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;
746 td_verror(td, EINVAL, "OS does not support atomic IO");
749 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
751 flags |= td->o.sync_io;
752 if (td->o.create_on_open && td->o.allow_create)
755 if (f->filetype != FIO_TYPE_FILE)
756 flags |= FIO_O_NOATIME;
763 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
767 f->fd = dup(STDOUT_FILENO);
769 from_hash = file_lookup_open(f, flags);
770 } else if (td_read(td)) {
771 if (f->filetype == FIO_TYPE_CHAR && !read_only)
777 f->fd = dup(STDIN_FILENO);
779 from_hash = file_lookup_open(f, flags);
780 } else if (td_trim(td)) {
781 assert(!td_rw(td)); /* should have matched above */
784 from_hash = file_lookup_open(f, flags);
788 char buf[FIO_VERROR_SIZE];
791 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
792 flags &= ~FIO_O_NOATIME;
795 if (__e == EMFILE && file_close_shadow_fds(td))
798 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
800 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
801 log_err("fio: looks like your file system does not " \
802 "support direct=1/buffered=0\n");
805 td_verror(td, __e, buf);
809 if (!from_hash && f->fd != -1) {
810 if (add_file_hash(f)) {
814 * Stash away descriptor for later close. This is to
815 * work-around a "feature" on Linux, where a close of
816 * an fd that has been opened for write will trigger
817 * udev to call blkid to check partitions, fs id, etc.
818 * That pollutes the device cache, which can slow down
819 * unbuffered accesses.
821 if (f->shadow_fd == -1)
822 f->shadow_fd = f->fd;
825 * OK to ignore, we haven't done anything
828 ret = generic_close_file(td, f);
838 * This function i.e. get_file_size() is the default .get_file_size
839 * implementation of majority of I/O engines.
841 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
843 return get_file_size(td, f);
847 * open/close all files, so that ->real_file_size gets set
849 static int get_file_sizes(struct thread_data *td)
855 for_each_file(td, f, i) {
856 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
859 if (td_io_get_file_size(td, f)) {
860 if (td->error != ENOENT) {
861 log_err("%s\n", td->verror);
869 * There are corner cases where we end up with -1 for
870 * ->real_file_size due to unsupported file type, etc.
871 * We then just set to size option value divided by number
872 * of files, similar to the way file ->io_size is set.
873 * stat(2) failure doesn't set ->real_file_size to -1.
875 if (f->real_file_size == -1ULL && td->o.size)
876 f->real_file_size = td->o.size / td->o.nr_files;
883 struct flist_head list;
890 * Get free number of bytes for each file on each unique mount.
892 static unsigned long long get_fs_free_counts(struct thread_data *td)
894 struct flist_head *n, *tmp;
895 unsigned long long ret = 0;
896 struct fio_mount *fm;
901 for_each_file(td, f, i) {
905 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
906 if (f->real_file_size != -1ULL)
907 ret += f->real_file_size;
909 } else if (f->filetype != FIO_TYPE_FILE)
912 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
914 if (stat(buf, &sb) < 0) {
918 if (stat(buf, &sb) < 0)
923 flist_for_each(n, &list) {
924 fm = flist_entry(n, struct fio_mount, list);
925 if (fm->key == sb.st_dev)
934 fm = calloc(1, sizeof(*fm));
935 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
936 fm->base = basename(fm->__base);
938 flist_add(&fm->list, &list);
941 flist_for_each_safe(n, tmp, &list) {
942 unsigned long long sz;
944 fm = flist_entry(n, struct fio_mount, list);
945 flist_del(&fm->list);
947 sz = get_fs_free_size(fm->base);
948 if (sz && sz != -1ULL)
957 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
960 struct thread_options *o = &td->o;
961 unsigned long long align_bs;
962 unsigned long long offset;
963 unsigned long long increment;
965 if (o->file_append && f->filetype == FIO_TYPE_FILE)
966 return f->real_file_size;
968 if (o->offset_increment_percent) {
969 assert(!o->offset_increment);
970 increment = o->offset_increment_percent * f->real_file_size / 100;
973 increment = o->offset_increment;
975 if (o->start_offset_percent > 0) {
976 /* calculate the raw offset */
977 offset = (f->real_file_size * o->start_offset_percent / 100) +
978 (td->subjob_number * increment);
982 /* start_offset_percent not set */
983 offset = o->start_offset +
984 td->subjob_number * increment;
989 * if offset_align is provided, use it
991 if (fio_option_is_set(o, start_offset_align)) {
992 align_bs = o->start_offset_align;
994 /* else take the minimum block size */
995 align_bs = td_min_bs(td);
999 * block align the offset at the next available boundary at
1000 * ceiling(offset / align_bs) * align_bs
1002 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
1009 * Find longest path component that exists and return its length
1011 int longest_existing_path(char *path) {
1022 sprintf(buf, "%s", path);
1025 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1031 *(buf_pos + 1) = '\0';
1034 dwAttr = GetFileAttributesA(buf);
1035 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1039 if (stat(buf, &sb) == 0)
1043 offset = buf_pos - buf;
1051 static bool create_work_dirs(struct thread_data *td, const char *fname)
1053 char path[PATH_MAX];
1057 snprintf(path, PATH_MAX, "%s", fname);
1060 offset = longest_existing_path(path);
1061 end = start + offset;
1062 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1069 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1070 log_err("fio: failed to create dir (%s): %s\n",
1071 start, strerror(errno));
1074 *end = FIO_OS_PATH_SEPARATOR;
1077 td->flags |= TD_F_DIRS_CREATED;
1082 * Open the files and setup files sizes, creating files if necessary.
1084 int setup_files(struct thread_data *td)
1086 unsigned long long total_size, extend_size;
1087 struct thread_options *o = &td->o;
1089 unsigned int i, nr_fs_extra = 0;
1090 int err = 0, need_extend;
1092 const unsigned long long bs = td_min_bs(td);
1095 dprint(FD_FILE, "setup files\n");
1097 old_state = td_bump_runstate(td, TD_SETTING_UP);
1099 for_each_file(td, f, i) {
1100 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1101 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1102 !(td->flags & TD_F_DIRS_CREATED) &&
1103 !create_work_dirs(td, f->file_name))
1108 * Find out physical size of files or devices for this thread,
1109 * before we determine I/O size and range of our targets.
1110 * If ioengine defines a setup() method, it's responsible for
1111 * opening the files and setting f->real_file_size to indicate
1112 * the valid range for that file.
1114 if (td->io_ops->setup)
1115 err = td->io_ops->setup(td);
1117 err = get_file_sizes(td);
1122 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1123 err = zbd_init_files(td);
1127 zbd_recalc_options_with_zone_granularity(td);
1129 if (o->read_iolog_file)
1133 * check sizes. if the files/devices do not exist and the size
1134 * isn't passed to fio, abort.
1137 for_each_file(td, f, i) {
1139 if (f->real_file_size == -1ULL)
1142 total_size += f->real_file_size;
1146 td->fill_device_size = get_fs_free_counts(td);
1149 * device/file sizes are zero and no size given, punt
1151 if ((!total_size || total_size == -1ULL) && !o->size &&
1152 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1153 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1154 log_err("%s: you need to specify size=\n", o->name);
1155 td_verror(td, EINVAL, "total_file_size");
1160 * Calculate per-file size and potential extra size for the
1161 * first files, if needed (i.e. if we don't have a fixed size).
1163 if (!o->file_size_low && o->nr_files) {
1166 fs = o->size / o->nr_files;
1167 all_fs = fs * o->nr_files;
1169 if (all_fs < o->size)
1170 nr_fs_extra = (o->size - all_fs) / bs;
1174 * now file sizes are known, so we can set ->io_size. if size= is
1175 * not given, ->io_size is just equal to ->real_file_size. if size
1176 * is given, ->io_size is size / nr_files.
1178 extend_size = total_size = 0;
1180 for_each_file(td, f, i) {
1181 f->file_offset = get_start_offset(td, f);
1184 * Update ->io_size depending on options specified.
1185 * ->file_size_low being 0 means filesize option isn't set.
1186 * Non zero ->file_size_low equals ->file_size_high means
1187 * filesize option is set in a fixed size format.
1188 * Non zero ->file_size_low not equals ->file_size_high means
1189 * filesize option is set in a range format.
1191 if (!o->file_size_low) {
1193 * no file size or range given, file size is equal to
1194 * total size divided by number of files. If the size
1195 * doesn't divide nicely with the min blocksize,
1196 * make the first files bigger.
1205 * We normally don't come here for regular files, but
1206 * if the result is 0 for a regular file, set it to the
1207 * real file size. This could be size of the existing
1208 * one if it already exists, but otherwise will be set
1209 * to 0. A new file won't be created because
1210 * ->io_size + ->file_offset equals ->real_file_size.
1213 if (f->file_offset > f->real_file_size)
1215 f->io_size = f->real_file_size - f->file_offset;
1217 log_info("fio: file %s may be ignored\n",
1220 } else if (f->real_file_size < o->file_size_low ||
1221 f->real_file_size > o->file_size_high) {
1222 if (f->file_offset > o->file_size_low)
1225 * file size given. if it's fixed, use that. if it's a
1226 * range, generate a random size in-between.
1228 if (o->file_size_low == o->file_size_high)
1229 f->io_size = o->file_size_low - f->file_offset;
1231 f->io_size = get_rand_file_size(td)
1235 f->io_size = f->real_file_size - f->file_offset;
1237 if (f->io_size == -1ULL)
1242 if (o->size_percent && o->size_percent != 100) {
1245 file_size = f->io_size + f->file_offset;
1246 f->io_size = (file_size *
1247 o->size_percent) / 100;
1248 if (f->io_size > (file_size - f->file_offset))
1249 f->io_size = file_size - f->file_offset;
1251 f->io_size -= (f->io_size % td_min_bs(td));
1254 io_size = f->io_size;
1255 if (o->io_size_percent && o->io_size_percent != 100) {
1256 io_size *= o->io_size_percent;
1260 total_size += io_size;
1263 if (f->filetype == FIO_TYPE_FILE &&
1264 (f->io_size + f->file_offset) > f->real_file_size) {
1265 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1266 !o->create_on_open) {
1268 extend_size += (f->io_size + f->file_offset);
1269 fio_file_set_extend(f);
1270 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1271 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1272 td_ioengine_flagged(td, FIO_FAKEIO)))
1273 f->real_file_size = f->io_size + f->file_offset;
1277 if (td->o.block_error_hist) {
1280 assert(td->o.nr_files == 1); /* checked in fixup_options */
1282 len = f->io_size / td->o.bs[DDIR_TRIM];
1283 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1284 log_err("fio: cannot calculate block histogram with "
1285 "%d trim blocks, maximum %d\n",
1286 len, MAX_NR_BLOCK_INFOS);
1287 td_verror(td, EINVAL, "block_error_hist");
1291 td->ts.nr_block_infos = len;
1292 for (i = 0; i < len; i++)
1293 td->ts.block_infos[i] =
1294 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1296 td->ts.nr_block_infos = 0;
1298 if (!o->size || (total_size && o->size > total_size))
1299 o->size = total_size;
1301 if (o->size < td_min_bs(td)) {
1302 log_err("fio: blocksize is larger than data set range\n");
1307 * See if we need to extend some files, typically needed when our
1308 * target regular files don't exist yet, but our jobs require them
1309 * initially due to read I/Os.
1313 if (output_format & FIO_OUTPUT_NORMAL) {
1314 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1316 need_extend > 1 ? "s" : "",
1318 need_extend > 1 ? "s" : "",
1319 need_extend > 1 ? "total " : "",
1323 for_each_file(td, f, i) {
1324 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1326 if (!fio_file_extend(f))
1329 assert(f->filetype == FIO_TYPE_FILE);
1330 fio_file_clear_extend(f);
1331 if (!o->fill_device) {
1332 old_len = f->real_file_size;
1333 extend_len = f->io_size + f->file_offset -
1336 f->real_file_size = (f->io_size + f->file_offset);
1337 err = extend_file(td, f);
1341 err = __file_invalidate_cache(td, f, old_len,
1345 * Shut up static checker
1361 * Prepopulate files with data. It might be expected to read some
1362 * "real" data instead of zero'ed files (if no writes to file occurred
1363 * prior to a read job). Engine has to provide a way to do that.
1365 if (td->io_ops->prepopulate_file) {
1368 for_each_file(td, f, i) {
1369 if (output_format & FIO_OUTPUT_NORMAL) {
1370 log_info("%s: Prepopulating IO file (%s)\n",
1371 o->name, f->file_name);
1374 err = td->io_ops->prepopulate_file(td, f);
1378 err = __file_invalidate_cache(td, f, f->file_offset,
1382 * Shut up static checker
1398 * iolog already set the total io size, if we read back
1401 if (!o->read_iolog_file) {
1403 td->total_io_size = o->io_size * o->loops;
1405 td->total_io_size = o->size * o->loops;
1409 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1410 err = zbd_setup_files(td);
1418 td_restore_runstate(td, old_state);
1423 log_err("%s: you need to specify valid offset=\n", o->name);
1425 td_restore_runstate(td, old_state);
1429 bool pre_read_files(struct thread_data *td)
1434 dprint(FD_FILE, "pre_read files\n");
1436 for_each_file(td, f, i) {
1437 if (!pre_read_file(td, f))
1444 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1446 unsigned int range_size, seed;
1450 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1451 fsize = min(f->real_file_size, f->io_size);
1453 nranges = (fsize + range_size - 1ULL) / range_size;
1455 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1456 if (!td->o.rand_repeatable)
1457 seed = td->rand_seeds[4];
1459 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1460 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1461 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1462 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1463 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1464 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1467 static bool init_rand_distribution(struct thread_data *td)
1473 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1474 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1475 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1478 state = td_bump_runstate(td, TD_SETTING_UP);
1480 for_each_file(td, f, i)
1481 __init_rand_distribution(td, f);
1483 td_restore_runstate(td, state);
1488 * Check if the number of blocks exceeds the randomness capability of
1489 * the selected generator. Tausworthe is 32-bit, the others are fullly
1492 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1495 if (blocks <= FRAND32_MAX)
1497 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1501 * If the user hasn't specified a random generator, switch
1502 * to tausworthe64 with informational warning. If the user did
1503 * specify one, just warn.
1505 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1508 if (!fio_option_is_set(&td->o, random_generator)) {
1509 log_info("fio: Switching to tausworthe64. Use the "
1510 "random_generator= option to get rid of this "
1512 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1517 * Just make this information to avoid breaking scripts.
1519 log_info("fio: Use the random_generator= option to switch to lfsr or "
1524 bool init_random_map(struct thread_data *td)
1526 unsigned long long blocks;
1530 if (init_rand_distribution(td))
1535 for_each_file(td, f, i) {
1536 uint64_t fsize = min(f->real_file_size, f->io_size);
1538 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1539 fsize = td->o.zone_range;
1541 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1543 if (check_rand_gen_limits(td, f, blocks))
1546 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1549 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1551 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1552 fio_file_set_lfsr(f);
1555 log_err("fio: failed initializing LFSR\n");
1558 } else if (!td->o.norandommap) {
1559 f->io_axmap = axmap_new(blocks);
1561 fio_file_set_axmap(f);
1564 } else if (td->o.norandommap)
1567 if (!td->o.softrandommap) {
1568 log_err("fio: failed allocating random map. If running"
1569 " a large number of jobs, try the 'norandommap'"
1570 " option or set 'softrandommap'. Or give"
1571 " a larger --alloc-size to fio.\n");
1575 log_info("fio: file %s failed allocating random map. Running "
1576 "job without.\n", f->file_name);
1582 void close_files(struct thread_data *td)
1587 for_each_file(td, f, i) {
1588 if (fio_file_open(f))
1589 td_io_close_file(td, f);
1593 void fio_file_free(struct fio_file *f)
1595 if (fio_file_axmap(f))
1596 axmap_free(f->io_axmap);
1597 if (!fio_file_smalloc(f)) {
1601 sfree(f->file_name);
1606 void close_and_free_files(struct thread_data *td)
1611 dprint(FD_FILE, "close files\n");
1613 for_each_file(td, f, i) {
1614 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1615 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1616 td_io_unlink_file(td, f);
1619 if (fio_file_open(f))
1620 td_io_close_file(td, f);
1622 remove_file_hash(f);
1624 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1625 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1626 td_io_unlink_file(td, f);
1633 td->o.filename = NULL;
1635 free(td->file_locks);
1636 td->files_index = 0;
1638 td->file_locks = NULL;
1639 td->o.file_lock_mode = FILE_LOCK_NONE;
1643 static void get_file_type(struct fio_file *f)
1647 if (!strcmp(f->file_name, "-"))
1648 f->filetype = FIO_TYPE_PIPE;
1650 f->filetype = FIO_TYPE_FILE;
1653 /* \\.\ is the device namespace in Windows, where every file is
1655 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1656 f->filetype = FIO_TYPE_BLOCK;
1659 if (!stat(f->file_name, &sb)) {
1660 if (S_ISBLK(sb.st_mode))
1661 f->filetype = FIO_TYPE_BLOCK;
1662 else if (S_ISCHR(sb.st_mode))
1663 f->filetype = FIO_TYPE_CHAR;
1664 else if (S_ISFIFO(sb.st_mode))
1665 f->filetype = FIO_TYPE_PIPE;
1669 static bool __is_already_allocated(const char *fname, bool set)
1671 struct flist_head *entry;
1674 ret = file_bloom_exists(fname, set);
1678 flist_for_each(entry, &filename_list) {
1679 struct file_name *fn;
1681 fn = flist_entry(entry, struct file_name, list);
1683 if (!strcmp(fn->filename, fname))
1690 static bool is_already_allocated(const char *fname)
1694 fio_file_hash_lock();
1695 ret = __is_already_allocated(fname, false);
1696 fio_file_hash_unlock();
1701 static void set_already_allocated(const char *fname)
1703 struct file_name *fn;
1705 fn = malloc(sizeof(struct file_name));
1706 fn->filename = strdup(fname);
1708 fio_file_hash_lock();
1709 if (!__is_already_allocated(fname, true)) {
1710 flist_add_tail(&fn->list, &filename_list);
1713 fio_file_hash_unlock();
1721 static void free_already_allocated(void)
1723 struct flist_head *entry, *tmp;
1724 struct file_name *fn;
1726 if (flist_empty(&filename_list))
1729 fio_file_hash_lock();
1730 flist_for_each_safe(entry, tmp, &filename_list) {
1731 fn = flist_entry(entry, struct file_name, list);
1733 flist_del(&fn->list);
1737 fio_file_hash_unlock();
1740 static struct fio_file *alloc_new_file(struct thread_data *td)
1744 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1745 f = calloc(1, sizeof(*f));
1747 f = scalloc(1, sizeof(*f));
1755 fio_file_reset(td, f);
1756 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1757 fio_file_set_smalloc(f);
1761 bool exists_and_not_regfile(const char *filename)
1765 if (lstat(filename, &sb) == -1)
1768 #ifndef WIN32 /* NOT Windows */
1769 if (S_ISREG(sb.st_mode))
1772 /* \\.\ is the device namespace in Windows, where every file
1773 * is a device node */
1774 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1781 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1783 int cur_files = td->files_index;
1784 char file_name[PATH_MAX];
1788 dprint(FD_FILE, "add file %s\n", fname);
1790 if (td->o.directory)
1791 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1792 td->o.unique_filename);
1794 sprintf(file_name + len, "%s", fname);
1796 /* clean cloned siblings using existing files */
1797 if (numjob && is_already_allocated(file_name) &&
1798 !exists_and_not_regfile(fname))
1801 f = alloc_new_file(td);
1803 if (td->files_size <= td->files_index) {
1804 unsigned int new_size = td->o.nr_files + 1;
1806 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1808 td->files = realloc(td->files, new_size * sizeof(f));
1809 if (td->files == NULL) {
1810 log_err("fio: realloc OOM\n");
1813 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1814 td->file_locks = realloc(td->file_locks, new_size);
1815 if (!td->file_locks) {
1816 log_err("fio: realloc OOM\n");
1819 td->file_locks[cur_files] = FILE_LOCK_NONE;
1821 td->files_size = new_size;
1823 td->files[cur_files] = f;
1824 f->fileno = cur_files;
1827 * init function, io engine may not be loaded yet
1829 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1830 f->real_file_size = -1ULL;
1832 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1833 f->file_name = strdup(file_name);
1835 f->file_name = smalloc_strdup(file_name);
1837 /* can't handle smalloc failure from here */
1838 assert(f->file_name);
1842 switch (td->o.file_lock_mode) {
1843 case FILE_LOCK_NONE:
1845 case FILE_LOCK_READWRITE:
1846 f->rwlock = fio_rwlock_init();
1848 case FILE_LOCK_EXCLUSIVE:
1849 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1852 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1858 if (td->o.numjobs > 1)
1859 set_already_allocated(file_name);
1864 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1870 int add_file_exclusive(struct thread_data *td, const char *fname)
1875 for_each_file(td, f, i) {
1876 if (!strcmp(f->file_name, fname))
1880 return add_file(td, fname, 0, 1);
1883 void get_file(struct fio_file *f)
1885 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1886 assert(fio_file_open(f));
1890 int put_file(struct thread_data *td, struct fio_file *f)
1892 int f_ret = 0, ret = 0;
1894 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1896 if (!fio_file_open(f)) {
1897 assert(f->fd == -1);
1901 assert(f->references);
1902 if (--f->references)
1905 disk_util_dec(f->du);
1907 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1908 unlock_file_all(td, f);
1910 if (should_fsync(td) && td->o.fsync_on_close) {
1911 f_ret = fsync(f->fd);
1916 if (td->io_ops->close_file)
1917 ret = td->io_ops->close_file(td, f);
1922 td->nr_open_files--;
1923 fio_file_clear_closing(f);
1924 fio_file_clear_open(f);
1925 assert(f->fd == -1);
1929 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1931 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1934 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1935 if (ddir == DDIR_READ)
1936 fio_rwlock_read(f->rwlock);
1938 fio_rwlock_write(f->rwlock);
1939 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1940 fio_sem_down(f->lock);
1942 td->file_locks[f->fileno] = td->o.file_lock_mode;
1945 void unlock_file(struct thread_data *td, struct fio_file *f)
1947 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1950 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1951 fio_rwlock_unlock(f->rwlock);
1952 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1953 fio_sem_up(f->lock);
1955 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1958 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1960 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1962 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1966 static bool recurse_dir(struct thread_data *td, const char *dirname)
1972 D = opendir(dirname);
1974 char buf[FIO_VERROR_SIZE];
1976 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1977 td_verror(td, errno, buf);
1981 while ((dir = readdir(D)) != NULL) {
1982 char full_path[PATH_MAX];
1985 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1988 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1990 if (lstat(full_path, &sb) == -1) {
1991 if (errno != ENOENT) {
1992 td_verror(td, errno, "stat");
1998 if (S_ISREG(sb.st_mode)) {
1999 add_file(td, full_path, 0, 1);
2002 if (!S_ISDIR(sb.st_mode))
2005 ret = recurse_dir(td, full_path);
2014 int add_dir_files(struct thread_data *td, const char *path)
2016 int ret = recurse_dir(td, path);
2019 log_info("fio: opendir added %d files\n", td->o.nr_files);
2024 void dup_files(struct thread_data *td, struct thread_data *org)
2029 dprint(FD_FILE, "dup files: %d\n", org->files_index);
2034 td->files = malloc(org->files_index * sizeof(f));
2036 if (td->o.file_lock_mode != FILE_LOCK_NONE)
2037 td->file_locks = malloc(org->files_index);
2039 for_each_file(org, f, i) {
2040 struct fio_file *__f;
2042 __f = alloc_new_file(td);
2045 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2046 __f->file_name = strdup(f->file_name);
2048 __f->file_name = smalloc_strdup(f->file_name);
2050 /* can't handle smalloc failure from here */
2051 assert(__f->file_name);
2052 __f->filetype = f->filetype;
2055 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2056 __f->lock = f->lock;
2057 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2058 __f->rwlock = f->rwlock;
2065 * Returns the index that matches the filename, or -1 if not there
2067 int get_fileno(struct thread_data *td, const char *fname)
2072 for_each_file(td, f, i)
2073 if (!strcmp(f->file_name, fname))
2080 * For log usage, where we add/open/close files automatically
2082 void free_release_files(struct thread_data *td)
2086 td->o.open_files = 0;
2087 td->files_index = 0;
2090 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2094 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2095 f->last_pos[i] = f->file_offset;
2096 f->last_start[i] = -1ULL;
2099 if (fio_file_axmap(f))
2100 axmap_reset(f->io_axmap);
2101 else if (fio_file_lfsr(f))
2102 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2104 zbd_file_reset(td, f);
2107 bool fio_files_done(struct thread_data *td)
2112 for_each_file(td, f, i)
2113 if (!fio_file_done(f))
2119 /* free memory used in initialization phase only */
2120 void filesetup_mem_free(void)
2122 free_already_allocated();
2126 * This function is for platforms which support direct I/O but not O_DIRECT.
2128 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2130 #ifdef FIO_OS_DIRECTIO
2131 int ret = fio_set_odirect(f);
2134 td_verror(td, ret, "fio_set_directio");
2135 #if defined(__sun__)
2136 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2137 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2139 log_err("fio: the file system does not seem to support direct IO\n");
2142 log_err("fio: the file system does not seem to support direct IO\n");
2149 log_err("fio: direct IO is not supported on this host operating system\n");