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/compiler.h>
57 #include <linux/spinlock.h>
58 #include <linux/skbuff.h>
59 #include <linux/string.h>
60 #include <linux/slab.h>
61 #include <linux/list.h>
62 #include <linux/rbtree.h>
63 #include <linux/init.h>
64 #include <linux/rtnetlink.h>
65 #include <linux/pkt_sched.h>
66 #include <net/netlink.h>
67 #include <net/pkt_sched.h>
68 #include <net/pkt_cls.h>
69 #include <asm/div64.h>
70
71 /*
72 * kernel internal service curve representation:
73 * coordinates are given by 64 bit unsigned integers.
74 * x-axis: unit is clock count.
75 * y-axis: unit is byte.
76 *
77 * The service curve parameters are converted to the internal
78 * representation. The slope values are scaled to avoid overflow.
79 * the inverse slope values as well as the y-projection of the 1st
80 * segment are kept in order to avoid 64-bit divide operations
81 * that are expensive on 32-bit architectures.
82 */
83
84 struct internal_sc {
85 u64 sm1; /* scaled slope of the 1st segment */
86 u64 ism1; /* scaled inverse-slope of the 1st segment */
87 u64 dx; /* the x-projection of the 1st segment */
88 u64 dy; /* the y-projection of the 1st segment */
89 u64 sm2; /* scaled slope of the 2nd segment */
90 u64 ism2; /* scaled inverse-slope of the 2nd segment */
91 };
92
93 /* runtime service curve */
94 struct runtime_sc {
95 u64 x; /* current starting position on x-axis */
96 u64 y; /* current starting position on y-axis */
97 u64 sm1; /* scaled slope of the 1st segment */
98 u64 ism1; /* scaled inverse-slope of the 1st segment */
99 u64 dx; /* the x-projection of the 1st segment */
100 u64 dy; /* the y-projection of the 1st segment */
101 u64 sm2; /* scaled slope of the 2nd segment */
102 u64 ism2; /* scaled inverse-slope of the 2nd segment */
103 };
104
105 enum hfsc_class_flags {
106 HFSC_RSC = 0x1,
107 HFSC_FSC = 0x2,
108 HFSC_USC = 0x4
109 };
110
111 struct hfsc_class {
112 struct Qdisc_class_common cl_common;
113
114 struct gnet_stats_basic_sync bstats;
115 struct gnet_stats_queue qstats;
116 struct net_rate_estimator __rcu *rate_est;
117 struct tcf_proto __rcu *filter_list; /* filter list */
118 struct tcf_block *block;
119 unsigned int level; /* class level in hierarchy */
120
121 struct hfsc_sched *sched; /* scheduler data */
122 struct hfsc_class *cl_parent; /* parent class */
123 struct list_head siblings; /* sibling classes */
124 struct list_head children; /* child classes */
125 struct Qdisc *qdisc; /* leaf qdisc */
126
127 struct rb_node el_node; /* qdisc's eligible tree member */
128 struct rb_root vt_tree; /* active children sorted by cl_vt */
129 struct rb_node vt_node; /* parent's vt_tree member */
130 struct rb_root cf_tree; /* active children sorted by cl_f */
131 struct rb_node cf_node; /* parent's cf_heap member */
132
133 u64 cl_total; /* total work in bytes */
134 u64 cl_cumul; /* cumulative work in bytes done by
135 real-time criteria */
136
137 u64 cl_d; /* deadline*/
138 u64 cl_e; /* eligible time */
139 u64 cl_vt; /* virtual time */
140 u64 cl_f; /* time when this class will fit for
141 link-sharing, max(myf, cfmin) */
142 u64 cl_myf; /* my fit-time (calculated from this
143 class's own upperlimit curve) */
144 u64 cl_cfmin; /* earliest children's fit-time (used
145 with cl_myf to obtain cl_f) */
146 u64 cl_cvtmin; /* minimal virtual time among the
147 children fit for link-sharing
148 (monotonic within a period) */
149 u64 cl_vtadj; /* intra-period cumulative vt
150 adjustment */
151 u64 cl_cvtoff; /* largest virtual time seen among
152 the children */
153
154 struct internal_sc cl_rsc; /* internal real-time service curve */
155 struct internal_sc cl_fsc; /* internal fair service curve */
156 struct internal_sc cl_usc; /* internal upperlimit service curve */
157 struct runtime_sc cl_deadline; /* deadline curve */
158 struct runtime_sc cl_eligible; /* eligible curve */
159 struct runtime_sc cl_virtual; /* virtual curve */
160 struct runtime_sc cl_ulimit; /* upperlimit curve */
161
162 u8 cl_flags; /* which curves are valid */
163 u32 cl_vtperiod; /* vt period sequence number */
164 u32 cl_parentperiod;/* parent's vt period sequence number*/
165 u32 cl_nactive; /* number of active children */
166 };
167
168 struct hfsc_sched {
169 u16 defcls; /* default class id */
170 struct hfsc_class root; /* root class */
171 struct Qdisc_class_hash clhash; /* class hash */
172 struct rb_root eligible; /* eligible tree */
173 struct qdisc_watchdog watchdog; /* watchdog timer */
174 };
175
176 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */
177
178
179 /*
180 * eligible tree holds backlogged classes being sorted by their eligible times.
181 * there is one eligible tree per hfsc instance.
182 */
183
184 static void
eltree_insert(struct hfsc_class * cl)185 eltree_insert(struct hfsc_class *cl)
186 {
187 struct rb_node **p = &cl->sched->eligible.rb_node;
188 struct rb_node *parent = NULL;
189 struct hfsc_class *cl1;
190
191 while (*p != NULL) {
192 parent = *p;
193 cl1 = rb_entry(parent, struct hfsc_class, el_node);
194 if (cl->cl_e >= cl1->cl_e)
195 p = &parent->rb_right;
196 else
197 p = &parent->rb_left;
198 }
199 rb_link_node(&cl->el_node, parent, p);
200 rb_insert_color(&cl->el_node, &cl->sched->eligible);
201 }
202
203 static inline void
eltree_remove(struct hfsc_class * cl)204 eltree_remove(struct hfsc_class *cl)
205 {
206 rb_erase(&cl->el_node, &cl->sched->eligible);
207 }
208
209 static inline void
eltree_update(struct hfsc_class * cl)210 eltree_update(struct hfsc_class *cl)
211 {
212 eltree_remove(cl);
213 eltree_insert(cl);
214 }
215
216 /* find the class with the minimum deadline among the eligible classes */
217 static inline struct hfsc_class *
eltree_get_mindl(struct hfsc_sched * q,u64 cur_time)218 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time)
219 {
220 struct hfsc_class *p, *cl = NULL;
221 struct rb_node *n;
222
223 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) {
224 p = rb_entry(n, struct hfsc_class, el_node);
225 if (p->cl_e > cur_time)
226 break;
227 if (cl == NULL || p->cl_d < cl->cl_d)
228 cl = p;
229 }
230 return cl;
231 }
232
233 /* find the class with minimum eligible time among the eligible classes */
234 static inline struct hfsc_class *
eltree_get_minel(struct hfsc_sched * q)235 eltree_get_minel(struct hfsc_sched *q)
236 {
237 struct rb_node *n;
238
239 n = rb_first(&q->eligible);
240 if (n == NULL)
241 return NULL;
242 return rb_entry(n, struct hfsc_class, el_node);
243 }
244
245 /*
246 * vttree holds holds backlogged child classes being sorted by their virtual
247 * time. each intermediate class has one vttree.
248 */
249 static void
vttree_insert(struct hfsc_class * cl)250 vttree_insert(struct hfsc_class *cl)
251 {
252 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node;
253 struct rb_node *parent = NULL;
254 struct hfsc_class *cl1;
255
256 while (*p != NULL) {
257 parent = *p;
258 cl1 = rb_entry(parent, struct hfsc_class, vt_node);
259 if (cl->cl_vt >= cl1->cl_vt)
260 p = &parent->rb_right;
261 else
262 p = &parent->rb_left;
263 }
264 rb_link_node(&cl->vt_node, parent, p);
265 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree);
266 }
267
268 static inline void
vttree_remove(struct hfsc_class * cl)269 vttree_remove(struct hfsc_class *cl)
270 {
271 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree);
272 }
273
274 static inline void
vttree_update(struct hfsc_class * cl)275 vttree_update(struct hfsc_class *cl)
276 {
277 vttree_remove(cl);
278 vttree_insert(cl);
279 }
280
281 static inline struct hfsc_class *
vttree_firstfit(struct hfsc_class * cl,u64 cur_time)282 vttree_firstfit(struct hfsc_class *cl, u64 cur_time)
283 {
284 struct hfsc_class *p;
285 struct rb_node *n;
286
287 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) {
288 p = rb_entry(n, struct hfsc_class, vt_node);
289 if (p->cl_f <= cur_time)
290 return p;
291 }
292 return NULL;
293 }
294
295 /*
296 * get the leaf class with the minimum vt in the hierarchy
297 */
298 static struct hfsc_class *
vttree_get_minvt(struct hfsc_class * cl,u64 cur_time)299 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time)
300 {
301 /* if root-class's cfmin is bigger than cur_time nothing to do */
302 if (cl->cl_cfmin > cur_time)
303 return NULL;
304
305 while (cl->level > 0) {
306 cl = vttree_firstfit(cl, cur_time);
307 if (cl == NULL)
308 return NULL;
309 /*
310 * update parent's cl_cvtmin.
311 */
312 if (cl->cl_parent->cl_cvtmin < cl->cl_vt)
313 cl->cl_parent->cl_cvtmin = cl->cl_vt;
314 }
315 return cl;
316 }
317
318 static void
cftree_insert(struct hfsc_class * cl)319 cftree_insert(struct hfsc_class *cl)
320 {
321 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node;
322 struct rb_node *parent = NULL;
323 struct hfsc_class *cl1;
324
325 while (*p != NULL) {
326 parent = *p;
327 cl1 = rb_entry(parent, struct hfsc_class, cf_node);
328 if (cl->cl_f >= cl1->cl_f)
329 p = &parent->rb_right;
330 else
331 p = &parent->rb_left;
332 }
333 rb_link_node(&cl->cf_node, parent, p);
334 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree);
335 }
336
337 static inline void
cftree_remove(struct hfsc_class * cl)338 cftree_remove(struct hfsc_class *cl)
339 {
340 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree);
341 }
342
343 static inline void
cftree_update(struct hfsc_class * cl)344 cftree_update(struct hfsc_class *cl)
345 {
346 cftree_remove(cl);
347 cftree_insert(cl);
348 }
349
350 /*
351 * service curve support functions
352 *
353 * external service curve parameters
354 * m: bps
355 * d: us
356 * internal service curve parameters
357 * sm: (bytes/psched_us) << SM_SHIFT
358 * ism: (psched_us/byte) << ISM_SHIFT
359 * dx: psched_us
360 *
361 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us.
362 *
363 * sm and ism are scaled in order to keep effective digits.
364 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective
365 * digits in decimal using the following table.
366 *
367 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps
368 * ------------+-------------------------------------------------------
369 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3
370 *
371 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125
372 *
373 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18.
374 */
375 #define SM_SHIFT (30 - PSCHED_SHIFT)
376 #define ISM_SHIFT (8 + PSCHED_SHIFT)
377
378 #define SM_MASK ((1ULL << SM_SHIFT) - 1)
379 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1)
380
381 static inline u64
seg_x2y(u64 x,u64 sm)382 seg_x2y(u64 x, u64 sm)
383 {
384 u64 y;
385
386 /*
387 * compute
388 * y = x * sm >> SM_SHIFT
389 * but divide it for the upper and lower bits to avoid overflow
390 */
391 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT);
392 return y;
393 }
394
395 static inline u64
seg_y2x(u64 y,u64 ism)396 seg_y2x(u64 y, u64 ism)
397 {
398 u64 x;
399
400 if (y == 0)
401 x = 0;
402 else if (ism == HT_INFINITY)
403 x = HT_INFINITY;
404 else {
405 x = (y >> ISM_SHIFT) * ism
406 + (((y & ISM_MASK) * ism) >> ISM_SHIFT);
407 }
408 return x;
409 }
410
411 /* Convert m (bps) into sm (bytes/psched us) */
412 static u64
m2sm(u32 m)413 m2sm(u32 m)
414 {
415 u64 sm;
416
417 sm = ((u64)m << SM_SHIFT);
418 sm += PSCHED_TICKS_PER_SEC - 1;
419 do_div(sm, PSCHED_TICKS_PER_SEC);
420 return sm;
421 }
422
423 /* convert m (bps) into ism (psched us/byte) */
424 static u64
m2ism(u32 m)425 m2ism(u32 m)
426 {
427 u64 ism;
428
429 if (m == 0)
430 ism = HT_INFINITY;
431 else {
432 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT);
433 ism += m - 1;
434 do_div(ism, m);
435 }
436 return ism;
437 }
438
439 /* convert d (us) into dx (psched us) */
440 static u64
d2dx(u32 d)441 d2dx(u32 d)
442 {
443 u64 dx;
444
445 dx = ((u64)d * PSCHED_TICKS_PER_SEC);
446 dx += USEC_PER_SEC - 1;
447 do_div(dx, USEC_PER_SEC);
448 return dx;
449 }
450
451 /* convert sm (bytes/psched us) into m (bps) */
452 static u32
sm2m(u64 sm)453 sm2m(u64 sm)
454 {
455 u64 m;
456
457 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT;
458 return (u32)m;
459 }
460
461 /* convert dx (psched us) into d (us) */
462 static u32
dx2d(u64 dx)463 dx2d(u64 dx)
464 {
465 u64 d;
466
467 d = dx * USEC_PER_SEC;
468 do_div(d, PSCHED_TICKS_PER_SEC);
469 return (u32)d;
470 }
471
472 static void
sc2isc(struct tc_service_curve * sc,struct internal_sc * isc)473 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc)
474 {
475 isc->sm1 = m2sm(sc->m1);
476 isc->ism1 = m2ism(sc->m1);
477 isc->dx = d2dx(sc->d);
478 isc->dy = seg_x2y(isc->dx, isc->sm1);
479 isc->sm2 = m2sm(sc->m2);
480 isc->ism2 = m2ism(sc->m2);
481 }
482
483 /*
484 * initialize the runtime service curve with the given internal
485 * service curve starting at (x, y).
486 */
487 static void
rtsc_init(struct runtime_sc * rtsc,struct internal_sc * isc,u64 x,u64 y)488 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
489 {
490 rtsc->x = x;
491 rtsc->y = y;
492 rtsc->sm1 = isc->sm1;
493 rtsc->ism1 = isc->ism1;
494 rtsc->dx = isc->dx;
495 rtsc->dy = isc->dy;
496 rtsc->sm2 = isc->sm2;
497 rtsc->ism2 = isc->ism2;
498 }
499
500 /*
501 * calculate the y-projection of the runtime service curve by the
502 * given x-projection value
503 */
504 static u64
rtsc_y2x(struct runtime_sc * rtsc,u64 y)505 rtsc_y2x(struct runtime_sc *rtsc, u64 y)
506 {
507 u64 x;
508
509 if (y < rtsc->y)
510 x = rtsc->x;
511 else if (y <= rtsc->y + rtsc->dy) {
512 /* x belongs to the 1st segment */
513 if (rtsc->dy == 0)
514 x = rtsc->x + rtsc->dx;
515 else
516 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1);
517 } else {
518 /* x belongs to the 2nd segment */
519 x = rtsc->x + rtsc->dx
520 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2);
521 }
522 return x;
523 }
524
525 static u64
rtsc_x2y(struct runtime_sc * rtsc,u64 x)526 rtsc_x2y(struct runtime_sc *rtsc, u64 x)
527 {
528 u64 y;
529
530 if (x <= rtsc->x)
531 y = rtsc->y;
532 else if (x <= rtsc->x + rtsc->dx)
533 /* y belongs to the 1st segment */
534 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1);
535 else
536 /* y belongs to the 2nd segment */
537 y = rtsc->y + rtsc->dy
538 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2);
539 return y;
540 }
541
542 /*
543 * update the runtime service curve by taking the minimum of the current
544 * runtime service curve and the service curve starting at (x, y).
545 */
546 static void
rtsc_min(struct runtime_sc * rtsc,struct internal_sc * isc,u64 x,u64 y)547 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y)
548 {
549 u64 y1, y2, dx, dy;
550 u32 dsm;
551
552 if (isc->sm1 <= isc->sm2) {
553 /* service curve is convex */
554 y1 = rtsc_x2y(rtsc, x);
555 if (y1 < y)
556 /* the current rtsc is smaller */
557 return;
558 rtsc->x = x;
559 rtsc->y = y;
560 return;
561 }
562
563 /*
564 * service curve is concave
565 * compute the two y values of the current rtsc
566 * y1: at x
567 * y2: at (x + dx)
568 */
569 y1 = rtsc_x2y(rtsc, x);
570 if (y1 <= y) {
571 /* rtsc is below isc, no change to rtsc */
572 return;
573 }
574
575 y2 = rtsc_x2y(rtsc, x + isc->dx);
576 if (y2 >= y + isc->dy) {
577 /* rtsc is above isc, replace rtsc by isc */
578 rtsc->x = x;
579 rtsc->y = y;
580 rtsc->dx = isc->dx;
581 rtsc->dy = isc->dy;
582 return;
583 }
584
585 /*
586 * the two curves intersect
587 * compute the offsets (dx, dy) using the reverse
588 * function of seg_x2y()
589 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y)
590 */
591 dx = (y1 - y) << SM_SHIFT;
592 dsm = isc->sm1 - isc->sm2;
593 do_div(dx, dsm);
594 /*
595 * check if (x, y1) belongs to the 1st segment of rtsc.
596 * if so, add the offset.
597 */
598 if (rtsc->x + rtsc->dx > x)
599 dx += rtsc->x + rtsc->dx - x;
600 dy = seg_x2y(dx, isc->sm1);
601
602 rtsc->x = x;
603 rtsc->y = y;
604 rtsc->dx = dx;
605 rtsc->dy = dy;
606 }
607
608 static void
init_ed(struct hfsc_class * cl,unsigned int next_len)609 init_ed(struct hfsc_class *cl, unsigned int next_len)
610 {
611 u64 cur_time = psched_get_time();
612
613 /* update the deadline curve */
614 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
615
616 /*
617 * update the eligible curve.
618 * for concave, it is equal to the deadline curve.
619 * for convex, it is a linear curve with slope m2.
620 */
621 cl->cl_eligible = cl->cl_deadline;
622 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
623 cl->cl_eligible.dx = 0;
624 cl->cl_eligible.dy = 0;
625 }
626
627 /* compute e and d */
628 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
629 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
630
631 eltree_insert(cl);
632 }
633
634 static void
update_ed(struct hfsc_class * cl,unsigned int next_len)635 update_ed(struct hfsc_class *cl, unsigned int next_len)
636 {
637 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul);
638 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
639
640 eltree_update(cl);
641 }
642
643 static inline void
update_d(struct hfsc_class * cl,unsigned int next_len)644 update_d(struct hfsc_class *cl, unsigned int next_len)
645 {
646 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len);
647 }
648
649 static inline void
update_cfmin(struct hfsc_class * cl)650 update_cfmin(struct hfsc_class *cl)
651 {
652 struct rb_node *n = rb_first(&cl->cf_tree);
653 struct hfsc_class *p;
654
655 if (n == NULL) {
656 cl->cl_cfmin = 0;
657 return;
658 }
659 p = rb_entry(n, struct hfsc_class, cf_node);
660 cl->cl_cfmin = p->cl_f;
661 }
662
663 static void
init_vf(struct hfsc_class * cl,unsigned int len)664 init_vf(struct hfsc_class *cl, unsigned int len)
665 {
666 struct hfsc_class *max_cl;
667 struct rb_node *n;
668 u64 vt, f, cur_time;
669 int go_active;
670
671 cur_time = 0;
672 go_active = 1;
673 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
674 if (go_active && cl->cl_nactive++ == 0)
675 go_active = 1;
676 else
677 go_active = 0;
678
679 if (go_active) {
680 n = rb_last(&cl->cl_parent->vt_tree);
681 if (n != NULL) {
682 max_cl = rb_entry(n, struct hfsc_class, vt_node);
683 /*
684 * set vt to the average of the min and max
685 * classes. if the parent's period didn't
686 * change, don't decrease vt of the class.
687 */
688 vt = max_cl->cl_vt;
689 if (cl->cl_parent->cl_cvtmin != 0)
690 vt = (cl->cl_parent->cl_cvtmin + vt)/2;
691
692 if (cl->cl_parent->cl_vtperiod !=
693 cl->cl_parentperiod || vt > cl->cl_vt)
694 cl->cl_vt = vt;
695 } else {
696 /*
697 * first child for a new parent backlog period.
698 * initialize cl_vt to the highest value seen
699 * among the siblings. this is analogous to
700 * what cur_time would provide in realtime case.
701 */
702 cl->cl_vt = cl->cl_parent->cl_cvtoff;
703 cl->cl_parent->cl_cvtmin = 0;
704 }
705
706 /* update the virtual curve */
707 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
708 cl->cl_vtadj = 0;
709
710 cl->cl_vtperiod++; /* increment vt period */
711 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod;
712 if (cl->cl_parent->cl_nactive == 0)
713 cl->cl_parentperiod++;
714 cl->cl_f = 0;
715
716 vttree_insert(cl);
717 cftree_insert(cl);
718
719 if (cl->cl_flags & HFSC_USC) {
720 /* class has upper limit curve */
721 if (cur_time == 0)
722 cur_time = psched_get_time();
723
724 /* update the ulimit curve */
725 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time,
726 cl->cl_total);
727 /* compute myf */
728 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit,
729 cl->cl_total);
730 }
731 }
732
733 f = max(cl->cl_myf, cl->cl_cfmin);
734 if (f != cl->cl_f) {
735 cl->cl_f = f;
736 cftree_update(cl);
737 }
738 update_cfmin(cl->cl_parent);
739 }
740 }
741
742 static void
update_vf(struct hfsc_class * cl,unsigned int len,u64 cur_time)743 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time)
744 {
745 u64 f; /* , myf_bound, delta; */
746 int go_passive = 0;
747
748 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
749 go_passive = 1;
750
751 for (; cl->cl_parent != NULL; cl = cl->cl_parent) {
752 cl->cl_total += len;
753
754 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0)
755 continue;
756
757 if (go_passive && --cl->cl_nactive == 0)
758 go_passive = 1;
759 else
760 go_passive = 0;
761
762 /* update vt */
763 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) + cl->cl_vtadj;
764
765 /*
766 * if vt of the class is smaller than cvtmin,
767 * the class was skipped in the past due to non-fit.
768 * if so, we need to adjust vtadj.
769 */
770 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) {
771 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt;
772 cl->cl_vt = cl->cl_parent->cl_cvtmin;
773 }
774
775 if (go_passive) {
776 /* no more active child, going passive */
777
778 /* update cvtoff of the parent class */
779 if (cl->cl_vt > cl->cl_parent->cl_cvtoff)
780 cl->cl_parent->cl_cvtoff = cl->cl_vt;
781
782 /* remove this class from the vt tree */
783 vttree_remove(cl);
784
785 cftree_remove(cl);
786 update_cfmin(cl->cl_parent);
787
788 continue;
789 }
790
791 /* update the vt tree */
792 vttree_update(cl);
793
794 /* update f */
795 if (cl->cl_flags & HFSC_USC) {
796 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, cl->cl_total);
797 #if 0
798 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit,
799 cl->cl_total);
800 /*
801 * This code causes classes to stay way under their
802 * limit when multiple classes are used at gigabit
803 * speed. needs investigation. -kaber
804 */
805 /*
806 * if myf lags behind by more than one clock tick
807 * from the current time, adjust myfadj to prevent
808 * a rate-limited class from going greedy.
809 * in a steady state under rate-limiting, myf
810 * fluctuates within one clock tick.
811 */
812 myf_bound = cur_time - PSCHED_JIFFIE2US(1);
813 if (cl->cl_myf < myf_bound) {
814 delta = cur_time - cl->cl_myf;
815 cl->cl_myfadj += delta;
816 cl->cl_myf += delta;
817 }
818 #endif
819 }
820
821 f = max(cl->cl_myf, cl->cl_cfmin);
822 if (f != cl->cl_f) {
823 cl->cl_f = f;
824 cftree_update(cl);
825 update_cfmin(cl->cl_parent);
826 }
827 }
828 }
829
830 static unsigned int
qdisc_peek_len(struct Qdisc * sch)831 qdisc_peek_len(struct Qdisc *sch)
832 {
833 struct sk_buff *skb;
834 unsigned int len;
835
836 skb = sch->ops->peek(sch);
837 if (unlikely(skb == NULL)) {
838 qdisc_warn_nonwc("qdisc_peek_len", sch);
839 return 0;
840 }
841 len = qdisc_pkt_len(skb);
842
843 return len;
844 }
845
846 static void
hfsc_adjust_levels(struct hfsc_class * cl)847 hfsc_adjust_levels(struct hfsc_class *cl)
848 {
849 struct hfsc_class *p;
850 unsigned int level;
851
852 do {
853 level = 0;
854 list_for_each_entry(p, &cl->children, siblings) {
855 if (p->level >= level)
856 level = p->level + 1;
857 }
858 cl->level = level;
859 } while ((cl = cl->cl_parent) != NULL);
860 }
861
862 static inline struct hfsc_class *
hfsc_find_class(u32 classid,struct Qdisc * sch)863 hfsc_find_class(u32 classid, struct Qdisc *sch)
864 {
865 struct hfsc_sched *q = qdisc_priv(sch);
866 struct Qdisc_class_common *clc;
867
868 clc = qdisc_class_find(&q->clhash, classid);
869 if (clc == NULL)
870 return NULL;
871 return container_of(clc, struct hfsc_class, cl_common);
872 }
873
874 static void
hfsc_change_rsc(struct hfsc_class * cl,struct tc_service_curve * rsc,u64 cur_time)875 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc,
876 u64 cur_time)
877 {
878 sc2isc(rsc, &cl->cl_rsc);
879 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul);
880 cl->cl_eligible = cl->cl_deadline;
881 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) {
882 cl->cl_eligible.dx = 0;
883 cl->cl_eligible.dy = 0;
884 }
885 cl->cl_flags |= HFSC_RSC;
886 }
887
888 static void
hfsc_change_fsc(struct hfsc_class * cl,struct tc_service_curve * fsc)889 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc)
890 {
891 sc2isc(fsc, &cl->cl_fsc);
892 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
893 cl->cl_flags |= HFSC_FSC;
894 }
895
896 static void
hfsc_change_usc(struct hfsc_class * cl,struct tc_service_curve * usc,u64 cur_time)897 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc,
898 u64 cur_time)
899 {
900 sc2isc(usc, &cl->cl_usc);
901 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total);
902 cl->cl_flags |= HFSC_USC;
903 }
904
905 static void
hfsc_upgrade_rt(struct hfsc_class * cl)906 hfsc_upgrade_rt(struct hfsc_class *cl)
907 {
908 cl->cl_fsc = cl->cl_rsc;
909 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total);
910 cl->cl_flags |= HFSC_FSC;
911 }
912
913 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = {
914 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) },
915 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) },
916 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) },
917 };
918
919 static int
hfsc_change_class(struct Qdisc * sch,u32 classid,u32 parentid,struct nlattr ** tca,unsigned long * arg,struct netlink_ext_ack * extack)920 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
921 struct nlattr **tca, unsigned long *arg,
922 struct netlink_ext_ack *extack)
923 {
924 struct hfsc_sched *q = qdisc_priv(sch);
925 struct hfsc_class *cl = (struct hfsc_class *)*arg;
926 struct hfsc_class *parent = NULL;
927 struct nlattr *opt = tca[TCA_OPTIONS];
928 struct nlattr *tb[TCA_HFSC_MAX + 1];
929 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL;
930 u64 cur_time;
931 int err;
932
933 if (opt == NULL)
934 return -EINVAL;
935
936 err = nla_parse_nested_deprecated(tb, TCA_HFSC_MAX, opt, hfsc_policy,
937 NULL);
938 if (err < 0)
939 return err;
940
941 if (tb[TCA_HFSC_RSC]) {
942 rsc = nla_data(tb[TCA_HFSC_RSC]);
943 if (rsc->m1 == 0 && rsc->m2 == 0)
944 rsc = NULL;
945 }
946
947 if (tb[TCA_HFSC_FSC]) {
948 fsc = nla_data(tb[TCA_HFSC_FSC]);
949 if (fsc->m1 == 0 && fsc->m2 == 0)
950 fsc = NULL;
951 }
952
953 if (tb[TCA_HFSC_USC]) {
954 usc = nla_data(tb[TCA_HFSC_USC]);
955 if (usc->m1 == 0 && usc->m2 == 0)
956 usc = NULL;
957 }
958
959 if (cl != NULL) {
960 int old_flags;
961
962 if (parentid) {
963 if (cl->cl_parent &&
964 cl->cl_parent->cl_common.classid != parentid)
965 return -EINVAL;
966 if (cl->cl_parent == NULL && parentid != TC_H_ROOT)
967 return -EINVAL;
968 }
969 cur_time = psched_get_time();
970
971 if (tca[TCA_RATE]) {
972 err = gen_replace_estimator(&cl->bstats, NULL,
973 &cl->rate_est,
974 NULL,
975 true,
976 tca[TCA_RATE]);
977 if (err)
978 return err;
979 }
980
981 sch_tree_lock(sch);
982 old_flags = cl->cl_flags;
983
984 if (rsc != NULL)
985 hfsc_change_rsc(cl, rsc, cur_time);
986 if (fsc != NULL)
987 hfsc_change_fsc(cl, fsc);
988 if (usc != NULL)
989 hfsc_change_usc(cl, usc, cur_time);
990
991 if (cl->qdisc->q.qlen != 0) {
992 int len = qdisc_peek_len(cl->qdisc);
993
994 if (cl->cl_flags & HFSC_RSC) {
995 if (old_flags & HFSC_RSC)
996 update_ed(cl, len);
997 else
998 init_ed(cl, len);
999 }
1000
1001 if (cl->cl_flags & HFSC_FSC) {
1002 if (old_flags & HFSC_FSC)
1003 update_vf(cl, 0, cur_time);
1004 else
1005 init_vf(cl, len);
1006 }
1007 }
1008 sch_tree_unlock(sch);
1009
1010 return 0;
1011 }
1012
1013 if (parentid == TC_H_ROOT)
1014 return -EEXIST;
1015
1016 parent = &q->root;
1017 if (parentid) {
1018 parent = hfsc_find_class(parentid, sch);
1019 if (parent == NULL)
1020 return -ENOENT;
1021 }
1022
1023 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0)
1024 return -EINVAL;
1025 if (hfsc_find_class(classid, sch))
1026 return -EEXIST;
1027
1028 if (rsc == NULL && fsc == NULL)
1029 return -EINVAL;
1030
1031 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL);
1032 if (cl == NULL)
1033 return -ENOBUFS;
1034
1035 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1036 if (err) {
1037 kfree(cl);
1038 return err;
1039 }
1040
1041 if (tca[TCA_RATE]) {
1042 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1043 NULL, true, tca[TCA_RATE]);
1044 if (err) {
1045 tcf_block_put(cl->block);
1046 kfree(cl);
1047 return err;
1048 }
1049 }
1050
1051 if (rsc != NULL)
1052 hfsc_change_rsc(cl, rsc, 0);
1053 if (fsc != NULL)
1054 hfsc_change_fsc(cl, fsc);
1055 if (usc != NULL)
1056 hfsc_change_usc(cl, usc, 0);
1057
1058 cl->cl_common.classid = classid;
1059 cl->sched = q;
1060 cl->cl_parent = parent;
1061 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1062 classid, NULL);
1063 if (cl->qdisc == NULL)
1064 cl->qdisc = &noop_qdisc;
1065 else
1066 qdisc_hash_add(cl->qdisc, true);
1067 INIT_LIST_HEAD(&cl->children);
1068 cl->vt_tree = RB_ROOT;
1069 cl->cf_tree = RB_ROOT;
1070
1071 sch_tree_lock(sch);
1072 /* Check if the inner class is a misconfigured 'rt' */
1073 if (!(parent->cl_flags & HFSC_FSC) && parent != &q->root) {
1074 NL_SET_ERR_MSG(extack,
1075 "Forced curve change on parent 'rt' to 'sc'");
1076 hfsc_upgrade_rt(parent);
1077 }
1078 qdisc_class_hash_insert(&q->clhash, &cl->cl_common);
1079 list_add_tail(&cl->siblings, &parent->children);
1080 if (parent->level == 0)
1081 qdisc_purge_queue(parent->qdisc);
1082 hfsc_adjust_levels(parent);
1083 sch_tree_unlock(sch);
1084
1085 qdisc_class_hash_grow(sch, &q->clhash);
1086
1087 *arg = (unsigned long)cl;
1088 return 0;
1089 }
1090
1091 static void
hfsc_destroy_class(struct Qdisc * sch,struct hfsc_class * cl)1092 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl)
1093 {
1094 struct hfsc_sched *q = qdisc_priv(sch);
1095
1096 tcf_block_put(cl->block);
1097 qdisc_put(cl->qdisc);
1098 gen_kill_estimator(&cl->rate_est);
1099 if (cl != &q->root)
1100 kfree(cl);
1101 }
1102
1103 static int
hfsc_delete_class(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)1104 hfsc_delete_class(struct Qdisc *sch, unsigned long arg,
1105 struct netlink_ext_ack *extack)
1106 {
1107 struct hfsc_sched *q = qdisc_priv(sch);
1108 struct hfsc_class *cl = (struct hfsc_class *)arg;
1109
1110 if (cl->level > 0 || qdisc_class_in_use(&cl->cl_common) ||
1111 cl == &q->root) {
1112 NL_SET_ERR_MSG(extack, "HFSC class in use");
1113 return -EBUSY;
1114 }
1115
1116 sch_tree_lock(sch);
1117
1118 list_del(&cl->siblings);
1119 hfsc_adjust_levels(cl->cl_parent);
1120
1121 qdisc_purge_queue(cl->qdisc);
1122 qdisc_class_hash_remove(&q->clhash, &cl->cl_common);
1123
1124 sch_tree_unlock(sch);
1125
1126 hfsc_destroy_class(sch, cl);
1127 return 0;
1128 }
1129
1130 static struct hfsc_class *
hfsc_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)1131 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
1132 {
1133 struct hfsc_sched *q = qdisc_priv(sch);
1134 struct hfsc_class *head, *cl;
1135 struct tcf_result res;
1136 struct tcf_proto *tcf;
1137 int result;
1138
1139 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 &&
1140 (cl = hfsc_find_class(skb->priority, sch)) != NULL)
1141 if (cl->level == 0)
1142 return cl;
1143
1144 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
1145 head = &q->root;
1146 tcf = rcu_dereference_bh(q->root.filter_list);
1147 while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
1148 #ifdef CONFIG_NET_CLS_ACT
1149 switch (result) {
1150 case TC_ACT_QUEUED:
1151 case TC_ACT_STOLEN:
1152 case TC_ACT_TRAP:
1153 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
1154 fallthrough;
1155 case TC_ACT_SHOT:
1156 return NULL;
1157 }
1158 #endif
1159 cl = (struct hfsc_class *)res.class;
1160 if (!cl) {
1161 cl = hfsc_find_class(res.classid, sch);
1162 if (!cl)
1163 break; /* filter selected invalid classid */
1164 if (cl->level >= head->level)
1165 break; /* filter may only point downwards */
1166 }
1167
1168 if (cl->level == 0)
1169 return cl; /* hit leaf class */
1170
1171 /* apply inner filter chain */
1172 tcf = rcu_dereference_bh(cl->filter_list);
1173 head = cl;
1174 }
1175
1176 /* classification failed, try default class */
1177 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle),
1178 READ_ONCE(q->defcls)), sch);
1179 if (cl == NULL || cl->level > 0)
1180 return NULL;
1181
1182 return cl;
1183 }
1184
1185 static int
hfsc_graft_class(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)1186 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1187 struct Qdisc **old, struct netlink_ext_ack *extack)
1188 {
1189 struct hfsc_class *cl = (struct hfsc_class *)arg;
1190
1191 if (cl->level > 0)
1192 return -EINVAL;
1193 if (new == NULL) {
1194 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1195 cl->cl_common.classid, NULL);
1196 if (new == NULL)
1197 new = &noop_qdisc;
1198 }
1199
1200 *old = qdisc_replace(sch, new, &cl->qdisc);
1201 return 0;
1202 }
1203
1204 static struct Qdisc *
hfsc_class_leaf(struct Qdisc * sch,unsigned long arg)1205 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg)
1206 {
1207 struct hfsc_class *cl = (struct hfsc_class *)arg;
1208
1209 if (cl->level == 0)
1210 return cl->qdisc;
1211
1212 return NULL;
1213 }
1214
1215 static void
hfsc_qlen_notify(struct Qdisc * sch,unsigned long arg)1216 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg)
1217 {
1218 struct hfsc_class *cl = (struct hfsc_class *)arg;
1219
1220 /* vttree is now handled in update_vf() so that update_vf(cl, 0, 0)
1221 * needs to be called explicitly to remove a class from vttree.
1222 */
1223 update_vf(cl, 0, 0);
1224 if (cl->cl_flags & HFSC_RSC)
1225 eltree_remove(cl);
1226 }
1227
1228 static unsigned long
hfsc_search_class(struct Qdisc * sch,u32 classid)1229 hfsc_search_class(struct Qdisc *sch, u32 classid)
1230 {
1231 return (unsigned long)hfsc_find_class(classid, sch);
1232 }
1233
1234 static unsigned long
hfsc_bind_tcf(struct Qdisc * sch,unsigned long parent,u32 classid)1235 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid)
1236 {
1237 struct hfsc_class *p = (struct hfsc_class *)parent;
1238 struct hfsc_class *cl = hfsc_find_class(classid, sch);
1239
1240 if (cl != NULL) {
1241 if (p != NULL && p->level <= cl->level)
1242 return 0;
1243 qdisc_class_get(&cl->cl_common);
1244 }
1245
1246 return (unsigned long)cl;
1247 }
1248
1249 static void
hfsc_unbind_tcf(struct Qdisc * sch,unsigned long arg)1250 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg)
1251 {
1252 struct hfsc_class *cl = (struct hfsc_class *)arg;
1253
1254 qdisc_class_put(&cl->cl_common);
1255 }
1256
hfsc_tcf_block(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)1257 static struct tcf_block *hfsc_tcf_block(struct Qdisc *sch, unsigned long arg,
1258 struct netlink_ext_ack *extack)
1259 {
1260 struct hfsc_sched *q = qdisc_priv(sch);
1261 struct hfsc_class *cl = (struct hfsc_class *)arg;
1262
1263 if (cl == NULL)
1264 cl = &q->root;
1265
1266 return cl->block;
1267 }
1268
1269 static int
hfsc_dump_sc(struct sk_buff * skb,int attr,struct internal_sc * sc)1270 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc)
1271 {
1272 struct tc_service_curve tsc;
1273
1274 tsc.m1 = sm2m(sc->sm1);
1275 tsc.d = dx2d(sc->dx);
1276 tsc.m2 = sm2m(sc->sm2);
1277 if (nla_put(skb, attr, sizeof(tsc), &tsc))
1278 goto nla_put_failure;
1279
1280 return skb->len;
1281
1282 nla_put_failure:
1283 return -1;
1284 }
1285
1286 static int
hfsc_dump_curves(struct sk_buff * skb,struct hfsc_class * cl)1287 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl)
1288 {
1289 if ((cl->cl_flags & HFSC_RSC) &&
1290 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0))
1291 goto nla_put_failure;
1292
1293 if ((cl->cl_flags & HFSC_FSC) &&
1294 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0))
1295 goto nla_put_failure;
1296
1297 if ((cl->cl_flags & HFSC_USC) &&
1298 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0))
1299 goto nla_put_failure;
1300
1301 return skb->len;
1302
1303 nla_put_failure:
1304 return -1;
1305 }
1306
1307 static int
hfsc_dump_class(struct Qdisc * sch,unsigned long arg,struct sk_buff * skb,struct tcmsg * tcm)1308 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb,
1309 struct tcmsg *tcm)
1310 {
1311 struct hfsc_class *cl = (struct hfsc_class *)arg;
1312 struct nlattr *nest;
1313
1314 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid :
1315 TC_H_ROOT;
1316 tcm->tcm_handle = cl->cl_common.classid;
1317 if (cl->level == 0)
1318 tcm->tcm_info = cl->qdisc->handle;
1319
1320 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1321 if (nest == NULL)
1322 goto nla_put_failure;
1323 if (hfsc_dump_curves(skb, cl) < 0)
1324 goto nla_put_failure;
1325 return nla_nest_end(skb, nest);
1326
1327 nla_put_failure:
1328 nla_nest_cancel(skb, nest);
1329 return -EMSGSIZE;
1330 }
1331
1332 static int
hfsc_dump_class_stats(struct Qdisc * sch,unsigned long arg,struct gnet_dump * d)1333 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1334 struct gnet_dump *d)
1335 {
1336 struct hfsc_class *cl = (struct hfsc_class *)arg;
1337 struct tc_hfsc_stats xstats;
1338 __u32 qlen;
1339
1340 qdisc_qstats_qlen_backlog(cl->qdisc, &qlen, &cl->qstats.backlog);
1341 xstats.level = cl->level;
1342 xstats.period = cl->cl_vtperiod;
1343 xstats.work = cl->cl_total;
1344 xstats.rtwork = cl->cl_cumul;
1345
1346 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1347 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1348 gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1349 return -1;
1350
1351 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
1352 }
1353
1354
1355
1356 static void
hfsc_walk(struct Qdisc * sch,struct qdisc_walker * arg)1357 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1358 {
1359 struct hfsc_sched *q = qdisc_priv(sch);
1360 struct hfsc_class *cl;
1361 unsigned int i;
1362
1363 if (arg->stop)
1364 return;
1365
1366 for (i = 0; i < q->clhash.hashsize; i++) {
1367 hlist_for_each_entry(cl, &q->clhash.hash[i],
1368 cl_common.hnode) {
1369 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
1370 return;
1371 }
1372 }
1373 }
1374
1375 static void
hfsc_schedule_watchdog(struct Qdisc * sch)1376 hfsc_schedule_watchdog(struct Qdisc *sch)
1377 {
1378 struct hfsc_sched *q = qdisc_priv(sch);
1379 struct hfsc_class *cl;
1380 u64 next_time = 0;
1381
1382 cl = eltree_get_minel(q);
1383 if (cl)
1384 next_time = cl->cl_e;
1385 if (q->root.cl_cfmin != 0) {
1386 if (next_time == 0 || next_time > q->root.cl_cfmin)
1387 next_time = q->root.cl_cfmin;
1388 }
1389 if (next_time)
1390 qdisc_watchdog_schedule(&q->watchdog, next_time);
1391 }
1392
1393 static int
hfsc_init_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1394 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1395 struct netlink_ext_ack *extack)
1396 {
1397 struct hfsc_sched *q = qdisc_priv(sch);
1398 struct tc_hfsc_qopt *qopt;
1399 int err;
1400
1401 qdisc_watchdog_init(&q->watchdog, sch);
1402
1403 if (!opt || nla_len(opt) < sizeof(*qopt))
1404 return -EINVAL;
1405 qopt = nla_data(opt);
1406
1407 q->defcls = qopt->defcls;
1408 err = qdisc_class_hash_init(&q->clhash);
1409 if (err < 0)
1410 return err;
1411 q->eligible = RB_ROOT;
1412
1413 err = tcf_block_get(&q->root.block, &q->root.filter_list, sch, extack);
1414 if (err)
1415 return err;
1416
1417 gnet_stats_basic_sync_init(&q->root.bstats);
1418 q->root.cl_common.classid = sch->handle;
1419 q->root.sched = q;
1420 q->root.qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1421 sch->handle, NULL);
1422 if (q->root.qdisc == NULL)
1423 q->root.qdisc = &noop_qdisc;
1424 else
1425 qdisc_hash_add(q->root.qdisc, true);
1426 INIT_LIST_HEAD(&q->root.children);
1427 q->root.vt_tree = RB_ROOT;
1428 q->root.cf_tree = RB_ROOT;
1429
1430 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common);
1431 qdisc_class_hash_grow(sch, &q->clhash);
1432
1433 return 0;
1434 }
1435
1436 static int
hfsc_change_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1437 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt,
1438 struct netlink_ext_ack *extack)
1439 {
1440 struct hfsc_sched *q = qdisc_priv(sch);
1441 struct tc_hfsc_qopt *qopt;
1442
1443 if (nla_len(opt) < sizeof(*qopt))
1444 return -EINVAL;
1445 qopt = nla_data(opt);
1446
1447 WRITE_ONCE(q->defcls, qopt->defcls);
1448
1449 return 0;
1450 }
1451
1452 static void
hfsc_reset_class(struct hfsc_class * cl)1453 hfsc_reset_class(struct hfsc_class *cl)
1454 {
1455 cl->cl_total = 0;
1456 cl->cl_cumul = 0;
1457 cl->cl_d = 0;
1458 cl->cl_e = 0;
1459 cl->cl_vt = 0;
1460 cl->cl_vtadj = 0;
1461 cl->cl_cvtmin = 0;
1462 cl->cl_cvtoff = 0;
1463 cl->cl_vtperiod = 0;
1464 cl->cl_parentperiod = 0;
1465 cl->cl_f = 0;
1466 cl->cl_myf = 0;
1467 cl->cl_cfmin = 0;
1468 cl->cl_nactive = 0;
1469
1470 cl->vt_tree = RB_ROOT;
1471 cl->cf_tree = RB_ROOT;
1472 qdisc_reset(cl->qdisc);
1473
1474 if (cl->cl_flags & HFSC_RSC)
1475 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0);
1476 if (cl->cl_flags & HFSC_FSC)
1477 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0);
1478 if (cl->cl_flags & HFSC_USC)
1479 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0);
1480 }
1481
1482 static void
hfsc_reset_qdisc(struct Qdisc * sch)1483 hfsc_reset_qdisc(struct Qdisc *sch)
1484 {
1485 struct hfsc_sched *q = qdisc_priv(sch);
1486 struct hfsc_class *cl;
1487 unsigned int i;
1488
1489 for (i = 0; i < q->clhash.hashsize; i++) {
1490 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode)
1491 hfsc_reset_class(cl);
1492 }
1493 q->eligible = RB_ROOT;
1494 qdisc_watchdog_cancel(&q->watchdog);
1495 }
1496
1497 static void
hfsc_destroy_qdisc(struct Qdisc * sch)1498 hfsc_destroy_qdisc(struct Qdisc *sch)
1499 {
1500 struct hfsc_sched *q = qdisc_priv(sch);
1501 struct hlist_node *next;
1502 struct hfsc_class *cl;
1503 unsigned int i;
1504
1505 for (i = 0; i < q->clhash.hashsize; i++) {
1506 hlist_for_each_entry(cl, &q->clhash.hash[i], cl_common.hnode) {
1507 tcf_block_put(cl->block);
1508 cl->block = NULL;
1509 }
1510 }
1511 for (i = 0; i < q->clhash.hashsize; i++) {
1512 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1513 cl_common.hnode)
1514 hfsc_destroy_class(sch, cl);
1515 }
1516 qdisc_class_hash_destroy(&q->clhash);
1517 qdisc_watchdog_cancel(&q->watchdog);
1518 }
1519
1520 static int
hfsc_dump_qdisc(struct Qdisc * sch,struct sk_buff * skb)1521 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb)
1522 {
1523 struct hfsc_sched *q = qdisc_priv(sch);
1524 unsigned char *b = skb_tail_pointer(skb);
1525 struct tc_hfsc_qopt qopt;
1526
1527 qopt.defcls = READ_ONCE(q->defcls);
1528 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1529 goto nla_put_failure;
1530 return skb->len;
1531
1532 nla_put_failure:
1533 nlmsg_trim(skb, b);
1534 return -1;
1535 }
1536
1537 static int
hfsc_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)1538 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free)
1539 {
1540 unsigned int len = qdisc_pkt_len(skb);
1541 struct hfsc_class *cl;
1542 int err;
1543 bool first;
1544
1545 cl = hfsc_classify(skb, sch, &err);
1546 if (cl == NULL) {
1547 if (err & __NET_XMIT_BYPASS)
1548 qdisc_qstats_drop(sch);
1549 __qdisc_drop(skb, to_free);
1550 return err;
1551 }
1552
1553 first = !cl->qdisc->q.qlen;
1554 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1555 if (unlikely(err != NET_XMIT_SUCCESS)) {
1556 if (net_xmit_drop_count(err)) {
1557 cl->qstats.drops++;
1558 qdisc_qstats_drop(sch);
1559 }
1560 return err;
1561 }
1562
1563 if (first) {
1564 if (cl->cl_flags & HFSC_RSC)
1565 init_ed(cl, len);
1566 if (cl->cl_flags & HFSC_FSC)
1567 init_vf(cl, len);
1568 /*
1569 * If this is the first packet, isolate the head so an eventual
1570 * head drop before the first dequeue operation has no chance
1571 * to invalidate the deadline.
1572 */
1573 if (cl->cl_flags & HFSC_RSC)
1574 cl->qdisc->ops->peek(cl->qdisc);
1575
1576 }
1577
1578 sch->qstats.backlog += len;
1579 sch->q.qlen++;
1580
1581 return NET_XMIT_SUCCESS;
1582 }
1583
1584 static struct sk_buff *
hfsc_dequeue(struct Qdisc * sch)1585 hfsc_dequeue(struct Qdisc *sch)
1586 {
1587 struct hfsc_sched *q = qdisc_priv(sch);
1588 struct hfsc_class *cl;
1589 struct sk_buff *skb;
1590 u64 cur_time;
1591 unsigned int next_len;
1592 int realtime = 0;
1593
1594 if (sch->q.qlen == 0)
1595 return NULL;
1596
1597 cur_time = psched_get_time();
1598
1599 /*
1600 * if there are eligible classes, use real-time criteria.
1601 * find the class with the minimum deadline among
1602 * the eligible classes.
1603 */
1604 cl = eltree_get_mindl(q, cur_time);
1605 if (cl) {
1606 realtime = 1;
1607 } else {
1608 /*
1609 * use link-sharing criteria
1610 * get the class with the minimum vt in the hierarchy
1611 */
1612 cl = vttree_get_minvt(&q->root, cur_time);
1613 if (cl == NULL) {
1614 qdisc_qstats_overlimit(sch);
1615 hfsc_schedule_watchdog(sch);
1616 return NULL;
1617 }
1618 }
1619
1620 skb = qdisc_dequeue_peeked(cl->qdisc);
1621 if (skb == NULL) {
1622 qdisc_warn_nonwc("HFSC", cl->qdisc);
1623 return NULL;
1624 }
1625
1626 bstats_update(&cl->bstats, skb);
1627 update_vf(cl, qdisc_pkt_len(skb), cur_time);
1628 if (realtime)
1629 cl->cl_cumul += qdisc_pkt_len(skb);
1630
1631 if (cl->cl_flags & HFSC_RSC) {
1632 if (cl->qdisc->q.qlen != 0) {
1633 /* update ed */
1634 next_len = qdisc_peek_len(cl->qdisc);
1635 if (realtime)
1636 update_ed(cl, next_len);
1637 else
1638 update_d(cl, next_len);
1639 } else {
1640 /* the class becomes passive */
1641 eltree_remove(cl);
1642 }
1643 }
1644
1645 qdisc_bstats_update(sch, skb);
1646 qdisc_qstats_backlog_dec(sch, skb);
1647 sch->q.qlen--;
1648
1649 return skb;
1650 }
1651
1652 static const struct Qdisc_class_ops hfsc_class_ops = {
1653 .change = hfsc_change_class,
1654 .delete = hfsc_delete_class,
1655 .graft = hfsc_graft_class,
1656 .leaf = hfsc_class_leaf,
1657 .qlen_notify = hfsc_qlen_notify,
1658 .find = hfsc_search_class,
1659 .bind_tcf = hfsc_bind_tcf,
1660 .unbind_tcf = hfsc_unbind_tcf,
1661 .tcf_block = hfsc_tcf_block,
1662 .dump = hfsc_dump_class,
1663 .dump_stats = hfsc_dump_class_stats,
1664 .walk = hfsc_walk
1665 };
1666
1667 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = {
1668 .id = "hfsc",
1669 .init = hfsc_init_qdisc,
1670 .change = hfsc_change_qdisc,
1671 .reset = hfsc_reset_qdisc,
1672 .destroy = hfsc_destroy_qdisc,
1673 .dump = hfsc_dump_qdisc,
1674 .enqueue = hfsc_enqueue,
1675 .dequeue = hfsc_dequeue,
1676 .peek = qdisc_peek_dequeued,
1677 .cl_ops = &hfsc_class_ops,
1678 .priv_size = sizeof(struct hfsc_sched),
1679 .owner = THIS_MODULE
1680 };
1681 MODULE_ALIAS_NET_SCH("hfsc");
1682
1683 static int __init
hfsc_init(void)1684 hfsc_init(void)
1685 {
1686 return register_qdisc(&hfsc_qdisc_ops);
1687 }
1688
1689 static void __exit
hfsc_cleanup(void)1690 hfsc_cleanup(void)
1691 {
1692 unregister_qdisc(&hfsc_qdisc_ops);
1693 }
1694
1695 MODULE_LICENSE("GPL");
1696 MODULE_DESCRIPTION("Hierarchical Fair Service Curve scheduler");
1697 module_init(hfsc_init);
1698 module_exit(hfsc_cleanup);
1699