1 /* $KAME: altq_rmclass.c,v 1.18 2003/11/06 06:32:53 kjc Exp $ */
2 /* $DragonFly: src/sys/net/altq/altq_rmclass.c,v 1.2 2005/05/25 01:44:33 dillon Exp $ */
5 * Copyright (c) 1991-1997 Regents of the University of California.
8 * Redistribution and use in source and binary forms, with or without
9 * 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 the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the Network Research
19 * Group at Lawrence Berkeley Laboratory.
20 * 4. Neither the name of the University nor of the Laboratory may be used
21 * to endorse or promote products derived from this software without
22 * specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * LBL code modified by speer@eng.sun.com, May 1977.
37 * For questions and/or comments, please send mail to cbq@ee.lbl.gov
40 #ident "@(#)rm_class.c 1.48 97/12/05 SMI"
44 #include "opt_inet6.h"
46 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
48 #include <sys/param.h>
49 #include <sys/malloc.h>
51 #include <sys/socket.h>
52 #include <sys/systm.h>
53 #include <sys/callout.h>
54 #include <sys/errno.h>
56 #include <sys/thread.h>
60 #include <net/altq/altq.h>
61 #include <net/altq/altq_rmclass.h>
62 #include <net/altq/altq_rmclass_debug.h>
63 #include <net/altq/altq_red.h>
64 #include <net/altq/altq_rio.h>
66 #include <sys/thread2.h>
69 static struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
70 static struct cbqtrace *cbqtrace_ptr = NULL;
71 static int cbqtrace_count;
78 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
84 static int rmc_satisfied(struct rm_class *, struct timeval *);
85 static void rmc_wrr_set_weights(struct rm_ifdat *);
86 static void rmc_depth_compute(struct rm_class *);
87 static void rmc_depth_recompute(rm_class_t *);
89 static struct mbuf *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
90 static struct mbuf *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
92 static int _rmc_addq(rm_class_t *, struct mbuf *);
93 static void _rmc_dropq(rm_class_t *);
94 static struct mbuf *_rmc_getq(rm_class_t *);
95 static struct mbuf *_rmc_pollq(rm_class_t *);
97 static int rmc_under_limit(struct rm_class *, struct timeval *);
98 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *);
99 static void rmc_drop_action(struct rm_class *);
100 static void rmc_restart(void *);
101 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
103 #define BORROW_OFFTIME
105 * BORROW_OFFTIME (experimental):
106 * borrow the offtime of the class borrowing from.
107 * the reason is that when its own offtime is set, the class is unable
108 * to borrow much, especially when cutoff is taking effect.
109 * but when the borrowed class is overloaded (advidle is close to minidle),
110 * use the borrowing class's offtime to avoid overload.
112 #define ADJUST_CUTOFF
114 * ADJUST_CUTOFF (experimental):
115 * if no underlimit class is found due to cutoff, increase cutoff and
116 * retry the scheduling loop.
117 * also, don't invoke delay_actions while cutoff is taking effect,
118 * since a sleeping class won't have a chance to be scheduled in the
121 * now heuristics for setting the top-level variable (cutoff_) becomes:
122 * 1. if a packet arrives for a not-overlimit class, set cutoff
123 * to the depth of the class.
124 * 2. if cutoff is i, and a packet arrives for an overlimit class
125 * with an underlimit ancestor at a lower level than i (say j),
126 * then set cutoff to j.
127 * 3. at scheduling a packet, if there is no underlimit class
128 * due to the current cutoff level, increase cutoff by 1 and
129 * then try to schedule again.
134 * rmc_newclass(...) - Create a new resource management class at priority
135 * 'pri' on the interface given by 'ifd'.
137 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
138 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
139 * than 100% of the bandwidth, this number should be the
140 * 'effective' rate for the class. Let f be the
141 * bandwidth fraction allocated to this class, and let
142 * nsPerByte be the data rate of the output link in
143 * nanoseconds/byte. Then nsecPerByte is set to
144 * nsPerByte / f. E.g., 1600 (= 800 / .5)
145 * for a class that gets 50% of an ethernet's bandwidth.
147 * action the routine to call when the class is over limit.
149 * maxq max allowable queue size for class (in packets).
151 * parent parent class pointer.
153 * borrow class to borrow from (should be either 'parent' or null).
155 * maxidle max value allowed for class 'idle' time estimate (this
156 * parameter determines how large an initial burst of packets
157 * can be before overlimit action is invoked.
159 * offtime how long 'delay' action will delay when class goes over
160 * limit (this parameter determines the steady-state burst
161 * size when a class is running over its limit).
163 * Maxidle and offtime have to be computed from the following: If the
164 * average packet size is s, the bandwidth fraction allocated to this
165 * class is f, we want to allow b packet bursts, and the gain of the
166 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
168 * ptime = s * nsPerByte * (1 - f) / f
169 * maxidle = ptime * (1 - g^b) / g^b
170 * minidle = -ptime * (1 / (f - 1))
171 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
173 * Operationally, it's convenient to specify maxidle & offtime in units
174 * independent of the link bandwidth so the maxidle & offtime passed to
175 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
176 * (The constant factor is a scale factor needed to make the parameters
177 * integers. This scaling also means that the 'unscaled' values of
178 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
179 * not nanoseconds.) Also note that the 'idle' filter computation keeps
180 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
181 * maxidle also must be scaled upward by this value. Thus, the passed
182 * values for maxidle and offtime can be computed as follows:
184 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
185 * offtime = offtime * 8 / (1000 * nsecPerByte)
187 * When USE_HRTIME is employed, then maxidle and offtime become:
188 * maxidle = maxilde * (8.0 / nsecPerByte);
189 * offtime = offtime * (8.0 / nsecPerByte);
192 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
193 void (*action)(rm_class_t *, rm_class_t *), int maxq,
194 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
195 int minidle, u_int offtime, int pktsize, int flags)
198 struct rm_class *peer;
201 if (pri >= RM_MAXPRIO)
204 if (flags & RMCF_RED) {
206 printf("rmc_newclass: RED not configured for CBQ!\n");
212 if (flags & RMCF_RIO) {
214 printf("rmc_newclass: RIO not configured for CBQ!\n");
220 cl = malloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
221 callout_init(&cl->callout_);
222 cl->q_ = malloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO);
225 * Class initialization.
227 cl->children_ = NULL;
228 cl->parent_ = parent;
229 cl->borrow_ = borrow;
233 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
236 cl->ns_per_byte_ = nsecPerByte;
238 qlimit(cl->q_) = maxq;
239 qtype(cl->q_) = Q_DROPHEAD;
243 #if 1 /* minidle is also scaled in ALTQ */
244 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
245 if (cl->minidle_ > 0)
248 cl->minidle_ = minidle;
250 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
251 if (cl->maxidle_ == 0)
253 #if 1 /* offtime is also scaled in ALTQ */
254 cl->avgidle_ = cl->maxidle_;
255 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
256 if (cl->offtime_ == 0)
260 cl->offtime_ = (offtime * nsecPerByte) / 8;
262 cl->overlimit = action;
265 if (flags & (RMCF_RED|RMCF_RIO)) {
266 int red_flags, red_pkttime;
269 if (flags & RMCF_ECN)
270 red_flags |= REDF_ECN;
272 if (flags & RMCF_CLEARDSCP)
273 red_flags |= RIOF_CLEARDSCP;
275 red_pkttime = nsecPerByte * pktsize / 1000;
277 if (flags & RMCF_RED) {
278 cl->red_ = red_alloc(0, 0,
279 qlimit(cl->q_) * 10/100,
280 qlimit(cl->q_) * 30/100,
281 red_flags, red_pkttime);
282 if (cl->red_ != NULL)
283 qtype(cl->q_) = Q_RED;
287 cl->red_ = (red_t *)rio_alloc(0, NULL,
288 red_flags, red_pkttime);
289 if (cl->red_ != NULL)
290 qtype(cl->q_) = Q_RIO;
294 #endif /* ALTQ_RED */
297 * put the class into the class tree
300 if ((peer = ifd->active_[pri]) != NULL) {
301 /* find the last class at this pri */
303 while (peer->peer_ != ifd->active_[pri])
307 ifd->active_[pri] = cl;
312 cl->next_ = parent->children_;
313 parent->children_ = cl;
318 * Compute the depth of this class and its ancestors in the class
321 rmc_depth_compute(cl);
324 * If CBQ's WRR is enabled, then initialize the class WRR state.
328 ifd->alloc_[pri] += cl->allotment_;
329 rmc_wrr_set_weights(ifd);
336 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
337 int minidle, u_int offtime, int pktsize)
339 struct rm_ifdat *ifd;
344 old_allotment = cl->allotment_;
347 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
349 cl->ns_per_byte_ = nsecPerByte;
351 qlimit(cl->q_) = maxq;
353 #if 1 /* minidle is also scaled in ALTQ */
354 cl->minidle_ = (minidle * nsecPerByte) / 8;
355 if (cl->minidle_ > 0)
358 cl->minidle_ = minidle;
360 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
361 if (cl->maxidle_ == 0)
363 #if 1 /* offtime is also scaled in ALTQ */
364 cl->avgidle_ = cl->maxidle_;
365 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
366 if (cl->offtime_ == 0)
370 cl->offtime_ = (offtime * nsecPerByte) / 8;
374 * If CBQ's WRR is enabled, then initialize the class WRR state.
377 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
378 rmc_wrr_set_weights(ifd);
386 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
387 * the appropriate run robin weights for the CBQ weighted round robin
394 rmc_wrr_set_weights(struct rm_ifdat *ifd)
397 struct rm_class *cl, *clh;
399 for (i = 0; i < RM_MAXPRIO; i++) {
401 * This is inverted from that of the simulator to
402 * maintain precision.
404 if (ifd->num_[i] == 0)
407 ifd->M_[i] = ifd->alloc_[i] /
408 (ifd->num_[i] * ifd->maxpkt_);
410 * Compute the weighted allotment for each class.
411 * This takes the expensive div instruction out
412 * of the main loop for the wrr scheduling path.
413 * These only get recomputed when a class comes or
416 if (ifd->active_[i] != NULL) {
417 clh = cl = ifd->active_[i];
419 /* safe-guard for slow link or alloc_ == 0 */
421 cl->w_allotment_ = 0;
423 cl->w_allotment_ = cl->allotment_ /
426 } while ((cl != NULL) && (cl != clh));
432 rmc_get_weight(struct rm_ifdat *ifd, int pri)
434 if ((pri >= 0) && (pri < RM_MAXPRIO))
435 return (ifd->M_[pri]);
442 * rmc_depth_compute(struct rm_class *cl) - This function computes the
443 * appropriate depth of class 'cl' and its ancestors.
449 rmc_depth_compute(struct rm_class *cl)
451 rm_class_t *t = cl, *p;
454 * Recompute the depth for the branch of the tree.
458 if (p && (t->depth_ >= p->depth_)) {
459 p->depth_ = t->depth_ + 1;
468 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
469 * the depth of the tree after a class has been deleted.
475 rmc_depth_recompute(rm_class_t *cl)
482 if ((t = p->children_) == NULL) {
488 if (t->depth_ > cdepth)
493 if (p->depth_ == cdepth + 1)
494 /* no change to this parent */
497 p->depth_ = cdepth + 1;
505 if (cl->depth_ >= 1) {
506 if (cl->children_ == NULL) {
508 } else if ((t = cl->children_) != NULL) {
510 if (t->children_ != NULL)
511 rmc_depth_recompute(t);
515 rmc_depth_compute(cl);
522 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
523 * function deletes a class from the link-sharing structure and frees
524 * all resources associated with the class.
530 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
532 struct rm_class *p, *head, *previous;
535 KKASSERT(cl->children_ == NULL);
538 callout_stop(&cl->callout_);
542 * Free packets in the packet queue.
543 * XXX - this may not be a desired behavior. Packets should be
549 * If the class has a parent, then remove the class from the
550 * class from the parent's children chain.
552 if (cl->parent_ != NULL) {
553 head = cl->parent_->children_;
555 if (head->next_ == NULL) {
556 KKASSERT(head == cl);
557 cl->parent_->children_ = NULL;
558 cl->parent_->leaf_ = 1;
559 } else while (p != NULL) {
562 cl->parent_->children_ = cl->next_;
564 previous->next_ = cl->next_;
575 * Delete class from class priority peer list.
577 if ((p = ifd->active_[cl->pri_]) != NULL) {
579 * If there is more than one member of this priority
580 * level, then look for class(cl) in the priority level.
583 while (p->peer_ != cl)
585 p->peer_ = cl->peer_;
587 if (ifd->active_[cl->pri_] == cl)
588 ifd->active_[cl->pri_] = cl->peer_;
591 ifd->active_[cl->pri_] = NULL;
596 * Recompute the WRR weights.
599 ifd->alloc_[cl->pri_] -= cl->allotment_;
600 ifd->num_[cl->pri_]--;
601 rmc_wrr_set_weights(ifd);
605 * Re-compute the depth of the tree.
608 rmc_depth_recompute(cl->parent_);
610 rmc_depth_recompute(ifd->root_);
616 * Free the class structure.
618 if (cl->red_ != NULL) {
620 if (q_is_rio(cl->q_))
621 rio_destroy((rio_t *)cl->red_);
624 if (q_is_red(cl->q_))
625 red_destroy(cl->red_);
628 free(cl->q_, M_ALTQ);
634 * rmc_init(...) - Initialize the resource management data structures
635 * associated with the output portion of interface 'ifp'. 'ifd' is
636 * where the structures will be built (for backwards compatibility, the
637 * structures aren't kept in the ifnet struct). 'nsecPerByte'
638 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
639 * 'restart' is the driver-specific routine that the generic 'delay
640 * until under limit' action will call to restart output. `maxq'
641 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
642 * is the maximum number of packets that the resource management
643 * code will allow to be queued 'downstream' (this is typically 1).
649 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
650 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
651 int minidle, u_int offtime, int flags)
656 * Initialize the CBQ tracing/debug facility.
660 bzero(ifd, sizeof (*ifd));
661 mtu = ifq->altq_ifp->if_mtu;
663 ifd->restart = restart;
664 ifd->maxqueued_ = maxqueued;
665 ifd->ns_per_byte_ = nsecPerByte;
667 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
668 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
670 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
671 if (mtu * nsecPerByte > 10 * 1000000)
672 ifd->maxiftime_ /= 4;
676 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
679 * Initialize the CBQ's WRR state.
681 for (i = 0; i < RM_MAXPRIO; i++) {
686 ifd->active_[i] = NULL;
690 * Initialize current packet state.
694 for (i = 0; i < RM_MAXQUEUED; i++) {
695 ifd->class_[i] = NULL;
697 ifd->borrowed_[i] = NULL;
701 * Create the root class of the link-sharing structure.
703 ifd->root_ = rmc_newclass(0, ifd, nsecPerByte, rmc_root_overlimit,
704 maxq, 0, 0, maxidle, minidle, offtime, 0, 0);
705 if (ifd->root_ == NULL) {
706 printf("rmc_init: root class not allocated\n");
709 ifd->root_->depth_ = 0;
714 * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by
715 * mbuf 'm' to queue for resource class 'cl'. This routine is called
716 * by a driver's if_output routine. This routine must be called with
717 * output packet completion interrupts locked out (to avoid racing with
720 * Returns: 0 on successful queueing
721 * -1 when packet drop occurs
724 rmc_queue_packet(struct rm_class *cl, struct mbuf *m)
727 struct rm_ifdat *ifd = cl->ifdat_;
729 int is_empty = qempty(cl->q_);
732 if (ifd->cutoff_ > 0) {
733 if (TV_LT(&cl->undertime_, &now)) {
734 if (ifd->cutoff_ > cl->depth_)
735 ifd->cutoff_ = cl->depth_;
736 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
741 * the class is overlimit. if the class has
742 * underlimit ancestors, set cutoff to the lowest
745 struct rm_class *borrow = cl->borrow_;
747 while (borrow != NULL &&
748 borrow->depth_ < ifd->cutoff_) {
749 if (TV_LT(&borrow->undertime_, &now)) {
750 ifd->cutoff_ = borrow->depth_;
751 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
754 borrow = borrow->borrow_;
758 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
759 if (TV_LT(&cl->borrow_->undertime_, &now)) {
760 ifd->cutoff_ = cl->borrow_->depth_;
761 CBQTRACE(rmc_queue_packet, 'ffob',
762 cl->borrow_->depth_);
768 if (_rmc_addq(cl, m) < 0)
773 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
777 if (qlen(cl->q_) > qlimit(cl->q_)) {
778 /* note: qlimit can be set to 0 or 1 */
787 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
788 * classes to see if there are satified.
792 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
797 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
798 if ((bp = ifd->active_[i]) != NULL) {
801 if (!rmc_satisfied(p, now)) {
802 ifd->cutoff_ = p->depth_;
814 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
818 rmc_satisfied(struct rm_class *cl, struct timeval *now)
824 if (TV_LT(now, &cl->undertime_))
826 if (cl->depth_ == 0) {
827 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
832 if (cl->children_ != NULL) {
835 if (!rmc_satisfied(p, now))
845 * Return 1 if class 'cl' is under limit or can borrow from a parent,
846 * 0 if overlimit. As a side-effect, this routine will invoke the
847 * class overlimit action if the class if overlimit.
851 rmc_under_limit(struct rm_class *cl, struct timeval *now)
855 struct rm_ifdat *ifd = cl->ifdat_;
857 ifd->borrowed_[ifd->qi_] = NULL;
859 * If cl is the root class, then always return that it is
860 * underlimit. Otherwise, check to see if the class is underlimit.
862 if (cl->parent_ == NULL)
866 if (TV_LT(now, &cl->undertime_))
869 callout_stop(&cl->callout_);
871 cl->undertime_.tv_sec = 0;
876 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
877 if (((cl = cl->borrow_) == NULL) ||
878 (cl->depth_ > ifd->cutoff_)) {
881 /* cutoff is taking effect, just
882 return false without calling
886 #ifdef BORROW_OFFTIME
888 * check if the class can borrow offtime too.
889 * borrow offtime from the top of the borrow
890 * chain if the top class is not overloaded.
893 /* cutoff is taking effect, use this class as top. */
895 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
897 if (top != NULL && top->avgidle_ == top->minidle_)
900 (p->overlimit)(p, top);
903 (p->overlimit)(p, NULL);
911 ifd->borrowed_[ifd->qi_] = cl;
916 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
917 * Packet-by-packet round robin.
919 * The heart of the weighted round-robin scheduler, which decides which
920 * class next gets to send a packet. Highest priority first, then
921 * weighted round-robin within priorites.
923 * Each able-to-send class gets to send until its byte allocation is
924 * exhausted. Thus, the active pointer is only changed after a class has
925 * exhausted its allocation.
927 * If the scheduler finds no class that is underlimit or able to borrow,
928 * then the first class found that had a nonzero queue and is allowed to
929 * borrow gets to send.
933 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
935 struct rm_class *cl = NULL, *first = NULL;
944 * if the driver polls the top of the queue and then removes
945 * the polled packet, we must return the same packet.
947 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
948 cl = ifd->pollcache_;
950 if (ifd->efficient_) {
951 /* check if this class is overlimit */
952 if (cl->undertime_.tv_sec != 0 &&
953 rmc_under_limit(cl, &now) == 0)
956 ifd->pollcache_ = NULL;
960 /* mode == ALTDQ_POLL || pollcache == NULL */
961 ifd->pollcache_ = NULL;
962 ifd->borrowed_[ifd->qi_] = NULL;
967 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
968 if (ifd->na_[cpri] == 0)
972 * Loop through twice for a priority level, if some class
973 * was unable to send a packet the first round because
974 * of the weighted round-robin mechanism.
975 * During the second loop at this level, deficit==2.
976 * (This second loop is not needed if for every class,
977 * "M[cl->pri_])" times "cl->allotment" is greater than
978 * the byte size for the largest packet in the class.)
981 cl = ifd->active_[cpri];
982 KKASSERT(cl != NULL);
984 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
985 cl->bytes_alloc_ += cl->w_allotment_;
986 if (!qempty(cl->q_)) {
987 if ((cl->undertime_.tv_sec == 0) ||
988 rmc_under_limit(cl, &now)) {
989 if (cl->bytes_alloc_ > 0 || deficit > 1)
992 /* underlimit but no alloc */
995 ifd->borrowed_[ifd->qi_] = NULL;
998 else if (first == NULL && cl->borrow_ != NULL)
999 first = cl; /* borrowing candidate */
1002 cl->bytes_alloc_ = 0;
1004 } while (cl != ifd->active_[cpri]);
1007 /* first loop found an underlimit class with deficit */
1008 /* Loop on same priority level, with new deficit. */
1014 #ifdef ADJUST_CUTOFF
1016 * no underlimit class found. if cutoff is taking effect,
1017 * increase cutoff and try again.
1019 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1021 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1024 #endif /* ADJUST_CUTOFF */
1026 * If LINK_EFFICIENCY is turned on, then the first overlimit
1027 * class we encounter will send a packet if all the classes
1028 * of the link-sharing structure are overlimit.
1031 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1033 if (!ifd->efficient_ || first == NULL)
1038 #if 0 /* too time-consuming for nothing */
1040 callout_stop(&cl->callout_);
1042 cl->undertime_.tv_sec = 0;
1044 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1045 ifd->cutoff_ = cl->borrow_->depth_;
1048 * Deque the packet and do the book keeping...
1051 if (op == ALTDQ_REMOVE) {
1054 panic("_rmc_wrr_dequeue_next");
1059 * Update class statistics and link data.
1061 if (cl->bytes_alloc_ > 0)
1062 cl->bytes_alloc_ -= m_pktlen(m);
1064 if ((cl->bytes_alloc_ <= 0) || first == cl)
1065 ifd->active_[cl->pri_] = cl->peer_;
1067 ifd->active_[cl->pri_] = cl;
1069 ifd->class_[ifd->qi_] = cl;
1070 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1071 ifd->now_[ifd->qi_] = now;
1072 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1075 /* mode == ALTDQ_PPOLL */
1077 ifd->pollcache_ = cl;
1083 * Dequeue & return next packet from the highest priority class that
1084 * has a packet to send & has enough allocation to send it. This
1085 * routine is called by a driver whenever it needs a new packet to
1088 static struct mbuf *
1089 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1093 struct rm_class *cl, *first = NULL;
1099 * if the driver polls the top of the queue and then removes
1100 * the polled packet, we must return the same packet.
1102 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1103 cl = ifd->pollcache_;
1105 ifd->pollcache_ = NULL;
1108 /* mode == ALTDQ_POLL || pollcache == NULL */
1109 ifd->pollcache_ = NULL;
1110 ifd->borrowed_[ifd->qi_] = NULL;
1112 #ifdef ADJUST_CUTOFF
1115 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1116 if (ifd->na_[cpri] == 0)
1118 cl = ifd->active_[cpri];
1119 KKASSERT(cl != NULL);
1121 if (!qempty(cl->q_)) {
1122 if ((cl->undertime_.tv_sec == 0) ||
1123 rmc_under_limit(cl, &now))
1125 if (first == NULL && cl->borrow_ != NULL)
1129 } while (cl != ifd->active_[cpri]);
1132 #ifdef ADJUST_CUTOFF
1134 * no underlimit class found. if cutoff is taking effect, increase
1135 * cutoff and try again.
1137 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1141 #endif /* ADJUST_CUTOFF */
1143 * If LINK_EFFICIENCY is turned on, then the first overlimit
1144 * class we encounter will send a packet if all the classes
1145 * of the link-sharing structure are overlimit.
1148 if (!ifd->efficient_ || first == NULL)
1153 #if 0 /* too time-consuming for nothing */
1155 callout_stop(&cl->callout_);
1157 cl->undertime_.tv_sec = 0;
1159 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1160 ifd->cutoff_ = cl->borrow_->depth_;
1163 * Deque the packet and do the book keeping...
1166 if (op == ALTDQ_REMOVE) {
1169 panic("_rmc_prr_dequeue_next");
1173 ifd->active_[cpri] = cl->peer_;
1175 ifd->class_[ifd->qi_] = cl;
1176 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1177 ifd->now_[ifd->qi_] = now;
1178 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1181 /* mode == ALTDQ_POLL */
1183 ifd->pollcache_ = cl;
1190 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1191 * is invoked by the packet driver to get the next packet to be
1192 * dequeued and output on the link. If WRR is enabled, then the
1193 * WRR dequeue next routine will determine the next packet to sent.
1194 * Otherwise, packet-by-packet round robin is invoked.
1196 * Returns: NULL, if a packet is not available or if all
1197 * classes are overlimit.
1199 * Otherwise, Pointer to the next packet.
1203 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1205 if (ifd->queued_ >= ifd->maxqueued_)
1208 return (_rmc_wrr_dequeue_next(ifd, mode));
1210 return (_rmc_prr_dequeue_next(ifd, mode));
1214 * Update the utilization estimate for the packet that just completed.
1215 * The packet's class & the parent(s) of that class all get their
1216 * estimators updated. This routine is called by the driver's output-
1217 * packet-completion interrupt service routine.
1221 * a macro to approximate "divide by 1000" that gives 0.000999,
1222 * if a value has enough effective digits.
1223 * (on pentium, mul takes 9 cycles but div takes 46!)
1225 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1227 rmc_update_class_util(struct rm_ifdat *ifd)
1229 int idle, avgidle, pktlen;
1230 int pkt_time, tidle;
1231 rm_class_t *cl, *borrowed;
1232 rm_class_t *borrows;
1233 struct timeval *nowp;
1236 * Get the most recent completed class.
1238 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1241 pktlen = ifd->curlen_[ifd->qo_];
1242 borrowed = ifd->borrowed_[ifd->qo_];
1245 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1248 * Run estimator on class and its ancestors.
1251 * rm_update_class_util is designed to be called when the
1252 * transfer is completed from a xmit complete interrupt,
1253 * but most drivers don't implement an upcall for that.
1254 * so, just use estimated completion time.
1255 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1257 nowp = &ifd->now_[ifd->qo_];
1258 /* get pkt_time (for link) in usec */
1259 #if 1 /* use approximation */
1260 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1261 pkt_time = NSEC_TO_USEC(pkt_time);
1263 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1265 #if 1 /* ALTQ4PPP */
1266 if (TV_LT(nowp, &ifd->ifnow_)) {
1270 * make sure the estimated completion time does not go
1271 * too far. it can happen when the link layer supports
1272 * data compression or the interface speed is set to
1273 * a much lower value.
1275 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1276 if (iftime+pkt_time < ifd->maxiftime_) {
1277 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1279 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1282 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1285 if (TV_LT(nowp, &ifd->ifnow_)) {
1286 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1288 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1292 while (cl != NULL) {
1293 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1294 if (idle >= 2000000)
1296 * this class is idle enough, reset avgidle.
1297 * (TV_DELTA returns 2000000 us when delta is large.)
1299 cl->avgidle_ = cl->maxidle_;
1301 /* get pkt_time (for class) in usec */
1302 #if 1 /* use approximation */
1303 pkt_time = pktlen * cl->ns_per_byte_;
1304 pkt_time = NSEC_TO_USEC(pkt_time);
1306 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1310 avgidle = cl->avgidle_;
1311 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1312 cl->avgidle_ = avgidle;
1314 /* Are we overlimit ? */
1316 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1319 * need some lower bound for avgidle, otherwise
1320 * a borrowing class gets unbounded penalty.
1322 if (avgidle < cl->minidle_)
1323 avgidle = cl->avgidle_ = cl->minidle_;
1325 /* set next idle to make avgidle 0 */
1327 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1328 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1332 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1333 cl->undertime_.tv_sec = 0;
1334 if (cl->sleeping_) {
1335 callout_stop(&cl->callout_);
1340 if (borrows != NULL) {
1342 ++cl->stats_.borrows;
1346 cl->last_ = ifd->ifnow_;
1347 cl->last_pkttime_ = pkt_time;
1350 if (cl->parent_ == NULL) {
1351 /* take stats of root class */
1352 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1360 * Check to see if cutoff needs to set to a new level.
1362 cl = ifd->class_[ifd->qo_];
1363 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1365 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1366 rmc_tl_satisfied(ifd, nowp);
1367 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1369 ifd->cutoff_ = borrowed->depth_;
1370 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1373 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1376 rmc_tl_satisfied(ifd, &now);
1378 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1380 ifd->cutoff_ = borrowed->depth_;
1381 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1387 * Release class slot
1389 ifd->borrowed_[ifd->qo_] = NULL;
1390 ifd->class_[ifd->qo_] = NULL;
1391 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1397 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1398 * over-limit action routines. These get invoked by rmc_under_limit()
1399 * if a class with packets to send if over its bandwidth limit & can't
1400 * borrow from a parent class.
1406 rmc_drop_action(struct rm_class *cl)
1408 struct rm_ifdat *ifd = cl->ifdat_;
1410 KKASSERT(qlen(cl->q_) > 0);
1413 ifd->na_[cl->pri_]--;
1416 void rmc_dropall(struct rm_class *cl)
1418 struct rm_ifdat *ifd = cl->ifdat_;
1420 if (!qempty(cl->q_)) {
1423 ifd->na_[cl->pri_]--;
1429 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1430 * delay action routine. It is invoked via rmc_under_limit when the
1431 * packet is discoverd to be overlimit.
1433 * If the delay action is result of borrow class being overlimit, then
1434 * delay for the offtime of the borrowing class that is overlimit.
1440 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1442 int delay, t, extradelay;
1444 cl->stats_.overactions++;
1445 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1446 #ifndef BORROW_OFFTIME
1447 delay += cl->offtime_;
1450 if (!cl->sleeping_) {
1451 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1452 #ifdef BORROW_OFFTIME
1454 extradelay = borrow->offtime_;
1457 extradelay = cl->offtime_;
1461 * XXX recalculate suspend time:
1462 * current undertime is (tidle + pkt_time) calculated
1463 * from the last transmission.
1464 * tidle: time required to bring avgidle back to 0
1465 * pkt_time: target waiting time for this class
1466 * we need to replace pkt_time by offtime
1468 extradelay -= cl->last_pkttime_;
1470 if (extradelay > 0) {
1471 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1472 delay += extradelay;
1476 cl->stats_.delays++;
1479 * Since packets are phased randomly with respect to the
1480 * clock, 1 tick (the next clock tick) can be an arbitrarily
1481 * short time so we have to wait for at least two ticks.
1482 * NOTE: If there's no other traffic, we need the timer as
1483 * a 'backstop' to restart this class.
1485 if (delay > tick * 2)
1486 t = (delay + tick - 1) / tick;
1489 callout_reset(&cl->callout_, t, rmc_restart, cl);
1495 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1496 * called by the system timer code & is responsible checking if the
1497 * class is still sleeping (it might have been restarted as a side
1498 * effect of the queue scan on a packet arrival) and, if so, restarting
1499 * output for the class. Inspecting the class state & restarting output
1500 * require locking the class structure. In general the driver is
1501 * responsible for locking but this is the only routine that is not
1502 * called directly or indirectly from the interface driver so it has
1503 * know about system locking conventions. Under bsd, locking is done
1504 * by raising IPL to splimp so that's what's implemented here. On a
1505 * different system this would probably need to be changed.
1511 rmc_restart(void *arg)
1513 struct rm_class *cl = arg;
1514 struct rm_ifdat *ifd = cl->ifdat_;
1517 if (cl->sleeping_) {
1519 cl->undertime_.tv_sec = 0;
1521 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1522 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1523 (ifd->restart)(ifd->ifq_);
1531 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1532 * handling routine for the root class of the link sharing structure.
1538 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1540 panic("rmc_root_overlimit");
1544 * Packet Queue handling routines. Eventually, this is to localize the
1545 * effects on the code whether queues are red queues or droptail
1550 _rmc_addq(rm_class_t *cl, struct mbuf *m)
1553 if (q_is_rio(cl->q_))
1554 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1557 if (q_is_red(cl->q_))
1558 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1559 #endif /* ALTQ_RED */
1561 if (cl->flags_ & RMCF_CLEARDSCP)
1562 write_dsfield(m, cl->pktattr_, 0);
1568 /* note: _rmc_dropq is not called for red */
1570 _rmc_dropq(rm_class_t *cl)
1574 if ((m = _getq(cl->q_)) != NULL)
1578 static struct mbuf *
1579 _rmc_getq(rm_class_t *cl)
1582 if (q_is_rio(cl->q_))
1583 return rio_getq((rio_t *)cl->red_, cl->q_);
1586 if (q_is_red(cl->q_))
1587 return red_getq(cl->red_, cl->q_);
1589 return _getq(cl->q_);
1592 static struct mbuf *
1593 _rmc_pollq(rm_class_t *cl)
1595 return qhead(cl->q_);
1600 * DDB hook to trace cbq events:
1601 * the last 1024 events are held in a circular buffer.
1602 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1604 void cbqtrace_dump(int);
1605 static char *rmc_funcname(void *);
1607 static struct rmc_funcs {
1611 rmc_init, "rmc_init",
1612 rmc_queue_packet, "rmc_queue_packet",
1613 rmc_under_limit, "rmc_under_limit",
1614 rmc_update_class_util, "rmc_update_class_util",
1615 rmc_delay_action, "rmc_delay_action",
1616 rmc_restart, "rmc_restart",
1617 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1621 static char *rmc_funcname(void *func)
1623 struct rmc_funcs *fp;
1625 for (fp = rmc_funcs; fp->func != NULL; fp++) {
1626 if (fp->func == func)
1634 cbqtrace_dump(int counter)
1639 counter = counter % NCBQTRACE;
1640 p = (int *)&cbqtrace_buffer[counter];
1642 for (i=0; i<20; i++) {
1643 printf("[0x%x] ", *p++);
1644 printf("%s: ", rmc_funcname((void *)*p++));
1646 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1647 printf("%d\n",*p++);
1649 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1650 p = (int *)cbqtrace_buffer;
1653 #endif /* CBQ_TRACE */
1654 #endif /* ALTQ_CBQ */