Commit | Line | Data |
---|---|---|
2874c5fd | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
afe4fd06 ED |
2 | /* |
3 | * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing) | |
4 | * | |
86b3bfe9 | 5 | * Copyright (C) 2013-2015 Eric Dumazet <edumazet@google.com> |
afe4fd06 | 6 | * |
05e8bb86 | 7 | * Meant to be mostly used for locally generated traffic : |
afe4fd06 ED |
8 | * Fast classification depends on skb->sk being set before reaching us. |
9 | * If not, (router workload), we use rxhash as fallback, with 32 bits wide hash. | |
10 | * All packets belonging to a socket are considered as a 'flow'. | |
11 | * | |
12 | * Flows are dynamically allocated and stored in a hash table of RB trees | |
13 | * They are also part of one Round Robin 'queues' (new or old flows) | |
14 | * | |
15 | * Burst avoidance (aka pacing) capability : | |
16 | * | |
17 | * Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a | |
18 | * bunch of packets, and this packet scheduler adds delay between | |
19 | * packets to respect rate limitation. | |
20 | * | |
21 | * enqueue() : | |
22 | * - lookup one RB tree (out of 1024 or more) to find the flow. | |
23 | * If non existent flow, create it, add it to the tree. | |
24 | * Add skb to the per flow list of skb (fifo). | |
25 | * - Use a special fifo for high prio packets | |
26 | * | |
27 | * dequeue() : serves flows in Round Robin | |
28 | * Note : When a flow becomes empty, we do not immediately remove it from | |
29 | * rb trees, for performance reasons (its expected to send additional packets, | |
30 | * or SLAB cache will reuse socket for another flow) | |
31 | */ | |
32 | ||
33 | #include <linux/module.h> | |
34 | #include <linux/types.h> | |
35 | #include <linux/kernel.h> | |
36 | #include <linux/jiffies.h> | |
37 | #include <linux/string.h> | |
38 | #include <linux/in.h> | |
39 | #include <linux/errno.h> | |
40 | #include <linux/init.h> | |
41 | #include <linux/skbuff.h> | |
42 | #include <linux/slab.h> | |
43 | #include <linux/rbtree.h> | |
44 | #include <linux/hash.h> | |
08f89b98 | 45 | #include <linux/prefetch.h> |
c3bd8549 | 46 | #include <linux/vmalloc.h> |
afe4fd06 ED |
47 | #include <net/netlink.h> |
48 | #include <net/pkt_sched.h> | |
49 | #include <net/sock.h> | |
50 | #include <net/tcp_states.h> | |
98781965 | 51 | #include <net/tcp.h> |
afe4fd06 | 52 | |
eeb84aa0 ED |
53 | struct fq_skb_cb { |
54 | u64 time_to_send; | |
55 | }; | |
56 | ||
57 | static inline struct fq_skb_cb *fq_skb_cb(struct sk_buff *skb) | |
58 | { | |
59 | qdisc_cb_private_validate(skb, sizeof(struct fq_skb_cb)); | |
60 | return (struct fq_skb_cb *)qdisc_skb_cb(skb)->data; | |
61 | } | |
62 | ||
afe4fd06 | 63 | /* |
eeb84aa0 ED |
64 | * Per flow structure, dynamically allocated. |
65 | * If packets have monotically increasing time_to_send, they are placed in O(1) | |
66 | * in linear list (head,tail), otherwise are placed in a rbtree (t_root). | |
afe4fd06 ED |
67 | */ |
68 | struct fq_flow { | |
7ba0537c | 69 | /* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */ |
eeb84aa0 | 70 | struct rb_root t_root; |
afe4fd06 ED |
71 | struct sk_buff *head; /* list of skbs for this flow : first skb */ |
72 | union { | |
73 | struct sk_buff *tail; /* last skb in the list */ | |
dde0a648 | 74 | unsigned long age; /* (jiffies | 1UL) when flow was emptied, for gc */ |
afe4fd06 | 75 | }; |
05e8bb86 | 76 | struct rb_node fq_node; /* anchor in fq_root[] trees */ |
afe4fd06 | 77 | struct sock *sk; |
7ba0537c | 78 | u32 socket_hash; /* sk_hash */ |
afe4fd06 | 79 | int qlen; /* number of packets in flow queue */ |
7ba0537c ED |
80 | |
81 | /* Second cache line, used in fq_dequeue() */ | |
afe4fd06 | 82 | int credit; |
7ba0537c ED |
83 | /* 32bit hole on 64bit arches */ |
84 | ||
dde0a648 | 85 | struct fq_flow *next; /* next pointer in RR lists */ |
afe4fd06 ED |
86 | |
87 | struct rb_node rate_node; /* anchor in q->delayed tree */ | |
88 | u64 time_next_packet; | |
7ba0537c | 89 | } ____cacheline_aligned_in_smp; |
afe4fd06 ED |
90 | |
91 | struct fq_flow_head { | |
92 | struct fq_flow *first; | |
93 | struct fq_flow *last; | |
94 | }; | |
95 | ||
96 | struct fq_sched_data { | |
97 | struct fq_flow_head new_flows; | |
98 | ||
99 | struct fq_flow_head old_flows; | |
100 | ||
101 | struct rb_root delayed; /* for rate limited flows */ | |
102 | u64 time_next_delayed_flow; | |
39d01050 | 103 | u64 ktime_cache; /* copy of last ktime_get_ns() */ |
fefa569a | 104 | unsigned long unthrottle_latency_ns; |
afe4fd06 ED |
105 | |
106 | struct fq_flow internal; /* for non classified or high prio packets */ | |
107 | u32 quantum; | |
108 | u32 initial_quantum; | |
f52ed899 | 109 | u32 flow_refill_delay; |
afe4fd06 | 110 | u32 flow_plimit; /* max packets per flow */ |
76a9ebe8 | 111 | unsigned long flow_max_rate; /* optional max rate per flow */ |
48872c11 | 112 | u64 ce_threshold; |
39d01050 | 113 | u64 horizon; /* horizon in ns */ |
06eb395f | 114 | u32 orphan_mask; /* mask for orphaned skb */ |
77879147 | 115 | u32 low_rate_threshold; |
afe4fd06 ED |
116 | struct rb_root *fq_root; |
117 | u8 rate_enable; | |
118 | u8 fq_trees_log; | |
39d01050 | 119 | u8 horizon_drop; |
afe4fd06 ED |
120 | u32 flows; |
121 | u32 inactive_flows; | |
122 | u32 throttled_flows; | |
123 | ||
124 | u64 stat_gc_flows; | |
125 | u64 stat_internal_packets; | |
afe4fd06 | 126 | u64 stat_throttled; |
48872c11 | 127 | u64 stat_ce_mark; |
39d01050 ED |
128 | u64 stat_horizon_drops; |
129 | u64 stat_horizon_caps; | |
afe4fd06 ED |
130 | u64 stat_flows_plimit; |
131 | u64 stat_pkts_too_long; | |
132 | u64 stat_allocation_errors; | |
583396f4 ED |
133 | |
134 | u32 timer_slack; /* hrtimer slack in ns */ | |
afe4fd06 ED |
135 | struct qdisc_watchdog watchdog; |
136 | }; | |
137 | ||
dde0a648 ED |
138 | /* |
139 | * f->tail and f->age share the same location. | |
140 | * We can use the low order bit to differentiate if this location points | |
141 | * to a sk_buff or contains a jiffies value, if we force this value to be odd. | |
142 | * This assumes f->tail low order bit must be 0 since alignof(struct sk_buff) >= 2 | |
143 | */ | |
afe4fd06 ED |
144 | static void fq_flow_set_detached(struct fq_flow *f) |
145 | { | |
dde0a648 | 146 | f->age = jiffies | 1UL; |
afe4fd06 ED |
147 | } |
148 | ||
149 | static bool fq_flow_is_detached(const struct fq_flow *f) | |
150 | { | |
dde0a648 | 151 | return !!(f->age & 1UL); |
afe4fd06 ED |
152 | } |
153 | ||
dde0a648 ED |
154 | /* special value to mark a throttled flow (not on old/new list) */ |
155 | static struct fq_flow throttled; | |
156 | ||
7df40c26 ED |
157 | static bool fq_flow_is_throttled(const struct fq_flow *f) |
158 | { | |
159 | return f->next == &throttled; | |
160 | } | |
161 | ||
162 | static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) | |
163 | { | |
164 | if (head->first) | |
165 | head->last->next = flow; | |
166 | else | |
167 | head->first = flow; | |
168 | head->last = flow; | |
169 | flow->next = NULL; | |
170 | } | |
171 | ||
172 | static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f) | |
173 | { | |
174 | rb_erase(&f->rate_node, &q->delayed); | |
175 | q->throttled_flows--; | |
176 | fq_flow_add_tail(&q->old_flows, f); | |
177 | } | |
178 | ||
afe4fd06 ED |
179 | static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) |
180 | { | |
181 | struct rb_node **p = &q->delayed.rb_node, *parent = NULL; | |
182 | ||
183 | while (*p) { | |
184 | struct fq_flow *aux; | |
185 | ||
186 | parent = *p; | |
e124557d | 187 | aux = rb_entry(parent, struct fq_flow, rate_node); |
afe4fd06 ED |
188 | if (f->time_next_packet >= aux->time_next_packet) |
189 | p = &parent->rb_right; | |
190 | else | |
191 | p = &parent->rb_left; | |
192 | } | |
193 | rb_link_node(&f->rate_node, parent, p); | |
194 | rb_insert_color(&f->rate_node, &q->delayed); | |
195 | q->throttled_flows++; | |
196 | q->stat_throttled++; | |
197 | ||
198 | f->next = &throttled; | |
199 | if (q->time_next_delayed_flow > f->time_next_packet) | |
200 | q->time_next_delayed_flow = f->time_next_packet; | |
201 | } | |
202 | ||
203 | ||
204 | static struct kmem_cache *fq_flow_cachep __read_mostly; | |
205 | ||
afe4fd06 ED |
206 | |
207 | /* limit number of collected flows per round */ | |
208 | #define FQ_GC_MAX 8 | |
209 | #define FQ_GC_AGE (3*HZ) | |
210 | ||
211 | static bool fq_gc_candidate(const struct fq_flow *f) | |
212 | { | |
213 | return fq_flow_is_detached(f) && | |
214 | time_after(jiffies, f->age + FQ_GC_AGE); | |
215 | } | |
216 | ||
217 | static void fq_gc(struct fq_sched_data *q, | |
218 | struct rb_root *root, | |
219 | struct sock *sk) | |
220 | { | |
afe4fd06 | 221 | struct rb_node **p, *parent; |
82a0aa53 ED |
222 | void *tofree[FQ_GC_MAX]; |
223 | struct fq_flow *f; | |
224 | int i, fcnt = 0; | |
afe4fd06 ED |
225 | |
226 | p = &root->rb_node; | |
227 | parent = NULL; | |
228 | while (*p) { | |
229 | parent = *p; | |
230 | ||
e124557d | 231 | f = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
232 | if (f->sk == sk) |
233 | break; | |
234 | ||
235 | if (fq_gc_candidate(f)) { | |
236 | tofree[fcnt++] = f; | |
237 | if (fcnt == FQ_GC_MAX) | |
238 | break; | |
239 | } | |
240 | ||
241 | if (f->sk > sk) | |
242 | p = &parent->rb_right; | |
243 | else | |
244 | p = &parent->rb_left; | |
245 | } | |
246 | ||
82a0aa53 ED |
247 | if (!fcnt) |
248 | return; | |
249 | ||
250 | for (i = fcnt; i > 0; ) { | |
251 | f = tofree[--i]; | |
252 | rb_erase(&f->fq_node, root); | |
253 | } | |
afe4fd06 ED |
254 | q->flows -= fcnt; |
255 | q->inactive_flows -= fcnt; | |
256 | q->stat_gc_flows += fcnt; | |
afe4fd06 | 257 | |
82a0aa53 | 258 | kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree); |
afe4fd06 ED |
259 | } |
260 | ||
afe4fd06 ED |
261 | static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) |
262 | { | |
263 | struct rb_node **p, *parent; | |
264 | struct sock *sk = skb->sk; | |
265 | struct rb_root *root; | |
266 | struct fq_flow *f; | |
afe4fd06 ED |
267 | |
268 | /* warning: no starvation prevention... */ | |
2abc2f07 | 269 | if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) |
afe4fd06 ED |
270 | return &q->internal; |
271 | ||
ca6fb065 | 272 | /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket |
e446f9df | 273 | * or a listener (SYNCOOKIE mode) |
ca6fb065 ED |
274 | * 1) request sockets are not full blown, |
275 | * they do not contain sk_pacing_rate | |
276 | * 2) They are not part of a 'flow' yet | |
277 | * 3) We do not want to rate limit them (eg SYNFLOOD attack), | |
06eb395f | 278 | * especially if the listener set SO_MAX_PACING_RATE |
ca6fb065 | 279 | * 4) We pretend they are orphaned |
06eb395f | 280 | */ |
e446f9df | 281 | if (!sk || sk_listener(sk)) { |
06eb395f ED |
282 | unsigned long hash = skb_get_hash(skb) & q->orphan_mask; |
283 | ||
afe4fd06 ED |
284 | /* By forcing low order bit to 1, we make sure to not |
285 | * collide with a local flow (socket pointers are word aligned) | |
286 | */ | |
06eb395f ED |
287 | sk = (struct sock *)((hash << 1) | 1UL); |
288 | skb_orphan(skb); | |
37c0aead ED |
289 | } else if (sk->sk_state == TCP_CLOSE) { |
290 | unsigned long hash = skb_get_hash(skb) & q->orphan_mask; | |
291 | /* | |
292 | * Sockets in TCP_CLOSE are non connected. | |
293 | * Typical use case is UDP sockets, they can send packets | |
294 | * with sendto() to many different destinations. | |
295 | * We probably could use a generic bit advertising | |
296 | * non connected sockets, instead of sk_state == TCP_CLOSE, | |
297 | * if we care enough. | |
298 | */ | |
299 | sk = (struct sock *)((hash << 1) | 1UL); | |
afe4fd06 ED |
300 | } |
301 | ||
29c58472 | 302 | root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)]; |
afe4fd06 ED |
303 | |
304 | if (q->flows >= (2U << q->fq_trees_log) && | |
305 | q->inactive_flows > q->flows/2) | |
306 | fq_gc(q, root, sk); | |
307 | ||
308 | p = &root->rb_node; | |
309 | parent = NULL; | |
310 | while (*p) { | |
311 | parent = *p; | |
312 | ||
e124557d | 313 | f = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
314 | if (f->sk == sk) { |
315 | /* socket might have been reallocated, so check | |
316 | * if its sk_hash is the same. | |
317 | * It not, we need to refill credit with | |
318 | * initial quantum | |
319 | */ | |
37c0aead | 320 | if (unlikely(skb->sk == sk && |
afe4fd06 ED |
321 | f->socket_hash != sk->sk_hash)) { |
322 | f->credit = q->initial_quantum; | |
323 | f->socket_hash = sk->sk_hash; | |
bb3d0b8b ED |
324 | if (q->rate_enable) |
325 | smp_store_release(&sk->sk_pacing_status, | |
326 | SK_PACING_FQ); | |
7df40c26 ED |
327 | if (fq_flow_is_throttled(f)) |
328 | fq_flow_unset_throttled(q, f); | |
fc59d5bd | 329 | f->time_next_packet = 0ULL; |
afe4fd06 ED |
330 | } |
331 | return f; | |
332 | } | |
333 | if (f->sk > sk) | |
334 | p = &parent->rb_right; | |
335 | else | |
336 | p = &parent->rb_left; | |
337 | } | |
338 | ||
339 | f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); | |
340 | if (unlikely(!f)) { | |
341 | q->stat_allocation_errors++; | |
342 | return &q->internal; | |
343 | } | |
eeb84aa0 ED |
344 | /* f->t_root is already zeroed after kmem_cache_zalloc() */ |
345 | ||
afe4fd06 ED |
346 | fq_flow_set_detached(f); |
347 | f->sk = sk; | |
bb3d0b8b | 348 | if (skb->sk == sk) { |
afe4fd06 | 349 | f->socket_hash = sk->sk_hash; |
bb3d0b8b ED |
350 | if (q->rate_enable) |
351 | smp_store_release(&sk->sk_pacing_status, | |
352 | SK_PACING_FQ); | |
353 | } | |
afe4fd06 ED |
354 | f->credit = q->initial_quantum; |
355 | ||
356 | rb_link_node(&f->fq_node, parent, p); | |
357 | rb_insert_color(&f->fq_node, root); | |
358 | ||
359 | q->flows++; | |
360 | q->inactive_flows++; | |
361 | return f; | |
362 | } | |
363 | ||
eeb84aa0 ED |
364 | static struct sk_buff *fq_peek(struct fq_flow *flow) |
365 | { | |
366 | struct sk_buff *skb = skb_rb_first(&flow->t_root); | |
367 | struct sk_buff *head = flow->head; | |
368 | ||
369 | if (!skb) | |
370 | return head; | |
371 | ||
372 | if (!head) | |
373 | return skb; | |
374 | ||
375 | if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send) | |
376 | return skb; | |
377 | return head; | |
378 | } | |
379 | ||
380 | static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow, | |
381 | struct sk_buff *skb) | |
382 | { | |
383 | if (skb == flow->head) { | |
384 | flow->head = skb->next; | |
385 | } else { | |
386 | rb_erase(&skb->rbnode, &flow->t_root); | |
387 | skb->dev = qdisc_dev(sch); | |
388 | } | |
389 | } | |
afe4fd06 | 390 | |
c288b0ca ED |
391 | /* Remove one skb from flow queue. |
392 | * This skb must be the return value of prior fq_peek(). | |
393 | */ | |
394 | static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow, | |
395 | struct sk_buff *skb) | |
afe4fd06 | 396 | { |
c288b0ca ED |
397 | fq_erase_head(sch, flow, skb); |
398 | skb_mark_not_on_list(skb); | |
399 | flow->qlen--; | |
400 | qdisc_qstats_backlog_dec(sch, skb); | |
401 | sch->q.qlen--; | |
afe4fd06 ED |
402 | } |
403 | ||
afe4fd06 ED |
404 | static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) |
405 | { | |
eeb84aa0 ED |
406 | struct rb_node **p, *parent; |
407 | struct sk_buff *head, *aux; | |
afe4fd06 | 408 | |
eeb84aa0 ED |
409 | head = flow->head; |
410 | if (!head || | |
411 | fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) { | |
412 | if (!head) | |
413 | flow->head = skb; | |
414 | else | |
415 | flow->tail->next = skb; | |
416 | flow->tail = skb; | |
417 | skb->next = NULL; | |
418 | return; | |
419 | } | |
420 | ||
421 | p = &flow->t_root.rb_node; | |
422 | parent = NULL; | |
afe4fd06 | 423 | |
eeb84aa0 ED |
424 | while (*p) { |
425 | parent = *p; | |
426 | aux = rb_to_skb(parent); | |
427 | if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send) | |
428 | p = &parent->rb_right; | |
429 | else | |
430 | p = &parent->rb_left; | |
431 | } | |
432 | rb_link_node(&skb->rbnode, parent, p); | |
433 | rb_insert_color(&skb->rbnode, &flow->t_root); | |
afe4fd06 ED |
434 | } |
435 | ||
39d01050 ED |
436 | static bool fq_packet_beyond_horizon(const struct sk_buff *skb, |
437 | const struct fq_sched_data *q) | |
438 | { | |
439 | return unlikely((s64)skb->tstamp > (s64)(q->ktime_cache + q->horizon)); | |
440 | } | |
441 | ||
520ac30f ED |
442 | static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
443 | struct sk_buff **to_free) | |
afe4fd06 ED |
444 | { |
445 | struct fq_sched_data *q = qdisc_priv(sch); | |
446 | struct fq_flow *f; | |
447 | ||
448 | if (unlikely(sch->q.qlen >= sch->limit)) | |
520ac30f | 449 | return qdisc_drop(skb, sch, to_free); |
afe4fd06 | 450 | |
39d01050 ED |
451 | if (!skb->tstamp) { |
452 | fq_skb_cb(skb)->time_to_send = q->ktime_cache = ktime_get_ns(); | |
453 | } else { | |
454 | /* Check if packet timestamp is too far in the future. | |
455 | * Try first if our cached value, to avoid ktime_get_ns() | |
456 | * cost in most cases. | |
457 | */ | |
458 | if (fq_packet_beyond_horizon(skb, q)) { | |
459 | /* Refresh our cache and check another time */ | |
460 | q->ktime_cache = ktime_get_ns(); | |
461 | if (fq_packet_beyond_horizon(skb, q)) { | |
462 | if (q->horizon_drop) { | |
463 | q->stat_horizon_drops++; | |
464 | return qdisc_drop(skb, sch, to_free); | |
465 | } | |
466 | q->stat_horizon_caps++; | |
467 | skb->tstamp = q->ktime_cache + q->horizon; | |
468 | } | |
469 | } | |
470 | fq_skb_cb(skb)->time_to_send = skb->tstamp; | |
471 | } | |
472 | ||
afe4fd06 ED |
473 | f = fq_classify(skb, q); |
474 | if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { | |
475 | q->stat_flows_plimit++; | |
520ac30f | 476 | return qdisc_drop(skb, sch, to_free); |
afe4fd06 ED |
477 | } |
478 | ||
479 | f->qlen++; | |
25331d6c | 480 | qdisc_qstats_backlog_inc(sch, skb); |
afe4fd06 ED |
481 | if (fq_flow_is_detached(f)) { |
482 | fq_flow_add_tail(&q->new_flows, f); | |
f52ed899 ED |
483 | if (time_after(jiffies, f->age + q->flow_refill_delay)) |
484 | f->credit = max_t(u32, f->credit, q->quantum); | |
afe4fd06 | 485 | q->inactive_flows--; |
afe4fd06 | 486 | } |
f52ed899 ED |
487 | |
488 | /* Note: this overwrites f->age */ | |
489 | flow_queue_add(f, skb); | |
490 | ||
afe4fd06 ED |
491 | if (unlikely(f == &q->internal)) { |
492 | q->stat_internal_packets++; | |
afe4fd06 ED |
493 | } |
494 | sch->q.qlen++; | |
495 | ||
496 | return NET_XMIT_SUCCESS; | |
497 | } | |
498 | ||
499 | static void fq_check_throttled(struct fq_sched_data *q, u64 now) | |
500 | { | |
fefa569a | 501 | unsigned long sample; |
afe4fd06 ED |
502 | struct rb_node *p; |
503 | ||
504 | if (q->time_next_delayed_flow > now) | |
505 | return; | |
506 | ||
fefa569a ED |
507 | /* Update unthrottle latency EWMA. |
508 | * This is cheap and can help diagnosing timer/latency problems. | |
509 | */ | |
510 | sample = (unsigned long)(now - q->time_next_delayed_flow); | |
511 | q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3; | |
512 | q->unthrottle_latency_ns += sample >> 3; | |
513 | ||
afe4fd06 ED |
514 | q->time_next_delayed_flow = ~0ULL; |
515 | while ((p = rb_first(&q->delayed)) != NULL) { | |
e124557d | 516 | struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node); |
afe4fd06 ED |
517 | |
518 | if (f->time_next_packet > now) { | |
519 | q->time_next_delayed_flow = f->time_next_packet; | |
520 | break; | |
521 | } | |
7df40c26 | 522 | fq_flow_unset_throttled(q, f); |
afe4fd06 ED |
523 | } |
524 | } | |
525 | ||
526 | static struct sk_buff *fq_dequeue(struct Qdisc *sch) | |
527 | { | |
528 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 ED |
529 | struct fq_flow_head *head; |
530 | struct sk_buff *skb; | |
531 | struct fq_flow *f; | |
76a9ebe8 ED |
532 | unsigned long rate; |
533 | u32 plen; | |
6b015a52 ED |
534 | u64 now; |
535 | ||
536 | if (!sch->q.qlen) | |
537 | return NULL; | |
afe4fd06 | 538 | |
c288b0ca ED |
539 | skb = fq_peek(&q->internal); |
540 | if (unlikely(skb)) { | |
541 | fq_dequeue_skb(sch, &q->internal, skb); | |
afe4fd06 | 542 | goto out; |
c288b0ca | 543 | } |
6b015a52 | 544 | |
39d01050 | 545 | q->ktime_cache = now = ktime_get_ns(); |
afe4fd06 ED |
546 | fq_check_throttled(q, now); |
547 | begin: | |
548 | head = &q->new_flows; | |
549 | if (!head->first) { | |
550 | head = &q->old_flows; | |
551 | if (!head->first) { | |
552 | if (q->time_next_delayed_flow != ~0ULL) | |
583396f4 ED |
553 | qdisc_watchdog_schedule_range_ns(&q->watchdog, |
554 | q->time_next_delayed_flow, | |
555 | q->timer_slack); | |
afe4fd06 ED |
556 | return NULL; |
557 | } | |
558 | } | |
559 | f = head->first; | |
560 | ||
561 | if (f->credit <= 0) { | |
562 | f->credit += q->quantum; | |
563 | head->first = f->next; | |
564 | fq_flow_add_tail(&q->old_flows, f); | |
565 | goto begin; | |
566 | } | |
567 | ||
eeb84aa0 | 568 | skb = fq_peek(f); |
7baf33bd | 569 | if (skb) { |
eeb84aa0 | 570 | u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send, |
ab408b6d ED |
571 | f->time_next_packet); |
572 | ||
573 | if (now < time_next_packet) { | |
574 | head->first = f->next; | |
575 | f->time_next_packet = time_next_packet; | |
576 | fq_flow_set_throttled(q, f); | |
577 | goto begin; | |
578 | } | |
348e289b | 579 | prefetch(&skb->end); |
e9c43add | 580 | if ((s64)(now - time_next_packet - q->ce_threshold) > 0) { |
48872c11 ED |
581 | INET_ECN_set_ce(skb); |
582 | q->stat_ce_mark++; | |
583 | } | |
c288b0ca ED |
584 | fq_dequeue_skb(sch, f, skb); |
585 | } else { | |
afe4fd06 ED |
586 | head->first = f->next; |
587 | /* force a pass through old_flows to prevent starvation */ | |
588 | if ((head == &q->new_flows) && q->old_flows.first) { | |
589 | fq_flow_add_tail(&q->old_flows, f); | |
590 | } else { | |
591 | fq_flow_set_detached(f); | |
afe4fd06 ED |
592 | q->inactive_flows++; |
593 | } | |
594 | goto begin; | |
595 | } | |
08e14fe4 ED |
596 | plen = qdisc_pkt_len(skb); |
597 | f->credit -= plen; | |
afe4fd06 | 598 | |
08e14fe4 | 599 | if (!q->rate_enable) |
98781965 ED |
600 | goto out; |
601 | ||
7eec4174 | 602 | rate = q->flow_max_rate; |
08e14fe4 ED |
603 | |
604 | /* If EDT time was provided for this skb, we need to | |
605 | * update f->time_next_packet only if this qdisc enforces | |
606 | * a flow max rate. | |
607 | */ | |
608 | if (!skb->tstamp) { | |
609 | if (skb->sk) | |
610 | rate = min(skb->sk->sk_pacing_rate, rate); | |
611 | ||
612 | if (rate <= q->low_rate_threshold) { | |
613 | f->credit = 0; | |
614 | } else { | |
615 | plen = max(plen, q->quantum); | |
616 | if (f->credit > 0) | |
617 | goto out; | |
618 | } | |
77879147 | 619 | } |
76a9ebe8 | 620 | if (rate != ~0UL) { |
0eab5eb7 ED |
621 | u64 len = (u64)plen * NSEC_PER_SEC; |
622 | ||
7eec4174 | 623 | if (likely(rate)) |
76a9ebe8 | 624 | len = div64_ul(len, rate); |
0eab5eb7 | 625 | /* Since socket rate can change later, |
ced7a04e ED |
626 | * clamp the delay to 1 second. |
627 | * Really, providers of too big packets should be fixed ! | |
0eab5eb7 | 628 | */ |
ced7a04e ED |
629 | if (unlikely(len > NSEC_PER_SEC)) { |
630 | len = NSEC_PER_SEC; | |
0eab5eb7 | 631 | q->stat_pkts_too_long++; |
afe4fd06 | 632 | } |
fefa569a ED |
633 | /* Account for schedule/timers drifts. |
634 | * f->time_next_packet was set when prior packet was sent, | |
635 | * and current time (@now) can be too late by tens of us. | |
636 | */ | |
637 | if (f->time_next_packet) | |
638 | len -= min(len/2, now - f->time_next_packet); | |
0eab5eb7 | 639 | f->time_next_packet = now + len; |
afe4fd06 ED |
640 | } |
641 | out: | |
afe4fd06 | 642 | qdisc_bstats_update(sch, skb); |
afe4fd06 ED |
643 | return skb; |
644 | } | |
645 | ||
e14ffdfd ED |
646 | static void fq_flow_purge(struct fq_flow *flow) |
647 | { | |
eeb84aa0 ED |
648 | struct rb_node *p = rb_first(&flow->t_root); |
649 | ||
650 | while (p) { | |
651 | struct sk_buff *skb = rb_to_skb(p); | |
652 | ||
653 | p = rb_next(p); | |
654 | rb_erase(&skb->rbnode, &flow->t_root); | |
655 | rtnl_kfree_skbs(skb, skb); | |
656 | } | |
e14ffdfd ED |
657 | rtnl_kfree_skbs(flow->head, flow->tail); |
658 | flow->head = NULL; | |
659 | flow->qlen = 0; | |
660 | } | |
661 | ||
afe4fd06 ED |
662 | static void fq_reset(struct Qdisc *sch) |
663 | { | |
8d34ce10 ED |
664 | struct fq_sched_data *q = qdisc_priv(sch); |
665 | struct rb_root *root; | |
8d34ce10 ED |
666 | struct rb_node *p; |
667 | struct fq_flow *f; | |
668 | unsigned int idx; | |
afe4fd06 | 669 | |
e14ffdfd ED |
670 | sch->q.qlen = 0; |
671 | sch->qstats.backlog = 0; | |
672 | ||
673 | fq_flow_purge(&q->internal); | |
8d34ce10 ED |
674 | |
675 | if (!q->fq_root) | |
676 | return; | |
677 | ||
678 | for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { | |
679 | root = &q->fq_root[idx]; | |
680 | while ((p = rb_first(root)) != NULL) { | |
e124557d | 681 | f = rb_entry(p, struct fq_flow, fq_node); |
8d34ce10 ED |
682 | rb_erase(p, root); |
683 | ||
e14ffdfd | 684 | fq_flow_purge(f); |
8d34ce10 ED |
685 | |
686 | kmem_cache_free(fq_flow_cachep, f); | |
687 | } | |
688 | } | |
689 | q->new_flows.first = NULL; | |
690 | q->old_flows.first = NULL; | |
691 | q->delayed = RB_ROOT; | |
692 | q->flows = 0; | |
693 | q->inactive_flows = 0; | |
694 | q->throttled_flows = 0; | |
afe4fd06 ED |
695 | } |
696 | ||
697 | static void fq_rehash(struct fq_sched_data *q, | |
698 | struct rb_root *old_array, u32 old_log, | |
699 | struct rb_root *new_array, u32 new_log) | |
700 | { | |
701 | struct rb_node *op, **np, *parent; | |
702 | struct rb_root *oroot, *nroot; | |
703 | struct fq_flow *of, *nf; | |
704 | int fcnt = 0; | |
705 | u32 idx; | |
706 | ||
707 | for (idx = 0; idx < (1U << old_log); idx++) { | |
708 | oroot = &old_array[idx]; | |
709 | while ((op = rb_first(oroot)) != NULL) { | |
710 | rb_erase(op, oroot); | |
e124557d | 711 | of = rb_entry(op, struct fq_flow, fq_node); |
afe4fd06 ED |
712 | if (fq_gc_candidate(of)) { |
713 | fcnt++; | |
714 | kmem_cache_free(fq_flow_cachep, of); | |
715 | continue; | |
716 | } | |
29c58472 | 717 | nroot = &new_array[hash_ptr(of->sk, new_log)]; |
afe4fd06 ED |
718 | |
719 | np = &nroot->rb_node; | |
720 | parent = NULL; | |
721 | while (*np) { | |
722 | parent = *np; | |
723 | ||
e124557d | 724 | nf = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
725 | BUG_ON(nf->sk == of->sk); |
726 | ||
727 | if (nf->sk > of->sk) | |
728 | np = &parent->rb_right; | |
729 | else | |
730 | np = &parent->rb_left; | |
731 | } | |
732 | ||
733 | rb_link_node(&of->fq_node, parent, np); | |
734 | rb_insert_color(&of->fq_node, nroot); | |
735 | } | |
736 | } | |
737 | q->flows -= fcnt; | |
738 | q->inactive_flows -= fcnt; | |
739 | q->stat_gc_flows += fcnt; | |
740 | } | |
741 | ||
c3bd8549 ED |
742 | static void fq_free(void *addr) |
743 | { | |
4cb28970 | 744 | kvfree(addr); |
c3bd8549 ED |
745 | } |
746 | ||
747 | static int fq_resize(struct Qdisc *sch, u32 log) | |
748 | { | |
749 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 750 | struct rb_root *array; |
2d8d40af | 751 | void *old_fq_root; |
afe4fd06 ED |
752 | u32 idx; |
753 | ||
754 | if (q->fq_root && log == q->fq_trees_log) | |
755 | return 0; | |
756 | ||
c3bd8549 | 757 | /* If XPS was setup, we can allocate memory on right NUMA node */ |
dcda9b04 | 758 | array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL, |
c3bd8549 | 759 | netdev_queue_numa_node_read(sch->dev_queue)); |
afe4fd06 ED |
760 | if (!array) |
761 | return -ENOMEM; | |
762 | ||
763 | for (idx = 0; idx < (1U << log); idx++) | |
764 | array[idx] = RB_ROOT; | |
765 | ||
2d8d40af ED |
766 | sch_tree_lock(sch); |
767 | ||
768 | old_fq_root = q->fq_root; | |
769 | if (old_fq_root) | |
770 | fq_rehash(q, old_fq_root, q->fq_trees_log, array, log); | |
771 | ||
afe4fd06 ED |
772 | q->fq_root = array; |
773 | q->fq_trees_log = log; | |
774 | ||
2d8d40af ED |
775 | sch_tree_unlock(sch); |
776 | ||
777 | fq_free(old_fq_root); | |
778 | ||
afe4fd06 ED |
779 | return 0; |
780 | } | |
781 | ||
782 | static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { | |
583396f4 ED |
783 | [TCA_FQ_UNSPEC] = { .strict_start_type = TCA_FQ_TIMER_SLACK }, |
784 | ||
afe4fd06 ED |
785 | [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, |
786 | [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, | |
787 | [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, | |
788 | [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, | |
789 | [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, | |
790 | [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, | |
791 | [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, | |
792 | [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, | |
f52ed899 | 793 | [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, |
7e6dc03e | 794 | [TCA_FQ_ORPHAN_MASK] = { .type = NLA_U32 }, |
77879147 | 795 | [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 }, |
48872c11 | 796 | [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 }, |
583396f4 | 797 | [TCA_FQ_TIMER_SLACK] = { .type = NLA_U32 }, |
39d01050 ED |
798 | [TCA_FQ_HORIZON] = { .type = NLA_U32 }, |
799 | [TCA_FQ_HORIZON_DROP] = { .type = NLA_U8 }, | |
afe4fd06 ED |
800 | }; |
801 | ||
2030721c AA |
802 | static int fq_change(struct Qdisc *sch, struct nlattr *opt, |
803 | struct netlink_ext_ack *extack) | |
afe4fd06 ED |
804 | { |
805 | struct fq_sched_data *q = qdisc_priv(sch); | |
806 | struct nlattr *tb[TCA_FQ_MAX + 1]; | |
807 | int err, drop_count = 0; | |
2ccccf5f | 808 | unsigned drop_len = 0; |
afe4fd06 ED |
809 | u32 fq_log; |
810 | ||
8cb08174 JB |
811 | err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy, |
812 | NULL); | |
afe4fd06 ED |
813 | if (err < 0) |
814 | return err; | |
815 | ||
816 | sch_tree_lock(sch); | |
817 | ||
818 | fq_log = q->fq_trees_log; | |
819 | ||
820 | if (tb[TCA_FQ_BUCKETS_LOG]) { | |
821 | u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); | |
822 | ||
823 | if (nval >= 1 && nval <= ilog2(256*1024)) | |
824 | fq_log = nval; | |
825 | else | |
826 | err = -EINVAL; | |
827 | } | |
828 | if (tb[TCA_FQ_PLIMIT]) | |
829 | sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); | |
830 | ||
831 | if (tb[TCA_FQ_FLOW_PLIMIT]) | |
832 | q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); | |
833 | ||
3725a269 KKJ |
834 | if (tb[TCA_FQ_QUANTUM]) { |
835 | u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); | |
836 | ||
d9e15a27 | 837 | if (quantum > 0 && quantum <= (1 << 20)) { |
3725a269 | 838 | q->quantum = quantum; |
d9e15a27 ED |
839 | } else { |
840 | NL_SET_ERR_MSG_MOD(extack, "invalid quantum"); | |
3725a269 | 841 | err = -EINVAL; |
d9e15a27 | 842 | } |
3725a269 | 843 | } |
afe4fd06 ED |
844 | |
845 | if (tb[TCA_FQ_INITIAL_QUANTUM]) | |
ede869cd | 846 | q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); |
afe4fd06 ED |
847 | |
848 | if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) | |
65c5189a ED |
849 | pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", |
850 | nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); | |
afe4fd06 | 851 | |
76a9ebe8 ED |
852 | if (tb[TCA_FQ_FLOW_MAX_RATE]) { |
853 | u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); | |
afe4fd06 | 854 | |
76a9ebe8 ED |
855 | q->flow_max_rate = (rate == ~0U) ? ~0UL : rate; |
856 | } | |
77879147 ED |
857 | if (tb[TCA_FQ_LOW_RATE_THRESHOLD]) |
858 | q->low_rate_threshold = | |
859 | nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]); | |
860 | ||
afe4fd06 ED |
861 | if (tb[TCA_FQ_RATE_ENABLE]) { |
862 | u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); | |
863 | ||
864 | if (enable <= 1) | |
865 | q->rate_enable = enable; | |
866 | else | |
867 | err = -EINVAL; | |
868 | } | |
869 | ||
f52ed899 ED |
870 | if (tb[TCA_FQ_FLOW_REFILL_DELAY]) { |
871 | u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ; | |
872 | ||
873 | q->flow_refill_delay = usecs_to_jiffies(usecs_delay); | |
874 | } | |
875 | ||
06eb395f ED |
876 | if (tb[TCA_FQ_ORPHAN_MASK]) |
877 | q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]); | |
878 | ||
48872c11 ED |
879 | if (tb[TCA_FQ_CE_THRESHOLD]) |
880 | q->ce_threshold = (u64)NSEC_PER_USEC * | |
881 | nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]); | |
882 | ||
583396f4 ED |
883 | if (tb[TCA_FQ_TIMER_SLACK]) |
884 | q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]); | |
885 | ||
39d01050 ED |
886 | if (tb[TCA_FQ_HORIZON]) |
887 | q->horizon = (u64)NSEC_PER_USEC * | |
888 | nla_get_u32(tb[TCA_FQ_HORIZON]); | |
889 | ||
890 | if (tb[TCA_FQ_HORIZON_DROP]) | |
891 | q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]); | |
892 | ||
2d8d40af | 893 | if (!err) { |
39d01050 | 894 | |
2d8d40af | 895 | sch_tree_unlock(sch); |
c3bd8549 | 896 | err = fq_resize(sch, fq_log); |
2d8d40af ED |
897 | sch_tree_lock(sch); |
898 | } | |
afe4fd06 ED |
899 | while (sch->q.qlen > sch->limit) { |
900 | struct sk_buff *skb = fq_dequeue(sch); | |
901 | ||
8d34ce10 ED |
902 | if (!skb) |
903 | break; | |
2ccccf5f | 904 | drop_len += qdisc_pkt_len(skb); |
e14ffdfd | 905 | rtnl_kfree_skbs(skb, skb); |
afe4fd06 ED |
906 | drop_count++; |
907 | } | |
2ccccf5f | 908 | qdisc_tree_reduce_backlog(sch, drop_count, drop_len); |
afe4fd06 ED |
909 | |
910 | sch_tree_unlock(sch); | |
911 | return err; | |
912 | } | |
913 | ||
914 | static void fq_destroy(struct Qdisc *sch) | |
915 | { | |
916 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 917 | |
8d34ce10 | 918 | fq_reset(sch); |
c3bd8549 | 919 | fq_free(q->fq_root); |
afe4fd06 ED |
920 | qdisc_watchdog_cancel(&q->watchdog); |
921 | } | |
922 | ||
e63d7dfd AA |
923 | static int fq_init(struct Qdisc *sch, struct nlattr *opt, |
924 | struct netlink_ext_ack *extack) | |
afe4fd06 ED |
925 | { |
926 | struct fq_sched_data *q = qdisc_priv(sch); | |
927 | int err; | |
928 | ||
929 | sch->limit = 10000; | |
930 | q->flow_plimit = 100; | |
931 | q->quantum = 2 * psched_mtu(qdisc_dev(sch)); | |
932 | q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); | |
f52ed899 | 933 | q->flow_refill_delay = msecs_to_jiffies(40); |
76a9ebe8 | 934 | q->flow_max_rate = ~0UL; |
fefa569a | 935 | q->time_next_delayed_flow = ~0ULL; |
afe4fd06 ED |
936 | q->rate_enable = 1; |
937 | q->new_flows.first = NULL; | |
938 | q->old_flows.first = NULL; | |
939 | q->delayed = RB_ROOT; | |
940 | q->fq_root = NULL; | |
941 | q->fq_trees_log = ilog2(1024); | |
06eb395f | 942 | q->orphan_mask = 1024 - 1; |
77879147 | 943 | q->low_rate_threshold = 550000 / 8; |
48872c11 | 944 | |
583396f4 ED |
945 | q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */ |
946 | ||
39d01050 ED |
947 | q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */ |
948 | q->horizon_drop = 1; /* by default, drop packets beyond horizon */ | |
949 | ||
48872c11 ED |
950 | /* Default ce_threshold of 4294 seconds */ |
951 | q->ce_threshold = (u64)NSEC_PER_USEC * ~0U; | |
952 | ||
fb420d5d | 953 | qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC); |
afe4fd06 ED |
954 | |
955 | if (opt) | |
2030721c | 956 | err = fq_change(sch, opt, extack); |
afe4fd06 | 957 | else |
c3bd8549 | 958 | err = fq_resize(sch, q->fq_trees_log); |
afe4fd06 ED |
959 | |
960 | return err; | |
961 | } | |
962 | ||
963 | static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) | |
964 | { | |
965 | struct fq_sched_data *q = qdisc_priv(sch); | |
48872c11 | 966 | u64 ce_threshold = q->ce_threshold; |
39d01050 | 967 | u64 horizon = q->horizon; |
afe4fd06 ED |
968 | struct nlattr *opts; |
969 | ||
ae0be8de | 970 | opts = nla_nest_start_noflag(skb, TCA_OPTIONS); |
afe4fd06 ED |
971 | if (opts == NULL) |
972 | goto nla_put_failure; | |
973 | ||
65c5189a ED |
974 | /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ |
975 | ||
48872c11 | 976 | do_div(ce_threshold, NSEC_PER_USEC); |
39d01050 | 977 | do_div(horizon, NSEC_PER_USEC); |
48872c11 | 978 | |
afe4fd06 ED |
979 | if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || |
980 | nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || | |
981 | nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || | |
982 | nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || | |
983 | nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || | |
76a9ebe8 ED |
984 | nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, |
985 | min_t(unsigned long, q->flow_max_rate, ~0U)) || | |
f52ed899 ED |
986 | nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, |
987 | jiffies_to_usecs(q->flow_refill_delay)) || | |
06eb395f | 988 | nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) || |
77879147 ED |
989 | nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD, |
990 | q->low_rate_threshold) || | |
48872c11 | 991 | nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) || |
583396f4 | 992 | nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) || |
39d01050 ED |
993 | nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) || |
994 | nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) || | |
995 | nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop)) | |
afe4fd06 ED |
996 | goto nla_put_failure; |
997 | ||
d59b7d80 | 998 | return nla_nest_end(skb, opts); |
afe4fd06 ED |
999 | |
1000 | nla_put_failure: | |
1001 | return -1; | |
1002 | } | |
1003 | ||
1004 | static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) | |
1005 | { | |
1006 | struct fq_sched_data *q = qdisc_priv(sch); | |
695b4ec0 ED |
1007 | struct tc_fq_qd_stats st; |
1008 | ||
1009 | sch_tree_lock(sch); | |
1010 | ||
1011 | st.gc_flows = q->stat_gc_flows; | |
1012 | st.highprio_packets = q->stat_internal_packets; | |
90caf67b | 1013 | st.tcp_retrans = 0; |
695b4ec0 ED |
1014 | st.throttled = q->stat_throttled; |
1015 | st.flows_plimit = q->stat_flows_plimit; | |
1016 | st.pkts_too_long = q->stat_pkts_too_long; | |
1017 | st.allocation_errors = q->stat_allocation_errors; | |
583396f4 ED |
1018 | st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack - |
1019 | ktime_get_ns(); | |
695b4ec0 ED |
1020 | st.flows = q->flows; |
1021 | st.inactive_flows = q->inactive_flows; | |
1022 | st.throttled_flows = q->throttled_flows; | |
fefa569a ED |
1023 | st.unthrottle_latency_ns = min_t(unsigned long, |
1024 | q->unthrottle_latency_ns, ~0U); | |
48872c11 | 1025 | st.ce_mark = q->stat_ce_mark; |
39d01050 ED |
1026 | st.horizon_drops = q->stat_horizon_drops; |
1027 | st.horizon_caps = q->stat_horizon_caps; | |
695b4ec0 | 1028 | sch_tree_unlock(sch); |
afe4fd06 ED |
1029 | |
1030 | return gnet_stats_copy_app(d, &st, sizeof(st)); | |
1031 | } | |
1032 | ||
1033 | static struct Qdisc_ops fq_qdisc_ops __read_mostly = { | |
1034 | .id = "fq", | |
1035 | .priv_size = sizeof(struct fq_sched_data), | |
1036 | ||
1037 | .enqueue = fq_enqueue, | |
1038 | .dequeue = fq_dequeue, | |
1039 | .peek = qdisc_peek_dequeued, | |
1040 | .init = fq_init, | |
1041 | .reset = fq_reset, | |
1042 | .destroy = fq_destroy, | |
1043 | .change = fq_change, | |
1044 | .dump = fq_dump, | |
1045 | .dump_stats = fq_dump_stats, | |
1046 | .owner = THIS_MODULE, | |
1047 | }; | |
1048 | ||
1049 | static int __init fq_module_init(void) | |
1050 | { | |
1051 | int ret; | |
1052 | ||
1053 | fq_flow_cachep = kmem_cache_create("fq_flow_cache", | |
1054 | sizeof(struct fq_flow), | |
1055 | 0, 0, NULL); | |
1056 | if (!fq_flow_cachep) | |
1057 | return -ENOMEM; | |
1058 | ||
1059 | ret = register_qdisc(&fq_qdisc_ops); | |
1060 | if (ret) | |
1061 | kmem_cache_destroy(fq_flow_cachep); | |
1062 | return ret; | |
1063 | } | |
1064 | ||
1065 | static void __exit fq_module_exit(void) | |
1066 | { | |
1067 | unregister_qdisc(&fq_qdisc_ops); | |
1068 | kmem_cache_destroy(fq_flow_cachep); | |
1069 | } | |
1070 | ||
1071 | module_init(fq_module_init) | |
1072 | module_exit(fq_module_exit) | |
1073 | MODULE_AUTHOR("Eric Dumazet"); | |
1074 | MODULE_LICENSE("GPL"); | |
67c20de3 | 1075 | MODULE_DESCRIPTION("Fair Queue Packet Scheduler"); |