Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/sched/gen_estimator.c Simple rate estimator.
4 *
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Eric Dumazet <edumazet@google.com>
7 *
8 * Changes:
9 * Jamal Hadi Salim - moved it to net/core and reshulfed
10 * names to make it usable in general net subsystem.
11 */
12
13#include <linux/uaccess.h>
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/jiffies.h>
19#include <linux/string.h>
20#include <linux/mm.h>
21#include <linux/socket.h>
22#include <linux/sockios.h>
23#include <linux/in.h>
24#include <linux/errno.h>
25#include <linux/interrupt.h>
26#include <linux/netdevice.h>
27#include <linux/skbuff.h>
28#include <linux/rtnetlink.h>
29#include <linux/init.h>
30#include <linux/slab.h>
31#include <linux/seqlock.h>
32#include <net/sock.h>
33#include <net/gen_stats.h>
34
35/* This code is NOT intended to be used for statistics collection,
36 * its purpose is to provide a base for statistical multiplexing
37 * for controlled load service.
38 * If you need only statistics, run a user level daemon which
39 * periodically reads byte counters.
40 */
41
42struct net_rate_estimator {
43 struct gnet_stats_basic_sync *bstats;
44 spinlock_t *stats_lock;
45 bool running;
46 struct gnet_stats_basic_sync __percpu *cpu_bstats;
47 u8 ewma_log;
48 u8 intvl_log; /* period : (250ms << intvl_log) */
49
50 seqcount_t seq;
51 u64 last_packets;
52 u64 last_bytes;
53
54 u64 avpps;
55 u64 avbps;
56
57 unsigned long next_jiffies;
58 struct timer_list timer;
59 struct rcu_head rcu;
60};
61
62static void est_fetch_counters(struct net_rate_estimator *e,
63 struct gnet_stats_basic_sync *b)
64{
65 gnet_stats_basic_sync_init(b);
66 if (e->stats_lock)
67 spin_lock(e->stats_lock);
68
69 gnet_stats_add_basic(b, e->cpu_bstats, e->bstats, e->running);
70
71 if (e->stats_lock)
72 spin_unlock(e->stats_lock);
73
74}
75
76static void est_timer(struct timer_list *t)
77{
78 struct net_rate_estimator *est = from_timer(est, t, timer);
79 struct gnet_stats_basic_sync b;
80 u64 b_bytes, b_packets;
81 u64 rate, brate;
82
83 est_fetch_counters(est, &b);
84 b_bytes = u64_stats_read(&b.bytes);
85 b_packets = u64_stats_read(&b.packets);
86
87 brate = (b_bytes - est->last_bytes) << (10 - est->intvl_log);
88 brate = (brate >> est->ewma_log) - (est->avbps >> est->ewma_log);
89
90 rate = (b_packets - est->last_packets) << (10 - est->intvl_log);
91 rate = (rate >> est->ewma_log) - (est->avpps >> est->ewma_log);
92
93 write_seqcount_begin(&est->seq);
94 est->avbps += brate;
95 est->avpps += rate;
96 write_seqcount_end(&est->seq);
97
98 est->last_bytes = b_bytes;
99 est->last_packets = b_packets;
100
101 est->next_jiffies += ((HZ/4) << est->intvl_log);
102
103 if (unlikely(time_after_eq(jiffies, est->next_jiffies))) {
104 /* Ouch... timer was delayed. */
105 est->next_jiffies = jiffies + 1;
106 }
107 mod_timer(&est->timer, est->next_jiffies);
108}
109
110/**
111 * gen_new_estimator - create a new rate estimator
112 * @bstats: basic statistics
113 * @cpu_bstats: bstats per cpu
114 * @rate_est: rate estimator statistics
115 * @lock: lock for statistics and control path
116 * @running: true if @bstats represents a running qdisc, thus @bstats'
117 * internal values might change during basic reads. Only used
118 * if @bstats_cpu is NULL
119 * @opt: rate estimator configuration TLV
120 *
121 * Creates a new rate estimator with &bstats as source and &rate_est
122 * as destination. A new timer with the interval specified in the
123 * configuration TLV is created. Upon each interval, the latest statistics
124 * will be read from &bstats and the estimated rate will be stored in
125 * &rate_est with the statistics lock grabbed during this period.
126 *
127 * Returns 0 on success or a negative error code.
128 *
129 */
130int gen_new_estimator(struct gnet_stats_basic_sync *bstats,
131 struct gnet_stats_basic_sync __percpu *cpu_bstats,
132 struct net_rate_estimator __rcu **rate_est,
133 spinlock_t *lock,
134 bool running,
135 struct nlattr *opt)
136{
137 struct gnet_estimator *parm = nla_data(opt);
138 struct net_rate_estimator *old, *est;
139 struct gnet_stats_basic_sync b;
140 int intvl_log;
141
142 if (nla_len(opt) < sizeof(*parm))
143 return -EINVAL;
144
145 /* allowed timer periods are :
146 * -2 : 250ms, -1 : 500ms, 0 : 1 sec
147 * 1 : 2 sec, 2 : 4 sec, 3 : 8 sec
148 */
149 if (parm->interval < -2 || parm->interval > 3)
150 return -EINVAL;
151
152 if (parm->ewma_log == 0 || parm->ewma_log >= 31)
153 return -EINVAL;
154
155 est = kzalloc(sizeof(*est), GFP_KERNEL);
156 if (!est)
157 return -ENOBUFS;
158
159 seqcount_init(&est->seq);
160 intvl_log = parm->interval + 2;
161 est->bstats = bstats;
162 est->stats_lock = lock;
163 est->running = running;
164 est->ewma_log = parm->ewma_log;
165 est->intvl_log = intvl_log;
166 est->cpu_bstats = cpu_bstats;
167
168 if (lock)
169 local_bh_disable();
170 est_fetch_counters(est, &b);
171 if (lock)
172 local_bh_enable();
173 est->last_bytes = u64_stats_read(&b.bytes);
174 est->last_packets = u64_stats_read(&b.packets);
175
176 if (lock)
177 spin_lock_bh(lock);
178 old = rcu_dereference_protected(*rate_est, 1);
179 if (old) {
180 del_timer_sync(&old->timer);
181 est->avbps = old->avbps;
182 est->avpps = old->avpps;
183 }
184
185 est->next_jiffies = jiffies + ((HZ/4) << intvl_log);
186 timer_setup(&est->timer, est_timer, 0);
187 mod_timer(&est->timer, est->next_jiffies);
188
189 rcu_assign_pointer(*rate_est, est);
190 if (lock)
191 spin_unlock_bh(lock);
192 if (old)
193 kfree_rcu(old, rcu);
194 return 0;
195}
196EXPORT_SYMBOL(gen_new_estimator);
197
198/**
199 * gen_kill_estimator - remove a rate estimator
200 * @rate_est: rate estimator
201 *
202 * Removes the rate estimator.
203 *
204 */
205void gen_kill_estimator(struct net_rate_estimator __rcu **rate_est)
206{
207 struct net_rate_estimator *est;
208
209 est = xchg((__force struct net_rate_estimator **)rate_est, NULL);
210 if (est) {
211 timer_shutdown_sync(&est->timer);
212 kfree_rcu(est, rcu);
213 }
214}
215EXPORT_SYMBOL(gen_kill_estimator);
216
217/**
218 * gen_replace_estimator - replace rate estimator configuration
219 * @bstats: basic statistics
220 * @cpu_bstats: bstats per cpu
221 * @rate_est: rate estimator statistics
222 * @lock: lock for statistics and control path
223 * @running: true if @bstats represents a running qdisc, thus @bstats'
224 * internal values might change during basic reads. Only used
225 * if @cpu_bstats is NULL
226 * @opt: rate estimator configuration TLV
227 *
228 * Replaces the configuration of a rate estimator by calling
229 * gen_kill_estimator() and gen_new_estimator().
230 *
231 * Returns 0 on success or a negative error code.
232 */
233int gen_replace_estimator(struct gnet_stats_basic_sync *bstats,
234 struct gnet_stats_basic_sync __percpu *cpu_bstats,
235 struct net_rate_estimator __rcu **rate_est,
236 spinlock_t *lock,
237 bool running, struct nlattr *opt)
238{
239 return gen_new_estimator(bstats, cpu_bstats, rate_est,
240 lock, running, opt);
241}
242EXPORT_SYMBOL(gen_replace_estimator);
243
244/**
245 * gen_estimator_active - test if estimator is currently in use
246 * @rate_est: rate estimator
247 *
248 * Returns true if estimator is active, and false if not.
249 */
250bool gen_estimator_active(struct net_rate_estimator __rcu **rate_est)
251{
252 return !!rcu_access_pointer(*rate_est);
253}
254EXPORT_SYMBOL(gen_estimator_active);
255
256bool gen_estimator_read(struct net_rate_estimator __rcu **rate_est,
257 struct gnet_stats_rate_est64 *sample)
258{
259 struct net_rate_estimator *est;
260 unsigned seq;
261
262 rcu_read_lock();
263 est = rcu_dereference(*rate_est);
264 if (!est) {
265 rcu_read_unlock();
266 return false;
267 }
268
269 do {
270 seq = read_seqcount_begin(&est->seq);
271 sample->bps = est->avbps >> 8;
272 sample->pps = est->avpps >> 8;
273 } while (read_seqcount_retry(&est->seq, seq));
274
275 rcu_read_unlock();
276 return true;
277}
278EXPORT_SYMBOL(gen_estimator_read);
1/*
2 * net/sched/gen_estimator.c Simple rate estimator.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 *
11 * Changes:
12 * Jamal Hadi Salim - moved it to net/core and reshulfed
13 * names to make it usable in general net subsystem.
14 */
15
16#include <asm/uaccess.h>
17#include <linux/bitops.h>
18#include <linux/module.h>
19#include <linux/types.h>
20#include <linux/kernel.h>
21#include <linux/jiffies.h>
22#include <linux/string.h>
23#include <linux/mm.h>
24#include <linux/socket.h>
25#include <linux/sockios.h>
26#include <linux/in.h>
27#include <linux/errno.h>
28#include <linux/interrupt.h>
29#include <linux/netdevice.h>
30#include <linux/skbuff.h>
31#include <linux/rtnetlink.h>
32#include <linux/init.h>
33#include <linux/rbtree.h>
34#include <linux/slab.h>
35#include <net/sock.h>
36#include <net/gen_stats.h>
37
38/*
39 This code is NOT intended to be used for statistics collection,
40 its purpose is to provide a base for statistical multiplexing
41 for controlled load service.
42 If you need only statistics, run a user level daemon which
43 periodically reads byte counters.
44
45 Unfortunately, rate estimation is not a very easy task.
46 F.e. I did not find a simple way to estimate the current peak rate
47 and even failed to formulate the problem 8)8)
48
49 So I preferred not to built an estimator into the scheduler,
50 but run this task separately.
51 Ideally, it should be kernel thread(s), but for now it runs
52 from timers, which puts apparent top bounds on the number of rated
53 flows, has minimal overhead on small, but is enough
54 to handle controlled load service, sets of aggregates.
55
56 We measure rate over A=(1<<interval) seconds and evaluate EWMA:
57
58 avrate = avrate*(1-W) + rate*W
59
60 where W is chosen as negative power of 2: W = 2^(-ewma_log)
61
62 The resulting time constant is:
63
64 T = A/(-ln(1-W))
65
66
67 NOTES.
68
69 * avbps is scaled by 2^5, avpps is scaled by 2^10.
70 * both values are reported as 32 bit unsigned values. bps can
71 overflow for fast links : max speed being 34360Mbit/sec
72 * Minimal interval is HZ/4=250msec (it is the greatest common divisor
73 for HZ=100 and HZ=1024 8)), maximal interval
74 is (HZ*2^EST_MAX_INTERVAL)/4 = 8sec. Shorter intervals
75 are too expensive, longer ones can be implemented
76 at user level painlessly.
77 */
78
79#define EST_MAX_INTERVAL 5
80
81struct gen_estimator
82{
83 struct list_head list;
84 struct gnet_stats_basic_packed *bstats;
85 struct gnet_stats_rate_est *rate_est;
86 spinlock_t *stats_lock;
87 int ewma_log;
88 u64 last_bytes;
89 u64 avbps;
90 u32 last_packets;
91 u32 avpps;
92 struct rcu_head e_rcu;
93 struct rb_node node;
94};
95
96struct gen_estimator_head
97{
98 struct timer_list timer;
99 struct list_head list;
100};
101
102static struct gen_estimator_head elist[EST_MAX_INTERVAL+1];
103
104/* Protects against NULL dereference */
105static DEFINE_RWLOCK(est_lock);
106
107/* Protects against soft lockup during large deletion */
108static struct rb_root est_root = RB_ROOT;
109static DEFINE_SPINLOCK(est_tree_lock);
110
111static void est_timer(unsigned long arg)
112{
113 int idx = (int)arg;
114 struct gen_estimator *e;
115
116 rcu_read_lock();
117 list_for_each_entry_rcu(e, &elist[idx].list, list) {
118 u64 nbytes;
119 u64 brate;
120 u32 npackets;
121 u32 rate;
122
123 spin_lock(e->stats_lock);
124 read_lock(&est_lock);
125 if (e->bstats == NULL)
126 goto skip;
127
128 nbytes = e->bstats->bytes;
129 npackets = e->bstats->packets;
130 brate = (nbytes - e->last_bytes)<<(7 - idx);
131 e->last_bytes = nbytes;
132 e->avbps += (brate >> e->ewma_log) - (e->avbps >> e->ewma_log);
133 e->rate_est->bps = (e->avbps+0xF)>>5;
134
135 rate = (npackets - e->last_packets)<<(12 - idx);
136 e->last_packets = npackets;
137 e->avpps += (rate >> e->ewma_log) - (e->avpps >> e->ewma_log);
138 e->rate_est->pps = (e->avpps+0x1FF)>>10;
139skip:
140 read_unlock(&est_lock);
141 spin_unlock(e->stats_lock);
142 }
143
144 if (!list_empty(&elist[idx].list))
145 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
146 rcu_read_unlock();
147}
148
149static void gen_add_node(struct gen_estimator *est)
150{
151 struct rb_node **p = &est_root.rb_node, *parent = NULL;
152
153 while (*p) {
154 struct gen_estimator *e;
155
156 parent = *p;
157 e = rb_entry(parent, struct gen_estimator, node);
158
159 if (est->bstats > e->bstats)
160 p = &parent->rb_right;
161 else
162 p = &parent->rb_left;
163 }
164 rb_link_node(&est->node, parent, p);
165 rb_insert_color(&est->node, &est_root);
166}
167
168static
169struct gen_estimator *gen_find_node(const struct gnet_stats_basic_packed *bstats,
170 const struct gnet_stats_rate_est *rate_est)
171{
172 struct rb_node *p = est_root.rb_node;
173
174 while (p) {
175 struct gen_estimator *e;
176
177 e = rb_entry(p, struct gen_estimator, node);
178
179 if (bstats > e->bstats)
180 p = p->rb_right;
181 else if (bstats < e->bstats || rate_est != e->rate_est)
182 p = p->rb_left;
183 else
184 return e;
185 }
186 return NULL;
187}
188
189/**
190 * gen_new_estimator - create a new rate estimator
191 * @bstats: basic statistics
192 * @rate_est: rate estimator statistics
193 * @stats_lock: statistics lock
194 * @opt: rate estimator configuration TLV
195 *
196 * Creates a new rate estimator with &bstats as source and &rate_est
197 * as destination. A new timer with the interval specified in the
198 * configuration TLV is created. Upon each interval, the latest statistics
199 * will be read from &bstats and the estimated rate will be stored in
200 * &rate_est with the statistics lock grabed during this period.
201 *
202 * Returns 0 on success or a negative error code.
203 *
204 */
205int gen_new_estimator(struct gnet_stats_basic_packed *bstats,
206 struct gnet_stats_rate_est *rate_est,
207 spinlock_t *stats_lock,
208 struct nlattr *opt)
209{
210 struct gen_estimator *est;
211 struct gnet_estimator *parm = nla_data(opt);
212 int idx;
213
214 if (nla_len(opt) < sizeof(*parm))
215 return -EINVAL;
216
217 if (parm->interval < -2 || parm->interval > 3)
218 return -EINVAL;
219
220 est = kzalloc(sizeof(*est), GFP_KERNEL);
221 if (est == NULL)
222 return -ENOBUFS;
223
224 idx = parm->interval + 2;
225 est->bstats = bstats;
226 est->rate_est = rate_est;
227 est->stats_lock = stats_lock;
228 est->ewma_log = parm->ewma_log;
229 est->last_bytes = bstats->bytes;
230 est->avbps = rate_est->bps<<5;
231 est->last_packets = bstats->packets;
232 est->avpps = rate_est->pps<<10;
233
234 spin_lock_bh(&est_tree_lock);
235 if (!elist[idx].timer.function) {
236 INIT_LIST_HEAD(&elist[idx].list);
237 setup_timer(&elist[idx].timer, est_timer, idx);
238 }
239
240 if (list_empty(&elist[idx].list))
241 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
242
243 list_add_rcu(&est->list, &elist[idx].list);
244 gen_add_node(est);
245 spin_unlock_bh(&est_tree_lock);
246
247 return 0;
248}
249EXPORT_SYMBOL(gen_new_estimator);
250
251/**
252 * gen_kill_estimator - remove a rate estimator
253 * @bstats: basic statistics
254 * @rate_est: rate estimator statistics
255 *
256 * Removes the rate estimator specified by &bstats and &rate_est.
257 *
258 * Note : Caller should respect an RCU grace period before freeing stats_lock
259 */
260void gen_kill_estimator(struct gnet_stats_basic_packed *bstats,
261 struct gnet_stats_rate_est *rate_est)
262{
263 struct gen_estimator *e;
264
265 spin_lock_bh(&est_tree_lock);
266 while ((e = gen_find_node(bstats, rate_est))) {
267 rb_erase(&e->node, &est_root);
268
269 write_lock(&est_lock);
270 e->bstats = NULL;
271 write_unlock(&est_lock);
272
273 list_del_rcu(&e->list);
274 kfree_rcu(e, e_rcu);
275 }
276 spin_unlock_bh(&est_tree_lock);
277}
278EXPORT_SYMBOL(gen_kill_estimator);
279
280/**
281 * gen_replace_estimator - replace rate estimator configuration
282 * @bstats: basic statistics
283 * @rate_est: rate estimator statistics
284 * @stats_lock: statistics lock
285 * @opt: rate estimator configuration TLV
286 *
287 * Replaces the configuration of a rate estimator by calling
288 * gen_kill_estimator() and gen_new_estimator().
289 *
290 * Returns 0 on success or a negative error code.
291 */
292int gen_replace_estimator(struct gnet_stats_basic_packed *bstats,
293 struct gnet_stats_rate_est *rate_est,
294 spinlock_t *stats_lock, struct nlattr *opt)
295{
296 gen_kill_estimator(bstats, rate_est);
297 return gen_new_estimator(bstats, rate_est, stats_lock, opt);
298}
299EXPORT_SYMBOL(gen_replace_estimator);
300
301/**
302 * gen_estimator_active - test if estimator is currently in use
303 * @bstats: basic statistics
304 * @rate_est: rate estimator statistics
305 *
306 * Returns true if estimator is active, and false if not.
307 */
308bool gen_estimator_active(const struct gnet_stats_basic_packed *bstats,
309 const struct gnet_stats_rate_est *rate_est)
310{
311 bool res;
312
313 ASSERT_RTNL();
314
315 spin_lock_bh(&est_tree_lock);
316 res = gen_find_node(bstats, rate_est) != NULL;
317 spin_unlock_bh(&est_tree_lock);
318
319 return res;
320}
321EXPORT_SYMBOL(gen_estimator_active);