xref: /openbmc/linux/kernel/irq/timings.c (revision 2d972b6a)
1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (C) 2016, Linaro Ltd - Daniel Lezcano <daniel.lezcano@linaro.org>
3 
4 #include <linux/kernel.h>
5 #include <linux/percpu.h>
6 #include <linux/slab.h>
7 #include <linux/static_key.h>
8 #include <linux/interrupt.h>
9 #include <linux/idr.h>
10 #include <linux/irq.h>
11 #include <linux/math64.h>
12 
13 #include <trace/events/irq.h>
14 
15 #include "internals.h"
16 
17 DEFINE_STATIC_KEY_FALSE(irq_timing_enabled);
18 
19 DEFINE_PER_CPU(struct irq_timings, irq_timings);
20 
21 struct irqt_stat {
22 	u64	next_evt;
23 	u64	last_ts;
24 	u64	variance;
25 	u32	avg;
26 	u32	nr_samples;
27 	int	anomalies;
28 	int	valid;
29 };
30 
31 static DEFINE_IDR(irqt_stats);
32 
33 void irq_timings_enable(void)
34 {
35 	static_branch_enable(&irq_timing_enabled);
36 }
37 
38 void irq_timings_disable(void)
39 {
40 	static_branch_disable(&irq_timing_enabled);
41 }
42 
43 /**
44  * irqs_update - update the irq timing statistics with a new timestamp
45  *
46  * @irqs: an irqt_stat struct pointer
47  * @ts: the new timestamp
48  *
49  * The statistics are computed online, in other words, the code is
50  * designed to compute the statistics on a stream of values rather
51  * than doing multiple passes on the values to compute the average,
52  * then the variance. The integer division introduces a loss of
53  * precision but with an acceptable error margin regarding the results
54  * we would have with the double floating precision: we are dealing
55  * with nanosec, so big numbers, consequently the mantisse is
56  * negligeable, especially when converting the time in usec
57  * afterwards.
58  *
59  * The computation happens at idle time. When the CPU is not idle, the
60  * interrupts' timestamps are stored in the circular buffer, when the
61  * CPU goes idle and this routine is called, all the buffer's values
62  * are injected in the statistical model continuying to extend the
63  * statistics from the previous busy-idle cycle.
64  *
65  * The observations showed a device will trigger a burst of periodic
66  * interrupts followed by one or two peaks of longer time, for
67  * instance when a SD card device flushes its cache, then the periodic
68  * intervals occur again. A one second inactivity period resets the
69  * stats, that gives us the certitude the statistical values won't
70  * exceed 1x10^9, thus the computation won't overflow.
71  *
72  * Basically, the purpose of the algorithm is to watch the periodic
73  * interrupts and eliminate the peaks.
74  *
75  * An interrupt is considered periodically stable if the interval of
76  * its occurences follow the normal distribution, thus the values
77  * comply with:
78  *
79  *      avg - 3 x stddev < value < avg + 3 x stddev
80  *
81  * Which can be simplified to:
82  *
83  *      -3 x stddev < value - avg < 3 x stddev
84  *
85  *      abs(value - avg) < 3 x stddev
86  *
87  * In order to save a costly square root computation, we use the
88  * variance. For the record, stddev = sqrt(variance). The equation
89  * above becomes:
90  *
91  *      abs(value - avg) < 3 x sqrt(variance)
92  *
93  * And finally we square it:
94  *
95  *      (value - avg) ^ 2 < (3 x sqrt(variance)) ^ 2
96  *
97  *      (value - avg) x (value - avg) < 9 x variance
98  *
99  * Statistically speaking, any values out of this interval is
100  * considered as an anomaly and is discarded. However, a normal
101  * distribution appears when the number of samples is 30 (it is the
102  * rule of thumb in statistics, cf. "30 samples" on Internet). When
103  * there are three consecutive anomalies, the statistics are resetted.
104  *
105  */
106 static void irqs_update(struct irqt_stat *irqs, u64 ts)
107 {
108 	u64 old_ts = irqs->last_ts;
109 	u64 variance = 0;
110 	u64 interval;
111 	s64 diff;
112 
113 	/*
114 	 * The timestamps are absolute time values, we need to compute
115 	 * the timing interval between two interrupts.
116 	 */
117 	irqs->last_ts = ts;
118 
119 	/*
120 	 * The interval type is u64 in order to deal with the same
121 	 * type in our computation, that prevent mindfuck issues with
122 	 * overflow, sign and division.
123 	 */
124 	interval = ts - old_ts;
125 
126 	/*
127 	 * The interrupt triggered more than one second apart, that
128 	 * ends the sequence as predictible for our purpose. In this
129 	 * case, assume we have the beginning of a sequence and the
130 	 * timestamp is the first value. As it is impossible to
131 	 * predict anything at this point, return.
132 	 *
133 	 * Note the first timestamp of the sequence will always fall
134 	 * in this test because the old_ts is zero. That is what we
135 	 * want as we need another timestamp to compute an interval.
136 	 */
137 	if (interval >= NSEC_PER_SEC) {
138 		memset(irqs, 0, sizeof(*irqs));
139 		irqs->last_ts = ts;
140 		return;
141 	}
142 
143 	/*
144 	 * Pre-compute the delta with the average as the result is
145 	 * used several times in this function.
146 	 */
147 	diff = interval - irqs->avg;
148 
149 	/*
150 	 * Increment the number of samples.
151 	 */
152 	irqs->nr_samples++;
153 
154 	/*
155 	 * Online variance divided by the number of elements if there
156 	 * is more than one sample.  Normally the formula is division
157 	 * by nr_samples - 1 but we assume the number of element will be
158 	 * more than 32 and dividing by 32 instead of 31 is enough
159 	 * precise.
160 	 */
161 	if (likely(irqs->nr_samples > 1))
162 		variance = irqs->variance >> IRQ_TIMINGS_SHIFT;
163 
164 	/*
165 	 * The rule of thumb in statistics for the normal distribution
166 	 * is having at least 30 samples in order to have the model to
167 	 * apply. Values outside the interval are considered as an
168 	 * anomaly.
169 	 */
170 	if ((irqs->nr_samples >= 30) && ((diff * diff) > (9 * variance))) {
171 		/*
172 		 * After three consecutive anomalies, we reset the
173 		 * stats as it is no longer stable enough.
174 		 */
175 		if (irqs->anomalies++ >= 3) {
176 			memset(irqs, 0, sizeof(*irqs));
177 			irqs->last_ts = ts;
178 			return;
179 		}
180 	} else {
181 		/*
182 		 * The anomalies must be consecutives, so at this
183 		 * point, we reset the anomalies counter.
184 		 */
185 		irqs->anomalies = 0;
186 	}
187 
188 	/*
189 	 * The interrupt is considered stable enough to try to predict
190 	 * the next event on it.
191 	 */
192 	irqs->valid = 1;
193 
194 	/*
195 	 * Online average algorithm:
196 	 *
197 	 *  new_average = average + ((value - average) / count)
198 	 *
199 	 * The variance computation depends on the new average
200 	 * to be computed here first.
201 	 *
202 	 */
203 	irqs->avg = irqs->avg + (diff >> IRQ_TIMINGS_SHIFT);
204 
205 	/*
206 	 * Online variance algorithm:
207 	 *
208 	 *  new_variance = variance + (value - average) x (value - new_average)
209 	 *
210 	 * Warning: irqs->avg is updated with the line above, hence
211 	 * 'interval - irqs->avg' is no longer equal to 'diff'
212 	 */
213 	irqs->variance = irqs->variance + (diff * (interval - irqs->avg));
214 
215 	/*
216 	 * Update the next event
217 	 */
218 	irqs->next_evt = ts + irqs->avg;
219 }
220 
221 /**
222  * irq_timings_next_event - Return when the next event is supposed to arrive
223  *
224  * During the last busy cycle, the number of interrupts is incremented
225  * and stored in the irq_timings structure. This information is
226  * necessary to:
227  *
228  * - know if the index in the table wrapped up:
229  *
230  *      If more than the array size interrupts happened during the
231  *      last busy/idle cycle, the index wrapped up and we have to
232  *      begin with the next element in the array which is the last one
233  *      in the sequence, otherwise it is a the index 0.
234  *
235  * - have an indication of the interrupts activity on this CPU
236  *   (eg. irq/sec)
237  *
238  * The values are 'consumed' after inserting in the statistical model,
239  * thus the count is reinitialized.
240  *
241  * The array of values **must** be browsed in the time direction, the
242  * timestamp must increase between an element and the next one.
243  *
244  * Returns a nanosec time based estimation of the earliest interrupt,
245  * U64_MAX otherwise.
246  */
247 u64 irq_timings_next_event(u64 now)
248 {
249 	struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
250 	struct irqt_stat *irqs;
251 	struct irqt_stat __percpu *s;
252 	u64 ts, next_evt = U64_MAX;
253 	int i, irq = 0;
254 
255 	/*
256 	 * This function must be called with the local irq disabled in
257 	 * order to prevent the timings circular buffer to be updated
258 	 * while we are reading it.
259 	 */
260 	lockdep_assert_irqs_disabled();
261 
262 	/*
263 	 * Number of elements in the circular buffer: If it happens it
264 	 * was flushed before, then the number of elements could be
265 	 * smaller than IRQ_TIMINGS_SIZE, so the count is used,
266 	 * otherwise the array size is used as we wrapped. The index
267 	 * begins from zero when we did not wrap. That could be done
268 	 * in a nicer way with the proper circular array structure
269 	 * type but with the cost of extra computation in the
270 	 * interrupt handler hot path. We choose efficiency.
271 	 *
272 	 * Inject measured irq/timestamp to the statistical model
273 	 * while decrementing the counter because we consume the data
274 	 * from our circular buffer.
275 	 */
276 	for (i = irqts->count & IRQ_TIMINGS_MASK,
277 		     irqts->count = min(IRQ_TIMINGS_SIZE, irqts->count);
278 	     irqts->count > 0; irqts->count--, i = (i + 1) & IRQ_TIMINGS_MASK) {
279 
280 		irq = irq_timing_decode(irqts->values[i], &ts);
281 
282 		s = idr_find(&irqt_stats, irq);
283 		if (s) {
284 			irqs = this_cpu_ptr(s);
285 			irqs_update(irqs, ts);
286 		}
287 	}
288 
289 	/*
290 	 * Look in the list of interrupts' statistics, the earliest
291 	 * next event.
292 	 */
293 	idr_for_each_entry(&irqt_stats, s, i) {
294 
295 		irqs = this_cpu_ptr(s);
296 
297 		if (!irqs->valid)
298 			continue;
299 
300 		if (irqs->next_evt <= now) {
301 			irq = i;
302 			next_evt = now;
303 
304 			/*
305 			 * This interrupt mustn't use in the future
306 			 * until new events occur and update the
307 			 * statistics.
308 			 */
309 			irqs->valid = 0;
310 			break;
311 		}
312 
313 		if (irqs->next_evt < next_evt) {
314 			irq = i;
315 			next_evt = irqs->next_evt;
316 		}
317 	}
318 
319 	return next_evt;
320 }
321 
322 void irq_timings_free(int irq)
323 {
324 	struct irqt_stat __percpu *s;
325 
326 	s = idr_find(&irqt_stats, irq);
327 	if (s) {
328 		free_percpu(s);
329 		idr_remove(&irqt_stats, irq);
330 	}
331 }
332 
333 int irq_timings_alloc(int irq)
334 {
335 	struct irqt_stat __percpu *s;
336 	int id;
337 
338 	/*
339 	 * Some platforms can have the same private interrupt per cpu,
340 	 * so this function may be be called several times with the
341 	 * same interrupt number. Just bail out in case the per cpu
342 	 * stat structure is already allocated.
343 	 */
344 	s = idr_find(&irqt_stats, irq);
345 	if (s)
346 		return 0;
347 
348 	s = alloc_percpu(*s);
349 	if (!s)
350 		return -ENOMEM;
351 
352 	idr_preload(GFP_KERNEL);
353 	id = idr_alloc(&irqt_stats, s, irq, irq + 1, GFP_NOWAIT);
354 	idr_preload_end();
355 
356 	if (id < 0) {
357 		free_percpu(s);
358 		return id;
359 	}
360 
361 	return 0;
362 }
363