projects/linux·blob v6.10 · refs

1168 lines · 30.1 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#include <linux/btf_ids.h>
52
53#define DEV_CREATE_FLAG_MASK \
54 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
55
56struct xdp_dev_bulk_queue {
57 struct xdp_frame *q[DEV_MAP_BULK_SIZE];
58 struct list_head flush_node;
59 struct net_device *dev;
60 struct net_device *dev_rx;
61 struct bpf_prog *xdp_prog;
62 unsigned int count;
63};
64
65struct bpf_dtab_netdev {
66 struct net_device *dev; /* must be first member, due to tracepoint */
67 struct hlist_node index_hlist;
68 struct bpf_prog *xdp_prog;
69 struct rcu_head rcu;
70 unsigned int idx;
71 struct bpf_devmap_val val;
72};
73
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 /* hash table size must be power of 2; roundup_pow_of_two() can
134 * overflow into UB on 32-bit arches, so check that first
135 */
136 if (dtab->map.max_entries > 1UL << 31)
137 return -EINVAL;
138
139 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
140
141 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
142 dtab->map.numa_node);
143 if (!dtab->dev_index_head)
144 return -ENOMEM;
145
146 spin_lock_init(&dtab->index_lock);
147 } else {
148 dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
149 sizeof(struct bpf_dtab_netdev *),
150 dtab->map.numa_node);
151 if (!dtab->netdev_map)
152 return -ENOMEM;
153 }
154
155 return 0;
156}
157
158static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
159{
160 struct bpf_dtab *dtab;
161 int err;
162
163 dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
164 if (!dtab)
165 return ERR_PTR(-ENOMEM);
166
167 err = dev_map_init_map(dtab, attr);
168 if (err) {
169 bpf_map_area_free(dtab);
170 return ERR_PTR(err);
171 }
172
173 spin_lock(&dev_map_lock);
174 list_add_tail_rcu(&dtab->list, &dev_map_list);
175 spin_unlock(&dev_map_lock);
176
177 return &dtab->map;
178}
179
180static void dev_map_free(struct bpf_map *map)
181{
182 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
183 int i;
184
185 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
186 * so the programs (can be more than one that used this map) were
187 * disconnected from events. The following synchronize_rcu() guarantees
188 * both rcu read critical sections complete and waits for
189 * preempt-disable regions (NAPI being the relevant context here) so we
190 * are certain there will be no further reads against the netdev_map and
191 * all flush operations are complete. Flush operations can only be done
192 * from NAPI context for this reason.
193 */
194
195 spin_lock(&dev_map_lock);
196 list_del_rcu(&dtab->list);
197 spin_unlock(&dev_map_lock);
198
199 bpf_clear_redirect_map(map);
200 synchronize_rcu();
201
202 /* Make sure prior __dev_map_entry_free() have completed. */
203 rcu_barrier();
204
205 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
206 for (i = 0; i < dtab->n_buckets; i++) {
207 struct bpf_dtab_netdev *dev;
208 struct hlist_head *head;
209 struct hlist_node *next;
210
211 head = dev_map_index_hash(dtab, i);
212
213 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
214 hlist_del_rcu(&dev->index_hlist);
215 if (dev->xdp_prog)
216 bpf_prog_put(dev->xdp_prog);
217 dev_put(dev->dev);
218 kfree(dev);
219 }
220 }
221
222 bpf_map_area_free(dtab->dev_index_head);
223 } else {
224 for (i = 0; i < dtab->map.max_entries; i++) {
225 struct bpf_dtab_netdev *dev;
226
227 dev = rcu_dereference_raw(dtab->netdev_map[i]);
228 if (!dev)
229 continue;
230
231 if (dev->xdp_prog)
232 bpf_prog_put(dev->xdp_prog);
233 dev_put(dev->dev);
234 kfree(dev);
235 }
236
237 bpf_map_area_free(dtab->netdev_map);
238 }
239
240 bpf_map_area_free(dtab);
241}
242
243static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
244{
245 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
246 u32 index = key ? *(u32 *)key : U32_MAX;
247 u32 *next = next_key;
248
249 if (index >= dtab->map.max_entries) {
250 *next = 0;
251 return 0;
252 }
253
254 if (index == dtab->map.max_entries - 1)
255 return -ENOENT;
256 *next = index + 1;
257 return 0;
258}
259
260/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
261 * by local_bh_disable() (from XDP calls inside NAPI). The
262 * rcu_read_lock_bh_held() below makes lockdep accept both.
263 */
264static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
265{
266 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
267 struct hlist_head *head = dev_map_index_hash(dtab, key);
268 struct bpf_dtab_netdev *dev;
269
270 hlist_for_each_entry_rcu(dev, head, index_hlist,
271 lockdep_is_held(&dtab->index_lock))
272 if (dev->idx == key)
273 return dev;
274
275 return NULL;
276}
277
278static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
279 void *next_key)
280{
281 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
282 u32 idx, *next = next_key;
283 struct bpf_dtab_netdev *dev, *next_dev;
284 struct hlist_head *head;
285 int i = 0;
286
287 if (!key)
288 goto find_first;
289
290 idx = *(u32 *)key;
291
292 dev = __dev_map_hash_lookup_elem(map, idx);
293 if (!dev)
294 goto find_first;
295
296 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
297 struct bpf_dtab_netdev, index_hlist);
298
299 if (next_dev) {
300 *next = next_dev->idx;
301 return 0;
302 }
303
304 i = idx & (dtab->n_buckets - 1);
305 i++;
306
307 find_first:
308 for (; i < dtab->n_buckets; i++) {
309 head = dev_map_index_hash(dtab, i);
310
311 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
312 struct bpf_dtab_netdev,
313 index_hlist);
314 if (next_dev) {
315 *next = next_dev->idx;
316 return 0;
317 }
318 }
319
320 return -ENOENT;
321}
322
323static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
324 struct xdp_frame **frames, int n,
325 struct net_device *dev)
326{
327 struct xdp_txq_info txq = { .dev = dev };
328 struct xdp_buff xdp;
329 int i, nframes = 0;
330
331 for (i = 0; i < n; i++) {
332 struct xdp_frame *xdpf = frames[i];
333 u32 act;
334 int err;
335
336 xdp_convert_frame_to_buff(xdpf, &xdp);
337 xdp.txq = &txq;
338
339 act = bpf_prog_run_xdp(xdp_prog, &xdp);
340 switch (act) {
341 case XDP_PASS:
342 err = xdp_update_frame_from_buff(&xdp, xdpf);
343 if (unlikely(err < 0))
344 xdp_return_frame_rx_napi(xdpf);
345 else
346 frames[nframes++] = xdpf;
347 break;
348 default:
349 bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
350 fallthrough;
351 case XDP_ABORTED:
352 trace_xdp_exception(dev, xdp_prog, act);
353 fallthrough;
354 case XDP_DROP:
355 xdp_return_frame_rx_napi(xdpf);
356 break;
357 }
358 }
359 return nframes; /* sent frames count */
360}
361
362static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
363{
364 struct net_device *dev = bq->dev;
365 unsigned int cnt = bq->count;
366 int sent = 0, err = 0;
367 int to_send = cnt;
368 int i;
369
370 if (unlikely(!cnt))
371 return;
372
373 for (i = 0; i < cnt; i++) {
374 struct xdp_frame *xdpf = bq->q[i];
375
376 prefetch(xdpf);
377 }
378
379 if (bq->xdp_prog) {
380 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
381 if (!to_send)
382 goto out;
383 }
384
385 sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
386 if (sent < 0) {
387 /* If ndo_xdp_xmit fails with an errno, no frames have
388 * been xmit'ed.
389 */
390 err = sent;
391 sent = 0;
392 }
393
394 /* If not all frames have been transmitted, it is our
395 * responsibility to free them
396 */
397 for (i = sent; unlikely(i < to_send); i++)
398 xdp_return_frame_rx_napi(bq->q[i]);
399
400out:
401 bq->count = 0;
402 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
403}
404
405/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
406 * driver before returning from its napi->poll() routine. See the comment above
407 * xdp_do_flush() in filter.c.
408 */
409void __dev_flush(void)
410{
411 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
412 struct xdp_dev_bulk_queue *bq, *tmp;
413
414 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
415 bq_xmit_all(bq, XDP_XMIT_FLUSH);
416 bq->dev_rx = NULL;
417 bq->xdp_prog = NULL;
418 __list_del_clearprev(&bq->flush_node);
419 }
420}
421
422#ifdef CONFIG_DEBUG_NET
423bool dev_check_flush(void)
424{
425 if (list_empty(this_cpu_ptr(&dev_flush_list)))
426 return false;
427 __dev_flush();
428 return true;
429}
430#endif
431
432/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
433 * by local_bh_disable() (from XDP calls inside NAPI). The
434 * rcu_read_lock_bh_held() below makes lockdep accept both.
435 */
436static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
437{
438 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
439 struct bpf_dtab_netdev *obj;
440
441 if (key >= map->max_entries)
442 return NULL;
443
444 obj = rcu_dereference_check(dtab->netdev_map[key],
445 rcu_read_lock_bh_held());
446 return obj;
447}
448
449/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
450 * variable access, and map elements stick around. See comment above
451 * xdp_do_flush() in filter.c.
452 */
453static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
454 struct net_device *dev_rx, struct bpf_prog *xdp_prog)
455{
456 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
457 struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
458
459 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
460 bq_xmit_all(bq, 0);
461
462 /* Ingress dev_rx will be the same for all xdp_frame's in
463 * bulk_queue, because bq stored per-CPU and must be flushed
464 * from net_device drivers NAPI func end.
465 *
466 * Do the same with xdp_prog and flush_list since these fields
467 * are only ever modified together.
468 */
469 if (!bq->dev_rx) {
470 bq->dev_rx = dev_rx;
471 bq->xdp_prog = xdp_prog;
472 list_add(&bq->flush_node, flush_list);
473 }
474
475 bq->q[bq->count++] = xdpf;
476}
477
478static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
479 struct net_device *dev_rx,
480 struct bpf_prog *xdp_prog)
481{
482 int err;
483
484 if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
485 return -EOPNOTSUPP;
486
487 if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
488 xdp_frame_has_frags(xdpf)))
489 return -EOPNOTSUPP;
490
491 err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
492 if (unlikely(err))
493 return err;
494
495 bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
496 return 0;
497}
498
499static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
500{
501 struct xdp_txq_info txq = { .dev = dst->dev };
502 struct xdp_buff xdp;
503 u32 act;
504
505 if (!dst->xdp_prog)
506 return XDP_PASS;
507
508 __skb_pull(skb, skb->mac_len);
509 xdp.txq = &txq;
510
511 act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
512 switch (act) {
513 case XDP_PASS:
514 __skb_push(skb, skb->mac_len);
515 break;
516 default:
517 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
518 fallthrough;
519 case XDP_ABORTED:
520 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
521 fallthrough;
522 case XDP_DROP:
523 kfree_skb(skb);
524 break;
525 }
526
527 return act;
528}
529
530int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
531 struct net_device *dev_rx)
532{
533 return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
534}
535
536int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
537 struct net_device *dev_rx)
538{
539 struct net_device *dev = dst->dev;
540
541 return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
542}
543
544static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
545{
546 if (!obj)
547 return false;
548
549 if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
550 return false;
551
552 if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
553 xdp_frame_has_frags(xdpf)))
554 return false;
555
556 if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
557 return false;
558
559 return true;
560}
561
562static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
563 struct net_device *dev_rx,
564 struct xdp_frame *xdpf)
565{
566 struct xdp_frame *nxdpf;
567
568 nxdpf = xdpf_clone(xdpf);
569 if (!nxdpf)
570 return -ENOMEM;
571
572 bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
573
574 return 0;
575}
576
577static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
578{
579 while (num_excluded--) {
580 if (ifindex == excluded[num_excluded])
581 return true;
582 }
583 return false;
584}
585
586/* Get ifindex of each upper device. 'indexes' must be able to hold at
587 * least MAX_NEST_DEV elements.
588 * Returns the number of ifindexes added.
589 */
590static int get_upper_ifindexes(struct net_device *dev, int *indexes)
591{
592 struct net_device *upper;
593 struct list_head *iter;
594 int n = 0;
595
596 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
597 indexes[n++] = upper->ifindex;
598 }
599 return n;
600}
601
602int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
603 struct bpf_map *map, bool exclude_ingress)
604{
605 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
606 struct bpf_dtab_netdev *dst, *last_dst = NULL;
607 int excluded_devices[1+MAX_NEST_DEV];
608 struct hlist_head *head;
609 int num_excluded = 0;
610 unsigned int i;
611 int err;
612
613 if (exclude_ingress) {
614 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
615 excluded_devices[num_excluded++] = dev_rx->ifindex;
616 }
617
618 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
619 for (i = 0; i < map->max_entries; i++) {
620 dst = rcu_dereference_check(dtab->netdev_map[i],
621 rcu_read_lock_bh_held());
622 if (!is_valid_dst(dst, xdpf))
623 continue;
624
625 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
626 continue;
627
628 /* we only need n-1 clones; last_dst enqueued below */
629 if (!last_dst) {
630 last_dst = dst;
631 continue;
632 }
633
634 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
635 if (err)
636 return err;
637
638 last_dst = dst;
639 }
640 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
641 for (i = 0; i < dtab->n_buckets; i++) {
642 head = dev_map_index_hash(dtab, i);
643 hlist_for_each_entry_rcu(dst, head, index_hlist,
644 lockdep_is_held(&dtab->index_lock)) {
645 if (!is_valid_dst(dst, xdpf))
646 continue;
647
648 if (is_ifindex_excluded(excluded_devices, num_excluded,
649 dst->dev->ifindex))
650 continue;
651
652 /* we only need n-1 clones; last_dst enqueued below */
653 if (!last_dst) {
654 last_dst = dst;
655 continue;
656 }
657
658 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
659 if (err)
660 return err;
661
662 last_dst = dst;
663 }
664 }
665 }
666
667 /* consume the last copy of the frame */
668 if (last_dst)
669 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
670 else
671 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
672
673 return 0;
674}
675
676int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
677 struct bpf_prog *xdp_prog)
678{
679 int err;
680
681 err = xdp_ok_fwd_dev(dst->dev, skb->len);
682 if (unlikely(err))
683 return err;
684
685 /* Redirect has already succeeded semantically at this point, so we just
686 * return 0 even if packet is dropped. Helper below takes care of
687 * freeing skb.
688 */
689 if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
690 return 0;
691
692 skb->dev = dst->dev;
693 generic_xdp_tx(skb, xdp_prog);
694
695 return 0;
696}
697
698static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
699 struct sk_buff *skb,
700 struct bpf_prog *xdp_prog)
701{
702 struct sk_buff *nskb;
703 int err;
704
705 nskb = skb_clone(skb, GFP_ATOMIC);
706 if (!nskb)
707 return -ENOMEM;
708
709 err = dev_map_generic_redirect(dst, nskb, xdp_prog);
710 if (unlikely(err)) {
711 consume_skb(nskb);
712 return err;
713 }
714
715 return 0;
716}
717
718int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
719 struct bpf_prog *xdp_prog, struct bpf_map *map,
720 bool exclude_ingress)
721{
722 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
723 struct bpf_dtab_netdev *dst, *last_dst = NULL;
724 int excluded_devices[1+MAX_NEST_DEV];
725 struct hlist_head *head;
726 struct hlist_node *next;
727 int num_excluded = 0;
728 unsigned int i;
729 int err;
730
731 if (exclude_ingress) {
732 num_excluded = get_upper_ifindexes(dev, excluded_devices);
733 excluded_devices[num_excluded++] = dev->ifindex;
734 }
735
736 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
737 for (i = 0; i < map->max_entries; i++) {
738 dst = rcu_dereference_check(dtab->netdev_map[i],
739 rcu_read_lock_bh_held());
740 if (!dst)
741 continue;
742
743 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
744 continue;
745
746 /* we only need n-1 clones; last_dst enqueued below */
747 if (!last_dst) {
748 last_dst = dst;
749 continue;
750 }
751
752 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
753 if (err)
754 return err;
755
756 last_dst = dst;
757
758 }
759 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
760 for (i = 0; i < dtab->n_buckets; i++) {
761 head = dev_map_index_hash(dtab, i);
762 hlist_for_each_entry_safe(dst, next, head, index_hlist) {
763 if (is_ifindex_excluded(excluded_devices, num_excluded,
764 dst->dev->ifindex))
765 continue;
766
767 /* we only need n-1 clones; last_dst enqueued below */
768 if (!last_dst) {
769 last_dst = dst;
770 continue;
771 }
772
773 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
774 if (err)
775 return err;
776
777 last_dst = dst;
778 }
779 }
780 }
781
782 /* consume the first skb and return */
783 if (last_dst)
784 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
785
786 /* dtab is empty */
787 consume_skb(skb);
788 return 0;
789}
790
791static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
792{
793 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
794
795 return obj ? &obj->val : NULL;
796}
797
798static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
799{
800 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
801 *(u32 *)key);
802 return obj ? &obj->val : NULL;
803}
804
805static void __dev_map_entry_free(struct rcu_head *rcu)
806{
807 struct bpf_dtab_netdev *dev;
808
809 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
810 if (dev->xdp_prog)
811 bpf_prog_put(dev->xdp_prog);
812 dev_put(dev->dev);
813 kfree(dev);
814}
815
816static long dev_map_delete_elem(struct bpf_map *map, void *key)
817{
818 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
819 struct bpf_dtab_netdev *old_dev;
820 int k = *(u32 *)key;
821
822 if (k >= map->max_entries)
823 return -EINVAL;
824
825 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
826 if (old_dev) {
827 call_rcu(&old_dev->rcu, __dev_map_entry_free);
828 atomic_dec((atomic_t *)&dtab->items);
829 }
830 return 0;
831}
832
833static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
834{
835 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
836 struct bpf_dtab_netdev *old_dev;
837 int k = *(u32 *)key;
838 unsigned long flags;
839 int ret = -ENOENT;
840
841 spin_lock_irqsave(&dtab->index_lock, flags);
842
843 old_dev = __dev_map_hash_lookup_elem(map, k);
844 if (old_dev) {
845 dtab->items--;
846 hlist_del_init_rcu(&old_dev->index_hlist);
847 call_rcu(&old_dev->rcu, __dev_map_entry_free);
848 ret = 0;
849 }
850 spin_unlock_irqrestore(&dtab->index_lock, flags);
851
852 return ret;
853}
854
855static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
856 struct bpf_dtab *dtab,
857 struct bpf_devmap_val *val,
858 unsigned int idx)
859{
860 struct bpf_prog *prog = NULL;
861 struct bpf_dtab_netdev *dev;
862
863 dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
864 GFP_NOWAIT | __GFP_NOWARN,
865 dtab->map.numa_node);
866 if (!dev)
867 return ERR_PTR(-ENOMEM);
868
869 dev->dev = dev_get_by_index(net, val->ifindex);
870 if (!dev->dev)
871 goto err_out;
872
873 if (val->bpf_prog.fd > 0) {
874 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
875 BPF_PROG_TYPE_XDP, false);
876 if (IS_ERR(prog))
877 goto err_put_dev;
878 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
879 !bpf_prog_map_compatible(&dtab->map, prog))
880 goto err_put_prog;
881 }
882
883 dev->idx = idx;
884 if (prog) {
885 dev->xdp_prog = prog;
886 dev->val.bpf_prog.id = prog->aux->id;
887 } else {
888 dev->xdp_prog = NULL;
889 dev->val.bpf_prog.id = 0;
890 }
891 dev->val.ifindex = val->ifindex;
892
893 return dev;
894err_put_prog:
895 bpf_prog_put(prog);
896err_put_dev:
897 dev_put(dev->dev);
898err_out:
899 kfree(dev);
900 return ERR_PTR(-EINVAL);
901}
902
903static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
904 void *key, void *value, u64 map_flags)
905{
906 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
907 struct bpf_dtab_netdev *dev, *old_dev;
908 struct bpf_devmap_val val = {};
909 u32 i = *(u32 *)key;
910
911 if (unlikely(map_flags > BPF_EXIST))
912 return -EINVAL;
913 if (unlikely(i >= dtab->map.max_entries))
914 return -E2BIG;
915 if (unlikely(map_flags == BPF_NOEXIST))
916 return -EEXIST;
917
918 /* already verified value_size <= sizeof val */
919 memcpy(&val, value, map->value_size);
920
921 if (!val.ifindex) {
922 dev = NULL;
923 /* can not specify fd if ifindex is 0 */
924 if (val.bpf_prog.fd > 0)
925 return -EINVAL;
926 } else {
927 dev = __dev_map_alloc_node(net, dtab, &val, i);
928 if (IS_ERR(dev))
929 return PTR_ERR(dev);
930 }
931
932 /* Use call_rcu() here to ensure rcu critical sections have completed
933 * Remembering the driver side flush operation will happen before the
934 * net device is removed.
935 */
936 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
937 if (old_dev)
938 call_rcu(&old_dev->rcu, __dev_map_entry_free);
939 else
940 atomic_inc((atomic_t *)&dtab->items);
941
942 return 0;
943}
944
945static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
946 u64 map_flags)
947{
948 return __dev_map_update_elem(current->nsproxy->net_ns,
949 map, key, value, map_flags);
950}
951
952static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
953 void *key, void *value, u64 map_flags)
954{
955 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
956 struct bpf_dtab_netdev *dev, *old_dev;
957 struct bpf_devmap_val val = {};
958 u32 idx = *(u32 *)key;
959 unsigned long flags;
960 int err = -EEXIST;
961
962 /* already verified value_size <= sizeof val */
963 memcpy(&val, value, map->value_size);
964
965 if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
966 return -EINVAL;
967
968 spin_lock_irqsave(&dtab->index_lock, flags);
969
970 old_dev = __dev_map_hash_lookup_elem(map, idx);
971 if (old_dev && (map_flags & BPF_NOEXIST))
972 goto out_err;
973
974 dev = __dev_map_alloc_node(net, dtab, &val, idx);
975 if (IS_ERR(dev)) {
976 err = PTR_ERR(dev);
977 goto out_err;
978 }
979
980 if (old_dev) {
981 hlist_del_rcu(&old_dev->index_hlist);
982 } else {
983 if (dtab->items >= dtab->map.max_entries) {
984 spin_unlock_irqrestore(&dtab->index_lock, flags);
985 call_rcu(&dev->rcu, __dev_map_entry_free);
986 return -E2BIG;
987 }
988 dtab->items++;
989 }
990
991 hlist_add_head_rcu(&dev->index_hlist,
992 dev_map_index_hash(dtab, idx));
993 spin_unlock_irqrestore(&dtab->index_lock, flags);
994
995 if (old_dev)
996 call_rcu(&old_dev->rcu, __dev_map_entry_free);
997
998 return 0;
999
1000out_err:
1001 spin_unlock_irqrestore(&dtab->index_lock, flags);
1002 return err;
1003}
1004
1005static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1006 u64 map_flags)
1007{
1008 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1009 map, key, value, map_flags);
1010}
1011
1012static long dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1013{
1014 return __bpf_xdp_redirect_map(map, ifindex, flags,
1015 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1016 __dev_map_lookup_elem);
1017}
1018
1019static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1020{
1021 return __bpf_xdp_redirect_map(map, ifindex, flags,
1022 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1023 __dev_map_hash_lookup_elem);
1024}
1025
1026static u64 dev_map_mem_usage(const struct bpf_map *map)
1027{
1028 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1029 u64 usage = sizeof(struct bpf_dtab);
1030
1031 if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1032 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1033 else
1034 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1035 usage += atomic_read((atomic_t *)&dtab->items) *
1036 (u64)sizeof(struct bpf_dtab_netdev);
1037 return usage;
1038}
1039
1040BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1041const struct bpf_map_ops dev_map_ops = {
1042 .map_meta_equal = bpf_map_meta_equal,
1043 .map_alloc = dev_map_alloc,
1044 .map_free = dev_map_free,
1045 .map_get_next_key = dev_map_get_next_key,
1046 .map_lookup_elem = dev_map_lookup_elem,
1047 .map_update_elem = dev_map_update_elem,
1048 .map_delete_elem = dev_map_delete_elem,
1049 .map_check_btf = map_check_no_btf,
1050 .map_mem_usage = dev_map_mem_usage,
1051 .map_btf_id = &dev_map_btf_ids[0],
1052 .map_redirect = dev_map_redirect,
1053};
1054
1055const struct bpf_map_ops dev_map_hash_ops = {
1056 .map_meta_equal = bpf_map_meta_equal,
1057 .map_alloc = dev_map_alloc,
1058 .map_free = dev_map_free,
1059 .map_get_next_key = dev_map_hash_get_next_key,
1060 .map_lookup_elem = dev_map_hash_lookup_elem,
1061 .map_update_elem = dev_map_hash_update_elem,
1062 .map_delete_elem = dev_map_hash_delete_elem,
1063 .map_check_btf = map_check_no_btf,
1064 .map_mem_usage = dev_map_mem_usage,
1065 .map_btf_id = &dev_map_btf_ids[0],
1066 .map_redirect = dev_hash_map_redirect,
1067};
1068
1069static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1070 struct net_device *netdev)
1071{
1072 unsigned long flags;
1073 u32 i;
1074
1075 spin_lock_irqsave(&dtab->index_lock, flags);
1076 for (i = 0; i < dtab->n_buckets; i++) {
1077 struct bpf_dtab_netdev *dev;
1078 struct hlist_head *head;
1079 struct hlist_node *next;
1080
1081 head = dev_map_index_hash(dtab, i);
1082
1083 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1084 if (netdev != dev->dev)
1085 continue;
1086
1087 dtab->items--;
1088 hlist_del_rcu(&dev->index_hlist);
1089 call_rcu(&dev->rcu, __dev_map_entry_free);
1090 }
1091 }
1092 spin_unlock_irqrestore(&dtab->index_lock, flags);
1093}
1094
1095static int dev_map_notification(struct notifier_block *notifier,
1096 ulong event, void *ptr)
1097{
1098 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1099 struct bpf_dtab *dtab;
1100 int i, cpu;
1101
1102 switch (event) {
1103 case NETDEV_REGISTER:
1104 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1105 break;
1106
1107 /* will be freed in free_netdev() */
1108 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1109 if (!netdev->xdp_bulkq)
1110 return NOTIFY_BAD;
1111
1112 for_each_possible_cpu(cpu)
1113 per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1114 break;
1115 case NETDEV_UNREGISTER:
1116 /* This rcu_read_lock/unlock pair is needed because
1117 * dev_map_list is an RCU list AND to ensure a delete
1118 * operation does not free a netdev_map entry while we
1119 * are comparing it against the netdev being unregistered.
1120 */
1121 rcu_read_lock();
1122 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1123 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1124 dev_map_hash_remove_netdev(dtab, netdev);
1125 continue;
1126 }
1127
1128 for (i = 0; i < dtab->map.max_entries; i++) {
1129 struct bpf_dtab_netdev *dev, *odev;
1130
1131 dev = rcu_dereference(dtab->netdev_map[i]);
1132 if (!dev || netdev != dev->dev)
1133 continue;
1134 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1135 if (dev == odev) {
1136 call_rcu(&dev->rcu,
1137 __dev_map_entry_free);
1138 atomic_dec((atomic_t *)&dtab->items);
1139 }
1140 }
1141 }
1142 rcu_read_unlock();
1143 break;
1144 default:
1145 break;
1146 }
1147 return NOTIFY_OK;
1148}
1149
1150static struct notifier_block dev_map_notifier = {
1151 .notifier_call = dev_map_notification,
1152};
1153
1154static int __init dev_map_init(void)
1155{
1156 int cpu;
1157
1158 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1159 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1160 offsetof(struct _bpf_dtab_netdev, dev));
1161 register_netdevice_notifier(&dev_map_notifier);
1162
1163 for_each_possible_cpu(cpu)
1164 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1165 return 0;
1166}
1167
1168subsys_initcall(dev_map_init);