xref: /openbmc/linux/net/dccp/ackvec.c (revision c21b37f6)
1 /*
2  *  net/dccp/ackvec.c
3  *
4  *  An implementation of the DCCP protocol
5  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
6  *
7  *      This program is free software; you can redistribute it and/or modify it
8  *      under the terms of the GNU General Public License as published by the
9  *      Free Software Foundation; version 2 of the License;
10  */
11 
12 #include "ackvec.h"
13 #include "dccp.h"
14 
15 #include <linux/dccp.h>
16 #include <linux/init.h>
17 #include <linux/errno.h>
18 #include <linux/kernel.h>
19 #include <linux/skbuff.h>
20 #include <linux/slab.h>
21 
22 #include <net/sock.h>
23 
24 static struct kmem_cache *dccp_ackvec_slab;
25 static struct kmem_cache *dccp_ackvec_record_slab;
26 
27 static struct dccp_ackvec_record *dccp_ackvec_record_new(void)
28 {
29 	struct dccp_ackvec_record *avr =
30 			kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC);
31 
32 	if (avr != NULL)
33 		INIT_LIST_HEAD(&avr->dccpavr_node);
34 
35 	return avr;
36 }
37 
38 static void dccp_ackvec_record_delete(struct dccp_ackvec_record *avr)
39 {
40 	if (unlikely(avr == NULL))
41 		return;
42 	/* Check if deleting a linked record */
43 	WARN_ON(!list_empty(&avr->dccpavr_node));
44 	kmem_cache_free(dccp_ackvec_record_slab, avr);
45 }
46 
47 static void dccp_ackvec_insert_avr(struct dccp_ackvec *av,
48 				   struct dccp_ackvec_record *avr)
49 {
50 	/*
51 	 * AVRs are sorted by seqno. Since we are sending them in order, we
52 	 * just add the AVR at the head of the list.
53 	 * -sorbo.
54 	 */
55 	if (!list_empty(&av->dccpav_records)) {
56 		const struct dccp_ackvec_record *head =
57 					list_entry(av->dccpav_records.next,
58 						   struct dccp_ackvec_record,
59 						   dccpavr_node);
60 		BUG_ON(before48(avr->dccpavr_ack_seqno,
61 				head->dccpavr_ack_seqno));
62 	}
63 
64 	list_add(&avr->dccpavr_node, &av->dccpav_records);
65 }
66 
67 int dccp_insert_option_ackvec(struct sock *sk, struct sk_buff *skb)
68 {
69 	struct dccp_sock *dp = dccp_sk(sk);
70 	struct dccp_ackvec *av = dp->dccps_hc_rx_ackvec;
71 	/* Figure out how many options do we need to represent the ackvec */
72 	const u16 nr_opts = (av->dccpav_vec_len +
73 			     DCCP_MAX_ACKVEC_OPT_LEN - 1) /
74 			    DCCP_MAX_ACKVEC_OPT_LEN;
75 	u16 len = av->dccpav_vec_len + 2 * nr_opts, i;
76 	struct timeval now;
77 	u32 elapsed_time;
78 	const unsigned char *tail, *from;
79 	unsigned char *to;
80 	struct dccp_ackvec_record *avr;
81 
82 	if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
83 		return -1;
84 
85 	dccp_timestamp(sk, &now);
86 	elapsed_time = timeval_delta(&now, &av->dccpav_time) / 10;
87 
88 	if (elapsed_time != 0 &&
89 	    dccp_insert_option_elapsed_time(sk, skb, elapsed_time))
90 		return -1;
91 
92 	avr = dccp_ackvec_record_new();
93 	if (avr == NULL)
94 		return -1;
95 
96 	DCCP_SKB_CB(skb)->dccpd_opt_len += len;
97 
98 	to   = skb_push(skb, len);
99 	len  = av->dccpav_vec_len;
100 	from = av->dccpav_buf + av->dccpav_buf_head;
101 	tail = av->dccpav_buf + DCCP_MAX_ACKVEC_LEN;
102 
103 	for (i = 0; i < nr_opts; ++i) {
104 		int copylen = len;
105 
106 		if (len > DCCP_MAX_ACKVEC_OPT_LEN)
107 			copylen = DCCP_MAX_ACKVEC_OPT_LEN;
108 
109 		*to++ = DCCPO_ACK_VECTOR_0;
110 		*to++ = copylen + 2;
111 
112 		/* Check if buf_head wraps */
113 		if (from + copylen > tail) {
114 			const u16 tailsize = tail - from;
115 
116 			memcpy(to, from, tailsize);
117 			to	+= tailsize;
118 			len	-= tailsize;
119 			copylen	-= tailsize;
120 			from	= av->dccpav_buf;
121 		}
122 
123 		memcpy(to, from, copylen);
124 		from += copylen;
125 		to   += copylen;
126 		len  -= copylen;
127 	}
128 
129 	/*
130 	 *	From RFC 4340, A.2:
131 	 *
132 	 *	For each acknowledgement it sends, the HC-Receiver will add an
133 	 *	acknowledgement record.  ack_seqno will equal the HC-Receiver
134 	 *	sequence number it used for the ack packet; ack_ptr will equal
135 	 *	buf_head; ack_ackno will equal buf_ackno; and ack_nonce will
136 	 *	equal buf_nonce.
137 	 */
138 	avr->dccpavr_ack_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
139 	avr->dccpavr_ack_ptr   = av->dccpav_buf_head;
140 	avr->dccpavr_ack_ackno = av->dccpav_buf_ackno;
141 	avr->dccpavr_ack_nonce = av->dccpav_buf_nonce;
142 	avr->dccpavr_sent_len  = av->dccpav_vec_len;
143 
144 	dccp_ackvec_insert_avr(av, avr);
145 
146 	dccp_pr_debug("%s ACK Vector 0, len=%d, ack_seqno=%llu, "
147 		      "ack_ackno=%llu\n",
148 		      dccp_role(sk), avr->dccpavr_sent_len,
149 		      (unsigned long long)avr->dccpavr_ack_seqno,
150 		      (unsigned long long)avr->dccpavr_ack_ackno);
151 	return 0;
152 }
153 
154 struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority)
155 {
156 	struct dccp_ackvec *av = kmem_cache_alloc(dccp_ackvec_slab, priority);
157 
158 	if (av != NULL) {
159 		av->dccpav_buf_head	= DCCP_MAX_ACKVEC_LEN - 1;
160 		av->dccpav_buf_ackno	= UINT48_MAX + 1;
161 		av->dccpav_buf_nonce = av->dccpav_buf_nonce = 0;
162 		av->dccpav_time.tv_sec	= 0;
163 		av->dccpav_time.tv_usec	= 0;
164 		av->dccpav_vec_len	= 0;
165 		INIT_LIST_HEAD(&av->dccpav_records);
166 	}
167 
168 	return av;
169 }
170 
171 void dccp_ackvec_free(struct dccp_ackvec *av)
172 {
173 	if (unlikely(av == NULL))
174 		return;
175 
176 	if (!list_empty(&av->dccpav_records)) {
177 		struct dccp_ackvec_record *avr, *next;
178 
179 		list_for_each_entry_safe(avr, next, &av->dccpav_records,
180 					 dccpavr_node) {
181 			list_del_init(&avr->dccpavr_node);
182 			dccp_ackvec_record_delete(avr);
183 		}
184 	}
185 
186 	kmem_cache_free(dccp_ackvec_slab, av);
187 }
188 
189 static inline u8 dccp_ackvec_state(const struct dccp_ackvec *av,
190 				   const u32 index)
191 {
192 	return av->dccpav_buf[index] & DCCP_ACKVEC_STATE_MASK;
193 }
194 
195 static inline u8 dccp_ackvec_len(const struct dccp_ackvec *av,
196 				 const u32 index)
197 {
198 	return av->dccpav_buf[index] & DCCP_ACKVEC_LEN_MASK;
199 }
200 
201 /*
202  * If several packets are missing, the HC-Receiver may prefer to enter multiple
203  * bytes with run length 0, rather than a single byte with a larger run length;
204  * this simplifies table updates if one of the missing packets arrives.
205  */
206 static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av,
207 						 const unsigned int packets,
208 						 const unsigned char state)
209 {
210 	unsigned int gap;
211 	long new_head;
212 
213 	if (av->dccpav_vec_len + packets > DCCP_MAX_ACKVEC_LEN)
214 		return -ENOBUFS;
215 
216 	gap	 = packets - 1;
217 	new_head = av->dccpav_buf_head - packets;
218 
219 	if (new_head < 0) {
220 		if (gap > 0) {
221 			memset(av->dccpav_buf, DCCP_ACKVEC_STATE_NOT_RECEIVED,
222 			       gap + new_head + 1);
223 			gap = -new_head;
224 		}
225 		new_head += DCCP_MAX_ACKVEC_LEN;
226 	}
227 
228 	av->dccpav_buf_head = new_head;
229 
230 	if (gap > 0)
231 		memset(av->dccpav_buf + av->dccpav_buf_head + 1,
232 		       DCCP_ACKVEC_STATE_NOT_RECEIVED, gap);
233 
234 	av->dccpav_buf[av->dccpav_buf_head] = state;
235 	av->dccpav_vec_len += packets;
236 	return 0;
237 }
238 
239 /*
240  * Implements the RFC 4340, Appendix A
241  */
242 int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk,
243 		    const u64 ackno, const u8 state)
244 {
245 	/*
246 	 * Check at the right places if the buffer is full, if it is, tell the
247 	 * caller to start dropping packets till the HC-Sender acks our ACK
248 	 * vectors, when we will free up space in dccpav_buf.
249 	 *
250 	 * We may well decide to do buffer compression, etc, but for now lets
251 	 * just drop.
252 	 *
253 	 * From Appendix A.1.1 (`New Packets'):
254 	 *
255 	 *	Of course, the circular buffer may overflow, either when the
256 	 *	HC-Sender is sending data at a very high rate, when the
257 	 *	HC-Receiver's acknowledgements are not reaching the HC-Sender,
258 	 *	or when the HC-Sender is forgetting to acknowledge those acks
259 	 *	(so the HC-Receiver is unable to clean up old state). In this
260 	 *	case, the HC-Receiver should either compress the buffer (by
261 	 *	increasing run lengths when possible), transfer its state to
262 	 *	a larger buffer, or, as a last resort, drop all received
263 	 *	packets, without processing them whatsoever, until its buffer
264 	 *	shrinks again.
265 	 */
266 
267 	/* See if this is the first ackno being inserted */
268 	if (av->dccpav_vec_len == 0) {
269 		av->dccpav_buf[av->dccpav_buf_head] = state;
270 		av->dccpav_vec_len = 1;
271 	} else if (after48(ackno, av->dccpav_buf_ackno)) {
272 		const u64 delta = dccp_delta_seqno(av->dccpav_buf_ackno,
273 						   ackno);
274 
275 		/*
276 		 * Look if the state of this packet is the same as the
277 		 * previous ackno and if so if we can bump the head len.
278 		 */
279 		if (delta == 1 &&
280 		    dccp_ackvec_state(av, av->dccpav_buf_head) == state &&
281 		    (dccp_ackvec_len(av, av->dccpav_buf_head) <
282 		     DCCP_ACKVEC_LEN_MASK))
283 			av->dccpav_buf[av->dccpav_buf_head]++;
284 		else if (dccp_ackvec_set_buf_head_state(av, delta, state))
285 			return -ENOBUFS;
286 	} else {
287 		/*
288 		 * A.1.2.  Old Packets
289 		 *
290 		 *	When a packet with Sequence Number S <= buf_ackno
291 		 *	arrives, the HC-Receiver will scan the table for
292 		 *	the byte corresponding to S. (Indexing structures
293 		 *	could reduce the complexity of this scan.)
294 		 */
295 		u64 delta = dccp_delta_seqno(ackno, av->dccpav_buf_ackno);
296 		u32 index = av->dccpav_buf_head;
297 
298 		while (1) {
299 			const u8 len = dccp_ackvec_len(av, index);
300 			const u8 state = dccp_ackvec_state(av, index);
301 			/*
302 			 * valid packets not yet in dccpav_buf have a reserved
303 			 * entry, with a len equal to 0.
304 			 */
305 			if (state == DCCP_ACKVEC_STATE_NOT_RECEIVED &&
306 			    len == 0 && delta == 0) { /* Found our
307 							 reserved seat! */
308 				dccp_pr_debug("Found %llu reserved seat!\n",
309 					      (unsigned long long)ackno);
310 				av->dccpav_buf[index] = state;
311 				goto out;
312 			}
313 			/* len == 0 means one packet */
314 			if (delta < len + 1)
315 				goto out_duplicate;
316 
317 			delta -= len + 1;
318 			if (++index == DCCP_MAX_ACKVEC_LEN)
319 				index = 0;
320 		}
321 	}
322 
323 	av->dccpav_buf_ackno = ackno;
324 	dccp_timestamp(sk, &av->dccpav_time);
325 out:
326 	return 0;
327 
328 out_duplicate:
329 	/* Duplicate packet */
330 	dccp_pr_debug("Received a dup or already considered lost "
331 		      "packet: %llu\n", (unsigned long long)ackno);
332 	return -EILSEQ;
333 }
334 
335 #ifdef CONFIG_IP_DCCP_DEBUG
336 void dccp_ackvector_print(const u64 ackno, const unsigned char *vector, int len)
337 {
338 	dccp_pr_debug_cat("ACK vector len=%d, ackno=%llu |", len,
339 			 (unsigned long long)ackno);
340 
341 	while (len--) {
342 		const u8 state = (*vector & DCCP_ACKVEC_STATE_MASK) >> 6;
343 		const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
344 
345 		dccp_pr_debug_cat("%d,%d|", state, rl);
346 		++vector;
347 	}
348 
349 	dccp_pr_debug_cat("\n");
350 }
351 
352 void dccp_ackvec_print(const struct dccp_ackvec *av)
353 {
354 	dccp_ackvector_print(av->dccpav_buf_ackno,
355 			     av->dccpav_buf + av->dccpav_buf_head,
356 			     av->dccpav_vec_len);
357 }
358 #endif
359 
360 static void dccp_ackvec_throw_record(struct dccp_ackvec *av,
361 				     struct dccp_ackvec_record *avr)
362 {
363 	struct dccp_ackvec_record *next;
364 
365 	/* sort out vector length */
366 	if (av->dccpav_buf_head <= avr->dccpavr_ack_ptr)
367 		av->dccpav_vec_len = avr->dccpavr_ack_ptr - av->dccpav_buf_head;
368 	else
369 		av->dccpav_vec_len = DCCP_MAX_ACKVEC_LEN - 1
370 				     - av->dccpav_buf_head
371 				     + avr->dccpavr_ack_ptr;
372 
373 	/* free records */
374 	list_for_each_entry_safe_from(avr, next, &av->dccpav_records,
375 				      dccpavr_node) {
376 		list_del_init(&avr->dccpavr_node);
377 		dccp_ackvec_record_delete(avr);
378 	}
379 }
380 
381 void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk,
382 				 const u64 ackno)
383 {
384 	struct dccp_ackvec_record *avr;
385 
386 	/*
387 	 * If we traverse backwards, it should be faster when we have large
388 	 * windows. We will be receiving ACKs for stuff we sent a while back
389 	 * -sorbo.
390 	 */
391 	list_for_each_entry_reverse(avr, &av->dccpav_records, dccpavr_node) {
392 		if (ackno == avr->dccpavr_ack_seqno) {
393 			dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, "
394 				      "ack_ackno=%llu, ACKED!\n",
395 				      dccp_role(sk), 1,
396 				      (unsigned long long)avr->dccpavr_ack_seqno,
397 				      (unsigned long long)avr->dccpavr_ack_ackno);
398 			dccp_ackvec_throw_record(av, avr);
399 			break;
400 		} else if (avr->dccpavr_ack_seqno > ackno)
401 			break; /* old news */
402 	}
403 }
404 
405 static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av,
406 					    struct sock *sk, u64 *ackno,
407 					    const unsigned char len,
408 					    const unsigned char *vector)
409 {
410 	unsigned char i;
411 	struct dccp_ackvec_record *avr;
412 
413 	/* Check if we actually sent an ACK vector */
414 	if (list_empty(&av->dccpav_records))
415 		return;
416 
417 	i = len;
418 	/*
419 	 * XXX
420 	 * I think it might be more efficient to work backwards. See comment on
421 	 * rcv_ackno. -sorbo.
422 	 */
423 	avr = list_entry(av->dccpav_records.next, struct dccp_ackvec_record,
424 			 dccpavr_node);
425 	while (i--) {
426 		const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
427 		u64 ackno_end_rl;
428 
429 		dccp_set_seqno(&ackno_end_rl, *ackno - rl);
430 
431 		/*
432 		 * If our AVR sequence number is greater than the ack, go
433 		 * forward in the AVR list until it is not so.
434 		 */
435 		list_for_each_entry_from(avr, &av->dccpav_records,
436 					 dccpavr_node) {
437 			if (!after48(avr->dccpavr_ack_seqno, *ackno))
438 				goto found;
439 		}
440 		/* End of the dccpav_records list, not found, exit */
441 		break;
442 found:
443 		if (between48(avr->dccpavr_ack_seqno, ackno_end_rl, *ackno)) {
444 			const u8 state = *vector & DCCP_ACKVEC_STATE_MASK;
445 			if (state != DCCP_ACKVEC_STATE_NOT_RECEIVED) {
446 				dccp_pr_debug("%s ACK vector 0, len=%d, "
447 					      "ack_seqno=%llu, ack_ackno=%llu, "
448 					      "ACKED!\n",
449 					      dccp_role(sk), len,
450 					      (unsigned long long)
451 					      avr->dccpavr_ack_seqno,
452 					      (unsigned long long)
453 					      avr->dccpavr_ack_ackno);
454 				dccp_ackvec_throw_record(av, avr);
455 				break;
456 			}
457 			/*
458 			 * If it wasn't received, continue scanning... we might
459 			 * find another one.
460 			 */
461 		}
462 
463 		dccp_set_seqno(ackno, ackno_end_rl - 1);
464 		++vector;
465 	}
466 }
467 
468 int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb,
469 		      u64 *ackno, const u8 opt, const u8 *value, const u8 len)
470 {
471 	if (len > DCCP_MAX_ACKVEC_OPT_LEN)
472 		return -1;
473 
474 	/* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */
475 	dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk,
476 					ackno, len, value);
477 	return 0;
478 }
479 
480 int __init dccp_ackvec_init(void)
481 {
482 	dccp_ackvec_slab = kmem_cache_create("dccp_ackvec",
483 					     sizeof(struct dccp_ackvec), 0,
484 					     SLAB_HWCACHE_ALIGN, NULL);
485 	if (dccp_ackvec_slab == NULL)
486 		goto out_err;
487 
488 	dccp_ackvec_record_slab =
489 			kmem_cache_create("dccp_ackvec_record",
490 					  sizeof(struct dccp_ackvec_record),
491 					  0, SLAB_HWCACHE_ALIGN, NULL);
492 	if (dccp_ackvec_record_slab == NULL)
493 		goto out_destroy_slab;
494 
495 	return 0;
496 
497 out_destroy_slab:
498 	kmem_cache_destroy(dccp_ackvec_slab);
499 	dccp_ackvec_slab = NULL;
500 out_err:
501 	DCCP_CRIT("Unable to create Ack Vector slab cache");
502 	return -ENOBUFS;
503 }
504 
505 void dccp_ackvec_exit(void)
506 {
507 	if (dccp_ackvec_slab != NULL) {
508 		kmem_cache_destroy(dccp_ackvec_slab);
509 		dccp_ackvec_slab = NULL;
510 	}
511 	if (dccp_ackvec_record_slab != NULL) {
512 		kmem_cache_destroy(dccp_ackvec_record_slab);
513 		dccp_ackvec_record_slab = NULL;
514 	}
515 }
516