17 #include "lib/axmap.h"
20 #ifdef CONFIG_LINUX_FALLOCATE
21 #include <linux/falloc.h>
24 static FLIST_HEAD(filename_list);
27 * List entry for filename_list
30 struct flist_head list;
34 static inline void clear_error(struct thread_data *td)
40 static int native_fallocate(struct thread_data *td, struct fio_file *f)
44 success = fio_fallocate(f, 0, f->real_file_size);
45 dprint(FD_FILE, "native fallocate of file %s size %llu was "
46 "%ssuccessful\n", f->file_name,
47 (unsigned long long) f->real_file_size,
54 dprint(FD_FILE, "native fallocate is not implemented\n");
59 static void fallocate_file(struct thread_data *td, struct fio_file *f)
61 if (td->o.fill_device)
64 switch (td->o.fallocate_mode) {
65 case FIO_FALLOCATE_NATIVE:
66 native_fallocate(td, f);
68 case FIO_FALLOCATE_NONE:
70 #ifdef CONFIG_POSIX_FALLOCATE
71 case FIO_FALLOCATE_POSIX: {
74 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
76 (unsigned long long) f->real_file_size);
78 r = posix_fallocate(f->fd, 0, f->real_file_size);
80 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
83 #endif /* CONFIG_POSIX_FALLOCATE */
84 #ifdef CONFIG_LINUX_FALLOCATE
85 case FIO_FALLOCATE_KEEP_SIZE: {
88 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
89 "file %s size %llu\n", f->file_name,
90 (unsigned long long) f->real_file_size);
92 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
94 td_verror(td, errno, "fallocate");
98 #endif /* CONFIG_LINUX_FALLOCATE */
100 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
106 * Leaves f->fd open on success, caller must close
108 static int extend_file(struct thread_data *td, struct fio_file *f)
110 int new_layout = 0, unlink_file = 0, flags;
111 unsigned long long left;
116 log_err("fio: refusing extend of file due to read-only\n");
121 * check if we need to lay the file out complete again. fio
122 * does that for operations involving reads, or for writes
123 * where overwrite is set
126 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
127 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
129 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
132 if (unlink_file || new_layout) {
135 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
137 ret = td_io_unlink_file(td, f);
138 if (ret != 0 && ret != ENOENT) {
139 td_verror(td, errno, "unlink");
145 if (td->o.allow_create)
154 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
155 f->fd = open(f->file_name, flags, 0644);
159 if (err == ENOENT && !td->o.allow_create)
160 log_err("fio: file creation disallowed by "
161 "allow_file_create=0\n");
163 td_verror(td, err, "open");
167 fallocate_file(td, f);
170 * If our jobs don't require regular files initially, we're done.
176 * The size will be -1ULL when fill_device is used, so don't truncate
177 * or fallocate this file, just write it
179 if (!td->o.fill_device) {
180 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
181 (unsigned long long) f->real_file_size);
182 if (ftruncate(f->fd, f->real_file_size) == -1) {
183 if (errno != EFBIG) {
184 td_verror(td, errno, "ftruncate");
190 left = f->real_file_size;
191 bs = td->o.max_bs[DDIR_WRITE];
197 td_verror(td, errno, "malloc");
201 while (left && !td->terminate) {
207 fill_io_buffer(td, b, bs, bs);
209 r = write(f->fd, b, bs);
219 if (td->o.fill_device)
221 log_info("fio: ENOSPC on laying out "
225 td_verror(td, errno, "write");
227 td_verror(td, EIO, "write");
234 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
235 td_io_unlink_file(td, f);
236 } else if (td->o.create_fsync) {
237 if (fsync(f->fd) < 0) {
238 td_verror(td, errno, "fsync");
242 if (td->o.fill_device && !td_write(td)) {
243 fio_file_clear_size_known(f);
244 if (td_io_get_file_size(td, f))
246 if (f->io_size > f->real_file_size)
247 f->io_size = f->real_file_size;
261 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
263 int r, did_open = 0, old_runstate;
264 unsigned long long left;
269 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
270 td_ioengine_flagged(td, FIO_NOIO))
273 if (f->filetype == FIO_TYPE_CHAR)
276 if (!fio_file_open(f)) {
277 if (td->io_ops->open_file(td, f)) {
278 log_err("fio: cannot pre-read, failed to open file\n");
284 old_runstate = td_bump_runstate(td, TD_PRE_READING);
287 bs = td->o.max_bs[DDIR_READ];
293 td_verror(td, errno, "malloc");
299 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
300 td_verror(td, errno, "lseek");
301 log_err("fio: failed to lseek pre-read file\n");
306 while (left && !td->terminate) {
310 r = read(f->fd, b, bs);
316 td_verror(td, EIO, "pre_read");
322 td_restore_runstate(td, old_runstate);
325 td->io_ops->close_file(td, f);
331 unsigned long long get_rand_file_size(struct thread_data *td)
333 unsigned long long ret, sized;
337 frand_max = rand_max(&td->file_size_state);
338 r = __rand(&td->file_size_state);
339 sized = td->o.file_size_high - td->o.file_size_low;
340 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
341 ret += td->o.file_size_low;
342 ret -= (ret % td->o.rw_min_bs);
346 static int file_size(struct thread_data *td, struct fio_file *f)
350 if (stat(f->file_name, &st) == -1) {
351 td_verror(td, errno, "fstat");
355 f->real_file_size = st.st_size;
359 static int bdev_size(struct thread_data *td, struct fio_file *f)
361 unsigned long long bytes = 0;
364 if (td->io_ops->open_file(td, f)) {
365 log_err("fio: failed opening blockdev %s for size check\n",
370 r = blockdev_size(f, &bytes);
372 td_verror(td, r, "blockdev_size");
377 log_err("%s: zero sized block device?\n", f->file_name);
381 f->real_file_size = bytes;
382 td->io_ops->close_file(td, f);
385 td->io_ops->close_file(td, f);
389 static int char_size(struct thread_data *td, struct fio_file *f)
391 #ifdef FIO_HAVE_CHARDEV_SIZE
392 unsigned long long bytes = 0;
395 if (td->io_ops->open_file(td, f)) {
396 log_err("fio: failed opening chardev %s for size check\n",
401 r = chardev_size(f, &bytes);
403 td_verror(td, r, "chardev_size");
408 log_err("%s: zero sized char device?\n", f->file_name);
412 f->real_file_size = bytes;
413 td->io_ops->close_file(td, f);
416 td->io_ops->close_file(td, f);
419 f->real_file_size = -1ULL;
424 static int get_file_size(struct thread_data *td, struct fio_file *f)
428 if (fio_file_size_known(f))
431 if (f->filetype == FIO_TYPE_FILE)
432 ret = file_size(td, f);
433 else if (f->filetype == FIO_TYPE_BLOCK)
434 ret = bdev_size(td, f);
435 else if (f->filetype == FIO_TYPE_CHAR)
436 ret = char_size(td, f);
438 f->real_file_size = -1;
439 log_info("%s: failed to get file size of %s\n", td->o.name,
441 return 1; /* avoid offset extends end error message */
445 * Leave ->real_file_size with 0 since it could be expectation
446 * of initial setup for regular files.
452 * ->file_offset normally hasn't been initialized yet, so this
453 * is basically always false unless ->real_file_size is -1, but
454 * if ->real_file_size is -1 this message doesn't make sense.
455 * As a result, this message is basically useless.
457 if (f->file_offset > f->real_file_size) {
458 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
459 (unsigned long long) f->file_offset,
460 (unsigned long long) f->real_file_size);
464 fio_file_set_size_known(f);
468 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
469 unsigned long long off,
470 unsigned long long len)
472 int errval = 0, ret = 0;
481 off = f->file_offset;
483 if (len == -1ULL || off == -1ULL)
486 if (td->io_ops->invalidate) {
487 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
489 ret = td->io_ops->invalidate(td, f);
492 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
493 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
496 } else if (f->filetype == FIO_TYPE_FILE) {
497 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
498 f->file_name, off, len);
499 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
502 } else if (f->filetype == FIO_TYPE_BLOCK) {
505 dprint(FD_IO, "drop page cache %s\n", f->file_name);
506 ret = blockdev_invalidate_cache(f);
507 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
509 * Linux multipath devices reject ioctl while
510 * the maps are being updated. That window can
511 * last tens of milliseconds; we'll try up to
512 * a quarter of a second.
515 ret = blockdev_invalidate_cache(f);
517 if (ret < 0 && errno == EACCES && geteuid()) {
518 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
519 log_err("fio: only root may flush block "
520 "devices. Cache flush bypassed!\n");
526 } else if (f->filetype == FIO_TYPE_CHAR ||
527 f->filetype == FIO_TYPE_PIPE) {
528 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
533 * Cache flushing isn't a fatal condition, and we know it will
534 * happen on some platforms where we don't have the proper
535 * function to flush eg block device caches. So just warn and
536 * continue on our way.
539 log_info("fio: cache invalidation of %s failed: %s\n",
540 f->file_name, strerror(errval));
546 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
548 if (!fio_file_open(f))
551 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
554 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
558 dprint(FD_FILE, "fd close %s\n", f->file_name);
562 if (close(f->fd) < 0)
567 if (f->shadow_fd != -1) {
576 int file_lookup_open(struct fio_file *f, int flags)
578 struct fio_file *__f;
581 __f = lookup_file_hash(f->file_name);
583 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
587 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
595 f->fd = open(f->file_name, flags, 0600);
599 static int file_close_shadow_fds(struct thread_data *td)
605 for_each_file(td, f, i) {
606 if (f->shadow_fd == -1)
617 int generic_open_file(struct thread_data *td, struct fio_file *f)
623 dprint(FD_FILE, "fd open %s\n", f->file_name);
625 if (!strcmp(f->file_name, "-")) {
627 log_err("fio: can't read/write to stdin/out\n");
633 * move output logging to stderr, if we are writing to stdout
642 flags |= OS_O_DIRECT;
645 td_verror(td, EINVAL, "OS does not support atomic IO");
648 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
652 if (td->o.create_on_open && td->o.allow_create)
655 if (f->filetype != FIO_TYPE_FILE)
656 flags |= FIO_O_NOATIME;
663 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
667 f->fd = dup(STDOUT_FILENO);
669 from_hash = file_lookup_open(f, flags);
670 } else if (td_read(td)) {
671 if (f->filetype == FIO_TYPE_CHAR && !read_only)
677 f->fd = dup(STDIN_FILENO);
679 from_hash = file_lookup_open(f, flags);
680 } else if (td_trim(td)) {
681 assert(!td_rw(td)); /* should have matched above */
683 from_hash = file_lookup_open(f, flags);
687 char buf[FIO_VERROR_SIZE];
690 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
691 flags &= ~FIO_O_NOATIME;
694 if (__e == EMFILE && file_close_shadow_fds(td))
697 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
699 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
700 log_err("fio: looks like your file system does not " \
701 "support direct=1/buffered=0\n");
704 td_verror(td, __e, buf);
708 if (!from_hash && f->fd != -1) {
709 if (add_file_hash(f)) {
713 * Stash away descriptor for later close. This is to
714 * work-around a "feature" on Linux, where a close of
715 * an fd that has been opened for write will trigger
716 * udev to call blkid to check partitions, fs id, etc.
717 * That pollutes the device cache, which can slow down
718 * unbuffered accesses.
720 if (f->shadow_fd == -1)
721 f->shadow_fd = f->fd;
724 * OK to ignore, we haven't done anything
727 ret = generic_close_file(td, f);
737 * This function i.e. get_file_size() is the default .get_file_size
738 * implementation of majority of I/O engines.
740 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
742 return get_file_size(td, f);
746 * open/close all files, so that ->real_file_size gets set
748 static int get_file_sizes(struct thread_data *td)
754 for_each_file(td, f, i) {
755 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
758 if (td_io_get_file_size(td, f)) {
759 if (td->error != ENOENT) {
760 log_err("%s\n", td->verror);
768 * There are corner cases where we end up with -1 for
769 * ->real_file_size due to unsupported file type, etc.
770 * We then just set to size option value divided by number
771 * of files, similar to the way file ->io_size is set.
772 * stat(2) failure doesn't set ->real_file_size to -1.
774 if (f->real_file_size == -1ULL && td->o.size)
775 f->real_file_size = td->o.size / td->o.nr_files;
782 struct flist_head list;
789 * Get free number of bytes for each file on each unique mount.
791 static unsigned long long get_fs_free_counts(struct thread_data *td)
793 struct flist_head *n, *tmp;
794 unsigned long long ret = 0;
795 struct fio_mount *fm;
800 for_each_file(td, f, i) {
804 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
805 if (f->real_file_size != -1ULL)
806 ret += f->real_file_size;
808 } else if (f->filetype != FIO_TYPE_FILE)
812 strncpy(buf, f->file_name, 255);
814 if (stat(buf, &sb) < 0) {
818 if (stat(buf, &sb) < 0)
823 flist_for_each(n, &list) {
824 fm = flist_entry(n, struct fio_mount, list);
825 if (fm->key == sb.st_dev)
834 fm = calloc(1, sizeof(*fm));
835 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
836 fm->base = basename(fm->__base);
838 flist_add(&fm->list, &list);
841 flist_for_each_safe(n, tmp, &list) {
842 unsigned long long sz;
844 fm = flist_entry(n, struct fio_mount, list);
845 flist_del(&fm->list);
847 sz = get_fs_free_size(fm->base);
848 if (sz && sz != -1ULL)
857 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
859 struct thread_options *o = &td->o;
860 unsigned long long align_bs;
861 unsigned long long offset;
863 if (o->file_append && f->filetype == FIO_TYPE_FILE)
864 return f->real_file_size;
866 if (o->start_offset_percent > 0) {
868 * if offset_align is provided, set initial offset
870 if (fio_option_is_set(o, start_offset_align)) {
871 align_bs = o->start_offset_align;
873 /* else take the minimum block size */
874 align_bs = td_min_bs(td);
877 /* calculate the raw offset */
878 offset = (f->real_file_size * o->start_offset_percent / 100) +
879 (td->subjob_number * o->offset_increment);
882 * block align the offset at the next available boundary at
883 * ceiling(offset / align_bs) * align_bs
885 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
888 /* start_offset_percent not set */
889 offset = o->start_offset +
890 td->subjob_number * o->offset_increment;
897 * Open the files and setup files sizes, creating files if necessary.
899 int setup_files(struct thread_data *td)
901 unsigned long long total_size, extend_size;
902 struct thread_options *o = &td->o;
904 unsigned int i, nr_fs_extra = 0;
905 int err = 0, need_extend;
907 const unsigned int bs = td_min_bs(td);
910 dprint(FD_FILE, "setup files\n");
912 old_state = td_bump_runstate(td, TD_SETTING_UP);
914 if (o->read_iolog_file)
918 * Find out physical size of files or devices for this thread,
919 * before we determine I/O size and range of our targets.
920 * If ioengine defines a setup() method, it's responsible for
921 * opening the files and setting f->real_file_size to indicate
922 * the valid range for that file.
924 if (td->io_ops->setup)
925 err = td->io_ops->setup(td);
927 err = get_file_sizes(td);
933 * check sizes. if the files/devices do not exist and the size
934 * isn't passed to fio, abort.
937 for_each_file(td, f, i) {
939 if (f->real_file_size == -1ULL)
942 total_size += f->real_file_size;
946 td->fill_device_size = get_fs_free_counts(td);
949 * device/file sizes are zero and no size given, punt
951 if ((!total_size || total_size == -1ULL) && !o->size &&
952 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
953 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
954 log_err("%s: you need to specify size=\n", o->name);
955 td_verror(td, EINVAL, "total_file_size");
960 * Calculate per-file size and potential extra size for the
961 * first files, if needed (i.e. if we don't have a fixed size).
963 if (!o->file_size_low && o->nr_files) {
966 fs = o->size / o->nr_files;
967 all_fs = fs * o->nr_files;
969 if (all_fs < o->size)
970 nr_fs_extra = (o->size - all_fs) / bs;
974 * now file sizes are known, so we can set ->io_size. if size= is
975 * not given, ->io_size is just equal to ->real_file_size. if size
976 * is given, ->io_size is size / nr_files.
978 extend_size = total_size = 0;
980 for_each_file(td, f, i) {
981 f->file_offset = get_start_offset(td, f);
984 * Update ->io_size depending on options specified.
985 * ->file_size_low being 0 means filesize option isn't set.
986 * Non zero ->file_size_low equals ->file_size_high means
987 * filesize option is set in a fixed size format.
988 * Non zero ->file_size_low not equals ->file_size_high means
989 * filesize option is set in a range format.
991 if (!o->file_size_low) {
993 * no file size or range given, file size is equal to
994 * total size divided by number of files. If the size
995 * doesn't divide nicely with the min blocksize,
996 * make the first files bigger.
1005 * We normally don't come here for regular files, but
1006 * if the result is 0 for a regular file, set it to the
1007 * real file size. This could be size of the existing
1008 * one if it already exists, but otherwise will be set
1009 * to 0. A new file won't be created because
1010 * ->io_size + ->file_offset equals ->real_file_size.
1013 if (f->file_offset > f->real_file_size)
1015 f->io_size = f->real_file_size - f->file_offset;
1017 log_info("fio: file %s may be ignored\n",
1020 } else if (f->real_file_size < o->file_size_low ||
1021 f->real_file_size > o->file_size_high) {
1022 if (f->file_offset > o->file_size_low)
1025 * file size given. if it's fixed, use that. if it's a
1026 * range, generate a random size in-between.
1028 if (o->file_size_low == o->file_size_high)
1029 f->io_size = o->file_size_low - f->file_offset;
1031 f->io_size = get_rand_file_size(td)
1035 f->io_size = f->real_file_size - f->file_offset;
1037 if (f->io_size == -1ULL)
1040 if (o->size_percent && o->size_percent != 100) {
1043 file_size = f->io_size + f->file_offset;
1044 f->io_size = (file_size *
1045 o->size_percent) / 100;
1046 if (f->io_size > (file_size - f->file_offset))
1047 f->io_size = file_size - f->file_offset;
1049 f->io_size -= (f->io_size % td_min_bs(td));
1051 total_size += f->io_size;
1054 if (f->filetype == FIO_TYPE_FILE &&
1055 (f->io_size + f->file_offset) > f->real_file_size &&
1056 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1057 if (!o->create_on_open) {
1059 extend_size += (f->io_size + f->file_offset);
1060 fio_file_set_extend(f);
1062 f->real_file_size = f->io_size + f->file_offset;
1066 if (td->o.block_error_hist) {
1069 assert(td->o.nr_files == 1); /* checked in fixup_options */
1071 len = f->io_size / td->o.bs[DDIR_TRIM];
1072 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1073 log_err("fio: cannot calculate block histogram with "
1074 "%d trim blocks, maximum %d\n",
1075 len, MAX_NR_BLOCK_INFOS);
1076 td_verror(td, EINVAL, "block_error_hist");
1080 td->ts.nr_block_infos = len;
1081 for (i = 0; i < len; i++)
1082 td->ts.block_infos[i] =
1083 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1085 td->ts.nr_block_infos = 0;
1087 if (!o->size || (total_size && o->size > total_size))
1088 o->size = total_size;
1090 if (o->size < td_min_bs(td)) {
1091 log_err("fio: blocksize too large for data set\n");
1096 * See if we need to extend some files, typically needed when our
1097 * target regular files don't exist yet, but our jobs require them
1098 * initially due to read I/Os.
1102 if (output_format & FIO_OUTPUT_NORMAL) {
1103 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1105 need_extend > 1 ? "s" : "",
1107 need_extend > 1 ? "s" : "",
1108 need_extend > 1 ? "total " : "",
1112 for_each_file(td, f, i) {
1113 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1115 if (!fio_file_extend(f))
1118 assert(f->filetype == FIO_TYPE_FILE);
1119 fio_file_clear_extend(f);
1120 if (!o->fill_device) {
1121 old_len = f->real_file_size;
1122 extend_len = f->io_size + f->file_offset -
1125 f->real_file_size = (f->io_size + f->file_offset);
1126 err = extend_file(td, f);
1130 err = __file_invalidate_cache(td, f, old_len,
1134 * Shut up static checker
1150 o->zone_size = o->size;
1153 * iolog already set the total io size, if we read back
1156 if (!o->read_iolog_file) {
1158 td->total_io_size = o->io_size * o->loops;
1160 td->total_io_size = o->size * o->loops;
1167 td_restore_runstate(td, old_state);
1170 log_err("%s: you need to specify valid offset=\n", o->name);
1172 td_restore_runstate(td, old_state);
1176 bool pre_read_files(struct thread_data *td)
1181 dprint(FD_FILE, "pre_read files\n");
1183 for_each_file(td, f, i) {
1184 if (!pre_read_file(td, f))
1191 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1193 unsigned int range_size, seed;
1194 unsigned long nranges;
1197 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1198 fsize = min(f->real_file_size, f->io_size);
1200 nranges = (fsize + range_size - 1) / range_size;
1202 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1203 if (!td->o.rand_repeatable)
1204 seed = td->rand_seeds[4];
1206 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1207 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1208 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1209 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1210 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1211 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1214 static bool init_rand_distribution(struct thread_data *td)
1220 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1223 state = td_bump_runstate(td, TD_SETTING_UP);
1225 for_each_file(td, f, i)
1226 __init_rand_distribution(td, f);
1228 td_restore_runstate(td, state);
1233 * Check if the number of blocks exceeds the randomness capability of
1234 * the selected generator. Tausworthe is 32-bit, the others are fullly
1237 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1240 if (blocks <= FRAND32_MAX)
1242 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1246 * If the user hasn't specified a random generator, switch
1247 * to tausworthe64 with informational warning. If the user did
1248 * specify one, just warn.
1250 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1253 if (!fio_option_is_set(&td->o, random_generator)) {
1254 log_info("fio: Switching to tausworthe64. Use the "
1255 "random_generator= option to get rid of this "
1257 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1262 * Just make this information to avoid breaking scripts.
1264 log_info("fio: Use the random_generator= option to switch to lfsr or "
1269 bool init_random_map(struct thread_data *td)
1271 unsigned long long blocks;
1275 if (init_rand_distribution(td))
1280 for_each_file(td, f, i) {
1281 uint64_t fsize = min(f->real_file_size, f->io_size);
1283 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1285 if (check_rand_gen_limits(td, f, blocks))
1288 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1291 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1293 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1294 fio_file_set_lfsr(f);
1297 } else if (!td->o.norandommap) {
1298 f->io_axmap = axmap_new(blocks);
1300 fio_file_set_axmap(f);
1303 } else if (td->o.norandommap)
1306 if (!td->o.softrandommap) {
1307 log_err("fio: failed allocating random map. If running"
1308 " a large number of jobs, try the 'norandommap'"
1309 " option or set 'softrandommap'. Or give"
1310 " a larger --alloc-size to fio.\n");
1314 log_info("fio: file %s failed allocating random map. Running "
1315 "job without.\n", f->file_name);
1321 void close_files(struct thread_data *td)
1326 for_each_file(td, f, i) {
1327 if (fio_file_open(f))
1328 td_io_close_file(td, f);
1332 void close_and_free_files(struct thread_data *td)
1336 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1338 dprint(FD_FILE, "close files\n");
1340 for_each_file(td, f, i) {
1341 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1342 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1343 td_io_unlink_file(td, f);
1346 if (fio_file_open(f))
1347 td_io_close_file(td, f);
1349 remove_file_hash(f);
1351 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1352 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1353 td_io_unlink_file(td, f);
1359 sfree(f->file_name);
1360 f->file_name = NULL;
1361 if (fio_file_axmap(f)) {
1362 axmap_free(f->io_axmap);
1371 td->o.filename = NULL;
1373 free(td->file_locks);
1374 td->files_index = 0;
1376 td->file_locks = NULL;
1377 td->o.file_lock_mode = FILE_LOCK_NONE;
1381 static void get_file_type(struct fio_file *f)
1385 if (!strcmp(f->file_name, "-"))
1386 f->filetype = FIO_TYPE_PIPE;
1388 f->filetype = FIO_TYPE_FILE;
1391 /* \\.\ is the device namespace in Windows, where every file is
1393 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1394 f->filetype = FIO_TYPE_BLOCK;
1397 if (!stat(f->file_name, &sb)) {
1398 if (S_ISBLK(sb.st_mode))
1399 f->filetype = FIO_TYPE_BLOCK;
1400 else if (S_ISCHR(sb.st_mode))
1401 f->filetype = FIO_TYPE_CHAR;
1402 else if (S_ISFIFO(sb.st_mode))
1403 f->filetype = FIO_TYPE_PIPE;
1407 static bool __is_already_allocated(const char *fname, bool set)
1409 struct flist_head *entry;
1412 ret = file_bloom_exists(fname, set);
1416 flist_for_each(entry, &filename_list) {
1417 struct file_name *fn;
1419 fn = flist_entry(entry, struct file_name, list);
1421 if (!strcmp(fn->filename, fname))
1428 static bool is_already_allocated(const char *fname)
1432 fio_file_hash_lock();
1433 ret = __is_already_allocated(fname, false);
1434 fio_file_hash_unlock();
1439 static void set_already_allocated(const char *fname)
1441 struct file_name *fn;
1443 fn = malloc(sizeof(struct file_name));
1444 fn->filename = strdup(fname);
1446 fio_file_hash_lock();
1447 if (!__is_already_allocated(fname, true)) {
1448 flist_add_tail(&fn->list, &filename_list);
1451 fio_file_hash_unlock();
1459 static void free_already_allocated(void)
1461 struct flist_head *entry, *tmp;
1462 struct file_name *fn;
1464 if (flist_empty(&filename_list))
1467 fio_file_hash_lock();
1468 flist_for_each_safe(entry, tmp, &filename_list) {
1469 fn = flist_entry(entry, struct file_name, list);
1471 flist_del(&fn->list);
1475 fio_file_hash_unlock();
1478 static struct fio_file *alloc_new_file(struct thread_data *td)
1482 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1483 f = calloc(1, sizeof(*f));
1485 f = scalloc(1, sizeof(*f));
1493 fio_file_reset(td, f);
1497 bool exists_and_not_regfile(const char *filename)
1501 if (lstat(filename, &sb) == -1)
1504 #ifndef WIN32 /* NOT Windows */
1505 if (S_ISREG(sb.st_mode))
1508 /* \\.\ is the device namespace in Windows, where every file
1509 * is a device node */
1510 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1517 static bool create_work_dirs(struct thread_data *td, const char *fname)
1519 char path[PATH_MAX];
1522 if (td->o.directory) {
1523 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1524 FIO_OS_PATH_SEPARATOR, fname);
1525 start = strstr(path, fname);
1527 snprintf(path, PATH_MAX, "%s", fname);
1532 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1537 #ifdef CONFIG_HAVE_MKDIR_TWO
1538 if (mkdir(path, 0600) && errno != EEXIST) {
1540 if (mkdir(path) && errno != EEXIST) {
1542 log_err("fio: failed to create dir (%s): %d\n",
1546 *end = FIO_OS_PATH_SEPARATOR;
1549 td->flags |= TD_F_DIRS_CREATED;
1553 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1555 int cur_files = td->files_index;
1556 char file_name[PATH_MAX];
1560 dprint(FD_FILE, "add file %s\n", fname);
1562 if (td->o.directory)
1563 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1564 td->o.unique_filename);
1566 sprintf(file_name + len, "%s", fname);
1568 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1569 !(td->flags & TD_F_DIRS_CREATED) &&
1570 !create_work_dirs(td, fname))
1573 /* clean cloned siblings using existing files */
1574 if (numjob && is_already_allocated(file_name) &&
1575 !exists_and_not_regfile(fname))
1578 f = alloc_new_file(td);
1580 if (td->files_size <= td->files_index) {
1581 unsigned int new_size = td->o.nr_files + 1;
1583 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1585 td->files = realloc(td->files, new_size * sizeof(f));
1586 if (td->files == NULL) {
1587 log_err("fio: realloc OOM\n");
1590 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1591 td->file_locks = realloc(td->file_locks, new_size);
1592 if (!td->file_locks) {
1593 log_err("fio: realloc OOM\n");
1596 td->file_locks[cur_files] = FILE_LOCK_NONE;
1598 td->files_size = new_size;
1600 td->files[cur_files] = f;
1601 f->fileno = cur_files;
1604 * init function, io engine may not be loaded yet
1606 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1607 f->real_file_size = -1ULL;
1609 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1610 f->file_name = strdup(file_name);
1612 f->file_name = smalloc_strdup(file_name);
1618 switch (td->o.file_lock_mode) {
1619 case FILE_LOCK_NONE:
1621 case FILE_LOCK_READWRITE:
1622 f->rwlock = fio_rwlock_init();
1624 case FILE_LOCK_EXCLUSIVE:
1625 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1628 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1634 if (td->o.numjobs > 1)
1635 set_already_allocated(file_name);
1640 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1646 int add_file_exclusive(struct thread_data *td, const char *fname)
1651 for_each_file(td, f, i) {
1652 if (!strcmp(f->file_name, fname))
1656 return add_file(td, fname, 0, 1);
1659 void get_file(struct fio_file *f)
1661 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1662 assert(fio_file_open(f));
1666 int put_file(struct thread_data *td, struct fio_file *f)
1668 int f_ret = 0, ret = 0;
1670 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1672 if (!fio_file_open(f)) {
1673 assert(f->fd == -1);
1677 assert(f->references);
1678 if (--f->references)
1681 if (should_fsync(td) && td->o.fsync_on_close) {
1682 f_ret = fsync(f->fd);
1687 if (td->io_ops->close_file)
1688 ret = td->io_ops->close_file(td, f);
1693 td->nr_open_files--;
1694 fio_file_clear_open(f);
1695 assert(f->fd == -1);
1699 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1701 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1704 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1705 if (ddir == DDIR_READ)
1706 fio_rwlock_read(f->rwlock);
1708 fio_rwlock_write(f->rwlock);
1709 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1710 fio_sem_down(f->lock);
1712 td->file_locks[f->fileno] = td->o.file_lock_mode;
1715 void unlock_file(struct thread_data *td, struct fio_file *f)
1717 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1720 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1721 fio_rwlock_unlock(f->rwlock);
1722 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1723 fio_sem_up(f->lock);
1725 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1728 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1730 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1732 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1736 static bool recurse_dir(struct thread_data *td, const char *dirname)
1742 D = opendir(dirname);
1744 char buf[FIO_VERROR_SIZE];
1746 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1747 td_verror(td, errno, buf);
1751 while ((dir = readdir(D)) != NULL) {
1752 char full_path[PATH_MAX];
1755 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1758 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1760 if (lstat(full_path, &sb) == -1) {
1761 if (errno != ENOENT) {
1762 td_verror(td, errno, "stat");
1768 if (S_ISREG(sb.st_mode)) {
1769 add_file(td, full_path, 0, 1);
1772 if (!S_ISDIR(sb.st_mode))
1775 ret = recurse_dir(td, full_path);
1784 int add_dir_files(struct thread_data *td, const char *path)
1786 int ret = recurse_dir(td, path);
1789 log_info("fio: opendir added %d files\n", td->o.nr_files);
1794 void dup_files(struct thread_data *td, struct thread_data *org)
1799 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1804 td->files = malloc(org->files_index * sizeof(f));
1806 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1807 td->file_locks = malloc(org->files_index);
1809 for_each_file(org, f, i) {
1810 struct fio_file *__f;
1812 __f = alloc_new_file(td);
1815 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1816 __f->file_name = strdup(f->file_name);
1818 __f->file_name = smalloc_strdup(f->file_name);
1819 if (!__f->file_name)
1822 __f->filetype = f->filetype;
1825 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1826 __f->lock = f->lock;
1827 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1828 __f->rwlock = f->rwlock;
1835 * Returns the index that matches the filename, or -1 if not there
1837 int get_fileno(struct thread_data *td, const char *fname)
1842 for_each_file(td, f, i)
1843 if (!strcmp(f->file_name, fname))
1850 * For log usage, where we add/open/close files automatically
1852 void free_release_files(struct thread_data *td)
1856 td->o.open_files = 0;
1857 td->files_index = 0;
1860 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1864 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1865 f->last_pos[i] = f->file_offset;
1866 f->last_start[i] = -1ULL;
1869 if (fio_file_axmap(f))
1870 axmap_reset(f->io_axmap);
1871 else if (fio_file_lfsr(f))
1872 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1875 bool fio_files_done(struct thread_data *td)
1880 for_each_file(td, f, i)
1881 if (!fio_file_done(f))
1887 /* free memory used in initialization phase only */
1888 void filesetup_mem_free(void)
1890 free_already_allocated();
1894 * This function is for platforms which support direct I/O but not O_DIRECT.
1896 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1898 #ifdef FIO_OS_DIRECTIO
1899 int ret = fio_set_odirect(f);
1902 td_verror(td, ret, "fio_set_directio");
1903 #if defined(__sun__)
1904 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1905 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1907 log_err("fio: the file system does not seem to support direct IO\n");
1910 log_err("fio: the file system does not seem to support direct IO\n");
1917 log_err("fio: direct IO is not supported on this host operating system\n");