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 %d unlink %d %s\n", new_layout, unlink_file, 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");
202 dprint(FD_FILE, "fill file %s, size %llu\n", f->file_name, (unsigned long long) f->real_file_size);
204 left = f->real_file_size;
205 bs = td->o.max_bs[DDIR_WRITE];
211 td_verror(td, errno, "malloc");
215 while (left && !td->terminate) {
221 fill_io_buffer(td, b, bs, bs);
223 r = write(f->fd, b, bs);
232 if (__e == ENOSPC || __e == EDQUOT) {
233 const char *__e_name;
234 if (td->o.fill_device)
240 log_info("fio: %s on laying out "
241 "file, stopping\n", __e_name);
243 td_verror(td, errno, "write");
245 td_verror(td, EIO, "write");
252 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
253 td_io_unlink_file(td, f);
254 } else if (td->o.create_fsync) {
255 if (fsync(f->fd) < 0) {
256 td_verror(td, errno, "fsync");
260 if (td->o.fill_device && !td_write(td)) {
261 fio_file_clear_size_known(f);
262 if (td_io_get_file_size(td, f))
264 if (f->io_size > f->real_file_size)
265 f->io_size = f->real_file_size;
279 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
281 int r, did_open = 0, old_runstate;
282 unsigned long long left;
283 unsigned long long bs;
287 if (td_ioengine_flagged(td, FIO_PIPEIO) ||
288 td_ioengine_flagged(td, FIO_NOIO))
291 if (f->filetype == FIO_TYPE_CHAR)
294 if (!fio_file_open(f)) {
295 if (td->io_ops->open_file(td, f)) {
296 log_err("fio: cannot pre-read, failed to open file\n");
302 old_runstate = td_bump_runstate(td, TD_PRE_READING);
305 bs = td->o.max_bs[DDIR_READ];
311 td_verror(td, errno, "malloc");
317 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
318 td_verror(td, errno, "lseek");
319 log_err("fio: failed to lseek pre-read file\n");
324 while (left && !td->terminate) {
328 r = read(f->fd, b, bs);
334 td_verror(td, EIO, "pre_read");
340 td_restore_runstate(td, old_runstate);
343 td->io_ops->close_file(td, f);
350 * Generic function to prepopulate regular file with data.
351 * Useful if you want to make sure I/O engine has data to read.
352 * Leaves f->fd open on success, caller must close.
354 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
357 unsigned long long left, bs;
360 /* generic function for regular files only */
361 assert(f->filetype == FIO_TYPE_FILE);
364 log_err("fio: refusing to write a file due to read-only\n");
369 if (td->o.allow_create)
376 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
377 f->fd = open(f->file_name, flags, 0644);
381 if (err == ENOENT && !td->o.allow_create)
382 log_err("fio: file creation disallowed by "
383 "allow_file_create=0\n");
385 td_verror(td, err, "open");
389 left = f->real_file_size;
390 bs = td->o.max_bs[DDIR_WRITE];
396 td_verror(td, errno, "malloc");
400 while (left && !td->terminate) {
406 fill_io_buffer(td, b, bs, bs);
408 r = write(f->fd, b, bs);
413 td_verror(td, errno, "write");
419 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
420 td_io_unlink_file(td, f);
421 } else if (td->o.create_fsync) {
422 if (fsync(f->fd) < 0) {
423 td_verror(td, errno, "fsync");
438 unsigned long long get_rand_file_size(struct thread_data *td)
440 unsigned long long ret, sized;
444 frand_max = rand_max(&td->file_size_state);
445 r = __rand(&td->file_size_state);
446 sized = td->o.file_size_high - td->o.file_size_low;
447 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
448 ret += td->o.file_size_low;
449 ret -= (ret % td->o.rw_min_bs);
453 static int file_size(struct thread_data *td, struct fio_file *f)
457 if (stat(f->file_name, &st) == -1) {
458 td_verror(td, errno, "fstat");
462 f->real_file_size = st.st_size;
466 static int bdev_size(struct thread_data *td, struct fio_file *f)
468 unsigned long long bytes = 0;
471 if (td->io_ops->open_file(td, f)) {
472 log_err("fio: failed opening blockdev %s for size check\n",
477 r = blockdev_size(f, &bytes);
479 td_verror(td, r, "blockdev_size");
484 log_err("%s: zero sized block device?\n", f->file_name);
488 f->real_file_size = bytes;
489 td->io_ops->close_file(td, f);
492 td->io_ops->close_file(td, f);
496 static int char_size(struct thread_data *td, struct fio_file *f)
498 #ifdef FIO_HAVE_CHARDEV_SIZE
499 unsigned long long bytes = 0;
502 if (td->io_ops->open_file(td, f)) {
503 log_err("fio: failed opening chardev %s for size check\n",
508 r = chardev_size(f, &bytes);
510 td_verror(td, r, "chardev_size");
515 log_err("%s: zero sized char device?\n", f->file_name);
519 f->real_file_size = bytes;
520 td->io_ops->close_file(td, f);
523 td->io_ops->close_file(td, f);
526 f->real_file_size = -1ULL;
531 static int get_file_size(struct thread_data *td, struct fio_file *f)
535 if (fio_file_size_known(f))
538 if (f->filetype == FIO_TYPE_FILE)
539 ret = file_size(td, f);
540 else if (f->filetype == FIO_TYPE_BLOCK)
541 ret = bdev_size(td, f);
542 else if (f->filetype == FIO_TYPE_CHAR)
543 ret = char_size(td, f);
545 f->real_file_size = -1;
546 log_info("%s: failed to get file size of %s\n", td->o.name,
548 return 1; /* avoid offset extends end error message */
552 * Leave ->real_file_size with 0 since it could be expectation
553 * of initial setup for regular files.
559 * ->file_offset normally hasn't been initialized yet, so this
560 * is basically always false unless ->real_file_size is -1, but
561 * if ->real_file_size is -1 this message doesn't make sense.
562 * As a result, this message is basically useless.
564 if (f->file_offset > f->real_file_size) {
565 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
566 (unsigned long long) f->file_offset,
567 (unsigned long long) f->real_file_size);
571 fio_file_set_size_known(f);
575 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
576 unsigned long long off,
577 unsigned long long len)
579 int errval = 0, ret = 0;
588 off = f->file_offset;
590 if (len == -1ULL || off == -1ULL)
593 if (td->io_ops->invalidate) {
594 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
596 ret = td->io_ops->invalidate(td, f);
599 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
600 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
602 } else if (f->filetype == FIO_TYPE_FILE) {
603 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
604 f->file_name, off, len);
605 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
608 } else if (f->filetype == FIO_TYPE_BLOCK) {
611 dprint(FD_IO, "drop page cache %s\n", f->file_name);
612 ret = blockdev_invalidate_cache(f);
613 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
615 * Linux multipath devices reject ioctl while
616 * the maps are being updated. That window can
617 * last tens of milliseconds; we'll try up to
618 * a quarter of a second.
621 ret = blockdev_invalidate_cache(f);
623 if (ret < 0 && errno == EACCES && geteuid()) {
624 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
625 log_err("fio: only root may flush block "
626 "devices. Cache flush bypassed!\n");
631 } else if (f->filetype == FIO_TYPE_CHAR ||
632 f->filetype == FIO_TYPE_PIPE) {
633 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
637 * Cache flushing isn't a fatal condition, and we know it will
638 * happen on some platforms where we don't have the proper
639 * function to flush eg block device caches. So just warn and
640 * continue on our way.
643 log_info("fio: cache invalidation of %s failed: %s\n",
644 f->file_name, strerror(errval));
650 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
652 if (!fio_file_open(f))
655 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
658 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
662 dprint(FD_FILE, "fd close %s\n", f->file_name);
666 if (close(f->fd) < 0)
671 if (f->shadow_fd != -1) {
680 int file_lookup_open(struct fio_file *f, int flags)
682 struct fio_file *__f;
685 __f = lookup_file_hash(f->file_name);
687 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
691 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
699 f->fd = open(f->file_name, flags, 0600);
703 static int file_close_shadow_fds(struct thread_data *td)
709 for_each_file(td, f, i) {
710 if (f->shadow_fd == -1)
721 int generic_open_file(struct thread_data *td, struct fio_file *f)
727 dprint(FD_FILE, "fd open %s\n", f->file_name);
729 if (!strcmp(f->file_name, "-")) {
731 log_err("fio: can't read/write to stdin/out\n");
737 * move output logging to stderr, if we are writing to stdout
746 flags |= OS_O_DIRECT;
749 td_verror(td, EINVAL, "OS does not support atomic IO");
752 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
754 flags |= td->o.sync_io;
755 if (td->o.create_on_open && td->o.allow_create)
758 if (f->filetype != FIO_TYPE_FILE)
759 flags |= FIO_O_NOATIME;
766 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
770 f->fd = dup(STDOUT_FILENO);
772 from_hash = file_lookup_open(f, flags);
773 } else if (td_read(td)) {
774 if (f->filetype == FIO_TYPE_CHAR && !read_only)
780 f->fd = dup(STDIN_FILENO);
782 from_hash = file_lookup_open(f, flags);
783 } else if (td_trim(td)) {
784 assert(!td_rw(td)); /* should have matched above */
787 from_hash = file_lookup_open(f, flags);
791 char buf[FIO_VERROR_SIZE];
794 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
795 flags &= ~FIO_O_NOATIME;
798 if (__e == EMFILE && file_close_shadow_fds(td))
801 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
803 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
804 log_err("fio: looks like your file system does not " \
805 "support direct=1/buffered=0\n");
808 td_verror(td, __e, buf);
812 if (!from_hash && f->fd != -1) {
813 if (add_file_hash(f)) {
817 * Stash away descriptor for later close. This is to
818 * work-around a "feature" on Linux, where a close of
819 * an fd that has been opened for write will trigger
820 * udev to call blkid to check partitions, fs id, etc.
821 * That pollutes the device cache, which can slow down
822 * unbuffered accesses.
824 if (f->shadow_fd == -1)
825 f->shadow_fd = f->fd;
828 * OK to ignore, we haven't done anything
831 ret = generic_close_file(td, f);
841 * This function i.e. get_file_size() is the default .get_file_size
842 * implementation of majority of I/O engines.
844 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
846 return get_file_size(td, f);
850 * open/close all files, so that ->real_file_size gets set
852 static int get_file_sizes(struct thread_data *td)
858 for_each_file(td, f, i) {
859 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
862 if (td_io_get_file_size(td, f)) {
863 if (td->error != ENOENT) {
864 log_err("%s\n", td->verror);
872 * There are corner cases where we end up with -1 for
873 * ->real_file_size due to unsupported file type, etc.
874 * We then just set to size option value divided by number
875 * of files, similar to the way file ->io_size is set.
876 * stat(2) failure doesn't set ->real_file_size to -1.
878 if (f->real_file_size == -1ULL && td->o.size)
879 f->real_file_size = td->o.size / td->o.nr_files;
886 struct flist_head list;
893 * Get free number of bytes for each file on each unique mount.
895 static unsigned long long get_fs_free_counts(struct thread_data *td)
897 struct flist_head *n, *tmp;
898 unsigned long long ret = 0;
899 struct fio_mount *fm;
904 for_each_file(td, f, i) {
908 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
909 if (f->real_file_size != -1ULL)
910 ret += f->real_file_size;
912 } else if (f->filetype != FIO_TYPE_FILE)
915 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
917 if (stat(buf, &sb) < 0) {
921 if (stat(buf, &sb) < 0)
926 flist_for_each(n, &list) {
927 fm = flist_entry(n, struct fio_mount, list);
928 if (fm->key == sb.st_dev)
937 fm = calloc(1, sizeof(*fm));
938 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
939 fm->base = basename(fm->__base);
941 flist_add(&fm->list, &list);
944 flist_for_each_safe(n, tmp, &list) {
945 unsigned long long sz;
947 fm = flist_entry(n, struct fio_mount, list);
948 flist_del(&fm->list);
950 sz = get_fs_free_size(fm->base);
951 if (sz && sz != -1ULL)
960 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
963 struct thread_options *o = &td->o;
964 unsigned long long align_bs;
965 unsigned long long offset;
966 unsigned long long increment;
968 if (o->file_append && f->filetype == FIO_TYPE_FILE)
969 return f->real_file_size;
971 if (o->offset_increment_percent) {
972 assert(!o->offset_increment);
973 increment = o->offset_increment_percent * f->real_file_size / 100;
976 increment = o->offset_increment;
978 if (o->start_offset_percent > 0) {
979 /* calculate the raw offset */
980 offset = (f->real_file_size * o->start_offset_percent / 100) +
981 (td->subjob_number * increment);
985 /* start_offset_percent not set */
986 offset = o->start_offset +
987 td->subjob_number * increment;
992 * if offset_align is provided, use it
994 if (fio_option_is_set(o, start_offset_align)) {
995 align_bs = o->start_offset_align;
997 /* else take the minimum block size */
998 align_bs = td_min_bs(td);
1002 * block align the offset at the next available boundary at
1003 * ceiling(offset / align_bs) * align_bs
1005 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
1012 * Find longest path component that exists and return its length
1014 int longest_existing_path(char *path) {
1025 sprintf(buf, "%s", path);
1028 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1034 *(buf_pos + 1) = '\0';
1037 dwAttr = GetFileAttributesA(buf);
1038 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1042 if (stat(buf, &sb) == 0)
1046 offset = buf_pos - buf;
1054 static bool create_work_dirs(struct thread_data *td, const char *fname)
1056 char path[PATH_MAX];
1060 snprintf(path, PATH_MAX, "%s", fname);
1063 offset = longest_existing_path(path);
1064 end = start + offset;
1065 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1072 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1073 log_err("fio: failed to create dir (%s): %s\n",
1074 start, strerror(errno));
1077 *end = FIO_OS_PATH_SEPARATOR;
1080 td->flags |= TD_F_DIRS_CREATED;
1084 int setup_shared_file(struct thread_data *td)
1090 if (td->o.nr_files > 1) {
1091 log_err("fio: shared file setup called for multiple files\n");
1100 log_err("fio: NULL shared file\n");
1104 file_size = thread_number * td->o.size;
1105 dprint(FD_FILE, "shared setup %s real_file_size=%llu, desired=%llu\n",
1106 f->file_name, (unsigned long long)f->real_file_size, (unsigned long long)file_size);
1108 if (f->real_file_size < file_size) {
1109 dprint(FD_FILE, "fio: extending shared file\n");
1110 f->real_file_size = file_size;
1111 err = extend_file(td, f);
1113 err = __file_invalidate_cache(td, f, 0, f->real_file_size);
1115 dprint(FD_FILE, "shared setup new real_file_size=%llu\n",
1116 (unsigned long long)f->real_file_size);
1123 * Open the files and setup files sizes, creating files if necessary.
1125 int setup_files(struct thread_data *td)
1127 unsigned long long total_size, extend_size;
1128 struct thread_options *o = &td->o;
1130 unsigned int i, nr_fs_extra = 0;
1131 int err = 0, need_extend;
1133 const unsigned long long bs = td_min_bs(td);
1136 dprint(FD_FILE, "setup files (thread_number=%d, subjob_number=%d)\n", td->thread_number, td->subjob_number);
1138 old_state = td_bump_runstate(td, TD_SETTING_UP);
1140 for_each_file(td, f, i) {
1141 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1142 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1143 !(td->flags & TD_F_DIRS_CREATED) &&
1144 !create_work_dirs(td, f->file_name))
1149 * Find out physical size of files or devices for this thread,
1150 * before we determine I/O size and range of our targets.
1151 * If ioengine defines a setup() method, it's responsible for
1152 * opening the files and setting f->real_file_size to indicate
1153 * the valid range for that file.
1155 if (td->io_ops->setup)
1156 err = td->io_ops->setup(td);
1158 err = get_file_sizes(td);
1163 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1164 err = zbd_init_files(td);
1168 zbd_recalc_options_with_zone_granularity(td);
1170 if (o->read_iolog_file)
1174 * check sizes. if the files/devices do not exist and the size
1175 * isn't passed to fio, abort.
1178 for_each_file(td, f, i) {
1180 if (f->real_file_size == -1ULL)
1183 total_size += f->real_file_size;
1187 td->fill_device_size = get_fs_free_counts(td);
1190 * device/file sizes are zero and no size given, punt
1192 if ((!total_size || total_size == -1ULL) && !o->size &&
1193 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1194 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1195 log_err("%s: you need to specify size=\n", o->name);
1196 td_verror(td, EINVAL, "total_file_size");
1201 * Calculate per-file size and potential extra size for the
1202 * first files, if needed (i.e. if we don't have a fixed size).
1204 if (!o->file_size_low && o->nr_files) {
1207 fs = o->size / o->nr_files;
1208 all_fs = fs * o->nr_files;
1210 if (all_fs < o->size)
1211 nr_fs_extra = (o->size - all_fs) / bs;
1215 * now file sizes are known, so we can set ->io_size. if size= is
1216 * not given, ->io_size is just equal to ->real_file_size. if size
1217 * is given, ->io_size is size / nr_files.
1219 extend_size = total_size = 0;
1221 for_each_file(td, f, i) {
1222 f->file_offset = get_start_offset(td, f);
1225 * Update ->io_size depending on options specified.
1226 * ->file_size_low being 0 means filesize option isn't set.
1227 * Non zero ->file_size_low equals ->file_size_high means
1228 * filesize option is set in a fixed size format.
1229 * Non zero ->file_size_low not equals ->file_size_high means
1230 * filesize option is set in a range format.
1232 if (!o->file_size_low) {
1234 * no file size or range given, file size is equal to
1235 * total size divided by number of files. If the size
1236 * doesn't divide nicely with the min blocksize,
1237 * make the first files bigger.
1246 * We normally don't come here for regular files, but
1247 * if the result is 0 for a regular file, set it to the
1248 * real file size. This could be size of the existing
1249 * one if it already exists, but otherwise will be set
1250 * to 0. A new file won't be created because
1251 * ->io_size + ->file_offset equals ->real_file_size.
1254 if (f->file_offset > f->real_file_size)
1256 f->io_size = f->real_file_size - f->file_offset;
1258 log_info("fio: file %s may be ignored\n",
1261 } else if (f->real_file_size < o->file_size_low ||
1262 f->real_file_size > o->file_size_high) {
1263 if (f->file_offset > o->file_size_low)
1266 * file size given. if it's fixed, use that. if it's a
1267 * range, generate a random size in-between.
1269 if (o->file_size_low == o->file_size_high)
1270 f->io_size = o->file_size_low - f->file_offset;
1272 f->io_size = get_rand_file_size(td)
1276 f->io_size = f->real_file_size - f->file_offset;
1278 if (f->io_size == -1ULL)
1283 if (o->size_percent && o->size_percent != 100) {
1286 file_size = f->io_size + f->file_offset;
1287 f->io_size = (file_size *
1288 o->size_percent) / 100;
1289 if (f->io_size > (file_size - f->file_offset))
1290 f->io_size = file_size - f->file_offset;
1292 f->io_size -= (f->io_size % td_min_bs(td));
1295 io_size = f->io_size;
1296 if (o->io_size_percent && o->io_size_percent != 100) {
1297 io_size *= o->io_size_percent;
1301 total_size += io_size;
1304 if (f->filetype == FIO_TYPE_FILE &&
1305 (f->io_size + f->file_offset) > f->real_file_size) {
1306 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1307 !o->create_on_open) {
1309 extend_size += (f->io_size + f->file_offset);
1310 fio_file_set_extend(f);
1311 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1312 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1313 td_ioengine_flagged(td, FIO_FAKEIO)))
1314 f->real_file_size = f->io_size + f->file_offset;
1318 if (td->o.block_error_hist) {
1321 assert(td->o.nr_files == 1); /* checked in fixup_options */
1323 len = f->io_size / td->o.bs[DDIR_TRIM];
1324 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1325 log_err("fio: cannot calculate block histogram with "
1326 "%d trim blocks, maximum %d\n",
1327 len, MAX_NR_BLOCK_INFOS);
1328 td_verror(td, EINVAL, "block_error_hist");
1332 td->ts.nr_block_infos = len;
1333 for (i = 0; i < len; i++)
1334 td->ts.block_infos[i] =
1335 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1337 td->ts.nr_block_infos = 0;
1339 if (!o->size || (total_size && o->size > total_size))
1340 o->size = total_size;
1342 if (o->size < td_min_bs(td)) {
1343 log_err("fio: blocksize is larger than data set range\n");
1348 * See if we need to extend some files, typically needed when our
1349 * target regular files don't exist yet, but our jobs require them
1350 * initially due to read I/Os.
1354 if (output_format & FIO_OUTPUT_NORMAL) {
1355 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1357 need_extend > 1 ? "s" : "",
1359 need_extend > 1 ? "s" : "",
1360 need_extend > 1 ? "total " : "",
1364 for_each_file(td, f, i) {
1365 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1367 if (!fio_file_extend(f))
1370 assert(f->filetype == FIO_TYPE_FILE);
1371 fio_file_clear_extend(f);
1372 if (!o->fill_device) {
1373 old_len = f->real_file_size;
1374 extend_len = f->io_size + f->file_offset -
1377 f->real_file_size = (f->io_size + f->file_offset);
1378 err = extend_file(td, f);
1382 err = __file_invalidate_cache(td, f, old_len,
1386 * Shut up static checker
1402 * Prepopulate files with data. It might be expected to read some
1403 * "real" data instead of zero'ed files (if no writes to file occurred
1404 * prior to a read job). Engine has to provide a way to do that.
1406 if (td->io_ops->prepopulate_file) {
1409 for_each_file(td, f, i) {
1410 if (output_format & FIO_OUTPUT_NORMAL) {
1411 log_info("%s: Prepopulating IO file (%s)\n",
1412 o->name, f->file_name);
1415 err = td->io_ops->prepopulate_file(td, f);
1419 err = __file_invalidate_cache(td, f, f->file_offset,
1423 * Shut up static checker
1439 * iolog already set the total io size, if we read back
1442 if (!o->read_iolog_file) {
1444 td->total_io_size = o->io_size * o->loops;
1446 td->total_io_size = o->size * o->loops;
1450 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1451 err = zbd_setup_files(td);
1459 td_restore_runstate(td, old_state);
1464 log_err("%s: you need to specify valid offset=\n", o->name);
1466 td_restore_runstate(td, old_state);
1470 bool pre_read_files(struct thread_data *td)
1475 dprint(FD_FILE, "pre_read files\n");
1477 for_each_file(td, f, i) {
1478 if (!pre_read_file(td, f))
1485 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1487 unsigned int range_size, seed;
1491 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1492 fsize = min(f->real_file_size, f->io_size);
1494 nranges = (fsize + range_size - 1ULL) / range_size;
1496 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1497 if (!td->o.rand_repeatable)
1498 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1500 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1501 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1502 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1503 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1504 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1505 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1508 static bool init_rand_distribution(struct thread_data *td)
1514 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1515 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1516 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1519 state = td_bump_runstate(td, TD_SETTING_UP);
1521 for_each_file(td, f, i)
1522 __init_rand_distribution(td, f);
1524 td_restore_runstate(td, state);
1529 * Check if the number of blocks exceeds the randomness capability of
1530 * the selected generator. Tausworthe is 32-bit, the others are fully
1533 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1536 if (blocks <= FRAND32_MAX)
1538 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1542 * If the user hasn't specified a random generator, switch
1543 * to tausworthe64 with informational warning. If the user did
1544 * specify one, just warn.
1546 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1549 if (!fio_option_is_set(&td->o, random_generator)) {
1550 log_info("fio: Switching to tausworthe64. Use the "
1551 "random_generator= option to get rid of this "
1553 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1558 * Just make this information to avoid breaking scripts.
1560 log_info("fio: Use the random_generator= option to switch to lfsr or "
1565 bool init_random_map(struct thread_data *td)
1567 unsigned long long blocks;
1571 if (init_rand_distribution(td))
1576 for_each_file(td, f, i) {
1577 uint64_t fsize = min(f->real_file_size, f->io_size);
1579 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1580 fsize = td->o.zone_range;
1582 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1584 if (check_rand_gen_limits(td, f, blocks))
1587 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1590 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1592 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1593 fio_file_set_lfsr(f);
1596 log_err("fio: failed initializing LFSR\n");
1599 } else if (!td->o.norandommap) {
1600 f->io_axmap = axmap_new(blocks);
1602 fio_file_set_axmap(f);
1605 } else if (td->o.norandommap)
1608 if (!td->o.softrandommap) {
1609 log_err("fio: failed allocating random map. If running"
1610 " a large number of jobs, try the 'norandommap'"
1611 " option or set 'softrandommap'. Or give"
1612 " a larger --alloc-size to fio.\n");
1616 log_info("fio: file %s failed allocating random map. Running "
1617 "job without.\n", f->file_name);
1623 void close_files(struct thread_data *td)
1628 for_each_file(td, f, i) {
1629 if (fio_file_open(f))
1630 td_io_close_file(td, f);
1634 void fio_file_free(struct fio_file *f)
1636 if (fio_file_axmap(f))
1637 axmap_free(f->io_axmap);
1638 if (!fio_file_smalloc(f)) {
1642 sfree(f->file_name);
1647 void close_and_free_files(struct thread_data *td)
1652 dprint(FD_FILE, "close files\n");
1654 for_each_file(td, f, i) {
1655 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1656 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1657 td_io_unlink_file(td, f);
1660 if (fio_file_open(f))
1661 td_io_close_file(td, f);
1663 remove_file_hash(f);
1665 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1666 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1667 td_io_unlink_file(td, f);
1674 td->o.filename = NULL;
1676 free(td->file_locks);
1677 td->files_index = 0;
1679 td->file_locks = NULL;
1680 td->o.file_lock_mode = FILE_LOCK_NONE;
1684 static void get_file_type(struct fio_file *f)
1688 if (!strcmp(f->file_name, "-"))
1689 f->filetype = FIO_TYPE_PIPE;
1691 f->filetype = FIO_TYPE_FILE;
1694 /* \\.\ is the device namespace in Windows, where every file is
1696 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1697 f->filetype = FIO_TYPE_BLOCK;
1700 if (!stat(f->file_name, &sb)) {
1701 if (S_ISBLK(sb.st_mode))
1702 f->filetype = FIO_TYPE_BLOCK;
1703 else if (S_ISCHR(sb.st_mode))
1704 f->filetype = FIO_TYPE_CHAR;
1705 else if (S_ISFIFO(sb.st_mode))
1706 f->filetype = FIO_TYPE_PIPE;
1710 static bool __is_already_allocated(const char *fname, bool set)
1712 struct flist_head *entry;
1715 ret = file_bloom_exists(fname, set);
1719 flist_for_each(entry, &filename_list) {
1720 struct file_name *fn;
1722 fn = flist_entry(entry, struct file_name, list);
1724 if (!strcmp(fn->filename, fname))
1731 static bool is_already_allocated(const char *fname)
1735 fio_file_hash_lock();
1736 ret = __is_already_allocated(fname, false);
1737 fio_file_hash_unlock();
1742 static void set_already_allocated(const char *fname)
1744 struct file_name *fn;
1746 fn = malloc(sizeof(struct file_name));
1747 fn->filename = strdup(fname);
1749 fio_file_hash_lock();
1750 if (!__is_already_allocated(fname, true)) {
1751 flist_add_tail(&fn->list, &filename_list);
1754 fio_file_hash_unlock();
1762 static void free_already_allocated(void)
1764 struct flist_head *entry, *tmp;
1765 struct file_name *fn;
1767 if (flist_empty(&filename_list))
1770 fio_file_hash_lock();
1771 flist_for_each_safe(entry, tmp, &filename_list) {
1772 fn = flist_entry(entry, struct file_name, list);
1774 flist_del(&fn->list);
1778 fio_file_hash_unlock();
1781 static struct fio_file *alloc_new_file(struct thread_data *td)
1785 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1786 f = calloc(1, sizeof(*f));
1788 f = scalloc(1, sizeof(*f));
1796 fio_file_reset(td, f);
1797 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1798 fio_file_set_smalloc(f);
1802 bool exists_and_not_regfile(const char *filename)
1806 if (lstat(filename, &sb) == -1)
1809 #ifndef WIN32 /* NOT Windows */
1810 if (S_ISREG(sb.st_mode))
1813 /* \\.\ is the device namespace in Windows, where every file
1814 * is a device node */
1815 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1822 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1824 int cur_files = td->files_index;
1825 char file_name[PATH_MAX];
1829 dprint(FD_FILE, "add file %s\n", fname);
1831 if (td->o.directory)
1832 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1833 td->o.unique_filename);
1835 sprintf(file_name + len, "%s", fname);
1837 /* clean cloned siblings using existing files */
1838 if (numjob && is_already_allocated(file_name) &&
1839 !exists_and_not_regfile(fname))
1842 f = alloc_new_file(td);
1844 if (td->files_size <= td->files_index) {
1845 unsigned int new_size = td->o.nr_files + 1;
1847 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1849 td->files = realloc(td->files, new_size * sizeof(f));
1850 if (td->files == NULL) {
1851 log_err("fio: realloc OOM\n");
1854 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1855 td->file_locks = realloc(td->file_locks, new_size);
1856 if (!td->file_locks) {
1857 log_err("fio: realloc OOM\n");
1860 td->file_locks[cur_files] = FILE_LOCK_NONE;
1862 td->files_size = new_size;
1864 td->files[cur_files] = f;
1865 f->fileno = cur_files;
1868 * init function, io engine may not be loaded yet
1870 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1871 f->real_file_size = -1ULL;
1873 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1874 f->file_name = strdup(file_name);
1876 f->file_name = smalloc_strdup(file_name);
1878 /* can't handle smalloc failure from here */
1879 assert(f->file_name);
1883 switch (td->o.file_lock_mode) {
1884 case FILE_LOCK_NONE:
1886 case FILE_LOCK_READWRITE:
1887 f->rwlock = fio_rwlock_init();
1889 case FILE_LOCK_EXCLUSIVE:
1890 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1893 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1899 if (td->o.numjobs > 1)
1900 set_already_allocated(file_name);
1905 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1911 int add_file_exclusive(struct thread_data *td, const char *fname)
1916 for_each_file(td, f, i) {
1917 if (!strcmp(f->file_name, fname))
1921 return add_file(td, fname, 0, 1);
1924 void get_file(struct fio_file *f)
1926 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1927 assert(fio_file_open(f));
1931 int put_file(struct thread_data *td, struct fio_file *f)
1933 int f_ret = 0, ret = 0;
1935 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1937 if (!fio_file_open(f)) {
1938 assert(f->fd == -1);
1942 assert(f->references);
1943 if (--f->references)
1946 disk_util_dec(f->du);
1948 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1949 unlock_file_all(td, f);
1951 if (should_fsync(td) && td->o.fsync_on_close) {
1952 f_ret = fsync(f->fd);
1957 if (td->io_ops->close_file)
1958 ret = td->io_ops->close_file(td, f);
1963 td->nr_open_files--;
1964 fio_file_clear_closing(f);
1965 fio_file_clear_open(f);
1966 assert(f->fd == -1);
1970 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1972 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1975 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1976 if (ddir == DDIR_READ)
1977 fio_rwlock_read(f->rwlock);
1979 fio_rwlock_write(f->rwlock);
1980 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1981 fio_sem_down(f->lock);
1983 td->file_locks[f->fileno] = td->o.file_lock_mode;
1986 void unlock_file(struct thread_data *td, struct fio_file *f)
1988 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1991 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1992 fio_rwlock_unlock(f->rwlock);
1993 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1994 fio_sem_up(f->lock);
1996 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1999 void unlock_file_all(struct thread_data *td, struct fio_file *f)
2001 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
2003 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
2007 static bool recurse_dir(struct thread_data *td, const char *dirname)
2013 D = opendir(dirname);
2015 char buf[FIO_VERROR_SIZE];
2017 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
2018 td_verror(td, errno, buf);
2022 while ((dir = readdir(D)) != NULL) {
2023 char full_path[PATH_MAX];
2026 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
2029 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
2031 if (lstat(full_path, &sb) == -1) {
2032 if (errno != ENOENT) {
2033 td_verror(td, errno, "stat");
2039 if (S_ISREG(sb.st_mode)) {
2040 add_file(td, full_path, 0, 1);
2043 if (!S_ISDIR(sb.st_mode))
2046 ret = recurse_dir(td, full_path);
2055 int add_dir_files(struct thread_data *td, const char *path)
2057 int ret = recurse_dir(td, path);
2060 log_info("fio: opendir added %d files\n", td->o.nr_files);
2065 void dup_files(struct thread_data *td, struct thread_data *org)
2070 dprint(FD_FILE, "dup files: %d\n", org->files_index);
2075 td->files = calloc(org->files_index, sizeof(f));
2077 if (td->o.file_lock_mode != FILE_LOCK_NONE)
2078 td->file_locks = malloc(org->files_index);
2080 assert(org->files_index >= org->o.nr_files);
2081 for_each_file(org, f, i) {
2082 struct fio_file *__f;
2084 __f = alloc_new_file(td);
2087 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2088 __f->file_name = strdup(f->file_name);
2090 __f->file_name = smalloc_strdup(f->file_name);
2092 /* can't handle smalloc failure from here */
2093 assert(__f->file_name);
2094 __f->filetype = f->filetype;
2097 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2098 __f->lock = f->lock;
2099 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2100 __f->rwlock = f->rwlock;
2107 * Returns the index that matches the filename, or -1 if not there
2109 int get_fileno(struct thread_data *td, const char *fname)
2114 for_each_file(td, f, i)
2115 if (!strcmp(f->file_name, fname))
2122 * For log usage, where we add/open/close files automatically
2124 void free_release_files(struct thread_data *td)
2128 td->o.open_files = 0;
2129 td->files_index = 0;
2132 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2136 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2137 f->last_pos[i] = f->file_offset;
2138 f->last_start[i] = -1ULL;
2141 if (fio_file_axmap(f))
2142 axmap_reset(f->io_axmap);
2143 else if (fio_file_lfsr(f))
2144 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2146 zbd_file_reset(td, f);
2149 bool fio_files_done(struct thread_data *td)
2154 for_each_file(td, f, i)
2155 if (!fio_file_done(f))
2161 /* free memory used in initialization phase only */
2162 void filesetup_mem_free(void)
2164 free_already_allocated();
2168 * This function is for platforms which support direct I/O but not O_DIRECT.
2170 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2172 #ifdef FIO_OS_DIRECTIO
2173 int ret = fio_set_odirect(f);
2176 td_verror(td, ret, "fio_set_directio");
2177 #if defined(__sun__)
2178 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2179 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2181 log_err("fio: the file system does not seem to support direct IO\n");
2184 log_err("fio: the file system does not seem to support direct IO\n");
2191 log_err("fio: direct IO is not supported on this host operating system\n");