xref: /openbmc/linux/kernel/irq/timings.c (revision 6aed82de719b424bd5548aa4179e95f34fd779ab)
1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (C) 2016, Linaro Ltd - Daniel Lezcano <daniel.lezcano@linaro.org>
3 #define pr_fmt(fmt) "irq_timings: " fmt
4 
5 #include <linux/kernel.h>
6 #include <linux/percpu.h>
7 #include <linux/slab.h>
8 #include <linux/static_key.h>
9 #include <linux/init.h>
10 #include <linux/interrupt.h>
11 #include <linux/idr.h>
12 #include <linux/irq.h>
13 #include <linux/math64.h>
14 #include <linux/log2.h>
15 
16 #include <trace/events/irq.h>
17 
18 #include "internals.h"
19 
20 DEFINE_STATIC_KEY_FALSE(irq_timing_enabled);
21 
22 DEFINE_PER_CPU(struct irq_timings, irq_timings);
23 
24 static DEFINE_IDR(irqt_stats);
25 
26 void irq_timings_enable(void)
27 {
28 	static_branch_enable(&irq_timing_enabled);
29 }
30 
31 void irq_timings_disable(void)
32 {
33 	static_branch_disable(&irq_timing_enabled);
34 }
35 
36 /*
37  * The main goal of this algorithm is to predict the next interrupt
38  * occurrence on the current CPU.
39  *
40  * Currently, the interrupt timings are stored in a circular array
41  * buffer every time there is an interrupt, as a tuple: the interrupt
42  * number and the associated timestamp when the event occurred <irq,
43  * timestamp>.
44  *
45  * For every interrupt occurring in a short period of time, we can
46  * measure the elapsed time between the occurrences for the same
47  * interrupt and we end up with a suite of intervals. The experience
48  * showed the interrupts are often coming following a periodic
49  * pattern.
50  *
51  * The objective of the algorithm is to find out this periodic pattern
52  * in a fastest way and use its period to predict the next irq event.
53  *
54  * When the next interrupt event is requested, we are in the situation
55  * where the interrupts are disabled and the circular buffer
56  * containing the timings is filled with the events which happened
57  * after the previous next-interrupt-event request.
58  *
59  * At this point, we read the circular buffer and we fill the irq
60  * related statistics structure. After this step, the circular array
61  * containing the timings is empty because all the values are
62  * dispatched in their corresponding buffers.
63  *
64  * Now for each interrupt, we can predict the next event by using the
65  * suffix array, log interval and exponential moving average
66  *
67  * 1. Suffix array
68  *
69  * Suffix array is an array of all the suffixes of a string. It is
70  * widely used as a data structure for compression, text search, ...
71  * For instance for the word 'banana', the suffixes will be: 'banana'
72  * 'anana' 'nana' 'ana' 'na' 'a'
73  *
74  * Usually, the suffix array is sorted but for our purpose it is
75  * not necessary and won't provide any improvement in the context of
76  * the solved problem where we clearly define the boundaries of the
77  * search by a max period and min period.
78  *
79  * The suffix array will build a suite of intervals of different
80  * length and will look for the repetition of each suite. If the suite
81  * is repeating then we have the period because it is the length of
82  * the suite whatever its position in the buffer.
83  *
84  * 2. Log interval
85  *
86  * We saw the irq timings allow to compute the interval of the
87  * occurrences for a specific interrupt. We can reasonibly assume the
88  * longer is the interval, the higher is the error for the next event
89  * and we can consider storing those interval values into an array
90  * where each slot in the array correspond to an interval at the power
91  * of 2 of the index. For example, index 12 will contain values
92  * between 2^11 and 2^12.
93  *
94  * At the end we have an array of values where at each index defines a
95  * [2^index - 1, 2 ^ index] interval values allowing to store a large
96  * number of values inside a small array.
97  *
98  * For example, if we have the value 1123, then we store it at
99  * ilog2(1123) = 10 index value.
100  *
101  * Storing those value at the specific index is done by computing an
102  * exponential moving average for this specific slot. For instance,
103  * for values 1800, 1123, 1453, ... fall under the same slot (10) and
104  * the exponential moving average is computed every time a new value
105  * is stored at this slot.
106  *
107  * 3. Exponential Moving Average
108  *
109  * The EMA is largely used to track a signal for stocks or as a low
110  * pass filter. The magic of the formula, is it is very simple and the
111  * reactivity of the average can be tuned with the factors called
112  * alpha.
113  *
114  * The higher the alphas are, the faster the average respond to the
115  * signal change. In our case, if a slot in the array is a big
116  * interval, we can have numbers with a big difference between
117  * them. The impact of those differences in the average computation
118  * can be tuned by changing the alpha value.
119  *
120  *
121  *  -- The algorithm --
122  *
123  * We saw the different processing above, now let's see how they are
124  * used together.
125  *
126  * For each interrupt:
127  *	For each interval:
128  *		Compute the index = ilog2(interval)
129  *		Compute a new_ema(buffer[index], interval)
130  *		Store the index in a circular buffer
131  *
132  *	Compute the suffix array of the indexes
133  *
134  *	For each suffix:
135  *		If the suffix is reverse-found 3 times
136  *			Return suffix
137  *
138  *	Return Not found
139  *
140  * However we can not have endless suffix array to be build, it won't
141  * make sense and it will add an extra overhead, so we can restrict
142  * this to a maximum suffix length of 5 and a minimum suffix length of
143  * 2. The experience showed 5 is the majority of the maximum pattern
144  * period found for different devices.
145  *
146  * The result is a pattern finding less than 1us for an interrupt.
147  *
148  * Example based on real values:
149  *
150  * Example 1 : MMC write/read interrupt interval:
151  *
152  *	223947, 1240, 1384, 1386, 1386,
153  *	217416, 1236, 1384, 1386, 1387,
154  *	214719, 1241, 1386, 1387, 1384,
155  *	213696, 1234, 1384, 1386, 1388,
156  *	219904, 1240, 1385, 1389, 1385,
157  *	212240, 1240, 1386, 1386, 1386,
158  *	214415, 1236, 1384, 1386, 1387,
159  *	214276, 1234, 1384, 1388, ?
160  *
161  * For each element, apply ilog2(value)
162  *
163  *	15, 8, 8, 8, 8,
164  *	15, 8, 8, 8, 8,
165  *	15, 8, 8, 8, 8,
166  *	15, 8, 8, 8, 8,
167  *	15, 8, 8, 8, 8,
168  *	15, 8, 8, 8, 8,
169  *	15, 8, 8, 8, 8,
170  *	15, 8, 8, 8, ?
171  *
172  * Max period of 5, we take the last (max_period * 3) 15 elements as
173  * we can be confident if the pattern repeats itself three times it is
174  * a repeating pattern.
175  *
176  *	             8,
177  *	15, 8, 8, 8, 8,
178  *	15, 8, 8, 8, 8,
179  *	15, 8, 8, 8, ?
180  *
181  * Suffixes are:
182  *
183  *  1) 8, 15, 8, 8, 8  <- max period
184  *  2) 8, 15, 8, 8
185  *  3) 8, 15, 8
186  *  4) 8, 15           <- min period
187  *
188  * From there we search the repeating pattern for each suffix.
189  *
190  * buffer: 8, 15, 8, 8, 8, 8, 15, 8, 8, 8, 8, 15, 8, 8, 8
191  *         |   |  |  |  |  |   |  |  |  |  |   |  |  |  |
192  *         8, 15, 8, 8, 8  |   |  |  |  |  |   |  |  |  |
193  *                         8, 15, 8, 8, 8  |   |  |  |  |
194  *                                         8, 15, 8, 8, 8
195  *
196  * When moving the suffix, we found exactly 3 matches.
197  *
198  * The first suffix with period 5 is repeating.
199  *
200  * The next event is (3 * max_period) % suffix_period
201  *
202  * In this example, the result 0, so the next event is suffix[0] => 8
203  *
204  * However, 8 is the index in the array of exponential moving average
205  * which was calculated on the fly when storing the values, so the
206  * interval is ema[8] = 1366
207  *
208  *
209  * Example 2:
210  *
211  *	4, 3, 5, 100,
212  *	3, 3, 5, 117,
213  *	4, 4, 5, 112,
214  *	4, 3, 4, 110,
215  *	3, 5, 3, 117,
216  *	4, 4, 5, 112,
217  *	4, 3, 4, 110,
218  *	3, 4, 5, 112,
219  *	4, 3, 4, 110
220  *
221  * ilog2
222  *
223  *	0, 0, 0, 4,
224  *	0, 0, 0, 4,
225  *	0, 0, 0, 4,
226  *	0, 0, 0, 4,
227  *	0, 0, 0, 4,
228  *	0, 0, 0, 4,
229  *	0, 0, 0, 4,
230  *	0, 0, 0, 4,
231  *	0, 0, 0, 4
232  *
233  * Max period 5:
234  *	   0, 0, 4,
235  *	0, 0, 0, 4,
236  *	0, 0, 0, 4,
237  *	0, 0, 0, 4
238  *
239  * Suffixes:
240  *
241  *  1) 0, 0, 4, 0, 0
242  *  2) 0, 0, 4, 0
243  *  3) 0, 0, 4
244  *  4) 0, 0
245  *
246  * buffer: 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4
247  *         |  |  |  |  |  |  X
248  *         0, 0, 4, 0, 0, |  X
249  *                        0, 0
250  *
251  * buffer: 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4
252  *         |  |  |  |  |  |  |  |  |  |  |  |  |  |  |
253  *         0, 0, 4, 0, |  |  |  |  |  |  |  |  |  |  |
254  *                     0, 0, 4, 0, |  |  |  |  |  |  |
255  *                                 0, 0, 4, 0, |  |  |
256  *                                             0  0  4
257  *
258  * Pattern is found 3 times, the remaining is 1 which results from
259  * (max_period * 3) % suffix_period. This value is the index in the
260  * suffix arrays. The suffix array for a period 4 has the value 4
261  * at index 1.
262  */
263 #define EMA_ALPHA_VAL		64
264 #define EMA_ALPHA_SHIFT		7
265 
266 #define PREDICTION_PERIOD_MIN	3
267 #define PREDICTION_PERIOD_MAX	5
268 #define PREDICTION_FACTOR	4
269 #define PREDICTION_MAX		10 /* 2 ^ PREDICTION_MAX useconds */
270 #define PREDICTION_BUFFER_SIZE	16 /* slots for EMAs, hardly more than 16 */
271 
272 /*
273  * Number of elements in the circular buffer: If it happens it was
274  * flushed before, then the number of elements could be smaller than
275  * IRQ_TIMINGS_SIZE, so the count is used, otherwise the array size is
276  * used as we wrapped. The index begins from zero when we did not
277  * wrap. That could be done in a nicer way with the proper circular
278  * array structure type but with the cost of extra computation in the
279  * interrupt handler hot path. We choose efficiency.
280  */
281 #define for_each_irqts(i, irqts)					\
282 	for (i = irqts->count < IRQ_TIMINGS_SIZE ?			\
283 		     0 : irqts->count & IRQ_TIMINGS_MASK,		\
284 		     irqts->count = min(IRQ_TIMINGS_SIZE,		\
285 					irqts->count);			\
286 	     irqts->count > 0; irqts->count--,				\
287 		     i = (i + 1) & IRQ_TIMINGS_MASK)
288 
289 struct irqt_stat {
290 	u64	last_ts;
291 	u64	ema_time[PREDICTION_BUFFER_SIZE];
292 	int	timings[IRQ_TIMINGS_SIZE];
293 	int	circ_timings[IRQ_TIMINGS_SIZE];
294 	int	count;
295 };
296 
297 /*
298  * Exponential moving average computation
299  */
300 static u64 irq_timings_ema_new(u64 value, u64 ema_old)
301 {
302 	s64 diff;
303 
304 	if (unlikely(!ema_old))
305 		return value;
306 
307 	diff = (value - ema_old) * EMA_ALPHA_VAL;
308 	/*
309 	 * We can use a s64 type variable to be added with the u64
310 	 * ema_old variable as this one will never have its topmost
311 	 * bit set, it will be always smaller than 2^63 nanosec
312 	 * interrupt interval (292 years).
313 	 */
314 	return ema_old + (diff >> EMA_ALPHA_SHIFT);
315 }
316 
317 static int irq_timings_next_event_index(int *buffer, size_t len, int period_max)
318 {
319 	int period;
320 
321 	/*
322 	 * Move the beginning pointer to the end minus the max period x 3.
323 	 * We are at the point we can begin searching the pattern
324 	 */
325 	buffer = &buffer[len - (period_max * 3)];
326 
327 	/* Adjust the length to the maximum allowed period x 3 */
328 	len = period_max * 3;
329 
330 	/*
331 	 * The buffer contains the suite of intervals, in a ilog2
332 	 * basis, we are looking for a repetition. We point the
333 	 * beginning of the search three times the length of the
334 	 * period beginning at the end of the buffer. We do that for
335 	 * each suffix.
336 	 */
337 	for (period = period_max; period >= PREDICTION_PERIOD_MIN; period--) {
338 
339 		/*
340 		 * The first comparison always succeed because the
341 		 * suffix is deduced from the first n-period bytes of
342 		 * the buffer and we compare the initial suffix with
343 		 * itself, so we can skip the first iteration.
344 		 */
345 		int idx = period;
346 		size_t size = period;
347 
348 		/*
349 		 * We look if the suite with period 'i' repeat
350 		 * itself. If it is truncated at the end, as it
351 		 * repeats we can use the period to find out the next
352 		 * element with the modulo.
353 		 */
354 		while (!memcmp(buffer, &buffer[idx], size * sizeof(int))) {
355 
356 			/*
357 			 * Move the index in a period basis
358 			 */
359 			idx += size;
360 
361 			/*
362 			 * If this condition is reached, all previous
363 			 * memcmp were successful, so the period is
364 			 * found.
365 			 */
366 			if (idx == len)
367 				return buffer[len % period];
368 
369 			/*
370 			 * If the remaining elements to compare are
371 			 * smaller than the period, readjust the size
372 			 * of the comparison for the last iteration.
373 			 */
374 			if (len - idx < period)
375 				size = len - idx;
376 		}
377 	}
378 
379 	return -1;
380 }
381 
382 static u64 __irq_timings_next_event(struct irqt_stat *irqs, int irq, u64 now)
383 {
384 	int index, i, period_max, count, start, min = INT_MAX;
385 
386 	if ((now - irqs->last_ts) >= NSEC_PER_SEC) {
387 		irqs->count = irqs->last_ts = 0;
388 		return U64_MAX;
389 	}
390 
391 	/*
392 	 * As we want to find three times the repetition, we need a
393 	 * number of intervals greater or equal to three times the
394 	 * maximum period, otherwise we truncate the max period.
395 	 */
396 	period_max = irqs->count > (3 * PREDICTION_PERIOD_MAX) ?
397 		PREDICTION_PERIOD_MAX : irqs->count / 3;
398 
399 	/*
400 	 * If we don't have enough irq timings for this prediction,
401 	 * just bail out.
402 	 */
403 	if (period_max <= PREDICTION_PERIOD_MIN)
404 		return U64_MAX;
405 
406 	/*
407 	 * 'count' will depends if the circular buffer wrapped or not
408 	 */
409 	count = irqs->count < IRQ_TIMINGS_SIZE ?
410 		irqs->count : IRQ_TIMINGS_SIZE;
411 
412 	start = irqs->count < IRQ_TIMINGS_SIZE ?
413 		0 : (irqs->count & IRQ_TIMINGS_MASK);
414 
415 	/*
416 	 * Copy the content of the circular buffer into another buffer
417 	 * in order to linearize the buffer instead of dealing with
418 	 * wrapping indexes and shifted array which will be prone to
419 	 * error and extremelly difficult to debug.
420 	 */
421 	for (i = 0; i < count; i++) {
422 		int index = (start + i) & IRQ_TIMINGS_MASK;
423 
424 		irqs->timings[i] = irqs->circ_timings[index];
425 		min = min_t(int, irqs->timings[i], min);
426 	}
427 
428 	index = irq_timings_next_event_index(irqs->timings, count, period_max);
429 	if (index < 0)
430 		return irqs->last_ts + irqs->ema_time[min];
431 
432 	return irqs->last_ts + irqs->ema_time[index];
433 }
434 
435 static __always_inline int irq_timings_interval_index(u64 interval)
436 {
437 	/*
438 	 * The PREDICTION_FACTOR increase the interval size for the
439 	 * array of exponential average.
440 	 */
441 	u64 interval_us = (interval >> 10) / PREDICTION_FACTOR;
442 
443 	return likely(interval_us) ? ilog2(interval_us) : 0;
444 }
445 
446 static __always_inline void __irq_timings_store(int irq, struct irqt_stat *irqs,
447 						u64 interval)
448 {
449 	int index;
450 
451 	/*
452 	 * Get the index in the ema table for this interrupt.
453 	 */
454 	index = irq_timings_interval_index(interval);
455 
456 	/*
457 	 * Store the index as an element of the pattern in another
458 	 * circular array.
459 	 */
460 	irqs->circ_timings[irqs->count & IRQ_TIMINGS_MASK] = index;
461 
462 	irqs->ema_time[index] = irq_timings_ema_new(interval,
463 						    irqs->ema_time[index]);
464 
465 	irqs->count++;
466 }
467 
468 static inline void irq_timings_store(int irq, struct irqt_stat *irqs, u64 ts)
469 {
470 	u64 old_ts = irqs->last_ts;
471 	u64 interval;
472 
473 	/*
474 	 * The timestamps are absolute time values, we need to compute
475 	 * the timing interval between two interrupts.
476 	 */
477 	irqs->last_ts = ts;
478 
479 	/*
480 	 * The interval type is u64 in order to deal with the same
481 	 * type in our computation, that prevent mindfuck issues with
482 	 * overflow, sign and division.
483 	 */
484 	interval = ts - old_ts;
485 
486 	/*
487 	 * The interrupt triggered more than one second apart, that
488 	 * ends the sequence as predictible for our purpose. In this
489 	 * case, assume we have the beginning of a sequence and the
490 	 * timestamp is the first value. As it is impossible to
491 	 * predict anything at this point, return.
492 	 *
493 	 * Note the first timestamp of the sequence will always fall
494 	 * in this test because the old_ts is zero. That is what we
495 	 * want as we need another timestamp to compute an interval.
496 	 */
497 	if (interval >= NSEC_PER_SEC) {
498 		irqs->count = 0;
499 		return;
500 	}
501 
502 	__irq_timings_store(irq, irqs, interval);
503 }
504 
505 /**
506  * irq_timings_next_event - Return when the next event is supposed to arrive
507  *
508  * During the last busy cycle, the number of interrupts is incremented
509  * and stored in the irq_timings structure. This information is
510  * necessary to:
511  *
512  * - know if the index in the table wrapped up:
513  *
514  *      If more than the array size interrupts happened during the
515  *      last busy/idle cycle, the index wrapped up and we have to
516  *      begin with the next element in the array which is the last one
517  *      in the sequence, otherwise it is a the index 0.
518  *
519  * - have an indication of the interrupts activity on this CPU
520  *   (eg. irq/sec)
521  *
522  * The values are 'consumed' after inserting in the statistical model,
523  * thus the count is reinitialized.
524  *
525  * The array of values **must** be browsed in the time direction, the
526  * timestamp must increase between an element and the next one.
527  *
528  * Returns a nanosec time based estimation of the earliest interrupt,
529  * U64_MAX otherwise.
530  */
531 u64 irq_timings_next_event(u64 now)
532 {
533 	struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
534 	struct irqt_stat *irqs;
535 	struct irqt_stat __percpu *s;
536 	u64 ts, next_evt = U64_MAX;
537 	int i, irq = 0;
538 
539 	/*
540 	 * This function must be called with the local irq disabled in
541 	 * order to prevent the timings circular buffer to be updated
542 	 * while we are reading it.
543 	 */
544 	lockdep_assert_irqs_disabled();
545 
546 	if (!irqts->count)
547 		return next_evt;
548 
549 	/*
550 	 * Number of elements in the circular buffer: If it happens it
551 	 * was flushed before, then the number of elements could be
552 	 * smaller than IRQ_TIMINGS_SIZE, so the count is used,
553 	 * otherwise the array size is used as we wrapped. The index
554 	 * begins from zero when we did not wrap. That could be done
555 	 * in a nicer way with the proper circular array structure
556 	 * type but with the cost of extra computation in the
557 	 * interrupt handler hot path. We choose efficiency.
558 	 *
559 	 * Inject measured irq/timestamp to the pattern prediction
560 	 * model while decrementing the counter because we consume the
561 	 * data from our circular buffer.
562 	 */
563 	for_each_irqts(i, irqts) {
564 		irq = irq_timing_decode(irqts->values[i], &ts);
565 		s = idr_find(&irqt_stats, irq);
566 		if (s)
567 			irq_timings_store(irq, this_cpu_ptr(s), ts);
568 	}
569 
570 	/*
571 	 * Look in the list of interrupts' statistics, the earliest
572 	 * next event.
573 	 */
574 	idr_for_each_entry(&irqt_stats, s, i) {
575 
576 		irqs = this_cpu_ptr(s);
577 
578 		ts = __irq_timings_next_event(irqs, i, now);
579 		if (ts <= now)
580 			return now;
581 
582 		if (ts < next_evt)
583 			next_evt = ts;
584 	}
585 
586 	return next_evt;
587 }
588 
589 void irq_timings_free(int irq)
590 {
591 	struct irqt_stat __percpu *s;
592 
593 	s = idr_find(&irqt_stats, irq);
594 	if (s) {
595 		free_percpu(s);
596 		idr_remove(&irqt_stats, irq);
597 	}
598 }
599 
600 int irq_timings_alloc(int irq)
601 {
602 	struct irqt_stat __percpu *s;
603 	int id;
604 
605 	/*
606 	 * Some platforms can have the same private interrupt per cpu,
607 	 * so this function may be be called several times with the
608 	 * same interrupt number. Just bail out in case the per cpu
609 	 * stat structure is already allocated.
610 	 */
611 	s = idr_find(&irqt_stats, irq);
612 	if (s)
613 		return 0;
614 
615 	s = alloc_percpu(*s);
616 	if (!s)
617 		return -ENOMEM;
618 
619 	idr_preload(GFP_KERNEL);
620 	id = idr_alloc(&irqt_stats, s, irq, irq + 1, GFP_NOWAIT);
621 	idr_preload_end();
622 
623 	if (id < 0) {
624 		free_percpu(s);
625 		return id;
626 	}
627 
628 	return 0;
629 }
630 
631 #ifdef CONFIG_TEST_IRQ_TIMINGS
632 static int __init irq_timings_test_irqts(struct irq_timings *irqts,
633 					 unsigned count)
634 {
635 	int start = count >= IRQ_TIMINGS_SIZE ? count - IRQ_TIMINGS_SIZE : 0;
636 	int i, irq, oirq = 0xBEEF;
637 	u64 ots = 0xDEAD, ts;
638 
639 	/*
640 	 * Fill the circular buffer by using the dedicated function.
641 	 */
642 	for (i = 0; i < count; i++) {
643 		pr_debug("%d: index=%d, ts=%llX irq=%X\n",
644 			 i, i & IRQ_TIMINGS_MASK, ots + i, oirq + i);
645 
646 		irq_timings_push(ots + i, oirq + i);
647 	}
648 
649 	/*
650 	 * Compute the first elements values after the index wrapped
651 	 * up or not.
652 	 */
653 	ots += start;
654 	oirq += start;
655 
656 	/*
657 	 * Test the circular buffer count is correct.
658 	 */
659 	pr_debug("---> Checking timings array count (%d) is right\n", count);
660 	if (WARN_ON(irqts->count != count))
661 		return -EINVAL;
662 
663 	/*
664 	 * Test the macro allowing to browse all the irqts.
665 	 */
666 	pr_debug("---> Checking the for_each_irqts() macro\n");
667 	for_each_irqts(i, irqts) {
668 
669 		irq = irq_timing_decode(irqts->values[i], &ts);
670 
671 		pr_debug("index=%d, ts=%llX / %llX, irq=%X / %X\n",
672 			 i, ts, ots, irq, oirq);
673 
674 		if (WARN_ON(ts != ots || irq != oirq))
675 			return -EINVAL;
676 
677 		ots++; oirq++;
678 	}
679 
680 	/*
681 	 * The circular buffer should have be flushed when browsed
682 	 * with for_each_irqts
683 	 */
684 	pr_debug("---> Checking timings array is empty after browsing it\n");
685 	if (WARN_ON(irqts->count))
686 		return -EINVAL;
687 
688 	return 0;
689 }
690 
691 static int __init irq_timings_irqts_selftest(void)
692 {
693 	struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
694 	int i, ret;
695 
696 	/*
697 	 * Test the circular buffer with different number of
698 	 * elements. The purpose is to test at the limits (empty, half
699 	 * full, full, wrapped with the cursor at the boundaries,
700 	 * wrapped several times, etc ...
701 	 */
702 	int count[] = { 0,
703 			IRQ_TIMINGS_SIZE >> 1,
704 			IRQ_TIMINGS_SIZE,
705 			IRQ_TIMINGS_SIZE + (IRQ_TIMINGS_SIZE >> 1),
706 			2 * IRQ_TIMINGS_SIZE,
707 			(2 * IRQ_TIMINGS_SIZE) + 3,
708 	};
709 
710 	for (i = 0; i < ARRAY_SIZE(count); i++) {
711 
712 		pr_info("---> Checking the timings with %d/%d values\n",
713 			count[i], IRQ_TIMINGS_SIZE);
714 
715 		ret = irq_timings_test_irqts(irqts, count[i]);
716 		if (ret)
717 			break;
718 	}
719 
720 	return ret;
721 }
722 
723 static int __init irq_timings_selftest(void)
724 {
725 	int ret;
726 
727 	pr_info("------------------- selftest start -----------------\n");
728 
729 	/*
730 	 * At this point, we don't except any subsystem to use the irq
731 	 * timings but us, so it should not be enabled.
732 	 */
733 	if (static_branch_unlikely(&irq_timing_enabled)) {
734 		pr_warn("irq timings already initialized, skipping selftest\n");
735 		return 0;
736 	}
737 
738 	ret = irq_timings_irqts_selftest();
739 
740 	pr_info("---------- selftest end with %s -----------\n",
741 		ret ? "failure" : "success");
742 
743 	return ret;
744 }
745 early_initcall(irq_timings_selftest);
746 #endif
747