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; /* DEVMAP type only */
78 struct list_head list;
80 /* these are only used for DEVMAP_HASH type maps */
81 struct hlist_head *dev_index_head;
82 spinlock_t index_lock;
87 static DEFINE_PER_CPU(struct list_head, dev_map_flush_list);
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 inline struct hlist_head *dev_map_index_hash(struct bpf_dtab *dtab,
107 return &dtab->dev_index_head[idx & (dtab->n_buckets - 1)];
110 static int dev_map_init_map(struct bpf_dtab *dtab, union bpf_attr *attr)
115 /* check sanity of attributes */
116 if (attr->max_entries == 0 || attr->key_size != 4 ||
117 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
120 /* Lookup returns a pointer straight to dev->ifindex, so make sure the
121 * verifier prevents writes from the BPF side
123 attr->map_flags |= BPF_F_RDONLY_PROG;
126 bpf_map_init_from_attr(&dtab->map, attr);
128 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
129 dtab->n_buckets = roundup_pow_of_two(dtab->map.max_entries);
131 if (!dtab->n_buckets) /* Overflow check */
133 cost += (u64) sizeof(struct hlist_head) * dtab->n_buckets;
135 cost += (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
138 /* if map size is larger than memlock limit, reject it */
139 err = bpf_map_charge_init(&dtab->map.memory, cost);
143 if (attr->map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
144 dtab->dev_index_head = dev_map_create_hash(dtab->n_buckets);
145 if (!dtab->dev_index_head)
148 spin_lock_init(&dtab->index_lock);
150 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
151 sizeof(struct bpf_dtab_netdev *),
152 dtab->map.numa_node);
153 if (!dtab->netdev_map)
160 bpf_map_charge_finish(&dtab->map.memory);
164 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
166 struct bpf_dtab *dtab;
169 if (!capable(CAP_NET_ADMIN))
170 return ERR_PTR(-EPERM);
172 dtab = kzalloc(sizeof(*dtab), GFP_USER);
174 return ERR_PTR(-ENOMEM);
176 err = dev_map_init_map(dtab, attr);
182 spin_lock(&dev_map_lock);
183 list_add_tail_rcu(&dtab->list, &dev_map_list);
184 spin_unlock(&dev_map_lock);
189 static void dev_map_free(struct bpf_map *map)
191 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
194 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
195 * so the programs (can be more than one that used this map) were
196 * disconnected from events. Wait for outstanding critical sections in
197 * these programs to complete. The rcu critical section only guarantees
198 * no further reads against netdev_map. It does __not__ ensure pending
199 * flush operations (if any) are complete.
202 spin_lock(&dev_map_lock);
203 list_del_rcu(&dtab->list);
204 spin_unlock(&dev_map_lock);
206 bpf_clear_redirect_map(map);
209 /* Make sure prior __dev_map_entry_free() have completed. */
212 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
213 for (i = 0; i < dtab->n_buckets; i++) {
214 struct bpf_dtab_netdev *dev;
215 struct hlist_head *head;
216 struct hlist_node *next;
218 head = dev_map_index_hash(dtab, i);
220 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
221 hlist_del_rcu(&dev->index_hlist);
222 free_percpu(dev->bulkq);
228 kfree(dtab->dev_index_head);
230 for (i = 0; i < dtab->map.max_entries; i++) {
231 struct bpf_dtab_netdev *dev;
233 dev = dtab->netdev_map[i];
237 free_percpu(dev->bulkq);
242 bpf_map_area_free(dtab->netdev_map);
248 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
250 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
251 u32 index = key ? *(u32 *)key : U32_MAX;
252 u32 *next = next_key;
254 if (index >= dtab->map.max_entries) {
259 if (index == dtab->map.max_entries - 1)
265 struct bpf_dtab_netdev *__dev_map_hash_lookup_elem(struct bpf_map *map, u32 key)
267 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
268 struct hlist_head *head = dev_map_index_hash(dtab, key);
269 struct bpf_dtab_netdev *dev;
271 hlist_for_each_entry_rcu(dev, head, index_hlist)
278 static int dev_map_hash_get_next_key(struct bpf_map *map, void *key,
281 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
282 u32 idx, *next = next_key;
283 struct bpf_dtab_netdev *dev, *next_dev;
284 struct hlist_head *head;
292 dev = __dev_map_hash_lookup_elem(map, idx);
296 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&dev->index_hlist)),
297 struct bpf_dtab_netdev, index_hlist);
300 *next = next_dev->idx;
304 i = idx & (dtab->n_buckets - 1);
308 for (; i < dtab->n_buckets; i++) {
309 head = dev_map_index_hash(dtab, i);
311 next_dev = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
312 struct bpf_dtab_netdev,
315 *next = next_dev->idx;
323 static int bq_xmit_all(struct xdp_bulk_queue *bq, u32 flags)
325 struct bpf_dtab_netdev *obj = bq->obj;
326 struct net_device *dev = obj->dev;
327 int sent = 0, drops = 0, err = 0;
330 if (unlikely(!bq->count))
333 for (i = 0; i < bq->count; i++) {
334 struct xdp_frame *xdpf = bq->q[i];
339 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
345 drops = bq->count - sent;
349 trace_xdp_devmap_xmit(&obj->dtab->map, obj->idx,
350 sent, drops, bq->dev_rx, dev, err);
352 __list_del_clearprev(&bq->flush_node);
355 /* If ndo_xdp_xmit fails with an errno, no frames have been
356 * xmit'ed and it's our responsibility to them free all.
358 for (i = 0; i < bq->count; i++) {
359 struct xdp_frame *xdpf = bq->q[i];
361 xdp_return_frame_rx_napi(xdpf);
367 /* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
368 * from the driver before returning from its napi->poll() routine. The poll()
369 * routine is called either from busy_poll context or net_rx_action signaled
370 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
371 * net device can be torn down. On devmap tear down we ensure the flush list
372 * is empty before completing to ensure all flush operations have completed.
374 void __dev_map_flush(void)
376 struct list_head *flush_list = this_cpu_ptr(&dev_map_flush_list);
377 struct xdp_bulk_queue *bq, *tmp;
380 list_for_each_entry_safe(bq, tmp, flush_list, flush_node)
381 bq_xmit_all(bq, XDP_XMIT_FLUSH);
385 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
386 * update happens in parallel here a dev_put wont happen until after reading the
389 struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
391 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
392 struct bpf_dtab_netdev *obj;
394 if (key >= map->max_entries)
397 obj = READ_ONCE(dtab->netdev_map[key]);
401 /* Runs under RCU-read-side, plus in softirq under NAPI protection.
402 * Thus, safe percpu variable access.
404 static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
405 struct net_device *dev_rx)
408 struct list_head *flush_list = this_cpu_ptr(&dev_map_flush_list);
409 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
411 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
414 /* Ingress dev_rx will be the same for all xdp_frame's in
415 * bulk_queue, because bq stored per-CPU and must be flushed
416 * from net_device drivers NAPI func end.
421 bq->q[bq->count++] = xdpf;
423 if (!bq->flush_node.prev)
424 list_add(&bq->flush_node, flush_list);
429 int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
430 struct net_device *dev_rx)
432 struct net_device *dev = dst->dev;
433 struct xdp_frame *xdpf;
436 if (!dev->netdev_ops->ndo_xdp_xmit)
439 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
443 xdpf = convert_to_xdp_frame(xdp);
447 return bq_enqueue(dst, xdpf, dev_rx);
450 int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
451 struct bpf_prog *xdp_prog)
455 err = xdp_ok_fwd_dev(dst->dev, skb->len);
459 generic_xdp_tx(skb, xdp_prog);
464 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
466 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
467 struct net_device *dev = obj ? obj->dev : NULL;
469 return dev ? &dev->ifindex : NULL;
472 static void *dev_map_hash_lookup_elem(struct bpf_map *map, void *key)
474 struct bpf_dtab_netdev *obj = __dev_map_hash_lookup_elem(map,
476 struct net_device *dev = obj ? obj->dev : NULL;
478 return dev ? &dev->ifindex : NULL;
481 static void __dev_map_entry_free(struct rcu_head *rcu)
483 struct bpf_dtab_netdev *dev;
485 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
486 free_percpu(dev->bulkq);
491 static int dev_map_delete_elem(struct bpf_map *map, void *key)
493 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
494 struct bpf_dtab_netdev *old_dev;
497 if (k >= map->max_entries)
500 /* Use call_rcu() here to ensure any rcu critical sections have
501 * completed, but this does not guarantee a flush has happened
502 * yet. Because driver side rcu_read_lock/unlock only protects the
503 * running XDP program. However, for pending flush operations the
504 * dev and ctx are stored in another per cpu map. And additionally,
505 * the driver tear down ensures all soft irqs are complete before
506 * removing the net device in the case of dev_put equals zero.
508 old_dev = xchg(&dtab->netdev_map[k], NULL);
510 call_rcu(&old_dev->rcu, __dev_map_entry_free);
514 static int dev_map_hash_delete_elem(struct bpf_map *map, void *key)
516 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
517 struct bpf_dtab_netdev *old_dev;
522 spin_lock_irqsave(&dtab->index_lock, flags);
524 old_dev = __dev_map_hash_lookup_elem(map, k);
527 hlist_del_init_rcu(&old_dev->index_hlist);
528 call_rcu(&old_dev->rcu, __dev_map_entry_free);
531 spin_unlock_irqrestore(&dtab->index_lock, flags);
536 static struct bpf_dtab_netdev *__dev_map_alloc_node(struct net *net,
537 struct bpf_dtab *dtab,
541 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
542 struct bpf_dtab_netdev *dev;
543 struct xdp_bulk_queue *bq;
546 dev = kmalloc_node(sizeof(*dev), gfp, dtab->map.numa_node);
548 return ERR_PTR(-ENOMEM);
550 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
551 sizeof(void *), gfp);
554 return ERR_PTR(-ENOMEM);
557 for_each_possible_cpu(cpu) {
558 bq = per_cpu_ptr(dev->bulkq, cpu);
562 dev->dev = dev_get_by_index(net, ifindex);
564 free_percpu(dev->bulkq);
566 return ERR_PTR(-EINVAL);
575 static int __dev_map_update_elem(struct net *net, struct bpf_map *map,
576 void *key, void *value, u64 map_flags)
578 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
579 struct bpf_dtab_netdev *dev, *old_dev;
580 u32 ifindex = *(u32 *)value;
583 if (unlikely(map_flags > BPF_EXIST))
585 if (unlikely(i >= dtab->map.max_entries))
587 if (unlikely(map_flags == BPF_NOEXIST))
593 dev = __dev_map_alloc_node(net, dtab, ifindex, i);
598 /* Use call_rcu() here to ensure rcu critical sections have completed
599 * Remembering the driver side flush operation will happen before the
600 * net device is removed.
602 old_dev = xchg(&dtab->netdev_map[i], dev);
604 call_rcu(&old_dev->rcu, __dev_map_entry_free);
609 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
612 return __dev_map_update_elem(current->nsproxy->net_ns,
613 map, key, value, map_flags);
616 static int __dev_map_hash_update_elem(struct net *net, struct bpf_map *map,
617 void *key, void *value, u64 map_flags)
619 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
620 struct bpf_dtab_netdev *dev, *old_dev;
621 u32 ifindex = *(u32 *)value;
622 u32 idx = *(u32 *)key;
626 if (unlikely(map_flags > BPF_EXIST || !ifindex))
629 spin_lock_irqsave(&dtab->index_lock, flags);
631 old_dev = __dev_map_hash_lookup_elem(map, idx);
632 if (old_dev && (map_flags & BPF_NOEXIST))
635 dev = __dev_map_alloc_node(net, dtab, ifindex, idx);
642 hlist_del_rcu(&old_dev->index_hlist);
644 if (dtab->items >= dtab->map.max_entries) {
645 spin_unlock_irqrestore(&dtab->index_lock, flags);
646 call_rcu(&dev->rcu, __dev_map_entry_free);
652 hlist_add_head_rcu(&dev->index_hlist,
653 dev_map_index_hash(dtab, idx));
654 spin_unlock_irqrestore(&dtab->index_lock, flags);
657 call_rcu(&old_dev->rcu, __dev_map_entry_free);
662 spin_unlock_irqrestore(&dtab->index_lock, flags);
666 static int dev_map_hash_update_elem(struct bpf_map *map, void *key, void *value,
669 return __dev_map_hash_update_elem(current->nsproxy->net_ns,
670 map, key, value, map_flags);
673 const struct bpf_map_ops dev_map_ops = {
674 .map_alloc = dev_map_alloc,
675 .map_free = dev_map_free,
676 .map_get_next_key = dev_map_get_next_key,
677 .map_lookup_elem = dev_map_lookup_elem,
678 .map_update_elem = dev_map_update_elem,
679 .map_delete_elem = dev_map_delete_elem,
680 .map_check_btf = map_check_no_btf,
683 const struct bpf_map_ops dev_map_hash_ops = {
684 .map_alloc = dev_map_alloc,
685 .map_free = dev_map_free,
686 .map_get_next_key = dev_map_hash_get_next_key,
687 .map_lookup_elem = dev_map_hash_lookup_elem,
688 .map_update_elem = dev_map_hash_update_elem,
689 .map_delete_elem = dev_map_hash_delete_elem,
690 .map_check_btf = map_check_no_btf,
693 static void dev_map_hash_remove_netdev(struct bpf_dtab *dtab,
694 struct net_device *netdev)
699 spin_lock_irqsave(&dtab->index_lock, flags);
700 for (i = 0; i < dtab->n_buckets; i++) {
701 struct bpf_dtab_netdev *dev;
702 struct hlist_head *head;
703 struct hlist_node *next;
705 head = dev_map_index_hash(dtab, i);
707 hlist_for_each_entry_safe(dev, next, head, index_hlist) {
708 if (netdev != dev->dev)
712 hlist_del_rcu(&dev->index_hlist);
713 call_rcu(&dev->rcu, __dev_map_entry_free);
716 spin_unlock_irqrestore(&dtab->index_lock, flags);
719 static int dev_map_notification(struct notifier_block *notifier,
720 ulong event, void *ptr)
722 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
723 struct bpf_dtab *dtab;
727 case NETDEV_UNREGISTER:
728 /* This rcu_read_lock/unlock pair is needed because
729 * dev_map_list is an RCU list AND to ensure a delete
730 * operation does not free a netdev_map entry while we
731 * are comparing it against the netdev being unregistered.
734 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
735 if (dtab->map.map_type == BPF_MAP_TYPE_DEVMAP_HASH) {
736 dev_map_hash_remove_netdev(dtab, netdev);
740 for (i = 0; i < dtab->map.max_entries; i++) {
741 struct bpf_dtab_netdev *dev, *odev;
743 dev = READ_ONCE(dtab->netdev_map[i]);
744 if (!dev || netdev != dev->dev)
746 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
749 __dev_map_entry_free);
760 static struct notifier_block dev_map_notifier = {
761 .notifier_call = dev_map_notification,
764 static int __init dev_map_init(void)
768 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
769 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
770 offsetof(struct _bpf_dtab_netdev, dev));
771 register_netdevice_notifier(&dev_map_notifier);
773 for_each_possible_cpu(cpu)
774 INIT_LIST_HEAD(&per_cpu(dev_map_flush_list, cpu));
778 subsys_initcall(dev_map_init);