]>
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> |
ddecf0f4 | 26 | #include <net/red.h> |
1da177e4 LT |
27 | |
28 | ||
29 | /* Stochastic Fairness Queuing algorithm. | |
30 | ======================================= | |
31 | ||
32 | Source: | |
33 | Paul E. McKenney "Stochastic Fairness Queuing", | |
34 | IEEE INFOCOMM'90 Proceedings, San Francisco, 1990. | |
35 | ||
36 | Paul E. McKenney "Stochastic Fairness Queuing", | |
37 | "Interworking: Research and Experience", v.2, 1991, p.113-131. | |
38 | ||
39 | ||
40 | See also: | |
41 | M. Shreedhar and George Varghese "Efficient Fair | |
42 | Queuing using Deficit Round Robin", Proc. SIGCOMM 95. | |
43 | ||
44 | ||
10297b99 | 45 | This is not the thing that is usually called (W)FQ nowadays. |
1da177e4 LT |
46 | It does not use any timestamp mechanism, but instead |
47 | processes queues in round-robin order. | |
48 | ||
49 | ADVANTAGE: | |
50 | ||
51 | - It is very cheap. Both CPU and memory requirements are minimal. | |
52 | ||
53 | DRAWBACKS: | |
54 | ||
10297b99 | 55 | - "Stochastic" -> It is not 100% fair. |
1da177e4 LT |
56 | When hash collisions occur, several flows are considered as one. |
57 | ||
58 | - "Round-robin" -> It introduces larger delays than virtual clock | |
59 | based schemes, and should not be used for isolating interactive | |
60 | traffic from non-interactive. It means, that this scheduler | |
61 | should be used as leaf of CBQ or P3, which put interactive traffic | |
62 | to higher priority band. | |
63 | ||
64 | We still need true WFQ for top level CSZ, but using WFQ | |
65 | for the best effort traffic is absolutely pointless: | |
66 | SFQ is superior for this purpose. | |
67 | ||
68 | IMPLEMENTATION: | |
18cb8098 ED |
69 | This implementation limits : |
70 | - maximal queue length per flow to 127 packets. | |
71 | - max mtu to 2^18-1; | |
72 | - max 65408 flows, | |
73 | - number of hash buckets to 65536. | |
1da177e4 LT |
74 | |
75 | It is easy to increase these values, but not in flight. */ | |
76 | ||
18cb8098 ED |
77 | #define SFQ_MAX_DEPTH 127 /* max number of packets per flow */ |
78 | #define SFQ_DEFAULT_FLOWS 128 | |
79 | #define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */ | |
80 | #define SFQ_EMPTY_SLOT 0xffff | |
817fb15d ED |
81 | #define SFQ_DEFAULT_HASH_DIVISOR 1024 |
82 | ||
eeaeb068 ED |
83 | /* We use 16 bits to store allot, and want to handle packets up to 64K |
84 | * Scale allot by 8 (1<<3) so that no overflow occurs. | |
85 | */ | |
86 | #define SFQ_ALLOT_SHIFT 3 | |
87 | #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT) | |
1da177e4 | 88 | |
18cb8098 ED |
89 | /* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */ |
90 | typedef u16 sfq_index; | |
1da177e4 | 91 | |
eda83e3b ED |
92 | /* |
93 | * We dont use pointers to save space. | |
18cb8098 ED |
94 | * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array |
95 | * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH] | |
eda83e3b ED |
96 | * are 'pointers' to dep[] array |
97 | */ | |
cc7ec456 | 98 | struct sfq_head { |
1da177e4 LT |
99 | sfq_index next; |
100 | sfq_index prev; | |
101 | }; | |
102 | ||
eda83e3b ED |
103 | struct sfq_slot { |
104 | struct sk_buff *skblist_next; | |
105 | struct sk_buff *skblist_prev; | |
106 | sfq_index qlen; /* number of skbs in skblist */ | |
18cb8098 | 107 | sfq_index next; /* next slot in sfq RR chain */ |
eda83e3b ED |
108 | struct sfq_head dep; /* anchor in dep[] chains */ |
109 | unsigned short hash; /* hash value (index in ht[]) */ | |
110 | short allot; /* credit for this slot */ | |
ddecf0f4 ED |
111 | |
112 | unsigned int backlog; | |
113 | struct red_vars vars; | |
eda83e3b ED |
114 | }; |
115 | ||
cc7ec456 | 116 | struct sfq_sched_data { |
18cb8098 ED |
117 | /* frequently used fields */ |
118 | int limit; /* limit of total number of packets in this qdisc */ | |
817fb15d | 119 | unsigned int divisor; /* number of slots in hash table */ |
ddecf0f4 ED |
120 | u8 headdrop; |
121 | u8 maxdepth; /* limit of packets per flow */ | |
18cb8098 | 122 | |
32740ddc | 123 | u32 perturbation; |
ddecf0f4 ED |
124 | u8 cur_depth; /* depth of longest slot */ |
125 | u8 flags; | |
eeaeb068 | 126 | unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */ |
25d8c0d5 | 127 | struct tcf_proto __rcu *filter_list; |
18cb8098 ED |
128 | sfq_index *ht; /* Hash table ('divisor' slots) */ |
129 | struct sfq_slot *slots; /* Flows table ('maxflows' entries) */ | |
130 | ||
ddecf0f4 ED |
131 | struct red_parms *red_parms; |
132 | struct tc_sfqred_stats stats; | |
133 | struct sfq_slot *tail; /* current slot in round */ | |
134 | ||
18cb8098 ED |
135 | struct sfq_head dep[SFQ_MAX_DEPTH + 1]; |
136 | /* Linked lists of slots, indexed by depth | |
137 | * dep[0] : list of unused flows | |
138 | * dep[1] : list of flows with 1 packet | |
139 | * dep[X] : list of flows with X packets | |
140 | */ | |
141 | ||
ddecf0f4 | 142 | unsigned int maxflows; /* number of flows in flows array */ |
18cb8098 ED |
143 | int perturb_period; |
144 | unsigned int quantum; /* Allotment per round: MUST BE >= MTU */ | |
145 | struct timer_list perturb_timer; | |
1da177e4 LT |
146 | }; |
147 | ||
eda83e3b ED |
148 | /* |
149 | * sfq_head are either in a sfq_slot or in dep[] array | |
150 | */ | |
151 | static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val) | |
152 | { | |
18cb8098 | 153 | if (val < SFQ_MAX_FLOWS) |
eda83e3b | 154 | return &q->slots[val].dep; |
18cb8098 | 155 | return &q->dep[val - SFQ_MAX_FLOWS]; |
eda83e3b ED |
156 | } |
157 | ||
11fca931 ED |
158 | static unsigned int sfq_hash(const struct sfq_sched_data *q, |
159 | const struct sk_buff *skb) | |
1da177e4 | 160 | { |
ada1dba0 | 161 | return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1); |
1da177e4 LT |
162 | } |
163 | ||
7d2681a6 PM |
164 | static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch, |
165 | int *qerr) | |
166 | { | |
167 | struct sfq_sched_data *q = qdisc_priv(sch); | |
168 | struct tcf_result res; | |
25d8c0d5 | 169 | struct tcf_proto *fl; |
7d2681a6 PM |
170 | int result; |
171 | ||
172 | if (TC_H_MAJ(skb->priority) == sch->handle && | |
173 | TC_H_MIN(skb->priority) > 0 && | |
817fb15d | 174 | TC_H_MIN(skb->priority) <= q->divisor) |
7d2681a6 PM |
175 | return TC_H_MIN(skb->priority); |
176 | ||
25d8c0d5 | 177 | fl = rcu_dereference_bh(q->filter_list); |
ada1dba0 | 178 | if (!fl) |
7d2681a6 PM |
179 | return sfq_hash(q, skb) + 1; |
180 | ||
c27f339a | 181 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; |
25d8c0d5 | 182 | result = tc_classify(skb, fl, &res); |
7d2681a6 PM |
183 | if (result >= 0) { |
184 | #ifdef CONFIG_NET_CLS_ACT | |
185 | switch (result) { | |
186 | case TC_ACT_STOLEN: | |
187 | case TC_ACT_QUEUED: | |
378a2f09 | 188 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
7d2681a6 PM |
189 | case TC_ACT_SHOT: |
190 | return 0; | |
191 | } | |
192 | #endif | |
817fb15d | 193 | if (TC_H_MIN(res.classid) <= q->divisor) |
7d2681a6 PM |
194 | return TC_H_MIN(res.classid); |
195 | } | |
196 | return 0; | |
197 | } | |
198 | ||
eda83e3b | 199 | /* |
18cb8098 | 200 | * x : slot number [0 .. SFQ_MAX_FLOWS - 1] |
eda83e3b | 201 | */ |
1da177e4 LT |
202 | static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) |
203 | { | |
204 | sfq_index p, n; | |
18cb8098 ED |
205 | struct sfq_slot *slot = &q->slots[x]; |
206 | int qlen = slot->qlen; | |
eda83e3b | 207 | |
18cb8098 | 208 | p = qlen + SFQ_MAX_FLOWS; |
eda83e3b | 209 | n = q->dep[qlen].next; |
1da177e4 | 210 | |
18cb8098 ED |
211 | slot->dep.next = n; |
212 | slot->dep.prev = p; | |
eda83e3b ED |
213 | |
214 | q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */ | |
215 | sfq_dep_head(q, n)->prev = x; | |
1da177e4 LT |
216 | } |
217 | ||
eda83e3b | 218 | #define sfq_unlink(q, x, n, p) \ |
fa08943b YY |
219 | do { \ |
220 | n = q->slots[x].dep.next; \ | |
221 | p = q->slots[x].dep.prev; \ | |
222 | sfq_dep_head(q, p)->next = n; \ | |
223 | sfq_dep_head(q, n)->prev = p; \ | |
224 | } while (0) | |
eda83e3b ED |
225 | |
226 | ||
1da177e4 LT |
227 | static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) |
228 | { | |
229 | sfq_index p, n; | |
eda83e3b | 230 | int d; |
1da177e4 | 231 | |
eda83e3b | 232 | sfq_unlink(q, x, n, p); |
1da177e4 | 233 | |
eda83e3b ED |
234 | d = q->slots[x].qlen--; |
235 | if (n == p && q->cur_depth == d) | |
236 | q->cur_depth--; | |
1da177e4 LT |
237 | sfq_link(q, x); |
238 | } | |
239 | ||
240 | static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) | |
241 | { | |
242 | sfq_index p, n; | |
243 | int d; | |
244 | ||
eda83e3b | 245 | sfq_unlink(q, x, n, p); |
1da177e4 | 246 | |
eda83e3b ED |
247 | d = ++q->slots[x].qlen; |
248 | if (q->cur_depth < d) | |
249 | q->cur_depth = d; | |
1da177e4 LT |
250 | sfq_link(q, x); |
251 | } | |
252 | ||
eda83e3b ED |
253 | /* helper functions : might be changed when/if skb use a standard list_head */ |
254 | ||
255 | /* remove one skb from tail of slot queue */ | |
256 | static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot) | |
257 | { | |
258 | struct sk_buff *skb = slot->skblist_prev; | |
259 | ||
260 | slot->skblist_prev = skb->prev; | |
ee09b3c1 | 261 | skb->prev->next = (struct sk_buff *)slot; |
eda83e3b ED |
262 | skb->next = skb->prev = NULL; |
263 | return skb; | |
264 | } | |
265 | ||
266 | /* remove one skb from head of slot queue */ | |
267 | static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot) | |
268 | { | |
269 | struct sk_buff *skb = slot->skblist_next; | |
270 | ||
271 | slot->skblist_next = skb->next; | |
18c8d82a | 272 | skb->next->prev = (struct sk_buff *)slot; |
eda83e3b ED |
273 | skb->next = skb->prev = NULL; |
274 | return skb; | |
275 | } | |
276 | ||
277 | static inline void slot_queue_init(struct sfq_slot *slot) | |
278 | { | |
18cb8098 | 279 | memset(slot, 0, sizeof(*slot)); |
eda83e3b ED |
280 | slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot; |
281 | } | |
282 | ||
283 | /* add skb to slot queue (tail add) */ | |
284 | static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb) | |
285 | { | |
286 | skb->prev = slot->skblist_prev; | |
287 | skb->next = (struct sk_buff *)slot; | |
288 | slot->skblist_prev->next = skb; | |
289 | slot->skblist_prev = skb; | |
290 | } | |
291 | ||
1da177e4 LT |
292 | static unsigned int sfq_drop(struct Qdisc *sch) |
293 | { | |
294 | struct sfq_sched_data *q = qdisc_priv(sch); | |
eda83e3b | 295 | sfq_index x, d = q->cur_depth; |
1da177e4 LT |
296 | struct sk_buff *skb; |
297 | unsigned int len; | |
eda83e3b | 298 | struct sfq_slot *slot; |
1da177e4 | 299 | |
eda83e3b | 300 | /* Queue is full! Find the longest slot and drop tail packet from it */ |
1da177e4 | 301 | if (d > 1) { |
eda83e3b ED |
302 | x = q->dep[d].next; |
303 | slot = &q->slots[x]; | |
304 | drop: | |
18cb8098 | 305 | skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot); |
0abf77e5 | 306 | len = qdisc_pkt_len(skb); |
ddecf0f4 | 307 | slot->backlog -= len; |
1da177e4 | 308 | sfq_dec(q, x); |
eda83e3b | 309 | kfree_skb(skb); |
1da177e4 | 310 | sch->q.qlen--; |
25331d6c JF |
311 | qdisc_qstats_drop(sch); |
312 | qdisc_qstats_backlog_dec(sch, skb); | |
1da177e4 LT |
313 | return len; |
314 | } | |
315 | ||
316 | if (d == 1) { | |
317 | /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */ | |
eda83e3b ED |
318 | x = q->tail->next; |
319 | slot = &q->slots[x]; | |
320 | q->tail->next = slot->next; | |
321 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; | |
322 | goto drop; | |
1da177e4 LT |
323 | } |
324 | ||
325 | return 0; | |
326 | } | |
327 | ||
ddecf0f4 ED |
328 | /* Is ECN parameter configured */ |
329 | static int sfq_prob_mark(const struct sfq_sched_data *q) | |
330 | { | |
331 | return q->flags & TC_RED_ECN; | |
332 | } | |
333 | ||
334 | /* Should packets over max threshold just be marked */ | |
335 | static int sfq_hard_mark(const struct sfq_sched_data *q) | |
336 | { | |
337 | return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN; | |
338 | } | |
339 | ||
340 | static int sfq_headdrop(const struct sfq_sched_data *q) | |
341 | { | |
342 | return q->headdrop; | |
343 | } | |
344 | ||
1da177e4 | 345 | static int |
6f9e98f7 | 346 | sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch) |
1da177e4 LT |
347 | { |
348 | struct sfq_sched_data *q = qdisc_priv(sch); | |
7d2681a6 | 349 | unsigned int hash; |
8efa8854 | 350 | sfq_index x, qlen; |
eda83e3b | 351 | struct sfq_slot *slot; |
7f3ff4f6 | 352 | int uninitialized_var(ret); |
ddecf0f4 ED |
353 | struct sk_buff *head; |
354 | int delta; | |
7d2681a6 PM |
355 | |
356 | hash = sfq_classify(skb, sch, &ret); | |
357 | if (hash == 0) { | |
c27f339a | 358 | if (ret & __NET_XMIT_BYPASS) |
25331d6c | 359 | qdisc_qstats_drop(sch); |
7d2681a6 PM |
360 | kfree_skb(skb); |
361 | return ret; | |
362 | } | |
363 | hash--; | |
1da177e4 LT |
364 | |
365 | x = q->ht[hash]; | |
eda83e3b ED |
366 | slot = &q->slots[x]; |
367 | if (x == SFQ_EMPTY_SLOT) { | |
368 | x = q->dep[0].next; /* get a free slot */ | |
18cb8098 ED |
369 | if (x >= SFQ_MAX_FLOWS) |
370 | return qdisc_drop(skb, sch); | |
eda83e3b ED |
371 | q->ht[hash] = x; |
372 | slot = &q->slots[x]; | |
373 | slot->hash = hash; | |
ddecf0f4 ED |
374 | slot->backlog = 0; /* should already be 0 anyway... */ |
375 | red_set_vars(&slot->vars); | |
376 | goto enqueue; | |
377 | } | |
378 | if (q->red_parms) { | |
379 | slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms, | |
380 | &slot->vars, | |
381 | slot->backlog); | |
382 | switch (red_action(q->red_parms, | |
383 | &slot->vars, | |
384 | slot->vars.qavg)) { | |
385 | case RED_DONT_MARK: | |
386 | break; | |
387 | ||
388 | case RED_PROB_MARK: | |
25331d6c | 389 | qdisc_qstats_overlimit(sch); |
ddecf0f4 ED |
390 | if (sfq_prob_mark(q)) { |
391 | /* We know we have at least one packet in queue */ | |
392 | if (sfq_headdrop(q) && | |
393 | INET_ECN_set_ce(slot->skblist_next)) { | |
394 | q->stats.prob_mark_head++; | |
395 | break; | |
396 | } | |
397 | if (INET_ECN_set_ce(skb)) { | |
398 | q->stats.prob_mark++; | |
399 | break; | |
400 | } | |
401 | } | |
402 | q->stats.prob_drop++; | |
403 | goto congestion_drop; | |
404 | ||
405 | case RED_HARD_MARK: | |
25331d6c | 406 | qdisc_qstats_overlimit(sch); |
ddecf0f4 ED |
407 | if (sfq_hard_mark(q)) { |
408 | /* We know we have at least one packet in queue */ | |
409 | if (sfq_headdrop(q) && | |
410 | INET_ECN_set_ce(slot->skblist_next)) { | |
411 | q->stats.forced_mark_head++; | |
412 | break; | |
413 | } | |
414 | if (INET_ECN_set_ce(skb)) { | |
415 | q->stats.forced_mark++; | |
416 | break; | |
417 | } | |
418 | } | |
419 | q->stats.forced_drop++; | |
420 | goto congestion_drop; | |
421 | } | |
1da177e4 | 422 | } |
6f9e98f7 | 423 | |
18cb8098 | 424 | if (slot->qlen >= q->maxdepth) { |
ddecf0f4 ED |
425 | congestion_drop: |
426 | if (!sfq_headdrop(q)) | |
18cb8098 ED |
427 | return qdisc_drop(skb, sch); |
428 | ||
ddecf0f4 | 429 | /* We know we have at least one packet in queue */ |
18cb8098 | 430 | head = slot_dequeue_head(slot); |
ddecf0f4 ED |
431 | delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb); |
432 | sch->qstats.backlog -= delta; | |
433 | slot->backlog -= delta; | |
18cb8098 ED |
434 | qdisc_drop(head, sch); |
435 | ||
18cb8098 ED |
436 | slot_queue_add(slot, skb); |
437 | return NET_XMIT_CN; | |
438 | } | |
32740ddc | 439 | |
ddecf0f4 | 440 | enqueue: |
25331d6c | 441 | qdisc_qstats_backlog_inc(sch, skb); |
ddecf0f4 | 442 | slot->backlog += qdisc_pkt_len(skb); |
eda83e3b | 443 | slot_queue_add(slot, skb); |
1da177e4 | 444 | sfq_inc(q, x); |
eda83e3b ED |
445 | if (slot->qlen == 1) { /* The flow is new */ |
446 | if (q->tail == NULL) { /* It is the first flow */ | |
447 | slot->next = x; | |
1da177e4 | 448 | } else { |
eda83e3b ED |
449 | slot->next = q->tail->next; |
450 | q->tail->next = x; | |
1da177e4 | 451 | } |
cc34eb67 ED |
452 | /* We put this flow at the end of our flow list. |
453 | * This might sound unfair for a new flow to wait after old ones, | |
454 | * but we could endup servicing new flows only, and freeze old ones. | |
455 | */ | |
456 | q->tail = slot; | |
ddecf0f4 | 457 | /* We could use a bigger initial quantum for new flows */ |
eeaeb068 | 458 | slot->allot = q->scaled_quantum; |
1da177e4 | 459 | } |
9190b3b3 | 460 | if (++sch->q.qlen <= q->limit) |
9871e50e | 461 | return NET_XMIT_SUCCESS; |
1da177e4 | 462 | |
8efa8854 | 463 | qlen = slot->qlen; |
1da177e4 | 464 | sfq_drop(sch); |
8efa8854 ED |
465 | /* Return Congestion Notification only if we dropped a packet |
466 | * from this flow. | |
467 | */ | |
e1738bd9 ED |
468 | if (qlen != slot->qlen) |
469 | return NET_XMIT_CN; | |
470 | ||
471 | /* As we dropped a packet, better let upper stack know this */ | |
472 | qdisc_tree_decrease_qlen(sch, 1); | |
473 | return NET_XMIT_SUCCESS; | |
1da177e4 LT |
474 | } |
475 | ||
1da177e4 | 476 | static struct sk_buff * |
6f9e98f7 | 477 | sfq_dequeue(struct Qdisc *sch) |
1da177e4 LT |
478 | { |
479 | struct sfq_sched_data *q = qdisc_priv(sch); | |
480 | struct sk_buff *skb; | |
aa3e2199 | 481 | sfq_index a, next_a; |
eda83e3b | 482 | struct sfq_slot *slot; |
1da177e4 LT |
483 | |
484 | /* No active slots */ | |
eda83e3b | 485 | if (q->tail == NULL) |
1da177e4 LT |
486 | return NULL; |
487 | ||
eeaeb068 | 488 | next_slot: |
eda83e3b ED |
489 | a = q->tail->next; |
490 | slot = &q->slots[a]; | |
eeaeb068 ED |
491 | if (slot->allot <= 0) { |
492 | q->tail = slot; | |
493 | slot->allot += q->scaled_quantum; | |
494 | goto next_slot; | |
495 | } | |
eda83e3b | 496 | skb = slot_dequeue_head(slot); |
1da177e4 | 497 | sfq_dec(q, a); |
9190b3b3 | 498 | qdisc_bstats_update(sch, skb); |
1da177e4 | 499 | sch->q.qlen--; |
25331d6c | 500 | qdisc_qstats_backlog_dec(sch, skb); |
ddecf0f4 | 501 | slot->backlog -= qdisc_pkt_len(skb); |
1da177e4 | 502 | /* Is the slot empty? */ |
eda83e3b ED |
503 | if (slot->qlen == 0) { |
504 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; | |
505 | next_a = slot->next; | |
aa3e2199 | 506 | if (a == next_a) { |
eda83e3b | 507 | q->tail = NULL; /* no more active slots */ |
1da177e4 LT |
508 | return skb; |
509 | } | |
eda83e3b | 510 | q->tail->next = next_a; |
eeaeb068 ED |
511 | } else { |
512 | slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb)); | |
1da177e4 LT |
513 | } |
514 | return skb; | |
515 | } | |
516 | ||
517 | static void | |
6f9e98f7 | 518 | sfq_reset(struct Qdisc *sch) |
1da177e4 LT |
519 | { |
520 | struct sk_buff *skb; | |
521 | ||
522 | while ((skb = sfq_dequeue(sch)) != NULL) | |
523 | kfree_skb(skb); | |
524 | } | |
525 | ||
225d9b89 ED |
526 | /* |
527 | * When q->perturbation is changed, we rehash all queued skbs | |
528 | * to avoid OOO (Out Of Order) effects. | |
529 | * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change | |
530 | * counters. | |
531 | */ | |
18cb8098 | 532 | static void sfq_rehash(struct Qdisc *sch) |
225d9b89 | 533 | { |
18cb8098 | 534 | struct sfq_sched_data *q = qdisc_priv(sch); |
225d9b89 ED |
535 | struct sk_buff *skb; |
536 | int i; | |
537 | struct sfq_slot *slot; | |
538 | struct sk_buff_head list; | |
18cb8098 | 539 | int dropped = 0; |
225d9b89 ED |
540 | |
541 | __skb_queue_head_init(&list); | |
542 | ||
18cb8098 | 543 | for (i = 0; i < q->maxflows; i++) { |
225d9b89 ED |
544 | slot = &q->slots[i]; |
545 | if (!slot->qlen) | |
546 | continue; | |
547 | while (slot->qlen) { | |
548 | skb = slot_dequeue_head(slot); | |
549 | sfq_dec(q, i); | |
550 | __skb_queue_tail(&list, skb); | |
551 | } | |
ddecf0f4 ED |
552 | slot->backlog = 0; |
553 | red_set_vars(&slot->vars); | |
225d9b89 ED |
554 | q->ht[slot->hash] = SFQ_EMPTY_SLOT; |
555 | } | |
556 | q->tail = NULL; | |
557 | ||
558 | while ((skb = __skb_dequeue(&list)) != NULL) { | |
559 | unsigned int hash = sfq_hash(q, skb); | |
560 | sfq_index x = q->ht[hash]; | |
561 | ||
562 | slot = &q->slots[x]; | |
563 | if (x == SFQ_EMPTY_SLOT) { | |
564 | x = q->dep[0].next; /* get a free slot */ | |
18cb8098 | 565 | if (x >= SFQ_MAX_FLOWS) { |
25331d6c JF |
566 | drop: |
567 | qdisc_qstats_backlog_dec(sch, skb); | |
18cb8098 ED |
568 | kfree_skb(skb); |
569 | dropped++; | |
570 | continue; | |
571 | } | |
225d9b89 ED |
572 | q->ht[hash] = x; |
573 | slot = &q->slots[x]; | |
574 | slot->hash = hash; | |
575 | } | |
18cb8098 ED |
576 | if (slot->qlen >= q->maxdepth) |
577 | goto drop; | |
225d9b89 | 578 | slot_queue_add(slot, skb); |
ddecf0f4 ED |
579 | if (q->red_parms) |
580 | slot->vars.qavg = red_calc_qavg(q->red_parms, | |
581 | &slot->vars, | |
582 | slot->backlog); | |
583 | slot->backlog += qdisc_pkt_len(skb); | |
225d9b89 ED |
584 | sfq_inc(q, x); |
585 | if (slot->qlen == 1) { /* The flow is new */ | |
586 | if (q->tail == NULL) { /* It is the first flow */ | |
587 | slot->next = x; | |
588 | } else { | |
589 | slot->next = q->tail->next; | |
590 | q->tail->next = x; | |
591 | } | |
592 | q->tail = slot; | |
593 | slot->allot = q->scaled_quantum; | |
594 | } | |
595 | } | |
18cb8098 ED |
596 | sch->q.qlen -= dropped; |
597 | qdisc_tree_decrease_qlen(sch, dropped); | |
225d9b89 ED |
598 | } |
599 | ||
1da177e4 LT |
600 | static void sfq_perturbation(unsigned long arg) |
601 | { | |
6f9e98f7 | 602 | struct Qdisc *sch = (struct Qdisc *)arg; |
1da177e4 | 603 | struct sfq_sched_data *q = qdisc_priv(sch); |
225d9b89 | 604 | spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch)); |
1da177e4 | 605 | |
225d9b89 | 606 | spin_lock(root_lock); |
63862b5b | 607 | q->perturbation = prandom_u32(); |
225d9b89 | 608 | if (!q->filter_list && q->tail) |
18cb8098 | 609 | sfq_rehash(sch); |
225d9b89 | 610 | spin_unlock(root_lock); |
1da177e4 | 611 | |
32740ddc AK |
612 | if (q->perturb_period) |
613 | mod_timer(&q->perturb_timer, jiffies + q->perturb_period); | |
1da177e4 LT |
614 | } |
615 | ||
1e90474c | 616 | static int sfq_change(struct Qdisc *sch, struct nlattr *opt) |
1da177e4 LT |
617 | { |
618 | struct sfq_sched_data *q = qdisc_priv(sch); | |
1e90474c | 619 | struct tc_sfq_qopt *ctl = nla_data(opt); |
18cb8098 | 620 | struct tc_sfq_qopt_v1 *ctl_v1 = NULL; |
5e50da01 | 621 | unsigned int qlen; |
ddecf0f4 | 622 | struct red_parms *p = NULL; |
1da177e4 | 623 | |
1e90474c | 624 | if (opt->nla_len < nla_attr_size(sizeof(*ctl))) |
1da177e4 | 625 | return -EINVAL; |
18cb8098 ED |
626 | if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1))) |
627 | ctl_v1 = nla_data(opt); | |
119b3d38 | 628 | if (ctl->divisor && |
629 | (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536)) | |
630 | return -EINVAL; | |
ddecf0f4 ED |
631 | if (ctl_v1 && ctl_v1->qth_min) { |
632 | p = kmalloc(sizeof(*p), GFP_KERNEL); | |
633 | if (!p) | |
634 | return -ENOMEM; | |
635 | } | |
1da177e4 | 636 | sch_tree_lock(sch); |
18cb8098 ED |
637 | if (ctl->quantum) { |
638 | q->quantum = ctl->quantum; | |
639 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
640 | } | |
6f9e98f7 | 641 | q->perturb_period = ctl->perturb_period * HZ; |
18cb8098 ED |
642 | if (ctl->flows) |
643 | q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS); | |
644 | if (ctl->divisor) { | |
817fb15d | 645 | q->divisor = ctl->divisor; |
18cb8098 ED |
646 | q->maxflows = min_t(u32, q->maxflows, q->divisor); |
647 | } | |
648 | if (ctl_v1) { | |
649 | if (ctl_v1->depth) | |
650 | q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH); | |
ddecf0f4 ED |
651 | if (p) { |
652 | swap(q->red_parms, p); | |
653 | red_set_parms(q->red_parms, | |
654 | ctl_v1->qth_min, ctl_v1->qth_max, | |
655 | ctl_v1->Wlog, | |
656 | ctl_v1->Plog, ctl_v1->Scell_log, | |
657 | NULL, | |
658 | ctl_v1->max_P); | |
659 | } | |
660 | q->flags = ctl_v1->flags; | |
18cb8098 ED |
661 | q->headdrop = ctl_v1->headdrop; |
662 | } | |
663 | if (ctl->limit) { | |
664 | q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows); | |
665 | q->maxflows = min_t(u32, q->maxflows, q->limit); | |
666 | } | |
667 | ||
5e50da01 | 668 | qlen = sch->q.qlen; |
5588b40d | 669 | while (sch->q.qlen > q->limit) |
1da177e4 | 670 | sfq_drop(sch); |
5e50da01 | 671 | qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen); |
1da177e4 LT |
672 | |
673 | del_timer(&q->perturb_timer); | |
674 | if (q->perturb_period) { | |
32740ddc | 675 | mod_timer(&q->perturb_timer, jiffies + q->perturb_period); |
63862b5b | 676 | q->perturbation = prandom_u32(); |
1da177e4 LT |
677 | } |
678 | sch_tree_unlock(sch); | |
ddecf0f4 | 679 | kfree(p); |
1da177e4 LT |
680 | return 0; |
681 | } | |
682 | ||
bd16a6cc ED |
683 | static void *sfq_alloc(size_t sz) |
684 | { | |
685 | void *ptr = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN); | |
686 | ||
687 | if (!ptr) | |
688 | ptr = vmalloc(sz); | |
689 | return ptr; | |
690 | } | |
691 | ||
692 | static void sfq_free(void *addr) | |
693 | { | |
4cb28970 | 694 | kvfree(addr); |
bd16a6cc ED |
695 | } |
696 | ||
697 | static void sfq_destroy(struct Qdisc *sch) | |
698 | { | |
699 | struct sfq_sched_data *q = qdisc_priv(sch); | |
700 | ||
701 | tcf_destroy_chain(&q->filter_list); | |
702 | q->perturb_period = 0; | |
703 | del_timer_sync(&q->perturb_timer); | |
704 | sfq_free(q->ht); | |
18cb8098 | 705 | sfq_free(q->slots); |
ddecf0f4 | 706 | kfree(q->red_parms); |
bd16a6cc ED |
707 | } |
708 | ||
1e90474c | 709 | static int sfq_init(struct Qdisc *sch, struct nlattr *opt) |
1da177e4 LT |
710 | { |
711 | struct sfq_sched_data *q = qdisc_priv(sch); | |
712 | int i; | |
713 | ||
d3e99483 | 714 | q->perturb_timer.function = sfq_perturbation; |
c19a28e1 | 715 | q->perturb_timer.data = (unsigned long)sch; |
d3e99483 | 716 | init_timer_deferrable(&q->perturb_timer); |
1da177e4 | 717 | |
18cb8098 ED |
718 | for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) { |
719 | q->dep[i].next = i + SFQ_MAX_FLOWS; | |
720 | q->dep[i].prev = i + SFQ_MAX_FLOWS; | |
1da177e4 | 721 | } |
6f9e98f7 | 722 | |
18cb8098 ED |
723 | q->limit = SFQ_MAX_DEPTH; |
724 | q->maxdepth = SFQ_MAX_DEPTH; | |
eda83e3b ED |
725 | q->cur_depth = 0; |
726 | q->tail = NULL; | |
817fb15d | 727 | q->divisor = SFQ_DEFAULT_HASH_DIVISOR; |
18cb8098 | 728 | q->maxflows = SFQ_DEFAULT_FLOWS; |
02a9098e ED |
729 | q->quantum = psched_mtu(qdisc_dev(sch)); |
730 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
731 | q->perturb_period = 0; | |
63862b5b | 732 | q->perturbation = prandom_u32(); |
02a9098e ED |
733 | |
734 | if (opt) { | |
1da177e4 LT |
735 | int err = sfq_change(sch, opt); |
736 | if (err) | |
737 | return err; | |
738 | } | |
6f9e98f7 | 739 | |
bd16a6cc | 740 | q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor); |
18cb8098 ED |
741 | q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows); |
742 | if (!q->ht || !q->slots) { | |
bd16a6cc | 743 | sfq_destroy(sch); |
817fb15d | 744 | return -ENOMEM; |
bd16a6cc | 745 | } |
817fb15d ED |
746 | for (i = 0; i < q->divisor; i++) |
747 | q->ht[i] = SFQ_EMPTY_SLOT; | |
748 | ||
18cb8098 | 749 | for (i = 0; i < q->maxflows; i++) { |
18c8d82a | 750 | slot_queue_init(&q->slots[i]); |
1da177e4 | 751 | sfq_link(q, i); |
18c8d82a | 752 | } |
23624935 ED |
753 | if (q->limit >= 1) |
754 | sch->flags |= TCQ_F_CAN_BYPASS; | |
755 | else | |
756 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
1da177e4 LT |
757 | return 0; |
758 | } | |
759 | ||
1da177e4 LT |
760 | static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) |
761 | { | |
762 | struct sfq_sched_data *q = qdisc_priv(sch); | |
27a884dc | 763 | unsigned char *b = skb_tail_pointer(skb); |
18cb8098 | 764 | struct tc_sfq_qopt_v1 opt; |
ddecf0f4 | 765 | struct red_parms *p = q->red_parms; |
18cb8098 ED |
766 | |
767 | memset(&opt, 0, sizeof(opt)); | |
768 | opt.v0.quantum = q->quantum; | |
769 | opt.v0.perturb_period = q->perturb_period / HZ; | |
770 | opt.v0.limit = q->limit; | |
771 | opt.v0.divisor = q->divisor; | |
772 | opt.v0.flows = q->maxflows; | |
773 | opt.depth = q->maxdepth; | |
774 | opt.headdrop = q->headdrop; | |
1da177e4 | 775 | |
ddecf0f4 ED |
776 | if (p) { |
777 | opt.qth_min = p->qth_min >> p->Wlog; | |
778 | opt.qth_max = p->qth_max >> p->Wlog; | |
779 | opt.Wlog = p->Wlog; | |
780 | opt.Plog = p->Plog; | |
781 | opt.Scell_log = p->Scell_log; | |
782 | opt.max_P = p->max_P; | |
783 | } | |
784 | memcpy(&opt.stats, &q->stats, sizeof(opt.stats)); | |
785 | opt.flags = q->flags; | |
786 | ||
1b34ec43 DM |
787 | if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) |
788 | goto nla_put_failure; | |
1da177e4 LT |
789 | |
790 | return skb->len; | |
791 | ||
1e90474c | 792 | nla_put_failure: |
dc5fc579 | 793 | nlmsg_trim(skb, b); |
1da177e4 LT |
794 | return -1; |
795 | } | |
796 | ||
41065fba JP |
797 | static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg) |
798 | { | |
799 | return NULL; | |
800 | } | |
801 | ||
7d2681a6 PM |
802 | static unsigned long sfq_get(struct Qdisc *sch, u32 classid) |
803 | { | |
804 | return 0; | |
805 | } | |
806 | ||
eb4a5527 JP |
807 | static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent, |
808 | u32 classid) | |
809 | { | |
23624935 ED |
810 | /* we cannot bypass queue discipline anymore */ |
811 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
eb4a5527 JP |
812 | return 0; |
813 | } | |
814 | ||
da7115d9 JP |
815 | static void sfq_put(struct Qdisc *q, unsigned long cl) |
816 | { | |
817 | } | |
818 | ||
25d8c0d5 JF |
819 | static struct tcf_proto __rcu **sfq_find_tcf(struct Qdisc *sch, |
820 | unsigned long cl) | |
7d2681a6 PM |
821 | { |
822 | struct sfq_sched_data *q = qdisc_priv(sch); | |
823 | ||
824 | if (cl) | |
825 | return NULL; | |
826 | return &q->filter_list; | |
827 | } | |
828 | ||
94de78d1 PM |
829 | static int sfq_dump_class(struct Qdisc *sch, unsigned long cl, |
830 | struct sk_buff *skb, struct tcmsg *tcm) | |
831 | { | |
832 | tcm->tcm_handle |= TC_H_MIN(cl); | |
833 | return 0; | |
834 | } | |
835 | ||
836 | static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl, | |
837 | struct gnet_dump *d) | |
838 | { | |
839 | struct sfq_sched_data *q = qdisc_priv(sch); | |
ee09b3c1 ED |
840 | sfq_index idx = q->ht[cl - 1]; |
841 | struct gnet_stats_queue qs = { 0 }; | |
842 | struct tc_sfq_xstats xstats = { 0 }; | |
c4266263 | 843 | |
ee09b3c1 ED |
844 | if (idx != SFQ_EMPTY_SLOT) { |
845 | const struct sfq_slot *slot = &q->slots[idx]; | |
94de78d1 | 846 | |
eeaeb068 | 847 | xstats.allot = slot->allot << SFQ_ALLOT_SHIFT; |
ee09b3c1 | 848 | qs.qlen = slot->qlen; |
ddecf0f4 | 849 | qs.backlog = slot->backlog; |
ee09b3c1 | 850 | } |
b0ab6f92 | 851 | if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0) |
94de78d1 PM |
852 | return -1; |
853 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
854 | } | |
855 | ||
7d2681a6 PM |
856 | static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) |
857 | { | |
94de78d1 PM |
858 | struct sfq_sched_data *q = qdisc_priv(sch); |
859 | unsigned int i; | |
860 | ||
861 | if (arg->stop) | |
862 | return; | |
863 | ||
817fb15d | 864 | for (i = 0; i < q->divisor; i++) { |
eda83e3b | 865 | if (q->ht[i] == SFQ_EMPTY_SLOT || |
94de78d1 PM |
866 | arg->count < arg->skip) { |
867 | arg->count++; | |
868 | continue; | |
869 | } | |
870 | if (arg->fn(sch, i + 1, arg) < 0) { | |
871 | arg->stop = 1; | |
872 | break; | |
873 | } | |
874 | arg->count++; | |
875 | } | |
7d2681a6 PM |
876 | } |
877 | ||
878 | static const struct Qdisc_class_ops sfq_class_ops = { | |
41065fba | 879 | .leaf = sfq_leaf, |
7d2681a6 | 880 | .get = sfq_get, |
da7115d9 | 881 | .put = sfq_put, |
7d2681a6 | 882 | .tcf_chain = sfq_find_tcf, |
eb4a5527 | 883 | .bind_tcf = sfq_bind, |
da7115d9 | 884 | .unbind_tcf = sfq_put, |
94de78d1 PM |
885 | .dump = sfq_dump_class, |
886 | .dump_stats = sfq_dump_class_stats, | |
7d2681a6 PM |
887 | .walk = sfq_walk, |
888 | }; | |
889 | ||
20fea08b | 890 | static struct Qdisc_ops sfq_qdisc_ops __read_mostly = { |
7d2681a6 | 891 | .cl_ops = &sfq_class_ops, |
1da177e4 LT |
892 | .id = "sfq", |
893 | .priv_size = sizeof(struct sfq_sched_data), | |
894 | .enqueue = sfq_enqueue, | |
895 | .dequeue = sfq_dequeue, | |
07bd8df5 | 896 | .peek = qdisc_peek_dequeued, |
1da177e4 LT |
897 | .drop = sfq_drop, |
898 | .init = sfq_init, | |
899 | .reset = sfq_reset, | |
900 | .destroy = sfq_destroy, | |
901 | .change = NULL, | |
902 | .dump = sfq_dump, | |
903 | .owner = THIS_MODULE, | |
904 | }; | |
905 | ||
906 | static int __init sfq_module_init(void) | |
907 | { | |
908 | return register_qdisc(&sfq_qdisc_ops); | |
909 | } | |
10297b99 | 910 | static void __exit sfq_module_exit(void) |
1da177e4 LT |
911 | { |
912 | unregister_qdisc(&sfq_qdisc_ops); | |
913 | } | |
914 | module_init(sfq_module_init) | |
915 | module_exit(sfq_module_exit) | |
916 | MODULE_LICENSE("GPL"); |