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/percpu.h>
  15 #include <linux/preempt.h>
  16 #include <linux/ratelimit.h>
  17 #include <linux/slab.h>
  18 #include <linux/vmalloc.h>
  19 #include <linux/workqueue.h>
  20 #include <linux/mean_and_variance.h>
  21
  22 #include "darray.h"
  23
  24 struct closure;
  25
  26 #ifdef CONFIG_BCACHEFS_DEBUG
  27 #define EBUG_ON(cond)           BUG_ON(cond)
  28 #else
  29 #define EBUG_ON(cond)
  30 #endif
  31
  32 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
  33 #define CPU_BIG_ENDIAN          0
  34 #elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
  35 #define CPU_BIG_ENDIAN          1
  36 #endif
  37
  38 /* type hackery */
  39
  40 #define type_is_exact(_val, _type)                                      \
  41         __builtin_types_compatible_p(typeof(_val), _type)
  42
  43 #define type_is(_val, _type)                                            \
  44         (__builtin_types_compatible_p(typeof(_val), _type) ||           \
  45          __builtin_types_compatible_p(typeof(_val), const _type))
  46
  47 /* Userspace doesn't align allocations as nicely as the kernel allocators: */
  48 static inline size_t buf_pages(void *p, size_t len)
  49 {
  50         return DIV_ROUND_UP(len +
  51                             ((unsigned long) p & (PAGE_SIZE - 1)),
  52                             PAGE_SIZE);
  53 }
  54
  55 static inline void vpfree(void *p, size_t size)
  56 {
  57         if (is_vmalloc_addr(p))
  58                 vfree(p);
  59         else
  60                 free_pages((unsigned long) p, get_order(size));
  61 }
  62
  63 static inline void *vpmalloc(size_t size, gfp_t gfp_mask)
  64 {
  65         return (void *) __get_free_pages(gfp_mask|__GFP_NOWARN,
  66                                          get_order(size)) ?:
  67                 __vmalloc(size, gfp_mask);
  68 }
  69
  70 static inline void kvpfree(void *p, size_t size)
  71 {
  72         if (size < PAGE_SIZE)
  73                 kfree(p);
  74         else
  75                 vpfree(p, size);
  76 }
  77
  78 static inline void *kvpmalloc(size_t size, gfp_t gfp_mask)
  79 {
  80         return size < PAGE_SIZE
  81                 ? kmalloc(size, gfp_mask)
  82                 : vpmalloc(size, gfp_mask);
  83 }
  84
  85 int mempool_init_kvpmalloc_pool(mempool_t *, int, size_t);
  86
  87 #define HEAP(type)                                                      \
  88 struct {                                                                \
  89         size_t size, used;                                              \
  90         type *data;                                                     \
  91 }
  92
  93 #define DECLARE_HEAP(type, name) HEAP(type) name
  94
  95 #define init_heap(heap, _size, gfp)                                     \
  96 ({                                                                      \
  97         (heap)->used = 0;                                               \
  98         (heap)->size = (_size);                                         \
  99         (heap)->data = kvpmalloc((heap)->size * sizeof((heap)->data[0]),\
 100                                  (gfp));                                \
 101 })
 102
 103 #define free_heap(heap)                                                 \
 104 do {                                                                    \
 105         kvpfree((heap)->data, (heap)->size * sizeof((heap)->data[0]));  \
 106         (heap)->data = NULL;                                            \
 107 } while (0)
 108
 109 #define heap_set_backpointer(h, i, _fn)                                 \
 110 do {                                                                    \
 111         void (*fn)(typeof(h), size_t) = _fn;                            \
 112         if (fn)                                                         \
 113                 fn(h, i);                                               \
 114 } while (0)
 115
 116 #define heap_swap(h, i, j, set_backpointer)                             \
 117 do {                                                                    \
 118         swap((h)->data[i], (h)->data[j]);                               \
 119         heap_set_backpointer(h, i, set_backpointer);                    \
 120         heap_set_backpointer(h, j, set_backpointer);                    \
 121 } while (0)
 122
 123 #define heap_peek(h)                                                    \
 124 ({                                                                      \
 125         EBUG_ON(!(h)->used);                                            \
 126         (h)->data[0];                                                   \
 127 })
 128
 129 #define heap_full(h)    ((h)->used == (h)->size)
 130
 131 #define heap_sift_down(h, i, cmp, set_backpointer)                      \
 132 do {                                                                    \
 133         size_t _c, _j = i;                                              \
 134                                                                         \
 135         for (; _j * 2 + 1 < (h)->used; _j = _c) {                       \
 136                 _c = _j * 2 + 1;                                        \
 137                 if (_c + 1 < (h)->used &&                               \
 138                     cmp(h, (h)->data[_c], (h)->data[_c + 1]) >= 0)      \
 139                         _c++;                                           \
 140                                                                         \
 141                 if (cmp(h, (h)->data[_c], (h)->data[_j]) >= 0)          \
 142                         break;                                          \
 143                 heap_swap(h, _c, _j, set_backpointer);                  \
 144         }                                                               \
 145 } while (0)
 146
 147 #define heap_sift_up(h, i, cmp, set_backpointer)                        \
 148 do {                                                                    \
 149         while (i) {                                                     \
 150                 size_t p = (i - 1) / 2;                                 \
 151                 if (cmp(h, (h)->data[i], (h)->data[p]) >= 0)            \
 152                         break;                                          \
 153                 heap_swap(h, i, p, set_backpointer);                    \
 154                 i = p;                                                  \
 155         }                                                               \
 156 } while (0)
 157
 158 #define __heap_add(h, d, cmp, set_backpointer)                          \
 159 ({                                                                      \
 160         size_t _i = (h)->used++;                                        \
 161         (h)->data[_i] = d;                                              \
 162         heap_set_backpointer(h, _i, set_backpointer);                   \
 163                                                                         \
 164         heap_sift_up(h, _i, cmp, set_backpointer);                      \
 165         _i;                                                             \
 166 })
 167
 168 #define heap_add(h, d, cmp, set_backpointer)                            \
 169 ({                                                                      \
 170         bool _r = !heap_full(h);                                        \
 171         if (_r)                                                         \
 172                 __heap_add(h, d, cmp, set_backpointer);                 \
 173         _r;                                                             \
 174 })
 175
 176 #define heap_add_or_replace(h, new, cmp, set_backpointer)               \
 177 do {                                                                    \
 178         if (!heap_add(h, new, cmp, set_backpointer) &&                  \
 179             cmp(h, new, heap_peek(h)) >= 0) {                           \
 180                 (h)->data[0] = new;                                     \
 181                 heap_set_backpointer(h, 0, set_backpointer);            \
 182                 heap_sift_down(h, 0, cmp, set_backpointer);             \
 183         }                                                               \
 184 } while (0)
 185
 186 #define heap_del(h, i, cmp, set_backpointer)                            \
 187 do {                                                                    \
 188         size_t _i = (i);                                                \
 189                                                                         \
 190         BUG_ON(_i >= (h)->used);                                        \
 191         (h)->used--;                                                    \
 192         if ((_i) < (h)->used) {                                         \
 193                 heap_swap(h, _i, (h)->used, set_backpointer);           \
 194                 heap_sift_up(h, _i, cmp, set_backpointer);              \
 195                 heap_sift_down(h, _i, cmp, set_backpointer);            \
 196         }                                                               \
 197 } while (0)
 198
 199 #define heap_pop(h, d, cmp, set_backpointer)                            \
 200 ({                                                                      \
 201         bool _r = (h)->used;                                            \
 202         if (_r) {                                                       \
 203                 (d) = (h)->data[0];                                     \
 204                 heap_del(h, 0, cmp, set_backpointer);                   \
 205         }                                                               \
 206         _r;                                                             \
 207 })
 208
 209 #define heap_resort(heap, cmp, set_backpointer)                         \
 210 do {                                                                    \
 211         ssize_t _i;                                                     \
 212         for (_i = (ssize_t) (heap)->used / 2 -  1; _i >= 0; --_i)       \
 213                 heap_sift_down(heap, _i, cmp, set_backpointer);         \
 214 } while (0)
 215
 216 #define ANYSINT_MAX(t)                                                  \
 217         ((((t) 1 << (sizeof(t) * 8 - 2)) - (t) 1) * (t) 2 + (t) 1)
 218
 219 #include "printbuf.h"
 220
 221 #define prt_vprintf(_out, ...)          bch2_prt_vprintf(_out, __VA_ARGS__)
 222 #define prt_printf(_out, ...)           bch2_prt_printf(_out, __VA_ARGS__)
 223 #define printbuf_str(_buf)              bch2_printbuf_str(_buf)
 224 #define printbuf_exit(_buf)             bch2_printbuf_exit(_buf)
 225
 226 #define printbuf_tabstops_reset(_buf)   bch2_printbuf_tabstops_reset(_buf)
 227 #define printbuf_tabstop_pop(_buf)      bch2_printbuf_tabstop_pop(_buf)
 228 #define printbuf_tabstop_push(_buf, _n) bch2_printbuf_tabstop_push(_buf, _n)
 229
 230 #define printbuf_indent_add(_out, _n)   bch2_printbuf_indent_add(_out, _n)
 231 #define printbuf_indent_sub(_out, _n)   bch2_printbuf_indent_sub(_out, _n)
 232
 233 #define prt_newline(_out)               bch2_prt_newline(_out)
 234 #define prt_tab(_out)                   bch2_prt_tab(_out)
 235 #define prt_tab_rjust(_out)             bch2_prt_tab_rjust(_out)
 236
 237 #define prt_bytes_indented(...)         bch2_prt_bytes_indented(__VA_ARGS__)
 238 #define prt_u64(_out, _v)               prt_printf(_out, "%llu", (u64) (_v))
 239 #define prt_human_readable_u64(...)     bch2_prt_human_readable_u64(__VA_ARGS__)
 240 #define prt_human_readable_s64(...)     bch2_prt_human_readable_s64(__VA_ARGS__)
 241 #define prt_units_u64(...)              bch2_prt_units_u64(__VA_ARGS__)
 242 #define prt_units_s64(...)              bch2_prt_units_s64(__VA_ARGS__)
 243 #define prt_string_option(...)          bch2_prt_string_option(__VA_ARGS__)
 244 #define prt_bitflags(...)               bch2_prt_bitflags(__VA_ARGS__)
 245
 246 void bch2_pr_time_units(struct printbuf *, u64);
 247 void bch2_prt_date_seconds(struct printbuf *, time64_t);
 248
 249 #ifdef __KERNEL__
 250 static inline void uuid_unparse_lower(u8 *uuid, char *out)
 251 {
 252         sprintf(out, "%pUb", uuid);
 253 }
 254 #else
 255 #include <uuid/uuid.h>
 256 #endif
 257
 258 static inline void pr_uuid(struct printbuf *out, u8 *uuid)
 259 {
 260         char uuid_str[40];
 261
 262         uuid_unparse_lower(uuid, uuid_str);
 263         prt_printf(out, "%s", uuid_str);
 264 }
 265
 266 int bch2_strtoint_h(const char *, int *);
 267 int bch2_strtouint_h(const char *, unsigned int *);
 268 int bch2_strtoll_h(const char *, long long *);
 269 int bch2_strtoull_h(const char *, unsigned long long *);
 270 int bch2_strtou64_h(const char *, u64 *);
 271
 272 static inline int bch2_strtol_h(const char *cp, long *res)
 273 {
 274 #if BITS_PER_LONG == 32
 275         return bch2_strtoint_h(cp, (int *) res);
 276 #else
 277         return bch2_strtoll_h(cp, (long long *) res);
 278 #endif
 279 }
 280
 281 static inline int bch2_strtoul_h(const char *cp, long *res)
 282 {
 283 #if BITS_PER_LONG == 32
 284         return bch2_strtouint_h(cp, (unsigned int *) res);
 285 #else
 286         return bch2_strtoull_h(cp, (unsigned long long *) res);
 287 #endif
 288 }
 289
 290 #define strtoi_h(cp, res)                                               \
 291         ( type_is(*res, int)            ? bch2_strtoint_h(cp, (void *) res)\
 292         : type_is(*res, long)           ? bch2_strtol_h(cp, (void *) res)\
 293         : type_is(*res, long long)      ? bch2_strtoll_h(cp, (void *) res)\
 294         : type_is(*res, unsigned)       ? bch2_strtouint_h(cp, (void *) res)\
 295         : type_is(*res, unsigned long)  ? bch2_strtoul_h(cp, (void *) res)\
 296         : type_is(*res, unsigned long long) ? bch2_strtoull_h(cp, (void *) res)\
 297         : -EINVAL)
 298
 299 #define strtoul_safe(cp, var)                                           \
 300 ({                                                                      \
 301         unsigned long _v;                                               \
 302         int _r = kstrtoul(cp, 10, &_v);                                 \
 303         if (!_r)                                                        \
 304                 var = _v;                                               \
 305         _r;                                                             \
 306 })
 307
 308 #define strtoul_safe_clamp(cp, var, min, max)                           \
 309 ({                                                                      \
 310         unsigned long _v;                                               \
 311         int _r = kstrtoul(cp, 10, &_v);                                 \
 312         if (!_r)                                                        \
 313                 var = clamp_t(typeof(var), _v, min, max);               \
 314         _r;                                                             \
 315 })
 316
 317 #define strtoul_safe_restrict(cp, var, min, max)                        \
 318 ({                                                                      \
 319         unsigned long _v;                                               \
 320         int _r = kstrtoul(cp, 10, &_v);                                 \
 321         if (!_r && _v >= min && _v <= max)                              \
 322                 var = _v;                                               \
 323         else                                                            \
 324                 _r = -EINVAL;                                           \
 325         _r;                                                             \
 326 })
 327
 328 #define snprint(out, var)                                               \
 329         prt_printf(out,                                                 \
 330                    type_is(var, int)            ? "%i\n"                \
 331                  : type_is(var, unsigned)       ? "%u\n"                \
 332                  : type_is(var, long)           ? "%li\n"               \
 333                  : type_is(var, unsigned long)  ? "%lu\n"               \
 334                  : type_is(var, s64)            ? "%lli\n"              \
 335                  : type_is(var, u64)            ? "%llu\n"              \
 336                  : type_is(var, char *)         ? "%s\n"                \
 337                  : "%i\n", var)
 338
 339 bool bch2_is_zero(const void *, size_t);
 340
 341 u64 bch2_read_flag_list(char *, const char * const[]);
 342
 343 void bch2_prt_u64_binary(struct printbuf *, u64, unsigned);
 344
 345 void bch2_print_string_as_lines(const char *prefix, const char *lines);
 346
 347 typedef DARRAY(unsigned long) bch_stacktrace;
 348 int bch2_save_backtrace(bch_stacktrace *stack, struct task_struct *);
 349 void bch2_prt_backtrace(struct printbuf *, bch_stacktrace *);
 350 int bch2_prt_task_backtrace(struct printbuf *, struct task_struct *);
 351
 352 #define NR_QUANTILES    15
 353 #define QUANTILE_IDX(i) inorder_to_eytzinger0(i, NR_QUANTILES)
 354 #define QUANTILE_FIRST  eytzinger0_first(NR_QUANTILES)
 355 #define QUANTILE_LAST   eytzinger0_last(NR_QUANTILES)
 356
 357 struct bch2_quantiles {
 358         struct bch2_quantile_entry {
 359                 u64     m;
 360                 u64     step;
 361         }               entries[NR_QUANTILES];
 362 };
 363
 364 struct bch2_time_stat_buffer {
 365         unsigned        nr;
 366         struct bch2_time_stat_buffer_entry {
 367                 u64     start;
 368                 u64     end;
 369         }               entries[32];
 370 };
 371
 372 struct bch2_time_stats {
 373         spinlock_t      lock;
 374         /* all fields are in nanoseconds */
 375         u64             max_duration;
 376         u64             min_duration;
 377         u64             max_freq;
 378         u64             min_freq;
 379         u64             last_event;
 380         struct bch2_quantiles quantiles;
 381
 382         struct mean_and_variance          duration_stats;
 383         struct mean_and_variance_weighted duration_stats_weighted;
 384         struct mean_and_variance          freq_stats;
 385         struct mean_and_variance_weighted freq_stats_weighted;
 386         struct bch2_time_stat_buffer __percpu *buffer;
 387 };
 388
 389 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
 390 void __bch2_time_stats_update(struct bch2_time_stats *stats, u64, u64);
 391 #else
 392 static inline void __bch2_time_stats_update(struct bch2_time_stats *stats, u64 start, u64 end) {}
 393 #endif
 394
 395 static inline void bch2_time_stats_update(struct bch2_time_stats *stats, u64 start)
 396 {
 397         __bch2_time_stats_update(stats, start, local_clock());
 398 }
 399
 400 void bch2_time_stats_to_text(struct printbuf *, struct bch2_time_stats *);
 401
 402 void bch2_time_stats_exit(struct bch2_time_stats *);
 403 void bch2_time_stats_init(struct bch2_time_stats *);
 404
 405 #define ewma_add(ewma, val, weight)                                     \
 406 ({                                                                      \
 407         typeof(ewma) _ewma = (ewma);                                    \
 408         typeof(weight) _weight = (weight);                              \
 409                                                                         \
 410         (((_ewma << _weight) - _ewma) + (val)) >> _weight;              \
 411 })
 412
 413 struct bch_ratelimit {
 414         /* Next time we want to do some work, in nanoseconds */
 415         u64                     next;
 416
 417         /*
 418          * Rate at which we want to do work, in units per nanosecond
 419          * The units here correspond to the units passed to
 420          * bch2_ratelimit_increment()
 421          */
 422         unsigned                rate;
 423 };
 424
 425 static inline void bch2_ratelimit_reset(struct bch_ratelimit *d)
 426 {
 427         d->next = local_clock();
 428 }
 429
 430 u64 bch2_ratelimit_delay(struct bch_ratelimit *);
 431 void bch2_ratelimit_increment(struct bch_ratelimit *, u64);
 432
 433 struct bch_pd_controller {
 434         struct bch_ratelimit    rate;
 435         unsigned long           last_update;
 436
 437         s64                     last_actual;
 438         s64                     smoothed_derivative;
 439
 440         unsigned                p_term_inverse;
 441         unsigned                d_smooth;
 442         unsigned                d_term;
 443
 444         /* for exporting to sysfs (no effect on behavior) */
 445         s64                     last_derivative;
 446         s64                     last_proportional;
 447         s64                     last_change;
 448         s64                     last_target;
 449
 450         /*
 451          * If true, the rate will not increase if bch2_ratelimit_delay()
 452          * is not being called often enough.
 453          */
 454         bool                    backpressure;
 455 };
 456
 457 void bch2_pd_controller_update(struct bch_pd_controller *, s64, s64, int);
 458 void bch2_pd_controller_init(struct bch_pd_controller *);
 459 void bch2_pd_controller_debug_to_text(struct printbuf *, struct bch_pd_controller *);
 460
 461 #define sysfs_pd_controller_attribute(name)                             \
 462         rw_attribute(name##_rate);                                      \
 463         rw_attribute(name##_rate_bytes);                                \
 464         rw_attribute(name##_rate_d_term);                               \
 465         rw_attribute(name##_rate_p_term_inverse);                       \
 466         read_attribute(name##_rate_debug)
 467
 468 #define sysfs_pd_controller_files(name)                                 \
 469         &sysfs_##name##_rate,                                           \
 470         &sysfs_##name##_rate_bytes,                                     \
 471         &sysfs_##name##_rate_d_term,                                    \
 472         &sysfs_##name##_rate_p_term_inverse,                            \
 473         &sysfs_##name##_rate_debug
 474
 475 #define sysfs_pd_controller_show(name, var)                             \
 476 do {                                                                    \
 477         sysfs_hprint(name##_rate,               (var)->rate.rate);      \
 478         sysfs_print(name##_rate_bytes,          (var)->rate.rate);      \
 479         sysfs_print(name##_rate_d_term,         (var)->d_term);         \
 480         sysfs_print(name##_rate_p_term_inverse, (var)->p_term_inverse); \
 481                                                                         \
 482         if (attr == &sysfs_##name##_rate_debug)                         \
 483                 bch2_pd_controller_debug_to_text(out, var);             \
 484 } while (0)
 485
 486 #define sysfs_pd_controller_store(name, var)                            \
 487 do {                                                                    \
 488         sysfs_strtoul_clamp(name##_rate,                                \
 489                             (var)->rate.rate, 1, UINT_MAX);             \
 490         sysfs_strtoul_clamp(name##_rate_bytes,                          \
 491                             (var)->rate.rate, 1, UINT_MAX);             \
 492         sysfs_strtoul(name##_rate_d_term,       (var)->d_term);         \
 493         sysfs_strtoul_clamp(name##_rate_p_term_inverse,                 \
 494                             (var)->p_term_inverse, 1, INT_MAX);         \
 495 } while (0)
 496
 497 #define container_of_or_null(ptr, type, member)                         \
 498 ({                                                                      \
 499         typeof(ptr) _ptr = ptr;                                         \
 500         _ptr ? container_of(_ptr, type, member) : NULL;                 \
 501 })
 502
 503 /* Does linear interpolation between powers of two */
 504 static inline unsigned fract_exp_two(unsigned x, unsigned fract_bits)
 505 {
 506         unsigned fract = x & ~(~0 << fract_bits);
 507
 508         x >>= fract_bits;
 509         x   = 1 << x;
 510         x  += (x * fract) >> fract_bits;
 511
 512         return x;
 513 }
 514
 515 void bch2_bio_map(struct bio *bio, void *base, size_t);
 516 int bch2_bio_alloc_pages(struct bio *, size_t, gfp_t);
 517
 518 static inline sector_t bdev_sectors(struct block_device *bdev)
 519 {
 520         return bdev->bd_inode->i_size >> 9;
 521 }
 522
 523 #define closure_bio_submit(bio, cl)                                     \
 524 do {                                                                    \
 525         closure_get(cl);                                                \
 526         submit_bio(bio);                                                \
 527 } while (0)
 528
 529 #define kthread_wait(cond)                                              \
 530 ({                                                                      \
 531         int _ret = 0;                                                   \
 532                                                                         \
 533         while (1) {                                                     \
 534                 set_current_state(TASK_INTERRUPTIBLE);                  \
 535                 if (kthread_should_stop()) {                            \
 536                         _ret = -1;                                      \
 537                         break;                                          \
 538                 }                                                       \
 539                                                                         \
 540                 if (cond)                                               \
 541                         break;                                          \
 542                                                                         \
 543                 schedule();                                             \
 544         }                                                               \
 545         set_current_state(TASK_RUNNING);                                \
 546         _ret;                                                           \
 547 })
 548
 549 #define kthread_wait_freezable(cond)                                    \
 550 ({                                                                      \
 551         int _ret = 0;                                                   \
 552         while (1) {                                                     \
 553                 set_current_state(TASK_INTERRUPTIBLE);                  \
 554                 if (kthread_should_stop()) {                            \
 555                         _ret = -1;                                      \
 556                         break;                                          \
 557                 }                                                       \
 558                                                                         \
 559                 if (cond)                                               \
 560                         break;                                          \
 561                                                                         \
 562                 schedule();                                             \
 563                 try_to_freeze();                                        \
 564         }                                                               \
 565         set_current_state(TASK_RUNNING);                                \
 566         _ret;                                                           \
 567 })
 568
 569 size_t bch2_rand_range(size_t);
 570
 571 void memcpy_to_bio(struct bio *, struct bvec_iter, const void *);
 572 void memcpy_from_bio(void *, struct bio *, struct bvec_iter);
 573
 574 static inline void memcpy_u64s_small(void *dst, const void *src,
 575                                      unsigned u64s)
 576 {
 577         u64 *d = dst;
 578         const u64 *s = src;
 579
 580         while (u64s--)
 581                 *d++ = *s++;
 582 }
 583
 584 static inline void __memcpy_u64s(void *dst, const void *src,
 585                                  unsigned u64s)
 586 {
 587 #ifdef CONFIG_X86_64
 588         long d0, d1, d2;
 589
 590         asm volatile("rep ; movsq"
 591                      : "=&c" (d0), "=&D" (d1), "=&S" (d2)
 592                      : "0" (u64s), "1" (dst), "2" (src)
 593                      : "memory");
 594 #else
 595         u64 *d = dst;
 596         const u64 *s = src;
 597
 598         while (u64s--)
 599                 *d++ = *s++;
 600 #endif
 601 }
 602
 603 static inline void memcpy_u64s(void *dst, const void *src,
 604                                unsigned u64s)
 605 {
 606         EBUG_ON(!(dst >= src + u64s * sizeof(u64) ||
 607                  dst + u64s * sizeof(u64) <= src));
 608
 609         __memcpy_u64s(dst, src, u64s);
 610 }
 611
 612 static inline void __memmove_u64s_down(void *dst, const void *src,
 613                                        unsigned u64s)
 614 {
 615         __memcpy_u64s(dst, src, u64s);
 616 }
 617
 618 static inline void memmove_u64s_down(void *dst, const void *src,
 619                                      unsigned u64s)
 620 {
 621         EBUG_ON(dst > src);
 622
 623         __memmove_u64s_down(dst, src, u64s);
 624 }
 625
 626 static inline void __memmove_u64s_down_small(void *dst, const void *src,
 627                                        unsigned u64s)
 628 {
 629         memcpy_u64s_small(dst, src, u64s);
 630 }
 631
 632 static inline void memmove_u64s_down_small(void *dst, const void *src,
 633                                      unsigned u64s)
 634 {
 635         EBUG_ON(dst > src);
 636
 637         __memmove_u64s_down_small(dst, src, u64s);
 638 }
 639
 640 static inline void __memmove_u64s_up_small(void *_dst, const void *_src,
 641                                            unsigned u64s)
 642 {
 643         u64 *dst = (u64 *) _dst + u64s;
 644         u64 *src = (u64 *) _src + u64s;
 645
 646         while (u64s--)
 647                 *--dst = *--src;
 648 }
 649
 650 static inline void memmove_u64s_up_small(void *dst, const void *src,
 651                                          unsigned u64s)
 652 {
 653         EBUG_ON(dst < src);
 654
 655         __memmove_u64s_up_small(dst, src, u64s);
 656 }
 657
 658 static inline void __memmove_u64s_up(void *_dst, const void *_src,
 659                                      unsigned u64s)
 660 {
 661         u64 *dst = (u64 *) _dst + u64s - 1;
 662         u64 *src = (u64 *) _src + u64s - 1;
 663
 664 #ifdef CONFIG_X86_64
 665         long d0, d1, d2;
 666
 667         asm volatile("std ;\n"
 668                      "rep ; movsq\n"
 669                      "cld ;\n"
 670                      : "=&c" (d0), "=&D" (d1), "=&S" (d2)
 671                      : "0" (u64s), "1" (dst), "2" (src)
 672                      : "memory");
 673 #else
 674         while (u64s--)
 675                 *dst-- = *src--;
 676 #endif
 677 }
 678
 679 static inline void memmove_u64s_up(void *dst, const void *src,
 680                                    unsigned u64s)
 681 {
 682         EBUG_ON(dst < src);
 683
 684         __memmove_u64s_up(dst, src, u64s);
 685 }
 686
 687 static inline void memmove_u64s(void *dst, const void *src,
 688                                 unsigned u64s)
 689 {
 690         if (dst < src)
 691                 __memmove_u64s_down(dst, src, u64s);
 692         else
 693                 __memmove_u64s_up(dst, src, u64s);
 694 }
 695
 696 /* Set the last few bytes up to a u64 boundary given an offset into a buffer. */
 697 static inline void memset_u64s_tail(void *s, int c, unsigned bytes)
 698 {
 699         unsigned rem = round_up(bytes, sizeof(u64)) - bytes;
 700
 701         memset(s + bytes, c, rem);
 702 }
 703
 704 void sort_cmp_size(void *base, size_t num, size_t size,
 705           int (*cmp_func)(const void *, const void *, size_t),
 706           void (*swap_func)(void *, void *, size_t));
 707
 708 /* just the memmove, doesn't update @_nr */
 709 #define __array_insert_item(_array, _nr, _pos)                          \
 710         memmove(&(_array)[(_pos) + 1],                                  \
 711                 &(_array)[(_pos)],                                      \
 712                 sizeof((_array)[0]) * ((_nr) - (_pos)))
 713
 714 #define array_insert_item(_array, _nr, _pos, _new_item)                 \
 715 do {                                                                    \
 716         __array_insert_item(_array, _nr, _pos);                         \
 717         (_nr)++;                                                        \
 718         (_array)[(_pos)] = (_new_item);                                 \
 719 } while (0)
 720
 721 #define array_remove_items(_array, _nr, _pos, _nr_to_remove)            \
 722 do {                                                                    \
 723         (_nr) -= (_nr_to_remove);                                       \
 724         memmove(&(_array)[(_pos)],                                      \
 725                 &(_array)[(_pos) + (_nr_to_remove)],                    \
 726                 sizeof((_array)[0]) * ((_nr) - (_pos)));                \
 727 } while (0)
 728
 729 #define array_remove_item(_array, _nr, _pos)                            \
 730         array_remove_items(_array, _nr, _pos, 1)
 731
 732 static inline void __move_gap(void *array, size_t element_size,
 733                               size_t nr, size_t size,
 734                               size_t old_gap, size_t new_gap)
 735 {
 736         size_t gap_end = old_gap + size - nr;
 737
 738         if (new_gap < old_gap) {
 739                 size_t move = old_gap - new_gap;
 740
 741                 memmove(array + element_size * (gap_end - move),
 742                         array + element_size * (old_gap - move),
 743                                 element_size * move);
 744         } else if (new_gap > old_gap) {
 745                 size_t move = new_gap - old_gap;
 746
 747                 memmove(array + element_size * old_gap,
 748                         array + element_size * gap_end,
 749                                 element_size * move);
 750         }
 751 }
 752
 753 /* Move the gap in a gap buffer: */
 754 #define move_gap(_array, _nr, _size, _old_gap, _new_gap)        \
 755         __move_gap(_array, sizeof(_array[0]), _nr, _size, _old_gap, _new_gap)
 756
 757 #define bubble_sort(_base, _nr, _cmp)                                   \
 758 do {                                                                    \
 759         ssize_t _i, _last;                                              \
 760         bool _swapped = true;                                           \
 761                                                                         \
 762         for (_last= (ssize_t) (_nr) - 1; _last > 0 && _swapped; --_last) {\
 763                 _swapped = false;                                       \
 764                 for (_i = 0; _i < _last; _i++)                          \
 765                         if (_cmp((_base)[_i], (_base)[_i + 1]) > 0) {   \
 766                                 swap((_base)[_i], (_base)[_i + 1]);     \
 767                                 _swapped = true;                        \
 768                         }                                               \
 769         }                                                               \
 770 } while (0)
 771
 772 static inline u64 percpu_u64_get(u64 __percpu *src)
 773 {
 774         u64 ret = 0;
 775         int cpu;
 776
 777         for_each_possible_cpu(cpu)
 778                 ret += *per_cpu_ptr(src, cpu);
 779         return ret;
 780 }
 781
 782 static inline void percpu_u64_set(u64 __percpu *dst, u64 src)
 783 {
 784         int cpu;
 785
 786         for_each_possible_cpu(cpu)
 787                 *per_cpu_ptr(dst, cpu) = 0;
 788         this_cpu_write(*dst, src);
 789 }
 790
 791 static inline void acc_u64s(u64 *acc, const u64 *src, unsigned nr)
 792 {
 793         unsigned i;
 794
 795         for (i = 0; i < nr; i++)
 796                 acc[i] += src[i];
 797 }
 798
 799 static inline void acc_u64s_percpu(u64 *acc, const u64 __percpu *src,
 800                                    unsigned nr)
 801 {
 802         int cpu;
 803
 804         for_each_possible_cpu(cpu)
 805                 acc_u64s(acc, per_cpu_ptr(src, cpu), nr);
 806 }
 807
 808 static inline void percpu_memset(void __percpu *p, int c, size_t bytes)
 809 {
 810         int cpu;
 811
 812         for_each_possible_cpu(cpu)
 813                 memset(per_cpu_ptr(p, cpu), c, bytes);
 814 }
 815
 816 u64 *bch2_acc_percpu_u64s(u64 __percpu *, unsigned);
 817
 818 #define cmp_int(l, r)           ((l > r) - (l < r))
 819
 820 static inline int u8_cmp(u8 l, u8 r)
 821 {
 822         return cmp_int(l, r);
 823 }
 824
 825 static inline int cmp_le32(__le32 l, __le32 r)
 826 {
 827         return cmp_int(le32_to_cpu(l), le32_to_cpu(r));
 828 }
 829
 830 #include <linux/uuid.h>
 831
 832 #endif /* _BCACHEFS_UTIL_H */