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);
273 static void bch2_time_stats_update_one(struct time_stats *stats,
278 duration = time_after64(end, start)
280 freq = time_after64(end, stats->last_event)
281 ? end - stats->last_event : 0;
285 stats->average_duration = stats->average_duration
286 ? ewma_add(stats->average_duration, duration, 6)
289 stats->average_frequency = stats->average_frequency
290 ? ewma_add(stats->average_frequency, freq, 6)
293 stats->max_duration = max(stats->max_duration, duration);
295 stats->last_event = end;
297 bch2_quantiles_update(&stats->quantiles, duration);
300 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
304 if (!stats->buffer) {
305 spin_lock_irqsave(&stats->lock, flags);
306 bch2_time_stats_update_one(stats, start, end);
308 if (stats->average_frequency < 32 &&
311 alloc_percpu_gfp(struct time_stat_buffer,
313 spin_unlock_irqrestore(&stats->lock, flags);
315 struct time_stat_buffer_entry *i;
316 struct time_stat_buffer *b;
319 b = this_cpu_ptr(stats->buffer);
321 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
322 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
327 if (b->nr == ARRAY_SIZE(b->entries)) {
328 spin_lock_irqsave(&stats->lock, flags);
330 i < b->entries + ARRAY_SIZE(b->entries);
332 bch2_time_stats_update_one(stats, i->start, i->end);
333 spin_unlock_irqrestore(&stats->lock, flags);
342 static const struct time_unit {
347 { "us", NSEC_PER_USEC },
348 { "ms", NSEC_PER_MSEC },
349 { "sec", NSEC_PER_SEC },
352 static const struct time_unit *pick_time_units(u64 ns)
354 const struct time_unit *u;
357 u + 1 < time_units + ARRAY_SIZE(time_units) &&
358 ns >= u[1].nsecs << 1;
365 static void pr_time_units(struct printbuf *out, u64 ns)
367 const struct time_unit *u = pick_time_units(ns);
369 prt_printf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
372 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
374 const struct time_unit *u;
375 u64 freq = READ_ONCE(stats->average_frequency);
379 prt_printf(out, "count:\t\t%llu\n",
381 prt_printf(out, "rate:\t\t%llu/sec\n",
382 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
384 prt_printf(out, "frequency:\t");
385 pr_time_units(out, freq);
387 prt_printf(out, "\navg duration:\t");
388 pr_time_units(out, stats->average_duration);
390 prt_printf(out, "\nmax duration:\t");
391 pr_time_units(out, stats->max_duration);
393 i = eytzinger0_first(NR_QUANTILES);
394 u = pick_time_units(stats->quantiles.entries[i].m);
396 prt_printf(out, "\nquantiles (%s):\t", u->name);
397 eytzinger0_for_each(i, NR_QUANTILES) {
398 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
400 q = max(stats->quantiles.entries[i].m, last_q);
401 prt_printf(out, "%llu%s",
402 div_u64(q, u->nsecs),
403 is_last ? "\n" : " ");
408 void bch2_time_stats_exit(struct time_stats *stats)
410 free_percpu(stats->buffer);
413 void bch2_time_stats_init(struct time_stats *stats)
415 memset(stats, 0, sizeof(*stats));
416 spin_lock_init(&stats->lock);
422 * bch2_ratelimit_delay() - return how long to delay until the next time to do
425 * @d - the struct bch_ratelimit to update
427 * Returns the amount of time to delay by, in jiffies
429 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
431 u64 now = local_clock();
433 return time_after64(d->next, now)
434 ? nsecs_to_jiffies(d->next - now)
439 * bch2_ratelimit_increment() - increment @d by the amount of work done
441 * @d - the struct bch_ratelimit to update
442 * @done - the amount of work done, in arbitrary units
444 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
446 u64 now = local_clock();
448 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
450 if (time_before64(now + NSEC_PER_SEC, d->next))
451 d->next = now + NSEC_PER_SEC;
453 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
454 d->next = now - NSEC_PER_SEC * 2;
460 * Updates pd_controller. Attempts to scale inputed values to units per second.
461 * @target: desired value
462 * @actual: current value
464 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
465 * it makes actual go down.
467 void bch2_pd_controller_update(struct bch_pd_controller *pd,
468 s64 target, s64 actual, int sign)
470 s64 proportional, derivative, change;
472 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
474 if (seconds_since_update == 0)
477 pd->last_update = jiffies;
479 proportional = actual - target;
480 proportional *= seconds_since_update;
481 proportional = div_s64(proportional, pd->p_term_inverse);
483 derivative = actual - pd->last_actual;
484 derivative = div_s64(derivative, seconds_since_update);
485 derivative = ewma_add(pd->smoothed_derivative, derivative,
486 (pd->d_term / seconds_since_update) ?: 1);
487 derivative = derivative * pd->d_term;
488 derivative = div_s64(derivative, pd->p_term_inverse);
490 change = proportional + derivative;
492 /* Don't increase rate if not keeping up */
495 time_after64(local_clock(),
496 pd->rate.next + NSEC_PER_MSEC))
499 change *= (sign * -1);
501 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
504 pd->last_actual = actual;
505 pd->last_derivative = derivative;
506 pd->last_proportional = proportional;
507 pd->last_change = change;
508 pd->last_target = target;
511 void bch2_pd_controller_init(struct bch_pd_controller *pd)
513 pd->rate.rate = 1024;
514 pd->last_update = jiffies;
515 pd->p_term_inverse = 6000;
517 pd->d_smooth = pd->d_term;
518 pd->backpressure = 1;
521 void bch2_pd_controller_debug_to_text(struct printbuf *out, struct bch_pd_controller *pd)
523 out->tabstops[0] = 20;
525 prt_printf(out, "rate:");
527 prt_human_readable_s64(out, pd->rate.rate);
530 prt_printf(out, "target:");
532 prt_human_readable_u64(out, pd->last_target);
535 prt_printf(out, "actual:");
537 prt_human_readable_u64(out, pd->last_actual);
540 prt_printf(out, "proportional:");
542 prt_human_readable_s64(out, pd->last_proportional);
545 prt_printf(out, "derivative:");
547 prt_human_readable_s64(out, pd->last_derivative);
550 prt_printf(out, "change:");
552 prt_human_readable_s64(out, pd->last_change);
555 prt_printf(out, "next io:");
557 prt_printf(out, "%llims", div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC));
563 void bch2_bio_map(struct bio *bio, void *base, size_t size)
566 struct page *page = is_vmalloc_addr(base)
567 ? vmalloc_to_page(base)
568 : virt_to_page(base);
569 unsigned offset = offset_in_page(base);
570 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
572 BUG_ON(!bio_add_page(bio, page, len, offset));
578 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
581 struct page *page = alloc_page(gfp_mask);
582 unsigned len = min_t(size_t, PAGE_SIZE, size);
587 if (unlikely(!bio_add_page(bio, page, len, 0))) {
598 size_t bch2_rand_range(size_t max)
606 rand = get_random_long();
607 rand &= roundup_pow_of_two(max) - 1;
608 } while (rand >= max);
613 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
616 struct bvec_iter iter;
618 __bio_for_each_segment(bv, dst, iter, dst_iter) {
619 void *dstp = kmap_atomic(bv.bv_page);
620 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
627 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
630 struct bvec_iter iter;
632 __bio_for_each_segment(bv, src, iter, src_iter) {
633 void *srcp = kmap_atomic(bv.bv_page);
634 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
641 #include "eytzinger.h"
643 static int alignment_ok(const void *base, size_t align)
645 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
646 ((unsigned long)base & (align - 1)) == 0;
649 static void u32_swap(void *a, void *b, size_t size)
652 *(u32 *)a = *(u32 *)b;
656 static void u64_swap(void *a, void *b, size_t size)
659 *(u64 *)a = *(u64 *)b;
663 static void generic_swap(void *a, void *b, size_t size)
669 *(char *)a++ = *(char *)b;
671 } while (--size > 0);
674 static inline int do_cmp(void *base, size_t n, size_t size,
675 int (*cmp_func)(const void *, const void *, size_t),
678 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
679 base + inorder_to_eytzinger0(r, n) * size,
683 static inline void do_swap(void *base, size_t n, size_t size,
684 void (*swap_func)(void *, void *, size_t),
687 swap_func(base + inorder_to_eytzinger0(l, n) * size,
688 base + inorder_to_eytzinger0(r, n) * size,
692 void eytzinger0_sort(void *base, size_t n, size_t size,
693 int (*cmp_func)(const void *, const void *, size_t),
694 void (*swap_func)(void *, void *, size_t))
699 if (size == 4 && alignment_ok(base, 4))
700 swap_func = u32_swap;
701 else if (size == 8 && alignment_ok(base, 8))
702 swap_func = u64_swap;
704 swap_func = generic_swap;
708 for (i = n / 2 - 1; i >= 0; --i) {
709 for (r = i; r * 2 + 1 < n; r = c) {
713 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
716 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
719 do_swap(base, n, size, swap_func, r, c);
724 for (i = n - 1; i > 0; --i) {
725 do_swap(base, n, size, swap_func, 0, i);
727 for (r = 0; r * 2 + 1 < i; r = c) {
731 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
734 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
737 do_swap(base, n, size, swap_func, r, c);
742 void sort_cmp_size(void *base, size_t num, size_t size,
743 int (*cmp_func)(const void *, const void *, size_t),
744 void (*swap_func)(void *, void *, size_t size))
746 /* pre-scale counters for performance */
747 int i = (num/2 - 1) * size, n = num * size, c, r;
750 if (size == 4 && alignment_ok(base, 4))
751 swap_func = u32_swap;
752 else if (size == 8 && alignment_ok(base, 8))
753 swap_func = u64_swap;
755 swap_func = generic_swap;
759 for ( ; i >= 0; i -= size) {
760 for (r = i; r * 2 + size < n; r = c) {
763 cmp_func(base + c, base + c + size, size) < 0)
765 if (cmp_func(base + r, base + c, size) >= 0)
767 swap_func(base + r, base + c, size);
772 for (i = n - size; i > 0; i -= size) {
773 swap_func(base, base + i, size);
774 for (r = 0; r * 2 + size < i; r = c) {
777 cmp_func(base + c, base + c + size, size) < 0)
779 if (cmp_func(base + r, base + c, size) >= 0)
781 swap_func(base + r, base + c, size);
786 static void mempool_free_vp(void *element, void *pool_data)
788 size_t size = (size_t) pool_data;
790 vpfree(element, size);
793 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
795 size_t size = (size_t) pool_data;
797 return vpmalloc(size, gfp_mask);
800 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
802 return size < PAGE_SIZE
803 ? mempool_init_kmalloc_pool(pool, min_nr, size)
804 : mempool_init(pool, min_nr, mempool_alloc_vp,
805 mempool_free_vp, (void *) size);
809 void eytzinger1_test(void)
811 unsigned inorder, eytz, size;
813 pr_info("1 based eytzinger test:");
818 unsigned extra = eytzinger1_extra(size);
821 pr_info("tree size %u", size);
823 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
824 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
826 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
827 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
830 eytzinger1_for_each(eytz, size) {
831 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
832 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
833 BUG_ON(eytz != eytzinger1_last(size) &&
834 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
841 void eytzinger0_test(void)
844 unsigned inorder, eytz, size;
846 pr_info("0 based eytzinger test:");
851 unsigned extra = eytzinger0_extra(size);
854 pr_info("tree size %u", size);
856 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
857 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
859 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
860 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
863 eytzinger0_for_each(eytz, size) {
864 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
865 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
866 BUG_ON(eytz != eytzinger0_last(size) &&
867 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
874 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
876 const u16 *l = _l, *r = _r;
878 return (*l > *r) - (*r - *l);
881 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
883 int i, c1 = -1, c2 = -1;
886 r = eytzinger0_find_le(test_array, nr,
887 sizeof(test_array[0]),
892 for (i = 0; i < nr; i++)
893 if (test_array[i] <= search && test_array[i] > c2)
897 eytzinger0_for_each(i, nr)
898 pr_info("[%3u] = %12u", i, test_array[i]);
899 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
904 void eytzinger0_find_test(void)
906 unsigned i, nr, allocated = 1 << 12;
907 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
909 for (nr = 1; nr < allocated; nr++) {
910 pr_info("testing %u elems", nr);
912 get_random_bytes(test_array, nr * sizeof(test_array[0]));
913 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
915 /* verify array is sorted correctly: */
916 eytzinger0_for_each(i, nr)
917 BUG_ON(i != eytzinger0_last(nr) &&
918 test_array[i] > test_array[eytzinger0_next(i, nr)]);
920 for (i = 0; i < U16_MAX; i += 1 << 12)
921 eytzinger0_find_test_val(test_array, nr, i);
923 for (i = 0; i < nr; i++) {
924 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
925 eytzinger0_find_test_val(test_array, nr, test_array[i]);
926 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
935 * Accumulate percpu counters onto one cpu's copy - only valid when access
936 * against any percpu counter is guarded against
938 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
943 /* access to pcpu vars has to be blocked by other locking */
945 ret = this_cpu_ptr(p);
948 for_each_possible_cpu(cpu) {
949 u64 *i = per_cpu_ptr(p, cpu);
952 acc_u64s(ret, i, nr);
953 memset(i, 0, nr * sizeof(u64));