Commit | Line | Data |
---|---|---|
1da177e4 LT |
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 | ||
1da177e4 | 14 | #include <linux/bitops.h> |
1da177e4 LT |
15 | #include <linux/module.h> |
16 | #include <linux/types.h> | |
17 | #include <linux/kernel.h> | |
18 | #include <linux/sched.h> | |
19 | #include <linux/string.h> | |
1da177e4 | 20 | #include <linux/errno.h> |
1da177e4 LT |
21 | #include <linux/netdevice.h> |
22 | #include <linux/skbuff.h> | |
23 | #include <linux/rtnetlink.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/rcupdate.h> | |
26 | #include <linux/list.h> | |
1da177e4 LT |
27 | #include <net/pkt_sched.h> |
28 | ||
29 | /* Main transmission queue. */ | |
30 | ||
0463d4ae | 31 | /* Modifications to data participating in scheduling must be protected with |
5fb66229 | 32 | * qdisc_lock(qdisc) spinlock. |
0463d4ae PM |
33 | * |
34 | * The idea is the following: | |
c7e4f3bb DM |
35 | * - enqueue, dequeue are serialized via qdisc root lock |
36 | * - ingress filtering is also serialized via qdisc root lock | |
0463d4ae | 37 | * - updates to tree and tree walking are only done under the rtnl mutex. |
1da177e4 | 38 | */ |
1da177e4 | 39 | |
37437bb2 | 40 | static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q) |
c716a81a | 41 | { |
6252352d | 42 | q->gso_skb = skb; |
53e91503 | 43 | q->qstats.requeues++; |
bbd8a0d3 | 44 | q->q.qlen++; /* it's still part of the queue */ |
37437bb2 | 45 | __netif_schedule(q); |
6252352d | 46 | |
c716a81a JHS |
47 | return 0; |
48 | } | |
49 | ||
d3b753db | 50 | static inline struct sk_buff *dequeue_skb(struct Qdisc *q) |
c716a81a | 51 | { |
554794de JP |
52 | struct sk_buff *skb = q->gso_skb; |
53 | ||
ebf05982 JP |
54 | if (unlikely(skb)) { |
55 | struct net_device *dev = qdisc_dev(q); | |
56 | struct netdev_queue *txq; | |
57 | ||
58 | /* check the reason of requeuing without tx lock first */ | |
59 | txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb)); | |
bbd8a0d3 KK |
60 | if (!netif_tx_queue_stopped(txq) && |
61 | !netif_tx_queue_frozen(txq)) { | |
6252352d | 62 | q->gso_skb = NULL; |
bbd8a0d3 KK |
63 | q->q.qlen--; |
64 | } else | |
ebf05982 JP |
65 | skb = NULL; |
66 | } else { | |
c716a81a | 67 | skb = q->dequeue(q); |
ebf05982 | 68 | } |
c716a81a JHS |
69 | |
70 | return skb; | |
71 | } | |
72 | ||
6c1361a6 | 73 | static inline int handle_dev_cpu_collision(struct sk_buff *skb, |
970565bb | 74 | struct netdev_queue *dev_queue, |
6c1361a6 | 75 | struct Qdisc *q) |
c716a81a | 76 | { |
6c1361a6 | 77 | int ret; |
c716a81a | 78 | |
c773e847 | 79 | if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) { |
6c1361a6 KK |
80 | /* |
81 | * Same CPU holding the lock. It may be a transient | |
82 | * configuration error, when hard_start_xmit() recurses. We | |
83 | * detect it by checking xmit owner and drop the packet when | |
84 | * deadloop is detected. Return OK to try the next skb. | |
85 | */ | |
c716a81a | 86 | kfree_skb(skb); |
6c1361a6 KK |
87 | if (net_ratelimit()) |
88 | printk(KERN_WARNING "Dead loop on netdevice %s, " | |
c773e847 | 89 | "fix it urgently!\n", dev_queue->dev->name); |
6c1361a6 KK |
90 | ret = qdisc_qlen(q); |
91 | } else { | |
92 | /* | |
93 | * Another cpu is holding lock, requeue & delay xmits for | |
94 | * some time. | |
95 | */ | |
96 | __get_cpu_var(netdev_rx_stat).cpu_collision++; | |
37437bb2 | 97 | ret = dev_requeue_skb(skb, q); |
c716a81a JHS |
98 | } |
99 | ||
6c1361a6 | 100 | return ret; |
c716a81a JHS |
101 | } |
102 | ||
10297b99 | 103 | /* |
bbd8a0d3 KK |
104 | * Transmit one skb, and handle the return status as required. Holding the |
105 | * __QDISC_STATE_RUNNING bit guarantees that only one CPU can execute this | |
106 | * function. | |
6c1361a6 KK |
107 | * |
108 | * Returns to the caller: | |
109 | * 0 - queue is empty or throttled. | |
110 | * >0 - queue is not empty. | |
6c1361a6 | 111 | */ |
bbd8a0d3 KK |
112 | int sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q, |
113 | struct net_device *dev, struct netdev_queue *txq, | |
114 | spinlock_t *root_lock) | |
1da177e4 | 115 | { |
5f1a485d | 116 | int ret = NETDEV_TX_BUSY; |
7698b4fc DM |
117 | |
118 | /* And release qdisc */ | |
119 | spin_unlock(root_lock); | |
c716a81a | 120 | |
c773e847 | 121 | HARD_TX_LOCK(dev, txq, smp_processor_id()); |
c3f26a26 DM |
122 | if (!netif_tx_queue_stopped(txq) && |
123 | !netif_tx_queue_frozen(txq)) | |
fd2ea0a7 | 124 | ret = dev_hard_start_xmit(skb, dev, txq); |
c773e847 | 125 | HARD_TX_UNLOCK(dev, txq); |
c716a81a | 126 | |
7698b4fc | 127 | spin_lock(root_lock); |
c716a81a | 128 | |
6c1361a6 KK |
129 | switch (ret) { |
130 | case NETDEV_TX_OK: | |
131 | /* Driver sent out skb successfully */ | |
132 | ret = qdisc_qlen(q); | |
133 | break; | |
134 | ||
135 | case NETDEV_TX_LOCKED: | |
136 | /* Driver try lock failed */ | |
eb6aafe3 | 137 | ret = handle_dev_cpu_collision(skb, txq, q); |
6c1361a6 KK |
138 | break; |
139 | ||
140 | default: | |
141 | /* Driver returned NETDEV_TX_BUSY - requeue skb */ | |
142 | if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit())) | |
143 | printk(KERN_WARNING "BUG %s code %d qlen %d\n", | |
144 | dev->name, ret, q->q.qlen); | |
145 | ||
37437bb2 | 146 | ret = dev_requeue_skb(skb, q); |
6c1361a6 KK |
147 | break; |
148 | } | |
c716a81a | 149 | |
c3f26a26 DM |
150 | if (ret && (netif_tx_queue_stopped(txq) || |
151 | netif_tx_queue_frozen(txq))) | |
37437bb2 DM |
152 | ret = 0; |
153 | ||
6c1361a6 | 154 | return ret; |
1da177e4 LT |
155 | } |
156 | ||
bbd8a0d3 KK |
157 | /* |
158 | * NOTE: Called under qdisc_lock(q) with locally disabled BH. | |
159 | * | |
160 | * __QDISC_STATE_RUNNING guarantees only one CPU can process | |
161 | * this qdisc at a time. qdisc_lock(q) serializes queue accesses for | |
162 | * this queue. | |
163 | * | |
164 | * netif_tx_lock serializes accesses to device driver. | |
165 | * | |
166 | * qdisc_lock(q) and netif_tx_lock are mutually exclusive, | |
167 | * if one is grabbed, another must be free. | |
168 | * | |
169 | * Note, that this procedure can be called by a watchdog timer | |
170 | * | |
171 | * Returns to the caller: | |
172 | * 0 - queue is empty or throttled. | |
173 | * >0 - queue is not empty. | |
174 | * | |
175 | */ | |
176 | static inline int qdisc_restart(struct Qdisc *q) | |
177 | { | |
178 | struct netdev_queue *txq; | |
179 | struct net_device *dev; | |
180 | spinlock_t *root_lock; | |
181 | struct sk_buff *skb; | |
182 | ||
183 | /* Dequeue packet */ | |
184 | skb = dequeue_skb(q); | |
185 | if (unlikely(!skb)) | |
186 | return 0; | |
187 | ||
188 | root_lock = qdisc_lock(q); | |
189 | dev = qdisc_dev(q); | |
190 | txq = netdev_get_tx_queue(dev, skb_get_queue_mapping(skb)); | |
191 | ||
192 | return sch_direct_xmit(skb, q, dev, txq, root_lock); | |
193 | } | |
194 | ||
37437bb2 | 195 | void __qdisc_run(struct Qdisc *q) |
48d83325 | 196 | { |
2ba2506c | 197 | unsigned long start_time = jiffies; |
2ba2506c | 198 | |
37437bb2 | 199 | while (qdisc_restart(q)) { |
2ba2506c HX |
200 | /* |
201 | * Postpone processing if | |
202 | * 1. another process needs the CPU; | |
203 | * 2. we've been doing it for too long. | |
204 | */ | |
205 | if (need_resched() || jiffies != start_time) { | |
37437bb2 | 206 | __netif_schedule(q); |
d90df3ad | 207 | break; |
2ba2506c HX |
208 | } |
209 | } | |
48d83325 | 210 | |
e2627c8c | 211 | clear_bit(__QDISC_STATE_RUNNING, &q->state); |
48d83325 HX |
212 | } |
213 | ||
9d21493b ED |
214 | unsigned long dev_trans_start(struct net_device *dev) |
215 | { | |
216 | unsigned long val, res = dev->trans_start; | |
217 | unsigned int i; | |
218 | ||
219 | for (i = 0; i < dev->num_tx_queues; i++) { | |
220 | val = netdev_get_tx_queue(dev, i)->trans_start; | |
221 | if (val && time_after(val, res)) | |
222 | res = val; | |
223 | } | |
224 | dev->trans_start = res; | |
225 | return res; | |
226 | } | |
227 | EXPORT_SYMBOL(dev_trans_start); | |
228 | ||
1da177e4 LT |
229 | static void dev_watchdog(unsigned long arg) |
230 | { | |
231 | struct net_device *dev = (struct net_device *)arg; | |
232 | ||
932ff279 | 233 | netif_tx_lock(dev); |
e8a0464c | 234 | if (!qdisc_tx_is_noop(dev)) { |
1da177e4 LT |
235 | if (netif_device_present(dev) && |
236 | netif_running(dev) && | |
237 | netif_carrier_ok(dev)) { | |
9d21493b | 238 | int some_queue_timedout = 0; |
e8a0464c | 239 | unsigned int i; |
9d21493b | 240 | unsigned long trans_start; |
e8a0464c DM |
241 | |
242 | for (i = 0; i < dev->num_tx_queues; i++) { | |
243 | struct netdev_queue *txq; | |
244 | ||
245 | txq = netdev_get_tx_queue(dev, i); | |
9d21493b ED |
246 | /* |
247 | * old device drivers set dev->trans_start | |
248 | */ | |
249 | trans_start = txq->trans_start ? : dev->trans_start; | |
250 | if (netif_tx_queue_stopped(txq) && | |
251 | time_after(jiffies, (trans_start + | |
252 | dev->watchdog_timeo))) { | |
253 | some_queue_timedout = 1; | |
e8a0464c DM |
254 | break; |
255 | } | |
256 | } | |
338f7566 | 257 | |
9d21493b | 258 | if (some_queue_timedout) { |
6579e57b | 259 | char drivername[64]; |
9d21493b ED |
260 | WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n", |
261 | dev->name, netdev_drivername(dev, drivername, 64), i); | |
d314774c | 262 | dev->netdev_ops->ndo_tx_timeout(dev); |
1da177e4 | 263 | } |
e8a0464c DM |
264 | if (!mod_timer(&dev->watchdog_timer, |
265 | round_jiffies(jiffies + | |
266 | dev->watchdog_timeo))) | |
1da177e4 LT |
267 | dev_hold(dev); |
268 | } | |
269 | } | |
932ff279 | 270 | netif_tx_unlock(dev); |
1da177e4 LT |
271 | |
272 | dev_put(dev); | |
273 | } | |
274 | ||
1da177e4 LT |
275 | void __netdev_watchdog_up(struct net_device *dev) |
276 | { | |
d314774c | 277 | if (dev->netdev_ops->ndo_tx_timeout) { |
1da177e4 LT |
278 | if (dev->watchdog_timeo <= 0) |
279 | dev->watchdog_timeo = 5*HZ; | |
60468d5b VP |
280 | if (!mod_timer(&dev->watchdog_timer, |
281 | round_jiffies(jiffies + dev->watchdog_timeo))) | |
1da177e4 LT |
282 | dev_hold(dev); |
283 | } | |
284 | } | |
285 | ||
286 | static void dev_watchdog_up(struct net_device *dev) | |
287 | { | |
1da177e4 | 288 | __netdev_watchdog_up(dev); |
1da177e4 LT |
289 | } |
290 | ||
291 | static void dev_watchdog_down(struct net_device *dev) | |
292 | { | |
932ff279 | 293 | netif_tx_lock_bh(dev); |
1da177e4 | 294 | if (del_timer(&dev->watchdog_timer)) |
15333061 | 295 | dev_put(dev); |
932ff279 | 296 | netif_tx_unlock_bh(dev); |
1da177e4 LT |
297 | } |
298 | ||
bea3348e SH |
299 | /** |
300 | * netif_carrier_on - set carrier | |
301 | * @dev: network device | |
302 | * | |
303 | * Device has detected that carrier. | |
304 | */ | |
0a242efc DV |
305 | void netif_carrier_on(struct net_device *dev) |
306 | { | |
bfaae0f0 | 307 | if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) { |
b4730016 DM |
308 | if (dev->reg_state == NETREG_UNINITIALIZED) |
309 | return; | |
0a242efc | 310 | linkwatch_fire_event(dev); |
bfaae0f0 JG |
311 | if (netif_running(dev)) |
312 | __netdev_watchdog_up(dev); | |
313 | } | |
0a242efc | 314 | } |
62e3ba1b | 315 | EXPORT_SYMBOL(netif_carrier_on); |
0a242efc | 316 | |
bea3348e SH |
317 | /** |
318 | * netif_carrier_off - clear carrier | |
319 | * @dev: network device | |
320 | * | |
321 | * Device has detected loss of carrier. | |
322 | */ | |
0a242efc DV |
323 | void netif_carrier_off(struct net_device *dev) |
324 | { | |
b4730016 DM |
325 | if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) { |
326 | if (dev->reg_state == NETREG_UNINITIALIZED) | |
327 | return; | |
0a242efc | 328 | linkwatch_fire_event(dev); |
b4730016 | 329 | } |
0a242efc | 330 | } |
62e3ba1b | 331 | EXPORT_SYMBOL(netif_carrier_off); |
0a242efc | 332 | |
1da177e4 LT |
333 | /* "NOOP" scheduler: the best scheduler, recommended for all interfaces |
334 | under all circumstances. It is difficult to invent anything faster or | |
335 | cheaper. | |
336 | */ | |
337 | ||
94df109a | 338 | static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc) |
1da177e4 LT |
339 | { |
340 | kfree_skb(skb); | |
341 | return NET_XMIT_CN; | |
342 | } | |
343 | ||
94df109a | 344 | static struct sk_buff *noop_dequeue(struct Qdisc * qdisc) |
1da177e4 LT |
345 | { |
346 | return NULL; | |
347 | } | |
348 | ||
20fea08b | 349 | struct Qdisc_ops noop_qdisc_ops __read_mostly = { |
1da177e4 LT |
350 | .id = "noop", |
351 | .priv_size = 0, | |
352 | .enqueue = noop_enqueue, | |
353 | .dequeue = noop_dequeue, | |
99c0db26 | 354 | .peek = noop_dequeue, |
1da177e4 LT |
355 | .owner = THIS_MODULE, |
356 | }; | |
357 | ||
7698b4fc | 358 | static struct netdev_queue noop_netdev_queue = { |
7698b4fc | 359 | .qdisc = &noop_qdisc, |
9f3ffae0 | 360 | .qdisc_sleeping = &noop_qdisc, |
7698b4fc DM |
361 | }; |
362 | ||
1da177e4 LT |
363 | struct Qdisc noop_qdisc = { |
364 | .enqueue = noop_enqueue, | |
365 | .dequeue = noop_dequeue, | |
366 | .flags = TCQ_F_BUILTIN, | |
10297b99 | 367 | .ops = &noop_qdisc_ops, |
1da177e4 | 368 | .list = LIST_HEAD_INIT(noop_qdisc.list), |
83874000 | 369 | .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock), |
7698b4fc | 370 | .dev_queue = &noop_netdev_queue, |
1da177e4 | 371 | }; |
62e3ba1b | 372 | EXPORT_SYMBOL(noop_qdisc); |
1da177e4 | 373 | |
20fea08b | 374 | static struct Qdisc_ops noqueue_qdisc_ops __read_mostly = { |
1da177e4 LT |
375 | .id = "noqueue", |
376 | .priv_size = 0, | |
377 | .enqueue = noop_enqueue, | |
378 | .dequeue = noop_dequeue, | |
99c0db26 | 379 | .peek = noop_dequeue, |
1da177e4 LT |
380 | .owner = THIS_MODULE, |
381 | }; | |
382 | ||
30ee42be DM |
383 | static struct Qdisc noqueue_qdisc; |
384 | static struct netdev_queue noqueue_netdev_queue = { | |
385 | .qdisc = &noqueue_qdisc, | |
9f3ffae0 | 386 | .qdisc_sleeping = &noqueue_qdisc, |
30ee42be DM |
387 | }; |
388 | ||
1da177e4 LT |
389 | static struct Qdisc noqueue_qdisc = { |
390 | .enqueue = NULL, | |
391 | .dequeue = noop_dequeue, | |
392 | .flags = TCQ_F_BUILTIN, | |
393 | .ops = &noqueue_qdisc_ops, | |
394 | .list = LIST_HEAD_INIT(noqueue_qdisc.list), | |
30ee42be DM |
395 | .q.lock = __SPIN_LOCK_UNLOCKED(noqueue_qdisc.q.lock), |
396 | .dev_queue = &noqueue_netdev_queue, | |
1da177e4 LT |
397 | }; |
398 | ||
399 | ||
d3678b46 DM |
400 | static const u8 prio2band[TC_PRIO_MAX+1] = |
401 | { 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1 }; | |
402 | ||
403 | /* 3-band FIFO queue: old style, but should be a bit faster than | |
404 | generic prio+fifo combination. | |
405 | */ | |
406 | ||
407 | #define PFIFO_FAST_BANDS 3 | |
408 | ||
fd3ae5e8 KK |
409 | /* |
410 | * Private data for a pfifo_fast scheduler containing: | |
411 | * - queues for the three band | |
412 | * - bitmap indicating which of the bands contain skbs | |
413 | */ | |
414 | struct pfifo_fast_priv { | |
415 | u32 bitmap; | |
416 | struct sk_buff_head q[PFIFO_FAST_BANDS]; | |
417 | }; | |
418 | ||
419 | /* | |
420 | * Convert a bitmap to the first band number where an skb is queued, where: | |
421 | * bitmap=0 means there are no skbs on any band. | |
422 | * bitmap=1 means there is an skb on band 0. | |
423 | * bitmap=7 means there are skbs on all 3 bands, etc. | |
424 | */ | |
425 | static const int bitmap2band[] = {-1, 0, 1, 0, 2, 0, 1, 0}; | |
426 | ||
427 | static inline struct sk_buff_head *band2list(struct pfifo_fast_priv *priv, | |
428 | int band) | |
d3678b46 | 429 | { |
fd3ae5e8 | 430 | return priv->q + band; |
d3678b46 DM |
431 | } |
432 | ||
433 | static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc) | |
321090e7 | 434 | { |
fd3ae5e8 KK |
435 | int band = prio2band[skb->priority & TC_PRIO_MAX]; |
436 | struct pfifo_fast_priv *priv = qdisc_priv(qdisc); | |
437 | struct sk_buff_head *list = band2list(priv, band); | |
1da177e4 | 438 | |
d3678b46 | 439 | if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) { |
fd3ae5e8 | 440 | priv->bitmap |= (1 << band); |
d3678b46 | 441 | qdisc->q.qlen++; |
821d24ae | 442 | return __qdisc_enqueue_tail(skb, qdisc, list); |
d3678b46 | 443 | } |
821d24ae TG |
444 | |
445 | return qdisc_drop(skb, qdisc); | |
1da177e4 LT |
446 | } |
447 | ||
d3678b46 | 448 | static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc) |
1da177e4 | 449 | { |
fd3ae5e8 KK |
450 | struct pfifo_fast_priv *priv = qdisc_priv(qdisc); |
451 | int band = bitmap2band[priv->bitmap]; | |
1da177e4 | 452 | |
fd3ae5e8 KK |
453 | if (likely(band >= 0)) { |
454 | struct sk_buff_head *list = band2list(priv, band); | |
455 | struct sk_buff *skb = __qdisc_dequeue_head(qdisc, list); | |
456 | ||
457 | qdisc->q.qlen--; | |
458 | if (skb_queue_empty(list)) | |
459 | priv->bitmap &= ~(1 << band); | |
460 | ||
461 | return skb; | |
d3678b46 | 462 | } |
f87a9c3d | 463 | |
1da177e4 LT |
464 | return NULL; |
465 | } | |
466 | ||
99c0db26 JP |
467 | static struct sk_buff *pfifo_fast_peek(struct Qdisc* qdisc) |
468 | { | |
fd3ae5e8 KK |
469 | struct pfifo_fast_priv *priv = qdisc_priv(qdisc); |
470 | int band = bitmap2band[priv->bitmap]; | |
471 | ||
472 | if (band >= 0) { | |
473 | struct sk_buff_head *list = band2list(priv, band); | |
99c0db26 | 474 | |
fd3ae5e8 | 475 | return skb_peek(list); |
99c0db26 JP |
476 | } |
477 | ||
478 | return NULL; | |
479 | } | |
480 | ||
d3678b46 | 481 | static void pfifo_fast_reset(struct Qdisc* qdisc) |
1da177e4 | 482 | { |
d3678b46 | 483 | int prio; |
fd3ae5e8 | 484 | struct pfifo_fast_priv *priv = qdisc_priv(qdisc); |
d3678b46 DM |
485 | |
486 | for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) | |
fd3ae5e8 | 487 | __qdisc_reset_queue(qdisc, band2list(priv, prio)); |
d3678b46 | 488 | |
fd3ae5e8 | 489 | priv->bitmap = 0; |
821d24ae | 490 | qdisc->qstats.backlog = 0; |
d3678b46 | 491 | qdisc->q.qlen = 0; |
1da177e4 LT |
492 | } |
493 | ||
d3678b46 DM |
494 | static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb) |
495 | { | |
496 | struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS }; | |
497 | ||
498 | memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1); | |
499 | NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt); | |
500 | return skb->len; | |
501 | ||
502 | nla_put_failure: | |
503 | return -1; | |
504 | } | |
505 | ||
506 | static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt) | |
507 | { | |
508 | int prio; | |
fd3ae5e8 | 509 | struct pfifo_fast_priv *priv = qdisc_priv(qdisc); |
d3678b46 DM |
510 | |
511 | for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) | |
fd3ae5e8 | 512 | skb_queue_head_init(band2list(priv, prio)); |
d3678b46 DM |
513 | |
514 | return 0; | |
515 | } | |
516 | ||
517 | static struct Qdisc_ops pfifo_fast_ops __read_mostly = { | |
518 | .id = "pfifo_fast", | |
fd3ae5e8 | 519 | .priv_size = sizeof(struct pfifo_fast_priv), |
d3678b46 DM |
520 | .enqueue = pfifo_fast_enqueue, |
521 | .dequeue = pfifo_fast_dequeue, | |
99c0db26 | 522 | .peek = pfifo_fast_peek, |
d3678b46 DM |
523 | .init = pfifo_fast_init, |
524 | .reset = pfifo_fast_reset, | |
525 | .dump = pfifo_fast_dump, | |
1da177e4 LT |
526 | .owner = THIS_MODULE, |
527 | }; | |
528 | ||
5ce2d488 | 529 | struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue, |
bb949fbd | 530 | struct Qdisc_ops *ops) |
1da177e4 LT |
531 | { |
532 | void *p; | |
533 | struct Qdisc *sch; | |
3d54b82f TG |
534 | unsigned int size; |
535 | int err = -ENOBUFS; | |
1da177e4 LT |
536 | |
537 | /* ensure that the Qdisc and the private data are 32-byte aligned */ | |
3d54b82f TG |
538 | size = QDISC_ALIGN(sizeof(*sch)); |
539 | size += ops->priv_size + (QDISC_ALIGNTO - 1); | |
1da177e4 | 540 | |
0da974f4 | 541 | p = kzalloc(size, GFP_KERNEL); |
1da177e4 | 542 | if (!p) |
3d54b82f | 543 | goto errout; |
3d54b82f TG |
544 | sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p); |
545 | sch->padded = (char *) sch - (char *) p; | |
1da177e4 LT |
546 | |
547 | INIT_LIST_HEAD(&sch->list); | |
548 | skb_queue_head_init(&sch->q); | |
549 | sch->ops = ops; | |
550 | sch->enqueue = ops->enqueue; | |
551 | sch->dequeue = ops->dequeue; | |
bb949fbd | 552 | sch->dev_queue = dev_queue; |
5ce2d488 | 553 | dev_hold(qdisc_dev(sch)); |
1da177e4 | 554 | atomic_set(&sch->refcnt, 1); |
3d54b82f TG |
555 | |
556 | return sch; | |
557 | errout: | |
01e123d7 | 558 | return ERR_PTR(err); |
3d54b82f TG |
559 | } |
560 | ||
bb949fbd DM |
561 | struct Qdisc * qdisc_create_dflt(struct net_device *dev, |
562 | struct netdev_queue *dev_queue, | |
563 | struct Qdisc_ops *ops, | |
9f9afec4 | 564 | unsigned int parentid) |
3d54b82f TG |
565 | { |
566 | struct Qdisc *sch; | |
10297b99 | 567 | |
5ce2d488 | 568 | sch = qdisc_alloc(dev_queue, ops); |
3d54b82f TG |
569 | if (IS_ERR(sch)) |
570 | goto errout; | |
9f9afec4 | 571 | sch->parent = parentid; |
3d54b82f | 572 | |
1da177e4 LT |
573 | if (!ops->init || ops->init(sch, NULL) == 0) |
574 | return sch; | |
575 | ||
0fbbeb1b | 576 | qdisc_destroy(sch); |
3d54b82f | 577 | errout: |
1da177e4 LT |
578 | return NULL; |
579 | } | |
62e3ba1b | 580 | EXPORT_SYMBOL(qdisc_create_dflt); |
1da177e4 | 581 | |
5fb66229 | 582 | /* Under qdisc_lock(qdisc) and BH! */ |
1da177e4 LT |
583 | |
584 | void qdisc_reset(struct Qdisc *qdisc) | |
585 | { | |
20fea08b | 586 | const struct Qdisc_ops *ops = qdisc->ops; |
1da177e4 LT |
587 | |
588 | if (ops->reset) | |
589 | ops->reset(qdisc); | |
67305ebc | 590 | |
bbd8a0d3 KK |
591 | if (qdisc->gso_skb) { |
592 | kfree_skb(qdisc->gso_skb); | |
593 | qdisc->gso_skb = NULL; | |
594 | qdisc->q.qlen = 0; | |
595 | } | |
1da177e4 | 596 | } |
62e3ba1b | 597 | EXPORT_SYMBOL(qdisc_reset); |
1da177e4 | 598 | |
1e0d5a57 | 599 | void qdisc_destroy(struct Qdisc *qdisc) |
1da177e4 | 600 | { |
8a34c5dc DM |
601 | const struct Qdisc_ops *ops = qdisc->ops; |
602 | ||
1e0d5a57 DM |
603 | if (qdisc->flags & TCQ_F_BUILTIN || |
604 | !atomic_dec_and_test(&qdisc->refcnt)) | |
605 | return; | |
606 | ||
3a682fbd | 607 | #ifdef CONFIG_NET_SCHED |
f6e0b239 JP |
608 | qdisc_list_del(qdisc); |
609 | ||
175f9c1b | 610 | qdisc_put_stab(qdisc->stab); |
3a682fbd | 611 | #endif |
8a34c5dc DM |
612 | gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est); |
613 | if (ops->reset) | |
614 | ops->reset(qdisc); | |
615 | if (ops->destroy) | |
616 | ops->destroy(qdisc); | |
617 | ||
618 | module_put(ops->owner); | |
619 | dev_put(qdisc_dev(qdisc)); | |
620 | ||
554794de | 621 | kfree_skb(qdisc->gso_skb); |
1da177e4 LT |
622 | kfree((char *) qdisc - qdisc->padded); |
623 | } | |
62e3ba1b | 624 | EXPORT_SYMBOL(qdisc_destroy); |
1da177e4 | 625 | |
e8a0464c DM |
626 | static bool dev_all_qdisc_sleeping_noop(struct net_device *dev) |
627 | { | |
628 | unsigned int i; | |
629 | ||
630 | for (i = 0; i < dev->num_tx_queues; i++) { | |
631 | struct netdev_queue *txq = netdev_get_tx_queue(dev, i); | |
632 | ||
633 | if (txq->qdisc_sleeping != &noop_qdisc) | |
634 | return false; | |
635 | } | |
636 | return true; | |
637 | } | |
638 | ||
639 | static void attach_one_default_qdisc(struct net_device *dev, | |
640 | struct netdev_queue *dev_queue, | |
641 | void *_unused) | |
642 | { | |
643 | struct Qdisc *qdisc; | |
644 | ||
645 | if (dev->tx_queue_len) { | |
646 | qdisc = qdisc_create_dflt(dev, dev_queue, | |
d3678b46 | 647 | &pfifo_fast_ops, TC_H_ROOT); |
e8a0464c DM |
648 | if (!qdisc) { |
649 | printk(KERN_INFO "%s: activation failed\n", dev->name); | |
650 | return; | |
651 | } | |
bbd8a0d3 KK |
652 | |
653 | /* Can by-pass the queue discipline for default qdisc */ | |
654 | qdisc->flags |= TCQ_F_CAN_BYPASS; | |
e8a0464c DM |
655 | } else { |
656 | qdisc = &noqueue_qdisc; | |
657 | } | |
658 | dev_queue->qdisc_sleeping = qdisc; | |
659 | } | |
660 | ||
661 | static void transition_one_qdisc(struct net_device *dev, | |
662 | struct netdev_queue *dev_queue, | |
663 | void *_need_watchdog) | |
664 | { | |
83874000 | 665 | struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping; |
e8a0464c DM |
666 | int *need_watchdog_p = _need_watchdog; |
667 | ||
a9312ae8 DM |
668 | if (!(new_qdisc->flags & TCQ_F_BUILTIN)) |
669 | clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state); | |
670 | ||
83874000 | 671 | rcu_assign_pointer(dev_queue->qdisc, new_qdisc); |
9d21493b ED |
672 | if (need_watchdog_p && new_qdisc != &noqueue_qdisc) { |
673 | dev_queue->trans_start = 0; | |
e8a0464c | 674 | *need_watchdog_p = 1; |
9d21493b | 675 | } |
e8a0464c DM |
676 | } |
677 | ||
1da177e4 LT |
678 | void dev_activate(struct net_device *dev) |
679 | { | |
e8a0464c | 680 | int need_watchdog; |
b0e1e646 | 681 | |
1da177e4 | 682 | /* No queueing discipline is attached to device; |
d3678b46 DM |
683 | create default one i.e. pfifo_fast for devices, |
684 | which need queueing and noqueue_qdisc for | |
685 | virtual interfaces | |
1da177e4 LT |
686 | */ |
687 | ||
e8a0464c DM |
688 | if (dev_all_qdisc_sleeping_noop(dev)) |
689 | netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL); | |
1da177e4 | 690 | |
cacaddf5 TC |
691 | if (!netif_carrier_ok(dev)) |
692 | /* Delay activation until next carrier-on event */ | |
693 | return; | |
694 | ||
e8a0464c DM |
695 | need_watchdog = 0; |
696 | netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog); | |
8d50b53d | 697 | transition_one_qdisc(dev, &dev->rx_queue, NULL); |
e8a0464c DM |
698 | |
699 | if (need_watchdog) { | |
1da177e4 LT |
700 | dev->trans_start = jiffies; |
701 | dev_watchdog_up(dev); | |
702 | } | |
b0e1e646 DM |
703 | } |
704 | ||
e8a0464c DM |
705 | static void dev_deactivate_queue(struct net_device *dev, |
706 | struct netdev_queue *dev_queue, | |
707 | void *_qdisc_default) | |
b0e1e646 | 708 | { |
e8a0464c | 709 | struct Qdisc *qdisc_default = _qdisc_default; |
970565bb | 710 | struct Qdisc *qdisc; |
970565bb | 711 | |
970565bb | 712 | qdisc = dev_queue->qdisc; |
b0e1e646 | 713 | if (qdisc) { |
83874000 DM |
714 | spin_lock_bh(qdisc_lock(qdisc)); |
715 | ||
a9312ae8 DM |
716 | if (!(qdisc->flags & TCQ_F_BUILTIN)) |
717 | set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state); | |
718 | ||
f7a54c13 | 719 | rcu_assign_pointer(dev_queue->qdisc, qdisc_default); |
b0e1e646 | 720 | qdisc_reset(qdisc); |
d3b753db | 721 | |
83874000 | 722 | spin_unlock_bh(qdisc_lock(qdisc)); |
b0e1e646 | 723 | } |
1da177e4 LT |
724 | } |
725 | ||
4335cd2d | 726 | static bool some_qdisc_is_busy(struct net_device *dev) |
e8a0464c DM |
727 | { |
728 | unsigned int i; | |
729 | ||
730 | for (i = 0; i < dev->num_tx_queues; i++) { | |
731 | struct netdev_queue *dev_queue; | |
7698b4fc | 732 | spinlock_t *root_lock; |
e2627c8c | 733 | struct Qdisc *q; |
e8a0464c DM |
734 | int val; |
735 | ||
736 | dev_queue = netdev_get_tx_queue(dev, i); | |
b9a3b110 | 737 | q = dev_queue->qdisc_sleeping; |
5fb66229 | 738 | root_lock = qdisc_lock(q); |
e8a0464c | 739 | |
4335cd2d | 740 | spin_lock_bh(root_lock); |
e8a0464c | 741 | |
b9a3b110 DM |
742 | val = (test_bit(__QDISC_STATE_RUNNING, &q->state) || |
743 | test_bit(__QDISC_STATE_SCHED, &q->state)); | |
e8a0464c | 744 | |
4335cd2d | 745 | spin_unlock_bh(root_lock); |
e8a0464c DM |
746 | |
747 | if (val) | |
748 | return true; | |
749 | } | |
750 | return false; | |
751 | } | |
752 | ||
1da177e4 LT |
753 | void dev_deactivate(struct net_device *dev) |
754 | { | |
e8a0464c | 755 | netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc); |
8d50b53d | 756 | dev_deactivate_queue(dev, &dev->rx_queue, &noop_qdisc); |
41a23b07 | 757 | |
1da177e4 LT |
758 | dev_watchdog_down(dev); |
759 | ||
ce0e32e6 | 760 | /* Wait for outstanding qdisc-less dev_queue_xmit calls. */ |
d4828d85 | 761 | synchronize_rcu(); |
1da177e4 | 762 | |
d4828d85 | 763 | /* Wait for outstanding qdisc_run calls. */ |
4335cd2d DM |
764 | while (some_qdisc_is_busy(dev)) |
765 | yield(); | |
1da177e4 LT |
766 | } |
767 | ||
b0e1e646 DM |
768 | static void dev_init_scheduler_queue(struct net_device *dev, |
769 | struct netdev_queue *dev_queue, | |
e8a0464c | 770 | void *_qdisc) |
b0e1e646 | 771 | { |
e8a0464c DM |
772 | struct Qdisc *qdisc = _qdisc; |
773 | ||
b0e1e646 DM |
774 | dev_queue->qdisc = qdisc; |
775 | dev_queue->qdisc_sleeping = qdisc; | |
b0e1e646 DM |
776 | } |
777 | ||
1da177e4 LT |
778 | void dev_init_scheduler(struct net_device *dev) |
779 | { | |
e8a0464c | 780 | netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc); |
8d50b53d | 781 | dev_init_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc); |
1da177e4 | 782 | |
b24b8a24 | 783 | setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev); |
1da177e4 LT |
784 | } |
785 | ||
e8a0464c DM |
786 | static void shutdown_scheduler_queue(struct net_device *dev, |
787 | struct netdev_queue *dev_queue, | |
788 | void *_qdisc_default) | |
1da177e4 | 789 | { |
b0e1e646 | 790 | struct Qdisc *qdisc = dev_queue->qdisc_sleeping; |
e8a0464c | 791 | struct Qdisc *qdisc_default = _qdisc_default; |
b0e1e646 DM |
792 | |
793 | if (qdisc) { | |
f7a54c13 | 794 | rcu_assign_pointer(dev_queue->qdisc, qdisc_default); |
b0e1e646 | 795 | dev_queue->qdisc_sleeping = qdisc_default; |
1da177e4 | 796 | |
1da177e4 | 797 | qdisc_destroy(qdisc); |
10297b99 | 798 | } |
b0e1e646 DM |
799 | } |
800 | ||
801 | void dev_shutdown(struct net_device *dev) | |
802 | { | |
e8a0464c | 803 | netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc); |
8d50b53d | 804 | shutdown_scheduler_queue(dev, &dev->rx_queue, &noop_qdisc); |
547b792c | 805 | WARN_ON(timer_pending(&dev->watchdog_timer)); |
1da177e4 | 806 | } |