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 && v < 100 && v > -100)
118 pr_buf(buf, ".%i", t / 103);
120 pr_buf(buf, "%c", si_units[u]);
123 void bch2_string_opt_to_text(struct printbuf *out,
124 const char * const list[],
129 for (i = 0; list[i]; i++)
130 pr_buf(out, i == selected ? "[%s] " : "%s ", list[i]);
133 void bch2_flags_to_text(struct printbuf *out,
134 const char * const list[], u64 flags)
136 unsigned bit, nr = 0;
139 if (out->pos != out->end)
145 while (flags && (bit = __ffs(flags)) < nr) {
149 pr_buf(out, "%s", list[bit]);
154 u64 bch2_read_flag_list(char *opt, const char * const list[])
157 char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
164 while ((p = strsep(&s, ","))) {
165 int flag = match_string(list, -1, p);
179 bool bch2_is_zero(const void *_p, size_t n)
184 for (i = 0; i < n; i++)
190 static void bch2_quantiles_update(struct quantiles *q, u64 v)
194 while (i < ARRAY_SIZE(q->entries)) {
195 struct quantile_entry *e = q->entries + i;
197 if (unlikely(!e->step)) {
199 e->step = max_t(unsigned, v / 2, 1024);
200 } else if (e->m > v) {
201 e->m = e->m >= e->step
204 } else if (e->m < v) {
205 e->m = e->m + e->step > e->m
210 if ((e->m > v ? e->m - v : v - e->m) < e->step)
211 e->step = max_t(unsigned, e->step / 2, 1);
216 i = eytzinger0_child(i, v > e->m);
222 static void bch2_time_stats_update_one(struct time_stats *stats,
227 duration = time_after64(end, start)
229 freq = time_after64(end, stats->last_event)
230 ? end - stats->last_event : 0;
234 stats->average_duration = stats->average_duration
235 ? ewma_add(stats->average_duration, duration, 6)
238 stats->average_frequency = stats->average_frequency
239 ? ewma_add(stats->average_frequency, freq, 6)
242 stats->max_duration = max(stats->max_duration, duration);
244 stats->last_event = end;
246 bch2_quantiles_update(&stats->quantiles, duration);
249 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
253 if (!stats->buffer) {
254 spin_lock_irqsave(&stats->lock, flags);
255 bch2_time_stats_update_one(stats, start, end);
257 if (stats->average_frequency < 32 &&
260 alloc_percpu_gfp(struct time_stat_buffer,
262 spin_unlock_irqrestore(&stats->lock, flags);
264 struct time_stat_buffer_entry *i;
265 struct time_stat_buffer *b;
268 b = this_cpu_ptr(stats->buffer);
270 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
271 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
276 if (b->nr == ARRAY_SIZE(b->entries)) {
277 spin_lock_irqsave(&stats->lock, flags);
279 i < b->entries + ARRAY_SIZE(b->entries);
281 bch2_time_stats_update_one(stats, i->start, i->end);
282 spin_unlock_irqrestore(&stats->lock, flags);
291 static const struct time_unit {
296 { "us", NSEC_PER_USEC },
297 { "ms", NSEC_PER_MSEC },
298 { "sec", NSEC_PER_SEC },
301 static const struct time_unit *pick_time_units(u64 ns)
303 const struct time_unit *u;
306 u + 1 < time_units + ARRAY_SIZE(time_units) &&
307 ns >= u[1].nsecs << 1;
314 static void pr_time_units(struct printbuf *out, u64 ns)
316 const struct time_unit *u = pick_time_units(ns);
318 pr_buf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
321 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
323 const struct time_unit *u;
324 u64 freq = READ_ONCE(stats->average_frequency);
328 pr_buf(out, "count:\t\t%llu\n",
330 pr_buf(out, "rate:\t\t%llu/sec\n",
331 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
333 pr_buf(out, "frequency:\t");
334 pr_time_units(out, freq);
336 pr_buf(out, "\navg duration:\t");
337 pr_time_units(out, stats->average_duration);
339 pr_buf(out, "\nmax duration:\t");
340 pr_time_units(out, stats->max_duration);
342 i = eytzinger0_first(NR_QUANTILES);
343 u = pick_time_units(stats->quantiles.entries[i].m);
345 pr_buf(out, "\nquantiles (%s):\t", u->name);
346 eytzinger0_for_each(i, NR_QUANTILES) {
347 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
349 q = max(stats->quantiles.entries[i].m, last_q);
350 pr_buf(out, "%llu%s",
351 div_u64(q, u->nsecs),
352 is_last ? "\n" : " ");
357 void bch2_time_stats_exit(struct time_stats *stats)
359 free_percpu(stats->buffer);
362 void bch2_time_stats_init(struct time_stats *stats)
364 memset(stats, 0, sizeof(*stats));
365 spin_lock_init(&stats->lock);
371 * bch2_ratelimit_delay() - return how long to delay until the next time to do
374 * @d - the struct bch_ratelimit to update
376 * Returns the amount of time to delay by, in jiffies
378 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
380 u64 now = local_clock();
382 return time_after64(d->next, now)
383 ? nsecs_to_jiffies(d->next - now)
388 * bch2_ratelimit_increment() - increment @d by the amount of work done
390 * @d - the struct bch_ratelimit to update
391 * @done - the amount of work done, in arbitrary units
393 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
395 u64 now = local_clock();
397 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
399 if (time_before64(now + NSEC_PER_SEC, d->next))
400 d->next = now + NSEC_PER_SEC;
402 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
403 d->next = now - NSEC_PER_SEC * 2;
409 * Updates pd_controller. Attempts to scale inputed values to units per second.
410 * @target: desired value
411 * @actual: current value
413 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
414 * it makes actual go down.
416 void bch2_pd_controller_update(struct bch_pd_controller *pd,
417 s64 target, s64 actual, int sign)
419 s64 proportional, derivative, change;
421 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
423 if (seconds_since_update == 0)
426 pd->last_update = jiffies;
428 proportional = actual - target;
429 proportional *= seconds_since_update;
430 proportional = div_s64(proportional, pd->p_term_inverse);
432 derivative = actual - pd->last_actual;
433 derivative = div_s64(derivative, seconds_since_update);
434 derivative = ewma_add(pd->smoothed_derivative, derivative,
435 (pd->d_term / seconds_since_update) ?: 1);
436 derivative = derivative * pd->d_term;
437 derivative = div_s64(derivative, pd->p_term_inverse);
439 change = proportional + derivative;
441 /* Don't increase rate if not keeping up */
444 time_after64(local_clock(),
445 pd->rate.next + NSEC_PER_MSEC))
448 change *= (sign * -1);
450 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
453 pd->last_actual = actual;
454 pd->last_derivative = derivative;
455 pd->last_proportional = proportional;
456 pd->last_change = change;
457 pd->last_target = target;
460 void bch2_pd_controller_init(struct bch_pd_controller *pd)
462 pd->rate.rate = 1024;
463 pd->last_update = jiffies;
464 pd->p_term_inverse = 6000;
466 pd->d_smooth = pd->d_term;
467 pd->backpressure = 1;
470 size_t bch2_pd_controller_print_debug(struct bch_pd_controller *pd, char *buf)
472 /* 2^64 - 1 is 20 digits, plus null byte */
476 char proportional[21];
481 bch2_hprint(&PBUF(rate), pd->rate.rate);
482 bch2_hprint(&PBUF(actual), pd->last_actual);
483 bch2_hprint(&PBUF(target), pd->last_target);
484 bch2_hprint(&PBUF(proportional), pd->last_proportional);
485 bch2_hprint(&PBUF(derivative), pd->last_derivative);
486 bch2_hprint(&PBUF(change), pd->last_change);
488 next_io = div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC);
494 "proportional:\t%s\n"
496 "change:\t\t%s/sec\n"
497 "next io:\t%llims\n",
498 rate, target, actual, proportional,
499 derivative, change, next_io);
504 void bch2_bio_map(struct bio *bio, void *base, size_t size)
507 struct page *page = is_vmalloc_addr(base)
508 ? vmalloc_to_page(base)
509 : virt_to_page(base);
510 unsigned offset = offset_in_page(base);
511 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
513 BUG_ON(!bio_add_page(bio, page, len, offset));
519 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
522 struct page *page = alloc_page(gfp_mask);
523 unsigned len = min_t(size_t, PAGE_SIZE, size);
528 BUG_ON(!bio_add_page(bio, page, len, 0));
535 size_t bch2_rand_range(size_t max)
543 rand = get_random_long();
544 rand &= roundup_pow_of_two(max) - 1;
545 } while (rand >= max);
550 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
553 struct bvec_iter iter;
555 __bio_for_each_segment(bv, dst, iter, dst_iter) {
556 void *dstp = kmap_atomic(bv.bv_page);
557 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
564 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
567 struct bvec_iter iter;
569 __bio_for_each_segment(bv, src, iter, src_iter) {
570 void *srcp = kmap_atomic(bv.bv_page);
571 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
578 void bch_scnmemcpy(struct printbuf *out,
579 const char *src, size_t len)
581 size_t n = printbuf_remaining(out);
585 memcpy(out->pos, src, n);
591 #include "eytzinger.h"
593 static int alignment_ok(const void *base, size_t align)
595 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
596 ((unsigned long)base & (align - 1)) == 0;
599 static void u32_swap(void *a, void *b, size_t size)
602 *(u32 *)a = *(u32 *)b;
606 static void u64_swap(void *a, void *b, size_t size)
609 *(u64 *)a = *(u64 *)b;
613 static void generic_swap(void *a, void *b, size_t size)
619 *(char *)a++ = *(char *)b;
621 } while (--size > 0);
624 static inline int do_cmp(void *base, size_t n, size_t size,
625 int (*cmp_func)(const void *, const void *, size_t),
628 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
629 base + inorder_to_eytzinger0(r, n) * size,
633 static inline void do_swap(void *base, size_t n, size_t size,
634 void (*swap_func)(void *, void *, size_t),
637 swap_func(base + inorder_to_eytzinger0(l, n) * size,
638 base + inorder_to_eytzinger0(r, n) * size,
642 void eytzinger0_sort(void *base, size_t n, size_t size,
643 int (*cmp_func)(const void *, const void *, size_t),
644 void (*swap_func)(void *, void *, size_t))
649 if (size == 4 && alignment_ok(base, 4))
650 swap_func = u32_swap;
651 else if (size == 8 && alignment_ok(base, 8))
652 swap_func = u64_swap;
654 swap_func = generic_swap;
658 for (i = n / 2 - 1; i >= 0; --i) {
659 for (r = i; r * 2 + 1 < n; r = c) {
663 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
666 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
669 do_swap(base, n, size, swap_func, r, c);
674 for (i = n - 1; i > 0; --i) {
675 do_swap(base, n, size, swap_func, 0, i);
677 for (r = 0; r * 2 + 1 < i; r = c) {
681 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
684 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
687 do_swap(base, n, size, swap_func, r, c);
692 void sort_cmp_size(void *base, size_t num, size_t size,
693 int (*cmp_func)(const void *, const void *, size_t),
694 void (*swap_func)(void *, void *, size_t size))
696 /* pre-scale counters for performance */
697 int i = (num/2 - 1) * size, n = num * size, c, r;
700 if (size == 4 && alignment_ok(base, 4))
701 swap_func = u32_swap;
702 else if (size == 8 && alignment_ok(base, 8))
703 swap_func = u64_swap;
705 swap_func = generic_swap;
709 for ( ; i >= 0; i -= size) {
710 for (r = i; r * 2 + size < n; r = c) {
713 cmp_func(base + c, base + c + size, size) < 0)
715 if (cmp_func(base + r, base + c, size) >= 0)
717 swap_func(base + r, base + c, size);
722 for (i = n - size; i > 0; i -= size) {
723 swap_func(base, base + i, size);
724 for (r = 0; r * 2 + size < i; r = c) {
727 cmp_func(base + c, base + c + size, size) < 0)
729 if (cmp_func(base + r, base + c, size) >= 0)
731 swap_func(base + r, base + c, size);
736 static void mempool_free_vp(void *element, void *pool_data)
738 size_t size = (size_t) pool_data;
740 vpfree(element, size);
743 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
745 size_t size = (size_t) pool_data;
747 return vpmalloc(size, gfp_mask);
750 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
752 return size < PAGE_SIZE
753 ? mempool_init_kmalloc_pool(pool, min_nr, size)
754 : mempool_init(pool, min_nr, mempool_alloc_vp,
755 mempool_free_vp, (void *) size);
759 void eytzinger1_test(void)
761 unsigned inorder, eytz, size;
763 pr_info("1 based eytzinger test:");
768 unsigned extra = eytzinger1_extra(size);
771 pr_info("tree size %u", size);
773 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
774 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
776 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
777 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
780 eytzinger1_for_each(eytz, size) {
781 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
782 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
783 BUG_ON(eytz != eytzinger1_last(size) &&
784 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
791 void eytzinger0_test(void)
794 unsigned inorder, eytz, size;
796 pr_info("0 based eytzinger test:");
801 unsigned extra = eytzinger0_extra(size);
804 pr_info("tree size %u", size);
806 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
807 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
809 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
810 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
813 eytzinger0_for_each(eytz, size) {
814 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
815 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
816 BUG_ON(eytz != eytzinger0_last(size) &&
817 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
824 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
826 const u16 *l = _l, *r = _r;
828 return (*l > *r) - (*r - *l);
831 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
833 int i, c1 = -1, c2 = -1;
836 r = eytzinger0_find_le(test_array, nr,
837 sizeof(test_array[0]),
842 for (i = 0; i < nr; i++)
843 if (test_array[i] <= search && test_array[i] > c2)
847 eytzinger0_for_each(i, nr)
848 pr_info("[%3u] = %12u", i, test_array[i]);
849 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
854 void eytzinger0_find_test(void)
856 unsigned i, nr, allocated = 1 << 12;
857 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
859 for (nr = 1; nr < allocated; nr++) {
860 pr_info("testing %u elems", nr);
862 get_random_bytes(test_array, nr * sizeof(test_array[0]));
863 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
865 /* verify array is sorted correctly: */
866 eytzinger0_for_each(i, nr)
867 BUG_ON(i != eytzinger0_last(nr) &&
868 test_array[i] > test_array[eytzinger0_next(i, nr)]);
870 for (i = 0; i < U16_MAX; i += 1 << 12)
871 eytzinger0_find_test_val(test_array, nr, i);
873 for (i = 0; i < nr; i++) {
874 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
875 eytzinger0_find_test_val(test_array, nr, test_array[i]);
876 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
885 * Accumulate percpu counters onto one cpu's copy - only valid when access
886 * against any percpu counter is guarded against
888 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
894 ret = this_cpu_ptr(p);
897 for_each_possible_cpu(cpu) {
898 u64 *i = per_cpu_ptr(p, cpu);
901 acc_u64s(ret, i, nr);
902 memset(i, 0, nr * sizeof(u64));