1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * net/sched/sch_cbs.c Credit Based Shaper 4 * 5 * Authors: Vinicius Costa Gomes <vinicius.gomes@intel.com> 6 */ 7 8 /* Credit Based Shaper (CBS) 9 * ========================= 10 * 11 * This is a simple rate-limiting shaper aimed at TSN applications on 12 * systems with known traffic workloads. 13 * 14 * Its algorithm is defined by the IEEE 802.1Q-2014 Specification, 15 * Section 8.6.8.2, and explained in more detail in the Annex L of the 16 * same specification. 17 * 18 * There are four tunables to be considered: 19 * 20 * 'idleslope': Idleslope is the rate of credits that is 21 * accumulated (in kilobits per second) when there is at least 22 * one packet waiting for transmission. Packets are transmitted 23 * when the current value of credits is equal or greater than 24 * zero. When there is no packet to be transmitted the amount of 25 * credits is set to zero. This is the main tunable of the CBS 26 * algorithm. 27 * 28 * 'sendslope': 29 * Sendslope is the rate of credits that is depleted (it should be a 30 * negative number of kilobits per second) when a transmission is 31 * ocurring. It can be calculated as follows, (IEEE 802.1Q-2014 Section 32 * 8.6.8.2 item g): 33 * 34 * sendslope = idleslope - port_transmit_rate 35 * 36 * 'hicredit': Hicredit defines the maximum amount of credits (in 37 * bytes) that can be accumulated. Hicredit depends on the 38 * characteristics of interfering traffic, 39 * 'max_interference_size' is the maximum size of any burst of 40 * traffic that can delay the transmission of a frame that is 41 * available for transmission for this traffic class, (IEEE 42 * 802.1Q-2014 Annex L, Equation L-3): 43 * 44 * hicredit = max_interference_size * (idleslope / port_transmit_rate) 45 * 46 * 'locredit': Locredit is the minimum amount of credits that can 47 * be reached. It is a function of the traffic flowing through 48 * this qdisc (IEEE 802.1Q-2014 Annex L, Equation L-2): 49 * 50 * locredit = max_frame_size * (sendslope / port_transmit_rate) 51 */ 52 53 #include <linux/ethtool.h> 54 #include <linux/module.h> 55 #include <linux/types.h> 56 #include <linux/kernel.h> 57 #include <linux/string.h> 58 #include <linux/errno.h> 59 #include <linux/skbuff.h> 60 #include <linux/units.h> 61 62 #include <net/netevent.h> 63 #include <net/netlink.h> 64 #include <net/sch_generic.h> 65 #include <net/pkt_sched.h> 66 67 static LIST_HEAD(cbs_list); 68 static DEFINE_SPINLOCK(cbs_list_lock); 69 70 struct cbs_sched_data { 71 bool offload; 72 int queue; 73 atomic64_t port_rate; /* in bytes/s */ 74 s64 last; /* timestamp in ns */ 75 s64 credits; /* in bytes */ 76 s32 locredit; /* in bytes */ 77 s32 hicredit; /* in bytes */ 78 s64 sendslope; /* in bytes/s */ 79 s64 idleslope; /* in bytes/s */ 80 struct qdisc_watchdog watchdog; 81 int (*enqueue)(struct sk_buff *skb, struct Qdisc *sch, 82 struct sk_buff **to_free); 83 struct sk_buff *(*dequeue)(struct Qdisc *sch); 84 struct Qdisc *qdisc; 85 struct list_head cbs_list; 86 }; 87 88 static int cbs_child_enqueue(struct sk_buff *skb, struct Qdisc *sch, 89 struct Qdisc *child, 90 struct sk_buff **to_free) 91 { 92 unsigned int len = qdisc_pkt_len(skb); 93 int err; 94 95 err = child->ops->enqueue(skb, child, to_free); 96 if (err != NET_XMIT_SUCCESS) 97 return err; 98 99 sch->qstats.backlog += len; 100 sch->q.qlen++; 101 102 return NET_XMIT_SUCCESS; 103 } 104 105 static int cbs_enqueue_offload(struct sk_buff *skb, struct Qdisc *sch, 106 struct sk_buff **to_free) 107 { 108 struct cbs_sched_data *q = qdisc_priv(sch); 109 struct Qdisc *qdisc = q->qdisc; 110 111 return cbs_child_enqueue(skb, sch, qdisc, to_free); 112 } 113 114 static int cbs_enqueue_soft(struct sk_buff *skb, struct Qdisc *sch, 115 struct sk_buff **to_free) 116 { 117 struct cbs_sched_data *q = qdisc_priv(sch); 118 struct Qdisc *qdisc = q->qdisc; 119 120 if (sch->q.qlen == 0 && q->credits > 0) { 121 /* We need to stop accumulating credits when there's 122 * no enqueued packets and q->credits is positive. 123 */ 124 q->credits = 0; 125 q->last = ktime_get_ns(); 126 } 127 128 return cbs_child_enqueue(skb, sch, qdisc, to_free); 129 } 130 131 static int cbs_enqueue(struct sk_buff *skb, struct Qdisc *sch, 132 struct sk_buff **to_free) 133 { 134 struct cbs_sched_data *q = qdisc_priv(sch); 135 136 return q->enqueue(skb, sch, to_free); 137 } 138 139 /* timediff is in ns, slope is in bytes/s */ 140 static s64 timediff_to_credits(s64 timediff, s64 slope) 141 { 142 return div64_s64(timediff * slope, NSEC_PER_SEC); 143 } 144 145 static s64 delay_from_credits(s64 credits, s64 slope) 146 { 147 if (unlikely(slope == 0)) 148 return S64_MAX; 149 150 return div64_s64(-credits * NSEC_PER_SEC, slope); 151 } 152 153 static s64 credits_from_len(unsigned int len, s64 slope, s64 port_rate) 154 { 155 if (unlikely(port_rate == 0)) 156 return S64_MAX; 157 158 return div64_s64(len * slope, port_rate); 159 } 160 161 static struct sk_buff *cbs_child_dequeue(struct Qdisc *sch, struct Qdisc *child) 162 { 163 struct sk_buff *skb; 164 165 skb = child->ops->dequeue(child); 166 if (!skb) 167 return NULL; 168 169 qdisc_qstats_backlog_dec(sch, skb); 170 qdisc_bstats_update(sch, skb); 171 sch->q.qlen--; 172 173 return skb; 174 } 175 176 static struct sk_buff *cbs_dequeue_soft(struct Qdisc *sch) 177 { 178 struct cbs_sched_data *q = qdisc_priv(sch); 179 struct Qdisc *qdisc = q->qdisc; 180 s64 now = ktime_get_ns(); 181 struct sk_buff *skb; 182 s64 credits; 183 int len; 184 185 /* The previous packet is still being sent */ 186 if (now < q->last) { 187 qdisc_watchdog_schedule_ns(&q->watchdog, q->last); 188 return NULL; 189 } 190 if (q->credits < 0) { 191 credits = timediff_to_credits(now - q->last, q->idleslope); 192 193 credits = q->credits + credits; 194 q->credits = min_t(s64, credits, q->hicredit); 195 196 if (q->credits < 0) { 197 s64 delay; 198 199 delay = delay_from_credits(q->credits, q->idleslope); 200 qdisc_watchdog_schedule_ns(&q->watchdog, now + delay); 201 202 q->last = now; 203 204 return NULL; 205 } 206 } 207 skb = cbs_child_dequeue(sch, qdisc); 208 if (!skb) 209 return NULL; 210 211 len = qdisc_pkt_len(skb); 212 213 /* As sendslope is a negative number, this will decrease the 214 * amount of q->credits. 215 */ 216 credits = credits_from_len(len, q->sendslope, 217 atomic64_read(&q->port_rate)); 218 credits += q->credits; 219 220 q->credits = max_t(s64, credits, q->locredit); 221 /* Estimate of the transmission of the last byte of the packet in ns */ 222 if (unlikely(atomic64_read(&q->port_rate) == 0)) 223 q->last = now; 224 else 225 q->last = now + div64_s64(len * NSEC_PER_SEC, 226 atomic64_read(&q->port_rate)); 227 228 return skb; 229 } 230 231 static struct sk_buff *cbs_dequeue_offload(struct Qdisc *sch) 232 { 233 struct cbs_sched_data *q = qdisc_priv(sch); 234 struct Qdisc *qdisc = q->qdisc; 235 236 return cbs_child_dequeue(sch, qdisc); 237 } 238 239 static struct sk_buff *cbs_dequeue(struct Qdisc *sch) 240 { 241 struct cbs_sched_data *q = qdisc_priv(sch); 242 243 return q->dequeue(sch); 244 } 245 246 static void cbs_reset(struct Qdisc *sch) 247 { 248 struct cbs_sched_data *q = qdisc_priv(sch); 249 250 /* Nothing to do if we couldn't create the underlying qdisc */ 251 if (!q->qdisc) 252 return; 253 254 qdisc_reset(q->qdisc); 255 qdisc_watchdog_cancel(&q->watchdog); 256 q->credits = 0; 257 q->last = 0; 258 } 259 260 static const struct nla_policy cbs_policy[TCA_CBS_MAX + 1] = { 261 [TCA_CBS_PARMS] = { .len = sizeof(struct tc_cbs_qopt) }, 262 }; 263 264 static void cbs_disable_offload(struct net_device *dev, 265 struct cbs_sched_data *q) 266 { 267 struct tc_cbs_qopt_offload cbs = { }; 268 const struct net_device_ops *ops; 269 int err; 270 271 if (!q->offload) 272 return; 273 274 q->enqueue = cbs_enqueue_soft; 275 q->dequeue = cbs_dequeue_soft; 276 277 ops = dev->netdev_ops; 278 if (!ops->ndo_setup_tc) 279 return; 280 281 cbs.queue = q->queue; 282 cbs.enable = 0; 283 284 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs); 285 if (err < 0) 286 pr_warn("Couldn't disable CBS offload for queue %d\n", 287 cbs.queue); 288 } 289 290 static int cbs_enable_offload(struct net_device *dev, struct cbs_sched_data *q, 291 const struct tc_cbs_qopt *opt, 292 struct netlink_ext_ack *extack) 293 { 294 const struct net_device_ops *ops = dev->netdev_ops; 295 struct tc_cbs_qopt_offload cbs = { }; 296 int err; 297 298 if (!ops->ndo_setup_tc) { 299 NL_SET_ERR_MSG(extack, "Specified device does not support cbs offload"); 300 return -EOPNOTSUPP; 301 } 302 303 cbs.queue = q->queue; 304 305 cbs.enable = 1; 306 cbs.hicredit = opt->hicredit; 307 cbs.locredit = opt->locredit; 308 cbs.idleslope = opt->idleslope; 309 cbs.sendslope = opt->sendslope; 310 311 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs); 312 if (err < 0) { 313 NL_SET_ERR_MSG(extack, "Specified device failed to setup cbs hardware offload"); 314 return err; 315 } 316 317 q->enqueue = cbs_enqueue_offload; 318 q->dequeue = cbs_dequeue_offload; 319 320 return 0; 321 } 322 323 static void cbs_set_port_rate(struct net_device *dev, struct cbs_sched_data *q) 324 { 325 struct ethtool_link_ksettings ecmd; 326 int speed = SPEED_10; 327 s64 port_rate; 328 int err; 329 330 err = __ethtool_get_link_ksettings(dev, &ecmd); 331 if (err < 0) 332 goto skip; 333 334 if (ecmd.base.speed && ecmd.base.speed != SPEED_UNKNOWN) 335 speed = ecmd.base.speed; 336 337 skip: 338 port_rate = speed * 1000 * BYTES_PER_KBIT; 339 340 atomic64_set(&q->port_rate, port_rate); 341 netdev_dbg(dev, "cbs: set %s's port_rate to: %lld, linkspeed: %d\n", 342 dev->name, (long long)atomic64_read(&q->port_rate), 343 ecmd.base.speed); 344 } 345 346 static int cbs_dev_notifier(struct notifier_block *nb, unsigned long event, 347 void *ptr) 348 { 349 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 350 struct cbs_sched_data *q; 351 struct net_device *qdev; 352 bool found = false; 353 354 ASSERT_RTNL(); 355 356 if (event != NETDEV_UP && event != NETDEV_CHANGE) 357 return NOTIFY_DONE; 358 359 spin_lock(&cbs_list_lock); 360 list_for_each_entry(q, &cbs_list, cbs_list) { 361 qdev = qdisc_dev(q->qdisc); 362 if (qdev == dev) { 363 found = true; 364 break; 365 } 366 } 367 spin_unlock(&cbs_list_lock); 368 369 if (found) 370 cbs_set_port_rate(dev, q); 371 372 return NOTIFY_DONE; 373 } 374 375 static int cbs_change(struct Qdisc *sch, struct nlattr *opt, 376 struct netlink_ext_ack *extack) 377 { 378 struct cbs_sched_data *q = qdisc_priv(sch); 379 struct net_device *dev = qdisc_dev(sch); 380 struct nlattr *tb[TCA_CBS_MAX + 1]; 381 struct tc_cbs_qopt *qopt; 382 int err; 383 384 err = nla_parse_nested_deprecated(tb, TCA_CBS_MAX, opt, cbs_policy, 385 extack); 386 if (err < 0) 387 return err; 388 389 if (!tb[TCA_CBS_PARMS]) { 390 NL_SET_ERR_MSG(extack, "Missing CBS parameter which are mandatory"); 391 return -EINVAL; 392 } 393 394 qopt = nla_data(tb[TCA_CBS_PARMS]); 395 396 if (!qopt->offload) { 397 cbs_set_port_rate(dev, q); 398 cbs_disable_offload(dev, q); 399 } else { 400 err = cbs_enable_offload(dev, q, qopt, extack); 401 if (err < 0) 402 return err; 403 } 404 405 /* Everything went OK, save the parameters used. */ 406 WRITE_ONCE(q->hicredit, qopt->hicredit); 407 WRITE_ONCE(q->locredit, qopt->locredit); 408 WRITE_ONCE(q->idleslope, qopt->idleslope * BYTES_PER_KBIT); 409 WRITE_ONCE(q->sendslope, qopt->sendslope * BYTES_PER_KBIT); 410 WRITE_ONCE(q->offload, qopt->offload); 411 412 return 0; 413 } 414 415 static int cbs_init(struct Qdisc *sch, struct nlattr *opt, 416 struct netlink_ext_ack *extack) 417 { 418 struct cbs_sched_data *q = qdisc_priv(sch); 419 struct net_device *dev = qdisc_dev(sch); 420 421 if (!opt) { 422 NL_SET_ERR_MSG(extack, "Missing CBS qdisc options which are mandatory"); 423 return -EINVAL; 424 } 425 426 q->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, 427 sch->handle, extack); 428 if (!q->qdisc) 429 return -ENOMEM; 430 431 spin_lock(&cbs_list_lock); 432 list_add(&q->cbs_list, &cbs_list); 433 spin_unlock(&cbs_list_lock); 434 435 qdisc_hash_add(q->qdisc, false); 436 437 q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0); 438 439 q->enqueue = cbs_enqueue_soft; 440 q->dequeue = cbs_dequeue_soft; 441 442 qdisc_watchdog_init(&q->watchdog, sch); 443 444 return cbs_change(sch, opt, extack); 445 } 446 447 static void cbs_destroy(struct Qdisc *sch) 448 { 449 struct cbs_sched_data *q = qdisc_priv(sch); 450 struct net_device *dev = qdisc_dev(sch); 451 452 /* Nothing to do if we couldn't create the underlying qdisc */ 453 if (!q->qdisc) 454 return; 455 456 qdisc_watchdog_cancel(&q->watchdog); 457 cbs_disable_offload(dev, q); 458 459 spin_lock(&cbs_list_lock); 460 list_del(&q->cbs_list); 461 spin_unlock(&cbs_list_lock); 462 463 qdisc_put(q->qdisc); 464 } 465 466 static int cbs_dump(struct Qdisc *sch, struct sk_buff *skb) 467 { 468 struct cbs_sched_data *q = qdisc_priv(sch); 469 struct tc_cbs_qopt opt = { }; 470 struct nlattr *nest; 471 472 nest = nla_nest_start_noflag(skb, TCA_OPTIONS); 473 if (!nest) 474 goto nla_put_failure; 475 476 opt.hicredit = READ_ONCE(q->hicredit); 477 opt.locredit = READ_ONCE(q->locredit); 478 opt.sendslope = div64_s64(READ_ONCE(q->sendslope), BYTES_PER_KBIT); 479 opt.idleslope = div64_s64(READ_ONCE(q->idleslope), BYTES_PER_KBIT); 480 opt.offload = READ_ONCE(q->offload); 481 482 if (nla_put(skb, TCA_CBS_PARMS, sizeof(opt), &opt)) 483 goto nla_put_failure; 484 485 return nla_nest_end(skb, nest); 486 487 nla_put_failure: 488 nla_nest_cancel(skb, nest); 489 return -1; 490 } 491 492 static int cbs_dump_class(struct Qdisc *sch, unsigned long cl, 493 struct sk_buff *skb, struct tcmsg *tcm) 494 { 495 struct cbs_sched_data *q = qdisc_priv(sch); 496 497 if (cl != 1 || !q->qdisc) /* only one class */ 498 return -ENOENT; 499 500 tcm->tcm_handle |= TC_H_MIN(1); 501 tcm->tcm_info = q->qdisc->handle; 502 503 return 0; 504 } 505 506 static int cbs_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 507 struct Qdisc **old, struct netlink_ext_ack *extack) 508 { 509 struct cbs_sched_data *q = qdisc_priv(sch); 510 511 if (!new) { 512 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, 513 sch->handle, NULL); 514 if (!new) 515 new = &noop_qdisc; 516 } 517 518 *old = qdisc_replace(sch, new, &q->qdisc); 519 return 0; 520 } 521 522 static struct Qdisc *cbs_leaf(struct Qdisc *sch, unsigned long arg) 523 { 524 struct cbs_sched_data *q = qdisc_priv(sch); 525 526 return q->qdisc; 527 } 528 529 static unsigned long cbs_find(struct Qdisc *sch, u32 classid) 530 { 531 return 1; 532 } 533 534 static void cbs_walk(struct Qdisc *sch, struct qdisc_walker *walker) 535 { 536 if (!walker->stop) { 537 tc_qdisc_stats_dump(sch, 1, walker); 538 } 539 } 540 541 static const struct Qdisc_class_ops cbs_class_ops = { 542 .graft = cbs_graft, 543 .leaf = cbs_leaf, 544 .find = cbs_find, 545 .walk = cbs_walk, 546 .dump = cbs_dump_class, 547 }; 548 549 static struct Qdisc_ops cbs_qdisc_ops __read_mostly = { 550 .id = "cbs", 551 .cl_ops = &cbs_class_ops, 552 .priv_size = sizeof(struct cbs_sched_data), 553 .enqueue = cbs_enqueue, 554 .dequeue = cbs_dequeue, 555 .peek = qdisc_peek_dequeued, 556 .init = cbs_init, 557 .reset = cbs_reset, 558 .destroy = cbs_destroy, 559 .change = cbs_change, 560 .dump = cbs_dump, 561 .owner = THIS_MODULE, 562 }; 563 MODULE_ALIAS_NET_SCH("cbs"); 564 565 static struct notifier_block cbs_device_notifier = { 566 .notifier_call = cbs_dev_notifier, 567 }; 568 569 static int __init cbs_module_init(void) 570 { 571 int err; 572 573 err = register_netdevice_notifier(&cbs_device_notifier); 574 if (err) 575 return err; 576 577 err = register_qdisc(&cbs_qdisc_ops); 578 if (err) 579 unregister_netdevice_notifier(&cbs_device_notifier); 580 581 return err; 582 } 583 584 static void __exit cbs_module_exit(void) 585 { 586 unregister_qdisc(&cbs_qdisc_ops); 587 unregister_netdevice_notifier(&cbs_device_notifier); 588 } 589 module_init(cbs_module_init) 590 module_exit(cbs_module_exit) 591 MODULE_LICENSE("GPL"); 592 MODULE_DESCRIPTION("Credit Based shaper"); 593