if: Multiple TX queue support step 3 of 3; map CPUID to subqueue
[dragonfly.git] / sys / net / altq / altq_hfsc.c
CommitLineData
4d723e5a 1/* $KAME: altq_hfsc.c,v 1.25 2004/04/17 10:54:48 kjc Exp $ */
4d723e5a
JS
2
3/*
4 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
5 *
6 * Permission to use, copy, modify, and distribute this software and
7 * its documentation is hereby granted (including for commercial or
8 * for-profit use), provided that both the copyright notice and this
9 * permission notice appear in all copies of the software, derivative
10 * works, or modified versions, and any portions thereof.
11 *
12 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
13 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
14 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
15 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
16 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
17 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
19 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
20 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
21 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
22 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
24 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
25 * DAMAGE.
26 *
27 * Carnegie Mellon encourages (but does not require) users of this
28 * software to return any improvements or extensions that they make,
29 * and to grant Carnegie Mellon the rights to redistribute these
30 * changes without encumbrance.
31 */
32/*
33 * H-FSC is described in Proceedings of SIGCOMM'97,
34 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
35 * Real-Time and Priority Service"
36 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
37 *
38 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
39 * when a class has an upperlimit, the fit-time is computed from the
40 * upperlimit service curve. the link-sharing scheduler does not schedule
41 * a class whose fit-time exceeds the current time.
42 */
43
44#include "opt_altq.h"
45#include "opt_inet.h"
46#include "opt_inet6.h"
47
48#ifdef ALTQ_HFSC /* hfsc is enabled by ALTQ_HFSC option in opt_altq.h */
49
50#include <sys/param.h>
51#include <sys/malloc.h>
52#include <sys/mbuf.h>
53#include <sys/socket.h>
54#include <sys/systm.h>
55#include <sys/errno.h>
56#include <sys/queue.h>
e9cb6d99 57#include <sys/thread.h>
4d723e5a
JS
58
59#include <net/if.h>
60#include <net/ifq_var.h>
61#include <netinet/in.h>
62
63#include <net/pf/pfvar.h>
64#include <net/altq/altq.h>
65#include <net/altq/altq_hfsc.h>
66
e9cb6d99
MD
67#include <sys/thread2.h>
68
f0a26983
SZ
69#define HFSC_SUBQ_INDEX ALTQ_SUBQ_INDEX_DEFAULT
70#define HFSC_LOCK(ifq) \
71 ALTQ_SQ_LOCK(&(ifq)->altq_subq[HFSC_SUBQ_INDEX])
72#define HFSC_UNLOCK(ifq) \
73 ALTQ_SQ_UNLOCK(&(ifq)->altq_subq[HFSC_SUBQ_INDEX])
74
4d723e5a
JS
75/*
76 * function prototypes
77 */
78static int hfsc_clear_interface(struct hfsc_if *);
f0a26983 79static int hfsc_request(struct ifaltq_subque *, int, void *);
4d723e5a
JS
80static void hfsc_purge(struct hfsc_if *);
81static struct hfsc_class *hfsc_class_create(struct hfsc_if *,
82 struct service_curve *,
83 struct service_curve *,
84 struct service_curve *,
85 struct hfsc_class *, int, int, int);
86static int hfsc_class_destroy(struct hfsc_class *);
87static struct hfsc_class *hfsc_nextclass(struct hfsc_class *);
f0a26983 88static int hfsc_enqueue(struct ifaltq_subque *, struct mbuf *,
4d723e5a 89 struct altq_pktattr *);
f0a26983 90static struct mbuf *hfsc_dequeue(struct ifaltq_subque *, struct mbuf *, int);
4d723e5a
JS
91
92static int hfsc_addq(struct hfsc_class *, struct mbuf *);
93static struct mbuf *hfsc_getq(struct hfsc_class *);
94static struct mbuf *hfsc_pollq(struct hfsc_class *);
95static void hfsc_purgeq(struct hfsc_class *);
96
97static void update_cfmin(struct hfsc_class *);
98static void set_active(struct hfsc_class *, int);
99static void set_passive(struct hfsc_class *);
100
101static void init_ed(struct hfsc_class *, int);
102static void update_ed(struct hfsc_class *, int);
103static void update_d(struct hfsc_class *, int);
104static void init_vf(struct hfsc_class *, int);
105static void update_vf(struct hfsc_class *, int, uint64_t);
106static ellist_t *ellist_alloc(void);
107static void ellist_destroy(ellist_t *);
108static void ellist_insert(struct hfsc_class *);
109static void ellist_remove(struct hfsc_class *);
110static void ellist_update(struct hfsc_class *);
111struct hfsc_class *ellist_get_mindl(ellist_t *, uint64_t);
112static actlist_t *actlist_alloc(void);
113static void actlist_destroy(actlist_t *);
114static void actlist_insert(struct hfsc_class *);
115static void actlist_remove(struct hfsc_class *);
116static void actlist_update(struct hfsc_class *);
117
118static struct hfsc_class *actlist_firstfit(struct hfsc_class *, uint64_t);
119
120static __inline uint64_t seg_x2y(uint64_t, uint64_t);
121static __inline uint64_t seg_y2x(uint64_t, uint64_t);
122static __inline uint64_t m2sm(u_int);
123static __inline uint64_t m2ism(u_int);
124static __inline uint64_t d2dx(u_int);
125static u_int sm2m(uint64_t);
126static u_int dx2d(uint64_t);
127
128static void sc2isc(struct service_curve *, struct internal_sc *);
129static void rtsc_init(struct runtime_sc *, struct internal_sc *,
130 uint64_t, uint64_t);
131static uint64_t rtsc_y2x(struct runtime_sc *, uint64_t);
132static uint64_t rtsc_x2y(struct runtime_sc *, uint64_t);
133static void rtsc_min(struct runtime_sc *, struct internal_sc *,
134 uint64_t, uint64_t);
135
136static void get_class_stats(struct hfsc_classstats *, struct hfsc_class *);
137static struct hfsc_class *clh_to_clp(struct hfsc_if *, uint32_t);
138
139/*
140 * macros
141 */
142#define is_a_parent_class(cl) ((cl)->cl_children != NULL)
143
144#define HT_INFINITY 0xffffffffffffffffLL /* infinite time value */
145
146int
9db4b353 147hfsc_pfattach(struct pf_altq *a, struct ifaltq *ifq)
4d723e5a 148{
2cc2f639 149 return altq_attach(ifq, ALTQT_HFSC, a->altq_disc, ifq_mapsubq_default,
4d723e5a 150 hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
4d723e5a
JS
151}
152
153int
154hfsc_add_altq(struct pf_altq *a)
155{
156 struct hfsc_if *hif;
157 struct ifnet *ifp;
158
159 if ((ifp = ifunit(a->ifname)) == NULL)
160 return (EINVAL);
161 if (!ifq_is_ready(&ifp->if_snd))
162 return (ENODEV);
163
efda3bd0 164 hif = kmalloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
4d723e5a
JS
165
166 hif->hif_eligible = ellist_alloc();
167 hif->hif_ifq = &ifp->if_snd;
9275f515 168 ifq_purge_all(&ifp->if_snd);
4d723e5a
JS
169
170 /* keep the state in pf_altq */
171 a->altq_disc = hif;
172
173 return (0);
174}
175
176int
177hfsc_remove_altq(struct pf_altq *a)
178{
179 struct hfsc_if *hif;
180
181 if ((hif = a->altq_disc) == NULL)
182 return (EINVAL);
183 a->altq_disc = NULL;
184
185 hfsc_clear_interface(hif);
186 hfsc_class_destroy(hif->hif_rootclass);
187
188 ellist_destroy(hif->hif_eligible);
189
efda3bd0 190 kfree(hif, M_ALTQ);
4d723e5a
JS
191
192 return (0);
193}
194
9db4b353
SZ
195static int
196hfsc_add_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
4d723e5a 197{
4d723e5a
JS
198 struct hfsc_class *cl, *parent;
199 struct hfsc_opts *opts;
200 struct service_curve rtsc, lssc, ulsc;
201
9db4b353 202 KKASSERT(a->qid != 0);
4d723e5a
JS
203
204 opts = &a->pq_u.hfsc_opts;
205
206 if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
207 parent = NULL;
208 else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
209 return (EINVAL);
210
4d723e5a
JS
211 if (clh_to_clp(hif, a->qid) != NULL)
212 return (EBUSY);
213
214 rtsc.m1 = opts->rtsc_m1;
215 rtsc.d = opts->rtsc_d;
216 rtsc.m2 = opts->rtsc_m2;
217 lssc.m1 = opts->lssc_m1;
218 lssc.d = opts->lssc_d;
219 lssc.m2 = opts->lssc_m2;
220 ulsc.m1 = opts->ulsc_m1;
221 ulsc.d = opts->ulsc_d;
222 ulsc.m2 = opts->ulsc_m2;
223
224 cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
225 opts->flags, a->qid);
226 if (cl == NULL)
227 return (ENOMEM);
228
229 return (0);
230}
231
232int
9db4b353 233hfsc_add_queue(struct pf_altq *a)
4d723e5a
JS
234{
235 struct hfsc_if *hif;
9db4b353
SZ
236 struct ifaltq *ifq;
237 int error;
4d723e5a 238
9db4b353
SZ
239 if (a->qid == 0)
240 return (EINVAL);
241
242 /* XXX not MP safe */
4d723e5a
JS
243 if ((hif = a->altq_disc) == NULL)
244 return (EINVAL);
9db4b353
SZ
245 ifq = hif->hif_ifq;
246
f0a26983 247 HFSC_LOCK(ifq);
9db4b353 248 error = hfsc_add_queue_locked(a, hif);
f0a26983 249 HFSC_UNLOCK(ifq);
9db4b353
SZ
250
251 return error;
252}
253
254static int
255hfsc_remove_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
256{
257 struct hfsc_class *cl;
4d723e5a
JS
258
259 if ((cl = clh_to_clp(hif, a->qid)) == NULL)
260 return (EINVAL);
261
262 return (hfsc_class_destroy(cl));
263}
264
265int
9db4b353
SZ
266hfsc_remove_queue(struct pf_altq *a)
267{
268 struct hfsc_if *hif;
269 struct ifaltq *ifq;
270 int error;
271
272 /* XXX not MP safe */
273 if ((hif = a->altq_disc) == NULL)
274 return (EINVAL);
275 ifq = hif->hif_ifq;
276
f0a26983 277 HFSC_LOCK(ifq);
9db4b353 278 error = hfsc_remove_queue_locked(a, hif);
f0a26983 279 HFSC_UNLOCK(ifq);
9db4b353
SZ
280
281 return error;
282}
283
284int
4d723e5a
JS
285hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
286{
287 struct hfsc_if *hif;
288 struct hfsc_class *cl;
289 struct hfsc_classstats stats;
9db4b353 290 struct ifaltq *ifq;
4d723e5a
JS
291 int error = 0;
292
9db4b353
SZ
293 if (*nbytes < sizeof(stats))
294 return (EINVAL);
295
296 /* XXX not MP safe */
4d723e5a
JS
297 if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL)
298 return (EBADF);
9db4b353 299 ifq = hif->hif_ifq;
4d723e5a 300
f0a26983 301 HFSC_LOCK(ifq);
4d723e5a 302
9db4b353 303 if ((cl = clh_to_clp(hif, a->qid)) == NULL) {
f0a26983 304 HFSC_UNLOCK(ifq);
4d723e5a 305 return (EINVAL);
9db4b353 306 }
4d723e5a
JS
307
308 get_class_stats(&stats, cl);
309
f0a26983 310 HFSC_UNLOCK(ifq);
9db4b353 311
4d723e5a
JS
312 if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
313 return (error);
314 *nbytes = sizeof(stats);
315 return (0);
316}
317
318/*
319 * bring the interface back to the initial state by discarding
320 * all the filters and classes except the root class.
321 */
322static int
323hfsc_clear_interface(struct hfsc_if *hif)
324{
325 struct hfsc_class *cl;
326
327 if (hif->hif_rootclass == NULL)
328 return (0);
329
330
331 /* clear out the classes */
332 while ((cl = hif->hif_rootclass->cl_children) != NULL) {
333 /*
334 * remove the first leaf class found in the hierarchy
335 * then start over
336 */
337 for (; cl != NULL; cl = hfsc_nextclass(cl)) {
338 if (!is_a_parent_class(cl)) {
339 hfsc_class_destroy(cl);
340 break;
341 }
342 }
343 }
344
345 return (0);
346}
347
348static int
f0a26983 349hfsc_request(struct ifaltq_subque *ifsq, int req, void *arg)
4d723e5a 350{
f0a26983 351 struct ifaltq *ifq = ifsq->ifsq_altq;
4d723e5a
JS
352 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
353
e9cb6d99 354 crit_enter();
4d723e5a
JS
355 switch (req) {
356 case ALTRQ_PURGE:
f0a26983
SZ
357 if (ifsq_get_index(ifsq) == HFSC_SUBQ_INDEX) {
358 hfsc_purge(hif);
359 } else {
360 /*
361 * Race happened, the unrelated subqueue was
362 * picked during the packet scheduler transition.
363 */
364 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
365 }
4d723e5a
JS
366 break;
367 }
e9cb6d99 368 crit_exit();
4d723e5a
JS
369 return (0);
370}
371
372/* discard all the queued packets on the interface */
373static void
374hfsc_purge(struct hfsc_if *hif)
375{
376 struct hfsc_class *cl;
377
378 for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
379 if (!qempty(cl->cl_q))
380 hfsc_purgeq(cl);
381 }
382 if (ifq_is_enabled(hif->hif_ifq))
f0a26983 383 hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX].ifq_len = 0;
4d723e5a
JS
384}
385
386struct hfsc_class *
387hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
388 struct service_curve *fsc, struct service_curve *usc,
389 struct hfsc_class *parent, int qlimit, int flags, int qid)
390{
391 struct hfsc_class *cl, *p;
0b31d406 392 int i;
4d723e5a
JS
393
394 if (hif->hif_classes >= HFSC_MAX_CLASSES)
395 return (NULL);
396
397#ifndef ALTQ_RED
398 if (flags & HFCF_RED) {
399#ifdef ALTQ_DEBUG
4b1cf444 400 kprintf("hfsc_class_create: RED not configured for HFSC!\n");
4d723e5a
JS
401#endif
402 return (NULL);
403 }
404#endif
405
efda3bd0
MD
406 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
407 cl->cl_q = kmalloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
4d723e5a
JS
408 cl->cl_actc = actlist_alloc();
409
410 if (qlimit == 0)
411 qlimit = 50; /* use default */
412 qlimit(cl->cl_q) = qlimit;
413 qtype(cl->cl_q) = Q_DROPTAIL;
414 qlen(cl->cl_q) = 0;
415 cl->cl_flags = flags;
416#ifdef ALTQ_RED
417 if (flags & (HFCF_RED|HFCF_RIO)) {
418 int red_flags, red_pkttime;
419 u_int m2;
420
421 m2 = 0;
422 if (rsc != NULL && rsc->m2 > m2)
423 m2 = rsc->m2;
424 if (fsc != NULL && fsc->m2 > m2)
425 m2 = fsc->m2;
426 if (usc != NULL && usc->m2 > m2)
427 m2 = usc->m2;
428
429 red_flags = 0;
430 if (flags & HFCF_ECN)
431 red_flags |= REDF_ECN;
432#ifdef ALTQ_RIO
433 if (flags & HFCF_CLEARDSCP)
434 red_flags |= RIOF_CLEARDSCP;
435#endif
436 if (m2 < 8)
437 red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
438 else
439 red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
440 * 1000 * 1000 * 1000 / (m2 / 8);
441 if (flags & HFCF_RED) {
442 cl->cl_red = red_alloc(0, 0,
443 qlimit(cl->cl_q) * 10/100,
444 qlimit(cl->cl_q) * 30/100,
445 red_flags, red_pkttime);
446 if (cl->cl_red != NULL)
447 qtype(cl->cl_q) = Q_RED;
448 }
449#ifdef ALTQ_RIO
450 else {
451 cl->cl_red = (red_t *)rio_alloc(0, NULL,
452 red_flags, red_pkttime);
453 if (cl->cl_red != NULL)
454 qtype(cl->cl_q) = Q_RIO;
455 }
456#endif
457 }
458#endif /* ALTQ_RED */
459
460 if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
efda3bd0 461 cl->cl_rsc = kmalloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
4d723e5a
JS
462 sc2isc(rsc, cl->cl_rsc);
463 rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
464 rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
465 }
466 if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
efda3bd0 467 cl->cl_fsc = kmalloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
4d723e5a
JS
468 sc2isc(fsc, cl->cl_fsc);
469 rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
470 }
471 if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
efda3bd0 472 cl->cl_usc = kmalloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
4d723e5a
JS
473 sc2isc(usc, cl->cl_usc);
474 rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
475 }
476
477 cl->cl_id = hif->hif_classid++;
478 cl->cl_handle = qid;
479 cl->cl_hif = hif;
480 cl->cl_parent = parent;
481
0b31d406 482 crit_enter();
4d723e5a
JS
483 hif->hif_classes++;
484
485 /*
486 * find a free slot in the class table. if the slot matching
487 * the lower bits of qid is free, use this slot. otherwise,
488 * use the first free slot.
489 */
490 i = qid % HFSC_MAX_CLASSES;
491 if (hif->hif_class_tbl[i] == NULL)
492 hif->hif_class_tbl[i] = cl;
493 else {
494 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
495 if (hif->hif_class_tbl[i] == NULL) {
496 hif->hif_class_tbl[i] = cl;
497 break;
498 }
499 }
500 if (i == HFSC_MAX_CLASSES) {
0b31d406 501 crit_exit();
4d723e5a
JS
502 goto err_ret;
503 }
504 }
505
506 if (flags & HFCF_DEFAULTCLASS)
507 hif->hif_defaultclass = cl;
508
509 if (parent == NULL) {
510 /* this is root class */
511 hif->hif_rootclass = cl;
512 } else if (parent->cl_children == NULL) {
513 /* add this class to the children list of the parent */
514 parent->cl_children = cl;
515 } else {
516 p = parent->cl_children;
517 while (p->cl_siblings != NULL)
518 p = p->cl_siblings;
519 p->cl_siblings = cl;
520 }
0b31d406 521 crit_exit();
4d723e5a
JS
522
523 return (cl);
524
525 err_ret:
526 if (cl->cl_actc != NULL)
527 actlist_destroy(cl->cl_actc);
528 if (cl->cl_red != NULL) {
529#ifdef ALTQ_RIO
530 if (q_is_rio(cl->cl_q))
531 rio_destroy((rio_t *)cl->cl_red);
532#endif
533#ifdef ALTQ_RED
534 if (q_is_red(cl->cl_q))
535 red_destroy(cl->cl_red);
536#endif
537 }
538 if (cl->cl_fsc != NULL)
efda3bd0 539 kfree(cl->cl_fsc, M_ALTQ);
4d723e5a 540 if (cl->cl_rsc != NULL)
efda3bd0 541 kfree(cl->cl_rsc, M_ALTQ);
4d723e5a 542 if (cl->cl_usc != NULL)
efda3bd0 543 kfree(cl->cl_usc, M_ALTQ);
4d723e5a 544 if (cl->cl_q != NULL)
efda3bd0
MD
545 kfree(cl->cl_q, M_ALTQ);
546 kfree(cl, M_ALTQ);
4d723e5a
JS
547 return (NULL);
548}
549
550static int
551hfsc_class_destroy(struct hfsc_class *cl)
552{
b621ed1c 553 struct hfsc_if *hif;
0b31d406 554 int i;
4d723e5a
JS
555
556 if (cl == NULL)
557 return (0);
b621ed1c 558 hif = cl->cl_hif;
4d723e5a
JS
559
560 if (is_a_parent_class(cl))
561 return (EBUSY);
562
0b31d406 563 crit_enter();
4d723e5a
JS
564
565 if (!qempty(cl->cl_q))
566 hfsc_purgeq(cl);
567
568 if (cl->cl_parent == NULL) {
569 /* this is root class */
570 } else {
571 struct hfsc_class *p = cl->cl_parent->cl_children;
572
573 if (p == cl) {
574 cl->cl_parent->cl_children = cl->cl_siblings;
575 } else {
576 do {
577 if (p->cl_siblings == cl) {
578 p->cl_siblings = cl->cl_siblings;
579 break;
580 }
581 } while ((p = p->cl_siblings) != NULL);
582 }
583 KKASSERT(p != NULL);
584 }
585
586 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
0f117493
SZ
587 if (hif->hif_class_tbl[i] == cl) {
588 hif->hif_class_tbl[i] = NULL;
4d723e5a
JS
589 break;
590 }
591 }
592
0f117493 593 hif->hif_classes--;
0b31d406 594 crit_exit();
4d723e5a
JS
595
596 actlist_destroy(cl->cl_actc);
597
598 if (cl->cl_red != NULL) {
599#ifdef ALTQ_RIO
600 if (q_is_rio(cl->cl_q))
601 rio_destroy((rio_t *)cl->cl_red);
602#endif
603#ifdef ALTQ_RED
604 if (q_is_red(cl->cl_q))
605 red_destroy(cl->cl_red);
606#endif
607 }
608
0f117493
SZ
609 if (cl == hif->hif_rootclass)
610 hif->hif_rootclass = NULL;
611 if (cl == hif->hif_defaultclass)
612 hif->hif_defaultclass = NULL;
613 if (cl == hif->hif_pollcache)
614 hif->hif_pollcache = NULL;
4d723e5a
JS
615
616 if (cl->cl_usc != NULL)
efda3bd0 617 kfree(cl->cl_usc, M_ALTQ);
4d723e5a 618 if (cl->cl_fsc != NULL)
efda3bd0 619 kfree(cl->cl_fsc, M_ALTQ);
4d723e5a 620 if (cl->cl_rsc != NULL)
efda3bd0
MD
621 kfree(cl->cl_rsc, M_ALTQ);
622 kfree(cl->cl_q, M_ALTQ);
623 kfree(cl, M_ALTQ);
4d723e5a
JS
624
625 return (0);
626}
627
628/*
629 * hfsc_nextclass returns the next class in the tree.
630 * usage:
631 * for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
632 * do_something;
633 */
634static struct hfsc_class *
635hfsc_nextclass(struct hfsc_class *cl)
636{
637 if (cl->cl_children != NULL) {
638 cl = cl->cl_children;
639 } else if (cl->cl_siblings != NULL) {
640 cl = cl->cl_siblings;
641 } else {
642 while ((cl = cl->cl_parent) != NULL) {
643 if (cl->cl_siblings != NULL) {
644 cl = cl->cl_siblings;
645 break;
646 }
647 }
648 }
649
650 return (cl);
651}
652
653/*
654 * hfsc_enqueue is an enqueue function to be registered to
655 * (*altq_enqueue) in struct ifaltq.
656 */
657static int
f0a26983
SZ
658hfsc_enqueue(struct ifaltq_subque *ifsq, struct mbuf *m,
659 struct altq_pktattr *pktattr)
4d723e5a 660{
f0a26983 661 struct ifaltq *ifq = ifsq->ifsq_altq;
4d723e5a
JS
662 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
663 struct hfsc_class *cl;
664 int len;
665
f0a26983
SZ
666 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
667 /*
668 * Race happened, the unrelated subqueue was
669 * picked during the packet scheduler transition.
670 */
671 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
672 m_freem(m);
673 return ENOBUFS;
674 }
675
4d723e5a
JS
676 /* grab class set by classifier */
677 if ((m->m_flags & M_PKTHDR) == 0) {
678 /* should not happen */
679 if_printf(ifq->altq_ifp, "altq: packet does not have pkthdr\n");
680 m_freem(m);
681 return (ENOBUFS);
682 }
e9cb6d99 683 crit_enter();
315a7da3
JL
684 if (m->m_pkthdr.fw_flags & PF_MBUF_STRUCTURE)
685 cl = clh_to_clp(hif, m->m_pkthdr.pf.qid);
4d723e5a
JS
686 else
687 cl = NULL;
688 if (cl == NULL || is_a_parent_class(cl)) {
689 cl = hif->hif_defaultclass;
690 if (cl == NULL) {
691 m_freem(m);
e9cb6d99 692 crit_exit();
4d723e5a
JS
693 return (ENOBUFS);
694 }
695 }
696 cl->cl_pktattr = NULL;
697 len = m_pktlen(m);
698 if (hfsc_addq(cl, m) != 0) {
699 /* drop occurred. mbuf was freed in hfsc_addq. */
700 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
e9cb6d99 701 crit_exit();
4d723e5a
JS
702 return (ENOBUFS);
703 }
f0a26983 704 ifsq->ifq_len++;
4d723e5a
JS
705 cl->cl_hif->hif_packets++;
706
707 /* successfully queued. */
708 if (qlen(cl->cl_q) == 1)
709 set_active(cl, m_pktlen(m));
e9cb6d99 710 crit_exit();
4d723e5a
JS
711 return (0);
712}
713
714/*
715 * hfsc_dequeue is a dequeue function to be registered to
716 * (*altq_dequeue) in struct ifaltq.
717 *
718 * note: ALTDQ_POLL returns the next packet without removing the packet
719 * from the queue. ALTDQ_REMOVE is a normal dequeue operation.
720 * ALTDQ_REMOVE must return the same packet if called immediately
721 * after ALTDQ_POLL.
722 */
723static struct mbuf *
f0a26983 724hfsc_dequeue(struct ifaltq_subque *ifsq, struct mbuf *mpolled, int op)
4d723e5a 725{
f0a26983 726 struct ifaltq *ifq = ifsq->ifsq_altq;
4d723e5a
JS
727 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
728 struct hfsc_class *cl;
729 struct mbuf *m;
730 int len, next_len;
731 int realtime = 0;
732 uint64_t cur_time;
733
f0a26983
SZ
734 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
735 /*
736 * Race happened, the unrelated subqueue was
737 * picked during the packet scheduler transition.
738 */
739 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
740 return NULL;
741 }
742
4d723e5a
JS
743 if (hif->hif_packets == 0) {
744 /* no packet in the tree */
745 return (NULL);
746 }
747
e9cb6d99 748 crit_enter();
4d723e5a
JS
749 cur_time = read_machclk();
750
751 if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
752 cl = hif->hif_pollcache;
753 hif->hif_pollcache = NULL;
754 /* check if the class was scheduled by real-time criteria */
755 if (cl->cl_rsc != NULL)
756 realtime = (cl->cl_e <= cur_time);
757 } else {
758 /*
759 * if there are eligible classes, use real-time criteria.
760 * find the class with the minimum deadline among
761 * the eligible classes.
762 */
763 if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
764 realtime = 1;
765 } else {
766#ifdef ALTQ_DEBUG
767 int fits = 0;
768#endif
769 /*
770 * use link-sharing criteria
771 * get the class with the minimum vt in the hierarchy
772 */
773 cl = hif->hif_rootclass;
774 while (is_a_parent_class(cl)) {
775
776 cl = actlist_firstfit(cl, cur_time);
777 if (cl == NULL) {
778#ifdef ALTQ_DEBUG
779 if (fits > 0)
4b1cf444 780 kprintf("%d fit but none found\n",fits);
4d723e5a 781#endif
e9cb6d99
MD
782 m = NULL;
783 goto done;
4d723e5a
JS
784 }
785 /*
786 * update parent's cl_cvtmin.
787 * don't update if the new vt is smaller.
788 */
789 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
790 cl->cl_parent->cl_cvtmin = cl->cl_vt;
791#ifdef ALTQ_DEBUG
792 fits++;
793#endif
794 }
795 }
796
797 if (op == ALTDQ_POLL) {
798 hif->hif_pollcache = cl;
799 m = hfsc_pollq(cl);
e9cb6d99 800 goto done;
4d723e5a
JS
801 }
802 }
803
804 m = hfsc_getq(cl);
805 if (m == NULL)
806 panic("hfsc_dequeue:");
807 len = m_pktlen(m);
808 cl->cl_hif->hif_packets--;
f0a26983 809 ifsq->ifq_len--;
4d723e5a
JS
810 PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
811
812 update_vf(cl, len, cur_time);
813 if (realtime)
814 cl->cl_cumul += len;
815
816 if (!qempty(cl->cl_q)) {
817 if (cl->cl_rsc != NULL) {
818 /* update ed */
819 next_len = m_pktlen(qhead(cl->cl_q));
820
821 if (realtime)
822 update_ed(cl, next_len);
823 else
824 update_d(cl, next_len);
825 }
826 } else {
827 /* the class becomes passive */
828 set_passive(cl);
829 }
e9cb6d99
MD
830done:
831 crit_exit();
d2c71fa0 832 KKASSERT(mpolled == NULL || m == mpolled);
4d723e5a
JS
833 return (m);
834}
835
836static int
837hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
838{
839
840#ifdef ALTQ_RIO
841 if (q_is_rio(cl->cl_q))
842 return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
843 m, cl->cl_pktattr);
844#endif
845#ifdef ALTQ_RED
846 if (q_is_red(cl->cl_q))
847 return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
848#endif
849 if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
850 m_freem(m);
851 return (-1);
852 }
853
854 if (cl->cl_flags & HFCF_CLEARDSCP)
855 write_dsfield(m, cl->cl_pktattr, 0);
856
857 _addq(cl->cl_q, m);
858
859 return (0);
860}
861
862static struct mbuf *
863hfsc_getq(struct hfsc_class *cl)
864{
865#ifdef ALTQ_RIO
866 if (q_is_rio(cl->cl_q))
867 return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
868#endif
869#ifdef ALTQ_RED
870 if (q_is_red(cl->cl_q))
871 return red_getq(cl->cl_red, cl->cl_q);
872#endif
873 return _getq(cl->cl_q);
874}
875
876static struct mbuf *
877hfsc_pollq(struct hfsc_class *cl)
878{
879 return qhead(cl->cl_q);
880}
881
882static void
883hfsc_purgeq(struct hfsc_class *cl)
884{
885 struct mbuf *m;
886
887 if (qempty(cl->cl_q))
888 return;
889
890 while ((m = _getq(cl->cl_q)) != NULL) {
891 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
892 m_freem(m);
893 cl->cl_hif->hif_packets--;
f0a26983 894 cl->cl_hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX].ifq_len--;
4d723e5a
JS
895 }
896 KKASSERT(qlen(cl->cl_q) == 0);
897
898 update_vf(cl, 0, 0); /* remove cl from the actlist */
899 set_passive(cl);
900}
901
902static void
903set_active(struct hfsc_class *cl, int len)
904{
905 if (cl->cl_rsc != NULL)
906 init_ed(cl, len);
907 if (cl->cl_fsc != NULL)
908 init_vf(cl, len);
909
910 cl->cl_stats.period++;
911}
912
913static void
914set_passive(struct hfsc_class *cl)
915{
916 if (cl->cl_rsc != NULL)
917 ellist_remove(cl);
918
919 /*
920 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
921 * needs to be called explicitly to remove a class from actlist
922 */
923}
924
925static void
926init_ed(struct hfsc_class *cl, int next_len)
927{
928 uint64_t cur_time;
929
930 cur_time = read_machclk();
931
932 /* update the deadline curve */
933 rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
934
935 /*
936 * update the eligible curve.
937 * for concave, it is equal to the deadline curve.
938 * for convex, it is a linear curve with slope m2.
939 */
940 cl->cl_eligible = cl->cl_deadline;
941 if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
942 cl->cl_eligible.dx = 0;
943 cl->cl_eligible.dy = 0;
944 }
945
946 /* compute e and d */
947 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
948 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
949
950 ellist_insert(cl);
951}
952
953static void
954update_ed(struct hfsc_class *cl, int next_len)
955{
956 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
957 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
958
959 ellist_update(cl);
960}
961
962static void
963update_d(struct hfsc_class *cl, int next_len)
964{
965 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
966}
967
968static void
969init_vf(struct hfsc_class *cl, int len)
970{
971 struct hfsc_class *max_cl, *p;
972 uint64_t vt, f, cur_time;
973 int go_active;
974
975 cur_time = 0;
976 go_active = 1;
977 for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
978 if (go_active && cl->cl_nactive++ == 0)
979 go_active = 1;
980 else
981 go_active = 0;
982
983 if (go_active) {
984 max_cl = actlist_last(cl->cl_parent->cl_actc);
985 if (max_cl != NULL) {
986 /*
987 * set vt to the average of the min and max
988 * classes. if the parent's period didn't
989 * change, don't decrease vt of the class.
990 */
991 vt = max_cl->cl_vt;
992 if (cl->cl_parent->cl_cvtmin != 0)
993 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
994
995 if (cl->cl_parent->cl_vtperiod !=
996 cl->cl_parentperiod || vt > cl->cl_vt)
997 cl->cl_vt = vt;
998 } else {
999 /*
1000 * first child for a new parent backlog period.
1001 * add parent's cvtmax to vtoff of children
1002 * to make a new vt (vtoff + vt) larger than
1003 * the vt in the last period for all children.
1004 */
1005 vt = cl->cl_parent->cl_cvtmax;
1006 for (p = cl->cl_parent->cl_children; p != NULL;
1007 p = p->cl_siblings)
1008 p->cl_vtoff += vt;
1009 cl->cl_vt = 0;
1010 cl->cl_parent->cl_cvtmax = 0;
1011 cl->cl_parent->cl_cvtmin = 0;
1012 }
1013 cl->cl_initvt = cl->cl_vt;
1014
1015 /* update the virtual curve */
1016 vt = cl->cl_vt + cl->cl_vtoff;
1017 rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
1018 if (cl->cl_virtual.x == vt) {
1019 cl->cl_virtual.x -= cl->cl_vtoff;
1020 cl->cl_vtoff = 0;
1021 }
1022 cl->cl_vtadj = 0;
1023
1024 cl->cl_vtperiod++; /* increment vt period */
1025 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
1026 if (cl->cl_parent->cl_nactive == 0)
1027 cl->cl_parentperiod++;
1028 cl->cl_f = 0;
1029
1030 actlist_insert(cl);
1031
1032 if (cl->cl_usc != NULL) {
1033 /* class has upper limit curve */
1034 if (cur_time == 0)
1035 cur_time = read_machclk();
1036
1037 /* update the ulimit curve */
1038 rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
1039 cl->cl_total);
1040 /* compute myf */
1041 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
1042 cl->cl_total);
1043 cl->cl_myfadj = 0;
1044 }
1045 }
1046
1047 if (cl->cl_myf > cl->cl_cfmin)
1048 f = cl->cl_myf;
1049 else
1050 f = cl->cl_cfmin;
1051 if (f != cl->cl_f) {
1052 cl->cl_f = f;
1053 update_cfmin(cl->cl_parent);
1054 }
1055 }
1056}
1057
1058static void
1059update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
1060{
1061 uint64_t f, myf_bound, delta;
1062 int go_passive;
1063
1064 go_passive = qempty(cl->cl_q);
1065
1066 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
1067 cl->cl_total += len;
1068
1069 if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1070 continue;
1071
1072 if (go_passive && --cl->cl_nactive == 0)
1073 go_passive = 1;
1074 else
1075 go_passive = 0;
1076
1077 if (go_passive) {
1078 /* no more active child, going passive */
1079
1080 /* update cvtmax of the parent class */
1081 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1082 cl->cl_parent->cl_cvtmax = cl->cl_vt;
1083
1084 /* remove this class from the vt list */
1085 actlist_remove(cl);
1086
1087 update_cfmin(cl->cl_parent);
1088
1089 continue;
1090 }
1091
1092 /*
1093 * update vt and f
1094 */
1095 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1096 - cl->cl_vtoff + cl->cl_vtadj;
1097
1098 /*
1099 * if vt of the class is smaller than cvtmin,
1100 * the class was skipped in the past due to non-fit.
1101 * if so, we need to adjust vtadj.
1102 */
1103 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1104 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1105 cl->cl_vt = cl->cl_parent->cl_cvtmin;
1106 }
1107
1108 /* update the vt list */
1109 actlist_update(cl);
1110
1111 if (cl->cl_usc != NULL) {
1112 cl->cl_myf = cl->cl_myfadj
1113 + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1114
1115 /*
1116 * if myf lags behind by more than one clock tick
1117 * from the current time, adjust myfadj to prevent
1118 * a rate-limited class from going greedy.
1119 * in a steady state under rate-limiting, myf
1120 * fluctuates within one clock tick.
1121 */
1122 myf_bound = cur_time - machclk_per_tick;
1123 if (cl->cl_myf < myf_bound) {
1124 delta = cur_time - cl->cl_myf;
1125 cl->cl_myfadj += delta;
1126 cl->cl_myf += delta;
1127 }
1128 }
1129
1130 /* cl_f is max(cl_myf, cl_cfmin) */
1131 if (cl->cl_myf > cl->cl_cfmin)
1132 f = cl->cl_myf;
1133 else
1134 f = cl->cl_cfmin;
1135 if (f != cl->cl_f) {
1136 cl->cl_f = f;
1137 update_cfmin(cl->cl_parent);
1138 }
1139 }
1140}
1141
1142static void
1143update_cfmin(struct hfsc_class *cl)
1144{
1145 struct hfsc_class *p;
1146 uint64_t cfmin;
1147
1148 if (TAILQ_EMPTY(cl->cl_actc)) {
1149 cl->cl_cfmin = 0;
1150 return;
1151 }
1152 cfmin = HT_INFINITY;
1153 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1154 if (p->cl_f == 0) {
1155 cl->cl_cfmin = 0;
1156 return;
1157 }
1158 if (p->cl_f < cfmin)
1159 cfmin = p->cl_f;
1160 }
1161 cl->cl_cfmin = cfmin;
1162}
1163
1164/*
1165 * TAILQ based ellist and actlist implementation
1166 * (ion wanted to make a calendar queue based implementation)
1167 */
1168/*
1169 * eligible list holds backlogged classes being sorted by their eligible times.
1170 * there is one eligible list per interface.
1171 */
1172
1173static ellist_t *
1174ellist_alloc(void)
1175{
1176 ellist_t *head;
1177
efda3bd0 1178 head = kmalloc(sizeof(ellist_t *), M_ALTQ, M_WAITOK);
4d723e5a
JS
1179 TAILQ_INIT(head);
1180 return (head);
1181}
1182
1183static void
1184ellist_destroy(ellist_t *head)
1185{
efda3bd0 1186 kfree(head, M_ALTQ);
4d723e5a
JS
1187}
1188
1189static void
1190ellist_insert(struct hfsc_class *cl)
1191{
1192 struct hfsc_if *hif = cl->cl_hif;
1193 struct hfsc_class *p;
1194
1195 /* check the last entry first */
1196 if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1197 p->cl_e <= cl->cl_e) {
1198 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1199 return;
1200 }
1201
1202 TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1203 if (cl->cl_e < p->cl_e) {
1204 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1205 return;
1206 }
1207 }
1208 KKASSERT(0); /* should not reach here */
1209}
1210
1211static void
1212ellist_remove(struct hfsc_class *cl)
1213{
1214 struct hfsc_if *hif = cl->cl_hif;
1215
1216 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1217}
1218
1219static void
1220ellist_update(struct hfsc_class *cl)
1221{
1222 struct hfsc_if *hif = cl->cl_hif;
1223 struct hfsc_class *p, *last;
1224
1225 /*
1226 * the eligible time of a class increases monotonically.
1227 * if the next entry has a larger eligible time, nothing to do.
1228 */
1229 p = TAILQ_NEXT(cl, cl_ellist);
1230 if (p == NULL || cl->cl_e <= p->cl_e)
1231 return;
1232
1233 /* check the last entry */
1234 last = TAILQ_LAST(hif->hif_eligible, _eligible);
1235 KKASSERT(last != NULL);
1236 if (last->cl_e <= cl->cl_e) {
1237 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1238 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1239 return;
1240 }
1241
1242 /*
1243 * the new position must be between the next entry
1244 * and the last entry
1245 */
1246 while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1247 if (cl->cl_e < p->cl_e) {
1248 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1249 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1250 return;
1251 }
1252 }
1253 KKASSERT(0); /* should not reach here */
1254}
1255
1256/* find the class with the minimum deadline among the eligible classes */
1257struct hfsc_class *
1258ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1259{
1260 struct hfsc_class *p, *cl = NULL;
1261
1262 TAILQ_FOREACH(p, head, cl_ellist) {
1263 if (p->cl_e > cur_time)
1264 break;
1265 if (cl == NULL || p->cl_d < cl->cl_d)
1266 cl = p;
1267 }
1268 return (cl);
1269}
1270
1271/*
1272 * active children list holds backlogged child classes being sorted
1273 * by their virtual time.
1274 * each intermediate class has one active children list.
1275 */
1276static actlist_t *
1277actlist_alloc(void)
1278{
1279 actlist_t *head;
1280
efda3bd0 1281 head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
4d723e5a
JS
1282 TAILQ_INIT(head);
1283 return (head);
1284}
1285
1286static void
1287actlist_destroy(actlist_t *head)
1288{
efda3bd0 1289 kfree(head, M_ALTQ);
4d723e5a
JS
1290}
1291static void
1292actlist_insert(struct hfsc_class *cl)
1293{
1294 struct hfsc_class *p;
1295
1296 /* check the last entry first */
1297 if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1298 || p->cl_vt <= cl->cl_vt) {
1299 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1300 return;
1301 }
1302
1303 TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1304 if (cl->cl_vt < p->cl_vt) {
1305 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1306 return;
1307 }
1308 }
1309 KKASSERT(0); /* should not reach here */
1310}
1311
1312static void
1313actlist_remove(struct hfsc_class *cl)
1314{
1315 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1316}
1317
1318static void
1319actlist_update(struct hfsc_class *cl)
1320{
1321 struct hfsc_class *p, *last;
1322
1323 /*
1324 * the virtual time of a class increases monotonically during its
1325 * backlogged period.
1326 * if the next entry has a larger virtual time, nothing to do.
1327 */
1328 p = TAILQ_NEXT(cl, cl_actlist);
1329 if (p == NULL || cl->cl_vt < p->cl_vt)
1330 return;
1331
1332 /* check the last entry */
1333 last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1334 KKASSERT(last != NULL);
1335 if (last->cl_vt <= cl->cl_vt) {
1336 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1337 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1338 return;
1339 }
1340
1341 /*
1342 * the new position must be between the next entry
1343 * and the last entry
1344 */
1345 while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1346 if (cl->cl_vt < p->cl_vt) {
1347 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1348 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1349 return;
1350 }
1351 }
1352 KKASSERT(0); /* should not reach here */
1353}
1354
1355static struct hfsc_class *
1356actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1357{
1358 struct hfsc_class *p;
1359
1360 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1361 if (p->cl_f <= cur_time)
1362 return (p);
1363 }
1364 return (NULL);
1365}
1366
1367/*
1368 * service curve support functions
1369 *
1370 * external service curve parameters
1371 * m: bits/sec
1372 * d: msec
1373 * internal service curve parameters
1374 * sm: (bytes/tsc_interval) << SM_SHIFT
1375 * ism: (tsc_count/byte) << ISM_SHIFT
1376 * dx: tsc_count
1377 *
1378 * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1379 * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1380 * speed. SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1381 * digits in decimal using the following table.
1382 *
1383 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
1384 * ----------+-------------------------------------------------------
1385 * bytes/nsec 12.5e-6 125e-6 1250e-6 12500e-6 125000e-6
1386 * sm(500MHz) 25.0e-6 250e-6 2500e-6 25000e-6 250000e-6
1387 * sm(200MHz) 62.5e-6 625e-6 6250e-6 62500e-6 625000e-6
1388 *
1389 * nsec/byte 80000 8000 800 80 8
1390 * ism(500MHz) 40000 4000 400 40 4
1391 * ism(200MHz) 16000 1600 160 16 1.6
1392 */
1393#define SM_SHIFT 24
1394#define ISM_SHIFT 10
1395
1396#define SM_MASK ((1LL << SM_SHIFT) - 1)
1397#define ISM_MASK ((1LL << ISM_SHIFT) - 1)
1398
1399static __inline uint64_t
1400seg_x2y(uint64_t x, uint64_t sm)
1401{
1402 uint64_t y;
1403
1404 /*
1405 * compute
1406 * y = x * sm >> SM_SHIFT
1407 * but divide it for the upper and lower bits to avoid overflow
1408 */
1409 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1410 return (y);
1411}
1412
1413static __inline uint64_t
1414seg_y2x(uint64_t y, uint64_t ism)
1415{
1416 uint64_t x;
1417
1418 if (y == 0)
1419 x = 0;
1420 else if (ism == HT_INFINITY)
1421 x = HT_INFINITY;
1422 else
1423 x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1424
1425 return (x);
1426}
1427
1428static __inline uint64_t
1429m2sm(u_int m)
1430{
1431 uint64_t sm;
1432
1433 sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1434 return (sm);
1435}
1436
1437static __inline uint64_t
1438m2ism(u_int m)
1439{
1440 uint64_t ism;
1441
1442 if (m == 0)
1443 ism = HT_INFINITY;
1444 else
1445 ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1446 return (ism);
1447}
1448
1449static __inline uint64_t
1450d2dx(u_int d)
1451{
1452 uint64_t dx;
1453
1454 dx = ((uint64_t)d * machclk_freq) / 1000;
1455 return (dx);
1456}
1457
1458static u_int
1459sm2m(uint64_t sm)
1460{
1461 uint64_t m;
1462
1463 m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1464 return ((u_int)m);
1465}
1466
1467static u_int
1468dx2d(uint64_t dx)
1469{
1470 uint64_t d;
1471
1472 d = dx * 1000 / machclk_freq;
1473 return ((u_int)d);
1474}
1475
1476static void
1477sc2isc(struct service_curve *sc, struct internal_sc *isc)
1478{
1479 isc->sm1 = m2sm(sc->m1);
1480 isc->ism1 = m2ism(sc->m1);
1481 isc->dx = d2dx(sc->d);
1482 isc->dy = seg_x2y(isc->dx, isc->sm1);
1483 isc->sm2 = m2sm(sc->m2);
1484 isc->ism2 = m2ism(sc->m2);
1485}
1486
1487/*
1488 * initialize the runtime service curve with the given internal
1489 * service curve starting at (x, y).
1490 */
1491static void
1492rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1493{
1494 rtsc->x = x;
1495 rtsc->y = y;
1496 rtsc->sm1 = isc->sm1;
1497 rtsc->ism1 = isc->ism1;
1498 rtsc->dx = isc->dx;
1499 rtsc->dy = isc->dy;
1500 rtsc->sm2 = isc->sm2;
1501 rtsc->ism2 = isc->ism2;
1502}
1503
1504/*
1505 * calculate the y-projection of the runtime service curve by the
1506 * given x-projection value
1507 */
1508static uint64_t
1509rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1510{
1511 uint64_t x;
1512
1513 if (y < rtsc->y) {
1514 x = rtsc->x;
1515 } else if (y <= rtsc->y + rtsc->dy) {
1516 /* x belongs to the 1st segment */
1517 if (rtsc->dy == 0)
1518 x = rtsc->x + rtsc->dx;
1519 else
1520 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1521 } else {
1522 /* x belongs to the 2nd segment */
1523 x = rtsc->x + rtsc->dx
1524 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1525 }
1526 return (x);
1527}
1528
1529static uint64_t
1530rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1531{
1532 uint64_t y;
1533
1534 if (x <= rtsc->x) {
1535 y = rtsc->y;
1536 } else if (x <= rtsc->x + rtsc->dx) {
1537 /* y belongs to the 1st segment */
1538 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1539 } else
1540 /* y belongs to the 2nd segment */
1541 y = rtsc->y + rtsc->dy
1542 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1543 return (y);
1544}
1545
1546/*
1547 * update the runtime service curve by taking the minimum of the current
1548 * runtime service curve and the service curve starting at (x, y).
1549 */
1550static void
1551rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1552{
1553 uint64_t y1, y2, dx, dy;
1554
1555 if (isc->sm1 <= isc->sm2) {
1556 /* service curve is convex */
1557 y1 = rtsc_x2y(rtsc, x);
1558 if (y1 < y)
1559 /* the current rtsc is smaller */
1560 return;
1561 rtsc->x = x;
1562 rtsc->y = y;
1563 return;
1564 }
1565
1566 /*
1567 * service curve is concave
1568 * compute the two y values of the current rtsc
1569 * y1: at x
1570 * y2: at (x + dx)
1571 */
1572 y1 = rtsc_x2y(rtsc, x);
1573 if (y1 <= y) {
1574 /* rtsc is below isc, no change to rtsc */
1575 return;
1576 }
1577
1578 y2 = rtsc_x2y(rtsc, x + isc->dx);
1579 if (y2 >= y + isc->dy) {
1580 /* rtsc is above isc, replace rtsc by isc */
1581 rtsc->x = x;
1582 rtsc->y = y;
1583 rtsc->dx = isc->dx;
1584 rtsc->dy = isc->dy;
1585 return;
1586 }
1587
1588 /*
1589 * the two curves intersect
1590 * compute the offsets (dx, dy) using the reverse
1591 * function of seg_x2y()
1592 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1593 */
1594 dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1595 /*
1596 * check if (x, y1) belongs to the 1st segment of rtsc.
1597 * if so, add the offset.
1598 */
1599 if (rtsc->x + rtsc->dx > x)
1600 dx += rtsc->x + rtsc->dx - x;
1601 dy = seg_x2y(dx, isc->sm1);
1602
1603 rtsc->x = x;
1604 rtsc->y = y;
1605 rtsc->dx = dx;
1606 rtsc->dy = dy;
1607}
1608
1609static void
1610get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1611{
1612 sp->class_id = cl->cl_id;
1613 sp->class_handle = cl->cl_handle;
1614
1615 if (cl->cl_rsc != NULL) {
1616 sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1617 sp->rsc.d = dx2d(cl->cl_rsc->dx);
1618 sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1619 } else {
1620 sp->rsc.m1 = 0;
1621 sp->rsc.d = 0;
1622 sp->rsc.m2 = 0;
1623 }
1624 if (cl->cl_fsc != NULL) {
1625 sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1626 sp->fsc.d = dx2d(cl->cl_fsc->dx);
1627 sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1628 } else {
1629 sp->fsc.m1 = 0;
1630 sp->fsc.d = 0;
1631 sp->fsc.m2 = 0;
1632 }
1633 if (cl->cl_usc != NULL) {
1634 sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1635 sp->usc.d = dx2d(cl->cl_usc->dx);
1636 sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1637 } else {
1638 sp->usc.m1 = 0;
1639 sp->usc.d = 0;
1640 sp->usc.m2 = 0;
1641 }
1642
1643 sp->total = cl->cl_total;
1644 sp->cumul = cl->cl_cumul;
1645
1646 sp->d = cl->cl_d;
1647 sp->e = cl->cl_e;
1648 sp->vt = cl->cl_vt;
1649 sp->f = cl->cl_f;
1650
1651 sp->initvt = cl->cl_initvt;
1652 sp->vtperiod = cl->cl_vtperiod;
1653 sp->parentperiod = cl->cl_parentperiod;
1654 sp->nactive = cl->cl_nactive;
1655 sp->vtoff = cl->cl_vtoff;
1656 sp->cvtmax = cl->cl_cvtmax;
1657 sp->myf = cl->cl_myf;
1658 sp->cfmin = cl->cl_cfmin;
1659 sp->cvtmin = cl->cl_cvtmin;
1660 sp->myfadj = cl->cl_myfadj;
1661 sp->vtadj = cl->cl_vtadj;
1662
1663 sp->cur_time = read_machclk();
1664 sp->machclk_freq = machclk_freq;
1665
1666 sp->qlength = qlen(cl->cl_q);
1667 sp->qlimit = qlimit(cl->cl_q);
1668 sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1669 sp->drop_cnt = cl->cl_stats.drop_cnt;
1670 sp->period = cl->cl_stats.period;
1671
1672 sp->qtype = qtype(cl->cl_q);
1673#ifdef ALTQ_RED
1674 if (q_is_red(cl->cl_q))
1675 red_getstats(cl->cl_red, &sp->red[0]);
1676#endif
1677#ifdef ALTQ_RIO
1678 if (q_is_rio(cl->cl_q))
1679 rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1680#endif
1681}
1682
1683/* convert a class handle to the corresponding class pointer */
1684static struct hfsc_class *
1685clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1686{
1687 int i;
1688 struct hfsc_class *cl;
1689
1690 if (chandle == 0)
1691 return (NULL);
1692 /*
1693 * first, try optimistically the slot matching the lower bits of
1694 * the handle. if it fails, do the linear table search.
1695 */
1696 i = chandle % HFSC_MAX_CLASSES;
1697 if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1698 return (cl);
1699 for (i = 0; i < HFSC_MAX_CLASSES; i++)
1700 if ((cl = hif->hif_class_tbl[i]) != NULL &&
1701 cl->cl_handle == chandle)
1702 return (cl);
1703 return (NULL);
1704}
1705
1706#endif /* ALTQ_HFSC */