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 */
98 case FIO_FALLOCATE_TRUNCATE: {
101 dprint(FD_FILE, "ftruncate file %s size %llu\n",
103 (unsigned long long) f->real_file_size);
104 r = ftruncate(f->fd, f->real_file_size);
106 td_verror(td, errno, "ftruncate");
111 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
117 * Leaves f->fd open on success, caller must close
119 static int extend_file(struct thread_data *td, struct fio_file *f)
121 int new_layout = 0, unlink_file = 0, flags;
122 unsigned long long left;
123 unsigned long long bs;
127 log_err("fio: refusing extend of file due to read-only\n");
132 * check if we need to lay the file out complete again. fio
133 * does that for operations involving reads, or for writes
134 * where overwrite is set
137 (td_write(td) && td->o.overwrite && !td->o.file_append) ||
138 (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
140 if (td_write(td) && !td->o.overwrite && !td->o.file_append)
143 if (unlink_file || new_layout) {
146 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
148 ret = td_io_unlink_file(td, f);
149 if (ret != 0 && ret != ENOENT) {
150 td_verror(td, errno, "unlink");
156 if (td->o.allow_create)
165 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
166 f->fd = open(f->file_name, flags, 0644);
170 if (err == ENOENT && !td->o.allow_create)
171 log_err("fio: file creation disallowed by "
172 "allow_file_create=0\n");
174 td_verror(td, err, "open");
178 fallocate_file(td, f);
181 * If our jobs don't require regular files initially, we're done.
187 * The size will be -1ULL when fill_device is used, so don't truncate
188 * or fallocate this file, just write it
190 if (!td->o.fill_device) {
191 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
192 (unsigned long long) f->real_file_size);
193 if (ftruncate(f->fd, f->real_file_size) == -1) {
194 if (errno != EFBIG) {
195 td_verror(td, errno, "ftruncate");
201 left = f->real_file_size;
202 bs = td->o.max_bs[DDIR_WRITE];
208 td_verror(td, errno, "malloc");
212 while (left && !td->terminate) {
218 fill_io_buffer(td, b, bs, bs);
220 r = write(f->fd, b, bs);
229 if (__e == ENOSPC || __e == EDQUOT) {
230 const char *__e_name;
231 if (td->o.fill_device)
237 log_info("fio: %s on laying out "
238 "file, stopping\n", __e_name);
240 td_verror(td, errno, "write");
242 td_verror(td, EIO, "write");
249 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
250 td_io_unlink_file(td, f);
251 } else if (td->o.create_fsync) {
252 if (fsync(f->fd) < 0) {
253 td_verror(td, errno, "fsync");
257 if (td->o.fill_device && !td_write(td)) {
258 fio_file_clear_size_known(f);
259 if (td_io_get_file_size(td, f))
261 if (f->io_size > f->real_file_size)
262 f->io_size = f->real_file_size;
276 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
278 int r, did_open = 0, old_runstate;
279 unsigned long long left;
280 unsigned long long bs;
284 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
285 td_ioengine_flagged(td, FIO_NOIO))
288 if (f->filetype == FIO_TYPE_CHAR)
291 if (!fio_file_open(f)) {
292 if (td->io_ops->open_file(td, f)) {
293 log_err("fio: cannot pre-read, failed to open file\n");
299 old_runstate = td_bump_runstate(td, TD_PRE_READING);
302 bs = td->o.max_bs[DDIR_READ];
308 td_verror(td, errno, "malloc");
314 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
315 td_verror(td, errno, "lseek");
316 log_err("fio: failed to lseek pre-read file\n");
321 while (left && !td->terminate) {
325 r = read(f->fd, b, bs);
331 td_verror(td, EIO, "pre_read");
337 td_restore_runstate(td, old_runstate);
340 td->io_ops->close_file(td, f);
347 * Generic function to prepopulate regular file with data.
348 * Useful if you want to make sure I/O engine has data to read.
349 * Leaves f->fd open on success, caller must close.
351 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
354 unsigned long long left, bs;
357 /* generic function for regular files only */
358 assert(f->filetype == FIO_TYPE_FILE);
361 log_err("fio: refusing to write a file due to read-only\n");
366 if (td->o.allow_create)
373 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
374 f->fd = open(f->file_name, flags, 0644);
378 if (err == ENOENT && !td->o.allow_create)
379 log_err("fio: file creation disallowed by "
380 "allow_file_create=0\n");
382 td_verror(td, err, "open");
386 left = f->real_file_size;
387 bs = td->o.max_bs[DDIR_WRITE];
393 td_verror(td, errno, "malloc");
397 while (left && !td->terminate) {
403 fill_io_buffer(td, b, bs, bs);
405 r = write(f->fd, b, bs);
410 td_verror(td, errno, "write");
416 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
417 td_io_unlink_file(td, f);
418 } else if (td->o.create_fsync) {
419 if (fsync(f->fd) < 0) {
420 td_verror(td, errno, "fsync");
435 unsigned long long get_rand_file_size(struct thread_data *td)
437 unsigned long long ret, sized;
441 frand_max = rand_max(&td->file_size_state);
442 r = __rand(&td->file_size_state);
443 sized = td->o.file_size_high - td->o.file_size_low;
444 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
445 ret += td->o.file_size_low;
446 ret -= (ret % td->o.rw_min_bs);
450 static int file_size(struct thread_data *td, struct fio_file *f)
454 if (stat(f->file_name, &st) == -1) {
455 td_verror(td, errno, "fstat");
459 f->real_file_size = st.st_size;
463 static int bdev_size(struct thread_data *td, struct fio_file *f)
465 unsigned long long bytes = 0;
468 if (td->io_ops->open_file(td, f)) {
469 log_err("fio: failed opening blockdev %s for size check\n",
474 r = blockdev_size(f, &bytes);
476 td_verror(td, r, "blockdev_size");
481 log_err("%s: zero sized block device?\n", f->file_name);
485 f->real_file_size = bytes;
486 td->io_ops->close_file(td, f);
489 td->io_ops->close_file(td, f);
493 static int char_size(struct thread_data *td, struct fio_file *f)
495 #ifdef FIO_HAVE_CHARDEV_SIZE
496 unsigned long long bytes = 0;
499 if (td->io_ops->open_file(td, f)) {
500 log_err("fio: failed opening chardev %s for size check\n",
505 r = chardev_size(f, &bytes);
507 td_verror(td, r, "chardev_size");
512 log_err("%s: zero sized char device?\n", f->file_name);
516 f->real_file_size = bytes;
517 td->io_ops->close_file(td, f);
520 td->io_ops->close_file(td, f);
523 f->real_file_size = -1ULL;
528 static int get_file_size(struct thread_data *td, struct fio_file *f)
532 if (fio_file_size_known(f))
535 if (f->filetype == FIO_TYPE_FILE)
536 ret = file_size(td, f);
537 else if (f->filetype == FIO_TYPE_BLOCK)
538 ret = bdev_size(td, f);
539 else if (f->filetype == FIO_TYPE_CHAR)
540 ret = char_size(td, f);
542 f->real_file_size = -1;
543 log_info("%s: failed to get file size of %s\n", td->o.name,
545 return 1; /* avoid offset extends end error message */
549 * Leave ->real_file_size with 0 since it could be expectation
550 * of initial setup for regular files.
556 * ->file_offset normally hasn't been initialized yet, so this
557 * is basically always false unless ->real_file_size is -1, but
558 * if ->real_file_size is -1 this message doesn't make sense.
559 * As a result, this message is basically useless.
561 if (f->file_offset > f->real_file_size) {
562 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
563 (unsigned long long) f->file_offset,
564 (unsigned long long) f->real_file_size);
568 fio_file_set_size_known(f);
572 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
573 unsigned long long off,
574 unsigned long long len)
576 int errval = 0, ret = 0;
585 off = f->file_offset;
587 if (len == -1ULL || off == -1ULL)
590 if (td->io_ops->invalidate) {
591 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
593 ret = td->io_ops->invalidate(td, f);
596 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
597 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
599 } else if (f->filetype == FIO_TYPE_FILE) {
600 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
601 f->file_name, off, len);
602 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
605 } else if (f->filetype == FIO_TYPE_BLOCK) {
608 dprint(FD_IO, "drop page cache %s\n", f->file_name);
609 ret = blockdev_invalidate_cache(f);
610 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
612 * Linux multipath devices reject ioctl while
613 * the maps are being updated. That window can
614 * last tens of milliseconds; we'll try up to
615 * a quarter of a second.
618 ret = blockdev_invalidate_cache(f);
620 if (ret < 0 && errno == EACCES && geteuid()) {
621 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
622 log_err("fio: only root may flush block "
623 "devices. Cache flush bypassed!\n");
628 } else if (f->filetype == FIO_TYPE_CHAR ||
629 f->filetype == FIO_TYPE_PIPE) {
630 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
634 * Cache flushing isn't a fatal condition, and we know it will
635 * happen on some platforms where we don't have the proper
636 * function to flush eg block device caches. So just warn and
637 * continue on our way.
640 log_info("fio: cache invalidation of %s failed: %s\n",
641 f->file_name, strerror(errval));
647 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
649 if (!fio_file_open(f))
652 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
655 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
659 dprint(FD_FILE, "fd close %s\n", f->file_name);
663 if (close(f->fd) < 0)
668 if (f->shadow_fd != -1) {
677 int file_lookup_open(struct fio_file *f, int flags)
679 struct fio_file *__f;
682 __f = lookup_file_hash(f->file_name);
684 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
688 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
696 f->fd = open(f->file_name, flags, 0600);
700 static int file_close_shadow_fds(struct thread_data *td)
706 for_each_file(td, f, i) {
707 if (f->shadow_fd == -1)
718 int generic_open_file(struct thread_data *td, struct fio_file *f)
724 dprint(FD_FILE, "fd open %s\n", f->file_name);
726 if (!strcmp(f->file_name, "-")) {
728 log_err("fio: can't read/write to stdin/out\n");
734 * move output logging to stderr, if we are writing to stdout
743 flags |= OS_O_DIRECT;
744 flags |= td->o.sync_io;
745 if (td->o.create_on_open && td->o.allow_create)
748 if (f->filetype != FIO_TYPE_FILE)
749 flags |= FIO_O_NOATIME;
756 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
760 f->fd = dup(STDOUT_FILENO);
762 from_hash = file_lookup_open(f, flags);
763 } else if (td_read(td)) {
764 if (td_ioengine_flagged(td, FIO_RO_NEEDS_RW_OPEN) && !read_only)
770 f->fd = dup(STDIN_FILENO);
772 from_hash = file_lookup_open(f, flags);
773 } else if (td_trim(td)) {
774 assert(!td_rw(td)); /* should have matched above */
777 from_hash = file_lookup_open(f, flags);
781 char buf[FIO_VERROR_SIZE];
784 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
785 flags &= ~FIO_O_NOATIME;
788 if (__e == EMFILE && file_close_shadow_fds(td))
791 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
793 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
794 log_err("fio: looks like your file system does not " \
795 "support direct=1/buffered=0\n");
798 td_verror(td, __e, buf);
802 if (!from_hash && f->fd != -1) {
803 if (add_file_hash(f)) {
807 * Stash away descriptor for later close. This is to
808 * work-around a "feature" on Linux, where a close of
809 * an fd that has been opened for write will trigger
810 * udev to call blkid to check partitions, fs id, etc.
811 * That pollutes the device cache, which can slow down
812 * unbuffered accesses.
814 if (f->shadow_fd == -1)
815 f->shadow_fd = f->fd;
818 * OK to ignore, we haven't done anything
821 ret = generic_close_file(td, f);
831 * This function i.e. get_file_size() is the default .get_file_size
832 * implementation of majority of I/O engines.
834 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
836 return get_file_size(td, f);
840 * open/close all files, so that ->real_file_size gets set
842 static int get_file_sizes(struct thread_data *td)
848 for_each_file(td, f, i) {
849 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
852 if (td_io_get_file_size(td, f)) {
853 if (td->error != ENOENT) {
854 log_err("%s\n", td->verror);
862 * There are corner cases where we end up with -1 for
863 * ->real_file_size due to unsupported file type, etc.
864 * We then just set to size option value divided by number
865 * of files, similar to the way file ->io_size is set.
866 * stat(2) failure doesn't set ->real_file_size to -1.
868 if (f->real_file_size == -1ULL && td->o.size)
869 f->real_file_size = td->o.size / td->o.nr_files;
876 struct flist_head list;
883 * Get free number of bytes for each file on each unique mount.
885 static unsigned long long get_fs_free_counts(struct thread_data *td)
887 struct flist_head *n, *tmp;
888 unsigned long long ret = 0;
889 struct fio_mount *fm;
894 for_each_file(td, f, i) {
898 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
899 if (f->real_file_size != -1ULL)
900 ret += f->real_file_size;
902 } else if (f->filetype != FIO_TYPE_FILE)
905 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
907 if (stat(buf, &sb) < 0) {
911 if (stat(buf, &sb) < 0)
916 flist_for_each(n, &list) {
917 fm = flist_entry(n, struct fio_mount, list);
918 if (fm->key == sb.st_dev)
927 fm = calloc(1, sizeof(*fm));
928 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
929 fm->base = basename(fm->__base);
931 flist_add(&fm->list, &list);
934 flist_for_each_safe(n, tmp, &list) {
935 unsigned long long sz;
937 fm = flist_entry(n, struct fio_mount, list);
938 flist_del(&fm->list);
940 sz = get_fs_free_size(fm->base);
941 if (sz && sz != -1ULL)
950 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
953 struct thread_options *o = &td->o;
954 unsigned long long align_bs;
955 unsigned long long offset;
956 unsigned long long increment;
958 if (o->file_append && f->filetype == FIO_TYPE_FILE)
959 return f->real_file_size;
961 if (o->offset_increment_percent) {
962 assert(!o->offset_increment);
963 increment = o->offset_increment_percent * f->real_file_size / 100;
966 increment = o->offset_increment;
968 if (o->start_offset_percent > 0) {
969 /* calculate the raw offset */
970 offset = (f->real_file_size * o->start_offset_percent / 100) +
971 (td->subjob_number * increment);
975 /* start_offset_percent not set */
976 offset = o->start_offset +
977 td->subjob_number * increment;
982 * if offset_align is provided, use it
984 if (fio_option_is_set(o, start_offset_align)) {
985 align_bs = o->start_offset_align;
987 /* else take the minimum block size */
988 align_bs = td_min_bs(td);
992 * block align the offset at the next available boundary at
993 * ceiling(offset / align_bs) * align_bs
995 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
1002 * Find longest path component that exists and return its length
1004 int longest_existing_path(char *path) {
1015 sprintf(buf, "%s", path);
1018 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1024 *(buf_pos + 1) = '\0';
1027 dwAttr = GetFileAttributesA(buf);
1028 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1032 if (stat(buf, &sb) == 0)
1036 offset = buf_pos - buf;
1044 static bool create_work_dirs(struct thread_data *td, const char *fname)
1046 char path[PATH_MAX];
1050 snprintf(path, PATH_MAX, "%s", fname);
1053 offset = longest_existing_path(path);
1054 end = start + offset;
1055 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1062 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1063 log_err("fio: failed to create dir (%s): %s\n",
1064 start, strerror(errno));
1067 *end = FIO_OS_PATH_SEPARATOR;
1070 td->flags |= TD_F_DIRS_CREATED;
1075 * Open the files and setup files sizes, creating files if necessary.
1077 int setup_files(struct thread_data *td)
1079 unsigned long long total_size, extend_size;
1080 struct thread_options *o = &td->o;
1082 unsigned int i, nr_fs_extra = 0;
1083 int err = 0, need_extend;
1085 const unsigned long long bs = td_min_bs(td);
1088 dprint(FD_FILE, "setup files\n");
1090 old_state = td_bump_runstate(td, TD_SETTING_UP);
1092 for_each_file(td, f, i) {
1093 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1094 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1095 !(td->flags & TD_F_DIRS_CREATED) &&
1096 !create_work_dirs(td, f->file_name))
1101 * Find out physical size of files or devices for this thread,
1102 * before we determine I/O size and range of our targets.
1103 * If ioengine defines a setup() method, it's responsible for
1104 * opening the files and setting f->real_file_size to indicate
1105 * the valid range for that file.
1107 if (td->io_ops->setup)
1108 err = td->io_ops->setup(td);
1110 err = get_file_sizes(td);
1115 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1116 err = zbd_init_files(td);
1120 zbd_recalc_options_with_zone_granularity(td);
1122 if (o->read_iolog_file)
1126 * check sizes. if the files/devices do not exist and the size
1127 * isn't passed to fio, abort.
1130 for_each_file(td, f, i) {
1132 if (f->real_file_size == -1ULL)
1135 total_size += f->real_file_size;
1139 td->fill_device_size = get_fs_free_counts(td);
1142 * device/file sizes are zero and no size given, punt
1144 if ((!total_size || total_size == -1ULL) && !o->size &&
1145 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1146 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1147 log_err("%s: you need to specify size=\n", o->name);
1148 td_verror(td, EINVAL, "total_file_size");
1153 * Calculate per-file size and potential extra size for the
1154 * first files, if needed (i.e. if we don't have a fixed size).
1156 if (!o->file_size_low && o->nr_files) {
1159 fs = o->size / o->nr_files;
1160 all_fs = fs * o->nr_files;
1162 if (all_fs < o->size)
1163 nr_fs_extra = (o->size - all_fs) / bs;
1167 * now file sizes are known, so we can set ->io_size. if size= is
1168 * not given, ->io_size is just equal to ->real_file_size. if size
1169 * is given, ->io_size is size / nr_files.
1171 extend_size = total_size = 0;
1173 for_each_file(td, f, i) {
1174 f->file_offset = get_start_offset(td, f);
1177 * Update ->io_size depending on options specified.
1178 * ->file_size_low being 0 means filesize option isn't set.
1179 * Non zero ->file_size_low equals ->file_size_high means
1180 * filesize option is set in a fixed size format.
1181 * Non zero ->file_size_low not equals ->file_size_high means
1182 * filesize option is set in a range format.
1184 if (!o->file_size_low) {
1186 * no file size or range given, file size is equal to
1187 * total size divided by number of files. If the size
1188 * doesn't divide nicely with the min blocksize,
1189 * make the first files bigger.
1198 * We normally don't come here for regular files, but
1199 * if the result is 0 for a regular file, set it to the
1200 * real file size. This could be size of the existing
1201 * one if it already exists, but otherwise will be set
1202 * to 0. A new file won't be created because
1203 * ->io_size + ->file_offset equals ->real_file_size.
1206 if (f->file_offset > f->real_file_size)
1208 f->io_size = f->real_file_size - f->file_offset;
1210 log_info("fio: file %s may be ignored\n",
1213 } else if (f->real_file_size < o->file_size_low ||
1214 f->real_file_size > o->file_size_high) {
1215 if (f->file_offset > o->file_size_low)
1218 * file size given. if it's fixed, use that. if it's a
1219 * range, generate a random size in-between.
1221 if (o->file_size_low == o->file_size_high)
1222 f->io_size = o->file_size_low - f->file_offset;
1224 f->io_size = get_rand_file_size(td)
1228 f->io_size = f->real_file_size - f->file_offset;
1230 if (f->io_size == -1ULL)
1235 if (o->size_percent && o->size_percent != 100) {
1238 file_size = f->io_size + f->file_offset;
1239 f->io_size = (file_size *
1240 o->size_percent) / 100;
1241 if (f->io_size > (file_size - f->file_offset))
1242 f->io_size = file_size - f->file_offset;
1244 f->io_size -= (f->io_size % td_min_bs(td));
1247 io_size = f->io_size;
1248 if (o->io_size_percent && o->io_size_percent != 100) {
1249 io_size *= o->io_size_percent;
1253 total_size += io_size;
1256 if (f->filetype == FIO_TYPE_FILE &&
1257 (f->io_size + f->file_offset) > f->real_file_size) {
1258 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1259 !o->create_on_open) {
1261 extend_size += (f->io_size + f->file_offset);
1262 fio_file_set_extend(f);
1263 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1264 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1265 td_ioengine_flagged(td, FIO_FAKEIO)))
1266 f->real_file_size = f->io_size + f->file_offset;
1270 if (td->o.block_error_hist) {
1273 assert(td->o.nr_files == 1); /* checked in fixup_options */
1275 len = f->io_size / td->o.bs[DDIR_TRIM];
1276 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1277 log_err("fio: cannot calculate block histogram with "
1278 "%d trim blocks, maximum %d\n",
1279 len, MAX_NR_BLOCK_INFOS);
1280 td_verror(td, EINVAL, "block_error_hist");
1284 td->ts.nr_block_infos = len;
1285 for (i = 0; i < len; i++)
1286 td->ts.block_infos[i] =
1287 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1289 td->ts.nr_block_infos = 0;
1291 if (!o->size || (total_size && o->size > total_size))
1292 o->size = total_size;
1294 if (o->size < td_min_bs(td)) {
1295 log_err("fio: blocksize is larger than data set range\n");
1300 * See if we need to extend some files, typically needed when our
1301 * target regular files don't exist yet, but our jobs require them
1302 * initially due to read I/Os.
1306 if (output_format & FIO_OUTPUT_NORMAL) {
1307 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1309 need_extend > 1 ? "s" : "",
1311 need_extend > 1 ? "s" : "",
1312 need_extend > 1 ? "total " : "",
1316 for_each_file(td, f, i) {
1317 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1319 if (!fio_file_extend(f))
1322 assert(f->filetype == FIO_TYPE_FILE);
1323 fio_file_clear_extend(f);
1324 if (!o->fill_device) {
1325 old_len = f->real_file_size;
1326 extend_len = f->io_size + f->file_offset -
1329 f->real_file_size = (f->io_size + f->file_offset);
1330 err = extend_file(td, f);
1334 err = __file_invalidate_cache(td, f, old_len,
1338 * Shut up static checker
1354 * Prepopulate files with data. It might be expected to read some
1355 * "real" data instead of zero'ed files (if no writes to file occurred
1356 * prior to a read job). Engine has to provide a way to do that.
1358 if (td->io_ops->prepopulate_file) {
1361 for_each_file(td, f, i) {
1362 if (output_format & FIO_OUTPUT_NORMAL) {
1363 log_info("%s: Prepopulating IO file (%s)\n",
1364 o->name, f->file_name);
1367 err = td->io_ops->prepopulate_file(td, f);
1371 err = __file_invalidate_cache(td, f, f->file_offset,
1375 * Shut up static checker
1391 * iolog already set the total io size, if we read back
1394 if (!o->read_iolog_file) {
1396 td->total_io_size = o->io_size * o->loops;
1398 td->total_io_size = o->size * o->loops;
1402 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1403 err = zbd_setup_files(td);
1411 td_restore_runstate(td, old_state);
1416 log_err("%s: you need to specify valid offset=\n", o->name);
1418 td_restore_runstate(td, old_state);
1422 bool pre_read_files(struct thread_data *td)
1427 dprint(FD_FILE, "pre_read files\n");
1429 for_each_file(td, f, i) {
1430 if (!pre_read_file(td, f))
1437 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1439 unsigned int range_size, seed;
1443 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1444 fsize = min(f->real_file_size, f->io_size);
1446 nranges = (fsize + range_size - 1ULL) / range_size;
1448 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1449 if (!td->o.rand_repeatable)
1450 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1452 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1453 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1454 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1455 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1456 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1457 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1460 static bool init_rand_distribution(struct thread_data *td)
1466 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1467 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1468 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1471 state = td_bump_runstate(td, TD_SETTING_UP);
1473 for_each_file(td, f, i)
1474 __init_rand_distribution(td, f);
1476 td_restore_runstate(td, state);
1481 * Check if the number of blocks exceeds the randomness capability of
1482 * the selected generator. Tausworthe is 32-bit, the others are fully
1485 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1488 if (blocks <= FRAND32_MAX)
1490 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1494 * If the user hasn't specified a random generator, switch
1495 * to tausworthe64 with informational warning. If the user did
1496 * specify one, just warn.
1498 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1501 if (!fio_option_is_set(&td->o, random_generator)) {
1502 log_info("fio: Switching to tausworthe64. Use the "
1503 "random_generator= option to get rid of this "
1505 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1510 * Just make this information to avoid breaking scripts.
1512 log_info("fio: Use the random_generator= option to switch to lfsr or "
1517 bool init_random_map(struct thread_data *td)
1519 unsigned long long blocks;
1523 if (init_rand_distribution(td))
1528 for_each_file(td, f, i) {
1529 uint64_t fsize = min(f->real_file_size, f->io_size);
1531 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1532 fsize = td->o.zone_range;
1534 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1536 if (check_rand_gen_limits(td, f, blocks))
1539 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1542 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1544 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1545 fio_file_set_lfsr(f);
1548 log_err("fio: failed initializing LFSR\n");
1551 } else if (!td->o.norandommap) {
1552 f->io_axmap = axmap_new(blocks);
1554 fio_file_set_axmap(f);
1557 } else if (td->o.norandommap)
1560 if (!td->o.softrandommap) {
1561 log_err("fio: failed allocating random map. If running"
1562 " a large number of jobs, try the 'norandommap'"
1563 " option or set 'softrandommap'. Or give"
1564 " a larger --alloc-size to fio.\n");
1568 log_info("fio: file %s failed allocating random map. Running "
1569 "job without.\n", f->file_name);
1575 void close_files(struct thread_data *td)
1580 for_each_file(td, f, i) {
1581 if (fio_file_open(f))
1582 td_io_close_file(td, f);
1586 void fio_file_free(struct fio_file *f)
1588 if (fio_file_axmap(f))
1589 axmap_free(f->io_axmap);
1590 if (!fio_file_smalloc(f)) {
1594 sfree(f->file_name);
1599 void close_and_free_files(struct thread_data *td)
1604 dprint(FD_FILE, "close files\n");
1606 for_each_file(td, f, i) {
1607 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1608 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1609 td_io_unlink_file(td, f);
1612 if (fio_file_open(f))
1613 td_io_close_file(td, f);
1615 remove_file_hash(f);
1617 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1618 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1619 td_io_unlink_file(td, f);
1626 td->o.filename = NULL;
1628 free(td->file_locks);
1629 td->files_index = 0;
1631 td->file_locks = NULL;
1632 td->o.file_lock_mode = FILE_LOCK_NONE;
1636 static void get_file_type(struct fio_file *f)
1640 if (!strcmp(f->file_name, "-"))
1641 f->filetype = FIO_TYPE_PIPE;
1643 f->filetype = FIO_TYPE_FILE;
1646 /* \\.\ is the device namespace in Windows, where every file is
1648 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1649 f->filetype = FIO_TYPE_BLOCK;
1652 if (!stat(f->file_name, &sb)) {
1653 if (S_ISBLK(sb.st_mode))
1654 f->filetype = FIO_TYPE_BLOCK;
1655 else if (S_ISCHR(sb.st_mode))
1656 f->filetype = FIO_TYPE_CHAR;
1657 else if (S_ISFIFO(sb.st_mode))
1658 f->filetype = FIO_TYPE_PIPE;
1662 static bool __is_already_allocated(const char *fname, bool set)
1664 struct flist_head *entry;
1667 ret = file_bloom_exists(fname, set);
1671 flist_for_each(entry, &filename_list) {
1672 struct file_name *fn;
1674 fn = flist_entry(entry, struct file_name, list);
1676 if (!strcmp(fn->filename, fname))
1683 static bool is_already_allocated(const char *fname)
1687 fio_file_hash_lock();
1688 ret = __is_already_allocated(fname, false);
1689 fio_file_hash_unlock();
1694 static void set_already_allocated(const char *fname)
1696 struct file_name *fn;
1698 fn = malloc(sizeof(struct file_name));
1699 fn->filename = strdup(fname);
1701 fio_file_hash_lock();
1702 if (!__is_already_allocated(fname, true)) {
1703 flist_add_tail(&fn->list, &filename_list);
1706 fio_file_hash_unlock();
1714 static void free_already_allocated(void)
1716 struct flist_head *entry, *tmp;
1717 struct file_name *fn;
1719 if (flist_empty(&filename_list))
1722 fio_file_hash_lock();
1723 flist_for_each_safe(entry, tmp, &filename_list) {
1724 fn = flist_entry(entry, struct file_name, list);
1726 flist_del(&fn->list);
1730 fio_file_hash_unlock();
1733 static struct fio_file *alloc_new_file(struct thread_data *td)
1737 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1738 f = calloc(1, sizeof(*f));
1740 f = scalloc(1, sizeof(*f));
1748 fio_file_reset(td, f);
1749 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1750 fio_file_set_smalloc(f);
1754 bool exists_and_not_regfile(const char *filename)
1758 if (lstat(filename, &sb) == -1)
1761 #ifndef WIN32 /* NOT Windows */
1762 if (S_ISREG(sb.st_mode))
1765 /* \\.\ is the device namespace in Windows, where every file
1766 * is a device node */
1767 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1774 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1776 int cur_files = td->files_index;
1777 char file_name[PATH_MAX];
1781 dprint(FD_FILE, "add file %s\n", fname);
1783 if (td->o.directory)
1784 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1785 td->o.unique_filename);
1787 sprintf(file_name + len, "%s", fname);
1789 /* clean cloned siblings using existing files */
1790 if (numjob && is_already_allocated(file_name) &&
1791 !exists_and_not_regfile(fname))
1794 f = alloc_new_file(td);
1796 if (td->files_size <= td->files_index) {
1797 unsigned int new_size = td->o.nr_files + 1;
1799 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1801 td->files = realloc(td->files, new_size * sizeof(f));
1802 if (td->files == NULL) {
1803 log_err("fio: realloc OOM\n");
1806 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1807 td->file_locks = realloc(td->file_locks, new_size);
1808 if (!td->file_locks) {
1809 log_err("fio: realloc OOM\n");
1812 td->file_locks[cur_files] = FILE_LOCK_NONE;
1814 td->files_size = new_size;
1816 td->files[cur_files] = f;
1817 f->fileno = cur_files;
1820 * init function, io engine may not be loaded yet
1822 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1823 f->real_file_size = -1ULL;
1825 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1826 f->file_name = strdup(file_name);
1828 f->file_name = smalloc_strdup(file_name);
1830 /* can't handle smalloc failure from here */
1831 assert(f->file_name);
1835 switch (td->o.file_lock_mode) {
1836 case FILE_LOCK_NONE:
1838 case FILE_LOCK_READWRITE:
1839 f->rwlock = fio_rwlock_init();
1841 case FILE_LOCK_EXCLUSIVE:
1842 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1845 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1851 if (td->o.numjobs > 1)
1852 set_already_allocated(file_name);
1857 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1863 int add_file_exclusive(struct thread_data *td, const char *fname)
1868 for_each_file(td, f, i) {
1869 if (!strcmp(f->file_name, fname))
1873 return add_file(td, fname, 0, 1);
1876 void get_file(struct fio_file *f)
1878 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1879 assert(fio_file_open(f));
1883 int put_file(struct thread_data *td, struct fio_file *f)
1885 int f_ret = 0, ret = 0;
1887 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1889 if (!fio_file_open(f)) {
1890 assert(f->fd == -1);
1894 assert(f->references);
1895 if (--f->references)
1898 disk_util_dec(f->du);
1900 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1901 unlock_file_all(td, f);
1903 if (should_fsync(td) && td->o.fsync_on_close) {
1904 f_ret = fsync(f->fd);
1909 if (td->io_ops->close_file)
1910 ret = td->io_ops->close_file(td, f);
1915 td->nr_open_files--;
1916 fio_file_clear_closing(f);
1917 fio_file_clear_open(f);
1918 assert(f->fd == -1);
1922 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1924 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1927 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1928 if (ddir == DDIR_READ)
1929 fio_rwlock_read(f->rwlock);
1931 fio_rwlock_write(f->rwlock);
1932 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1933 fio_sem_down(f->lock);
1935 td->file_locks[f->fileno] = td->o.file_lock_mode;
1938 void unlock_file(struct thread_data *td, struct fio_file *f)
1940 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1943 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1944 fio_rwlock_unlock(f->rwlock);
1945 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1946 fio_sem_up(f->lock);
1948 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1951 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1953 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1955 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1959 static bool recurse_dir(struct thread_data *td, const char *dirname)
1965 D = opendir(dirname);
1967 char buf[FIO_VERROR_SIZE];
1969 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1970 td_verror(td, errno, buf);
1974 while ((dir = readdir(D)) != NULL) {
1975 char full_path[PATH_MAX];
1978 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1981 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1983 if (lstat(full_path, &sb) == -1) {
1984 if (errno != ENOENT) {
1985 td_verror(td, errno, "stat");
1991 if (S_ISREG(sb.st_mode)) {
1992 add_file(td, full_path, 0, 1);
1995 if (!S_ISDIR(sb.st_mode))
1998 ret = recurse_dir(td, full_path);
2007 int add_dir_files(struct thread_data *td, const char *path)
2009 int ret = recurse_dir(td, path);
2012 log_info("fio: opendir added %d files\n", td->o.nr_files);
2017 void dup_files(struct thread_data *td, struct thread_data *org)
2022 dprint(FD_FILE, "dup files: %d\n", org->files_index);
2027 td->files = calloc(org->files_index, sizeof(f));
2029 if (td->o.file_lock_mode != FILE_LOCK_NONE)
2030 td->file_locks = malloc(org->files_index);
2032 assert(org->files_index >= org->o.nr_files);
2033 for_each_file(org, f, i) {
2034 struct fio_file *__f;
2036 __f = alloc_new_file(td);
2039 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2040 __f->file_name = strdup(f->file_name);
2042 __f->file_name = smalloc_strdup(f->file_name);
2044 /* can't handle smalloc failure from here */
2045 assert(__f->file_name);
2046 __f->filetype = f->filetype;
2049 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2050 __f->lock = f->lock;
2051 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2052 __f->rwlock = f->rwlock;
2059 * Returns the index that matches the filename, or -1 if not there
2061 int get_fileno(struct thread_data *td, const char *fname)
2066 for_each_file(td, f, i)
2067 if (!strcmp(f->file_name, fname))
2074 * For log usage, where we add/open/close files automatically
2076 void free_release_files(struct thread_data *td)
2080 td->o.open_files = 0;
2081 td->files_index = 0;
2084 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2088 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2089 f->last_pos[i] = f->file_offset;
2090 f->last_start[i] = -1ULL;
2093 if (fio_file_axmap(f))
2094 axmap_reset(f->io_axmap);
2095 else if (fio_file_lfsr(f))
2096 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2098 zbd_file_reset(td, f);
2101 bool fio_files_done(struct thread_data *td)
2106 for_each_file(td, f, i)
2107 if (!fio_file_done(f))
2113 /* free memory used in initialization phase only */
2114 void filesetup_mem_free(void)
2116 free_already_allocated();
2120 * This function is for platforms which support direct I/O but not O_DIRECT.
2122 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2124 #ifdef FIO_OS_DIRECTIO
2125 int ret = fio_set_odirect(f);
2128 td_verror(td, ret, "fio_set_directio");
2129 #if defined(__sun__)
2130 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2131 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2133 log_err("fio: the file system does not seem to support direct IO\n");
2136 log_err("fio: the file system does not seem to support direct IO\n");
2143 log_err("fio: direct IO is not supported on this host operating system\n");