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