]>
Commit | Line | Data |
---|---|---|
afe4fd06 ED |
1 | /* |
2 | * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing) | |
3 | * | |
4 | * Copyright (C) 2013 Eric Dumazet <edumazet@google.com> | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License | |
8 | * as published by the Free Software Foundation; either version | |
9 | * 2 of the License, or (at your option) any later version. | |
10 | * | |
11 | * Meant to be mostly used for localy generated traffic : | |
12 | * Fast classification depends on skb->sk being set before reaching us. | |
13 | * If not, (router workload), we use rxhash as fallback, with 32 bits wide hash. | |
14 | * All packets belonging to a socket are considered as a 'flow'. | |
15 | * | |
16 | * Flows are dynamically allocated and stored in a hash table of RB trees | |
17 | * They are also part of one Round Robin 'queues' (new or old flows) | |
18 | * | |
19 | * Burst avoidance (aka pacing) capability : | |
20 | * | |
21 | * Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a | |
22 | * bunch of packets, and this packet scheduler adds delay between | |
23 | * packets to respect rate limitation. | |
24 | * | |
25 | * enqueue() : | |
26 | * - lookup one RB tree (out of 1024 or more) to find the flow. | |
27 | * If non existent flow, create it, add it to the tree. | |
28 | * Add skb to the per flow list of skb (fifo). | |
29 | * - Use a special fifo for high prio packets | |
30 | * | |
31 | * dequeue() : serves flows in Round Robin | |
32 | * Note : When a flow becomes empty, we do not immediately remove it from | |
33 | * rb trees, for performance reasons (its expected to send additional packets, | |
34 | * or SLAB cache will reuse socket for another flow) | |
35 | */ | |
36 | ||
37 | #include <linux/module.h> | |
38 | #include <linux/types.h> | |
39 | #include <linux/kernel.h> | |
40 | #include <linux/jiffies.h> | |
41 | #include <linux/string.h> | |
42 | #include <linux/in.h> | |
43 | #include <linux/errno.h> | |
44 | #include <linux/init.h> | |
45 | #include <linux/skbuff.h> | |
46 | #include <linux/slab.h> | |
47 | #include <linux/rbtree.h> | |
48 | #include <linux/hash.h> | |
08f89b98 | 49 | #include <linux/prefetch.h> |
afe4fd06 ED |
50 | #include <net/netlink.h> |
51 | #include <net/pkt_sched.h> | |
52 | #include <net/sock.h> | |
53 | #include <net/tcp_states.h> | |
54 | ||
55 | /* | |
56 | * Per flow structure, dynamically allocated | |
57 | */ | |
58 | struct fq_flow { | |
59 | struct sk_buff *head; /* list of skbs for this flow : first skb */ | |
60 | union { | |
61 | struct sk_buff *tail; /* last skb in the list */ | |
62 | unsigned long age; /* jiffies when flow was emptied, for gc */ | |
63 | }; | |
64 | struct rb_node fq_node; /* anchor in fq_root[] trees */ | |
65 | struct sock *sk; | |
66 | int qlen; /* number of packets in flow queue */ | |
67 | int credit; | |
68 | u32 socket_hash; /* sk_hash */ | |
69 | struct fq_flow *next; /* next pointer in RR lists, or &detached */ | |
70 | ||
71 | struct rb_node rate_node; /* anchor in q->delayed tree */ | |
72 | u64 time_next_packet; | |
73 | }; | |
74 | ||
75 | struct fq_flow_head { | |
76 | struct fq_flow *first; | |
77 | struct fq_flow *last; | |
78 | }; | |
79 | ||
80 | struct fq_sched_data { | |
81 | struct fq_flow_head new_flows; | |
82 | ||
83 | struct fq_flow_head old_flows; | |
84 | ||
85 | struct rb_root delayed; /* for rate limited flows */ | |
86 | u64 time_next_delayed_flow; | |
87 | ||
88 | struct fq_flow internal; /* for non classified or high prio packets */ | |
89 | u32 quantum; | |
90 | u32 initial_quantum; | |
afe4fd06 ED |
91 | u32 flow_max_rate; /* optional max rate per flow */ |
92 | u32 flow_plimit; /* max packets per flow */ | |
93 | struct rb_root *fq_root; | |
94 | u8 rate_enable; | |
95 | u8 fq_trees_log; | |
96 | ||
97 | u32 flows; | |
98 | u32 inactive_flows; | |
99 | u32 throttled_flows; | |
100 | ||
101 | u64 stat_gc_flows; | |
102 | u64 stat_internal_packets; | |
103 | u64 stat_tcp_retrans; | |
104 | u64 stat_throttled; | |
105 | u64 stat_flows_plimit; | |
106 | u64 stat_pkts_too_long; | |
107 | u64 stat_allocation_errors; | |
108 | struct qdisc_watchdog watchdog; | |
109 | }; | |
110 | ||
111 | /* special value to mark a detached flow (not on old/new list) */ | |
112 | static struct fq_flow detached, throttled; | |
113 | ||
114 | static void fq_flow_set_detached(struct fq_flow *f) | |
115 | { | |
116 | f->next = &detached; | |
117 | } | |
118 | ||
119 | static bool fq_flow_is_detached(const struct fq_flow *f) | |
120 | { | |
121 | return f->next == &detached; | |
122 | } | |
123 | ||
124 | static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) | |
125 | { | |
126 | struct rb_node **p = &q->delayed.rb_node, *parent = NULL; | |
127 | ||
128 | while (*p) { | |
129 | struct fq_flow *aux; | |
130 | ||
131 | parent = *p; | |
132 | aux = container_of(parent, struct fq_flow, rate_node); | |
133 | if (f->time_next_packet >= aux->time_next_packet) | |
134 | p = &parent->rb_right; | |
135 | else | |
136 | p = &parent->rb_left; | |
137 | } | |
138 | rb_link_node(&f->rate_node, parent, p); | |
139 | rb_insert_color(&f->rate_node, &q->delayed); | |
140 | q->throttled_flows++; | |
141 | q->stat_throttled++; | |
142 | ||
143 | f->next = &throttled; | |
144 | if (q->time_next_delayed_flow > f->time_next_packet) | |
145 | q->time_next_delayed_flow = f->time_next_packet; | |
146 | } | |
147 | ||
148 | ||
149 | static struct kmem_cache *fq_flow_cachep __read_mostly; | |
150 | ||
151 | static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) | |
152 | { | |
153 | if (head->first) | |
154 | head->last->next = flow; | |
155 | else | |
156 | head->first = flow; | |
157 | head->last = flow; | |
158 | flow->next = NULL; | |
159 | } | |
160 | ||
161 | /* limit number of collected flows per round */ | |
162 | #define FQ_GC_MAX 8 | |
163 | #define FQ_GC_AGE (3*HZ) | |
164 | ||
165 | static bool fq_gc_candidate(const struct fq_flow *f) | |
166 | { | |
167 | return fq_flow_is_detached(f) && | |
168 | time_after(jiffies, f->age + FQ_GC_AGE); | |
169 | } | |
170 | ||
171 | static void fq_gc(struct fq_sched_data *q, | |
172 | struct rb_root *root, | |
173 | struct sock *sk) | |
174 | { | |
175 | struct fq_flow *f, *tofree[FQ_GC_MAX]; | |
176 | struct rb_node **p, *parent; | |
177 | int fcnt = 0; | |
178 | ||
179 | p = &root->rb_node; | |
180 | parent = NULL; | |
181 | while (*p) { | |
182 | parent = *p; | |
183 | ||
184 | f = container_of(parent, struct fq_flow, fq_node); | |
185 | if (f->sk == sk) | |
186 | break; | |
187 | ||
188 | if (fq_gc_candidate(f)) { | |
189 | tofree[fcnt++] = f; | |
190 | if (fcnt == FQ_GC_MAX) | |
191 | break; | |
192 | } | |
193 | ||
194 | if (f->sk > sk) | |
195 | p = &parent->rb_right; | |
196 | else | |
197 | p = &parent->rb_left; | |
198 | } | |
199 | ||
200 | q->flows -= fcnt; | |
201 | q->inactive_flows -= fcnt; | |
202 | q->stat_gc_flows += fcnt; | |
203 | while (fcnt) { | |
204 | struct fq_flow *f = tofree[--fcnt]; | |
205 | ||
206 | rb_erase(&f->fq_node, root); | |
207 | kmem_cache_free(fq_flow_cachep, f); | |
208 | } | |
209 | } | |
210 | ||
afe4fd06 ED |
211 | static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) |
212 | { | |
213 | struct rb_node **p, *parent; | |
214 | struct sock *sk = skb->sk; | |
215 | struct rb_root *root; | |
216 | struct fq_flow *f; | |
afe4fd06 ED |
217 | |
218 | /* warning: no starvation prevention... */ | |
2abc2f07 | 219 | if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) |
afe4fd06 ED |
220 | return &q->internal; |
221 | ||
222 | if (unlikely(!sk)) { | |
223 | /* By forcing low order bit to 1, we make sure to not | |
224 | * collide with a local flow (socket pointers are word aligned) | |
225 | */ | |
226 | sk = (struct sock *)(skb_get_rxhash(skb) | 1L); | |
227 | } | |
228 | ||
229 | root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)]; | |
230 | ||
231 | if (q->flows >= (2U << q->fq_trees_log) && | |
232 | q->inactive_flows > q->flows/2) | |
233 | fq_gc(q, root, sk); | |
234 | ||
235 | p = &root->rb_node; | |
236 | parent = NULL; | |
237 | while (*p) { | |
238 | parent = *p; | |
239 | ||
240 | f = container_of(parent, struct fq_flow, fq_node); | |
241 | if (f->sk == sk) { | |
242 | /* socket might have been reallocated, so check | |
243 | * if its sk_hash is the same. | |
244 | * It not, we need to refill credit with | |
245 | * initial quantum | |
246 | */ | |
247 | if (unlikely(skb->sk && | |
248 | f->socket_hash != sk->sk_hash)) { | |
249 | f->credit = q->initial_quantum; | |
250 | f->socket_hash = sk->sk_hash; | |
fc59d5bd | 251 | f->time_next_packet = 0ULL; |
afe4fd06 ED |
252 | } |
253 | return f; | |
254 | } | |
255 | if (f->sk > sk) | |
256 | p = &parent->rb_right; | |
257 | else | |
258 | p = &parent->rb_left; | |
259 | } | |
260 | ||
261 | f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); | |
262 | if (unlikely(!f)) { | |
263 | q->stat_allocation_errors++; | |
264 | return &q->internal; | |
265 | } | |
266 | fq_flow_set_detached(f); | |
267 | f->sk = sk; | |
268 | if (skb->sk) | |
269 | f->socket_hash = sk->sk_hash; | |
270 | f->credit = q->initial_quantum; | |
271 | ||
272 | rb_link_node(&f->fq_node, parent, p); | |
273 | rb_insert_color(&f->fq_node, root); | |
274 | ||
275 | q->flows++; | |
276 | q->inactive_flows++; | |
277 | return f; | |
278 | } | |
279 | ||
280 | ||
281 | /* remove one skb from head of flow queue */ | |
8d34ce10 | 282 | static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow) |
afe4fd06 ED |
283 | { |
284 | struct sk_buff *skb = flow->head; | |
285 | ||
286 | if (skb) { | |
287 | flow->head = skb->next; | |
288 | skb->next = NULL; | |
289 | flow->qlen--; | |
8d34ce10 ED |
290 | sch->qstats.backlog -= qdisc_pkt_len(skb); |
291 | sch->q.qlen--; | |
afe4fd06 ED |
292 | } |
293 | return skb; | |
294 | } | |
295 | ||
296 | /* We might add in the future detection of retransmits | |
297 | * For the time being, just return false | |
298 | */ | |
299 | static bool skb_is_retransmit(struct sk_buff *skb) | |
300 | { | |
301 | return false; | |
302 | } | |
303 | ||
304 | /* add skb to flow queue | |
305 | * flow queue is a linked list, kind of FIFO, except for TCP retransmits | |
306 | * We special case tcp retransmits to be transmitted before other packets. | |
307 | * We rely on fact that TCP retransmits are unlikely, so we do not waste | |
308 | * a separate queue or a pointer. | |
309 | * head-> [retrans pkt 1] | |
310 | * [retrans pkt 2] | |
311 | * [ normal pkt 1] | |
312 | * [ normal pkt 2] | |
313 | * [ normal pkt 3] | |
314 | * tail-> [ normal pkt 4] | |
315 | */ | |
316 | static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) | |
317 | { | |
318 | struct sk_buff *prev, *head = flow->head; | |
319 | ||
320 | skb->next = NULL; | |
321 | if (!head) { | |
322 | flow->head = skb; | |
323 | flow->tail = skb; | |
324 | return; | |
325 | } | |
326 | if (likely(!skb_is_retransmit(skb))) { | |
327 | flow->tail->next = skb; | |
328 | flow->tail = skb; | |
329 | return; | |
330 | } | |
331 | ||
332 | /* This skb is a tcp retransmit, | |
333 | * find the last retrans packet in the queue | |
334 | */ | |
335 | prev = NULL; | |
336 | while (skb_is_retransmit(head)) { | |
337 | prev = head; | |
338 | head = head->next; | |
339 | if (!head) | |
340 | break; | |
341 | } | |
342 | if (!prev) { /* no rtx packet in queue, become the new head */ | |
343 | skb->next = flow->head; | |
344 | flow->head = skb; | |
345 | } else { | |
346 | if (prev == flow->tail) | |
347 | flow->tail = skb; | |
348 | else | |
349 | skb->next = prev->next; | |
350 | prev->next = skb; | |
351 | } | |
352 | } | |
353 | ||
354 | static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch) | |
355 | { | |
356 | struct fq_sched_data *q = qdisc_priv(sch); | |
357 | struct fq_flow *f; | |
358 | ||
359 | if (unlikely(sch->q.qlen >= sch->limit)) | |
360 | return qdisc_drop(skb, sch); | |
361 | ||
362 | f = fq_classify(skb, q); | |
363 | if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { | |
364 | q->stat_flows_plimit++; | |
365 | return qdisc_drop(skb, sch); | |
366 | } | |
367 | ||
368 | f->qlen++; | |
369 | flow_queue_add(f, skb); | |
370 | if (skb_is_retransmit(skb)) | |
371 | q->stat_tcp_retrans++; | |
372 | sch->qstats.backlog += qdisc_pkt_len(skb); | |
373 | if (fq_flow_is_detached(f)) { | |
374 | fq_flow_add_tail(&q->new_flows, f); | |
375 | if (q->quantum > f->credit) | |
376 | f->credit = q->quantum; | |
377 | q->inactive_flows--; | |
378 | qdisc_unthrottled(sch); | |
379 | } | |
380 | if (unlikely(f == &q->internal)) { | |
381 | q->stat_internal_packets++; | |
382 | qdisc_unthrottled(sch); | |
383 | } | |
384 | sch->q.qlen++; | |
385 | ||
386 | return NET_XMIT_SUCCESS; | |
387 | } | |
388 | ||
389 | static void fq_check_throttled(struct fq_sched_data *q, u64 now) | |
390 | { | |
391 | struct rb_node *p; | |
392 | ||
393 | if (q->time_next_delayed_flow > now) | |
394 | return; | |
395 | ||
396 | q->time_next_delayed_flow = ~0ULL; | |
397 | while ((p = rb_first(&q->delayed)) != NULL) { | |
398 | struct fq_flow *f = container_of(p, struct fq_flow, rate_node); | |
399 | ||
400 | if (f->time_next_packet > now) { | |
401 | q->time_next_delayed_flow = f->time_next_packet; | |
402 | break; | |
403 | } | |
404 | rb_erase(p, &q->delayed); | |
405 | q->throttled_flows--; | |
406 | fq_flow_add_tail(&q->old_flows, f); | |
407 | } | |
408 | } | |
409 | ||
410 | static struct sk_buff *fq_dequeue(struct Qdisc *sch) | |
411 | { | |
412 | struct fq_sched_data *q = qdisc_priv(sch); | |
413 | u64 now = ktime_to_ns(ktime_get()); | |
414 | struct fq_flow_head *head; | |
415 | struct sk_buff *skb; | |
416 | struct fq_flow *f; | |
0eab5eb7 | 417 | u32 rate; |
afe4fd06 | 418 | |
8d34ce10 | 419 | skb = fq_dequeue_head(sch, &q->internal); |
afe4fd06 ED |
420 | if (skb) |
421 | goto out; | |
422 | fq_check_throttled(q, now); | |
423 | begin: | |
424 | head = &q->new_flows; | |
425 | if (!head->first) { | |
426 | head = &q->old_flows; | |
427 | if (!head->first) { | |
428 | if (q->time_next_delayed_flow != ~0ULL) | |
429 | qdisc_watchdog_schedule_ns(&q->watchdog, | |
430 | q->time_next_delayed_flow); | |
431 | return NULL; | |
432 | } | |
433 | } | |
434 | f = head->first; | |
435 | ||
436 | if (f->credit <= 0) { | |
437 | f->credit += q->quantum; | |
438 | head->first = f->next; | |
439 | fq_flow_add_tail(&q->old_flows, f); | |
440 | goto begin; | |
441 | } | |
442 | ||
443 | if (unlikely(f->head && now < f->time_next_packet)) { | |
444 | head->first = f->next; | |
445 | fq_flow_set_throttled(q, f); | |
446 | goto begin; | |
447 | } | |
448 | ||
8d34ce10 | 449 | skb = fq_dequeue_head(sch, f); |
afe4fd06 ED |
450 | if (!skb) { |
451 | head->first = f->next; | |
452 | /* force a pass through old_flows to prevent starvation */ | |
453 | if ((head == &q->new_flows) && q->old_flows.first) { | |
454 | fq_flow_add_tail(&q->old_flows, f); | |
455 | } else { | |
456 | fq_flow_set_detached(f); | |
457 | f->age = jiffies; | |
458 | q->inactive_flows++; | |
459 | } | |
460 | goto begin; | |
461 | } | |
08f89b98 | 462 | prefetch(&skb->end); |
afe4fd06 ED |
463 | f->time_next_packet = now; |
464 | f->credit -= qdisc_pkt_len(skb); | |
465 | ||
0eab5eb7 ED |
466 | if (f->credit > 0 || !q->rate_enable) |
467 | goto out; | |
afe4fd06 | 468 | |
7eec4174 ED |
469 | rate = q->flow_max_rate; |
470 | if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT) | |
471 | rate = min(skb->sk->sk_pacing_rate, rate); | |
afe4fd06 | 472 | |
7eec4174 | 473 | if (rate != ~0U) { |
0eab5eb7 ED |
474 | u32 plen = max(qdisc_pkt_len(skb), q->quantum); |
475 | u64 len = (u64)plen * NSEC_PER_SEC; | |
476 | ||
7eec4174 ED |
477 | if (likely(rate)) |
478 | do_div(len, rate); | |
0eab5eb7 ED |
479 | /* Since socket rate can change later, |
480 | * clamp the delay to 125 ms. | |
481 | * TODO: maybe segment the too big skb, as in commit | |
482 | * e43ac79a4bc ("sch_tbf: segment too big GSO packets") | |
483 | */ | |
484 | if (unlikely(len > 125 * NSEC_PER_MSEC)) { | |
485 | len = 125 * NSEC_PER_MSEC; | |
486 | q->stat_pkts_too_long++; | |
afe4fd06 | 487 | } |
0eab5eb7 ED |
488 | |
489 | f->time_next_packet = now + len; | |
afe4fd06 ED |
490 | } |
491 | out: | |
afe4fd06 | 492 | qdisc_bstats_update(sch, skb); |
afe4fd06 ED |
493 | qdisc_unthrottled(sch); |
494 | return skb; | |
495 | } | |
496 | ||
497 | static void fq_reset(struct Qdisc *sch) | |
498 | { | |
8d34ce10 ED |
499 | struct fq_sched_data *q = qdisc_priv(sch); |
500 | struct rb_root *root; | |
afe4fd06 | 501 | struct sk_buff *skb; |
8d34ce10 ED |
502 | struct rb_node *p; |
503 | struct fq_flow *f; | |
504 | unsigned int idx; | |
afe4fd06 | 505 | |
8d34ce10 | 506 | while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL) |
afe4fd06 | 507 | kfree_skb(skb); |
8d34ce10 ED |
508 | |
509 | if (!q->fq_root) | |
510 | return; | |
511 | ||
512 | for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { | |
513 | root = &q->fq_root[idx]; | |
514 | while ((p = rb_first(root)) != NULL) { | |
515 | f = container_of(p, struct fq_flow, fq_node); | |
516 | rb_erase(p, root); | |
517 | ||
518 | while ((skb = fq_dequeue_head(sch, f)) != NULL) | |
519 | kfree_skb(skb); | |
520 | ||
521 | kmem_cache_free(fq_flow_cachep, f); | |
522 | } | |
523 | } | |
524 | q->new_flows.first = NULL; | |
525 | q->old_flows.first = NULL; | |
526 | q->delayed = RB_ROOT; | |
527 | q->flows = 0; | |
528 | q->inactive_flows = 0; | |
529 | q->throttled_flows = 0; | |
afe4fd06 ED |
530 | } |
531 | ||
532 | static void fq_rehash(struct fq_sched_data *q, | |
533 | struct rb_root *old_array, u32 old_log, | |
534 | struct rb_root *new_array, u32 new_log) | |
535 | { | |
536 | struct rb_node *op, **np, *parent; | |
537 | struct rb_root *oroot, *nroot; | |
538 | struct fq_flow *of, *nf; | |
539 | int fcnt = 0; | |
540 | u32 idx; | |
541 | ||
542 | for (idx = 0; idx < (1U << old_log); idx++) { | |
543 | oroot = &old_array[idx]; | |
544 | while ((op = rb_first(oroot)) != NULL) { | |
545 | rb_erase(op, oroot); | |
546 | of = container_of(op, struct fq_flow, fq_node); | |
547 | if (fq_gc_candidate(of)) { | |
548 | fcnt++; | |
549 | kmem_cache_free(fq_flow_cachep, of); | |
550 | continue; | |
551 | } | |
552 | nroot = &new_array[hash_32((u32)(long)of->sk, new_log)]; | |
553 | ||
554 | np = &nroot->rb_node; | |
555 | parent = NULL; | |
556 | while (*np) { | |
557 | parent = *np; | |
558 | ||
559 | nf = container_of(parent, struct fq_flow, fq_node); | |
560 | BUG_ON(nf->sk == of->sk); | |
561 | ||
562 | if (nf->sk > of->sk) | |
563 | np = &parent->rb_right; | |
564 | else | |
565 | np = &parent->rb_left; | |
566 | } | |
567 | ||
568 | rb_link_node(&of->fq_node, parent, np); | |
569 | rb_insert_color(&of->fq_node, nroot); | |
570 | } | |
571 | } | |
572 | q->flows -= fcnt; | |
573 | q->inactive_flows -= fcnt; | |
574 | q->stat_gc_flows += fcnt; | |
575 | } | |
576 | ||
577 | static int fq_resize(struct fq_sched_data *q, u32 log) | |
578 | { | |
579 | struct rb_root *array; | |
580 | u32 idx; | |
581 | ||
582 | if (q->fq_root && log == q->fq_trees_log) | |
583 | return 0; | |
584 | ||
585 | array = kmalloc(sizeof(struct rb_root) << log, GFP_KERNEL); | |
586 | if (!array) | |
587 | return -ENOMEM; | |
588 | ||
589 | for (idx = 0; idx < (1U << log); idx++) | |
590 | array[idx] = RB_ROOT; | |
591 | ||
592 | if (q->fq_root) { | |
593 | fq_rehash(q, q->fq_root, q->fq_trees_log, array, log); | |
594 | kfree(q->fq_root); | |
595 | } | |
596 | q->fq_root = array; | |
597 | q->fq_trees_log = log; | |
598 | ||
599 | return 0; | |
600 | } | |
601 | ||
602 | static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { | |
603 | [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, | |
604 | [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, | |
605 | [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, | |
606 | [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, | |
607 | [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, | |
608 | [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, | |
609 | [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, | |
610 | [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, | |
611 | }; | |
612 | ||
613 | static int fq_change(struct Qdisc *sch, struct nlattr *opt) | |
614 | { | |
615 | struct fq_sched_data *q = qdisc_priv(sch); | |
616 | struct nlattr *tb[TCA_FQ_MAX + 1]; | |
617 | int err, drop_count = 0; | |
618 | u32 fq_log; | |
619 | ||
620 | if (!opt) | |
621 | return -EINVAL; | |
622 | ||
623 | err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy); | |
624 | if (err < 0) | |
625 | return err; | |
626 | ||
627 | sch_tree_lock(sch); | |
628 | ||
629 | fq_log = q->fq_trees_log; | |
630 | ||
631 | if (tb[TCA_FQ_BUCKETS_LOG]) { | |
632 | u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); | |
633 | ||
634 | if (nval >= 1 && nval <= ilog2(256*1024)) | |
635 | fq_log = nval; | |
636 | else | |
637 | err = -EINVAL; | |
638 | } | |
639 | if (tb[TCA_FQ_PLIMIT]) | |
640 | sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); | |
641 | ||
642 | if (tb[TCA_FQ_FLOW_PLIMIT]) | |
643 | q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); | |
644 | ||
645 | if (tb[TCA_FQ_QUANTUM]) | |
646 | q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); | |
647 | ||
648 | if (tb[TCA_FQ_INITIAL_QUANTUM]) | |
ede869cd | 649 | q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); |
afe4fd06 ED |
650 | |
651 | if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) | |
65c5189a ED |
652 | pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", |
653 | nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); | |
afe4fd06 ED |
654 | |
655 | if (tb[TCA_FQ_FLOW_MAX_RATE]) | |
656 | q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); | |
657 | ||
658 | if (tb[TCA_FQ_RATE_ENABLE]) { | |
659 | u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); | |
660 | ||
661 | if (enable <= 1) | |
662 | q->rate_enable = enable; | |
663 | else | |
664 | err = -EINVAL; | |
665 | } | |
666 | ||
667 | if (!err) | |
668 | err = fq_resize(q, fq_log); | |
669 | ||
670 | while (sch->q.qlen > sch->limit) { | |
671 | struct sk_buff *skb = fq_dequeue(sch); | |
672 | ||
8d34ce10 ED |
673 | if (!skb) |
674 | break; | |
afe4fd06 ED |
675 | kfree_skb(skb); |
676 | drop_count++; | |
677 | } | |
678 | qdisc_tree_decrease_qlen(sch, drop_count); | |
679 | ||
680 | sch_tree_unlock(sch); | |
681 | return err; | |
682 | } | |
683 | ||
684 | static void fq_destroy(struct Qdisc *sch) | |
685 | { | |
686 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 687 | |
8d34ce10 ED |
688 | fq_reset(sch); |
689 | kfree(q->fq_root); | |
afe4fd06 ED |
690 | qdisc_watchdog_cancel(&q->watchdog); |
691 | } | |
692 | ||
693 | static int fq_init(struct Qdisc *sch, struct nlattr *opt) | |
694 | { | |
695 | struct fq_sched_data *q = qdisc_priv(sch); | |
696 | int err; | |
697 | ||
698 | sch->limit = 10000; | |
699 | q->flow_plimit = 100; | |
700 | q->quantum = 2 * psched_mtu(qdisc_dev(sch)); | |
701 | q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); | |
afe4fd06 ED |
702 | q->flow_max_rate = ~0U; |
703 | q->rate_enable = 1; | |
704 | q->new_flows.first = NULL; | |
705 | q->old_flows.first = NULL; | |
706 | q->delayed = RB_ROOT; | |
707 | q->fq_root = NULL; | |
708 | q->fq_trees_log = ilog2(1024); | |
709 | qdisc_watchdog_init(&q->watchdog, sch); | |
710 | ||
711 | if (opt) | |
712 | err = fq_change(sch, opt); | |
713 | else | |
714 | err = fq_resize(q, q->fq_trees_log); | |
715 | ||
716 | return err; | |
717 | } | |
718 | ||
719 | static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) | |
720 | { | |
721 | struct fq_sched_data *q = qdisc_priv(sch); | |
722 | struct nlattr *opts; | |
723 | ||
724 | opts = nla_nest_start(skb, TCA_OPTIONS); | |
725 | if (opts == NULL) | |
726 | goto nla_put_failure; | |
727 | ||
65c5189a ED |
728 | /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ |
729 | ||
afe4fd06 ED |
730 | if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || |
731 | nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || | |
732 | nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || | |
733 | nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || | |
734 | nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || | |
afe4fd06 ED |
735 | nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) || |
736 | nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) | |
737 | goto nla_put_failure; | |
738 | ||
739 | nla_nest_end(skb, opts); | |
740 | return skb->len; | |
741 | ||
742 | nla_put_failure: | |
743 | return -1; | |
744 | } | |
745 | ||
746 | static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) | |
747 | { | |
748 | struct fq_sched_data *q = qdisc_priv(sch); | |
749 | u64 now = ktime_to_ns(ktime_get()); | |
750 | struct tc_fq_qd_stats st = { | |
751 | .gc_flows = q->stat_gc_flows, | |
752 | .highprio_packets = q->stat_internal_packets, | |
753 | .tcp_retrans = q->stat_tcp_retrans, | |
754 | .throttled = q->stat_throttled, | |
755 | .flows_plimit = q->stat_flows_plimit, | |
756 | .pkts_too_long = q->stat_pkts_too_long, | |
757 | .allocation_errors = q->stat_allocation_errors, | |
758 | .flows = q->flows, | |
759 | .inactive_flows = q->inactive_flows, | |
760 | .throttled_flows = q->throttled_flows, | |
761 | .time_next_delayed_flow = q->time_next_delayed_flow - now, | |
762 | }; | |
763 | ||
764 | return gnet_stats_copy_app(d, &st, sizeof(st)); | |
765 | } | |
766 | ||
767 | static struct Qdisc_ops fq_qdisc_ops __read_mostly = { | |
768 | .id = "fq", | |
769 | .priv_size = sizeof(struct fq_sched_data), | |
770 | ||
771 | .enqueue = fq_enqueue, | |
772 | .dequeue = fq_dequeue, | |
773 | .peek = qdisc_peek_dequeued, | |
774 | .init = fq_init, | |
775 | .reset = fq_reset, | |
776 | .destroy = fq_destroy, | |
777 | .change = fq_change, | |
778 | .dump = fq_dump, | |
779 | .dump_stats = fq_dump_stats, | |
780 | .owner = THIS_MODULE, | |
781 | }; | |
782 | ||
783 | static int __init fq_module_init(void) | |
784 | { | |
785 | int ret; | |
786 | ||
787 | fq_flow_cachep = kmem_cache_create("fq_flow_cache", | |
788 | sizeof(struct fq_flow), | |
789 | 0, 0, NULL); | |
790 | if (!fq_flow_cachep) | |
791 | return -ENOMEM; | |
792 | ||
793 | ret = register_qdisc(&fq_qdisc_ops); | |
794 | if (ret) | |
795 | kmem_cache_destroy(fq_flow_cachep); | |
796 | return ret; | |
797 | } | |
798 | ||
799 | static void __exit fq_module_exit(void) | |
800 | { | |
801 | unregister_qdisc(&fq_qdisc_ops); | |
802 | kmem_cache_destroy(fq_flow_cachep); | |
803 | } | |
804 | ||
805 | module_init(fq_module_init) | |
806 | module_exit(fq_module_exit) | |
807 | MODULE_AUTHOR("Eric Dumazet"); | |
808 | MODULE_LICENSE("GPL"); |