]>
Commit | Line | Data |
---|---|---|
afe4fd06 ED |
1 | /* |
2 | * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing) | |
3 | * | |
86b3bfe9 | 4 | * Copyright (C) 2013-2015 Eric Dumazet <edumazet@google.com> |
afe4fd06 ED |
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 | * | |
05e8bb86 | 11 | * Meant to be mostly used for locally generated traffic : |
afe4fd06 ED |
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> |
c3bd8549 | 50 | #include <linux/vmalloc.h> |
afe4fd06 ED |
51 | #include <net/netlink.h> |
52 | #include <net/pkt_sched.h> | |
53 | #include <net/sock.h> | |
54 | #include <net/tcp_states.h> | |
98781965 | 55 | #include <net/tcp.h> |
afe4fd06 ED |
56 | |
57 | /* | |
58 | * Per flow structure, dynamically allocated | |
59 | */ | |
60 | struct fq_flow { | |
61 | struct sk_buff *head; /* list of skbs for this flow : first skb */ | |
62 | union { | |
63 | struct sk_buff *tail; /* last skb in the list */ | |
64 | unsigned long age; /* jiffies when flow was emptied, for gc */ | |
65 | }; | |
05e8bb86 | 66 | struct rb_node fq_node; /* anchor in fq_root[] trees */ |
afe4fd06 ED |
67 | struct sock *sk; |
68 | int qlen; /* number of packets in flow queue */ | |
69 | int credit; | |
70 | u32 socket_hash; /* sk_hash */ | |
71 | struct fq_flow *next; /* next pointer in RR lists, or &detached */ | |
72 | ||
73 | struct rb_node rate_node; /* anchor in q->delayed tree */ | |
74 | u64 time_next_packet; | |
75 | }; | |
76 | ||
77 | struct fq_flow_head { | |
78 | struct fq_flow *first; | |
79 | struct fq_flow *last; | |
80 | }; | |
81 | ||
82 | struct fq_sched_data { | |
83 | struct fq_flow_head new_flows; | |
84 | ||
85 | struct fq_flow_head old_flows; | |
86 | ||
87 | struct rb_root delayed; /* for rate limited flows */ | |
88 | u64 time_next_delayed_flow; | |
89 | ||
90 | struct fq_flow internal; /* for non classified or high prio packets */ | |
91 | u32 quantum; | |
92 | u32 initial_quantum; | |
f52ed899 | 93 | u32 flow_refill_delay; |
afe4fd06 ED |
94 | u32 flow_max_rate; /* optional max rate per flow */ |
95 | u32 flow_plimit; /* max packets per flow */ | |
06eb395f | 96 | u32 orphan_mask; /* mask for orphaned skb */ |
afe4fd06 ED |
97 | struct rb_root *fq_root; |
98 | u8 rate_enable; | |
99 | u8 fq_trees_log; | |
100 | ||
101 | u32 flows; | |
102 | u32 inactive_flows; | |
103 | u32 throttled_flows; | |
104 | ||
105 | u64 stat_gc_flows; | |
106 | u64 stat_internal_packets; | |
107 | u64 stat_tcp_retrans; | |
108 | u64 stat_throttled; | |
109 | u64 stat_flows_plimit; | |
110 | u64 stat_pkts_too_long; | |
111 | u64 stat_allocation_errors; | |
112 | struct qdisc_watchdog watchdog; | |
113 | }; | |
114 | ||
115 | /* special value to mark a detached flow (not on old/new list) */ | |
116 | static struct fq_flow detached, throttled; | |
117 | ||
118 | static void fq_flow_set_detached(struct fq_flow *f) | |
119 | { | |
120 | f->next = &detached; | |
f52ed899 | 121 | f->age = jiffies; |
afe4fd06 ED |
122 | } |
123 | ||
124 | static bool fq_flow_is_detached(const struct fq_flow *f) | |
125 | { | |
126 | return f->next == &detached; | |
127 | } | |
128 | ||
129 | static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) | |
130 | { | |
131 | struct rb_node **p = &q->delayed.rb_node, *parent = NULL; | |
132 | ||
133 | while (*p) { | |
134 | struct fq_flow *aux; | |
135 | ||
136 | parent = *p; | |
137 | aux = container_of(parent, struct fq_flow, rate_node); | |
138 | if (f->time_next_packet >= aux->time_next_packet) | |
139 | p = &parent->rb_right; | |
140 | else | |
141 | p = &parent->rb_left; | |
142 | } | |
143 | rb_link_node(&f->rate_node, parent, p); | |
144 | rb_insert_color(&f->rate_node, &q->delayed); | |
145 | q->throttled_flows++; | |
146 | q->stat_throttled++; | |
147 | ||
148 | f->next = &throttled; | |
149 | if (q->time_next_delayed_flow > f->time_next_packet) | |
150 | q->time_next_delayed_flow = f->time_next_packet; | |
151 | } | |
152 | ||
153 | ||
154 | static struct kmem_cache *fq_flow_cachep __read_mostly; | |
155 | ||
156 | static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) | |
157 | { | |
158 | if (head->first) | |
159 | head->last->next = flow; | |
160 | else | |
161 | head->first = flow; | |
162 | head->last = flow; | |
163 | flow->next = NULL; | |
164 | } | |
165 | ||
166 | /* limit number of collected flows per round */ | |
167 | #define FQ_GC_MAX 8 | |
168 | #define FQ_GC_AGE (3*HZ) | |
169 | ||
170 | static bool fq_gc_candidate(const struct fq_flow *f) | |
171 | { | |
172 | return fq_flow_is_detached(f) && | |
173 | time_after(jiffies, f->age + FQ_GC_AGE); | |
174 | } | |
175 | ||
176 | static void fq_gc(struct fq_sched_data *q, | |
177 | struct rb_root *root, | |
178 | struct sock *sk) | |
179 | { | |
180 | struct fq_flow *f, *tofree[FQ_GC_MAX]; | |
181 | struct rb_node **p, *parent; | |
182 | int fcnt = 0; | |
183 | ||
184 | p = &root->rb_node; | |
185 | parent = NULL; | |
186 | while (*p) { | |
187 | parent = *p; | |
188 | ||
189 | f = container_of(parent, struct fq_flow, fq_node); | |
190 | if (f->sk == sk) | |
191 | break; | |
192 | ||
193 | if (fq_gc_candidate(f)) { | |
194 | tofree[fcnt++] = f; | |
195 | if (fcnt == FQ_GC_MAX) | |
196 | break; | |
197 | } | |
198 | ||
199 | if (f->sk > sk) | |
200 | p = &parent->rb_right; | |
201 | else | |
202 | p = &parent->rb_left; | |
203 | } | |
204 | ||
205 | q->flows -= fcnt; | |
206 | q->inactive_flows -= fcnt; | |
207 | q->stat_gc_flows += fcnt; | |
208 | while (fcnt) { | |
209 | struct fq_flow *f = tofree[--fcnt]; | |
210 | ||
211 | rb_erase(&f->fq_node, root); | |
212 | kmem_cache_free(fq_flow_cachep, f); | |
213 | } | |
214 | } | |
215 | ||
afe4fd06 ED |
216 | static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) |
217 | { | |
218 | struct rb_node **p, *parent; | |
219 | struct sock *sk = skb->sk; | |
220 | struct rb_root *root; | |
221 | struct fq_flow *f; | |
afe4fd06 ED |
222 | |
223 | /* warning: no starvation prevention... */ | |
2abc2f07 | 224 | if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) |
afe4fd06 ED |
225 | return &q->internal; |
226 | ||
ca6fb065 | 227 | /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket |
e446f9df | 228 | * or a listener (SYNCOOKIE mode) |
ca6fb065 ED |
229 | * 1) request sockets are not full blown, |
230 | * they do not contain sk_pacing_rate | |
231 | * 2) They are not part of a 'flow' yet | |
232 | * 3) We do not want to rate limit them (eg SYNFLOOD attack), | |
06eb395f | 233 | * especially if the listener set SO_MAX_PACING_RATE |
ca6fb065 | 234 | * 4) We pretend they are orphaned |
06eb395f | 235 | */ |
e446f9df | 236 | if (!sk || sk_listener(sk)) { |
06eb395f ED |
237 | unsigned long hash = skb_get_hash(skb) & q->orphan_mask; |
238 | ||
afe4fd06 ED |
239 | /* By forcing low order bit to 1, we make sure to not |
240 | * collide with a local flow (socket pointers are word aligned) | |
241 | */ | |
06eb395f ED |
242 | sk = (struct sock *)((hash << 1) | 1UL); |
243 | skb_orphan(skb); | |
afe4fd06 ED |
244 | } |
245 | ||
246 | root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)]; | |
247 | ||
248 | if (q->flows >= (2U << q->fq_trees_log) && | |
249 | q->inactive_flows > q->flows/2) | |
250 | fq_gc(q, root, sk); | |
251 | ||
252 | p = &root->rb_node; | |
253 | parent = NULL; | |
254 | while (*p) { | |
255 | parent = *p; | |
256 | ||
257 | f = container_of(parent, struct fq_flow, fq_node); | |
258 | if (f->sk == sk) { | |
259 | /* socket might have been reallocated, so check | |
260 | * if its sk_hash is the same. | |
261 | * It not, we need to refill credit with | |
262 | * initial quantum | |
263 | */ | |
264 | if (unlikely(skb->sk && | |
265 | f->socket_hash != sk->sk_hash)) { | |
266 | f->credit = q->initial_quantum; | |
267 | f->socket_hash = sk->sk_hash; | |
fc59d5bd | 268 | f->time_next_packet = 0ULL; |
afe4fd06 ED |
269 | } |
270 | return f; | |
271 | } | |
272 | if (f->sk > sk) | |
273 | p = &parent->rb_right; | |
274 | else | |
275 | p = &parent->rb_left; | |
276 | } | |
277 | ||
278 | f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); | |
279 | if (unlikely(!f)) { | |
280 | q->stat_allocation_errors++; | |
281 | return &q->internal; | |
282 | } | |
283 | fq_flow_set_detached(f); | |
284 | f->sk = sk; | |
285 | if (skb->sk) | |
286 | f->socket_hash = sk->sk_hash; | |
287 | f->credit = q->initial_quantum; | |
288 | ||
289 | rb_link_node(&f->fq_node, parent, p); | |
290 | rb_insert_color(&f->fq_node, root); | |
291 | ||
292 | q->flows++; | |
293 | q->inactive_flows++; | |
294 | return f; | |
295 | } | |
296 | ||
297 | ||
298 | /* remove one skb from head of flow queue */ | |
8d34ce10 | 299 | static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow) |
afe4fd06 ED |
300 | { |
301 | struct sk_buff *skb = flow->head; | |
302 | ||
303 | if (skb) { | |
304 | flow->head = skb->next; | |
305 | skb->next = NULL; | |
306 | flow->qlen--; | |
25331d6c | 307 | qdisc_qstats_backlog_dec(sch, skb); |
8d34ce10 | 308 | sch->q.qlen--; |
afe4fd06 ED |
309 | } |
310 | return skb; | |
311 | } | |
312 | ||
313 | /* We might add in the future detection of retransmits | |
314 | * For the time being, just return false | |
315 | */ | |
316 | static bool skb_is_retransmit(struct sk_buff *skb) | |
317 | { | |
318 | return false; | |
319 | } | |
320 | ||
321 | /* add skb to flow queue | |
322 | * flow queue is a linked list, kind of FIFO, except for TCP retransmits | |
323 | * We special case tcp retransmits to be transmitted before other packets. | |
324 | * We rely on fact that TCP retransmits are unlikely, so we do not waste | |
325 | * a separate queue or a pointer. | |
326 | * head-> [retrans pkt 1] | |
327 | * [retrans pkt 2] | |
328 | * [ normal pkt 1] | |
329 | * [ normal pkt 2] | |
330 | * [ normal pkt 3] | |
331 | * tail-> [ normal pkt 4] | |
332 | */ | |
333 | static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) | |
334 | { | |
335 | struct sk_buff *prev, *head = flow->head; | |
336 | ||
337 | skb->next = NULL; | |
338 | if (!head) { | |
339 | flow->head = skb; | |
340 | flow->tail = skb; | |
341 | return; | |
342 | } | |
343 | if (likely(!skb_is_retransmit(skb))) { | |
344 | flow->tail->next = skb; | |
345 | flow->tail = skb; | |
346 | return; | |
347 | } | |
348 | ||
349 | /* This skb is a tcp retransmit, | |
350 | * find the last retrans packet in the queue | |
351 | */ | |
352 | prev = NULL; | |
353 | while (skb_is_retransmit(head)) { | |
354 | prev = head; | |
355 | head = head->next; | |
356 | if (!head) | |
357 | break; | |
358 | } | |
359 | if (!prev) { /* no rtx packet in queue, become the new head */ | |
360 | skb->next = flow->head; | |
361 | flow->head = skb; | |
362 | } else { | |
363 | if (prev == flow->tail) | |
364 | flow->tail = skb; | |
365 | else | |
366 | skb->next = prev->next; | |
367 | prev->next = skb; | |
368 | } | |
369 | } | |
370 | ||
371 | static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch) | |
372 | { | |
373 | struct fq_sched_data *q = qdisc_priv(sch); | |
374 | struct fq_flow *f; | |
375 | ||
376 | if (unlikely(sch->q.qlen >= sch->limit)) | |
377 | return qdisc_drop(skb, sch); | |
378 | ||
379 | f = fq_classify(skb, q); | |
380 | if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { | |
381 | q->stat_flows_plimit++; | |
382 | return qdisc_drop(skb, sch); | |
383 | } | |
384 | ||
385 | f->qlen++; | |
afe4fd06 ED |
386 | if (skb_is_retransmit(skb)) |
387 | q->stat_tcp_retrans++; | |
25331d6c | 388 | qdisc_qstats_backlog_inc(sch, skb); |
afe4fd06 ED |
389 | if (fq_flow_is_detached(f)) { |
390 | fq_flow_add_tail(&q->new_flows, f); | |
f52ed899 ED |
391 | if (time_after(jiffies, f->age + q->flow_refill_delay)) |
392 | f->credit = max_t(u32, f->credit, q->quantum); | |
afe4fd06 | 393 | q->inactive_flows--; |
afe4fd06 | 394 | } |
f52ed899 ED |
395 | |
396 | /* Note: this overwrites f->age */ | |
397 | flow_queue_add(f, skb); | |
398 | ||
afe4fd06 ED |
399 | if (unlikely(f == &q->internal)) { |
400 | q->stat_internal_packets++; | |
afe4fd06 ED |
401 | } |
402 | sch->q.qlen++; | |
403 | ||
404 | return NET_XMIT_SUCCESS; | |
405 | } | |
406 | ||
407 | static void fq_check_throttled(struct fq_sched_data *q, u64 now) | |
408 | { | |
409 | struct rb_node *p; | |
410 | ||
411 | if (q->time_next_delayed_flow > now) | |
412 | return; | |
413 | ||
414 | q->time_next_delayed_flow = ~0ULL; | |
415 | while ((p = rb_first(&q->delayed)) != NULL) { | |
416 | struct fq_flow *f = container_of(p, struct fq_flow, rate_node); | |
417 | ||
418 | if (f->time_next_packet > now) { | |
419 | q->time_next_delayed_flow = f->time_next_packet; | |
420 | break; | |
421 | } | |
422 | rb_erase(p, &q->delayed); | |
423 | q->throttled_flows--; | |
424 | fq_flow_add_tail(&q->old_flows, f); | |
425 | } | |
426 | } | |
427 | ||
428 | static struct sk_buff *fq_dequeue(struct Qdisc *sch) | |
429 | { | |
430 | struct fq_sched_data *q = qdisc_priv(sch); | |
d2de875c | 431 | u64 now = ktime_get_ns(); |
afe4fd06 ED |
432 | struct fq_flow_head *head; |
433 | struct sk_buff *skb; | |
434 | struct fq_flow *f; | |
0eab5eb7 | 435 | u32 rate; |
afe4fd06 | 436 | |
8d34ce10 | 437 | skb = fq_dequeue_head(sch, &q->internal); |
afe4fd06 ED |
438 | if (skb) |
439 | goto out; | |
440 | fq_check_throttled(q, now); | |
441 | begin: | |
442 | head = &q->new_flows; | |
443 | if (!head->first) { | |
444 | head = &q->old_flows; | |
445 | if (!head->first) { | |
446 | if (q->time_next_delayed_flow != ~0ULL) | |
447 | qdisc_watchdog_schedule_ns(&q->watchdog, | |
45f50bed | 448 | q->time_next_delayed_flow); |
afe4fd06 ED |
449 | return NULL; |
450 | } | |
451 | } | |
452 | f = head->first; | |
453 | ||
454 | if (f->credit <= 0) { | |
455 | f->credit += q->quantum; | |
456 | head->first = f->next; | |
457 | fq_flow_add_tail(&q->old_flows, f); | |
458 | goto begin; | |
459 | } | |
460 | ||
98781965 ED |
461 | skb = f->head; |
462 | if (unlikely(skb && now < f->time_next_packet && | |
463 | !skb_is_tcp_pure_ack(skb))) { | |
afe4fd06 ED |
464 | head->first = f->next; |
465 | fq_flow_set_throttled(q, f); | |
466 | goto begin; | |
467 | } | |
468 | ||
8d34ce10 | 469 | skb = fq_dequeue_head(sch, f); |
afe4fd06 ED |
470 | if (!skb) { |
471 | head->first = f->next; | |
472 | /* force a pass through old_flows to prevent starvation */ | |
473 | if ((head == &q->new_flows) && q->old_flows.first) { | |
474 | fq_flow_add_tail(&q->old_flows, f); | |
475 | } else { | |
476 | fq_flow_set_detached(f); | |
afe4fd06 ED |
477 | q->inactive_flows++; |
478 | } | |
479 | goto begin; | |
480 | } | |
08f89b98 | 481 | prefetch(&skb->end); |
afe4fd06 ED |
482 | f->credit -= qdisc_pkt_len(skb); |
483 | ||
0eab5eb7 ED |
484 | if (f->credit > 0 || !q->rate_enable) |
485 | goto out; | |
afe4fd06 | 486 | |
98781965 ED |
487 | /* Do not pace locally generated ack packets */ |
488 | if (skb_is_tcp_pure_ack(skb)) | |
489 | goto out; | |
490 | ||
7eec4174 | 491 | rate = q->flow_max_rate; |
86b3bfe9 | 492 | if (skb->sk) |
7eec4174 | 493 | rate = min(skb->sk->sk_pacing_rate, rate); |
afe4fd06 | 494 | |
7eec4174 | 495 | if (rate != ~0U) { |
0eab5eb7 ED |
496 | u32 plen = max(qdisc_pkt_len(skb), q->quantum); |
497 | u64 len = (u64)plen * NSEC_PER_SEC; | |
498 | ||
7eec4174 ED |
499 | if (likely(rate)) |
500 | do_div(len, rate); | |
0eab5eb7 | 501 | /* Since socket rate can change later, |
ced7a04e ED |
502 | * clamp the delay to 1 second. |
503 | * Really, providers of too big packets should be fixed ! | |
0eab5eb7 | 504 | */ |
ced7a04e ED |
505 | if (unlikely(len > NSEC_PER_SEC)) { |
506 | len = NSEC_PER_SEC; | |
0eab5eb7 | 507 | q->stat_pkts_too_long++; |
afe4fd06 | 508 | } |
0eab5eb7 ED |
509 | |
510 | f->time_next_packet = now + len; | |
afe4fd06 ED |
511 | } |
512 | out: | |
afe4fd06 | 513 | qdisc_bstats_update(sch, skb); |
afe4fd06 ED |
514 | return skb; |
515 | } | |
516 | ||
e14ffdfd ED |
517 | static void fq_flow_purge(struct fq_flow *flow) |
518 | { | |
519 | rtnl_kfree_skbs(flow->head, flow->tail); | |
520 | flow->head = NULL; | |
521 | flow->qlen = 0; | |
522 | } | |
523 | ||
afe4fd06 ED |
524 | static void fq_reset(struct Qdisc *sch) |
525 | { | |
8d34ce10 ED |
526 | struct fq_sched_data *q = qdisc_priv(sch); |
527 | struct rb_root *root; | |
8d34ce10 ED |
528 | struct rb_node *p; |
529 | struct fq_flow *f; | |
530 | unsigned int idx; | |
afe4fd06 | 531 | |
e14ffdfd ED |
532 | sch->q.qlen = 0; |
533 | sch->qstats.backlog = 0; | |
534 | ||
535 | fq_flow_purge(&q->internal); | |
8d34ce10 ED |
536 | |
537 | if (!q->fq_root) | |
538 | return; | |
539 | ||
540 | for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { | |
541 | root = &q->fq_root[idx]; | |
542 | while ((p = rb_first(root)) != NULL) { | |
543 | f = container_of(p, struct fq_flow, fq_node); | |
544 | rb_erase(p, root); | |
545 | ||
e14ffdfd | 546 | fq_flow_purge(f); |
8d34ce10 ED |
547 | |
548 | kmem_cache_free(fq_flow_cachep, f); | |
549 | } | |
550 | } | |
551 | q->new_flows.first = NULL; | |
552 | q->old_flows.first = NULL; | |
553 | q->delayed = RB_ROOT; | |
554 | q->flows = 0; | |
555 | q->inactive_flows = 0; | |
556 | q->throttled_flows = 0; | |
afe4fd06 ED |
557 | } |
558 | ||
559 | static void fq_rehash(struct fq_sched_data *q, | |
560 | struct rb_root *old_array, u32 old_log, | |
561 | struct rb_root *new_array, u32 new_log) | |
562 | { | |
563 | struct rb_node *op, **np, *parent; | |
564 | struct rb_root *oroot, *nroot; | |
565 | struct fq_flow *of, *nf; | |
566 | int fcnt = 0; | |
567 | u32 idx; | |
568 | ||
569 | for (idx = 0; idx < (1U << old_log); idx++) { | |
570 | oroot = &old_array[idx]; | |
571 | while ((op = rb_first(oroot)) != NULL) { | |
572 | rb_erase(op, oroot); | |
573 | of = container_of(op, struct fq_flow, fq_node); | |
574 | if (fq_gc_candidate(of)) { | |
575 | fcnt++; | |
576 | kmem_cache_free(fq_flow_cachep, of); | |
577 | continue; | |
578 | } | |
579 | nroot = &new_array[hash_32((u32)(long)of->sk, new_log)]; | |
580 | ||
581 | np = &nroot->rb_node; | |
582 | parent = NULL; | |
583 | while (*np) { | |
584 | parent = *np; | |
585 | ||
586 | nf = container_of(parent, struct fq_flow, fq_node); | |
587 | BUG_ON(nf->sk == of->sk); | |
588 | ||
589 | if (nf->sk > of->sk) | |
590 | np = &parent->rb_right; | |
591 | else | |
592 | np = &parent->rb_left; | |
593 | } | |
594 | ||
595 | rb_link_node(&of->fq_node, parent, np); | |
596 | rb_insert_color(&of->fq_node, nroot); | |
597 | } | |
598 | } | |
599 | q->flows -= fcnt; | |
600 | q->inactive_flows -= fcnt; | |
601 | q->stat_gc_flows += fcnt; | |
602 | } | |
603 | ||
c3bd8549 | 604 | static void *fq_alloc_node(size_t sz, int node) |
afe4fd06 | 605 | { |
c3bd8549 ED |
606 | void *ptr; |
607 | ||
608 | ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node); | |
609 | if (!ptr) | |
610 | ptr = vmalloc_node(sz, node); | |
611 | return ptr; | |
612 | } | |
613 | ||
614 | static void fq_free(void *addr) | |
615 | { | |
4cb28970 | 616 | kvfree(addr); |
c3bd8549 ED |
617 | } |
618 | ||
619 | static int fq_resize(struct Qdisc *sch, u32 log) | |
620 | { | |
621 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 622 | struct rb_root *array; |
2d8d40af | 623 | void *old_fq_root; |
afe4fd06 ED |
624 | u32 idx; |
625 | ||
626 | if (q->fq_root && log == q->fq_trees_log) | |
627 | return 0; | |
628 | ||
c3bd8549 ED |
629 | /* If XPS was setup, we can allocate memory on right NUMA node */ |
630 | array = fq_alloc_node(sizeof(struct rb_root) << log, | |
631 | netdev_queue_numa_node_read(sch->dev_queue)); | |
afe4fd06 ED |
632 | if (!array) |
633 | return -ENOMEM; | |
634 | ||
635 | for (idx = 0; idx < (1U << log); idx++) | |
636 | array[idx] = RB_ROOT; | |
637 | ||
2d8d40af ED |
638 | sch_tree_lock(sch); |
639 | ||
640 | old_fq_root = q->fq_root; | |
641 | if (old_fq_root) | |
642 | fq_rehash(q, old_fq_root, q->fq_trees_log, array, log); | |
643 | ||
afe4fd06 ED |
644 | q->fq_root = array; |
645 | q->fq_trees_log = log; | |
646 | ||
2d8d40af ED |
647 | sch_tree_unlock(sch); |
648 | ||
649 | fq_free(old_fq_root); | |
650 | ||
afe4fd06 ED |
651 | return 0; |
652 | } | |
653 | ||
654 | static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { | |
655 | [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, | |
656 | [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, | |
657 | [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, | |
658 | [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, | |
659 | [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, | |
660 | [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, | |
661 | [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, | |
662 | [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, | |
f52ed899 | 663 | [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, |
afe4fd06 ED |
664 | }; |
665 | ||
666 | static int fq_change(struct Qdisc *sch, struct nlattr *opt) | |
667 | { | |
668 | struct fq_sched_data *q = qdisc_priv(sch); | |
669 | struct nlattr *tb[TCA_FQ_MAX + 1]; | |
670 | int err, drop_count = 0; | |
2ccccf5f | 671 | unsigned drop_len = 0; |
afe4fd06 ED |
672 | u32 fq_log; |
673 | ||
674 | if (!opt) | |
675 | return -EINVAL; | |
676 | ||
677 | err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy); | |
678 | if (err < 0) | |
679 | return err; | |
680 | ||
681 | sch_tree_lock(sch); | |
682 | ||
683 | fq_log = q->fq_trees_log; | |
684 | ||
685 | if (tb[TCA_FQ_BUCKETS_LOG]) { | |
686 | u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); | |
687 | ||
688 | if (nval >= 1 && nval <= ilog2(256*1024)) | |
689 | fq_log = nval; | |
690 | else | |
691 | err = -EINVAL; | |
692 | } | |
693 | if (tb[TCA_FQ_PLIMIT]) | |
694 | sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); | |
695 | ||
696 | if (tb[TCA_FQ_FLOW_PLIMIT]) | |
697 | q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); | |
698 | ||
3725a269 KKJ |
699 | if (tb[TCA_FQ_QUANTUM]) { |
700 | u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); | |
701 | ||
702 | if (quantum > 0) | |
703 | q->quantum = quantum; | |
704 | else | |
705 | err = -EINVAL; | |
706 | } | |
afe4fd06 ED |
707 | |
708 | if (tb[TCA_FQ_INITIAL_QUANTUM]) | |
ede869cd | 709 | q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); |
afe4fd06 ED |
710 | |
711 | if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) | |
65c5189a ED |
712 | pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", |
713 | nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); | |
afe4fd06 ED |
714 | |
715 | if (tb[TCA_FQ_FLOW_MAX_RATE]) | |
716 | q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); | |
717 | ||
718 | if (tb[TCA_FQ_RATE_ENABLE]) { | |
719 | u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); | |
720 | ||
721 | if (enable <= 1) | |
722 | q->rate_enable = enable; | |
723 | else | |
724 | err = -EINVAL; | |
725 | } | |
726 | ||
f52ed899 ED |
727 | if (tb[TCA_FQ_FLOW_REFILL_DELAY]) { |
728 | u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ; | |
729 | ||
730 | q->flow_refill_delay = usecs_to_jiffies(usecs_delay); | |
731 | } | |
732 | ||
06eb395f ED |
733 | if (tb[TCA_FQ_ORPHAN_MASK]) |
734 | q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]); | |
735 | ||
2d8d40af ED |
736 | if (!err) { |
737 | sch_tree_unlock(sch); | |
c3bd8549 | 738 | err = fq_resize(sch, fq_log); |
2d8d40af ED |
739 | sch_tree_lock(sch); |
740 | } | |
afe4fd06 ED |
741 | while (sch->q.qlen > sch->limit) { |
742 | struct sk_buff *skb = fq_dequeue(sch); | |
743 | ||
8d34ce10 ED |
744 | if (!skb) |
745 | break; | |
2ccccf5f | 746 | drop_len += qdisc_pkt_len(skb); |
e14ffdfd | 747 | rtnl_kfree_skbs(skb, skb); |
afe4fd06 ED |
748 | drop_count++; |
749 | } | |
2ccccf5f | 750 | qdisc_tree_reduce_backlog(sch, drop_count, drop_len); |
afe4fd06 ED |
751 | |
752 | sch_tree_unlock(sch); | |
753 | return err; | |
754 | } | |
755 | ||
756 | static void fq_destroy(struct Qdisc *sch) | |
757 | { | |
758 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 759 | |
8d34ce10 | 760 | fq_reset(sch); |
c3bd8549 | 761 | fq_free(q->fq_root); |
afe4fd06 ED |
762 | qdisc_watchdog_cancel(&q->watchdog); |
763 | } | |
764 | ||
765 | static int fq_init(struct Qdisc *sch, struct nlattr *opt) | |
766 | { | |
767 | struct fq_sched_data *q = qdisc_priv(sch); | |
768 | int err; | |
769 | ||
770 | sch->limit = 10000; | |
771 | q->flow_plimit = 100; | |
772 | q->quantum = 2 * psched_mtu(qdisc_dev(sch)); | |
773 | q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); | |
f52ed899 | 774 | q->flow_refill_delay = msecs_to_jiffies(40); |
afe4fd06 ED |
775 | q->flow_max_rate = ~0U; |
776 | q->rate_enable = 1; | |
777 | q->new_flows.first = NULL; | |
778 | q->old_flows.first = NULL; | |
779 | q->delayed = RB_ROOT; | |
780 | q->fq_root = NULL; | |
781 | q->fq_trees_log = ilog2(1024); | |
06eb395f | 782 | q->orphan_mask = 1024 - 1; |
afe4fd06 ED |
783 | qdisc_watchdog_init(&q->watchdog, sch); |
784 | ||
785 | if (opt) | |
786 | err = fq_change(sch, opt); | |
787 | else | |
c3bd8549 | 788 | err = fq_resize(sch, q->fq_trees_log); |
afe4fd06 ED |
789 | |
790 | return err; | |
791 | } | |
792 | ||
793 | static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) | |
794 | { | |
795 | struct fq_sched_data *q = qdisc_priv(sch); | |
796 | struct nlattr *opts; | |
797 | ||
798 | opts = nla_nest_start(skb, TCA_OPTIONS); | |
799 | if (opts == NULL) | |
800 | goto nla_put_failure; | |
801 | ||
65c5189a ED |
802 | /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ |
803 | ||
afe4fd06 ED |
804 | if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || |
805 | nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || | |
806 | nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || | |
807 | nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || | |
808 | nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || | |
afe4fd06 | 809 | nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) || |
f52ed899 ED |
810 | nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, |
811 | jiffies_to_usecs(q->flow_refill_delay)) || | |
06eb395f | 812 | nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) || |
afe4fd06 ED |
813 | nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) |
814 | goto nla_put_failure; | |
815 | ||
d59b7d80 | 816 | return nla_nest_end(skb, opts); |
afe4fd06 ED |
817 | |
818 | nla_put_failure: | |
819 | return -1; | |
820 | } | |
821 | ||
822 | static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) | |
823 | { | |
824 | struct fq_sched_data *q = qdisc_priv(sch); | |
d2de875c | 825 | u64 now = ktime_get_ns(); |
afe4fd06 ED |
826 | struct tc_fq_qd_stats st = { |
827 | .gc_flows = q->stat_gc_flows, | |
828 | .highprio_packets = q->stat_internal_packets, | |
829 | .tcp_retrans = q->stat_tcp_retrans, | |
830 | .throttled = q->stat_throttled, | |
831 | .flows_plimit = q->stat_flows_plimit, | |
832 | .pkts_too_long = q->stat_pkts_too_long, | |
833 | .allocation_errors = q->stat_allocation_errors, | |
834 | .flows = q->flows, | |
835 | .inactive_flows = q->inactive_flows, | |
836 | .throttled_flows = q->throttled_flows, | |
837 | .time_next_delayed_flow = q->time_next_delayed_flow - now, | |
838 | }; | |
839 | ||
840 | return gnet_stats_copy_app(d, &st, sizeof(st)); | |
841 | } | |
842 | ||
843 | static struct Qdisc_ops fq_qdisc_ops __read_mostly = { | |
844 | .id = "fq", | |
845 | .priv_size = sizeof(struct fq_sched_data), | |
846 | ||
847 | .enqueue = fq_enqueue, | |
848 | .dequeue = fq_dequeue, | |
849 | .peek = qdisc_peek_dequeued, | |
850 | .init = fq_init, | |
851 | .reset = fq_reset, | |
852 | .destroy = fq_destroy, | |
853 | .change = fq_change, | |
854 | .dump = fq_dump, | |
855 | .dump_stats = fq_dump_stats, | |
856 | .owner = THIS_MODULE, | |
857 | }; | |
858 | ||
859 | static int __init fq_module_init(void) | |
860 | { | |
861 | int ret; | |
862 | ||
863 | fq_flow_cachep = kmem_cache_create("fq_flow_cache", | |
864 | sizeof(struct fq_flow), | |
865 | 0, 0, NULL); | |
866 | if (!fq_flow_cachep) | |
867 | return -ENOMEM; | |
868 | ||
869 | ret = register_qdisc(&fq_qdisc_ops); | |
870 | if (ret) | |
871 | kmem_cache_destroy(fq_flow_cachep); | |
872 | return ret; | |
873 | } | |
874 | ||
875 | static void __exit fq_module_exit(void) | |
876 | { | |
877 | unregister_qdisc(&fq_qdisc_ops); | |
878 | kmem_cache_destroy(fq_flow_cachep); | |
879 | } | |
880 | ||
881 | module_init(fq_module_init) | |
882 | module_exit(fq_module_exit) | |
883 | MODULE_AUTHOR("Eric Dumazet"); | |
884 | MODULE_LICENSE("GPL"); |