projects/linux·blob v5.17 · refs

1134 lines · 29.0 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
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
325static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
326 struct xdp_frame **frames, int n,
327 struct net_device *dev)
328{
329 struct xdp_txq_info txq = { .dev = dev };
330 struct xdp_buff xdp;
331 int i, nframes = 0;
332
333 for (i = 0; i < n; i++) {
334 struct xdp_frame *xdpf = frames[i];
335 u32 act;
336 int err;
337
338 xdp_convert_frame_to_buff(xdpf, &xdp);
339 xdp.txq = &txq;
340
341 act = bpf_prog_run_xdp(xdp_prog, &xdp);
342 switch (act) {
343 case XDP_PASS:
344 err = xdp_update_frame_from_buff(&xdp, xdpf);
345 if (unlikely(err < 0))
346 xdp_return_frame_rx_napi(xdpf);
347 else
348 frames[nframes++] = xdpf;
349 break;
350 default:
351 bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
352 fallthrough;
353 case XDP_ABORTED:
354 trace_xdp_exception(dev, xdp_prog, act);
355 fallthrough;
356 case XDP_DROP:
357 xdp_return_frame_rx_napi(xdpf);
358 break;
359 }
360 }
361 return nframes; /* sent frames count */
362}
363
364static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
365{
366 struct net_device *dev = bq->dev;
367 unsigned int cnt = bq->count;
368 int sent = 0, err = 0;
369 int to_send = cnt;
370 int i;
371
372 if (unlikely(!cnt))
373 return;
374
375 for (i = 0; i < cnt; i++) {
376 struct xdp_frame *xdpf = bq->q[i];
377
378 prefetch(xdpf);
379 }
380
381 if (bq->xdp_prog) {
382 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
383 if (!to_send)
384 goto out;
385 }
386
387 sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
388 if (sent < 0) {
389 /* If ndo_xdp_xmit fails with an errno, no frames have
390 * been xmit'ed.
391 */
392 err = sent;
393 sent = 0;
394 }
395
396 /* If not all frames have been transmitted, it is our
397 * responsibility to free them
398 */
399 for (i = sent; unlikely(i < to_send); i++)
400 xdp_return_frame_rx_napi(bq->q[i]);
401
402out:
403 bq->count = 0;
404 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
405}
406
407/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
408 * driver before returning from its napi->poll() routine. See the comment above
409 * xdp_do_flush() in filter.c.
410 */
411void __dev_flush(void)
412{
413 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
414 struct xdp_dev_bulk_queue *bq, *tmp;
415
416 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
417 bq_xmit_all(bq, XDP_XMIT_FLUSH);
418 bq->dev_rx = NULL;
419 bq->xdp_prog = NULL;
420 __list_del_clearprev(&bq->flush_node);
421 }
422}
423
424/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
425 * by local_bh_disable() (from XDP calls inside NAPI). The
426 * rcu_read_lock_bh_held() below makes lockdep accept both.
427 */
428static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
429{
430 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
431 struct bpf_dtab_netdev *obj;
432
433 if (key >= map->max_entries)
434 return NULL;
435
436 obj = rcu_dereference_check(dtab->netdev_map[key],
437 rcu_read_lock_bh_held());
438 return obj;
439}
440
441/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
442 * variable access, and map elements stick around. See comment above
443 * xdp_do_flush() in filter.c.
444 */
445static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
446 struct net_device *dev_rx, struct bpf_prog *xdp_prog)
447{
448 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
449 struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
450
451 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
452 bq_xmit_all(bq, 0);
453
454 /* Ingress dev_rx will be the same for all xdp_frame's in
455 * bulk_queue, because bq stored per-CPU and must be flushed
456 * from net_device drivers NAPI func end.
457 *
458 * Do the same with xdp_prog and flush_list since these fields
459 * are only ever modified together.
460 */
461 if (!bq->dev_rx) {
462 bq->dev_rx = dev_rx;
463 bq->xdp_prog = xdp_prog;
464 list_add(&bq->flush_node, flush_list);
465 }
466
467 bq->q[bq->count++] = xdpf;
468}
469
470static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
471 struct net_device *dev_rx,
472 struct bpf_prog *xdp_prog)
473{
474 int err;
475
476 if (!dev->netdev_ops->ndo_xdp_xmit)
477 return -EOPNOTSUPP;
478
479 err = xdp_ok_fwd_dev(dev, xdpf->len);
480 if (unlikely(err))
481 return err;
482
483 bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
484 return 0;
485}
486
487static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
488{
489 struct xdp_txq_info txq = { .dev = dst->dev };
490 struct xdp_buff xdp;
491 u32 act;
492
493 if (!dst->xdp_prog)
494 return XDP_PASS;
495
496 __skb_pull(skb, skb->mac_len);
497 xdp.txq = &txq;
498
499 act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
500 switch (act) {
501 case XDP_PASS:
502 __skb_push(skb, skb->mac_len);
503 break;
504 default:
505 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
506 fallthrough;
507 case XDP_ABORTED:
508 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
509 fallthrough;
510 case XDP_DROP:
511 kfree_skb(skb);
512 break;
513 }
514
515 return act;
516}
517
518int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
519 struct net_device *dev_rx)
520{
521 return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
522}
523
524int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
525 struct net_device *dev_rx)
526{
527 struct net_device *dev = dst->dev;
528
529 return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
530}
531
532static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
533{
534 if (!obj ||
535 !obj->dev->netdev_ops->ndo_xdp_xmit)
536 return false;
537
538 if (xdp_ok_fwd_dev(obj->dev, xdpf->len))
539 return false;
540
541 return true;
542}
543
544static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
545 struct net_device *dev_rx,
546 struct xdp_frame *xdpf)
547{
548 struct xdp_frame *nxdpf;
549
550 nxdpf = xdpf_clone(xdpf);
551 if (!nxdpf)
552 return -ENOMEM;
553
554 bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
555
556 return 0;
557}
558
559static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
560{
561 while (num_excluded--) {
562 if (ifindex == excluded[num_excluded])
563 return true;
564 }
565 return false;
566}
567
568/* Get ifindex of each upper device. 'indexes' must be able to hold at
569 * least MAX_NEST_DEV elements.
570 * Returns the number of ifindexes added.
571 */
572static int get_upper_ifindexes(struct net_device *dev, int *indexes)
573{
574 struct net_device *upper;
575 struct list_head *iter;
576 int n = 0;
577
578 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
579 indexes[n++] = upper->ifindex;
580 }
581 return n;
582}
583
584int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
585 struct bpf_map *map, bool exclude_ingress)
586{
587 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
588 struct bpf_dtab_netdev *dst, *last_dst = NULL;
589 int excluded_devices[1+MAX_NEST_DEV];
590 struct hlist_head *head;
591 int num_excluded = 0;
592 unsigned int i;
593 int err;
594
595 if (exclude_ingress) {
596 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
597 excluded_devices[num_excluded++] = dev_rx->ifindex;
598 }
599
600 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
601 for (i = 0; i < map->max_entries; i++) {
602 dst = rcu_dereference_check(dtab->netdev_map[i],
603 rcu_read_lock_bh_held());
604 if (!is_valid_dst(dst, xdpf))
605 continue;
606
607 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
608 continue;
609
610 /* we only need n-1 clones; last_dst enqueued below */
611 if (!last_dst) {
612 last_dst = dst;
613 continue;
614 }
615
616 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
617 if (err)
618 return err;
619
620 last_dst = dst;
621 }
622 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
623 for (i = 0; i < dtab->n_buckets; i++) {
624 head = dev_map_index_hash(dtab, i);
625 hlist_for_each_entry_rcu(dst, head, index_hlist,
626 lockdep_is_held(&dtab->index_lock)) {
627 if (!is_valid_dst(dst, xdpf))
628 continue;
629
630 if (is_ifindex_excluded(excluded_devices, num_excluded,
631 dst->dev->ifindex))
632 continue;
633
634 /* we only need n-1 clones; last_dst enqueued below */
635 if (!last_dst) {
636 last_dst = dst;
637 continue;
638 }
639
640 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
641 if (err)
642 return err;
643
644 last_dst = dst;
645 }
646 }
647 }
648
649 /* consume the last copy of the frame */
650 if (last_dst)
651 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
652 else
653 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
654
655 return 0;
656}
657
658int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
659 struct bpf_prog *xdp_prog)
660{
661 int err;
662
663 err = xdp_ok_fwd_dev(dst->dev, skb->len);
664 if (unlikely(err))
665 return err;
666
667 /* Redirect has already succeeded semantically at this point, so we just
668 * return 0 even if packet is dropped. Helper below takes care of
669 * freeing skb.
670 */
671 if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
672 return 0;
673
674 skb->dev = dst->dev;
675 generic_xdp_tx(skb, xdp_prog);
676
677 return 0;
678}
679
680static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
681 struct sk_buff *skb,
682 struct bpf_prog *xdp_prog)
683{
684 struct sk_buff *nskb;
685 int err;
686
687 nskb = skb_clone(skb, GFP_ATOMIC);
688 if (!nskb)
689 return -ENOMEM;
690
691 err = dev_map_generic_redirect(dst, nskb, xdp_prog);
692 if (unlikely(err)) {
693 consume_skb(nskb);
694 return err;
695 }
696
697 return 0;
698}
699
700int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
701 struct bpf_prog *xdp_prog, struct bpf_map *map,
702 bool exclude_ingress)
703{
704 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
705 struct bpf_dtab_netdev *dst, *last_dst = NULL;
706 int excluded_devices[1+MAX_NEST_DEV];
707 struct hlist_head *head;
708 struct hlist_node *next;
709 int num_excluded = 0;
710 unsigned int i;
711 int err;
712
713 if (exclude_ingress) {
714 num_excluded = get_upper_ifindexes(dev, excluded_devices);
715 excluded_devices[num_excluded++] = dev->ifindex;
716 }
717
718 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
719 for (i = 0; i < map->max_entries; i++) {
720 dst = rcu_dereference_check(dtab->netdev_map[i],
721 rcu_read_lock_bh_held());
722 if (!dst)
723 continue;
724
725 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
726 continue;
727
728 /* we only need n-1 clones; last_dst enqueued below */
729 if (!last_dst) {
730 last_dst = dst;
731 continue;
732 }
733
734 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
735 if (err)
736 return err;
737
738 last_dst = dst;
739
740 }
741 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
742 for (i = 0; i < dtab->n_buckets; i++) {
743 head = dev_map_index_hash(dtab, i);
744 hlist_for_each_entry_safe(dst, next, head, index_hlist) {
745 if (!dst)
746 continue;
747
748 if (is_ifindex_excluded(excluded_devices, num_excluded,
749 dst->dev->ifindex))
750 continue;
751
752 /* we only need n-1 clones; last_dst enqueued below */
753 if (!last_dst) {
754 last_dst = dst;
755 continue;
756 }
757
758 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
759 if (err)
760 return err;
761
762 last_dst = dst;
763 }
764 }
765 }
766
767 /* consume the first skb and return */
768 if (last_dst)
769 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
770
771 /* dtab is empty */
772 consume_skb(skb);
773 return 0;
774}
775
776static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
777{
778 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
779
780 return obj ? &obj->val : NULL;
781}
782
783static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
784{
785 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
786 *(u32 *)key);
787 return obj ? &obj->val : NULL;
788}
789
790static void __dev_map_entry_free(struct rcu_head *rcu)
791{
792 struct bpf_dtab_netdev *dev;
793
794 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
795 if (dev->xdp_prog)
796 bpf_prog_put(dev->xdp_prog);
797 dev_put(dev->dev);
798 kfree(dev);
799}
800
801static int dev_map_delete_elem(struct bpf_map *map, void *key)
802{
803 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
804 struct bpf_dtab_netdev *old_dev;
805 int k = *(u32 *)key;
806
807 if (k >= map->max_entries)
808 return -EINVAL;
809
810 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
811 if (old_dev)
812 call_rcu(&old_dev->rcu, __dev_map_entry_free);
813 return 0;
814}
815
816static int dev_map_hash_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 unsigned long flags;
822 int ret = -ENOENT;
823
824 spin_lock_irqsave(&dtab->index_lock, flags);
825
826 old_dev = __dev_map_hash_lookup_elem(map, k);
827 if (old_dev) {
828 dtab->items--;
829 hlist_del_init_rcu(&old_dev->index_hlist);
830 call_rcu(&old_dev->rcu, __dev_map_entry_free);
831 ret = 0;
832 }
833 spin_unlock_irqrestore(&dtab->index_lock, flags);
834
835 return ret;
836}
837
838static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
839 struct bpf_dtab *dtab,
840 struct bpf_devmap_val *val,
841 unsigned int idx)
842{
843 struct bpf_prog *prog = NULL;
844 struct bpf_dtab_netdev *dev;
845
846 dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
847 GFP_ATOMIC | __GFP_NOWARN,
848 dtab->map.numa_node);
849 if (!dev)
850 return ERR_PTR(-ENOMEM);
851
852 dev->dev = dev_get_by_index(net, val->ifindex);
853 if (!dev->dev)
854 goto err_out;
855
856 if (val->bpf_prog.fd > 0) {
857 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
858 BPF_PROG_TYPE_XDP, false);
859 if (IS_ERR(prog))
860 goto err_put_dev;
861 if (prog->expected_attach_type != BPF_XDP_DEVMAP)
862 goto err_put_prog;
863 }
864
865 dev->idx = idx;
866 dev->dtab = dtab;
867 if (prog) {
868 dev->xdp_prog = prog;
869 dev->val.bpf_prog.id = prog->aux->id;
870 } else {
871 dev->xdp_prog = NULL;
872 dev->val.bpf_prog.id = 0;
873 }
874 dev->val.ifindex = val->ifindex;
875
876 return dev;
877err_put_prog:
878 bpf_prog_put(prog);
879err_put_dev:
880 dev_put(dev->dev);
881err_out:
882 kfree(dev);
883 return ERR_PTR(-EINVAL);
884}
885
886static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
887 void *key, void *value, u64 map_flags)
888{
889 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
890 struct bpf_dtab_netdev *dev, *old_dev;
891 struct bpf_devmap_val val = {};
892 u32 i = *(u32 *)key;
893
894 if (unlikely(map_flags > BPF_EXIST))
895 return -EINVAL;
896 if (unlikely(i >= dtab->map.max_entries))
897 return -E2BIG;
898 if (unlikely(map_flags == BPF_NOEXIST))
899 return -EEXIST;
900
901 /* already verified value_size <= sizeof val */
902 memcpy(&val, value, map->value_size);
903
904 if (!val.ifindex) {
905 dev = NULL;
906 /* can not specify fd if ifindex is 0 */
907 if (val.bpf_prog.fd > 0)
908 return -EINVAL;
909 } else {
910 dev = __dev_map_alloc_node(net, dtab, &val, i);
911 if (IS_ERR(dev))
912 return PTR_ERR(dev);
913 }
914
915 /* Use call_rcu() here to ensure rcu critical sections have completed
916 * Remembering the driver side flush operation will happen before the
917 * net device is removed.
918 */
919 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
920 if (old_dev)
921 call_rcu(&old_dev->rcu, __dev_map_entry_free);
922
923 return 0;
924}
925
926static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
927 u64 map_flags)
928{
929 return __dev_map_update_elem(current->nsproxy->net_ns,
930 map, key, value, map_flags);
931}
932
933static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
934 void *key, void *value, u64 map_flags)
935{
936 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
937 struct bpf_dtab_netdev *dev, *old_dev;
938 struct bpf_devmap_val val = {};
939 u32 idx = *(u32 *)key;
940 unsigned long flags;
941 int err = -EEXIST;
942
943 /* already verified value_size <= sizeof val */
944 memcpy(&val, value, map->value_size);
945
946 if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
947 return -EINVAL;
948
949 spin_lock_irqsave(&dtab->index_lock, flags);
950
951 old_dev = __dev_map_hash_lookup_elem(map, idx);
952 if (old_dev && (map_flags & BPF_NOEXIST))
953 goto out_err;
954
955 dev = __dev_map_alloc_node(net, dtab, &val, idx);
956 if (IS_ERR(dev)) {
957 err = PTR_ERR(dev);
958 goto out_err;
959 }
960
961 if (old_dev) {
962 hlist_del_rcu(&old_dev->index_hlist);
963 } else {
964 if (dtab->items >= dtab->map.max_entries) {
965 spin_unlock_irqrestore(&dtab->index_lock, flags);
966 call_rcu(&dev->rcu, __dev_map_entry_free);
967 return -E2BIG;
968 }
969 dtab->items++;
970 }
971
972 hlist_add_head_rcu(&dev->index_hlist,
973 dev_map_index_hash(dtab, idx));
974 spin_unlock_irqrestore(&dtab->index_lock, flags);
975
976 if (old_dev)
977 call_rcu(&old_dev->rcu, __dev_map_entry_free);
978
979 return 0;
980
981out_err:
982 spin_unlock_irqrestore(&dtab->index_lock, flags);
983 return err;
984}
985
986static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
987 u64 map_flags)
988{
989 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
990 map, key, value, map_flags);
991}
992
993static int dev_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
994{
995 return __bpf_xdp_redirect_map(map, ifindex, flags,
996 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
997 __dev_map_lookup_elem);
998}
999
1000static int dev_hash_map_redirect(struct bpf_map *map, u32 ifindex, u64 flags)
1001{
1002 return __bpf_xdp_redirect_map(map, ifindex, flags,
1003 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1004 __dev_map_hash_lookup_elem);
1005}
1006
1007static int dev_map_btf_id;
1008const struct bpf_map_ops dev_map_ops = {
1009 .map_meta_equal = bpf_map_meta_equal,
1010 .map_alloc = dev_map_alloc,
1011 .map_free = dev_map_free,
1012 .map_get_next_key = dev_map_get_next_key,
1013 .map_lookup_elem = dev_map_lookup_elem,
1014 .map_update_elem = dev_map_update_elem,
1015 .map_delete_elem = dev_map_delete_elem,
1016 .map_check_btf = map_check_no_btf,
1017 .map_btf_name = "bpf_dtab",
1018 .map_btf_id = &dev_map_btf_id,
1019 .map_redirect = dev_map_redirect,
1020};
1021
1022static int dev_map_hash_map_btf_id;
1023const struct bpf_map_ops dev_map_hash_ops = {
1024 .map_meta_equal = bpf_map_meta_equal,
1025 .map_alloc = dev_map_alloc,
1026 .map_free = dev_map_free,
1027 .map_get_next_key = dev_map_hash_get_next_key,
1028 .map_lookup_elem = dev_map_hash_lookup_elem,
1029 .map_update_elem = dev_map_hash_update_elem,
1030 .map_delete_elem = dev_map_hash_delete_elem,
1031 .map_check_btf = map_check_no_btf,
1032 .map_btf_name = "bpf_dtab",
1033 .map_btf_id = &dev_map_hash_map_btf_id,
1034 .map_redirect = dev_hash_map_redirect,
1035};
1036
1037static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1038 struct net_device *netdev)
1039{
1040 unsigned long flags;
1041 u32 i;
1042
1043 spin_lock_irqsave(&dtab->index_lock, flags);
1044 for (i = 0; i < dtab->n_buckets; i++) {
1045 struct bpf_dtab_netdev *dev;
1046 struct hlist_head *head;
1047 struct hlist_node *next;
1048
1049 head = dev_map_index_hash(dtab, i);
1050
1051 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1052 if (netdev != dev->dev)
1053 continue;
1054
1055 dtab->items--;
1056 hlist_del_rcu(&dev->index_hlist);
1057 call_rcu(&dev->rcu, __dev_map_entry_free);
1058 }
1059 }
1060 spin_unlock_irqrestore(&dtab->index_lock, flags);
1061}
1062
1063static int dev_map_notification(struct notifier_block *notifier,
1064 ulong event, void *ptr)
1065{
1066 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1067 struct bpf_dtab *dtab;
1068 int i, cpu;
1069
1070 switch (event) {
1071 case NETDEV_REGISTER:
1072 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1073 break;
1074
1075 /* will be freed in free_netdev() */
1076 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1077 if (!netdev->xdp_bulkq)
1078 return NOTIFY_BAD;
1079
1080 for_each_possible_cpu(cpu)
1081 per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1082 break;
1083 case NETDEV_UNREGISTER:
1084 /* This rcu_read_lock/unlock pair is needed because
1085 * dev_map_list is an RCU list AND to ensure a delete
1086 * operation does not free a netdev_map entry while we
1087 * are comparing it against the netdev being unregistered.
1088 */
1089 rcu_read_lock();
1090 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1091 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1092 dev_map_hash_remove_netdev(dtab, netdev);
1093 continue;
1094 }
1095
1096 for (i = 0; i < dtab->map.max_entries; i++) {
1097 struct bpf_dtab_netdev *dev, *odev;
1098
1099 dev = rcu_dereference(dtab->netdev_map[i]);
1100 if (!dev || netdev != dev->dev)
1101 continue;
1102 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1103 if (dev == odev)
1104 call_rcu(&dev->rcu,
1105 __dev_map_entry_free);
1106 }
1107 }
1108 rcu_read_unlock();
1109 break;
1110 default:
1111 break;
1112 }
1113 return NOTIFY_OK;
1114}
1115
1116static struct notifier_block dev_map_notifier = {
1117 .notifier_call = dev_map_notification,
1118};
1119
1120static int __init dev_map_init(void)
1121{
1122 int cpu;
1123
1124 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1125 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1126 offsetof(struct _bpf_dtab_netdev, dev));
1127 register_netdevice_notifier(&dev_map_notifier);
1128
1129 for_each_possible_cpu(cpu)
1130 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1131 return 0;
1132}
1133
1134subsys_initcall(dev_map_init);