xref: /linux/lib/flex_proportions.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /*
2  *  Floating proportions with flexible aging period
3  *
4  *   Copyright (C) 2011, SUSE, Jan Kara <jack@suse.cz>
5  *
6  * The goal of this code is: Given different types of event, measure proportion
7  * of each type of event over time. The proportions are measured with
8  * exponentially decaying history to give smooth transitions. A formula
9  * expressing proportion of event of type 'j' is:
10  *
11  *   p_{j} = (\Sum_{i>=0} x_{i,j}/2^{i+1})/(\Sum_{i>=0} x_i/2^{i+1})
12  *
13  * Where x_{i,j} is j's number of events in i-th last time period and x_i is
14  * total number of events in i-th last time period.
15  *
16  * Note that p_{j}'s are normalised, i.e.
17  *
18  *   \Sum_{j} p_{j} = 1,
19  *
20  * This formula can be straightforwardly computed by maintaining denominator
21  * (let's call it 'd') and for each event type its numerator (let's call it
22  * 'n_j'). When an event of type 'j' happens, we simply need to do:
23  *   n_j++; d++;
24  *
25  * When a new period is declared, we could do:
26  *   d /= 2
27  *   for each j
28  *     n_j /= 2
29  *
30  * To avoid iteration over all event types, we instead shift numerator of event
31  * j lazily when someone asks for a proportion of event j or when event j
32  * occurs. This can bit trivially implemented by remembering last period in
33  * which something happened with proportion of type j.
34  */
35 #include <linux/flex_proportions.h>
36 
37 int fprop_global_init(struct fprop_global *p, gfp_t gfp)
38 {
39 	int err;
40 
41 	p->period = 0;
42 	/* Use 1 to avoid dealing with periods with 0 events... */
43 	err = percpu_counter_init(&p->events, 1, gfp);
44 	if (err)
45 		return err;
46 	seqcount_init(&p->sequence);
47 	return 0;
48 }
49 
50 void fprop_global_destroy(struct fprop_global *p)
51 {
52 	percpu_counter_destroy(&p->events);
53 }
54 
55 /*
56  * Declare @periods new periods. It is upto the caller to make sure period
57  * transitions cannot happen in parallel.
58  *
59  * The function returns true if the proportions are still defined and false
60  * if aging zeroed out all events. This can be used to detect whether declaring
61  * further periods has any effect.
62  */
63 bool fprop_new_period(struct fprop_global *p, int periods)
64 {
65 	s64 events;
66 	unsigned long flags;
67 
68 	local_irq_save(flags);
69 	events = percpu_counter_sum(&p->events);
70 	/*
71 	 * Don't do anything if there are no events.
72 	 */
73 	if (events <= 1) {
74 		local_irq_restore(flags);
75 		return false;
76 	}
77 	write_seqcount_begin(&p->sequence);
78 	if (periods < 64)
79 		events -= events >> periods;
80 	/* Use addition to avoid losing events happening between sum and set */
81 	percpu_counter_add(&p->events, -events);
82 	p->period += periods;
83 	write_seqcount_end(&p->sequence);
84 	local_irq_restore(flags);
85 
86 	return true;
87 }
88 
89 /*
90  * ---- SINGLE ----
91  */
92 
93 int fprop_local_init_single(struct fprop_local_single *pl)
94 {
95 	pl->events = 0;
96 	pl->period = 0;
97 	raw_spin_lock_init(&pl->lock);
98 	return 0;
99 }
100 
101 void fprop_local_destroy_single(struct fprop_local_single *pl)
102 {
103 }
104 
105 static void fprop_reflect_period_single(struct fprop_global *p,
106 					struct fprop_local_single *pl)
107 {
108 	unsigned int period = p->period;
109 	unsigned long flags;
110 
111 	/* Fast path - period didn't change */
112 	if (pl->period == period)
113 		return;
114 	raw_spin_lock_irqsave(&pl->lock, flags);
115 	/* Someone updated pl->period while we were spinning? */
116 	if (pl->period >= period) {
117 		raw_spin_unlock_irqrestore(&pl->lock, flags);
118 		return;
119 	}
120 	/* Aging zeroed our fraction? */
121 	if (period - pl->period < BITS_PER_LONG)
122 		pl->events >>= period - pl->period;
123 	else
124 		pl->events = 0;
125 	pl->period = period;
126 	raw_spin_unlock_irqrestore(&pl->lock, flags);
127 }
128 
129 /* Event of type pl happened */
130 void __fprop_inc_single(struct fprop_global *p, struct fprop_local_single *pl)
131 {
132 	fprop_reflect_period_single(p, pl);
133 	pl->events++;
134 	percpu_counter_add(&p->events, 1);
135 }
136 
137 /* Return fraction of events of type pl */
138 void fprop_fraction_single(struct fprop_global *p,
139 			   struct fprop_local_single *pl,
140 			   unsigned long *numerator, unsigned long *denominator)
141 {
142 	unsigned int seq;
143 	s64 num, den;
144 
145 	do {
146 		seq = read_seqcount_begin(&p->sequence);
147 		fprop_reflect_period_single(p, pl);
148 		num = pl->events;
149 		den = percpu_counter_read_positive(&p->events);
150 	} while (read_seqcount_retry(&p->sequence, seq));
151 
152 	/*
153 	 * Make fraction <= 1 and denominator > 0 even in presence of percpu
154 	 * counter errors
155 	 */
156 	if (den <= num) {
157 		if (num)
158 			den = num;
159 		else
160 			den = 1;
161 	}
162 	*denominator = den;
163 	*numerator = num;
164 }
165 
166 /*
167  * ---- PERCPU ----
168  */
169 #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids)))
170 
171 int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp)
172 {
173 	int err;
174 
175 	err = percpu_counter_init(&pl->events, 0, gfp);
176 	if (err)
177 		return err;
178 	pl->period = 0;
179 	raw_spin_lock_init(&pl->lock);
180 	return 0;
181 }
182 
183 void fprop_local_destroy_percpu(struct fprop_local_percpu *pl)
184 {
185 	percpu_counter_destroy(&pl->events);
186 }
187 
188 static void fprop_reflect_period_percpu(struct fprop_global *p,
189 					struct fprop_local_percpu *pl)
190 {
191 	unsigned int period = p->period;
192 	unsigned long flags;
193 
194 	/* Fast path - period didn't change */
195 	if (pl->period == period)
196 		return;
197 	raw_spin_lock_irqsave(&pl->lock, flags);
198 	/* Someone updated pl->period while we were spinning? */
199 	if (pl->period >= period) {
200 		raw_spin_unlock_irqrestore(&pl->lock, flags);
201 		return;
202 	}
203 	/* Aging zeroed our fraction? */
204 	if (period - pl->period < BITS_PER_LONG) {
205 		s64 val = percpu_counter_read(&pl->events);
206 
207 		if (val < (nr_cpu_ids * PROP_BATCH))
208 			val = percpu_counter_sum(&pl->events);
209 
210 		__percpu_counter_add(&pl->events,
211 			-val + (val >> (period-pl->period)), PROP_BATCH);
212 	} else
213 		percpu_counter_set(&pl->events, 0);
214 	pl->period = period;
215 	raw_spin_unlock_irqrestore(&pl->lock, flags);
216 }
217 
218 /* Event of type pl happened */
219 void __fprop_inc_percpu(struct fprop_global *p, struct fprop_local_percpu *pl)
220 {
221 	fprop_reflect_period_percpu(p, pl);
222 	__percpu_counter_add(&pl->events, 1, PROP_BATCH);
223 	percpu_counter_add(&p->events, 1);
224 }
225 
226 void fprop_fraction_percpu(struct fprop_global *p,
227 			   struct fprop_local_percpu *pl,
228 			   unsigned long *numerator, unsigned long *denominator)
229 {
230 	unsigned int seq;
231 	s64 num, den;
232 
233 	do {
234 		seq = read_seqcount_begin(&p->sequence);
235 		fprop_reflect_period_percpu(p, pl);
236 		num = percpu_counter_read_positive(&pl->events);
237 		den = percpu_counter_read_positive(&p->events);
238 	} while (read_seqcount_retry(&p->sequence, seq));
239 
240 	/*
241 	 * Make fraction <= 1 and denominator > 0 even in presence of percpu
242 	 * counter errors
243 	 */
244 	if (den <= num) {
245 		if (num)
246 			den = num;
247 		else
248 			den = 1;
249 	}
250 	*denominator = den;
251 	*numerator = num;
252 }
253 
254 /*
255  * Like __fprop_inc_percpu() except that event is counted only if the given
256  * type has fraction smaller than @max_frac/FPROP_FRAC_BASE
257  */
258 void __fprop_inc_percpu_max(struct fprop_global *p,
259 			    struct fprop_local_percpu *pl, int max_frac)
260 {
261 	if (unlikely(max_frac < FPROP_FRAC_BASE)) {
262 		unsigned long numerator, denominator;
263 
264 		fprop_fraction_percpu(p, pl, &numerator, &denominator);
265 		if (numerator >
266 		    (((u64)denominator) * max_frac) >> FPROP_FRAC_SHIFT)
267 			return;
268 	} else
269 		fprop_reflect_period_percpu(p, pl);
270 	__percpu_counter_add(&pl->events, 1, PROP_BATCH);
271 	percpu_counter_add(&p->events, 1);
272 }
273