projects/linux·blob v4.18 · refs

543 lines · 15.3 KiB

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 <net/xdp.h>
52#include <linux/filter.h>
53#include <trace/events/xdp.h>
54
55#define DEV_CREATE_FLAG_MASK \
56 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
57
58#define DEV_MAP_BULK_SIZE 16
59struct xdp_bulk_queue {
60 struct xdp_frame *q[DEV_MAP_BULK_SIZE];
61 struct net_device *dev_rx;
62 unsigned int count;
63};
64
65struct bpf_dtab_netdev {
66 struct net_device *dev; /* must be first member, due to tracepoint */
67 struct bpf_dtab *dtab;
68 unsigned int bit;
69 struct xdp_bulk_queue __percpu *bulkq;
70 struct rcu_head rcu;
71};
72
73struct bpf_dtab {
74 struct bpf_map map;
75 struct bpf_dtab_netdev **netdev_map;
76 unsigned long __percpu *flush_needed;
77 struct list_head list;
78};
79
80static DEFINE_SPINLOCK(dev_map_lock);
81static LIST_HEAD(dev_map_list);
82
83static u64 dev_map_bitmap_size(const union bpf_attr *attr)
84{
85 return BITS_TO_LONGS((u64) attr->max_entries) * sizeof(unsigned long);
86}
87
88static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
89{
90 struct bpf_dtab *dtab;
91 int err = -EINVAL;
92 u64 cost;
93
94 if (!capable(CAP_NET_ADMIN))
95 return ERR_PTR(-EPERM);
96
97 /* check sanity of attributes */
98 if (attr->max_entries == 0 || attr->key_size != 4 ||
99 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
100 return ERR_PTR(-EINVAL);
101
102 dtab = kzalloc(sizeof(*dtab), GFP_USER);
103 if (!dtab)
104 return ERR_PTR(-ENOMEM);
105
106 bpf_map_init_from_attr(&dtab->map, attr);
107
108 /* make sure page count doesn't overflow */
109 cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
110 cost += dev_map_bitmap_size(attr) * num_possible_cpus();
111 if (cost >= U32_MAX - PAGE_SIZE)
112 goto free_dtab;
113
114 dtab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
115
116 /* if map size is larger than memlock limit, reject it early */
117 err = bpf_map_precharge_memlock(dtab->map.pages);
118 if (err)
119 goto free_dtab;
120
121 err = -ENOMEM;
122
123 /* A per cpu bitfield with a bit per possible net device */
124 dtab->flush_needed = __alloc_percpu_gfp(dev_map_bitmap_size(attr),
125 __alignof__(unsigned long),
126 GFP_KERNEL | __GFP_NOWARN);
127 if (!dtab->flush_needed)
128 goto free_dtab;
129
130 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
131 sizeof(struct bpf_dtab_netdev *),
132 dtab->map.numa_node);
133 if (!dtab->netdev_map)
134 goto free_dtab;
135
136 spin_lock(&dev_map_lock);
137 list_add_tail_rcu(&dtab->list, &dev_map_list);
138 spin_unlock(&dev_map_lock);
139
140 return &dtab->map;
141free_dtab:
142 free_percpu(dtab->flush_needed);
143 kfree(dtab);
144 return ERR_PTR(err);
145}
146
147static void dev_map_free(struct bpf_map *map)
148{
149 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
150 int i, cpu;
151
152 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
153 * so the programs (can be more than one that used this map) were
154 * disconnected from events. Wait for outstanding critical sections in
155 * these programs to complete. The rcu critical section only guarantees
156 * no further reads against netdev_map. It does __not__ ensure pending
157 * flush operations (if any) are complete.
158 */
159
160 spin_lock(&dev_map_lock);
161 list_del_rcu(&dtab->list);
162 spin_unlock(&dev_map_lock);
163
164 synchronize_rcu();
165
166 /* To ensure all pending flush operations have completed wait for flush
167 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
168 * Because the above synchronize_rcu() ensures the map is disconnected
169 * from the program we can assume no new bits will be set.
170 */
171 for_each_online_cpu(cpu) {
172 unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
173
174 while (!bitmap_empty(bitmap, dtab->map.max_entries))
175 cond_resched();
176 }
177
178 for (i = 0; i < dtab->map.max_entries; i++) {
179 struct bpf_dtab_netdev *dev;
180
181 dev = dtab->netdev_map[i];
182 if (!dev)
183 continue;
184
185 dev_put(dev->dev);
186 kfree(dev);
187 }
188
189 free_percpu(dtab->flush_needed);
190 bpf_map_area_free(dtab->netdev_map);
191 kfree(dtab);
192}
193
194static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
195{
196 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
197 u32 index = key ? *(u32 *)key : U32_MAX;
198 u32 *next = next_key;
199
200 if (index >= dtab->map.max_entries) {
201 *next = 0;
202 return 0;
203 }
204
205 if (index == dtab->map.max_entries - 1)
206 return -ENOENT;
207 *next = index + 1;
208 return 0;
209}
210
211void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
212{
213 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
214 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
215
216 __set_bit(bit, bitmap);
217}
218
219static int bq_xmit_all(struct bpf_dtab_netdev *obj,
220 struct xdp_bulk_queue *bq, u32 flags,
221 bool in_napi_ctx)
222{
223 struct net_device *dev = obj->dev;
224 int sent = 0, drops = 0, err = 0;
225 int i;
226
227 if (unlikely(!bq->count))
228 return 0;
229
230 for (i = 0; i < bq->count; i++) {
231 struct xdp_frame *xdpf = bq->q[i];
232
233 prefetch(xdpf);
234 }
235
236 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
237 if (sent < 0) {
238 err = sent;
239 sent = 0;
240 goto error;
241 }
242 drops = bq->count - sent;
243out:
244 bq->count = 0;
245
246 trace_xdp_devmap_xmit(&obj->dtab->map, obj->bit,
247 sent, drops, bq->dev_rx, dev, err);
248 bq->dev_rx = NULL;
249 return 0;
250error:
251 /* If ndo_xdp_xmit fails with an errno, no frames have been
252 * xmit'ed and it's our responsibility to them free all.
253 */
254 for (i = 0; i < bq->count; i++) {
255 struct xdp_frame *xdpf = bq->q[i];
256
257 /* RX path under NAPI protection, can return frames faster */
258 if (likely(in_napi_ctx))
259 xdp_return_frame_rx_napi(xdpf);
260 else
261 xdp_return_frame(xdpf);
262 drops++;
263 }
264 goto out;
265}
266
267/* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
268 * from the driver before returning from its napi->poll() routine. The poll()
269 * routine is called either from busy_poll context or net_rx_action signaled
270 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
271 * net device can be torn down. On devmap tear down we ensure the ctx bitmap
272 * is zeroed before completing to ensure all flush operations have completed.
273 */
274void __dev_map_flush(struct bpf_map *map)
275{
276 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
277 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
278 u32 bit;
279
280 for_each_set_bit(bit, bitmap, map->max_entries) {
281 struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
282 struct xdp_bulk_queue *bq;
283
284 /* This is possible if the dev entry is removed by user space
285 * between xdp redirect and flush op.
286 */
287 if (unlikely(!dev))
288 continue;
289
290 __clear_bit(bit, bitmap);
291
292 bq = this_cpu_ptr(dev->bulkq);
293 bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, true);
294 }
295}
296
297/* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
298 * update happens in parallel here a dev_put wont happen until after reading the
299 * ifindex.
300 */
301struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
302{
303 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
304 struct bpf_dtab_netdev *obj;
305
306 if (key >= map->max_entries)
307 return NULL;
308
309 obj = READ_ONCE(dtab->netdev_map[key]);
310 return obj;
311}
312
313/* Runs under RCU-read-side, plus in softirq under NAPI protection.
314 * Thus, safe percpu variable access.
315 */
316static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
317 struct net_device *dev_rx)
318
319{
320 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
321
322 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
323 bq_xmit_all(obj, bq, 0, true);
324
325 /* Ingress dev_rx will be the same for all xdp_frame's in
326 * bulk_queue, because bq stored per-CPU and must be flushed
327 * from net_device drivers NAPI func end.
328 */
329 if (!bq->dev_rx)
330 bq->dev_rx = dev_rx;
331
332 bq->q[bq->count++] = xdpf;
333 return 0;
334}
335
336int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
337 struct net_device *dev_rx)
338{
339 struct net_device *dev = dst->dev;
340 struct xdp_frame *xdpf;
341 int err;
342
343 if (!dev->netdev_ops->ndo_xdp_xmit)
344 return -EOPNOTSUPP;
345
346 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
347 if (unlikely(err))
348 return err;
349
350 xdpf = convert_to_xdp_frame(xdp);
351 if (unlikely(!xdpf))
352 return -EOVERFLOW;
353
354 return bq_enqueue(dst, xdpf, dev_rx);
355}
356
357int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
358 struct bpf_prog *xdp_prog)
359{
360 int err;
361
362 err = xdp_ok_fwd_dev(dst->dev, skb->len);
363 if (unlikely(err))
364 return err;
365 skb->dev = dst->dev;
366 generic_xdp_tx(skb, xdp_prog);
367
368 return 0;
369}
370
371static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
372{
373 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
374 struct net_device *dev = obj ? obj->dev : NULL;
375
376 return dev ? &dev->ifindex : NULL;
377}
378
379static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
380{
381 if (dev->dev->netdev_ops->ndo_xdp_xmit) {
382 struct xdp_bulk_queue *bq;
383 unsigned long *bitmap;
384
385 int cpu;
386
387 for_each_online_cpu(cpu) {
388 bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
389 __clear_bit(dev->bit, bitmap);
390
391 bq = per_cpu_ptr(dev->bulkq, cpu);
392 bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, false);
393 }
394 }
395}
396
397static void __dev_map_entry_free(struct rcu_head *rcu)
398{
399 struct bpf_dtab_netdev *dev;
400
401 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
402 dev_map_flush_old(dev);
403 free_percpu(dev->bulkq);
404 dev_put(dev->dev);
405 kfree(dev);
406}
407
408static int dev_map_delete_elem(struct bpf_map *map, void *key)
409{
410 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
411 struct bpf_dtab_netdev *old_dev;
412 int k = *(u32 *)key;
413
414 if (k >= map->max_entries)
415 return -EINVAL;
416
417 /* Use call_rcu() here to ensure any rcu critical sections have
418 * completed, but this does not guarantee a flush has happened
419 * yet. Because driver side rcu_read_lock/unlock only protects the
420 * running XDP program. However, for pending flush operations the
421 * dev and ctx are stored in another per cpu map. And additionally,
422 * the driver tear down ensures all soft irqs are complete before
423 * removing the net device in the case of dev_put equals zero.
424 */
425 old_dev = xchg(&dtab->netdev_map[k], NULL);
426 if (old_dev)
427 call_rcu(&old_dev->rcu, __dev_map_entry_free);
428 return 0;
429}
430
431static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
432 u64 map_flags)
433{
434 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
435 struct net *net = current->nsproxy->net_ns;
436 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
437 struct bpf_dtab_netdev *dev, *old_dev;
438 u32 i = *(u32 *)key;
439 u32 ifindex = *(u32 *)value;
440
441 if (unlikely(map_flags > BPF_EXIST))
442 return -EINVAL;
443 if (unlikely(i >= dtab->map.max_entries))
444 return -E2BIG;
445 if (unlikely(map_flags == BPF_NOEXIST))
446 return -EEXIST;
447
448 if (!ifindex) {
449 dev = NULL;
450 } else {
451 dev = kmalloc_node(sizeof(*dev), gfp, map->numa_node);
452 if (!dev)
453 return -ENOMEM;
454
455 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
456 sizeof(void *), gfp);
457 if (!dev->bulkq) {
458 kfree(dev);
459 return -ENOMEM;
460 }
461
462 dev->dev = dev_get_by_index(net, ifindex);
463 if (!dev->dev) {
464 free_percpu(dev->bulkq);
465 kfree(dev);
466 return -EINVAL;
467 }
468
469 dev->bit = i;
470 dev->dtab = dtab;
471 }
472
473 /* Use call_rcu() here to ensure rcu critical sections have completed
474 * Remembering the driver side flush operation will happen before the
475 * net device is removed.
476 */
477 old_dev = xchg(&dtab->netdev_map[i], dev);
478 if (old_dev)
479 call_rcu(&old_dev->rcu, __dev_map_entry_free);
480
481 return 0;
482}
483
484const struct bpf_map_ops dev_map_ops = {
485 .map_alloc = dev_map_alloc,
486 .map_free = dev_map_free,
487 .map_get_next_key = dev_map_get_next_key,
488 .map_lookup_elem = dev_map_lookup_elem,
489 .map_update_elem = dev_map_update_elem,
490 .map_delete_elem = dev_map_delete_elem,
491};
492
493static int dev_map_notification(struct notifier_block *notifier,
494 ulong event, void *ptr)
495{
496 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
497 struct bpf_dtab *dtab;
498 int i;
499
500 switch (event) {
501 case NETDEV_UNREGISTER:
502 /* This rcu_read_lock/unlock pair is needed because
503 * dev_map_list is an RCU list AND to ensure a delete
504 * operation does not free a netdev_map entry while we
505 * are comparing it against the netdev being unregistered.
506 */
507 rcu_read_lock();
508 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
509 for (i = 0; i < dtab->map.max_entries; i++) {
510 struct bpf_dtab_netdev *dev, *odev;
511
512 dev = READ_ONCE(dtab->netdev_map[i]);
513 if (!dev ||
514 dev->dev->ifindex != netdev->ifindex)
515 continue;
516 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
517 if (dev == odev)
518 call_rcu(&dev->rcu,
519 __dev_map_entry_free);
520 }
521 }
522 rcu_read_unlock();
523 break;
524 default:
525 break;
526 }
527 return NOTIFY_OK;
528}
529
530static struct notifier_block dev_map_notifier = {
531 .notifier_call = dev_map_notification,
532};
533
534static int __init dev_map_init(void)
535{
536 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
537 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
538 offsetof(struct _bpf_dtab_netdev, dev));
539 register_netdevice_notifier(&dev_map_notifier);
540 return 0;
541}
542
543subsys_initcall(dev_map_init);