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");
313 if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
314 td_verror(td, errno, "lseek");
315 log_err("fio: failed to lseek pre-read file\n");
320 while (left && !td->terminate) {
324 r = read(f->fd, b, bs);
330 td_verror(td, EIO, "pre_read");
336 td_restore_runstate(td, old_runstate);
339 td->io_ops->close_file(td, f);
346 * Generic function to prepopulate regular file with data.
347 * Useful if you want to make sure I/O engine has data to read.
348 * Leaves f->fd open on success, caller must close.
350 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
353 unsigned long long left, bs;
356 /* generic function for regular files only */
357 assert(f->filetype == FIO_TYPE_FILE);
360 log_err("fio: refusing to write a file due to read-only\n");
365 if (td->o.allow_create)
372 dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
373 f->fd = open(f->file_name, flags, 0644);
377 if (err == ENOENT && !td->o.allow_create)
378 log_err("fio: file creation disallowed by "
379 "allow_file_create=0\n");
381 td_verror(td, err, "open");
385 left = f->real_file_size;
386 bs = td->o.max_bs[DDIR_WRITE];
392 td_verror(td, errno, "malloc");
396 while (left && !td->terminate) {
402 fill_io_buffer(td, b, bs, bs);
404 r = write(f->fd, b, bs);
409 td_verror(td, errno, "write");
415 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
416 td_io_unlink_file(td, f);
417 } else if (td->o.create_fsync) {
418 if (fsync(f->fd) < 0) {
419 td_verror(td, errno, "fsync");
434 unsigned long long get_rand_file_size(struct thread_data *td)
436 unsigned long long ret, sized;
440 frand_max = rand_max(&td->file_size_state);
441 r = __rand(&td->file_size_state);
442 sized = td->o.file_size_high - td->o.file_size_low;
443 ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
444 ret += td->o.file_size_low;
445 ret -= (ret % td->o.rw_min_bs);
449 static int file_size(struct thread_data *td, struct fio_file *f)
453 if (stat(f->file_name, &st) == -1) {
454 td_verror(td, errno, "fstat");
458 f->real_file_size = st.st_size;
462 static int bdev_size(struct thread_data *td, struct fio_file *f)
464 unsigned long long bytes = 0;
467 if (td->io_ops->open_file(td, f)) {
468 log_err("fio: failed opening blockdev %s for size check\n",
473 r = blockdev_size(f, &bytes);
475 td_verror(td, r, "blockdev_size");
480 log_err("%s: zero sized block device?\n", f->file_name);
484 f->real_file_size = bytes;
485 td->io_ops->close_file(td, f);
488 td->io_ops->close_file(td, f);
492 static int char_size(struct thread_data *td, struct fio_file *f)
494 #ifdef FIO_HAVE_CHARDEV_SIZE
495 unsigned long long bytes = 0;
498 if (td->io_ops->open_file(td, f)) {
499 log_err("fio: failed opening chardev %s for size check\n",
504 r = chardev_size(f, &bytes);
506 td_verror(td, r, "chardev_size");
511 log_err("%s: zero sized char device?\n", f->file_name);
515 f->real_file_size = bytes;
516 td->io_ops->close_file(td, f);
519 td->io_ops->close_file(td, f);
522 f->real_file_size = -1ULL;
527 static int get_file_size(struct thread_data *td, struct fio_file *f)
531 if (fio_file_size_known(f))
534 if (f->filetype == FIO_TYPE_FILE)
535 ret = file_size(td, f);
536 else if (f->filetype == FIO_TYPE_BLOCK)
537 ret = bdev_size(td, f);
538 else if (f->filetype == FIO_TYPE_CHAR)
539 ret = char_size(td, f);
541 f->real_file_size = -1;
542 log_info("%s: failed to get file size of %s\n", td->o.name,
544 return 1; /* avoid offset extends end error message */
548 * Leave ->real_file_size with 0 since it could be expectation
549 * of initial setup for regular files.
555 * ->file_offset normally hasn't been initialized yet, so this
556 * is basically always false unless ->real_file_size is -1, but
557 * if ->real_file_size is -1 this message doesn't make sense.
558 * As a result, this message is basically useless.
560 if (f->file_offset > f->real_file_size) {
561 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
562 (unsigned long long) f->file_offset,
563 (unsigned long long) f->real_file_size);
567 fio_file_set_size_known(f);
571 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
572 unsigned long long off,
573 unsigned long long len)
575 int errval = 0, ret = 0;
584 off = f->file_offset;
586 if (len == -1ULL || off == -1ULL)
589 if (td->io_ops->invalidate) {
590 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
592 ret = td->io_ops->invalidate(td, f);
595 } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
596 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
598 } else if (f->filetype == FIO_TYPE_FILE) {
599 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
600 f->file_name, off, len);
601 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
604 } else if (f->filetype == FIO_TYPE_BLOCK) {
607 dprint(FD_IO, "drop page cache %s\n", f->file_name);
608 ret = blockdev_invalidate_cache(f);
609 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
611 * Linux multipath devices reject ioctl while
612 * the maps are being updated. That window can
613 * last tens of milliseconds; we'll try up to
614 * a quarter of a second.
617 ret = blockdev_invalidate_cache(f);
619 if (ret < 0 && errno == EACCES && geteuid()) {
620 if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
621 log_err("fio: only root may flush block "
622 "devices. Cache flush bypassed!\n");
627 } else if (f->filetype == FIO_TYPE_CHAR ||
628 f->filetype == FIO_TYPE_PIPE) {
629 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
633 * Cache flushing isn't a fatal condition, and we know it will
634 * happen on some platforms where we don't have the proper
635 * function to flush eg block device caches. So just warn and
636 * continue on our way.
639 log_info("fio: cache invalidation of %s failed: %s\n",
640 f->file_name, strerror(errval));
646 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
648 if (!fio_file_open(f))
651 return __file_invalidate_cache(td, f, -1ULL, -1ULL);
654 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
658 dprint(FD_FILE, "fd close %s\n", f->file_name);
662 if (close(f->fd) < 0)
667 if (f->shadow_fd != -1) {
676 int file_lookup_open(struct fio_file *f, int flags)
678 struct fio_file *__f;
681 __f = lookup_file_hash(f->file_name);
683 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
687 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
695 f->fd = open(f->file_name, flags, 0600);
699 static int file_close_shadow_fds(struct thread_data *td)
705 for_each_file(td, f, i) {
706 if (f->shadow_fd == -1)
717 int generic_open_file(struct thread_data *td, struct fio_file *f)
723 dprint(FD_FILE, "fd open %s\n", f->file_name);
725 if (!strcmp(f->file_name, "-")) {
727 log_err("fio: can't read/write to stdin/out\n");
733 * move output logging to stderr, if we are writing to stdout
740 flags |= OS_O_DIRECT;
741 flags |= td->o.sync_io;
742 if (td->o.create_on_open && td->o.allow_create)
744 if (f->filetype != FIO_TYPE_FILE)
745 flags |= FIO_O_NOATIME;
752 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
756 f->fd = dup(STDOUT_FILENO);
758 from_hash = file_lookup_open(f, flags);
759 } else if (td_read(td)) {
760 if (td_ioengine_flagged(td, FIO_RO_NEEDS_RW_OPEN) && !read_only)
766 f->fd = dup(STDIN_FILENO);
768 from_hash = file_lookup_open(f, flags);
769 } else if (td_trim(td)) {
770 assert(!td_rw(td)); /* should have matched above */
773 from_hash = file_lookup_open(f, flags);
777 char buf[FIO_VERROR_SIZE];
780 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
781 flags &= ~FIO_O_NOATIME;
784 if (__e == EMFILE && file_close_shadow_fds(td))
787 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
789 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
790 log_err("fio: looks like your file system does not " \
791 "support direct=1/buffered=0\n");
794 td_verror(td, __e, buf);
798 if (!from_hash && f->fd != -1) {
799 if (add_file_hash(f)) {
803 * Stash away descriptor for later close. This is to
804 * work-around a "feature" on Linux, where a close of
805 * an fd that has been opened for write will trigger
806 * udev to call blkid to check partitions, fs id, etc.
807 * That pollutes the device cache, which can slow down
808 * unbuffered accesses.
810 if (f->shadow_fd == -1)
811 f->shadow_fd = f->fd;
814 * OK to ignore, we haven't done anything
817 ret = generic_close_file(td, f);
827 * This function i.e. get_file_size() is the default .get_file_size
828 * implementation of majority of I/O engines.
830 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
832 return get_file_size(td, f);
836 * open/close all files, so that ->real_file_size gets set
838 static int get_file_sizes(struct thread_data *td)
844 for_each_file(td, f, i) {
845 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
848 if (td_io_get_file_size(td, f)) {
849 if (td->error != ENOENT) {
850 log_err("%s\n", td->verror);
858 * There are corner cases where we end up with -1 for
859 * ->real_file_size due to unsupported file type, etc.
860 * We then just set to size option value divided by number
861 * of files, similar to the way file ->io_size is set.
862 * stat(2) failure doesn't set ->real_file_size to -1.
864 if (f->real_file_size == -1ULL && td->o.size)
865 f->real_file_size = td->o.size / td->o.nr_files;
872 struct flist_head list;
879 * Get free number of bytes for each file on each unique mount.
881 static unsigned long long get_fs_free_counts(struct thread_data *td)
883 struct flist_head *n, *tmp;
884 unsigned long long ret = 0;
885 struct fio_mount *fm;
890 for_each_file(td, f, i) {
894 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
895 if (f->real_file_size != -1ULL)
896 ret += f->real_file_size;
898 } else if (f->filetype != FIO_TYPE_FILE)
901 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
903 if (stat(buf, &sb) < 0) {
907 if (stat(buf, &sb) < 0)
912 flist_for_each(n, &list) {
913 fm = flist_entry(n, struct fio_mount, list);
914 if (fm->key == sb.st_dev)
923 fm = calloc(1, sizeof(*fm));
924 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
925 fm->base = basename(fm->__base);
927 flist_add(&fm->list, &list);
930 flist_for_each_safe(n, tmp, &list) {
931 unsigned long long sz;
933 fm = flist_entry(n, struct fio_mount, list);
934 flist_del(&fm->list);
936 sz = get_fs_free_size(fm->base);
937 if (sz && sz != -1ULL)
946 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
949 struct thread_options *o = &td->o;
950 unsigned long long align_bs;
951 unsigned long long offset;
952 unsigned long long increment;
954 if (o->file_append && f->filetype == FIO_TYPE_FILE)
955 return f->real_file_size;
957 if (o->offset_increment_percent) {
958 assert(!o->offset_increment);
959 increment = o->offset_increment_percent * f->real_file_size / 100;
962 increment = o->offset_increment;
964 if (o->start_offset_percent > 0) {
965 /* calculate the raw offset */
966 offset = (f->real_file_size * o->start_offset_percent / 100) +
967 (td->subjob_number * increment);
971 /* start_offset_percent not set */
972 offset = o->start_offset +
973 td->subjob_number * increment;
978 * if offset_align is provided, use it
980 if (fio_option_is_set(o, start_offset_align)) {
981 align_bs = o->start_offset_align;
983 /* else take the minimum block size */
984 align_bs = td_min_bs(td);
988 * block align the offset at the next available boundary at
989 * ceiling(offset / align_bs) * align_bs
991 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
998 * Find longest path component that exists and return its length
1000 int longest_existing_path(char *path) {
1011 sprintf(buf, "%s", path);
1014 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1020 *(buf_pos + 1) = '\0';
1023 dwAttr = GetFileAttributesA(buf);
1024 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1028 if (stat(buf, &sb) == 0)
1032 offset = buf_pos - buf;
1040 static bool create_work_dirs(struct thread_data *td, const char *fname)
1042 char path[PATH_MAX];
1046 snprintf(path, PATH_MAX, "%s", fname);
1049 offset = longest_existing_path(path);
1050 end = start + offset;
1051 while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1058 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1059 log_err("fio: failed to create dir (%s): %s\n",
1060 start, strerror(errno));
1063 *end = FIO_OS_PATH_SEPARATOR;
1066 td->flags |= TD_F_DIRS_CREATED;
1071 * Open the files and setup files sizes, creating files if necessary.
1073 int setup_files(struct thread_data *td)
1075 unsigned long long total_size, extend_size;
1076 struct thread_options *o = &td->o;
1078 unsigned int i, nr_fs_extra = 0;
1079 int err = 0, need_extend;
1081 const unsigned long long bs = td_min_bs(td);
1084 dprint(FD_FILE, "setup files\n");
1086 old_state = td_bump_runstate(td, TD_SETTING_UP);
1088 for_each_file(td, f, i) {
1089 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1090 strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1091 !(td->flags & TD_F_DIRS_CREATED) &&
1092 !create_work_dirs(td, f->file_name))
1097 * Find out physical size of files or devices for this thread,
1098 * before we determine I/O size and range of our targets.
1099 * If ioengine defines a setup() method, it's responsible for
1100 * opening the files and setting f->real_file_size to indicate
1101 * the valid range for that file.
1103 if (td->io_ops->setup)
1104 err = td->io_ops->setup(td);
1106 err = get_file_sizes(td);
1111 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1112 err = zbd_init_files(td);
1116 zbd_recalc_options_with_zone_granularity(td);
1118 if (o->read_iolog_file)
1122 * check sizes. if the files/devices do not exist and the size
1123 * isn't passed to fio, abort.
1126 for_each_file(td, f, i) {
1128 if (f->real_file_size == -1ULL)
1131 total_size += f->real_file_size;
1135 td->fill_device_size = get_fs_free_counts(td);
1138 * device/file sizes are zero and no size given, punt
1140 if ((!total_size || total_size == -1ULL) && !o->size &&
1141 !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1142 !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1143 log_err("%s: you need to specify size=\n", o->name);
1144 td_verror(td, EINVAL, "total_file_size");
1149 * Calculate per-file size and potential extra size for the
1150 * first files, if needed (i.e. if we don't have a fixed size).
1152 if (!o->file_size_low && o->nr_files) {
1155 fs = o->size / o->nr_files;
1156 all_fs = fs * o->nr_files;
1158 if (all_fs < o->size)
1159 nr_fs_extra = (o->size - all_fs) / bs;
1163 * now file sizes are known, so we can set ->io_size. if size= is
1164 * not given, ->io_size is just equal to ->real_file_size. if size
1165 * is given, ->io_size is size / nr_files.
1167 extend_size = total_size = 0;
1169 for_each_file(td, f, i) {
1170 f->file_offset = get_start_offset(td, f);
1173 * Update ->io_size depending on options specified.
1174 * ->file_size_low being 0 means filesize option isn't set.
1175 * Non zero ->file_size_low equals ->file_size_high means
1176 * filesize option is set in a fixed size format.
1177 * Non zero ->file_size_low not equals ->file_size_high means
1178 * filesize option is set in a range format.
1180 if (!o->file_size_low) {
1182 * no file size or range given, file size is equal to
1183 * total size divided by number of files. If the size
1184 * doesn't divide nicely with the min blocksize,
1185 * make the first files bigger.
1194 * We normally don't come here for regular files, but
1195 * if the result is 0 for a regular file, set it to the
1196 * real file size. This could be size of the existing
1197 * one if it already exists, but otherwise will be set
1198 * to 0. A new file won't be created because
1199 * ->io_size + ->file_offset equals ->real_file_size.
1202 if (f->file_offset > f->real_file_size)
1204 f->io_size = f->real_file_size - f->file_offset;
1206 log_info("fio: file %s may be ignored\n",
1209 } else if (f->real_file_size < o->file_size_low ||
1210 f->real_file_size > o->file_size_high) {
1211 if (f->file_offset > o->file_size_low)
1214 * file size given. if it's fixed, use that. if it's a
1215 * range, generate a random size in-between.
1217 if (o->file_size_low == o->file_size_high)
1218 f->io_size = o->file_size_low - f->file_offset;
1220 f->io_size = get_rand_file_size(td)
1224 f->io_size = f->real_file_size - f->file_offset;
1226 if (f->io_size == -1ULL)
1231 if (o->size_percent && o->size_percent != 100) {
1234 file_size = f->io_size + f->file_offset;
1235 f->io_size = (file_size *
1236 o->size_percent) / 100;
1237 if (f->io_size > (file_size - f->file_offset))
1238 f->io_size = file_size - f->file_offset;
1240 f->io_size -= (f->io_size % td_min_bs(td));
1243 io_size = f->io_size;
1244 if (o->io_size_percent && o->io_size_percent != 100) {
1245 io_size *= o->io_size_percent;
1249 total_size += io_size;
1252 if (f->filetype == FIO_TYPE_FILE &&
1253 (f->io_size + f->file_offset) > f->real_file_size) {
1254 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1255 !o->create_on_open) {
1257 extend_size += (f->io_size + f->file_offset);
1258 fio_file_set_extend(f);
1259 } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1260 (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1261 td_ioengine_flagged(td, FIO_FAKEIO)))
1262 f->real_file_size = f->io_size + f->file_offset;
1266 if (td->o.block_error_hist) {
1269 assert(td->o.nr_files == 1); /* checked in fixup_options */
1271 len = f->io_size / td->o.bs[DDIR_TRIM];
1272 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1273 log_err("fio: cannot calculate block histogram with "
1274 "%d trim blocks, maximum %d\n",
1275 len, MAX_NR_BLOCK_INFOS);
1276 td_verror(td, EINVAL, "block_error_hist");
1280 td->ts.nr_block_infos = len;
1281 for (i = 0; i < len; i++)
1282 td->ts.block_infos[i] =
1283 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1285 td->ts.nr_block_infos = 0;
1287 if (!o->size || (total_size && o->size > total_size))
1288 o->size = total_size;
1290 if (o->size < td_min_bs(td)) {
1291 log_err("fio: blocksize is larger than data set range\n");
1296 * See if we need to extend some files, typically needed when our
1297 * target regular files don't exist yet, but our jobs require them
1298 * initially due to read I/Os.
1302 if (output_format & FIO_OUTPUT_NORMAL) {
1303 log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1305 need_extend > 1 ? "s" : "",
1307 need_extend > 1 ? "s" : "",
1308 need_extend > 1 ? "total " : "",
1312 for_each_file(td, f, i) {
1313 unsigned long long old_len = -1ULL, extend_len = -1ULL;
1315 if (!fio_file_extend(f))
1318 assert(f->filetype == FIO_TYPE_FILE);
1319 fio_file_clear_extend(f);
1320 if (!o->fill_device) {
1321 old_len = f->real_file_size;
1322 extend_len = f->io_size + f->file_offset -
1325 f->real_file_size = (f->io_size + f->file_offset);
1326 err = extend_file(td, f);
1330 err = __file_invalidate_cache(td, f, old_len,
1334 * Shut up static checker
1350 * Prepopulate files with data. It might be expected to read some
1351 * "real" data instead of zero'ed files (if no writes to file occurred
1352 * prior to a read job). Engine has to provide a way to do that.
1354 if (td->io_ops->prepopulate_file) {
1357 for_each_file(td, f, i) {
1358 if (output_format & FIO_OUTPUT_NORMAL) {
1359 log_info("%s: Prepopulating IO file (%s)\n",
1360 o->name, f->file_name);
1363 err = td->io_ops->prepopulate_file(td, f);
1367 err = __file_invalidate_cache(td, f, f->file_offset,
1371 * Shut up static checker
1387 * iolog already set the total io size, if we read back
1390 if (!o->read_iolog_file) {
1392 td->total_io_size = o->io_size * o->loops;
1394 td->total_io_size = o->size * o->loops;
1398 if (td->o.zone_mode == ZONE_MODE_ZBD) {
1399 err = zbd_setup_files(td);
1407 td_restore_runstate(td, old_state);
1418 log_err("%s: you need to specify valid offset=\n", o->name);
1420 td_restore_runstate(td, old_state);
1424 bool pre_read_files(struct thread_data *td)
1429 dprint(FD_FILE, "pre_read files\n");
1431 for_each_file(td, f, i) {
1432 if (!pre_read_file(td, f))
1439 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1441 unsigned int range_size, seed;
1445 range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1446 fsize = min(f->real_file_size, f->io_size);
1448 nranges = (fsize + range_size - 1ULL) / range_size;
1450 seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number *
1451 td->rand_seeds[FIO_RAND_BLOCK_OFF];
1453 if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1454 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1455 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1456 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1457 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1458 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1461 static bool init_rand_distribution(struct thread_data *td)
1467 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1468 td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1469 td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1472 state = td_bump_runstate(td, TD_SETTING_UP);
1474 for_each_file(td, f, i)
1475 __init_rand_distribution(td, f);
1477 td_restore_runstate(td, state);
1482 * Check if the number of blocks exceeds the randomness capability of
1483 * the selected generator. Tausworthe is 32-bit, the others are fully
1486 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1489 if (blocks <= FRAND32_MAX)
1491 if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1495 * If the user hasn't specified a random generator, switch
1496 * to tausworthe64 with informational warning. If the user did
1497 * specify one, just warn.
1499 log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1502 if (!fio_option_is_set(&td->o, random_generator)) {
1503 log_info("fio: Switching to tausworthe64. Use the "
1504 "random_generator= option to get rid of this "
1506 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1511 * Just make this information to avoid breaking scripts.
1513 log_info("fio: Use the random_generator= option to switch to lfsr or "
1518 bool init_random_map(struct thread_data *td)
1520 unsigned long long blocks;
1524 if (init_rand_distribution(td))
1529 for_each_file(td, f, i) {
1530 uint64_t fsize = min(f->real_file_size, f->io_size);
1532 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1533 fsize = td->o.zone_range;
1535 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1537 if (check_rand_gen_limits(td, f, blocks))
1540 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1543 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1545 if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1546 fio_file_set_lfsr(f);
1549 log_err("fio: failed initializing LFSR\n");
1552 } else if (!td->o.norandommap) {
1553 f->io_axmap = axmap_new(blocks);
1555 fio_file_set_axmap(f);
1558 } else if (td->o.norandommap)
1561 if (!td->o.softrandommap) {
1562 log_err("fio: failed allocating random map. If running"
1563 " a large number of jobs, try the 'norandommap'"
1564 " option or set 'softrandommap'. Or give"
1565 " a larger --alloc-size to fio.\n");
1569 log_info("fio: file %s failed allocating random map. Running "
1570 "job without.\n", f->file_name);
1576 void close_files(struct thread_data *td)
1581 for_each_file(td, f, i) {
1582 if (fio_file_open(f))
1583 td_io_close_file(td, f);
1587 void fio_file_free(struct fio_file *f)
1589 if (fio_file_axmap(f))
1590 axmap_free(f->io_axmap);
1592 sfree(f->ruhs_info);
1593 if (!fio_file_smalloc(f)) {
1597 sfree(f->file_name);
1602 void close_and_free_files(struct thread_data *td)
1607 dprint(FD_FILE, "close files\n");
1609 for_each_file(td, f, i) {
1610 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1611 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1612 td_io_unlink_file(td, f);
1615 if (fio_file_open(f))
1616 td_io_close_file(td, f);
1618 remove_file_hash(f);
1620 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1621 dprint(FD_FILE, "free unlink %s\n", f->file_name);
1622 td_io_unlink_file(td, f);
1626 fdp_free_ruhs_info(f);
1630 td->o.filename = NULL;
1632 free(td->file_locks);
1633 td->files_index = 0;
1635 td->file_locks = NULL;
1636 td->o.file_lock_mode = FILE_LOCK_NONE;
1640 static void get_file_type(struct fio_file *f)
1644 if (!strcmp(f->file_name, "-"))
1645 f->filetype = FIO_TYPE_PIPE;
1647 f->filetype = FIO_TYPE_FILE;
1650 /* \\.\ is the device namespace in Windows, where every file is
1652 if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1653 f->filetype = FIO_TYPE_BLOCK;
1656 if (!stat(f->file_name, &sb)) {
1657 if (S_ISBLK(sb.st_mode))
1658 f->filetype = FIO_TYPE_BLOCK;
1659 else if (S_ISCHR(sb.st_mode))
1660 f->filetype = FIO_TYPE_CHAR;
1661 else if (S_ISFIFO(sb.st_mode))
1662 f->filetype = FIO_TYPE_PIPE;
1666 static bool __is_already_allocated(const char *fname, bool set)
1668 struct flist_head *entry;
1671 ret = file_bloom_exists(fname, set);
1675 flist_for_each(entry, &filename_list) {
1676 struct file_name *fn;
1678 fn = flist_entry(entry, struct file_name, list);
1680 if (!strcmp(fn->filename, fname))
1687 static bool is_already_allocated(const char *fname)
1691 fio_file_hash_lock();
1692 ret = __is_already_allocated(fname, false);
1693 fio_file_hash_unlock();
1698 static void set_already_allocated(const char *fname)
1700 struct file_name *fn;
1702 fn = malloc(sizeof(struct file_name));
1703 fn->filename = strdup(fname);
1705 fio_file_hash_lock();
1706 if (!__is_already_allocated(fname, true)) {
1707 flist_add_tail(&fn->list, &filename_list);
1710 fio_file_hash_unlock();
1718 static void free_already_allocated(void)
1720 struct flist_head *entry, *tmp;
1721 struct file_name *fn;
1723 if (flist_empty(&filename_list))
1726 fio_file_hash_lock();
1727 flist_for_each_safe(entry, tmp, &filename_list) {
1728 fn = flist_entry(entry, struct file_name, list);
1730 flist_del(&fn->list);
1734 fio_file_hash_unlock();
1737 static struct fio_file *alloc_new_file(struct thread_data *td)
1741 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1742 f = calloc(1, sizeof(*f));
1744 f = scalloc(1, sizeof(*f));
1752 fio_file_reset(td, f);
1753 if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1754 fio_file_set_smalloc(f);
1758 bool exists_and_not_regfile(const char *filename)
1762 if (lstat(filename, &sb) == -1)
1765 #ifndef WIN32 /* NOT Windows */
1766 if (S_ISREG(sb.st_mode))
1769 /* \\.\ is the device namespace in Windows, where every file
1770 * is a device node */
1771 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1778 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1780 int cur_files = td->files_index;
1781 char file_name[PATH_MAX];
1785 dprint(FD_FILE, "add file %s\n", fname);
1787 if (td->o.directory)
1788 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1789 td->o.unique_filename);
1791 sprintf(file_name + len, "%s", fname);
1793 /* clean cloned siblings using existing files */
1794 if (numjob && is_already_allocated(file_name) &&
1795 !exists_and_not_regfile(fname))
1798 f = alloc_new_file(td);
1800 if (td->files_size <= td->files_index) {
1801 unsigned int new_size = td->o.nr_files + 1;
1803 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1805 td->files = realloc(td->files, new_size * sizeof(f));
1806 if (td->files == NULL) {
1807 log_err("fio: realloc OOM\n");
1810 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1811 td->file_locks = realloc(td->file_locks, new_size);
1812 if (!td->file_locks) {
1813 log_err("fio: realloc OOM\n");
1816 td->file_locks[cur_files] = FILE_LOCK_NONE;
1818 td->files_size = new_size;
1820 td->files[cur_files] = f;
1821 f->fileno = cur_files;
1824 * init function, io engine may not be loaded yet
1826 if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1827 f->real_file_size = -1ULL;
1829 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1830 f->file_name = strdup(file_name);
1832 f->file_name = smalloc_strdup(file_name);
1834 /* can't handle smalloc failure from here */
1835 assert(f->file_name);
1839 switch (td->o.file_lock_mode) {
1840 case FILE_LOCK_NONE:
1842 case FILE_LOCK_READWRITE:
1843 f->rwlock = fio_rwlock_init();
1845 case FILE_LOCK_EXCLUSIVE:
1846 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1849 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1855 if (td->o.numjobs > 1)
1856 set_already_allocated(file_name);
1861 dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1867 int add_file_exclusive(struct thread_data *td, const char *fname)
1872 for_each_file(td, f, i) {
1873 if (!strcmp(f->file_name, fname))
1877 return add_file(td, fname, 0, 1);
1880 void get_file(struct fio_file *f)
1882 dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1883 assert(fio_file_open(f));
1887 int put_file(struct thread_data *td, struct fio_file *f)
1889 int f_ret = 0, ret = 0;
1891 dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1893 if (!fio_file_open(f)) {
1894 assert(f->fd == -1);
1898 assert(f->references);
1899 if (--f->references)
1902 disk_util_dec(f->du);
1904 if (td->o.file_lock_mode != FILE_LOCK_NONE)
1905 unlock_file_all(td, f);
1907 if (should_fsync(td) && td->o.fsync_on_close) {
1908 f_ret = fsync(f->fd);
1913 if (td->io_ops->close_file)
1914 ret = td->io_ops->close_file(td, f);
1919 td->nr_open_files--;
1920 fio_file_clear_closing(f);
1921 fio_file_clear_open(f);
1922 assert(f->fd == -1);
1926 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1928 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1931 if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1932 if (ddir == DDIR_READ)
1933 fio_rwlock_read(f->rwlock);
1935 fio_rwlock_write(f->rwlock);
1936 } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1937 fio_sem_down(f->lock);
1939 td->file_locks[f->fileno] = td->o.file_lock_mode;
1942 void unlock_file(struct thread_data *td, struct fio_file *f)
1944 if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1947 if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1948 fio_rwlock_unlock(f->rwlock);
1949 else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1950 fio_sem_up(f->lock);
1952 td->file_locks[f->fileno] = FILE_LOCK_NONE;
1955 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1957 if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1959 if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1963 static bool recurse_dir(struct thread_data *td, const char *dirname)
1969 D = opendir(dirname);
1971 char buf[FIO_VERROR_SIZE];
1973 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1974 td_verror(td, errno, buf);
1978 while ((dir = readdir(D)) != NULL) {
1979 char full_path[PATH_MAX];
1982 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1985 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1987 if (lstat(full_path, &sb) == -1) {
1988 if (errno != ENOENT) {
1989 td_verror(td, errno, "stat");
1995 if (S_ISREG(sb.st_mode)) {
1996 add_file(td, full_path, 0, 1);
1999 if (!S_ISDIR(sb.st_mode))
2002 ret = recurse_dir(td, full_path);
2011 int add_dir_files(struct thread_data *td, const char *path)
2013 int ret = recurse_dir(td, path);
2016 log_info("fio: opendir added %d files\n", td->o.nr_files);
2021 void dup_files(struct thread_data *td, struct thread_data *org)
2026 dprint(FD_FILE, "dup files: %d\n", org->files_index);
2031 td->files = calloc(org->files_index, sizeof(f));
2033 if (td->o.file_lock_mode != FILE_LOCK_NONE)
2034 td->file_locks = malloc(org->files_index);
2036 assert(org->files_index >= org->o.nr_files);
2037 for_each_file(org, f, i) {
2038 struct fio_file *__f;
2040 __f = alloc_new_file(td);
2043 if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2044 __f->file_name = strdup(f->file_name);
2046 __f->file_name = smalloc_strdup(f->file_name);
2048 /* can't handle smalloc failure from here */
2049 assert(__f->file_name);
2050 __f->filetype = f->filetype;
2053 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2054 __f->lock = f->lock;
2055 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2056 __f->rwlock = f->rwlock;
2063 * Returns the index that matches the filename, or -1 if not there
2065 int get_fileno(struct thread_data *td, const char *fname)
2070 for_each_file(td, f, i)
2071 if (!strcmp(f->file_name, fname))
2078 * For log usage, where we add/open/close files automatically
2080 void free_release_files(struct thread_data *td)
2084 td->o.open_files = 0;
2085 td->files_index = 0;
2088 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2092 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2093 f->last_pos[i] = f->file_offset;
2094 f->last_start[i] = -1ULL;
2097 if (fio_file_axmap(f))
2098 axmap_reset(f->io_axmap);
2099 else if (fio_file_lfsr(f))
2100 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2102 zbd_file_reset(td, f);
2105 bool fio_files_done(struct thread_data *td)
2110 for_each_file(td, f, i)
2111 if (!fio_file_done(f))
2117 /* free memory used in initialization phase only */
2118 void filesetup_mem_free(void)
2120 free_already_allocated();
2124 * This function is for platforms which support direct I/O but not O_DIRECT.
2126 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2128 #ifdef FIO_OS_DIRECTIO
2129 int ret = fio_set_odirect(f);
2132 td_verror(td, ret, "fio_set_directio");
2133 #if defined(__sun__)
2134 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2135 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2137 log_err("fio: the file system does not seem to support direct IO\n");
2140 log_err("fio: the file system does not seem to support direct IO\n");
2147 log_err("fio: direct IO is not supported on this host operating system\n");