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> | |
34f79502 | 42 | #include <net/tcp.h> |
174a79ff | 43 | |
cace572e CF |
44 | #define SOCK_CREATE_FLAG_MASK \ |
45 | (BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY) | |
46 | ||
174a79ff JF |
47 | struct bpf_stab { |
48 | struct bpf_map map; | |
49 | struct sock **sock_map; | |
50 | struct bpf_prog *bpf_parse; | |
51 | struct bpf_prog *bpf_verdict; | |
174a79ff JF |
52 | }; |
53 | ||
54 | enum smap_psock_state { | |
55 | SMAP_TX_RUNNING, | |
56 | }; | |
57 | ||
2f857d04 JF |
58 | struct smap_psock_map_entry { |
59 | struct list_head list; | |
60 | struct sock **entry; | |
61 | }; | |
62 | ||
174a79ff JF |
63 | struct smap_psock { |
64 | struct rcu_head rcu; | |
2f857d04 JF |
65 | /* refcnt is used inside sk_callback_lock */ |
66 | u32 refcnt; | |
174a79ff JF |
67 | |
68 | /* datapath variables */ | |
69 | struct sk_buff_head rxqueue; | |
70 | bool strp_enabled; | |
71 | ||
72 | /* datapath error path cache across tx work invocations */ | |
73 | int save_rem; | |
74 | int save_off; | |
75 | struct sk_buff *save_skb; | |
76 | ||
77 | struct strparser strp; | |
78 | struct bpf_prog *bpf_parse; | |
79 | struct bpf_prog *bpf_verdict; | |
2f857d04 | 80 | struct list_head maps; |
174a79ff JF |
81 | |
82 | /* Back reference used when sock callback trigger sockmap operations */ | |
174a79ff JF |
83 | struct sock *sock; |
84 | unsigned long state; | |
85 | ||
86 | struct work_struct tx_work; | |
87 | struct work_struct gc_work; | |
88 | ||
89 | void (*save_data_ready)(struct sock *sk); | |
90 | void (*save_write_space)(struct sock *sk); | |
91 | void (*save_state_change)(struct sock *sk); | |
92 | }; | |
93 | ||
94 | static inline struct smap_psock *smap_psock_sk(const struct sock *sk) | |
95 | { | |
2f857d04 | 96 | return rcu_dereference_sk_user_data(sk); |
174a79ff JF |
97 | } |
98 | ||
8108a775 JF |
99 | /* compute the linear packet data range [data, data_end) for skb when |
100 | * sk_skb type programs are in use. | |
101 | */ | |
102 | static inline void bpf_compute_data_end_sk_skb(struct sk_buff *skb) | |
103 | { | |
104 | TCP_SKB_CB(skb)->bpf.data_end = skb->data + skb_headlen(skb); | |
105 | } | |
106 | ||
04686ef2 JF |
107 | enum __sk_action { |
108 | __SK_DROP = 0, | |
109 | __SK_PASS, | |
110 | __SK_REDIRECT, | |
111 | }; | |
112 | ||
174a79ff JF |
113 | static int smap_verdict_func(struct smap_psock *psock, struct sk_buff *skb) |
114 | { | |
115 | struct bpf_prog *prog = READ_ONCE(psock->bpf_verdict); | |
116 | int rc; | |
117 | ||
118 | if (unlikely(!prog)) | |
04686ef2 | 119 | return __SK_DROP; |
174a79ff JF |
120 | |
121 | skb_orphan(skb); | |
34f79502 JF |
122 | /* We need to ensure that BPF metadata for maps is also cleared |
123 | * when we orphan the skb so that we don't have the possibility | |
124 | * to reference a stale map. | |
125 | */ | |
126 | TCP_SKB_CB(skb)->bpf.map = NULL; | |
174a79ff | 127 | skb->sk = psock->sock; |
8108a775 | 128 | bpf_compute_data_end_sk_skb(skb); |
34f79502 | 129 | preempt_disable(); |
174a79ff | 130 | rc = (*prog->bpf_func)(skb, prog->insnsi); |
34f79502 | 131 | preempt_enable(); |
174a79ff JF |
132 | skb->sk = NULL; |
133 | ||
04686ef2 | 134 | /* Moving return codes from UAPI namespace into internal namespace */ |
bfa64075 | 135 | return rc == SK_PASS ? |
04686ef2 JF |
136 | (TCP_SKB_CB(skb)->bpf.map ? __SK_REDIRECT : __SK_PASS) : |
137 | __SK_DROP; | |
174a79ff JF |
138 | } |
139 | ||
140 | static void smap_do_verdict(struct smap_psock *psock, struct sk_buff *skb) | |
141 | { | |
90a9631c | 142 | struct sock *sk; |
174a79ff JF |
143 | int rc; |
144 | ||
174a79ff JF |
145 | rc = smap_verdict_func(psock, skb); |
146 | switch (rc) { | |
04686ef2 | 147 | case __SK_REDIRECT: |
34f79502 | 148 | sk = do_sk_redirect_map(skb); |
90a9631c JF |
149 | if (likely(sk)) { |
150 | struct smap_psock *peer = smap_psock_sk(sk); | |
174a79ff JF |
151 | |
152 | if (likely(peer && | |
153 | test_bit(SMAP_TX_RUNNING, &peer->state) && | |
90a9631c JF |
154 | !sock_flag(sk, SOCK_DEAD) && |
155 | sock_writeable(sk))) { | |
156 | skb_set_owner_w(skb, sk); | |
174a79ff JF |
157 | skb_queue_tail(&peer->rxqueue, skb); |
158 | schedule_work(&peer->tx_work); | |
159 | break; | |
160 | } | |
161 | } | |
162 | /* Fall through and free skb otherwise */ | |
04686ef2 | 163 | case __SK_DROP: |
174a79ff | 164 | default: |
174a79ff JF |
165 | kfree_skb(skb); |
166 | } | |
167 | } | |
168 | ||
169 | static void smap_report_sk_error(struct smap_psock *psock, int err) | |
170 | { | |
171 | struct sock *sk = psock->sock; | |
172 | ||
173 | sk->sk_err = err; | |
174 | sk->sk_error_report(sk); | |
175 | } | |
176 | ||
2f857d04 | 177 | static void smap_release_sock(struct smap_psock *psock, struct sock *sock); |
174a79ff JF |
178 | |
179 | /* Called with lock_sock(sk) held */ | |
180 | static void smap_state_change(struct sock *sk) | |
181 | { | |
2f857d04 | 182 | struct smap_psock_map_entry *e, *tmp; |
174a79ff | 183 | struct smap_psock *psock; |
78aeaaef | 184 | struct socket_wq *wq; |
174a79ff JF |
185 | struct sock *osk; |
186 | ||
187 | rcu_read_lock(); | |
188 | ||
189 | /* Allowing transitions into an established syn_recv states allows | |
190 | * for early binding sockets to a smap object before the connection | |
191 | * is established. | |
192 | */ | |
193 | switch (sk->sk_state) { | |
78aeaaef | 194 | case TCP_SYN_SENT: |
174a79ff JF |
195 | case TCP_SYN_RECV: |
196 | case TCP_ESTABLISHED: | |
197 | break; | |
198 | case TCP_CLOSE_WAIT: | |
199 | case TCP_CLOSING: | |
200 | case TCP_LAST_ACK: | |
201 | case TCP_FIN_WAIT1: | |
202 | case TCP_FIN_WAIT2: | |
203 | case TCP_LISTEN: | |
204 | break; | |
205 | case TCP_CLOSE: | |
206 | /* Only release if the map entry is in fact the sock in | |
207 | * question. There is a case where the operator deletes | |
208 | * the sock from the map, but the TCP sock is closed before | |
209 | * the psock is detached. Use cmpxchg to verify correct | |
210 | * sock is removed. | |
211 | */ | |
212 | psock = smap_psock_sk(sk); | |
213 | if (unlikely(!psock)) | |
214 | break; | |
2f857d04 JF |
215 | write_lock_bh(&sk->sk_callback_lock); |
216 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { | |
217 | osk = cmpxchg(e->entry, sk, NULL); | |
218 | if (osk == sk) { | |
219 | list_del(&e->list); | |
220 | smap_release_sock(psock, sk); | |
221 | } | |
222 | } | |
223 | write_unlock_bh(&sk->sk_callback_lock); | |
174a79ff JF |
224 | break; |
225 | default: | |
cf56e3b9 JF |
226 | psock = smap_psock_sk(sk); |
227 | if (unlikely(!psock)) | |
228 | break; | |
174a79ff JF |
229 | smap_report_sk_error(psock, EPIPE); |
230 | break; | |
231 | } | |
78aeaaef JF |
232 | |
233 | wq = rcu_dereference(sk->sk_wq); | |
234 | if (skwq_has_sleeper(wq)) | |
235 | wake_up_interruptible_all(&wq->wait); | |
174a79ff JF |
236 | rcu_read_unlock(); |
237 | } | |
238 | ||
239 | static void smap_read_sock_strparser(struct strparser *strp, | |
240 | struct sk_buff *skb) | |
241 | { | |
242 | struct smap_psock *psock; | |
243 | ||
244 | rcu_read_lock(); | |
245 | psock = container_of(strp, struct smap_psock, strp); | |
246 | smap_do_verdict(psock, skb); | |
247 | rcu_read_unlock(); | |
248 | } | |
249 | ||
250 | /* Called with lock held on socket */ | |
251 | static void smap_data_ready(struct sock *sk) | |
252 | { | |
253 | struct smap_psock *psock; | |
254 | ||
d26e597d | 255 | rcu_read_lock(); |
174a79ff | 256 | psock = smap_psock_sk(sk); |
d26e597d JF |
257 | if (likely(psock)) { |
258 | write_lock_bh(&sk->sk_callback_lock); | |
174a79ff | 259 | strp_data_ready(&psock->strp); |
d26e597d JF |
260 | write_unlock_bh(&sk->sk_callback_lock); |
261 | } | |
262 | rcu_read_unlock(); | |
174a79ff JF |
263 | } |
264 | ||
265 | static void smap_tx_work(struct work_struct *w) | |
266 | { | |
267 | struct smap_psock *psock; | |
268 | struct sk_buff *skb; | |
269 | int rem, off, n; | |
270 | ||
271 | psock = container_of(w, struct smap_psock, tx_work); | |
272 | ||
273 | /* lock sock to avoid losing sk_socket at some point during loop */ | |
274 | lock_sock(psock->sock); | |
275 | if (psock->save_skb) { | |
276 | skb = psock->save_skb; | |
277 | rem = psock->save_rem; | |
278 | off = psock->save_off; | |
279 | psock->save_skb = NULL; | |
280 | goto start; | |
281 | } | |
282 | ||
283 | while ((skb = skb_dequeue(&psock->rxqueue))) { | |
284 | rem = skb->len; | |
285 | off = 0; | |
286 | start: | |
287 | do { | |
288 | if (likely(psock->sock->sk_socket)) | |
289 | n = skb_send_sock_locked(psock->sock, | |
290 | skb, off, rem); | |
291 | else | |
292 | n = -EINVAL; | |
293 | if (n <= 0) { | |
294 | if (n == -EAGAIN) { | |
295 | /* Retry when space is available */ | |
296 | psock->save_skb = skb; | |
297 | psock->save_rem = rem; | |
298 | psock->save_off = off; | |
299 | goto out; | |
300 | } | |
301 | /* Hard errors break pipe and stop xmit */ | |
302 | smap_report_sk_error(psock, n ? -n : EPIPE); | |
303 | clear_bit(SMAP_TX_RUNNING, &psock->state); | |
174a79ff JF |
304 | kfree_skb(skb); |
305 | goto out; | |
306 | } | |
307 | rem -= n; | |
308 | off += n; | |
309 | } while (rem); | |
174a79ff JF |
310 | kfree_skb(skb); |
311 | } | |
312 | out: | |
313 | release_sock(psock->sock); | |
314 | } | |
315 | ||
316 | static void smap_write_space(struct sock *sk) | |
317 | { | |
318 | struct smap_psock *psock; | |
319 | ||
320 | rcu_read_lock(); | |
321 | psock = smap_psock_sk(sk); | |
322 | if (likely(psock && test_bit(SMAP_TX_RUNNING, &psock->state))) | |
323 | schedule_work(&psock->tx_work); | |
324 | rcu_read_unlock(); | |
325 | } | |
326 | ||
327 | static void smap_stop_sock(struct smap_psock *psock, struct sock *sk) | |
328 | { | |
174a79ff | 329 | if (!psock->strp_enabled) |
2f857d04 | 330 | return; |
174a79ff JF |
331 | sk->sk_data_ready = psock->save_data_ready; |
332 | sk->sk_write_space = psock->save_write_space; | |
333 | sk->sk_state_change = psock->save_state_change; | |
334 | psock->save_data_ready = NULL; | |
335 | psock->save_write_space = NULL; | |
336 | psock->save_state_change = NULL; | |
337 | strp_stop(&psock->strp); | |
338 | psock->strp_enabled = false; | |
174a79ff JF |
339 | } |
340 | ||
341 | static void smap_destroy_psock(struct rcu_head *rcu) | |
342 | { | |
343 | struct smap_psock *psock = container_of(rcu, | |
344 | struct smap_psock, rcu); | |
345 | ||
346 | /* Now that a grace period has passed there is no longer | |
347 | * any reference to this sock in the sockmap so we can | |
348 | * destroy the psock, strparser, and bpf programs. But, | |
349 | * because we use workqueue sync operations we can not | |
350 | * do it in rcu context | |
351 | */ | |
352 | schedule_work(&psock->gc_work); | |
353 | } | |
354 | ||
2f857d04 | 355 | static void smap_release_sock(struct smap_psock *psock, struct sock *sock) |
174a79ff | 356 | { |
2f857d04 JF |
357 | psock->refcnt--; |
358 | if (psock->refcnt) | |
359 | return; | |
174a79ff JF |
360 | |
361 | smap_stop_sock(psock, sock); | |
362 | clear_bit(SMAP_TX_RUNNING, &psock->state); | |
363 | rcu_assign_sk_user_data(sock, NULL); | |
364 | call_rcu_sched(&psock->rcu, smap_destroy_psock); | |
365 | } | |
366 | ||
367 | static int smap_parse_func_strparser(struct strparser *strp, | |
368 | struct sk_buff *skb) | |
369 | { | |
370 | struct smap_psock *psock; | |
371 | struct bpf_prog *prog; | |
372 | int rc; | |
373 | ||
374 | rcu_read_lock(); | |
375 | psock = container_of(strp, struct smap_psock, strp); | |
376 | prog = READ_ONCE(psock->bpf_parse); | |
377 | ||
378 | if (unlikely(!prog)) { | |
379 | rcu_read_unlock(); | |
380 | return skb->len; | |
381 | } | |
382 | ||
383 | /* Attach socket for bpf program to use if needed we can do this | |
384 | * because strparser clones the skb before handing it to a upper | |
385 | * layer, meaning skb_orphan has been called. We NULL sk on the | |
386 | * way out to ensure we don't trigger a BUG_ON in skb/sk operations | |
387 | * later and because we are not charging the memory of this skb to | |
388 | * any socket yet. | |
389 | */ | |
390 | skb->sk = psock->sock; | |
8108a775 | 391 | bpf_compute_data_end_sk_skb(skb); |
174a79ff JF |
392 | rc = (*prog->bpf_func)(skb, prog->insnsi); |
393 | skb->sk = NULL; | |
394 | rcu_read_unlock(); | |
395 | return rc; | |
396 | } | |
397 | ||
398 | ||
399 | static int smap_read_sock_done(struct strparser *strp, int err) | |
400 | { | |
401 | return err; | |
402 | } | |
403 | ||
404 | static int smap_init_sock(struct smap_psock *psock, | |
405 | struct sock *sk) | |
406 | { | |
3fd87127 EB |
407 | static const struct strp_callbacks cb = { |
408 | .rcv_msg = smap_read_sock_strparser, | |
409 | .parse_msg = smap_parse_func_strparser, | |
410 | .read_sock_done = smap_read_sock_done, | |
411 | }; | |
174a79ff | 412 | |
174a79ff JF |
413 | return strp_init(&psock->strp, sk, &cb); |
414 | } | |
415 | ||
416 | static void smap_init_progs(struct smap_psock *psock, | |
417 | struct bpf_stab *stab, | |
418 | struct bpf_prog *verdict, | |
419 | struct bpf_prog *parse) | |
420 | { | |
421 | struct bpf_prog *orig_parse, *orig_verdict; | |
422 | ||
423 | orig_parse = xchg(&psock->bpf_parse, parse); | |
424 | orig_verdict = xchg(&psock->bpf_verdict, verdict); | |
425 | ||
426 | if (orig_verdict) | |
427 | bpf_prog_put(orig_verdict); | |
428 | if (orig_parse) | |
429 | bpf_prog_put(orig_parse); | |
430 | } | |
431 | ||
432 | static void smap_start_sock(struct smap_psock *psock, struct sock *sk) | |
433 | { | |
434 | if (sk->sk_data_ready == smap_data_ready) | |
435 | return; | |
436 | psock->save_data_ready = sk->sk_data_ready; | |
437 | psock->save_write_space = sk->sk_write_space; | |
438 | psock->save_state_change = sk->sk_state_change; | |
439 | sk->sk_data_ready = smap_data_ready; | |
440 | sk->sk_write_space = smap_write_space; | |
441 | sk->sk_state_change = smap_state_change; | |
442 | psock->strp_enabled = true; | |
443 | } | |
444 | ||
445 | static void sock_map_remove_complete(struct bpf_stab *stab) | |
446 | { | |
447 | bpf_map_area_free(stab->sock_map); | |
448 | kfree(stab); | |
449 | } | |
450 | ||
451 | static void smap_gc_work(struct work_struct *w) | |
452 | { | |
2f857d04 | 453 | struct smap_psock_map_entry *e, *tmp; |
174a79ff JF |
454 | struct smap_psock *psock; |
455 | ||
456 | psock = container_of(w, struct smap_psock, gc_work); | |
457 | ||
458 | /* no callback lock needed because we already detached sockmap ops */ | |
459 | if (psock->strp_enabled) | |
460 | strp_done(&psock->strp); | |
461 | ||
462 | cancel_work_sync(&psock->tx_work); | |
463 | __skb_queue_purge(&psock->rxqueue); | |
464 | ||
465 | /* At this point all strparser and xmit work must be complete */ | |
466 | if (psock->bpf_parse) | |
467 | bpf_prog_put(psock->bpf_parse); | |
468 | if (psock->bpf_verdict) | |
469 | bpf_prog_put(psock->bpf_verdict); | |
470 | ||
2f857d04 JF |
471 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { |
472 | list_del(&e->list); | |
473 | kfree(e); | |
474 | } | |
174a79ff JF |
475 | |
476 | sock_put(psock->sock); | |
477 | kfree(psock); | |
478 | } | |
479 | ||
480 | static struct smap_psock *smap_init_psock(struct sock *sock, | |
481 | struct bpf_stab *stab) | |
482 | { | |
483 | struct smap_psock *psock; | |
484 | ||
96eabe7a MKL |
485 | psock = kzalloc_node(sizeof(struct smap_psock), |
486 | GFP_ATOMIC | __GFP_NOWARN, | |
487 | stab->map.numa_node); | |
174a79ff JF |
488 | if (!psock) |
489 | return ERR_PTR(-ENOMEM); | |
490 | ||
491 | psock->sock = sock; | |
492 | skb_queue_head_init(&psock->rxqueue); | |
493 | INIT_WORK(&psock->tx_work, smap_tx_work); | |
494 | INIT_WORK(&psock->gc_work, smap_gc_work); | |
2f857d04 JF |
495 | INIT_LIST_HEAD(&psock->maps); |
496 | psock->refcnt = 1; | |
174a79ff JF |
497 | |
498 | rcu_assign_sk_user_data(sock, psock); | |
499 | sock_hold(sock); | |
500 | return psock; | |
501 | } | |
502 | ||
503 | static struct bpf_map *sock_map_alloc(union bpf_attr *attr) | |
504 | { | |
505 | struct bpf_stab *stab; | |
506 | int err = -EINVAL; | |
507 | u64 cost; | |
508 | ||
fb50df8d JF |
509 | if (!capable(CAP_NET_ADMIN)) |
510 | return ERR_PTR(-EPERM); | |
511 | ||
174a79ff JF |
512 | /* check sanity of attributes */ |
513 | if (attr->max_entries == 0 || attr->key_size != 4 || | |
cace572e | 514 | attr->value_size != 4 || attr->map_flags & ~SOCK_CREATE_FLAG_MASK) |
174a79ff JF |
515 | return ERR_PTR(-EINVAL); |
516 | ||
517 | if (attr->value_size > KMALLOC_MAX_SIZE) | |
518 | return ERR_PTR(-E2BIG); | |
519 | ||
520 | stab = kzalloc(sizeof(*stab), GFP_USER); | |
521 | if (!stab) | |
522 | return ERR_PTR(-ENOMEM); | |
523 | ||
524 | /* mandatory map attributes */ | |
525 | stab->map.map_type = attr->map_type; | |
526 | stab->map.key_size = attr->key_size; | |
527 | stab->map.value_size = attr->value_size; | |
528 | stab->map.max_entries = attr->max_entries; | |
529 | stab->map.map_flags = attr->map_flags; | |
96eabe7a | 530 | stab->map.numa_node = bpf_map_attr_numa_node(attr); |
174a79ff JF |
531 | |
532 | /* make sure page count doesn't overflow */ | |
533 | cost = (u64) stab->map.max_entries * sizeof(struct sock *); | |
534 | if (cost >= U32_MAX - PAGE_SIZE) | |
535 | goto free_stab; | |
536 | ||
537 | stab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT; | |
538 | ||
539 | /* if map size is larger than memlock limit, reject it early */ | |
540 | err = bpf_map_precharge_memlock(stab->map.pages); | |
541 | if (err) | |
542 | goto free_stab; | |
543 | ||
f740c34e | 544 | err = -ENOMEM; |
174a79ff | 545 | stab->sock_map = bpf_map_area_alloc(stab->map.max_entries * |
96eabe7a MKL |
546 | sizeof(struct sock *), |
547 | stab->map.numa_node); | |
174a79ff JF |
548 | if (!stab->sock_map) |
549 | goto free_stab; | |
550 | ||
174a79ff JF |
551 | return &stab->map; |
552 | free_stab: | |
553 | kfree(stab); | |
554 | return ERR_PTR(err); | |
555 | } | |
556 | ||
2f857d04 JF |
557 | static void smap_list_remove(struct smap_psock *psock, struct sock **entry) |
558 | { | |
559 | struct smap_psock_map_entry *e, *tmp; | |
560 | ||
561 | list_for_each_entry_safe(e, tmp, &psock->maps, list) { | |
562 | if (e->entry == entry) { | |
563 | list_del(&e->list); | |
564 | break; | |
565 | } | |
566 | } | |
567 | } | |
568 | ||
174a79ff JF |
569 | static void sock_map_free(struct bpf_map *map) |
570 | { | |
571 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
572 | int i; | |
573 | ||
574 | synchronize_rcu(); | |
575 | ||
576 | /* At this point no update, lookup or delete operations can happen. | |
577 | * However, be aware we can still get a socket state event updates, | |
578 | * and data ready callabacks that reference the psock from sk_user_data | |
579 | * Also psock worker threads are still in-flight. So smap_release_sock | |
580 | * will only free the psock after cancel_sync on the worker threads | |
581 | * and a grace period expire to ensure psock is really safe to remove. | |
582 | */ | |
583 | rcu_read_lock(); | |
584 | for (i = 0; i < stab->map.max_entries; i++) { | |
2f857d04 | 585 | struct smap_psock *psock; |
174a79ff JF |
586 | struct sock *sock; |
587 | ||
588 | sock = xchg(&stab->sock_map[i], NULL); | |
589 | if (!sock) | |
590 | continue; | |
591 | ||
2f857d04 JF |
592 | write_lock_bh(&sock->sk_callback_lock); |
593 | psock = smap_psock_sk(sock); | |
be2b8690 JF |
594 | /* This check handles a racing sock event that can get the |
595 | * sk_callback_lock before this case but after xchg happens | |
596 | * causing the refcnt to hit zero and sock user data (psock) | |
597 | * to be null and queued for garbage collection. | |
598 | */ | |
599 | if (likely(psock)) { | |
600 | smap_list_remove(psock, &stab->sock_map[i]); | |
601 | smap_release_sock(psock, sock); | |
602 | } | |
2f857d04 | 603 | write_unlock_bh(&sock->sk_callback_lock); |
174a79ff JF |
604 | } |
605 | rcu_read_unlock(); | |
606 | ||
607 | if (stab->bpf_verdict) | |
608 | bpf_prog_put(stab->bpf_verdict); | |
609 | if (stab->bpf_parse) | |
610 | bpf_prog_put(stab->bpf_parse); | |
611 | ||
2f857d04 | 612 | sock_map_remove_complete(stab); |
174a79ff JF |
613 | } |
614 | ||
615 | static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next_key) | |
616 | { | |
617 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
618 | u32 i = key ? *(u32 *)key : U32_MAX; | |
619 | u32 *next = (u32 *)next_key; | |
620 | ||
621 | if (i >= stab->map.max_entries) { | |
622 | *next = 0; | |
623 | return 0; | |
624 | } | |
625 | ||
626 | if (i == stab->map.max_entries - 1) | |
627 | return -ENOENT; | |
628 | ||
629 | *next = i + 1; | |
630 | return 0; | |
631 | } | |
632 | ||
633 | struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key) | |
634 | { | |
635 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
636 | ||
637 | if (key >= map->max_entries) | |
638 | return NULL; | |
639 | ||
640 | return READ_ONCE(stab->sock_map[key]); | |
641 | } | |
642 | ||
643 | static int sock_map_delete_elem(struct bpf_map *map, void *key) | |
644 | { | |
645 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
2f857d04 | 646 | struct smap_psock *psock; |
174a79ff JF |
647 | int k = *(u32 *)key; |
648 | struct sock *sock; | |
649 | ||
650 | if (k >= map->max_entries) | |
651 | return -EINVAL; | |
652 | ||
653 | sock = xchg(&stab->sock_map[k], NULL); | |
654 | if (!sock) | |
655 | return -EINVAL; | |
656 | ||
2f857d04 JF |
657 | write_lock_bh(&sock->sk_callback_lock); |
658 | psock = smap_psock_sk(sock); | |
659 | if (!psock) | |
660 | goto out; | |
661 | ||
662 | if (psock->bpf_parse) | |
663 | smap_stop_sock(psock, sock); | |
664 | smap_list_remove(psock, &stab->sock_map[k]); | |
665 | smap_release_sock(psock, sock); | |
666 | out: | |
667 | write_unlock_bh(&sock->sk_callback_lock); | |
174a79ff JF |
668 | return 0; |
669 | } | |
670 | ||
671 | /* Locking notes: Concurrent updates, deletes, and lookups are allowed and are | |
672 | * done inside rcu critical sections. This ensures on updates that the psock | |
673 | * will not be released via smap_release_sock() until concurrent updates/deletes | |
674 | * complete. All operations operate on sock_map using cmpxchg and xchg | |
675 | * operations to ensure we do not get stale references. Any reads into the | |
676 | * map must be done with READ_ONCE() because of this. | |
677 | * | |
678 | * A psock is destroyed via call_rcu and after any worker threads are cancelled | |
679 | * and syncd so we are certain all references from the update/lookup/delete | |
680 | * operations as well as references in the data path are no longer in use. | |
681 | * | |
2f857d04 JF |
682 | * Psocks may exist in multiple maps, but only a single set of parse/verdict |
683 | * programs may be inherited from the maps it belongs to. A reference count | |
684 | * is kept with the total number of references to the psock from all maps. The | |
685 | * psock will not be released until this reaches zero. The psock and sock | |
686 | * user data data use the sk_callback_lock to protect critical data structures | |
687 | * from concurrent access. This allows us to avoid two updates from modifying | |
688 | * the user data in sock and the lock is required anyways for modifying | |
689 | * callbacks, we simply increase its scope slightly. | |
174a79ff | 690 | * |
2f857d04 JF |
691 | * Rules to follow, |
692 | * - psock must always be read inside RCU critical section | |
693 | * - sk_user_data must only be modified inside sk_callback_lock and read | |
694 | * inside RCU critical section. | |
695 | * - psock->maps list must only be read & modified inside sk_callback_lock | |
696 | * - sock_map must use READ_ONCE and (cmp)xchg operations | |
697 | * - BPF verdict/parse programs must use READ_ONCE and xchg operations | |
174a79ff JF |
698 | */ |
699 | static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops, | |
700 | struct bpf_map *map, | |
2f857d04 | 701 | void *key, u64 flags) |
174a79ff JF |
702 | { |
703 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
2f857d04 | 704 | struct smap_psock_map_entry *e = NULL; |
174a79ff | 705 | struct bpf_prog *verdict, *parse; |
2f857d04 JF |
706 | struct sock *osock, *sock; |
707 | struct smap_psock *psock; | |
174a79ff | 708 | u32 i = *(u32 *)key; |
2f857d04 | 709 | int err; |
174a79ff JF |
710 | |
711 | if (unlikely(flags > BPF_EXIST)) | |
712 | return -EINVAL; | |
713 | ||
714 | if (unlikely(i >= stab->map.max_entries)) | |
715 | return -E2BIG; | |
716 | ||
174a79ff | 717 | sock = READ_ONCE(stab->sock_map[i]); |
2f857d04 JF |
718 | if (flags == BPF_EXIST && !sock) |
719 | return -ENOENT; | |
720 | else if (flags == BPF_NOEXIST && sock) | |
174a79ff | 721 | return -EEXIST; |
174a79ff | 722 | |
2f857d04 | 723 | sock = skops->sk; |
174a79ff | 724 | |
2f857d04 JF |
725 | /* 1. If sock map has BPF programs those will be inherited by the |
726 | * sock being added. If the sock is already attached to BPF programs | |
727 | * this results in an error. | |
728 | */ | |
729 | verdict = READ_ONCE(stab->bpf_verdict); | |
730 | parse = READ_ONCE(stab->bpf_parse); | |
174a79ff | 731 | |
2f857d04 | 732 | if (parse && verdict) { |
174a79ff JF |
733 | /* bpf prog refcnt may be zero if a concurrent attach operation |
734 | * removes the program after the above READ_ONCE() but before | |
735 | * we increment the refcnt. If this is the case abort with an | |
736 | * error. | |
737 | */ | |
738 | verdict = bpf_prog_inc_not_zero(stab->bpf_verdict); | |
739 | if (IS_ERR(verdict)) | |
740 | return PTR_ERR(verdict); | |
741 | ||
742 | parse = bpf_prog_inc_not_zero(stab->bpf_parse); | |
743 | if (IS_ERR(parse)) { | |
744 | bpf_prog_put(verdict); | |
745 | return PTR_ERR(parse); | |
746 | } | |
747 | } | |
748 | ||
2f857d04 JF |
749 | write_lock_bh(&sock->sk_callback_lock); |
750 | psock = smap_psock_sk(sock); | |
751 | ||
752 | /* 2. Do not allow inheriting programs if psock exists and has | |
753 | * already inherited programs. This would create confusion on | |
754 | * which parser/verdict program is running. If no psock exists | |
755 | * create one. Inside sk_callback_lock to ensure concurrent create | |
756 | * doesn't update user data. | |
757 | */ | |
758 | if (psock) { | |
759 | if (READ_ONCE(psock->bpf_parse) && parse) { | |
760 | err = -EBUSY; | |
761 | goto out_progs; | |
762 | } | |
763 | psock->refcnt++; | |
764 | } else { | |
174a79ff JF |
765 | psock = smap_init_psock(sock, stab); |
766 | if (IS_ERR(psock)) { | |
2f857d04 JF |
767 | err = PTR_ERR(psock); |
768 | goto out_progs; | |
174a79ff | 769 | } |
2f857d04 | 770 | |
174a79ff JF |
771 | set_bit(SMAP_TX_RUNNING, &psock->state); |
772 | } | |
773 | ||
2f857d04 JF |
774 | e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN); |
775 | if (!e) { | |
776 | err = -ENOMEM; | |
777 | goto out_progs; | |
778 | } | |
779 | e->entry = &stab->sock_map[i]; | |
780 | ||
781 | /* 3. At this point we have a reference to a valid psock that is | |
782 | * running. Attach any BPF programs needed. | |
783 | */ | |
784 | if (parse && verdict && !psock->strp_enabled) { | |
174a79ff JF |
785 | err = smap_init_sock(psock, sock); |
786 | if (err) | |
2f857d04 | 787 | goto out_free; |
174a79ff JF |
788 | smap_init_progs(psock, stab, verdict, parse); |
789 | smap_start_sock(psock, sock); | |
174a79ff JF |
790 | } |
791 | ||
2f857d04 JF |
792 | /* 4. Place psock in sockmap for use and stop any programs on |
793 | * the old sock assuming its not the same sock we are replacing | |
794 | * it with. Because we can only have a single set of programs if | |
795 | * old_sock has a strp we can stop it. | |
796 | */ | |
797 | list_add_tail(&e->list, &psock->maps); | |
798 | write_unlock_bh(&sock->sk_callback_lock); | |
174a79ff | 799 | |
2f857d04 JF |
800 | osock = xchg(&stab->sock_map[i], sock); |
801 | if (osock) { | |
802 | struct smap_psock *opsock = smap_psock_sk(osock); | |
803 | ||
804 | write_lock_bh(&osock->sk_callback_lock); | |
805 | if (osock != sock && parse) | |
806 | smap_stop_sock(opsock, osock); | |
807 | smap_list_remove(opsock, &stab->sock_map[i]); | |
808 | smap_release_sock(opsock, osock); | |
809 | write_unlock_bh(&osock->sk_callback_lock); | |
810 | } | |
174a79ff | 811 | return 0; |
2f857d04 JF |
812 | out_free: |
813 | smap_release_sock(psock, sock); | |
814 | out_progs: | |
815 | if (verdict) | |
816 | bpf_prog_put(verdict); | |
817 | if (parse) | |
818 | bpf_prog_put(parse); | |
174a79ff | 819 | write_unlock_bh(&sock->sk_callback_lock); |
2f857d04 | 820 | kfree(e); |
174a79ff JF |
821 | return err; |
822 | } | |
823 | ||
5a67da2a | 824 | int sock_map_prog(struct bpf_map *map, struct bpf_prog *prog, u32 type) |
174a79ff JF |
825 | { |
826 | struct bpf_stab *stab = container_of(map, struct bpf_stab, map); | |
464bc0fd | 827 | struct bpf_prog *orig; |
174a79ff | 828 | |
81374aaa JF |
829 | if (unlikely(map->map_type != BPF_MAP_TYPE_SOCKMAP)) |
830 | return -EINVAL; | |
831 | ||
464bc0fd JF |
832 | switch (type) { |
833 | case BPF_SK_SKB_STREAM_PARSER: | |
834 | orig = xchg(&stab->bpf_parse, prog); | |
835 | break; | |
836 | case BPF_SK_SKB_STREAM_VERDICT: | |
837 | orig = xchg(&stab->bpf_verdict, prog); | |
838 | break; | |
839 | default: | |
840 | return -EOPNOTSUPP; | |
841 | } | |
174a79ff | 842 | |
464bc0fd JF |
843 | if (orig) |
844 | bpf_prog_put(orig); | |
174a79ff JF |
845 | |
846 | return 0; | |
847 | } | |
848 | ||
849 | static void *sock_map_lookup(struct bpf_map *map, void *key) | |
850 | { | |
851 | return NULL; | |
852 | } | |
853 | ||
854 | static int sock_map_update_elem(struct bpf_map *map, | |
855 | void *key, void *value, u64 flags) | |
856 | { | |
857 | struct bpf_sock_ops_kern skops; | |
858 | u32 fd = *(u32 *)value; | |
859 | struct socket *socket; | |
860 | int err; | |
861 | ||
862 | socket = sockfd_lookup(fd, &err); | |
863 | if (!socket) | |
864 | return err; | |
865 | ||
866 | skops.sk = socket->sk; | |
867 | if (!skops.sk) { | |
868 | fput(socket->file); | |
869 | return -EINVAL; | |
870 | } | |
871 | ||
435bf0d3 JF |
872 | if (skops.sk->sk_type != SOCK_STREAM || |
873 | skops.sk->sk_protocol != IPPROTO_TCP) { | |
874 | fput(socket->file); | |
875 | return -EOPNOTSUPP; | |
876 | } | |
877 | ||
2f857d04 | 878 | err = sock_map_ctx_update_elem(&skops, map, key, flags); |
174a79ff JF |
879 | fput(socket->file); |
880 | return err; | |
881 | } | |
882 | ||
883 | const struct bpf_map_ops sock_map_ops = { | |
884 | .map_alloc = sock_map_alloc, | |
885 | .map_free = sock_map_free, | |
886 | .map_lookup_elem = sock_map_lookup, | |
887 | .map_get_next_key = sock_map_get_next_key, | |
888 | .map_update_elem = sock_map_update_elem, | |
889 | .map_delete_elem = sock_map_delete_elem, | |
174a79ff JF |
890 | }; |
891 | ||
2f857d04 JF |
892 | BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, bpf_sock, |
893 | struct bpf_map *, map, void *, key, u64, flags) | |
174a79ff JF |
894 | { |
895 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
2f857d04 | 896 | return sock_map_ctx_update_elem(bpf_sock, map, key, flags); |
174a79ff JF |
897 | } |
898 | ||
899 | const struct bpf_func_proto bpf_sock_map_update_proto = { | |
900 | .func = bpf_sock_map_update, | |
901 | .gpl_only = false, | |
902 | .pkt_access = true, | |
903 | .ret_type = RET_INTEGER, | |
904 | .arg1_type = ARG_PTR_TO_CTX, | |
905 | .arg2_type = ARG_CONST_MAP_PTR, | |
906 | .arg3_type = ARG_PTR_TO_MAP_KEY, | |
907 | .arg4_type = ARG_ANYTHING, | |
174a79ff | 908 | }; |