2 /* $KAME: altq_rmclass.c,v 1.19 2005/04/13 03:44:25 suz 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
39 * @(#)rm_class.c 1.48 97/12/05 SMI
41 #if defined(__FreeBSD__) || defined(__NetBSD__)
45 #include "opt_inet6.h"
47 #endif /* __FreeBSD__ || __NetBSD__ */
48 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */
50 #include <sys/param.h>
51 #include <sys/malloc.h>
53 #include <sys/socket.h>
54 #include <sys/systm.h>
55 #include <sys/errno.h>
58 #include <sys/kernel.h>
63 #include <netinet/in.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
68 #include <altq/altq.h>
69 #include <altq/altq_rmclass.h>
70 #include <altq/altq_rmclass_debug.h>
71 #include <altq/altq_red.h>
72 #include <altq/altq_rio.h>
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 mbuf_t *_rmc_wrr_dequeue_next(struct rm_ifdat *, int);
90 static mbuf_t *_rmc_prr_dequeue_next(struct rm_ifdat *, int);
92 static int _rmc_addq(rm_class_t *, mbuf_t *);
93 static void _rmc_dropq(rm_class_t *);
94 static mbuf_t *_rmc_getq(rm_class_t *);
95 static mbuf_t *_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(struct rm_class *);
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(struct rm_class), M_DEVBUF, M_NOWAIT | M_ZERO);
223 CALLOUT_INIT(&cl->callout_);
224 cl->q_ = malloc(sizeof(class_queue_t), M_DEVBUF, M_NOWAIT | M_ZERO);
225 if (cl->q_ == NULL) {
231 * Class initialization.
233 cl->children_ = NULL;
234 cl->parent_ = parent;
235 cl->borrow_ = borrow;
239 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
242 cl->ns_per_byte_ = nsecPerByte;
244 qlimit(cl->q_) = maxq;
245 qtype(cl->q_) = Q_DROPHEAD;
249 #if 1 /* minidle is also scaled in ALTQ */
250 cl->minidle_ = (minidle * (int)nsecPerByte) / 8;
251 if (cl->minidle_ > 0)
254 cl->minidle_ = minidle;
256 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
257 if (cl->maxidle_ == 0)
259 #if 1 /* offtime is also scaled in ALTQ */
260 cl->avgidle_ = cl->maxidle_;
261 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
262 if (cl->offtime_ == 0)
266 cl->offtime_ = (offtime * nsecPerByte) / 8;
268 cl->overlimit = action;
271 if (flags & (RMCF_RED|RMCF_RIO)) {
272 int red_flags, red_pkttime;
275 if (flags & RMCF_ECN)
276 red_flags |= REDF_ECN;
277 if (flags & RMCF_FLOWVALVE)
278 red_flags |= REDF_FLOWVALVE;
280 if (flags & RMCF_CLEARDSCP)
281 red_flags |= RIOF_CLEARDSCP;
283 red_pkttime = nsecPerByte * pktsize / 1000;
285 if (flags & RMCF_RED) {
286 cl->red_ = red_alloc(0, 0,
287 qlimit(cl->q_) * 10/100,
288 qlimit(cl->q_) * 30/100,
289 red_flags, red_pkttime);
290 if (cl->red_ != NULL)
291 qtype(cl->q_) = Q_RED;
295 cl->red_ = (red_t *)rio_alloc(0, NULL,
296 red_flags, red_pkttime);
297 if (cl->red_ != NULL)
298 qtype(cl->q_) = Q_RIO;
302 #endif /* ALTQ_RED */
305 * put the class into the class tree
313 if ((peer = ifd->active_[pri]) != NULL) {
314 /* find the last class at this pri */
316 while (peer->peer_ != ifd->active_[pri])
320 ifd->active_[pri] = cl;
325 cl->next_ = parent->children_;
326 parent->children_ = cl;
331 * Compute the depth of this class and its ancestors in the class
334 rmc_depth_compute(cl);
337 * If CBQ's WRR is enabled, then initialize the class WRR state.
341 ifd->alloc_[pri] += cl->allotment_;
342 rmc_wrr_set_weights(ifd);
344 IFQ_UNLOCK(ifd->ifq_);
350 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle,
351 int minidle, u_int offtime, int pktsize)
353 struct rm_ifdat *ifd;
358 old_allotment = cl->allotment_;
366 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */
368 cl->ns_per_byte_ = nsecPerByte;
370 qlimit(cl->q_) = maxq;
372 #if 1 /* minidle is also scaled in ALTQ */
373 cl->minidle_ = (minidle * nsecPerByte) / 8;
374 if (cl->minidle_ > 0)
377 cl->minidle_ = minidle;
379 cl->maxidle_ = (maxidle * nsecPerByte) / 8;
380 if (cl->maxidle_ == 0)
382 #if 1 /* offtime is also scaled in ALTQ */
383 cl->avgidle_ = cl->maxidle_;
384 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN;
385 if (cl->offtime_ == 0)
389 cl->offtime_ = (offtime * nsecPerByte) / 8;
393 * If CBQ's WRR is enabled, then initialize the class WRR state.
396 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment;
397 rmc_wrr_set_weights(ifd);
399 IFQ_UNLOCK(ifd->ifq_);
406 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
407 * the appropriate run robin weights for the CBQ weighted round robin
414 rmc_wrr_set_weights(struct rm_ifdat *ifd)
417 struct rm_class *cl, *clh;
419 for (i = 0; i < RM_MAXPRIO; i++) {
421 * This is inverted from that of the simulator to
422 * maintain precision.
424 if (ifd->num_[i] == 0)
427 ifd->M_[i] = ifd->alloc_[i] /
428 (ifd->num_[i] * ifd->maxpkt_);
430 * Compute the weighted allotment for each class.
431 * This takes the expensive div instruction out
432 * of the main loop for the wrr scheduling path.
433 * These only get recomputed when a class comes or
436 if (ifd->active_[i] != NULL) {
437 clh = cl = ifd->active_[i];
439 /* safe-guard for slow link or alloc_ == 0 */
441 cl->w_allotment_ = 0;
443 cl->w_allotment_ = cl->allotment_ /
446 } while ((cl != NULL) && (cl != clh));
452 rmc_get_weight(struct rm_ifdat *ifd, int pri)
454 if ((pri >= 0) && (pri < RM_MAXPRIO))
455 return (ifd->M_[pri]);
462 * rmc_depth_compute(struct rm_class *cl) - This function computes the
463 * appropriate depth of class 'cl' and its ancestors.
469 rmc_depth_compute(struct rm_class *cl)
471 rm_class_t *t = cl, *p;
474 * Recompute the depth for the branch of the tree.
478 if (p && (t->depth_ >= p->depth_)) {
479 p->depth_ = t->depth_ + 1;
488 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
489 * the depth of the tree after a class has been deleted.
495 rmc_depth_recompute(rm_class_t *cl)
502 if ((t = p->children_) == NULL) {
508 if (t->depth_ > cdepth)
513 if (p->depth_ == cdepth + 1)
514 /* no change to this parent */
517 p->depth_ = cdepth + 1;
525 if (cl->depth_ >= 1) {
526 if (cl->children_ == NULL) {
528 } else if ((t = cl->children_) != NULL) {
530 if (t->children_ != NULL)
531 rmc_depth_recompute(t);
535 rmc_depth_compute(cl);
542 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
543 * function deletes a class from the link-sharing structure and frees
544 * all resources associated with the class.
550 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl)
552 struct rm_class *p, *head, *previous;
555 ASSERT(cl->children_ == NULL);
558 CALLOUT_STOP(&cl->callout_);
567 * Free packets in the packet queue.
568 * XXX - this may not be a desired behavior. Packets should be
574 * If the class has a parent, then remove the class from the
575 * class from the parent's children chain.
577 if (cl->parent_ != NULL) {
578 head = cl->parent_->children_;
580 if (head->next_ == NULL) {
582 cl->parent_->children_ = NULL;
583 cl->parent_->leaf_ = 1;
584 } else while (p != NULL) {
587 cl->parent_->children_ = cl->next_;
589 previous->next_ = cl->next_;
600 * Delete class from class priority peer list.
602 if ((p = ifd->active_[cl->pri_]) != NULL) {
604 * If there is more than one member of this priority
605 * level, then look for class(cl) in the priority level.
608 while (p->peer_ != cl)
610 p->peer_ = cl->peer_;
612 if (ifd->active_[cl->pri_] == cl)
613 ifd->active_[cl->pri_] = cl->peer_;
616 ifd->active_[cl->pri_] = NULL;
621 * Recompute the WRR weights.
624 ifd->alloc_[cl->pri_] -= cl->allotment_;
625 ifd->num_[cl->pri_]--;
626 rmc_wrr_set_weights(ifd);
630 * Re-compute the depth of the tree.
633 rmc_depth_recompute(cl->parent_);
635 rmc_depth_recompute(ifd->root_);
638 IFQ_UNLOCK(ifd->ifq_);
642 * Free the class structure.
644 if (cl->red_ != NULL) {
646 if (q_is_rio(cl->q_))
647 rio_destroy((rio_t *)cl->red_);
650 if (q_is_red(cl->q_))
651 red_destroy(cl->red_);
654 free(cl->q_, M_DEVBUF);
661 * rmc_init(...) - Initialize the resource management data structures
662 * associated with the output portion of interface 'ifp'. 'ifd' is
663 * where the structures will be built (for backwards compatibility, the
664 * structures aren't kept in the ifnet struct). 'nsecPerByte'
665 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
666 * 'restart' is the driver-specific routine that the generic 'delay
667 * until under limit' action will call to restart output. `maxq'
668 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
669 * is the maximum number of packets that the resource management
670 * code will allow to be queued 'downstream' (this is typically 1).
676 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte,
677 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle,
678 int minidle, u_int offtime, int flags)
683 * Initialize the CBQ tracing/debug facility.
687 bzero((char *)ifd, sizeof (*ifd));
688 mtu = ifq->altq_ifp->if_mtu;
690 ifd->restart = restart;
691 ifd->maxqueued_ = maxqueued;
692 ifd->ns_per_byte_ = nsecPerByte;
694 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0;
695 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0;
697 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16;
698 if (mtu * nsecPerByte > 10 * 1000000)
699 ifd->maxiftime_ /= 4;
703 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_);
706 * Initialize the CBQ's WRR state.
708 for (i = 0; i < RM_MAXPRIO; i++) {
713 ifd->active_[i] = NULL;
717 * Initialize current packet state.
721 for (i = 0; i < RM_MAXQUEUED; i++) {
722 ifd->class_[i] = NULL;
724 ifd->borrowed_[i] = NULL;
728 * Create the root class of the link-sharing structure.
730 if ((ifd->root_ = rmc_newclass(0, ifd,
732 rmc_root_overlimit, maxq, 0, 0,
733 maxidle, minidle, offtime,
735 printf("rmc_init: root class not allocated\n");
738 ifd->root_->depth_ = 0;
743 * rmc_queue_packet(struct rm_class *cl, mbuf_t *m) - Add packet given by
744 * mbuf 'm' to queue for resource class 'cl'. This routine is called
745 * by a driver's if_output routine. This routine must be called with
746 * output packet completion interrupts locked out (to avoid racing with
749 * Returns: 0 on successful queueing
750 * -1 when packet drop occurs
753 rmc_queue_packet(struct rm_class *cl, mbuf_t *m)
756 struct rm_ifdat *ifd = cl->ifdat_;
758 int is_empty = qempty(cl->q_);
761 if (ifd->cutoff_ > 0) {
762 if (TV_LT(&cl->undertime_, &now)) {
763 if (ifd->cutoff_ > cl->depth_)
764 ifd->cutoff_ = cl->depth_;
765 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_);
770 * the class is overlimit. if the class has
771 * underlimit ancestors, set cutoff to the lowest
774 struct rm_class *borrow = cl->borrow_;
776 while (borrow != NULL &&
777 borrow->depth_ < ifd->cutoff_) {
778 if (TV_LT(&borrow->undertime_, &now)) {
779 ifd->cutoff_ = borrow->depth_;
780 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_);
783 borrow = borrow->borrow_;
787 else if ((ifd->cutoff_ > 1) && cl->borrow_) {
788 if (TV_LT(&cl->borrow_->undertime_, &now)) {
789 ifd->cutoff_ = cl->borrow_->depth_;
790 CBQTRACE(rmc_queue_packet, 'ffob',
791 cl->borrow_->depth_);
797 if (_rmc_addq(cl, m) < 0)
802 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle);
806 if (qlen(cl->q_) > qlimit(cl->q_)) {
807 /* note: qlimit can be set to 0 or 1 */
816 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
817 * classes to see if there are satified.
821 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now)
826 for (i = RM_MAXPRIO - 1; i >= 0; i--) {
827 if ((bp = ifd->active_[i]) != NULL) {
830 if (!rmc_satisfied(p, now)) {
831 ifd->cutoff_ = p->depth_;
843 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
847 rmc_satisfied(struct rm_class *cl, struct timeval *now)
853 if (TV_LT(now, &cl->undertime_))
855 if (cl->depth_ == 0) {
856 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_))
861 if (cl->children_ != NULL) {
864 if (!rmc_satisfied(p, now))
874 * Return 1 if class 'cl' is under limit or can borrow from a parent,
875 * 0 if overlimit. As a side-effect, this routine will invoke the
876 * class overlimit action if the class if overlimit.
880 rmc_under_limit(struct rm_class *cl, struct timeval *now)
884 struct rm_ifdat *ifd = cl->ifdat_;
886 ifd->borrowed_[ifd->qi_] = NULL;
888 * If cl is the root class, then always return that it is
889 * underlimit. Otherwise, check to see if the class is underlimit.
891 if (cl->parent_ == NULL)
895 if (TV_LT(now, &cl->undertime_))
898 CALLOUT_STOP(&cl->callout_);
900 cl->undertime_.tv_sec = 0;
905 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) {
906 if (((cl = cl->borrow_) == NULL) ||
907 (cl->depth_ > ifd->cutoff_)) {
910 /* cutoff is taking effect, just
911 return false without calling
915 #ifdef BORROW_OFFTIME
917 * check if the class can borrow offtime too.
918 * borrow offtime from the top of the borrow
919 * chain if the top class is not overloaded.
922 /* cutoff is taking effect, use this class as top. */
924 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_);
926 if (top != NULL && top->avgidle_ == top->minidle_)
929 (p->overlimit)(p, top);
932 (p->overlimit)(p, NULL);
940 ifd->borrowed_[ifd->qi_] = cl;
945 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
946 * Packet-by-packet round robin.
948 * The heart of the weighted round-robin scheduler, which decides which
949 * class next gets to send a packet. Highest priority first, then
950 * weighted round-robin within priorites.
952 * Each able-to-send class gets to send until its byte allocation is
953 * exhausted. Thus, the active pointer is only changed after a class has
954 * exhausted its allocation.
956 * If the scheduler finds no class that is underlimit or able to borrow,
957 * then the first class found that had a nonzero queue and is allowed to
958 * borrow gets to send.
962 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op)
964 struct rm_class *cl = NULL, *first = NULL;
973 * if the driver polls the top of the queue and then removes
974 * the polled packet, we must return the same packet.
976 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
977 cl = ifd->pollcache_;
979 if (ifd->efficient_) {
980 /* check if this class is overlimit */
981 if (cl->undertime_.tv_sec != 0 &&
982 rmc_under_limit(cl, &now) == 0)
985 ifd->pollcache_ = NULL;
989 /* mode == ALTDQ_POLL || pollcache == NULL */
990 ifd->pollcache_ = NULL;
991 ifd->borrowed_[ifd->qi_] = NULL;
996 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
997 if (ifd->na_[cpri] == 0)
1001 * Loop through twice for a priority level, if some class
1002 * was unable to send a packet the first round because
1003 * of the weighted round-robin mechanism.
1004 * During the second loop at this level, deficit==2.
1005 * (This second loop is not needed if for every class,
1006 * "M[cl->pri_])" times "cl->allotment" is greater than
1007 * the byte size for the largest packet in the class.)
1010 cl = ifd->active_[cpri];
1013 if ((deficit < 2) && (cl->bytes_alloc_ <= 0))
1014 cl->bytes_alloc_ += cl->w_allotment_;
1015 if (!qempty(cl->q_)) {
1016 if ((cl->undertime_.tv_sec == 0) ||
1017 rmc_under_limit(cl, &now)) {
1018 if (cl->bytes_alloc_ > 0 || deficit > 1)
1021 /* underlimit but no alloc */
1024 ifd->borrowed_[ifd->qi_] = NULL;
1027 else if (first == NULL && cl->borrow_ != NULL)
1028 first = cl; /* borrowing candidate */
1031 cl->bytes_alloc_ = 0;
1033 } while (cl != ifd->active_[cpri]);
1036 /* first loop found an underlimit class with deficit */
1037 /* Loop on same priority level, with new deficit. */
1043 #ifdef ADJUST_CUTOFF
1045 * no underlimit class found. if cutoff is taking effect,
1046 * increase cutoff and try again.
1048 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1050 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_);
1053 #endif /* ADJUST_CUTOFF */
1055 * If LINK_EFFICIENCY is turned on, then the first overlimit
1056 * class we encounter will send a packet if all the classes
1057 * of the link-sharing structure are overlimit.
1060 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_);
1062 if (!ifd->efficient_ || first == NULL)
1067 #if 0 /* too time-consuming for nothing */
1069 CALLOUT_STOP(&cl->callout_);
1071 cl->undertime_.tv_sec = 0;
1073 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1074 ifd->cutoff_ = cl->borrow_->depth_;
1077 * Deque the packet and do the book keeping...
1080 if (op == ALTDQ_REMOVE) {
1083 panic("_rmc_wrr_dequeue_next");
1088 * Update class statistics and link data.
1090 if (cl->bytes_alloc_ > 0)
1091 cl->bytes_alloc_ -= m_pktlen(m);
1093 if ((cl->bytes_alloc_ <= 0) || first == cl)
1094 ifd->active_[cl->pri_] = cl->peer_;
1096 ifd->active_[cl->pri_] = cl;
1098 ifd->class_[ifd->qi_] = cl;
1099 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1100 ifd->now_[ifd->qi_] = now;
1101 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1104 /* mode == ALTDQ_PPOLL */
1106 ifd->pollcache_ = cl;
1112 * Dequeue & return next packet from the highest priority class that
1113 * has a packet to send & has enough allocation to send it. This
1114 * routine is called by a driver whenever it needs a new packet to
1118 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op)
1122 struct rm_class *cl, *first = NULL;
1128 * if the driver polls the top of the queue and then removes
1129 * the polled packet, we must return the same packet.
1131 if (op == ALTDQ_REMOVE && ifd->pollcache_) {
1132 cl = ifd->pollcache_;
1134 ifd->pollcache_ = NULL;
1137 /* mode == ALTDQ_POLL || pollcache == NULL */
1138 ifd->pollcache_ = NULL;
1139 ifd->borrowed_[ifd->qi_] = NULL;
1141 #ifdef ADJUST_CUTOFF
1144 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) {
1145 if (ifd->na_[cpri] == 0)
1147 cl = ifd->active_[cpri];
1150 if (!qempty(cl->q_)) {
1151 if ((cl->undertime_.tv_sec == 0) ||
1152 rmc_under_limit(cl, &now))
1154 if (first == NULL && cl->borrow_ != NULL)
1158 } while (cl != ifd->active_[cpri]);
1161 #ifdef ADJUST_CUTOFF
1163 * no underlimit class found. if cutoff is taking effect, increase
1164 * cutoff and try again.
1166 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) {
1170 #endif /* ADJUST_CUTOFF */
1172 * If LINK_EFFICIENCY is turned on, then the first overlimit
1173 * class we encounter will send a packet if all the classes
1174 * of the link-sharing structure are overlimit.
1177 if (!ifd->efficient_ || first == NULL)
1182 #if 0 /* too time-consuming for nothing */
1184 CALLOUT_STOP(&cl->callout_);
1186 cl->undertime_.tv_sec = 0;
1188 ifd->borrowed_[ifd->qi_] = cl->borrow_;
1189 ifd->cutoff_ = cl->borrow_->depth_;
1192 * Deque the packet and do the book keeping...
1195 if (op == ALTDQ_REMOVE) {
1198 panic("_rmc_prr_dequeue_next");
1202 ifd->active_[cpri] = cl->peer_;
1204 ifd->class_[ifd->qi_] = cl;
1205 ifd->curlen_[ifd->qi_] = m_pktlen(m);
1206 ifd->now_[ifd->qi_] = now;
1207 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_;
1210 /* mode == ALTDQ_POLL */
1212 ifd->pollcache_ = cl;
1219 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1220 * is invoked by the packet driver to get the next packet to be
1221 * dequeued and output on the link. If WRR is enabled, then the
1222 * WRR dequeue next routine will determine the next packet to sent.
1223 * Otherwise, packet-by-packet round robin is invoked.
1225 * Returns: NULL, if a packet is not available or if all
1226 * classes are overlimit.
1228 * Otherwise, Pointer to the next packet.
1232 rmc_dequeue_next(struct rm_ifdat *ifd, int mode)
1234 if (ifd->queued_ >= ifd->maxqueued_)
1237 return (_rmc_wrr_dequeue_next(ifd, mode));
1239 return (_rmc_prr_dequeue_next(ifd, mode));
1243 * Update the utilization estimate for the packet that just completed.
1244 * The packet's class & the parent(s) of that class all get their
1245 * estimators updated. This routine is called by the driver's output-
1246 * packet-completion interrupt service routine.
1250 * a macro to approximate "divide by 1000" that gives 0.000999,
1251 * if a value has enough effective digits.
1252 * (on pentium, mul takes 9 cycles but div takes 46!)
1254 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1256 rmc_update_class_util(struct rm_ifdat *ifd)
1258 int idle, avgidle, pktlen;
1259 int pkt_time, tidle;
1260 rm_class_t *cl, *borrowed;
1261 rm_class_t *borrows;
1262 struct timeval *nowp;
1265 * Get the most recent completed class.
1267 if ((cl = ifd->class_[ifd->qo_]) == NULL)
1270 pktlen = ifd->curlen_[ifd->qo_];
1271 borrowed = ifd->borrowed_[ifd->qo_];
1274 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1277 * Run estimator on class and its ancestors.
1280 * rm_update_class_util is designed to be called when the
1281 * transfer is completed from a xmit complete interrupt,
1282 * but most drivers don't implement an upcall for that.
1283 * so, just use estimated completion time.
1284 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1286 nowp = &ifd->now_[ifd->qo_];
1287 /* get pkt_time (for link) in usec */
1288 #if 1 /* use approximation */
1289 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_;
1290 pkt_time = NSEC_TO_USEC(pkt_time);
1292 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000;
1294 #if 1 /* ALTQ4PPP */
1295 if (TV_LT(nowp, &ifd->ifnow_)) {
1299 * make sure the estimated completion time does not go
1300 * too far. it can happen when the link layer supports
1301 * data compression or the interface speed is set to
1302 * a much lower value.
1304 TV_DELTA(&ifd->ifnow_, nowp, iftime);
1305 if (iftime+pkt_time < ifd->maxiftime_) {
1306 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1308 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_);
1311 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1314 if (TV_LT(nowp, &ifd->ifnow_)) {
1315 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_);
1317 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_);
1321 while (cl != NULL) {
1322 TV_DELTA(&ifd->ifnow_, &cl->last_, idle);
1323 if (idle >= 2000000)
1325 * this class is idle enough, reset avgidle.
1326 * (TV_DELTA returns 2000000 us when delta is large.)
1328 cl->avgidle_ = cl->maxidle_;
1330 /* get pkt_time (for class) in usec */
1331 #if 1 /* use approximation */
1332 pkt_time = pktlen * cl->ns_per_byte_;
1333 pkt_time = NSEC_TO_USEC(pkt_time);
1335 pkt_time = pktlen * cl->ns_per_byte_ / 1000;
1339 avgidle = cl->avgidle_;
1340 avgidle += idle - (avgidle >> RM_FILTER_GAIN);
1341 cl->avgidle_ = avgidle;
1343 /* Are we overlimit ? */
1345 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle);
1348 * need some lower bound for avgidle, otherwise
1349 * a borrowing class gets unbounded penalty.
1351 if (avgidle < cl->minidle_)
1352 avgidle = cl->avgidle_ = cl->minidle_;
1354 /* set next idle to make avgidle 0 */
1356 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN);
1357 TV_ADD_DELTA(nowp, tidle, &cl->undertime_);
1361 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle;
1362 cl->undertime_.tv_sec = 0;
1363 if (cl->sleeping_) {
1364 CALLOUT_STOP(&cl->callout_);
1369 if (borrows != NULL) {
1371 ++cl->stats_.borrows;
1375 cl->last_ = ifd->ifnow_;
1376 cl->last_pkttime_ = pkt_time;
1379 if (cl->parent_ == NULL) {
1380 /* take stats of root class */
1381 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen);
1389 * Check to see if cutoff needs to set to a new level.
1391 cl = ifd->class_[ifd->qo_];
1392 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) {
1394 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) {
1395 rmc_tl_satisfied(ifd, nowp);
1396 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1398 ifd->cutoff_ = borrowed->depth_;
1399 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1402 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) {
1405 rmc_tl_satisfied(ifd, &now);
1407 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_);
1409 ifd->cutoff_ = borrowed->depth_;
1410 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_);
1416 * Release class slot
1418 ifd->borrowed_[ifd->qo_] = NULL;
1419 ifd->class_[ifd->qo_] = NULL;
1420 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_;
1426 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1427 * over-limit action routines. These get invoked by rmc_under_limit()
1428 * if a class with packets to send if over its bandwidth limit & can't
1429 * borrow from a parent class.
1435 rmc_drop_action(struct rm_class *cl)
1437 struct rm_ifdat *ifd = cl->ifdat_;
1439 ASSERT(qlen(cl->q_) > 0);
1442 ifd->na_[cl->pri_]--;
1445 void rmc_dropall(struct rm_class *cl)
1447 struct rm_ifdat *ifd = cl->ifdat_;
1449 if (!qempty(cl->q_)) {
1452 ifd->na_[cl->pri_]--;
1456 #if (__FreeBSD_version > 300000)
1457 /* hzto() is removed from FreeBSD-3.0 */
1458 static int hzto(struct timeval *);
1467 t2.tv_sec = tv->tv_sec - t2.tv_sec;
1468 t2.tv_usec = tv->tv_usec - t2.tv_usec;
1469 return (tvtohz(&t2));
1471 #endif /* __FreeBSD_version > 300000 */
1475 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1476 * delay action routine. It is invoked via rmc_under_limit when the
1477 * packet is discoverd to be overlimit.
1479 * If the delay action is result of borrow class being overlimit, then
1480 * delay for the offtime of the borrowing class that is overlimit.
1486 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow)
1488 int delay, t, extradelay;
1490 cl->stats_.overactions++;
1491 TV_DELTA(&cl->undertime_, &cl->overtime_, delay);
1492 #ifndef BORROW_OFFTIME
1493 delay += cl->offtime_;
1496 if (!cl->sleeping_) {
1497 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle);
1498 #ifdef BORROW_OFFTIME
1500 extradelay = borrow->offtime_;
1503 extradelay = cl->offtime_;
1507 * XXX recalculate suspend time:
1508 * current undertime is (tidle + pkt_time) calculated
1509 * from the last transmission.
1510 * tidle: time required to bring avgidle back to 0
1511 * pkt_time: target waiting time for this class
1512 * we need to replace pkt_time by offtime
1514 extradelay -= cl->last_pkttime_;
1516 if (extradelay > 0) {
1517 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_);
1518 delay += extradelay;
1522 cl->stats_.delays++;
1525 * Since packets are phased randomly with respect to the
1526 * clock, 1 tick (the next clock tick) can be an arbitrarily
1527 * short time so we have to wait for at least two ticks.
1528 * NOTE: If there's no other traffic, we need the timer as
1529 * a 'backstop' to restart this class.
1531 if (delay > tick * 2) {
1533 /* FreeBSD rounds up the tick */
1534 t = hzto(&cl->undertime_);
1536 /* other BSDs round down the tick */
1537 t = hzto(&cl->undertime_) + 1;
1541 CALLOUT_RESET(&cl->callout_, t,
1542 (timeout_t *)rmc_restart, (caddr_t)cl);
1548 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1549 * called by the system timer code & is responsible checking if the
1550 * class is still sleeping (it might have been restarted as a side
1551 * effect of the queue scan on a packet arrival) and, if so, restarting
1552 * output for the class. Inspecting the class state & restarting output
1553 * require locking the class structure. In general the driver is
1554 * responsible for locking but this is the only routine that is not
1555 * called directly or indirectly from the interface driver so it has
1556 * know about system locking conventions. Under bsd, locking is done
1557 * by raising IPL to splimp so that's what's implemented here. On a
1558 * different system this would probably need to be changed.
1564 rmc_restart(struct rm_class *cl)
1566 struct rm_ifdat *ifd = cl->ifdat_;
1574 IFQ_LOCK(ifd->ifq_);
1575 if (cl->sleeping_) {
1577 cl->undertime_.tv_sec = 0;
1579 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) {
1580 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle);
1581 (ifd->restart)(ifd->ifq_);
1584 IFQ_UNLOCK(ifd->ifq_);
1590 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1591 * handling routine for the root class of the link sharing structure.
1597 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow)
1599 panic("rmc_root_overlimit");
1603 * Packet Queue handling routines. Eventually, this is to localize the
1604 * effects on the code whether queues are red queues or droptail
1609 _rmc_addq(rm_class_t *cl, mbuf_t *m)
1612 if (q_is_rio(cl->q_))
1613 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_);
1616 if (q_is_red(cl->q_))
1617 return red_addq(cl->red_, cl->q_, m, cl->pktattr_);
1618 #endif /* ALTQ_RED */
1620 if (cl->flags_ & RMCF_CLEARDSCP)
1621 write_dsfield(m, cl->pktattr_, 0);
1627 /* note: _rmc_dropq is not called for red */
1629 _rmc_dropq(rm_class_t *cl)
1633 if ((m = _getq(cl->q_)) != NULL)
1638 _rmc_getq(rm_class_t *cl)
1641 if (q_is_rio(cl->q_))
1642 return rio_getq((rio_t *)cl->red_, cl->q_);
1645 if (q_is_red(cl->q_))
1646 return red_getq(cl->red_, cl->q_);
1648 return _getq(cl->q_);
1652 _rmc_pollq(rm_class_t *cl)
1654 return qhead(cl->q_);
1659 struct cbqtrace cbqtrace_buffer[NCBQTRACE+1];
1660 struct cbqtrace *cbqtrace_ptr = NULL;
1664 * DDB hook to trace cbq events:
1665 * the last 1024 events are held in a circular buffer.
1666 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1668 void cbqtrace_dump(int);
1669 static char *rmc_funcname(void *);
1671 static struct rmc_funcs {
1676 rmc_init, "rmc_init",
1677 rmc_queue_packet, "rmc_queue_packet",
1678 rmc_under_limit, "rmc_under_limit",
1679 rmc_update_class_util, "rmc_update_class_util",
1680 rmc_delay_action, "rmc_delay_action",
1681 rmc_restart, "rmc_restart",
1682 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next",
1686 static char *rmc_funcname(void *func)
1688 struct rmc_funcs *fp;
1690 for (fp = rmc_funcs; fp->func != NULL; fp++)
1691 if (fp->func == func)
1696 void cbqtrace_dump(int counter)
1701 counter = counter % NCBQTRACE;
1702 p = (int *)&cbqtrace_buffer[counter];
1704 for (i=0; i<20; i++) {
1705 printf("[0x%x] ", *p++);
1706 printf("%s: ", rmc_funcname((void *)*p++));
1708 printf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]);
1709 printf("%d\n",*p++);
1711 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE])
1712 p = (int *)cbqtrace_buffer;
1715 #endif /* CBQ_TRACE */
1716 #endif /* ALTQ_CBQ */
1718 #if defined(ALTQ_CBQ) || defined(ALTQ_RED) || defined(ALTQ_RIO) || defined(ALTQ_HFSC) || defined(ALTQ_PRIQ)
1719 #if !defined(__GNUC__) || defined(ALTQ_DEBUG)
1722 _addq(class_queue_t *q, mbuf_t *m)
1726 if ((m0 = qtail(q)) != NULL)
1727 m->m_nextpkt = m0->m_nextpkt;
1736 _getq(class_queue_t *q)
1740 if ((m = qtail(q)) == NULL)
1742 if ((m0 = m->m_nextpkt) != m)
1743 m->m_nextpkt = m0->m_nextpkt;
1745 ASSERT(qlen(q) == 1);
1749 m0->m_nextpkt = NULL;
1753 /* drop a packet at the tail of the queue */
1755 _getq_tail(class_queue_t *q)
1757 mbuf_t *m, *m0, *prev;
1759 if ((m = m0 = qtail(q)) == NULL)
1765 prev->m_nextpkt = m->m_nextpkt;
1767 ASSERT(qlen(q) == 1);
1772 m->m_nextpkt = NULL;
1776 /* randomly select a packet in the queue */
1778 _getq_random(class_queue_t *q)
1783 if ((m = qtail(q)) == NULL)
1785 if (m->m_nextpkt == m) {
1786 ASSERT(qlen(q) == 1);
1789 struct mbuf *prev = NULL;
1791 n = arc4random() % qlen(q) + 1;
1792 for (i = 0; i < n; i++) {
1796 prev->m_nextpkt = m->m_nextpkt;
1801 m->m_nextpkt = NULL;
1806 _removeq(class_queue_t *q, mbuf_t *m)
1815 prev->m_nextpkt = m->m_nextpkt;
1818 else if (qtail(q) == m)
1824 _flushq(class_queue_t *q)
1828 while ((m = _getq(q)) != NULL)
1830 ASSERT(qlen(q) == 0);
1833 #endif /* !__GNUC__ || ALTQ_DEBUG */
1834 #endif /* ALTQ_CBQ || ALTQ_RED || ALTQ_RIO || ALTQ_HFSC || ALTQ_PRIQ */