Loading...
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);
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_packed *bstats;
44 spinlock_t *stats_lock;
45 seqcount_t *running;
46 struct gnet_stats_basic_cpu __percpu *cpu_bstats;
47 u8 ewma_log;
48 u8 intvl_log; /* period : (250ms << intvl_log) */
49
50 seqcount_t seq;
51 u32 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_packed *b)
64{
65 memset(b, 0, sizeof(*b));
66 if (e->stats_lock)
67 spin_lock(e->stats_lock);
68
69 __gnet_stats_copy_basic(e->running, b, e->cpu_bstats, e->bstats);
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_packed b;
80 u64 rate, brate;
81
82 est_fetch_counters(est, &b);
83 brate = (b.bytes - est->last_bytes) << (10 - est->ewma_log - est->intvl_log);
84 brate -= (est->avbps >> est->ewma_log);
85
86 rate = (u64)(b.packets - est->last_packets) << (10 - est->ewma_log - est->intvl_log);
87 rate -= (est->avpps >> est->ewma_log);
88
89 write_seqcount_begin(&est->seq);
90 est->avbps += brate;
91 est->avpps += rate;
92 write_seqcount_end(&est->seq);
93
94 est->last_bytes = b.bytes;
95 est->last_packets = b.packets;
96
97 est->next_jiffies += ((HZ/4) << est->intvl_log);
98
99 if (unlikely(time_after_eq(jiffies, est->next_jiffies))) {
100 /* Ouch... timer was delayed. */
101 est->next_jiffies = jiffies + 1;
102 }
103 mod_timer(&est->timer, est->next_jiffies);
104}
105
106/**
107 * gen_new_estimator - create a new rate estimator
108 * @bstats: basic statistics
109 * @cpu_bstats: bstats per cpu
110 * @rate_est: rate estimator statistics
111 * @lock: lock for statistics and control path
112 * @running: qdisc running seqcount
113 * @opt: rate estimator configuration TLV
114 *
115 * Creates a new rate estimator with &bstats as source and &rate_est
116 * as destination. A new timer with the interval specified in the
117 * configuration TLV is created. Upon each interval, the latest statistics
118 * will be read from &bstats and the estimated rate will be stored in
119 * &rate_est with the statistics lock grabbed during this period.
120 *
121 * Returns 0 on success or a negative error code.
122 *
123 */
124int gen_new_estimator(struct gnet_stats_basic_packed *bstats,
125 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
126 struct net_rate_estimator __rcu **rate_est,
127 spinlock_t *lock,
128 seqcount_t *running,
129 struct nlattr *opt)
130{
131 struct gnet_estimator *parm = nla_data(opt);
132 struct net_rate_estimator *old, *est;
133 struct gnet_stats_basic_packed b;
134 int intvl_log;
135
136 if (nla_len(opt) < sizeof(*parm))
137 return -EINVAL;
138
139 /* allowed timer periods are :
140 * -2 : 250ms, -1 : 500ms, 0 : 1 sec
141 * 1 : 2 sec, 2 : 4 sec, 3 : 8 sec
142 */
143 if (parm->interval < -2 || parm->interval > 3)
144 return -EINVAL;
145
146 est = kzalloc(sizeof(*est), GFP_KERNEL);
147 if (!est)
148 return -ENOBUFS;
149
150 seqcount_init(&est->seq);
151 intvl_log = parm->interval + 2;
152 est->bstats = bstats;
153 est->stats_lock = lock;
154 est->running = running;
155 est->ewma_log = parm->ewma_log;
156 est->intvl_log = intvl_log;
157 est->cpu_bstats = cpu_bstats;
158
159 if (lock)
160 local_bh_disable();
161 est_fetch_counters(est, &b);
162 if (lock)
163 local_bh_enable();
164 est->last_bytes = b.bytes;
165 est->last_packets = b.packets;
166
167 if (lock)
168 spin_lock_bh(lock);
169 old = rcu_dereference_protected(*rate_est, 1);
170 if (old) {
171 del_timer_sync(&old->timer);
172 est->avbps = old->avbps;
173 est->avpps = old->avpps;
174 }
175
176 est->next_jiffies = jiffies + ((HZ/4) << intvl_log);
177 timer_setup(&est->timer, est_timer, 0);
178 mod_timer(&est->timer, est->next_jiffies);
179
180 rcu_assign_pointer(*rate_est, est);
181 if (lock)
182 spin_unlock_bh(lock);
183 if (old)
184 kfree_rcu(old, rcu);
185 return 0;
186}
187EXPORT_SYMBOL(gen_new_estimator);
188
189/**
190 * gen_kill_estimator - remove a rate estimator
191 * @rate_est: rate estimator
192 *
193 * Removes the rate estimator.
194 *
195 */
196void gen_kill_estimator(struct net_rate_estimator __rcu **rate_est)
197{
198 struct net_rate_estimator *est;
199
200 est = xchg((__force struct net_rate_estimator **)rate_est, NULL);
201 if (est) {
202 del_timer_sync(&est->timer);
203 kfree_rcu(est, rcu);
204 }
205}
206EXPORT_SYMBOL(gen_kill_estimator);
207
208/**
209 * gen_replace_estimator - replace rate estimator configuration
210 * @bstats: basic statistics
211 * @cpu_bstats: bstats per cpu
212 * @rate_est: rate estimator statistics
213 * @lock: lock for statistics and control path
214 * @running: qdisc running seqcount (might be NULL)
215 * @opt: rate estimator configuration TLV
216 *
217 * Replaces the configuration of a rate estimator by calling
218 * gen_kill_estimator() and gen_new_estimator().
219 *
220 * Returns 0 on success or a negative error code.
221 */
222int gen_replace_estimator(struct gnet_stats_basic_packed *bstats,
223 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
224 struct net_rate_estimator __rcu **rate_est,
225 spinlock_t *lock,
226 seqcount_t *running, struct nlattr *opt)
227{
228 return gen_new_estimator(bstats, cpu_bstats, rate_est,
229 lock, running, opt);
230}
231EXPORT_SYMBOL(gen_replace_estimator);
232
233/**
234 * gen_estimator_active - test if estimator is currently in use
235 * @rate_est: rate estimator
236 *
237 * Returns true if estimator is active, and false if not.
238 */
239bool gen_estimator_active(struct net_rate_estimator __rcu **rate_est)
240{
241 return !!rcu_access_pointer(*rate_est);
242}
243EXPORT_SYMBOL(gen_estimator_active);
244
245bool gen_estimator_read(struct net_rate_estimator __rcu **rate_est,
246 struct gnet_stats_rate_est64 *sample)
247{
248 struct net_rate_estimator *est;
249 unsigned seq;
250
251 rcu_read_lock();
252 est = rcu_dereference(*rate_est);
253 if (!est) {
254 rcu_read_unlock();
255 return false;
256 }
257
258 do {
259 seq = read_seqcount_begin(&est->seq);
260 sample->bps = est->avbps >> 8;
261 sample->pps = est->avpps >> 8;
262 } while (read_seqcount_retry(&est->seq, seq));
263
264 rcu_read_unlock();
265 return true;
266}
267EXPORT_SYMBOL(gen_estimator_read);