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 22 struct mqprio_sched { 23 struct Qdisc **qdiscs; 24 int hw_owned; 25 }; 26 27 static void mqprio_destroy(struct Qdisc *sch) 28 { 29 struct net_device *dev = qdisc_dev(sch); 30 struct mqprio_sched *priv = qdisc_priv(sch); 31 struct tc_to_netdev tc = {.type = TC_SETUP_MQPRIO}; 32 unsigned int ntx; 33 34 if (priv->qdiscs) { 35 for (ntx = 0; 36 ntx < dev->num_tx_queues && priv->qdiscs[ntx]; 37 ntx++) 38 qdisc_destroy(priv->qdiscs[ntx]); 39 kfree(priv->qdiscs); 40 } 41 42 if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc) 43 dev->netdev_ops->ndo_setup_tc(dev, sch->handle, 0, &tc); 44 else 45 netdev_set_num_tc(dev, 0); 46 } 47 48 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt) 49 { 50 int i, j; 51 52 /* Verify num_tc is not out of max range */ 53 if (qopt->num_tc > TC_MAX_QUEUE) 54 return -EINVAL; 55 56 /* Verify priority mapping uses valid tcs */ 57 for (i = 0; i < TC_BITMASK + 1; i++) { 58 if (qopt->prio_tc_map[i] >= qopt->num_tc) 59 return -EINVAL; 60 } 61 62 /* net_device does not support requested operation */ 63 if (qopt->hw && !dev->netdev_ops->ndo_setup_tc) 64 return -EINVAL; 65 66 /* if hw owned qcount and qoffset are taken from LLD so 67 * no reason to verify them here 68 */ 69 if (qopt->hw) 70 return 0; 71 72 for (i = 0; i < qopt->num_tc; i++) { 73 unsigned int last = qopt->offset[i] + qopt->count[i]; 74 75 /* Verify the queue count is in tx range being equal to the 76 * real_num_tx_queues indicates the last queue is in use. 77 */ 78 if (qopt->offset[i] >= dev->real_num_tx_queues || 79 !qopt->count[i] || 80 last > dev->real_num_tx_queues) 81 return -EINVAL; 82 83 /* Verify that the offset and counts do not overlap */ 84 for (j = i + 1; j < qopt->num_tc; j++) { 85 if (last > qopt->offset[j]) 86 return -EINVAL; 87 } 88 } 89 90 return 0; 91 } 92 93 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt) 94 { 95 struct net_device *dev = qdisc_dev(sch); 96 struct mqprio_sched *priv = qdisc_priv(sch); 97 struct netdev_queue *dev_queue; 98 struct Qdisc *qdisc; 99 int i, err = -EOPNOTSUPP; 100 struct tc_mqprio_qopt *qopt = NULL; 101 102 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE); 103 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK); 104 105 if (sch->parent != TC_H_ROOT) 106 return -EOPNOTSUPP; 107 108 if (!netif_is_multiqueue(dev)) 109 return -EOPNOTSUPP; 110 111 if (!opt || nla_len(opt) < sizeof(*qopt)) 112 return -EINVAL; 113 114 qopt = nla_data(opt); 115 if (mqprio_parse_opt(dev, qopt)) 116 return -EINVAL; 117 118 /* pre-allocate qdisc, attachment can't fail */ 119 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]), 120 GFP_KERNEL); 121 if (!priv->qdiscs) 122 return -ENOMEM; 123 124 for (i = 0; i < dev->num_tx_queues; i++) { 125 dev_queue = netdev_get_tx_queue(dev, i); 126 qdisc = qdisc_create_dflt(dev_queue, 127 get_default_qdisc_ops(dev, i), 128 TC_H_MAKE(TC_H_MAJ(sch->handle), 129 TC_H_MIN(i + 1))); 130 if (!qdisc) 131 return -ENOMEM; 132 133 priv->qdiscs[i] = qdisc; 134 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 135 } 136 137 /* If the mqprio options indicate that hardware should own 138 * the queue mapping then run ndo_setup_tc otherwise use the 139 * supplied and verified mapping 140 */ 141 if (qopt->hw) { 142 struct tc_to_netdev tc = {.type = TC_SETUP_MQPRIO, 143 { .tc = qopt->num_tc }}; 144 145 priv->hw_owned = 1; 146 err = dev->netdev_ops->ndo_setup_tc(dev, sch->handle, 0, &tc); 147 if (err) 148 return err; 149 } else { 150 netdev_set_num_tc(dev, qopt->num_tc); 151 for (i = 0; i < qopt->num_tc; i++) 152 netdev_set_tc_queue(dev, i, 153 qopt->count[i], qopt->offset[i]); 154 } 155 156 /* Always use supplied priority mappings */ 157 for (i = 0; i < TC_BITMASK + 1; i++) 158 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]); 159 160 sch->flags |= TCQ_F_MQROOT; 161 return 0; 162 } 163 164 static void mqprio_attach(struct Qdisc *sch) 165 { 166 struct net_device *dev = qdisc_dev(sch); 167 struct mqprio_sched *priv = qdisc_priv(sch); 168 struct Qdisc *qdisc, *old; 169 unsigned int ntx; 170 171 /* Attach underlying qdisc */ 172 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) { 173 qdisc = priv->qdiscs[ntx]; 174 old = dev_graft_qdisc(qdisc->dev_queue, qdisc); 175 if (old) 176 qdisc_destroy(old); 177 if (ntx < dev->real_num_tx_queues) 178 qdisc_hash_add(qdisc); 179 } 180 kfree(priv->qdiscs); 181 priv->qdiscs = NULL; 182 } 183 184 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch, 185 unsigned long cl) 186 { 187 struct net_device *dev = qdisc_dev(sch); 188 unsigned long ntx = cl - 1 - netdev_get_num_tc(dev); 189 190 if (ntx >= dev->num_tx_queues) 191 return NULL; 192 return netdev_get_tx_queue(dev, ntx); 193 } 194 195 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new, 196 struct Qdisc **old) 197 { 198 struct net_device *dev = qdisc_dev(sch); 199 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 200 201 if (!dev_queue) 202 return -EINVAL; 203 204 if (dev->flags & IFF_UP) 205 dev_deactivate(dev); 206 207 *old = dev_graft_qdisc(dev_queue, new); 208 209 if (new) 210 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 211 212 if (dev->flags & IFF_UP) 213 dev_activate(dev); 214 215 return 0; 216 } 217 218 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb) 219 { 220 struct net_device *dev = qdisc_dev(sch); 221 struct mqprio_sched *priv = qdisc_priv(sch); 222 unsigned char *b = skb_tail_pointer(skb); 223 struct tc_mqprio_qopt opt = { 0 }; 224 struct Qdisc *qdisc; 225 unsigned int i; 226 227 sch->q.qlen = 0; 228 memset(&sch->bstats, 0, sizeof(sch->bstats)); 229 memset(&sch->qstats, 0, sizeof(sch->qstats)); 230 231 for (i = 0; i < dev->num_tx_queues; i++) { 232 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc); 233 spin_lock_bh(qdisc_lock(qdisc)); 234 sch->q.qlen += qdisc->q.qlen; 235 sch->bstats.bytes += qdisc->bstats.bytes; 236 sch->bstats.packets += qdisc->bstats.packets; 237 sch->qstats.backlog += qdisc->qstats.backlog; 238 sch->qstats.drops += qdisc->qstats.drops; 239 sch->qstats.requeues += qdisc->qstats.requeues; 240 sch->qstats.overlimits += qdisc->qstats.overlimits; 241 spin_unlock_bh(qdisc_lock(qdisc)); 242 } 243 244 opt.num_tc = netdev_get_num_tc(dev); 245 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map)); 246 opt.hw = priv->hw_owned; 247 248 for (i = 0; i < netdev_get_num_tc(dev); i++) { 249 opt.count[i] = dev->tc_to_txq[i].count; 250 opt.offset[i] = dev->tc_to_txq[i].offset; 251 } 252 253 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 254 goto nla_put_failure; 255 256 return skb->len; 257 nla_put_failure: 258 nlmsg_trim(skb, b); 259 return -1; 260 } 261 262 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl) 263 { 264 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 265 266 if (!dev_queue) 267 return NULL; 268 269 return dev_queue->qdisc_sleeping; 270 } 271 272 static unsigned long mqprio_get(struct Qdisc *sch, u32 classid) 273 { 274 struct net_device *dev = qdisc_dev(sch); 275 unsigned int ntx = TC_H_MIN(classid); 276 277 if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev)) 278 return 0; 279 return ntx; 280 } 281 282 static void mqprio_put(struct Qdisc *sch, unsigned long cl) 283 { 284 } 285 286 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl, 287 struct sk_buff *skb, struct tcmsg *tcm) 288 { 289 struct net_device *dev = qdisc_dev(sch); 290 291 if (cl <= netdev_get_num_tc(dev)) { 292 tcm->tcm_parent = TC_H_ROOT; 293 tcm->tcm_info = 0; 294 } else { 295 int i; 296 struct netdev_queue *dev_queue; 297 298 dev_queue = mqprio_queue_get(sch, cl); 299 tcm->tcm_parent = 0; 300 for (i = 0; i < netdev_get_num_tc(dev); i++) { 301 struct netdev_tc_txq tc = dev->tc_to_txq[i]; 302 int q_idx = cl - netdev_get_num_tc(dev); 303 304 if (q_idx > tc.offset && 305 q_idx <= tc.offset + tc.count) { 306 tcm->tcm_parent = 307 TC_H_MAKE(TC_H_MAJ(sch->handle), 308 TC_H_MIN(i + 1)); 309 break; 310 } 311 } 312 tcm->tcm_info = dev_queue->qdisc_sleeping->handle; 313 } 314 tcm->tcm_handle |= TC_H_MIN(cl); 315 return 0; 316 } 317 318 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl, 319 struct gnet_dump *d) 320 __releases(d->lock) 321 __acquires(d->lock) 322 { 323 struct net_device *dev = qdisc_dev(sch); 324 325 if (cl <= netdev_get_num_tc(dev)) { 326 int i; 327 __u32 qlen = 0; 328 struct Qdisc *qdisc; 329 struct gnet_stats_queue qstats = {0}; 330 struct gnet_stats_basic_packed bstats = {0}; 331 struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1]; 332 333 /* Drop lock here it will be reclaimed before touching 334 * statistics this is required because the d->lock we 335 * hold here is the look on dev_queue->qdisc_sleeping 336 * also acquired below. 337 */ 338 if (d->lock) 339 spin_unlock_bh(d->lock); 340 341 for (i = tc.offset; i < tc.offset + tc.count; i++) { 342 struct netdev_queue *q = netdev_get_tx_queue(dev, i); 343 344 qdisc = rtnl_dereference(q->qdisc); 345 spin_lock_bh(qdisc_lock(qdisc)); 346 qlen += qdisc->q.qlen; 347 bstats.bytes += qdisc->bstats.bytes; 348 bstats.packets += qdisc->bstats.packets; 349 qstats.backlog += qdisc->qstats.backlog; 350 qstats.drops += qdisc->qstats.drops; 351 qstats.requeues += qdisc->qstats.requeues; 352 qstats.overlimits += qdisc->qstats.overlimits; 353 spin_unlock_bh(qdisc_lock(qdisc)); 354 } 355 /* Reclaim root sleeping lock before completing stats */ 356 if (d->lock) 357 spin_lock_bh(d->lock); 358 if (gnet_stats_copy_basic(NULL, d, NULL, &bstats) < 0 || 359 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0) 360 return -1; 361 } else { 362 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 363 364 sch = dev_queue->qdisc_sleeping; 365 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), 366 d, NULL, &sch->bstats) < 0 || 367 gnet_stats_copy_queue(d, NULL, 368 &sch->qstats, sch->q.qlen) < 0) 369 return -1; 370 } 371 return 0; 372 } 373 374 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg) 375 { 376 struct net_device *dev = qdisc_dev(sch); 377 unsigned long ntx; 378 379 if (arg->stop) 380 return; 381 382 /* Walk hierarchy with a virtual class per tc */ 383 arg->count = arg->skip; 384 for (ntx = arg->skip; 385 ntx < dev->num_tx_queues + netdev_get_num_tc(dev); 386 ntx++) { 387 if (arg->fn(sch, ntx + 1, arg) < 0) { 388 arg->stop = 1; 389 break; 390 } 391 arg->count++; 392 } 393 } 394 395 static const struct Qdisc_class_ops mqprio_class_ops = { 396 .graft = mqprio_graft, 397 .leaf = mqprio_leaf, 398 .get = mqprio_get, 399 .put = mqprio_put, 400 .walk = mqprio_walk, 401 .dump = mqprio_dump_class, 402 .dump_stats = mqprio_dump_class_stats, 403 }; 404 405 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = { 406 .cl_ops = &mqprio_class_ops, 407 .id = "mqprio", 408 .priv_size = sizeof(struct mqprio_sched), 409 .init = mqprio_init, 410 .destroy = mqprio_destroy, 411 .attach = mqprio_attach, 412 .dump = mqprio_dump, 413 .owner = THIS_MODULE, 414 }; 415 416 static int __init mqprio_module_init(void) 417 { 418 return register_qdisc(&mqprio_qdisc_ops); 419 } 420 421 static void __exit mqprio_module_exit(void) 422 { 423 unregister_qdisc(&mqprio_qdisc_ops); 424 } 425 426 module_init(mqprio_module_init); 427 module_exit(mqprio_module_exit); 428 429 MODULE_LICENSE("GPL"); 430