xref: /freebsd/sys/netinet/siftr.c (revision ff0ba87247820afbdfdc1b307c803f7923d0e4d3)
1 /*-
2  * Copyright (c) 2007-2009
3  * 	Swinburne University of Technology, Melbourne, Australia.
4  * Copyright (c) 2009-2010, The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * Portions of this software were developed at the Centre for Advanced
8  * Internet Architectures, Swinburne University of Technology, Melbourne,
9  * Australia by Lawrence Stewart under sponsorship from the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 /******************************************************
34  * Statistical Information For TCP Research (SIFTR)
35  *
36  * A FreeBSD kernel module that adds very basic intrumentation to the
37  * TCP stack, allowing internal stats to be recorded to a log file
38  * for experimental, debugging and performance analysis purposes.
39  *
40  * SIFTR was first released in 2007 by James Healy and Lawrence Stewart whilst
41  * working on the NewTCP research project at Swinburne University of
42  * Technology's Centre for Advanced Internet Architectures, Melbourne,
43  * Australia, which was made possible in part by a grant from the Cisco
44  * University Research Program Fund at Community Foundation Silicon Valley.
45  * More details are available at:
46  *   http://caia.swin.edu.au/urp/newtcp/
47  *
48  * Work on SIFTR v1.2.x was sponsored by the FreeBSD Foundation as part of
49  * the "Enhancing the FreeBSD TCP Implementation" project 2008-2009.
50  * More details are available at:
51  *   http://www.freebsdfoundation.org/
52  *   http://caia.swin.edu.au/freebsd/etcp09/
53  *
54  * Lawrence Stewart is the current maintainer, and all contact regarding
55  * SIFTR should be directed to him via email: lastewart@swin.edu.au
56  *
57  * Initial release date: June 2007
58  * Most recent update: September 2010
59  ******************************************************/
60 
61 #include <sys/cdefs.h>
62 __FBSDID("$FreeBSD$");
63 
64 #include <sys/param.h>
65 #include <sys/alq.h>
66 #include <sys/errno.h>
67 #include <sys/eventhandler.h>
68 #include <sys/hash.h>
69 #include <sys/kernel.h>
70 #include <sys/kthread.h>
71 #include <sys/lock.h>
72 #include <sys/mbuf.h>
73 #include <sys/module.h>
74 #include <sys/mutex.h>
75 #include <sys/pcpu.h>
76 #include <sys/proc.h>
77 #include <sys/sbuf.h>
78 #include <sys/smp.h>
79 #include <sys/socket.h>
80 #include <sys/socketvar.h>
81 #include <sys/sysctl.h>
82 #include <sys/unistd.h>
83 
84 #include <net/if.h>
85 #include <net/if_var.h>
86 #include <net/pfil.h>
87 
88 #include <netinet/in.h>
89 #include <netinet/in_pcb.h>
90 #include <netinet/in_systm.h>
91 #include <netinet/in_var.h>
92 #include <netinet/ip.h>
93 #include <netinet/tcp_var.h>
94 
95 #ifdef SIFTR_IPV6
96 #include <netinet/ip6.h>
97 #include <netinet6/in6_pcb.h>
98 #endif /* SIFTR_IPV6 */
99 
100 #include <machine/in_cksum.h>
101 
102 /*
103  * Three digit version number refers to X.Y.Z where:
104  * X is the major version number
105  * Y is bumped to mark backwards incompatible changes
106  * Z is bumped to mark backwards compatible changes
107  */
108 #define V_MAJOR		1
109 #define V_BACKBREAK	2
110 #define V_BACKCOMPAT	4
111 #define MODVERSION	__CONCAT(V_MAJOR, __CONCAT(V_BACKBREAK, V_BACKCOMPAT))
112 #define MODVERSION_STR	__XSTRING(V_MAJOR) "." __XSTRING(V_BACKBREAK) "." \
113     __XSTRING(V_BACKCOMPAT)
114 
115 #define HOOK 0
116 #define UNHOOK 1
117 #define SIFTR_EXPECTED_MAX_TCP_FLOWS 65536
118 #define SYS_NAME "FreeBSD"
119 #define PACKET_TAG_SIFTR 100
120 #define PACKET_COOKIE_SIFTR 21749576
121 #define SIFTR_LOG_FILE_MODE 0644
122 #define SIFTR_DISABLE 0
123 #define SIFTR_ENABLE 1
124 
125 /*
126  * Hard upper limit on the length of log messages. Bump this up if you add new
127  * data fields such that the line length could exceed the below value.
128  */
129 #define MAX_LOG_MSG_LEN 200
130 /* XXX: Make this a sysctl tunable. */
131 #define SIFTR_ALQ_BUFLEN (1000*MAX_LOG_MSG_LEN)
132 
133 /*
134  * 1 byte for IP version
135  * IPv4: src/dst IP (4+4) + src/dst port (2+2) = 12 bytes
136  * IPv6: src/dst IP (16+16) + src/dst port (2+2) = 36 bytes
137  */
138 #ifdef SIFTR_IPV6
139 #define FLOW_KEY_LEN 37
140 #else
141 #define FLOW_KEY_LEN 13
142 #endif
143 
144 #ifdef SIFTR_IPV6
145 #define SIFTR_IPMODE 6
146 #else
147 #define SIFTR_IPMODE 4
148 #endif
149 
150 /* useful macros */
151 #define CAST_PTR_INT(X) (*((int*)(X)))
152 
153 #define UPPER_SHORT(X)	(((X) & 0xFFFF0000) >> 16)
154 #define LOWER_SHORT(X)	((X) & 0x0000FFFF)
155 
156 #define FIRST_OCTET(X)	(((X) & 0xFF000000) >> 24)
157 #define SECOND_OCTET(X)	(((X) & 0x00FF0000) >> 16)
158 #define THIRD_OCTET(X)	(((X) & 0x0000FF00) >> 8)
159 #define FOURTH_OCTET(X)	((X) & 0x000000FF)
160 
161 static MALLOC_DEFINE(M_SIFTR, "siftr", "dynamic memory used by SIFTR");
162 static MALLOC_DEFINE(M_SIFTR_PKTNODE, "siftr_pktnode",
163     "SIFTR pkt_node struct");
164 static MALLOC_DEFINE(M_SIFTR_HASHNODE, "siftr_hashnode",
165     "SIFTR flow_hash_node struct");
166 
167 /* Used as links in the pkt manager queue. */
168 struct pkt_node {
169 	/* Timestamp of pkt as noted in the pfil hook. */
170 	struct timeval		tval;
171 	/* Direction pkt is travelling; either PFIL_IN or PFIL_OUT. */
172 	uint8_t			direction;
173 	/* IP version pkt_node relates to; either INP_IPV4 or INP_IPV6. */
174 	uint8_t			ipver;
175 	/* Hash of the pkt which triggered the log message. */
176 	uint32_t		hash;
177 	/* Local/foreign IP address. */
178 #ifdef SIFTR_IPV6
179 	uint32_t		ip_laddr[4];
180 	uint32_t		ip_faddr[4];
181 #else
182 	uint8_t			ip_laddr[4];
183 	uint8_t			ip_faddr[4];
184 #endif
185 	/* Local TCP port. */
186 	uint16_t		tcp_localport;
187 	/* Foreign TCP port. */
188 	uint16_t		tcp_foreignport;
189 	/* Congestion Window (bytes). */
190 	u_long			snd_cwnd;
191 	/* Sending Window (bytes). */
192 	u_long			snd_wnd;
193 	/* Receive Window (bytes). */
194 	u_long			rcv_wnd;
195 	/* Unused (was: Bandwidth Controlled Window (bytes)). */
196 	u_long			snd_bwnd;
197 	/* Slow Start Threshold (bytes). */
198 	u_long			snd_ssthresh;
199 	/* Current state of the TCP FSM. */
200 	int			conn_state;
201 	/* Max Segment Size (bytes). */
202 	u_int			max_seg_size;
203 	/*
204 	 * Smoothed RTT stored as found in the TCP control block
205 	 * in units of (TCP_RTT_SCALE*hz).
206 	 */
207 	int			smoothed_rtt;
208 	/* Is SACK enabled? */
209 	u_char			sack_enabled;
210 	/* Window scaling for snd window. */
211 	u_char			snd_scale;
212 	/* Window scaling for recv window. */
213 	u_char			rcv_scale;
214 	/* TCP control block flags. */
215 	u_int			flags;
216 	/* Retransmit timeout length. */
217 	int			rxt_length;
218 	/* Size of the TCP send buffer in bytes. */
219 	u_int			snd_buf_hiwater;
220 	/* Current num bytes in the send socket buffer. */
221 	u_int			snd_buf_cc;
222 	/* Size of the TCP receive buffer in bytes. */
223 	u_int			rcv_buf_hiwater;
224 	/* Current num bytes in the receive socket buffer. */
225 	u_int			rcv_buf_cc;
226 	/* Number of bytes inflight that we are waiting on ACKs for. */
227 	u_int			sent_inflight_bytes;
228 	/* Number of segments currently in the reassembly queue. */
229 	int			t_segqlen;
230 	/* Link to next pkt_node in the list. */
231 	STAILQ_ENTRY(pkt_node)	nodes;
232 };
233 
234 struct flow_hash_node
235 {
236 	uint16_t counter;
237 	uint8_t key[FLOW_KEY_LEN];
238 	LIST_ENTRY(flow_hash_node) nodes;
239 };
240 
241 struct siftr_stats
242 {
243 	/* # TCP pkts seen by the SIFTR PFIL hooks, including any skipped. */
244 	uint64_t n_in;
245 	uint64_t n_out;
246 	/* # pkts skipped due to failed malloc calls. */
247 	uint32_t nskip_in_malloc;
248 	uint32_t nskip_out_malloc;
249 	/* # pkts skipped due to failed mtx acquisition. */
250 	uint32_t nskip_in_mtx;
251 	uint32_t nskip_out_mtx;
252 	/* # pkts skipped due to failed inpcb lookups. */
253 	uint32_t nskip_in_inpcb;
254 	uint32_t nskip_out_inpcb;
255 	/* # pkts skipped due to failed tcpcb lookups. */
256 	uint32_t nskip_in_tcpcb;
257 	uint32_t nskip_out_tcpcb;
258 	/* # pkts skipped due to stack reinjection. */
259 	uint32_t nskip_in_dejavu;
260 	uint32_t nskip_out_dejavu;
261 };
262 
263 static DPCPU_DEFINE(struct siftr_stats, ss);
264 
265 static volatile unsigned int siftr_exit_pkt_manager_thread = 0;
266 static unsigned int siftr_enabled = 0;
267 static unsigned int siftr_pkts_per_log = 1;
268 static unsigned int siftr_generate_hashes = 0;
269 /* static unsigned int siftr_binary_log = 0; */
270 static char siftr_logfile[PATH_MAX] = "/var/log/siftr.log";
271 static char siftr_logfile_shadow[PATH_MAX] = "/var/log/siftr.log";
272 static u_long siftr_hashmask;
273 STAILQ_HEAD(pkthead, pkt_node) pkt_queue = STAILQ_HEAD_INITIALIZER(pkt_queue);
274 LIST_HEAD(listhead, flow_hash_node) *counter_hash;
275 static int wait_for_pkt;
276 static struct alq *siftr_alq = NULL;
277 static struct mtx siftr_pkt_queue_mtx;
278 static struct mtx siftr_pkt_mgr_mtx;
279 static struct thread *siftr_pkt_manager_thr = NULL;
280 /*
281  * pfil.h defines PFIL_IN as 1 and PFIL_OUT as 2,
282  * which we use as an index into this array.
283  */
284 static char direction[3] = {'\0', 'i','o'};
285 
286 /* Required function prototypes. */
287 static int siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS);
288 static int siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS);
289 
290 
291 /* Declare the net.inet.siftr sysctl tree and populate it. */
292 
293 SYSCTL_DECL(_net_inet_siftr);
294 
295 SYSCTL_NODE(_net_inet, OID_AUTO, siftr, CTLFLAG_RW, NULL,
296     "siftr related settings");
297 
298 SYSCTL_PROC(_net_inet_siftr, OID_AUTO, enabled, CTLTYPE_UINT|CTLFLAG_RW,
299     &siftr_enabled, 0, &siftr_sysctl_enabled_handler, "IU",
300     "switch siftr module operations on/off");
301 
302 SYSCTL_PROC(_net_inet_siftr, OID_AUTO, logfile, CTLTYPE_STRING|CTLFLAG_RW,
303     &siftr_logfile_shadow, sizeof(siftr_logfile_shadow), &siftr_sysctl_logfile_name_handler,
304     "A", "file to save siftr log messages to");
305 
306 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, ppl, CTLFLAG_RW,
307     &siftr_pkts_per_log, 1,
308     "number of packets between generating a log message");
309 
310 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, genhashes, CTLFLAG_RW,
311     &siftr_generate_hashes, 0,
312     "enable packet hash generation");
313 
314 /* XXX: TODO
315 SYSCTL_UINT(_net_inet_siftr, OID_AUTO, binary, CTLFLAG_RW,
316     &siftr_binary_log, 0,
317     "write log files in binary instead of ascii");
318 */
319 
320 
321 /* Begin functions. */
322 
323 static void
324 siftr_process_pkt(struct pkt_node * pkt_node)
325 {
326 	struct flow_hash_node *hash_node;
327 	struct listhead *counter_list;
328 	struct siftr_stats *ss;
329 	struct ale *log_buf;
330 	uint8_t key[FLOW_KEY_LEN];
331 	uint8_t found_match, key_offset;
332 
333 	hash_node = NULL;
334 	ss = DPCPU_PTR(ss);
335 	found_match = 0;
336 	key_offset = 1;
337 
338 	/*
339 	 * Create the key that will be used to create a hash index
340 	 * into our hash table. Our key consists of:
341 	 * ipversion, localip, localport, foreignip, foreignport
342 	 */
343 	key[0] = pkt_node->ipver;
344 	memcpy(key + key_offset, &pkt_node->ip_laddr,
345 	    sizeof(pkt_node->ip_laddr));
346 	key_offset += sizeof(pkt_node->ip_laddr);
347 	memcpy(key + key_offset, &pkt_node->tcp_localport,
348 	    sizeof(pkt_node->tcp_localport));
349 	key_offset += sizeof(pkt_node->tcp_localport);
350 	memcpy(key + key_offset, &pkt_node->ip_faddr,
351 	    sizeof(pkt_node->ip_faddr));
352 	key_offset += sizeof(pkt_node->ip_faddr);
353 	memcpy(key + key_offset, &pkt_node->tcp_foreignport,
354 	    sizeof(pkt_node->tcp_foreignport));
355 
356 	counter_list = counter_hash +
357 	    (hash32_buf(key, sizeof(key), 0) & siftr_hashmask);
358 
359 	/*
360 	 * If the list is not empty i.e. the hash index has
361 	 * been used by another flow previously.
362 	 */
363 	if (LIST_FIRST(counter_list) != NULL) {
364 		/*
365 		 * Loop through the hash nodes in the list.
366 		 * There should normally only be 1 hash node in the list,
367 		 * except if there have been collisions at the hash index
368 		 * computed by hash32_buf().
369 		 */
370 		LIST_FOREACH(hash_node, counter_list, nodes) {
371 			/*
372 			 * Check if the key for the pkt we are currently
373 			 * processing is the same as the key stored in the
374 			 * hash node we are currently processing.
375 			 * If they are the same, then we've found the
376 			 * hash node that stores the counter for the flow
377 			 * the pkt belongs to.
378 			 */
379 			if (memcmp(hash_node->key, key, sizeof(key)) == 0) {
380 				found_match = 1;
381 				break;
382 			}
383 		}
384 	}
385 
386 	/* If this flow hash hasn't been seen before or we have a collision. */
387 	if (hash_node == NULL || !found_match) {
388 		/* Create a new hash node to store the flow's counter. */
389 		hash_node = malloc(sizeof(struct flow_hash_node),
390 		    M_SIFTR_HASHNODE, M_WAITOK);
391 
392 		if (hash_node != NULL) {
393 			/* Initialise our new hash node list entry. */
394 			hash_node->counter = 0;
395 			memcpy(hash_node->key, key, sizeof(key));
396 			LIST_INSERT_HEAD(counter_list, hash_node, nodes);
397 		} else {
398 			/* Malloc failed. */
399 			if (pkt_node->direction == PFIL_IN)
400 				ss->nskip_in_malloc++;
401 			else
402 				ss->nskip_out_malloc++;
403 
404 			return;
405 		}
406 	} else if (siftr_pkts_per_log > 1) {
407 		/*
408 		 * Taking the remainder of the counter divided
409 		 * by the current value of siftr_pkts_per_log
410 		 * and storing that in counter provides a neat
411 		 * way to modulate the frequency of log
412 		 * messages being written to the log file.
413 		 */
414 		hash_node->counter = (hash_node->counter + 1) %
415 		    siftr_pkts_per_log;
416 
417 		/*
418 		 * If we have not seen enough packets since the last time
419 		 * we wrote a log message for this connection, return.
420 		 */
421 		if (hash_node->counter > 0)
422 			return;
423 	}
424 
425 	log_buf = alq_getn(siftr_alq, MAX_LOG_MSG_LEN, ALQ_WAITOK);
426 
427 	if (log_buf == NULL)
428 		return; /* Should only happen if the ALQ is shutting down. */
429 
430 #ifdef SIFTR_IPV6
431 	pkt_node->ip_laddr[3] = ntohl(pkt_node->ip_laddr[3]);
432 	pkt_node->ip_faddr[3] = ntohl(pkt_node->ip_faddr[3]);
433 
434 	if (pkt_node->ipver == INP_IPV6) { /* IPv6 packet */
435 		pkt_node->ip_laddr[0] = ntohl(pkt_node->ip_laddr[0]);
436 		pkt_node->ip_laddr[1] = ntohl(pkt_node->ip_laddr[1]);
437 		pkt_node->ip_laddr[2] = ntohl(pkt_node->ip_laddr[2]);
438 		pkt_node->ip_faddr[0] = ntohl(pkt_node->ip_faddr[0]);
439 		pkt_node->ip_faddr[1] = ntohl(pkt_node->ip_faddr[1]);
440 		pkt_node->ip_faddr[2] = ntohl(pkt_node->ip_faddr[2]);
441 
442 		/* Construct an IPv6 log message. */
443 		log_buf->ae_bytesused = snprintf(log_buf->ae_data,
444 		    MAX_LOG_MSG_LEN,
445 		    "%c,0x%08x,%zd.%06ld,%x:%x:%x:%x:%x:%x:%x:%x,%u,%x:%x:%x:"
446 		    "%x:%x:%x:%x:%x,%u,%ld,%ld,%ld,%ld,%ld,%u,%u,%u,%u,%u,%u,"
447 		    "%u,%d,%u,%u,%u,%u,%u,%u\n",
448 		    direction[pkt_node->direction],
449 		    pkt_node->hash,
450 		    pkt_node->tval.tv_sec,
451 		    pkt_node->tval.tv_usec,
452 		    UPPER_SHORT(pkt_node->ip_laddr[0]),
453 		    LOWER_SHORT(pkt_node->ip_laddr[0]),
454 		    UPPER_SHORT(pkt_node->ip_laddr[1]),
455 		    LOWER_SHORT(pkt_node->ip_laddr[1]),
456 		    UPPER_SHORT(pkt_node->ip_laddr[2]),
457 		    LOWER_SHORT(pkt_node->ip_laddr[2]),
458 		    UPPER_SHORT(pkt_node->ip_laddr[3]),
459 		    LOWER_SHORT(pkt_node->ip_laddr[3]),
460 		    ntohs(pkt_node->tcp_localport),
461 		    UPPER_SHORT(pkt_node->ip_faddr[0]),
462 		    LOWER_SHORT(pkt_node->ip_faddr[0]),
463 		    UPPER_SHORT(pkt_node->ip_faddr[1]),
464 		    LOWER_SHORT(pkt_node->ip_faddr[1]),
465 		    UPPER_SHORT(pkt_node->ip_faddr[2]),
466 		    LOWER_SHORT(pkt_node->ip_faddr[2]),
467 		    UPPER_SHORT(pkt_node->ip_faddr[3]),
468 		    LOWER_SHORT(pkt_node->ip_faddr[3]),
469 		    ntohs(pkt_node->tcp_foreignport),
470 		    pkt_node->snd_ssthresh,
471 		    pkt_node->snd_cwnd,
472 		    pkt_node->snd_bwnd,
473 		    pkt_node->snd_wnd,
474 		    pkt_node->rcv_wnd,
475 		    pkt_node->snd_scale,
476 		    pkt_node->rcv_scale,
477 		    pkt_node->conn_state,
478 		    pkt_node->max_seg_size,
479 		    pkt_node->smoothed_rtt,
480 		    pkt_node->sack_enabled,
481 		    pkt_node->flags,
482 		    pkt_node->rxt_length,
483 		    pkt_node->snd_buf_hiwater,
484 		    pkt_node->snd_buf_cc,
485 		    pkt_node->rcv_buf_hiwater,
486 		    pkt_node->rcv_buf_cc,
487 		    pkt_node->sent_inflight_bytes,
488 		    pkt_node->t_segqlen);
489 	} else { /* IPv4 packet */
490 		pkt_node->ip_laddr[0] = FIRST_OCTET(pkt_node->ip_laddr[3]);
491 		pkt_node->ip_laddr[1] = SECOND_OCTET(pkt_node->ip_laddr[3]);
492 		pkt_node->ip_laddr[2] = THIRD_OCTET(pkt_node->ip_laddr[3]);
493 		pkt_node->ip_laddr[3] = FOURTH_OCTET(pkt_node->ip_laddr[3]);
494 		pkt_node->ip_faddr[0] = FIRST_OCTET(pkt_node->ip_faddr[3]);
495 		pkt_node->ip_faddr[1] = SECOND_OCTET(pkt_node->ip_faddr[3]);
496 		pkt_node->ip_faddr[2] = THIRD_OCTET(pkt_node->ip_faddr[3]);
497 		pkt_node->ip_faddr[3] = FOURTH_OCTET(pkt_node->ip_faddr[3]);
498 #endif /* SIFTR_IPV6 */
499 
500 		/* Construct an IPv4 log message. */
501 		log_buf->ae_bytesused = snprintf(log_buf->ae_data,
502 		    MAX_LOG_MSG_LEN,
503 		    "%c,0x%08x,%jd.%06ld,%u.%u.%u.%u,%u,%u.%u.%u.%u,%u,%ld,%ld,"
504 		    "%ld,%ld,%ld,%u,%u,%u,%u,%u,%u,%u,%d,%u,%u,%u,%u,%u,%u\n",
505 		    direction[pkt_node->direction],
506 		    pkt_node->hash,
507 		    (intmax_t)pkt_node->tval.tv_sec,
508 		    pkt_node->tval.tv_usec,
509 		    pkt_node->ip_laddr[0],
510 		    pkt_node->ip_laddr[1],
511 		    pkt_node->ip_laddr[2],
512 		    pkt_node->ip_laddr[3],
513 		    ntohs(pkt_node->tcp_localport),
514 		    pkt_node->ip_faddr[0],
515 		    pkt_node->ip_faddr[1],
516 		    pkt_node->ip_faddr[2],
517 		    pkt_node->ip_faddr[3],
518 		    ntohs(pkt_node->tcp_foreignport),
519 		    pkt_node->snd_ssthresh,
520 		    pkt_node->snd_cwnd,
521 		    pkt_node->snd_bwnd,
522 		    pkt_node->snd_wnd,
523 		    pkt_node->rcv_wnd,
524 		    pkt_node->snd_scale,
525 		    pkt_node->rcv_scale,
526 		    pkt_node->conn_state,
527 		    pkt_node->max_seg_size,
528 		    pkt_node->smoothed_rtt,
529 		    pkt_node->sack_enabled,
530 		    pkt_node->flags,
531 		    pkt_node->rxt_length,
532 		    pkt_node->snd_buf_hiwater,
533 		    pkt_node->snd_buf_cc,
534 		    pkt_node->rcv_buf_hiwater,
535 		    pkt_node->rcv_buf_cc,
536 		    pkt_node->sent_inflight_bytes,
537 		    pkt_node->t_segqlen);
538 #ifdef SIFTR_IPV6
539 	}
540 #endif
541 
542 	alq_post_flags(siftr_alq, log_buf, 0);
543 }
544 
545 
546 static void
547 siftr_pkt_manager_thread(void *arg)
548 {
549 	STAILQ_HEAD(pkthead, pkt_node) tmp_pkt_queue =
550 	    STAILQ_HEAD_INITIALIZER(tmp_pkt_queue);
551 	struct pkt_node *pkt_node, *pkt_node_temp;
552 	uint8_t draining;
553 
554 	draining = 2;
555 
556 	mtx_lock(&siftr_pkt_mgr_mtx);
557 
558 	/* draining == 0 when queue has been flushed and it's safe to exit. */
559 	while (draining) {
560 		/*
561 		 * Sleep until we are signalled to wake because thread has
562 		 * been told to exit or until 1 tick has passed.
563 		 */
564 		mtx_sleep(&wait_for_pkt, &siftr_pkt_mgr_mtx, PWAIT, "pktwait",
565 		    1);
566 
567 		/* Gain exclusive access to the pkt_node queue. */
568 		mtx_lock(&siftr_pkt_queue_mtx);
569 
570 		/*
571 		 * Move pkt_queue to tmp_pkt_queue, which leaves
572 		 * pkt_queue empty and ready to receive more pkt_nodes.
573 		 */
574 		STAILQ_CONCAT(&tmp_pkt_queue, &pkt_queue);
575 
576 		/*
577 		 * We've finished making changes to the list. Unlock it
578 		 * so the pfil hooks can continue queuing pkt_nodes.
579 		 */
580 		mtx_unlock(&siftr_pkt_queue_mtx);
581 
582 		/*
583 		 * We can't hold a mutex whilst calling siftr_process_pkt
584 		 * because ALQ might sleep waiting for buffer space.
585 		 */
586 		mtx_unlock(&siftr_pkt_mgr_mtx);
587 
588 		/* Flush all pkt_nodes to the log file. */
589 		STAILQ_FOREACH_SAFE(pkt_node, &tmp_pkt_queue, nodes,
590 		    pkt_node_temp) {
591 			siftr_process_pkt(pkt_node);
592 			STAILQ_REMOVE_HEAD(&tmp_pkt_queue, nodes);
593 			free(pkt_node, M_SIFTR_PKTNODE);
594 		}
595 
596 		KASSERT(STAILQ_EMPTY(&tmp_pkt_queue),
597 		    ("SIFTR tmp_pkt_queue not empty after flush"));
598 
599 		mtx_lock(&siftr_pkt_mgr_mtx);
600 
601 		/*
602 		 * If siftr_exit_pkt_manager_thread gets set during the window
603 		 * where we are draining the tmp_pkt_queue above, there might
604 		 * still be pkts in pkt_queue that need to be drained.
605 		 * Allow one further iteration to occur after
606 		 * siftr_exit_pkt_manager_thread has been set to ensure
607 		 * pkt_queue is completely empty before we kill the thread.
608 		 *
609 		 * siftr_exit_pkt_manager_thread is set only after the pfil
610 		 * hooks have been removed, so only 1 extra iteration
611 		 * is needed to drain the queue.
612 		 */
613 		if (siftr_exit_pkt_manager_thread)
614 			draining--;
615 	}
616 
617 	mtx_unlock(&siftr_pkt_mgr_mtx);
618 
619 	/* Calls wakeup on this thread's struct thread ptr. */
620 	kthread_exit();
621 }
622 
623 
624 static uint32_t
625 hash_pkt(struct mbuf *m, uint32_t offset)
626 {
627 	uint32_t hash;
628 
629 	hash = 0;
630 
631 	while (m != NULL && offset > m->m_len) {
632 		/*
633 		 * The IP packet payload does not start in this mbuf, so
634 		 * need to figure out which mbuf it starts in and what offset
635 		 * into the mbuf's data region the payload starts at.
636 		 */
637 		offset -= m->m_len;
638 		m = m->m_next;
639 	}
640 
641 	while (m != NULL) {
642 		/* Ensure there is data in the mbuf */
643 		if ((m->m_len - offset) > 0)
644 			hash = hash32_buf(m->m_data + offset,
645 			    m->m_len - offset, hash);
646 
647 		m = m->m_next;
648 		offset = 0;
649         }
650 
651 	return (hash);
652 }
653 
654 
655 /*
656  * Check if a given mbuf has the SIFTR mbuf tag. If it does, log the fact that
657  * it's a reinjected packet and return. If it doesn't, tag the mbuf and return.
658  * Return value >0 means the caller should skip processing this mbuf.
659  */
660 static inline int
661 siftr_chkreinject(struct mbuf *m, int dir, struct siftr_stats *ss)
662 {
663 	if (m_tag_locate(m, PACKET_COOKIE_SIFTR, PACKET_TAG_SIFTR, NULL)
664 	    != NULL) {
665 		if (dir == PFIL_IN)
666 			ss->nskip_in_dejavu++;
667 		else
668 			ss->nskip_out_dejavu++;
669 
670 		return (1);
671 	} else {
672 		struct m_tag *tag = m_tag_alloc(PACKET_COOKIE_SIFTR,
673 		    PACKET_TAG_SIFTR, 0, M_NOWAIT);
674 		if (tag == NULL) {
675 			if (dir == PFIL_IN)
676 				ss->nskip_in_malloc++;
677 			else
678 				ss->nskip_out_malloc++;
679 
680 			return (1);
681 		}
682 
683 		m_tag_prepend(m, tag);
684 	}
685 
686 	return (0);
687 }
688 
689 
690 /*
691  * Look up an inpcb for a packet. Return the inpcb pointer if found, or NULL
692  * otherwise.
693  */
694 static inline struct inpcb *
695 siftr_findinpcb(int ipver, struct ip *ip, struct mbuf *m, uint16_t sport,
696     uint16_t dport, int dir, struct siftr_stats *ss)
697 {
698 	struct inpcb *inp;
699 
700 	/* We need the tcbinfo lock. */
701 	INP_INFO_UNLOCK_ASSERT(&V_tcbinfo);
702 
703 	if (dir == PFIL_IN)
704 		inp = (ipver == INP_IPV4 ?
705 		    in_pcblookup(&V_tcbinfo, ip->ip_src, sport, ip->ip_dst,
706 		    dport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif)
707 		    :
708 #ifdef SIFTR_IPV6
709 		    in6_pcblookup(&V_tcbinfo,
710 		    &((struct ip6_hdr *)ip)->ip6_src, sport,
711 		    &((struct ip6_hdr *)ip)->ip6_dst, dport, INPLOOKUP_RLOCKPCB,
712 		    m->m_pkthdr.rcvif)
713 #else
714 		    NULL
715 #endif
716 		    );
717 
718 	else
719 		inp = (ipver == INP_IPV4 ?
720 		    in_pcblookup(&V_tcbinfo, ip->ip_dst, dport, ip->ip_src,
721 		    sport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif)
722 		    :
723 #ifdef SIFTR_IPV6
724 		    in6_pcblookup(&V_tcbinfo,
725 		    &((struct ip6_hdr *)ip)->ip6_dst, dport,
726 		    &((struct ip6_hdr *)ip)->ip6_src, sport, INPLOOKUP_RLOCKPCB,
727 		    m->m_pkthdr.rcvif)
728 #else
729 		    NULL
730 #endif
731 		    );
732 
733 	/* If we can't find the inpcb, bail. */
734 	if (inp == NULL) {
735 		if (dir == PFIL_IN)
736 			ss->nskip_in_inpcb++;
737 		else
738 			ss->nskip_out_inpcb++;
739 	}
740 
741 	return (inp);
742 }
743 
744 
745 static inline void
746 siftr_siftdata(struct pkt_node *pn, struct inpcb *inp, struct tcpcb *tp,
747     int ipver, int dir, int inp_locally_locked)
748 {
749 #ifdef SIFTR_IPV6
750 	if (ipver == INP_IPV4) {
751 		pn->ip_laddr[3] = inp->inp_laddr.s_addr;
752 		pn->ip_faddr[3] = inp->inp_faddr.s_addr;
753 #else
754 		*((uint32_t *)pn->ip_laddr) = inp->inp_laddr.s_addr;
755 		*((uint32_t *)pn->ip_faddr) = inp->inp_faddr.s_addr;
756 #endif
757 #ifdef SIFTR_IPV6
758 	} else {
759 		pn->ip_laddr[0] = inp->in6p_laddr.s6_addr32[0];
760 		pn->ip_laddr[1] = inp->in6p_laddr.s6_addr32[1];
761 		pn->ip_laddr[2] = inp->in6p_laddr.s6_addr32[2];
762 		pn->ip_laddr[3] = inp->in6p_laddr.s6_addr32[3];
763 		pn->ip_faddr[0] = inp->in6p_faddr.s6_addr32[0];
764 		pn->ip_faddr[1] = inp->in6p_faddr.s6_addr32[1];
765 		pn->ip_faddr[2] = inp->in6p_faddr.s6_addr32[2];
766 		pn->ip_faddr[3] = inp->in6p_faddr.s6_addr32[3];
767 	}
768 #endif
769 	pn->tcp_localport = inp->inp_lport;
770 	pn->tcp_foreignport = inp->inp_fport;
771 	pn->snd_cwnd = tp->snd_cwnd;
772 	pn->snd_wnd = tp->snd_wnd;
773 	pn->rcv_wnd = tp->rcv_wnd;
774 	pn->snd_bwnd = 0;		/* Unused, kept for compat. */
775 	pn->snd_ssthresh = tp->snd_ssthresh;
776 	pn->snd_scale = tp->snd_scale;
777 	pn->rcv_scale = tp->rcv_scale;
778 	pn->conn_state = tp->t_state;
779 	pn->max_seg_size = tp->t_maxseg;
780 	pn->smoothed_rtt = tp->t_srtt;
781 	pn->sack_enabled = (tp->t_flags & TF_SACK_PERMIT) != 0;
782 	pn->flags = tp->t_flags;
783 	pn->rxt_length = tp->t_rxtcur;
784 	pn->snd_buf_hiwater = inp->inp_socket->so_snd.sb_hiwat;
785 	pn->snd_buf_cc = sbused(&inp->inp_socket->so_snd);
786 	pn->rcv_buf_hiwater = inp->inp_socket->so_rcv.sb_hiwat;
787 	pn->rcv_buf_cc = sbused(&inp->inp_socket->so_rcv);
788 	pn->sent_inflight_bytes = tp->snd_max - tp->snd_una;
789 	pn->t_segqlen = tp->t_segqlen;
790 
791 	/* We've finished accessing the tcb so release the lock. */
792 	if (inp_locally_locked)
793 		INP_RUNLOCK(inp);
794 
795 	pn->ipver = ipver;
796 	pn->direction = dir;
797 
798 	/*
799 	 * Significantly more accurate than using getmicrotime(), but slower!
800 	 * Gives true microsecond resolution at the expense of a hit to
801 	 * maximum pps throughput processing when SIFTR is loaded and enabled.
802 	 */
803 	microtime(&pn->tval);
804 }
805 
806 
807 /*
808  * pfil hook that is called for each IPv4 packet making its way through the
809  * stack in either direction.
810  * The pfil subsystem holds a non-sleepable mutex somewhere when
811  * calling our hook function, so we can't sleep at all.
812  * It's very important to use the M_NOWAIT flag with all function calls
813  * that support it so that they won't sleep, otherwise you get a panic.
814  */
815 static int
816 siftr_chkpkt(void *arg, struct mbuf **m, struct ifnet *ifp, int dir,
817     struct inpcb *inp)
818 {
819 	struct pkt_node *pn;
820 	struct ip *ip;
821 	struct tcphdr *th;
822 	struct tcpcb *tp;
823 	struct siftr_stats *ss;
824 	unsigned int ip_hl;
825 	int inp_locally_locked;
826 
827 	inp_locally_locked = 0;
828 	ss = DPCPU_PTR(ss);
829 
830 	/*
831 	 * m_pullup is not required here because ip_{input|output}
832 	 * already do the heavy lifting for us.
833 	 */
834 
835 	ip = mtod(*m, struct ip *);
836 
837 	/* Only continue processing if the packet is TCP. */
838 	if (ip->ip_p != IPPROTO_TCP)
839 		goto ret;
840 
841 	/*
842 	 * If a kernel subsystem reinjects packets into the stack, our pfil
843 	 * hook will be called multiple times for the same packet.
844 	 * Make sure we only process unique packets.
845 	 */
846 	if (siftr_chkreinject(*m, dir, ss))
847 		goto ret;
848 
849 	if (dir == PFIL_IN)
850 		ss->n_in++;
851 	else
852 		ss->n_out++;
853 
854 	/*
855 	 * Create a tcphdr struct starting at the correct offset
856 	 * in the IP packet. ip->ip_hl gives the ip header length
857 	 * in 4-byte words, so multiply it to get the size in bytes.
858 	 */
859 	ip_hl = (ip->ip_hl << 2);
860 	th = (struct tcphdr *)((caddr_t)ip + ip_hl);
861 
862 	/*
863 	 * If the pfil hooks don't provide a pointer to the
864 	 * inpcb, we need to find it ourselves and lock it.
865 	 */
866 	if (!inp) {
867 		/* Find the corresponding inpcb for this pkt. */
868 		inp = siftr_findinpcb(INP_IPV4, ip, *m, th->th_sport,
869 		    th->th_dport, dir, ss);
870 
871 		if (inp == NULL)
872 			goto ret;
873 		else
874 			inp_locally_locked = 1;
875 	}
876 
877 	INP_LOCK_ASSERT(inp);
878 
879 	/* Find the TCP control block that corresponds with this packet */
880 	tp = intotcpcb(inp);
881 
882 	/*
883 	 * If we can't find the TCP control block (happens occasionaly for a
884 	 * packet sent during the shutdown phase of a TCP connection),
885 	 * or we're in the timewait state, bail
886 	 */
887 	if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) {
888 		if (dir == PFIL_IN)
889 			ss->nskip_in_tcpcb++;
890 		else
891 			ss->nskip_out_tcpcb++;
892 
893 		goto inp_unlock;
894 	}
895 
896 	pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO);
897 
898 	if (pn == NULL) {
899 		if (dir == PFIL_IN)
900 			ss->nskip_in_malloc++;
901 		else
902 			ss->nskip_out_malloc++;
903 
904 		goto inp_unlock;
905 	}
906 
907 	siftr_siftdata(pn, inp, tp, INP_IPV4, dir, inp_locally_locked);
908 
909 	if (siftr_generate_hashes) {
910 		if ((*m)->m_pkthdr.csum_flags & CSUM_TCP) {
911 			/*
912 			 * For outbound packets, the TCP checksum isn't
913 			 * calculated yet. This is a problem for our packet
914 			 * hashing as the receiver will calc a different hash
915 			 * to ours if we don't include the correct TCP checksum
916 			 * in the bytes being hashed. To work around this
917 			 * problem, we manually calc the TCP checksum here in
918 			 * software. We unset the CSUM_TCP flag so the lower
919 			 * layers don't recalc it.
920 			 */
921 			(*m)->m_pkthdr.csum_flags &= ~CSUM_TCP;
922 
923 			/*
924 			 * Calculate the TCP checksum in software and assign
925 			 * to correct TCP header field, which will follow the
926 			 * packet mbuf down the stack. The trick here is that
927 			 * tcp_output() sets th->th_sum to the checksum of the
928 			 * pseudo header for us already. Because of the nature
929 			 * of the checksumming algorithm, we can sum over the
930 			 * entire IP payload (i.e. TCP header and data), which
931 			 * will include the already calculated pseduo header
932 			 * checksum, thus giving us the complete TCP checksum.
933 			 *
934 			 * To put it in simple terms, if checksum(1,2,3,4)=10,
935 			 * then checksum(1,2,3,4,5) == checksum(10,5).
936 			 * This property is what allows us to "cheat" and
937 			 * checksum only the IP payload which has the TCP
938 			 * th_sum field populated with the pseudo header's
939 			 * checksum, and not need to futz around checksumming
940 			 * pseudo header bytes and TCP header/data in one hit.
941 			 * Refer to RFC 1071 for more info.
942 			 *
943 			 * NB: in_cksum_skip(struct mbuf *m, int len, int skip)
944 			 * in_cksum_skip 2nd argument is NOT the number of
945 			 * bytes to read from the mbuf at "skip" bytes offset
946 			 * from the start of the mbuf (very counter intuitive!).
947 			 * The number of bytes to read is calculated internally
948 			 * by the function as len-skip i.e. to sum over the IP
949 			 * payload (TCP header + data) bytes, it is INCORRECT
950 			 * to call the function like this:
951 			 * in_cksum_skip(at, ip->ip_len - offset, offset)
952 			 * Rather, it should be called like this:
953 			 * in_cksum_skip(at, ip->ip_len, offset)
954 			 * which means read "ip->ip_len - offset" bytes from
955 			 * the mbuf cluster "at" at offset "offset" bytes from
956 			 * the beginning of the "at" mbuf's data pointer.
957 			 */
958 			th->th_sum = in_cksum_skip(*m, ntohs(ip->ip_len),
959 			    ip_hl);
960 		}
961 
962 		/*
963 		 * XXX: Having to calculate the checksum in software and then
964 		 * hash over all bytes is really inefficient. Would be nice to
965 		 * find a way to create the hash and checksum in the same pass
966 		 * over the bytes.
967 		 */
968 		pn->hash = hash_pkt(*m, ip_hl);
969 	}
970 
971 	mtx_lock(&siftr_pkt_queue_mtx);
972 	STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes);
973 	mtx_unlock(&siftr_pkt_queue_mtx);
974 	goto ret;
975 
976 inp_unlock:
977 	if (inp_locally_locked)
978 		INP_RUNLOCK(inp);
979 
980 ret:
981 	/* Returning 0 ensures pfil will not discard the pkt */
982 	return (0);
983 }
984 
985 
986 #ifdef SIFTR_IPV6
987 static int
988 siftr_chkpkt6(void *arg, struct mbuf **m, struct ifnet *ifp, int dir,
989     struct inpcb *inp)
990 {
991 	struct pkt_node *pn;
992 	struct ip6_hdr *ip6;
993 	struct tcphdr *th;
994 	struct tcpcb *tp;
995 	struct siftr_stats *ss;
996 	unsigned int ip6_hl;
997 	int inp_locally_locked;
998 
999 	inp_locally_locked = 0;
1000 	ss = DPCPU_PTR(ss);
1001 
1002 	/*
1003 	 * m_pullup is not required here because ip6_{input|output}
1004 	 * already do the heavy lifting for us.
1005 	 */
1006 
1007 	ip6 = mtod(*m, struct ip6_hdr *);
1008 
1009 	/*
1010 	 * Only continue processing if the packet is TCP
1011 	 * XXX: We should follow the next header fields
1012 	 * as shown on Pg 6 RFC 2460, but right now we'll
1013 	 * only check pkts that have no extension headers.
1014 	 */
1015 	if (ip6->ip6_nxt != IPPROTO_TCP)
1016 		goto ret6;
1017 
1018 	/*
1019 	 * If a kernel subsystem reinjects packets into the stack, our pfil
1020 	 * hook will be called multiple times for the same packet.
1021 	 * Make sure we only process unique packets.
1022 	 */
1023 	if (siftr_chkreinject(*m, dir, ss))
1024 		goto ret6;
1025 
1026 	if (dir == PFIL_IN)
1027 		ss->n_in++;
1028 	else
1029 		ss->n_out++;
1030 
1031 	ip6_hl = sizeof(struct ip6_hdr);
1032 
1033 	/*
1034 	 * Create a tcphdr struct starting at the correct offset
1035 	 * in the ipv6 packet. ip->ip_hl gives the ip header length
1036 	 * in 4-byte words, so multiply it to get the size in bytes.
1037 	 */
1038 	th = (struct tcphdr *)((caddr_t)ip6 + ip6_hl);
1039 
1040 	/*
1041 	 * For inbound packets, the pfil hooks don't provide a pointer to the
1042 	 * inpcb, so we need to find it ourselves and lock it.
1043 	 */
1044 	if (!inp) {
1045 		/* Find the corresponding inpcb for this pkt. */
1046 		inp = siftr_findinpcb(INP_IPV6, (struct ip *)ip6, *m,
1047 		    th->th_sport, th->th_dport, dir, ss);
1048 
1049 		if (inp == NULL)
1050 			goto ret6;
1051 		else
1052 			inp_locally_locked = 1;
1053 	}
1054 
1055 	/* Find the TCP control block that corresponds with this packet. */
1056 	tp = intotcpcb(inp);
1057 
1058 	/*
1059 	 * If we can't find the TCP control block (happens occasionaly for a
1060 	 * packet sent during the shutdown phase of a TCP connection),
1061 	 * or we're in the timewait state, bail.
1062 	 */
1063 	if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) {
1064 		if (dir == PFIL_IN)
1065 			ss->nskip_in_tcpcb++;
1066 		else
1067 			ss->nskip_out_tcpcb++;
1068 
1069 		goto inp_unlock6;
1070 	}
1071 
1072 	pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO);
1073 
1074 	if (pn == NULL) {
1075 		if (dir == PFIL_IN)
1076 			ss->nskip_in_malloc++;
1077 		else
1078 			ss->nskip_out_malloc++;
1079 
1080 		goto inp_unlock6;
1081 	}
1082 
1083 	siftr_siftdata(pn, inp, tp, INP_IPV6, dir, inp_locally_locked);
1084 
1085 	/* XXX: Figure out how to generate hashes for IPv6 packets. */
1086 
1087 	mtx_lock(&siftr_pkt_queue_mtx);
1088 	STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes);
1089 	mtx_unlock(&siftr_pkt_queue_mtx);
1090 	goto ret6;
1091 
1092 inp_unlock6:
1093 	if (inp_locally_locked)
1094 		INP_RUNLOCK(inp);
1095 
1096 ret6:
1097 	/* Returning 0 ensures pfil will not discard the pkt. */
1098 	return (0);
1099 }
1100 #endif /* #ifdef SIFTR_IPV6 */
1101 
1102 
1103 static int
1104 siftr_pfil(int action)
1105 {
1106 	struct pfil_head *pfh_inet;
1107 #ifdef SIFTR_IPV6
1108 	struct pfil_head *pfh_inet6;
1109 #endif
1110 	VNET_ITERATOR_DECL(vnet_iter);
1111 
1112 	VNET_LIST_RLOCK();
1113 	VNET_FOREACH(vnet_iter) {
1114 		CURVNET_SET(vnet_iter);
1115 		pfh_inet = pfil_head_get(PFIL_TYPE_AF, AF_INET);
1116 #ifdef SIFTR_IPV6
1117 		pfh_inet6 = pfil_head_get(PFIL_TYPE_AF, AF_INET6);
1118 #endif
1119 
1120 		if (action == HOOK) {
1121 			pfil_add_hook(siftr_chkpkt, NULL,
1122 			    PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet);
1123 #ifdef SIFTR_IPV6
1124 			pfil_add_hook(siftr_chkpkt6, NULL,
1125 			    PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet6);
1126 #endif
1127 		} else if (action == UNHOOK) {
1128 			pfil_remove_hook(siftr_chkpkt, NULL,
1129 			    PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet);
1130 #ifdef SIFTR_IPV6
1131 			pfil_remove_hook(siftr_chkpkt6, NULL,
1132 			    PFIL_IN | PFIL_OUT | PFIL_WAITOK, pfh_inet6);
1133 #endif
1134 		}
1135 		CURVNET_RESTORE();
1136 	}
1137 	VNET_LIST_RUNLOCK();
1138 
1139 	return (0);
1140 }
1141 
1142 
1143 static int
1144 siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS)
1145 {
1146 	struct alq *new_alq;
1147 	int error;
1148 
1149 	error = sysctl_handle_string(oidp, arg1, arg2, req);
1150 
1151 	/* Check for error or same filename */
1152 	if (error != 0 || req->newptr == NULL ||
1153 	    strncmp(siftr_logfile, arg1, arg2) == 0)
1154 		goto done;
1155 
1156 	/* Filname changed */
1157 	error = alq_open(&new_alq, arg1, curthread->td_ucred,
1158 	    SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0);
1159 	if (error != 0)
1160 		goto done;
1161 
1162 	/*
1163 	 * If disabled, siftr_alq == NULL so we simply close
1164 	 * the alq as we've proved it can be opened.
1165 	 * If enabled, close the existing alq and switch the old
1166 	 * for the new.
1167 	 */
1168 	if (siftr_alq == NULL) {
1169 		alq_close(new_alq);
1170 	} else {
1171 		alq_close(siftr_alq);
1172 		siftr_alq = new_alq;
1173 	}
1174 
1175 	/* Update filename upon success */
1176 	strlcpy(siftr_logfile, arg1, arg2);
1177 done:
1178 	return (error);
1179 }
1180 
1181 static int
1182 siftr_manage_ops(uint8_t action)
1183 {
1184 	struct siftr_stats totalss;
1185 	struct timeval tval;
1186 	struct flow_hash_node *counter, *tmp_counter;
1187 	struct sbuf *s;
1188 	int i, key_index, ret, error;
1189 	uint32_t bytes_to_write, total_skipped_pkts;
1190 	uint16_t lport, fport;
1191 	uint8_t *key, ipver;
1192 
1193 #ifdef SIFTR_IPV6
1194 	uint32_t laddr[4];
1195 	uint32_t faddr[4];
1196 #else
1197 	uint8_t laddr[4];
1198 	uint8_t faddr[4];
1199 #endif
1200 
1201 	error = 0;
1202 	total_skipped_pkts = 0;
1203 
1204 	/* Init an autosizing sbuf that initially holds 200 chars. */
1205 	if ((s = sbuf_new(NULL, NULL, 200, SBUF_AUTOEXTEND)) == NULL)
1206 		return (-1);
1207 
1208 	if (action == SIFTR_ENABLE) {
1209 		/*
1210 		 * Create our alq
1211 		 * XXX: We should abort if alq_open fails!
1212 		 */
1213 		alq_open(&siftr_alq, siftr_logfile, curthread->td_ucred,
1214 		    SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0);
1215 
1216 		STAILQ_INIT(&pkt_queue);
1217 
1218 		DPCPU_ZERO(ss);
1219 
1220 		siftr_exit_pkt_manager_thread = 0;
1221 
1222 		ret = kthread_add(&siftr_pkt_manager_thread, NULL, NULL,
1223 		    &siftr_pkt_manager_thr, RFNOWAIT, 0,
1224 		    "siftr_pkt_manager_thr");
1225 
1226 		siftr_pfil(HOOK);
1227 
1228 		microtime(&tval);
1229 
1230 		sbuf_printf(s,
1231 		    "enable_time_secs=%jd\tenable_time_usecs=%06ld\t"
1232 		    "siftrver=%s\thz=%u\ttcp_rtt_scale=%u\tsysname=%s\t"
1233 		    "sysver=%u\tipmode=%u\n",
1234 		    (intmax_t)tval.tv_sec, tval.tv_usec, MODVERSION_STR, hz,
1235 		    TCP_RTT_SCALE, SYS_NAME, __FreeBSD_version, SIFTR_IPMODE);
1236 
1237 		sbuf_finish(s);
1238 		alq_writen(siftr_alq, sbuf_data(s), sbuf_len(s), ALQ_WAITOK);
1239 
1240 	} else if (action == SIFTR_DISABLE && siftr_pkt_manager_thr != NULL) {
1241 		/*
1242 		 * Remove the pfil hook functions. All threads currently in
1243 		 * the hook functions are allowed to exit before siftr_pfil()
1244 		 * returns.
1245 		 */
1246 		siftr_pfil(UNHOOK);
1247 
1248 		/* This will block until the pkt manager thread unlocks it. */
1249 		mtx_lock(&siftr_pkt_mgr_mtx);
1250 
1251 		/* Tell the pkt manager thread that it should exit now. */
1252 		siftr_exit_pkt_manager_thread = 1;
1253 
1254 		/*
1255 		 * Wake the pkt_manager thread so it realises that
1256 		 * siftr_exit_pkt_manager_thread == 1 and exits gracefully.
1257 		 * The wakeup won't be delivered until we unlock
1258 		 * siftr_pkt_mgr_mtx so this isn't racy.
1259 		 */
1260 		wakeup(&wait_for_pkt);
1261 
1262 		/* Wait for the pkt_manager thread to exit. */
1263 		mtx_sleep(siftr_pkt_manager_thr, &siftr_pkt_mgr_mtx, PWAIT,
1264 		    "thrwait", 0);
1265 
1266 		siftr_pkt_manager_thr = NULL;
1267 		mtx_unlock(&siftr_pkt_mgr_mtx);
1268 
1269 		totalss.n_in = DPCPU_VARSUM(ss, n_in);
1270 		totalss.n_out = DPCPU_VARSUM(ss, n_out);
1271 		totalss.nskip_in_malloc = DPCPU_VARSUM(ss, nskip_in_malloc);
1272 		totalss.nskip_out_malloc = DPCPU_VARSUM(ss, nskip_out_malloc);
1273 		totalss.nskip_in_mtx = DPCPU_VARSUM(ss, nskip_in_mtx);
1274 		totalss.nskip_out_mtx = DPCPU_VARSUM(ss, nskip_out_mtx);
1275 		totalss.nskip_in_tcpcb = DPCPU_VARSUM(ss, nskip_in_tcpcb);
1276 		totalss.nskip_out_tcpcb = DPCPU_VARSUM(ss, nskip_out_tcpcb);
1277 		totalss.nskip_in_inpcb = DPCPU_VARSUM(ss, nskip_in_inpcb);
1278 		totalss.nskip_out_inpcb = DPCPU_VARSUM(ss, nskip_out_inpcb);
1279 
1280 		total_skipped_pkts = totalss.nskip_in_malloc +
1281 		    totalss.nskip_out_malloc + totalss.nskip_in_mtx +
1282 		    totalss.nskip_out_mtx + totalss.nskip_in_tcpcb +
1283 		    totalss.nskip_out_tcpcb + totalss.nskip_in_inpcb +
1284 		    totalss.nskip_out_inpcb;
1285 
1286 		microtime(&tval);
1287 
1288 		sbuf_printf(s,
1289 		    "disable_time_secs=%jd\tdisable_time_usecs=%06ld\t"
1290 		    "num_inbound_tcp_pkts=%ju\tnum_outbound_tcp_pkts=%ju\t"
1291 		    "total_tcp_pkts=%ju\tnum_inbound_skipped_pkts_malloc=%u\t"
1292 		    "num_outbound_skipped_pkts_malloc=%u\t"
1293 		    "num_inbound_skipped_pkts_mtx=%u\t"
1294 		    "num_outbound_skipped_pkts_mtx=%u\t"
1295 		    "num_inbound_skipped_pkts_tcpcb=%u\t"
1296 		    "num_outbound_skipped_pkts_tcpcb=%u\t"
1297 		    "num_inbound_skipped_pkts_inpcb=%u\t"
1298 		    "num_outbound_skipped_pkts_inpcb=%u\t"
1299 		    "total_skipped_tcp_pkts=%u\tflow_list=",
1300 		    (intmax_t)tval.tv_sec,
1301 		    tval.tv_usec,
1302 		    (uintmax_t)totalss.n_in,
1303 		    (uintmax_t)totalss.n_out,
1304 		    (uintmax_t)(totalss.n_in + totalss.n_out),
1305 		    totalss.nskip_in_malloc,
1306 		    totalss.nskip_out_malloc,
1307 		    totalss.nskip_in_mtx,
1308 		    totalss.nskip_out_mtx,
1309 		    totalss.nskip_in_tcpcb,
1310 		    totalss.nskip_out_tcpcb,
1311 		    totalss.nskip_in_inpcb,
1312 		    totalss.nskip_out_inpcb,
1313 		    total_skipped_pkts);
1314 
1315 		/*
1316 		 * Iterate over the flow hash, printing a summary of each
1317 		 * flow seen and freeing any malloc'd memory.
1318 		 * The hash consists of an array of LISTs (man 3 queue).
1319 		 */
1320 		for (i = 0; i <= siftr_hashmask; i++) {
1321 			LIST_FOREACH_SAFE(counter, counter_hash + i, nodes,
1322 			    tmp_counter) {
1323 				key = counter->key;
1324 				key_index = 1;
1325 
1326 				ipver = key[0];
1327 
1328 				memcpy(laddr, key + key_index, sizeof(laddr));
1329 				key_index += sizeof(laddr);
1330 				memcpy(&lport, key + key_index, sizeof(lport));
1331 				key_index += sizeof(lport);
1332 				memcpy(faddr, key + key_index, sizeof(faddr));
1333 				key_index += sizeof(faddr);
1334 				memcpy(&fport, key + key_index, sizeof(fport));
1335 
1336 #ifdef SIFTR_IPV6
1337 				laddr[3] = ntohl(laddr[3]);
1338 				faddr[3] = ntohl(faddr[3]);
1339 
1340 				if (ipver == INP_IPV6) {
1341 					laddr[0] = ntohl(laddr[0]);
1342 					laddr[1] = ntohl(laddr[1]);
1343 					laddr[2] = ntohl(laddr[2]);
1344 					faddr[0] = ntohl(faddr[0]);
1345 					faddr[1] = ntohl(faddr[1]);
1346 					faddr[2] = ntohl(faddr[2]);
1347 
1348 					sbuf_printf(s,
1349 					    "%x:%x:%x:%x:%x:%x:%x:%x;%u-"
1350 					    "%x:%x:%x:%x:%x:%x:%x:%x;%u,",
1351 					    UPPER_SHORT(laddr[0]),
1352 					    LOWER_SHORT(laddr[0]),
1353 					    UPPER_SHORT(laddr[1]),
1354 					    LOWER_SHORT(laddr[1]),
1355 					    UPPER_SHORT(laddr[2]),
1356 					    LOWER_SHORT(laddr[2]),
1357 					    UPPER_SHORT(laddr[3]),
1358 					    LOWER_SHORT(laddr[3]),
1359 					    ntohs(lport),
1360 					    UPPER_SHORT(faddr[0]),
1361 					    LOWER_SHORT(faddr[0]),
1362 					    UPPER_SHORT(faddr[1]),
1363 					    LOWER_SHORT(faddr[1]),
1364 					    UPPER_SHORT(faddr[2]),
1365 					    LOWER_SHORT(faddr[2]),
1366 					    UPPER_SHORT(faddr[3]),
1367 					    LOWER_SHORT(faddr[3]),
1368 					    ntohs(fport));
1369 				} else {
1370 					laddr[0] = FIRST_OCTET(laddr[3]);
1371 					laddr[1] = SECOND_OCTET(laddr[3]);
1372 					laddr[2] = THIRD_OCTET(laddr[3]);
1373 					laddr[3] = FOURTH_OCTET(laddr[3]);
1374 					faddr[0] = FIRST_OCTET(faddr[3]);
1375 					faddr[1] = SECOND_OCTET(faddr[3]);
1376 					faddr[2] = THIRD_OCTET(faddr[3]);
1377 					faddr[3] = FOURTH_OCTET(faddr[3]);
1378 #endif
1379 					sbuf_printf(s,
1380 					    "%u.%u.%u.%u;%u-%u.%u.%u.%u;%u,",
1381 					    laddr[0],
1382 					    laddr[1],
1383 					    laddr[2],
1384 					    laddr[3],
1385 					    ntohs(lport),
1386 					    faddr[0],
1387 					    faddr[1],
1388 					    faddr[2],
1389 					    faddr[3],
1390 					    ntohs(fport));
1391 #ifdef SIFTR_IPV6
1392 				}
1393 #endif
1394 
1395 				free(counter, M_SIFTR_HASHNODE);
1396 			}
1397 
1398 			LIST_INIT(counter_hash + i);
1399 		}
1400 
1401 		sbuf_printf(s, "\n");
1402 		sbuf_finish(s);
1403 
1404 		i = 0;
1405 		do {
1406 			bytes_to_write = min(SIFTR_ALQ_BUFLEN, sbuf_len(s)-i);
1407 			alq_writen(siftr_alq, sbuf_data(s)+i, bytes_to_write, ALQ_WAITOK);
1408 			i += bytes_to_write;
1409 		} while (i < sbuf_len(s));
1410 
1411 		alq_close(siftr_alq);
1412 		siftr_alq = NULL;
1413 	}
1414 
1415 	sbuf_delete(s);
1416 
1417 	/*
1418 	 * XXX: Should be using ret to check if any functions fail
1419 	 * and set error appropriately
1420 	 */
1421 
1422 	return (error);
1423 }
1424 
1425 
1426 static int
1427 siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS)
1428 {
1429 	if (req->newptr == NULL)
1430 		goto skip;
1431 
1432 	/* If the value passed in isn't 0 or 1, return an error. */
1433 	if (CAST_PTR_INT(req->newptr) != 0 && CAST_PTR_INT(req->newptr) != 1)
1434 		return (1);
1435 
1436 	/* If we are changing state (0 to 1 or 1 to 0). */
1437 	if (CAST_PTR_INT(req->newptr) != siftr_enabled )
1438 		if (siftr_manage_ops(CAST_PTR_INT(req->newptr))) {
1439 			siftr_manage_ops(SIFTR_DISABLE);
1440 			return (1);
1441 		}
1442 
1443 skip:
1444 	return (sysctl_handle_int(oidp, arg1, arg2, req));
1445 }
1446 
1447 
1448 static void
1449 siftr_shutdown_handler(void *arg)
1450 {
1451 	siftr_manage_ops(SIFTR_DISABLE);
1452 }
1453 
1454 
1455 /*
1456  * Module is being unloaded or machine is shutting down. Take care of cleanup.
1457  */
1458 static int
1459 deinit_siftr(void)
1460 {
1461 	/* Cleanup. */
1462 	siftr_manage_ops(SIFTR_DISABLE);
1463 	hashdestroy(counter_hash, M_SIFTR, siftr_hashmask);
1464 	mtx_destroy(&siftr_pkt_queue_mtx);
1465 	mtx_destroy(&siftr_pkt_mgr_mtx);
1466 
1467 	return (0);
1468 }
1469 
1470 
1471 /*
1472  * Module has just been loaded into the kernel.
1473  */
1474 static int
1475 init_siftr(void)
1476 {
1477 	EVENTHANDLER_REGISTER(shutdown_pre_sync, siftr_shutdown_handler, NULL,
1478 	    SHUTDOWN_PRI_FIRST);
1479 
1480 	/* Initialise our flow counter hash table. */
1481 	counter_hash = hashinit(SIFTR_EXPECTED_MAX_TCP_FLOWS, M_SIFTR,
1482 	    &siftr_hashmask);
1483 
1484 	mtx_init(&siftr_pkt_queue_mtx, "siftr_pkt_queue_mtx", NULL, MTX_DEF);
1485 	mtx_init(&siftr_pkt_mgr_mtx, "siftr_pkt_mgr_mtx", NULL, MTX_DEF);
1486 
1487 	/* Print message to the user's current terminal. */
1488 	uprintf("\nStatistical Information For TCP Research (SIFTR) %s\n"
1489 	    "          http://caia.swin.edu.au/urp/newtcp\n\n",
1490 	    MODVERSION_STR);
1491 
1492 	return (0);
1493 }
1494 
1495 
1496 /*
1497  * This is the function that is called to load and unload the module.
1498  * When the module is loaded, this function is called once with
1499  * "what" == MOD_LOAD
1500  * When the module is unloaded, this function is called twice with
1501  * "what" = MOD_QUIESCE first, followed by "what" = MOD_UNLOAD second
1502  * When the system is shut down e.g. CTRL-ALT-DEL or using the shutdown command,
1503  * this function is called once with "what" = MOD_SHUTDOWN
1504  * When the system is shut down, the handler isn't called until the very end
1505  * of the shutdown sequence i.e. after the disks have been synced.
1506  */
1507 static int
1508 siftr_load_handler(module_t mod, int what, void *arg)
1509 {
1510 	int ret;
1511 
1512 	switch (what) {
1513 	case MOD_LOAD:
1514 		ret = init_siftr();
1515 		break;
1516 
1517 	case MOD_QUIESCE:
1518 	case MOD_SHUTDOWN:
1519 		ret = deinit_siftr();
1520 		break;
1521 
1522 	case MOD_UNLOAD:
1523 		ret = 0;
1524 		break;
1525 
1526 	default:
1527 		ret = EINVAL;
1528 		break;
1529 	}
1530 
1531 	return (ret);
1532 }
1533 
1534 
1535 static moduledata_t siftr_mod = {
1536 	.name = "siftr",
1537 	.evhand = siftr_load_handler,
1538 };
1539 
1540 /*
1541  * Param 1: name of the kernel module
1542  * Param 2: moduledata_t struct containing info about the kernel module
1543  *          and the execution entry point for the module
1544  * Param 3: From sysinit_sub_id enumeration in /usr/include/sys/kernel.h
1545  *          Defines the module initialisation order
1546  * Param 4: From sysinit_elem_order enumeration in /usr/include/sys/kernel.h
1547  *          Defines the initialisation order of this kld relative to others
1548  *          within the same subsystem as defined by param 3
1549  */
1550 DECLARE_MODULE(siftr, siftr_mod, SI_SUB_SMP, SI_ORDER_ANY);
1551 MODULE_DEPEND(siftr, alq, 1, 1, 1);
1552 MODULE_VERSION(siftr, MODVERSION);
1553