dev(dst->dev, skb->len);
616 if (unlikely(err))
617 return err;
618 skb->dev = dst->dev;
619 generic_xdp_tx(skb, xdp_prog);
620
621 return 0;
622}
623
624static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
625 struct sk_buff *skb,
626 struct bpf_prog *xdp_prog)
627{
628 struct sk_buff *nskb;
629 int err;
630
631 nskb = skb_clone(skb, GFP_ATOMIC);
632 if (!nskb)
633 return -ENOMEM;
634
635 err = dev_map_generic_redirect(dst, nskb, xdp_prog);
636 if (unlikely(err)) {
637 consume_skb(nskb);
638 return err;
639 }
640
641 return 0;
642}
643
644int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
645 struct bpf_prog *xdp_prog, struct bpf_map *map,
646 bool exclude_ingress)
647{
648 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
649 int exclude_ifindex = exclude_ingress ? dev->ifindex : 0;
650 struct bpf_dtab_netdev *dst, *last_dst = NULL;
651 struct hlist_head *head;
652 struct hlist_node *next;
653 unsigned int i;
654 int err;
655
656 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
657 for (i = 0; i < map->max_entries; i++) {
658 dst = rcu_dereference_check(dtab->netdev_map[i],
659 rcu_read_lock_bh_held());
660 if (!dst || dst->dev->ifindex == exclude_ifindex)
661 continue;
662
663 /* we only need n-1 clones; last_dst enqueued below */
664 if (!last_dst) {
665 last_dst = dst;
666 continue;
667 }
668
669 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
670 if (err)
671 return err;
672
673 last_dst = dst;
674 }
675 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
676 for (i = 0; i < dtab->n_buckets; i++) {
677 head = dev_map_index_hash(dtab, i);
678 hlist_for_each_entry_safe(dst, next, head, index_hlist) {
679 if (!dst || dst->dev->ifindex == exclude_ifindex)
680 continue;
681
682 /* we only need n-1 clones; last_dst enqueued below */
683 if (!last_dst) {
684 last_dst = dst;
685 continue;
686 }
687
688 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
689 if (err)
690 return err;
691
692 last_dst = dst;
693 }
694 }
695 }
696
697 /* consume the first skb and return */
698 if (last_dst)
699 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
700
701 /* dtab is empty */
702 consume_skb(skb);
703 return 0;
704}
705
706static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
707{
708 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
709
710 return obj ? &obj->val : NULL;
711}
712
713static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
714{
715 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
716 *(u32 *)key);
717 return obj ? &obj->val : NULL;
718}
719
720static void __dev_map_entry_free(struct rcu_head *rcu)
721{
722 struct bpf_dtab_netdev *dev;
723
724 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
725 if (dev->xdp_prog)
726 bpf_prog_put(dev->xdp_prog);
727 dev_put(dev->dev);
728 kfree(dev);
729}
730
731static int dev_map_delete_elem(struct bpf_map *map, void *key)
732{
733 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
734 struct bpf_dtab_netdev *old_dev;
735 int k = *(u32 *)key;
736
737 if (k >= map->max_entries)
738 return -EINVAL;
739
740 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
741 if (old_dev)
742 call_rcu(&old_dev->rcu, __dev_map_entry_free);
743 return 0;
744}
745
746static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
747{
748 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
749 struct bpf_dtab_netdev *old_dev;
750 int k = *(u32 *)key;
751 unsigned long flags;
752 int ret = -ENOENT;
753
754 spin_lock_irqsave(&dtab->index_lock, flags);
755
756 old_dev = __dev_map_hash_lookup_elem(map, k);
757 if (old_dev) {
758 dtab->items--;
759 hlist_del_init_rcu(&old_dev->index_hlist);
760 call_rcu(&old_dev->rcu, __dev_map_entry_free);
761 ret = 0;
762 }
763 spin_unlock_irqrestore(&dtab->index_lock, flags);
764
765 return ret;
766}
767
768static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
769 struct bpf_dtab *dtab,
770 struct bpf_devmap_val *val,
771 unsigned int idx)
772{
773 struct bpf_prog *prog = NULL;
774 struct bpf_dtab_netdev *dev;
775
776 dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
777 GFP_ATOMIC | __GFP_NOWARN,
778 dtab->map.numa_node);
779 if (!dev)
780 return ERR_PTR(-ENOMEM);
781
782 dev->dev = dev_get_by_index(net, val->ifindex);
783 if (!dev->dev)
784 goto err_out;
785
786 if (val->bpf_prog.fd > 0) {
787 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
788 BPF_PROG_TYPE_XDP, false);
789 if (IS_ERR(prog))
790 goto err_put_dev;
791 if (prog->expected_attach_type != BPF_XDP_DEVMAP)
792 goto err_put_prog;
793 }
794
795 dev->idx = idx;
796 dev->dtab = dtab;
797 if (prog) {
798 dev->xdp_prog = prog;
799 dev->val.bpf_prog.id = prog->aux->id;
800 } else {
801 dev->xdp_prog = NULL;
802 dev->val.bpf_prog.id = 0;
803 }
804 dev->val.ifindex = val->ifindex;
805
806 return dev;
807err_put_prog:
808 bpf_prog_put(prog);
809err_put_dev:
810 dev_put(dev->dev);
811err_out:
812 kfree(dev);
813 return ERR_PTR(-EINVAL);
814}
815
816static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
817 void *key, void *value, u64 map_flags)
818{
819 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
820 struct bpf_dtab_netdev *dev, *old_dev;
821 struct bpf_devmap_val val = {};
822 u32 i = *(u32 *)key;
823
824 if (unlikely(map_flags > BPF_EXIST))
825 return -EINVAL;
826 if (unlikely(i >= dtab->map.max_entries))
827 return -E2BIG;
828 if (unlikely(map_flags == BPF_NOEXIST))
829 return -EEXIST;
830
831 /* already verified value_size <= sizeof val */
832 memcpy(&val, value, map->value_size);
833
834 if (!val.ifindex) {
835 dev = NULL;
836 /* can not specify fd if ifindex is 0 */
837 if (val.bpf_prog.fd > 0)
838 return -EINVAL;
839 } else {
840 dev = __dev_map_alloc_node(net, dtab, &val, i);
841 if (IS_ERR(dev))
842 return PTR_ERR(dev);
843 }
844
845 /* Use call_rcu() here to ensure rcu critical sections have completed
846 * Remembering the driver side flush operation will happen before the
847 * net device is removed.
848 */
849 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
850 if (old_dev)
851 call_rcu(&old_dev->rcu, __dev_map_entry_free);
852
853 return 0;
854}
855
856static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
857 u64 map_flags)
858{
859 return __dev_map_update_elem(current->nsproxy->net_ns,
860 map, key, value, map_flags);
861}
862
863static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
864 void *key, void *value, u64 map_flags)
865{
866 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
867 struct bpf_dtab_netdev *dev, *old_dev;
868 struct bpf_devmap_val val = {};
869 u32 idx = *(u32 *)key;
870 unsigned long flags;
871 int err = -EEXIST;
872
873 /* already verified value_size <= sizeof val */
874 memcpy(&val, value, map->value_size);
875
876 if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
877 return -EINVAL;
878
879 spin_lock_irqsave(&dtab->index_lock, flags);
880
881 old_dev = __dev_map_hash_lookup_elem(map, idx);
882 if (old_dev && (map_flags & BPF_NOEXIST))
883 goto out_err;
884
885 dev = __dev_map_alloc_node(net, dtab, &val, idx);
886 if (IS_ERR(dev)) {
887 err = PTR_ERR(dev);
888 goto out_err;
889 }
890
891 if (old_dev) {
892 hlist_del_rcu(&old_dev->index_hlist);
893 } else {
894 if (dtab->items >= dtab->map.max_entries) {
895 spin_unlock_irqrestore(&dtab->index_lock, flags);
896 call_rcu(&dev->rcu, __dev_map_entry_free);
897 return -E2BIG;
898 }
899 dtab->items++;
900 }
901
902 hlist_add_head_rcu(&dev->index_hlist,
903 dev_map_index_hash(dtab, idx));
904 spin_unlock_irqrestore(&dtab->index_lock, flags);
905
906 if (old_dev)
907 call_rcu(&old_dev->rcu, __dev_map_entry_free);
908
909 return 0;
910
911out_err:
912 spin_unlock_irqrestore(&dtab->index_lock, flags);
913 return err;
914}
915
916static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
917 u64 map_flags)
918{
919 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
920 map, key, value, map_flags);
921}
922
923static int dev_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
924{
925 return __bpf_xdp_redirect_map(map, ifindex, flags,
926 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
927 __dev_map_lookup_elem);
928}
929
930static int dev_hash_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
931{
932 return __bpf_xdp_redirect_map(map, ifindex, flags,
933 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
934 __dev_map_hash_lookup_elem);
935}
936
937static int dev_map_btf_id;
938const struct bpf_map_ops dev_map_ops = {
939 .map_meta_equal = bpf_map_meta_equal,
940 .map_alloc = dev_map_alloc,
941 .map_free = dev_map_free,
942 .map_get_next_key = dev_map_get_next_key,
943 .map_lookup_elem = dev_map_lookup_elem,
944 .map_update_elem = dev_map_update_elem,
945 .map_delete_elem = dev_map_delete_elem,
946 .map_check_btf = map_check_no_btf,
947 .map_btf_name = "bpf_dtab",
948 .map_btf_id = &dev_map_btf_id,
949 .map_redirect = dev_map_redirect,
950};
951
952static int dev_map_hash_map_btf_id;
953const struct bpf_map_ops dev_map_hash_ops = {
954 .map_meta_equal = bpf_map_meta_equal,
955 .map_alloc = dev_map_alloc,
956 .map_free = dev_map_free,
957 .map_get_next_key = dev_map_hash_get_next_key,
958 .map_lookup_elem = dev_map_hash_lookup_elem,
959 .map_update_elem = dev_map_hash_update_elem,
960 .map_delete_elem = dev_map_hash_delete_elem,
961 .map_check_btf = map_check_no_btf,
962 .map_btf_name = "bpf_dtab",
963 .map_btf_id = &dev_map_hash_map_btf_id,
964 .map_redirect = dev_hash_map_redirect,
965};
966
967static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
968 struct net_device *netdev)
969{
970 unsigned long flags;
971 u32 i;
972
973 spin_lock_irqsave(&dtab->index_lock, flags);
974 for (i = 0; i < dtab->n_buckets; i++) {
975 struct bpf_dtab_netdev *dev;
976 struct hlist_head *head;
977 struct hlist_node *next;
978
979 head = dev_map_index_hash(dtab, i);
980
981 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
982 if (netdev != dev->dev)
983 continue;
984
985 dtab->items--;
986 hlist_del_rcu(&dev->index_hlist);
987 call_rcu(&dev->rcu, __dev_map_entry_free);
988 }
989 }
990 spin_unlock_irqrestore(&dtab->index_lock, flags);
991}
992
993static int dev_map_notification(struct notifier_block *notifier,
994 ulong event, void *ptr)
995{
996 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
997 struct bpf_dtab *dtab;
998 int i, cpu;
999
1000 switch (event) {
1001 case NETDEV_REGISTER:
1002 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1003 break;
1004
1005 /* will be freed in free_netdev() */
1006 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1007 if (!netdev->xdp_bulkq)
1008 return NOTIFY_BAD;
1009
1010 for_each_possible_cpu(cpu)
1011 per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1012 break;
1013 case NETDEV_UNREGISTER:
1014 /* This rcu_read_lock/unlock pair is needed because
1015 * dev_map_list is an RCU list AND to ensure a delete
1016 * operation does not free a netdev_map entry while we
1017 * are comparing it against the netdev being unregistered.
1018 */
1019 rcu_read_lock();
1020 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1021 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1022 dev_map_hash_remove_netdev(dtab, netdev);
1023 continue;
1024 }
1025
1026 for (i = 0; i < dtab->map.max_entries; i++) {
1027 struct bpf_dtab_netdev *dev, *odev;
1028
1029 dev = rcu_dereference(dtab->netdev_map[i]);
1030 if (!dev || netdev != dev->dev)
1031 continue;
1032 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1033 if (dev == odev)
1034 call_rcu(&dev->rcu,
1035 __dev_map_entry_free);
1036 }
1037 }
1038 rcu_read_unlock();
1039 break;
1040 default:
1041 break;
1042 }
1043 return NOTIFY_OK;
1044}
1045
1046static struct notifier_block dev_map_notifier = {
1047 .notifier_call = dev_map_notification,
1048};
1049
1050static int __init dev_map_init(void)
1051{
1052 int cpu;
1053
1054 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1055 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1056 offsetof(struct _bpf_dtab_netdev, dev));
1057 register_netdevice_notifier(&dev_map_notifier);
1058
1059 for_each_possible_cpu(cpu)
1060 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1061 return 0;
1062}
1063
1064subsys_initcall(dev_map_init);