Loading...
1/* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
2 *
3 * This program is free software; you can redistribute it and/or
4 * modify it under the terms of version 2 of the GNU General Public
5 * License as published by the Free Software Foundation.
6 *
7 * This program is distributed in the hope that it will be useful, but
8 * WITHOUT ANY WARRANTY; without even the implied warranty of
9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10 * General Public License for more details.
11 */
12
13/* Devmaps primary use is as a backend map for XDP BPF helper call
14 * bpf_redirect_map(). Because XDP is mostly concerned with performance we
15 * spent some effort to ensure the datapath with redirect maps does not use
16 * any locking. This is a quick note on the details.
17 *
18 * We have three possible paths to get into the devmap control plane bpf
19 * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
20 * will invoke an update, delete, or lookup operation. To ensure updates and
21 * deletes appear atomic from the datapath side xchg() is used to modify the
22 * netdev_map array. Then because the datapath does a lookup into the netdev_map
23 * array (read-only) from an RCU critical section we use call_rcu() to wait for
24 * an rcu grace period before free'ing the old data structures. This ensures the
25 * datapath always has a valid copy. However, the datapath does a "flush"
26 * operation that pushes any pending packets in the driver outside the RCU
27 * critical section. Each bpf_dtab_netdev tracks these pending operations using
28 * an atomic per-cpu bitmap. The bpf_dtab_netdev object will not be destroyed
29 * until all bits are cleared indicating outstanding flush operations have
30 * completed.
31 *
32 * BPF syscalls may race with BPF program calls on any of the update, delete
33 * or lookup operations. As noted above the xchg() operation also keep the
34 * netdev_map consistent in this case. From the devmap side BPF programs
35 * calling into these operations are the same as multiple user space threads
36 * making system calls.
37 *
38 * Finally, any of the above may race with a netdev_unregister notifier. The
39 * unregister notifier must search for net devices in the map structure that
40 * contain a reference to the net device and remove them. This is a two step
41 * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
42 * check to see if the ifindex is the same as the net_device being removed.
43 * When removing the dev a cmpxchg() is used to ensure the correct dev is
44 * removed, in the case of a concurrent update or delete operation it is
45 * possible that the initially referenced dev is no longer in the map. As the
46 * notifier hook walks the map we know that new dev references can not be
47 * added by the user because core infrastructure ensures dev_get_by_index()
48 * calls will fail at this point.
49 */
50#include <linux/bpf.h>
51#include <linux/filter.h>
52
53#define DEV_CREATE_FLAG_MASK \
54 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
55
56struct bpf_dtab_netdev {
57 struct net_device *dev;
58 struct bpf_dtab *dtab;
59 unsigned int bit;
60 struct rcu_head rcu;
61};
62
63struct bpf_dtab {
64 struct bpf_map map;
65 struct bpf_dtab_netdev **netdev_map;
66 unsigned long __percpu *flush_needed;
67 struct list_head list;
68};
69
70static DEFINE_SPINLOCK(dev_map_lock);
71static LIST_HEAD(dev_map_list);
72
73static u64 dev_map_bitmap_size(const union bpf_attr *attr)
74{
75 return BITS_TO_LONGS((u64) attr->max_entries) * sizeof(unsigned long);
76}
77
78static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
79{
80 struct bpf_dtab *dtab;
81 int err = -EINVAL;
82 u64 cost;
83
84 if (!capable(CAP_NET_ADMIN))
85 return ERR_PTR(-EPERM);
86
87 /* check sanity of attributes */
88 if (attr->max_entries == 0 || attr->key_size != 4 ||
89 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
90 return ERR_PTR(-EINVAL);
91
92 dtab = kzalloc(sizeof(*dtab), GFP_USER);
93 if (!dtab)
94 return ERR_PTR(-ENOMEM);
95
96 bpf_map_init_from_attr(&dtab->map, attr);
97
98 /* make sure page count doesn't overflow */
99 cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
100 cost += dev_map_bitmap_size(attr) * num_possible_cpus();
101 if (cost >= U32_MAX - PAGE_SIZE)
102 goto free_dtab;
103
104 dtab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
105
106 /* if map size is larger than memlock limit, reject it early */
107 err = bpf_map_precharge_memlock(dtab->map.pages);
108 if (err)
109 goto free_dtab;
110
111 err = -ENOMEM;
112
113 /* A per cpu bitfield with a bit per possible net device */
114 dtab->flush_needed = __alloc_percpu_gfp(dev_map_bitmap_size(attr),
115 __alignof__(unsigned long),
116 GFP_KERNEL | __GFP_NOWARN);
117 if (!dtab->flush_needed)
118 goto free_dtab;
119
120 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
121 sizeof(struct bpf_dtab_netdev *),
122 dtab->map.numa_node);
123 if (!dtab->netdev_map)
124 goto free_dtab;
125
126 spin_lock(&dev_map_lock);
127 list_add_tail_rcu(&dtab->list, &dev_map_list);
128 spin_unlock(&dev_map_lock);
129
130 return &dtab->map;
131free_dtab:
132 free_percpu(dtab->flush_needed);
133 kfree(dtab);
134 return ERR_PTR(err);
135}
136
137static void dev_map_free(struct bpf_map *map)
138{
139 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
140 int i, cpu;
141
142 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
143 * so the programs (can be more than one that used this map) were
144 * disconnected from events. Wait for outstanding critical sections in
145 * these programs to complete. The rcu critical section only guarantees
146 * no further reads against netdev_map. It does __not__ ensure pending
147 * flush operations (if any) are complete.
148 */
149
150 spin_lock(&dev_map_lock);
151 list_del_rcu(&dtab->list);
152 spin_unlock(&dev_map_lock);
153
154 synchronize_rcu();
155
156 /* To ensure all pending flush operations have completed wait for flush
157 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
158 * Because the above synchronize_rcu() ensures the map is disconnected
159 * from the program we can assume no new bits will be set.
160 */
161 for_each_online_cpu(cpu) {
162 unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
163
164 while (!bitmap_empty(bitmap, dtab->map.max_entries))
165 cond_resched();
166 }
167
168 for (i = 0; i < dtab->map.max_entries; i++) {
169 struct bpf_dtab_netdev *dev;
170
171 dev = dtab->netdev_map[i];
172 if (!dev)
173 continue;
174
175 dev_put(dev->dev);
176 kfree(dev);
177 }
178
179 free_percpu(dtab->flush_needed);
180 bpf_map_area_free(dtab->netdev_map);
181 kfree(dtab);
182}
183
184static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
185{
186 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
187 u32 index = key ? *(u32 *)key : U32_MAX;
188 u32 *next = next_key;
189
190 if (index >= dtab->map.max_entries) {
191 *next = 0;
192 return 0;
193 }
194
195 if (index == dtab->map.max_entries - 1)
196 return -ENOENT;
197 *next = index + 1;
198 return 0;
199}
200
201void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
202{
203 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
204 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
205
206 __set_bit(bit, bitmap);
207}
208
209/* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
210 * from the driver before returning from its napi->poll() routine. The poll()
211 * routine is called either from busy_poll context or net_rx_action signaled
212 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
213 * net device can be torn down. On devmap tear down we ensure the ctx bitmap
214 * is zeroed before completing to ensure all flush operations have completed.
215 */
216void __dev_map_flush(struct bpf_map *map)
217{
218 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
219 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
220 u32 bit;
221
222 for_each_set_bit(bit, bitmap, map->max_entries) {
223 struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
224 struct net_device *netdev;
225
226 /* This is possible if the dev entry is removed by user space
227 * between xdp redirect and flush op.
228 */
229 if (unlikely(!dev))
230 continue;
231
232 __clear_bit(bit, bitmap);
233 netdev = dev->dev;
234 if (likely(netdev->netdev_ops->ndo_xdp_flush))
235 netdev->netdev_ops->ndo_xdp_flush(netdev);
236 }
237}
238
239/* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
240 * update happens in parallel here a dev_put wont happen until after reading the
241 * ifindex.
242 */
243struct net_device *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
244{
245 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
246 struct bpf_dtab_netdev *dev;
247
248 if (key >= map->max_entries)
249 return NULL;
250
251 dev = READ_ONCE(dtab->netdev_map[key]);
252 return dev ? dev->dev : NULL;
253}
254
255static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
256{
257 struct net_device *dev = __dev_map_lookup_elem(map, *(u32 *)key);
258
259 return dev ? &dev->ifindex : NULL;
260}
261
262static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
263{
264 if (dev->dev->netdev_ops->ndo_xdp_flush) {
265 struct net_device *fl = dev->dev;
266 unsigned long *bitmap;
267 int cpu;
268
269 for_each_online_cpu(cpu) {
270 bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
271 __clear_bit(dev->bit, bitmap);
272
273 fl->netdev_ops->ndo_xdp_flush(dev->dev);
274 }
275 }
276}
277
278static void __dev_map_entry_free(struct rcu_head *rcu)
279{
280 struct bpf_dtab_netdev *dev;
281
282 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
283 dev_map_flush_old(dev);
284 dev_put(dev->dev);
285 kfree(dev);
286}
287
288static int dev_map_delete_elem(struct bpf_map *map, void *key)
289{
290 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
291 struct bpf_dtab_netdev *old_dev;
292 int k = *(u32 *)key;
293
294 if (k >= map->max_entries)
295 return -EINVAL;
296
297 /* Use call_rcu() here to ensure any rcu critical sections have
298 * completed, but this does not guarantee a flush has happened
299 * yet. Because driver side rcu_read_lock/unlock only protects the
300 * running XDP program. However, for pending flush operations the
301 * dev and ctx are stored in another per cpu map. And additionally,
302 * the driver tear down ensures all soft irqs are complete before
303 * removing the net device in the case of dev_put equals zero.
304 */
305 old_dev = xchg(&dtab->netdev_map[k], NULL);
306 if (old_dev)
307 call_rcu(&old_dev->rcu, __dev_map_entry_free);
308 return 0;
309}
310
311static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
312 u64 map_flags)
313{
314 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
315 struct net *net = current->nsproxy->net_ns;
316 struct bpf_dtab_netdev *dev, *old_dev;
317 u32 i = *(u32 *)key;
318 u32 ifindex = *(u32 *)value;
319
320 if (unlikely(map_flags > BPF_EXIST))
321 return -EINVAL;
322 if (unlikely(i >= dtab->map.max_entries))
323 return -E2BIG;
324 if (unlikely(map_flags == BPF_NOEXIST))
325 return -EEXIST;
326
327 if (!ifindex) {
328 dev = NULL;
329 } else {
330 dev = kmalloc_node(sizeof(*dev), GFP_ATOMIC | __GFP_NOWARN,
331 map->numa_node);
332 if (!dev)
333 return -ENOMEM;
334
335 dev->dev = dev_get_by_index(net, ifindex);
336 if (!dev->dev) {
337 kfree(dev);
338 return -EINVAL;
339 }
340
341 dev->bit = i;
342 dev->dtab = dtab;
343 }
344
345 /* Use call_rcu() here to ensure rcu critical sections have completed
346 * Remembering the driver side flush operation will happen before the
347 * net device is removed.
348 */
349 old_dev = xchg(&dtab->netdev_map[i], dev);
350 if (old_dev)
351 call_rcu(&old_dev->rcu, __dev_map_entry_free);
352
353 return 0;
354}
355
356const struct bpf_map_ops dev_map_ops = {
357 .map_alloc = dev_map_alloc,
358 .map_free = dev_map_free,
359 .map_get_next_key = dev_map_get_next_key,
360 .map_lookup_elem = dev_map_lookup_elem,
361 .map_update_elem = dev_map_update_elem,
362 .map_delete_elem = dev_map_delete_elem,
363};
364
365static int dev_map_notification(struct notifier_block *notifier,
366 ulong event, void *ptr)
367{
368 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
369 struct bpf_dtab *dtab;
370 int i;
371
372 switch (event) {
373 case NETDEV_UNREGISTER:
374 /* This rcu_read_lock/unlock pair is needed because
375 * dev_map_list is an RCU list AND to ensure a delete
376 * operation does not free a netdev_map entry while we
377 * are comparing it against the netdev being unregistered.
378 */
379 rcu_read_lock();
380 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
381 for (i = 0; i < dtab->map.max_entries; i++) {
382 struct bpf_dtab_netdev *dev, *odev;
383
384 dev = READ_ONCE(dtab->netdev_map[i]);
385 if (!dev ||
386 dev->dev->ifindex != netdev->ifindex)
387 continue;
388 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
389 if (dev == odev)
390 call_rcu(&dev->rcu,
391 __dev_map_entry_free);
392 }
393 }
394 rcu_read_unlock();
395 break;
396 default:
397 break;
398 }
399 return NOTIFY_OK;
400}
401
402static struct notifier_block dev_map_notifier = {
403 .notifier_call = dev_map_notification,
404};
405
406static int __init dev_map_init(void)
407{
408 register_netdevice_notifier(&dev_map_notifier);
409 return 0;
410}
411
412subsys_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);