xref: /freebsd/sys/netpfil/ipfw/ip_dn_io.c (revision 02923dd9b0de050b989b5a4103e402055cb7f61a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet6.h"
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/malloc.h>
38 #include <sys/mbuf.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/module.h>
42 #include <sys/mutex.h>
43 #include <sys/priv.h>
44 #include <sys/proc.h>
45 #include <sys/rwlock.h>
46 #include <sys/sdt.h>
47 #include <sys/socket.h>
48 #include <sys/time.h>
49 #include <sys/sysctl.h>
50 
51 #include <net/if.h>	/* IFNAMSIZ, struct ifaddr, ifq head, lock.h mutex.h */
52 #include <net/if_var.h>	/* NET_EPOCH_... */
53 #include <net/if_private.h>
54 #include <net/netisr.h>
55 #include <net/vnet.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/ip.h>		/* ip_len, ip_off */
59 #include <netinet/ip_var.h>	/* ip_output(), IP_FORWARDING */
60 #include <netinet/ip_fw.h>
61 #include <netinet/ip_dummynet.h>
62 #include <netinet/if_ether.h> /* various ether_* routines */
63 #include <netinet/ip6.h>       /* for ip6_input, ip6_output prototypes */
64 #include <netinet6/ip6_var.h>
65 
66 #include <netpfil/ipfw/ip_fw_private.h>
67 #include <netpfil/ipfw/dn_heap.h>
68 #include <netpfil/ipfw/ip_dn_private.h>
69 #ifdef NEW_AQM
70 #include <netpfil/ipfw/dn_aqm.h>
71 #endif
72 #include <netpfil/ipfw/dn_sched.h>
73 
74 SDT_PROVIDER_DEFINE(dummynet);
75 SDT_PROBE_DEFINE2(dummynet, , , drop, "struct mbuf *", "struct dn_queue *");
76 
77 /*
78  * We keep a private variable for the simulation time, but we could
79  * probably use an existing one ("softticks" in sys/kern/kern_timeout.c)
80  * instead of V_dn_cfg.curr_time
81  */
82 VNET_DEFINE(struct dn_parms, dn_cfg);
83 #define V_dn_cfg VNET(dn_cfg)
84 
85 /*
86  * We use a heap to store entities for which we have pending timer events.
87  * The heap is checked at every tick and all entities with expired events
88  * are extracted.
89  */
90 
91 MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap");
92 
93 extern	void (*bridge_dn_p)(struct mbuf *, struct ifnet *);
94 
95 #ifdef SYSCTL_NODE
96 
97 /*
98  * Because of the way the SYSBEGIN/SYSEND macros work on other
99  * platforms, there should not be functions between them.
100  * So keep the handlers outside the block.
101  */
102 static int
sysctl_hash_size(SYSCTL_HANDLER_ARGS)103 sysctl_hash_size(SYSCTL_HANDLER_ARGS)
104 {
105 	int error, value;
106 
107 	value = V_dn_cfg.hash_size;
108 	error = sysctl_handle_int(oidp, &value, 0, req);
109 	if (error != 0 || req->newptr == NULL)
110 		return (error);
111 	if (value < 16 || value > 65536)
112 		return (EINVAL);
113 	V_dn_cfg.hash_size = value;
114 	return (0);
115 }
116 
117 static int
sysctl_limits(SYSCTL_HANDLER_ARGS)118 sysctl_limits(SYSCTL_HANDLER_ARGS)
119 {
120 	int error;
121 	long value;
122 
123 	if (arg2 != 0)
124 		value = V_dn_cfg.slot_limit;
125 	else
126 		value = V_dn_cfg.byte_limit;
127 	error = sysctl_handle_long(oidp, &value, 0, req);
128 
129 	if (error != 0 || req->newptr == NULL)
130 		return (error);
131 	if (arg2 != 0) {
132 		if (value < 1)
133 			return (EINVAL);
134 		V_dn_cfg.slot_limit = value;
135 	} else {
136 		if (value < 1500)
137 			return (EINVAL);
138 		V_dn_cfg.byte_limit = value;
139 	}
140 	return (0);
141 }
142 
143 SYSBEGIN(f4)
144 
145 SYSCTL_DECL(_net_inet);
146 SYSCTL_DECL(_net_inet_ip);
147 #ifdef NEW_AQM
148 SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
149     "Dummynet");
150 #else
151 static SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet,
152     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
153     "Dummynet");
154 #endif
155 
156 /* wrapper to pass V_dn_cfg fields to SYSCTL_* */
157 #define DC(x)	(&(VNET_NAME(dn_cfg).x))
158 
159 /* parameters */
160 
161 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, hash_size,
162     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
163     0, 0, sysctl_hash_size, "I",
164     "Default hash table size");
165 
166 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_slot_limit,
167     CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
168     0, 1, sysctl_limits, "L",
169     "Upper limit in slots for pipe queue.");
170 SYSCTL_PROC(_net_inet_ip_dummynet, OID_AUTO, pipe_byte_limit,
171     CTLTYPE_LONG | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
172     0, 0, sysctl_limits, "L",
173     "Upper limit in bytes for pipe queue.");
174 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, io_fast,
175     CTLFLAG_RW | CTLFLAG_VNET, DC(io_fast), 0, "Enable fast dummynet io.");
176 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, debug,
177     CTLFLAG_RW | CTLFLAG_VNET, DC(debug), 0, "Dummynet debug level");
178 
179 /* RED parameters */
180 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth,
181     CTLFLAG_RD | CTLFLAG_VNET, DC(red_lookup_depth), 0, "Depth of RED lookup table");
182 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size,
183     CTLFLAG_RD | CTLFLAG_VNET, DC(red_avg_pkt_size), 0, "RED Medium packet size");
184 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size,
185     CTLFLAG_RD | CTLFLAG_VNET, DC(red_max_pkt_size), 0, "RED Max packet size");
186 
187 /* time adjustment */
188 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta,
189     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta), 0, "Last vs standard tick difference (usec).");
190 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_delta_sum,
191     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_delta_sum), 0, "Accumulated tick difference (usec).");
192 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_adjustment,
193     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_adjustment), 0, "Tick adjustments done.");
194 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_diff,
195     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_diff), 0,
196     "Adjusted vs non-adjusted curr_time difference (ticks).");
197 SYSCTL_LONG(_net_inet_ip_dummynet, OID_AUTO, tick_lost,
198     CTLFLAG_RD | CTLFLAG_VNET, DC(tick_lost), 0,
199     "Number of ticks coalesced by dummynet taskqueue.");
200 
201 /* Drain parameters */
202 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire,
203     CTLFLAG_RW | CTLFLAG_VNET, DC(expire), 0, "Expire empty queues/pipes");
204 SYSCTL_UINT(_net_inet_ip_dummynet, OID_AUTO, expire_cycle,
205     CTLFLAG_RD | CTLFLAG_VNET, DC(expire_cycle), 0, "Expire cycle for queues/pipes");
206 
207 /* statistics */
208 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, schk_count,
209     CTLFLAG_RD | CTLFLAG_VNET, DC(schk_count), 0, "Number of schedulers");
210 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, si_count,
211     CTLFLAG_RD | CTLFLAG_VNET, DC(si_count), 0, "Number of scheduler instances");
212 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, fsk_count,
213     CTLFLAG_RD | CTLFLAG_VNET, DC(fsk_count), 0, "Number of flowsets");
214 SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, queue_count,
215     CTLFLAG_RD | CTLFLAG_VNET, DC(queue_count), 0, "Number of queues");
216 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt,
217     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt), 0,
218     "Number of packets passed to dummynet.");
219 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_fast,
220     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_fast), 0,
221     "Number of packets bypassed dummynet scheduler.");
222 SYSCTL_ULONG(_net_inet_ip_dummynet, OID_AUTO, io_pkt_drop,
223     CTLFLAG_RD | CTLFLAG_VNET, DC(io_pkt_drop), 0,
224     "Number of packets dropped by dummynet.");
225 #undef DC
226 SYSEND
227 
228 #endif
229 
230 static void	dummynet_send(struct mbuf *);
231 
232 /*
233  * Return the mbuf tag holding the dummynet state (it should
234  * be the first one on the list).
235  */
236 struct dn_pkt_tag *
dn_tag_get(struct mbuf * m)237 dn_tag_get(struct mbuf *m)
238 {
239 	struct m_tag *mtag = m_tag_first(m);
240 #ifdef NEW_AQM
241 	/* XXX: to skip ts m_tag. For Debugging only*/
242 	if (mtag != NULL && mtag->m_tag_id == DN_AQM_MTAG_TS) {
243 		m_tag_delete(m,mtag);
244 		mtag = m_tag_first(m);
245 		D("skip TS tag");
246 	}
247 #endif
248 	KASSERT(mtag != NULL &&
249 	    mtag->m_tag_cookie == MTAG_ABI_COMPAT &&
250 	    mtag->m_tag_id == PACKET_TAG_DUMMYNET,
251 	    ("packet on dummynet queue w/o dummynet tag!"));
252 	return (struct dn_pkt_tag *)(mtag+1);
253 }
254 
255 #ifndef NEW_AQM
256 static inline void
mq_append(struct mq * q,struct mbuf * m)257 mq_append(struct mq *q, struct mbuf *m)
258 {
259 #ifdef USERSPACE
260 	// buffers from netmap need to be copied
261 	// XXX note that the routine is not expected to fail
262 	ND("append %p to %p", m, q);
263 	if (m->m_flags & M_STACK) {
264 		struct mbuf *m_new;
265 		void *p;
266 		int l, ofs;
267 
268 		ofs = m->m_data - m->__m_extbuf;
269 		// XXX allocate
270 		MGETHDR(m_new, M_NOWAIT, MT_DATA);
271 		ND("*** WARNING, volatile buf %p ext %p %d dofs %d m_new %p",
272 			m, m->__m_extbuf, m->__m_extlen, ofs, m_new);
273 		p = m_new->__m_extbuf;	/* new pointer */
274 		l = m_new->__m_extlen;	/* new len */
275 		if (l <= m->__m_extlen) {
276 			panic("extlen too large");
277 		}
278 
279 		*m_new = *m;	// copy
280 		m_new->m_flags &= ~M_STACK;
281 		m_new->__m_extbuf = p; // point to new buffer
282 		_pkt_copy(m->__m_extbuf, p, m->__m_extlen);
283 		m_new->m_data = p + ofs;
284 		m = m_new;
285 	}
286 #endif /* USERSPACE */
287 	if (q->head == NULL)
288 		q->head = m;
289 	else
290 		q->tail->m_nextpkt = m;
291 	q->count++;
292 	q->tail = m;
293 	m->m_nextpkt = NULL;
294 }
295 #endif
296 
297 /*
298  * Dispose a list of packet. Use a functions so if we need to do
299  * more work, this is a central point to do it.
300  */
dn_free_pkts(struct mbuf * mnext)301 void dn_free_pkts(struct mbuf *mnext)
302 {
303         struct mbuf *m;
304 
305         while ((m = mnext) != NULL) {
306                 mnext = m->m_nextpkt;
307                 FREE_PKT(m);
308         }
309 }
310 
311 static int
red_drops(struct dn_queue * q,int len)312 red_drops (struct dn_queue *q, int len)
313 {
314 	/*
315 	 * RED algorithm
316 	 *
317 	 * RED calculates the average queue size (avg) using a low-pass filter
318 	 * with an exponential weighted (w_q) moving average:
319 	 * 	avg  <-  (1-w_q) * avg + w_q * q_size
320 	 * where q_size is the queue length (measured in bytes or * packets).
321 	 *
322 	 * If q_size == 0, we compute the idle time for the link, and set
323 	 *	avg = (1 - w_q)^(idle/s)
324 	 * where s is the time needed for transmitting a medium-sized packet.
325 	 *
326 	 * Now, if avg < min_th the packet is enqueued.
327 	 * If avg > max_th the packet is dropped. Otherwise, the packet is
328 	 * dropped with probability P function of avg.
329 	 */
330 
331 	struct dn_fsk *fs = q->fs;
332 	int64_t p_b = 0;
333 
334 	/* Queue in bytes or packets? */
335 	uint32_t q_size = (fs->fs.flags & DN_QSIZE_BYTES) ?
336 	    q->ni.len_bytes : q->ni.length;
337 
338 	/* Average queue size estimation. */
339 	if (q_size != 0) {
340 		/* Queue is not empty, avg <- avg + (q_size - avg) * w_q */
341 		int diff = SCALE(q_size) - q->avg;
342 		int64_t v = SCALE_MUL((int64_t)diff, (int64_t)fs->w_q);
343 
344 		q->avg += (int)v;
345 	} else {
346 		/*
347 		 * Queue is empty, find for how long the queue has been
348 		 * empty and use a lookup table for computing
349 		 * (1 - * w_q)^(idle_time/s) where s is the time to send a
350 		 * (small) packet.
351 		 * XXX check wraps...
352 		 */
353 		if (q->avg) {
354 			u_int t = div64((V_dn_cfg.curr_time - q->q_time), fs->lookup_step);
355 
356 			q->avg = (t < fs->lookup_depth) ?
357 			    SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0;
358 		}
359 	}
360 
361 	/* Should i drop? */
362 	if (q->avg < fs->min_th) {
363 		q->count = -1;
364 		return (0);	/* accept packet */
365 	}
366 	if (q->avg >= fs->max_th) {	/* average queue >=  max threshold */
367 		if (fs->fs.flags & DN_IS_ECN)
368 			return (1);
369 		if (fs->fs.flags & DN_IS_GENTLE_RED) {
370 			/*
371 			 * According to Gentle-RED, if avg is greater than
372 			 * max_th the packet is dropped with a probability
373 			 *	 p_b = c_3 * avg - c_4
374 			 * where c_3 = (1 - max_p) / max_th
375 			 *       c_4 = 1 - 2 * max_p
376 			 */
377 			p_b = SCALE_MUL((int64_t)fs->c_3, (int64_t)q->avg) -
378 			    fs->c_4;
379 		} else {
380 			q->count = -1;
381 			return (1);
382 		}
383 	} else if (q->avg > fs->min_th) {
384 		if (fs->fs.flags & DN_IS_ECN)
385 			return (1);
386 		/*
387 		 * We compute p_b using the linear dropping function
388 		 *	 p_b = c_1 * avg - c_2
389 		 * where c_1 = max_p / (max_th - min_th)
390 		 * 	 c_2 = max_p * min_th / (max_th - min_th)
391 		 */
392 		p_b = SCALE_MUL((int64_t)fs->c_1, (int64_t)q->avg) - fs->c_2;
393 	}
394 
395 	if (fs->fs.flags & DN_QSIZE_BYTES)
396 		p_b = div64((p_b * len) , fs->max_pkt_size);
397 	if (++q->count == 0)
398 		q->random = random() & 0xffff;
399 	else {
400 		/*
401 		 * q->count counts packets arrived since last drop, so a greater
402 		 * value of q->count means a greater packet drop probability.
403 		 */
404 		if (SCALE_MUL(p_b, SCALE((int64_t)q->count)) > q->random) {
405 			q->count = 0;
406 			/* After a drop we calculate a new random value. */
407 			q->random = random() & 0xffff;
408 			return (1);	/* drop */
409 		}
410 	}
411 	/* End of RED algorithm. */
412 
413 	return (0);	/* accept */
414 
415 }
416 
417 /*
418  * ECN/ECT Processing (partially adopted from altq)
419  */
420 #ifndef NEW_AQM
421 static
422 #endif
423 int
ecn_mark(struct mbuf * m)424 ecn_mark(struct mbuf* m)
425 {
426 	struct ip *ip;
427 	ip = (struct ip *)mtodo(m, dn_tag_get(m)->iphdr_off);
428 
429 	switch (ip->ip_v) {
430 	case IPVERSION:
431 	{
432 		uint16_t old;
433 
434 		if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_NOTECT)
435 			return (0);	/* not-ECT */
436 		if ((ip->ip_tos & IPTOS_ECN_MASK) == IPTOS_ECN_CE)
437 			return (1);	/* already marked */
438 
439 		/*
440 		 * ecn-capable but not marked,
441 		 * mark CE and update checksum
442 		 */
443 		old = *(uint16_t *)ip;
444 		ip->ip_tos |= IPTOS_ECN_CE;
445 		ip->ip_sum = cksum_adjust(ip->ip_sum, old, *(uint16_t *)ip);
446 		return (1);
447 	}
448 #ifdef INET6
449 	case (IPV6_VERSION >> 4):
450 	{
451 		struct ip6_hdr *ip6 = (struct ip6_hdr *)ip;
452 		u_int32_t flowlabel;
453 
454 		flowlabel = ntohl(ip6->ip6_flow);
455 		if ((flowlabel >> 28) != 6)
456 			return (0);	/* version mismatch! */
457 		if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
458 		    (IPTOS_ECN_NOTECT << 20))
459 			return (0);	/* not-ECT */
460 		if ((flowlabel & (IPTOS_ECN_MASK << 20)) ==
461 		    (IPTOS_ECN_CE << 20))
462 			return (1);	/* already marked */
463 		/*
464 		 * ecn-capable but not marked, mark CE
465 		 */
466 		flowlabel |= (IPTOS_ECN_CE << 20);
467 		ip6->ip6_flow = htonl(flowlabel);
468 		return (1);
469 	}
470 #endif
471 	}
472 	return (0);
473 }
474 
475 /*
476  * Enqueue a packet in q, subject to space and queue management policy
477  * (whose parameters are in q->fs).
478  * Update stats for the queue and the scheduler.
479  * Return 0 on success, 1 on drop. The packet is consumed anyways.
480  */
481 int
dn_enqueue(struct dn_queue * q,struct mbuf * m,int drop)482 dn_enqueue(struct dn_queue *q, struct mbuf* m, int drop)
483 {
484 	struct dn_fs *f;
485 	struct dn_flow *ni;	/* stats for scheduler instance */
486 	uint64_t len;
487 
488 	if (q->fs == NULL || q->_si == NULL) {
489 		printf("%s fs %p si %p, dropping\n",
490 			__FUNCTION__, q->fs, q->_si);
491 		FREE_PKT(m);
492 		return 1;
493 	}
494 	f = &(q->fs->fs);
495 	ni = &q->_si->ni;
496 	len = m->m_pkthdr.len;
497 	/* Update statistics, then check reasons to drop pkt. */
498 	q->ni.tot_bytes += len;
499 	q->ni.tot_pkts++;
500 	ni->tot_bytes += len;
501 	ni->tot_pkts++;
502 	if (drop)
503 		goto drop;
504 	if (f->plr[0] || f->plr[1]) {
505 		if (__predict_true(f->plr[1] == 0)) {
506 			if (random() < f->plr[0])
507 				goto drop;
508 		} else {
509 			switch (f->pl_state) {
510 			case PLR_STATE_B:
511 				if (random() < f->plr[3])
512 					f->pl_state = PLR_STATE_G;
513 				if (random() < f->plr[2])
514 					goto drop;
515 				break;
516 			case PLR_STATE_G: /* FALLTHROUGH */
517 			default:
518 				if (random() < f->plr[1])
519 					f->pl_state = PLR_STATE_B;
520 				if (random() < f->plr[0])
521 					goto drop;
522 				break;
523 			}
524 		}
525 	}
526 	if (m->m_pkthdr.rcvif != NULL)
527 		m_rcvif_serialize(m);
528 #ifdef NEW_AQM
529 	/* Call AQM enqueue function */
530 	if (q->fs->aqmfp)
531 		return q->fs->aqmfp->enqueue(q ,m);
532 #endif
533 	if (f->flags & DN_IS_RED && red_drops(q, m->m_pkthdr.len)) {
534 		if (!(f->flags & DN_IS_ECN) || !ecn_mark(m))
535 			goto drop;
536 	}
537 	if (f->flags & DN_QSIZE_BYTES) {
538 		if (q->ni.len_bytes > f->qsize)
539 			goto drop;
540 	} else if (q->ni.length >= f->qsize) {
541 		goto drop;
542 	}
543 	mq_append(&q->mq, m);
544 	q->ni.length++;
545 	q->ni.len_bytes += len;
546 	ni->length++;
547 	ni->len_bytes += len;
548 	return (0);
549 
550 drop:
551 	V_dn_cfg.io_pkt_drop++;
552 	SDT_PROBE2(dummynet, , , drop, m, q);
553 	q->ni.drops++;
554 	ni->drops++;
555 	FREE_PKT(m);
556 	return (1);
557 }
558 
559 /*
560  * Fetch packets from the delay line which are due now. If there are
561  * leftover packets, reinsert the delay line in the heap.
562  * Runs under scheduler lock.
563  */
564 static void
transmit_event(struct mq * q,struct delay_line * dline,uint64_t now)565 transmit_event(struct mq *q, struct delay_line *dline, uint64_t now)
566 {
567 	struct mbuf *m;
568 	struct dn_pkt_tag *pkt = NULL;
569 
570 	dline->oid.subtype = 0; /* not in heap */
571 	while ((m = dline->mq.head) != NULL) {
572 		pkt = dn_tag_get(m);
573 		if (!DN_KEY_LEQ(pkt->output_time, now))
574 			break;
575 		dline->mq.head = m->m_nextpkt;
576 		dline->mq.count--;
577 		if (m->m_pkthdr.rcvif != NULL &&
578 		  __predict_false(m_rcvif_restore(m) == NULL))
579 			m_freem(m);
580 		else
581 			mq_append(q, m);
582 	}
583 	if (m != NULL) {
584 		dline->oid.subtype = 1; /* in heap */
585 		heap_insert(&V_dn_cfg.evheap, pkt->output_time, dline);
586 	}
587 }
588 
589 /*
590  * Convert the additional MAC overheads/delays into an equivalent
591  * number of bits for the given data rate. The samples are
592  * in milliseconds so we need to divide by 1000.
593  */
594 static uint64_t
extra_bits(struct mbuf * m,struct dn_schk * s)595 extra_bits(struct mbuf *m, struct dn_schk *s)
596 {
597 	int index;
598 	uint64_t bits;
599 	struct dn_profile *pf = s->profile;
600 
601 	if (!pf || pf->samples_no == 0)
602 		return 0;
603 	index  = random() % pf->samples_no;
604 	bits = div64((uint64_t)pf->samples[index] * s->link.bandwidth, 1000);
605 	if (index >= pf->loss_level) {
606 		struct dn_pkt_tag *dt = dn_tag_get(m);
607 		if (dt)
608 			dt->dn_dir = DIR_DROP;
609 	}
610 	return bits;
611 }
612 
613 /*
614  * Send traffic from a scheduler instance due by 'now'.
615  * Return a pointer to the head of the queue.
616  */
617 static struct mbuf *
serve_sched(struct mq * q,struct dn_sch_inst * si,uint64_t now)618 serve_sched(struct mq *q, struct dn_sch_inst *si, uint64_t now)
619 {
620 	struct mq def_q;
621 	struct dn_schk *s = si->sched;
622 	struct mbuf *m = NULL;
623 	int delay_line_idle = (si->dline.mq.head == NULL);
624 	int done;
625 	uint32_t bw;
626 
627 	if (q == NULL) {
628 		q = &def_q;
629 		q->head = NULL;
630 	}
631 
632 	bw = s->link.bandwidth;
633 	si->kflags &= ~DN_ACTIVE;
634 
635 	if (bw > 0)
636 		si->credit += (now - si->sched_time) * bw;
637 	else
638 		si->credit = 0;
639 	si->sched_time = now;
640 	done = 0;
641 	while (si->credit >= 0 && (m = s->fp->dequeue(si)) != NULL) {
642 		uint64_t len_scaled;
643 
644 		done++;
645 		len_scaled = (bw == 0) ? 0 : hz *
646 			(m->m_pkthdr.len * 8 + extra_bits(m, s));
647 		si->credit -= len_scaled;
648 		/* Move packet in the delay line */
649 		dn_tag_get(m)->output_time = V_dn_cfg.curr_time + s->link.delay ;
650 		if (m->m_pkthdr.rcvif != NULL)
651 			m_rcvif_serialize(m);
652 		mq_append(&si->dline.mq, m);
653 	}
654 
655 	/*
656 	 * If credit >= 0 the instance is idle, mark time.
657 	 * Otherwise put back in the heap, and adjust the output
658 	 * time of the last inserted packet, m, which was too early.
659 	 */
660 	if (si->credit >= 0) {
661 		si->idle_time = now;
662 	} else {
663 		uint64_t t;
664 		KASSERT (bw > 0, ("bw=0 and credit<0 ?"));
665 		t = div64(bw - 1 - si->credit, bw);
666 		if (m)
667 			dn_tag_get(m)->output_time += t;
668 		si->kflags |= DN_ACTIVE;
669 		heap_insert(&V_dn_cfg.evheap, now + t, si);
670 	}
671 	if (delay_line_idle && done)
672 		transmit_event(q, &si->dline, now);
673 	return q->head;
674 }
675 
676 /*
677  * The timer handler for dummynet. Time is computed in ticks, but
678  * but the code is tolerant to the actual rate at which this is called.
679  * Once complete, the function reschedules itself for the next tick.
680  */
681 void
dummynet_task(void * context,int pending)682 dummynet_task(void *context, int pending)
683 {
684 	struct timeval t;
685 	struct mq q = { NULL, NULL }; /* queue to accumulate results */
686 	struct epoch_tracker et;
687 
688 	VNET_ITERATOR_DECL(vnet_iter);
689 	VNET_LIST_RLOCK();
690 	NET_EPOCH_ENTER(et);
691 
692 	VNET_FOREACH(vnet_iter) {
693 		memset(&q, 0, sizeof(struct mq));
694 		CURVNET_SET(vnet_iter);
695 
696 		if (! V_dn_cfg.init_done) {
697 			CURVNET_RESTORE();
698 			continue;
699 		}
700 
701 		DN_BH_WLOCK();
702 
703 		/* Update number of lost(coalesced) ticks. */
704 		V_dn_cfg.tick_lost += pending - 1;
705 
706 		getmicrouptime(&t);
707 		/* Last tick duration (usec). */
708 		V_dn_cfg.tick_last = (t.tv_sec - V_dn_cfg.prev_t.tv_sec) * 1000000 +
709 		(t.tv_usec - V_dn_cfg.prev_t.tv_usec);
710 		/* Last tick vs standard tick difference (usec). */
711 		V_dn_cfg.tick_delta = (V_dn_cfg.tick_last * hz - 1000000) / hz;
712 		/* Accumulated tick difference (usec). */
713 		V_dn_cfg.tick_delta_sum += V_dn_cfg.tick_delta;
714 
715 		V_dn_cfg.prev_t = t;
716 
717 		/*
718 		* Adjust curr_time if the accumulated tick difference is
719 		* greater than the 'standard' tick. Since curr_time should
720 		* be monotonically increasing, we do positive adjustments
721 		* as required, and throttle curr_time in case of negative
722 		* adjustment.
723 		*/
724 		V_dn_cfg.curr_time++;
725 		if (V_dn_cfg.tick_delta_sum - tick >= 0) {
726 			int diff = V_dn_cfg.tick_delta_sum / tick;
727 
728 			V_dn_cfg.curr_time += diff;
729 			V_dn_cfg.tick_diff += diff;
730 			V_dn_cfg.tick_delta_sum %= tick;
731 			V_dn_cfg.tick_adjustment++;
732 		} else if (V_dn_cfg.tick_delta_sum + tick <= 0) {
733 			V_dn_cfg.curr_time--;
734 			V_dn_cfg.tick_diff--;
735 			V_dn_cfg.tick_delta_sum += tick;
736 			V_dn_cfg.tick_adjustment++;
737 		}
738 
739 		/* serve pending events, accumulate in q */
740 		for (;;) {
741 			struct dn_id *p;    /* generic parameter to handler */
742 
743 			if (V_dn_cfg.evheap.elements == 0 ||
744 			    DN_KEY_LT(V_dn_cfg.curr_time, HEAP_TOP(&V_dn_cfg.evheap)->key))
745 				break;
746 			p = HEAP_TOP(&V_dn_cfg.evheap)->object;
747 			heap_extract(&V_dn_cfg.evheap, NULL);
748 			if (p->type == DN_SCH_I) {
749 				serve_sched(&q, (struct dn_sch_inst *)p, V_dn_cfg.curr_time);
750 			} else { /* extracted a delay line */
751 				transmit_event(&q, (struct delay_line *)p, V_dn_cfg.curr_time);
752 			}
753 		}
754 		if (V_dn_cfg.expire && ++V_dn_cfg.expire_cycle >= V_dn_cfg.expire) {
755 			V_dn_cfg.expire_cycle = 0;
756 			dn_drain_scheduler();
757 			dn_drain_queue();
758 		}
759 		DN_BH_WUNLOCK();
760 		if (q.head != NULL)
761 			dummynet_send(q.head);
762 
763 		CURVNET_RESTORE();
764 	}
765 	NET_EPOCH_EXIT(et);
766 	VNET_LIST_RUNLOCK();
767 
768 	/* Schedule our next run. */
769 	dn_reschedule();
770 }
771 
772 /*
773  * forward a chain of packets to the proper destination.
774  * This runs outside the dummynet lock.
775  */
776 static void
dummynet_send(struct mbuf * m)777 dummynet_send(struct mbuf *m)
778 {
779 	struct mbuf *n;
780 
781 	NET_EPOCH_ASSERT();
782 
783 	for (; m != NULL; m = n) {
784 		struct ifnet *ifp = NULL;	/* gcc 3.4.6 complains */
785         	struct m_tag *tag;
786 		int dst;
787 
788 		n = m->m_nextpkt;
789 		m->m_nextpkt = NULL;
790 		tag = m_tag_first(m);
791 		if (tag == NULL) { /* should not happen */
792 			dst = DIR_DROP;
793 		} else {
794 			struct dn_pkt_tag *pkt = dn_tag_get(m);
795 			/* extract the dummynet info, rename the tag
796 			 * to carry reinject info.
797 			 */
798 			ifp = ifnet_byindexgen(pkt->if_index, pkt->if_idxgen);
799 			if (((pkt->dn_dir == (DIR_OUT | PROTO_LAYER2)) ||
800 			    (pkt->dn_dir == (DIR_OUT | PROTO_LAYER2 | PROTO_IPV6))) &&
801 				ifp == NULL) {
802 				dst = DIR_DROP;
803 			} else {
804 				dst = pkt->dn_dir;
805 				tag->m_tag_cookie = MTAG_IPFW_RULE;
806 				tag->m_tag_id = 0;
807 			}
808 		}
809 
810 		switch (dst) {
811 		case DIR_OUT:
812 			ip_output(m, NULL, NULL, IP_FORWARDING, NULL, NULL);
813 			break ;
814 
815 		case DIR_IN :
816 			netisr_dispatch(NETISR_IP, m);
817 			break;
818 
819 #ifdef INET6
820 		case DIR_IN | PROTO_IPV6:
821 			netisr_dispatch(NETISR_IPV6, m);
822 			break;
823 
824 		case DIR_OUT | PROTO_IPV6:
825 			ip6_output(m, NULL, NULL, IPV6_FORWARDING, NULL, NULL, NULL);
826 			break;
827 #endif
828 
829 		case DIR_FWD | PROTO_IFB: /* DN_TO_IFB_FWD: */
830 			if (bridge_dn_p != NULL)
831 				((*bridge_dn_p)(m, ifp));
832 			else
833 				printf("dummynet: if_bridge not loaded\n");
834 
835 			break;
836 
837 		case DIR_IN | PROTO_LAYER2 | PROTO_IPV6:
838 		case DIR_IN | PROTO_LAYER2: /* DN_TO_ETH_DEMUX: */
839 			/*
840 			 * The Ethernet code assumes the Ethernet header is
841 			 * contiguous in the first mbuf header.
842 			 * Insure this is true.
843 			 */
844 			if (m->m_len < ETHER_HDR_LEN &&
845 			    (m = m_pullup(m, ETHER_HDR_LEN)) == NULL) {
846 				printf("dummynet/ether: pullup failed, "
847 				    "dropping packet\n");
848 				break;
849 			}
850 			ether_demux(m->m_pkthdr.rcvif, m);
851 			break;
852 
853 		case DIR_OUT | PROTO_LAYER2 | PROTO_IPV6:
854 		case DIR_OUT | PROTO_LAYER2: /* DN_TO_ETH_OUT: */
855 			MPASS(ifp != NULL);
856 			ether_output_frame(ifp, m);
857 			break;
858 
859 		case DIR_DROP:
860 			/* drop the packet after some time */
861 			FREE_PKT(m);
862 			break;
863 
864 		default:
865 			printf("dummynet: bad switch %d!\n", dst);
866 			FREE_PKT(m);
867 			break;
868 		}
869 	}
870 }
871 
872 static inline int
tag_mbuf(struct mbuf * m,int dir,struct ip_fw_args * fwa)873 tag_mbuf(struct mbuf *m, int dir, struct ip_fw_args *fwa)
874 {
875 	struct dn_pkt_tag *dt;
876 	struct m_tag *mtag;
877 
878 	mtag = m_tag_get(PACKET_TAG_DUMMYNET,
879 		    sizeof(*dt), M_NOWAIT | M_ZERO);
880 	if (mtag == NULL)
881 		return 1;		/* Cannot allocate packet header. */
882 	m_tag_prepend(m, mtag);		/* Attach to mbuf chain. */
883 	dt = (struct dn_pkt_tag *)(mtag + 1);
884 	dt->rule = fwa->rule;
885 	/* only keep this info */
886 	dt->rule.info &= (IPFW_ONEPASS | IPFW_IS_DUMMYNET);
887 	dt->dn_dir = dir;
888 	if (fwa->flags & IPFW_ARGS_OUT && fwa->ifp != NULL) {
889 		NET_EPOCH_ASSERT();
890 		dt->if_index = fwa->ifp->if_index;
891 		dt->if_idxgen = fwa->ifp->if_idxgen;
892 	}
893 	/* dt->output_time is updated as we move through */
894 	dt->output_time = V_dn_cfg.curr_time;
895 	dt->iphdr_off = (dir & PROTO_LAYER2) ? ETHER_HDR_LEN : 0;
896 	return 0;
897 }
898 
899 /*
900  * dummynet hook for packets.
901  * We use the argument to locate the flowset fs and the sched_set sch
902  * associated to it. The we apply flow_mask and sched_mask to
903  * determine the queue and scheduler instances.
904  */
905 int
dummynet_io(struct mbuf ** m0,struct ip_fw_args * fwa)906 dummynet_io(struct mbuf **m0, struct ip_fw_args *fwa)
907 {
908 	struct mbuf *m = *m0;
909 	struct dn_fsk *fs = NULL;
910 	struct dn_sch_inst *si;
911 	struct dn_queue *q = NULL;	/* default */
912 	int fs_id, dir;
913 
914 	fs_id = (fwa->rule.info & IPFW_INFO_MASK) +
915 		((fwa->rule.info & IPFW_IS_PIPE) ? 2*DN_MAX_ID : 0);
916 	/* XXXGL: convert args to dir */
917 	if (fwa->flags & IPFW_ARGS_IN)
918 		dir = DIR_IN;
919 	else
920 		dir = DIR_OUT;
921 	if (fwa->flags & IPFW_ARGS_ETHER)
922 		dir |= PROTO_LAYER2;
923 	else if (fwa->flags & IPFW_ARGS_IP6)
924 		dir |= PROTO_IPV6;
925 	DN_BH_WLOCK();
926 	V_dn_cfg.io_pkt++;
927 	/* we could actually tag outside the lock, but who cares... */
928 	if (tag_mbuf(m, dir, fwa))
929 		goto dropit;
930 	/* XXX locate_flowset could be optimised with a direct ref. */
931 	fs = dn_ht_find(V_dn_cfg.fshash, fs_id, 0, NULL);
932 	if (fs == NULL)
933 		goto dropit;	/* This queue/pipe does not exist! */
934 	if (fs->sched == NULL)	/* should not happen */
935 		goto dropit;
936 	/* find scheduler instance, possibly applying sched_mask */
937 	si = ipdn_si_find(fs->sched, &(fwa->f_id));
938 	if (si == NULL)
939 		goto dropit;
940 	/*
941 	 * If the scheduler supports multiple queues, find the right one
942 	 * (otherwise it will be ignored by enqueue).
943 	 */
944 	if (fs->sched->fp->flags & DN_MULTIQUEUE) {
945 		q = ipdn_q_find(fs, si, &(fwa->f_id));
946 		if (q == NULL)
947 			goto dropit;
948 	}
949 	if (fs->sched->fp->enqueue(si, q, m)) {
950 		/* packet was dropped by enqueue() */
951 		m = *m0 = NULL;
952 
953 		/* dn_enqueue already increases io_pkt_drop */
954 		V_dn_cfg.io_pkt_drop--;
955 
956 		goto dropit;
957 	}
958 
959 	if (si->kflags & DN_ACTIVE) {
960 		m = *m0 = NULL; /* consumed */
961 		goto done; /* already active, nothing to do */
962 	}
963 
964 	/* compute the initial allowance */
965 	if (si->idle_time < V_dn_cfg.curr_time) {
966 	    /* Do this only on the first packet on an idle pipe */
967 	    struct dn_link *p = &fs->sched->link;
968 
969 	    si->sched_time = V_dn_cfg.curr_time;
970 	    si->credit = V_dn_cfg.io_fast ? p->bandwidth : 0;
971 	    if (p->burst) {
972 		uint64_t burst = (V_dn_cfg.curr_time - si->idle_time) * p->bandwidth;
973 		if (burst > p->burst)
974 			burst = p->burst;
975 		si->credit += burst;
976 	    }
977 	}
978 	/* pass through scheduler and delay line */
979 	m = serve_sched(NULL, si, V_dn_cfg.curr_time);
980 
981 	/* optimization -- pass it back to ipfw for immediate send */
982 	/* XXX Don't call dummynet_send() if scheduler return the packet
983 	 *     just enqueued. This avoid a lock order reversal.
984 	 *
985 	 */
986 	if (/*V_dn_cfg.io_fast &&*/ m == *m0 && (dir & PROTO_LAYER2) == 0 ) {
987 		/* fast io, rename the tag * to carry reinject info. */
988 		struct m_tag *tag = m_tag_first(m);
989 
990 		tag->m_tag_cookie = MTAG_IPFW_RULE;
991 		tag->m_tag_id = 0;
992 		V_dn_cfg.io_pkt_fast++;
993 		if (m->m_nextpkt != NULL) {
994 			printf("dummynet: fast io: pkt chain detected!\n");
995 			m->m_nextpkt = NULL;
996 		}
997 		m = NULL;
998 	} else {
999 		*m0 = NULL;
1000 	}
1001 done:
1002 	DN_BH_WUNLOCK();
1003 	if (m)
1004 		dummynet_send(m);
1005 	return 0;
1006 
1007 dropit:
1008 	V_dn_cfg.io_pkt_drop++;
1009 	SDT_PROBE2(dummynet, , , drop, m, q);
1010 	DN_BH_WUNLOCK();
1011 	if (m)
1012 		FREE_PKT(m);
1013 	*m0 = NULL;
1014 	return (fs && (fs->fs.flags & DN_NOERROR)) ? 0 : ENOBUFS;
1015 }
1016