17 #include "lib/axmap.h"
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
25 static FLIST_HEAD(filename_list);
28 * List entry for filename_list
31 struct flist_head list;
35 static inline void clear_error(struct thread_data *td)
41 static inline int native_fallocate(struct thread_data *td, struct fio_file *f)
45 success = fio_fallocate(f, 0, f->real_file_size);
46 dprint(FD_FILE, "native fallocate of file %s size %llu was "
47 "%ssuccessful\n", f->file_name,
48 (unsigned long long) f->real_file_size,
55 dprint(FD_FILE, "native fallocate is not implemented\n");
60 static void fallocate_file(struct thread_data *td, struct fio_file *f)
64 if (td->o.fill_device)
67 switch (td->o.fallocate_mode) {
68 case FIO_FALLOCATE_NATIVE:
69 native_fallocate(td, f);
71 case FIO_FALLOCATE_NONE:
73 #ifdef CONFIG_POSIX_FALLOCATE
74 case FIO_FALLOCATE_POSIX:
75 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
77 (unsigned long long) f->real_file_size);
79 r = posix_fallocate(f->fd, 0, f->real_file_size);
81 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:
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");
95 #endif /* CONFIG_LINUX_FALLOCATE */
97 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
103 * Leaves f->fd open on success, caller must close
105 static int extend_file(struct thread_data *td, struct fio_file *f)
107 int new_layout = 0, unlink_file = 0, flags;
108 unsigned long long left;
113 log_err("fio: refusing extend of file due to read-only\n");
118 * check if we need to lay the file out complete again. fio
119 * does that for operations involving reads, or for writes
120 * where overwrite is set
123 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
124 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
126 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
129 if (unlink_file || new_layout) {
132 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
134 ret = td_io_unlink_file(td, f);
135 if (ret != 0 && ret != ENOENT) {
136 td_verror(td, errno, "unlink");
142 if (td->o.allow_create)
151 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
152 f->fd = open(f->file_name, flags, 0644);
156 if (err == ENOENT && !td->o.allow_create)
157 log_err("fio: file creation disallowed by "
158 "allow_file_create=0\n");
160 td_verror(td, err, "open");
164 fallocate_file(td, f);
167 * If our jobs don't require regular files initially, we're done.
173 * The size will be -1ULL when fill_device is used, so don't truncate
174 * or fallocate this file, just write it
176 if (!td->o.fill_device) {
177 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
178 (unsigned long long) f->real_file_size);
179 if (ftruncate(f->fd, f->real_file_size) == -1) {
180 if (errno != EFBIG) {
181 td_verror(td, errno, "ftruncate");
187 left = f->real_file_size;
188 bs = td->o.max_bs[DDIR_WRITE];
194 td_verror(td, errno, "malloc");
198 while (left && !td->terminate) {
204 fill_io_buffer(td, b, bs, bs);
206 r = write(f->fd, b, bs);
216 if (td->o.fill_device)
218 log_info("fio: ENOSPC on laying out "
222 td_verror(td, errno, "write");
224 td_verror(td, EIO, "write");
231 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
232 td_io_unlink_file(td, f);
233 } else if (td->o.create_fsync) {
234 if (fsync(f->fd) < 0) {
235 td_verror(td, errno, "fsync");
239 if (td->o.fill_device && !td_write(td)) {
240 fio_file_clear_size_known(f);
241 if (td_io_get_file_size(td, f))
243 if (f->io_size > f->real_file_size)
244 f->io_size = f->real_file_size;
258 static int pre_read_file(struct thread_data *td, struct fio_file *f)
260 int ret = 0, r, did_open = 0, old_runstate;
261 unsigned long long left;
265 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
266 td_ioengine_flagged(td, FIO_NOIO))
269 if (f->filetype == FIO_TYPE_CHAR)
272 if (!fio_file_open(f)) {
273 if (td->io_ops->open_file(td, f)) {
274 log_err("fio: cannot pre-read, failed to open file\n");
280 old_runstate = td_bump_runstate(td, TD_PRE_READING);
283 bs = td->o.max_bs[DDIR_READ];
289 td_verror(td, errno, "malloc");
295 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
296 td_verror(td, errno, "lseek");
297 log_err("fio: failed to lseek pre-read file\n");
302 while (left && !td->terminate) {
306 r = read(f->fd, b, bs);
312 td_verror(td, EIO, "pre_read");
318 td_restore_runstate(td, old_runstate);
321 td->io_ops->close_file(td, f);
327 unsigned long long get_rand_file_size(struct thread_data *td)
329 unsigned long long ret, sized;
333 frand_max = rand_max(&td->file_size_state);
334 r = __rand(&td->file_size_state);
335 sized = td->o.file_size_high - td->o.file_size_low;
336 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
337 ret += td->o.file_size_low;
338 ret -= (ret % td->o.rw_min_bs);
342 static int file_size(struct thread_data *td, struct fio_file *f)
346 if (stat(f->file_name, &st) == -1) {
347 td_verror(td, errno, "fstat");
351 f->real_file_size = st.st_size;
355 static int bdev_size(struct thread_data *td, struct fio_file *f)
357 unsigned long long bytes = 0;
360 if (td->io_ops->open_file(td, f)) {
361 log_err("fio: failed opening blockdev %s for size check\n",
366 r = blockdev_size(f, &bytes);
368 td_verror(td, r, "blockdev_size");
373 log_err("%s: zero sized block device?\n", f->file_name);
377 f->real_file_size = bytes;
378 td->io_ops->close_file(td, f);
381 td->io_ops->close_file(td, f);
385 static int char_size(struct thread_data *td, struct fio_file *f)
387 #ifdef FIO_HAVE_CHARDEV_SIZE
388 unsigned long long bytes = 0;
391 if (td->io_ops->open_file(td, f)) {
392 log_err("fio: failed opening chardev %s for size check\n",
397 r = chardev_size(f, &bytes);
399 td_verror(td, r, "chardev_size");
404 log_err("%s: zero sized char device?\n", f->file_name);
408 f->real_file_size = bytes;
409 td->io_ops->close_file(td, f);
412 td->io_ops->close_file(td, f);
415 f->real_file_size = -1ULL;
420 static int get_file_size(struct thread_data *td, struct fio_file *f)
424 if (fio_file_size_known(f))
427 if (f->filetype == FIO_TYPE_FILE)
428 ret = file_size(td, f);
429 else if (f->filetype == FIO_TYPE_BLOCK)
430 ret = bdev_size(td, f);
431 else if (f->filetype == FIO_TYPE_CHAR)
432 ret = char_size(td, f);
434 f->real_file_size = -1ULL;
437 * Leave ->real_file_size with 0 since it could be expectation
438 * of initial setup for regular files.
444 * If ->real_file_size is -1, a conditional for the message
445 * "offset extends end" is always true, but it makes no sense,
446 * so just return the same value here.
448 if (f->real_file_size == -1ULL) {
449 log_info("%s: failed to get file size of %s\n", td->o.name,
454 if (td->o.start_offset && f->file_offset == 0)
455 dprint(FD_FILE, "offset of file %s not initialized yet\n",
458 * ->file_offset normally hasn't been initialized yet, so this
459 * is basically always false.
461 if (f->file_offset > f->real_file_size) {
462 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
463 (unsigned long long) f->file_offset,
464 (unsigned long long) f->real_file_size);
468 fio_file_set_size_known(f);
472 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
473 unsigned long long off,
474 unsigned long long len)
476 int errval = 0, ret = 0;
485 off = f->file_offset;
487 if (len == -1ULL || off == -1ULL)
490 if (td->io_ops->invalidate) {
491 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
493 ret = td->io_ops->invalidate(td, f);
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()) {
519 log_err("fio: only root may flush block "
520 "devices. Cache flush bypassed!\n");
527 } else if (f->filetype == FIO_TYPE_CHAR ||
528 f->filetype == FIO_TYPE_PIPE) {
529 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
534 * Cache flushing isn't a fatal condition, and we know it will
535 * happen on some platforms where we don't have the proper
536 * function to flush eg block device caches. So just warn and
537 * continue on our way.
540 log_info("fio: cache invalidation of %s failed: %s\n",
541 f->file_name, strerror(errval));
547 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
549 if (!fio_file_open(f))
552 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
555 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
559 dprint(FD_FILE, "fd close %s\n", f->file_name);
563 if (close(f->fd) < 0)
568 if (f->shadow_fd != -1) {
577 int file_lookup_open(struct fio_file *f, int flags)
579 struct fio_file *__f;
582 __f = lookup_file_hash(f->file_name);
584 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
588 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
596 f->fd = open(f->file_name, flags, 0600);
600 static int file_close_shadow_fds(struct thread_data *td)
606 for_each_file(td, f, i) {
607 if (f->shadow_fd == -1)
618 int generic_open_file(struct thread_data *td, struct fio_file *f)
624 dprint(FD_FILE, "fd open %s\n", f->file_name);
626 if (!strcmp(f->file_name, "-")) {
628 log_err("fio: can't read/write to stdin/out\n");
634 * move output logging to stderr, if we are writing to stdout
643 flags |= OS_O_DIRECT;
646 td_verror(td, EINVAL, "OS does not support atomic IO");
649 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
653 if (td->o.create_on_open && td->o.allow_create)
656 if (f->filetype != FIO_TYPE_FILE)
657 flags |= FIO_O_NOATIME;
664 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
668 f->fd = dup(STDOUT_FILENO);
670 from_hash = file_lookup_open(f, flags);
671 } else if (td_read(td)) {
672 if (f->filetype == FIO_TYPE_CHAR && !read_only)
678 f->fd = dup(STDIN_FILENO);
680 from_hash = file_lookup_open(f, flags);
681 } else if (td_trim(td)) {
682 assert(!td_rw(td)); /* should have matched above */
684 from_hash = file_lookup_open(f, flags);
688 char buf[FIO_VERROR_SIZE];
691 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
692 flags &= ~FIO_O_NOATIME;
695 if (__e == EMFILE && file_close_shadow_fds(td))
698 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
700 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
701 log_err("fio: looks like your file system does not " \
702 "support direct=1/buffered=0\n");
705 td_verror(td, __e, buf);
709 if (!from_hash && f->fd != -1) {
710 if (add_file_hash(f)) {
714 * Stash away descriptor for later close. This is to
715 * work-around a "feature" on Linux, where a close of
716 * an fd that has been opened for write will trigger
717 * udev to call blkid to check partitions, fs id, etc.
718 * That pollutes the device cache, which can slow down
719 * unbuffered accesses.
721 if (f->shadow_fd == -1)
722 f->shadow_fd = f->fd;
725 * OK to ignore, we haven't done anything
728 ret = generic_close_file(td, f);
738 * This function i.e. get_file_size() is the default .get_file_size
739 * implementation of majority of I/O engines.
741 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
743 return get_file_size(td, f);
747 * open/close all files, so that ->real_file_size gets set
749 static int get_file_sizes(struct thread_data *td)
755 for_each_file(td, f, i) {
756 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
759 if (td_io_get_file_size(td, f)) {
760 if (td->error != ENOENT) {
761 log_err("%s\n", td->verror);
769 * There are corner cases where we end up with -1 for
770 * ->real_file_size due to unsupported file type, etc.
771 * We then just set to size option value divided by number
772 * of files, similar to the way file ->io_size is set.
773 * stat(2) failure doesn't set ->real_file_size to -1.
775 if (f->real_file_size == -1ULL && td->o.size)
776 f->real_file_size = td->o.size / td->o.nr_files;
783 struct flist_head list;
790 * Get free number of bytes for each file on each unique mount.
792 static unsigned long long get_fs_free_counts(struct thread_data *td)
794 struct flist_head *n, *tmp;
795 unsigned long long ret = 0;
796 struct fio_mount *fm;
801 for_each_file(td, f, i) {
805 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
806 if (f->real_file_size != -1ULL)
807 ret += f->real_file_size;
809 } else if (f->filetype != FIO_TYPE_FILE)
813 strncpy(buf, f->file_name, 255);
815 if (stat(buf, &sb) < 0) {
819 if (stat(buf, &sb) < 0)
824 flist_for_each(n, &list) {
825 fm = flist_entry(n, struct fio_mount, list);
826 if (fm->key == sb.st_dev)
835 fm = calloc(1, sizeof(*fm));
836 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
837 fm->base = basename(fm->__base);
839 flist_add(&fm->list, &list);
842 flist_for_each_safe(n, tmp, &list) {
843 unsigned long long sz;
845 fm = flist_entry(n, struct fio_mount, list);
846 flist_del(&fm->list);
848 sz = get_fs_free_size(fm->base);
849 if (sz && sz != -1ULL)
858 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
860 struct thread_options *o = &td->o;
861 unsigned long long align_bs;
862 unsigned long long offset;
864 if (o->file_append && f->filetype == FIO_TYPE_FILE)
865 return f->real_file_size;
867 if (o->start_offset_percent > 0) {
869 * if offset_align is provided, set initial offset
871 if (fio_option_is_set(o, start_offset_align)) {
872 align_bs = o->start_offset_align;
874 /* else take the minimum block size */
875 align_bs = td_min_bs(td);
878 /* calculate the raw offset */
879 offset = (f->real_file_size * o->start_offset_percent / 100) +
880 (td->subjob_number * o->offset_increment);
883 * block align the offset at the next available boundary at
884 * ceiling(offset / align_bs) * align_bs
886 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
889 /* start_offset_percent not set */
890 offset = o->start_offset +
891 td->subjob_number * o->offset_increment;
898 * Open the files and setup files sizes, creating files if necessary.
900 int setup_files(struct thread_data *td)
902 unsigned long long total_size, extend_size;
903 struct thread_options *o = &td->o;
905 unsigned int i, nr_fs_extra = 0;
906 int err = 0, need_extend;
908 const unsigned int bs = td_min_bs(td);
911 dprint(FD_FILE, "setup files\n");
913 old_state = td_bump_runstate(td, TD_SETTING_UP);
915 if (o->read_iolog_file)
919 * Find out physical size of files or devices for this thread,
920 * before we determine I/O size and range of our targets.
921 * If ioengine defines a setup() method, it's responsible for
922 * opening the files and setting f->real_file_size to indicate
923 * the valid range for that file.
925 if (td->io_ops->setup)
926 err = td->io_ops->setup(td);
928 err = get_file_sizes(td);
934 * check sizes. if the files/devices do not exist and the size
935 * isn't passed to fio, abort.
938 for_each_file(td, f, i) {
940 if (f->real_file_size == -1ULL)
943 total_size += f->real_file_size;
947 td->fill_device_size = get_fs_free_counts(td);
950 * device/file sizes are zero and no size given, punt
952 if ((!total_size || total_size == -1ULL) && !o->size &&
953 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
954 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
955 log_err("%s: you need to specify size=\n", o->name);
956 td_verror(td, EINVAL, "total_file_size");
961 * Calculate per-file size and potential extra size for the
962 * first files, if needed (i.e. if we don't have a fixed size).
964 if (!o->file_size_low && o->nr_files) {
967 fs = o->size / o->nr_files;
968 all_fs = fs * o->nr_files;
970 if (all_fs < o->size)
971 nr_fs_extra = (o->size - all_fs) / bs;
975 * now file sizes are known, so we can set ->io_size. if size= is
976 * not given, ->io_size is just equal to ->real_file_size. if size
977 * is given, ->io_size is size / nr_files.
979 extend_size = total_size = 0;
981 for_each_file(td, f, i) {
982 f->file_offset = get_start_offset(td, f);
985 * Update ->io_size depending on options specified.
986 * ->file_size_low being 0 means filesize option isn't set.
987 * Non zero ->file_size_low equals ->file_size_high means
988 * filesize option is set in a fixed size format.
989 * Non zero ->file_size_low not equals ->file_size_high means
990 * filesize option is set in a range format.
992 if (!o->file_size_low) {
994 * no file size or range given, file size is equal to
995 * total size divided by number of files. If the size
996 * doesn't divide nicely with the min blocksize,
997 * make the first files bigger.
1006 * We normally don't come here for regular files, but
1007 * if the result is 0 for a regular file, set it to the
1008 * real file size. This could be size of the existing
1009 * one if it already exists, but otherwise will be set
1010 * to 0. A new file won't be created because
1011 * ->io_size + ->file_offset equals ->real_file_size.
1014 if (f->file_offset > f->real_file_size)
1016 f->io_size = f->real_file_size - f->file_offset;
1018 log_info("fio: file %s may be ignored\n",
1021 } else if (f->real_file_size < o->file_size_low ||
1022 f->real_file_size > o->file_size_high) {
1023 if (f->file_offset > o->file_size_low)
1026 * file size given. if it's fixed, use that. if it's a
1027 * range, generate a random size in-between.
1029 if (o->file_size_low == o->file_size_high)
1030 f->io_size = o->file_size_low - f->file_offset;
1032 f->io_size = get_rand_file_size(td)
1036 f->io_size = f->real_file_size - f->file_offset;
1038 if (f->io_size == -1ULL)
1041 if (o->size_percent) {
1044 file_size = f->io_size + f->file_offset;
1045 f->io_size = (file_size *
1046 o->size_percent) / 100;
1047 if (f->io_size > (file_size - f->file_offset))
1048 f->io_size = file_size - f->file_offset;
1050 f->io_size -= (f->io_size % td_min_bs(td));
1052 total_size += f->io_size;
1055 if (f->filetype == FIO_TYPE_FILE &&
1056 (f->io_size + f->file_offset) > f->real_file_size &&
1057 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1058 if (!o->create_on_open) {
1060 extend_size += (f->io_size + f->file_offset);
1061 fio_file_set_extend(f);
1063 f->real_file_size = f->io_size + f->file_offset;
1067 if (td->o.block_error_hist) {
1070 assert(td->o.nr_files == 1); /* checked in fixup_options */
1072 len = f->io_size / td->o.bs[DDIR_TRIM];
1073 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1074 log_err("fio: cannot calculate block histogram with "
1075 "%d trim blocks, maximum %d\n",
1076 len, MAX_NR_BLOCK_INFOS);
1077 td_verror(td, EINVAL, "block_error_hist");
1081 td->ts.nr_block_infos = len;
1082 for (i = 0; i < len; i++)
1083 td->ts.block_infos[i] =
1084 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1086 td->ts.nr_block_infos = 0;
1088 if (!o->size || (total_size && o->size > total_size))
1089 o->size = total_size;
1091 if (o->size < td_min_bs(td)) {
1092 log_err("fio: blocksize too large for data set\n");
1097 * See if we need to extend some files, typically needed when our
1098 * target regular files don't exist yet, but our jobs require them
1099 * initially due to read I/Os.
1103 if (output_format & FIO_OUTPUT_NORMAL) {
1104 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1106 need_extend > 1 ? "s" : "",
1108 need_extend > 1 ? "s" : "",
1109 need_extend > 1 ? "total " : "",
1113 for_each_file(td, f, i) {
1114 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1116 if (!fio_file_extend(f))
1119 assert(f->filetype == FIO_TYPE_FILE);
1120 fio_file_clear_extend(f);
1121 if (!o->fill_device) {
1122 old_len = f->real_file_size;
1123 extend_len = f->io_size + f->file_offset -
1126 f->real_file_size = (f->io_size + f->file_offset);
1127 err = extend_file(td, f);
1131 err = __file_invalidate_cache(td, f, old_len,
1135 * Shut up static checker
1151 o->zone_size = o->size;
1154 * iolog already set the total io size, if we read back
1157 if (!o->read_iolog_file) {
1159 td->total_io_size = o->io_size * o->loops;
1161 td->total_io_size = o->size * o->loops;
1168 td_restore_runstate(td, old_state);
1171 log_err("%s: you need to specify valid offset=\n", o->name);
1173 td_restore_runstate(td, old_state);
1177 int pre_read_files(struct thread_data *td)
1182 dprint(FD_FILE, "pre_read files\n");
1184 for_each_file(td, f, i) {
1185 if (pre_read_file(td, f))
1192 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1194 unsigned int range_size, seed;
1195 unsigned long nranges;
1198 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1199 fsize = min(f->real_file_size, f->io_size);
1201 nranges = (fsize + range_size - 1) / range_size;
1203 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1204 if (!td->o.rand_repeatable)
1205 seed = td->rand_seeds[4];
1207 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1208 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1209 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1210 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1211 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1212 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1217 static int init_rand_distribution(struct thread_data *td)
1223 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1226 state = td_bump_runstate(td, TD_SETTING_UP);
1228 for_each_file(td, f, i)
1229 __init_rand_distribution(td, f);
1231 td_restore_runstate(td, state);
1237 * Check if the number of blocks exceeds the randomness capability of
1238 * the selected generator. Tausworthe is 32-bit, the others are fullly
1241 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1244 if (blocks <= FRAND32_MAX)
1246 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1250 * If the user hasn't specified a random generator, switch
1251 * to tausworthe64 with informational warning. If the user did
1252 * specify one, just warn.
1254 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1257 if (!fio_option_is_set(&td->o, random_generator)) {
1258 log_info("fio: Switching to tausworthe64. Use the "
1259 "random_generator= option to get rid of this "
1261 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1266 * Just make this information to avoid breaking scripts.
1268 log_info("fio: Use the random_generator= option to switch to lfsr or "
1273 int init_random_map(struct thread_data *td)
1275 unsigned long long blocks;
1279 if (init_rand_distribution(td))
1284 for_each_file(td, f, i) {
1285 uint64_t fsize = min(f->real_file_size, f->io_size);
1287 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1289 if (check_rand_gen_limits(td, f, blocks))
1292 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1295 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1297 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1298 fio_file_set_lfsr(f);
1301 } else if (!td->o.norandommap) {
1302 f->io_axmap = axmap_new(blocks);
1304 fio_file_set_axmap(f);
1307 } else if (td->o.norandommap)
1310 if (!td->o.softrandommap) {
1311 log_err("fio: failed allocating random map. If running"
1312 " a large number of jobs, try the 'norandommap'"
1313 " option or set 'softrandommap'. Or give"
1314 " a larger --alloc-size to fio.\n");
1318 log_info("fio: file %s failed allocating random map. Running "
1319 "job without.\n", f->file_name);
1325 void close_files(struct thread_data *td)
1330 for_each_file(td, f, i) {
1331 if (fio_file_open(f))
1332 td_io_close_file(td, f);
1336 void close_and_free_files(struct thread_data *td)
1340 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1342 dprint(FD_FILE, "close files\n");
1344 for_each_file(td, f, i) {
1345 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1346 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1347 td_io_unlink_file(td, f);
1350 if (fio_file_open(f))
1351 td_io_close_file(td, f);
1353 remove_file_hash(f);
1355 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1356 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1357 td_io_unlink_file(td, f);
1363 sfree(f->file_name);
1364 f->file_name = NULL;
1365 if (fio_file_axmap(f)) {
1366 axmap_free(f->io_axmap);
1375 td->o.filename = NULL;
1377 free(td->file_locks);
1378 td->files_index = 0;
1380 td->file_locks = NULL;
1381 td->o.file_lock_mode = FILE_LOCK_NONE;
1385 static void get_file_type(struct fio_file *f)
1389 if (!strcmp(f->file_name, "-"))
1390 f->filetype = FIO_TYPE_PIPE;
1392 f->filetype = FIO_TYPE_FILE;
1395 /* \\.\ is the device namespace in Windows, where every file is
1397 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1398 f->filetype = FIO_TYPE_BLOCK;
1401 if (!stat(f->file_name, &sb)) {
1402 if (S_ISBLK(sb.st_mode))
1403 f->filetype = FIO_TYPE_BLOCK;
1404 else if (S_ISCHR(sb.st_mode))
1405 f->filetype = FIO_TYPE_CHAR;
1406 else if (S_ISFIFO(sb.st_mode))
1407 f->filetype = FIO_TYPE_PIPE;
1411 static bool __is_already_allocated(const char *fname, bool set)
1413 struct flist_head *entry;
1416 ret = file_bloom_exists(fname, set);
1420 flist_for_each(entry, &filename_list) {
1421 struct file_name *fn;
1423 fn = flist_entry(entry, struct file_name, list);
1425 if (!strcmp(fn->filename, fname))
1432 static bool is_already_allocated(const char *fname)
1436 fio_file_hash_lock();
1437 ret = __is_already_allocated(fname, false);
1438 fio_file_hash_unlock();
1443 static void set_already_allocated(const char *fname)
1445 struct file_name *fn;
1447 fn = malloc(sizeof(struct file_name));
1448 fn->filename = strdup(fname);
1450 fio_file_hash_lock();
1451 if (!__is_already_allocated(fname, true)) {
1452 flist_add_tail(&fn->list, &filename_list);
1455 fio_file_hash_unlock();
1463 static void free_already_allocated(void)
1465 struct flist_head *entry, *tmp;
1466 struct file_name *fn;
1468 if (flist_empty(&filename_list))
1471 fio_file_hash_lock();
1472 flist_for_each_safe(entry, tmp, &filename_list) {
1473 fn = flist_entry(entry, struct file_name, list);
1475 flist_del(&fn->list);
1479 fio_file_hash_unlock();
1482 static struct fio_file *alloc_new_file(struct thread_data *td)
1486 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1487 f = calloc(1, sizeof(*f));
1489 f = smalloc(sizeof(*f));
1497 fio_file_reset(td, f);
1501 bool exists_and_not_regfile(const char *filename)
1505 if (lstat(filename, &sb) == -1)
1508 #ifndef WIN32 /* NOT Windows */
1509 if (S_ISREG(sb.st_mode))
1512 /* \\.\ is the device namespace in Windows, where every file
1513 * is a device node */
1514 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1521 static int create_work_dirs(struct thread_data *td, const char *fname)
1523 char path[PATH_MAX];
1526 if (td->o.directory) {
1527 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1528 FIO_OS_PATH_SEPARATOR, fname);
1529 start = strstr(path, fname);
1531 snprintf(path, PATH_MAX, "%s", fname);
1536 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1541 #ifdef CONFIG_HAVE_MKDIR_TWO
1542 if (mkdir(path, 0600) && errno != EEXIST) {
1544 if (mkdir(path) && errno != EEXIST) {
1546 log_err("fio: failed to create dir (%s): %d\n",
1550 *end = FIO_OS_PATH_SEPARATOR;
1553 td->flags |= TD_F_DIRS_CREATED;
1557 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1559 int cur_files = td->files_index;
1560 char file_name[PATH_MAX];
1564 dprint(FD_FILE, "add file %s\n", fname);
1566 if (td->o.directory)
1567 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1568 td->o.unique_filename);
1570 sprintf(file_name + len, "%s", fname);
1572 if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1573 !(td->flags & TD_F_DIRS_CREATED) &&
1574 create_work_dirs(td, fname))
1577 /* clean cloned siblings using existing files */
1578 if (numjob && is_already_allocated(file_name) &&
1579 !exists_and_not_regfile(fname))
1582 f = alloc_new_file(td);
1584 if (td->files_size <= td->files_index) {
1585 unsigned int new_size = td->o.nr_files + 1;
1587 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1589 td->files = realloc(td->files, new_size * sizeof(f));
1590 if (td->files == NULL) {
1591 log_err("fio: realloc OOM\n");
1594 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1595 td->file_locks = realloc(td->file_locks, new_size);
1596 if (!td->file_locks) {
1597 log_err("fio: realloc OOM\n");
1600 td->file_locks[cur_files] = FILE_LOCK_NONE;
1602 td->files_size = new_size;
1604 td->files[cur_files] = f;
1605 f->fileno = cur_files;
1608 * init function, io engine may not be loaded yet
1610 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1611 f->real_file_size = -1ULL;
1613 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1614 f->file_name = strdup(file_name);
1616 f->file_name = smalloc_strdup(file_name);
1622 switch (td->o.file_lock_mode) {
1623 case FILE_LOCK_NONE:
1625 case FILE_LOCK_READWRITE:
1626 f->rwlock = fio_rwlock_init();
1628 case FILE_LOCK_EXCLUSIVE:
1629 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1632 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1637 if (f->filetype == FIO_TYPE_FILE)
1638 td->nr_normal_files++;
1640 if (td->o.numjobs > 1)
1641 set_already_allocated(file_name);
1646 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1652 int add_file_exclusive(struct thread_data *td, const char *fname)
1657 for_each_file(td, f, i) {
1658 if (!strcmp(f->file_name, fname))
1662 return add_file(td, fname, 0, 1);
1665 void get_file(struct fio_file *f)
1667 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1668 assert(fio_file_open(f));
1672 int put_file(struct thread_data *td, struct fio_file *f)
1674 int f_ret = 0, ret = 0;
1676 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1678 if (!fio_file_open(f)) {
1679 assert(f->fd == -1);
1683 assert(f->references);
1684 if (--f->references)
1687 if (should_fsync(td) && td->o.fsync_on_close) {
1688 f_ret = fsync(f->fd);
1693 if (td->io_ops->close_file)
1694 ret = td->io_ops->close_file(td, f);
1699 td->nr_open_files--;
1700 fio_file_clear_open(f);
1701 assert(f->fd == -1);
1705 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1707 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1710 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1711 if (ddir == DDIR_READ)
1712 fio_rwlock_read(f->rwlock);
1714 fio_rwlock_write(f->rwlock);
1715 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1716 fio_mutex_down(f->lock);
1718 td->file_locks[f->fileno] = td->o.file_lock_mode;
1721 void unlock_file(struct thread_data *td, struct fio_file *f)
1723 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1726 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1727 fio_rwlock_unlock(f->rwlock);
1728 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1729 fio_mutex_up(f->lock);
1731 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1734 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1736 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1738 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1742 static int recurse_dir(struct thread_data *td, const char *dirname)
1748 D = opendir(dirname);
1750 char buf[FIO_VERROR_SIZE];
1752 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1753 td_verror(td, errno, buf);
1757 while ((dir = readdir(D)) != NULL) {
1758 char full_path[PATH_MAX];
1761 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1764 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1766 if (lstat(full_path, &sb) == -1) {
1767 if (errno != ENOENT) {
1768 td_verror(td, errno, "stat");
1774 if (S_ISREG(sb.st_mode)) {
1775 add_file(td, full_path, 0, 1);
1778 if (!S_ISDIR(sb.st_mode))
1781 ret = recurse_dir(td, full_path);
1790 int add_dir_files(struct thread_data *td, const char *path)
1792 int ret = recurse_dir(td, path);
1795 log_info("fio: opendir added %d files\n", td->o.nr_files);
1800 void dup_files(struct thread_data *td, struct thread_data *org)
1805 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1810 td->files = malloc(org->files_index * sizeof(f));
1812 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1813 td->file_locks = malloc(org->files_index);
1815 for_each_file(org, f, i) {
1816 struct fio_file *__f;
1818 __f = alloc_new_file(td);
1821 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1822 __f->file_name = strdup(f->file_name);
1824 __f->file_name = smalloc_strdup(f->file_name);
1825 if (!__f->file_name)
1828 __f->filetype = f->filetype;
1831 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1832 __f->lock = f->lock;
1833 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1834 __f->rwlock = f->rwlock;
1841 * Returns the index that matches the filename, or -1 if not there
1843 int get_fileno(struct thread_data *td, const char *fname)
1848 for_each_file(td, f, i)
1849 if (!strcmp(f->file_name, fname))
1856 * For log usage, where we add/open/close files automatically
1858 void free_release_files(struct thread_data *td)
1862 td->o.open_files = 0;
1863 td->files_index = 0;
1864 td->nr_normal_files = 0;
1867 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1871 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1872 f->last_pos[i] = f->file_offset;
1873 f->last_start[i] = -1ULL;
1876 if (fio_file_axmap(f))
1877 axmap_reset(f->io_axmap);
1878 else if (fio_file_lfsr(f))
1879 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1882 bool fio_files_done(struct thread_data *td)
1887 for_each_file(td, f, i)
1888 if (!fio_file_done(f))
1894 /* free memory used in initialization phase only */
1895 void filesetup_mem_free(void)
1897 free_already_allocated();
1901 * This function is for platforms which support direct I/O but not O_DIRECT.
1903 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1905 #ifdef FIO_OS_DIRECTIO
1906 int ret = fio_set_odirect(f);
1909 td_verror(td, ret, "fio_set_directio");
1910 #if defined(__sun__)
1911 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1912 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1914 log_err("fio: the file system does not seem to support direct IO\n");
1917 log_err("fio: the file system does not seem to support direct IO\n");
1924 log_err("fio: direct IO is not supported on this host operating system\n");