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 struct timings_intervals {
633 u64 *intervals;
634 size_t count;
635 };
636
637 /*
638 * Intervals are given in nanosecond base
639 */
640 static u64 intervals0[] __initdata = {
641 10000, 50000, 200000, 500000,
642 10000, 50000, 200000, 500000,
643 10000, 50000, 200000, 500000,
644 10000, 50000, 200000, 500000,
645 10000, 50000, 200000, 500000,
646 10000, 50000, 200000, 500000,
647 10000, 50000, 200000, 500000,
648 10000, 50000, 200000, 500000,
649 10000, 50000, 200000,
650 };
651
652 static u64 intervals1[] __initdata = {
653 223947000, 1240000, 1384000, 1386000, 1386000,
654 217416000, 1236000, 1384000, 1386000, 1387000,
655 214719000, 1241000, 1386000, 1387000, 1384000,
656 213696000, 1234000, 1384000, 1386000, 1388000,
657 219904000, 1240000, 1385000, 1389000, 1385000,
658 212240000, 1240000, 1386000, 1386000, 1386000,
659 214415000, 1236000, 1384000, 1386000, 1387000,
660 214276000, 1234000,
661 };
662
663 static u64 intervals2[] __initdata = {
664 4000, 3000, 5000, 100000,
665 3000, 3000, 5000, 117000,
666 4000, 4000, 5000, 112000,
667 4000, 3000, 4000, 110000,
668 3000, 5000, 3000, 117000,
669 4000, 4000, 5000, 112000,
670 4000, 3000, 4000, 110000,
671 3000, 4000, 5000, 112000,
672 4000,
673 };
674
675 static u64 intervals3[] __initdata = {
676 1385000, 212240000, 1240000,
677 1386000, 214415000, 1236000,
678 1384000, 214276000, 1234000,
679 1386000, 214415000, 1236000,
680 1385000, 212240000, 1240000,
681 1386000, 214415000, 1236000,
682 1384000, 214276000, 1234000,
683 1386000, 214415000, 1236000,
684 1385000, 212240000, 1240000,
685 };
686
687 static u64 intervals4[] __initdata = {
688 10000, 50000, 10000, 50000,
689 10000, 50000, 10000, 50000,
690 10000, 50000, 10000, 50000,
691 10000, 50000, 10000, 50000,
692 10000, 50000, 10000, 50000,
693 10000, 50000, 10000, 50000,
694 10000, 50000, 10000, 50000,
695 10000, 50000, 10000, 50000,
696 10000,
697 };
698
699 static struct timings_intervals tis[] __initdata = {
700 { intervals0, ARRAY_SIZE(intervals0) },
701 { intervals1, ARRAY_SIZE(intervals1) },
702 { intervals2, ARRAY_SIZE(intervals2) },
703 { intervals3, ARRAY_SIZE(intervals3) },
704 { intervals4, ARRAY_SIZE(intervals4) },
705 };
706
707 static int __init irq_timings_test_next_index(struct timings_intervals *ti)
708 {
709 int _buffer[IRQ_TIMINGS_SIZE];
710 int buffer[IRQ_TIMINGS_SIZE];
711 int index, start, i, count, period_max;
712
713 count = ti->count - 1;
714
715 period_max = count > (3 * PREDICTION_PERIOD_MAX) ?
716 PREDICTION_PERIOD_MAX : count / 3;
717
718 /*
719 * Inject all values except the last one which will be used
720 * to compare with the next index result.
721 */
722 pr_debug("index suite: ");
723
724 for (i = 0; i < count; i++) {
725 index = irq_timings_interval_index(ti->intervals[i]);
726 _buffer[i & IRQ_TIMINGS_MASK] = index;
727 pr_cont("%d ", index);
728 }
729
730 start = count < IRQ_TIMINGS_SIZE ? 0 :
731 count & IRQ_TIMINGS_MASK;
732
733 count = min_t(int, count, IRQ_TIMINGS_SIZE);
734
735 for (i = 0; i < count; i++) {
736 int index = (start + i) & IRQ_TIMINGS_MASK;
737 buffer[i] = _buffer[index];
738 }
739
740 index = irq_timings_next_event_index(buffer, count, period_max);
741 i = irq_timings_interval_index(ti->intervals[ti->count - 1]);
742
743 if (index != i) {
744 pr_err("Expected (%d) and computed (%d) next indexes differ\n",
745 i, index);
746 return -EINVAL;
747 }
748
749 return 0;
750 }
751
752 static int __init irq_timings_next_index_selftest(void)
753 {
754 int i, ret;
755
756 for (i = 0; i < ARRAY_SIZE(tis); i++) {
757
758 pr_info("---> Injecting intervals number #%d (count=%zd)\n",
759 i, tis[i].count);
760
761 ret = irq_timings_test_next_index(&tis[i]);
762 if (ret)
763 break;
764 }
765
766 return ret;
767 }
768
769 static int __init irq_timings_test_irqs(struct timings_intervals *ti)
770 {
771 struct irqt_stat __percpu *s;
772 struct irqt_stat *irqs;
773 int i, index, ret, irq = 0xACE5;
774
775 ret = irq_timings_alloc(irq);
776 if (ret) {
777 pr_err("Failed to allocate irq timings\n");
778 return ret;
779 }
780
781 s = idr_find(&irqt_stats, irq);
782 if (!s) {
783 ret = -EIDRM;
784 goto out;
785 }
786
787 irqs = this_cpu_ptr(s);
788
789 for (i = 0; i < ti->count; i++) {
790
791 index = irq_timings_interval_index(ti->intervals[i]);
792 pr_debug("%d: interval=%llu ema_index=%d\n",
793 i, ti->intervals[i], index);
794
795 __irq_timings_store(irq, irqs, ti->intervals[i]);
796 if (irqs->circ_timings[i & IRQ_TIMINGS_MASK] != index) {
797 pr_err("Failed to store in the circular buffer\n");
798 goto out;
799 }
800 }
801
802 if (irqs->count != ti->count) {
803 pr_err("Count differs\n");
804 goto out;
805 }
806
807 ret = 0;
808 out:
809 irq_timings_free(irq);
810
811 return ret;
812 }
813
814 static int __init irq_timings_irqs_selftest(void)
815 {
816 int i, ret;
817
818 for (i = 0; i < ARRAY_SIZE(tis); i++) {
819 pr_info("---> Injecting intervals number #%d (count=%zd)\n",
820 i, tis[i].count);
821 ret = irq_timings_test_irqs(&tis[i]);
822 if (ret)
823 break;
824 }
825
826 return ret;
827 }
828
829 static int __init irq_timings_test_irqts(struct irq_timings *irqts,
830 unsigned count)
831 {
832 int start = count >= IRQ_TIMINGS_SIZE ? count - IRQ_TIMINGS_SIZE : 0;
833 int i, irq, oirq = 0xBEEF;
834 u64 ots = 0xDEAD, ts;
835
836 /*
837 * Fill the circular buffer by using the dedicated function.
838 */
839 for (i = 0; i < count; i++) {
840 pr_debug("%d: index=%d, ts=%llX irq=%X\n",
841 i, i & IRQ_TIMINGS_MASK, ots + i, oirq + i);
842
843 irq_timings_push(ots + i, oirq + i);
844 }
845
846 /*
847 * Compute the first elements values after the index wrapped
848 * up or not.
849 */
850 ots += start;
851 oirq += start;
852
853 /*
854 * Test the circular buffer count is correct.
855 */
856 pr_debug("---> Checking timings array count (%d) is right\n", count);
857 if (WARN_ON(irqts->count != count))
858 return -EINVAL;
859
860 /*
861 * Test the macro allowing to browse all the irqts.
862 */
863 pr_debug("---> Checking the for_each_irqts() macro\n");
864 for_each_irqts(i, irqts) {
865
866 irq = irq_timing_decode(irqts->values[i], &ts);
867
868 pr_debug("index=%d, ts=%llX / %llX, irq=%X / %X\n",
869 i, ts, ots, irq, oirq);
870
871 if (WARN_ON(ts != ots || irq != oirq))
872 return -EINVAL;
873
874 ots++; oirq++;
875 }
876
877 /*
878 * The circular buffer should have be flushed when browsed
879 * with for_each_irqts
880 */
881 pr_debug("---> Checking timings array is empty after browsing it\n");
882 if (WARN_ON(irqts->count))
883 return -EINVAL;
884
885 return 0;
886 }
887
888 static int __init irq_timings_irqts_selftest(void)
889 {
890 struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
891 int i, ret;
892
893 /*
894 * Test the circular buffer with different number of
895 * elements. The purpose is to test at the limits (empty, half
896 * full, full, wrapped with the cursor at the boundaries,
897 * wrapped several times, etc ...
898 */
899 int count[] = { 0,
900 IRQ_TIMINGS_SIZE >> 1,
901 IRQ_TIMINGS_SIZE,
902 IRQ_TIMINGS_SIZE + (IRQ_TIMINGS_SIZE >> 1),
903 2 * IRQ_TIMINGS_SIZE,
904 (2 * IRQ_TIMINGS_SIZE) + 3,
905 };
906
907 for (i = 0; i < ARRAY_SIZE(count); i++) {
908
909 pr_info("---> Checking the timings with %d/%d values\n",
910 count[i], IRQ_TIMINGS_SIZE);
911
912 ret = irq_timings_test_irqts(irqts, count[i]);
913 if (ret)
914 break;
915 }
916
917 return ret;
918 }
919
920 static int __init irq_timings_selftest(void)
921 {
922 int ret;
923
924 pr_info("------------------- selftest start -----------------\n");
925
926 /*
927 * At this point, we don't except any subsystem to use the irq
928 * timings but us, so it should not be enabled.
929 */
930 if (static_branch_unlikely(&irq_timing_enabled)) {
931 pr_warn("irq timings already initialized, skipping selftest\n");
932 return 0;
933 }
934
935 ret = irq_timings_irqts_selftest();
936 if (ret)
937 goto out;
938
939 ret = irq_timings_irqs_selftest();
940 if (ret)
941 goto out;
942
943 ret = irq_timings_next_index_selftest();
944 out:
945 pr_info("---------- selftest end with %s -----------\n",
946 ret ? "failure" : "success");
947
948 return ret;
949 }
950 early_initcall(irq_timings_selftest);
951 #endif