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 static int bch2_printbuf_realloc(struct printbuf *out, unsigned extra)
104 unsigned new_size = roundup_pow_of_two(out->size + extra);
105 char *buf = krealloc(out->buf, new_size, !out->atomic ? GFP_KERNEL : GFP_ATOMIC);
108 out->allocation_failure = true;
113 out->size = new_size;
117 void bch2_pr_buf(struct printbuf *out, const char *fmt, ...)
124 len = vsnprintf(out->buf + out->pos, printbuf_remaining(out), fmt, args);
126 } while (len + 1 >= printbuf_remaining(out) &&
127 !bch2_printbuf_realloc(out, len + 1));
129 len = min_t(size_t, len,
130 printbuf_remaining(out) ? printbuf_remaining(out) - 1 : 0);
134 void bch2_hprint(struct printbuf *buf, s64 v)
138 for (u = 0; v >= 1024 || v <= -1024; u++) {
139 t = v & ~(~0U << 10);
143 pr_buf(buf, "%lli", v);
146 * 103 is magic: t is in the range [-1023, 1023] and we want
147 * to turn it into [-9, 9]
149 if (u && t && v < 100 && v > -100)
150 pr_buf(buf, ".%i", t / 103);
152 pr_char(buf, si_units[u]);
155 void bch2_pr_units(struct printbuf *out, s64 raw, s64 bytes)
157 switch (out->units) {
158 case PRINTBUF_UNITS_RAW:
159 pr_buf(out, "%llu", raw);
161 case PRINTBUF_UNITS_BYTES:
162 pr_buf(out, "%llu", bytes);
164 case PRINTBUF_UNITS_HUMAN_READABLE:
165 bch2_hprint(out, bytes);
170 void bch2_string_opt_to_text(struct printbuf *out,
171 const char * const list[],
176 for (i = 0; list[i]; i++)
177 pr_buf(out, i == selected ? "[%s] " : "%s ", list[i]);
180 void bch2_flags_to_text(struct printbuf *out,
181 const char * const list[], u64 flags)
183 unsigned bit, nr = 0;
189 while (flags && (bit = __ffs(flags)) < nr) {
193 pr_buf(out, "%s", list[bit]);
198 u64 bch2_read_flag_list(char *opt, const char * const list[])
201 char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
208 while ((p = strsep(&s, ","))) {
209 int flag = match_string(list, -1, p);
223 bool bch2_is_zero(const void *_p, size_t n)
228 for (i = 0; i < n; i++)
234 static void bch2_quantiles_update(struct quantiles *q, u64 v)
238 while (i < ARRAY_SIZE(q->entries)) {
239 struct quantile_entry *e = q->entries + i;
241 if (unlikely(!e->step)) {
243 e->step = max_t(unsigned, v / 2, 1024);
244 } else if (e->m > v) {
245 e->m = e->m >= e->step
248 } else if (e->m < v) {
249 e->m = e->m + e->step > e->m
254 if ((e->m > v ? e->m - v : v - e->m) < e->step)
255 e->step = max_t(unsigned, e->step / 2, 1);
260 i = eytzinger0_child(i, v > e->m);
266 static void bch2_time_stats_update_one(struct time_stats *stats,
271 duration = time_after64(end, start)
273 freq = time_after64(end, stats->last_event)
274 ? end - stats->last_event : 0;
278 stats->average_duration = stats->average_duration
279 ? ewma_add(stats->average_duration, duration, 6)
282 stats->average_frequency = stats->average_frequency
283 ? ewma_add(stats->average_frequency, freq, 6)
286 stats->max_duration = max(stats->max_duration, duration);
288 stats->last_event = end;
290 bch2_quantiles_update(&stats->quantiles, duration);
293 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
297 if (!stats->buffer) {
298 spin_lock_irqsave(&stats->lock, flags);
299 bch2_time_stats_update_one(stats, start, end);
301 if (stats->average_frequency < 32 &&
304 alloc_percpu_gfp(struct time_stat_buffer,
306 spin_unlock_irqrestore(&stats->lock, flags);
308 struct time_stat_buffer_entry *i;
309 struct time_stat_buffer *b;
312 b = this_cpu_ptr(stats->buffer);
314 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
315 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
320 if (b->nr == ARRAY_SIZE(b->entries)) {
321 spin_lock_irqsave(&stats->lock, flags);
323 i < b->entries + ARRAY_SIZE(b->entries);
325 bch2_time_stats_update_one(stats, i->start, i->end);
326 spin_unlock_irqrestore(&stats->lock, flags);
335 static const struct time_unit {
340 { "us", NSEC_PER_USEC },
341 { "ms", NSEC_PER_MSEC },
342 { "sec", NSEC_PER_SEC },
345 static const struct time_unit *pick_time_units(u64 ns)
347 const struct time_unit *u;
350 u + 1 < time_units + ARRAY_SIZE(time_units) &&
351 ns >= u[1].nsecs << 1;
358 static void pr_time_units(struct printbuf *out, u64 ns)
360 const struct time_unit *u = pick_time_units(ns);
362 pr_buf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
365 void bch2_time_stats_to_text(struct printbuf *out, struct time_stats *stats)
367 const struct time_unit *u;
368 u64 freq = READ_ONCE(stats->average_frequency);
372 pr_buf(out, "count:\t\t%llu\n",
374 pr_buf(out, "rate:\t\t%llu/sec\n",
375 freq ? div64_u64(NSEC_PER_SEC, freq) : 0);
377 pr_buf(out, "frequency:\t");
378 pr_time_units(out, freq);
380 pr_buf(out, "\navg duration:\t");
381 pr_time_units(out, stats->average_duration);
383 pr_buf(out, "\nmax duration:\t");
384 pr_time_units(out, stats->max_duration);
386 i = eytzinger0_first(NR_QUANTILES);
387 u = pick_time_units(stats->quantiles.entries[i].m);
389 pr_buf(out, "\nquantiles (%s):\t", u->name);
390 eytzinger0_for_each(i, NR_QUANTILES) {
391 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
393 q = max(stats->quantiles.entries[i].m, last_q);
394 pr_buf(out, "%llu%s",
395 div_u64(q, u->nsecs),
396 is_last ? "\n" : " ");
401 void bch2_time_stats_exit(struct time_stats *stats)
403 free_percpu(stats->buffer);
406 void bch2_time_stats_init(struct time_stats *stats)
408 memset(stats, 0, sizeof(*stats));
409 spin_lock_init(&stats->lock);
415 * bch2_ratelimit_delay() - return how long to delay until the next time to do
418 * @d - the struct bch_ratelimit to update
420 * Returns the amount of time to delay by, in jiffies
422 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
424 u64 now = local_clock();
426 return time_after64(d->next, now)
427 ? nsecs_to_jiffies(d->next - now)
432 * bch2_ratelimit_increment() - increment @d by the amount of work done
434 * @d - the struct bch_ratelimit to update
435 * @done - the amount of work done, in arbitrary units
437 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
439 u64 now = local_clock();
441 d->next += div_u64(done * NSEC_PER_SEC, d->rate);
443 if (time_before64(now + NSEC_PER_SEC, d->next))
444 d->next = now + NSEC_PER_SEC;
446 if (time_after64(now - NSEC_PER_SEC * 2, d->next))
447 d->next = now - NSEC_PER_SEC * 2;
453 * Updates pd_controller. Attempts to scale inputed values to units per second.
454 * @target: desired value
455 * @actual: current value
457 * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
458 * it makes actual go down.
460 void bch2_pd_controller_update(struct bch_pd_controller *pd,
461 s64 target, s64 actual, int sign)
463 s64 proportional, derivative, change;
465 unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
467 if (seconds_since_update == 0)
470 pd->last_update = jiffies;
472 proportional = actual - target;
473 proportional *= seconds_since_update;
474 proportional = div_s64(proportional, pd->p_term_inverse);
476 derivative = actual - pd->last_actual;
477 derivative = div_s64(derivative, seconds_since_update);
478 derivative = ewma_add(pd->smoothed_derivative, derivative,
479 (pd->d_term / seconds_since_update) ?: 1);
480 derivative = derivative * pd->d_term;
481 derivative = div_s64(derivative, pd->p_term_inverse);
483 change = proportional + derivative;
485 /* Don't increase rate if not keeping up */
488 time_after64(local_clock(),
489 pd->rate.next + NSEC_PER_MSEC))
492 change *= (sign * -1);
494 pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
497 pd->last_actual = actual;
498 pd->last_derivative = derivative;
499 pd->last_proportional = proportional;
500 pd->last_change = change;
501 pd->last_target = target;
504 void bch2_pd_controller_init(struct bch_pd_controller *pd)
506 pd->rate.rate = 1024;
507 pd->last_update = jiffies;
508 pd->p_term_inverse = 6000;
510 pd->d_smooth = pd->d_term;
511 pd->backpressure = 1;
514 void bch2_pd_controller_debug_to_text(struct printbuf *out, struct bch_pd_controller *pd)
516 out->tabstops[0] = 20;
518 pr_buf(out, "rate:");
520 bch2_hprint(out, pd->rate.rate);
523 pr_buf(out, "target:");
525 bch2_hprint(out, pd->last_target);
528 pr_buf(out, "actual:");
530 bch2_hprint(out, pd->last_actual);
533 pr_buf(out, "proportional:");
535 bch2_hprint(out, pd->last_proportional);
538 pr_buf(out, "derivative:");
540 bch2_hprint(out, pd->last_derivative);
543 pr_buf(out, "change:");
545 bch2_hprint(out, pd->last_change);
548 pr_buf(out, "next io:");
550 pr_buf(out, "%llims", div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC));
556 void bch2_bio_map(struct bio *bio, void *base, size_t size)
559 struct page *page = is_vmalloc_addr(base)
560 ? vmalloc_to_page(base)
561 : virt_to_page(base);
562 unsigned offset = offset_in_page(base);
563 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
565 BUG_ON(!bio_add_page(bio, page, len, offset));
571 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
574 struct page *page = alloc_page(gfp_mask);
575 unsigned len = min_t(size_t, PAGE_SIZE, size);
580 if (unlikely(!bio_add_page(bio, page, len, 0))) {
591 size_t bch2_rand_range(size_t max)
599 rand = get_random_long();
600 rand &= roundup_pow_of_two(max) - 1;
601 } while (rand >= max);
606 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
609 struct bvec_iter iter;
611 __bio_for_each_segment(bv, dst, iter, dst_iter) {
612 void *dstp = kmap_atomic(bv.bv_page);
613 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
620 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
623 struct bvec_iter iter;
625 __bio_for_each_segment(bv, src, iter, src_iter) {
626 void *srcp = kmap_atomic(bv.bv_page);
627 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
634 #include "eytzinger.h"
636 static int alignment_ok(const void *base, size_t align)
638 return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
639 ((unsigned long)base & (align - 1)) == 0;
642 static void u32_swap(void *a, void *b, size_t size)
645 *(u32 *)a = *(u32 *)b;
649 static void u64_swap(void *a, void *b, size_t size)
652 *(u64 *)a = *(u64 *)b;
656 static void generic_swap(void *a, void *b, size_t size)
662 *(char *)a++ = *(char *)b;
664 } while (--size > 0);
667 static inline int do_cmp(void *base, size_t n, size_t size,
668 int (*cmp_func)(const void *, const void *, size_t),
671 return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
672 base + inorder_to_eytzinger0(r, n) * size,
676 static inline void do_swap(void *base, size_t n, size_t size,
677 void (*swap_func)(void *, void *, size_t),
680 swap_func(base + inorder_to_eytzinger0(l, n) * size,
681 base + inorder_to_eytzinger0(r, n) * size,
685 void eytzinger0_sort(void *base, size_t n, size_t size,
686 int (*cmp_func)(const void *, const void *, size_t),
687 void (*swap_func)(void *, void *, size_t))
692 if (size == 4 && alignment_ok(base, 4))
693 swap_func = u32_swap;
694 else if (size == 8 && alignment_ok(base, 8))
695 swap_func = u64_swap;
697 swap_func = generic_swap;
701 for (i = n / 2 - 1; i >= 0; --i) {
702 for (r = i; r * 2 + 1 < n; r = c) {
706 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
709 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
712 do_swap(base, n, size, swap_func, r, c);
717 for (i = n - 1; i > 0; --i) {
718 do_swap(base, n, size, swap_func, 0, i);
720 for (r = 0; r * 2 + 1 < i; r = c) {
724 do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
727 if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
730 do_swap(base, n, size, swap_func, r, c);
735 void sort_cmp_size(void *base, size_t num, size_t size,
736 int (*cmp_func)(const void *, const void *, size_t),
737 void (*swap_func)(void *, void *, size_t size))
739 /* pre-scale counters for performance */
740 int i = (num/2 - 1) * size, n = num * size, c, r;
743 if (size == 4 && alignment_ok(base, 4))
744 swap_func = u32_swap;
745 else if (size == 8 && alignment_ok(base, 8))
746 swap_func = u64_swap;
748 swap_func = generic_swap;
752 for ( ; i >= 0; i -= size) {
753 for (r = i; r * 2 + size < n; r = c) {
756 cmp_func(base + c, base + c + size, size) < 0)
758 if (cmp_func(base + r, base + c, size) >= 0)
760 swap_func(base + r, base + c, size);
765 for (i = n - size; i > 0; i -= size) {
766 swap_func(base, base + i, size);
767 for (r = 0; r * 2 + size < i; r = c) {
770 cmp_func(base + c, base + c + size, size) < 0)
772 if (cmp_func(base + r, base + c, size) >= 0)
774 swap_func(base + r, base + c, size);
779 static void mempool_free_vp(void *element, void *pool_data)
781 size_t size = (size_t) pool_data;
783 vpfree(element, size);
786 static void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
788 size_t size = (size_t) pool_data;
790 return vpmalloc(size, gfp_mask);
793 int mempool_init_kvpmalloc_pool(mempool_t *pool, int min_nr, size_t size)
795 return size < PAGE_SIZE
796 ? mempool_init_kmalloc_pool(pool, min_nr, size)
797 : mempool_init(pool, min_nr, mempool_alloc_vp,
798 mempool_free_vp, (void *) size);
802 void eytzinger1_test(void)
804 unsigned inorder, eytz, size;
806 pr_info("1 based eytzinger test:");
811 unsigned extra = eytzinger1_extra(size);
814 pr_info("tree size %u", size);
816 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
817 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
819 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size) != 0);
820 BUG_ON(eytzinger1_next(eytzinger1_last(size), size) != 0);
823 eytzinger1_for_each(eytz, size) {
824 BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
825 BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
826 BUG_ON(eytz != eytzinger1_last(size) &&
827 eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
834 void eytzinger0_test(void)
837 unsigned inorder, eytz, size;
839 pr_info("0 based eytzinger test:");
844 unsigned extra = eytzinger0_extra(size);
847 pr_info("tree size %u", size);
849 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
850 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
852 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size) != -1);
853 BUG_ON(eytzinger0_next(eytzinger0_last(size), size) != -1);
856 eytzinger0_for_each(eytz, size) {
857 BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
858 BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
859 BUG_ON(eytz != eytzinger0_last(size) &&
860 eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
867 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
869 const u16 *l = _l, *r = _r;
871 return (*l > *r) - (*r - *l);
874 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
876 int i, c1 = -1, c2 = -1;
879 r = eytzinger0_find_le(test_array, nr,
880 sizeof(test_array[0]),
885 for (i = 0; i < nr; i++)
886 if (test_array[i] <= search && test_array[i] > c2)
890 eytzinger0_for_each(i, nr)
891 pr_info("[%3u] = %12u", i, test_array[i]);
892 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
897 void eytzinger0_find_test(void)
899 unsigned i, nr, allocated = 1 << 12;
900 u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
902 for (nr = 1; nr < allocated; nr++) {
903 pr_info("testing %u elems", nr);
905 get_random_bytes(test_array, nr * sizeof(test_array[0]));
906 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
908 /* verify array is sorted correctly: */
909 eytzinger0_for_each(i, nr)
910 BUG_ON(i != eytzinger0_last(nr) &&
911 test_array[i] > test_array[eytzinger0_next(i, nr)]);
913 for (i = 0; i < U16_MAX; i += 1 << 12)
914 eytzinger0_find_test_val(test_array, nr, i);
916 for (i = 0; i < nr; i++) {
917 eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
918 eytzinger0_find_test_val(test_array, nr, test_array[i]);
919 eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
928 * Accumulate percpu counters onto one cpu's copy - only valid when access
929 * against any percpu counter is guarded against
931 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
936 /* access to pcpu vars has to be blocked by other locking */
938 ret = this_cpu_ptr(p);
941 for_each_possible_cpu(cpu) {
942 u64 *i = per_cpu_ptr(p, cpu);
945 acc_u64s(ret, i, nr);
946 memset(i, 0, nr * sizeof(u64));