2 * net/sched/sch_mqprio.c
4 * Copyright (c) 2010 John Fastabend <john.r.fastabend@intel.com>
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.
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>
23 struct Qdisc **qdiscs;
27 static void mqprio_destroy(struct Qdisc *sch)
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};
36 ntx < dev->num_tx_queues && priv->qdiscs[ntx];
38 qdisc_destroy(priv->qdiscs[ntx]);
42 if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc)
43 dev->netdev_ops->ndo_setup_tc(dev, sch->handle, 0, &tc);
45 netdev_set_num_tc(dev, 0);
48 static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
52 /* Verify num_tc is not out of max range */
53 if (qopt->num_tc > TC_MAX_QUEUE)
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)
62 /* net_device does not support requested operation */
63 if (qopt->hw && !dev->netdev_ops->ndo_setup_tc)
66 /* if hw owned qcount and qoffset are taken from LLD so
67 * no reason to verify them here
72 for (i = 0; i < qopt->num_tc; i++) {
73 unsigned int last = qopt->offset[i] + qopt->count[i];
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.
78 if (qopt->offset[i] >= dev->real_num_tx_queues ||
80 last > dev->real_num_tx_queues)
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])
93 static int mqprio_init(struct Qdisc *sch, struct nlattr *opt)
95 struct net_device *dev = qdisc_dev(sch);
96 struct mqprio_sched *priv = qdisc_priv(sch);
97 struct netdev_queue *dev_queue;
99 int i, err = -EOPNOTSUPP;
100 struct tc_mqprio_qopt *qopt = NULL;
102 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
103 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
105 if (sch->parent != TC_H_ROOT)
108 if (!netif_is_multiqueue(dev))
111 if (!opt || nla_len(opt) < sizeof(*qopt))
114 qopt = nla_data(opt);
115 if (mqprio_parse_opt(dev, qopt))
118 /* pre-allocate qdisc, attachment can't fail */
119 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
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),
133 priv->qdiscs[i] = qdisc;
134 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
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
142 struct tc_to_netdev tc = {.type = TC_SETUP_MQPRIO,
143 { .tc = qopt->num_tc }};
146 err = dev->netdev_ops->ndo_setup_tc(dev, sch->handle, 0, &tc);
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]);
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]);
160 sch->flags |= TCQ_F_MQROOT;
164 static void mqprio_attach(struct Qdisc *sch)
166 struct net_device *dev = qdisc_dev(sch);
167 struct mqprio_sched *priv = qdisc_priv(sch);
168 struct Qdisc *qdisc, *old;
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);
177 if (ntx < dev->real_num_tx_queues)
178 qdisc_hash_add(qdisc);
184 static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
187 struct net_device *dev = qdisc_dev(sch);
188 unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
190 if (ntx >= dev->num_tx_queues)
192 return netdev_get_tx_queue(dev, ntx);
195 static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
198 struct net_device *dev = qdisc_dev(sch);
199 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
204 if (dev->flags & IFF_UP)
207 *old = dev_graft_qdisc(dev_queue, new);
210 new->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
212 if (dev->flags & IFF_UP)
218 static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
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 };
228 memset(&sch->bstats, 0, sizeof(sch->bstats));
229 memset(&sch->qstats, 0, sizeof(sch->qstats));
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));
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;
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;
253 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
254 goto nla_put_failure;
262 static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
264 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
269 return dev_queue->qdisc_sleeping;
272 static unsigned long mqprio_get(struct Qdisc *sch, u32 classid)
274 struct net_device *dev = qdisc_dev(sch);
275 unsigned int ntx = TC_H_MIN(classid);
277 if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev))
282 static void mqprio_put(struct Qdisc *sch, unsigned long cl)
286 static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
287 struct sk_buff *skb, struct tcmsg *tcm)
289 struct net_device *dev = qdisc_dev(sch);
291 if (cl <= netdev_get_num_tc(dev)) {
292 tcm->tcm_parent = TC_H_ROOT;
296 struct netdev_queue *dev_queue;
298 dev_queue = mqprio_queue_get(sch, cl);
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);
304 if (q_idx > tc.offset &&
305 q_idx <= tc.offset + tc.count) {
307 TC_H_MAKE(TC_H_MAJ(sch->handle),
312 tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
314 tcm->tcm_handle |= TC_H_MIN(cl);
318 static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
323 struct net_device *dev = qdisc_dev(sch);
325 if (cl <= netdev_get_num_tc(dev)) {
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];
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.
339 spin_unlock_bh(d->lock);
341 for (i = tc.offset; i < tc.offset + tc.count; i++) {
342 struct netdev_queue *q = netdev_get_tx_queue(dev, i);
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));
355 /* Reclaim root sleeping lock before completing stats */
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)
362 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
364 sch = dev_queue->qdisc_sleeping;
365 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), d,
366 sch->cpu_bstats, &sch->bstats) < 0 ||
367 gnet_stats_copy_queue(d, NULL,
368 &sch->qstats, sch->q.qlen) < 0)
374 static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
376 struct net_device *dev = qdisc_dev(sch);
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);
387 if (arg->fn(sch, ntx + 1, arg) < 0) {
395 static const struct Qdisc_class_ops mqprio_class_ops = {
396 .graft = mqprio_graft,
401 .dump = mqprio_dump_class,
402 .dump_stats = mqprio_dump_class_stats,
405 static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
406 .cl_ops = &mqprio_class_ops,
408 .priv_size = sizeof(struct mqprio_sched),
410 .destroy = mqprio_destroy,
411 .attach = mqprio_attach,
413 .owner = THIS_MODULE,
416 static int __init mqprio_module_init(void)
418 return register_qdisc(&mqprio_qdisc_ops);
421 static void __exit mqprio_module_exit(void)
423 unregister_qdisc(&mqprio_qdisc_ops);
426 module_init(mqprio_module_init);
427 module_exit(mqprio_module_exit);
429 MODULE_LICENSE("GPL");