xref: /freebsd/sys/netpfil/ipfw/ip_dn_io.c (revision 8c784bb8cf36911b828652f0bf7e88f443abec50)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2010 Luigi Rizzo, Riccardo Panicucci, Universita` di Pisa
5  * All rights reserved
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * Dummynet portions related to packet handling.
31  */
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet6.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
40 #include <sys/mbuf.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/module.h>
44 #include <sys/mutex.h>
45 #include <sys/priv.h>
46 #include <sys/proc.h>
47 #include <sys/rwlock.h>
48 #include <sys/socket.h>
49 #include <sys/time.h>
50 #include <sys/sysctl.h>
51 
52 #include <net/if.h>	/* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */
53 #include <net/if_var.h>	/* NET_EPOCH_... */
54 #include <net/if_private.h>
55 #include <net/netisr.h>
56 #include <net/vnet.h>
57 
58 #include <netinet/in.h>
59 #include <netinet/ip.h>		/* ip_len, ip_off */
60 #include <netinet/ip_var.h>	/* ip_output(), IP_FORWARDING */
61 #include <netinet/ip_fw.h>
62 #include <netinet/ip_dummynet.h>
63 #include <netinet/if_ether.h> /* various ether_* routines */
64 #include <netinet/ip6.h>       /* for ip6_input, ip6_output prototypes */
65 #include <netinet6/ip6_var.h>
66 
67 #include <netpfil/ipfw/ip_fw_private.h>
68 #include <netpfil/ipfw/dn_heap.h>
69 #include <netpfil/ipfw/ip_dn_private.h>
70 #ifdef NEW_AQM
71 #include <netpfil/ipfw/dn_aqm.h>
72 #endif
73 #include <netpfil/ipfw/dn_sched.h>
74 
75 /*
76  * We keep a private variable for the simulation time, but we could
77  * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
78  * instead of V_dn_cfg.curr_time
79  */
80 VNET_DEFINE(struct dn_parms, dn_cfg);
81 #define V_dn_cfg VNET(dn_cfg)
82 
83 /*
84  * We use a heap to store entities for which we have pending timer events.
85  * The heap is checked at every tick and all entities with expired events
86  * are extracted.
87  */
88 
89 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
90 
91 extern	void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
92 
93 #ifdef SYSCTL_NODE
94 
95 /*
96  * Because of the way the SYSBEGIN/SYSEND macros work on other
97  * platforms, there should not be functions between them.
98  * So keep the handlers outside the block.
99  */
100 static int
101 sysctl_hash_size(SYSCTL_HANDLER_ARGS)
102 {
103 	int error, value;
104 
105 	value = V_dn_cfg.hash_size;
106 	error = sysctl_handle_int(oidp, &value, 0, req);
107 	if (error != 0 || req->newptr == NULL)
108 		return (error);
109 	if (value < 16 || value > 65536)
110 		return (EINVAL);
111 	V_dn_cfg.hash_size = value;
112 	return (0);
113 }
114 
115 static int
116 sysctl_limits(SYSCTL_HANDLER_ARGS)
117 {
118 	int error;
119 	long value;
120 
121 	if (arg2 != 0)
122 		value = V_dn_cfg.slot_limit;
123 	else
124 		value = V_dn_cfg.byte_limit;
125 	error = sysctl_handle_long(oidp, &value, 0, req);
126 
127 	if (error != 0 || req->newptr == NULL)
128 		return (error);
129 	if (arg2 != 0) {
130 		if (value < 1)
131 			return (EINVAL);
132 		V_dn_cfg.slot_limit = value;
133 	} else {
134 		if (value < 1500)
135 			return (EINVAL);
136 		V_dn_cfg.byte_limit = value;
137 	}
138 	return (0);
139 }
140 
141 SYSBEGIN(f4)
142 
143 SYSCTL_DECL(_net_inet);
144 SYSCTL_DECL(_net_inet_ip);
145 #ifdef NEW_AQM
146 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
147     "Dummynet");
148 #else
149 static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
150     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
151     "Dummynet");
152 #endif
153 
154 /* wrapper to pass V_dn_cfg fields to SYSCTL_* */
155 #define DC(x)	(&(VNET_NAME(dn_cfg).x))
156 
157 /* parameters */
158 
159 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size,
160     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
161     0, 0, sysctl_hash_size, "I",
162     "Default hash table size");
163 
164 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
165     CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
166     0, 1, sysctl_limits, "L",
167     "Upper limit in slots for pipe queue.");
168 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
169     CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
170     0, 0, sysctl_limits, "L",
171     "Upper limit in bytes for pipe queue.");
172 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
173     CTLFLAG_RW | CTLFLAG_VNET, DC(io_fast), 0, "Enable fast dummynet io.");
174 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug,
175     CTLFLAG_RW | CTLFLAG_VNET, DC(debug), 0, "Dummynet debug level");
176 
177 /* RED parameters */
178 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
179     CTLFLAG_RD | CTLFLAG_VNET, DC(red_lookup_depth), 0, "Depth of RED lookup table");
180 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
181     CTLFLAG_RD | CTLFLAG_VNET, DC(red_avg_pkt_size), 0, "RED Medium packet size");
182 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
183     CTLFLAG_RD | CTLFLAG_VNET, DC(red_max_pkt_size), 0, "RED Max packet size");
184 
185 /* time adjustment */
186 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
187     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta), 0, "Last vs standard tick difference (usec).");
188 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
189     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta_sum), 0, "Accumulated tick difference (usec).");
190 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
191     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_adjustment), 0, "Tick adjustments done.");
192 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
193     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_diff), 0,
194     "Adjusted vs non-adjusted curr_time difference (ticks).");
195 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
196     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_lost), 0,
197     "Number of ticks coalesced by dummynet taskqueue.");
198 
199 /* Drain parameters */
200 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire,
201     CTLFLAG_RW | CTLFLAG_VNET, DC(expire), 0, "Expire empty queues/pipes");
202 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle,
203     CTLFLAG_RD | CTLFLAG_VNET, DC(expire_cycle), 0, "Expire cycle for queues/pipes");
204 
205 /* statistics */
206 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count,
207     CTLFLAG_RD | CTLFLAG_VNET, DC(schk_count), 0, "Number of schedulers");
208 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count,
209     CTLFLAG_RD | CTLFLAG_VNET, DC(si_count), 0, "Number of scheduler instances");
210 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count,
211     CTLFLAG_RD | CTLFLAG_VNET, DC(fsk_count), 0, "Number of flowsets");
212 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count,
213     CTLFLAG_RD | CTLFLAG_VNET, DC(queue_count), 0, "Number of queues");
214 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
215     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt), 0,
216     "Number of packets passed to dummynet.");
217 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
218     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_fast), 0,
219     "Number of packets bypassed dummynet scheduler.");
220 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
221     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_drop), 0,
222     "Number of packets dropped by dummynet.");
223 #undef DC
224 SYSEND
225 
226 #endif
227 
228 static void	dummynet_send(struct mbuf *);
229 
230 /*
231  * Return the mbuf tag holding the dummynet state (it should
232  * be the first one on the list).
233  */
234 struct dn_pkt_tag *
235 dn_tag_get(struct mbuf *m)
236 {
237 	struct m_tag *mtag = m_tag_first(m);
238 #ifdef NEW_AQM
239 	/* XXX: to skip ts m_tag. For Debugging only*/
240 	if (mtag != NULL && mtag->m_tag_id == DN_AQM_MTAG_TS) {
241 		m_tag_delete(m,mtag);
242 		mtag = m_tag_first(m);
243 		D("skip TS tag");
244 	}
245 #endif
246 	KASSERT(mtag != NULL &&
247 	    mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
248 	    mtag->m_tag_id == PACKET_TAG_DUMMYNET,
249 	    ("packet on dummynet queue w/o dummynet tag!"));
250 	return (struct dn_pkt_tag *)(mtag+1);
251 }
252 
253 #ifndef NEW_AQM
254 static inline void
255 mq_append(struct mq *q, struct mbuf *m)
256 {
257 #ifdef USERSPACE
258 	// buffers from netmap need to be copied
259 	// XXX note that the routine is not expected to fail
260 	ND("append %p to %p", m, q);
261 	if (m->m_flags & M_STACK) {
262 		struct mbuf *m_new;
263 		void *p;
264 		int l, ofs;
265 
266 		ofs = m->m_data - m->__m_extbuf;
267 		// XXX allocate
268 		MGETHDR(m_new, M_NOWAIT, MT_DATA);
269 		ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p",
270 			m, m->__m_extbuf, m->__m_extlen, ofs, m_new);
271 		p = m_new->__m_extbuf;	/* new pointer */
272 		l = m_new->__m_extlen;	/* new len */
273 		if (l <= m->__m_extlen) {
274 			panic("extlen too large");
275 		}
276 
277 		*m_new = *m;	// copy
278 		m_new->m_flags &= ~M_STACK;
279 		m_new->__m_extbuf = p; // point to new buffer
280 		_pkt_copy(m->__m_extbuf, p, m->__m_extlen);
281 		m_new->m_data = p + ofs;
282 		m = m_new;
283 	}
284 #endif /* USERSPACE */
285 	if (q->head == NULL)
286 		q->head = m;
287 	else
288 		q->tail->m_nextpkt = m;
289 	q->count++;
290 	q->tail = m;
291 	m->m_nextpkt = NULL;
292 }
293 #endif
294 
295 /*
296  * Dispose a list of packet. Use a functions so if we need to do
297  * more work, this is a central point to do it.
298  */
299 void dn_free_pkts(struct mbuf *mnext)
300 {
301         struct mbuf *m;
302 
303         while ((m = mnext) != NULL) {
304                 mnext = m->m_nextpkt;
305                 FREE_PKT(m);
306         }
307 }
308 
309 static int
310 red_drops (struct dn_queue *q, int len)
311 {
312 	/*
313 	 * RED algorithm
314 	 *
315 	 * RED calculates the average queue size (avg) using a low-pass filter
316 	 * with an exponential weighted (w_q) moving average:
317 	 * 	avg  <-  (1-w_q) * avg + w_q * q_size
318 	 * where q_size is the queue length (measured in bytes or * packets).
319 	 *
320 	 * If q_size == 0, we compute the idle time for the link, and set
321 	 *	avg = (1 - w_q)^(idle/s)
322 	 * where s is the time needed for transmitting a medium-sized packet.
323 	 *
324 	 * Now, if avg < min_th the packet is enqueued.
325 	 * If avg > max_th the packet is dropped. Otherwise, the packet is
326 	 * dropped with probability P function of avg.
327 	 */
328 
329 	struct dn_fsk *fs = q->fs;
330 	int64_t p_b = 0;
331 
332 	/* Queue in bytes or packets? */
333 	uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ?
334 	    q->ni.len_bytes : q->ni.length;
335 
336 	/* Average queue size estimation. */
337 	if (q_size != 0) {
338 		/* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
339 		int diff = SCALE(q_size) - q->avg;
340 		int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
341 
342 		q->avg += (int)v;
343 	} else {
344 		/*
345 		 * Queue is empty, find for how long the queue has been
346 		 * empty and use a lookup table for computing
347 		 * (1 - * w_q)^(idle_time/s) where s is the time to send a
348 		 * (small) packet.
349 		 * XXX check wraps...
350 		 */
351 		if (q->avg) {
352 			u_int t = div64((V_dn_cfg.curr_time - q->q_time), fs->lookup_step);
353 
354 			q->avg = (t < fs->lookup_depth) ?
355 			    SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
356 		}
357 	}
358 
359 	/* Should i drop? */
360 	if (q->avg < fs->min_th) {
361 		q->count = -1;
362 		return (0);	/* accept packet */
363 	}
364 	if (q->avg >= fs->max_th) {	/* average queue >=  max threshold */
365 		if (fs->fs.flags & DN_IS_ECN)
366 			return (1);
367 		if (fs->fs.flags & DN_IS_GENTLE_RED) {
368 			/*
369 			 * According to Gentle-RED, if avg is greater than
370 			 * max_th the packet is dropped with a probability
371 			 *	 p_b = c_3 * avg - c_4
372 			 * where c_3 = (1 - max_p) / max_th
373 			 *       c_4 = 1 - 2 * max_p
374 			 */
375 			p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
376 			    fs->c_4;
377 		} else {
378 			q->count = -1;
379 			return (1);
380 		}
381 	} else if (q->avg > fs->min_th) {
382 		if (fs->fs.flags & DN_IS_ECN)
383 			return (1);
384 		/*
385 		 * We compute p_b using the linear dropping function
386 		 *	 p_b = c_1 * avg - c_2
387 		 * where c_1 = max_p / (max_th - min_th)
388 		 * 	 c_2 = max_p * min_th / (max_th - min_th)
389 		 */
390 		p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
391 	}
392 
393 	if (fs->fs.flags & DN_QSIZE_BYTES)
394 		p_b = div64((p_b * len) , fs->max_pkt_size);
395 	if (++q->count == 0)
396 		q->random = random() & 0xffff;
397 	else {
398 		/*
399 		 * q->count counts packets arrived since last drop, so a greater
400 		 * value of q->count means a greater packet drop probability.
401 		 */
402 		if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
403 			q->count = 0;
404 			/* After a drop we calculate a new random value. */
405 			q->random = random() & 0xffff;
406 			return (1);	/* drop */
407 		}
408 	}
409 	/* End of RED algorithm. */
410 
411 	return (0);	/* accept */
412 
413 }
414 
415 /*
416  * ECN/ECT Processing (partially adopted from altq)
417  */
418 #ifndef NEW_AQM
419 static
420 #endif
421 int
422 ecn_mark(struct mbuf* m)
423 {
424 	struct ip *ip;
425 	ip = (struct ip *)mtodo(m, dn_tag_get(m)->iphdr_off);
426 
427 	switch (ip->ip_v) {
428 	case IPVERSION:
429 	{
430 		uint16_t old;
431 
432 		if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT)
433 			return (0);	/* not-ECT */
434 		if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE)
435 			return (1);	/* already marked */
436 
437 		/*
438 		 * ecn-capable but not marked,
439 		 * mark CE and update checksum
440 		 */
441 		old = *(uint16_t *)ip;
442 		ip->ip_tos |= IPTOS_ECN_CE;
443 		ip->ip_sum = cksum_adjust(ip->ip_sum, old, *(uint16_t *)ip);
444 		return (1);
445 	}
446 #ifdef INET6
447 	case (IPV6_VERSION >> 4):
448 	{
449 		struct ip6_hdr *ip6 = (struct ip6_hdr *)ip;
450 		u_int32_t flowlabel;
451 
452 		flowlabel = ntohl(ip6->ip6_flow);
453 		if ((flowlabel >> 28) != 6)
454 			return (0);	/* version mismatch! */
455 		if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
456 		    (IPTOS_ECN_NOTECT << 20))
457 			return (0);	/* not-ECT */
458 		if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
459 		    (IPTOS_ECN_CE << 20))
460 			return (1);	/* already marked */
461 		/*
462 		 * ecn-capable but not marked, mark CE
463 		 */
464 		flowlabel |= (IPTOS_ECN_CE << 20);
465 		ip6->ip6_flow = htonl(flowlabel);
466 		return (1);
467 	}
468 #endif
469 	}
470 	return (0);
471 }
472 
473 /*
474  * Enqueue a packet in q, subject to space and queue management policy
475  * (whose parameters are in q->fs).
476  * Update stats for the queue and the scheduler.
477  * Return 0 on success, 1 on drop. The packet is consumed anyways.
478  */
479 int
480 dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop)
481 {
482 	struct dn_fs *f;
483 	struct dn_flow *ni;	/* stats for scheduler instance */
484 	uint64_t len;
485 
486 	if (q->fs == NULL || q->_si == NULL) {
487 		printf("%s fs %p si %p, dropping\n",
488 			__FUNCTION__, q->fs, q->_si);
489 		FREE_PKT(m);
490 		return 1;
491 	}
492 	f = &(q->fs->fs);
493 	ni = &q->_si->ni;
494 	len = m->m_pkthdr.len;
495 	/* Update statistics, then check reasons to drop pkt. */
496 	q->ni.tot_bytes += len;
497 	q->ni.tot_pkts++;
498 	ni->tot_bytes += len;
499 	ni->tot_pkts++;
500 	if (drop)
501 		goto drop;
502 	if (f->plr && random() < f->plr)
503 		goto drop;
504 	if (m->m_pkthdr.rcvif != NULL)
505 		m_rcvif_serialize(m);
506 #ifdef NEW_AQM
507 	/* Call AQM enqueue function */
508 	if (q->fs->aqmfp)
509 		return q->fs->aqmfp->enqueue(q ,m);
510 #endif
511 	if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) {
512 		if (!(f->flags & DN_IS_ECN) || !ecn_mark(m))
513 			goto drop;
514 	}
515 	if (f->flags & DN_QSIZE_BYTES) {
516 		if (q->ni.len_bytes > f->qsize)
517 			goto drop;
518 	} else if (q->ni.length >= f->qsize) {
519 		goto drop;
520 	}
521 	mq_append(&q->mq, m);
522 	q->ni.length++;
523 	q->ni.len_bytes += len;
524 	ni->length++;
525 	ni->len_bytes += len;
526 	return (0);
527 
528 drop:
529 	V_dn_cfg.io_pkt_drop++;
530 	q->ni.drops++;
531 	ni->drops++;
532 	FREE_PKT(m);
533 	return (1);
534 }
535 
536 /*
537  * Fetch packets from the delay line which are due now. If there are
538  * leftover packets, reinsert the delay line in the heap.
539  * Runs under scheduler lock.
540  */
541 static void
542 transmit_event(struct mq *q, struct delay_line *dline, uint64_t now)
543 {
544 	struct mbuf *m;
545 	struct dn_pkt_tag *pkt = NULL;
546 
547 	dline->oid.subtype = 0; /* not in heap */
548 	while ((m = dline->mq.head) != NULL) {
549 		pkt = dn_tag_get(m);
550 		if (!DN_KEY_LEQ(pkt->output_time, now))
551 			break;
552 		dline->mq.head = m->m_nextpkt;
553 		dline->mq.count--;
554 		if (m->m_pkthdr.rcvif != NULL &&
555 		  __predict_false(m_rcvif_restore(m) == NULL))
556 			m_freem(m);
557 		else
558 			mq_append(q, m);
559 	}
560 	if (m != NULL) {
561 		dline->oid.subtype = 1; /* in heap */
562 		heap_insert(&V_dn_cfg.evheap, pkt->output_time, dline);
563 	}
564 }
565 
566 /*
567  * Convert the additional MAC overheads/delays into an equivalent
568  * number of bits for the given data rate. The samples are
569  * in milliseconds so we need to divide by 1000.
570  */
571 static uint64_t
572 extra_bits(struct mbuf *m, struct dn_schk *s)
573 {
574 	int index;
575 	uint64_t bits;
576 	struct dn_profile *pf = s->profile;
577 
578 	if (!pf || pf->samples_no == 0)
579 		return 0;
580 	index  = random() % pf->samples_no;
581 	bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000);
582 	if (index >= pf->loss_level) {
583 		struct dn_pkt_tag *dt = dn_tag_get(m);
584 		if (dt)
585 			dt->dn_dir = DIR_DROP;
586 	}
587 	return bits;
588 }
589 
590 /*
591  * Send traffic from a scheduler instance due by 'now'.
592  * Return a pointer to the head of the queue.
593  */
594 static struct mbuf *
595 serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now)
596 {
597 	struct mq def_q;
598 	struct dn_schk *s = si->sched;
599 	struct mbuf *m = NULL;
600 	int delay_line_idle = (si->dline.mq.head == NULL);
601 	int done;
602 	uint32_t bw;
603 
604 	if (q == NULL) {
605 		q = &def_q;
606 		q->head = NULL;
607 	}
608 
609 	bw = s->link.bandwidth;
610 	si->kflags &= ~DN_ACTIVE;
611 
612 	if (bw > 0)
613 		si->credit += (now - si->sched_time) * bw;
614 	else
615 		si->credit = 0;
616 	si->sched_time = now;
617 	done = 0;
618 	while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) {
619 		uint64_t len_scaled;
620 
621 		done++;
622 		len_scaled = (bw == 0) ? 0 : hz *
623 			(m->m_pkthdr.len * 8 + extra_bits(m, s));
624 		si->credit -= len_scaled;
625 		/* Move packet in the delay line */
626 		dn_tag_get(m)->output_time = V_dn_cfg.curr_time + s->link.delay ;
627 		if (m->m_pkthdr.rcvif != NULL)
628 			m_rcvif_serialize(m);
629 		mq_append(&si->dline.mq, m);
630 	}
631 
632 	/*
633 	 * If credit >= 0 the instance is idle, mark time.
634 	 * Otherwise put back in the heap, and adjust the output
635 	 * time of the last inserted packet, m, which was too early.
636 	 */
637 	if (si->credit >= 0) {
638 		si->idle_time = now;
639 	} else {
640 		uint64_t t;
641 		KASSERT (bw > 0, ("bw=0 and credit<0 ?"));
642 		t = div64(bw - 1 - si->credit, bw);
643 		if (m)
644 			dn_tag_get(m)->output_time += t;
645 		si->kflags |= DN_ACTIVE;
646 		heap_insert(&V_dn_cfg.evheap, now + t, si);
647 	}
648 	if (delay_line_idle && done)
649 		transmit_event(q, &si->dline, now);
650 	return q->head;
651 }
652 
653 /*
654  * The timer handler for dummynet. Time is computed in ticks, but
655  * but the code is tolerant to the actual rate at which this is called.
656  * Once complete, the function reschedules itself for the next tick.
657  */
658 void
659 dummynet_task(void *context, int pending)
660 {
661 	struct timeval t;
662 	struct mq q = { NULL, NULL }; /* queue to accumulate results */
663 	struct epoch_tracker et;
664 
665 	VNET_ITERATOR_DECL(vnet_iter);
666 	VNET_LIST_RLOCK();
667 	NET_EPOCH_ENTER(et);
668 
669 	VNET_FOREACH(vnet_iter) {
670 		memset(&q, 0, sizeof(struct mq));
671 		CURVNET_SET(vnet_iter);
672 
673 		if (! V_dn_cfg.init_done) {
674 			CURVNET_RESTORE();
675 			continue;
676 		}
677 
678 		DN_BH_WLOCK();
679 
680 		/* Update number of lost(coalesced) ticks. */
681 		V_dn_cfg.tick_lost += pending - 1;
682 
683 		getmicrouptime(&t);
684 		/* Last tick duration (usec). */
685 		V_dn_cfg.tick_last = (t.tv_sec - V_dn_cfg.prev_t.tv_sec) * 1000000 +
686 		(t.tv_usec - V_dn_cfg.prev_t.tv_usec);
687 		/* Last tick vs standard tick difference (usec). */
688 		V_dn_cfg.tick_delta = (V_dn_cfg.tick_last * hz - 1000000) / hz;
689 		/* Accumulated tick difference (usec). */
690 		V_dn_cfg.tick_delta_sum += V_dn_cfg.tick_delta;
691 
692 		V_dn_cfg.prev_t = t;
693 
694 		/*
695 		* Adjust curr_time if the accumulated tick difference is
696 		* greater than the 'standard' tick. Since curr_time should
697 		* be monotonically increasing, we do positive adjustments
698 		* as required, and throttle curr_time in case of negative
699 		* adjustment.
700 		*/
701 		V_dn_cfg.curr_time++;
702 		if (V_dn_cfg.tick_delta_sum - tick >= 0) {
703 			int diff = V_dn_cfg.tick_delta_sum / tick;
704 
705 			V_dn_cfg.curr_time += diff;
706 			V_dn_cfg.tick_diff += diff;
707 			V_dn_cfg.tick_delta_sum %= tick;
708 			V_dn_cfg.tick_adjustment++;
709 		} else if (V_dn_cfg.tick_delta_sum + tick <= 0) {
710 			V_dn_cfg.curr_time--;
711 			V_dn_cfg.tick_diff--;
712 			V_dn_cfg.tick_delta_sum += tick;
713 			V_dn_cfg.tick_adjustment++;
714 		}
715 
716 		/* serve pending events, accumulate in q */
717 		for (;;) {
718 			struct dn_id *p;    /* generic parameter to handler */
719 
720 			if (V_dn_cfg.evheap.elements == 0 ||
721 			    DN_KEY_LT(V_dn_cfg.curr_time, HEAP_TOP(&V_dn_cfg.evheap)->key))
722 				break;
723 			p = HEAP_TOP(&V_dn_cfg.evheap)->object;
724 			heap_extract(&V_dn_cfg.evheap, NULL);
725 			if (p->type == DN_SCH_I) {
726 				serve_sched(&q, (struct dn_sch_inst *)p, V_dn_cfg.curr_time);
727 			} else { /* extracted a delay line */
728 				transmit_event(&q, (struct delay_line *)p, V_dn_cfg.curr_time);
729 			}
730 		}
731 		if (V_dn_cfg.expire && ++V_dn_cfg.expire_cycle >= V_dn_cfg.expire) {
732 			V_dn_cfg.expire_cycle = 0;
733 			dn_drain_scheduler();
734 			dn_drain_queue();
735 		}
736 		DN_BH_WUNLOCK();
737 		if (q.head != NULL)
738 			dummynet_send(q.head);
739 
740 		CURVNET_RESTORE();
741 	}
742 	NET_EPOCH_EXIT(et);
743 	VNET_LIST_RUNLOCK();
744 
745 	/* Schedule our next run. */
746 	dn_reschedule();
747 }
748 
749 /*
750  * forward a chain of packets to the proper destination.
751  * This runs outside the dummynet lock.
752  */
753 static void
754 dummynet_send(struct mbuf *m)
755 {
756 	struct mbuf *n;
757 
758 	NET_EPOCH_ASSERT();
759 
760 	for (; m != NULL; m = n) {
761 		struct ifnet *ifp = NULL;	/* gcc 3.4.6 complains */
762         	struct m_tag *tag;
763 		int dst;
764 
765 		n = m->m_nextpkt;
766 		m->m_nextpkt = NULL;
767 		tag = m_tag_first(m);
768 		if (tag == NULL) { /* should not happen */
769 			dst = DIR_DROP;
770 		} else {
771 			struct dn_pkt_tag *pkt = dn_tag_get(m);
772 			/* extract the dummynet info, rename the tag
773 			 * to carry reinject info.
774 			 */
775 			ifp = ifnet_byindexgen(pkt->if_index, pkt->if_idxgen);
776 			if (((pkt->dn_dir == (DIR_OUT | PROTO_LAYER2)) ||
777 			    (pkt->dn_dir == (DIR_OUT | PROTO_LAYER2 | PROTO_IPV6))) &&
778 				ifp == NULL) {
779 				dst = DIR_DROP;
780 			} else {
781 				dst = pkt->dn_dir;
782 				tag->m_tag_cookie = MTAG_IPFW_RULE;
783 				tag->m_tag_id = 0;
784 			}
785 		}
786 
787 		switch (dst) {
788 		case DIR_OUT:
789 			ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
790 			break ;
791 
792 		case DIR_IN :
793 			netisr_dispatch(NETISR_IP, m);
794 			break;
795 
796 #ifdef INET6
797 		case DIR_IN | PROTO_IPV6:
798 			netisr_dispatch(NETISR_IPV6, m);
799 			break;
800 
801 		case DIR_OUT | PROTO_IPV6:
802 			ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
803 			break;
804 #endif
805 
806 		case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */
807 			if (bridge_dn_p != NULL)
808 				((*bridge_dn_p)(m, ifp));
809 			else
810 				printf("dummynet: if_bridge not loaded\n");
811 
812 			break;
813 
814 		case DIR_IN | PROTO_LAYER2 | PROTO_IPV6:
815 		case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */
816 			/*
817 			 * The Ethernet code assumes the Ethernet header is
818 			 * contiguous in the first mbuf header.
819 			 * Insure this is true.
820 			 */
821 			if (m->m_len < ETHER_HDR_LEN &&
822 			    (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
823 				printf("dummynet/ether: pullup failed, "
824 				    "dropping packet\n");
825 				break;
826 			}
827 			ether_demux(m->m_pkthdr.rcvif, m);
828 			break;
829 
830 		case DIR_OUT | PROTO_LAYER2 | PROTO_IPV6:
831 		case DIR_OUT | PROTO_LAYER2: /* DN_TO_ETH_OUT: */
832 			MPASS(ifp != NULL);
833 			ether_output_frame(ifp, m);
834 			break;
835 
836 		case DIR_DROP:
837 			/* drop the packet after some time */
838 			FREE_PKT(m);
839 			break;
840 
841 		default:
842 			printf("dummynet: bad switch %d!\n", dst);
843 			FREE_PKT(m);
844 			break;
845 		}
846 	}
847 }
848 
849 static inline int
850 tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa)
851 {
852 	struct dn_pkt_tag *dt;
853 	struct m_tag *mtag;
854 
855 	mtag = m_tag_get(PACKET_TAG_DUMMYNET,
856 		    sizeof(*dt), M_NOWAIT | M_ZERO);
857 	if (mtag == NULL)
858 		return 1;		/* Cannot allocate packet header. */
859 	m_tag_prepend(m, mtag);		/* Attach to mbuf chain. */
860 	dt = (struct dn_pkt_tag *)(mtag + 1);
861 	dt->rule = fwa->rule;
862 	/* only keep this info */
863 	dt->rule.info &= (IPFW_ONEPASS | IPFW_IS_DUMMYNET);
864 	dt->dn_dir = dir;
865 	if (fwa->flags & IPFW_ARGS_OUT && fwa->ifp != NULL) {
866 		NET_EPOCH_ASSERT();
867 		dt->if_index = fwa->ifp->if_index;
868 		dt->if_idxgen = fwa->ifp->if_idxgen;
869 	}
870 	/* dt->output tame is updated as we move through */
871 	dt->output_time = V_dn_cfg.curr_time;
872 	dt->iphdr_off = (dir & PROTO_LAYER2) ? ETHER_HDR_LEN : 0;
873 	return 0;
874 }
875 
876 /*
877  * dummynet hook for packets.
878  * We use the argument to locate the flowset fs and the sched_set sch
879  * associated to it. The we apply flow_mask and sched_mask to
880  * determine the queue and scheduler instances.
881  */
882 int
883 dummynet_io(struct mbuf **m0, struct ip_fw_args *fwa)
884 {
885 	struct mbuf *m = *m0;
886 	struct dn_fsk *fs = NULL;
887 	struct dn_sch_inst *si;
888 	struct dn_queue *q = NULL;	/* default */
889 	int fs_id, dir;
890 
891 	fs_id = (fwa->rule.info & IPFW_INFO_MASK) +
892 		((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0);
893 	/* XXXGL: convert args to dir */
894 	if (fwa->flags & IPFW_ARGS_IN)
895 		dir = DIR_IN;
896 	else
897 		dir = DIR_OUT;
898 	if (fwa->flags & IPFW_ARGS_ETHER)
899 		dir |= PROTO_LAYER2;
900 	else if (fwa->flags & IPFW_ARGS_IP6)
901 		dir |= PROTO_IPV6;
902 	DN_BH_WLOCK();
903 	V_dn_cfg.io_pkt++;
904 	/* we could actually tag outside the lock, but who cares... */
905 	if (tag_mbuf(m, dir, fwa))
906 		goto dropit;
907 	/* XXX locate_flowset could be optimised with a direct ref. */
908 	fs = dn_ht_find(V_dn_cfg.fshash, fs_id, 0, NULL);
909 	if (fs == NULL)
910 		goto dropit;	/* This queue/pipe does not exist! */
911 	if (fs->sched == NULL)	/* should not happen */
912 		goto dropit;
913 	/* find scheduler instance, possibly applying sched_mask */
914 	si = ipdn_si_find(fs->sched, &(fwa->f_id));
915 	if (si == NULL)
916 		goto dropit;
917 	/*
918 	 * If the scheduler supports multiple queues, find the right one
919 	 * (otherwise it will be ignored by enqueue).
920 	 */
921 	if (fs->sched->fp->flags & DN_MULTIQUEUE) {
922 		q = ipdn_q_find(fs, si, &(fwa->f_id));
923 		if (q == NULL)
924 			goto dropit;
925 	}
926 	if (fs->sched->fp->enqueue(si, q, m)) {
927 		/* packet was dropped by enqueue() */
928 		m = *m0 = NULL;
929 
930 		/* dn_enqueue already increases io_pkt_drop */
931 		V_dn_cfg.io_pkt_drop--;
932 
933 		goto dropit;
934 	}
935 
936 	if (si->kflags & DN_ACTIVE) {
937 		m = *m0 = NULL; /* consumed */
938 		goto done; /* already active, nothing to do */
939 	}
940 
941 	/* compute the initial allowance */
942 	if (si->idle_time < V_dn_cfg.curr_time) {
943 	    /* Do this only on the first packet on an idle pipe */
944 	    struct dn_link *p = &fs->sched->link;
945 
946 	    si->sched_time = V_dn_cfg.curr_time;
947 	    si->credit = V_dn_cfg.io_fast ? p->bandwidth : 0;
948 	    if (p->burst) {
949 		uint64_t burst = (V_dn_cfg.curr_time - si->idle_time) * p->bandwidth;
950 		if (burst > p->burst)
951 			burst = p->burst;
952 		si->credit += burst;
953 	    }
954 	}
955 	/* pass through scheduler and delay line */
956 	m = serve_sched(NULL, si, V_dn_cfg.curr_time);
957 
958 	/* optimization -- pass it back to ipfw for immediate send */
959 	/* XXX Don't call dummynet_send() if scheduler return the packet
960 	 *     just enqueued. This avoid a lock order reversal.
961 	 *
962 	 */
963 	if (/*V_dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) {
964 		/* fast io, rename the tag * to carry reinject info. */
965 		struct m_tag *tag = m_tag_first(m);
966 
967 		tag->m_tag_cookie = MTAG_IPFW_RULE;
968 		tag->m_tag_id = 0;
969 		V_dn_cfg.io_pkt_fast++;
970 		if (m->m_nextpkt != NULL) {
971 			printf("dummynet: fast io: pkt chain detected!\n");
972 			m->m_nextpkt = NULL;
973 		}
974 		m = NULL;
975 	} else {
976 		*m0 = NULL;
977 	}
978 done:
979 	DN_BH_WUNLOCK();
980 	if (m)
981 		dummynet_send(m);
982 	return 0;
983 
984 dropit:
985 	V_dn_cfg.io_pkt_drop++;
986 	DN_BH_WUNLOCK();
987 	if (m)
988 		FREE_PKT(m);
989 	*m0 = NULL;
990 	return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS;
991 }
992