projects/linux·blob v7.2-rc5 · refs

1217 lines · 31.6 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 <linux/local_lock.h>
49#include <net/xdp.h>
50#include <linux/filter.h>
51#include <trace/events/xdp.h>
52#include <linux/btf_ids.h>
53
54#define DEV_CREATE_FLAG_MASK \
55 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
56
57struct xdp_dev_bulk_queue {
58 struct xdp_frame *q[DEV_MAP_BULK_SIZE];
59 struct list_head flush_node;
60 struct net_device *dev;
61 struct net_device *dev_rx;
62 struct bpf_prog *xdp_prog;
63 unsigned int count;
64 local_lock_t bq_lock;
65};
66
67struct bpf_dtab_netdev {
68 struct net_device *dev; /* must be first member, due to tracepoint */
69 struct hlist_node index_hlist;
70 struct bpf_prog *xdp_prog;
71 struct rcu_head rcu;
72 unsigned int idx;
73 struct bpf_devmap_val val;
74};
75
76struct bpf_dtab {
77 struct bpf_map map;
78 struct bpf_dtab_netdev __rcu **netdev_map; /* DEVMAP type only */
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 int numa_node)
93{
94 int i;
95 struct hlist_head *hash;
96
97 hash = bpf_map_area_alloc((u64) entries * sizeof(*hash), numa_node);
98 if (hash != NULL)
99 for (i = 0; i < entries; i++)
100 INIT_HLIST_HEAD(&hash[i]);
101
102 return hash;
103}
104
105static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
106 int idx)
107{
108 return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
109}
110
111static int dev_map_alloc_check(union bpf_attr *attr)
112{
113 u32 valsize = attr->value_size;
114
115 /* check sanity of attributes. 2 value sizes supported:
116 * 4 bytes: ifindex
117 * 8 bytes: ifindex + prog fd
118 */
119 if (attr->max_entries == 0 || attr->key_size != 4 ||
120 (valsize != offsetofend(struct bpf_devmap_val, ifindex) &&
121 valsize != offsetofend(struct bpf_devmap_val, bpf_prog.fd)) ||
122 attr->map_flags & ~DEV_CREATE_FLAG_MASK)
123 return -EINVAL;
124
125 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
126 /* Hash table size must be power of 2; roundup_pow_of_two()
127 * can overflow into UB on 32-bit arches
128 */
129 if (attr->max_entries > 1UL << 31)
130 return -EINVAL;
131 }
132
133 return 0;
134}
135
136static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
137{
138 /* Lookup returns a pointer straight to dev->ifindex, so make sure the
139 * verifier prevents writes from the BPF side
140 */
141 attr->map_flags |= BPF_F_RDONLY_PROG;
142 bpf_map_init_from_attr(&dtab->map, attr);
143
144 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
145 /* Hash table size must be power of 2 */
146 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
147 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets,
148 dtab->map.numa_node);
149 if (!dtab->dev_index_head)
150 return -ENOMEM;
151
152 spin_lock_init(&dtab->index_lock);
153 } else {
154 dtab->netdev_map = bpf_map_area_alloc((u64) dtab->map.max_entries *
155 sizeof(struct bpf_dtab_netdev *),
156 dtab->map.numa_node);
157 if (!dtab->netdev_map)
158 return -ENOMEM;
159 }
160
161 return 0;
162}
163
164static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
165{
166 struct bpf_dtab *dtab;
167 int err;
168
169 dtab = bpf_map_area_alloc(sizeof(*dtab), NUMA_NO_NODE);
170 if (!dtab)
171 return ERR_PTR(-ENOMEM);
172
173 err = dev_map_init_map(dtab, attr);
174 if (err) {
175 bpf_map_area_free(dtab);
176 return ERR_PTR(err);
177 }
178
179 spin_lock(&dev_map_lock);
180 list_add_tail_rcu(&dtab->list, &dev_map_list);
181 spin_unlock(&dev_map_lock);
182
183 return &dtab->map;
184}
185
186static void dev_map_free(struct bpf_map *map)
187{
188 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
189 u32 i;
190
191 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
192 * so the programs (can be more than one that used this map) were
193 * disconnected from events. The following synchronize_rcu() guarantees
194 * both rcu read critical sections complete and waits for
195 * preempt-disable regions (NAPI being the relevant context here) so we
196 * are certain there will be no further reads against the netdev_map and
197 * all flush operations are complete. Flush operations can only be done
198 * from NAPI context for this reason.
199 */
200
201 spin_lock(&dev_map_lock);
202 list_del_rcu(&dtab->list);
203 spin_unlock(&dev_map_lock);
204
205 /* bpf_redirect_info->map is assigned in __bpf_xdp_redirect_map()
206 * during NAPI callback and cleared after the XDP redirect. There is no
207 * explicit RCU read section which protects bpf_redirect_info->map but
208 * local_bh_disable() also marks the beginning an RCU section. This
209 * makes the complete softirq callback RCU protected. Thus after
210 * following synchronize_rcu() there no bpf_redirect_info->map == map
211 * assignment.
212 */
213 synchronize_rcu();
214
215 /* Make sure prior __dev_map_entry_free() have completed. */
216 rcu_barrier();
217
218 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
219 for (i = 0; i < dtab->n_buckets; i++) {
220 struct bpf_dtab_netdev *dev;
221 struct hlist_head *head;
222 struct hlist_node *next;
223
224 head = dev_map_index_hash(dtab, i);
225
226 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
227 hlist_del_rcu(&dev->index_hlist);
228 if (dev->xdp_prog)
229 bpf_prog_put(dev->xdp_prog);
230 dev_put(dev->dev);
231 kfree(dev);
232 }
233 }
234
235 bpf_map_area_free(dtab->dev_index_head);
236 } else {
237 for (i = 0; i < dtab->map.max_entries; i++) {
238 struct bpf_dtab_netdev *dev;
239
240 dev = rcu_dereference_raw(dtab->netdev_map[i]);
241 if (!dev)
242 continue;
243
244 if (dev->xdp_prog)
245 bpf_prog_put(dev->xdp_prog);
246 dev_put(dev->dev);
247 kfree(dev);
248 }
249
250 bpf_map_area_free(dtab->netdev_map);
251 }
252
253 bpf_map_area_free(dtab);
254}
255
256static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
257{
258 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
259 u32 index = key ? *(u32 *)key : U32_MAX;
260 u32 *next = next_key;
261
262 if (index >= dtab->map.max_entries) {
263 *next = 0;
264 return 0;
265 }
266
267 if (index == dtab->map.max_entries - 1)
268 return -ENOENT;
269 *next = index + 1;
270 return 0;
271}
272
273/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
274 * by local_bh_disable() (from XDP calls inside NAPI). The
275 * rcu_read_lock_bh_held() below makes lockdep accept both.
276 */
277static void *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
278{
279 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
280 struct hlist_head *head = dev_map_index_hash(dtab, key);
281 struct bpf_dtab_netdev *dev;
282
283 hlist_for_each_entry_rcu(dev, head, index_hlist,
284 lockdep_is_held(&dtab->index_lock))
285 if (dev->idx == key)
286 return dev;
287
288 return NULL;
289}
290
291static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
292 void *next_key)
293{
294 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
295 u32 idx, *next = next_key;
296 struct bpf_dtab_netdev *dev, *next_dev;
297 struct hlist_head *head;
298 int i = 0;
299
300 if (!key)
301 goto find_first;
302
303 idx = *(u32 *)key;
304
305 dev = __dev_map_hash_lookup_elem(map, idx);
306 if (!dev)
307 goto find_first;
308
309 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
310 struct bpf_dtab_netdev, index_hlist);
311
312 if (next_dev) {
313 *next = next_dev->idx;
314 return 0;
315 }
316
317 i = idx & (dtab->n_buckets - 1);
318 i++;
319
320 find_first:
321 for (; i < dtab->n_buckets; i++) {
322 head = dev_map_index_hash(dtab, i);
323
324 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
325 struct bpf_dtab_netdev,
326 index_hlist);
327 if (next_dev) {
328 *next = next_dev->idx;
329 return 0;
330 }
331 }
332
333 return -ENOENT;
334}
335
336static int dev_map_bpf_prog_run(struct bpf_prog *xdp_prog,
337 struct xdp_frame **frames, int n,
338 struct net_device *tx_dev,
339 struct net_device *rx_dev)
340{
341 struct xdp_txq_info txq = { .dev = tx_dev };
342 struct xdp_rxq_info rxq = { .dev = rx_dev };
343 struct xdp_buff xdp;
344 int i, nframes = 0;
345
346 for (i = 0; i < n; i++) {
347 struct xdp_frame *xdpf = frames[i];
348 u32 act;
349 int err;
350
351 xdp_convert_frame_to_buff(xdpf, &xdp);
352 xdp.txq = &txq;
353 xdp.rxq = &rxq;
354
355 act = bpf_prog_run_xdp(xdp_prog, &xdp);
356 switch (act) {
357 case XDP_PASS:
358 err = xdp_update_frame_from_buff(&xdp, xdpf);
359 if (unlikely(err < 0))
360 xdp_return_frame_rx_napi(xdpf);
361 else
362 frames[nframes++] = xdpf;
363 break;
364 default:
365 bpf_warn_invalid_xdp_action(NULL, xdp_prog, act);
366 fallthrough;
367 case XDP_ABORTED:
368 trace_xdp_exception(tx_dev, xdp_prog, act);
369 fallthrough;
370 case XDP_DROP:
371 xdp_return_frame_rx_napi(xdpf);
372 break;
373 }
374 }
375 return nframes; /* sent frames count */
376}
377
378static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
379{
380 struct net_device *dev = bq->dev;
381 unsigned int cnt = bq->count;
382 int sent = 0, err = 0;
383 int to_send = cnt;
384 int i;
385
386 lockdep_assert_held(&bq->bq_lock);
387
388 if (unlikely(!cnt))
389 return;
390
391 for (i = 0; i < cnt; i++) {
392 struct xdp_frame *xdpf = bq->q[i];
393
394 prefetch(xdpf);
395 }
396
397 if (bq->xdp_prog) {
398 to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q, cnt, dev, bq->dev_rx);
399 if (!to_send)
400 goto out;
401 }
402
403 sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
404 if (sent < 0) {
405 /* If ndo_xdp_xmit fails with an errno, no frames have
406 * been xmit'ed.
407 */
408 err = sent;
409 sent = 0;
410 }
411
412 /* If not all frames have been transmitted, it is our
413 * responsibility to free them
414 */
415 for (i = sent; unlikely(i < to_send); i++)
416 xdp_return_frame_rx_napi(bq->q[i]);
417
418out:
419 bq->count = 0;
420 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, cnt - sent, err);
421}
422
423/* __dev_flush is called from xdp_do_flush() which _must_ be signalled from the
424 * driver before returning from its napi->poll() routine. See the comment above
425 * xdp_do_flush() in filter.c.
426 */
427void __dev_flush(struct list_head *flush_list)
428{
429 struct xdp_dev_bulk_queue *bq, *tmp;
430
431 list_for_each_entry_safe(bq, tmp, flush_list, flush_node) {
432 local_lock_nested_bh(&bq->dev->xdp_bulkq->bq_lock);
433 bq_xmit_all(bq, XDP_XMIT_FLUSH);
434 bq->dev_rx = NULL;
435 bq->xdp_prog = NULL;
436 __list_del_clearprev(&bq->flush_node);
437 local_unlock_nested_bh(&bq->dev->xdp_bulkq->bq_lock);
438 }
439}
440
441/* Elements are kept alive by RCU; either by rcu_read_lock() (from syscall) or
442 * by local_bh_disable() (from XDP calls inside NAPI). The
443 * rcu_read_lock_bh_held() below makes lockdep accept both.
444 */
445static void *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
446{
447 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
448 struct bpf_dtab_netdev *obj;
449
450 if (key >= map->max_entries)
451 return NULL;
452
453 obj = rcu_dereference_check(dtab->netdev_map[key],
454 rcu_read_lock_bh_held());
455 return obj;
456}
457
458/* Runs in NAPI, i.e., softirq under local_bh_disable(). Thus, safe percpu
459 * variable access, and map elements stick around. See comment above
460 * xdp_do_flush() in filter.c. PREEMPT_RT relies on local_lock_nested_bh()
461 * to serialise access to the per-CPU bq.
462 */
463static void bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
464 struct net_device *dev_rx, struct bpf_prog *xdp_prog)
465{
466 struct xdp_dev_bulk_queue *bq;
467
468 local_lock_nested_bh(&dev->xdp_bulkq->bq_lock);
469 bq = this_cpu_ptr(dev->xdp_bulkq);
470
471 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
472 bq_xmit_all(bq, 0);
473
474 /* Ingress dev_rx will be the same for all xdp_frame's in
475 * bulk_queue, because bq stored per-CPU and must be flushed
476 * from net_device drivers NAPI func end.
477 *
478 * Do the same with xdp_prog and flush_list since these fields
479 * are only ever modified together.
480 */
481 if (!bq->dev_rx) {
482 struct list_head *flush_list = bpf_net_ctx_get_dev_flush_list();
483
484 bq->dev_rx = dev_rx;
485 bq->xdp_prog = xdp_prog;
486 list_add(&bq->flush_node, flush_list);
487 }
488
489 bq->q[bq->count++] = xdpf;
490
491 local_unlock_nested_bh(&dev->xdp_bulkq->bq_lock);
492}
493
494static inline int __xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
495 struct net_device *dev_rx,
496 struct bpf_prog *xdp_prog)
497{
498 int err;
499
500 if (!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
501 return -EOPNOTSUPP;
502
503 if (unlikely(!(dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
504 xdp_frame_has_frags(xdpf)))
505 return -EOPNOTSUPP;
506
507 err = xdp_ok_fwd_dev(dev, xdp_get_frame_len(xdpf));
508 if (unlikely(err))
509 return err;
510
511 bq_enqueue(dev, xdpf, dev_rx, xdp_prog);
512 return 0;
513}
514
515static u32 dev_map_bpf_prog_run_skb(struct sk_buff *skb, struct bpf_dtab_netdev *dst)
516{
517 struct xdp_txq_info txq = { .dev = dst->dev };
518 struct xdp_buff xdp;
519 u32 act;
520
521 if (!dst->xdp_prog)
522 return XDP_PASS;
523
524 __skb_pull(skb, skb->mac_len);
525 xdp.txq = &txq;
526
527 act = bpf_prog_run_generic_xdp(skb, &xdp, dst->xdp_prog);
528 switch (act) {
529 case XDP_PASS:
530 __skb_push(skb, skb->mac_len);
531 break;
532 default:
533 bpf_warn_invalid_xdp_action(NULL, dst->xdp_prog, act);
534 fallthrough;
535 case XDP_ABORTED:
536 trace_xdp_exception(dst->dev, dst->xdp_prog, act);
537 fallthrough;
538 case XDP_DROP:
539 kfree_skb(skb);
540 break;
541 }
542
543 return act;
544}
545
546int dev_xdp_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
547 struct net_device *dev_rx)
548{
549 return __xdp_enqueue(dev, xdpf, dev_rx, NULL);
550}
551
552int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_frame *xdpf,
553 struct net_device *dev_rx)
554{
555 struct net_device *dev = dst->dev;
556
557 return __xdp_enqueue(dev, xdpf, dev_rx, dst->xdp_prog);
558}
559
560static bool is_valid_dst(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf)
561{
562 if (!obj)
563 return false;
564
565 if (!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT))
566 return false;
567
568 if (unlikely(!(obj->dev->xdp_features & NETDEV_XDP_ACT_NDO_XMIT_SG) &&
569 xdp_frame_has_frags(xdpf)))
570 return false;
571
572 if (xdp_ok_fwd_dev(obj->dev, xdp_get_frame_len(xdpf)))
573 return false;
574
575 return true;
576}
577
578static int dev_map_enqueue_clone(struct bpf_dtab_netdev *obj,
579 struct net_device *dev_rx,
580 struct xdp_frame *xdpf)
581{
582 struct xdp_frame *nxdpf;
583
584 /* Frags live outside the linear frame and cannot be cloned safely. */
585 if (unlikely(xdp_frame_has_frags(xdpf)))
586 return -EOPNOTSUPP;
587
588 nxdpf = xdpf_clone(xdpf);
589 if (!nxdpf)
590 return -ENOMEM;
591
592 bq_enqueue(obj->dev, nxdpf, dev_rx, obj->xdp_prog);
593
594 return 0;
595}
596
597static inline bool is_ifindex_excluded(int *excluded, int num_excluded, int ifindex)
598{
599 while (num_excluded--) {
600 if (ifindex == excluded[num_excluded])
601 return true;
602 }
603 return false;
604}
605
606/* Get ifindex of each upper device. 'indexes' must be able to hold at
607 * least 'max' elements.
608 * Returns the number of ifindexes added, or -EOVERFLOW if there are too
609 * many upper devices.
610 */
611static int get_upper_ifindexes(struct net_device *dev, int *indexes, int max)
612{
613 struct net_device *upper;
614 struct list_head *iter;
615 int n = 0;
616
617 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
618 if (n >= max)
619 return -EOVERFLOW;
620 indexes[n++] = upper->ifindex;
621 }
622
623 return n;
624}
625
626int dev_map_enqueue_multi(struct xdp_frame *xdpf, struct net_device *dev_rx,
627 struct bpf_map *map, bool exclude_ingress)
628{
629 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
630 struct bpf_dtab_netdev *dst, *last_dst = NULL;
631 int excluded_devices[1+MAX_NEST_DEV];
632 struct hlist_head *head;
633 int num_excluded = 0;
634 unsigned int i;
635 int err;
636
637 if (exclude_ingress) {
638 num_excluded = get_upper_ifindexes(dev_rx, excluded_devices,
639 ARRAY_SIZE(excluded_devices) - 1);
640 if (num_excluded < 0)
641 return num_excluded;
642
643 excluded_devices[num_excluded++] = dev_rx->ifindex;
644 }
645
646 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
647 for (i = 0; i < map->max_entries; i++) {
648 dst = rcu_dereference_check(dtab->netdev_map[i],
649 rcu_read_lock_bh_held());
650 if (!is_valid_dst(dst, xdpf))
651 continue;
652
653 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
654 continue;
655
656 /* we only need n-1 clones; last_dst enqueued below */
657 if (!last_dst) {
658 last_dst = dst;
659 continue;
660 }
661
662 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
663 if (err)
664 return err;
665
666 last_dst = dst;
667 }
668 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
669 for (i = 0; i < dtab->n_buckets; i++) {
670 head = dev_map_index_hash(dtab, i);
671 hlist_for_each_entry_rcu(dst, head, index_hlist,
672 rcu_read_lock_bh_held()) {
673 if (!is_valid_dst(dst, xdpf))
674 continue;
675
676 if (is_ifindex_excluded(excluded_devices, num_excluded,
677 dst->dev->ifindex))
678 continue;
679
680 /* we only need n-1 clones; last_dst enqueued below */
681 if (!last_dst) {
682 last_dst = dst;
683 continue;
684 }
685
686 err = dev_map_enqueue_clone(last_dst, dev_rx, xdpf);
687 if (err)
688 return err;
689
690 last_dst = dst;
691 }
692 }
693 }
694
695 /* consume the last copy of the frame */
696 if (last_dst)
697 bq_enqueue(last_dst->dev, xdpf, dev_rx, last_dst->xdp_prog);
698 else
699 xdp_return_frame_rx_napi(xdpf); /* dtab is empty */
700
701 return 0;
702}
703
704int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
705 const struct bpf_prog *xdp_prog)
706{
707 int err;
708
709 err = xdp_ok_fwd_dev(dst->dev, skb->len);
710 if (unlikely(err))
711 return err;
712
713 if (dst->xdp_prog && skb_cloned(skb)) {
714 struct sk_buff *nskb;
715
716 nskb = skb_copy(skb, GFP_ATOMIC);
717 if (!nskb)
718 return -ENOMEM;
719
720 nskb->mac_len = skb->mac_len;
721 consume_skb(skb);
722 skb = nskb;
723 }
724
725 /* Redirect has already succeeded semantically at this point, so we just
726 * return 0 even if packet is dropped. Helper below takes care of
727 * freeing skb.
728 */
729 if (dev_map_bpf_prog_run_skb(skb, dst) != XDP_PASS)
730 return 0;
731
732 skb->dev = dst->dev;
733 generic_xdp_tx(skb, xdp_prog);
734
735 return 0;
736}
737
738static int dev_map_redirect_clone(struct bpf_dtab_netdev *dst,
739 struct sk_buff *skb,
740 const struct bpf_prog *xdp_prog)
741{
742 struct sk_buff *nskb;
743 int err;
744
745 if (unlikely(skb_is_nonlinear(skb)))
746 return -EOPNOTSUPP;
747
748 nskb = skb_clone(skb, GFP_ATOMIC);
749 if (!nskb)
750 return -ENOMEM;
751
752 err = dev_map_generic_redirect(dst, nskb, xdp_prog);
753 if (unlikely(err)) {
754 consume_skb(nskb);
755 return err;
756 }
757
758 return 0;
759}
760
761int dev_map_redirect_multi(struct net_device *dev, struct sk_buff *skb,
762 const struct bpf_prog *xdp_prog,
763 struct bpf_map *map, bool exclude_ingress)
764{
765 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
766 struct bpf_dtab_netdev *dst, *last_dst = NULL;
767 int excluded_devices[1+MAX_NEST_DEV];
768 struct hlist_head *head;
769 int num_excluded = 0;
770 unsigned int i;
771 int err;
772
773 if (exclude_ingress) {
774 num_excluded = get_upper_ifindexes(dev, excluded_devices,
775 ARRAY_SIZE(excluded_devices) - 1);
776 if (num_excluded < 0)
777 return num_excluded;
778
779 excluded_devices[num_excluded++] = dev->ifindex;
780 }
781
782 if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
783 for (i = 0; i < map->max_entries; i++) {
784 dst = rcu_dereference_check(dtab->netdev_map[i],
785 rcu_read_lock_bh_held());
786 if (!dst)
787 continue;
788
789 if (is_ifindex_excluded(excluded_devices, num_excluded, dst->dev->ifindex))
790 continue;
791
792 /* we only need n-1 clones; last_dst enqueued below */
793 if (!last_dst) {
794 last_dst = dst;
795 continue;
796 }
797
798 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
799 if (err)
800 return err;
801
802 last_dst = dst;
803
804 }
805 } else { /* BPF_MAP_TYPE_DEVMAP_HASH */
806 for (i = 0; i < dtab->n_buckets; i++) {
807 head = dev_map_index_hash(dtab, i);
808 hlist_for_each_entry_rcu(dst, head, index_hlist, rcu_read_lock_bh_held()) {
809 if (is_ifindex_excluded(excluded_devices, num_excluded,
810 dst->dev->ifindex))
811 continue;
812
813 /* we only need n-1 clones; last_dst enqueued below */
814 if (!last_dst) {
815 last_dst = dst;
816 continue;
817 }
818
819 err = dev_map_redirect_clone(last_dst, skb, xdp_prog);
820 if (err)
821 return err;
822
823 last_dst = dst;
824 }
825 }
826 }
827
828 /* consume the first skb and return */
829 if (last_dst)
830 return dev_map_generic_redirect(last_dst, skb, xdp_prog);
831
832 /* dtab is empty */
833 consume_skb(skb);
834 return 0;
835}
836
837static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
838{
839 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
840
841 return obj ? &obj->val : NULL;
842}
843
844static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
845{
846 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
847 *(u32 *)key);
848 return obj ? &obj->val : NULL;
849}
850
851static void __dev_map_entry_free(struct rcu_head *rcu)
852{
853 struct bpf_dtab_netdev *dev;
854
855 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
856 if (dev->xdp_prog)
857 bpf_prog_put(dev->xdp_prog);
858 dev_put(dev->dev);
859 kfree(dev);
860}
861
862static long dev_map_delete_elem(struct bpf_map *map, void *key)
863{
864 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
865 struct bpf_dtab_netdev *old_dev;
866 u32 k = *(u32 *)key;
867
868 if (k >= map->max_entries)
869 return -EINVAL;
870
871 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[k], NULL));
872 if (old_dev) {
873 call_rcu(&old_dev->rcu, __dev_map_entry_free);
874 atomic_dec((atomic_t *)&dtab->items);
875 }
876 return 0;
877}
878
879static long dev_map_hash_delete_elem(struct bpf_map *map, void *key)
880{
881 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
882 struct bpf_dtab_netdev *old_dev;
883 u32 k = *(u32 *)key;
884 unsigned long flags;
885 int ret = -ENOENT;
886
887 spin_lock_irqsave(&dtab->index_lock, flags);
888
889 old_dev = __dev_map_hash_lookup_elem(map, k);
890 if (old_dev) {
891 dtab->items--;
892 hlist_del_init_rcu(&old_dev->index_hlist);
893 call_rcu(&old_dev->rcu, __dev_map_entry_free);
894 ret = 0;
895 }
896 spin_unlock_irqrestore(&dtab->index_lock, flags);
897
898 return ret;
899}
900
901static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
902 struct bpf_dtab *dtab,
903 struct bpf_devmap_val *val,
904 unsigned int idx)
905{
906 struct bpf_prog *prog = NULL;
907 struct bpf_dtab_netdev *dev;
908
909 dev = bpf_map_kmalloc_node(&dtab->map, sizeof(*dev),
910 GFP_NOWAIT,
911 dtab->map.numa_node);
912 if (!dev)
913 return ERR_PTR(-ENOMEM);
914
915 dev->dev = dev_get_by_index(net, val->ifindex);
916 if (!dev->dev)
917 goto err_out;
918
919 if (val->bpf_prog.fd > 0) {
920 prog = bpf_prog_get_type_dev(val->bpf_prog.fd,
921 BPF_PROG_TYPE_XDP, false);
922 if (IS_ERR(prog))
923 goto err_put_dev;
924 if (prog->expected_attach_type != BPF_XDP_DEVMAP ||
925 !bpf_prog_map_compatible(&dtab->map, prog))
926 goto err_put_prog;
927 }
928
929 dev->idx = idx;
930 if (prog) {
931 dev->xdp_prog = prog;
932 dev->val.bpf_prog.id = prog->aux->id;
933 } else {
934 dev->xdp_prog = NULL;
935 dev->val.bpf_prog.id = 0;
936 }
937 dev->val.ifindex = val->ifindex;
938
939 return dev;
940err_put_prog:
941 bpf_prog_put(prog);
942err_put_dev:
943 dev_put(dev->dev);
944err_out:
945 kfree(dev);
946 return ERR_PTR(-EINVAL);
947}
948
949static long __dev_map_update_elem(struct net *net, struct bpf_map *map,
950 void *key, void *value, u64 map_flags)
951{
952 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
953 struct bpf_dtab_netdev *dev, *old_dev;
954 struct bpf_devmap_val val = {};
955 u32 i = *(u32 *)key;
956
957 if (unlikely(map_flags > BPF_EXIST))
958 return -EINVAL;
959 if (unlikely(i >= dtab->map.max_entries))
960 return -E2BIG;
961 if (unlikely(map_flags == BPF_NOEXIST))
962 return -EEXIST;
963
964 /* already verified value_size <= sizeof val */
965 memcpy(&val, value, map->value_size);
966
967 if (!val.ifindex) {
968 dev = NULL;
969 /* can not specify fd if ifindex is 0 */
970 if (val.bpf_prog.fd > 0)
971 return -EINVAL;
972 } else {
973 dev = __dev_map_alloc_node(net, dtab, &val, i);
974 if (IS_ERR(dev))
975 return PTR_ERR(dev);
976 }
977
978 /* Use call_rcu() here to ensure rcu critical sections have completed
979 * Remembering the driver side flush operation will happen before the
980 * net device is removed.
981 */
982 old_dev = unrcu_pointer(xchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev)));
983 if (old_dev)
984 call_rcu(&old_dev->rcu, __dev_map_entry_free);
985 else
986 atomic_inc((atomic_t *)&dtab->items);
987
988 return 0;
989}
990
991static long dev_map_update_elem(struct bpf_map *map, void *key, void *value,
992 u64 map_flags)
993{
994 return __dev_map_update_elem(current->nsproxy->net_ns,
995 map, key, value, map_flags);
996}
997
998static long __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
999 void *key, void *value, u64 map_flags)
1000{
1001 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1002 struct bpf_dtab_netdev *dev, *old_dev;
1003 struct bpf_devmap_val val = {};
1004 u32 idx = *(u32 *)key;
1005 unsigned long flags;
1006 int err = -EEXIST;
1007
1008 /* already verified value_size <= sizeof val */
1009 memcpy(&val, value, map->value_size);
1010
1011 if (unlikely(map_flags > BPF_EXIST || !val.ifindex))
1012 return -EINVAL;
1013
1014 spin_lock_irqsave(&dtab->index_lock, flags);
1015
1016 old_dev = __dev_map_hash_lookup_elem(map, idx);
1017 if (old_dev && (map_flags & BPF_NOEXIST))
1018 goto out_err;
1019
1020 dev = __dev_map_alloc_node(net, dtab, &val, idx);
1021 if (IS_ERR(dev)) {
1022 err = PTR_ERR(dev);
1023 goto out_err;
1024 }
1025
1026 if (old_dev) {
1027 hlist_del_rcu(&old_dev->index_hlist);
1028 } else {
1029 if (dtab->items >= dtab->map.max_entries) {
1030 spin_unlock_irqrestore(&dtab->index_lock, flags);
1031 call_rcu(&dev->rcu, __dev_map_entry_free);
1032 return -E2BIG;
1033 }
1034 dtab->items++;
1035 }
1036
1037 hlist_add_head_rcu(&dev->index_hlist,
1038 dev_map_index_hash(dtab, idx));
1039 spin_unlock_irqrestore(&dtab->index_lock, flags);
1040
1041 if (old_dev)
1042 call_rcu(&old_dev->rcu, __dev_map_entry_free);
1043
1044 return 0;
1045
1046out_err:
1047 spin_unlock_irqrestore(&dtab->index_lock, flags);
1048 return err;
1049}
1050
1051static long dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
1052 u64 map_flags)
1053{
1054 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
1055 map, key, value, map_flags);
1056}
1057
1058static long dev_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1059{
1060 return __bpf_xdp_redirect_map(map, ifindex, flags,
1061 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1062 __dev_map_lookup_elem);
1063}
1064
1065static long dev_hash_map_redirect(struct bpf_map *map, u64 ifindex, u64 flags)
1066{
1067 return __bpf_xdp_redirect_map(map, ifindex, flags,
1068 BPF_F_BROADCAST | BPF_F_EXCLUDE_INGRESS,
1069 __dev_map_hash_lookup_elem);
1070}
1071
1072static u64 dev_map_mem_usage(const struct bpf_map *map)
1073{
1074 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
1075 u64 usage = sizeof(struct bpf_dtab);
1076
1077 if (map->map_type == BPF_MAP_TYPE_DEVMAP_HASH)
1078 usage += (u64)dtab->n_buckets * sizeof(struct hlist_head);
1079 else
1080 usage += (u64)map->max_entries * sizeof(struct bpf_dtab_netdev *);
1081 usage += atomic_read((atomic_t *)&dtab->items) *
1082 (u64)sizeof(struct bpf_dtab_netdev);
1083 return usage;
1084}
1085
1086BTF_ID_LIST_SINGLE(dev_map_btf_ids, struct, bpf_dtab)
1087const struct bpf_map_ops dev_map_ops = {
1088 .map_meta_equal = bpf_map_meta_equal,
1089 .map_alloc_check = dev_map_alloc_check,
1090 .map_alloc = dev_map_alloc,
1091 .map_free = dev_map_free,
1092 .map_get_next_key = dev_map_get_next_key,
1093 .map_lookup_elem = dev_map_lookup_elem,
1094 .map_update_elem = dev_map_update_elem,
1095 .map_delete_elem = dev_map_delete_elem,
1096 .map_check_btf = map_check_no_btf,
1097 .map_mem_usage = dev_map_mem_usage,
1098 .map_btf_id = &dev_map_btf_ids[0],
1099 .map_redirect = dev_map_redirect,
1100};
1101
1102const struct bpf_map_ops dev_map_hash_ops = {
1103 .map_meta_equal = bpf_map_meta_equal,
1104 .map_alloc_check = dev_map_alloc_check,
1105 .map_alloc = dev_map_alloc,
1106 .map_free = dev_map_free,
1107 .map_get_next_key = dev_map_hash_get_next_key,
1108 .map_lookup_elem = dev_map_hash_lookup_elem,
1109 .map_update_elem = dev_map_hash_update_elem,
1110 .map_delete_elem = dev_map_hash_delete_elem,
1111 .map_check_btf = map_check_no_btf,
1112 .map_mem_usage = dev_map_mem_usage,
1113 .map_btf_id = &dev_map_btf_ids[0],
1114 .map_redirect = dev_hash_map_redirect,
1115};
1116
1117static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
1118 struct net_device *netdev)
1119{
1120 unsigned long flags;
1121 u32 i;
1122
1123 spin_lock_irqsave(&dtab->index_lock, flags);
1124 for (i = 0; i < dtab->n_buckets; i++) {
1125 struct bpf_dtab_netdev *dev;
1126 struct hlist_head *head;
1127 struct hlist_node *next;
1128
1129 head = dev_map_index_hash(dtab, i);
1130
1131 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
1132 if (netdev != dev->dev)
1133 continue;
1134
1135 dtab->items--;
1136 hlist_del_rcu(&dev->index_hlist);
1137 call_rcu(&dev->rcu, __dev_map_entry_free);
1138 }
1139 }
1140 spin_unlock_irqrestore(&dtab->index_lock, flags);
1141}
1142
1143static int dev_map_notification(struct notifier_block *notifier,
1144 ulong event, void *ptr)
1145{
1146 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
1147 struct bpf_dtab *dtab;
1148 int i, cpu;
1149
1150 switch (event) {
1151 case NETDEV_REGISTER:
1152 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
1153 break;
1154
1155 /* will be freed in free_netdev() */
1156 netdev->xdp_bulkq = alloc_percpu(struct xdp_dev_bulk_queue);
1157 if (!netdev->xdp_bulkq)
1158 return NOTIFY_BAD;
1159
1160 for_each_possible_cpu(cpu) {
1161 struct xdp_dev_bulk_queue *bq;
1162
1163 bq = per_cpu_ptr(netdev->xdp_bulkq, cpu);
1164 bq->dev = netdev;
1165 local_lock_init(&bq->bq_lock);
1166 }
1167 break;
1168 case NETDEV_UNREGISTER:
1169 /* This rcu_read_lock/unlock pair is needed because
1170 * dev_map_list is an RCU list AND to ensure a delete
1171 * operation does not free a netdev_map entry while we
1172 * are comparing it against the netdev being unregistered.
1173 */
1174 rcu_read_lock();
1175 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
1176 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
1177 dev_map_hash_remove_netdev(dtab, netdev);
1178 continue;
1179 }
1180
1181 for (i = 0; i < dtab->map.max_entries; i++) {
1182 struct bpf_dtab_netdev *dev, *odev;
1183
1184 dev = rcu_dereference(dtab->netdev_map[i]);
1185 if (!dev || netdev != dev->dev)
1186 continue;
1187 odev = unrcu_pointer(cmpxchg(&dtab->netdev_map[i], RCU_INITIALIZER(dev), NULL));
1188 if (dev == odev) {
1189 call_rcu(&dev->rcu,
1190 __dev_map_entry_free);
1191 atomic_dec((atomic_t *)&dtab->items);
1192 }
1193 }
1194 }
1195 rcu_read_unlock();
1196 break;
1197 default:
1198 break;
1199 }
1200 return NOTIFY_OK;
1201}
1202
1203static struct notifier_block dev_map_notifier = {
1204 .notifier_call = dev_map_notification,
1205};
1206
1207static int __init dev_map_init(void)
1208{
1209 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
1210 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
1211 offsetof(struct _bpf_dtab_netdev, dev));
1212 register_netdevice_notifier(&dev_map_notifier);
1213
1214 return 0;
1215}
1216
1217subsys_initcall(dev_map_init);