xref: /freebsd/sys/net/altq/altq_fairq.c (revision b2d48be1bc7df45ddd13b143a160d0acb5a383c5)
1 /*
2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
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  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * $DragonFly: src/sys/net/altq/altq_fairq.c,v 1.1 2008/04/06 18:58:15 dillon Exp $
35  * $FreeBSD$
36  */
37 /*
38  * Matt: I gutted altq_priq.c and used it as a skeleton on which to build
39  * fairq.  The fairq algorithm is completely different then priq, of course,
40  * but because I used priq's skeleton I believe I should include priq's
41  * copyright.
42  *
43  * Copyright (C) 2000-2003
44  *	Sony Computer Science Laboratories Inc.  All rights reserved.
45  *
46  * Redistribution and use in source and binary forms, with or without
47  * modification, are permitted provided that the following conditions
48  * are met:
49  * 1. Redistributions of source code must retain the above copyright
50  *    notice, this list of conditions and the following disclaimer.
51  * 2. Redistributions in binary form must reproduce the above copyright
52  *    notice, this list of conditions and the following disclaimer in the
53  *    documentation and/or other materials provided with the distribution.
54  *
55  * THIS SOFTWARE IS PROVIDED BY SONY CSL AND CONTRIBUTORS ``AS IS'' AND
56  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
57  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
58  * ARE DISCLAIMED.  IN NO EVENT SHALL SONY CSL OR CONTRIBUTORS BE LIABLE
59  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
60  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
61  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
62  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
63  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
64  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
65  * SUCH DAMAGE.
66  */
67 
68 /*
69  * FAIRQ - take traffic classified by keep state (hashed into
70  * mbuf->m_pkthdr.altq_state_hash) and bucketize it.  Fairly extract
71  * the first packet from each bucket in a round-robin fashion.
72  *
73  * TODO - better overall qlimit support (right now it is per-bucket).
74  *	- NOTE: red etc is per bucket, not overall.
75  *	- better service curve support.
76  *
77  * EXAMPLE:
78  *
79  *  altq on em0 fairq bandwidth 650Kb queue { std, bulk }
80  *  queue std  priority 3 bandwidth 400Kb \
81  *	fairq (buckets 64, default, hogs 1Kb) qlimit 50
82  *  queue bulk priority 2 bandwidth 100Kb \
83  *	fairq (buckets 64, hogs 1Kb) qlimit 50
84  *
85  *  pass out on em0 from any to any keep state queue std
86  *  pass out on em0 inet proto tcp ..... port ... keep state queue bulk
87  */
88 #include "opt_altq.h"
89 #include "opt_inet.h"
90 #include "opt_inet6.h"
91 
92 #ifdef ALTQ_FAIRQ  /* fairq is enabled in the kernel conf */
93 
94 #include <sys/param.h>
95 #include <sys/malloc.h>
96 #include <sys/mbuf.h>
97 #include <sys/socket.h>
98 #include <sys/sockio.h>
99 #include <sys/systm.h>
100 #include <sys/proc.h>
101 #include <sys/errno.h>
102 #include <sys/kernel.h>
103 #include <sys/queue.h>
104 
105 #include <net/if.h>
106 #include <net/if_var.h>
107 #include <netinet/in.h>
108 
109 #include <netpfil/pf/pf.h>
110 #include <netpfil/pf/pf_altq.h>
111 #include <netpfil/pf/pf_mtag.h>
112 #include <net/altq/altq.h>
113 #include <net/altq/altq_fairq.h>
114 
115 /*
116  * function prototypes
117  */
118 static int	fairq_clear_interface(struct fairq_if *);
119 static int	fairq_request(struct ifaltq *, int, void *);
120 static void	fairq_purge(struct fairq_if *);
121 static struct fairq_class *fairq_class_create(struct fairq_if *, int, int, u_int, struct fairq_opts *, int);
122 static int	fairq_class_destroy(struct fairq_class *);
123 static int	fairq_enqueue(struct ifaltq *, struct mbuf *, struct altq_pktattr *);
124 static struct mbuf *fairq_dequeue(struct ifaltq *, int);
125 
126 static int	fairq_addq(struct fairq_class *, struct mbuf *, u_int32_t);
127 static struct mbuf *fairq_getq(struct fairq_class *, uint64_t);
128 static struct mbuf *fairq_pollq(struct fairq_class *, uint64_t, int *);
129 static fairq_bucket_t *fairq_selectq(struct fairq_class *, int);
130 static void	fairq_purgeq(struct fairq_class *);
131 
132 static void	get_class_stats(struct fairq_classstats *, struct fairq_class *);
133 static struct fairq_class *clh_to_clp(struct fairq_if *, uint32_t);
134 
135 int
136 fairq_pfattach(struct pf_altq *a)
137 {
138 	struct ifnet *ifp;
139 	int error;
140 
141 	if ((ifp = ifunit(a->ifname)) == NULL || a->altq_disc == NULL)
142 		return (EINVAL);
143 
144 	error = altq_attach(&ifp->if_snd, ALTQT_FAIRQ, a->altq_disc,
145 	    fairq_enqueue, fairq_dequeue, fairq_request, NULL, NULL);
146 
147 	return (error);
148 }
149 
150 int
151 fairq_add_altq(struct pf_altq *a)
152 {
153 	struct fairq_if *pif;
154 	struct ifnet *ifp;
155 
156 	if ((ifp = ifunit(a->ifname)) == NULL)
157 		return (EINVAL);
158 	if (!ALTQ_IS_READY(&ifp->if_snd))
159 		return (ENODEV);
160 
161 
162 	pif = malloc(sizeof(struct fairq_if),
163 			M_DEVBUF, M_WAITOK | M_ZERO);
164 	pif->pif_bandwidth = a->ifbandwidth;
165 	pif->pif_maxpri = -1;
166 	pif->pif_ifq = &ifp->if_snd;
167 
168 	/* keep the state in pf_altq */
169 	a->altq_disc = pif;
170 
171 	return (0);
172 }
173 
174 int
175 fairq_remove_altq(struct pf_altq *a)
176 {
177 	struct fairq_if *pif;
178 
179 	if ((pif = a->altq_disc) == NULL)
180 		return (EINVAL);
181 	a->altq_disc = NULL;
182 
183 	fairq_clear_interface(pif);
184 
185 	free(pif, M_DEVBUF);
186 	return (0);
187 }
188 
189 int
190 fairq_add_queue(struct pf_altq *a)
191 {
192 	struct fairq_if *pif;
193 	struct fairq_class *cl;
194 
195 	if ((pif = a->altq_disc) == NULL)
196 		return (EINVAL);
197 
198 	/* check parameters */
199 	if (a->priority >= FAIRQ_MAXPRI)
200 		return (EINVAL);
201 	if (a->qid == 0)
202 		return (EINVAL);
203 	if (pif->pif_classes[a->priority] != NULL)
204 		return (EBUSY);
205 	if (clh_to_clp(pif, a->qid) != NULL)
206 		return (EBUSY);
207 
208 	cl = fairq_class_create(pif, a->priority, a->qlimit, a->bandwidth,
209 			       &a->pq_u.fairq_opts, a->qid);
210 	if (cl == NULL)
211 		return (ENOMEM);
212 
213 	return (0);
214 }
215 
216 int
217 fairq_remove_queue(struct pf_altq *a)
218 {
219 	struct fairq_if *pif;
220 	struct fairq_class *cl;
221 
222 	if ((pif = a->altq_disc) == NULL)
223 		return (EINVAL);
224 
225 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
226 		return (EINVAL);
227 
228 	return (fairq_class_destroy(cl));
229 }
230 
231 int
232 fairq_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
233 {
234 	struct fairq_if *pif;
235 	struct fairq_class *cl;
236 	struct fairq_classstats stats;
237 	int error = 0;
238 
239 	if ((pif = altq_lookup(a->ifname, ALTQT_FAIRQ)) == NULL)
240 		return (EBADF);
241 
242 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
243 		return (EINVAL);
244 
245 	if (*nbytes < sizeof(stats))
246 		return (EINVAL);
247 
248 	get_class_stats(&stats, cl);
249 
250 	if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
251 		return (error);
252 	*nbytes = sizeof(stats);
253 	return (0);
254 }
255 
256 /*
257  * bring the interface back to the initial state by discarding
258  * all the filters and classes.
259  */
260 static int
261 fairq_clear_interface(struct fairq_if *pif)
262 {
263 	struct fairq_class *cl;
264 	int pri;
265 
266 	/* clear out the classes */
267 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
268 		if ((cl = pif->pif_classes[pri]) != NULL)
269 			fairq_class_destroy(cl);
270 	}
271 
272 	return (0);
273 }
274 
275 static int
276 fairq_request(struct ifaltq *ifq, int req, void *arg)
277 {
278 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
279 
280 	IFQ_LOCK_ASSERT(ifq);
281 
282 	switch (req) {
283 	case ALTRQ_PURGE:
284 		fairq_purge(pif);
285 		break;
286 	}
287 	return (0);
288 }
289 
290 /* discard all the queued packets on the interface */
291 static void
292 fairq_purge(struct fairq_if *pif)
293 {
294 	struct fairq_class *cl;
295 	int pri;
296 
297 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
298 		if ((cl = pif->pif_classes[pri]) != NULL && cl->cl_head)
299 			fairq_purgeq(cl);
300 	}
301 	if (ALTQ_IS_ENABLED(pif->pif_ifq))
302 		pif->pif_ifq->ifq_len = 0;
303 }
304 
305 static struct fairq_class *
306 fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
307 		   u_int bandwidth, struct fairq_opts *opts, int qid)
308 {
309 	struct fairq_class *cl;
310 	int flags = opts->flags;
311 	u_int nbuckets = opts->nbuckets;
312 	int i;
313 
314 #ifndef ALTQ_RED
315 	if (flags & FARF_RED) {
316 #ifdef ALTQ_DEBUG
317 		printf("fairq_class_create: RED not configured for FAIRQ!\n");
318 #endif
319 		return (NULL);
320 	}
321 #endif
322 #ifndef ALTQ_CODEL
323 	if (flags & FARF_CODEL) {
324 #ifdef ALTQ_DEBUG
325 		printf("fairq_class_create: CODEL not configured for FAIRQ!\n");
326 #endif
327 		return (NULL);
328 	}
329 #endif
330 	if (nbuckets == 0)
331 		nbuckets = 256;
332 	if (nbuckets > FAIRQ_MAX_BUCKETS)
333 		nbuckets = FAIRQ_MAX_BUCKETS;
334 	/* enforce power-of-2 size */
335 	while ((nbuckets ^ (nbuckets - 1)) != ((nbuckets << 1) - 1))
336 		++nbuckets;
337 
338 	if ((cl = pif->pif_classes[pri]) != NULL) {
339 		/* modify the class instead of creating a new one */
340 		IFQ_LOCK(cl->cl_pif->pif_ifq);
341 		if (cl->cl_head)
342 			fairq_purgeq(cl);
343 		IFQ_UNLOCK(cl->cl_pif->pif_ifq);
344 #ifdef ALTQ_RIO
345 		if (cl->cl_qtype == Q_RIO)
346 			rio_destroy((rio_t *)cl->cl_red);
347 #endif
348 #ifdef ALTQ_RED
349 		if (cl->cl_qtype == Q_RED)
350 			red_destroy(cl->cl_red);
351 #endif
352 #ifdef ALTQ_CODEL
353 		if (cl->cl_qtype == Q_CODEL)
354 			codel_destroy(cl->cl_codel);
355 #endif
356 	} else {
357 		cl = malloc(sizeof(struct fairq_class),
358 				M_DEVBUF, M_WAITOK | M_ZERO);
359 		cl->cl_nbuckets = nbuckets;
360 		cl->cl_nbucket_mask = nbuckets - 1;
361 
362 		cl->cl_buckets = malloc(
363 			sizeof(struct fairq_bucket) * cl->cl_nbuckets,
364 			M_DEVBUF, M_WAITOK | M_ZERO);
365 		cl->cl_head = NULL;
366 	}
367 
368 	pif->pif_classes[pri] = cl;
369 	if (flags & FARF_DEFAULTCLASS)
370 		pif->pif_default = cl;
371 	if (qlimit == 0)
372 		qlimit = 50;  /* use default */
373 	cl->cl_qlimit = qlimit;
374 	for (i = 0; i < cl->cl_nbuckets; ++i) {
375 		qlimit(&cl->cl_buckets[i].queue) = qlimit;
376 	}
377 	cl->cl_bandwidth = bandwidth / 8;
378 	cl->cl_qtype = Q_DROPTAIL;
379 	cl->cl_flags = flags & FARF_USERFLAGS;
380 	cl->cl_pri = pri;
381 	if (pri > pif->pif_maxpri)
382 		pif->pif_maxpri = pri;
383 	cl->cl_pif = pif;
384 	cl->cl_handle = qid;
385 	cl->cl_hogs_m1 = opts->hogs_m1 / 8;
386 	cl->cl_lssc_m1 = opts->lssc_m1 / 8;	/* NOT YET USED */
387 
388 #ifdef ALTQ_RED
389 	if (flags & (FARF_RED|FARF_RIO)) {
390 		int red_flags, red_pkttime;
391 
392 		red_flags = 0;
393 		if (flags & FARF_ECN)
394 			red_flags |= REDF_ECN;
395 #ifdef ALTQ_RIO
396 		if (flags & FARF_CLEARDSCP)
397 			red_flags |= RIOF_CLEARDSCP;
398 #endif
399 		if (pif->pif_bandwidth < 8)
400 			red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
401 		else
402 			red_pkttime = (int64_t)pif->pif_ifq->altq_ifp->if_mtu
403 			  * 1000 * 1000 * 1000 / (pif->pif_bandwidth / 8);
404 #ifdef ALTQ_RIO
405 		if (flags & FARF_RIO) {
406 			cl->cl_red = (red_t *)rio_alloc(0, NULL,
407 						red_flags, red_pkttime);
408 			if (cl->cl_red != NULL)
409 				cl->cl_qtype = Q_RIO;
410 		} else
411 #endif
412 		if (flags & FARF_RED) {
413 			cl->cl_red = red_alloc(0, 0,
414 			    cl->cl_qlimit * 10/100,
415 			    cl->cl_qlimit * 30/100,
416 			    red_flags, red_pkttime);
417 			if (cl->cl_red != NULL)
418 				cl->cl_qtype = Q_RED;
419 		}
420 	}
421 #endif /* ALTQ_RED */
422 #ifdef ALTQ_CODEL
423 	if (flags & FARF_CODEL) {
424 		cl->cl_codel = codel_alloc(5, 100, 0);
425 		if (cl->cl_codel != NULL)
426 			cl->cl_qtype = Q_CODEL;
427 	}
428 #endif
429 
430 	return (cl);
431 }
432 
433 static int
434 fairq_class_destroy(struct fairq_class *cl)
435 {
436 	struct fairq_if *pif;
437 	int pri;
438 
439 	IFQ_LOCK(cl->cl_pif->pif_ifq);
440 
441 	if (cl->cl_head)
442 		fairq_purgeq(cl);
443 
444 	pif = cl->cl_pif;
445 	pif->pif_classes[cl->cl_pri] = NULL;
446 	if (pif->pif_poll_cache == cl)
447 		pif->pif_poll_cache = NULL;
448 	if (pif->pif_maxpri == cl->cl_pri) {
449 		for (pri = cl->cl_pri; pri >= 0; pri--)
450 			if (pif->pif_classes[pri] != NULL) {
451 				pif->pif_maxpri = pri;
452 				break;
453 			}
454 		if (pri < 0)
455 			pif->pif_maxpri = -1;
456 	}
457 	IFQ_UNLOCK(cl->cl_pif->pif_ifq);
458 
459 	if (cl->cl_red != NULL) {
460 #ifdef ALTQ_RIO
461 		if (cl->cl_qtype == Q_RIO)
462 			rio_destroy((rio_t *)cl->cl_red);
463 #endif
464 #ifdef ALTQ_RED
465 		if (cl->cl_qtype == Q_RED)
466 			red_destroy(cl->cl_red);
467 #endif
468 #ifdef ALTQ_CODEL
469 		if (cl->cl_qtype == Q_CODEL)
470 			codel_destroy(cl->cl_codel);
471 #endif
472 	}
473 	free(cl->cl_buckets, M_DEVBUF);
474 	free(cl, M_DEVBUF);
475 
476 	return (0);
477 }
478 
479 /*
480  * fairq_enqueue is an enqueue function to be registered to
481  * (*altq_enqueue) in struct ifaltq.
482  */
483 static int
484 fairq_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
485 {
486 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
487 	struct fairq_class *cl = NULL; /* Make compiler happy */
488 	struct pf_mtag *t;
489 	u_int32_t qid_hash = 0;
490 	int len;
491 
492 	IFQ_LOCK_ASSERT(ifq);
493 
494 	/* grab class set by classifier */
495 	if ((m->m_flags & M_PKTHDR) == 0) {
496 		/* should not happen */
497 		printf("altq: packet for %s does not have pkthdr\n",
498 			ifq->altq_ifp->if_xname);
499 		m_freem(m);
500 		return (ENOBUFS);
501 	}
502 
503 	if ((t = pf_find_mtag(m)) != NULL) {
504 		cl = clh_to_clp(pif, t->qid);
505 		qid_hash = t->qid_hash;
506 	}
507 	if (cl == NULL) {
508 		cl = pif->pif_default;
509 		if (cl == NULL) {
510 			m_freem(m);
511 			return (ENOBUFS);
512 		}
513 	}
514 	cl->cl_flags |= FARF_HAS_PACKETS;
515 	cl->cl_pktattr = NULL;
516 	len = m_pktlen(m);
517 	if (fairq_addq(cl, m, qid_hash) != 0) {
518 		/* drop occurred.  mbuf was freed in fairq_addq. */
519 		PKTCNTR_ADD(&cl->cl_dropcnt, len);
520 		return (ENOBUFS);
521 	}
522 	IFQ_INC_LEN(ifq);
523 
524 	return (0);
525 }
526 
527 /*
528  * fairq_dequeue is a dequeue function to be registered to
529  * (*altq_dequeue) in struct ifaltq.
530  *
531  * note: ALTDQ_POLL returns the next packet without removing the packet
532  *	from the queue.  ALTDQ_REMOVE is a normal dequeue operation.
533  *	ALTDQ_REMOVE must return the same packet if called immediately
534  *	after ALTDQ_POLL.
535  */
536 static struct mbuf *
537 fairq_dequeue(struct ifaltq *ifq, int op)
538 {
539 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
540 	struct fairq_class *cl;
541 	struct fairq_class *best_cl;
542 	struct mbuf *best_m;
543 	struct mbuf *m = NULL;
544 	uint64_t cur_time = read_machclk();
545 	int pri;
546 	int hit_limit;
547 
548 	IFQ_LOCK_ASSERT(ifq);
549 
550 	if (IFQ_IS_EMPTY(ifq)) {
551 		return (NULL);
552 	}
553 
554 	if (pif->pif_poll_cache && op == ALTDQ_REMOVE) {
555 		best_cl = pif->pif_poll_cache;
556 		m = fairq_getq(best_cl, cur_time);
557 		pif->pif_poll_cache = NULL;
558 		if (m) {
559 			IFQ_DEC_LEN(ifq);
560 			PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
561 			return (m);
562 		}
563 	} else {
564 		best_cl = NULL;
565 		best_m = NULL;
566 
567 		for (pri = pif->pif_maxpri;  pri >= 0; pri--) {
568 			if ((cl = pif->pif_classes[pri]) == NULL)
569 				continue;
570 			if ((cl->cl_flags & FARF_HAS_PACKETS) == 0)
571 				continue;
572 			m = fairq_pollq(cl, cur_time, &hit_limit);
573 			if (m == NULL) {
574 				cl->cl_flags &= ~FARF_HAS_PACKETS;
575 				continue;
576 			}
577 
578 			/*
579 			 * Only override the best choice if we are under
580 			 * the BW limit.
581 			 */
582 			if (hit_limit == 0 || best_cl == NULL) {
583 				best_cl = cl;
584 				best_m = m;
585 			}
586 
587 			/*
588 			 * Remember the highest priority mbuf in case we
589 			 * do not find any lower priority mbufs.
590 			 */
591 			if (hit_limit)
592 				continue;
593 			break;
594 		}
595 		if (op == ALTDQ_POLL) {
596 			pif->pif_poll_cache = best_cl;
597 			m = best_m;
598 		} else if (best_cl) {
599 			m = fairq_getq(best_cl, cur_time);
600 			if (m != NULL) {
601 				IFQ_DEC_LEN(ifq);
602 				PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
603 			}
604 		}
605 		return (m);
606 	}
607 	return (NULL);
608 }
609 
610 static int
611 fairq_addq(struct fairq_class *cl, struct mbuf *m, u_int32_t bucketid)
612 {
613 	fairq_bucket_t *b;
614 	u_int hindex;
615 	uint64_t bw;
616 
617 	/*
618 	 * If the packet doesn't have any keep state put it on the end of
619 	 * our queue.  XXX this can result in out of order delivery.
620 	 */
621 	if (bucketid == 0) {
622 		if (cl->cl_head)
623 			b = cl->cl_head->prev;
624 		else
625 			b = &cl->cl_buckets[0];
626 	} else {
627 		hindex = bucketid & cl->cl_nbucket_mask;
628 		b = &cl->cl_buckets[hindex];
629 	}
630 
631 	/*
632 	 * Add the bucket to the end of the circular list of active buckets.
633 	 *
634 	 * As a special case we add the bucket to the beginning of the list
635 	 * instead of the end if it was not previously on the list and if
636 	 * its traffic is less then the hog level.
637 	 */
638 	if (b->in_use == 0) {
639 		b->in_use = 1;
640 		if (cl->cl_head == NULL) {
641 			cl->cl_head = b;
642 			b->next = b;
643 			b->prev = b;
644 		} else {
645 			b->next = cl->cl_head;
646 			b->prev = cl->cl_head->prev;
647 			b->prev->next = b;
648 			b->next->prev = b;
649 
650 			if (b->bw_delta && cl->cl_hogs_m1) {
651 				bw = b->bw_bytes * machclk_freq / b->bw_delta;
652 				if (bw < cl->cl_hogs_m1)
653 					cl->cl_head = b;
654 			}
655 		}
656 	}
657 
658 #ifdef ALTQ_RIO
659 	if (cl->cl_qtype == Q_RIO)
660 		return rio_addq((rio_t *)cl->cl_red, &b->queue, m, cl->cl_pktattr);
661 #endif
662 #ifdef ALTQ_RED
663 	if (cl->cl_qtype == Q_RED)
664 		return red_addq(cl->cl_red, &b->queue, m, cl->cl_pktattr);
665 #endif
666 #ifdef ALTQ_CODEL
667 	if (cl->cl_qtype == Q_CODEL)
668 		return codel_addq(cl->cl_codel, &b->queue, m);
669 #endif
670 	if (qlen(&b->queue) >= qlimit(&b->queue)) {
671 		m_freem(m);
672 		return (-1);
673 	}
674 
675 	if (cl->cl_flags & FARF_CLEARDSCP)
676 		write_dsfield(m, cl->cl_pktattr, 0);
677 
678 	_addq(&b->queue, m);
679 
680 	return (0);
681 }
682 
683 static struct mbuf *
684 fairq_getq(struct fairq_class *cl, uint64_t cur_time)
685 {
686 	fairq_bucket_t *b;
687 	struct mbuf *m;
688 
689 	b = fairq_selectq(cl, 0);
690 	if (b == NULL)
691 		m = NULL;
692 #ifdef ALTQ_RIO
693 	else if (cl->cl_qtype == Q_RIO)
694 		m = rio_getq((rio_t *)cl->cl_red, &b->queue);
695 #endif
696 #ifdef ALTQ_RED
697 	else if (cl->cl_qtype == Q_RED)
698 		m = red_getq(cl->cl_red, &b->queue);
699 #endif
700 #ifdef ALTQ_CODEL
701 	else if (cl->cl_qtype == Q_CODEL)
702 		m = codel_getq(cl->cl_codel, &b->queue);
703 #endif
704 	else
705 		m = _getq(&b->queue);
706 
707 	/*
708 	 * Calculate the BW change
709 	 */
710 	if (m != NULL) {
711 		uint64_t delta;
712 
713 		/*
714 		 * Per-class bandwidth calculation
715 		 */
716 		delta = (cur_time - cl->cl_last_time);
717 		if (delta > machclk_freq * 8)
718 			delta = machclk_freq * 8;
719 		cl->cl_bw_delta += delta;
720 		cl->cl_bw_bytes += m->m_pkthdr.len;
721 		cl->cl_last_time = cur_time;
722 		cl->cl_bw_delta -= cl->cl_bw_delta >> 3;
723 		cl->cl_bw_bytes -= cl->cl_bw_bytes >> 3;
724 
725 		/*
726 		 * Per-bucket bandwidth calculation
727 		 */
728 		delta = (cur_time - b->last_time);
729 		if (delta > machclk_freq * 8)
730 			delta = machclk_freq * 8;
731 		b->bw_delta += delta;
732 		b->bw_bytes += m->m_pkthdr.len;
733 		b->last_time = cur_time;
734 		b->bw_delta -= b->bw_delta >> 3;
735 		b->bw_bytes -= b->bw_bytes >> 3;
736 	}
737 	return(m);
738 }
739 
740 /*
741  * Figure out what the next packet would be if there were no limits.  If
742  * this class hits its bandwidth limit *hit_limit is set to no-zero, otherwise
743  * it is set to 0.  A non-NULL mbuf is returned either way.
744  */
745 static struct mbuf *
746 fairq_pollq(struct fairq_class *cl, uint64_t cur_time, int *hit_limit)
747 {
748 	fairq_bucket_t *b;
749 	struct mbuf *m;
750 	uint64_t delta;
751 	uint64_t bw;
752 
753 	*hit_limit = 0;
754 	b = fairq_selectq(cl, 1);
755 	if (b == NULL)
756 		return(NULL);
757 	m = qhead(&b->queue);
758 
759 	/*
760 	 * Did this packet exceed the class bandwidth?  Calculate the
761 	 * bandwidth component of the packet.
762 	 *
763 	 * - Calculate bytes per second
764 	 */
765 	delta = cur_time - cl->cl_last_time;
766 	if (delta > machclk_freq * 8)
767 		delta = machclk_freq * 8;
768 	cl->cl_bw_delta += delta;
769 	cl->cl_last_time = cur_time;
770 	if (cl->cl_bw_delta) {
771 		bw = cl->cl_bw_bytes * machclk_freq / cl->cl_bw_delta;
772 
773 		if (bw > cl->cl_bandwidth)
774 			*hit_limit = 1;
775 #ifdef ALTQ_DEBUG
776 		printf("BW %6ju relative to %6u %d queue %p\n",
777 			(uintmax_t)bw, cl->cl_bandwidth, *hit_limit, b);
778 #endif
779 	}
780 	return(m);
781 }
782 
783 /*
784  * Locate the next queue we want to pull a packet out of.  This code
785  * is also responsible for removing empty buckets from the circular list.
786  */
787 static
788 fairq_bucket_t *
789 fairq_selectq(struct fairq_class *cl, int ispoll)
790 {
791 	fairq_bucket_t *b;
792 	uint64_t bw;
793 
794 	if (ispoll == 0 && cl->cl_polled) {
795 		b = cl->cl_polled;
796 		cl->cl_polled = NULL;
797 		return(b);
798 	}
799 
800 	while ((b = cl->cl_head) != NULL) {
801 		/*
802 		 * Remove empty queues from consideration
803 		 */
804 		if (qempty(&b->queue)) {
805 			b->in_use = 0;
806 			cl->cl_head = b->next;
807 			if (cl->cl_head == b) {
808 				cl->cl_head = NULL;
809 			} else {
810 				b->next->prev = b->prev;
811 				b->prev->next = b->next;
812 			}
813 			continue;
814 		}
815 
816 		/*
817 		 * Advance the round robin.  Queues with bandwidths less
818 		 * then the hog bandwidth are allowed to burst.
819 		 */
820 		if (cl->cl_hogs_m1 == 0) {
821 			cl->cl_head = b->next;
822 		} else if (b->bw_delta) {
823 			bw = b->bw_bytes * machclk_freq / b->bw_delta;
824 			if (bw >= cl->cl_hogs_m1) {
825 				cl->cl_head = b->next;
826 			}
827 			/*
828 			 * XXX TODO -
829 			 */
830 		}
831 
832 		/*
833 		 * Return bucket b.
834 		 */
835 		break;
836 	}
837 	if (ispoll)
838 		cl->cl_polled = b;
839 	return(b);
840 }
841 
842 static void
843 fairq_purgeq(struct fairq_class *cl)
844 {
845 	fairq_bucket_t *b;
846 	struct mbuf *m;
847 
848 	while ((b = fairq_selectq(cl, 0)) != NULL) {
849 		while ((m = _getq(&b->queue)) != NULL) {
850 			PKTCNTR_ADD(&cl->cl_dropcnt, m_pktlen(m));
851 			m_freem(m);
852 		}
853 		ASSERT(qlen(&b->queue) == 0);
854 	}
855 }
856 
857 static void
858 get_class_stats(struct fairq_classstats *sp, struct fairq_class *cl)
859 {
860 	fairq_bucket_t *b;
861 
862 	sp->class_handle = cl->cl_handle;
863 	sp->qlimit = cl->cl_qlimit;
864 	sp->xmit_cnt = cl->cl_xmitcnt;
865 	sp->drop_cnt = cl->cl_dropcnt;
866 	sp->qtype = cl->cl_qtype;
867 	sp->qlength = 0;
868 
869 	if (cl->cl_head) {
870 		b = cl->cl_head;
871 		do {
872 			sp->qlength += qlen(&b->queue);
873 			b = b->next;
874 		} while (b != cl->cl_head);
875 	}
876 
877 #ifdef ALTQ_RED
878 	if (cl->cl_qtype == Q_RED)
879 		red_getstats(cl->cl_red, &sp->red[0]);
880 #endif
881 #ifdef ALTQ_RIO
882 	if (cl->cl_qtype == Q_RIO)
883 		rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
884 #endif
885 #ifdef ALTQ_CODEL
886 	if (cl->cl_qtype == Q_CODEL)
887 		codel_getstats(cl->cl_codel, &sp->codel);
888 #endif
889 }
890 
891 /* convert a class handle to the corresponding class pointer */
892 static struct fairq_class *
893 clh_to_clp(struct fairq_if *pif, uint32_t chandle)
894 {
895 	struct fairq_class *cl;
896 	int idx;
897 
898 	if (chandle == 0)
899 		return (NULL);
900 
901 	for (idx = pif->pif_maxpri; idx >= 0; idx--)
902 		if ((cl = pif->pif_classes[idx]) != NULL &&
903 		    cl->cl_handle == chandle)
904 			return (cl);
905 
906 	return (NULL);
907 }
908 
909 #endif /* ALTQ_FAIRQ */
910