2 * Copyright (c) 1991-1997 Regents of the University of California.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. All advertising materials mentioning features or use of this software
14 * must display the following acknowledgement:
15 * This product includes software developed by the Network Research
16 * Group at Lawrence Berkeley Laboratory.
17 * 4. Neither the name of the University nor of the Laboratory may be used
18 * to endorse or promote products derived from this software without
19 * specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * LBL code modified by speer@eng.sun.com, May 1977.
34 * For questions and/or comments, please send mail to cbq@ee.lbl.gov
35 * $KAME: altq_rmclass.c,v 1.19 2005/04/13 03:44:25 suz Exp $
39 #include "opt_inet6.h"
40 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
42 #include <sys/param.h>
43 #include <sys/malloc.h>
45 #include <sys/socket.h>
46 #include <sys/systm.h>
47 #include <sys/errno.h>
51 #include <net/if_var.h>
52 #include <net/if_private.h>
54 #include <net/altq/if_altq.h>
55 #include <net/altq/altq.h>
56 #include <net/altq/altq_codel.h>
57 #include <net/altq/altq_rmclass.h>
58 #include <net/altq/altq_rmclass_debug.h>
59 #include <net/altq/altq_red.h>
60 #include <net/altq/altq_rio.h>
65 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
71 static int rmc_satisfied(struct rm_class *, struct timeval *);
72 static void rmc_wrr_set_weights(struct rm_ifdat *);
73 static void rmc_depth_compute(struct rm_class *);
74 static void rmc_depth_recompute(rm_class_t *);
76 static mbuf_t *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
77 static mbuf_t *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
79 static int _rmc_addq(rm_class_t *, mbuf_t *);
80 static void _rmc_dropq(rm_class_t *);
81 static mbuf_t *_rmc_getq(rm_class_t *);
82 static mbuf_t *_rmc_pollq(rm_class_t *);
84 static int rmc_under_limit(struct rm_class *, struct timeval *);
85 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
86 static void rmc_drop_action(struct rm_class *);
87 static void rmc_restart(void *);
88 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
90 #define BORROW_OFFTIME
92 * BORROW_OFFTIME (experimental):
93 * borrow the offtime of the class borrowing from.
94 * the reason is that when its own offtime is set, the class is unable
95 * to borrow much, especially when cutoff is taking effect.
96 * but when the borrowed class is overloaded (advidle is close to minidle),
97 * use the borrowing class's offtime to avoid overload.
101 * ADJUST_CUTOFF (experimental):
102 * if no underlimit class is found due to cutoff, increase cutoff and
103 * retry the scheduling loop.
104 * also, don't invoke delay_actions while cutoff is taking effect,
105 * since a sleeping class won't have a chance to be scheduled in the
108 * now heuristics for setting the top-level variable (cutoff_) becomes:
109 * 1. if a packet arrives for a not-overlimit class, set cutoff
110 * to the depth of the class.
111 * 2. if cutoff is i, and a packet arrives for an overlimit class
112 * with an underlimit ancestor at a lower level than i (say j),
113 * then set cutoff to j.
114 * 3. at scheduling a packet, if there is no underlimit class
115 * due to the current cutoff level, increase cutoff by 1 and
116 * then try to schedule again.
121 * rmc_newclass(...) - Create a new resource management class at priority
122 * 'pri' on the interface given by 'ifd'.
124 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
125 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
126 * than 100% of the bandwidth, this number should be the
127 * 'effective' rate for the class. Let f be the
128 * bandwidth fraction allocated to this class, and let
129 * nsPerByte be the data rate of the output link in
130 * nanoseconds/byte. Then nsecPerByte is set to
131 * nsPerByte / f. E.g., 1600 (= 800 / .5)
132 * for a class that gets 50% of an ethernet's bandwidth.
134 * action the routine to call when the class is over limit.
136 * maxq max allowable queue size for class (in packets).
138 * parent parent class pointer.
140 * borrow class to borrow from (should be either 'parent' or null).
142 * maxidle max value allowed for class 'idle' time estimate (this
143 * parameter determines how large an initial burst of packets
144 * can be before overlimit action is invoked.
146 * offtime how long 'delay' action will delay when class goes over
147 * limit (this parameter determines the steady-state burst
148 * size when a class is running over its limit).
150 * Maxidle and offtime have to be computed from the following: If the
151 * average packet size is s, the bandwidth fraction allocated to this
152 * class is f, we want to allow b packet bursts, and the gain of the
153 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
155 * ptime = s * nsPerByte * (1 - f) / f
156 * maxidle = ptime * (1 - g^b) / g^b
157 * minidle = -ptime * (1 / (f - 1))
158 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
160 * Operationally, it's convenient to specify maxidle & offtime in units
161 * independent of the link bandwidth so the maxidle & offtime passed to
162 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
163 * (The constant factor is a scale factor needed to make the parameters
164 * integers. This scaling also means that the 'unscaled' values of
165 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
166 * not nanoseconds.) Also note that the 'idle' filter computation keeps
167 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
168 * maxidle also must be scaled upward by this value. Thus, the passed
169 * values for maxidle and offtime can be computed as follows:
171 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
172 * offtime = offtime * 8 / (1000 * nsecPerByte)
174 * When USE_HRTIME is employed, then maxidle and offtime become:
175 * maxidle = maxilde * (8.0 / nsecPerByte);
176 * offtime = offtime * (8.0 / nsecPerByte);
179 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
180 void (*action)(rm_class_t *, rm_class_t *), int maxq,
181 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
182 int minidle, u_int offtime, int pktsize, int flags)
185 struct rm_class *peer;
188 if (pri >= RM_MAXPRIO)
191 if (flags & RMCF_RED) {
193 printf("rmc_newclass: RED not configured for CBQ!\n");
199 if (flags & RMCF_RIO) {
201 printf("rmc_newclass: RIO not configured for CBQ!\n");
207 if (flags & RMCF_CODEL) {
209 printf("rmc_newclass: CODEL not configured for CBQ!\n");
215 cl = malloc(sizeof(struct rm_class), M_DEVBUF, M_NOWAIT | M_ZERO);
218 CALLOUT_INIT(&cl->callout_);
219 cl->q_ = malloc(sizeof(class_queue_t), M_DEVBUF, M_NOWAIT | M_ZERO);
220 if (cl->q_ == NULL) {
226 * Class initialization.
228 cl->children_ = NULL;
229 cl->parent_ = parent;
230 cl->borrow_ = borrow;
234 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
237 cl->ns_per_byte_ = nsecPerByte;
239 qlimit(cl->q_) = maxq;
240 qtype(cl->q_) = Q_DROPHEAD;
244 #if 1 /* minidle is also scaled in ALTQ */
245 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
246 if (cl->minidle_ > 0)
249 cl->minidle_ = minidle;
251 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
252 if (cl->maxidle_ == 0)
254 #if 1 /* offtime is also scaled in ALTQ */
255 cl->avgidle_ = cl->maxidle_;
256 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
257 if (cl->offtime_ == 0)
261 cl->offtime_ = (offtime * nsecPerByte) / 8;
263 cl->overlimit = action;
266 if (flags & (RMCF_RED|RMCF_RIO)) {
267 int red_flags, red_pkttime;
270 if (flags & RMCF_ECN)
271 red_flags |= REDF_ECN;
272 if (flags & RMCF_FLOWVALVE)
273 red_flags |= REDF_FLOWVALVE;
275 if (flags & RMCF_CLEARDSCP)
276 red_flags |= RIOF_CLEARDSCP;
278 red_pkttime = nsecPerByte * pktsize / 1000;
280 if (flags & RMCF_RED) {
281 cl->red_ = red_alloc(0, 0,
282 qlimit(cl->q_) * 10/100,
283 qlimit(cl->q_) * 30/100,
284 red_flags, red_pkttime);
285 if (cl->red_ != NULL)
286 qtype(cl->q_) = Q_RED;
290 cl->red_ = (red_t *)rio_alloc(0, NULL,
291 red_flags, red_pkttime);
292 if (cl->red_ != NULL)
293 qtype(cl->q_) = Q_RIO;
297 #endif /* ALTQ_RED */
299 if (flags & RMCF_CODEL) {
300 cl->codel_ = codel_alloc(5, 100, 0);
301 if (cl->codel_ != NULL)
302 qtype(cl->q_) = Q_CODEL;
307 * put the class into the class tree
311 if ((peer = ifd->active_[pri]) != NULL) {
312 /* find the last class at this pri */
314 while (peer->peer_ != ifd->active_[pri])
318 ifd->active_[pri] = cl;
323 cl->next_ = parent->children_;
324 parent->children_ = cl;
329 * Compute the depth of this class and its ancestors in the class
332 rmc_depth_compute(cl);
335 * If CBQ's WRR is enabled, then initialize the class WRR state.
339 ifd->alloc_[pri] += cl->allotment_;
340 rmc_wrr_set_weights(ifd);
342 IFQ_UNLOCK(ifd->ifq_);
348 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
349 int minidle, u_int offtime, int pktsize)
351 struct rm_ifdat *ifd;
356 old_allotment = cl->allotment_;
360 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
362 cl->ns_per_byte_ = nsecPerByte;
364 qlimit(cl->q_) = maxq;
366 #if 1 /* minidle is also scaled in ALTQ */
367 cl->minidle_ = (minidle * nsecPerByte) / 8;
368 if (cl->minidle_ > 0)
371 cl->minidle_ = minidle;
373 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
374 if (cl->maxidle_ == 0)
376 #if 1 /* offtime is also scaled in ALTQ */
377 cl->avgidle_ = cl->maxidle_;
378 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
379 if (cl->offtime_ == 0)
383 cl->offtime_ = (offtime * nsecPerByte) / 8;
387 * If CBQ's WRR is enabled, then initialize the class WRR state.
390 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
391 rmc_wrr_set_weights(ifd);
393 IFQ_UNLOCK(ifd->ifq_);
400 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
401 * the appropriate run robin weights for the CBQ weighted round robin
408 rmc_wrr_set_weights(struct rm_ifdat *ifd)
411 struct rm_class *cl, *clh;
413 for (i = 0; i < RM_MAXPRIO; i++) {
415 * This is inverted from that of the simulator to
416 * maintain precision.
418 if (ifd->num_[i] == 0)
421 ifd->M_[i] = ifd->alloc_[i] /
422 (ifd->num_[i] * ifd->maxpkt_);
424 * Compute the weighted allotment for each class.
425 * This takes the expensive div instruction out
426 * of the main loop for the wrr scheduling path.
427 * These only get recomputed when a class comes or
430 if (ifd->active_[i] != NULL) {
431 clh = cl = ifd->active_[i];
433 /* safe-guard for slow link or alloc_ == 0 */
435 cl->w_allotment_ = 0;
437 cl->w_allotment_ = cl->allotment_ /
440 } while ((cl != NULL) && (cl != clh));
446 rmc_get_weight(struct rm_ifdat *ifd, int pri)
448 if ((pri >= 0) && (pri < RM_MAXPRIO))
449 return (ifd->M_[pri]);
456 * rmc_depth_compute(struct rm_class *cl) - This function computes the
457 * appropriate depth of class 'cl' and its ancestors.
463 rmc_depth_compute(struct rm_class *cl)
465 rm_class_t *t = cl, *p;
468 * Recompute the depth for the branch of the tree.
472 if (p && (t->depth_ >= p->depth_)) {
473 p->depth_ = t->depth_ + 1;
482 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
483 * the depth of the tree after a class has been deleted.
489 rmc_depth_recompute(rm_class_t *cl)
496 if ((t = p->children_) == NULL) {
502 if (t->depth_ > cdepth)
507 if (p->depth_ == cdepth + 1)
508 /* no change to this parent */
511 p->depth_ = cdepth + 1;
519 if (cl->depth_ >= 1) {
520 if (cl->children_ == NULL) {
522 } else if ((t = cl->children_) != NULL) {
524 if (t->children_ != NULL)
525 rmc_depth_recompute(t);
529 rmc_depth_compute(cl);
536 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
537 * function deletes a class from the link-sharing structure and frees
538 * all resources associated with the class.
544 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
546 struct rm_class *p, *head, *previous;
549 ASSERT(cl->children_ == NULL);
552 CALLOUT_STOP(&cl->callout_);
557 * Free packets in the packet queue.
558 * XXX - this may not be a desired behavior. Packets should be
564 * If the class has a parent, then remove the class from the
565 * class from the parent's children chain.
567 if (cl->parent_ != NULL) {
568 head = cl->parent_->children_;
570 if (head->next_ == NULL) {
572 cl->parent_->children_ = NULL;
573 cl->parent_->leaf_ = 1;
574 } else while (p != NULL) {
577 cl->parent_->children_ = cl->next_;
579 previous->next_ = cl->next_;
590 * Delete class from class priority peer list.
592 if ((p = ifd->active_[cl->pri_]) != NULL) {
594 * If there is more than one member of this priority
595 * level, then look for class(cl) in the priority level.
598 while (p->peer_ != cl)
600 p->peer_ = cl->peer_;
602 if (ifd->active_[cl->pri_] == cl)
603 ifd->active_[cl->pri_] = cl->peer_;
606 ifd->active_[cl->pri_] = NULL;
611 * Recompute the WRR weights.
614 ifd->alloc_[cl->pri_] -= cl->allotment_;
615 ifd->num_[cl->pri_]--;
616 rmc_wrr_set_weights(ifd);
620 * Re-compute the depth of the tree.
623 rmc_depth_recompute(cl->parent_);
625 rmc_depth_recompute(ifd->root_);
628 IFQ_UNLOCK(ifd->ifq_);
632 * Free the class structure.
634 if (cl->red_ != NULL) {
636 if (q_is_rio(cl->q_))
637 rio_destroy((rio_t *)cl->red_);
640 if (q_is_red(cl->q_))
641 red_destroy(cl->red_);
644 if (q_is_codel(cl->q_))
645 codel_destroy(cl->codel_);
648 free(cl->q_, M_DEVBUF);
654 * rmc_init(...) - Initialize the resource management data structures
655 * associated with the output portion of interface 'ifp'. 'ifd' is
656 * where the structures will be built (for backwards compatibility, the
657 * structures aren't kept in the ifnet struct). 'nsecPerByte'
658 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
659 * 'restart' is the driver-specific routine that the generic 'delay
660 * until under limit' action will call to restart output. `maxq'
661 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
662 * is the maximum number of packets that the resource management
663 * code will allow to be queued 'downstream' (this is typically 1).
669 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
670 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
671 int minidle, u_int offtime, int flags)
676 * Initialize the CBQ tracing/debug facility.
680 bzero((char *)ifd, sizeof (*ifd));
681 mtu = ifq->altq_ifp->if_mtu;
683 ifd->restart = restart;
684 ifd->maxqueued_ = maxqueued;
685 ifd->ns_per_byte_ = nsecPerByte;
687 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
688 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
690 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
691 if (mtu * nsecPerByte > 10 * 1000000)
692 ifd->maxiftime_ /= 4;
696 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
699 * Initialize the CBQ's WRR state.
701 for (i = 0; i < RM_MAXPRIO; i++) {
706 ifd->active_[i] = NULL;
710 * Initialize current packet state.
714 for (i = 0; i < RM_MAXQUEUED; i++) {
715 ifd->class_[i] = NULL;
717 ifd->borrowed_[i] = NULL;
721 * Create the root class of the link-sharing structure.
723 if ((ifd->root_ = rmc_newclass(0, ifd,
725 rmc_root_overlimit, maxq, 0, 0,
726 maxidle, minidle, offtime,
728 printf("rmc_init: root class not allocated\n");
731 ifd->root_->depth_ = 0;
736 * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
737 * mbuf 'm' to queue for resource class 'cl'. This routine is called
738 * by a driver's if_output routine. This routine must be called with
739 * output packet completion interrupts locked out (to avoid racing with
742 * Returns: 0 on successful queueing
743 * -1 when packet drop occurs
746 rmc_queue_packet(struct rm_class *cl, mbuf_t *m)
749 struct rm_ifdat *ifd = cl->ifdat_;
751 int is_empty = qempty(cl->q_);
754 if (ifd->cutoff_ > 0) {
755 if (TV_LT(&cl->undertime_, &now)) {
756 if (ifd->cutoff_ > cl->depth_)
757 ifd->cutoff_ = cl->depth_;
758 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
763 * the class is overlimit. if the class has
764 * underlimit ancestors, set cutoff to the lowest
767 struct rm_class *borrow = cl->borrow_;
769 while (borrow != NULL &&
770 borrow->depth_ < ifd->cutoff_) {
771 if (TV_LT(&borrow->undertime_, &now)) {
772 ifd->cutoff_ = borrow->depth_;
773 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
776 borrow = borrow->borrow_;
780 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
781 if (TV_LT(&cl->borrow_->undertime_, &now)) {
782 ifd->cutoff_ = cl->borrow_->depth_;
783 CBQTRACE(rmc_queue_packet, 'ffob',
784 cl->borrow_->depth_);
790 if (_rmc_addq(cl, m) < 0)
795 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
799 if (qlen(cl->q_) > qlimit(cl->q_)) {
800 /* note: qlimit can be set to 0 or 1 */
809 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
810 * classes to see if there are satified.
814 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
819 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
820 if ((bp = ifd->active_[i]) != NULL) {
823 if (!rmc_satisfied(p, now)) {
824 ifd->cutoff_ = p->depth_;
836 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
840 rmc_satisfied(struct rm_class *cl, struct timeval *now)
846 if (TV_LT(now, &cl->undertime_))
848 if (cl->depth_ == 0) {
849 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
854 if (cl->children_ != NULL) {
857 if (!rmc_satisfied(p, now))
867 * Return 1 if class 'cl' is under limit or can borrow from a parent,
868 * 0 if overlimit. As a side-effect, this routine will invoke the
869 * class overlimit action if the class if overlimit.
873 rmc_under_limit(struct rm_class *cl, struct timeval *now)
877 struct rm_ifdat *ifd = cl->ifdat_;
879 ifd->borrowed_[ifd->qi_] = NULL;
881 * If cl is the root class, then always return that it is
882 * underlimit. Otherwise, check to see if the class is underlimit.
884 if (cl->parent_ == NULL)
888 if (TV_LT(now, &cl->undertime_))
891 CALLOUT_STOP(&cl->callout_);
893 cl->undertime_.tv_sec = 0;
898 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
899 if (((cl = cl->borrow_) == NULL) ||
900 (cl->depth_ > ifd->cutoff_)) {
903 /* cutoff is taking effect, just
904 return false without calling
908 #ifdef BORROW_OFFTIME
910 * check if the class can borrow offtime too.
911 * borrow offtime from the top of the borrow
912 * chain if the top class is not overloaded.
915 /* cutoff is taking effect, use this class as top. */
917 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
919 if (top != NULL && top->avgidle_ == top->minidle_)
922 (p->overlimit)(p, top);
925 (p->overlimit)(p, NULL);
933 ifd->borrowed_[ifd->qi_] = cl;
938 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
939 * Packet-by-packet round robin.
941 * The heart of the weighted round-robin scheduler, which decides which
942 * class next gets to send a packet. Highest priority first, then
943 * weighted round-robin within priorites.
945 * Each able-to-send class gets to send until its byte allocation is
946 * exhausted. Thus, the active pointer is only changed after a class has
947 * exhausted its allocation.
949 * If the scheduler finds no class that is underlimit or able to borrow,
950 * then the first class found that had a nonzero queue and is allowed to
951 * borrow gets to send.
955 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
957 struct rm_class *cl = NULL, *first = NULL;
966 * if the driver polls the top of the queue and then removes
967 * the polled packet, we must return the same packet.
969 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
970 cl = ifd->pollcache_;
972 if (ifd->efficient_) {
973 /* check if this class is overlimit */
974 if (cl->undertime_.tv_sec != 0 &&
975 rmc_under_limit(cl, &now) == 0)
978 ifd->pollcache_ = NULL;
982 /* mode == ALTDQ_POLL || pollcache == NULL */
983 ifd->pollcache_ = NULL;
984 ifd->borrowed_[ifd->qi_] = NULL;
989 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
990 if (ifd->na_[cpri] == 0)
994 * Loop through twice for a priority level, if some class
995 * was unable to send a packet the first round because
996 * of the weighted round-robin mechanism.
997 * During the second loop at this level, deficit==2.
998 * (This second loop is not needed if for every class,
999 * "M[cl->pri_])" times "cl->allotment" is greater than
1000 * the byte size for the largest packet in the class.)
1003 cl = ifd->active_[cpri];
1006 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
1007 cl->bytes_alloc_ += cl->w_allotment_;
1008 if (!qempty(cl->q_)) {
1009 if ((cl->undertime_.tv_sec == 0) ||
1010 rmc_under_limit(cl, &now)) {
1011 if (cl->bytes_alloc_ > 0 || deficit > 1)
1014 /* underlimit but no alloc */
1017 ifd->borrowed_[ifd->qi_] = NULL;
1020 else if (first == NULL && cl->borrow_ != NULL)
1021 first = cl; /* borrowing candidate */
1024 cl->bytes_alloc_ = 0;
1026 } while (cl != ifd->active_[cpri]);
1029 /* first loop found an underlimit class with deficit */
1030 /* Loop on same priority level, with new deficit. */
1036 #ifdef ADJUST_CUTOFF
1038 * no underlimit class found. if cutoff is taking effect,
1039 * increase cutoff and try again.
1041 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1043 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1046 #endif /* ADJUST_CUTOFF */
1048 * If LINK_EFFICIENCY is turned on, then the first overlimit
1049 * class we encounter will send a packet if all the classes
1050 * of the link-sharing structure are overlimit.
1053 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1055 if (!ifd->efficient_ || first == NULL)
1060 #if 0 /* too time-consuming for nothing */
1062 CALLOUT_STOP(&cl->callout_);
1064 cl->undertime_.tv_sec = 0;
1066 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1067 ifd->cutoff_ = cl->borrow_->depth_;
1070 * Deque the packet and do the book keeping...
1073 if (op == ALTDQ_REMOVE) {
1076 panic("_rmc_wrr_dequeue_next");
1081 * Update class statistics and link data.
1083 if (cl->bytes_alloc_ > 0)
1084 cl->bytes_alloc_ -= m_pktlen(m);
1086 if ((cl->bytes_alloc_ <= 0) || first == cl)
1087 ifd->active_[cl->pri_] = cl->peer_;
1089 ifd->active_[cl->pri_] = cl;
1091 ifd->class_[ifd->qi_] = cl;
1092 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1093 ifd->now_[ifd->qi_] = now;
1094 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1097 /* mode == ALTDQ_PPOLL */
1099 ifd->pollcache_ = cl;
1105 * Dequeue & return next packet from the highest priority class that
1106 * has a packet to send & has enough allocation to send it. This
1107 * routine is called by a driver whenever it needs a new packet to
1111 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1115 struct rm_class *cl, *first = NULL;
1121 * if the driver polls the top of the queue and then removes
1122 * the polled packet, we must return the same packet.
1124 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1125 cl = ifd->pollcache_;
1127 ifd->pollcache_ = NULL;
1130 /* mode == ALTDQ_POLL || pollcache == NULL */
1131 ifd->pollcache_ = NULL;
1132 ifd->borrowed_[ifd->qi_] = NULL;
1134 #ifdef ADJUST_CUTOFF
1137 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1138 if (ifd->na_[cpri] == 0)
1140 cl = ifd->active_[cpri];
1143 if (!qempty(cl->q_)) {
1144 if ((cl->undertime_.tv_sec == 0) ||
1145 rmc_under_limit(cl, &now))
1147 if (first == NULL && cl->borrow_ != NULL)
1151 } while (cl != ifd->active_[cpri]);
1154 #ifdef ADJUST_CUTOFF
1156 * no underlimit class found. if cutoff is taking effect, increase
1157 * cutoff and try again.
1159 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1163 #endif /* ADJUST_CUTOFF */
1165 * If LINK_EFFICIENCY is turned on, then the first overlimit
1166 * class we encounter will send a packet if all the classes
1167 * of the link-sharing structure are overlimit.
1170 if (!ifd->efficient_ || first == NULL)
1175 #if 0 /* too time-consuming for nothing */
1177 CALLOUT_STOP(&cl->callout_);
1179 cl->undertime_.tv_sec = 0;
1181 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1182 ifd->cutoff_ = cl->borrow_->depth_;
1185 * Deque the packet and do the book keeping...
1188 if (op == ALTDQ_REMOVE) {
1191 panic("_rmc_prr_dequeue_next");
1195 ifd->active_[cpri] = cl->peer_;
1197 ifd->class_[ifd->qi_] = cl;
1198 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1199 ifd->now_[ifd->qi_] = now;
1200 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1203 /* mode == ALTDQ_POLL */
1205 ifd->pollcache_ = cl;
1212 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1213 * is invoked by the packet driver to get the next packet to be
1214 * dequeued and output on the link. If WRR is enabled, then the
1215 * WRR dequeue next routine will determine the next packet to sent.
1216 * Otherwise, packet-by-packet round robin is invoked.
1218 * Returns: NULL, if a packet is not available or if all
1219 * classes are overlimit.
1221 * Otherwise, Pointer to the next packet.
1225 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1227 if (ifd->queued_ >= ifd->maxqueued_)
1230 return (_rmc_wrr_dequeue_next(ifd, mode));
1232 return (_rmc_prr_dequeue_next(ifd, mode));
1236 * Update the utilization estimate for the packet that just completed.
1237 * The packet's class & the parent(s) of that class all get their
1238 * estimators updated. This routine is called by the driver's output-
1239 * packet-completion interrupt service routine.
1243 * a macro to approximate "divide by 1000" that gives 0.000999,
1244 * if a value has enough effective digits.
1245 * (on pentium, mul takes 9 cycles but div takes 46!)
1247 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1249 rmc_update_class_util(struct rm_ifdat *ifd)
1251 int idle, avgidle, pktlen;
1252 int pkt_time, tidle;
1253 rm_class_t *cl, *borrowed;
1254 rm_class_t *borrows;
1255 struct timeval *nowp;
1258 * Get the most recent completed class.
1260 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1263 pktlen = ifd->curlen_[ifd->qo_];
1264 borrowed = ifd->borrowed_[ifd->qo_];
1267 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1270 * Run estimator on class and its ancestors.
1273 * rm_update_class_util is designed to be called when the
1274 * transfer is completed from a xmit complete interrupt,
1275 * but most drivers don't implement an upcall for that.
1276 * so, just use estimated completion time.
1277 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1279 nowp = &ifd->now_[ifd->qo_];
1280 /* get pkt_time (for link) in usec */
1281 #if 1 /* use approximation */
1282 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1283 pkt_time = NSEC_TO_USEC(pkt_time);
1285 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1287 #if 1 /* ALTQ4PPP */
1288 if (TV_LT(nowp, &ifd->ifnow_)) {
1292 * make sure the estimated completion time does not go
1293 * too far. it can happen when the link layer supports
1294 * data compression or the interface speed is set to
1295 * a much lower value.
1297 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1298 if (iftime+pkt_time < ifd->maxiftime_) {
1299 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1301 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1304 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1307 if (TV_LT(nowp, &ifd->ifnow_)) {
1308 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1310 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1314 while (cl != NULL) {
1315 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1316 if (idle >= 2000000)
1318 * this class is idle enough, reset avgidle.
1319 * (TV_DELTA returns 2000000 us when delta is large.)
1321 cl->avgidle_ = cl->maxidle_;
1323 /* get pkt_time (for class) in usec */
1324 #if 1 /* use approximation */
1325 pkt_time = pktlen * cl->ns_per_byte_;
1326 pkt_time = NSEC_TO_USEC(pkt_time);
1328 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1332 avgidle = cl->avgidle_;
1333 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1334 cl->avgidle_ = avgidle;
1336 /* Are we overlimit ? */
1338 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1341 * need some lower bound for avgidle, otherwise
1342 * a borrowing class gets unbounded penalty.
1344 if (avgidle < cl->minidle_)
1345 avgidle = cl->avgidle_ = cl->minidle_;
1347 /* set next idle to make avgidle 0 */
1349 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1350 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1354 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1355 cl->undertime_.tv_sec = 0;
1356 if (cl->sleeping_) {
1357 CALLOUT_STOP(&cl->callout_);
1362 if (borrows != NULL) {
1364 ++cl->stats_.borrows;
1368 cl->last_ = ifd->ifnow_;
1369 cl->last_pkttime_ = pkt_time;
1372 if (cl->parent_ == NULL) {
1373 /* take stats of root class */
1374 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1382 * Check to see if cutoff needs to set to a new level.
1384 cl = ifd->class_[ifd->qo_];
1385 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1387 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1388 rmc_tl_satisfied(ifd, nowp);
1389 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1391 ifd->cutoff_ = borrowed->depth_;
1392 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1395 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1398 rmc_tl_satisfied(ifd, &now);
1400 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1402 ifd->cutoff_ = borrowed->depth_;
1403 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1409 * Release class slot
1411 ifd->borrowed_[ifd->qo_] = NULL;
1412 ifd->class_[ifd->qo_] = NULL;
1413 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1419 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1420 * over-limit action routines. These get invoked by rmc_under_limit()
1421 * if a class with packets to send if over its bandwidth limit & can't
1422 * borrow from a parent class.
1428 rmc_drop_action(struct rm_class *cl)
1430 struct rm_ifdat *ifd = cl->ifdat_;
1432 ASSERT(qlen(cl->q_) > 0);
1435 ifd->na_[cl->pri_]--;
1438 void rmc_dropall(struct rm_class *cl)
1440 struct rm_ifdat *ifd = cl->ifdat_;
1442 if (!qempty(cl->q_)) {
1445 ifd->na_[cl->pri_]--;
1450 hzto(struct timeval *tv)
1455 t2.tv_sec = tv->tv_sec - t2.tv_sec;
1456 t2.tv_usec = tv->tv_usec - t2.tv_usec;
1457 return (tvtohz(&t2));
1462 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1463 * delay action routine. It is invoked via rmc_under_limit when the
1464 * packet is discoverd to be overlimit.
1466 * If the delay action is result of borrow class being overlimit, then
1467 * delay for the offtime of the borrowing class that is overlimit.
1473 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1475 int delay, t, extradelay;
1477 cl->stats_.overactions++;
1478 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1479 #ifndef BORROW_OFFTIME
1480 delay += cl->offtime_;
1483 if (!cl->sleeping_) {
1484 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1485 #ifdef BORROW_OFFTIME
1487 extradelay = borrow->offtime_;
1490 extradelay = cl->offtime_;
1494 * XXX recalculate suspend time:
1495 * current undertime is (tidle + pkt_time) calculated
1496 * from the last transmission.
1497 * tidle: time required to bring avgidle back to 0
1498 * pkt_time: target waiting time for this class
1499 * we need to replace pkt_time by offtime
1501 extradelay -= cl->last_pkttime_;
1503 if (extradelay > 0) {
1504 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1505 delay += extradelay;
1509 cl->stats_.delays++;
1512 * Since packets are phased randomly with respect to the
1513 * clock, 1 tick (the next clock tick) can be an arbitrarily
1514 * short time so we have to wait for at least two ticks.
1515 * NOTE: If there's no other traffic, we need the timer as
1516 * a 'backstop' to restart this class.
1518 if (delay > tick * 2) {
1519 /* FreeBSD rounds up the tick */
1520 t = hzto(&cl->undertime_);
1523 CALLOUT_RESET(&cl->callout_, t, rmc_restart, cl);
1529 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1530 * called by the system timer code & is responsible checking if the
1531 * class is still sleeping (it might have been restarted as a side
1532 * effect of the queue scan on a packet arrival) and, if so, restarting
1533 * output for the class. Inspecting the class state & restarting output
1534 * require locking the class structure. In general the driver is
1535 * responsible for locking but this is the only routine that is not
1536 * called directly or indirectly from the interface driver so it has
1537 * know about system locking conventions. Under bsd, locking is done
1538 * by raising IPL to splimp so that's what's implemented here. On a
1539 * different system this would probably need to be changed.
1545 rmc_restart(void *arg)
1547 struct rm_class *cl = arg;
1548 struct rm_ifdat *ifd = cl->ifdat_;
1549 struct epoch_tracker et;
1553 NET_EPOCH_ENTER(et);
1554 IFQ_LOCK(ifd->ifq_);
1555 CURVNET_SET(ifd->ifq_->altq_ifp->if_vnet);
1556 if (cl->sleeping_) {
1558 cl->undertime_.tv_sec = 0;
1560 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1561 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1562 (ifd->restart)(ifd->ifq_);
1566 IFQ_UNLOCK(ifd->ifq_);
1573 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1574 * handling routine for the root class of the link sharing structure.
1580 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1582 panic("rmc_root_overlimit");
1586 * Packet Queue handling routines. Eventually, this is to localize the
1587 * effects on the code whether queues are red queues or droptail
1592 _rmc_addq(rm_class_t *cl, mbuf_t *m)
1595 if (q_is_rio(cl->q_))
1596 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1599 if (q_is_red(cl->q_))
1600 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1601 #endif /* ALTQ_RED */
1603 if (q_is_codel(cl->q_))
1604 return codel_addq(cl->codel_, cl->q_, m);
1607 if (cl->flags_ & RMCF_CLEARDSCP)
1608 write_dsfield(m, cl->pktattr_, 0);
1614 /* note: _rmc_dropq is not called for red */
1616 _rmc_dropq(rm_class_t *cl)
1620 if ((m = _getq(cl->q_)) != NULL)
1625 _rmc_getq(rm_class_t *cl)
1628 if (q_is_rio(cl->q_))
1629 return rio_getq((rio_t *)cl->red_, cl->q_);
1632 if (q_is_red(cl->q_))
1633 return red_getq(cl->red_, cl->q_);
1636 if (q_is_codel(cl->q_))
1637 return codel_getq(cl->codel_, cl->q_);
1639 return _getq(cl->q_);
1643 _rmc_pollq(rm_class_t *cl)
1645 return qhead(cl->q_);
1650 struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
1651 struct cbqtrace *cbqtrace_ptr = NULL;
1655 * DDB hook to trace cbq events:
1656 * the last 1024 events are held in a circular buffer.
1657 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1659 void cbqtrace_dump(int);
1660 static char *rmc_funcname(void *);
1662 static struct rmc_funcs {
1667 rmc_init, "rmc_init",
1668 rmc_queue_packet, "rmc_queue_packet",
1669 rmc_under_limit, "rmc_under_limit",
1670 rmc_update_class_util, "rmc_update_class_util",
1671 rmc_delay_action, "rmc_delay_action",
1672 rmc_restart, "rmc_restart",
1673 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1677 static char *rmc_funcname(void *func)
1679 struct rmc_funcs *fp;
1681 for (fp = rmc_funcs; fp->func != NULL; fp++)
1682 if (fp->func == func)
1687 void cbqtrace_dump(int counter)
1692 counter = counter % NCBQTRACE;
1693 p = (int *)&cbqtrace_buffer[counter];
1695 for (i=0; i<20; i++) {
1696 printf("[0x%x] ", *p++);
1697 printf("%s: ", rmc_funcname((void *)*p++));
1699 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1700 printf("%d\n",*p++);
1702 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1703 p = (int *)cbqtrace_buffer;
1706 #endif /* CBQ_TRACE */
1707 #endif /* ALTQ_CBQ */
1709 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || \
1710 defined(ALTQ_HFSC) || defined(ALTQ_PRIQ) || defined(ALTQ_CODEL)
1711 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1714 _addq(class_queue_t *q, mbuf_t *m)
1718 if ((m0 = qtail(q)) != NULL)
1719 m->m_nextpkt = m0->m_nextpkt;
1728 _getq(class_queue_t *q)
1732 if ((m = qtail(q)) == NULL)
1734 if ((m0 = m->m_nextpkt) != m)
1735 m->m_nextpkt = m0->m_nextpkt;
1737 ASSERT(qlen(q) == 1);
1741 m0->m_nextpkt = NULL;
1745 /* drop a packet at the tail of the queue */
1747 _getq_tail(class_queue_t *q)
1749 mbuf_t *m, *m0, *prev;
1751 if ((m = m0 = qtail(q)) == NULL)
1757 prev->m_nextpkt = m->m_nextpkt;
1759 ASSERT(qlen(q) == 1);
1764 m->m_nextpkt = NULL;
1768 /* randomly select a packet in the queue */
1770 _getq_random(class_queue_t *q)
1775 if ((m = qtail(q)) == NULL)
1777 if (m->m_nextpkt == m) {
1778 ASSERT(qlen(q) == 1);
1781 struct mbuf *prev = NULL;
1783 n = arc4random() % qlen(q) + 1;
1784 for (i = 0; i < n; i++) {
1788 prev->m_nextpkt = m->m_nextpkt;
1793 m->m_nextpkt = NULL;
1798 _removeq(class_queue_t *q, mbuf_t *m)
1807 prev->m_nextpkt = m->m_nextpkt;
1810 else if (qtail(q) == m)
1816 _flushq(class_queue_t *q)
1820 while ((m = _getq(q)) != NULL)
1822 ASSERT(qlen(q) == 0);
1825 #endif /* !__GNUC__ || ALTQ_DEBUG */
1826 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */