2 * net/sched/gen_estimator.c Simple rate estimator.
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.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
12 * Jamal Hadi Salim - moved it to net/core and reshulfed
13 * names to make it usable in general net subsystem.
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>
24 #include <linux/socket.h>
25 #include <linux/sockios.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>
36 #include <net/gen_stats.h>
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.
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)
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.
56 We measure rate over A=(1<<interval) seconds and evaluate EWMA:
58 avrate = avrate*(1-W) + rate*W
60 where W is chosen as negative power of 2: W = 2^(-ewma_log)
62 The resulting time constant is:
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.
79 #define EST_MAX_INTERVAL 5
81 struct gen_estimator {
82 struct list_head list;
83 struct gnet_stats_basic_packed *bstats;
84 struct gnet_stats_rate_est64 *rate_est;
85 spinlock_t *stats_lock;
92 struct rcu_head e_rcu;
94 struct gnet_stats_basic_cpu __percpu *cpu_bstats;
98 struct gen_estimator_head {
99 unsigned long next_jiffies;
100 struct timer_list timer;
101 struct list_head list;
104 static struct gen_estimator_head elist[EST_MAX_INTERVAL+1];
106 /* Protects against NULL dereference */
107 static DEFINE_RWLOCK(est_lock);
109 /* Protects against soft lockup during large deletion */
110 static struct rb_root est_root = RB_ROOT;
111 static DEFINE_SPINLOCK(est_tree_lock);
113 static void est_timer(unsigned long arg)
116 struct gen_estimator *e;
119 list_for_each_entry_rcu(e, &elist[idx].list, list) {
120 struct gnet_stats_basic_packed b = {0};
125 spin_lock(e->stats_lock);
126 read_lock(&est_lock);
127 if (e->bstats == NULL)
130 __gnet_stats_copy_basic(e->running, &b, e->cpu_bstats, e->bstats);
132 brate = (b.bytes - e->last_bytes)<<(7 - idx);
133 e->last_bytes = b.bytes;
134 e->avbps += (brate >> e->ewma_log) - (e->avbps >> e->ewma_log);
135 WRITE_ONCE(e->rate_est->bps, (e->avbps + 0xF) >> 5);
137 rate = b.packets - e->last_packets;
139 e->last_packets = b.packets;
140 e->avpps += (rate >> e->ewma_log) - (e->avpps >> e->ewma_log);
141 WRITE_ONCE(e->rate_est->pps, (e->avpps + 0xF) >> 5);
143 read_unlock(&est_lock);
145 spin_unlock(e->stats_lock);
148 if (!list_empty(&elist[idx].list)) {
149 elist[idx].next_jiffies += ((HZ/4) << idx);
151 if (unlikely(time_after_eq(jiffies, elist[idx].next_jiffies))) {
152 /* Ouch... timer was delayed. */
153 elist[idx].next_jiffies = jiffies + 1;
155 mod_timer(&elist[idx].timer, elist[idx].next_jiffies);
160 static void gen_add_node(struct gen_estimator *est)
162 struct rb_node **p = &est_root.rb_node, *parent = NULL;
165 struct gen_estimator *e;
168 e = rb_entry(parent, struct gen_estimator, node);
170 if (est->bstats > e->bstats)
171 p = &parent->rb_right;
173 p = &parent->rb_left;
175 rb_link_node(&est->node, parent, p);
176 rb_insert_color(&est->node, &est_root);
180 struct gen_estimator *gen_find_node(const struct gnet_stats_basic_packed *bstats,
181 const struct gnet_stats_rate_est64 *rate_est)
183 struct rb_node *p = est_root.rb_node;
186 struct gen_estimator *e;
188 e = rb_entry(p, struct gen_estimator, node);
190 if (bstats > e->bstats)
192 else if (bstats < e->bstats || rate_est != e->rate_est)
201 * gen_new_estimator - create a new rate estimator
202 * @bstats: basic statistics
203 * @cpu_bstats: bstats per cpu
204 * @rate_est: rate estimator statistics
205 * @stats_lock: statistics lock
206 * @running: qdisc running seqcount
207 * @opt: rate estimator configuration TLV
209 * Creates a new rate estimator with &bstats as source and &rate_est
210 * as destination. A new timer with the interval specified in the
211 * configuration TLV is created. Upon each interval, the latest statistics
212 * will be read from &bstats and the estimated rate will be stored in
213 * &rate_est with the statistics lock grabbed during this period.
215 * Returns 0 on success or a negative error code.
218 int gen_new_estimator(struct gnet_stats_basic_packed *bstats,
219 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
220 struct gnet_stats_rate_est64 *rate_est,
221 spinlock_t *stats_lock,
225 struct gen_estimator *est;
226 struct gnet_estimator *parm = nla_data(opt);
227 struct gnet_stats_basic_packed b = {0};
230 if (nla_len(opt) < sizeof(*parm))
233 if (parm->interval < -2 || parm->interval > 3)
236 est = kzalloc(sizeof(*est), GFP_KERNEL);
240 __gnet_stats_copy_basic(running, &b, cpu_bstats, bstats);
242 idx = parm->interval + 2;
243 est->bstats = bstats;
244 est->rate_est = rate_est;
245 est->stats_lock = stats_lock;
246 est->running = running;
247 est->ewma_log = parm->ewma_log;
248 est->last_bytes = b.bytes;
249 est->avbps = rate_est->bps<<5;
250 est->last_packets = b.packets;
251 est->avpps = rate_est->pps<<10;
252 est->cpu_bstats = cpu_bstats;
254 spin_lock_bh(&est_tree_lock);
255 if (!elist[idx].timer.function) {
256 INIT_LIST_HEAD(&elist[idx].list);
257 setup_timer(&elist[idx].timer, est_timer, idx);
260 if (list_empty(&elist[idx].list)) {
261 elist[idx].next_jiffies = jiffies + ((HZ/4) << idx);
262 mod_timer(&elist[idx].timer, elist[idx].next_jiffies);
264 list_add_rcu(&est->list, &elist[idx].list);
266 spin_unlock_bh(&est_tree_lock);
270 EXPORT_SYMBOL(gen_new_estimator);
273 * gen_kill_estimator - remove a rate estimator
274 * @bstats: basic statistics
275 * @rate_est: rate estimator statistics
277 * Removes the rate estimator specified by &bstats and &rate_est.
279 * Note : Caller should respect an RCU grace period before freeing stats_lock
281 void gen_kill_estimator(struct gnet_stats_basic_packed *bstats,
282 struct gnet_stats_rate_est64 *rate_est)
284 struct gen_estimator *e;
286 spin_lock_bh(&est_tree_lock);
287 while ((e = gen_find_node(bstats, rate_est))) {
288 rb_erase(&e->node, &est_root);
290 write_lock(&est_lock);
292 write_unlock(&est_lock);
294 list_del_rcu(&e->list);
297 spin_unlock_bh(&est_tree_lock);
299 EXPORT_SYMBOL(gen_kill_estimator);
302 * gen_replace_estimator - replace rate estimator configuration
303 * @bstats: basic statistics
304 * @cpu_bstats: bstats per cpu
305 * @rate_est: rate estimator statistics
306 * @stats_lock: statistics lock
307 * @running: qdisc running seqcount (might be NULL)
308 * @opt: rate estimator configuration TLV
310 * Replaces the configuration of a rate estimator by calling
311 * gen_kill_estimator() and gen_new_estimator().
313 * Returns 0 on success or a negative error code.
315 int gen_replace_estimator(struct gnet_stats_basic_packed *bstats,
316 struct gnet_stats_basic_cpu __percpu *cpu_bstats,
317 struct gnet_stats_rate_est64 *rate_est,
318 spinlock_t *stats_lock,
319 seqcount_t *running, struct nlattr *opt)
321 gen_kill_estimator(bstats, rate_est);
322 return gen_new_estimator(bstats, cpu_bstats, rate_est, stats_lock, running, opt);
324 EXPORT_SYMBOL(gen_replace_estimator);
327 * gen_estimator_active - test if estimator is currently in use
328 * @bstats: basic statistics
329 * @rate_est: rate estimator statistics
331 * Returns true if estimator is active, and false if not.
333 bool gen_estimator_active(const struct gnet_stats_basic_packed *bstats,
334 const struct gnet_stats_rate_est64 *rate_est)
340 spin_lock_bh(&est_tree_lock);
341 res = gen_find_node(bstats, rate_est) != NULL;
342 spin_unlock_bh(&est_tree_lock);
346 EXPORT_SYMBOL(gen_estimator_active);