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

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _BCACHEFS_UTIL_H
   3 #define _BCACHEFS_UTIL_H
   4
   5 #include <linux/bio.h>
   6 #include <linux/blkdev.h>
   7 #include <linux/closure.h>
   8 #include <linux/errno.h>
   9 #include <linux/freezer.h>
  10 #include <linux/kernel.h>
  11 #include <linux/sched/clock.h>
  12 #include <linux/llist.h>
  13 #include <linux/log2.h>
  14 #include <linux/printbuf.h>
  15 #include <linux/percpu.h>
  16 #include <linux/preempt.h>
  17 #include <linux/ratelimit.h>
  18 #include <linux/slab.h>
  19 #include <linux/vmalloc.h>
  20 #include <linux/workqueue.h>
  21 #include <linux/mean_and_variance.h>
  22
  23 struct closure;
  24
  25 #ifdef CONFIG_BCACHEFS_DEBUG
  26
  27 #define EBUG_ON(cond)           BUG_ON(cond)
  28 #define atomic_dec_bug(v)       BUG_ON(atomic_dec_return(v) < 0)
  29 #define atomic_inc_bug(v, i)    BUG_ON(atomic_inc_return(v) <= i)
  30 #define atomic_sub_bug(i, v)    BUG_ON(atomic_sub_return(i, v) < 0)
  31 #define atomic_add_bug(i, v)    BUG_ON(atomic_add_return(i, v) < 0)
  32 #define atomic_long_dec_bug(v)          BUG_ON(atomic_long_dec_return(v) < 0)
  33 #define atomic_long_sub_bug(i, v)       BUG_ON(atomic_long_sub_return(i, v) < 0)
  34 #define atomic64_dec_bug(v)     BUG_ON(atomic64_dec_return(v) < 0)
  35 #define atomic64_inc_bug(v, i)  BUG_ON(atomic64_inc_return(v) <= i)
  36 #define atomic64_sub_bug(i, v)  BUG_ON(atomic64_sub_return(i, v) < 0)
  37 #define atomic64_add_bug(i, v)  BUG_ON(atomic64_add_return(i, v) < 0)
  38
  39 #else /* DEBUG */
  40
  41 #define EBUG_ON(cond)
  42 #define atomic_dec_bug(v)       atomic_dec(v)
  43 #define atomic_inc_bug(v, i)    atomic_inc(v)
  44 #define atomic_sub_bug(i, v)    atomic_sub(i, v)
  45 #define atomic_add_bug(i, v)    atomic_add(i, v)
  46 #define atomic_long_dec_bug(v)          atomic_long_dec(v)
  47 #define atomic_long_sub_bug(i, v)       atomic_long_sub(i, v)
  48 #define atomic64_dec_bug(v)     atomic64_dec(v)
  49 #define atomic64_inc_bug(v, i)  atomic64_inc(v)
  50 #define atomic64_sub_bug(i, v)  atomic64_sub(i, v)
  51 #define atomic64_add_bug(i, v)  atomic64_add(i, v)
  52
  53 #endif
  54
  55 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  56 #define CPU_BIG_ENDIAN          0
  57 #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
  58 #define CPU_BIG_ENDIAN          1
  59 #endif
  60
  61 /* type hackery */
  62
  63 #define type_is_exact(_val, _type)                                      \
  64         __builtin_types_compatible_p(typeof(_val), _type)
  65
  66 #define type_is(_val, _type)                                            \
  67         (__builtin_types_compatible_p(typeof(_val), _type) ||           \
  68          __builtin_types_compatible_p(typeof(_val), const _type))
  69
  70 /* Userspace doesn't align allocations as nicely as the kernel allocators: */
  71 static inline size_t buf_pages(void *p, size_t len)
  72 {
  73         return DIV_ROUND_UP(len +
  74                             ((unsigned long) p & (PAGE_SIZE - 1)),
  75                             PAGE_SIZE);
  76 }
  77
  78 static inline void vpfree(void *p, size_t size)
  79 {
  80         if (is_vmalloc_addr(p))
  81                 vfree(p);
  82         else
  83                 free_pages((unsigned long) p, get_order(size));
  84 }
  85
  86 static inline void *vpmalloc(size_t size, gfp_t gfp_mask)
  87 {
  88         return (void *) __get_free_pages(gfp_mask|__GFP_NOWARN,
  89                                          get_order(size)) ?:
  90                 __vmalloc(size, gfp_mask);
  91 }
  92
  93 static inline void kvpfree(void *p, size_t size)
  94 {
  95         if (size < PAGE_SIZE)
  96                 kfree(p);
  97         else
  98                 vpfree(p, size);
  99 }
 100
 101 static inline void *kvpmalloc(size_t size, gfp_t gfp_mask)
 102 {
 103         return size < PAGE_SIZE
 104                 ? kmalloc(size, gfp_mask)
 105                 : vpmalloc(size, gfp_mask);
 106 }
 107
 108 int mempool_init_kvpmalloc_pool(mempool_t *, int, size_t);
 109
 110 #define HEAP(type)                                                      \
 111 struct {                                                                \
 112         size_t size, used;                                              \
 113         type *data;                                                     \
 114 }
 115
 116 #define DECLARE_HEAP(type, name) HEAP(type) name
 117
 118 #define init_heap(heap, _size, gfp)                                     \
 119 ({                                                                      \
 120         (heap)->used = 0;                                               \
 121         (heap)->size = (_size);                                         \
 122         (heap)->data = kvpmalloc((heap)->size * sizeof((heap)->data[0]),\
 123                                  (gfp));                                \
 124 })
 125
 126 #define free_heap(heap)                                                 \
 127 do {                                                                    \
 128         kvpfree((heap)->data, (heap)->size * sizeof((heap)->data[0]));  \
 129         (heap)->data = NULL;                                            \
 130 } while (0)
 131
 132 #define heap_set_backpointer(h, i, _fn)                                 \
 133 do {                                                                    \
 134         void (*fn)(typeof(h), size_t) = _fn;                            \
 135         if (fn)                                                         \
 136                 fn(h, i);                                               \
 137 } while (0)
 138
 139 #define heap_swap(h, i, j, set_backpointer)                             \
 140 do {                                                                    \
 141         swap((h)->data[i], (h)->data[j]);                               \
 142         heap_set_backpointer(h, i, set_backpointer);                    \
 143         heap_set_backpointer(h, j, set_backpointer);                    \
 144 } while (0)
 145
 146 #define heap_peek(h)                                                    \
 147 ({                                                                      \
 148         EBUG_ON(!(h)->used);                                            \
 149         (h)->data[0];                                                   \
 150 })
 151
 152 #define heap_full(h)    ((h)->used == (h)->size)
 153
 154 #define heap_sift_down(h, i, cmp, set_backpointer)                      \
 155 do {                                                                    \
 156         size_t _c, _j = i;                                              \
 157                                                                         \
 158         for (; _j * 2 + 1 < (h)->used; _j = _c) {                       \
 159                 _c = _j * 2 + 1;                                        \
 160                 if (_c + 1 < (h)->used &&                               \
 161                     cmp(h, (h)->data[_c], (h)->data[_c + 1]) >= 0)      \
 162                         _c++;                                           \
 163                                                                         \
 164                 if (cmp(h, (h)->data[_c], (h)->data[_j]) >= 0)          \
 165                         break;                                          \
 166                 heap_swap(h, _c, _j, set_backpointer);                  \
 167         }                                                               \
 168 } while (0)
 169
 170 #define heap_sift_up(h, i, cmp, set_backpointer)                        \
 171 do {                                                                    \
 172         while (i) {                                                     \
 173                 size_t p = (i - 1) / 2;                                 \
 174                 if (cmp(h, (h)->data[i], (h)->data[p]) >= 0)            \
 175                         break;                                          \
 176                 heap_swap(h, i, p, set_backpointer);                    \
 177                 i = p;                                                  \
 178         }                                                               \
 179 } while (0)
 180
 181 #define __heap_add(h, d, cmp, set_backpointer)                          \
 182 ({                                                                      \
 183         size_t _i = (h)->used++;                                        \
 184         (h)->data[_i] = d;                                              \
 185         heap_set_backpointer(h, _i, set_backpointer);                   \
 186                                                                         \
 187         heap_sift_up(h, _i, cmp, set_backpointer);                      \
 188         _i;                                                             \
 189 })
 190
 191 #define heap_add(h, d, cmp, set_backpointer)                            \
 192 ({                                                                      \
 193         bool _r = !heap_full(h);                                        \
 194         if (_r)                                                         \
 195                 __heap_add(h, d, cmp, set_backpointer);                 \
 196         _r;                                                             \
 197 })
 198
 199 #define heap_add_or_replace(h, new, cmp, set_backpointer)               \
 200 do {                                                                    \
 201         if (!heap_add(h, new, cmp, set_backpointer) &&                  \
 202             cmp(h, new, heap_peek(h)) >= 0) {                           \
 203                 (h)->data[0] = new;                                     \
 204                 heap_set_backpointer(h, 0, set_backpointer);            \
 205                 heap_sift_down(h, 0, cmp, set_backpointer);             \
 206         }                                                               \
 207 } while (0)
 208
 209 #define heap_del(h, i, cmp, set_backpointer)                            \
 210 do {                                                                    \
 211         size_t _i = (i);                                                \
 212                                                                         \
 213         BUG_ON(_i >= (h)->used);                                        \
 214         (h)->used--;                                                    \
 215         if ((_i) < (h)->used) {                                         \
 216                 heap_swap(h, _i, (h)->used, set_backpointer);           \
 217                 heap_sift_up(h, _i, cmp, set_backpointer);              \
 218                 heap_sift_down(h, _i, cmp, set_backpointer);            \
 219         }                                                               \
 220 } while (0)
 221
 222 #define heap_pop(h, d, cmp, set_backpointer)                            \
 223 ({                                                                      \
 224         bool _r = (h)->used;                                            \
 225         if (_r) {                                                       \
 226                 (d) = (h)->data[0];                                     \
 227                 heap_del(h, 0, cmp, set_backpointer);                   \
 228         }                                                               \
 229         _r;                                                             \
 230 })
 231
 232 #define heap_resort(heap, cmp, set_backpointer)                         \
 233 do {                                                                    \
 234         ssize_t _i;                                                     \
 235         for (_i = (ssize_t) (heap)->used / 2 -  1; _i >= 0; --_i)       \
 236                 heap_sift_down(heap, _i, cmp, set_backpointer);         \
 237 } while (0)
 238
 239 #define ANYSINT_MAX(t)                                                  \
 240         ((((t) 1 << (sizeof(t) * 8 - 2)) - (t) 1) * (t) 2 + (t) 1)
 241
 242
 243 #ifdef __KERNEL__
 244 static inline void pr_time(struct printbuf *out, u64 time)
 245 {
 246         prt_printf(out, "%llu", time);
 247 }
 248 #else
 249 #include <time.h>
 250 static inline void pr_time(struct printbuf *out, u64 _time)
 251 {
 252         char time_str[64];
 253         time_t time = _time;
 254         struct tm *tm = localtime(&time);
 255         size_t err = strftime(time_str, sizeof(time_str), "%c", tm);
 256         if (!err)
 257                 prt_printf(out, "(formatting error)");
 258         else
 259                 prt_printf(out, "%s", time_str);
 260 }
 261 #endif
 262
 263 #ifdef __KERNEL__
 264 static inline void uuid_unparse_lower(u8 *uuid, char *out)
 265 {
 266         sprintf(out, "%pUb", uuid);
 267 }
 268 #else
 269 #include <uuid/uuid.h>
 270 #endif
 271
 272 static inline void pr_uuid(struct printbuf *out, u8 *uuid)
 273 {
 274         char uuid_str[40];
 275
 276         uuid_unparse_lower(uuid, uuid_str);
 277         prt_printf(out, "%s", uuid_str);
 278 }
 279
 280 int bch2_strtoint_h(const char *, int *);
 281 int bch2_strtouint_h(const char *, unsigned int *);
 282 int bch2_strtoll_h(const char *, long long *);
 283 int bch2_strtoull_h(const char *, unsigned long long *);
 284 int bch2_strtou64_h(const char *, u64 *);
 285
 286 static inline int bch2_strtol_h(const char *cp, long *res)
 287 {
 288 #if BITS_PER_LONG == 32
 289         return bch2_strtoint_h(cp, (int *) res);
 290 #else
 291         return bch2_strtoll_h(cp, (long long *) res);
 292 #endif
 293 }
 294
 295 static inline int bch2_strtoul_h(const char *cp, long *res)
 296 {
 297 #if BITS_PER_LONG == 32
 298         return bch2_strtouint_h(cp, (unsigned int *) res);
 299 #else
 300         return bch2_strtoull_h(cp, (unsigned long long *) res);
 301 #endif
 302 }
 303
 304 #define strtoi_h(cp, res)                                               \
 305         ( type_is(*res, int)            ? bch2_strtoint_h(cp, (void *) res)\
 306         : type_is(*res, long)           ? bch2_strtol_h(cp, (void *) res)\
 307         : type_is(*res, long long)      ? bch2_strtoll_h(cp, (void *) res)\
 308         : type_is(*res, unsigned)       ? bch2_strtouint_h(cp, (void *) res)\
 309         : type_is(*res, unsigned long)  ? bch2_strtoul_h(cp, (void *) res)\
 310         : type_is(*res, unsigned long long) ? bch2_strtoull_h(cp, (void *) res)\
 311         : -EINVAL)
 312
 313 #define strtoul_safe(cp, var)                                           \
 314 ({                                                                      \
 315         unsigned long _v;                                               \
 316         int _r = kstrtoul(cp, 10, &_v);                                 \
 317         if (!_r)                                                        \
 318                 var = _v;                                               \
 319         _r;                                                             \
 320 })
 321
 322 #define strtoul_safe_clamp(cp, var, min, max)                           \
 323 ({                                                                      \
 324         unsigned long _v;                                               \
 325         int _r = kstrtoul(cp, 10, &_v);                                 \
 326         if (!_r)                                                        \
 327                 var = clamp_t(typeof(var), _v, min, max);               \
 328         _r;                                                             \
 329 })
 330
 331 #define strtoul_safe_restrict(cp, var, min, max)                        \
 332 ({                                                                      \
 333         unsigned long _v;                                               \
 334         int _r = kstrtoul(cp, 10, &_v);                                 \
 335         if (!_r && _v >= min && _v <= max)                              \
 336                 var = _v;                                               \
 337         else                                                            \
 338                 _r = -EINVAL;                                           \
 339         _r;                                                             \
 340 })
 341
 342 #define snprint(out, var)                                               \
 343         prt_printf(out,                                                 \
 344                    type_is(var, int)            ? "%i\n"                \
 345                  : type_is(var, unsigned)       ? "%u\n"                \
 346                  : type_is(var, long)           ? "%li\n"               \
 347                  : type_is(var, unsigned long)  ? "%lu\n"               \
 348                  : type_is(var, s64)            ? "%lli\n"              \
 349                  : type_is(var, u64)            ? "%llu\n"              \
 350                  : type_is(var, char *)         ? "%s\n"                \
 351                  : "%i\n", var)
 352
 353 bool bch2_is_zero(const void *, size_t);
 354
 355 u64 bch2_read_flag_list(char *, const char * const[]);
 356
 357 void bch2_prt_u64_binary(struct printbuf *, u64, unsigned);
 358
 359 void bch2_print_string_as_lines(const char *prefix, const char *lines);
 360 int bch2_prt_backtrace(struct printbuf *, struct task_struct *);
 361
 362 #define NR_QUANTILES    15
 363 #define QUANTILE_IDX(i) inorder_to_eytzinger0(i, NR_QUANTILES)
 364 #define QUANTILE_FIRST  eytzinger0_first(NR_QUANTILES)
 365 #define QUANTILE_LAST   eytzinger0_last(NR_QUANTILES)
 366
 367 struct quantiles {
 368         struct quantile_entry {
 369                 u64     m;
 370                 u64     step;
 371         }               entries[NR_QUANTILES];
 372 };
 373
 374 struct time_stat_buffer {
 375         unsigned        nr;
 376         struct time_stat_buffer_entry {
 377                 u64     start;
 378                 u64     end;
 379         }               entries[32];
 380 };
 381
 382 struct time_stats {
 383         spinlock_t      lock;
 384         /* all fields are in nanoseconds */
 385         u64             max_duration;
 386         u64             min_duration;
 387         u64             max_freq;
 388         u64             min_freq;
 389         u64             last_event;
 390         struct quantiles quantiles;
 391
 392         struct mean_and_variance          duration_stats;
 393         struct mean_and_variance_weighted duration_stats_weighted;
 394         struct mean_and_variance          freq_stats;
 395         struct mean_and_variance_weighted freq_stats_weighted;
 396         struct time_stat_buffer __percpu *buffer;
 397 };
 398
 399 void __bch2_time_stats_update(struct time_stats *stats, u64, u64);
 400
 401 static inline void bch2_time_stats_update(struct time_stats *stats, u64 start)
 402 {
 403         __bch2_time_stats_update(stats, start, local_clock());
 404 }
 405
 406 void bch2_time_stats_to_text(struct printbuf *, struct time_stats *);
 407
 408 void bch2_time_stats_exit(struct time_stats *);
 409 void bch2_time_stats_init(struct time_stats *);
 410
 411 #define ewma_add(ewma, val, weight)                                     \
 412 ({                                                                      \
 413         typeof(ewma) _ewma = (ewma);                                    \
 414         typeof(weight) _weight = (weight);                              \
 415                                                                         \
 416         (((_ewma << _weight) - _ewma) + (val)) >> _weight;              \
 417 })
 418
 419 struct bch_ratelimit {
 420         /* Next time we want to do some work, in nanoseconds */
 421         u64                     next;
 422
 423         /*
 424          * Rate at which we want to do work, in units per nanosecond
 425          * The units here correspond to the units passed to
 426          * bch2_ratelimit_increment()
 427          */
 428         unsigned                rate;
 429 };
 430
 431 static inline void bch2_ratelimit_reset(struct bch_ratelimit *d)
 432 {
 433         d->next = local_clock();
 434 }
 435
 436 u64 bch2_ratelimit_delay(struct bch_ratelimit *);
 437 void bch2_ratelimit_increment(struct bch_ratelimit *, u64);
 438
 439 struct bch_pd_controller {
 440         struct bch_ratelimit    rate;
 441         unsigned long           last_update;
 442
 443         s64                     last_actual;
 444         s64                     smoothed_derivative;
 445
 446         unsigned                p_term_inverse;
 447         unsigned                d_smooth;
 448         unsigned                d_term;
 449
 450         /* for exporting to sysfs (no effect on behavior) */
 451         s64                     last_derivative;
 452         s64                     last_proportional;
 453         s64                     last_change;
 454         s64                     last_target;
 455
 456         /* If true, the rate will not increase if bch2_ratelimit_delay()
 457          * is not being called often enough. */
 458         bool                    backpressure;
 459 };
 460
 461 void bch2_pd_controller_update(struct bch_pd_controller *, s64, s64, int);
 462 void bch2_pd_controller_init(struct bch_pd_controller *);
 463 void bch2_pd_controller_debug_to_text(struct printbuf *, struct bch_pd_controller *);
 464
 465 #define sysfs_pd_controller_attribute(name)                             \
 466         rw_attribute(name##_rate);                                      \
 467         rw_attribute(name##_rate_bytes);                                \
 468         rw_attribute(name##_rate_d_term);                               \
 469         rw_attribute(name##_rate_p_term_inverse);                       \
 470         read_attribute(name##_rate_debug)
 471
 472 #define sysfs_pd_controller_files(name)                                 \
 473         &sysfs_##name##_rate,                                           \
 474         &sysfs_##name##_rate_bytes,                                     \
 475         &sysfs_##name##_rate_d_term,                                    \
 476         &sysfs_##name##_rate_p_term_inverse,                            \
 477         &sysfs_##name##_rate_debug
 478
 479 #define sysfs_pd_controller_show(name, var)                             \
 480 do {                                                                    \
 481         sysfs_hprint(name##_rate,               (var)->rate.rate);      \
 482         sysfs_print(name##_rate_bytes,          (var)->rate.rate);      \
 483         sysfs_print(name##_rate_d_term,         (var)->d_term);         \
 484         sysfs_print(name##_rate_p_term_inverse, (var)->p_term_inverse); \
 485                                                                         \
 486         if (attr == &sysfs_##name##_rate_debug)                         \
 487                 bch2_pd_controller_debug_to_text(out, var);             \
 488 } while (0)
 489
 490 #define sysfs_pd_controller_store(name, var)                            \
 491 do {                                                                    \
 492         sysfs_strtoul_clamp(name##_rate,                                \
 493                             (var)->rate.rate, 1, UINT_MAX);             \
 494         sysfs_strtoul_clamp(name##_rate_bytes,                          \
 495                             (var)->rate.rate, 1, UINT_MAX);             \
 496         sysfs_strtoul(name##_rate_d_term,       (var)->d_term);         \
 497         sysfs_strtoul_clamp(name##_rate_p_term_inverse,                 \
 498                             (var)->p_term_inverse, 1, INT_MAX);         \
 499 } while (0)
 500
 501 #define container_of_or_null(ptr, type, member)                         \
 502 ({                                                                      \
 503         typeof(ptr) _ptr = ptr;                                         \
 504         _ptr ? container_of(_ptr, type, member) : NULL;                 \
 505 })
 506
 507 /* Does linear interpolation between powers of two */
 508 static inline unsigned fract_exp_two(unsigned x, unsigned fract_bits)
 509 {
 510         unsigned fract = x & ~(~0 << fract_bits);
 511
 512         x >>= fract_bits;
 513         x   = 1 << x;
 514         x  += (x * fract) >> fract_bits;
 515
 516         return x;
 517 }
 518
 519 void bch2_bio_map(struct bio *bio, void *base, size_t);
 520 int bch2_bio_alloc_pages(struct bio *, size_t, gfp_t);
 521
 522 static inline sector_t bdev_sectors(struct block_device *bdev)
 523 {
 524         return bdev->bd_inode->i_size >> 9;
 525 }
 526
 527 #define closure_bio_submit(bio, cl)                                     \
 528 do {                                                                    \
 529         closure_get(cl);                                                \
 530         submit_bio(bio);                                                \
 531 } while (0)
 532
 533 #define kthread_wait_freezable(cond)                                    \
 534 ({                                                                      \
 535         int _ret = 0;                                                   \
 536         while (1) {                                                     \
 537                 set_current_state(TASK_INTERRUPTIBLE);                  \
 538                 if (kthread_should_stop()) {                            \
 539                         _ret = -1;                                      \
 540                         break;                                          \
 541                 }                                                       \
 542                                                                         \
 543                 if (cond)                                               \
 544                         break;                                          \
 545                                                                         \
 546                 schedule();                                             \
 547                 try_to_freeze();                                        \
 548         }                                                               \
 549         set_current_state(TASK_RUNNING);                                \
 550         _ret;                                                           \
 551 })
 552
 553 size_t bch2_rand_range(size_t);
 554
 555 void memcpy_to_bio(struct bio *, struct bvec_iter, const void *);
 556 void memcpy_from_bio(void *, struct bio *, struct bvec_iter);
 557
 558 static inline void memcpy_u64s_small(void *dst, const void *src,
 559                                      unsigned u64s)
 560 {
 561         u64 *d = dst;
 562         const u64 *s = src;
 563
 564         while (u64s--)
 565                 *d++ = *s++;
 566 }
 567
 568 static inline void __memcpy_u64s(void *dst, const void *src,
 569                                  unsigned u64s)
 570 {
 571 #ifdef CONFIG_X86_64
 572         long d0, d1, d2;
 573         asm volatile("rep ; movsq"
 574                      : "=&c" (d0), "=&D" (d1), "=&S" (d2)
 575                      : "0" (u64s), "1" (dst), "2" (src)
 576                      : "memory");
 577 #else
 578         u64 *d = dst;
 579         const u64 *s = src;
 580
 581         while (u64s--)
 582                 *d++ = *s++;
 583 #endif
 584 }
 585
 586 static inline void memcpy_u64s(void *dst, const void *src,
 587                                unsigned u64s)
 588 {
 589         EBUG_ON(!(dst >= src + u64s * sizeof(u64) ||
 590                  dst + u64s * sizeof(u64) <= src));
 591
 592         __memcpy_u64s(dst, src, u64s);
 593 }
 594
 595 static inline void __memmove_u64s_down(void *dst, const void *src,
 596                                        unsigned u64s)
 597 {
 598         __memcpy_u64s(dst, src, u64s);
 599 }
 600
 601 static inline void memmove_u64s_down(void *dst, const void *src,
 602                                      unsigned u64s)
 603 {
 604         EBUG_ON(dst > src);
 605
 606         __memmove_u64s_down(dst, src, u64s);
 607 }
 608
 609 static inline void __memmove_u64s_up_small(void *_dst, const void *_src,
 610                                            unsigned u64s)
 611 {
 612         u64 *dst = (u64 *) _dst + u64s;
 613         u64 *src = (u64 *) _src + u64s;
 614
 615         while (u64s--)
 616                 *--dst = *--src;
 617 }
 618
 619 static inline void memmove_u64s_up_small(void *dst, const void *src,
 620                                          unsigned u64s)
 621 {
 622         EBUG_ON(dst < src);
 623
 624         __memmove_u64s_up_small(dst, src, u64s);
 625 }
 626
 627 static inline void __memmove_u64s_up(void *_dst, const void *_src,
 628                                      unsigned u64s)
 629 {
 630         u64 *dst = (u64 *) _dst + u64s - 1;
 631         u64 *src = (u64 *) _src + u64s - 1;
 632
 633 #ifdef CONFIG_X86_64
 634         long d0, d1, d2;
 635         asm volatile("std ;\n"
 636                      "rep ; movsq\n"
 637                      "cld ;\n"
 638                      : "=&c" (d0), "=&D" (d1), "=&S" (d2)
 639                      : "0" (u64s), "1" (dst), "2" (src)
 640                      : "memory");
 641 #else
 642         while (u64s--)
 643                 *dst-- = *src--;
 644 #endif
 645 }
 646
 647 static inline void memmove_u64s_up(void *dst, const void *src,
 648                                    unsigned u64s)
 649 {
 650         EBUG_ON(dst < src);
 651
 652         __memmove_u64s_up(dst, src, u64s);
 653 }
 654
 655 static inline void memmove_u64s(void *dst, const void *src,
 656                                 unsigned u64s)
 657 {
 658         if (dst < src)
 659                 __memmove_u64s_down(dst, src, u64s);
 660         else
 661                 __memmove_u64s_up(dst, src, u64s);
 662 }
 663
 664 /* Set the last few bytes up to a u64 boundary given an offset into a buffer. */
 665 static inline void memset_u64s_tail(void *s, int c, unsigned bytes)
 666 {
 667         unsigned rem = round_up(bytes, sizeof(u64)) - bytes;
 668
 669         memset(s + bytes, c, rem);
 670 }
 671
 672 void sort_cmp_size(void *base, size_t num, size_t size,
 673           int (*cmp_func)(const void *, const void *, size_t),
 674           void (*swap_func)(void *, void *, size_t));
 675
 676 /* just the memmove, doesn't update @_nr */
 677 #define __array_insert_item(_array, _nr, _pos)                          \
 678         memmove(&(_array)[(_pos) + 1],                                  \
 679                 &(_array)[(_pos)],                                      \
 680                 sizeof((_array)[0]) * ((_nr) - (_pos)))
 681
 682 #define array_insert_item(_array, _nr, _pos, _new_item)                 \
 683 do {                                                                    \
 684         __array_insert_item(_array, _nr, _pos);                         \
 685         (_nr)++;                                                        \
 686         (_array)[(_pos)] = (_new_item);                                 \
 687 } while (0)
 688
 689 #define array_remove_items(_array, _nr, _pos, _nr_to_remove)            \
 690 do {                                                                    \
 691         (_nr) -= (_nr_to_remove);                                       \
 692         memmove(&(_array)[(_pos)],                                      \
 693                 &(_array)[(_pos) + (_nr_to_remove)],                    \
 694                 sizeof((_array)[0]) * ((_nr) - (_pos)));                \
 695 } while (0)
 696
 697 #define array_remove_item(_array, _nr, _pos)                            \
 698         array_remove_items(_array, _nr, _pos, 1)
 699
 700 static inline void __move_gap(void *array, size_t element_size,
 701                               size_t nr, size_t size,
 702                               size_t old_gap, size_t new_gap)
 703 {
 704         size_t gap_end = old_gap + size - nr;
 705
 706         if (new_gap < old_gap) {
 707                 size_t move = old_gap - new_gap;
 708
 709                 memmove(array + element_size * (gap_end - move),
 710                         array + element_size * (old_gap - move),
 711                                 element_size * move);
 712         } else if (new_gap > old_gap) {
 713                 size_t move = new_gap - old_gap;
 714
 715                 memmove(array + element_size * old_gap,
 716                         array + element_size * gap_end,
 717                                 element_size * move);
 718         }
 719 }
 720
 721 /* Move the gap in a gap buffer: */
 722 #define move_gap(_array, _nr, _size, _old_gap, _new_gap)        \
 723         __move_gap(_array, sizeof(_array[0]), _nr, _size, _old_gap, _new_gap)
 724
 725 #define bubble_sort(_base, _nr, _cmp)                                   \
 726 do {                                                                    \
 727         ssize_t _i, _end;                                               \
 728         bool _swapped = true;                                           \
 729                                                                         \
 730         for (_end = (ssize_t) (_nr) - 1; _end > 0 && _swapped; --_end) {\
 731                 _swapped = false;                                       \
 732                 for (_i = 0; _i < _end; _i++)                           \
 733                         if (_cmp((_base)[_i], (_base)[_i + 1]) > 0) {   \
 734                                 swap((_base)[_i], (_base)[_i + 1]);     \
 735                                 _swapped = true;                        \
 736                         }                                               \
 737         }                                                               \
 738 } while (0)
 739
 740 static inline u64 percpu_u64_get(u64 __percpu *src)
 741 {
 742         u64 ret = 0;
 743         int cpu;
 744
 745         for_each_possible_cpu(cpu)
 746                 ret += *per_cpu_ptr(src, cpu);
 747         return ret;
 748 }
 749
 750 static inline void percpu_u64_set(u64 __percpu *dst, u64 src)
 751 {
 752         int cpu;
 753
 754         for_each_possible_cpu(cpu)
 755                 *per_cpu_ptr(dst, cpu) = 0;
 756         this_cpu_write(*dst, src);
 757 }
 758
 759 static inline void acc_u64s(u64 *acc, const u64 *src, unsigned nr)
 760 {
 761         unsigned i;
 762
 763         for (i = 0; i < nr; i++)
 764                 acc[i] += src[i];
 765 }
 766
 767 static inline void acc_u64s_percpu(u64 *acc, const u64 __percpu *src,
 768                                    unsigned nr)
 769 {
 770         int cpu;
 771
 772         for_each_possible_cpu(cpu)
 773                 acc_u64s(acc, per_cpu_ptr(src, cpu), nr);
 774 }
 775
 776 static inline void percpu_memset(void __percpu *p, int c, size_t bytes)
 777 {
 778         int cpu;
 779
 780         for_each_possible_cpu(cpu)
 781                 memset(per_cpu_ptr(p, cpu), c, bytes);
 782 }
 783
 784 u64 *bch2_acc_percpu_u64s(u64 __percpu *, unsigned);
 785
 786 #define cmp_int(l, r)           ((l > r) - (l < r))
 787
 788 static inline int u8_cmp(u8 l, u8 r)
 789 {
 790         return cmp_int(l, r);
 791 }
 792
 793 #endif /* _BCACHEFS_UTIL_H */