1 /* @(#)rm_class.c 1.48 97/12/05 SMI */ 2 /* $KAME: altq_rmclass.c,v 1.18 2003/11/06 06:32:53 kjc Exp $ */ 3 4 /* 5 * Copyright (c) 1991-1997 Regents of the University of California. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the Network Research 19 * Group at Lawrence Berkeley Laboratory. 20 * 4. Neither the name of the University nor of the Laboratory may be used 21 * to endorse or promote products derived from this software without 22 * specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * LBL code modified by speer@eng.sun.com, May 1977. 37 * For questions and/or comments, please send mail to cbq@ee.lbl.gov 38 */ 39 40 #include "opt_altq.h" 41 #include "opt_inet.h" 42 #include "opt_inet6.h" 43 44 #ifdef ALTQ_CBQ /* cbq is enabled by ALTQ_CBQ option in opt_altq.h */ 45 46 #include <sys/param.h> 47 #include <sys/malloc.h> 48 #include <sys/mbuf.h> 49 #include <sys/socket.h> 50 #include <sys/systm.h> 51 #include <sys/callout.h> 52 #include <sys/errno.h> 53 #include <sys/time.h> 54 #include <sys/thread.h> 55 56 #include <net/if.h> 57 58 #include <net/altq/altq.h> 59 #include <net/altq/altq_rmclass.h> 60 #include <net/altq/altq_rmclass_debug.h> 61 #include <net/altq/altq_red.h> 62 #include <net/altq/altq_rio.h> 63 64 #include <sys/thread2.h> 65 66 #ifdef CBQ_TRACE 67 static struct cbqtrace cbqtrace_buffer[NCBQTRACE+1]; 68 static struct cbqtrace *cbqtrace_ptr = NULL; 69 static int cbqtrace_count; 70 #endif 71 72 /* 73 * Local Macros 74 */ 75 76 #define reset_cutoff(ifd) { ifd->cutoff_ = RM_MAXDEPTH; } 77 78 /* 79 * Local routines. 80 */ 81 82 static int rmc_satisfied(struct rm_class *, struct timeval *); 83 static void rmc_wrr_set_weights(struct rm_ifdat *); 84 static void rmc_depth_compute(struct rm_class *); 85 static void rmc_depth_recompute(rm_class_t *); 86 87 static struct mbuf *_rmc_wrr_dequeue_next(struct rm_ifdat *, int); 88 static struct mbuf *_rmc_prr_dequeue_next(struct rm_ifdat *, int); 89 90 static int _rmc_addq(rm_class_t *, struct mbuf *); 91 static void _rmc_dropq(rm_class_t *); 92 static struct mbuf *_rmc_getq(rm_class_t *); 93 static struct mbuf *_rmc_pollq(rm_class_t *); 94 95 static int rmc_under_limit(struct rm_class *, struct timeval *); 96 static void rmc_tl_satisfied(struct rm_ifdat *, struct timeval *); 97 static void rmc_drop_action(struct rm_class *); 98 static void rmc_restart(void *); 99 static void rmc_root_overlimit(struct rm_class *, struct rm_class *); 100 101 #define BORROW_OFFTIME 102 /* 103 * BORROW_OFFTIME (experimental): 104 * borrow the offtime of the class borrowing from. 105 * the reason is that when its own offtime is set, the class is unable 106 * to borrow much, especially when cutoff is taking effect. 107 * but when the borrowed class is overloaded (advidle is close to minidle), 108 * use the borrowing class's offtime to avoid overload. 109 */ 110 #define ADJUST_CUTOFF 111 /* 112 * ADJUST_CUTOFF (experimental): 113 * if no underlimit class is found due to cutoff, increase cutoff and 114 * retry the scheduling loop. 115 * also, don't invoke delay_actions while cutoff is taking effect, 116 * since a sleeping class won't have a chance to be scheduled in the 117 * next loop. 118 * 119 * now heuristics for setting the top-level variable (cutoff_) becomes: 120 * 1. if a packet arrives for a not-overlimit class, set cutoff 121 * to the depth of the class. 122 * 2. if cutoff is i, and a packet arrives for an overlimit class 123 * with an underlimit ancestor at a lower level than i (say j), 124 * then set cutoff to j. 125 * 3. at scheduling a packet, if there is no underlimit class 126 * due to the current cutoff level, increase cutoff by 1 and 127 * then try to schedule again. 128 */ 129 130 /* 131 * rm_class_t * 132 * rmc_newclass(...) - Create a new resource management class at priority 133 * 'pri' on the interface given by 'ifd'. 134 * 135 * nsecPerByte is the data rate of the interface in nanoseconds/byte. 136 * E.g., 800 for a 10Mb/s ethernet. If the class gets less 137 * than 100% of the bandwidth, this number should be the 138 * 'effective' rate for the class. Let f be the 139 * bandwidth fraction allocated to this class, and let 140 * nsPerByte be the data rate of the output link in 141 * nanoseconds/byte. Then nsecPerByte is set to 142 * nsPerByte / f. E.g., 1600 (= 800 / .5) 143 * for a class that gets 50% of an ethernet's bandwidth. 144 * 145 * action the routine to call when the class is over limit. 146 * 147 * maxq max allowable queue size for class (in packets). 148 * 149 * parent parent class pointer. 150 * 151 * borrow class to borrow from (should be either 'parent' or null). 152 * 153 * maxidle max value allowed for class 'idle' time estimate (this 154 * parameter determines how large an initial burst of packets 155 * can be before overlimit action is invoked. 156 * 157 * offtime how long 'delay' action will delay when class goes over 158 * limit (this parameter determines the steady-state burst 159 * size when a class is running over its limit). 160 * 161 * Maxidle and offtime have to be computed from the following: If the 162 * average packet size is s, the bandwidth fraction allocated to this 163 * class is f, we want to allow b packet bursts, and the gain of the 164 * averaging filter is g (= 1 - 2^(-RM_FILTER_GAIN)), then: 165 * 166 * ptime = s * nsPerByte * (1 - f) / f 167 * maxidle = ptime * (1 - g^b) / g^b 168 * minidle = -ptime * (1 / (f - 1)) 169 * offtime = ptime * (1 + 1/(1 - g) * (1 - g^(b - 1)) / g^(b - 1) 170 * 171 * Operationally, it's convenient to specify maxidle & offtime in units 172 * independent of the link bandwidth so the maxidle & offtime passed to 173 * this routine are the above values multiplied by 8*f/(1000*nsPerByte). 174 * (The constant factor is a scale factor needed to make the parameters 175 * integers. This scaling also means that the 'unscaled' values of 176 * maxidle*nsecPerByte/8 and offtime*nsecPerByte/8 will be in microseconds, 177 * not nanoseconds.) Also note that the 'idle' filter computation keeps 178 * an estimate scaled upward by 2^RM_FILTER_GAIN so the passed value of 179 * maxidle also must be scaled upward by this value. Thus, the passed 180 * values for maxidle and offtime can be computed as follows: 181 * 182 * maxidle = maxidle * 2^RM_FILTER_GAIN * 8 / (1000 * nsecPerByte) 183 * offtime = offtime * 8 / (1000 * nsecPerByte) 184 * 185 * When USE_HRTIME is employed, then maxidle and offtime become: 186 * maxidle = maxilde * (8.0 / nsecPerByte); 187 * offtime = offtime * (8.0 / nsecPerByte); 188 */ 189 struct rm_class * 190 rmc_newclass(int pri, struct rm_ifdat *ifd, u_int nsecPerByte, 191 void (*action)(rm_class_t *, rm_class_t *), int maxq, 192 struct rm_class *parent, struct rm_class *borrow, u_int maxidle, 193 int minidle, u_int offtime, int pktsize, int flags) 194 { 195 struct rm_class *cl; 196 struct rm_class *peer; 197 198 if (pri >= RM_MAXPRIO) 199 return (NULL); 200 #ifndef ALTQ_RED 201 if (flags & RMCF_RED) { 202 #ifdef ALTQ_DEBUG 203 kprintf("rmc_newclass: RED not configured for CBQ!\n"); 204 #endif 205 return (NULL); 206 } 207 #endif 208 #ifndef ALTQ_RIO 209 if (flags & RMCF_RIO) { 210 #ifdef ALTQ_DEBUG 211 kprintf("rmc_newclass: RIO not configured for CBQ!\n"); 212 #endif 213 return (NULL); 214 } 215 #endif 216 217 cl = kmalloc(sizeof(*cl), M_ALTQ, M_WAITOK | M_ZERO); 218 callout_init(&cl->callout_); 219 cl->q_ = kmalloc(sizeof(*cl->q_), M_ALTQ, M_WAITOK | M_ZERO); 220 221 /* 222 * Class initialization. 223 */ 224 cl->children_ = NULL; 225 cl->parent_ = parent; 226 cl->borrow_ = borrow; 227 cl->leaf_ = 1; 228 cl->ifdat_ = ifd; 229 cl->pri_ = pri; 230 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */ 231 cl->depth_ = 0; 232 cl->qthresh_ = 0; 233 cl->ns_per_byte_ = nsecPerByte; 234 235 qlimit(cl->q_) = maxq; 236 qtype(cl->q_) = Q_DROPHEAD; 237 qlen(cl->q_) = 0; 238 cl->flags_ = flags; 239 240 #if 1 /* minidle is also scaled in ALTQ */ 241 cl->minidle_ = (minidle * (int)nsecPerByte) / 8; 242 if (cl->minidle_ > 0) 243 cl->minidle_ = 0; 244 #else 245 cl->minidle_ = minidle; 246 #endif 247 cl->maxidle_ = (maxidle * nsecPerByte) / 8; 248 if (cl->maxidle_ == 0) 249 cl->maxidle_ = 1; 250 #if 1 /* offtime is also scaled in ALTQ */ 251 cl->avgidle_ = cl->maxidle_; 252 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN; 253 if (cl->offtime_ == 0) 254 cl->offtime_ = 1; 255 #else 256 cl->avgidle_ = 0; 257 cl->offtime_ = (offtime * nsecPerByte) / 8; 258 #endif 259 cl->overlimit = action; 260 261 #ifdef ALTQ_RED 262 if (flags & (RMCF_RED|RMCF_RIO)) { 263 int red_flags, red_pkttime; 264 265 red_flags = 0; 266 if (flags & RMCF_ECN) 267 red_flags |= REDF_ECN; 268 #ifdef ALTQ_RIO 269 if (flags & RMCF_CLEARDSCP) 270 red_flags |= RIOF_CLEARDSCP; 271 #endif 272 red_pkttime = nsecPerByte * pktsize / 1000; 273 274 if (flags & RMCF_RED) { 275 cl->red_ = red_alloc(0, 0, 276 qlimit(cl->q_) * 10/100, 277 qlimit(cl->q_) * 30/100, 278 red_flags, red_pkttime); 279 if (cl->red_ != NULL) 280 qtype(cl->q_) = Q_RED; 281 } 282 #ifdef ALTQ_RIO 283 else { 284 cl->red_ = (red_t *)rio_alloc(0, NULL, 285 red_flags, red_pkttime); 286 if (cl->red_ != NULL) 287 qtype(cl->q_) = Q_RIO; 288 } 289 #endif 290 } 291 #endif /* ALTQ_RED */ 292 293 /* 294 * put the class into the class tree 295 */ 296 crit_enter(); 297 if ((peer = ifd->active_[pri]) != NULL) { 298 /* find the last class at this pri */ 299 cl->peer_ = peer; 300 while (peer->peer_ != ifd->active_[pri]) 301 peer = peer->peer_; 302 peer->peer_ = cl; 303 } else { 304 ifd->active_[pri] = cl; 305 cl->peer_ = cl; 306 } 307 308 if (cl->parent_) { 309 cl->next_ = parent->children_; 310 parent->children_ = cl; 311 parent->leaf_ = 0; 312 } 313 314 /* 315 * Compute the depth of this class and its ancestors in the class 316 * hierarchy. 317 */ 318 rmc_depth_compute(cl); 319 320 /* 321 * If CBQ's WRR is enabled, then initialize the class WRR state. 322 */ 323 if (ifd->wrr_) { 324 ifd->num_[pri]++; 325 ifd->alloc_[pri] += cl->allotment_; 326 rmc_wrr_set_weights(ifd); 327 } 328 crit_exit(); 329 return (cl); 330 } 331 332 int 333 rmc_modclass(struct rm_class *cl, u_int nsecPerByte, int maxq, u_int maxidle, 334 int minidle, u_int offtime, int pktsize) 335 { 336 struct rm_ifdat *ifd; 337 u_int old_allotment; 338 339 ifd = cl->ifdat_; 340 old_allotment = cl->allotment_; 341 342 crit_enter(); 343 cl->allotment_ = RM_NS_PER_SEC / nsecPerByte; /* Bytes per sec */ 344 cl->qthresh_ = 0; 345 cl->ns_per_byte_ = nsecPerByte; 346 347 qlimit(cl->q_) = maxq; 348 349 #if 1 /* minidle is also scaled in ALTQ */ 350 cl->minidle_ = (minidle * nsecPerByte) / 8; 351 if (cl->minidle_ > 0) 352 cl->minidle_ = 0; 353 #else 354 cl->minidle_ = minidle; 355 #endif 356 cl->maxidle_ = (maxidle * nsecPerByte) / 8; 357 if (cl->maxidle_ == 0) 358 cl->maxidle_ = 1; 359 #if 1 /* offtime is also scaled in ALTQ */ 360 cl->avgidle_ = cl->maxidle_; 361 cl->offtime_ = ((offtime * nsecPerByte) / 8) >> RM_FILTER_GAIN; 362 if (cl->offtime_ == 0) 363 cl->offtime_ = 1; 364 #else 365 cl->avgidle_ = 0; 366 cl->offtime_ = (offtime * nsecPerByte) / 8; 367 #endif 368 369 /* 370 * If CBQ's WRR is enabled, then initialize the class WRR state. 371 */ 372 if (ifd->wrr_) { 373 ifd->alloc_[cl->pri_] += cl->allotment_ - old_allotment; 374 rmc_wrr_set_weights(ifd); 375 } 376 crit_exit(); 377 return (0); 378 } 379 380 /* 381 * static void 382 * rmc_wrr_set_weights(struct rm_ifdat *ifdat) - This function computes 383 * the appropriate run robin weights for the CBQ weighted round robin 384 * algorithm. 385 * 386 * Returns: NONE 387 */ 388 389 static void 390 rmc_wrr_set_weights(struct rm_ifdat *ifd) 391 { 392 int i; 393 struct rm_class *cl, *clh; 394 395 for (i = 0; i < RM_MAXPRIO; i++) { 396 /* 397 * This is inverted from that of the simulator to 398 * maintain precision. 399 */ 400 if (ifd->num_[i] == 0) 401 ifd->M_[i] = 0; 402 else 403 ifd->M_[i] = ifd->alloc_[i] / 404 (ifd->num_[i] * ifd->maxpkt_); 405 /* 406 * Compute the weighted allotment for each class. 407 * This takes the expensive div instruction out 408 * of the main loop for the wrr scheduling path. 409 * These only get recomputed when a class comes or 410 * goes. 411 */ 412 if (ifd->active_[i] != NULL) { 413 clh = cl = ifd->active_[i]; 414 do { 415 /* safe-guard for slow link or alloc_ == 0 */ 416 if (ifd->M_[i] == 0) 417 cl->w_allotment_ = 0; 418 else 419 cl->w_allotment_ = cl->allotment_ / 420 ifd->M_[i]; 421 cl = cl->peer_; 422 } while ((cl != NULL) && (cl != clh)); 423 } 424 } 425 } 426 427 int 428 rmc_get_weight(struct rm_ifdat *ifd, int pri) 429 { 430 if ((pri >= 0) && (pri < RM_MAXPRIO)) 431 return (ifd->M_[pri]); 432 else 433 return (0); 434 } 435 436 /* 437 * static void 438 * rmc_depth_compute(struct rm_class *cl) - This function computes the 439 * appropriate depth of class 'cl' and its ancestors. 440 * 441 * Returns: NONE 442 */ 443 444 static void 445 rmc_depth_compute(struct rm_class *cl) 446 { 447 rm_class_t *t = cl, *p; 448 449 /* 450 * Recompute the depth for the branch of the tree. 451 */ 452 while (t != NULL) { 453 p = t->parent_; 454 if (p && (t->depth_ >= p->depth_)) { 455 p->depth_ = t->depth_ + 1; 456 t = p; 457 } else 458 t = NULL; 459 } 460 } 461 462 /* 463 * static void 464 * rmc_depth_recompute(struct rm_class *cl) - This function re-computes 465 * the depth of the tree after a class has been deleted. 466 * 467 * Returns: NONE 468 */ 469 470 static void 471 rmc_depth_recompute(rm_class_t *cl) 472 { 473 #if 1 /* ALTQ */ 474 rm_class_t *p, *t; 475 476 p = cl; 477 while (p != NULL) { 478 if ((t = p->children_) == NULL) { 479 p->depth_ = 0; 480 } else { 481 int cdepth = 0; 482 483 while (t != NULL) { 484 if (t->depth_ > cdepth) 485 cdepth = t->depth_; 486 t = t->next_; 487 } 488 489 if (p->depth_ == cdepth + 1) 490 /* no change to this parent */ 491 return; 492 493 p->depth_ = cdepth + 1; 494 } 495 496 p = p->parent_; 497 } 498 #else 499 rm_class_t *t; 500 501 if (cl->depth_ >= 1) { 502 if (cl->children_ == NULL) { 503 cl->depth_ = 0; 504 } else if ((t = cl->children_) != NULL) { 505 while (t != NULL) { 506 if (t->children_ != NULL) 507 rmc_depth_recompute(t); 508 t = t->next_; 509 } 510 } else 511 rmc_depth_compute(cl); 512 } 513 #endif 514 } 515 516 /* 517 * void 518 * rmc_delete_class(struct rm_ifdat *ifdat, struct rm_class *cl) - This 519 * function deletes a class from the link-sharing structure and frees 520 * all resources associated with the class. 521 * 522 * Returns: NONE 523 */ 524 525 void 526 rmc_delete_class(struct rm_ifdat *ifd, struct rm_class *cl) 527 { 528 struct rm_class *p, *head, *previous; 529 530 KKASSERT(cl->children_ == NULL); 531 532 if (cl->sleeping_) 533 callout_stop(&cl->callout_); 534 535 crit_enter(); 536 /* 537 * Free packets in the packet queue. 538 * XXX - this may not be a desired behavior. Packets should be 539 * re-queued. 540 */ 541 rmc_dropall(cl); 542 543 /* 544 * If the class has a parent, then remove the class from the 545 * class from the parent's children chain. 546 */ 547 if (cl->parent_ != NULL) { 548 head = cl->parent_->children_; 549 p = previous = head; 550 if (head->next_ == NULL) { 551 KKASSERT(head == cl); 552 cl->parent_->children_ = NULL; 553 cl->parent_->leaf_ = 1; 554 } else while (p != NULL) { 555 if (p == cl) { 556 if (cl == head) 557 cl->parent_->children_ = cl->next_; 558 else 559 previous->next_ = cl->next_; 560 cl->next_ = NULL; 561 p = NULL; 562 } else { 563 previous = p; 564 p = p->next_; 565 } 566 } 567 } 568 569 /* 570 * Delete class from class priority peer list. 571 */ 572 if ((p = ifd->active_[cl->pri_]) != NULL) { 573 /* 574 * If there is more than one member of this priority 575 * level, then look for class(cl) in the priority level. 576 */ 577 if (p != p->peer_) { 578 while (p->peer_ != cl) 579 p = p->peer_; 580 p->peer_ = cl->peer_; 581 582 if (ifd->active_[cl->pri_] == cl) 583 ifd->active_[cl->pri_] = cl->peer_; 584 } else { 585 KKASSERT(p == cl); 586 ifd->active_[cl->pri_] = NULL; 587 } 588 } 589 590 /* 591 * Recompute the WRR weights. 592 */ 593 if (ifd->wrr_) { 594 ifd->alloc_[cl->pri_] -= cl->allotment_; 595 ifd->num_[cl->pri_]--; 596 rmc_wrr_set_weights(ifd); 597 } 598 599 /* 600 * Re-compute the depth of the tree. 601 */ 602 #if 1 /* ALTQ */ 603 rmc_depth_recompute(cl->parent_); 604 #else 605 rmc_depth_recompute(ifd->root_); 606 #endif 607 608 crit_exit(); 609 610 /* 611 * Free the class structure. 612 */ 613 if (cl->red_ != NULL) { 614 #ifdef ALTQ_RIO 615 if (q_is_rio(cl->q_)) 616 rio_destroy((rio_t *)cl->red_); 617 #endif 618 #ifdef ALTQ_RED 619 if (q_is_red(cl->q_)) 620 red_destroy(cl->red_); 621 #endif 622 } 623 kfree(cl->q_, M_ALTQ); 624 kfree(cl, M_ALTQ); 625 } 626 627 /* 628 * void 629 * rmc_init(...) - Initialize the resource management data structures 630 * associated with the output portion of interface 'ifp'. 'ifd' is 631 * where the structures will be built (for backwards compatibility, the 632 * structures aren't kept in the ifnet struct). 'nsecPerByte' 633 * gives the link speed (inverse of bandwidth) in nanoseconds/byte. 634 * 'restart' is the driver-specific routine that the generic 'delay 635 * until under limit' action will call to restart output. `maxq' 636 * is the queue size of the 'link' & 'default' classes. 'maxqueued' 637 * is the maximum number of packets that the resource management 638 * code will allow to be queued 'downstream' (this is typically 1). 639 * 640 * Returns: NONE 641 */ 642 643 void 644 rmc_init(struct ifaltq *ifq, struct rm_ifdat *ifd, u_int nsecPerByte, 645 void (*restart)(struct ifaltq *), int maxq, int maxqueued, u_int maxidle, 646 int minidle, u_int offtime, int flags) 647 { 648 int i, mtu; 649 650 /* 651 * Initialize the CBQ tracing/debug facility. 652 */ 653 CBQTRACEINIT(); 654 655 bzero(ifd, sizeof (*ifd)); 656 mtu = ifq->altq_ifp->if_mtu; 657 ifd->ifq_ = ifq; 658 ifd->restart = restart; 659 ifd->maxqueued_ = maxqueued; 660 ifd->ns_per_byte_ = nsecPerByte; 661 ifd->maxpkt_ = mtu; 662 ifd->wrr_ = (flags & RMCF_WRR) ? 1 : 0; 663 ifd->efficient_ = (flags & RMCF_EFFICIENT) ? 1 : 0; 664 #if 1 665 ifd->maxiftime_ = mtu * nsecPerByte / 1000 * 16; 666 if (mtu * nsecPerByte > 10 * 1000000) 667 ifd->maxiftime_ /= 4; 668 #endif 669 670 reset_cutoff(ifd); 671 CBQTRACE(rmc_init, 'INIT', ifd->cutoff_); 672 673 /* 674 * Initialize the CBQ's WRR state. 675 */ 676 for (i = 0; i < RM_MAXPRIO; i++) { 677 ifd->alloc_[i] = 0; 678 ifd->M_[i] = 0; 679 ifd->num_[i] = 0; 680 ifd->na_[i] = 0; 681 ifd->active_[i] = NULL; 682 } 683 684 /* 685 * Initialize current packet state. 686 */ 687 ifd->qi_ = 0; 688 ifd->qo_ = 0; 689 for (i = 0; i < RM_MAXQUEUED; i++) { 690 ifd->class_[i] = NULL; 691 ifd->curlen_[i] = 0; 692 ifd->borrowed_[i] = NULL; 693 } 694 695 /* 696 * Create the root class of the link-sharing structure. 697 */ 698 ifd->root_ = rmc_newclass(0, ifd, nsecPerByte, rmc_root_overlimit, 699 maxq, 0, 0, maxidle, minidle, offtime, 0, 0); 700 if (ifd->root_ == NULL) { 701 kprintf("rmc_init: root class not allocated\n"); 702 return ; 703 } 704 ifd->root_->depth_ = 0; 705 } 706 707 /* 708 * void 709 * rmc_queue_packet(struct rm_class *cl, struct mbuf *m) - Add packet given by 710 * mbuf 'm' to queue for resource class 'cl'. This routine is called 711 * by a driver's if_output routine. This routine must be called with 712 * output packet completion interrupts locked out (to avoid racing with 713 * rmc_dequeue_next). 714 * 715 * Returns: 0 on successful queueing 716 * -1 when packet drop occurs 717 */ 718 int 719 rmc_queue_packet(struct rm_class *cl, struct mbuf *m) 720 { 721 struct timeval now; 722 struct rm_ifdat *ifd = cl->ifdat_; 723 int cpri = cl->pri_; 724 int is_empty = qempty(cl->q_); 725 726 RM_GETTIME(now); 727 if (ifd->cutoff_ > 0) { 728 if (TV_LT(&cl->undertime_, &now)) { 729 if (ifd->cutoff_ > cl->depth_) 730 ifd->cutoff_ = cl->depth_; 731 CBQTRACE(rmc_queue_packet, 'ffoc', cl->depth_); 732 } 733 #if 1 /* ALTQ */ 734 else { 735 /* 736 * the class is overlimit. if the class has 737 * underlimit ancestors, set cutoff to the lowest 738 * depth among them. 739 */ 740 struct rm_class *borrow = cl->borrow_; 741 742 while (borrow != NULL && 743 borrow->depth_ < ifd->cutoff_) { 744 if (TV_LT(&borrow->undertime_, &now)) { 745 ifd->cutoff_ = borrow->depth_; 746 CBQTRACE(rmc_queue_packet, 'ffob', ifd->cutoff_); 747 break; 748 } 749 borrow = borrow->borrow_; 750 } 751 } 752 #else /* !ALTQ */ 753 else if ((ifd->cutoff_ > 1) && cl->borrow_) { 754 if (TV_LT(&cl->borrow_->undertime_, &now)) { 755 ifd->cutoff_ = cl->borrow_->depth_; 756 CBQTRACE(rmc_queue_packet, 'ffob', 757 cl->borrow_->depth_); 758 } 759 } 760 #endif /* !ALTQ */ 761 } 762 763 if (_rmc_addq(cl, m) < 0) 764 /* failed */ 765 return (-1); 766 767 if (is_empty) { 768 CBQTRACE(rmc_queue_packet, 'ytpe', cl->stats_.handle); 769 ifd->na_[cpri]++; 770 } 771 772 if (qlen(cl->q_) > qlimit(cl->q_)) { 773 /* note: qlimit can be set to 0 or 1 */ 774 rmc_drop_action(cl); 775 return (-1); 776 } 777 return (0); 778 } 779 780 /* 781 * void 782 * rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) - Check all 783 * classes to see if there are satified. 784 */ 785 786 static void 787 rmc_tl_satisfied(struct rm_ifdat *ifd, struct timeval *now) 788 { 789 int i; 790 rm_class_t *p, *bp; 791 792 for (i = RM_MAXPRIO - 1; i >= 0; i--) { 793 if ((bp = ifd->active_[i]) != NULL) { 794 p = bp; 795 do { 796 if (!rmc_satisfied(p, now)) { 797 ifd->cutoff_ = p->depth_; 798 return; 799 } 800 p = p->peer_; 801 } while (p != bp); 802 } 803 } 804 805 reset_cutoff(ifd); 806 } 807 808 /* 809 * rmc_satisfied - Return 1 of the class is satisfied. O, otherwise. 810 */ 811 812 static int 813 rmc_satisfied(struct rm_class *cl, struct timeval *now) 814 { 815 rm_class_t *p; 816 817 if (cl == NULL) 818 return (1); 819 if (TV_LT(now, &cl->undertime_)) 820 return (1); 821 if (cl->depth_ == 0) { 822 if (!cl->sleeping_ && (qlen(cl->q_) > cl->qthresh_)) 823 return (0); 824 else 825 return (1); 826 } 827 if (cl->children_ != NULL) { 828 p = cl->children_; 829 while (p != NULL) { 830 if (!rmc_satisfied(p, now)) 831 return (0); 832 p = p->next_; 833 } 834 } 835 836 return (1); 837 } 838 839 /* 840 * Return 1 if class 'cl' is under limit or can borrow from a parent, 841 * 0 if overlimit. As a side-effect, this routine will invoke the 842 * class overlimit action if the class if overlimit. 843 */ 844 845 static int 846 rmc_under_limit(struct rm_class *cl, struct timeval *now) 847 { 848 rm_class_t *p = cl; 849 rm_class_t *top; 850 struct rm_ifdat *ifd = cl->ifdat_; 851 852 ifd->borrowed_[ifd->qi_] = NULL; 853 /* 854 * If cl is the root class, then always return that it is 855 * underlimit. Otherwise, check to see if the class is underlimit. 856 */ 857 if (cl->parent_ == NULL) 858 return (1); 859 860 if (cl->sleeping_) { 861 if (TV_LT(now, &cl->undertime_)) 862 return (0); 863 864 callout_stop(&cl->callout_); 865 cl->sleeping_ = 0; 866 cl->undertime_.tv_sec = 0; 867 return (1); 868 } 869 870 top = NULL; 871 while (cl->undertime_.tv_sec && TV_LT(now, &cl->undertime_)) { 872 if (((cl = cl->borrow_) == NULL) || 873 (cl->depth_ > ifd->cutoff_)) { 874 #ifdef ADJUST_CUTOFF 875 if (cl != NULL) 876 /* cutoff is taking effect, just 877 return false without calling 878 the delay action. */ 879 return (0); 880 #endif 881 #ifdef BORROW_OFFTIME 882 /* 883 * check if the class can borrow offtime too. 884 * borrow offtime from the top of the borrow 885 * chain if the top class is not overloaded. 886 */ 887 if (cl != NULL) { 888 /* cutoff is taking effect, use this class as top. */ 889 top = cl; 890 CBQTRACE(rmc_under_limit, 'ffou', ifd->cutoff_); 891 } 892 if (top != NULL && top->avgidle_ == top->minidle_) 893 top = NULL; 894 p->overtime_ = *now; 895 (p->overlimit)(p, top); 896 #else 897 p->overtime_ = *now; 898 (p->overlimit)(p, NULL); 899 #endif 900 return (0); 901 } 902 top = cl; 903 } 904 905 if (cl != p) 906 ifd->borrowed_[ifd->qi_] = cl; 907 return (1); 908 } 909 910 /* 911 * _rmc_wrr_dequeue_next() - This is scheduler for WRR as opposed to 912 * Packet-by-packet round robin. 913 * 914 * The heart of the weighted round-robin scheduler, which decides which 915 * class next gets to send a packet. Highest priority first, then 916 * weighted round-robin within priorites. 917 * 918 * Each able-to-send class gets to send until its byte allocation is 919 * exhausted. Thus, the active pointer is only changed after a class has 920 * exhausted its allocation. 921 * 922 * If the scheduler finds no class that is underlimit or able to borrow, 923 * then the first class found that had a nonzero queue and is allowed to 924 * borrow gets to send. 925 */ 926 927 static struct mbuf * 928 _rmc_wrr_dequeue_next(struct rm_ifdat *ifd, int op) 929 { 930 struct rm_class *cl = NULL, *first = NULL; 931 u_int deficit; 932 int cpri; 933 struct mbuf *m; 934 struct timeval now; 935 936 RM_GETTIME(now); 937 938 /* 939 * if the driver polls the top of the queue and then removes 940 * the polled packet, we must return the same packet. 941 */ 942 if (op == ALTDQ_REMOVE && ifd->pollcache_) { 943 cl = ifd->pollcache_; 944 cpri = cl->pri_; 945 if (ifd->efficient_) { 946 /* check if this class is overlimit */ 947 if (cl->undertime_.tv_sec != 0 && 948 rmc_under_limit(cl, &now) == 0) 949 first = cl; 950 } 951 ifd->pollcache_ = NULL; 952 goto _wrr_out; 953 } 954 else { 955 /* mode == ALTDQ_POLL || pollcache == NULL */ 956 ifd->pollcache_ = NULL; 957 ifd->borrowed_[ifd->qi_] = NULL; 958 } 959 #ifdef ADJUST_CUTOFF 960 _again: 961 #endif 962 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) { 963 if (ifd->na_[cpri] == 0) 964 continue; 965 deficit = 0; 966 /* 967 * Loop through twice for a priority level, if some class 968 * was unable to send a packet the first round because 969 * of the weighted round-robin mechanism. 970 * During the second loop at this level, deficit==2. 971 * (This second loop is not needed if for every class, 972 * "M[cl->pri_])" times "cl->allotment" is greater than 973 * the byte size for the largest packet in the class.) 974 */ 975 _wrr_loop: 976 cl = ifd->active_[cpri]; 977 KKASSERT(cl != NULL); 978 do { 979 if ((deficit < 2) && (cl->bytes_alloc_ <= 0)) 980 cl->bytes_alloc_ += cl->w_allotment_; 981 if (!qempty(cl->q_)) { 982 if ((cl->undertime_.tv_sec == 0) || 983 rmc_under_limit(cl, &now)) { 984 if (cl->bytes_alloc_ > 0 || deficit > 1) 985 goto _wrr_out; 986 987 /* underlimit but no alloc */ 988 deficit = 1; 989 #if 1 990 ifd->borrowed_[ifd->qi_] = NULL; 991 #endif 992 } 993 else if (first == NULL && cl->borrow_ != NULL) 994 first = cl; /* borrowing candidate */ 995 } 996 997 cl->bytes_alloc_ = 0; 998 cl = cl->peer_; 999 } while (cl != ifd->active_[cpri]); 1000 1001 if (deficit == 1) { 1002 /* first loop found an underlimit class with deficit */ 1003 /* Loop on same priority level, with new deficit. */ 1004 deficit = 2; 1005 goto _wrr_loop; 1006 } 1007 } 1008 1009 #ifdef ADJUST_CUTOFF 1010 /* 1011 * no underlimit class found. if cutoff is taking effect, 1012 * increase cutoff and try again. 1013 */ 1014 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) { 1015 ifd->cutoff_++; 1016 CBQTRACE(_rmc_wrr_dequeue_next, 'ojda', ifd->cutoff_); 1017 goto _again; 1018 } 1019 #endif /* ADJUST_CUTOFF */ 1020 /* 1021 * If LINK_EFFICIENCY is turned on, then the first overlimit 1022 * class we encounter will send a packet if all the classes 1023 * of the link-sharing structure are overlimit. 1024 */ 1025 reset_cutoff(ifd); 1026 CBQTRACE(_rmc_wrr_dequeue_next, 'otsr', ifd->cutoff_); 1027 1028 if (!ifd->efficient_ || first == NULL) 1029 return (NULL); 1030 1031 cl = first; 1032 cpri = cl->pri_; 1033 #if 0 /* too time-consuming for nothing */ 1034 if (cl->sleeping_) 1035 callout_stop(&cl->callout_); 1036 cl->sleeping_ = 0; 1037 cl->undertime_.tv_sec = 0; 1038 #endif 1039 ifd->borrowed_[ifd->qi_] = cl->borrow_; 1040 ifd->cutoff_ = cl->borrow_->depth_; 1041 1042 /* 1043 * Deque the packet and do the book keeping... 1044 */ 1045 _wrr_out: 1046 if (op == ALTDQ_REMOVE) { 1047 m = _rmc_getq(cl); 1048 if (m == NULL) 1049 panic("_rmc_wrr_dequeue_next"); 1050 if (qempty(cl->q_)) 1051 ifd->na_[cpri]--; 1052 1053 /* 1054 * Update class statistics and link data. 1055 */ 1056 if (cl->bytes_alloc_ > 0) 1057 cl->bytes_alloc_ -= m_pktlen(m); 1058 1059 if ((cl->bytes_alloc_ <= 0) || first == cl) 1060 ifd->active_[cl->pri_] = cl->peer_; 1061 else 1062 ifd->active_[cl->pri_] = cl; 1063 1064 ifd->class_[ifd->qi_] = cl; 1065 ifd->curlen_[ifd->qi_] = m_pktlen(m); 1066 ifd->now_[ifd->qi_] = now; 1067 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_; 1068 ifd->queued_++; 1069 } else { 1070 /* mode == ALTDQ_PPOLL */ 1071 m = _rmc_pollq(cl); 1072 ifd->pollcache_ = cl; 1073 } 1074 return (m); 1075 } 1076 1077 /* 1078 * Dequeue & return next packet from the highest priority class that 1079 * has a packet to send & has enough allocation to send it. This 1080 * routine is called by a driver whenever it needs a new packet to 1081 * output. 1082 */ 1083 static struct mbuf * 1084 _rmc_prr_dequeue_next(struct rm_ifdat *ifd, int op) 1085 { 1086 struct mbuf *m; 1087 int cpri; 1088 struct rm_class *cl, *first = NULL; 1089 struct timeval now; 1090 1091 RM_GETTIME(now); 1092 1093 /* 1094 * if the driver polls the top of the queue and then removes 1095 * the polled packet, we must return the same packet. 1096 */ 1097 if (op == ALTDQ_REMOVE && ifd->pollcache_) { 1098 cl = ifd->pollcache_; 1099 cpri = cl->pri_; 1100 ifd->pollcache_ = NULL; 1101 goto _prr_out; 1102 } else { 1103 /* mode == ALTDQ_POLL || pollcache == NULL */ 1104 ifd->pollcache_ = NULL; 1105 ifd->borrowed_[ifd->qi_] = NULL; 1106 } 1107 #ifdef ADJUST_CUTOFF 1108 _again: 1109 #endif 1110 for (cpri = RM_MAXPRIO - 1; cpri >= 0; cpri--) { 1111 if (ifd->na_[cpri] == 0) 1112 continue; 1113 cl = ifd->active_[cpri]; 1114 KKASSERT(cl != NULL); 1115 do { 1116 if (!qempty(cl->q_)) { 1117 if ((cl->undertime_.tv_sec == 0) || 1118 rmc_under_limit(cl, &now)) 1119 goto _prr_out; 1120 if (first == NULL && cl->borrow_ != NULL) 1121 first = cl; 1122 } 1123 cl = cl->peer_; 1124 } while (cl != ifd->active_[cpri]); 1125 } 1126 1127 #ifdef ADJUST_CUTOFF 1128 /* 1129 * no underlimit class found. if cutoff is taking effect, increase 1130 * cutoff and try again. 1131 */ 1132 if (first != NULL && ifd->cutoff_ < ifd->root_->depth_) { 1133 ifd->cutoff_++; 1134 goto _again; 1135 } 1136 #endif /* ADJUST_CUTOFF */ 1137 /* 1138 * If LINK_EFFICIENCY is turned on, then the first overlimit 1139 * class we encounter will send a packet if all the classes 1140 * of the link-sharing structure are overlimit. 1141 */ 1142 reset_cutoff(ifd); 1143 if (!ifd->efficient_ || first == NULL) 1144 return (NULL); 1145 1146 cl = first; 1147 cpri = cl->pri_; 1148 #if 0 /* too time-consuming for nothing */ 1149 if (cl->sleeping_) 1150 callout_stop(&cl->callout_); 1151 cl->sleeping_ = 0; 1152 cl->undertime_.tv_sec = 0; 1153 #endif 1154 ifd->borrowed_[ifd->qi_] = cl->borrow_; 1155 ifd->cutoff_ = cl->borrow_->depth_; 1156 1157 /* 1158 * Deque the packet and do the book keeping... 1159 */ 1160 _prr_out: 1161 if (op == ALTDQ_REMOVE) { 1162 m = _rmc_getq(cl); 1163 if (m == NULL) 1164 panic("_rmc_prr_dequeue_next"); 1165 if (qempty(cl->q_)) 1166 ifd->na_[cpri]--; 1167 1168 ifd->active_[cpri] = cl->peer_; 1169 1170 ifd->class_[ifd->qi_] = cl; 1171 ifd->curlen_[ifd->qi_] = m_pktlen(m); 1172 ifd->now_[ifd->qi_] = now; 1173 ifd->qi_ = (ifd->qi_ + 1) % ifd->maxqueued_; 1174 ifd->queued_++; 1175 } else { 1176 /* mode == ALTDQ_POLL */ 1177 m = _rmc_pollq(cl); 1178 ifd->pollcache_ = cl; 1179 } 1180 return (m); 1181 } 1182 1183 /* 1184 * struct mbuf * 1185 * rmc_dequeue_next(struct rm_ifdat *ifd, struct timeval *now) - this function 1186 * is invoked by the packet driver to get the next packet to be 1187 * dequeued and output on the link. If WRR is enabled, then the 1188 * WRR dequeue next routine will determine the next packet to sent. 1189 * Otherwise, packet-by-packet round robin is invoked. 1190 * 1191 * Returns: NULL, if a packet is not available or if all 1192 * classes are overlimit. 1193 * 1194 * Otherwise, Pointer to the next packet. 1195 */ 1196 1197 struct mbuf * 1198 rmc_dequeue_next(struct rm_ifdat *ifd, int mode) 1199 { 1200 if (ifd->queued_ >= ifd->maxqueued_) 1201 return (NULL); 1202 else if (ifd->wrr_) 1203 return (_rmc_wrr_dequeue_next(ifd, mode)); 1204 else 1205 return (_rmc_prr_dequeue_next(ifd, mode)); 1206 } 1207 1208 /* 1209 * Update the utilization estimate for the packet that just completed. 1210 * The packet's class & the parent(s) of that class all get their 1211 * estimators updated. This routine is called by the driver's output- 1212 * packet-completion interrupt service routine. 1213 */ 1214 1215 /* 1216 * a macro to approximate "divide by 1000" that gives 0.000999, 1217 * if a value has enough effective digits. 1218 * (on pentium, mul takes 9 cycles but div takes 46!) 1219 */ 1220 #define NSEC_TO_USEC(t) (((t) >> 10) + ((t) >> 16) + ((t) >> 17)) 1221 void 1222 rmc_update_class_util(struct rm_ifdat *ifd) 1223 { 1224 int idle, avgidle, pktlen; 1225 int pkt_time, tidle; 1226 rm_class_t *cl, *borrowed; 1227 rm_class_t *borrows; 1228 struct timeval *nowp; 1229 1230 /* 1231 * Get the most recent completed class. 1232 */ 1233 if ((cl = ifd->class_[ifd->qo_]) == NULL) 1234 return; 1235 1236 pktlen = ifd->curlen_[ifd->qo_]; 1237 borrowed = ifd->borrowed_[ifd->qo_]; 1238 borrows = borrowed; 1239 1240 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen); 1241 1242 /* 1243 * Run estimator on class and its ancestors. 1244 */ 1245 /* 1246 * rm_update_class_util is designed to be called when the 1247 * transfer is completed from a xmit complete interrupt, 1248 * but most drivers don't implement an upcall for that. 1249 * so, just use estimated completion time. 1250 * as a result, ifd->qi_ and ifd->qo_ are always synced. 1251 */ 1252 nowp = &ifd->now_[ifd->qo_]; 1253 /* get pkt_time (for link) in usec */ 1254 #if 1 /* use approximation */ 1255 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_; 1256 pkt_time = NSEC_TO_USEC(pkt_time); 1257 #else 1258 pkt_time = ifd->curlen_[ifd->qo_] * ifd->ns_per_byte_ / 1000; 1259 #endif 1260 #if 1 /* ALTQ4PPP */ 1261 if (TV_LT(nowp, &ifd->ifnow_)) { 1262 int iftime; 1263 1264 /* 1265 * make sure the estimated completion time does not go 1266 * too far. it can happen when the link layer supports 1267 * data compression or the interface speed is set to 1268 * a much lower value. 1269 */ 1270 TV_DELTA(&ifd->ifnow_, nowp, iftime); 1271 if (iftime+pkt_time < ifd->maxiftime_) { 1272 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_); 1273 } else { 1274 TV_ADD_DELTA(nowp, ifd->maxiftime_, &ifd->ifnow_); 1275 } 1276 } else { 1277 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_); 1278 } 1279 #else 1280 if (TV_LT(nowp, &ifd->ifnow_)) { 1281 TV_ADD_DELTA(&ifd->ifnow_, pkt_time, &ifd->ifnow_); 1282 } else { 1283 TV_ADD_DELTA(nowp, pkt_time, &ifd->ifnow_); 1284 } 1285 #endif 1286 1287 while (cl != NULL) { 1288 TV_DELTA(&ifd->ifnow_, &cl->last_, idle); 1289 if (idle >= 2000000) 1290 /* 1291 * this class is idle enough, reset avgidle. 1292 * (TV_DELTA returns 2000000 us when delta is large.) 1293 */ 1294 cl->avgidle_ = cl->maxidle_; 1295 1296 /* get pkt_time (for class) in usec */ 1297 #if 1 /* use approximation */ 1298 pkt_time = pktlen * cl->ns_per_byte_; 1299 pkt_time = NSEC_TO_USEC(pkt_time); 1300 #else 1301 pkt_time = pktlen * cl->ns_per_byte_ / 1000; 1302 #endif 1303 idle -= pkt_time; 1304 1305 avgidle = cl->avgidle_; 1306 avgidle += idle - (avgidle >> RM_FILTER_GAIN); 1307 cl->avgidle_ = avgidle; 1308 1309 /* Are we overlimit ? */ 1310 if (avgidle <= 0) { 1311 CBQTRACE(rmc_update_class_util, 'milo', cl->stats_.handle); 1312 #if 1 /* ALTQ */ 1313 /* 1314 * need some lower bound for avgidle, otherwise 1315 * a borrowing class gets unbounded penalty. 1316 */ 1317 if (avgidle < cl->minidle_) 1318 avgidle = cl->avgidle_ = cl->minidle_; 1319 #endif 1320 /* set next idle to make avgidle 0 */ 1321 tidle = pkt_time + 1322 (((1 - RM_POWER) * avgidle) >> RM_FILTER_GAIN); 1323 TV_ADD_DELTA(nowp, tidle, &cl->undertime_); 1324 ++cl->stats_.over; 1325 } else { 1326 cl->avgidle_ = 1327 (avgidle > cl->maxidle_) ? cl->maxidle_ : avgidle; 1328 cl->undertime_.tv_sec = 0; 1329 if (cl->sleeping_) { 1330 callout_stop(&cl->callout_); 1331 cl->sleeping_ = 0; 1332 } 1333 } 1334 1335 if (borrows != NULL) { 1336 if (borrows != cl) 1337 ++cl->stats_.borrows; 1338 else 1339 borrows = NULL; 1340 } 1341 cl->last_ = ifd->ifnow_; 1342 cl->last_pkttime_ = pkt_time; 1343 1344 #if 1 1345 if (cl->parent_ == NULL) { 1346 /* take stats of root class */ 1347 PKTCNTR_ADD(&cl->stats_.xmit_cnt, pktlen); 1348 } 1349 #endif 1350 1351 cl = cl->parent_; 1352 } 1353 1354 /* 1355 * Check to see if cutoff needs to set to a new level. 1356 */ 1357 cl = ifd->class_[ifd->qo_]; 1358 if (borrowed && (ifd->cutoff_ >= borrowed->depth_)) { 1359 #if 1 /* ALTQ */ 1360 if ((qlen(cl->q_) <= 0) || TV_LT(nowp, &borrowed->undertime_)) { 1361 rmc_tl_satisfied(ifd, nowp); 1362 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_); 1363 } else { 1364 ifd->cutoff_ = borrowed->depth_; 1365 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_); 1366 } 1367 #else /* !ALTQ */ 1368 if ((qlen(cl->q_) <= 1) || TV_LT(&now, &borrowed->undertime_)) { 1369 reset_cutoff(ifd); 1370 #ifdef notdef 1371 rmc_tl_satisfied(ifd, &now); 1372 #endif 1373 CBQTRACE(rmc_update_class_util, 'broe', ifd->cutoff_); 1374 } else { 1375 ifd->cutoff_ = borrowed->depth_; 1376 CBQTRACE(rmc_update_class_util, 'ffob', borrowed->depth_); 1377 } 1378 #endif /* !ALTQ */ 1379 } 1380 1381 /* 1382 * Release class slot 1383 */ 1384 ifd->borrowed_[ifd->qo_] = NULL; 1385 ifd->class_[ifd->qo_] = NULL; 1386 ifd->qo_ = (ifd->qo_ + 1) % ifd->maxqueued_; 1387 ifd->queued_--; 1388 } 1389 1390 /* 1391 * void 1392 * rmc_drop_action(struct rm_class *cl) - Generic (not protocol-specific) 1393 * over-limit action routines. These get invoked by rmc_under_limit() 1394 * if a class with packets to send if over its bandwidth limit & can't 1395 * borrow from a parent class. 1396 * 1397 * Returns: NONE 1398 */ 1399 1400 static void 1401 rmc_drop_action(struct rm_class *cl) 1402 { 1403 struct rm_ifdat *ifd = cl->ifdat_; 1404 1405 KKASSERT(qlen(cl->q_) > 0); 1406 _rmc_dropq(cl); 1407 if (qempty(cl->q_)) 1408 ifd->na_[cl->pri_]--; 1409 } 1410 1411 void 1412 rmc_dropall(struct rm_class *cl) 1413 { 1414 struct rm_ifdat *ifd = cl->ifdat_; 1415 1416 if (!qempty(cl->q_)) { 1417 _flushq(cl->q_); 1418 1419 ifd->na_[cl->pri_]--; 1420 } 1421 } 1422 1423 /* 1424 * void 1425 * rmc_delay_action(struct rm_class *cl) - This function is the generic CBQ 1426 * delay action routine. It is invoked via rmc_under_limit when the 1427 * packet is discoverd to be overlimit. 1428 * 1429 * If the delay action is result of borrow class being overlimit, then 1430 * delay for the offtime of the borrowing class that is overlimit. 1431 * 1432 * Returns: NONE 1433 */ 1434 1435 void 1436 rmc_delay_action(struct rm_class *cl, struct rm_class *borrow) 1437 { 1438 int delay, t, extradelay; 1439 1440 cl->stats_.overactions++; 1441 TV_DELTA(&cl->undertime_, &cl->overtime_, delay); 1442 #ifndef BORROW_OFFTIME 1443 delay += cl->offtime_; 1444 #endif 1445 1446 if (!cl->sleeping_) { 1447 CBQTRACE(rmc_delay_action, 'yled', cl->stats_.handle); 1448 #ifdef BORROW_OFFTIME 1449 if (borrow != NULL) 1450 extradelay = borrow->offtime_; 1451 else 1452 #endif 1453 extradelay = cl->offtime_; 1454 1455 #ifdef ALTQ 1456 /* 1457 * XXX recalculate suspend time: 1458 * current undertime is (tidle + pkt_time) calculated 1459 * from the last transmission. 1460 * tidle: time required to bring avgidle back to 0 1461 * pkt_time: target waiting time for this class 1462 * we need to replace pkt_time by offtime 1463 */ 1464 extradelay -= cl->last_pkttime_; 1465 #endif 1466 if (extradelay > 0) { 1467 TV_ADD_DELTA(&cl->undertime_, extradelay, &cl->undertime_); 1468 delay += extradelay; 1469 } 1470 1471 cl->sleeping_ = 1; 1472 cl->stats_.delays++; 1473 1474 /* 1475 * Since packets are phased randomly with respect to the 1476 * clock, 1 tick (the next clock tick) can be an arbitrarily 1477 * short time so we have to wait for at least two ticks. 1478 * NOTE: If there's no other traffic, we need the timer as 1479 * a 'backstop' to restart this class. 1480 */ 1481 if (delay > ustick * 2) 1482 t = (delay + ustick - 1) / ustick; 1483 else 1484 t = 2; 1485 callout_reset(&cl->callout_, t, rmc_restart, cl); 1486 } 1487 } 1488 1489 /* 1490 * void 1491 * rmc_restart() - is just a helper routine for rmc_delay_action -- it is 1492 * called by the system timer code & is responsible checking if the 1493 * class is still sleeping (it might have been restarted as a side 1494 * effect of the queue scan on a packet arrival) and, if so, restarting 1495 * output for the class. Inspecting the class state & restarting output 1496 * require locking the class structure. In general the driver is 1497 * responsible for locking but this is the only routine that is not 1498 * called directly or indirectly from the interface driver so it has 1499 * know about system locking conventions. Under bsd, locking is done 1500 * by raising IPL to splimp so that's what's implemented here. On a 1501 * different system this would probably need to be changed. 1502 * 1503 * Since this function is called from an independant timeout, we 1504 * have to set up the lock conditions expected for the ALTQ operation. 1505 * Note that the restart will probably fall through to an if_start. 1506 * 1507 * Returns: NONE 1508 */ 1509 1510 static void 1511 rmc_restart(void *arg) 1512 { 1513 struct rm_class *cl = arg; 1514 struct rm_ifdat *ifd = cl->ifdat_; 1515 1516 ALTQ_LOCK(ifd->ifq_); 1517 if (cl->sleeping_) { 1518 cl->sleeping_ = 0; 1519 cl->undertime_.tv_sec = 0; 1520 1521 if (ifd->queued_ < ifd->maxqueued_ && ifd->restart != NULL) { 1522 CBQTRACE(rmc_restart, 'trts', cl->stats_.handle); 1523 (ifd->restart)(ifd->ifq_); 1524 } 1525 } 1526 ALTQ_UNLOCK(ifd->ifq_); 1527 } 1528 1529 /* 1530 * void 1531 * rmc_root_overlimit(struct rm_class *cl) - This the generic overlimit 1532 * handling routine for the root class of the link sharing structure. 1533 * 1534 * Returns: NONE 1535 */ 1536 1537 static void 1538 rmc_root_overlimit(struct rm_class *cl, struct rm_class *borrow) 1539 { 1540 panic("rmc_root_overlimit"); 1541 } 1542 1543 /* 1544 * Packet Queue handling routines. Eventually, this is to localize the 1545 * effects on the code whether queues are red queues or droptail 1546 * queues. 1547 */ 1548 1549 static int 1550 _rmc_addq(rm_class_t *cl, struct mbuf *m) 1551 { 1552 #ifdef ALTQ_RIO 1553 if (q_is_rio(cl->q_)) 1554 return rio_addq((rio_t *)cl->red_, cl->q_, m, cl->pktattr_); 1555 #endif 1556 #ifdef ALTQ_RED 1557 if (q_is_red(cl->q_)) 1558 return red_addq(cl->red_, cl->q_, m, cl->pktattr_); 1559 #endif /* ALTQ_RED */ 1560 1561 if (cl->flags_ & RMCF_CLEARDSCP) 1562 write_dsfield(m, cl->pktattr_, 0); 1563 1564 _addq(cl->q_, m); 1565 return (0); 1566 } 1567 1568 /* note: _rmc_dropq is not called for red */ 1569 static void 1570 _rmc_dropq(rm_class_t *cl) 1571 { 1572 struct mbuf *m; 1573 1574 if ((m = _getq(cl->q_)) != NULL) 1575 m_freem(m); 1576 } 1577 1578 static struct mbuf * 1579 _rmc_getq(rm_class_t *cl) 1580 { 1581 #ifdef ALTQ_RIO 1582 if (q_is_rio(cl->q_)) 1583 return rio_getq((rio_t *)cl->red_, cl->q_); 1584 #endif 1585 #ifdef ALTQ_RED 1586 if (q_is_red(cl->q_)) 1587 return red_getq(cl->red_, cl->q_); 1588 #endif 1589 return _getq(cl->q_); 1590 } 1591 1592 static struct mbuf * 1593 _rmc_pollq(rm_class_t *cl) 1594 { 1595 return qhead(cl->q_); 1596 } 1597 1598 #ifdef CBQ_TRACE 1599 /* 1600 * DDB hook to trace cbq events: 1601 * the last 1024 events are held in a circular buffer. 1602 * use "call cbqtrace_dump(N)" to display 20 events from Nth event. 1603 */ 1604 void cbqtrace_dump(int); 1605 static char *rmc_funcname(void *); 1606 1607 static struct rmc_funcs { 1608 void *func; 1609 char *name; 1610 } rmc_funcs[] = { 1611 rmc_init, "rmc_init", 1612 rmc_queue_packet, "rmc_queue_packet", 1613 rmc_under_limit, "rmc_under_limit", 1614 rmc_update_class_util, "rmc_update_class_util", 1615 rmc_delay_action, "rmc_delay_action", 1616 rmc_restart, "rmc_restart", 1617 _rmc_wrr_dequeue_next, "_rmc_wrr_dequeue_next", 1618 NULL, NULL 1619 }; 1620 1621 static chari * 1622 rmc_funcname(void *func) 1623 { 1624 struct rmc_funcs *fp; 1625 1626 for (fp = rmc_funcs; fp->func != NULL; fp++) { 1627 if (fp->func == func) 1628 return (fp->name); 1629 } 1630 1631 return ("unknown"); 1632 } 1633 1634 void 1635 cbqtrace_dump(int counter) 1636 { 1637 int i, *p; 1638 char *cp; 1639 1640 counter = counter % NCBQTRACE; 1641 p = (int *)&cbqtrace_buffer[counter]; 1642 1643 for (i=0; i<20; i++) { 1644 kprintf("[0x%x] ", *p++); 1645 kprintf("%s: ", rmc_funcname((void *)*p++)); 1646 cp = (char *)p++; 1647 kprintf("%c%c%c%c: ", cp[0], cp[1], cp[2], cp[3]); 1648 kprintf("%d\n",*p++); 1649 1650 if (p >= (int *)&cbqtrace_buffer[NCBQTRACE]) 1651 p = (int *)cbqtrace_buffer; 1652 } 1653 } 1654 #endif /* CBQ_TRACE */ 1655 #endif /* ALTQ_CBQ */ 1656