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 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;
64 struct bpf_dtab_netdev {
65 struct net_device *dev; /* must be first member, due to tracepoint */
66 struct hlist_node index_hlist;
67 struct bpf_dtab *dtab;
74 struct bpf_dtab_netdev **netdev_map; /* DEVMAP type only */
75 struct list_head list;
77 /* these are only used for DEVMAP_HASH type maps */
78 struct hlist_head *dev_index_head;
79 spinlock_t index_lock;
84 static DEFINE_PER_CPU(struct list_head, dev_flush_list);
85 static DEFINE_SPINLOCK(dev_map_lock);
86 static LIST_HEAD(dev_map_list);
88 static struct hlist_head *dev_map_create_hash(unsigned int entries)
91 struct hlist_head *hash;
93 hash = kmalloc_array(entries, sizeof(*hash), GFP_KERNEL);
95 for (i = 0; i < entries; i++)
96 INIT_HLIST_HEAD(&hash[i]);
101 static inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
104 return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
107 static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
112 /* check sanity of attributes */
113 if (attr->max_entries == 0 || attr->key_size != 4 ||
114 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
117 /* Lookup returns a pointer straight to dev->ifindex, so make sure the
118 * verifier prevents writes from the BPF side
120 attr->map_flags |= BPF_F_RDONLY_PROG;
123 bpf_map_init_from_attr(&dtab->map, attr);
125 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
126 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
128 if (!dtab->n_buckets) /* Overflow check */
130 cost += (u64) sizeof(struct hlist_head) * dtab->n_buckets;
132 cost += (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
135 /* if map size is larger than memlock limit, reject it */
136 err = bpf_map_charge_init(&dtab->map.memory, cost);
140 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
141 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets);
142 if (!dtab->dev_index_head)
145 spin_lock_init(&dtab->index_lock);
147 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
148 sizeof(struct bpf_dtab_netdev *),
149 dtab->map.numa_node);
150 if (!dtab->netdev_map)
157 bpf_map_charge_finish(&dtab->map.memory);
161 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
163 struct bpf_dtab *dtab;
166 if (!capable(CAP_NET_ADMIN))
167 return ERR_PTR(-EPERM);
169 dtab = kzalloc(sizeof(*dtab), GFP_USER);
171 return ERR_PTR(-ENOMEM);
173 err = dev_map_init_map(dtab, attr);
179 spin_lock(&dev_map_lock);
180 list_add_tail_rcu(&dtab->list, &dev_map_list);
181 spin_unlock(&dev_map_lock);
186 static void dev_map_free(struct bpf_map *map)
188 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
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. Wait for outstanding critical sections in
194 * these programs to complete. The rcu critical section only guarantees
195 * no further reads against netdev_map. It does __not__ ensure pending
196 * flush operations (if any) are complete.
199 spin_lock(&dev_map_lock);
200 list_del_rcu(&dtab->list);
201 spin_unlock(&dev_map_lock);
203 bpf_clear_redirect_map(map);
206 /* Make sure prior __dev_map_entry_free() have completed. */
209 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
210 for (i = 0; i < dtab->n_buckets; i++) {
211 struct bpf_dtab_netdev *dev;
212 struct hlist_head *head;
213 struct hlist_node *next;
215 head = dev_map_index_hash(dtab, i);
217 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
218 hlist_del_rcu(&dev->index_hlist);
224 kfree(dtab->dev_index_head);
226 for (i = 0; i < dtab->map.max_entries; i++) {
227 struct bpf_dtab_netdev *dev;
229 dev = dtab->netdev_map[i];
237 bpf_map_area_free(dtab->netdev_map);
243 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
245 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
246 u32 index = key ? *(u32 *)key : U32_MAX;
247 u32 *next = next_key;
249 if (index >= dtab->map.max_entries) {
254 if (index == dtab->map.max_entries - 1)
260 struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
262 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
263 struct hlist_head *head = dev_map_index_hash(dtab, key);
264 struct bpf_dtab_netdev *dev;
266 hlist_for_each_entry_rcu(dev, head, index_hlist)
273 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
276 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
277 u32 idx, *next = next_key;
278 struct bpf_dtab_netdev *dev, *next_dev;
279 struct hlist_head *head;
287 dev = __dev_map_hash_lookup_elem(map, idx);
291 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
292 struct bpf_dtab_netdev, index_hlist);
295 *next = next_dev->idx;
299 i = idx & (dtab->n_buckets - 1);
303 for (; i < dtab->n_buckets; i++) {
304 head = dev_map_index_hash(dtab, i);
306 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
307 struct bpf_dtab_netdev,
310 *next = next_dev->idx;
318 static int bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
320 struct net_device *dev = bq->dev;
321 int sent = 0, drops = 0, err = 0;
324 if (unlikely(!bq->count))
327 for (i = 0; i < bq->count; i++) {
328 struct xdp_frame *xdpf = bq->q[i];
333 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
339 drops = bq->count - sent;
343 trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, drops, err);
345 __list_del_clearprev(&bq->flush_node);
348 /* If ndo_xdp_xmit fails with an errno, no frames have been
349 * xmit'ed and it's our responsibility to them free all.
351 for (i = 0; i < bq->count; i++) {
352 struct xdp_frame *xdpf = bq->q[i];
354 xdp_return_frame_rx_napi(xdpf);
360 /* __dev_flush is called from xdp_do_flush() which _must_ be signaled
361 * from the driver before returning from its napi->poll() routine. The poll()
362 * routine is called either from busy_poll context or net_rx_action signaled
363 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
364 * net device can be torn down. On devmap tear down we ensure the flush list
365 * is empty before completing to ensure all flush operations have completed.
367 void __dev_flush(void)
369 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
370 struct xdp_dev_bulk_queue *bq, *tmp;
373 list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
374 bq_xmit_all(bq, XDP_XMIT_FLUSH);
378 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
379 * update happens in parallel here a dev_put wont happen until after reading the
382 struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
384 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
385 struct bpf_dtab_netdev *obj;
387 if (key >= map->max_entries)
390 obj = READ_ONCE(dtab->netdev_map[key]);
394 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
395 * Thus, safe percpu variable access.
397 static int bq_enqueue(struct net_device *dev, struct xdp_frame *xdpf,
398 struct net_device *dev_rx)
400 struct list_head *flush_list = this_cpu_ptr(&dev_flush_list);
401 struct xdp_dev_bulk_queue *bq = this_cpu_ptr(dev->xdp_bulkq);
403 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
406 /* Ingress dev_rx will be the same for all xdp_frame's in
407 * bulk_queue, because bq stored per-CPU and must be flushed
408 * from net_device drivers NAPI func end.
413 bq->q[bq->count++] = xdpf;
415 if (!bq->flush_node.prev)
416 list_add(&bq->flush_node, flush_list);
421 static inline int __xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
422 struct net_device *dev_rx)
424 struct xdp_frame *xdpf;
427 if (!dev->netdev_ops->ndo_xdp_xmit)
430 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
434 xdpf = convert_to_xdp_frame(xdp);
438 return bq_enqueue(dev, xdpf, dev_rx);
441 int dev_xdp_enqueue(struct net_device *dev, struct xdp_buff *xdp,
442 struct net_device *dev_rx)
444 return __xdp_enqueue(dev, xdp, dev_rx);
447 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
448 struct net_device *dev_rx)
450 struct net_device *dev = dst->dev;
452 return __xdp_enqueue(dev, xdp, dev_rx);
455 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
456 struct bpf_prog *xdp_prog)
460 err = xdp_ok_fwd_dev(dst->dev, skb->len);
464 generic_xdp_tx(skb, xdp_prog);
469 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
471 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
472 struct net_device *dev = obj ? obj->dev : NULL;
474 return dev ? &dev->ifindex : NULL;
477 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
479 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
481 struct net_device *dev = obj ? obj->dev : NULL;
483 return dev ? &dev->ifindex : NULL;
486 static void __dev_map_entry_free(struct rcu_head *rcu)
488 struct bpf_dtab_netdev *dev;
490 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
495 static int dev_map_delete_elem(struct bpf_map *map, void *key)
497 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
498 struct bpf_dtab_netdev *old_dev;
501 if (k >= map->max_entries)
504 /* Use call_rcu() here to ensure any rcu critical sections have
505 * completed, but this does not guarantee a flush has happened
506 * yet. Because driver side rcu_read_lock/unlock only protects the
507 * running XDP program. However, for pending flush operations the
508 * dev and ctx are stored in another per cpu map. And additionally,
509 * the driver tear down ensures all soft irqs are complete before
510 * removing the net device in the case of dev_put equals zero.
512 old_dev = xchg(&dtab->netdev_map[k], NULL);
514 call_rcu(&old_dev->rcu, __dev_map_entry_free);
518 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
520 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
521 struct bpf_dtab_netdev *old_dev;
526 spin_lock_irqsave(&dtab->index_lock, flags);
528 old_dev = __dev_map_hash_lookup_elem(map, k);
531 hlist_del_init_rcu(&old_dev->index_hlist);
532 call_rcu(&old_dev->rcu, __dev_map_entry_free);
535 spin_unlock_irqrestore(&dtab->index_lock, flags);
540 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
541 struct bpf_dtab *dtab,
545 struct bpf_dtab_netdev *dev;
547 dev = kmalloc_node(sizeof(*dev), GFP_ATOMIC | __GFP_NOWARN,
548 dtab->map.numa_node);
550 return ERR_PTR(-ENOMEM);
552 dev->dev = dev_get_by_index(net, ifindex);
555 return ERR_PTR(-EINVAL);
564 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
565 void *key, void *value, u64 map_flags)
567 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
568 struct bpf_dtab_netdev *dev, *old_dev;
569 u32 ifindex = *(u32 *)value;
572 if (unlikely(map_flags > BPF_EXIST))
574 if (unlikely(i >= dtab->map.max_entries))
576 if (unlikely(map_flags == BPF_NOEXIST))
582 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
587 /* Use call_rcu() here to ensure rcu critical sections have completed
588 * Remembering the driver side flush operation will happen before the
589 * net device is removed.
591 old_dev = xchg(&dtab->netdev_map[i], dev);
593 call_rcu(&old_dev->rcu, __dev_map_entry_free);
598 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
601 return __dev_map_update_elem(current->nsproxy->net_ns,
602 map, key, value, map_flags);
605 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
606 void *key, void *value, u64 map_flags)
608 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
609 struct bpf_dtab_netdev *dev, *old_dev;
610 u32 ifindex = *(u32 *)value;
611 u32 idx = *(u32 *)key;
615 if (unlikely(map_flags > BPF_EXIST || !ifindex))
618 spin_lock_irqsave(&dtab->index_lock, flags);
620 old_dev = __dev_map_hash_lookup_elem(map, idx);
621 if (old_dev && (map_flags & BPF_NOEXIST))
624 dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
631 hlist_del_rcu(&old_dev->index_hlist);
633 if (dtab->items >= dtab->map.max_entries) {
634 spin_unlock_irqrestore(&dtab->index_lock, flags);
635 call_rcu(&dev->rcu, __dev_map_entry_free);
641 hlist_add_head_rcu(&dev->index_hlist,
642 dev_map_index_hash(dtab, idx));
643 spin_unlock_irqrestore(&dtab->index_lock, flags);
646 call_rcu(&old_dev->rcu, __dev_map_entry_free);
651 spin_unlock_irqrestore(&dtab->index_lock, flags);
655 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
658 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
659 map, key, value, map_flags);
662 const struct bpf_map_ops dev_map_ops = {
663 .map_alloc = dev_map_alloc,
664 .map_free = dev_map_free,
665 .map_get_next_key = dev_map_get_next_key,
666 .map_lookup_elem = dev_map_lookup_elem,
667 .map_update_elem = dev_map_update_elem,
668 .map_delete_elem = dev_map_delete_elem,
669 .map_check_btf = map_check_no_btf,
672 const struct bpf_map_ops dev_map_hash_ops = {
673 .map_alloc = dev_map_alloc,
674 .map_free = dev_map_free,
675 .map_get_next_key = dev_map_hash_get_next_key,
676 .map_lookup_elem = dev_map_hash_lookup_elem,
677 .map_update_elem = dev_map_hash_update_elem,
678 .map_delete_elem = dev_map_hash_delete_elem,
679 .map_check_btf = map_check_no_btf,
682 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
683 struct net_device *netdev)
688 spin_lock_irqsave(&dtab->index_lock, flags);
689 for (i = 0; i < dtab->n_buckets; i++) {
690 struct bpf_dtab_netdev *dev;
691 struct hlist_head *head;
692 struct hlist_node *next;
694 head = dev_map_index_hash(dtab, i);
696 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
697 if (netdev != dev->dev)
701 hlist_del_rcu(&dev->index_hlist);
702 call_rcu(&dev->rcu, __dev_map_entry_free);
705 spin_unlock_irqrestore(&dtab->index_lock, flags);
708 static int dev_map_notification(struct notifier_block *notifier,
709 ulong event, void *ptr)
711 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
712 struct bpf_dtab *dtab;
716 case NETDEV_REGISTER:
717 if (!netdev->netdev_ops->ndo_xdp_xmit || netdev->xdp_bulkq)
720 /* will be freed in free_netdev() */
722 __alloc_percpu_gfp(sizeof(struct xdp_dev_bulk_queue),
723 sizeof(void *), GFP_ATOMIC);
724 if (!netdev->xdp_bulkq)
727 for_each_possible_cpu(cpu)
728 per_cpu_ptr(netdev->xdp_bulkq, cpu)->dev = netdev;
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 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
739 dev_map_hash_remove_netdev(dtab, netdev);
743 for (i = 0; i < dtab->map.max_entries; i++) {
744 struct bpf_dtab_netdev *dev, *odev;
746 dev = READ_ONCE(dtab->netdev_map[i]);
747 if (!dev || netdev != dev->dev)
749 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
752 __dev_map_entry_free);
763 static struct notifier_block dev_map_notifier = {
764 .notifier_call = dev_map_notification,
767 static int __init dev_map_init(void)
771 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
772 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
773 offsetof(struct _bpf_dtab_netdev, dev));
774 register_netdevice_notifier(&dev_map_notifier);
776 for_each_possible_cpu(cpu)
777 INIT_LIST_HEAD(&per_cpu(dev_flush_list, cpu));
781 subsys_initcall(dev_map_init);