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 /* string_get_size units: */
31 static const char *const units_2[] = {
32 "B", "KiB", "MiB", "GiB", "TiB", "PiB", "EiB", "ZiB", "YiB"
34 static const char *const units_10[] = {
35 "B", "kB", "MB", "GB", "TB", "PB", "EB", "ZB", "YB"
38 static int parse_u64(const char *cp, u64 *res)
40 const char *start = cp;
50 if (v > U64_MAX - (*cp - '0'))
54 } while (isdigit(*cp));
60 static int bch2_pow(u64 n, u64 p, u64 *res)
65 if (*res > div_u64(U64_MAX, n))
72 static int parse_unit_suffix(const char *cp, u64 *res)
74 const char *start = cp;
82 for (u = 1; u < strlen(si_units); u++)
83 if (*cp == si_units[u]) {
88 for (u = 0; u < ARRAY_SIZE(units_2); u++)
89 if (!strncmp(cp, units_2[u], strlen(units_2[u]))) {
90 cp += strlen(units_2[u]);
94 for (u = 0; u < ARRAY_SIZE(units_10); u++)
95 if (!strncmp(cp, units_10[u], strlen(units_10[u]))) {
96 cp += strlen(units_10[u]);
104 ret = bch2_pow(base, u, res);
111 #define parse_or_ret(cp, _f) \
119 static int __bch2_strtou64_h(const char *cp, u64 *res)
121 const char *start = cp;
122 u64 v = 0, b, f_n = 0, f_d = 1;
125 parse_or_ret(cp, parse_u64(cp, &v));
129 ret = parse_u64(cp, &f_n);
134 ret = bch2_pow(10, ret, &f_d);
139 parse_or_ret(cp, parse_unit_suffix(cp, &b));
141 if (v > div_u64(U64_MAX, b))
145 if (f_n > div_u64(U64_MAX, b))
148 f_n = div_u64(f_n * b, f_d);
157 static int __bch2_strtoh(const char *cp, u64 *res,
158 u64 t_max, bool t_signed)
160 bool positive = *cp != '-';
163 if (*cp == '+' || *cp == '-')
166 parse_or_ret(cp, __bch2_strtou64_h(cp, &v));
189 #define STRTO_H(name, type) \
190 int bch2_ ## name ## _h(const char *cp, type *res) \
193 int ret = __bch2_strtoh(cp, &v, ANYSINT_MAX(type), \
194 ANYSINT_MAX(type) != ((type) ~0ULL)); \
199 STRTO_H(strtoint, int)
200 STRTO_H(strtouint, unsigned int)
201 STRTO_H(strtoll, long long)
202 STRTO_H(strtoull, unsigned long long)
203 STRTO_H(strtou64, u64)
205 u64 bch2_read_flag_list(char *opt, const char * const list[])
208 char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
215 while ((p = strsep(&s, ","))) {
216 int flag = match_string(list, -1, p);
230 bool bch2_is_zero(const void *_p, size_t n)
235 for (i = 0; i < n; i++)
241 static void bch2_quantiles_update(struct quantiles *q, u64 v)
245 while (i < ARRAY_SIZE(q->entries)) {
246 struct quantile_entry *e = q->entries + i;
248 if (unlikely(!e->step)) {
250 e->step = max_t(unsigned, v / 2, 1024);
251 } else if (e->m > v) {
252 e->m = e->m >= e->step
255 } else if (e->m < v) {
256 e->m = e->m + e->step > e->m
261 if ((e->m > v ? e->m - v : v - e->m) < e->step)
262 e->step = max_t(unsigned, e->step / 2, 1);
267 i = eytzinger0_child(i, v > e->m);
271 void bch2_prt_u64_binary(struct printbuf *out, u64 v, unsigned nr_bits)
274 prt_char(out, '0' + ((v >> --nr_bits) & 1));
279 static void bch2_time_stats_update_one(struct time_stats *stats,
284 duration = time_after64(end, start)
286 freq = time_after64(end, stats->last_event)
287 ? end - stats->last_event : 0;
291 stats->average_duration = stats->average_duration
292 ? ewma_add(stats->average_duration, duration, 6)
295 stats->average_frequency = stats->average_frequency
296 ? ewma_add(stats->average_frequency, freq, 6)
299 stats->max_duration = max(stats->max_duration, duration);
301 stats->last_event = end;
303 bch2_quantiles_update(&stats->quantiles, duration);
306 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
310 if (!stats->buffer) {
311 spin_lock_irqsave(&stats->lock, flags);
312 bch2_time_stats_update_one(stats, start, end);
314 if (stats->average_frequency < 32 &&
317 alloc_percpu_gfp(struct time_stat_buffer,
319 spin_unlock_irqrestore(&stats->lock, flags);
321 struct time_stat_buffer_entry *i;
322 struct time_stat_buffer *b;
325 b = this_cpu_ptr(stats->buffer);
327 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
328 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
333 if (b->nr == ARRAY_SIZE(b->entries)) {
334 spin_lock_irqsave(&stats->lock, flags);
336 i < b->entries + ARRAY_SIZE(b->entries);
338 bch2_time_stats_update_one(stats, i->start, i->end);
339 spin_unlock_irqrestore(&stats->lock, flags);
348 static const struct time_unit {
353 { "us", NSEC_PER_USEC },
354 { "ms", NSEC_PER_MSEC },
355 { "sec", NSEC_PER_SEC },
358 static const struct time_unit *pick_time_units(u64 ns)
360 const struct time_unit *u;
363 u + 1 < time_units + ARRAY_SIZE(time_units) &&
364 ns >= u[1].nsecs << 1;
371 static void pr_time_units(struct printbuf *out, u64 ns)
373 const struct time_unit *u = pick_time_units(ns);
375 prt_printf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
378 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
380 const struct time_unit *u;
381 u64 freq = READ_ONCE(stats->average_frequency);
385 prt_printf(out, "count:\t\t%llu",
388 prt_printf(out, "rate:\t\t%llu/sec",
389 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
392 prt_printf(out, "frequency:\t");
393 pr_time_units(out, freq);
396 prt_printf(out, "avg duration:\t");
397 pr_time_units(out, stats->average_duration);
400 prt_printf(out, "max duration:\t");
401 pr_time_units(out, stats->max_duration);
403 i = eytzinger0_first(NR_QUANTILES);
404 u = pick_time_units(stats->quantiles.entries[i].m);
407 prt_printf(out, "quantiles (%s):\t", u->name);
408 eytzinger0_for_each(i, NR_QUANTILES) {
409 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
411 q = max(stats->quantiles.entries[i].m, last_q);
412 prt_printf(out, "%llu ",
413 div_u64(q, u->nsecs));
420 void bch2_time_stats_exit(struct time_stats *stats)
422 free_percpu(stats->buffer);
425 void bch2_time_stats_init(struct time_stats *stats)
427 memset(stats, 0, sizeof(*stats));
428 spin_lock_init(&stats->lock);
434 * bch2_ratelimit_delay() - return how long to delay until the next time to do
437 * @d - the struct bch_ratelimit to update
439 * Returns the amount of time to delay by, in jiffies
441 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
443 u64 now = local_clock();
445 return time_after64(d->next, now)
446 ? nsecs_to_jiffies(d->next - now)
451 * bch2_ratelimit_increment() - increment @d by the amount of work done
453 * @d - the struct bch_ratelimit to update
454 * @done - the amount of work done, in arbitrary units
456 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
458 u64 now = local_clock();
460 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
462 if (time_before64(now + NSEC_PER_SEC, d->next))
463 d->next = now + NSEC_PER_SEC;
465 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
466 d->next = now - NSEC_PER_SEC * 2;
472 * Updates pd_controller. Attempts to scale inputed values to units per second.
473 * @target: desired value
474 * @actual: current value
476 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
477 * it makes actual go down.
479 void bch2_pd_controller_update(struct bch_pd_controller *pd,
480 s64 target, s64 actual, int sign)
482 s64 proportional, derivative, change;
484 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
486 if (seconds_since_update == 0)
489 pd->last_update = jiffies;
491 proportional = actual - target;
492 proportional *= seconds_since_update;
493 proportional = div_s64(proportional, pd->p_term_inverse);
495 derivative = actual - pd->last_actual;
496 derivative = div_s64(derivative, seconds_since_update);
497 derivative = ewma_add(pd->smoothed_derivative, derivative,
498 (pd->d_term / seconds_since_update) ?: 1);
499 derivative = derivative * pd->d_term;
500 derivative = div_s64(derivative, pd->p_term_inverse);
502 change = proportional + derivative;
504 /* Don't increase rate if not keeping up */
507 time_after64(local_clock(),
508 pd->rate.next + NSEC_PER_MSEC))
511 change *= (sign * -1);
513 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
516 pd->last_actual = actual;
517 pd->last_derivative = derivative;
518 pd->last_proportional = proportional;
519 pd->last_change = change;
520 pd->last_target = target;
523 void bch2_pd_controller_init(struct bch_pd_controller *pd)
525 pd->rate.rate = 1024;
526 pd->last_update = jiffies;
527 pd->p_term_inverse = 6000;
529 pd->d_smooth = pd->d_term;
530 pd->backpressure = 1;
533 void bch2_pd_controller_debug_to_text(struct printbuf *out, struct bch_pd_controller *pd)
535 if (!out->nr_tabstops)
536 printbuf_tabstop_push(out, 20);
538 prt_printf(out, "rate:");
540 prt_human_readable_s64(out, pd->rate.rate);
543 prt_printf(out, "target:");
545 prt_human_readable_u64(out, pd->last_target);
548 prt_printf(out, "actual:");
550 prt_human_readable_u64(out, pd->last_actual);
553 prt_printf(out, "proportional:");
555 prt_human_readable_s64(out, pd->last_proportional);
558 prt_printf(out, "derivative:");
560 prt_human_readable_s64(out, pd->last_derivative);
563 prt_printf(out, "change:");
565 prt_human_readable_s64(out, pd->last_change);
568 prt_printf(out, "next io:");
570 prt_printf(out, "%llims", div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC));
576 void bch2_bio_map(struct bio *bio, void *base, size_t size)
579 struct page *page = is_vmalloc_addr(base)
580 ? vmalloc_to_page(base)
581 : virt_to_page(base);
582 unsigned offset = offset_in_page(base);
583 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
585 BUG_ON(!bio_add_page(bio, page, len, offset));
591 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
594 struct page *page = alloc_page(gfp_mask);
595 unsigned len = min_t(size_t, PAGE_SIZE, size);
600 if (unlikely(!bio_add_page(bio, page, len, 0))) {
611 size_t bch2_rand_range(size_t max)
619 rand = get_random_long();
620 rand &= roundup_pow_of_two(max) - 1;
621 } while (rand >= max);
626 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
629 struct bvec_iter iter;
631 __bio_for_each_segment(bv, dst, iter, dst_iter) {
632 void *dstp = kmap_atomic(bv.bv_page);
633 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
640 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
643 struct bvec_iter iter;
645 __bio_for_each_segment(bv, src, iter, src_iter) {
646 void *srcp = kmap_atomic(bv.bv_page);
647 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
654 #include "eytzinger.h"
656 static int alignment_ok(const void *base, size_t align)
658 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
659 ((unsigned long)base & (align - 1)) == 0;
662 static void u32_swap(void *a, void *b, size_t size)
665 *(u32 *)a = *(u32 *)b;
669 static void u64_swap(void *a, void *b, size_t size)
672 *(u64 *)a = *(u64 *)b;
676 static void generic_swap(void *a, void *b, size_t size)
682 *(char *)a++ = *(char *)b;
684 } while (--size > 0);
687 static inline int do_cmp(void *base, size_t n, size_t size,
688 int (*cmp_func)(const void *, const void *, size_t),
691 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
692 base + inorder_to_eytzinger0(r, n) * size,
696 static inline void do_swap(void *base, size_t n, size_t size,
697 void (*swap_func)(void *, void *, size_t),
700 swap_func(base + inorder_to_eytzinger0(l, n) * size,
701 base + inorder_to_eytzinger0(r, n) * size,
705 void eytzinger0_sort(void *base, size_t n, size_t size,
706 int (*cmp_func)(const void *, const void *, size_t),
707 void (*swap_func)(void *, void *, size_t))
712 if (size == 4 && alignment_ok(base, 4))
713 swap_func = u32_swap;
714 else if (size == 8 && alignment_ok(base, 8))
715 swap_func = u64_swap;
717 swap_func = generic_swap;
721 for (i = n / 2 - 1; i >= 0; --i) {
722 for (r = i; r * 2 + 1 < n; r = c) {
726 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
729 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
732 do_swap(base, n, size, swap_func, r, c);
737 for (i = n - 1; i > 0; --i) {
738 do_swap(base, n, size, swap_func, 0, i);
740 for (r = 0; r * 2 + 1 < i; r = c) {
744 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
747 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
750 do_swap(base, n, size, swap_func, r, c);
755 void sort_cmp_size(void *base, size_t num, size_t size,
756 int (*cmp_func)(const void *, const void *, size_t),
757 void (*swap_func)(void *, void *, size_t size))
759 /* pre-scale counters for performance */
760 int i = (num/2 - 1) * size, n = num * size, c, r;
763 if (size == 4 && alignment_ok(base, 4))
764 swap_func = u32_swap;
765 else if (size == 8 && alignment_ok(base, 8))
766 swap_func = u64_swap;
768 swap_func = generic_swap;
772 for ( ; i >= 0; i -= size) {
773 for (r = i; r * 2 + size < n; r = c) {
776 cmp_func(base + c, base + c + size, size) < 0)
778 if (cmp_func(base + r, base + c, size) >= 0)
780 swap_func(base + r, base + c, size);
785 for (i = n - size; i > 0; i -= size) {
786 swap_func(base, base + i, size);
787 for (r = 0; r * 2 + size < i; r = c) {
790 cmp_func(base + c, base + c + size, size) < 0)
792 if (cmp_func(base + r, base + c, size) >= 0)
794 swap_func(base + r, base + c, size);
799 static void mempool_free_vp(void *element, void *pool_data)
801 size_t size = (size_t) pool_data;
803 vpfree(element, size);
806 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
808 size_t size = (size_t) pool_data;
810 return vpmalloc(size, gfp_mask);
813 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
815 return size < PAGE_SIZE
816 ? mempool_init_kmalloc_pool(pool, min_nr, size)
817 : mempool_init(pool, min_nr, mempool_alloc_vp,
818 mempool_free_vp, (void *) size);
822 void eytzinger1_test(void)
824 unsigned inorder, eytz, size;
826 pr_info("1 based eytzinger test:");
831 unsigned extra = eytzinger1_extra(size);
834 pr_info("tree size %u", size);
836 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
837 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
839 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
840 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
843 eytzinger1_for_each(eytz, size) {
844 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
845 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
846 BUG_ON(eytz != eytzinger1_last(size) &&
847 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
854 void eytzinger0_test(void)
857 unsigned inorder, eytz, size;
859 pr_info("0 based eytzinger test:");
864 unsigned extra = eytzinger0_extra(size);
867 pr_info("tree size %u", size);
869 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
870 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
872 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
873 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
876 eytzinger0_for_each(eytz, size) {
877 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
878 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
879 BUG_ON(eytz != eytzinger0_last(size) &&
880 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
887 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
889 const u16 *l = _l, *r = _r;
891 return (*l > *r) - (*r - *l);
894 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
896 int i, c1 = -1, c2 = -1;
899 r = eytzinger0_find_le(test_array, nr,
900 sizeof(test_array[0]),
905 for (i = 0; i < nr; i++)
906 if (test_array[i] <= search && test_array[i] > c2)
910 eytzinger0_for_each(i, nr)
911 pr_info("[%3u] = %12u", i, test_array[i]);
912 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
917 void eytzinger0_find_test(void)
919 unsigned i, nr, allocated = 1 << 12;
920 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
922 for (nr = 1; nr < allocated; nr++) {
923 pr_info("testing %u elems", nr);
925 get_random_bytes(test_array, nr * sizeof(test_array[0]));
926 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
928 /* verify array is sorted correctly: */
929 eytzinger0_for_each(i, nr)
930 BUG_ON(i != eytzinger0_last(nr) &&
931 test_array[i] > test_array[eytzinger0_next(i, nr)]);
933 for (i = 0; i < U16_MAX; i += 1 << 12)
934 eytzinger0_find_test_val(test_array, nr, i);
936 for (i = 0; i < nr; i++) {
937 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
938 eytzinger0_find_test_val(test_array, nr, test_array[i]);
939 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
948 * Accumulate percpu counters onto one cpu's copy - only valid when access
949 * against any percpu counter is guarded against
951 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
956 /* access to pcpu vars has to be blocked by other locking */
958 ret = this_cpu_ptr(p);
961 for_each_possible_cpu(cpu) {
962 u64 *i = per_cpu_ptr(p, cpu);
965 acc_u64s(ret, i, nr);
966 memset(i, 0, nr * sizeof(u64));