xref: /freebsd/sys/net/altq/altq_fairq.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
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 <net/if_private.h>
108 #include <netinet/in.h>
109 
110 #include <netpfil/pf/pf.h>
111 #include <netpfil/pf/pf_altq.h>
112 #include <netpfil/pf/pf_mtag.h>
113 #include <net/altq/altq.h>
114 #include <net/altq/altq_fairq.h>
115 
116 /*
117  * function prototypes
118  */
119 static int	fairq_clear_interface(struct fairq_if *);
120 static int	fairq_request(struct ifaltq *, int, void *);
121 static void	fairq_purge(struct fairq_if *);
122 static struct fairq_class *fairq_class_create(struct fairq_if *, int, int, u_int, struct fairq_opts *, int);
123 static int	fairq_class_destroy(struct fairq_class *);
124 static int	fairq_enqueue(struct ifaltq *, struct mbuf *, struct altq_pktattr *);
125 static struct mbuf *fairq_dequeue(struct ifaltq *, int);
126 
127 static int	fairq_addq(struct fairq_class *, struct mbuf *, u_int32_t);
128 static struct mbuf *fairq_getq(struct fairq_class *, uint64_t);
129 static struct mbuf *fairq_pollq(struct fairq_class *, uint64_t, int *);
130 static fairq_bucket_t *fairq_selectq(struct fairq_class *, int);
131 static void	fairq_purgeq(struct fairq_class *);
132 
133 static void	get_class_stats(struct fairq_classstats *, struct fairq_class *);
134 static struct fairq_class *clh_to_clp(struct fairq_if *, uint32_t);
135 
136 int
137 fairq_pfattach(struct pf_altq *a)
138 {
139 	struct ifnet *ifp;
140 	int error;
141 
142 	if ((ifp = ifunit(a->ifname)) == NULL || a->altq_disc == NULL)
143 		return (EINVAL);
144 
145 	error = altq_attach(&ifp->if_snd, ALTQT_FAIRQ, a->altq_disc,
146 	    fairq_enqueue, fairq_dequeue, fairq_request);
147 
148 	return (error);
149 }
150 
151 int
152 fairq_add_altq(struct ifnet *ifp, struct pf_altq *a)
153 {
154 	struct fairq_if *pif;
155 
156 	if (ifp == NULL)
157 		return (EINVAL);
158 	if (!ALTQ_IS_READY(&ifp->if_snd))
159 		return (ENODEV);
160 
161 	pif = malloc(sizeof(struct fairq_if),
162 			M_DEVBUF, M_WAITOK | M_ZERO);
163 	pif->pif_bandwidth = a->ifbandwidth;
164 	pif->pif_maxpri = -1;
165 	pif->pif_ifq = &ifp->if_snd;
166 
167 	/* keep the state in pf_altq */
168 	a->altq_disc = pif;
169 
170 	return (0);
171 }
172 
173 int
174 fairq_remove_altq(struct pf_altq *a)
175 {
176 	struct fairq_if *pif;
177 
178 	if ((pif = a->altq_disc) == NULL)
179 		return (EINVAL);
180 	a->altq_disc = NULL;
181 
182 	fairq_clear_interface(pif);
183 
184 	free(pif, M_DEVBUF);
185 	return (0);
186 }
187 
188 int
189 fairq_add_queue(struct pf_altq *a)
190 {
191 	struct fairq_if *pif;
192 	struct fairq_class *cl;
193 
194 	if ((pif = a->altq_disc) == NULL)
195 		return (EINVAL);
196 
197 	/* check parameters */
198 	if (a->priority >= FAIRQ_MAXPRI)
199 		return (EINVAL);
200 	if (a->qid == 0)
201 		return (EINVAL);
202 	if (pif->pif_classes[a->priority] != NULL)
203 		return (EBUSY);
204 	if (clh_to_clp(pif, a->qid) != NULL)
205 		return (EBUSY);
206 
207 	cl = fairq_class_create(pif, a->priority, a->qlimit, a->bandwidth,
208 			       &a->pq_u.fairq_opts, a->qid);
209 	if (cl == NULL)
210 		return (ENOMEM);
211 
212 	return (0);
213 }
214 
215 int
216 fairq_remove_queue(struct pf_altq *a)
217 {
218 	struct fairq_if *pif;
219 	struct fairq_class *cl;
220 
221 	if ((pif = a->altq_disc) == NULL)
222 		return (EINVAL);
223 
224 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
225 		return (EINVAL);
226 
227 	return (fairq_class_destroy(cl));
228 }
229 
230 int
231 fairq_getqstats(struct pf_altq *a, void *ubuf, int *nbytes, int version)
232 {
233 	struct fairq_if *pif;
234 	struct fairq_class *cl;
235 	struct fairq_classstats stats;
236 	int error = 0;
237 
238 	if ((pif = altq_lookup(a->ifname, ALTQT_FAIRQ)) == NULL)
239 		return (EBADF);
240 
241 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
242 		return (EINVAL);
243 
244 	if (*nbytes < sizeof(stats))
245 		return (EINVAL);
246 
247 	get_class_stats(&stats, cl);
248 
249 	if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
250 		return (error);
251 	*nbytes = sizeof(stats);
252 	return (0);
253 }
254 
255 /*
256  * bring the interface back to the initial state by discarding
257  * all the filters and classes.
258  */
259 static int
260 fairq_clear_interface(struct fairq_if *pif)
261 {
262 	struct fairq_class *cl;
263 	int pri;
264 
265 	/* clear out the classes */
266 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
267 		if ((cl = pif->pif_classes[pri]) != NULL)
268 			fairq_class_destroy(cl);
269 	}
270 
271 	return (0);
272 }
273 
274 static int
275 fairq_request(struct ifaltq *ifq, int req, void *arg)
276 {
277 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
278 
279 	IFQ_LOCK_ASSERT(ifq);
280 
281 	switch (req) {
282 	case ALTRQ_PURGE:
283 		fairq_purge(pif);
284 		break;
285 	}
286 	return (0);
287 }
288 
289 /* discard all the queued packets on the interface */
290 static void
291 fairq_purge(struct fairq_if *pif)
292 {
293 	struct fairq_class *cl;
294 	int pri;
295 
296 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
297 		if ((cl = pif->pif_classes[pri]) != NULL && cl->cl_head)
298 			fairq_purgeq(cl);
299 	}
300 	if (ALTQ_IS_ENABLED(pif->pif_ifq))
301 		pif->pif_ifq->ifq_len = 0;
302 }
303 
304 static struct fairq_class *
305 fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
306 		   u_int bandwidth, struct fairq_opts *opts, int qid)
307 {
308 	struct fairq_class *cl;
309 	int flags = opts->flags;
310 	u_int nbuckets = opts->nbuckets;
311 	int i;
312 
313 #ifndef ALTQ_RED
314 	if (flags & FARF_RED) {
315 #ifdef ALTQ_DEBUG
316 		printf("fairq_class_create: RED not configured for FAIRQ!\n");
317 #endif
318 		return (NULL);
319 	}
320 #endif
321 #ifndef ALTQ_CODEL
322 	if (flags & FARF_CODEL) {
323 #ifdef ALTQ_DEBUG
324 		printf("fairq_class_create: CODEL not configured for FAIRQ!\n");
325 #endif
326 		return (NULL);
327 	}
328 #endif
329 	if (nbuckets == 0)
330 		nbuckets = 256;
331 	if (nbuckets > FAIRQ_MAX_BUCKETS)
332 		nbuckets = FAIRQ_MAX_BUCKETS;
333 	/* enforce power-of-2 size */
334 	while ((nbuckets ^ (nbuckets - 1)) != ((nbuckets << 1) - 1))
335 		++nbuckets;
336 
337 	if ((cl = pif->pif_classes[pri]) != NULL) {
338 		/* modify the class instead of creating a new one */
339 		IFQ_LOCK(cl->cl_pif->pif_ifq);
340 		if (cl->cl_head)
341 			fairq_purgeq(cl);
342 		IFQ_UNLOCK(cl->cl_pif->pif_ifq);
343 #ifdef ALTQ_RIO
344 		if (cl->cl_qtype == Q_RIO)
345 			rio_destroy((rio_t *)cl->cl_red);
346 #endif
347 #ifdef ALTQ_RED
348 		if (cl->cl_qtype == Q_RED)
349 			red_destroy(cl->cl_red);
350 #endif
351 #ifdef ALTQ_CODEL
352 		if (cl->cl_qtype == Q_CODEL)
353 			codel_destroy(cl->cl_codel);
354 #endif
355 	} else {
356 		cl = malloc(sizeof(struct fairq_class),
357 				M_DEVBUF, M_WAITOK | M_ZERO);
358 		cl->cl_nbuckets = nbuckets;
359 		cl->cl_nbucket_mask = nbuckets - 1;
360 
361 		cl->cl_buckets = malloc(
362 			sizeof(struct fairq_bucket) * cl->cl_nbuckets,
363 			M_DEVBUF, M_WAITOK | M_ZERO);
364 		cl->cl_head = NULL;
365 	}
366 
367 	pif->pif_classes[pri] = cl;
368 	if (flags & FARF_DEFAULTCLASS)
369 		pif->pif_default = cl;
370 	if (qlimit == 0)
371 		qlimit = 50;  /* use default */
372 	cl->cl_qlimit = qlimit;
373 	for (i = 0; i < cl->cl_nbuckets; ++i) {
374 		qlimit(&cl->cl_buckets[i].queue) = qlimit;
375 	}
376 	cl->cl_bandwidth = bandwidth / 8;
377 	cl->cl_qtype = Q_DROPTAIL;
378 	cl->cl_flags = flags & FARF_USERFLAGS;
379 	cl->cl_pri = pri;
380 	if (pri > pif->pif_maxpri)
381 		pif->pif_maxpri = pri;
382 	cl->cl_pif = pif;
383 	cl->cl_handle = qid;
384 	cl->cl_hogs_m1 = opts->hogs_m1 / 8;
385 	cl->cl_lssc_m1 = opts->lssc_m1 / 8;	/* NOT YET USED */
386 
387 #ifdef ALTQ_RED
388 	if (flags & (FARF_RED|FARF_RIO)) {
389 		int red_flags, red_pkttime;
390 
391 		red_flags = 0;
392 		if (flags & FARF_ECN)
393 			red_flags |= REDF_ECN;
394 #ifdef ALTQ_RIO
395 		if (flags & FARF_CLEARDSCP)
396 			red_flags |= RIOF_CLEARDSCP;
397 #endif
398 		if (pif->pif_bandwidth < 8)
399 			red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
400 		else
401 			red_pkttime = (int64_t)pif->pif_ifq->altq_ifp->if_mtu
402 			  * 1000 * 1000 * 1000 / (pif->pif_bandwidth / 8);
403 #ifdef ALTQ_RIO
404 		if (flags & FARF_RIO) {
405 			cl->cl_red = (red_t *)rio_alloc(0, NULL,
406 						red_flags, red_pkttime);
407 			if (cl->cl_red != NULL)
408 				cl->cl_qtype = Q_RIO;
409 		} else
410 #endif
411 		if (flags & FARF_RED) {
412 			cl->cl_red = red_alloc(0, 0,
413 			    cl->cl_qlimit * 10/100,
414 			    cl->cl_qlimit * 30/100,
415 			    red_flags, red_pkttime);
416 			if (cl->cl_red != NULL)
417 				cl->cl_qtype = Q_RED;
418 		}
419 	}
420 #endif /* ALTQ_RED */
421 #ifdef ALTQ_CODEL
422 	if (flags & FARF_CODEL) {
423 		cl->cl_codel = codel_alloc(5, 100, 0);
424 		if (cl->cl_codel != NULL)
425 			cl->cl_qtype = Q_CODEL;
426 	}
427 #endif
428 
429 	return (cl);
430 }
431 
432 static int
433 fairq_class_destroy(struct fairq_class *cl)
434 {
435 	struct fairq_if *pif;
436 	int pri;
437 
438 	IFQ_LOCK(cl->cl_pif->pif_ifq);
439 
440 	if (cl->cl_head)
441 		fairq_purgeq(cl);
442 
443 	pif = cl->cl_pif;
444 	pif->pif_classes[cl->cl_pri] = NULL;
445 	if (pif->pif_poll_cache == cl)
446 		pif->pif_poll_cache = NULL;
447 	if (pif->pif_maxpri == cl->cl_pri) {
448 		for (pri = cl->cl_pri; pri >= 0; pri--)
449 			if (pif->pif_classes[pri] != NULL) {
450 				pif->pif_maxpri = pri;
451 				break;
452 			}
453 		if (pri < 0)
454 			pif->pif_maxpri = -1;
455 	}
456 	IFQ_UNLOCK(cl->cl_pif->pif_ifq);
457 
458 	if (cl->cl_red != NULL) {
459 #ifdef ALTQ_RIO
460 		if (cl->cl_qtype == Q_RIO)
461 			rio_destroy((rio_t *)cl->cl_red);
462 #endif
463 #ifdef ALTQ_RED
464 		if (cl->cl_qtype == Q_RED)
465 			red_destroy(cl->cl_red);
466 #endif
467 #ifdef ALTQ_CODEL
468 		if (cl->cl_qtype == Q_CODEL)
469 			codel_destroy(cl->cl_codel);
470 #endif
471 	}
472 	free(cl->cl_buckets, M_DEVBUF);
473 	free(cl, M_DEVBUF);
474 
475 	return (0);
476 }
477 
478 /*
479  * fairq_enqueue is an enqueue function to be registered to
480  * (*altq_enqueue) in struct ifaltq.
481  */
482 static int
483 fairq_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
484 {
485 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
486 	struct fairq_class *cl = NULL; /* Make compiler happy */
487 	struct pf_mtag *t;
488 	u_int32_t qid_hash = 0;
489 	int len;
490 
491 	IFQ_LOCK_ASSERT(ifq);
492 
493 	/* grab class set by classifier */
494 	if ((m->m_flags & M_PKTHDR) == 0) {
495 		/* should not happen */
496 		printf("altq: packet for %s does not have pkthdr\n",
497 			ifq->altq_ifp->if_xname);
498 		m_freem(m);
499 		return (ENOBUFS);
500 	}
501 
502 	if ((t = pf_find_mtag(m)) != NULL) {
503 		cl = clh_to_clp(pif, t->qid);
504 		qid_hash = t->qid_hash;
505 	}
506 	if (cl == NULL) {
507 		cl = pif->pif_default;
508 		if (cl == NULL) {
509 			m_freem(m);
510 			return (ENOBUFS);
511 		}
512 	}
513 	cl->cl_flags |= FARF_HAS_PACKETS;
514 	cl->cl_pktattr = NULL;
515 	len = m_pktlen(m);
516 	if (fairq_addq(cl, m, qid_hash) != 0) {
517 		/* drop occurred.  mbuf was freed in fairq_addq. */
518 		PKTCNTR_ADD(&cl->cl_dropcnt, len);
519 		return (ENOBUFS);
520 	}
521 	IFQ_INC_LEN(ifq);
522 
523 	return (0);
524 }
525 
526 /*
527  * fairq_dequeue is a dequeue function to be registered to
528  * (*altq_dequeue) in struct ifaltq.
529  *
530  * note: ALTDQ_POLL returns the next packet without removing the packet
531  *	from the queue.  ALTDQ_REMOVE is a normal dequeue operation.
532  *	ALTDQ_REMOVE must return the same packet if called immediately
533  *	after ALTDQ_POLL.
534  */
535 static struct mbuf *
536 fairq_dequeue(struct ifaltq *ifq, int op)
537 {
538 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
539 	struct fairq_class *cl;
540 	struct fairq_class *best_cl;
541 	struct mbuf *best_m;
542 	struct mbuf *m = NULL;
543 	uint64_t cur_time = read_machclk();
544 	int pri;
545 	int hit_limit;
546 
547 	IFQ_LOCK_ASSERT(ifq);
548 
549 	if (IFQ_IS_EMPTY(ifq)) {
550 		return (NULL);
551 	}
552 
553 	if (pif->pif_poll_cache && op == ALTDQ_REMOVE) {
554 		best_cl = pif->pif_poll_cache;
555 		m = fairq_getq(best_cl, cur_time);
556 		pif->pif_poll_cache = NULL;
557 		if (m) {
558 			IFQ_DEC_LEN(ifq);
559 			PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
560 			return (m);
561 		}
562 	} else {
563 		best_cl = NULL;
564 		best_m = NULL;
565 
566 		for (pri = pif->pif_maxpri;  pri >= 0; pri--) {
567 			if ((cl = pif->pif_classes[pri]) == NULL)
568 				continue;
569 			if ((cl->cl_flags & FARF_HAS_PACKETS) == 0)
570 				continue;
571 			m = fairq_pollq(cl, cur_time, &hit_limit);
572 			if (m == NULL) {
573 				cl->cl_flags &= ~FARF_HAS_PACKETS;
574 				continue;
575 			}
576 
577 			/*
578 			 * Only override the best choice if we are under
579 			 * the BW limit.
580 			 */
581 			if (hit_limit == 0 || best_cl == NULL) {
582 				best_cl = cl;
583 				best_m = m;
584 			}
585 
586 			/*
587 			 * Remember the highest priority mbuf in case we
588 			 * do not find any lower priority mbufs.
589 			 */
590 			if (hit_limit)
591 				continue;
592 			break;
593 		}
594 		if (op == ALTDQ_POLL) {
595 			pif->pif_poll_cache = best_cl;
596 			m = best_m;
597 		} else if (best_cl) {
598 			m = fairq_getq(best_cl, cur_time);
599 			if (m != NULL) {
600 				IFQ_DEC_LEN(ifq);
601 				PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
602 			}
603 		}
604 		return (m);
605 	}
606 	return (NULL);
607 }
608 
609 static int
610 fairq_addq(struct fairq_class *cl, struct mbuf *m, u_int32_t bucketid)
611 {
612 	fairq_bucket_t *b;
613 	u_int hindex;
614 	uint64_t bw;
615 
616 	/*
617 	 * If the packet doesn't have any keep state put it on the end of
618 	 * our queue.  XXX this can result in out of order delivery.
619 	 */
620 	if (bucketid == 0) {
621 		if (cl->cl_head)
622 			b = cl->cl_head->prev;
623 		else
624 			b = &cl->cl_buckets[0];
625 	} else {
626 		hindex = bucketid & cl->cl_nbucket_mask;
627 		b = &cl->cl_buckets[hindex];
628 	}
629 
630 	/*
631 	 * Add the bucket to the end of the circular list of active buckets.
632 	 *
633 	 * As a special case we add the bucket to the beginning of the list
634 	 * instead of the end if it was not previously on the list and if
635 	 * its traffic is less then the hog level.
636 	 */
637 	if (b->in_use == 0) {
638 		b->in_use = 1;
639 		if (cl->cl_head == NULL) {
640 			cl->cl_head = b;
641 			b->next = b;
642 			b->prev = b;
643 		} else {
644 			b->next = cl->cl_head;
645 			b->prev = cl->cl_head->prev;
646 			b->prev->next = b;
647 			b->next->prev = b;
648 
649 			if (b->bw_delta && cl->cl_hogs_m1) {
650 				bw = b->bw_bytes * machclk_freq / b->bw_delta;
651 				if (bw < cl->cl_hogs_m1)
652 					cl->cl_head = b;
653 			}
654 		}
655 	}
656 
657 #ifdef ALTQ_RIO
658 	if (cl->cl_qtype == Q_RIO)
659 		return rio_addq((rio_t *)cl->cl_red, &b->queue, m, cl->cl_pktattr);
660 #endif
661 #ifdef ALTQ_RED
662 	if (cl->cl_qtype == Q_RED)
663 		return red_addq(cl->cl_red, &b->queue, m, cl->cl_pktattr);
664 #endif
665 #ifdef ALTQ_CODEL
666 	if (cl->cl_qtype == Q_CODEL)
667 		return codel_addq(cl->cl_codel, &b->queue, m);
668 #endif
669 	if (qlen(&b->queue) >= qlimit(&b->queue)) {
670 		m_freem(m);
671 		return (-1);
672 	}
673 
674 	if (cl->cl_flags & FARF_CLEARDSCP)
675 		write_dsfield(m, cl->cl_pktattr, 0);
676 
677 	_addq(&b->queue, m);
678 
679 	return (0);
680 }
681 
682 static struct mbuf *
683 fairq_getq(struct fairq_class *cl, uint64_t cur_time)
684 {
685 	fairq_bucket_t *b;
686 	struct mbuf *m;
687 
688 	b = fairq_selectq(cl, 0);
689 	if (b == NULL)
690 		m = NULL;
691 #ifdef ALTQ_RIO
692 	else if (cl->cl_qtype == Q_RIO)
693 		m = rio_getq((rio_t *)cl->cl_red, &b->queue);
694 #endif
695 #ifdef ALTQ_RED
696 	else if (cl->cl_qtype == Q_RED)
697 		m = red_getq(cl->cl_red, &b->queue);
698 #endif
699 #ifdef ALTQ_CODEL
700 	else if (cl->cl_qtype == Q_CODEL)
701 		m = codel_getq(cl->cl_codel, &b->queue);
702 #endif
703 	else
704 		m = _getq(&b->queue);
705 
706 	/*
707 	 * Calculate the BW change
708 	 */
709 	if (m != NULL) {
710 		uint64_t delta;
711 
712 		/*
713 		 * Per-class bandwidth calculation
714 		 */
715 		delta = (cur_time - cl->cl_last_time);
716 		if (delta > machclk_freq * 8)
717 			delta = machclk_freq * 8;
718 		cl->cl_bw_delta += delta;
719 		cl->cl_bw_bytes += m->m_pkthdr.len;
720 		cl->cl_last_time = cur_time;
721 		cl->cl_bw_delta -= cl->cl_bw_delta >> 3;
722 		cl->cl_bw_bytes -= cl->cl_bw_bytes >> 3;
723 
724 		/*
725 		 * Per-bucket bandwidth calculation
726 		 */
727 		delta = (cur_time - b->last_time);
728 		if (delta > machclk_freq * 8)
729 			delta = machclk_freq * 8;
730 		b->bw_delta += delta;
731 		b->bw_bytes += m->m_pkthdr.len;
732 		b->last_time = cur_time;
733 		b->bw_delta -= b->bw_delta >> 3;
734 		b->bw_bytes -= b->bw_bytes >> 3;
735 	}
736 	return(m);
737 }
738 
739 /*
740  * Figure out what the next packet would be if there were no limits.  If
741  * this class hits its bandwidth limit *hit_limit is set to no-zero, otherwise
742  * it is set to 0.  A non-NULL mbuf is returned either way.
743  */
744 static struct mbuf *
745 fairq_pollq(struct fairq_class *cl, uint64_t cur_time, int *hit_limit)
746 {
747 	fairq_bucket_t *b;
748 	struct mbuf *m;
749 	uint64_t delta;
750 	uint64_t bw;
751 
752 	*hit_limit = 0;
753 	b = fairq_selectq(cl, 1);
754 	if (b == NULL)
755 		return(NULL);
756 	m = qhead(&b->queue);
757 
758 	/*
759 	 * Did this packet exceed the class bandwidth?  Calculate the
760 	 * bandwidth component of the packet.
761 	 *
762 	 * - Calculate bytes per second
763 	 */
764 	delta = cur_time - cl->cl_last_time;
765 	if (delta > machclk_freq * 8)
766 		delta = machclk_freq * 8;
767 	cl->cl_bw_delta += delta;
768 	cl->cl_last_time = cur_time;
769 	if (cl->cl_bw_delta) {
770 		bw = cl->cl_bw_bytes * machclk_freq / cl->cl_bw_delta;
771 
772 		if (bw > cl->cl_bandwidth)
773 			*hit_limit = 1;
774 #ifdef ALTQ_DEBUG
775 		printf("BW %6ju relative to %6u %d queue %p\n",
776 			(uintmax_t)bw, cl->cl_bandwidth, *hit_limit, b);
777 #endif
778 	}
779 	return(m);
780 }
781 
782 /*
783  * Locate the next queue we want to pull a packet out of.  This code
784  * is also responsible for removing empty buckets from the circular list.
785  */
786 static
787 fairq_bucket_t *
788 fairq_selectq(struct fairq_class *cl, int ispoll)
789 {
790 	fairq_bucket_t *b;
791 	uint64_t bw;
792 
793 	if (ispoll == 0 && cl->cl_polled) {
794 		b = cl->cl_polled;
795 		cl->cl_polled = NULL;
796 		return(b);
797 	}
798 
799 	while ((b = cl->cl_head) != NULL) {
800 		/*
801 		 * Remove empty queues from consideration
802 		 */
803 		if (qempty(&b->queue)) {
804 			b->in_use = 0;
805 			cl->cl_head = b->next;
806 			if (cl->cl_head == b) {
807 				cl->cl_head = NULL;
808 			} else {
809 				b->next->prev = b->prev;
810 				b->prev->next = b->next;
811 			}
812 			continue;
813 		}
814 
815 		/*
816 		 * Advance the round robin.  Queues with bandwidths less
817 		 * then the hog bandwidth are allowed to burst.
818 		 */
819 		if (cl->cl_hogs_m1 == 0) {
820 			cl->cl_head = b->next;
821 		} else if (b->bw_delta) {
822 			bw = b->bw_bytes * machclk_freq / b->bw_delta;
823 			if (bw >= cl->cl_hogs_m1) {
824 				cl->cl_head = b->next;
825 			}
826 			/*
827 			 * XXX TODO -
828 			 */
829 		}
830 
831 		/*
832 		 * Return bucket b.
833 		 */
834 		break;
835 	}
836 	if (ispoll)
837 		cl->cl_polled = b;
838 	return(b);
839 }
840 
841 static void
842 fairq_purgeq(struct fairq_class *cl)
843 {
844 	fairq_bucket_t *b;
845 	struct mbuf *m;
846 
847 	while ((b = fairq_selectq(cl, 0)) != NULL) {
848 		while ((m = _getq(&b->queue)) != NULL) {
849 			PKTCNTR_ADD(&cl->cl_dropcnt, m_pktlen(m));
850 			m_freem(m);
851 		}
852 		ASSERT(qlen(&b->queue) == 0);
853 	}
854 }
855 
856 static void
857 get_class_stats(struct fairq_classstats *sp, struct fairq_class *cl)
858 {
859 	fairq_bucket_t *b;
860 
861 	sp->class_handle = cl->cl_handle;
862 	sp->qlimit = cl->cl_qlimit;
863 	sp->xmit_cnt = cl->cl_xmitcnt;
864 	sp->drop_cnt = cl->cl_dropcnt;
865 	sp->qtype = cl->cl_qtype;
866 	sp->qlength = 0;
867 
868 	if (cl->cl_head) {
869 		b = cl->cl_head;
870 		do {
871 			sp->qlength += qlen(&b->queue);
872 			b = b->next;
873 		} while (b != cl->cl_head);
874 	}
875 
876 #ifdef ALTQ_RED
877 	if (cl->cl_qtype == Q_RED)
878 		red_getstats(cl->cl_red, &sp->red[0]);
879 #endif
880 #ifdef ALTQ_RIO
881 	if (cl->cl_qtype == Q_RIO)
882 		rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
883 #endif
884 #ifdef ALTQ_CODEL
885 	if (cl->cl_qtype == Q_CODEL)
886 		codel_getstats(cl->cl_codel, &sp->codel);
887 #endif
888 }
889 
890 /* convert a class handle to the corresponding class pointer */
891 static struct fairq_class *
892 clh_to_clp(struct fairq_if *pif, uint32_t chandle)
893 {
894 	struct fairq_class *cl;
895 	int idx;
896 
897 	if (chandle == 0)
898 		return (NULL);
899 
900 	for (idx = pif->pif_maxpri; idx >= 0; idx--)
901 		if ((cl = pif->pif_classes[idx]) != NULL &&
902 		    cl->cl_handle == chandle)
903 			return (cl);
904 
905 	return (NULL);
906 }
907 
908 #endif /* ALTQ_FAIRQ */
909