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 if (v + (f_n * b) / f_d < v)
150 v += (f_n * b) / f_d;
156 static int __bch2_strtoh(const char *cp, u64 *res,
157 u64 t_max, bool t_signed)
159 bool positive = *cp != '-';
162 if (*cp == '+' || *cp == '-')
165 parse_or_ret(cp, __bch2_strtou64_h(cp, &v));
188 #define STRTO_H(name, type) \
189 int bch2_ ## name ## _h(const char *cp, type *res) \
192 int ret = __bch2_strtoh(cp, &v, ANYSINT_MAX(type), \
193 ANYSINT_MAX(type) != ((type) ~0ULL)); \
198 STRTO_H(strtoint, int)
199 STRTO_H(strtouint, unsigned int)
200 STRTO_H(strtoll, long long)
201 STRTO_H(strtoull, unsigned long long)
202 STRTO_H(strtou64, u64)
204 u64 bch2_read_flag_list(char *opt, const char * const list[])
207 char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
214 while ((p = strsep(&s, ","))) {
215 int flag = match_string(list, -1, p);
229 bool bch2_is_zero(const void *_p, size_t n)
234 for (i = 0; i < n; i++)
240 static void bch2_quantiles_update(struct quantiles *q, u64 v)
244 while (i < ARRAY_SIZE(q->entries)) {
245 struct quantile_entry *e = q->entries + i;
247 if (unlikely(!e->step)) {
249 e->step = max_t(unsigned, v / 2, 1024);
250 } else if (e->m > v) {
251 e->m = e->m >= e->step
254 } else if (e->m < v) {
255 e->m = e->m + e->step > e->m
260 if ((e->m > v ? e->m - v : v - e->m) < e->step)
261 e->step = max_t(unsigned, e->step / 2, 1);
266 i = eytzinger0_child(i, v > e->m);
272 static void bch2_time_stats_update_one(struct time_stats *stats,
277 duration = time_after64(end, start)
279 freq = time_after64(end, stats->last_event)
280 ? end - stats->last_event : 0;
284 stats->average_duration = stats->average_duration
285 ? ewma_add(stats->average_duration, duration, 6)
288 stats->average_frequency = stats->average_frequency
289 ? ewma_add(stats->average_frequency, freq, 6)
292 stats->max_duration = max(stats->max_duration, duration);
294 stats->last_event = end;
296 bch2_quantiles_update(&stats->quantiles, duration);
299 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
303 if (!stats->buffer) {
304 spin_lock_irqsave(&stats->lock, flags);
305 bch2_time_stats_update_one(stats, start, end);
307 if (stats->average_frequency < 32 &&
310 alloc_percpu_gfp(struct time_stat_buffer,
312 spin_unlock_irqrestore(&stats->lock, flags);
314 struct time_stat_buffer_entry *i;
315 struct time_stat_buffer *b;
318 b = this_cpu_ptr(stats->buffer);
320 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
321 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
326 if (b->nr == ARRAY_SIZE(b->entries)) {
327 spin_lock_irqsave(&stats->lock, flags);
329 i < b->entries + ARRAY_SIZE(b->entries);
331 bch2_time_stats_update_one(stats, i->start, i->end);
332 spin_unlock_irqrestore(&stats->lock, flags);
341 static const struct time_unit {
346 { "us", NSEC_PER_USEC },
347 { "ms", NSEC_PER_MSEC },
348 { "sec", NSEC_PER_SEC },
351 static const struct time_unit *pick_time_units(u64 ns)
353 const struct time_unit *u;
356 u + 1 < time_units + ARRAY_SIZE(time_units) &&
357 ns >= u[1].nsecs << 1;
364 static void pr_time_units(struct printbuf *out, u64 ns)
366 const struct time_unit *u = pick_time_units(ns);
368 prt_printf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
371 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
373 const struct time_unit *u;
374 u64 freq = READ_ONCE(stats->average_frequency);
378 prt_printf(out, "count:\t\t%llu\n",
380 prt_printf(out, "rate:\t\t%llu/sec\n",
381 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
383 prt_printf(out, "frequency:\t");
384 pr_time_units(out, freq);
386 prt_printf(out, "\navg duration:\t");
387 pr_time_units(out, stats->average_duration);
389 prt_printf(out, "\nmax duration:\t");
390 pr_time_units(out, stats->max_duration);
392 i = eytzinger0_first(NR_QUANTILES);
393 u = pick_time_units(stats->quantiles.entries[i].m);
395 prt_printf(out, "\nquantiles (%s):\t", u->name);
396 eytzinger0_for_each(i, NR_QUANTILES) {
397 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
399 q = max(stats->quantiles.entries[i].m, last_q);
400 prt_printf(out, "%llu%s",
401 div_u64(q, u->nsecs),
402 is_last ? "\n" : " ");
407 void bch2_time_stats_exit(struct time_stats *stats)
409 free_percpu(stats->buffer);
412 void bch2_time_stats_init(struct time_stats *stats)
414 memset(stats, 0, sizeof(*stats));
415 spin_lock_init(&stats->lock);
421 * bch2_ratelimit_delay() - return how long to delay until the next time to do
424 * @d - the struct bch_ratelimit to update
426 * Returns the amount of time to delay by, in jiffies
428 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
430 u64 now = local_clock();
432 return time_after64(d->next, now)
433 ? nsecs_to_jiffies(d->next - now)
438 * bch2_ratelimit_increment() - increment @d by the amount of work done
440 * @d - the struct bch_ratelimit to update
441 * @done - the amount of work done, in arbitrary units
443 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
445 u64 now = local_clock();
447 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
449 if (time_before64(now + NSEC_PER_SEC, d->next))
450 d->next = now + NSEC_PER_SEC;
452 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
453 d->next = now - NSEC_PER_SEC * 2;
459 * Updates pd_controller. Attempts to scale inputed values to units per second.
460 * @target: desired value
461 * @actual: current value
463 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
464 * it makes actual go down.
466 void bch2_pd_controller_update(struct bch_pd_controller *pd,
467 s64 target, s64 actual, int sign)
469 s64 proportional, derivative, change;
471 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
473 if (seconds_since_update == 0)
476 pd->last_update = jiffies;
478 proportional = actual - target;
479 proportional *= seconds_since_update;
480 proportional = div_s64(proportional, pd->p_term_inverse);
482 derivative = actual - pd->last_actual;
483 derivative = div_s64(derivative, seconds_since_update);
484 derivative = ewma_add(pd->smoothed_derivative, derivative,
485 (pd->d_term / seconds_since_update) ?: 1);
486 derivative = derivative * pd->d_term;
487 derivative = div_s64(derivative, pd->p_term_inverse);
489 change = proportional + derivative;
491 /* Don't increase rate if not keeping up */
494 time_after64(local_clock(),
495 pd->rate.next + NSEC_PER_MSEC))
498 change *= (sign * -1);
500 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
503 pd->last_actual = actual;
504 pd->last_derivative = derivative;
505 pd->last_proportional = proportional;
506 pd->last_change = change;
507 pd->last_target = target;
510 void bch2_pd_controller_init(struct bch_pd_controller *pd)
512 pd->rate.rate = 1024;
513 pd->last_update = jiffies;
514 pd->p_term_inverse = 6000;
516 pd->d_smooth = pd->d_term;
517 pd->backpressure = 1;
520 void bch2_pd_controller_debug_to_text(struct printbuf *out, struct bch_pd_controller *pd)
522 out->tabstops[0] = 20;
524 prt_printf(out, "rate:");
526 prt_human_readable_s64(out, pd->rate.rate);
529 prt_printf(out, "target:");
531 prt_human_readable_u64(out, pd->last_target);
534 prt_printf(out, "actual:");
536 prt_human_readable_u64(out, pd->last_actual);
539 prt_printf(out, "proportional:");
541 prt_human_readable_s64(out, pd->last_proportional);
544 prt_printf(out, "derivative:");
546 prt_human_readable_s64(out, pd->last_derivative);
549 prt_printf(out, "change:");
551 prt_human_readable_s64(out, pd->last_change);
554 prt_printf(out, "next io:");
556 prt_printf(out, "%llims", div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC));
562 void bch2_bio_map(struct bio *bio, void *base, size_t size)
565 struct page *page = is_vmalloc_addr(base)
566 ? vmalloc_to_page(base)
567 : virt_to_page(base);
568 unsigned offset = offset_in_page(base);
569 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
571 BUG_ON(!bio_add_page(bio, page, len, offset));
577 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
580 struct page *page = alloc_page(gfp_mask);
581 unsigned len = min_t(size_t, PAGE_SIZE, size);
586 if (unlikely(!bio_add_page(bio, page, len, 0))) {
597 size_t bch2_rand_range(size_t max)
605 rand = get_random_long();
606 rand &= roundup_pow_of_two(max) - 1;
607 } while (rand >= max);
612 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
615 struct bvec_iter iter;
617 __bio_for_each_segment(bv, dst, iter, dst_iter) {
618 void *dstp = kmap_atomic(bv.bv_page);
619 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
626 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
629 struct bvec_iter iter;
631 __bio_for_each_segment(bv, src, iter, src_iter) {
632 void *srcp = kmap_atomic(bv.bv_page);
633 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
640 #include "eytzinger.h"
642 static int alignment_ok(const void *base, size_t align)
644 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
645 ((unsigned long)base & (align - 1)) == 0;
648 static void u32_swap(void *a, void *b, size_t size)
651 *(u32 *)a = *(u32 *)b;
655 static void u64_swap(void *a, void *b, size_t size)
658 *(u64 *)a = *(u64 *)b;
662 static void generic_swap(void *a, void *b, size_t size)
668 *(char *)a++ = *(char *)b;
670 } while (--size > 0);
673 static inline int do_cmp(void *base, size_t n, size_t size,
674 int (*cmp_func)(const void *, const void *, size_t),
677 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
678 base + inorder_to_eytzinger0(r, n) * size,
682 static inline void do_swap(void *base, size_t n, size_t size,
683 void (*swap_func)(void *, void *, size_t),
686 swap_func(base + inorder_to_eytzinger0(l, n) * size,
687 base + inorder_to_eytzinger0(r, n) * size,
691 void eytzinger0_sort(void *base, size_t n, size_t size,
692 int (*cmp_func)(const void *, const void *, size_t),
693 void (*swap_func)(void *, void *, size_t))
698 if (size == 4 && alignment_ok(base, 4))
699 swap_func = u32_swap;
700 else if (size == 8 && alignment_ok(base, 8))
701 swap_func = u64_swap;
703 swap_func = generic_swap;
707 for (i = n / 2 - 1; i >= 0; --i) {
708 for (r = i; r * 2 + 1 < n; r = c) {
712 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
715 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
718 do_swap(base, n, size, swap_func, r, c);
723 for (i = n - 1; i > 0; --i) {
724 do_swap(base, n, size, swap_func, 0, i);
726 for (r = 0; r * 2 + 1 < i; r = c) {
730 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
733 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
736 do_swap(base, n, size, swap_func, r, c);
741 void sort_cmp_size(void *base, size_t num, size_t size,
742 int (*cmp_func)(const void *, const void *, size_t),
743 void (*swap_func)(void *, void *, size_t size))
745 /* pre-scale counters for performance */
746 int i = (num/2 - 1) * size, n = num * size, c, r;
749 if (size == 4 && alignment_ok(base, 4))
750 swap_func = u32_swap;
751 else if (size == 8 && alignment_ok(base, 8))
752 swap_func = u64_swap;
754 swap_func = generic_swap;
758 for ( ; i >= 0; i -= size) {
759 for (r = i; r * 2 + size < n; r = c) {
762 cmp_func(base + c, base + c + size, size) < 0)
764 if (cmp_func(base + r, base + c, size) >= 0)
766 swap_func(base + r, base + c, size);
771 for (i = n - size; i > 0; i -= size) {
772 swap_func(base, base + i, size);
773 for (r = 0; r * 2 + size < i; 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 static void mempool_free_vp(void *element, void *pool_data)
787 size_t size = (size_t) pool_data;
789 vpfree(element, size);
792 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
794 size_t size = (size_t) pool_data;
796 return vpmalloc(size, gfp_mask);
799 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
801 return size < PAGE_SIZE
802 ? mempool_init_kmalloc_pool(pool, min_nr, size)
803 : mempool_init(pool, min_nr, mempool_alloc_vp,
804 mempool_free_vp, (void *) size);
808 void eytzinger1_test(void)
810 unsigned inorder, eytz, size;
812 pr_info("1 based eytzinger test:");
817 unsigned extra = eytzinger1_extra(size);
820 pr_info("tree size %u", size);
822 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
823 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
825 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
826 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
829 eytzinger1_for_each(eytz, size) {
830 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
831 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
832 BUG_ON(eytz != eytzinger1_last(size) &&
833 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
840 void eytzinger0_test(void)
843 unsigned inorder, eytz, size;
845 pr_info("0 based eytzinger test:");
850 unsigned extra = eytzinger0_extra(size);
853 pr_info("tree size %u", size);
855 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
856 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
858 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
859 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
862 eytzinger0_for_each(eytz, size) {
863 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
864 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
865 BUG_ON(eytz != eytzinger0_last(size) &&
866 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
873 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
875 const u16 *l = _l, *r = _r;
877 return (*l > *r) - (*r - *l);
880 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
882 int i, c1 = -1, c2 = -1;
885 r = eytzinger0_find_le(test_array, nr,
886 sizeof(test_array[0]),
891 for (i = 0; i < nr; i++)
892 if (test_array[i] <= search && test_array[i] > c2)
896 eytzinger0_for_each(i, nr)
897 pr_info("[%3u] = %12u", i, test_array[i]);
898 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
903 void eytzinger0_find_test(void)
905 unsigned i, nr, allocated = 1 << 12;
906 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
908 for (nr = 1; nr < allocated; nr++) {
909 pr_info("testing %u elems", nr);
911 get_random_bytes(test_array, nr * sizeof(test_array[0]));
912 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
914 /* verify array is sorted correctly: */
915 eytzinger0_for_each(i, nr)
916 BUG_ON(i != eytzinger0_last(nr) &&
917 test_array[i] > test_array[eytzinger0_next(i, nr)]);
919 for (i = 0; i < U16_MAX; i += 1 << 12)
920 eytzinger0_find_test_val(test_array, nr, i);
922 for (i = 0; i < nr; i++) {
923 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
924 eytzinger0_find_test_val(test_array, nr, test_array[i]);
925 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
934 * Accumulate percpu counters onto one cpu's copy - only valid when access
935 * against any percpu counter is guarded against
937 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
942 /* access to pcpu vars has to be blocked by other locking */
944 ret = this_cpu_ptr(p);
947 for_each_possible_cpu(cpu) {
948 u64 *i = per_cpu_ptr(p, cpu);
951 acc_u64s(ret, i, nr);
952 memset(i, 0, nr * sizeof(u64));