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>
58 #include <net/altq/altq.h>
59 #include <net/altq/altq_rmclass.h>
60 #include <net/altq/altq_rmclass_debug.h>
61 #include <net/altq/altq_red.h>
62 #include <net/altq/altq_rio.h>
64 #include <sys/thread2.h>
67 static struct cbqtrace cbqtrace_buffer
[NCBQTRACE
+1];
68 static struct cbqtrace
*cbqtrace_ptr
= NULL
;
69 static int cbqtrace_count
;
76 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; }
82 static int rmc_satisfied(struct rm_class
*, struct timeval
*);
83 static void rmc_wrr_set_weights(struct rm_ifdat
*);
84 static void rmc_depth_compute(struct rm_class
*);
85 static void rmc_depth_recompute(rm_class_t
*);
87 static struct mbuf
*_rmc_wrr_dequeue_next(struct rm_ifdat
*, int);
88 static struct mbuf
*_rmc_prr_dequeue_next(struct rm_ifdat
*, int);
90 static int _rmc_addq(rm_class_t
*, struct mbuf
*);
91 static void _rmc_dropq(rm_class_t
*);
92 static struct mbuf
*_rmc_getq(rm_class_t
*);
93 static struct mbuf
*_rmc_pollq(rm_class_t
*);
95 static int rmc_under_limit(struct rm_class
*, struct timeval
*);
96 static void rmc_tl_satisfied(struct rm_ifdat
*, struct timeval
*);
97 static void rmc_drop_action(struct rm_class
*);
98 static void rmc_restart(void *);
99 static void rmc_root_overlimit(struct rm_class
*, struct rm_class
*);
101 #define BORROW_OFFTIME
103 * BORROW_OFFTIME (experimental):
104 * borrow the offtime of the class borrowing from.
105 * the reason is that when its own offtime is set, the class is unable
106 * to borrow much, especially when cutoff is taking effect.
107 * but when the borrowed class is overloaded (advidle is close to minidle),
108 * use the borrowing class's offtime to avoid overload.
110 #define ADJUST_CUTOFF
112 * ADJUST_CUTOFF (experimental):
113 * if no underlimit class is found due to cutoff, increase cutoff and
114 * retry the scheduling loop.
115 * also, don't invoke delay_actions while cutoff is taking effect,
116 * since a sleeping class won't have a chance to be scheduled in the
119 * now heuristics for setting the top-level variable (cutoff_) becomes:
120 * 1. if a packet arrives for a not-overlimit class, set cutoff
121 * to the depth of the class.
122 * 2. if cutoff is i, and a packet arrives for an overlimit class
123 * with an underlimit ancestor at a lower level than i (say j),
124 * then set cutoff to j.
125 * 3. at scheduling a packet, if there is no underlimit class
126 * due to the current cutoff level, increase cutoff by 1 and
127 * then try to schedule again.
132 * rmc_newclass(...) - Create a new resource management class at priority
133 * 'pri' on the interface given by 'ifd'.
135 * nsecPerByte is the data rate of the interface in nanoseconds/byte.
136 * E.g., 800 for a 10Mb/s ethernet. If the class gets less
137 * than 100% of the bandwidth, this number should be the
138 * 'effective' rate for the class. Let f be the
139 * bandwidth fraction allocated to this class, and let
140 * nsPerByte be the data rate of the output link in
141 * nanoseconds/byte. Then nsecPerByte is set to
142 * nsPerByte / f. E.g., 1600 (= 800 / .5)
143 * for a class that gets 50% of an ethernet's bandwidth.
145 * action the routine to call when the class is over limit.
147 * maxq max allowable queue size for class (in packets).
149 * parent parent class pointer.
151 * borrow class to borrow from (should be either 'parent' or null).
153 * maxidle max value allowed for class 'idle' time estimate (this
154 * parameter determines how large an initial burst of packets
155 * can be before overlimit action is invoked.
157 * offtime how long 'delay' action will delay when class goes over
158 * limit (this parameter determines the steady-state burst
159 * size when a class is running over its limit).
161 * Maxidle and offtime have to be computed from the following: If the
162 * average packet size is s, the bandwidth fraction allocated to this
163 * class is f, we want to allow b packet bursts, and the gain of the
164 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then:
166 * ptime = s * nsPerByte * (1 - f) / f
167 * maxidle = ptime * (1 - g^b) / g^b
168 * minidle = -ptime * (1 / (f - 1))
169 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1)
171 * Operationally, it's convenient to specify maxidle & offtime in units
172 * independent of the link bandwidth so the maxidle & offtime passed to
173 * this routine are the above values multiplied by 8*f/(1000*nsPerByte).
174 * (The constant factor is a scale factor needed to make the parameters
175 * integers. This scaling also means that the 'unscaled' values of
176 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds,
177 * not nanoseconds.) Also note that the 'idle' filter computation keeps
178 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of
179 * maxidle also must be scaled upward by this value. Thus, the passed
180 * values for maxidle and offtime can be computed as follows:
182 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte)
183 * offtime = offtime * 8 / (1000 * nsecPerByte)
185 * When USE_HRTIME is employed, then maxidle and offtime become:
186 * maxidle = maxilde * (8.0 / nsecPerByte);
187 * offtime = offtime * (8.0 / nsecPerByte);
190 rmc_newclass(int pri
, struct rm_ifdat
*ifd
, u_int nsecPerByte
,
191 void (*action
)(rm_class_t
*, rm_class_t
*), int maxq
,
192 struct rm_class
*parent
, struct rm_class
*borrow
, u_int maxidle
,
193 int minidle
, u_int offtime
, int pktsize
, int flags
)
196 struct rm_class
*peer
;
198 if (pri
>= RM_MAXPRIO
)
201 if (flags
& RMCF_RED
) {
203 kprintf("rmc_newclass: RED not configured for CBQ!\n");
209 if (flags
& RMCF_RIO
) {
211 kprintf("rmc_newclass: RIO not configured for CBQ!\n");
217 cl
= kmalloc(sizeof(*cl
), M_ALTQ
, M_WAITOK
| M_ZERO
);
218 callout_init(&cl
->callout_
);
219 cl
->q_
= kmalloc(sizeof(*cl
->q_
), M_ALTQ
, M_WAITOK
| M_ZERO
);
222 * Class initialization.
224 cl
->children_
= NULL
;
225 cl
->parent_
= parent
;
226 cl
->borrow_
= borrow
;
230 cl
->allotment_
= RM_NS_PER_SEC
/ nsecPerByte
; /* Bytes per sec */
233 cl
->ns_per_byte_
= nsecPerByte
;
235 qlimit(cl
->q_
) = maxq
;
236 qtype(cl
->q_
) = Q_DROPHEAD
;
240 #if 1 /* minidle is also scaled in ALTQ */
241 cl
->minidle_
= (minidle
* (int)nsecPerByte
) / 8;
242 if (cl
->minidle_
> 0)
245 cl
->minidle_
= minidle
;
247 cl
->maxidle_
= (maxidle
* nsecPerByte
) / 8;
248 if (cl
->maxidle_
== 0)
250 #if 1 /* offtime is also scaled in ALTQ */
251 cl
->avgidle_
= cl
->maxidle_
;
252 cl
->offtime_
= ((offtime
* nsecPerByte
) / 8) >> RM_FILTER_GAIN
;
253 if (cl
->offtime_
== 0)
257 cl
->offtime_
= (offtime
* nsecPerByte
) / 8;
259 cl
->overlimit
= action
;
262 if (flags
& (RMCF_RED
|RMCF_RIO
)) {
263 int red_flags
, red_pkttime
;
266 if (flags
& RMCF_ECN
)
267 red_flags
|= REDF_ECN
;
269 if (flags
& RMCF_CLEARDSCP
)
270 red_flags
|= RIOF_CLEARDSCP
;
272 red_pkttime
= nsecPerByte
* pktsize
/ 1000;
274 if (flags
& RMCF_RED
) {
275 cl
->red_
= red_alloc(0, 0,
276 qlimit(cl
->q_
) * 10/100,
277 qlimit(cl
->q_
) * 30/100,
278 red_flags
, red_pkttime
);
279 if (cl
->red_
!= NULL
)
280 qtype(cl
->q_
) = Q_RED
;
284 cl
->red_
= (red_t
*)rio_alloc(0, NULL
,
285 red_flags
, red_pkttime
);
286 if (cl
->red_
!= NULL
)
287 qtype(cl
->q_
) = Q_RIO
;
291 #endif /* ALTQ_RED */
294 * put the class into the class tree
297 if ((peer
= ifd
->active_
[pri
]) != NULL
) {
298 /* find the last class at this pri */
300 while (peer
->peer_
!= ifd
->active_
[pri
])
304 ifd
->active_
[pri
] = cl
;
309 cl
->next_
= parent
->children_
;
310 parent
->children_
= cl
;
315 * Compute the depth of this class and its ancestors in the class
318 rmc_depth_compute(cl
);
321 * If CBQ's WRR is enabled, then initialize the class WRR state.
325 ifd
->alloc_
[pri
] += cl
->allotment_
;
326 rmc_wrr_set_weights(ifd
);
333 rmc_modclass(struct rm_class
*cl
, u_int nsecPerByte
, int maxq
, u_int maxidle
,
334 int minidle
, u_int offtime
, int pktsize
)
336 struct rm_ifdat
*ifd
;
340 old_allotment
= cl
->allotment_
;
343 cl
->allotment_
= RM_NS_PER_SEC
/ nsecPerByte
; /* Bytes per sec */
345 cl
->ns_per_byte_
= nsecPerByte
;
347 qlimit(cl
->q_
) = maxq
;
349 #if 1 /* minidle is also scaled in ALTQ */
350 cl
->minidle_
= (minidle
* nsecPerByte
) / 8;
351 if (cl
->minidle_
> 0)
354 cl
->minidle_
= minidle
;
356 cl
->maxidle_
= (maxidle
* nsecPerByte
) / 8;
357 if (cl
->maxidle_
== 0)
359 #if 1 /* offtime is also scaled in ALTQ */
360 cl
->avgidle_
= cl
->maxidle_
;
361 cl
->offtime_
= ((offtime
* nsecPerByte
) / 8) >> RM_FILTER_GAIN
;
362 if (cl
->offtime_
== 0)
366 cl
->offtime_
= (offtime
* nsecPerByte
) / 8;
370 * If CBQ's WRR is enabled, then initialize the class WRR state.
373 ifd
->alloc_
[cl
->pri_
] += cl
->allotment_
- old_allotment
;
374 rmc_wrr_set_weights(ifd
);
382 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes
383 * the appropriate run robin weights for the CBQ weighted round robin
390 rmc_wrr_set_weights(struct rm_ifdat
*ifd
)
393 struct rm_class
*cl
, *clh
;
395 for (i
= 0; i
< RM_MAXPRIO
; i
++) {
397 * This is inverted from that of the simulator to
398 * maintain precision.
400 if (ifd
->num_
[i
] == 0)
403 ifd
->M_
[i
] = ifd
->alloc_
[i
] /
404 (ifd
->num_
[i
] * ifd
->maxpkt_
);
406 * Compute the weighted allotment for each class.
407 * This takes the expensive div instruction out
408 * of the main loop for the wrr scheduling path.
409 * These only get recomputed when a class comes or
412 if (ifd
->active_
[i
] != NULL
) {
413 clh
= cl
= ifd
->active_
[i
];
415 /* safe-guard for slow link or alloc_ == 0 */
417 cl
->w_allotment_
= 0;
419 cl
->w_allotment_
= cl
->allotment_
/
422 } while ((cl
!= NULL
) && (cl
!= clh
));
428 rmc_get_weight(struct rm_ifdat
*ifd
, int pri
)
430 if ((pri
>= 0) && (pri
< RM_MAXPRIO
))
431 return (ifd
->M_
[pri
]);
438 * rmc_depth_compute(struct rm_class *cl) - This function computes the
439 * appropriate depth of class 'cl' and its ancestors.
445 rmc_depth_compute(struct rm_class
*cl
)
447 rm_class_t
*t
= cl
, *p
;
450 * Recompute the depth for the branch of the tree.
454 if (p
&& (t
->depth_
>= p
->depth_
)) {
455 p
->depth_
= t
->depth_
+ 1;
464 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes
465 * the depth of the tree after a class has been deleted.
471 rmc_depth_recompute(rm_class_t
*cl
)
478 if ((t
= p
->children_
) == NULL
) {
484 if (t
->depth_
> cdepth
)
489 if (p
->depth_
== cdepth
+ 1)
490 /* no change to this parent */
493 p
->depth_
= cdepth
+ 1;
501 if (cl
->depth_
>= 1) {
502 if (cl
->children_
== NULL
) {
504 } else if ((t
= cl
->children_
) != NULL
) {
506 if (t
->children_
!= NULL
)
507 rmc_depth_recompute(t
);
511 rmc_depth_compute(cl
);
518 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This
519 * function deletes a class from the link-sharing structure and frees
520 * all resources associated with the class.
526 rmc_delete_class(struct rm_ifdat
*ifd
, struct rm_class
*cl
)
528 struct rm_class
*p
, *head
, *previous
;
530 KKASSERT(cl
->children_
== NULL
);
533 callout_stop(&cl
->callout_
);
537 if (ifd
->pollcache_
== cl
)
538 ifd
->pollcache_
= NULL
;
541 * Free packets in the packet queue.
542 * XXX - this may not be a desired behavior. Packets should be
548 * If the class has a parent, then remove the class from the
549 * class from the parent's children chain.
551 if (cl
->parent_
!= NULL
) {
552 head
= cl
->parent_
->children_
;
554 if (head
->next_
== NULL
) {
555 KKASSERT(head
== cl
);
556 cl
->parent_
->children_
= NULL
;
557 cl
->parent_
->leaf_
= 1;
558 } else while (p
!= NULL
) {
561 cl
->parent_
->children_
= cl
->next_
;
563 previous
->next_
= cl
->next_
;
574 * Delete class from class priority peer list.
576 if ((p
= ifd
->active_
[cl
->pri_
]) != NULL
) {
578 * If there is more than one member of this priority
579 * level, then look for class(cl) in the priority level.
582 while (p
->peer_
!= cl
)
584 p
->peer_
= cl
->peer_
;
586 if (ifd
->active_
[cl
->pri_
] == cl
)
587 ifd
->active_
[cl
->pri_
] = cl
->peer_
;
590 ifd
->active_
[cl
->pri_
] = NULL
;
595 * Recompute the WRR weights.
598 ifd
->alloc_
[cl
->pri_
] -= cl
->allotment_
;
599 ifd
->num_
[cl
->pri_
]--;
600 rmc_wrr_set_weights(ifd
);
604 * Re-compute the depth of the tree.
607 rmc_depth_recompute(cl
->parent_
);
609 rmc_depth_recompute(ifd
->root_
);
615 * Free the class structure.
617 if (cl
->red_
!= NULL
) {
619 if (q_is_rio(cl
->q_
))
620 rio_destroy((rio_t
*)cl
->red_
);
623 if (q_is_red(cl
->q_
))
624 red_destroy(cl
->red_
);
627 kfree(cl
->q_
, M_ALTQ
);
633 * rmc_init(...) - Initialize the resource management data structures
634 * associated with the output portion of interface 'ifp'. 'ifd' is
635 * where the structures will be built (for backwards compatibility, the
636 * structures aren't kept in the ifnet struct). 'nsecPerByte'
637 * gives the link speed (inverse of bandwidth) in nanoseconds/byte.
638 * 'restart' is the driver-specific routine that the generic 'delay
639 * until under limit' action will call to restart output. `maxq'
640 * is the queue size of the 'link' & 'default' classes. 'maxqueued'
641 * is the maximum number of packets that the resource management
642 * code will allow to be queued 'downstream' (this is typically 1).
648 rmc_init(struct ifaltq
*ifq
, struct rm_ifdat
*ifd
, u_int nsecPerByte
,
649 void (*restart
)(struct ifaltq
*), int maxq
, int maxqueued
, u_int maxidle
,
650 int minidle
, u_int offtime
, int flags
)
655 * Initialize the CBQ tracing/debug facility.
659 bzero(ifd
, sizeof (*ifd
));
660 mtu
= ifq
->altq_ifp
->if_mtu
;
662 ifd
->restart
= restart
;
663 ifd
->maxqueued_
= maxqueued
;
664 ifd
->ns_per_byte_
= nsecPerByte
;
666 ifd
->wrr_
= (flags
& RMCF_WRR
) ? 1 : 0;
667 ifd
->efficient_
= (flags
& RMCF_EFFICIENT
) ? 1 : 0;
669 ifd
->maxiftime_
= mtu
* nsecPerByte
/ 1000 * 16;
670 if (mtu
* nsecPerByte
> 10 * 1000000)
671 ifd
->maxiftime_
/= 4;
675 CBQTRACE(rmc_init
, 'INIT', ifd
->cutoff_
);
678 * Initialize the CBQ's WRR state.
680 for (i
= 0; i
< RM_MAXPRIO
; i
++) {
685 ifd
->active_
[i
] = NULL
;
689 * Initialize current packet state.
693 for (i
= 0; i
< RM_MAXQUEUED
; i
++) {
694 ifd
->class_
[i
] = NULL
;
696 ifd
->borrowed_
[i
] = NULL
;
700 * Create the root class of the link-sharing structure.
702 ifd
->root_
= rmc_newclass(0, ifd
, nsecPerByte
, rmc_root_overlimit
,
703 maxq
, 0, 0, maxidle
, minidle
, offtime
, 0, 0);
704 if (ifd
->root_
== NULL
) {
705 kprintf("rmc_init: root class not allocated\n");
708 ifd
->root_
->depth_
= 0;
713 * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by
714 * mbuf 'm' to queue for resource class 'cl'. This routine is called
715 * by a driver's if_output routine. This routine must be called with
716 * output packet completion interrupts locked out (to avoid racing with
719 * Returns: 0 on successful queueing
720 * -1 when packet drop occurs
723 rmc_queue_packet(struct rm_class
*cl
, struct mbuf
*m
)
726 struct rm_ifdat
*ifd
= cl
->ifdat_
;
728 int is_empty
= qempty(cl
->q_
);
731 if (ifd
->cutoff_
> 0) {
732 if (TV_LT(&cl
->undertime_
, &now
)) {
733 if (ifd
->cutoff_
> cl
->depth_
)
734 ifd
->cutoff_
= cl
->depth_
;
735 CBQTRACE(rmc_queue_packet
, 'ffoc', cl
->depth_
);
740 * the class is overlimit. if the class has
741 * underlimit ancestors, set cutoff to the lowest
744 struct rm_class
*borrow
= cl
->borrow_
;
746 while (borrow
!= NULL
&&
747 borrow
->depth_
< ifd
->cutoff_
) {
748 if (TV_LT(&borrow
->undertime_
, &now
)) {
749 ifd
->cutoff_
= borrow
->depth_
;
750 CBQTRACE(rmc_queue_packet
, 'ffob', ifd
->cutoff_
);
753 borrow
= borrow
->borrow_
;
757 else if ((ifd
->cutoff_
> 1) && cl
->borrow_
) {
758 if (TV_LT(&cl
->borrow_
->undertime_
, &now
)) {
759 ifd
->cutoff_
= cl
->borrow_
->depth_
;
760 CBQTRACE(rmc_queue_packet
, 'ffob',
761 cl
->borrow_
->depth_
);
767 if (_rmc_addq(cl
, m
) < 0)
772 CBQTRACE(rmc_queue_packet
, 'ytpe', cl
->stats_
.handle
);
776 if (qlen(cl
->q_
) > qlimit(cl
->q_
)) {
777 /* note: qlimit can be set to 0 or 1 */
786 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all
787 * classes to see if there are satified.
791 rmc_tl_satisfied(struct rm_ifdat
*ifd
, struct timeval
*now
)
796 for (i
= RM_MAXPRIO
- 1; i
>= 0; i
--) {
797 if ((bp
= ifd
->active_
[i
]) != NULL
) {
800 if (!rmc_satisfied(p
, now
)) {
801 ifd
->cutoff_
= p
->depth_
;
813 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise.
817 rmc_satisfied(struct rm_class
*cl
, struct timeval
*now
)
823 if (TV_LT(now
, &cl
->undertime_
))
825 if (cl
->depth_
== 0) {
826 if (!cl
->sleeping_
&& (qlen(cl
->q_
) > cl
->qthresh_
))
831 if (cl
->children_
!= NULL
) {
834 if (!rmc_satisfied(p
, now
))
844 * Return 1 if class 'cl' is under limit or can borrow from a parent,
845 * 0 if overlimit. As a side-effect, this routine will invoke the
846 * class overlimit action if the class if overlimit.
850 rmc_under_limit(struct rm_class
*cl
, struct timeval
*now
)
854 struct rm_ifdat
*ifd
= cl
->ifdat_
;
856 ifd
->borrowed_
[ifd
->qi_
] = NULL
;
858 * If cl is the root class, then always return that it is
859 * underlimit. Otherwise, check to see if the class is underlimit.
861 if (cl
->parent_
== NULL
)
865 if (TV_LT(now
, &cl
->undertime_
))
868 callout_stop(&cl
->callout_
);
870 cl
->undertime_
.tv_sec
= 0;
875 while (cl
->undertime_
.tv_sec
&& TV_LT(now
, &cl
->undertime_
)) {
876 if (((cl
= cl
->borrow_
) == NULL
) ||
877 (cl
->depth_
> ifd
->cutoff_
)) {
880 /* cutoff is taking effect, just
881 return false without calling
885 #ifdef BORROW_OFFTIME
887 * check if the class can borrow offtime too.
888 * borrow offtime from the top of the borrow
889 * chain if the top class is not overloaded.
892 /* cutoff is taking effect, use this class as top. */
894 CBQTRACE(rmc_under_limit
, 'ffou', ifd
->cutoff_
);
896 if (top
!= NULL
&& top
->avgidle_
== top
->minidle_
)
899 (p
->overlimit
)(p
, top
);
902 (p
->overlimit
)(p
, NULL
);
910 ifd
->borrowed_
[ifd
->qi_
] = cl
;
915 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to
916 * Packet-by-packet round robin.
918 * The heart of the weighted round-robin scheduler, which decides which
919 * class next gets to send a packet. Highest priority first, then
920 * weighted round-robin within priorites.
922 * Each able-to-send class gets to send until its byte allocation is
923 * exhausted. Thus, the active pointer is only changed after a class has
924 * exhausted its allocation.
926 * If the scheduler finds no class that is underlimit or able to borrow,
927 * then the first class found that had a nonzero queue and is allowed to
928 * borrow gets to send.
932 _rmc_wrr_dequeue_next(struct rm_ifdat
*ifd
, int op
)
934 struct rm_class
*cl
= NULL
, *first
= NULL
;
943 * if the driver polls the top of the queue and then removes
944 * the polled packet, we must return the same packet.
946 if (op
== ALTDQ_REMOVE
&& ifd
->pollcache_
) {
947 cl
= ifd
->pollcache_
;
949 if (ifd
->efficient_
) {
950 /* check if this class is overlimit */
951 if (cl
->undertime_
.tv_sec
!= 0 &&
952 rmc_under_limit(cl
, &now
) == 0)
955 ifd
->pollcache_
= NULL
;
958 /* mode == ALTDQ_POLL || pollcache == NULL */
959 ifd
->pollcache_
= NULL
;
960 ifd
->borrowed_
[ifd
->qi_
] = NULL
;
964 for (cpri
= RM_MAXPRIO
- 1; cpri
>= 0; cpri
--) {
965 if (ifd
->na_
[cpri
] == 0)
969 * Loop through twice for a priority level, if some class
970 * was unable to send a packet the first round because
971 * of the weighted round-robin mechanism.
972 * During the second loop at this level, deficit==2.
973 * (This second loop is not needed if for every class,
974 * "M[cl->pri_])" times "cl->allotment" is greater than
975 * the byte size for the largest packet in the class.)
978 cl
= ifd
->active_
[cpri
];
979 KKASSERT(cl
!= NULL
);
981 if ((deficit
< 2) && (cl
->bytes_alloc_
<= 0))
982 cl
->bytes_alloc_
+= cl
->w_allotment_
;
983 if (!qempty(cl
->q_
)) {
984 if ((cl
->undertime_
.tv_sec
== 0) ||
985 rmc_under_limit(cl
, &now
)) {
986 if (cl
->bytes_alloc_
> 0 || deficit
> 1)
989 /* underlimit but no alloc */
992 ifd
->borrowed_
[ifd
->qi_
] = NULL
;
995 else if (first
== NULL
&& cl
->borrow_
!= NULL
)
996 first
= cl
; /* borrowing candidate */
999 cl
->bytes_alloc_
= 0;
1001 } while (cl
!= ifd
->active_
[cpri
]);
1004 /* first loop found an underlimit class with deficit */
1005 /* Loop on same priority level, with new deficit. */
1011 #ifdef ADJUST_CUTOFF
1013 * no underlimit class found. if cutoff is taking effect,
1014 * increase cutoff and try again.
1016 if (first
!= NULL
&& ifd
->cutoff_
< ifd
->root_
->depth_
) {
1018 CBQTRACE(_rmc_wrr_dequeue_next
, 'ojda', ifd
->cutoff_
);
1021 #endif /* ADJUST_CUTOFF */
1023 * If LINK_EFFICIENCY is turned on, then the first overlimit
1024 * class we encounter will send a packet if all the classes
1025 * of the link-sharing structure are overlimit.
1028 CBQTRACE(_rmc_wrr_dequeue_next
, 'otsr', ifd
->cutoff_
);
1030 if (!ifd
->efficient_
|| first
== NULL
)
1035 #if 0 /* too time-consuming for nothing */
1037 callout_stop(&cl
->callout_
);
1039 cl
->undertime_
.tv_sec
= 0;
1041 ifd
->borrowed_
[ifd
->qi_
] = cl
->borrow_
;
1042 ifd
->cutoff_
= cl
->borrow_
->depth_
;
1045 * Deque the packet and do the book keeping...
1048 if (op
== ALTDQ_REMOVE
) {
1051 panic("_rmc_wrr_dequeue_next");
1056 * Update class statistics and link data.
1058 if (cl
->bytes_alloc_
> 0)
1059 cl
->bytes_alloc_
-= m_pktlen(m
);
1061 if ((cl
->bytes_alloc_
<= 0) || first
== cl
)
1062 ifd
->active_
[cl
->pri_
] = cl
->peer_
;
1064 ifd
->active_
[cl
->pri_
] = cl
;
1066 ifd
->class_
[ifd
->qi_
] = cl
;
1067 ifd
->curlen_
[ifd
->qi_
] = m_pktlen(m
);
1068 ifd
->now_
[ifd
->qi_
] = now
;
1069 ifd
->qi_
= (ifd
->qi_
+ 1) % ifd
->maxqueued_
;
1072 /* mode == ALTDQ_PPOLL */
1076 * Don't use poll cache; the poll/dequeue
1077 * model is no longer applicable to SMP
1085 * The dequeue at (+) will hit the poll
1086 * cache set by CPU-B.
1088 ifd
->pollcache_
= cl
;
1095 * Dequeue & return next packet from the highest priority class that
1096 * has a packet to send & has enough allocation to send it. This
1097 * routine is called by a driver whenever it needs a new packet to
1100 static struct mbuf
*
1101 _rmc_prr_dequeue_next(struct rm_ifdat
*ifd
, int op
)
1105 struct rm_class
*cl
, *first
= NULL
;
1111 * if the driver polls the top of the queue and then removes
1112 * the polled packet, we must return the same packet.
1114 if (op
== ALTDQ_REMOVE
&& ifd
->pollcache_
) {
1115 cl
= ifd
->pollcache_
;
1117 ifd
->pollcache_
= NULL
;
1120 /* mode == ALTDQ_POLL || pollcache == NULL */
1121 ifd
->pollcache_
= NULL
;
1122 ifd
->borrowed_
[ifd
->qi_
] = NULL
;
1123 #ifdef ADJUST_CUTOFF
1126 for (cpri
= RM_MAXPRIO
- 1; cpri
>= 0; cpri
--) {
1127 if (ifd
->na_
[cpri
] == 0)
1129 cl
= ifd
->active_
[cpri
];
1130 KKASSERT(cl
!= NULL
);
1132 if (!qempty(cl
->q_
)) {
1133 if ((cl
->undertime_
.tv_sec
== 0) ||
1134 rmc_under_limit(cl
, &now
))
1136 if (first
== NULL
&& cl
->borrow_
!= NULL
)
1140 } while (cl
!= ifd
->active_
[cpri
]);
1143 #ifdef ADJUST_CUTOFF
1145 * no underlimit class found. if cutoff is taking effect, increase
1146 * cutoff and try again.
1148 if (first
!= NULL
&& ifd
->cutoff_
< ifd
->root_
->depth_
) {
1152 #endif /* ADJUST_CUTOFF */
1154 * If LINK_EFFICIENCY is turned on, then the first overlimit
1155 * class we encounter will send a packet if all the classes
1156 * of the link-sharing structure are overlimit.
1159 if (!ifd
->efficient_
|| first
== NULL
)
1164 #if 0 /* too time-consuming for nothing */
1166 callout_stop(&cl
->callout_
);
1168 cl
->undertime_
.tv_sec
= 0;
1170 ifd
->borrowed_
[ifd
->qi_
] = cl
->borrow_
;
1171 ifd
->cutoff_
= cl
->borrow_
->depth_
;
1174 * Deque the packet and do the book keeping...
1177 if (op
== ALTDQ_REMOVE
) {
1180 panic("_rmc_prr_dequeue_next");
1184 ifd
->active_
[cpri
] = cl
->peer_
;
1186 ifd
->class_
[ifd
->qi_
] = cl
;
1187 ifd
->curlen_
[ifd
->qi_
] = m_pktlen(m
);
1188 ifd
->now_
[ifd
->qi_
] = now
;
1189 ifd
->qi_
= (ifd
->qi_
+ 1) % ifd
->maxqueued_
;
1192 /* mode == ALTDQ_POLL */
1196 * Don't use poll cache; the poll/dequeue
1197 * model is no longer applicable to SMP
1205 * The dequeue at (+) will hit the poll
1206 * cache set by CPU-B.
1208 ifd
->pollcache_
= cl
;
1216 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function
1217 * is invoked by the packet driver to get the next packet to be
1218 * dequeued and output on the link. If WRR is enabled, then the
1219 * WRR dequeue next routine will determine the next packet to sent.
1220 * Otherwise, packet-by-packet round robin is invoked.
1222 * Returns: NULL, if a packet is not available or if all
1223 * classes are overlimit.
1225 * Otherwise, Pointer to the next packet.
1229 rmc_dequeue_next(struct rm_ifdat
*ifd
, int mode
)
1231 if (ifd
->queued_
>= ifd
->maxqueued_
)
1234 return (_rmc_wrr_dequeue_next(ifd
, mode
));
1236 return (_rmc_prr_dequeue_next(ifd
, mode
));
1240 * Update the utilization estimate for the packet that just completed.
1241 * The packet's class & the parent(s) of that class all get their
1242 * estimators updated. This routine is called by the driver's output-
1243 * packet-completion interrupt service routine.
1247 * a macro to approximate "divide by 1000" that gives 0.000999,
1248 * if a value has enough effective digits.
1249 * (on pentium, mul takes 9 cycles but div takes 46!)
1251 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17))
1253 rmc_update_class_util(struct rm_ifdat
*ifd
)
1255 int idle
, avgidle
, pktlen
;
1256 int pkt_time
, tidle
;
1257 rm_class_t
*cl
, *borrowed
;
1258 rm_class_t
*borrows
;
1259 struct timeval
*nowp
;
1262 * Get the most recent completed class.
1264 if ((cl
= ifd
->class_
[ifd
->qo_
]) == NULL
)
1267 pktlen
= ifd
->curlen_
[ifd
->qo_
];
1268 borrowed
= ifd
->borrowed_
[ifd
->qo_
];
1271 PKTCNTR_ADD(&cl
->stats_
.xmit_cnt
, pktlen
);
1274 * Run estimator on class and its ancestors.
1277 * rm_update_class_util is designed to be called when the
1278 * transfer is completed from a xmit complete interrupt,
1279 * but most drivers don't implement an upcall for that.
1280 * so, just use estimated completion time.
1281 * as a result, ifd->qi_ and ifd->qo_ are always synced.
1283 nowp
= &ifd
->now_
[ifd
->qo_
];
1284 /* get pkt_time (for link) in usec */
1285 #if 1 /* use approximation */
1286 pkt_time
= ifd
->curlen_
[ifd
->qo_
] * ifd
->ns_per_byte_
;
1287 pkt_time
= NSEC_TO_USEC(pkt_time
);
1289 pkt_time
= ifd
->curlen_
[ifd
->qo_
] * ifd
->ns_per_byte_
/ 1000;
1291 #if 1 /* ALTQ4PPP */
1292 if (TV_LT(nowp
, &ifd
->ifnow_
)) {
1296 * make sure the estimated completion time does not go
1297 * too far. it can happen when the link layer supports
1298 * data compression or the interface speed is set to
1299 * a much lower value.
1301 TV_DELTA(&ifd
->ifnow_
, nowp
, iftime
);
1302 if (iftime
+pkt_time
< ifd
->maxiftime_
) {
1303 TV_ADD_DELTA(&ifd
->ifnow_
, pkt_time
, &ifd
->ifnow_
);
1305 TV_ADD_DELTA(nowp
, ifd
->maxiftime_
, &ifd
->ifnow_
);
1308 TV_ADD_DELTA(nowp
, pkt_time
, &ifd
->ifnow_
);
1311 if (TV_LT(nowp
, &ifd
->ifnow_
)) {
1312 TV_ADD_DELTA(&ifd
->ifnow_
, pkt_time
, &ifd
->ifnow_
);
1314 TV_ADD_DELTA(nowp
, pkt_time
, &ifd
->ifnow_
);
1318 while (cl
!= NULL
) {
1319 TV_DELTA(&ifd
->ifnow_
, &cl
->last_
, idle
);
1320 if (idle
>= 2000000)
1322 * this class is idle enough, reset avgidle.
1323 * (TV_DELTA returns 2000000 us when delta is large.)
1325 cl
->avgidle_
= cl
->maxidle_
;
1327 /* get pkt_time (for class) in usec */
1328 #if 1 /* use approximation */
1329 pkt_time
= pktlen
* cl
->ns_per_byte_
;
1330 pkt_time
= NSEC_TO_USEC(pkt_time
);
1332 pkt_time
= pktlen
* cl
->ns_per_byte_
/ 1000;
1336 avgidle
= cl
->avgidle_
;
1337 avgidle
+= idle
- (avgidle
>> RM_FILTER_GAIN
);
1338 cl
->avgidle_
= avgidle
;
1340 /* Are we overlimit ? */
1342 CBQTRACE(rmc_update_class_util
, 'milo', cl
->stats_
.handle
);
1345 * need some lower bound for avgidle, otherwise
1346 * a borrowing class gets unbounded penalty.
1348 if (avgidle
< cl
->minidle_
)
1349 avgidle
= cl
->avgidle_
= cl
->minidle_
;
1351 /* set next idle to make avgidle 0 */
1353 (((1 - RM_POWER
) * avgidle
) >> RM_FILTER_GAIN
);
1354 TV_ADD_DELTA(nowp
, tidle
, &cl
->undertime_
);
1358 (avgidle
> cl
->maxidle_
) ? cl
->maxidle_
: avgidle
;
1359 cl
->undertime_
.tv_sec
= 0;
1360 if (cl
->sleeping_
) {
1361 callout_stop(&cl
->callout_
);
1366 if (borrows
!= NULL
) {
1368 ++cl
->stats_
.borrows
;
1372 cl
->last_
= ifd
->ifnow_
;
1373 cl
->last_pkttime_
= pkt_time
;
1376 if (cl
->parent_
== NULL
) {
1377 /* take stats of root class */
1378 PKTCNTR_ADD(&cl
->stats_
.xmit_cnt
, pktlen
);
1386 * Check to see if cutoff needs to set to a new level.
1388 cl
= ifd
->class_
[ifd
->qo_
];
1389 if (borrowed
&& (ifd
->cutoff_
>= borrowed
->depth_
)) {
1391 if ((qlen(cl
->q_
) <= 0) || TV_LT(nowp
, &borrowed
->undertime_
)) {
1392 rmc_tl_satisfied(ifd
, nowp
);
1393 CBQTRACE(rmc_update_class_util
, 'broe', ifd
->cutoff_
);
1395 ifd
->cutoff_
= borrowed
->depth_
;
1396 CBQTRACE(rmc_update_class_util
, 'ffob', borrowed
->depth_
);
1399 if ((qlen(cl
->q_
) <= 1) || TV_LT(&now
, &borrowed
->undertime_
)) {
1402 rmc_tl_satisfied(ifd
, &now
);
1404 CBQTRACE(rmc_update_class_util
, 'broe', ifd
->cutoff_
);
1406 ifd
->cutoff_
= borrowed
->depth_
;
1407 CBQTRACE(rmc_update_class_util
, 'ffob', borrowed
->depth_
);
1413 * Release class slot
1415 ifd
->borrowed_
[ifd
->qo_
] = NULL
;
1416 ifd
->class_
[ifd
->qo_
] = NULL
;
1417 ifd
->qo_
= (ifd
->qo_
+ 1) % ifd
->maxqueued_
;
1423 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific)
1424 * over-limit action routines. These get invoked by rmc_under_limit()
1425 * if a class with packets to send if over its bandwidth limit & can't
1426 * borrow from a parent class.
1432 rmc_drop_action(struct rm_class
*cl
)
1434 struct rm_ifdat
*ifd
= cl
->ifdat_
;
1436 KKASSERT(qlen(cl
->q_
) > 0);
1439 ifd
->na_
[cl
->pri_
]--;
1443 rmc_dropall(struct rm_class
*cl
)
1445 struct rm_ifdat
*ifd
= cl
->ifdat_
;
1447 if (!qempty(cl
->q_
)) {
1450 ifd
->na_
[cl
->pri_
]--;
1456 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ
1457 * delay action routine. It is invoked via rmc_under_limit when the
1458 * packet is discoverd to be overlimit.
1460 * If the delay action is result of borrow class being overlimit, then
1461 * delay for the offtime of the borrowing class that is overlimit.
1467 rmc_delay_action(struct rm_class
*cl
, struct rm_class
*borrow
)
1469 int delay
, t
, extradelay
;
1471 cl
->stats_
.overactions
++;
1472 TV_DELTA(&cl
->undertime_
, &cl
->overtime_
, delay
);
1473 #ifndef BORROW_OFFTIME
1474 delay
+= cl
->offtime_
;
1477 if (!cl
->sleeping_
) {
1478 CBQTRACE(rmc_delay_action
, 'yled', cl
->stats_
.handle
);
1479 #ifdef BORROW_OFFTIME
1481 extradelay
= borrow
->offtime_
;
1484 extradelay
= cl
->offtime_
;
1488 * XXX recalculate suspend time:
1489 * current undertime is (tidle + pkt_time) calculated
1490 * from the last transmission.
1491 * tidle: time required to bring avgidle back to 0
1492 * pkt_time: target waiting time for this class
1493 * we need to replace pkt_time by offtime
1495 extradelay
-= cl
->last_pkttime_
;
1497 if (extradelay
> 0) {
1498 TV_ADD_DELTA(&cl
->undertime_
, extradelay
, &cl
->undertime_
);
1499 delay
+= extradelay
;
1503 cl
->stats_
.delays
++;
1506 * Since packets are phased randomly with respect to the
1507 * clock, 1 tick (the next clock tick) can be an arbitrarily
1508 * short time so we have to wait for at least two ticks.
1509 * NOTE: If there's no other traffic, we need the timer as
1510 * a 'backstop' to restart this class.
1512 if (delay
> ustick
* 2)
1513 t
= (delay
+ ustick
- 1) / ustick
;
1516 callout_reset(&cl
->callout_
, t
, rmc_restart
, cl
);
1522 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is
1523 * called by the system timer code & is responsible checking if the
1524 * class is still sleeping (it might have been restarted as a side
1525 * effect of the queue scan on a packet arrival) and, if so, restarting
1526 * output for the class. Inspecting the class state & restarting output
1527 * require locking the class structure. In general the driver is
1528 * responsible for locking but this is the only routine that is not
1529 * called directly or indirectly from the interface driver so it has
1530 * know about system locking conventions. Under bsd, locking is done
1531 * by raising IPL to splimp so that's what's implemented here. On a
1532 * different system this would probably need to be changed.
1534 * Since this function is called from an independant timeout, we
1535 * have to set up the lock conditions expected for the ALTQ operation.
1536 * Note that the restart will probably fall through to an if_start.
1542 rmc_restart(void *arg
)
1544 struct rm_class
*cl
= arg
;
1545 struct rm_ifdat
*ifd
= cl
->ifdat_
;
1546 struct ifaltq_subque
*ifsq
= &ifd
->ifq_
->altq_subq
[0];
1549 if (cl
->sleeping_
) {
1551 cl
->undertime_
.tv_sec
= 0;
1553 if (ifd
->queued_
< ifd
->maxqueued_
&& ifd
->restart
!= NULL
) {
1554 CBQTRACE(rmc_restart
, 'trts', cl
->stats_
.handle
);
1555 (ifd
->restart
)(ifd
->ifq_
);
1558 ALTQ_SQ_UNLOCK(ifsq
);
1563 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit
1564 * handling routine for the root class of the link sharing structure.
1570 rmc_root_overlimit(struct rm_class
*cl
, struct rm_class
*borrow
)
1572 panic("rmc_root_overlimit");
1576 * Packet Queue handling routines. Eventually, this is to localize the
1577 * effects on the code whether queues are red queues or droptail
1582 _rmc_addq(rm_class_t
*cl
, struct mbuf
*m
)
1585 if (q_is_rio(cl
->q_
))
1586 return rio_addq((rio_t
*)cl
->red_
, cl
->q_
, m
, cl
->pktattr_
);
1589 if (q_is_red(cl
->q_
))
1590 return red_addq(cl
->red_
, cl
->q_
, m
, cl
->pktattr_
);
1591 #endif /* ALTQ_RED */
1593 if (cl
->flags_
& RMCF_CLEARDSCP
)
1594 write_dsfield(m
, cl
->pktattr_
, 0);
1600 /* note: _rmc_dropq is not called for red */
1602 _rmc_dropq(rm_class_t
*cl
)
1606 if ((m
= _getq(cl
->q_
)) != NULL
)
1610 static struct mbuf
*
1611 _rmc_getq(rm_class_t
*cl
)
1614 if (q_is_rio(cl
->q_
))
1615 return rio_getq((rio_t
*)cl
->red_
, cl
->q_
);
1618 if (q_is_red(cl
->q_
))
1619 return red_getq(cl
->red_
, cl
->q_
);
1621 return _getq(cl
->q_
);
1624 static struct mbuf
*
1625 _rmc_pollq(rm_class_t
*cl
)
1627 return qhead(cl
->q_
);
1632 * DDB hook to trace cbq events:
1633 * the last 1024 events are held in a circular buffer.
1634 * use "call cbqtrace_dump(N)" to display 20 events from Nth event.
1636 void cbqtrace_dump(int);
1637 static char *rmc_funcname(void *);
1639 static struct rmc_funcs
{
1643 rmc_init
, "rmc_init",
1644 rmc_queue_packet
, "rmc_queue_packet",
1645 rmc_under_limit
, "rmc_under_limit",
1646 rmc_update_class_util
, "rmc_update_class_util",
1647 rmc_delay_action
, "rmc_delay_action",
1648 rmc_restart
, "rmc_restart",
1649 _rmc_wrr_dequeue_next
, "_rmc_wrr_dequeue_next",
1654 rmc_funcname(void *func
)
1656 struct rmc_funcs
*fp
;
1658 for (fp
= rmc_funcs
; fp
->func
!= NULL
; fp
++) {
1659 if (fp
->func
== func
)
1667 cbqtrace_dump(int counter
)
1672 counter
= counter
% NCBQTRACE
;
1673 p
= (int *)&cbqtrace_buffer
[counter
];
1675 for (i
=0; i
<20; i
++) {
1676 kprintf("[0x%x] ", *p
++);
1677 kprintf("%s: ", rmc_funcname((void *)*p
++));
1679 kprintf("%c%c%c%c: ", cp
[0], cp
[1], cp
[2], cp
[3]);
1680 kprintf("%d\n",*p
++);
1682 if (p
>= (int *)&cbqtrace_buffer
[NCBQTRACE
])
1683 p
= (int *)cbqtrace_buffer
;
1686 #endif /* CBQ_TRACE */
1687 #endif /* ALTQ_CBQ */