projects/linux·blob v5.2-rc5 · refs

539 lines · 15.0 KiB

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 * an atomic per-cpu bitmap. The bpf_dtab_netdev object will not be destroyed
21 * until all bits are cleared indicating outstanding flush operations have
22 * completed.
23 *
24 * BPF syscalls may race with BPF program calls on any of the update, delete
25 * or lookup operations. As noted above the xchg() operation also keep the
26 * netdev_map consistent in this case. From the devmap side BPF programs
27 * calling into these operations are the same as multiple user space threads
28 * making system calls.
29 *
30 * Finally, any of the above may race with a netdev_unregister notifier. The
31 * unregister notifier must search for net devices in the map structure that
32 * contain a reference to the net device and remove them. This is a two step
33 * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
34 * check to see if the ifindex is the same as the net_device being removed.
35 * When removing the dev a cmpxchg() is used to ensure the correct dev is
36 * removed, in the case of a concurrent update or delete operation it is
37 * possible that the initially referenced dev is no longer in the map. As the
38 * notifier hook walks the map we know that new dev references can not be
39 * added by the user because core infrastructure ensures dev_get_by_index()
40 * calls will fail at this point.
41 */
42#include <linux/bpf.h>
43#include <net/xdp.h>
44#include <linux/filter.h>
45#include <trace/events/xdp.h>
46
47#define DEV_CREATE_FLAG_MASK \
48 (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
49
50#define DEV_MAP_BULK_SIZE 16
51struct xdp_bulk_queue {
52 struct xdp_frame *q[DEV_MAP_BULK_SIZE];
53 struct net_device *dev_rx;
54 unsigned int count;
55};
56
57struct bpf_dtab_netdev {
58 struct net_device *dev; /* must be first member, due to tracepoint */
59 struct bpf_dtab *dtab;
60 unsigned int bit;
61 struct xdp_bulk_queue __percpu *bulkq;
62 struct rcu_head rcu;
63};
64
65struct bpf_dtab {
66 struct bpf_map map;
67 struct bpf_dtab_netdev **netdev_map;
68 unsigned long __percpu *flush_needed;
69 struct list_head list;
70};
71
72static DEFINE_SPINLOCK(dev_map_lock);
73static LIST_HEAD(dev_map_list);
74
75static u64 dev_map_bitmap_size(const union bpf_attr *attr)
76{
77 return BITS_TO_LONGS((u64) attr->max_entries) * sizeof(unsigned long);
78}
79
80static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
81{
82 struct bpf_dtab *dtab;
83 int err = -EINVAL;
84 u64 cost;
85
86 if (!capable(CAP_NET_ADMIN))
87 return ERR_PTR(-EPERM);
88
89 /* check sanity of attributes */
90 if (attr->max_entries == 0 || attr->key_size != 4 ||
91 attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
92 return ERR_PTR(-EINVAL);
93
94 dtab = kzalloc(sizeof(*dtab), GFP_USER);
95 if (!dtab)
96 return ERR_PTR(-ENOMEM);
97
98 bpf_map_init_from_attr(&dtab->map, attr);
99
100 /* make sure page count doesn't overflow */
101 cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
102 cost += dev_map_bitmap_size(attr) * num_possible_cpus();
103 if (cost >= U32_MAX - PAGE_SIZE)
104 goto free_dtab;
105
106 dtab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
107
108 /* if map size is larger than memlock limit, reject it early */
109 err = bpf_map_precharge_memlock(dtab->map.pages);
110 if (err)
111 goto free_dtab;
112
113 err = -ENOMEM;
114
115 /* A per cpu bitfield with a bit per possible net device */
116 dtab->flush_needed = __alloc_percpu_gfp(dev_map_bitmap_size(attr),
117 __alignof__(unsigned long),
118 GFP_KERNEL | __GFP_NOWARN);
119 if (!dtab->flush_needed)
120 goto free_dtab;
121
122 dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
123 sizeof(struct bpf_dtab_netdev *),
124 dtab->map.numa_node);
125 if (!dtab->netdev_map)
126 goto free_dtab;
127
128 spin_lock(&dev_map_lock);
129 list_add_tail_rcu(&dtab->list, &dev_map_list);
130 spin_unlock(&dev_map_lock);
131
132 return &dtab->map;
133free_dtab:
134 free_percpu(dtab->flush_needed);
135 kfree(dtab);
136 return ERR_PTR(err);
137}
138
139static void dev_map_free(struct bpf_map *map)
140{
141 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
142 int i, cpu;
143
144 /* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
145 * so the programs (can be more than one that used this map) were
146 * disconnected from events. Wait for outstanding critical sections in
147 * these programs to complete. The rcu critical section only guarantees
148 * no further reads against netdev_map. It does __not__ ensure pending
149 * flush operations (if any) are complete.
150 */
151
152 spin_lock(&dev_map_lock);
153 list_del_rcu(&dtab->list);
154 spin_unlock(&dev_map_lock);
155
156 bpf_clear_redirect_map(map);
157 synchronize_rcu();
158
159 /* Make sure prior __dev_map_entry_free() have completed. */
160 rcu_barrier();
161
162 /* To ensure all pending flush operations have completed wait for flush
163 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
164 * Because the above synchronize_rcu() ensures the map is disconnected
165 * from the program we can assume no new bits will be set.
166 */
167 for_each_online_cpu(cpu) {
168 unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
169
170 while (!bitmap_empty(bitmap, dtab->map.max_entries))
171 cond_resched();
172 }
173
174 for (i = 0; i < dtab->map.max_entries; i++) {
175 struct bpf_dtab_netdev *dev;
176
177 dev = dtab->netdev_map[i];
178 if (!dev)
179 continue;
180
181 dev_put(dev->dev);
182 kfree(dev);
183 }
184
185 free_percpu(dtab->flush_needed);
186 bpf_map_area_free(dtab->netdev_map);
187 kfree(dtab);
188}
189
190static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
191{
192 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
193 u32 index = key ? *(u32 *)key : U32_MAX;
194 u32 *next = next_key;
195
196 if (index >= dtab->map.max_entries) {
197 *next = 0;
198 return 0;
199 }
200
201 if (index == dtab->map.max_entries - 1)
202 return -ENOENT;
203 *next = index + 1;
204 return 0;
205}
206
207void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
208{
209 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
210 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
211
212 __set_bit(bit, bitmap);
213}
214
215static int bq_xmit_all(struct bpf_dtab_netdev *obj,
216 struct xdp_bulk_queue *bq, u32 flags,
217 bool in_napi_ctx)
218{
219 struct net_device *dev = obj->dev;
220 int sent = 0, drops = 0, err = 0;
221 int i;
222
223 if (unlikely(!bq->count))
224 return 0;
225
226 for (i = 0; i < bq->count; i++) {
227 struct xdp_frame *xdpf = bq->q[i];
228
229 prefetch(xdpf);
230 }
231
232 sent = dev->netdev_ops->ndo_xdp_xmit(dev, bq->count, bq->q, flags);
233 if (sent < 0) {
234 err = sent;
235 sent = 0;
236 goto error;
237 }
238 drops = bq->count - sent;
239out:
240 bq->count = 0;
241
242 trace_xdp_devmap_xmit(&obj->dtab->map, obj->bit,
243 sent, drops, bq->dev_rx, dev, err);
244 bq->dev_rx = NULL;
245 return 0;
246error:
247 /* If ndo_xdp_xmit fails with an errno, no frames have been
248 * xmit'ed and it's our responsibility to them free all.
249 */
250 for (i = 0; i < bq->count; i++) {
251 struct xdp_frame *xdpf = bq->q[i];
252
253 /* RX path under NAPI protection, can return frames faster */
254 if (likely(in_napi_ctx))
255 xdp_return_frame_rx_napi(xdpf);
256 else
257 xdp_return_frame(xdpf);
258 drops++;
259 }
260 goto out;
261}
262
263/* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
264 * from the driver before returning from its napi->poll() routine. The poll()
265 * routine is called either from busy_poll context or net_rx_action signaled
266 * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
267 * net device can be torn down. On devmap tear down we ensure the ctx bitmap
268 * is zeroed before completing to ensure all flush operations have completed.
269 */
270void __dev_map_flush(struct bpf_map *map)
271{
272 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
273 unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
274 u32 bit;
275
276 for_each_set_bit(bit, bitmap, map->max_entries) {
277 struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
278 struct xdp_bulk_queue *bq;
279
280 /* This is possible if the dev entry is removed by user space
281 * between xdp redirect and flush op.
282 */
283 if (unlikely(!dev))
284 continue;
285
286 __clear_bit(bit, bitmap);
287
288 bq = this_cpu_ptr(dev->bulkq);
289 bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, true);
290 }
291}
292
293/* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
294 * update happens in parallel here a dev_put wont happen until after reading the
295 * ifindex.
296 */
297struct bpf_dtab_netdev *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
298{
299 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
300 struct bpf_dtab_netdev *obj;
301
302 if (key >= map->max_entries)
303 return NULL;
304
305 obj = READ_ONCE(dtab->netdev_map[key]);
306 return obj;
307}
308
309/* Runs under RCU-read-side, plus in softirq under NAPI protection.
310 * Thus, safe percpu variable access.
311 */
312static int bq_enqueue(struct bpf_dtab_netdev *obj, struct xdp_frame *xdpf,
313 struct net_device *dev_rx)
314
315{
316 struct xdp_bulk_queue *bq = this_cpu_ptr(obj->bulkq);
317
318 if (unlikely(bq->count == DEV_MAP_BULK_SIZE))
319 bq_xmit_all(obj, bq, 0, true);
320
321 /* Ingress dev_rx will be the same for all xdp_frame's in
322 * bulk_queue, because bq stored per-CPU and must be flushed
323 * from net_device drivers NAPI func end.
324 */
325 if (!bq->dev_rx)
326 bq->dev_rx = dev_rx;
327
328 bq->q[bq->count++] = xdpf;
329 return 0;
330}
331
332int dev_map_enqueue(struct bpf_dtab_netdev *dst, struct xdp_buff *xdp,
333 struct net_device *dev_rx)
334{
335 struct net_device *dev = dst->dev;
336 struct xdp_frame *xdpf;
337 int err;
338
339 if (!dev->netdev_ops->ndo_xdp_xmit)
340 return -EOPNOTSUPP;
341
342 err = xdp_ok_fwd_dev(dev, xdp->data_end - xdp->data);
343 if (unlikely(err))
344 return err;
345
346 xdpf = convert_to_xdp_frame(xdp);
347 if (unlikely(!xdpf))
348 return -EOVERFLOW;
349
350 return bq_enqueue(dst, xdpf, dev_rx);
351}
352
353int dev_map_generic_redirect(struct bpf_dtab_netdev *dst, struct sk_buff *skb,
354 struct bpf_prog *xdp_prog)
355{
356 int err;
357
358 err = xdp_ok_fwd_dev(dst->dev, skb->len);
359 if (unlikely(err))
360 return err;
361 skb->dev = dst->dev;
362 generic_xdp_tx(skb, xdp_prog);
363
364 return 0;
365}
366
367static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
368{
369 struct bpf_dtab_netdev *obj = __dev_map_lookup_elem(map, *(u32 *)key);
370 struct net_device *dev = obj ? obj->dev : NULL;
371
372 return dev ? &dev->ifindex : NULL;
373}
374
375static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
376{
377 if (dev->dev->netdev_ops->ndo_xdp_xmit) {
378 struct xdp_bulk_queue *bq;
379 unsigned long *bitmap;
380
381 int cpu;
382
383 for_each_online_cpu(cpu) {
384 bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
385 __clear_bit(dev->bit, bitmap);
386
387 bq = per_cpu_ptr(dev->bulkq, cpu);
388 bq_xmit_all(dev, bq, XDP_XMIT_FLUSH, false);
389 }
390 }
391}
392
393static void __dev_map_entry_free(struct rcu_head *rcu)
394{
395 struct bpf_dtab_netdev *dev;
396
397 dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
398 dev_map_flush_old(dev);
399 free_percpu(dev->bulkq);
400 dev_put(dev->dev);
401 kfree(dev);
402}
403
404static int dev_map_delete_elem(struct bpf_map *map, void *key)
405{
406 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
407 struct bpf_dtab_netdev *old_dev;
408 int k = *(u32 *)key;
409
410 if (k >= map->max_entries)
411 return -EINVAL;
412
413 /* Use call_rcu() here to ensure any rcu critical sections have
414 * completed, but this does not guarantee a flush has happened
415 * yet. Because driver side rcu_read_lock/unlock only protects the
416 * running XDP program. However, for pending flush operations the
417 * dev and ctx are stored in another per cpu map. And additionally,
418 * the driver tear down ensures all soft irqs are complete before
419 * removing the net device in the case of dev_put equals zero.
420 */
421 old_dev = xchg(&dtab->netdev_map[k], NULL);
422 if (old_dev)
423 call_rcu(&old_dev->rcu, __dev_map_entry_free);
424 return 0;
425}
426
427static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
428 u64 map_flags)
429{
430 struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
431 struct net *net = current->nsproxy->net_ns;
432 gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
433 struct bpf_dtab_netdev *dev, *old_dev;
434 u32 i = *(u32 *)key;
435 u32 ifindex = *(u32 *)value;
436
437 if (unlikely(map_flags > BPF_EXIST))
438 return -EINVAL;
439 if (unlikely(i >= dtab->map.max_entries))
440 return -E2BIG;
441 if (unlikely(map_flags == BPF_NOEXIST))
442 return -EEXIST;
443
444 if (!ifindex) {
445 dev = NULL;
446 } else {
447 dev = kmalloc_node(sizeof(*dev), gfp, map->numa_node);
448 if (!dev)
449 return -ENOMEM;
450
451 dev->bulkq = __alloc_percpu_gfp(sizeof(*dev->bulkq),
452 sizeof(void *), gfp);
453 if (!dev->bulkq) {
454 kfree(dev);
455 return -ENOMEM;
456 }
457
458 dev->dev = dev_get_by_index(net, ifindex);
459 if (!dev->dev) {
460 free_percpu(dev->bulkq);
461 kfree(dev);
462 return -EINVAL;
463 }
464
465 dev->bit = i;
466 dev->dtab = dtab;
467 }
468
469 /* Use call_rcu() here to ensure rcu critical sections have completed
470 * Remembering the driver side flush operation will happen before the
471 * net device is removed.
472 */
473 old_dev = xchg(&dtab->netdev_map[i], dev);
474 if (old_dev)
475 call_rcu(&old_dev->rcu, __dev_map_entry_free);
476
477 return 0;
478}
479
480const struct bpf_map_ops dev_map_ops = {
481 .map_alloc = dev_map_alloc,
482 .map_free = dev_map_free,
483 .map_get_next_key = dev_map_get_next_key,
484 .map_lookup_elem = dev_map_lookup_elem,
485 .map_update_elem = dev_map_update_elem,
486 .map_delete_elem = dev_map_delete_elem,
487 .map_check_btf = map_check_no_btf,
488};
489
490static int dev_map_notification(struct notifier_block *notifier,
491 ulong event, void *ptr)
492{
493 struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
494 struct bpf_dtab *dtab;
495 int i;
496
497 switch (event) {
498 case NETDEV_UNREGISTER:
499 /* This rcu_read_lock/unlock pair is needed because
500 * dev_map_list is an RCU list AND to ensure a delete
501 * operation does not free a netdev_map entry while we
502 * are comparing it against the netdev being unregistered.
503 */
504 rcu_read_lock();
505 list_for_each_entry_rcu(dtab, &dev_map_list, list) {
506 for (i = 0; i < dtab->map.max_entries; i++) {
507 struct bpf_dtab_netdev *dev, *odev;
508
509 dev = READ_ONCE(dtab->netdev_map[i]);
510 if (!dev || netdev != dev->dev)
511 continue;
512 odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
513 if (dev == odev)
514 call_rcu(&dev->rcu,
515 __dev_map_entry_free);
516 }
517 }
518 rcu_read_unlock();
519 break;
520 default:
521 break;
522 }
523 return NOTIFY_OK;
524}
525
526static struct notifier_block dev_map_notifier = {
527 .notifier_call = dev_map_notification,
528};
529
530static int __init dev_map_init(void)
531{
532 /* Assure tracepoint shadow struct _bpf_dtab_netdev is in sync */
533 BUILD_BUG_ON(offsetof(struct bpf_dtab_netdev, dev) !=
534 offsetof(struct _bpf_dtab_netdev, dev));
535 register_netdevice_notifier(&dev_map_notifier);
536 return 0;
537}
538
539subsys_initcall(dev_map_init);