git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/util.c

   1 /*
   2  * random utiility code, for bcache but in theory not specific to bcache
   3  *
   4  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include <linux/bio.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/ctype.h>
  11 #include <linux/debugfs.h>
  12 #include <linux/freezer.h>
  13 #include <linux/kthread.h>
  14 #include <linux/log2.h>
  15 #include <linux/math64.h>
  16 #include <linux/percpu.h>
  17 #include <linux/preempt.h>
  18 #include <linux/random.h>
  19 #include <linux/seq_file.h>
  20 #include <linux/string.h>
  21 #include <linux/types.h>
  22 #include <linux/sched/clock.h>
  23
  24 #include "eytzinger.h"
  25 #include "util.h"
  26
  27 #define simple_strtoint(c, end, base)   simple_strtol(c, end, base)
  28 #define simple_strtouint(c, end, base)  simple_strtoul(c, end, base)
  29
  30 #define STRTO_H(name, type)                                     \
  31 int bch2_ ## name ## _h(const char *cp, type *res)              \
  32 {                                                               \
  33         int u = 0;                                              \
  34         char *e;                                                \
  35         type i = simple_ ## name(cp, &e, 10);                   \
  36                                                                 \
  37         switch (tolower(*e)) {                                  \
  38         default:                                                \
  39                 return -EINVAL;                                 \
  40         case 'y':                                               \
  41         case 'z':                                               \
  42                 u++;                                            \
  43         case 'e':                                               \
  44                 u++;                                            \
  45         case 'p':                                               \
  46                 u++;                                            \
  47         case 't':                                               \
  48                 u++;                                            \
  49         case 'g':                                               \
  50                 u++;                                            \
  51         case 'm':                                               \
  52                 u++;                                            \
  53         case 'k':                                               \
  54                 u++;                                            \
  55                 if (e++ == cp)                                  \
  56                         return -EINVAL;                         \
  57         case '\n':                                              \
  58         case '\0':                                              \
  59                 if (*e == '\n')                                 \
  60                         e++;                                    \
  61         }                                                       \
  62                                                                 \
  63         if (*e)                                                 \
  64                 return -EINVAL;                                 \
  65                                                                 \
  66         while (u--) {                                           \
  67                 if ((type) ~0 > 0 &&                            \
  68                     (type) ~0 / 1024 <= i)                      \
  69                         return -EINVAL;                         \
  70                 if ((i > 0 && ANYSINT_MAX(type) / 1024 < i) ||  \
  71                     (i < 0 && -ANYSINT_MAX(type) / 1024 > i))   \
  72                         return -EINVAL;                         \
  73                 i *= 1024;                                      \
  74         }                                                       \
  75                                                                 \
  76         *res = i;                                               \
  77         return 0;                                               \
  78 }                                                               \
  79
  80 STRTO_H(strtoint, int)
  81 STRTO_H(strtouint, unsigned int)
  82 STRTO_H(strtoll, long long)
  83 STRTO_H(strtoull, unsigned long long)
  84
  85 ssize_t bch2_hprint(char *buf, s64 v)
  86 {
  87         static const char units[] = "?kMGTPEZY";
  88         char dec[4] = "";
  89         int u, t = 0;
  90
  91         for (u = 0; v >= 1024 || v <= -1024; u++) {
  92                 t = v & ~(~0U << 10);
  93                 v >>= 10;
  94         }
  95
  96         if (!u)
  97                 return sprintf(buf, "%lli", v);
  98
  99         /*
 100          * 103 is magic: t is in the range [-1023, 1023] and we want
 101          * to turn it into [-9, 9]
 102          */
 103         if (v < 100 && v > -100)
 104                 scnprintf(dec, sizeof(dec), ".%i", t / 103);
 105
 106         return sprintf(buf, "%lli%s%c", v, dec, units[u]);
 107 }
 108
 109 ssize_t bch2_scnprint_string_list(char *buf, size_t size,
 110                                   const char * const list[],
 111                                   size_t selected)
 112 {
 113         char *out = buf;
 114         size_t i;
 115
 116         if (size)
 117                 *out = '\0';
 118
 119         for (i = 0; list[i]; i++)
 120                 out += scnprintf(out, buf + size - out,
 121                                  i == selected ? "[%s] " : "%s ", list[i]);
 122
 123         if (out != buf)
 124                 *--out = '\0';
 125
 126         return out - buf;
 127 }
 128
 129 ssize_t bch2_read_string_list(const char *buf, const char * const list[])
 130 {
 131         size_t i, len;
 132
 133         buf = skip_spaces(buf);
 134
 135         len = strlen(buf);
 136         while (len && isspace(buf[len - 1]))
 137                 --len;
 138
 139         for (i = 0; list[i]; i++)
 140                 if (strlen(list[i]) == len &&
 141                     !memcmp(buf, list[i], len))
 142                         break;
 143
 144         return list[i] ? i : -EINVAL;
 145 }
 146
 147 ssize_t bch2_scnprint_flag_list(char *buf, size_t size,
 148                                 const char * const list[], u64 flags)
 149 {
 150         char *out = buf, *end = buf + size;
 151         unsigned bit, nr = 0;
 152
 153         while (list[nr])
 154                 nr++;
 155
 156         if (size)
 157                 *out = '\0';
 158
 159         while (flags && (bit = __ffs(flags)) < nr) {
 160                 out += scnprintf(out, end - out, "%s,", list[bit]);
 161                 flags ^= 1 << bit;
 162         }
 163
 164         if (out != buf)
 165                 *--out = '\0';
 166
 167         return out - buf;
 168 }
 169
 170 u64 bch2_read_flag_list(char *opt, const char * const list[])
 171 {
 172         u64 ret = 0;
 173         char *p, *s, *d = kstrndup(opt, PAGE_SIZE - 1, GFP_KERNEL);
 174
 175         if (!d)
 176                 return -ENOMEM;
 177
 178         s = strim(d);
 179
 180         while ((p = strsep(&s, ","))) {
 181                 int flag = bch2_read_string_list(p, list);
 182                 if (flag < 0) {
 183                         ret = -1;
 184                         break;
 185                 }
 186
 187                 ret |= 1 << flag;
 188         }
 189
 190         kfree(d);
 191
 192         return ret;
 193 }
 194
 195 bool bch2_is_zero(const void *_p, size_t n)
 196 {
 197         const char *p = _p;
 198         size_t i;
 199
 200         for (i = 0; i < n; i++)
 201                 if (p[i])
 202                         return false;
 203         return true;
 204 }
 205
 206 void bch2_quantiles_update(struct quantiles *q, u64 v)
 207 {
 208         unsigned i = 0;
 209
 210         while (i < ARRAY_SIZE(q->entries)) {
 211                 struct quantile_entry *e = q->entries + i;
 212
 213                 if (unlikely(!e->step)) {
 214                         e->m = v;
 215                         e->step = max_t(unsigned, v / 2, 1024);
 216                 } else if (e->m > v) {
 217                         e->m = e->m >= e->step
 218                                 ? e->m - e->step
 219                                 : 0;
 220                 } else if (e->m < v) {
 221                         e->m = e->m + e->step > e->m
 222                                 ? e->m + e->step
 223                                 : U32_MAX;
 224                 }
 225
 226                 if ((e->m > v ? e->m - v : v - e->m) < e->step)
 227                         e->step = max_t(unsigned, e->step / 2, 1);
 228
 229                 if (v >= e->m)
 230                         break;
 231
 232                 i = eytzinger0_child(i, v > e->m);
 233         }
 234 }
 235
 236 /* time stats: */
 237
 238 static void bch2_time_stats_update_one(struct time_stats *stats,
 239                                        u64 start, u64 end)
 240 {
 241         u64 duration, freq;
 242
 243         duration        = time_after64(end, start)
 244                 ? end - start : 0;
 245         freq            = time_after64(end, stats->last_event)
 246                 ? end - stats->last_event : 0;
 247
 248         stats->count++;
 249
 250         stats->average_duration = stats->average_duration
 251                 ? ewma_add(stats->average_duration, duration, 6)
 252                 : duration;
 253
 254         stats->average_frequency = stats->average_frequency
 255                 ? ewma_add(stats->average_frequency, freq, 6)
 256                 : freq;
 257
 258         stats->max_duration = max(stats->max_duration, duration);
 259
 260         stats->last_event = end;
 261
 262         bch2_quantiles_update(&stats->quantiles, duration);
 263 }
 264
 265 void __bch2_time_stats_update(struct time_stats *stats, u64 start, u64 end)
 266 {
 267         unsigned long flags;
 268
 269         if (!stats->buffer) {
 270                 spin_lock_irqsave(&stats->lock, flags);
 271                 bch2_time_stats_update_one(stats, start, end);
 272
 273                 if (stats->average_frequency < 32 &&
 274                     stats->count > 1024)
 275                         stats->buffer =
 276                                 alloc_percpu_gfp(struct time_stat_buffer,
 277                                                  GFP_ATOMIC);
 278                 spin_unlock_irqrestore(&stats->lock, flags);
 279         } else {
 280                 struct time_stat_buffer_entry *i;
 281                 struct time_stat_buffer *b;
 282
 283                 preempt_disable();
 284                 b = this_cpu_ptr(stats->buffer);
 285
 286                 BUG_ON(b->nr >= ARRAY_SIZE(b->entries));
 287                 b->entries[b->nr++] = (struct time_stat_buffer_entry) {
 288                         .start = start,
 289                         .end = end
 290                 };
 291
 292                 if (b->nr == ARRAY_SIZE(b->entries)) {
 293                         spin_lock_irqsave(&stats->lock, flags);
 294                         for (i = b->entries;
 295                              i < b->entries + ARRAY_SIZE(b->entries);
 296                              i++)
 297                                 bch2_time_stats_update_one(stats, i->start, i->end);
 298                         spin_unlock_irqrestore(&stats->lock, flags);
 299
 300                         b->nr = 0;
 301                 }
 302
 303                 preempt_enable();
 304         }
 305 }
 306
 307 static const struct time_unit {
 308         const char      *name;
 309         u32             nsecs;
 310 } time_units[] = {
 311         { "ns",         1               },
 312         { "us",         NSEC_PER_USEC   },
 313         { "ms",         NSEC_PER_MSEC   },
 314         { "sec",        NSEC_PER_SEC    },
 315 };
 316
 317 static const struct time_unit *pick_time_units(u64 ns)
 318 {
 319         const struct time_unit *u;
 320
 321         for (u = time_units;
 322              u + 1 < time_units + ARRAY_SIZE(time_units) &&
 323              ns >= u[1].nsecs << 1;
 324              u++)
 325                 ;
 326
 327         return u;
 328 }
 329
 330 static size_t pr_time_units(char *buf, size_t len, u64 ns)
 331 {
 332         const struct time_unit *u = pick_time_units(ns);
 333
 334         return scnprintf(buf, len, "%llu %s", div_u64(ns, u->nsecs), u->name);
 335 }
 336
 337 size_t bch2_time_stats_print(struct time_stats *stats, char *buf, size_t len)
 338 {
 339         char *out = buf, *end = buf + len;
 340         const struct time_unit *u;
 341         u64 freq = READ_ONCE(stats->average_frequency);
 342         u64 q, last_q = 0;
 343         int i;
 344
 345         out += scnprintf(out, end - out, "count:\t\t%llu\n",
 346                          stats->count);
 347         out += scnprintf(out, end - out, "rate:\t\t%llu/sec\n",
 348                          freq ?  div64_u64(NSEC_PER_SEC, freq) : 0);
 349
 350         out += scnprintf(out, end - out, "frequency:\t");
 351         out += pr_time_units(out, end - out, freq);
 352
 353         out += scnprintf(out, end - out, "\navg duration:\t");
 354         out += pr_time_units(out, end - out, stats->average_duration);
 355
 356         out += scnprintf(out, end - out, "\nmax duration:\t");
 357         out += pr_time_units(out, end - out, stats->max_duration);
 358
 359         i = eytzinger0_first(NR_QUANTILES);
 360         u = pick_time_units(stats->quantiles.entries[i].m);
 361
 362         out += scnprintf(out, end - out, "\nquantiles (%s):\t", u->name);
 363         eytzinger0_for_each(i, NR_QUANTILES) {
 364                 bool is_last = eytzinger0_next(i, NR_QUANTILES) == -1;
 365
 366                 q = max(stats->quantiles.entries[i].m, last_q);
 367                 out += scnprintf(out, end - out, "%llu%s",
 368                                  div_u64(q, u->nsecs),
 369                                  is_last ? "\n" : " ");
 370                 last_q = q;
 371         }
 372
 373         return out - buf;
 374 }
 375
 376 void bch2_time_stats_exit(struct time_stats *stats)
 377 {
 378         free_percpu(stats->buffer);
 379 }
 380
 381 void bch2_time_stats_init(struct time_stats *stats)
 382 {
 383         memset(stats, 0, sizeof(*stats));
 384         spin_lock_init(&stats->lock);
 385 }
 386
 387 /* ratelimit: */
 388
 389 /**
 390  * bch2_ratelimit_delay() - return how long to delay until the next time to do
 391  * some work
 392  *
 393  * @d - the struct bch_ratelimit to update
 394  *
 395  * Returns the amount of time to delay by, in jiffies
 396  */
 397 u64 bch2_ratelimit_delay(struct bch_ratelimit *d)
 398 {
 399         u64 now = local_clock();
 400
 401         return time_after64(d->next, now)
 402                 ? nsecs_to_jiffies(d->next - now)
 403                 : 0;
 404 }
 405
 406 /**
 407  * bch2_ratelimit_increment() - increment @d by the amount of work done
 408  *
 409  * @d - the struct bch_ratelimit to update
 410  * @done - the amount of work done, in arbitrary units
 411  */
 412 void bch2_ratelimit_increment(struct bch_ratelimit *d, u64 done)
 413 {
 414         u64 now = local_clock();
 415
 416         d->next += div_u64(done * NSEC_PER_SEC, d->rate);
 417
 418         if (time_before64(now + NSEC_PER_SEC, d->next))
 419                 d->next = now + NSEC_PER_SEC;
 420
 421         if (time_after64(now - NSEC_PER_SEC * 2, d->next))
 422                 d->next = now - NSEC_PER_SEC * 2;
 423 }
 424
 425 int bch2_ratelimit_wait_freezable_stoppable(struct bch_ratelimit *d)
 426 {
 427         bool kthread = (current->flags & PF_KTHREAD) != 0;
 428
 429         while (1) {
 430                 u64 delay = bch2_ratelimit_delay(d);
 431
 432                 if (delay)
 433                         set_current_state(TASK_INTERRUPTIBLE);
 434
 435                 if (kthread && kthread_should_stop())
 436                         return 1;
 437
 438                 if (!delay)
 439                         return 0;
 440
 441                 schedule_timeout(delay);
 442                 try_to_freeze();
 443         }
 444 }
 445
 446 /* pd controller: */
 447
 448 /*
 449  * Updates pd_controller. Attempts to scale inputed values to units per second.
 450  * @target: desired value
 451  * @actual: current value
 452  *
 453  * @sign: 1 or -1; 1 if increasing the rate makes actual go up, -1 if increasing
 454  * it makes actual go down.
 455  */
 456 void bch2_pd_controller_update(struct bch_pd_controller *pd,
 457                               s64 target, s64 actual, int sign)
 458 {
 459         s64 proportional, derivative, change;
 460
 461         unsigned long seconds_since_update = (jiffies - pd->last_update) / HZ;
 462
 463         if (seconds_since_update == 0)
 464                 return;
 465
 466         pd->last_update = jiffies;
 467
 468         proportional = actual - target;
 469         proportional *= seconds_since_update;
 470         proportional = div_s64(proportional, pd->p_term_inverse);
 471
 472         derivative = actual - pd->last_actual;
 473         derivative = div_s64(derivative, seconds_since_update);
 474         derivative = ewma_add(pd->smoothed_derivative, derivative,
 475                               (pd->d_term / seconds_since_update) ?: 1);
 476         derivative = derivative * pd->d_term;
 477         derivative = div_s64(derivative, pd->p_term_inverse);
 478
 479         change = proportional + derivative;
 480
 481         /* Don't increase rate if not keeping up */
 482         if (change > 0 &&
 483             pd->backpressure &&
 484             time_after64(local_clock(),
 485                          pd->rate.next + NSEC_PER_MSEC))
 486                 change = 0;
 487
 488         change *= (sign * -1);
 489
 490         pd->rate.rate = clamp_t(s64, (s64) pd->rate.rate + change,
 491                                 1, UINT_MAX);
 492
 493         pd->last_actual         = actual;
 494         pd->last_derivative     = derivative;
 495         pd->last_proportional   = proportional;
 496         pd->last_change         = change;
 497         pd->last_target         = target;
 498 }
 499
 500 void bch2_pd_controller_init(struct bch_pd_controller *pd)
 501 {
 502         pd->rate.rate           = 1024;
 503         pd->last_update         = jiffies;
 504         pd->p_term_inverse      = 6000;
 505         pd->d_term              = 30;
 506         pd->d_smooth            = pd->d_term;
 507         pd->backpressure        = 1;
 508 }
 509
 510 size_t bch2_pd_controller_print_debug(struct bch_pd_controller *pd, char *buf)
 511 {
 512         /* 2^64 - 1 is 20 digits, plus null byte */
 513         char rate[21];
 514         char actual[21];
 515         char target[21];
 516         char proportional[21];
 517         char derivative[21];
 518         char change[21];
 519         s64 next_io;
 520
 521         bch2_hprint(rate,       pd->rate.rate);
 522         bch2_hprint(actual,     pd->last_actual);
 523         bch2_hprint(target,     pd->last_target);
 524         bch2_hprint(proportional, pd->last_proportional);
 525         bch2_hprint(derivative, pd->last_derivative);
 526         bch2_hprint(change,     pd->last_change);
 527
 528         next_io = div64_s64(pd->rate.next - local_clock(), NSEC_PER_MSEC);
 529
 530         return sprintf(buf,
 531                        "rate:\t\t%s/sec\n"
 532                        "target:\t\t%s\n"
 533                        "actual:\t\t%s\n"
 534                        "proportional:\t%s\n"
 535                        "derivative:\t%s\n"
 536                        "change:\t\t%s/sec\n"
 537                        "next io:\t%llims\n",
 538                        rate, target, actual, proportional,
 539                        derivative, change, next_io);
 540 }
 541
 542 /* misc: */
 543
 544 void bch2_bio_map(struct bio *bio, void *base)
 545 {
 546         size_t size = bio->bi_iter.bi_size;
 547         struct bio_vec *bv = bio->bi_io_vec;
 548
 549         BUG_ON(!bio->bi_iter.bi_size);
 550         BUG_ON(bio->bi_vcnt);
 551
 552         bv->bv_offset = base ? offset_in_page(base) : 0;
 553         goto start;
 554
 555         for (; size; bio->bi_vcnt++, bv++) {
 556                 bv->bv_offset   = 0;
 557 start:          bv->bv_len      = min_t(size_t, PAGE_SIZE - bv->bv_offset,
 558                                         size);
 559                 BUG_ON(bio->bi_vcnt >= bio->bi_max_vecs);
 560                 if (base) {
 561                         bv->bv_page = is_vmalloc_addr(base)
 562                                 ? vmalloc_to_page(base)
 563                                 : virt_to_page(base);
 564
 565                         base += bv->bv_len;
 566                 }
 567
 568                 size -= bv->bv_len;
 569         }
 570 }
 571
 572 size_t bch2_rand_range(size_t max)
 573 {
 574         size_t rand;
 575
 576         if (!max)
 577                 return 0;
 578
 579         do {
 580                 rand = get_random_long();
 581                 rand &= roundup_pow_of_two(max) - 1;
 582         } while (rand >= max);
 583
 584         return rand;
 585 }
 586
 587 void memcpy_to_bio(struct bio *dst, struct bvec_iter dst_iter, void *src)
 588 {
 589         struct bio_vec bv;
 590         struct bvec_iter iter;
 591
 592         __bio_for_each_segment(bv, dst, iter, dst_iter) {
 593                 void *dstp = kmap_atomic(bv.bv_page);
 594                 memcpy(dstp + bv.bv_offset, src, bv.bv_len);
 595                 kunmap_atomic(dstp);
 596
 597                 src += bv.bv_len;
 598         }
 599 }
 600
 601 void memcpy_from_bio(void *dst, struct bio *src, struct bvec_iter src_iter)
 602 {
 603         struct bio_vec bv;
 604         struct bvec_iter iter;
 605
 606         __bio_for_each_segment(bv, src, iter, src_iter) {
 607                 void *srcp = kmap_atomic(bv.bv_page);
 608                 memcpy(dst, srcp + bv.bv_offset, bv.bv_len);
 609                 kunmap_atomic(srcp);
 610
 611                 dst += bv.bv_len;
 612         }
 613 }
 614
 615 size_t bch_scnmemcpy(char *buf, size_t size, const char *src, size_t len)
 616 {
 617         size_t n;
 618
 619         if (!size)
 620                 return 0;
 621
 622         n = min(size - 1, len);
 623         memcpy(buf, src, n);
 624         buf[n] = '\0';
 625
 626         return n;
 627 }
 628
 629 #include "eytzinger.h"
 630
 631 static int alignment_ok(const void *base, size_t align)
 632 {
 633         return IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) ||
 634                 ((unsigned long)base & (align - 1)) == 0;
 635 }
 636
 637 static void u32_swap(void *a, void *b, size_t size)
 638 {
 639         u32 t = *(u32 *)a;
 640         *(u32 *)a = *(u32 *)b;
 641         *(u32 *)b = t;
 642 }
 643
 644 static void u64_swap(void *a, void *b, size_t size)
 645 {
 646         u64 t = *(u64 *)a;
 647         *(u64 *)a = *(u64 *)b;
 648         *(u64 *)b = t;
 649 }
 650
 651 static void generic_swap(void *a, void *b, size_t size)
 652 {
 653         char t;
 654
 655         do {
 656                 t = *(char *)a;
 657                 *(char *)a++ = *(char *)b;
 658                 *(char *)b++ = t;
 659         } while (--size > 0);
 660 }
 661
 662 static inline int do_cmp(void *base, size_t n, size_t size,
 663                          int (*cmp_func)(const void *, const void *, size_t),
 664                          size_t l, size_t r)
 665 {
 666         return cmp_func(base + inorder_to_eytzinger0(l, n) * size,
 667                         base + inorder_to_eytzinger0(r, n) * size,
 668                         size);
 669 }
 670
 671 static inline void do_swap(void *base, size_t n, size_t size,
 672                            void (*swap_func)(void *, void *, size_t),
 673                            size_t l, size_t r)
 674 {
 675         swap_func(base + inorder_to_eytzinger0(l, n) * size,
 676                   base + inorder_to_eytzinger0(r, n) * size,
 677                   size);
 678 }
 679
 680 void eytzinger0_sort(void *base, size_t n, size_t size,
 681                      int (*cmp_func)(const void *, const void *, size_t),
 682                      void (*swap_func)(void *, void *, size_t))
 683 {
 684         int i, c, r;
 685
 686         if (!swap_func) {
 687                 if (size == 4 && alignment_ok(base, 4))
 688                         swap_func = u32_swap;
 689                 else if (size == 8 && alignment_ok(base, 8))
 690                         swap_func = u64_swap;
 691                 else
 692                         swap_func = generic_swap;
 693         }
 694
 695         /* heapify */
 696         for (i = n / 2 - 1; i >= 0; --i) {
 697                 for (r = i; r * 2 + 1 < n; r = c) {
 698                         c = r * 2 + 1;
 699
 700                         if (c + 1 < n &&
 701                             do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
 702                                 c++;
 703
 704                         if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
 705                                 break;
 706
 707                         do_swap(base, n, size, swap_func, r, c);
 708                 }
 709         }
 710
 711         /* sort */
 712         for (i = n - 1; i > 0; --i) {
 713                 do_swap(base, n, size, swap_func, 0, i);
 714
 715                 for (r = 0; r * 2 + 1 < i; r = c) {
 716                         c = r * 2 + 1;
 717
 718                         if (c + 1 < i &&
 719                             do_cmp(base, n, size, cmp_func, c, c + 1) < 0)
 720                                 c++;
 721
 722                         if (do_cmp(base, n, size, cmp_func, r, c) >= 0)
 723                                 break;
 724
 725                         do_swap(base, n, size, swap_func, r, c);
 726                 }
 727         }
 728 }
 729
 730 void sort_cmp_size(void *base, size_t num, size_t size,
 731           int (*cmp_func)(const void *, const void *, size_t),
 732           void (*swap_func)(void *, void *, size_t size))
 733 {
 734         /* pre-scale counters for performance */
 735         int i = (num/2 - 1) * size, n = num * size, c, r;
 736
 737         if (!swap_func) {
 738                 if (size == 4 && alignment_ok(base, 4))
 739                         swap_func = u32_swap;
 740                 else if (size == 8 && alignment_ok(base, 8))
 741                         swap_func = u64_swap;
 742                 else
 743                         swap_func = generic_swap;
 744         }
 745
 746         /* heapify */
 747         for ( ; i >= 0; i -= size) {
 748                 for (r = i; r * 2 + size < n; r  = c) {
 749                         c = r * 2 + size;
 750                         if (c < n - size &&
 751                             cmp_func(base + c, base + c + size, size) < 0)
 752                                 c += size;
 753                         if (cmp_func(base + r, base + c, size) >= 0)
 754                                 break;
 755                         swap_func(base + r, base + c, size);
 756                 }
 757         }
 758
 759         /* sort */
 760         for (i = n - size; i > 0; i -= size) {
 761                 swap_func(base, base + i, size);
 762                 for (r = 0; r * 2 + size < i; r = c) {
 763                         c = r * 2 + size;
 764                         if (c < i - size &&
 765                             cmp_func(base + c, base + c + size, size) < 0)
 766                                 c += size;
 767                         if (cmp_func(base + r, base + c, size) >= 0)
 768                                 break;
 769                         swap_func(base + r, base + c, size);
 770                 }
 771         }
 772 }
 773
 774 void mempool_free_vp(void *element, void *pool_data)
 775 {
 776         size_t size = (size_t) pool_data;
 777
 778         vpfree(element, size);
 779 }
 780
 781 void *mempool_alloc_vp(gfp_t gfp_mask, void *pool_data)
 782 {
 783         size_t size = (size_t) pool_data;
 784
 785         return vpmalloc(size, gfp_mask);
 786 }
 787
 788 #if 0
 789 void eytzinger1_test(void)
 790 {
 791         unsigned inorder, eytz, size;
 792
 793         pr_info("1 based eytzinger test:");
 794
 795         for (size = 2;
 796              size < 65536;
 797              size++) {
 798                 unsigned extra = eytzinger1_extra(size);
 799
 800                 if (!(size % 4096))
 801                         pr_info("tree size %u", size);
 802
 803                 BUG_ON(eytzinger1_prev(0, size) != eytzinger1_last(size));
 804                 BUG_ON(eytzinger1_next(0, size) != eytzinger1_first(size));
 805
 806                 BUG_ON(eytzinger1_prev(eytzinger1_first(size), size)    != 0);
 807                 BUG_ON(eytzinger1_next(eytzinger1_last(size), size)     != 0);
 808
 809                 inorder = 1;
 810                 eytzinger1_for_each(eytz, size) {
 811                         BUG_ON(__inorder_to_eytzinger1(inorder, size, extra) != eytz);
 812                         BUG_ON(__eytzinger1_to_inorder(eytz, size, extra) != inorder);
 813                         BUG_ON(eytz != eytzinger1_last(size) &&
 814                                eytzinger1_prev(eytzinger1_next(eytz, size), size) != eytz);
 815
 816                         inorder++;
 817                 }
 818         }
 819 }
 820
 821 void eytzinger0_test(void)
 822 {
 823
 824         unsigned inorder, eytz, size;
 825
 826         pr_info("0 based eytzinger test:");
 827
 828         for (size = 1;
 829              size < 65536;
 830              size++) {
 831                 unsigned extra = eytzinger0_extra(size);
 832
 833                 if (!(size % 4096))
 834                         pr_info("tree size %u", size);
 835
 836                 BUG_ON(eytzinger0_prev(-1, size) != eytzinger0_last(size));
 837                 BUG_ON(eytzinger0_next(-1, size) != eytzinger0_first(size));
 838
 839                 BUG_ON(eytzinger0_prev(eytzinger0_first(size), size)    != -1);
 840                 BUG_ON(eytzinger0_next(eytzinger0_last(size), size)     != -1);
 841
 842                 inorder = 0;
 843                 eytzinger0_for_each(eytz, size) {
 844                         BUG_ON(__inorder_to_eytzinger0(inorder, size, extra) != eytz);
 845                         BUG_ON(__eytzinger0_to_inorder(eytz, size, extra) != inorder);
 846                         BUG_ON(eytz != eytzinger0_last(size) &&
 847                                eytzinger0_prev(eytzinger0_next(eytz, size), size) != eytz);
 848
 849                         inorder++;
 850                 }
 851         }
 852 }
 853
 854 static inline int cmp_u16(const void *_l, const void *_r, size_t size)
 855 {
 856         const u16 *l = _l, *r = _r;
 857
 858         return (*l > *r) - (*r - *l);
 859 }
 860
 861 static void eytzinger0_find_test_val(u16 *test_array, unsigned nr, u16 search)
 862 {
 863         int i, c1 = -1, c2 = -1;
 864         ssize_t r;
 865
 866         r = eytzinger0_find_le(test_array, nr,
 867                                sizeof(test_array[0]),
 868                                cmp_u16, &search);
 869         if (r >= 0)
 870                 c1 = test_array[r];
 871
 872         for (i = 0; i < nr; i++)
 873                 if (test_array[i] <= search && test_array[i] > c2)
 874                         c2 = test_array[i];
 875
 876         if (c1 != c2) {
 877                 eytzinger0_for_each(i, nr)
 878                         pr_info("[%3u] = %12u", i, test_array[i]);
 879                 pr_info("find_le(%2u) -> [%2zi] = %2i should be %2i",
 880                         i, r, c1, c2);
 881         }
 882 }
 883
 884 void eytzinger0_find_test(void)
 885 {
 886         unsigned i, nr, allocated = 1 << 12;
 887         u16 *test_array = kmalloc_array(allocated, sizeof(test_array[0]), GFP_KERNEL);
 888
 889         for (nr = 1; nr < allocated; nr++) {
 890                 pr_info("testing %u elems", nr);
 891
 892                 get_random_bytes(test_array, nr * sizeof(test_array[0]));
 893                 eytzinger0_sort(test_array, nr, sizeof(test_array[0]), cmp_u16, NULL);
 894
 895                 /* verify array is sorted correctly: */
 896                 eytzinger0_for_each(i, nr)
 897                         BUG_ON(i != eytzinger0_last(nr) &&
 898                                test_array[i] > test_array[eytzinger0_next(i, nr)]);
 899
 900                 for (i = 0; i < U16_MAX; i += 1 << 12)
 901                         eytzinger0_find_test_val(test_array, nr, i);
 902
 903                 for (i = 0; i < nr; i++) {
 904                         eytzinger0_find_test_val(test_array, nr, test_array[i] - 1);
 905                         eytzinger0_find_test_val(test_array, nr, test_array[i]);
 906                         eytzinger0_find_test_val(test_array, nr, test_array[i] + 1);
 907                 }
 908         }
 909
 910         kfree(test_array);
 911 }
 912 #endif