1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * net/sched/sch_mqprio.c 4 * 5 * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com> 6 */ 7 8 #include <linux/types.h> 9 #include <linux/slab.h> 10 #include <linux/kernel.h> 11 #include <linux/string.h> 12 #include <linux/errno.h> 13 #include <linux/skbuff.h> 14 #include <linux/module.h> 15 #include <net/netlink.h> 16 #include <net/pkt_sched.h> 17 #include <net/sch_generic.h> 18 #include <net/pkt_cls.h> 19 20 struct mqprio_sched { 21 struct Qdisc **qdiscs; 22 u16 mode; 23 u16 shaper; 24 int hw_offload; 25 u32 flags; 26 u64 min_rate[TC_QOPT_MAX_QUEUE]; 27 u64 max_rate[TC_QOPT_MAX_QUEUE]; 28 }; 29 30 static void mqprio_destroy(struct Qdisc *sch) 31 { 32 struct net_device *dev = qdisc_dev(sch); 33 struct mqprio_sched *priv = qdisc_priv(sch); 34 unsigned int ntx; 35 36 if (priv->qdiscs) { 37 for (ntx = 0; 38 ntx < dev->num_tx_queues && priv->qdiscs[ntx]; 39 ntx++) 40 qdisc_put(priv->qdiscs[ntx]); 41 kfree(priv->qdiscs); 42 } 43 44 if (priv->hw_offload && dev->netdev_ops->ndo_setup_tc) { 45 struct tc_mqprio_qopt_offload mqprio = { { 0 } }; 46 47 switch (priv->mode) { 48 case TC_MQPRIO_MODE_DCB: 49 case TC_MQPRIO_MODE_CHANNEL: 50 dev->netdev_ops->ndo_setup_tc(dev, 51 TC_SETUP_QDISC_MQPRIO, 52 &mqprio); 53 break; 54 default: 55 return; 56 } 57 } else { 58 netdev_set_num_tc(dev, 0); 59 } 60 } 61 62 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt) 63 { 64 int i, j; 65 66 /* Verify num_tc is not out of max range */ 67 if (qopt->num_tc > TC_MAX_QUEUE) 68 return -EINVAL; 69 70 /* Verify priority mapping uses valid tcs */ 71 for (i = 0; i < TC_BITMASK + 1; i++) { 72 if (qopt->prio_tc_map[i] >= qopt->num_tc) 73 return -EINVAL; 74 } 75 76 /* Limit qopt->hw to maximum supported offload value. Drivers have 77 * the option of overriding this later if they don't support the a 78 * given offload type. 79 */ 80 if (qopt->hw > TC_MQPRIO_HW_OFFLOAD_MAX) 81 qopt->hw = TC_MQPRIO_HW_OFFLOAD_MAX; 82 83 /* If hardware offload is requested we will leave it to the device 84 * to either populate the queue counts itself or to validate the 85 * provided queue counts. If ndo_setup_tc is not present then 86 * hardware doesn't support offload and we should return an error. 87 */ 88 if (qopt->hw) 89 return dev->netdev_ops->ndo_setup_tc ? 0 : -EINVAL; 90 91 for (i = 0; i < qopt->num_tc; i++) { 92 unsigned int last = qopt->offset[i] + qopt->count[i]; 93 94 /* Verify the queue count is in tx range being equal to the 95 * real_num_tx_queues indicates the last queue is in use. 96 */ 97 if (qopt->offset[i] >= dev->real_num_tx_queues || 98 !qopt->count[i] || 99 last > dev->real_num_tx_queues) 100 return -EINVAL; 101 102 /* Verify that the offset and counts do not overlap */ 103 for (j = i + 1; j < qopt->num_tc; j++) { 104 if (last > qopt->offset[j]) 105 return -EINVAL; 106 } 107 } 108 109 return 0; 110 } 111 112 static const struct nla_policy mqprio_policy[TCA_MQPRIO_MAX + 1] = { 113 [TCA_MQPRIO_MODE] = { .len = sizeof(u16) }, 114 [TCA_MQPRIO_SHAPER] = { .len = sizeof(u16) }, 115 [TCA_MQPRIO_MIN_RATE64] = { .type = NLA_NESTED }, 116 [TCA_MQPRIO_MAX_RATE64] = { .type = NLA_NESTED }, 117 }; 118 119 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 120 const struct nla_policy *policy, int len) 121 { 122 int nested_len = nla_len(nla) - NLA_ALIGN(len); 123 124 if (nested_len >= nla_attr_size(0)) 125 return nla_parse_deprecated(tb, maxtype, 126 nla_data(nla) + NLA_ALIGN(len), 127 nested_len, policy, NULL); 128 129 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 130 return 0; 131 } 132 133 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt, 134 struct netlink_ext_ack *extack) 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)), extack); 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_QDISC_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_put(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, struct netlink_ext_ack *extack) 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_noflag(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_noflag(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 ntx, tc; 391 392 sch->q.qlen = 0; 393 gnet_stats_basic_sync_init(&sch->bstats); 394 memset(&sch->qstats, 0, sizeof(sch->qstats)); 395 396 /* MQ supports lockless qdiscs. However, statistics accounting needs 397 * to account for all, none, or a mix of locked and unlocked child 398 * qdiscs. Percpu stats are added to counters in-band and locking 399 * qdisc totals are added at end. 400 */ 401 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) { 402 qdisc = netdev_get_tx_queue(dev, ntx)->qdisc_sleeping; 403 spin_lock_bh(qdisc_lock(qdisc)); 404 405 gnet_stats_add_basic(&sch->bstats, qdisc->cpu_bstats, 406 &qdisc->bstats, false); 407 gnet_stats_add_queue(&sch->qstats, qdisc->cpu_qstats, 408 &qdisc->qstats); 409 sch->q.qlen += qdisc_qlen(qdisc); 410 411 spin_unlock_bh(qdisc_lock(qdisc)); 412 } 413 414 opt.num_tc = netdev_get_num_tc(dev); 415 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map)); 416 opt.hw = priv->hw_offload; 417 418 for (tc = 0; tc < netdev_get_num_tc(dev); tc++) { 419 opt.count[tc] = dev->tc_to_txq[tc].count; 420 opt.offset[tc] = dev->tc_to_txq[tc].offset; 421 } 422 423 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 424 goto nla_put_failure; 425 426 if ((priv->flags & TC_MQPRIO_F_MODE) && 427 nla_put_u16(skb, TCA_MQPRIO_MODE, priv->mode)) 428 goto nla_put_failure; 429 430 if ((priv->flags & TC_MQPRIO_F_SHAPER) && 431 nla_put_u16(skb, TCA_MQPRIO_SHAPER, priv->shaper)) 432 goto nla_put_failure; 433 434 if ((priv->flags & TC_MQPRIO_F_MIN_RATE || 435 priv->flags & TC_MQPRIO_F_MAX_RATE) && 436 (dump_rates(priv, &opt, skb) != 0)) 437 goto nla_put_failure; 438 439 return nla_nest_end(skb, nla); 440 nla_put_failure: 441 nlmsg_trim(skb, nla); 442 return -1; 443 } 444 445 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl) 446 { 447 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 448 449 if (!dev_queue) 450 return NULL; 451 452 return dev_queue->qdisc_sleeping; 453 } 454 455 static unsigned long mqprio_find(struct Qdisc *sch, u32 classid) 456 { 457 struct net_device *dev = qdisc_dev(sch); 458 unsigned int ntx = TC_H_MIN(classid); 459 460 /* There are essentially two regions here that have valid classid 461 * values. The first region will have a classid value of 1 through 462 * num_tx_queues. All of these are backed by actual Qdiscs. 463 */ 464 if (ntx < TC_H_MIN_PRIORITY) 465 return (ntx <= dev->num_tx_queues) ? ntx : 0; 466 467 /* The second region represents the hardware traffic classes. These 468 * are represented by classid values of TC_H_MIN_PRIORITY through 469 * TC_H_MIN_PRIORITY + netdev_get_num_tc - 1 470 */ 471 return ((ntx - TC_H_MIN_PRIORITY) < netdev_get_num_tc(dev)) ? ntx : 0; 472 } 473 474 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl, 475 struct sk_buff *skb, struct tcmsg *tcm) 476 { 477 if (cl < TC_H_MIN_PRIORITY) { 478 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 479 struct net_device *dev = qdisc_dev(sch); 480 int tc = netdev_txq_to_tc(dev, cl - 1); 481 482 tcm->tcm_parent = (tc < 0) ? 0 : 483 TC_H_MAKE(TC_H_MAJ(sch->handle), 484 TC_H_MIN(tc + TC_H_MIN_PRIORITY)); 485 tcm->tcm_info = dev_queue->qdisc_sleeping->handle; 486 } else { 487 tcm->tcm_parent = TC_H_ROOT; 488 tcm->tcm_info = 0; 489 } 490 tcm->tcm_handle |= TC_H_MIN(cl); 491 return 0; 492 } 493 494 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl, 495 struct gnet_dump *d) 496 __releases(d->lock) 497 __acquires(d->lock) 498 { 499 if (cl >= TC_H_MIN_PRIORITY) { 500 int i; 501 __u32 qlen; 502 struct gnet_stats_queue qstats = {0}; 503 struct gnet_stats_basic_sync bstats; 504 struct net_device *dev = qdisc_dev(sch); 505 struct netdev_tc_txq tc = dev->tc_to_txq[cl & TC_BITMASK]; 506 507 gnet_stats_basic_sync_init(&bstats); 508 /* Drop lock here it will be reclaimed before touching 509 * statistics this is required because the d->lock we 510 * hold here is the look on dev_queue->qdisc_sleeping 511 * also acquired below. 512 */ 513 if (d->lock) 514 spin_unlock_bh(d->lock); 515 516 for (i = tc.offset; i < tc.offset + tc.count; i++) { 517 struct netdev_queue *q = netdev_get_tx_queue(dev, i); 518 struct Qdisc *qdisc = rtnl_dereference(q->qdisc); 519 520 spin_lock_bh(qdisc_lock(qdisc)); 521 522 gnet_stats_add_basic(&bstats, qdisc->cpu_bstats, 523 &qdisc->bstats, false); 524 gnet_stats_add_queue(&qstats, qdisc->cpu_qstats, 525 &qdisc->qstats); 526 sch->q.qlen += qdisc_qlen(qdisc); 527 528 spin_unlock_bh(qdisc_lock(qdisc)); 529 } 530 qlen = qdisc_qlen(sch) + qstats.qlen; 531 532 /* Reclaim root sleeping lock before completing stats */ 533 if (d->lock) 534 spin_lock_bh(d->lock); 535 if (gnet_stats_copy_basic(d, NULL, &bstats, false) < 0 || 536 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0) 537 return -1; 538 } else { 539 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 540 541 sch = dev_queue->qdisc_sleeping; 542 if (gnet_stats_copy_basic(d, sch->cpu_bstats, 543 &sch->bstats, true) < 0 || 544 qdisc_qstats_copy(d, sch) < 0) 545 return -1; 546 } 547 return 0; 548 } 549 550 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg) 551 { 552 struct net_device *dev = qdisc_dev(sch); 553 unsigned long ntx; 554 555 if (arg->stop) 556 return; 557 558 /* Walk hierarchy with a virtual class per tc */ 559 arg->count = arg->skip; 560 for (ntx = arg->skip; ntx < netdev_get_num_tc(dev); ntx++) { 561 if (arg->fn(sch, ntx + TC_H_MIN_PRIORITY, arg) < 0) { 562 arg->stop = 1; 563 return; 564 } 565 arg->count++; 566 } 567 568 /* Pad the values and skip over unused traffic classes */ 569 if (ntx < TC_MAX_QUEUE) { 570 arg->count = TC_MAX_QUEUE; 571 ntx = TC_MAX_QUEUE; 572 } 573 574 /* Reset offset, sort out remaining per-queue qdiscs */ 575 for (ntx -= TC_MAX_QUEUE; ntx < dev->num_tx_queues; ntx++) { 576 if (arg->fn(sch, ntx + 1, arg) < 0) { 577 arg->stop = 1; 578 return; 579 } 580 arg->count++; 581 } 582 } 583 584 static struct netdev_queue *mqprio_select_queue(struct Qdisc *sch, 585 struct tcmsg *tcm) 586 { 587 return mqprio_queue_get(sch, TC_H_MIN(tcm->tcm_parent)); 588 } 589 590 static const struct Qdisc_class_ops mqprio_class_ops = { 591 .graft = mqprio_graft, 592 .leaf = mqprio_leaf, 593 .find = mqprio_find, 594 .walk = mqprio_walk, 595 .dump = mqprio_dump_class, 596 .dump_stats = mqprio_dump_class_stats, 597 .select_queue = mqprio_select_queue, 598 }; 599 600 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = { 601 .cl_ops = &mqprio_class_ops, 602 .id = "mqprio", 603 .priv_size = sizeof(struct mqprio_sched), 604 .init = mqprio_init, 605 .destroy = mqprio_destroy, 606 .attach = mqprio_attach, 607 .change_real_num_tx = mq_change_real_num_tx, 608 .dump = mqprio_dump, 609 .owner = THIS_MODULE, 610 }; 611 612 static int __init mqprio_module_init(void) 613 { 614 return register_qdisc(&mqprio_qdisc_ops); 615 } 616 617 static void __exit mqprio_module_exit(void) 618 { 619 unregister_qdisc(&mqprio_qdisc_ops); 620 } 621 622 module_init(mqprio_module_init); 623 module_exit(mqprio_module_exit); 624 625 MODULE_LICENSE("GPL"); 626