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