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 == NULL) { 122 err = -ENOMEM; 123 goto err; 124 } 125 126 for (i = 0; i < dev->num_tx_queues; i++) { 127 dev_queue = netdev_get_tx_queue(dev, i); 128 qdisc = qdisc_create_dflt(dev_queue, 129 get_default_qdisc_ops(dev, i), 130 TC_H_MAKE(TC_H_MAJ(sch->handle), 131 TC_H_MIN(i + 1))); 132 if (qdisc == NULL) { 133 err = -ENOMEM; 134 goto err; 135 } 136 priv->qdiscs[i] = qdisc; 137 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 138 } 139 140 /* If the mqprio options indicate that hardware should own 141 * the queue mapping then run ndo_setup_tc otherwise use the 142 * supplied and verified mapping 143 */ 144 if (qopt->hw) { 145 struct tc_to_netdev tc = {.type = TC_SETUP_MQPRIO, 146 { .tc = qopt->num_tc }}; 147 148 priv->hw_owned = 1; 149 err = dev->netdev_ops->ndo_setup_tc(dev, sch->handle, 0, &tc); 150 if (err) 151 goto err; 152 } else { 153 netdev_set_num_tc(dev, qopt->num_tc); 154 for (i = 0; i < qopt->num_tc; i++) 155 netdev_set_tc_queue(dev, i, 156 qopt->count[i], qopt->offset[i]); 157 } 158 159 /* Always use supplied priority mappings */ 160 for (i = 0; i < TC_BITMASK + 1; i++) 161 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]); 162 163 sch->flags |= TCQ_F_MQROOT; 164 return 0; 165 166 err: 167 mqprio_destroy(sch); 168 return err; 169 } 170 171 static void mqprio_attach(struct Qdisc *sch) 172 { 173 struct net_device *dev = qdisc_dev(sch); 174 struct mqprio_sched *priv = qdisc_priv(sch); 175 struct Qdisc *qdisc, *old; 176 unsigned int ntx; 177 178 /* Attach underlying qdisc */ 179 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) { 180 qdisc = priv->qdiscs[ntx]; 181 old = dev_graft_qdisc(qdisc->dev_queue, qdisc); 182 if (old) 183 qdisc_destroy(old); 184 if (ntx < dev->real_num_tx_queues) 185 qdisc_hash_add(qdisc); 186 } 187 kfree(priv->qdiscs); 188 priv->qdiscs = NULL; 189 } 190 191 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch, 192 unsigned long cl) 193 { 194 struct net_device *dev = qdisc_dev(sch); 195 unsigned long ntx = cl - 1 - netdev_get_num_tc(dev); 196 197 if (ntx >= dev->num_tx_queues) 198 return NULL; 199 return netdev_get_tx_queue(dev, ntx); 200 } 201 202 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new, 203 struct Qdisc **old) 204 { 205 struct net_device *dev = qdisc_dev(sch); 206 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 207 208 if (!dev_queue) 209 return -EINVAL; 210 211 if (dev->flags & IFF_UP) 212 dev_deactivate(dev); 213 214 *old = dev_graft_qdisc(dev_queue, new); 215 216 if (new) 217 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; 218 219 if (dev->flags & IFF_UP) 220 dev_activate(dev); 221 222 return 0; 223 } 224 225 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb) 226 { 227 struct net_device *dev = qdisc_dev(sch); 228 struct mqprio_sched *priv = qdisc_priv(sch); 229 unsigned char *b = skb_tail_pointer(skb); 230 struct tc_mqprio_qopt opt = { 0 }; 231 struct Qdisc *qdisc; 232 unsigned int i; 233 234 sch->q.qlen = 0; 235 memset(&sch->bstats, 0, sizeof(sch->bstats)); 236 memset(&sch->qstats, 0, sizeof(sch->qstats)); 237 238 for (i = 0; i < dev->num_tx_queues; i++) { 239 qdisc = rtnl_dereference(netdev_get_tx_queue(dev, i)->qdisc); 240 spin_lock_bh(qdisc_lock(qdisc)); 241 sch->q.qlen += qdisc->q.qlen; 242 sch->bstats.bytes += qdisc->bstats.bytes; 243 sch->bstats.packets += qdisc->bstats.packets; 244 sch->qstats.backlog += qdisc->qstats.backlog; 245 sch->qstats.drops += qdisc->qstats.drops; 246 sch->qstats.requeues += qdisc->qstats.requeues; 247 sch->qstats.overlimits += qdisc->qstats.overlimits; 248 spin_unlock_bh(qdisc_lock(qdisc)); 249 } 250 251 opt.num_tc = netdev_get_num_tc(dev); 252 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map)); 253 opt.hw = priv->hw_owned; 254 255 for (i = 0; i < netdev_get_num_tc(dev); i++) { 256 opt.count[i] = dev->tc_to_txq[i].count; 257 opt.offset[i] = dev->tc_to_txq[i].offset; 258 } 259 260 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) 261 goto nla_put_failure; 262 263 return skb->len; 264 nla_put_failure: 265 nlmsg_trim(skb, b); 266 return -1; 267 } 268 269 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl) 270 { 271 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 272 273 if (!dev_queue) 274 return NULL; 275 276 return dev_queue->qdisc_sleeping; 277 } 278 279 static unsigned long mqprio_get(struct Qdisc *sch, u32 classid) 280 { 281 struct net_device *dev = qdisc_dev(sch); 282 unsigned int ntx = TC_H_MIN(classid); 283 284 if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev)) 285 return 0; 286 return ntx; 287 } 288 289 static void mqprio_put(struct Qdisc *sch, unsigned long cl) 290 { 291 } 292 293 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl, 294 struct sk_buff *skb, struct tcmsg *tcm) 295 { 296 struct net_device *dev = qdisc_dev(sch); 297 298 if (cl <= netdev_get_num_tc(dev)) { 299 tcm->tcm_parent = TC_H_ROOT; 300 tcm->tcm_info = 0; 301 } else { 302 int i; 303 struct netdev_queue *dev_queue; 304 305 dev_queue = mqprio_queue_get(sch, cl); 306 tcm->tcm_parent = 0; 307 for (i = 0; i < netdev_get_num_tc(dev); i++) { 308 struct netdev_tc_txq tc = dev->tc_to_txq[i]; 309 int q_idx = cl - netdev_get_num_tc(dev); 310 311 if (q_idx > tc.offset && 312 q_idx <= tc.offset + tc.count) { 313 tcm->tcm_parent = 314 TC_H_MAKE(TC_H_MAJ(sch->handle), 315 TC_H_MIN(i + 1)); 316 break; 317 } 318 } 319 tcm->tcm_info = dev_queue->qdisc_sleeping->handle; 320 } 321 tcm->tcm_handle |= TC_H_MIN(cl); 322 return 0; 323 } 324 325 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl, 326 struct gnet_dump *d) 327 __releases(d->lock) 328 __acquires(d->lock) 329 { 330 struct net_device *dev = qdisc_dev(sch); 331 332 if (cl <= netdev_get_num_tc(dev)) { 333 int i; 334 __u32 qlen = 0; 335 struct Qdisc *qdisc; 336 struct gnet_stats_queue qstats = {0}; 337 struct gnet_stats_basic_packed bstats = {0}; 338 struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1]; 339 340 /* Drop lock here it will be reclaimed before touching 341 * statistics this is required because the d->lock we 342 * hold here is the look on dev_queue->qdisc_sleeping 343 * also acquired below. 344 */ 345 if (d->lock) 346 spin_unlock_bh(d->lock); 347 348 for (i = tc.offset; i < tc.offset + tc.count; i++) { 349 struct netdev_queue *q = netdev_get_tx_queue(dev, i); 350 351 qdisc = rtnl_dereference(q->qdisc); 352 spin_lock_bh(qdisc_lock(qdisc)); 353 qlen += qdisc->q.qlen; 354 bstats.bytes += qdisc->bstats.bytes; 355 bstats.packets += qdisc->bstats.packets; 356 qstats.backlog += qdisc->qstats.backlog; 357 qstats.drops += qdisc->qstats.drops; 358 qstats.requeues += qdisc->qstats.requeues; 359 qstats.overlimits += qdisc->qstats.overlimits; 360 spin_unlock_bh(qdisc_lock(qdisc)); 361 } 362 /* Reclaim root sleeping lock before completing stats */ 363 if (d->lock) 364 spin_lock_bh(d->lock); 365 if (gnet_stats_copy_basic(NULL, d, NULL, &bstats) < 0 || 366 gnet_stats_copy_queue(d, NULL, &qstats, qlen) < 0) 367 return -1; 368 } else { 369 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl); 370 371 sch = dev_queue->qdisc_sleeping; 372 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), 373 d, NULL, &sch->bstats) < 0 || 374 gnet_stats_copy_queue(d, NULL, 375 &sch->qstats, sch->q.qlen) < 0) 376 return -1; 377 } 378 return 0; 379 } 380 381 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg) 382 { 383 struct net_device *dev = qdisc_dev(sch); 384 unsigned long ntx; 385 386 if (arg->stop) 387 return; 388 389 /* Walk hierarchy with a virtual class per tc */ 390 arg->count = arg->skip; 391 for (ntx = arg->skip; 392 ntx < dev->num_tx_queues + netdev_get_num_tc(dev); 393 ntx++) { 394 if (arg->fn(sch, ntx + 1, arg) < 0) { 395 arg->stop = 1; 396 break; 397 } 398 arg->count++; 399 } 400 } 401 402 static const struct Qdisc_class_ops mqprio_class_ops = { 403 .graft = mqprio_graft, 404 .leaf = mqprio_leaf, 405 .get = mqprio_get, 406 .put = mqprio_put, 407 .walk = mqprio_walk, 408 .dump = mqprio_dump_class, 409 .dump_stats = mqprio_dump_class_stats, 410 }; 411 412 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = { 413 .cl_ops = &mqprio_class_ops, 414 .id = "mqprio", 415 .priv_size = sizeof(struct mqprio_sched), 416 .init = mqprio_init, 417 .destroy = mqprio_destroy, 418 .attach = mqprio_attach, 419 .dump = mqprio_dump, 420 .owner = THIS_MODULE, 421 }; 422 423 static int __init mqprio_module_init(void) 424 { 425 return register_qdisc(&mqprio_qdisc_ops); 426 } 427 428 static void __exit mqprio_module_exit(void) 429 { 430 unregister_qdisc(&mqprio_qdisc_ops); 431 } 432 433 module_init(mqprio_module_init); 434 module_exit(mqprio_module_exit); 435 436 MODULE_LICENSE("GPL"); 437