2 * Copyright (c) 2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
34 * $DragonFly: src/sys/net/altq/altq_fairq.c,v 1.2 2008/05/14 11:59:23 sephe Exp $
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
42 * Copyright (C) 2000-2003
43 * Sony Computer Science Laboratories Inc. All rights reserved.
45 * Redistribution and use in source and binary forms, with or without
46 * modification, are permitted provided that the following conditions
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.
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
68 * FAIRQ - take traffic classified by keep state (hashed into
69 * pf->state_hash) and bucketize it. Fairly extract
70 * the first packet from each bucket in a round-robin fashion.
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.
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
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
89 #include "opt_inet6.h"
91 #ifdef ALTQ_FAIRQ /* fairq is enabled in the kernel conf */
93 #include <sys/param.h>
94 #include <sys/malloc.h>
96 #include <sys/socket.h>
97 #include <sys/sockio.h>
98 #include <sys/systm.h>
100 #include <sys/errno.h>
101 #include <sys/kernel.h>
102 #include <sys/queue.h>
103 #include <sys/thread.h>
106 #include <net/ifq_var.h>
107 #include <netinet/in.h>
109 #include <net/pf/pfvar.h>
110 #include <net/altq/altq.h>
111 #include <net/altq/altq_fairq.h>
113 #include <sys/thread2.h>
116 * function prototypes
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 *, struct mbuf *, int);
126 static int fairq_addq(struct fairq_class *, struct mbuf *, int hash);
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 *);
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);
136 fairq_pfattach(struct pf_altq *a, struct ifaltq *ifq)
138 return altq_attach(ifq, ALTQT_FAIRQ, a->altq_disc,
139 fairq_enqueue, fairq_dequeue, fairq_request, NULL, NULL);
143 fairq_add_altq(struct pf_altq *a)
145 struct fairq_if *pif;
148 if ((ifp = ifunit(a->ifname)) == NULL)
150 if (!ifq_is_ready(&ifp->if_snd))
153 pif = kmalloc(sizeof(*pif), M_ALTQ, M_WAITOK | M_ZERO);
154 pif->pif_bandwidth = a->ifbandwidth;
155 pif->pif_maxpri = -1;
156 pif->pif_ifq = &ifp->if_snd;
157 ifq_purge(&ifp->if_snd);
159 /* keep the state in pf_altq */
166 fairq_remove_altq(struct pf_altq *a)
168 struct fairq_if *pif;
170 if ((pif = a->altq_disc) == NULL)
174 fairq_clear_interface(pif);
181 fairq_add_queue_locked(struct pf_altq *a, struct fairq_if *pif)
183 struct fairq_class *cl;
185 KKASSERT(a->priority < FAIRQ_MAXPRI);
186 KKASSERT(a->qid != 0);
188 if (pif->pif_classes[a->priority] != NULL)
190 if (clh_to_clp(pif, a->qid) != NULL)
193 cl = fairq_class_create(pif, a->priority, a->qlimit, a->bandwidth,
194 &a->pq_u.fairq_opts, a->qid);
202 fairq_add_queue(struct pf_altq *a)
204 struct fairq_if *pif;
208 /* check parameters */
209 if (a->priority >= FAIRQ_MAXPRI)
214 /* XXX not MP safe */
215 if ((pif = a->altq_disc) == NULL)
220 error = fairq_add_queue_locked(a, pif);
227 fairq_remove_queue_locked(struct pf_altq *a, struct fairq_if *pif)
229 struct fairq_class *cl;
231 if ((cl = clh_to_clp(pif, a->qid)) == NULL)
234 return (fairq_class_destroy(cl));
238 fairq_remove_queue(struct pf_altq *a)
240 struct fairq_if *pif;
244 /* XXX not MP safe */
245 if ((pif = a->altq_disc) == NULL)
250 error = fairq_remove_queue_locked(a, pif);
257 fairq_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
259 struct fairq_if *pif;
260 struct fairq_class *cl;
261 struct fairq_classstats stats;
265 if (*nbytes < sizeof(stats))
268 /* XXX not MP safe */
269 if ((pif = altq_lookup(a->ifname, ALTQT_FAIRQ)) == NULL)
275 if ((cl = clh_to_clp(pif, a->qid)) == NULL) {
280 get_class_stats(&stats, cl);
284 if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
286 *nbytes = sizeof(stats);
291 * bring the interface back to the initial state by discarding
292 * all the filters and classes.
295 fairq_clear_interface(struct fairq_if *pif)
297 struct fairq_class *cl;
300 /* clear out the classes */
301 for (pri = 0; pri <= pif->pif_maxpri; pri++) {
302 if ((cl = pif->pif_classes[pri]) != NULL)
303 fairq_class_destroy(cl);
310 fairq_request(struct ifaltq *ifq, int req, void *arg)
312 struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
324 /* discard all the queued packets on the interface */
326 fairq_purge(struct fairq_if *pif)
328 struct fairq_class *cl;
331 for (pri = 0; pri <= pif->pif_maxpri; pri++) {
332 if ((cl = pif->pif_classes[pri]) != NULL && cl->cl_head)
335 if (ifq_is_enabled(pif->pif_ifq))
336 pif->pif_ifq->ifq_len = 0;
339 static struct fairq_class *
340 fairq_class_create(struct fairq_if *pif, int pri, int qlimit,
341 u_int bandwidth, struct fairq_opts *opts, int qid)
343 struct fairq_class *cl;
344 int flags = opts->flags;
345 u_int nbuckets = opts->nbuckets;
349 if (flags & FARF_RED) {
351 kprintf("fairq_class_create: RED not configured for FAIRQ!\n");
358 if (nbuckets > FAIRQ_MAX_BUCKETS)
359 nbuckets = FAIRQ_MAX_BUCKETS;
360 /* enforce power-of-2 size */
361 while ((nbuckets ^ (nbuckets - 1)) != ((nbuckets << 1) - 1))
364 if ((cl = pif->pif_classes[pri]) != NULL) {
365 /* modify the class instead of creating a new one */
371 if (cl->cl_qtype == Q_RIO)
372 rio_destroy((rio_t *)cl->cl_red);
375 if (cl->cl_qtype == Q_RED)
376 red_destroy(cl->cl_red);
379 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
380 cl->cl_nbuckets = nbuckets;
381 cl->cl_nbucket_mask = nbuckets - 1;
383 cl->cl_buckets = kmalloc(sizeof(*cl->cl_buckets) *
385 M_ALTQ, M_WAITOK | M_ZERO);
389 pif->pif_classes[pri] = cl;
390 if (flags & FARF_DEFAULTCLASS)
391 pif->pif_default = cl;
393 qlimit = 50; /* use default */
394 cl->cl_qlimit = qlimit;
395 for (i = 0; i < cl->cl_nbuckets; ++i) {
396 qlimit(&cl->cl_buckets[i].queue) = qlimit;
398 cl->cl_bandwidth = bandwidth / 8;
399 cl->cl_qtype = Q_DROPTAIL;
400 cl->cl_flags = flags & FARF_USERFLAGS;
402 if (pri > pif->pif_maxpri)
403 pif->pif_maxpri = pri;
406 cl->cl_hogs_m1 = opts->hogs_m1 / 8;
407 cl->cl_lssc_m1 = opts->lssc_m1 / 8; /* NOT YET USED */
410 if (flags & (FARF_RED|FARF_RIO)) {
411 int red_flags, red_pkttime;
414 if (flags & FARF_ECN)
415 red_flags |= REDF_ECN;
417 if (flags & FARF_CLEARDSCP)
418 red_flags |= RIOF_CLEARDSCP;
420 if (pif->pif_bandwidth < 8)
421 red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
423 red_pkttime = (int64_t)pif->pif_ifq->altq_ifp->if_mtu
424 * 1000 * 1000 * 1000 / (pif->pif_bandwidth / 8);
426 if (flags & FARF_RIO) {
427 cl->cl_red = (red_t *)rio_alloc(0, NULL,
428 red_flags, red_pkttime);
429 if (cl->cl_red != NULL)
430 cl->cl_qtype = Q_RIO;
433 if (flags & FARF_RED) {
434 cl->cl_red = red_alloc(0, 0,
435 cl->cl_qlimit * 10/100,
436 cl->cl_qlimit * 30/100,
437 red_flags, red_pkttime);
438 if (cl->cl_red != NULL)
439 cl->cl_qtype = Q_RED;
442 #endif /* ALTQ_RED */
448 fairq_class_destroy(struct fairq_class *cl)
450 struct fairq_if *pif;
459 pif->pif_classes[cl->cl_pri] = NULL;
460 if (pif->pif_poll_cache == cl)
461 pif->pif_poll_cache = NULL;
462 if (pif->pif_maxpri == cl->cl_pri) {
463 for (pri = cl->cl_pri; pri >= 0; pri--)
464 if (pif->pif_classes[pri] != NULL) {
465 pif->pif_maxpri = pri;
469 pif->pif_maxpri = -1;
473 if (cl->cl_red != NULL) {
475 if (cl->cl_qtype == Q_RIO)
476 rio_destroy((rio_t *)cl->cl_red);
479 if (cl->cl_qtype == Q_RED)
480 red_destroy(cl->cl_red);
483 kfree(cl->cl_buckets, M_ALTQ);
484 cl->cl_head = NULL; /* sanity */
485 cl->cl_polled = NULL; /* sanity */
486 cl->cl_buckets = NULL; /* sanity */
493 * fairq_enqueue is an enqueue function to be registered to
494 * (*altq_enqueue) in struct ifaltq.
497 fairq_enqueue(struct ifaltq *ifq, struct mbuf *m, struct altq_pktattr *pktattr)
499 struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
500 struct fairq_class *cl;
507 /* grab class set by classifier */
508 if ((m->m_flags & M_PKTHDR) == 0) {
509 /* should not happen */
510 if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
516 if (m->m_pkthdr.fw_flags & PF_MBUF_STRUCTURE) {
517 cl = clh_to_clp(pif, m->m_pkthdr.pf.qid);
518 if (m->m_pkthdr.pf.flags & PF_TAG_STATE_HASHED)
519 hash = (int)m->m_pkthdr.pf.state_hash;
527 cl = pif->pif_default;
534 cl->cl_flags |= FARF_HAS_PACKETS;
535 cl->cl_pktattr = NULL;
537 if (fairq_addq(cl, m, hash) != 0) {
538 /* drop occurred. mbuf was freed in fairq_addq. */
539 PKTCNTR_ADD(&cl->cl_dropcnt, len);
551 * fairq_dequeue is a dequeue function to be registered to
552 * (*altq_dequeue) in struct ifaltq.
554 * note: ALTDQ_POLL returns the next packet without removing the packet
555 * from the queue. ALTDQ_REMOVE is a normal dequeue operation.
556 * ALTDQ_REMOVE must return the same packet if called immediately
560 fairq_dequeue(struct ifaltq *ifq, struct mbuf *mpolled, int op)
562 struct fairq_if *pif = (struct fairq_if *)ifq->altq_disc;
563 struct fairq_class *cl;
564 struct fairq_class *best_cl;
567 uint64_t cur_time = read_machclk();
571 if (ifq_is_empty(ifq)) {
572 /* no packet in the queue */
573 KKASSERT(mpolled == NULL);
578 if (pif->pif_poll_cache && op == ALTDQ_REMOVE) {
579 best_cl = pif->pif_poll_cache;
580 m = fairq_getq(best_cl, cur_time);
581 pif->pif_poll_cache = NULL;
584 PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
590 for (pri = pif->pif_maxpri; pri >= 0; pri--) {
591 if ((cl = pif->pif_classes[pri]) == NULL)
593 if ((cl->cl_flags & FARF_HAS_PACKETS) == 0)
595 m = fairq_pollq(cl, cur_time, &hit_limit);
597 cl->cl_flags &= ~FARF_HAS_PACKETS;
602 * Only override the best choice if we are under
605 if (hit_limit == 0 || best_cl == NULL) {
611 * Remember the highest priority mbuf in case we
612 * do not find any lower priority mbufs.
618 if (op == ALTDQ_POLL) {
619 pif->pif_poll_cache = best_cl;
621 } else if (best_cl) {
622 m = fairq_getq(best_cl, cur_time);
623 KKASSERT(best_m == m);
625 PKTCNTR_ADD(&best_cl->cl_xmitcnt, m_pktlen(m));
631 KKASSERT(mpolled == NULL || mpolled == m);
636 fairq_addq(struct fairq_class *cl, struct mbuf *m, int hash)
643 * If the packet doesn't have any keep state put it on the end of
644 * our queue. XXX this can result in out of order delivery.
648 b = cl->cl_head->prev;
650 b = &cl->cl_buckets[0];
652 hindex = hash & cl->cl_nbucket_mask;
653 b = &cl->cl_buckets[hindex];
657 * Add the bucket to the end of the circular list of active buckets.
659 * As a special case we add the bucket to the beginning of the list
660 * instead of the end if it was not previously on the list and if
661 * its traffic is less then the hog level.
663 if (b->in_use == 0) {
665 if (cl->cl_head == NULL) {
670 b->next = cl->cl_head;
671 b->prev = cl->cl_head->prev;
675 if (b->bw_delta && cl->cl_hogs_m1) {
676 bw = b->bw_bytes * machclk_freq / b->bw_delta;
677 if (bw < cl->cl_hogs_m1)
684 if (cl->cl_qtype == Q_RIO)
685 return rio_addq((rio_t *)cl->cl_red, &b->queue, m, cl->cl_pktattr);
688 if (cl->cl_qtype == Q_RED)
689 return red_addq(cl->cl_red, &b->queue, m, cl->cl_pktattr);
691 if (qlen(&b->queue) >= qlimit(&b->queue)) {
696 if (cl->cl_flags & FARF_CLEARDSCP)
697 write_dsfield(m, cl->cl_pktattr, 0);
705 fairq_getq(struct fairq_class *cl, uint64_t cur_time)
710 b = fairq_selectq(cl, 0);
714 else if (cl->cl_qtype == Q_RIO)
715 m = rio_getq((rio_t *)cl->cl_red, &b->queue);
718 else if (cl->cl_qtype == Q_RED)
719 m = red_getq(cl->cl_red, &b->queue);
722 m = _getq(&b->queue);
725 * Calculate the BW change
731 * Per-class bandwidth calculation
733 delta = (cur_time - cl->cl_last_time);
734 if (delta > machclk_freq * 8)
735 delta = machclk_freq * 8;
736 cl->cl_bw_delta += delta;
737 cl->cl_bw_bytes += m->m_pkthdr.len;
738 cl->cl_last_time = cur_time;
739 cl->cl_bw_delta -= cl->cl_bw_delta >> 3;
740 cl->cl_bw_bytes -= cl->cl_bw_bytes >> 3;
743 * Per-bucket bandwidth calculation
745 delta = (cur_time - b->last_time);
746 if (delta > machclk_freq * 8)
747 delta = machclk_freq * 8;
748 b->bw_delta += delta;
749 b->bw_bytes += m->m_pkthdr.len;
750 b->last_time = cur_time;
751 b->bw_delta -= b->bw_delta >> 3;
752 b->bw_bytes -= b->bw_bytes >> 3;
758 * Figure out what the next packet would be if there were no limits. If
759 * this class hits its bandwidth limit *hit_limit is set to no-zero, otherwise
760 * it is set to 0. A non-NULL mbuf is returned either way.
763 fairq_pollq(struct fairq_class *cl, uint64_t cur_time, int *hit_limit)
771 b = fairq_selectq(cl, 1);
774 m = qhead(&b->queue);
777 * Did this packet exceed the class bandwidth? Calculate the
778 * bandwidth component of the packet.
780 * - Calculate bytes per second
782 delta = cur_time - cl->cl_last_time;
783 if (delta > machclk_freq * 8)
784 delta = machclk_freq * 8;
785 cl->cl_bw_delta += delta;
786 cl->cl_last_time = cur_time;
787 if (cl->cl_bw_delta) {
788 bw = cl->cl_bw_bytes * machclk_freq / cl->cl_bw_delta;
790 if (bw > cl->cl_bandwidth)
793 kprintf("BW %6lld relative to %6u %d queue %p\n",
794 bw, cl->cl_bandwidth, *hit_limit, b);
801 * Locate the next queue we want to pull a packet out of. This code
802 * is also responsible for removing empty buckets from the circular list.
806 fairq_selectq(struct fairq_class *cl, int ispoll)
811 if (ispoll == 0 && cl->cl_polled) {
813 cl->cl_polled = NULL;
817 while ((b = cl->cl_head) != NULL) {
819 * Remove empty queues from consideration
821 if (qempty(&b->queue)) {
823 cl->cl_head = b->next;
824 if (cl->cl_head == b) {
827 b->next->prev = b->prev;
828 b->prev->next = b->next;
834 * Advance the round robin. Queues with bandwidths less
835 * then the hog bandwidth are allowed to burst.
837 if (cl->cl_hogs_m1 == 0) {
838 cl->cl_head = b->next;
839 } else if (b->bw_delta) {
840 bw = b->bw_bytes * machclk_freq / b->bw_delta;
841 if (bw >= cl->cl_hogs_m1) {
842 cl->cl_head = b->next;
860 fairq_purgeq(struct fairq_class *cl)
865 while ((b = fairq_selectq(cl, 0)) != NULL) {
866 while ((m = _getq(&b->queue)) != NULL) {
867 PKTCNTR_ADD(&cl->cl_dropcnt, m_pktlen(m));
870 KKASSERT(qlen(&b->queue) == 0);
875 get_class_stats(struct fairq_classstats *sp, struct fairq_class *cl)
879 sp->class_handle = cl->cl_handle;
880 sp->qlimit = cl->cl_qlimit;
881 sp->xmit_cnt = cl->cl_xmitcnt;
882 sp->drop_cnt = cl->cl_dropcnt;
883 sp->qtype = cl->cl_qtype;
889 sp->qlength += qlen(&b->queue);
891 } while (b != cl->cl_head);
895 if (cl->cl_qtype == Q_RED)
896 red_getstats(cl->cl_red, &sp->red[0]);
899 if (cl->cl_qtype == Q_RIO)
900 rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
904 /* convert a class handle to the corresponding class pointer */
905 static struct fairq_class *
906 clh_to_clp(struct fairq_if *pif, uint32_t chandle)
908 struct fairq_class *cl;
914 for (idx = pif->pif_maxpri; idx >= 0; idx--)
915 if ((cl = pif->pif_classes[idx]) != NULL &&
916 cl->cl_handle == chandle)
922 #endif /* ALTQ_FAIRQ */