blob: 332d8c58184d4964406dfe71179b877fd9f927f6 [file] [log] [blame]
Peter Zijlstra145ca252007-10-16 23:25:49 -07001/*
2 * Floating proportions
3 *
4 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
5 *
6 * Description:
7 *
8 * The floating proportion is a time derivative with an exponentially decaying
9 * history:
10 *
11 * p_{j} = \Sum_{i=0} (dx_{j}/dt_{-i}) / 2^(1+i)
12 *
13 * Where j is an element from {prop_local}, x_{j} is j's number of events,
14 * and i the time period over which the differential is taken. So d/dt_{-i} is
15 * the differential over the i-th last period.
16 *
17 * The decaying history gives smooth transitions. The time differential carries
18 * the notion of speed.
19 *
20 * The denominator is 2^(1+i) because we want the series to be normalised, ie.
21 *
22 * \Sum_{i=0} 1/2^(1+i) = 1
23 *
24 * Further more, if we measure time (t) in the same events as x; so that:
25 *
26 * t = \Sum_{j} x_{j}
27 *
28 * we get that:
29 *
30 * \Sum_{j} p_{j} = 1
31 *
32 * Writing this in an iterative fashion we get (dropping the 'd's):
33 *
34 * if (++x_{j}, ++t > period)
35 * t /= 2;
36 * for_each (j)
37 * x_{j} /= 2;
38 *
39 * so that:
40 *
41 * p_{j} = x_{j} / t;
42 *
43 * We optimize away the '/= 2' for the global time delta by noting that:
44 *
45 * if (++t > period) t /= 2:
46 *
47 * Can be approximated by:
48 *
49 * period/2 + (++t % period/2)
50 *
51 * [ Furthermore, when we choose period to be 2^n it can be written in terms of
52 * binary operations and wraparound artefacts disappear. ]
53 *
54 * Also note that this yields a natural counter of the elapsed periods:
55 *
56 * c = t / (period/2)
57 *
58 * [ Its monotonic increasing property can be applied to mitigate the wrap-
59 * around issue. ]
60 *
61 * This allows us to do away with the loop over all prop_locals on each period
62 * expiration. By remembering the period count under which it was last accessed
63 * as c_{j}, we can obtain the number of 'missed' cycles from:
64 *
65 * c - c_{j}
66 *
67 * We can then lazily catch up to the global period count every time we are
68 * going to use x_{j}, by doing:
69 *
70 * x_{j} /= 2^(c - c_{j}), c_{j} = c
71 */
72
73#include <linux/proportions.h>
74#include <linux/rcupdate.h>
75
76/*
77 * Limit the time part in order to ensure there are some bits left for the
78 * cycle counter.
79 */
80#define PROP_MAX_SHIFT (3*BITS_PER_LONG/4)
81
82int prop_descriptor_init(struct prop_descriptor *pd, int shift)
83{
84 int err;
85
86 if (shift > PROP_MAX_SHIFT)
87 shift = PROP_MAX_SHIFT;
88
89 pd->index = 0;
90 pd->pg[0].shift = shift;
91 mutex_init(&pd->mutex);
92 err = percpu_counter_init_irq(&pd->pg[0].events, 0);
93 if (err)
94 goto out;
95
96 err = percpu_counter_init_irq(&pd->pg[1].events, 0);
97 if (err)
98 percpu_counter_destroy(&pd->pg[0].events);
99
100out:
101 return err;
102}
103
104/*
105 * We have two copies, and flip between them to make it seem like an atomic
106 * update. The update is not really atomic wrt the events counter, but
107 * it is internally consistent with the bit layout depending on shift.
108 *
109 * We copy the events count, move the bits around and flip the index.
110 */
111void prop_change_shift(struct prop_descriptor *pd, int shift)
112{
113 int index;
114 int offset;
115 u64 events;
116 unsigned long flags;
117
118 if (shift > PROP_MAX_SHIFT)
119 shift = PROP_MAX_SHIFT;
120
121 mutex_lock(&pd->mutex);
122
123 index = pd->index ^ 1;
124 offset = pd->pg[pd->index].shift - shift;
125 if (!offset)
126 goto out;
127
128 pd->pg[index].shift = shift;
129
130 local_irq_save(flags);
131 events = percpu_counter_sum(&pd->pg[pd->index].events);
132 if (offset < 0)
133 events <<= -offset;
134 else
135 events >>= offset;
136 percpu_counter_set(&pd->pg[index].events, events);
137
138 /*
139 * ensure the new pg is fully written before the switch
140 */
141 smp_wmb();
142 pd->index = index;
143 local_irq_restore(flags);
144
145 synchronize_rcu();
146
147out:
148 mutex_unlock(&pd->mutex);
149}
150
151/*
152 * wrap the access to the data in an rcu_read_lock() section;
153 * this is used to track the active references.
154 */
155static struct prop_global *prop_get_global(struct prop_descriptor *pd)
156{
157 int index;
158
159 rcu_read_lock();
160 index = pd->index;
161 /*
162 * match the wmb from vcd_flip()
163 */
164 smp_rmb();
165 return &pd->pg[index];
166}
167
168static void prop_put_global(struct prop_descriptor *pd, struct prop_global *pg)
169{
170 rcu_read_unlock();
171}
172
173static void
174prop_adjust_shift(int *pl_shift, unsigned long *pl_period, int new_shift)
175{
176 int offset = *pl_shift - new_shift;
177
178 if (!offset)
179 return;
180
181 if (offset < 0)
182 *pl_period <<= -offset;
183 else
184 *pl_period >>= offset;
185
186 *pl_shift = new_shift;
187}
188
189/*
190 * PERCPU
191 */
192
193int prop_local_init_percpu(struct prop_local_percpu *pl)
194{
195 spin_lock_init(&pl->lock);
196 pl->shift = 0;
197 pl->period = 0;
198 return percpu_counter_init_irq(&pl->events, 0);
199}
200
201void prop_local_destroy_percpu(struct prop_local_percpu *pl)
202{
203 percpu_counter_destroy(&pl->events);
204}
205
206/*
207 * Catch up with missed period expirations.
208 *
209 * until (c_{j} == c)
210 * x_{j} -= x_{j}/2;
211 * c_{j}++;
212 */
213static
214void prop_norm_percpu(struct prop_global *pg, struct prop_local_percpu *pl)
215{
216 unsigned long period = 1UL << (pg->shift - 1);
217 unsigned long period_mask = ~(period - 1);
218 unsigned long global_period;
219 unsigned long flags;
220
221 global_period = percpu_counter_read(&pg->events);
222 global_period &= period_mask;
223
224 /*
225 * Fast path - check if the local and global period count still match
226 * outside of the lock.
227 */
228 if (pl->period == global_period)
229 return;
230
231 spin_lock_irqsave(&pl->lock, flags);
232 prop_adjust_shift(&pl->shift, &pl->period, pg->shift);
233 /*
234 * For each missed period, we half the local counter.
235 * basically:
236 * pl->events >> (global_period - pl->period);
237 *
238 * but since the distributed nature of percpu counters make division
239 * rather hard, use a regular subtraction loop. This is safe, because
240 * the events will only every be incremented, hence the subtraction
241 * can never result in a negative number.
242 */
243 while (pl->period != global_period) {
244 unsigned long val = percpu_counter_read(&pl->events);
245 unsigned long half = (val + 1) >> 1;
246
247 /*
248 * Half of zero won't be much less, break out.
249 * This limits the loop to shift iterations, even
250 * if we missed a million.
251 */
252 if (!val)
253 break;
254
255 percpu_counter_add(&pl->events, -half);
256 pl->period += period;
257 }
258 pl->period = global_period;
259 spin_unlock_irqrestore(&pl->lock, flags);
260}
261
262/*
263 * ++x_{j}, ++t
264 */
265void __prop_inc_percpu(struct prop_descriptor *pd, struct prop_local_percpu *pl)
266{
267 struct prop_global *pg = prop_get_global(pd);
268
269 prop_norm_percpu(pg, pl);
270 percpu_counter_add(&pl->events, 1);
271 percpu_counter_add(&pg->events, 1);
272 prop_put_global(pd, pg);
273}
274
275/*
276 * Obtain a fraction of this proportion
277 *
278 * p_{j} = x_{j} / (period/2 + t % period/2)
279 */
280void prop_fraction_percpu(struct prop_descriptor *pd,
281 struct prop_local_percpu *pl,
282 long *numerator, long *denominator)
283{
284 struct prop_global *pg = prop_get_global(pd);
285 unsigned long period_2 = 1UL << (pg->shift - 1);
286 unsigned long counter_mask = period_2 - 1;
287 unsigned long global_count;
288
289 prop_norm_percpu(pg, pl);
290 *numerator = percpu_counter_read_positive(&pl->events);
291
292 global_count = percpu_counter_read(&pg->events);
293 *denominator = period_2 + (global_count & counter_mask);
294
295 prop_put_global(pd, pg);
296}
297
298/*
299 * SINGLE
300 */
301
302int prop_local_init_single(struct prop_local_single *pl)
303{
304 spin_lock_init(&pl->lock);
305 pl->shift = 0;
306 pl->period = 0;
307 pl->events = 0;
308 return 0;
309}
310
311void prop_local_destroy_single(struct prop_local_single *pl)
312{
313}
314
315/*
316 * Catch up with missed period expirations.
317 */
318static
319void prop_norm_single(struct prop_global *pg, struct prop_local_single *pl)
320{
321 unsigned long period = 1UL << (pg->shift - 1);
322 unsigned long period_mask = ~(period - 1);
323 unsigned long global_period;
324 unsigned long flags;
325
326 global_period = percpu_counter_read(&pg->events);
327 global_period &= period_mask;
328
329 /*
330 * Fast path - check if the local and global period count still match
331 * outside of the lock.
332 */
333 if (pl->period == global_period)
334 return;
335
336 spin_lock_irqsave(&pl->lock, flags);
337 prop_adjust_shift(&pl->shift, &pl->period, pg->shift);
338 /*
339 * For each missed period, we half the local counter.
340 */
341 period = (global_period - pl->period) >> (pg->shift - 1);
342 if (likely(period < BITS_PER_LONG))
343 pl->events >>= period;
344 else
345 pl->events = 0;
346 pl->period = global_period;
347 spin_unlock_irqrestore(&pl->lock, flags);
348}
349
350/*
351 * ++x_{j}, ++t
352 */
353void __prop_inc_single(struct prop_descriptor *pd, struct prop_local_single *pl)
354{
355 struct prop_global *pg = prop_get_global(pd);
356
357 prop_norm_single(pg, pl);
358 pl->events++;
359 percpu_counter_add(&pg->events, 1);
360 prop_put_global(pd, pg);
361}
362
363/*
364 * Obtain a fraction of this proportion
365 *
366 * p_{j} = x_{j} / (period/2 + t % period/2)
367 */
368void prop_fraction_single(struct prop_descriptor *pd,
369 struct prop_local_single *pl,
370 long *numerator, long *denominator)
371{
372 struct prop_global *pg = prop_get_global(pd);
373 unsigned long period_2 = 1UL << (pg->shift - 1);
374 unsigned long counter_mask = period_2 - 1;
375 unsigned long global_count;
376
377 prop_norm_single(pg, pl);
378 *numerator = pl->events;
379
380 global_count = percpu_counter_read(&pg->events);
381 *denominator = period_2 + (global_count & counter_mask);
382
383 prop_put_global(pd, pg);
384}