1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Floating proportions with flexible aging period |
4 | * |
5 | * Copyright (C) 2011, SUSE, Jan Kara <jack@suse.cz> |
6 | * |
7 | * The goal of this code is: Given different types of event, measure proportion |
8 | * of each type of event over time. The proportions are measured with |
9 | * exponentially decaying history to give smooth transitions. A formula |
10 | * expressing proportion of event of type 'j' is: |
11 | * |
12 | * p_{j} = (\Sum_{i>=0} x_{i,j}/2^{i+1})/(\Sum_{i>=0} x_i/2^{i+1}) |
13 | * |
14 | * Where x_{i,j} is j's number of events in i-th last time period and x_i is |
15 | * total number of events in i-th last time period. |
16 | * |
17 | * Note that p_{j}'s are normalised, i.e. |
18 | * |
19 | * \Sum_{j} p_{j} = 1, |
20 | * |
21 | * This formula can be straightforwardly computed by maintaining denominator |
22 | * (let's call it 'd') and for each event type its numerator (let's call it |
23 | * 'n_j'). When an event of type 'j' happens, we simply need to do: |
24 | * n_j++; d++; |
25 | * |
26 | * When a new period is declared, we could do: |
27 | * d /= 2 |
28 | * for each j |
29 | * n_j /= 2 |
30 | * |
31 | * To avoid iteration over all event types, we instead shift numerator of event |
32 | * j lazily when someone asks for a proportion of event j or when event j |
33 | * occurs. This can bit trivially implemented by remembering last period in |
34 | * which something happened with proportion of type j. |
35 | */ |
36 | #include <linux/flex_proportions.h> |
37 | |
38 | int fprop_global_init(struct fprop_global *p, gfp_t gfp) |
39 | { |
40 | int err; |
41 | |
42 | p->period = 0; |
43 | /* Use 1 to avoid dealing with periods with 0 events... */ |
44 | err = percpu_counter_init(&p->events, 1, gfp); |
45 | if (err) |
46 | return err; |
47 | seqcount_init(&p->sequence); |
48 | return 0; |
49 | } |
50 | |
51 | void fprop_global_destroy(struct fprop_global *p) |
52 | { |
53 | percpu_counter_destroy(fbc: &p->events); |
54 | } |
55 | |
56 | /* |
57 | * Declare @periods new periods. It is upto the caller to make sure period |
58 | * transitions cannot happen in parallel. |
59 | * |
60 | * The function returns true if the proportions are still defined and false |
61 | * if aging zeroed out all events. This can be used to detect whether declaring |
62 | * further periods has any effect. |
63 | */ |
64 | bool fprop_new_period(struct fprop_global *p, int periods) |
65 | { |
66 | s64 events = percpu_counter_sum(fbc: &p->events); |
67 | |
68 | /* |
69 | * Don't do anything if there are no events. |
70 | */ |
71 | if (events <= 1) |
72 | return false; |
73 | preempt_disable_nested(); |
74 | write_seqcount_begin(&p->sequence); |
75 | if (periods < 64) |
76 | events -= events >> periods; |
77 | /* Use addition to avoid losing events happening between sum and set */ |
78 | percpu_counter_add(fbc: &p->events, amount: -events); |
79 | p->period += periods; |
80 | write_seqcount_end(&p->sequence); |
81 | preempt_enable_nested(); |
82 | |
83 | return true; |
84 | } |
85 | |
86 | /* |
87 | * ---- SINGLE ---- |
88 | */ |
89 | |
90 | int fprop_local_init_single(struct fprop_local_single *pl) |
91 | { |
92 | pl->events = 0; |
93 | pl->period = 0; |
94 | raw_spin_lock_init(&pl->lock); |
95 | return 0; |
96 | } |
97 | |
98 | void fprop_local_destroy_single(struct fprop_local_single *pl) |
99 | { |
100 | } |
101 | |
102 | static void fprop_reflect_period_single(struct fprop_global *p, |
103 | struct fprop_local_single *pl) |
104 | { |
105 | unsigned int period = p->period; |
106 | unsigned long flags; |
107 | |
108 | /* Fast path - period didn't change */ |
109 | if (pl->period == period) |
110 | return; |
111 | raw_spin_lock_irqsave(&pl->lock, flags); |
112 | /* Someone updated pl->period while we were spinning? */ |
113 | if (pl->period >= period) { |
114 | raw_spin_unlock_irqrestore(&pl->lock, flags); |
115 | return; |
116 | } |
117 | /* Aging zeroed our fraction? */ |
118 | if (period - pl->period < BITS_PER_LONG) |
119 | pl->events >>= period - pl->period; |
120 | else |
121 | pl->events = 0; |
122 | pl->period = period; |
123 | raw_spin_unlock_irqrestore(&pl->lock, flags); |
124 | } |
125 | |
126 | /* Event of type pl happened */ |
127 | void __fprop_inc_single(struct fprop_global *p, struct fprop_local_single *pl) |
128 | { |
129 | fprop_reflect_period_single(p, pl); |
130 | pl->events++; |
131 | percpu_counter_add(fbc: &p->events, amount: 1); |
132 | } |
133 | |
134 | /* Return fraction of events of type pl */ |
135 | void fprop_fraction_single(struct fprop_global *p, |
136 | struct fprop_local_single *pl, |
137 | unsigned long *numerator, unsigned long *denominator) |
138 | { |
139 | unsigned int seq; |
140 | s64 num, den; |
141 | |
142 | do { |
143 | seq = read_seqcount_begin(&p->sequence); |
144 | fprop_reflect_period_single(p, pl); |
145 | num = pl->events; |
146 | den = percpu_counter_read_positive(fbc: &p->events); |
147 | } while (read_seqcount_retry(&p->sequence, seq)); |
148 | |
149 | /* |
150 | * Make fraction <= 1 and denominator > 0 even in presence of percpu |
151 | * counter errors |
152 | */ |
153 | if (den <= num) { |
154 | if (num) |
155 | den = num; |
156 | else |
157 | den = 1; |
158 | } |
159 | *denominator = den; |
160 | *numerator = num; |
161 | } |
162 | |
163 | /* |
164 | * ---- PERCPU ---- |
165 | */ |
166 | #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids))) |
167 | |
168 | int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp) |
169 | { |
170 | int err; |
171 | |
172 | err = percpu_counter_init(&pl->events, 0, gfp); |
173 | if (err) |
174 | return err; |
175 | pl->period = 0; |
176 | raw_spin_lock_init(&pl->lock); |
177 | return 0; |
178 | } |
179 | |
180 | void fprop_local_destroy_percpu(struct fprop_local_percpu *pl) |
181 | { |
182 | percpu_counter_destroy(fbc: &pl->events); |
183 | } |
184 | |
185 | static void fprop_reflect_period_percpu(struct fprop_global *p, |
186 | struct fprop_local_percpu *pl) |
187 | { |
188 | unsigned int period = p->period; |
189 | unsigned long flags; |
190 | |
191 | /* Fast path - period didn't change */ |
192 | if (pl->period == period) |
193 | return; |
194 | raw_spin_lock_irqsave(&pl->lock, flags); |
195 | /* Someone updated pl->period while we were spinning? */ |
196 | if (pl->period >= period) { |
197 | raw_spin_unlock_irqrestore(&pl->lock, flags); |
198 | return; |
199 | } |
200 | /* Aging zeroed our fraction? */ |
201 | if (period - pl->period < BITS_PER_LONG) { |
202 | s64 val = percpu_counter_read(fbc: &pl->events); |
203 | |
204 | if (val < (nr_cpu_ids * PROP_BATCH)) |
205 | val = percpu_counter_sum(fbc: &pl->events); |
206 | |
207 | percpu_counter_add_batch(fbc: &pl->events, |
208 | amount: -val + (val >> (period-pl->period)), PROP_BATCH); |
209 | } else |
210 | percpu_counter_set(fbc: &pl->events, amount: 0); |
211 | pl->period = period; |
212 | raw_spin_unlock_irqrestore(&pl->lock, flags); |
213 | } |
214 | |
215 | /* Event of type pl happened */ |
216 | void __fprop_add_percpu(struct fprop_global *p, struct fprop_local_percpu *pl, |
217 | long nr) |
218 | { |
219 | fprop_reflect_period_percpu(p, pl); |
220 | percpu_counter_add_batch(fbc: &pl->events, amount: nr, PROP_BATCH); |
221 | percpu_counter_add(fbc: &p->events, amount: nr); |
222 | } |
223 | |
224 | void fprop_fraction_percpu(struct fprop_global *p, |
225 | struct fprop_local_percpu *pl, |
226 | unsigned long *numerator, unsigned long *denominator) |
227 | { |
228 | unsigned int seq; |
229 | s64 num, den; |
230 | |
231 | do { |
232 | seq = read_seqcount_begin(&p->sequence); |
233 | fprop_reflect_period_percpu(p, pl); |
234 | num = percpu_counter_read_positive(fbc: &pl->events); |
235 | den = percpu_counter_read_positive(fbc: &p->events); |
236 | } while (read_seqcount_retry(&p->sequence, seq)); |
237 | |
238 | /* |
239 | * Make fraction <= 1 and denominator > 0 even in presence of percpu |
240 | * counter errors |
241 | */ |
242 | if (den <= num) { |
243 | if (num) |
244 | den = num; |
245 | else |
246 | den = 1; |
247 | } |
248 | *denominator = den; |
249 | *numerator = num; |
250 | } |
251 | |
252 | /* |
253 | * Like __fprop_add_percpu() except that event is counted only if the given |
254 | * type has fraction smaller than @max_frac/FPROP_FRAC_BASE |
255 | */ |
256 | void __fprop_add_percpu_max(struct fprop_global *p, |
257 | struct fprop_local_percpu *pl, int max_frac, long nr) |
258 | { |
259 | if (unlikely(max_frac < FPROP_FRAC_BASE)) { |
260 | unsigned long numerator, denominator; |
261 | s64 tmp; |
262 | |
263 | fprop_fraction_percpu(p, pl, numerator: &numerator, denominator: &denominator); |
264 | /* Adding 'nr' to fraction exceeds max_frac/FPROP_FRAC_BASE? */ |
265 | tmp = (u64)denominator * max_frac - |
266 | ((u64)numerator << FPROP_FRAC_SHIFT); |
267 | if (tmp < 0) { |
268 | /* Maximum fraction already exceeded? */ |
269 | return; |
270 | } else if (tmp < nr * (FPROP_FRAC_BASE - max_frac)) { |
271 | /* Add just enough for the fraction to saturate */ |
272 | nr = div_u64(dividend: tmp + FPROP_FRAC_BASE - max_frac - 1, |
273 | FPROP_FRAC_BASE - max_frac); |
274 | } |
275 | } |
276 | |
277 | __fprop_add_percpu(p, pl, nr); |
278 | } |
279 | |