17 #include "lib/axmap.h"
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
23 static FLIST_HEAD(filename_list);
26 * List entry for filename_list
29 struct flist_head list;
33 static inline void clear_error(struct thread_data *td)
39 static int native_fallocate(struct thread_data *td, struct fio_file *f)
43 success = fio_fallocate(f, 0, f->real_file_size);
44 dprint(FD_FILE, "native fallocate of file %s size %llu was "
45 "%ssuccessful\n", f->file_name,
46 (unsigned long long) f->real_file_size,
53 dprint(FD_FILE, "native fallocate is not implemented\n");
58 static void fallocate_file(struct thread_data *td, struct fio_file *f)
60 if (td->o.fill_device)
63 switch (td->o.fallocate_mode) {
64 case FIO_FALLOCATE_NATIVE:
65 native_fallocate(td, f);
67 case FIO_FALLOCATE_NONE:
69 #ifdef CONFIG_POSIX_FALLOCATE
70 case FIO_FALLOCATE_POSIX: {
73 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
75 (unsigned long long) f->real_file_size);
77 r = posix_fallocate(f->fd, 0, f->real_file_size);
79 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
82 #endif /* CONFIG_POSIX_FALLOCATE */
83 #ifdef CONFIG_LINUX_FALLOCATE
84 case FIO_FALLOCATE_KEEP_SIZE: {
87 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
88 "file %s size %llu\n", f->file_name,
89 (unsigned long long) f->real_file_size);
91 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
93 td_verror(td, errno, "fallocate");
97 #endif /* CONFIG_LINUX_FALLOCATE */
99 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
105 * Leaves f->fd open on success, caller must close
107 static int extend_file(struct thread_data *td, struct fio_file *f)
109 int new_layout = 0, unlink_file = 0, flags;
110 unsigned long long left;
115 log_err("fio: refusing extend of file due to read-only\n");
120 * check if we need to lay the file out complete again. fio
121 * does that for operations involving reads, or for writes
122 * where overwrite is set
125 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
126 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
128 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
131 if (unlink_file || new_layout) {
134 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
136 ret = td_io_unlink_file(td, f);
137 if (ret != 0 && ret != ENOENT) {
138 td_verror(td, errno, "unlink");
144 if (td->o.allow_create)
153 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
154 f->fd = open(f->file_name, flags, 0644);
158 if (err == ENOENT && !td->o.allow_create)
159 log_err("fio: file creation disallowed by "
160 "allow_file_create=0\n");
162 td_verror(td, err, "open");
166 fallocate_file(td, f);
169 * If our jobs don't require regular files initially, we're done.
175 * The size will be -1ULL when fill_device is used, so don't truncate
176 * or fallocate this file, just write it
178 if (!td->o.fill_device) {
179 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
180 (unsigned long long) f->real_file_size);
181 if (ftruncate(f->fd, f->real_file_size) == -1) {
182 if (errno != EFBIG) {
183 td_verror(td, errno, "ftruncate");
189 left = f->real_file_size;
190 bs = td->o.max_bs[DDIR_WRITE];
196 td_verror(td, errno, "malloc");
200 while (left && !td->terminate) {
206 fill_io_buffer(td, b, bs, bs);
208 r = write(f->fd, b, bs);
218 if (td->o.fill_device)
220 log_info("fio: ENOSPC on laying out "
224 td_verror(td, errno, "write");
226 td_verror(td, EIO, "write");
233 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
234 td_io_unlink_file(td, f);
235 } else if (td->o.create_fsync) {
236 if (fsync(f->fd) < 0) {
237 td_verror(td, errno, "fsync");
241 if (td->o.fill_device && !td_write(td)) {
242 fio_file_clear_size_known(f);
243 if (td_io_get_file_size(td, f))
245 if (f->io_size > f->real_file_size)
246 f->io_size = f->real_file_size;
260 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
262 int r, did_open = 0, old_runstate;
263 unsigned long long left;
268 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
269 td_ioengine_flagged(td, FIO_NOIO))
272 if (f->filetype == FIO_TYPE_CHAR)
275 if (!fio_file_open(f)) {
276 if (td->io_ops->open_file(td, f)) {
277 log_err("fio: cannot pre-read, failed to open file\n");
283 old_runstate = td_bump_runstate(td, TD_PRE_READING);
286 bs = td->o.max_bs[DDIR_READ];
292 td_verror(td, errno, "malloc");
298 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
299 td_verror(td, errno, "lseek");
300 log_err("fio: failed to lseek pre-read file\n");
305 while (left && !td->terminate) {
309 r = read(f->fd, b, bs);
315 td_verror(td, EIO, "pre_read");
321 td_restore_runstate(td, old_runstate);
324 td->io_ops->close_file(td, f);
330 unsigned long long get_rand_file_size(struct thread_data *td)
332 unsigned long long ret, sized;
336 frand_max = rand_max(&td->file_size_state);
337 r = __rand(&td->file_size_state);
338 sized = td->o.file_size_high - td->o.file_size_low;
339 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
340 ret += td->o.file_size_low;
341 ret -= (ret % td->o.rw_min_bs);
345 static int file_size(struct thread_data *td, struct fio_file *f)
349 if (stat(f->file_name, &st) == -1) {
350 td_verror(td, errno, "fstat");
354 f->real_file_size = st.st_size;
358 static int bdev_size(struct thread_data *td, struct fio_file *f)
360 unsigned long long bytes = 0;
363 if (td->io_ops->open_file(td, f)) {
364 log_err("fio: failed opening blockdev %s for size check\n",
369 r = blockdev_size(f, &bytes);
371 td_verror(td, r, "blockdev_size");
376 log_err("%s: zero sized block device?\n", f->file_name);
380 f->real_file_size = bytes;
381 td->io_ops->close_file(td, f);
384 td->io_ops->close_file(td, f);
388 static int char_size(struct thread_data *td, struct fio_file *f)
390 #ifdef FIO_HAVE_CHARDEV_SIZE
391 unsigned long long bytes = 0;
394 if (td->io_ops->open_file(td, f)) {
395 log_err("fio: failed opening chardev %s for size check\n",
400 r = chardev_size(f, &bytes);
402 td_verror(td, r, "chardev_size");
407 log_err("%s: zero sized char device?\n", f->file_name);
411 f->real_file_size = bytes;
412 td->io_ops->close_file(td, f);
415 td->io_ops->close_file(td, f);
418 f->real_file_size = -1ULL;
423 static int get_file_size(struct thread_data *td, struct fio_file *f)
427 if (fio_file_size_known(f))
430 if (f->filetype == FIO_TYPE_FILE)
431 ret = file_size(td, f);
432 else if (f->filetype == FIO_TYPE_BLOCK)
433 ret = bdev_size(td, f);
434 else if (f->filetype == FIO_TYPE_CHAR)
435 ret = char_size(td, f);
437 f->real_file_size = -1;
438 log_info("%s: failed to get file size of %s\n", td->o.name,
440 return 1; /* avoid offset extends end error message */
444 * Leave ->real_file_size with 0 since it could be expectation
445 * of initial setup for regular files.
451 * ->file_offset normally hasn't been initialized yet, so this
452 * is basically always false unless ->real_file_size is -1, but
453 * if ->real_file_size is -1 this message doesn't make sense.
454 * As a result, this message is basically useless.
456 if (f->file_offset > f->real_file_size) {
457 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
458 (unsigned long long) f->file_offset,
459 (unsigned long long) f->real_file_size);
463 fio_file_set_size_known(f);
467 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
468 unsigned long long off,
469 unsigned long long len)
471 int errval = 0, ret = 0;
480 off = f->file_offset;
482 if (len == -1ULL || off == -1ULL)
485 if (td->io_ops->invalidate) {
486 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
488 ret = td->io_ops->invalidate(td, f);
491 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
492 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
495 } else if (f->filetype == FIO_TYPE_FILE) {
496 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
497 f->file_name, off, len);
498 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
501 } else if (f->filetype == FIO_TYPE_BLOCK) {
504 dprint(FD_IO, "drop page cache %s\n", f->file_name);
505 ret = blockdev_invalidate_cache(f);
506 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
508 * Linux multipath devices reject ioctl while
509 * the maps are being updated. That window can
510 * last tens of milliseconds; we'll try up to
511 * a quarter of a second.
514 ret = blockdev_invalidate_cache(f);
516 if (ret < 0 && errno == EACCES && geteuid()) {
517 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
518 log_err("fio: only root may flush block "
519 "devices. Cache flush bypassed!\n");
525 } else if (f->filetype == FIO_TYPE_CHAR ||
526 f->filetype == FIO_TYPE_PIPE) {
527 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
532 * Cache flushing isn't a fatal condition, and we know it will
533 * happen on some platforms where we don't have the proper
534 * function to flush eg block device caches. So just warn and
535 * continue on our way.
538 log_info("fio: cache invalidation of %s failed: %s\n",
539 f->file_name, strerror(errval));
545 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
547 if (!fio_file_open(f))
550 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
553 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
557 dprint(FD_FILE, "fd close %s\n", f->file_name);
561 if (close(f->fd) < 0)
566 if (f->shadow_fd != -1) {
575 int file_lookup_open(struct fio_file *f, int flags)
577 struct fio_file *__f;
580 __f = lookup_file_hash(f->file_name);
582 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
586 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
594 f->fd = open(f->file_name, flags, 0600);
598 static int file_close_shadow_fds(struct thread_data *td)
604 for_each_file(td, f, i) {
605 if (f->shadow_fd == -1)
616 int generic_open_file(struct thread_data *td, struct fio_file *f)
622 dprint(FD_FILE, "fd open %s\n", f->file_name);
624 if (!strcmp(f->file_name, "-")) {
626 log_err("fio: can't read/write to stdin/out\n");
632 * move output logging to stderr, if we are writing to stdout
641 flags |= OS_O_DIRECT;
644 td_verror(td, EINVAL, "OS does not support atomic IO");
647 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
651 if (td->o.create_on_open && td->o.allow_create)
654 if (f->filetype != FIO_TYPE_FILE)
655 flags |= FIO_O_NOATIME;
662 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
666 f->fd = dup(STDOUT_FILENO);
668 from_hash = file_lookup_open(f, flags);
669 } else if (td_read(td)) {
670 if (f->filetype == FIO_TYPE_CHAR && !read_only)
676 f->fd = dup(STDIN_FILENO);
678 from_hash = file_lookup_open(f, flags);
679 } else if (td_trim(td)) {
680 assert(!td_rw(td)); /* should have matched above */
682 from_hash = file_lookup_open(f, flags);
686 char buf[FIO_VERROR_SIZE];
689 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
690 flags &= ~FIO_O_NOATIME;
693 if (__e == EMFILE && file_close_shadow_fds(td))
696 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
698 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
699 log_err("fio: looks like your file system does not " \
700 "support direct=1/buffered=0\n");
703 td_verror(td, __e, buf);
707 if (!from_hash && f->fd != -1) {
708 if (add_file_hash(f)) {
712 * Stash away descriptor for later close. This is to
713 * work-around a "feature" on Linux, where a close of
714 * an fd that has been opened for write will trigger
715 * udev to call blkid to check partitions, fs id, etc.
716 * That pollutes the device cache, which can slow down
717 * unbuffered accesses.
719 if (f->shadow_fd == -1)
720 f->shadow_fd = f->fd;
723 * OK to ignore, we haven't done anything
726 ret = generic_close_file(td, f);
736 * This function i.e. get_file_size() is the default .get_file_size
737 * implementation of majority of I/O engines.
739 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
741 return get_file_size(td, f);
745 * open/close all files, so that ->real_file_size gets set
747 static int get_file_sizes(struct thread_data *td)
753 for_each_file(td, f, i) {
754 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
757 if (td_io_get_file_size(td, f)) {
758 if (td->error != ENOENT) {
759 log_err("%s\n", td->verror);
767 * There are corner cases where we end up with -1 for
768 * ->real_file_size due to unsupported file type, etc.
769 * We then just set to size option value divided by number
770 * of files, similar to the way file ->io_size is set.
771 * stat(2) failure doesn't set ->real_file_size to -1.
773 if (f->real_file_size == -1ULL && td->o.size)
774 f->real_file_size = td->o.size / td->o.nr_files;
781 struct flist_head list;
788 * Get free number of bytes for each file on each unique mount.
790 static unsigned long long get_fs_free_counts(struct thread_data *td)
792 struct flist_head *n, *tmp;
793 unsigned long long ret = 0;
794 struct fio_mount *fm;
799 for_each_file(td, f, i) {
803 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
804 if (f->real_file_size != -1ULL)
805 ret += f->real_file_size;
807 } else if (f->filetype != FIO_TYPE_FILE)
811 strncpy(buf, f->file_name, 255);
813 if (stat(buf, &sb) < 0) {
817 if (stat(buf, &sb) < 0)
822 flist_for_each(n, &list) {
823 fm = flist_entry(n, struct fio_mount, list);
824 if (fm->key == sb.st_dev)
833 fm = calloc(1, sizeof(*fm));
834 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
835 fm->base = basename(fm->__base);
837 flist_add(&fm->list, &list);
840 flist_for_each_safe(n, tmp, &list) {
841 unsigned long long sz;
843 fm = flist_entry(n, struct fio_mount, list);
844 flist_del(&fm->list);
846 sz = get_fs_free_size(fm->base);
847 if (sz && sz != -1ULL)
856 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
858 struct thread_options *o = &td->o;
859 unsigned long long align_bs;
860 unsigned long long offset;
862 if (o->file_append && f->filetype == FIO_TYPE_FILE)
863 return f->real_file_size;
865 if (o->start_offset_percent > 0) {
867 * if offset_align is provided, set initial offset
869 if (fio_option_is_set(o, start_offset_align)) {
870 align_bs = o->start_offset_align;
872 /* else take the minimum block size */
873 align_bs = td_min_bs(td);
876 /* calculate the raw offset */
877 offset = (f->real_file_size * o->start_offset_percent / 100) +
878 (td->subjob_number * o->offset_increment);
881 * block align the offset at the next available boundary at
882 * ceiling(offset / align_bs) * align_bs
884 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
887 /* start_offset_percent not set */
888 offset = o->start_offset +
889 td->subjob_number * o->offset_increment;
896 * Open the files and setup files sizes, creating files if necessary.
898 int setup_files(struct thread_data *td)
900 unsigned long long total_size, extend_size;
901 struct thread_options *o = &td->o;
903 unsigned int i, nr_fs_extra = 0;
904 int err = 0, need_extend;
906 const unsigned int bs = td_min_bs(td);
909 dprint(FD_FILE, "setup files\n");
911 old_state = td_bump_runstate(td, TD_SETTING_UP);
913 if (o->read_iolog_file)
917 * Find out physical size of files or devices for this thread,
918 * before we determine I/O size and range of our targets.
919 * If ioengine defines a setup() method, it's responsible for
920 * opening the files and setting f->real_file_size to indicate
921 * the valid range for that file.
923 if (td->io_ops->setup)
924 err = td->io_ops->setup(td);
926 err = get_file_sizes(td);
932 * check sizes. if the files/devices do not exist and the size
933 * isn't passed to fio, abort.
936 for_each_file(td, f, i) {
938 if (f->real_file_size == -1ULL)
941 total_size += f->real_file_size;
945 td->fill_device_size = get_fs_free_counts(td);
948 * device/file sizes are zero and no size given, punt
950 if ((!total_size || total_size == -1ULL) && !o->size &&
951 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
952 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
953 log_err("%s: you need to specify size=\n", o->name);
954 td_verror(td, EINVAL, "total_file_size");
959 * Calculate per-file size and potential extra size for the
960 * first files, if needed (i.e. if we don't have a fixed size).
962 if (!o->file_size_low && o->nr_files) {
965 fs = o->size / o->nr_files;
966 all_fs = fs * o->nr_files;
968 if (all_fs < o->size)
969 nr_fs_extra = (o->size - all_fs) / bs;
973 * now file sizes are known, so we can set ->io_size. if size= is
974 * not given, ->io_size is just equal to ->real_file_size. if size
975 * is given, ->io_size is size / nr_files.
977 extend_size = total_size = 0;
979 for_each_file(td, f, i) {
980 f->file_offset = get_start_offset(td, f);
983 * Update ->io_size depending on options specified.
984 * ->file_size_low being 0 means filesize option isn't set.
985 * Non zero ->file_size_low equals ->file_size_high means
986 * filesize option is set in a fixed size format.
987 * Non zero ->file_size_low not equals ->file_size_high means
988 * filesize option is set in a range format.
990 if (!o->file_size_low) {
992 * no file size or range given, file size is equal to
993 * total size divided by number of files. If the size
994 * doesn't divide nicely with the min blocksize,
995 * make the first files bigger.
1004 * We normally don't come here for regular files, but
1005 * if the result is 0 for a regular file, set it to the
1006 * real file size. This could be size of the existing
1007 * one if it already exists, but otherwise will be set
1008 * to 0. A new file won't be created because
1009 * ->io_size + ->file_offset equals ->real_file_size.
1012 if (f->file_offset > f->real_file_size)
1014 f->io_size = f->real_file_size - f->file_offset;
1016 log_info("fio: file %s may be ignored\n",
1019 } else if (f->real_file_size < o->file_size_low ||
1020 f->real_file_size > o->file_size_high) {
1021 if (f->file_offset > o->file_size_low)
1024 * file size given. if it's fixed, use that. if it's a
1025 * range, generate a random size in-between.
1027 if (o->file_size_low == o->file_size_high)
1028 f->io_size = o->file_size_low - f->file_offset;
1030 f->io_size = get_rand_file_size(td)
1034 f->io_size = f->real_file_size - f->file_offset;
1036 if (f->io_size == -1ULL)
1039 if (o->size_percent) {
1042 file_size = f->io_size + f->file_offset;
1043 f->io_size = (file_size *
1044 o->size_percent) / 100;
1045 if (f->io_size > (file_size - f->file_offset))
1046 f->io_size = file_size - f->file_offset;
1048 f->io_size -= (f->io_size % td_min_bs(td));
1050 total_size += f->io_size;
1053 if (f->filetype == FIO_TYPE_FILE &&
1054 (f->io_size + f->file_offset) > f->real_file_size &&
1055 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1056 if (!o->create_on_open) {
1058 extend_size += (f->io_size + f->file_offset);
1059 fio_file_set_extend(f);
1061 f->real_file_size = f->io_size + f->file_offset;
1065 if (td->o.block_error_hist) {
1068 assert(td->o.nr_files == 1); /* checked in fixup_options */
1070 len = f->io_size / td->o.bs[DDIR_TRIM];
1071 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1072 log_err("fio: cannot calculate block histogram with "
1073 "%d trim blocks, maximum %d\n",
1074 len, MAX_NR_BLOCK_INFOS);
1075 td_verror(td, EINVAL, "block_error_hist");
1079 td->ts.nr_block_infos = len;
1080 for (i = 0; i < len; i++)
1081 td->ts.block_infos[i] =
1082 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1084 td->ts.nr_block_infos = 0;
1086 if (!o->size || (total_size && o->size > total_size))
1087 o->size = total_size;
1089 if (o->size < td_min_bs(td)) {
1090 log_err("fio: blocksize too large for data set\n");
1095 * See if we need to extend some files, typically needed when our
1096 * target regular files don't exist yet, but our jobs require them
1097 * initially due to read I/Os.
1101 if (output_format & FIO_OUTPUT_NORMAL) {
1102 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1104 need_extend > 1 ? "s" : "",
1106 need_extend > 1 ? "s" : "",
1107 need_extend > 1 ? "total " : "",
1111 for_each_file(td, f, i) {
1112 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1114 if (!fio_file_extend(f))
1117 assert(f->filetype == FIO_TYPE_FILE);
1118 fio_file_clear_extend(f);
1119 if (!o->fill_device) {
1120 old_len = f->real_file_size;
1121 extend_len = f->io_size + f->file_offset -
1124 f->real_file_size = (f->io_size + f->file_offset);
1125 err = extend_file(td, f);
1129 err = __file_invalidate_cache(td, f, old_len,
1133 * Shut up static checker
1149 o->zone_size = o->size;
1152 * iolog already set the total io size, if we read back
1155 if (!o->read_iolog_file) {
1157 td->total_io_size = o->io_size * o->loops;
1159 td->total_io_size = o->size * o->loops;
1166 td_restore_runstate(td, old_state);
1169 log_err("%s: you need to specify valid offset=\n", o->name);
1171 td_restore_runstate(td, old_state);
1175 bool pre_read_files(struct thread_data *td)
1180 dprint(FD_FILE, "pre_read files\n");
1182 for_each_file(td, f, i) {
1183 if (!pre_read_file(td, f))
1190 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1192 unsigned int range_size, seed;
1193 unsigned long nranges;
1196 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1197 fsize = min(f->real_file_size, f->io_size);
1199 nranges = (fsize + range_size - 1) / range_size;
1201 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1202 if (!td->o.rand_repeatable)
1203 seed = td->rand_seeds[4];
1205 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1206 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1207 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1208 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1209 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1210 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1213 static bool init_rand_distribution(struct thread_data *td)
1219 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1222 state = td_bump_runstate(td, TD_SETTING_UP);
1224 for_each_file(td, f, i)
1225 __init_rand_distribution(td, f);
1227 td_restore_runstate(td, state);
1232 * Check if the number of blocks exceeds the randomness capability of
1233 * the selected generator. Tausworthe is 32-bit, the others are fullly
1236 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1239 if (blocks <= FRAND32_MAX)
1241 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1245 * If the user hasn't specified a random generator, switch
1246 * to tausworthe64 with informational warning. If the user did
1247 * specify one, just warn.
1249 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1252 if (!fio_option_is_set(&td->o, random_generator)) {
1253 log_info("fio: Switching to tausworthe64. Use the "
1254 "random_generator= option to get rid of this "
1256 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1261 * Just make this information to avoid breaking scripts.
1263 log_info("fio: Use the random_generator= option to switch to lfsr or "
1268 bool init_random_map(struct thread_data *td)
1270 unsigned long long blocks;
1274 if (init_rand_distribution(td))
1279 for_each_file(td, f, i) {
1280 uint64_t fsize = min(f->real_file_size, f->io_size);
1282 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1284 if (check_rand_gen_limits(td, f, blocks))
1287 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1290 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1292 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1293 fio_file_set_lfsr(f);
1296 } else if (!td->o.norandommap) {
1297 f->io_axmap = axmap_new(blocks);
1299 fio_file_set_axmap(f);
1302 } else if (td->o.norandommap)
1305 if (!td->o.softrandommap) {
1306 log_err("fio: failed allocating random map. If running"
1307 " a large number of jobs, try the 'norandommap'"
1308 " option or set 'softrandommap'. Or give"
1309 " a larger --alloc-size to fio.\n");
1313 log_info("fio: file %s failed allocating random map. Running "
1314 "job without.\n", f->file_name);
1320 void close_files(struct thread_data *td)
1325 for_each_file(td, f, i) {
1326 if (fio_file_open(f))
1327 td_io_close_file(td, f);
1331 void close_and_free_files(struct thread_data *td)
1335 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1337 dprint(FD_FILE, "close files\n");
1339 for_each_file(td, f, i) {
1340 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1341 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1342 td_io_unlink_file(td, f);
1345 if (fio_file_open(f))
1346 td_io_close_file(td, f);
1348 remove_file_hash(f);
1350 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1351 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1352 td_io_unlink_file(td, f);
1358 sfree(f->file_name);
1359 f->file_name = NULL;
1360 if (fio_file_axmap(f)) {
1361 axmap_free(f->io_axmap);
1370 td->o.filename = NULL;
1372 free(td->file_locks);
1373 td->files_index = 0;
1375 td->file_locks = NULL;
1376 td->o.file_lock_mode = FILE_LOCK_NONE;
1380 static void get_file_type(struct fio_file *f)
1384 if (!strcmp(f->file_name, "-"))
1385 f->filetype = FIO_TYPE_PIPE;
1387 f->filetype = FIO_TYPE_FILE;
1390 /* \\.\ is the device namespace in Windows, where every file is
1392 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1393 f->filetype = FIO_TYPE_BLOCK;
1396 if (!stat(f->file_name, &sb)) {
1397 if (S_ISBLK(sb.st_mode))
1398 f->filetype = FIO_TYPE_BLOCK;
1399 else if (S_ISCHR(sb.st_mode))
1400 f->filetype = FIO_TYPE_CHAR;
1401 else if (S_ISFIFO(sb.st_mode))
1402 f->filetype = FIO_TYPE_PIPE;
1406 static bool __is_already_allocated(const char *fname, bool set)
1408 struct flist_head *entry;
1411 ret = file_bloom_exists(fname, set);
1415 flist_for_each(entry, &filename_list) {
1416 struct file_name *fn;
1418 fn = flist_entry(entry, struct file_name, list);
1420 if (!strcmp(fn->filename, fname))
1427 static bool is_already_allocated(const char *fname)
1431 fio_file_hash_lock();
1432 ret = __is_already_allocated(fname, false);
1433 fio_file_hash_unlock();
1438 static void set_already_allocated(const char *fname)
1440 struct file_name *fn;
1442 fn = malloc(sizeof(struct file_name));
1443 fn->filename = strdup(fname);
1445 fio_file_hash_lock();
1446 if (!__is_already_allocated(fname, true)) {
1447 flist_add_tail(&fn->list, &filename_list);
1450 fio_file_hash_unlock();
1458 static void free_already_allocated(void)
1460 struct flist_head *entry, *tmp;
1461 struct file_name *fn;
1463 if (flist_empty(&filename_list))
1466 fio_file_hash_lock();
1467 flist_for_each_safe(entry, tmp, &filename_list) {
1468 fn = flist_entry(entry, struct file_name, list);
1470 flist_del(&fn->list);
1474 fio_file_hash_unlock();
1477 static struct fio_file *alloc_new_file(struct thread_data *td)
1481 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1482 f = calloc(1, sizeof(*f));
1484 f = smalloc(sizeof(*f));
1492 fio_file_reset(td, f);
1496 bool exists_and_not_regfile(const char *filename)
1500 if (lstat(filename, &sb) == -1)
1503 #ifndef WIN32 /* NOT Windows */
1504 if (S_ISREG(sb.st_mode))
1507 /* \\.\ is the device namespace in Windows, where every file
1508 * is a device node */
1509 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1516 static bool create_work_dirs(struct thread_data *td, const char *fname)
1518 char path[PATH_MAX];
1521 if (td->o.directory) {
1522 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1523 FIO_OS_PATH_SEPARATOR, fname);
1524 start = strstr(path, fname);
1526 snprintf(path, PATH_MAX, "%s", fname);
1531 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1536 #ifdef CONFIG_HAVE_MKDIR_TWO
1537 if (mkdir(path, 0600) && errno != EEXIST) {
1539 if (mkdir(path) && errno != EEXIST) {
1541 log_err("fio: failed to create dir (%s): %d\n",
1545 *end = FIO_OS_PATH_SEPARATOR;
1548 td->flags |= TD_F_DIRS_CREATED;
1552 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1554 int cur_files = td->files_index;
1555 char file_name[PATH_MAX];
1559 dprint(FD_FILE, "add file %s\n", fname);
1561 if (td->o.directory)
1562 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1563 td->o.unique_filename);
1565 sprintf(file_name + len, "%s", fname);
1567 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1568 !(td->flags & TD_F_DIRS_CREATED) &&
1569 !create_work_dirs(td, fname))
1572 /* clean cloned siblings using existing files */
1573 if (numjob && is_already_allocated(file_name) &&
1574 !exists_and_not_regfile(fname))
1577 f = alloc_new_file(td);
1579 if (td->files_size <= td->files_index) {
1580 unsigned int new_size = td->o.nr_files + 1;
1582 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1584 td->files = realloc(td->files, new_size * sizeof(f));
1585 if (td->files == NULL) {
1586 log_err("fio: realloc OOM\n");
1589 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1590 td->file_locks = realloc(td->file_locks, new_size);
1591 if (!td->file_locks) {
1592 log_err("fio: realloc OOM\n");
1595 td->file_locks[cur_files] = FILE_LOCK_NONE;
1597 td->files_size = new_size;
1599 td->files[cur_files] = f;
1600 f->fileno = cur_files;
1603 * init function, io engine may not be loaded yet
1605 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1606 f->real_file_size = -1ULL;
1608 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1609 f->file_name = strdup(file_name);
1611 f->file_name = smalloc_strdup(file_name);
1617 switch (td->o.file_lock_mode) {
1618 case FILE_LOCK_NONE:
1620 case FILE_LOCK_READWRITE:
1621 f->rwlock = fio_rwlock_init();
1623 case FILE_LOCK_EXCLUSIVE:
1624 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1627 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1633 if (td->o.numjobs > 1)
1634 set_already_allocated(file_name);
1639 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1645 int add_file_exclusive(struct thread_data *td, const char *fname)
1650 for_each_file(td, f, i) {
1651 if (!strcmp(f->file_name, fname))
1655 return add_file(td, fname, 0, 1);
1658 void get_file(struct fio_file *f)
1660 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1661 assert(fio_file_open(f));
1665 int put_file(struct thread_data *td, struct fio_file *f)
1667 int f_ret = 0, ret = 0;
1669 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1671 if (!fio_file_open(f)) {
1672 assert(f->fd == -1);
1676 assert(f->references);
1677 if (--f->references)
1680 if (should_fsync(td) && td->o.fsync_on_close) {
1681 f_ret = fsync(f->fd);
1686 if (td->io_ops->close_file)
1687 ret = td->io_ops->close_file(td, f);
1692 td->nr_open_files--;
1693 fio_file_clear_open(f);
1694 assert(f->fd == -1);
1698 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1700 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1703 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1704 if (ddir == DDIR_READ)
1705 fio_rwlock_read(f->rwlock);
1707 fio_rwlock_write(f->rwlock);
1708 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1709 fio_mutex_down(f->lock);
1711 td->file_locks[f->fileno] = td->o.file_lock_mode;
1714 void unlock_file(struct thread_data *td, struct fio_file *f)
1716 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1719 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1720 fio_rwlock_unlock(f->rwlock);
1721 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1722 fio_mutex_up(f->lock);
1724 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1727 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1729 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1731 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1735 static bool recurse_dir(struct thread_data *td, const char *dirname)
1741 D = opendir(dirname);
1743 char buf[FIO_VERROR_SIZE];
1745 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1746 td_verror(td, errno, buf);
1750 while ((dir = readdir(D)) != NULL) {
1751 char full_path[PATH_MAX];
1754 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1757 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1759 if (lstat(full_path, &sb) == -1) {
1760 if (errno != ENOENT) {
1761 td_verror(td, errno, "stat");
1767 if (S_ISREG(sb.st_mode)) {
1768 add_file(td, full_path, 0, 1);
1771 if (!S_ISDIR(sb.st_mode))
1774 ret = recurse_dir(td, full_path);
1783 int add_dir_files(struct thread_data *td, const char *path)
1785 int ret = recurse_dir(td, path);
1788 log_info("fio: opendir added %d files\n", td->o.nr_files);
1793 void dup_files(struct thread_data *td, struct thread_data *org)
1798 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1803 td->files = malloc(org->files_index * sizeof(f));
1805 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1806 td->file_locks = malloc(org->files_index);
1808 for_each_file(org, f, i) {
1809 struct fio_file *__f;
1811 __f = alloc_new_file(td);
1814 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1815 __f->file_name = strdup(f->file_name);
1817 __f->file_name = smalloc_strdup(f->file_name);
1818 if (!__f->file_name)
1821 __f->filetype = f->filetype;
1824 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1825 __f->lock = f->lock;
1826 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1827 __f->rwlock = f->rwlock;
1834 * Returns the index that matches the filename, or -1 if not there
1836 int get_fileno(struct thread_data *td, const char *fname)
1841 for_each_file(td, f, i)
1842 if (!strcmp(f->file_name, fname))
1849 * For log usage, where we add/open/close files automatically
1851 void free_release_files(struct thread_data *td)
1855 td->o.open_files = 0;
1856 td->files_index = 0;
1859 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1863 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1864 f->last_pos[i] = f->file_offset;
1865 f->last_start[i] = -1ULL;
1868 if (fio_file_axmap(f))
1869 axmap_reset(f->io_axmap);
1870 else if (fio_file_lfsr(f))
1871 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1874 bool fio_files_done(struct thread_data *td)
1879 for_each_file(td, f, i)
1880 if (!fio_file_done(f))
1886 /* free memory used in initialization phase only */
1887 void filesetup_mem_free(void)
1889 free_already_allocated();
1893 * This function is for platforms which support direct I/O but not O_DIRECT.
1895 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1897 #ifdef FIO_OS_DIRECTIO
1898 int ret = fio_set_odirect(f);
1901 td_verror(td, ret, "fio_set_directio");
1902 #if defined(__sun__)
1903 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1904 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1906 log_err("fio: the file system does not seem to support direct IO\n");
1909 log_err("fio: the file system does not seem to support direct IO\n");
1916 log_err("fio: direct IO is not supported on this host operating system\n");