blob: f59a2cb2c803d79bd1f0eb1806464a0220824f9e [file] [log] [blame]
Thomas Gleixner2874c5f2019-05-27 08:55:01 +02001// SPDX-License-Identifier: GPL-2.0-or-later
Eric Dumazetafe4fd02013-08-29 15:49:55 -07002/*
3 * net/sched/sch_fq.c Fair Queue Packet Scheduler (per flow pacing)
4 *
Eric Dumazet86b3bfe2015-01-28 06:06:36 -08005 * Copyright (C) 2013-2015 Eric Dumazet <edumazet@google.com>
Eric Dumazetafe4fd02013-08-29 15:49:55 -07006 *
Simon Horman05e8bb82015-04-02 11:20:23 +09007 * Meant to be mostly used for locally generated traffic :
Eric Dumazetafe4fd02013-08-29 15:49:55 -07008 * 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>
Eric Dumazet08f89b92013-08-30 09:46:43 -070045#include <linux/prefetch.h>
Eric Dumazetc3bd8542013-12-15 13:15:25 -080046#include <linux/vmalloc.h>
Eric Dumazetafe4fd02013-08-29 15:49:55 -070047#include <net/netlink.h>
48#include <net/pkt_sched.h>
49#include <net/sock.h>
50#include <net/tcp_states.h>
Eric Dumazet98781962015-02-03 18:31:53 -080051#include <net/tcp.h>
Eric Dumazetafe4fd02013-08-29 15:49:55 -070052
Eric Dumazeteeb84aa2019-05-04 16:48:53 -070053struct fq_skb_cb {
54 u64 time_to_send;
55};
56
57static 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
Eric Dumazetafe4fd02013-08-29 15:49:55 -070063/*
Eric Dumazeteeb84aa2019-05-04 16:48:53 -070064 * 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).
Eric Dumazetafe4fd02013-08-29 15:49:55 -070067 */
68struct fq_flow {
Eric Dumazet7ba05372020-05-02 19:54:19 -070069/* First cache line : used in fq_gc(), fq_enqueue(), fq_dequeue() */
Eric Dumazeteeb84aa2019-05-04 16:48:53 -070070 struct rb_root t_root;
Eric Dumazetafe4fd02013-08-29 15:49:55 -070071 struct sk_buff *head; /* list of skbs for this flow : first skb */
72 union {
73 struct sk_buff *tail; /* last skb in the list */
Eric Dumazetdde0a642020-05-02 19:54:18 -070074 unsigned long age; /* (jiffies | 1UL) when flow was emptied, for gc */
Eric Dumazetafe4fd02013-08-29 15:49:55 -070075 };
Simon Horman05e8bb82015-04-02 11:20:23 +090076 struct rb_node fq_node; /* anchor in fq_root[] trees */
Eric Dumazetafe4fd02013-08-29 15:49:55 -070077 struct sock *sk;
Eric Dumazetafe4fd02013-08-29 15:49:55 -070078 u32 socket_hash; /* sk_hash */
Eric Dumazet7ba05372020-05-02 19:54:19 -070079 int qlen; /* number of packets in flow queue */
80
81/* Second cache line, used in fq_dequeue() */
82 int credit;
83 /* 32bit hole on 64bit arches */
84
Eric Dumazetdde0a642020-05-02 19:54:18 -070085 struct fq_flow *next; /* next pointer in RR lists */
Eric Dumazetafe4fd02013-08-29 15:49:55 -070086
87 struct rb_node rate_node; /* anchor in q->delayed tree */
88 u64 time_next_packet;
Eric Dumazet7ba05372020-05-02 19:54:19 -070089} ____cacheline_aligned_in_smp;
Eric Dumazetafe4fd02013-08-29 15:49:55 -070090
91struct fq_flow_head {
92 struct fq_flow *first;
93 struct fq_flow *last;
94};
95
96struct 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;
Eric Dumazet39d01052020-05-01 07:07:41 -0700103 u64 ktime_cache; /* copy of last ktime_get_ns() */
Eric Dumazetfefa5692016-09-22 08:58:55 -0700104 unsigned long unthrottle_latency_ns;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700105
106 struct fq_flow internal; /* for non classified or high prio packets */
107 u32 quantum;
108 u32 initial_quantum;
Eric Dumazetf52ed892013-11-15 08:58:14 -0800109 u32 flow_refill_delay;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700110 u32 flow_plimit; /* max packets per flow */
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700111 unsigned long flow_max_rate; /* optional max rate per flow */
Eric Dumazet48872c12018-11-11 09:11:31 -0800112 u64 ce_threshold;
Eric Dumazet39d01052020-05-01 07:07:41 -0700113 u64 horizon; /* horizon in ns */
Eric Dumazet06eb3952015-02-04 21:30:40 -0800114 u32 orphan_mask; /* mask for orphaned skb */
Eric Dumazet77879142016-09-19 23:39:11 -0400115 u32 low_rate_threshold;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700116 struct rb_root *fq_root;
117 u8 rate_enable;
118 u8 fq_trees_log;
Eric Dumazet39d01052020-05-01 07:07:41 -0700119 u8 horizon_drop;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700120 u32 flows;
121 u32 inactive_flows;
122 u32 throttled_flows;
123
124 u64 stat_gc_flows;
125 u64 stat_internal_packets;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700126 u64 stat_throttled;
Eric Dumazet48872c12018-11-11 09:11:31 -0800127 u64 stat_ce_mark;
Eric Dumazet39d01052020-05-01 07:07:41 -0700128 u64 stat_horizon_drops;
129 u64 stat_horizon_caps;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700130 u64 stat_flows_plimit;
131 u64 stat_pkts_too_long;
132 u64 stat_allocation_errors;
Eric Dumazet583396f2020-03-16 19:12:51 -0700133
134 u32 timer_slack; /* hrtimer slack in ns */
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700135 struct qdisc_watchdog watchdog;
136};
137
Eric Dumazetdde0a642020-05-02 19:54:18 -0700138/*
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 */
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700144static void fq_flow_set_detached(struct fq_flow *f)
145{
Eric Dumazetdde0a642020-05-02 19:54:18 -0700146 f->age = jiffies | 1UL;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700147}
148
149static bool fq_flow_is_detached(const struct fq_flow *f)
150{
Eric Dumazetdde0a642020-05-02 19:54:18 -0700151 return !!(f->age & 1UL);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700152}
153
Eric Dumazetdde0a642020-05-02 19:54:18 -0700154/* special value to mark a throttled flow (not on old/new list) */
155static struct fq_flow throttled;
156
Eric Dumazet7df40c22018-05-02 10:03:30 -0700157static bool fq_flow_is_throttled(const struct fq_flow *f)
158{
159 return f->next == &throttled;
160}
161
162static 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
172static 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
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700179static 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;
Geliang Tange1245572016-12-20 22:02:15 +0800187 aux = rb_entry(parent, struct fq_flow, rate_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700188 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
204static struct kmem_cache *fq_flow_cachep __read_mostly;
205
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700206
207/* limit number of collected flows per round */
208#define FQ_GC_MAX 8
209#define FQ_GC_AGE (3*HZ)
210
211static 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
217static void fq_gc(struct fq_sched_data *q,
218 struct rb_root *root,
219 struct sock *sk)
220{
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700221 struct rb_node **p, *parent;
Eric Dumazet82a0aa52020-05-02 19:54:20 -0700222 void *tofree[FQ_GC_MAX];
223 struct fq_flow *f;
224 int i, fcnt = 0;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700225
226 p = &root->rb_node;
227 parent = NULL;
228 while (*p) {
229 parent = *p;
230
Geliang Tange1245572016-12-20 22:02:15 +0800231 f = rb_entry(parent, struct fq_flow, fq_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700232 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
Eric Dumazet82a0aa52020-05-02 19:54:20 -0700247 if (!fcnt)
248 return;
249
250 for (i = fcnt; i > 0; ) {
251 f = tofree[--i];
252 rb_erase(&f->fq_node, root);
253 }
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700254 q->flows -= fcnt;
255 q->inactive_flows -= fcnt;
256 q->stat_gc_flows += fcnt;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700257
Eric Dumazet82a0aa52020-05-02 19:54:20 -0700258 kmem_cache_free_bulk(fq_flow_cachep, fcnt, tofree);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700259}
260
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700261static 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;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700267
268 /* warning: no starvation prevention... */
Maciej Żenczykowski2abc2f02013-11-14 08:50:43 -0800269 if (unlikely((skb->priority & TC_PRIO_MAX) == TC_PRIO_CONTROL))
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700270 return &q->internal;
271
Eric Dumazetca6fb062015-10-02 11:43:35 -0700272 /* SYNACK messages are attached to a TCP_NEW_SYN_RECV request socket
Eric Dumazete446f9d2015-10-08 05:01:55 -0700273 * or a listener (SYNCOOKIE mode)
Eric Dumazetca6fb062015-10-02 11:43:35 -0700274 * 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),
Eric Dumazet06eb3952015-02-04 21:30:40 -0800278 * especially if the listener set SO_MAX_PACING_RATE
Eric Dumazetca6fb062015-10-02 11:43:35 -0700279 * 4) We pretend they are orphaned
Eric Dumazet06eb3952015-02-04 21:30:40 -0800280 */
Eric Dumazete446f9d2015-10-08 05:01:55 -0700281 if (!sk || sk_listener(sk)) {
Eric Dumazet06eb3952015-02-04 21:30:40 -0800282 unsigned long hash = skb_get_hash(skb) & q->orphan_mask;
283
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700284 /* By forcing low order bit to 1, we make sure to not
285 * collide with a local flow (socket pointers are word aligned)
286 */
Eric Dumazet06eb3952015-02-04 21:30:40 -0800287 sk = (struct sock *)((hash << 1) | 1UL);
288 skb_orphan(skb);
Eric Dumazet37c0aea2019-05-04 16:48:54 -0700289 } 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);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700300 }
301
Eric Dumazet29c58472016-11-17 09:48:30 -0800302 root = &q->fq_root[hash_ptr(sk, q->fq_trees_log)];
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700303
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
Geliang Tange1245572016-12-20 22:02:15 +0800313 f = rb_entry(parent, struct fq_flow, fq_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700314 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 */
Eric Dumazet37c0aea2019-05-04 16:48:54 -0700320 if (unlikely(skb->sk == sk &&
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700321 f->socket_hash != sk->sk_hash)) {
322 f->credit = q->initial_quantum;
323 f->socket_hash = sk->sk_hash;
Eric Dumazetbb3d0b82019-12-23 11:13:24 -0800324 if (q->rate_enable)
325 smp_store_release(&sk->sk_pacing_status,
326 SK_PACING_FQ);
Eric Dumazet7df40c22018-05-02 10:03:30 -0700327 if (fq_flow_is_throttled(f))
328 fq_flow_unset_throttled(q, f);
Eric Dumazetfc59d5b2013-10-27 16:26:43 -0700329 f->time_next_packet = 0ULL;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700330 }
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 }
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700344 /* f->t_root is already zeroed after kmem_cache_zalloc() */
345
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700346 fq_flow_set_detached(f);
347 f->sk = sk;
Eric Dumazetbb3d0b82019-12-23 11:13:24 -0800348 if (skb->sk == sk) {
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700349 f->socket_hash = sk->sk_hash;
Eric Dumazetbb3d0b82019-12-23 11:13:24 -0800350 if (q->rate_enable)
351 smp_store_release(&sk->sk_pacing_status,
352 SK_PACING_FQ);
353 }
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700354 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
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700364static 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
380static 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}
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700390
Eric Dumazetc288b0c2020-05-02 19:54:21 -0700391/* Remove one skb from flow queue.
392 * This skb must be the return value of prior fq_peek().
393 */
394static void fq_dequeue_skb(struct Qdisc *sch, struct fq_flow *flow,
395 struct sk_buff *skb)
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700396{
Eric Dumazetc288b0c2020-05-02 19:54:21 -0700397 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--;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700402}
403
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700404static void flow_queue_add(struct fq_flow *flow, struct sk_buff *skb)
405{
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700406 struct rb_node **p, *parent;
407 struct sk_buff *head, *aux;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700408
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700409 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;
423
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);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700434}
435
Eric Dumazet39d01052020-05-01 07:07:41 -0700436static 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
Eric Dumazet520ac302016-06-21 23:16:49 -0700442static int fq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
443 struct sk_buff **to_free)
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700444{
445 struct fq_sched_data *q = qdisc_priv(sch);
446 struct fq_flow *f;
447
448 if (unlikely(sch->q.qlen >= sch->limit))
Eric Dumazet520ac302016-06-21 23:16:49 -0700449 return qdisc_drop(skb, sch, to_free);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700450
Eric Dumazet39d01052020-05-01 07:07:41 -0700451 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
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700473 f = fq_classify(skb, q);
474 if (unlikely(f->qlen >= q->flow_plimit && f != &q->internal)) {
475 q->stat_flows_plimit++;
Eric Dumazet520ac302016-06-21 23:16:49 -0700476 return qdisc_drop(skb, sch, to_free);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700477 }
478
479 f->qlen++;
John Fastabend25331d62014-09-28 11:53:29 -0700480 qdisc_qstats_backlog_inc(sch, skb);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700481 if (fq_flow_is_detached(f)) {
482 fq_flow_add_tail(&q->new_flows, f);
Eric Dumazetf52ed892013-11-15 08:58:14 -0800483 if (time_after(jiffies, f->age + q->flow_refill_delay))
484 f->credit = max_t(u32, f->credit, q->quantum);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700485 q->inactive_flows--;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700486 }
Eric Dumazetf52ed892013-11-15 08:58:14 -0800487
488 /* Note: this overwrites f->age */
489 flow_queue_add(f, skb);
490
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700491 if (unlikely(f == &q->internal)) {
492 q->stat_internal_packets++;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700493 }
494 sch->q.qlen++;
495
496 return NET_XMIT_SUCCESS;
497}
498
499static void fq_check_throttled(struct fq_sched_data *q, u64 now)
500{
Eric Dumazetfefa5692016-09-22 08:58:55 -0700501 unsigned long sample;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700502 struct rb_node *p;
503
504 if (q->time_next_delayed_flow > now)
505 return;
506
Eric Dumazetfefa5692016-09-22 08:58:55 -0700507 /* 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
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700514 q->time_next_delayed_flow = ~0ULL;
515 while ((p = rb_first(&q->delayed)) != NULL) {
Geliang Tange1245572016-12-20 22:02:15 +0800516 struct fq_flow *f = rb_entry(p, struct fq_flow, rate_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700517
518 if (f->time_next_packet > now) {
519 q->time_next_delayed_flow = f->time_next_packet;
520 break;
521 }
Eric Dumazet7df40c22018-05-02 10:03:30 -0700522 fq_flow_unset_throttled(q, f);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700523 }
524}
525
526static struct sk_buff *fq_dequeue(struct Qdisc *sch)
527{
528 struct fq_sched_data *q = qdisc_priv(sch);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700529 struct fq_flow_head *head;
530 struct sk_buff *skb;
531 struct fq_flow *f;
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700532 unsigned long rate;
533 u32 plen;
Eric Dumazet6b015a52018-11-19 17:30:19 -0800534 u64 now;
535
536 if (!sch->q.qlen)
537 return NULL;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700538
Eric Dumazetc288b0c2020-05-02 19:54:21 -0700539 skb = fq_peek(&q->internal);
540 if (unlikely(skb)) {
541 fq_dequeue_skb(sch, &q->internal, skb);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700542 goto out;
Eric Dumazetc288b0c2020-05-02 19:54:21 -0700543 }
Eric Dumazet6b015a52018-11-19 17:30:19 -0800544
Eric Dumazet39d01052020-05-01 07:07:41 -0700545 q->ktime_cache = now = ktime_get_ns();
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700546 fq_check_throttled(q, now);
547begin:
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)
Eric Dumazet583396f2020-03-16 19:12:51 -0700553 qdisc_watchdog_schedule_range_ns(&q->watchdog,
554 q->time_next_delayed_flow,
555 q->timer_slack);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700556 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
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700568 skb = fq_peek(f);
Eric Dumazet7baf33b2018-10-15 09:37:55 -0700569 if (skb) {
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700570 u64 time_next_packet = max_t(u64, fq_skb_cb(skb)->time_to_send,
Eric Dumazetab408b62018-09-21 08:51:52 -0700571 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 }
Eric Dumazet348e2892020-05-02 19:54:22 -0700579 prefetch(&skb->end);
Eric Dumazete9c43ad2019-10-14 10:40:32 -0700580 if ((s64)(now - time_next_packet - q->ce_threshold) > 0) {
Eric Dumazet48872c12018-11-11 09:11:31 -0800581 INET_ECN_set_ce(skb);
582 q->stat_ce_mark++;
583 }
Eric Dumazetc288b0c2020-05-02 19:54:21 -0700584 fq_dequeue_skb(sch, f, skb);
585 } else {
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700586 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);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700592 q->inactive_flows++;
593 }
594 goto begin;
595 }
Eric Dumazet08e14fe2018-11-12 16:17:16 -0800596 plen = qdisc_pkt_len(skb);
597 f->credit -= plen;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700598
Eric Dumazet08e14fe2018-11-12 16:17:16 -0800599 if (!q->rate_enable)
Eric Dumazet98781962015-02-03 18:31:53 -0800600 goto out;
601
Eric Dumazet7eec4172013-10-08 15:16:00 -0700602 rate = q->flow_max_rate;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700603
Eric Dumazet08e14fe2018-11-12 16:17:16 -0800604 /* 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 }
Eric Dumazet77879142016-09-19 23:39:11 -0400619 }
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700620 if (rate != ~0UL) {
Eric Dumazet0eab5eb2013-10-01 09:10:16 -0700621 u64 len = (u64)plen * NSEC_PER_SEC;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700622
Eric Dumazet7eec4172013-10-08 15:16:00 -0700623 if (likely(rate))
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700624 len = div64_ul(len, rate);
Eric Dumazet0eab5eb2013-10-01 09:10:16 -0700625 /* Since socket rate can change later,
Eric Dumazetced7a042014-11-25 08:57:29 -0800626 * clamp the delay to 1 second.
627 * Really, providers of too big packets should be fixed !
Eric Dumazet0eab5eb2013-10-01 09:10:16 -0700628 */
Eric Dumazetced7a042014-11-25 08:57:29 -0800629 if (unlikely(len > NSEC_PER_SEC)) {
630 len = NSEC_PER_SEC;
Eric Dumazet0eab5eb2013-10-01 09:10:16 -0700631 q->stat_pkts_too_long++;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700632 }
Eric Dumazetfefa5692016-09-22 08:58:55 -0700633 /* 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);
Eric Dumazet0eab5eb2013-10-01 09:10:16 -0700639 f->time_next_packet = now + len;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700640 }
641out:
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700642 qdisc_bstats_update(sch, skb);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700643 return skb;
644}
645
Eric Dumazete14ffdf2016-06-13 20:21:53 -0700646static void fq_flow_purge(struct fq_flow *flow)
647{
Eric Dumazeteeb84aa2019-05-04 16:48:53 -0700648 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 }
Eric Dumazete14ffdf2016-06-13 20:21:53 -0700657 rtnl_kfree_skbs(flow->head, flow->tail);
658 flow->head = NULL;
659 flow->qlen = 0;
660}
661
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700662static void fq_reset(struct Qdisc *sch)
663{
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700664 struct fq_sched_data *q = qdisc_priv(sch);
665 struct rb_root *root;
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700666 struct rb_node *p;
667 struct fq_flow *f;
668 unsigned int idx;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700669
Eric Dumazete14ffdf2016-06-13 20:21:53 -0700670 sch->q.qlen = 0;
671 sch->qstats.backlog = 0;
672
673 fq_flow_purge(&q->internal);
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700674
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) {
Geliang Tange1245572016-12-20 22:02:15 +0800681 f = rb_entry(p, struct fq_flow, fq_node);
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700682 rb_erase(p, root);
683
Eric Dumazete14ffdf2016-06-13 20:21:53 -0700684 fq_flow_purge(f);
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700685
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;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700695}
696
697static 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);
Geliang Tange1245572016-12-20 22:02:15 +0800711 of = rb_entry(op, struct fq_flow, fq_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700712 if (fq_gc_candidate(of)) {
713 fcnt++;
714 kmem_cache_free(fq_flow_cachep, of);
715 continue;
716 }
Eric Dumazet29c58472016-11-17 09:48:30 -0800717 nroot = &new_array[hash_ptr(of->sk, new_log)];
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700718
719 np = &nroot->rb_node;
720 parent = NULL;
721 while (*np) {
722 parent = *np;
723
Geliang Tange1245572016-12-20 22:02:15 +0800724 nf = rb_entry(parent, struct fq_flow, fq_node);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700725 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
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800742static void fq_free(void *addr)
743{
WANG Cong4cb28972014-06-02 15:55:22 -0700744 kvfree(addr);
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800745}
746
747static int fq_resize(struct Qdisc *sch, u32 log)
748{
749 struct fq_sched_data *q = qdisc_priv(sch);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700750 struct rb_root *array;
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800751 void *old_fq_root;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700752 u32 idx;
753
754 if (q->fq_root && log == q->fq_trees_log)
755 return 0;
756
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800757 /* If XPS was setup, we can allocate memory on right NUMA node */
Michal Hockodcda9b02017-07-12 14:36:45 -0700758 array = kvmalloc_node(sizeof(struct rb_root) << log, GFP_KERNEL | __GFP_RETRY_MAYFAIL,
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800759 netdev_queue_numa_node_read(sch->dev_queue));
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700760 if (!array)
761 return -ENOMEM;
762
763 for (idx = 0; idx < (1U << log); idx++)
764 array[idx] = RB_ROOT;
765
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800766 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
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700772 q->fq_root = array;
773 q->fq_trees_log = log;
774
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800775 sch_tree_unlock(sch);
776
777 fq_free(old_fq_root);
778
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700779 return 0;
780}
781
Davide Caratti70411012023-04-20 16:59:46 +0200782static struct netlink_range_validation iq_range = {
783 .max = INT_MAX,
784};
785
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700786static const struct nla_policy fq_policy[TCA_FQ_MAX + 1] = {
Eric Dumazet583396f2020-03-16 19:12:51 -0700787 [TCA_FQ_UNSPEC] = { .strict_start_type = TCA_FQ_TIMER_SLACK },
788
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700789 [TCA_FQ_PLIMIT] = { .type = NLA_U32 },
790 [TCA_FQ_FLOW_PLIMIT] = { .type = NLA_U32 },
791 [TCA_FQ_QUANTUM] = { .type = NLA_U32 },
Davide Caratti70411012023-04-20 16:59:46 +0200792 [TCA_FQ_INITIAL_QUANTUM] = NLA_POLICY_FULL_RANGE(NLA_U32, &iq_range),
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700793 [TCA_FQ_RATE_ENABLE] = { .type = NLA_U32 },
794 [TCA_FQ_FLOW_DEFAULT_RATE] = { .type = NLA_U32 },
795 [TCA_FQ_FLOW_MAX_RATE] = { .type = NLA_U32 },
796 [TCA_FQ_BUCKETS_LOG] = { .type = NLA_U32 },
Eric Dumazetf52ed892013-11-15 08:58:14 -0800797 [TCA_FQ_FLOW_REFILL_DELAY] = { .type = NLA_U32 },
Jakub Kicinski7e6dc032020-03-02 21:05:19 -0800798 [TCA_FQ_ORPHAN_MASK] = { .type = NLA_U32 },
Eric Dumazet77879142016-09-19 23:39:11 -0400799 [TCA_FQ_LOW_RATE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet48872c12018-11-11 09:11:31 -0800800 [TCA_FQ_CE_THRESHOLD] = { .type = NLA_U32 },
Eric Dumazet583396f2020-03-16 19:12:51 -0700801 [TCA_FQ_TIMER_SLACK] = { .type = NLA_U32 },
Eric Dumazet39d01052020-05-01 07:07:41 -0700802 [TCA_FQ_HORIZON] = { .type = NLA_U32 },
803 [TCA_FQ_HORIZON_DROP] = { .type = NLA_U8 },
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700804};
805
Alexander Aring20307212017-12-20 12:35:14 -0500806static int fq_change(struct Qdisc *sch, struct nlattr *opt,
807 struct netlink_ext_ack *extack)
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700808{
809 struct fq_sched_data *q = qdisc_priv(sch);
810 struct nlattr *tb[TCA_FQ_MAX + 1];
811 int err, drop_count = 0;
WANG Cong2ccccf52016-02-25 14:55:01 -0800812 unsigned drop_len = 0;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700813 u32 fq_log;
814
Johannes Berg8cb08172019-04-26 14:07:28 +0200815 err = nla_parse_nested_deprecated(tb, TCA_FQ_MAX, opt, fq_policy,
816 NULL);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700817 if (err < 0)
818 return err;
819
820 sch_tree_lock(sch);
821
822 fq_log = q->fq_trees_log;
823
824 if (tb[TCA_FQ_BUCKETS_LOG]) {
825 u32 nval = nla_get_u32(tb[TCA_FQ_BUCKETS_LOG]);
826
827 if (nval >= 1 && nval <= ilog2(256*1024))
828 fq_log = nval;
829 else
830 err = -EINVAL;
831 }
832 if (tb[TCA_FQ_PLIMIT])
833 sch->limit = nla_get_u32(tb[TCA_FQ_PLIMIT]);
834
835 if (tb[TCA_FQ_FLOW_PLIMIT])
836 q->flow_plimit = nla_get_u32(tb[TCA_FQ_FLOW_PLIMIT]);
837
Kenneth Klette Jonassen3725a262015-02-03 17:49:18 +0100838 if (tb[TCA_FQ_QUANTUM]) {
839 u32 quantum = nla_get_u32(tb[TCA_FQ_QUANTUM]);
840
Eric Dumazetd9e15a22020-01-06 06:10:39 -0800841 if (quantum > 0 && quantum <= (1 << 20)) {
Kenneth Klette Jonassen3725a262015-02-03 17:49:18 +0100842 q->quantum = quantum;
Eric Dumazetd9e15a22020-01-06 06:10:39 -0800843 } else {
844 NL_SET_ERR_MSG_MOD(extack, "invalid quantum");
Kenneth Klette Jonassen3725a262015-02-03 17:49:18 +0100845 err = -EINVAL;
Eric Dumazetd9e15a22020-01-06 06:10:39 -0800846 }
Kenneth Klette Jonassen3725a262015-02-03 17:49:18 +0100847 }
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700848
849 if (tb[TCA_FQ_INITIAL_QUANTUM])
Eric Dumazetede869c2013-10-07 12:50:18 -0700850 q->initial_quantum = nla_get_u32(tb[TCA_FQ_INITIAL_QUANTUM]);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700851
852 if (tb[TCA_FQ_FLOW_DEFAULT_RATE])
Eric Dumazet65c51892013-11-15 08:57:26 -0800853 pr_warn_ratelimited("sch_fq: defrate %u ignored.\n",
854 nla_get_u32(tb[TCA_FQ_FLOW_DEFAULT_RATE]));
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700855
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700856 if (tb[TCA_FQ_FLOW_MAX_RATE]) {
857 u32 rate = nla_get_u32(tb[TCA_FQ_FLOW_MAX_RATE]);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700858
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700859 q->flow_max_rate = (rate == ~0U) ? ~0UL : rate;
860 }
Eric Dumazet77879142016-09-19 23:39:11 -0400861 if (tb[TCA_FQ_LOW_RATE_THRESHOLD])
862 q->low_rate_threshold =
863 nla_get_u32(tb[TCA_FQ_LOW_RATE_THRESHOLD]);
864
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700865 if (tb[TCA_FQ_RATE_ENABLE]) {
866 u32 enable = nla_get_u32(tb[TCA_FQ_RATE_ENABLE]);
867
868 if (enable <= 1)
869 q->rate_enable = enable;
870 else
871 err = -EINVAL;
872 }
873
Eric Dumazetf52ed892013-11-15 08:58:14 -0800874 if (tb[TCA_FQ_FLOW_REFILL_DELAY]) {
875 u32 usecs_delay = nla_get_u32(tb[TCA_FQ_FLOW_REFILL_DELAY]) ;
876
877 q->flow_refill_delay = usecs_to_jiffies(usecs_delay);
878 }
879
Eric Dumazet06eb3952015-02-04 21:30:40 -0800880 if (tb[TCA_FQ_ORPHAN_MASK])
881 q->orphan_mask = nla_get_u32(tb[TCA_FQ_ORPHAN_MASK]);
882
Eric Dumazet48872c12018-11-11 09:11:31 -0800883 if (tb[TCA_FQ_CE_THRESHOLD])
884 q->ce_threshold = (u64)NSEC_PER_USEC *
885 nla_get_u32(tb[TCA_FQ_CE_THRESHOLD]);
886
Eric Dumazet583396f2020-03-16 19:12:51 -0700887 if (tb[TCA_FQ_TIMER_SLACK])
888 q->timer_slack = nla_get_u32(tb[TCA_FQ_TIMER_SLACK]);
889
Eric Dumazet39d01052020-05-01 07:07:41 -0700890 if (tb[TCA_FQ_HORIZON])
891 q->horizon = (u64)NSEC_PER_USEC *
892 nla_get_u32(tb[TCA_FQ_HORIZON]);
893
894 if (tb[TCA_FQ_HORIZON_DROP])
895 q->horizon_drop = nla_get_u8(tb[TCA_FQ_HORIZON_DROP]);
896
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800897 if (!err) {
Eric Dumazet39d01052020-05-01 07:07:41 -0700898
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800899 sch_tree_unlock(sch);
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800900 err = fq_resize(sch, fq_log);
Eric Dumazet2d8d40a2014-03-06 22:57:52 -0800901 sch_tree_lock(sch);
902 }
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700903 while (sch->q.qlen > sch->limit) {
904 struct sk_buff *skb = fq_dequeue(sch);
905
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700906 if (!skb)
907 break;
WANG Cong2ccccf52016-02-25 14:55:01 -0800908 drop_len += qdisc_pkt_len(skb);
Eric Dumazete14ffdf2016-06-13 20:21:53 -0700909 rtnl_kfree_skbs(skb, skb);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700910 drop_count++;
911 }
WANG Cong2ccccf52016-02-25 14:55:01 -0800912 qdisc_tree_reduce_backlog(sch, drop_count, drop_len);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700913
914 sch_tree_unlock(sch);
915 return err;
916}
917
918static void fq_destroy(struct Qdisc *sch)
919{
920 struct fq_sched_data *q = qdisc_priv(sch);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700921
Eric Dumazet8d34ce12013-09-27 14:20:01 -0700922 fq_reset(sch);
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800923 fq_free(q->fq_root);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700924 qdisc_watchdog_cancel(&q->watchdog);
925}
926
Alexander Aringe63d7dfd2017-12-20 12:35:13 -0500927static int fq_init(struct Qdisc *sch, struct nlattr *opt,
928 struct netlink_ext_ack *extack)
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700929{
930 struct fq_sched_data *q = qdisc_priv(sch);
931 int err;
932
933 sch->limit = 10000;
934 q->flow_plimit = 100;
935 q->quantum = 2 * psched_mtu(qdisc_dev(sch));
936 q->initial_quantum = 10 * psched_mtu(qdisc_dev(sch));
Eric Dumazetf52ed892013-11-15 08:58:14 -0800937 q->flow_refill_delay = msecs_to_jiffies(40);
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700938 q->flow_max_rate = ~0UL;
Eric Dumazetfefa5692016-09-22 08:58:55 -0700939 q->time_next_delayed_flow = ~0ULL;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700940 q->rate_enable = 1;
941 q->new_flows.first = NULL;
942 q->old_flows.first = NULL;
943 q->delayed = RB_ROOT;
944 q->fq_root = NULL;
945 q->fq_trees_log = ilog2(1024);
Eric Dumazet06eb3952015-02-04 21:30:40 -0800946 q->orphan_mask = 1024 - 1;
Eric Dumazet77879142016-09-19 23:39:11 -0400947 q->low_rate_threshold = 550000 / 8;
Eric Dumazet48872c12018-11-11 09:11:31 -0800948
Eric Dumazet583396f2020-03-16 19:12:51 -0700949 q->timer_slack = 10 * NSEC_PER_USEC; /* 10 usec of hrtimer slack */
950
Eric Dumazet39d01052020-05-01 07:07:41 -0700951 q->horizon = 10ULL * NSEC_PER_SEC; /* 10 seconds */
952 q->horizon_drop = 1; /* by default, drop packets beyond horizon */
953
Eric Dumazet48872c12018-11-11 09:11:31 -0800954 /* Default ce_threshold of 4294 seconds */
955 q->ce_threshold = (u64)NSEC_PER_USEC * ~0U;
956
Eric Dumazetfb420d52018-09-28 10:28:44 -0700957 qdisc_watchdog_init_clockid(&q->watchdog, sch, CLOCK_MONOTONIC);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700958
959 if (opt)
Alexander Aring20307212017-12-20 12:35:14 -0500960 err = fq_change(sch, opt, extack);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700961 else
Eric Dumazetc3bd8542013-12-15 13:15:25 -0800962 err = fq_resize(sch, q->fq_trees_log);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700963
964 return err;
965}
966
967static int fq_dump(struct Qdisc *sch, struct sk_buff *skb)
968{
969 struct fq_sched_data *q = qdisc_priv(sch);
Eric Dumazet48872c12018-11-11 09:11:31 -0800970 u64 ce_threshold = q->ce_threshold;
Eric Dumazet39d01052020-05-01 07:07:41 -0700971 u64 horizon = q->horizon;
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700972 struct nlattr *opts;
973
Michal Kubecekae0be8d2019-04-26 11:13:06 +0200974 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700975 if (opts == NULL)
976 goto nla_put_failure;
977
Eric Dumazet65c51892013-11-15 08:57:26 -0800978 /* TCA_FQ_FLOW_DEFAULT_RATE is not used anymore */
979
Eric Dumazet48872c12018-11-11 09:11:31 -0800980 do_div(ce_threshold, NSEC_PER_USEC);
Eric Dumazet39d01052020-05-01 07:07:41 -0700981 do_div(horizon, NSEC_PER_USEC);
Eric Dumazet48872c12018-11-11 09:11:31 -0800982
Eric Dumazetafe4fd02013-08-29 15:49:55 -0700983 if (nla_put_u32(skb, TCA_FQ_PLIMIT, sch->limit) ||
984 nla_put_u32(skb, TCA_FQ_FLOW_PLIMIT, q->flow_plimit) ||
985 nla_put_u32(skb, TCA_FQ_QUANTUM, q->quantum) ||
986 nla_put_u32(skb, TCA_FQ_INITIAL_QUANTUM, q->initial_quantum) ||
987 nla_put_u32(skb, TCA_FQ_RATE_ENABLE, q->rate_enable) ||
Eric Dumazet76a9ebe2018-10-15 09:37:53 -0700988 nla_put_u32(skb, TCA_FQ_FLOW_MAX_RATE,
989 min_t(unsigned long, q->flow_max_rate, ~0U)) ||
Eric Dumazetf52ed892013-11-15 08:58:14 -0800990 nla_put_u32(skb, TCA_FQ_FLOW_REFILL_DELAY,
991 jiffies_to_usecs(q->flow_refill_delay)) ||
Eric Dumazet06eb3952015-02-04 21:30:40 -0800992 nla_put_u32(skb, TCA_FQ_ORPHAN_MASK, q->orphan_mask) ||
Eric Dumazet77879142016-09-19 23:39:11 -0400993 nla_put_u32(skb, TCA_FQ_LOW_RATE_THRESHOLD,
994 q->low_rate_threshold) ||
Eric Dumazet48872c12018-11-11 09:11:31 -0800995 nla_put_u32(skb, TCA_FQ_CE_THRESHOLD, (u32)ce_threshold) ||
Eric Dumazet583396f2020-03-16 19:12:51 -0700996 nla_put_u32(skb, TCA_FQ_BUCKETS_LOG, q->fq_trees_log) ||
Eric Dumazet39d01052020-05-01 07:07:41 -0700997 nla_put_u32(skb, TCA_FQ_TIMER_SLACK, q->timer_slack) ||
998 nla_put_u32(skb, TCA_FQ_HORIZON, (u32)horizon) ||
999 nla_put_u8(skb, TCA_FQ_HORIZON_DROP, q->horizon_drop))
Eric Dumazetafe4fd02013-08-29 15:49:55 -07001000 goto nla_put_failure;
1001
Yang Yingliangd59b7d82014-03-12 10:20:32 +08001002 return nla_nest_end(skb, opts);
Eric Dumazetafe4fd02013-08-29 15:49:55 -07001003
1004nla_put_failure:
1005 return -1;
1006}
1007
1008static int fq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1009{
1010 struct fq_sched_data *q = qdisc_priv(sch);
Eric Dumazet695b4ec2016-09-15 16:20:01 -07001011 struct tc_fq_qd_stats st;
1012
1013 sch_tree_lock(sch);
1014
1015 st.gc_flows = q->stat_gc_flows;
1016 st.highprio_packets = q->stat_internal_packets;
Eric Dumazet90caf672018-09-21 08:51:54 -07001017 st.tcp_retrans = 0;
Eric Dumazet695b4ec2016-09-15 16:20:01 -07001018 st.throttled = q->stat_throttled;
1019 st.flows_plimit = q->stat_flows_plimit;
1020 st.pkts_too_long = q->stat_pkts_too_long;
1021 st.allocation_errors = q->stat_allocation_errors;
Eric Dumazet583396f2020-03-16 19:12:51 -07001022 st.time_next_delayed_flow = q->time_next_delayed_flow + q->timer_slack -
1023 ktime_get_ns();
Eric Dumazet695b4ec2016-09-15 16:20:01 -07001024 st.flows = q->flows;
1025 st.inactive_flows = q->inactive_flows;
1026 st.throttled_flows = q->throttled_flows;
Eric Dumazetfefa5692016-09-22 08:58:55 -07001027 st.unthrottle_latency_ns = min_t(unsigned long,
1028 q->unthrottle_latency_ns, ~0U);
Eric Dumazet48872c12018-11-11 09:11:31 -08001029 st.ce_mark = q->stat_ce_mark;
Eric Dumazet39d01052020-05-01 07:07:41 -07001030 st.horizon_drops = q->stat_horizon_drops;
1031 st.horizon_caps = q->stat_horizon_caps;
Eric Dumazet695b4ec2016-09-15 16:20:01 -07001032 sch_tree_unlock(sch);
Eric Dumazetafe4fd02013-08-29 15:49:55 -07001033
1034 return gnet_stats_copy_app(d, &st, sizeof(st));
1035}
1036
1037static struct Qdisc_ops fq_qdisc_ops __read_mostly = {
1038 .id = "fq",
1039 .priv_size = sizeof(struct fq_sched_data),
1040
1041 .enqueue = fq_enqueue,
1042 .dequeue = fq_dequeue,
1043 .peek = qdisc_peek_dequeued,
1044 .init = fq_init,
1045 .reset = fq_reset,
1046 .destroy = fq_destroy,
1047 .change = fq_change,
1048 .dump = fq_dump,
1049 .dump_stats = fq_dump_stats,
1050 .owner = THIS_MODULE,
1051};
1052
1053static int __init fq_module_init(void)
1054{
1055 int ret;
1056
1057 fq_flow_cachep = kmem_cache_create("fq_flow_cache",
1058 sizeof(struct fq_flow),
1059 0, 0, NULL);
1060 if (!fq_flow_cachep)
1061 return -ENOMEM;
1062
1063 ret = register_qdisc(&fq_qdisc_ops);
1064 if (ret)
1065 kmem_cache_destroy(fq_flow_cachep);
1066 return ret;
1067}
1068
1069static void __exit fq_module_exit(void)
1070{
1071 unregister_qdisc(&fq_qdisc_ops);
1072 kmem_cache_destroy(fq_flow_cachep);
1073}
1074
1075module_init(fq_module_init)
1076module_exit(fq_module_exit)
1077MODULE_AUTHOR("Eric Dumazet");
1078MODULE_LICENSE("GPL");
Rob Gill67c20de2020-06-20 02:08:25 +00001079MODULE_DESCRIPTION("Fair Queue Packet Scheduler");