1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
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.
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.
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.
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.
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.
47 #include <linux/bpf.h>
49 #include <linux/filter.h>
50 #include <trace/events/xdp.h>
52 #define DEV_CREATE_FLAG_MASK \
53 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
55 #define DEV_MAP_BULK_SIZE 16
56 struct bpf_dtab_netdev;
58 struct 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;
66 struct 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;
72 unsigned int idx; /* keep track of map index for tracepoint */
77 struct bpf_dtab_netdev **netdev_map;
78 struct list_head __percpu *flush_list;
79 struct list_head list;
81 /* these are only used for DEVMAP_HASH type maps */
82 struct hlist_head *dev_index_head;
83 spinlock_t index_lock;
88 static DEFINE_SPINLOCK(dev_map_lock);
89 static LIST_HEAD(dev_map_list);
91 static struct hlist_head *dev_map_create_hash(unsigned int entries)
94 struct hlist_head *hash;
96 hash = kmalloc_array(entries, sizeof(*hash), GFP_KERNEL);
98 for (i = 0; i < entries; i++)
99 INIT_HLIST_HEAD(&hash[i]);
104 static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
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)
114 /* Lookup returns a pointer straight to dev->ifindex, so make sure the
115 * verifier prevents writes from the BPF side
117 attr->map_flags |= BPF_F_RDONLY_PROG;
120 bpf_map_init_from_attr(&dtab->map, attr);
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();
126 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
127 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
129 if (!dtab->n_buckets) /* Overflow check */
131 cost += sizeof(struct hlist_head) * dtab->n_buckets;
134 /* if map size is larger than memlock limit, reject it */
135 err = bpf_map_charge_init(&dtab->map.memory, cost);
139 dtab->flush_list = alloc_percpu(struct list_head);
140 if (!dtab->flush_list)
143 for_each_possible_cpu(cpu)
144 INIT_LIST_HEAD(per_cpu_ptr(dtab->flush_list, cpu));
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)
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)
157 spin_lock_init(&dtab->index_lock);
163 bpf_map_area_free(dtab->netdev_map);
165 free_percpu(dtab->flush_list);
167 bpf_map_charge_finish(&dtab->map.memory);
171 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
173 struct bpf_dtab *dtab;
176 if (!capable(CAP_NET_ADMIN))
177 return ERR_PTR(-EPERM);
179 dtab = kzalloc(sizeof(*dtab), GFP_USER);
181 return ERR_PTR(-ENOMEM);
183 err = dev_map_init_map(dtab, attr);
189 spin_lock(&dev_map_lock);
190 list_add_tail_rcu(&dtab->list, &dev_map_list);
191 spin_unlock(&dev_map_lock);
196 static void dev_map_free(struct bpf_map *map)
198 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
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.
209 spin_lock(&dev_map_lock);
210 list_del_rcu(&dtab->list);
211 spin_unlock(&dev_map_lock);
213 bpf_clear_redirect_map(map);
216 /* Make sure prior __dev_map_entry_free() have completed. */
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.
224 for_each_online_cpu(cpu) {
225 struct list_head *flush_list = per_cpu_ptr(dtab->flush_list, cpu);
227 while (!list_empty(flush_list))
231 for (i = 0; i < dtab->map.max_entries; i++) {
232 struct bpf_dtab_netdev *dev;
234 dev = dtab->netdev_map[i];
238 free_percpu(dev->bulkq);
243 free_percpu(dtab->flush_list);
244 bpf_map_area_free(dtab->netdev_map);
245 kfree(dtab->dev_index_head);
249 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
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;
255 if (index >= dtab->map.max_entries) {
260 if (index == dtab->map.max_entries - 1)
266 static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
269 return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
272 struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
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;
278 hlist_for_each_entry_rcu(dev, head, index_hlist)
285 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
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;
299 dev = __dev_map_hash_lookup_elem(map, idx);
303 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
304 struct bpf_dtab_netdev, index_hlist);
307 *next = next_dev->idx;
311 i = idx & (dtab->n_buckets - 1);
315 for (; i < dtab->n_buckets; i++) {
316 head = dev_map_index_hash(dtab, i);
318 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
319 struct bpf_dtab_netdev,
322 *next = next_dev->idx;
330 static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags,
333 struct bpf_dtab_netdev *obj = bq->obj;
334 struct net_device *dev = obj->dev;
335 int sent = 0, drops = 0, err = 0;
338 if (unlikely(!bq->count))
341 for (i = 0; i < bq->count; i++) {
342 struct xdp_frame *xdpf = bq->q[i];
347 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
353 drops = bq->count - sent;
357 trace_xdp_devmap_xmit(&obj->dtab->map, obj->idx,
358 sent, drops, bq->dev_rx, dev, err);
360 __list_del_clearprev(&bq->flush_node);
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.
366 for (i = 0; i < bq->count; i++) {
367 struct xdp_frame *xdpf = bq->q[i];
369 /* RX path under NAPI protection, can return frames faster */
370 if (likely(in_napi_ctx))
371 xdp_return_frame_rx_napi(xdpf);
373 xdp_return_frame(xdpf);
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.
386 void __dev_map_flush(struct bpf_map *map)
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;
393 list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
394 bq_xmit_all(bq, XDP_XMIT_FLUSH, true);
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
402 struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
404 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
405 struct bpf_dtab_netdev *obj;
407 if (key >= map->max_entries)
410 obj = READ_ONCE(dtab->netdev_map[key]);
414 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
415 * Thus, safe percpu variable access.
417 static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
418 struct net_device *dev_rx)
421 struct list_head *flush_list = this_cpu_ptr(obj->dtab->flush_list);
422 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
424 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
425 bq_xmit_all(bq, 0, true);
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.
434 bq->q[bq->count++] = xdpf;
436 if (!bq->flush_node.prev)
437 list_add(&bq->flush_node, flush_list);
442 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
443 struct net_device *dev_rx)
445 struct net_device *dev = dst->dev;
446 struct xdp_frame *xdpf;
449 if (!dev->netdev_ops->ndo_xdp_xmit)
452 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
456 xdpf = convert_to_xdp_frame(xdp);
460 return bq_enqueue(dst, xdpf, dev_rx);
463 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
464 struct bpf_prog *xdp_prog)
468 err = xdp_ok_fwd_dev(dst->dev, skb->len);
472 generic_xdp_tx(skb, xdp_prog);
477 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
479 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
480 struct net_device *dev = obj ? obj->dev : NULL;
482 return dev ? &dev->ifindex : NULL;
485 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
487 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
489 struct net_device *dev = obj ? obj->dev : NULL;
491 return dev ? &dev->ifindex : NULL;
494 static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
496 if (dev->dev->netdev_ops->ndo_xdp_xmit) {
497 struct xdp_bulk_queue *bq;
501 for_each_online_cpu(cpu) {
502 bq = per_cpu_ptr(dev->bulkq, cpu);
503 bq_xmit_all(bq, XDP_XMIT_FLUSH, false);
509 static void __dev_map_entry_free(struct rcu_head *rcu)
511 struct bpf_dtab_netdev *dev;
513 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
514 dev_map_flush_old(dev);
515 free_percpu(dev->bulkq);
520 static int dev_map_delete_elem(struct bpf_map *map, void *key)
522 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
523 struct bpf_dtab_netdev *old_dev;
526 if (k >= map->max_entries)
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.
537 old_dev = xchg(&dtab->netdev_map[k], NULL);
539 call_rcu(&old_dev->rcu, __dev_map_entry_free);
543 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
545 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
546 struct bpf_dtab_netdev *old_dev;
551 spin_lock_irqsave(&dtab->index_lock, flags);
553 old_dev = __dev_map_hash_lookup_elem(map, k);
556 hlist_del_init_rcu(&old_dev->index_hlist);
557 call_rcu(&old_dev->rcu, __dev_map_entry_free);
560 spin_unlock_irqrestore(&dtab->index_lock, flags);
565 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
566 struct bpf_dtab *dtab,
570 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
571 struct bpf_dtab_netdev *dev;
572 struct xdp_bulk_queue *bq;
575 dev = kmalloc_node(sizeof(*dev), gfp, dtab->map.numa_node);
577 return ERR_PTR(-ENOMEM);
579 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
580 sizeof(void *), gfp);
583 return ERR_PTR(-ENOMEM);
586 for_each_possible_cpu(cpu) {
587 bq = per_cpu_ptr(dev->bulkq, cpu);
591 dev->dev = dev_get_by_index(net, ifindex);
593 free_percpu(dev->bulkq);
595 return ERR_PTR(-EINVAL);
604 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
605 void *key, void *value, u64 map_flags)
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;
612 if (unlikely(map_flags > BPF_EXIST))
614 if (unlikely(i >= dtab->map.max_entries))
616 if (unlikely(map_flags == BPF_NOEXIST))
622 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
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.
631 old_dev = xchg(&dtab->netdev_map[i], dev);
633 call_rcu(&old_dev->rcu, __dev_map_entry_free);
638 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
641 return __dev_map_update_elem(current->nsproxy->net_ns,
642 map, key, value, map_flags);
645 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
646 void *key, void *value, u64 map_flags)
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;
655 if (unlikely(map_flags > BPF_EXIST || !ifindex))
658 spin_lock_irqsave(&dtab->index_lock, flags);
660 old_dev = __dev_map_hash_lookup_elem(map, idx);
661 if (old_dev && (map_flags & BPF_NOEXIST))
664 dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
671 hlist_del_rcu(&old_dev->index_hlist);
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);
681 hlist_add_head_rcu(&dev->index_hlist,
682 dev_map_index_hash(dtab, idx));
683 spin_unlock_irqrestore(&dtab->index_lock, flags);
686 call_rcu(&old_dev->rcu, __dev_map_entry_free);
691 spin_unlock_irqrestore(&dtab->index_lock, flags);
695 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
698 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
699 map, key, value, map_flags);
702 const 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,
712 const 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,
722 static int dev_map_notification(struct notifier_block *notifier,
723 ulong event, void *ptr)
725 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
726 struct bpf_dtab *dtab;
730 case NETDEV_UNREGISTER:
731 /* This rcu_read_lock/unlock pair is needed because
732 * dev_map_list is an RCU list AND to ensure a delete
733 * operation does not free a netdev_map entry while we
734 * are comparing it against the netdev being unregistered.
737 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
738 for (i = 0; i < dtab->map.max_entries; i++) {
739 struct bpf_dtab_netdev *dev, *odev;
741 dev = READ_ONCE(dtab->netdev_map[i]);
742 if (!dev || netdev != dev->dev)
744 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
747 __dev_map_entry_free);
758 static struct notifier_block dev_map_notifier = {
759 .notifier_call = dev_map_notification,
762 static int __init dev_map_init(void)
764 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
765 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
766 offsetof(struct _bpf_dtab_netdev, dev));
767 register_netdevice_notifier(&dev_map_notifier);
771 subsys_initcall(dev_map_init);