1 /* 2 * net/sched/sch_mqprio.c 3 * 4 * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com> 5 * 6 * This program is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU General Public License 8 * version 2 as published by the Free Software Foundation. 9 */ 10 11 #include <linux/types.h> 12 #include <linux/slab.h> 13 #include <linux/kernel.h> 14 #include <linux/string.h> 15 #include <linux/errno.h> 16 #include <linux/skbuff.h> 17 #include <linux/module.h> 18 #include <net/netlink.h> 19 #include <net/pkt_sched.h> 20 #include <net/sch_generic.h> 21 #include <net/pkt_cls.h> 22 23 struct mqprio_sched { 24 struct Qdisc **qdiscs; 25 u16 mode; 26 u16 shaper; 27 int hw_offload; 28 u32 flags; 29 u64 min_rate[TC_QOPT_MAX_QUEUE]; 30 u64 max_rate[TC_QOPT_MAX_QUEUE]; 31 }; 32 33 static void mqprio_destroy(struct Qdisc *sch) 34 { 35 struct net_device *dev = qdisc_dev(sch); 36 struct mqprio_sched *priv = qdisc_priv(sch); 37 unsigned int ntx; 38 39 if (priv->qdiscs) { 40 for (ntx = 0; 41 ntx < dev->num_tx_queues && priv->qdiscs[ntx]; 42 ntx++) 43 qdisc_destroy(priv->qdiscs[ntx]); 44 kfree(priv->qdiscs); 45 } 46 47 if (priv->hw_offload && dev->netdev_ops->ndo_setup_tc) { 48 struct tc_mqprio_qopt_offload mqprio = { { 0 } }; 49 50 switch (priv->mode) { 51 case TC_MQPRIO_MODE_DCB: 52 case TC_MQPRIO_MODE_CHANNEL: 53 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_MQPRIO, 54 &mqprio); 55 break; 56 default: 57 return; 58 } 59 } else { 60 netdev_set_num_tc(dev, 0); 61 } 62 } 63 64 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt) 65 { 66 int i, j; 67 68 /* Verify num_tc is not out of max range */ 69 if (qopt->num_tc > TC_MAX_QUEUE) 70 return -EINVAL; 71 72 /* Verify priority mapping uses valid tcs */ 73 for (i = 0; i < TC_BITMASK + 1; i++) { 74 if (qopt->prio_tc_map[i] >= qopt->num_tc) 75 return -EINVAL; 76 } 77 78 /* Limit qopt->hw to maximum supported offload value. Drivers have 79 * the option of overriding this later if they don't support the a 80 * given offload type. 81 */ 82 if (qopt->hw > TC_MQPRIO_HW_OFFLOAD_MAX) 83 qopt->hw = TC_MQPRIO_HW_OFFLOAD_MAX; 84 85 /* If hardware offload is requested we will leave it to the device 86 * to either populate the queue counts itself or to validate the 87 * provided queue counts. If ndo_setup_tc is not present then 88 * hardware doesn't support offload and we should return an error. 89 */ 90 if (qopt->hw) 91 return dev->netdev_ops->ndo_setup_tc ? 0 : -EINVAL; 92 93 for (i = 0; i < qopt->num_tc; i++) { 94 unsigned int last = qopt->offset[i] + qopt->count[i]; 95 96 /* Verify the queue count is in tx range being equal to the 97 * real_num_tx_queues indicates the last queue is in use. 98 */ 99 if (qopt->offset[i] >= dev->real_num_tx_queues || 100 !qopt->count[i] || 101 last > dev->real_num_tx_queues) 102 return -EINVAL; 103 104 /* Verify that the offset and counts do not overlap */ 105 for (j = i + 1; j < qopt->num_tc; j++) { 106 if (last > qopt->offset[j]) 107 return -EINVAL; 108 } 109 } 110 111 return 0; 112 } 113 114 static const struct nla_policy mqprio_policy[TCA_MQPRIO_MAX + 1] = { 115 [TCA_MQPRIO_MODE] = { .len = sizeof(u16) }, 116 [TCA_MQPRIO_SHAPER] = { .len = sizeof(u16) }, 117 [TCA_MQPRIO_MIN_RATE64] = { .type = NLA_NESTED }, 118 [TCA_MQPRIO_MAX_RATE64] = { .type = NLA_NESTED }, 119 }; 120 121 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 122 const struct nla_policy *policy, int len) 123 { 124 int nested_len = nla_len(nla) - NLA_ALIGN(len); 125 126 if (nested_len >= nla_attr_size(0)) 127 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len), 128 nested_len, policy, NULL); 129 130 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 131 return 0; 132 } 133 134 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt) 135 { 136 struct net_device *dev = qdisc_dev(sch); 137 struct mqprio_sched *priv = qdisc_priv(sch); 138 struct netdev_queue *dev_queue; 139 struct Qdisc *qdisc; 140 int i, err = -EOPNOTSUPP; 141 struct tc_mqprio_qopt *qopt = NULL; 142 struct nlattr *tb[TCA_MQPRIO_MAX + 1]; 143 struct nlattr *attr; 144 int rem; 145 int len; 146 147 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE); 148 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK); 149 150 if (sch->parent != TC_H_ROOT) 151 return -EOPNOTSUPP; 152 153 if (!netif_is_multiqueue(dev)) 154 return -EOPNOTSUPP; 155 156 /* make certain can allocate enough classids to handle queues */ 157 if (dev->num_tx_queues >= TC_H_MIN_PRIORITY) 158 return -ENOMEM; 159 160 if (!opt || nla_len(opt) < sizeof(*qopt)) 161 return -EINVAL; 162 163 qopt = nla_data(opt); 164 if (mqprio_parse_opt(dev, qopt)) 165 return -EINVAL; 166 167 len = nla_len(opt) - NLA_ALIGN(sizeof(*qopt)); 168 if (len > 0) { 169 err = parse_attr(tb, TCA_MQPRIO_MAX, opt, mqprio_policy, 170 sizeof(*qopt)); 171 if (err < 0) 172 return err; 173 174 if (!qopt->hw) 175 return -EINVAL; 176 177 if (tb[TCA_MQPRIO_MODE]) { 178 priv->flags |= TC_MQPRIO_F_MODE; 179 priv->mode = *(u16 *)nla_data(tb[TCA_MQPRIO_MODE]); 180 } 181 182 if (tb[TCA_MQPRIO_SHAPER]) { 183 priv->flags |= TC_MQPRIO_F_SHAPER; 184 priv->shaper = *(u16 *)nla_data(tb[TCA_MQPRIO_SHAPER]); 185 } 186 187 if (tb[TCA_MQPRIO_MIN_RATE64]) { 188 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) 189 return -EINVAL; 190 i = 0; 191 nla_for_each_nested(attr, tb[TCA_MQPRIO_MIN_RATE64], 192 rem) { 193 if (nla_type(attr) != TCA_MQPRIO_MIN_RATE64) 194 return -EINVAL; 195 if (i >= qopt->num_tc) 196 break; 197 priv->min_rate[i] = *(u64 *)nla_data(attr); 198 i++; 199 } 200 priv->flags |= TC_MQPRIO_F_MIN_RATE; 201 } 202 203 if (tb[TCA_MQPRIO_MAX_RATE64]) { 204 if (priv->shaper != TC_MQPRIO_SHAPER_BW_RATE) 205 return -EINVAL; 206 i = 0; 207 nla_for_each_nested(attr, tb[TCA_MQPRIO_MAX_RATE64], 208 rem) { 209 if (nla_type(attr) != TCA_MQPRIO_MAX_RATE64) 210 return -EINVAL; 211 if (i >= qopt->num_tc) 212 break; 213 priv->max_rate[i] = *(u64 *)nla_data(attr); 214 i++; 215 } 216 priv->flags |= TC_MQPRIO_F_MAX_RATE; 217 } 218 } 219 220 /* pre-allocate qdisc, attachment can't fail */ 221 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]), 222 GFP_KERNEL); 223 if (!priv->qdiscs) 224 return -ENOMEM; 225 226 for (i = 0; i < dev->num_tx_queues; i++) { 227 dev_queue = netdev_get_tx_queue(dev, i); 228 qdisc = qdisc_create_dflt(dev_queue, 229 get_default_qdisc_ops(dev, i), 230 TC_H_MAKE(TC_H_MAJ(sch->handle), 231 TC_H_MIN(i + 1))); 232 if (!qdisc) 233 return -ENOMEM; 234 235 priv->qdiscs[i] = qdisc; 236 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 237 } 238 239 /* If the mqprio options indicate that hardware should own 240 * the queue mapping then run ndo_setup_tc otherwise use the 241 * supplied and verified mapping 242 */ 243 if (qopt->hw) { 244 struct tc_mqprio_qopt_offload mqprio = {.qopt = *qopt}; 245 246 switch (priv->mode) { 247 case TC_MQPRIO_MODE_DCB: 248 if (priv->shaper != TC_MQPRIO_SHAPER_DCB) 249 return -EINVAL; 250 break; 251 case TC_MQPRIO_MODE_CHANNEL: 252 mqprio.flags = priv->flags; 253 if (priv->flags & TC_MQPRIO_F_MODE) 254 mqprio.mode = priv->mode; 255 if (priv->flags & TC_MQPRIO_F_SHAPER) 256 mqprio.shaper = priv->shaper; 257 if (priv->flags & TC_MQPRIO_F_MIN_RATE) 258 for (i = 0; i < mqprio.qopt.num_tc; i++) 259 mqprio.min_rate[i] = priv->min_rate[i]; 260 if (priv->flags & TC_MQPRIO_F_MAX_RATE) 261 for (i = 0; i < mqprio.qopt.num_tc; i++) 262 mqprio.max_rate[i] = priv->max_rate[i]; 263 break; 264 default: 265 return -EINVAL; 266 } 267 err = dev->netdev_ops->ndo_setup_tc(dev, 268 TC_SETUP_MQPRIO, 269 &mqprio); 270 if (err) 271 return err; 272 273 priv->hw_offload = mqprio.qopt.hw; 274 } else { 275 netdev_set_num_tc(dev, qopt->num_tc); 276 for (i = 0; i < qopt->num_tc; i++) 277 netdev_set_tc_queue(dev, i, 278 qopt->count[i], qopt->offset[i]); 279 } 280 281 /* Always use supplied priority mappings */ 282 for (i = 0; i < TC_BITMASK + 1; i++) 283 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]); 284 285 sch->flags |= TCQ_F_MQROOT; 286 return 0; 287 } 288 289 static void mqprio_attach(struct Qdisc *sch) 290 { 291 struct net_device *dev = qdisc_dev(sch); 292 struct mqprio_sched *priv = qdisc_priv(sch); 293 struct Qdisc *qdisc, *old; 294 unsigned int ntx; 295 296 /* Attach underlying qdisc */ 297 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) { 298 qdisc = priv->qdiscs[ntx]; 299 old = dev_graft_qdisc(qdisc->dev_queue, qdisc); 300 if (old) 301 qdisc_destroy(old); 302 if (ntx < dev->real_num_tx_queues) 303 qdisc_hash_add(qdisc, false); 304 } 305 kfree(priv->qdiscs); 306 priv->qdiscs = NULL; 307 } 308 309 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch, 310 unsigned long cl) 311 { 312 struct net_device *dev = qdisc_dev(sch); 313 unsigned long ntx = cl - 1; 314 315 if (ntx >= dev->num_tx_queues) 316 return NULL; 317 return netdev_get_tx_queue(dev, ntx); 318 } 319 320 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new, 321 struct Qdisc **old) 322 { 323 struct net_device *dev = qdisc_dev(sch); 324 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 325 326 if (!dev_queue) 327 return -EINVAL; 328 329 if (dev->flags & IFF_UP) 330 dev_deactivate(dev); 331 332 *old = dev_graft_qdisc(dev_queue, new); 333 334 if (new) 335 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 336 337 if (dev->flags & IFF_UP) 338 dev_activate(dev); 339 340 return 0; 341 } 342 343 static int dump_rates(struct mqprio_sched *priv, 344 struct tc_mqprio_qopt *opt, struct sk_buff *skb) 345 { 346 struct nlattr *nest; 347 int i; 348 349 if (priv->flags & TC_MQPRIO_F_MIN_RATE) { 350 nest = nla_nest_start(skb, TCA_MQPRIO_MIN_RATE64); 351 if (!nest) 352 goto nla_put_failure; 353 354 for (i = 0; i < opt->num_tc; i++) { 355 if (nla_put(skb, TCA_MQPRIO_MIN_RATE64, 356 sizeof(priv->min_rate[i]), 357 &priv->min_rate[i])) 358 goto nla_put_failure; 359 } 360 nla_nest_end(skb, nest); 361 } 362 363 if (priv->flags & TC_MQPRIO_F_MAX_RATE) { 364 nest = nla_nest_start(skb, TCA_MQPRIO_MAX_RATE64); 365 if (!nest) 366 goto nla_put_failure; 367 368 for (i = 0; i < opt->num_tc; i++) { 369 if (nla_put(skb, TCA_MQPRIO_MAX_RATE64, 370 sizeof(priv->max_rate[i]), 371 &priv->max_rate[i])) 372 goto nla_put_failure; 373 } 374 nla_nest_end(skb, nest); 375 } 376 return 0; 377 378 nla_put_failure: 379 nla_nest_cancel(skb, nest); 380 return -1; 381 } 382 383 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb) 384 { 385 struct net_device *dev = qdisc_dev(sch); 386 struct mqprio_sched *priv = qdisc_priv(sch); 387 struct nlattr *nla = (struct nlattr *)skb_tail_pointer(skb); 388 struct tc_mqprio_qopt opt = { 0 }; 389 struct Qdisc *qdisc; 390 unsigned int i; 391 392 sch->q.qlen = 0; 393 memset(&sch->bstats, 0, sizeof(sch->bstats)); 394 memset(&sch->qstats, 0, sizeof(sch->qstats)); 395 396 for (i = 0; i < dev->num_tx_queues; i++) { 397 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc); 398 spin_lock_bh(qdisc_lock(qdisc)); 399 sch->q.qlen += qdisc->q.qlen; 400 sch->bstats.bytes += qdisc->bstats.bytes; 401 sch->bstats.packets += qdisc->bstats.packets; 402 sch->qstats.backlog += qdisc->qstats.backlog; 403 sch->qstats.drops += qdisc->qstats.drops; 404 sch->qstats.requeues += qdisc->qstats.requeues; 405 sch->qstats.overlimits += qdisc->qstats.overlimits; 406 spin_unlock_bh(qdisc_lock(qdisc)); 407 } 408 409 opt.num_tc = netdev_get_num_tc(dev); 410 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map)); 411 opt.hw = priv->hw_offload; 412 413 for (i = 0; i < netdev_get_num_tc(dev); i++) { 414 opt.count[i] = dev->tc_to_txq[i].count; 415 opt.offset[i] = dev->tc_to_txq[i].offset; 416 } 417 418 if (nla_put(skb, TCA_OPTIONS, NLA_ALIGN(sizeof(opt)), &opt)) 419 goto nla_put_failure; 420 421 if ((priv->flags & TC_MQPRIO_F_MODE) && 422 nla_put_u16(skb, TCA_MQPRIO_MODE, priv->mode)) 423 goto nla_put_failure; 424 425 if ((priv->flags & TC_MQPRIO_F_SHAPER) && 426 nla_put_u16(skb, TCA_MQPRIO_SHAPER, priv->shaper)) 427 goto nla_put_failure; 428 429 if ((priv->flags & TC_MQPRIO_F_MIN_RATE || 430 priv->flags & TC_MQPRIO_F_MAX_RATE) && 431 (dump_rates(priv, &opt, skb) != 0)) 432 goto nla_put_failure; 433 434 return nla_nest_end(skb, nla); 435 nla_put_failure: 436 nlmsg_trim(skb, nla); 437 return -1; 438 } 439 440 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl) 441 { 442 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 443 444 if (!dev_queue) 445 return NULL; 446 447 return dev_queue->qdisc_sleeping; 448 } 449 450 static unsigned long mqprio_find(struct Qdisc *sch, u32 classid) 451 { 452 struct net_device *dev = qdisc_dev(sch); 453 unsigned int ntx = TC_H_MIN(classid); 454 455 /* There are essentially two regions here that have valid classid 456 * values. The first region will have a classid value of 1 through 457 * num_tx_queues. All of these are backed by actual Qdiscs. 458 */ 459 if (ntx < TC_H_MIN_PRIORITY) 460 return (ntx <= dev->num_tx_queues) ? ntx : 0; 461 462 /* The second region represents the hardware traffic classes. These 463 * are represented by classid values of TC_H_MIN_PRIORITY through 464 * TC_H_MIN_PRIORITY + netdev_get_num_tc - 1 465 */ 466 return ((ntx - TC_H_MIN_PRIORITY) < netdev_get_num_tc(dev)) ? ntx : 0; 467 } 468 469 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl, 470 struct sk_buff *skb, struct tcmsg *tcm) 471 { 472 if (cl < TC_H_MIN_PRIORITY) { 473 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 474 struct net_device *dev = qdisc_dev(sch); 475 int tc = netdev_txq_to_tc(dev, cl - 1); 476 477 tcm->tcm_parent = (tc < 0) ? 0 : 478 TC_H_MAKE(TC_H_MAJ(sch->handle), 479 TC_H_MIN(tc + TC_H_MIN_PRIORITY)); 480 tcm->tcm_info = dev_queue->qdisc_sleeping->handle; 481 } else { 482 tcm->tcm_parent = TC_H_ROOT; 483 tcm->tcm_info = 0; 484 } 485 tcm->tcm_handle |= TC_H_MIN(cl); 486 return 0; 487 } 488 489 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl, 490 struct gnet_dump *d) 491 __releases(d->lock) 492 __acquires(d->lock) 493 { 494 if (cl >= TC_H_MIN_PRIORITY) { 495 int i; 496 __u32 qlen = 0; 497 struct Qdisc *qdisc; 498 struct gnet_stats_queue qstats = {0}; 499 struct gnet_stats_basic_packed bstats = {0}; 500 struct net_device *dev = qdisc_dev(sch); 501 struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK]; 502 503 /* Drop lock here it will be reclaimed before touching 504 * statistics this is required because the d->lock we 505 * hold here is the look on dev_queue->qdisc_sleeping 506 * also acquired below. 507 */ 508 if (d->lock) 509 spin_unlock_bh(d->lock); 510 511 for (i = tc.offset; i < tc.offset + tc.count; i++) { 512 struct netdev_queue *q = netdev_get_tx_queue(dev, i); 513 514 qdisc = rtnl_dereference(q->qdisc); 515 spin_lock_bh(qdisc_lock(qdisc)); 516 qlen += qdisc->q.qlen; 517 bstats.bytes += qdisc->bstats.bytes; 518 bstats.packets += qdisc->bstats.packets; 519 qstats.backlog += qdisc->qstats.backlog; 520 qstats.drops += qdisc->qstats.drops; 521 qstats.requeues += qdisc->qstats.requeues; 522 qstats.overlimits += qdisc->qstats.overlimits; 523 spin_unlock_bh(qdisc_lock(qdisc)); 524 } 525 /* Reclaim root sleeping lock before completing stats */ 526 if (d->lock) 527 spin_lock_bh(d->lock); 528 if (gnet_stats_copy_basic(NULL, d, NULL, &bstats) < 0 || 529 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0) 530 return -1; 531 } else { 532 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 533 534 sch = dev_queue->qdisc_sleeping; 535 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), 536 d, NULL, &sch->bstats) < 0 || 537 gnet_stats_copy_queue(d, NULL, 538 &sch->qstats, sch->q.qlen) < 0) 539 return -1; 540 } 541 return 0; 542 } 543 544 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg) 545 { 546 struct net_device *dev = qdisc_dev(sch); 547 unsigned long ntx; 548 549 if (arg->stop) 550 return; 551 552 /* Walk hierarchy with a virtual class per tc */ 553 arg->count = arg->skip; 554 for (ntx = arg->skip; ntx < netdev_get_num_tc(dev); ntx++) { 555 if (arg->fn(sch, ntx + TC_H_MIN_PRIORITY, arg) < 0) { 556 arg->stop = 1; 557 return; 558 } 559 arg->count++; 560 } 561 562 /* Pad the values and skip over unused traffic classes */ 563 if (ntx < TC_MAX_QUEUE) { 564 arg->count = TC_MAX_QUEUE; 565 ntx = TC_MAX_QUEUE; 566 } 567 568 /* Reset offset, sort out remaining per-queue qdiscs */ 569 for (ntx -= TC_MAX_QUEUE; ntx < dev->num_tx_queues; ntx++) { 570 if (arg->fn(sch, ntx + 1, arg) < 0) { 571 arg->stop = 1; 572 return; 573 } 574 arg->count++; 575 } 576 } 577 578 static struct netdev_queue *mqprio_select_queue(struct Qdisc *sch, 579 struct tcmsg *tcm) 580 { 581 return mqprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent)); 582 } 583 584 static const struct Qdisc_class_ops mqprio_class_ops = { 585 .graft = mqprio_graft, 586 .leaf = mqprio_leaf, 587 .find = mqprio_find, 588 .walk = mqprio_walk, 589 .dump = mqprio_dump_class, 590 .dump_stats = mqprio_dump_class_stats, 591 .select_queue = mqprio_select_queue, 592 }; 593 594 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = { 595 .cl_ops = &mqprio_class_ops, 596 .id = "mqprio", 597 .priv_size = sizeof(struct mqprio_sched), 598 .init = mqprio_init, 599 .destroy = mqprio_destroy, 600 .attach = mqprio_attach, 601 .dump = mqprio_dump, 602 .owner = THIS_MODULE, 603 }; 604 605 static int __init mqprio_module_init(void) 606 { 607 return register_qdisc(&mqprio_qdisc_ops); 608 } 609 610 static void __exit mqprio_module_exit(void) 611 { 612 unregister_qdisc(&mqprio_qdisc_ops); 613 } 614 615 module_init(mqprio_module_init); 616 module_exit(mqprio_module_exit); 617 618 MODULE_LICENSE("GPL"); 619