700 lines · c
1/* SPDX-License-Identifier: GPL-2.0 */2#ifndef _BCACHEFS_UTIL_H3#define _BCACHEFS_UTIL_H4 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/min_heap.h>12#include <linux/sched/clock.h>13#include <linux/llist.h>14#include <linux/log2.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 22#include "mean_and_variance.h"23 24#include "darray.h"25#include "time_stats.h"26 27struct closure;28 29#ifdef CONFIG_BCACHEFS_DEBUG30#define EBUG_ON(cond) BUG_ON(cond)31#else32#define EBUG_ON(cond)33#endif34 35#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__36#define CPU_BIG_ENDIAN 037#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__38#define CPU_BIG_ENDIAN 139#endif40 41/* type hackery */42 43#define type_is_exact(_val, _type) \44 __builtin_types_compatible_p(typeof(_val), _type)45 46#define type_is(_val, _type) \47 (__builtin_types_compatible_p(typeof(_val), _type) || \48 __builtin_types_compatible_p(typeof(_val), const _type))49 50/* Userspace doesn't align allocations as nicely as the kernel allocators: */51static inline size_t buf_pages(void *p, size_t len)52{53 return DIV_ROUND_UP(len +54 ((unsigned long) p & (PAGE_SIZE - 1)),55 PAGE_SIZE);56}57 58#define init_heap(heap, _size, gfp) \59({ \60 (heap)->nr = 0; \61 (heap)->size = (_size); \62 (heap)->data = kvmalloc((heap)->size * sizeof((heap)->data[0]),\63 (gfp)); \64})65 66#define free_heap(heap) \67do { \68 kvfree((heap)->data); \69 (heap)->data = NULL; \70} while (0)71 72#define ANYSINT_MAX(t) \73 ((((t) 1 << (sizeof(t) * 8 - 2)) - (t) 1) * (t) 2 + (t) 1)74 75#include "printbuf.h"76 77#define prt_vprintf(_out, ...) bch2_prt_vprintf(_out, __VA_ARGS__)78#define prt_printf(_out, ...) bch2_prt_printf(_out, __VA_ARGS__)79#define printbuf_str(_buf) bch2_printbuf_str(_buf)80#define printbuf_exit(_buf) bch2_printbuf_exit(_buf)81 82#define printbuf_tabstops_reset(_buf) bch2_printbuf_tabstops_reset(_buf)83#define printbuf_tabstop_pop(_buf) bch2_printbuf_tabstop_pop(_buf)84#define printbuf_tabstop_push(_buf, _n) bch2_printbuf_tabstop_push(_buf, _n)85 86#define printbuf_indent_add(_out, _n) bch2_printbuf_indent_add(_out, _n)87#define printbuf_indent_sub(_out, _n) bch2_printbuf_indent_sub(_out, _n)88 89#define prt_newline(_out) bch2_prt_newline(_out)90#define prt_tab(_out) bch2_prt_tab(_out)91#define prt_tab_rjust(_out) bch2_prt_tab_rjust(_out)92 93#define prt_bytes_indented(...) bch2_prt_bytes_indented(__VA_ARGS__)94#define prt_u64(_out, _v) prt_printf(_out, "%llu", (u64) (_v))95#define prt_human_readable_u64(...) bch2_prt_human_readable_u64(__VA_ARGS__)96#define prt_human_readable_s64(...) bch2_prt_human_readable_s64(__VA_ARGS__)97#define prt_units_u64(...) bch2_prt_units_u64(__VA_ARGS__)98#define prt_units_s64(...) bch2_prt_units_s64(__VA_ARGS__)99#define prt_string_option(...) bch2_prt_string_option(__VA_ARGS__)100#define prt_bitflags(...) bch2_prt_bitflags(__VA_ARGS__)101#define prt_bitflags_vector(...) bch2_prt_bitflags_vector(__VA_ARGS__)102 103void bch2_pr_time_units(struct printbuf *, u64);104void bch2_prt_datetime(struct printbuf *, time64_t);105 106#ifdef __KERNEL__107static inline void uuid_unparse_lower(u8 *uuid, char *out)108{109 sprintf(out, "%pUb", uuid);110}111#else112#include <uuid/uuid.h>113#endif114 115static inline void pr_uuid(struct printbuf *out, u8 *uuid)116{117 char uuid_str[40];118 119 uuid_unparse_lower(uuid, uuid_str);120 prt_printf(out, "%s", uuid_str);121}122 123int bch2_strtoint_h(const char *, int *);124int bch2_strtouint_h(const char *, unsigned int *);125int bch2_strtoll_h(const char *, long long *);126int bch2_strtoull_h(const char *, unsigned long long *);127int bch2_strtou64_h(const char *, u64 *);128 129static inline int bch2_strtol_h(const char *cp, long *res)130{131#if BITS_PER_LONG == 32132 return bch2_strtoint_h(cp, (int *) res);133#else134 return bch2_strtoll_h(cp, (long long *) res);135#endif136}137 138static inline int bch2_strtoul_h(const char *cp, long *res)139{140#if BITS_PER_LONG == 32141 return bch2_strtouint_h(cp, (unsigned int *) res);142#else143 return bch2_strtoull_h(cp, (unsigned long long *) res);144#endif145}146 147#define strtoi_h(cp, res) \148 ( type_is(*res, int) ? bch2_strtoint_h(cp, (void *) res)\149 : type_is(*res, long) ? bch2_strtol_h(cp, (void *) res)\150 : type_is(*res, long long) ? bch2_strtoll_h(cp, (void *) res)\151 : type_is(*res, unsigned) ? bch2_strtouint_h(cp, (void *) res)\152 : type_is(*res, unsigned long) ? bch2_strtoul_h(cp, (void *) res)\153 : type_is(*res, unsigned long long) ? bch2_strtoull_h(cp, (void *) res)\154 : -EINVAL)155 156#define strtoul_safe(cp, var) \157({ \158 unsigned long _v; \159 int _r = kstrtoul(cp, 10, &_v); \160 if (!_r) \161 var = _v; \162 _r; \163})164 165#define strtoul_safe_clamp(cp, var, min, max) \166({ \167 unsigned long _v; \168 int _r = kstrtoul(cp, 10, &_v); \169 if (!_r) \170 var = clamp_t(typeof(var), _v, min, max); \171 _r; \172})173 174#define strtoul_safe_restrict(cp, var, min, max) \175({ \176 unsigned long _v; \177 int _r = kstrtoul(cp, 10, &_v); \178 if (!_r && _v >= min && _v <= max) \179 var = _v; \180 else \181 _r = -EINVAL; \182 _r; \183})184 185#define snprint(out, var) \186 prt_printf(out, \187 type_is(var, int) ? "%i\n" \188 : type_is(var, unsigned) ? "%u\n" \189 : type_is(var, long) ? "%li\n" \190 : type_is(var, unsigned long) ? "%lu\n" \191 : type_is(var, s64) ? "%lli\n" \192 : type_is(var, u64) ? "%llu\n" \193 : type_is(var, char *) ? "%s\n" \194 : "%i\n", var)195 196bool bch2_is_zero(const void *, size_t);197 198u64 bch2_read_flag_list(const char *, const char * const[]);199 200void bch2_prt_u64_base2_nbits(struct printbuf *, u64, unsigned);201void bch2_prt_u64_base2(struct printbuf *, u64);202 203void bch2_print_string_as_lines(const char *prefix, const char *lines);204void bch2_print_string_as_lines_nonblocking(const char *prefix, const char *lines);205 206typedef DARRAY(unsigned long) bch_stacktrace;207int bch2_save_backtrace(bch_stacktrace *stack, struct task_struct *, unsigned, gfp_t);208void bch2_prt_backtrace(struct printbuf *, bch_stacktrace *);209int bch2_prt_task_backtrace(struct printbuf *, struct task_struct *, unsigned, gfp_t);210 211static inline void prt_bdevname(struct printbuf *out, struct block_device *bdev)212{213#ifdef __KERNEL__214 prt_printf(out, "%pg", bdev);215#else216 prt_str(out, bdev->name);217#endif218}219 220void bch2_time_stats_to_text(struct printbuf *, struct bch2_time_stats *);221 222#define ewma_add(ewma, val, weight) \223({ \224 typeof(ewma) _ewma = (ewma); \225 typeof(weight) _weight = (weight); \226 \227 (((_ewma << _weight) - _ewma) + (val)) >> _weight; \228})229 230struct bch_ratelimit {231 /* Next time we want to do some work, in nanoseconds */232 u64 next;233 234 /*235 * Rate at which we want to do work, in units per nanosecond236 * The units here correspond to the units passed to237 * bch2_ratelimit_increment()238 */239 unsigned rate;240};241 242static inline void bch2_ratelimit_reset(struct bch_ratelimit *d)243{244 d->next = local_clock();245}246 247u64 bch2_ratelimit_delay(struct bch_ratelimit *);248void bch2_ratelimit_increment(struct bch_ratelimit *, u64);249 250struct bch_pd_controller {251 struct bch_ratelimit rate;252 unsigned long last_update;253 254 s64 last_actual;255 s64 smoothed_derivative;256 257 unsigned p_term_inverse;258 unsigned d_smooth;259 unsigned d_term;260 261 /* for exporting to sysfs (no effect on behavior) */262 s64 last_derivative;263 s64 last_proportional;264 s64 last_change;265 s64 last_target;266 267 /*268 * If true, the rate will not increase if bch2_ratelimit_delay()269 * is not being called often enough.270 */271 bool backpressure;272};273 274void bch2_pd_controller_update(struct bch_pd_controller *, s64, s64, int);275void bch2_pd_controller_init(struct bch_pd_controller *);276void bch2_pd_controller_debug_to_text(struct printbuf *, struct bch_pd_controller *);277 278#define sysfs_pd_controller_attribute(name) \279 rw_attribute(name##_rate); \280 rw_attribute(name##_rate_bytes); \281 rw_attribute(name##_rate_d_term); \282 rw_attribute(name##_rate_p_term_inverse); \283 read_attribute(name##_rate_debug)284 285#define sysfs_pd_controller_files(name) \286 &sysfs_##name##_rate, \287 &sysfs_##name##_rate_bytes, \288 &sysfs_##name##_rate_d_term, \289 &sysfs_##name##_rate_p_term_inverse, \290 &sysfs_##name##_rate_debug291 292#define sysfs_pd_controller_show(name, var) \293do { \294 sysfs_hprint(name##_rate, (var)->rate.rate); \295 sysfs_print(name##_rate_bytes, (var)->rate.rate); \296 sysfs_print(name##_rate_d_term, (var)->d_term); \297 sysfs_print(name##_rate_p_term_inverse, (var)->p_term_inverse); \298 \299 if (attr == &sysfs_##name##_rate_debug) \300 bch2_pd_controller_debug_to_text(out, var); \301} while (0)302 303#define sysfs_pd_controller_store(name, var) \304do { \305 sysfs_strtoul_clamp(name##_rate, \306 (var)->rate.rate, 1, UINT_MAX); \307 sysfs_strtoul_clamp(name##_rate_bytes, \308 (var)->rate.rate, 1, UINT_MAX); \309 sysfs_strtoul(name##_rate_d_term, (var)->d_term); \310 sysfs_strtoul_clamp(name##_rate_p_term_inverse, \311 (var)->p_term_inverse, 1, INT_MAX); \312} while (0)313 314#define container_of_or_null(ptr, type, member) \315({ \316 typeof(ptr) _ptr = ptr; \317 _ptr ? container_of(_ptr, type, member) : NULL; \318})319 320/* Does linear interpolation between powers of two */321static inline unsigned fract_exp_two(unsigned x, unsigned fract_bits)322{323 unsigned fract = x & ~(~0 << fract_bits);324 325 x >>= fract_bits;326 x = 1 << x;327 x += (x * fract) >> fract_bits;328 329 return x;330}331 332void bch2_bio_map(struct bio *bio, void *base, size_t);333int bch2_bio_alloc_pages(struct bio *, size_t, gfp_t);334 335#define closure_bio_submit(bio, cl) \336do { \337 closure_get(cl); \338 submit_bio(bio); \339} while (0)340 341#define kthread_wait(cond) \342({ \343 int _ret = 0; \344 \345 while (1) { \346 set_current_state(TASK_INTERRUPTIBLE); \347 if (kthread_should_stop()) { \348 _ret = -1; \349 break; \350 } \351 \352 if (cond) \353 break; \354 \355 schedule(); \356 } \357 set_current_state(TASK_RUNNING); \358 _ret; \359})360 361#define kthread_wait_freezable(cond) \362({ \363 int _ret = 0; \364 while (1) { \365 set_current_state(TASK_INTERRUPTIBLE); \366 if (kthread_should_stop()) { \367 _ret = -1; \368 break; \369 } \370 \371 if (cond) \372 break; \373 \374 schedule(); \375 try_to_freeze(); \376 } \377 set_current_state(TASK_RUNNING); \378 _ret; \379})380 381size_t bch2_rand_range(size_t);382 383void memcpy_to_bio(struct bio *, struct bvec_iter, const void *);384void memcpy_from_bio(void *, struct bio *, struct bvec_iter);385 386static inline void memcpy_u64s_small(void *dst, const void *src,387 unsigned u64s)388{389 u64 *d = dst;390 const u64 *s = src;391 392 while (u64s--)393 *d++ = *s++;394}395 396static inline void __memcpy_u64s(void *dst, const void *src,397 unsigned u64s)398{399#ifdef CONFIG_X86_64400 long d0, d1, d2;401 402 asm volatile("rep ; movsq"403 : "=&c" (d0), "=&D" (d1), "=&S" (d2)404 : "0" (u64s), "1" (dst), "2" (src)405 : "memory");406#else407 u64 *d = dst;408 const u64 *s = src;409 410 while (u64s--)411 *d++ = *s++;412#endif413}414 415static inline void memcpy_u64s(void *dst, const void *src,416 unsigned u64s)417{418 EBUG_ON(!(dst >= src + u64s * sizeof(u64) ||419 dst + u64s * sizeof(u64) <= src));420 421 __memcpy_u64s(dst, src, u64s);422}423 424static inline void __memmove_u64s_down(void *dst, const void *src,425 unsigned u64s)426{427 __memcpy_u64s(dst, src, u64s);428}429 430static inline void memmove_u64s_down(void *dst, const void *src,431 unsigned u64s)432{433 EBUG_ON(dst > src);434 435 __memmove_u64s_down(dst, src, u64s);436}437 438static inline void __memmove_u64s_down_small(void *dst, const void *src,439 unsigned u64s)440{441 memcpy_u64s_small(dst, src, u64s);442}443 444static inline void memmove_u64s_down_small(void *dst, const void *src,445 unsigned u64s)446{447 EBUG_ON(dst > src);448 449 __memmove_u64s_down_small(dst, src, u64s);450}451 452static inline void __memmove_u64s_up_small(void *_dst, const void *_src,453 unsigned u64s)454{455 u64 *dst = (u64 *) _dst + u64s;456 u64 *src = (u64 *) _src + u64s;457 458 while (u64s--)459 *--dst = *--src;460}461 462static inline void memmove_u64s_up_small(void *dst, const void *src,463 unsigned u64s)464{465 EBUG_ON(dst < src);466 467 __memmove_u64s_up_small(dst, src, u64s);468}469 470static inline void __memmove_u64s_up(void *_dst, const void *_src,471 unsigned u64s)472{473 u64 *dst = (u64 *) _dst + u64s - 1;474 u64 *src = (u64 *) _src + u64s - 1;475 476#ifdef CONFIG_X86_64477 long d0, d1, d2;478 479 asm volatile("std ;\n"480 "rep ; movsq\n"481 "cld ;\n"482 : "=&c" (d0), "=&D" (d1), "=&S" (d2)483 : "0" (u64s), "1" (dst), "2" (src)484 : "memory");485#else486 while (u64s--)487 *dst-- = *src--;488#endif489}490 491static inline void memmove_u64s_up(void *dst, const void *src,492 unsigned u64s)493{494 EBUG_ON(dst < src);495 496 __memmove_u64s_up(dst, src, u64s);497}498 499static inline void memmove_u64s(void *dst, const void *src,500 unsigned u64s)501{502 if (dst < src)503 __memmove_u64s_down(dst, src, u64s);504 else505 __memmove_u64s_up(dst, src, u64s);506}507 508/* Set the last few bytes up to a u64 boundary given an offset into a buffer. */509static inline void memset_u64s_tail(void *s, int c, unsigned bytes)510{511 unsigned rem = round_up(bytes, sizeof(u64)) - bytes;512 513 memset(s + bytes, c, rem);514}515 516/* just the memmove, doesn't update @_nr */517#define __array_insert_item(_array, _nr, _pos) \518 memmove(&(_array)[(_pos) + 1], \519 &(_array)[(_pos)], \520 sizeof((_array)[0]) * ((_nr) - (_pos)))521 522#define array_insert_item(_array, _nr, _pos, _new_item) \523do { \524 __array_insert_item(_array, _nr, _pos); \525 (_nr)++; \526 (_array)[(_pos)] = (_new_item); \527} while (0)528 529#define array_remove_items(_array, _nr, _pos, _nr_to_remove) \530do { \531 (_nr) -= (_nr_to_remove); \532 memmove(&(_array)[(_pos)], \533 &(_array)[(_pos) + (_nr_to_remove)], \534 sizeof((_array)[0]) * ((_nr) - (_pos))); \535} while (0)536 537#define array_remove_item(_array, _nr, _pos) \538 array_remove_items(_array, _nr, _pos, 1)539 540static inline void __move_gap(void *array, size_t element_size,541 size_t nr, size_t size,542 size_t old_gap, size_t new_gap)543{544 size_t gap_end = old_gap + size - nr;545 546 if (new_gap < old_gap) {547 size_t move = old_gap - new_gap;548 549 memmove(array + element_size * (gap_end - move),550 array + element_size * (old_gap - move),551 element_size * move);552 } else if (new_gap > old_gap) {553 size_t move = new_gap - old_gap;554 555 memmove(array + element_size * old_gap,556 array + element_size * gap_end,557 element_size * move);558 }559}560 561/* Move the gap in a gap buffer: */562#define move_gap(_d, _new_gap) \563do { \564 BUG_ON(_new_gap > (_d)->nr); \565 BUG_ON((_d)->gap > (_d)->nr); \566 \567 __move_gap((_d)->data, sizeof((_d)->data[0]), \568 (_d)->nr, (_d)->size, (_d)->gap, _new_gap); \569 (_d)->gap = _new_gap; \570} while (0)571 572#define bubble_sort(_base, _nr, _cmp) \573do { \574 ssize_t _i, _last; \575 bool _swapped = true; \576 \577 for (_last= (ssize_t) (_nr) - 1; _last > 0 && _swapped; --_last) {\578 _swapped = false; \579 for (_i = 0; _i < _last; _i++) \580 if (_cmp((_base)[_i], (_base)[_i + 1]) > 0) { \581 swap((_base)[_i], (_base)[_i + 1]); \582 _swapped = true; \583 } \584 } \585} while (0)586 587#define per_cpu_sum(_p) \588({ \589 typeof(*_p) _ret = 0; \590 \591 int cpu; \592 for_each_possible_cpu(cpu) \593 _ret += *per_cpu_ptr(_p, cpu); \594 _ret; \595})596 597static inline u64 percpu_u64_get(u64 __percpu *src)598{599 return per_cpu_sum(src);600}601 602static inline void percpu_u64_set(u64 __percpu *dst, u64 src)603{604 int cpu;605 606 for_each_possible_cpu(cpu)607 *per_cpu_ptr(dst, cpu) = 0;608 this_cpu_write(*dst, src);609}610 611static inline void acc_u64s(u64 *acc, const u64 *src, unsigned nr)612{613 for (unsigned i = 0; i < nr; i++)614 acc[i] += src[i];615}616 617static inline void acc_u64s_percpu(u64 *acc, const u64 __percpu *src,618 unsigned nr)619{620 int cpu;621 622 for_each_possible_cpu(cpu)623 acc_u64s(acc, per_cpu_ptr(src, cpu), nr);624}625 626static inline void percpu_memset(void __percpu *p, int c, size_t bytes)627{628 int cpu;629 630 for_each_possible_cpu(cpu)631 memset(per_cpu_ptr(p, cpu), c, bytes);632}633 634u64 *bch2_acc_percpu_u64s(u64 __percpu *, unsigned);635 636#define cmp_int(l, r) ((l > r) - (l < r))637 638static inline int u8_cmp(u8 l, u8 r)639{640 return cmp_int(l, r);641}642 643static inline int cmp_le32(__le32 l, __le32 r)644{645 return cmp_int(le32_to_cpu(l), le32_to_cpu(r));646}647 648#include <linux/uuid.h>649 650#define QSTR(n) { { { .len = strlen(n) } }, .name = n }651 652static inline bool qstr_eq(const struct qstr l, const struct qstr r)653{654 return l.len == r.len && !memcmp(l.name, r.name, l.len);655}656 657void bch2_darray_str_exit(darray_str *);658int bch2_split_devs(const char *, darray_str *);659 660#ifdef __KERNEL__661 662__must_check663static inline int copy_to_user_errcode(void __user *to, const void *from, unsigned long n)664{665 return copy_to_user(to, from, n) ? -EFAULT : 0;666}667 668__must_check669static inline int copy_from_user_errcode(void *to, const void __user *from, unsigned long n)670{671 return copy_from_user(to, from, n) ? -EFAULT : 0;672}673 674#endif675 676static inline void mod_bit(long nr, volatile unsigned long *addr, bool v)677{678 if (v)679 set_bit(nr, addr);680 else681 clear_bit(nr, addr);682}683 684static inline void __set_bit_le64(size_t bit, __le64 *addr)685{686 addr[bit / 64] |= cpu_to_le64(BIT_ULL(bit % 64));687}688 689static inline void __clear_bit_le64(size_t bit, __le64 *addr)690{691 addr[bit / 64] &= ~cpu_to_le64(BIT_ULL(bit % 64));692}693 694static inline bool test_bit_le64(size_t bit, __le64 *addr)695{696 return (addr[bit / 64] & cpu_to_le64(BIT_ULL(bit % 64))) != 0;697}698 699#endif /* _BCACHEFS_UTIL_H */700