1 /* $KAME: altq_hfsc.c,v 1.25 2004/04/17 10:54:48 kjc Exp $ */
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>
57 #include <sys/thread.h>
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
67 #include <sys/thread2.h>
68
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
75 /*
76 * function prototypes
77 */
78 static int hfsc_clear_interface(struct hfsc_if *);
79 static int hfsc_request(struct ifaltq_subque *, int, void *);
80 static void hfsc_purge(struct hfsc_if *);
81 static 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);
86 static int hfsc_class_destroy(struct hfsc_class *);
87 static struct hfsc_class *hfsc_nextclass(struct hfsc_class *);
88 static int hfsc_enqueue(struct ifaltq_subque *, struct mbuf *,
89 struct altq_pktattr *);
90 static struct mbuf *hfsc_dequeue(struct ifaltq_subque *, int);
91
92 static int hfsc_addq(struct hfsc_class *, struct mbuf *);
93 static struct mbuf *hfsc_getq(struct hfsc_class *);
94 static struct mbuf *hfsc_pollq(struct hfsc_class *);
95 static void hfsc_purgeq(struct hfsc_class *);
96
97 static void update_cfmin(struct hfsc_class *);
98 static void set_active(struct hfsc_class *, int);
99 static void set_passive(struct hfsc_class *);
100
101 static void init_ed(struct hfsc_class *, int);
102 static void update_ed(struct hfsc_class *, int);
103 static void update_d(struct hfsc_class *, int);
104 static void init_vf(struct hfsc_class *, int);
105 static void update_vf(struct hfsc_class *, int, uint64_t);
106 static ellist_t *ellist_alloc(void);
107 static void ellist_destroy(ellist_t *);
108 static void ellist_insert(struct hfsc_class *);
109 static void ellist_remove(struct hfsc_class *);
110 static void ellist_update(struct hfsc_class *);
111 struct hfsc_class *ellist_get_mindl(ellist_t *, uint64_t);
112 static actlist_t *actlist_alloc(void);
113 static void actlist_destroy(actlist_t *);
114 static void actlist_insert(struct hfsc_class *);
115 static void actlist_remove(struct hfsc_class *);
116 static void actlist_update(struct hfsc_class *);
117
118 static struct hfsc_class *actlist_firstfit(struct hfsc_class *, uint64_t);
119
120 static __inline uint64_t seg_x2y(uint64_t, uint64_t);
121 static __inline uint64_t seg_y2x(uint64_t, uint64_t);
122 static __inline uint64_t m2sm(u_int);
123 static __inline uint64_t m2ism(u_int);
124 static __inline uint64_t d2dx(u_int);
125 static u_int sm2m(uint64_t);
126 static u_int dx2d(uint64_t);
127
128 static void sc2isc(struct service_curve *, struct internal_sc *);
129 static void rtsc_init(struct runtime_sc *, struct internal_sc *,
130 uint64_t, uint64_t);
131 static uint64_t rtsc_y2x(struct runtime_sc *, uint64_t);
132 static uint64_t rtsc_x2y(struct runtime_sc *, uint64_t);
133 static void rtsc_min(struct runtime_sc *, struct internal_sc *,
134 uint64_t, uint64_t);
135
136 static void get_class_stats(struct hfsc_classstats *, struct hfsc_class *);
137 static 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
146 int
hfsc_pfattach(struct pf_altq * a,struct ifaltq * ifq)147 hfsc_pfattach(struct pf_altq *a, struct ifaltq *ifq)
148 {
149 return altq_attach(ifq, ALTQT_HFSC, a->altq_disc, ifq_mapsubq_default,
150 hfsc_enqueue, hfsc_dequeue, hfsc_request, NULL, NULL);
151 }
152
153 int
hfsc_add_altq(struct pf_altq * a)154 hfsc_add_altq(struct pf_altq *a)
155 {
156 struct hfsc_if *hif;
157 struct ifnet *ifp;
158
159 ifnet_lock();
160
161 if ((ifp = ifunit(a->ifname)) == NULL) {
162 ifnet_unlock();
163 return (EINVAL);
164 }
165 if (!ifq_is_ready(&ifp->if_snd)) {
166 ifnet_unlock();
167 return (ENODEV);
168 }
169
170 hif = kmalloc(sizeof(struct hfsc_if), M_ALTQ, M_WAITOK | M_ZERO);
171
172 hif->hif_eligible = ellist_alloc();
173 hif->hif_ifq = &ifp->if_snd;
174 ifq_purge_all(&ifp->if_snd);
175
176 ifnet_unlock();
177
178 /* keep the state in pf_altq */
179 a->altq_disc = hif;
180
181 return (0);
182 }
183
184 int
hfsc_remove_altq(struct pf_altq * a)185 hfsc_remove_altq(struct pf_altq *a)
186 {
187 struct hfsc_if *hif;
188
189 if ((hif = a->altq_disc) == NULL)
190 return (EINVAL);
191 a->altq_disc = NULL;
192
193 hfsc_clear_interface(hif);
194 hfsc_class_destroy(hif->hif_rootclass);
195
196 ellist_destroy(hif->hif_eligible);
197
198 kfree(hif, M_ALTQ);
199
200 return (0);
201 }
202
203 static int
hfsc_add_queue_locked(struct pf_altq * a,struct hfsc_if * hif)204 hfsc_add_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
205 {
206 struct hfsc_class *cl, *parent;
207 struct hfsc_opts *opts;
208 struct service_curve rtsc, lssc, ulsc;
209
210 KKASSERT(a->qid != 0);
211
212 opts = &a->pq_u.hfsc_opts;
213
214 if (a->parent_qid == HFSC_NULLCLASS_HANDLE && hif->hif_rootclass == NULL)
215 parent = NULL;
216 else if ((parent = clh_to_clp(hif, a->parent_qid)) == NULL)
217 return (EINVAL);
218
219 if (clh_to_clp(hif, a->qid) != NULL)
220 return (EBUSY);
221
222 rtsc.m1 = opts->rtsc_m1;
223 rtsc.d = opts->rtsc_d;
224 rtsc.m2 = opts->rtsc_m2;
225 lssc.m1 = opts->lssc_m1;
226 lssc.d = opts->lssc_d;
227 lssc.m2 = opts->lssc_m2;
228 ulsc.m1 = opts->ulsc_m1;
229 ulsc.d = opts->ulsc_d;
230 ulsc.m2 = opts->ulsc_m2;
231
232 cl = hfsc_class_create(hif, &rtsc, &lssc, &ulsc, parent, a->qlimit,
233 opts->flags, a->qid);
234 if (cl == NULL)
235 return (ENOMEM);
236
237 return (0);
238 }
239
240 int
hfsc_add_queue(struct pf_altq * a)241 hfsc_add_queue(struct pf_altq *a)
242 {
243 struct hfsc_if *hif;
244 struct ifaltq *ifq;
245 int error;
246
247 if (a->qid == 0)
248 return (EINVAL);
249
250 /* XXX not MP safe */
251 if ((hif = a->altq_disc) == NULL)
252 return (EINVAL);
253 ifq = hif->hif_ifq;
254
255 HFSC_LOCK(ifq);
256 error = hfsc_add_queue_locked(a, hif);
257 HFSC_UNLOCK(ifq);
258
259 return error;
260 }
261
262 static int
hfsc_remove_queue_locked(struct pf_altq * a,struct hfsc_if * hif)263 hfsc_remove_queue_locked(struct pf_altq *a, struct hfsc_if *hif)
264 {
265 struct hfsc_class *cl;
266
267 if ((cl = clh_to_clp(hif, a->qid)) == NULL)
268 return (EINVAL);
269
270 return (hfsc_class_destroy(cl));
271 }
272
273 int
hfsc_remove_queue(struct pf_altq * a)274 hfsc_remove_queue(struct pf_altq *a)
275 {
276 struct hfsc_if *hif;
277 struct ifaltq *ifq;
278 int error;
279
280 /* XXX not MP safe */
281 if ((hif = a->altq_disc) == NULL)
282 return (EINVAL);
283 ifq = hif->hif_ifq;
284
285 HFSC_LOCK(ifq);
286 error = hfsc_remove_queue_locked(a, hif);
287 HFSC_UNLOCK(ifq);
288
289 return error;
290 }
291
292 int
hfsc_getqstats(struct pf_altq * a,void * ubuf,int * nbytes)293 hfsc_getqstats(struct pf_altq *a, void *ubuf, int *nbytes)
294 {
295 struct hfsc_if *hif;
296 struct hfsc_class *cl;
297 struct hfsc_classstats stats;
298 struct ifaltq *ifq;
299 int error = 0;
300
301 if (*nbytes < sizeof(stats))
302 return (EINVAL);
303
304 ifnet_lock();
305
306 /* XXX not MP safe */
307 if ((hif = altq_lookup(a->ifname, ALTQT_HFSC)) == NULL) {
308 ifnet_unlock();
309 return (EBADF);
310 }
311 ifq = hif->hif_ifq;
312
313 HFSC_LOCK(ifq);
314
315 if ((cl = clh_to_clp(hif, a->qid)) == NULL) {
316 HFSC_UNLOCK(ifq);
317 ifnet_unlock();
318 return (EINVAL);
319 }
320
321 get_class_stats(&stats, cl);
322
323 HFSC_UNLOCK(ifq);
324
325 ifnet_unlock();
326
327 if ((error = copyout((caddr_t)&stats, ubuf, sizeof(stats))) != 0)
328 return (error);
329 *nbytes = sizeof(stats);
330 return (0);
331 }
332
333 /*
334 * bring the interface back to the initial state by discarding
335 * all the filters and classes except the root class.
336 */
337 static int
hfsc_clear_interface(struct hfsc_if * hif)338 hfsc_clear_interface(struct hfsc_if *hif)
339 {
340 struct hfsc_class *cl;
341
342 if (hif->hif_rootclass == NULL)
343 return (0);
344
345
346 /* clear out the classes */
347 while ((cl = hif->hif_rootclass->cl_children) != NULL) {
348 /*
349 * remove the first leaf class found in the hierarchy
350 * then start over
351 */
352 for (; cl != NULL; cl = hfsc_nextclass(cl)) {
353 if (!is_a_parent_class(cl)) {
354 hfsc_class_destroy(cl);
355 break;
356 }
357 }
358 }
359
360 return (0);
361 }
362
363 static int
hfsc_request(struct ifaltq_subque * ifsq,int req,void * arg)364 hfsc_request(struct ifaltq_subque *ifsq, int req, void *arg)
365 {
366 struct ifaltq *ifq = ifsq->ifsq_altq;
367 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
368
369 crit_enter();
370 switch (req) {
371 case ALTRQ_PURGE:
372 if (ifsq_get_index(ifsq) == HFSC_SUBQ_INDEX) {
373 hfsc_purge(hif);
374 } else {
375 /*
376 * Race happened, the unrelated subqueue was
377 * picked during the packet scheduler transition.
378 */
379 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
380 }
381 break;
382 }
383 crit_exit();
384 return (0);
385 }
386
387 /* discard all the queued packets on the interface */
388 static void
hfsc_purge(struct hfsc_if * hif)389 hfsc_purge(struct hfsc_if *hif)
390 {
391 struct hfsc_class *cl;
392
393 for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl)) {
394 if (!qempty(cl->cl_q))
395 hfsc_purgeq(cl);
396 }
397 if (ifq_is_enabled(hif->hif_ifq))
398 ALTQ_SQ_CNTR_RESET(&hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX]);
399 }
400
401 static struct hfsc_class *
hfsc_class_create(struct hfsc_if * hif,struct service_curve * rsc,struct service_curve * fsc,struct service_curve * usc,struct hfsc_class * parent,int qlimit,int flags,int qid)402 hfsc_class_create(struct hfsc_if *hif, struct service_curve *rsc,
403 struct service_curve *fsc, struct service_curve *usc,
404 struct hfsc_class *parent, int qlimit, int flags, int qid)
405 {
406 struct hfsc_class *cl, *p;
407 int i;
408
409 if (hif->hif_classes >= HFSC_MAX_CLASSES)
410 return (NULL);
411
412 #ifndef ALTQ_RED
413 if (flags & HFCF_RED) {
414 #ifdef ALTQ_DEBUG
415 kprintf("hfsc_class_create: RED not configured for HFSC!\n");
416 #endif
417 return (NULL);
418 }
419 #endif
420
421 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO);
422 cl->cl_q = kmalloc(sizeof(*cl->cl_q), M_ALTQ, M_WAITOK | M_ZERO);
423 cl->cl_actc = actlist_alloc();
424
425 if (qlimit == 0)
426 qlimit = 50; /* use default */
427 qlimit(cl->cl_q) = qlimit;
428 qtype(cl->cl_q) = Q_DROPTAIL;
429 qlen(cl->cl_q) = 0;
430 cl->cl_flags = flags;
431 #ifdef ALTQ_RED
432 if (flags & (HFCF_RED|HFCF_RIO)) {
433 int red_flags, red_pkttime;
434 u_int m2;
435
436 m2 = 0;
437 if (rsc != NULL && rsc->m2 > m2)
438 m2 = rsc->m2;
439 if (fsc != NULL && fsc->m2 > m2)
440 m2 = fsc->m2;
441 if (usc != NULL && usc->m2 > m2)
442 m2 = usc->m2;
443
444 red_flags = 0;
445 if (flags & HFCF_ECN)
446 red_flags |= REDF_ECN;
447 #ifdef ALTQ_RIO
448 if (flags & HFCF_CLEARDSCP)
449 red_flags |= RIOF_CLEARDSCP;
450 #endif
451 if (m2 < 8)
452 red_pkttime = 1000 * 1000 * 1000; /* 1 sec */
453 else
454 red_pkttime = (int64_t)hif->hif_ifq->altq_ifp->if_mtu
455 * 1000 * 1000 * 1000 / (m2 / 8);
456 if (flags & HFCF_RED) {
457 cl->cl_red = red_alloc(0, 0,
458 qlimit(cl->cl_q) * 10/100,
459 qlimit(cl->cl_q) * 30/100,
460 red_flags, red_pkttime);
461 if (cl->cl_red != NULL)
462 qtype(cl->cl_q) = Q_RED;
463 }
464 #ifdef ALTQ_RIO
465 else {
466 cl->cl_red = (red_t *)rio_alloc(0, NULL,
467 red_flags, red_pkttime);
468 if (cl->cl_red != NULL)
469 qtype(cl->cl_q) = Q_RIO;
470 }
471 #endif
472 }
473 #endif /* ALTQ_RED */
474
475 if (rsc != NULL && (rsc->m1 != 0 || rsc->m2 != 0)) {
476 cl->cl_rsc = kmalloc(sizeof(*cl->cl_rsc), M_ALTQ, M_WAITOK);
477 sc2isc(rsc, cl->cl_rsc);
478 rtsc_init(&cl->cl_deadline, cl->cl_rsc, 0, 0);
479 rtsc_init(&cl->cl_eligible, cl->cl_rsc, 0, 0);
480 }
481 if (fsc != NULL && (fsc->m1 != 0 || fsc->m2 != 0)) {
482 cl->cl_fsc = kmalloc(sizeof(*cl->cl_fsc), M_ALTQ, M_WAITOK);
483 sc2isc(fsc, cl->cl_fsc);
484 rtsc_init(&cl->cl_virtual, cl->cl_fsc, 0, 0);
485 }
486 if (usc != NULL && (usc->m1 != 0 || usc->m2 != 0)) {
487 cl->cl_usc = kmalloc(sizeof(*cl->cl_usc), M_ALTQ, M_WAITOK);
488 sc2isc(usc, cl->cl_usc);
489 rtsc_init(&cl->cl_ulimit, cl->cl_usc, 0, 0);
490 }
491
492 cl->cl_id = hif->hif_classid++;
493 cl->cl_handle = qid;
494 cl->cl_hif = hif;
495 cl->cl_parent = parent;
496
497 crit_enter();
498 hif->hif_classes++;
499
500 /*
501 * find a free slot in the class table. if the slot matching
502 * the lower bits of qid is free, use this slot. otherwise,
503 * use the first free slot.
504 */
505 i = qid % HFSC_MAX_CLASSES;
506 if (hif->hif_class_tbl[i] == NULL)
507 hif->hif_class_tbl[i] = cl;
508 else {
509 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
510 if (hif->hif_class_tbl[i] == NULL) {
511 hif->hif_class_tbl[i] = cl;
512 break;
513 }
514 }
515 if (i == HFSC_MAX_CLASSES) {
516 crit_exit();
517 goto err_ret;
518 }
519 }
520
521 if (flags & HFCF_DEFAULTCLASS)
522 hif->hif_defaultclass = cl;
523
524 if (parent == NULL) {
525 /* this is root class */
526 hif->hif_rootclass = cl;
527 } else if (parent->cl_children == NULL) {
528 /* add this class to the children list of the parent */
529 parent->cl_children = cl;
530 } else {
531 p = parent->cl_children;
532 while (p->cl_siblings != NULL)
533 p = p->cl_siblings;
534 p->cl_siblings = cl;
535 }
536 crit_exit();
537
538 return (cl);
539
540 err_ret:
541 if (cl->cl_actc != NULL)
542 actlist_destroy(cl->cl_actc);
543 if (cl->cl_red != NULL) {
544 #ifdef ALTQ_RIO
545 if (q_is_rio(cl->cl_q))
546 rio_destroy((rio_t *)cl->cl_red);
547 #endif
548 #ifdef ALTQ_RED
549 if (q_is_red(cl->cl_q))
550 red_destroy(cl->cl_red);
551 #endif
552 }
553 if (cl->cl_fsc != NULL)
554 kfree(cl->cl_fsc, M_ALTQ);
555 if (cl->cl_rsc != NULL)
556 kfree(cl->cl_rsc, M_ALTQ);
557 if (cl->cl_usc != NULL)
558 kfree(cl->cl_usc, M_ALTQ);
559 if (cl->cl_q != NULL)
560 kfree(cl->cl_q, M_ALTQ);
561 kfree(cl, M_ALTQ);
562 return (NULL);
563 }
564
565 static int
hfsc_class_destroy(struct hfsc_class * cl)566 hfsc_class_destroy(struct hfsc_class *cl)
567 {
568 struct hfsc_if *hif;
569 int i;
570
571 if (cl == NULL)
572 return (0);
573 hif = cl->cl_hif;
574
575 if (is_a_parent_class(cl))
576 return (EBUSY);
577
578 crit_enter();
579
580 if (!qempty(cl->cl_q))
581 hfsc_purgeq(cl);
582
583 if (cl->cl_parent == NULL) {
584 /* this is root class */
585 } else {
586 struct hfsc_class *p = cl->cl_parent->cl_children;
587
588 if (p == cl) {
589 cl->cl_parent->cl_children = cl->cl_siblings;
590 } else {
591 do {
592 if (p->cl_siblings == cl) {
593 p->cl_siblings = cl->cl_siblings;
594 break;
595 }
596 } while ((p = p->cl_siblings) != NULL);
597 }
598 KKASSERT(p != NULL);
599 }
600
601 for (i = 0; i < HFSC_MAX_CLASSES; i++) {
602 if (hif->hif_class_tbl[i] == cl) {
603 hif->hif_class_tbl[i] = NULL;
604 break;
605 }
606 }
607
608 hif->hif_classes--;
609 crit_exit();
610
611 actlist_destroy(cl->cl_actc);
612
613 if (cl->cl_red != NULL) {
614 #ifdef ALTQ_RIO
615 if (q_is_rio(cl->cl_q))
616 rio_destroy((rio_t *)cl->cl_red);
617 #endif
618 #ifdef ALTQ_RED
619 if (q_is_red(cl->cl_q))
620 red_destroy(cl->cl_red);
621 #endif
622 }
623
624 if (cl == hif->hif_rootclass)
625 hif->hif_rootclass = NULL;
626 if (cl == hif->hif_defaultclass)
627 hif->hif_defaultclass = NULL;
628 if (cl == hif->hif_pollcache)
629 hif->hif_pollcache = NULL;
630
631 if (cl->cl_usc != NULL)
632 kfree(cl->cl_usc, M_ALTQ);
633 if (cl->cl_fsc != NULL)
634 kfree(cl->cl_fsc, M_ALTQ);
635 if (cl->cl_rsc != NULL)
636 kfree(cl->cl_rsc, M_ALTQ);
637 kfree(cl->cl_q, M_ALTQ);
638 kfree(cl, M_ALTQ);
639
640 return (0);
641 }
642
643 /*
644 * hfsc_nextclass returns the next class in the tree.
645 * usage:
646 * for (cl = hif->hif_rootclass; cl != NULL; cl = hfsc_nextclass(cl))
647 * do_something;
648 */
649 static struct hfsc_class *
hfsc_nextclass(struct hfsc_class * cl)650 hfsc_nextclass(struct hfsc_class *cl)
651 {
652 if (cl->cl_children != NULL) {
653 cl = cl->cl_children;
654 } else if (cl->cl_siblings != NULL) {
655 cl = cl->cl_siblings;
656 } else {
657 while ((cl = cl->cl_parent) != NULL) {
658 if (cl->cl_siblings != NULL) {
659 cl = cl->cl_siblings;
660 break;
661 }
662 }
663 }
664
665 return (cl);
666 }
667
668 /*
669 * hfsc_enqueue is an enqueue function to be registered to
670 * (*ifsq_enqueue) in struct ifaltq_subque.
671 */
672 static int
hfsc_enqueue(struct ifaltq_subque * ifsq,struct mbuf * m,struct altq_pktattr * pktattr)673 hfsc_enqueue(struct ifaltq_subque *ifsq, struct mbuf *m,
674 struct altq_pktattr *pktattr)
675 {
676 struct ifaltq *ifq = ifsq->ifsq_altq;
677 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
678 struct hfsc_class *cl;
679 int len;
680
681 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
682 /*
683 * Race happened, the unrelated subqueue was
684 * picked during the packet scheduler transition.
685 */
686 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
687 m_freem(m);
688 return ENOBUFS;
689 }
690
691 /* grab class set by classifier */
692 M_ASSERTPKTHDR(m);
693 crit_enter();
694 if (m->m_pkthdr.fw_flags & PF_MBUF_STRUCTURE)
695 cl = clh_to_clp(hif, m->m_pkthdr.pf.qid);
696 else
697 cl = NULL;
698 if (cl == NULL || is_a_parent_class(cl)) {
699 cl = hif->hif_defaultclass;
700 if (cl == NULL) {
701 m_freem(m);
702 crit_exit();
703 return (ENOBUFS);
704 }
705 }
706 cl->cl_pktattr = NULL;
707 len = m_pktlen(m);
708 if (hfsc_addq(cl, m) != 0) {
709 /* drop occurred. mbuf was freed in hfsc_addq. */
710 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, len);
711 crit_exit();
712 return (ENOBUFS);
713 }
714 ALTQ_SQ_PKTCNT_INC(ifsq);
715 cl->cl_hif->hif_packets++;
716
717 /* successfully queued. */
718 if (qlen(cl->cl_q) == 1)
719 set_active(cl, m_pktlen(m));
720 crit_exit();
721 return (0);
722 }
723
724 /*
725 * hfsc_dequeue is a dequeue function to be registered to
726 * (*ifsq_dequeue) in struct ifaltq_subque.
727 *
728 * note: ALTDQ_POLL returns the next packet without removing the packet
729 * from the queue. ALTDQ_REMOVE is a normal dequeue operation.
730 */
731 static struct mbuf *
hfsc_dequeue(struct ifaltq_subque * ifsq,int op)732 hfsc_dequeue(struct ifaltq_subque *ifsq, int op)
733 {
734 struct ifaltq *ifq = ifsq->ifsq_altq;
735 struct hfsc_if *hif = (struct hfsc_if *)ifq->altq_disc;
736 struct hfsc_class *cl;
737 struct mbuf *m;
738 int len, next_len;
739 int realtime = 0;
740 uint64_t cur_time;
741
742 if (ifsq_get_index(ifsq) != HFSC_SUBQ_INDEX) {
743 /*
744 * Race happened, the unrelated subqueue was
745 * picked during the packet scheduler transition.
746 */
747 ifsq_classic_request(ifsq, ALTRQ_PURGE, NULL);
748 return NULL;
749 }
750
751 if (hif->hif_packets == 0) {
752 /* no packet in the tree */
753 return (NULL);
754 }
755
756 crit_enter();
757 cur_time = read_machclk();
758
759 if (op == ALTDQ_REMOVE && hif->hif_pollcache != NULL) {
760 cl = hif->hif_pollcache;
761 hif->hif_pollcache = NULL;
762 /* check if the class was scheduled by real-time criteria */
763 if (cl->cl_rsc != NULL)
764 realtime = (cl->cl_e <= cur_time);
765 } else {
766 /*
767 * if there are eligible classes, use real-time criteria.
768 * find the class with the minimum deadline among
769 * the eligible classes.
770 */
771 if ((cl = ellist_get_mindl(hif->hif_eligible, cur_time)) != NULL) {
772 realtime = 1;
773 } else {
774 #ifdef ALTQ_DEBUG
775 int fits = 0;
776 #endif
777 /*
778 * use link-sharing criteria
779 * get the class with the minimum vt in the hierarchy
780 */
781 cl = hif->hif_rootclass;
782 while (is_a_parent_class(cl)) {
783
784 cl = actlist_firstfit(cl, cur_time);
785 if (cl == NULL) {
786 #ifdef ALTQ_DEBUG
787 if (fits > 0)
788 kprintf("%d fit but none found\n",fits);
789 #endif
790 m = NULL;
791 goto done;
792 }
793 /*
794 * update parent's cl_cvtmin.
795 * don't update if the new vt is smaller.
796 */
797 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
798 cl->cl_parent->cl_cvtmin = cl->cl_vt;
799 #ifdef ALTQ_DEBUG
800 fits++;
801 #endif
802 }
803 }
804
805 if (op == ALTDQ_POLL) {
806 #ifdef foo
807 /*
808 * Don't use poll cache; the poll/dequeue
809 * model is no longer applicable to SMP
810 * system. e.g.
811 * CPU-A CPU-B
812 * : :
813 * poll :
814 * : poll
815 * dequeue (+) :
816 *
817 * The dequeue at (+) will hit the poll
818 * cache set by CPU-B.
819 */
820 hif->hif_pollcache = cl;
821 #endif
822 m = hfsc_pollq(cl);
823 goto done;
824 }
825 }
826
827 m = hfsc_getq(cl);
828 if (m == NULL)
829 panic("hfsc_dequeue:");
830 len = m_pktlen(m);
831 cl->cl_hif->hif_packets--;
832 ALTQ_SQ_PKTCNT_DEC(ifsq);
833 PKTCNTR_ADD(&cl->cl_stats.xmit_cnt, len);
834
835 update_vf(cl, len, cur_time);
836 if (realtime)
837 cl->cl_cumul += len;
838
839 if (!qempty(cl->cl_q)) {
840 if (cl->cl_rsc != NULL) {
841 /* update ed */
842 next_len = m_pktlen(qhead(cl->cl_q));
843
844 if (realtime)
845 update_ed(cl, next_len);
846 else
847 update_d(cl, next_len);
848 }
849 } else {
850 /* the class becomes passive */
851 set_passive(cl);
852 }
853 done:
854 crit_exit();
855 return (m);
856 }
857
858 static int
hfsc_addq(struct hfsc_class * cl,struct mbuf * m)859 hfsc_addq(struct hfsc_class *cl, struct mbuf *m)
860 {
861
862 #ifdef ALTQ_RIO
863 if (q_is_rio(cl->cl_q))
864 return rio_addq((rio_t *)cl->cl_red, cl->cl_q,
865 m, cl->cl_pktattr);
866 #endif
867 #ifdef ALTQ_RED
868 if (q_is_red(cl->cl_q))
869 return red_addq(cl->cl_red, cl->cl_q, m, cl->cl_pktattr);
870 #endif
871 if (qlen(cl->cl_q) >= qlimit(cl->cl_q)) {
872 m_freem(m);
873 return (-1);
874 }
875
876 if (cl->cl_flags & HFCF_CLEARDSCP)
877 write_dsfield(m, cl->cl_pktattr, 0);
878
879 _addq(cl->cl_q, m);
880
881 return (0);
882 }
883
884 static struct mbuf *
hfsc_getq(struct hfsc_class * cl)885 hfsc_getq(struct hfsc_class *cl)
886 {
887 #ifdef ALTQ_RIO
888 if (q_is_rio(cl->cl_q))
889 return rio_getq((rio_t *)cl->cl_red, cl->cl_q);
890 #endif
891 #ifdef ALTQ_RED
892 if (q_is_red(cl->cl_q))
893 return red_getq(cl->cl_red, cl->cl_q);
894 #endif
895 return _getq(cl->cl_q);
896 }
897
898 static struct mbuf *
hfsc_pollq(struct hfsc_class * cl)899 hfsc_pollq(struct hfsc_class *cl)
900 {
901 return qhead(cl->cl_q);
902 }
903
904 static void
hfsc_purgeq(struct hfsc_class * cl)905 hfsc_purgeq(struct hfsc_class *cl)
906 {
907 struct mbuf *m;
908
909 if (qempty(cl->cl_q))
910 return;
911
912 while ((m = _getq(cl->cl_q)) != NULL) {
913 ALTQ_SQ_PKTCNT_DEC(
914 &cl->cl_hif->hif_ifq->altq_subq[HFSC_SUBQ_INDEX]);
915 PKTCNTR_ADD(&cl->cl_stats.drop_cnt, m_pktlen(m));
916 m_freem(m);
917 cl->cl_hif->hif_packets--;
918 }
919 KKASSERT(qlen(cl->cl_q) == 0);
920
921 update_vf(cl, 0, 0); /* remove cl from the actlist */
922 set_passive(cl);
923 }
924
925 static void
set_active(struct hfsc_class * cl,int len)926 set_active(struct hfsc_class *cl, int len)
927 {
928 if (cl->cl_rsc != NULL)
929 init_ed(cl, len);
930 if (cl->cl_fsc != NULL)
931 init_vf(cl, len);
932
933 cl->cl_stats.period++;
934 }
935
936 static void
set_passive(struct hfsc_class * cl)937 set_passive(struct hfsc_class *cl)
938 {
939 if (cl->cl_rsc != NULL)
940 ellist_remove(cl);
941
942 /*
943 * actlist is now handled in update_vf() so that update_vf(cl, 0, 0)
944 * needs to be called explicitly to remove a class from actlist
945 */
946 }
947
948 static void
init_ed(struct hfsc_class * cl,int next_len)949 init_ed(struct hfsc_class *cl, int next_len)
950 {
951 uint64_t cur_time;
952
953 cur_time = read_machclk();
954
955 /* update the deadline curve */
956 rtsc_min(&cl->cl_deadline, cl->cl_rsc, cur_time, cl->cl_cumul);
957
958 /*
959 * update the eligible curve.
960 * for concave, it is equal to the deadline curve.
961 * for convex, it is a linear curve with slope m2.
962 */
963 cl->cl_eligible = cl->cl_deadline;
964 if (cl->cl_rsc->sm1 <= cl->cl_rsc->sm2) {
965 cl->cl_eligible.dx = 0;
966 cl->cl_eligible.dy = 0;
967 }
968
969 /* compute e and d */
970 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
971 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
972
973 ellist_insert(cl);
974 }
975
976 static void
update_ed(struct hfsc_class * cl,int next_len)977 update_ed(struct hfsc_class *cl, int next_len)
978 {
979 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
980 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
981
982 ellist_update(cl);
983 }
984
985 static void
update_d(struct hfsc_class * cl,int next_len)986 update_d(struct hfsc_class *cl, int next_len)
987 {
988 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
989 }
990
991 static void
init_vf(struct hfsc_class * cl,int len)992 init_vf(struct hfsc_class *cl, int len)
993 {
994 struct hfsc_class *max_cl, *p;
995 uint64_t vt, f, cur_time;
996 int go_active;
997
998 cur_time = 0;
999 go_active = 1;
1000 for ( ; cl->cl_parent != NULL; cl = cl->cl_parent) {
1001 if (go_active && cl->cl_nactive++ == 0)
1002 go_active = 1;
1003 else
1004 go_active = 0;
1005
1006 if (go_active) {
1007 max_cl = actlist_last(cl->cl_parent->cl_actc);
1008 if (max_cl != NULL) {
1009 /*
1010 * set vt to the average of the min and max
1011 * classes. if the parent's period didn't
1012 * change, don't decrease vt of the class.
1013 */
1014 vt = max_cl->cl_vt;
1015 if (cl->cl_parent->cl_cvtmin != 0)
1016 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
1017
1018 if (cl->cl_parent->cl_vtperiod !=
1019 cl->cl_parentperiod || vt > cl->cl_vt)
1020 cl->cl_vt = vt;
1021 } else {
1022 /*
1023 * first child for a new parent backlog period.
1024 * add parent's cvtmax to vtoff of children
1025 * to make a new vt (vtoff + vt) larger than
1026 * the vt in the last period for all children.
1027 */
1028 vt = cl->cl_parent->cl_cvtmax;
1029 for (p = cl->cl_parent->cl_children; p != NULL;
1030 p = p->cl_siblings)
1031 p->cl_vtoff += vt;
1032 cl->cl_vt = 0;
1033 cl->cl_parent->cl_cvtmax = 0;
1034 cl->cl_parent->cl_cvtmin = 0;
1035 }
1036 cl->cl_initvt = cl->cl_vt;
1037
1038 /* update the virtual curve */
1039 vt = cl->cl_vt + cl->cl_vtoff;
1040 rtsc_min(&cl->cl_virtual, cl->cl_fsc, vt, cl->cl_total);
1041 if (cl->cl_virtual.x == vt) {
1042 cl->cl_virtual.x -= cl->cl_vtoff;
1043 cl->cl_vtoff = 0;
1044 }
1045 cl->cl_vtadj = 0;
1046
1047 cl->cl_vtperiod++; /* increment vt period */
1048 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
1049 if (cl->cl_parent->cl_nactive == 0)
1050 cl->cl_parentperiod++;
1051 cl->cl_f = 0;
1052
1053 actlist_insert(cl);
1054
1055 if (cl->cl_usc != NULL) {
1056 /* class has upper limit curve */
1057 if (cur_time == 0)
1058 cur_time = read_machclk();
1059
1060 /* update the ulimit curve */
1061 rtsc_min(&cl->cl_ulimit, cl->cl_usc, cur_time,
1062 cl->cl_total);
1063 /* compute myf */
1064 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
1065 cl->cl_total);
1066 cl->cl_myfadj = 0;
1067 }
1068 }
1069
1070 if (cl->cl_myf > cl->cl_cfmin)
1071 f = cl->cl_myf;
1072 else
1073 f = cl->cl_cfmin;
1074 if (f != cl->cl_f) {
1075 cl->cl_f = f;
1076 update_cfmin(cl->cl_parent);
1077 }
1078 }
1079 }
1080
1081 static void
update_vf(struct hfsc_class * cl,int len,uint64_t cur_time)1082 update_vf(struct hfsc_class *cl, int len, uint64_t cur_time)
1083 {
1084 uint64_t f, myf_bound, delta;
1085 int go_passive;
1086
1087 go_passive = qempty(cl->cl_q);
1088
1089 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
1090 cl->cl_total += len;
1091
1092 if (cl->cl_fsc == NULL || cl->cl_nactive == 0)
1093 continue;
1094
1095 if (go_passive && --cl->cl_nactive == 0)
1096 go_passive = 1;
1097 else
1098 go_passive = 0;
1099
1100 if (go_passive) {
1101 /* no more active child, going passive */
1102
1103 /* update cvtmax of the parent class */
1104 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
1105 cl->cl_parent->cl_cvtmax = cl->cl_vt;
1106
1107 /* remove this class from the vt list */
1108 actlist_remove(cl);
1109
1110 update_cfmin(cl->cl_parent);
1111
1112 continue;
1113 }
1114
1115 /*
1116 * update vt and f
1117 */
1118 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
1119 - cl->cl_vtoff + cl->cl_vtadj;
1120
1121 /*
1122 * if vt of the class is smaller than cvtmin,
1123 * the class was skipped in the past due to non-fit.
1124 * if so, we need to adjust vtadj.
1125 */
1126 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
1127 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
1128 cl->cl_vt = cl->cl_parent->cl_cvtmin;
1129 }
1130
1131 /* update the vt list */
1132 actlist_update(cl);
1133
1134 if (cl->cl_usc != NULL) {
1135 cl->cl_myf = cl->cl_myfadj
1136 + rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
1137
1138 /*
1139 * if myf lags behind by more than one clock tick
1140 * from the current time, adjust myfadj to prevent
1141 * a rate-limited class from going greedy.
1142 * in a steady state under rate-limiting, myf
1143 * fluctuates within one clock tick.
1144 */
1145 myf_bound = cur_time - machclk_per_tick;
1146 if (cl->cl_myf < myf_bound) {
1147 delta = cur_time - cl->cl_myf;
1148 cl->cl_myfadj += delta;
1149 cl->cl_myf += delta;
1150 }
1151 }
1152
1153 /* cl_f is max(cl_myf, cl_cfmin) */
1154 if (cl->cl_myf > cl->cl_cfmin)
1155 f = cl->cl_myf;
1156 else
1157 f = cl->cl_cfmin;
1158 if (f != cl->cl_f) {
1159 cl->cl_f = f;
1160 update_cfmin(cl->cl_parent);
1161 }
1162 }
1163 }
1164
1165 static void
update_cfmin(struct hfsc_class * cl)1166 update_cfmin(struct hfsc_class *cl)
1167 {
1168 struct hfsc_class *p;
1169 uint64_t cfmin;
1170
1171 if (TAILQ_EMPTY(cl->cl_actc)) {
1172 cl->cl_cfmin = 0;
1173 return;
1174 }
1175 cfmin = HT_INFINITY;
1176 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1177 if (p->cl_f == 0) {
1178 cl->cl_cfmin = 0;
1179 return;
1180 }
1181 if (p->cl_f < cfmin)
1182 cfmin = p->cl_f;
1183 }
1184 cl->cl_cfmin = cfmin;
1185 }
1186
1187 /*
1188 * TAILQ based ellist and actlist implementation
1189 * (ion wanted to make a calendar queue based implementation)
1190 */
1191 /*
1192 * eligible list holds backlogged classes being sorted by their eligible times.
1193 * there is one eligible list per interface.
1194 */
1195
1196 static ellist_t *
ellist_alloc(void)1197 ellist_alloc(void)
1198 {
1199 ellist_t *head;
1200
1201 head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
1202 TAILQ_INIT(head);
1203 return (head);
1204 }
1205
1206 static void
ellist_destroy(ellist_t * head)1207 ellist_destroy(ellist_t *head)
1208 {
1209 kfree(head, M_ALTQ);
1210 }
1211
1212 static void
ellist_insert(struct hfsc_class * cl)1213 ellist_insert(struct hfsc_class *cl)
1214 {
1215 struct hfsc_if *hif = cl->cl_hif;
1216 struct hfsc_class *p;
1217
1218 /* check the last entry first */
1219 if ((p = TAILQ_LAST(hif->hif_eligible, _eligible)) == NULL ||
1220 p->cl_e <= cl->cl_e) {
1221 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1222 return;
1223 }
1224
1225 TAILQ_FOREACH(p, hif->hif_eligible, cl_ellist) {
1226 if (cl->cl_e < p->cl_e) {
1227 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1228 return;
1229 }
1230 }
1231 KKASSERT(0); /* should not reach here */
1232 }
1233
1234 static void
ellist_remove(struct hfsc_class * cl)1235 ellist_remove(struct hfsc_class *cl)
1236 {
1237 struct hfsc_if *hif = cl->cl_hif;
1238
1239 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1240 }
1241
1242 static void
ellist_update(struct hfsc_class * cl)1243 ellist_update(struct hfsc_class *cl)
1244 {
1245 struct hfsc_if *hif = cl->cl_hif;
1246 struct hfsc_class *p, *last;
1247
1248 /*
1249 * the eligible time of a class increases monotonically.
1250 * if the next entry has a larger eligible time, nothing to do.
1251 */
1252 p = TAILQ_NEXT(cl, cl_ellist);
1253 if (p == NULL || cl->cl_e <= p->cl_e)
1254 return;
1255
1256 /* check the last entry */
1257 last = TAILQ_LAST(hif->hif_eligible, _eligible);
1258 KKASSERT(last != NULL);
1259 if (last->cl_e <= cl->cl_e) {
1260 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1261 TAILQ_INSERT_TAIL(hif->hif_eligible, cl, cl_ellist);
1262 return;
1263 }
1264
1265 /*
1266 * the new position must be between the next entry
1267 * and the last entry
1268 */
1269 while ((p = TAILQ_NEXT(p, cl_ellist)) != NULL) {
1270 if (cl->cl_e < p->cl_e) {
1271 TAILQ_REMOVE(hif->hif_eligible, cl, cl_ellist);
1272 TAILQ_INSERT_BEFORE(p, cl, cl_ellist);
1273 return;
1274 }
1275 }
1276 KKASSERT(0); /* should not reach here */
1277 }
1278
1279 /* find the class with the minimum deadline among the eligible classes */
1280 struct hfsc_class *
ellist_get_mindl(ellist_t * head,uint64_t cur_time)1281 ellist_get_mindl(ellist_t *head, uint64_t cur_time)
1282 {
1283 struct hfsc_class *p, *cl = NULL;
1284
1285 TAILQ_FOREACH(p, head, cl_ellist) {
1286 if (p->cl_e > cur_time)
1287 break;
1288 if (cl == NULL || p->cl_d < cl->cl_d)
1289 cl = p;
1290 }
1291 return (cl);
1292 }
1293
1294 /*
1295 * active children list holds backlogged child classes being sorted
1296 * by their virtual time.
1297 * each intermediate class has one active children list.
1298 */
1299 static actlist_t *
actlist_alloc(void)1300 actlist_alloc(void)
1301 {
1302 actlist_t *head;
1303
1304 head = kmalloc(sizeof(*head), M_ALTQ, M_WAITOK);
1305 TAILQ_INIT(head);
1306 return (head);
1307 }
1308
1309 static void
actlist_destroy(actlist_t * head)1310 actlist_destroy(actlist_t *head)
1311 {
1312 kfree(head, M_ALTQ);
1313 }
1314 static void
actlist_insert(struct hfsc_class * cl)1315 actlist_insert(struct hfsc_class *cl)
1316 {
1317 struct hfsc_class *p;
1318
1319 /* check the last entry first */
1320 if ((p = TAILQ_LAST(cl->cl_parent->cl_actc, _active)) == NULL
1321 || p->cl_vt <= cl->cl_vt) {
1322 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1323 return;
1324 }
1325
1326 TAILQ_FOREACH(p, cl->cl_parent->cl_actc, cl_actlist) {
1327 if (cl->cl_vt < p->cl_vt) {
1328 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1329 return;
1330 }
1331 }
1332 KKASSERT(0); /* should not reach here */
1333 }
1334
1335 static void
actlist_remove(struct hfsc_class * cl)1336 actlist_remove(struct hfsc_class *cl)
1337 {
1338 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1339 }
1340
1341 static void
actlist_update(struct hfsc_class * cl)1342 actlist_update(struct hfsc_class *cl)
1343 {
1344 struct hfsc_class *p, *last;
1345
1346 /*
1347 * the virtual time of a class increases monotonically during its
1348 * backlogged period.
1349 * if the next entry has a larger virtual time, nothing to do.
1350 */
1351 p = TAILQ_NEXT(cl, cl_actlist);
1352 if (p == NULL || cl->cl_vt < p->cl_vt)
1353 return;
1354
1355 /* check the last entry */
1356 last = TAILQ_LAST(cl->cl_parent->cl_actc, _active);
1357 KKASSERT(last != NULL);
1358 if (last->cl_vt <= cl->cl_vt) {
1359 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1360 TAILQ_INSERT_TAIL(cl->cl_parent->cl_actc, cl, cl_actlist);
1361 return;
1362 }
1363
1364 /*
1365 * the new position must be between the next entry
1366 * and the last entry
1367 */
1368 while ((p = TAILQ_NEXT(p, cl_actlist)) != NULL) {
1369 if (cl->cl_vt < p->cl_vt) {
1370 TAILQ_REMOVE(cl->cl_parent->cl_actc, cl, cl_actlist);
1371 TAILQ_INSERT_BEFORE(p, cl, cl_actlist);
1372 return;
1373 }
1374 }
1375 KKASSERT(0); /* should not reach here */
1376 }
1377
1378 static struct hfsc_class *
actlist_firstfit(struct hfsc_class * cl,uint64_t cur_time)1379 actlist_firstfit(struct hfsc_class *cl, uint64_t cur_time)
1380 {
1381 struct hfsc_class *p;
1382
1383 TAILQ_FOREACH(p, cl->cl_actc, cl_actlist) {
1384 if (p->cl_f <= cur_time)
1385 return (p);
1386 }
1387 return (NULL);
1388 }
1389
1390 /*
1391 * service curve support functions
1392 *
1393 * external service curve parameters
1394 * m: bits/sec
1395 * d: msec
1396 * internal service curve parameters
1397 * sm: (bytes/tsc_interval) << SM_SHIFT
1398 * ism: (tsc_count/byte) << ISM_SHIFT
1399 * dx: tsc_count
1400 *
1401 * SM_SHIFT and ISM_SHIFT are scaled in order to keep effective digits.
1402 * we should be able to handle 100K-1Gbps linkspeed with 200Hz-1GHz CPU
1403 * speed. SM_SHIFT and ISM_SHIFT are selected to have at least 3 effective
1404 * digits in decimal using the following table.
1405 *
1406 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
1407 * ----------+-------------------------------------------------------
1408 * bytes/nsec 12.5e-6 125e-6 1250e-6 12500e-6 125000e-6
1409 * sm(500MHz) 25.0e-6 250e-6 2500e-6 25000e-6 250000e-6
1410 * sm(200MHz) 62.5e-6 625e-6 6250e-6 62500e-6 625000e-6
1411 *
1412 * nsec/byte 80000 8000 800 80 8
1413 * ism(500MHz) 40000 4000 400 40 4
1414 * ism(200MHz) 16000 1600 160 16 1.6
1415 */
1416 #define SM_SHIFT 24
1417 #define ISM_SHIFT 10
1418
1419 #define SM_MASK ((1LL << SM_SHIFT) - 1)
1420 #define ISM_MASK ((1LL << ISM_SHIFT) - 1)
1421
1422 static __inline uint64_t
seg_x2y(uint64_t x,uint64_t sm)1423 seg_x2y(uint64_t x, uint64_t sm)
1424 {
1425 uint64_t y;
1426
1427 /*
1428 * compute
1429 * y = x * sm >> SM_SHIFT
1430 * but divide it for the upper and lower bits to avoid overflow
1431 */
1432 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
1433 return (y);
1434 }
1435
1436 static __inline uint64_t
seg_y2x(uint64_t y,uint64_t ism)1437 seg_y2x(uint64_t y, uint64_t ism)
1438 {
1439 uint64_t x;
1440
1441 if (y == 0)
1442 x = 0;
1443 else if (ism == HT_INFINITY)
1444 x = HT_INFINITY;
1445 else
1446 x = (y >> ISM_SHIFT) * ism + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
1447
1448 return (x);
1449 }
1450
1451 static __inline uint64_t
m2sm(u_int m)1452 m2sm(u_int m)
1453 {
1454 uint64_t sm;
1455
1456 sm = ((uint64_t)m << SM_SHIFT) / 8 / machclk_freq;
1457 return (sm);
1458 }
1459
1460 static __inline uint64_t
m2ism(u_int m)1461 m2ism(u_int m)
1462 {
1463 uint64_t ism;
1464
1465 if (m == 0)
1466 ism = HT_INFINITY;
1467 else
1468 ism = ((uint64_t)machclk_freq << ISM_SHIFT) * 8 / m;
1469 return (ism);
1470 }
1471
1472 static __inline uint64_t
d2dx(u_int d)1473 d2dx(u_int d)
1474 {
1475 uint64_t dx;
1476
1477 dx = ((uint64_t)d * machclk_freq) / 1000;
1478 return (dx);
1479 }
1480
1481 static u_int
sm2m(uint64_t sm)1482 sm2m(uint64_t sm)
1483 {
1484 uint64_t m;
1485
1486 m = (sm * 8 * machclk_freq) >> SM_SHIFT;
1487 return ((u_int)m);
1488 }
1489
1490 static u_int
dx2d(uint64_t dx)1491 dx2d(uint64_t dx)
1492 {
1493 uint64_t d;
1494
1495 d = dx * 1000 / machclk_freq;
1496 return ((u_int)d);
1497 }
1498
1499 static void
sc2isc(struct service_curve * sc,struct internal_sc * isc)1500 sc2isc(struct service_curve *sc, struct internal_sc *isc)
1501 {
1502 isc->sm1 = m2sm(sc->m1);
1503 isc->ism1 = m2ism(sc->m1);
1504 isc->dx = d2dx(sc->d);
1505 isc->dy = seg_x2y(isc->dx, isc->sm1);
1506 isc->sm2 = m2sm(sc->m2);
1507 isc->ism2 = m2ism(sc->m2);
1508 }
1509
1510 /*
1511 * initialize the runtime service curve with the given internal
1512 * service curve starting at (x, y).
1513 */
1514 static void
rtsc_init(struct runtime_sc * rtsc,struct internal_sc * isc,uint64_t x,uint64_t y)1515 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1516 {
1517 rtsc->x = x;
1518 rtsc->y = y;
1519 rtsc->sm1 = isc->sm1;
1520 rtsc->ism1 = isc->ism1;
1521 rtsc->dx = isc->dx;
1522 rtsc->dy = isc->dy;
1523 rtsc->sm2 = isc->sm2;
1524 rtsc->ism2 = isc->ism2;
1525 }
1526
1527 /*
1528 * calculate the y-projection of the runtime service curve by the
1529 * given x-projection value
1530 */
1531 static uint64_t
rtsc_y2x(struct runtime_sc * rtsc,uint64_t y)1532 rtsc_y2x(struct runtime_sc *rtsc, uint64_t y)
1533 {
1534 uint64_t x;
1535
1536 if (y < rtsc->y) {
1537 x = rtsc->x;
1538 } else if (y <= rtsc->y + rtsc->dy) {
1539 /* x belongs to the 1st segment */
1540 if (rtsc->dy == 0)
1541 x = rtsc->x + rtsc->dx;
1542 else
1543 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
1544 } else {
1545 /* x belongs to the 2nd segment */
1546 x = rtsc->x + rtsc->dx
1547 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
1548 }
1549 return (x);
1550 }
1551
1552 static uint64_t
rtsc_x2y(struct runtime_sc * rtsc,uint64_t x)1553 rtsc_x2y(struct runtime_sc *rtsc, uint64_t x)
1554 {
1555 uint64_t y;
1556
1557 if (x <= rtsc->x) {
1558 y = rtsc->y;
1559 } else if (x <= rtsc->x + rtsc->dx) {
1560 /* y belongs to the 1st segment */
1561 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
1562 } else
1563 /* y belongs to the 2nd segment */
1564 y = rtsc->y + rtsc->dy
1565 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
1566 return (y);
1567 }
1568
1569 /*
1570 * update the runtime service curve by taking the minimum of the current
1571 * runtime service curve and the service curve starting at (x, y).
1572 */
1573 static void
rtsc_min(struct runtime_sc * rtsc,struct internal_sc * isc,uint64_t x,uint64_t y)1574 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, uint64_t x, uint64_t y)
1575 {
1576 uint64_t y1, y2, dx, dy;
1577
1578 if (isc->sm1 <= isc->sm2) {
1579 /* service curve is convex */
1580 y1 = rtsc_x2y(rtsc, x);
1581 if (y1 < y)
1582 /* the current rtsc is smaller */
1583 return;
1584 rtsc->x = x;
1585 rtsc->y = y;
1586 return;
1587 }
1588
1589 /*
1590 * service curve is concave
1591 * compute the two y values of the current rtsc
1592 * y1: at x
1593 * y2: at (x + dx)
1594 */
1595 y1 = rtsc_x2y(rtsc, x);
1596 if (y1 <= y) {
1597 /* rtsc is below isc, no change to rtsc */
1598 return;
1599 }
1600
1601 y2 = rtsc_x2y(rtsc, x + isc->dx);
1602 if (y2 >= y + isc->dy) {
1603 /* rtsc is above isc, replace rtsc by isc */
1604 rtsc->x = x;
1605 rtsc->y = y;
1606 rtsc->dx = isc->dx;
1607 rtsc->dy = isc->dy;
1608 return;
1609 }
1610
1611 /*
1612 * the two curves intersect
1613 * compute the offsets (dx, dy) using the reverse
1614 * function of seg_x2y()
1615 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
1616 */
1617 dx = ((y1 - y) << SM_SHIFT) / (isc->sm1 - isc->sm2);
1618 /*
1619 * check if (x, y1) belongs to the 1st segment of rtsc.
1620 * if so, add the offset.
1621 */
1622 if (rtsc->x + rtsc->dx > x)
1623 dx += rtsc->x + rtsc->dx - x;
1624 dy = seg_x2y(dx, isc->sm1);
1625
1626 rtsc->x = x;
1627 rtsc->y = y;
1628 rtsc->dx = dx;
1629 rtsc->dy = dy;
1630 }
1631
1632 static void
get_class_stats(struct hfsc_classstats * sp,struct hfsc_class * cl)1633 get_class_stats(struct hfsc_classstats *sp, struct hfsc_class *cl)
1634 {
1635 sp->class_id = cl->cl_id;
1636 sp->class_handle = cl->cl_handle;
1637
1638 if (cl->cl_rsc != NULL) {
1639 sp->rsc.m1 = sm2m(cl->cl_rsc->sm1);
1640 sp->rsc.d = dx2d(cl->cl_rsc->dx);
1641 sp->rsc.m2 = sm2m(cl->cl_rsc->sm2);
1642 } else {
1643 sp->rsc.m1 = 0;
1644 sp->rsc.d = 0;
1645 sp->rsc.m2 = 0;
1646 }
1647 if (cl->cl_fsc != NULL) {
1648 sp->fsc.m1 = sm2m(cl->cl_fsc->sm1);
1649 sp->fsc.d = dx2d(cl->cl_fsc->dx);
1650 sp->fsc.m2 = sm2m(cl->cl_fsc->sm2);
1651 } else {
1652 sp->fsc.m1 = 0;
1653 sp->fsc.d = 0;
1654 sp->fsc.m2 = 0;
1655 }
1656 if (cl->cl_usc != NULL) {
1657 sp->usc.m1 = sm2m(cl->cl_usc->sm1);
1658 sp->usc.d = dx2d(cl->cl_usc->dx);
1659 sp->usc.m2 = sm2m(cl->cl_usc->sm2);
1660 } else {
1661 sp->usc.m1 = 0;
1662 sp->usc.d = 0;
1663 sp->usc.m2 = 0;
1664 }
1665
1666 sp->total = cl->cl_total;
1667 sp->cumul = cl->cl_cumul;
1668
1669 sp->d = cl->cl_d;
1670 sp->e = cl->cl_e;
1671 sp->vt = cl->cl_vt;
1672 sp->f = cl->cl_f;
1673
1674 sp->initvt = cl->cl_initvt;
1675 sp->vtperiod = cl->cl_vtperiod;
1676 sp->parentperiod = cl->cl_parentperiod;
1677 sp->nactive = cl->cl_nactive;
1678 sp->vtoff = cl->cl_vtoff;
1679 sp->cvtmax = cl->cl_cvtmax;
1680 sp->myf = cl->cl_myf;
1681 sp->cfmin = cl->cl_cfmin;
1682 sp->cvtmin = cl->cl_cvtmin;
1683 sp->myfadj = cl->cl_myfadj;
1684 sp->vtadj = cl->cl_vtadj;
1685
1686 sp->cur_time = read_machclk();
1687 sp->machclk_freq = machclk_freq;
1688
1689 sp->qlength = qlen(cl->cl_q);
1690 sp->qlimit = qlimit(cl->cl_q);
1691 sp->xmit_cnt = cl->cl_stats.xmit_cnt;
1692 sp->drop_cnt = cl->cl_stats.drop_cnt;
1693 sp->period = cl->cl_stats.period;
1694
1695 sp->qtype = qtype(cl->cl_q);
1696 #ifdef ALTQ_RED
1697 if (q_is_red(cl->cl_q))
1698 red_getstats(cl->cl_red, &sp->red[0]);
1699 #endif
1700 #ifdef ALTQ_RIO
1701 if (q_is_rio(cl->cl_q))
1702 rio_getstats((rio_t *)cl->cl_red, &sp->red[0]);
1703 #endif
1704 }
1705
1706 /* convert a class handle to the corresponding class pointer */
1707 static struct hfsc_class *
clh_to_clp(struct hfsc_if * hif,uint32_t chandle)1708 clh_to_clp(struct hfsc_if *hif, uint32_t chandle)
1709 {
1710 int i;
1711 struct hfsc_class *cl;
1712
1713 if (chandle == 0)
1714 return (NULL);
1715 /*
1716 * first, try optimistically the slot matching the lower bits of
1717 * the handle. if it fails, do the linear table search.
1718 */
1719 i = chandle % HFSC_MAX_CLASSES;
1720 if ((cl = hif->hif_class_tbl[i]) != NULL && cl->cl_handle == chandle)
1721 return (cl);
1722 for (i = 0; i < HFSC_MAX_CLASSES; i++)
1723 if ((cl = hif->hif_class_tbl[i]) != NULL &&
1724 cl->cl_handle == chandle)
1725 return (cl);
1726 return (NULL);
1727 }
1728
1729 #endif /* ALTQ_HFSC */
1730