Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * net/sched/sch_api.c Packet scheduler API. |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or |
| 5 | * modify it under the terms of the GNU General Public License |
| 6 | * as published by the Free Software Foundation; either version |
| 7 | * 2 of the License, or (at your option) any later version. |
| 8 | * |
| 9 | * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> |
| 10 | * |
| 11 | * Fixes: |
| 12 | * |
| 13 | * Rani Assaf <rani@magic.metawire.com> :980802: JIFFIES and CPU clock sources are repaired. |
| 14 | * Eduardo J. Blanco <ejbs@netlabs.com.uy> :990222: kmod support |
| 15 | * Jamal Hadi Salim <hadi@nortelnetworks.com>: 990601: ingress support |
| 16 | */ |
| 17 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 18 | #include <linux/module.h> |
| 19 | #include <linux/types.h> |
| 20 | #include <linux/kernel.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 21 | #include <linux/string.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 22 | #include <linux/errno.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 23 | #include <linux/skbuff.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 24 | #include <linux/init.h> |
| 25 | #include <linux/proc_fs.h> |
| 26 | #include <linux/seq_file.h> |
| 27 | #include <linux/kmod.h> |
| 28 | #include <linux/list.h> |
Patrick McHardy | 4179477 | 2007-03-16 01:19:15 -0700 | [diff] [blame] | 29 | #include <linux/hrtimer.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 30 | |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame] | 31 | #include <net/net_namespace.h> |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 32 | #include <net/sock.h> |
Arnaldo Carvalho de Melo | dc5fc57 | 2007-03-25 23:06:12 -0700 | [diff] [blame] | 33 | #include <net/netlink.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 34 | #include <net/pkt_sched.h> |
| 35 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 36 | static int qdisc_notify(struct sk_buff *oskb, struct nlmsghdr *n, u32 clid, |
| 37 | struct Qdisc *old, struct Qdisc *new); |
| 38 | static int tclass_notify(struct sk_buff *oskb, struct nlmsghdr *n, |
| 39 | struct Qdisc *q, unsigned long cl, int event); |
| 40 | |
| 41 | /* |
| 42 | |
| 43 | Short review. |
| 44 | ------------- |
| 45 | |
| 46 | This file consists of two interrelated parts: |
| 47 | |
| 48 | 1. queueing disciplines manager frontend. |
| 49 | 2. traffic classes manager frontend. |
| 50 | |
| 51 | Generally, queueing discipline ("qdisc") is a black box, |
| 52 | which is able to enqueue packets and to dequeue them (when |
| 53 | device is ready to send something) in order and at times |
| 54 | determined by algorithm hidden in it. |
| 55 | |
| 56 | qdisc's are divided to two categories: |
| 57 | - "queues", which have no internal structure visible from outside. |
| 58 | - "schedulers", which split all the packets to "traffic classes", |
| 59 | using "packet classifiers" (look at cls_api.c) |
| 60 | |
| 61 | In turn, classes may have child qdiscs (as rule, queues) |
| 62 | attached to them etc. etc. etc. |
| 63 | |
| 64 | The goal of the routines in this file is to translate |
| 65 | information supplied by user in the form of handles |
| 66 | to more intelligible for kernel form, to make some sanity |
| 67 | checks and part of work, which is common to all qdiscs |
| 68 | and to provide rtnetlink notifications. |
| 69 | |
| 70 | All real intelligent work is done inside qdisc modules. |
| 71 | |
| 72 | |
| 73 | |
| 74 | Every discipline has two major routines: enqueue and dequeue. |
| 75 | |
| 76 | ---dequeue |
| 77 | |
| 78 | dequeue usually returns a skb to send. It is allowed to return NULL, |
| 79 | but it does not mean that queue is empty, it just means that |
| 80 | discipline does not want to send anything this time. |
| 81 | Queue is really empty if q->q.qlen == 0. |
| 82 | For complicated disciplines with multiple queues q->q is not |
| 83 | real packet queue, but however q->q.qlen must be valid. |
| 84 | |
| 85 | ---enqueue |
| 86 | |
| 87 | enqueue returns 0, if packet was enqueued successfully. |
| 88 | If packet (this one or another one) was dropped, it returns |
| 89 | not zero error code. |
| 90 | NET_XMIT_DROP - this packet dropped |
| 91 | Expected action: do not backoff, but wait until queue will clear. |
| 92 | NET_XMIT_CN - probably this packet enqueued, but another one dropped. |
| 93 | Expected action: backoff or ignore |
| 94 | NET_XMIT_POLICED - dropped by police. |
| 95 | Expected action: backoff or error to real-time apps. |
| 96 | |
| 97 | Auxiliary routines: |
| 98 | |
| 99 | ---requeue |
| 100 | |
| 101 | requeues once dequeued packet. It is used for non-standard or |
| 102 | just buggy devices, which can defer output even if dev->tbusy=0. |
| 103 | |
| 104 | ---reset |
| 105 | |
| 106 | returns qdisc to initial state: purge all buffers, clear all |
| 107 | timers, counters (except for statistics) etc. |
| 108 | |
| 109 | ---init |
| 110 | |
| 111 | initializes newly created qdisc. |
| 112 | |
| 113 | ---destroy |
| 114 | |
| 115 | destroys resources allocated by init and during lifetime of qdisc. |
| 116 | |
| 117 | ---change |
| 118 | |
| 119 | changes qdisc parameters. |
| 120 | */ |
| 121 | |
| 122 | /* Protects list of registered TC modules. It is pure SMP lock. */ |
| 123 | static DEFINE_RWLOCK(qdisc_mod_lock); |
| 124 | |
| 125 | |
| 126 | /************************************************ |
| 127 | * Queueing disciplines manipulation. * |
| 128 | ************************************************/ |
| 129 | |
| 130 | |
| 131 | /* The list of all installed queueing disciplines. */ |
| 132 | |
| 133 | static struct Qdisc_ops *qdisc_base; |
| 134 | |
| 135 | /* Register/uregister queueing discipline */ |
| 136 | |
| 137 | int register_qdisc(struct Qdisc_ops *qops) |
| 138 | { |
| 139 | struct Qdisc_ops *q, **qp; |
| 140 | int rc = -EEXIST; |
| 141 | |
| 142 | write_lock(&qdisc_mod_lock); |
| 143 | for (qp = &qdisc_base; (q = *qp) != NULL; qp = &q->next) |
| 144 | if (!strcmp(qops->id, q->id)) |
| 145 | goto out; |
| 146 | |
| 147 | if (qops->enqueue == NULL) |
| 148 | qops->enqueue = noop_qdisc_ops.enqueue; |
| 149 | if (qops->requeue == NULL) |
| 150 | qops->requeue = noop_qdisc_ops.requeue; |
| 151 | if (qops->dequeue == NULL) |
| 152 | qops->dequeue = noop_qdisc_ops.dequeue; |
| 153 | |
| 154 | qops->next = NULL; |
| 155 | *qp = qops; |
| 156 | rc = 0; |
| 157 | out: |
| 158 | write_unlock(&qdisc_mod_lock); |
| 159 | return rc; |
| 160 | } |
Patrick McHardy | 62e3ba1 | 2008-01-22 22:10:23 -0800 | [diff] [blame] | 161 | EXPORT_SYMBOL(register_qdisc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | |
| 163 | int unregister_qdisc(struct Qdisc_ops *qops) |
| 164 | { |
| 165 | struct Qdisc_ops *q, **qp; |
| 166 | int err = -ENOENT; |
| 167 | |
| 168 | write_lock(&qdisc_mod_lock); |
| 169 | for (qp = &qdisc_base; (q=*qp)!=NULL; qp = &q->next) |
| 170 | if (q == qops) |
| 171 | break; |
| 172 | if (q) { |
| 173 | *qp = q->next; |
| 174 | q->next = NULL; |
| 175 | err = 0; |
| 176 | } |
| 177 | write_unlock(&qdisc_mod_lock); |
| 178 | return err; |
| 179 | } |
Patrick McHardy | 62e3ba1 | 2008-01-22 22:10:23 -0800 | [diff] [blame] | 180 | EXPORT_SYMBOL(unregister_qdisc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 181 | |
| 182 | /* We know handle. Find qdisc among all qdisc's attached to device |
| 183 | (root qdisc, all its children, children of children etc.) |
| 184 | */ |
| 185 | |
Patrick McHardy | 0463d4a | 2007-04-16 17:02:10 -0700 | [diff] [blame] | 186 | struct Qdisc *qdisc_lookup(struct net_device *dev, u32 handle) |
Patrick McHardy | 43effa1 | 2006-11-29 17:35:48 -0800 | [diff] [blame] | 187 | { |
| 188 | struct Qdisc *q; |
| 189 | |
| 190 | list_for_each_entry(q, &dev->qdisc_list, list) { |
| 191 | if (q->handle == handle) |
| 192 | return q; |
| 193 | } |
| 194 | return NULL; |
| 195 | } |
| 196 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 197 | static struct Qdisc *qdisc_leaf(struct Qdisc *p, u32 classid) |
| 198 | { |
| 199 | unsigned long cl; |
| 200 | struct Qdisc *leaf; |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 201 | const struct Qdisc_class_ops *cops = p->ops->cl_ops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 202 | |
| 203 | if (cops == NULL) |
| 204 | return NULL; |
| 205 | cl = cops->get(p, classid); |
| 206 | |
| 207 | if (cl == 0) |
| 208 | return NULL; |
| 209 | leaf = cops->leaf(p, cl); |
| 210 | cops->put(p, cl); |
| 211 | return leaf; |
| 212 | } |
| 213 | |
| 214 | /* Find queueing discipline by name */ |
| 215 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 216 | static struct Qdisc_ops *qdisc_lookup_ops(struct nlattr *kind) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 217 | { |
| 218 | struct Qdisc_ops *q = NULL; |
| 219 | |
| 220 | if (kind) { |
| 221 | read_lock(&qdisc_mod_lock); |
| 222 | for (q = qdisc_base; q; q = q->next) { |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 223 | if (nla_strcmp(kind, q->id) == 0) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 224 | if (!try_module_get(q->owner)) |
| 225 | q = NULL; |
| 226 | break; |
| 227 | } |
| 228 | } |
| 229 | read_unlock(&qdisc_mod_lock); |
| 230 | } |
| 231 | return q; |
| 232 | } |
| 233 | |
| 234 | static struct qdisc_rate_table *qdisc_rtab_list; |
| 235 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 236 | struct qdisc_rate_table *qdisc_get_rtab(struct tc_ratespec *r, struct nlattr *tab) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 237 | { |
| 238 | struct qdisc_rate_table *rtab; |
| 239 | |
| 240 | for (rtab = qdisc_rtab_list; rtab; rtab = rtab->next) { |
| 241 | if (memcmp(&rtab->rate, r, sizeof(struct tc_ratespec)) == 0) { |
| 242 | rtab->refcnt++; |
| 243 | return rtab; |
| 244 | } |
| 245 | } |
| 246 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 247 | if (tab == NULL || r->rate == 0 || r->cell_log == 0 || nla_len(tab) != 1024) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 248 | return NULL; |
| 249 | |
| 250 | rtab = kmalloc(sizeof(*rtab), GFP_KERNEL); |
| 251 | if (rtab) { |
| 252 | rtab->rate = *r; |
| 253 | rtab->refcnt = 1; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 254 | memcpy(rtab->data, nla_data(tab), 1024); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 255 | rtab->next = qdisc_rtab_list; |
| 256 | qdisc_rtab_list = rtab; |
| 257 | } |
| 258 | return rtab; |
| 259 | } |
Patrick McHardy | 62e3ba1 | 2008-01-22 22:10:23 -0800 | [diff] [blame] | 260 | EXPORT_SYMBOL(qdisc_get_rtab); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 261 | |
| 262 | void qdisc_put_rtab(struct qdisc_rate_table *tab) |
| 263 | { |
| 264 | struct qdisc_rate_table *rtab, **rtabp; |
| 265 | |
| 266 | if (!tab || --tab->refcnt) |
| 267 | return; |
| 268 | |
| 269 | for (rtabp = &qdisc_rtab_list; (rtab=*rtabp) != NULL; rtabp = &rtab->next) { |
| 270 | if (rtab == tab) { |
| 271 | *rtabp = rtab->next; |
| 272 | kfree(rtab); |
| 273 | return; |
| 274 | } |
| 275 | } |
| 276 | } |
Patrick McHardy | 62e3ba1 | 2008-01-22 22:10:23 -0800 | [diff] [blame] | 277 | EXPORT_SYMBOL(qdisc_put_rtab); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | |
Patrick McHardy | 4179477 | 2007-03-16 01:19:15 -0700 | [diff] [blame] | 279 | static enum hrtimer_restart qdisc_watchdog(struct hrtimer *timer) |
| 280 | { |
| 281 | struct qdisc_watchdog *wd = container_of(timer, struct qdisc_watchdog, |
| 282 | timer); |
Stephen Hemminger | 1936502 | 2007-03-22 12:18:35 -0700 | [diff] [blame] | 283 | struct net_device *dev = wd->qdisc->dev; |
Patrick McHardy | 4179477 | 2007-03-16 01:19:15 -0700 | [diff] [blame] | 284 | |
| 285 | wd->qdisc->flags &= ~TCQ_F_THROTTLED; |
Stephen Hemminger | 11274e5 | 2007-03-22 12:17:42 -0700 | [diff] [blame] | 286 | smp_wmb(); |
Patrick McHardy | 0621ed2 | 2007-07-14 20:49:26 -0700 | [diff] [blame] | 287 | netif_schedule(dev); |
Stephen Hemminger | 1936502 | 2007-03-22 12:18:35 -0700 | [diff] [blame] | 288 | |
Patrick McHardy | 4179477 | 2007-03-16 01:19:15 -0700 | [diff] [blame] | 289 | return HRTIMER_NORESTART; |
| 290 | } |
| 291 | |
| 292 | void qdisc_watchdog_init(struct qdisc_watchdog *wd, struct Qdisc *qdisc) |
| 293 | { |
| 294 | hrtimer_init(&wd->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS); |
| 295 | wd->timer.function = qdisc_watchdog; |
| 296 | wd->qdisc = qdisc; |
| 297 | } |
| 298 | EXPORT_SYMBOL(qdisc_watchdog_init); |
| 299 | |
| 300 | void qdisc_watchdog_schedule(struct qdisc_watchdog *wd, psched_time_t expires) |
| 301 | { |
| 302 | ktime_t time; |
| 303 | |
| 304 | wd->qdisc->flags |= TCQ_F_THROTTLED; |
| 305 | time = ktime_set(0, 0); |
| 306 | time = ktime_add_ns(time, PSCHED_US2NS(expires)); |
| 307 | hrtimer_start(&wd->timer, time, HRTIMER_MODE_ABS); |
| 308 | } |
| 309 | EXPORT_SYMBOL(qdisc_watchdog_schedule); |
| 310 | |
| 311 | void qdisc_watchdog_cancel(struct qdisc_watchdog *wd) |
| 312 | { |
| 313 | hrtimer_cancel(&wd->timer); |
| 314 | wd->qdisc->flags &= ~TCQ_F_THROTTLED; |
| 315 | } |
| 316 | EXPORT_SYMBOL(qdisc_watchdog_cancel); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 317 | |
| 318 | /* Allocate an unique handle from space managed by kernel */ |
| 319 | |
| 320 | static u32 qdisc_alloc_handle(struct net_device *dev) |
| 321 | { |
| 322 | int i = 0x10000; |
| 323 | static u32 autohandle = TC_H_MAKE(0x80000000U, 0); |
| 324 | |
| 325 | do { |
| 326 | autohandle += TC_H_MAKE(0x10000U, 0); |
| 327 | if (autohandle == TC_H_MAKE(TC_H_ROOT, 0)) |
| 328 | autohandle = TC_H_MAKE(0x80000000U, 0); |
| 329 | } while (qdisc_lookup(dev, autohandle) && --i > 0); |
| 330 | |
| 331 | return i>0 ? autohandle : 0; |
| 332 | } |
| 333 | |
| 334 | /* Attach toplevel qdisc to device dev */ |
| 335 | |
| 336 | static struct Qdisc * |
| 337 | dev_graft_qdisc(struct net_device *dev, struct Qdisc *qdisc) |
| 338 | { |
| 339 | struct Qdisc *oqdisc; |
| 340 | |
| 341 | if (dev->flags & IFF_UP) |
| 342 | dev_deactivate(dev); |
| 343 | |
| 344 | qdisc_lock_tree(dev); |
| 345 | if (qdisc && qdisc->flags&TCQ_F_INGRESS) { |
| 346 | oqdisc = dev->qdisc_ingress; |
| 347 | /* Prune old scheduler */ |
| 348 | if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1) { |
| 349 | /* delete */ |
| 350 | qdisc_reset(oqdisc); |
| 351 | dev->qdisc_ingress = NULL; |
| 352 | } else { /* new */ |
| 353 | dev->qdisc_ingress = qdisc; |
| 354 | } |
| 355 | |
| 356 | } else { |
| 357 | |
| 358 | oqdisc = dev->qdisc_sleeping; |
| 359 | |
| 360 | /* Prune old scheduler */ |
| 361 | if (oqdisc && atomic_read(&oqdisc->refcnt) <= 1) |
| 362 | qdisc_reset(oqdisc); |
| 363 | |
| 364 | /* ... and graft new one */ |
| 365 | if (qdisc == NULL) |
| 366 | qdisc = &noop_qdisc; |
| 367 | dev->qdisc_sleeping = qdisc; |
| 368 | dev->qdisc = &noop_qdisc; |
| 369 | } |
| 370 | |
| 371 | qdisc_unlock_tree(dev); |
| 372 | |
| 373 | if (dev->flags & IFF_UP) |
| 374 | dev_activate(dev); |
| 375 | |
| 376 | return oqdisc; |
| 377 | } |
| 378 | |
Patrick McHardy | 43effa1 | 2006-11-29 17:35:48 -0800 | [diff] [blame] | 379 | void qdisc_tree_decrease_qlen(struct Qdisc *sch, unsigned int n) |
| 380 | { |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 381 | const struct Qdisc_class_ops *cops; |
Patrick McHardy | 43effa1 | 2006-11-29 17:35:48 -0800 | [diff] [blame] | 382 | unsigned long cl; |
| 383 | u32 parentid; |
| 384 | |
| 385 | if (n == 0) |
| 386 | return; |
| 387 | while ((parentid = sch->parent)) { |
Patrick McHardy | 0463d4a | 2007-04-16 17:02:10 -0700 | [diff] [blame] | 388 | sch = qdisc_lookup(sch->dev, TC_H_MAJ(parentid)); |
Patrick McHardy | ffc8fef | 2007-07-30 17:11:50 -0700 | [diff] [blame] | 389 | if (sch == NULL) { |
| 390 | WARN_ON(parentid != TC_H_ROOT); |
| 391 | return; |
| 392 | } |
Patrick McHardy | 43effa1 | 2006-11-29 17:35:48 -0800 | [diff] [blame] | 393 | cops = sch->ops->cl_ops; |
| 394 | if (cops->qlen_notify) { |
| 395 | cl = cops->get(sch, parentid); |
| 396 | cops->qlen_notify(sch, cl); |
| 397 | cops->put(sch, cl); |
| 398 | } |
| 399 | sch->q.qlen -= n; |
| 400 | } |
| 401 | } |
| 402 | EXPORT_SYMBOL(qdisc_tree_decrease_qlen); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 403 | |
| 404 | /* Graft qdisc "new" to class "classid" of qdisc "parent" or |
| 405 | to device "dev". |
| 406 | |
| 407 | Old qdisc is not destroyed but returned in *old. |
| 408 | */ |
| 409 | |
| 410 | static int qdisc_graft(struct net_device *dev, struct Qdisc *parent, |
| 411 | u32 classid, |
| 412 | struct Qdisc *new, struct Qdisc **old) |
| 413 | { |
| 414 | int err = 0; |
| 415 | struct Qdisc *q = *old; |
| 416 | |
| 417 | |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 418 | if (parent == NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 419 | if (q && q->flags&TCQ_F_INGRESS) { |
| 420 | *old = dev_graft_qdisc(dev, q); |
| 421 | } else { |
| 422 | *old = dev_graft_qdisc(dev, new); |
| 423 | } |
| 424 | } else { |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 425 | const struct Qdisc_class_ops *cops = parent->ops->cl_ops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 426 | |
| 427 | err = -EINVAL; |
| 428 | |
| 429 | if (cops) { |
| 430 | unsigned long cl = cops->get(parent, classid); |
| 431 | if (cl) { |
| 432 | err = cops->graft(parent, cl, new, old); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 433 | cops->put(parent, cl); |
| 434 | } |
| 435 | } |
| 436 | } |
| 437 | return err; |
| 438 | } |
| 439 | |
| 440 | /* |
| 441 | Allocate and initialize new qdisc. |
| 442 | |
| 443 | Parameters are passed via opt. |
| 444 | */ |
| 445 | |
| 446 | static struct Qdisc * |
Patrick McHardy | ffc8fef | 2007-07-30 17:11:50 -0700 | [diff] [blame] | 447 | qdisc_create(struct net_device *dev, u32 parent, u32 handle, |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 448 | struct nlattr **tca, int *errp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 449 | { |
| 450 | int err; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 451 | struct nlattr *kind = tca[TCA_KIND]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 452 | struct Qdisc *sch; |
| 453 | struct Qdisc_ops *ops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 454 | |
| 455 | ops = qdisc_lookup_ops(kind); |
| 456 | #ifdef CONFIG_KMOD |
| 457 | if (ops == NULL && kind != NULL) { |
| 458 | char name[IFNAMSIZ]; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 459 | if (nla_strlcpy(name, kind, IFNAMSIZ) < IFNAMSIZ) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 460 | /* We dropped the RTNL semaphore in order to |
| 461 | * perform the module load. So, even if we |
| 462 | * succeeded in loading the module we have to |
| 463 | * tell the caller to replay the request. We |
| 464 | * indicate this using -EAGAIN. |
| 465 | * We replay the request because the device may |
| 466 | * go away in the mean time. |
| 467 | */ |
| 468 | rtnl_unlock(); |
| 469 | request_module("sch_%s", name); |
| 470 | rtnl_lock(); |
| 471 | ops = qdisc_lookup_ops(kind); |
| 472 | if (ops != NULL) { |
| 473 | /* We will try again qdisc_lookup_ops, |
| 474 | * so don't keep a reference. |
| 475 | */ |
| 476 | module_put(ops->owner); |
| 477 | err = -EAGAIN; |
| 478 | goto err_out; |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | #endif |
| 483 | |
Jamal Hadi Salim | b9e2cc0 | 2006-08-03 16:36:51 -0700 | [diff] [blame] | 484 | err = -ENOENT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 485 | if (ops == NULL) |
| 486 | goto err_out; |
| 487 | |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 488 | sch = qdisc_alloc(dev, ops); |
| 489 | if (IS_ERR(sch)) { |
| 490 | err = PTR_ERR(sch); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 491 | goto err_out2; |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 492 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 493 | |
Patrick McHardy | ffc8fef | 2007-07-30 17:11:50 -0700 | [diff] [blame] | 494 | sch->parent = parent; |
| 495 | |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 496 | if (handle == TC_H_INGRESS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 497 | sch->flags |= TCQ_F_INGRESS; |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 498 | sch->stats_lock = &dev->ingress_lock; |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 499 | handle = TC_H_MAKE(TC_H_INGRESS, 0); |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 500 | } else { |
| 501 | sch->stats_lock = &dev->queue_lock; |
| 502 | if (handle == 0) { |
| 503 | handle = qdisc_alloc_handle(dev); |
| 504 | err = -ENOMEM; |
| 505 | if (handle == 0) |
| 506 | goto err_out3; |
| 507 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 508 | } |
| 509 | |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 510 | sch->handle = handle; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 511 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 512 | if (!ops->init || (err = ops->init(sch, tca[TCA_OPTIONS])) == 0) { |
| 513 | if (tca[TCA_RATE]) { |
Thomas Graf | 023e09a | 2005-07-05 14:15:53 -0700 | [diff] [blame] | 514 | err = gen_new_estimator(&sch->bstats, &sch->rate_est, |
| 515 | sch->stats_lock, |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 516 | tca[TCA_RATE]); |
Thomas Graf | 023e09a | 2005-07-05 14:15:53 -0700 | [diff] [blame] | 517 | if (err) { |
| 518 | /* |
| 519 | * Any broken qdiscs that would require |
| 520 | * a ops->reset() here? The qdisc was never |
| 521 | * in action so it shouldn't be necessary. |
| 522 | */ |
| 523 | if (ops->destroy) |
| 524 | ops->destroy(sch); |
| 525 | goto err_out3; |
| 526 | } |
| 527 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 528 | qdisc_lock_tree(dev); |
| 529 | list_add_tail(&sch->list, &dev->qdisc_list); |
| 530 | qdisc_unlock_tree(dev); |
| 531 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 532 | return sch; |
| 533 | } |
| 534 | err_out3: |
| 535 | dev_put(dev); |
Thomas Graf | 3d54b82 | 2005-07-05 14:15:09 -0700 | [diff] [blame] | 536 | kfree((char *) sch - sch->padded); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 537 | err_out2: |
| 538 | module_put(ops->owner); |
| 539 | err_out: |
| 540 | *errp = err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 541 | return NULL; |
| 542 | } |
| 543 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 544 | static int qdisc_change(struct Qdisc *sch, struct nlattr **tca) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 545 | { |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 546 | if (tca[TCA_OPTIONS]) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 547 | int err; |
| 548 | |
| 549 | if (sch->ops->change == NULL) |
| 550 | return -EINVAL; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 551 | err = sch->ops->change(sch, tca[TCA_OPTIONS]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 552 | if (err) |
| 553 | return err; |
| 554 | } |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 555 | if (tca[TCA_RATE]) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 556 | gen_replace_estimator(&sch->bstats, &sch->rate_est, |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 557 | sch->stats_lock, tca[TCA_RATE]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 558 | return 0; |
| 559 | } |
| 560 | |
| 561 | struct check_loop_arg |
| 562 | { |
| 563 | struct qdisc_walker w; |
| 564 | struct Qdisc *p; |
| 565 | int depth; |
| 566 | }; |
| 567 | |
| 568 | static int check_loop_fn(struct Qdisc *q, unsigned long cl, struct qdisc_walker *w); |
| 569 | |
| 570 | static int check_loop(struct Qdisc *q, struct Qdisc *p, int depth) |
| 571 | { |
| 572 | struct check_loop_arg arg; |
| 573 | |
| 574 | if (q->ops->cl_ops == NULL) |
| 575 | return 0; |
| 576 | |
| 577 | arg.w.stop = arg.w.skip = arg.w.count = 0; |
| 578 | arg.w.fn = check_loop_fn; |
| 579 | arg.depth = depth; |
| 580 | arg.p = p; |
| 581 | q->ops->cl_ops->walk(q, &arg.w); |
| 582 | return arg.w.stop ? -ELOOP : 0; |
| 583 | } |
| 584 | |
| 585 | static int |
| 586 | check_loop_fn(struct Qdisc *q, unsigned long cl, struct qdisc_walker *w) |
| 587 | { |
| 588 | struct Qdisc *leaf; |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 589 | const struct Qdisc_class_ops *cops = q->ops->cl_ops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 590 | struct check_loop_arg *arg = (struct check_loop_arg *)w; |
| 591 | |
| 592 | leaf = cops->leaf(q, cl); |
| 593 | if (leaf) { |
| 594 | if (leaf == arg->p || arg->depth > 7) |
| 595 | return -ELOOP; |
| 596 | return check_loop(leaf, arg->p, arg->depth + 1); |
| 597 | } |
| 598 | return 0; |
| 599 | } |
| 600 | |
| 601 | /* |
| 602 | * Delete/get qdisc. |
| 603 | */ |
| 604 | |
| 605 | static int tc_get_qdisc(struct sk_buff *skb, struct nlmsghdr *n, void *arg) |
| 606 | { |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 607 | struct net *net = skb->sk->sk_net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 608 | struct tcmsg *tcm = NLMSG_DATA(n); |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 609 | struct nlattr *tca[TCA_MAX + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 610 | struct net_device *dev; |
| 611 | u32 clid = tcm->tcm_parent; |
| 612 | struct Qdisc *q = NULL; |
| 613 | struct Qdisc *p = NULL; |
| 614 | int err; |
| 615 | |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 616 | if (net != &init_net) |
| 617 | return -EINVAL; |
| 618 | |
Eric W. Biederman | 881d966 | 2007-09-17 11:56:21 -0700 | [diff] [blame] | 619 | if ((dev = __dev_get_by_index(&init_net, tcm->tcm_ifindex)) == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 620 | return -ENODEV; |
| 621 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 622 | err = nlmsg_parse(n, sizeof(*tcm), tca, TCA_MAX, NULL); |
| 623 | if (err < 0) |
| 624 | return err; |
| 625 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 626 | if (clid) { |
| 627 | if (clid != TC_H_ROOT) { |
| 628 | if (TC_H_MAJ(clid) != TC_H_MAJ(TC_H_INGRESS)) { |
| 629 | if ((p = qdisc_lookup(dev, TC_H_MAJ(clid))) == NULL) |
| 630 | return -ENOENT; |
| 631 | q = qdisc_leaf(p, clid); |
| 632 | } else { /* ingress */ |
| 633 | q = dev->qdisc_ingress; |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 634 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 635 | } else { |
| 636 | q = dev->qdisc_sleeping; |
| 637 | } |
| 638 | if (!q) |
| 639 | return -ENOENT; |
| 640 | |
| 641 | if (tcm->tcm_handle && q->handle != tcm->tcm_handle) |
| 642 | return -EINVAL; |
| 643 | } else { |
| 644 | if ((q = qdisc_lookup(dev, tcm->tcm_handle)) == NULL) |
| 645 | return -ENOENT; |
| 646 | } |
| 647 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 648 | if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], q->ops->id)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 649 | return -EINVAL; |
| 650 | |
| 651 | if (n->nlmsg_type == RTM_DELQDISC) { |
| 652 | if (!clid) |
| 653 | return -EINVAL; |
| 654 | if (q->handle == 0) |
| 655 | return -ENOENT; |
| 656 | if ((err = qdisc_graft(dev, p, clid, NULL, &q)) != 0) |
| 657 | return err; |
| 658 | if (q) { |
| 659 | qdisc_notify(skb, n, clid, q, NULL); |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 660 | qdisc_lock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 661 | qdisc_destroy(q); |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 662 | qdisc_unlock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 663 | } |
| 664 | } else { |
| 665 | qdisc_notify(skb, n, clid, NULL, q); |
| 666 | } |
| 667 | return 0; |
| 668 | } |
| 669 | |
| 670 | /* |
| 671 | Create/change qdisc. |
| 672 | */ |
| 673 | |
| 674 | static int tc_modify_qdisc(struct sk_buff *skb, struct nlmsghdr *n, void *arg) |
| 675 | { |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 676 | struct net *net = skb->sk->sk_net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | struct tcmsg *tcm; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 678 | struct nlattr *tca[TCA_MAX + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 679 | struct net_device *dev; |
| 680 | u32 clid; |
| 681 | struct Qdisc *q, *p; |
| 682 | int err; |
| 683 | |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 684 | if (net != &init_net) |
| 685 | return -EINVAL; |
| 686 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 687 | replay: |
| 688 | /* Reinit, just in case something touches this. */ |
| 689 | tcm = NLMSG_DATA(n); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 690 | clid = tcm->tcm_parent; |
| 691 | q = p = NULL; |
| 692 | |
Eric W. Biederman | 881d966 | 2007-09-17 11:56:21 -0700 | [diff] [blame] | 693 | if ((dev = __dev_get_by_index(&init_net, tcm->tcm_ifindex)) == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 694 | return -ENODEV; |
| 695 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 696 | err = nlmsg_parse(n, sizeof(*tcm), tca, TCA_MAX, NULL); |
| 697 | if (err < 0) |
| 698 | return err; |
| 699 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 700 | if (clid) { |
| 701 | if (clid != TC_H_ROOT) { |
| 702 | if (clid != TC_H_INGRESS) { |
| 703 | if ((p = qdisc_lookup(dev, TC_H_MAJ(clid))) == NULL) |
| 704 | return -ENOENT; |
| 705 | q = qdisc_leaf(p, clid); |
| 706 | } else { /*ingress */ |
| 707 | q = dev->qdisc_ingress; |
| 708 | } |
| 709 | } else { |
| 710 | q = dev->qdisc_sleeping; |
| 711 | } |
| 712 | |
| 713 | /* It may be default qdisc, ignore it */ |
| 714 | if (q && q->handle == 0) |
| 715 | q = NULL; |
| 716 | |
| 717 | if (!q || !tcm->tcm_handle || q->handle != tcm->tcm_handle) { |
| 718 | if (tcm->tcm_handle) { |
| 719 | if (q && !(n->nlmsg_flags&NLM_F_REPLACE)) |
| 720 | return -EEXIST; |
| 721 | if (TC_H_MIN(tcm->tcm_handle)) |
| 722 | return -EINVAL; |
| 723 | if ((q = qdisc_lookup(dev, tcm->tcm_handle)) == NULL) |
| 724 | goto create_n_graft; |
| 725 | if (n->nlmsg_flags&NLM_F_EXCL) |
| 726 | return -EEXIST; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 727 | if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], q->ops->id)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 728 | return -EINVAL; |
| 729 | if (q == p || |
| 730 | (p && check_loop(q, p, 0))) |
| 731 | return -ELOOP; |
| 732 | atomic_inc(&q->refcnt); |
| 733 | goto graft; |
| 734 | } else { |
| 735 | if (q == NULL) |
| 736 | goto create_n_graft; |
| 737 | |
| 738 | /* This magic test requires explanation. |
| 739 | * |
| 740 | * We know, that some child q is already |
| 741 | * attached to this parent and have choice: |
| 742 | * either to change it or to create/graft new one. |
| 743 | * |
| 744 | * 1. We are allowed to create/graft only |
| 745 | * if CREATE and REPLACE flags are set. |
| 746 | * |
| 747 | * 2. If EXCL is set, requestor wanted to say, |
| 748 | * that qdisc tcm_handle is not expected |
| 749 | * to exist, so that we choose create/graft too. |
| 750 | * |
| 751 | * 3. The last case is when no flags are set. |
| 752 | * Alas, it is sort of hole in API, we |
| 753 | * cannot decide what to do unambiguously. |
| 754 | * For now we select create/graft, if |
| 755 | * user gave KIND, which does not match existing. |
| 756 | */ |
| 757 | if ((n->nlmsg_flags&NLM_F_CREATE) && |
| 758 | (n->nlmsg_flags&NLM_F_REPLACE) && |
| 759 | ((n->nlmsg_flags&NLM_F_EXCL) || |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 760 | (tca[TCA_KIND] && |
| 761 | nla_strcmp(tca[TCA_KIND], q->ops->id)))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 762 | goto create_n_graft; |
| 763 | } |
| 764 | } |
| 765 | } else { |
| 766 | if (!tcm->tcm_handle) |
| 767 | return -EINVAL; |
| 768 | q = qdisc_lookup(dev, tcm->tcm_handle); |
| 769 | } |
| 770 | |
| 771 | /* Change qdisc parameters */ |
| 772 | if (q == NULL) |
| 773 | return -ENOENT; |
| 774 | if (n->nlmsg_flags&NLM_F_EXCL) |
| 775 | return -EEXIST; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 776 | if (tca[TCA_KIND] && nla_strcmp(tca[TCA_KIND], q->ops->id)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 777 | return -EINVAL; |
| 778 | err = qdisc_change(q, tca); |
| 779 | if (err == 0) |
| 780 | qdisc_notify(skb, n, clid, NULL, q); |
| 781 | return err; |
| 782 | |
| 783 | create_n_graft: |
| 784 | if (!(n->nlmsg_flags&NLM_F_CREATE)) |
| 785 | return -ENOENT; |
| 786 | if (clid == TC_H_INGRESS) |
Patrick McHardy | ffc8fef | 2007-07-30 17:11:50 -0700 | [diff] [blame] | 787 | q = qdisc_create(dev, tcm->tcm_parent, tcm->tcm_parent, |
| 788 | tca, &err); |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 789 | else |
Patrick McHardy | ffc8fef | 2007-07-30 17:11:50 -0700 | [diff] [blame] | 790 | q = qdisc_create(dev, tcm->tcm_parent, tcm->tcm_handle, |
| 791 | tca, &err); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 792 | if (q == NULL) { |
| 793 | if (err == -EAGAIN) |
| 794 | goto replay; |
| 795 | return err; |
| 796 | } |
| 797 | |
| 798 | graft: |
| 799 | if (1) { |
| 800 | struct Qdisc *old_q = NULL; |
| 801 | err = qdisc_graft(dev, p, clid, q, &old_q); |
| 802 | if (err) { |
| 803 | if (q) { |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 804 | qdisc_lock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 805 | qdisc_destroy(q); |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 806 | qdisc_unlock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 807 | } |
| 808 | return err; |
| 809 | } |
| 810 | qdisc_notify(skb, n, clid, old_q, q); |
| 811 | if (old_q) { |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 812 | qdisc_lock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 813 | qdisc_destroy(old_q); |
Patrick McHardy | fd44de7 | 2007-04-16 17:07:08 -0700 | [diff] [blame] | 814 | qdisc_unlock_tree(dev); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 815 | } |
| 816 | } |
| 817 | return 0; |
| 818 | } |
| 819 | |
| 820 | static int tc_fill_qdisc(struct sk_buff *skb, struct Qdisc *q, u32 clid, |
Jamal Hadi Salim | e431b8c | 2005-06-18 22:55:31 -0700 | [diff] [blame] | 821 | u32 pid, u32 seq, u16 flags, int event) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 822 | { |
| 823 | struct tcmsg *tcm; |
| 824 | struct nlmsghdr *nlh; |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 825 | unsigned char *b = skb_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 826 | struct gnet_dump d; |
| 827 | |
Jamal Hadi Salim | e431b8c | 2005-06-18 22:55:31 -0700 | [diff] [blame] | 828 | nlh = NLMSG_NEW(skb, pid, seq, event, sizeof(*tcm), flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 829 | tcm = NLMSG_DATA(nlh); |
| 830 | tcm->tcm_family = AF_UNSPEC; |
Patrick McHardy | 9ef1d4c | 2005-06-28 12:55:30 -0700 | [diff] [blame] | 831 | tcm->tcm__pad1 = 0; |
| 832 | tcm->tcm__pad2 = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 833 | tcm->tcm_ifindex = q->dev->ifindex; |
| 834 | tcm->tcm_parent = clid; |
| 835 | tcm->tcm_handle = q->handle; |
| 836 | tcm->tcm_info = atomic_read(&q->refcnt); |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 837 | NLA_PUT(skb, TCA_KIND, IFNAMSIZ, q->ops->id); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 838 | if (q->ops->dump && q->ops->dump(q, skb) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 839 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 840 | q->qstats.qlen = q->q.qlen; |
| 841 | |
| 842 | if (gnet_stats_start_copy_compat(skb, TCA_STATS2, TCA_STATS, |
| 843 | TCA_XSTATS, q->stats_lock, &d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 844 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 845 | |
| 846 | if (q->ops->dump_stats && q->ops->dump_stats(q, &d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 847 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | |
| 849 | if (gnet_stats_copy_basic(&d, &q->bstats) < 0 || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 850 | gnet_stats_copy_rate_est(&d, &q->rate_est) < 0 || |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 851 | gnet_stats_copy_queue(&d, &q->qstats) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 852 | goto nla_put_failure; |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 853 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 854 | if (gnet_stats_finish_copy(&d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 855 | goto nla_put_failure; |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 856 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 857 | nlh->nlmsg_len = skb_tail_pointer(skb) - b; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 858 | return skb->len; |
| 859 | |
| 860 | nlmsg_failure: |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 861 | nla_put_failure: |
Arnaldo Carvalho de Melo | dc5fc57 | 2007-03-25 23:06:12 -0700 | [diff] [blame] | 862 | nlmsg_trim(skb, b); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 863 | return -1; |
| 864 | } |
| 865 | |
| 866 | static int qdisc_notify(struct sk_buff *oskb, struct nlmsghdr *n, |
| 867 | u32 clid, struct Qdisc *old, struct Qdisc *new) |
| 868 | { |
| 869 | struct sk_buff *skb; |
| 870 | u32 pid = oskb ? NETLINK_CB(oskb).pid : 0; |
| 871 | |
| 872 | skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL); |
| 873 | if (!skb) |
| 874 | return -ENOBUFS; |
| 875 | |
| 876 | if (old && old->handle) { |
| 877 | if (tc_fill_qdisc(skb, old, clid, pid, n->nlmsg_seq, 0, RTM_DELQDISC) < 0) |
| 878 | goto err_out; |
| 879 | } |
| 880 | if (new) { |
| 881 | if (tc_fill_qdisc(skb, new, clid, pid, n->nlmsg_seq, old ? NLM_F_REPLACE : 0, RTM_NEWQDISC) < 0) |
| 882 | goto err_out; |
| 883 | } |
| 884 | |
| 885 | if (skb->len) |
Denis V. Lunev | 97c53ca | 2007-11-19 22:26:51 -0800 | [diff] [blame] | 886 | return rtnetlink_send(skb, &init_net, pid, RTNLGRP_TC, n->nlmsg_flags&NLM_F_ECHO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 887 | |
| 888 | err_out: |
| 889 | kfree_skb(skb); |
| 890 | return -EINVAL; |
| 891 | } |
| 892 | |
| 893 | static int tc_dump_qdisc(struct sk_buff *skb, struct netlink_callback *cb) |
| 894 | { |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 895 | struct net *net = skb->sk->sk_net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 896 | int idx, q_idx; |
| 897 | int s_idx, s_q_idx; |
| 898 | struct net_device *dev; |
| 899 | struct Qdisc *q; |
| 900 | |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 901 | if (net != &init_net) |
| 902 | return 0; |
| 903 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 904 | s_idx = cb->args[0]; |
| 905 | s_q_idx = q_idx = cb->args[1]; |
| 906 | read_lock(&dev_base_lock); |
Pavel Emelianov | 7562f87 | 2007-05-03 15:13:45 -0700 | [diff] [blame] | 907 | idx = 0; |
Eric W. Biederman | 881d966 | 2007-09-17 11:56:21 -0700 | [diff] [blame] | 908 | for_each_netdev(&init_net, dev) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 909 | if (idx < s_idx) |
Pavel Emelianov | 7562f87 | 2007-05-03 15:13:45 -0700 | [diff] [blame] | 910 | goto cont; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 911 | if (idx > s_idx) |
| 912 | s_q_idx = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 913 | q_idx = 0; |
| 914 | list_for_each_entry(q, &dev->qdisc_list, list) { |
| 915 | if (q_idx < s_q_idx) { |
| 916 | q_idx++; |
| 917 | continue; |
| 918 | } |
| 919 | if (tc_fill_qdisc(skb, q, q->parent, NETLINK_CB(cb->skb).pid, |
Patrick McHardy | 0463d4a | 2007-04-16 17:02:10 -0700 | [diff] [blame] | 920 | cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWQDISC) <= 0) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 921 | goto done; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 922 | q_idx++; |
| 923 | } |
Pavel Emelianov | 7562f87 | 2007-05-03 15:13:45 -0700 | [diff] [blame] | 924 | cont: |
| 925 | idx++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 926 | } |
| 927 | |
| 928 | done: |
| 929 | read_unlock(&dev_base_lock); |
| 930 | |
| 931 | cb->args[0] = idx; |
| 932 | cb->args[1] = q_idx; |
| 933 | |
| 934 | return skb->len; |
| 935 | } |
| 936 | |
| 937 | |
| 938 | |
| 939 | /************************************************ |
| 940 | * Traffic classes manipulation. * |
| 941 | ************************************************/ |
| 942 | |
| 943 | |
| 944 | |
| 945 | static int tc_ctl_tclass(struct sk_buff *skb, struct nlmsghdr *n, void *arg) |
| 946 | { |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 947 | struct net *net = skb->sk->sk_net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 948 | struct tcmsg *tcm = NLMSG_DATA(n); |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 949 | struct nlattr *tca[TCA_MAX + 1]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 950 | struct net_device *dev; |
| 951 | struct Qdisc *q = NULL; |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 952 | const struct Qdisc_class_ops *cops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 953 | unsigned long cl = 0; |
| 954 | unsigned long new_cl; |
| 955 | u32 pid = tcm->tcm_parent; |
| 956 | u32 clid = tcm->tcm_handle; |
| 957 | u32 qid = TC_H_MAJ(clid); |
| 958 | int err; |
| 959 | |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 960 | if (net != &init_net) |
| 961 | return -EINVAL; |
| 962 | |
Eric W. Biederman | 881d966 | 2007-09-17 11:56:21 -0700 | [diff] [blame] | 963 | if ((dev = __dev_get_by_index(&init_net, tcm->tcm_ifindex)) == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 964 | return -ENODEV; |
| 965 | |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 966 | err = nlmsg_parse(n, sizeof(*tcm), tca, TCA_MAX, NULL); |
| 967 | if (err < 0) |
| 968 | return err; |
| 969 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 970 | /* |
| 971 | parent == TC_H_UNSPEC - unspecified parent. |
| 972 | parent == TC_H_ROOT - class is root, which has no parent. |
| 973 | parent == X:0 - parent is root class. |
| 974 | parent == X:Y - parent is a node in hierarchy. |
| 975 | parent == 0:Y - parent is X:Y, where X:0 is qdisc. |
| 976 | |
| 977 | handle == 0:0 - generate handle from kernel pool. |
| 978 | handle == 0:Y - class is X:Y, where X:0 is qdisc. |
| 979 | handle == X:Y - clear. |
| 980 | handle == X:0 - root class. |
| 981 | */ |
| 982 | |
| 983 | /* Step 1. Determine qdisc handle X:0 */ |
| 984 | |
| 985 | if (pid != TC_H_ROOT) { |
| 986 | u32 qid1 = TC_H_MAJ(pid); |
| 987 | |
| 988 | if (qid && qid1) { |
| 989 | /* If both majors are known, they must be identical. */ |
| 990 | if (qid != qid1) |
| 991 | return -EINVAL; |
| 992 | } else if (qid1) { |
| 993 | qid = qid1; |
| 994 | } else if (qid == 0) |
| 995 | qid = dev->qdisc_sleeping->handle; |
| 996 | |
| 997 | /* Now qid is genuine qdisc handle consistent |
| 998 | both with parent and child. |
| 999 | |
| 1000 | TC_H_MAJ(pid) still may be unspecified, complete it now. |
| 1001 | */ |
| 1002 | if (pid) |
| 1003 | pid = TC_H_MAKE(qid, pid); |
| 1004 | } else { |
| 1005 | if (qid == 0) |
| 1006 | qid = dev->qdisc_sleeping->handle; |
| 1007 | } |
| 1008 | |
| 1009 | /* OK. Locate qdisc */ |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 1010 | if ((q = qdisc_lookup(dev, qid)) == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1011 | return -ENOENT; |
| 1012 | |
| 1013 | /* An check that it supports classes */ |
| 1014 | cops = q->ops->cl_ops; |
| 1015 | if (cops == NULL) |
| 1016 | return -EINVAL; |
| 1017 | |
| 1018 | /* Now try to get class */ |
| 1019 | if (clid == 0) { |
| 1020 | if (pid == TC_H_ROOT) |
| 1021 | clid = qid; |
| 1022 | } else |
| 1023 | clid = TC_H_MAKE(qid, clid); |
| 1024 | |
| 1025 | if (clid) |
| 1026 | cl = cops->get(q, clid); |
| 1027 | |
| 1028 | if (cl == 0) { |
| 1029 | err = -ENOENT; |
| 1030 | if (n->nlmsg_type != RTM_NEWTCLASS || !(n->nlmsg_flags&NLM_F_CREATE)) |
| 1031 | goto out; |
| 1032 | } else { |
| 1033 | switch (n->nlmsg_type) { |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 1034 | case RTM_NEWTCLASS: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1035 | err = -EEXIST; |
| 1036 | if (n->nlmsg_flags&NLM_F_EXCL) |
| 1037 | goto out; |
| 1038 | break; |
| 1039 | case RTM_DELTCLASS: |
| 1040 | err = cops->delete(q, cl); |
| 1041 | if (err == 0) |
| 1042 | tclass_notify(skb, n, q, cl, RTM_DELTCLASS); |
| 1043 | goto out; |
| 1044 | case RTM_GETTCLASS: |
| 1045 | err = tclass_notify(skb, n, q, cl, RTM_NEWTCLASS); |
| 1046 | goto out; |
| 1047 | default: |
| 1048 | err = -EINVAL; |
| 1049 | goto out; |
| 1050 | } |
| 1051 | } |
| 1052 | |
| 1053 | new_cl = cl; |
| 1054 | err = cops->change(q, clid, pid, tca, &new_cl); |
| 1055 | if (err == 0) |
| 1056 | tclass_notify(skb, n, q, new_cl, RTM_NEWTCLASS); |
| 1057 | |
| 1058 | out: |
| 1059 | if (cl) |
| 1060 | cops->put(q, cl); |
| 1061 | |
| 1062 | return err; |
| 1063 | } |
| 1064 | |
| 1065 | |
| 1066 | static int tc_fill_tclass(struct sk_buff *skb, struct Qdisc *q, |
| 1067 | unsigned long cl, |
Jamal Hadi Salim | e431b8c | 2005-06-18 22:55:31 -0700 | [diff] [blame] | 1068 | u32 pid, u32 seq, u16 flags, int event) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1069 | { |
| 1070 | struct tcmsg *tcm; |
| 1071 | struct nlmsghdr *nlh; |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 1072 | unsigned char *b = skb_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1073 | struct gnet_dump d; |
Eric Dumazet | 20fea08 | 2007-11-14 01:44:41 -0800 | [diff] [blame] | 1074 | const struct Qdisc_class_ops *cl_ops = q->ops->cl_ops; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1075 | |
Jamal Hadi Salim | e431b8c | 2005-06-18 22:55:31 -0700 | [diff] [blame] | 1076 | nlh = NLMSG_NEW(skb, pid, seq, event, sizeof(*tcm), flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1077 | tcm = NLMSG_DATA(nlh); |
| 1078 | tcm->tcm_family = AF_UNSPEC; |
| 1079 | tcm->tcm_ifindex = q->dev->ifindex; |
| 1080 | tcm->tcm_parent = q->handle; |
| 1081 | tcm->tcm_handle = q->handle; |
| 1082 | tcm->tcm_info = 0; |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1083 | NLA_PUT(skb, TCA_KIND, IFNAMSIZ, q->ops->id); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1084 | if (cl_ops->dump && cl_ops->dump(q, cl, skb, tcm) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1085 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1086 | |
| 1087 | if (gnet_stats_start_copy_compat(skb, TCA_STATS2, TCA_STATS, |
| 1088 | TCA_XSTATS, q->stats_lock, &d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1089 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1090 | |
| 1091 | if (cl_ops->dump_stats && cl_ops->dump_stats(q, cl, &d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1092 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1093 | |
| 1094 | if (gnet_stats_finish_copy(&d) < 0) |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1095 | goto nla_put_failure; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1096 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 1097 | nlh->nlmsg_len = skb_tail_pointer(skb) - b; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1098 | return skb->len; |
| 1099 | |
| 1100 | nlmsg_failure: |
Patrick McHardy | 1e90474 | 2008-01-22 22:11:17 -0800 | [diff] [blame^] | 1101 | nla_put_failure: |
Arnaldo Carvalho de Melo | dc5fc57 | 2007-03-25 23:06:12 -0700 | [diff] [blame] | 1102 | nlmsg_trim(skb, b); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1103 | return -1; |
| 1104 | } |
| 1105 | |
| 1106 | static int tclass_notify(struct sk_buff *oskb, struct nlmsghdr *n, |
| 1107 | struct Qdisc *q, unsigned long cl, int event) |
| 1108 | { |
| 1109 | struct sk_buff *skb; |
| 1110 | u32 pid = oskb ? NETLINK_CB(oskb).pid : 0; |
| 1111 | |
| 1112 | skb = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL); |
| 1113 | if (!skb) |
| 1114 | return -ENOBUFS; |
| 1115 | |
| 1116 | if (tc_fill_tclass(skb, q, cl, pid, n->nlmsg_seq, 0, event) < 0) { |
| 1117 | kfree_skb(skb); |
| 1118 | return -EINVAL; |
| 1119 | } |
| 1120 | |
Denis V. Lunev | 97c53ca | 2007-11-19 22:26:51 -0800 | [diff] [blame] | 1121 | return rtnetlink_send(skb, &init_net, pid, RTNLGRP_TC, n->nlmsg_flags&NLM_F_ECHO); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1122 | } |
| 1123 | |
| 1124 | struct qdisc_dump_args |
| 1125 | { |
| 1126 | struct qdisc_walker w; |
| 1127 | struct sk_buff *skb; |
| 1128 | struct netlink_callback *cb; |
| 1129 | }; |
| 1130 | |
| 1131 | static int qdisc_class_dump(struct Qdisc *q, unsigned long cl, struct qdisc_walker *arg) |
| 1132 | { |
| 1133 | struct qdisc_dump_args *a = (struct qdisc_dump_args *)arg; |
| 1134 | |
| 1135 | return tc_fill_tclass(a->skb, q, cl, NETLINK_CB(a->cb->skb).pid, |
| 1136 | a->cb->nlh->nlmsg_seq, NLM_F_MULTI, RTM_NEWTCLASS); |
| 1137 | } |
| 1138 | |
| 1139 | static int tc_dump_tclass(struct sk_buff *skb, struct netlink_callback *cb) |
| 1140 | { |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 1141 | struct net *net = skb->sk->sk_net; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1142 | int t; |
| 1143 | int s_t; |
| 1144 | struct net_device *dev; |
| 1145 | struct Qdisc *q; |
| 1146 | struct tcmsg *tcm = (struct tcmsg*)NLMSG_DATA(cb->nlh); |
| 1147 | struct qdisc_dump_args arg; |
| 1148 | |
Denis V. Lunev | b854272 | 2007-12-01 00:21:31 +1100 | [diff] [blame] | 1149 | if (net != &init_net) |
| 1150 | return 0; |
| 1151 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1152 | if (cb->nlh->nlmsg_len < NLMSG_LENGTH(sizeof(*tcm))) |
| 1153 | return 0; |
Eric W. Biederman | 881d966 | 2007-09-17 11:56:21 -0700 | [diff] [blame] | 1154 | if ((dev = dev_get_by_index(&init_net, tcm->tcm_ifindex)) == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1155 | return 0; |
| 1156 | |
| 1157 | s_t = cb->args[0]; |
| 1158 | t = 0; |
| 1159 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1160 | list_for_each_entry(q, &dev->qdisc_list, list) { |
| 1161 | if (t < s_t || !q->ops->cl_ops || |
| 1162 | (tcm->tcm_parent && |
| 1163 | TC_H_MAJ(tcm->tcm_parent) != q->handle)) { |
| 1164 | t++; |
| 1165 | continue; |
| 1166 | } |
| 1167 | if (t > s_t) |
| 1168 | memset(&cb->args[1], 0, sizeof(cb->args)-sizeof(cb->args[0])); |
| 1169 | arg.w.fn = qdisc_class_dump; |
| 1170 | arg.skb = skb; |
| 1171 | arg.cb = cb; |
| 1172 | arg.w.stop = 0; |
| 1173 | arg.w.skip = cb->args[1]; |
| 1174 | arg.w.count = 0; |
| 1175 | q->ops->cl_ops->walk(q, &arg.w); |
| 1176 | cb->args[1] = arg.w.count; |
| 1177 | if (arg.w.stop) |
| 1178 | break; |
| 1179 | t++; |
| 1180 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1181 | |
| 1182 | cb->args[0] = t; |
| 1183 | |
| 1184 | dev_put(dev); |
| 1185 | return skb->len; |
| 1186 | } |
| 1187 | |
| 1188 | /* Main classifier routine: scans classifier chain attached |
| 1189 | to this qdisc, (optionally) tests for protocol and asks |
| 1190 | specific classifiers. |
| 1191 | */ |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1192 | int tc_classify_compat(struct sk_buff *skb, struct tcf_proto *tp, |
| 1193 | struct tcf_result *res) |
| 1194 | { |
| 1195 | __be16 protocol = skb->protocol; |
| 1196 | int err = 0; |
| 1197 | |
| 1198 | for (; tp; tp = tp->next) { |
| 1199 | if ((tp->protocol == protocol || |
| 1200 | tp->protocol == htons(ETH_P_ALL)) && |
| 1201 | (err = tp->classify(skb, tp, res)) >= 0) { |
| 1202 | #ifdef CONFIG_NET_CLS_ACT |
| 1203 | if (err != TC_ACT_RECLASSIFY && skb->tc_verd) |
| 1204 | skb->tc_verd = SET_TC_VERD(skb->tc_verd, 0); |
| 1205 | #endif |
| 1206 | return err; |
| 1207 | } |
| 1208 | } |
| 1209 | return -1; |
| 1210 | } |
| 1211 | EXPORT_SYMBOL(tc_classify_compat); |
| 1212 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1213 | int tc_classify(struct sk_buff *skb, struct tcf_proto *tp, |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1214 | struct tcf_result *res) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1215 | { |
| 1216 | int err = 0; |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1217 | __be16 protocol; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1218 | #ifdef CONFIG_NET_CLS_ACT |
| 1219 | struct tcf_proto *otp = tp; |
| 1220 | reclassify: |
| 1221 | #endif |
| 1222 | protocol = skb->protocol; |
| 1223 | |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1224 | err = tc_classify_compat(skb, tp, res); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1225 | #ifdef CONFIG_NET_CLS_ACT |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1226 | if (err == TC_ACT_RECLASSIFY) { |
| 1227 | u32 verd = G_TC_VERD(skb->tc_verd); |
| 1228 | tp = otp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1229 | |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1230 | if (verd++ >= MAX_REC_LOOP) { |
| 1231 | printk("rule prio %u protocol %02x reclassify loop, " |
| 1232 | "packet dropped\n", |
| 1233 | tp->prio&0xffff, ntohs(tp->protocol)); |
| 1234 | return TC_ACT_SHOT; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1235 | } |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1236 | skb->tc_verd = SET_TC_VERD(skb->tc_verd, verd); |
| 1237 | goto reclassify; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1238 | } |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1239 | #endif |
| 1240 | return err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1241 | } |
Patrick McHardy | 73ca491 | 2007-07-15 00:02:31 -0700 | [diff] [blame] | 1242 | EXPORT_SYMBOL(tc_classify); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1243 | |
Patrick McHardy | a48b5a6 | 2007-03-23 11:29:43 -0700 | [diff] [blame] | 1244 | void tcf_destroy(struct tcf_proto *tp) |
| 1245 | { |
| 1246 | tp->ops->destroy(tp); |
| 1247 | module_put(tp->ops->owner); |
| 1248 | kfree(tp); |
| 1249 | } |
| 1250 | |
| 1251 | void tcf_destroy_chain(struct tcf_proto *fl) |
| 1252 | { |
| 1253 | struct tcf_proto *tp; |
| 1254 | |
| 1255 | while ((tp = fl) != NULL) { |
| 1256 | fl = tp->next; |
| 1257 | tcf_destroy(tp); |
| 1258 | } |
| 1259 | } |
| 1260 | EXPORT_SYMBOL(tcf_destroy_chain); |
| 1261 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1262 | #ifdef CONFIG_PROC_FS |
| 1263 | static int psched_show(struct seq_file *seq, void *v) |
| 1264 | { |
Patrick McHardy | 3c0cfc1 | 2007-10-10 16:32:41 -0700 | [diff] [blame] | 1265 | struct timespec ts; |
| 1266 | |
| 1267 | hrtimer_get_res(CLOCK_MONOTONIC, &ts); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1268 | seq_printf(seq, "%08x %08x %08x %08x\n", |
Patrick McHardy | 641b9e0 | 2007-03-16 01:18:42 -0700 | [diff] [blame] | 1269 | (u32)NSEC_PER_USEC, (u32)PSCHED_US2NS(1), |
Patrick McHardy | 514bca3 | 2007-03-16 12:34:52 -0700 | [diff] [blame] | 1270 | 1000000, |
Patrick McHardy | 3c0cfc1 | 2007-10-10 16:32:41 -0700 | [diff] [blame] | 1271 | (u32)NSEC_PER_SEC/(u32)ktime_to_ns(timespec_to_ktime(ts))); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1272 | |
| 1273 | return 0; |
| 1274 | } |
| 1275 | |
| 1276 | static int psched_open(struct inode *inode, struct file *file) |
| 1277 | { |
| 1278 | return single_open(file, psched_show, PDE(inode)->data); |
| 1279 | } |
| 1280 | |
Arjan van de Ven | da7071d | 2007-02-12 00:55:36 -0800 | [diff] [blame] | 1281 | static const struct file_operations psched_fops = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1282 | .owner = THIS_MODULE, |
| 1283 | .open = psched_open, |
| 1284 | .read = seq_read, |
| 1285 | .llseek = seq_lseek, |
| 1286 | .release = single_release, |
YOSHIFUJI Hideaki | 10297b9 | 2007-02-09 23:25:16 +0900 | [diff] [blame] | 1287 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1288 | #endif |
| 1289 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1290 | static int __init pktsched_init(void) |
| 1291 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1292 | register_qdisc(&pfifo_qdisc_ops); |
| 1293 | register_qdisc(&bfifo_qdisc_ops); |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame] | 1294 | proc_net_fops_create(&init_net, "psched", 0, &psched_fops); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1295 | |
Thomas Graf | be577dd | 2007-03-22 11:55:50 -0700 | [diff] [blame] | 1296 | rtnl_register(PF_UNSPEC, RTM_NEWQDISC, tc_modify_qdisc, NULL); |
| 1297 | rtnl_register(PF_UNSPEC, RTM_DELQDISC, tc_get_qdisc, NULL); |
| 1298 | rtnl_register(PF_UNSPEC, RTM_GETQDISC, tc_get_qdisc, tc_dump_qdisc); |
| 1299 | rtnl_register(PF_UNSPEC, RTM_NEWTCLASS, tc_ctl_tclass, NULL); |
| 1300 | rtnl_register(PF_UNSPEC, RTM_DELTCLASS, tc_ctl_tclass, NULL); |
| 1301 | rtnl_register(PF_UNSPEC, RTM_GETTCLASS, tc_ctl_tclass, tc_dump_tclass); |
| 1302 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1303 | return 0; |
| 1304 | } |
| 1305 | |
| 1306 | subsys_initcall(pktsched_init); |