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 { 86 u64 sm1; /* scaled slope of the 1st segment */ 87 u64 ism1; /* scaled inverse-slope of the 1st segment */ 88 u64 dx; /* the x-projection of the 1st segment */ 89 u64 dy; /* the y-projection of the 1st segment */ 90 u64 sm2; /* scaled slope of the 2nd segment */ 91 u64 ism2; /* scaled inverse-slope of the 2nd segment */ 92 }; 93 94 /* runtime service curve */ 95 struct runtime_sc 96 { 97 u64 x; /* current starting position on x-axis */ 98 u64 y; /* current starting position on y-axis */ 99 u64 sm1; /* scaled slope of the 1st segment */ 100 u64 ism1; /* scaled inverse-slope of the 1st segment */ 101 u64 dx; /* the x-projection of the 1st segment */ 102 u64 dy; /* the y-projection of the 1st segment */ 103 u64 sm2; /* scaled slope of the 2nd segment */ 104 u64 ism2; /* scaled inverse-slope of the 2nd segment */ 105 }; 106 107 enum hfsc_class_flags 108 { 109 HFSC_RSC = 0x1, 110 HFSC_FSC = 0x2, 111 HFSC_USC = 0x4 112 }; 113 114 struct hfsc_class 115 { 116 struct Qdisc_class_common cl_common; 117 unsigned int refcnt; /* usage count */ 118 119 struct gnet_stats_basic_packed bstats; 120 struct gnet_stats_queue qstats; 121 struct gnet_stats_rate_est rate_est; 122 unsigned int level; /* class level in hierarchy */ 123 struct tcf_proto *filter_list; /* filter list */ 124 unsigned int filter_cnt; /* filter count */ 125 126 struct hfsc_sched *sched; /* scheduler data */ 127 struct hfsc_class *cl_parent; /* parent class */ 128 struct list_head siblings; /* sibling classes */ 129 struct list_head children; /* child classes */ 130 struct Qdisc *qdisc; /* leaf qdisc */ 131 132 struct rb_node el_node; /* qdisc's eligible tree member */ 133 struct rb_root vt_tree; /* active children sorted by cl_vt */ 134 struct rb_node vt_node; /* parent's vt_tree member */ 135 struct rb_root cf_tree; /* active children sorted by cl_f */ 136 struct rb_node cf_node; /* parent's cf_heap member */ 137 struct list_head dlist; /* drop list member */ 138 139 u64 cl_total; /* total work in bytes */ 140 u64 cl_cumul; /* cumulative work in bytes done by 141 real-time criteria */ 142 143 u64 cl_d; /* deadline*/ 144 u64 cl_e; /* eligible time */ 145 u64 cl_vt; /* virtual time */ 146 u64 cl_f; /* time when this class will fit for 147 link-sharing, max(myf, cfmin) */ 148 u64 cl_myf; /* my fit-time (calculated from this 149 class's own upperlimit curve) */ 150 u64 cl_myfadj; /* my fit-time adjustment (to cancel 151 history dependence) */ 152 u64 cl_cfmin; /* earliest children's fit-time (used 153 with cl_myf to obtain cl_f) */ 154 u64 cl_cvtmin; /* minimal virtual time among the 155 children fit for link-sharing 156 (monotonic within a period) */ 157 u64 cl_vtadj; /* intra-period cumulative vt 158 adjustment */ 159 u64 cl_vtoff; /* inter-period cumulative vt offset */ 160 u64 cl_cvtmax; /* max child's vt in the last period */ 161 u64 cl_cvtoff; /* cumulative cvtmax of all periods */ 162 u64 cl_pcvtoff; /* parent's cvtoff at initialization 163 time */ 164 165 struct internal_sc cl_rsc; /* internal real-time service curve */ 166 struct internal_sc cl_fsc; /* internal fair service curve */ 167 struct internal_sc cl_usc; /* internal upperlimit service curve */ 168 struct runtime_sc cl_deadline; /* deadline curve */ 169 struct runtime_sc cl_eligible; /* eligible curve */ 170 struct runtime_sc cl_virtual; /* virtual curve */ 171 struct runtime_sc cl_ulimit; /* upperlimit curve */ 172 173 unsigned long cl_flags; /* which curves are valid */ 174 unsigned long cl_vtperiod; /* vt period sequence number */ 175 unsigned long cl_parentperiod;/* parent's vt period sequence number*/ 176 unsigned long cl_nactive; /* number of active children */ 177 }; 178 179 struct hfsc_sched 180 { 181 u16 defcls; /* default class id */ 182 struct hfsc_class root; /* root class */ 183 struct Qdisc_class_hash clhash; /* class hash */ 184 struct rb_root eligible; /* eligible tree */ 185 struct list_head droplist; /* active leaf class list (for 186 dropping) */ 187 struct qdisc_watchdog watchdog; /* watchdog timer */ 188 }; 189 190 #define HT_INFINITY 0xffffffffffffffffULL /* infinite time value */ 191 192 193 /* 194 * eligible tree holds backlogged classes being sorted by their eligible times. 195 * there is one eligible tree per hfsc instance. 196 */ 197 198 static void 199 eltree_insert(struct hfsc_class *cl) 200 { 201 struct rb_node **p = &cl->sched->eligible.rb_node; 202 struct rb_node *parent = NULL; 203 struct hfsc_class *cl1; 204 205 while (*p != NULL) { 206 parent = *p; 207 cl1 = rb_entry(parent, struct hfsc_class, el_node); 208 if (cl->cl_e >= cl1->cl_e) 209 p = &parent->rb_right; 210 else 211 p = &parent->rb_left; 212 } 213 rb_link_node(&cl->el_node, parent, p); 214 rb_insert_color(&cl->el_node, &cl->sched->eligible); 215 } 216 217 static inline void 218 eltree_remove(struct hfsc_class *cl) 219 { 220 rb_erase(&cl->el_node, &cl->sched->eligible); 221 } 222 223 static inline void 224 eltree_update(struct hfsc_class *cl) 225 { 226 eltree_remove(cl); 227 eltree_insert(cl); 228 } 229 230 /* find the class with the minimum deadline among the eligible classes */ 231 static inline struct hfsc_class * 232 eltree_get_mindl(struct hfsc_sched *q, u64 cur_time) 233 { 234 struct hfsc_class *p, *cl = NULL; 235 struct rb_node *n; 236 237 for (n = rb_first(&q->eligible); n != NULL; n = rb_next(n)) { 238 p = rb_entry(n, struct hfsc_class, el_node); 239 if (p->cl_e > cur_time) 240 break; 241 if (cl == NULL || p->cl_d < cl->cl_d) 242 cl = p; 243 } 244 return cl; 245 } 246 247 /* find the class with minimum eligible time among the eligible classes */ 248 static inline struct hfsc_class * 249 eltree_get_minel(struct hfsc_sched *q) 250 { 251 struct rb_node *n; 252 253 n = rb_first(&q->eligible); 254 if (n == NULL) 255 return NULL; 256 return rb_entry(n, struct hfsc_class, el_node); 257 } 258 259 /* 260 * vttree holds holds backlogged child classes being sorted by their virtual 261 * time. each intermediate class has one vttree. 262 */ 263 static void 264 vttree_insert(struct hfsc_class *cl) 265 { 266 struct rb_node **p = &cl->cl_parent->vt_tree.rb_node; 267 struct rb_node *parent = NULL; 268 struct hfsc_class *cl1; 269 270 while (*p != NULL) { 271 parent = *p; 272 cl1 = rb_entry(parent, struct hfsc_class, vt_node); 273 if (cl->cl_vt >= cl1->cl_vt) 274 p = &parent->rb_right; 275 else 276 p = &parent->rb_left; 277 } 278 rb_link_node(&cl->vt_node, parent, p); 279 rb_insert_color(&cl->vt_node, &cl->cl_parent->vt_tree); 280 } 281 282 static inline void 283 vttree_remove(struct hfsc_class *cl) 284 { 285 rb_erase(&cl->vt_node, &cl->cl_parent->vt_tree); 286 } 287 288 static inline void 289 vttree_update(struct hfsc_class *cl) 290 { 291 vttree_remove(cl); 292 vttree_insert(cl); 293 } 294 295 static inline struct hfsc_class * 296 vttree_firstfit(struct hfsc_class *cl, u64 cur_time) 297 { 298 struct hfsc_class *p; 299 struct rb_node *n; 300 301 for (n = rb_first(&cl->vt_tree); n != NULL; n = rb_next(n)) { 302 p = rb_entry(n, struct hfsc_class, vt_node); 303 if (p->cl_f <= cur_time) 304 return p; 305 } 306 return NULL; 307 } 308 309 /* 310 * get the leaf class with the minimum vt in the hierarchy 311 */ 312 static struct hfsc_class * 313 vttree_get_minvt(struct hfsc_class *cl, u64 cur_time) 314 { 315 /* if root-class's cfmin is bigger than cur_time nothing to do */ 316 if (cl->cl_cfmin > cur_time) 317 return NULL; 318 319 while (cl->level > 0) { 320 cl = vttree_firstfit(cl, cur_time); 321 if (cl == NULL) 322 return NULL; 323 /* 324 * update parent's cl_cvtmin. 325 */ 326 if (cl->cl_parent->cl_cvtmin < cl->cl_vt) 327 cl->cl_parent->cl_cvtmin = cl->cl_vt; 328 } 329 return cl; 330 } 331 332 static void 333 cftree_insert(struct hfsc_class *cl) 334 { 335 struct rb_node **p = &cl->cl_parent->cf_tree.rb_node; 336 struct rb_node *parent = NULL; 337 struct hfsc_class *cl1; 338 339 while (*p != NULL) { 340 parent = *p; 341 cl1 = rb_entry(parent, struct hfsc_class, cf_node); 342 if (cl->cl_f >= cl1->cl_f) 343 p = &parent->rb_right; 344 else 345 p = &parent->rb_left; 346 } 347 rb_link_node(&cl->cf_node, parent, p); 348 rb_insert_color(&cl->cf_node, &cl->cl_parent->cf_tree); 349 } 350 351 static inline void 352 cftree_remove(struct hfsc_class *cl) 353 { 354 rb_erase(&cl->cf_node, &cl->cl_parent->cf_tree); 355 } 356 357 static inline void 358 cftree_update(struct hfsc_class *cl) 359 { 360 cftree_remove(cl); 361 cftree_insert(cl); 362 } 363 364 /* 365 * service curve support functions 366 * 367 * external service curve parameters 368 * m: bps 369 * d: us 370 * internal service curve parameters 371 * sm: (bytes/psched_us) << SM_SHIFT 372 * ism: (psched_us/byte) << ISM_SHIFT 373 * dx: psched_us 374 * 375 * The clock source resolution with ktime and PSCHED_SHIFT 10 is 1.024us. 376 * 377 * sm and ism are scaled in order to keep effective digits. 378 * SM_SHIFT and ISM_SHIFT are selected to keep at least 4 effective 379 * digits in decimal using the following table. 380 * 381 * bits/sec 100Kbps 1Mbps 10Mbps 100Mbps 1Gbps 382 * ------------+------------------------------------------------------- 383 * bytes/1.024us 12.8e-3 128e-3 1280e-3 12800e-3 128000e-3 384 * 385 * 1.024us/byte 78.125 7.8125 0.78125 0.078125 0.0078125 386 * 387 * So, for PSCHED_SHIFT 10 we need: SM_SHIFT 20, ISM_SHIFT 18. 388 */ 389 #define SM_SHIFT (30 - PSCHED_SHIFT) 390 #define ISM_SHIFT (8 + PSCHED_SHIFT) 391 392 #define SM_MASK ((1ULL << SM_SHIFT) - 1) 393 #define ISM_MASK ((1ULL << ISM_SHIFT) - 1) 394 395 static inline u64 396 seg_x2y(u64 x, u64 sm) 397 { 398 u64 y; 399 400 /* 401 * compute 402 * y = x * sm >> SM_SHIFT 403 * but divide it for the upper and lower bits to avoid overflow 404 */ 405 y = (x >> SM_SHIFT) * sm + (((x & SM_MASK) * sm) >> SM_SHIFT); 406 return y; 407 } 408 409 static inline u64 410 seg_y2x(u64 y, u64 ism) 411 { 412 u64 x; 413 414 if (y == 0) 415 x = 0; 416 else if (ism == HT_INFINITY) 417 x = HT_INFINITY; 418 else { 419 x = (y >> ISM_SHIFT) * ism 420 + (((y & ISM_MASK) * ism) >> ISM_SHIFT); 421 } 422 return x; 423 } 424 425 /* Convert m (bps) into sm (bytes/psched us) */ 426 static u64 427 m2sm(u32 m) 428 { 429 u64 sm; 430 431 sm = ((u64)m << SM_SHIFT); 432 sm += PSCHED_TICKS_PER_SEC - 1; 433 do_div(sm, PSCHED_TICKS_PER_SEC); 434 return sm; 435 } 436 437 /* convert m (bps) into ism (psched us/byte) */ 438 static u64 439 m2ism(u32 m) 440 { 441 u64 ism; 442 443 if (m == 0) 444 ism = HT_INFINITY; 445 else { 446 ism = ((u64)PSCHED_TICKS_PER_SEC << ISM_SHIFT); 447 ism += m - 1; 448 do_div(ism, m); 449 } 450 return ism; 451 } 452 453 /* convert d (us) into dx (psched us) */ 454 static u64 455 d2dx(u32 d) 456 { 457 u64 dx; 458 459 dx = ((u64)d * PSCHED_TICKS_PER_SEC); 460 dx += USEC_PER_SEC - 1; 461 do_div(dx, USEC_PER_SEC); 462 return dx; 463 } 464 465 /* convert sm (bytes/psched us) into m (bps) */ 466 static u32 467 sm2m(u64 sm) 468 { 469 u64 m; 470 471 m = (sm * PSCHED_TICKS_PER_SEC) >> SM_SHIFT; 472 return (u32)m; 473 } 474 475 /* convert dx (psched us) into d (us) */ 476 static u32 477 dx2d(u64 dx) 478 { 479 u64 d; 480 481 d = dx * USEC_PER_SEC; 482 do_div(d, PSCHED_TICKS_PER_SEC); 483 return (u32)d; 484 } 485 486 static void 487 sc2isc(struct tc_service_curve *sc, struct internal_sc *isc) 488 { 489 isc->sm1 = m2sm(sc->m1); 490 isc->ism1 = m2ism(sc->m1); 491 isc->dx = d2dx(sc->d); 492 isc->dy = seg_x2y(isc->dx, isc->sm1); 493 isc->sm2 = m2sm(sc->m2); 494 isc->ism2 = m2ism(sc->m2); 495 } 496 497 /* 498 * initialize the runtime service curve with the given internal 499 * service curve starting at (x, y). 500 */ 501 static void 502 rtsc_init(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y) 503 { 504 rtsc->x = x; 505 rtsc->y = y; 506 rtsc->sm1 = isc->sm1; 507 rtsc->ism1 = isc->ism1; 508 rtsc->dx = isc->dx; 509 rtsc->dy = isc->dy; 510 rtsc->sm2 = isc->sm2; 511 rtsc->ism2 = isc->ism2; 512 } 513 514 /* 515 * calculate the y-projection of the runtime service curve by the 516 * given x-projection value 517 */ 518 static u64 519 rtsc_y2x(struct runtime_sc *rtsc, u64 y) 520 { 521 u64 x; 522 523 if (y < rtsc->y) 524 x = rtsc->x; 525 else if (y <= rtsc->y + rtsc->dy) { 526 /* x belongs to the 1st segment */ 527 if (rtsc->dy == 0) 528 x = rtsc->x + rtsc->dx; 529 else 530 x = rtsc->x + seg_y2x(y - rtsc->y, rtsc->ism1); 531 } else { 532 /* x belongs to the 2nd segment */ 533 x = rtsc->x + rtsc->dx 534 + seg_y2x(y - rtsc->y - rtsc->dy, rtsc->ism2); 535 } 536 return x; 537 } 538 539 static u64 540 rtsc_x2y(struct runtime_sc *rtsc, u64 x) 541 { 542 u64 y; 543 544 if (x <= rtsc->x) 545 y = rtsc->y; 546 else if (x <= rtsc->x + rtsc->dx) 547 /* y belongs to the 1st segment */ 548 y = rtsc->y + seg_x2y(x - rtsc->x, rtsc->sm1); 549 else 550 /* y belongs to the 2nd segment */ 551 y = rtsc->y + rtsc->dy 552 + seg_x2y(x - rtsc->x - rtsc->dx, rtsc->sm2); 553 return y; 554 } 555 556 /* 557 * update the runtime service curve by taking the minimum of the current 558 * runtime service curve and the service curve starting at (x, y). 559 */ 560 static void 561 rtsc_min(struct runtime_sc *rtsc, struct internal_sc *isc, u64 x, u64 y) 562 { 563 u64 y1, y2, dx, dy; 564 u32 dsm; 565 566 if (isc->sm1 <= isc->sm2) { 567 /* service curve is convex */ 568 y1 = rtsc_x2y(rtsc, x); 569 if (y1 < y) 570 /* the current rtsc is smaller */ 571 return; 572 rtsc->x = x; 573 rtsc->y = y; 574 return; 575 } 576 577 /* 578 * service curve is concave 579 * compute the two y values of the current rtsc 580 * y1: at x 581 * y2: at (x + dx) 582 */ 583 y1 = rtsc_x2y(rtsc, x); 584 if (y1 <= y) { 585 /* rtsc is below isc, no change to rtsc */ 586 return; 587 } 588 589 y2 = rtsc_x2y(rtsc, x + isc->dx); 590 if (y2 >= y + isc->dy) { 591 /* rtsc is above isc, replace rtsc by isc */ 592 rtsc->x = x; 593 rtsc->y = y; 594 rtsc->dx = isc->dx; 595 rtsc->dy = isc->dy; 596 return; 597 } 598 599 /* 600 * the two curves intersect 601 * compute the offsets (dx, dy) using the reverse 602 * function of seg_x2y() 603 * seg_x2y(dx, sm1) == seg_x2y(dx, sm2) + (y1 - y) 604 */ 605 dx = (y1 - y) << SM_SHIFT; 606 dsm = isc->sm1 - isc->sm2; 607 do_div(dx, dsm); 608 /* 609 * check if (x, y1) belongs to the 1st segment of rtsc. 610 * if so, add the offset. 611 */ 612 if (rtsc->x + rtsc->dx > x) 613 dx += rtsc->x + rtsc->dx - x; 614 dy = seg_x2y(dx, isc->sm1); 615 616 rtsc->x = x; 617 rtsc->y = y; 618 rtsc->dx = dx; 619 rtsc->dy = dy; 620 return; 621 } 622 623 static void 624 init_ed(struct hfsc_class *cl, unsigned int next_len) 625 { 626 u64 cur_time = psched_get_time(); 627 628 /* update the deadline curve */ 629 rtsc_min(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul); 630 631 /* 632 * update the eligible curve. 633 * for concave, it is equal to the deadline curve. 634 * for convex, it is a linear curve with slope m2. 635 */ 636 cl->cl_eligible = cl->cl_deadline; 637 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) { 638 cl->cl_eligible.dx = 0; 639 cl->cl_eligible.dy = 0; 640 } 641 642 /* compute e and d */ 643 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul); 644 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len); 645 646 eltree_insert(cl); 647 } 648 649 static void 650 update_ed(struct hfsc_class *cl, unsigned int next_len) 651 { 652 cl->cl_e = rtsc_y2x(&cl->cl_eligible, cl->cl_cumul); 653 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len); 654 655 eltree_update(cl); 656 } 657 658 static inline void 659 update_d(struct hfsc_class *cl, unsigned int next_len) 660 { 661 cl->cl_d = rtsc_y2x(&cl->cl_deadline, cl->cl_cumul + next_len); 662 } 663 664 static inline void 665 update_cfmin(struct hfsc_class *cl) 666 { 667 struct rb_node *n = rb_first(&cl->cf_tree); 668 struct hfsc_class *p; 669 670 if (n == NULL) { 671 cl->cl_cfmin = 0; 672 return; 673 } 674 p = rb_entry(n, struct hfsc_class, cf_node); 675 cl->cl_cfmin = p->cl_f; 676 } 677 678 static void 679 init_vf(struct hfsc_class *cl, unsigned int len) 680 { 681 struct hfsc_class *max_cl; 682 struct rb_node *n; 683 u64 vt, f, cur_time; 684 int go_active; 685 686 cur_time = 0; 687 go_active = 1; 688 for (; cl->cl_parent != NULL; cl = cl->cl_parent) { 689 if (go_active && cl->cl_nactive++ == 0) 690 go_active = 1; 691 else 692 go_active = 0; 693 694 if (go_active) { 695 n = rb_last(&cl->cl_parent->vt_tree); 696 if (n != NULL) { 697 max_cl = rb_entry(n, struct hfsc_class,vt_node); 698 /* 699 * set vt to the average of the min and max 700 * classes. if the parent's period didn't 701 * change, don't decrease vt of the class. 702 */ 703 vt = max_cl->cl_vt; 704 if (cl->cl_parent->cl_cvtmin != 0) 705 vt = (cl->cl_parent->cl_cvtmin + vt)/2; 706 707 if (cl->cl_parent->cl_vtperiod != 708 cl->cl_parentperiod || vt > cl->cl_vt) 709 cl->cl_vt = vt; 710 } else { 711 /* 712 * first child for a new parent backlog period. 713 * add parent's cvtmax to cvtoff to make a new 714 * vt (vtoff + vt) larger than the vt in the 715 * last period for all children. 716 */ 717 vt = cl->cl_parent->cl_cvtmax; 718 cl->cl_parent->cl_cvtoff += vt; 719 cl->cl_parent->cl_cvtmax = 0; 720 cl->cl_parent->cl_cvtmin = 0; 721 cl->cl_vt = 0; 722 } 723 724 cl->cl_vtoff = cl->cl_parent->cl_cvtoff - 725 cl->cl_pcvtoff; 726 727 /* update the virtual curve */ 728 vt = cl->cl_vt + cl->cl_vtoff; 729 rtsc_min(&cl->cl_virtual, &cl->cl_fsc, vt, 730 cl->cl_total); 731 if (cl->cl_virtual.x == vt) { 732 cl->cl_virtual.x -= cl->cl_vtoff; 733 cl->cl_vtoff = 0; 734 } 735 cl->cl_vtadj = 0; 736 737 cl->cl_vtperiod++; /* increment vt period */ 738 cl->cl_parentperiod = cl->cl_parent->cl_vtperiod; 739 if (cl->cl_parent->cl_nactive == 0) 740 cl->cl_parentperiod++; 741 cl->cl_f = 0; 742 743 vttree_insert(cl); 744 cftree_insert(cl); 745 746 if (cl->cl_flags & HFSC_USC) { 747 /* class has upper limit curve */ 748 if (cur_time == 0) 749 cur_time = psched_get_time(); 750 751 /* update the ulimit curve */ 752 rtsc_min(&cl->cl_ulimit, &cl->cl_usc, cur_time, 753 cl->cl_total); 754 /* compute myf */ 755 cl->cl_myf = rtsc_y2x(&cl->cl_ulimit, 756 cl->cl_total); 757 cl->cl_myfadj = 0; 758 } 759 } 760 761 f = max(cl->cl_myf, cl->cl_cfmin); 762 if (f != cl->cl_f) { 763 cl->cl_f = f; 764 cftree_update(cl); 765 update_cfmin(cl->cl_parent); 766 } 767 } 768 } 769 770 static void 771 update_vf(struct hfsc_class *cl, unsigned int len, u64 cur_time) 772 { 773 u64 f; /* , myf_bound, delta; */ 774 int go_passive = 0; 775 776 if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC) 777 go_passive = 1; 778 779 for (; cl->cl_parent != NULL; cl = cl->cl_parent) { 780 cl->cl_total += len; 781 782 if (!(cl->cl_flags & HFSC_FSC) || cl->cl_nactive == 0) 783 continue; 784 785 if (go_passive && --cl->cl_nactive == 0) 786 go_passive = 1; 787 else 788 go_passive = 0; 789 790 if (go_passive) { 791 /* no more active child, going passive */ 792 793 /* update cvtmax of the parent class */ 794 if (cl->cl_vt > cl->cl_parent->cl_cvtmax) 795 cl->cl_parent->cl_cvtmax = cl->cl_vt; 796 797 /* remove this class from the vt tree */ 798 vttree_remove(cl); 799 800 cftree_remove(cl); 801 update_cfmin(cl->cl_parent); 802 803 continue; 804 } 805 806 /* 807 * update vt and f 808 */ 809 cl->cl_vt = rtsc_y2x(&cl->cl_virtual, cl->cl_total) 810 - cl->cl_vtoff + cl->cl_vtadj; 811 812 /* 813 * if vt of the class is smaller than cvtmin, 814 * the class was skipped in the past due to non-fit. 815 * if so, we need to adjust vtadj. 816 */ 817 if (cl->cl_vt < cl->cl_parent->cl_cvtmin) { 818 cl->cl_vtadj += cl->cl_parent->cl_cvtmin - cl->cl_vt; 819 cl->cl_vt = cl->cl_parent->cl_cvtmin; 820 } 821 822 /* update the vt tree */ 823 vttree_update(cl); 824 825 if (cl->cl_flags & HFSC_USC) { 826 cl->cl_myf = cl->cl_myfadj + rtsc_y2x(&cl->cl_ulimit, 827 cl->cl_total); 828 #if 0 829 /* 830 * This code causes classes to stay way under their 831 * limit when multiple classes are used at gigabit 832 * speed. needs investigation. -kaber 833 */ 834 /* 835 * if myf lags behind by more than one clock tick 836 * from the current time, adjust myfadj to prevent 837 * a rate-limited class from going greedy. 838 * in a steady state under rate-limiting, myf 839 * fluctuates within one clock tick. 840 */ 841 myf_bound = cur_time - PSCHED_JIFFIE2US(1); 842 if (cl->cl_myf < myf_bound) { 843 delta = cur_time - cl->cl_myf; 844 cl->cl_myfadj += delta; 845 cl->cl_myf += delta; 846 } 847 #endif 848 } 849 850 f = max(cl->cl_myf, cl->cl_cfmin); 851 if (f != cl->cl_f) { 852 cl->cl_f = f; 853 cftree_update(cl); 854 update_cfmin(cl->cl_parent); 855 } 856 } 857 } 858 859 static void 860 set_active(struct hfsc_class *cl, unsigned int len) 861 { 862 if (cl->cl_flags & HFSC_RSC) 863 init_ed(cl, len); 864 if (cl->cl_flags & HFSC_FSC) 865 init_vf(cl, len); 866 867 list_add_tail(&cl->dlist, &cl->sched->droplist); 868 } 869 870 static void 871 set_passive(struct hfsc_class *cl) 872 { 873 if (cl->cl_flags & HFSC_RSC) 874 eltree_remove(cl); 875 876 list_del(&cl->dlist); 877 878 /* 879 * vttree is now handled in update_vf() so that update_vf(cl, 0, 0) 880 * needs to be called explicitly to remove a class from vttree. 881 */ 882 } 883 884 static unsigned int 885 qdisc_peek_len(struct Qdisc *sch) 886 { 887 struct sk_buff *skb; 888 unsigned int len; 889 890 skb = sch->ops->peek(sch); 891 if (skb == NULL) { 892 qdisc_warn_nonwc("qdisc_peek_len", sch); 893 return 0; 894 } 895 len = qdisc_pkt_len(skb); 896 897 return len; 898 } 899 900 static void 901 hfsc_purge_queue(struct Qdisc *sch, struct hfsc_class *cl) 902 { 903 unsigned int len = cl->qdisc->q.qlen; 904 905 qdisc_reset(cl->qdisc); 906 qdisc_tree_decrease_qlen(cl->qdisc, len); 907 } 908 909 static void 910 hfsc_adjust_levels(struct hfsc_class *cl) 911 { 912 struct hfsc_class *p; 913 unsigned int level; 914 915 do { 916 level = 0; 917 list_for_each_entry(p, &cl->children, siblings) { 918 if (p->level >= level) 919 level = p->level + 1; 920 } 921 cl->level = level; 922 } while ((cl = cl->cl_parent) != NULL); 923 } 924 925 static inline struct hfsc_class * 926 hfsc_find_class(u32 classid, struct Qdisc *sch) 927 { 928 struct hfsc_sched *q = qdisc_priv(sch); 929 struct Qdisc_class_common *clc; 930 931 clc = qdisc_class_find(&q->clhash, classid); 932 if (clc == NULL) 933 return NULL; 934 return container_of(clc, struct hfsc_class, cl_common); 935 } 936 937 static void 938 hfsc_change_rsc(struct hfsc_class *cl, struct tc_service_curve *rsc, 939 u64 cur_time) 940 { 941 sc2isc(rsc, &cl->cl_rsc); 942 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, cur_time, cl->cl_cumul); 943 cl->cl_eligible = cl->cl_deadline; 944 if (cl->cl_rsc.sm1 <= cl->cl_rsc.sm2) { 945 cl->cl_eligible.dx = 0; 946 cl->cl_eligible.dy = 0; 947 } 948 cl->cl_flags |= HFSC_RSC; 949 } 950 951 static void 952 hfsc_change_fsc(struct hfsc_class *cl, struct tc_service_curve *fsc) 953 { 954 sc2isc(fsc, &cl->cl_fsc); 955 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, cl->cl_vt, cl->cl_total); 956 cl->cl_flags |= HFSC_FSC; 957 } 958 959 static void 960 hfsc_change_usc(struct hfsc_class *cl, struct tc_service_curve *usc, 961 u64 cur_time) 962 { 963 sc2isc(usc, &cl->cl_usc); 964 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, cur_time, cl->cl_total); 965 cl->cl_flags |= HFSC_USC; 966 } 967 968 static const struct nla_policy hfsc_policy[TCA_HFSC_MAX + 1] = { 969 [TCA_HFSC_RSC] = { .len = sizeof(struct tc_service_curve) }, 970 [TCA_HFSC_FSC] = { .len = sizeof(struct tc_service_curve) }, 971 [TCA_HFSC_USC] = { .len = sizeof(struct tc_service_curve) }, 972 }; 973 974 static int 975 hfsc_change_class(struct Qdisc *sch, u32 classid, u32 parentid, 976 struct nlattr **tca, unsigned long *arg) 977 { 978 struct hfsc_sched *q = qdisc_priv(sch); 979 struct hfsc_class *cl = (struct hfsc_class *)*arg; 980 struct hfsc_class *parent = NULL; 981 struct nlattr *opt = tca[TCA_OPTIONS]; 982 struct nlattr *tb[TCA_HFSC_MAX + 1]; 983 struct tc_service_curve *rsc = NULL, *fsc = NULL, *usc = NULL; 984 u64 cur_time; 985 int err; 986 987 if (opt == NULL) 988 return -EINVAL; 989 990 err = nla_parse_nested(tb, TCA_HFSC_MAX, opt, hfsc_policy); 991 if (err < 0) 992 return err; 993 994 if (tb[TCA_HFSC_RSC]) { 995 rsc = nla_data(tb[TCA_HFSC_RSC]); 996 if (rsc->m1 == 0 && rsc->m2 == 0) 997 rsc = NULL; 998 } 999 1000 if (tb[TCA_HFSC_FSC]) { 1001 fsc = nla_data(tb[TCA_HFSC_FSC]); 1002 if (fsc->m1 == 0 && fsc->m2 == 0) 1003 fsc = NULL; 1004 } 1005 1006 if (tb[TCA_HFSC_USC]) { 1007 usc = nla_data(tb[TCA_HFSC_USC]); 1008 if (usc->m1 == 0 && usc->m2 == 0) 1009 usc = NULL; 1010 } 1011 1012 if (cl != NULL) { 1013 if (parentid) { 1014 if (cl->cl_parent && 1015 cl->cl_parent->cl_common.classid != parentid) 1016 return -EINVAL; 1017 if (cl->cl_parent == NULL && parentid != TC_H_ROOT) 1018 return -EINVAL; 1019 } 1020 cur_time = psched_get_time(); 1021 1022 if (tca[TCA_RATE]) { 1023 err = gen_replace_estimator(&cl->bstats, &cl->rate_est, 1024 qdisc_root_sleeping_lock(sch), 1025 tca[TCA_RATE]); 1026 if (err) 1027 return err; 1028 } 1029 1030 sch_tree_lock(sch); 1031 if (rsc != NULL) 1032 hfsc_change_rsc(cl, rsc, cur_time); 1033 if (fsc != NULL) 1034 hfsc_change_fsc(cl, fsc); 1035 if (usc != NULL) 1036 hfsc_change_usc(cl, usc, cur_time); 1037 1038 if (cl->qdisc->q.qlen != 0) { 1039 if (cl->cl_flags & HFSC_RSC) 1040 update_ed(cl, qdisc_peek_len(cl->qdisc)); 1041 if (cl->cl_flags & HFSC_FSC) 1042 update_vf(cl, 0, cur_time); 1043 } 1044 sch_tree_unlock(sch); 1045 1046 return 0; 1047 } 1048 1049 if (parentid == TC_H_ROOT) 1050 return -EEXIST; 1051 1052 parent = &q->root; 1053 if (parentid) { 1054 parent = hfsc_find_class(parentid, sch); 1055 if (parent == NULL) 1056 return -ENOENT; 1057 } 1058 1059 if (classid == 0 || TC_H_MAJ(classid ^ sch->handle) != 0) 1060 return -EINVAL; 1061 if (hfsc_find_class(classid, sch)) 1062 return -EEXIST; 1063 1064 if (rsc == NULL && fsc == NULL) 1065 return -EINVAL; 1066 1067 cl = kzalloc(sizeof(struct hfsc_class), GFP_KERNEL); 1068 if (cl == NULL) 1069 return -ENOBUFS; 1070 1071 if (tca[TCA_RATE]) { 1072 err = gen_new_estimator(&cl->bstats, &cl->rate_est, 1073 qdisc_root_sleeping_lock(sch), 1074 tca[TCA_RATE]); 1075 if (err) { 1076 kfree(cl); 1077 return err; 1078 } 1079 } 1080 1081 if (rsc != NULL) 1082 hfsc_change_rsc(cl, rsc, 0); 1083 if (fsc != NULL) 1084 hfsc_change_fsc(cl, fsc); 1085 if (usc != NULL) 1086 hfsc_change_usc(cl, usc, 0); 1087 1088 cl->cl_common.classid = classid; 1089 cl->refcnt = 1; 1090 cl->sched = q; 1091 cl->cl_parent = parent; 1092 cl->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue, 1093 &pfifo_qdisc_ops, classid); 1094 if (cl->qdisc == NULL) 1095 cl->qdisc = &noop_qdisc; 1096 INIT_LIST_HEAD(&cl->children); 1097 cl->vt_tree = RB_ROOT; 1098 cl->cf_tree = RB_ROOT; 1099 1100 sch_tree_lock(sch); 1101 qdisc_class_hash_insert(&q->clhash, &cl->cl_common); 1102 list_add_tail(&cl->siblings, &parent->children); 1103 if (parent->level == 0) 1104 hfsc_purge_queue(sch, parent); 1105 hfsc_adjust_levels(parent); 1106 cl->cl_pcvtoff = parent->cl_cvtoff; 1107 sch_tree_unlock(sch); 1108 1109 qdisc_class_hash_grow(sch, &q->clhash); 1110 1111 *arg = (unsigned long)cl; 1112 return 0; 1113 } 1114 1115 static void 1116 hfsc_destroy_class(struct Qdisc *sch, struct hfsc_class *cl) 1117 { 1118 struct hfsc_sched *q = qdisc_priv(sch); 1119 1120 tcf_destroy_chain(&cl->filter_list); 1121 qdisc_destroy(cl->qdisc); 1122 gen_kill_estimator(&cl->bstats, &cl->rate_est); 1123 if (cl != &q->root) 1124 kfree(cl); 1125 } 1126 1127 static int 1128 hfsc_delete_class(struct Qdisc *sch, unsigned long arg) 1129 { 1130 struct hfsc_sched *q = qdisc_priv(sch); 1131 struct hfsc_class *cl = (struct hfsc_class *)arg; 1132 1133 if (cl->level > 0 || cl->filter_cnt > 0 || cl == &q->root) 1134 return -EBUSY; 1135 1136 sch_tree_lock(sch); 1137 1138 list_del(&cl->siblings); 1139 hfsc_adjust_levels(cl->cl_parent); 1140 1141 hfsc_purge_queue(sch, cl); 1142 qdisc_class_hash_remove(&q->clhash, &cl->cl_common); 1143 1144 BUG_ON(--cl->refcnt == 0); 1145 /* 1146 * This shouldn't happen: we "hold" one cops->get() when called 1147 * from tc_ctl_tclass; the destroy method is done from cops->put(). 1148 */ 1149 1150 sch_tree_unlock(sch); 1151 return 0; 1152 } 1153 1154 static struct hfsc_class * 1155 hfsc_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr) 1156 { 1157 struct hfsc_sched *q = qdisc_priv(sch); 1158 struct hfsc_class *cl; 1159 struct tcf_result res; 1160 struct tcf_proto *tcf; 1161 int result; 1162 1163 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0 && 1164 (cl = hfsc_find_class(skb->priority, sch)) != NULL) 1165 if (cl->level == 0) 1166 return cl; 1167 1168 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 1169 tcf = q->root.filter_list; 1170 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) { 1171 #ifdef CONFIG_NET_CLS_ACT 1172 switch (result) { 1173 case TC_ACT_QUEUED: 1174 case TC_ACT_STOLEN: 1175 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; 1176 case TC_ACT_SHOT: 1177 return NULL; 1178 } 1179 #endif 1180 if ((cl = (struct hfsc_class *)res.class) == NULL) { 1181 if ((cl = hfsc_find_class(res.classid, sch)) == NULL) 1182 break; /* filter selected invalid classid */ 1183 } 1184 1185 if (cl->level == 0) 1186 return cl; /* hit leaf class */ 1187 1188 /* apply inner filter chain */ 1189 tcf = cl->filter_list; 1190 } 1191 1192 /* classification failed, try default class */ 1193 cl = hfsc_find_class(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch); 1194 if (cl == NULL || cl->level > 0) 1195 return NULL; 1196 1197 return cl; 1198 } 1199 1200 static int 1201 hfsc_graft_class(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 1202 struct Qdisc **old) 1203 { 1204 struct hfsc_class *cl = (struct hfsc_class *)arg; 1205 1206 if (cl->level > 0) 1207 return -EINVAL; 1208 if (new == NULL) { 1209 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue, 1210 &pfifo_qdisc_ops, 1211 cl->cl_common.classid); 1212 if (new == NULL) 1213 new = &noop_qdisc; 1214 } 1215 1216 sch_tree_lock(sch); 1217 hfsc_purge_queue(sch, cl); 1218 *old = cl->qdisc; 1219 cl->qdisc = new; 1220 sch_tree_unlock(sch); 1221 return 0; 1222 } 1223 1224 static struct Qdisc * 1225 hfsc_class_leaf(struct Qdisc *sch, unsigned long arg) 1226 { 1227 struct hfsc_class *cl = (struct hfsc_class *)arg; 1228 1229 if (cl->level == 0) 1230 return cl->qdisc; 1231 1232 return NULL; 1233 } 1234 1235 static void 1236 hfsc_qlen_notify(struct Qdisc *sch, unsigned long arg) 1237 { 1238 struct hfsc_class *cl = (struct hfsc_class *)arg; 1239 1240 if (cl->qdisc->q.qlen == 0) { 1241 update_vf(cl, 0, 0); 1242 set_passive(cl); 1243 } 1244 } 1245 1246 static unsigned long 1247 hfsc_get_class(struct Qdisc *sch, u32 classid) 1248 { 1249 struct hfsc_class *cl = hfsc_find_class(classid, sch); 1250 1251 if (cl != NULL) 1252 cl->refcnt++; 1253 1254 return (unsigned long)cl; 1255 } 1256 1257 static void 1258 hfsc_put_class(struct Qdisc *sch, unsigned long arg) 1259 { 1260 struct hfsc_class *cl = (struct hfsc_class *)arg; 1261 1262 if (--cl->refcnt == 0) 1263 hfsc_destroy_class(sch, cl); 1264 } 1265 1266 static unsigned long 1267 hfsc_bind_tcf(struct Qdisc *sch, unsigned long parent, u32 classid) 1268 { 1269 struct hfsc_class *p = (struct hfsc_class *)parent; 1270 struct hfsc_class *cl = hfsc_find_class(classid, sch); 1271 1272 if (cl != NULL) { 1273 if (p != NULL && p->level <= cl->level) 1274 return 0; 1275 cl->filter_cnt++; 1276 } 1277 1278 return (unsigned long)cl; 1279 } 1280 1281 static void 1282 hfsc_unbind_tcf(struct Qdisc *sch, unsigned long arg) 1283 { 1284 struct hfsc_class *cl = (struct hfsc_class *)arg; 1285 1286 cl->filter_cnt--; 1287 } 1288 1289 static struct tcf_proto ** 1290 hfsc_tcf_chain(struct Qdisc *sch, unsigned long arg) 1291 { 1292 struct hfsc_sched *q = qdisc_priv(sch); 1293 struct hfsc_class *cl = (struct hfsc_class *)arg; 1294 1295 if (cl == NULL) 1296 cl = &q->root; 1297 1298 return &cl->filter_list; 1299 } 1300 1301 static int 1302 hfsc_dump_sc(struct sk_buff *skb, int attr, struct internal_sc *sc) 1303 { 1304 struct tc_service_curve tsc; 1305 1306 tsc.m1 = sm2m(sc->sm1); 1307 tsc.d = dx2d(sc->dx); 1308 tsc.m2 = sm2m(sc->sm2); 1309 NLA_PUT(skb, attr, sizeof(tsc), &tsc); 1310 1311 return skb->len; 1312 1313 nla_put_failure: 1314 return -1; 1315 } 1316 1317 static inline int 1318 hfsc_dump_curves(struct sk_buff *skb, struct hfsc_class *cl) 1319 { 1320 if ((cl->cl_flags & HFSC_RSC) && 1321 (hfsc_dump_sc(skb, TCA_HFSC_RSC, &cl->cl_rsc) < 0)) 1322 goto nla_put_failure; 1323 1324 if ((cl->cl_flags & HFSC_FSC) && 1325 (hfsc_dump_sc(skb, TCA_HFSC_FSC, &cl->cl_fsc) < 0)) 1326 goto nla_put_failure; 1327 1328 if ((cl->cl_flags & HFSC_USC) && 1329 (hfsc_dump_sc(skb, TCA_HFSC_USC, &cl->cl_usc) < 0)) 1330 goto nla_put_failure; 1331 1332 return skb->len; 1333 1334 nla_put_failure: 1335 return -1; 1336 } 1337 1338 static int 1339 hfsc_dump_class(struct Qdisc *sch, unsigned long arg, struct sk_buff *skb, 1340 struct tcmsg *tcm) 1341 { 1342 struct hfsc_class *cl = (struct hfsc_class *)arg; 1343 struct nlattr *nest; 1344 1345 tcm->tcm_parent = cl->cl_parent ? cl->cl_parent->cl_common.classid : 1346 TC_H_ROOT; 1347 tcm->tcm_handle = cl->cl_common.classid; 1348 if (cl->level == 0) 1349 tcm->tcm_info = cl->qdisc->handle; 1350 1351 nest = nla_nest_start(skb, TCA_OPTIONS); 1352 if (nest == NULL) 1353 goto nla_put_failure; 1354 if (hfsc_dump_curves(skb, cl) < 0) 1355 goto nla_put_failure; 1356 nla_nest_end(skb, nest); 1357 return skb->len; 1358 1359 nla_put_failure: 1360 nla_nest_cancel(skb, nest); 1361 return -EMSGSIZE; 1362 } 1363 1364 static int 1365 hfsc_dump_class_stats(struct Qdisc *sch, unsigned long arg, 1366 struct gnet_dump *d) 1367 { 1368 struct hfsc_class *cl = (struct hfsc_class *)arg; 1369 struct tc_hfsc_stats xstats; 1370 1371 cl->qstats.qlen = cl->qdisc->q.qlen; 1372 xstats.level = cl->level; 1373 xstats.period = cl->cl_vtperiod; 1374 xstats.work = cl->cl_total; 1375 xstats.rtwork = cl->cl_cumul; 1376 1377 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 || 1378 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 || 1379 gnet_stats_copy_queue(d, &cl->qstats) < 0) 1380 return -1; 1381 1382 return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); 1383 } 1384 1385 1386 1387 static void 1388 hfsc_walk(struct Qdisc *sch, struct qdisc_walker *arg) 1389 { 1390 struct hfsc_sched *q = qdisc_priv(sch); 1391 struct hlist_node *n; 1392 struct hfsc_class *cl; 1393 unsigned int i; 1394 1395 if (arg->stop) 1396 return; 1397 1398 for (i = 0; i < q->clhash.hashsize; i++) { 1399 hlist_for_each_entry(cl, n, &q->clhash.hash[i], 1400 cl_common.hnode) { 1401 if (arg->count < arg->skip) { 1402 arg->count++; 1403 continue; 1404 } 1405 if (arg->fn(sch, (unsigned long)cl, arg) < 0) { 1406 arg->stop = 1; 1407 return; 1408 } 1409 arg->count++; 1410 } 1411 } 1412 } 1413 1414 static void 1415 hfsc_schedule_watchdog(struct Qdisc *sch) 1416 { 1417 struct hfsc_sched *q = qdisc_priv(sch); 1418 struct hfsc_class *cl; 1419 u64 next_time = 0; 1420 1421 if ((cl = eltree_get_minel(q)) != NULL) 1422 next_time = cl->cl_e; 1423 if (q->root.cl_cfmin != 0) { 1424 if (next_time == 0 || next_time > q->root.cl_cfmin) 1425 next_time = q->root.cl_cfmin; 1426 } 1427 WARN_ON(next_time == 0); 1428 qdisc_watchdog_schedule(&q->watchdog, next_time); 1429 } 1430 1431 static int 1432 hfsc_init_qdisc(struct Qdisc *sch, struct nlattr *opt) 1433 { 1434 struct hfsc_sched *q = qdisc_priv(sch); 1435 struct tc_hfsc_qopt *qopt; 1436 int err; 1437 1438 if (opt == NULL || nla_len(opt) < sizeof(*qopt)) 1439 return -EINVAL; 1440 qopt = nla_data(opt); 1441 1442 q->defcls = qopt->defcls; 1443 err = qdisc_class_hash_init(&q->clhash); 1444 if (err < 0) 1445 return err; 1446 q->eligible = RB_ROOT; 1447 INIT_LIST_HEAD(&q->droplist); 1448 1449 q->root.cl_common.classid = sch->handle; 1450 q->root.refcnt = 1; 1451 q->root.sched = q; 1452 q->root.qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue, 1453 &pfifo_qdisc_ops, 1454 sch->handle); 1455 if (q->root.qdisc == NULL) 1456 q->root.qdisc = &noop_qdisc; 1457 INIT_LIST_HEAD(&q->root.children); 1458 q->root.vt_tree = RB_ROOT; 1459 q->root.cf_tree = RB_ROOT; 1460 1461 qdisc_class_hash_insert(&q->clhash, &q->root.cl_common); 1462 qdisc_class_hash_grow(sch, &q->clhash); 1463 1464 qdisc_watchdog_init(&q->watchdog, sch); 1465 1466 return 0; 1467 } 1468 1469 static int 1470 hfsc_change_qdisc(struct Qdisc *sch, struct nlattr *opt) 1471 { 1472 struct hfsc_sched *q = qdisc_priv(sch); 1473 struct tc_hfsc_qopt *qopt; 1474 1475 if (opt == NULL || nla_len(opt) < sizeof(*qopt)) 1476 return -EINVAL; 1477 qopt = nla_data(opt); 1478 1479 sch_tree_lock(sch); 1480 q->defcls = qopt->defcls; 1481 sch_tree_unlock(sch); 1482 1483 return 0; 1484 } 1485 1486 static void 1487 hfsc_reset_class(struct hfsc_class *cl) 1488 { 1489 cl->cl_total = 0; 1490 cl->cl_cumul = 0; 1491 cl->cl_d = 0; 1492 cl->cl_e = 0; 1493 cl->cl_vt = 0; 1494 cl->cl_vtadj = 0; 1495 cl->cl_vtoff = 0; 1496 cl->cl_cvtmin = 0; 1497 cl->cl_cvtmax = 0; 1498 cl->cl_cvtoff = 0; 1499 cl->cl_pcvtoff = 0; 1500 cl->cl_vtperiod = 0; 1501 cl->cl_parentperiod = 0; 1502 cl->cl_f = 0; 1503 cl->cl_myf = 0; 1504 cl->cl_myfadj = 0; 1505 cl->cl_cfmin = 0; 1506 cl->cl_nactive = 0; 1507 1508 cl->vt_tree = RB_ROOT; 1509 cl->cf_tree = RB_ROOT; 1510 qdisc_reset(cl->qdisc); 1511 1512 if (cl->cl_flags & HFSC_RSC) 1513 rtsc_init(&cl->cl_deadline, &cl->cl_rsc, 0, 0); 1514 if (cl->cl_flags & HFSC_FSC) 1515 rtsc_init(&cl->cl_virtual, &cl->cl_fsc, 0, 0); 1516 if (cl->cl_flags & HFSC_USC) 1517 rtsc_init(&cl->cl_ulimit, &cl->cl_usc, 0, 0); 1518 } 1519 1520 static void 1521 hfsc_reset_qdisc(struct Qdisc *sch) 1522 { 1523 struct hfsc_sched *q = qdisc_priv(sch); 1524 struct hfsc_class *cl; 1525 struct hlist_node *n; 1526 unsigned int i; 1527 1528 for (i = 0; i < q->clhash.hashsize; i++) { 1529 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode) 1530 hfsc_reset_class(cl); 1531 } 1532 q->eligible = RB_ROOT; 1533 INIT_LIST_HEAD(&q->droplist); 1534 qdisc_watchdog_cancel(&q->watchdog); 1535 sch->q.qlen = 0; 1536 } 1537 1538 static void 1539 hfsc_destroy_qdisc(struct Qdisc *sch) 1540 { 1541 struct hfsc_sched *q = qdisc_priv(sch); 1542 struct hlist_node *n, *next; 1543 struct hfsc_class *cl; 1544 unsigned int i; 1545 1546 for (i = 0; i < q->clhash.hashsize; i++) { 1547 hlist_for_each_entry(cl, n, &q->clhash.hash[i], cl_common.hnode) 1548 tcf_destroy_chain(&cl->filter_list); 1549 } 1550 for (i = 0; i < q->clhash.hashsize; i++) { 1551 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i], 1552 cl_common.hnode) 1553 hfsc_destroy_class(sch, cl); 1554 } 1555 qdisc_class_hash_destroy(&q->clhash); 1556 qdisc_watchdog_cancel(&q->watchdog); 1557 } 1558 1559 static int 1560 hfsc_dump_qdisc(struct Qdisc *sch, struct sk_buff *skb) 1561 { 1562 struct hfsc_sched *q = qdisc_priv(sch); 1563 unsigned char *b = skb_tail_pointer(skb); 1564 struct tc_hfsc_qopt qopt; 1565 1566 qopt.defcls = q->defcls; 1567 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt); 1568 return skb->len; 1569 1570 nla_put_failure: 1571 nlmsg_trim(skb, b); 1572 return -1; 1573 } 1574 1575 static int 1576 hfsc_enqueue(struct sk_buff *skb, struct Qdisc *sch) 1577 { 1578 struct hfsc_class *cl; 1579 int uninitialized_var(err); 1580 1581 cl = hfsc_classify(skb, sch, &err); 1582 if (cl == NULL) { 1583 if (err & __NET_XMIT_BYPASS) 1584 sch->qstats.drops++; 1585 kfree_skb(skb); 1586 return err; 1587 } 1588 1589 err = qdisc_enqueue(skb, cl->qdisc); 1590 if (unlikely(err != NET_XMIT_SUCCESS)) { 1591 if (net_xmit_drop_count(err)) { 1592 cl->qstats.drops++; 1593 sch->qstats.drops++; 1594 } 1595 return err; 1596 } 1597 1598 if (cl->qdisc->q.qlen == 1) 1599 set_active(cl, qdisc_pkt_len(skb)); 1600 1601 cl->bstats.packets++; 1602 cl->bstats.bytes += qdisc_pkt_len(skb); 1603 sch->bstats.packets++; 1604 sch->bstats.bytes += qdisc_pkt_len(skb); 1605 sch->q.qlen++; 1606 1607 return NET_XMIT_SUCCESS; 1608 } 1609 1610 static struct sk_buff * 1611 hfsc_dequeue(struct Qdisc *sch) 1612 { 1613 struct hfsc_sched *q = qdisc_priv(sch); 1614 struct hfsc_class *cl; 1615 struct sk_buff *skb; 1616 u64 cur_time; 1617 unsigned int next_len; 1618 int realtime = 0; 1619 1620 if (sch->q.qlen == 0) 1621 return NULL; 1622 1623 cur_time = psched_get_time(); 1624 1625 /* 1626 * if there are eligible classes, use real-time criteria. 1627 * find the class with the minimum deadline among 1628 * the eligible classes. 1629 */ 1630 if ((cl = eltree_get_mindl(q, cur_time)) != NULL) { 1631 realtime = 1; 1632 } else { 1633 /* 1634 * use link-sharing criteria 1635 * get the class with the minimum vt in the hierarchy 1636 */ 1637 cl = vttree_get_minvt(&q->root, cur_time); 1638 if (cl == NULL) { 1639 sch->qstats.overlimits++; 1640 hfsc_schedule_watchdog(sch); 1641 return NULL; 1642 } 1643 } 1644 1645 skb = qdisc_dequeue_peeked(cl->qdisc); 1646 if (skb == NULL) { 1647 qdisc_warn_nonwc("HFSC", cl->qdisc); 1648 return NULL; 1649 } 1650 1651 update_vf(cl, qdisc_pkt_len(skb), cur_time); 1652 if (realtime) 1653 cl->cl_cumul += qdisc_pkt_len(skb); 1654 1655 if (cl->qdisc->q.qlen != 0) { 1656 if (cl->cl_flags & HFSC_RSC) { 1657 /* update ed */ 1658 next_len = qdisc_peek_len(cl->qdisc); 1659 if (realtime) 1660 update_ed(cl, next_len); 1661 else 1662 update_d(cl, next_len); 1663 } 1664 } else { 1665 /* the class becomes passive */ 1666 set_passive(cl); 1667 } 1668 1669 sch->flags &= ~TCQ_F_THROTTLED; 1670 sch->q.qlen--; 1671 1672 return skb; 1673 } 1674 1675 static unsigned int 1676 hfsc_drop(struct Qdisc *sch) 1677 { 1678 struct hfsc_sched *q = qdisc_priv(sch); 1679 struct hfsc_class *cl; 1680 unsigned int len; 1681 1682 list_for_each_entry(cl, &q->droplist, dlist) { 1683 if (cl->qdisc->ops->drop != NULL && 1684 (len = cl->qdisc->ops->drop(cl->qdisc)) > 0) { 1685 if (cl->qdisc->q.qlen == 0) { 1686 update_vf(cl, 0, 0); 1687 set_passive(cl); 1688 } else { 1689 list_move_tail(&cl->dlist, &q->droplist); 1690 } 1691 cl->qstats.drops++; 1692 sch->qstats.drops++; 1693 sch->q.qlen--; 1694 return len; 1695 } 1696 } 1697 return 0; 1698 } 1699 1700 static const struct Qdisc_class_ops hfsc_class_ops = { 1701 .change = hfsc_change_class, 1702 .delete = hfsc_delete_class, 1703 .graft = hfsc_graft_class, 1704 .leaf = hfsc_class_leaf, 1705 .qlen_notify = hfsc_qlen_notify, 1706 .get = hfsc_get_class, 1707 .put = hfsc_put_class, 1708 .bind_tcf = hfsc_bind_tcf, 1709 .unbind_tcf = hfsc_unbind_tcf, 1710 .tcf_chain = hfsc_tcf_chain, 1711 .dump = hfsc_dump_class, 1712 .dump_stats = hfsc_dump_class_stats, 1713 .walk = hfsc_walk 1714 }; 1715 1716 static struct Qdisc_ops hfsc_qdisc_ops __read_mostly = { 1717 .id = "hfsc", 1718 .init = hfsc_init_qdisc, 1719 .change = hfsc_change_qdisc, 1720 .reset = hfsc_reset_qdisc, 1721 .destroy = hfsc_destroy_qdisc, 1722 .dump = hfsc_dump_qdisc, 1723 .enqueue = hfsc_enqueue, 1724 .dequeue = hfsc_dequeue, 1725 .peek = qdisc_peek_dequeued, 1726 .drop = hfsc_drop, 1727 .cl_ops = &hfsc_class_ops, 1728 .priv_size = sizeof(struct hfsc_sched), 1729 .owner = THIS_MODULE 1730 }; 1731 1732 static int __init 1733 hfsc_init(void) 1734 { 1735 return register_qdisc(&hfsc_qdisc_ops); 1736 } 1737 1738 static void __exit 1739 hfsc_cleanup(void) 1740 { 1741 unregister_qdisc(&hfsc_qdisc_ops); 1742 } 1743 1744 MODULE_LICENSE("GPL"); 1745 module_init(hfsc_init); 1746 module_exit(hfsc_cleanup); 1747