1 /* @(#)rm_class.c 1.48 97/12/05 SMI */
2 /* $KAME: altq_rmclass.c,v 1.18 2003/11/06 06:32:53 kjc 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
42 #include "opt_inet6.h"
44 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
46 #include <sys/param.h>
47 #include <sys/malloc.h>
49 #include <sys/socket.h>
50 #include <sys/systm.h>
51 #include <sys/callout.h>
52 #include <sys/errno.h>
54 #include <sys/thread.h>
55 #include <sys/thread2.h>
58 #include <net/ifq_var.h>
59 #include <net/netmsg2.h>
60 #include <net/netisr2.h>
62 #include <net/altq/altq.h>
63 #include <net/altq/altq_rmclass.h>
64 #include <net/altq/altq_rmclass_debug.h>
65 #include <net/altq/altq_red.h>
66 #include <net/altq/altq_rio.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_restart_dispatch(netmsg_t);
102 static void rmc_root_overlimit(struct rm_class *, struct rm_class *);
104 #define BORROW_OFFTIME
106 * BORROW_OFFTIME (experimental):
107 * borrow the offtime of the class borrowing from.
108 * the reason is that when its own offtime is set, the class is unable
109 * to borrow much, especially when cutoff is taking effect.
110 * but when the borrowed class is overloaded (advidle is close to minidle),
111 * use the borrowing class's offtime to avoid overload.
113 #define ADJUST_CUTOFF
115 * ADJUST_CUTOFF (experimental):
116 * if no underlimit class is found due to cutoff, increase cutoff and
117 * retry the scheduling loop.
118 * also, don't invoke delay_actions while cutoff is taking effect,
119 * since a sleeping class won't have a chance to be scheduled in the
122 * now heuristics for setting the top-level variable (cutoff_) becomes:
123 * 1. if a packet arrives for a not-overlimit class, set cutoff
124 * to the depth of the class.
125 * 2. if cutoff is i, and a packet arrives for an overlimit class
126 * with an underlimit ancestor at a lower level than i (say j),
127 * then set cutoff to j.
128 * 3. at scheduling a packet, if there is no underlimit class
129 * due to the current cutoff level, increase cutoff by 1 and
130 * then try to schedule again.
135 * rmc_newclass(...) - Create a new resource management class at priority
136 * 'pri' on the interface given by 'ifd'.
138 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
139 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
140 * than 100% of the bandwidth, this number should be the
141 * 'effective' rate for the class. Let f be the
142 * bandwidth fraction allocated to this class, and let
143 * nsPerByte be the data rate of the output link in
144 * nanoseconds/byte. Then nsecPerByte is set to
145 * nsPerByte / f. E.g., 1600 (= 800 / .5)
146 * for a class that gets 50% of an ethernet's bandwidth.
148 * action the routine to call when the class is over limit.
150 * maxq max allowable queue size for class (in packets).
152 * parent parent class pointer.
154 * borrow class to borrow from (should be either 'parent' or null).
156 * maxidle max value allowed for class 'idle' time estimate (this
157 * parameter determines how large an initial burst of packets
158 * can be before overlimit action is invoked.
160 * offtime how long 'delay' action will delay when class goes over
161 * limit (this parameter determines the steady-state burst
162 * size when a class is running over its limit).
164 * Maxidle and offtime have to be computed from the following: If the
165 * average packet size is s, the bandwidth fraction allocated to this
166 * class is f, we want to allow b packet bursts, and the gain of the
167 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
169 * ptime = s * nsPerByte * (1 - f) / f
170 * maxidle = ptime * (1 - g^b) / g^b
171 * minidle = -ptime * (1 / (f - 1))
172 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
174 * Operationally, it's convenient to specify maxidle & offtime in units
175 * independent of the link bandwidth so the maxidle & offtime passed to
176 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
177 * (The constant factor is a scale factor needed to make the parameters
178 * integers. This scaling also means that the 'unscaled' values of
179 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
180 * not nanoseconds.) Also note that the 'idle' filter computation keeps
181 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
182 * maxidle also must be scaled upward by this value. Thus, the passed
183 * values for maxidle and offtime can be computed as follows:
185 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
186 * offtime = offtime * 8 / (1000 * nsecPerByte)
188 * When USE_HRTIME is employed, then maxidle and offtime become:
189 * maxidle = maxilde * (8.0 / nsecPerByte);
190 * offtime = offtime * (8.0 / nsecPerByte);
193 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte,
194 void (*action)(rm_class_t *, rm_class_t *), int maxq,
195 struct rm_class *parent, struct rm_class *borrow, u_int maxidle,
196 int minidle, u_int offtime, int pktsize, int flags)
199 struct rm_class *peer;
201 if (pri >= RM_MAXPRIO)
204 if (flags & RMCF_RED) {
206 kprintf("rmc_newclass: RED not configured for CBQ!\n");
212 if (flags & RMCF_RIO) {
214 kprintf("rmc_newclass: RIO not configured for CBQ!\n");
220 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
221 callout_init(&cl->callout_);
222 netmsg_init(&cl->callout_nmsg_, NULL, &netisr_adone_rport,
223 MSGF_PRIORITY, rmc_restart_dispatch);
224 cl->callout_nmsg_.lmsg.u.ms_resultp = cl;
226 cl->q_ = kmalloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO);
229 * Class initialization.
231 cl->children_ = NULL;
232 cl->parent_ = parent;
233 cl->borrow_ = borrow;
237 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
240 cl->ns_per_byte_ = nsecPerByte;
242 qlimit(cl->q_) = maxq;
243 qtype(cl->q_) = Q_DROPHEAD;
247 #if 1 /* minidle is also scaled in ALTQ */
248 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
249 if (cl->minidle_ > 0)
252 cl->minidle_ = minidle;
254 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
255 if (cl->maxidle_ == 0)
257 #if 1 /* offtime is also scaled in ALTQ */
258 cl->avgidle_ = cl->maxidle_;
259 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
260 if (cl->offtime_ == 0)
264 cl->offtime_ = (offtime * nsecPerByte) / 8;
266 cl->overlimit = action;
269 if (flags & (RMCF_RED|RMCF_RIO)) {
270 int red_flags, red_pkttime;
273 if (flags & RMCF_ECN)
274 red_flags |= REDF_ECN;
276 if (flags & RMCF_CLEARDSCP)
277 red_flags |= RIOF_CLEARDSCP;
279 red_pkttime = nsecPerByte * pktsize / 1000;
281 if (flags & RMCF_RED) {
282 cl->red_ = red_alloc(0, 0,
283 qlimit(cl->q_) * 10/100,
284 qlimit(cl->q_) * 30/100,
285 red_flags, red_pkttime);
286 if (cl->red_ != NULL)
287 qtype(cl->q_) = Q_RED;
291 cl->red_ = (red_t *)rio_alloc(0, NULL,
292 red_flags, red_pkttime);
293 if (cl->red_ != NULL)
294 qtype(cl->q_) = Q_RIO;
298 #endif /* ALTQ_RED */
301 * put the class into the class tree
304 if ((peer = ifd->active_[pri]) != NULL) {
305 /* find the last class at this pri */
307 while (peer->peer_ != ifd->active_[pri])
311 ifd->active_[pri] = cl;
316 cl->next_ = parent->children_;
317 parent->children_ = cl;
322 * Compute the depth of this class and its ancestors in the class
325 rmc_depth_compute(cl);
328 * If CBQ's WRR is enabled, then initialize the class WRR state.
332 ifd->alloc_[pri] += cl->allotment_;
333 rmc_wrr_set_weights(ifd);
340 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
341 int minidle, u_int offtime, int pktsize)
343 struct rm_ifdat *ifd;
347 old_allotment = cl->allotment_;
350 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
352 cl->ns_per_byte_ = nsecPerByte;
354 qlimit(cl->q_) = maxq;
356 #if 1 /* minidle is also scaled in ALTQ */
357 cl->minidle_ = (minidle * nsecPerByte) / 8;
358 if (cl->minidle_ > 0)
361 cl->minidle_ = minidle;
363 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
364 if (cl->maxidle_ == 0)
366 #if 1 /* offtime is also scaled in ALTQ */
367 cl->avgidle_ = cl->maxidle_;
368 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
369 if (cl->offtime_ == 0)
373 cl->offtime_ = (offtime * nsecPerByte) / 8;
377 * If CBQ's WRR is enabled, then initialize the class WRR state.
380 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
381 rmc_wrr_set_weights(ifd);
389 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
390 * the appropriate run robin weights for the CBQ weighted round robin
397 rmc_wrr_set_weights(struct rm_ifdat *ifd)
400 struct rm_class *cl, *clh;
402 for (i = 0; i < RM_MAXPRIO; i++) {
404 * This is inverted from that of the simulator to
405 * maintain precision.
407 if (ifd->num_[i] == 0)
410 ifd->M_[i] = ifd->alloc_[i] /
411 (ifd->num_[i] * ifd->maxpkt_);
413 * Compute the weighted allotment for each class.
414 * This takes the expensive div instruction out
415 * of the main loop for the wrr scheduling path.
416 * These only get recomputed when a class comes or
419 if (ifd->active_[i] != NULL) {
420 clh = cl = ifd->active_[i];
422 /* safe-guard for slow link or alloc_ == 0 */
424 cl->w_allotment_ = 0;
426 cl->w_allotment_ = cl->allotment_ /
429 } while ((cl != NULL) && (cl != clh));
435 rmc_get_weight(struct rm_ifdat *ifd, int pri)
437 if ((pri >= 0) && (pri < RM_MAXPRIO))
438 return (ifd->M_[pri]);
445 * rmc_depth_compute(struct rm_class *cl) - This function computes the
446 * appropriate depth of class 'cl' and its ancestors.
452 rmc_depth_compute(struct rm_class *cl)
454 rm_class_t *t = cl, *p;
457 * Recompute the depth for the branch of the tree.
461 if (p && (t->depth_ >= p->depth_)) {
462 p->depth_ = t->depth_ + 1;
471 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
472 * the depth of the tree after a class has been deleted.
478 rmc_depth_recompute(rm_class_t *cl)
485 if ((t = p->children_) == NULL) {
491 if (t->depth_ > cdepth)
496 if (p->depth_ == cdepth + 1)
497 /* no change to this parent */
500 p->depth_ = cdepth + 1;
508 if (cl->depth_ >= 1) {
509 if (cl->children_ == NULL) {
511 } else if ((t = cl->children_) != NULL) {
513 if (t->children_ != NULL)
514 rmc_depth_recompute(t);
518 rmc_depth_compute(cl);
525 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
526 * function deletes a class from the link-sharing structure and frees
527 * all resources associated with the class.
533 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
535 struct rm_class *p, *head, *previous;
536 struct netmsg_base smsg;
537 struct ifaltq_subque *ifsq =
538 &ifd->ifq_->altq_subq[ALTQ_SUBQ_INDEX_DEFAULT];
540 KKASSERT(cl->children_ == NULL);
542 ALTQ_SQ_ASSERT_LOCKED(ifsq);
543 ALTQ_SQ_UNLOCK(ifsq);
544 callout_stop_sync(&cl->callout_);
545 /* Make sure that cl->callout_nmsg_ stops. */
546 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0,
547 netmsg_sync_handler);
548 lwkt_domsg(netisr_cpuport(0), &smsg.lmsg, 0);
553 if (ifd->pollcache_ == cl)
554 ifd->pollcache_ = NULL;
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) {
571 KKASSERT(head == cl);
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_);
631 * Free the class structure.
633 if (cl->red_ != NULL) {
635 if (q_is_rio(cl->q_))
636 rio_destroy((rio_t *)cl->red_);
639 if (q_is_red(cl->q_))
640 red_destroy(cl->red_);
643 kfree(cl->q_, M_ALTQ);
649 * rmc_init(...) - Initialize the resource management data structures
650 * associated with the output portion of interface 'ifp'. 'ifd' is
651 * where the structures will be built (for backwards compatibility, the
652 * structures aren't kept in the ifnet struct). 'nsecPerByte'
653 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
654 * 'restart' is the driver-specific routine that the generic 'delay
655 * until under limit' action will call to restart output. `maxq'
656 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
657 * is the maximum number of packets that the resource management
658 * code will allow to be queued 'downstream' (this is typically 1).
664 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
665 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
666 int minidle, u_int offtime, int flags)
671 * Initialize the CBQ tracing/debug facility.
675 bzero(ifd, sizeof (*ifd));
676 mtu = ifq->altq_ifp->if_mtu;
678 ifd->restart = restart;
679 ifd->maxqueued_ = maxqueued;
680 ifd->ns_per_byte_ = nsecPerByte;
682 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
683 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
685 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
686 if (mtu * nsecPerByte > 10 * 1000000)
687 ifd->maxiftime_ /= 4;
691 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
694 * Initialize the CBQ's WRR state.
696 for (i = 0; i < RM_MAXPRIO; i++) {
701 ifd->active_[i] = NULL;
705 * Initialize current packet state.
709 for (i = 0; i < RM_MAXQUEUED; i++) {
710 ifd->class_[i] = NULL;
712 ifd->borrowed_[i] = NULL;
716 * Create the root class of the link-sharing structure.
718 ifd->root_ = rmc_newclass(0, ifd, nsecPerByte, rmc_root_overlimit,
719 maxq, 0, 0, maxidle, minidle, offtime, 0, 0);
720 if (ifd->root_ == NULL) {
721 kprintf("rmc_init: root class not allocated\n");
724 ifd->root_->depth_ = 0;
729 * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by
730 * mbuf 'm' to queue for resource class 'cl'. This routine is called
731 * by a driver's if_output routine. This routine must be called with
732 * output packet completion interrupts locked out (to avoid racing with
735 * Returns: 0 on successful queueing
736 * -1 when packet drop occurs
739 rmc_queue_packet(struct rm_class *cl, struct mbuf *m)
742 struct rm_ifdat *ifd = cl->ifdat_;
744 int is_empty = qempty(cl->q_);
747 if (ifd->cutoff_ > 0) {
748 if (TV_LT(&cl->undertime_, &now)) {
749 if (ifd->cutoff_ > cl->depth_)
750 ifd->cutoff_ = cl->depth_;
751 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
756 * the class is overlimit. if the class has
757 * underlimit ancestors, set cutoff to the lowest
760 struct rm_class *borrow = cl->borrow_;
762 while (borrow != NULL &&
763 borrow->depth_ < ifd->cutoff_) {
764 if (TV_LT(&borrow->undertime_, &now)) {
765 ifd->cutoff_ = borrow->depth_;
766 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
769 borrow = borrow->borrow_;
773 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
774 if (TV_LT(&cl->borrow_->undertime_, &now)) {
775 ifd->cutoff_ = cl->borrow_->depth_;
776 CBQTRACE(rmc_queue_packet, 'ffob',
777 cl->borrow_->depth_);
783 if (_rmc_addq(cl, m) < 0)
788 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
792 if (qlen(cl->q_) > qlimit(cl->q_)) {
793 /* note: qlimit can be set to 0 or 1 */
802 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
803 * classes to see if there are satified.
807 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
812 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
813 if ((bp = ifd->active_[i]) != NULL) {
816 if (!rmc_satisfied(p, now)) {
817 ifd->cutoff_ = p->depth_;
829 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
833 rmc_satisfied(struct rm_class *cl, struct timeval *now)
839 if (TV_LT(now, &cl->undertime_))
841 if (cl->depth_ == 0) {
842 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
847 if (cl->children_ != NULL) {
850 if (!rmc_satisfied(p, now))
860 * Return 1 if class 'cl' is under limit or can borrow from a parent,
861 * 0 if overlimit. As a side-effect, this routine will invoke the
862 * class overlimit action if the class if overlimit.
866 rmc_under_limit(struct rm_class *cl, struct timeval *now)
870 struct rm_ifdat *ifd = cl->ifdat_;
872 ifd->borrowed_[ifd->qi_] = NULL;
874 * If cl is the root class, then always return that it is
875 * underlimit. Otherwise, check to see if the class is underlimit.
877 if (cl->parent_ == NULL)
881 if (TV_LT(now, &cl->undertime_))
884 callout_stop(&cl->callout_);
886 cl->undertime_.tv_sec = 0;
891 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
892 if (((cl = cl->borrow_) == NULL) ||
893 (cl->depth_ > ifd->cutoff_)) {
896 /* cutoff is taking effect, just
897 return false without calling
901 #ifdef BORROW_OFFTIME
903 * check if the class can borrow offtime too.
904 * borrow offtime from the top of the borrow
905 * chain if the top class is not overloaded.
908 /* cutoff is taking effect, use this class as top. */
910 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
912 if (top != NULL && top->avgidle_ == top->minidle_)
915 (p->overlimit)(p, top);
918 (p->overlimit)(p, NULL);
926 ifd->borrowed_[ifd->qi_] = cl;
931 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
932 * Packet-by-packet round robin.
934 * The heart of the weighted round-robin scheduler, which decides which
935 * class next gets to send a packet. Highest priority first, then
936 * weighted round-robin within priorites.
938 * Each able-to-send class gets to send until its byte allocation is
939 * exhausted. Thus, the active pointer is only changed after a class has
940 * exhausted its allocation.
942 * If the scheduler finds no class that is underlimit or able to borrow,
943 * then the first class found that had a nonzero queue and is allowed to
944 * borrow gets to send.
948 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
950 struct rm_class *cl = NULL, *first = NULL;
959 * if the driver polls the top of the queue and then removes
960 * the polled packet, we must return the same packet.
962 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
963 cl = ifd->pollcache_;
965 if (ifd->efficient_) {
966 /* check if this class is overlimit */
967 if (cl->undertime_.tv_sec != 0 &&
968 rmc_under_limit(cl, &now) == 0)
971 ifd->pollcache_ = NULL;
974 /* mode == ALTDQ_POLL || pollcache == NULL */
975 ifd->pollcache_ = NULL;
976 ifd->borrowed_[ifd->qi_] = NULL;
980 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
981 if (ifd->na_[cpri] == 0)
985 * Loop through twice for a priority level, if some class
986 * was unable to send a packet the first round because
987 * of the weighted round-robin mechanism.
988 * During the second loop at this level, deficit==2.
989 * (This second loop is not needed if for every class,
990 * "M[cl->pri_])" times "cl->allotment" is greater than
991 * the byte size for the largest packet in the class.)
994 cl = ifd->active_[cpri];
995 KKASSERT(cl != NULL);
997 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
998 cl->bytes_alloc_ += cl->w_allotment_;
999 if (!qempty(cl->q_)) {
1000 if ((cl->undertime_.tv_sec == 0) ||
1001 rmc_under_limit(cl, &now)) {
1002 if (cl->bytes_alloc_ > 0 || deficit > 1)
1005 /* underlimit but no alloc */
1008 ifd->borrowed_[ifd->qi_] = NULL;
1011 else if (first == NULL && cl->borrow_ != NULL)
1012 first = cl; /* borrowing candidate */
1015 cl->bytes_alloc_ = 0;
1017 } while (cl != ifd->active_[cpri]);
1020 /* first loop found an underlimit class with deficit */
1021 /* Loop on same priority level, with new deficit. */
1027 #ifdef ADJUST_CUTOFF
1029 * no underlimit class found. if cutoff is taking effect,
1030 * increase cutoff and try again.
1032 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1034 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1037 #endif /* ADJUST_CUTOFF */
1039 * If LINK_EFFICIENCY is turned on, then the first overlimit
1040 * class we encounter will send a packet if all the classes
1041 * of the link-sharing structure are overlimit.
1044 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1046 if (!ifd->efficient_ || first == NULL)
1051 #if 0 /* too time-consuming for nothing */
1053 callout_stop(&cl->callout_);
1055 cl->undertime_.tv_sec = 0;
1057 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1058 ifd->cutoff_ = cl->borrow_->depth_;
1061 * Deque the packet and do the book keeping...
1064 if (op == ALTDQ_REMOVE) {
1067 panic("_rmc_wrr_dequeue_next");
1072 * Update class statistics and link data.
1074 if (cl->bytes_alloc_ > 0)
1075 cl->bytes_alloc_ -= m_pktlen(m);
1077 if ((cl->bytes_alloc_ <= 0) || first == cl)
1078 ifd->active_[cl->pri_] = cl->peer_;
1080 ifd->active_[cl->pri_] = cl;
1082 ifd->class_[ifd->qi_] = cl;
1083 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1084 ifd->now_[ifd->qi_] = now;
1085 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1088 /* mode == ALTDQ_PPOLL */
1092 * Don't use poll cache; the poll/dequeue
1093 * model is no longer applicable to SMP
1101 * The dequeue at (+) will hit the poll
1102 * cache set by CPU-B.
1104 ifd->pollcache_ = cl;
1111 * Dequeue & return next packet from the highest priority class that
1112 * has a packet to send & has enough allocation to send it. This
1113 * routine is called by a driver whenever it needs a new packet to
1116 static struct mbuf *
1117 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1121 struct rm_class *cl, *first = NULL;
1127 * if the driver polls the top of the queue and then removes
1128 * the polled packet, we must return the same packet.
1130 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1131 cl = ifd->pollcache_;
1133 ifd->pollcache_ = NULL;
1136 /* mode == ALTDQ_POLL || pollcache == NULL */
1137 ifd->pollcache_ = NULL;
1138 ifd->borrowed_[ifd->qi_] = NULL;
1139 #ifdef ADJUST_CUTOFF
1142 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1143 if (ifd->na_[cpri] == 0)
1145 cl = ifd->active_[cpri];
1146 KKASSERT(cl != NULL);
1148 if (!qempty(cl->q_)) {
1149 if ((cl->undertime_.tv_sec == 0) ||
1150 rmc_under_limit(cl, &now))
1152 if (first == NULL && cl->borrow_ != NULL)
1156 } while (cl != ifd->active_[cpri]);
1159 #ifdef ADJUST_CUTOFF
1161 * no underlimit class found. if cutoff is taking effect, increase
1162 * cutoff and try again.
1164 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1168 #endif /* ADJUST_CUTOFF */
1170 * If LINK_EFFICIENCY is turned on, then the first overlimit
1171 * class we encounter will send a packet if all the classes
1172 * of the link-sharing structure are overlimit.
1175 if (!ifd->efficient_ || first == NULL)
1180 #if 0 /* too time-consuming for nothing */
1182 callout_stop(&cl->callout_);
1184 cl->undertime_.tv_sec = 0;
1186 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1187 ifd->cutoff_ = cl->borrow_->depth_;
1190 * Deque the packet and do the book keeping...
1193 if (op == ALTDQ_REMOVE) {
1196 panic("_rmc_prr_dequeue_next");
1200 ifd->active_[cpri] = cl->peer_;
1202 ifd->class_[ifd->qi_] = cl;
1203 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1204 ifd->now_[ifd->qi_] = now;
1205 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1208 /* mode == ALTDQ_POLL */
1212 * Don't use poll cache; the poll/dequeue
1213 * model is no longer applicable to SMP
1221 * The dequeue at (+) will hit the poll
1222 * cache set by CPU-B.
1224 ifd->pollcache_ = cl;
1232 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1233 * is invoked by the packet driver to get the next packet to be
1234 * dequeued and output on the link. If WRR is enabled, then the
1235 * WRR dequeue next routine will determine the next packet to sent.
1236 * Otherwise, packet-by-packet round robin is invoked.
1238 * Returns: NULL, if a packet is not available or if all
1239 * classes are overlimit.
1241 * Otherwise, Pointer to the next packet.
1245 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1247 if (ifd->queued_ >= ifd->maxqueued_)
1250 return (_rmc_wrr_dequeue_next(ifd, mode));
1252 return (_rmc_prr_dequeue_next(ifd, mode));
1256 * Update the utilization estimate for the packet that just completed.
1257 * The packet's class & the parent(s) of that class all get their
1258 * estimators updated. This routine is called by the driver's output-
1259 * packet-completion interrupt service routine.
1263 * a macro to approximate "divide by 1000" that gives 0.000999,
1264 * if a value has enough effective digits.
1265 * (on pentium, mul takes 9 cycles but div takes 46!)
1267 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1269 rmc_update_class_util(struct rm_ifdat *ifd)
1271 int idle, avgidle, pktlen;
1272 int pkt_time, tidle;
1273 rm_class_t *cl, *borrowed;
1274 rm_class_t *borrows;
1275 struct timeval *nowp;
1278 * Get the most recent completed class.
1280 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1283 pktlen = ifd->curlen_[ifd->qo_];
1284 borrowed = ifd->borrowed_[ifd->qo_];
1287 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1290 * Run estimator on class and its ancestors.
1293 * rm_update_class_util is designed to be called when the
1294 * transfer is completed from a xmit complete interrupt,
1295 * but most drivers don't implement an upcall for that.
1296 * so, just use estimated completion time.
1297 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1299 nowp = &ifd->now_[ifd->qo_];
1300 /* get pkt_time (for link) in usec */
1301 #if 1 /* use approximation */
1302 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1303 pkt_time = NSEC_TO_USEC(pkt_time);
1305 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1307 #if 1 /* ALTQ4PPP */
1308 if (TV_LT(nowp, &ifd->ifnow_)) {
1312 * make sure the estimated completion time does not go
1313 * too far. it can happen when the link layer supports
1314 * data compression or the interface speed is set to
1315 * a much lower value.
1317 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1318 if (iftime+pkt_time < ifd->maxiftime_) {
1319 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1321 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1324 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1327 if (TV_LT(nowp, &ifd->ifnow_)) {
1328 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1330 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1334 while (cl != NULL) {
1335 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1336 if (idle >= 2000000)
1338 * this class is idle enough, reset avgidle.
1339 * (TV_DELTA returns 2000000 us when delta is large.)
1341 cl->avgidle_ = cl->maxidle_;
1343 /* get pkt_time (for class) in usec */
1344 #if 1 /* use approximation */
1345 pkt_time = pktlen * cl->ns_per_byte_;
1346 pkt_time = NSEC_TO_USEC(pkt_time);
1348 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1352 avgidle = cl->avgidle_;
1353 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1354 cl->avgidle_ = avgidle;
1356 /* Are we overlimit ? */
1358 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1361 * need some lower bound for avgidle, otherwise
1362 * a borrowing class gets unbounded penalty.
1364 if (avgidle < cl->minidle_)
1365 avgidle = cl->avgidle_ = cl->minidle_;
1367 /* set next idle to make avgidle 0 */
1369 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1370 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1374 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1375 cl->undertime_.tv_sec = 0;
1376 if (cl->sleeping_) {
1377 callout_stop(&cl->callout_);
1382 if (borrows != NULL) {
1384 ++cl->stats_.borrows;
1388 cl->last_ = ifd->ifnow_;
1389 cl->last_pkttime_ = pkt_time;
1392 if (cl->parent_ == NULL) {
1393 /* take stats of root class */
1394 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1402 * Check to see if cutoff needs to set to a new level.
1404 cl = ifd->class_[ifd->qo_];
1405 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1407 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1408 rmc_tl_satisfied(ifd, nowp);
1409 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1411 ifd->cutoff_ = borrowed->depth_;
1412 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1415 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1418 rmc_tl_satisfied(ifd, &now);
1420 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1422 ifd->cutoff_ = borrowed->depth_;
1423 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1429 * Release class slot
1431 ifd->borrowed_[ifd->qo_] = NULL;
1432 ifd->class_[ifd->qo_] = NULL;
1433 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1439 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1440 * over-limit action routines. These get invoked by rmc_under_limit()
1441 * if a class with packets to send if over its bandwidth limit & can't
1442 * borrow from a parent class.
1448 rmc_drop_action(struct rm_class *cl)
1450 struct rm_ifdat *ifd = cl->ifdat_;
1452 KKASSERT(qlen(cl->q_) > 0);
1455 ifd->na_[cl->pri_]--;
1459 rmc_dropall(struct rm_class *cl)
1461 struct rm_ifdat *ifd = cl->ifdat_;
1463 if (!qempty(cl->q_)) {
1466 ifd->na_[cl->pri_]--;
1472 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1473 * delay action routine. It is invoked via rmc_under_limit when the
1474 * packet is discoverd to be overlimit.
1476 * If the delay action is result of borrow class being overlimit, then
1477 * delay for the offtime of the borrowing class that is overlimit.
1483 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1485 int delay, t, extradelay;
1487 cl->stats_.overactions++;
1488 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1489 #ifndef BORROW_OFFTIME
1490 delay += cl->offtime_;
1493 if (!cl->sleeping_) {
1494 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1495 #ifdef BORROW_OFFTIME
1497 extradelay = borrow->offtime_;
1500 extradelay = cl->offtime_;
1504 * XXX recalculate suspend time:
1505 * current undertime is (tidle + pkt_time) calculated
1506 * from the last transmission.
1507 * tidle: time required to bring avgidle back to 0
1508 * pkt_time: target waiting time for this class
1509 * we need to replace pkt_time by offtime
1511 extradelay -= cl->last_pkttime_;
1513 if (extradelay > 0) {
1514 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1515 delay += extradelay;
1519 cl->stats_.delays++;
1522 * Since packets are phased randomly with respect to the
1523 * clock, 1 tick (the next clock tick) can be an arbitrarily
1524 * short time so we have to wait for at least two ticks.
1525 * NOTE: If there's no other traffic, we need the timer as
1526 * a 'backstop' to restart this class.
1528 if (delay > ustick * 2)
1529 t = (delay + ustick - 1) / ustick;
1532 callout_reset_bycpu(&cl->callout_, t, rmc_restart, cl, 0);
1538 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1539 * called by the system timer code & is responsible checking if the
1540 * class is still sleeping (it might have been restarted as a side
1541 * effect of the queue scan on a packet arrival) and, if so, restarting
1542 * output for the class. Inspecting the class state & restarting output
1543 * require locking the class structure. In general the driver is
1544 * responsible for locking but this is the only routine that is not
1545 * called directly or indirectly from the interface driver so it has
1546 * know about system locking conventions. Under bsd, locking is done
1547 * by raising IPL to splimp so that's what's implemented here. On a
1548 * different system this would probably need to be changed.
1550 * Since this function is called from an independant timeout, we
1551 * have to set up the lock conditions expected for the ALTQ operation.
1552 * Note that the restart will probably fall through to an if_start.
1558 rmc_restart_dispatch(netmsg_t nmsg)
1560 struct rm_class *cl = nmsg->lmsg.u.ms_resultp;
1561 struct rm_ifdat *ifd = cl->ifdat_;
1562 struct ifaltq_subque *ifsq =
1563 &ifd->ifq_->altq_subq[ALTQ_SUBQ_INDEX_DEFAULT];
1568 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */
1572 if (cl->sleeping_) {
1574 cl->undertime_.tv_sec = 0;
1576 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1577 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1578 (ifd->restart)(ifd->ifq_);
1581 ALTQ_SQ_UNLOCK(ifsq);
1585 rmc_restart(void *xcl)
1587 struct rm_class *cl = xcl;
1588 struct lwkt_msg *lmsg = &cl->callout_nmsg_.lmsg;
1590 KASSERT(mycpuid == 0, ("not on cpu0"));
1592 if (lmsg->ms_flags & MSGF_DONE)
1593 lwkt_sendmsg_oncpu(netisr_cpuport(0), lmsg);
1599 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1600 * handling routine for the root class of the link sharing structure.
1606 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1608 panic("rmc_root_overlimit");
1612 * Packet Queue handling routines. Eventually, this is to localize the
1613 * effects on the code whether queues are red queues or droptail
1618 _rmc_addq(rm_class_t *cl, struct mbuf *m)
1621 if (q_is_rio(cl->q_))
1622 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1625 if (q_is_red(cl->q_))
1626 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1627 #endif /* ALTQ_RED */
1629 if (cl->flags_ & RMCF_CLEARDSCP)
1630 write_dsfield(m, cl->pktattr_, 0);
1636 /* note: _rmc_dropq is not called for red */
1638 _rmc_dropq(rm_class_t *cl)
1642 if ((m = _getq(cl->q_)) != NULL)
1646 static struct mbuf *
1647 _rmc_getq(rm_class_t *cl)
1650 if (q_is_rio(cl->q_))
1651 return rio_getq((rio_t *)cl->red_, cl->q_);
1654 if (q_is_red(cl->q_))
1655 return red_getq(cl->red_, cl->q_);
1657 return _getq(cl->q_);
1660 static struct mbuf *
1661 _rmc_pollq(rm_class_t *cl)
1663 return qhead(cl->q_);
1668 * DDB hook to trace cbq events:
1669 * the last 1024 events are held in a circular buffer.
1670 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1672 void cbqtrace_dump(int);
1673 static char *rmc_funcname(void *);
1675 static struct rmc_funcs {
1679 rmc_init, "rmc_init",
1680 rmc_queue_packet, "rmc_queue_packet",
1681 rmc_under_limit, "rmc_under_limit",
1682 rmc_update_class_util, "rmc_update_class_util",
1683 rmc_delay_action, "rmc_delay_action",
1684 rmc_restart, "rmc_restart",
1685 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1690 rmc_funcname(void *func)
1692 struct rmc_funcs *fp;
1694 for (fp = rmc_funcs; fp->func != NULL; fp++) {
1695 if (fp->func == func)
1703 cbqtrace_dump(int counter)
1708 counter = counter % NCBQTRACE;
1709 p = (int *)&cbqtrace_buffer[counter];
1711 for (i=0; i<20; i++) {
1712 kprintf("[0x%x] ", *p++);
1713 kprintf("%s: ", rmc_funcname((void *)*p++));
1715 kprintf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1716 kprintf("%d\n",*p++);
1718 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1719 p = (int *)cbqtrace_buffer;
1722 #endif /* CBQ_TRACE */
1723 #endif /* ALTQ_CBQ */