15 #include "lib/axmap.h"
19 #ifdef CONFIG_LINUX_FALLOCATE
20 #include <linux/falloc.h>
23 static FLIST_HEAD(filename_list);
26 * List entry for filename_list
29 struct flist_head list;
33 static inline void clear_error(struct thread_data *td)
39 static int native_fallocate(struct thread_data *td, struct fio_file *f)
43 success = fio_fallocate(f, 0, f->real_file_size);
44 dprint(FD_FILE, "native fallocate of file %s size %llu was "
45 "%ssuccessful\n", f->file_name,
46 (unsigned long long) f->real_file_size,
53 dprint(FD_FILE, "native fallocate is not implemented\n");
58 static void fallocate_file(struct thread_data *td, struct fio_file *f)
60 if (td->o.fill_device)
63 switch (td->o.fallocate_mode) {
64 case FIO_FALLOCATE_NATIVE:
65 native_fallocate(td, f);
67 case FIO_FALLOCATE_NONE:
69 #ifdef CONFIG_POSIX_FALLOCATE
70 case FIO_FALLOCATE_POSIX: {
73 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
75 (unsigned long long) f->real_file_size);
77 r = posix_fallocate(f->fd, 0, f->real_file_size);
79 log_err("fio: posix_fallocate fails: %s\n", strerror(r));
82 #endif /* CONFIG_POSIX_FALLOCATE */
83 #ifdef CONFIG_LINUX_FALLOCATE
84 case FIO_FALLOCATE_KEEP_SIZE: {
87 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
88 "file %s size %llu\n", f->file_name,
89 (unsigned long long) f->real_file_size);
91 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
93 td_verror(td, errno, "fallocate");
97 #endif /* CONFIG_LINUX_FALLOCATE */
99 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
105 * Leaves f->fd open on success, caller must close
107 static int extend_file(struct thread_data *td, struct fio_file *f)
109 int new_layout = 0, unlink_file = 0, flags;
110 unsigned long long left;
111 unsigned long long bs;
115 log_err("fio: refusing extend of file due to read-only\n");
120 * check if we need to lay the file out complete again. fio
121 * does that for operations involving reads, or for writes
122 * where overwrite is set
125 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
126 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
128 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
131 if (unlink_file || new_layout) {
134 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
136 ret = td_io_unlink_file(td, f);
137 if (ret != 0 && ret != ENOENT) {
138 td_verror(td, errno, "unlink");
144 if (td->o.allow_create)
153 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
154 f->fd = open(f->file_name, flags, 0644);
158 if (err == ENOENT && !td->o.allow_create)
159 log_err("fio: file creation disallowed by "
160 "allow_file_create=0\n");
162 td_verror(td, err, "open");
166 fallocate_file(td, f);
169 * If our jobs don't require regular files initially, we're done.
175 * The size will be -1ULL when fill_device is used, so don't truncate
176 * or fallocate this file, just write it
178 if (!td->o.fill_device) {
179 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
180 (unsigned long long) f->real_file_size);
181 if (ftruncate(f->fd, f->real_file_size) == -1) {
182 if (errno != EFBIG) {
183 td_verror(td, errno, "ftruncate");
189 left = f->real_file_size;
190 bs = td->o.max_bs[DDIR_WRITE];
196 td_verror(td, errno, "malloc");
200 while (left && !td->terminate) {
206 fill_io_buffer(td, b, bs, bs);
208 r = write(f->fd, b, bs);
218 if (td->o.fill_device)
220 log_info("fio: ENOSPC on laying out "
224 td_verror(td, errno, "write");
226 td_verror(td, EIO, "write");
233 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
234 td_io_unlink_file(td, f);
235 } else if (td->o.create_fsync) {
236 if (fsync(f->fd) < 0) {
237 td_verror(td, errno, "fsync");
241 if (td->o.fill_device && !td_write(td)) {
242 fio_file_clear_size_known(f);
243 if (td_io_get_file_size(td, f))
245 if (f->io_size > f->real_file_size)
246 f->io_size = f->real_file_size;
260 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
262 int r, did_open = 0, old_runstate;
263 unsigned long long left;
264 unsigned long long bs;
268 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
269 td_ioengine_flagged(td, FIO_NOIO))
272 if (f->filetype == FIO_TYPE_CHAR)
275 if (!fio_file_open(f)) {
276 if (td->io_ops->open_file(td, f)) {
277 log_err("fio: cannot pre-read, failed to open file\n");
283 old_runstate = td_bump_runstate(td, TD_PRE_READING);
286 bs = td->o.max_bs[DDIR_READ];
292 td_verror(td, errno, "malloc");
298 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
299 td_verror(td, errno, "lseek");
300 log_err("fio: failed to lseek pre-read file\n");
305 while (left && !td->terminate) {
309 r = read(f->fd, b, bs);
315 td_verror(td, EIO, "pre_read");
321 td_restore_runstate(td, old_runstate);
324 td->io_ops->close_file(td, f);
330 unsigned long long get_rand_file_size(struct thread_data *td)
332 unsigned long long ret, sized;
336 frand_max = rand_max(&td->file_size_state);
337 r = __rand(&td->file_size_state);
338 sized = td->o.file_size_high - td->o.file_size_low;
339 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
340 ret += td->o.file_size_low;
341 ret -= (ret % td->o.rw_min_bs);
345 static int file_size(struct thread_data *td, struct fio_file *f)
349 if (stat(f->file_name, &st) == -1) {
350 td_verror(td, errno, "fstat");
354 f->real_file_size = st.st_size;
358 static int bdev_size(struct thread_data *td, struct fio_file *f)
360 unsigned long long bytes = 0;
363 if (td->io_ops->open_file(td, f)) {
364 log_err("fio: failed opening blockdev %s for size check\n",
369 r = blockdev_size(f, &bytes);
371 td_verror(td, r, "blockdev_size");
376 log_err("%s: zero sized block device?\n", f->file_name);
380 f->real_file_size = bytes;
381 td->io_ops->close_file(td, f);
384 td->io_ops->close_file(td, f);
388 static int char_size(struct thread_data *td, struct fio_file *f)
390 #ifdef FIO_HAVE_CHARDEV_SIZE
391 unsigned long long bytes = 0;
394 if (td->io_ops->open_file(td, f)) {
395 log_err("fio: failed opening chardev %s for size check\n",
400 r = chardev_size(f, &bytes);
402 td_verror(td, r, "chardev_size");
407 log_err("%s: zero sized char device?\n", f->file_name);
411 f->real_file_size = bytes;
412 td->io_ops->close_file(td, f);
415 td->io_ops->close_file(td, f);
418 f->real_file_size = -1ULL;
423 static int get_file_size(struct thread_data *td, struct fio_file *f)
427 if (fio_file_size_known(f))
430 if (f->filetype == FIO_TYPE_FILE)
431 ret = file_size(td, f);
432 else if (f->filetype == FIO_TYPE_BLOCK)
433 ret = bdev_size(td, f);
434 else if (f->filetype == FIO_TYPE_CHAR)
435 ret = char_size(td, f);
437 f->real_file_size = -1;
438 log_info("%s: failed to get file size of %s\n", td->o.name,
440 return 1; /* avoid offset extends end error message */
444 * Leave ->real_file_size with 0 since it could be expectation
445 * of initial setup for regular files.
451 * ->file_offset normally hasn't been initialized yet, so this
452 * is basically always false unless ->real_file_size is -1, but
453 * if ->real_file_size is -1 this message doesn't make sense.
454 * As a result, this message is basically useless.
456 if (f->file_offset > f->real_file_size) {
457 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
458 (unsigned long long) f->file_offset,
459 (unsigned long long) f->real_file_size);
463 fio_file_set_size_known(f);
467 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
468 unsigned long long off,
469 unsigned long long len)
471 int errval = 0, ret = 0;
480 off = f->file_offset;
482 if (len == -1ULL || off == -1ULL)
485 if (td->io_ops->invalidate) {
486 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
488 ret = td->io_ops->invalidate(td, f);
491 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
492 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
494 } else if (f->filetype == FIO_TYPE_FILE) {
495 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
496 f->file_name, off, len);
497 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
500 } else if (f->filetype == FIO_TYPE_BLOCK) {
503 dprint(FD_IO, "drop page cache %s\n", f->file_name);
504 ret = blockdev_invalidate_cache(f);
505 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
507 * Linux multipath devices reject ioctl while
508 * the maps are being updated. That window can
509 * last tens of milliseconds; we'll try up to
510 * a quarter of a second.
513 ret = blockdev_invalidate_cache(f);
515 if (ret < 0 && errno == EACCES && geteuid()) {
516 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
517 log_err("fio: only root may flush block "
518 "devices. Cache flush bypassed!\n");
523 } else if (f->filetype == FIO_TYPE_CHAR ||
524 f->filetype == FIO_TYPE_PIPE) {
525 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
529 * Cache flushing isn't a fatal condition, and we know it will
530 * happen on some platforms where we don't have the proper
531 * function to flush eg block device caches. So just warn and
532 * continue on our way.
535 log_info("fio: cache invalidation of %s failed: %s\n",
536 f->file_name, strerror(errval));
542 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
544 if (!fio_file_open(f))
547 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
550 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
554 dprint(FD_FILE, "fd close %s\n", f->file_name);
558 if (close(f->fd) < 0)
563 if (f->shadow_fd != -1) {
572 int file_lookup_open(struct fio_file *f, int flags)
574 struct fio_file *__f;
577 __f = lookup_file_hash(f->file_name);
579 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
583 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
591 f->fd = open(f->file_name, flags, 0600);
595 static int file_close_shadow_fds(struct thread_data *td)
601 for_each_file(td, f, i) {
602 if (f->shadow_fd == -1)
613 int generic_open_file(struct thread_data *td, struct fio_file *f)
619 dprint(FD_FILE, "fd open %s\n", f->file_name);
621 if (!strcmp(f->file_name, "-")) {
623 log_err("fio: can't read/write to stdin/out\n");
629 * move output logging to stderr, if we are writing to stdout
638 flags |= OS_O_DIRECT;
641 td_verror(td, EINVAL, "OS does not support atomic IO");
644 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
648 if (td->o.create_on_open && td->o.allow_create)
651 if (f->filetype != FIO_TYPE_FILE)
652 flags |= FIO_O_NOATIME;
659 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
663 f->fd = dup(STDOUT_FILENO);
665 from_hash = file_lookup_open(f, flags);
666 } else if (td_read(td)) {
667 if (f->filetype == FIO_TYPE_CHAR && !read_only)
673 f->fd = dup(STDIN_FILENO);
675 from_hash = file_lookup_open(f, flags);
676 } else if (td_trim(td)) {
677 assert(!td_rw(td)); /* should have matched above */
680 from_hash = file_lookup_open(f, flags);
684 char buf[FIO_VERROR_SIZE];
687 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
688 flags &= ~FIO_O_NOATIME;
691 if (__e == EMFILE && file_close_shadow_fds(td))
694 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
696 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
697 log_err("fio: looks like your file system does not " \
698 "support direct=1/buffered=0\n");
701 td_verror(td, __e, buf);
705 if (!from_hash && f->fd != -1) {
706 if (add_file_hash(f)) {
710 * Stash away descriptor for later close. This is to
711 * work-around a "feature" on Linux, where a close of
712 * an fd that has been opened for write will trigger
713 * udev to call blkid to check partitions, fs id, etc.
714 * That pollutes the device cache, which can slow down
715 * unbuffered accesses.
717 if (f->shadow_fd == -1)
718 f->shadow_fd = f->fd;
721 * OK to ignore, we haven't done anything
724 ret = generic_close_file(td, f);
734 * This function i.e. get_file_size() is the default .get_file_size
735 * implementation of majority of I/O engines.
737 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
739 return get_file_size(td, f);
743 * open/close all files, so that ->real_file_size gets set
745 static int get_file_sizes(struct thread_data *td)
751 for_each_file(td, f, i) {
752 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
755 if (td_io_get_file_size(td, f)) {
756 if (td->error != ENOENT) {
757 log_err("%s\n", td->verror);
765 * There are corner cases where we end up with -1 for
766 * ->real_file_size due to unsupported file type, etc.
767 * We then just set to size option value divided by number
768 * of files, similar to the way file ->io_size is set.
769 * stat(2) failure doesn't set ->real_file_size to -1.
771 if (f->real_file_size == -1ULL && td->o.size)
772 f->real_file_size = td->o.size / td->o.nr_files;
779 struct flist_head list;
786 * Get free number of bytes for each file on each unique mount.
788 static unsigned long long get_fs_free_counts(struct thread_data *td)
790 struct flist_head *n, *tmp;
791 unsigned long long ret = 0;
792 struct fio_mount *fm;
797 for_each_file(td, f, i) {
801 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
802 if (f->real_file_size != -1ULL)
803 ret += f->real_file_size;
805 } else if (f->filetype != FIO_TYPE_FILE)
808 snprintf(buf, ARRAY_SIZE(buf), "%s", f->file_name);
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 snprintf(fm->__base, ARRAY_SIZE(fm->__base), "%s", buf);
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;
892 static bool create_work_dirs(struct thread_data *td, const char *fname)
897 if (td->o.directory) {
898 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
899 FIO_OS_PATH_SEPARATOR, fname);
900 start = strstr(path, fname);
902 snprintf(path, PATH_MAX, "%s", fname);
907 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
912 #ifdef CONFIG_HAVE_MKDIR_TWO
913 if (mkdir(path, 0600) && errno != EEXIST) {
915 if (mkdir(path) && errno != EEXIST) {
917 log_err("fio: failed to create dir (%s): %d\n",
921 *end = FIO_OS_PATH_SEPARATOR;
924 td->flags |= TD_F_DIRS_CREATED;
929 * Open the files and setup files sizes, creating files if necessary.
931 int setup_files(struct thread_data *td)
933 unsigned long long total_size, extend_size;
934 struct thread_options *o = &td->o;
936 unsigned int i, nr_fs_extra = 0;
937 int err = 0, need_extend;
939 const unsigned long long bs = td_min_bs(td);
942 dprint(FD_FILE, "setup files\n");
944 old_state = td_bump_runstate(td, TD_SETTING_UP);
946 for_each_file(td, f, i) {
947 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
948 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
949 !(td->flags & TD_F_DIRS_CREATED) &&
950 !create_work_dirs(td, f->file_name))
955 * Find out physical size of files or devices for this thread,
956 * before we determine I/O size and range of our targets.
957 * If ioengine defines a setup() method, it's responsible for
958 * opening the files and setting f->real_file_size to indicate
959 * the valid range for that file.
961 if (td->io_ops->setup)
962 err = td->io_ops->setup(td);
964 err = get_file_sizes(td);
969 if (o->read_iolog_file)
973 * check sizes. if the files/devices do not exist and the size
974 * isn't passed to fio, abort.
977 for_each_file(td, f, i) {
979 if (f->real_file_size == -1ULL)
982 total_size += f->real_file_size;
986 td->fill_device_size = get_fs_free_counts(td);
989 * device/file sizes are zero and no size given, punt
991 if ((!total_size || total_size == -1ULL) && !o->size &&
992 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
993 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
994 log_err("%s: you need to specify size=\n", o->name);
995 td_verror(td, EINVAL, "total_file_size");
1000 * Calculate per-file size and potential extra size for the
1001 * first files, if needed (i.e. if we don't have a fixed size).
1003 if (!o->file_size_low && o->nr_files) {
1006 fs = o->size / o->nr_files;
1007 all_fs = fs * o->nr_files;
1009 if (all_fs < o->size)
1010 nr_fs_extra = (o->size - all_fs) / bs;
1014 * now file sizes are known, so we can set ->io_size. if size= is
1015 * not given, ->io_size is just equal to ->real_file_size. if size
1016 * is given, ->io_size is size / nr_files.
1018 extend_size = total_size = 0;
1020 for_each_file(td, f, i) {
1021 f->file_offset = get_start_offset(td, f);
1024 * Update ->io_size depending on options specified.
1025 * ->file_size_low being 0 means filesize option isn't set.
1026 * Non zero ->file_size_low equals ->file_size_high means
1027 * filesize option is set in a fixed size format.
1028 * Non zero ->file_size_low not equals ->file_size_high means
1029 * filesize option is set in a range format.
1031 if (!o->file_size_low) {
1033 * no file size or range given, file size is equal to
1034 * total size divided by number of files. If the size
1035 * doesn't divide nicely with the min blocksize,
1036 * make the first files bigger.
1045 * We normally don't come here for regular files, but
1046 * if the result is 0 for a regular file, set it to the
1047 * real file size. This could be size of the existing
1048 * one if it already exists, but otherwise will be set
1049 * to 0. A new file won't be created because
1050 * ->io_size + ->file_offset equals ->real_file_size.
1053 if (f->file_offset > f->real_file_size)
1055 f->io_size = f->real_file_size - f->file_offset;
1057 log_info("fio: file %s may be ignored\n",
1060 } else if (f->real_file_size < o->file_size_low ||
1061 f->real_file_size > o->file_size_high) {
1062 if (f->file_offset > o->file_size_low)
1065 * file size given. if it's fixed, use that. if it's a
1066 * range, generate a random size in-between.
1068 if (o->file_size_low == o->file_size_high)
1069 f->io_size = o->file_size_low - f->file_offset;
1071 f->io_size = get_rand_file_size(td)
1075 f->io_size = f->real_file_size - f->file_offset;
1077 if (f->io_size == -1ULL)
1080 if (o->size_percent && o->size_percent != 100) {
1083 file_size = f->io_size + f->file_offset;
1084 f->io_size = (file_size *
1085 o->size_percent) / 100;
1086 if (f->io_size > (file_size - f->file_offset))
1087 f->io_size = file_size - f->file_offset;
1089 f->io_size -= (f->io_size % td_min_bs(td));
1091 total_size += f->io_size;
1094 if (f->filetype == FIO_TYPE_FILE &&
1095 (f->io_size + f->file_offset) > f->real_file_size &&
1096 !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1097 if (!o->create_on_open) {
1099 extend_size += (f->io_size + f->file_offset);
1100 fio_file_set_extend(f);
1102 f->real_file_size = f->io_size + f->file_offset;
1106 if (td->o.block_error_hist) {
1109 assert(td->o.nr_files == 1); /* checked in fixup_options */
1111 len = f->io_size / td->o.bs[DDIR_TRIM];
1112 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1113 log_err("fio: cannot calculate block histogram with "
1114 "%d trim blocks, maximum %d\n",
1115 len, MAX_NR_BLOCK_INFOS);
1116 td_verror(td, EINVAL, "block_error_hist");
1120 td->ts.nr_block_infos = len;
1121 for (i = 0; i < len; i++)
1122 td->ts.block_infos[i] =
1123 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1125 td->ts.nr_block_infos = 0;
1127 if (!o->size || (total_size && o->size > total_size))
1128 o->size = total_size;
1130 if (o->size < td_min_bs(td)) {
1131 log_err("fio: blocksize too large for data set\n");
1136 * See if we need to extend some files, typically needed when our
1137 * target regular files don't exist yet, but our jobs require them
1138 * initially due to read I/Os.
1142 if (output_format & FIO_OUTPUT_NORMAL) {
1143 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1145 need_extend > 1 ? "s" : "",
1147 need_extend > 1 ? "s" : "",
1148 need_extend > 1 ? "total " : "",
1152 for_each_file(td, f, i) {
1153 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1155 if (!fio_file_extend(f))
1158 assert(f->filetype == FIO_TYPE_FILE);
1159 fio_file_clear_extend(f);
1160 if (!o->fill_device) {
1161 old_len = f->real_file_size;
1162 extend_len = f->io_size + f->file_offset -
1165 f->real_file_size = (f->io_size + f->file_offset);
1166 err = extend_file(td, f);
1170 err = __file_invalidate_cache(td, f, old_len,
1174 * Shut up static checker
1190 * iolog already set the total io size, if we read back
1193 if (!o->read_iolog_file) {
1195 td->total_io_size = o->io_size * o->loops;
1197 td->total_io_size = o->size * o->loops;
1204 td_restore_runstate(td, old_state);
1206 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1214 log_err("%s: you need to specify valid offset=\n", o->name);
1216 td_restore_runstate(td, old_state);
1220 bool pre_read_files(struct thread_data *td)
1225 dprint(FD_FILE, "pre_read files\n");
1227 for_each_file(td, f, i) {
1228 if (!pre_read_file(td, f))
1235 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1237 unsigned int range_size, seed;
1241 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1242 fsize = min(f->real_file_size, f->io_size);
1244 nranges = (fsize + range_size - 1ULL) / range_size;
1246 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1247 if (!td->o.rand_repeatable)
1248 seed = td->rand_seeds[4];
1250 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1251 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1252 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1253 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1254 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1255 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1258 static bool init_rand_distribution(struct thread_data *td)
1264 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1267 state = td_bump_runstate(td, TD_SETTING_UP);
1269 for_each_file(td, f, i)
1270 __init_rand_distribution(td, f);
1272 td_restore_runstate(td, state);
1277 * Check if the number of blocks exceeds the randomness capability of
1278 * the selected generator. Tausworthe is 32-bit, the others are fullly
1281 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1284 if (blocks <= FRAND32_MAX)
1286 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1290 * If the user hasn't specified a random generator, switch
1291 * to tausworthe64 with informational warning. If the user did
1292 * specify one, just warn.
1294 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1297 if (!fio_option_is_set(&td->o, random_generator)) {
1298 log_info("fio: Switching to tausworthe64. Use the "
1299 "random_generator= option to get rid of this "
1301 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1306 * Just make this information to avoid breaking scripts.
1308 log_info("fio: Use the random_generator= option to switch to lfsr or "
1313 bool init_random_map(struct thread_data *td)
1315 unsigned long long blocks;
1319 if (init_rand_distribution(td))
1324 for_each_file(td, f, i) {
1325 uint64_t fsize = min(f->real_file_size, f->io_size);
1327 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1329 if (check_rand_gen_limits(td, f, blocks))
1332 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1335 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1337 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1338 fio_file_set_lfsr(f);
1341 } else if (!td->o.norandommap) {
1342 f->io_axmap = axmap_new(blocks);
1344 fio_file_set_axmap(f);
1347 } else if (td->o.norandommap)
1350 if (!td->o.softrandommap) {
1351 log_err("fio: failed allocating random map. If running"
1352 " a large number of jobs, try the 'norandommap'"
1353 " option or set 'softrandommap'. Or give"
1354 " a larger --alloc-size to fio.\n");
1358 log_info("fio: file %s failed allocating random map. Running "
1359 "job without.\n", f->file_name);
1365 void close_files(struct thread_data *td)
1370 for_each_file(td, f, i) {
1371 if (fio_file_open(f))
1372 td_io_close_file(td, f);
1376 void close_and_free_files(struct thread_data *td)
1380 bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1382 dprint(FD_FILE, "close files\n");
1384 for_each_file(td, f, i) {
1385 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1386 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1387 td_io_unlink_file(td, f);
1390 if (fio_file_open(f))
1391 td_io_close_file(td, f);
1393 remove_file_hash(f);
1395 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1396 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1397 td_io_unlink_file(td, f);
1400 zbd_free_zone_info(f);
1405 sfree(f->file_name);
1406 f->file_name = NULL;
1407 if (fio_file_axmap(f)) {
1408 axmap_free(f->io_axmap);
1417 td->o.filename = NULL;
1419 free(td->file_locks);
1420 td->files_index = 0;
1422 td->file_locks = NULL;
1423 td->o.file_lock_mode = FILE_LOCK_NONE;
1427 static void get_file_type(struct fio_file *f)
1431 if (!strcmp(f->file_name, "-"))
1432 f->filetype = FIO_TYPE_PIPE;
1434 f->filetype = FIO_TYPE_FILE;
1437 /* \\.\ is the device namespace in Windows, where every file is
1439 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1440 f->filetype = FIO_TYPE_BLOCK;
1443 if (!stat(f->file_name, &sb)) {
1444 if (S_ISBLK(sb.st_mode))
1445 f->filetype = FIO_TYPE_BLOCK;
1446 else if (S_ISCHR(sb.st_mode))
1447 f->filetype = FIO_TYPE_CHAR;
1448 else if (S_ISFIFO(sb.st_mode))
1449 f->filetype = FIO_TYPE_PIPE;
1453 static bool __is_already_allocated(const char *fname, bool set)
1455 struct flist_head *entry;
1458 ret = file_bloom_exists(fname, set);
1462 flist_for_each(entry, &filename_list) {
1463 struct file_name *fn;
1465 fn = flist_entry(entry, struct file_name, list);
1467 if (!strcmp(fn->filename, fname))
1474 static bool is_already_allocated(const char *fname)
1478 fio_file_hash_lock();
1479 ret = __is_already_allocated(fname, false);
1480 fio_file_hash_unlock();
1485 static void set_already_allocated(const char *fname)
1487 struct file_name *fn;
1489 fn = malloc(sizeof(struct file_name));
1490 fn->filename = strdup(fname);
1492 fio_file_hash_lock();
1493 if (!__is_already_allocated(fname, true)) {
1494 flist_add_tail(&fn->list, &filename_list);
1497 fio_file_hash_unlock();
1505 static void free_already_allocated(void)
1507 struct flist_head *entry, *tmp;
1508 struct file_name *fn;
1510 if (flist_empty(&filename_list))
1513 fio_file_hash_lock();
1514 flist_for_each_safe(entry, tmp, &filename_list) {
1515 fn = flist_entry(entry, struct file_name, list);
1517 flist_del(&fn->list);
1521 fio_file_hash_unlock();
1524 static struct fio_file *alloc_new_file(struct thread_data *td)
1528 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1529 f = calloc(1, sizeof(*f));
1531 f = scalloc(1, sizeof(*f));
1539 fio_file_reset(td, f);
1543 bool exists_and_not_regfile(const char *filename)
1547 if (lstat(filename, &sb) == -1)
1550 #ifndef WIN32 /* NOT Windows */
1551 if (S_ISREG(sb.st_mode))
1554 /* \\.\ is the device namespace in Windows, where every file
1555 * is a device node */
1556 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1563 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1565 int cur_files = td->files_index;
1566 char file_name[PATH_MAX];
1570 dprint(FD_FILE, "add file %s\n", fname);
1572 if (td->o.directory)
1573 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1574 td->o.unique_filename);
1576 sprintf(file_name + len, "%s", fname);
1578 /* clean cloned siblings using existing files */
1579 if (numjob && is_already_allocated(file_name) &&
1580 !exists_and_not_regfile(fname))
1583 f = alloc_new_file(td);
1585 if (td->files_size <= td->files_index) {
1586 unsigned int new_size = td->o.nr_files + 1;
1588 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1590 td->files = realloc(td->files, new_size * sizeof(f));
1591 if (td->files == NULL) {
1592 log_err("fio: realloc OOM\n");
1595 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1596 td->file_locks = realloc(td->file_locks, new_size);
1597 if (!td->file_locks) {
1598 log_err("fio: realloc OOM\n");
1601 td->file_locks[cur_files] = FILE_LOCK_NONE;
1603 td->files_size = new_size;
1605 td->files[cur_files] = f;
1606 f->fileno = cur_files;
1609 * init function, io engine may not be loaded yet
1611 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1612 f->real_file_size = -1ULL;
1614 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1615 f->file_name = strdup(file_name);
1617 f->file_name = smalloc_strdup(file_name);
1619 /* can't handle smalloc failure from here */
1620 assert(f->file_name);
1624 switch (td->o.file_lock_mode) {
1625 case FILE_LOCK_NONE:
1627 case FILE_LOCK_READWRITE:
1628 f->rwlock = fio_rwlock_init();
1630 case FILE_LOCK_EXCLUSIVE:
1631 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1634 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
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 disk_util_dec(f->du);
1689 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1690 unlock_file_all(td, f);
1692 if (should_fsync(td) && td->o.fsync_on_close) {
1693 f_ret = fsync(f->fd);
1698 if (td->io_ops->close_file)
1699 ret = td->io_ops->close_file(td, f);
1704 td->nr_open_files--;
1705 fio_file_clear_closing(f);
1706 fio_file_clear_open(f);
1707 assert(f->fd == -1);
1711 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1713 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1716 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1717 if (ddir == DDIR_READ)
1718 fio_rwlock_read(f->rwlock);
1720 fio_rwlock_write(f->rwlock);
1721 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1722 fio_sem_down(f->lock);
1724 td->file_locks[f->fileno] = td->o.file_lock_mode;
1727 void unlock_file(struct thread_data *td, struct fio_file *f)
1729 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1732 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1733 fio_rwlock_unlock(f->rwlock);
1734 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1735 fio_sem_up(f->lock);
1737 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1740 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1742 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1744 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1748 static bool recurse_dir(struct thread_data *td, const char *dirname)
1754 D = opendir(dirname);
1756 char buf[FIO_VERROR_SIZE];
1758 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1759 td_verror(td, errno, buf);
1763 while ((dir = readdir(D)) != NULL) {
1764 char full_path[PATH_MAX];
1767 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1770 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1772 if (lstat(full_path, &sb) == -1) {
1773 if (errno != ENOENT) {
1774 td_verror(td, errno, "stat");
1780 if (S_ISREG(sb.st_mode)) {
1781 add_file(td, full_path, 0, 1);
1784 if (!S_ISDIR(sb.st_mode))
1787 ret = recurse_dir(td, full_path);
1796 int add_dir_files(struct thread_data *td, const char *path)
1798 int ret = recurse_dir(td, path);
1801 log_info("fio: opendir added %d files\n", td->o.nr_files);
1806 void dup_files(struct thread_data *td, struct thread_data *org)
1811 dprint(FD_FILE, "dup files: %d\n", org->files_index);
1816 td->files = malloc(org->files_index * sizeof(f));
1818 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1819 td->file_locks = malloc(org->files_index);
1821 for_each_file(org, f, i) {
1822 struct fio_file *__f;
1824 __f = alloc_new_file(td);
1827 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1828 __f->file_name = strdup(f->file_name);
1830 __f->file_name = smalloc_strdup(f->file_name);
1832 /* can't handle smalloc failure from here */
1833 assert(__f->file_name);
1834 __f->filetype = f->filetype;
1837 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1838 __f->lock = f->lock;
1839 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1840 __f->rwlock = f->rwlock;
1847 * Returns the index that matches the filename, or -1 if not there
1849 int get_fileno(struct thread_data *td, const char *fname)
1854 for_each_file(td, f, i)
1855 if (!strcmp(f->file_name, fname))
1862 * For log usage, where we add/open/close files automatically
1864 void free_release_files(struct thread_data *td)
1868 td->o.open_files = 0;
1869 td->files_index = 0;
1872 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1876 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1877 f->last_pos[i] = f->file_offset;
1878 f->last_start[i] = -1ULL;
1881 if (fio_file_axmap(f))
1882 axmap_reset(f->io_axmap);
1883 else if (fio_file_lfsr(f))
1884 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1886 zbd_file_reset(td, f);
1889 bool fio_files_done(struct thread_data *td)
1894 for_each_file(td, f, i)
1895 if (!fio_file_done(f))
1901 /* free memory used in initialization phase only */
1902 void filesetup_mem_free(void)
1904 free_already_allocated();
1908 * This function is for platforms which support direct I/O but not O_DIRECT.
1910 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1912 #ifdef FIO_OS_DIRECTIO
1913 int ret = fio_set_odirect(f);
1916 td_verror(td, ret, "fio_set_directio");
1917 #if defined(__sun__)
1918 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1919 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1921 log_err("fio: the file system does not seem to support direct IO\n");
1924 log_err("fio: the file system does not seem to support direct IO\n");
1931 log_err("fio: direct IO is not supported on this host operating system\n");