1 /* 2 * net/sched/sch_cbs.c Credit Based Shaper 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License, or (at your option) any later version. 8 * 9 * Authors: Vinicius Costa Gomes <vinicius.gomes@intel.com> 10 * 11 */ 12 13 /* Credit Based Shaper (CBS) 14 * ========================= 15 * 16 * This is a simple rate-limiting shaper aimed at TSN applications on 17 * systems with known traffic workloads. 18 * 19 * Its algorithm is defined by the IEEE 802.1Q-2014 Specification, 20 * Section 8.6.8.2, and explained in more detail in the Annex L of the 21 * same specification. 22 * 23 * There are four tunables to be considered: 24 * 25 * 'idleslope': Idleslope is the rate of credits that is 26 * accumulated (in kilobits per second) when there is at least 27 * one packet waiting for transmission. Packets are transmitted 28 * when the current value of credits is equal or greater than 29 * zero. When there is no packet to be transmitted the amount of 30 * credits is set to zero. This is the main tunable of the CBS 31 * algorithm. 32 * 33 * 'sendslope': 34 * Sendslope is the rate of credits that is depleted (it should be a 35 * negative number of kilobits per second) when a transmission is 36 * ocurring. It can be calculated as follows, (IEEE 802.1Q-2014 Section 37 * 8.6.8.2 item g): 38 * 39 * sendslope = idleslope - port_transmit_rate 40 * 41 * 'hicredit': Hicredit defines the maximum amount of credits (in 42 * bytes) that can be accumulated. Hicredit depends on the 43 * characteristics of interfering traffic, 44 * 'max_interference_size' is the maximum size of any burst of 45 * traffic that can delay the transmission of a frame that is 46 * available for transmission for this traffic class, (IEEE 47 * 802.1Q-2014 Annex L, Equation L-3): 48 * 49 * hicredit = max_interference_size * (idleslope / port_transmit_rate) 50 * 51 * 'locredit': Locredit is the minimum amount of credits that can 52 * be reached. It is a function of the traffic flowing through 53 * this qdisc (IEEE 802.1Q-2014 Annex L, Equation L-2): 54 * 55 * locredit = max_frame_size * (sendslope / port_transmit_rate) 56 */ 57 58 #include <linux/module.h> 59 #include <linux/types.h> 60 #include <linux/kernel.h> 61 #include <linux/string.h> 62 #include <linux/errno.h> 63 #include <linux/skbuff.h> 64 #include <net/netlink.h> 65 #include <net/sch_generic.h> 66 #include <net/pkt_sched.h> 67 68 #define BYTES_PER_KBIT (1000LL / 8) 69 70 struct cbs_sched_data { 71 bool offload; 72 int queue; 73 s64 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 *(*dequeue)(struct Qdisc *sch); 83 }; 84 85 static int cbs_enqueue_offload(struct sk_buff *skb, struct Qdisc *sch) 86 { 87 return qdisc_enqueue_tail(skb, sch); 88 } 89 90 static int cbs_enqueue_soft(struct sk_buff *skb, struct Qdisc *sch) 91 { 92 struct cbs_sched_data *q = qdisc_priv(sch); 93 94 if (sch->q.qlen == 0 && q->credits > 0) { 95 /* We need to stop accumulating credits when there's 96 * no enqueued packets and q->credits is positive. 97 */ 98 q->credits = 0; 99 q->last = ktime_get_ns(); 100 } 101 102 return qdisc_enqueue_tail(skb, sch); 103 } 104 105 static int cbs_enqueue(struct sk_buff *skb, struct Qdisc *sch, 106 struct sk_buff **to_free) 107 { 108 struct cbs_sched_data *q = qdisc_priv(sch); 109 110 return q->enqueue(skb, sch); 111 } 112 113 /* timediff is in ns, slope is in bytes/s */ 114 static s64 timediff_to_credits(s64 timediff, s64 slope) 115 { 116 return div64_s64(timediff * slope, NSEC_PER_SEC); 117 } 118 119 static s64 delay_from_credits(s64 credits, s64 slope) 120 { 121 if (unlikely(slope == 0)) 122 return S64_MAX; 123 124 return div64_s64(-credits * NSEC_PER_SEC, slope); 125 } 126 127 static s64 credits_from_len(unsigned int len, s64 slope, s64 port_rate) 128 { 129 if (unlikely(port_rate == 0)) 130 return S64_MAX; 131 132 return div64_s64(len * slope, port_rate); 133 } 134 135 static struct sk_buff *cbs_dequeue_soft(struct Qdisc *sch) 136 { 137 struct cbs_sched_data *q = qdisc_priv(sch); 138 s64 now = ktime_get_ns(); 139 struct sk_buff *skb; 140 s64 credits; 141 int len; 142 143 if (q->credits < 0) { 144 credits = timediff_to_credits(now - q->last, q->idleslope); 145 146 credits = q->credits + credits; 147 q->credits = min_t(s64, credits, q->hicredit); 148 149 if (q->credits < 0) { 150 s64 delay; 151 152 delay = delay_from_credits(q->credits, q->idleslope); 153 qdisc_watchdog_schedule_ns(&q->watchdog, now + delay); 154 155 q->last = now; 156 157 return NULL; 158 } 159 } 160 161 skb = qdisc_dequeue_head(sch); 162 if (!skb) 163 return NULL; 164 165 len = qdisc_pkt_len(skb); 166 167 /* As sendslope is a negative number, this will decrease the 168 * amount of q->credits. 169 */ 170 credits = credits_from_len(len, q->sendslope, q->port_rate); 171 credits += q->credits; 172 173 q->credits = max_t(s64, credits, q->locredit); 174 q->last = now; 175 176 return skb; 177 } 178 179 static struct sk_buff *cbs_dequeue_offload(struct Qdisc *sch) 180 { 181 return qdisc_dequeue_head(sch); 182 } 183 184 static struct sk_buff *cbs_dequeue(struct Qdisc *sch) 185 { 186 struct cbs_sched_data *q = qdisc_priv(sch); 187 188 return q->dequeue(sch); 189 } 190 191 static const struct nla_policy cbs_policy[TCA_CBS_MAX + 1] = { 192 [TCA_CBS_PARMS] = { .len = sizeof(struct tc_cbs_qopt) }, 193 }; 194 195 static void cbs_disable_offload(struct net_device *dev, 196 struct cbs_sched_data *q) 197 { 198 struct tc_cbs_qopt_offload cbs = { }; 199 const struct net_device_ops *ops; 200 int err; 201 202 if (!q->offload) 203 return; 204 205 q->enqueue = cbs_enqueue_soft; 206 q->dequeue = cbs_dequeue_soft; 207 208 ops = dev->netdev_ops; 209 if (!ops->ndo_setup_tc) 210 return; 211 212 cbs.queue = q->queue; 213 cbs.enable = 0; 214 215 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs); 216 if (err < 0) 217 pr_warn("Couldn't disable CBS offload for queue %d\n", 218 cbs.queue); 219 } 220 221 static int cbs_enable_offload(struct net_device *dev, struct cbs_sched_data *q, 222 const struct tc_cbs_qopt *opt) 223 { 224 const struct net_device_ops *ops = dev->netdev_ops; 225 struct tc_cbs_qopt_offload cbs = { }; 226 int err; 227 228 if (!ops->ndo_setup_tc) 229 return -EOPNOTSUPP; 230 231 cbs.queue = q->queue; 232 233 cbs.enable = 1; 234 cbs.hicredit = opt->hicredit; 235 cbs.locredit = opt->locredit; 236 cbs.idleslope = opt->idleslope; 237 cbs.sendslope = opt->sendslope; 238 239 err = ops->ndo_setup_tc(dev, TC_SETUP_QDISC_CBS, &cbs); 240 if (err < 0) 241 return err; 242 243 q->enqueue = cbs_enqueue_offload; 244 q->dequeue = cbs_dequeue_offload; 245 246 return 0; 247 } 248 249 static int cbs_change(struct Qdisc *sch, struct nlattr *opt) 250 { 251 struct cbs_sched_data *q = qdisc_priv(sch); 252 struct net_device *dev = qdisc_dev(sch); 253 struct nlattr *tb[TCA_CBS_MAX + 1]; 254 struct tc_cbs_qopt *qopt; 255 int err; 256 257 err = nla_parse_nested(tb, TCA_CBS_MAX, opt, cbs_policy, NULL); 258 if (err < 0) 259 return err; 260 261 if (!tb[TCA_CBS_PARMS]) 262 return -EINVAL; 263 264 qopt = nla_data(tb[TCA_CBS_PARMS]); 265 266 if (!qopt->offload) { 267 struct ethtool_link_ksettings ecmd; 268 s64 link_speed; 269 270 if (!__ethtool_get_link_ksettings(dev, &ecmd)) 271 link_speed = ecmd.base.speed; 272 else 273 link_speed = SPEED_1000; 274 275 q->port_rate = link_speed * 1000 * BYTES_PER_KBIT; 276 277 cbs_disable_offload(dev, q); 278 } else { 279 err = cbs_enable_offload(dev, q, qopt); 280 if (err < 0) 281 return err; 282 } 283 284 /* Everything went OK, save the parameters used. */ 285 q->hicredit = qopt->hicredit; 286 q->locredit = qopt->locredit; 287 q->idleslope = qopt->idleslope * BYTES_PER_KBIT; 288 q->sendslope = qopt->sendslope * BYTES_PER_KBIT; 289 q->offload = qopt->offload; 290 291 return 0; 292 } 293 294 static int cbs_init(struct Qdisc *sch, struct nlattr *opt) 295 { 296 struct cbs_sched_data *q = qdisc_priv(sch); 297 struct net_device *dev = qdisc_dev(sch); 298 299 if (!opt) 300 return -EINVAL; 301 302 q->queue = sch->dev_queue - netdev_get_tx_queue(dev, 0); 303 304 q->enqueue = cbs_enqueue_soft; 305 q->dequeue = cbs_dequeue_soft; 306 307 qdisc_watchdog_init(&q->watchdog, sch); 308 309 return cbs_change(sch, opt); 310 } 311 312 static void cbs_destroy(struct Qdisc *sch) 313 { 314 struct cbs_sched_data *q = qdisc_priv(sch); 315 struct net_device *dev = qdisc_dev(sch); 316 317 qdisc_watchdog_cancel(&q->watchdog); 318 319 cbs_disable_offload(dev, q); 320 } 321 322 static int cbs_dump(struct Qdisc *sch, struct sk_buff *skb) 323 { 324 struct cbs_sched_data *q = qdisc_priv(sch); 325 struct tc_cbs_qopt opt = { }; 326 struct nlattr *nest; 327 328 nest = nla_nest_start(skb, TCA_OPTIONS); 329 if (!nest) 330 goto nla_put_failure; 331 332 opt.hicredit = q->hicredit; 333 opt.locredit = q->locredit; 334 opt.sendslope = div64_s64(q->sendslope, BYTES_PER_KBIT); 335 opt.idleslope = div64_s64(q->idleslope, BYTES_PER_KBIT); 336 opt.offload = q->offload; 337 338 if (nla_put(skb, TCA_CBS_PARMS, sizeof(opt), &opt)) 339 goto nla_put_failure; 340 341 return nla_nest_end(skb, nest); 342 343 nla_put_failure: 344 nla_nest_cancel(skb, nest); 345 return -1; 346 } 347 348 static struct Qdisc_ops cbs_qdisc_ops __read_mostly = { 349 .id = "cbs", 350 .priv_size = sizeof(struct cbs_sched_data), 351 .enqueue = cbs_enqueue, 352 .dequeue = cbs_dequeue, 353 .peek = qdisc_peek_dequeued, 354 .init = cbs_init, 355 .reset = qdisc_reset_queue, 356 .destroy = cbs_destroy, 357 .change = cbs_change, 358 .dump = cbs_dump, 359 .owner = THIS_MODULE, 360 }; 361 362 static int __init cbs_module_init(void) 363 { 364 return register_qdisc(&cbs_qdisc_ops); 365 } 366 367 static void __exit cbs_module_exit(void) 368 { 369 unregister_qdisc(&cbs_qdisc_ops); 370 } 371 module_init(cbs_module_init) 372 module_exit(cbs_module_exit) 373 MODULE_LICENSE("GPL"); 374