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#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 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 dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
163 if (!dtab)
164 return ERR_PTR(-ENOMEM);
165
166 err = dev_map_init_map(dtab, attr);
167 if (err) {
168 bpf_map_area_free(dtab);
169 return ERR_PTR(err);
170 }
171
172 spin_lock(&dev_map_lock);
173 list_add_tail_rcu(&dtab->list, &dev_map_list);
174 spin_unlock(&dev_map_lock);
175
176 return &dtab->map;
177}
178
179static void dev_map_free(struct bpf_map *map)
180{
181 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
182 int i;
183
184 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
185 * so the programs (can be more than one that used this map) were
186 * disconnected from events. The following synchronize_rcu() guarantees
187 * both rcu read critical sections complete and waits for
188 * preempt-disable regions (NAPI being the relevant context here) so we
189 * are certain there will be no further reads against the netdev_map and
190 * all flush operations are complete. Flush operations can only be done
191 * from NAPI context for this reason.
192 */
193
194 spin_lock(&dev_map_lock);
195 list_del_rcu(&dtab->list);
196 spin_unlock(&dev_map_lock);
197
198 bpf_clear_redirect_map(map);
199 synchronize_rcu();
200
201 /* Make sure prior __dev_map_entry_free() have completed. */
202 rcu_barrier();
203
204 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
205 for (i = 0; i < dtab->n_buckets; i++) {
206 struct bpf_dtab_netdev *dev;
207 struct hlist_head *head;
208 struct hlist_node *next;
209
210 head = dev_map_index_hash(dtab, i);
211
212 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
213 hlist_del_rcu(&dev->index_hlist);
214 if (dev->xdp_prog)
215 bpf_prog_put(dev->xdp_prog);
216 dev_put(dev->dev);
217 kfree(dev);
218 }
219 }
220
221 bpf_map_area_free(dtab->dev_index_head);
222 } else {
223 for (i = 0; i < dtab->map.max_entries; i++) {
224 struct bpf_dtab_netdev *dev;
225
226 dev = rcu_dereference_raw(dtab->netdev_map[i]);
227 if (!dev)
228 continue;
229
230 if (dev->xdp_prog)
231 bpf_prog_put(dev->xdp_prog);
232 dev_put(dev->dev);
233 kfree(dev);
234 }
235
236 bpf_map_area_free(dtab->netdev_map);
237 }
238
239 bpf_map_area_free(dtab);
240}
241
242static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
243{
244 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
245 u32 index = key ? *(u32 *)key : U32_MAX;
246 u32 *next = next_key;
247
248 if (index >= dtab->map.max_entries) {
249 *next = 0;
250 return 0;
251 }
252
253 if (index == dtab->map.max_entries - 1)
254 return -ENOENT;
255 *next = index + 1;
256 return 0;
257}
258
259/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
260 * by local_bh_disable() (from XDP calls inside NAPI). The
261 * rcu_read_lock_bh_held() below makes lockdep accept both.
262 */
263static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
264{
265 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
266 struct hlist_head *head = dev_map_index_hash(dtab, key);
267 struct bpf_dtab_netdev *dev;
268
269 hlist_for_each_entry_rcu(dev, head, index_hlist,
270 lockdep_is_held(&dtab->index_lock))
271 if (dev->idx == key)
272 return dev;
273
274 return NULL;
275}
276
277static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
278 void *next_key)
279{
280 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
281 u32 idx, *next = next_key;
282 struct bpf_dtab_netdev *dev, *next_dev;
283 struct hlist_head *head;
284 int i = 0;
285
286 if (!key)
287 goto find_first;
288
289 idx = *(u32 *)key;
290
291 dev = __dev_map_hash_lookup_elem(map, idx);
292 if (!dev)
293 goto find_first;
294
295 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
296 struct bpf_dtab_netdev, index_hlist);
297
298 if (next_dev) {
299 *next = next_dev->idx;
300 return 0;
301 }
302
303 i = idx & (dtab->n_buckets - 1);
304 i++;
305
306 find_first:
307 for (; i < dtab->n_buckets; i++) {
308 head = dev_map_index_hash(dtab, i);
309
310 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
311 struct bpf_dtab_netdev,
312 index_hlist);
313 if (next_dev) {
314 *next = next_dev->idx;
315 return 0;
316 }
317 }
318
319 return -ENOENT;
320}
321
322static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
323 struct xdp_frame **frames, int n,
324 struct net_device *dev)
325{
326 struct xdp_txq_info txq = { .dev = dev };
327 struct xdp_buff xdp;
328 int i, nframes = 0;
329
330 for (i = 0; i < n; i++) {
331 struct xdp_frame *xdpf = frames[i];
332 u32 act;
333 int err;
334
335 xdp_convert_frame_to_buff(xdpf, &xdp);
336 xdp.txq = &txq;
337
338 act = bpf_prog_run_xdp(xdp_prog, &xdp);
339 switch (act) {
340 case XDP_PASS:
341 err = xdp_update_frame_from_buff(&xdp, xdpf);
342 if (unlikely(err < 0))
343 xdp_return_frame_rx_napi(xdpf);
344 else
345 frames[nframes++] = xdpf;
346 break;
347 default:
348 bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
349 fallthrough;
350 case XDP_ABORTED:
351 trace_xdp_exception(dev, xdp_prog, act);
352 fallthrough;
353 case XDP_DROP:
354 xdp_return_frame_rx_napi(xdpf);
355 break;
356 }
357 }
358 return nframes; /* sent frames count */
359}
360
361static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
362{
363 struct net_device *dev = bq->dev;
364 unsigned int cnt = bq->count;
365 int sent = 0, err = 0;
366 int to_send = cnt;
367 int i;
368
369 if (unlikely(!cnt))
370 return;
371
372 for (i = 0; i < cnt; i++) {
373 struct xdp_frame *xdpf = bq->q[i];
374
375 prefetch(xdpf);
376 }
377
378 if (bq->xdp_prog) {
379 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev);
380 if (!to_send)
381 goto out;
382 }
383
384 sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
385 if (sent < 0) {
386 /* If ndo_xdp_xmit fails with an errno, no frames have
387 * been xmit'ed.
388 */
389 err = sent;
390 sent = 0;
391 }
392
393 /* If not all frames have been transmitted, it is our
394 * responsibility to free them
395 */
396 for (i = sent; unlikely(i < to_send); i++)
397 xdp_return_frame_rx_napi(bq->q[i]);
398
399out:
400 bq->count = 0;
401 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
402}
403
404/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
405 * driver before returning from its napi->poll() routine. See the comment above
406 * xdp_do_flush() in filter.c.
407 */
408void __dev_flush(void)
409{
410 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
411 struct xdp_dev_bulk_queue *bq, *tmp;
412
413 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
414 bq_xmit_all(bq, XDP_XMIT_FLUSH);
415 bq->dev_rx = NULL;
416 bq->xdp_prog = NULL;
417 __list_del_clearprev(&bq->flush_node);
418 }
419}
420
421#ifdef CONFIG_DEBUG_NET
422bool dev_check_flush(void)
423{
424 if (list_empty(this_cpu_ptr(&dev_flush_list)))
425 return false;
426 __dev_flush();
427 return true;
428}
429#endif
430
431/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
432 * by local_bh_disable() (from XDP calls inside NAPI). The
433 * rcu_read_lock_bh_held() below makes lockdep accept both.
434 */
435static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
436{
437 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
438 struct bpf_dtab_netdev *obj;
439
440 if (key >= map->max_entries)
441 return NULL;
442
443 obj = rcu_dereference_check(dtab->netdev_map[key],
444 rcu_read_lock_bh_held());
445 return obj;
446}
447
448/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
449 * variable access, and map elements stick around. See comment above
450 * xdp_do_flush() in filter.c.
451 */
452static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
453 struct net_device *dev_rx, struct bpf_prog *xdp_prog)
454{
455 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
456 struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
457
458 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
459 bq_xmit_all(bq, 0);
460
461 /* Ingress dev_rx will be the same for all xdp_frame's in
462 * bulk_queue, because bq stored per-CPU and must be flushed
463 * from net_device drivers NAPI func end.
464 *
465 * Do the same with xdp_prog and flush_list since these fields
466 * are only ever modified together.
467 */
468 if (!bq->dev_rx) {
469 bq->dev_rx = dev_rx;
470 bq->xdp_prog = xdp_prog;
471 list_add(&bq->flush_node, flush_list);
472 }
473
474 bq->q[bq->count++] = xdpf;
475}
476
477static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
478 struct net_device *dev_rx,
479 struct bpf_prog *xdp_prog)
480{
481 int err;
482
483 if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
484 return -EOPNOTSUPP;
485
486 if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
487 xdp_frame_has_frags(xdpf)))
488 return -EOPNOTSUPP;
489
490 err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
491 if (unlikely(err))
492 return err;
493
494 bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
495 return 0;
496}
497
498static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
499{
500 struct xdp_txq_info txq = { .dev = dst->dev };
501 struct xdp_buff xdp;
502 u32 act;
503
504 if (!dst->xdp_prog)
505 return XDP_PASS;
506
507 __skb_pull(skb, skb->mac_len);
508 xdp.txq = &txq;
509
510 act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
511 switch (act) {
512 case XDP_PASS:
513 __skb_push(skb, skb->mac_len);
514 break;
515 default:
516 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
517 fallthrough;
518 case XDP_ABORTED:
519 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
520 fallthrough;
521 case XDP_DROP:
522 kfree_skb(skb);
523 break;
524 }
525
526 return act;
527}
528
529int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
530 struct net_device *dev_rx)
531{
532 return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
533}
534
535int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
536 struct net_device *dev_rx)
537{
538 struct net_device *dev = dst->dev;
539
540 return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
541}
542
543static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
544{
545 if (!obj)
546 return false;
547
548 if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
549 return false;
550
551 if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
552 xdp_frame_has_frags(xdpf)))
553 return false;
554
555 if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
556 return false;
557
558 return true;
559}
560
561static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
562 struct net_device *dev_rx,
563 struct xdp_frame *xdpf)
564{
565 struct xdp_frame *nxdpf;
566
567 nxdpf = xdpf_clone(xdpf);
568 if (!nxdpf)
569 return -ENOMEM;
570
571 bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
572
573 return 0;
574}
575
576static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
577{
578 while (num_excluded--) {
579 if (ifindex == excluded[num_excluded])
580 return true;
581 }
582 return false;
583}
584
585/* Get ifindex of each upper device. 'indexes' must be able to hold at
586 * least MAX_NEST_DEV elements.
587 * Returns the number of ifindexes added.
588 */
589static int get_upper_ifindexes(struct net_device *dev, int *indexes)
590{
591 struct net_device *upper;
592 struct list_head *iter;
593 int n = 0;
594
595 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
596 indexes[n++] = upper->ifindex;
597 }
598 return n;
599}
600
601int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
602 struct bpf_map *map, bool exclude_ingress)
603{
604 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
605 struct bpf_dtab_netdev *dst, *last_dst = NULL;
606 int excluded_devices[1+MAX_NEST_DEV];
607 struct hlist_head *head;
608 int num_excluded = 0;
609 unsigned int i;
610 int err;
611
612 if (exclude_ingress) {
613 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices);
614 excluded_devices[num_excluded++] = dev_rx->ifindex;
615 }
616
617 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
618 for (i = 0; i < map->max_entries; i++) {
619 dst = rcu_dereference_check(dtab->netdev_map[i],
620 rcu_read_lock_bh_held());
621 if (!is_valid_dst(dst, xdpf))
622 continue;
623
624 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
625 continue;
626
627 /* we only need n-1 clones; last_dst enqueued below */
628 if (!last_dst) {
629 last_dst = dst;
630 continue;
631 }
632
633 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
634 if (err)
635 return err;
636
637 last_dst = dst;
638 }
639 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
640 for (i = 0; i < dtab->n_buckets; i++) {
641 head = dev_map_index_hash(dtab, i);
642 hlist_for_each_entry_rcu(dst, head, index_hlist,
643 lockdep_is_held(&dtab->index_lock)) {
644 if (!is_valid_dst(dst, xdpf))
645 continue;
646
647 if (is_ifindex_excluded(excluded_devices, num_excluded,
648 dst->dev->ifindex))
649 continue;
650
651 /* we only need n-1 clones; last_dst enqueued below */
652 if (!last_dst) {
653 last_dst = dst;
654 continue;
655 }
656
657 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
658 if (err)
659 return err;
660
661 last_dst = dst;
662 }
663 }
664 }
665
666 /* consume the last copy of the frame */
667 if (last_dst)
668 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
669 else
670 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
671
672 return 0;
673}
674
675int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
676 struct bpf_prog *xdp_prog)
677{
678 int err;
679
680 err = xdp_ok_fwd_dev(dst->dev, skb->len);
681 if (unlikely(err))
682 return err;
683
684 /* Redirect has already succeeded semantically at this point, so we just
685 * return 0 even if packet is dropped. Helper below takes care of
686 * freeing skb.
687 */
688 if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
689 return 0;
690
691 skb->dev = dst->dev;
692 generic_xdp_tx(skb, xdp_prog);
693
694 return 0;
695}
696
697static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
698 struct sk_buff *skb,
699 struct bpf_prog *xdp_prog)
700{
701 struct sk_buff *nskb;
702 int err;
703
704 nskb = skb_clone(skb, GFP_ATOMIC);
705 if (!nskb)
706 return -ENOMEM;
707
708 err = dev_map_generic_redirect(dst, nskb, xdp_prog);
709 if (unlikely(err)) {
710 consume_skb(nskb);
711 return err;
712 }
713
714 return 0;
715}
716
717int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
718 struct bpf_prog *xdp_prog, struct bpf_map *map,
719 bool exclude_ingress)
720{
721 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
722 struct bpf_dtab_netdev *dst, *last_dst = NULL;
723 int excluded_devices[1+MAX_NEST_DEV];
724 struct hlist_head *head;
725 struct hlist_node *next;
726 int num_excluded = 0;
727 unsigned int i;
728 int err;
729
730 if (exclude_ingress) {
731 num_excluded = get_upper_ifindexes(dev, excluded_devices);
732 excluded_devices[num_excluded++] = dev->ifindex;
733 }
734
735 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
736 for (i = 0; i < map->max_entries; i++) {
737 dst = rcu_dereference_check(dtab->netdev_map[i],
738 rcu_read_lock_bh_held());
739 if (!dst)
740 continue;
741
742 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
743 continue;
744
745 /* we only need n-1 clones; last_dst enqueued below */
746 if (!last_dst) {
747 last_dst = dst;
748 continue;
749 }
750
751 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
752 if (err)
753 return err;
754
755 last_dst = dst;
756
757 }
758 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
759 for (i = 0; i < dtab->n_buckets; i++) {
760 head = dev_map_index_hash(dtab, i);
761 hlist_for_each_entry_safe(dst, next, head, index_hlist) {
762 if (!dst)
763 continue;
764
765 if (is_ifindex_excluded(excluded_devices, num_excluded,
766 dst->dev->ifindex))
767 continue;
768
769 /* we only need n-1 clones; last_dst enqueued below */
770 if (!last_dst) {
771 last_dst = dst;
772 continue;
773 }
774
775 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
776 if (err)
777 return err;
778
779 last_dst = dst;
780 }
781 }
782 }
783
784 /* consume the first skb and return */
785 if (last_dst)
786 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
787
788 /* dtab is empty */
789 consume_skb(skb);
790 return 0;
791}
792
793static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
794{
795 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
796
797 return obj ? &obj->val : NULL;
798}
799
800static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
801{
802 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
803 *(u32 *)key);
804 return obj ? &obj->val : NULL;
805}
806
807static void __dev_map_entry_free(struct rcu_head *rcu)
808{
809 struct bpf_dtab_netdev *dev;
810
811 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
812 if (dev->xdp_prog)
813 bpf_prog_put(dev->xdp_prog);
814 dev_put(dev->dev);
815 kfree(dev);
816}
817
818static long dev_map_delete_elem(struct bpf_map *map, void *key)
819{
820 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
821 struct bpf_dtab_netdev *old_dev;
822 int k = *(u32 *)key;
823
824 if (k >= map->max_entries)
825 return -EINVAL;
826
827 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
828 if (old_dev) {
829 call_rcu(&old_dev->rcu, __dev_map_entry_free);
830 atomic_dec((atomic_t *)&dtab->items);
831 }
832 return 0;
833}
834
835static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
836{
837 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
838 struct bpf_dtab_netdev *old_dev;
839 int k = *(u32 *)key;
840 unsigned long flags;
841 int ret = -ENOENT;
842
843 spin_lock_irqsave(&dtab->index_lock, flags);
844
845 old_dev = __dev_map_hash_lookup_elem(map, k);
846 if (old_dev) {
847 dtab->items--;
848 hlist_del_init_rcu(&old_dev->index_hlist);
849 call_rcu(&old_dev->rcu, __dev_map_entry_free);
850 ret = 0;
851 }
852 spin_unlock_irqrestore(&dtab->index_lock, flags);
853
854 return ret;
855}
856
857static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
858 struct bpf_dtab *dtab,
859 struct bpf_devmap_val *val,
860 unsigned int idx)
861{
862 struct bpf_prog *prog = NULL;
863 struct bpf_dtab_netdev *dev;
864
865 dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
866 GFP_NOWAIT | __GFP_NOWARN,
867 dtab->map.numa_node);
868 if (!dev)
869 return ERR_PTR(-ENOMEM);
870
871 dev->dev = dev_get_by_index(net, val->ifindex);
872 if (!dev->dev)
873 goto err_out;
874
875 if (val->bpf_prog.fd > 0) {
876 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
877 BPF_PROG_TYPE_XDP, false);
878 if (IS_ERR(prog))
879 goto err_put_dev;
880 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
881 !bpf_prog_map_compatible(&dtab->map, prog))
882 goto err_put_prog;
883 }
884
885 dev->idx = idx;
886 if (prog) {
887 dev->xdp_prog = prog;
888 dev->val.bpf_prog.id = prog->aux->id;
889 } else {
890 dev->xdp_prog = NULL;
891 dev->val.bpf_prog.id = 0;
892 }
893 dev->val.ifindex = val->ifindex;
894
895 return dev;
896err_put_prog:
897 bpf_prog_put(prog);
898err_put_dev:
899 dev_put(dev->dev);
900err_out:
901 kfree(dev);
902 return ERR_PTR(-EINVAL);
903}
904
905static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
906 void *key, void *value, u64 map_flags)
907{
908 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
909 struct bpf_dtab_netdev *dev, *old_dev;
910 struct bpf_devmap_val val = {};
911 u32 i = *(u32 *)key;
912
913 if (unlikely(map_flags > BPF_EXIST))
914 return -EINVAL;
915 if (unlikely(i >= dtab->map.max_entries))
916 return -E2BIG;
917 if (unlikely(map_flags == BPF_NOEXIST))
918 return -EEXIST;
919
920 /* already verified value_size <= sizeof val */
921 memcpy(&val, value, map->value_size);
922
923 if (!val.ifindex) {
924 dev = NULL;
925 /* can not specify fd if ifindex is 0 */
926 if (val.bpf_prog.fd > 0)
927 return -EINVAL;
928 } else {
929 dev = __dev_map_alloc_node(net, dtab, &val, i);
930 if (IS_ERR(dev))
931 return PTR_ERR(dev);
932 }
933
934 /* Use call_rcu() here to ensure rcu critical sections have completed
935 * Remembering the driver side flush operation will happen before the
936 * net device is removed.
937 */
938 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
939 if (old_dev)
940 call_rcu(&old_dev->rcu, __dev_map_entry_free);
941 else
942 atomic_inc((atomic_t *)&dtab->items);
943
944 return 0;
945}
946
947static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
948 u64 map_flags)
949{
950 return __dev_map_update_elem(current->nsproxy->net_ns,
951 map, key, value, map_flags);
952}
953
954static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
955 void *key, void *value, u64 map_flags)
956{
957 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
958 struct bpf_dtab_netdev *dev, *old_dev;
959 struct bpf_devmap_val val = {};
960 u32 idx = *(u32 *)key;
961 unsigned long flags;
962 int err = -EEXIST;
963
964 /* already verified value_size <= sizeof val */
965 memcpy(&val, value, map->value_size);
966
967 if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
968 return -EINVAL;
969
970 spin_lock_irqsave(&dtab->index_lock, flags);
971
972 old_dev = __dev_map_hash_lookup_elem(map, idx);
973 if (old_dev && (map_flags & BPF_NOEXIST))
974 goto out_err;
975
976 dev = __dev_map_alloc_node(net, dtab, &val, idx);
977 if (IS_ERR(dev)) {
978 err = PTR_ERR(dev);
979 goto out_err;
980 }
981
982 if (old_dev) {
983 hlist_del_rcu(&old_dev->index_hlist);
984 } else {
985 if (dtab->items >= dtab->map.max_entries) {
986 spin_unlock_irqrestore(&dtab->index_lock, flags);
987 call_rcu(&dev->rcu, __dev_map_entry_free);
988 return -E2BIG;
989 }
990 dtab->items++;
991 }
992
993 hlist_add_head_rcu(&dev->index_hlist,
994 dev_map_index_hash(dtab, idx));
995 spin_unlock_irqrestore(&dtab->index_lock, flags);
996
997 if (old_dev)
998 call_rcu(&old_dev->rcu, __dev_map_entry_free);
999
1000 return 0;
1001
1002out_err:
1003 spin_unlock_irqrestore(&dtab->index_lock, flags);
1004 return err;
1005}
1006
1007static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1008 u64 map_flags)
1009{
1010 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1011 map, key, value, map_flags);
1012}
1013
1014static long dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1015{
1016 return __bpf_xdp_redirect_map(map, ifindex, flags,
1017 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1018 __dev_map_lookup_elem);
1019}
1020
1021static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1022{
1023 return __bpf_xdp_redirect_map(map, ifindex, flags,
1024 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1025 __dev_map_hash_lookup_elem);
1026}
1027
1028static u64 dev_map_mem_usage(const struct bpf_map *map)
1029{
1030 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1031 u64 usage = sizeof(struct bpf_dtab);
1032
1033 if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1034 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1035 else
1036 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1037 usage += atomic_read((atomic_t *)&dtab->items) *
1038 (u64)sizeof(struct bpf_dtab_netdev);
1039 return usage;
1040}
1041
1042BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1043const struct bpf_map_ops dev_map_ops = {
1044 .map_meta_equal = bpf_map_meta_equal,
1045 .map_alloc = dev_map_alloc,
1046 .map_free = dev_map_free,
1047 .map_get_next_key = dev_map_get_next_key,
1048 .map_lookup_elem = dev_map_lookup_elem,
1049 .map_update_elem = dev_map_update_elem,
1050 .map_delete_elem = dev_map_delete_elem,
1051 .map_check_btf = map_check_no_btf,
1052 .map_mem_usage = dev_map_mem_usage,
1053 .map_btf_id = &dev_map_btf_ids[0],
1054 .map_redirect = dev_map_redirect,
1055};
1056
1057const struct bpf_map_ops dev_map_hash_ops = {
1058 .map_meta_equal = bpf_map_meta_equal,
1059 .map_alloc = dev_map_alloc,
1060 .map_free = dev_map_free,
1061 .map_get_next_key = dev_map_hash_get_next_key,
1062 .map_lookup_elem = dev_map_hash_lookup_elem,
1063 .map_update_elem = dev_map_hash_update_elem,
1064 .map_delete_elem = dev_map_hash_delete_elem,
1065 .map_check_btf = map_check_no_btf,
1066 .map_mem_usage = dev_map_mem_usage,
1067 .map_btf_id = &dev_map_btf_ids[0],
1068 .map_redirect = dev_hash_map_redirect,
1069};
1070
1071static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1072 struct net_device *netdev)
1073{
1074 unsigned long flags;
1075 u32 i;
1076
1077 spin_lock_irqsave(&dtab->index_lock, flags);
1078 for (i = 0; i < dtab->n_buckets; i++) {
1079 struct bpf_dtab_netdev *dev;
1080 struct hlist_head *head;
1081 struct hlist_node *next;
1082
1083 head = dev_map_index_hash(dtab, i);
1084
1085 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1086 if (netdev != dev->dev)
1087 continue;
1088
1089 dtab->items--;
1090 hlist_del_rcu(&dev->index_hlist);
1091 call_rcu(&dev->rcu, __dev_map_entry_free);
1092 }
1093 }
1094 spin_unlock_irqrestore(&dtab->index_lock, flags);
1095}
1096
1097static int dev_map_notification(struct notifier_block *notifier,
1098 ulong event, void *ptr)
1099{
1100 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1101 struct bpf_dtab *dtab;
1102 int i, cpu;
1103
1104 switch (event) {
1105 case NETDEV_REGISTER:
1106 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1107 break;
1108
1109 /* will be freed in free_netdev() */
1110 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1111 if (!netdev->xdp_bulkq)
1112 return NOTIFY_BAD;
1113
1114 for_each_possible_cpu(cpu)
1115 per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
1116 break;
1117 case NETDEV_UNREGISTER:
1118 /* This rcu_read_lock/unlock pair is needed because
1119 * dev_map_list is an RCU list AND to ensure a delete
1120 * operation does not free a netdev_map entry while we
1121 * are comparing it against the netdev being unregistered.
1122 */
1123 rcu_read_lock();
1124 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1125 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1126 dev_map_hash_remove_netdev(dtab, netdev);
1127 continue;
1128 }
1129
1130 for (i = 0; i < dtab->map.max_entries; i++) {
1131 struct bpf_dtab_netdev *dev, *odev;
1132
1133 dev = rcu_dereference(dtab->netdev_map[i]);
1134 if (!dev || netdev != dev->dev)
1135 continue;
1136 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1137 if (dev == odev) {
1138 call_rcu(&dev->rcu,
1139 __dev_map_entry_free);
1140 atomic_dec((atomic_t *)&dtab->items);
1141 }
1142 }
1143 }
1144 rcu_read_unlock();
1145 break;
1146 default:
1147 break;
1148 }
1149 return NOTIFY_OK;
1150}
1151
1152static struct notifier_block dev_map_notifier = {
1153 .notifier_call = dev_map_notification,
1154};
1155
1156static int __init dev_map_init(void)
1157{
1158 int cpu;
1159
1160 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1161 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1162 offsetof(struct _bpf_dtab_netdev, dev));
1163 register_netdevice_notifier(&dev_map_notifier);
1164
1165 for_each_possible_cpu(cpu)
1166 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
1167 return 0;
1168}
1169
1170subsys_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
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);