xref: /openbmc/linux/kernel/irq/timings.c (revision ddc141e5)
1 /*
2  * linux/kernel/irq/timings.c
3  *
4  * Copyright (C) 2016, Linaro Ltd - Daniel Lezcano <daniel.lezcano@linaro.org>
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  *
10  */
11 #include <linux/kernel.h>
12 #include <linux/percpu.h>
13 #include <linux/slab.h>
14 #include <linux/static_key.h>
15 #include <linux/interrupt.h>
16 #include <linux/idr.h>
17 #include <linux/irq.h>
18 #include <linux/math64.h>
19 
20 #include <trace/events/irq.h>
21 
22 #include "internals.h"
23 
24 DEFINE_STATIC_KEY_FALSE(irq_timing_enabled);
25 
26 DEFINE_PER_CPU(struct irq_timings, irq_timings);
27 
28 struct irqt_stat {
29 	u64	next_evt;
30 	u64	last_ts;
31 	u64	variance;
32 	u32	avg;
33 	u32	nr_samples;
34 	int	anomalies;
35 	int	valid;
36 };
37 
38 static DEFINE_IDR(irqt_stats);
39 
40 void irq_timings_enable(void)
41 {
42 	static_branch_enable(&irq_timing_enabled);
43 }
44 
45 void irq_timings_disable(void)
46 {
47 	static_branch_disable(&irq_timing_enabled);
48 }
49 
50 /**
51  * irqs_update - update the irq timing statistics with a new timestamp
52  *
53  * @irqs: an irqt_stat struct pointer
54  * @ts: the new timestamp
55  *
56  * The statistics are computed online, in other words, the code is
57  * designed to compute the statistics on a stream of values rather
58  * than doing multiple passes on the values to compute the average,
59  * then the variance. The integer division introduces a loss of
60  * precision but with an acceptable error margin regarding the results
61  * we would have with the double floating precision: we are dealing
62  * with nanosec, so big numbers, consequently the mantisse is
63  * negligeable, especially when converting the time in usec
64  * afterwards.
65  *
66  * The computation happens at idle time. When the CPU is not idle, the
67  * interrupts' timestamps are stored in the circular buffer, when the
68  * CPU goes idle and this routine is called, all the buffer's values
69  * are injected in the statistical model continuying to extend the
70  * statistics from the previous busy-idle cycle.
71  *
72  * The observations showed a device will trigger a burst of periodic
73  * interrupts followed by one or two peaks of longer time, for
74  * instance when a SD card device flushes its cache, then the periodic
75  * intervals occur again. A one second inactivity period resets the
76  * stats, that gives us the certitude the statistical values won't
77  * exceed 1x10^9, thus the computation won't overflow.
78  *
79  * Basically, the purpose of the algorithm is to watch the periodic
80  * interrupts and eliminate the peaks.
81  *
82  * An interrupt is considered periodically stable if the interval of
83  * its occurences follow the normal distribution, thus the values
84  * comply with:
85  *
86  *      avg - 3 x stddev < value < avg + 3 x stddev
87  *
88  * Which can be simplified to:
89  *
90  *      -3 x stddev < value - avg < 3 x stddev
91  *
92  *      abs(value - avg) < 3 x stddev
93  *
94  * In order to save a costly square root computation, we use the
95  * variance. For the record, stddev = sqrt(variance). The equation
96  * above becomes:
97  *
98  *      abs(value - avg) < 3 x sqrt(variance)
99  *
100  * And finally we square it:
101  *
102  *      (value - avg) ^ 2 < (3 x sqrt(variance)) ^ 2
103  *
104  *      (value - avg) x (value - avg) < 9 x variance
105  *
106  * Statistically speaking, any values out of this interval is
107  * considered as an anomaly and is discarded. However, a normal
108  * distribution appears when the number of samples is 30 (it is the
109  * rule of thumb in statistics, cf. "30 samples" on Internet). When
110  * there are three consecutive anomalies, the statistics are resetted.
111  *
112  */
113 static void irqs_update(struct irqt_stat *irqs, u64 ts)
114 {
115 	u64 old_ts = irqs->last_ts;
116 	u64 variance = 0;
117 	u64 interval;
118 	s64 diff;
119 
120 	/*
121 	 * The timestamps are absolute time values, we need to compute
122 	 * the timing interval between two interrupts.
123 	 */
124 	irqs->last_ts = ts;
125 
126 	/*
127 	 * The interval type is u64 in order to deal with the same
128 	 * type in our computation, that prevent mindfuck issues with
129 	 * overflow, sign and division.
130 	 */
131 	interval = ts - old_ts;
132 
133 	/*
134 	 * The interrupt triggered more than one second apart, that
135 	 * ends the sequence as predictible for our purpose. In this
136 	 * case, assume we have the beginning of a sequence and the
137 	 * timestamp is the first value. As it is impossible to
138 	 * predict anything at this point, return.
139 	 *
140 	 * Note the first timestamp of the sequence will always fall
141 	 * in this test because the old_ts is zero. That is what we
142 	 * want as we need another timestamp to compute an interval.
143 	 */
144 	if (interval >= NSEC_PER_SEC) {
145 		memset(irqs, 0, sizeof(*irqs));
146 		irqs->last_ts = ts;
147 		return;
148 	}
149 
150 	/*
151 	 * Pre-compute the delta with the average as the result is
152 	 * used several times in this function.
153 	 */
154 	diff = interval - irqs->avg;
155 
156 	/*
157 	 * Increment the number of samples.
158 	 */
159 	irqs->nr_samples++;
160 
161 	/*
162 	 * Online variance divided by the number of elements if there
163 	 * is more than one sample.  Normally the formula is division
164 	 * by nr_samples - 1 but we assume the number of element will be
165 	 * more than 32 and dividing by 32 instead of 31 is enough
166 	 * precise.
167 	 */
168 	if (likely(irqs->nr_samples > 1))
169 		variance = irqs->variance >> IRQ_TIMINGS_SHIFT;
170 
171 	/*
172 	 * The rule of thumb in statistics for the normal distribution
173 	 * is having at least 30 samples in order to have the model to
174 	 * apply. Values outside the interval are considered as an
175 	 * anomaly.
176 	 */
177 	if ((irqs->nr_samples >= 30) && ((diff * diff) > (9 * variance))) {
178 		/*
179 		 * After three consecutive anomalies, we reset the
180 		 * stats as it is no longer stable enough.
181 		 */
182 		if (irqs->anomalies++ >= 3) {
183 			memset(irqs, 0, sizeof(*irqs));
184 			irqs->last_ts = ts;
185 			return;
186 		}
187 	} else {
188 		/*
189 		 * The anomalies must be consecutives, so at this
190 		 * point, we reset the anomalies counter.
191 		 */
192 		irqs->anomalies = 0;
193 	}
194 
195 	/*
196 	 * The interrupt is considered stable enough to try to predict
197 	 * the next event on it.
198 	 */
199 	irqs->valid = 1;
200 
201 	/*
202 	 * Online average algorithm:
203 	 *
204 	 *  new_average = average + ((value - average) / count)
205 	 *
206 	 * The variance computation depends on the new average
207 	 * to be computed here first.
208 	 *
209 	 */
210 	irqs->avg = irqs->avg + (diff >> IRQ_TIMINGS_SHIFT);
211 
212 	/*
213 	 * Online variance algorithm:
214 	 *
215 	 *  new_variance = variance + (value - average) x (value - new_average)
216 	 *
217 	 * Warning: irqs->avg is updated with the line above, hence
218 	 * 'interval - irqs->avg' is no longer equal to 'diff'
219 	 */
220 	irqs->variance = irqs->variance + (diff * (interval - irqs->avg));
221 
222 	/*
223 	 * Update the next event
224 	 */
225 	irqs->next_evt = ts + irqs->avg;
226 }
227 
228 /**
229  * irq_timings_next_event - Return when the next event is supposed to arrive
230  *
231  * During the last busy cycle, the number of interrupts is incremented
232  * and stored in the irq_timings structure. This information is
233  * necessary to:
234  *
235  * - know if the index in the table wrapped up:
236  *
237  *      If more than the array size interrupts happened during the
238  *      last busy/idle cycle, the index wrapped up and we have to
239  *      begin with the next element in the array which is the last one
240  *      in the sequence, otherwise it is a the index 0.
241  *
242  * - have an indication of the interrupts activity on this CPU
243  *   (eg. irq/sec)
244  *
245  * The values are 'consumed' after inserting in the statistical model,
246  * thus the count is reinitialized.
247  *
248  * The array of values **must** be browsed in the time direction, the
249  * timestamp must increase between an element and the next one.
250  *
251  * Returns a nanosec time based estimation of the earliest interrupt,
252  * U64_MAX otherwise.
253  */
254 u64 irq_timings_next_event(u64 now)
255 {
256 	struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
257 	struct irqt_stat *irqs;
258 	struct irqt_stat __percpu *s;
259 	u64 ts, next_evt = U64_MAX;
260 	int i, irq = 0;
261 
262 	/*
263 	 * This function must be called with the local irq disabled in
264 	 * order to prevent the timings circular buffer to be updated
265 	 * while we are reading it.
266 	 */
267 	lockdep_assert_irqs_disabled();
268 
269 	/*
270 	 * Number of elements in the circular buffer: If it happens it
271 	 * was flushed before, then the number of elements could be
272 	 * smaller than IRQ_TIMINGS_SIZE, so the count is used,
273 	 * otherwise the array size is used as we wrapped. The index
274 	 * begins from zero when we did not wrap. That could be done
275 	 * in a nicer way with the proper circular array structure
276 	 * type but with the cost of extra computation in the
277 	 * interrupt handler hot path. We choose efficiency.
278 	 *
279 	 * Inject measured irq/timestamp to the statistical model
280 	 * while decrementing the counter because we consume the data
281 	 * from our circular buffer.
282 	 */
283 	for (i = irqts->count & IRQ_TIMINGS_MASK,
284 		     irqts->count = min(IRQ_TIMINGS_SIZE, irqts->count);
285 	     irqts->count > 0; irqts->count--, i = (i + 1) & IRQ_TIMINGS_MASK) {
286 
287 		irq = irq_timing_decode(irqts->values[i], &ts);
288 
289 		s = idr_find(&irqt_stats, irq);
290 		if (s) {
291 			irqs = this_cpu_ptr(s);
292 			irqs_update(irqs, ts);
293 		}
294 	}
295 
296 	/*
297 	 * Look in the list of interrupts' statistics, the earliest
298 	 * next event.
299 	 */
300 	idr_for_each_entry(&irqt_stats, s, i) {
301 
302 		irqs = this_cpu_ptr(s);
303 
304 		if (!irqs->valid)
305 			continue;
306 
307 		if (irqs->next_evt <= now) {
308 			irq = i;
309 			next_evt = now;
310 
311 			/*
312 			 * This interrupt mustn't use in the future
313 			 * until new events occur and update the
314 			 * statistics.
315 			 */
316 			irqs->valid = 0;
317 			break;
318 		}
319 
320 		if (irqs->next_evt < next_evt) {
321 			irq = i;
322 			next_evt = irqs->next_evt;
323 		}
324 	}
325 
326 	return next_evt;
327 }
328 
329 void irq_timings_free(int irq)
330 {
331 	struct irqt_stat __percpu *s;
332 
333 	s = idr_find(&irqt_stats, irq);
334 	if (s) {
335 		free_percpu(s);
336 		idr_remove(&irqt_stats, irq);
337 	}
338 }
339 
340 int irq_timings_alloc(int irq)
341 {
342 	struct irqt_stat __percpu *s;
343 	int id;
344 
345 	/*
346 	 * Some platforms can have the same private interrupt per cpu,
347 	 * so this function may be be called several times with the
348 	 * same interrupt number. Just bail out in case the per cpu
349 	 * stat structure is already allocated.
350 	 */
351 	s = idr_find(&irqt_stats, irq);
352 	if (s)
353 		return 0;
354 
355 	s = alloc_percpu(*s);
356 	if (!s)
357 		return -ENOMEM;
358 
359 	idr_preload(GFP_KERNEL);
360 	id = idr_alloc(&irqt_stats, s, irq, irq + 1, GFP_NOWAIT);
361 	idr_preload_end();
362 
363 	if (id < 0) {
364 		free_percpu(s);
365 		return id;
366 	}
367 
368 	return 0;
369 }
370