xref: /freebsd/sys/netgraph/ng_pipe.c (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 
33 /*
34  * This node permits simple traffic shaping by emulating bandwidth
35  * and delay, as well as random packet losses.
36  * The node has two hooks, upper and lower. Traffic flowing from upper to
37  * lower hook is referenced as downstream, and vice versa. Parameters for
38  * both directions can be set separately, except for delay.
39  */
40 
41 #include <sys/param.h>
42 #include <sys/errno.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/mbuf.h>
47 #include <sys/prng.h>
48 #include <sys/time.h>
49 
50 #include <vm/uma.h>
51 
52 #include <net/vnet.h>
53 
54 #include <netinet/in.h>
55 #include <netinet/in_systm.h>
56 #include <netinet/ip.h>
57 
58 #include <netgraph/ng_message.h>
59 #include <netgraph/netgraph.h>
60 #include <netgraph/ng_parse.h>
61 #include <netgraph/ng_pipe.h>
62 
63 static MALLOC_DEFINE(M_NG_PIPE, "ng_pipe", "ng_pipe");
64 
65 /* Packet header struct */
66 struct ngp_hdr {
67 	TAILQ_ENTRY(ngp_hdr)	ngp_link;	/* next pkt in queue */
68 	struct timeval		when;		/* this packet's due time */
69 	struct mbuf		*m;		/* ptr to the packet data */
70 };
71 TAILQ_HEAD(p_head, ngp_hdr);
72 
73 /* FIFO queue struct */
74 struct ngp_fifo {
75 	TAILQ_ENTRY(ngp_fifo)	fifo_le;	/* list of active queues only */
76 	struct p_head		packet_head;	/* FIFO queue head */
77 	u_int32_t		hash;		/* flow signature */
78 	struct timeval		vtime;		/* virtual time, for WFQ */
79 	u_int32_t		rr_deficit;	/* for DRR */
80 	u_int32_t		packets;	/* # of packets in this queue */
81 };
82 
83 /* Per hook info */
84 struct hookinfo {
85 	hook_p			hook;
86 	int			noqueue;	/* bypass any processing */
87 	TAILQ_HEAD(, ngp_fifo)	fifo_head;	/* FIFO queues */
88 	TAILQ_HEAD(, ngp_hdr)	qout_head;	/* delay queue head */
89 	struct timeval		qin_utime;
90 	struct ng_pipe_hookcfg	cfg;
91 	struct ng_pipe_hookrun	run;
92 	struct ng_pipe_hookstat	stats;
93 	uint64_t		*ber_p;		/* loss_p(BER,psize) map */
94 };
95 
96 /* Per node info */
97 struct node_priv {
98 	u_int64_t		delay;
99 	u_int32_t		overhead;
100 	u_int32_t		header_offset;
101 	struct hookinfo		lower;
102 	struct hookinfo		upper;
103 	struct callout		timer;
104 	int			timer_scheduled;
105 };
106 typedef struct node_priv *priv_p;
107 
108 /* Macro for calculating the virtual time for packet dequeueing in WFQ */
109 #define FIFO_VTIME_SORT(plen)						\
110 	if (hinfo->cfg.wfq && hinfo->cfg.bandwidth) {			\
111 		ngp_f->vtime.tv_usec = now->tv_usec + ((uint64_t) (plen) \
112 			+ priv->overhead ) * hinfo->run.fifo_queues *	\
113 			8000000 / hinfo->cfg.bandwidth;			\
114 		ngp_f->vtime.tv_sec = now->tv_sec +			\
115 			ngp_f->vtime.tv_usec / 1000000;			\
116 		ngp_f->vtime.tv_usec = ngp_f->vtime.tv_usec % 1000000;	\
117 		TAILQ_FOREACH(ngp_f1, &hinfo->fifo_head, fifo_le)	\
118 			if (ngp_f1->vtime.tv_sec > ngp_f->vtime.tv_sec || \
119 			    (ngp_f1->vtime.tv_sec == ngp_f->vtime.tv_sec && \
120 			    ngp_f1->vtime.tv_usec > ngp_f->vtime.tv_usec)) \
121 				break;					\
122 		if (ngp_f1 == NULL)					\
123 			TAILQ_INSERT_TAIL(&hinfo->fifo_head, ngp_f, fifo_le); \
124 		else							\
125 			TAILQ_INSERT_BEFORE(ngp_f1, ngp_f, fifo_le);	\
126 	} else								\
127 		TAILQ_INSERT_TAIL(&hinfo->fifo_head, ngp_f, fifo_le);	\
128 
129 static void	parse_cfg(struct ng_pipe_hookcfg *, struct ng_pipe_hookcfg *,
130 			struct hookinfo *, priv_p);
131 static void	pipe_dequeue(struct hookinfo *, struct timeval *);
132 static void	ngp_callout(node_p, hook_p, void *, int);
133 static int	ngp_modevent(module_t, int, void *);
134 
135 /* zone for storing ngp_hdr-s */
136 static uma_zone_t ngp_zone;
137 
138 /* Netgraph methods */
139 static ng_constructor_t	ngp_constructor;
140 static ng_rcvmsg_t	ngp_rcvmsg;
141 static ng_shutdown_t	ngp_shutdown;
142 static ng_newhook_t	ngp_newhook;
143 static ng_rcvdata_t	ngp_rcvdata;
144 static ng_disconnect_t	ngp_disconnect;
145 
146 /* Parse type for struct ng_pipe_hookstat */
147 static const struct ng_parse_struct_field
148 	ng_pipe_hookstat_type_fields[] = NG_PIPE_HOOKSTAT_INFO;
149 static const struct ng_parse_type ng_pipe_hookstat_type = {
150 	&ng_parse_struct_type,
151 	&ng_pipe_hookstat_type_fields
152 };
153 
154 /* Parse type for struct ng_pipe_stats */
155 static const struct ng_parse_struct_field ng_pipe_stats_type_fields[] =
156 	NG_PIPE_STATS_INFO(&ng_pipe_hookstat_type);
157 static const struct ng_parse_type ng_pipe_stats_type = {
158 	&ng_parse_struct_type,
159 	&ng_pipe_stats_type_fields
160 };
161 
162 /* Parse type for struct ng_pipe_hookrun */
163 static const struct ng_parse_struct_field
164 	ng_pipe_hookrun_type_fields[] = NG_PIPE_HOOKRUN_INFO;
165 static const struct ng_parse_type ng_pipe_hookrun_type = {
166 	&ng_parse_struct_type,
167 	&ng_pipe_hookrun_type_fields
168 };
169 
170 /* Parse type for struct ng_pipe_run */
171 static const struct ng_parse_struct_field
172 	ng_pipe_run_type_fields[] = NG_PIPE_RUN_INFO(&ng_pipe_hookrun_type);
173 static const struct ng_parse_type ng_pipe_run_type = {
174 	&ng_parse_struct_type,
175 	&ng_pipe_run_type_fields
176 };
177 
178 /* Parse type for struct ng_pipe_hookcfg */
179 static const struct ng_parse_struct_field
180 	ng_pipe_hookcfg_type_fields[] = NG_PIPE_HOOKCFG_INFO;
181 static const struct ng_parse_type ng_pipe_hookcfg_type = {
182 	&ng_parse_struct_type,
183 	&ng_pipe_hookcfg_type_fields
184 };
185 
186 /* Parse type for struct ng_pipe_cfg */
187 static const struct ng_parse_struct_field
188 	ng_pipe_cfg_type_fields[] = NG_PIPE_CFG_INFO(&ng_pipe_hookcfg_type);
189 static const struct ng_parse_type ng_pipe_cfg_type = {
190 	&ng_parse_struct_type,
191 	&ng_pipe_cfg_type_fields
192 };
193 
194 /* List of commands and how to convert arguments to/from ASCII */
195 static const struct ng_cmdlist ngp_cmds[] = {
196 	{
197 		.cookie =	NGM_PIPE_COOKIE,
198 		.cmd =		NGM_PIPE_GET_STATS,
199 		.name =		"getstats",
200 		.respType =	&ng_pipe_stats_type
201 	},
202 	{
203 		.cookie =	NGM_PIPE_COOKIE,
204 		.cmd =		NGM_PIPE_CLR_STATS,
205 		.name =		"clrstats"
206 	},
207 	{
208 		.cookie =	NGM_PIPE_COOKIE,
209 		.cmd =		NGM_PIPE_GETCLR_STATS,
210 		.name =		"getclrstats",
211 		.respType =	&ng_pipe_stats_type
212 	},
213 	{
214 		.cookie =	NGM_PIPE_COOKIE,
215 		.cmd =		NGM_PIPE_GET_RUN,
216 		.name =		"getrun",
217 		.respType =	&ng_pipe_run_type
218 	},
219 	{
220 		.cookie =	NGM_PIPE_COOKIE,
221 		.cmd =		NGM_PIPE_GET_CFG,
222 		.name =		"getcfg",
223 		.respType =	&ng_pipe_cfg_type
224 	},
225 	{
226 		.cookie =	NGM_PIPE_COOKIE,
227 		.cmd =		NGM_PIPE_SET_CFG,
228 		.name =		"setcfg",
229 		.mesgType =	&ng_pipe_cfg_type,
230 	},
231 	{ 0 }
232 };
233 
234 /* Netgraph type descriptor */
235 static struct ng_type ng_pipe_typestruct = {
236 	.version =	NG_ABI_VERSION,
237 	.name =		NG_PIPE_NODE_TYPE,
238 	.mod_event =	ngp_modevent,
239 	.constructor =	ngp_constructor,
240 	.shutdown =	ngp_shutdown,
241 	.rcvmsg =	ngp_rcvmsg,
242 	.newhook =	ngp_newhook,
243 	.rcvdata =	ngp_rcvdata,
244 	.disconnect =	ngp_disconnect,
245 	.cmdlist =	ngp_cmds
246 };
247 NETGRAPH_INIT(pipe, &ng_pipe_typestruct);
248 
249 /* Node constructor */
250 static int
251 ngp_constructor(node_p node)
252 {
253 	priv_p priv;
254 
255 	priv = malloc(sizeof(*priv), M_NG_PIPE, M_ZERO | M_WAITOK);
256 	NG_NODE_SET_PRIVATE(node, priv);
257 
258 	/* Mark node as single-threaded */
259 	NG_NODE_FORCE_WRITER(node);
260 
261 	ng_callout_init(&priv->timer);
262 
263 	return (0);
264 }
265 
266 /* Add a hook */
267 static int
268 ngp_newhook(node_p node, hook_p hook, const char *name)
269 {
270 	const priv_p priv = NG_NODE_PRIVATE(node);
271 	struct hookinfo *hinfo;
272 
273 	if (strcmp(name, NG_PIPE_HOOK_UPPER) == 0) {
274 		bzero(&priv->upper, sizeof(priv->upper));
275 		priv->upper.hook = hook;
276 		NG_HOOK_SET_PRIVATE(hook, &priv->upper);
277 	} else if (strcmp(name, NG_PIPE_HOOK_LOWER) == 0) {
278 		bzero(&priv->lower, sizeof(priv->lower));
279 		priv->lower.hook = hook;
280 		NG_HOOK_SET_PRIVATE(hook, &priv->lower);
281 	} else
282 		return (EINVAL);
283 
284 	/* Load non-zero initial cfg values */
285 	hinfo = NG_HOOK_PRIVATE(hook);
286 	hinfo->cfg.qin_size_limit = 50;
287 	hinfo->cfg.fifo = 1;
288 	hinfo->cfg.droptail = 1;
289 	TAILQ_INIT(&hinfo->fifo_head);
290 	TAILQ_INIT(&hinfo->qout_head);
291 	return (0);
292 }
293 
294 /* Receive a control message */
295 static int
296 ngp_rcvmsg(node_p node, item_p item, hook_p lasthook)
297 {
298 	const priv_p priv = NG_NODE_PRIVATE(node);
299 	struct ng_mesg *resp = NULL;
300 	struct ng_mesg *msg, *flow_msg;
301 	struct ng_pipe_stats *stats;
302 	struct ng_pipe_run *run;
303 	struct ng_pipe_cfg *cfg;
304 	int error = 0;
305 	int prev_down, now_down, cmd;
306 
307 	NGI_GET_MSG(item, msg);
308 	switch (msg->header.typecookie) {
309 	case NGM_PIPE_COOKIE:
310 		switch (msg->header.cmd) {
311 		case NGM_PIPE_GET_STATS:
312 		case NGM_PIPE_CLR_STATS:
313 		case NGM_PIPE_GETCLR_STATS:
314 			if (msg->header.cmd != NGM_PIPE_CLR_STATS) {
315 				NG_MKRESPONSE(resp, msg,
316 				    sizeof(*stats), M_NOWAIT);
317 				if (resp == NULL) {
318 					error = ENOMEM;
319 					break;
320 				}
321 				stats = (struct ng_pipe_stats *) resp->data;
322 				bcopy(&priv->upper.stats, &stats->downstream,
323 				    sizeof(stats->downstream));
324 				bcopy(&priv->lower.stats, &stats->upstream,
325 				    sizeof(stats->upstream));
326 			}
327 			if (msg->header.cmd != NGM_PIPE_GET_STATS) {
328 				bzero(&priv->upper.stats,
329 				    sizeof(priv->upper.stats));
330 				bzero(&priv->lower.stats,
331 				    sizeof(priv->lower.stats));
332 			}
333 			break;
334 		case NGM_PIPE_GET_RUN:
335 			NG_MKRESPONSE(resp, msg, sizeof(*run), M_NOWAIT);
336 			if (resp == NULL) {
337 				error = ENOMEM;
338 				break;
339 			}
340 			run = (struct ng_pipe_run *) resp->data;
341 			bcopy(&priv->upper.run, &run->downstream,
342 				sizeof(run->downstream));
343 			bcopy(&priv->lower.run, &run->upstream,
344 				sizeof(run->upstream));
345 			break;
346 		case NGM_PIPE_GET_CFG:
347 			NG_MKRESPONSE(resp, msg, sizeof(*cfg), M_NOWAIT);
348 			if (resp == NULL) {
349 				error = ENOMEM;
350 				break;
351 			}
352 			cfg = (struct ng_pipe_cfg *) resp->data;
353 			bcopy(&priv->upper.cfg, &cfg->downstream,
354 				sizeof(cfg->downstream));
355 			bcopy(&priv->lower.cfg, &cfg->upstream,
356 				sizeof(cfg->upstream));
357 			cfg->delay = priv->delay;
358 			cfg->overhead = priv->overhead;
359 			cfg->header_offset = priv->header_offset;
360 			if (cfg->upstream.bandwidth ==
361 			    cfg->downstream.bandwidth) {
362 				cfg->bandwidth = cfg->upstream.bandwidth;
363 				cfg->upstream.bandwidth = 0;
364 				cfg->downstream.bandwidth = 0;
365 			} else
366 				cfg->bandwidth = 0;
367 			break;
368 		case NGM_PIPE_SET_CFG:
369 			cfg = (struct ng_pipe_cfg *) msg->data;
370 			if (msg->header.arglen != sizeof(*cfg)) {
371 				error = EINVAL;
372 				break;
373 			}
374 
375 			if (cfg->delay == -1)
376 				priv->delay = 0;
377 			else if (cfg->delay > 0 && cfg->delay < 10000000)
378 				priv->delay = cfg->delay;
379 
380 			if (cfg->bandwidth == -1) {
381 				priv->upper.cfg.bandwidth = 0;
382 				priv->lower.cfg.bandwidth = 0;
383 				priv->overhead = 0;
384 			} else if (cfg->bandwidth >= 100 &&
385 			    cfg->bandwidth <= 1000000000) {
386 				priv->upper.cfg.bandwidth = cfg->bandwidth;
387 				priv->lower.cfg.bandwidth = cfg->bandwidth;
388 				if (cfg->bandwidth >= 10000000)
389 					priv->overhead = 8+4+12; /* Ethernet */
390 				else
391 					priv->overhead = 10; /* HDLC */
392 			}
393 
394 			if (cfg->overhead == -1)
395 				priv->overhead = 0;
396 			else if (cfg->overhead > 0 &&
397 			    cfg->overhead < MAX_OHSIZE)
398 				priv->overhead = cfg->overhead;
399 
400 			if (cfg->header_offset == -1)
401 				priv->header_offset = 0;
402 			else if (cfg->header_offset > 0 &&
403 			    cfg->header_offset < 64)
404 				priv->header_offset = cfg->header_offset;
405 
406 			prev_down = priv->upper.cfg.ber == 1 ||
407 			    priv->lower.cfg.ber == 1;
408 			parse_cfg(&priv->upper.cfg, &cfg->downstream,
409 			    &priv->upper, priv);
410 			parse_cfg(&priv->lower.cfg, &cfg->upstream,
411 			    &priv->lower, priv);
412 			now_down = priv->upper.cfg.ber == 1 ||
413 			    priv->lower.cfg.ber == 1;
414 
415 			if (prev_down != now_down) {
416 				if (now_down)
417 					cmd = NGM_LINK_IS_DOWN;
418 				else
419 					cmd = NGM_LINK_IS_UP;
420 
421 				if (priv->lower.hook != NULL) {
422 					NG_MKMESSAGE(flow_msg, NGM_FLOW_COOKIE,
423 					    cmd, 0, M_NOWAIT);
424 					if (flow_msg != NULL)
425 						NG_SEND_MSG_HOOK(error, node,
426 						    flow_msg, priv->lower.hook,
427 						    0);
428 				}
429 				if (priv->upper.hook != NULL) {
430 					NG_MKMESSAGE(flow_msg, NGM_FLOW_COOKIE,
431 					    cmd, 0, M_NOWAIT);
432 					if (flow_msg != NULL)
433 						NG_SEND_MSG_HOOK(error, node,
434 						    flow_msg, priv->upper.hook,
435 						    0);
436 				}
437 			}
438 			break;
439 		default:
440 			error = EINVAL;
441 			break;
442 		}
443 		break;
444 	default:
445 		error = EINVAL;
446 		break;
447 	}
448 	NG_RESPOND_MSG(error, node, item, resp);
449 	NG_FREE_MSG(msg);
450 
451 	return (error);
452 }
453 
454 static void
455 parse_cfg(struct ng_pipe_hookcfg *current, struct ng_pipe_hookcfg *new,
456 	struct hookinfo *hinfo, priv_p priv)
457 {
458 
459 	if (new->ber == -1) {
460 		current->ber = 0;
461 		if (hinfo->ber_p) {
462 			free(hinfo->ber_p, M_NG_PIPE);
463 			hinfo->ber_p = NULL;
464 		}
465 	} else if (new->ber >= 1 && new->ber <= 1000000000000) {
466 		static const uint64_t one = 0x1000000000000; /* = 2^48 */
467 		uint64_t p0, p;
468 		uint32_t fsize, i;
469 
470 		if (hinfo->ber_p == NULL)
471 			hinfo->ber_p =
472 			    malloc((MAX_FSIZE + MAX_OHSIZE) * sizeof(uint64_t),
473 			    M_NG_PIPE, M_WAITOK);
474 		current->ber = new->ber;
475 
476 		/*
477 		 * For given BER and each frame size N (in bytes) calculate
478 		 * the probability P_OK that the frame is clean:
479 		 *
480 		 * P_OK(BER,N) = (1 - 1/BER)^(N*8)
481 		 *
482 		 * We use a 64-bit fixed-point format with decimal point
483 		 * positioned between bits 47 and 48.
484 		 */
485 		p0 = one - one / new->ber;
486 		p = one;
487 		for (fsize = 0; fsize < MAX_FSIZE + MAX_OHSIZE; fsize++) {
488 			hinfo->ber_p[fsize] = p;
489 			for (i = 0; i < 8; i++)
490 				p = (p * (p0 & 0xffff) >> 48) +
491 				    (p * ((p0 >> 16) & 0xffff) >> 32) +
492 				    (p * (p0 >> 32) >> 16);
493 		}
494 	}
495 
496 	if (new->qin_size_limit == -1)
497 		current->qin_size_limit = 0;
498 	else if (new->qin_size_limit >= 5)
499 		current->qin_size_limit = new->qin_size_limit;
500 
501 	if (new->qout_size_limit == -1)
502 		current->qout_size_limit = 0;
503 	else if (new->qout_size_limit >= 5)
504 		current->qout_size_limit = new->qout_size_limit;
505 
506 	if (new->duplicate == -1)
507 		current->duplicate = 0;
508 	else if (new->duplicate > 0 && new->duplicate <= 50)
509 		current->duplicate = new->duplicate;
510 
511 	if (new->fifo) {
512 		current->fifo = 1;
513 		current->wfq = 0;
514 		current->drr = 0;
515 	}
516 
517 	if (new->wfq) {
518 		current->fifo = 0;
519 		current->wfq = 1;
520 		current->drr = 0;
521 	}
522 
523 	if (new->drr) {
524 		current->fifo = 0;
525 		current->wfq = 0;
526 		/* DRR quantum */
527 		if (new->drr >= 32)
528 			current->drr = new->drr;
529 		else
530 			current->drr = 2048;		/* default quantum */
531 	}
532 
533 	if (new->droptail) {
534 		current->droptail = 1;
535 		current->drophead = 0;
536 	}
537 
538 	if (new->drophead) {
539 		current->droptail = 0;
540 		current->drophead = 1;
541 	}
542 
543 	if (new->bandwidth == -1) {
544 		current->bandwidth = 0;
545 		current->fifo = 1;
546 		current->wfq = 0;
547 		current->drr = 0;
548 	} else if (new->bandwidth >= 100 && new->bandwidth <= 1000000000)
549 		current->bandwidth = new->bandwidth;
550 
551 	if (current->bandwidth | priv->delay |
552 	    current->duplicate | current->ber)
553 		hinfo->noqueue = 0;
554 	else
555 		hinfo->noqueue = 1;
556 }
557 
558 /*
559  * Compute a hash signature for a packet. This function suffers from the
560  * NIH sindrome, so probably it would be wise to look around what other
561  * folks have found out to be a good and efficient IP hash function...
562  */
563 static int
564 ip_hash(struct mbuf *m, int offset)
565 {
566 	u_int64_t i;
567 	struct ip *ip = (struct ip *)(mtod(m, u_char *) + offset);
568 
569 	if (m->m_len < sizeof(struct ip) + offset ||
570 	    ip->ip_v != 4 || ip->ip_hl << 2 != sizeof(struct ip))
571 		return 0;
572 
573 	i = ((u_int64_t) ip->ip_src.s_addr ^
574 	    ((u_int64_t) ip->ip_src.s_addr << 13) ^
575 	    ((u_int64_t) ip->ip_dst.s_addr << 7) ^
576 	    ((u_int64_t) ip->ip_dst.s_addr << 19));
577 	return (i ^ (i >> 32));
578 }
579 
580 /*
581  * Receive data on a hook - both in upstream and downstream direction.
582  * We put the frame on the inbound queue, and try to initiate dequeuing
583  * sequence immediately. If inbound queue is full, discard one frame
584  * depending on dropping policy (from the head or from the tail of the
585  * queue).
586  */
587 static int
588 ngp_rcvdata(hook_p hook, item_p item)
589 {
590 	struct hookinfo *const hinfo = NG_HOOK_PRIVATE(hook);
591 	const priv_p priv = NG_NODE_PRIVATE(NG_HOOK_NODE(hook));
592 	struct timeval uuptime;
593 	struct timeval *now = &uuptime;
594 	struct ngp_fifo *ngp_f = NULL, *ngp_f1;
595 	struct ngp_hdr *ngp_h = NULL;
596 	struct mbuf *m;
597 	int hash, plen;
598 	int error = 0;
599 
600 	/*
601 	 * Shortcut from inbound to outbound hook when neither of
602 	 * bandwidth, delay, BER or duplication probability is
603 	 * configured, nor we have queued frames to drain.
604 	 */
605 	if (hinfo->run.qin_frames == 0 && hinfo->run.qout_frames == 0 &&
606 	    hinfo->noqueue) {
607 		struct hookinfo *dest;
608 		if (hinfo == &priv->lower)
609 			dest = &priv->upper;
610 		else
611 			dest = &priv->lower;
612 
613 		/* Send the frame. */
614 		plen = NGI_M(item)->m_pkthdr.len;
615 		NG_FWD_ITEM_HOOK(error, item, dest->hook);
616 
617 		/* Update stats. */
618 		if (error) {
619 			hinfo->stats.out_disc_frames++;
620 			hinfo->stats.out_disc_octets += plen;
621 		} else {
622 			hinfo->stats.fwd_frames++;
623 			hinfo->stats.fwd_octets += plen;
624 		}
625 
626 		return (error);
627 	}
628 
629 	microuptime(now);
630 
631 	/*
632 	 * If this was an empty queue, update service deadline time.
633 	 */
634 	if (hinfo->run.qin_frames == 0) {
635 		struct timeval *when = &hinfo->qin_utime;
636 		if (when->tv_sec < now->tv_sec || (when->tv_sec == now->tv_sec
637 		    && when->tv_usec < now->tv_usec)) {
638 			when->tv_sec = now->tv_sec;
639 			when->tv_usec = now->tv_usec;
640 		}
641 	}
642 
643 	/* Populate the packet header */
644 	NGI_GET_M(item, m);
645 	NG_FREE_ITEM(item);
646 	ngp_h = uma_zalloc(ngp_zone, M_NOWAIT);
647 	if (ngp_h == NULL) {
648 		NG_FREE_M(m);
649 		return (ENOMEM);
650 	}
651 	ngp_h->m = m;
652 
653 	if (hinfo->cfg.fifo)
654 		hash = 0;	/* all packets go into a single FIFO queue */
655 	else
656 		hash = ip_hash(m, priv->header_offset);
657 
658 	/* Find the appropriate FIFO queue for the packet and enqueue it*/
659 	TAILQ_FOREACH(ngp_f, &hinfo->fifo_head, fifo_le)
660 		if (hash == ngp_f->hash)
661 			break;
662 	if (ngp_f == NULL) {
663 		ngp_f = uma_zalloc(ngp_zone, M_NOWAIT);
664 		if (ngp_f == NULL) {
665 			NG_FREE_M(ngp_h->m);
666 			uma_zfree(ngp_zone, ngp_h);
667 			return (ENOMEM);
668 		}
669 		TAILQ_INIT(&ngp_f->packet_head);
670 		ngp_f->hash = hash;
671 		ngp_f->packets = 1;
672 		ngp_f->rr_deficit = hinfo->cfg.drr;	/* DRR quantum */
673 		hinfo->run.fifo_queues++;
674 		TAILQ_INSERT_TAIL(&ngp_f->packet_head, ngp_h, ngp_link);
675 		FIFO_VTIME_SORT(m->m_pkthdr.len);
676 	} else {
677 		TAILQ_INSERT_TAIL(&ngp_f->packet_head, ngp_h, ngp_link);
678 		ngp_f->packets++;
679 	}
680 	hinfo->run.qin_frames++;
681 	hinfo->run.qin_octets += m->m_pkthdr.len;
682 
683 	/* Discard a frame if inbound queue limit has been reached */
684 	if (hinfo->run.qin_frames > hinfo->cfg.qin_size_limit) {
685 		struct mbuf *m1;
686 		int longest = 0;
687 
688 		/* Find the longest queue */
689 		TAILQ_FOREACH(ngp_f1, &hinfo->fifo_head, fifo_le)
690 			if (ngp_f1->packets > longest) {
691 				longest = ngp_f1->packets;
692 				ngp_f = ngp_f1;
693 			}
694 
695 		/* Drop a frame from the queue head/tail, depending on cfg */
696 		if (hinfo->cfg.drophead)
697 			ngp_h = TAILQ_FIRST(&ngp_f->packet_head);
698 		else
699 			ngp_h = TAILQ_LAST(&ngp_f->packet_head, p_head);
700 		TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
701 		m1 = ngp_h->m;
702 		uma_zfree(ngp_zone, ngp_h);
703 		hinfo->run.qin_octets -= m1->m_pkthdr.len;
704 		hinfo->stats.in_disc_octets += m1->m_pkthdr.len;
705 		m_freem(m1);
706 		if (--(ngp_f->packets) == 0) {
707 			TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
708 			uma_zfree(ngp_zone, ngp_f);
709 			hinfo->run.fifo_queues--;
710 		}
711 		hinfo->run.qin_frames--;
712 		hinfo->stats.in_disc_frames++;
713 	}
714 
715 	/*
716 	 * Try to start the dequeuing process immediately.
717 	 */
718 	pipe_dequeue(hinfo, now);
719 
720 	return (0);
721 }
722 
723 /*
724  * Dequeueing sequence - we basically do the following:
725  *  1) Try to extract the frame from the inbound (bandwidth) queue;
726  *  2) In accordance to BER specified, discard the frame randomly;
727  *  3) If the frame survives BER, prepend it with delay info and move it
728  *     to outbound (delay) queue;
729  *  4) Loop to 2) until bandwidth quota for this timeslice is reached, or
730  *     inbound queue is flushed completely;
731  *  5) Dequeue frames from the outbound queue and send them downstream until
732  *     outbound queue is flushed completely, or the next frame in the queue
733  *     is not due to be dequeued yet
734  */
735 static void
736 pipe_dequeue(struct hookinfo *hinfo, struct timeval *now) {
737 	static uint64_t rand, oldrand;
738 	const node_p node = NG_HOOK_NODE(hinfo->hook);
739 	const priv_p priv = NG_NODE_PRIVATE(node);
740 	struct hookinfo *dest;
741 	struct ngp_fifo *ngp_f, *ngp_f1;
742 	struct ngp_hdr *ngp_h, *ngp_dup;
743 	struct timeval *when;
744 	struct mbuf *m;
745 	int plen, error = 0;
746 
747 	/* Which one is the destination hook? */
748 	if (hinfo == &priv->lower)
749 		dest = &priv->upper;
750 	else
751 		dest = &priv->lower;
752 
753 	/* Bandwidth queue processing */
754 	while ((ngp_f = TAILQ_FIRST(&hinfo->fifo_head))) {
755 		when = &hinfo->qin_utime;
756 		if (when->tv_sec > now->tv_sec || (when->tv_sec == now->tv_sec
757 		    && when->tv_usec > now->tv_usec))
758 			break;
759 
760 		ngp_h = TAILQ_FIRST(&ngp_f->packet_head);
761 		m = ngp_h->m;
762 
763 		/* Deficit Round Robin (DRR) processing */
764 		if (hinfo->cfg.drr) {
765 			if (ngp_f->rr_deficit >= m->m_pkthdr.len) {
766 				ngp_f->rr_deficit -= m->m_pkthdr.len;
767 			} else {
768 				ngp_f->rr_deficit += hinfo->cfg.drr;
769 				TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
770 				TAILQ_INSERT_TAIL(&hinfo->fifo_head,
771 				    ngp_f, fifo_le);
772 				continue;
773 			}
774 		}
775 
776 		/*
777 		 * Either create a duplicate and pass it on, or
778 		 * dequeue the original packet.  When any of the
779 		 * memory allocations for the duplicate package fail,
780 		 * simply do not duplicate it at all.
781 		 */
782 		ngp_dup = NULL;
783 		if (hinfo->cfg.duplicate &&
784 		    prng32_bounded(100) <= hinfo->cfg.duplicate) {
785 			ngp_dup = uma_zalloc(ngp_zone, M_NOWAIT);
786 			if (ngp_dup != NULL) {
787 				ngp_dup->m = m_dup(m, M_NOWAIT);
788 				if (ngp_dup->m == NULL) {
789 					uma_zfree(ngp_zone, ngp_dup);
790 					ngp_dup = NULL;
791 				}
792 			}
793 		}
794 
795 		if (ngp_dup != NULL) {
796 			ngp_h = ngp_dup;
797 			m = ngp_h->m;
798 		} else {
799 			TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
800 			hinfo->run.qin_frames--;
801 			hinfo->run.qin_octets -= m->m_pkthdr.len;
802 			ngp_f->packets--;
803 		}
804 
805 		/* Calculate the serialization delay */
806 		if (hinfo->cfg.bandwidth) {
807 			hinfo->qin_utime.tv_usec +=
808 			    ((uint64_t) m->m_pkthdr.len + priv->overhead ) *
809 			    8000000 / hinfo->cfg.bandwidth;
810 			hinfo->qin_utime.tv_sec +=
811 			    hinfo->qin_utime.tv_usec / 1000000;
812 			hinfo->qin_utime.tv_usec =
813 			    hinfo->qin_utime.tv_usec % 1000000;
814 		}
815 		when = &ngp_h->when;
816 		when->tv_sec = hinfo->qin_utime.tv_sec;
817 		when->tv_usec = hinfo->qin_utime.tv_usec;
818 
819 		/* Sort / rearrange inbound queues */
820 		if (ngp_f->packets) {
821 			if (hinfo->cfg.wfq) {
822 				TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
823 				FIFO_VTIME_SORT(TAILQ_FIRST(
824 				    &ngp_f->packet_head)->m->m_pkthdr.len)
825 			}
826 		} else {
827 			TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
828 			uma_zfree(ngp_zone, ngp_f);
829 			hinfo->run.fifo_queues--;
830 		}
831 
832 		/* Randomly discard the frame, according to BER setting */
833 		if (hinfo->cfg.ber) {
834 			oldrand = rand;
835 			rand = prng32_bounded(1U << 31);
836 			if (((oldrand ^ rand) << 17) >=
837 			    hinfo->ber_p[priv->overhead + m->m_pkthdr.len]) {
838 				hinfo->stats.out_disc_frames++;
839 				hinfo->stats.out_disc_octets += m->m_pkthdr.len;
840 				uma_zfree(ngp_zone, ngp_h);
841 				m_freem(m);
842 				continue;
843 			}
844 		}
845 
846 		/* Discard frame if outbound queue size limit exceeded */
847 		if (hinfo->cfg.qout_size_limit &&
848 		    hinfo->run.qout_frames>=hinfo->cfg.qout_size_limit) {
849 			hinfo->stats.out_disc_frames++;
850 			hinfo->stats.out_disc_octets += m->m_pkthdr.len;
851 			uma_zfree(ngp_zone, ngp_h);
852 			m_freem(m);
853 			continue;
854 		}
855 
856 		/* Calculate the propagation delay */
857 		when->tv_usec += priv->delay;
858 		when->tv_sec += when->tv_usec / 1000000;
859 		when->tv_usec = when->tv_usec % 1000000;
860 
861 		/* Put the frame into the delay queue */
862 		TAILQ_INSERT_TAIL(&hinfo->qout_head, ngp_h, ngp_link);
863 		hinfo->run.qout_frames++;
864 		hinfo->run.qout_octets += m->m_pkthdr.len;
865 	}
866 
867 	/* Delay queue processing */
868 	while ((ngp_h = TAILQ_FIRST(&hinfo->qout_head))) {
869 		when = &ngp_h->when;
870 		m = ngp_h->m;
871 		if (when->tv_sec > now->tv_sec ||
872 		    (when->tv_sec == now->tv_sec &&
873 		    when->tv_usec > now->tv_usec))
874 			break;
875 
876 		/* Update outbound queue stats */
877 		plen = m->m_pkthdr.len;
878 		hinfo->run.qout_frames--;
879 		hinfo->run.qout_octets -= plen;
880 
881 		/* Dequeue the packet from qout */
882 		TAILQ_REMOVE(&hinfo->qout_head, ngp_h, ngp_link);
883 		uma_zfree(ngp_zone, ngp_h);
884 
885 		NG_SEND_DATA(error, dest->hook, m, meta);
886 		if (error) {
887 			hinfo->stats.out_disc_frames++;
888 			hinfo->stats.out_disc_octets += plen;
889 		} else {
890 			hinfo->stats.fwd_frames++;
891 			hinfo->stats.fwd_octets += plen;
892 		}
893 	}
894 
895 	if ((hinfo->run.qin_frames != 0 || hinfo->run.qout_frames != 0) &&
896 	    !priv->timer_scheduled) {
897 		ng_callout(&priv->timer, node, NULL, 1, ngp_callout, NULL, 0);
898 		priv->timer_scheduled = 1;
899 	}
900 }
901 
902 /*
903  * This routine is called on every clock tick.  We poll connected hooks
904  * for queued frames by calling pipe_dequeue().
905  */
906 static void
907 ngp_callout(node_p node, hook_p hook, void *arg1, int arg2)
908 {
909 	const priv_p priv = NG_NODE_PRIVATE(node);
910 	struct timeval now;
911 
912 	priv->timer_scheduled = 0;
913 	microuptime(&now);
914 	if (priv->upper.hook != NULL)
915 		pipe_dequeue(&priv->upper, &now);
916 	if (priv->lower.hook != NULL)
917 		pipe_dequeue(&priv->lower, &now);
918 }
919 
920 /*
921  * Shutdown processing
922  *
923  * This is tricky. If we have both a lower and upper hook, then we
924  * probably want to extricate ourselves and leave the two peers
925  * still linked to each other. Otherwise we should just shut down as
926  * a normal node would.
927  */
928 static int
929 ngp_shutdown(node_p node)
930 {
931 	const priv_p priv = NG_NODE_PRIVATE(node);
932 
933 	if (priv->timer_scheduled)
934 		ng_uncallout(&priv->timer, node);
935 	if (priv->lower.hook && priv->upper.hook)
936 		ng_bypass(priv->lower.hook, priv->upper.hook);
937 	else {
938 		if (priv->upper.hook != NULL)
939 			ng_rmhook_self(priv->upper.hook);
940 		if (priv->lower.hook != NULL)
941 			ng_rmhook_self(priv->lower.hook);
942 	}
943 	NG_NODE_UNREF(node);
944 	free(priv, M_NG_PIPE);
945 	return (0);
946 }
947 
948 /*
949  * Hook disconnection
950  */
951 static int
952 ngp_disconnect(hook_p hook)
953 {
954 	struct hookinfo *const hinfo = NG_HOOK_PRIVATE(hook);
955 	struct ngp_fifo *ngp_f;
956 	struct ngp_hdr *ngp_h;
957 	priv_p priv;
958 
959 	KASSERT(hinfo != NULL, ("%s: null info", __FUNCTION__));
960 	hinfo->hook = NULL;
961 
962 	/* Flush all fifo queues associated with the hook */
963 	while ((ngp_f = TAILQ_FIRST(&hinfo->fifo_head))) {
964 		while ((ngp_h = TAILQ_FIRST(&ngp_f->packet_head))) {
965 			TAILQ_REMOVE(&ngp_f->packet_head, ngp_h, ngp_link);
966 			m_freem(ngp_h->m);
967 			uma_zfree(ngp_zone, ngp_h);
968 		}
969 		TAILQ_REMOVE(&hinfo->fifo_head, ngp_f, fifo_le);
970 		uma_zfree(ngp_zone, ngp_f);
971 	}
972 
973 	/* Flush the delay queue */
974 	while ((ngp_h = TAILQ_FIRST(&hinfo->qout_head))) {
975 		TAILQ_REMOVE(&hinfo->qout_head, ngp_h, ngp_link);
976 		m_freem(ngp_h->m);
977 		uma_zfree(ngp_zone, ngp_h);
978 	}
979 
980 	/* Release the packet loss probability table (BER) */
981 	if (hinfo->ber_p)
982 		free(hinfo->ber_p, M_NG_PIPE);
983 
984 	/* Destroy the node if all hooks are disconnected */
985 	priv = NG_NODE_PRIVATE(NG_HOOK_NODE(hook));
986 
987 	if (priv->upper.hook == NULL && priv->lower.hook == NULL)
988 		ng_rmnode_self(NG_HOOK_NODE(hook));
989 
990 	return (0);
991 }
992 
993 static int
994 ngp_modevent(module_t mod, int type, void *unused)
995 {
996 	int error = 0;
997 
998 	switch (type) {
999 	case MOD_LOAD:
1000 		ngp_zone = uma_zcreate("ng_pipe", max(sizeof(struct ngp_hdr),
1001 		    sizeof (struct ngp_fifo)), NULL, NULL, NULL, NULL,
1002 		    UMA_ALIGN_PTR, 0);
1003 		break;
1004 	case MOD_UNLOAD:
1005 		uma_zdestroy(ngp_zone);
1006 		break;
1007 	default:
1008 		error = EOPNOTSUPP;
1009 		break;
1010 	}
1011 
1012 	return (error);
1013 }
1014