projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge git://git.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[deliverable/linux.git] / net / sched / sch_hfsc.c
1 /*
2 * Copyright (c) 2003 Patrick McHardy, <kaber@trash.net>
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version 2
7 * of the License, or (at your option) any later version.
8 *
9 * 2003-10-17 - Ported from altq
10 */
11 /*
12 * Copyright (c) 1997-1999 Carnegie Mellon University. All Rights Reserved.
13 *
14 * Permission to use, copy, modify, and distribute this software and
15 * its documentation is hereby granted (including for commercial or
16 * for-profit use), provided that both the copyright notice and this
17 * permission notice appear in all copies of the software, derivative
18 * works, or modified versions, and any portions thereof.
19 *
20 * THIS SOFTWARE IS EXPERIMENTAL AND IS KNOWN TO HAVE BUGS, SOME OF
21 * WHICH MAY HAVE SERIOUS CONSEQUENCES. CARNEGIE MELLON PROVIDES THIS
22 * SOFTWARE IN ITS ``AS IS'' CONDITION, AND ANY EXPRESS OR IMPLIED
23 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
24 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25 * DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT
28 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
29 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
30 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
32 * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
33 * DAMAGE.
34 *
35 * Carnegie Mellon encourages (but does not require) users of this
36 * software to return any improvements or extensions that they make,
37 * and to grant Carnegie Mellon the rights to redistribute these
38 * changes without encumbrance.
39 */
40 /*
41 * H-FSC is described in Proceedings of SIGCOMM'97,
42 * "A Hierarchical Fair Service Curve Algorithm for Link-Sharing,
43 * Real-Time and Priority Service"
44 * by Ion Stoica, Hui Zhang, and T. S. Eugene Ng.
45 *
46 * Oleg Cherevko <olwi@aq.ml.com.ua> added the upperlimit for link-sharing.
47 * when a class has an upperlimit, the fit-time is computed from the
48 * upperlimit service curve. the link-sharing scheduler does not schedule
49 * a class whose fit-time exceeds the current time.
50 */
51
52 #include <linux/kernel.h>
53 #include <linux/module.h>
54 #include <linux/types.h>
55 #include <linux/errno.h>
56 #include <linux/jiffies.h>
57 #include <linux/compiler.h>
58 #include <linux/spinlock.h>
59 #include <linux/skbuff.h>
60 #include <linux/string.h>
61 #include <linux/slab.h>
62 #include <linux/timer.h>
63 #include <linux/list.h>
64 #include <linux/rbtree.h>
65 #include <linux/init.h>
66 #include <linux/netdevice.h>
67 #include <linux/rtnetlink.h>
68 #include <linux/pkt_sched.h>
69 #include <net/pkt_sched.h>
70 #include <net/pkt_cls.h>
71 #include <asm/system.h>
72 #include <asm/div64.h>
73
74 #define HFSC_DEBUG 1
75
76 /*
77 * kernel internal service curve representation:
78 * coordinates are given by 64 bit unsigned integers.
79 * x-axis: unit is clock count.
80 * y-axis: unit is byte.
81 *
82 * The service curve parameters are converted to the internal
83 * representation. The slope values are scaled to avoid overflow.
84 * the inverse slope values as well as the y-projection of the 1st
85 * segment are kept in order to to avoid 64-bit divide operations
86 * that are expensive on 32-bit architectures.
87 */
88
89 struct internal_sc
90 {
91 u64 sm1; /* scaled slope of the 1st segment */
92 u64 ism1; /* scaled inverse-slope of the 1st segment */
93 u64 dx; /* the x-projection of the 1st segment */
94 u64 dy; /* the y-projection of the 1st segment */
95 u64 sm2; /* scaled slope of the 2nd segment */
96 u64 ism2; /* scaled inverse-slope of the 2nd segment */
97 };
98
99 /* runtime service curve */
100 struct runtime_sc
101 {
102 u64 x; /* current starting position on x-axis */
103 u64 y; /* current starting position on y-axis */
104 u64 sm1; /* scaled slope of the 1st segment */
105 u64 ism1; /* scaled inverse-slope of the 1st segment */
106 u64 dx; /* the x-projection of the 1st segment */
107 u64 dy; /* the y-projection of the 1st segment */
108 u64 sm2; /* scaled slope of the 2nd segment */
109 u64 ism2; /* scaled inverse-slope of the 2nd segment */
110 };
111
112 enum hfsc_class_flags
113 {
114 HFSC_RSC = 0x1,
115 HFSC_FSC = 0x2,
116 HFSC_USC = 0x4
117 };
118
119 struct hfsc_class
120 {
121 u32 classid; /* class id */
122 unsigned int refcnt; /* usage count */
123
124 struct gnet_stats_basic bstats;
125 struct gnet_stats_queue qstats;
126 struct gnet_stats_rate_est rate_est;
127 spinlock_t *stats_lock;
128 unsigned int level; /* class level in hierarchy */
129 struct tcf_proto *filter_list; /* filter list */
130 unsigned int filter_cnt; /* filter count */
131
132 struct hfsc_sched *sched; /* scheduler data */
133 struct hfsc_class *cl_parent; /* parent class */
134 struct list_head siblings; /* sibling classes */
135 struct list_head children; /* child classes */
136 struct Qdisc *qdisc; /* leaf qdisc */
137
138 struct rb_node el_node; /* qdisc's eligible tree member */
139 struct rb_root vt_tree; /* active children sorted by cl_vt */
140 struct rb_node vt_node; /* parent's vt_tree member */
141 struct rb_root cf_tree; /* active children sorted by cl_f */
142 struct rb_node cf_node; /* parent's cf_heap member */
143 struct list_head hlist; /* hash list member */
144 struct list_head dlist; /* drop list member */
145
146 u64 cl_total; /* total work in bytes */
147 u64 cl_cumul; /* cumulative work in bytes done by
148 real-time criteria */
149
150 u64 cl_d; /* deadline*/
151 u64 cl_e; /* eligible time */
152 u64 cl_vt; /* virtual time */
153 u64 cl_f; /* time when this class will fit for
154 link-sharing, max(myf, cfmin) */
155 u64 cl_myf; /* my fit-time (calculated from this
156 class's own upperlimit curve) */
157 u64 cl_myfadj; /* my fit-time adjustment (to cancel
158 history dependence) */
159 u64 cl_cfmin; /* earliest children's fit-time (used
160 with cl_myf to obtain cl_f) */
161 u64 cl_cvtmin; /* minimal virtual time among the
162 children fit for link-sharing
163 (monotonic within a period) */
164 u64 cl_vtadj; /* intra-period cumulative vt
165 adjustment */
166 u64 cl_vtoff; /* inter-period cumulative vt offset */
167 u64 cl_cvtmax; /* max child's vt in the last period */
168 u64 cl_cvtoff; /* cumulative cvtmax of all periods */
169 u64 cl_pcvtoff; /* parent's cvtoff at initalization
170 time */
171
172 struct internal_sc cl_rsc; /* internal real-time service curve */
173 struct internal_sc cl_fsc; /* internal fair service curve */
174 struct internal_sc cl_usc; /* internal upperlimit service curve */
175 struct runtime_sc cl_deadline; /* deadline curve */
176 struct runtime_sc cl_eligible; /* eligible curve */
177 struct runtime_sc cl_virtual; /* virtual curve */
178 struct runtime_sc cl_ulimit; /* upperlimit curve */
179
180 unsigned long cl_flags; /* which curves are valid */
181 unsigned long cl_vtperiod; /* vt period sequence number */
182 unsigned long cl_parentperiod;/* parent's vt period sequence number*/
183 unsigned long cl_nactive; /* number of active children */
184 };
185
186 #define HFSC_HSIZE 16
187
188 struct hfsc_sched
189 {
190 u16 defcls; /* default class id */
191 struct hfsc_class root; /* root class */
192 struct list_head clhash[HFSC_HSIZE]; /* class hash */
193 struct rb_root eligible; /* eligible tree */
194 struct list_head droplist; /* active leaf class list (for
195 dropping) */
196 struct sk_buff_head requeue; /* requeued packet */
197 struct timer_list wd_timer; /* watchdog timer */
198 };
199
200 /*
201 * macros
202 */
203 #ifdef CONFIG_NET_SCH_CLK_GETTIMEOFDAY
204 #include <linux/time.h>
205 #undef PSCHED_GET_TIME
206 #define PSCHED_GET_TIME(stamp) \
207 do { \
208 struct timeval tv; \
209 do_gettimeofday(&tv); \
210 (stamp) = 1ULL * USEC_PER_SEC * tv.tv_sec + tv.tv_usec; \
211 } while (0)
212 #endif
213
214 #if HFSC_DEBUG
215 #define ASSERT(cond) \
216 do { \
217 if (unlikely(!(cond))) \
218 printk("assertion %s failed at %s:%i (%s)\n", \
219 #cond, __FILE__, __LINE__, __FUNCTION__); \
220 } while (0)
221 #else
222 #define ASSERT(cond)
223 #endif /* HFSC_DEBUG */
224
225 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
226
227
228 /*
229 * eligible tree holds backlogged classes being sorted by their eligible times.
230 * there is one eligible tree per hfsc instance.
231 */
232
233 static void
234 eltree_insert(struct hfsc_class *cl)
235 {
236 struct rb_node **p = &cl->sched->eligible.rb_node;
237 struct rb_node *parent = NULL;
238 struct hfsc_class *cl1;
239
240 while (*p != NULL) {
241 parent = *p;
242 cl1 = rb_entry(parent, struct hfsc_class, el_node);
243 if (cl->cl_e >= cl1->cl_e)
244 p = &parent->rb_right;
245 else
246 p = &parent->rb_left;
247 }
248 rb_link_node(&cl->el_node, parent, p);
249 rb_insert_color(&cl->el_node, &cl->sched->eligible);
250 }
251
252 static inline void
253 eltree_remove(struct hfsc_class *cl)
254 {
255 rb_erase(&cl->el_node, &cl->sched->eligible);
256 }
257
258 static inline void
259 eltree_update(struct hfsc_class *cl)
260 {
261 eltree_remove(cl);
262 eltree_insert(cl);
263 }
264
265 /* find the class with the minimum deadline among the eligible classes */
266 static inline struct hfsc_class *
267 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
268 {
269 struct hfsc_class *p, *cl = NULL;
270 struct rb_node *n;
271
272 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
273 p = rb_entry(n, struct hfsc_class, el_node);
274 if (p->cl_e > cur_time)
275 break;
276 if (cl == NULL || p->cl_d < cl->cl_d)
277 cl = p;
278 }
279 return cl;
280 }
281
282 /* find the class with minimum eligible time among the eligible classes */
283 static inline struct hfsc_class *
284 eltree_get_minel(struct hfsc_sched *q)
285 {
286 struct rb_node *n;
287
288 n = rb_first(&q->eligible);
289 if (n == NULL)
290 return NULL;
291 return rb_entry(n, struct hfsc_class, el_node);
292 }
293
294 /*
295 * vttree holds holds backlogged child classes being sorted by their virtual
296 * time. each intermediate class has one vttree.
297 */
298 static void
299 vttree_insert(struct hfsc_class *cl)
300 {
301 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
302 struct rb_node *parent = NULL;
303 struct hfsc_class *cl1;
304
305 while (*p != NULL) {
306 parent = *p;
307 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
308 if (cl->cl_vt >= cl1->cl_vt)
309 p = &parent->rb_right;
310 else
311 p = &parent->rb_left;
312 }
313 rb_link_node(&cl->vt_node, parent, p);
314 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
315 }
316
317 static inline void
318 vttree_remove(struct hfsc_class *cl)
319 {
320 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
321 }
322
323 static inline void
324 vttree_update(struct hfsc_class *cl)
325 {
326 vttree_remove(cl);
327 vttree_insert(cl);
328 }
329
330 static inline struct hfsc_class *
331 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
332 {
333 struct hfsc_class *p;
334 struct rb_node *n;
335
336 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
337 p = rb_entry(n, struct hfsc_class, vt_node);
338 if (p->cl_f <= cur_time)
339 return p;
340 }
341 return NULL;
342 }
343
344 /*
345 * get the leaf class with the minimum vt in the hierarchy
346 */
347 static struct hfsc_class *
348 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
349 {
350 /* if root-class's cfmin is bigger than cur_time nothing to do */
351 if (cl->cl_cfmin > cur_time)
352 return NULL;
353
354 while (cl->level > 0) {
355 cl = vttree_firstfit(cl, cur_time);
356 if (cl == NULL)
357 return NULL;
358 /*
359 * update parent's cl_cvtmin.
360 */
361 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
362 cl->cl_parent->cl_cvtmin = cl->cl_vt;
363 }
364 return cl;
365 }
366
367 static void
368 cftree_insert(struct hfsc_class *cl)
369 {
370 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
371 struct rb_node *parent = NULL;
372 struct hfsc_class *cl1;
373
374 while (*p != NULL) {
375 parent = *p;
376 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
377 if (cl->cl_f >= cl1->cl_f)
378 p = &parent->rb_right;
379 else
380 p = &parent->rb_left;
381 }
382 rb_link_node(&cl->cf_node, parent, p);
383 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
384 }
385
386 static inline void
387 cftree_remove(struct hfsc_class *cl)
388 {
389 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
390 }
391
392 static inline void
393 cftree_update(struct hfsc_class *cl)
394 {
395 cftree_remove(cl);
396 cftree_insert(cl);
397 }
398
399 /*
400 * service curve support functions
401 *
402 * external service curve parameters
403 * m: bps
404 * d: us
405 * internal service curve parameters
406 * sm: (bytes/psched_us) << SM_SHIFT
407 * ism: (psched_us/byte) << ISM_SHIFT
408 * dx: psched_us
409 *
410 * Clock source resolution (CONFIG_NET_SCH_CLK_*)
411 * JIFFIES: for 48<=HZ<=1534 resolution is between 0.63us and 1.27us.
412 * CPU: resolution is between 0.5us and 1us.
413 * GETTIMEOFDAY: resolution is exactly 1us.
414 *
415 * sm and ism are scaled in order to keep effective digits.
416 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
417 * digits in decimal using the following table.
418 *
419 * Note: We can afford the additional accuracy (altq hfsc keeps at most
420 * 3 effective digits) thanks to the fact that linux clock is bounded
421 * much more tightly.
422 *
423 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
424 * ------------+-------------------------------------------------------
425 * bytes/0.5us 6.25e-3 62.5e-3 625e-3 6250e-e 62500e-3
426 * bytes/us 12.5e-3 125e-3 1250e-3 12500e-3 125000e-3
427 * bytes/1.27us 15.875e-3 158.75e-3 1587.5e-3 15875e-3 158750e-3
428 *
429 * 0.5us/byte 160 16 1.6 0.16 0.016
430 * us/byte 80 8 0.8 0.08 0.008
431 * 1.27us/byte 63 6.3 0.63 0.063 0.0063
432 */
433 #define SM_SHIFT 20
434 #define ISM_SHIFT 18
435
436 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
437 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
438
439 static inline u64
440 seg_x2y(u64 x, u64 sm)
441 {
442 u64 y;
443
444 /*
445 * compute
446 * y = x * sm >> SM_SHIFT
447 * but divide it for the upper and lower bits to avoid overflow
448 */
449 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
450 return y;
451 }
452
453 static inline u64
454 seg_y2x(u64 y, u64 ism)
455 {
456 u64 x;
457
458 if (y == 0)
459 x = 0;
460 else if (ism == HT_INFINITY)
461 x = HT_INFINITY;
462 else {
463 x = (y >> ISM_SHIFT) * ism
464 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
465 }
466 return x;
467 }
468
469 /* Convert m (bps) into sm (bytes/psched us) */
470 static u64
471 m2sm(u32 m)
472 {
473 u64 sm;
474
475 sm = ((u64)m << SM_SHIFT);
476 sm += PSCHED_JIFFIE2US(HZ) - 1;
477 do_div(sm, PSCHED_JIFFIE2US(HZ));
478 return sm;
479 }
480
481 /* convert m (bps) into ism (psched us/byte) */
482 static u64
483 m2ism(u32 m)
484 {
485 u64 ism;
486
487 if (m == 0)
488 ism = HT_INFINITY;
489 else {
490 ism = ((u64)PSCHED_JIFFIE2US(HZ) << ISM_SHIFT);
491 ism += m - 1;
492 do_div(ism, m);
493 }
494 return ism;
495 }
496
497 /* convert d (us) into dx (psched us) */
498 static u64
499 d2dx(u32 d)
500 {
501 u64 dx;
502
503 dx = ((u64)d * PSCHED_JIFFIE2US(HZ));
504 dx += USEC_PER_SEC - 1;
505 do_div(dx, USEC_PER_SEC);
506 return dx;
507 }
508
509 /* convert sm (bytes/psched us) into m (bps) */
510 static u32
511 sm2m(u64 sm)
512 {
513 u64 m;
514
515 m = (sm * PSCHED_JIFFIE2US(HZ)) >> SM_SHIFT;
516 return (u32)m;
517 }
518
519 /* convert dx (psched us) into d (us) */
520 static u32
521 dx2d(u64 dx)
522 {
523 u64 d;
524
525 d = dx * USEC_PER_SEC;
526 do_div(d, PSCHED_JIFFIE2US(HZ));
527 return (u32)d;
528 }
529
530 static void
531 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
532 {
533 isc->sm1 = m2sm(sc->m1);
534 isc->ism1 = m2ism(sc->m1);
535 isc->dx = d2dx(sc->d);
536 isc->dy = seg_x2y(isc->dx, isc->sm1);
537 isc->sm2 = m2sm(sc->m2);
538 isc->ism2 = m2ism(sc->m2);
539 }
540
541 /*
542 * initialize the runtime service curve with the given internal
543 * service curve starting at (x, y).
544 */
545 static void
546 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
547 {
548 rtsc->x = x;
549 rtsc->y = y;
550 rtsc->sm1 = isc->sm1;
551 rtsc->ism1 = isc->ism1;
552 rtsc->dx = isc->dx;
553 rtsc->dy = isc->dy;
554 rtsc->sm2 = isc->sm2;
555 rtsc->ism2 = isc->ism2;
556 }
557
558 /*
559 * calculate the y-projection of the runtime service curve by the
560 * given x-projection value
561 */
562 static u64
563 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
564 {
565 u64 x;
566
567 if (y < rtsc->y)
568 x = rtsc->x;
569 else if (y <= rtsc->y + rtsc->dy) {
570 /* x belongs to the 1st segment */
571 if (rtsc->dy == 0)
572 x = rtsc->x + rtsc->dx;
573 else
574 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
575 } else {
576 /* x belongs to the 2nd segment */
577 x = rtsc->x + rtsc->dx
578 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
579 }
580 return x;
581 }
582
583 static u64
584 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
585 {
586 u64 y;
587
588 if (x <= rtsc->x)
589 y = rtsc->y;
590 else if (x <= rtsc->x + rtsc->dx)
591 /* y belongs to the 1st segment */
592 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
593 else
594 /* y belongs to the 2nd segment */
595 y = rtsc->y + rtsc->dy
596 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
597 return y;
598 }
599
600 /*
601 * update the runtime service curve by taking the minimum of the current
602 * runtime service curve and the service curve starting at (x, y).
603 */
604 static void
605 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
606 {
607 u64 y1, y2, dx, dy;
608 u32 dsm;
609
610 if (isc->sm1 <= isc->sm2) {
611 /* service curve is convex */
612 y1 = rtsc_x2y(rtsc, x);
613 if (y1 < y)
614 /* the current rtsc is smaller */
615 return;
616 rtsc->x = x;
617 rtsc->y = y;
618 return;
619 }
620
621 /*
622 * service curve is concave
623 * compute the two y values of the current rtsc
624 * y1: at x
625 * y2: at (x + dx)
626 */
627 y1 = rtsc_x2y(rtsc, x);
628 if (y1 <= y) {
629 /* rtsc is below isc, no change to rtsc */
630 return;
631 }
632
633 y2 = rtsc_x2y(rtsc, x + isc->dx);
634 if (y2 >= y + isc->dy) {
635 /* rtsc is above isc, replace rtsc by isc */
636 rtsc->x = x;
637 rtsc->y = y;
638 rtsc->dx = isc->dx;
639 rtsc->dy = isc->dy;
640 return;
641 }
642
643 /*
644 * the two curves intersect
645 * compute the offsets (dx, dy) using the reverse
646 * function of seg_x2y()
647 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
648 */
649 dx = (y1 - y) << SM_SHIFT;
650 dsm = isc->sm1 - isc->sm2;
651 do_div(dx, dsm);
652 /*
653 * check if (x, y1) belongs to the 1st segment of rtsc.
654 * if so, add the offset.
655 */
656 if (rtsc->x + rtsc->dx > x)
657 dx += rtsc->x + rtsc->dx - x;
658 dy = seg_x2y(dx, isc->sm1);
659
660 rtsc->x = x;
661 rtsc->y = y;
662 rtsc->dx = dx;
663 rtsc->dy = dy;
664 return;
665 }
666
667 static void
668 init_ed(struct hfsc_class *cl, unsigned int next_len)
669 {
670 u64 cur_time;
671
672 PSCHED_GET_TIME(cur_time);
673
674 /* update the deadline curve */
675 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
676
677 /*
678 * update the eligible curve.
679 * for concave, it is equal to the deadline curve.
680 * for convex, it is a linear curve with slope m2.
681 */
682 cl->cl_eligible = cl->cl_deadline;
683 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
684 cl->cl_eligible.dx = 0;
685 cl->cl_eligible.dy = 0;
686 }
687
688 /* compute e and d */
689 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
690 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
691
692 eltree_insert(cl);
693 }
694
695 static void
696 update_ed(struct hfsc_class *cl, unsigned int next_len)
697 {
698 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
699 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
700
701 eltree_update(cl);
702 }
703
704 static inline void
705 update_d(struct hfsc_class *cl, unsigned int next_len)
706 {
707 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
708 }
709
710 static inline void
711 update_cfmin(struct hfsc_class *cl)
712 {
713 struct rb_node *n = rb_first(&cl->cf_tree);
714 struct hfsc_class *p;
715
716 if (n == NULL) {
717 cl->cl_cfmin = 0;
718 return;
719 }
720 p = rb_entry(n, struct hfsc_class, cf_node);
721 cl->cl_cfmin = p->cl_f;
722 }
723
724 static void
725 init_vf(struct hfsc_class *cl, unsigned int len)
726 {
727 struct hfsc_class *max_cl;
728 struct rb_node *n;
729 u64 vt, f, cur_time;
730 int go_active;
731
732 cur_time = 0;
733 go_active = 1;
734 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
735 if (go_active && cl->cl_nactive++ == 0)
736 go_active = 1;
737 else
738 go_active = 0;
739
740 if (go_active) {
741 n = rb_last(&cl->cl_parent->vt_tree);
742 if (n != NULL) {
743 max_cl = rb_entry(n, struct hfsc_class,vt_node);
744 /*
745 * set vt to the average of the min and max
746 * classes. if the parent's period didn't
747 * change, don't decrease vt of the class.
748 */
749 vt = max_cl->cl_vt;
750 if (cl->cl_parent->cl_cvtmin != 0)
751 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
752
753 if (cl->cl_parent->cl_vtperiod !=
754 cl->cl_parentperiod || vt > cl->cl_vt)
755 cl->cl_vt = vt;
756 } else {
757 /*
758 * first child for a new parent backlog period.
759 * add parent's cvtmax to cvtoff to make a new
760 * vt (vtoff + vt) larger than the vt in the
761 * last period for all children.
762 */
763 vt = cl->cl_parent->cl_cvtmax;
764 cl->cl_parent->cl_cvtoff += vt;
765 cl->cl_parent->cl_cvtmax = 0;
766 cl->cl_parent->cl_cvtmin = 0;
767 cl->cl_vt = 0;
768 }
769
770 cl->cl_vtoff = cl->cl_parent->cl_cvtoff -
771 cl->cl_pcvtoff;
772
773 /* update the virtual curve */
774 vt = cl->cl_vt + cl->cl_vtoff;
775 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt,
776 cl->cl_total);
777 if (cl->cl_virtual.x == vt) {
778 cl->cl_virtual.x -= cl->cl_vtoff;
779 cl->cl_vtoff = 0;
780 }
781 cl->cl_vtadj = 0;
782
783 cl->cl_vtperiod++; /* increment vt period */
784 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
785 if (cl->cl_parent->cl_nactive == 0)
786 cl->cl_parentperiod++;
787 cl->cl_f = 0;
788
789 vttree_insert(cl);
790 cftree_insert(cl);
791
792 if (cl->cl_flags & HFSC_USC) {
793 /* class has upper limit curve */
794 if (cur_time == 0)
795 PSCHED_GET_TIME(cur_time);
796
797 /* update the ulimit curve */
798 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
799 cl->cl_total);
800 /* compute myf */
801 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
802 cl->cl_total);
803 cl->cl_myfadj = 0;
804 }
805 }
806
807 f = max(cl->cl_myf, cl->cl_cfmin);
808 if (f != cl->cl_f) {
809 cl->cl_f = f;
810 cftree_update(cl);
811 update_cfmin(cl->cl_parent);
812 }
813 }
814 }
815
816 static void
817 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
818 {
819 u64 f; /* , myf_bound, delta; */
820 int go_passive = 0;
821
822 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
823 go_passive = 1;
824
825 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
826 cl->cl_total += len;
827
828 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
829 continue;
830
831 if (go_passive && --cl->cl_nactive == 0)
832 go_passive = 1;
833 else
834 go_passive = 0;
835
836 if (go_passive) {
837 /* no more active child, going passive */
838
839 /* update cvtmax of the parent class */
840 if (cl->cl_vt > cl->cl_parent->cl_cvtmax)
841 cl->cl_parent->cl_cvtmax = cl->cl_vt;
842
843 /* remove this class from the vt tree */
844 vttree_remove(cl);
845
846 cftree_remove(cl);
847 update_cfmin(cl->cl_parent);
848
849 continue;
850 }
851
852 /*
853 * update vt and f
854 */
855 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total)
856 - cl->cl_vtoff + cl->cl_vtadj;
857
858 /*
859 * if vt of the class is smaller than cvtmin,
860 * the class was skipped in the past due to non-fit.
861 * if so, we need to adjust vtadj.
862 */
863 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
864 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
865 cl->cl_vt = cl->cl_parent->cl_cvtmin;
866 }
867
868 /* update the vt tree */
869 vttree_update(cl);
870
871 if (cl->cl_flags & HFSC_USC) {
872 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
873 cl->cl_total);
874 #if 0
875 /*
876 * This code causes classes to stay way under their
877 * limit when multiple classes are used at gigabit
878 * speed. needs investigation. -kaber
879 */
880 /*
881 * if myf lags behind by more than one clock tick
882 * from the current time, adjust myfadj to prevent
883 * a rate-limited class from going greedy.
884 * in a steady state under rate-limiting, myf
885 * fluctuates within one clock tick.
886 */
887 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
888 if (cl->cl_myf < myf_bound) {
889 delta = cur_time - cl->cl_myf;
890 cl->cl_myfadj += delta;
891 cl->cl_myf += delta;
892 }
893 #endif
894 }
895
896 f = max(cl->cl_myf, cl->cl_cfmin);
897 if (f != cl->cl_f) {
898 cl->cl_f = f;
899 cftree_update(cl);
900 update_cfmin(cl->cl_parent);
901 }
902 }
903 }
904
905 static void
906 set_active(struct hfsc_class *cl, unsigned int len)
907 {
908 if (cl->cl_flags & HFSC_RSC)
909 init_ed(cl, len);
910 if (cl->cl_flags & HFSC_FSC)
911 init_vf(cl, len);
912
913 list_add_tail(&cl->dlist, &cl->sched->droplist);
914 }
915
916 static void
917 set_passive(struct hfsc_class *cl)
918 {
919 if (cl->cl_flags & HFSC_RSC)
920 eltree_remove(cl);
921
922 list_del(&cl->dlist);
923
924 /*
925 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
926 * needs to be called explicitly to remove a class from vttree.
927 */
928 }
929
930 /*
931 * hack to get length of first packet in queue.
932 */
933 static unsigned int
934 qdisc_peek_len(struct Qdisc *sch)
935 {
936 struct sk_buff *skb;
937 unsigned int len;
938
939 skb = sch->dequeue(sch);
940 if (skb == NULL) {
941 if (net_ratelimit())
942 printk("qdisc_peek_len: non work-conserving qdisc ?\n");
943 return 0;
944 }
945 len = skb->len;
946 if (unlikely(sch->ops->requeue(skb, sch) != NET_XMIT_SUCCESS)) {
947 if (net_ratelimit())
948 printk("qdisc_peek_len: failed to requeue\n");
949 return 0;
950 }
951 return len;
952 }
953
954 static void
955 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl)
956 {
957 unsigned int len = cl->qdisc->q.qlen;
958
959 qdisc_reset(cl->qdisc);
960 if (len > 0) {
961 update_vf(cl, 0, 0);
962 set_passive(cl);
963 sch->q.qlen -= len;
964 }
965 }
966
967 static void
968 hfsc_adjust_levels(struct hfsc_class *cl)
969 {
970 struct hfsc_class *p;
971 unsigned int level;
972
973 do {
974 level = 0;
975 list_for_each_entry(p, &cl->children, siblings) {
976 if (p->level >= level)
977 level = p->level + 1;
978 }
979 cl->level = level;
980 } while ((cl = cl->cl_parent) != NULL);
981 }
982
983 static inline unsigned int
984 hfsc_hash(u32 h)
985 {
986 h ^= h >> 8;
987 h ^= h >> 4;
988
989 return h & (HFSC_HSIZE - 1);
990 }
991
992 static inline struct hfsc_class *
993 hfsc_find_class(u32 classid, struct Qdisc *sch)
994 {
995 struct hfsc_sched *q = qdisc_priv(sch);
996 struct hfsc_class *cl;
997
998 list_for_each_entry(cl, &q->clhash[hfsc_hash(classid)], hlist) {
999 if (cl->classid == classid)
1000 return cl;
1001 }
1002 return NULL;
1003 }
1004
1005 static void
1006 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
1007 u64 cur_time)
1008 {
1009 sc2isc(rsc, &cl->cl_rsc);
1010 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
1011 cl->cl_eligible = cl->cl_deadline;
1012 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
1013 cl->cl_eligible.dx = 0;
1014 cl->cl_eligible.dy = 0;
1015 }
1016 cl->cl_flags |= HFSC_RSC;
1017 }
1018
1019 static void
1020 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
1021 {
1022 sc2isc(fsc, &cl->cl_fsc);
1023 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
1024 cl->cl_flags |= HFSC_FSC;
1025 }
1026
1027 static void
1028 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
1029 u64 cur_time)
1030 {
1031 sc2isc(usc, &cl->cl_usc);
1032 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
1033 cl->cl_flags |= HFSC_USC;
1034 }
1035
1036 static int
1037 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
1038 struct rtattr **tca, unsigned long *arg)
1039 {
1040 struct hfsc_sched *q = qdisc_priv(sch);
1041 struct hfsc_class *cl = (struct hfsc_class *)*arg;
1042 struct hfsc_class *parent = NULL;
1043 struct rtattr *opt = tca[TCA_OPTIONS-1];
1044 struct rtattr *tb[TCA_HFSC_MAX];
1045 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
1046 u64 cur_time;
1047
1048 if (opt == NULL || rtattr_parse_nested(tb, TCA_HFSC_MAX, opt))
1049 return -EINVAL;
1050
1051 if (tb[TCA_HFSC_RSC-1]) {
1052 if (RTA_PAYLOAD(tb[TCA_HFSC_RSC-1]) < sizeof(*rsc))
1053 return -EINVAL;
1054 rsc = RTA_DATA(tb[TCA_HFSC_RSC-1]);
1055 if (rsc->m1 == 0 && rsc->m2 == 0)
1056 rsc = NULL;
1057 }
1058
1059 if (tb[TCA_HFSC_FSC-1]) {
1060 if (RTA_PAYLOAD(tb[TCA_HFSC_FSC-1]) < sizeof(*fsc))
1061 return -EINVAL;
1062 fsc = RTA_DATA(tb[TCA_HFSC_FSC-1]);
1063 if (fsc->m1 == 0 && fsc->m2 == 0)
1064 fsc = NULL;
1065 }
1066
1067 if (tb[TCA_HFSC_USC-1]) {
1068 if (RTA_PAYLOAD(tb[TCA_HFSC_USC-1]) < sizeof(*usc))
1069 return -EINVAL;
1070 usc = RTA_DATA(tb[TCA_HFSC_USC-1]);
1071 if (usc->m1 == 0 && usc->m2 == 0)
1072 usc = NULL;
1073 }
1074
1075 if (cl != NULL) {
1076 if (parentid) {
1077 if (cl->cl_parent && cl->cl_parent->classid != parentid)
1078 return -EINVAL;
1079 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
1080 return -EINVAL;
1081 }
1082 PSCHED_GET_TIME(cur_time);
1083
1084 sch_tree_lock(sch);
1085 if (rsc != NULL)
1086 hfsc_change_rsc(cl, rsc, cur_time);
1087 if (fsc != NULL)
1088 hfsc_change_fsc(cl, fsc);
1089 if (usc != NULL)
1090 hfsc_change_usc(cl, usc, cur_time);
1091
1092 if (cl->qdisc->q.qlen != 0) {
1093 if (cl->cl_flags & HFSC_RSC)
1094 update_ed(cl, qdisc_peek_len(cl->qdisc));
1095 if (cl->cl_flags & HFSC_FSC)
1096 update_vf(cl, 0, cur_time);
1097 }
1098 sch_tree_unlock(sch);
1099
1100 #ifdef CONFIG_NET_ESTIMATOR
1101 if (tca[TCA_RATE-1])
1102 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1103 cl->stats_lock, tca[TCA_RATE-1]);
1104 #endif
1105 return 0;
1106 }
1107
1108 if (parentid == TC_H_ROOT)
1109 return -EEXIST;
1110
1111 parent = &q->root;
1112 if (parentid) {
1113 parent = hfsc_find_class(parentid, sch);
1114 if (parent == NULL)
1115 return -ENOENT;
1116 }
1117
1118 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1119 return -EINVAL;
1120 if (hfsc_find_class(classid, sch))
1121 return -EEXIST;
1122
1123 if (rsc == NULL && fsc == NULL)
1124 return -EINVAL;
1125
1126 cl = kmalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1127 if (cl == NULL)
1128 return -ENOBUFS;
1129 memset(cl, 0, sizeof(struct hfsc_class));
1130
1131 if (rsc != NULL)
1132 hfsc_change_rsc(cl, rsc, 0);
1133 if (fsc != NULL)
1134 hfsc_change_fsc(cl, fsc);
1135 if (usc != NULL)
1136 hfsc_change_usc(cl, usc, 0);
1137
1138 cl->refcnt = 1;
1139 cl->classid = classid;
1140 cl->sched = q;
1141 cl->cl_parent = parent;
1142 cl->qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1143 if (cl->qdisc == NULL)
1144 cl->qdisc = &noop_qdisc;
1145 cl->stats_lock = &sch->dev->queue_lock;
1146 INIT_LIST_HEAD(&cl->children);
1147 cl->vt_tree = RB_ROOT;
1148 cl->cf_tree = RB_ROOT;
1149
1150 sch_tree_lock(sch);
1151 list_add_tail(&cl->hlist, &q->clhash[hfsc_hash(classid)]);
1152 list_add_tail(&cl->siblings, &parent->children);
1153 if (parent->level == 0)
1154 hfsc_purge_queue(sch, parent);
1155 hfsc_adjust_levels(parent);
1156 cl->cl_pcvtoff = parent->cl_cvtoff;
1157 sch_tree_unlock(sch);
1158
1159 #ifdef CONFIG_NET_ESTIMATOR
1160 if (tca[TCA_RATE-1])
1161 gen_new_estimator(&cl->bstats, &cl->rate_est,
1162 cl->stats_lock, tca[TCA_RATE-1]);
1163 #endif
1164 *arg = (unsigned long)cl;
1165 return 0;
1166 }
1167
1168 static void
1169 hfsc_destroy_filters(struct tcf_proto **fl)
1170 {
1171 struct tcf_proto *tp;
1172
1173 while ((tp = *fl) != NULL) {
1174 *fl = tp->next;
1175 tcf_destroy(tp);
1176 }
1177 }
1178
1179 static void
1180 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1181 {
1182 struct hfsc_sched *q = qdisc_priv(sch);
1183
1184 hfsc_destroy_filters(&cl->filter_list);
1185 qdisc_destroy(cl->qdisc);
1186 #ifdef CONFIG_NET_ESTIMATOR
1187 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1188 #endif
1189 if (cl != &q->root)
1190 kfree(cl);
1191 }
1192
1193 static int
1194 hfsc_delete_class(struct Qdisc *sch, unsigned long arg)
1195 {
1196 struct hfsc_sched *q = qdisc_priv(sch);
1197 struct hfsc_class *cl = (struct hfsc_class *)arg;
1198
1199 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root)
1200 return -EBUSY;
1201
1202 sch_tree_lock(sch);
1203
1204 list_del(&cl->hlist);
1205 list_del(&cl->siblings);
1206 hfsc_adjust_levels(cl->cl_parent);
1207 hfsc_purge_queue(sch, cl);
1208 if (--cl->refcnt == 0)
1209 hfsc_destroy_class(sch, cl);
1210
1211 sch_tree_unlock(sch);
1212 return 0;
1213 }
1214
1215 static struct hfsc_class *
1216 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1217 {
1218 struct hfsc_sched *q = qdisc_priv(sch);
1219 struct hfsc_class *cl;
1220 struct tcf_result res;
1221 struct tcf_proto *tcf;
1222 int result;
1223
1224 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1225 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1226 if (cl->level == 0)
1227 return cl;
1228
1229 *qerr = NET_XMIT_BYPASS;
1230 tcf = q->root.filter_list;
1231 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
1232 #ifdef CONFIG_NET_CLS_ACT
1233 switch (result) {
1234 case TC_ACT_QUEUED:
1235 case TC_ACT_STOLEN:
1236 *qerr = NET_XMIT_SUCCESS;
1237 case TC_ACT_SHOT:
1238 return NULL;
1239 }
1240 #elif defined(CONFIG_NET_CLS_POLICE)
1241 if (result == TC_POLICE_SHOT)
1242 return NULL;
1243 #endif
1244 if ((cl = (struct hfsc_class *)res.class) == NULL) {
1245 if ((cl = hfsc_find_class(res.classid, sch)) == NULL)
1246 break; /* filter selected invalid classid */
1247 }
1248
1249 if (cl->level == 0)
1250 return cl; /* hit leaf class */
1251
1252 /* apply inner filter chain */
1253 tcf = cl->filter_list;
1254 }
1255
1256 /* classification failed, try default class */
1257 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
1258 if (cl == NULL || cl->level > 0)
1259 return NULL;
1260
1261 return cl;
1262 }
1263
1264 static int
1265 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1266 struct Qdisc **old)
1267 {
1268 struct hfsc_class *cl = (struct hfsc_class *)arg;
1269
1270 if (cl == NULL)
1271 return -ENOENT;
1272 if (cl->level > 0)
1273 return -EINVAL;
1274 if (new == NULL) {
1275 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1276 if (new == NULL)
1277 new = &noop_qdisc;
1278 }
1279
1280 sch_tree_lock(sch);
1281 hfsc_purge_queue(sch, cl);
1282 *old = xchg(&cl->qdisc, new);
1283 sch_tree_unlock(sch);
1284 return 0;
1285 }
1286
1287 static struct Qdisc *
1288 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1289 {
1290 struct hfsc_class *cl = (struct hfsc_class *)arg;
1291
1292 if (cl != NULL && cl->level == 0)
1293 return cl->qdisc;
1294
1295 return NULL;
1296 }
1297
1298 static unsigned long
1299 hfsc_get_class(struct Qdisc *sch, u32 classid)
1300 {
1301 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1302
1303 if (cl != NULL)
1304 cl->refcnt++;
1305
1306 return (unsigned long)cl;
1307 }
1308
1309 static void
1310 hfsc_put_class(struct Qdisc *sch, unsigned long arg)
1311 {
1312 struct hfsc_class *cl = (struct hfsc_class *)arg;
1313
1314 if (--cl->refcnt == 0)
1315 hfsc_destroy_class(sch, cl);
1316 }
1317
1318 static unsigned long
1319 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1320 {
1321 struct hfsc_class *p = (struct hfsc_class *)parent;
1322 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1323
1324 if (cl != NULL) {
1325 if (p != NULL && p->level <= cl->level)
1326 return 0;
1327 cl->filter_cnt++;
1328 }
1329
1330 return (unsigned long)cl;
1331 }
1332
1333 static void
1334 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1335 {
1336 struct hfsc_class *cl = (struct hfsc_class *)arg;
1337
1338 cl->filter_cnt--;
1339 }
1340
1341 static struct tcf_proto **
1342 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg)
1343 {
1344 struct hfsc_sched *q = qdisc_priv(sch);
1345 struct hfsc_class *cl = (struct hfsc_class *)arg;
1346
1347 if (cl == NULL)
1348 cl = &q->root;
1349
1350 return &cl->filter_list;
1351 }
1352
1353 static int
1354 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1355 {
1356 struct tc_service_curve tsc;
1357
1358 tsc.m1 = sm2m(sc->sm1);
1359 tsc.d = dx2d(sc->dx);
1360 tsc.m2 = sm2m(sc->sm2);
1361 RTA_PUT(skb, attr, sizeof(tsc), &tsc);
1362
1363 return skb->len;
1364
1365 rtattr_failure:
1366 return -1;
1367 }
1368
1369 static inline int
1370 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1371 {
1372 if ((cl->cl_flags & HFSC_RSC) &&
1373 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1374 goto rtattr_failure;
1375
1376 if ((cl->cl_flags & HFSC_FSC) &&
1377 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1378 goto rtattr_failure;
1379
1380 if ((cl->cl_flags & HFSC_USC) &&
1381 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1382 goto rtattr_failure;
1383
1384 return skb->len;
1385
1386 rtattr_failure:
1387 return -1;
1388 }
1389
1390 static int
1391 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1392 struct tcmsg *tcm)
1393 {
1394 struct hfsc_class *cl = (struct hfsc_class *)arg;
1395 unsigned char *b = skb->tail;
1396 struct rtattr *rta = (struct rtattr *)b;
1397
1398 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->classid : TC_H_ROOT;
1399 tcm->tcm_handle = cl->classid;
1400 if (cl->level == 0)
1401 tcm->tcm_info = cl->qdisc->handle;
1402
1403 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1404 if (hfsc_dump_curves(skb, cl) < 0)
1405 goto rtattr_failure;
1406 rta->rta_len = skb->tail - b;
1407 return skb->len;
1408
1409 rtattr_failure:
1410 skb_trim(skb, b - skb->data);
1411 return -1;
1412 }
1413
1414 static int
1415 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1416 struct gnet_dump *d)
1417 {
1418 struct hfsc_class *cl = (struct hfsc_class *)arg;
1419 struct tc_hfsc_stats xstats;
1420
1421 cl->qstats.qlen = cl->qdisc->q.qlen;
1422 xstats.level = cl->level;
1423 xstats.period = cl->cl_vtperiod;
1424 xstats.work = cl->cl_total;
1425 xstats.rtwork = cl->cl_cumul;
1426
1427 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1428 #ifdef CONFIG_NET_ESTIMATOR
1429 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1430 #endif
1431 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1432 return -1;
1433
1434 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1435 }
1436
1437
1438
1439 static void
1440 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1441 {
1442 struct hfsc_sched *q = qdisc_priv(sch);
1443 struct hfsc_class *cl;
1444 unsigned int i;
1445
1446 if (arg->stop)
1447 return;
1448
1449 for (i = 0; i < HFSC_HSIZE; i++) {
1450 list_for_each_entry(cl, &q->clhash[i], hlist) {
1451 if (arg->count < arg->skip) {
1452 arg->count++;
1453 continue;
1454 }
1455 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1456 arg->stop = 1;
1457 return;
1458 }
1459 arg->count++;
1460 }
1461 }
1462 }
1463
1464 static void
1465 hfsc_watchdog(unsigned long arg)
1466 {
1467 struct Qdisc *sch = (struct Qdisc *)arg;
1468
1469 sch->flags &= ~TCQ_F_THROTTLED;
1470 netif_schedule(sch->dev);
1471 }
1472
1473 static void
1474 hfsc_schedule_watchdog(struct Qdisc *sch, u64 cur_time)
1475 {
1476 struct hfsc_sched *q = qdisc_priv(sch);
1477 struct hfsc_class *cl;
1478 u64 next_time = 0;
1479 long delay;
1480
1481 if ((cl = eltree_get_minel(q)) != NULL)
1482 next_time = cl->cl_e;
1483 if (q->root.cl_cfmin != 0) {
1484 if (next_time == 0 || next_time > q->root.cl_cfmin)
1485 next_time = q->root.cl_cfmin;
1486 }
1487 ASSERT(next_time != 0);
1488 delay = next_time - cur_time;
1489 delay = PSCHED_US2JIFFIE(delay);
1490
1491 sch->flags |= TCQ_F_THROTTLED;
1492 mod_timer(&q->wd_timer, jiffies + delay);
1493 }
1494
1495 static int
1496 hfsc_init_qdisc(struct Qdisc *sch, struct rtattr *opt)
1497 {
1498 struct hfsc_sched *q = qdisc_priv(sch);
1499 struct tc_hfsc_qopt *qopt;
1500 unsigned int i;
1501
1502 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1503 return -EINVAL;
1504 qopt = RTA_DATA(opt);
1505
1506 sch->stats_lock = &sch->dev->queue_lock;
1507
1508 q->defcls = qopt->defcls;
1509 for (i = 0; i < HFSC_HSIZE; i++)
1510 INIT_LIST_HEAD(&q->clhash[i]);
1511 q->eligible = RB_ROOT;
1512 INIT_LIST_HEAD(&q->droplist);
1513 skb_queue_head_init(&q->requeue);
1514
1515 q->root.refcnt = 1;
1516 q->root.classid = sch->handle;
1517 q->root.sched = q;
1518 q->root.qdisc = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1519 if (q->root.qdisc == NULL)
1520 q->root.qdisc = &noop_qdisc;
1521 q->root.stats_lock = &sch->dev->queue_lock;
1522 INIT_LIST_HEAD(&q->root.children);
1523 q->root.vt_tree = RB_ROOT;
1524 q->root.cf_tree = RB_ROOT;
1525
1526 list_add(&q->root.hlist, &q->clhash[hfsc_hash(q->root.classid)]);
1527
1528 init_timer(&q->wd_timer);
1529 q->wd_timer.function = hfsc_watchdog;
1530 q->wd_timer.data = (unsigned long)sch;
1531
1532 return 0;
1533 }
1534
1535 static int
1536 hfsc_change_qdisc(struct Qdisc *sch, struct rtattr *opt)
1537 {
1538 struct hfsc_sched *q = qdisc_priv(sch);
1539 struct tc_hfsc_qopt *qopt;
1540
1541 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
1542 return -EINVAL;
1543 qopt = RTA_DATA(opt);
1544
1545 sch_tree_lock(sch);
1546 q->defcls = qopt->defcls;
1547 sch_tree_unlock(sch);
1548
1549 return 0;
1550 }
1551
1552 static void
1553 hfsc_reset_class(struct hfsc_class *cl)
1554 {
1555 cl->cl_total = 0;
1556 cl->cl_cumul = 0;
1557 cl->cl_d = 0;
1558 cl->cl_e = 0;
1559 cl->cl_vt = 0;
1560 cl->cl_vtadj = 0;
1561 cl->cl_vtoff = 0;
1562 cl->cl_cvtmin = 0;
1563 cl->cl_cvtmax = 0;
1564 cl->cl_cvtoff = 0;
1565 cl->cl_pcvtoff = 0;
1566 cl->cl_vtperiod = 0;
1567 cl->cl_parentperiod = 0;
1568 cl->cl_f = 0;
1569 cl->cl_myf = 0;
1570 cl->cl_myfadj = 0;
1571 cl->cl_cfmin = 0;
1572 cl->cl_nactive = 0;
1573
1574 cl->vt_tree = RB_ROOT;
1575 cl->cf_tree = RB_ROOT;
1576 qdisc_reset(cl->qdisc);
1577
1578 if (cl->cl_flags & HFSC_RSC)
1579 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1580 if (cl->cl_flags & HFSC_FSC)
1581 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1582 if (cl->cl_flags & HFSC_USC)
1583 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1584 }
1585
1586 static void
1587 hfsc_reset_qdisc(struct Qdisc *sch)
1588 {
1589 struct hfsc_sched *q = qdisc_priv(sch);
1590 struct hfsc_class *cl;
1591 unsigned int i;
1592
1593 for (i = 0; i < HFSC_HSIZE; i++) {
1594 list_for_each_entry(cl, &q->clhash[i], hlist)
1595 hfsc_reset_class(cl);
1596 }
1597 __skb_queue_purge(&q->requeue);
1598 q->eligible = RB_ROOT;
1599 INIT_LIST_HEAD(&q->droplist);
1600 del_timer(&q->wd_timer);
1601 sch->flags &= ~TCQ_F_THROTTLED;
1602 sch->q.qlen = 0;
1603 }
1604
1605 static void
1606 hfsc_destroy_qdisc(struct Qdisc *sch)
1607 {
1608 struct hfsc_sched *q = qdisc_priv(sch);
1609 struct hfsc_class *cl, *next;
1610 unsigned int i;
1611
1612 for (i = 0; i < HFSC_HSIZE; i++) {
1613 list_for_each_entry_safe(cl, next, &q->clhash[i], hlist)
1614 hfsc_destroy_class(sch, cl);
1615 }
1616 __skb_queue_purge(&q->requeue);
1617 del_timer(&q->wd_timer);
1618 }
1619
1620 static int
1621 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1622 {
1623 struct hfsc_sched *q = qdisc_priv(sch);
1624 unsigned char *b = skb->tail;
1625 struct tc_hfsc_qopt qopt;
1626
1627 qopt.defcls = q->defcls;
1628 RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
1629 return skb->len;
1630
1631 rtattr_failure:
1632 skb_trim(skb, b - skb->data);
1633 return -1;
1634 }
1635
1636 static int
1637 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1638 {
1639 struct hfsc_class *cl;
1640 unsigned int len;
1641 int err;
1642
1643 cl = hfsc_classify(skb, sch, &err);
1644 if (cl == NULL) {
1645 if (err == NET_XMIT_BYPASS)
1646 sch->qstats.drops++;
1647 kfree_skb(skb);
1648 return err;
1649 }
1650
1651 len = skb->len;
1652 err = cl->qdisc->enqueue(skb, cl->qdisc);
1653 if (unlikely(err != NET_XMIT_SUCCESS)) {
1654 cl->qstats.drops++;
1655 sch->qstats.drops++;
1656 return err;
1657 }
1658
1659 if (cl->qdisc->q.qlen == 1)
1660 set_active(cl, len);
1661
1662 cl->bstats.packets++;
1663 cl->bstats.bytes += len;
1664 sch->bstats.packets++;
1665 sch->bstats.bytes += len;
1666 sch->q.qlen++;
1667
1668 return NET_XMIT_SUCCESS;
1669 }
1670
1671 static struct sk_buff *
1672 hfsc_dequeue(struct Qdisc *sch)
1673 {
1674 struct hfsc_sched *q = qdisc_priv(sch);
1675 struct hfsc_class *cl;
1676 struct sk_buff *skb;
1677 u64 cur_time;
1678 unsigned int next_len;
1679 int realtime = 0;
1680
1681 if (sch->q.qlen == 0)
1682 return NULL;
1683 if ((skb = __skb_dequeue(&q->requeue)))
1684 goto out;
1685
1686 PSCHED_GET_TIME(cur_time);
1687
1688 /*
1689 * if there are eligible classes, use real-time criteria.
1690 * find the class with the minimum deadline among
1691 * the eligible classes.
1692 */
1693 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) {
1694 realtime = 1;
1695 } else {
1696 /*
1697 * use link-sharing criteria
1698 * get the class with the minimum vt in the hierarchy
1699 */
1700 cl = vttree_get_minvt(&q->root, cur_time);
1701 if (cl == NULL) {
1702 sch->qstats.overlimits++;
1703 hfsc_schedule_watchdog(sch, cur_time);
1704 return NULL;
1705 }
1706 }
1707
1708 skb = cl->qdisc->dequeue(cl->qdisc);
1709 if (skb == NULL) {
1710 if (net_ratelimit())
1711 printk("HFSC: Non-work-conserving qdisc ?\n");
1712 return NULL;
1713 }
1714
1715 update_vf(cl, skb->len, cur_time);
1716 if (realtime)
1717 cl->cl_cumul += skb->len;
1718
1719 if (cl->qdisc->q.qlen != 0) {
1720 if (cl->cl_flags & HFSC_RSC) {
1721 /* update ed */
1722 next_len = qdisc_peek_len(cl->qdisc);
1723 if (realtime)
1724 update_ed(cl, next_len);
1725 else
1726 update_d(cl, next_len);
1727 }
1728 } else {
1729 /* the class becomes passive */
1730 set_passive(cl);
1731 }
1732
1733 out:
1734 sch->flags &= ~TCQ_F_THROTTLED;
1735 sch->q.qlen--;
1736
1737 return skb;
1738 }
1739
1740 static int
1741 hfsc_requeue(struct sk_buff *skb, struct Qdisc *sch)
1742 {
1743 struct hfsc_sched *q = qdisc_priv(sch);
1744
1745 __skb_queue_head(&q->requeue, skb);
1746 sch->q.qlen++;
1747 sch->qstats.requeues++;
1748 return NET_XMIT_SUCCESS;
1749 }
1750
1751 static unsigned int
1752 hfsc_drop(struct Qdisc *sch)
1753 {
1754 struct hfsc_sched *q = qdisc_priv(sch);
1755 struct hfsc_class *cl;
1756 unsigned int len;
1757
1758 list_for_each_entry(cl, &q->droplist, dlist) {
1759 if (cl->qdisc->ops->drop != NULL &&
1760 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) {
1761 if (cl->qdisc->q.qlen == 0) {
1762 update_vf(cl, 0, 0);
1763 set_passive(cl);
1764 } else {
1765 list_move_tail(&cl->dlist, &q->droplist);
1766 }
1767 cl->qstats.drops++;
1768 sch->qstats.drops++;
1769 sch->q.qlen--;
1770 return len;
1771 }
1772 }
1773 return 0;
1774 }
1775
1776 static struct Qdisc_class_ops hfsc_class_ops = {
1777 .change = hfsc_change_class,
1778 .delete = hfsc_delete_class,
1779 .graft = hfsc_graft_class,
1780 .leaf = hfsc_class_leaf,
1781 .get = hfsc_get_class,
1782 .put = hfsc_put_class,
1783 .bind_tcf = hfsc_bind_tcf,
1784 .unbind_tcf = hfsc_unbind_tcf,
1785 .tcf_chain = hfsc_tcf_chain,
1786 .dump = hfsc_dump_class,
1787 .dump_stats = hfsc_dump_class_stats,
1788 .walk = hfsc_walk
1789 };
1790
1791 static struct Qdisc_ops hfsc_qdisc_ops = {
1792 .id = "hfsc",
1793 .init = hfsc_init_qdisc,
1794 .change = hfsc_change_qdisc,
1795 .reset = hfsc_reset_qdisc,
1796 .destroy = hfsc_destroy_qdisc,
1797 .dump = hfsc_dump_qdisc,
1798 .enqueue = hfsc_enqueue,
1799 .dequeue = hfsc_dequeue,
1800 .requeue = hfsc_requeue,
1801 .drop = hfsc_drop,
1802 .cl_ops = &hfsc_class_ops,
1803 .priv_size = sizeof(struct hfsc_sched),
1804 .owner = THIS_MODULE
1805 };
1806
1807 static int __init
1808 hfsc_init(void)
1809 {
1810 return register_qdisc(&hfsc_qdisc_ops);
1811 }
1812
1813 static void __exit
1814 hfsc_cleanup(void)
1815 {
1816 unregister_qdisc(&hfsc_qdisc_ops);
1817 }
1818
1819 MODULE_LICENSE("GPL");
1820 module_init(hfsc_init);
1821 module_exit(hfsc_cleanup);
Linux kernel - Deliverables
This page took 0.071177 seconds and 6 git commands to generate.