io_uring: don't assume mm is constant across submits
[linux-2.6-microblaze.git] / lib / flex_proportions.c
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(&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;
67         unsigned long flags;
68
69         local_irq_save(flags);
70         events = percpu_counter_sum(&p->events);
71         /*
72          * Don't do anything if there are no events.
73          */
74         if (events <= 1) {
75                 local_irq_restore(flags);
76                 return false;
77         }
78         write_seqcount_begin(&p->sequence);
79         if (periods < 64)
80                 events -= events >> periods;
81         /* Use addition to avoid losing events happening between sum and set */
82         percpu_counter_add(&p->events, -events);
83         p->period += periods;
84         write_seqcount_end(&p->sequence);
85         local_irq_restore(flags);
86
87         return true;
88 }
89
90 /*
91  * ---- SINGLE ----
92  */
93
94 int fprop_local_init_single(struct fprop_local_single *pl)
95 {
96         pl->events = 0;
97         pl->period = 0;
98         raw_spin_lock_init(&pl->lock);
99         return 0;
100 }
101
102 void fprop_local_destroy_single(struct fprop_local_single *pl)
103 {
104 }
105
106 static void fprop_reflect_period_single(struct fprop_global *p,
107                                         struct fprop_local_single *pl)
108 {
109         unsigned int period = p->period;
110         unsigned long flags;
111
112         /* Fast path - period didn't change */
113         if (pl->period == period)
114                 return;
115         raw_spin_lock_irqsave(&pl->lock, flags);
116         /* Someone updated pl->period while we were spinning? */
117         if (pl->period >= period) {
118                 raw_spin_unlock_irqrestore(&pl->lock, flags);
119                 return;
120         }
121         /* Aging zeroed our fraction? */
122         if (period - pl->period < BITS_PER_LONG)
123                 pl->events >>= period - pl->period;
124         else
125                 pl->events = 0;
126         pl->period = period;
127         raw_spin_unlock_irqrestore(&pl->lock, flags);
128 }
129
130 /* Event of type pl happened */
131 void __fprop_inc_single(struct fprop_global *p, struct fprop_local_single *pl)
132 {
133         fprop_reflect_period_single(p, pl);
134         pl->events++;
135         percpu_counter_add(&p->events, 1);
136 }
137
138 /* Return fraction of events of type pl */
139 void fprop_fraction_single(struct fprop_global *p,
140                            struct fprop_local_single *pl,
141                            unsigned long *numerator, unsigned long *denominator)
142 {
143         unsigned int seq;
144         s64 num, den;
145
146         do {
147                 seq = read_seqcount_begin(&p->sequence);
148                 fprop_reflect_period_single(p, pl);
149                 num = pl->events;
150                 den = percpu_counter_read_positive(&p->events);
151         } while (read_seqcount_retry(&p->sequence, seq));
152
153         /*
154          * Make fraction <= 1 and denominator > 0 even in presence of percpu
155          * counter errors
156          */
157         if (den <= num) {
158                 if (num)
159                         den = num;
160                 else
161                         den = 1;
162         }
163         *denominator = den;
164         *numerator = num;
165 }
166
167 /*
168  * ---- PERCPU ----
169  */
170 #define PROP_BATCH (8*(1+ilog2(nr_cpu_ids)))
171
172 int fprop_local_init_percpu(struct fprop_local_percpu *pl, gfp_t gfp)
173 {
174         int err;
175
176         err = percpu_counter_init(&pl->events, 0, gfp);
177         if (err)
178                 return err;
179         pl->period = 0;
180         raw_spin_lock_init(&pl->lock);
181         return 0;
182 }
183
184 void fprop_local_destroy_percpu(struct fprop_local_percpu *pl)
185 {
186         percpu_counter_destroy(&pl->events);
187 }
188
189 static void fprop_reflect_period_percpu(struct fprop_global *p,
190                                         struct fprop_local_percpu *pl)
191 {
192         unsigned int period = p->period;
193         unsigned long flags;
194
195         /* Fast path - period didn't change */
196         if (pl->period == period)
197                 return;
198         raw_spin_lock_irqsave(&pl->lock, flags);
199         /* Someone updated pl->period while we were spinning? */
200         if (pl->period >= period) {
201                 raw_spin_unlock_irqrestore(&pl->lock, flags);
202                 return;
203         }
204         /* Aging zeroed our fraction? */
205         if (period - pl->period < BITS_PER_LONG) {
206                 s64 val = percpu_counter_read(&pl->events);
207
208                 if (val < (nr_cpu_ids * PROP_BATCH))
209                         val = percpu_counter_sum(&pl->events);
210
211                 percpu_counter_add_batch(&pl->events,
212                         -val + (val >> (period-pl->period)), PROP_BATCH);
213         } else
214                 percpu_counter_set(&pl->events, 0);
215         pl->period = period;
216         raw_spin_unlock_irqrestore(&pl->lock, flags);
217 }
218
219 /* Event of type pl happened */
220 void __fprop_inc_percpu(struct fprop_global *p, struct fprop_local_percpu *pl)
221 {
222         fprop_reflect_period_percpu(p, pl);
223         percpu_counter_add_batch(&pl->events, 1, PROP_BATCH);
224         percpu_counter_add(&p->events, 1);
225 }
226
227 void fprop_fraction_percpu(struct fprop_global *p,
228                            struct fprop_local_percpu *pl,
229                            unsigned long *numerator, unsigned long *denominator)
230 {
231         unsigned int seq;
232         s64 num, den;
233
234         do {
235                 seq = read_seqcount_begin(&p->sequence);
236                 fprop_reflect_period_percpu(p, pl);
237                 num = percpu_counter_read_positive(&pl->events);
238                 den = percpu_counter_read_positive(&p->events);
239         } while (read_seqcount_retry(&p->sequence, seq));
240
241         /*
242          * Make fraction <= 1 and denominator > 0 even in presence of percpu
243          * counter errors
244          */
245         if (den <= num) {
246                 if (num)
247                         den = num;
248                 else
249                         den = 1;
250         }
251         *denominator = den;
252         *numerator = num;
253 }
254
255 /*
256  * Like __fprop_inc_percpu() except that event is counted only if the given
257  * type has fraction smaller than @max_frac/FPROP_FRAC_BASE
258  */
259 void __fprop_inc_percpu_max(struct fprop_global *p,
260                             struct fprop_local_percpu *pl, int max_frac)
261 {
262         if (unlikely(max_frac < FPROP_FRAC_BASE)) {
263                 unsigned long numerator, denominator;
264
265                 fprop_fraction_percpu(p, pl, &numerator, &denominator);
266                 if (numerator >
267                     (((u64)denominator) * max_frac) >> FPROP_FRAC_SHIFT)
268                         return;
269         }
270
271         __fprop_inc_percpu(p, pl);
272 }