15 #include "lib/axmap.h"
18 #ifdef CONFIG_LINUX_FALLOCATE
19 #include <linux/falloc.h>
22 static FLIST_HEAD(filename_list);
25 * List entry for filename_list
28 struct flist_head list;
32 static inline void clear_error(struct thread_data *td)
38 static int native_fallocate(struct thread_data *td, struct fio_file *f)
42 success = fio_fallocate(f, 0, f->real_file_size);
43 dprint(FD_FILE, "native fallocate of file %s size %llu was "
44 "%ssuccessful\n", f->file_name,
45 (unsigned long long) f->real_file_size,
52 dprint(FD_FILE, "native fallocate is not implemented\n");
57 static void fallocate_file(struct thread_data *td, struct fio_file *f)
59 if (td->o.fill_device)
62 switch (td->o.fallocate_mode) {
63 case FIO_FALLOCATE_NATIVE:
64 native_fallocate(td, f);
66 case FIO_FALLOCATE_NONE:
68 #ifdef CONFIG_POSIX_FALLOCATE
69 case FIO_FALLOCATE_POSIX: {
72 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
74 (unsigned long long) f->real_file_size);
76 r = posix_fallocate(f->fd, 0, f->real_file_size);
78 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
81 #endif /* CONFIG_POSIX_FALLOCATE */
82 #ifdef CONFIG_LINUX_FALLOCATE
83 case FIO_FALLOCATE_KEEP_SIZE: {
86 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
87 "file %s size %llu\n", f->file_name,
88 (unsigned long long) f->real_file_size);
90 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
92 td_verror(td, errno, "fallocate");
96 #endif /* CONFIG_LINUX_FALLOCATE */
98 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
104 * Leaves f->fd open on success, caller must close
106 static int extend_file(struct thread_data *td, struct fio_file *f)
108 int new_layout = 0, unlink_file = 0, flags;
109 unsigned long long left;
114 log_err("fio: refusing extend of file due to read-only\n");
119 * check if we need to lay the file out complete again. fio
120 * does that for operations involving reads, or for writes
121 * where overwrite is set
124 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
125 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
127 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
130 if (unlink_file || new_layout) {
133 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
135 ret = td_io_unlink_file(td, f);
136 if (ret != 0 && ret != ENOENT) {
137 td_verror(td, errno, "unlink");
143 if (td->o.allow_create)
152 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
153 f->fd = open(f->file_name, flags, 0644);
157 if (err == ENOENT && !td->o.allow_create)
158 log_err("fio: file creation disallowed by "
159 "allow_file_create=0\n");
161 td_verror(td, err, "open");
165 fallocate_file(td, f);
168 * If our jobs don't require regular files initially, we're done.
174 * The size will be -1ULL when fill_device is used, so don't truncate
175 * or fallocate this file, just write it
177 if (!td->o.fill_device) {
178 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
179 (unsigned long long) f->real_file_size);
180 if (ftruncate(f->fd, f->real_file_size) == -1) {
181 if (errno != EFBIG) {
182 td_verror(td, errno, "ftruncate");
188 left = f->real_file_size;
189 bs = td->o.max_bs[DDIR_WRITE];
195 td_verror(td, errno, "malloc");
199 while (left && !td->terminate) {
205 fill_io_buffer(td, b, bs, bs);
207 r = write(f->fd, b, bs);
217 if (td->o.fill_device)
219 log_info("fio: ENOSPC on laying out "
223 td_verror(td, errno, "write");
225 td_verror(td, EIO, "write");
232 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
233 td_io_unlink_file(td, f);
234 } else if (td->o.create_fsync) {
235 if (fsync(f->fd) < 0) {
236 td_verror(td, errno, "fsync");
240 if (td->o.fill_device && !td_write(td)) {
241 fio_file_clear_size_known(f);
242 if (td_io_get_file_size(td, f))
244 if (f->io_size > f->real_file_size)
245 f->io_size = f->real_file_size;
259 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
261 int r, did_open = 0, old_runstate;
262 unsigned long long left;
267 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
268 td_ioengine_flagged(td, FIO_NOIO))
271 if (f->filetype == FIO_TYPE_CHAR)
274 if (!fio_file_open(f)) {
275 if (td->io_ops->open_file(td, f)) {
276 log_err("fio: cannot pre-read, failed to open file\n");
282 old_runstate = td_bump_runstate(td, TD_PRE_READING);
285 bs = td->o.max_bs[DDIR_READ];
291 td_verror(td, errno, "malloc");
297 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
298 td_verror(td, errno, "lseek");
299 log_err("fio: failed to lseek pre-read file\n");
304 while (left && !td->terminate) {
308 r = read(f->fd, b, bs);
314 td_verror(td, EIO, "pre_read");
320 td_restore_runstate(td, old_runstate);
323 td->io_ops->close_file(td, f);
329 unsigned long long get_rand_file_size(struct thread_data *td)
331 unsigned long long ret, sized;
335 frand_max = rand_max(&td->file_size_state);
336 r = __rand(&td->file_size_state);
337 sized = td->o.file_size_high - td->o.file_size_low;
338 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
339 ret += td->o.file_size_low;
340 ret -= (ret % td->o.rw_min_bs);
344 static int file_size(struct thread_data *td, struct fio_file *f)
348 if (stat(f->file_name, &st) == -1) {
349 td_verror(td, errno, "fstat");
353 f->real_file_size = st.st_size;
357 static int bdev_size(struct thread_data *td, struct fio_file *f)
359 unsigned long long bytes = 0;
362 if (td->io_ops->open_file(td, f)) {
363 log_err("fio: failed opening blockdev %s for size check\n",
368 r = blockdev_size(f, &bytes);
370 td_verror(td, r, "blockdev_size");
375 log_err("%s: zero sized block device?\n", f->file_name);
379 f->real_file_size = bytes;
380 td->io_ops->close_file(td, f);
383 td->io_ops->close_file(td, f);
387 static int char_size(struct thread_data *td, struct fio_file *f)
389 #ifdef FIO_HAVE_CHARDEV_SIZE
390 unsigned long long bytes = 0;
393 if (td->io_ops->open_file(td, f)) {
394 log_err("fio: failed opening chardev %s for size check\n",
399 r = chardev_size(f, &bytes);
401 td_verror(td, r, "chardev_size");
406 log_err("%s: zero sized char device?\n", f->file_name);
410 f->real_file_size = bytes;
411 td->io_ops->close_file(td, f);
414 td->io_ops->close_file(td, f);
417 f->real_file_size = -1ULL;
422 static int get_file_size(struct thread_data *td, struct fio_file *f)
426 if (fio_file_size_known(f))
429 if (f->filetype == FIO_TYPE_FILE)
430 ret = file_size(td, f);
431 else if (f->filetype == FIO_TYPE_BLOCK)
432 ret = bdev_size(td, f);
433 else if (f->filetype == FIO_TYPE_CHAR)
434 ret = char_size(td, f);
436 f->real_file_size = -1;
437 log_info("%s: failed to get file size of %s\n", td->o.name,
439 return 1; /* avoid offset extends end error message */
443 * Leave ->real_file_size with 0 since it could be expectation
444 * of initial setup for regular files.
450 * ->file_offset normally hasn't been initialized yet, so this
451 * is basically always false unless ->real_file_size is -1, but
452 * if ->real_file_size is -1 this message doesn't make sense.
453 * As a result, this message is basically useless.
455 if (f->file_offset > f->real_file_size) {
456 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
457 (unsigned long long) f->file_offset,
458 (unsigned long long) f->real_file_size);
462 fio_file_set_size_known(f);
466 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
467 unsigned long long off,
468 unsigned long long len)
470 int errval = 0, ret = 0;
479 off = f->file_offset;
481 if (len == -1ULL || off == -1ULL)
484 if (td->io_ops->invalidate) {
485 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
487 ret = td->io_ops->invalidate(td, f);
490 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
491 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
493 } else if (f->filetype == FIO_TYPE_FILE) {
494 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
495 f->file_name, off, len);
496 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
499 } else if (f->filetype == FIO_TYPE_BLOCK) {
502 dprint(FD_IO, "drop page cache %s\n", f->file_name);
503 ret = blockdev_invalidate_cache(f);
504 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
506 * Linux multipath devices reject ioctl while
507 * the maps are being updated. That window can
508 * last tens of milliseconds; we'll try up to
509 * a quarter of a second.
512 ret = blockdev_invalidate_cache(f);
514 if (ret < 0 && errno == EACCES && geteuid()) {
515 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
516 log_err("fio: only root may flush block "
517 "devices. Cache flush bypassed!\n");
522 } else if (f->filetype == FIO_TYPE_CHAR ||
523 f->filetype == FIO_TYPE_PIPE) {
524 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
528 * Cache flushing isn't a fatal condition, and we know it will
529 * happen on some platforms where we don't have the proper
530 * function to flush eg block device caches. So just warn and
531 * continue on our way.
534 log_info("fio: cache invalidation of %s failed: %s\n",
535 f->file_name, strerror(errval));
541 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
543 if (!fio_file_open(f))
546 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
549 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
553 dprint(FD_FILE, "fd close %s\n", f->file_name);
557 if (close(f->fd) < 0)
562 if (f->shadow_fd != -1) {
571 int file_lookup_open(struct fio_file *f, int flags)
573 struct fio_file *__f;
576 __f = lookup_file_hash(f->file_name);
578 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
582 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
590 f->fd = open(f->file_name, flags, 0600);
594 static int file_close_shadow_fds(struct thread_data *td)
600 for_each_file(td, f, i) {
601 if (f->shadow_fd == -1)
612 int generic_open_file(struct thread_data *td, struct fio_file *f)
618 dprint(FD_FILE, "fd open %s\n", f->file_name);
620 if (!strcmp(f->file_name, "-")) {
622 log_err("fio: can't read/write to stdin/out\n");
628 * move output logging to stderr, if we are writing to stdout
637 flags |= OS_O_DIRECT;
640 td_verror(td, EINVAL, "OS does not support atomic IO");
643 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
647 if (td->o.create_on_open && td->o.allow_create)
650 if (f->filetype != FIO_TYPE_FILE)
651 flags |= FIO_O_NOATIME;
658 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
662 f->fd = dup(STDOUT_FILENO);
664 from_hash = file_lookup_open(f, flags);
665 } else if (td_read(td)) {
666 if (f->filetype == FIO_TYPE_CHAR && !read_only)
672 f->fd = dup(STDIN_FILENO);
674 from_hash = file_lookup_open(f, flags);
675 } else if (td_trim(td)) {
676 assert(!td_rw(td)); /* should have matched above */
679 from_hash = file_lookup_open(f, flags);
683 char buf[FIO_VERROR_SIZE];
686 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
687 flags &= ~FIO_O_NOATIME;
690 if (__e == EMFILE && file_close_shadow_fds(td))
693 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
695 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
696 log_err("fio: looks like your file system does not " \
697 "support direct=1/buffered=0\n");
700 td_verror(td, __e, buf);
704 if (!from_hash && f->fd != -1) {
705 if (add_file_hash(f)) {
709 * Stash away descriptor for later close. This is to
710 * work-around a "feature" on Linux, where a close of
711 * an fd that has been opened for write will trigger
712 * udev to call blkid to check partitions, fs id, etc.
713 * That pollutes the device cache, which can slow down
714 * unbuffered accesses.
716 if (f->shadow_fd == -1)
717 f->shadow_fd = f->fd;
720 * OK to ignore, we haven't done anything
723 ret = generic_close_file(td, f);
733 * This function i.e. get_file_size() is the default .get_file_size
734 * implementation of majority of I/O engines.
736 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
738 return get_file_size(td, f);
742 * open/close all files, so that ->real_file_size gets set
744 static int get_file_sizes(struct thread_data *td)
750 for_each_file(td, f, i) {
751 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
754 if (td_io_get_file_size(td, f)) {
755 if (td->error != ENOENT) {
756 log_err("%s\n", td->verror);
764 * There are corner cases where we end up with -1 for
765 * ->real_file_size due to unsupported file type, etc.
766 * We then just set to size option value divided by number
767 * of files, similar to the way file ->io_size is set.
768 * stat(2) failure doesn't set ->real_file_size to -1.
770 if (f->real_file_size == -1ULL && td->o.size)
771 f->real_file_size = td->o.size / td->o.nr_files;
778 struct flist_head list;
785 * Get free number of bytes for each file on each unique mount.
787 static unsigned long long get_fs_free_counts(struct thread_data *td)
789 struct flist_head *n, *tmp;
790 unsigned long long ret = 0;
791 struct fio_mount *fm;
796 for_each_file(td, f, i) {
800 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
801 if (f->real_file_size != -1ULL)
802 ret += f->real_file_size;
804 } else if (f->filetype != FIO_TYPE_FILE)
808 strncpy(buf, f->file_name, 255);
810 if (stat(buf, &sb) < 0) {
814 if (stat(buf, &sb) < 0)
819 flist_for_each(n, &list) {
820 fm = flist_entry(n, struct fio_mount, list);
821 if (fm->key == sb.st_dev)
830 fm = calloc(1, sizeof(*fm));
831 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
832 fm->base = basename(fm->__base);
834 flist_add(&fm->list, &list);
837 flist_for_each_safe(n, tmp, &list) {
838 unsigned long long sz;
840 fm = flist_entry(n, struct fio_mount, list);
841 flist_del(&fm->list);
843 sz = get_fs_free_size(fm->base);
844 if (sz && sz != -1ULL)
853 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
855 struct thread_options *o = &td->o;
856 unsigned long long align_bs;
857 unsigned long long offset;
859 if (o->file_append && f->filetype == FIO_TYPE_FILE)
860 return f->real_file_size;
862 if (o->start_offset_percent > 0) {
864 * if offset_align is provided, set initial offset
866 if (fio_option_is_set(o, start_offset_align)) {
867 align_bs = o->start_offset_align;
869 /* else take the minimum block size */
870 align_bs = td_min_bs(td);
873 /* calculate the raw offset */
874 offset = (f->real_file_size * o->start_offset_percent / 100) +
875 (td->subjob_number * o->offset_increment);
878 * block align the offset at the next available boundary at
879 * ceiling(offset / align_bs) * align_bs
881 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
884 /* start_offset_percent not set */
885 offset = o->start_offset +
886 td->subjob_number * o->offset_increment;
893 * Open the files and setup files sizes, creating files if necessary.
895 int setup_files(struct thread_data *td)
897 unsigned long long total_size, extend_size;
898 struct thread_options *o = &td->o;
900 unsigned int i, nr_fs_extra = 0;
901 int err = 0, need_extend;
903 const unsigned int bs = td_min_bs(td);
906 dprint(FD_FILE, "setup files\n");
908 old_state = td_bump_runstate(td, TD_SETTING_UP);
910 if (o->read_iolog_file)
914 * Find out physical size of files or devices for this thread,
915 * before we determine I/O size and range of our targets.
916 * If ioengine defines a setup() method, it's responsible for
917 * opening the files and setting f->real_file_size to indicate
918 * the valid range for that file.
920 if (td->io_ops->setup)
921 err = td->io_ops->setup(td);
923 err = get_file_sizes(td);
929 * check sizes. if the files/devices do not exist and the size
930 * isn't passed to fio, abort.
933 for_each_file(td, f, i) {
935 if (f->real_file_size == -1ULL)
938 total_size += f->real_file_size;
942 td->fill_device_size = get_fs_free_counts(td);
945 * device/file sizes are zero and no size given, punt
947 if ((!total_size || total_size == -1ULL) && !o->size &&
948 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
949 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
950 log_err("%s: you need to specify size=\n", o->name);
951 td_verror(td, EINVAL, "total_file_size");
956 * Calculate per-file size and potential extra size for the
957 * first files, if needed (i.e. if we don't have a fixed size).
959 if (!o->file_size_low && o->nr_files) {
962 fs = o->size / o->nr_files;
963 all_fs = fs * o->nr_files;
965 if (all_fs < o->size)
966 nr_fs_extra = (o->size - all_fs) / bs;
970 * now file sizes are known, so we can set ->io_size. if size= is
971 * not given, ->io_size is just equal to ->real_file_size. if size
972 * is given, ->io_size is size / nr_files.
974 extend_size = total_size = 0;
976 for_each_file(td, f, i) {
977 f->file_offset = get_start_offset(td, f);
980 * Update ->io_size depending on options specified.
981 * ->file_size_low being 0 means filesize option isn't set.
982 * Non zero ->file_size_low equals ->file_size_high means
983 * filesize option is set in a fixed size format.
984 * Non zero ->file_size_low not equals ->file_size_high means
985 * filesize option is set in a range format.
987 if (!o->file_size_low) {
989 * no file size or range given, file size is equal to
990 * total size divided by number of files. If the size
991 * doesn't divide nicely with the min blocksize,
992 * make the first files bigger.
1001 * We normally don't come here for regular files, but
1002 * if the result is 0 for a regular file, set it to the
1003 * real file size. This could be size of the existing
1004 * one if it already exists, but otherwise will be set
1005 * to 0. A new file won't be created because
1006 * ->io_size + ->file_offset equals ->real_file_size.
1009 if (f->file_offset > f->real_file_size)
1011 f->io_size = f->real_file_size - f->file_offset;
1013 log_info("fio: file %s may be ignored\n",
1016 } else if (f->real_file_size < o->file_size_low ||
1017 f->real_file_size > o->file_size_high) {
1018 if (f->file_offset > o->file_size_low)
1021 * file size given. if it's fixed, use that. if it's a
1022 * range, generate a random size in-between.
1024 if (o->file_size_low == o->file_size_high)
1025 f->io_size = o->file_size_low - f->file_offset;
1027 f->io_size = get_rand_file_size(td)
1031 f->io_size = f->real_file_size - f->file_offset;
1033 if (f->io_size == -1ULL)
1036 if (o->size_percent && o->size_percent != 100) {
1039 file_size = f->io_size + f->file_offset;
1040 f->io_size = (file_size *
1041 o->size_percent) / 100;
1042 if (f->io_size > (file_size - f->file_offset))
1043 f->io_size = file_size - f->file_offset;
1045 f->io_size -= (f->io_size % td_min_bs(td));
1047 total_size += f->io_size;
1050 if (f->filetype == FIO_TYPE_FILE &&
1051 (f->io_size + f->file_offset) > f->real_file_size &&
1052 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1053 if (!o->create_on_open) {
1055 extend_size += (f->io_size + f->file_offset);
1056 fio_file_set_extend(f);
1058 f->real_file_size = f->io_size + f->file_offset;
1062 if (td->o.block_error_hist) {
1065 assert(td->o.nr_files == 1); /* checked in fixup_options */
1067 len = f->io_size / td->o.bs[DDIR_TRIM];
1068 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1069 log_err("fio: cannot calculate block histogram with "
1070 "%d trim blocks, maximum %d\n",
1071 len, MAX_NR_BLOCK_INFOS);
1072 td_verror(td, EINVAL, "block_error_hist");
1076 td->ts.nr_block_infos = len;
1077 for (i = 0; i < len; i++)
1078 td->ts.block_infos[i] =
1079 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1081 td->ts.nr_block_infos = 0;
1083 if (!o->size || (total_size && o->size > total_size))
1084 o->size = total_size;
1086 if (o->size < td_min_bs(td)) {
1087 log_err("fio: blocksize too large for data set\n");
1092 * See if we need to extend some files, typically needed when our
1093 * target regular files don't exist yet, but our jobs require them
1094 * initially due to read I/Os.
1098 if (output_format & FIO_OUTPUT_NORMAL) {
1099 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1101 need_extend > 1 ? "s" : "",
1103 need_extend > 1 ? "s" : "",
1104 need_extend > 1 ? "total " : "",
1108 for_each_file(td, f, i) {
1109 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1111 if (!fio_file_extend(f))
1114 assert(f->filetype == FIO_TYPE_FILE);
1115 fio_file_clear_extend(f);
1116 if (!o->fill_device) {
1117 old_len = f->real_file_size;
1118 extend_len = f->io_size + f->file_offset -
1121 f->real_file_size = (f->io_size + f->file_offset);
1122 err = extend_file(td, f);
1126 err = __file_invalidate_cache(td, f, old_len,
1130 * Shut up static checker
1146 o->zone_size = o->size;
1149 * iolog already set the total io size, if we read back
1152 if (!o->read_iolog_file) {
1154 td->total_io_size = o->io_size * o->loops;
1156 td->total_io_size = o->size * o->loops;
1163 td_restore_runstate(td, old_state);
1166 log_err("%s: you need to specify valid offset=\n", o->name);
1168 td_restore_runstate(td, old_state);
1172 bool pre_read_files(struct thread_data *td)
1177 dprint(FD_FILE, "pre_read files\n");
1179 for_each_file(td, f, i) {
1180 if (!pre_read_file(td, f))
1187 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1189 unsigned int range_size, seed;
1190 unsigned long nranges;
1193 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1194 fsize = min(f->real_file_size, f->io_size);
1196 nranges = (fsize + range_size - 1) / range_size;
1198 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1199 if (!td->o.rand_repeatable)
1200 seed = td->rand_seeds[4];
1202 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1203 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1204 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1205 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1206 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1207 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1210 static bool init_rand_distribution(struct thread_data *td)
1216 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1219 state = td_bump_runstate(td, TD_SETTING_UP);
1221 for_each_file(td, f, i)
1222 __init_rand_distribution(td, f);
1224 td_restore_runstate(td, state);
1229 * Check if the number of blocks exceeds the randomness capability of
1230 * the selected generator. Tausworthe is 32-bit, the others are fullly
1233 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1236 if (blocks <= FRAND32_MAX)
1238 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1242 * If the user hasn't specified a random generator, switch
1243 * to tausworthe64 with informational warning. If the user did
1244 * specify one, just warn.
1246 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1249 if (!fio_option_is_set(&td->o, random_generator)) {
1250 log_info("fio: Switching to tausworthe64. Use the "
1251 "random_generator= option to get rid of this "
1253 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1258 * Just make this information to avoid breaking scripts.
1260 log_info("fio: Use the random_generator= option to switch to lfsr or "
1265 bool init_random_map(struct thread_data *td)
1267 unsigned long long blocks;
1271 if (init_rand_distribution(td))
1276 for_each_file(td, f, i) {
1277 uint64_t fsize = min(f->real_file_size, f->io_size);
1279 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1281 if (check_rand_gen_limits(td, f, blocks))
1284 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1287 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1289 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1290 fio_file_set_lfsr(f);
1293 } else if (!td->o.norandommap) {
1294 f->io_axmap = axmap_new(blocks);
1296 fio_file_set_axmap(f);
1299 } else if (td->o.norandommap)
1302 if (!td->o.softrandommap) {
1303 log_err("fio: failed allocating random map. If running"
1304 " a large number of jobs, try the 'norandommap'"
1305 " option or set 'softrandommap'. Or give"
1306 " a larger --alloc-size to fio.\n");
1310 log_info("fio: file %s failed allocating random map. Running "
1311 "job without.\n", f->file_name);
1317 void close_files(struct thread_data *td)
1322 for_each_file(td, f, i) {
1323 if (fio_file_open(f))
1324 td_io_close_file(td, f);
1328 void close_and_free_files(struct thread_data *td)
1332 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1334 dprint(FD_FILE, "close files\n");
1336 for_each_file(td, f, i) {
1337 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1338 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1339 td_io_unlink_file(td, f);
1342 if (fio_file_open(f))
1343 td_io_close_file(td, f);
1345 remove_file_hash(f);
1347 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1348 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1349 td_io_unlink_file(td, f);
1355 sfree(f->file_name);
1356 f->file_name = NULL;
1357 if (fio_file_axmap(f)) {
1358 axmap_free(f->io_axmap);
1367 td->o.filename = NULL;
1369 free(td->file_locks);
1370 td->files_index = 0;
1372 td->file_locks = NULL;
1373 td->o.file_lock_mode = FILE_LOCK_NONE;
1377 static void get_file_type(struct fio_file *f)
1381 if (!strcmp(f->file_name, "-"))
1382 f->filetype = FIO_TYPE_PIPE;
1384 f->filetype = FIO_TYPE_FILE;
1387 /* \\.\ is the device namespace in Windows, where every file is
1389 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1390 f->filetype = FIO_TYPE_BLOCK;
1393 if (!stat(f->file_name, &sb)) {
1394 if (S_ISBLK(sb.st_mode))
1395 f->filetype = FIO_TYPE_BLOCK;
1396 else if (S_ISCHR(sb.st_mode))
1397 f->filetype = FIO_TYPE_CHAR;
1398 else if (S_ISFIFO(sb.st_mode))
1399 f->filetype = FIO_TYPE_PIPE;
1403 static bool __is_already_allocated(const char *fname, bool set)
1405 struct flist_head *entry;
1408 ret = file_bloom_exists(fname, set);
1412 flist_for_each(entry, &filename_list) {
1413 struct file_name *fn;
1415 fn = flist_entry(entry, struct file_name, list);
1417 if (!strcmp(fn->filename, fname))
1424 static bool is_already_allocated(const char *fname)
1428 fio_file_hash_lock();
1429 ret = __is_already_allocated(fname, false);
1430 fio_file_hash_unlock();
1435 static void set_already_allocated(const char *fname)
1437 struct file_name *fn;
1439 fn = malloc(sizeof(struct file_name));
1440 fn->filename = strdup(fname);
1442 fio_file_hash_lock();
1443 if (!__is_already_allocated(fname, true)) {
1444 flist_add_tail(&fn->list, &filename_list);
1447 fio_file_hash_unlock();
1455 static void free_already_allocated(void)
1457 struct flist_head *entry, *tmp;
1458 struct file_name *fn;
1460 if (flist_empty(&filename_list))
1463 fio_file_hash_lock();
1464 flist_for_each_safe(entry, tmp, &filename_list) {
1465 fn = flist_entry(entry, struct file_name, list);
1467 flist_del(&fn->list);
1471 fio_file_hash_unlock();
1474 static struct fio_file *alloc_new_file(struct thread_data *td)
1478 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1479 f = calloc(1, sizeof(*f));
1481 f = scalloc(1, sizeof(*f));
1489 fio_file_reset(td, f);
1493 bool exists_and_not_regfile(const char *filename)
1497 if (lstat(filename, &sb) == -1)
1500 #ifndef WIN32 /* NOT Windows */
1501 if (S_ISREG(sb.st_mode))
1504 /* \\.\ is the device namespace in Windows, where every file
1505 * is a device node */
1506 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1513 static bool create_work_dirs(struct thread_data *td, const char *fname)
1515 char path[PATH_MAX];
1518 if (td->o.directory) {
1519 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1520 FIO_OS_PATH_SEPARATOR, fname);
1521 start = strstr(path, fname);
1523 snprintf(path, PATH_MAX, "%s", fname);
1528 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1533 #ifdef CONFIG_HAVE_MKDIR_TWO
1534 if (mkdir(path, 0600) && errno != EEXIST) {
1536 if (mkdir(path) && errno != EEXIST) {
1538 log_err("fio: failed to create dir (%s): %d\n",
1542 *end = FIO_OS_PATH_SEPARATOR;
1545 td->flags |= TD_F_DIRS_CREATED;
1549 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1551 int cur_files = td->files_index;
1552 char file_name[PATH_MAX];
1556 dprint(FD_FILE, "add file %s\n", fname);
1558 if (td->o.directory)
1559 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1560 td->o.unique_filename);
1562 sprintf(file_name + len, "%s", fname);
1564 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1565 !(td->flags & TD_F_DIRS_CREATED) &&
1566 !create_work_dirs(td, fname))
1569 /* clean cloned siblings using existing files */
1570 if (numjob && is_already_allocated(file_name) &&
1571 !exists_and_not_regfile(fname))
1574 f = alloc_new_file(td);
1576 if (td->files_size <= td->files_index) {
1577 unsigned int new_size = td->o.nr_files + 1;
1579 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1581 td->files = realloc(td->files, new_size * sizeof(f));
1582 if (td->files == NULL) {
1583 log_err("fio: realloc OOM\n");
1586 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1587 td->file_locks = realloc(td->file_locks, new_size);
1588 if (!td->file_locks) {
1589 log_err("fio: realloc OOM\n");
1592 td->file_locks[cur_files] = FILE_LOCK_NONE;
1594 td->files_size = new_size;
1596 td->files[cur_files] = f;
1597 f->fileno = cur_files;
1600 * init function, io engine may not be loaded yet
1602 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1603 f->real_file_size = -1ULL;
1605 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1606 f->file_name = strdup(file_name);
1608 f->file_name = smalloc_strdup(file_name);
1610 /* can't handle smalloc failure from here */
1611 assert(f->file_name);
1615 switch (td->o.file_lock_mode) {
1616 case FILE_LOCK_NONE:
1618 case FILE_LOCK_READWRITE:
1619 f->rwlock = fio_rwlock_init();
1621 case FILE_LOCK_EXCLUSIVE:
1622 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1625 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1631 if (td->o.numjobs > 1)
1632 set_already_allocated(file_name);
1637 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1643 int add_file_exclusive(struct thread_data *td, const char *fname)
1648 for_each_file(td, f, i) {
1649 if (!strcmp(f->file_name, fname))
1653 return add_file(td, fname, 0, 1);
1656 void get_file(struct fio_file *f)
1658 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1659 assert(fio_file_open(f));
1663 int put_file(struct thread_data *td, struct fio_file *f)
1665 int f_ret = 0, ret = 0;
1667 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1669 if (!fio_file_open(f)) {
1670 assert(f->fd == -1);
1674 assert(f->references);
1675 if (--f->references)
1678 if (should_fsync(td) && td->o.fsync_on_close) {
1679 f_ret = fsync(f->fd);
1684 if (td->io_ops->close_file)
1685 ret = td->io_ops->close_file(td, f);
1690 td->nr_open_files--;
1691 fio_file_clear_open(f);
1692 assert(f->fd == -1);
1696 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1698 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1701 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1702 if (ddir == DDIR_READ)
1703 fio_rwlock_read(f->rwlock);
1705 fio_rwlock_write(f->rwlock);
1706 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1707 fio_sem_down(f->lock);
1709 td->file_locks[f->fileno] = td->o.file_lock_mode;
1712 void unlock_file(struct thread_data *td, struct fio_file *f)
1714 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1717 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1718 fio_rwlock_unlock(f->rwlock);
1719 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1720 fio_sem_up(f->lock);
1722 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1725 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1727 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1729 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1733 static bool recurse_dir(struct thread_data *td, const char *dirname)
1739 D = opendir(dirname);
1741 char buf[FIO_VERROR_SIZE];
1743 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1744 td_verror(td, errno, buf);
1748 while ((dir = readdir(D)) != NULL) {
1749 char full_path[PATH_MAX];
1752 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1755 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1757 if (lstat(full_path, &sb) == -1) {
1758 if (errno != ENOENT) {
1759 td_verror(td, errno, "stat");
1765 if (S_ISREG(sb.st_mode)) {
1766 add_file(td, full_path, 0, 1);
1769 if (!S_ISDIR(sb.st_mode))
1772 ret = recurse_dir(td, full_path);
1781 int add_dir_files(struct thread_data *td, const char *path)
1783 int ret = recurse_dir(td, path);
1786 log_info("fio: opendir added %d files\n", td->o.nr_files);
1791 void dup_files(struct thread_data *td, struct thread_data *org)
1796 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1801 td->files = malloc(org->files_index * sizeof(f));
1803 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1804 td->file_locks = malloc(org->files_index);
1806 for_each_file(org, f, i) {
1807 struct fio_file *__f;
1809 __f = alloc_new_file(td);
1812 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1813 __f->file_name = strdup(f->file_name);
1815 __f->file_name = smalloc_strdup(f->file_name);
1817 /* can't handle smalloc failure from here */
1818 assert(__f->file_name);
1819 __f->filetype = f->filetype;
1822 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1823 __f->lock = f->lock;
1824 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1825 __f->rwlock = f->rwlock;
1832 * Returns the index that matches the filename, or -1 if not there
1834 int get_fileno(struct thread_data *td, const char *fname)
1839 for_each_file(td, f, i)
1840 if (!strcmp(f->file_name, fname))
1847 * For log usage, where we add/open/close files automatically
1849 void free_release_files(struct thread_data *td)
1853 td->o.open_files = 0;
1854 td->files_index = 0;
1857 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1861 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1862 f->last_pos[i] = f->file_offset;
1863 f->last_start[i] = -1ULL;
1866 if (fio_file_axmap(f))
1867 axmap_reset(f->io_axmap);
1868 else if (fio_file_lfsr(f))
1869 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1872 bool fio_files_done(struct thread_data *td)
1877 for_each_file(td, f, i)
1878 if (!fio_file_done(f))
1884 /* free memory used in initialization phase only */
1885 void filesetup_mem_free(void)
1887 free_already_allocated();
1891 * This function is for platforms which support direct I/O but not O_DIRECT.
1893 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1895 #ifdef FIO_OS_DIRECTIO
1896 int ret = fio_set_odirect(f);
1899 td_verror(td, ret, "fio_set_directio");
1900 #if defined(__sun__)
1901 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1902 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1904 log_err("fio: the file system does not seem to support direct IO\n");
1907 log_err("fio: the file system does not seem to support direct IO\n");
1914 log_err("fio: direct IO is not supported on this host operating system\n");