]>
Commit | Line | Data |
---|---|---|
174a79ff JF |
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 | /* A BPF sock_map is used to store sock objects. This is primarly used | |
14 | * for doing socket redirect with BPF helper routines. | |
15 | * | |
2f857d04 JF |
16 | * A sock map may have BPF programs attached to it, currently a program |
17 | * used to parse packets and a program to provide a verdict and redirect | |
18 | * decision on the packet are supported. Any programs attached to a sock | |
19 | * map are inherited by sock objects when they are added to the map. If | |
20 | * no BPF programs are attached the sock object may only be used for sock | |
21 | * redirect. | |
22 | * | |
23 | * A sock object may be in multiple maps, but can only inherit a single | |
24 | * parse or verdict program. If adding a sock object to a map would result | |
25 | * in having multiple parsing programs the update will return an EBUSY error. | |
174a79ff JF |
26 | * |
27 | * For reference this program is similar to devmap used in XDP context | |
28 | * reviewing these together may be useful. For an example please review | |
29 | * ./samples/bpf/sockmap/. | |
30 | */ | |
31 | #include <linux/bpf.h> | |
32 | #include <net/sock.h> | |
33 | #include <linux/filter.h> | |
34 | #include <linux/errno.h> | |
35 | #include <linux/file.h> | |
36 | #include <linux/kernel.h> | |
37 | #include <linux/net.h> | |
38 | #include <linux/skbuff.h> | |
39 | #include <linux/workqueue.h> | |
40 | #include <linux/list.h> | |
41 | #include <net/strparser.h> | |
42 | ||
43 | struct bpf_stab { | |
44 | struct bpf_map map; | |
45 | struct sock **sock_map; | |
46 | struct bpf_prog *bpf_parse; | |
47 | struct bpf_prog *bpf_verdict; | |
174a79ff JF |
48 | }; |
49 | ||
50 | enum smap_psock_state { | |
51 | SMAP_TX_RUNNING, | |
52 | }; | |
53 | ||
2f857d04 JF |
54 | struct smap_psock_map_entry { |
55 | struct list_head list; | |
56 | struct sock **entry; | |
57 | }; | |
58 | ||
174a79ff JF |
59 | struct smap_psock { |
60 | struct rcu_head rcu; | |
2f857d04 JF |
61 | /* refcnt is used inside sk_callback_lock */ |
62 | u32 refcnt; | |
174a79ff JF |
63 | |
64 | /* datapath variables */ | |
65 | struct sk_buff_head rxqueue; | |
66 | bool strp_enabled; | |
67 | ||
68 | /* datapath error path cache across tx work invocations */ | |
69 | int save_rem; | |
70 | int save_off; | |
71 | struct sk_buff *save_skb; | |
72 | ||
73 | struct strparser strp; | |
74 | struct bpf_prog *bpf_parse; | |
75 | struct bpf_prog *bpf_verdict; | |
2f857d04 | 76 | struct list_head maps; |
174a79ff JF |
77 | |
78 | /* Back reference used when sock callback trigger sockmap operations */ | |
174a79ff JF |
79 | struct sock *sock; |
80 | unsigned long state; | |
81 | ||
82 | struct work_struct tx_work; | |
83 | struct work_struct gc_work; | |
84 | ||
85 | void (*save_data_ready)(struct sock *sk); | |
86 | void (*save_write_space)(struct sock *sk); | |
87 | void (*save_state_change)(struct sock *sk); | |
88 | }; | |
89 | ||
90 | static inline struct smap_psock *smap_psock_sk(const struct sock *sk) | |
91 | { | |
2f857d04 | 92 | return rcu_dereference_sk_user_data(sk); |
174a79ff JF |
93 | } |
94 | ||
95 | static int smap_verdict_func(struct smap_psock *psock, struct sk_buff *skb) | |
96 | { | |
97 | struct bpf_prog *prog = READ_ONCE(psock->bpf_verdict); | |
98 | int rc; | |
99 | ||
100 | if (unlikely(!prog)) | |
101 | return SK_DROP; | |
102 | ||
103 | skb_orphan(skb); | |
104 | skb->sk = psock->sock; | |
105 | bpf_compute_data_end(skb); | |
106 | rc = (*prog->bpf_func)(skb, prog->insnsi); | |
107 | skb->sk = NULL; | |
108 | ||
109 | return rc; | |
110 | } | |
111 | ||
112 | static void smap_do_verdict(struct smap_psock *psock, struct sk_buff *skb) | |
113 | { | |
114 | struct sock *sock; | |
115 | int rc; | |
116 | ||
117 | /* Because we use per cpu values to feed input from sock redirect | |
118 | * in BPF program to do_sk_redirect_map() call we need to ensure we | |
119 | * are not preempted. RCU read lock is not sufficient in this case | |
120 | * with CONFIG_PREEMPT_RCU enabled so we must be explicit here. | |
121 | */ | |
122 | preempt_disable(); | |
123 | rc = smap_verdict_func(psock, skb); | |
124 | switch (rc) { | |
125 | case SK_REDIRECT: | |
126 | sock = do_sk_redirect_map(); | |
127 | preempt_enable(); | |
128 | if (likely(sock)) { | |
129 | struct smap_psock *peer = smap_psock_sk(sock); | |
130 | ||
131 | if (likely(peer && | |
132 | test_bit(SMAP_TX_RUNNING, &peer->state) && | |
133 | sk_stream_memory_free(peer->sock))) { | |
134 | peer->sock->sk_wmem_queued += skb->truesize; | |
135 | sk_mem_charge(peer->sock, skb->truesize); | |
136 | skb_queue_tail(&peer->rxqueue, skb); | |
137 | schedule_work(&peer->tx_work); | |
138 | break; | |
139 | } | |
140 | } | |
141 | /* Fall through and free skb otherwise */ | |
142 | case SK_DROP: | |
143 | default: | |
976d28bf DB |
144 | if (rc != SK_REDIRECT) |
145 | preempt_enable(); | |
174a79ff JF |
146 | kfree_skb(skb); |
147 | } | |
148 | } | |
149 | ||
150 | static void smap_report_sk_error(struct smap_psock *psock, int err) | |
151 | { | |
152 | struct sock *sk = psock->sock; | |
153 | ||
154 | sk->sk_err = err; | |
155 | sk->sk_error_report(sk); | |
156 | } | |
157 | ||
2f857d04 | 158 | static void smap_release_sock(struct smap_psock *psock, struct sock *sock); |
174a79ff JF |
159 | |
160 | /* Called with lock_sock(sk) held */ | |
161 | static void smap_state_change(struct sock *sk) | |
162 | { | |
2f857d04 | 163 | struct smap_psock_map_entry *e, *tmp; |
174a79ff | 164 | struct smap_psock *psock; |
78aeaaef | 165 | struct socket_wq *wq; |
174a79ff JF |
166 | struct sock *osk; |
167 | ||
168 | rcu_read_lock(); | |
169 | ||
170 | /* Allowing transitions into an established syn_recv states allows | |
171 | * for early binding sockets to a smap object before the connection | |
172 | * is established. | |
173 | */ | |
174 | switch (sk->sk_state) { | |
78aeaaef | 175 | case TCP_SYN_SENT: |
174a79ff JF |
176 | case TCP_SYN_RECV: |
177 | case TCP_ESTABLISHED: | |
178 | break; | |
179 | case TCP_CLOSE_WAIT: | |
180 | case TCP_CLOSING: | |
181 | case TCP_LAST_ACK: | |
182 | case TCP_FIN_WAIT1: | |
183 | case TCP_FIN_WAIT2: | |
184 | case TCP_LISTEN: | |
185 | break; | |
186 | case TCP_CLOSE: | |
187 | /* Only release if the map entry is in fact the sock in | |
188 | * question. There is a case where the operator deletes | |
189 | * the sock from the map, but the TCP sock is closed before | |
190 | * the psock is detached. Use cmpxchg to verify correct | |
191 | * sock is removed. | |
192 | */ | |
193 | psock = smap_psock_sk(sk); | |
194 | if (unlikely(!psock)) | |
195 | break; | |
2f857d04 JF |
196 | write_lock_bh(&sk->sk_callback_lock); |
197 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { | |
198 | osk = cmpxchg(e->entry, sk, NULL); | |
199 | if (osk == sk) { | |
200 | list_del(&e->list); | |
201 | smap_release_sock(psock, sk); | |
202 | } | |
203 | } | |
204 | write_unlock_bh(&sk->sk_callback_lock); | |
174a79ff JF |
205 | break; |
206 | default: | |
cf56e3b9 JF |
207 | psock = smap_psock_sk(sk); |
208 | if (unlikely(!psock)) | |
209 | break; | |
174a79ff JF |
210 | smap_report_sk_error(psock, EPIPE); |
211 | break; | |
212 | } | |
78aeaaef JF |
213 | |
214 | wq = rcu_dereference(sk->sk_wq); | |
215 | if (skwq_has_sleeper(wq)) | |
216 | wake_up_interruptible_all(&wq->wait); | |
174a79ff JF |
217 | rcu_read_unlock(); |
218 | } | |
219 | ||
220 | static void smap_read_sock_strparser(struct strparser *strp, | |
221 | struct sk_buff *skb) | |
222 | { | |
223 | struct smap_psock *psock; | |
224 | ||
225 | rcu_read_lock(); | |
226 | psock = container_of(strp, struct smap_psock, strp); | |
227 | smap_do_verdict(psock, skb); | |
228 | rcu_read_unlock(); | |
229 | } | |
230 | ||
231 | /* Called with lock held on socket */ | |
232 | static void smap_data_ready(struct sock *sk) | |
233 | { | |
234 | struct smap_psock *psock; | |
235 | ||
d26e597d | 236 | rcu_read_lock(); |
174a79ff | 237 | psock = smap_psock_sk(sk); |
d26e597d JF |
238 | if (likely(psock)) { |
239 | write_lock_bh(&sk->sk_callback_lock); | |
174a79ff | 240 | strp_data_ready(&psock->strp); |
d26e597d JF |
241 | write_unlock_bh(&sk->sk_callback_lock); |
242 | } | |
243 | rcu_read_unlock(); | |
174a79ff JF |
244 | } |
245 | ||
246 | static void smap_tx_work(struct work_struct *w) | |
247 | { | |
248 | struct smap_psock *psock; | |
249 | struct sk_buff *skb; | |
250 | int rem, off, n; | |
251 | ||
252 | psock = container_of(w, struct smap_psock, tx_work); | |
253 | ||
254 | /* lock sock to avoid losing sk_socket at some point during loop */ | |
255 | lock_sock(psock->sock); | |
256 | if (psock->save_skb) { | |
257 | skb = psock->save_skb; | |
258 | rem = psock->save_rem; | |
259 | off = psock->save_off; | |
260 | psock->save_skb = NULL; | |
261 | goto start; | |
262 | } | |
263 | ||
264 | while ((skb = skb_dequeue(&psock->rxqueue))) { | |
265 | rem = skb->len; | |
266 | off = 0; | |
267 | start: | |
268 | do { | |
269 | if (likely(psock->sock->sk_socket)) | |
270 | n = skb_send_sock_locked(psock->sock, | |
271 | skb, off, rem); | |
272 | else | |
273 | n = -EINVAL; | |
274 | if (n <= 0) { | |
275 | if (n == -EAGAIN) { | |
276 | /* Retry when space is available */ | |
277 | psock->save_skb = skb; | |
278 | psock->save_rem = rem; | |
279 | psock->save_off = off; | |
280 | goto out; | |
281 | } | |
282 | /* Hard errors break pipe and stop xmit */ | |
283 | smap_report_sk_error(psock, n ? -n : EPIPE); | |
284 | clear_bit(SMAP_TX_RUNNING, &psock->state); | |
285 | sk_mem_uncharge(psock->sock, skb->truesize); | |
286 | psock->sock->sk_wmem_queued -= skb->truesize; | |
287 | kfree_skb(skb); | |
288 | goto out; | |
289 | } | |
290 | rem -= n; | |
291 | off += n; | |
292 | } while (rem); | |
293 | sk_mem_uncharge(psock->sock, skb->truesize); | |
294 | psock->sock->sk_wmem_queued -= skb->truesize; | |
295 | kfree_skb(skb); | |
296 | } | |
297 | out: | |
298 | release_sock(psock->sock); | |
299 | } | |
300 | ||
301 | static void smap_write_space(struct sock *sk) | |
302 | { | |
303 | struct smap_psock *psock; | |
304 | ||
305 | rcu_read_lock(); | |
306 | psock = smap_psock_sk(sk); | |
307 | if (likely(psock && test_bit(SMAP_TX_RUNNING, &psock->state))) | |
308 | schedule_work(&psock->tx_work); | |
309 | rcu_read_unlock(); | |
310 | } | |
311 | ||
312 | static void smap_stop_sock(struct smap_psock *psock, struct sock *sk) | |
313 | { | |
174a79ff | 314 | if (!psock->strp_enabled) |
2f857d04 | 315 | return; |
174a79ff JF |
316 | sk->sk_data_ready = psock->save_data_ready; |
317 | sk->sk_write_space = psock->save_write_space; | |
318 | sk->sk_state_change = psock->save_state_change; | |
319 | psock->save_data_ready = NULL; | |
320 | psock->save_write_space = NULL; | |
321 | psock->save_state_change = NULL; | |
322 | strp_stop(&psock->strp); | |
323 | psock->strp_enabled = false; | |
174a79ff JF |
324 | } |
325 | ||
326 | static void smap_destroy_psock(struct rcu_head *rcu) | |
327 | { | |
328 | struct smap_psock *psock = container_of(rcu, | |
329 | struct smap_psock, rcu); | |
330 | ||
331 | /* Now that a grace period has passed there is no longer | |
332 | * any reference to this sock in the sockmap so we can | |
333 | * destroy the psock, strparser, and bpf programs. But, | |
334 | * because we use workqueue sync operations we can not | |
335 | * do it in rcu context | |
336 | */ | |
337 | schedule_work(&psock->gc_work); | |
338 | } | |
339 | ||
2f857d04 | 340 | static void smap_release_sock(struct smap_psock *psock, struct sock *sock) |
174a79ff | 341 | { |
2f857d04 JF |
342 | psock->refcnt--; |
343 | if (psock->refcnt) | |
344 | return; | |
174a79ff JF |
345 | |
346 | smap_stop_sock(psock, sock); | |
347 | clear_bit(SMAP_TX_RUNNING, &psock->state); | |
348 | rcu_assign_sk_user_data(sock, NULL); | |
349 | call_rcu_sched(&psock->rcu, smap_destroy_psock); | |
350 | } | |
351 | ||
352 | static int smap_parse_func_strparser(struct strparser *strp, | |
353 | struct sk_buff *skb) | |
354 | { | |
355 | struct smap_psock *psock; | |
356 | struct bpf_prog *prog; | |
357 | int rc; | |
358 | ||
359 | rcu_read_lock(); | |
360 | psock = container_of(strp, struct smap_psock, strp); | |
361 | prog = READ_ONCE(psock->bpf_parse); | |
362 | ||
363 | if (unlikely(!prog)) { | |
364 | rcu_read_unlock(); | |
365 | return skb->len; | |
366 | } | |
367 | ||
368 | /* Attach socket for bpf program to use if needed we can do this | |
369 | * because strparser clones the skb before handing it to a upper | |
370 | * layer, meaning skb_orphan has been called. We NULL sk on the | |
371 | * way out to ensure we don't trigger a BUG_ON in skb/sk operations | |
372 | * later and because we are not charging the memory of this skb to | |
373 | * any socket yet. | |
374 | */ | |
375 | skb->sk = psock->sock; | |
376 | bpf_compute_data_end(skb); | |
377 | rc = (*prog->bpf_func)(skb, prog->insnsi); | |
378 | skb->sk = NULL; | |
379 | rcu_read_unlock(); | |
380 | return rc; | |
381 | } | |
382 | ||
383 | ||
384 | static int smap_read_sock_done(struct strparser *strp, int err) | |
385 | { | |
386 | return err; | |
387 | } | |
388 | ||
389 | static int smap_init_sock(struct smap_psock *psock, | |
390 | struct sock *sk) | |
391 | { | |
3fd87127 EB |
392 | static const struct strp_callbacks cb = { |
393 | .rcv_msg = smap_read_sock_strparser, | |
394 | .parse_msg = smap_parse_func_strparser, | |
395 | .read_sock_done = smap_read_sock_done, | |
396 | }; | |
174a79ff | 397 | |
174a79ff JF |
398 | return strp_init(&psock->strp, sk, &cb); |
399 | } | |
400 | ||
401 | static void smap_init_progs(struct smap_psock *psock, | |
402 | struct bpf_stab *stab, | |
403 | struct bpf_prog *verdict, | |
404 | struct bpf_prog *parse) | |
405 | { | |
406 | struct bpf_prog *orig_parse, *orig_verdict; | |
407 | ||
408 | orig_parse = xchg(&psock->bpf_parse, parse); | |
409 | orig_verdict = xchg(&psock->bpf_verdict, verdict); | |
410 | ||
411 | if (orig_verdict) | |
412 | bpf_prog_put(orig_verdict); | |
413 | if (orig_parse) | |
414 | bpf_prog_put(orig_parse); | |
415 | } | |
416 | ||
417 | static void smap_start_sock(struct smap_psock *psock, struct sock *sk) | |
418 | { | |
419 | if (sk->sk_data_ready == smap_data_ready) | |
420 | return; | |
421 | psock->save_data_ready = sk->sk_data_ready; | |
422 | psock->save_write_space = sk->sk_write_space; | |
423 | psock->save_state_change = sk->sk_state_change; | |
424 | sk->sk_data_ready = smap_data_ready; | |
425 | sk->sk_write_space = smap_write_space; | |
426 | sk->sk_state_change = smap_state_change; | |
427 | psock->strp_enabled = true; | |
428 | } | |
429 | ||
430 | static void sock_map_remove_complete(struct bpf_stab *stab) | |
431 | { | |
432 | bpf_map_area_free(stab->sock_map); | |
433 | kfree(stab); | |
434 | } | |
435 | ||
436 | static void smap_gc_work(struct work_struct *w) | |
437 | { | |
2f857d04 | 438 | struct smap_psock_map_entry *e, *tmp; |
174a79ff JF |
439 | struct smap_psock *psock; |
440 | ||
441 | psock = container_of(w, struct smap_psock, gc_work); | |
442 | ||
443 | /* no callback lock needed because we already detached sockmap ops */ | |
444 | if (psock->strp_enabled) | |
445 | strp_done(&psock->strp); | |
446 | ||
447 | cancel_work_sync(&psock->tx_work); | |
448 | __skb_queue_purge(&psock->rxqueue); | |
449 | ||
450 | /* At this point all strparser and xmit work must be complete */ | |
451 | if (psock->bpf_parse) | |
452 | bpf_prog_put(psock->bpf_parse); | |
453 | if (psock->bpf_verdict) | |
454 | bpf_prog_put(psock->bpf_verdict); | |
455 | ||
2f857d04 JF |
456 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { |
457 | list_del(&e->list); | |
458 | kfree(e); | |
459 | } | |
174a79ff JF |
460 | |
461 | sock_put(psock->sock); | |
462 | kfree(psock); | |
463 | } | |
464 | ||
465 | static struct smap_psock *smap_init_psock(struct sock *sock, | |
466 | struct bpf_stab *stab) | |
467 | { | |
468 | struct smap_psock *psock; | |
469 | ||
96eabe7a MKL |
470 | psock = kzalloc_node(sizeof(struct smap_psock), |
471 | GFP_ATOMIC | __GFP_NOWARN, | |
472 | stab->map.numa_node); | |
174a79ff JF |
473 | if (!psock) |
474 | return ERR_PTR(-ENOMEM); | |
475 | ||
476 | psock->sock = sock; | |
477 | skb_queue_head_init(&psock->rxqueue); | |
478 | INIT_WORK(&psock->tx_work, smap_tx_work); | |
479 | INIT_WORK(&psock->gc_work, smap_gc_work); | |
2f857d04 JF |
480 | INIT_LIST_HEAD(&psock->maps); |
481 | psock->refcnt = 1; | |
174a79ff JF |
482 | |
483 | rcu_assign_sk_user_data(sock, psock); | |
484 | sock_hold(sock); | |
485 | return psock; | |
486 | } | |
487 | ||
488 | static struct bpf_map *sock_map_alloc(union bpf_attr *attr) | |
489 | { | |
490 | struct bpf_stab *stab; | |
491 | int err = -EINVAL; | |
492 | u64 cost; | |
493 | ||
494 | /* check sanity of attributes */ | |
495 | if (attr->max_entries == 0 || attr->key_size != 4 || | |
96eabe7a | 496 | attr->value_size != 4 || attr->map_flags & ~BPF_F_NUMA_NODE) |
174a79ff JF |
497 | return ERR_PTR(-EINVAL); |
498 | ||
499 | if (attr->value_size > KMALLOC_MAX_SIZE) | |
500 | return ERR_PTR(-E2BIG); | |
501 | ||
502 | stab = kzalloc(sizeof(*stab), GFP_USER); | |
503 | if (!stab) | |
504 | return ERR_PTR(-ENOMEM); | |
505 | ||
506 | /* mandatory map attributes */ | |
507 | stab->map.map_type = attr->map_type; | |
508 | stab->map.key_size = attr->key_size; | |
509 | stab->map.value_size = attr->value_size; | |
510 | stab->map.max_entries = attr->max_entries; | |
511 | stab->map.map_flags = attr->map_flags; | |
96eabe7a | 512 | stab->map.numa_node = bpf_map_attr_numa_node(attr); |
174a79ff JF |
513 | |
514 | /* make sure page count doesn't overflow */ | |
515 | cost = (u64) stab->map.max_entries * sizeof(struct sock *); | |
516 | if (cost >= U32_MAX - PAGE_SIZE) | |
517 | goto free_stab; | |
518 | ||
519 | stab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT; | |
520 | ||
521 | /* if map size is larger than memlock limit, reject it early */ | |
522 | err = bpf_map_precharge_memlock(stab->map.pages); | |
523 | if (err) | |
524 | goto free_stab; | |
525 | ||
526 | stab->sock_map = bpf_map_area_alloc(stab->map.max_entries * | |
96eabe7a MKL |
527 | sizeof(struct sock *), |
528 | stab->map.numa_node); | |
174a79ff JF |
529 | if (!stab->sock_map) |
530 | goto free_stab; | |
531 | ||
174a79ff JF |
532 | return &stab->map; |
533 | free_stab: | |
534 | kfree(stab); | |
535 | return ERR_PTR(err); | |
536 | } | |
537 | ||
2f857d04 JF |
538 | static void smap_list_remove(struct smap_psock *psock, struct sock **entry) |
539 | { | |
540 | struct smap_psock_map_entry *e, *tmp; | |
541 | ||
542 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { | |
543 | if (e->entry == entry) { | |
544 | list_del(&e->list); | |
545 | break; | |
546 | } | |
547 | } | |
548 | } | |
549 | ||
174a79ff JF |
550 | static void sock_map_free(struct bpf_map *map) |
551 | { | |
552 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
553 | int i; | |
554 | ||
555 | synchronize_rcu(); | |
556 | ||
557 | /* At this point no update, lookup or delete operations can happen. | |
558 | * However, be aware we can still get a socket state event updates, | |
559 | * and data ready callabacks that reference the psock from sk_user_data | |
560 | * Also psock worker threads are still in-flight. So smap_release_sock | |
561 | * will only free the psock after cancel_sync on the worker threads | |
562 | * and a grace period expire to ensure psock is really safe to remove. | |
563 | */ | |
564 | rcu_read_lock(); | |
565 | for (i = 0; i < stab->map.max_entries; i++) { | |
2f857d04 | 566 | struct smap_psock *psock; |
174a79ff JF |
567 | struct sock *sock; |
568 | ||
569 | sock = xchg(&stab->sock_map[i], NULL); | |
570 | if (!sock) | |
571 | continue; | |
572 | ||
2f857d04 JF |
573 | write_lock_bh(&sock->sk_callback_lock); |
574 | psock = smap_psock_sk(sock); | |
575 | smap_list_remove(psock, &stab->sock_map[i]); | |
576 | smap_release_sock(psock, sock); | |
577 | write_unlock_bh(&sock->sk_callback_lock); | |
174a79ff JF |
578 | } |
579 | rcu_read_unlock(); | |
580 | ||
581 | if (stab->bpf_verdict) | |
582 | bpf_prog_put(stab->bpf_verdict); | |
583 | if (stab->bpf_parse) | |
584 | bpf_prog_put(stab->bpf_parse); | |
585 | ||
2f857d04 | 586 | sock_map_remove_complete(stab); |
174a79ff JF |
587 | } |
588 | ||
589 | static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next_key) | |
590 | { | |
591 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
592 | u32 i = key ? *(u32 *)key : U32_MAX; | |
593 | u32 *next = (u32 *)next_key; | |
594 | ||
595 | if (i >= stab->map.max_entries) { | |
596 | *next = 0; | |
597 | return 0; | |
598 | } | |
599 | ||
600 | if (i == stab->map.max_entries - 1) | |
601 | return -ENOENT; | |
602 | ||
603 | *next = i + 1; | |
604 | return 0; | |
605 | } | |
606 | ||
607 | struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key) | |
608 | { | |
609 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
610 | ||
611 | if (key >= map->max_entries) | |
612 | return NULL; | |
613 | ||
614 | return READ_ONCE(stab->sock_map[key]); | |
615 | } | |
616 | ||
617 | static int sock_map_delete_elem(struct bpf_map *map, void *key) | |
618 | { | |
619 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
2f857d04 | 620 | struct smap_psock *psock; |
174a79ff JF |
621 | int k = *(u32 *)key; |
622 | struct sock *sock; | |
623 | ||
624 | if (k >= map->max_entries) | |
625 | return -EINVAL; | |
626 | ||
627 | sock = xchg(&stab->sock_map[k], NULL); | |
628 | if (!sock) | |
629 | return -EINVAL; | |
630 | ||
2f857d04 JF |
631 | write_lock_bh(&sock->sk_callback_lock); |
632 | psock = smap_psock_sk(sock); | |
633 | if (!psock) | |
634 | goto out; | |
635 | ||
636 | if (psock->bpf_parse) | |
637 | smap_stop_sock(psock, sock); | |
638 | smap_list_remove(psock, &stab->sock_map[k]); | |
639 | smap_release_sock(psock, sock); | |
640 | out: | |
641 | write_unlock_bh(&sock->sk_callback_lock); | |
174a79ff JF |
642 | return 0; |
643 | } | |
644 | ||
645 | /* Locking notes: Concurrent updates, deletes, and lookups are allowed and are | |
646 | * done inside rcu critical sections. This ensures on updates that the psock | |
647 | * will not be released via smap_release_sock() until concurrent updates/deletes | |
648 | * complete. All operations operate on sock_map using cmpxchg and xchg | |
649 | * operations to ensure we do not get stale references. Any reads into the | |
650 | * map must be done with READ_ONCE() because of this. | |
651 | * | |
652 | * A psock is destroyed via call_rcu and after any worker threads are cancelled | |
653 | * and syncd so we are certain all references from the update/lookup/delete | |
654 | * operations as well as references in the data path are no longer in use. | |
655 | * | |
2f857d04 JF |
656 | * Psocks may exist in multiple maps, but only a single set of parse/verdict |
657 | * programs may be inherited from the maps it belongs to. A reference count | |
658 | * is kept with the total number of references to the psock from all maps. The | |
659 | * psock will not be released until this reaches zero. The psock and sock | |
660 | * user data data use the sk_callback_lock to protect critical data structures | |
661 | * from concurrent access. This allows us to avoid two updates from modifying | |
662 | * the user data in sock and the lock is required anyways for modifying | |
663 | * callbacks, we simply increase its scope slightly. | |
174a79ff | 664 | * |
2f857d04 JF |
665 | * Rules to follow, |
666 | * - psock must always be read inside RCU critical section | |
667 | * - sk_user_data must only be modified inside sk_callback_lock and read | |
668 | * inside RCU critical section. | |
669 | * - psock->maps list must only be read & modified inside sk_callback_lock | |
670 | * - sock_map must use READ_ONCE and (cmp)xchg operations | |
671 | * - BPF verdict/parse programs must use READ_ONCE and xchg operations | |
174a79ff JF |
672 | */ |
673 | static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops, | |
674 | struct bpf_map *map, | |
2f857d04 | 675 | void *key, u64 flags) |
174a79ff JF |
676 | { |
677 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
2f857d04 | 678 | struct smap_psock_map_entry *e = NULL; |
174a79ff | 679 | struct bpf_prog *verdict, *parse; |
2f857d04 JF |
680 | struct sock *osock, *sock; |
681 | struct smap_psock *psock; | |
174a79ff | 682 | u32 i = *(u32 *)key; |
2f857d04 | 683 | int err; |
174a79ff JF |
684 | |
685 | if (unlikely(flags > BPF_EXIST)) | |
686 | return -EINVAL; | |
687 | ||
688 | if (unlikely(i >= stab->map.max_entries)) | |
689 | return -E2BIG; | |
690 | ||
174a79ff | 691 | sock = READ_ONCE(stab->sock_map[i]); |
2f857d04 JF |
692 | if (flags == BPF_EXIST && !sock) |
693 | return -ENOENT; | |
694 | else if (flags == BPF_NOEXIST && sock) | |
174a79ff | 695 | return -EEXIST; |
174a79ff | 696 | |
2f857d04 | 697 | sock = skops->sk; |
174a79ff | 698 | |
2f857d04 JF |
699 | /* 1. If sock map has BPF programs those will be inherited by the |
700 | * sock being added. If the sock is already attached to BPF programs | |
701 | * this results in an error. | |
702 | */ | |
703 | verdict = READ_ONCE(stab->bpf_verdict); | |
704 | parse = READ_ONCE(stab->bpf_parse); | |
174a79ff | 705 | |
2f857d04 | 706 | if (parse && verdict) { |
174a79ff JF |
707 | /* bpf prog refcnt may be zero if a concurrent attach operation |
708 | * removes the program after the above READ_ONCE() but before | |
709 | * we increment the refcnt. If this is the case abort with an | |
710 | * error. | |
711 | */ | |
712 | verdict = bpf_prog_inc_not_zero(stab->bpf_verdict); | |
713 | if (IS_ERR(verdict)) | |
714 | return PTR_ERR(verdict); | |
715 | ||
716 | parse = bpf_prog_inc_not_zero(stab->bpf_parse); | |
717 | if (IS_ERR(parse)) { | |
718 | bpf_prog_put(verdict); | |
719 | return PTR_ERR(parse); | |
720 | } | |
721 | } | |
722 | ||
2f857d04 JF |
723 | write_lock_bh(&sock->sk_callback_lock); |
724 | psock = smap_psock_sk(sock); | |
725 | ||
726 | /* 2. Do not allow inheriting programs if psock exists and has | |
727 | * already inherited programs. This would create confusion on | |
728 | * which parser/verdict program is running. If no psock exists | |
729 | * create one. Inside sk_callback_lock to ensure concurrent create | |
730 | * doesn't update user data. | |
731 | */ | |
732 | if (psock) { | |
733 | if (READ_ONCE(psock->bpf_parse) && parse) { | |
734 | err = -EBUSY; | |
735 | goto out_progs; | |
736 | } | |
737 | psock->refcnt++; | |
738 | } else { | |
174a79ff JF |
739 | psock = smap_init_psock(sock, stab); |
740 | if (IS_ERR(psock)) { | |
2f857d04 JF |
741 | err = PTR_ERR(psock); |
742 | goto out_progs; | |
174a79ff | 743 | } |
2f857d04 | 744 | |
174a79ff JF |
745 | set_bit(SMAP_TX_RUNNING, &psock->state); |
746 | } | |
747 | ||
2f857d04 JF |
748 | e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN); |
749 | if (!e) { | |
750 | err = -ENOMEM; | |
751 | goto out_progs; | |
752 | } | |
753 | e->entry = &stab->sock_map[i]; | |
754 | ||
755 | /* 3. At this point we have a reference to a valid psock that is | |
756 | * running. Attach any BPF programs needed. | |
757 | */ | |
758 | if (parse && verdict && !psock->strp_enabled) { | |
174a79ff JF |
759 | err = smap_init_sock(psock, sock); |
760 | if (err) | |
2f857d04 | 761 | goto out_free; |
174a79ff JF |
762 | smap_init_progs(psock, stab, verdict, parse); |
763 | smap_start_sock(psock, sock); | |
174a79ff JF |
764 | } |
765 | ||
2f857d04 JF |
766 | /* 4. Place psock in sockmap for use and stop any programs on |
767 | * the old sock assuming its not the same sock we are replacing | |
768 | * it with. Because we can only have a single set of programs if | |
769 | * old_sock has a strp we can stop it. | |
770 | */ | |
771 | list_add_tail(&e->list, &psock->maps); | |
772 | write_unlock_bh(&sock->sk_callback_lock); | |
174a79ff | 773 | |
2f857d04 JF |
774 | osock = xchg(&stab->sock_map[i], sock); |
775 | if (osock) { | |
776 | struct smap_psock *opsock = smap_psock_sk(osock); | |
777 | ||
778 | write_lock_bh(&osock->sk_callback_lock); | |
779 | if (osock != sock && parse) | |
780 | smap_stop_sock(opsock, osock); | |
781 | smap_list_remove(opsock, &stab->sock_map[i]); | |
782 | smap_release_sock(opsock, osock); | |
783 | write_unlock_bh(&osock->sk_callback_lock); | |
784 | } | |
174a79ff | 785 | return 0; |
2f857d04 JF |
786 | out_free: |
787 | smap_release_sock(psock, sock); | |
788 | out_progs: | |
789 | if (verdict) | |
790 | bpf_prog_put(verdict); | |
791 | if (parse) | |
792 | bpf_prog_put(parse); | |
174a79ff | 793 | write_unlock_bh(&sock->sk_callback_lock); |
2f857d04 | 794 | kfree(e); |
174a79ff JF |
795 | return err; |
796 | } | |
797 | ||
464bc0fd | 798 | int sock_map_attach_prog(struct bpf_map *map, struct bpf_prog *prog, u32 type) |
174a79ff JF |
799 | { |
800 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
464bc0fd | 801 | struct bpf_prog *orig; |
174a79ff | 802 | |
81374aaa JF |
803 | if (unlikely(map->map_type != BPF_MAP_TYPE_SOCKMAP)) |
804 | return -EINVAL; | |
805 | ||
464bc0fd JF |
806 | switch (type) { |
807 | case BPF_SK_SKB_STREAM_PARSER: | |
808 | orig = xchg(&stab->bpf_parse, prog); | |
809 | break; | |
810 | case BPF_SK_SKB_STREAM_VERDICT: | |
811 | orig = xchg(&stab->bpf_verdict, prog); | |
812 | break; | |
813 | default: | |
814 | return -EOPNOTSUPP; | |
815 | } | |
174a79ff | 816 | |
464bc0fd JF |
817 | if (orig) |
818 | bpf_prog_put(orig); | |
174a79ff JF |
819 | |
820 | return 0; | |
821 | } | |
822 | ||
823 | static void *sock_map_lookup(struct bpf_map *map, void *key) | |
824 | { | |
825 | return NULL; | |
826 | } | |
827 | ||
828 | static int sock_map_update_elem(struct bpf_map *map, | |
829 | void *key, void *value, u64 flags) | |
830 | { | |
831 | struct bpf_sock_ops_kern skops; | |
832 | u32 fd = *(u32 *)value; | |
833 | struct socket *socket; | |
834 | int err; | |
835 | ||
836 | socket = sockfd_lookup(fd, &err); | |
837 | if (!socket) | |
838 | return err; | |
839 | ||
840 | skops.sk = socket->sk; | |
841 | if (!skops.sk) { | |
842 | fput(socket->file); | |
843 | return -EINVAL; | |
844 | } | |
845 | ||
2f857d04 | 846 | err = sock_map_ctx_update_elem(&skops, map, key, flags); |
174a79ff JF |
847 | fput(socket->file); |
848 | return err; | |
849 | } | |
850 | ||
851 | const struct bpf_map_ops sock_map_ops = { | |
852 | .map_alloc = sock_map_alloc, | |
853 | .map_free = sock_map_free, | |
854 | .map_lookup_elem = sock_map_lookup, | |
855 | .map_get_next_key = sock_map_get_next_key, | |
856 | .map_update_elem = sock_map_update_elem, | |
857 | .map_delete_elem = sock_map_delete_elem, | |
174a79ff JF |
858 | }; |
859 | ||
2f857d04 JF |
860 | BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, bpf_sock, |
861 | struct bpf_map *, map, void *, key, u64, flags) | |
174a79ff JF |
862 | { |
863 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
2f857d04 | 864 | return sock_map_ctx_update_elem(bpf_sock, map, key, flags); |
174a79ff JF |
865 | } |
866 | ||
867 | const struct bpf_func_proto bpf_sock_map_update_proto = { | |
868 | .func = bpf_sock_map_update, | |
869 | .gpl_only = false, | |
870 | .pkt_access = true, | |
871 | .ret_type = RET_INTEGER, | |
872 | .arg1_type = ARG_PTR_TO_CTX, | |
873 | .arg2_type = ARG_CONST_MAP_PTR, | |
874 | .arg3_type = ARG_PTR_TO_MAP_KEY, | |
875 | .arg4_type = ARG_ANYTHING, | |
174a79ff | 876 | }; |