]>
Commit | Line | Data |
---|---|---|
2874c5fd | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
1da177e4 LT |
2 | /* |
3 | * net/sched/sch_sfq.c Stochastic Fairness Queueing discipline. | |
4 | * | |
1da177e4 LT |
5 | * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> |
6 | */ | |
7 | ||
1da177e4 | 8 | #include <linux/module.h> |
1da177e4 LT |
9 | #include <linux/types.h> |
10 | #include <linux/kernel.h> | |
11 | #include <linux/jiffies.h> | |
12 | #include <linux/string.h> | |
1da177e4 LT |
13 | #include <linux/in.h> |
14 | #include <linux/errno.h> | |
1da177e4 | 15 | #include <linux/init.h> |
1da177e4 | 16 | #include <linux/skbuff.h> |
32740ddc | 17 | #include <linux/jhash.h> |
5a0e3ad6 | 18 | #include <linux/slab.h> |
817fb15d | 19 | #include <linux/vmalloc.h> |
0ba48053 | 20 | #include <net/netlink.h> |
1da177e4 | 21 | #include <net/pkt_sched.h> |
cf1facda | 22 | #include <net/pkt_cls.h> |
ddecf0f4 | 23 | #include <net/red.h> |
1da177e4 LT |
24 | |
25 | ||
26 | /* Stochastic Fairness Queuing algorithm. | |
27 | ======================================= | |
28 | ||
29 | Source: | |
30 | Paul E. McKenney "Stochastic Fairness Queuing", | |
31 | IEEE INFOCOMM'90 Proceedings, San Francisco, 1990. | |
32 | ||
33 | Paul E. McKenney "Stochastic Fairness Queuing", | |
34 | "Interworking: Research and Experience", v.2, 1991, p.113-131. | |
35 | ||
36 | ||
37 | See also: | |
38 | M. Shreedhar and George Varghese "Efficient Fair | |
39 | Queuing using Deficit Round Robin", Proc. SIGCOMM 95. | |
40 | ||
41 | ||
10297b99 | 42 | This is not the thing that is usually called (W)FQ nowadays. |
1da177e4 LT |
43 | It does not use any timestamp mechanism, but instead |
44 | processes queues in round-robin order. | |
45 | ||
46 | ADVANTAGE: | |
47 | ||
48 | - It is very cheap. Both CPU and memory requirements are minimal. | |
49 | ||
50 | DRAWBACKS: | |
51 | ||
10297b99 | 52 | - "Stochastic" -> It is not 100% fair. |
1da177e4 LT |
53 | When hash collisions occur, several flows are considered as one. |
54 | ||
55 | - "Round-robin" -> It introduces larger delays than virtual clock | |
56 | based schemes, and should not be used for isolating interactive | |
57 | traffic from non-interactive. It means, that this scheduler | |
58 | should be used as leaf of CBQ or P3, which put interactive traffic | |
59 | to higher priority band. | |
60 | ||
61 | We still need true WFQ for top level CSZ, but using WFQ | |
62 | for the best effort traffic is absolutely pointless: | |
63 | SFQ is superior for this purpose. | |
64 | ||
65 | IMPLEMENTATION: | |
18cb8098 ED |
66 | This implementation limits : |
67 | - maximal queue length per flow to 127 packets. | |
68 | - max mtu to 2^18-1; | |
69 | - max 65408 flows, | |
70 | - number of hash buckets to 65536. | |
1da177e4 LT |
71 | |
72 | It is easy to increase these values, but not in flight. */ | |
73 | ||
18cb8098 ED |
74 | #define SFQ_MAX_DEPTH 127 /* max number of packets per flow */ |
75 | #define SFQ_DEFAULT_FLOWS 128 | |
76 | #define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */ | |
77 | #define SFQ_EMPTY_SLOT 0xffff | |
817fb15d ED |
78 | #define SFQ_DEFAULT_HASH_DIVISOR 1024 |
79 | ||
eeaeb068 ED |
80 | /* We use 16 bits to store allot, and want to handle packets up to 64K |
81 | * Scale allot by 8 (1<<3) so that no overflow occurs. | |
82 | */ | |
83 | #define SFQ_ALLOT_SHIFT 3 | |
84 | #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT) | |
1da177e4 | 85 | |
18cb8098 ED |
86 | /* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */ |
87 | typedef u16 sfq_index; | |
1da177e4 | 88 | |
eda83e3b ED |
89 | /* |
90 | * We dont use pointers to save space. | |
18cb8098 ED |
91 | * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array |
92 | * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH] | |
eda83e3b ED |
93 | * are 'pointers' to dep[] array |
94 | */ | |
cc7ec456 | 95 | struct sfq_head { |
1da177e4 LT |
96 | sfq_index next; |
97 | sfq_index prev; | |
98 | }; | |
99 | ||
eda83e3b ED |
100 | struct sfq_slot { |
101 | struct sk_buff *skblist_next; | |
102 | struct sk_buff *skblist_prev; | |
103 | sfq_index qlen; /* number of skbs in skblist */ | |
18cb8098 | 104 | sfq_index next; /* next slot in sfq RR chain */ |
eda83e3b ED |
105 | struct sfq_head dep; /* anchor in dep[] chains */ |
106 | unsigned short hash; /* hash value (index in ht[]) */ | |
107 | short allot; /* credit for this slot */ | |
ddecf0f4 ED |
108 | |
109 | unsigned int backlog; | |
110 | struct red_vars vars; | |
eda83e3b ED |
111 | }; |
112 | ||
cc7ec456 | 113 | struct sfq_sched_data { |
18cb8098 ED |
114 | /* frequently used fields */ |
115 | int limit; /* limit of total number of packets in this qdisc */ | |
817fb15d | 116 | unsigned int divisor; /* number of slots in hash table */ |
ddecf0f4 ED |
117 | u8 headdrop; |
118 | u8 maxdepth; /* limit of packets per flow */ | |
18cb8098 | 119 | |
32740ddc | 120 | u32 perturbation; |
ddecf0f4 ED |
121 | u8 cur_depth; /* depth of longest slot */ |
122 | u8 flags; | |
eeaeb068 | 123 | unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */ |
25d8c0d5 | 124 | struct tcf_proto __rcu *filter_list; |
6529eaba | 125 | struct tcf_block *block; |
18cb8098 ED |
126 | sfq_index *ht; /* Hash table ('divisor' slots) */ |
127 | struct sfq_slot *slots; /* Flows table ('maxflows' entries) */ | |
128 | ||
ddecf0f4 ED |
129 | struct red_parms *red_parms; |
130 | struct tc_sfqred_stats stats; | |
131 | struct sfq_slot *tail; /* current slot in round */ | |
132 | ||
18cb8098 ED |
133 | struct sfq_head dep[SFQ_MAX_DEPTH + 1]; |
134 | /* Linked lists of slots, indexed by depth | |
135 | * dep[0] : list of unused flows | |
136 | * dep[1] : list of flows with 1 packet | |
137 | * dep[X] : list of flows with X packets | |
138 | */ | |
139 | ||
ddecf0f4 | 140 | unsigned int maxflows; /* number of flows in flows array */ |
18cb8098 ED |
141 | int perturb_period; |
142 | unsigned int quantum; /* Allotment per round: MUST BE >= MTU */ | |
143 | struct timer_list perturb_timer; | |
cdeabbb8 | 144 | struct Qdisc *sch; |
1da177e4 LT |
145 | }; |
146 | ||
eda83e3b ED |
147 | /* |
148 | * sfq_head are either in a sfq_slot or in dep[] array | |
149 | */ | |
150 | static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val) | |
151 | { | |
18cb8098 | 152 | if (val < SFQ_MAX_FLOWS) |
eda83e3b | 153 | return &q->slots[val].dep; |
18cb8098 | 154 | return &q->dep[val - SFQ_MAX_FLOWS]; |
eda83e3b ED |
155 | } |
156 | ||
11fca931 ED |
157 | static unsigned int sfq_hash(const struct sfq_sched_data *q, |
158 | const struct sk_buff *skb) | |
1da177e4 | 159 | { |
ada1dba0 | 160 | return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1); |
1da177e4 LT |
161 | } |
162 | ||
7d2681a6 PM |
163 | static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch, |
164 | int *qerr) | |
165 | { | |
166 | struct sfq_sched_data *q = qdisc_priv(sch); | |
167 | struct tcf_result res; | |
25d8c0d5 | 168 | struct tcf_proto *fl; |
7d2681a6 PM |
169 | int result; |
170 | ||
171 | if (TC_H_MAJ(skb->priority) == sch->handle && | |
172 | TC_H_MIN(skb->priority) > 0 && | |
817fb15d | 173 | TC_H_MIN(skb->priority) <= q->divisor) |
7d2681a6 PM |
174 | return TC_H_MIN(skb->priority); |
175 | ||
25d8c0d5 | 176 | fl = rcu_dereference_bh(q->filter_list); |
ada1dba0 | 177 | if (!fl) |
7d2681a6 PM |
178 | return sfq_hash(q, skb) + 1; |
179 | ||
c27f339a | 180 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; |
87d83093 | 181 | result = tcf_classify(skb, fl, &res, false); |
7d2681a6 PM |
182 | if (result >= 0) { |
183 | #ifdef CONFIG_NET_CLS_ACT | |
184 | switch (result) { | |
185 | case TC_ACT_STOLEN: | |
186 | case TC_ACT_QUEUED: | |
e25ea21f | 187 | case TC_ACT_TRAP: |
378a2f09 | 188 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
f3ae608e | 189 | /* fall through */ |
7d2681a6 PM |
190 | case TC_ACT_SHOT: |
191 | return 0; | |
192 | } | |
193 | #endif | |
817fb15d | 194 | if (TC_H_MIN(res.classid) <= q->divisor) |
7d2681a6 PM |
195 | return TC_H_MIN(res.classid); |
196 | } | |
197 | return 0; | |
198 | } | |
199 | ||
eda83e3b | 200 | /* |
18cb8098 | 201 | * x : slot number [0 .. SFQ_MAX_FLOWS - 1] |
eda83e3b | 202 | */ |
1da177e4 LT |
203 | static inline void sfq_link(struct sfq_sched_data *q, sfq_index x) |
204 | { | |
205 | sfq_index p, n; | |
18cb8098 ED |
206 | struct sfq_slot *slot = &q->slots[x]; |
207 | int qlen = slot->qlen; | |
eda83e3b | 208 | |
18cb8098 | 209 | p = qlen + SFQ_MAX_FLOWS; |
eda83e3b | 210 | n = q->dep[qlen].next; |
1da177e4 | 211 | |
18cb8098 ED |
212 | slot->dep.next = n; |
213 | slot->dep.prev = p; | |
eda83e3b ED |
214 | |
215 | q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */ | |
216 | sfq_dep_head(q, n)->prev = x; | |
1da177e4 LT |
217 | } |
218 | ||
eda83e3b | 219 | #define sfq_unlink(q, x, n, p) \ |
fa08943b YY |
220 | do { \ |
221 | n = q->slots[x].dep.next; \ | |
222 | p = q->slots[x].dep.prev; \ | |
223 | sfq_dep_head(q, p)->next = n; \ | |
224 | sfq_dep_head(q, n)->prev = p; \ | |
225 | } while (0) | |
eda83e3b ED |
226 | |
227 | ||
1da177e4 LT |
228 | static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x) |
229 | { | |
230 | sfq_index p, n; | |
eda83e3b | 231 | int d; |
1da177e4 | 232 | |
eda83e3b | 233 | sfq_unlink(q, x, n, p); |
1da177e4 | 234 | |
eda83e3b ED |
235 | d = q->slots[x].qlen--; |
236 | if (n == p && q->cur_depth == d) | |
237 | q->cur_depth--; | |
1da177e4 LT |
238 | sfq_link(q, x); |
239 | } | |
240 | ||
241 | static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x) | |
242 | { | |
243 | sfq_index p, n; | |
244 | int d; | |
245 | ||
eda83e3b | 246 | sfq_unlink(q, x, n, p); |
1da177e4 | 247 | |
eda83e3b ED |
248 | d = ++q->slots[x].qlen; |
249 | if (q->cur_depth < d) | |
250 | q->cur_depth = d; | |
1da177e4 LT |
251 | sfq_link(q, x); |
252 | } | |
253 | ||
eda83e3b ED |
254 | /* helper functions : might be changed when/if skb use a standard list_head */ |
255 | ||
256 | /* remove one skb from tail of slot queue */ | |
257 | static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot) | |
258 | { | |
259 | struct sk_buff *skb = slot->skblist_prev; | |
260 | ||
261 | slot->skblist_prev = skb->prev; | |
ee09b3c1 | 262 | skb->prev->next = (struct sk_buff *)slot; |
eda83e3b ED |
263 | skb->next = skb->prev = NULL; |
264 | return skb; | |
265 | } | |
266 | ||
267 | /* remove one skb from head of slot queue */ | |
268 | static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot) | |
269 | { | |
270 | struct sk_buff *skb = slot->skblist_next; | |
271 | ||
272 | slot->skblist_next = skb->next; | |
18c8d82a | 273 | skb->next->prev = (struct sk_buff *)slot; |
eda83e3b ED |
274 | skb->next = skb->prev = NULL; |
275 | return skb; | |
276 | } | |
277 | ||
278 | static inline void slot_queue_init(struct sfq_slot *slot) | |
279 | { | |
18cb8098 | 280 | memset(slot, 0, sizeof(*slot)); |
eda83e3b ED |
281 | slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot; |
282 | } | |
283 | ||
284 | /* add skb to slot queue (tail add) */ | |
285 | static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb) | |
286 | { | |
287 | skb->prev = slot->skblist_prev; | |
288 | skb->next = (struct sk_buff *)slot; | |
289 | slot->skblist_prev->next = skb; | |
290 | slot->skblist_prev = skb; | |
291 | } | |
292 | ||
f9ab7425 | 293 | static unsigned int sfq_drop(struct Qdisc *sch, struct sk_buff **to_free) |
1da177e4 LT |
294 | { |
295 | struct sfq_sched_data *q = qdisc_priv(sch); | |
eda83e3b | 296 | sfq_index x, d = q->cur_depth; |
1da177e4 LT |
297 | struct sk_buff *skb; |
298 | unsigned int len; | |
eda83e3b | 299 | struct sfq_slot *slot; |
1da177e4 | 300 | |
eda83e3b | 301 | /* Queue is full! Find the longest slot and drop tail packet from it */ |
1da177e4 | 302 | if (d > 1) { |
eda83e3b ED |
303 | x = q->dep[d].next; |
304 | slot = &q->slots[x]; | |
305 | drop: | |
18cb8098 | 306 | skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot); |
0abf77e5 | 307 | len = qdisc_pkt_len(skb); |
ddecf0f4 | 308 | slot->backlog -= len; |
1da177e4 LT |
309 | sfq_dec(q, x); |
310 | sch->q.qlen--; | |
25331d6c | 311 | qdisc_qstats_backlog_dec(sch, skb); |
f9ab7425 | 312 | qdisc_drop(skb, sch, to_free); |
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 |
520ac30f | 346 | sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, struct sk_buff **to_free) |
1da177e4 LT |
347 | { |
348 | struct sfq_sched_data *q = qdisc_priv(sch); | |
2ccccf5f | 349 | unsigned int hash, dropped; |
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); |
f9ab7425 | 360 | __qdisc_drop(skb, to_free); |
7d2681a6 PM |
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 | 369 | if (x >= SFQ_MAX_FLOWS) |
520ac30f | 370 | return qdisc_drop(skb, sch, to_free); |
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)) | |
520ac30f | 427 | return qdisc_drop(skb, sch, to_free); |
18cb8098 | 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; | |
520ac30f | 434 | qdisc_drop(head, sch, to_free); |
18cb8098 | 435 | |
18cb8098 | 436 | slot_queue_add(slot, skb); |
325d5dc3 | 437 | qdisc_tree_reduce_backlog(sch, 0, delta); |
18cb8098 ED |
438 | return NET_XMIT_CN; |
439 | } | |
32740ddc | 440 | |
ddecf0f4 | 441 | enqueue: |
25331d6c | 442 | qdisc_qstats_backlog_inc(sch, skb); |
ddecf0f4 | 443 | slot->backlog += qdisc_pkt_len(skb); |
eda83e3b | 444 | slot_queue_add(slot, skb); |
1da177e4 | 445 | sfq_inc(q, x); |
eda83e3b ED |
446 | if (slot->qlen == 1) { /* The flow is new */ |
447 | if (q->tail == NULL) { /* It is the first flow */ | |
448 | slot->next = x; | |
1da177e4 | 449 | } else { |
eda83e3b ED |
450 | slot->next = q->tail->next; |
451 | q->tail->next = x; | |
1da177e4 | 452 | } |
cc34eb67 ED |
453 | /* We put this flow at the end of our flow list. |
454 | * This might sound unfair for a new flow to wait after old ones, | |
455 | * but we could endup servicing new flows only, and freeze old ones. | |
456 | */ | |
457 | q->tail = slot; | |
ddecf0f4 | 458 | /* We could use a bigger initial quantum for new flows */ |
eeaeb068 | 459 | slot->allot = q->scaled_quantum; |
1da177e4 | 460 | } |
9190b3b3 | 461 | if (++sch->q.qlen <= q->limit) |
9871e50e | 462 | return NET_XMIT_SUCCESS; |
1da177e4 | 463 | |
8efa8854 | 464 | qlen = slot->qlen; |
f9ab7425 | 465 | dropped = sfq_drop(sch, to_free); |
8efa8854 ED |
466 | /* Return Congestion Notification only if we dropped a packet |
467 | * from this flow. | |
468 | */ | |
325d5dc3 KK |
469 | if (qlen != slot->qlen) { |
470 | qdisc_tree_reduce_backlog(sch, 0, dropped - qdisc_pkt_len(skb)); | |
e1738bd9 | 471 | return NET_XMIT_CN; |
325d5dc3 | 472 | } |
e1738bd9 ED |
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 | ||
cdeabbb8 | 605 | static void sfq_perturbation(struct timer_list *t) |
1da177e4 | 606 | { |
cdeabbb8 KC |
607 | struct sfq_sched_data *q = from_timer(q, t, perturb_timer); |
608 | struct Qdisc *sch = q->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; |
f9ab7425 GF |
628 | struct sk_buff *to_free = NULL; |
629 | struct sk_buff *tail = NULL; | |
1da177e4 | 630 | |
1e90474c | 631 | if (opt->nla_len < nla_attr_size(sizeof(*ctl))) |
1da177e4 | 632 | return -EINVAL; |
18cb8098 ED |
633 | if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1))) |
634 | ctl_v1 = nla_data(opt); | |
119b3d38 | 635 | if (ctl->divisor && |
636 | (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536)) | |
637 | return -EINVAL; | |
8afa10cb NF |
638 | if (ctl_v1 && !red_check_params(ctl_v1->qth_min, ctl_v1->qth_max, |
639 | ctl_v1->Wlog)) | |
640 | return -EINVAL; | |
ddecf0f4 ED |
641 | if (ctl_v1 && ctl_v1->qth_min) { |
642 | p = kmalloc(sizeof(*p), GFP_KERNEL); | |
643 | if (!p) | |
644 | return -ENOMEM; | |
645 | } | |
1da177e4 | 646 | sch_tree_lock(sch); |
18cb8098 ED |
647 | if (ctl->quantum) { |
648 | q->quantum = ctl->quantum; | |
649 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
650 | } | |
6f9e98f7 | 651 | q->perturb_period = ctl->perturb_period * HZ; |
18cb8098 ED |
652 | if (ctl->flows) |
653 | q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS); | |
654 | if (ctl->divisor) { | |
817fb15d | 655 | q->divisor = ctl->divisor; |
18cb8098 ED |
656 | q->maxflows = min_t(u32, q->maxflows, q->divisor); |
657 | } | |
658 | if (ctl_v1) { | |
659 | if (ctl_v1->depth) | |
660 | q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH); | |
ddecf0f4 ED |
661 | if (p) { |
662 | swap(q->red_parms, p); | |
663 | red_set_parms(q->red_parms, | |
664 | ctl_v1->qth_min, ctl_v1->qth_max, | |
665 | ctl_v1->Wlog, | |
666 | ctl_v1->Plog, ctl_v1->Scell_log, | |
667 | NULL, | |
668 | ctl_v1->max_P); | |
669 | } | |
670 | q->flags = ctl_v1->flags; | |
18cb8098 ED |
671 | q->headdrop = ctl_v1->headdrop; |
672 | } | |
673 | if (ctl->limit) { | |
674 | q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows); | |
675 | q->maxflows = min_t(u32, q->maxflows, q->limit); | |
676 | } | |
677 | ||
5e50da01 | 678 | qlen = sch->q.qlen; |
f9ab7425 GF |
679 | while (sch->q.qlen > q->limit) { |
680 | dropped += sfq_drop(sch, &to_free); | |
681 | if (!tail) | |
682 | tail = to_free; | |
683 | } | |
684 | ||
685 | rtnl_kfree_skbs(to_free, tail); | |
2ccccf5f | 686 | qdisc_tree_reduce_backlog(sch, qlen - sch->q.qlen, dropped); |
1da177e4 LT |
687 | |
688 | del_timer(&q->perturb_timer); | |
689 | if (q->perturb_period) { | |
32740ddc | 690 | mod_timer(&q->perturb_timer, jiffies + q->perturb_period); |
63862b5b | 691 | q->perturbation = prandom_u32(); |
1da177e4 LT |
692 | } |
693 | sch_tree_unlock(sch); | |
ddecf0f4 | 694 | kfree(p); |
1da177e4 LT |
695 | return 0; |
696 | } | |
697 | ||
bd16a6cc ED |
698 | static void *sfq_alloc(size_t sz) |
699 | { | |
752ade68 | 700 | return kvmalloc(sz, GFP_KERNEL); |
bd16a6cc ED |
701 | } |
702 | ||
703 | static void sfq_free(void *addr) | |
704 | { | |
4cb28970 | 705 | kvfree(addr); |
bd16a6cc ED |
706 | } |
707 | ||
708 | static void sfq_destroy(struct Qdisc *sch) | |
709 | { | |
710 | struct sfq_sched_data *q = qdisc_priv(sch); | |
711 | ||
6529eaba | 712 | tcf_block_put(q->block); |
bd16a6cc ED |
713 | q->perturb_period = 0; |
714 | del_timer_sync(&q->perturb_timer); | |
715 | sfq_free(q->ht); | |
18cb8098 | 716 | sfq_free(q->slots); |
ddecf0f4 | 717 | kfree(q->red_parms); |
bd16a6cc ED |
718 | } |
719 | ||
e63d7dfd AA |
720 | static int sfq_init(struct Qdisc *sch, struct nlattr *opt, |
721 | struct netlink_ext_ack *extack) | |
1da177e4 LT |
722 | { |
723 | struct sfq_sched_data *q = qdisc_priv(sch); | |
724 | int i; | |
6529eaba JP |
725 | int err; |
726 | ||
f85729d0 | 727 | q->sch = sch; |
cdeabbb8 | 728 | timer_setup(&q->perturb_timer, sfq_perturbation, TIMER_DEFERRABLE); |
e2326576 | 729 | |
8d1a77f9 | 730 | err = tcf_block_get(&q->block, &q->filter_list, sch, extack); |
6529eaba JP |
731 | if (err) |
732 | return err; | |
1da177e4 | 733 | |
18cb8098 ED |
734 | for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) { |
735 | q->dep[i].next = i + SFQ_MAX_FLOWS; | |
736 | q->dep[i].prev = i + SFQ_MAX_FLOWS; | |
1da177e4 | 737 | } |
6f9e98f7 | 738 | |
18cb8098 ED |
739 | q->limit = SFQ_MAX_DEPTH; |
740 | q->maxdepth = SFQ_MAX_DEPTH; | |
eda83e3b ED |
741 | q->cur_depth = 0; |
742 | q->tail = NULL; | |
817fb15d | 743 | q->divisor = SFQ_DEFAULT_HASH_DIVISOR; |
18cb8098 | 744 | q->maxflows = SFQ_DEFAULT_FLOWS; |
02a9098e ED |
745 | q->quantum = psched_mtu(qdisc_dev(sch)); |
746 | q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum); | |
747 | q->perturb_period = 0; | |
63862b5b | 748 | q->perturbation = prandom_u32(); |
02a9098e ED |
749 | |
750 | if (opt) { | |
1da177e4 LT |
751 | int err = sfq_change(sch, opt); |
752 | if (err) | |
753 | return err; | |
754 | } | |
6f9e98f7 | 755 | |
bd16a6cc | 756 | q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor); |
18cb8098 ED |
757 | q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows); |
758 | if (!q->ht || !q->slots) { | |
87b60cfa | 759 | /* Note: sfq_destroy() will be called by our caller */ |
817fb15d | 760 | return -ENOMEM; |
bd16a6cc | 761 | } |
87b60cfa | 762 | |
817fb15d ED |
763 | for (i = 0; i < q->divisor; i++) |
764 | q->ht[i] = SFQ_EMPTY_SLOT; | |
765 | ||
18cb8098 | 766 | for (i = 0; i < q->maxflows; i++) { |
18c8d82a | 767 | slot_queue_init(&q->slots[i]); |
1da177e4 | 768 | sfq_link(q, i); |
18c8d82a | 769 | } |
23624935 ED |
770 | if (q->limit >= 1) |
771 | sch->flags |= TCQ_F_CAN_BYPASS; | |
772 | else | |
773 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
1da177e4 LT |
774 | return 0; |
775 | } | |
776 | ||
1da177e4 LT |
777 | static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb) |
778 | { | |
779 | struct sfq_sched_data *q = qdisc_priv(sch); | |
27a884dc | 780 | unsigned char *b = skb_tail_pointer(skb); |
18cb8098 | 781 | struct tc_sfq_qopt_v1 opt; |
ddecf0f4 | 782 | struct red_parms *p = q->red_parms; |
18cb8098 ED |
783 | |
784 | memset(&opt, 0, sizeof(opt)); | |
785 | opt.v0.quantum = q->quantum; | |
786 | opt.v0.perturb_period = q->perturb_period / HZ; | |
787 | opt.v0.limit = q->limit; | |
788 | opt.v0.divisor = q->divisor; | |
789 | opt.v0.flows = q->maxflows; | |
790 | opt.depth = q->maxdepth; | |
791 | opt.headdrop = q->headdrop; | |
1da177e4 | 792 | |
ddecf0f4 ED |
793 | if (p) { |
794 | opt.qth_min = p->qth_min >> p->Wlog; | |
795 | opt.qth_max = p->qth_max >> p->Wlog; | |
796 | opt.Wlog = p->Wlog; | |
797 | opt.Plog = p->Plog; | |
798 | opt.Scell_log = p->Scell_log; | |
799 | opt.max_P = p->max_P; | |
800 | } | |
801 | memcpy(&opt.stats, &q->stats, sizeof(opt.stats)); | |
802 | opt.flags = q->flags; | |
803 | ||
1b34ec43 DM |
804 | if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt)) |
805 | goto nla_put_failure; | |
1da177e4 LT |
806 | |
807 | return skb->len; | |
808 | ||
1e90474c | 809 | nla_put_failure: |
dc5fc579 | 810 | nlmsg_trim(skb, b); |
1da177e4 LT |
811 | return -1; |
812 | } | |
813 | ||
41065fba JP |
814 | static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg) |
815 | { | |
816 | return NULL; | |
817 | } | |
818 | ||
143976ce | 819 | static unsigned long sfq_find(struct Qdisc *sch, u32 classid) |
7d2681a6 PM |
820 | { |
821 | return 0; | |
822 | } | |
823 | ||
eb4a5527 JP |
824 | static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent, |
825 | u32 classid) | |
826 | { | |
23624935 ED |
827 | /* we cannot bypass queue discipline anymore */ |
828 | sch->flags &= ~TCQ_F_CAN_BYPASS; | |
eb4a5527 JP |
829 | return 0; |
830 | } | |
831 | ||
143976ce | 832 | static void sfq_unbind(struct Qdisc *q, unsigned long cl) |
da7115d9 JP |
833 | { |
834 | } | |
835 | ||
cbaacc4e AA |
836 | static struct tcf_block *sfq_tcf_block(struct Qdisc *sch, unsigned long cl, |
837 | struct netlink_ext_ack *extack) | |
7d2681a6 PM |
838 | { |
839 | struct sfq_sched_data *q = qdisc_priv(sch); | |
840 | ||
841 | if (cl) | |
842 | return NULL; | |
6529eaba | 843 | return q->block; |
7d2681a6 PM |
844 | } |
845 | ||
94de78d1 PM |
846 | static int sfq_dump_class(struct Qdisc *sch, unsigned long cl, |
847 | struct sk_buff *skb, struct tcmsg *tcm) | |
848 | { | |
849 | tcm->tcm_handle |= TC_H_MIN(cl); | |
850 | return 0; | |
851 | } | |
852 | ||
853 | static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl, | |
854 | struct gnet_dump *d) | |
855 | { | |
856 | struct sfq_sched_data *q = qdisc_priv(sch); | |
ee09b3c1 ED |
857 | sfq_index idx = q->ht[cl - 1]; |
858 | struct gnet_stats_queue qs = { 0 }; | |
859 | struct tc_sfq_xstats xstats = { 0 }; | |
c4266263 | 860 | |
ee09b3c1 ED |
861 | if (idx != SFQ_EMPTY_SLOT) { |
862 | const struct sfq_slot *slot = &q->slots[idx]; | |
94de78d1 | 863 | |
eeaeb068 | 864 | xstats.allot = slot->allot << SFQ_ALLOT_SHIFT; |
ee09b3c1 | 865 | qs.qlen = slot->qlen; |
ddecf0f4 | 866 | qs.backlog = slot->backlog; |
ee09b3c1 | 867 | } |
b0ab6f92 | 868 | if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0) |
94de78d1 PM |
869 | return -1; |
870 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
871 | } | |
872 | ||
7d2681a6 PM |
873 | static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) |
874 | { | |
94de78d1 PM |
875 | struct sfq_sched_data *q = qdisc_priv(sch); |
876 | unsigned int i; | |
877 | ||
878 | if (arg->stop) | |
879 | return; | |
880 | ||
817fb15d | 881 | for (i = 0; i < q->divisor; i++) { |
eda83e3b | 882 | if (q->ht[i] == SFQ_EMPTY_SLOT || |
94de78d1 PM |
883 | arg->count < arg->skip) { |
884 | arg->count++; | |
885 | continue; | |
886 | } | |
887 | if (arg->fn(sch, i + 1, arg) < 0) { | |
888 | arg->stop = 1; | |
889 | break; | |
890 | } | |
891 | arg->count++; | |
892 | } | |
7d2681a6 PM |
893 | } |
894 | ||
895 | static const struct Qdisc_class_ops sfq_class_ops = { | |
41065fba | 896 | .leaf = sfq_leaf, |
143976ce | 897 | .find = sfq_find, |
6529eaba | 898 | .tcf_block = sfq_tcf_block, |
eb4a5527 | 899 | .bind_tcf = sfq_bind, |
143976ce | 900 | .unbind_tcf = sfq_unbind, |
94de78d1 PM |
901 | .dump = sfq_dump_class, |
902 | .dump_stats = sfq_dump_class_stats, | |
7d2681a6 PM |
903 | .walk = sfq_walk, |
904 | }; | |
905 | ||
20fea08b | 906 | static struct Qdisc_ops sfq_qdisc_ops __read_mostly = { |
7d2681a6 | 907 | .cl_ops = &sfq_class_ops, |
1da177e4 LT |
908 | .id = "sfq", |
909 | .priv_size = sizeof(struct sfq_sched_data), | |
910 | .enqueue = sfq_enqueue, | |
911 | .dequeue = sfq_dequeue, | |
07bd8df5 | 912 | .peek = qdisc_peek_dequeued, |
1da177e4 LT |
913 | .init = sfq_init, |
914 | .reset = sfq_reset, | |
915 | .destroy = sfq_destroy, | |
916 | .change = NULL, | |
917 | .dump = sfq_dump, | |
918 | .owner = THIS_MODULE, | |
919 | }; | |
920 | ||
921 | static int __init sfq_module_init(void) | |
922 | { | |
923 | return register_qdisc(&sfq_qdisc_ops); | |
924 | } | |
10297b99 | 925 | static void __exit sfq_module_exit(void) |
1da177e4 LT |
926 | { |
927 | unregister_qdisc(&sfq_qdisc_ops); | |
928 | } | |
929 | module_init(sfq_module_init) | |
930 | module_exit(sfq_module_exit) | |
931 | MODULE_LICENSE("GPL"); |