Commit | Line | Data |
---|---|---|
604326b4 DB |
1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | /* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */ | |
3 | ||
4 | #ifndef _LINUX_SKMSG_H | |
5 | #define _LINUX_SKMSG_H | |
6 | ||
7 | #include <linux/bpf.h> | |
8 | #include <linux/filter.h> | |
9 | #include <linux/scatterlist.h> | |
10 | #include <linux/skbuff.h> | |
11 | ||
12 | #include <net/sock.h> | |
13 | #include <net/tcp.h> | |
14 | #include <net/strparser.h> | |
15 | ||
16 | #define MAX_MSG_FRAGS MAX_SKB_FRAGS | |
17 | ||
18 | enum __sk_action { | |
19 | __SK_DROP = 0, | |
20 | __SK_PASS, | |
21 | __SK_REDIRECT, | |
22 | __SK_NONE, | |
23 | }; | |
24 | ||
25 | struct sk_msg_sg { | |
26 | u32 start; | |
27 | u32 curr; | |
28 | u32 end; | |
29 | u32 size; | |
30 | u32 copybreak; | |
31 | bool copy[MAX_MSG_FRAGS]; | |
d3b18ad3 JF |
32 | /* The extra element is used for chaining the front and sections when |
33 | * the list becomes partitioned (e.g. end < start). The crypto APIs | |
34 | * require the chaining. | |
35 | */ | |
36 | struct scatterlist data[MAX_MSG_FRAGS + 1]; | |
604326b4 DB |
37 | }; |
38 | ||
39 | struct sk_msg { | |
40 | struct sk_msg_sg sg; | |
41 | void *data; | |
42 | void *data_end; | |
43 | u32 apply_bytes; | |
44 | u32 cork_bytes; | |
45 | u32 flags; | |
46 | struct sk_buff *skb; | |
47 | struct sock *sk_redir; | |
48 | struct sock *sk; | |
49 | struct list_head list; | |
50 | }; | |
51 | ||
52 | struct sk_psock_progs { | |
53 | struct bpf_prog *msg_parser; | |
54 | struct bpf_prog *skb_parser; | |
55 | struct bpf_prog *skb_verdict; | |
56 | }; | |
57 | ||
58 | enum sk_psock_state_bits { | |
59 | SK_PSOCK_TX_ENABLED, | |
60 | }; | |
61 | ||
62 | struct sk_psock_link { | |
63 | struct list_head list; | |
64 | struct bpf_map *map; | |
65 | void *link_raw; | |
66 | }; | |
67 | ||
68 | struct sk_psock_parser { | |
69 | struct strparser strp; | |
70 | bool enabled; | |
71 | void (*saved_data_ready)(struct sock *sk); | |
72 | }; | |
73 | ||
74 | struct sk_psock_work_state { | |
75 | struct sk_buff *skb; | |
76 | u32 len; | |
77 | u32 off; | |
78 | }; | |
79 | ||
80 | struct sk_psock { | |
81 | struct sock *sk; | |
82 | struct sock *sk_redir; | |
83 | u32 apply_bytes; | |
84 | u32 cork_bytes; | |
85 | u32 eval; | |
86 | struct sk_msg *cork; | |
87 | struct sk_psock_progs progs; | |
88 | struct sk_psock_parser parser; | |
89 | struct sk_buff_head ingress_skb; | |
90 | struct list_head ingress_msg; | |
91 | unsigned long state; | |
92 | struct list_head link; | |
93 | spinlock_t link_lock; | |
94 | refcount_t refcnt; | |
95 | void (*saved_unhash)(struct sock *sk); | |
96 | void (*saved_close)(struct sock *sk, long timeout); | |
97 | void (*saved_write_space)(struct sock *sk); | |
98 | struct proto *sk_proto; | |
99 | struct sk_psock_work_state work_state; | |
100 | struct work_struct work; | |
101 | union { | |
102 | struct rcu_head rcu; | |
103 | struct work_struct gc; | |
104 | }; | |
105 | }; | |
106 | ||
107 | int sk_msg_alloc(struct sock *sk, struct sk_msg *msg, int len, | |
108 | int elem_first_coalesce); | |
d829e9c4 DB |
109 | int sk_msg_clone(struct sock *sk, struct sk_msg *dst, struct sk_msg *src, |
110 | u32 off, u32 len); | |
604326b4 DB |
111 | void sk_msg_trim(struct sock *sk, struct sk_msg *msg, int len); |
112 | int sk_msg_free(struct sock *sk, struct sk_msg *msg); | |
113 | int sk_msg_free_nocharge(struct sock *sk, struct sk_msg *msg); | |
114 | void sk_msg_free_partial(struct sock *sk, struct sk_msg *msg, u32 bytes); | |
115 | void sk_msg_free_partial_nocharge(struct sock *sk, struct sk_msg *msg, | |
116 | u32 bytes); | |
117 | ||
118 | void sk_msg_return(struct sock *sk, struct sk_msg *msg, int bytes); | |
d3b18ad3 | 119 | void sk_msg_return_zero(struct sock *sk, struct sk_msg *msg, int bytes); |
604326b4 DB |
120 | |
121 | int sk_msg_zerocopy_from_iter(struct sock *sk, struct iov_iter *from, | |
122 | struct sk_msg *msg, u32 bytes); | |
123 | int sk_msg_memcopy_from_iter(struct sock *sk, struct iov_iter *from, | |
124 | struct sk_msg *msg, u32 bytes); | |
125 | ||
126 | static inline void sk_msg_check_to_free(struct sk_msg *msg, u32 i, u32 bytes) | |
127 | { | |
128 | WARN_ON(i == msg->sg.end && bytes); | |
129 | } | |
130 | ||
131 | static inline void sk_msg_apply_bytes(struct sk_psock *psock, u32 bytes) | |
132 | { | |
133 | if (psock->apply_bytes) { | |
134 | if (psock->apply_bytes < bytes) | |
135 | psock->apply_bytes = 0; | |
136 | else | |
137 | psock->apply_bytes -= bytes; | |
138 | } | |
139 | } | |
140 | ||
141 | #define sk_msg_iter_var_prev(var) \ | |
142 | do { \ | |
143 | if (var == 0) \ | |
144 | var = MAX_MSG_FRAGS - 1; \ | |
145 | else \ | |
146 | var--; \ | |
147 | } while (0) | |
148 | ||
149 | #define sk_msg_iter_var_next(var) \ | |
150 | do { \ | |
151 | var++; \ | |
152 | if (var == MAX_MSG_FRAGS) \ | |
153 | var = 0; \ | |
154 | } while (0) | |
155 | ||
156 | #define sk_msg_iter_prev(msg, which) \ | |
157 | sk_msg_iter_var_prev(msg->sg.which) | |
158 | ||
159 | #define sk_msg_iter_next(msg, which) \ | |
160 | sk_msg_iter_var_next(msg->sg.which) | |
161 | ||
162 | static inline void sk_msg_clear_meta(struct sk_msg *msg) | |
163 | { | |
164 | memset(&msg->sg, 0, offsetofend(struct sk_msg_sg, copy)); | |
165 | } | |
166 | ||
167 | static inline void sk_msg_init(struct sk_msg *msg) | |
168 | { | |
d3b18ad3 | 169 | BUILD_BUG_ON(ARRAY_SIZE(msg->sg.data) - 1 != MAX_MSG_FRAGS); |
604326b4 | 170 | memset(msg, 0, sizeof(*msg)); |
d3b18ad3 | 171 | sg_init_marker(msg->sg.data, MAX_MSG_FRAGS); |
604326b4 DB |
172 | } |
173 | ||
174 | static inline void sk_msg_xfer(struct sk_msg *dst, struct sk_msg *src, | |
175 | int which, u32 size) | |
176 | { | |
177 | dst->sg.data[which] = src->sg.data[which]; | |
178 | dst->sg.data[which].length = size; | |
3f4c3127 | 179 | dst->sg.size += size; |
604326b4 DB |
180 | src->sg.data[which].length -= size; |
181 | src->sg.data[which].offset += size; | |
182 | } | |
183 | ||
d3b18ad3 JF |
184 | static inline void sk_msg_xfer_full(struct sk_msg *dst, struct sk_msg *src) |
185 | { | |
186 | memcpy(dst, src, sizeof(*src)); | |
187 | sk_msg_init(src); | |
188 | } | |
189 | ||
8734a162 JF |
190 | static inline bool sk_msg_full(const struct sk_msg *msg) |
191 | { | |
192 | return (msg->sg.end == msg->sg.start) && msg->sg.size; | |
193 | } | |
194 | ||
604326b4 DB |
195 | static inline u32 sk_msg_elem_used(const struct sk_msg *msg) |
196 | { | |
8734a162 JF |
197 | if (sk_msg_full(msg)) |
198 | return MAX_MSG_FRAGS; | |
199 | ||
604326b4 DB |
200 | return msg->sg.end >= msg->sg.start ? |
201 | msg->sg.end - msg->sg.start : | |
202 | msg->sg.end + (MAX_MSG_FRAGS - msg->sg.start); | |
203 | } | |
204 | ||
604326b4 DB |
205 | static inline struct scatterlist *sk_msg_elem(struct sk_msg *msg, int which) |
206 | { | |
207 | return &msg->sg.data[which]; | |
208 | } | |
209 | ||
6fff607e JF |
210 | static inline struct scatterlist sk_msg_elem_cpy(struct sk_msg *msg, int which) |
211 | { | |
212 | return msg->sg.data[which]; | |
213 | } | |
214 | ||
604326b4 DB |
215 | static inline struct page *sk_msg_page(struct sk_msg *msg, int which) |
216 | { | |
217 | return sg_page(sk_msg_elem(msg, which)); | |
218 | } | |
219 | ||
220 | static inline bool sk_msg_to_ingress(const struct sk_msg *msg) | |
221 | { | |
222 | return msg->flags & BPF_F_INGRESS; | |
223 | } | |
224 | ||
225 | static inline void sk_msg_compute_data_pointers(struct sk_msg *msg) | |
226 | { | |
227 | struct scatterlist *sge = sk_msg_elem(msg, msg->sg.start); | |
228 | ||
229 | if (msg->sg.copy[msg->sg.start]) { | |
230 | msg->data = NULL; | |
231 | msg->data_end = NULL; | |
232 | } else { | |
233 | msg->data = sg_virt(sge); | |
234 | msg->data_end = msg->data + sge->length; | |
235 | } | |
236 | } | |
237 | ||
238 | static inline void sk_msg_page_add(struct sk_msg *msg, struct page *page, | |
239 | u32 len, u32 offset) | |
240 | { | |
241 | struct scatterlist *sge; | |
242 | ||
243 | get_page(page); | |
244 | sge = sk_msg_elem(msg, msg->sg.end); | |
245 | sg_set_page(sge, page, len, offset); | |
246 | sg_unmark_end(sge); | |
247 | ||
248 | msg->sg.copy[msg->sg.end] = true; | |
249 | msg->sg.size += len; | |
250 | sk_msg_iter_next(msg, end); | |
251 | } | |
252 | ||
d3b18ad3 JF |
253 | static inline void sk_msg_sg_copy(struct sk_msg *msg, u32 i, bool copy_state) |
254 | { | |
255 | do { | |
256 | msg->sg.copy[i] = copy_state; | |
257 | sk_msg_iter_var_next(i); | |
258 | if (i == msg->sg.end) | |
259 | break; | |
260 | } while (1); | |
261 | } | |
262 | ||
263 | static inline void sk_msg_sg_copy_set(struct sk_msg *msg, u32 start) | |
264 | { | |
265 | sk_msg_sg_copy(msg, start, true); | |
266 | } | |
267 | ||
268 | static inline void sk_msg_sg_copy_clear(struct sk_msg *msg, u32 start) | |
269 | { | |
270 | sk_msg_sg_copy(msg, start, false); | |
271 | } | |
272 | ||
604326b4 DB |
273 | static inline struct sk_psock *sk_psock(const struct sock *sk) |
274 | { | |
275 | return rcu_dereference_sk_user_data(sk); | |
276 | } | |
277 | ||
604326b4 DB |
278 | static inline void sk_psock_queue_msg(struct sk_psock *psock, |
279 | struct sk_msg *msg) | |
280 | { | |
281 | list_add_tail(&msg->list, &psock->ingress_msg); | |
282 | } | |
283 | ||
d3b18ad3 JF |
284 | static inline bool sk_psock_queue_empty(const struct sk_psock *psock) |
285 | { | |
286 | return psock ? list_empty(&psock->ingress_msg) : true; | |
287 | } | |
288 | ||
604326b4 DB |
289 | static inline void sk_psock_report_error(struct sk_psock *psock, int err) |
290 | { | |
291 | struct sock *sk = psock->sk; | |
292 | ||
293 | sk->sk_err = err; | |
294 | sk->sk_error_report(sk); | |
295 | } | |
296 | ||
297 | struct sk_psock *sk_psock_init(struct sock *sk, int node); | |
298 | ||
299 | int sk_psock_init_strp(struct sock *sk, struct sk_psock *psock); | |
300 | void sk_psock_start_strp(struct sock *sk, struct sk_psock *psock); | |
301 | void sk_psock_stop_strp(struct sock *sk, struct sk_psock *psock); | |
302 | ||
303 | int sk_psock_msg_verdict(struct sock *sk, struct sk_psock *psock, | |
304 | struct sk_msg *msg); | |
305 | ||
306 | static inline struct sk_psock_link *sk_psock_init_link(void) | |
307 | { | |
308 | return kzalloc(sizeof(struct sk_psock_link), | |
309 | GFP_ATOMIC | __GFP_NOWARN); | |
310 | } | |
311 | ||
312 | static inline void sk_psock_free_link(struct sk_psock_link *link) | |
313 | { | |
314 | kfree(link); | |
315 | } | |
316 | ||
317 | struct sk_psock_link *sk_psock_link_pop(struct sk_psock *psock); | |
318 | #if defined(CONFIG_BPF_STREAM_PARSER) | |
319 | void sk_psock_unlink(struct sock *sk, struct sk_psock_link *link); | |
320 | #else | |
321 | static inline void sk_psock_unlink(struct sock *sk, | |
322 | struct sk_psock_link *link) | |
323 | { | |
324 | } | |
325 | #endif | |
326 | ||
327 | void __sk_psock_purge_ingress_msg(struct sk_psock *psock); | |
328 | ||
329 | static inline void sk_psock_cork_free(struct sk_psock *psock) | |
330 | { | |
331 | if (psock->cork) { | |
332 | sk_msg_free(psock->sk, psock->cork); | |
333 | kfree(psock->cork); | |
334 | psock->cork = NULL; | |
335 | } | |
336 | } | |
337 | ||
338 | static inline void sk_psock_update_proto(struct sock *sk, | |
339 | struct sk_psock *psock, | |
340 | struct proto *ops) | |
341 | { | |
342 | psock->saved_unhash = sk->sk_prot->unhash; | |
343 | psock->saved_close = sk->sk_prot->close; | |
344 | psock->saved_write_space = sk->sk_write_space; | |
345 | ||
346 | psock->sk_proto = sk->sk_prot; | |
347 | sk->sk_prot = ops; | |
348 | } | |
349 | ||
350 | static inline void sk_psock_restore_proto(struct sock *sk, | |
351 | struct sk_psock *psock) | |
352 | { | |
353 | if (psock->sk_proto) { | |
354 | sk->sk_prot = psock->sk_proto; | |
355 | psock->sk_proto = NULL; | |
356 | } | |
357 | } | |
358 | ||
359 | static inline void sk_psock_set_state(struct sk_psock *psock, | |
360 | enum sk_psock_state_bits bit) | |
361 | { | |
362 | set_bit(bit, &psock->state); | |
363 | } | |
364 | ||
365 | static inline void sk_psock_clear_state(struct sk_psock *psock, | |
366 | enum sk_psock_state_bits bit) | |
367 | { | |
368 | clear_bit(bit, &psock->state); | |
369 | } | |
370 | ||
371 | static inline bool sk_psock_test_state(const struct sk_psock *psock, | |
372 | enum sk_psock_state_bits bit) | |
373 | { | |
374 | return test_bit(bit, &psock->state); | |
375 | } | |
376 | ||
5032d079 JF |
377 | static inline struct sk_psock *sk_psock_get_checked(struct sock *sk) |
378 | { | |
379 | struct sk_psock *psock; | |
380 | ||
381 | rcu_read_lock(); | |
382 | psock = sk_psock(sk); | |
383 | if (psock) { | |
384 | if (sk->sk_prot->recvmsg != tcp_bpf_recvmsg) { | |
385 | psock = ERR_PTR(-EBUSY); | |
386 | goto out; | |
387 | } | |
388 | ||
389 | if (!refcount_inc_not_zero(&psock->refcnt)) | |
390 | psock = ERR_PTR(-EBUSY); | |
391 | } | |
392 | out: | |
393 | rcu_read_unlock(); | |
394 | return psock; | |
395 | } | |
396 | ||
604326b4 DB |
397 | static inline struct sk_psock *sk_psock_get(struct sock *sk) |
398 | { | |
399 | struct sk_psock *psock; | |
400 | ||
401 | rcu_read_lock(); | |
402 | psock = sk_psock(sk); | |
403 | if (psock && !refcount_inc_not_zero(&psock->refcnt)) | |
404 | psock = NULL; | |
405 | rcu_read_unlock(); | |
406 | return psock; | |
407 | } | |
408 | ||
409 | void sk_psock_stop(struct sock *sk, struct sk_psock *psock); | |
410 | void sk_psock_destroy(struct rcu_head *rcu); | |
411 | void sk_psock_drop(struct sock *sk, struct sk_psock *psock); | |
412 | ||
413 | static inline void sk_psock_put(struct sock *sk, struct sk_psock *psock) | |
414 | { | |
415 | if (refcount_dec_and_test(&psock->refcnt)) | |
416 | sk_psock_drop(sk, psock); | |
417 | } | |
418 | ||
419 | static inline void psock_set_prog(struct bpf_prog **pprog, | |
420 | struct bpf_prog *prog) | |
421 | { | |
422 | prog = xchg(pprog, prog); | |
423 | if (prog) | |
424 | bpf_prog_put(prog); | |
425 | } | |
426 | ||
427 | static inline void psock_progs_drop(struct sk_psock_progs *progs) | |
428 | { | |
429 | psock_set_prog(&progs->msg_parser, NULL); | |
430 | psock_set_prog(&progs->skb_parser, NULL); | |
431 | psock_set_prog(&progs->skb_verdict, NULL); | |
432 | } | |
433 | ||
434 | #endif /* _LINUX_SKMSG_H */ |