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/*
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 and avpps are scaled by 2^5.
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_est64 *rate_est;
86 spinlock_t *stats_lock;
87 int ewma_log;
88 u32 last_packets;
89 unsigned long avpps;
90 u64 last_bytes;
91 u64 avbps;
92 struct rcu_head e_rcu;
93 struct rb_node node;
94 struct gnet_stats_basic_cpu __percpu *cpu_bstats;
95 struct rcu_head head;
96};
97
98struct gen_estimator_head
99{
100 struct timer_list timer;
101 struct list_head list;
102};
103
104static struct gen_estimator_head elist[EST_MAX_INTERVAL+1];
105
106/* Protects against NULL dereference */
107static DEFINE_RWLOCK(est_lock);
108
109/* Protects against soft lockup during large deletion */
110static struct rb_root est_root = RB_ROOT;
111static DEFINE_SPINLOCK(est_tree_lock);
112
113static void est_timer(unsigned long arg)
114{
115 int idx = (int)arg;
116 struct gen_estimator *e;
117
118 rcu_read_lock();
119 list_for_each_entry_rcu(e, &elist[idx].list, list) {
120 struct gnet_stats_basic_packed b = {0};
121 unsigned long rate;
122 u64 brate;
123
124 spin_lock(e->stats_lock);
125 read_lock(&est_lock);
126 if (e->bstats == NULL)
127 goto skip;
128
129 __gnet_stats_copy_basic(&b, e->cpu_bstats, e->bstats);
130
131 brate = (b.bytes - e->last_bytes)<<(7 - idx);
132 e->last_bytes = b.bytes;
133 e->avbps += (brate >> e->ewma_log) - (e->avbps >> e->ewma_log);
134 e->rate_est->bps = (e->avbps+0xF)>>5;
135
136 rate = b.packets - e->last_packets;
137 rate <<= (7 - idx);
138 e->last_packets = b.packets;
139 e->avpps += (rate >> e->ewma_log) - (e->avpps >> e->ewma_log);
140 e->rate_est->pps = (e->avpps + 0xF) >> 5;
141skip:
142 read_unlock(&est_lock);
143 spin_unlock(e->stats_lock);
144 }
145
146 if (!list_empty(&elist[idx].list))
147 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
148 rcu_read_unlock();
149}
150
151static void gen_add_node(struct gen_estimator *est)
152{
153 struct rb_node **p = &est_root.rb_node, *parent = NULL;
154
155 while (*p) {
156 struct gen_estimator *e;
157
158 parent = *p;
159 e = rb_entry(parent, struct gen_estimator, node);
160
161 if (est->bstats > e->bstats)
162 p = &parent->rb_right;
163 else
164 p = &parent->rb_left;
165 }
166 rb_link_node(&est->node, parent, p);
167 rb_insert_color(&est->node, &est_root);
168}
169
170static
171struct gen_estimator *gen_find_node(const struct gnet_stats_basic_packed *bstats,
172 const struct gnet_stats_rate_est64 *rate_est)
173{
174 struct rb_node *p = est_root.rb_node;
175
176 while (p) {
177 struct gen_estimator *e;
178
179 e = rb_entry(p, struct gen_estimator, node);
180
181 if (bstats > e->bstats)
182 p = p->rb_right;
183 else if (bstats < e->bstats || rate_est != e->rate_est)
184 p = p->rb_left;
185 else
186 return e;
187 }
188 return NULL;
189}
190
191/**
192 * gen_new_estimator - create a new rate estimator
193 * @bstats: basic statistics
194 * @cpu_bstats: bstats per cpu
195 * @rate_est: rate estimator statistics
196 * @stats_lock: statistics lock
197 * @opt: rate estimator configuration TLV
198 *
199 * Creates a new rate estimator with &bstats as source and &rate_est
200 * as destination. A new timer with the interval specified in the
201 * configuration TLV is created. Upon each interval, the latest statistics
202 * will be read from &bstats and the estimated rate will be stored in
203 * &rate_est with the statistics lock grabbed during this period.
204 *
205 * Returns 0 on success or a negative error code.
206 *
207 */
208int gen_new_estimator(struct gnet_stats_basic_packed *bstats,
209 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
210 struct gnet_stats_rate_est64 *rate_est,
211 spinlock_t *stats_lock,
212 struct nlattr *opt)
213{
214 struct gen_estimator *est;
215 struct gnet_estimator *parm = nla_data(opt);
216 struct gnet_stats_basic_packed b = {0};
217 int idx;
218
219 if (nla_len(opt) < sizeof(*parm))
220 return -EINVAL;
221
222 if (parm->interval < -2 || parm->interval > 3)
223 return -EINVAL;
224
225 est = kzalloc(sizeof(*est), GFP_KERNEL);
226 if (est == NULL)
227 return -ENOBUFS;
228
229 __gnet_stats_copy_basic(&b, cpu_bstats, bstats);
230
231 idx = parm->interval + 2;
232 est->bstats = bstats;
233 est->rate_est = rate_est;
234 est->stats_lock = stats_lock;
235 est->ewma_log = parm->ewma_log;
236 est->last_bytes = b.bytes;
237 est->avbps = rate_est->bps<<5;
238 est->last_packets = b.packets;
239 est->avpps = rate_est->pps<<10;
240 est->cpu_bstats = cpu_bstats;
241
242 spin_lock_bh(&est_tree_lock);
243 if (!elist[idx].timer.function) {
244 INIT_LIST_HEAD(&elist[idx].list);
245 setup_timer(&elist[idx].timer, est_timer, idx);
246 }
247
248 if (list_empty(&elist[idx].list))
249 mod_timer(&elist[idx].timer, jiffies + ((HZ/4) << idx));
250
251 list_add_rcu(&est->list, &elist[idx].list);
252 gen_add_node(est);
253 spin_unlock_bh(&est_tree_lock);
254
255 return 0;
256}
257EXPORT_SYMBOL(gen_new_estimator);
258
259/**
260 * gen_kill_estimator - remove a rate estimator
261 * @bstats: basic statistics
262 * @rate_est: rate estimator statistics
263 *
264 * Removes the rate estimator specified by &bstats and &rate_est.
265 *
266 * Note : Caller should respect an RCU grace period before freeing stats_lock
267 */
268void gen_kill_estimator(struct gnet_stats_basic_packed *bstats,
269 struct gnet_stats_rate_est64 *rate_est)
270{
271 struct gen_estimator *e;
272
273 spin_lock_bh(&est_tree_lock);
274 while ((e = gen_find_node(bstats, rate_est))) {
275 rb_erase(&e->node, &est_root);
276
277 write_lock(&est_lock);
278 e->bstats = NULL;
279 write_unlock(&est_lock);
280
281 list_del_rcu(&e->list);
282 kfree_rcu(e, e_rcu);
283 }
284 spin_unlock_bh(&est_tree_lock);
285}
286EXPORT_SYMBOL(gen_kill_estimator);
287
288/**
289 * gen_replace_estimator - replace rate estimator configuration
290 * @bstats: basic statistics
291 * @cpu_bstats: bstats per cpu
292 * @rate_est: rate estimator statistics
293 * @stats_lock: statistics lock
294 * @opt: rate estimator configuration TLV
295 *
296 * Replaces the configuration of a rate estimator by calling
297 * gen_kill_estimator() and gen_new_estimator().
298 *
299 * Returns 0 on success or a negative error code.
300 */
301int gen_replace_estimator(struct gnet_stats_basic_packed *bstats,
302 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
303 struct gnet_stats_rate_est64 *rate_est,
304 spinlock_t *stats_lock, struct nlattr *opt)
305{
306 gen_kill_estimator(bstats, rate_est);
307 return gen_new_estimator(bstats, cpu_bstats, rate_est, stats_lock, opt);
308}
309EXPORT_SYMBOL(gen_replace_estimator);
310
311/**
312 * gen_estimator_active - test if estimator is currently in use
313 * @bstats: basic statistics
314 * @rate_est: rate estimator statistics
315 *
316 * Returns true if estimator is active, and false if not.
317 */
318bool gen_estimator_active(const struct gnet_stats_basic_packed *bstats,
319 const struct gnet_stats_rate_est64 *rate_est)
320{
321 bool res;
322
323 ASSERT_RTNL();
324
325 spin_lock_bh(&est_tree_lock);
326 res = gen_find_node(bstats, rate_est) != NULL;
327 spin_unlock_bh(&est_tree_lock);
328
329 return res;
330}
331EXPORT_SYMBOL(gen_estimator_active);