xref: /linux/net/sched/sch_generic.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
1 /*
2  * net/sched/sch_generic.c	Generic packet scheduler routines.
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  *              Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11  *              - Ingress support
12  */
13 
14 #include <asm/uaccess.h>
15 #include <asm/system.h>
16 #include <linux/bitops.h>
17 #include <linux/config.h>
18 #include <linux/module.h>
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/string.h>
23 #include <linux/mm.h>
24 #include <linux/socket.h>
25 #include <linux/sockios.h>
26 #include <linux/in.h>
27 #include <linux/errno.h>
28 #include <linux/interrupt.h>
29 #include <linux/netdevice.h>
30 #include <linux/skbuff.h>
31 #include <linux/rtnetlink.h>
32 #include <linux/init.h>
33 #include <linux/rcupdate.h>
34 #include <linux/list.h>
35 #include <net/sock.h>
36 #include <net/pkt_sched.h>
37 
38 /* Main transmission queue. */
39 
40 /* Main qdisc structure lock.
41 
42    However, modifications
43    to data, participating in scheduling must be additionally
44    protected with dev->queue_lock spinlock.
45 
46    The idea is the following:
47    - enqueue, dequeue are serialized via top level device
48      spinlock dev->queue_lock.
49    - tree walking is protected by read_lock_bh(qdisc_tree_lock)
50      and this lock is used only in process context.
51    - updates to tree are made under rtnl semaphore or
52      from softirq context (__qdisc_destroy rcu-callback)
53      hence this lock needs local bh disabling.
54 
55    qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
56  */
57 DEFINE_RWLOCK(qdisc_tree_lock);
58 
59 void qdisc_lock_tree(struct net_device *dev)
60 {
61 	write_lock_bh(&qdisc_tree_lock);
62 	spin_lock_bh(&dev->queue_lock);
63 }
64 
65 void qdisc_unlock_tree(struct net_device *dev)
66 {
67 	spin_unlock_bh(&dev->queue_lock);
68 	write_unlock_bh(&qdisc_tree_lock);
69 }
70 
71 /*
72    dev->queue_lock serializes queue accesses for this device
73    AND dev->qdisc pointer itself.
74 
75    netif_tx_lock serializes accesses to device driver.
76 
77    dev->queue_lock and netif_tx_lock are mutually exclusive,
78    if one is grabbed, another must be free.
79  */
80 
81 
82 /* Kick device.
83    Note, that this procedure can be called by a watchdog timer, so that
84    we do not check dev->tbusy flag here.
85 
86    Returns:  0  - queue is empty.
87             >0  - queue is not empty, but throttled.
88 	    <0  - queue is not empty. Device is throttled, if dev->tbusy != 0.
89 
90    NOTE: Called under dev->queue_lock with locally disabled BH.
91 */
92 
93 static inline int qdisc_restart(struct net_device *dev)
94 {
95 	struct Qdisc *q = dev->qdisc;
96 	struct sk_buff *skb;
97 
98 	/* Dequeue packet */
99 	if ((skb = q->dequeue(q)) != NULL) {
100 		unsigned nolock = (dev->features & NETIF_F_LLTX);
101 		/*
102 		 * When the driver has LLTX set it does its own locking
103 		 * in start_xmit. No need to add additional overhead by
104 		 * locking again. These checks are worth it because
105 		 * even uncongested locks can be quite expensive.
106 		 * The driver can do trylock like here too, in case
107 		 * of lock congestion it should return -1 and the packet
108 		 * will be requeued.
109 		 */
110 		if (!nolock) {
111 			if (!netif_tx_trylock(dev)) {
112 			collision:
113 				/* So, someone grabbed the driver. */
114 
115 				/* It may be transient configuration error,
116 				   when hard_start_xmit() recurses. We detect
117 				   it by checking xmit owner and drop the
118 				   packet when deadloop is detected.
119 				*/
120 				if (dev->xmit_lock_owner == smp_processor_id()) {
121 					kfree_skb(skb);
122 					if (net_ratelimit())
123 						printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
124 					return -1;
125 				}
126 				__get_cpu_var(netdev_rx_stat).cpu_collision++;
127 				goto requeue;
128 			}
129 		}
130 
131 		{
132 			/* And release queue */
133 			spin_unlock(&dev->queue_lock);
134 
135 			if (!netif_queue_stopped(dev)) {
136 				int ret;
137 				if (netdev_nit)
138 					dev_queue_xmit_nit(skb, dev);
139 
140 				ret = dev->hard_start_xmit(skb, dev);
141 				if (ret == NETDEV_TX_OK) {
142 					if (!nolock) {
143 						netif_tx_unlock(dev);
144 					}
145 					spin_lock(&dev->queue_lock);
146 					return -1;
147 				}
148 				if (ret == NETDEV_TX_LOCKED && nolock) {
149 					spin_lock(&dev->queue_lock);
150 					goto collision;
151 				}
152 			}
153 
154 			/* NETDEV_TX_BUSY - we need to requeue */
155 			/* Release the driver */
156 			if (!nolock) {
157 				netif_tx_unlock(dev);
158 			}
159 			spin_lock(&dev->queue_lock);
160 			q = dev->qdisc;
161 		}
162 
163 		/* Device kicked us out :(
164 		   This is possible in three cases:
165 
166 		   0. driver is locked
167 		   1. fastroute is enabled
168 		   2. device cannot determine busy state
169 		      before start of transmission (f.e. dialout)
170 		   3. device is buggy (ppp)
171 		 */
172 
173 requeue:
174 		q->ops->requeue(skb, q);
175 		netif_schedule(dev);
176 		return 1;
177 	}
178 	BUG_ON((int) q->q.qlen < 0);
179 	return q->q.qlen;
180 }
181 
182 void __qdisc_run(struct net_device *dev)
183 {
184 	while (qdisc_restart(dev) < 0 && !netif_queue_stopped(dev))
185 		/* NOTHING */;
186 
187 	clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
188 }
189 
190 static void dev_watchdog(unsigned long arg)
191 {
192 	struct net_device *dev = (struct net_device *)arg;
193 
194 	netif_tx_lock(dev);
195 	if (dev->qdisc != &noop_qdisc) {
196 		if (netif_device_present(dev) &&
197 		    netif_running(dev) &&
198 		    netif_carrier_ok(dev)) {
199 			if (netif_queue_stopped(dev) &&
200 			    time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
201 
202 				printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
203 				       dev->name);
204 				dev->tx_timeout(dev);
205 			}
206 			if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
207 				dev_hold(dev);
208 		}
209 	}
210 	netif_tx_unlock(dev);
211 
212 	dev_put(dev);
213 }
214 
215 static void dev_watchdog_init(struct net_device *dev)
216 {
217 	init_timer(&dev->watchdog_timer);
218 	dev->watchdog_timer.data = (unsigned long)dev;
219 	dev->watchdog_timer.function = dev_watchdog;
220 }
221 
222 void __netdev_watchdog_up(struct net_device *dev)
223 {
224 	if (dev->tx_timeout) {
225 		if (dev->watchdog_timeo <= 0)
226 			dev->watchdog_timeo = 5*HZ;
227 		if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
228 			dev_hold(dev);
229 	}
230 }
231 
232 static void dev_watchdog_up(struct net_device *dev)
233 {
234 	netif_tx_lock_bh(dev);
235 	__netdev_watchdog_up(dev);
236 	netif_tx_unlock_bh(dev);
237 }
238 
239 static void dev_watchdog_down(struct net_device *dev)
240 {
241 	netif_tx_lock_bh(dev);
242 	if (del_timer(&dev->watchdog_timer))
243 		dev_put(dev);
244 	netif_tx_unlock_bh(dev);
245 }
246 
247 void netif_carrier_on(struct net_device *dev)
248 {
249 	if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
250 		linkwatch_fire_event(dev);
251 	if (netif_running(dev))
252 		__netdev_watchdog_up(dev);
253 }
254 
255 void netif_carrier_off(struct net_device *dev)
256 {
257 	if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
258 		linkwatch_fire_event(dev);
259 }
260 
261 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
262    under all circumstances. It is difficult to invent anything faster or
263    cheaper.
264  */
265 
266 static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
267 {
268 	kfree_skb(skb);
269 	return NET_XMIT_CN;
270 }
271 
272 static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
273 {
274 	return NULL;
275 }
276 
277 static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
278 {
279 	if (net_ratelimit())
280 		printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
281 		       skb->dev->name);
282 	kfree_skb(skb);
283 	return NET_XMIT_CN;
284 }
285 
286 struct Qdisc_ops noop_qdisc_ops = {
287 	.id		=	"noop",
288 	.priv_size	=	0,
289 	.enqueue	=	noop_enqueue,
290 	.dequeue	=	noop_dequeue,
291 	.requeue	=	noop_requeue,
292 	.owner		=	THIS_MODULE,
293 };
294 
295 struct Qdisc noop_qdisc = {
296 	.enqueue	=	noop_enqueue,
297 	.dequeue	=	noop_dequeue,
298 	.flags		=	TCQ_F_BUILTIN,
299 	.ops		=	&noop_qdisc_ops,
300 	.list		=	LIST_HEAD_INIT(noop_qdisc.list),
301 };
302 
303 static struct Qdisc_ops noqueue_qdisc_ops = {
304 	.id		=	"noqueue",
305 	.priv_size	=	0,
306 	.enqueue	=	noop_enqueue,
307 	.dequeue	=	noop_dequeue,
308 	.requeue	=	noop_requeue,
309 	.owner		=	THIS_MODULE,
310 };
311 
312 static struct Qdisc noqueue_qdisc = {
313 	.enqueue	=	NULL,
314 	.dequeue	=	noop_dequeue,
315 	.flags		=	TCQ_F_BUILTIN,
316 	.ops		=	&noqueue_qdisc_ops,
317 	.list		=	LIST_HEAD_INIT(noqueue_qdisc.list),
318 };
319 
320 
321 static const u8 prio2band[TC_PRIO_MAX+1] =
322 	{ 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 };
323 
324 /* 3-band FIFO queue: old style, but should be a bit faster than
325    generic prio+fifo combination.
326  */
327 
328 #define PFIFO_FAST_BANDS 3
329 
330 static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
331 					     struct Qdisc *qdisc)
332 {
333 	struct sk_buff_head *list = qdisc_priv(qdisc);
334 	return list + prio2band[skb->priority & TC_PRIO_MAX];
335 }
336 
337 static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
338 {
339 	struct sk_buff_head *list = prio2list(skb, qdisc);
340 
341 	if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
342 		qdisc->q.qlen++;
343 		return __qdisc_enqueue_tail(skb, qdisc, list);
344 	}
345 
346 	return qdisc_drop(skb, qdisc);
347 }
348 
349 static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
350 {
351 	int prio;
352 	struct sk_buff_head *list = qdisc_priv(qdisc);
353 
354 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
355 		if (!skb_queue_empty(list + prio)) {
356 			qdisc->q.qlen--;
357 			return __qdisc_dequeue_head(qdisc, list + prio);
358 		}
359 	}
360 
361 	return NULL;
362 }
363 
364 static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
365 {
366 	qdisc->q.qlen++;
367 	return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
368 }
369 
370 static void pfifo_fast_reset(struct Qdisc* qdisc)
371 {
372 	int prio;
373 	struct sk_buff_head *list = qdisc_priv(qdisc);
374 
375 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
376 		__qdisc_reset_queue(qdisc, list + prio);
377 
378 	qdisc->qstats.backlog = 0;
379 	qdisc->q.qlen = 0;
380 }
381 
382 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
383 {
384 	struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
385 
386 	memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
387 	RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
388 	return skb->len;
389 
390 rtattr_failure:
391 	return -1;
392 }
393 
394 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
395 {
396 	int prio;
397 	struct sk_buff_head *list = qdisc_priv(qdisc);
398 
399 	for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
400 		skb_queue_head_init(list + prio);
401 
402 	return 0;
403 }
404 
405 static struct Qdisc_ops pfifo_fast_ops = {
406 	.id		=	"pfifo_fast",
407 	.priv_size	=	PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
408 	.enqueue	=	pfifo_fast_enqueue,
409 	.dequeue	=	pfifo_fast_dequeue,
410 	.requeue	=	pfifo_fast_requeue,
411 	.init		=	pfifo_fast_init,
412 	.reset		=	pfifo_fast_reset,
413 	.dump		=	pfifo_fast_dump,
414 	.owner		=	THIS_MODULE,
415 };
416 
417 struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
418 {
419 	void *p;
420 	struct Qdisc *sch;
421 	unsigned int size;
422 	int err = -ENOBUFS;
423 
424 	/* ensure that the Qdisc and the private data are 32-byte aligned */
425 	size = QDISC_ALIGN(sizeof(*sch));
426 	size += ops->priv_size + (QDISC_ALIGNTO - 1);
427 
428 	p = kmalloc(size, GFP_KERNEL);
429 	if (!p)
430 		goto errout;
431 	memset(p, 0, size);
432 	sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
433 	sch->padded = (char *) sch - (char *) p;
434 
435 	INIT_LIST_HEAD(&sch->list);
436 	skb_queue_head_init(&sch->q);
437 	sch->ops = ops;
438 	sch->enqueue = ops->enqueue;
439 	sch->dequeue = ops->dequeue;
440 	sch->dev = dev;
441 	dev_hold(dev);
442 	sch->stats_lock = &dev->queue_lock;
443 	atomic_set(&sch->refcnt, 1);
444 
445 	return sch;
446 errout:
447 	return ERR_PTR(-err);
448 }
449 
450 struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
451 {
452 	struct Qdisc *sch;
453 
454 	sch = qdisc_alloc(dev, ops);
455 	if (IS_ERR(sch))
456 		goto errout;
457 
458 	if (!ops->init || ops->init(sch, NULL) == 0)
459 		return sch;
460 
461 	qdisc_destroy(sch);
462 errout:
463 	return NULL;
464 }
465 
466 /* Under dev->queue_lock and BH! */
467 
468 void qdisc_reset(struct Qdisc *qdisc)
469 {
470 	struct Qdisc_ops *ops = qdisc->ops;
471 
472 	if (ops->reset)
473 		ops->reset(qdisc);
474 }
475 
476 /* this is the rcu callback function to clean up a qdisc when there
477  * are no further references to it */
478 
479 static void __qdisc_destroy(struct rcu_head *head)
480 {
481 	struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
482 	struct Qdisc_ops  *ops = qdisc->ops;
483 
484 #ifdef CONFIG_NET_ESTIMATOR
485 	gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
486 #endif
487 	write_lock(&qdisc_tree_lock);
488 	if (ops->reset)
489 		ops->reset(qdisc);
490 	if (ops->destroy)
491 		ops->destroy(qdisc);
492 	write_unlock(&qdisc_tree_lock);
493 	module_put(ops->owner);
494 
495 	dev_put(qdisc->dev);
496 	kfree((char *) qdisc - qdisc->padded);
497 }
498 
499 /* Under dev->queue_lock and BH! */
500 
501 void qdisc_destroy(struct Qdisc *qdisc)
502 {
503 	struct list_head cql = LIST_HEAD_INIT(cql);
504 	struct Qdisc *cq, *q, *n;
505 
506 	if (qdisc->flags & TCQ_F_BUILTIN ||
507 		!atomic_dec_and_test(&qdisc->refcnt))
508 		return;
509 
510 	if (!list_empty(&qdisc->list)) {
511 		if (qdisc->ops->cl_ops == NULL)
512 			list_del(&qdisc->list);
513 		else
514 			list_move(&qdisc->list, &cql);
515 	}
516 
517 	/* unlink inner qdiscs from dev->qdisc_list immediately */
518 	list_for_each_entry(cq, &cql, list)
519 		list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
520 			if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
521 				if (q->ops->cl_ops == NULL)
522 					list_del_init(&q->list);
523 				else
524 					list_move_tail(&q->list, &cql);
525 			}
526 	list_for_each_entry_safe(cq, n, &cql, list)
527 		list_del_init(&cq->list);
528 
529 	call_rcu(&qdisc->q_rcu, __qdisc_destroy);
530 }
531 
532 void dev_activate(struct net_device *dev)
533 {
534 	/* No queueing discipline is attached to device;
535 	   create default one i.e. pfifo_fast for devices,
536 	   which need queueing and noqueue_qdisc for
537 	   virtual interfaces
538 	 */
539 
540 	if (dev->qdisc_sleeping == &noop_qdisc) {
541 		struct Qdisc *qdisc;
542 		if (dev->tx_queue_len) {
543 			qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
544 			if (qdisc == NULL) {
545 				printk(KERN_INFO "%s: activation failed\n", dev->name);
546 				return;
547 			}
548 			write_lock_bh(&qdisc_tree_lock);
549 			list_add_tail(&qdisc->list, &dev->qdisc_list);
550 			write_unlock_bh(&qdisc_tree_lock);
551 		} else {
552 			qdisc =  &noqueue_qdisc;
553 		}
554 		write_lock_bh(&qdisc_tree_lock);
555 		dev->qdisc_sleeping = qdisc;
556 		write_unlock_bh(&qdisc_tree_lock);
557 	}
558 
559 	if (!netif_carrier_ok(dev))
560 		/* Delay activation until next carrier-on event */
561 		return;
562 
563 	spin_lock_bh(&dev->queue_lock);
564 	rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
565 	if (dev->qdisc != &noqueue_qdisc) {
566 		dev->trans_start = jiffies;
567 		dev_watchdog_up(dev);
568 	}
569 	spin_unlock_bh(&dev->queue_lock);
570 }
571 
572 void dev_deactivate(struct net_device *dev)
573 {
574 	struct Qdisc *qdisc;
575 
576 	spin_lock_bh(&dev->queue_lock);
577 	qdisc = dev->qdisc;
578 	dev->qdisc = &noop_qdisc;
579 
580 	qdisc_reset(qdisc);
581 
582 	spin_unlock_bh(&dev->queue_lock);
583 
584 	dev_watchdog_down(dev);
585 
586 	while (test_bit(__LINK_STATE_SCHED, &dev->state))
587 		yield();
588 
589 	spin_unlock_wait(&dev->_xmit_lock);
590 }
591 
592 void dev_init_scheduler(struct net_device *dev)
593 {
594 	qdisc_lock_tree(dev);
595 	dev->qdisc = &noop_qdisc;
596 	dev->qdisc_sleeping = &noop_qdisc;
597 	INIT_LIST_HEAD(&dev->qdisc_list);
598 	qdisc_unlock_tree(dev);
599 
600 	dev_watchdog_init(dev);
601 }
602 
603 void dev_shutdown(struct net_device *dev)
604 {
605 	struct Qdisc *qdisc;
606 
607 	qdisc_lock_tree(dev);
608 	qdisc = dev->qdisc_sleeping;
609 	dev->qdisc = &noop_qdisc;
610 	dev->qdisc_sleeping = &noop_qdisc;
611 	qdisc_destroy(qdisc);
612 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
613         if ((qdisc = dev->qdisc_ingress) != NULL) {
614 		dev->qdisc_ingress = NULL;
615 		qdisc_destroy(qdisc);
616         }
617 #endif
618 	BUG_TRAP(!timer_pending(&dev->watchdog_timer));
619 	qdisc_unlock_tree(dev);
620 }
621 
622 EXPORT_SYMBOL(__netdev_watchdog_up);
623 EXPORT_SYMBOL(netif_carrier_on);
624 EXPORT_SYMBOL(netif_carrier_off);
625 EXPORT_SYMBOL(noop_qdisc);
626 EXPORT_SYMBOL(noop_qdisc_ops);
627 EXPORT_SYMBOL(qdisc_create_dflt);
628 EXPORT_SYMBOL(qdisc_alloc);
629 EXPORT_SYMBOL(qdisc_destroy);
630 EXPORT_SYMBOL(qdisc_reset);
631 EXPORT_SYMBOL(qdisc_lock_tree);
632 EXPORT_SYMBOL(qdisc_unlock_tree);
633