net/sched/sch_multiq.c

   1 /*
   2  * Copyright (c) 2008, Intel Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, see <http://www.gnu.org/licenses/>.
  15  *
  16  * Author: Alexander Duyck <alexander.h.duyck@intel.com>
  17  */
  18
  19 #include <linux/module.h>
  20 #include <linux/slab.h>
  21 #include <linux/types.h>
  22 #include <linux/kernel.h>
  23 #include <linux/string.h>
  24 #include <linux/errno.h>
  25 #include <linux/skbuff.h>
  26 #include <net/netlink.h>
  27 #include <net/pkt_sched.h>
  28
  29
  30 struct multiq_sched_data {
  31         u16 bands;
  32         u16 max_bands;
  33         u16 curband;
  34         struct tcf_proto __rcu *filter_list;
  35         struct Qdisc **queues;
  36 };
  37
  38
  39 static struct Qdisc *
  40 multiq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
  41 {
  42         struct multiq_sched_data *q = qdisc_priv(sch);
  43         u32 band;
  44         struct tcf_result res;
  45         struct tcf_proto *fl = rcu_dereference_bh(q->filter_list);
  46         int err;
  47
  48         *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
  49         err = tc_classify(skb, fl, &res, false);
  50 #ifdef CONFIG_NET_CLS_ACT
  51         switch (err) {
  52         case TC_ACT_STOLEN:
  53         case TC_ACT_QUEUED:
  54                 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
  55         case TC_ACT_SHOT:
  56                 return NULL;
  57         }
  58 #endif
  59         band = skb_get_queue_mapping(skb);
  60
  61         if (band >= q->bands)
  62                 return q->queues[0];
  63
  64         return q->queues[band];
  65 }
  66
  67 static int
  68 multiq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
  69                struct sk_buff **to_free)
  70 {
  71         struct Qdisc *qdisc;
  72         int ret;
  73
  74         qdisc = multiq_classify(skb, sch, &ret);
  75 #ifdef CONFIG_NET_CLS_ACT
  76         if (qdisc == NULL) {
  77
  78                 if (ret & __NET_XMIT_BYPASS)
  79                         qdisc_qstats_drop(sch);
  80                 __qdisc_drop(skb, to_free);
  81                 return ret;
  82         }
  83 #endif
  84
  85         ret = qdisc_enqueue(skb, qdisc, to_free);
  86         if (ret == NET_XMIT_SUCCESS) {
  87                 sch->q.qlen++;
  88                 return NET_XMIT_SUCCESS;
  89         }
  90         if (net_xmit_drop_count(ret))
  91                 qdisc_qstats_drop(sch);
  92         return ret;
  93 }
  94
  95 static struct sk_buff *multiq_dequeue(struct Qdisc *sch)
  96 {
  97         struct multiq_sched_data *q = qdisc_priv(sch);
  98         struct Qdisc *qdisc;
  99         struct sk_buff *skb;
 100         int band;
 101
 102         for (band = 0; band < q->bands; band++) {
 103                 /* cycle through bands to ensure fairness */
 104                 q->curband++;
 105                 if (q->curband >= q->bands)
 106                         q->curband = 0;
 107
 108                 /* Check that target subqueue is available before
 109                  * pulling an skb to avoid head-of-line blocking.
 110                  */
 111                 if (!netif_xmit_stopped(
 112                     netdev_get_tx_queue(qdisc_dev(sch), q->curband))) {
 113                         qdisc = q->queues[q->curband];
 114                         skb = qdisc->dequeue(qdisc);
 115                         if (skb) {
 116                                 qdisc_bstats_update(sch, skb);
 117                                 sch->q.qlen--;
 118                                 return skb;
 119                         }
 120                 }
 121         }
 122         return NULL;
 123
 124 }
 125
 126 static struct sk_buff *multiq_peek(struct Qdisc *sch)
 127 {
 128         struct multiq_sched_data *q = qdisc_priv(sch);
 129         unsigned int curband = q->curband;
 130         struct Qdisc *qdisc;
 131         struct sk_buff *skb;
 132         int band;
 133
 134         for (band = 0; band < q->bands; band++) {
 135                 /* cycle through bands to ensure fairness */
 136                 curband++;
 137                 if (curband >= q->bands)
 138                         curband = 0;
 139
 140                 /* Check that target subqueue is available before
 141                  * pulling an skb to avoid head-of-line blocking.
 142                  */
 143                 if (!netif_xmit_stopped(
 144                     netdev_get_tx_queue(qdisc_dev(sch), curband))) {
 145                         qdisc = q->queues[curband];
 146                         skb = qdisc->ops->peek(qdisc);
 147                         if (skb)
 148                                 return skb;
 149                 }
 150         }
 151         return NULL;
 152
 153 }
 154
 155 static void
 156 multiq_reset(struct Qdisc *sch)
 157 {
 158         u16 band;
 159         struct multiq_sched_data *q = qdisc_priv(sch);
 160
 161         for (band = 0; band < q->bands; band++)
 162                 qdisc_reset(q->queues[band]);
 163         sch->q.qlen = 0;
 164         q->curband = 0;
 165 }
 166
 167 static void
 168 multiq_destroy(struct Qdisc *sch)
 169 {
 170         int band;
 171         struct multiq_sched_data *q = qdisc_priv(sch);
 172
 173         tcf_destroy_chain(&q->filter_list);
 174         for (band = 0; band < q->bands; band++)
 175                 qdisc_destroy(q->queues[band]);
 176
 177         kfree(q->queues);
 178 }
 179
 180 static int multiq_tune(struct Qdisc *sch, struct nlattr *opt)
 181 {
 182         struct multiq_sched_data *q = qdisc_priv(sch);
 183         struct tc_multiq_qopt *qopt;
 184         int i;
 185
 186         if (!netif_is_multiqueue(qdisc_dev(sch)))
 187                 return -EOPNOTSUPP;
 188         if (nla_len(opt) < sizeof(*qopt))
 189                 return -EINVAL;
 190
 191         qopt = nla_data(opt);
 192
 193         qopt->bands = qdisc_dev(sch)->real_num_tx_queues;
 194
 195         sch_tree_lock(sch);
 196         q->bands = qopt->bands;
 197         for (i = q->bands; i < q->max_bands; i++) {
 198                 if (q->queues[i] != &noop_qdisc) {
 199                         struct Qdisc *child = q->queues[i];
 200                         q->queues[i] = &noop_qdisc;
 201                         qdisc_tree_reduce_backlog(child, child->q.qlen,
 202                                                   child->qstats.backlog);
 203                         qdisc_destroy(child);
 204                 }
 205         }
 206
 207         sch_tree_unlock(sch);
 208
 209         for (i = 0; i < q->bands; i++) {
 210                 if (q->queues[i] == &noop_qdisc) {
 211                         struct Qdisc *child, *old;
 212                         child = qdisc_create_dflt(sch->dev_queue,
 213                                                   &pfifo_qdisc_ops,
 214                                                   TC_H_MAKE(sch->handle,
 215                                                             i + 1));
 216                         if (child) {
 217                                 sch_tree_lock(sch);
 218                                 old = q->queues[i];
 219                                 q->queues[i] = child;
 220
 221                                 if (old != &noop_qdisc) {
 222                                         qdisc_tree_reduce_backlog(old,
 223                                                                   old->q.qlen,
 224                                                                   old->qstats.backlog);
 225                                         qdisc_destroy(old);
 226                                 }
 227                                 sch_tree_unlock(sch);
 228                         }
 229                 }
 230         }
 231         return 0;
 232 }
 233
 234 static int multiq_init(struct Qdisc *sch, struct nlattr *opt)
 235 {
 236         struct multiq_sched_data *q = qdisc_priv(sch);
 237         int i, err;
 238
 239         q->queues = NULL;
 240
 241         if (opt == NULL)
 242                 return -EINVAL;
 243
 244         q->max_bands = qdisc_dev(sch)->num_tx_queues;
 245
 246         q->queues = kcalloc(q->max_bands, sizeof(struct Qdisc *), GFP_KERNEL);
 247         if (!q->queues)
 248                 return -ENOBUFS;
 249         for (i = 0; i < q->max_bands; i++)
 250                 q->queues[i] = &noop_qdisc;
 251
 252         err = multiq_tune(sch, opt);
 253
 254         if (err)
 255                 kfree(q->queues);
 256
 257         return err;
 258 }
 259
 260 static int multiq_dump(struct Qdisc *sch, struct sk_buff *skb)
 261 {
 262         struct multiq_sched_data *q = qdisc_priv(sch);
 263         unsigned char *b = skb_tail_pointer(skb);
 264         struct tc_multiq_qopt opt;
 265
 266         opt.bands = q->bands;
 267         opt.max_bands = q->max_bands;
 268
 269         if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
 270                 goto nla_put_failure;
 271
 272         return skb->len;
 273
 274 nla_put_failure:
 275         nlmsg_trim(skb, b);
 276         return -1;
 277 }
 278
 279 static int multiq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 280                       struct Qdisc **old)
 281 {
 282         struct multiq_sched_data *q = qdisc_priv(sch);
 283         unsigned long band = arg - 1;
 284
 285         if (new == NULL)
 286                 new = &noop_qdisc;
 287
 288         *old = qdisc_replace(sch, new, &q->queues[band]);
 289         return 0;
 290 }
 291
 292 static struct Qdisc *
 293 multiq_leaf(struct Qdisc *sch, unsigned long arg)
 294 {
 295         struct multiq_sched_data *q = qdisc_priv(sch);
 296         unsigned long band = arg - 1;
 297
 298         return q->queues[band];
 299 }
 300
 301 static unsigned long multiq_get(struct Qdisc *sch, u32 classid)
 302 {
 303         struct multiq_sched_data *q = qdisc_priv(sch);
 304         unsigned long band = TC_H_MIN(classid);
 305
 306         if (band - 1 >= q->bands)
 307                 return 0;
 308         return band;
 309 }
 310
 311 static unsigned long multiq_bind(struct Qdisc *sch, unsigned long parent,
 312                                  u32 classid)
 313 {
 314         return multiq_get(sch, classid);
 315 }
 316
 317
 318 static void multiq_put(struct Qdisc *q, unsigned long cl)
 319 {
 320 }
 321
 322 static int multiq_dump_class(struct Qdisc *sch, unsigned long cl,
 323                              struct sk_buff *skb, struct tcmsg *tcm)
 324 {
 325         struct multiq_sched_data *q = qdisc_priv(sch);
 326
 327         tcm->tcm_handle |= TC_H_MIN(cl);
 328         tcm->tcm_info = q->queues[cl - 1]->handle;
 329         return 0;
 330 }
 331
 332 static int multiq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
 333                                  struct gnet_dump *d)
 334 {
 335         struct multiq_sched_data *q = qdisc_priv(sch);
 336         struct Qdisc *cl_q;
 337
 338         cl_q = q->queues[cl - 1];
 339         if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
 340                                   d, NULL, &cl_q->bstats) < 0 ||
 341             gnet_stats_copy_queue(d, NULL, &cl_q->qstats, cl_q->q.qlen) < 0)
 342                 return -1;
 343
 344         return 0;
 345 }
 346
 347 static void multiq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
 348 {
 349         struct multiq_sched_data *q = qdisc_priv(sch);
 350         int band;
 351
 352         if (arg->stop)
 353                 return;
 354
 355         for (band = 0; band < q->bands; band++) {
 356                 if (arg->count < arg->skip) {
 357                         arg->count++;
 358                         continue;
 359                 }
 360                 if (arg->fn(sch, band + 1, arg) < 0) {
 361                         arg->stop = 1;
 362                         break;
 363                 }
 364                 arg->count++;
 365         }
 366 }
 367
 368 static struct tcf_proto __rcu **multiq_find_tcf(struct Qdisc *sch,
 369                                                 unsigned long cl)
 370 {
 371         struct multiq_sched_data *q = qdisc_priv(sch);
 372
 373         if (cl)
 374                 return NULL;
 375         return &q->filter_list;
 376 }
 377
 378 static const struct Qdisc_class_ops multiq_class_ops = {
 379         .graft          =       multiq_graft,
 380         .leaf           =       multiq_leaf,
 381         .get            =       multiq_get,
 382         .put            =       multiq_put,
 383         .walk           =       multiq_walk,
 384         .tcf_chain      =       multiq_find_tcf,
 385         .bind_tcf       =       multiq_bind,
 386         .unbind_tcf     =       multiq_put,
 387         .dump           =       multiq_dump_class,
 388         .dump_stats     =       multiq_dump_class_stats,
 389 };
 390
 391 static struct Qdisc_ops multiq_qdisc_ops __read_mostly = {
 392         .next           =       NULL,
 393         .cl_ops         =       &multiq_class_ops,
 394         .id             =       "multiq",
 395         .priv_size      =       sizeof(struct multiq_sched_data),
 396         .enqueue        =       multiq_enqueue,
 397         .dequeue        =       multiq_dequeue,
 398         .peek           =       multiq_peek,
 399         .init           =       multiq_init,
 400         .reset          =       multiq_reset,
 401         .destroy        =       multiq_destroy,
 402         .change         =       multiq_tune,
 403         .dump           =       multiq_dump,
 404         .owner          =       THIS_MODULE,
 405 };
 406
 407 static int __init multiq_module_init(void)
 408 {
 409         return register_qdisc(&multiq_qdisc_ops);
 410 }
 411
 412 static void __exit multiq_module_exit(void)
 413 {
 414         unregister_qdisc(&multiq_qdisc_ops);
 415 }
 416
 417 module_init(multiq_module_init)
 418 module_exit(multiq_module_exit)
 419
 420 MODULE_LICENSE("GPL");