]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * net/sched/sch_sfq.c Stochastic Fairness Queueing discipline. | |
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 | ||
1da177e4 | 12 | #include <linux/module.h> |
1da177e4 LT |
13 | #include <linux/types.h> |
14 | #include <linux/kernel.h> | |
15 | #include <linux/jiffies.h> | |
16 | #include <linux/string.h> | |
1da177e4 LT |
17 | #include <linux/in.h> |
18 | #include <linux/errno.h> | |
1da177e4 | 19 | #include <linux/init.h> |
1da177e4 | 20 | #include <linux/skbuff.h> |
32740ddc | 21 | #include <linux/jhash.h> |
5a0e3ad6 | 22 | #include <linux/slab.h> |
817fb15d | 23 | #include <linux/vmalloc.h> |
0ba48053 | 24 | #include <net/netlink.h> |
1da177e4 | 25 | #include <net/pkt_sched.h> |
cf1facda | 26 | #include <net/pkt_cls.h> |
ddecf0f4 | 27 | #include <net/red.h> |
1da177e4 LT |
28 | |
29 | ||
30 | /* Stochastic Fairness Queuing algorithm. | |
31 | ======================================= | |
32 | ||
33 | Source: | |
34 | Paul E. McKenney "Stochastic Fairness Queuing", | |
35 | IEEE INFOCOMM'90 Proceedings, San Francisco, 1990. | |
36 | ||
37 | Paul E. McKenney "Stochastic Fairness Queuing", | |
38 | "Interworking: Research and Experience", v.2, 1991, p.113-131. | |
39 | ||
40 | ||
41 | See also: | |
42 | M. Shreedhar and George Varghese "Efficient Fair | |
43 | Queuing using Deficit Round Robin", Proc. SIGCOMM 95. | |
44 | ||
45 | ||
10297b99 | 46 | This is not the thing that is usually called (W)FQ nowadays. |
1da177e4 LT |
47 | It does not use any timestamp mechanism, but instead |
48 | processes queues in round-robin order. | |
49 | ||
50 | ADVANTAGE: | |
51 | ||
52 | - It is very cheap. Both CPU and memory requirements are minimal. | |
53 | ||
54 | DRAWBACKS: | |
55 | ||
10297b99 | 56 | - "Stochastic" -> It is not 100% fair. |
1da177e4 LT |
57 | When hash collisions occur, several flows are considered as one. |
58 | ||
59 | - "Round-robin" -> It introduces larger delays than virtual clock | |
60 | based schemes, and should not be used for isolating interactive | |
61 | traffic from non-interactive. It means, that this scheduler | |
62 | should be used as leaf of CBQ or P3, which put interactive traffic | |
63 | to higher priority band. | |
64 | ||
65 | We still need true WFQ for top level CSZ, but using WFQ | |
66 | for the best effort traffic is absolutely pointless: | |
67 | SFQ is superior for this purpose. | |
68 | ||
69 | IMPLEMENTATION: | |
18cb8098 ED |
70 | This implementation limits : |
71 | - maximal queue length per flow to 127 packets. | |
72 | - max mtu to 2^18-1; | |
73 | - max 65408 flows, | |
74 | - number of hash buckets to 65536. | |
1da177e4 LT |
75 | |
76 | It is easy to increase these values, but not in flight. */ | |
77 | ||
18cb8098 ED |
78 | #define SFQ_MAX_DEPTH 127 /* max number of packets per flow */ |
79 | #define SFQ_DEFAULT_FLOWS 128 | |
80 | #define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */ | |
81 | #define SFQ_EMPTY_SLOT 0xffff | |
817fb15d ED |
82 | #define SFQ_DEFAULT_HASH_DIVISOR 1024 |
83 | ||
eeaeb068 ED |
84 | /* We use 16 bits to store allot, and want to handle packets up to 64K |
85 | * Scale allot by 8 (1<<3) so that no overflow occurs. | |
86 | */ | |
87 | #define SFQ_ALLOT_SHIFT 3 | |
88 | #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT) | |
1da177e4 | 89 | |
18cb8098 ED |
90 | /* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */ |
91 | typedef u16 sfq_index; | |
1da177e4 | 92 | |
eda83e3b ED |
93 | /* |
94 | * We dont use pointers to save space. | |
18cb8098 ED |
95 | * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array |
96 | * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH] | |
eda83e3b ED |
97 | * are 'pointers' to dep[] array |
98 | */ | |
cc7ec456 | 99 | struct sfq_head { |
1da177e4 LT |
100 | sfq_index next; |
101 | sfq_index prev; | |
102 | }; | |
103 | ||
eda83e3b ED |
104 | struct sfq_slot { |
105 | struct sk_buff *skblist_next; | |
106 | struct sk_buff *skblist_prev; | |
107 | sfq_index qlen; /* number of skbs in skblist */ | |
18cb8098 | 108 | sfq_index next; /* next slot in sfq RR chain */ |
eda83e3b ED |
109 | struct sfq_head dep; /* anchor in dep[] chains */ |
110 | unsigned short hash; /* hash value (index in ht[]) */ | |
111 | short allot; /* credit for this slot */ | |
ddecf0f4 ED |
112 | |
113 | unsigned int backlog; | |
114 | struct red_vars vars; | |
eda83e3b ED |
115 | }; |
116 | ||
cc7ec456 | 117 | struct sfq_sched_data { |
18cb8098 ED |
118 | /* frequently used fields */ |
119 | int limit; /* limit of total number of packets in this qdisc */ | |
817fb15d | 120 | unsigned int divisor; /* number of slots in hash table */ |
ddecf0f4 ED |
121 | u8 headdrop; |
122 | u8 maxdepth; /* limit of packets per flow */ | |
18cb8098 | 123 | |
32740ddc | 124 | u32 perturbation; |
ddecf0f4 ED |
125 | u8 cur_depth; /* depth of longest slot */ |
126 | u8 flags; | |
eeaeb068 | 127 | unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */ |
25d8c0d5 | 128 | struct tcf_proto __rcu *filter_list; |
6529eaba | 129 | struct tcf_block *block; |
18cb8098 ED |
130 | sfq_index *ht; /* Hash table ('divisor' slots) */ |
131 | struct sfq_slot *slots; /* Flows table ('maxflows' entries) */ | |
132 | ||
ddecf0f4 ED |
133 | struct red_parms *red_parms; |
134 | struct tc_sfqred_stats stats; | |
135 | struct sfq_slot *tail; /* current slot in round */ | |
136 | ||
18cb8098 ED |
137 | struct sfq_head dep[SFQ_MAX_DEPTH + 1]; |
138 | /* Linked lists of slots, indexed by depth | |
139 | * dep[0] : list of unused flows | |
140 | * dep[1] : list of flows with 1 packet | |
141 | * dep[X] : list of flows with X packets | |
142 | */ | |
143 | ||
ddecf0f4 | 144 | unsigned int maxflows; /* number of flows in flows array */ |
18cb8098 ED |
145 | int perturb_period; |
146 | unsigned int quantum; /* Allotment per round: MUST BE >= MTU */ | |
147 | struct timer_list perturb_timer; | |
1da177e4 LT |
148 | }; |
149 | ||
eda83e3b ED |
150 | /* |
151 | * sfq_head are either in a sfq_slot or in dep[] array | |
152 | */ | |
153 | static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val) | |
154 | { | |
18cb8098 | 155 | if (val < SFQ_MAX_FLOWS) |
eda83e3b | 156 | return &q->slots[val].dep; |
18cb8098 | 157 | return &q->dep[val - SFQ_MAX_FLOWS]; |
eda83e3b ED |
158 | } |
159 | ||
11fca931 ED |
160 | static unsigned int sfq_hash(const struct sfq_sched_data *q, |
161 | const struct sk_buff *skb) | |
1da177e4 | 162 | { |
ada1dba0 | 163 | return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1); |
1da177e4 LT |
164 | } |
165 | ||
7d2681a6 PM |
166 | static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch, |
167 | int *qerr) | |
168 | { | |
169 | struct sfq_sched_data *q = qdisc_priv(sch); | |
170 | struct tcf_result res; | |
25d8c0d5 | 171 | struct tcf_proto *fl; |
7d2681a6 PM |
172 | int result; |
173 | ||
174 | if (TC_H_MAJ(skb->priority) == sch->handle && | |
175 | TC_H_MIN(skb->priority) > 0 && | |
817fb15d | 176 | TC_H_MIN(skb->priority) <= q->divisor) |
7d2681a6 PM |
177 | return TC_H_MIN(skb->priority); |
178 | ||
25d8c0d5 | 179 | fl = rcu_dereference_bh(q->filter_list); |
ada1dba0 | 180 | if (!fl) |
7d2681a6 PM |
181 | return sfq_hash(q, skb) + 1; |
182 | ||
c27f339a | 183 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; |
87d83093 | 184 | result = tcf_classify(skb, fl, &res, false); |
7d2681a6 PM |
185 | if (result >= 0) { |
186 | #ifdef CONFIG_NET_CLS_ACT | |
187 | switch (result) { | |
188 | case TC_ACT_STOLEN: | |
189 | case TC_ACT_QUEUED: | |
e25ea21f | 190 | case TC_ACT_TRAP: |
378a2f09 | 191 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
7d2681a6 PM |
192 | case TC_ACT_SHOT: |
193 | return 0; | |
194 | } | |
195 | #endif | |
817fb15d | 196 | if (TC_H_MIN(res.classid) <= q->divisor) |
7d2681a6 PM |
197 | return TC_H_MIN(res.classid); |
198 | } | |
199 | return 0; | |
200 | } | |
201 | ||
eda83e3b | 202 | /* |
18cb8098 | 203 | * x : slot number [0 .. SFQ_MAX_FLOWS - 1] |
eda83e3b | 204 | */ |
1da177e4 LT |
205 | static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) |
206 | { | |
207 | sfq_index p, n; | |
18cb8098 ED |
208 | struct sfq_slot *slot = &q->slots[x]; |
209 | int qlen = slot->qlen; | |
eda83e3b | 210 | |
18cb8098 | 211 | p = qlen + SFQ_MAX_FLOWS; |
eda83e3b | 212 | n = q->dep[qlen].next; |
1da177e4 | 213 | |
18cb8098 ED |
214 | slot->dep.next = n; |
215 | slot->dep.prev = p; | |
eda83e3b ED |
216 | |
217 | q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */ | |
218 | sfq_dep_head(q, n)->prev = x; | |
1da177e4 LT |
219 | } |
220 | ||
eda83e3b | 221 | #define sfq_unlink(q, x, n, p) \ |
fa08943b YY |
222 | do { \ |
223 | n = q->slots[x].dep.next; \ | |
224 | p = q->slots[x].dep.prev; \ | |
225 | sfq_dep_head(q, p)->next = n; \ | |
226 | sfq_dep_head(q, n)->prev = p; \ | |
227 | } while (0) | |
eda83e3b ED |
228 | |
229 | ||
1da177e4 LT |
230 | static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) |
231 | { | |
232 | sfq_index p, n; | |
eda83e3b | 233 | int d; |
1da177e4 | 234 | |
eda83e3b | 235 | sfq_unlink(q, x, n, p); |
1da177e4 | 236 | |
eda83e3b ED |
237 | d = q->slots[x].qlen--; |
238 | if (n == p && q->cur_depth == d) | |
239 | q->cur_depth--; | |
1da177e4 LT |
240 | sfq_link(q, x); |
241 | } | |
242 | ||
243 | static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) | |
244 | { | |
245 | sfq_index p, n; | |
246 | int d; | |
247 | ||
eda83e3b | 248 | sfq_unlink(q, x, n, p); |
1da177e4 | 249 | |
eda83e3b ED |
250 | d = ++q->slots[x].qlen; |
251 | if (q->cur_depth < d) | |
252 | q->cur_depth = d; | |
1da177e4 LT |
253 | sfq_link(q, x); |
254 | } | |
255 | ||
eda83e3b ED |
256 | /* helper functions : might be changed when/if skb use a standard list_head */ |
257 | ||
258 | /* remove one skb from tail of slot queue */ | |
259 | static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot) | |
260 | { | |
261 | struct sk_buff *skb = slot->skblist_prev; | |
262 | ||
263 | slot->skblist_prev = skb->prev; | |
ee09b3c1 | 264 | skb->prev->next = (struct sk_buff *)slot; |
eda83e3b ED |
265 | skb->next = skb->prev = NULL; |
266 | return skb; | |
267 | } | |
268 | ||
269 | /* remove one skb from head of slot queue */ | |
270 | static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot) | |
271 | { | |
272 | struct sk_buff *skb = slot->skblist_next; | |
273 | ||
274 | slot->skblist_next = skb->next; | |
18c8d82a | 275 | skb->next->prev = (struct sk_buff *)slot; |
eda83e3b ED |
276 | skb->next = skb->prev = NULL; |
277 | return skb; | |
278 | } | |
279 | ||
280 | static inline void slot_queue_init(struct sfq_slot *slot) | |
281 | { | |
18cb8098 | 282 | memset(slot, 0, sizeof(*slot)); |
eda83e3b ED |
283 | slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot; |
284 | } | |
285 | ||
286 | /* add skb to slot queue (tail add) */ | |
287 | static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb) | |
288 | { | |
289 | skb->prev = slot->skblist_prev; | |
290 | skb->next = (struct sk_buff *)slot; | |
291 | slot->skblist_prev->next = skb; | |
292 | slot->skblist_prev = skb; | |
293 | } | |
294 | ||
1da177e4 LT |
295 | static unsigned int sfq_drop(struct Qdisc *sch) |
296 | { | |
297 | struct sfq_sched_data *q = qdisc_priv(sch); | |
eda83e3b | 298 | sfq_index x, d = q->cur_depth; |
1da177e4 LT |
299 | struct sk_buff *skb; |
300 | unsigned int len; | |
eda83e3b | 301 | struct sfq_slot *slot; |
1da177e4 | 302 | |
eda83e3b | 303 | /* Queue is full! Find the longest slot and drop tail packet from it */ |
1da177e4 | 304 | if (d > 1) { |
eda83e3b ED |
305 | x = q->dep[d].next; |
306 | slot = &q->slots[x]; | |
307 | drop: | |
18cb8098 | 308 | skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot); |
0abf77e5 | 309 | len = qdisc_pkt_len(skb); |
ddecf0f4 | 310 | slot->backlog -= len; |
1da177e4 LT |
311 | sfq_dec(q, x); |
312 | sch->q.qlen--; | |
25331d6c JF |
313 | qdisc_qstats_drop(sch); |
314 | qdisc_qstats_backlog_dec(sch, skb); | |
e8d092aa | 315 | kfree_skb(skb); |
1da177e4 LT |
316 | return len; |
317 | } | |
318 | ||
319 | if (d == 1) { | |
320 | /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */ | |
eda83e3b ED |
321 | x = q->tail->next; |
322 | slot = &q->slots[x]; | |
323 | q->tail->next = slot->next; | |
324 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; | |
325 | goto drop; | |
1da177e4 LT |
326 | } |
327 | ||
328 | return 0; | |
329 | } | |
330 | ||
ddecf0f4 ED |
331 | /* Is ECN parameter configured */ |
332 | static int sfq_prob_mark(const struct sfq_sched_data *q) | |
333 | { | |
334 | return q->flags & TC_RED_ECN; | |
335 | } | |
336 | ||
337 | /* Should packets over max threshold just be marked */ | |
338 | static int sfq_hard_mark(const struct sfq_sched_data *q) | |
339 | { | |
340 | return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN; | |
341 | } | |
342 | ||
343 | static int sfq_headdrop(const struct sfq_sched_data *q) | |
344 | { | |
345 | return q->headdrop; | |
346 | } | |
347 | ||
1da177e4 | 348 | static int |
520ac30f | 349 | sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free) |
1da177e4 LT |
350 | { |
351 | struct sfq_sched_data *q = qdisc_priv(sch); | |
2ccccf5f | 352 | unsigned int hash, dropped; |
8efa8854 | 353 | sfq_index x, qlen; |
eda83e3b | 354 | struct sfq_slot *slot; |
7f3ff4f6 | 355 | int uninitialized_var(ret); |
ddecf0f4 ED |
356 | struct sk_buff *head; |
357 | int delta; | |
7d2681a6 PM |
358 | |
359 | hash = sfq_classify(skb, sch, &ret); | |
360 | if (hash == 0) { | |
c27f339a | 361 | if (ret & __NET_XMIT_BYPASS) |
25331d6c | 362 | qdisc_qstats_drop(sch); |
7d2681a6 PM |
363 | kfree_skb(skb); |
364 | return ret; | |
365 | } | |
366 | hash--; | |
1da177e4 LT |
367 | |
368 | x = q->ht[hash]; | |
eda83e3b ED |
369 | slot = &q->slots[x]; |
370 | if (x == SFQ_EMPTY_SLOT) { | |
371 | x = q->dep[0].next; /* get a free slot */ | |
18cb8098 | 372 | if (x >= SFQ_MAX_FLOWS) |
520ac30f | 373 | return qdisc_drop(skb, sch, to_free); |
eda83e3b ED |
374 | q->ht[hash] = x; |
375 | slot = &q->slots[x]; | |
376 | slot->hash = hash; | |
ddecf0f4 ED |
377 | slot->backlog = 0; /* should already be 0 anyway... */ |
378 | red_set_vars(&slot->vars); | |
379 | goto enqueue; | |
380 | } | |
381 | if (q->red_parms) { | |
382 | slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms, | |
383 | &slot->vars, | |
384 | slot->backlog); | |
385 | switch (red_action(q->red_parms, | |
386 | &slot->vars, | |
387 | slot->vars.qavg)) { | |
388 | case RED_DONT_MARK: | |
389 | break; | |
390 | ||
391 | case RED_PROB_MARK: | |
25331d6c | 392 | qdisc_qstats_overlimit(sch); |
ddecf0f4 ED |
393 | if (sfq_prob_mark(q)) { |
394 | /* We know we have at least one packet in queue */ | |
395 | if (sfq_headdrop(q) && | |
396 | INET_ECN_set_ce(slot->skblist_next)) { | |
397 | q->stats.prob_mark_head++; | |
398 | break; | |
399 | } | |
400 | if (INET_ECN_set_ce(skb)) { | |
401 | q->stats.prob_mark++; | |
402 | break; | |
403 | } | |
404 | } | |
405 | q->stats.prob_drop++; | |
406 | goto congestion_drop; | |
407 | ||
408 | case RED_HARD_MARK: | |
25331d6c | 409 | qdisc_qstats_overlimit(sch); |
ddecf0f4 ED |
410 | if (sfq_hard_mark(q)) { |
411 | /* We know we have at least one packet in queue */ | |
412 | if (sfq_headdrop(q) && | |
413 | INET_ECN_set_ce(slot->skblist_next)) { | |
414 | q->stats.forced_mark_head++; | |
415 | break; | |
416 | } | |
417 | if (INET_ECN_set_ce(skb)) { | |
418 | q->stats.forced_mark++; | |
419 | break; | |
420 | } | |
421 | } | |
422 | q->stats.forced_drop++; | |
423 | goto congestion_drop; | |
424 | } | |
1da177e4 | 425 | } |
6f9e98f7 | 426 | |
18cb8098 | 427 | if (slot->qlen >= q->maxdepth) { |
ddecf0f4 ED |
428 | congestion_drop: |
429 | if (!sfq_headdrop(q)) | |
520ac30f | 430 | return qdisc_drop(skb, sch, to_free); |
18cb8098 | 431 | |
ddecf0f4 | 432 | /* We know we have at least one packet in queue */ |
18cb8098 | 433 | head = slot_dequeue_head(slot); |
ddecf0f4 ED |
434 | delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb); |
435 | sch->qstats.backlog -= delta; | |
436 | slot->backlog -= delta; | |
520ac30f | 437 | qdisc_drop(head, sch, to_free); |
18cb8098 | 438 | |
18cb8098 ED |
439 | slot_queue_add(slot, skb); |
440 | return NET_XMIT_CN; | |
441 | } | |
32740ddc | 442 | |
ddecf0f4 | 443 | enqueue: |
25331d6c | 444 | qdisc_qstats_backlog_inc(sch, skb); |
ddecf0f4 | 445 | slot->backlog += qdisc_pkt_len(skb); |
eda83e3b | 446 | slot_queue_add(slot, skb); |
1da177e4 | 447 | sfq_inc(q, x); |
eda83e3b ED |
448 | if (slot->qlen == 1) { /* The flow is new */ |
449 | if (q->tail == NULL) { /* It is the first flow */ | |
450 | slot->next = x; | |
1da177e4 | 451 | } else { |
eda83e3b ED |
452 | slot->next = q->tail->next; |
453 | q->tail->next = x; | |
1da177e4 | 454 | } |
cc34eb67 ED |
455 | /* We put this flow at the end of our flow list. |
456 | * This might sound unfair for a new flow to wait after old ones, | |
457 | * but we could endup servicing new flows only, and freeze old ones. | |
458 | */ | |
459 | q->tail = slot; | |
ddecf0f4 | 460 | /* We could use a bigger initial quantum for new flows */ |
eeaeb068 | 461 | slot->allot = q->scaled_quantum; |
1da177e4 | 462 | } |
9190b3b3 | 463 | if (++sch->q.qlen <= q->limit) |
9871e50e | 464 | return NET_XMIT_SUCCESS; |
1da177e4 | 465 | |
8efa8854 | 466 | qlen = slot->qlen; |
2ccccf5f | 467 | dropped = sfq_drop(sch); |
8efa8854 ED |
468 | /* Return Congestion Notification only if we dropped a packet |
469 | * from this flow. | |
470 | */ | |
e1738bd9 ED |
471 | if (qlen != slot->qlen) |
472 | return NET_XMIT_CN; | |
473 | ||
474 | /* As we dropped a packet, better let upper stack know this */ | |
2ccccf5f | 475 | qdisc_tree_reduce_backlog(sch, 1, dropped); |
e1738bd9 | 476 | return NET_XMIT_SUCCESS; |
1da177e4 LT |
477 | } |
478 | ||
1da177e4 | 479 | static struct sk_buff * |
6f9e98f7 | 480 | sfq_dequeue(struct Qdisc *sch) |
1da177e4 LT |
481 | { |
482 | struct sfq_sched_data *q = qdisc_priv(sch); | |
483 | struct sk_buff *skb; | |
aa3e2199 | 484 | sfq_index a, next_a; |
eda83e3b | 485 | struct sfq_slot *slot; |
1da177e4 LT |
486 | |
487 | /* No active slots */ | |
eda83e3b | 488 | if (q->tail == NULL) |
1da177e4 LT |
489 | return NULL; |
490 | ||
eeaeb068 | 491 | next_slot: |
eda83e3b ED |
492 | a = q->tail->next; |
493 | slot = &q->slots[a]; | |
eeaeb068 ED |
494 | if (slot->allot <= 0) { |
495 | q->tail = slot; | |
496 | slot->allot += q->scaled_quantum; | |
497 | goto next_slot; | |
498 | } | |
eda83e3b | 499 | skb = slot_dequeue_head(slot); |
1da177e4 | 500 | sfq_dec(q, a); |
9190b3b3 | 501 | qdisc_bstats_update(sch, skb); |
1da177e4 | 502 | sch->q.qlen--; |
25331d6c | 503 | qdisc_qstats_backlog_dec(sch, skb); |
ddecf0f4 | 504 | slot->backlog -= qdisc_pkt_len(skb); |
1da177e4 | 505 | /* Is the slot empty? */ |
eda83e3b ED |
506 | if (slot->qlen == 0) { |
507 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; | |
508 | next_a = slot->next; | |
aa3e2199 | 509 | if (a == next_a) { |
eda83e3b | 510 | q->tail = NULL; /* no more active slots */ |
1da177e4 LT |
511 | return skb; |
512 | } | |
eda83e3b | 513 | q->tail->next = next_a; |
eeaeb068 ED |
514 | } else { |
515 | slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb)); | |
1da177e4 LT |
516 | } |
517 | return skb; | |
518 | } | |
519 | ||
520 | static void | |
6f9e98f7 | 521 | sfq_reset(struct Qdisc *sch) |
1da177e4 LT |
522 | { |
523 | struct sk_buff *skb; | |
524 | ||
525 | while ((skb = sfq_dequeue(sch)) != NULL) | |
fea02478 | 526 | rtnl_kfree_skbs(skb, skb); |
1da177e4 LT |
527 | } |
528 | ||
225d9b89 ED |
529 | /* |
530 | * When q->perturbation is changed, we rehash all queued skbs | |
531 | * to avoid OOO (Out Of Order) effects. | |
532 | * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change | |
533 | * counters. | |
534 | */ | |
18cb8098 | 535 | static void sfq_rehash(struct Qdisc *sch) |
225d9b89 | 536 | { |
18cb8098 | 537 | struct sfq_sched_data *q = qdisc_priv(sch); |
225d9b89 ED |
538 | struct sk_buff *skb; |
539 | int i; | |
540 | struct sfq_slot *slot; | |
541 | struct sk_buff_head list; | |
18cb8098 | 542 | int dropped = 0; |
2ccccf5f | 543 | unsigned int drop_len = 0; |
225d9b89 ED |
544 | |
545 | __skb_queue_head_init(&list); | |
546 | ||
18cb8098 | 547 | for (i = 0; i < q->maxflows; i++) { |
225d9b89 ED |
548 | slot = &q->slots[i]; |
549 | if (!slot->qlen) | |
550 | continue; | |
551 | while (slot->qlen) { | |
552 | skb = slot_dequeue_head(slot); | |
553 | sfq_dec(q, i); | |
554 | __skb_queue_tail(&list, skb); | |
555 | } | |
ddecf0f4 ED |
556 | slot->backlog = 0; |
557 | red_set_vars(&slot->vars); | |
225d9b89 ED |
558 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; |
559 | } | |
560 | q->tail = NULL; | |
561 | ||
562 | while ((skb = __skb_dequeue(&list)) != NULL) { | |
563 | unsigned int hash = sfq_hash(q, skb); | |
564 | sfq_index x = q->ht[hash]; | |
565 | ||
566 | slot = &q->slots[x]; | |
567 | if (x == SFQ_EMPTY_SLOT) { | |
568 | x = q->dep[0].next; /* get a free slot */ | |
18cb8098 | 569 | if (x >= SFQ_MAX_FLOWS) { |
25331d6c JF |
570 | drop: |
571 | qdisc_qstats_backlog_dec(sch, skb); | |
2ccccf5f | 572 | drop_len += qdisc_pkt_len(skb); |
18cb8098 ED |
573 | kfree_skb(skb); |
574 | dropped++; | |
575 | continue; | |
576 | } | |
225d9b89 ED |
577 | q->ht[hash] = x; |
578 | slot = &q->slots[x]; | |
579 | slot->hash = hash; | |
580 | } | |
18cb8098 ED |
581 | if (slot->qlen >= q->maxdepth) |
582 | goto drop; | |
225d9b89 | 583 | slot_queue_add(slot, skb); |
ddecf0f4 ED |
584 | if (q->red_parms) |
585 | slot->vars.qavg = red_calc_qavg(q->red_parms, | |
586 | &slot->vars, | |
587 | slot->backlog); | |
588 | slot->backlog += qdisc_pkt_len(skb); | |
225d9b89 ED |
589 | sfq_inc(q, x); |
590 | if (slot->qlen == 1) { /* The flow is new */ | |
591 | if (q->tail == NULL) { /* It is the first flow */ | |
592 | slot->next = x; | |
593 | } else { | |
594 | slot->next = q->tail->next; | |
595 | q->tail->next = x; | |
596 | } | |
597 | q->tail = slot; | |
598 | slot->allot = q->scaled_quantum; | |
599 | } | |
600 | } | |
18cb8098 | 601 | sch->q.qlen -= dropped; |
2ccccf5f | 602 | qdisc_tree_reduce_backlog(sch, dropped, drop_len); |
225d9b89 ED |
603 | } |
604 | ||
1da177e4 LT |
605 | static void sfq_perturbation(unsigned long arg) |
606 | { | |
6f9e98f7 | 607 | struct Qdisc *sch = (struct Qdisc *)arg; |
1da177e4 | 608 | struct sfq_sched_data *q = qdisc_priv(sch); |
225d9b89 | 609 | spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch)); |
1da177e4 | 610 | |
225d9b89 | 611 | spin_lock(root_lock); |
63862b5b | 612 | q->perturbation = prandom_u32(); |
225d9b89 | 613 | if (!q->filter_list && q->tail) |
18cb8098 | 614 | sfq_rehash(sch); |
225d9b89 | 615 | spin_unlock(root_lock); |
1da177e4 | 616 | |
32740ddc AK |
617 | if (q->perturb_period) |
618 | mod_timer(&q->perturb_timer, jiffies + q->perturb_period); | |
1da177e4 LT |
619 | } |
620 | ||
1e90474c | 621 | static int sfq_change(struct Qdisc *sch, struct nlattr *opt) |
1da177e4 LT |
622 | { |
623 | struct sfq_sched_data *q = qdisc_priv(sch); | |
1e90474c | 624 | struct tc_sfq_qopt *ctl = nla_data(opt); |
18cb8098 | 625 | struct tc_sfq_qopt_v1 *ctl_v1 = NULL; |
2ccccf5f | 626 | unsigned int qlen, dropped = 0; |
ddecf0f4 | 627 | struct red_parms *p = NULL; |
1da177e4 | 628 | |
1e90474c | 629 | if (opt->nla_len < nla_attr_size(sizeof(*ctl))) |
1da177e4 | 630 | return -EINVAL; |
18cb8098 ED |
631 | if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1))) |
632 | ctl_v1 = nla_data(opt); | |
119b3d38 | 633 | if (ctl->divisor && |
634 | (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536)) | |
635 | return -EINVAL; | |
ddecf0f4 ED |
636 | if (ctl_v1 && ctl_v1->qth_min) { |
637 | p = kmalloc(sizeof(*p), GFP_KERNEL); | |
638 | if (!p) | |
639 | return -ENOMEM; | |
640 | } | |
1da177e4 | 641 | sch_tree_lock(sch); |
18cb8098 ED |
642 | if (ctl->quantum) { |
643 | q->quantum = ctl->quantum; | |
644 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
645 | } | |
6f9e98f7 | 646 | q->perturb_period = ctl->perturb_period * HZ; |
18cb8098 ED |
647 | if (ctl->flows) |
648 | q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS); | |
649 | if (ctl->divisor) { | |
817fb15d | 650 | q->divisor = ctl->divisor; |
18cb8098 ED |
651 | q->maxflows = min_t(u32, q->maxflows, q->divisor); |
652 | } | |
653 | if (ctl_v1) { | |
654 | if (ctl_v1->depth) | |
655 | q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH); | |
ddecf0f4 ED |
656 | if (p) { |
657 | swap(q->red_parms, p); | |
658 | red_set_parms(q->red_parms, | |
659 | ctl_v1->qth_min, ctl_v1->qth_max, | |
660 | ctl_v1->Wlog, | |
661 | ctl_v1->Plog, ctl_v1->Scell_log, | |
662 | NULL, | |
663 | ctl_v1->max_P); | |
664 | } | |
665 | q->flags = ctl_v1->flags; | |
18cb8098 ED |
666 | q->headdrop = ctl_v1->headdrop; |
667 | } | |
668 | if (ctl->limit) { | |
669 | q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows); | |
670 | q->maxflows = min_t(u32, q->maxflows, q->limit); | |
671 | } | |
672 | ||
5e50da01 | 673 | qlen = sch->q.qlen; |
5588b40d | 674 | while (sch->q.qlen > q->limit) |
2ccccf5f WC |
675 | dropped += sfq_drop(sch); |
676 | qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped); | |
1da177e4 LT |
677 | |
678 | del_timer(&q->perturb_timer); | |
679 | if (q->perturb_period) { | |
32740ddc | 680 | mod_timer(&q->perturb_timer, jiffies + q->perturb_period); |
63862b5b | 681 | q->perturbation = prandom_u32(); |
1da177e4 LT |
682 | } |
683 | sch_tree_unlock(sch); | |
ddecf0f4 | 684 | kfree(p); |
1da177e4 LT |
685 | return 0; |
686 | } | |
687 | ||
bd16a6cc ED |
688 | static void *sfq_alloc(size_t sz) |
689 | { | |
752ade68 | 690 | return kvmalloc(sz, GFP_KERNEL); |
bd16a6cc ED |
691 | } |
692 | ||
693 | static void sfq_free(void *addr) | |
694 | { | |
4cb28970 | 695 | kvfree(addr); |
bd16a6cc ED |
696 | } |
697 | ||
698 | static void sfq_destroy(struct Qdisc *sch) | |
699 | { | |
700 | struct sfq_sched_data *q = qdisc_priv(sch); | |
701 | ||
6529eaba | 702 | tcf_block_put(q->block); |
bd16a6cc ED |
703 | q->perturb_period = 0; |
704 | del_timer_sync(&q->perturb_timer); | |
705 | sfq_free(q->ht); | |
18cb8098 | 706 | sfq_free(q->slots); |
ddecf0f4 | 707 | kfree(q->red_parms); |
bd16a6cc ED |
708 | } |
709 | ||
1e90474c | 710 | static int sfq_init(struct Qdisc *sch, struct nlattr *opt) |
1da177e4 LT |
711 | { |
712 | struct sfq_sched_data *q = qdisc_priv(sch); | |
713 | int i; | |
6529eaba JP |
714 | int err; |
715 | ||
716 | err = tcf_block_get(&q->block, &q->filter_list); | |
717 | if (err) | |
718 | return err; | |
1da177e4 | 719 | |
3b1af93c GT |
720 | setup_deferrable_timer(&q->perturb_timer, sfq_perturbation, |
721 | (unsigned long)sch); | |
1da177e4 | 722 | |
18cb8098 ED |
723 | for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) { |
724 | q->dep[i].next = i + SFQ_MAX_FLOWS; | |
725 | q->dep[i].prev = i + SFQ_MAX_FLOWS; | |
1da177e4 | 726 | } |
6f9e98f7 | 727 | |
18cb8098 ED |
728 | q->limit = SFQ_MAX_DEPTH; |
729 | q->maxdepth = SFQ_MAX_DEPTH; | |
eda83e3b ED |
730 | q->cur_depth = 0; |
731 | q->tail = NULL; | |
817fb15d | 732 | q->divisor = SFQ_DEFAULT_HASH_DIVISOR; |
18cb8098 | 733 | q->maxflows = SFQ_DEFAULT_FLOWS; |
02a9098e ED |
734 | q->quantum = psched_mtu(qdisc_dev(sch)); |
735 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
736 | q->perturb_period = 0; | |
63862b5b | 737 | q->perturbation = prandom_u32(); |
02a9098e ED |
738 | |
739 | if (opt) { | |
1da177e4 LT |
740 | int err = sfq_change(sch, opt); |
741 | if (err) | |
742 | return err; | |
743 | } | |
6f9e98f7 | 744 | |
bd16a6cc | 745 | q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor); |
18cb8098 ED |
746 | q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows); |
747 | if (!q->ht || !q->slots) { | |
87b60cfa | 748 | /* Note: sfq_destroy() will be called by our caller */ |
817fb15d | 749 | return -ENOMEM; |
bd16a6cc | 750 | } |
87b60cfa | 751 | |
817fb15d ED |
752 | for (i = 0; i < q->divisor; i++) |
753 | q->ht[i] = SFQ_EMPTY_SLOT; | |
754 | ||
18cb8098 | 755 | for (i = 0; i < q->maxflows; i++) { |
18c8d82a | 756 | slot_queue_init(&q->slots[i]); |
1da177e4 | 757 | sfq_link(q, i); |
18c8d82a | 758 | } |
23624935 ED |
759 | if (q->limit >= 1) |
760 | sch->flags |= TCQ_F_CAN_BYPASS; | |
761 | else | |
762 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
1da177e4 LT |
763 | return 0; |
764 | } | |
765 | ||
1da177e4 LT |
766 | static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) |
767 | { | |
768 | struct sfq_sched_data *q = qdisc_priv(sch); | |
27a884dc | 769 | unsigned char *b = skb_tail_pointer(skb); |
18cb8098 | 770 | struct tc_sfq_qopt_v1 opt; |
ddecf0f4 | 771 | struct red_parms *p = q->red_parms; |
18cb8098 ED |
772 | |
773 | memset(&opt, 0, sizeof(opt)); | |
774 | opt.v0.quantum = q->quantum; | |
775 | opt.v0.perturb_period = q->perturb_period / HZ; | |
776 | opt.v0.limit = q->limit; | |
777 | opt.v0.divisor = q->divisor; | |
778 | opt.v0.flows = q->maxflows; | |
779 | opt.depth = q->maxdepth; | |
780 | opt.headdrop = q->headdrop; | |
1da177e4 | 781 | |
ddecf0f4 ED |
782 | if (p) { |
783 | opt.qth_min = p->qth_min >> p->Wlog; | |
784 | opt.qth_max = p->qth_max >> p->Wlog; | |
785 | opt.Wlog = p->Wlog; | |
786 | opt.Plog = p->Plog; | |
787 | opt.Scell_log = p->Scell_log; | |
788 | opt.max_P = p->max_P; | |
789 | } | |
790 | memcpy(&opt.stats, &q->stats, sizeof(opt.stats)); | |
791 | opt.flags = q->flags; | |
792 | ||
1b34ec43 DM |
793 | if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) |
794 | goto nla_put_failure; | |
1da177e4 LT |
795 | |
796 | return skb->len; | |
797 | ||
1e90474c | 798 | nla_put_failure: |
dc5fc579 | 799 | nlmsg_trim(skb, b); |
1da177e4 LT |
800 | return -1; |
801 | } | |
802 | ||
41065fba JP |
803 | static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg) |
804 | { | |
805 | return NULL; | |
806 | } | |
807 | ||
7d2681a6 PM |
808 | static unsigned long sfq_get(struct Qdisc *sch, u32 classid) |
809 | { | |
810 | return 0; | |
811 | } | |
812 | ||
eb4a5527 JP |
813 | static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent, |
814 | u32 classid) | |
815 | { | |
23624935 ED |
816 | /* we cannot bypass queue discipline anymore */ |
817 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
eb4a5527 JP |
818 | return 0; |
819 | } | |
820 | ||
da7115d9 JP |
821 | static void sfq_put(struct Qdisc *q, unsigned long cl) |
822 | { | |
823 | } | |
824 | ||
6529eaba | 825 | static struct tcf_block *sfq_tcf_block(struct Qdisc *sch, unsigned long cl) |
7d2681a6 PM |
826 | { |
827 | struct sfq_sched_data *q = qdisc_priv(sch); | |
828 | ||
829 | if (cl) | |
830 | return NULL; | |
6529eaba | 831 | return q->block; |
7d2681a6 PM |
832 | } |
833 | ||
94de78d1 PM |
834 | static int sfq_dump_class(struct Qdisc *sch, unsigned long cl, |
835 | struct sk_buff *skb, struct tcmsg *tcm) | |
836 | { | |
837 | tcm->tcm_handle |= TC_H_MIN(cl); | |
838 | return 0; | |
839 | } | |
840 | ||
841 | static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl, | |
842 | struct gnet_dump *d) | |
843 | { | |
844 | struct sfq_sched_data *q = qdisc_priv(sch); | |
ee09b3c1 ED |
845 | sfq_index idx = q->ht[cl - 1]; |
846 | struct gnet_stats_queue qs = { 0 }; | |
847 | struct tc_sfq_xstats xstats = { 0 }; | |
c4266263 | 848 | |
ee09b3c1 ED |
849 | if (idx != SFQ_EMPTY_SLOT) { |
850 | const struct sfq_slot *slot = &q->slots[idx]; | |
94de78d1 | 851 | |
eeaeb068 | 852 | xstats.allot = slot->allot << SFQ_ALLOT_SHIFT; |
ee09b3c1 | 853 | qs.qlen = slot->qlen; |
ddecf0f4 | 854 | qs.backlog = slot->backlog; |
ee09b3c1 | 855 | } |
b0ab6f92 | 856 | if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0) |
94de78d1 PM |
857 | return -1; |
858 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
859 | } | |
860 | ||
7d2681a6 PM |
861 | static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) |
862 | { | |
94de78d1 PM |
863 | struct sfq_sched_data *q = qdisc_priv(sch); |
864 | unsigned int i; | |
865 | ||
866 | if (arg->stop) | |
867 | return; | |
868 | ||
817fb15d | 869 | for (i = 0; i < q->divisor; i++) { |
eda83e3b | 870 | if (q->ht[i] == SFQ_EMPTY_SLOT || |
94de78d1 PM |
871 | arg->count < arg->skip) { |
872 | arg->count++; | |
873 | continue; | |
874 | } | |
875 | if (arg->fn(sch, i + 1, arg) < 0) { | |
876 | arg->stop = 1; | |
877 | break; | |
878 | } | |
879 | arg->count++; | |
880 | } | |
7d2681a6 PM |
881 | } |
882 | ||
883 | static const struct Qdisc_class_ops sfq_class_ops = { | |
41065fba | 884 | .leaf = sfq_leaf, |
7d2681a6 | 885 | .get = sfq_get, |
da7115d9 | 886 | .put = sfq_put, |
6529eaba | 887 | .tcf_block = sfq_tcf_block, |
eb4a5527 | 888 | .bind_tcf = sfq_bind, |
da7115d9 | 889 | .unbind_tcf = sfq_put, |
94de78d1 PM |
890 | .dump = sfq_dump_class, |
891 | .dump_stats = sfq_dump_class_stats, | |
7d2681a6 PM |
892 | .walk = sfq_walk, |
893 | }; | |
894 | ||
20fea08b | 895 | static struct Qdisc_ops sfq_qdisc_ops __read_mostly = { |
7d2681a6 | 896 | .cl_ops = &sfq_class_ops, |
1da177e4 LT |
897 | .id = "sfq", |
898 | .priv_size = sizeof(struct sfq_sched_data), | |
899 | .enqueue = sfq_enqueue, | |
900 | .dequeue = sfq_dequeue, | |
07bd8df5 | 901 | .peek = qdisc_peek_dequeued, |
1da177e4 LT |
902 | .init = sfq_init, |
903 | .reset = sfq_reset, | |
904 | .destroy = sfq_destroy, | |
905 | .change = NULL, | |
906 | .dump = sfq_dump, | |
907 | .owner = THIS_MODULE, | |
908 | }; | |
909 | ||
910 | static int __init sfq_module_init(void) | |
911 | { | |
912 | return register_qdisc(&sfq_qdisc_ops); | |
913 | } | |
10297b99 | 914 | static void __exit sfq_module_exit(void) |
1da177e4 LT |
915 | { |
916 | unregister_qdisc(&sfq_qdisc_ops); | |
917 | } | |
918 | module_init(sfq_module_init) | |
919 | module_exit(sfq_module_exit) | |
920 | MODULE_LICENSE("GPL"); |