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 { | |
eeb84aa0 | 69 | struct rb_root t_root; |
afe4fd06 ED |
70 | struct sk_buff *head; /* list of skbs for this flow : first skb */ |
71 | union { | |
72 | struct sk_buff *tail; /* last skb in the list */ | |
73 | unsigned long age; /* jiffies when flow was emptied, for gc */ | |
74 | }; | |
05e8bb86 | 75 | struct rb_node fq_node; /* anchor in fq_root[] trees */ |
afe4fd06 ED |
76 | struct sock *sk; |
77 | int qlen; /* number of packets in flow queue */ | |
78 | int credit; | |
79 | u32 socket_hash; /* sk_hash */ | |
80 | struct fq_flow *next; /* next pointer in RR lists, or &detached */ | |
81 | ||
82 | struct rb_node rate_node; /* anchor in q->delayed tree */ | |
83 | u64 time_next_packet; | |
84 | }; | |
85 | ||
86 | struct fq_flow_head { | |
87 | struct fq_flow *first; | |
88 | struct fq_flow *last; | |
89 | }; | |
90 | ||
91 | struct fq_sched_data { | |
92 | struct fq_flow_head new_flows; | |
93 | ||
94 | struct fq_flow_head old_flows; | |
95 | ||
96 | struct rb_root delayed; /* for rate limited flows */ | |
97 | u64 time_next_delayed_flow; | |
fefa569a | 98 | unsigned long unthrottle_latency_ns; |
afe4fd06 ED |
99 | |
100 | struct fq_flow internal; /* for non classified or high prio packets */ | |
101 | u32 quantum; | |
102 | u32 initial_quantum; | |
f52ed899 | 103 | u32 flow_refill_delay; |
afe4fd06 | 104 | u32 flow_plimit; /* max packets per flow */ |
76a9ebe8 | 105 | unsigned long flow_max_rate; /* optional max rate per flow */ |
48872c11 | 106 | u64 ce_threshold; |
06eb395f | 107 | u32 orphan_mask; /* mask for orphaned skb */ |
77879147 | 108 | u32 low_rate_threshold; |
afe4fd06 ED |
109 | struct rb_root *fq_root; |
110 | u8 rate_enable; | |
111 | u8 fq_trees_log; | |
112 | ||
113 | u32 flows; | |
114 | u32 inactive_flows; | |
115 | u32 throttled_flows; | |
116 | ||
117 | u64 stat_gc_flows; | |
118 | u64 stat_internal_packets; | |
afe4fd06 | 119 | u64 stat_throttled; |
48872c11 | 120 | u64 stat_ce_mark; |
afe4fd06 ED |
121 | u64 stat_flows_plimit; |
122 | u64 stat_pkts_too_long; | |
123 | u64 stat_allocation_errors; | |
124 | struct qdisc_watchdog watchdog; | |
125 | }; | |
126 | ||
127 | /* special value to mark a detached flow (not on old/new list) */ | |
128 | static struct fq_flow detached, throttled; | |
129 | ||
130 | static void fq_flow_set_detached(struct fq_flow *f) | |
131 | { | |
132 | f->next = &detached; | |
f52ed899 | 133 | f->age = jiffies; |
afe4fd06 ED |
134 | } |
135 | ||
136 | static bool fq_flow_is_detached(const struct fq_flow *f) | |
137 | { | |
138 | return f->next == &detached; | |
139 | } | |
140 | ||
7df40c26 ED |
141 | static bool fq_flow_is_throttled(const struct fq_flow *f) |
142 | { | |
143 | return f->next == &throttled; | |
144 | } | |
145 | ||
146 | static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) | |
147 | { | |
148 | if (head->first) | |
149 | head->last->next = flow; | |
150 | else | |
151 | head->first = flow; | |
152 | head->last = flow; | |
153 | flow->next = NULL; | |
154 | } | |
155 | ||
156 | static void fq_flow_unset_throttled(struct fq_sched_data *q, struct fq_flow *f) | |
157 | { | |
158 | rb_erase(&f->rate_node, &q->delayed); | |
159 | q->throttled_flows--; | |
160 | fq_flow_add_tail(&q->old_flows, f); | |
161 | } | |
162 | ||
afe4fd06 ED |
163 | static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) |
164 | { | |
165 | struct rb_node **p = &q->delayed.rb_node, *parent = NULL; | |
166 | ||
167 | while (*p) { | |
168 | struct fq_flow *aux; | |
169 | ||
170 | parent = *p; | |
e124557d | 171 | aux = rb_entry(parent, struct fq_flow, rate_node); |
afe4fd06 ED |
172 | if (f->time_next_packet >= aux->time_next_packet) |
173 | p = &parent->rb_right; | |
174 | else | |
175 | p = &parent->rb_left; | |
176 | } | |
177 | rb_link_node(&f->rate_node, parent, p); | |
178 | rb_insert_color(&f->rate_node, &q->delayed); | |
179 | q->throttled_flows++; | |
180 | q->stat_throttled++; | |
181 | ||
182 | f->next = &throttled; | |
183 | if (q->time_next_delayed_flow > f->time_next_packet) | |
184 | q->time_next_delayed_flow = f->time_next_packet; | |
185 | } | |
186 | ||
187 | ||
188 | static struct kmem_cache *fq_flow_cachep __read_mostly; | |
189 | ||
afe4fd06 ED |
190 | |
191 | /* limit number of collected flows per round */ | |
192 | #define FQ_GC_MAX 8 | |
193 | #define FQ_GC_AGE (3*HZ) | |
194 | ||
195 | static bool fq_gc_candidate(const struct fq_flow *f) | |
196 | { | |
197 | return fq_flow_is_detached(f) && | |
198 | time_after(jiffies, f->age + FQ_GC_AGE); | |
199 | } | |
200 | ||
201 | static void fq_gc(struct fq_sched_data *q, | |
202 | struct rb_root *root, | |
203 | struct sock *sk) | |
204 | { | |
205 | struct fq_flow *f, *tofree[FQ_GC_MAX]; | |
206 | struct rb_node **p, *parent; | |
207 | int fcnt = 0; | |
208 | ||
209 | p = &root->rb_node; | |
210 | parent = NULL; | |
211 | while (*p) { | |
212 | parent = *p; | |
213 | ||
e124557d | 214 | f = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
215 | if (f->sk == sk) |
216 | break; | |
217 | ||
218 | if (fq_gc_candidate(f)) { | |
219 | tofree[fcnt++] = f; | |
220 | if (fcnt == FQ_GC_MAX) | |
221 | break; | |
222 | } | |
223 | ||
224 | if (f->sk > sk) | |
225 | p = &parent->rb_right; | |
226 | else | |
227 | p = &parent->rb_left; | |
228 | } | |
229 | ||
230 | q->flows -= fcnt; | |
231 | q->inactive_flows -= fcnt; | |
232 | q->stat_gc_flows += fcnt; | |
233 | while (fcnt) { | |
234 | struct fq_flow *f = tofree[--fcnt]; | |
235 | ||
236 | rb_erase(&f->fq_node, root); | |
237 | kmem_cache_free(fq_flow_cachep, f); | |
238 | } | |
239 | } | |
240 | ||
afe4fd06 ED |
241 | static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) |
242 | { | |
243 | struct rb_node **p, *parent; | |
244 | struct sock *sk = skb->sk; | |
245 | struct rb_root *root; | |
246 | struct fq_flow *f; | |
afe4fd06 ED |
247 | |
248 | /* warning: no starvation prevention... */ | |
2abc2f07 | 249 | if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) |
afe4fd06 ED |
250 | return &q->internal; |
251 | ||
ca6fb065 | 252 | /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket |
e446f9df | 253 | * or a listener (SYNCOOKIE mode) |
ca6fb065 ED |
254 | * 1) request sockets are not full blown, |
255 | * they do not contain sk_pacing_rate | |
256 | * 2) They are not part of a 'flow' yet | |
257 | * 3) We do not want to rate limit them (eg SYNFLOOD attack), | |
06eb395f | 258 | * especially if the listener set SO_MAX_PACING_RATE |
ca6fb065 | 259 | * 4) We pretend they are orphaned |
06eb395f | 260 | */ |
e446f9df | 261 | if (!sk || sk_listener(sk)) { |
06eb395f ED |
262 | unsigned long hash = skb_get_hash(skb) & q->orphan_mask; |
263 | ||
afe4fd06 ED |
264 | /* By forcing low order bit to 1, we make sure to not |
265 | * collide with a local flow (socket pointers are word aligned) | |
266 | */ | |
06eb395f ED |
267 | sk = (struct sock *)((hash << 1) | 1UL); |
268 | skb_orphan(skb); | |
37c0aead ED |
269 | } else if (sk->sk_state == TCP_CLOSE) { |
270 | unsigned long hash = skb_get_hash(skb) & q->orphan_mask; | |
271 | /* | |
272 | * Sockets in TCP_CLOSE are non connected. | |
273 | * Typical use case is UDP sockets, they can send packets | |
274 | * with sendto() to many different destinations. | |
275 | * We probably could use a generic bit advertising | |
276 | * non connected sockets, instead of sk_state == TCP_CLOSE, | |
277 | * if we care enough. | |
278 | */ | |
279 | sk = (struct sock *)((hash << 1) | 1UL); | |
afe4fd06 ED |
280 | } |
281 | ||
29c58472 | 282 | root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)]; |
afe4fd06 ED |
283 | |
284 | if (q->flows >= (2U << q->fq_trees_log) && | |
285 | q->inactive_flows > q->flows/2) | |
286 | fq_gc(q, root, sk); | |
287 | ||
288 | p = &root->rb_node; | |
289 | parent = NULL; | |
290 | while (*p) { | |
291 | parent = *p; | |
292 | ||
e124557d | 293 | f = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
294 | if (f->sk == sk) { |
295 | /* socket might have been reallocated, so check | |
296 | * if its sk_hash is the same. | |
297 | * It not, we need to refill credit with | |
298 | * initial quantum | |
299 | */ | |
37c0aead | 300 | if (unlikely(skb->sk == sk && |
afe4fd06 ED |
301 | f->socket_hash != sk->sk_hash)) { |
302 | f->credit = q->initial_quantum; | |
303 | f->socket_hash = sk->sk_hash; | |
bb3d0b8b ED |
304 | if (q->rate_enable) |
305 | smp_store_release(&sk->sk_pacing_status, | |
306 | SK_PACING_FQ); | |
7df40c26 ED |
307 | if (fq_flow_is_throttled(f)) |
308 | fq_flow_unset_throttled(q, f); | |
fc59d5bd | 309 | f->time_next_packet = 0ULL; |
afe4fd06 ED |
310 | } |
311 | return f; | |
312 | } | |
313 | if (f->sk > sk) | |
314 | p = &parent->rb_right; | |
315 | else | |
316 | p = &parent->rb_left; | |
317 | } | |
318 | ||
319 | f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); | |
320 | if (unlikely(!f)) { | |
321 | q->stat_allocation_errors++; | |
322 | return &q->internal; | |
323 | } | |
eeb84aa0 ED |
324 | /* f->t_root is already zeroed after kmem_cache_zalloc() */ |
325 | ||
afe4fd06 ED |
326 | fq_flow_set_detached(f); |
327 | f->sk = sk; | |
bb3d0b8b | 328 | if (skb->sk == sk) { |
afe4fd06 | 329 | f->socket_hash = sk->sk_hash; |
bb3d0b8b ED |
330 | if (q->rate_enable) |
331 | smp_store_release(&sk->sk_pacing_status, | |
332 | SK_PACING_FQ); | |
333 | } | |
afe4fd06 ED |
334 | f->credit = q->initial_quantum; |
335 | ||
336 | rb_link_node(&f->fq_node, parent, p); | |
337 | rb_insert_color(&f->fq_node, root); | |
338 | ||
339 | q->flows++; | |
340 | q->inactive_flows++; | |
341 | return f; | |
342 | } | |
343 | ||
eeb84aa0 ED |
344 | static struct sk_buff *fq_peek(struct fq_flow *flow) |
345 | { | |
346 | struct sk_buff *skb = skb_rb_first(&flow->t_root); | |
347 | struct sk_buff *head = flow->head; | |
348 | ||
349 | if (!skb) | |
350 | return head; | |
351 | ||
352 | if (!head) | |
353 | return skb; | |
354 | ||
355 | if (fq_skb_cb(skb)->time_to_send < fq_skb_cb(head)->time_to_send) | |
356 | return skb; | |
357 | return head; | |
358 | } | |
359 | ||
360 | static void fq_erase_head(struct Qdisc *sch, struct fq_flow *flow, | |
361 | struct sk_buff *skb) | |
362 | { | |
363 | if (skb == flow->head) { | |
364 | flow->head = skb->next; | |
365 | } else { | |
366 | rb_erase(&skb->rbnode, &flow->t_root); | |
367 | skb->dev = qdisc_dev(sch); | |
368 | } | |
369 | } | |
afe4fd06 ED |
370 | |
371 | /* remove one skb from head of flow queue */ | |
8d34ce10 | 372 | static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow) |
afe4fd06 | 373 | { |
eeb84aa0 | 374 | struct sk_buff *skb = fq_peek(flow); |
afe4fd06 ED |
375 | |
376 | if (skb) { | |
eeb84aa0 | 377 | fq_erase_head(sch, flow, skb); |
a8305bff | 378 | skb_mark_not_on_list(skb); |
afe4fd06 | 379 | flow->qlen--; |
25331d6c | 380 | qdisc_qstats_backlog_dec(sch, skb); |
8d34ce10 | 381 | sch->q.qlen--; |
afe4fd06 ED |
382 | } |
383 | return skb; | |
384 | } | |
385 | ||
afe4fd06 ED |
386 | static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) |
387 | { | |
eeb84aa0 ED |
388 | struct rb_node **p, *parent; |
389 | struct sk_buff *head, *aux; | |
afe4fd06 | 390 | |
eeb84aa0 ED |
391 | fq_skb_cb(skb)->time_to_send = skb->tstamp ?: ktime_get_ns(); |
392 | ||
393 | head = flow->head; | |
394 | if (!head || | |
395 | fq_skb_cb(skb)->time_to_send >= fq_skb_cb(flow->tail)->time_to_send) { | |
396 | if (!head) | |
397 | flow->head = skb; | |
398 | else | |
399 | flow->tail->next = skb; | |
400 | flow->tail = skb; | |
401 | skb->next = NULL; | |
402 | return; | |
403 | } | |
404 | ||
405 | p = &flow->t_root.rb_node; | |
406 | parent = NULL; | |
afe4fd06 | 407 | |
eeb84aa0 ED |
408 | while (*p) { |
409 | parent = *p; | |
410 | aux = rb_to_skb(parent); | |
411 | if (fq_skb_cb(skb)->time_to_send >= fq_skb_cb(aux)->time_to_send) | |
412 | p = &parent->rb_right; | |
413 | else | |
414 | p = &parent->rb_left; | |
415 | } | |
416 | rb_link_node(&skb->rbnode, parent, p); | |
417 | rb_insert_color(&skb->rbnode, &flow->t_root); | |
afe4fd06 ED |
418 | } |
419 | ||
520ac30f ED |
420 | static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
421 | struct sk_buff **to_free) | |
afe4fd06 ED |
422 | { |
423 | struct fq_sched_data *q = qdisc_priv(sch); | |
424 | struct fq_flow *f; | |
425 | ||
426 | if (unlikely(sch->q.qlen >= sch->limit)) | |
520ac30f | 427 | return qdisc_drop(skb, sch, to_free); |
afe4fd06 ED |
428 | |
429 | f = fq_classify(skb, q); | |
430 | if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { | |
431 | q->stat_flows_plimit++; | |
520ac30f | 432 | return qdisc_drop(skb, sch, to_free); |
afe4fd06 ED |
433 | } |
434 | ||
435 | f->qlen++; | |
25331d6c | 436 | qdisc_qstats_backlog_inc(sch, skb); |
afe4fd06 ED |
437 | if (fq_flow_is_detached(f)) { |
438 | fq_flow_add_tail(&q->new_flows, f); | |
f52ed899 ED |
439 | if (time_after(jiffies, f->age + q->flow_refill_delay)) |
440 | f->credit = max_t(u32, f->credit, q->quantum); | |
afe4fd06 | 441 | q->inactive_flows--; |
afe4fd06 | 442 | } |
f52ed899 ED |
443 | |
444 | /* Note: this overwrites f->age */ | |
445 | flow_queue_add(f, skb); | |
446 | ||
afe4fd06 ED |
447 | if (unlikely(f == &q->internal)) { |
448 | q->stat_internal_packets++; | |
afe4fd06 ED |
449 | } |
450 | sch->q.qlen++; | |
451 | ||
452 | return NET_XMIT_SUCCESS; | |
453 | } | |
454 | ||
455 | static void fq_check_throttled(struct fq_sched_data *q, u64 now) | |
456 | { | |
fefa569a | 457 | unsigned long sample; |
afe4fd06 ED |
458 | struct rb_node *p; |
459 | ||
460 | if (q->time_next_delayed_flow > now) | |
461 | return; | |
462 | ||
fefa569a ED |
463 | /* Update unthrottle latency EWMA. |
464 | * This is cheap and can help diagnosing timer/latency problems. | |
465 | */ | |
466 | sample = (unsigned long)(now - q->time_next_delayed_flow); | |
467 | q->unthrottle_latency_ns -= q->unthrottle_latency_ns >> 3; | |
468 | q->unthrottle_latency_ns += sample >> 3; | |
469 | ||
afe4fd06 ED |
470 | q->time_next_delayed_flow = ~0ULL; |
471 | while ((p = rb_first(&q->delayed)) != NULL) { | |
e124557d | 472 | struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node); |
afe4fd06 ED |
473 | |
474 | if (f->time_next_packet > now) { | |
475 | q->time_next_delayed_flow = f->time_next_packet; | |
476 | break; | |
477 | } | |
7df40c26 | 478 | fq_flow_unset_throttled(q, f); |
afe4fd06 ED |
479 | } |
480 | } | |
481 | ||
482 | static struct sk_buff *fq_dequeue(struct Qdisc *sch) | |
483 | { | |
484 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 ED |
485 | struct fq_flow_head *head; |
486 | struct sk_buff *skb; | |
487 | struct fq_flow *f; | |
76a9ebe8 ED |
488 | unsigned long rate; |
489 | u32 plen; | |
6b015a52 ED |
490 | u64 now; |
491 | ||
492 | if (!sch->q.qlen) | |
493 | return NULL; | |
afe4fd06 | 494 | |
8d34ce10 | 495 | skb = fq_dequeue_head(sch, &q->internal); |
afe4fd06 ED |
496 | if (skb) |
497 | goto out; | |
6b015a52 ED |
498 | |
499 | now = ktime_get_ns(); | |
afe4fd06 ED |
500 | fq_check_throttled(q, now); |
501 | begin: | |
502 | head = &q->new_flows; | |
503 | if (!head->first) { | |
504 | head = &q->old_flows; | |
505 | if (!head->first) { | |
506 | if (q->time_next_delayed_flow != ~0ULL) | |
507 | qdisc_watchdog_schedule_ns(&q->watchdog, | |
45f50bed | 508 | q->time_next_delayed_flow); |
afe4fd06 ED |
509 | return NULL; |
510 | } | |
511 | } | |
512 | f = head->first; | |
513 | ||
514 | if (f->credit <= 0) { | |
515 | f->credit += q->quantum; | |
516 | head->first = f->next; | |
517 | fq_flow_add_tail(&q->old_flows, f); | |
518 | goto begin; | |
519 | } | |
520 | ||
eeb84aa0 | 521 | skb = fq_peek(f); |
7baf33bd | 522 | if (skb) { |
eeb84aa0 | 523 | u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send, |
ab408b6d ED |
524 | f->time_next_packet); |
525 | ||
526 | if (now < time_next_packet) { | |
527 | head->first = f->next; | |
528 | f->time_next_packet = time_next_packet; | |
529 | fq_flow_set_throttled(q, f); | |
530 | goto begin; | |
531 | } | |
e9c43add | 532 | if ((s64)(now - time_next_packet - q->ce_threshold) > 0) { |
48872c11 ED |
533 | INET_ECN_set_ce(skb); |
534 | q->stat_ce_mark++; | |
535 | } | |
afe4fd06 ED |
536 | } |
537 | ||
8d34ce10 | 538 | skb = fq_dequeue_head(sch, f); |
afe4fd06 ED |
539 | if (!skb) { |
540 | head->first = f->next; | |
541 | /* force a pass through old_flows to prevent starvation */ | |
542 | if ((head == &q->new_flows) && q->old_flows.first) { | |
543 | fq_flow_add_tail(&q->old_flows, f); | |
544 | } else { | |
545 | fq_flow_set_detached(f); | |
afe4fd06 ED |
546 | q->inactive_flows++; |
547 | } | |
548 | goto begin; | |
549 | } | |
08f89b98 | 550 | prefetch(&skb->end); |
08e14fe4 ED |
551 | plen = qdisc_pkt_len(skb); |
552 | f->credit -= plen; | |
afe4fd06 | 553 | |
08e14fe4 | 554 | if (!q->rate_enable) |
98781965 ED |
555 | goto out; |
556 | ||
7eec4174 | 557 | rate = q->flow_max_rate; |
08e14fe4 ED |
558 | |
559 | /* If EDT time was provided for this skb, we need to | |
560 | * update f->time_next_packet only if this qdisc enforces | |
561 | * a flow max rate. | |
562 | */ | |
563 | if (!skb->tstamp) { | |
564 | if (skb->sk) | |
565 | rate = min(skb->sk->sk_pacing_rate, rate); | |
566 | ||
567 | if (rate <= q->low_rate_threshold) { | |
568 | f->credit = 0; | |
569 | } else { | |
570 | plen = max(plen, q->quantum); | |
571 | if (f->credit > 0) | |
572 | goto out; | |
573 | } | |
77879147 | 574 | } |
76a9ebe8 | 575 | if (rate != ~0UL) { |
0eab5eb7 ED |
576 | u64 len = (u64)plen * NSEC_PER_SEC; |
577 | ||
7eec4174 | 578 | if (likely(rate)) |
76a9ebe8 | 579 | len = div64_ul(len, rate); |
0eab5eb7 | 580 | /* Since socket rate can change later, |
ced7a04e ED |
581 | * clamp the delay to 1 second. |
582 | * Really, providers of too big packets should be fixed ! | |
0eab5eb7 | 583 | */ |
ced7a04e ED |
584 | if (unlikely(len > NSEC_PER_SEC)) { |
585 | len = NSEC_PER_SEC; | |
0eab5eb7 | 586 | q->stat_pkts_too_long++; |
afe4fd06 | 587 | } |
fefa569a ED |
588 | /* Account for schedule/timers drifts. |
589 | * f->time_next_packet was set when prior packet was sent, | |
590 | * and current time (@now) can be too late by tens of us. | |
591 | */ | |
592 | if (f->time_next_packet) | |
593 | len -= min(len/2, now - f->time_next_packet); | |
0eab5eb7 | 594 | f->time_next_packet = now + len; |
afe4fd06 ED |
595 | } |
596 | out: | |
afe4fd06 | 597 | qdisc_bstats_update(sch, skb); |
afe4fd06 ED |
598 | return skb; |
599 | } | |
600 | ||
e14ffdfd ED |
601 | static void fq_flow_purge(struct fq_flow *flow) |
602 | { | |
eeb84aa0 ED |
603 | struct rb_node *p = rb_first(&flow->t_root); |
604 | ||
605 | while (p) { | |
606 | struct sk_buff *skb = rb_to_skb(p); | |
607 | ||
608 | p = rb_next(p); | |
609 | rb_erase(&skb->rbnode, &flow->t_root); | |
610 | rtnl_kfree_skbs(skb, skb); | |
611 | } | |
e14ffdfd ED |
612 | rtnl_kfree_skbs(flow->head, flow->tail); |
613 | flow->head = NULL; | |
614 | flow->qlen = 0; | |
615 | } | |
616 | ||
afe4fd06 ED |
617 | static void fq_reset(struct Qdisc *sch) |
618 | { | |
8d34ce10 ED |
619 | struct fq_sched_data *q = qdisc_priv(sch); |
620 | struct rb_root *root; | |
8d34ce10 ED |
621 | struct rb_node *p; |
622 | struct fq_flow *f; | |
623 | unsigned int idx; | |
afe4fd06 | 624 | |
e14ffdfd ED |
625 | sch->q.qlen = 0; |
626 | sch->qstats.backlog = 0; | |
627 | ||
628 | fq_flow_purge(&q->internal); | |
8d34ce10 ED |
629 | |
630 | if (!q->fq_root) | |
631 | return; | |
632 | ||
633 | for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { | |
634 | root = &q->fq_root[idx]; | |
635 | while ((p = rb_first(root)) != NULL) { | |
e124557d | 636 | f = rb_entry(p, struct fq_flow, fq_node); |
8d34ce10 ED |
637 | rb_erase(p, root); |
638 | ||
e14ffdfd | 639 | fq_flow_purge(f); |
8d34ce10 ED |
640 | |
641 | kmem_cache_free(fq_flow_cachep, f); | |
642 | } | |
643 | } | |
644 | q->new_flows.first = NULL; | |
645 | q->old_flows.first = NULL; | |
646 | q->delayed = RB_ROOT; | |
647 | q->flows = 0; | |
648 | q->inactive_flows = 0; | |
649 | q->throttled_flows = 0; | |
afe4fd06 ED |
650 | } |
651 | ||
652 | static void fq_rehash(struct fq_sched_data *q, | |
653 | struct rb_root *old_array, u32 old_log, | |
654 | struct rb_root *new_array, u32 new_log) | |
655 | { | |
656 | struct rb_node *op, **np, *parent; | |
657 | struct rb_root *oroot, *nroot; | |
658 | struct fq_flow *of, *nf; | |
659 | int fcnt = 0; | |
660 | u32 idx; | |
661 | ||
662 | for (idx = 0; idx < (1U << old_log); idx++) { | |
663 | oroot = &old_array[idx]; | |
664 | while ((op = rb_first(oroot)) != NULL) { | |
665 | rb_erase(op, oroot); | |
e124557d | 666 | of = rb_entry(op, struct fq_flow, fq_node); |
afe4fd06 ED |
667 | if (fq_gc_candidate(of)) { |
668 | fcnt++; | |
669 | kmem_cache_free(fq_flow_cachep, of); | |
670 | continue; | |
671 | } | |
29c58472 | 672 | nroot = &new_array[hash_ptr(of->sk, new_log)]; |
afe4fd06 ED |
673 | |
674 | np = &nroot->rb_node; | |
675 | parent = NULL; | |
676 | while (*np) { | |
677 | parent = *np; | |
678 | ||
e124557d | 679 | nf = rb_entry(parent, struct fq_flow, fq_node); |
afe4fd06 ED |
680 | BUG_ON(nf->sk == of->sk); |
681 | ||
682 | if (nf->sk > of->sk) | |
683 | np = &parent->rb_right; | |
684 | else | |
685 | np = &parent->rb_left; | |
686 | } | |
687 | ||
688 | rb_link_node(&of->fq_node, parent, np); | |
689 | rb_insert_color(&of->fq_node, nroot); | |
690 | } | |
691 | } | |
692 | q->flows -= fcnt; | |
693 | q->inactive_flows -= fcnt; | |
694 | q->stat_gc_flows += fcnt; | |
695 | } | |
696 | ||
c3bd8549 ED |
697 | static void fq_free(void *addr) |
698 | { | |
4cb28970 | 699 | kvfree(addr); |
c3bd8549 ED |
700 | } |
701 | ||
702 | static int fq_resize(struct Qdisc *sch, u32 log) | |
703 | { | |
704 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 705 | struct rb_root *array; |
2d8d40af | 706 | void *old_fq_root; |
afe4fd06 ED |
707 | u32 idx; |
708 | ||
709 | if (q->fq_root && log == q->fq_trees_log) | |
710 | return 0; | |
711 | ||
c3bd8549 | 712 | /* If XPS was setup, we can allocate memory on right NUMA node */ |
dcda9b04 | 713 | array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL, |
c3bd8549 | 714 | netdev_queue_numa_node_read(sch->dev_queue)); |
afe4fd06 ED |
715 | if (!array) |
716 | return -ENOMEM; | |
717 | ||
718 | for (idx = 0; idx < (1U << log); idx++) | |
719 | array[idx] = RB_ROOT; | |
720 | ||
2d8d40af ED |
721 | sch_tree_lock(sch); |
722 | ||
723 | old_fq_root = q->fq_root; | |
724 | if (old_fq_root) | |
725 | fq_rehash(q, old_fq_root, q->fq_trees_log, array, log); | |
726 | ||
afe4fd06 ED |
727 | q->fq_root = array; |
728 | q->fq_trees_log = log; | |
729 | ||
2d8d40af ED |
730 | sch_tree_unlock(sch); |
731 | ||
732 | fq_free(old_fq_root); | |
733 | ||
afe4fd06 ED |
734 | return 0; |
735 | } | |
736 | ||
737 | static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { | |
738 | [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, | |
739 | [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, | |
740 | [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, | |
741 | [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, | |
742 | [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, | |
743 | [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, | |
744 | [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, | |
745 | [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, | |
f52ed899 | 746 | [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, |
77879147 | 747 | [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 }, |
48872c11 | 748 | [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 }, |
afe4fd06 ED |
749 | }; |
750 | ||
2030721c AA |
751 | static int fq_change(struct Qdisc *sch, struct nlattr *opt, |
752 | struct netlink_ext_ack *extack) | |
afe4fd06 ED |
753 | { |
754 | struct fq_sched_data *q = qdisc_priv(sch); | |
755 | struct nlattr *tb[TCA_FQ_MAX + 1]; | |
756 | int err, drop_count = 0; | |
2ccccf5f | 757 | unsigned drop_len = 0; |
afe4fd06 ED |
758 | u32 fq_log; |
759 | ||
760 | if (!opt) | |
761 | return -EINVAL; | |
762 | ||
8cb08174 JB |
763 | err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy, |
764 | NULL); | |
afe4fd06 ED |
765 | if (err < 0) |
766 | return err; | |
767 | ||
768 | sch_tree_lock(sch); | |
769 | ||
770 | fq_log = q->fq_trees_log; | |
771 | ||
772 | if (tb[TCA_FQ_BUCKETS_LOG]) { | |
773 | u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); | |
774 | ||
775 | if (nval >= 1 && nval <= ilog2(256*1024)) | |
776 | fq_log = nval; | |
777 | else | |
778 | err = -EINVAL; | |
779 | } | |
780 | if (tb[TCA_FQ_PLIMIT]) | |
781 | sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); | |
782 | ||
783 | if (tb[TCA_FQ_FLOW_PLIMIT]) | |
784 | q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); | |
785 | ||
3725a269 KKJ |
786 | if (tb[TCA_FQ_QUANTUM]) { |
787 | u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); | |
788 | ||
789 | if (quantum > 0) | |
790 | q->quantum = quantum; | |
791 | else | |
792 | err = -EINVAL; | |
793 | } | |
afe4fd06 ED |
794 | |
795 | if (tb[TCA_FQ_INITIAL_QUANTUM]) | |
ede869cd | 796 | q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); |
afe4fd06 ED |
797 | |
798 | if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) | |
65c5189a ED |
799 | pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", |
800 | nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); | |
afe4fd06 | 801 | |
76a9ebe8 ED |
802 | if (tb[TCA_FQ_FLOW_MAX_RATE]) { |
803 | u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); | |
afe4fd06 | 804 | |
76a9ebe8 ED |
805 | q->flow_max_rate = (rate == ~0U) ? ~0UL : rate; |
806 | } | |
77879147 ED |
807 | if (tb[TCA_FQ_LOW_RATE_THRESHOLD]) |
808 | q->low_rate_threshold = | |
809 | nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]); | |
810 | ||
afe4fd06 ED |
811 | if (tb[TCA_FQ_RATE_ENABLE]) { |
812 | u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); | |
813 | ||
814 | if (enable <= 1) | |
815 | q->rate_enable = enable; | |
816 | else | |
817 | err = -EINVAL; | |
818 | } | |
819 | ||
f52ed899 ED |
820 | if (tb[TCA_FQ_FLOW_REFILL_DELAY]) { |
821 | u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ; | |
822 | ||
823 | q->flow_refill_delay = usecs_to_jiffies(usecs_delay); | |
824 | } | |
825 | ||
06eb395f ED |
826 | if (tb[TCA_FQ_ORPHAN_MASK]) |
827 | q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]); | |
828 | ||
48872c11 ED |
829 | if (tb[TCA_FQ_CE_THRESHOLD]) |
830 | q->ce_threshold = (u64)NSEC_PER_USEC * | |
831 | nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]); | |
832 | ||
2d8d40af ED |
833 | if (!err) { |
834 | sch_tree_unlock(sch); | |
c3bd8549 | 835 | err = fq_resize(sch, fq_log); |
2d8d40af ED |
836 | sch_tree_lock(sch); |
837 | } | |
afe4fd06 ED |
838 | while (sch->q.qlen > sch->limit) { |
839 | struct sk_buff *skb = fq_dequeue(sch); | |
840 | ||
8d34ce10 ED |
841 | if (!skb) |
842 | break; | |
2ccccf5f | 843 | drop_len += qdisc_pkt_len(skb); |
e14ffdfd | 844 | rtnl_kfree_skbs(skb, skb); |
afe4fd06 ED |
845 | drop_count++; |
846 | } | |
2ccccf5f | 847 | qdisc_tree_reduce_backlog(sch, drop_count, drop_len); |
afe4fd06 ED |
848 | |
849 | sch_tree_unlock(sch); | |
850 | return err; | |
851 | } | |
852 | ||
853 | static void fq_destroy(struct Qdisc *sch) | |
854 | { | |
855 | struct fq_sched_data *q = qdisc_priv(sch); | |
afe4fd06 | 856 | |
8d34ce10 | 857 | fq_reset(sch); |
c3bd8549 | 858 | fq_free(q->fq_root); |
afe4fd06 ED |
859 | qdisc_watchdog_cancel(&q->watchdog); |
860 | } | |
861 | ||
e63d7dfd AA |
862 | static int fq_init(struct Qdisc *sch, struct nlattr *opt, |
863 | struct netlink_ext_ack *extack) | |
afe4fd06 ED |
864 | { |
865 | struct fq_sched_data *q = qdisc_priv(sch); | |
866 | int err; | |
867 | ||
868 | sch->limit = 10000; | |
869 | q->flow_plimit = 100; | |
870 | q->quantum = 2 * psched_mtu(qdisc_dev(sch)); | |
871 | q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); | |
f52ed899 | 872 | q->flow_refill_delay = msecs_to_jiffies(40); |
76a9ebe8 | 873 | q->flow_max_rate = ~0UL; |
fefa569a | 874 | q->time_next_delayed_flow = ~0ULL; |
afe4fd06 ED |
875 | q->rate_enable = 1; |
876 | q->new_flows.first = NULL; | |
877 | q->old_flows.first = NULL; | |
878 | q->delayed = RB_ROOT; | |
879 | q->fq_root = NULL; | |
880 | q->fq_trees_log = ilog2(1024); | |
06eb395f | 881 | q->orphan_mask = 1024 - 1; |
77879147 | 882 | q->low_rate_threshold = 550000 / 8; |
48872c11 ED |
883 | |
884 | /* Default ce_threshold of 4294 seconds */ | |
885 | q->ce_threshold = (u64)NSEC_PER_USEC * ~0U; | |
886 | ||
fb420d5d | 887 | qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC); |
afe4fd06 ED |
888 | |
889 | if (opt) | |
2030721c | 890 | err = fq_change(sch, opt, extack); |
afe4fd06 | 891 | else |
c3bd8549 | 892 | err = fq_resize(sch, q->fq_trees_log); |
afe4fd06 ED |
893 | |
894 | return err; | |
895 | } | |
896 | ||
897 | static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) | |
898 | { | |
899 | struct fq_sched_data *q = qdisc_priv(sch); | |
48872c11 | 900 | u64 ce_threshold = q->ce_threshold; |
afe4fd06 ED |
901 | struct nlattr *opts; |
902 | ||
ae0be8de | 903 | opts = nla_nest_start_noflag(skb, TCA_OPTIONS); |
afe4fd06 ED |
904 | if (opts == NULL) |
905 | goto nla_put_failure; | |
906 | ||
65c5189a ED |
907 | /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ |
908 | ||
48872c11 ED |
909 | do_div(ce_threshold, NSEC_PER_USEC); |
910 | ||
afe4fd06 ED |
911 | if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || |
912 | nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || | |
913 | nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || | |
914 | nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || | |
915 | nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || | |
76a9ebe8 ED |
916 | nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, |
917 | min_t(unsigned long, q->flow_max_rate, ~0U)) || | |
f52ed899 ED |
918 | nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, |
919 | jiffies_to_usecs(q->flow_refill_delay)) || | |
06eb395f | 920 | nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) || |
77879147 ED |
921 | nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD, |
922 | q->low_rate_threshold) || | |
48872c11 | 923 | nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) || |
afe4fd06 ED |
924 | nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) |
925 | goto nla_put_failure; | |
926 | ||
d59b7d80 | 927 | return nla_nest_end(skb, opts); |
afe4fd06 ED |
928 | |
929 | nla_put_failure: | |
930 | return -1; | |
931 | } | |
932 | ||
933 | static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) | |
934 | { | |
935 | struct fq_sched_data *q = qdisc_priv(sch); | |
695b4ec0 ED |
936 | struct tc_fq_qd_stats st; |
937 | ||
938 | sch_tree_lock(sch); | |
939 | ||
940 | st.gc_flows = q->stat_gc_flows; | |
941 | st.highprio_packets = q->stat_internal_packets; | |
90caf67b | 942 | st.tcp_retrans = 0; |
695b4ec0 ED |
943 | st.throttled = q->stat_throttled; |
944 | st.flows_plimit = q->stat_flows_plimit; | |
945 | st.pkts_too_long = q->stat_pkts_too_long; | |
946 | st.allocation_errors = q->stat_allocation_errors; | |
fb420d5d | 947 | st.time_next_delayed_flow = q->time_next_delayed_flow - ktime_get_ns(); |
695b4ec0 ED |
948 | st.flows = q->flows; |
949 | st.inactive_flows = q->inactive_flows; | |
950 | st.throttled_flows = q->throttled_flows; | |
fefa569a ED |
951 | st.unthrottle_latency_ns = min_t(unsigned long, |
952 | q->unthrottle_latency_ns, ~0U); | |
48872c11 | 953 | st.ce_mark = q->stat_ce_mark; |
695b4ec0 | 954 | sch_tree_unlock(sch); |
afe4fd06 ED |
955 | |
956 | return gnet_stats_copy_app(d, &st, sizeof(st)); | |
957 | } | |
958 | ||
959 | static struct Qdisc_ops fq_qdisc_ops __read_mostly = { | |
960 | .id = "fq", | |
961 | .priv_size = sizeof(struct fq_sched_data), | |
962 | ||
963 | .enqueue = fq_enqueue, | |
964 | .dequeue = fq_dequeue, | |
965 | .peek = qdisc_peek_dequeued, | |
966 | .init = fq_init, | |
967 | .reset = fq_reset, | |
968 | .destroy = fq_destroy, | |
969 | .change = fq_change, | |
970 | .dump = fq_dump, | |
971 | .dump_stats = fq_dump_stats, | |
972 | .owner = THIS_MODULE, | |
973 | }; | |
974 | ||
975 | static int __init fq_module_init(void) | |
976 | { | |
977 | int ret; | |
978 | ||
979 | fq_flow_cachep = kmem_cache_create("fq_flow_cache", | |
980 | sizeof(struct fq_flow), | |
981 | 0, 0, NULL); | |
982 | if (!fq_flow_cachep) | |
983 | return -ENOMEM; | |
984 | ||
985 | ret = register_qdisc(&fq_qdisc_ops); | |
986 | if (ret) | |
987 | kmem_cache_destroy(fq_flow_cachep); | |
988 | return ret; | |
989 | } | |
990 | ||
991 | static void __exit fq_module_exit(void) | |
992 | { | |
993 | unregister_qdisc(&fq_qdisc_ops); | |
994 | kmem_cache_destroy(fq_flow_cachep); | |
995 | } | |
996 | ||
997 | module_init(fq_module_init) | |
998 | module_exit(fq_module_exit) | |
999 | MODULE_AUTHOR("Eric Dumazet"); | |
1000 | MODULE_LICENSE("GPL"); |