xref: /freebsd/sys/netinet/tcp_pcap.c (revision 3e5d0784b9b5296bda801add034b057ad68237f7)
1 /*-
2  * Copyright (c) 2015
3  *	Jonathan Looney. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 #include <sys/queue.h>
30 #include <sys/param.h>
31 #include <sys/types.h>
32 #include <sys/socket.h>
33 #include <sys/socketvar.h>
34 #include <sys/sysctl.h>
35 #include <sys/systm.h>
36 #include <sys/mbuf.h>
37 #include <sys/eventhandler.h>
38 #include <machine/atomic.h>
39 #include <netinet/in.h>
40 #include <netinet/in_pcb.h>
41 #include <netinet/tcp_var.h>
42 #include <netinet/tcp_pcap.h>
43 
44 #define M_LEADINGSPACE_NOWRITE(m)					\
45 	((m)->m_data - M_START(m))
46 
47 int tcp_pcap_aggressive_free = 1;
48 static int tcp_pcap_clusters_referenced_cur = 0;
49 static int tcp_pcap_clusters_referenced_max = 0;
50 
51 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_aggressive_free,
52 	CTLFLAG_RW, &tcp_pcap_aggressive_free, 0,
53 	"Free saved packets when the memory system comes under pressure");
54 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_clusters_referenced_cur,
55 	CTLFLAG_RD, &tcp_pcap_clusters_referenced_cur, 0,
56 	"Number of clusters currently referenced on TCP PCAP queues");
57 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_clusters_referenced_max,
58 	CTLFLAG_RW, &tcp_pcap_clusters_referenced_max, 0,
59 	"Maximum number of clusters allowed to be referenced on TCP PCAP "
60 	"queues");
61 
62 static int tcp_pcap_alloc_reuse_ext = 0;
63 static int tcp_pcap_alloc_reuse_mbuf = 0;
64 static int tcp_pcap_alloc_new_mbuf = 0;
65 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_reuse_ext,
66 	CTLFLAG_RD, &tcp_pcap_alloc_reuse_ext, 0,
67 	"Number of mbufs with external storage reused for the TCP PCAP "
68 	"functionality");
69 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_reuse_mbuf,
70 	CTLFLAG_RD, &tcp_pcap_alloc_reuse_mbuf, 0,
71 	"Number of mbufs with internal storage reused for the TCP PCAP "
72 	"functionality");
73 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_alloc_new_mbuf,
74 	CTLFLAG_RD, &tcp_pcap_alloc_new_mbuf, 0,
75 	"Number of new mbufs allocated for the TCP PCAP functionality");
76 
77 VNET_DEFINE(int, tcp_pcap_packets) = 0;
78 #define V_tcp_pcap_packets	VNET(tcp_pcap_packets)
79 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_pcap_packets,
80 	CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_pcap_packets), 0,
81 	"Default number of packets saved per direction per TCPCB");
82 
83 /* Initialize the values. */
84 static void
85 tcp_pcap_max_set(void)
86 {
87 
88 	tcp_pcap_clusters_referenced_max = nmbclusters / 4;
89 }
90 
91 void
92 tcp_pcap_init(void)
93 {
94 
95 	tcp_pcap_max_set();
96 	EVENTHANDLER_REGISTER(nmbclusters_change, tcp_pcap_max_set,
97 		NULL, EVENTHANDLER_PRI_ANY);
98 }
99 
100 /*
101  * If we are below the maximum allowed cluster references,
102  * increment the reference count and return TRUE. Otherwise,
103  * leave the reference count alone and return FALSE.
104  */
105 static __inline bool
106 tcp_pcap_take_cluster_reference(void)
107 {
108 	if (atomic_fetchadd_int(&tcp_pcap_clusters_referenced_cur, 1) >=
109 		tcp_pcap_clusters_referenced_max) {
110 		atomic_add_int(&tcp_pcap_clusters_referenced_cur, -1);
111 		return FALSE;
112 	}
113 	return TRUE;
114 }
115 
116 /*
117  * For all the external entries in m, apply the given adjustment.
118  * This can be used to adjust the counter when an mbuf chain is
119  * copied or freed.
120  */
121 static __inline void
122 tcp_pcap_adj_cluster_reference(struct mbuf *m, int adj)
123 {
124 	while (m) {
125 		if (m->m_flags & M_EXT)
126 			atomic_add_int(&tcp_pcap_clusters_referenced_cur, adj);
127 
128 		m = m->m_next;
129 	}
130 }
131 
132 /*
133  * Free all mbufs in a chain, decrementing the reference count as
134  * necessary.
135  *
136  * Functions in this file should use this instead of m_freem() when
137  * they are freeing mbuf chains that may contain clusters that were
138  * already included in tcp_pcap_clusters_referenced_cur.
139  */
140 static void
141 tcp_pcap_m_freem(struct mbuf *mb)
142 {
143 	while (mb != NULL) {
144 		if (mb->m_flags & M_EXT)
145 			atomic_subtract_int(&tcp_pcap_clusters_referenced_cur,
146 			    1);
147 		mb = m_free(mb);
148 	}
149 }
150 
151 /*
152  * Copy data from m to n, where n cannot fit all the data we might
153  * want from m.
154  *
155  * Prioritize data like this:
156  * 1. TCP header
157  * 2. IP header
158  * 3. Data
159  */
160 static void
161 tcp_pcap_copy_bestfit(struct tcphdr *th, struct mbuf *m, struct mbuf *n)
162 {
163 	struct mbuf *m_cur = m;
164 	int bytes_to_copy=0, trailing_data, skip=0, tcp_off;
165 
166 	/* Below, we assume these will be non-NULL. */
167 	KASSERT(th, ("%s: called with th == NULL", __func__));
168 	KASSERT(m, ("%s: called with m == NULL", __func__));
169 	KASSERT(n, ("%s: called with n == NULL", __func__));
170 
171 	/* We assume this initialization occurred elsewhere. */
172 	KASSERT(n->m_len == 0, ("%s: called with n->m_len=%d (expected 0)",
173 		__func__, n->m_len));
174 	KASSERT(n->m_data == M_START(n),
175 		("%s: called with n->m_data != M_START(n)", __func__));
176 
177 	/*
178 	 * Calculate the size of the TCP header. We use this often
179 	 * enough that it is worth just calculating at the start.
180 	 */
181 	tcp_off = th->th_off << 2;
182 
183 	/* Trim off leading empty mbufs. */
184 	while (m && m->m_len == 0)
185 		m = m->m_next;
186 
187 	if (m) {
188 		m_cur = m;
189 	}
190 	else {
191 		/*
192 		 * No data? Highly unusual. We would expect to at
193 		 * least see a TCP header in the mbuf.
194 		 * As we have a pointer to the TCP header, I guess
195 		 * we should just copy that. (???)
196 		 */
197 fallback:
198 		bytes_to_copy = tcp_off;
199 		if (bytes_to_copy > M_SIZE(n))
200 			bytes_to_copy = M_SIZE(n);
201 		bcopy(th, n->m_data, bytes_to_copy);
202 		n->m_len = bytes_to_copy;
203 		return;
204 	}
205 
206 	/*
207 	 * Find TCP header. Record the total number of bytes up to,
208 	 * and including, the TCP header.
209 	 */
210 	while (m_cur) {
211 		if ((caddr_t) th >= (caddr_t) m_cur->m_data &&
212 			(caddr_t) th < (caddr_t) (m_cur->m_data + m_cur->m_len))
213 			break;
214 		bytes_to_copy += m_cur->m_len;
215 		m_cur = m_cur->m_next;
216 	}
217 	if (m_cur)
218 		bytes_to_copy += (caddr_t) th - (caddr_t) m_cur->m_data;
219 	else
220 		goto fallback;
221 	bytes_to_copy += tcp_off;
222 
223 	/*
224 	 * If we already want to copy more bytes than we can hold
225 	 * in the destination mbuf, skip leading bytes and copy
226 	 * what we can.
227 	 *
228 	 * Otherwise, consider trailing data.
229 	 */
230 	if (bytes_to_copy > M_SIZE(n)) {
231 		skip  = bytes_to_copy - M_SIZE(n);
232 		bytes_to_copy = M_SIZE(n);
233 	}
234 	else {
235 		/*
236 		 * Determine how much trailing data is in the chain.
237 		 * We start with the length of this mbuf (the one
238 		 * containing th) and subtract the size of the TCP
239 		 * header (tcp_off) and the size of the data prior
240 		 * to th (th - m_cur->m_data).
241 		 *
242 		 * This *should not* be negative, as the TCP code
243 		 * should put the whole TCP header in a single
244 		 * mbuf. But, it isn't a problem if it is. We will
245 		 * simple work off our negative balance as we look
246 		 * at subsequent mbufs.
247 		 */
248 		trailing_data = m_cur->m_len - tcp_off;
249 		trailing_data -= (caddr_t) th - (caddr_t) m_cur->m_data;
250 		m_cur = m_cur->m_next;
251 		while (m_cur) {
252 			trailing_data += m_cur->m_len;
253 			m_cur = m_cur->m_next;
254 		}
255 		if ((bytes_to_copy + trailing_data) > M_SIZE(n))
256 			bytes_to_copy = M_SIZE(n);
257 		else
258 			bytes_to_copy += trailing_data;
259 	}
260 
261 	m_copydata(m, skip, bytes_to_copy, n->m_data);
262 	n->m_len = bytes_to_copy;
263 }
264 
265 void
266 tcp_pcap_add(struct tcphdr *th, struct mbuf *m, struct mbufq *queue)
267 {
268 	struct mbuf *n = NULL, *mhead;
269 
270 	KASSERT(th, ("%s: called with th == NULL", __func__));
271 	KASSERT(m, ("%s: called with m == NULL", __func__));
272 	KASSERT(queue, ("%s: called with queue == NULL", __func__));
273 
274 	/* We only care about data packets. */
275 	while (m && m->m_type != MT_DATA)
276 		m = m->m_next;
277 
278 	/* We only need to do something if we still have an mbuf. */
279 	if (!m)
280 		return;
281 
282 	/* If we are not saving mbufs, return now. */
283 	if (queue->mq_maxlen == 0)
284 		return;
285 
286 	/*
287 	 * Check to see if we will need to recycle mbufs.
288 	 *
289 	 * If we need to get rid of mbufs to stay below
290 	 * our packet count, try to reuse the mbuf. Once
291 	 * we already have a new mbuf (n), then we can
292 	 * simply free subsequent mbufs.
293 	 *
294 	 * Note that most of the logic in here is to deal
295 	 * with the reuse. If we are fine with constant
296 	 * mbuf allocs/deallocs, we could ditch this logic.
297 	 * But, it only seems to make sense to reuse
298 	 * mbufs we already have.
299 	 */
300 	while (mbufq_full(queue)) {
301 		mhead = mbufq_dequeue(queue);
302 
303 		if (n) {
304 			tcp_pcap_m_freem(mhead);
305 		}
306 		else {
307 			/*
308 			 * If this held an external cluster, try to
309 			 * detach the cluster. But, if we held the
310 			 * last reference, go through the normal
311 			 * free-ing process.
312 			 */
313 			if (mhead->m_flags & M_EXTPG) {
314 				/* Don't mess around with these. */
315 				tcp_pcap_m_freem(mhead);
316 				continue;
317 			} else if (mhead->m_flags & M_EXT) {
318 				switch (mhead->m_ext.ext_type) {
319 				case EXT_SFBUF:
320 					/* Don't mess around with these. */
321 					tcp_pcap_m_freem(mhead);
322 					continue;
323 				default:
324 					if (atomic_fetchadd_int(
325 						mhead->m_ext.ext_cnt, -1) == 1)
326 					{
327 						/*
328 						 * We held the last reference
329 						 * on this cluster. Restore
330 						 * the reference count and put
331 						 * it back in the pool.
332 				 		 */
333 						*(mhead->m_ext.ext_cnt) = 1;
334 						tcp_pcap_m_freem(mhead);
335 						continue;
336 					}
337 					/*
338 					 * We were able to cleanly free the
339 					 * reference.
340 				 	 */
341 					atomic_subtract_int(
342 					    &tcp_pcap_clusters_referenced_cur,
343 					    1);
344 					tcp_pcap_alloc_reuse_ext++;
345 					break;
346 				}
347 			} else {
348 				tcp_pcap_alloc_reuse_mbuf++;
349 			}
350 
351 			n = mhead;
352 			tcp_pcap_m_freem(n->m_next);
353 			m_init(n, M_NOWAIT, MT_DATA, 0);
354 		}
355 	}
356 
357 	/* Check to see if we need to get a new mbuf. */
358 	if (!n) {
359 		if (!(n = m_get(M_NOWAIT, MT_DATA)))
360 			return;
361 		tcp_pcap_alloc_new_mbuf++;
362 	}
363 
364 	/*
365 	 * What are we dealing with? If a cluster, attach it. Otherwise,
366 	 * try to copy the data from the beginning of the mbuf to the
367 	 * end of data. (There may be data between the start of the data
368 	 * area and the current data pointer. We want to get this, because
369 	 * it may contain header information that is useful.)
370 	 * In cases where that isn't possible, settle for what we can
371 	 * get.
372 	 */
373 	if ((m->m_flags & (M_EXT|M_EXTPG)) &&
374 	    tcp_pcap_take_cluster_reference()) {
375 		n->m_data = m->m_data;
376 		n->m_len = m->m_len;
377 		mb_dupcl(n, m);
378 	}
379 	else if (((m->m_data + m->m_len) - M_START(m)) <= M_SIZE(n)) {
380 		/*
381 		 * At this point, n is guaranteed to be a normal mbuf
382 		 * with no cluster and no packet header. Because the
383 		 * logic in this code block requires this, the assert
384 		 * is here to catch any instances where someone
385 		 * changes the logic to invalidate that assumption.
386 		 */
387 		KASSERT((n->m_flags & (M_EXT | M_PKTHDR)) == 0,
388 			("%s: Unexpected flags (%#x) for mbuf",
389 			__func__, n->m_flags));
390 		n->m_data = n->m_dat + M_LEADINGSPACE_NOWRITE(m);
391 		n->m_len = m->m_len;
392 		if (m->m_flags & M_EXTPG)
393 			m_copydata(m, 0, m->m_len, n->m_data);
394 		else
395 			bcopy(M_START(m), n->m_dat,
396 			    m->m_len + M_LEADINGSPACE_NOWRITE(m));
397 	}
398 	else {
399 		/*
400 		 * This is the case where we need to "settle for what
401 		 * we can get". The most probable way to this code
402 		 * path is that we've already taken references to the
403 		 * maximum number of mbuf clusters we can, and the data
404 		 * is too long to fit in an mbuf's internal storage.
405 		 * Try for a "best fit".
406 		 */
407 		tcp_pcap_copy_bestfit(th, m, n);
408 
409 		/* Don't try to get additional data. */
410 		goto add_to_queue;
411 	}
412 
413 	if (m->m_next) {
414 		n->m_next = m_copym(m->m_next, 0, M_COPYALL, M_NOWAIT);
415 		tcp_pcap_adj_cluster_reference(n->m_next, 1);
416 	}
417 
418 add_to_queue:
419 	/* Add the new mbuf to the list. */
420 	if (mbufq_enqueue(queue, n)) {
421 		/* This shouldn't happen. If INVARIANTS is defined, panic. */
422 		KASSERT(0, ("%s: mbufq was unexpectedly full!", __func__));
423 		tcp_pcap_m_freem(n);
424 	}
425 }
426 
427 void
428 tcp_pcap_drain(struct mbufq *queue)
429 {
430 	struct mbuf *m;
431 	while ((m = mbufq_dequeue(queue)))
432 		tcp_pcap_m_freem(m);
433 }
434 
435 void
436 tcp_pcap_tcpcb_init(struct tcpcb *tp)
437 {
438 	mbufq_init(&(tp->t_inpkts), V_tcp_pcap_packets);
439 	mbufq_init(&(tp->t_outpkts), V_tcp_pcap_packets);
440 }
441 
442 void
443 tcp_pcap_set_sock_max(struct mbufq *queue, int newval)
444 {
445 	queue->mq_maxlen = newval;
446 	while (queue->mq_len > queue->mq_maxlen)
447 		tcp_pcap_m_freem(mbufq_dequeue(queue));
448 }
449 
450 int
451 tcp_pcap_get_sock_max(struct mbufq *queue)
452 {
453 	return queue->mq_maxlen;
454 }
455