blob: 28de43092330abc125423d5328babc709b1f986f [file] [log] [blame]
John Fastabendb8970f02011-01-17 08:06:09 +00001/*
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>
Paul Gortmaker3a9a2312011-05-27 09:12:25 -040017#include <linux/module.h>
John Fastabendb8970f02011-01-17 08:06:09 +000018#include <net/netlink.h>
19#include <net/pkt_sched.h>
20#include <net/sch_generic.h>
21
22struct mqprio_sched {
23 struct Qdisc **qdiscs;
24 int hw_owned;
25};
26
27static void mqprio_destroy(struct Qdisc *sch)
28{
29 struct net_device *dev = qdisc_dev(sch);
30 struct mqprio_sched *priv = qdisc_priv(sch);
31 unsigned int ntx;
32
Ben Hutchingsac7100b2011-02-14 19:02:23 +000033 if (priv->qdiscs) {
34 for (ntx = 0;
35 ntx < dev->num_tx_queues && priv->qdiscs[ntx];
36 ntx++)
37 qdisc_destroy(priv->qdiscs[ntx]);
38 kfree(priv->qdiscs);
39 }
John Fastabendb8970f02011-01-17 08:06:09 +000040
41 if (priv->hw_owned && dev->netdev_ops->ndo_setup_tc)
42 dev->netdev_ops->ndo_setup_tc(dev, 0);
43 else
44 netdev_set_num_tc(dev, 0);
John Fastabendb8970f02011-01-17 08:06:09 +000045}
46
47static int mqprio_parse_opt(struct net_device *dev, struct tc_mqprio_qopt *qopt)
48{
49 int i, j;
50
51 /* Verify num_tc is not out of max range */
52 if (qopt->num_tc > TC_MAX_QUEUE)
53 return -EINVAL;
54
55 /* Verify priority mapping uses valid tcs */
56 for (i = 0; i < TC_BITMASK + 1; i++) {
57 if (qopt->prio_tc_map[i] >= qopt->num_tc)
58 return -EINVAL;
59 }
60
61 /* net_device does not support requested operation */
62 if (qopt->hw && !dev->netdev_ops->ndo_setup_tc)
63 return -EINVAL;
64
65 /* if hw owned qcount and qoffset are taken from LLD so
66 * no reason to verify them here
67 */
68 if (qopt->hw)
69 return 0;
70
71 for (i = 0; i < qopt->num_tc; i++) {
72 unsigned int last = qopt->offset[i] + qopt->count[i];
73
74 /* Verify the queue count is in tx range being equal to the
75 * real_num_tx_queues indicates the last queue is in use.
76 */
77 if (qopt->offset[i] >= dev->real_num_tx_queues ||
78 !qopt->count[i] ||
79 last > dev->real_num_tx_queues)
80 return -EINVAL;
81
82 /* Verify that the offset and counts do not overlap */
83 for (j = i + 1; j < qopt->num_tc; j++) {
84 if (last > qopt->offset[j])
85 return -EINVAL;
86 }
87 }
88
89 return 0;
90}
91
92static int mqprio_init(struct Qdisc *sch, struct nlattr *opt)
93{
94 struct net_device *dev = qdisc_dev(sch);
95 struct mqprio_sched *priv = qdisc_priv(sch);
96 struct netdev_queue *dev_queue;
97 struct Qdisc *qdisc;
98 int i, err = -EOPNOTSUPP;
99 struct tc_mqprio_qopt *qopt = NULL;
100
101 BUILD_BUG_ON(TC_MAX_QUEUE != TC_QOPT_MAX_QUEUE);
102 BUILD_BUG_ON(TC_BITMASK != TC_QOPT_BITMASK);
103
104 if (sch->parent != TC_H_ROOT)
105 return -EOPNOTSUPP;
106
107 if (!netif_is_multiqueue(dev))
108 return -EOPNOTSUPP;
109
Thomas Graf7838f2c2011-12-22 02:05:07 +0000110 if (!opt || nla_len(opt) < sizeof(*qopt))
John Fastabendb8970f02011-01-17 08:06:09 +0000111 return -EINVAL;
112
113 qopt = nla_data(opt);
114 if (mqprio_parse_opt(dev, qopt))
115 return -EINVAL;
116
117 /* pre-allocate qdisc, attachment can't fail */
118 priv->qdiscs = kcalloc(dev->num_tx_queues, sizeof(priv->qdiscs[0]),
119 GFP_KERNEL);
120 if (priv->qdiscs == NULL) {
121 err = -ENOMEM;
122 goto err;
123 }
124
125 for (i = 0; i < dev->num_tx_queues; i++) {
126 dev_queue = netdev_get_tx_queue(dev, i);
127 qdisc = qdisc_create_dflt(dev_queue, &pfifo_fast_ops,
128 TC_H_MAKE(TC_H_MAJ(sch->handle),
129 TC_H_MIN(i + 1)));
130 if (qdisc == NULL) {
131 err = -ENOMEM;
132 goto err;
133 }
John Fastabendb8970f02011-01-17 08:06:09 +0000134 priv->qdiscs[i] = qdisc;
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 priv->hw_owned = 1;
143 err = dev->netdev_ops->ndo_setup_tc(dev, qopt->num_tc);
144 if (err)
145 goto err;
146 } else {
147 netdev_set_num_tc(dev, qopt->num_tc);
148 for (i = 0; i < qopt->num_tc; i++)
149 netdev_set_tc_queue(dev, i,
150 qopt->count[i], qopt->offset[i]);
151 }
152
153 /* Always use supplied priority mappings */
154 for (i = 0; i < TC_BITMASK + 1; i++)
155 netdev_set_prio_tc_map(dev, i, qopt->prio_tc_map[i]);
156
157 sch->flags |= TCQ_F_MQROOT;
158 return 0;
159
160err:
161 mqprio_destroy(sch);
162 return err;
163}
164
165static void mqprio_attach(struct Qdisc *sch)
166{
167 struct net_device *dev = qdisc_dev(sch);
168 struct mqprio_sched *priv = qdisc_priv(sch);
169 struct Qdisc *qdisc;
170 unsigned int ntx;
171
172 /* Attach underlying qdisc */
173 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
174 qdisc = priv->qdiscs[ntx];
175 qdisc = dev_graft_qdisc(qdisc->dev_queue, qdisc);
176 if (qdisc)
177 qdisc_destroy(qdisc);
178 }
179 kfree(priv->qdiscs);
180 priv->qdiscs = NULL;
181}
182
183static struct netdev_queue *mqprio_queue_get(struct Qdisc *sch,
184 unsigned long cl)
185{
186 struct net_device *dev = qdisc_dev(sch);
187 unsigned long ntx = cl - 1 - netdev_get_num_tc(dev);
188
189 if (ntx >= dev->num_tx_queues)
190 return NULL;
191 return netdev_get_tx_queue(dev, ntx);
192}
193
194static int mqprio_graft(struct Qdisc *sch, unsigned long cl, struct Qdisc *new,
195 struct Qdisc **old)
196{
197 struct net_device *dev = qdisc_dev(sch);
198 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
199
200 if (!dev_queue)
201 return -EINVAL;
202
203 if (dev->flags & IFF_UP)
204 dev_deactivate(dev);
205
206 *old = dev_graft_qdisc(dev_queue, new);
207
208 if (dev->flags & IFF_UP)
209 dev_activate(dev);
210
211 return 0;
212}
213
214static int mqprio_dump(struct Qdisc *sch, struct sk_buff *skb)
215{
216 struct net_device *dev = qdisc_dev(sch);
217 struct mqprio_sched *priv = qdisc_priv(sch);
218 unsigned char *b = skb_tail_pointer(skb);
Eric Dumazet144ce872011-01-26 07:21:57 +0000219 struct tc_mqprio_qopt opt = { 0 };
John Fastabendb8970f02011-01-17 08:06:09 +0000220 struct Qdisc *qdisc;
221 unsigned int i;
222
223 sch->q.qlen = 0;
224 memset(&sch->bstats, 0, sizeof(sch->bstats));
225 memset(&sch->qstats, 0, sizeof(sch->qstats));
226
227 for (i = 0; i < dev->num_tx_queues; i++) {
228 qdisc = netdev_get_tx_queue(dev, i)->qdisc;
229 spin_lock_bh(qdisc_lock(qdisc));
230 sch->q.qlen += qdisc->q.qlen;
231 sch->bstats.bytes += qdisc->bstats.bytes;
232 sch->bstats.packets += qdisc->bstats.packets;
233 sch->qstats.qlen += qdisc->qstats.qlen;
234 sch->qstats.backlog += qdisc->qstats.backlog;
235 sch->qstats.drops += qdisc->qstats.drops;
236 sch->qstats.requeues += qdisc->qstats.requeues;
237 sch->qstats.overlimits += qdisc->qstats.overlimits;
238 spin_unlock_bh(qdisc_lock(qdisc));
239 }
240
241 opt.num_tc = netdev_get_num_tc(dev);
242 memcpy(opt.prio_tc_map, dev->prio_tc_map, sizeof(opt.prio_tc_map));
243 opt.hw = priv->hw_owned;
244
245 for (i = 0; i < netdev_get_num_tc(dev); i++) {
246 opt.count[i] = dev->tc_to_txq[i].count;
247 opt.offset[i] = dev->tc_to_txq[i].offset;
248 }
249
250 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
251
252 return skb->len;
253nla_put_failure:
254 nlmsg_trim(skb, b);
255 return -1;
256}
257
258static struct Qdisc *mqprio_leaf(struct Qdisc *sch, unsigned long cl)
259{
260 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
261
262 if (!dev_queue)
263 return NULL;
264
265 return dev_queue->qdisc_sleeping;
266}
267
268static unsigned long mqprio_get(struct Qdisc *sch, u32 classid)
269{
270 struct net_device *dev = qdisc_dev(sch);
271 unsigned int ntx = TC_H_MIN(classid);
272
273 if (ntx > dev->num_tx_queues + netdev_get_num_tc(dev))
274 return 0;
275 return ntx;
276}
277
278static void mqprio_put(struct Qdisc *sch, unsigned long cl)
279{
280}
281
282static int mqprio_dump_class(struct Qdisc *sch, unsigned long cl,
283 struct sk_buff *skb, struct tcmsg *tcm)
284{
285 struct net_device *dev = qdisc_dev(sch);
286
287 if (cl <= netdev_get_num_tc(dev)) {
288 tcm->tcm_parent = TC_H_ROOT;
289 tcm->tcm_info = 0;
290 } else {
291 int i;
292 struct netdev_queue *dev_queue;
293
294 dev_queue = mqprio_queue_get(sch, cl);
295 tcm->tcm_parent = 0;
296 for (i = 0; i < netdev_get_num_tc(dev); i++) {
297 struct netdev_tc_txq tc = dev->tc_to_txq[i];
298 int q_idx = cl - netdev_get_num_tc(dev);
299
300 if (q_idx > tc.offset &&
301 q_idx <= tc.offset + tc.count) {
302 tcm->tcm_parent =
303 TC_H_MAKE(TC_H_MAJ(sch->handle),
304 TC_H_MIN(i + 1));
305 break;
306 }
307 }
308 tcm->tcm_info = dev_queue->qdisc_sleeping->handle;
309 }
310 tcm->tcm_handle |= TC_H_MIN(cl);
311 return 0;
312}
313
314static int mqprio_dump_class_stats(struct Qdisc *sch, unsigned long cl,
stephen hemmingerea18fd92011-02-23 09:06:51 +0000315 struct gnet_dump *d)
316 __releases(d->lock)
317 __acquires(d->lock)
John Fastabendb8970f02011-01-17 08:06:09 +0000318{
319 struct net_device *dev = qdisc_dev(sch);
320
321 if (cl <= netdev_get_num_tc(dev)) {
322 int i;
323 struct Qdisc *qdisc;
324 struct gnet_stats_queue qstats = {0};
325 struct gnet_stats_basic_packed bstats = {0};
326 struct netdev_tc_txq tc = dev->tc_to_txq[cl - 1];
327
328 /* Drop lock here it will be reclaimed before touching
329 * statistics this is required because the d->lock we
330 * hold here is the look on dev_queue->qdisc_sleeping
331 * also acquired below.
332 */
333 spin_unlock_bh(d->lock);
334
335 for (i = tc.offset; i < tc.offset + tc.count; i++) {
336 qdisc = netdev_get_tx_queue(dev, i)->qdisc;
337 spin_lock_bh(qdisc_lock(qdisc));
338 bstats.bytes += qdisc->bstats.bytes;
339 bstats.packets += qdisc->bstats.packets;
340 qstats.qlen += qdisc->qstats.qlen;
341 qstats.backlog += qdisc->qstats.backlog;
342 qstats.drops += qdisc->qstats.drops;
343 qstats.requeues += qdisc->qstats.requeues;
344 qstats.overlimits += qdisc->qstats.overlimits;
345 spin_unlock_bh(qdisc_lock(qdisc));
346 }
347 /* Reclaim root sleeping lock before completing stats */
348 spin_lock_bh(d->lock);
349 if (gnet_stats_copy_basic(d, &bstats) < 0 ||
350 gnet_stats_copy_queue(d, &qstats) < 0)
351 return -1;
352 } else {
353 struct netdev_queue *dev_queue = mqprio_queue_get(sch, cl);
354
355 sch = dev_queue->qdisc_sleeping;
356 sch->qstats.qlen = sch->q.qlen;
357 if (gnet_stats_copy_basic(d, &sch->bstats) < 0 ||
358 gnet_stats_copy_queue(d, &sch->qstats) < 0)
359 return -1;
360 }
361 return 0;
362}
363
364static void mqprio_walk(struct Qdisc *sch, struct qdisc_walker *arg)
365{
366 struct net_device *dev = qdisc_dev(sch);
367 unsigned long ntx;
368
369 if (arg->stop)
370 return;
371
372 /* Walk hierarchy with a virtual class per tc */
373 arg->count = arg->skip;
374 for (ntx = arg->skip;
375 ntx < dev->num_tx_queues + netdev_get_num_tc(dev);
376 ntx++) {
377 if (arg->fn(sch, ntx + 1, arg) < 0) {
378 arg->stop = 1;
379 break;
380 }
381 arg->count++;
382 }
383}
384
385static const struct Qdisc_class_ops mqprio_class_ops = {
386 .graft = mqprio_graft,
387 .leaf = mqprio_leaf,
388 .get = mqprio_get,
389 .put = mqprio_put,
390 .walk = mqprio_walk,
391 .dump = mqprio_dump_class,
392 .dump_stats = mqprio_dump_class_stats,
393};
394
stephen hemmingerea18fd92011-02-23 09:06:51 +0000395static struct Qdisc_ops mqprio_qdisc_ops __read_mostly = {
John Fastabendb8970f02011-01-17 08:06:09 +0000396 .cl_ops = &mqprio_class_ops,
397 .id = "mqprio",
398 .priv_size = sizeof(struct mqprio_sched),
399 .init = mqprio_init,
400 .destroy = mqprio_destroy,
401 .attach = mqprio_attach,
402 .dump = mqprio_dump,
403 .owner = THIS_MODULE,
404};
405
406static int __init mqprio_module_init(void)
407{
408 return register_qdisc(&mqprio_qdisc_ops);
409}
410
411static void __exit mqprio_module_exit(void)
412{
413 unregister_qdisc(&mqprio_qdisc_ops);
414}
415
416module_init(mqprio_module_init);
417module_exit(mqprio_module_exit);
418
419MODULE_LICENSE("GPL");