1 // SPDX-License-Identifier: GPL-2.0
3 * random utiility code, for bcache but in theory not specific to bcache
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
10 #include <linux/blkdev.h>
11 #include <linux/ctype.h>
12 #include <linux/debugfs.h>
13 #include <linux/freezer.h>
14 #include <linux/kthread.h>
15 #include <linux/log2.h>
16 #include <linux/math64.h>
17 #include <linux/percpu.h>
18 #include <linux/preempt.h>
19 #include <linux/random.h>
20 #include <linux/seq_file.h>
21 #include <linux/string.h>
22 #include <linux/types.h>
23 #include <linux/sched/clock.h>
25 #include "eytzinger.h"
28 static const char si_units[] = "?kMGTPEZY";
30 static int __bch2_strtoh(const char *cp, u64 *res,
31 u64 t_max, bool t_signed)
33 bool positive = *cp != '-';
37 if (*cp == '+' || *cp == '-')
47 if (v > U64_MAX - (*cp - '0'))
51 } while (isdigit(*cp));
53 for (u = 1; u < strlen(si_units); u++)
54 if (*cp == si_units[u]) {
65 if (fls64(v) + u * 10 > 64)
86 #define STRTO_H(name, type) \
87 int bch2_ ## name ## _h(const char *cp, type *res) \
90 int ret = __bch2_strtoh(cp, &v, ANYSINT_MAX(type), \
91 ANYSINT_MAX(type) != ((type) ~0ULL)); \
96 STRTO_H(strtoint, int)
97 STRTO_H(strtouint, unsigned int)
98 STRTO_H(strtoll, long long)
99 STRTO_H(strtoull, unsigned long long)
100 STRTO_H(strtou64, u64)
102 void bch2_hprint(struct printbuf *buf, s64 v)
106 for (u = 0; v >= 1024 || v <= -1024; u++) {
107 t = v & ~(~0U << 10);
111 pr_buf(buf, "%lli", v);
114 * 103 is magic: t is in the range [-1023, 1023] and we want
115 * to turn it into [-9, 9]
117 if (u && t && v < 100 && v > -100)
118 pr_buf(buf, ".%i", t / 103);
120 pr_char(buf, si_units[u]);
123 void bch2_pr_units(struct printbuf *out, s64 raw, s64 bytes)
125 switch (out->units) {
126 case PRINTBUF_UNITS_RAW:
127 pr_buf(out, "%llu", raw);
129 case PRINTBUF_UNITS_BYTES:
130 pr_buf(out, "%llu", bytes);
132 case PRINTBUF_UNITS_HUMAN_READABLE:
133 bch2_hprint(out, bytes);
138 void bch2_string_opt_to_text(struct printbuf *out,
139 const char * const list[],
144 for (i = 0; list[i]; i++)
145 pr_buf(out, i == selected ? "[%s] " : "%s ", list[i]);
148 void bch2_flags_to_text(struct printbuf *out,
149 const char * const list[], u64 flags)
151 unsigned bit, nr = 0;
154 if (out->pos != out->end)
160 while (flags && (bit = __ffs(flags)) < nr) {
164 pr_buf(out, "%s", list[bit]);
169 u64 bch2_read_flag_list(char *opt, const char * const list[])
172 char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
179 while ((p = strsep(&s, ","))) {
180 int flag = match_string(list, -1, p);
194 bool bch2_is_zero(const void *_p, size_t n)
199 for (i = 0; i < n; i++)
205 static void bch2_quantiles_update(struct quantiles *q, u64 v)
209 while (i < ARRAY_SIZE(q->entries)) {
210 struct quantile_entry *e = q->entries + i;
212 if (unlikely(!e->step)) {
214 e->step = max_t(unsigned, v / 2, 1024);
215 } else if (e->m > v) {
216 e->m = e->m >= e->step
219 } else if (e->m < v) {
220 e->m = e->m + e->step > e->m
225 if ((e->m > v ? e->m - v : v - e->m) < e->step)
226 e->step = max_t(unsigned, e->step / 2, 1);
231 i = eytzinger0_child(i, v > e->m);
237 static void bch2_time_stats_update_one(struct time_stats *stats,
242 duration = time_after64(end, start)
244 freq = time_after64(end, stats->last_event)
245 ? end - stats->last_event : 0;
249 stats->average_duration = stats->average_duration
250 ? ewma_add(stats->average_duration, duration, 6)
253 stats->average_frequency = stats->average_frequency
254 ? ewma_add(stats->average_frequency, freq, 6)
257 stats->max_duration = max(stats->max_duration, duration);
259 stats->last_event = end;
261 bch2_quantiles_update(&stats->quantiles, duration);
264 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
268 if (!stats->buffer) {
269 spin_lock_irqsave(&stats->lock, flags);
270 bch2_time_stats_update_one(stats, start, end);
272 if (stats->average_frequency < 32 &&
275 alloc_percpu_gfp(struct time_stat_buffer,
277 spin_unlock_irqrestore(&stats->lock, flags);
279 struct time_stat_buffer_entry *i;
280 struct time_stat_buffer *b;
283 b = this_cpu_ptr(stats->buffer);
285 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
286 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
291 if (b->nr == ARRAY_SIZE(b->entries)) {
292 spin_lock_irqsave(&stats->lock, flags);
294 i < b->entries + ARRAY_SIZE(b->entries);
296 bch2_time_stats_update_one(stats, i->start, i->end);
297 spin_unlock_irqrestore(&stats->lock, flags);
306 static const struct time_unit {
311 { "us", NSEC_PER_USEC },
312 { "ms", NSEC_PER_MSEC },
313 { "sec", NSEC_PER_SEC },
316 static const struct time_unit *pick_time_units(u64 ns)
318 const struct time_unit *u;
321 u + 1 < time_units + ARRAY_SIZE(time_units) &&
322 ns >= u[1].nsecs << 1;
329 static void pr_time_units(struct printbuf *out, u64 ns)
331 const struct time_unit *u = pick_time_units(ns);
333 pr_buf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
336 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
338 const struct time_unit *u;
339 u64 freq = READ_ONCE(stats->average_frequency);
343 pr_buf(out, "count:\t\t%llu\n",
345 pr_buf(out, "rate:\t\t%llu/sec\n",
346 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
348 pr_buf(out, "frequency:\t");
349 pr_time_units(out, freq);
351 pr_buf(out, "\navg duration:\t");
352 pr_time_units(out, stats->average_duration);
354 pr_buf(out, "\nmax duration:\t");
355 pr_time_units(out, stats->max_duration);
357 i = eytzinger0_first(NR_QUANTILES);
358 u = pick_time_units(stats->quantiles.entries[i].m);
360 pr_buf(out, "\nquantiles (%s):\t", u->name);
361 eytzinger0_for_each(i, NR_QUANTILES) {
362 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
364 q = max(stats->quantiles.entries[i].m, last_q);
365 pr_buf(out, "%llu%s",
366 div_u64(q, u->nsecs),
367 is_last ? "\n" : " ");
372 void bch2_time_stats_exit(struct time_stats *stats)
374 free_percpu(stats->buffer);
377 void bch2_time_stats_init(struct time_stats *stats)
379 memset(stats, 0, sizeof(*stats));
380 spin_lock_init(&stats->lock);
386 * bch2_ratelimit_delay() - return how long to delay until the next time to do
389 * @d - the struct bch_ratelimit to update
391 * Returns the amount of time to delay by, in jiffies
393 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
395 u64 now = local_clock();
397 return time_after64(d->next, now)
398 ? nsecs_to_jiffies(d->next - now)
403 * bch2_ratelimit_increment() - increment @d by the amount of work done
405 * @d - the struct bch_ratelimit to update
406 * @done - the amount of work done, in arbitrary units
408 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
410 u64 now = local_clock();
412 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
414 if (time_before64(now + NSEC_PER_SEC, d->next))
415 d->next = now + NSEC_PER_SEC;
417 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
418 d->next = now - NSEC_PER_SEC * 2;
424 * Updates pd_controller. Attempts to scale inputed values to units per second.
425 * @target: desired value
426 * @actual: current value
428 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
429 * it makes actual go down.
431 void bch2_pd_controller_update(struct bch_pd_controller *pd,
432 s64 target, s64 actual, int sign)
434 s64 proportional, derivative, change;
436 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
438 if (seconds_since_update == 0)
441 pd->last_update = jiffies;
443 proportional = actual - target;
444 proportional *= seconds_since_update;
445 proportional = div_s64(proportional, pd->p_term_inverse);
447 derivative = actual - pd->last_actual;
448 derivative = div_s64(derivative, seconds_since_update);
449 derivative = ewma_add(pd->smoothed_derivative, derivative,
450 (pd->d_term / seconds_since_update) ?: 1);
451 derivative = derivative * pd->d_term;
452 derivative = div_s64(derivative, pd->p_term_inverse);
454 change = proportional + derivative;
456 /* Don't increase rate if not keeping up */
459 time_after64(local_clock(),
460 pd->rate.next + NSEC_PER_MSEC))
463 change *= (sign * -1);
465 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
468 pd->last_actual = actual;
469 pd->last_derivative = derivative;
470 pd->last_proportional = proportional;
471 pd->last_change = change;
472 pd->last_target = target;
475 void bch2_pd_controller_init(struct bch_pd_controller *pd)
477 pd->rate.rate = 1024;
478 pd->last_update = jiffies;
479 pd->p_term_inverse = 6000;
481 pd->d_smooth = pd->d_term;
482 pd->backpressure = 1;
485 size_t bch2_pd_controller_print_debug(struct bch_pd_controller *pd, char *buf)
487 /* 2^64 - 1 is 20 digits, plus null byte */
491 char proportional[21];
496 bch2_hprint(&PBUF(rate), pd->rate.rate);
497 bch2_hprint(&PBUF(actual), pd->last_actual);
498 bch2_hprint(&PBUF(target), pd->last_target);
499 bch2_hprint(&PBUF(proportional), pd->last_proportional);
500 bch2_hprint(&PBUF(derivative), pd->last_derivative);
501 bch2_hprint(&PBUF(change), pd->last_change);
503 next_io = div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC);
509 "proportional:\t%s\n"
511 "change:\t\t%s/sec\n"
512 "next io:\t%llims\n",
513 rate, target, actual, proportional,
514 derivative, change, next_io);
519 void bch2_bio_map(struct bio *bio, void *base, size_t size)
522 struct page *page = is_vmalloc_addr(base)
523 ? vmalloc_to_page(base)
524 : virt_to_page(base);
525 unsigned offset = offset_in_page(base);
526 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
528 BUG_ON(!bio_add_page(bio, page, len, offset));
534 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
537 struct page *page = alloc_page(gfp_mask);
538 unsigned len = min_t(size_t, PAGE_SIZE, size);
543 if (unlikely(!bio_add_page(bio, page, len, 0))) {
554 size_t bch2_rand_range(size_t max)
562 rand = get_random_long();
563 rand &= roundup_pow_of_two(max) - 1;
564 } while (rand >= max);
569 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
572 struct bvec_iter iter;
574 __bio_for_each_segment(bv, dst, iter, dst_iter) {
575 void *dstp = kmap_atomic(bv.bv_page);
576 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
583 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
586 struct bvec_iter iter;
588 __bio_for_each_segment(bv, src, iter, src_iter) {
589 void *srcp = kmap_atomic(bv.bv_page);
590 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
597 #include "eytzinger.h"
599 static int alignment_ok(const void *base, size_t align)
601 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
602 ((unsigned long)base & (align - 1)) == 0;
605 static void u32_swap(void *a, void *b, size_t size)
608 *(u32 *)a = *(u32 *)b;
612 static void u64_swap(void *a, void *b, size_t size)
615 *(u64 *)a = *(u64 *)b;
619 static void generic_swap(void *a, void *b, size_t size)
625 *(char *)a++ = *(char *)b;
627 } while (--size > 0);
630 static inline int do_cmp(void *base, size_t n, size_t size,
631 int (*cmp_func)(const void *, const void *, size_t),
634 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
635 base + inorder_to_eytzinger0(r, n) * size,
639 static inline void do_swap(void *base, size_t n, size_t size,
640 void (*swap_func)(void *, void *, size_t),
643 swap_func(base + inorder_to_eytzinger0(l, n) * size,
644 base + inorder_to_eytzinger0(r, n) * size,
648 void eytzinger0_sort(void *base, size_t n, size_t size,
649 int (*cmp_func)(const void *, const void *, size_t),
650 void (*swap_func)(void *, void *, size_t))
655 if (size == 4 && alignment_ok(base, 4))
656 swap_func = u32_swap;
657 else if (size == 8 && alignment_ok(base, 8))
658 swap_func = u64_swap;
660 swap_func = generic_swap;
664 for (i = n / 2 - 1; i >= 0; --i) {
665 for (r = i; r * 2 + 1 < n; r = c) {
669 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
672 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
675 do_swap(base, n, size, swap_func, r, c);
680 for (i = n - 1; i > 0; --i) {
681 do_swap(base, n, size, swap_func, 0, i);
683 for (r = 0; r * 2 + 1 < i; r = c) {
687 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
690 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
693 do_swap(base, n, size, swap_func, r, c);
698 void sort_cmp_size(void *base, size_t num, size_t size,
699 int (*cmp_func)(const void *, const void *, size_t),
700 void (*swap_func)(void *, void *, size_t size))
702 /* pre-scale counters for performance */
703 int i = (num/2 - 1) * size, n = num * size, c, r;
706 if (size == 4 && alignment_ok(base, 4))
707 swap_func = u32_swap;
708 else if (size == 8 && alignment_ok(base, 8))
709 swap_func = u64_swap;
711 swap_func = generic_swap;
715 for ( ; i >= 0; i -= size) {
716 for (r = i; r * 2 + size < n; r = c) {
719 cmp_func(base + c, base + c + size, size) < 0)
721 if (cmp_func(base + r, base + c, size) >= 0)
723 swap_func(base + r, base + c, size);
728 for (i = n - size; i > 0; i -= size) {
729 swap_func(base, base + i, size);
730 for (r = 0; r * 2 + size < i; r = c) {
733 cmp_func(base + c, base + c + size, size) < 0)
735 if (cmp_func(base + r, base + c, size) >= 0)
737 swap_func(base + r, base + c, size);
742 static void mempool_free_vp(void *element, void *pool_data)
744 size_t size = (size_t) pool_data;
746 vpfree(element, size);
749 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
751 size_t size = (size_t) pool_data;
753 return vpmalloc(size, gfp_mask);
756 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
758 return size < PAGE_SIZE
759 ? mempool_init_kmalloc_pool(pool, min_nr, size)
760 : mempool_init(pool, min_nr, mempool_alloc_vp,
761 mempool_free_vp, (void *) size);
765 void eytzinger1_test(void)
767 unsigned inorder, eytz, size;
769 pr_info("1 based eytzinger test:");
774 unsigned extra = eytzinger1_extra(size);
777 pr_info("tree size %u", size);
779 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
780 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
782 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
783 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
786 eytzinger1_for_each(eytz, size) {
787 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
788 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
789 BUG_ON(eytz != eytzinger1_last(size) &&
790 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
797 void eytzinger0_test(void)
800 unsigned inorder, eytz, size;
802 pr_info("0 based eytzinger test:");
807 unsigned extra = eytzinger0_extra(size);
810 pr_info("tree size %u", size);
812 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
813 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
815 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
816 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
819 eytzinger0_for_each(eytz, size) {
820 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
821 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
822 BUG_ON(eytz != eytzinger0_last(size) &&
823 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
830 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
832 const u16 *l = _l, *r = _r;
834 return (*l > *r) - (*r - *l);
837 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
839 int i, c1 = -1, c2 = -1;
842 r = eytzinger0_find_le(test_array, nr,
843 sizeof(test_array[0]),
848 for (i = 0; i < nr; i++)
849 if (test_array[i] <= search && test_array[i] > c2)
853 eytzinger0_for_each(i, nr)
854 pr_info("[%3u] = %12u", i, test_array[i]);
855 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
860 void eytzinger0_find_test(void)
862 unsigned i, nr, allocated = 1 << 12;
863 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
865 for (nr = 1; nr < allocated; nr++) {
866 pr_info("testing %u elems", nr);
868 get_random_bytes(test_array, nr * sizeof(test_array[0]));
869 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
871 /* verify array is sorted correctly: */
872 eytzinger0_for_each(i, nr)
873 BUG_ON(i != eytzinger0_last(nr) &&
874 test_array[i] > test_array[eytzinger0_next(i, nr)]);
876 for (i = 0; i < U16_MAX; i += 1 << 12)
877 eytzinger0_find_test_val(test_array, nr, i);
879 for (i = 0; i < nr; i++) {
880 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
881 eytzinger0_find_test_val(test_array, nr, test_array[i]);
882 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
891 * Accumulate percpu counters onto one cpu's copy - only valid when access
892 * against any percpu counter is guarded against
894 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
899 /* access to pcpu vars has to be blocked by other locking */
901 ret = this_cpu_ptr(p);
904 for_each_possible_cpu(cpu) {
905 u64 *i = per_cpu_ptr(p, cpu);
908 acc_u64s(ret, i, nr);
909 memset(i, 0, nr * sizeof(u64));