xref: /freebsd/sys/netgraph/ng_pipe.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004-2010 University of Zagreb
5  * Copyright (c) 2007-2008 FreeBSD Foundation
6  *
7  * This software was developed by the University of Zagreb and the
8  * FreeBSD Foundation under sponsorship by the Stichting NLnet and the
9  * 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 AUTHOR 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 AUTHOR 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  * $FreeBSD$
33  */
34 
35 /*
36  * This node permits simple traffic shaping by emulating bandwidth
37  * and delay, as well as random packet losses.
38  * The node has two hooks, upper and lower. Traffic flowing from upper to
39  * lower hook is referenced as downstream, and vice versa. Parameters for
40  * both directions can be set separately, except for delay.
41  */
42 
43 
44 #include <sys/param.h>
45 #include <sys/errno.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/time.h>
51 
52 #include <vm/uma.h>
53 
54 #include <net/vnet.h>
55 
56 #include <netinet/in.h>
57 #include <netinet/in_systm.h>
58 #include <netinet/ip.h>
59 
60 #include <netgraph/ng_message.h>
61 #include <netgraph/netgraph.h>
62 #include <netgraph/ng_parse.h>
63 #include <netgraph/ng_pipe.h>
64 
65 static MALLOC_DEFINE(M_NG_PIPE, "ng_pipe", "ng_pipe");
66 
67 /* Packet header struct */
68 struct ngp_hdr {
69 	TAILQ_ENTRY(ngp_hdr)	ngp_link;	/* next pkt in queue */
70 	struct timeval		when;		/* this packet's due time */
71 	struct mbuf		*m;		/* ptr to the packet data */
72 };
73 TAILQ_HEAD(p_head, ngp_hdr);
74 
75 /* FIFO queue struct */
76 struct ngp_fifo {
77 	TAILQ_ENTRY(ngp_fifo)	fifo_le;	/* list of active queues only */
78 	struct p_head		packet_head;	/* FIFO queue head */
79 	u_int32_t		hash;		/* flow signature */
80 	struct timeval		vtime;		/* virtual time, for WFQ */
81 	u_int32_t		rr_deficit;	/* for DRR */
82 	u_int32_t		packets;	/* # of packets in this queue */
83 };
84 
85 /* Per hook info */
86 struct hookinfo {
87 	hook_p			hook;
88 	int			noqueue;	/* bypass any processing */
89 	TAILQ_HEAD(, ngp_fifo)	fifo_head;	/* FIFO queues */
90 	TAILQ_HEAD(, ngp_hdr)	qout_head;	/* delay queue head */
91 	struct timeval		qin_utime;
92 	struct ng_pipe_hookcfg	cfg;
93 	struct ng_pipe_hookrun	run;
94 	struct ng_pipe_hookstat	stats;
95 	uint64_t		*ber_p;		/* loss_p(BER,psize) map */
96 };
97 
98 /* Per node info */
99 struct node_priv {
100 	u_int64_t		delay;
101 	u_int32_t		overhead;
102 	u_int32_t		header_offset;
103 	struct hookinfo		lower;
104 	struct hookinfo		upper;
105 	struct callout		timer;
106 	int			timer_scheduled;
107 };
108 typedef struct node_priv *priv_p;
109 
110 /* Macro for calculating the virtual time for packet dequeueing in WFQ */
111 #define FIFO_VTIME_SORT(plen)						\
112 	if (hinfo->cfg.wfq && hinfo->cfg.bandwidth) {			\
113 		ngp_f->vtime.tv_usec = now->tv_usec + ((uint64_t) (plen) \
114 			+ priv->overhead ) * hinfo->run.fifo_queues *	\
115 			8000000 / hinfo->cfg.bandwidth;			\
116 		ngp_f->vtime.tv_sec = now->tv_sec +			\
117 			ngp_f->vtime.tv_usec / 1000000;			\
118 		ngp_f->vtime.tv_usec = ngp_f->vtime.tv_usec % 1000000;	\
119 		TAILQ_FOREACH(ngp_f1, &hinfo->fifo_head, fifo_le)	\
120 			if (ngp_f1->vtime.tv_sec > ngp_f->vtime.tv_sec || \
121 			    (ngp_f1->vtime.tv_sec == ngp_f->vtime.tv_sec && \
122 			    ngp_f1->vtime.tv_usec > ngp_f->vtime.tv_usec)) \
123 				break;					\
124 		if (ngp_f1 == NULL)					\
125 			TAILQ_INSERT_TAIL(&hinfo->fifo_head, ngp_f, fifo_le); \
126 		else							\
127 			TAILQ_INSERT_BEFORE(ngp_f1, ngp_f, fifo_le);	\
128 	} else								\
129 		TAILQ_INSERT_TAIL(&hinfo->fifo_head, ngp_f, fifo_le);	\
130 
131 
132 static void	parse_cfg(struct ng_pipe_hookcfg *, struct ng_pipe_hookcfg *,
133 			struct hookinfo *, priv_p);
134 static void	pipe_dequeue(struct hookinfo *, struct timeval *);
135 static void	ngp_callout(node_p, hook_p, void *, int);
136 static int	ngp_modevent(module_t, int, void *);
137 
138 /* zone for storing ngp_hdr-s */
139 static uma_zone_t ngp_zone;
140 
141 /* Netgraph methods */
142 static ng_constructor_t	ngp_constructor;
143 static ng_rcvmsg_t	ngp_rcvmsg;
144 static ng_shutdown_t	ngp_shutdown;
145 static ng_newhook_t	ngp_newhook;
146 static ng_rcvdata_t	ngp_rcvdata;
147 static ng_disconnect_t	ngp_disconnect;
148 
149 /* Parse type for struct ng_pipe_hookstat */
150 static const struct ng_parse_struct_field
151 	ng_pipe_hookstat_type_fields[] = NG_PIPE_HOOKSTAT_INFO;
152 static const struct ng_parse_type ng_pipe_hookstat_type = {
153 	&ng_parse_struct_type,
154 	&ng_pipe_hookstat_type_fields
155 };
156 
157 /* Parse type for struct ng_pipe_stats */
158 static const struct ng_parse_struct_field ng_pipe_stats_type_fields[] =
159 	NG_PIPE_STATS_INFO(&ng_pipe_hookstat_type);
160 static const struct ng_parse_type ng_pipe_stats_type = {
161 	&ng_parse_struct_type,
162 	&ng_pipe_stats_type_fields
163 };
164 
165 /* Parse type for struct ng_pipe_hookrun */
166 static const struct ng_parse_struct_field
167 	ng_pipe_hookrun_type_fields[] = NG_PIPE_HOOKRUN_INFO;
168 static const struct ng_parse_type ng_pipe_hookrun_type = {
169 	&ng_parse_struct_type,
170 	&ng_pipe_hookrun_type_fields
171 };
172 
173 /* Parse type for struct ng_pipe_run */
174 static const struct ng_parse_struct_field
175 	ng_pipe_run_type_fields[] = NG_PIPE_RUN_INFO(&ng_pipe_hookrun_type);
176 static const struct ng_parse_type ng_pipe_run_type = {
177 	&ng_parse_struct_type,
178 	&ng_pipe_run_type_fields
179 };
180 
181 /* Parse type for struct ng_pipe_hookcfg */
182 static const struct ng_parse_struct_field
183 	ng_pipe_hookcfg_type_fields[] = NG_PIPE_HOOKCFG_INFO;
184 static const struct ng_parse_type ng_pipe_hookcfg_type = {
185 	&ng_parse_struct_type,
186 	&ng_pipe_hookcfg_type_fields
187 };
188 
189 /* Parse type for struct ng_pipe_cfg */
190 static const struct ng_parse_struct_field
191 	ng_pipe_cfg_type_fields[] = NG_PIPE_CFG_INFO(&ng_pipe_hookcfg_type);
192 static const struct ng_parse_type ng_pipe_cfg_type = {
193 	&ng_parse_struct_type,
194 	&ng_pipe_cfg_type_fields
195 };
196 
197 /* List of commands and how to convert arguments to/from ASCII */
198 static const struct ng_cmdlist ngp_cmds[] = {
199 	{
200 		.cookie =	NGM_PIPE_COOKIE,
201 		.cmd =		NGM_PIPE_GET_STATS,
202 		.name = 	"getstats",
203 		.respType =	 &ng_pipe_stats_type
204 	},
205 	{
206 		.cookie =	NGM_PIPE_COOKIE,
207 		.cmd =		NGM_PIPE_CLR_STATS,
208 		.name =		"clrstats"
209 	},
210 	{
211 		.cookie =	NGM_PIPE_COOKIE,
212 		.cmd =		NGM_PIPE_GETCLR_STATS,
213 		.name =		"getclrstats",
214 		.respType =	&ng_pipe_stats_type
215 	},
216 	{
217 		.cookie =	NGM_PIPE_COOKIE,
218 		.cmd =		NGM_PIPE_GET_RUN,
219 		.name =		"getrun",
220 		.respType =	&ng_pipe_run_type
221 	},
222 	{
223 		.cookie =	NGM_PIPE_COOKIE,
224 		.cmd =		NGM_PIPE_GET_CFG,
225 		.name =		"getcfg",
226 		.respType =	&ng_pipe_cfg_type
227 	},
228 	{
229 		.cookie =	NGM_PIPE_COOKIE,
230 		.cmd =		NGM_PIPE_SET_CFG,
231 		.name =		"setcfg",
232 		.mesgType =	&ng_pipe_cfg_type,
233 	},
234 	{ 0 }
235 };
236 
237 /* Netgraph type descriptor */
238 static struct ng_type ng_pipe_typestruct = {
239 	.version =	NG_ABI_VERSION,
240 	.name =		NG_PIPE_NODE_TYPE,
241 	.mod_event =	ngp_modevent,
242 	.constructor =	ngp_constructor,
243 	.shutdown =	ngp_shutdown,
244 	.rcvmsg =	ngp_rcvmsg,
245 	.newhook =	ngp_newhook,
246 	.rcvdata =	ngp_rcvdata,
247 	.disconnect =	ngp_disconnect,
248 	.cmdlist =	ngp_cmds
249 };
250 NETGRAPH_INIT(pipe, &ng_pipe_typestruct);
251 
252 /* Node constructor */
253 static int
254 ngp_constructor(node_p node)
255 {
256 	priv_p priv;
257 
258 	priv = malloc(sizeof(*priv), M_NG_PIPE, M_ZERO | M_WAITOK);
259 	NG_NODE_SET_PRIVATE(node, priv);
260 
261 	/* Mark node as single-threaded */
262 	NG_NODE_FORCE_WRITER(node);
263 
264 	ng_callout_init(&priv->timer);
265 
266 	return (0);
267 }
268 
269 /* Add a hook */
270 static int
271 ngp_newhook(node_p node, hook_p hook, const char *name)
272 {
273 	const priv_p priv = NG_NODE_PRIVATE(node);
274 	struct hookinfo *hinfo;
275 
276 	if (strcmp(name, NG_PIPE_HOOK_UPPER) == 0) {
277 		bzero(&priv->upper, sizeof(priv->upper));
278 		priv->upper.hook = hook;
279 		NG_HOOK_SET_PRIVATE(hook, &priv->upper);
280 	} else if (strcmp(name, NG_PIPE_HOOK_LOWER) == 0) {
281 		bzero(&priv->lower, sizeof(priv->lower));
282 		priv->lower.hook = hook;
283 		NG_HOOK_SET_PRIVATE(hook, &priv->lower);
284 	} else
285 		return (EINVAL);
286 
287 	/* Load non-zero initial cfg values */
288 	hinfo = NG_HOOK_PRIVATE(hook);
289 	hinfo->cfg.qin_size_limit = 50;
290 	hinfo->cfg.fifo = 1;
291 	hinfo->cfg.droptail = 1;
292 	TAILQ_INIT(&hinfo->fifo_head);
293 	TAILQ_INIT(&hinfo->qout_head);
294 	return (0);
295 }
296 
297 /* Receive a control message */
298 static int
299 ngp_rcvmsg(node_p node, item_p item, hook_p lasthook)
300 {
301 	const priv_p priv = NG_NODE_PRIVATE(node);
302 	struct ng_mesg *resp = NULL;
303 	struct ng_mesg *msg, *flow_msg;
304 	struct ng_pipe_stats *stats;
305 	struct ng_pipe_run *run;
306 	struct ng_pipe_cfg *cfg;
307 	int error = 0;
308 	int prev_down, now_down, cmd;
309 
310 	NGI_GET_MSG(item, msg);
311 	switch (msg->header.typecookie) {
312 	case NGM_PIPE_COOKIE:
313 		switch (msg->header.cmd) {
314 		case NGM_PIPE_GET_STATS:
315 		case NGM_PIPE_CLR_STATS:
316 		case NGM_PIPE_GETCLR_STATS:
317 			if (msg->header.cmd != NGM_PIPE_CLR_STATS) {
318 				NG_MKRESPONSE(resp, msg,
319 				    sizeof(*stats), M_NOWAIT);
320 				if (resp == NULL) {
321 					error = ENOMEM;
322 					break;
323 				}
324 				stats = (struct ng_pipe_stats *) resp->data;
325 				bcopy(&priv->upper.stats, &stats->downstream,
326 				    sizeof(stats->downstream));
327 				bcopy(&priv->lower.stats, &stats->upstream,
328 				    sizeof(stats->upstream));
329 			}
330 			if (msg->header.cmd != NGM_PIPE_GET_STATS) {
331 				bzero(&priv->upper.stats,
332 				    sizeof(priv->upper.stats));
333 				bzero(&priv->lower.stats,
334 				    sizeof(priv->lower.stats));
335 			}
336 			break;
337 		case NGM_PIPE_GET_RUN:
338 			NG_MKRESPONSE(resp, msg, sizeof(*run), M_NOWAIT);
339 			if (resp == NULL) {
340 				error = ENOMEM;
341 				break;
342 			}
343 			run = (struct ng_pipe_run *) resp->data;
344 			bcopy(&priv->upper.run, &run->downstream,
345 				sizeof(run->downstream));
346 			bcopy(&priv->lower.run, &run->upstream,
347 				sizeof(run->upstream));
348 			break;
349 		case NGM_PIPE_GET_CFG:
350 			NG_MKRESPONSE(resp, msg, sizeof(*cfg), M_NOWAIT);
351 			if (resp == NULL) {
352 				error = ENOMEM;
353 				break;
354 			}
355 			cfg = (struct ng_pipe_cfg *) resp->data;
356 			bcopy(&priv->upper.cfg, &cfg->downstream,
357 				sizeof(cfg->downstream));
358 			bcopy(&priv->lower.cfg, &cfg->upstream,
359 				sizeof(cfg->upstream));
360 			cfg->delay = priv->delay;
361 			cfg->overhead = priv->overhead;
362 			cfg->header_offset = priv->header_offset;
363 			if (cfg->upstream.bandwidth ==
364 			    cfg->downstream.bandwidth) {
365 				cfg->bandwidth = cfg->upstream.bandwidth;
366 				cfg->upstream.bandwidth = 0;
367 				cfg->downstream.bandwidth = 0;
368 			} else
369 				cfg->bandwidth = 0;
370 			break;
371 		case NGM_PIPE_SET_CFG:
372 			cfg = (struct ng_pipe_cfg *) msg->data;
373 			if (msg->header.arglen != sizeof(*cfg)) {
374 				error = EINVAL;
375 				break;
376 			}
377 
378 			if (cfg->delay == -1)
379 				priv->delay = 0;
380 			else if (cfg->delay > 0 && cfg->delay < 10000000)
381 				priv->delay = cfg->delay;
382 
383 			if (cfg->bandwidth == -1) {
384 				priv->upper.cfg.bandwidth = 0;
385 				priv->lower.cfg.bandwidth = 0;
386 				priv->overhead = 0;
387 			} else if (cfg->bandwidth >= 100 &&
388 			    cfg->bandwidth <= 1000000000) {
389 				priv->upper.cfg.bandwidth = cfg->bandwidth;
390 				priv->lower.cfg.bandwidth = cfg->bandwidth;
391 				if (cfg->bandwidth >= 10000000)
392 					priv->overhead = 8+4+12; /* Ethernet */
393 				else
394 					priv->overhead = 10; /* HDLC */
395 			}
396 
397 			if (cfg->overhead == -1)
398 				priv->overhead = 0;
399 			else if (cfg->overhead > 0 &&
400 			    cfg->overhead < MAX_OHSIZE)
401 				priv->overhead = cfg->overhead;
402 
403 			if (cfg->header_offset == -1)
404 				priv->header_offset = 0;
405 			else if (cfg->header_offset > 0 &&
406 			    cfg->header_offset < 64)
407 				priv->header_offset = cfg->header_offset;
408 
409 			prev_down = priv->upper.cfg.ber == 1 ||
410 			    priv->lower.cfg.ber == 1;
411 			parse_cfg(&priv->upper.cfg, &cfg->downstream,
412 			    &priv->upper, priv);
413 			parse_cfg(&priv->lower.cfg, &cfg->upstream,
414 			    &priv->lower, priv);
415 			now_down = priv->upper.cfg.ber == 1 ||
416 			    priv->lower.cfg.ber == 1;
417 
418 			if (prev_down != now_down) {
419 				if (now_down)
420 					cmd = NGM_LINK_IS_DOWN;
421 				else
422 					cmd = NGM_LINK_IS_UP;
423 
424 				if (priv->lower.hook != NULL) {
425 					NG_MKMESSAGE(flow_msg, NGM_FLOW_COOKIE,
426 					    cmd, 0, M_NOWAIT);
427 					if (flow_msg != NULL)
428 						NG_SEND_MSG_HOOK(error, node,
429 						    flow_msg, priv->lower.hook,
430 						    0);
431 				}
432 				if (priv->upper.hook != NULL) {
433 					NG_MKMESSAGE(flow_msg, NGM_FLOW_COOKIE,
434 					    cmd, 0, M_NOWAIT);
435 					if (flow_msg != NULL)
436 						NG_SEND_MSG_HOOK(error, node,
437 						    flow_msg, priv->upper.hook,
438 						    0);
439 				}
440 			}
441 			break;
442 		default:
443 			error = EINVAL;
444 			break;
445 		}
446 		break;
447 	default:
448 		error = EINVAL;
449 		break;
450 	}
451 	NG_RESPOND_MSG(error, node, item, resp);
452 	NG_FREE_MSG(msg);
453 
454 	return (error);
455 }
456 
457 static void
458 parse_cfg(struct ng_pipe_hookcfg *current, struct ng_pipe_hookcfg *new,
459 	struct hookinfo *hinfo, priv_p priv)
460 {
461 
462 	if (new->ber == -1) {
463 		current->ber = 0;
464 		if (hinfo->ber_p) {
465 			free(hinfo->ber_p, M_NG_PIPE);
466 			hinfo->ber_p = NULL;
467 		}
468 	} else if (new->ber >= 1 && new->ber <= 1000000000000) {
469 		static const uint64_t one = 0x1000000000000; /* = 2^48 */
470 		uint64_t p0, p;
471 		uint32_t fsize, i;
472 
473 		if (hinfo->ber_p == NULL)
474 			hinfo->ber_p =
475 			    malloc((MAX_FSIZE + MAX_OHSIZE) * sizeof(uint64_t),
476 			    M_NG_PIPE, M_WAITOK);
477 		current->ber = new->ber;
478 
479 		/*
480 		 * For given BER and each frame size N (in bytes) calculate
481 		 * the probability P_OK that the frame is clean:
482 		 *
483 		 * P_OK(BER,N) = (1 - 1/BER)^(N*8)
484 		 *
485 		 * We use a 64-bit fixed-point format with decimal point
486 		 * positioned between bits 47 and 48.
487 		 */
488 		p0 = one - one / new->ber;
489 		p = one;
490 		for (fsize = 0; fsize < MAX_FSIZE + MAX_OHSIZE; fsize++) {
491 			hinfo->ber_p[fsize] = p;
492 			for (i = 0; i < 8; i++)
493 				p = (p * (p0 & 0xffff) >> 48) +
494 				    (p * ((p0 >> 16) & 0xffff) >> 32) +
495 				    (p * (p0 >> 32) >> 16);
496 		}
497 	}
498 
499 	if (new->qin_size_limit == -1)
500 		current->qin_size_limit = 0;
501 	else if (new->qin_size_limit >= 5)
502 		current->qin_size_limit = new->qin_size_limit;
503 
504 	if (new->qout_size_limit == -1)
505 		current->qout_size_limit = 0;
506 	else if (new->qout_size_limit >= 5)
507 		current->qout_size_limit = new->qout_size_limit;
508 
509 	if (new->duplicate == -1)
510 		current->duplicate = 0;
511 	else if (new->duplicate > 0 && new->duplicate <= 50)
512 		current->duplicate = new->duplicate;
513 
514 	if (new->fifo) {
515 		current->fifo = 1;
516 		current->wfq = 0;
517 		current->drr = 0;
518 	}
519 
520 	if (new->wfq) {
521 		current->fifo = 0;
522 		current->wfq = 1;
523 		current->drr = 0;
524 	}
525 
526 	if (new->drr) {
527 		current->fifo = 0;
528 		current->wfq = 0;
529 		/* DRR quantum */
530 		if (new->drr >= 32)
531 			current->drr = new->drr;
532 		else
533 			current->drr = 2048;		/* default quantum */
534 	}
535 
536 	if (new->droptail) {
537 		current->droptail = 1;
538 		current->drophead = 0;
539 	}
540 
541 	if (new->drophead) {
542 		current->droptail = 0;
543 		current->drophead = 1;
544 	}
545 
546 	if (new->bandwidth == -1) {
547 		current->bandwidth = 0;
548 		current->fifo = 1;
549 		current->wfq = 0;
550 		current->drr = 0;
551 	} else if (new->bandwidth >= 100 && new->bandwidth <= 1000000000)
552 		current->bandwidth = new->bandwidth;
553 
554 	if (current->bandwidth | priv->delay |
555 	    current->duplicate | current->ber)
556 		hinfo->noqueue = 0;
557 	else
558 		hinfo->noqueue = 1;
559 }
560 
561 /*
562  * Compute a hash signature for a packet. This function suffers from the
563  * NIH sindrome, so probably it would be wise to look around what other
564  * folks have found out to be a good and efficient IP hash function...
565  */
566 static int
567 ip_hash(struct mbuf *m, int offset)
568 {
569 	u_int64_t i;
570 	struct ip *ip = (struct ip *)(mtod(m, u_char *) + offset);
571 
572 	if (m->m_len < sizeof(struct ip) + offset ||
573 	    ip->ip_v != 4 || ip->ip_hl << 2 != sizeof(struct ip))
574 		return 0;
575 
576 	i = ((u_int64_t) ip->ip_src.s_addr ^
577 	    ((u_int64_t) ip->ip_src.s_addr << 13) ^
578 	    ((u_int64_t) ip->ip_dst.s_addr << 7) ^
579 	    ((u_int64_t) ip->ip_dst.s_addr << 19));
580 	return (i ^ (i >> 32));
581 }
582 
583 /*
584  * Receive data on a hook - both in upstream and downstream direction.
585  * We put the frame on the inbound queue, and try to initiate dequeuing
586  * sequence immediately. If inbound queue is full, discard one frame
587  * depending on dropping policy (from the head or from the tail of the
588  * queue).
589  */
590 static int
591 ngp_rcvdata(hook_p hook, item_p item)
592 {
593 	struct hookinfo *const hinfo = NG_HOOK_PRIVATE(hook);
594 	const priv_p priv = NG_NODE_PRIVATE(NG_HOOK_NODE(hook));
595 	struct timeval uuptime;
596 	struct timeval *now = &uuptime;
597 	struct ngp_fifo *ngp_f = NULL, *ngp_f1;
598 	struct ngp_hdr *ngp_h = NULL;
599 	struct mbuf *m;
600 	int hash, plen;
601 	int error = 0;
602 
603 	/*
604 	 * Shortcut from inbound to outbound hook when neither of
605 	 * bandwidth, delay, BER or duplication probability is
606 	 * configured, nor we have queued frames to drain.
607 	 */
608 	if (hinfo->run.qin_frames == 0 && hinfo->run.qout_frames == 0 &&
609 	    hinfo->noqueue) {
610 		struct hookinfo *dest;
611 		if (hinfo == &priv->lower)
612 			dest = &priv->upper;
613 		else
614 			dest = &priv->lower;
615 
616 		/* Send the frame. */
617 		plen = NGI_M(item)->m_pkthdr.len;
618 		NG_FWD_ITEM_HOOK(error, item, dest->hook);
619 
620 		/* Update stats. */
621 		if (error) {
622 			hinfo->stats.out_disc_frames++;
623 			hinfo->stats.out_disc_octets += plen;
624 		} else {
625 			hinfo->stats.fwd_frames++;
626 			hinfo->stats.fwd_octets += plen;
627 		}
628 
629 		return (error);
630 	}
631 
632 	microuptime(now);
633 
634 	/*
635 	 * If this was an empty queue, update service deadline time.
636 	 */
637 	if (hinfo->run.qin_frames == 0) {
638 		struct timeval *when = &hinfo->qin_utime;
639 		if (when->tv_sec < now->tv_sec || (when->tv_sec == now->tv_sec
640 		    && when->tv_usec < now->tv_usec)) {
641 			when->tv_sec = now->tv_sec;
642 			when->tv_usec = now->tv_usec;
643 		}
644 	}
645 
646 	/* Populate the packet header */
647 	ngp_h = uma_zalloc(ngp_zone, M_NOWAIT);
648 	KASSERT((ngp_h != NULL), ("ngp_h zalloc failed (1)"));
649 	NGI_GET_M(item, m);
650 	KASSERT(m != NULL, ("NGI_GET_M failed"));
651 	ngp_h->m = m;
652 	NG_FREE_ITEM(item);
653 
654 	if (hinfo->cfg.fifo)
655 		hash = 0;	/* all packets go into a single FIFO queue */
656 	else
657 		hash = ip_hash(m, priv->header_offset);
658 
659 	/* Find the appropriate FIFO queue for the packet and enqueue it*/
660 	TAILQ_FOREACH(ngp_f, &hinfo->fifo_head, fifo_le)
661 		if (hash == ngp_f->hash)
662 			break;
663 	if (ngp_f == NULL) {
664 		ngp_f = uma_zalloc(ngp_zone, M_NOWAIT);
665 		KASSERT(ngp_h != NULL, ("ngp_h zalloc failed (2)"));
666 		TAILQ_INIT(&ngp_f->packet_head);
667 		ngp_f->hash = hash;
668 		ngp_f->packets = 1;
669 		ngp_f->rr_deficit = hinfo->cfg.drr;	/* DRR quantum */
670 		hinfo->run.fifo_queues++;
671 		TAILQ_INSERT_TAIL(&ngp_f->packet_head, ngp_h, ngp_link);
672 		FIFO_VTIME_SORT(m->m_pkthdr.len);
673 	} else {
674 		TAILQ_INSERT_TAIL(&ngp_f->packet_head, ngp_h, ngp_link);
675 		ngp_f->packets++;
676 	}
677 	hinfo->run.qin_frames++;
678 	hinfo->run.qin_octets += m->m_pkthdr.len;
679 
680 	/* Discard a frame if inbound queue limit has been reached */
681 	if (hinfo->run.qin_frames > hinfo->cfg.qin_size_limit) {
682 		struct mbuf *m1;
683 		int longest = 0;
684 
685 		/* Find the longest queue */
686 		TAILQ_FOREACH(ngp_f1, &hinfo->fifo_head, fifo_le)
687 			if (ngp_f1->packets > longest) {
688 				longest = ngp_f1->packets;
689 				ngp_f = ngp_f1;
690 			}
691 
692 		/* Drop a frame from the queue head/tail, depending on cfg */
693 		if (hinfo->cfg.drophead)
694 			ngp_h = TAILQ_FIRST(&ngp_f->packet_head);
695 		else
696 			ngp_h = TAILQ_LAST(&ngp_f->packet_head, p_head);
697 		TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
698 		m1 = ngp_h->m;
699 		uma_zfree(ngp_zone, ngp_h);
700 		hinfo->run.qin_octets -= m1->m_pkthdr.len;
701 		hinfo->stats.in_disc_octets += m1->m_pkthdr.len;
702 		m_freem(m1);
703 		if (--(ngp_f->packets) == 0) {
704 			TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
705 			uma_zfree(ngp_zone, ngp_f);
706 			hinfo->run.fifo_queues--;
707 		}
708 		hinfo->run.qin_frames--;
709 		hinfo->stats.in_disc_frames++;
710 	}
711 
712 	/*
713 	 * Try to start the dequeuing process immediately.
714 	 */
715 	pipe_dequeue(hinfo, now);
716 
717 	return (0);
718 }
719 
720 
721 /*
722  * Dequeueing sequence - we basically do the following:
723  *  1) Try to extract the frame from the inbound (bandwidth) queue;
724  *  2) In accordance to BER specified, discard the frame randomly;
725  *  3) If the frame survives BER, prepend it with delay info and move it
726  *     to outbound (delay) queue;
727  *  4) Loop to 2) until bandwidth quota for this timeslice is reached, or
728  *     inbound queue is flushed completely;
729  *  5) Dequeue frames from the outbound queue and send them downstream until
730  *     outbound queue is flushed completely, or the next frame in the queue
731  *     is not due to be dequeued yet
732  */
733 static void
734 pipe_dequeue(struct hookinfo *hinfo, struct timeval *now) {
735 	static uint64_t rand, oldrand;
736 	const node_p node = NG_HOOK_NODE(hinfo->hook);
737 	const priv_p priv = NG_NODE_PRIVATE(node);
738 	struct hookinfo *dest;
739 	struct ngp_fifo *ngp_f, *ngp_f1;
740 	struct ngp_hdr *ngp_h;
741 	struct timeval *when;
742 	struct mbuf *m;
743 	int plen, error = 0;
744 
745 	/* Which one is the destination hook? */
746 	if (hinfo == &priv->lower)
747 		dest = &priv->upper;
748 	else
749 		dest = &priv->lower;
750 
751 	/* Bandwidth queue processing */
752 	while ((ngp_f = TAILQ_FIRST(&hinfo->fifo_head))) {
753 		when = &hinfo->qin_utime;
754 		if (when->tv_sec > now->tv_sec || (when->tv_sec == now->tv_sec
755 		    && when->tv_usec > now->tv_usec))
756 			break;
757 
758 		ngp_h = TAILQ_FIRST(&ngp_f->packet_head);
759 		m = ngp_h->m;
760 
761 		/* Deficit Round Robin (DRR) processing */
762 		if (hinfo->cfg.drr) {
763 			if (ngp_f->rr_deficit >= m->m_pkthdr.len) {
764 				ngp_f->rr_deficit -= m->m_pkthdr.len;
765 			} else {
766 				ngp_f->rr_deficit += hinfo->cfg.drr;
767 				TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
768 				TAILQ_INSERT_TAIL(&hinfo->fifo_head,
769 				    ngp_f, fifo_le);
770 				continue;
771 			}
772 		}
773 
774 		/*
775 		 * Either create a duplicate and pass it on, or dequeue
776 		 * the original packet...
777 		 */
778 		if (hinfo->cfg.duplicate &&
779 		    random() % 100 <= hinfo->cfg.duplicate) {
780 			ngp_h = uma_zalloc(ngp_zone, M_NOWAIT);
781 			KASSERT(ngp_h != NULL, ("ngp_h zalloc failed (3)"));
782 			m = m_dup(m, M_NOWAIT);
783 			KASSERT(m != NULL, ("m_dup failed"));
784 			ngp_h->m = m;
785 		} else {
786 			TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
787 			hinfo->run.qin_frames--;
788 			hinfo->run.qin_octets -= m->m_pkthdr.len;
789 			ngp_f->packets--;
790 		}
791 
792 		/* Calculate the serialization delay */
793 		if (hinfo->cfg.bandwidth) {
794 			hinfo->qin_utime.tv_usec +=
795 			    ((uint64_t) m->m_pkthdr.len + priv->overhead ) *
796 			    8000000 / hinfo->cfg.bandwidth;
797 			hinfo->qin_utime.tv_sec +=
798 			    hinfo->qin_utime.tv_usec / 1000000;
799 			hinfo->qin_utime.tv_usec =
800 			    hinfo->qin_utime.tv_usec % 1000000;
801 		}
802 		when = &ngp_h->when;
803 		when->tv_sec = hinfo->qin_utime.tv_sec;
804 		when->tv_usec = hinfo->qin_utime.tv_usec;
805 
806 		/* Sort / rearrange inbound queues */
807 		if (ngp_f->packets) {
808 			if (hinfo->cfg.wfq) {
809 				TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
810 				FIFO_VTIME_SORT(TAILQ_FIRST(
811 				    &ngp_f->packet_head)->m->m_pkthdr.len)
812 			}
813 		} else {
814 			TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
815 			uma_zfree(ngp_zone, ngp_f);
816 			hinfo->run.fifo_queues--;
817 		}
818 
819 		/* Randomly discard the frame, according to BER setting */
820 		if (hinfo->cfg.ber) {
821 			oldrand = rand;
822 			rand = random();
823 			if (((oldrand ^ rand) << 17) >=
824 			    hinfo->ber_p[priv->overhead + m->m_pkthdr.len]) {
825 				hinfo->stats.out_disc_frames++;
826 				hinfo->stats.out_disc_octets += m->m_pkthdr.len;
827 				uma_zfree(ngp_zone, ngp_h);
828 				m_freem(m);
829 				continue;
830 			}
831 		}
832 
833 		/* Discard frame if outbound queue size limit exceeded */
834 		if (hinfo->cfg.qout_size_limit &&
835 		    hinfo->run.qout_frames>=hinfo->cfg.qout_size_limit) {
836 			hinfo->stats.out_disc_frames++;
837 			hinfo->stats.out_disc_octets += m->m_pkthdr.len;
838 			uma_zfree(ngp_zone, ngp_h);
839 			m_freem(m);
840 			continue;
841 		}
842 
843 		/* Calculate the propagation delay */
844 		when->tv_usec += priv->delay;
845 		when->tv_sec += when->tv_usec / 1000000;
846 		when->tv_usec = when->tv_usec % 1000000;
847 
848 		/* Put the frame into the delay queue */
849 		TAILQ_INSERT_TAIL(&hinfo->qout_head, ngp_h, ngp_link);
850 		hinfo->run.qout_frames++;
851 		hinfo->run.qout_octets += m->m_pkthdr.len;
852 	}
853 
854 	/* Delay queue processing */
855 	while ((ngp_h = TAILQ_FIRST(&hinfo->qout_head))) {
856 		when = &ngp_h->when;
857 		m = ngp_h->m;
858 		if (when->tv_sec > now->tv_sec ||
859 		    (when->tv_sec == now->tv_sec &&
860 		    when->tv_usec > now->tv_usec))
861 			break;
862 
863 		/* Update outbound queue stats */
864 		plen = m->m_pkthdr.len;
865 		hinfo->run.qout_frames--;
866 		hinfo->run.qout_octets -= plen;
867 
868 		/* Dequeue the packet from qout */
869 		TAILQ_REMOVE(&hinfo->qout_head, ngp_h, ngp_link);
870 		uma_zfree(ngp_zone, ngp_h);
871 
872 		NG_SEND_DATA(error, dest->hook, m, meta);
873 		if (error) {
874 			hinfo->stats.out_disc_frames++;
875 			hinfo->stats.out_disc_octets += plen;
876 		} else {
877 			hinfo->stats.fwd_frames++;
878 			hinfo->stats.fwd_octets += plen;
879 		}
880 	}
881 
882 	if ((hinfo->run.qin_frames != 0 || hinfo->run.qout_frames != 0) &&
883 	    !priv->timer_scheduled) {
884 		ng_callout(&priv->timer, node, NULL, 1, ngp_callout, NULL, 0);
885 		priv->timer_scheduled = 1;
886 	}
887 }
888 
889 /*
890  * This routine is called on every clock tick.  We poll connected hooks
891  * for queued frames by calling pipe_dequeue().
892  */
893 static void
894 ngp_callout(node_p node, hook_p hook, void *arg1, int arg2)
895 {
896 	const priv_p priv = NG_NODE_PRIVATE(node);
897 	struct timeval now;
898 
899 	priv->timer_scheduled = 0;
900 	microuptime(&now);
901 	if (priv->upper.hook != NULL)
902 		pipe_dequeue(&priv->upper, &now);
903 	if (priv->lower.hook != NULL)
904 		pipe_dequeue(&priv->lower, &now);
905 }
906 
907 /*
908  * Shutdown processing
909  *
910  * This is tricky. If we have both a lower and upper hook, then we
911  * probably want to extricate ourselves and leave the two peers
912  * still linked to each other. Otherwise we should just shut down as
913  * a normal node would.
914  */
915 static int
916 ngp_shutdown(node_p node)
917 {
918 	const priv_p priv = NG_NODE_PRIVATE(node);
919 
920 	if (priv->timer_scheduled)
921 		ng_uncallout(&priv->timer, node);
922 	if (priv->lower.hook && priv->upper.hook)
923 		ng_bypass(priv->lower.hook, priv->upper.hook);
924 	else {
925 		if (priv->upper.hook != NULL)
926 			ng_rmhook_self(priv->upper.hook);
927 		if (priv->lower.hook != NULL)
928 			ng_rmhook_self(priv->lower.hook);
929 	}
930 	NG_NODE_UNREF(node);
931 	free(priv, M_NG_PIPE);
932 	return (0);
933 }
934 
935 
936 /*
937  * Hook disconnection
938  */
939 static int
940 ngp_disconnect(hook_p hook)
941 {
942 	struct hookinfo *const hinfo = NG_HOOK_PRIVATE(hook);
943 	struct ngp_fifo *ngp_f;
944 	struct ngp_hdr *ngp_h;
945 
946 	KASSERT(hinfo != NULL, ("%s: null info", __FUNCTION__));
947 	hinfo->hook = NULL;
948 
949 	/* Flush all fifo queues associated with the hook */
950 	while ((ngp_f = TAILQ_FIRST(&hinfo->fifo_head))) {
951 		while ((ngp_h = TAILQ_FIRST(&ngp_f->packet_head))) {
952 			TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
953 			m_freem(ngp_h->m);
954 			uma_zfree(ngp_zone, ngp_h);
955 		}
956 		TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
957 		uma_zfree(ngp_zone, ngp_f);
958 	}
959 
960 	/* Flush the delay queue */
961 	while ((ngp_h = TAILQ_FIRST(&hinfo->qout_head))) {
962 		TAILQ_REMOVE(&hinfo->qout_head, ngp_h, ngp_link);
963 		m_freem(ngp_h->m);
964 		uma_zfree(ngp_zone, ngp_h);
965 	}
966 
967 	/* Release the packet loss probability table (BER) */
968 	if (hinfo->ber_p)
969 		free(hinfo->ber_p, M_NG_PIPE);
970 
971 	return (0);
972 }
973 
974 static int
975 ngp_modevent(module_t mod, int type, void *unused)
976 {
977 	int error = 0;
978 
979 	switch (type) {
980 	case MOD_LOAD:
981 		ngp_zone = uma_zcreate("ng_pipe", max(sizeof(struct ngp_hdr),
982 		    sizeof (struct ngp_fifo)), NULL, NULL, NULL, NULL,
983 		    UMA_ALIGN_PTR, 0);
984 		if (ngp_zone == NULL)
985 			panic("ng_pipe: couldn't allocate descriptor zone");
986 		break;
987 	case MOD_UNLOAD:
988 		uma_zdestroy(ngp_zone);
989 		break;
990 	default:
991 		error = EOPNOTSUPP;
992 		break;
993 	}
994 
995 	return (error);
996 }
997