Merge branch 'x86-cleanups-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-block.git] / net / sched / sch_multiq.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2008, Intel Corporation.
4  *
5  * Author: Alexander Duyck <alexander.h.duyck@intel.com>
6  */
7
8 #include <linux/module.h>
9 #include <linux/slab.h>
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/string.h>
13 #include <linux/errno.h>
14 #include <linux/skbuff.h>
15 #include <net/netlink.h>
16 #include <net/pkt_sched.h>
17 #include <net/pkt_cls.h>
18
19 struct multiq_sched_data {
20         u16 bands;
21         u16 max_bands;
22         u16 curband;
23         struct tcf_proto __rcu *filter_list;
24         struct tcf_block *block;
25         struct Qdisc **queues;
26 };
27
28
29 static struct Qdisc *
30 multiq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
31 {
32         struct multiq_sched_data *q = qdisc_priv(sch);
33         u32 band;
34         struct tcf_result res;
35         struct tcf_proto *fl = rcu_dereference_bh(q->filter_list);
36         int err;
37
38         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
39         err = tcf_classify(skb, fl, &res, false);
40 #ifdef CONFIG_NET_CLS_ACT
41         switch (err) {
42         case TC_ACT_STOLEN:
43         case TC_ACT_QUEUED:
44         case TC_ACT_TRAP:
45                 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
46                 /* fall through */
47         case TC_ACT_SHOT:
48                 return NULL;
49         }
50 #endif
51         band = skb_get_queue_mapping(skb);
52
53         if (band >= q->bands)
54                 return q->queues[0];
55
56         return q->queues[band];
57 }
58
59 static int
60 multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
61                struct sk_buff **to_free)
62 {
63         struct Qdisc *qdisc;
64         int ret;
65
66         qdisc = multiq_classify(skb, sch, &ret);
67 #ifdef CONFIG_NET_CLS_ACT
68         if (qdisc == NULL) {
69
70                 if (ret & __NET_XMIT_BYPASS)
71                         qdisc_qstats_drop(sch);
72                 __qdisc_drop(skb, to_free);
73                 return ret;
74         }
75 #endif
76
77         ret = qdisc_enqueue(skb, qdisc, to_free);
78         if (ret == NET_XMIT_SUCCESS) {
79                 sch->q.qlen++;
80                 return NET_XMIT_SUCCESS;
81         }
82         if (net_xmit_drop_count(ret))
83                 qdisc_qstats_drop(sch);
84         return ret;
85 }
86
87 static struct sk_buff *multiq_dequeue(struct Qdisc *sch)
88 {
89         struct multiq_sched_data *q = qdisc_priv(sch);
90         struct Qdisc *qdisc;
91         struct sk_buff *skb;
92         int band;
93
94         for (band = 0; band < q->bands; band++) {
95                 /* cycle through bands to ensure fairness */
96                 q->curband++;
97                 if (q->curband >= q->bands)
98                         q->curband = 0;
99
100                 /* Check that target subqueue is available before
101                  * pulling an skb to avoid head-of-line blocking.
102                  */
103                 if (!netif_xmit_stopped(
104                     netdev_get_tx_queue(qdisc_dev(sch), q->curband))) {
105                         qdisc = q->queues[q->curband];
106                         skb = qdisc->dequeue(qdisc);
107                         if (skb) {
108                                 qdisc_bstats_update(sch, skb);
109                                 sch->q.qlen--;
110                                 return skb;
111                         }
112                 }
113         }
114         return NULL;
115
116 }
117
118 static struct sk_buff *multiq_peek(struct Qdisc *sch)
119 {
120         struct multiq_sched_data *q = qdisc_priv(sch);
121         unsigned int curband = q->curband;
122         struct Qdisc *qdisc;
123         struct sk_buff *skb;
124         int band;
125
126         for (band = 0; band < q->bands; band++) {
127                 /* cycle through bands to ensure fairness */
128                 curband++;
129                 if (curband >= q->bands)
130                         curband = 0;
131
132                 /* Check that target subqueue is available before
133                  * pulling an skb to avoid head-of-line blocking.
134                  */
135                 if (!netif_xmit_stopped(
136                     netdev_get_tx_queue(qdisc_dev(sch), curband))) {
137                         qdisc = q->queues[curband];
138                         skb = qdisc->ops->peek(qdisc);
139                         if (skb)
140                                 return skb;
141                 }
142         }
143         return NULL;
144
145 }
146
147 static void
148 multiq_reset(struct Qdisc *sch)
149 {
150         u16 band;
151         struct multiq_sched_data *q = qdisc_priv(sch);
152
153         for (band = 0; band < q->bands; band++)
154                 qdisc_reset(q->queues[band]);
155         sch->q.qlen = 0;
156         q->curband = 0;
157 }
158
159 static void
160 multiq_destroy(struct Qdisc *sch)
161 {
162         int band;
163         struct multiq_sched_data *q = qdisc_priv(sch);
164
165         tcf_block_put(q->block);
166         for (band = 0; band < q->bands; band++)
167                 qdisc_put(q->queues[band]);
168
169         kfree(q->queues);
170 }
171
172 static int multiq_tune(struct Qdisc *sch, struct nlattr *opt,
173                        struct netlink_ext_ack *extack)
174 {
175         struct multiq_sched_data *q = qdisc_priv(sch);
176         struct tc_multiq_qopt *qopt;
177         int i;
178
179         if (!netif_is_multiqueue(qdisc_dev(sch)))
180                 return -EOPNOTSUPP;
181         if (nla_len(opt) < sizeof(*qopt))
182                 return -EINVAL;
183
184         qopt = nla_data(opt);
185
186         qopt->bands = qdisc_dev(sch)->real_num_tx_queues;
187
188         sch_tree_lock(sch);
189         q->bands = qopt->bands;
190         for (i = q->bands; i < q->max_bands; i++) {
191                 if (q->queues[i] != &noop_qdisc) {
192                         struct Qdisc *child = q->queues[i];
193
194                         q->queues[i] = &noop_qdisc;
195                         qdisc_tree_flush_backlog(child);
196                         qdisc_put(child);
197                 }
198         }
199
200         sch_tree_unlock(sch);
201
202         for (i = 0; i < q->bands; i++) {
203                 if (q->queues[i] == &noop_qdisc) {
204                         struct Qdisc *child, *old;
205                         child = qdisc_create_dflt(sch->dev_queue,
206                                                   &pfifo_qdisc_ops,
207                                                   TC_H_MAKE(sch->handle,
208                                                             i + 1), extack);
209                         if (child) {
210                                 sch_tree_lock(sch);
211                                 old = q->queues[i];
212                                 q->queues[i] = child;
213                                 if (child != &noop_qdisc)
214                                         qdisc_hash_add(child, true);
215
216                                 if (old != &noop_qdisc) {
217                                         qdisc_tree_flush_backlog(old);
218                                         qdisc_put(old);
219                                 }
220                                 sch_tree_unlock(sch);
221                         }
222                 }
223         }
224         return 0;
225 }
226
227 static int multiq_init(struct Qdisc *sch, struct nlattr *opt,
228                        struct netlink_ext_ack *extack)
229 {
230         struct multiq_sched_data *q = qdisc_priv(sch);
231         int i, err;
232
233         q->queues = NULL;
234
235         if (!opt)
236                 return -EINVAL;
237
238         err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
239         if (err)
240                 return err;
241
242         q->max_bands = qdisc_dev(sch)->num_tx_queues;
243
244         q->queues = kcalloc(q->max_bands, sizeof(struct Qdisc *), GFP_KERNEL);
245         if (!q->queues)
246                 return -ENOBUFS;
247         for (i = 0; i < q->max_bands; i++)
248                 q->queues[i] = &noop_qdisc;
249
250         return multiq_tune(sch, opt, extack);
251 }
252
253 static int multiq_dump(struct Qdisc *sch, struct sk_buff *skb)
254 {
255         struct multiq_sched_data *q = qdisc_priv(sch);
256         unsigned char *b = skb_tail_pointer(skb);
257         struct tc_multiq_qopt opt;
258
259         opt.bands = q->bands;
260         opt.max_bands = q->max_bands;
261
262         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
263                 goto nla_put_failure;
264
265         return skb->len;
266
267 nla_put_failure:
268         nlmsg_trim(skb, b);
269         return -1;
270 }
271
272 static int multiq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
273                         struct Qdisc **old, struct netlink_ext_ack *extack)
274 {
275         struct multiq_sched_data *q = qdisc_priv(sch);
276         unsigned long band = arg - 1;
277
278         if (new == NULL)
279                 new = &noop_qdisc;
280
281         *old = qdisc_replace(sch, new, &q->queues[band]);
282         return 0;
283 }
284
285 static struct Qdisc *
286 multiq_leaf(struct Qdisc *sch, unsigned long arg)
287 {
288         struct multiq_sched_data *q = qdisc_priv(sch);
289         unsigned long band = arg - 1;
290
291         return q->queues[band];
292 }
293
294 static unsigned long multiq_find(struct Qdisc *sch, u32 classid)
295 {
296         struct multiq_sched_data *q = qdisc_priv(sch);
297         unsigned long band = TC_H_MIN(classid);
298
299         if (band - 1 >= q->bands)
300                 return 0;
301         return band;
302 }
303
304 static unsigned long multiq_bind(struct Qdisc *sch, unsigned long parent,
305                                  u32 classid)
306 {
307         return multiq_find(sch, classid);
308 }
309
310
311 static void multiq_unbind(struct Qdisc *q, unsigned long cl)
312 {
313 }
314
315 static int multiq_dump_class(struct Qdisc *sch, unsigned long cl,
316                              struct sk_buff *skb, struct tcmsg *tcm)
317 {
318         struct multiq_sched_data *q = qdisc_priv(sch);
319
320         tcm->tcm_handle |= TC_H_MIN(cl);
321         tcm->tcm_info = q->queues[cl - 1]->handle;
322         return 0;
323 }
324
325 static int multiq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
326                                  struct gnet_dump *d)
327 {
328         struct multiq_sched_data *q = qdisc_priv(sch);
329         struct Qdisc *cl_q;
330
331         cl_q = q->queues[cl - 1];
332         if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
333                                   d, NULL, &cl_q->bstats) < 0 ||
334             qdisc_qstats_copy(d, cl_q) < 0)
335                 return -1;
336
337         return 0;
338 }
339
340 static void multiq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
341 {
342         struct multiq_sched_data *q = qdisc_priv(sch);
343         int band;
344
345         if (arg->stop)
346                 return;
347
348         for (band = 0; band < q->bands; band++) {
349                 if (arg->count < arg->skip) {
350                         arg->count++;
351                         continue;
352                 }
353                 if (arg->fn(sch, band + 1, arg) < 0) {
354                         arg->stop = 1;
355                         break;
356                 }
357                 arg->count++;
358         }
359 }
360
361 static struct tcf_block *multiq_tcf_block(struct Qdisc *sch, unsigned long cl,
362                                           struct netlink_ext_ack *extack)
363 {
364         struct multiq_sched_data *q = qdisc_priv(sch);
365
366         if (cl)
367                 return NULL;
368         return q->block;
369 }
370
371 static const struct Qdisc_class_ops multiq_class_ops = {
372         .graft          =       multiq_graft,
373         .leaf           =       multiq_leaf,
374         .find           =       multiq_find,
375         .walk           =       multiq_walk,
376         .tcf_block      =       multiq_tcf_block,
377         .bind_tcf       =       multiq_bind,
378         .unbind_tcf     =       multiq_unbind,
379         .dump           =       multiq_dump_class,
380         .dump_stats     =       multiq_dump_class_stats,
381 };
382
383 static struct Qdisc_ops multiq_qdisc_ops __read_mostly = {
384         .next           =       NULL,
385         .cl_ops         =       &multiq_class_ops,
386         .id             =       "multiq",
387         .priv_size      =       sizeof(struct multiq_sched_data),
388         .enqueue        =       multiq_enqueue,
389         .dequeue        =       multiq_dequeue,
390         .peek           =       multiq_peek,
391         .init           =       multiq_init,
392         .reset          =       multiq_reset,
393         .destroy        =       multiq_destroy,
394         .change         =       multiq_tune,
395         .dump           =       multiq_dump,
396         .owner          =       THIS_MODULE,
397 };
398
399 static int __init multiq_module_init(void)
400 {
401         return register_qdisc(&multiq_qdisc_ops);
402 }
403
404 static void __exit multiq_module_exit(void)
405 {
406         unregister_qdisc(&multiq_qdisc_ops);
407 }
408
409 module_init(multiq_module_init)
410 module_exit(multiq_module_exit)
411
412 MODULE_LICENSE("GPL");