xref: /freebsd-13-stable/sys/net/altq/altq_fairq.c (revision 4b40a16f0d188422227478889b38cc341d50f88f)
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  */
36 /*
37  * Matt: I gutted altq_priq.c and used it as a skeleton on which to build
38  * fairq.  The fairq algorithm is completely different then priq, of course,
39  * but because I used priq's skeleton I believe I should include priq's
40  * copyright.
41  *
42  * Copyright (C) 2000-2003
43  *	Sony Computer Science Laboratories Inc.  All rights reserved.
44  *
45  * Redistribution and use in source and binary forms, with or without
46  * modification, are permitted provided that the following conditions
47  * are met:
48  * 1. Redistributions of source code must retain the above copyright
49  *    notice, this list of conditions and the following disclaimer.
50  * 2. Redistributions in binary form must reproduce the above copyright
51  *    notice, this list of conditions and the following disclaimer in the
52  *    documentation and/or other materials provided with the distribution.
53  *
54  * THIS SOFTWARE IS PROVIDED BY SONY CSL AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL SONY CSL OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  */
66 
67 /*
68  * FAIRQ - take traffic classified by keep state (hashed into
69  * mbuf->m_pkthdr.altq_state_hash) and bucketize it.  Fairly extract
70  * the first packet from each bucket in a round-robin fashion.
71  *
72  * TODO - better overall qlimit support (right now it is per-bucket).
73  *	- NOTE: red etc is per bucket, not overall.
74  *	- better service curve support.
75  *
76  * EXAMPLE:
77  *
78  *  altq on em0 fairq bandwidth 650Kb queue { std, bulk }
79  *  queue std  priority 3 bandwidth 400Kb \
80  *	fairq (buckets 64, default, hogs 1Kb) qlimit 50
81  *  queue bulk priority 2 bandwidth 100Kb \
82  *	fairq (buckets 64, hogs 1Kb) qlimit 50
83  *
84  *  pass out on em0 from any to any keep state queue std
85  *  pass out on em0 inet proto tcp ..... port ... keep state queue bulk
86  */
87 #include "opt_altq.h"
88 #include "opt_inet.h"
89 #include "opt_inet6.h"
90 
91 #ifdef ALTQ_FAIRQ  /* fairq is enabled in the kernel conf */
92 
93 #include <sys/param.h>
94 #include <sys/malloc.h>
95 #include <sys/mbuf.h>
96 #include <sys/socket.h>
97 #include <sys/sockio.h>
98 #include <sys/systm.h>
99 #include <sys/proc.h>
100 #include <sys/errno.h>
101 #include <sys/kernel.h>
102 #include <sys/queue.h>
103 
104 #include <net/if.h>
105 #include <net/if_var.h>
106 #include <netinet/in.h>
107 
108 #include <netpfil/pf/pf.h>
109 #include <netpfil/pf/pf_altq.h>
110 #include <netpfil/pf/pf_mtag.h>
111 #include <net/altq/altq.h>
112 #include <net/altq/altq_fairq.h>
113 
114 /*
115  * function prototypes
116  */
117 static int	fairq_clear_interface(struct fairq_if *);
118 static int	fairq_request(struct ifaltq *, int, void *);
119 static void	fairq_purge(struct fairq_if *);
120 static struct fairq_class *fairq_class_create(struct fairq_if *, int, int, u_int, struct fairq_opts *, int);
121 static int	fairq_class_destroy(struct fairq_class *);
122 static int	fairq_enqueue(struct ifaltq *, struct mbuf *, struct altq_pktattr *);
123 static struct mbuf *fairq_dequeue(struct ifaltq *, int);
124 
125 static int	fairq_addq(struct fairq_class *, struct mbuf *, u_int32_t);
126 static struct mbuf *fairq_getq(struct fairq_class *, uint64_t);
127 static struct mbuf *fairq_pollq(struct fairq_class *, uint64_t, int *);
128 static fairq_bucket_t *fairq_selectq(struct fairq_class *, int);
129 static void	fairq_purgeq(struct fairq_class *);
130 
131 static void	get_class_stats(struct fairq_classstats *, struct fairq_class *);
132 static struct fairq_class *clh_to_clp(struct fairq_if *, uint32_t);
133 
134 int
fairq_pfattach(struct pf_altq * a)135 fairq_pfattach(struct pf_altq *a)
136 {
137 	struct ifnet *ifp;
138 	int error;
139 
140 	if ((ifp = ifunit(a->ifname)) == NULL || a->altq_disc == NULL)
141 		return (EINVAL);
142 
143 	error = altq_attach(&ifp->if_snd, ALTQT_FAIRQ, a->altq_disc,
144 	    fairq_enqueue, fairq_dequeue, fairq_request, NULL, NULL);
145 
146 	return (error);
147 }
148 
149 int
fairq_add_altq(struct ifnet * ifp,struct pf_altq * a)150 fairq_add_altq(struct ifnet *ifp, struct pf_altq *a)
151 {
152 	struct fairq_if *pif;
153 
154 	if (ifp == NULL)
155 		return (EINVAL);
156 	if (!ALTQ_IS_READY(&ifp->if_snd))
157 		return (ENODEV);
158 
159 	pif = malloc(sizeof(struct fairq_if),
160 			M_DEVBUF, M_WAITOK | M_ZERO);
161 	pif->pif_bandwidth = a->ifbandwidth;
162 	pif->pif_maxpri = -1;
163 	pif->pif_ifq = &ifp->if_snd;
164 
165 	/* keep the state in pf_altq */
166 	a->altq_disc = pif;
167 
168 	return (0);
169 }
170 
171 int
fairq_remove_altq(struct pf_altq * a)172 fairq_remove_altq(struct pf_altq *a)
173 {
174 	struct fairq_if *pif;
175 
176 	if ((pif = a->altq_disc) == NULL)
177 		return (EINVAL);
178 	a->altq_disc = NULL;
179 
180 	fairq_clear_interface(pif);
181 
182 	free(pif, M_DEVBUF);
183 	return (0);
184 }
185 
186 int
fairq_add_queue(struct pf_altq * a)187 fairq_add_queue(struct pf_altq *a)
188 {
189 	struct fairq_if *pif;
190 	struct fairq_class *cl;
191 
192 	if ((pif = a->altq_disc) == NULL)
193 		return (EINVAL);
194 
195 	/* check parameters */
196 	if (a->priority >= FAIRQ_MAXPRI)
197 		return (EINVAL);
198 	if (a->qid == 0)
199 		return (EINVAL);
200 	if (pif->pif_classes[a->priority] != NULL)
201 		return (EBUSY);
202 	if (clh_to_clp(pif, a->qid) != NULL)
203 		return (EBUSY);
204 
205 	cl = fairq_class_create(pif, a->priority, a->qlimit, a->bandwidth,
206 			       &a->pq_u.fairq_opts, a->qid);
207 	if (cl == NULL)
208 		return (ENOMEM);
209 
210 	return (0);
211 }
212 
213 int
fairq_remove_queue(struct pf_altq * a)214 fairq_remove_queue(struct pf_altq *a)
215 {
216 	struct fairq_if *pif;
217 	struct fairq_class *cl;
218 
219 	if ((pif = a->altq_disc) == NULL)
220 		return (EINVAL);
221 
222 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
223 		return (EINVAL);
224 
225 	return (fairq_class_destroy(cl));
226 }
227 
228 int
fairq_getqstats(struct pf_altq * a,void * ubuf,int * nbytes,int version)229 fairq_getqstats(struct pf_altq *a, void *ubuf, int *nbytes, int version)
230 {
231 	struct fairq_if *pif;
232 	struct fairq_class *cl;
233 	struct fairq_classstats stats;
234 	int error = 0;
235 
236 	if ((pif = altq_lookup(a->ifname, ALTQT_FAIRQ)) == NULL)
237 		return (EBADF);
238 
239 	if ((cl = clh_to_clp(pif, a->qid)) == NULL)
240 		return (EINVAL);
241 
242 	if (*nbytes < sizeof(stats))
243 		return (EINVAL);
244 
245 	get_class_stats(&stats, cl);
246 
247 	if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
248 		return (error);
249 	*nbytes = sizeof(stats);
250 	return (0);
251 }
252 
253 /*
254  * bring the interface back to the initial state by discarding
255  * all the filters and classes.
256  */
257 static int
fairq_clear_interface(struct fairq_if * pif)258 fairq_clear_interface(struct fairq_if *pif)
259 {
260 	struct fairq_class *cl;
261 	int pri;
262 
263 	/* clear out the classes */
264 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
265 		if ((cl = pif->pif_classes[pri]) != NULL)
266 			fairq_class_destroy(cl);
267 	}
268 
269 	return (0);
270 }
271 
272 static int
fairq_request(struct ifaltq * ifq,int req,void * arg)273 fairq_request(struct ifaltq *ifq, int req, void *arg)
274 {
275 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
276 
277 	IFQ_LOCK_ASSERT(ifq);
278 
279 	switch (req) {
280 	case ALTRQ_PURGE:
281 		fairq_purge(pif);
282 		break;
283 	}
284 	return (0);
285 }
286 
287 /* discard all the queued packets on the interface */
288 static void
fairq_purge(struct fairq_if * pif)289 fairq_purge(struct fairq_if *pif)
290 {
291 	struct fairq_class *cl;
292 	int pri;
293 
294 	for (pri = 0; pri <= pif->pif_maxpri; pri++) {
295 		if ((cl = pif->pif_classes[pri]) != NULL && cl->cl_head)
296 			fairq_purgeq(cl);
297 	}
298 	if (ALTQ_IS_ENABLED(pif->pif_ifq))
299 		pif->pif_ifq->ifq_len = 0;
300 }
301 
302 static struct fairq_class *
fairq_class_create(struct fairq_if * pif,int pri,int qlimit,u_int bandwidth,struct fairq_opts * opts,int qid)303 fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
304 		   u_int bandwidth, struct fairq_opts *opts, int qid)
305 {
306 	struct fairq_class *cl;
307 	int flags = opts->flags;
308 	u_int nbuckets = opts->nbuckets;
309 	int i;
310 
311 #ifndef ALTQ_RED
312 	if (flags & FARF_RED) {
313 #ifdef ALTQ_DEBUG
314 		printf("fairq_class_create: RED not configured for FAIRQ!\n");
315 #endif
316 		return (NULL);
317 	}
318 #endif
319 #ifndef ALTQ_CODEL
320 	if (flags & FARF_CODEL) {
321 #ifdef ALTQ_DEBUG
322 		printf("fairq_class_create: CODEL not configured for FAIRQ!\n");
323 #endif
324 		return (NULL);
325 	}
326 #endif
327 	if (nbuckets == 0)
328 		nbuckets = 256;
329 	if (nbuckets > FAIRQ_MAX_BUCKETS)
330 		nbuckets = FAIRQ_MAX_BUCKETS;
331 	/* enforce power-of-2 size */
332 	while ((nbuckets ^ (nbuckets - 1)) != ((nbuckets << 1) - 1))
333 		++nbuckets;
334 
335 	if ((cl = pif->pif_classes[pri]) != NULL) {
336 		/* modify the class instead of creating a new one */
337 		IFQ_LOCK(cl->cl_pif->pif_ifq);
338 		if (cl->cl_head)
339 			fairq_purgeq(cl);
340 		IFQ_UNLOCK(cl->cl_pif->pif_ifq);
341 #ifdef ALTQ_RIO
342 		if (cl->cl_qtype == Q_RIO)
343 			rio_destroy((rio_t *)cl->cl_red);
344 #endif
345 #ifdef ALTQ_RED
346 		if (cl->cl_qtype == Q_RED)
347 			red_destroy(cl->cl_red);
348 #endif
349 #ifdef ALTQ_CODEL
350 		if (cl->cl_qtype == Q_CODEL)
351 			codel_destroy(cl->cl_codel);
352 #endif
353 	} else {
354 		cl = malloc(sizeof(struct fairq_class),
355 				M_DEVBUF, M_WAITOK | M_ZERO);
356 		cl->cl_nbuckets = nbuckets;
357 		cl->cl_nbucket_mask = nbuckets - 1;
358 
359 		cl->cl_buckets = malloc(
360 			sizeof(struct fairq_bucket) * cl->cl_nbuckets,
361 			M_DEVBUF, M_WAITOK | M_ZERO);
362 		cl->cl_head = NULL;
363 	}
364 
365 	pif->pif_classes[pri] = cl;
366 	if (flags & FARF_DEFAULTCLASS)
367 		pif->pif_default = cl;
368 	if (qlimit == 0)
369 		qlimit = 50;  /* use default */
370 	cl->cl_qlimit = qlimit;
371 	for (i = 0; i < cl->cl_nbuckets; ++i) {
372 		qlimit(&cl->cl_buckets[i].queue) = qlimit;
373 	}
374 	cl->cl_bandwidth = bandwidth / 8;
375 	cl->cl_qtype = Q_DROPTAIL;
376 	cl->cl_flags = flags & FARF_USERFLAGS;
377 	cl->cl_pri = pri;
378 	if (pri > pif->pif_maxpri)
379 		pif->pif_maxpri = pri;
380 	cl->cl_pif = pif;
381 	cl->cl_handle = qid;
382 	cl->cl_hogs_m1 = opts->hogs_m1 / 8;
383 	cl->cl_lssc_m1 = opts->lssc_m1 / 8;	/* NOT YET USED */
384 
385 #ifdef ALTQ_RED
386 	if (flags & (FARF_RED|FARF_RIO)) {
387 		int red_flags, red_pkttime;
388 
389 		red_flags = 0;
390 		if (flags & FARF_ECN)
391 			red_flags |= REDF_ECN;
392 #ifdef ALTQ_RIO
393 		if (flags & FARF_CLEARDSCP)
394 			red_flags |= RIOF_CLEARDSCP;
395 #endif
396 		if (pif->pif_bandwidth < 8)
397 			red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
398 		else
399 			red_pkttime = (int64_t)pif->pif_ifq->altq_ifp->if_mtu
400 			  * 1000 * 1000 * 1000 / (pif->pif_bandwidth / 8);
401 #ifdef ALTQ_RIO
402 		if (flags & FARF_RIO) {
403 			cl->cl_red = (red_t *)rio_alloc(0, NULL,
404 						red_flags, red_pkttime);
405 			if (cl->cl_red != NULL)
406 				cl->cl_qtype = Q_RIO;
407 		} else
408 #endif
409 		if (flags & FARF_RED) {
410 			cl->cl_red = red_alloc(0, 0,
411 			    cl->cl_qlimit * 10/100,
412 			    cl->cl_qlimit * 30/100,
413 			    red_flags, red_pkttime);
414 			if (cl->cl_red != NULL)
415 				cl->cl_qtype = Q_RED;
416 		}
417 	}
418 #endif /* ALTQ_RED */
419 #ifdef ALTQ_CODEL
420 	if (flags & FARF_CODEL) {
421 		cl->cl_codel = codel_alloc(5, 100, 0);
422 		if (cl->cl_codel != NULL)
423 			cl->cl_qtype = Q_CODEL;
424 	}
425 #endif
426 
427 	return (cl);
428 }
429 
430 static int
fairq_class_destroy(struct fairq_class * cl)431 fairq_class_destroy(struct fairq_class *cl)
432 {
433 	struct fairq_if *pif;
434 	int pri;
435 
436 	IFQ_LOCK(cl->cl_pif->pif_ifq);
437 
438 	if (cl->cl_head)
439 		fairq_purgeq(cl);
440 
441 	pif = cl->cl_pif;
442 	pif->pif_classes[cl->cl_pri] = NULL;
443 	if (pif->pif_poll_cache == cl)
444 		pif->pif_poll_cache = NULL;
445 	if (pif->pif_maxpri == cl->cl_pri) {
446 		for (pri = cl->cl_pri; pri >= 0; pri--)
447 			if (pif->pif_classes[pri] != NULL) {
448 				pif->pif_maxpri = pri;
449 				break;
450 			}
451 		if (pri < 0)
452 			pif->pif_maxpri = -1;
453 	}
454 	IFQ_UNLOCK(cl->cl_pif->pif_ifq);
455 
456 	if (cl->cl_red != NULL) {
457 #ifdef ALTQ_RIO
458 		if (cl->cl_qtype == Q_RIO)
459 			rio_destroy((rio_t *)cl->cl_red);
460 #endif
461 #ifdef ALTQ_RED
462 		if (cl->cl_qtype == Q_RED)
463 			red_destroy(cl->cl_red);
464 #endif
465 #ifdef ALTQ_CODEL
466 		if (cl->cl_qtype == Q_CODEL)
467 			codel_destroy(cl->cl_codel);
468 #endif
469 	}
470 	free(cl->cl_buckets, M_DEVBUF);
471 	free(cl, M_DEVBUF);
472 
473 	return (0);
474 }
475 
476 /*
477  * fairq_enqueue is an enqueue function to be registered to
478  * (*altq_enqueue) in struct ifaltq.
479  */
480 static int
fairq_enqueue(struct ifaltq * ifq,struct mbuf * m,struct altq_pktattr * pktattr)481 fairq_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
482 {
483 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
484 	struct fairq_class *cl = NULL; /* Make compiler happy */
485 	struct pf_mtag *t;
486 	u_int32_t qid_hash = 0;
487 	int len;
488 
489 	IFQ_LOCK_ASSERT(ifq);
490 
491 	/* grab class set by classifier */
492 	if ((m->m_flags & M_PKTHDR) == 0) {
493 		/* should not happen */
494 		printf("altq: packet for %s does not have pkthdr\n",
495 			ifq->altq_ifp->if_xname);
496 		m_freem(m);
497 		return (ENOBUFS);
498 	}
499 
500 	if ((t = pf_find_mtag(m)) != NULL) {
501 		cl = clh_to_clp(pif, t->qid);
502 		qid_hash = t->qid_hash;
503 	}
504 	if (cl == NULL) {
505 		cl = pif->pif_default;
506 		if (cl == NULL) {
507 			m_freem(m);
508 			return (ENOBUFS);
509 		}
510 	}
511 	cl->cl_flags |= FARF_HAS_PACKETS;
512 	cl->cl_pktattr = NULL;
513 	len = m_pktlen(m);
514 	if (fairq_addq(cl, m, qid_hash) != 0) {
515 		/* drop occurred.  mbuf was freed in fairq_addq. */
516 		PKTCNTR_ADD(&cl->cl_dropcnt, len);
517 		return (ENOBUFS);
518 	}
519 	IFQ_INC_LEN(ifq);
520 
521 	return (0);
522 }
523 
524 /*
525  * fairq_dequeue is a dequeue function to be registered to
526  * (*altq_dequeue) in struct ifaltq.
527  *
528  * note: ALTDQ_POLL returns the next packet without removing the packet
529  *	from the queue.  ALTDQ_REMOVE is a normal dequeue operation.
530  *	ALTDQ_REMOVE must return the same packet if called immediately
531  *	after ALTDQ_POLL.
532  */
533 static struct mbuf *
fairq_dequeue(struct ifaltq * ifq,int op)534 fairq_dequeue(struct ifaltq *ifq, int op)
535 {
536 	struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
537 	struct fairq_class *cl;
538 	struct fairq_class *best_cl;
539 	struct mbuf *best_m;
540 	struct mbuf *m = NULL;
541 	uint64_t cur_time = read_machclk();
542 	int pri;
543 	int hit_limit;
544 
545 	IFQ_LOCK_ASSERT(ifq);
546 
547 	if (IFQ_IS_EMPTY(ifq)) {
548 		return (NULL);
549 	}
550 
551 	if (pif->pif_poll_cache && op == ALTDQ_REMOVE) {
552 		best_cl = pif->pif_poll_cache;
553 		m = fairq_getq(best_cl, cur_time);
554 		pif->pif_poll_cache = NULL;
555 		if (m) {
556 			IFQ_DEC_LEN(ifq);
557 			PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
558 			return (m);
559 		}
560 	} else {
561 		best_cl = NULL;
562 		best_m = NULL;
563 
564 		for (pri = pif->pif_maxpri;  pri >= 0; pri--) {
565 			if ((cl = pif->pif_classes[pri]) == NULL)
566 				continue;
567 			if ((cl->cl_flags & FARF_HAS_PACKETS) == 0)
568 				continue;
569 			m = fairq_pollq(cl, cur_time, &hit_limit);
570 			if (m == NULL) {
571 				cl->cl_flags &= ~FARF_HAS_PACKETS;
572 				continue;
573 			}
574 
575 			/*
576 			 * Only override the best choice if we are under
577 			 * the BW limit.
578 			 */
579 			if (hit_limit == 0 || best_cl == NULL) {
580 				best_cl = cl;
581 				best_m = m;
582 			}
583 
584 			/*
585 			 * Remember the highest priority mbuf in case we
586 			 * do not find any lower priority mbufs.
587 			 */
588 			if (hit_limit)
589 				continue;
590 			break;
591 		}
592 		if (op == ALTDQ_POLL) {
593 			pif->pif_poll_cache = best_cl;
594 			m = best_m;
595 		} else if (best_cl) {
596 			m = fairq_getq(best_cl, cur_time);
597 			if (m != NULL) {
598 				IFQ_DEC_LEN(ifq);
599 				PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
600 			}
601 		}
602 		return (m);
603 	}
604 	return (NULL);
605 }
606 
607 static int
fairq_addq(struct fairq_class * cl,struct mbuf * m,u_int32_t bucketid)608 fairq_addq(struct fairq_class *cl, struct mbuf *m, u_int32_t bucketid)
609 {
610 	fairq_bucket_t *b;
611 	u_int hindex;
612 	uint64_t bw;
613 
614 	/*
615 	 * If the packet doesn't have any keep state put it on the end of
616 	 * our queue.  XXX this can result in out of order delivery.
617 	 */
618 	if (bucketid == 0) {
619 		if (cl->cl_head)
620 			b = cl->cl_head->prev;
621 		else
622 			b = &cl->cl_buckets[0];
623 	} else {
624 		hindex = bucketid & cl->cl_nbucket_mask;
625 		b = &cl->cl_buckets[hindex];
626 	}
627 
628 	/*
629 	 * Add the bucket to the end of the circular list of active buckets.
630 	 *
631 	 * As a special case we add the bucket to the beginning of the list
632 	 * instead of the end if it was not previously on the list and if
633 	 * its traffic is less then the hog level.
634 	 */
635 	if (b->in_use == 0) {
636 		b->in_use = 1;
637 		if (cl->cl_head == NULL) {
638 			cl->cl_head = b;
639 			b->next = b;
640 			b->prev = b;
641 		} else {
642 			b->next = cl->cl_head;
643 			b->prev = cl->cl_head->prev;
644 			b->prev->next = b;
645 			b->next->prev = b;
646 
647 			if (b->bw_delta && cl->cl_hogs_m1) {
648 				bw = b->bw_bytes * machclk_freq / b->bw_delta;
649 				if (bw < cl->cl_hogs_m1)
650 					cl->cl_head = b;
651 			}
652 		}
653 	}
654 
655 #ifdef ALTQ_RIO
656 	if (cl->cl_qtype == Q_RIO)
657 		return rio_addq((rio_t *)cl->cl_red, &b->queue, m, cl->cl_pktattr);
658 #endif
659 #ifdef ALTQ_RED
660 	if (cl->cl_qtype == Q_RED)
661 		return red_addq(cl->cl_red, &b->queue, m, cl->cl_pktattr);
662 #endif
663 #ifdef ALTQ_CODEL
664 	if (cl->cl_qtype == Q_CODEL)
665 		return codel_addq(cl->cl_codel, &b->queue, m);
666 #endif
667 	if (qlen(&b->queue) >= qlimit(&b->queue)) {
668 		m_freem(m);
669 		return (-1);
670 	}
671 
672 	if (cl->cl_flags & FARF_CLEARDSCP)
673 		write_dsfield(m, cl->cl_pktattr, 0);
674 
675 	_addq(&b->queue, m);
676 
677 	return (0);
678 }
679 
680 static struct mbuf *
fairq_getq(struct fairq_class * cl,uint64_t cur_time)681 fairq_getq(struct fairq_class *cl, uint64_t cur_time)
682 {
683 	fairq_bucket_t *b;
684 	struct mbuf *m;
685 
686 	b = fairq_selectq(cl, 0);
687 	if (b == NULL)
688 		m = NULL;
689 #ifdef ALTQ_RIO
690 	else if (cl->cl_qtype == Q_RIO)
691 		m = rio_getq((rio_t *)cl->cl_red, &b->queue);
692 #endif
693 #ifdef ALTQ_RED
694 	else if (cl->cl_qtype == Q_RED)
695 		m = red_getq(cl->cl_red, &b->queue);
696 #endif
697 #ifdef ALTQ_CODEL
698 	else if (cl->cl_qtype == Q_CODEL)
699 		m = codel_getq(cl->cl_codel, &b->queue);
700 #endif
701 	else
702 		m = _getq(&b->queue);
703 
704 	/*
705 	 * Calculate the BW change
706 	 */
707 	if (m != NULL) {
708 		uint64_t delta;
709 
710 		/*
711 		 * Per-class bandwidth calculation
712 		 */
713 		delta = (cur_time - cl->cl_last_time);
714 		if (delta > machclk_freq * 8)
715 			delta = machclk_freq * 8;
716 		cl->cl_bw_delta += delta;
717 		cl->cl_bw_bytes += m->m_pkthdr.len;
718 		cl->cl_last_time = cur_time;
719 		cl->cl_bw_delta -= cl->cl_bw_delta >> 3;
720 		cl->cl_bw_bytes -= cl->cl_bw_bytes >> 3;
721 
722 		/*
723 		 * Per-bucket bandwidth calculation
724 		 */
725 		delta = (cur_time - b->last_time);
726 		if (delta > machclk_freq * 8)
727 			delta = machclk_freq * 8;
728 		b->bw_delta += delta;
729 		b->bw_bytes += m->m_pkthdr.len;
730 		b->last_time = cur_time;
731 		b->bw_delta -= b->bw_delta >> 3;
732 		b->bw_bytes -= b->bw_bytes >> 3;
733 	}
734 	return(m);
735 }
736 
737 /*
738  * Figure out what the next packet would be if there were no limits.  If
739  * this class hits its bandwidth limit *hit_limit is set to no-zero, otherwise
740  * it is set to 0.  A non-NULL mbuf is returned either way.
741  */
742 static struct mbuf *
fairq_pollq(struct fairq_class * cl,uint64_t cur_time,int * hit_limit)743 fairq_pollq(struct fairq_class *cl, uint64_t cur_time, int *hit_limit)
744 {
745 	fairq_bucket_t *b;
746 	struct mbuf *m;
747 	uint64_t delta;
748 	uint64_t bw;
749 
750 	*hit_limit = 0;
751 	b = fairq_selectq(cl, 1);
752 	if (b == NULL)
753 		return(NULL);
754 	m = qhead(&b->queue);
755 
756 	/*
757 	 * Did this packet exceed the class bandwidth?  Calculate the
758 	 * bandwidth component of the packet.
759 	 *
760 	 * - Calculate bytes per second
761 	 */
762 	delta = cur_time - cl->cl_last_time;
763 	if (delta > machclk_freq * 8)
764 		delta = machclk_freq * 8;
765 	cl->cl_bw_delta += delta;
766 	cl->cl_last_time = cur_time;
767 	if (cl->cl_bw_delta) {
768 		bw = cl->cl_bw_bytes * machclk_freq / cl->cl_bw_delta;
769 
770 		if (bw > cl->cl_bandwidth)
771 			*hit_limit = 1;
772 #ifdef ALTQ_DEBUG
773 		printf("BW %6ju relative to %6u %d queue %p\n",
774 			(uintmax_t)bw, cl->cl_bandwidth, *hit_limit, b);
775 #endif
776 	}
777 	return(m);
778 }
779 
780 /*
781  * Locate the next queue we want to pull a packet out of.  This code
782  * is also responsible for removing empty buckets from the circular list.
783  */
784 static
785 fairq_bucket_t *
fairq_selectq(struct fairq_class * cl,int ispoll)786 fairq_selectq(struct fairq_class *cl, int ispoll)
787 {
788 	fairq_bucket_t *b;
789 	uint64_t bw;
790 
791 	if (ispoll == 0 && cl->cl_polled) {
792 		b = cl->cl_polled;
793 		cl->cl_polled = NULL;
794 		return(b);
795 	}
796 
797 	while ((b = cl->cl_head) != NULL) {
798 		/*
799 		 * Remove empty queues from consideration
800 		 */
801 		if (qempty(&b->queue)) {
802 			b->in_use = 0;
803 			cl->cl_head = b->next;
804 			if (cl->cl_head == b) {
805 				cl->cl_head = NULL;
806 			} else {
807 				b->next->prev = b->prev;
808 				b->prev->next = b->next;
809 			}
810 			continue;
811 		}
812 
813 		/*
814 		 * Advance the round robin.  Queues with bandwidths less
815 		 * then the hog bandwidth are allowed to burst.
816 		 */
817 		if (cl->cl_hogs_m1 == 0) {
818 			cl->cl_head = b->next;
819 		} else if (b->bw_delta) {
820 			bw = b->bw_bytes * machclk_freq / b->bw_delta;
821 			if (bw >= cl->cl_hogs_m1) {
822 				cl->cl_head = b->next;
823 			}
824 			/*
825 			 * XXX TODO -
826 			 */
827 		}
828 
829 		/*
830 		 * Return bucket b.
831 		 */
832 		break;
833 	}
834 	if (ispoll)
835 		cl->cl_polled = b;
836 	return(b);
837 }
838 
839 static void
fairq_purgeq(struct fairq_class * cl)840 fairq_purgeq(struct fairq_class *cl)
841 {
842 	fairq_bucket_t *b;
843 	struct mbuf *m;
844 
845 	while ((b = fairq_selectq(cl, 0)) != NULL) {
846 		while ((m = _getq(&b->queue)) != NULL) {
847 			PKTCNTR_ADD(&cl->cl_dropcnt, m_pktlen(m));
848 			m_freem(m);
849 		}
850 		ASSERT(qlen(&b->queue) == 0);
851 	}
852 }
853 
854 static void
get_class_stats(struct fairq_classstats * sp,struct fairq_class * cl)855 get_class_stats(struct fairq_classstats *sp, struct fairq_class *cl)
856 {
857 	fairq_bucket_t *b;
858 
859 	sp->class_handle = cl->cl_handle;
860 	sp->qlimit = cl->cl_qlimit;
861 	sp->xmit_cnt = cl->cl_xmitcnt;
862 	sp->drop_cnt = cl->cl_dropcnt;
863 	sp->qtype = cl->cl_qtype;
864 	sp->qlength = 0;
865 
866 	if (cl->cl_head) {
867 		b = cl->cl_head;
868 		do {
869 			sp->qlength += qlen(&b->queue);
870 			b = b->next;
871 		} while (b != cl->cl_head);
872 	}
873 
874 #ifdef ALTQ_RED
875 	if (cl->cl_qtype == Q_RED)
876 		red_getstats(cl->cl_red, &sp->red[0]);
877 #endif
878 #ifdef ALTQ_RIO
879 	if (cl->cl_qtype == Q_RIO)
880 		rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
881 #endif
882 #ifdef ALTQ_CODEL
883 	if (cl->cl_qtype == Q_CODEL)
884 		codel_getstats(cl->cl_codel, &sp->codel);
885 #endif
886 }
887 
888 /* convert a class handle to the corresponding class pointer */
889 static struct fairq_class *
clh_to_clp(struct fairq_if * pif,uint32_t chandle)890 clh_to_clp(struct fairq_if *pif, uint32_t chandle)
891 {
892 	struct fairq_class *cl;
893 	int idx;
894 
895 	if (chandle == 0)
896 		return (NULL);
897 
898 	for (idx = pif->pif_maxpri; idx >= 0; idx--)
899 		if ((cl = pif->pif_classes[idx]) != NULL &&
900 		    cl->cl_handle == chandle)
901 			return (cl);
902 
903 	return (NULL);
904 }
905 
906 #endif /* ALTQ_FAIRQ */
907