Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 1 | /* |
| 2 | * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing) |
| 3 | * |
| 4 | * Copyright (C) 2013 Eric Dumazet <edumazet@google.com> |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or |
| 7 | * modify it under the terms of the GNU General Public License |
| 8 | * as published by the Free Software Foundation; either version |
| 9 | * 2 of the License, or (at your option) any later version. |
| 10 | * |
| 11 | * Meant to be mostly used for localy generated traffic : |
| 12 | * Fast classification depends on skb->sk being set before reaching us. |
| 13 | * If not, (router workload), we use rxhash as fallback, with 32 bits wide hash. |
| 14 | * All packets belonging to a socket are considered as a 'flow'. |
| 15 | * |
| 16 | * Flows are dynamically allocated and stored in a hash table of RB trees |
| 17 | * They are also part of one Round Robin 'queues' (new or old flows) |
| 18 | * |
| 19 | * Burst avoidance (aka pacing) capability : |
| 20 | * |
| 21 | * Transport (eg TCP) can set in sk->sk_pacing_rate a rate, enqueue a |
| 22 | * bunch of packets, and this packet scheduler adds delay between |
| 23 | * packets to respect rate limitation. |
| 24 | * |
| 25 | * enqueue() : |
| 26 | * - lookup one RB tree (out of 1024 or more) to find the flow. |
| 27 | * If non existent flow, create it, add it to the tree. |
| 28 | * Add skb to the per flow list of skb (fifo). |
| 29 | * - Use a special fifo for high prio packets |
| 30 | * |
| 31 | * dequeue() : serves flows in Round Robin |
| 32 | * Note : When a flow becomes empty, we do not immediately remove it from |
| 33 | * rb trees, for performance reasons (its expected to send additional packets, |
| 34 | * or SLAB cache will reuse socket for another flow) |
| 35 | */ |
| 36 | |
| 37 | #include <linux/module.h> |
| 38 | #include <linux/types.h> |
| 39 | #include <linux/kernel.h> |
| 40 | #include <linux/jiffies.h> |
| 41 | #include <linux/string.h> |
| 42 | #include <linux/in.h> |
| 43 | #include <linux/errno.h> |
| 44 | #include <linux/init.h> |
| 45 | #include <linux/skbuff.h> |
| 46 | #include <linux/slab.h> |
| 47 | #include <linux/rbtree.h> |
| 48 | #include <linux/hash.h> |
Eric Dumazet | 08f89b9 | 2013-08-30 09:46:43 -0700 | [diff] [blame] | 49 | #include <linux/prefetch.h> |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 50 | #include <linux/vmalloc.h> |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 51 | #include <net/netlink.h> |
| 52 | #include <net/pkt_sched.h> |
| 53 | #include <net/sock.h> |
| 54 | #include <net/tcp_states.h> |
| 55 | |
| 56 | /* |
| 57 | * Per flow structure, dynamically allocated |
| 58 | */ |
| 59 | struct fq_flow { |
| 60 | struct sk_buff *head; /* list of skbs for this flow : first skb */ |
| 61 | union { |
| 62 | struct sk_buff *tail; /* last skb in the list */ |
| 63 | unsigned long age; /* jiffies when flow was emptied, for gc */ |
| 64 | }; |
| 65 | struct rb_node fq_node; /* anchor in fq_root[] trees */ |
| 66 | struct sock *sk; |
| 67 | int qlen; /* number of packets in flow queue */ |
| 68 | int credit; |
| 69 | u32 socket_hash; /* sk_hash */ |
| 70 | struct fq_flow *next; /* next pointer in RR lists, or &detached */ |
| 71 | |
| 72 | struct rb_node rate_node; /* anchor in q->delayed tree */ |
| 73 | u64 time_next_packet; |
| 74 | }; |
| 75 | |
| 76 | struct fq_flow_head { |
| 77 | struct fq_flow *first; |
| 78 | struct fq_flow *last; |
| 79 | }; |
| 80 | |
| 81 | struct fq_sched_data { |
| 82 | struct fq_flow_head new_flows; |
| 83 | |
| 84 | struct fq_flow_head old_flows; |
| 85 | |
| 86 | struct rb_root delayed; /* for rate limited flows */ |
| 87 | u64 time_next_delayed_flow; |
| 88 | |
| 89 | struct fq_flow internal; /* for non classified or high prio packets */ |
| 90 | u32 quantum; |
| 91 | u32 initial_quantum; |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 92 | u32 flow_refill_delay; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 93 | u32 flow_max_rate; /* optional max rate per flow */ |
| 94 | u32 flow_plimit; /* max packets per flow */ |
| 95 | struct rb_root *fq_root; |
| 96 | u8 rate_enable; |
| 97 | u8 fq_trees_log; |
| 98 | |
| 99 | u32 flows; |
| 100 | u32 inactive_flows; |
| 101 | u32 throttled_flows; |
| 102 | |
| 103 | u64 stat_gc_flows; |
| 104 | u64 stat_internal_packets; |
| 105 | u64 stat_tcp_retrans; |
| 106 | u64 stat_throttled; |
| 107 | u64 stat_flows_plimit; |
| 108 | u64 stat_pkts_too_long; |
| 109 | u64 stat_allocation_errors; |
| 110 | struct qdisc_watchdog watchdog; |
| 111 | }; |
| 112 | |
| 113 | /* special value to mark a detached flow (not on old/new list) */ |
| 114 | static struct fq_flow detached, throttled; |
| 115 | |
| 116 | static void fq_flow_set_detached(struct fq_flow *f) |
| 117 | { |
| 118 | f->next = &detached; |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 119 | f->age = jiffies; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 120 | } |
| 121 | |
| 122 | static bool fq_flow_is_detached(const struct fq_flow *f) |
| 123 | { |
| 124 | return f->next == &detached; |
| 125 | } |
| 126 | |
| 127 | static void fq_flow_set_throttled(struct fq_sched_data *q, struct fq_flow *f) |
| 128 | { |
| 129 | struct rb_node **p = &q->delayed.rb_node, *parent = NULL; |
| 130 | |
| 131 | while (*p) { |
| 132 | struct fq_flow *aux; |
| 133 | |
| 134 | parent = *p; |
| 135 | aux = container_of(parent, struct fq_flow, rate_node); |
| 136 | if (f->time_next_packet >= aux->time_next_packet) |
| 137 | p = &parent->rb_right; |
| 138 | else |
| 139 | p = &parent->rb_left; |
| 140 | } |
| 141 | rb_link_node(&f->rate_node, parent, p); |
| 142 | rb_insert_color(&f->rate_node, &q->delayed); |
| 143 | q->throttled_flows++; |
| 144 | q->stat_throttled++; |
| 145 | |
| 146 | f->next = &throttled; |
| 147 | if (q->time_next_delayed_flow > f->time_next_packet) |
| 148 | q->time_next_delayed_flow = f->time_next_packet; |
| 149 | } |
| 150 | |
| 151 | |
| 152 | static struct kmem_cache *fq_flow_cachep __read_mostly; |
| 153 | |
| 154 | static void fq_flow_add_tail(struct fq_flow_head *head, struct fq_flow *flow) |
| 155 | { |
| 156 | if (head->first) |
| 157 | head->last->next = flow; |
| 158 | else |
| 159 | head->first = flow; |
| 160 | head->last = flow; |
| 161 | flow->next = NULL; |
| 162 | } |
| 163 | |
| 164 | /* limit number of collected flows per round */ |
| 165 | #define FQ_GC_MAX 8 |
| 166 | #define FQ_GC_AGE (3*HZ) |
| 167 | |
| 168 | static bool fq_gc_candidate(const struct fq_flow *f) |
| 169 | { |
| 170 | return fq_flow_is_detached(f) && |
| 171 | time_after(jiffies, f->age + FQ_GC_AGE); |
| 172 | } |
| 173 | |
| 174 | static void fq_gc(struct fq_sched_data *q, |
| 175 | struct rb_root *root, |
| 176 | struct sock *sk) |
| 177 | { |
| 178 | struct fq_flow *f, *tofree[FQ_GC_MAX]; |
| 179 | struct rb_node **p, *parent; |
| 180 | int fcnt = 0; |
| 181 | |
| 182 | p = &root->rb_node; |
| 183 | parent = NULL; |
| 184 | while (*p) { |
| 185 | parent = *p; |
| 186 | |
| 187 | f = container_of(parent, struct fq_flow, fq_node); |
| 188 | if (f->sk == sk) |
| 189 | break; |
| 190 | |
| 191 | if (fq_gc_candidate(f)) { |
| 192 | tofree[fcnt++] = f; |
| 193 | if (fcnt == FQ_GC_MAX) |
| 194 | break; |
| 195 | } |
| 196 | |
| 197 | if (f->sk > sk) |
| 198 | p = &parent->rb_right; |
| 199 | else |
| 200 | p = &parent->rb_left; |
| 201 | } |
| 202 | |
| 203 | q->flows -= fcnt; |
| 204 | q->inactive_flows -= fcnt; |
| 205 | q->stat_gc_flows += fcnt; |
| 206 | while (fcnt) { |
| 207 | struct fq_flow *f = tofree[--fcnt]; |
| 208 | |
| 209 | rb_erase(&f->fq_node, root); |
| 210 | kmem_cache_free(fq_flow_cachep, f); |
| 211 | } |
| 212 | } |
| 213 | |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 214 | static struct fq_flow *fq_classify(struct sk_buff *skb, struct fq_sched_data *q) |
| 215 | { |
| 216 | struct rb_node **p, *parent; |
| 217 | struct sock *sk = skb->sk; |
| 218 | struct rb_root *root; |
| 219 | struct fq_flow *f; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 220 | |
| 221 | /* warning: no starvation prevention... */ |
Maciej Żenczykowski | 2abc2f0 | 2013-11-14 08:50:43 -0800 | [diff] [blame] | 222 | if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL)) |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 223 | return &q->internal; |
| 224 | |
| 225 | if (unlikely(!sk)) { |
| 226 | /* By forcing low order bit to 1, we make sure to not |
| 227 | * collide with a local flow (socket pointers are word aligned) |
| 228 | */ |
Tom Herbert | 3958afa1b | 2013-12-15 22:12:06 -0800 | [diff] [blame] | 229 | sk = (struct sock *)(skb_get_hash(skb) | 1L); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 230 | } |
| 231 | |
| 232 | root = &q->fq_root[hash_32((u32)(long)sk, q->fq_trees_log)]; |
| 233 | |
| 234 | if (q->flows >= (2U << q->fq_trees_log) && |
| 235 | q->inactive_flows > q->flows/2) |
| 236 | fq_gc(q, root, sk); |
| 237 | |
| 238 | p = &root->rb_node; |
| 239 | parent = NULL; |
| 240 | while (*p) { |
| 241 | parent = *p; |
| 242 | |
| 243 | f = container_of(parent, struct fq_flow, fq_node); |
| 244 | if (f->sk == sk) { |
| 245 | /* socket might have been reallocated, so check |
| 246 | * if its sk_hash is the same. |
| 247 | * It not, we need to refill credit with |
| 248 | * initial quantum |
| 249 | */ |
| 250 | if (unlikely(skb->sk && |
| 251 | f->socket_hash != sk->sk_hash)) { |
| 252 | f->credit = q->initial_quantum; |
| 253 | f->socket_hash = sk->sk_hash; |
Eric Dumazet | fc59d5b | 2013-10-27 16:26:43 -0700 | [diff] [blame] | 254 | f->time_next_packet = 0ULL; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 255 | } |
| 256 | return f; |
| 257 | } |
| 258 | if (f->sk > sk) |
| 259 | p = &parent->rb_right; |
| 260 | else |
| 261 | p = &parent->rb_left; |
| 262 | } |
| 263 | |
| 264 | f = kmem_cache_zalloc(fq_flow_cachep, GFP_ATOMIC | __GFP_NOWARN); |
| 265 | if (unlikely(!f)) { |
| 266 | q->stat_allocation_errors++; |
| 267 | return &q->internal; |
| 268 | } |
| 269 | fq_flow_set_detached(f); |
| 270 | f->sk = sk; |
| 271 | if (skb->sk) |
| 272 | f->socket_hash = sk->sk_hash; |
| 273 | f->credit = q->initial_quantum; |
| 274 | |
| 275 | rb_link_node(&f->fq_node, parent, p); |
| 276 | rb_insert_color(&f->fq_node, root); |
| 277 | |
| 278 | q->flows++; |
| 279 | q->inactive_flows++; |
| 280 | return f; |
| 281 | } |
| 282 | |
| 283 | |
| 284 | /* remove one skb from head of flow queue */ |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 285 | static struct sk_buff *fq_dequeue_head(struct Qdisc *sch, struct fq_flow *flow) |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 286 | { |
| 287 | struct sk_buff *skb = flow->head; |
| 288 | |
| 289 | if (skb) { |
| 290 | flow->head = skb->next; |
| 291 | skb->next = NULL; |
| 292 | flow->qlen--; |
John Fastabend | 25331d6 | 2014-09-28 11:53:29 -0700 | [diff] [blame] | 293 | qdisc_qstats_backlog_dec(sch, skb); |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 294 | sch->q.qlen--; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 295 | } |
| 296 | return skb; |
| 297 | } |
| 298 | |
| 299 | /* We might add in the future detection of retransmits |
| 300 | * For the time being, just return false |
| 301 | */ |
| 302 | static bool skb_is_retransmit(struct sk_buff *skb) |
| 303 | { |
| 304 | return false; |
| 305 | } |
| 306 | |
| 307 | /* add skb to flow queue |
| 308 | * flow queue is a linked list, kind of FIFO, except for TCP retransmits |
| 309 | * We special case tcp retransmits to be transmitted before other packets. |
| 310 | * We rely on fact that TCP retransmits are unlikely, so we do not waste |
| 311 | * a separate queue or a pointer. |
| 312 | * head-> [retrans pkt 1] |
| 313 | * [retrans pkt 2] |
| 314 | * [ normal pkt 1] |
| 315 | * [ normal pkt 2] |
| 316 | * [ normal pkt 3] |
| 317 | * tail-> [ normal pkt 4] |
| 318 | */ |
| 319 | static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb) |
| 320 | { |
| 321 | struct sk_buff *prev, *head = flow->head; |
| 322 | |
| 323 | skb->next = NULL; |
| 324 | if (!head) { |
| 325 | flow->head = skb; |
| 326 | flow->tail = skb; |
| 327 | return; |
| 328 | } |
| 329 | if (likely(!skb_is_retransmit(skb))) { |
| 330 | flow->tail->next = skb; |
| 331 | flow->tail = skb; |
| 332 | return; |
| 333 | } |
| 334 | |
| 335 | /* This skb is a tcp retransmit, |
| 336 | * find the last retrans packet in the queue |
| 337 | */ |
| 338 | prev = NULL; |
| 339 | while (skb_is_retransmit(head)) { |
| 340 | prev = head; |
| 341 | head = head->next; |
| 342 | if (!head) |
| 343 | break; |
| 344 | } |
| 345 | if (!prev) { /* no rtx packet in queue, become the new head */ |
| 346 | skb->next = flow->head; |
| 347 | flow->head = skb; |
| 348 | } else { |
| 349 | if (prev == flow->tail) |
| 350 | flow->tail = skb; |
| 351 | else |
| 352 | skb->next = prev->next; |
| 353 | prev->next = skb; |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch) |
| 358 | { |
| 359 | struct fq_sched_data *q = qdisc_priv(sch); |
| 360 | struct fq_flow *f; |
| 361 | |
| 362 | if (unlikely(sch->q.qlen >= sch->limit)) |
| 363 | return qdisc_drop(skb, sch); |
| 364 | |
| 365 | f = fq_classify(skb, q); |
| 366 | if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) { |
| 367 | q->stat_flows_plimit++; |
| 368 | return qdisc_drop(skb, sch); |
| 369 | } |
| 370 | |
| 371 | f->qlen++; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 372 | if (skb_is_retransmit(skb)) |
| 373 | q->stat_tcp_retrans++; |
John Fastabend | 25331d6 | 2014-09-28 11:53:29 -0700 | [diff] [blame] | 374 | qdisc_qstats_backlog_inc(sch, skb); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 375 | if (fq_flow_is_detached(f)) { |
| 376 | fq_flow_add_tail(&q->new_flows, f); |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 377 | if (time_after(jiffies, f->age + q->flow_refill_delay)) |
| 378 | f->credit = max_t(u32, f->credit, q->quantum); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 379 | q->inactive_flows--; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 380 | } |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 381 | |
| 382 | /* Note: this overwrites f->age */ |
| 383 | flow_queue_add(f, skb); |
| 384 | |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 385 | if (unlikely(f == &q->internal)) { |
| 386 | q->stat_internal_packets++; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 387 | } |
| 388 | sch->q.qlen++; |
| 389 | |
| 390 | return NET_XMIT_SUCCESS; |
| 391 | } |
| 392 | |
| 393 | static void fq_check_throttled(struct fq_sched_data *q, u64 now) |
| 394 | { |
| 395 | struct rb_node *p; |
| 396 | |
| 397 | if (q->time_next_delayed_flow > now) |
| 398 | return; |
| 399 | |
| 400 | q->time_next_delayed_flow = ~0ULL; |
| 401 | while ((p = rb_first(&q->delayed)) != NULL) { |
| 402 | struct fq_flow *f = container_of(p, struct fq_flow, rate_node); |
| 403 | |
| 404 | if (f->time_next_packet > now) { |
| 405 | q->time_next_delayed_flow = f->time_next_packet; |
| 406 | break; |
| 407 | } |
| 408 | rb_erase(p, &q->delayed); |
| 409 | q->throttled_flows--; |
| 410 | fq_flow_add_tail(&q->old_flows, f); |
| 411 | } |
| 412 | } |
| 413 | |
| 414 | static struct sk_buff *fq_dequeue(struct Qdisc *sch) |
| 415 | { |
| 416 | struct fq_sched_data *q = qdisc_priv(sch); |
Eric Dumazet | d2de875 | 2014-08-22 18:32:09 -0700 | [diff] [blame] | 417 | u64 now = ktime_get_ns(); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 418 | struct fq_flow_head *head; |
| 419 | struct sk_buff *skb; |
| 420 | struct fq_flow *f; |
Eric Dumazet | 0eab5eb | 2013-10-01 09:10:16 -0700 | [diff] [blame] | 421 | u32 rate; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 422 | |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 423 | skb = fq_dequeue_head(sch, &q->internal); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 424 | if (skb) |
| 425 | goto out; |
| 426 | fq_check_throttled(q, now); |
| 427 | begin: |
| 428 | head = &q->new_flows; |
| 429 | if (!head->first) { |
| 430 | head = &q->old_flows; |
| 431 | if (!head->first) { |
| 432 | if (q->time_next_delayed_flow != ~0ULL) |
| 433 | qdisc_watchdog_schedule_ns(&q->watchdog, |
Eric Dumazet | f2600cf | 2014-10-04 10:11:31 -0700 | [diff] [blame] | 434 | q->time_next_delayed_flow, |
| 435 | false); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 436 | return NULL; |
| 437 | } |
| 438 | } |
| 439 | f = head->first; |
| 440 | |
| 441 | if (f->credit <= 0) { |
| 442 | f->credit += q->quantum; |
| 443 | head->first = f->next; |
| 444 | fq_flow_add_tail(&q->old_flows, f); |
| 445 | goto begin; |
| 446 | } |
| 447 | |
| 448 | if (unlikely(f->head && now < f->time_next_packet)) { |
| 449 | head->first = f->next; |
| 450 | fq_flow_set_throttled(q, f); |
| 451 | goto begin; |
| 452 | } |
| 453 | |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 454 | skb = fq_dequeue_head(sch, f); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 455 | if (!skb) { |
| 456 | head->first = f->next; |
| 457 | /* force a pass through old_flows to prevent starvation */ |
| 458 | if ((head == &q->new_flows) && q->old_flows.first) { |
| 459 | fq_flow_add_tail(&q->old_flows, f); |
| 460 | } else { |
| 461 | fq_flow_set_detached(f); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 462 | q->inactive_flows++; |
| 463 | } |
| 464 | goto begin; |
| 465 | } |
Eric Dumazet | 08f89b9 | 2013-08-30 09:46:43 -0700 | [diff] [blame] | 466 | prefetch(&skb->end); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 467 | f->time_next_packet = now; |
| 468 | f->credit -= qdisc_pkt_len(skb); |
| 469 | |
Eric Dumazet | 0eab5eb | 2013-10-01 09:10:16 -0700 | [diff] [blame] | 470 | if (f->credit > 0 || !q->rate_enable) |
| 471 | goto out; |
| 472 | |
Eric Dumazet | 7eec417 | 2013-10-08 15:16:00 -0700 | [diff] [blame] | 473 | rate = q->flow_max_rate; |
| 474 | if (skb->sk && skb->sk->sk_state != TCP_TIME_WAIT) |
| 475 | rate = min(skb->sk->sk_pacing_rate, rate); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 476 | |
Eric Dumazet | 7eec417 | 2013-10-08 15:16:00 -0700 | [diff] [blame] | 477 | if (rate != ~0U) { |
Eric Dumazet | 0eab5eb | 2013-10-01 09:10:16 -0700 | [diff] [blame] | 478 | u32 plen = max(qdisc_pkt_len(skb), q->quantum); |
| 479 | u64 len = (u64)plen * NSEC_PER_SEC; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 480 | |
Eric Dumazet | 7eec417 | 2013-10-08 15:16:00 -0700 | [diff] [blame] | 481 | if (likely(rate)) |
| 482 | do_div(len, rate); |
Eric Dumazet | 0eab5eb | 2013-10-01 09:10:16 -0700 | [diff] [blame] | 483 | /* Since socket rate can change later, |
| 484 | * clamp the delay to 125 ms. |
| 485 | * TODO: maybe segment the too big skb, as in commit |
| 486 | * e43ac79a4bc ("sch_tbf: segment too big GSO packets") |
| 487 | */ |
| 488 | if (unlikely(len > 125 * NSEC_PER_MSEC)) { |
| 489 | len = 125 * NSEC_PER_MSEC; |
| 490 | q->stat_pkts_too_long++; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 491 | } |
Eric Dumazet | 0eab5eb | 2013-10-01 09:10:16 -0700 | [diff] [blame] | 492 | |
| 493 | f->time_next_packet = now + len; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 494 | } |
| 495 | out: |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 496 | qdisc_bstats_update(sch, skb); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 497 | return skb; |
| 498 | } |
| 499 | |
| 500 | static void fq_reset(struct Qdisc *sch) |
| 501 | { |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 502 | struct fq_sched_data *q = qdisc_priv(sch); |
| 503 | struct rb_root *root; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 504 | struct sk_buff *skb; |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 505 | struct rb_node *p; |
| 506 | struct fq_flow *f; |
| 507 | unsigned int idx; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 508 | |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 509 | while ((skb = fq_dequeue_head(sch, &q->internal)) != NULL) |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 510 | kfree_skb(skb); |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 511 | |
| 512 | if (!q->fq_root) |
| 513 | return; |
| 514 | |
| 515 | for (idx = 0; idx < (1U << q->fq_trees_log); idx++) { |
| 516 | root = &q->fq_root[idx]; |
| 517 | while ((p = rb_first(root)) != NULL) { |
| 518 | f = container_of(p, struct fq_flow, fq_node); |
| 519 | rb_erase(p, root); |
| 520 | |
| 521 | while ((skb = fq_dequeue_head(sch, f)) != NULL) |
| 522 | kfree_skb(skb); |
| 523 | |
| 524 | kmem_cache_free(fq_flow_cachep, f); |
| 525 | } |
| 526 | } |
| 527 | q->new_flows.first = NULL; |
| 528 | q->old_flows.first = NULL; |
| 529 | q->delayed = RB_ROOT; |
| 530 | q->flows = 0; |
| 531 | q->inactive_flows = 0; |
| 532 | q->throttled_flows = 0; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 533 | } |
| 534 | |
| 535 | static void fq_rehash(struct fq_sched_data *q, |
| 536 | struct rb_root *old_array, u32 old_log, |
| 537 | struct rb_root *new_array, u32 new_log) |
| 538 | { |
| 539 | struct rb_node *op, **np, *parent; |
| 540 | struct rb_root *oroot, *nroot; |
| 541 | struct fq_flow *of, *nf; |
| 542 | int fcnt = 0; |
| 543 | u32 idx; |
| 544 | |
| 545 | for (idx = 0; idx < (1U << old_log); idx++) { |
| 546 | oroot = &old_array[idx]; |
| 547 | while ((op = rb_first(oroot)) != NULL) { |
| 548 | rb_erase(op, oroot); |
| 549 | of = container_of(op, struct fq_flow, fq_node); |
| 550 | if (fq_gc_candidate(of)) { |
| 551 | fcnt++; |
| 552 | kmem_cache_free(fq_flow_cachep, of); |
| 553 | continue; |
| 554 | } |
| 555 | nroot = &new_array[hash_32((u32)(long)of->sk, new_log)]; |
| 556 | |
| 557 | np = &nroot->rb_node; |
| 558 | parent = NULL; |
| 559 | while (*np) { |
| 560 | parent = *np; |
| 561 | |
| 562 | nf = container_of(parent, struct fq_flow, fq_node); |
| 563 | BUG_ON(nf->sk == of->sk); |
| 564 | |
| 565 | if (nf->sk > of->sk) |
| 566 | np = &parent->rb_right; |
| 567 | else |
| 568 | np = &parent->rb_left; |
| 569 | } |
| 570 | |
| 571 | rb_link_node(&of->fq_node, parent, np); |
| 572 | rb_insert_color(&of->fq_node, nroot); |
| 573 | } |
| 574 | } |
| 575 | q->flows -= fcnt; |
| 576 | q->inactive_flows -= fcnt; |
| 577 | q->stat_gc_flows += fcnt; |
| 578 | } |
| 579 | |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 580 | static void *fq_alloc_node(size_t sz, int node) |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 581 | { |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 582 | void *ptr; |
| 583 | |
| 584 | ptr = kmalloc_node(sz, GFP_KERNEL | __GFP_REPEAT | __GFP_NOWARN, node); |
| 585 | if (!ptr) |
| 586 | ptr = vmalloc_node(sz, node); |
| 587 | return ptr; |
| 588 | } |
| 589 | |
| 590 | static void fq_free(void *addr) |
| 591 | { |
WANG Cong | 4cb2897 | 2014-06-02 15:55:22 -0700 | [diff] [blame] | 592 | kvfree(addr); |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 593 | } |
| 594 | |
| 595 | static int fq_resize(struct Qdisc *sch, u32 log) |
| 596 | { |
| 597 | struct fq_sched_data *q = qdisc_priv(sch); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 598 | struct rb_root *array; |
Eric Dumazet | 2d8d40a | 2014-03-06 22:57:52 -0800 | [diff] [blame] | 599 | void *old_fq_root; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 600 | u32 idx; |
| 601 | |
| 602 | if (q->fq_root && log == q->fq_trees_log) |
| 603 | return 0; |
| 604 | |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 605 | /* If XPS was setup, we can allocate memory on right NUMA node */ |
| 606 | array = fq_alloc_node(sizeof(struct rb_root) << log, |
| 607 | netdev_queue_numa_node_read(sch->dev_queue)); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 608 | if (!array) |
| 609 | return -ENOMEM; |
| 610 | |
| 611 | for (idx = 0; idx < (1U << log); idx++) |
| 612 | array[idx] = RB_ROOT; |
| 613 | |
Eric Dumazet | 2d8d40a | 2014-03-06 22:57:52 -0800 | [diff] [blame] | 614 | sch_tree_lock(sch); |
| 615 | |
| 616 | old_fq_root = q->fq_root; |
| 617 | if (old_fq_root) |
| 618 | fq_rehash(q, old_fq_root, q->fq_trees_log, array, log); |
| 619 | |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 620 | q->fq_root = array; |
| 621 | q->fq_trees_log = log; |
| 622 | |
Eric Dumazet | 2d8d40a | 2014-03-06 22:57:52 -0800 | [diff] [blame] | 623 | sch_tree_unlock(sch); |
| 624 | |
| 625 | fq_free(old_fq_root); |
| 626 | |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 627 | return 0; |
| 628 | } |
| 629 | |
| 630 | static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = { |
| 631 | [TCA_FQ_PLIMIT] = { .type = NLA_U32 }, |
| 632 | [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 }, |
| 633 | [TCA_FQ_QUANTUM] = { .type = NLA_U32 }, |
| 634 | [TCA_FQ_INITIAL_QUANTUM] = { .type = NLA_U32 }, |
| 635 | [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 }, |
| 636 | [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 }, |
| 637 | [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 }, |
| 638 | [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 }, |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 639 | [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 }, |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 640 | }; |
| 641 | |
| 642 | static int fq_change(struct Qdisc *sch, struct nlattr *opt) |
| 643 | { |
| 644 | struct fq_sched_data *q = qdisc_priv(sch); |
| 645 | struct nlattr *tb[TCA_FQ_MAX + 1]; |
| 646 | int err, drop_count = 0; |
| 647 | u32 fq_log; |
| 648 | |
| 649 | if (!opt) |
| 650 | return -EINVAL; |
| 651 | |
| 652 | err = nla_parse_nested(tb, TCA_FQ_MAX, opt, fq_policy); |
| 653 | if (err < 0) |
| 654 | return err; |
| 655 | |
| 656 | sch_tree_lock(sch); |
| 657 | |
| 658 | fq_log = q->fq_trees_log; |
| 659 | |
| 660 | if (tb[TCA_FQ_BUCKETS_LOG]) { |
| 661 | u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]); |
| 662 | |
| 663 | if (nval >= 1 && nval <= ilog2(256*1024)) |
| 664 | fq_log = nval; |
| 665 | else |
| 666 | err = -EINVAL; |
| 667 | } |
| 668 | if (tb[TCA_FQ_PLIMIT]) |
| 669 | sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]); |
| 670 | |
| 671 | if (tb[TCA_FQ_FLOW_PLIMIT]) |
| 672 | q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]); |
| 673 | |
| 674 | if (tb[TCA_FQ_QUANTUM]) |
| 675 | q->quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]); |
| 676 | |
| 677 | if (tb[TCA_FQ_INITIAL_QUANTUM]) |
Eric Dumazet | ede869c | 2013-10-07 12:50:18 -0700 | [diff] [blame] | 678 | q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 679 | |
| 680 | if (tb[TCA_FQ_FLOW_DEFAULT_RATE]) |
Eric Dumazet | 65c5189a | 2013-11-15 08:57:26 -0800 | [diff] [blame] | 681 | pr_warn_ratelimited("sch_fq: defrate %u ignored.\n", |
| 682 | nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE])); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 683 | |
| 684 | if (tb[TCA_FQ_FLOW_MAX_RATE]) |
| 685 | q->flow_max_rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]); |
| 686 | |
| 687 | if (tb[TCA_FQ_RATE_ENABLE]) { |
| 688 | u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]); |
| 689 | |
| 690 | if (enable <= 1) |
| 691 | q->rate_enable = enable; |
| 692 | else |
| 693 | err = -EINVAL; |
| 694 | } |
| 695 | |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 696 | if (tb[TCA_FQ_FLOW_REFILL_DELAY]) { |
| 697 | u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ; |
| 698 | |
| 699 | q->flow_refill_delay = usecs_to_jiffies(usecs_delay); |
| 700 | } |
| 701 | |
Eric Dumazet | 2d8d40a | 2014-03-06 22:57:52 -0800 | [diff] [blame] | 702 | if (!err) { |
| 703 | sch_tree_unlock(sch); |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 704 | err = fq_resize(sch, fq_log); |
Eric Dumazet | 2d8d40a | 2014-03-06 22:57:52 -0800 | [diff] [blame] | 705 | sch_tree_lock(sch); |
| 706 | } |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 707 | while (sch->q.qlen > sch->limit) { |
| 708 | struct sk_buff *skb = fq_dequeue(sch); |
| 709 | |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 710 | if (!skb) |
| 711 | break; |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 712 | kfree_skb(skb); |
| 713 | drop_count++; |
| 714 | } |
| 715 | qdisc_tree_decrease_qlen(sch, drop_count); |
| 716 | |
| 717 | sch_tree_unlock(sch); |
| 718 | return err; |
| 719 | } |
| 720 | |
| 721 | static void fq_destroy(struct Qdisc *sch) |
| 722 | { |
| 723 | struct fq_sched_data *q = qdisc_priv(sch); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 724 | |
Eric Dumazet | 8d34ce1 | 2013-09-27 14:20:01 -0700 | [diff] [blame] | 725 | fq_reset(sch); |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 726 | fq_free(q->fq_root); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 727 | qdisc_watchdog_cancel(&q->watchdog); |
| 728 | } |
| 729 | |
| 730 | static int fq_init(struct Qdisc *sch, struct nlattr *opt) |
| 731 | { |
| 732 | struct fq_sched_data *q = qdisc_priv(sch); |
| 733 | int err; |
| 734 | |
| 735 | sch->limit = 10000; |
| 736 | q->flow_plimit = 100; |
| 737 | q->quantum = 2 * psched_mtu(qdisc_dev(sch)); |
| 738 | q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch)); |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 739 | q->flow_refill_delay = msecs_to_jiffies(40); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 740 | q->flow_max_rate = ~0U; |
| 741 | q->rate_enable = 1; |
| 742 | q->new_flows.first = NULL; |
| 743 | q->old_flows.first = NULL; |
| 744 | q->delayed = RB_ROOT; |
| 745 | q->fq_root = NULL; |
| 746 | q->fq_trees_log = ilog2(1024); |
| 747 | qdisc_watchdog_init(&q->watchdog, sch); |
| 748 | |
| 749 | if (opt) |
| 750 | err = fq_change(sch, opt); |
| 751 | else |
Eric Dumazet | c3bd854 | 2013-12-15 13:15:25 -0800 | [diff] [blame] | 752 | err = fq_resize(sch, q->fq_trees_log); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 753 | |
| 754 | return err; |
| 755 | } |
| 756 | |
| 757 | static int fq_dump(struct Qdisc *sch, struct sk_buff *skb) |
| 758 | { |
| 759 | struct fq_sched_data *q = qdisc_priv(sch); |
| 760 | struct nlattr *opts; |
| 761 | |
| 762 | opts = nla_nest_start(skb, TCA_OPTIONS); |
| 763 | if (opts == NULL) |
| 764 | goto nla_put_failure; |
| 765 | |
Eric Dumazet | 65c5189a | 2013-11-15 08:57:26 -0800 | [diff] [blame] | 766 | /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */ |
| 767 | |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 768 | if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) || |
| 769 | nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) || |
| 770 | nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) || |
| 771 | nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) || |
| 772 | nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) || |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 773 | nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE, q->flow_max_rate) || |
Eric Dumazet | f52ed89 | 2013-11-15 08:58:14 -0800 | [diff] [blame] | 774 | nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY, |
| 775 | jiffies_to_usecs(q->flow_refill_delay)) || |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 776 | nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log)) |
| 777 | goto nla_put_failure; |
| 778 | |
Yang Yingliang | d59b7d8 | 2014-03-12 10:20:32 +0800 | [diff] [blame] | 779 | return nla_nest_end(skb, opts); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 780 | |
| 781 | nla_put_failure: |
| 782 | return -1; |
| 783 | } |
| 784 | |
| 785 | static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d) |
| 786 | { |
| 787 | struct fq_sched_data *q = qdisc_priv(sch); |
Eric Dumazet | d2de875 | 2014-08-22 18:32:09 -0700 | [diff] [blame] | 788 | u64 now = ktime_get_ns(); |
Eric Dumazet | afe4fd0 | 2013-08-29 15:49:55 -0700 | [diff] [blame] | 789 | struct tc_fq_qd_stats st = { |
| 790 | .gc_flows = q->stat_gc_flows, |
| 791 | .highprio_packets = q->stat_internal_packets, |
| 792 | .tcp_retrans = q->stat_tcp_retrans, |
| 793 | .throttled = q->stat_throttled, |
| 794 | .flows_plimit = q->stat_flows_plimit, |
| 795 | .pkts_too_long = q->stat_pkts_too_long, |
| 796 | .allocation_errors = q->stat_allocation_errors, |
| 797 | .flows = q->flows, |
| 798 | .inactive_flows = q->inactive_flows, |
| 799 | .throttled_flows = q->throttled_flows, |
| 800 | .time_next_delayed_flow = q->time_next_delayed_flow - now, |
| 801 | }; |
| 802 | |
| 803 | return gnet_stats_copy_app(d, &st, sizeof(st)); |
| 804 | } |
| 805 | |
| 806 | static struct Qdisc_ops fq_qdisc_ops __read_mostly = { |
| 807 | .id = "fq", |
| 808 | .priv_size = sizeof(struct fq_sched_data), |
| 809 | |
| 810 | .enqueue = fq_enqueue, |
| 811 | .dequeue = fq_dequeue, |
| 812 | .peek = qdisc_peek_dequeued, |
| 813 | .init = fq_init, |
| 814 | .reset = fq_reset, |
| 815 | .destroy = fq_destroy, |
| 816 | .change = fq_change, |
| 817 | .dump = fq_dump, |
| 818 | .dump_stats = fq_dump_stats, |
| 819 | .owner = THIS_MODULE, |
| 820 | }; |
| 821 | |
| 822 | static int __init fq_module_init(void) |
| 823 | { |
| 824 | int ret; |
| 825 | |
| 826 | fq_flow_cachep = kmem_cache_create("fq_flow_cache", |
| 827 | sizeof(struct fq_flow), |
| 828 | 0, 0, NULL); |
| 829 | if (!fq_flow_cachep) |
| 830 | return -ENOMEM; |
| 831 | |
| 832 | ret = register_qdisc(&fq_qdisc_ops); |
| 833 | if (ret) |
| 834 | kmem_cache_destroy(fq_flow_cachep); |
| 835 | return ret; |
| 836 | } |
| 837 | |
| 838 | static void __exit fq_module_exit(void) |
| 839 | { |
| 840 | unregister_qdisc(&fq_qdisc_ops); |
| 841 | kmem_cache_destroy(fq_flow_cachep); |
| 842 | } |
| 843 | |
| 844 | module_init(fq_module_init) |
| 845 | module_exit(fq_module_exit) |
| 846 | MODULE_AUTHOR("Eric Dumazet"); |
| 847 | MODULE_LICENSE("GPL"); |