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 = unrcu_pointer(xchg(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 <asm/system.h>
18#include <linux/bitops.h>
19#include <linux/module.h>
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/jiffies.h>
23#include <linux/string.h>
24#include <linux/mm.h>
25#include <linux/socket.h>
26#include <linux/sockios.h>
27#include <linux/in.h>
28#include <linux/errno.h>
29#include <linux/interrupt.h>
30#include <linux/netdevice.h>
31#include <linux/skbuff.h>
32#include <linux/rtnetlink.h>
33#include <linux/init.h>
34#include <linux/rbtree.h>
35#include <linux/slab.h>
36#include <net/sock.h>
37#include <net/gen_stats.h>
38
39/*
40 This code is NOT intended to be used for statistics collection,
41 its purpose is to provide a base for statistical multiplexing
42 for controlled load service.
43 If you need only statistics, run a user level daemon which
44 periodically reads byte counters.
45
46 Unfortunately, rate estimation is not a very easy task.
47 F.e. I did not find a simple way to estimate the current peak rate
48 and even failed to formulate the problem 8)8)
49
50 So I preferred not to built an estimator into the scheduler,
51 but run this task separately.
52 Ideally, it should be kernel thread(s), but for now it runs
53 from timers, which puts apparent top bounds on the number of rated
54 flows, has minimal overhead on small, but is enough
55 to handle controlled load service, sets of aggregates.
56
57 We measure rate over A=(1<<interval) seconds and evaluate EWMA:
58
59 avrate = avrate*(1-W) + rate*W
60
61 where W is chosen as negative power of 2: W = 2^(-ewma_log)
62
63 The resulting time constant is:
64
65 T = A/(-ln(1-W))
66
67
68 NOTES.
69
70 * avbps is scaled by 2^5, avpps is scaled by 2^10.
71 * both values are reported as 32 bit unsigned values. bps can
72 overflow for fast links : max speed being 34360Mbit/sec
73 * Minimal interval is HZ/4=250msec (it is the greatest common divisor
74 for HZ=100 and HZ=1024 8)), maximal interval
75 is (HZ*2^EST_MAX_INTERVAL)/4 = 8sec. Shorter intervals
76 are too expensive, longer ones can be implemented
77 at user level painlessly.
78 */
79
80#define EST_MAX_INTERVAL 5
81
82struct gen_estimator
83{
84 struct list_head list;
85 struct gnet_stats_basic_packed *bstats;
86 struct gnet_stats_rate_est *rate_est;
87 spinlock_t *stats_lock;
88 int ewma_log;
89 u64 last_bytes;
90 u64 avbps;
91 u32 last_packets;
92 u32 avpps;
93 struct rcu_head e_rcu;
94 struct rb_node node;
95};
96
97struct gen_estimator_head
98{
99 struct timer_list timer;
100 struct list_head list;
101};
102
103static struct gen_estimator_head elist[EST_MAX_INTERVAL+1];
104
105/* Protects against NULL dereference */
106static DEFINE_RWLOCK(est_lock);
107
108/* Protects against soft lockup during large deletion */
109static struct rb_root est_root = RB_ROOT;
110static DEFINE_SPINLOCK(est_tree_lock);
111
112static void est_timer(unsigned long arg)
113{
114 int idx = (int)arg;
115 struct gen_estimator *e;
116
117 rcu_read_lock();
118 list_for_each_entry_rcu(e, &elist[idx].list, list) {
119 u64 nbytes;
120 u64 brate;
121 u32 npackets;
122 u32 rate;
123
124 spin_lock(e->stats_lock);
125 read_lock(&est_lock);
126 if (e->bstats == NULL)
127 goto skip;
128
129 nbytes = e->bstats->bytes;
130 npackets = e->bstats->packets;
131 brate = (nbytes - e->last_bytes)<<(7 - idx);
132 e->last_bytes = nbytes;
133 e->avbps += (brate >> e->ewma_log) - (e->avbps >> e->ewma_log);
134 e->rate_est->bps = (e->avbps+0xF)>>5;
135
136 rate = (npackets - e->last_packets)<<(12 - idx);
137 e->last_packets = npackets;
138 e->avpps += (rate >> e->ewma_log) - (e->avpps >> e->ewma_log);
139 e->rate_est->pps = (e->avpps+0x1FF)>>10;
140skip:
141 read_unlock(&est_lock);
142 spin_unlock(e->stats_lock);
143 }
144
145 if (!list_empty(&elist[idx].list))
146 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
147 rcu_read_unlock();
148}
149
150static void gen_add_node(struct gen_estimator *est)
151{
152 struct rb_node **p = &est_root.rb_node, *parent = NULL;
153
154 while (*p) {
155 struct gen_estimator *e;
156
157 parent = *p;
158 e = rb_entry(parent, struct gen_estimator, node);
159
160 if (est->bstats > e->bstats)
161 p = &parent->rb_right;
162 else
163 p = &parent->rb_left;
164 }
165 rb_link_node(&est->node, parent, p);
166 rb_insert_color(&est->node, &est_root);
167}
168
169static
170struct gen_estimator *gen_find_node(const struct gnet_stats_basic_packed *bstats,
171 const struct gnet_stats_rate_est *rate_est)
172{
173 struct rb_node *p = est_root.rb_node;
174
175 while (p) {
176 struct gen_estimator *e;
177
178 e = rb_entry(p, struct gen_estimator, node);
179
180 if (bstats > e->bstats)
181 p = p->rb_right;
182 else if (bstats < e->bstats || rate_est != e->rate_est)
183 p = p->rb_left;
184 else
185 return e;
186 }
187 return NULL;
188}
189
190/**
191 * gen_new_estimator - create a new rate estimator
192 * @bstats: basic statistics
193 * @rate_est: rate estimator statistics
194 * @stats_lock: statistics lock
195 * @opt: rate estimator configuration TLV
196 *
197 * Creates a new rate estimator with &bstats as source and &rate_est
198 * as destination. A new timer with the interval specified in the
199 * configuration TLV is created. Upon each interval, the latest statistics
200 * will be read from &bstats and the estimated rate will be stored in
201 * &rate_est with the statistics lock grabed during this period.
202 *
203 * Returns 0 on success or a negative error code.
204 *
205 */
206int gen_new_estimator(struct gnet_stats_basic_packed *bstats,
207 struct gnet_stats_rate_est *rate_est,
208 spinlock_t *stats_lock,
209 struct nlattr *opt)
210{
211 struct gen_estimator *est;
212 struct gnet_estimator *parm = nla_data(opt);
213 int idx;
214
215 if (nla_len(opt) < sizeof(*parm))
216 return -EINVAL;
217
218 if (parm->interval < -2 || parm->interval > 3)
219 return -EINVAL;
220
221 est = kzalloc(sizeof(*est), GFP_KERNEL);
222 if (est == NULL)
223 return -ENOBUFS;
224
225 idx = parm->interval + 2;
226 est->bstats = bstats;
227 est->rate_est = rate_est;
228 est->stats_lock = stats_lock;
229 est->ewma_log = parm->ewma_log;
230 est->last_bytes = bstats->bytes;
231 est->avbps = rate_est->bps<<5;
232 est->last_packets = bstats->packets;
233 est->avpps = rate_est->pps<<10;
234
235 spin_lock_bh(&est_tree_lock);
236 if (!elist[idx].timer.function) {
237 INIT_LIST_HEAD(&elist[idx].list);
238 setup_timer(&elist[idx].timer, est_timer, idx);
239 }
240
241 if (list_empty(&elist[idx].list))
242 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
243
244 list_add_rcu(&est->list, &elist[idx].list);
245 gen_add_node(est);
246 spin_unlock_bh(&est_tree_lock);
247
248 return 0;
249}
250EXPORT_SYMBOL(gen_new_estimator);
251
252/**
253 * gen_kill_estimator - remove a rate estimator
254 * @bstats: basic statistics
255 * @rate_est: rate estimator statistics
256 *
257 * Removes the rate estimator specified by &bstats and &rate_est.
258 *
259 * Note : Caller should respect an RCU grace period before freeing stats_lock
260 */
261void gen_kill_estimator(struct gnet_stats_basic_packed *bstats,
262 struct gnet_stats_rate_est *rate_est)
263{
264 struct gen_estimator *e;
265
266 spin_lock_bh(&est_tree_lock);
267 while ((e = gen_find_node(bstats, rate_est))) {
268 rb_erase(&e->node, &est_root);
269
270 write_lock(&est_lock);
271 e->bstats = NULL;
272 write_unlock(&est_lock);
273
274 list_del_rcu(&e->list);
275 kfree_rcu(e, e_rcu);
276 }
277 spin_unlock_bh(&est_tree_lock);
278}
279EXPORT_SYMBOL(gen_kill_estimator);
280
281/**
282 * gen_replace_estimator - replace rate estimator configuration
283 * @bstats: basic statistics
284 * @rate_est: rate estimator statistics
285 * @stats_lock: statistics lock
286 * @opt: rate estimator configuration TLV
287 *
288 * Replaces the configuration of a rate estimator by calling
289 * gen_kill_estimator() and gen_new_estimator().
290 *
291 * Returns 0 on success or a negative error code.
292 */
293int gen_replace_estimator(struct gnet_stats_basic_packed *bstats,
294 struct gnet_stats_rate_est *rate_est,
295 spinlock_t *stats_lock, struct nlattr *opt)
296{
297 gen_kill_estimator(bstats, rate_est);
298 return gen_new_estimator(bstats, rate_est, stats_lock, opt);
299}
300EXPORT_SYMBOL(gen_replace_estimator);
301
302/**
303 * gen_estimator_active - test if estimator is currently in use
304 * @bstats: basic statistics
305 * @rate_est: rate estimator statistics
306 *
307 * Returns true if estimator is active, and false if not.
308 */
309bool gen_estimator_active(const struct gnet_stats_basic_packed *bstats,
310 const struct gnet_stats_rate_est *rate_est)
311{
312 bool res;
313
314 ASSERT_RTNL();
315
316 spin_lock_bh(&est_tree_lock);
317 res = gen_find_node(bstats, rate_est) != NULL;
318 spin_unlock_bh(&est_tree_lock);
319
320 return res;
321}
322EXPORT_SYMBOL(gen_estimator_active);