fs/bcachefs/util.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * random utiility code, for bcache but in theory not specific to bcache
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <[email protected]>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include <linux/bio.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/console.h>
  12 #include <linux/ctype.h>
  13 #include <linux/debugfs.h>
  14 #include <linux/freezer.h>
  15 #include <linux/kthread.h>
  16 #include <linux/log2.h>
  17 #include <linux/math64.h>
  18 #include <linux/percpu.h>
  19 #include <linux/preempt.h>
  20 #include <linux/random.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/string.h>
  23 #include <linux/types.h>
  24 #include <linux/sched/clock.h>
  25
  26 #include "eytzinger.h"
  27 #include "mean_and_variance.h"
  28 #include "util.h"
  29
  30 static const char si_units[] = "?kMGTPEZY";
  31
  32 /* string_get_size units: */
  33 static const char *const units_2[] = {
  34         "B", "KiB", "MiB", "GiB", "TiB", "PiB", "EiB", "ZiB", "YiB"
  35 };
  36 static const char *const units_10[] = {
  37         "B", "kB", "MB", "GB", "TB", "PB", "EB", "ZB", "YB"
  38 };
  39
  40 static int parse_u64(const char *cp, u64 *res)
  41 {
  42         const char *start = cp;
  43         u64 v = 0;
  44
  45         if (!isdigit(*cp))
  46                 return -EINVAL;
  47
  48         do {
  49                 if (v > U64_MAX / 10)
  50                         return -ERANGE;
  51                 v *= 10;
  52                 if (v > U64_MAX - (*cp - '0'))
  53                         return -ERANGE;
  54                 v += *cp - '0';
  55                 cp++;
  56         } while (isdigit(*cp));
  57
  58         *res = v;
  59         return cp - start;
  60 }
  61
  62 static int bch2_pow(u64 n, u64 p, u64 *res)
  63 {
  64         *res = 1;
  65
  66         while (p--) {
  67                 if (*res > div_u64(U64_MAX, n))
  68                         return -ERANGE;
  69                 *res *= n;
  70         }
  71         return 0;
  72 }
  73
  74 static int parse_unit_suffix(const char *cp, u64 *res)
  75 {
  76         const char *start = cp;
  77         u64 base = 1024;
  78         unsigned u;
  79         int ret;
  80
  81         if (*cp == ' ')
  82                 cp++;
  83
  84         for (u = 1; u < strlen(si_units); u++)
  85                 if (*cp == si_units[u]) {
  86                         cp++;
  87                         goto got_unit;
  88                 }
  89
  90         for (u = 0; u < ARRAY_SIZE(units_2); u++)
  91                 if (!strncmp(cp, units_2[u], strlen(units_2[u]))) {
  92                         cp += strlen(units_2[u]);
  93                         goto got_unit;
  94                 }
  95
  96         for (u = 0; u < ARRAY_SIZE(units_10); u++)
  97                 if (!strncmp(cp, units_10[u], strlen(units_10[u]))) {
  98                         cp += strlen(units_10[u]);
  99                         base = 1000;
 100                         goto got_unit;
 101                 }
 102
 103         *res = 1;
 104         return 0;
 105 got_unit:
 106         ret = bch2_pow(base, u, res);
 107         if (ret)
 108                 return ret;
 109
 110         return cp - start;
 111 }
 112
 113 #define parse_or_ret(cp, _f)                    \
 114 do {                                            \
 115         int _ret = _f;                          \
 116         if (_ret < 0)                           \
 117                 return _ret;                    \
 118         cp += _ret;                             \
 119 } while (0)
 120
 121 static int __bch2_strtou64_h(const char *cp, u64 *res)
 122 {
 123         const char *start = cp;
 124         u64 v = 0, b, f_n = 0, f_d = 1;
 125         int ret;
 126
 127         parse_or_ret(cp, parse_u64(cp, &v));
 128
 129         if (*cp == '.') {
 130                 cp++;
 131                 ret = parse_u64(cp, &f_n);
 132                 if (ret < 0)
 133                         return ret;
 134                 cp += ret;
 135
 136                 ret = bch2_pow(10, ret, &f_d);
 137                 if (ret)
 138                         return ret;
 139         }
 140
 141         parse_or_ret(cp, parse_unit_suffix(cp, &b));
 142
 143         if (v > div_u64(U64_MAX, b))
 144                 return -ERANGE;
 145         v *= b;
 146
 147         if (f_n > div_u64(U64_MAX, b))
 148                 return -ERANGE;
 149
 150         f_n = div_u64(f_n * b, f_d);
 151         if (v + f_n < v)
 152                 return -ERANGE;
 153         v += f_n;
 154
 155         *res = v;
 156         return cp - start;
 157 }
 158
 159 static int __bch2_strtoh(const char *cp, u64 *res,
 160                          u64 t_max, bool t_signed)
 161 {
 162         bool positive = *cp != '-';
 163         u64 v = 0;
 164
 165         if (*cp == '+' || *cp == '-')
 166                 cp++;
 167
 168         parse_or_ret(cp, __bch2_strtou64_h(cp, &v));
 169
 170         if (*cp == '\n')
 171                 cp++;
 172         if (*cp)
 173                 return -EINVAL;
 174
 175         if (positive) {
 176                 if (v > t_max)
 177                         return -ERANGE;
 178         } else {
 179                 if (v && !t_signed)
 180                         return -ERANGE;
 181
 182                 if (v > t_max + 1)
 183                         return -ERANGE;
 184                 v = -v;
 185         }
 186
 187         *res = v;
 188         return 0;
 189 }
 190
 191 #define STRTO_H(name, type)                                     \
 192 int bch2_ ## name ## _h(const char *cp, type *res)              \
 193 {                                                               \
 194         u64 v = 0;                                              \
 195         int ret = __bch2_strtoh(cp, &v, ANYSINT_MAX(type),      \
 196                         ANYSINT_MAX(type) != ((type) ~0ULL));   \
 197         *res = v;                                               \
 198         return ret;                                             \
 199 }
 200
 201 STRTO_H(strtoint, int)
 202 STRTO_H(strtouint, unsigned int)
 203 STRTO_H(strtoll, long long)
 204 STRTO_H(strtoull, unsigned long long)
 205 STRTO_H(strtou64, u64)
 206
 207 u64 bch2_read_flag_list(char *opt, const char * const list[])
 208 {
 209         u64 ret = 0;
 210         char *p, *s, *d = kstrdup(opt, GFP_KERNEL);
 211
 212         if (!d)
 213                 return -ENOMEM;
 214
 215         s = strim(d);
 216
 217         while ((p = strsep(&s, ","))) {
 218                 int flag = match_string(list, -1, p);
 219
 220                 if (flag < 0) {
 221                         ret = -1;
 222                         break;
 223                 }
 224
 225                 ret |= 1 << flag;
 226         }
 227
 228         kfree(d);
 229
 230         return ret;
 231 }
 232
 233 bool bch2_is_zero(const void *_p, size_t n)
 234 {
 235         const char *p = _p;
 236         size_t i;
 237
 238         for (i = 0; i < n; i++)
 239                 if (p[i])
 240                         return false;
 241         return true;
 242 }
 243
 244 void bch2_prt_u64_base2_nbits(struct printbuf *out, u64 v, unsigned nr_bits)
 245 {
 246         while (nr_bits)
 247                 prt_char(out, '0' + ((v >> --nr_bits) & 1));
 248 }
 249
 250 void bch2_prt_u64_base2(struct printbuf *out, u64 v)
 251 {
 252         bch2_prt_u64_base2_nbits(out, v, fls64(v) ?: 1);
 253 }
 254
 255 void bch2_print_string_as_lines(const char *prefix, const char *lines)
 256 {
 257         const char *p;
 258
 259         if (!lines) {
 260                 printk("%s (null)\n", prefix);
 261                 return;
 262         }
 263
 264         console_lock();
 265         while (1) {
 266                 p = strchrnul(lines, '\n');
 267                 printk("%s%.*s\n", prefix, (int) (p - lines), lines);
 268                 if (!*p)
 269                         break;
 270                 lines = p + 1;
 271         }
 272         console_unlock();
 273 }
 274
 275 int bch2_save_backtrace(bch_stacktrace *stack, struct task_struct *task, unsigned skipnr,
 276                         gfp_t gfp)
 277 {
 278 #ifdef CONFIG_STACKTRACE
 279         unsigned nr_entries = 0;
 280
 281         stack->nr = 0;
 282         int ret = darray_make_room_gfp(stack, 32, gfp);
 283         if (ret)
 284                 return ret;
 285
 286         if (!down_read_trylock(&task->signal->exec_update_lock))
 287                 return -1;
 288
 289         do {
 290                 nr_entries = stack_trace_save_tsk(task, stack->data, stack->size, skipnr + 1);
 291         } while (nr_entries == stack->size &&
 292                  !(ret = darray_make_room_gfp(stack, stack->size * 2, gfp)));
 293
 294         stack->nr = nr_entries;
 295         up_read(&task->signal->exec_update_lock);
 296
 297         return ret;
 298 #else
 299         return 0;
 300 #endif
 301 }
 302
 303 void bch2_prt_backtrace(struct printbuf *out, bch_stacktrace *stack)
 304 {
 305         darray_for_each(*stack, i) {
 306                 prt_printf(out, "[<0>] %pB", (void *) *i);
 307                 prt_newline(out);
 308         }
 309 }
 310
 311 int bch2_prt_task_backtrace(struct printbuf *out, struct task_struct *task, unsigned skipnr, gfp_t gfp)
 312 {
 313         bch_stacktrace stack = { 0 };
 314         int ret = bch2_save_backtrace(&stack, task, skipnr + 1, gfp);
 315
 316         bch2_prt_backtrace(out, &stack);
 317         darray_exit(&stack);
 318         return ret;
 319 }
 320
 321 #ifndef __KERNEL__
 322 #include <time.h>
 323 void bch2_prt_datetime(struct printbuf *out, time64_t sec)
 324 {
 325         time_t t = sec;
 326         char buf[64];
 327         ctime_r(&t, buf);
 328         strim(buf);
 329         prt_str(out, buf);
 330 }
 331 #else
 332 void bch2_prt_datetime(struct printbuf *out, time64_t sec)
 333 {
 334         char buf[64];
 335         snprintf(buf, sizeof(buf), "%ptT", &sec);
 336         prt_u64(out, sec);
 337 }
 338 #endif
 339
 340 void bch2_pr_time_units(struct printbuf *out, u64 ns)
 341 {
 342         const struct time_unit *u = bch2_pick_time_units(ns);
 343
 344         prt_printf(out, "%llu %s", div_u64(ns, u->nsecs), u->name);
 345 }
 346
 347 static void bch2_pr_time_units_aligned(struct printbuf *out, u64 ns)
 348 {
 349         const struct time_unit *u = bch2_pick_time_units(ns);
 350
 351         prt_printf(out, "%llu \r%s", div64_u64(ns, u->nsecs), u->name);
 352 }
 353
 354 static inline void pr_name_and_units(struct printbuf *out, const char *name, u64 ns)
 355 {
 356         prt_printf(out, "%s\t", name);
 357         bch2_pr_time_units_aligned(out, ns);
 358         prt_newline(out);
 359 }
 360
 361 #define TABSTOP_SIZE 12
 362
 363 void bch2_time_stats_to_text(struct printbuf *out, struct bch2_time_stats *stats)
 364 {
 365         struct quantiles *quantiles = time_stats_to_quantiles(stats);
 366         s64 f_mean = 0, d_mean = 0;
 367         u64 f_stddev = 0, d_stddev = 0;
 368
 369         if (stats->buffer) {
 370                 int cpu;
 371
 372                 spin_lock_irq(&stats->lock);
 373                 for_each_possible_cpu(cpu)
 374                         __bch2_time_stats_clear_buffer(stats, per_cpu_ptr(stats->buffer, cpu));
 375                 spin_unlock_irq(&stats->lock);
 376         }
 377
 378         /*
 379          * avoid divide by zero
 380          */
 381         if (stats->freq_stats.n) {
 382                 f_mean = mean_and_variance_get_mean(stats->freq_stats);
 383                 f_stddev = mean_and_variance_get_stddev(stats->freq_stats);
 384                 d_mean = mean_and_variance_get_mean(stats->duration_stats);
 385                 d_stddev = mean_and_variance_get_stddev(stats->duration_stats);
 386         }
 387
 388         printbuf_tabstop_push(out, out->indent + TABSTOP_SIZE);
 389         prt_printf(out, "count:\t%llu\n", stats->duration_stats.n);
 390         printbuf_tabstop_pop(out);
 391
 392         printbuf_tabstops_reset(out);
 393
 394         printbuf_tabstop_push(out, out->indent + 20);
 395         printbuf_tabstop_push(out, TABSTOP_SIZE + 2);
 396         printbuf_tabstop_push(out, 0);
 397         printbuf_tabstop_push(out, TABSTOP_SIZE + 2);
 398
 399         prt_printf(out, "\tsince mount\r\trecent\r\n");
 400         prt_printf(out, "recent");
 401
 402         printbuf_tabstops_reset(out);
 403         printbuf_tabstop_push(out, out->indent + 20);
 404         printbuf_tabstop_push(out, TABSTOP_SIZE);
 405         printbuf_tabstop_push(out, 2);
 406         printbuf_tabstop_push(out, TABSTOP_SIZE);
 407
 408         prt_printf(out, "duration of events\n");
 409         printbuf_indent_add(out, 2);
 410
 411         pr_name_and_units(out, "min:", stats->min_duration);
 412         pr_name_and_units(out, "max:", stats->max_duration);
 413         pr_name_and_units(out, "total:", stats->total_duration);
 414
 415         prt_printf(out, "mean:\t");
 416         bch2_pr_time_units_aligned(out, d_mean);
 417         prt_tab(out);
 418         bch2_pr_time_units_aligned(out, mean_and_variance_weighted_get_mean(stats->duration_stats_weighted, TIME_STATS_MV_WEIGHT));
 419         prt_newline(out);
 420
 421         prt_printf(out, "stddev:\t");
 422         bch2_pr_time_units_aligned(out, d_stddev);
 423         prt_tab(out);
 424         bch2_pr_time_units_aligned(out, mean_and_variance_weighted_get_stddev(stats->duration_stats_weighted, TIME_STATS_MV_WEIGHT));
 425
 426         printbuf_indent_sub(out, 2);
 427         prt_newline(out);
 428
 429         prt_printf(out, "time between events\n");
 430         printbuf_indent_add(out, 2);
 431
 432         pr_name_and_units(out, "min:", stats->min_freq);
 433         pr_name_and_units(out, "max:", stats->max_freq);
 434
 435         prt_printf(out, "mean:\t");
 436         bch2_pr_time_units_aligned(out, f_mean);
 437         prt_tab(out);
 438         bch2_pr_time_units_aligned(out, mean_and_variance_weighted_get_mean(stats->freq_stats_weighted, TIME_STATS_MV_WEIGHT));
 439         prt_newline(out);
 440
 441         prt_printf(out, "stddev:\t");
 442         bch2_pr_time_units_aligned(out, f_stddev);
 443         prt_tab(out);
 444         bch2_pr_time_units_aligned(out, mean_and_variance_weighted_get_stddev(stats->freq_stats_weighted, TIME_STATS_MV_WEIGHT));
 445
 446         printbuf_indent_sub(out, 2);
 447         prt_newline(out);
 448
 449         printbuf_tabstops_reset(out);
 450
 451         if (quantiles) {
 452                 int i = eytzinger0_first(NR_QUANTILES);
 453                 const struct time_unit *u =
 454                         bch2_pick_time_units(quantiles->entries[i].m);
 455                 u64 last_q = 0;
 456
 457                 prt_printf(out, "quantiles (%s):\t", u->name);
 458                 eytzinger0_for_each(i, NR_QUANTILES) {
 459                         bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
 460
 461                         u64 q = max(quantiles->entries[i].m, last_q);
 462                         prt_printf(out, "%llu ", div_u64(q, u->nsecs));
 463                         if (is_last)
 464                                 prt_newline(out);
 465                         last_q = q;
 466                 }
 467         }
 468 }
 469
 470 /* ratelimit: */
 471
 472 /**
 473  * bch2_ratelimit_delay() - return how long to delay until the next time to do
 474  *              some work
 475  * @d:          the struct bch_ratelimit to update
 476  * Returns:     the amount of time to delay by, in jiffies
 477  */
 478 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
 479 {
 480         u64 now = local_clock();
 481
 482         return time_after64(d->next, now)
 483                 ? nsecs_to_jiffies(d->next - now)
 484                 : 0;
 485 }
 486
 487 /**
 488  * bch2_ratelimit_increment() - increment @d by the amount of work done
 489  * @d:          the struct bch_ratelimit to update
 490  * @done:       the amount of work done, in arbitrary units
 491  */
 492 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
 493 {
 494         u64 now = local_clock();
 495
 496         d->next += div_u64(done * NSEC_PER_SEC, d->rate);
 497
 498         if (time_before64(now + NSEC_PER_SEC, d->next))
 499                 d->next = now + NSEC_PER_SEC;
 500
 501         if (time_after64(now - NSEC_PER_SEC * 2, d->next))
 502                 d->next = now - NSEC_PER_SEC * 2;
 503 }
 504
 505 /* pd controller: */
 506
 507 /*
 508  * Updates pd_controller. Attempts to scale inputed values to units per second.
 509  * @target: desired value
 510  * @actual: current value
 511  *
 512  * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
 513  * it makes actual go down.
 514  */
 515 void bch2_pd_controller_update(struct bch_pd_controller *pd,
 516                               s64 target, s64 actual, int sign)
 517 {
 518         s64 proportional, derivative, change;
 519
 520         unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
 521
 522         if (seconds_since_update == 0)
 523                 return;
 524
 525         pd->last_update = jiffies;
 526
 527         proportional = actual - target;
 528         proportional *= seconds_since_update;
 529         proportional = div_s64(proportional, pd->p_term_inverse);
 530
 531         derivative = actual - pd->last_actual;
 532         derivative = div_s64(derivative, seconds_since_update);
 533         derivative = ewma_add(pd->smoothed_derivative, derivative,
 534                               (pd->d_term / seconds_since_update) ?: 1);
 535         derivative = derivative * pd->d_term;
 536         derivative = div_s64(derivative, pd->p_term_inverse);
 537
 538         change = proportional + derivative;
 539
 540         /* Don't increase rate if not keeping up */
 541         if (change > 0 &&
 542             pd->backpressure &&
 543             time_after64(local_clock(),
 544                          pd->rate.next + NSEC_PER_MSEC))
 545                 change = 0;
 546
 547         change *= (sign * -1);
 548
 549         pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
 550                                 1, UINT_MAX);
 551
 552         pd->last_actual         = actual;
 553         pd->last_derivative     = derivative;
 554         pd->last_proportional   = proportional;
 555         pd->last_change         = change;
 556         pd->last_target         = target;
 557 }
 558
 559 void bch2_pd_controller_init(struct bch_pd_controller *pd)
 560 {
 561         pd->rate.rate           = 1024;
 562         pd->last_update         = jiffies;
 563         pd->p_term_inverse      = 6000;
 564         pd->d_term              = 30;
 565         pd->d_smooth            = pd->d_term;
 566         pd->backpressure        = 1;
 567 }
 568
 569 void bch2_pd_controller_debug_to_text(struct printbuf *out, struct bch_pd_controller *pd)
 570 {
 571         if (!out->nr_tabstops)
 572                 printbuf_tabstop_push(out, 20);
 573
 574         prt_printf(out, "rate:\t");
 575         prt_human_readable_s64(out, pd->rate.rate);
 576         prt_newline(out);
 577
 578         prt_printf(out, "target:\t");
 579         prt_human_readable_u64(out, pd->last_target);
 580         prt_newline(out);
 581
 582         prt_printf(out, "actual:\t");
 583         prt_human_readable_u64(out, pd->last_actual);
 584         prt_newline(out);
 585
 586         prt_printf(out, "proportional:\t");
 587         prt_human_readable_s64(out, pd->last_proportional);
 588         prt_newline(out);
 589
 590         prt_printf(out, "derivative:\t");
 591         prt_human_readable_s64(out, pd->last_derivative);
 592         prt_newline(out);
 593
 594         prt_printf(out, "change:\t");
 595         prt_human_readable_s64(out, pd->last_change);
 596         prt_newline(out);
 597
 598         prt_printf(out, "next io:\t%llims\n", div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC));
 599 }
 600
 601 /* misc: */
 602
 603 void bch2_bio_map(struct bio *bio, void *base, size_t size)
 604 {
 605         while (size) {
 606                 struct page *page = is_vmalloc_addr(base)
 607                                 ? vmalloc_to_page(base)
 608                                 : virt_to_page(base);
 609                 unsigned offset = offset_in_page(base);
 610                 unsigned len = min_t(size_t, PAGE_SIZE - offset, size);
 611
 612                 BUG_ON(!bio_add_page(bio, page, len, offset));
 613                 size -= len;
 614                 base += len;
 615         }
 616 }
 617
 618 int bch2_bio_alloc_pages(struct bio *bio, size_t size, gfp_t gfp_mask)
 619 {
 620         while (size) {
 621                 struct page *page = alloc_pages(gfp_mask, 0);
 622                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 623
 624                 if (!page)
 625                         return -ENOMEM;
 626
 627                 if (unlikely(!bio_add_page(bio, page, len, 0))) {
 628                         __free_page(page);
 629                         break;
 630                 }
 631
 632                 size -= len;
 633         }
 634
 635         return 0;
 636 }
 637
 638 size_t bch2_rand_range(size_t max)
 639 {
 640         size_t rand;
 641
 642         if (!max)
 643                 return 0;
 644
 645         do {
 646                 rand = get_random_long();
 647                 rand &= roundup_pow_of_two(max) - 1;
 648         } while (rand >= max);
 649
 650         return rand;
 651 }
 652
 653 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, const void *src)
 654 {
 655         struct bio_vec bv;
 656         struct bvec_iter iter;
 657
 658         __bio_for_each_segment(bv, dst, iter, dst_iter) {
 659                 void *dstp = kmap_local_page(bv.bv_page);
 660
 661                 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
 662                 kunmap_local(dstp);
 663
 664                 src += bv.bv_len;
 665         }
 666 }
 667
 668 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
 669 {
 670         struct bio_vec bv;
 671         struct bvec_iter iter;
 672
 673         __bio_for_each_segment(bv, src, iter, src_iter) {
 674                 void *srcp = kmap_local_page(bv.bv_page);
 675
 676                 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
 677                 kunmap_local(srcp);
 678
 679                 dst += bv.bv_len;
 680         }
 681 }
 682
 683 #if 0
 684 void eytzinger1_test(void)
 685 {
 686         unsigned inorder, eytz, size;
 687
 688         pr_info("1 based eytzinger test:");
 689
 690         for (size = 2;
 691              size < 65536;
 692              size++) {
 693                 unsigned extra = eytzinger1_extra(size);
 694
 695                 if (!(size % 4096))
 696                         pr_info("tree size %u", size);
 697
 698                 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
 699                 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
 700
 701                 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size)    != 0);
 702                 BUG_ON(eytzinger1_next(eytzinger1_last(size), size)     != 0);
 703
 704                 inorder = 1;
 705                 eytzinger1_for_each(eytz, size) {
 706                         BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
 707                         BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
 708                         BUG_ON(eytz != eytzinger1_last(size) &&
 709                                eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
 710
 711                         inorder++;
 712                 }
 713         }
 714 }
 715
 716 void eytzinger0_test(void)
 717 {
 718
 719         unsigned inorder, eytz, size;
 720
 721         pr_info("0 based eytzinger test:");
 722
 723         for (size = 1;
 724              size < 65536;
 725              size++) {
 726                 unsigned extra = eytzinger0_extra(size);
 727
 728                 if (!(size % 4096))
 729                         pr_info("tree size %u", size);
 730
 731                 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
 732                 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
 733
 734                 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size)    != -1);
 735                 BUG_ON(eytzinger0_next(eytzinger0_last(size), size)     != -1);
 736
 737                 inorder = 0;
 738                 eytzinger0_for_each(eytz, size) {
 739                         BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
 740                         BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
 741                         BUG_ON(eytz != eytzinger0_last(size) &&
 742                                eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
 743
 744                         inorder++;
 745                 }
 746         }
 747 }
 748
 749 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
 750 {
 751         const u16 *l = _l, *r = _r;
 752
 753         return (*l > *r) - (*r - *l);
 754 }
 755
 756 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
 757 {
 758         int i, c1 = -1, c2 = -1;
 759         ssize_t r;
 760
 761         r = eytzinger0_find_le(test_array, nr,
 762                                sizeof(test_array[0]),
 763                                cmp_u16, &search);
 764         if (r >= 0)
 765                 c1 = test_array[r];
 766
 767         for (i = 0; i < nr; i++)
 768                 if (test_array[i] <= search && test_array[i] > c2)
 769                         c2 = test_array[i];
 770
 771         if (c1 != c2) {
 772                 eytzinger0_for_each(i, nr)
 773                         pr_info("[%3u] = %12u", i, test_array[i]);
 774                 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
 775                         i, r, c1, c2);
 776         }
 777 }
 778
 779 void eytzinger0_find_test(void)
 780 {
 781         unsigned i, nr, allocated = 1 << 12;
 782         u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
 783
 784         for (nr = 1; nr < allocated; nr++) {
 785                 pr_info("testing %u elems", nr);
 786
 787                 get_random_bytes(test_array, nr * sizeof(test_array[0]));
 788                 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
 789
 790                 /* verify array is sorted correctly: */
 791                 eytzinger0_for_each(i, nr)
 792                         BUG_ON(i != eytzinger0_last(nr) &&
 793                                test_array[i] > test_array[eytzinger0_next(i, nr)]);
 794
 795                 for (i = 0; i < U16_MAX; i += 1 << 12)
 796                         eytzinger0_find_test_val(test_array, nr, i);
 797
 798                 for (i = 0; i < nr; i++) {
 799                         eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
 800                         eytzinger0_find_test_val(test_array, nr, test_array[i]);
 801                         eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
 802                 }
 803         }
 804
 805         kfree(test_array);
 806 }
 807 #endif
 808
 809 /*
 810  * Accumulate percpu counters onto one cpu's copy - only valid when access
 811  * against any percpu counter is guarded against
 812  */
 813 u64 *bch2_acc_percpu_u64s(u64 __percpu *p, unsigned nr)
 814 {
 815         u64 *ret;
 816         int cpu;
 817
 818         /* access to pcpu vars has to be blocked by other locking */
 819         preempt_disable();
 820         ret = this_cpu_ptr(p);
 821         preempt_enable();
 822
 823         for_each_possible_cpu(cpu) {
 824                 u64 *i = per_cpu_ptr(p, cpu);
 825
 826                 if (i != ret) {
 827                         acc_u64s(ret, i, nr);
 828                         memset(i, 0, nr * sizeof(u64));
 829                 }
 830         }
 831
 832         return ret;
 833 }
 834
 835 void bch2_darray_str_exit(darray_str *d)
 836 {
 837         darray_for_each(*d, i)
 838                 kfree(*i);
 839         darray_exit(d);
 840 }
 841
 842 int bch2_split_devs(const char *_dev_name, darray_str *ret)
 843 {
 844         darray_init(ret);
 845
 846         char *dev_name, *s, *orig;
 847
 848         dev_name = orig = kstrdup(_dev_name, GFP_KERNEL);
 849         if (!dev_name)
 850                 return -ENOMEM;
 851
 852         while ((s = strsep(&dev_name, ":"))) {
 853                 char *p = kstrdup(s, GFP_KERNEL);
 854                 if (!p)
 855                         goto err;
 856
 857                 if (darray_push(ret, p)) {
 858                         kfree(p);
 859                         goto err;
 860                 }
 861         }
 862
 863         kfree(orig);
 864         return 0;
 865 err:
 866         bch2_darray_str_exit(ret);
 867         kfree(orig);
 868         return -ENOMEM;
 869 }